diff --git a/.gitattributes b/.gitattributes
index bed0738c7eeb449bca98b5d2f33c89a1ee56349a..0f9fe5ee3e46832a11d33bd6306b59d1c65d234a 100644
--- a/.gitattributes
+++ b/.gitattributes
@@ -1,4 +1,6 @@
*.7z filter=lfs diff=lfs merge=lfs -text
+*.pdf filter=lfs diff=lfs merge=lfs -text
+*.mov filter=lfs diff=lfs merge=lfs -text
*.arrow filter=lfs diff=lfs merge=lfs -text
*.avro filter=lfs diff=lfs merge=lfs -text
*.bin filter=lfs diff=lfs merge=lfs -text
diff --git a/LICENSE b/LICENSE
new file mode 100644
index 0000000000000000000000000000000000000000..c7e9d3ed2e2c914e536df3a030c676a47099488c
--- /dev/null
+++ b/LICENSE
@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2024 ICTNLP
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
diff --git a/README.md b/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..d9eb8251e70522f81bc2f3f667b79050a955cd0b
--- /dev/null
+++ b/README.md
@@ -0,0 +1,379 @@
+# StreamSpeech
+
+[](https://arxiv.org/abs/2406.03049)
+[](https://ictnlp.github.io/StreamSpeech-site/)
+[](https://huggingface.co/ICTNLP/StreamSpeech_Models/tree/main)
+[](https://hits.seeyoufarm.com)
+
+[](https://x.com/Gorden_Sun/status/1798742796524007845) [](https://x.com/imxiaohu/status/1798999363987124355)
+
+> **Authors**: **[Shaolei Zhang](https://zhangshaolei1998.github.io/), [Qingkai Fang](https://fangqingkai.github.io/), [Shoutao Guo](https://scholar.google.com.hk/citations?user=XwHtPyAAAAAJ&hl), [Zhengrui Ma](https://scholar.google.com.hk/citations?user=dUgq6tEAAAAJ), [Min Zhang](https://scholar.google.com.hk/citations?user=CncXH-YAAAAJ), [Yang Feng*](https://people.ucas.edu.cn/~yangfeng?language=en)**
+
+
+Code for ACL 2024 paper "[StreamSpeech: Simultaneous Speech-to-Speech Translation with Multi-task Learning](https://arxiv.org/pdf/2406.03049)".
+
+
+
+💡**Highlight**:
+1. StreamSpeech achieves **SOTA performance** on both offline and simultaneous speech-to-speech translation.
+2. StreamSpeech performs **streaming ASR**, **simultaneous speech-to-text translation** and **simultaneous speech-to-speech translation** via an "All in One" seamless model.
+3. StreamSpeech can present intermediate results (i.e., ASR or translation results) during simultaneous translation, offering a more comprehensive low-latency communication experience.
+
+## 🔥News
+- [2025.06.17] We are excited to extend the "All-in-One" feature of StreamSpeech to more general multimodal interactions via developing **Stream-Omni**. 👉Refer to [paper](https://arxiv.org/abs/2506.13642), [code & demo](https://github.com/ictnlp/Stream-Omni), [model](https://huggingface.co/ICTNLP/stream-omni-8b) for more details.
+ - Stream-Omni is an GPT-4o-like language-vision-speech chatbot that simultaneously supports interactions across any combination of text, vision, and speech modalities.
+ - Stream-Omni can simultaneously produce intermediate textual results (e.g., ASR transcriptions and model responses) during speech interactions, like the advanced voice service of GPT-4o.
+
+- [2024.06.17] Add [Web GUI demo](./demo), now you can experience StreamSpeech in your local browser.
+- [2024.06.05] [Paper](https://arxiv.org/pdf/2406.03049), [code](https://github.com/ictnlp/StreamSpeech), [models](https://huggingface.co/ICTNLP/StreamSpeech_Models/tree/main) and [demo](https://ictnlp.github.io/StreamSpeech-site/) of StreamSpeech are available!
+
+## ⭐Features
+
+### Support 8 Tasks
+- **Offline**: Speech Recognition (ASR)✅, Speech-to-Text Translation (S2TT)✅, Speech-to-Speech Translation (S2ST)✅, Speech Synthesis (TTS)✅
+- **Simultaneous**: Streaming ASR✅, Simultaneous S2TT✅, Simultaneous S2ST✅, Real-time TTS✅ under any latency (with one model)
+
+### GUI Demo
+
+https://github.com/ictnlp/StreamSpeech/assets/34680227/4d9bdabf-af66-4320-ae7d-0f23e721cd71
+
+ Simultaneously provide ASR, translation, and synthesis results via a seamless model
+
+
+### Case
+
+> **Speech Input**: [example/wavs/common_voice_fr_17301936.mp3](./example/wavs/common_voice_fr_17301936.mp3)
+>
+> **Transcription** (ground truth): jai donc lexpérience des années passées jen dirai un mot tout à lheure
+>
+> **Translation** (ground truth): i therefore have the experience of the passed years i'll say a few words about that later
+
+| StreamSpeech | Simultaneous | Offline |
+| ----------------------------------------------- | ------------------------------------------------------------ | ------------------------------------------------------------ |
+| **Speech Recognition** | jai donc expérience des années passé jen dirairai un mot tout à lheure | jai donc lexpérience des années passé jen dirairai un mot tout à lheure |
+| **Speech-to-Text Translation** | i therefore have an experience of last years i will tell a word later | so i have the experience in the past years i'll say a word later |
+| **Speech-to-Speech Translation** | | |
+| **Text-to-Speech Synthesis** (*incrementally synthesize speech word by word*) | | |
+
+
+
+## ⚙Requirements
+
+- Python == 3.10, PyTorch == 2.0.1, Install fairseq & SimulEval
+
+ ```bash
+ cd fairseq
+ pip install --editable ./ --no-build-isolation
+ cd SimulEval
+ pip install --editable ./
+ ```
+
+## 🚀Quick Start
+
+### 1. Model Download
+
+#### (1) StreamSpeech Models
+
+| Language | UnitY | StreamSpeech (offline) | StreamSpeech (simultaneous) |
+| -------- | ------------------------------------------------------------ | ------------------------------------------------------------ | ------------------------------------------------------------ |
+| Fr-En | unity.fr-en.pt [[Huggingface](https://huggingface.co/ICTNLP/StreamSpeech_Models/blob/main/unity.fr-en.pt)] [[Baidu](https://pan.baidu.com/s/10uGYgl0xTej9FP43iKx7Cg?pwd=nkvu)] | streamspeech.offline.fr-en.pt [[Huggingface](https://huggingface.co/ICTNLP/StreamSpeech_Models/blob/main/streamspeech.offline.fr-en.pt)] [[Baidu](https://pan.baidu.com/s/1GFckHGP5SNLuOEj6mbIWhQ?pwd=pwgq)] | streamspeech.simultaneous.fr-en.pt [[Huggingface](https://huggingface.co/ICTNLP/StreamSpeech_Models/blob/main/streamspeech.simultaneous.fr-en.pt)] [[Baidu](https://pan.baidu.com/s/1edCPFljogyDHgGXkUV8_3w?pwd=8gg3)] |
+| Es-En | unity.es-en.pt [[Huggingface](https://huggingface.co/ICTNLP/StreamSpeech_Models/blob/main/unity.es-en.pt)] [[Baidu](https://pan.baidu.com/s/1RwIEHye8jjw3kiIgrCHA3A?pwd=hde4)] | streamspeech.offline.es-en.pt [[Huggingface](https://huggingface.co/ICTNLP/StreamSpeech_Models/blob/main/streamspeech.offline.es-en.pt)] [[Baidu](https://pan.baidu.com/s/1T89G4NC4J0Ofzcsc8Rt2Ww?pwd=yuhd)] | streamspeech.simultaneous.es-en.pt [[Huggingface](https://huggingface.co/ICTNLP/StreamSpeech_Models/blob/main/streamspeech.simultaneous.es-en.pt)] [[Baidu](https://pan.baidu.com/s/1NbLEVcYWHIdqqLD17P1s9g?pwd=p1pc)] |
+| De-En | unity.de-en.pt [[Huggingface](https://huggingface.co/ICTNLP/StreamSpeech_Models/blob/main/unity.de-en.pt)] [[Baidu](https://pan.baidu.com/s/1Mg_PBeZ5acEDhl5wRJ_-7w?pwd=egvv)] | streamspeech.offline.de-en.pt [[Huggingface](https://huggingface.co/ICTNLP/StreamSpeech_Models/blob/main/streamspeech.offline.de-en.pt)] [[Baidu](https://pan.baidu.com/s/1mTE4eHuVLJPB7Yg9AackEg?pwd=6ga8)] | streamspeech.simultaneous.de-en.pt [[Huggingface](https://huggingface.co/ICTNLP/StreamSpeech_Models/blob/main/streamspeech.simultaneous.de-en.pt)] [[Baidu](https://pan.baidu.com/s/1DYPMg3mdDopLY70BYQTduQ?pwd=r7kw)] |
+
+#### (2) Unit-based HiFi-GAN Vocoder
+
+| Unit config | Unit size | Vocoder language | Dataset | Model |
+| ----------------- | --------- | ---------------- | --------------------------------------------------- | ------------------------------------------------------------ |
+| mHuBERT, layer 11 | 1000 | En | [LJSpeech](https://keithito.com/LJ-Speech-Dataset/) | [ckpt](https://dl.fbaipublicfiles.com/fairseq/speech_to_speech/vocoder/code_hifigan/mhubert_vp_en_es_fr_it3_400k_layer11_km1000_lj/g_00500000), [config](https://dl.fbaipublicfiles.com/fairseq/speech_to_speech/vocoder/code_hifigan/mhubert_vp_en_es_fr_it3_400k_layer11_km1000_lj/config.json) |
+
+### 2. Prepare Data and Config (only for test/inference)
+
+#### (1) Config Files
+
+Replace `/data/zhangshaolei/StreamSpeech` in files [configs/fr-en/config_gcmvn.yaml](./configs/fr-en/config_gcmvn.yaml) and [configs/fr-en/config_mtl_asr_st_ctcst.yaml](./configs/fr-en/config_mtl_asr_st_ctcst.yaml) with your local address of StreamSpeech repo.
+
+#### (2) Test Data
+
+Prepare test data following [SimulEval](https://github.com/facebookresearch/SimulEval) format. [example/](./example) provides an example:
+
+- [wav_list.txt](./example/wav_list.txt): Each line records the path of a source speech.
+- [target.txt](./example/target.txt): Each line records the reference text, e.g., target translation or source transcription (used to calculate the metrics).
+
+### 3. Inference with SimulEval
+
+Run these scripts to inference StreamSpeech on streaming ASR, simultaneous S2TT and simultaneous S2ST.
+
+> `--source-segment-size`: set the chunk size (millisecond) to any value to control the latency
+
+
+Simultaneous Speech-to-Speech Translation
+
+`--output-asr-translation`: whether to output the intermediate ASR and translated text results during simultaneous speech-to-speech translation.
+
+```shell
+export CUDA_VISIBLE_DEVICES=0
+
+ROOT=/data/zhangshaolei/StreamSpeech # path to StreamSpeech repo
+PRETRAIN_ROOT=/data/zhangshaolei/pretrain_models
+VOCODER_CKPT=$PRETRAIN_ROOT/unit-based_HiFi-GAN_vocoder/mHuBERT.layer11.km1000.en/g_00500000 # path to downloaded Unit-based HiFi-GAN Vocoder
+VOCODER_CFG=$PRETRAIN_ROOT/unit-based_HiFi-GAN_vocoder/mHuBERT.layer11.km1000.en/config.json # path to downloaded Unit-based HiFi-GAN Vocoder
+
+LANG=fr
+file=streamspeech.simultaneous.${LANG}-en.pt # path to downloaded StreamSpeech model
+output_dir=$ROOT/res/streamspeech.simultaneous.${LANG}-en/simul-s2st
+
+chunk_size=320 #ms
+PYTHONPATH=$ROOT/fairseq simuleval --data-bin ${ROOT}/configs/${LANG}-en \
+ --user-dir ${ROOT}/researches/ctc_unity --agent-dir ${ROOT}/agent \
+ --source example/wav_list.txt --target example/target.txt \
+ --model-path $file \
+ --config-yaml config_gcmvn.yaml --multitask-config-yaml config_mtl_asr_st_ctcst.yaml \
+ --agent $ROOT/agent/speech_to_speech.streamspeech.agent.py \
+ --vocoder $VOCODER_CKPT --vocoder-cfg $VOCODER_CFG --dur-prediction \
+ --output $output_dir/chunk_size=$chunk_size \
+ --source-segment-size $chunk_size \
+ --quality-metrics ASR_BLEU --target-speech-lang en --latency-metrics AL AP DAL StartOffset EndOffset LAAL ATD NumChunks DiscontinuitySum DiscontinuityAve DiscontinuityNum RTF \
+ --device gpu --computation-aware \
+ --output-asr-translation True
+```
+
+You should get the following outputs:
+
+```
+fairseq plugins loaded...
+fairseq plugins loaded...
+fairseq plugins loaded...
+fairseq plugins loaded...
+2024-06-06 09:45:46 | INFO | fairseq.tasks.speech_to_speech | dictionary size: 1,004
+import agents...
+Removing weight norm...
+2024-06-06 09:45:50 | INFO | agent.tts.vocoder | loaded CodeHiFiGAN checkpoint from /data/zhangshaolei/pretrain_models/unit-based_HiFi-GAN_vocoder/mHuBERT.layer11.km1000.en/g_00500000
+2024-06-06 09:45:50 | INFO | simuleval.utils.agent | System will run on device: gpu.
+2024-06-06 09:45:50 | INFO | simuleval.dataloader | Evaluating from speech to speech.
+ 0%| | 0/2 [00:00
+
+
+Simultaneous Speech-to-Text Translation
+
+```shell
+export CUDA_VISIBLE_DEVICES=0
+
+ROOT=/data/zhangshaolei/StreamSpeech # path to StreamSpeech repo
+
+LANG=fr
+file=streamspeech.simultaneous.${LANG}-en.pt # path to downloaded StreamSpeech model
+output_dir=$ROOT/res/streamspeech.simultaneous.${LANG}-en/simul-s2tt
+
+chunk_size=320 #ms
+PYTHONPATH=$ROOT/fairseq simuleval --data-bin ${ROOT}/configs/${LANG}-en \
+ --user-dir ${ROOT}/researches/ctc_unity --agent-dir ${ROOT}/agent \
+ --source example/wav_list.txt --target example/target.txt \
+ --model-path $file \
+ --config-yaml config_gcmvn.yaml --multitask-config-yaml config_mtl_asr_st_ctcst.yaml \
+ --agent $ROOT/agent/speech_to_text.s2tt.streamspeech.agent.py\
+ --output $output_dir/chunk_size=$chunk_size \
+ --source-segment-size $chunk_size \
+ --quality-metrics BLEU --latency-metrics AL AP DAL StartOffset EndOffset LAAL ATD NumChunks RTF \
+ --device gpu --computation-aware
+```
+
+
+
+Streaming ASR
+
+```shell
+export CUDA_VISIBLE_DEVICES=0
+
+ROOT=/data/zhangshaolei/StreamSpeech # path to StreamSpeech repo
+
+LANG=fr
+file=streamspeech.simultaneous.${LANG}-en.pt # path to downloaded StreamSpeech model
+output_dir=$ROOT/res/streamspeech.simultaneous.${LANG}-en/streaming-asr
+
+chunk_size=320 #ms
+PYTHONPATH=$ROOT/fairseq simuleval --data-bin ${ROOT}/configs/${LANG}-en \
+ --user-dir ${ROOT}/researches/ctc_unity --agent-dir ${ROOT}/agent \
+ --source example/wav_list.txt --target example/source.txt \
+ --model-path $file \
+ --config-yaml config_gcmvn.yaml --multitask-config-yaml config_mtl_asr_st_ctcst.yaml \
+ --agent $ROOT/agent/speech_to_text.asr.streamspeech.agent.py\
+ --output $output_dir/chunk_size=$chunk_size \
+ --source-segment-size $chunk_size \
+ --quality-metrics BLEU --latency-metrics AL AP DAL StartOffset EndOffset LAAL ATD NumChunks RTF \
+ --device gpu --computation-aware
+```
+
+
+## 🎈Develop Your Own StreamSpeech
+
+### 1. Data Preprocess
+
+- Follow [`./preprocess_scripts`](./preprocess_scripts) to process CVSS-C data.
+
+### 2. Training
+
+> [!Note]
+> You can directly use the [downloaded StreamSpeech model](#1-model-download) for evaluation and skip training.
+
+
+
+
+
+- Follow [`researches/ctc_unity/train_scripts/train.simul-s2st.sh`](./researches/ctc_unity/train_scripts/train.simul-s2st.sh) to train StreamSpeech for simultaneous speech-to-speech translation.
+- Follow [`researches/ctc_unity/train_scripts/train.offline-s2st.sh`](./researches/ctc_unity/train_scripts/train.offline-s2st.sh) to train StreamSpeech for offline speech-to-speech translation.
+- We also provide some other StreamSpeech variants and baseline implementations.
+
+| Model | --user-dir | --arch | Description |
+| ----------------- | -------------------------- | --------------------------------- | ------------------------------------------------------------ |
+| **Translatotron 2** | `researches/translatotron` | `s2spect2_conformer_modified` | [Translatotron 2](https://proceedings.mlr.press/v162/jia22b.html) |
+| **UnitY** | `researches/translatotron` | `unity_conformer_modified` | [UnitY](https://aclanthology.org/2023.acl-long.872/) |
+| **Uni-UnitY** | `researches/uni_unity` | `uni_unity_conformer` | Change all encoders in UnitY into unidirectional |
+| **Chunk-UnitY** | `researches/chunk_unity` | `chunk_unity_conformer` | Change the Conformer in UnitY into Chunk-based Conformer |
+| **StreamSpeech** | `researches/ctc_unity` | `streamspeech` | StreamSpeech |
+| **StreamSpeech (cascade)** | `researches/ctc_unity` | `streamspeech_cascade` | Cascaded StreamSpeech of S2TT and TTS. TTS module can be used independently for real-time TTS given incremental text. |
+| **HMT** | `researches/hmt` | `hmt_transformer_iwslt_de_en` | [HMT](https://openreview.net/forum?id=9y0HFvaAYD6): strong simultaneous text-to-text translation method |
+| **DiSeg** | `researches/diseg` | `convtransformer_espnet_base_seg` | [DiSeg](https://aclanthology.org/2023.findings-acl.485/): strong simultaneous speech-to-text translation method |
+
+> [!Tip]
+> The `train_scripts/` and `test_scripts/` in directory `--user-dir` give the training and testing scripts for each model.
+> Refer to official repo of [UnitY](https://github.com/facebookresearch/fairseq/blob/main/fairseq/models/speech_to_speech/s2s_conformer_unity.py), [Translatotron 2](https://github.com/facebookresearch/fairseq/blob/main/fairseq/models/speech_to_speech/s2s_conformer_translatotron2.py), [HMT](https://github.com/ictnlp/HMT) and [DiSeg](https://github.com/ictnlp/DiSeg) for more details.
+
+### 3. Evaluation
+
+#### (1) Offline Evaluation
+
+Follow [`pred.offline-s2st.sh`](./researches/ctc_unity/test_scripts/pred.offline-s2st.sh) to evaluate the offline performance of StreamSpeech on ASR, S2TT and S2ST.
+
+#### (2) Simultaneous Evaluation
+
+A trained StreamSpeech model can be used for streaming ASR, simultaneous speech-to-text translation and simultaneous speech-to-speech translation. We provide [agent/](./agent) for these three tasks:
+
+- `agent/speech_to_speech.streamspeech.agent.py`: simultaneous speech-to-speech translation
+- `agent/speech_to_text.s2tt.streamspeech.agent.py`: simultaneous speech-to-text translation
+- `agent/speech_to_text.asr.streamspeech.agent.py`: streaming ASR
+
+Follow [`simuleval.simul-s2st.sh`](./researches/ctc_unity/test_scripts/simuleval.simul-s2st.sh), [`simuleval.simul-s2tt.sh`](./researches/ctc_unity/test_scripts/simuleval.simul-s2tt.sh), [`simuleval.streaming-asr.sh`](./researches/ctc_unity/test_scripts/simuleval.streaming-asr.sh) to evaluate StreamSpeech.
+
+### 4. Our Results
+
+Our project page ([https://ictnlp.github.io/StreamSpeech-site/](https://ictnlp.github.io/StreamSpeech-site/)) provides some translated speech generated by StreamSpeech, listen to it 🎧.
+
+#### (1) Offline Speech-to-Speech Translation ( ASR-BLEU: quality )
+
+
+
+## 🖋Citation
+
+If you have any questions, please feel free to submit an issue or contact `zhangshaolei20z@ict.ac.cn`.
+
+If our work is useful for you, please cite as:
+
+```
+@inproceedings{streamspeech,
+ title={StreamSpeech: Simultaneous Speech-to-Speech Translation with Multi-task Learning},
+ author={Shaolei Zhang and Qingkai Fang and Shoutao Guo and Zhengrui Ma and Min Zhang and Yang Feng},
+ year={2024},
+ booktitle = {Proceedings of the 62th Annual Meeting of the Association for Computational Linguistics (Long Papers)},
+ publisher = {Association for Computational Linguistics}
+}
+```
diff --git a/SimulEval/.github/workflows/main.yml b/SimulEval/.github/workflows/main.yml
new file mode 100644
index 0000000000000000000000000000000000000000..0415649833d28f6e84941cd196c302141094d7c8
--- /dev/null
+++ b/SimulEval/.github/workflows/main.yml
@@ -0,0 +1,46 @@
+# This workflow will install Python dependencies, run tests and lint with a variety of Python versions
+# For more information see: https://help.github.com/actions/language-and-framework-guides/using-python-with-github-actions
+
+name: build
+
+on:
+ push:
+ branches:
+ - main
+ pull_request:
+
+jobs:
+ build:
+
+ runs-on: ubuntu-latest
+ strategy:
+ matrix:
+ python-version: [3.7, 3.8]
+
+ steps:
+ - uses: actions/checkout@v2
+ - name: Set up Python ${{ matrix.python-version}}
+ uses: actions/setup-python@v2
+ with:
+ python-version: ${{ matrix.python-version }}
+ - name: Install dependencies
+ run: |
+ sudo apt-get install libsndfile1
+ python -m pip install --upgrade pip
+ pip install flake8 pytest black
+ pip install -e .
+ if [ -f requirements.txt ]; then pip install -r requirements.txt; fi
+ - name: Lint with black
+ run: black --check --diff .
+ - name: Lint with flake8
+ run: |
+ # stop the build if there are Python syntax errors or undefined names
+ flake8 . --count --select=E9,F63,F7,F82 --show-source --statistics
+ # exit-zero treats all errors as warnings. The GitHub editor is 127 chars wide
+ flake8 . --count --exit-zero --max-complexity=10 --max-line-length=127 --statistics
+ - name: Test with pytest
+ run: |
+ pytest simuleval/test/test_agent.py
+ pytest simuleval/test/test_agent_pipeline.py
+ pytest simuleval/test/test_evaluator.py
+ pytest simuleval/test/test_remote_evaluation.py
diff --git a/SimulEval/.gitignore b/SimulEval/.gitignore
new file mode 100644
index 0000000000000000000000000000000000000000..1b244ec7787105ca4d11138cba295409ed1f6764
--- /dev/null
+++ b/SimulEval/.gitignore
@@ -0,0 +1,139 @@
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+# Usually these files are written by a python script from a template
+# before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+cover/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+.pybuilder/
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+# For a library or package, you might want to ignore these files since the code is
+# intended to run in multiple environments; otherwise, check them in:
+# .python-version
+
+# pipenv
+# According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+# However, in case of collaboration, if having platform-specific dependencies or dependencies
+# having no cross-platform support, pipenv may install dependencies that don't work, or not
+# install all needed dependencies.
+#Pipfile.lock
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+# pytype static type analyzer
+.pytype/
+
+# Cython debug symbols
+cython_debug/
+.vscode
diff --git a/SimulEval/CHANGELOG.md b/SimulEval/CHANGELOG.md
new file mode 100644
index 0000000000000000000000000000000000000000..133ca12f3d71bd4828bd2bdbce84cbe979587f45
--- /dev/null
+++ b/SimulEval/CHANGELOG.md
@@ -0,0 +1,10 @@
+1.0.0 (September 25, 2020)
+
+* Initial release.
+
+1.0.1 (February 8, 2021)
+
+* Change CLI command
+ * Change `simuleval-server` to `simuleval --server-only`
+ * Change `simuleval-client` to `simuleval --client-only`
+* Fix some typos
diff --git a/SimulEval/CODE_OF_CONDUCT.md b/SimulEval/CODE_OF_CONDUCT.md
new file mode 100644
index 0000000000000000000000000000000000000000..50bf79cf05470d70b3dcb5bcde0a0969cc451e3d
--- /dev/null
+++ b/SimulEval/CODE_OF_CONDUCT.md
@@ -0,0 +1,45 @@
+# Open Source Code of Conduct
+
+## Our Pledge
+
+In the interest of fostering an open and welcoming environment, we as contributors and maintainers pledge to make participation in our project and our community a harassment-free experience for everyone, regardless of age, body size, disability, ethnicity, sex characteristics, gender identity and expression, level of experience, education, socio-economic status, nationality, personal appearance, race, religion, or sexual identity and orientation.
+
+## Our Standards
+
+Examples of behavior that contributes to creating a positive environment include:
+
+Using welcoming and inclusive language
+Being respectful of differing viewpoints and experiences
+Gracefully accepting constructive criticism
+Focusing on what is best for the community
+Showing empathy towards other community members
+Examples of unacceptable behavior by participants include:
+
+The use of sexualized language or imagery and unwelcome sexual attention or advances
+Trolling, insulting/derogatory comments, and personal or political attacks
+Public or private harassment
+Publishing others’ private information, such as a physical or electronic address, without explicit permission
+Other conduct which could reasonably be considered inappropriate in a professional setting
+
+## Our Responsibilities
+
+Project maintainers are responsible for clarifying the standards of acceptable behavior and are expected to take appropriate and fair corrective action in response to any instances of unacceptable behavior.
+
+Project maintainers have the right and responsibility to remove, edit, or reject comments, commits, code, wiki edits, issues, and other contributions that are not aligned to this Code of Conduct, or to ban temporarily or permanently any contributor for other behaviors that they deem inappropriate, threatening, offensive, or harmful.
+
+## Scope
+
+This Code of Conduct applies within all project spaces, and it also applies when an individual is representing the project or its community in public spaces. Examples of representing a project or community include using an official project e-mail address, posting via an official social media account, or acting as an appointed representative at an online or offline event. Representation of a project may be further defined and clarified by project maintainers.
+
+## Enforcement
+
+Instances of abusive, harassing, or otherwise unacceptable behavior may be reported by contacting the project team at opensource-conduct@fb.com. All complaints will be reviewed and investigated and will result in a response that is deemed necessary and appropriate to the circumstances. The project team is obligated to maintain confidentiality with regard to the reporter of an incident. Further details of specific enforcement policies may be posted separately.
+
+Project maintainers who do not follow or enforce the Code of Conduct in good faith may face temporary or permanent repercussions as determined by other members of the project’s leadership.
+
+## Attribution
+
+This Code of Conduct is adapted from the [Contributor Covenant][homepage], version 1.4,
+available at https://www.contributor-covenant.org/version/1/4/code-of-conduct.html
+
+[homepage]: https://www.contributor-covenant.org
diff --git a/SimulEval/CONTRIBUTING.md b/SimulEval/CONTRIBUTING.md
new file mode 100644
index 0000000000000000000000000000000000000000..80de56bab331b14f416e85985c4adc148ab41046
--- /dev/null
+++ b/SimulEval/CONTRIBUTING.md
@@ -0,0 +1,20 @@
+# Contributing to Facebook AI SimulEval
+We want to make contributing to this project as easy and transparent as
+possible.
+
+## Pull Requests
+We actively welcome your pull requests.
+
+1. Fork the repo and create your branch from `master`.
+2. If you've added code that should be tested, add tests.
+3. If you've changed APIs, update the documentation.
+4. Ensure the test suite passes.
+5. Make sure your code lints.
+
+## Issues
+We use GitHub issues to track public bugs. Please ensure your description is
+clear and has sufficient instructions to be able to reproduce the issue.
+
+## License
+By contributing to Facebook AI SimulEval, you agree that your contributions will
+be licensed under the LICENSE file in the root directory of this source tree.
diff --git a/SimulEval/LICENSE b/SimulEval/LICENSE
new file mode 100644
index 0000000000000000000000000000000000000000..4e46ffb4ff36aa75a607b3284b2d773fe63cee01
--- /dev/null
+++ b/SimulEval/LICENSE
@@ -0,0 +1,427 @@
+Attribution-ShareAlike 4.0 International
+
+=======================================================================
+
+Creative Commons Corporation ("Creative Commons") is not a law firm and
+does not provide legal services or legal advice. Distribution of
+Creative Commons public licenses does not create a lawyer-client or
+other relationship. Creative Commons makes its licenses and related
+information available on an "as-is" basis. Creative Commons gives no
+warranties regarding its licenses, any material licensed under their
+terms and conditions, or any related information. Creative Commons
+disclaims all liability for damages resulting from their use to the
+fullest extent possible.
+
+Using Creative Commons Public Licenses
+
+Creative Commons public licenses provide a standard set of terms and
+conditions that creators and other rights holders may use to share
+original works of authorship and other material subject to copyright
+and certain other rights specified in the public license below. The
+following considerations are for informational purposes only, are not
+exhaustive, and do not form part of our licenses.
+
+ Considerations for licensors: Our public licenses are
+ intended for use by those authorized to give the public
+ permission to use material in ways otherwise restricted by
+ copyright and certain other rights. Our licenses are
+ irrevocable. Licensors should read and understand the terms
+ and conditions of the license they choose before applying it.
+ Licensors should also secure all rights necessary before
+ applying our licenses so that the public can reuse the
+ material as expected. Licensors should clearly mark any
+ material not subject to the license. This includes other CC-
+ licensed material, or material used under an exception or
+ limitation to copyright. More considerations for licensors:
+ wiki.creativecommons.org/Considerations_for_licensors
+
+ Considerations for the public: By using one of our public
+ licenses, a licensor grants the public permission to use the
+ licensed material under specified terms and conditions. If
+ the licensor's permission is not necessary for any reason--for
+ example, because of any applicable exception or limitation to
+ copyright--then that use is not regulated by the license. Our
+ licenses grant only permissions under copyright and certain
+ other rights that a licensor has authority to grant. Use of
+ the licensed material may still be restricted for other
+ reasons, including because others have copyright or other
+ rights in the material. A licensor may make special requests,
+ such as asking that all changes be marked or described.
+ Although not required by our licenses, you are encouraged to
+ respect those requests where reasonable. More_considerations
+ for the public:
+ wiki.creativecommons.org/Considerations_for_licensees
+
+=======================================================================
+
+Creative Commons Attribution-ShareAlike 4.0 International Public
+License
+
+By exercising the Licensed Rights (defined below), You accept and agree
+to be bound by the terms and conditions of this Creative Commons
+Attribution-ShareAlike 4.0 International Public License ("Public
+License"). To the extent this Public License may be interpreted as a
+contract, You are granted the Licensed Rights in consideration of Your
+acceptance of these terms and conditions, and the Licensor grants You
+such rights in consideration of benefits the Licensor receives from
+making the Licensed Material available under these terms and
+conditions.
+
+
+Section 1 -- Definitions.
+
+ a. Adapted Material means material subject to Copyright and Similar
+ Rights that is derived from or based upon the Licensed Material
+ and in which the Licensed Material is translated, altered,
+ arranged, transformed, or otherwise modified in a manner requiring
+ permission under the Copyright and Similar Rights held by the
+ Licensor. For purposes of this Public License, where the Licensed
+ Material is a musical work, performance, or sound recording,
+ Adapted Material is always produced where the Licensed Material is
+ synched in timed relation with a moving image.
+
+ b. Adapter's License means the license You apply to Your Copyright
+ and Similar Rights in Your contributions to Adapted Material in
+ accordance with the terms and conditions of this Public License.
+
+ c. BY-SA Compatible License means a license listed at
+ creativecommons.org/compatiblelicenses, approved by Creative
+ Commons as essentially the equivalent of this Public License.
+
+ d. Copyright and Similar Rights means copyright and/or similar rights
+ closely related to copyright including, without limitation,
+ performance, broadcast, sound recording, and Sui Generis Database
+ Rights, without regard to how the rights are labeled or
+ categorized. For purposes of this Public License, the rights
+ specified in Section 2(b)(1)-(2) are not Copyright and Similar
+ Rights.
+
+ e. Effective Technological Measures means those measures that, in the
+ absence of proper authority, may not be circumvented under laws
+ fulfilling obligations under Article 11 of the WIPO Copyright
+ Treaty adopted on December 20, 1996, and/or similar international
+ agreements.
+
+ f. Exceptions and Limitations means fair use, fair dealing, and/or
+ any other exception or limitation to Copyright and Similar Rights
+ that applies to Your use of the Licensed Material.
+
+ g. License Elements means the license attributes listed in the name
+ of a Creative Commons Public License. The License Elements of this
+ Public License are Attribution and ShareAlike.
+
+ h. Licensed Material means the artistic or literary work, database,
+ or other material to which the Licensor applied this Public
+ License.
+
+ i. Licensed Rights means the rights granted to You subject to the
+ terms and conditions of this Public License, which are limited to
+ all Copyright and Similar Rights that apply to Your use of the
+ Licensed Material and that the Licensor has authority to license.
+
+ j. Licensor means the individual(s) or entity(ies) granting rights
+ under this Public License.
+
+ k. Share means to provide material to the public by any means or
+ process that requires permission under the Licensed Rights, such
+ as reproduction, public display, public performance, distribution,
+ dissemination, communication, or importation, and to make material
+ available to the public including in ways that members of the
+ public may access the material from a place and at a time
+ individually chosen by them.
+
+ l. Sui Generis Database Rights means rights other than copyright
+ resulting from Directive 96/9/EC of the European Parliament and of
+ the Council of 11 March 1996 on the legal protection of databases,
+ as amended and/or succeeded, as well as other essentially
+ equivalent rights anywhere in the world.
+
+ m. You means the individual or entity exercising the Licensed Rights
+ under this Public License. Your has a corresponding meaning.
+
+
+Section 2 -- Scope.
+
+ a. License grant.
+
+ 1. Subject to the terms and conditions of this Public License,
+ the Licensor hereby grants You a worldwide, royalty-free,
+ non-sublicensable, non-exclusive, irrevocable license to
+ exercise the Licensed Rights in the Licensed Material to:
+
+ a. reproduce and Share the Licensed Material, in whole or
+ in part; and
+
+ b. produce, reproduce, and Share Adapted Material.
+
+ 2. Exceptions and Limitations. For the avoidance of doubt, where
+ Exceptions and Limitations apply to Your use, this Public
+ License does not apply, and You do not need to comply with
+ its terms and conditions.
+
+ 3. Term. The term of this Public License is specified in Section
+ 6(a).
+
+ 4. Media and formats; technical modifications allowed. The
+ Licensor authorizes You to exercise the Licensed Rights in
+ all media and formats whether now known or hereafter created,
+ and to make technical modifications necessary to do so. The
+ Licensor waives and/or agrees not to assert any right or
+ authority to forbid You from making technical modifications
+ necessary to exercise the Licensed Rights, including
+ technical modifications necessary to circumvent Effective
+ Technological Measures. For purposes of this Public License,
+ simply making modifications authorized by this Section 2(a)
+ (4) never produces Adapted Material.
+
+ 5. Downstream recipients.
+
+ a. Offer from the Licensor -- Licensed Material. Every
+ recipient of the Licensed Material automatically
+ receives an offer from the Licensor to exercise the
+ Licensed Rights under the terms and conditions of this
+ Public License.
+
+ b. Additional offer from the Licensor -- Adapted Material.
+ Every recipient of Adapted Material from You
+ automatically receives an offer from the Licensor to
+ exercise the Licensed Rights in the Adapted Material
+ under the conditions of the Adapter's License You apply.
+
+ c. No downstream restrictions. You may not offer or impose
+ any additional or different terms or conditions on, or
+ apply any Effective Technological Measures to, the
+ Licensed Material if doing so restricts exercise of the
+ Licensed Rights by any recipient of the Licensed
+ Material.
+
+ 6. No endorsement. Nothing in this Public License constitutes or
+ may be construed as permission to assert or imply that You
+ are, or that Your use of the Licensed Material is, connected
+ with, or sponsored, endorsed, or granted official status by,
+ the Licensor or others designated to receive attribution as
+ provided in Section 3(a)(1)(A)(i).
+
+ b. Other rights.
+
+ 1. Moral rights, such as the right of integrity, are not
+ licensed under this Public License, nor are publicity,
+ privacy, and/or other similar personality rights; however, to
+ the extent possible, the Licensor waives and/or agrees not to
+ assert any such rights held by the Licensor to the limited
+ extent necessary to allow You to exercise the Licensed
+ Rights, but not otherwise.
+
+ 2. Patent and trademark rights are not licensed under this
+ Public License.
+
+ 3. To the extent possible, the Licensor waives any right to
+ collect royalties from You for the exercise of the Licensed
+ Rights, whether directly or through a collecting society
+ under any voluntary or waivable statutory or compulsory
+ licensing scheme. In all other cases the Licensor expressly
+ reserves any right to collect such royalties.
+
+
+Section 3 -- License Conditions.
+
+Your exercise of the Licensed Rights is expressly made subject to the
+following conditions.
+
+ a. Attribution.
+
+ 1. If You Share the Licensed Material (including in modified
+ form), You must:
+
+ a. retain the following if it is supplied by the Licensor
+ with the Licensed Material:
+
+ i. identification of the creator(s) of the Licensed
+ Material and any others designated to receive
+ attribution, in any reasonable manner requested by
+ the Licensor (including by pseudonym if
+ designated);
+
+ ii. a copyright notice;
+
+ iii. a notice that refers to this Public License;
+
+ iv. a notice that refers to the disclaimer of
+ warranties;
+
+ v. a URI or hyperlink to the Licensed Material to the
+ extent reasonably practicable;
+
+ b. indicate if You modified the Licensed Material and
+ retain an indication of any previous modifications; and
+
+ c. indicate the Licensed Material is licensed under this
+ Public License, and include the text of, or the URI or
+ hyperlink to, this Public License.
+
+ 2. You may satisfy the conditions in Section 3(a)(1) in any
+ reasonable manner based on the medium, means, and context in
+ which You Share the Licensed Material. For example, it may be
+ reasonable to satisfy the conditions by providing a URI or
+ hyperlink to a resource that includes the required
+ information.
+
+ 3. If requested by the Licensor, You must remove any of the
+ information required by Section 3(a)(1)(A) to the extent
+ reasonably practicable.
+
+ b. ShareAlike.
+
+ In addition to the conditions in Section 3(a), if You Share
+ Adapted Material You produce, the following conditions also apply.
+
+ 1. The Adapter's License You apply must be a Creative Commons
+ license with the same License Elements, this version or
+ later, or a BY-SA Compatible License.
+
+ 2. You must include the text of, or the URI or hyperlink to, the
+ Adapter's License You apply. You may satisfy this condition
+ in any reasonable manner based on the medium, means, and
+ context in which You Share Adapted Material.
+
+ 3. You may not offer or impose any additional or different terms
+ or conditions on, or apply any Effective Technological
+ Measures to, Adapted Material that restrict exercise of the
+ rights granted under the Adapter's License You apply.
+
+
+Section 4 -- Sui Generis Database Rights.
+
+Where the Licensed Rights include Sui Generis Database Rights that
+apply to Your use of the Licensed Material:
+
+ a. for the avoidance of doubt, Section 2(a)(1) grants You the right
+ to extract, reuse, reproduce, and Share all or a substantial
+ portion of the contents of the database;
+
+ b. if You include all or a substantial portion of the database
+ contents in a database in which You have Sui Generis Database
+ Rights, then the database in which You have Sui Generis Database
+ Rights (but not its individual contents) is Adapted Material,
+
+ including for purposes of Section 3(b); and
+ c. You must comply with the conditions in Section 3(a) if You Share
+ all or a substantial portion of the contents of the database.
+
+For the avoidance of doubt, this Section 4 supplements and does not
+replace Your obligations under this Public License where the Licensed
+Rights include other Copyright and Similar Rights.
+
+
+Section 5 -- Disclaimer of Warranties and Limitation of Liability.
+
+ a. UNLESS OTHERWISE SEPARATELY UNDERTAKEN BY THE LICENSOR, TO THE
+ EXTENT POSSIBLE, THE LICENSOR OFFERS THE LICENSED MATERIAL AS-IS
+ AND AS-AVAILABLE, AND MAKES NO REPRESENTATIONS OR WARRANTIES OF
+ ANY KIND CONCERNING THE LICENSED MATERIAL, WHETHER EXPRESS,
+ IMPLIED, STATUTORY, OR OTHER. THIS INCLUDES, WITHOUT LIMITATION,
+ WARRANTIES OF TITLE, MERCHANTABILITY, FITNESS FOR A PARTICULAR
+ PURPOSE, NON-INFRINGEMENT, ABSENCE OF LATENT OR OTHER DEFECTS,
+ ACCURACY, OR THE PRESENCE OR ABSENCE OF ERRORS, WHETHER OR NOT
+ KNOWN OR DISCOVERABLE. WHERE DISCLAIMERS OF WARRANTIES ARE NOT
+ ALLOWED IN FULL OR IN PART, THIS DISCLAIMER MAY NOT APPLY TO YOU.
+
+ b. TO THE EXTENT POSSIBLE, IN NO EVENT WILL THE LICENSOR BE LIABLE
+ TO YOU ON ANY LEGAL THEORY (INCLUDING, WITHOUT LIMITATION,
+ NEGLIGENCE) OR OTHERWISE FOR ANY DIRECT, SPECIAL, INDIRECT,
+ INCIDENTAL, CONSEQUENTIAL, PUNITIVE, EXEMPLARY, OR OTHER LOSSES,
+ COSTS, EXPENSES, OR DAMAGES ARISING OUT OF THIS PUBLIC LICENSE OR
+ USE OF THE LICENSED MATERIAL, EVEN IF THE LICENSOR HAS BEEN
+ ADVISED OF THE POSSIBILITY OF SUCH LOSSES, COSTS, EXPENSES, OR
+ DAMAGES. WHERE A LIMITATION OF LIABILITY IS NOT ALLOWED IN FULL OR
+ IN PART, THIS LIMITATION MAY NOT APPLY TO YOU.
+
+ c. The disclaimer of warranties and limitation of liability provided
+ above shall be interpreted in a manner that, to the extent
+ possible, most closely approximates an absolute disclaimer and
+ waiver of all liability.
+
+
+Section 6 -- Term and Termination.
+
+ a. This Public License applies for the term of the Copyright and
+ Similar Rights licensed here. However, if You fail to comply with
+ this Public License, then Your rights under this Public License
+ terminate automatically.
+
+ b. Where Your right to use the Licensed Material has terminated under
+ Section 6(a), it reinstates:
+
+ 1. automatically as of the date the violation is cured, provided
+ it is cured within 30 days of Your discovery of the
+ violation; or
+
+ 2. upon express reinstatement by the Licensor.
+
+ For the avoidance of doubt, this Section 6(b) does not affect any
+ right the Licensor may have to seek remedies for Your violations
+ of this Public License.
+
+ c. For the avoidance of doubt, the Licensor may also offer the
+ Licensed Material under separate terms or conditions or stop
+ distributing the Licensed Material at any time; however, doing so
+ will not terminate this Public License.
+
+ d. Sections 1, 5, 6, 7, and 8 survive termination of this Public
+ License.
+
+
+Section 7 -- Other Terms and Conditions.
+
+ a. The Licensor shall not be bound by any additional or different
+ terms or conditions communicated by You unless expressly agreed.
+
+ b. Any arrangements, understandings, or agreements regarding the
+ Licensed Material not stated herein are separate from and
+ independent of the terms and conditions of this Public License.
+
+
+Section 8 -- Interpretation.
+
+ a. For the avoidance of doubt, this Public License does not, and
+ shall not be interpreted to, reduce, limit, restrict, or impose
+ conditions on any use of the Licensed Material that could lawfully
+ be made without permission under this Public License.
+
+ b. To the extent possible, if any provision of this Public License is
+ deemed unenforceable, it shall be automatically reformed to the
+ minimum extent necessary to make it enforceable. If the provision
+ cannot be reformed, it shall be severed from this Public License
+ without affecting the enforceability of the remaining terms and
+ conditions.
+
+ c. No term or condition of this Public License will be waived and no
+ failure to comply consented to unless expressly agreed to by the
+ Licensor.
+
+ d. Nothing in this Public License constitutes or may be interpreted
+ as a limitation upon, or waiver of, any privileges and immunities
+ that apply to the Licensor or You, including from the legal
+ processes of any jurisdiction or authority.
+
+
+=======================================================================
+
+Creative Commons is not a party to its public
+licenses. Notwithstanding, Creative Commons may elect to apply one of
+its public licenses to material it publishes and in those instances
+will be considered the “Licensor.” The text of the Creative Commons
+public licenses is dedicated to the public domain under the CC0 Public
+Domain Dedication. Except for the limited purpose of indicating that
+material is shared under a Creative Commons public license or as
+otherwise permitted by the Creative Commons policies published at
+creativecommons.org/policies, Creative Commons does not authorize the
+use of the trademark "Creative Commons" or any other trademark or logo
+of Creative Commons without its prior written consent including,
+without limitation, in connection with any unauthorized modifications
+to any of its public licenses or any other arrangements,
+understandings, or agreements concerning use of licensed material. For
+the avoidance of doubt, this paragraph does not form part of the
+public licenses.
+
+Creative Commons may be contacted at creativecommons.org.
diff --git a/SimulEval/README.md b/SimulEval/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..f1c4b146c953b36d9938e0162e633ecb807de37a
--- /dev/null
+++ b/SimulEval/README.md
@@ -0,0 +1,35 @@
+# SimulEval
+[](https://github.com/facebookresearch/SimulEval/actions)
+
+SimulEval is a general evaluation framework for simultaneous translation on text and speech. Full documentation can be found [here](https://simuleval.readthedocs.io/en/v1.1.0/).
+
+## Installation
+```
+git clone https://github.com/facebookresearch/SimulEval.git
+cd SimulEval
+pip install -e .
+```
+
+## Quick Start
+Following is the evaluation of a [dummy agent](examples/quick_start) which operates wait-k (k = 3) policy and generates random words until the length of the generated words is the same as the number of all the source words.
+```shell
+cd examples/quick_start
+simuleval --source source.txt --target target.txt --agent first_agent.py
+```
+
+# License
+
+SimulEval is licensed under Creative Commons BY-SA 4.0.
+
+# Citation
+
+Please cite as:
+
+```bibtex
+@inproceedings{simuleval2020,
+ title = {Simuleval: An evaluation toolkit for simultaneous translation},
+ author = {Xutai Ma, Mohammad Javad Dousti, Changhan Wang, Jiatao Gu, Juan Pino},
+ booktitle = {Proceedings of the EMNLP},
+ year = {2020},
+}
+```
diff --git a/SimulEval/docs/Makefile b/SimulEval/docs/Makefile
new file mode 100644
index 0000000000000000000000000000000000000000..ed88099027f775942fa65dce2314f1ae9675cb36
--- /dev/null
+++ b/SimulEval/docs/Makefile
@@ -0,0 +1,20 @@
+# Minimal makefile for Sphinx documentation
+#
+
+# You can set these variables from the command line, and also
+# from the environment for the first two.
+SPHINXOPTS ?=
+SPHINXBUILD ?= sphinx-build
+SOURCEDIR = .
+BUILDDIR = build
+
+# Put it first so that "make" without argument is like "make help".
+help:
+ @$(SPHINXBUILD) -M help "$(SOURCEDIR)" "$(BUILDDIR)" $(SPHINXOPTS) $(O)
+
+.PHONY: help Makefile
+
+# Catch-all target: route all unknown targets to Sphinx using the new
+# "make mode" option. $(O) is meant as a shortcut for $(SPHINXOPTS).
+%: Makefile
+ @$(SPHINXBUILD) -M $@ "$(SOURCEDIR)" "$(BUILDDIR)" $(SPHINXOPTS) $(O)
diff --git a/SimulEval/docs/conf.py b/SimulEval/docs/conf.py
new file mode 100644
index 0000000000000000000000000000000000000000..8266d55ea29a635a779e1787230c71eed058c79b
--- /dev/null
+++ b/SimulEval/docs/conf.py
@@ -0,0 +1,40 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+# Configuration file for the Sphinx documentation builder.
+#
+# For the full list of built-in configuration values, see the documentation:
+# https://www.sphinx-doc.org/en/master/usage/configuration.html
+
+# -- Project information -----------------------------------------------------
+# https://www.sphinx-doc.org/en/master/usage/configuration.html#project-information
+
+project = "SimulEval"
+copyright = "Facebook AI Research (FAIR)"
+author = "Facebook AI Research (FAIR)"
+release = "1.1.0"
+
+# -- General configuration ---------------------------------------------------
+# https://www.sphinx-doc.org/en/master/usage/configuration.html#general-configuration
+
+extensions = [
+ "sphinx_rtd_theme",
+ "sphinx.ext.autodoc",
+ "sphinx.ext.intersphinx",
+ "sphinx.ext.viewcode",
+ "sphinx.ext.napoleon",
+ "sphinxarg.ext",
+]
+
+# templates_path = ['_templates']
+exclude_patterns = []
+
+
+# -- Options for HTML output -------------------------------------------------
+# https://www.sphinx-doc.org/en/master/usage/configuration.html#options-for-html-output
+
+html_theme = "sphinx_rtd_theme"
+# html_static_path = ['_static']
diff --git a/SimulEval/docs/index.rst b/SimulEval/docs/index.rst
new file mode 100644
index 0000000000000000000000000000000000000000..5c095eab7f1dba976c28c5d989633c562dcdb747
--- /dev/null
+++ b/SimulEval/docs/index.rst
@@ -0,0 +1,30 @@
+SimulEval documentation
+=======================
+
+SimulEval is a general evaluation framework for simultaneous translation.
+
+.. toctree::
+ :maxdepth: 2
+ :glob:
+ :caption: Get started
+
+ installation
+ quick_start
+
+.. toctree::
+ :maxdepth: 2
+ :caption: User's guide
+
+ user_guide/introduction
+ user_guide/agent
+ user_guide/evaluator
+ user_guide/dataloader
+
+.. toctree::
+ :maxdepth: 2
+ :caption: Tutorials
+
+ tutorials/remote_evaluation
+ tutorials/speech_to_text
+ tutorials/speech_to_speech
+
diff --git a/SimulEval/docs/installation.rst b/SimulEval/docs/installation.rst
new file mode 100644
index 0000000000000000000000000000000000000000..8e3b4d1e0bf456277ebfc2bbed533ebdaa1bc957
--- /dev/null
+++ b/SimulEval/docs/installation.rst
@@ -0,0 +1,21 @@
+Installation
+============
+
+From pip
+--------
+
+.. code-block:: bash
+
+ pip install simuleval
+
+
+From source
+-----------
+
+.. code-block:: bash
+
+ git clone https://github.com/facebookresearch/SimulEval.git
+ cd SimulEval
+ pip install -e .
+
+
diff --git a/SimulEval/docs/make.bat b/SimulEval/docs/make.bat
new file mode 100644
index 0000000000000000000000000000000000000000..30395bd8a61ac4975bdb0d31172144fd1718979b
--- /dev/null
+++ b/SimulEval/docs/make.bat
@@ -0,0 +1,41 @@
+:: Copyright (c) Facebook, Inc. and its affiliates.
+:: All rights reserved.
+::
+:: This source code is licensed under the license found in the
+:: LICENSE file in the root directory of this source tree.
+
+@ECHO OFF
+
+pushd %~dp0
+
+REM Command file for Sphinx documentation
+
+if "%SPHINXBUILD%" == "" (
+ set SPHINXBUILD=sphinx-build
+)
+set SOURCEDIR=source
+set BUILDDIR=build
+
+%SPHINXBUILD% >NUL 2>NUL
+if errorlevel 9009 (
+ echo.
+ echo.The 'sphinx-build' command was not found. Make sure you have Sphinx
+ echo.installed, then set the SPHINXBUILD environment variable to point
+ echo.to the full path of the 'sphinx-build' executable. Alternatively you
+ echo.may add the Sphinx directory to PATH.
+ echo.
+ echo.If you don't have Sphinx installed, grab it from
+ echo.https://www.sphinx-doc.org/
+ exit /b 1
+)
+
+if "%1" == "" goto help
+
+%SPHINXBUILD% -M %1 %SOURCEDIR% %BUILDDIR% %SPHINXOPTS% %O%
+goto end
+
+:help
+%SPHINXBUILD% -M help %SOURCEDIR% %BUILDDIR% %SPHINXOPTS% %O%
+
+:end
+popd
diff --git a/SimulEval/docs/quick_start.rst b/SimulEval/docs/quick_start.rst
new file mode 100644
index 0000000000000000000000000000000000000000..678da9458b461528c6f3989ff85100c9814ed2c2
--- /dev/null
+++ b/SimulEval/docs/quick_start.rst
@@ -0,0 +1,49 @@
+.. _first-agent:
+
+Quick Start
+===========
+
+This section will introduce a minimal example on how to use SimulEval for simultaneous translation evaluation.
+The code in the example can be found in :code:`examples/quick_start`.
+
+The agent in SimulEval is core for simultaneous evaluation.
+It's a carrier of user's simultaneous system.
+The user has to implement the agent based on their system for evaluation.
+The example simultaneous system is a dummy wait-k agent, which
+
+- Runs `wait-k `_ policy.
+- Generates random characters the policy decide to write.
+- Stops the generation k predictions after source input. For simplicity, we just set :code:`k=3` in this example.
+
+The implementation of this agent is shown as follow.
+
+.. literalinclude:: ../examples/quick_start/first_agent.py
+ :language: python
+ :lines: 6-
+
+There two essential components for an agent:
+
+- :code:`states`: The attribute keeps track of the source and target information.
+- :code:`policy`: The method makes decisions when the there is a new source segment.
+
+Once the agent is implemented and saved at :code:`first_agent.py`,
+run the following command for latency evaluation on:
+
+.. code-block:: bash
+
+ simuleval --source source.txt --reference target.txt --agent first_agent.py
+
+where :code:`--source` is the input file while :code:`--target` is the reference file.
+
+By default, the SimulEval will give the following output --- one quality and three latency metrics.
+
+.. code-block:: bash
+
+ 2022-12-05 13:43:58 | INFO | simuleval.cli | Evaluate system: DummyWaitkTextAgent
+ 2022-12-05 13:43:58 | INFO | simuleval.dataloader | Evaluating from text to text.
+ 2022-12-05 13:43:58 | INFO | simuleval.sentence_level_evaluator | Results:
+ BLEU AL AP DAL
+ 1.541 3.0 0.688 3.0
+
+The average lagging is expected since we are running an wait-3 system where the source and target always have the same length.
+Notice that we have a very low yet random BLEU score. It's because we are randomly generate the output.
diff --git a/SimulEval/docs/requirements.txt b/SimulEval/docs/requirements.txt
new file mode 100644
index 0000000000000000000000000000000000000000..1ec40fe260a4497472bbad0be5e47682e7224441
--- /dev/null
+++ b/SimulEval/docs/requirements.txt
@@ -0,0 +1,2 @@
+sphinx-argparse==0.4.0
+-e .
diff --git a/SimulEval/docs/tutorials/remote_evaluation.rst b/SimulEval/docs/tutorials/remote_evaluation.rst
new file mode 100644
index 0000000000000000000000000000000000000000..77f7d1972c1fc8ce50fcb76447e5fc0b05792d9c
--- /dev/null
+++ b/SimulEval/docs/tutorials/remote_evaluation.rst
@@ -0,0 +1,44 @@
+Remote Evaluation
+=================
+
+Stand Alone Agent
+-----------------
+The agent can run in stand alone mode,
+by using :code:`--standalone` option.
+The SimulEval will kickoff a server that host the agent.
+For instance, with the agent in :ref:`first-agent`,
+
+.. code-block:: bash
+
+ > simuleval --standalone --remote-port 8888 --agent first_agent.py.py
+ 2022-12-06 19:12:26 | INFO | simuleval.cli | Evaluate system: DummyWaitkTextAgent
+ 2022-12-06 19:12:26 | INFO | simuleval.agent_server | Simultaneous Translation Server Started (process id 53902). Listening to port 8888
+
+For detailed RESTful APIs, please see (TODO)
+
+Docker
+-----------------
+You can also use a docker image to run the simuleval.
+An minimal example of :code:`Dockerfile` is
+
+.. literalinclude:: ../../examples/quick_start/Dockerfile
+ :language: docker
+
+Build and run the docker image:
+
+.. code-block:: bash
+
+ cd examples/quick_start && docker build -t simuleval_agent .
+ docker run -p 8888:8888 simuleval_agent:latest
+
+Remote Evaluation
+------------------
+If there is an agent server or docker image available,
+(let's say the one we just kickoff at localhost:8888)
+We can start a remote evaluator as follow. For simplicity we assume they are on the same machine
+
+.. code-block:: bash
+
+ simuleval --remote-eval --remote-port 8888 \
+ --source source.txt --target target.txt \
+ --source-type text --target-type text
diff --git a/SimulEval/docs/tutorials/speech_to_speech.rst b/SimulEval/docs/tutorials/speech_to_speech.rst
new file mode 100644
index 0000000000000000000000000000000000000000..2419928cc9c0b845f4d555cdc3df43e833df9eeb
--- /dev/null
+++ b/SimulEval/docs/tutorials/speech_to_speech.rst
@@ -0,0 +1,2 @@
+Speech-to-Speech
+================
\ No newline at end of file
diff --git a/SimulEval/docs/tutorials/speech_to_text.rst b/SimulEval/docs/tutorials/speech_to_text.rst
new file mode 100644
index 0000000000000000000000000000000000000000..6430f81a798d20231d3abd3394d9b79f870b4859
--- /dev/null
+++ b/SimulEval/docs/tutorials/speech_to_text.rst
@@ -0,0 +1,2 @@
+Speech-to-Text
+==============
\ No newline at end of file
diff --git a/SimulEval/docs/user_guide/agent.rst b/SimulEval/docs/user_guide/agent.rst
new file mode 100644
index 0000000000000000000000000000000000000000..91999dbb7dbf9cea9bd56e87f115d4bfd9063657
--- /dev/null
+++ b/SimulEval/docs/user_guide/agent.rst
@@ -0,0 +1,198 @@
+Agent
+=====
+
+To evaluate the simultaneous translation system,
+the users need to implement agent class which operate the system logics.
+This section will introduce how to implement an agent.
+
+Source-Target Types
+-------------------
+First of all,
+we must declare the source and target types of the agent class.
+It can be done by inheriting from
+
+- One of the following four built-in agent types
+
+ - :class:`simuleval.agents.TextToTextAgent`
+ - :class:`simuleval.agents.SpeechToTextAgent`
+ - :class:`simuleval.agents.TextToSpeechAgent`
+ - :class:`simuleval.agents.SpeechToSpeechAgent`
+
+- Or :class:`simuleval.agents.GenericAgent`, with explicit declaration of :code:`source_type` and :code:`target_type`.
+
+The follow two examples are equivalent.
+
+.. code-block:: python
+
+ from simuleval import simuleval
+ from simuleval.agents import GenericAgent
+
+ class MySpeechToTextAgent(GenericAgent):
+ source_type = "Speech"
+ target_type = "Text"
+ ....
+
+.. code-block:: python
+
+ from simuleval.agents import SpeechToSpeechAgent
+
+ class MySpeechToTextAgent(SpeechToSpeechAgent):
+ ....
+
+.. _agent_policy:
+
+Policy
+------
+
+The agent must have a :code:`policy` method which must return one of two actions, :code:`ReadAction` and :code:`WriteAction`.
+For example, an agent with a :code:`policy` method should look like this
+
+.. code-block:: python
+
+ class MySpeechToTextAgent(SpeechToSpeechAgent):
+ def policy(self):
+ if do_we_need_more_input(self.states):
+ return ReadAction()
+ else:
+ prediction = generate_a_token(self.states)
+ finished = is_sentence_finished(self.states)
+ return WriteAction(prediction, finished=finished)
+
+
+..
+ .. autoclass:: simuleval.agents.actions.WriteAction
+
+..
+ .. autoclass:: simuleval.agents.actions.ReadAction
+
+States
+------------
+Each agent has the attribute the :code:`states` to keep track of the progress of decoding.
+The :code:`states` attribute will be reset at the beginning of each sentence.
+SimulEval provide an built-in states :class:`simuleval.agents.states.AgentStates`,
+which has some basic attributes such source and target sequences.
+The users can also define customized states with :code:`Agent.build_states` method:
+
+.. code-block:: python
+
+ from simuleval.agents.states import AgentStates
+ from dataclasses import dataclass
+
+ @dataclass
+ class MyComplicatedStates(AgentStates)
+ some_very_useful_variable: int
+
+ def reset(self):
+ super().reset()
+ # also remember to reset the value
+ some_very_useful_variable = 0
+
+ class MySpeechToTextAgent(SpeechToSpeechAgent):
+ def build_states(self):
+ return MyComplicatedStates(0)
+
+ def policy(self):
+ some_very_useful_variable = self.states.some_very_useful_variable
+ ...
+ self.states.some_very_useful_variable = new_value
+ ...
+
+..
+ .. autoclass:: simuleval.agents.states.AgentStates
+ :members:
+
+
+Pipeline
+--------
+The simultaneous system can consist several different components.
+For instance, a simultaneous speech-to-text translation can have a streaming automatic speech recognition system and simultaneous text-to-text translation system.
+SimulEval introduces the agent pipeline to support this function.
+The following is a minimal example.
+We concatenate two wait-k systems with different rates (:code:`k=2` and :code:`k=3`)
+Note that if there are more than one agent class define,
+the :code:`@entrypoint` decorator has to be used to determine the entry point
+
+.. literalinclude:: ../../examples/quick_start/agent_pipeline.py
+ :language: python
+ :lines: 7-
+
+Customized Arguments
+-----------------------
+
+It is often the case that we need to pass some customized arguments for the system to configure different settings.
+The agent class has a built-in static method :code:`add_args` for this purpose.
+The following is an updated version of the dummy agent from :ref:`first-agent`.
+
+.. literalinclude:: ../../examples/quick_start/agent_with_configs.py
+ :language: python
+ :lines: 6-
+
+Then just simply pass the arguments through command line as follow.
+
+.. code-block:: bash
+
+ simuleval \
+ --source source.txt --source target.txt \ # data arguments
+ --agent dummy_waitk_text_agent_v2.py \
+ --waitk 3 --vocab data/dict.txt # agent arguments
+
+Load Agents from Python Class
+-----------------------------
+
+If you have the agent class in the python environment, for instance
+
+.. literalinclude:: ../../examples/quick_start/agent_with_configs.py
+ :language: python
+ :lines: 6-
+
+You can also start the evaluation with following command
+
+.. code-block:: bash
+
+ simuleval \
+ --source source.txt --source target.txt \ # data arguments
+ --agent-class DummyWaitkTextAgent \
+ --waitk 3 --vocab data/dict.txt # agent arguments
+
+
+Load Agents from Directory
+--------------------------
+
+Agent can also be loaded from a directory, which will be referred to as system directory.
+The system directory should have everything required to start the agent. Again use the following agent as example
+
+.. literalinclude:: ../../examples/quick_start/agent_with_configs.py
+ :language: python
+ :lines: 6-
+
+and the system directory has
+
+.. code-block:: bash
+
+ > ls ${system_dir}
+ main.yaml dict.txt
+
+Where the `main.yaml` has all the command line options. The path will be the relative path to the `${system_dir}`.
+
+.. code-block:: yaml
+
+ waitk: 3
+ vocab: dict.txt
+
+The agent can then be started as following
+
+.. code-block:: bash
+
+ simuleval \
+ --source source.txt --source target.txt \ # data arguments
+ --system-dir ${system_dir}
+
+By default, the `main.yaml` will be read. You can also have multiple YAML files in the system directory and pass them through command line arguments
+
+.. code-block:: bash
+ > ls ${system_dir}
+ main.yaml dict.txt v1.yaml
+
+ > simuleval \
+ --source source.txt --source target.txt \ # data arguments
+ --system-dir ${system_dir} --system-config v1.yaml
\ No newline at end of file
diff --git a/SimulEval/docs/user_guide/dataloader.rst b/SimulEval/docs/user_guide/dataloader.rst
new file mode 100644
index 0000000000000000000000000000000000000000..54f462b5aee6bf9d07d894b7447190314baba78b
--- /dev/null
+++ b/SimulEval/docs/user_guide/dataloader.rst
@@ -0,0 +1,5 @@
+Dataloader
+===========
+There are two ways to load data.
+
+.. autoclass:: simuleval.data.dataloader.GenericDataloader
\ No newline at end of file
diff --git a/SimulEval/docs/user_guide/evaluator.rst b/SimulEval/docs/user_guide/evaluator.rst
new file mode 100644
index 0000000000000000000000000000000000000000..c66c99d2baecc9669e3db52f29981dc687936911
--- /dev/null
+++ b/SimulEval/docs/user_guide/evaluator.rst
@@ -0,0 +1,44 @@
+Evaluator
+=========
+
+The evaluation in SimulEval implemented as the Evaluator shown below.
+It runs on sentence level, and will score the translation on quality and latency.
+The user can use :code:`--quality-metrics` and :code:`--latency-metrics` to choose the metrics.
+The final results along with the logs will be saved at :code:`--output` if given.
+
+.. autoclass:: simuleval.evaluator.evaluator.SentenceLevelEvaluator
+
+Quality Scorers
+---------------
+
+.. autoclass:: simuleval.evaluator.scorers.quality_scorer.SacreBLEUScorer
+.. autoclass:: simuleval.evaluator.scorers.quality_scorer.ASRSacreBLEUScorer
+
+Latency Scorers
+---------------
+
+.. autoclass:: simuleval.evaluator.scorers.latency_scorer.ALScorer
+ :members:
+
+.. autoclass:: simuleval.evaluator.scorers.latency_scorer.APScorer
+ :members:
+
+.. autoclass:: simuleval.evaluator.scorers.latency_scorer.DALScorer
+ :members:
+
+Customized Scorers
+------------------
+To add customized scorers, the user can use :code:`@register_latency_scorer` or :code:`@register_quality_scorer` to decorate a scorer class.
+and use :code:`--quality-metrics` and :code:`--latency-metrics` to call the scorer. For example:
+
+.. literalinclude:: ../../examples/quick_start/agent_with_new_metrics.py
+ :lines: 6-
+
+.. code-block:: bash
+
+ > simuleval --source source.txt --target target.txt --agent agent_with_new_metrics.py --latency-metrics RTF
+ 2022-12-06 12:56:01 | INFO | simuleval.cli | Evaluate system: DummyWaitkTextAgent
+ 2022-12-06 12:56:01 | INFO | simuleval.dataloader | Evaluating from text to text.
+ 2022-12-06 12:56:01 | INFO | simuleval.sentence_level_evaluator | Results:
+ BLEU RTF
+ 1.593 1.078
diff --git a/SimulEval/docs/user_guide/introduction.rst b/SimulEval/docs/user_guide/introduction.rst
new file mode 100644
index 0000000000000000000000000000000000000000..bbecc30ad8bed045b31e44aa1af7b26e57f17a2f
--- /dev/null
+++ b/SimulEval/docs/user_guide/introduction.rst
@@ -0,0 +1,35 @@
+Introduction
+============
+Different from offline translation system, the evaluation of simultaneous translation requires incremental decoding with an streaming input.
+The simultaneous introduce the a front-end / back-end setup, shown as follow.
+
+The back-end contains one or multiple user-defined agents which make decisions of whether to generate prediction at a certain point.
+The agent can also considered as queue, where the input are keep pushed in and policy decides the timing to pop the output.
+
+The front-end on the other side, represent the source of input and recipient of the system prediction.
+In deployment, the front-end can be web page or cell phone app.
+In SimulEval, the front-end is the evaluator , which feeds streaming input to back-end, receive prediction and track the delays.
+The front-end and back-end can run separately for different purpose.
+
+The evaluation process can summarized as follow pseudocode
+
+.. code-block:: python
+
+ for instance in evaluator.instances:
+ while not instance.finished:
+ input_segment = instance.send_source()
+ prediction = agent.pushpop(input_segment)
+ if prediction is not None:
+ instance.receive_prediction(prediction)
+
+ results = [scorer.score() for scorer in evaluate.scorers]
+
+
+
+The common usage of SimulEval is as follow
+
+.. code-block:: bash
+
+ simuleval DATALOADER_OPTIONS EVALUATOR_OPTIONS --agent $AGENT_FILE AGENT_OPTIONS
+
+We will introduce the usage of the toolkit based on these three major components: Agent, Dataloader and Evaluator.
\ No newline at end of file
diff --git a/SimulEval/examples/quick_start/Dockerfile b/SimulEval/examples/quick_start/Dockerfile
new file mode 100644
index 0000000000000000000000000000000000000000..bcb545abc4d4a6702882821193297e0e6b4400f1
--- /dev/null
+++ b/SimulEval/examples/quick_start/Dockerfile
@@ -0,0 +1,10 @@
+FROM python:3.8
+RUN apt-get update \
+&& apt-get upgrade -y \
+&& apt-get install -y \
+&& apt-get -y install apt-utils gcc libpq-dev libsndfile-dev
+RUN git clone https://github.com/facebookresearch/SimulEval.git
+WORKDIR SimulEval
+RUN git checkout v1.1.0
+RUN pip install -e .
+CMD ["simuleval", "--standalone", "--remote-port", "8888", "--agent", "examples/quick_start/first_agent.py.py"]
diff --git a/SimulEval/examples/quick_start/agent_pipeline.py b/SimulEval/examples/quick_start/agent_pipeline.py
new file mode 100644
index 0000000000000000000000000000000000000000..5cc523e649a4ec70f5b35f709f227b618beada63
--- /dev/null
+++ b/SimulEval/examples/quick_start/agent_pipeline.py
@@ -0,0 +1,39 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import random
+from simuleval.utils import entrypoint
+from simuleval.agents import TextToTextAgent
+from simuleval.agents.actions import ReadAction, WriteAction
+from simuleval.agents import AgentPipeline
+
+
+class DummyWaitkTextAgent(TextToTextAgent):
+ waitk = 0
+ vocab = [chr(i) for i in range(ord("A"), ord("Z") + 1)]
+
+ def policy(self):
+ lagging = len(self.states.source) - len(self.states.target)
+
+ if lagging >= self.waitk or self.states.source_finished:
+ prediction = random.choice(self.vocab)
+
+ return WriteAction(prediction, finished=(lagging <= 1))
+ else:
+ return ReadAction()
+
+
+class DummyWait2TextAgent(DummyWaitkTextAgent):
+ waitk = 2
+
+
+class DummyWait4TextAgent(DummyWaitkTextAgent):
+ waitk = 4
+
+
+@entrypoint
+class DummyPipeline(AgentPipeline):
+ pipeline = [DummyWait2TextAgent, DummyWait4TextAgent]
diff --git a/SimulEval/examples/quick_start/agent_with_configs.py b/SimulEval/examples/quick_start/agent_with_configs.py
new file mode 100644
index 0000000000000000000000000000000000000000..97f09469cd676404002eea057561ca72b11c8b67
--- /dev/null
+++ b/SimulEval/examples/quick_start/agent_with_configs.py
@@ -0,0 +1,39 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import random
+from simuleval.utils import entrypoint
+from simuleval.agents import TextToTextAgent
+from simuleval.agents.actions import ReadAction, WriteAction
+from argparse import Namespace, ArgumentParser
+
+
+@entrypoint
+class DummyWaitkTextAgent(TextToTextAgent):
+ def __init__(self, args: Namespace):
+ """Initialize your agent here.
+ For example loading model, vocab, etc
+ """
+ super().__init__(args)
+ self.waitk = args.waitk
+ with open(args.vocab) as f:
+ self.vocab = [line.strip() for line in f]
+
+ @staticmethod
+ def add_args(parser: ArgumentParser):
+ """Add customized command line arguments"""
+ parser.add_argument("--waitk", type=int, default=3)
+ parser.add_argument("--vocab", type=str)
+
+ def policy(self):
+ lagging = len(self.states.source) - len(self.states.target)
+
+ if lagging >= self.waitk or self.states.source_finished:
+ prediction = random.choice(self.vocab)
+
+ return WriteAction(prediction, finished=(lagging <= 1))
+ else:
+ return ReadAction()
diff --git a/SimulEval/examples/quick_start/agent_with_new_metrics.py b/SimulEval/examples/quick_start/agent_with_new_metrics.py
new file mode 100644
index 0000000000000000000000000000000000000000..ff38466483b1e214354198585461df89f4c024ab
--- /dev/null
+++ b/SimulEval/examples/quick_start/agent_with_new_metrics.py
@@ -0,0 +1,47 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import random
+from statistics import mean
+from simuleval.utils import entrypoint
+from simuleval.evaluator.scorers.latency_scorer import (
+ register_latency_scorer,
+ LatencyScorer,
+)
+from simuleval.agents import TextToTextAgent
+from simuleval.agents.actions import ReadAction, WriteAction
+
+
+@register_latency_scorer("RTF")
+class RTFScorer(LatencyScorer):
+ """Real time factor
+
+ Usage:
+ --latency-metrics RTF
+ """
+
+ def __call__(self, instances) -> float:
+ scores = []
+ for ins in instances.values():
+ scores.append(ins.delays[-1] / ins.source_length)
+
+ return mean(scores)
+
+
+@entrypoint
+class DummyWaitkTextAgent(TextToTextAgent):
+ waitk = 3
+ vocab = [chr(i) for i in range(ord("A"), ord("Z") + 1)]
+
+ def policy(self):
+ lagging = len(self.states.source) - len(self.states.target)
+
+ if lagging >= self.waitk or self.states.source_finished:
+ prediction = random.choice(self.vocab)
+
+ return WriteAction(prediction, finished=(lagging <= 1))
+ else:
+ return ReadAction()
diff --git a/SimulEval/examples/quick_start/dict.txt b/SimulEval/examples/quick_start/dict.txt
new file mode 100644
index 0000000000000000000000000000000000000000..a6f1d23fc5f9d1d41759f1679068071a140949b4
--- /dev/null
+++ b/SimulEval/examples/quick_start/dict.txt
@@ -0,0 +1,26 @@
+A
+B
+C
+D
+E
+F
+G
+H
+I
+J
+K
+L
+M
+N
+O
+P
+Q
+R
+S
+T
+U
+V
+W
+X
+Y
+Z
diff --git a/SimulEval/examples/quick_start/first_agent.py b/SimulEval/examples/quick_start/first_agent.py
new file mode 100644
index 0000000000000000000000000000000000000000..e75e5908aac1a56493720741335f3ca5ca96832d
--- /dev/null
+++ b/SimulEval/examples/quick_start/first_agent.py
@@ -0,0 +1,26 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import random
+from simuleval.utils import entrypoint
+from simuleval.agents import TextToTextAgent
+from simuleval.agents.actions import ReadAction, WriteAction
+
+
+@entrypoint
+class DummyWaitkTextAgent(TextToTextAgent):
+ waitk = 3
+ vocab = [chr(i) for i in range(ord("A"), ord("Z") + 1)]
+
+ def policy(self):
+ lagging = len(self.states.source) - len(self.states.target)
+
+ if lagging >= self.waitk or self.states.source_finished:
+ prediction = random.choice(self.vocab)
+
+ return WriteAction(prediction, finished=(lagging <= 1))
+ else:
+ return ReadAction()
diff --git a/SimulEval/examples/quick_start/readme.md b/SimulEval/examples/quick_start/readme.md
new file mode 100644
index 0000000000000000000000000000000000000000..0964761e045df4ae9cd519ff47c5128a0b370d91
--- /dev/null
+++ b/SimulEval/examples/quick_start/readme.md
@@ -0,0 +1,47 @@
+# Quick Start
+Following are some minimal examples to use SimulEval. More details can be found [here](https://simuleval.readthedocs.io/en/v1.1.0/quick_start.html).
+
+## First Agent
+To evaluate a text-to-text wait-3 system with random output:
+
+```
+> simuleval --source source.txt --target target.txt --agent first_agent.py
+
+2022-12-05 13:43:58 | INFO | simuleval.cli | Evaluate system: DummyWaitkTextAgent
+2022-12-05 13:43:58 | INFO | simuleval.dataloader | Evaluating from text to text.
+2022-12-05 13:43:58 | INFO | simuleval.sentence_level_evaluator | Results:
+BLEU AL AP DAL
+1.541 3.0 0.688 3.0
+
+```
+
+## Agent with Command Line Arguments
+```
+simuleval --source source.txt --target target.txt --agent agent_with_configs.py --waitk 3 --vocab dict.txt
+```
+
+## Agent Pipeline
+```
+simuleval --source source.txt --target target.txt --agent agent_pipeline.py
+```
+
+## Agent with New Metrics
+```
+simuleval --source source.txt --target target.txt --agent agent_with_new_metrics.py
+```
+
+## Standalone Agent & Remote Evaluation
+Start an agent server:
+```
+simuleval --standalone --remote-port 8888 --agent agent_with_new_metrics.py
+```
+Or with docker
+```
+docker build -t simuleval_agent .
+docker run -p 8888:8888 simuleval_agent:latest
+```
+
+Start a remote evaluator:
+```
+simuleval --remote-eval --source source.txt --target target.txt --source-type text --target-type text --remote-port 8888
+```
diff --git a/SimulEval/examples/quick_start/source.txt b/SimulEval/examples/quick_start/source.txt
new file mode 100644
index 0000000000000000000000000000000000000000..1a5cf9fdf2c7930b545c64c959a7a5f9585e64b5
--- /dev/null
+++ b/SimulEval/examples/quick_start/source.txt
@@ -0,0 +1,10 @@
+Z U S N B Y X Q L O T
+M A J F P G O Y V R H M Z O T A
+M O W A O I D H H B O F
+N Q N I P C O H A A
+G B O J H P W C I A L V
+P T Z D E E N T B Y G Z R K
+F S H U K R W K S B R K M B B Q F C O U
+M H O L W Z G J Y X J B I
+A V B F E S F E W Q C S
+F N O I E Z B R S C V N S
diff --git a/SimulEval/examples/quick_start/target.txt b/SimulEval/examples/quick_start/target.txt
new file mode 100644
index 0000000000000000000000000000000000000000..1cfac4ec4432c15948a61886611307b70ef882e8
--- /dev/null
+++ b/SimulEval/examples/quick_start/target.txt
@@ -0,0 +1,10 @@
+Z U S N B Y X Q L O T
+M A J F P G O Y V R H M Z O T A
+M O W A O I D H H B O F
+N Q N I P C O H A A
+G B O J H P W C I A L V
+P T Z D E E N T B Y G Z R K
+F S H U K R W K S B R K M B B Q F C O U
+M H O L W Z G J Y X J B I
+A V B F E S F E W Q C S
+F N O I E Z B R S C V N S
\ No newline at end of file
diff --git a/SimulEval/examples/speech_to_speech/english_counter_agent.py b/SimulEval/examples/speech_to_speech/english_counter_agent.py
new file mode 100644
index 0000000000000000000000000000000000000000..e335288bf7aa0e0a5e44fc48592caa0904f3afd3
--- /dev/null
+++ b/SimulEval/examples/speech_to_speech/english_counter_agent.py
@@ -0,0 +1,69 @@
+from simuleval.utils import entrypoint
+from simuleval.data.segments import SpeechSegment
+from simuleval.agents import SpeechToSpeechAgent
+from simuleval.agents.actions import WriteAction, ReadAction
+from fairseq.checkpoint_utils import load_model_ensemble_and_task_from_hf_hub
+from fairseq.models.text_to_speech.hub_interface import TTSHubInterface
+
+
+class TTSModel:
+ def __init__(self):
+ models, cfg, task = load_model_ensemble_and_task_from_hf_hub(
+ "facebook/fastspeech2-en-ljspeech",
+ arg_overrides={"vocoder": "hifigan", "fp16": False},
+ )
+ TTSHubInterface.update_cfg_with_data_cfg(cfg, task.data_cfg)
+ self.tts_generator = task.build_generator(models, cfg)
+ self.tts_task = task
+ self.tts_model = models[0]
+ self.tts_model.to("cpu")
+ self.tts_generator.vocoder.to("cpu")
+
+ def synthesize(self, text):
+ sample = TTSHubInterface.get_model_input(self.tts_task, text)
+ if sample["net_input"]["src_lengths"][0] == 0:
+ return [], 0
+ for key in sample["net_input"].keys():
+ if sample["net_input"][key] is not None:
+ sample["net_input"][key] = sample["net_input"][key].to("cpu")
+
+ wav, rate = TTSHubInterface.get_prediction(
+ self.tts_task, self.tts_model, self.tts_generator, sample
+ )
+ wav = wav.tolist()
+ return wav, rate
+
+
+@entrypoint
+class EnglishSpeechCounter(SpeechToSpeechAgent):
+ """
+ Incrementally feed text to this offline Fastspeech2 TTS model,
+ with a minimum numbers of phonemes every chunk.
+ """
+
+ def __init__(self, args):
+ super().__init__(args)
+ self.wait_seconds = args.wait_seconds
+ self.tts_model = TTSModel()
+
+ @staticmethod
+ def add_args(parser):
+ parser.add_argument("--wait-seconds", default=1, type=int)
+
+ def policy(self):
+ length_in_seconds = round(
+ len(self.states.source) / self.states.source_sample_rate
+ )
+ if not self.states.source_finished and length_in_seconds < self.wait_seconds:
+ return ReadAction()
+ samples, fs = self.tts_model.synthesize(f"{length_in_seconds} mississippi")
+
+ # A SpeechSegment has to be returned for speech-to-speech translation system
+ return WriteAction(
+ SpeechSegment(
+ content=samples,
+ sample_rate=fs,
+ finished=self.states.source_finished,
+ ),
+ finished=self.states.source_finished,
+ )
diff --git a/SimulEval/examples/speech_to_speech/eval.sh b/SimulEval/examples/speech_to_speech/eval.sh
new file mode 100644
index 0000000000000000000000000000000000000000..f0f08c0d5c0af6bb8cc0d08be7f32613f96d8ae5
--- /dev/null
+++ b/SimulEval/examples/speech_to_speech/eval.sh
@@ -0,0 +1,6 @@
+simuleval \
+ --agent english_counter_agent.py --output output \
+ --source source.txt --target reference/en.txt --source-segment-size 1000\
+ --quality-metrics WHISPER_ASR_BLEU \
+ --target-speech-lang en --transcript-lowercase --transcript-non-punctuation --whisper-model-size large \
+ --latency-metrics StartOffset EndOffset ATD
diff --git a/SimulEval/examples/speech_to_speech/readme.md b/SimulEval/examples/speech_to_speech/readme.md
new file mode 100644
index 0000000000000000000000000000000000000000..d328afcb0450165132a1c36b3d51568031795ada
--- /dev/null
+++ b/SimulEval/examples/speech_to_speech/readme.md
@@ -0,0 +1,66 @@
+## Simultaneous Speech-to-Speech Translation
+
+This tutorial provides a minimal example on how to evaluate a simultaneous speech-to-speech translation system.
+
+### Requirements
+
+To run this example, the following package is required
+
+- [`whisper`](https://github.com/openai/whisper): for quality evaluation (`WHISPER_ASR_BLEU`).
+
+### Agent
+
+The speech-to-speech agent ([english_counter_agent.py](english_counter_agent.py)) in this example is a counter, which generates a piece of audio every second after an initial wait.
+The policy of the agent is show follow. The agent will wait for `self.wait_seconds` seconds,
+and generate the audio of `{length_in_seconds} mississippi` every second afterward.
+
+```python
+ def policy(self):
+ length_in_seconds = round(
+ len(self.states.source) / self.states.source_sample_rate
+ )
+ if not self.states.source_finished and length_in_seconds < self.wait_seconds:
+ return ReadAction()
+ print(length_in_seconds)
+ samples, fs = self.tts_model.synthesize(f"{length_in_seconds} mississippi")
+
+ # A SpeechSegment has to be returned for speech-to-speech translation system
+ return WriteAction(
+ SpeechSegment(
+ content=samples,
+ sample_rate=fs,
+ finished=self.states.source_finished,
+ ),
+ finished=self.states.source_finished,
+ )
+```
+
+Notice that for speech output agent, the `WriteAction` has to contain a `SpeechSegment` class.
+
+### Evaluation
+
+The following command will start an evaluation
+
+```bash
+simuleval \
+ --agent english_counter_agent.py --output output \
+ --source source.txt --target reference/en.txt --source-segment-size 1000\
+ --quality-metrics WHISPER_ASR_BLEU \
+ --target-speech-lang en --transcript-lowercase --transcript-non-punctuation\
+ --latency-metrics StartOffset EndOffset ATD
+```
+
+For quality evaluation, we use ASR_BLEU, that is transcribing the speech output and compute BLEU score with the reference text. To use this feature, `whisper` has to be installed.
+
+We use three metrics for latency evaluation
+
+- `StartOffset`: The starting offset of translation comparing with source audio
+- `EndOffset`: The ending offset of translation comparing with source audio
+- `ATD`: Average Token Delay
+
+The results of the evaluation should be as following. The transcripts and alignments can be found in the `output` directory.
+
+```
+ WHISPER_ASR_BLEU StartOffset EndOffset ATD
+ 100.0 1000.0 1490.703 1248.261
+```
diff --git a/SimulEval/examples/speech_to_speech/reference/de.txt b/SimulEval/examples/speech_to_speech/reference/de.txt
new file mode 100644
index 0000000000000000000000000000000000000000..c8ffdacfb94d38c882113f910c3d737a12a51b86
--- /dev/null
+++ b/SimulEval/examples/speech_to_speech/reference/de.txt
@@ -0,0 +1 @@
+ein Mississippi zwei Mississippi drei Mississippi vier Mississippi fünf Mississippi sechs Mississippi sieben Mississippi
diff --git a/SimulEval/examples/speech_to_speech/reference/en.txt b/SimulEval/examples/speech_to_speech/reference/en.txt
new file mode 100644
index 0000000000000000000000000000000000000000..874a7d74113b6f64e76cc9666ff4b47cff53b93b
--- /dev/null
+++ b/SimulEval/examples/speech_to_speech/reference/en.txt
@@ -0,0 +1 @@
+one mississippi two mississippi three mississippi four mississippi five mississippi six mississippi seven mississippi
diff --git a/SimulEval/examples/speech_to_speech/reference/ja.txt b/SimulEval/examples/speech_to_speech/reference/ja.txt
new file mode 100644
index 0000000000000000000000000000000000000000..3c068c144446e0c077be79e170ed37c013e5b07b
--- /dev/null
+++ b/SimulEval/examples/speech_to_speech/reference/ja.txt
@@ -0,0 +1 @@
+1 ミシシッピ 2 ミシシッピ 3 ミシシッピ 4 ミシシッピ 5 ミシシッピ 6 ミシシッピ 7 ミシシッピ
diff --git a/SimulEval/examples/speech_to_speech/reference/zh.txt b/SimulEval/examples/speech_to_speech/reference/zh.txt
new file mode 100644
index 0000000000000000000000000000000000000000..b7ec39d1712aeb3a1846748f9e7b2f799d7e7808
--- /dev/null
+++ b/SimulEval/examples/speech_to_speech/reference/zh.txt
@@ -0,0 +1 @@
+一密西西比二密西西比三密西西比四密西西比五密西西比六密西西比七密西西比
diff --git a/SimulEval/examples/speech_to_speech/source.txt b/SimulEval/examples/speech_to_speech/source.txt
new file mode 100644
index 0000000000000000000000000000000000000000..607adc1323937ebc5ee0c5442a56a1b25cb9252b
--- /dev/null
+++ b/SimulEval/examples/speech_to_speech/source.txt
@@ -0,0 +1 @@
+test.wav
diff --git a/SimulEval/examples/speech_to_speech/test.wav b/SimulEval/examples/speech_to_speech/test.wav
new file mode 100644
index 0000000000000000000000000000000000000000..e4696d5662b2e2b3928107a55177cc3a0230f55c
--- /dev/null
+++ b/SimulEval/examples/speech_to_speech/test.wav
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:354152b48e1575f4ca3590b76700d46ef302dca9c183b5526f810c636c5d88f7
+size 302124
diff --git a/SimulEval/setup.cfg b/SimulEval/setup.cfg
new file mode 100644
index 0000000000000000000000000000000000000000..d884a39c56cc5f020cfca27534ec3b50490f4097
--- /dev/null
+++ b/SimulEval/setup.cfg
@@ -0,0 +1,2 @@
+[tool:pytest]
+flake8-max-line-length = 127
diff --git a/SimulEval/setup.py b/SimulEval/setup.py
new file mode 100644
index 0000000000000000000000000000000000000000..64aae337f81fc0c31251223179c68e45ec2972db
--- /dev/null
+++ b/SimulEval/setup.py
@@ -0,0 +1,43 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import setuptools
+
+setuptools.setup(
+ python_requires=">3.7.0",
+ name="simuleval",
+ version="1.1.0",
+ author="Xutai Ma",
+ entry_points={
+ "console_scripts": [
+ "simuleval = simuleval.cli:main",
+ ],
+ },
+ install_requires=[
+ "pytest",
+ "pytest-cov",
+ "sacrebleu==2.3.1",
+ "tornado",
+ "soundfile",
+ "pandas",
+ "requests",
+ "pytest-flake8",
+ "textgrid",
+ "tqdm==4.64.1",
+ "pyyaml",
+ "bitarray==2.6.0",
+ "yt-dlp",
+ "pydub",
+ ],
+ package=setuptools.find_packages(
+ exclude=[
+ "examples",
+ "examples.*",
+ "docs",
+ "docs.*",
+ ]
+ ),
+)
diff --git a/SimulEval/simuleval/__init__.py b/SimulEval/simuleval/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..5656d59e07f3fa33dd3bad1a0f9279ff4b8a6128
--- /dev/null
+++ b/SimulEval/simuleval/__init__.py
@@ -0,0 +1,5 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
diff --git a/SimulEval/simuleval/agents/__init__.py b/SimulEval/simuleval/agents/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..8fc3dfcb6bc95f452cc7b278fcf5315c815a36b6
--- /dev/null
+++ b/SimulEval/simuleval/agents/__init__.py
@@ -0,0 +1,16 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from .agent import ( # noqa
+ GenericAgent,
+ SpeechToTextAgent,
+ SpeechToSpeechAgent,
+ TextToSpeechAgent,
+ TextToTextAgent,
+)
+from .states import AgentStates # noqa
+from .actions import Action, ReadAction, WriteAction # noqa
+from .pipeline import AgentPipeline # noqa
diff --git a/SimulEval/simuleval/agents/actions.py b/SimulEval/simuleval/agents/actions.py
new file mode 100644
index 0000000000000000000000000000000000000000..81e3a4c50bb0b4e4f49fad76d5e209b9c0ebffcb
--- /dev/null
+++ b/SimulEval/simuleval/agents/actions.py
@@ -0,0 +1,60 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from typing import Union, List
+from dataclasses import dataclass
+from simuleval.data.segments import Segment
+
+
+class Action:
+ """
+ Abstract Action class
+ """
+
+ def is_read(self) -> bool:
+ """
+ Whether the action is a read action
+
+ Returns:
+ bool: True if the action is a read action.
+ """
+ assert NotImplementedError
+
+
+class ReadAction(Action):
+ """
+ Action to return when policy decide to read one more source segment.
+ The only way to use it is to return :code:`ReadAction()`
+ """
+
+ def is_read(self) -> bool:
+ return True
+
+ def __repr__(self) -> str:
+ return "ReadAction()"
+
+
+@dataclass
+class WriteAction(Action):
+ """
+ Action to return when policy decide to generate a prediction
+
+ Args:
+ content (Union[str, List[float]]): The prediction.
+ finished (bool): Indicates if current sentence is finished.
+
+ .. note:: For text the prediction a str; for speech, it's a list.
+
+ """
+
+ content: Union[str, List[float], Segment]
+ finished: bool
+
+ def is_read(self) -> bool:
+ return False
+
+ def __repr__(self) -> str:
+ return f"WriteAction({self.content}, finished={self.finished})"
diff --git a/SimulEval/simuleval/agents/agent.py b/SimulEval/simuleval/agents/agent.py
new file mode 100644
index 0000000000000000000000000000000000000000..c1e7b18383f21004030587234b097e95397013e8
--- /dev/null
+++ b/SimulEval/simuleval/agents/agent.py
@@ -0,0 +1,216 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from inspect import signature
+from argparse import Namespace, ArgumentParser
+from simuleval.data.segments import Segment, TextSegment, SpeechSegment, EmptySegment
+from typing import Optional
+from .states import AgentStates
+from .actions import Action
+
+
+SEGMENT_TYPE_DICT = {"text": TextSegment, "speech": SpeechSegment}
+
+
+class GenericAgent:
+ """
+ Generic Agent class.
+ """
+
+ source_type = None
+ target_type = None
+
+ def __init__(self, args: Optional[Namespace] = None) -> None:
+ if args is not None:
+ self.args = args
+ assert self.source_type
+ assert self.target_type
+ self.device = "cpu"
+
+ self.states = self.build_states()
+ self.reset()
+
+ def build_states(self) -> AgentStates:
+ """
+ Build states instance for agent
+
+ Returns:
+ AgentStates: agent states
+ """
+ return AgentStates()
+
+ def reset(self) -> None:
+ """
+ Reset agent, called every time when a new sentence coming in.
+ """
+ self.states.reset()
+
+ def policy(self, states: Optional[AgentStates] = None) -> Action:
+ """
+ The policy to make decision every time
+ when the system has new input.
+ The function has to return an Action instance
+
+ Args:
+ states (Optional[AgentStates]): an optional states for stateless agent
+
+ Returns:
+ Action: The actions to make at certain point.
+
+ .. note:
+
+ WriteAction means that the system has a prediction.
+ ReadAction means that the system needs more source.
+ When states are provided, the agent will become stateless and ignore self.states.
+ """
+ assert NotImplementedError
+
+ def push(
+ self, source_segment: Segment, states: Optional[AgentStates] = None
+ ) -> None:
+ """
+ The function to process the incoming information.
+
+ Args:
+ source_info (dict): incoming information dictionary
+ states (Optional[AgentStates]): an optional states for stateless agent
+ """
+ if states is None:
+ states = self.states
+ states.update_source(source_segment)
+
+ def pop(self, states: Optional[AgentStates] = None) -> Segment:
+ """
+ The function to generate system output.
+ By default, it first runs policy,
+ and than returns the output segment.
+ If the policy decide to read,
+ it will return an empty segment.
+
+ Args:
+ states (Optional[AgentStates]): an optional states for stateless agent
+
+ Returns:
+ Segment: segment to return.
+ """
+ if len(signature(self.policy).parameters) == 0:
+ is_stateless = False
+ if states:
+ raise RuntimeError("Feeding states to stateful agents.")
+ else:
+ is_stateless = True
+
+ if states is None:
+ states = self.states
+
+ if states.target_finished:
+ return EmptySegment(finished=True)
+
+ if is_stateless:
+ action = self.policy(states)
+ else:
+ action = self.policy()
+
+ if not isinstance(action, Action):
+ raise RuntimeError(
+ f"The return value of {self.policy.__qualname__} is not an {Action.__qualname__} instance"
+ )
+ if action.is_read():
+ return EmptySegment()
+ else:
+ if isinstance(action.content, Segment):
+ return action.content
+
+ segment = SEGMENT_TYPE_DICT[self.target_type](
+ index=0, content=action.content, finished=action.finished
+ )
+ states.update_target(segment)
+ return segment
+
+ def pushpop(
+ self, segment: Segment, states: Optional[AgentStates] = None
+ ) -> Segment:
+ """
+ Operate pop immediately after push.
+
+ Args:
+ segment (Segment): input segment
+
+ Returns:
+ Segment: output segment
+ """
+ self.push(segment, states)
+ return self.pop(states)
+
+ @staticmethod
+ def add_args(parser: ArgumentParser):
+ """
+ Add agent arguments to parser.
+ Has to be a static method.
+
+ Args:
+ parser (ArgumentParser): cli argument parser
+ """
+ pass
+
+ @classmethod
+ def from_args(cls, args):
+ return cls(args)
+
+ def to(self, device: str, *args, **kwargs) -> None:
+ """
+ Move agent to specified device.
+
+ Args:
+ device (str): Device to move agent to.
+ """
+ pass
+
+ def __repr__(self) -> str:
+ return f"{self.__class__.__name__}[{self.source_type} -> {self.target_type}]"
+
+ def __str__(self) -> str:
+ return self.__repr__()
+
+
+class SpeechToTextAgent(GenericAgent):
+ """
+ Same as generic agent, but with explicit types
+ speech -> text
+ """
+
+ source_type: str = "speech"
+ target_type: str = "text"
+
+
+class SpeechToSpeechAgent(GenericAgent):
+ """
+ Same as generic agent, but with explicit types
+ speech -> speech
+ """
+
+ source_type: str = "speech"
+ target_type: str = "speech"
+
+
+class TextToSpeechAgent(GenericAgent):
+ """
+ Same as generic agent, but with explicit types
+ text -> speech
+ """
+
+ source_type: str = "text"
+ target_type: str = "speech"
+
+
+class TextToTextAgent(GenericAgent):
+ """
+ Same as generic agent, but with explicit types
+ text -> text
+ """
+
+ source_type: str = "text"
+ target_type: str = "text"
diff --git a/SimulEval/simuleval/agents/pipeline.py b/SimulEval/simuleval/agents/pipeline.py
new file mode 100644
index 0000000000000000000000000000000000000000..1a7d9f4b9946d80437809dc99f6c084e84ef34f3
--- /dev/null
+++ b/SimulEval/simuleval/agents/pipeline.py
@@ -0,0 +1,90 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from typing import List, Optional
+from simuleval.data.segments import Segment
+from .agent import GenericAgent, AgentStates
+
+
+class AgentPipeline(GenericAgent):
+ """A pipeline of agents
+
+ Attributes:
+ pipeline (list): a list of agent classes.
+
+ """
+
+ pipeline: List = []
+
+ def __init__(self, module_list: List[GenericAgent]) -> None:
+ self.module_list = module_list
+ self.check_pipeline_types()
+
+ def check_pipeline_types(self):
+ if len(self.pipeline) > 1:
+ for i in range(1, len(self.pipeline)):
+ if (
+ self.module_list[i].source_type
+ != self.module_list[i - 1].target_type
+ ):
+ raise RuntimeError(
+ f"{self.module_list[i]}.source_type({self.module_list[i].source_type}) != {self.pipeline[i-1]}.target_type({self.pipeline[i - 1].target_type}" # noqa F401
+ )
+
+ @property
+ def source_type(self) -> Optional[str]:
+ return self.module_list[0].source_type
+
+ @property
+ def target_type(self) -> Optional[str]:
+ return self.module_list[-1].target_type
+
+ def reset(self) -> None:
+ for module in self.module_list:
+ module.reset()
+
+ def build_states(self) -> List[AgentStates]:
+ return [module.build_states() for module in self.module_list]
+
+ def push(
+ self, segment: Segment, states: Optional[List[Optional[AgentStates]]] = None
+ ) -> None:
+ if states is None:
+ states = [None for _ in self.module_list]
+ else:
+ assert len(states) == len(self.module_list)
+
+ for index, module in enumerate(self.module_list[:-1]):
+ segment = module.pushpop(segment, states[index])
+ self.module_list[-1].push(segment, states[-1])
+
+ def pop(self, states: Optional[List[Optional[AgentStates]]] = None) -> Segment:
+ if states is None:
+ last_states = None
+ else:
+ assert len(states) == len(self.module_list)
+ last_states = states[-1]
+
+ return self.module_list[-1].pop(last_states)
+
+ @classmethod
+ def add_args(cls, parser) -> None:
+ for module_class in cls.pipeline:
+ module_class.add_args(parser)
+
+ @classmethod
+ def from_args(cls, args):
+ assert len(cls.pipeline) > 0
+ return cls([module_class.from_args(args) for module_class in cls.pipeline])
+
+ def __repr__(self) -> str:
+ pipline_str = "\n\t".join(
+ "\t".join(str(module).splitlines(True)) for module in self.module_list
+ )
+ return f"{self.__class__.__name__}(\n\t{pipline_str}\n)"
+
+ def __str__(self) -> str:
+ return self.__repr__()
diff --git a/SimulEval/simuleval/agents/service.py b/SimulEval/simuleval/agents/service.py
new file mode 100644
index 0000000000000000000000000000000000000000..62e09c79fdab936f0574afe7a105d92ec0ec8a6a
--- /dev/null
+++ b/SimulEval/simuleval/agents/service.py
@@ -0,0 +1,59 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+import os
+import json
+import logging
+from tornado import web, ioloop
+from simuleval.data.segments import segment_from_json_string
+from simuleval import options
+
+logger = logging.getLogger("simuleval.agent_server")
+
+
+class SystemHandler(web.RequestHandler):
+ def initialize(self, system):
+ self.system = system
+
+ def get(self):
+ self.write(json.dumps({"info": str(self.system)}))
+
+
+class ResetHandle(SystemHandler):
+ def post(self):
+ self.system.reset()
+
+
+class OutputHandler(SystemHandler):
+ def get(self):
+ output_segment = self.system.pop()
+ self.write(output_segment.json())
+
+
+class InputHandler(SystemHandler):
+ def put(self):
+ segment = segment_from_json_string(self.request.body)
+ self.system.push(segment)
+
+
+def start_agent_service(system):
+ parser = options.general_parser()
+ options.add_evaluator_args(parser)
+ args, _ = parser.parse_known_args()
+ app = web.Application(
+ [
+ (r"/reset", ResetHandle, {"system": system}),
+ (r"/input", InputHandler, {"system": system}),
+ (r"/output", OutputHandler, {"system": system}),
+ (r"/", SystemHandler, {"system": system}),
+ ],
+ debug=False,
+ )
+
+ app.listen(args.remote_port, max_buffer_size=1024**3)
+
+ logger.info(
+ f"Simultaneous Translation Server Started (process id {os.getpid()}). Listening to port {args.remote_port} "
+ )
+ ioloop.IOLoop.current().start()
diff --git a/SimulEval/simuleval/agents/states.py b/SimulEval/simuleval/agents/states.py
new file mode 100644
index 0000000000000000000000000000000000000000..2c01e8e77da62b7c4d79b156d39c572fa40711c5
--- /dev/null
+++ b/SimulEval/simuleval/agents/states.py
@@ -0,0 +1,68 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from simuleval.data.segments import Segment, TextSegment, EmptySegment, SpeechSegment
+
+
+class AgentStates:
+ """
+ Tracker of the decoding progress.
+
+ Attributes:
+ source (list): current source sequence.
+ target (list): current target sequence.
+ source_finished (bool): if the source is finished.
+ target_finished (bool): if the target is finished.
+ """
+
+ def __init__(self) -> None:
+ self.reset()
+
+ def reset(self) -> None:
+ """Reset Agent states"""
+ self.source = []
+ self.target = []
+ self.source_finished = False
+ self.target_finished = False
+ self.source_sample_rate = 0
+ self.target_sample_rate = 0
+
+ def update_source(self, segment: Segment):
+ """
+ Update states from input segment
+
+ Args:
+ segment (~simuleval.agents.segments.Segment): input segment
+ """
+ self.source_finished = segment.finished
+ if isinstance(segment, EmptySegment):
+ return
+ elif isinstance(segment, TextSegment):
+ self.source.append(segment.content)
+ elif isinstance(segment, SpeechSegment):
+ self.source += segment.content
+ self.source_sample_rate = segment.sample_rate
+ else:
+ raise NotImplementedError
+
+ def update_target(self, segment: Segment):
+ """
+ Update states from output segment
+
+ Args:
+ segment (~simuleval.agents.segments.Segment): input segment
+ """
+ self.target_finished = segment.finished
+ if not self.target_finished:
+ if isinstance(segment, EmptySegment):
+ return
+ elif isinstance(segment, TextSegment):
+ self.target.append(segment.content)
+ elif isinstance(segment, SpeechSegment):
+ self.target += segment.content
+ self.target_sample_rate = segment.sample_rate
+ else:
+ raise NotImplementedError
diff --git a/SimulEval/simuleval/analysis/curve.py b/SimulEval/simuleval/analysis/curve.py
new file mode 100644
index 0000000000000000000000000000000000000000..b0ab879a5921e4616a268d27595fa4652a702e3f
--- /dev/null
+++ b/SimulEval/simuleval/analysis/curve.py
@@ -0,0 +1,129 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import json
+import pandas
+from pathlib import Path
+from typing import Dict, List, Union
+
+
+class SimulEvalResults:
+ def __init__(self, path: Union[Path, str]) -> None:
+ self.path = Path(path)
+ scores_path = self.path / "scores"
+ if scores_path.exists():
+ self.is_finished = True
+ with open(self.path / "scores") as f:
+ self.scores = json.load(f)
+ else:
+ self.is_finished = False
+ self.scores = {}
+
+ @property
+ def quality(self) -> float:
+ if self.is_finished:
+ if self.scores is None:
+ return 0
+ return self.scores["Quality"]["BLEU"]
+ else:
+ return 0
+
+ @property
+ def bleu(self) -> float:
+ return self.quality
+
+ @property
+ def latency(self) -> Dict[str, float]:
+ if self.is_finished:
+ return self.scores["Latency"]
+ else:
+ return {}
+
+ @property
+ def average_lagging(self):
+ return self.latency.get("AL", 0)
+
+ @property
+ def average_lagging_ca(self):
+ return self.latency.get("AL_CA", 0)
+
+ @property
+ def average_proportion(self):
+ return self.latency.get("AP", 0)
+
+ @property
+ def name(self):
+ return self.path.name
+
+
+class S2SSimulEvalResults(SimulEvalResults):
+ @property
+ def bow_average_lagging(self):
+ return self.latency.get("BOW", {}).get("AL", 0)
+
+ @property
+ def cow_average_lagging(self):
+ return self.latency.get("COW", {}).get("AL", 0)
+
+ @property
+ def eow_average_lagging(self):
+ return self.latency.get("EOW", {}).get("AL", 0)
+
+
+class QualityLatencyAnalyzer:
+ def __init__(self) -> None:
+ self.score_list: List[SimulEvalResults] = []
+
+ def add_scores_from_path(self, path: Path):
+ self.score_list.append(SimulEvalResults(path))
+
+ @classmethod
+ def from_paths(cls, path_list: List[Path]):
+ analyzer = cls()
+ for path in path_list:
+ analyzer.add_scores_from_path(path)
+ return analyzer
+
+ def summarize(self):
+ results = []
+ for score in self.score_list:
+ if score.bleu == 0:
+ continue
+ results.append(
+ [
+ score.name,
+ round(score.average_lagging / 1000, 2),
+ round(score.average_lagging_ca / 1000, 2),
+ round(score.average_proportion, 2),
+ round(score.bleu, 2),
+ ]
+ )
+ results.sort(key=lambda x: x[1])
+ return pandas.DataFrame(results, columns=["name", "AL", "AL(CA)", "AP", "BLEU"])
+
+
+class S2SQualityLatencyAnalyzer(QualityLatencyAnalyzer):
+ def add_scores_from_path(self, path: Path):
+ self.score_list.append(S2SSimulEvalResults(path))
+
+ def summarize(self):
+ results = []
+ for score in self.score_list:
+ if score.bleu == 0:
+ continue
+ results.append(
+ [
+ score.name,
+ round(score.bow_average_lagging / 1000, 2),
+ round(score.cow_average_lagging / 1000, 2),
+ round(score.eow_average_lagging / 1000, 2),
+ round(score.bleu, 2),
+ ]
+ )
+ results.sort(key=lambda x: x[1])
+ return pandas.DataFrame(
+ results, columns=["name", "BOW_AL", "COW_AL", "EOW_AL", "BLEU"]
+ )
diff --git a/SimulEval/simuleval/cli.py b/SimulEval/simuleval/cli.py
new file mode 100644
index 0000000000000000000000000000000000000000..5191ce561185e1b41af8870a94427d60211bad9c
--- /dev/null
+++ b/SimulEval/simuleval/cli.py
@@ -0,0 +1,98 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import sys
+import logging
+from simuleval import options
+from simuleval.utils.agent import build_system_args
+from simuleval.utils.slurm import submit_slurm_job
+from simuleval.utils.arguments import check_argument
+from simuleval.utils import EVALUATION_SYSTEM_LIST
+from simuleval.evaluator import (
+ build_evaluator,
+ build_remote_evaluator,
+ SentenceLevelEvaluator,
+)
+from simuleval.agents.service import start_agent_service
+from simuleval.agents import GenericAgent
+
+
+logging.basicConfig(
+ format="%(asctime)s | %(levelname)-8s | %(name)-16s | %(message)s",
+ datefmt="%Y-%m-%d %H:%M:%S",
+ level=logging.INFO,
+ stream=sys.stderr,
+)
+
+
+logger = logging.getLogger("simuleval.cli")
+
+
+def main():
+ if check_argument("remote_eval"):
+ remote_evaluate()
+ return
+
+ if check_argument("score_only"):
+ scoring()
+ return
+
+ if check_argument("slurm"):
+ submit_slurm_job()
+ return
+
+ system, args = build_system_args()
+
+ if check_argument("standalone"):
+ start_agent_service(system)
+ return
+
+ # build evaluator
+ evaluator = build_evaluator(args)
+ # evaluate system
+ evaluator(system)
+
+
+def evaluate(system_class: GenericAgent, config_dict: dict = {}):
+ EVALUATION_SYSTEM_LIST.append(system_class)
+
+ if check_argument("slurm", config_dict):
+ submit_slurm_job(config_dict)
+ return
+
+ system, args = build_system_args(config_dict)
+
+ # build evaluator
+ evaluator = build_evaluator(args)
+ # evaluate system
+ evaluator(system)
+
+
+def scoring():
+ parser = options.general_parser()
+ options.add_evaluator_args(parser)
+ options.add_scorer_args(parser)
+ options.add_dataloader_args(parser)
+ args = parser.parse_args()
+ evaluator = SentenceLevelEvaluator.from_args(args)
+ print(evaluator.results)
+
+
+def remote_evaluate():
+ # build evaluator
+ parser = options.general_parser()
+ options.add_dataloader_args(parser)
+ options.add_evaluator_args(parser)
+ options.add_scorer_args(parser)
+ args = parser.parse_args()
+ evaluator = build_remote_evaluator(args)
+
+ # evaluate system
+ evaluator.remote_eval()
+
+
+if __name__ == "__main__":
+ main()
diff --git a/SimulEval/simuleval/data/__init__.py b/SimulEval/simuleval/data/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..3ffe1c7bdb6d8273bdf52c6e4fbb0ab3c6864745
--- /dev/null
+++ b/SimulEval/simuleval/data/__init__.py
@@ -0,0 +1 @@
+from .dataloader import build_dataloader # noqa
diff --git a/SimulEval/simuleval/data/dataloader/__init__.py b/SimulEval/simuleval/data/dataloader/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..e4c1dea8e91402e8a5a159290623d9861875ef2c
--- /dev/null
+++ b/SimulEval/simuleval/data/dataloader/__init__.py
@@ -0,0 +1,35 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import logging
+from .dataloader import ( # noqa
+ GenericDataloader,
+ register_dataloader,
+ register_dataloader_class,
+ SUPPORTED_MEDIUM,
+ SUPPORTED_SOURCE_MEDIUM,
+ SUPPORTED_TARGET_MEDIUM,
+ DATALOADER_DICT,
+)
+from .t2t_dataloader import TextToTextDataloader # noqa
+from .s2t_dataloader import SpeechToTextDataloader # noqa
+
+
+logger = logging.getLogger("simuleval.dataloader")
+
+
+def build_dataloader(args) -> GenericDataloader:
+ dataloader_key = getattr(args, "dataloader", None)
+ if dataloader_key is not None:
+ assert dataloader_key in DATALOADER_DICT, f"{dataloader_key} is not defined"
+ logger.info(f"Evaluating from dataloader {dataloader_key}.")
+ return DATALOADER_DICT[dataloader_key].from_args(args)
+
+ assert args.source_type in SUPPORTED_SOURCE_MEDIUM
+ assert args.target_type in SUPPORTED_TARGET_MEDIUM
+
+ logger.info(f"Evaluating from {args.source_type} to {args.target_type}.")
+ return DATALOADER_DICT[f"{args.source_type}-to-{args.target_type}"].from_args(args)
diff --git a/SimulEval/simuleval/data/dataloader/dataloader.py b/SimulEval/simuleval/data/dataloader/dataloader.py
new file mode 100644
index 0000000000000000000000000000000000000000..518ed7514dd86c47f470ad0e5446d0dd4bd7a0b1
--- /dev/null
+++ b/SimulEval/simuleval/data/dataloader/dataloader.py
@@ -0,0 +1,97 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from typing import Any, Dict, List, Union
+from argparse import Namespace, ArgumentParser
+
+SUPPORTED_MEDIUM = ["text", "speech"]
+SUPPORTED_SOURCE_MEDIUM = ["youtube", "text", "speech"]
+SUPPORTED_TARGET_MEDIUM = ["text", "speech"]
+DATALOADER_DICT = {}
+
+
+def register_dataloader(name):
+ def register(cls):
+ DATALOADER_DICT[name] = cls
+ return cls
+
+ return register
+
+
+def register_dataloader_class(name, cls):
+ DATALOADER_DICT[name] = cls
+
+
+class GenericDataloader:
+ """
+ Load source and target data
+
+ .. argparse::
+ :ref: simuleval.options.add_data_args
+ :passparser:
+ :prog:
+
+ """
+
+ def __init__(
+ self, source_list: List[str], target_list: Union[List[str], List[None]]
+ ) -> None:
+ self.source_list = source_list
+ self.target_list = target_list
+ assert len(self.source_list) == len(self.target_list)
+
+ def __len__(self):
+ return len(self.source_list)
+
+ def get_source(self, index: int) -> Any:
+ return self.preprocess_source(self.source_list[index])
+
+ def get_target(self, index: int) -> Any:
+ return self.preprocess_target(self.target_list[index])
+
+ def __getitem__(self, index: int) -> Dict[str, Any]:
+ return {"source": self.get_source(index), "target": self.get_target(index)}
+
+ def preprocess_source(self, source: Any) -> Any:
+ raise NotImplementedError
+
+ def preprocess_target(self, target: Any) -> Any:
+ raise NotImplementedError
+
+ @classmethod
+ def from_args(cls, args: Namespace):
+ return cls(args.source, args.target)
+
+ @staticmethod
+ def add_args(parser: ArgumentParser):
+ parser.add_argument(
+ "--source",
+ type=str,
+ help="Source file.",
+ )
+ parser.add_argument(
+ "--target",
+ type=str,
+ help="Target file.",
+ )
+ parser.add_argument(
+ "--source-type",
+ type=str,
+ choices=SUPPORTED_SOURCE_MEDIUM,
+ help="Source Data type to evaluate.",
+ )
+ parser.add_argument(
+ "--target-type",
+ type=str,
+ choices=SUPPORTED_TARGET_MEDIUM,
+ help="Data type to evaluate.",
+ )
+ parser.add_argument(
+ "--source-segment-size",
+ type=int,
+ default=1,
+ help="Source segment size, For text the unit is # token, for speech is ms",
+ )
diff --git a/SimulEval/simuleval/data/dataloader/s2t_dataloader.py b/SimulEval/simuleval/data/dataloader/s2t_dataloader.py
new file mode 100644
index 0000000000000000000000000000000000000000..e5c238ccc28677bdc78c36adb1fc89dba310db8a
--- /dev/null
+++ b/SimulEval/simuleval/data/dataloader/s2t_dataloader.py
@@ -0,0 +1,138 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from __future__ import annotations
+from pathlib import Path
+from typing import List, Union
+from .dataloader import GenericDataloader
+from simuleval.data.dataloader import register_dataloader
+from argparse import Namespace
+from urllib.parse import urlparse, parse_qs
+import yt_dlp as youtube_dl
+from pydub import AudioSegment
+
+try:
+ import soundfile
+
+ IS_IMPORT_SOUNDFILE = True
+except Exception:
+ IS_IMPORT_SOUNDFILE = False
+
+
+def download_youtube_video(url):
+ def get_video_id(url):
+ url_data = urlparse(url)
+ query = parse_qs(url_data.query)
+ video = query.get("v", [])
+ if video:
+ return video[0]
+ else:
+ raise Exception("unrecoginzed url format.")
+
+ id = get_video_id(url)
+ name = f"{id}.wav"
+
+ if not Path(name).exists():
+ ydl_opts = {
+ "format": "bestaudio/best",
+ "postprocessors": [
+ {
+ "key": "FFmpegExtractAudio",
+ "preferredcodec": "wav",
+ "preferredquality": "192",
+ }
+ ],
+ "outtmpl": id, # name the file "downloaded_video" with original extension
+ }
+ with youtube_dl.YoutubeDL(ydl_opts) as ydl:
+ ydl.download([url])
+
+ sound = AudioSegment.from_wav(name)
+ sound = sound.set_channels(1).set_frame_rate(16000)
+ sound.export(name, format="wav")
+ return name
+
+
+@register_dataloader("speech-to-text")
+class SpeechToTextDataloader(GenericDataloader):
+ def preprocess_source(self, source: Union[Path, str]) -> List[float]:
+ assert IS_IMPORT_SOUNDFILE, "Please make sure soundfile is properly installed."
+ samples, _ = soundfile.read(source, dtype="float32")
+ samples = samples.tolist()
+ return samples
+
+ def preprocess_target(self, target: str) -> str:
+ return target
+
+ def get_source_audio_info(self, index: int) -> soundfile._SoundFileInfo:
+ return soundfile.info(self.get_source_audio_path(index))
+
+ def get_source_audio_path(self, index: int):
+ return self.source_list[index]
+
+ @classmethod
+ def from_files(
+ cls, source: Union[Path, str], target: Union[Path, str]
+ ) -> SpeechToTextDataloader:
+ with open(source) as f:
+ source_list = [line.strip() for line in f]
+ with open(target) as f:
+ target_list = [line.strip() for line in f]
+ dataloader = cls(source_list, target_list)
+ return dataloader
+
+ @classmethod
+ def from_args(cls, args: Namespace):
+ args.source_type = "speech"
+ args.target_type = "text"
+ return cls.from_files(args.source, args.target)
+
+
+@register_dataloader("speech-to-speech")
+class SpeechToSpeechDataloader(SpeechToTextDataloader):
+ @classmethod
+ def from_files(
+ cls, source: Union[Path, str], target: Union[Path, str]
+ ) -> SpeechToSpeechDataloader:
+ with open(source) as f:
+ source_list = [line.strip() for line in f]
+ with open(target) as f:
+ target_list = [line.strip() for line in f]
+ dataloader = cls(source_list, target_list)
+ return dataloader
+
+ @classmethod
+ def from_args(cls, args: Namespace):
+ args.source_type = "speech"
+ args.target_type = "speech"
+ return cls.from_files(args.source, args.target)
+
+
+@register_dataloader("youtube-to-text")
+class YoutubeToTextDataloader(SpeechToTextDataloader):
+ @classmethod
+ def from_youtube(
+ cls, source: Union[Path, str], target: Union[Path, str]
+ ) -> YoutubeToTextDataloader:
+ source_list = [download_youtube_video(source)]
+ target_list = [target]
+ dataloader = cls(source_list, target_list)
+ return dataloader
+
+ @classmethod
+ def from_args(cls, args: Namespace):
+ args.source_type = "youtube"
+ args.target_type = "text"
+ return cls.from_youtube(args.source, args.target)
+
+
+@register_dataloader("youtube-to-speech")
+class YoutubeToSpeechDataloader(YoutubeToTextDataloader):
+ @classmethod
+ def from_args(cls, args: Namespace):
+ args.source_type = "youtube"
+ args.target_type = "speech"
+ return cls.from_youtube(args.source, args.target)
diff --git a/SimulEval/simuleval/data/dataloader/t2t_dataloader.py b/SimulEval/simuleval/data/dataloader/t2t_dataloader.py
new file mode 100644
index 0000000000000000000000000000000000000000..16d5d41408165eee5565b5566cfad8aac8c39fbb
--- /dev/null
+++ b/SimulEval/simuleval/data/dataloader/t2t_dataloader.py
@@ -0,0 +1,53 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from __future__ import annotations
+from pathlib import Path
+from typing import Callable, List, Union, Optional
+from .dataloader import GenericDataloader
+from simuleval.data.dataloader import register_dataloader
+from argparse import Namespace
+
+
+@register_dataloader("text-to-text")
+class TextToTextDataloader(GenericDataloader):
+ def __init__(
+ self, source_list: List[str], target_list: Union[List[str], List[None]]
+ ) -> None:
+ super().__init__(source_list, target_list)
+ self.source_splitter = lambda x: x.split()
+ self.target_splitter = lambda x: x
+
+ def set_source_splitter(self, function: Callable) -> None:
+ # TODO, make is configurable
+ self.splitter = function
+
+ def preprocess_source(self, source: str) -> List:
+ return self.source_splitter(source)
+
+ def preprocess_target(self, target: str) -> List:
+ return self.target_splitter(target)
+
+ @classmethod
+ def from_files(
+ cls, source: Union[Path, str], target: Optional[Union[Path, str]]
+ ) -> TextToTextDataloader:
+ assert source
+ with open(source) as f:
+ source_list = f.readlines()
+ if target:
+ with open(target) as f:
+ target_list = f.readlines()
+ else:
+ target_list = [None for _ in source_list]
+ dataloader = cls(source_list, target_list)
+ return dataloader
+
+ @classmethod
+ def from_args(cls, args: Namespace):
+ args.source_type = "text"
+ args.target_type = "text"
+ return cls.from_files(args.source, args.target)
diff --git a/SimulEval/simuleval/data/segments.py b/SimulEval/simuleval/data/segments.py
new file mode 100644
index 0000000000000000000000000000000000000000..c823695ecbb97bab4bf8fa07d66a5ce3803e9fcb
--- /dev/null
+++ b/SimulEval/simuleval/data/segments.py
@@ -0,0 +1,52 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import json
+from dataclasses import dataclass, field
+
+
+@dataclass
+class Segment:
+ index: int = 0
+ content: list = field(default_factory=list)
+ finished: bool = False
+ is_empty: bool = False
+ data_type: str = None
+
+ def json(self) -> str:
+ info_dict = {attribute: value for attribute, value in self.__dict__.items()}
+ return json.dumps(info_dict)
+
+ @classmethod
+ def from_json(cls, json_string: str):
+ return cls(**json.loads(json_string))
+
+
+@dataclass
+class EmptySegment(Segment):
+ is_empty: bool = True
+
+
+@dataclass
+class TextSegment(Segment):
+ content: str = ""
+ data_type: str = "text"
+
+
+@dataclass
+class SpeechSegment(Segment):
+ sample_rate: int = -1
+ data_type: str = "speech"
+
+
+def segment_from_json_string(string: str):
+ info_dict = json.loads(string)
+ if info_dict["data_type"] == "text":
+ return TextSegment.from_json(string)
+ elif info_dict["data_type"] == "speech":
+ return SpeechSegment.from_json(string)
+ else:
+ return EmptySegment.from_json(string)
diff --git a/SimulEval/simuleval/evaluator/__init__.py b/SimulEval/simuleval/evaluator/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..8dc1c292ab4ea3dc8486a000033965031278abee
--- /dev/null
+++ b/SimulEval/simuleval/evaluator/__init__.py
@@ -0,0 +1,16 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from .evaluator import SentenceLevelEvaluator
+from .remote import RemoteEvaluator
+
+
+def build_evaluator(args):
+ return SentenceLevelEvaluator.from_args(args)
+
+
+def build_remote_evaluator(args):
+ return RemoteEvaluator(build_evaluator(args))
diff --git a/SimulEval/simuleval/evaluator/evaluator.py b/SimulEval/simuleval/evaluator/evaluator.py
new file mode 100644
index 0000000000000000000000000000000000000000..a9fdb48c0c5e0a6fe8acd02dc03a88b374e645d2
--- /dev/null
+++ b/SimulEval/simuleval/evaluator/evaluator.py
@@ -0,0 +1,261 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import pandas
+import os
+import numbers
+from argparse import Namespace
+from typing import Dict, Generator, Optional
+from .scorers import get_scorer_class
+from .scorers.latency_scorer import LatencyScorer
+from .scorers.quality_scorer import QualityScorer
+
+from .instance import INSTANCE_TYPE_DICT, LogInstance
+import yaml
+import logging
+import json
+from tqdm import tqdm
+from pathlib import Path
+from simuleval.data.dataloader import GenericDataloader, build_dataloader
+
+
+logger = logging.getLogger("simuleval.sentence_level_evaluator")
+
+
+class SentenceLevelEvaluator(object):
+ """
+ Sentence Level evaluator. It iterates over sentence pairs and run evaluation.
+
+
+ .. code-block:: python
+
+ for instance in self.maybe_tqdm(self.instances.values()):
+ agent.reset()
+ while not instance.finish_prediction:
+ input_segment = instance.send_source(self.source_segment_size)
+ output_segment = agent.pushpop(input_segment)
+ instance.receive_prediction(output_segment)
+
+
+ Attributes:
+ instances: collections of sentence pairs. Instances also keep track of delays.
+ latency_scorers (List[~simuleval.scorers.latency_scorer.LatencyScorer]): Scorers for latency evaluation.
+ quality_scorers (List[~simuleval.scorers.latency_scorer.QualityScorer]): Scorers for quality evaluation.
+ output: output directory
+
+ Evaluator related command line arguments:
+
+ .. argparse::
+ :ref: simuleval.options.add_evaluator_args
+ :passparser:
+ :prog:
+ """
+
+ def __init__(
+ self,
+ dataloader: Optional[GenericDataloader],
+ quality_scorers: Dict[str, QualityScorer],
+ latency_scorers: Dict[str, LatencyScorer],
+ args: Namespace,
+ ) -> None:
+ self.dataloader = dataloader
+ self.quality_scorers = quality_scorers
+ self.latency_scorers = latency_scorers
+ self.instances = {}
+
+ self.args = args
+ self.output = Path(args.output) if args.output else None
+ self.score_only = args.score_only
+ self.source_segment_size = getattr(args, "source_segment_size", 1)
+ self.source_type = getattr(args, "source_type", None)
+ self.target_type = getattr(args, "target_type", None)
+
+ if (
+ self.source_type is None
+ and self.target_type is None
+ and self.output is not None
+ ):
+ with open(self.output / "config.yaml") as f:
+ configs = yaml.safe_load(f)
+ self.source_type = configs["source_type"]
+ self.target_type = configs["target_type"]
+
+ assert self.source_type
+ assert self.target_type
+
+ if self.output is not None:
+ os.makedirs(self.output, exist_ok=True)
+ with open(self.output / "config.yaml", "w") as f:
+ yaml.dump(
+ {"source_type": self.source_type, "target_type": self.source_type},
+ f,
+ default_flow_style=False,
+ )
+
+ self.instance_class = INSTANCE_TYPE_DICT[
+ f"{self.source_type}-{self.target_type}"
+ ]
+ self.start_index = getattr(args, "start_index", 0)
+ self.end_index = getattr(args, "end_index", -1)
+
+ if not self.score_only:
+ if self.output:
+ if (
+ self.args.continue_unfinished
+ and (self.output / "instances.log").exists()
+ ):
+ with open(self.output / "instances.log", "r") as f:
+ line = None
+ for line in f: # noqa
+ pass
+ if line is not None:
+ last_info = json.loads(line.strip())
+ self.start_index = last_info["index"] + 1
+ else:
+ self.output.mkdir(exist_ok=True, parents=True)
+ open(self.output / "instances.log", "w").close()
+ if self.end_index < 0:
+ assert self.dataloader is not None
+ self.end_index = len(self.dataloader)
+
+ self.build_instances()
+
+ if not self.args.no_progress_bar and not self.score_only:
+ self.instance_iterator = tqdm(
+ self.instances.values(),
+ initial=self.start_index,
+ total=len(self.instances.values()),
+ )
+ else:
+ self.instance_iterator = self.instances.values()
+
+ def write_log(self, instance):
+ if self.output is not None:
+ with open(self.output / "instances.log", "a") as f:
+ f.write(json.dumps(instance.summarize()) + "\n")
+
+ def build_instances(self):
+ if self.score_only:
+ self.build_instances_from_log()
+ else:
+ self.build_instances_from_dataloader()
+
+ def build_instances_from_log(self):
+ self.instances = {}
+ if self.output is not None:
+ with open(self.output / "instances.log", "r") as f:
+ for line in f:
+ instance = LogInstance(line.strip())
+ self.instances[instance.index] = instance
+
+ def build_instances_from_dataloader(self):
+ for i in self.get_indices():
+ self.instances[i] = self.instance_class(i, self.dataloader, self.args)
+
+ def __len__(self) -> int:
+ return self.end_index - self.start_index
+
+ def get_indices(self) -> Generator:
+ if self.end_index < 0:
+ self.end_index = max(self.instances.keys()) + 1
+
+ if self.start_index > self.end_index:
+ return []
+
+ for index in range(self.start_index, self.end_index):
+ yield index
+
+ @property
+ def quality(self) -> Dict[str, float]:
+ return {
+ name: scorer(self.instances)
+ for name, scorer in self.quality_scorers.items()
+ }
+
+ @property
+ def latency(self) -> Dict[str, float]:
+ return {
+ name: scorer(self.instances)
+ for name, scorer in self.latency_scorers.items()
+ }
+
+ @property
+ def results(self):
+ scores = {**self.quality, **self.latency}
+ new_scores = {}
+ for name, value in scores.items():
+ if isinstance(value, numbers.Number):
+ value = round(value, 3)
+ new_scores[name] = [value]
+
+ df = pandas.DataFrame(new_scores)
+ return df
+
+ def dump_results(self) -> None:
+ results = self.results
+ if self.output:
+ results.to_csv(self.output / "scores.tsv", sep="\t", index=False)
+
+ logger.info("Results:")
+ print(results.to_string(index=False))
+
+ def dump_metrics(self) -> None:
+ metrics = pandas.DataFrame([ins.metrics for ins in self.instances.values()])
+ metrics = metrics.round(3)
+ if self.output:
+ metrics.to_csv(self.output / "metrics.tsv", sep="\t", index=False)
+
+ def is_finished(self, instance) -> bool:
+ if hasattr(instance, "source_finished_reading"):
+ return instance.source_finished_reading
+ return instance.finish_prediction
+
+ def __call__(self, system):
+ system.reset()
+ for instance in self.instance_iterator:
+ while not self.is_finished(instance):
+ input_segment = instance.send_source(self.source_segment_size)
+ output_segment = system.pushpop(input_segment)
+ instance.receive_prediction(output_segment)
+ if instance.finish_prediction:
+ # if instance.finish_prediction where set by the reader,
+ # source_finished_reading will be set as well. If it is
+ # set by any of the intermediate components, then we didn't
+ # end yet. We are going to clear the state and continue
+ # processing the rest of the input.
+ system.reset()
+
+ if not self.score_only:
+ self.write_log(instance)
+
+ self.dump_results()
+ self.dump_metrics()
+
+ @classmethod
+ def from_args(cls, args):
+ if not args.score_only:
+ dataloader = build_dataloader(args)
+ else:
+ dataloader = None
+
+ latency_scorers = {}
+ use_ref_len = not args.no_use_ref_len
+ for name in args.latency_metrics:
+ # latency_scorers[name] = get_scorer_class("latency", name).from_args(args)
+ from copy import deepcopy
+ _args=deepcopy(args)
+ _args.computation_aware=False
+ latency_scorers[name] = get_scorer_class("latency", name).from_args(_args)
+ if args.computation_aware:
+ latency_scorers[name + "_CA"] = get_scorer_class("latency", name)(
+ computation_aware=True, use_ref_len=use_ref_len
+ )
+
+ quality_scorers = {}
+ for name in args.quality_metrics:
+ quality_scorers[name] = get_scorer_class("quality", name).from_args(args)
+
+ return cls(dataloader, quality_scorers, latency_scorers, args)
diff --git a/SimulEval/simuleval/evaluator/instance.py b/SimulEval/simuleval/evaluator/instance.py
new file mode 100644
index 0000000000000000000000000000000000000000..23fb7b6ca977be79dbbe41a52b0e42b8f384104f
--- /dev/null
+++ b/SimulEval/simuleval/evaluator/instance.py
@@ -0,0 +1,457 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import json
+import time
+import math
+from typing import Dict, List, Optional, Union
+from pathlib import Path
+
+from simuleval.data.segments import TextSegment, SpeechSegment, EmptySegment
+
+from simuleval.data.dataloader import SpeechToTextDataloader, TextToTextDataloader
+from argparse import Namespace
+
+try:
+ import soundfile
+
+ IS_IMPORT_SOUNDFILE = True
+except Exception:
+ IS_IMPORT_SOUNDFILE = False
+
+
+class Instance(object):
+ """
+ Instance class. An instance class contains one source and target sentence pair.
+ it send the source to and read hypotheses from the agent.
+
+ Args:
+ index (int): the index of the sentence pair in the corpus.
+ dataloader (GenericDataloader): the dataloader used to load the sentence pair.
+ args (Namespace): command line arguments.
+ """
+
+ def __init__(
+ self,
+ index: int,
+ dataloader: Optional[Union[SpeechToTextDataloader, TextToTextDataloader]],
+ args: Optional[Namespace],
+ ):
+ self.index = index
+ self.finish_prediction = False
+ self.dataloader = dataloader
+ if self.dataloader is not None:
+ self.source = self.dataloader[self.index]["source"]
+ self.reference = self.dataloader[self.index]["target"]
+ self.reset()
+ if args is not None:
+ self.args = args
+ self.latency_unit = args.eval_latency_unit
+
+ def reset(self):
+ self.step = 0
+ self.elapsed = []
+ self.prediction_list = []
+ self.delays = []
+ self.start_time = None
+ self.metrics = {}
+
+ def step_to_elapsed(self, *args):
+ raise NotImplementedError
+
+ def step_to_delay(self, step):
+ raise NotImplementedError
+
+ @property
+ def finish(self):
+ return self.finish_prediction
+
+ @finish.setter
+ def finish(self, status: bool):
+ self.finish_prediction = status
+
+ def preprocess_target(self, target: str) -> str:
+ """
+ Preprocess the target, for example tokenization.
+ """
+ return target
+
+ def preprocess_source(self, source: str):
+ """
+ Preprocess the source, for example tokenization.
+ """
+ raise NotImplementedError
+
+ def receive_prediction(self, prediction: str):
+ raise NotImplementedError
+
+ def send_source(self, *args):
+ raise NotImplementedError
+
+ @property
+ def source_length(self):
+ raise NotImplementedError
+
+ @property
+ def prediction_length(self):
+ return len(self.prediction_list)
+
+ @property
+ def target_length_latency(self):
+ raise NotImplementedError
+
+ @property
+ def prediction(self):
+ raise NotImplementedError
+
+ @property
+ def source_info(self):
+ return self.source
+
+ @property
+ def reference_length(self) -> int:
+ if self.latency_unit == "word":
+ return len(self.reference.split(" "))
+ elif self.latency_unit == "char":
+ return len(self.reference.strip())
+ else:
+ raise NotImplementedError
+
+ def summarize(self):
+ return {
+ "index": self.index,
+ "prediction": self.prediction,
+ "delays": self.delays,
+ "elapsed": self.elapsed,
+ "prediction_length": self.prediction_length,
+ "reference": self.reference,
+ "source": self.source_info,
+ "source_length": self.source_length,
+ "metric": self.metrics,
+ }
+
+ @classmethod
+ def from_json(cls, json_string):
+ info = json.loads(json_string)
+ instance = cls(info["index"], None, None)
+ instance.prediction_list = info["prediction"].split()
+ instance.delays = info["delays"]
+ instance.elapsed = info["elapsed"]
+ instance.reference = info["reference"]
+ instance.metrics = info["metric"]
+ instance.finish_prediction = True
+ return instance
+
+
+class TextInputInstance(Instance):
+ @property
+ def source_length(self):
+ return len(self.source)
+
+ @property
+ def source_info(self):
+ return " ".join(self.source)
+
+ def step_to_elapsed(self, *args):
+ return 0
+
+ def step_to_delay(self, step):
+ return step
+
+ def send_source(self, config_dict: Optional[Dict]):
+ if self.step >= self.source_length:
+ segment = EmptySegment(finished=True)
+ else:
+ segment = TextSegment(
+ index=self.step,
+ content=self.source[self.step],
+ finished=(self.step == self.source_length - 1),
+ )
+ self.step += 1
+
+ return segment
+
+
+class TextOutputInstance(Instance):
+ def receive_prediction(self, prediction: TextSegment):
+ """
+ Handler for receiving new predictions
+ """
+
+ if self.finish_prediction or prediction.is_empty:
+ self.finish_prediction = prediction.finished
+ return
+
+ if self.start_time is None:
+ self.start_time = time.time()
+
+ self.finish_prediction = prediction.finished
+
+ if len(prediction.content) == 0:
+ return
+
+ current_time = time.time()
+
+ if self.latency_unit == "word":
+ prediction_list = prediction.content.strip().split()
+ elif self.latency_unit == "char":
+ prediction_list = list(prediction.content.replace(" ", ""))
+ else:
+ raise NotImplementedError
+
+ self.prediction_list += prediction_list
+
+ self.elapsed += [self.step_to_elapsed(self.step, current_time)] * len(
+ prediction_list
+ )
+ self.delays += [self.step_to_delay(self.step)] * len(prediction_list)
+
+ @property
+ def target_length_latency(self):
+ if self.latency_unit == "word":
+ return len(self.reference.split(" "))
+ elif self.latency_unit == "char":
+ return len(self.reference)
+ else:
+ raise NotImplementedError
+
+ @property
+ def prediction(self) -> str:
+ if self.latency_unit == "word":
+ return "".join(list(self.prediction_list)).replace("▁", " ")
+ elif self.latency_unit == "char":
+ return "".join(list(self.prediction_list)).replace("▁", "")
+ else:
+ raise NotImplementedError
+
+
+class SpeechInputInstance(Instance):
+ def __init__(
+ self,
+ index: int,
+ dataloader: Optional[SpeechToTextDataloader],
+ args: Optional[Namespace],
+ ):
+ super().__init__(index, dataloader, args)
+ self.sample_rate_value = None
+ self.sample_list = None
+ self.source_finished_reading = False
+ self.dataloader: SpeechToTextDataloader
+
+ @property
+ def sample_rate(self):
+ if self.sample_rate_value is None:
+ self.audio_info = self.dataloader.get_source_audio_info(self.index)
+ self.sample_rate_value = self.audio_info.samplerate
+ return self.sample_rate_value
+
+ @property
+ def samples(self) -> List[float]:
+ if self.sample_list is None:
+ self.sample_list = self.source
+ return self.sample_list
+
+ @property
+ def is_finish_source(self):
+ return self.step == len(self.samples)
+
+ def send_source(self, segment_size=10):
+ if self.step == 0:
+ self.start_time = time.time()
+ assert segment_size >= 1, "instance size has to larger than 1 ms"
+
+ num_samples = math.ceil(segment_size / 1000 * self.sample_rate)
+
+ if self.step < len(self.samples):
+ if self.step + num_samples >= len(self.samples):
+ # Pad zeros if the requested number of samples
+ # are more than available samples.
+ samples = self.samples[self.step :] # noqa E203
+ is_finished = True
+ self.source_finished_reading = True
+ else:
+ samples = self.samples[self.step : self.step + num_samples] # noqa E203
+ is_finished = False
+
+ self.step = min(self.step + num_samples, len(self.samples))
+
+ segment = SpeechSegment(
+ index=self.len_sample_to_ms(self.step),
+ content=samples,
+ sample_rate=self.audio_info.samplerate,
+ finished=is_finished,
+ )
+
+ else:
+ # Finish reading this audio
+ segment = EmptySegment(
+ index=self.len_sample_to_ms(self.step),
+ finished=True,
+ )
+ self.source_finished_reading = True
+
+ return segment
+
+ @property
+ def source_length(self):
+ # In milliseconds
+ return self.len_sample_to_ms(len(self.samples))
+
+ @property
+ def source_info(self):
+ return str(self.audio_info).split("\n")
+
+ def len_sample_to_ms(self, length):
+ assert getattr(self, "sample_rate", None), "Read a audio file first"
+ return length * 1000 / self.sample_rate
+
+ def len_ms_to_samples(self, length):
+ assert getattr(self, "sample_rate", None), "Read a audio file first"
+ return math.ceil(length / 1000 * self.sample_rate)
+
+ def step_to_delay(self, step):
+ return self.len_sample_to_ms(self.step)
+
+ def step_to_elapsed(self, step, current_time):
+ return self.len_sample_to_ms(step) + (current_time - self.start_time) * 1000
+
+
+class SpeechOutputInstance(Instance):
+ def __init__(self, index, dataloader, args):
+ super().__init__(index, dataloader, args)
+ self.prediction_time = 0
+ self.durations = []
+ self.intervals = []
+ self.target_sample_rate = -1
+ self.dataloader: SpeechToTextDataloader # For now we only support speech input.
+ assert IS_IMPORT_SOUNDFILE, "Please make sure soundfile is properly installed."
+ assert self.args.output is not None, "'output' is needed for speech output"
+
+ @property
+ def wav_path(self):
+ wav_dir_path = Path(self.args.output) / "wavs"
+ wav_dir_path.mkdir(exist_ok=True)
+ wav_path = wav_dir_path / f"{self.index}_pred.wav"
+ return wav_path.absolute()
+
+ @property
+ def prediction(self):
+ return self.wav_path
+
+ def summarize(self):
+ samples = []
+ self.intervals = []
+ self.silences = []
+
+ if len(self.prediction_list) > 0:
+ # start from the first segment offset
+ start = prev_end = prediction_offset = self.delays[0]
+
+ for i, delay in enumerate(self.delays):
+ start = max(prev_end, delay)
+
+ if start > prev_end:
+ # Wait source speech, add discontinuity with silence
+ samples += [0.0] * int(
+ self.target_sample_rate * (start - prev_end) / 1000
+ )
+ self.silences.append(start - prev_end)
+
+ samples += self.prediction_list[i]
+ duration = self.durations[i]
+ prev_end = start + duration
+ self.intervals.append([start, duration])
+ soundfile.write(self.wav_path, samples, self.target_sample_rate)
+ else:
+ # For empty prediction
+ prediction_offset = self.source_length
+
+ return {
+ "index": self.index,
+ "prediction": self.wav_path.as_posix(),
+ "delays": self.delays,
+ "durations": self.durations,
+ "prediction_offset": prediction_offset,
+ "elapsed": [],
+ "intervals": self.intervals,
+ "prediction_length": len(samples) / self.target_sample_rate,
+ "source_length": self.source_length,
+ "reference": self.reference,
+ "source": self.dataloader.get_source_audio_path(self.index),
+ }
+
+ def receive_prediction(self, segment: SpeechSegment):
+ """
+ Handler for receiving new predictions
+ """
+ if self.start_time is None:
+ self.start_time = time.time()
+
+ if self.finish_prediction and (
+ not hasattr(self, "source_finished_reading") or self.source_finished_reading
+ ):
+ return
+
+ self.finish_prediction = segment.finished
+
+ if segment.is_empty:
+ return
+
+ if len(segment.content) == 0:
+ return
+
+ current_time = time.time()
+
+ pred_duration = 1000 * len(segment.content) / segment.sample_rate
+
+ if self.target_sample_rate < 0:
+ self.target_sample_rate = segment.sample_rate
+
+ self.durations.append(pred_duration)
+ self.prediction_list.append(segment.content)
+ self.elapsed.append(self.step_to_elapsed(self.step, current_time))
+ self.delays.append(self.step_to_delay(self.step))
+
+ if self.finish_prediction:
+ self.summarize()
+
+
+class SpeechToTextInstance(SpeechInputInstance, TextOutputInstance):
+ pass
+
+
+class TextToTextInstance(TextInputInstance, TextOutputInstance):
+ pass
+
+
+class SpeechToSpeechInstance(SpeechInputInstance, SpeechOutputInstance):
+ pass
+
+
+INSTANCE_TYPE_DICT = {
+ "speech-text": SpeechToTextInstance,
+ "text-text": TextToTextInstance,
+ "speech-speech": SpeechToSpeechInstance,
+ "youtube-text": SpeechToTextInstance,
+ "youtube-speech": SpeechToSpeechInstance,
+}
+
+
+class LogInstance:
+ def __init__(self, info: str) -> None:
+ self.info = json.loads(info.strip())
+ self.intervals = []
+ for key, value in self.info.items():
+ setattr(self, key, value)
+
+ self.index = self.info["index"]
+ self.reference = self.info.get("reference", "")
+ self.reference_length = len(
+ self.reference.split(" ")
+ ) # ToDo: temporary solution, make it configurable
+ self.source_length = self.info.get("source_length") # just for testing!
+ self.finish_prediction = True
+ self.metrics = {}
diff --git a/SimulEval/simuleval/evaluator/remote.py b/SimulEval/simuleval/evaluator/remote.py
new file mode 100644
index 0000000000000000000000000000000000000000..d73c83262c61e441e64955b269845e4e3d5905a3
--- /dev/null
+++ b/SimulEval/simuleval/evaluator/remote.py
@@ -0,0 +1,47 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import logging
+from simuleval.data.segments import Segment, segment_from_json_string
+from simuleval.evaluator import SentenceLevelEvaluator
+import requests
+
+logger = logging.getLogger("simuleval.remote_evaluator")
+
+
+class RemoteEvaluator:
+ def __init__(self, evaluator: SentenceLevelEvaluator) -> None:
+ self.evaluator = evaluator
+ self.address = evaluator.args.remote_address
+ self.port = evaluator.args.remote_port
+ self.source_segment_size = evaluator.args.source_segment_size
+ self.base_url = f"http://{self.address}:{self.port}"
+
+ def send_source(self, segment: Segment):
+ url = f"{self.base_url}/input"
+ requests.put(url, data=segment.json())
+
+ def receive_prediction(self) -> Segment:
+ url = f"{self.base_url}/output"
+ r = requests.get(url)
+ return segment_from_json_string(r.text)
+
+ def system_reset(self):
+ requests.post(f"{self.base_url}/reset")
+
+ def results(self):
+ return self.evaluator.results()
+
+ def remote_eval(self):
+ for instance in self.evaluator.instance_iterator:
+ self.system_reset()
+ while not instance.finish_prediction:
+ self.send_source(instance.send_source(self.source_segment_size))
+ output_segment = self.receive_prediction()
+ instance.receive_prediction(output_segment)
+ self.evaluator.write_log(instance)
+
+ self.evaluator.dump_results()
diff --git a/SimulEval/simuleval/evaluator/scorers/__init__.py b/SimulEval/simuleval/evaluator/scorers/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..56abd10dd938b378387e2751b26284a1a2b4dc69
--- /dev/null
+++ b/SimulEval/simuleval/evaluator/scorers/__init__.py
@@ -0,0 +1,14 @@
+from .latency_scorer import LATENCY_SCORERS_DICT
+from .quality_scorer import QUALITY_SCORERS_DICT
+
+
+def get_scorer_class(scorer_type, name):
+ if scorer_type == "quality":
+ scorer_dict = QUALITY_SCORERS_DICT
+ else:
+ scorer_dict = LATENCY_SCORERS_DICT
+
+ if name not in scorer_dict:
+ raise RuntimeError(f"No {scorer_type} metric called {name}")
+
+ return scorer_dict[name]
diff --git a/SimulEval/simuleval/evaluator/scorers/latency_scorer.py b/SimulEval/simuleval/evaluator/scorers/latency_scorer.py
new file mode 100644
index 0000000000000000000000000000000000000000..8cebf01e2abd0b869e2ecaf868c926c72126c3f3
--- /dev/null
+++ b/SimulEval/simuleval/evaluator/scorers/latency_scorer.py
@@ -0,0 +1,686 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from statistics import mean
+from pathlib import Path
+import subprocess
+import logging
+import textgrid
+import sys
+import shutil
+from typing import List, Union, Dict
+from simuleval.evaluator.instance import (
+ TextInputInstance,
+ TextOutputInstance,
+ SpeechOutputInstance,
+ Instance,
+ LogInstance,
+ SpeechOutputInstance,
+)
+from argparse import ArgumentParser, Namespace
+from subprocess import Popen, PIPE
+
+logger = logging.getLogger("simuleval.latency_scorer")
+
+LATENCY_SCORERS_DICT = {}
+LATENCY_SCORERS_NAME_DICT = {}
+
+
+def register_latency_scorer(name):
+ def register(cls):
+ LATENCY_SCORERS_DICT[name] = cls
+ LATENCY_SCORERS_NAME_DICT[cls.__name__] = name
+ return cls
+
+ return register
+
+
+class LatencyScorer:
+ metric = None
+ add_duration = False
+
+ def __init__(
+ self, computation_aware: bool = False, use_ref_len: bool = True
+ ) -> None:
+ super().__init__()
+ self.use_ref_len = use_ref_len
+ self.computation_aware = computation_aware
+
+ @property
+ def timestamp_type(self):
+ return "delays" if not self.computation_aware else "elapsed"
+
+ def compute(self, *args):
+ raise NotImplementedError
+
+ def get_delays_lengths(self, ins: Instance):
+ """
+ Args:
+ ins Instance: one instance
+
+ Returns:
+ A tuple with the 3 elements:
+ delays (List[Union[float, int]]): Sequence of delays.
+ src_len (Union[float, int]): Length of source sequence.
+ tgt_len (Union[float, int]): Length of target sequence.
+ """
+ delays = getattr(ins, self.timestamp_type, None)
+ assert delays
+
+ if not self.use_ref_len or ins.reference is None:
+ tgt_len = len(delays)
+ else:
+ tgt_len = ins.reference_length
+ src_len = ins.source_length
+ return delays, src_len, tgt_len
+
+ @property
+ def metric_name(self) -> str:
+ return LATENCY_SCORERS_NAME_DICT[self.__class__.__name__]
+
+ def __call__(self, instances: Dict[int, Instance]) -> float:
+ scores = []
+ for index, ins in instances.items():
+ if isinstance(ins, TextInputInstance):
+ if self.computation_aware:
+ raise RuntimeError(
+ "The computation aware latency is not supported on text input."
+ )
+ delays = getattr(ins, self.timestamp_type, None)
+ if delays is None or len(delays) == 0:
+ logger.warn(f"Instance {index} has no delay information. Skipped")
+ continue
+ score = self.compute(ins)
+ ins.metrics[self.metric_name] = score
+ scores.append(score)
+
+ return mean(scores)
+
+ @staticmethod
+ def add_args(parser: ArgumentParser):
+ pass
+
+ @classmethod
+ def from_args(cls, args: Namespace):
+ return cls(
+ computation_aware=args.computation_aware,
+ use_ref_len=not args.no_use_ref_len,
+ )
+
+
+@register_latency_scorer("AL")
+class ALScorer(LatencyScorer):
+ r"""
+ Average Lagging (AL) from
+ `STACL: Simultaneous Translation with Implicit Anticipation and Controllable Latency using Prefix-to-Prefix Framework `_
+
+ Give source :math:`X`, target :math:`Y`, delays :math:`D`,
+
+ .. math::
+
+ AL = \frac{1}{\tau} \sum_i^\tau D_i - (i - 1) \frac{|X|}{|Y|}
+
+ Where
+
+ .. math::
+
+ \tau = argmin_i(D_i = |X|)
+
+ When reference was given, :math:`|Y|` would be the reference length
+
+ Usage:
+ ----latency-metrics AL
+ """ # noqa: E501
+
+ def compute(self, ins: Instance):
+ """
+ Function to compute latency on one sentence (instance).
+
+ Args:
+ ins Instance: one instance
+
+ Returns:
+ float: the latency score on one sentence.
+ """
+ delays, source_length, target_length = self.get_delays_lengths(ins)
+
+ if delays[0] > source_length:
+ return delays[0]
+
+ AL = 0
+ gamma = target_length / source_length
+ tau = 0
+ for t_minus_1, d in enumerate(delays):
+ AL += d - t_minus_1 / gamma
+ tau = t_minus_1 + 1
+
+ if d >= source_length:
+ break
+ AL /= tau
+ return AL
+
+
+@register_latency_scorer("LAAL")
+class LAALScorer(ALScorer):
+ r"""
+ Length Adaptive Average Lagging (LAAL) as proposed in
+ `CUNI-KIT System for Simultaneous Speech Translation Task at IWSLT 2022
+ `_.
+ The name was suggested in `Over-Generation Cannot Be Rewarded:
+ Length-Adaptive Average Lagging for Simultaneous Speech Translation
+ `_.
+ It is the original Average Lagging as proposed in
+ `Controllable Latency using Prefix-to-Prefix Framework
+ `_
+ but is robust to the length difference between the hypothesis and reference.
+
+ Give source :math:`X`, target :math:`Y`, delays :math:`D`,
+
+ .. math::
+
+ LAAL = \frac{1}{\tau} \sum_i^\tau D_i - (i - 1) \frac{|X|}{max(|Y|,|Y*|)}
+
+ Where
+
+ .. math::
+
+ \tau = argmin_i(D_i = |X|)
+
+ When reference was given, :math:`|Y|` would be the reference length, and :math:`|Y*|` is the length of the hypothesis.
+
+ Usage:
+ ----latency-metrics LAAL
+ """
+
+ def compute(self, ins: Instance):
+ """
+ Function to compute latency on one sentence (instance).
+
+ Args:
+ ins: Instance: one instance
+
+ Returns:
+ float: the latency score on one sentence.
+ """
+ delays, source_length, target_length = self.get_delays_lengths(ins)
+ if delays[0] > source_length:
+ return delays[0]
+
+ LAAL = 0
+ gamma = max(len(delays), target_length) / source_length
+ tau = 0
+ for t_minus_1, d in enumerate(delays):
+ LAAL += d - t_minus_1 / gamma
+ tau = t_minus_1 + 1
+
+ if d >= source_length:
+ break
+ LAAL /= tau
+ return LAAL
+
+
+@register_latency_scorer("AP")
+class APScorer(LatencyScorer):
+ r"""
+ Average Proportion (AP) from
+ `Can neural machine translation do simultaneous translation? `_
+
+ Give source :math:`X`, target :math:`Y`, delays :math:`D`,
+ the AP is calculated as:
+
+ .. math::
+
+ AP = \frac{1}{|X||Y]} \sum_i^{|Y|} D_i
+
+ Usage:
+ ----latency-metrics AP
+ """
+
+ def compute(self, ins: Instance) -> float:
+ """
+ Function to compute latency on one sentence (instance).
+
+ Args:
+ ins Instance: one instance
+
+ Returns:
+ float: the latency score on one sentence.
+ """
+ delays, source_length, target_length = self.get_delays_lengths(ins)
+ return sum(delays) / (source_length * target_length)
+
+
+@register_latency_scorer("DAL")
+class DALScorer(LatencyScorer):
+ r"""
+ Differentiable Average Lagging (DAL) from
+ Monotonic Infinite Lookback Attention for Simultaneous Machine Translation
+ (https://arxiv.org/abs/1906.05218)
+
+ Usage:
+ ----latency-metrics DAL
+ """
+
+ def compute(self, ins: Instance):
+ """
+ Function to compute latency on one sentence (instance).
+
+ Args:
+ ins Instance: one instance
+
+ Returns:
+ float: the latency score on one sentence.
+ """
+ delays, source_length, target_length = self.get_delays_lengths(ins)
+
+ DAL = 0
+ target_length = len(delays)
+ gamma = target_length / source_length
+ g_prime_last = 0
+ for i_minus_1, g in enumerate(delays):
+ if i_minus_1 + 1 == 1:
+ g_prime = g
+ else:
+ g_prime = max([g, g_prime_last + 1 / gamma])
+
+ DAL += g_prime - i_minus_1 / gamma
+ g_prime_last = g_prime
+
+ DAL /= target_length
+ return DAL
+
+
+@register_latency_scorer("ATD")
+class ATDScorer(LatencyScorer):
+ r"""
+ Average Token Delay (ATD) from
+ Average Token Delay: A Latency Metric for Simultaneous Translation
+ (https://arxiv.org/abs/2211.13173)
+
+ Different from speech segments, text tokens have no length
+ and multiple tokens can be output at the same time like subtitle.
+ Therefore, we set its length to be 0. However, to calculate latency in text-text,
+ we give virtual time 1 for the length of text tokens.
+
+ Usage:
+ ----latency-metrics ATD
+ """
+
+ def __call__(self, instances) -> float: # noqa C901
+ if isinstance(instances[0], TextInputInstance):
+ TGT_TOKEN_LEN = 1
+ SRC_TOKEN_LEN = 1
+ INPUT_TYPE = "text"
+ OUTPUT_TYPE = "text"
+ else:
+ SRC_TOKEN_LEN = 300 # 300ms per word
+ INPUT_TYPE = "speech"
+ if isinstance(instances[0], TextOutputInstance) or isinstance(
+ instances[0], LogInstance
+ ):
+ TGT_TOKEN_LEN = 0
+ OUTPUT_TYPE = "text"
+ else:
+ TGT_TOKEN_LEN = 300
+ OUTPUT_TYPE = "speech"
+
+ scores = []
+ for index, ins in instances.items():
+ delays = getattr(ins, "delays", None)
+ if delays is None or len(delays) == 0:
+ logger.warn(f"Instance {index} has no delay information. Skipped")
+ continue
+
+ if self.computation_aware:
+ elapsed = getattr(ins, "elapsed", None)
+ if elapsed is None or len(elapsed) == 0:
+ logger.warn(
+ f"Instance {index} has no computational delay information. Skipped"
+ )
+ continue
+ if elapsed != [0] * len(delays):
+ compute_elapsed = self.subtract(elapsed, delays)
+ compute_times = self.subtract(
+ compute_elapsed, [0] + compute_elapsed[:-1]
+ )
+ else:
+ compute_times = elapsed
+ else:
+ compute_times = [0] * len(delays)
+
+ chunk_sizes = {"src": [0], "tgt": [0]}
+ token_to_chunk = {"src": [0], "tgt": [0]}
+ token_to_time = {"src": [0], "tgt": [0]}
+
+ tgt_token_lens = []
+ delays_no_duplicate = sorted(set(delays), key=delays.index)
+
+ if OUTPUT_TYPE == "text":
+ prev_delay = None
+ for delay in delays:
+ if delay != prev_delay:
+ chunk_sizes["tgt"].append(1)
+ else:
+ chunk_sizes["tgt"][-1] += 1
+ prev_delay = delay
+ for i, chunk_size in enumerate(chunk_sizes["tgt"][1:], 1):
+ token_to_chunk["tgt"] += [i] * chunk_size
+ tgt_token_lens = [TGT_TOKEN_LEN] * len(delays)
+ else:
+ s2s_delays = []
+ s2s_compute_times = []
+ chunk_durations = []
+ chunk_compute_times = []
+ prev_delay = None
+ for delay, compute_time, duration in zip(
+ delays, compute_times, ins.durations
+ ):
+ if delay != prev_delay:
+ chunk_durations.append(duration)
+ chunk_compute_times.append(compute_time)
+ else:
+ chunk_durations[-1] += duration
+ chunk_compute_times[-1] += compute_time
+ prev_delay = delay
+ for i, chunk_duration in enumerate(chunk_durations, 1):
+ num_tokens, rest = divmod(chunk_duration, TGT_TOKEN_LEN)
+ token_lens = int(num_tokens) * [TGT_TOKEN_LEN] + (
+ [rest] if rest != 0 else []
+ )
+ tgt_token_lens += token_lens
+ chunk_sizes["tgt"] += [len(token_lens)]
+ token_to_chunk["tgt"] += [i] * len(token_lens)
+ s2s_delays += [delays_no_duplicate[i - 1]] * len(token_lens)
+ s2s_compute_times += [
+ chunk_compute_times[i - 1] / len(token_lens)
+ ] * len(token_lens)
+ delays = s2s_delays
+ compute_times = s2s_compute_times
+
+ if INPUT_TYPE == "text":
+ chunk_sizes["src"] += self.subtract(
+ delays_no_duplicate, [0] + delays_no_duplicate[:-1]
+ )
+ for i, chunk_size in enumerate(chunk_sizes["src"][1:], 1):
+ token_lens = chunk_size * [SRC_TOKEN_LEN]
+ for token_len in token_lens:
+ token_to_time["src"].append(
+ token_to_time["src"][-1] + token_len
+ )
+ token_to_chunk["src"].append(i)
+ else:
+ chunk_durations = self.subtract(
+ delays_no_duplicate, [0] + delays_no_duplicate[:-1]
+ )
+ for i, chunk_duration in enumerate(chunk_durations, 1):
+ num_tokens, rest = divmod(chunk_duration, SRC_TOKEN_LEN)
+ token_lens = int(num_tokens) * [SRC_TOKEN_LEN] + (
+ [rest] if rest != 0 else []
+ )
+ chunk_sizes["src"] += [len(token_lens)]
+ for token_len in token_lens:
+ token_to_time["src"].append(
+ token_to_time["src"][-1] + token_len
+ )
+ token_to_chunk["src"].append(i)
+
+ for delay, compute_time, token_len in zip(
+ delays, compute_times, tgt_token_lens
+ ):
+ tgt_start_time = max(delay, token_to_time["tgt"][-1])
+ token_to_time["tgt"].append(tgt_start_time + token_len + compute_time)
+
+ scores.append(self.compute(chunk_sizes, token_to_chunk, token_to_time))
+
+ return mean(scores)
+
+ def subtract(self, arr1, arr2):
+ return [x - y for x, y in zip(arr1, arr2)]
+
+ def compute(
+ self,
+ chunk_sizes: Dict[str, List[Union[float, int]]],
+ token_to_chunk: Dict[str, List[Union[float, int]]],
+ token_to_time: Dict[str, List[Union[float, int]]],
+ ) -> float:
+ """
+ Function to compute latency on one sentence (instance).
+ Args:
+ chunk_sizes Dict[str, List[Union[float, int]]]: Sequence of chunk sizes for source and target.
+ token_to_chunk Dict[str, List[Union[float, int]]]: Sequence of chunk indices to which the tokens belong for source and target.
+ token_to_time Dict[str, List[Union[float, int]]]: Sequence of ending times of tokens for source and target.
+
+ Returns:
+ float: the latency score on one sentence.
+ """ # noqa C501
+
+ tgt_to_src = []
+
+ for t in range(1, len(token_to_chunk["tgt"])):
+ chunk_id = token_to_chunk["tgt"][t]
+ AccSize_x = sum(chunk_sizes["src"][:chunk_id])
+ AccSize_y = sum(chunk_sizes["tgt"][:chunk_id])
+
+ S = t - max(0, AccSize_y - AccSize_x)
+ current_src_size = sum(chunk_sizes["src"][: chunk_id + 1])
+
+ if S < current_src_size:
+ tgt_to_src.append((t, S))
+ else:
+ tgt_to_src.append((t, current_src_size))
+
+ atd_delays = []
+
+ for t, s in tgt_to_src:
+ atd_delay = token_to_time["tgt"][t] - token_to_time["src"][s]
+ atd_delays.append(atd_delay)
+
+ return float(mean(atd_delays))
+
+
+@register_latency_scorer("NumChunks")
+class NumChunksScorer(LatencyScorer):
+ """Number of chunks (of speech/text) in output
+
+ Usage:
+ ----latency-metrics NumChunks
+
+ """
+
+ def compute(self, ins: Instance):
+ delays, _, _ = self.get_delays_lengths(ins)
+ return len(delays)
+
+
+@register_latency_scorer("DiscontinuitySum")
+class DiscontinuitySumScorer(LatencyScorer):
+ """Sum of discontinuity in speech output
+
+ Usage:
+ ----latency-metrics DiscontinuitySum
+
+ """
+
+ def compute(self, ins: Instance):
+ assert isinstance(ins, SpeechOutputInstance)
+ return sum(ins.silences)
+
+
+@register_latency_scorer("DiscontinuityAve")
+class DiscontinuityAveScorer(LatencyScorer):
+ """Average of discontinuities in speech output
+
+ Usage:
+ ----latency-metrics DiscontinuityAve
+
+ """
+
+ def compute(self, ins: Instance):
+ assert isinstance(ins, SpeechOutputInstance)
+ if len(ins.silences) == 0:
+ return 0
+ return sum(ins.silences) / len(ins.silences)
+
+
+@register_latency_scorer("DiscontinuityNum")
+class DiscontinuityNumScorer(LatencyScorer):
+ """Number of discontinuities in speech output
+
+ Usage:
+ ----latency-metrics DiscontinuityNum
+
+ """
+
+ def compute(self, ins: Instance):
+ assert isinstance(ins, SpeechOutputInstance)
+ return len(ins.silences)
+
+
+@register_latency_scorer("StartOffset")
+class StartOffsetScorer(LatencyScorer):
+ """Starting offset of the translation
+
+ Usage:
+ ----latency-metrics StartOffset
+
+ """
+
+ def compute(self, ins: Instance):
+ delays, _, _ = self.get_delays_lengths(ins)
+ return delays[0]
+
+
+@register_latency_scorer("EndOffset")
+class EndOffsetScorer(LatencyScorer):
+ """Ending offset of the translation
+
+ Usage:
+ ----latency-metrics EndOffset
+
+ """
+
+ def compute(self, ins: Instance):
+ delays, source_length, _ = self.get_delays_lengths(ins)
+ if isinstance(ins, SpeechOutputInstance) or (
+ isinstance(ins, LogInstance) and len(ins.intervals) > 0
+ ):
+ delays = [start + duration for start, duration in ins.intervals]
+ return delays[-1] - source_length
+
+
+@register_latency_scorer("RTF")
+class RTFScorer(LatencyScorer):
+ """Compute Real Time Factor (RTF)
+
+ Usage:
+ ----latency-metrics (RTF)
+
+ """
+
+ def compute(self, ins: Instance):
+ delays, source_length, _ = self.get_delays_lengths(ins)
+ if isinstance(ins, SpeechOutputInstance):
+ delays = [start + duration for start, duration in ins.intervals]
+ return delays[-1] / source_length
+
+
+def speechoutput_alignment_latency_scorer(scorer_class): # noqa C901
+ class Klass(scorer_class):
+ def __init__(self, **kargs) -> None:
+ assert getattr(self, "boundary_type", None) in [
+ "BOW",
+ "EOW",
+ "COW",
+ ], self.boundary_type
+ super().__init__(**kargs)
+ if self.computation_aware:
+ raise RuntimeError(
+ "The computation aware latency for speech output is not supported yet"
+ )
+
+ @property
+ def timestamp_type(self):
+ return "aligned_delays"
+
+ def __call__(self, instances) -> float:
+ self.prepare_alignment(instances)
+ return super().__call__(instances)
+
+ def prepare_alignment(self, instances):
+ try:
+ subprocess.check_output(
+ "mfa version", shell=True, stderr=subprocess.STDOUT
+ )
+ except subprocess.CalledProcessError as grepexc:
+ logger.error(grepexc.output.decode("utf-8").strip())
+ logger.error("Please make sure the mfa>=2.0.6 is correctly installed. ")
+ sys.exit(1)
+
+ output_dir = Path(instances[0].prediction).absolute().parent.parent
+ align_dir = output_dir / "align"
+ if not align_dir.exists():
+ logger.info("Align target transcripts with speech.")
+ temp_dir = Path(output_dir) / "mfa"
+ shutil.rmtree(temp_dir, ignore_errors=True)
+ temp_dir.mkdir(exist_ok=True)
+ original_model_path = Path.home() / "Documents/MFA/pretrained_models"
+ acoustic_model_path = temp_dir / "acoustic.zip"
+ acoustic_model_path.symlink_to(
+ original_model_path / "acoustic" / "english_mfa.zip"
+ )
+ dictionary_path = temp_dir / "dict"
+ dictionary_path.symlink_to(
+ original_model_path / "dictionary" / "english_mfa.dict"
+ )
+ mfa_command = (
+ f"mfa align {output_dir / 'wavs'} {dictionary_path.as_posix()} {acoustic_model_path.as_posix()}"
+ + f" {align_dir.as_posix()} --clean --overwrite --temporary_directory {temp_dir.as_posix()}"
+ )
+ logger.info(mfa_command)
+
+ subprocess.run(
+ mfa_command,
+ shell=True,
+ check=True,
+ )
+ else:
+ logger.info("Found existing alignment")
+
+ for file in align_dir.iterdir():
+ if file.name.endswith("TextGrid"):
+ index = int(file.name.split("_")[0])
+ target_offset = instances[index].delays[0]
+ info = textgrid.TextGrid.fromFile(file)
+ delays = []
+ for interval in info[0]:
+ if len(interval.mark) > 0:
+ if self.boundary_type == "BOW":
+ delays.append(target_offset + 1000 * interval.minTime)
+ elif self.boundary_type == "EOW":
+ delays.append(target_offset + 1000 * interval.maxTime)
+ else:
+ delays.append(
+ target_offset
+ + 0.5 * (interval.maxTime + interval.minTime) * 1000
+ )
+ setattr(instances[index], self.timestamp_type, delays)
+
+ return Klass
+
+
+for boundary_type in ["BOW", "COW", "EOW"]:
+ for metric in ["AL", "LAAL", "AP", "DAL", "ATD", "StartOffset", "EndOffset"]:
+
+ @register_latency_scorer(f"{metric}_SpeechAlign_{boundary_type}")
+ @speechoutput_alignment_latency_scorer
+ class SpeechAlignScorer(LATENCY_SCORERS_DICT[metric]): # type: ignore
+ f"""Compute {metric} based on alignment ({boundary_type})
+
+ Usage:
+ ----latency-metrics {metric}_SpeechAlign_{boundary_type}
+ """
+ boundary_type = boundary_type
+ __name__ = f"{metric}SpeechAlign{boundary_type}Scorer"
diff --git a/SimulEval/simuleval/evaluator/scorers/quality_scorer.py b/SimulEval/simuleval/evaluator/scorers/quality_scorer.py
new file mode 100644
index 0000000000000000000000000000000000000000..52c6a3cdec9a9db142ceb67814fa3d2553171776
--- /dev/null
+++ b/SimulEval/simuleval/evaluator/scorers/quality_scorer.py
@@ -0,0 +1,331 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import re
+import logging
+import sacrebleu
+from pathlib import Path
+from typing import Dict
+from sacrebleu.metrics.bleu import BLEU
+import subprocess
+import string
+import tqdm
+
+QUALITY_SCORERS_DICT = {}
+
+
+def register_quality_scorer(name):
+ def register(cls):
+ QUALITY_SCORERS_DICT[name] = cls
+ return cls
+
+ return register
+
+
+class QualityScorer:
+ def __init__(self) -> None:
+ pass
+
+ def __call__(self, instances: Dict) -> float:
+ raise NotImplementedError
+
+ @staticmethod
+ def add_args(parser):
+ pass
+
+
+def add_sacrebleu_args(parser):
+ parser.add_argument(
+ "--sacrebleu-tokenizer",
+ type=str,
+ default=sacrebleu.metrics.METRICS["BLEU"].TOKENIZER_DEFAULT,
+ choices=sacrebleu.metrics.METRICS["BLEU"].TOKENIZERS,
+ help="Tokenizer in sacrebleu",
+ )
+
+
+@register_quality_scorer("BLEU")
+class SacreBLEUScorer(QualityScorer):
+ """
+ SacreBLEU Scorer
+
+ Usage:
+ :code:`--quality-metrics BLEU`
+
+ Additional command line arguments:
+
+ .. argparse::
+ :ref: simuleval.evaluator.scorers.quality_scorer.add_sacrebleu_args
+ :passparser:
+ :prog:
+ """
+
+ def __init__(self, tokenizer: str = "13a") -> None:
+ super().__init__()
+ self.logger = logging.getLogger("simuleval.scorer.bleu")
+ self.tokenizer = tokenizer
+
+ def __call__(self, instances: Dict) -> float:
+ try:
+ return (
+ BLEU(tokenize=self.tokenizer)
+ .corpus_score(
+ [ins.prediction for ins in instances.values()],
+ [[ins.reference for ins in instances.values()]],
+ )
+ .score
+ )
+ except Exception as e:
+ self.logger.error(str(e))
+ return 0
+
+ @staticmethod
+ def add_args(parser):
+ add_sacrebleu_args(parser)
+
+ @classmethod
+ def from_args(cls, args):
+ return cls(args.sacrebleu_tokenizer)
+
+
+@register_quality_scorer("ASR_BLEU")
+class ASRSacreBLEUScorer(QualityScorer):
+ """
+ ASR + SacreBLEU Scorer (BETA version)
+
+ Usage:
+ :code:`--quality-metrics ASR_BLEU`
+
+ Additional command line arguments:
+
+ .. argparse::
+ :ref: simuleval.evaluator.scorers.quality_scorer.add_sacrebleu_args
+ :passparser:
+ :prog:
+ """
+
+ def __init__(self, tokenizer: str = "13a", target_lang: str = "en") -> None:
+ super().__init__()
+ self.logger = logging.getLogger("simuleval.scorer.asr_bleu")
+ self.tokenizer = tokenizer
+ self.target_lang = target_lang
+
+ def __call__(self, instances: Dict) -> float:
+ transcripts = self.asr_transcribe(instances)
+ score = (
+ BLEU(tokenize=self.tokenizer)
+ .corpus_score(
+ transcripts,
+ [[ins.reference for ins in instances.values()]],
+ )
+ .score
+ )
+ return score
+
+ def asr_transcribe(self, instances):
+ self.logger.warn("Beta feature: Evaluating speech output. Faieseq is required.")
+ try:
+ import fairseq
+
+ fairseq_path = Path(fairseq.__path__[0]).parent # type: ignore
+ except Exception:
+ self.logger.warn("Please install fairseq.")
+ return ["" for _ in instances.keys()]
+
+ wav_dir = Path(instances[0].prediction).absolute().parent
+ root_dir = wav_dir.parent
+ transcripts_path = root_dir / "asr_transcripts.txt"
+ asr_cmd_bash_path = root_dir / "asr_cmd.bash"
+
+ # This is a dummy reference. The bleu score will be compute separately.
+ reference_path = root_dir / "instances.log"
+
+ fairseq_asr_bleu_cmd = "\n".join(
+ [
+ f"cd {fairseq_path.as_posix()}/examples/speech_to_speech/asr_bleu/",
+ " ".join(
+ [
+ "python compute_asr_bleu.py",
+ f"--reference_path {reference_path.as_posix()}",
+ f"--lang {self.target_lang}",
+ f"--audio_dirpath {wav_dir.as_posix()}",
+ "--reference_format txt",
+ f"--transcripts_path {(root_dir / 'asr_transcripts.txt').as_posix()}",
+ ]
+ ),
+ ]
+ )
+ with open(asr_cmd_bash_path, "w") as f:
+ f.write(fairseq_asr_bleu_cmd + "\n")
+
+ process = subprocess.Popen(["bash", asr_cmd_bash_path], stdout=subprocess.PIPE)
+ _, stderr = process.communicate()
+
+ if process.returncode != 0:
+ self.logger.error("ASR on target speech failed:")
+ self.logger.error(str(stderr) + "\n")
+ return ["" for _ in instances.keys()]
+
+ with open(transcripts_path, "r") as f:
+ transcripts = [line.strip() for line in f]
+
+ for idx, item in enumerate(transcripts):
+ with open(wav_dir / f"{idx}_pred.txt", "w") as f:
+ f.write(item.lower() + "\n")
+
+ return transcripts
+
+ @staticmethod
+ def add_args(parser):
+ add_sacrebleu_args(parser)
+ parser.add_argument(
+ "--target-speech-lang",
+ type=str,
+ default="en",
+ help="The language of target speech",
+ )
+
+ @classmethod
+ def from_args(cls, args):
+ return cls(args.sacrebleu_tokenizer, args.target_speech_lang)
+
+
+PUNCTUATIONS_EXCLUDE_APOSTROPHE = (
+ string.punctuation.replace("'", "") + "¡¨«°³º»¿‘“”…♪♫ˆᵉ™，ʾ˚"
+)
+PUNCTUATIONS_TO_SPACE = "-/–·—•"
+
+
+def remove_punctuations(text, punctuations=string.punctuation):
+ text = text.translate(
+ str.maketrans(PUNCTUATIONS_TO_SPACE, " " * len(PUNCTUATIONS_TO_SPACE))
+ )
+ return text.translate(str.maketrans("", "", punctuations))
+
+
+@register_quality_scorer("WHISPER_ASR_BLEU")
+class WhisperASRSacreBLEUScorer(QualityScorer):
+ """
+ Whisper ASR + SacreBLEU Scorer with whisper model
+
+ Usage:
+ :code:`--quality-metrics ASR_BLEU`
+
+ Additional command line arguments:
+
+ .. argparse::
+ :ref: simuleval.evaluator.scorers.quality_scorer.add_sacrebleu_args
+ :passparser:
+ :prog:
+ """
+
+ def __init__(
+ self,
+ tokenizer: str = "13a",
+ target_lang: str = "en",
+ model_size: str = "base",
+ lowercase: bool = False,
+ remove_punctuations: bool = False,
+ ) -> None:
+ super().__init__()
+ self.logger = logging.getLogger("simuleval.scorer.whisper_asr_bleu")
+ self.tokenizer = tokenizer
+ self.target_lang = target_lang
+ self.model_size = model_size
+ self.lowercase = lowercase
+ self.remove_punctuations = remove_punctuations
+
+ def __call__(self, instances: Dict) -> float:
+ transcripts = self.asr_transcribe(instances)
+ score = (
+ BLEU(tokenize=self.tokenizer)
+ .corpus_score(
+ transcripts,
+ [[ins.reference for ins in instances.values()]],
+ )
+ .score
+ )
+ return score
+
+ def asr_transcribe(self, instances):
+ self.logger.info(
+ "Evaluating speech output by ASR BLEU. whisper and sacrebleu are required."
+ )
+ self.logger.info("Configs:")
+ self.logger.info(f"tokenizer = {self.tokenizer}")
+ self.logger.info(f"target_lang = {self.target_lang}")
+ self.logger.info(f"model_size = {self.model_size}")
+ self.logger.info(f"lowercase = {self.lowercase}")
+ self.logger.info(f"remove_punctuations = {self.remove_punctuations}")
+ try:
+ import whisper
+ except Exception:
+ self.logger.warn("Please install whisper.")
+ return ["" for _ in instances.keys()]
+
+ model = whisper.load_model(self.model_size)
+ wav_dir = Path(instances[0].prediction).absolute().parent
+
+ transcripts = []
+ for index in tqdm.tqdm(instances.keys()):
+ wav_path = wav_dir / f"{index}_pred.wav"
+ if wav_path.exists():
+ result = model.transcribe(
+ wav_path.as_posix(), language=self.target_lang
+ )
+ text = result["text"]
+ assert type(text) == str
+ if self.lowercase:
+ text = text.lower()
+ if self.remove_punctuations:
+ text = remove_punctuations(text)
+ transcripts.append(text.strip())
+ else:
+ transcripts.append("")
+
+ root_dir = wav_dir.parent
+ transcripts_path = root_dir / "asr_transcripts.txt"
+ with open(transcripts_path, "w") as f:
+ for line in transcripts:
+ f.write(line + "\n")
+
+ return transcripts
+
+ @staticmethod
+ def add_args(parser):
+ add_sacrebleu_args(parser)
+ parser.add_argument(
+ "--target-speech-lang",
+ type=str,
+ default="en",
+ help="The language of target speech",
+ )
+ parser.add_argument(
+ "--whisper-model-size",
+ type=str,
+ default="large",
+ help="The size of whisper asr model",
+ )
+ parser.add_argument(
+ "--transcript-lowercase",
+ action="store_true",
+ help="Lowercase the whisper output",
+ )
+ parser.add_argument(
+ "--transcript-non-punctuation",
+ action="store_true",
+ help="Remove punctuations in the whisper output",
+ )
+
+ @classmethod
+ def from_args(cls, args):
+ return cls(
+ args.sacrebleu_tokenizer,
+ args.target_speech_lang,
+ args.whisper_model_size,
+ args.transcript_lowercase,
+ args.transcript_non_punctuation,
+ )
diff --git a/SimulEval/simuleval/options.py b/SimulEval/simuleval/options.py
new file mode 100644
index 0000000000000000000000000000000000000000..650f4eed94e58eed22fb93deb3262e4dfd11e424
--- /dev/null
+++ b/SimulEval/simuleval/options.py
@@ -0,0 +1,168 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import logging
+import argparse
+from typing import List, Optional
+from simuleval.data.dataloader import DATALOADER_DICT, GenericDataloader
+from simuleval.evaluator.scorers import get_scorer_class
+
+
+def add_dataloader_args(
+ parser: argparse.ArgumentParser, cli_argument_list: Optional[List[str]] = None
+):
+ if cli_argument_list is None:
+ args, _ = parser.parse_known_args()
+ else:
+ args, _ = parser.parse_known_args(cli_argument_list)
+ dataloader_class = DATALOADER_DICT.get(args.dataloader)
+ if dataloader_class is None:
+ dataloader_class = GenericDataloader
+ dataloader_class.add_args(parser)
+
+
+def add_evaluator_args(parser: argparse.ArgumentParser):
+ parser.add_argument(
+ "--quality-metrics",
+ nargs="+",
+ default=["BLEU"],
+ help="Quality metrics",
+ )
+ parser.add_argument(
+ "--latency-metrics",
+ nargs="+",
+ default=["LAAL", "AL", "AP", "DAL", "ATD"],
+ help="Latency metrics",
+ )
+ parser.add_argument(
+ "--continue-unfinished",
+ action="store_true",
+ default=False,
+ help="Continue the experiments in output dir.",
+ )
+ parser.add_argument(
+ "--computation-aware",
+ action="store_true",
+ default=False,
+ help="Include computational latency.",
+ )
+ parser.add_argument(
+ "--no-use-ref-len",
+ action="store_true",
+ default=False,
+ help="Include computational latency.",
+ )
+ parser.add_argument(
+ "--eval-latency-unit",
+ type=str,
+ default="word",
+ choices=["word", "char"],
+ help="Basic unit used for latency calculation, choose from "
+ "words (detokenized) and characters.",
+ )
+ parser.add_argument(
+ "--remote-address",
+ default="localhost",
+ help="Address to client backend",
+ )
+ parser.add_argument(
+ "--remote-port",
+ default=12321,
+ help="Port to client backend",
+ )
+ parser.add_argument(
+ "--no-progress-bar",
+ action="store_true",
+ default=False,
+ help="Do not use progress bar",
+ )
+ parser.add_argument(
+ "--start-index",
+ type=int,
+ default=0,
+ help="Start index for evaluation.",
+ )
+ parser.add_argument(
+ "--end-index",
+ type=int,
+ default=-1,
+ help="The last index for evaluation.",
+ )
+ parser.add_argument("--output", type=str, default=None, help="Output directory")
+
+
+def add_scorer_args(
+ parser: argparse.ArgumentParser, cli_argument_list: Optional[List[str]] = None
+):
+ if cli_argument_list is None:
+ args, _ = parser.parse_known_args()
+ else:
+ args, _ = parser.parse_known_args(cli_argument_list)
+
+ for metric in args.latency_metrics:
+ get_scorer_class("latency", metric).add_args(parser)
+
+ for metric in args.quality_metrics:
+ get_scorer_class("quality", metric).add_args(parser)
+
+
+def general_parser():
+ parser = argparse.ArgumentParser()
+ parser.add_argument(
+ "--remote-eval",
+ action="store_true",
+ help="Evaluate a standalone agent",
+ )
+ parser.add_argument(
+ "--standalone",
+ action="store_true",
+ help="",
+ )
+ parser.add_argument(
+ "--slurm", action="store_true", default=False, help="Use slurm."
+ )
+ parser.add_argument("--agent", default=None, help="Agent file")
+ parser.add_argument(
+ "--agent-class",
+ default=None,
+ help="The full string of class of the agent.",
+ )
+ parser.add_argument(
+ "--system-dir",
+ default=None,
+ help="Directory that contains everything to start the simultaneous system.",
+ )
+ parser.add_argument(
+ "--system-config",
+ default="main.yaml",
+ help="Name of the config yaml of the system configs.",
+ )
+ parser.add_argument("--dataloader", default=None, help="Dataloader to use")
+ parser.add_argument(
+ "--log-level",
+ type=str,
+ default="info",
+ choices=[x.lower() for x in logging._levelToName.values()],
+ help="Log level.",
+ )
+ parser.add_argument(
+ "--score-only",
+ action="store_true",
+ default=False,
+ help="Only score the inference file.",
+ )
+ parser.add_argument(
+ "--device", type=str, default="cpu", help="Device to run the model."
+ )
+ return parser
+
+
+def add_slurm_args(parser):
+ parser.add_argument(
+ "--slurm-partition", default="learnaccel,ust", help="Slurm partition."
+ )
+ parser.add_argument("--slurm-job-name", default="simuleval", help="Slurm job name.")
+ parser.add_argument("--slurm-time", default="10:00:00", help="Slurm partition.")
diff --git a/SimulEval/simuleval/test/test_agent.py b/SimulEval/simuleval/test/test_agent.py
new file mode 100644
index 0000000000000000000000000000000000000000..1f01d061f796615d4cf5e1f9207e0052dadf7eb7
--- /dev/null
+++ b/SimulEval/simuleval/test/test_agent.py
@@ -0,0 +1,61 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import os
+import tempfile
+from pathlib import Path
+
+import simuleval.cli as cli
+from simuleval.agents import TextToTextAgent
+from simuleval.agents.actions import ReadAction, WriteAction
+from simuleval.data.segments import TextSegment
+
+ROOT_PATH = Path(__file__).parents[2]
+
+
+def test_agent(root_path=ROOT_PATH):
+ with tempfile.TemporaryDirectory() as tmpdirname:
+ cli.sys.argv[1:] = [
+ "--agent",
+ os.path.join(root_path, "examples", "quick_start", "first_agent.py"),
+ "--source",
+ os.path.join(root_path, "examples", "quick_start", "source.txt"),
+ "--target",
+ os.path.join(root_path, "examples", "quick_start", "target.txt"),
+ "--output",
+ tmpdirname,
+ ]
+ cli.main()
+
+
+def test_statelss_agent(root_path=ROOT_PATH):
+ class DummyWaitkTextAgent(TextToTextAgent):
+ waitk = 0
+ vocab = [chr(i) for i in range(ord("A"), ord("Z") + 1)]
+
+ def policy(self, states=None):
+ if states is None:
+ states = self.states
+
+ lagging = len(states.source) - len(states.target)
+
+ if lagging >= self.waitk or states.source_finished:
+ prediction = self.vocab[len(states.source)]
+
+ return WriteAction(prediction, finished=(lagging <= 1))
+ else:
+ return ReadAction()
+
+ args = None
+ agent_stateless = DummyWaitkTextAgent.from_args(args)
+ agent_state = agent_stateless.build_states()
+ agent_stateful = DummyWaitkTextAgent.from_args(args)
+
+ for _ in range(10):
+ segment = TextSegment(0, "A")
+ output_1 = agent_stateless.pushpop(segment, agent_state)
+ output_2 = agent_stateful.pushpop(segment)
+ assert output_1.content == output_2.content
diff --git a/SimulEval/simuleval/test/test_agent_pipeline.py b/SimulEval/simuleval/test/test_agent_pipeline.py
new file mode 100644
index 0000000000000000000000000000000000000000..9aa1b57b187406b486d1df6e9db6f256a9dd1d30
--- /dev/null
+++ b/SimulEval/simuleval/test/test_agent_pipeline.py
@@ -0,0 +1,62 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import os
+from pathlib import Path
+
+import simuleval.cli as cli
+from simuleval.agents import AgentPipeline, TextToTextAgent
+from simuleval.agents.actions import ReadAction, WriteAction
+from simuleval.data.segments import TextSegment
+
+ROOT_PATH = Path(__file__).parents[2]
+
+
+def test_pipeline_cmd(root_path=ROOT_PATH):
+ cli.sys.argv[1:] = [
+ "--agent",
+ os.path.join(root_path, "examples", "quick_start", "agent_pipeline.py"),
+ "--source",
+ os.path.join(root_path, "examples", "quick_start", "source.txt"),
+ "--target",
+ os.path.join(root_path, "examples", "quick_start", "target.txt"),
+ ]
+ cli.main()
+
+
+def test_pipeline():
+ class DummyWaitkTextAgent(TextToTextAgent):
+ waitk = 0
+ vocab = [chr(i) for i in range(ord("A"), ord("Z") + 1)]
+
+ def policy(self):
+ lagging = len(self.states.source) - len(self.states.target)
+
+ if lagging >= self.waitk or self.states.source_finished:
+ prediction = self.vocab[len(self.states.source)]
+
+ return WriteAction(prediction, finished=(lagging <= 1))
+ else:
+ return ReadAction()
+
+ class DummyWait2TextAgent(DummyWaitkTextAgent):
+ waitk = 2
+
+ class DummyWait4TextAgent(DummyWaitkTextAgent):
+ waitk = 4
+
+ class DummyPipeline(AgentPipeline):
+ pipeline = [DummyWait2TextAgent, DummyWait4TextAgent]
+
+ args = None
+ agent_1 = DummyPipeline.from_args(args)
+ agent_2 = DummyPipeline.from_args(args)
+ for _ in range(10):
+ segment = TextSegment(0, "A")
+ output_1 = agent_1.pushpop(segment)
+ agent_2.push(segment)
+ output_2 = agent_2.pop()
+ assert output_1.content == output_2.content
diff --git a/SimulEval/simuleval/test/test_evaluator.py b/SimulEval/simuleval/test/test_evaluator.py
new file mode 100644
index 0000000000000000000000000000000000000000..d16d0f6dfc761c23d6e7920ba9ccea8f16200955
--- /dev/null
+++ b/SimulEval/simuleval/test/test_evaluator.py
@@ -0,0 +1,30 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import os
+import tempfile
+from pathlib import Path
+
+import simuleval.cli as cli
+
+ROOT_PATH = Path(__file__).parents[2]
+
+
+def test_score_only(root_path=ROOT_PATH):
+ with tempfile.TemporaryDirectory() as tmpdirname:
+ cli.sys.argv[1:] = [
+ "--agent",
+ os.path.join(root_path, "examples", "quick_start", "first_agent.py"),
+ "--source",
+ os.path.join(root_path, "examples", "quick_start", "source.txt"),
+ "--target",
+ os.path.join(root_path, "examples", "quick_start", "target.txt"),
+ "--output",
+ tmpdirname,
+ ]
+ cli.main()
+ cli.sys.argv[1:] = ["--score-only", "--output", tmpdirname]
+ cli.main()
diff --git a/SimulEval/simuleval/test/test_remote_evaluation.py b/SimulEval/simuleval/test/test_remote_evaluation.py
new file mode 100644
index 0000000000000000000000000000000000000000..953dec1084abf049dcb9289fa0625cfad73d146e
--- /dev/null
+++ b/SimulEval/simuleval/test/test_remote_evaluation.py
@@ -0,0 +1,59 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import os
+import tempfile
+import time
+from multiprocessing import Process
+from pathlib import Path
+
+import simuleval.cli as cli
+from simuleval.utils.functional import find_free_port
+
+ROOT_PATH = Path(__file__).parents[2]
+
+
+def p1(port, root_path):
+ cli.sys.argv[1:] = [
+ "--standalone",
+ "--remote-port",
+ str(port),
+ "--agent",
+ os.path.join(root_path, "examples", "quick_start", "first_agent.py"),
+ ]
+ cli.main()
+ time.sleep(5)
+
+
+def p2(port, root_path):
+ with tempfile.TemporaryDirectory() as tmpdirname:
+ cli.sys.argv[1:] = [
+ "--remote-eval",
+ "--remote-port",
+ str(port),
+ "--source",
+ os.path.join(root_path, "examples", "quick_start", "source.txt"),
+ "--target",
+ os.path.join(root_path, "examples", "quick_start", "target.txt"),
+ "--dataloader",
+ "text-to-text",
+ "--output",
+ tmpdirname,
+ ]
+ cli.main()
+
+
+def test_remote_eval(root_path=ROOT_PATH):
+ port = find_free_port()
+
+ p_1 = Process(target=p1, args=(port, root_path))
+ p_1.start()
+
+ p_2 = Process(target=p2, args=(port, root_path))
+ p_2.start()
+
+ p_1.kill()
+ p_2.kill()
diff --git a/SimulEval/simuleval/utils/__init__.py b/SimulEval/simuleval/utils/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..2032486a831672d174bf27082b8fa783c72da667
--- /dev/null
+++ b/SimulEval/simuleval/utils/__init__.py
@@ -0,0 +1,12 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from .agent import build_system_from_dir, EVALUATION_SYSTEM_LIST # noqa F401
+
+
+def entrypoint(klass):
+ EVALUATION_SYSTEM_LIST.append(klass)
+ return klass
diff --git a/SimulEval/simuleval/utils/agent.py b/SimulEval/simuleval/utils/agent.py
new file mode 100644
index 0000000000000000000000000000000000000000..e4f5736cced4dd6a874184582afb185541df35f9
--- /dev/null
+++ b/SimulEval/simuleval/utils/agent.py
@@ -0,0 +1,140 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+
+import os
+import sys
+import yaml
+import logging
+import importlib
+from argparse import Namespace
+from typing import Union, Optional, Tuple
+from pathlib import Path
+from simuleval import options
+from simuleval.agents import GenericAgent
+from simuleval.utils.arguments import cli_argument_list, check_argument
+
+EVALUATION_SYSTEM_LIST = []
+
+logger = logging.getLogger("simuleval.utils.agent")
+
+
+def import_file(file_path):
+ spec = importlib.util.spec_from_file_location("agents", file_path)
+ agent_modules = importlib.util.module_from_spec(spec)
+ spec.loader.exec_module(agent_modules)
+
+
+def get_agent_class(config_dict: Optional[dict] = None) -> GenericAgent:
+ class_name = check_argument("agent_class", config_dict)
+
+ if class_name is not None:
+ if check_argument("agent"):
+ raise RuntimeError("Use either --agent or --agent-class, not both.")
+ EVALUATION_SYSTEM_LIST.append(get_agent_class_from_string(class_name))
+
+ system_dir = check_argument("system_dir")
+ config_name = check_argument("system_config")
+
+ if system_dir is not None:
+ EVALUATION_SYSTEM_LIST.append(get_agent_class_from_dir(system_dir, config_name))
+
+ agent_file = check_argument("agent")
+ if agent_file is not None:
+ import_file(agent_file)
+
+ if len(EVALUATION_SYSTEM_LIST) == 0:
+ raise RuntimeError(
+ "Please use @entrypoint decorator to indicate the system you want to evaluate."
+ )
+ if len(EVALUATION_SYSTEM_LIST) > 1:
+ raise RuntimeError("More than one system is not supported right now.")
+ return EVALUATION_SYSTEM_LIST[0]
+
+
+def get_system_config(path: Union[Path, str], config_name) -> dict:
+ path = Path(path)
+ with open(path / config_name, "r") as f:
+ try:
+ config_dict = yaml.safe_load(f)
+ except yaml.YAMLError as exc:
+ logging.error(f"Failed to load configs from {path / config_name}.")
+ logging.error(exc)
+ sys.exit(1)
+ return config_dict
+
+
+def get_agent_class_from_string(class_name: str) -> GenericAgent:
+ try:
+ agent_module = importlib.import_module(".".join(class_name.split(".")[:-1]))
+ agent_class = getattr(agent_module, class_name.split(".")[-1])
+ except Exception as e:
+ logger.error(f"Not able to load {class_name}.")
+ raise e
+ return agent_class
+
+
+def get_agent_class_from_dir(
+ path: Union[Path, str], config_name: str = "main.yaml"
+) -> GenericAgent:
+ config_dict = get_system_config(path, config_name)
+ assert "agent_class" in config_dict
+ class_name = config_dict["agent_class"]
+ return get_agent_class_from_string(class_name)
+
+
+def build_system_from_dir(
+ path: Union[Path, str],
+ config_name: str = "main.yaml",
+ overwrite_config_dict: Optional[dict] = None,
+) -> GenericAgent:
+ path = Path(path)
+ config_dict = get_system_config(path, config_name)
+ if overwrite_config_dict is not None:
+ for key, value in overwrite_config_dict:
+ config_dict[key] = value
+ agent_class = get_agent_class_from_dir(path, config_name)
+
+ parser = options.general_parser()
+ agent_class.add_args(parser)
+ args, _ = parser.parse_known_args(cli_argument_list(config_dict))
+ sys.path.append(path.as_posix())
+
+ cur_dir = os.getcwd()
+ os.chdir(path.as_posix())
+ system = agent_class.from_args(args)
+ os.chdir(cur_dir)
+ return system
+
+
+def build_system_args(
+ config_dict: Optional[dict] = None,
+) -> Tuple[GenericAgent, Namespace]:
+ parser = options.general_parser()
+ cli_arguments = cli_argument_list(config_dict)
+ options.add_evaluator_args(parser)
+ options.add_scorer_args(parser, cli_arguments)
+ options.add_slurm_args(parser)
+ options.add_dataloader_args(parser, cli_arguments)
+
+ if check_argument("system_dir"):
+ system = build_system_from_dir(
+ check_argument("system_dir"), check_argument("system_config"), config_dict
+ )
+ else:
+ system_class = get_agent_class(config_dict)
+ system_class.add_args(parser)
+ args, _ = parser.parse_known_args(cli_argument_list(config_dict))
+ system = system_class.from_args(args)
+
+ args = parser.parse_args(cli_argument_list(config_dict))
+
+ logger.info(f"System will run on device: {args.device}.")
+ system.to(args.device)
+
+ args.source_type = system.source_type
+ args.target_type = system.target_type
+ return system, args
diff --git a/SimulEval/simuleval/utils/arguments.py b/SimulEval/simuleval/utils/arguments.py
new file mode 100644
index 0000000000000000000000000000000000000000..b2e42119d0e5d66c01e7ccb6e8df67d699ffd33e
--- /dev/null
+++ b/SimulEval/simuleval/utils/arguments.py
@@ -0,0 +1,25 @@
+import sys
+from typing import Optional
+from simuleval import options
+
+
+def cli_argument_list(config_dict: Optional[dict]):
+ if config_dict is None:
+ return sys.argv[1:]
+ else:
+ string = ""
+ for key, value in config_dict.items():
+ if f"--{key.replace('_', '-')}" in sys.argv:
+ continue
+
+ if type(value) is not bool:
+ string += f" --{key.replace('_', '-')} {value}"
+ else:
+ string += f" --{key.replace('_', '-')}"
+ return sys.argv[1:] + string.split()
+
+
+def check_argument(name: str, config_dict: Optional[dict] = None):
+ parser = options.general_parser()
+ args, _ = parser.parse_known_args(cli_argument_list(config_dict))
+ return getattr(args, name)
diff --git a/SimulEval/simuleval/utils/functional.py b/SimulEval/simuleval/utils/functional.py
new file mode 100644
index 0000000000000000000000000000000000000000..aaffe993bdd42d7809f3e8fdc941c81bee31ac1e
--- /dev/null
+++ b/SimulEval/simuleval/utils/functional.py
@@ -0,0 +1,15 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from contextlib import closing
+import socket
+
+
+def find_free_port():
+ with closing(socket.socket(socket.AF_INET, socket.SOCK_STREAM)) as s:
+ s.bind(("", 0))
+ s.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEADDR, 1)
+ return s.getsockname()[1]
diff --git a/SimulEval/simuleval/utils/slurm.py b/SimulEval/simuleval/utils/slurm.py
new file mode 100644
index 0000000000000000000000000000000000000000..6031dd8f43c43460623285f8673eee65b6eab0ff
--- /dev/null
+++ b/SimulEval/simuleval/utils/slurm.py
@@ -0,0 +1,106 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import os
+import re
+import sys
+import logging
+import subprocess
+from typing import Optional, Dict
+from simuleval import options
+from simuleval.utils.arguments import cli_argument_list
+from simuleval.utils.agent import get_agent_class
+
+logger = logging.getLogger("simuleval.slurm")
+
+
+def mkdir_output_dir(path: str) -> bool:
+ try:
+ os.makedirs(path, exist_ok=True)
+ return True
+ except BaseException as be:
+ logger.error(f"Failed to write results to {path}.")
+ logger.error(be)
+ logger.error("Skip writing predictions.")
+ return False
+
+
+def submit_slurm_job(config_dict: Optional[Dict] = None) -> None:
+ if config_dict is not None and "slurm" in config_dict:
+ raise RuntimeError("--slurm is only available as a CLI argument")
+ parser = options.general_parser()
+ cli_arguments = cli_argument_list(config_dict)
+ options.add_evaluator_args(parser)
+ options.add_scorer_args(parser, cli_arguments)
+ options.add_slurm_args(parser)
+ options.add_dataloader_args(parser, cli_arguments)
+ system_class = get_agent_class(config_dict)
+ system_class.add_args(parser)
+ args = parser.parse_args(cli_argument_list(config_dict))
+ args.output = os.path.abspath(args.output)
+ assert mkdir_output_dir(args.output)
+
+ if args.agent is None:
+ args.agent = sys.argv[0]
+
+ os.system(f"cp {args.agent} {args.output}/agent.py")
+ _args = [sys.argv[0]]
+ for arg in sys.argv[1:]:
+ if str(arg).isdigit() or str(arg).startswith("--"):
+ _args.append(arg)
+ else:
+ _args.append(f'"{arg}"')
+ command = " ".join(_args).strip()
+ command = re.sub(r"(--slurm\S*(\s+[^-]\S+)*)", "", command).strip()
+ if subprocess.check_output(["which", "simuleval"]).decode().strip() in command:
+ command = re.sub(
+ r"--agent\s+\S+", f"--agent {args.output}/agent.py", command
+ ).strip()
+ else:
+ # Attention: not fully tested!
+ command = re.sub(
+ r"[^\"'\s]+\.py", f"{os.path.abspath(args.output)}/agent.py", command
+ ).strip()
+
+ if "--output" in command:
+ command = re.sub(r"--output\s+\S+", f"--output {args.output}", command).strip()
+ else:
+ command += f" --output {args.output}"
+
+ command = command.replace("--", "\\\n\t--")
+ script = f"""#!/bin/bash
+#SBATCH --time={args.slurm_time}
+#SBATCH --partition={args.slurm_partition}
+#SBATCH --nodes=1
+#SBATCH --gpus-per-node=1
+#SBATCH --ntasks-per-node=8
+#SBATCH --output="{args.output}/slurm-%j.log"
+#SBATCH --job-name="{args.slurm_job_name}"
+
+cd {os.path.abspath(args.output)}
+
+GPU_ID=$SLURM_LOCALID
+
+# Change to local a gpu id for debugging, e.g.
+# GPU_ID=0
+
+CUDA_VISIBLE_DEVICES=$GPU_ID {command}
+ """
+ script_file = os.path.join(args.output, "script.sh")
+ with open(script_file, "w") as f:
+ f.writelines(script)
+
+ process = subprocess.Popen(
+ ["sbatch", script_file],
+ stderr=subprocess.PIPE,
+ stdout=subprocess.PIPE,
+ )
+ stdout, stderr = process.communicate()
+ logger.info("Using slurm.")
+ logger.info(f"sbatch stdout: {stdout.decode('utf-8').strip()}")
+ stderr = stderr.decode("utf-8").strip()
+ if len(stderr) > 0:
+ logger.info(f"sbatch stderr: {stderr.decode('utf-8').strip()}")
diff --git "a/StreamSpeech\357\274\232Simultaneous Speech-to-Speech Translation with.pdf" "b/StreamSpeech\357\274\232Simultaneous Speech-to-Speech Translation with.pdf"
new file mode 100644
index 0000000000000000000000000000000000000000..3bf90a6aabd157bf35e4414066c5b608346a7885
--- /dev/null
+++ "b/StreamSpeech\357\274\232Simultaneous Speech-to-Speech Translation with.pdf"
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bcaac07cca64b6b9b2cf1cd5dd9d5cd56577ab5e1e06df5e36e73bb9c3aff1eb
+size 6568281
diff --git a/agent/__init__.py b/agent/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..cf73bf692e48ef6f73fe9596f21d98c1e207eac9
--- /dev/null
+++ b/agent/__init__.py
@@ -0,0 +1,5 @@
+import os
+import importlib
+
+importlib.import_module("agent.sequence_generator")
+print("import agents...")
diff --git a/agent/ctc_decoder.py b/agent/ctc_decoder.py
new file mode 100644
index 0000000000000000000000000000000000000000..79a67bafcef634b8c5180143df7f2bc91129b497
--- /dev/null
+++ b/agent/ctc_decoder.py
@@ -0,0 +1,111 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import json
+import logging
+import math
+from argparse import Namespace
+from pathlib import Path
+from typing import List
+
+import torch
+import torch.nn as nn
+from fairseq import utils
+from fairseq.data import Dictionary
+from fairseq.data.audio.data_cfg import MultitaskConfig, S2SDataConfig
+from fairseq.data.audio.speech_to_speech_dataset import SpeechToSpeechDatasetCreator
+from fairseq.data.audio.speech_to_text_dataset import (
+ SpeechToTextDataset,
+ TextTargetMultitaskData,
+)
+from fairseq.tasks import LegacyFairseqTask, register_task
+from fairseq.tasks.speech_to_text import DummyMultiTask
+from fairseq.tasks.text_to_speech import batch_mel_cepstral_distortion
+
+logger = logging.getLogger(__name__)
+
+
+class CTCDecoder(nn.Module):
+ def __init__(self, tgt_dict, models):
+ super().__init__()
+ self.pad = tgt_dict.pad()
+ self.eos = tgt_dict.eos()
+ self.unk = tgt_dict.unk()
+ self.models = models
+ self.tgt_dict = tgt_dict
+
+ @torch.no_grad()
+ def generate(self, encoder_out, prefix=None, aux_task_name=None, **kwargs):
+ model = self.models[0]
+ model.eval()
+
+ max_len = model.max_decoder_positions()
+ # TODO: incorporate max_len_a and max_len_b
+
+ incremental_state = {}
+ pred_out, attn, scores = [], [], []
+
+ prev_output_tokens = None
+ decoder_name = f"{aux_task_name}_decoder" if aux_task_name else "decoder"
+ ctc_decoder = getattr(model, decoder_name)
+ ctc_out = ctc_decoder(encoder_out["encoder_out"][0], **kwargs)
+ lprobs = model.get_normalized_probs(
+ [ctc_out["encoder_out"].transpose(0, 1)], log_probs=True
+ )
+ # never select pad, unk
+ lprobs[:, :, self.pad] = -math.inf
+ lprobs[:, :, self.unk] = -math.inf
+
+ cur_pred_lprob, cur_pred_out = torch.max(lprobs, dim=2)
+ scores = cur_pred_lprob
+ pred_out = cur_pred_out
+ attn = None
+ alignment = None
+
+ def _ctc_postprocess(tokens):
+ _toks = tokens.int().tolist()
+ deduplicated_toks = [
+ v for i, v in enumerate(_toks) if i == 0 or v != _toks[i - 1]
+ ]
+ hyp = [
+ v
+ for v in deduplicated_toks
+ if (v != 0) and (v != self.tgt_dict.pad_index)
+ ]
+ return torch.tensor(hyp)
+
+ def _ctc_postprocess_index(tokens):
+ _toks = tokens.int().tolist()
+ deduplicated_toks = [
+ (v, i) for i, v in enumerate(_toks) if i == 0 or v != _toks[i - 1]
+ ]
+ index = [
+ i
+ for v, i in deduplicated_toks
+ if (v != 0) and (v != self.tgt_dict.pad_index)
+ ]
+ return index
+
+ if prefix is not None:
+
+ pred_out = torch.cat((prefix, pred_out[:, prefix.size(1) :]), dim=1)
+
+ hypos = [
+ [
+ {
+ "tokens": _ctc_postprocess(pred_out[b]),
+ "org_tokens": pred_out[b],
+ "lprobs": lprobs,
+ "index": _ctc_postprocess_index(pred_out[b]),
+ "attn": None,
+ "alignment": None,
+ "positional_scores": scores[b],
+ "score": utils.item(scores[b].sum().data),
+ }
+ ]
+ for b in range(pred_out.size(0))
+ ]
+
+ return hypos
diff --git a/agent/ctc_generator.py b/agent/ctc_generator.py
new file mode 100644
index 0000000000000000000000000000000000000000..286a3ba358435f4bfc0db075f82ac3fb11bbdfb8
--- /dev/null
+++ b/agent/ctc_generator.py
@@ -0,0 +1,123 @@
+import json
+import logging
+import math
+from argparse import Namespace
+from pathlib import Path
+from typing import Dict, List, Optional
+
+import torch
+import torch.nn as nn
+from torch import Tensor
+from fairseq import utils
+from fairseq.data import Dictionary
+from fairseq.data.audio.data_cfg import MultitaskConfig, S2SDataConfig
+from fairseq.data.audio.speech_to_speech_dataset import SpeechToSpeechDatasetCreator
+from fairseq.data.audio.speech_to_text_dataset import (
+ SpeechToTextDataset,
+ TextTargetMultitaskData,
+)
+from fairseq.tasks import LegacyFairseqTask, register_task
+from fairseq.tasks.speech_to_text import DummyMultiTask
+from fairseq.tasks.text_to_speech import batch_mel_cepstral_distortion
+
+logger = logging.getLogger(__name__)
+
+
+class CTCSequenceGenerator(nn.Module):
+ def __init__(self, tgt_dict, models, use_incremental_states=False):
+ super().__init__()
+ self.pad = tgt_dict.pad()
+ self.eos = tgt_dict.eos()
+ self.unk = tgt_dict.unk()
+ self.models = models
+ self.tgt_dict = tgt_dict
+ self.use_incremental_states = use_incremental_states
+ self.incremental_states = None
+
+ def reset_incremental_states(self):
+ self.incremental_states = None
+
+ @torch.no_grad()
+ def generate(self, encoder_out, prefix=None, aux_task_name=None, **kwargs):
+ if self.use_incremental_states:
+ if self.incremental_states is None:
+ incremental_states = torch.jit.annotate(
+ List[Dict[str, Dict[str, Optional[Tensor]]]],
+ [
+ torch.jit.annotate(Dict[str, Dict[str, Optional[Tensor]]], {})
+ for i in range(1)
+ ],
+ )
+ self.incremental_states = incremental_states
+ else:
+ incremental_states = self.incremental_states
+ else:
+ incremental_states = None
+
+ # currently only support viterbi search for stacked units
+ model = self.models[0]
+ model.eval()
+
+ max_len = model.max_decoder_positions()
+ # TODO: incorporate max_len_a and max_len_b
+
+ incremental_state = {}
+ pred_out, attn, scores = [], [], []
+
+ prev_output_tokens = None
+ decoder_name = f"{aux_task_name}_decoder" if aux_task_name else "decoder"
+ ctc_decoder = getattr(model, decoder_name)
+ ctc_out, ctc_extra = ctc_decoder(
+ None,
+ encoder_out=encoder_out,
+ incremental_state=(
+ incremental_states[0] if self.use_incremental_states else None
+ ),
+ **kwargs,
+ )
+ lprobs = model.get_normalized_probs([ctc_out], log_probs=True)
+
+ # never select pad, unk
+ lprobs[:, :, self.pad] = -math.inf
+ lprobs[:, :, self.unk] = -math.inf
+
+ cur_pred_lprob, cur_pred_out = torch.max(lprobs, dim=2)
+ scores = cur_pred_lprob
+ pred_out = cur_pred_out
+
+ attn = ctc_extra["attn"][0]
+ alignment = None
+
+ def _ctc_postprocess(tokens):
+ _toks = tokens.int().tolist()
+ deduplicated_toks = [
+ v for i, v in enumerate(_toks) if i == 0 or v != _toks[i - 1]
+ ]
+ hyp = [
+ v
+ for v in deduplicated_toks
+ if (v != self.tgt_dict.blank_index) and (v != self.tgt_dict.pad_index)
+ ]
+ return torch.tensor(hyp)
+
+ if prefix is not None:
+ if self.use_incremental_states:
+ pred_out = torch.cat((prefix, pred_out), dim=1)
+ else:
+ pred_out = torch.cat((prefix, pred_out[:, prefix.size(1) :]), dim=1)
+
+ hypos = [
+ [
+ {
+ "tokens": _ctc_postprocess(pred_out[b]),
+ "org_tokens": pred_out[b],
+ "attn": None,
+ "alignment": None,
+ "positional_scores": scores[b],
+ "score": utils.item(scores[b].sum().data),
+ }
+ ]
+ for b in range(pred_out.size(0))
+ ]
+
+ return hypos
diff --git a/agent/mt/__init__.py b/agent/mt/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..044fe72a2de96d87bec0ddb8655a1ac349b7f89d
--- /dev/null
+++ b/agent/mt/__init__.py
@@ -0,0 +1,8 @@
+import os
+import importlib
+
+# automatically import any Python files in the criterions/ directory
+for file in os.listdir(os.path.dirname(__file__)):
+ if file.endswith(".py") and not file.startswith("_"):
+ file_name = file[: file.find(".py")]
+ importlib.import_module("agent.mt." + file_name)
diff --git a/agent/mt/hmt_sequence_generator.py b/agent/mt/hmt_sequence_generator.py
new file mode 100644
index 0000000000000000000000000000000000000000..a7d251beed7fb1c2e008f82b9c792e99b5857a88
--- /dev/null
+++ b/agent/mt/hmt_sequence_generator.py
@@ -0,0 +1,1149 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import math
+import sys
+from typing import Dict, List, Optional
+
+import torch
+import torch.nn as nn
+from torch import Tensor
+
+from fairseq import search, utils
+from fairseq.data import data_utils
+from fairseq.models import FairseqIncrementalDecoder
+from fairseq.ngram_repeat_block import NGramRepeatBlock
+from fairseq.sequence_generator import SequenceGenerator
+
+
+class HmtSequenceGenerator(nn.Module):
+ def __init__(
+ self,
+ models,
+ tgt_dict,
+ beam_size=1,
+ max_len_a=0,
+ max_len_b=200,
+ max_len=0,
+ min_len=1,
+ normalize_scores=True,
+ len_penalty=1.0,
+ unk_penalty=0.0,
+ temperature=1.0,
+ match_source_len=False,
+ no_repeat_ngram_size=0,
+ search_strategy=None,
+ eos=None,
+ symbols_to_strip_from_output=None,
+ lm_model=None,
+ lm_weight=1.0,
+ tokens_to_suppress=(),
+ ):
+ """Generates translations of a given source sentence.
+
+ Args:
+ models (List[~fairseq.models.FairseqModel]): ensemble of models,
+ currently support fairseq.models.TransformerModel for scripting
+ beam_size (int, optional): beam width (default: 1)
+ max_len_a/b (int, optional): generate sequences of maximum length
+ ax + b, where x is the source length
+ max_len (int, optional): the maximum length of the generated output
+ (not including end-of-sentence)
+ min_len (int, optional): the minimum length of the generated output
+ (not including end-of-sentence)
+ normalize_scores (bool, optional): normalize scores by the length
+ of the output (default: True)
+ len_penalty (float, optional): length penalty, where <1.0 favors
+ shorter, >1.0 favors longer sentences (default: 1.0)
+ unk_penalty (float, optional): unknown word penalty, where <0
+ produces more unks, >0 produces fewer (default: 0.0)
+ temperature (float, optional): temperature, where values
+ >1.0 produce more uniform samples and values <1.0 produce
+ sharper samples (default: 1.0)
+ match_source_len (bool, optional): outputs should match the source
+ length (default: False)
+ """
+ super().__init__()
+ if isinstance(models, EnsembleModel):
+ self.model = models
+ else:
+ self.model = EnsembleModel(models)
+ self.tgt_dict = tgt_dict
+ self.pad = tgt_dict.pad()
+ self.unk = tgt_dict.unk()
+ self.eos = tgt_dict.eos() if eos is None else eos
+ self.symbols_to_strip_from_output = (
+ symbols_to_strip_from_output.union({self.eos})
+ if symbols_to_strip_from_output is not None
+ else {self.eos}
+ )
+
+ self.token_indices_to_suppress: Optional[Tensor] = None
+ token_indices_to_suppress = []
+ for token_string in tokens_to_suppress:
+ token_index = tgt_dict.index(token_string)
+ assert token_index != self.unk
+ token_indices_to_suppress.append(token_index)
+ if len(token_indices_to_suppress) > 0:
+ self.token_indices_to_suppress = torch.Tensor(
+ token_indices_to_suppress
+ ).long()
+
+ self.vocab_size = len(tgt_dict)
+ self.beam_size = beam_size
+ # the max beam size is the dictionary size - 1, since we never select pad
+ self.beam_size = min(beam_size, self.vocab_size - 1)
+ self.model.set_decoder_beam_size(self.beam_size)
+ self.max_len_a = max_len_a
+ self.max_len_b = max_len_b
+ self.min_len = min_len
+ self.max_len = max_len or self.model.max_decoder_positions()
+
+ self.normalize_scores = normalize_scores
+ self.len_penalty = len_penalty
+ self.unk_penalty = unk_penalty
+ self.temperature = temperature
+ self.match_source_len = match_source_len
+
+ if no_repeat_ngram_size > 0:
+ self.repeat_ngram_blocker = NGramRepeatBlock(no_repeat_ngram_size)
+ else:
+ self.repeat_ngram_blocker = None
+
+ assert temperature > 0, "--temperature must be greater than 0"
+
+ self.search = (
+ search.BeamSearch(tgt_dict) if search_strategy is None else search_strategy
+ )
+ # We only need to set src_lengths in LengthConstrainedBeamSearch.
+ # As a module attribute, setting it would break in multithread
+ # settings when the model is shared.
+ self.should_set_src_lengths = (
+ hasattr(self.search, "needs_src_lengths") and self.search.needs_src_lengths
+ )
+
+ self.model.eval()
+
+ self.lm_model = lm_model
+ self.lm_weight = lm_weight
+ if self.lm_model is not None:
+ self.lm_model.eval()
+
+ def cuda(self):
+ self.model.cuda()
+ return self
+
+ @torch.no_grad()
+ def forward(
+ self,
+ sample: Dict[str, Dict[str, Tensor]],
+ prefix_tokens: Optional[Tensor] = None,
+ bos_token: Optional[int] = None,
+ ):
+ """Generate a batch of translations.
+
+ Args:
+ sample (dict): batch
+ prefix_tokens (torch.LongTensor, optional): force decoder to begin
+ with these tokens
+ bos_token (int, optional): beginning of sentence token
+ (default: self.eos)
+ """
+ return self._generate(sample, prefix_tokens, bos_token=bos_token)
+
+ # TODO(myleott): unused, deprecate after pytorch-translate migration
+ def generate_batched_itr(self, data_itr, beam_size=None, cuda=False, timer=None):
+ """Iterate over a batched dataset and yield individual translations.
+ Args:
+ cuda (bool, optional): use GPU for generation
+ timer (StopwatchMeter, optional): time generations
+ """
+ for sample in data_itr:
+ s = utils.move_to_cuda(sample) if cuda else sample
+ if "net_input" not in s:
+ continue
+ input = s["net_input"]
+ # model.forward normally channels prev_output_tokens into the decoder
+ # separately, but SequenceGenerator directly calls model.encoder
+ encoder_input = {
+ k: v for k, v in input.items() if k != "prev_output_tokens"
+ }
+ if timer is not None:
+ timer.start()
+ with torch.no_grad():
+ hypos = self.generate(encoder_input)
+ if timer is not None:
+ timer.stop(sum(len(h[0]["tokens"]) for h in hypos))
+ for i, id in enumerate(s["id"].data):
+ # remove padding
+ src = utils.strip_pad(input["src_tokens"].data[i, :], self.pad)
+ ref = (
+ utils.strip_pad(s["target"].data[i, :], self.pad)
+ if s["target"] is not None
+ else None
+ )
+ yield id, src, ref, hypos[i]
+
+ @torch.no_grad()
+ def generate(
+ self, models, sample: Dict[str, Dict[str, Tensor]], **kwargs
+ ) -> List[List[Dict[str, Tensor]]]:
+ """Generate translations. Match the api of other fairseq generators.
+
+ Args:
+ models (List[~fairseq.models.FairseqModel]): ensemble of models
+ sample (dict): batch
+ prefix_tokens (torch.LongTensor, optional): force decoder to begin
+ with these tokens
+ constraints (torch.LongTensor, optional): force decoder to include
+ the list of constraints
+ bos_token (int, optional): beginning of sentence token
+ (default: self.eos)
+ """
+ return self._generate(sample, **kwargs)
+
+ def _generate(
+ self,
+ sample: Dict[str, Dict[str, Tensor]],
+ prefix_tokens: Optional[Tensor] = None,
+ constraints: Optional[Tensor] = None,
+ bos_token: Optional[int] = None,
+ src_finished=None,
+ pre_src_prefix=None,
+ ):
+ incremental_states = torch.jit.annotate(
+ List[Dict[str, Dict[str, Optional[Tensor]]]],
+ [
+ torch.jit.annotate(Dict[str, Dict[str, Optional[Tensor]]], {})
+ for i in range(self.model.models_size)
+ ],
+ )
+ net_input = sample["net_input"]
+
+ if "src_tokens" in net_input:
+ src_tokens = net_input["src_tokens"]
+ # length of the source text being the character length except EndOfSentence and pad
+ src_lengths = (
+ (src_tokens.ne(self.eos) & src_tokens.ne(self.pad)).long().sum(dim=1)
+ )
+ elif "source" in net_input:
+ src_tokens = net_input["source"]
+ src_lengths = (
+ net_input["padding_mask"].size(-1) - net_input["padding_mask"].sum(-1)
+ if net_input["padding_mask"] is not None
+ else torch.tensor(src_tokens.size(-1)).to(src_tokens)
+ )
+ elif "features" in net_input:
+ src_tokens = net_input["features"]
+ src_lengths = (
+ net_input["padding_mask"].size(-1) - net_input["padding_mask"].sum(-1)
+ if net_input["padding_mask"] is not None
+ else torch.tensor(src_tokens.size(-1)).to(src_tokens)
+ )
+ else:
+ raise Exception(
+ "expected src_tokens or source in net input. input keys: "
+ + str(net_input.keys())
+ )
+
+ # bsz: total number of sentences in beam
+ # Note that src_tokens may have more than 2 dimensions (i.e. audio features)
+ bsz, src_len = src_tokens.size()[:2]
+ beam_size = self.beam_size
+
+ if constraints is not None and not self.search.supports_constraints:
+ raise NotImplementedError(
+ "Target-side constraints were provided, but search method doesn't support them"
+ )
+
+ # Initialize constraints, when active
+ self.search.init_constraints(constraints, beam_size)
+ start = prefix_tokens.size(-1) if prefix_tokens is not None else 0
+ if src_finished:
+ max_len: int = -1
+ if self.match_source_len:
+ max_len = src_lengths.max().item()
+ else:
+ max_len = min(
+ int(self.max_len_a * src_len + self.max_len_b),
+ self.max_len - 1,
+ )
+ else:
+ import pdb
+
+ pdb.set_trace()
+ first_read = self.model.models[0].decoder.first_read
+ cands_per_token = self.model.models[0].decoder.cands_per_token
+
+ max_len = start
+
+ assert (
+ self.min_len <= max_len
+ ), "min_len cannot be larger than max_len, please adjust these!"
+ # compute the encoder output for each beam
+ with torch.autograd.profiler.record_function("EnsembleModel: forward_encoder"):
+ encoder_outs = self.model.forward_encoder(net_input)
+
+ # placeholder of indices for bsz * beam_size to hold tokens and accumulative scores
+ new_order = torch.arange(bsz).view(-1, 1).repeat(1, beam_size).view(-1)
+ new_order = new_order.to(src_tokens.device).long()
+ encoder_outs = self.model.reorder_encoder_out(encoder_outs, new_order)
+ # ensure encoder_outs is a List.
+ assert encoder_outs is not None
+
+ # initialize buffers
+ scores = torch.zeros(bsz * beam_size, max_len + 1).to(src_tokens).float()
+
+ if prefix_tokens is None:
+ tokens = (
+ torch.zeros(bsz * beam_size, max_len + 2, device=src_tokens.device)
+ .long()
+ .fill_(self.pad)
+ ) # +2 for eos and pad
+ tokens[:, 0] = self.eos if bos_token is None else bos_token
+ else:
+ tokens = torch.cat(
+ (
+ torch.full(
+ (bsz, 1), self.eos, dtype=torch.long, device=src_tokens.device
+ ),
+ prefix_tokens,
+ ),
+ dim=-1,
+ )
+ tokens = tokens.repeat(beam_size, 1)
+ if tokens.size(-1) < max_len + 2:
+ tokens = torch.cat(
+ (
+ tokens,
+ torch.zeros(
+ tokens.size(0),
+ max_len + 2 - tokens.size(-1),
+ device=src_tokens.device,
+ ).long(),
+ ),
+ dim=-1,
+ )
+
+ attn: Optional[Tensor] = None
+
+ # A list that indicates candidates that should be ignored.
+ # For example, suppose we're sampling and have already finalized 2/5
+ # samples. Then cands_to_ignore would mark 2 positions as being ignored,
+ # so that we only finalize the remaining 3 samples.
+ cands_to_ignore = (
+ torch.zeros(bsz, beam_size).to(src_tokens).eq(-1)
+ ) # forward and backward-compatible False mask
+
+ # list of completed sentences
+ finalized = torch.jit.annotate(
+ List[List[Dict[str, Tensor]]],
+ [torch.jit.annotate(List[Dict[str, Tensor]], []) for i in range(bsz)],
+ ) # contains lists of dictionaries of infomation about the hypothesis being finalized at each step
+
+ # a boolean array indicating if the sentence at the index is finished or not
+ finished = [False for i in range(bsz)]
+ num_remaining_sent = bsz # number of sentences remaining
+
+ # number of candidate hypos per step
+ cand_size = 2 * beam_size # 2 x beam size in case half are EOS
+
+ # offset arrays for converting between different indexing schemes
+ bbsz_offsets = (
+ (torch.arange(0, bsz) * beam_size)
+ .unsqueeze(1)
+ .type_as(tokens)
+ .to(src_tokens.device)
+ )
+ cand_offsets = torch.arange(0, cand_size).type_as(tokens).to(src_tokens.device)
+
+ reorder_state: Optional[Tensor] = None
+ batch_idxs: Optional[Tensor] = None
+
+ original_batch_idxs: Optional[Tensor] = None
+ if "id" in sample and isinstance(sample["id"], Tensor):
+ original_batch_idxs = sample["id"]
+ else:
+ original_batch_idxs = torch.arange(0, bsz).type_as(tokens)
+
+ first_read = self.model.models[0].decoder.first_read
+ cands_per_token = self.model.models[0].decoder.cands_per_token
+ previous_selected_idx = None
+
+ ds = []
+ src_lens = sample["net_input"]["src_lengths"]
+ reads = None
+ read = None
+ source_length = sample["net_input"]["src_lengths"]
+ padding_length = (
+ sample["net_input"]["src_tokens"].size(1)
+ - sample["net_input"]["src_lengths"]
+ )
+ finalized_rw = {}
+ finalized_extra = {}
+ finalized_extra["transition_probs"] = {}
+ finalized_extra["attention"] = {}
+ finalized_extra["states_output"] = {}
+ finalized_src_length = sample["net_input"]["src_lengths"]
+
+ ids = sample["id"]
+ next_selected_idxs = None
+
+ for step in range(start, max_len + 1): # one extra step for EOS marker
+ # reorder decoder internal states based on the prev choice of beams
+ if reorder_state is not None:
+ if batch_idxs is not None:
+ # update beam indices to take into account removed sentences
+ corr = batch_idxs - torch.arange(batch_idxs.numel()).type_as(
+ batch_idxs
+ )
+ reorder_state.view(-1, beam_size).add_(
+ corr.unsqueeze(-1) * beam_size
+ )
+ original_batch_idxs = original_batch_idxs[batch_idxs]
+ self.model.reorder_incremental_state(incremental_states, reorder_state)
+ encoder_outs = self.model.reorder_encoder_out(
+ encoder_outs, reorder_state
+ )
+ if reads is not None:
+ reads = reads.index_select(0, reorder_state)
+ read = read.index_select(0, reorder_state)
+ padding_length = padding_length.index_select(0, reorder_state)
+ source_length = source_length.index_select(0, reorder_state)
+ ids = ids.index_select(0, reorder_state)
+ next_selected_idx = next_selected_idx.index_select(0, reorder_state)
+ src_lens = src_lens.index_select(0, reorder_state)
+ next_selected_idxs = next_selected_idxs.index_select(
+ 0, reorder_state
+ )
+ with torch.autograd.profiler.record_function(
+ "EnsembleModel: forward_decoder"
+ ):
+ pre_lprobs, avg_attn_scores, pre_transition_probs = (
+ self.model.forward_decoder(
+ tokens[:, : step + 1],
+ encoder_outs,
+ incremental_states,
+ self.temperature,
+ )
+ )
+ lprobs = pre_lprobs[:, -1, :, :]
+
+ transition_probs = pre_transition_probs[:, -1, :, :]
+
+ if read is not None:
+ cands = self.model.models[0].decoder.cands
+ cands = cands.contiguous().view(1, -1, cands_per_token, 1)
+ cands = cands.min(src_lens.unsqueeze(1).unsqueeze(2).unsqueeze(3))
+ transition_mask = cands[:, -1, :, :] < read.unsqueeze(2)
+ transition_probs = transition_probs.masked_fill(transition_mask, 0)
+
+ next_selected_idx = (transition_probs[:, :, 1:] >= 0.5).max(
+ dim=1, keepdim=True
+ )[1]
+
+ attention = avg_attn_scores.contiguous().view(
+ avg_attn_scores.size(0),
+ avg_attn_scores.size(1),
+ avg_attn_scores.size(2) // cands_per_token,
+ cands_per_token,
+ avg_attn_scores.size(3),
+ )[:, :, -1, :, :]
+ attention = attention.gather(
+ dim=2,
+ index=next_selected_idx.unsqueeze(3).repeat(
+ 1, attention.size(1), 1, attention.size(-1)
+ ),
+ )
+
+ read = (attention > 0).sum(dim=-1, keepdim=False)[:, :, -1]
+ read = read.max(dim=1, keepdim=True)[0]
+
+ if next_selected_idxs is None:
+ next_selected_idxs = next_selected_idx[:, :, 0]
+ else:
+ next_selected_idxs = torch.cat(
+ (next_selected_idxs, next_selected_idx[:, :, 0]), dim=1
+ )
+
+ if reads is None:
+ reads = read
+ else:
+ read = read.max(reads[:, -1:])
+ reads = torch.cat((reads, read), dim=1)
+
+ lprobs = lprobs.gather(
+ dim=1, index=next_selected_idx.repeat(1, 1, lprobs.size(-1))
+ ).squeeze(1)
+
+ if self.lm_model is not None:
+ lm_out = self.lm_model(tokens[:, : step + 1])
+ probs = self.lm_model.get_normalized_probs(
+ lm_out, log_probs=True, sample=None
+ )
+ probs = probs[:, -1, :] * self.lm_weight
+ lprobs += probs
+
+ lprobs[lprobs != lprobs] = torch.tensor(-math.inf).to(lprobs)
+
+ lprobs[:, self.pad] = -math.inf # never select pad
+ lprobs[:, self.unk] -= self.unk_penalty # apply unk penalty
+
+ # handle max length constraint
+ if step >= max_len:
+ lprobs[:, : self.eos] = -math.inf
+ lprobs[:, self.eos + 1 :] = -math.inf
+
+ # handle prefix tokens (possibly with different lengths)
+ if (
+ prefix_tokens is not None
+ and step < prefix_tokens.size(1)
+ and step < max_len
+ ):
+ lprobs, tokens, scores = self._prefix_tokens(
+ step, lprobs, scores, tokens, prefix_tokens, beam_size
+ )
+ else:
+ if step < self.min_len:
+ # minimum length constraint (does not apply if using prefix_tokens)
+ lprobs[:, self.eos] = -math.inf
+
+ if self.token_indices_to_suppress is not None:
+ lprobs[:, self.token_indices_to_suppress] = -math.inf
+
+ # Record attention scores, only support avg_attn_scores is a Tensor
+ """
+ if avg_attn_scores is not None:
+ if attn is None:
+ attn = torch.empty(
+ bsz * beam_size, avg_attn_scores.size(1), max_len + 2
+ ).to(scores)
+ attn[:, :, step + 1].copy_(avg_attn_scores)
+ """
+ attn = None
+
+ scores = scores.type_as(lprobs)
+ eos_bbsz_idx = torch.empty(0).to(
+ tokens
+ ) # indices of hypothesis ending with eos (finished sentences)
+ eos_scores = torch.empty(0).to(
+ scores
+ ) # scores of hypothesis ending with eos (finished sentences)
+
+ if self.should_set_src_lengths:
+ self.search.set_src_lengths(src_lengths)
+
+ if self.repeat_ngram_blocker is not None:
+ lprobs = self.repeat_ngram_blocker(tokens, lprobs, bsz, beam_size, step)
+
+ # Shape: (batch, cand_size)
+ cand_scores, cand_indices, cand_beams = self.search.step(
+ step,
+ lprobs.view(bsz, -1, self.vocab_size),
+ scores.view(bsz, beam_size, -1)[:, :, :step],
+ tokens[:, : step + 1],
+ original_batch_idxs,
+ )
+
+ # cand_bbsz_idx contains beam indices for the top candidate
+ # hypotheses, with a range of values: [0, bsz*beam_size),
+ # and dimensions: [bsz, cand_size]
+ cand_bbsz_idx = cand_beams.add(bbsz_offsets)
+
+ # finalize hypotheses that end in eos
+ # Shape of eos_mask: (batch size, beam size)
+ eos_mask = cand_indices.eq(self.eos) & cand_scores.ne(-math.inf)
+ eos_mask[:, :beam_size][cands_to_ignore] = torch.tensor(0).to(eos_mask)
+
+ # only consider eos when it's among the top beam_size indices
+ # Now we know what beam item(s) to finish
+ # Shape: 1d list of absolute-numbered
+ eos_bbsz_idx = torch.masked_select(
+ cand_bbsz_idx[:, :beam_size], mask=eos_mask[:, :beam_size]
+ )
+
+ finalized_sents: List[int] = []
+ if eos_bbsz_idx.numel() > 0:
+ eos_scores = torch.masked_select(
+ cand_scores[:, :beam_size], mask=eos_mask[:, :beam_size]
+ )
+
+ finalized_sents = self.finalize_hypos(
+ step,
+ eos_bbsz_idx,
+ eos_scores,
+ tokens,
+ scores,
+ finalized,
+ finished,
+ beam_size,
+ attn,
+ src_lengths,
+ max_len,
+ )
+
+ for idx in finalized_sents:
+ finalized_rw[ids[idx].item()] = reads[idx]
+ finalized_extra["transition_probs"][ids[idx].item()] = (
+ pre_transition_probs[idx][:, :, -1]
+ )
+ finalized_extra["attention"][ids[idx].item()] = avg_attn_scores[
+ idx
+ ].mean(dim=0, keepdim=False)[:, : source_length[idx]]
+ finalized_extra["states_output"][ids[idx].item()] = (
+ pre_lprobs[idx]
+ .max(dim=-1, keepdim=False)[1]
+ .contiguous()
+ .view(-1)
+ )
+
+ num_remaining_sent -= len(finalized_sents)
+
+ assert num_remaining_sent >= 0
+ if num_remaining_sent == 0:
+ break
+ if self.search.stop_on_max_len and step >= max_len:
+ break
+ assert step < max_len, f"{step} < {max_len}"
+
+ # Remove finalized sentences (ones for which {beam_size}
+ # finished hypotheses have been generated) from the batch.
+ if len(finalized_sents) > 0:
+ new_bsz = bsz - len(finalized_sents)
+
+ # construct batch_idxs which holds indices of batches to keep for the next pass
+ batch_mask = torch.ones(
+ bsz, dtype=torch.bool, device=cand_indices.device
+ )
+ batch_mask[finalized_sents] = False
+ # TODO replace `nonzero(as_tuple=False)` after TorchScript supports it
+ batch_idxs = torch.arange(
+ bsz, device=cand_indices.device
+ ).masked_select(batch_mask)
+
+ # Choose the subset of the hypothesized constraints that will continue
+ self.search.prune_sentences(batch_idxs)
+
+ eos_mask = eos_mask[batch_idxs]
+ cand_beams = cand_beams[batch_idxs]
+ bbsz_offsets.resize_(new_bsz, 1)
+ cand_bbsz_idx = cand_beams.add(bbsz_offsets)
+ cand_scores = cand_scores[batch_idxs]
+ cand_indices = cand_indices[batch_idxs]
+
+ if prefix_tokens is not None:
+ prefix_tokens = prefix_tokens[batch_idxs]
+ src_lengths = src_lengths[batch_idxs]
+ cands_to_ignore = cands_to_ignore[batch_idxs]
+
+ scores = scores.view(bsz, -1)[batch_idxs].view(new_bsz * beam_size, -1)
+ tokens = tokens.view(bsz, -1)[batch_idxs].view(new_bsz * beam_size, -1)
+ if attn is not None:
+ attn = attn.view(bsz, -1)[batch_idxs].view(
+ new_bsz * beam_size, attn.size(1), -1
+ )
+ bsz = new_bsz
+ else:
+ batch_idxs = None
+
+ # Set active_mask so that values > cand_size indicate eos hypos
+ # and values < cand_size indicate candidate active hypos.
+ # After, the min values per row are the top candidate active hypos
+
+ # Rewrite the operator since the element wise or is not supported in torchscript.
+
+ eos_mask[:, :beam_size] = ~((~cands_to_ignore) & (~eos_mask[:, :beam_size]))
+ active_mask = torch.add(
+ eos_mask.type_as(cand_offsets) * cand_size,
+ cand_offsets[: eos_mask.size(1)],
+ )
+
+ # get the top beam_size active hypotheses, which are just
+ # the hypos with the smallest values in active_mask.
+ # {active_hypos} indicates which {beam_size} hypotheses
+ # from the list of {2 * beam_size} candidates were
+ # selected. Shapes: (batch size, beam size)
+ new_cands_to_ignore, active_hypos = torch.topk(
+ active_mask, k=beam_size, dim=1, largest=False
+ )
+
+ # update cands_to_ignore to ignore any finalized hypos.
+ cands_to_ignore = new_cands_to_ignore.ge(cand_size)[:, :beam_size]
+ # Make sure there is at least one active item for each sentence in the batch.
+ assert (~cands_to_ignore).any(dim=1).all()
+
+ # update cands_to_ignore to ignore any finalized hypos
+
+ # {active_bbsz_idx} denotes which beam number is continued for each new hypothesis (a beam
+ # can be selected more than once).
+ active_bbsz_idx = torch.gather(cand_bbsz_idx, dim=1, index=active_hypos)
+ active_scores = torch.gather(cand_scores, dim=1, index=active_hypos)
+
+ active_bbsz_idx = active_bbsz_idx.view(-1)
+ active_scores = active_scores.view(-1)
+
+ # copy tokens and scores for active hypotheses
+
+ # Set the tokens for each beam (can select the same row more than once)
+ tokens[:, : step + 1] = torch.index_select(
+ tokens[:, : step + 1], dim=0, index=active_bbsz_idx
+ )
+ # Select the next token for each of them
+ tokens.view(bsz, beam_size, -1)[:, :, step + 1] = torch.gather(
+ cand_indices, dim=1, index=active_hypos
+ )
+ if step > 0:
+ scores[:, :step] = torch.index_select(
+ scores[:, :step], dim=0, index=active_bbsz_idx
+ )
+ scores.view(bsz, beam_size, -1)[:, :, step] = torch.gather(
+ cand_scores, dim=1, index=active_hypos
+ )
+
+ # Update constraints based on which candidates were selected for the next beam
+ self.search.update_constraints(active_hypos)
+
+ # copy attention for active hypotheses
+ if attn is not None:
+ attn[:, :, : step + 2] = torch.index_select(
+ attn[:, :, : step + 2], dim=0, index=active_bbsz_idx
+ )
+
+ # reorder incremental state in decoder
+ reorder_state = active_bbsz_idx
+
+ # sort by score descending
+ for sent in range(len(finalized)):
+ scores = torch.tensor(
+ [float(elem["score"].item()) for elem in finalized[sent]]
+ )
+ _, sorted_scores_indices = torch.sort(scores, descending=True)
+ finalized[sent] = [finalized[sent][ssi] for ssi in sorted_scores_indices]
+ finalized[sent] = torch.jit.annotate(
+ List[Dict[str, Tensor]], finalized[sent]
+ )
+ return finalized, finalized_rw, finalized_src_length, finalized_extra
+
+ def _prefix_tokens(
+ self, step: int, lprobs, scores, tokens, prefix_tokens, beam_size: int
+ ):
+ """Handle prefix tokens"""
+ prefix_toks = prefix_tokens[:, step].unsqueeze(-1).repeat(1, beam_size).view(-1)
+ prefix_lprobs = lprobs.gather(-1, prefix_toks.unsqueeze(-1))
+ prefix_mask = prefix_toks.ne(self.pad)
+ lprobs[prefix_mask] = torch.tensor(-math.inf).to(lprobs)
+ lprobs[prefix_mask] = lprobs[prefix_mask].scatter(
+ -1, prefix_toks[prefix_mask].unsqueeze(-1), prefix_lprobs[prefix_mask]
+ )
+ # if prefix includes eos, then we should make sure tokens and
+ # scores are the same across all beams
+ eos_mask = prefix_toks.eq(self.eos)
+ if eos_mask.any():
+ # validate that the first beam matches the prefix
+ first_beam = tokens[eos_mask].view(-1, beam_size, tokens.size(-1))[
+ :, 0, 1 : step + 1
+ ]
+ eos_mask_batch_dim = eos_mask.view(-1, beam_size)[:, 0]
+ target_prefix = prefix_tokens[eos_mask_batch_dim][:, :step]
+ assert (first_beam == target_prefix).all()
+
+ # copy tokens, scores and lprobs from the first beam to all beams
+ tokens = self.replicate_first_beam(tokens, eos_mask_batch_dim, beam_size)
+ scores = self.replicate_first_beam(scores, eos_mask_batch_dim, beam_size)
+ lprobs = self.replicate_first_beam(lprobs, eos_mask_batch_dim, beam_size)
+ return lprobs, tokens, scores
+
+ def replicate_first_beam(self, tensor, mask, beam_size: int):
+ tensor = tensor.view(-1, beam_size, tensor.size(-1))
+ tensor[mask] = tensor[mask][:, :1, :]
+ return tensor.view(-1, tensor.size(-1))
+
+ def finalize_hypos(
+ self,
+ step: int,
+ bbsz_idx,
+ eos_scores,
+ tokens,
+ scores,
+ finalized: List[List[Dict[str, Tensor]]],
+ finished: List[bool],
+ beam_size: int,
+ attn: Optional[Tensor],
+ src_lengths,
+ max_len: int,
+ ):
+ """Finalize hypothesis, store finalized information in `finalized`, and change `finished` accordingly.
+ A sentence is finalized when {beam_size} finished items have been collected for it.
+
+ Returns number of sentences (not beam items) being finalized.
+ These will be removed from the batch and not processed further.
+ Args:
+ bbsz_idx (Tensor):
+ """
+ assert bbsz_idx.numel() == eos_scores.numel()
+
+ # clone relevant token and attention tensors.
+ # tokens is (batch * beam, max_len). So the index_select
+ # gets the newly EOS rows, then selects cols 1..{step + 2}
+ tokens_clone = tokens.index_select(0, bbsz_idx)[
+ :, 1 : step + 2
+ ] # skip the first index, which is EOS
+
+ tokens_clone[:, step] = self.eos
+ attn_clone = (
+ attn.index_select(0, bbsz_idx)[:, :, 1 : step + 2]
+ if attn is not None
+ else None
+ )
+
+ # compute scores per token position
+ pos_scores = scores.index_select(0, bbsz_idx)[:, : step + 1]
+ pos_scores[:, step] = eos_scores
+ # convert from cumulative to per-position scores
+ pos_scores[:, 1:] = pos_scores[:, 1:] - pos_scores[:, :-1]
+
+ # normalize sentence-level scores
+ if self.normalize_scores:
+ eos_scores /= (step + 1) ** self.len_penalty
+
+ # cum_unfin records which sentences in the batch are finished.
+ # It helps match indexing between (a) the original sentences
+ # in the batch and (b) the current, possibly-reduced set of
+ # sentences.
+ cum_unfin: List[int] = []
+ prev = 0
+ for f in finished:
+ if f:
+ prev += 1
+ else:
+ cum_unfin.append(prev)
+ cum_fin_tensor = torch.tensor(cum_unfin, dtype=torch.int).to(bbsz_idx)
+
+ unfin_idx = torch.div(bbsz_idx, beam_size, rounding_mode="trunc")
+ sent = unfin_idx + torch.index_select(cum_fin_tensor, 0, unfin_idx)
+
+ # Create a set of "{sent}{unfin_idx}", where
+ # "unfin_idx" is the index in the current (possibly reduced)
+ # list of sentences, and "sent" is the index in the original,
+ # unreduced batch
+ # For every finished beam item
+ # sentence index in the current (possibly reduced) batch
+ seen = (sent << 32) + unfin_idx
+ unique_seen: List[int] = torch.unique(seen).tolist()
+
+ if self.match_source_len:
+ condition = step > torch.index_select(src_lengths, 0, unfin_idx)
+ eos_scores = torch.where(condition, torch.tensor(-math.inf), eos_scores)
+ sent_list: List[int] = sent.tolist()
+ for i in range(bbsz_idx.size()[0]):
+ # An input sentence (among those in a batch) is finished when
+ # beam_size hypotheses have been collected for it
+ if len(finalized[sent_list[i]]) < beam_size:
+ if attn_clone is not None:
+ # remove padding tokens from attn scores
+ hypo_attn = attn_clone[i]
+ else:
+ hypo_attn = torch.empty(0)
+
+ finalized[sent_list[i]].append(
+ {
+ "tokens": tokens_clone[i],
+ "score": eos_scores[i],
+ "attention": hypo_attn, # src_len x tgt_len
+ "alignment": torch.empty(0),
+ "positional_scores": pos_scores[i],
+ }
+ )
+
+ newly_finished: List[int] = []
+ for unique_s in unique_seen:
+ # check termination conditions for this sentence
+ unique_sent: int = unique_s >> 32
+ unique_unfin_idx: int = unique_s - (unique_sent << 32)
+
+ if not finished[unique_sent] and self.is_finished(
+ step, unique_unfin_idx, max_len, len(finalized[unique_sent]), beam_size
+ ):
+ finished[unique_sent] = True
+ newly_finished.append(unique_unfin_idx)
+
+ return newly_finished
+
+ def is_finished(
+ self,
+ step: int,
+ unfin_idx: int,
+ max_len: int,
+ finalized_sent_len: int,
+ beam_size: int,
+ ):
+ """
+ Check whether decoding for a sentence is finished, which
+ occurs when the list of finalized sentences has reached the
+ beam size, or when we reach the maximum length.
+ """
+ assert finalized_sent_len <= beam_size
+ if finalized_sent_len == beam_size or step == max_len:
+ return True
+ return False
+
+
+class EnsembleModel(nn.Module):
+ """A wrapper around an ensemble of models."""
+
+ def __init__(self, models):
+ super().__init__()
+ self.models_size = len(models)
+ # method '__len__' is not supported in ModuleList for torch script
+ self.single_model = models[0]
+ self.models = nn.ModuleList(models)
+
+ self.has_incremental: bool = False
+ if all(
+ hasattr(m, "decoder") and isinstance(m.decoder, FairseqIncrementalDecoder)
+ for m in models
+ ):
+ self.has_incremental = True
+
+ def forward(self):
+ pass
+
+ def has_encoder(self):
+ return hasattr(self.single_model, "encoder")
+
+ def has_incremental_states(self):
+ return self.has_incremental
+
+ def max_decoder_positions(self):
+ return min(
+ [
+ m.max_decoder_positions()
+ for m in self.models
+ if hasattr(m, "max_decoder_positions")
+ ]
+ + [sys.maxsize]
+ )
+
+ def set_decoder_beam_size(self, beam_size):
+ """Set beam size for efficient beamable enc-dec attention."""
+ if beam_size > 1:
+ for model in self.models:
+ if hasattr(model, "set_beam_size"):
+ model.set_beam_size(beam_size)
+
+ @torch.jit.export
+ def forward_encoder(self, net_input: Dict[str, Tensor]):
+ if not self.has_encoder():
+ return None
+ return [model.encoder.forward_torchscript(net_input) for model in self.models]
+
+ @torch.jit.export
+ def forward_decoder(
+ self,
+ tokens,
+ encoder_outs: List[Dict[str, List[Tensor]]],
+ incremental_states: List[Dict[str, Dict[str, Optional[Tensor]]]],
+ temperature: float = 1.0,
+ ):
+ log_probs = []
+ avg_attn: Optional[Tensor] = None
+ encoder_out: Optional[Dict[str, List[Tensor]]] = None
+ for i, model in enumerate(self.models):
+ if self.has_encoder():
+ encoder_out = encoder_outs[i]
+ # decode each model
+ # if self.has_incremental_states():
+ if False:
+ decoder_out = model.decoder.forward(
+ tokens,
+ encoder_out=encoder_out,
+ incremental_state=incremental_states[i],
+ )
+ else:
+ if hasattr(model, "decoder"):
+ decoder_out = model.decoder.forward(tokens, encoder_out=encoder_out)
+ else:
+ decoder_out = model.forward(tokens)
+
+ attn: Optional[Tensor] = None
+ decoder_len = len(decoder_out)
+ attn = decoder_out[-1]["attention"]
+ decoder_out_tuple = (
+ decoder_out[0].div_(temperature),
+ None if decoder_len <= 1 else decoder_out[1],
+ )
+
+ probs = model.get_normalized_probs(
+ decoder_out_tuple, log_probs=True, sample=None
+ )
+ # This should be attended
+ # probs = probs[:, -1, :]
+ transition_probs = model.get_normalized_probs(
+ (decoder_out[1],), log_probs=False, sample=None
+ )
+
+ if self.models_size == 1:
+ return probs, attn, transition_probs
+
+ log_probs.append(probs)
+ if attn is not None:
+ if avg_attn is None:
+ avg_attn = attn
+ else:
+ avg_attn.add_(attn)
+
+ avg_probs = torch.logsumexp(torch.stack(log_probs, dim=0), dim=0) - math.log(
+ self.models_size
+ )
+
+ if avg_attn is not None:
+ avg_attn.div_(self.models_size)
+ return avg_probs, avg_attn
+
+ @torch.jit.export
+ def reorder_encoder_out(
+ self, encoder_outs: Optional[List[Dict[str, List[Tensor]]]], new_order
+ ):
+ """
+ Reorder encoder output according to *new_order*.
+
+ Args:
+ encoder_out: output from the ``forward()`` method
+ new_order (LongTensor): desired order
+
+ Returns:
+ *encoder_out* rearranged according to *new_order*
+ """
+ new_outs: List[Dict[str, List[Tensor]]] = []
+ if not self.has_encoder():
+ return new_outs
+ for i, model in enumerate(self.models):
+ assert encoder_outs is not None
+ new_outs.append(
+ model.encoder.reorder_encoder_out(encoder_outs[i], new_order)
+ )
+ return new_outs
+
+ @torch.jit.export
+ def reorder_incremental_state(
+ self,
+ incremental_states: List[Dict[str, Dict[str, Optional[Tensor]]]],
+ new_order,
+ ):
+ if not self.has_incremental_states():
+ return
+ for i, model in enumerate(self.models):
+ model.decoder.reorder_incremental_state_scripting(
+ incremental_states[i], new_order
+ )
+
+
+class SequenceGeneratorWithAlignment(SequenceGenerator):
+ def __init__(
+ self, models, tgt_dict, left_pad_target=False, print_alignment="hard", **kwargs
+ ):
+ """Generates translations of a given source sentence.
+
+ Produces alignments following "Jointly Learning to Align and
+ Translate with Transformer Models" (Garg et al., EMNLP 2019).
+
+ Args:
+ left_pad_target (bool, optional): Whether or not the
+ hypothesis should be left padded or not when they are
+ teacher forced for generating alignments.
+ """
+ super().__init__(EnsembleModelWithAlignment(models), tgt_dict, **kwargs)
+ self.left_pad_target = left_pad_target
+
+ if print_alignment == "hard":
+ self.extract_alignment = utils.extract_hard_alignment
+ elif print_alignment == "soft":
+ self.extract_alignment = utils.extract_soft_alignment
+
+ @torch.no_grad()
+ def generate(self, models, sample, **kwargs):
+ finalized = super()._generate(sample, **kwargs)
+
+ src_tokens = sample["net_input"]["src_tokens"]
+ bsz = src_tokens.shape[0]
+ beam_size = self.beam_size
+ (
+ src_tokens,
+ src_lengths,
+ prev_output_tokens,
+ tgt_tokens,
+ ) = self._prepare_batch_for_alignment(sample, finalized)
+ if any(getattr(m, "full_context_alignment", False) for m in self.model.models):
+ attn = self.model.forward_align(src_tokens, src_lengths, prev_output_tokens)
+ else:
+ attn = [
+ finalized[i // beam_size][i % beam_size]["attention"].transpose(1, 0)
+ for i in range(bsz * beam_size)
+ ]
+
+ if src_tokens.device != "cpu":
+ src_tokens = src_tokens.to("cpu")
+ tgt_tokens = tgt_tokens.to("cpu")
+ attn = [i.to("cpu") for i in attn]
+
+ # Process the attn matrix to extract hard alignments.
+ for i in range(bsz * beam_size):
+ alignment = self.extract_alignment(
+ attn[i], src_tokens[i], tgt_tokens[i], self.pad, self.eos
+ )
+ finalized[i // beam_size][i % beam_size]["alignment"] = alignment
+ return finalized
+
+ def _prepare_batch_for_alignment(self, sample, hypothesis):
+ src_tokens = sample["net_input"]["src_tokens"]
+ bsz = src_tokens.shape[0]
+ src_tokens = (
+ src_tokens[:, None, :]
+ .expand(-1, self.beam_size, -1)
+ .contiguous()
+ .view(bsz * self.beam_size, -1)
+ )
+ src_lengths = sample["net_input"]["src_lengths"]
+ src_lengths = (
+ src_lengths[:, None]
+ .expand(-1, self.beam_size)
+ .contiguous()
+ .view(bsz * self.beam_size)
+ )
+ prev_output_tokens = data_utils.collate_tokens(
+ [beam["tokens"] for example in hypothesis for beam in example],
+ self.pad,
+ self.eos,
+ self.left_pad_target,
+ move_eos_to_beginning=True,
+ )
+ tgt_tokens = data_utils.collate_tokens(
+ [beam["tokens"] for example in hypothesis for beam in example],
+ self.pad,
+ self.eos,
+ self.left_pad_target,
+ move_eos_to_beginning=False,
+ )
+ return src_tokens, src_lengths, prev_output_tokens, tgt_tokens
+
+
+class EnsembleModelWithAlignment(EnsembleModel):
+ """A wrapper around an ensemble of models."""
+
+ def __init__(self, models):
+ super().__init__(models)
+
+ def forward_align(self, src_tokens, src_lengths, prev_output_tokens):
+ avg_attn = None
+ for model in self.models:
+ decoder_out = model(src_tokens, src_lengths, prev_output_tokens)
+ attn = decoder_out[1]["attn"][0]
+ if avg_attn is None:
+ avg_attn = attn
+ else:
+ avg_attn.add_(attn)
+ if len(self.models) > 1:
+ avg_attn.div_(len(self.models))
+ return avg_attn
diff --git a/agent/mt/hmt_sim_sequence_generator.py b/agent/mt/hmt_sim_sequence_generator.py
new file mode 100644
index 0000000000000000000000000000000000000000..313f432b63cb9749cdaa2902b12de75b11a32841
--- /dev/null
+++ b/agent/mt/hmt_sim_sequence_generator.py
@@ -0,0 +1,1117 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import math
+import sys
+from typing import Dict, List, Optional
+
+import torch
+import torch.nn as nn
+from torch import Tensor
+
+from fairseq import search, utils
+from fairseq.data import data_utils
+from fairseq.models import FairseqIncrementalDecoder
+from fairseq.ngram_repeat_block import NGramRepeatBlock
+
+
+class HmtSimSequenceGenerator(nn.Module):
+ def __init__(
+ self,
+ models,
+ tgt_dict,
+ beam_size=1,
+ max_len_a=0,
+ max_len_b=200,
+ max_len=0,
+ min_len=1,
+ normalize_scores=True,
+ len_penalty=1.0,
+ unk_penalty=0.0,
+ temperature=1.0,
+ match_source_len=False,
+ no_repeat_ngram_size=0,
+ search_strategy=None,
+ eos=None,
+ symbols_to_strip_from_output=None,
+ lm_model=None,
+ lm_weight=1.0,
+ tokens_to_suppress=(),
+ ):
+ """Generates translations of a given source sentence.
+
+ Args:
+ models (List[~fairseq.models.FairseqModel]): ensemble of models,
+ currently support fairseq.models.TransformerModel for scripting
+ beam_size (int, optional): beam width (default: 1)
+ max_len_a/b (int, optional): generate sequences of maximum length
+ ax + b, where x is the source length
+ max_len (int, optional): the maximum length of the generated output
+ (not including end-of-sentence)
+ min_len (int, optional): the minimum length of the generated output
+ (not including end-of-sentence)
+ normalize_scores (bool, optional): normalize scores by the length
+ of the output (default: True)
+ len_penalty (float, optional): length penalty, where <1.0 favors
+ shorter, >1.0 favors longer sentences (default: 1.0)
+ unk_penalty (float, optional): unknown word penalty, where <0
+ produces more unks, >0 produces fewer (default: 0.0)
+ temperature (float, optional): temperature, where values
+ >1.0 produce more uniform samples and values <1.0 produce
+ sharper samples (default: 1.0)
+ match_source_len (bool, optional): outputs should match the source
+ length (default: False)
+ """
+ super().__init__()
+ if isinstance(models, EnsembleModel):
+ self.model = models
+ else:
+ self.model = EnsembleModel(models)
+ self.tgt_dict = tgt_dict
+ self.pad = tgt_dict.pad()
+ self.unk = tgt_dict.unk()
+ self.eos = tgt_dict.eos() if eos is None else eos
+ self.symbols_to_strip_from_output = (
+ symbols_to_strip_from_output.union({self.eos})
+ if symbols_to_strip_from_output is not None
+ else {self.eos}
+ )
+
+ self.token_indices_to_suppress: Optional[Tensor] = None
+ token_indices_to_suppress = []
+ for token_string in tokens_to_suppress:
+ token_index = tgt_dict.index(token_string)
+ assert token_index != self.unk
+ token_indices_to_suppress.append(token_index)
+ if len(token_indices_to_suppress) > 0:
+ self.token_indices_to_suppress = torch.Tensor(
+ token_indices_to_suppress
+ ).long()
+
+ self.vocab_size = len(tgt_dict)
+ self.beam_size = beam_size
+ # the max beam size is the dictionary size - 1, since we never select pad
+ self.beam_size = min(beam_size, self.vocab_size - 1)
+ self.model.set_decoder_beam_size(self.beam_size)
+ self.max_len_a = max_len_a
+ self.max_len_b = max_len_b
+ self.min_len = 0
+ self.max_len = max_len or self.model.max_decoder_positions()
+
+ self.normalize_scores = normalize_scores
+ self.len_penalty = len_penalty
+ self.unk_penalty = unk_penalty
+ self.temperature = temperature
+ self.match_source_len = match_source_len
+
+ if no_repeat_ngram_size > 0:
+ self.repeat_ngram_blocker = NGramRepeatBlock(no_repeat_ngram_size)
+ else:
+ self.repeat_ngram_blocker = None
+
+ assert temperature > 0, "--temperature must be greater than 0"
+
+ self.search = (
+ search.BeamSearch(tgt_dict) if search_strategy is None else search_strategy
+ )
+ # We only need to set src_lengths in LengthConstrainedBeamSearch.
+ # As a module attribute, setting it would break in multithread
+ # settings when the model is shared.
+ self.should_set_src_lengths = (
+ hasattr(self.search, "needs_src_lengths") and self.search.needs_src_lengths
+ )
+
+ self.model.eval()
+
+ self.lm_model = lm_model
+ self.lm_weight = lm_weight
+ if self.lm_model is not None:
+ self.lm_model.eval()
+
+ def cuda(self):
+ self.model.cuda()
+ return self
+
+ @torch.no_grad()
+ def forward(
+ self,
+ sample: Dict[str, Dict[str, Tensor]],
+ prefix_tokens: Optional[Tensor] = None,
+ bos_token: Optional[int] = None,
+ ):
+ """Generate a batch of translations.
+
+ Args:
+ sample (dict): batch
+ prefix_tokens (torch.LongTensor, optional): force decoder to begin
+ with these tokens
+ bos_token (int, optional): beginning of sentence token
+ (default: self.eos)
+ """
+ return self._generate(sample, prefix_tokens, bos_token=bos_token)
+
+ # TODO(myleott): unused, deprecate after pytorch-translate migration
+ def generate_batched_itr(self, data_itr, beam_size=None, cuda=False, timer=None):
+ """Iterate over a batched dataset and yield individual translations.
+ Args:
+ cuda (bool, optional): use GPU for generation
+ timer (StopwatchMeter, optional): time generations
+ """
+ for sample in data_itr:
+ s = utils.move_to_cuda(sample) if cuda else sample
+ if "net_input" not in s:
+ continue
+ input = s["net_input"]
+ # model.forward normally channels prev_output_tokens into the decoder
+ # separately, but SequenceGenerator directly calls model.encoder
+ encoder_input = {
+ k: v for k, v in input.items() if k != "prev_output_tokens"
+ }
+ if timer is not None:
+ timer.start()
+ with torch.no_grad():
+ hypos = self.generate(encoder_input)
+ if timer is not None:
+ timer.stop(sum(len(h[0]["tokens"]) for h in hypos))
+ for i, id in enumerate(s["id"].data):
+ # remove padding
+ src = utils.strip_pad(input["src_tokens"].data[i, :], self.pad)
+ ref = (
+ utils.strip_pad(s["target"].data[i, :], self.pad)
+ if s["target"] is not None
+ else None
+ )
+ yield id, src, ref, hypos[i]
+
+ @torch.no_grad()
+ def generate(
+ self, models, sample: Dict[str, Dict[str, Tensor]], **kwargs
+ ) -> List[List[Dict[str, Tensor]]]:
+ """Generate translations. Match the api of other fairseq generators.
+
+ Args:
+ models (List[~fairseq.models.FairseqModel]): ensemble of models
+ sample (dict): batch
+ prefix_tokens (torch.LongTensor, optional): force decoder to begin
+ with these tokens
+ constraints (torch.LongTensor, optional): force decoder to include
+ the list of constraints
+ bos_token (int, optional): beginning of sentence token
+ (default: self.eos)
+ """
+ return self._generate(sample, **kwargs)
+
+ def _generate(
+ self,
+ sample: Dict[str, Dict[str, Tensor]],
+ prefix_tokens: Optional[Tensor] = None,
+ constraints: Optional[Tensor] = None,
+ bos_token: Optional[int] = None,
+ src_finished=False,
+ pre_src_prefix=None,
+ ):
+ incremental_states = torch.jit.annotate(
+ List[Dict[str, Dict[str, Optional[Tensor]]]],
+ [
+ torch.jit.annotate(Dict[str, Dict[str, Optional[Tensor]]], {})
+ for i in range(self.model.models_size)
+ ],
+ )
+ net_input = sample["net_input"]
+
+ if "src_tokens" in net_input:
+ src_tokens = net_input["src_tokens"]
+ # length of the source text being the character length except EndOfSentence and pad
+ src_lengths = (
+ (src_tokens.ne(self.eos) & src_tokens.ne(self.pad)).long().sum(dim=1)
+ )
+ elif "source" in net_input:
+ src_tokens = net_input["source"]
+ src_lengths = (
+ net_input["padding_mask"].size(-1) - net_input["padding_mask"].sum(-1)
+ if net_input["padding_mask"] is not None
+ else torch.tensor(src_tokens.size(-1)).to(src_tokens)
+ )
+ elif "features" in net_input:
+ src_tokens = net_input["features"]
+ src_lengths = (
+ net_input["padding_mask"].size(-1) - net_input["padding_mask"].sum(-1)
+ if net_input["padding_mask"] is not None
+ else torch.tensor(src_tokens.size(-1)).to(src_tokens)
+ )
+ else:
+ raise Exception(
+ "expected src_tokens or source in net input. input keys: "
+ + str(net_input.keys())
+ )
+
+ # bsz: total number of sentences in beam
+ # Note that src_tokens may have more than 2 dimensions (i.e. audio features)
+ bsz, src_len = src_tokens.size()[:2]
+ beam_size = self.beam_size
+
+ if constraints is not None and not self.search.supports_constraints:
+ raise NotImplementedError(
+ "Target-side constraints were provided, but search method doesn't support them"
+ )
+
+ # Initialize constraints, when active
+ self.search.init_constraints(constraints, beam_size)
+
+ start = prefix_tokens.size(-1) if prefix_tokens is not None else 0
+
+ max_len: int = -1
+ if self.match_source_len:
+ max_len = src_lengths.max().item()
+ else:
+ max_len = min(
+ int(self.max_len_a * src_len + self.max_len_b),
+ self.max_len - 1,
+ )
+ assert (
+ self.min_len <= max_len
+ ), "min_len cannot be larger than max_len, please adjust these!"
+ # compute the encoder output for each beam
+ """
+ with torch.autograd.profiler.record_function("EnsembleModel: forward_encoder"):
+ encoder_outs = self.model.forward_encoder(net_input)
+
+ # placeholder of indices for bsz * beam_size to hold tokens and accumulative scores
+ new_order = torch.arange(bsz).view(-1, 1).repeat(1, beam_size).view(-1)
+ new_order = new_order.to(src_tokens.device).long()
+ encoder_outs = self.model.reorder_encoder_out(encoder_outs, new_order)
+ # ensure encoder_outs is a List.
+ assert encoder_outs is not None
+ """
+ # initialize buffers
+ scores = (
+ torch.zeros(bsz * beam_size, max_len + 1).to(src_tokens).float()
+ ) # +1 for eos; pad is never chosen for scoring
+ attn: Optional[Tensor] = None
+
+ if prefix_tokens is None:
+ tokens = (
+ torch.zeros(bsz * beam_size, max_len + 2, device=src_tokens.device)
+ .long()
+ .fill_(self.pad)
+ ) # +2 for eos and pad
+ tokens[:, 0] = self.eos if bos_token is None else bos_token
+ else:
+ tokens = torch.cat(
+ (
+ torch.full(
+ (bsz, 1), self.eos, dtype=torch.long, device=src_tokens.device
+ ),
+ prefix_tokens,
+ ),
+ dim=-1,
+ )
+ tokens = tokens.repeat(beam_size, 1)
+ if tokens.size(-1) < max_len + 2:
+ tokens = torch.cat(
+ (
+ tokens,
+ torch.zeros(
+ tokens.size(0),
+ max_len + 2 - tokens.size(-1),
+ device=src_tokens.device,
+ ).long(),
+ ),
+ dim=-1,
+ )
+
+ # A list that indicates candidates that should be ignored.
+ # For example, suppose we're sampling and have already finalized 2/5
+ # samples. Then cands_to_ignore would mark 2 positions as being ignored,
+ # so that we only finalize the remaining 3 samples.
+ cands_to_ignore = (
+ torch.zeros(bsz, beam_size).to(src_tokens).eq(-1)
+ ) # forward and backward-compatible False mask
+
+ # list of completed sentences
+ finalized = torch.jit.annotate(
+ List[List[Dict[str, Tensor]]],
+ [torch.jit.annotate(List[Dict[str, Tensor]], []) for i in range(bsz)],
+ ) # contains lists of dictionaries of infomation about the hypothesis being finalized at each step
+
+ # a boolean array indicating if the sentence at the index is finished or not
+ finished = [False for i in range(bsz)]
+ num_remaining_sent = bsz # number of sentences remaining
+
+ # number of candidate hypos per step
+ cand_size = 2 * beam_size # 2 x beam size in case half are EOS
+
+ # offset arrays for converting between different indexing schemes
+ bbsz_offsets = (
+ (torch.arange(0, bsz) * beam_size)
+ .unsqueeze(1)
+ .type_as(tokens)
+ .to(src_tokens.device)
+ )
+ cand_offsets = torch.arange(0, cand_size).type_as(tokens).to(src_tokens.device)
+
+ reorder_state: Optional[Tensor] = None
+ batch_idxs: Optional[Tensor] = None
+
+ original_batch_idxs: Optional[Tensor] = None
+ if "id" in sample and isinstance(sample["id"], Tensor):
+ original_batch_idxs = sample["id"]
+ else:
+ original_batch_idxs = torch.arange(0, bsz).type_as(tokens)
+
+ first_read = self.model.models[0].decoder.first_read
+ cands_per_token = self.model.models[0].decoder.cands_per_token
+ previous_selected_idx = None
+
+ ds = []
+ src_lens = sample["net_input"]["src_lengths"]
+ reads = None
+ read = None
+ source_length = sample["net_input"]["src_lengths"]
+ padding_length = (
+ sample["net_input"]["src_tokens"].size(1)
+ - sample["net_input"]["src_lengths"]
+ )
+ finalized_rw = {}
+ finalized_extra = {}
+ finalized_extra["transition_probs"] = {}
+ finalized_extra["attention"] = {}
+ finalized_extra["states_output"] = {}
+ finalized_src_length = sample["net_input"]["src_lengths"]
+
+ ids = sample["id"]
+ if pre_src_prefix is None:
+ num_read = 1
+ else:
+ num_read = pre_src_prefix.size(-1) + 1
+
+ # partial_net_input = {
+ # "src_tokens": net_input["src_tokens"][:, :num_read],
+ # "src_lengths": net_input["src_lengths"].clamp(0, num_read),
+ # }
+ encoder_outs = self.model.forward_encoder(net_input)
+
+ step = start
+ return_read_flag = False
+ # for step in range(max_len + 1): # one extra step for EOS marker
+ while step < max_len + 1:
+ # reorder decoder internal states based on the prev choice of beams
+ partial_net_input = {
+ "src_tokens": net_input["src_tokens"][:, :num_read],
+ "src_lengths": net_input["src_lengths"].clamp(0, num_read),
+ }
+ # new_encoder_outs = self.model.forward_encoder(partial_net_input)
+
+ new_encoder_outs = self.bulid_partial_encoder_out(encoder_outs, num_read)
+
+ finish_read = partial_net_input["src_tokens"].size(-1) == net_input[
+ "src_tokens"
+ ].size(-1)
+
+ if reorder_state is not None:
+ if batch_idxs is not None:
+ # update beam indices to take into account removed sentences
+ corr = batch_idxs - torch.arange(batch_idxs.numel()).type_as(
+ batch_idxs
+ )
+ reorder_state.view(-1, beam_size).add_(
+ corr.unsqueeze(-1) * beam_size
+ )
+ original_batch_idxs = original_batch_idxs[batch_idxs]
+ self.model.reorder_incremental_state(incremental_states, reorder_state)
+ encoder_outs = self.model.reorder_encoder_out(
+ encoder_outs, reorder_state
+ )
+ if reads is not None:
+ reads = reads.index_select(0, reorder_state)
+ read = read.index_select(0, reorder_state)
+ padding_length = padding_length.index_select(0, reorder_state)
+ source_length = source_length.index_select(0, reorder_state)
+ ids = ids.index_select(0, reorder_state)
+ next_selected_idx = next_selected_idx.index_select(0, reorder_state)
+ if num_read - first_read - step < 0 and not finish_read:
+ num_read += 1
+ continue
+
+ if num_read - first_read - step < 0 and finish_read and not src_finished:
+ return_read_flag = True
+
+ with torch.autograd.profiler.record_function(
+ "EnsembleModel: forward_decoder"
+ ):
+ pre_lprobs, avg_attn_scores, pre_transition_probs = (
+ self.model.forward_decoder(
+ tokens[:, : step + 1],
+ new_encoder_outs,
+ incremental_states,
+ self.temperature,
+ )
+ )
+
+ lprobs = pre_lprobs[:, -1, :, :]
+ transition_probs = pre_transition_probs[:, -1, :, :]
+
+ if read is not None:
+ cands = self.model.models[0].decoder.cands
+ cands = cands.contiguous().view(1, -1, cands_per_token, 1)
+ # cands=cands.min(src_lens.unsqueeze(1).unsqueeze(2).unsqueeze(3))
+ cands = cands.clamp(0, new_encoder_outs[0]["encoder_out"][0].size(0))
+ transition_mask = cands[:, -1, :, :] < read.unsqueeze(2)
+ transition_probs = transition_probs.masked_fill(transition_mask, 0)
+
+ # not reach high boundary & not finish read & unselect
+ if (
+ num_read - first_read - step < cands_per_token
+ and not finish_read
+ and transition_probs[0][num_read - first_read - step][1] < 0.5
+ ):
+ num_read += 1
+ continue
+ try:
+ if (
+ num_read - first_read - step < cands_per_token
+ and not src_finished
+ and not return_read_flag
+ and finish_read
+ and transition_probs[0][num_read - first_read - step][1] < 0.5
+ ):
+ return_read_flag = True
+ except:
+ import pdb
+
+ pdb.set_trace()
+
+ next_selected_idx = (transition_probs[:, :, 1:] >= 0.5).max(
+ dim=1, keepdim=True
+ )[1]
+
+ read = torch.full([1, 1], num_read, device=avg_attn_scores.device)
+ if reads is None:
+ reads = read
+ else:
+ read = read.max(reads[:, -1:])
+ reads = torch.cat((reads, read), dim=1)
+
+ lprobs = lprobs.gather(
+ dim=1, index=next_selected_idx.repeat(1, 1, lprobs.size(-1))
+ ).squeeze(1)
+
+ if return_read_flag:
+ lprobs[:, : self.eos] = -math.inf
+ lprobs[:, self.eos + 1 :] = -math.inf
+
+ if self.lm_model is not None:
+ lm_out = self.lm_model(tokens[:, : step + 1])
+ probs = self.lm_model.get_normalized_probs(
+ lm_out, log_probs=True, sample=None
+ )
+ probs = probs[:, -1, :] * self.lm_weight
+ lprobs += probs
+
+ lprobs[lprobs != lprobs] = torch.tensor(-math.inf).to(lprobs)
+
+ lprobs[:, self.pad] = -math.inf # never select pad
+ lprobs[:, self.unk] -= self.unk_penalty # apply unk penalty
+
+ # handle max length constraint
+ if step >= max_len:
+ lprobs[:, : self.eos] = -math.inf
+ lprobs[:, self.eos + 1 :] = -math.inf
+
+ # handle prefix tokens (possibly with different lengths)
+ if (
+ prefix_tokens is not None
+ and step < prefix_tokens.size(1)
+ and step < max_len
+ ):
+ lprobs, tokens, scores = self._prefix_tokens(
+ step, lprobs, scores, tokens, prefix_tokens, beam_size
+ )
+ else:
+ if step < self.min_len:
+ # minimum length constraint (does not apply if using prefix_tokens)
+ lprobs[:, self.eos] = -math.inf
+
+ if self.token_indices_to_suppress is not None:
+ lprobs[:, self.token_indices_to_suppress] = -math.inf
+
+ # Record attention scores, only support avg_attn_scores is a Tensor
+ """
+ if avg_attn_scores is not None:
+ if attn is None:
+ attn = torch.empty(
+ bsz * beam_size, avg_attn_scores.size(1), max_len + 2
+ ).to(scores)
+ attn[:, :, step + 1].copy_(avg_attn_scores)
+ """
+ attn = None
+
+ scores = scores.type_as(lprobs)
+ eos_bbsz_idx = torch.empty(0).to(
+ tokens
+ ) # indices of hypothesis ending with eos (finished sentences)
+ eos_scores = torch.empty(0).to(
+ scores
+ ) # scores of hypothesis ending with eos (finished sentences)
+
+ if self.should_set_src_lengths:
+ self.search.set_src_lengths(src_lengths)
+
+ if self.repeat_ngram_blocker is not None:
+ lprobs = self.repeat_ngram_blocker(tokens, lprobs, bsz, beam_size, step)
+
+ # Shape: (batch, cand_size)
+ cand_scores, cand_indices, cand_beams = self.search.step(
+ step,
+ lprobs.view(bsz, -1, self.vocab_size),
+ scores.view(bsz, beam_size, -1)[:, :, :step],
+ tokens[:, : step + 1],
+ original_batch_idxs,
+ )
+
+ # cand_bbsz_idx contains beam indices for the top candidate
+ # hypotheses, with a range of values: [0, bsz*beam_size),
+ # and dimensions: [bsz, cand_size]
+ cand_bbsz_idx = cand_beams.add(bbsz_offsets)
+
+ # finalize hypotheses that end in eos
+ # Shape of eos_mask: (batch size, beam size)
+ eos_mask = cand_indices.eq(self.eos) & cand_scores.ne(-math.inf)
+ eos_mask[:, :beam_size][cands_to_ignore] = torch.tensor(0).to(eos_mask)
+
+ # only consider eos when it's among the top beam_size indices
+ # Now we know what beam item(s) to finish
+ # Shape: 1d list of absolute-numbered
+ eos_bbsz_idx = torch.masked_select(
+ cand_bbsz_idx[:, :beam_size], mask=eos_mask[:, :beam_size]
+ )
+
+ finalized_sents: List[int] = []
+ if eos_bbsz_idx.numel() > 0:
+ eos_scores = torch.masked_select(
+ cand_scores[:, :beam_size], mask=eos_mask[:, :beam_size]
+ )
+
+ finalized_sents = self.finalize_hypos(
+ step,
+ eos_bbsz_idx,
+ eos_scores,
+ tokens,
+ scores,
+ finalized,
+ finished,
+ beam_size,
+ attn,
+ src_lengths,
+ max_len,
+ )
+
+ for idx in finalized_sents:
+ finalized_rw[ids[idx].item()] = reads[idx]
+ finalized_extra["transition_probs"][ids[idx].item()] = (
+ pre_transition_probs[idx][:, :, -1]
+ )
+ finalized_extra["attention"][ids[idx].item()] = avg_attn_scores[
+ idx
+ ].mean(dim=0, keepdim=False)[:, : source_length[idx]]
+ finalized_extra["states_output"][ids[idx].item()] = (
+ pre_lprobs[idx]
+ .max(dim=-1, keepdim=False)[1]
+ .contiguous()
+ .view(-1)
+ )
+
+ num_remaining_sent -= len(finalized_sents)
+
+ assert num_remaining_sent >= 0
+ if num_remaining_sent == 0:
+ break
+ if self.search.stop_on_max_len and step >= max_len:
+ break
+ assert step < max_len, f"{step} < {max_len}"
+
+ # Remove finalized sentences (ones for which {beam_size}
+ # finished hypotheses have been generated) from the batch.
+ if len(finalized_sents) > 0:
+ new_bsz = bsz - len(finalized_sents)
+
+ # construct batch_idxs which holds indices of batches to keep for the next pass
+ batch_mask = torch.ones(
+ bsz, dtype=torch.bool, device=cand_indices.device
+ )
+ batch_mask[finalized_sents] = False
+ # TODO replace `nonzero(as_tuple=False)` after TorchScript supports it
+ batch_idxs = torch.arange(
+ bsz, device=cand_indices.device
+ ).masked_select(batch_mask)
+
+ # Choose the subset of the hypothesized constraints that will continue
+ self.search.prune_sentences(batch_idxs)
+
+ eos_mask = eos_mask[batch_idxs]
+ cand_beams = cand_beams[batch_idxs]
+ bbsz_offsets.resize_(new_bsz, 1)
+ cand_bbsz_idx = cand_beams.add(bbsz_offsets)
+ cand_scores = cand_scores[batch_idxs]
+ cand_indices = cand_indices[batch_idxs]
+
+ if prefix_tokens is not None:
+ prefix_tokens = prefix_tokens[batch_idxs]
+ src_lengths = src_lengths[batch_idxs]
+ cands_to_ignore = cands_to_ignore[batch_idxs]
+
+ scores = scores.view(bsz, -1)[batch_idxs].view(new_bsz * beam_size, -1)
+ tokens = tokens.view(bsz, -1)[batch_idxs].view(new_bsz * beam_size, -1)
+ if attn is not None:
+ attn = attn.view(bsz, -1)[batch_idxs].view(
+ new_bsz * beam_size, attn.size(1), -1
+ )
+ bsz = new_bsz
+ else:
+ batch_idxs = None
+
+ # Set active_mask so that values > cand_size indicate eos hypos
+ # and values < cand_size indicate candidate active hypos.
+ # After, the min values per row are the top candidate active hypos
+
+ # Rewrite the operator since the element wise or is not supported in torchscript.
+
+ eos_mask[:, :beam_size] = ~((~cands_to_ignore) & (~eos_mask[:, :beam_size]))
+ active_mask = torch.add(
+ eos_mask.type_as(cand_offsets) * cand_size,
+ cand_offsets[: eos_mask.size(1)],
+ )
+
+ # get the top beam_size active hypotheses, which are just
+ # the hypos with the smallest values in active_mask.
+ # {active_hypos} indicates which {beam_size} hypotheses
+ # from the list of {2 * beam_size} candidates were
+ # selected. Shapes: (batch size, beam size)
+ new_cands_to_ignore, active_hypos = torch.topk(
+ active_mask, k=beam_size, dim=1, largest=False
+ )
+
+ # update cands_to_ignore to ignore any finalized hypos.
+ cands_to_ignore = new_cands_to_ignore.ge(cand_size)[:, :beam_size]
+ # Make sure there is at least one active item for each sentence in the batch.
+ assert (~cands_to_ignore).any(dim=1).all()
+
+ # update cands_to_ignore to ignore any finalized hypos
+
+ # {active_bbsz_idx} denotes which beam number is continued for each new hypothesis (a beam
+ # can be selected more than once).
+ active_bbsz_idx = torch.gather(cand_bbsz_idx, dim=1, index=active_hypos)
+ active_scores = torch.gather(cand_scores, dim=1, index=active_hypos)
+
+ active_bbsz_idx = active_bbsz_idx.view(-1)
+ active_scores = active_scores.view(-1)
+
+ # copy tokens and scores for active hypotheses
+
+ # Set the tokens for each beam (can select the same row more than once)
+ tokens[:, : step + 1] = torch.index_select(
+ tokens[:, : step + 1], dim=0, index=active_bbsz_idx
+ )
+ # Select the next token for each of them
+ tokens.view(bsz, beam_size, -1)[:, :, step + 1] = torch.gather(
+ cand_indices, dim=1, index=active_hypos
+ )
+ if step > 0:
+ scores[:, :step] = torch.index_select(
+ scores[:, :step], dim=0, index=active_bbsz_idx
+ )
+ scores.view(bsz, beam_size, -1)[:, :, step] = torch.gather(
+ cand_scores, dim=1, index=active_hypos
+ )
+
+ # Update constraints based on which candidates were selected for the next beam
+ self.search.update_constraints(active_hypos)
+
+ # copy attention for active hypotheses
+ if attn is not None:
+ attn[:, :, : step + 2] = torch.index_select(
+ attn[:, :, : step + 2], dim=0, index=active_bbsz_idx
+ )
+
+ # reorder incremental state in decoder
+ reorder_state = active_bbsz_idx
+
+ step += 1
+
+ # sort by score descending
+ for sent in range(len(finalized)):
+ scores = torch.tensor(
+ [float(elem["score"].item()) for elem in finalized[sent]]
+ )
+ _, sorted_scores_indices = torch.sort(scores, descending=True)
+ finalized[sent] = [finalized[sent][ssi] for ssi in sorted_scores_indices]
+ finalized[sent] = torch.jit.annotate(
+ List[Dict[str, Tensor]], finalized[sent]
+ )
+ return finalized, finalized_rw, finalized_src_length, finalized_extra
+
+ def bulid_partial_encoder_out(self, encoder_outs, l):
+ new_encoder_outs = [
+ {
+ "encoder_out": [encoder_outs[0]["encoder_out"][0][:l]], # T x B x C
+ "encoder_padding_mask": [
+ encoder_outs[0]["encoder_padding_mask"][0][:, :l]
+ ], # B x T
+ "encoder_embedding": [
+ encoder_outs[0]["encoder_embedding"][0][:, :l, :]
+ ],
+ "encoder_states": encoder_outs[0]["encoder_states"],
+ "src_tokens": [],
+ "src_lengths": [encoder_outs[0]["src_lengths"][0].clamp(0, l)],
+ }
+ ]
+
+ return new_encoder_outs
+
+ def _prefix_tokens(
+ self, step: int, lprobs, scores, tokens, prefix_tokens, beam_size: int
+ ):
+ """Handle prefix tokens"""
+ prefix_toks = prefix_tokens[:, step].unsqueeze(-1).repeat(1, beam_size).view(-1)
+ prefix_lprobs = lprobs.gather(-1, prefix_toks.unsqueeze(-1))
+ prefix_mask = prefix_toks.ne(self.pad)
+ lprobs[prefix_mask] = torch.tensor(-math.inf).to(lprobs)
+ lprobs[prefix_mask] = lprobs[prefix_mask].scatter(
+ -1, prefix_toks[prefix_mask].unsqueeze(-1), prefix_lprobs[prefix_mask]
+ )
+ # if prefix includes eos, then we should make sure tokens and
+ # scores are the same across all beams
+ eos_mask = prefix_toks.eq(self.eos)
+ if eos_mask.any():
+ # validate that the first beam matches the prefix
+ first_beam = tokens[eos_mask].view(-1, beam_size, tokens.size(-1))[
+ :, 0, 1 : step + 1
+ ]
+ eos_mask_batch_dim = eos_mask.view(-1, beam_size)[:, 0]
+ target_prefix = prefix_tokens[eos_mask_batch_dim][:, :step]
+ assert (first_beam == target_prefix).all()
+
+ # copy tokens, scores and lprobs from the first beam to all beams
+ tokens = self.replicate_first_beam(tokens, eos_mask_batch_dim, beam_size)
+ scores = self.replicate_first_beam(scores, eos_mask_batch_dim, beam_size)
+ lprobs = self.replicate_first_beam(lprobs, eos_mask_batch_dim, beam_size)
+ return lprobs, tokens, scores
+
+ def replicate_first_beam(self, tensor, mask, beam_size: int):
+ tensor = tensor.view(-1, beam_size, tensor.size(-1))
+ tensor[mask] = tensor[mask][:, :1, :]
+ return tensor.view(-1, tensor.size(-1))
+
+ def finalize_hypos(
+ self,
+ step: int,
+ bbsz_idx,
+ eos_scores,
+ tokens,
+ scores,
+ finalized: List[List[Dict[str, Tensor]]],
+ finished: List[bool],
+ beam_size: int,
+ attn: Optional[Tensor],
+ src_lengths,
+ max_len: int,
+ ):
+ """Finalize hypothesis, store finalized information in `finalized`, and change `finished` accordingly.
+ A sentence is finalized when {beam_size} finished items have been collected for it.
+
+ Returns number of sentences (not beam items) being finalized.
+ These will be removed from the batch and not processed further.
+ Args:
+ bbsz_idx (Tensor):
+ """
+ assert bbsz_idx.numel() == eos_scores.numel()
+
+ # clone relevant token and attention tensors.
+ # tokens is (batch * beam, max_len). So the index_select
+ # gets the newly EOS rows, then selects cols 1..{step + 2}
+ tokens_clone = tokens.index_select(0, bbsz_idx)[
+ :, 1 : step + 2
+ ] # skip the first index, which is EOS
+
+ tokens_clone[:, step] = self.eos
+ attn_clone = (
+ attn.index_select(0, bbsz_idx)[:, :, 1 : step + 2]
+ if attn is not None
+ else None
+ )
+
+ # compute scores per token position
+ pos_scores = scores.index_select(0, bbsz_idx)[:, : step + 1]
+ pos_scores[:, step] = eos_scores
+ # convert from cumulative to per-position scores
+ pos_scores[:, 1:] = pos_scores[:, 1:] - pos_scores[:, :-1]
+
+ # normalize sentence-level scores
+ if self.normalize_scores:
+ eos_scores /= (step + 1) ** self.len_penalty
+
+ # cum_unfin records which sentences in the batch are finished.
+ # It helps match indexing between (a) the original sentences
+ # in the batch and (b) the current, possibly-reduced set of
+ # sentences.
+ cum_unfin: List[int] = []
+ prev = 0
+ for f in finished:
+ if f:
+ prev += 1
+ else:
+ cum_unfin.append(prev)
+ cum_fin_tensor = torch.tensor(cum_unfin, dtype=torch.int).to(bbsz_idx)
+
+ unfin_idx = torch.div(bbsz_idx, beam_size, rounding_mode="trunc")
+ sent = unfin_idx + torch.index_select(cum_fin_tensor, 0, unfin_idx)
+
+ # Create a set of "{sent}{unfin_idx}", where
+ # "unfin_idx" is the index in the current (possibly reduced)
+ # list of sentences, and "sent" is the index in the original,
+ # unreduced batch
+ # For every finished beam item
+ # sentence index in the current (possibly reduced) batch
+ seen = (sent << 32) + unfin_idx
+ unique_seen: List[int] = torch.unique(seen).tolist()
+
+ if self.match_source_len:
+ condition = step > torch.index_select(src_lengths, 0, unfin_idx)
+ eos_scores = torch.where(condition, torch.tensor(-math.inf), eos_scores)
+ sent_list: List[int] = sent.tolist()
+ for i in range(bbsz_idx.size()[0]):
+ # An input sentence (among those in a batch) is finished when
+ # beam_size hypotheses have been collected for it
+ if len(finalized[sent_list[i]]) < beam_size:
+ if attn_clone is not None:
+ # remove padding tokens from attn scores
+ hypo_attn = attn_clone[i]
+ else:
+ hypo_attn = torch.empty(0)
+
+ finalized[sent_list[i]].append(
+ {
+ "tokens": tokens_clone[i],
+ "score": eos_scores[i],
+ "attention": hypo_attn, # src_len x tgt_len
+ "alignment": torch.empty(0),
+ "positional_scores": pos_scores[i],
+ }
+ )
+
+ newly_finished: List[int] = []
+ for unique_s in unique_seen:
+ # check termination conditions for this sentence
+ unique_sent: int = unique_s >> 32
+ unique_unfin_idx: int = unique_s - (unique_sent << 32)
+
+ if not finished[unique_sent] and self.is_finished(
+ step, unique_unfin_idx, max_len, len(finalized[unique_sent]), beam_size
+ ):
+ finished[unique_sent] = True
+ newly_finished.append(unique_unfin_idx)
+
+ return newly_finished
+
+ def is_finished(
+ self,
+ step: int,
+ unfin_idx: int,
+ max_len: int,
+ finalized_sent_len: int,
+ beam_size: int,
+ ):
+ """
+ Check whether decoding for a sentence is finished, which
+ occurs when the list of finalized sentences has reached the
+ beam size, or when we reach the maximum length.
+ """
+ assert finalized_sent_len <= beam_size
+ if finalized_sent_len == beam_size or step == max_len:
+ return True
+ return False
+
+
+class EnsembleModel(nn.Module):
+ """A wrapper around an ensemble of models."""
+
+ def __init__(self, models):
+ super().__init__()
+ self.models_size = len(models)
+ # method '__len__' is not supported in ModuleList for torch script
+ self.single_model = models[0]
+ self.models = nn.ModuleList(models)
+
+ self.has_incremental: bool = False
+ if all(
+ hasattr(m, "decoder") and isinstance(m.decoder, FairseqIncrementalDecoder)
+ for m in models
+ ):
+ self.has_incremental = True
+
+ def forward(self):
+ pass
+
+ def has_encoder(self):
+ return hasattr(self.single_model, "encoder")
+
+ def has_incremental_states(self):
+ return self.has_incremental
+
+ def max_decoder_positions(self):
+ return min(
+ [
+ m.max_decoder_positions()
+ for m in self.models
+ if hasattr(m, "max_decoder_positions")
+ ]
+ + [sys.maxsize]
+ )
+
+ def set_decoder_beam_size(self, beam_size):
+ """Set beam size for efficient beamable enc-dec attention."""
+ if beam_size > 1:
+ for model in self.models:
+ if hasattr(model, "set_beam_size"):
+ model.set_beam_size(beam_size)
+
+ @torch.jit.export
+ def forward_encoder(self, net_input: Dict[str, Tensor]):
+ if not self.has_encoder():
+ return None
+ return [model.encoder.forward_torchscript(net_input) for model in self.models]
+
+ @torch.jit.export
+ def forward_decoder(
+ self,
+ tokens,
+ encoder_outs: List[Dict[str, List[Tensor]]],
+ incremental_states: List[Dict[str, Dict[str, Optional[Tensor]]]],
+ temperature: float = 1.0,
+ ):
+ log_probs = []
+ avg_attn: Optional[Tensor] = None
+ encoder_out: Optional[Dict[str, List[Tensor]]] = None
+ for i, model in enumerate(self.models):
+ if self.has_encoder():
+ encoder_out = encoder_outs[i]
+ # decode each model
+ # if self.has_incremental_states():
+ if False:
+ decoder_out = model.decoder.forward(
+ tokens,
+ encoder_out=encoder_out,
+ incremental_state=incremental_states[i],
+ )
+ else:
+ if hasattr(model, "decoder"):
+ decoder_out = model.decoder.forward(tokens, encoder_out=encoder_out)
+ else:
+ decoder_out = model.forward(tokens)
+
+ attn: Optional[Tensor] = None
+ decoder_len = len(decoder_out)
+ attn = decoder_out[-1]["attention"]
+ """
+ if decoder_len > 1 and decoder_out[1] is not None:
+ if isinstance(decoder_out[1], Tensor):
+ attn = decoder_out[1]
+ else:
+ attn_holder = decoder_out[1]["attn"]
+ if isinstance(attn_holder, Tensor):
+ attn = attn_holder
+ elif attn_holder is not None:
+ attn = attn_holder[0]
+ if attn is not None:
+ attn = attn[:, -1, :]
+ """
+ decoder_out_tuple = (
+ decoder_out[0].div_(temperature),
+ None if decoder_len <= 1 else decoder_out[1],
+ )
+
+ probs = model.get_normalized_probs(
+ decoder_out_tuple, log_probs=True, sample=None
+ )
+ # probs = probs[:, -1, :]
+
+ transition_probs = model.get_normalized_probs(
+ (decoder_out[1],), log_probs=False, sample=None
+ )
+
+ if self.models_size == 1:
+ return probs, attn, transition_probs
+
+ log_probs.append(probs)
+ if attn is not None:
+ if avg_attn is None:
+ avg_attn = attn
+ else:
+ avg_attn.add_(attn)
+
+ avg_probs = torch.logsumexp(torch.stack(log_probs, dim=0), dim=0) - math.log(
+ self.models_size
+ )
+
+ if avg_attn is not None:
+ avg_attn.div_(self.models_size)
+ return avg_probs, avg_attn
+
+ @torch.jit.export
+ def reorder_encoder_out(
+ self, encoder_outs: Optional[List[Dict[str, List[Tensor]]]], new_order
+ ):
+ """
+ Reorder encoder output according to *new_order*.
+
+ Args:
+ encoder_out: output from the ``forward()`` method
+ new_order (LongTensor): desired order
+
+ Returns:
+ *encoder_out* rearranged according to *new_order*
+ """
+ new_outs: List[Dict[str, List[Tensor]]] = []
+ if not self.has_encoder():
+ return new_outs
+ for i, model in enumerate(self.models):
+ assert encoder_outs is not None
+ new_outs.append(
+ model.encoder.reorder_encoder_out(encoder_outs[i], new_order)
+ )
+ return new_outs
+
+ @torch.jit.export
+ def reorder_incremental_state(
+ self,
+ incremental_states: List[Dict[str, Dict[str, Optional[Tensor]]]],
+ new_order,
+ ):
+ if not self.has_incremental_states():
+ return
+ for i, model in enumerate(self.models):
+ model.decoder.reorder_incremental_state_scripting(
+ incremental_states[i], new_order
+ )
+
+
+class EnsembleModelWithAlignment(EnsembleModel):
+ """A wrapper around an ensemble of models."""
+
+ def __init__(self, models):
+ super().__init__(models)
+
+ def forward_align(self, src_tokens, src_lengths, prev_output_tokens):
+ avg_attn = None
+ for model in self.models:
+ decoder_out = model(src_tokens, src_lengths, prev_output_tokens)
+ attn = decoder_out[1]["attn"][0]
+ if avg_attn is None:
+ avg_attn = attn
+ else:
+ avg_attn.add_(attn)
+ if len(self.models) > 1:
+ avg_attn.div_(len(self.models))
+ return avg_attn
diff --git a/agent/sequence_generator.py b/agent/sequence_generator.py
new file mode 100644
index 0000000000000000000000000000000000000000..e364eaf0f7f731dab9da70cf5d6f35c0ce16182a
--- /dev/null
+++ b/agent/sequence_generator.py
@@ -0,0 +1,687 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import math
+import sys
+from typing import Dict, List, Optional
+import torch
+from torch import Tensor
+
+from fairseq.sequence_generator import EnsembleModel as EnsembleModelBase
+from fairseq.sequence_generator import SequenceGenerator as SequenceGeneratorBase
+
+
+class SequenceGenerator(SequenceGeneratorBase):
+ def __init__(
+ self,
+ models,
+ tgt_dict,
+ beam_size=1,
+ max_len_a=0,
+ max_len_b=200,
+ max_len=0,
+ min_len=1,
+ normalize_scores=True,
+ len_penalty=1.0,
+ unk_penalty=0.0,
+ temperature=1.0,
+ match_source_len=False,
+ no_repeat_ngram_size=0,
+ search_strategy=None,
+ eos=None,
+ symbols_to_strip_from_output=None,
+ lm_model=None,
+ lm_weight=1.0,
+ tokens_to_suppress=(),
+ use_incremental_states=False,
+ ):
+ """Generates translations of a given source sentence.
+
+ Args:
+ models (List[~fairseq.models.FairseqModel]): ensemble of models,
+ currently support fairseq.models.TransformerModel for scripting
+ beam_size (int, optional): beam width (default: 1)
+ max_len_a/b (int, optional): generate sequences of maximum length
+ ax + b, where x is the source length
+ max_len (int, optional): the maximum length of the generated output
+ (not including end-of-sentence)
+ min_len (int, optional): the minimum length of the generated output
+ (not including end-of-sentence)
+ normalize_scores (bool, optional): normalize scores by the length
+ of the output (default: True)
+ len_penalty (float, optional): length penalty, where <1.0 favors
+ shorter, >1.0 favors longer sentences (default: 1.0)
+ unk_penalty (float, optional): unknown word penalty, where <0
+ produces more unks, >0 produces fewer (default: 0.0)
+ temperature (float, optional): temperature, where values
+ >1.0 produce more uniform samples and values <1.0 produce
+ sharper samples (default: 1.0)
+ match_source_len (bool, optional): outputs should match the source
+ length (default: False)
+ """
+ super().__init__(
+ models=models,
+ tgt_dict=tgt_dict,
+ beam_size=beam_size,
+ max_len_a=max_len_a,
+ max_len_b=max_len_b,
+ max_len=max_len,
+ min_len=min_len,
+ normalize_scores=normalize_scores,
+ len_penalty=len_penalty,
+ unk_penalty=unk_penalty,
+ temperature=temperature,
+ match_source_len=match_source_len,
+ no_repeat_ngram_size=no_repeat_ngram_size,
+ search_strategy=search_strategy,
+ eos=eos,
+ symbols_to_strip_from_output=symbols_to_strip_from_output,
+ lm_model=lm_model,
+ lm_weight=lm_weight,
+ tokens_to_suppress=tokens_to_suppress,
+ )
+
+ if isinstance(models, EnsembleModel):
+ self.model = models
+ else:
+ self.model = EnsembleModel(models)
+
+ self.model.set_decoder_beam_size(self.beam_size)
+ self.model.eval()
+ self.use_incremental_states = use_incremental_states
+ self.incremental_states = None
+ if not self.use_incremental_states:
+ self.model.has_incremental = False
+
+ def reset_incremental_states(self):
+ self.incremental_states = None
+
+ def _generate(
+ self,
+ sample: Dict[str, Dict[str, Tensor]],
+ prefix_tokens: Optional[Tensor] = None,
+ constraints: Optional[Tensor] = None,
+ bos_token: Optional[int] = None,
+ ):
+ net_input = sample["net_input"]
+
+ if "src_tokens" in net_input:
+ src_tokens = net_input["src_tokens"]
+ # length of the source text being the character length except EndOfSentence and pad
+ # if src_lengths exists in net_input (speech_to_text dataset case), then use it
+ if "src_lengths" in net_input:
+ src_lengths = net_input["src_lengths"]
+ else:
+ src_lengths = (
+ (src_tokens.ne(self.eos) & src_tokens.ne(self.pad))
+ .long()
+ .sum(dim=1)
+ )
+ elif "source" in net_input:
+ src_tokens = net_input["source"]
+ src_lengths = (
+ net_input["padding_mask"].size(-1) - net_input["padding_mask"].sum(-1)
+ if net_input["padding_mask"] is not None
+ else torch.tensor(src_tokens.size(-1)).to(src_tokens)
+ )
+ elif "features" in net_input:
+ src_tokens = net_input["features"]
+ src_lengths = (
+ net_input["padding_mask"].size(-1) - net_input["padding_mask"].sum(-1)
+ if net_input["padding_mask"] is not None
+ else torch.tensor(src_tokens.size(-1)).to(src_tokens)
+ )
+ else:
+ raise Exception(
+ "expected src_tokens or source in net input. input keys: "
+ + str(net_input.keys())
+ )
+
+ if constraints is not None and not self.search.supports_constraints:
+ raise NotImplementedError(
+ "Target-side constraints were provided, but search method doesn't support them"
+ )
+
+ # Initialize constraints, when active
+ self.search.init_constraints(constraints, self.beam_size)
+
+ # compute the encoder output for each beam
+ with torch.autograd.profiler.record_function("EnsembleModel: forward_encoder"):
+ encoder_outs = self.model.forward_encoder(net_input)
+
+ finalized = self.generate_decoder(
+ encoder_outs,
+ src_tokens,
+ src_lengths,
+ sample,
+ prefix_tokens,
+ constraints,
+ bos_token,
+ )
+ return finalized
+
+ def generate_decoder(
+ self,
+ encoder_outs,
+ src_tokens,
+ src_lengths,
+ sample: Dict[str, Dict[str, Tensor]],
+ prefix_tokens: Optional[Tensor] = None,
+ constraints: Optional[Tensor] = None,
+ bos_token: Optional[int] = None,
+ aux_task_name="",
+ encoder_outs_aug: Optional[
+ Tensor
+ ] = None, # an additional/augmented encoder_outs
+ max_new_tokens=-1,
+ **kwargs,
+ ):
+ if self.use_incremental_states:
+ if self.incremental_states is None:
+ incremental_states = torch.jit.annotate(
+ List[Dict[str, Dict[str, Optional[Tensor]]]],
+ [
+ torch.jit.annotate(Dict[str, Dict[str, Optional[Tensor]]], {})
+ for i in range(self.model.models_size)
+ ],
+ )
+ self.incremental_states = incremental_states
+ else:
+ incremental_states = self.incremental_states
+ else:
+ incremental_states = None
+
+ # bsz: total number of sentences in beam
+ # Note that src_tokens may have more than 2 dimensions (i.e. audio features)
+ bsz, src_len = src_tokens.size()[:2]
+ beam_size = self.beam_size
+
+ decoder_name = f"{aux_task_name}_decoder" if aux_task_name else "decoder"
+
+ start = prefix_tokens.size(-1) if prefix_tokens is not None else 0
+
+ if max_new_tokens == -1:
+ max_len: int = -1
+ if self.match_source_len:
+ max_len = src_lengths.max().item()
+ else:
+ max_len = min(
+ int(self.max_len_a * src_len + self.max_len_b),
+ self.max_len - 1,
+ )
+ else:
+ max_len = start + max_new_tokens
+ assert (
+ self.min_len <= max_len
+ ), "min_len cannot be larger than max_len, please adjust these!"
+
+ # placeholder of indices for bsz * beam_size to hold tokens and accumulative scores
+ new_order = (
+ torch.arange(bsz, device=src_tokens.device)
+ .view(-1, 1)
+ .repeat(1, beam_size)
+ .view(-1)
+ )
+ new_order = new_order.long()
+ encoder_outs = self.model.reorder_encoder_out(encoder_outs, new_order)
+ # ensure encoder_outs is a List.
+ assert encoder_outs is not None
+ if encoder_outs_aug is not None:
+ encoder_outs_aug = self.model.reorder_encoder_out(
+ encoder_outs_aug, new_order
+ )
+
+ # initialize buffers
+ scores = torch.zeros(
+ bsz * beam_size, max_len + 1, device=src_tokens.device
+ ).float() # +1 for eos; pad is never chosen for scoring
+ if prefix_tokens is None:
+ tokens = (
+ torch.zeros(bsz * beam_size, max_len + 2, device=src_tokens.device)
+ .long()
+ .fill_(self.pad)
+ ) # +2 for eos and pad
+ tokens[:, 0] = self.eos if bos_token is None else bos_token
+ else:
+ tokens = torch.cat(
+ (
+ torch.full(
+ (bsz, 1), self.eos, dtype=torch.long, device=src_tokens.device
+ ),
+ prefix_tokens,
+ ),
+ dim=-1,
+ )
+ tokens = tokens.repeat(beam_size, 1)
+ if tokens.size(-1) < max_len + 2:
+ tokens = torch.cat(
+ (
+ tokens,
+ torch.zeros(
+ tokens.size(0),
+ max_len + 2 - tokens.size(-1),
+ device=src_tokens.device,
+ ).long(),
+ ),
+ dim=-1,
+ )
+
+ attn: Optional[Tensor] = None
+
+ # A list that indicates candidates that should be ignored.
+ # For example, suppose we're sampling and have already finalized 2/5
+ # samples. Then cands_to_ignore would mark 2 positions as being ignored,
+ # so that we only finalize the remaining 3 samples.
+ cands_to_ignore = torch.zeros(bsz, beam_size, device=src_tokens.device).eq(
+ -1
+ ) # forward and backward-compatible False mask
+
+ # list of completed sentences
+ finalized = torch.jit.annotate(
+ List[List[Dict[str, Tensor]]],
+ [torch.jit.annotate(List[Dict[str, Tensor]], []) for i in range(bsz)],
+ ) # contains lists of dictionaries of infomation about the hypothesis being finalized at each step
+
+ # a boolean array indicating if the sentence at the index is finished or not
+ finished = [False for i in range(bsz)]
+ num_remaining_sent = bsz # number of sentences remaining
+
+ # number of candidate hypos per step
+ cand_size = 2 * beam_size # 2 x beam size in case half are EOS
+
+ # offset arrays for converting between different indexing schemes
+ bbsz_offsets = (
+ (torch.arange(0, bsz, device=src_tokens.device) * beam_size)
+ .unsqueeze(1)
+ .type_as(tokens)
+ )
+ cand_offsets = torch.arange(0, cand_size, device=src_tokens.device).type_as(
+ tokens
+ )
+
+ reorder_state: Optional[Tensor] = None
+ batch_idxs: Optional[Tensor] = None
+
+ original_batch_idxs: Optional[Tensor] = None
+ if "id" in sample and isinstance(sample["id"], Tensor):
+ original_batch_idxs = sample["id"]
+ else:
+ original_batch_idxs = torch.arange(0, bsz).type_as(tokens)
+
+ for step in range(start, max_len + 1): # one extra step for EOS marker
+ # reorder decoder internal states based on the prev choice of beams
+ if reorder_state is not None:
+ if batch_idxs is not None:
+ # update beam indices to take into account removed sentences
+ corr = batch_idxs - torch.arange(batch_idxs.numel()).type_as(
+ batch_idxs
+ )
+ reorder_state.view(-1, beam_size).add_(
+ corr.unsqueeze(-1) * beam_size
+ )
+ original_batch_idxs = original_batch_idxs[batch_idxs]
+ self.model.reorder_incremental_state(
+ incremental_states, reorder_state, decoder_name
+ )
+ encoder_outs = self.model.reorder_encoder_out(
+ encoder_outs, reorder_state
+ )
+ if encoder_outs_aug is not None:
+ encoder_outs_aug = self.model.reorder_encoder_out(
+ encoder_outs_aug, reorder_state
+ )
+ with torch.autograd.profiler.record_function(
+ "EnsembleModel: forward_decoder"
+ ):
+ lprobs, avg_attn_scores = self.model.forward_decoder(
+ tokens[:, : step + 1],
+ encoder_outs,
+ incremental_states,
+ self.temperature,
+ decoder_name=decoder_name,
+ encoder_outs_aug=encoder_outs_aug,
+ **kwargs,
+ )
+
+ if self.lm_model is not None and not aux_task_name:
+ lm_out = self.lm_model(tokens[:, : step + 1])
+ probs = self.lm_model.get_normalized_probs(
+ lm_out, log_probs=True, sample=None
+ )
+ probs = probs[:, -1, :] * self.lm_weight
+ lprobs += probs
+
+ lprobs[lprobs != lprobs] = torch.tensor(-math.inf, device=lprobs.device)
+
+ lprobs[:, self.pad] = -math.inf # never select pad
+ lprobs[:, self.unk] -= self.unk_penalty # apply unk penalty
+
+ # handle max length constraint
+ if step >= max_len:
+ lprobs[:, : self.eos] = -math.inf
+ lprobs[:, self.eos + 1 :] = -math.inf
+
+ # handle prefix tokens (possibly with different lengths)
+ if (
+ prefix_tokens is not None
+ and step < prefix_tokens.size(1)
+ and step < max_len
+ ):
+ lprobs, tokens, scores = self._prefix_tokens(
+ step, lprobs, scores, tokens, prefix_tokens, beam_size
+ )
+ else:
+ if step < self.min_len:
+ # minimum length constraint (does not apply if using prefix_tokens)
+ lprobs[:, self.eos] = -math.inf
+
+ if self.token_indices_to_suppress is not None:
+ lprobs[:, self.token_indices_to_suppress] = -math.inf
+
+ # Record attention scores, only support avg_attn_scores is a Tensor
+ if avg_attn_scores is not None:
+ if attn is None:
+ attn = torch.empty(
+ bsz * beam_size,
+ avg_attn_scores.size(1),
+ max_len + 2,
+ device=scores.device,
+ )
+ attn[:, :, step + 1].copy_(avg_attn_scores)
+
+ scores = scores.type_as(lprobs)
+ eos_bbsz_idx = torch.empty(0, device=tokens.device).type_as(
+ tokens
+ ) # indices of hypothesis ending with eos (finished sentences)
+ eos_scores = torch.empty(0, device=scores.device).type_as(
+ scores
+ ) # scores of hypothesis ending with eos (finished sentences)
+
+ if self.should_set_src_lengths:
+ self.search.set_src_lengths(src_lengths)
+
+ if self.repeat_ngram_blocker is not None:
+ lprobs = self.repeat_ngram_blocker(tokens, lprobs, bsz, beam_size, step)
+
+ # Shape: (batch, cand_size)
+ cand_scores, cand_indices, cand_beams = self.search.step(
+ step,
+ lprobs.view(bsz, -1, self.vocab_size),
+ scores.view(bsz, beam_size, -1)[:, :, :step],
+ tokens[:, : step + 1],
+ original_batch_idxs,
+ )
+
+ # cand_bbsz_idx contains beam indices for the top candidate
+ # hypotheses, with a range of values: [0, bsz*beam_size),
+ # and dimensions: [bsz, cand_size]
+ cand_bbsz_idx = cand_beams.add(bbsz_offsets)
+
+ # finalize hypotheses that end in eos
+ # Shape of eos_mask: (batch size, beam size)
+ eos_mask = cand_indices.eq(self.eos) & cand_scores.ne(-math.inf)
+ eos_mask[:, :beam_size][cands_to_ignore] = torch.tensor(
+ 0, device=eos_mask.device
+ ).type_as(eos_mask)
+
+ # only consider eos when it's among the top beam_size indices
+ # Now we know what beam item(s) to finish
+ # Shape: 1d list of absolute-numbered
+ eos_bbsz_idx = torch.masked_select(
+ cand_bbsz_idx[:, :beam_size], mask=eos_mask[:, :beam_size]
+ )
+
+ finalized_sents: List[int] = []
+ if eos_bbsz_idx.numel() > 0:
+ eos_scores = torch.masked_select(
+ cand_scores[:, :beam_size], mask=eos_mask[:, :beam_size]
+ )
+
+ finalized_sents = self.finalize_hypos(
+ step,
+ eos_bbsz_idx,
+ eos_scores,
+ tokens,
+ scores,
+ finalized,
+ finished,
+ beam_size,
+ attn,
+ src_lengths,
+ max_len,
+ )
+ num_remaining_sent -= len(finalized_sents)
+
+ assert num_remaining_sent >= 0
+ if num_remaining_sent == 0:
+ break
+ if self.search.stop_on_max_len and step >= max_len:
+ break
+ assert step < max_len, f"{step} < {max_len}"
+
+ # Remove finalized sentences (ones for which {beam_size}
+ # finished hypotheses have been generated) from the batch.
+ if len(finalized_sents) > 0:
+ new_bsz = bsz - len(finalized_sents)
+
+ # construct batch_idxs which holds indices of batches to keep for the next pass
+ batch_mask = torch.ones(
+ bsz, dtype=torch.bool, device=cand_indices.device
+ )
+ batch_mask[finalized_sents] = False
+ # TODO replace `nonzero(as_tuple=False)` after TorchScript supports it
+ batch_idxs = (
+ torch.arange(bsz, device=cand_indices.device)
+ .type_as(cand_indices)
+ .masked_select(batch_mask)
+ )
+
+ # Choose the subset of the hypothesized constraints that will continue
+ self.search.prune_sentences(batch_idxs)
+
+ eos_mask = eos_mask[batch_idxs]
+ cand_beams = cand_beams[batch_idxs]
+ bbsz_offsets.resize_(new_bsz, 1)
+ cand_bbsz_idx = cand_beams.add(bbsz_offsets)
+ cand_scores = cand_scores[batch_idxs]
+ cand_indices = cand_indices[batch_idxs]
+
+ if prefix_tokens is not None:
+ prefix_tokens = prefix_tokens[batch_idxs]
+ src_lengths = src_lengths[batch_idxs]
+ cands_to_ignore = cands_to_ignore[batch_idxs]
+
+ scores = scores.view(bsz, -1)[batch_idxs].view(new_bsz * beam_size, -1)
+ tokens = tokens.view(bsz, -1)[batch_idxs].view(new_bsz * beam_size, -1)
+ if attn is not None:
+ attn = attn.view(bsz, -1)[batch_idxs].view(
+ new_bsz * beam_size, attn.size(1), -1
+ )
+ bsz = new_bsz
+ else:
+ batch_idxs = None
+
+ # Set active_mask so that values > cand_size indicate eos hypos
+ # and values < cand_size indicate candidate active hypos.
+ # After, the min values per row are the top candidate active hypos
+
+ # Rewrite the operator since the element wise or is not supported in torchscript.
+
+ eos_mask[:, :beam_size] = ~((~cands_to_ignore) & (~eos_mask[:, :beam_size]))
+ active_mask = torch.add(
+ eos_mask.type_as(cand_offsets) * cand_size,
+ cand_offsets[: eos_mask.size(1)],
+ )
+
+ # get the top beam_size active hypotheses, which are just
+ # the hypos with the smallest values in active_mask.
+ # {active_hypos} indicates which {beam_size} hypotheses
+ # from the list of {2 * beam_size} candidates were
+ # selected. Shapes: (batch size, beam size)
+ new_cands_to_ignore, active_hypos = torch.topk(
+ active_mask, k=beam_size, dim=1, largest=False
+ )
+
+ # update cands_to_ignore to ignore any finalized hypos.
+ cands_to_ignore = new_cands_to_ignore.ge(cand_size)[:, :beam_size]
+ # Make sure there is at least one active item for each sentence in the batch.
+ assert (~cands_to_ignore).any(dim=1).all()
+
+ # update cands_to_ignore to ignore any finalized hypos
+
+ # {active_bbsz_idx} denotes which beam number is continued for each new hypothesis (a beam
+ # can be selected more than once).
+ active_bbsz_idx = torch.gather(cand_bbsz_idx, dim=1, index=active_hypos)
+ active_scores = torch.gather(cand_scores, dim=1, index=active_hypos)
+
+ active_bbsz_idx = active_bbsz_idx.view(-1)
+ active_scores = active_scores.view(-1)
+
+ # copy tokens and scores for active hypotheses
+
+ # Set the tokens for each beam (can select the same row more than once)
+ tokens[:, : step + 1] = torch.index_select(
+ tokens[:, : step + 1], dim=0, index=active_bbsz_idx
+ )
+ # Select the next token for each of them
+ tokens.view(bsz, beam_size, -1)[:, :, step + 1] = torch.gather(
+ cand_indices, dim=1, index=active_hypos
+ )
+ if step > 0:
+ scores[:, :step] = torch.index_select(
+ scores[:, :step], dim=0, index=active_bbsz_idx
+ )
+ scores.view(bsz, beam_size, -1)[:, :, step] = torch.gather(
+ cand_scores, dim=1, index=active_hypos
+ )
+
+ # Update constraints based on which candidates were selected for the next beam
+ self.search.update_constraints(active_hypos)
+
+ # copy attention for active hypotheses
+ if attn is not None:
+ attn[:, :, : step + 2] = torch.index_select(
+ attn[:, :, : step + 2], dim=0, index=active_bbsz_idx
+ )
+
+ # reorder incremental state in decoder
+ reorder_state = active_bbsz_idx
+
+ # sort by score descending
+ for sent in range(len(finalized)):
+ scores = torch.tensor(
+ [float(elem["score"].item()) for elem in finalized[sent]]
+ )
+ _, sorted_scores_indices = torch.sort(scores, descending=True)
+ finalized[sent] = [finalized[sent][ssi] for ssi in sorted_scores_indices]
+ finalized[sent] = torch.jit.annotate(
+ List[Dict[str, Tensor]], finalized[sent]
+ )
+ return finalized
+
+
+class EnsembleModel(EnsembleModelBase):
+ """A wrapper around an ensemble of models."""
+
+ def __init__(self, models):
+ super().__init__(models)
+
+ @torch.jit.export
+ def forward_decoder(
+ self,
+ tokens,
+ encoder_outs: List[Dict[str, List[Tensor]]],
+ incremental_states: List[Dict[str, Dict[str, Optional[Tensor]]]],
+ temperature: float = 1.0,
+ decoder_name="decoder",
+ encoder_outs_aug: List[Dict[str, List[Tensor]]] = None,
+ **kwargs,
+ ):
+ log_probs = []
+ avg_attn: Optional[Tensor] = None
+ encoder_out: Optional[Dict[str, List[Tensor]]] = None
+ encoder_out_aug: Optional[Dict[str, List[Tensor]]] = None
+ for i, model in enumerate(self.models):
+ if self.has_encoder():
+ encoder_out = encoder_outs[i]
+ if encoder_outs_aug is not None:
+ encoder_out_aug = encoder_outs_aug[i]
+ # decode each model
+ if self.has_incremental_states() and incremental_states is not None:
+ if encoder_out_aug is not None:
+ decoder_out = getattr(model, decoder_name).forward(
+ tokens,
+ encoder_out=encoder_out,
+ encoder_out_aug=encoder_out_aug,
+ incremental_state=incremental_states[i],
+ )
+ else:
+ decoder_out = getattr(model, decoder_name).forward(
+ tokens,
+ encoder_out=encoder_out,
+ incremental_state=incremental_states[i],
+ **kwargs,
+ )
+ else:
+ if hasattr(model, decoder_name):
+ decoder_out = getattr(model, decoder_name).forward(
+ tokens, encoder_out=encoder_out
+ )
+ else:
+ decoder_out = model.forward(tokens)
+
+ attn: Optional[Tensor] = None
+ decoder_len = len(decoder_out)
+ if decoder_len > 1 and decoder_out[1] is not None:
+ if isinstance(decoder_out[1], Tensor):
+ attn = decoder_out[1]
+ else:
+ attn_holder = decoder_out[1]["attn"]
+ if isinstance(attn_holder, Tensor):
+ attn = attn_holder
+ elif attn_holder is not None:
+ attn = attn_holder[0]
+ if attn is not None:
+ attn = attn[:, -1, :]
+
+ decoder_out_tuple = (
+ decoder_out[0][:, -1:, :].div_(temperature),
+ None if decoder_len <= 1 else decoder_out[1],
+ )
+ probs = getattr(model, decoder_name).get_normalized_probs(
+ decoder_out_tuple, log_probs=True, sample=None
+ )
+ probs = probs[:, -1, :]
+ if self.models_size == 1:
+ return probs, attn
+
+ log_probs.append(probs)
+ if attn is not None:
+ if avg_attn is None:
+ avg_attn = attn
+ else:
+ avg_attn.add_(attn)
+
+ avg_probs = torch.logsumexp(torch.stack(log_probs, dim=0), dim=0) - math.log(
+ self.models_size
+ )
+
+ if avg_attn is not None:
+ avg_attn.div_(self.models_size)
+ return avg_probs, avg_attn
+
+ @torch.jit.export
+ def reorder_incremental_state(
+ self,
+ incremental_states: List[Dict[str, Dict[str, Optional[Tensor]]]],
+ new_order,
+ decoder_name="decoder",
+ ):
+ if not self.has_incremental_states():
+ return
+ for i, model in enumerate(self.models):
+ getattr(model, decoder_name).reorder_incremental_state_scripting(
+ incremental_states[i], new_order
+ )
diff --git a/agent/speech_to_speech.streamspeech.agent.py b/agent/speech_to_speech.streamspeech.agent.py
new file mode 100644
index 0000000000000000000000000000000000000000..610e7133423098e46571d4dc95aa791ab347362b
--- /dev/null
+++ b/agent/speech_to_speech.streamspeech.agent.py
@@ -0,0 +1,770 @@
+##########################################
+# Simultaneous Speech-to-Speech Translation Agent for StreamSpeech
+#
+# StreamSpeech: Simultaneous Speech-to-Speech Translation with Multi-task Learning (ACL 2024)
+##########################################
+
+from simuleval.utils import entrypoint
+from simuleval.data.segments import SpeechSegment
+from simuleval.agents import SpeechToSpeechAgent
+from simuleval.agents.actions import WriteAction, ReadAction
+from fairseq.checkpoint_utils import load_model_ensemble_and_task
+from fairseq.models.text_to_speech.hub_interface import TTSHubInterface
+from pathlib import Path
+from typing import Any, Dict, Optional, Union
+from fairseq.data.audio.audio_utils import convert_waveform
+from examples.speech_to_text.data_utils import extract_fbank_features
+import ast
+import math
+import os
+import json
+import numpy as np
+from copy import deepcopy
+import torch
+import torchaudio.compliance.kaldi as kaldi
+import yaml
+from fairseq import checkpoint_utils, tasks, utils, options
+from fairseq.file_io import PathManager
+from fairseq import search
+from fairseq.data.audio.feature_transforms import CompositeAudioFeatureTransform
+
+
+SHIFT_SIZE = 10
+WINDOW_SIZE = 25
+ORG_SAMPLE_RATE = 48000
+SAMPLE_RATE = 16000
+FEATURE_DIM = 80
+BOW_PREFIX = "\u2581"
+DEFAULT_EOS = 2
+
+
+class OnlineFeatureExtractor:
+ """
+ Extract speech feature on the fly.
+ """
+
+ def __init__(self, args, cfg):
+ self.shift_size = args.shift_size
+ self.window_size = args.window_size
+ assert self.window_size >= self.shift_size
+
+ self.sample_rate = args.sample_rate
+ self.feature_dim = args.feature_dim
+ self.num_samples_per_shift = int(self.shift_size * self.sample_rate / 1000)
+ self.num_samples_per_window = int(self.window_size * self.sample_rate / 1000)
+ self.len_ms_to_samples = lambda x: x * self.sample_rate / 1000
+ self.previous_residual_samples = []
+ self.global_cmvn = args.global_cmvn
+ self.device = "cuda" if args.device == "gpu" else "cpu"
+ self.feature_transforms = CompositeAudioFeatureTransform.from_config_dict(
+ {"feature_transforms": ["utterance_cmvn"]}
+ )
+
+ def clear_cache(self):
+ self.previous_residual_samples = []
+
+ def __call__(self, new_samples, sr=ORG_SAMPLE_RATE):
+ samples = new_samples
+
+ # # num_frames is the number of frames from the new segment
+ num_frames = math.floor(
+ (len(samples) - self.len_ms_to_samples(self.window_size - self.shift_size))
+ / self.num_samples_per_shift
+ )
+
+ # # the number of frames used for feature extraction
+ # # including some part of thte previous segment
+ effective_num_samples = int(
+ num_frames * self.len_ms_to_samples(self.shift_size)
+ + self.len_ms_to_samples(self.window_size - self.shift_size)
+ )
+ samples = samples[:effective_num_samples]
+ waveform, sample_rate = convert_waveform(
+ torch.tensor([samples]), sr, to_mono=True, to_sample_rate=16000
+ )
+ output = extract_fbank_features(waveform, 16000)
+ output = self.transform(output)
+ return torch.tensor(output, device=self.device)
+
+ def transform(self, input):
+ if self.global_cmvn is None:
+ return input
+
+ mean = self.global_cmvn["mean"]
+ std = self.global_cmvn["std"]
+
+ x = np.subtract(input, mean)
+ x = np.divide(x, std)
+ return x
+
+
+@entrypoint
+class StreamSpeechS2STAgent(SpeechToSpeechAgent):
+ """
+ Incrementally feed text to this offline Fastspeech2 TTS model,
+ with a minimum numbers of phonemes every chunk.
+ """
+
+ def __init__(self, args):
+ super().__init__(args)
+ self.eos = DEFAULT_EOS
+
+ self.gpu = self.args.device == "gpu"
+ self.device = "cuda" if args.device == "gpu" else "cpu"
+
+ self.args = args
+
+ self.load_model_vocab(args)
+
+ self.max_len = args.max_len
+
+ self.force_finish = args.force_finish
+
+ torch.set_grad_enabled(False)
+
+ tgt_dict_mt = self.dict[f"{self.models[0].mt_task_name}"]
+ tgt_dict = self.dict["tgt"]
+ tgt_dict_asr = self.dict["source_unigram"]
+ tgt_dict_st = self.dict["ctc_target_unigram"]
+ args.user_dir=args.agent_dir
+ utils.import_user_module(args)
+ from agent.sequence_generator import SequenceGenerator
+ from agent.ctc_generator import CTCSequenceGenerator
+ from agent.ctc_decoder import CTCDecoder
+ from agent.tts.vocoder import CodeHiFiGANVocoderWithDur
+
+ self.ctc_generator = CTCSequenceGenerator(
+ tgt_dict, self.models, use_incremental_states=False
+ )
+
+ self.asr_ctc_generator = CTCDecoder(tgt_dict_asr, self.models)
+ self.st_ctc_generator = CTCDecoder(tgt_dict_st, self.models)
+
+ self.generator = SequenceGenerator(
+ self.models,
+ tgt_dict,
+ beam_size=1,
+ max_len_a=1,
+ max_len_b=200,
+ max_len=0,
+ min_len=1,
+ normalize_scores=True,
+ len_penalty=1.0,
+ unk_penalty=0.0,
+ temperature=1.0,
+ match_source_len=False,
+ no_repeat_ngram_size=0,
+ search_strategy=search.BeamSearch(tgt_dict),
+ eos=tgt_dict.eos(),
+ symbols_to_strip_from_output=None,
+ )
+
+ self.generator_mt = SequenceGenerator(
+ self.models,
+ tgt_dict_mt,
+ beam_size=1,
+ max_len_a=0,
+ max_len_b=100,
+ max_len=0,
+ min_len=1,
+ normalize_scores=True,
+ len_penalty=1.0,
+ unk_penalty=0.0,
+ temperature=1.0,
+ match_source_len=False,
+ no_repeat_ngram_size=0,
+ search_strategy=search.BeamSearch(tgt_dict_mt),
+ eos=tgt_dict_mt.eos(),
+ symbols_to_strip_from_output=None,
+ use_incremental_states=False,
+ )
+
+ with open(args.vocoder_cfg) as f:
+ vocoder_cfg = json.load(f)
+ self.vocoder = CodeHiFiGANVocoderWithDur(args.vocoder, vocoder_cfg)
+ if self.device == "cuda":
+ self.vocoder = self.vocoder.cuda()
+ self.dur_prediction = args.dur_prediction
+
+ self.lagging_k1 = args.lagging_k1
+ self.lagging_k2 = args.lagging_k2
+ self.segment_size = args.segment_size
+ self.stride_n = args.stride_n
+
+ self.unit_per_subword = args.unit_per_subword
+ self.stride_n2 = args.stride_n2
+
+ if args.extra_output_dir is not None:
+ self.asr_file = Path(args.extra_output_dir + "/asr.txt")
+ self.st_file = Path(args.extra_output_dir + "/st.txt")
+ self.unit_file = Path(args.extra_output_dir + "/unit.txt")
+ self.quiet = False
+ else:
+ self.quiet = True
+
+ self.output_asr_translation = args.output_asr_translation
+
+ if args.source_segment_size >= 640:
+ self.whole_word = True
+ else:
+ self.whole_word = False
+
+ self.reset()
+
+ @staticmethod
+ def add_args(parser):
+ parser.add_argument(
+ "--model-path",
+ type=str,
+ required=True,
+ help="path to your pretrained model.",
+ )
+ parser.add_argument(
+ "--data-bin", type=str, required=True, help="Path of data binary"
+ )
+ parser.add_argument(
+ "--config-yaml", type=str, default=None, help="Path to config yaml file"
+ )
+ parser.add_argument(
+ "--multitask-config-yaml",
+ type=str,
+ default=None,
+ help="Path to config yaml file",
+ )
+ parser.add_argument(
+ "--global-stats",
+ type=str,
+ default=None,
+ help="Path to json file containing cmvn stats",
+ )
+ parser.add_argument(
+ "--tgt-splitter-type",
+ type=str,
+ default="SentencePiece",
+ help="Subword splitter type for target text",
+ )
+ parser.add_argument(
+ "--tgt-splitter-path",
+ type=str,
+ default=None,
+ help="Subword splitter model path for target text",
+ )
+ parser.add_argument(
+ "--user-dir",
+ type=str,
+ default="researches/ctc_unity",
+ help="User directory for model",
+ )
+ parser.add_argument(
+ "--agent-dir",
+ type=str,
+ default="agent",
+ help="User directory for agents",
+ )
+ parser.add_argument(
+ "--max-len", type=int, default=200, help="Max length of translation"
+ )
+ parser.add_argument(
+ "--force-finish",
+ default=False,
+ action="store_true",
+ help="Force the model to finish the hypothsis if the source is not finished",
+ )
+ parser.add_argument(
+ "--shift-size",
+ type=int,
+ default=SHIFT_SIZE,
+ help="Shift size of feature extraction window.",
+ )
+ parser.add_argument(
+ "--window-size",
+ type=int,
+ default=WINDOW_SIZE,
+ help="Window size of feature extraction window.",
+ )
+ parser.add_argument(
+ "--sample-rate", type=int, default=ORG_SAMPLE_RATE, help="Sample rate"
+ )
+ parser.add_argument(
+ "--feature-dim",
+ type=int,
+ default=FEATURE_DIM,
+ help="Acoustic feature dimension.",
+ )
+ parser.add_argument(
+ "--vocoder", type=str, required=True, help="path to the CodeHiFiGAN vocoder"
+ )
+ parser.add_argument(
+ "--vocoder-cfg",
+ type=str,
+ required=True,
+ help="path to the CodeHiFiGAN vocoder config",
+ )
+ parser.add_argument(
+ "--dur-prediction",
+ action="store_true",
+ help="enable duration prediction (for reduced/unique code sequences)",
+ )
+ parser.add_argument("--lagging-k1", type=int, default=0, help="lagging number")
+ parser.add_argument("--lagging-k2", type=int, default=0, help="lagging number")
+ parser.add_argument(
+ "--segment-size", type=int, default=320, help="segment-size"
+ )
+ parser.add_argument("--stride-n", type=int, default=1, help="lagging number")
+ parser.add_argument("--stride-n2", type=int, default=1, help="lagging number")
+ parser.add_argument(
+ "--unit-per-subword", type=int, default=15, help="lagging number"
+ )
+ parser.add_argument(
+ "--extra-output-dir", type=str, default=None, help="extra output dir"
+ )
+ parser.add_argument(
+ "--output-asr-translation",
+ type=bool,
+ default=False,
+ help="extra output dir",
+ )
+
+ def reset(self):
+ self.src_seg_num = 0
+ self.tgt_subwords_indices = None
+ self.src_ctc_indices = None
+ self.src_ctc_prefix_length = 0
+ self.tgt_ctc_prefix_length = 0
+ self.tgt_units_indices = None
+ self.prev_output_tokens_mt = None
+ self.tgt_text = []
+ self.mt_decoder_out = None
+ self.unit = None
+ self.wav = []
+ self.post_transcription = ""
+ self.unfinished_wav = None
+ self.states.reset()
+ try:
+ self.generator_mt.reset_incremental_states()
+ self.ctc_generator.reset_incremental_states()
+ except:
+ pass
+
+ def to_device(self, tensor):
+ if self.gpu:
+ return tensor.cuda()
+ else:
+ return tensor.cpu()
+
+ def load_model_vocab(self, args):
+ filename = args.model_path
+ if not os.path.exists(filename):
+ raise IOError("Model file not found: {}".format(filename))
+
+ state = checkpoint_utils.load_checkpoint_to_cpu(filename)
+ state["cfg"].common['user_dir']=args.user_dir
+ utils.import_user_module(state["cfg"].common)
+
+ task_args = state["cfg"]["task"]
+ task_args.data = args.data_bin
+
+ args.global_cmvn = None
+ if args.config_yaml is not None:
+ task_args.config_yaml = args.config_yaml
+ with open(os.path.join(args.data_bin, args.config_yaml), "r") as f:
+ config = yaml.load(f, Loader=yaml.BaseLoader)
+
+ if "global_cmvn" in config:
+ args.global_cmvn = np.load(config["global_cmvn"]["stats_npz_path"])
+
+ self.feature_extractor = OnlineFeatureExtractor(args, config)
+
+ if args.multitask_config_yaml is not None:
+ task_args.multitask_config_yaml = args.multitask_config_yaml
+
+ task = tasks.setup_task(task_args)
+ self.task = task
+
+ overrides = ast.literal_eval(state["cfg"].common_eval.model_overrides)
+
+ models, saved_cfg = checkpoint_utils.load_model_ensemble(
+ utils.split_paths(filename),
+ arg_overrides=overrides,
+ task=task,
+ suffix=state["cfg"].checkpoint.checkpoint_suffix,
+ strict=(state["cfg"].checkpoint.checkpoint_shard_count == 1),
+ num_shards=state["cfg"].checkpoint.checkpoint_shard_count,
+ )
+
+ chunk_size = args.source_segment_size // 40
+
+ self.models = models
+
+ for model in self.models:
+ model.eval()
+ model.share_memory()
+ if self.gpu:
+ model.cuda()
+ model.encoder.chunk_size = chunk_size
+
+ if chunk_size >= 16:
+ chunk_size = 16
+ else:
+ chunk_size = 8
+ for conv in model.encoder.subsample.conv_layers:
+ conv.chunk_size = chunk_size
+ for layer in model.encoder.conformer_layers:
+ layer.conv_module.depthwise_conv.chunk_size = chunk_size
+
+ # Set dictionary
+ self.dict = {}
+ self.dict["tgt"] = task.target_dictionary
+
+ for k, v in task.multitask_tasks.items():
+ self.dict[k] = v.tgt_dict
+
+ @torch.inference_mode()
+ def policy(self):
+
+ feature = self.feature_extractor(self.states.source)
+
+ if feature.size(0) == 0 and not self.states.source_finished:
+ return ReadAction()
+
+ src_indices = feature.unsqueeze(0)
+ src_lengths = torch.tensor([feature.size(0)], device=self.device).long()
+
+ self.encoder_outs = self.generator.model.forward_encoder(
+ {"src_tokens": src_indices, "src_lengths": src_lengths}
+ )
+
+ finalized_asr = self.asr_ctc_generator.generate(
+ self.encoder_outs[0], aux_task_name="source_unigram"
+ )
+ asr_probs = torch.exp(finalized_asr[0][0]["lprobs"])
+
+ for i, hypo in enumerate(finalized_asr):
+ i_beam = 0
+ tmp = hypo[i_beam]["tokens"].int()
+ src_ctc_indices = tmp
+ src_ctc_index = hypo[i_beam]["index"]
+ text = "".join([self.dict["source_unigram"][c] for c in tmp])
+ text = text.replace("_", " ")
+ text = text.replace("▁", " ")
+ text = text.replace("", " ")
+ text = text.replace("", "")
+ text = text.replace("", "")
+ if len(text) > 0 and text[0] == " ":
+ text = text[1:]
+ if self.states.source_finished and not self.quiet:
+ with open(self.asr_file, "a") as file:
+ print(text, file=file)
+ if self.output_asr_translation:
+ print("Streaming ASR:", text)
+
+ finalized_st = self.st_ctc_generator.generate(
+ self.encoder_outs[0], aux_task_name="ctc_target_unigram"
+ )
+ st_probs = torch.exp(finalized_st[0][0]["lprobs"])
+
+ for i, hypo in enumerate(finalized_st):
+ i_beam = 0
+ tmp = hypo[i_beam]["tokens"].int()
+ tgt_ctc_indices = tmp
+ tgt_ctc_index = hypo[i_beam]["index"]
+ text = "".join([self.dict["ctc_target_unigram"][c] for c in tmp])
+ text = text.replace("_", " ")
+ text = text.replace("▁", " ")
+ text = text.replace("", " ")
+ text = text.replace("", "")
+ text = text.replace("", "")
+ if len(text) > 0 and text[0] == " ":
+ text = text[1:]
+
+ if not self.states.source_finished:
+ src_ctc_prefix_length = src_ctc_indices.size(-1)
+ tgt_ctc_prefix_length = tgt_ctc_indices.size(-1)
+
+ self.src_ctc_indices = src_ctc_indices
+ if (
+ src_ctc_prefix_length < self.src_ctc_prefix_length + self.stride_n
+ or tgt_ctc_prefix_length < self.tgt_ctc_prefix_length + self.stride_n
+ ):
+ return ReadAction()
+ self.src_ctc_prefix_length = max(
+ src_ctc_prefix_length, self.src_ctc_prefix_length
+ )
+ self.tgt_ctc_prefix_length = max(
+ tgt_ctc_prefix_length, self.tgt_ctc_prefix_length
+ )
+ subword_tokens = (
+ (tgt_ctc_prefix_length - self.lagging_k1) // self.stride_n
+ ) * self.stride_n
+
+ if self.whole_word:
+ subword_tokens += 1
+ new_subword_tokens = (
+ (subword_tokens - self.tgt_subwords_indices.size(-1))
+ if self.tgt_subwords_indices is not None
+ else subword_tokens
+ )
+
+ if new_subword_tokens < 1:
+ return ReadAction()
+ else:
+ self.src_ctc_indices = src_ctc_indices
+ new_subword_tokens = -1
+
+ new_subword_tokens = int(new_subword_tokens)
+
+ single_model = self.generator.model.single_model
+ mt_decoder = getattr(single_model, f"{single_model.mt_task_name}_decoder")
+
+ # 1. MT decoder
+ finalized_mt = self.generator_mt.generate_decoder(
+ self.encoder_outs,
+ src_indices,
+ src_lengths,
+ {
+ "id": 1,
+ "net_input": {"src_tokens": src_indices, "src_lengths": src_lengths},
+ },
+ self.tgt_subwords_indices,
+ None,
+ None,
+ aux_task_name=single_model.mt_task_name,
+ max_new_tokens=new_subword_tokens,
+ )
+
+ if finalized_mt[0][0]["tokens"][-1] == 2:
+ tgt_subwords_indices = finalized_mt[0][0]["tokens"][:-1].unsqueeze(0)
+ else:
+ tgt_subwords_indices = finalized_mt[0][0]["tokens"].unsqueeze(0)
+
+ if self.whole_word:
+ j = 999999
+ if not self.states.source_finished:
+ for j in range(tgt_subwords_indices.size(-1) - 1, -1, -1):
+ if self.generator_mt.tgt_dict[
+ tgt_subwords_indices[0][j]
+ ].startswith("▁"):
+ break
+ tgt_subwords_indices = tgt_subwords_indices[:, :j]
+ finalized_mt[0][0]["tokens"] = finalized_mt[0][0]["tokens"][:j]
+
+ if j == 0:
+ return ReadAction()
+
+ new_incremental_states = [{}]
+ if (
+ self.generator_mt.incremental_states is not None
+ and self.generator_mt.use_incremental_states
+ ):
+ for k, v in self.generator_mt.incremental_states[0].items():
+ if v["prev_key"].size(2) == v["prev_value"].size(2):
+ new_incremental_states[0][k] = {
+ "prev_key": v["prev_key"][:, :, :j, :].contiguous(),
+ "prev_value": v["prev_value"][:, :, :j, :].contiguous(),
+ "prev_key_padding_mask": None,
+ }
+ else:
+ new_incremental_states[0][k] = {
+ "prev_key": v["prev_key"],
+ "prev_value": v["prev_value"][:, :, :j, :].contiguous(),
+ "prev_key_padding_mask": None,
+ }
+ self.generator_mt.incremental_states = deepcopy(
+ new_incremental_states
+ )
+
+ max_tgt_len = max([len(hypo[0]["tokens"]) for hypo in finalized_mt])
+ if self.whole_word:
+ max_tgt_len += 1
+ prev_output_tokens_mt = (
+ src_indices.new_zeros(src_indices.shape[0], max_tgt_len)
+ .fill_(mt_decoder.padding_idx)
+ .int()
+ )
+
+ for i, hypo in enumerate(finalized_mt):
+ i_beam = 0
+ tmp = hypo[i_beam]["tokens"].int()
+ prev_output_tokens_mt[i, 0] = self.generator_mt.eos
+ if tmp[-1] == self.generator_mt.eos:
+ tmp = tmp[:-1]
+ prev_output_tokens_mt[i, 1 : len(tmp) + 1] = tmp
+
+ tokens = [self.generator_mt.tgt_dict[c] for c in tmp]
+
+ text = "".join(tokens)
+ text = text.replace("_", " ")
+ text = text.replace("▁", " ")
+ text = text.replace("", " ")
+ text = text.replace("", "")
+ text = text.replace("", "")
+ if len(text) > 0 and text[0] == " ":
+ text = text[1:]
+ if self.states.source_finished and not self.quiet:
+ with open(self.st_file, "a") as file:
+ print(text, file=file)
+ if self.output_asr_translation:
+ print("Simultaneous translation:", text)
+
+ if self.tgt_subwords_indices is not None and torch.equal(
+ self.tgt_subwords_indices, tgt_subwords_indices
+ ):
+ if not self.states.source_finished:
+ return ReadAction()
+ else:
+ return WriteAction(
+ SpeechSegment(
+ content=(
+ self.unfinished_wav.tolist()
+ if self.unfinished_wav is not None
+ else []
+ ),
+ sample_rate=SAMPLE_RATE,
+ finished=True,
+ ),
+ finished=True,
+ )
+ self.tgt_subwords_indices = tgt_subwords_indices
+
+ if not self.states.source_finished:
+ if self.prev_output_tokens_mt is not None:
+ if torch.equal(
+ self.prev_output_tokens_mt, prev_output_tokens_mt
+ ) or prev_output_tokens_mt.size(-1) <= self.prev_output_tokens_mt.size(
+ -1
+ ):
+ return ReadAction()
+ self.prev_output_tokens_mt = prev_output_tokens_mt
+ mt_decoder_out = mt_decoder(
+ prev_output_tokens_mt,
+ encoder_out=self.encoder_outs[0],
+ features_only=True,
+ )[0].transpose(0, 1)
+
+ if self.mt_decoder_out is None:
+ self.mt_decoder_out = mt_decoder_out
+ else:
+ self.mt_decoder_out = torch.cat(
+ (self.mt_decoder_out, mt_decoder_out[self.mt_decoder_out.size(0) :]),
+ dim=0,
+ )
+ self.mt_decoder_out = mt_decoder_out
+ x = self.mt_decoder_out
+
+ if getattr(single_model, "proj", None) is not None:
+ x = single_model.proj(x)
+
+ mt_decoder_padding_mask = None
+ if prev_output_tokens_mt.eq(mt_decoder.padding_idx).any():
+ mt_decoder_padding_mask = prev_output_tokens_mt.eq(mt_decoder.padding_idx)
+
+ # 2. T2U encoder
+ if getattr(single_model, "synthesizer_encoder", None) is not None:
+ t2u_encoder_out = single_model.synthesizer_encoder(
+ x,
+ mt_decoder_padding_mask,
+ )
+ else:
+ t2u_encoder_out = {
+ "encoder_out": [x], # T x B x C
+ "encoder_padding_mask": (
+ [mt_decoder_padding_mask]
+ if mt_decoder_padding_mask is not None
+ else []
+ ), # B x T
+ "encoder_embedding": [],
+ "encoder_states": [],
+ "src_tokens": [],
+ "src_lengths": [],
+ }
+
+ if getattr(single_model, "t2u_augmented_cross_attn", False):
+ encoder_outs_aug = [t2u_encoder_out]
+ else:
+ encoder_outs = [t2u_encoder_out]
+ encoder_outs_aug = None
+ finalized = self.ctc_generator.generate(
+ encoder_outs[0],
+ prefix=self.tgt_units_indices,
+ )
+
+ if len(finalized[0][0]["tokens"]) == 0:
+ if not self.states.source_finished:
+ return ReadAction()
+ else:
+ return WriteAction(
+ SpeechSegment(
+ content=(
+ self.unfinished_wav.tolist()
+ if self.unfinished_wav is not None
+ else []
+ ),
+ sample_rate=SAMPLE_RATE,
+ finished=True,
+ ),
+ finished=True,
+ )
+
+ for i, hypo in enumerate(finalized):
+ i_beam = 0
+ tmp = hypo[i_beam]["tokens"].int() # hyp + eos
+ if tmp[-1] == self.generator.eos:
+ tmp = tmp[:-1]
+ unit = []
+ for c in tmp:
+ u = self.generator.tgt_dict[c].replace("", "").replace("", "")
+ if u != "":
+ unit.append(int(u))
+
+ if len(unit) > 0 and unit[0] == " ":
+ unit = unit[1:]
+ text = " ".join([str(_) for _ in unit])
+ if self.states.source_finished and not self.quiet:
+ with open(self.unit_file, "a") as file:
+ print(text, file=file)
+ cur_unit = unit if self.unit is None else unit[len(self.unit) :]
+ if len(unit) < 1 or len(cur_unit) < 1:
+ if not self.states.source_finished:
+ return ReadAction()
+ else:
+ return WriteAction(
+ SpeechSegment(
+ content=(
+ self.unfinished_wav.tolist()
+ if self.unfinished_wav is not None
+ else []
+ ),
+ sample_rate=SAMPLE_RATE,
+ finished=True,
+ ),
+ finished=True,
+ )
+
+ x = {
+ "code": torch.tensor(unit, dtype=torch.long, device=self.device).view(
+ 1, -1
+ ),
+ }
+ wav, dur = self.vocoder(x, self.dur_prediction)
+
+ cur_wav_length = dur[:, -len(cur_unit) :].sum() * 320
+ new_wav = wav[-cur_wav_length:]
+ if self.unfinished_wav is not None and len(self.unfinished_wav) > 0:
+ new_wav = torch.cat((self.unfinished_wav, new_wav), dim=0)
+
+ self.wav = wav
+ self.unit = unit
+
+ # A SpeechSegment has to be returned for speech-to-speech translation system
+ if self.states.source_finished and new_subword_tokens == -1:
+ self.states.target_finished = True
+ self.reset()
+
+ return WriteAction(
+ SpeechSegment(
+ content=new_wav.tolist(),
+ sample_rate=SAMPLE_RATE,
+ finished=self.states.source_finished,
+ ),
+ finished=self.states.target_finished,
+ )
diff --git a/agent/speech_to_speech.wait-k-stride-n.agent.py b/agent/speech_to_speech.wait-k-stride-n.agent.py
new file mode 100644
index 0000000000000000000000000000000000000000..8101ae701a94b71f38debc590f9deb7a31e4b497
--- /dev/null
+++ b/agent/speech_to_speech.wait-k-stride-n.agent.py
@@ -0,0 +1,581 @@
+##########################################
+# Simultaneous Speech-to-Speech Translation Agent for Wait-k Policy
+#
+# StreamSpeech: Simultaneous Speech-to-Speech Translation with Multi-task Learning (ACL 2024)
+##########################################
+
+from simuleval.utils import entrypoint
+from simuleval.data.segments import SpeechSegment
+from simuleval.agents import SpeechToSpeechAgent
+from simuleval.agents.actions import WriteAction, ReadAction
+from fairseq.checkpoint_utils import load_model_ensemble_and_task
+from fairseq.models.text_to_speech.hub_interface import TTSHubInterface
+from pathlib import Path
+from typing import Any, Dict, Optional, Union
+from fairseq.data.audio.audio_utils import convert_waveform
+from examples.speech_to_text.data_utils import extract_fbank_features
+import ast
+import math
+import os
+import json
+import numpy as np
+import torch
+import torchaudio.compliance.kaldi as kaldi
+import yaml
+from fairseq import checkpoint_utils, tasks, utils, options
+from fairseq.file_io import PathManager
+from fairseq import search
+from fairseq.data.audio.feature_transforms import CompositeAudioFeatureTransform
+from fairseq.models.text_to_speech.vocoder import CodeHiFiGANVocoder
+
+SHIFT_SIZE = 10
+WINDOW_SIZE = 25
+ORG_SAMPLE_RATE = 48000
+SAMPLE_RATE = 16000
+FEATURE_DIM = 80
+BOW_PREFIX = "\u2581"
+DEFAULT_EOS = 2
+
+
+class OnlineFeatureExtractor:
+ """
+ Extract speech feature on the fly.
+ """
+
+ def __init__(self, args, cfg):
+ self.shift_size = args.shift_size
+ self.window_size = args.window_size
+ assert self.window_size >= self.shift_size
+
+ self.sample_rate = args.sample_rate
+ self.feature_dim = args.feature_dim
+ self.num_samples_per_shift = int(self.shift_size * self.sample_rate / 1000)
+ self.num_samples_per_window = int(self.window_size * self.sample_rate / 1000)
+ self.len_ms_to_samples = lambda x: x * self.sample_rate / 1000
+ self.previous_residual_samples = []
+ self.global_cmvn = args.global_cmvn
+ self.device = "cuda" if args.device == "gpu" else "cpu"
+ self.feature_transforms = CompositeAudioFeatureTransform.from_config_dict(
+ {"feature_transforms": ["utterance_cmvn"]}
+ )
+
+ def clear_cache(self):
+ self.previous_residual_samples = []
+
+ def __call__(self, new_samples, sr=ORG_SAMPLE_RATE):
+ samples = new_samples
+
+ # # num_frames is the number of frames from the new segment
+ num_frames = math.floor(
+ (len(samples) - self.len_ms_to_samples(self.window_size - self.shift_size))
+ / self.num_samples_per_shift
+ )
+
+ # # the number of frames used for feature extraction
+ # # including some part of thte previous segment
+ effective_num_samples = int(
+ num_frames * self.len_ms_to_samples(self.shift_size)
+ + self.len_ms_to_samples(self.window_size - self.shift_size)
+ )
+ samples = samples[:effective_num_samples]
+ waveform, sample_rate = convert_waveform(
+ torch.tensor([samples]), sr, to_mono=True, to_sample_rate=16000
+ )
+ output = extract_fbank_features(waveform, 16000)
+ output = self.transform(output)
+ return torch.tensor(output, device=self.device)
+
+ def transform(self, input):
+ if self.global_cmvn is None:
+ return input
+
+ mean = self.global_cmvn["mean"]
+ std = self.global_cmvn["std"]
+
+ x = np.subtract(input, mean)
+ x = np.divide(x, std)
+ return x
+
+
+@entrypoint
+class WaitkS2STAgent(SpeechToSpeechAgent):
+ """
+ Incrementally feed text to this offline Fastspeech2 TTS model,
+ with a minimum numbers of phonemes every chunk.
+ """
+
+ def __init__(self, args):
+ super().__init__(args)
+ self.eos = DEFAULT_EOS
+
+ self.gpu = self.args.device == "gpu"
+ self.device = "cuda" if args.device == "gpu" else "cpu"
+
+ self.args = args
+
+ self.load_model_vocab(args)
+
+ self.max_len = args.max_len
+
+ self.force_finish = args.force_finish
+
+ torch.set_grad_enabled(False)
+
+ tgt_dict_mt = self.dict[f"{self.models[0].mt_task_name}"]
+ tgt_dict = self.dict["tgt"]
+ args.user_dir=args.agent_dir
+ utils.import_user_module(args)
+ from agent.sequence_generator import SequenceGenerator
+
+ self.generator = SequenceGenerator(
+ self.models,
+ tgt_dict,
+ beam_size=1,
+ max_len_a=1,
+ max_len_b=200,
+ max_len=0,
+ min_len=1,
+ normalize_scores=True,
+ len_penalty=1.0,
+ unk_penalty=0.0,
+ temperature=1.0,
+ match_source_len=False,
+ no_repeat_ngram_size=0,
+ search_strategy=search.BeamSearch(tgt_dict),
+ eos=tgt_dict.eos(),
+ symbols_to_strip_from_output=None,
+ )
+
+ self.generator_mt = SequenceGenerator(
+ self.models,
+ tgt_dict_mt,
+ beam_size=1,
+ max_len_a=1,
+ max_len_b=200,
+ max_len=0,
+ min_len=1,
+ normalize_scores=True,
+ len_penalty=1.0,
+ unk_penalty=0.0,
+ temperature=1.0,
+ match_source_len=False,
+ no_repeat_ngram_size=0,
+ search_strategy=search.BeamSearch(tgt_dict_mt),
+ eos=tgt_dict_mt.eos(),
+ symbols_to_strip_from_output=None,
+ )
+
+ from agent.tts.vocoder import CodeHiFiGANVocoderWithDur
+
+ with open(args.vocoder_cfg) as f:
+ vocoder_cfg = json.load(f)
+ self.vocoder = CodeHiFiGANVocoderWithDur(args.vocoder, vocoder_cfg)
+ if self.device == "cuda":
+ self.vocoder = self.vocoder.cuda()
+ self.dur_prediction = args.dur_prediction
+
+ self.lagging_k1 = args.lagging_k1
+ self.lagging_k2 = args.lagging_k2
+ self.segment_size = args.segment_size
+ self.stride_n = args.stride_n
+ self.unit_per_subword = args.unit_per_subword
+ self.stride_n2 = args.stride_n2
+ self.reset()
+
+ @staticmethod
+ def add_args(parser):
+ parser.add_argument(
+ "--model-path",
+ type=str,
+ required=True,
+ help="path to your pretrained model.",
+ )
+ parser.add_argument(
+ "--data-bin", type=str, required=True, help="Path of data binary"
+ )
+ parser.add_argument(
+ "--config-yaml", type=str, default=None, help="Path to config yaml file"
+ )
+ parser.add_argument(
+ "--multitask-config-yaml",
+ type=str,
+ default=None,
+ help="Path to config yaml file",
+ )
+ parser.add_argument(
+ "--global-stats",
+ type=str,
+ default=None,
+ help="Path to json file containing cmvn stats",
+ )
+ parser.add_argument(
+ "--tgt-splitter-type",
+ type=str,
+ default="SentencePiece",
+ help="Subword splitter type for target text",
+ )
+ parser.add_argument(
+ "--tgt-splitter-path",
+ type=str,
+ default=None,
+ help="Subword splitter model path for target text",
+ )
+ parser.add_argument(
+ "--user-dir",
+ type=str,
+ default="researches/ctc_unity",
+ help="User directory for model",
+ )
+ parser.add_argument(
+ "--agent-dir",
+ type=str,
+ default="agent",
+ help="User directory for agents",
+ )
+ parser.add_argument(
+ "--max-len", type=int, default=200, help="Max length of translation"
+ )
+ parser.add_argument(
+ "--force-finish",
+ default=False,
+ action="store_true",
+ help="Force the model to finish the hypothsis if the source is not finished",
+ )
+ parser.add_argument(
+ "--shift-size",
+ type=int,
+ default=SHIFT_SIZE,
+ help="Shift size of feature extraction window.",
+ )
+ parser.add_argument(
+ "--window-size",
+ type=int,
+ default=WINDOW_SIZE,
+ help="Window size of feature extraction window.",
+ )
+ parser.add_argument(
+ "--sample-rate", type=int, default=SAMPLE_RATE, help="Sample rate"
+ )
+ parser.add_argument(
+ "--feature-dim",
+ type=int,
+ default=FEATURE_DIM,
+ help="Acoustic feature dimension.",
+ )
+ parser.add_argument(
+ "--vocoder", type=str, required=True, help="path to the CodeHiFiGAN vocoder"
+ )
+ parser.add_argument(
+ "--vocoder-cfg",
+ type=str,
+ required=True,
+ help="path to the CodeHiFiGAN vocoder config",
+ )
+ parser.add_argument(
+ "--dur-prediction",
+ action="store_true",
+ help="enable duration prediction (for reduced/unique code sequences)",
+ )
+ parser.add_argument("--lagging-k1", type=int, default=3, help="lagging number")
+ parser.add_argument("--lagging-k2", type=int, default=1, help="lagging number")
+ parser.add_argument(
+ "--segment-size", type=int, default=320, help="segment-size"
+ )
+ parser.add_argument("--stride-n", type=int, default=1, help="lagging number")
+ parser.add_argument("--stride-n2", type=int, default=1, help="lagging number")
+ parser.add_argument(
+ "--unit-per-subword", type=int, default=15, help="lagging number"
+ )
+
+ def reset(self):
+ self.src_seg_num = 0
+ self.tgt_subwords_indices = None
+ self.tgt_units_indices = None
+ self.unit = None
+ self.wav = []
+ self.states.reset()
+
+ def to_device(self, tensor):
+ if self.gpu:
+ return tensor.cuda()
+ else:
+ return tensor.cpu()
+
+ def load_model_vocab(self, args):
+ filename = args.model_path
+ if not os.path.exists(filename):
+ raise IOError("Model file not found: {}".format(filename))
+
+ state = checkpoint_utils.load_checkpoint_to_cpu(filename)
+ state["cfg"].common['user_dir']=args.user_dir
+ utils.import_user_module(state["cfg"].common)
+
+ task_args = state["cfg"]["task"]
+ task_args.data = args.data_bin
+
+ args.global_cmvn = None
+ if args.config_yaml is not None:
+ task_args.config_yaml = args.config_yaml
+ with open(os.path.join(args.data_bin, args.config_yaml), "r") as f:
+ config = yaml.load(f, Loader=yaml.BaseLoader)
+
+ if "global_cmvn" in config:
+ args.global_cmvn = np.load(config["global_cmvn"]["stats_npz_path"])
+
+ self.feature_extractor = OnlineFeatureExtractor(args, config)
+
+ if args.multitask_config_yaml is not None:
+ task_args.multitask_config_yaml = args.multitask_config_yaml
+
+ task = tasks.setup_task(task_args)
+
+ overrides = ast.literal_eval(state["cfg"].common_eval.model_overrides)
+
+ models, saved_cfg = checkpoint_utils.load_model_ensemble(
+ utils.split_paths(filename),
+ arg_overrides=overrides,
+ task=task,
+ suffix=state["cfg"].checkpoint.checkpoint_suffix,
+ strict=(state["cfg"].checkpoint.checkpoint_shard_count == 1),
+ num_shards=state["cfg"].checkpoint.checkpoint_shard_count,
+ )
+
+ self.models = models
+
+ for model in self.models:
+ model.eval()
+ model.share_memory()
+ if self.gpu:
+ model.cuda()
+
+ # Set dictionary
+ self.dict = {}
+ self.dict["tgt"] = task.target_dictionary
+
+ for k, v in task.multitask_tasks.items():
+ self.dict[k] = v.tgt_dict
+
+ def policy(self):
+ src_seg_num = len(self.states.source) // (
+ self.segment_size * ORG_SAMPLE_RATE / 1000
+ )
+ cur_tgt_subword_tokens = (
+ self.tgt_subwords_indices.size(-1)
+ if self.tgt_subwords_indices is not None
+ else 0
+ )
+ cur_tgt_unit_tokens = (
+ self.tgt_units_indices.size(-1) if self.tgt_units_indices is not None else 0
+ )
+ if (
+ src_seg_num <= self.src_seg_num or src_seg_num < self.lagging_k1
+ ) and not self.states.source_finished:
+ return ReadAction()
+ else:
+ self.src_seg_num = src_seg_num
+
+ subword_tokens = (
+ (self.src_seg_num - self.lagging_k1) // self.stride_n
+ ) * self.stride_n
+ unit_tokens = (
+ ((subword_tokens - self.lagging_k2) // self.stride_n2)
+ * self.stride_n2
+ * self.unit_per_subword
+ )
+
+ new_subword_tokens = (
+ (subword_tokens - self.tgt_subwords_indices.size(-1))
+ if self.tgt_subwords_indices is not None
+ else subword_tokens
+ )
+ new_unit_tokens = (
+ (unit_tokens - self.tgt_units_indices.size(-1))
+ if self.tgt_units_indices is not None
+ else unit_tokens
+ )
+
+ if (
+ new_subword_tokens < 1 or new_unit_tokens < 1
+ ) and not self.states.source_finished:
+ return ReadAction()
+ if self.states.source_finished:
+ new_subword_tokens = -1
+ new_unit_tokens = -1
+ new_subword_tokens = int(new_subword_tokens)
+ new_unit_tokens = int(new_unit_tokens)
+
+ feature = self.feature_extractor(self.states.source)
+ if feature.size(0) == 0 and not self.states.source_finished:
+ return ReadAction()
+
+ src_indices = feature.unsqueeze(0)
+ src_lengths = torch.tensor([feature.size(0)], device=self.device).long()
+
+ self.encoder_outs = self.generator.model.forward_encoder(
+ {"src_tokens": src_indices, "src_lengths": src_lengths}
+ )
+
+ single_model = self.generator.model.single_model
+ mt_decoder = getattr(single_model, f"{single_model.mt_task_name}_decoder")
+
+ # 1. MT decoder
+ finalized_mt = self.generator_mt.generate_decoder(
+ self.encoder_outs,
+ src_indices,
+ src_lengths,
+ {
+ "id": 1,
+ "net_input": {"src_tokens": src_indices, "src_lengths": src_lengths},
+ },
+ self.tgt_subwords_indices,
+ None,
+ None,
+ aux_task_name=single_model.mt_task_name,
+ max_new_tokens=new_subword_tokens,
+ )
+ self.tgt_subwords_indices = finalized_mt[0][0]["tokens"][:-1].unsqueeze(0)
+
+ unit_tokens = (
+ ((self.tgt_subwords_indices.size(-1) - self.lagging_k2) // self.stride_n2)
+ * self.stride_n2
+ * self.unit_per_subword
+ )
+ new_unit_tokens = (
+ (unit_tokens - self.tgt_units_indices.size(-1))
+ if self.tgt_units_indices is not None
+ else unit_tokens
+ )
+
+ if new_unit_tokens < 1 and not self.states.source_finished:
+ return ReadAction()
+ if self.states.source_finished:
+ new_unit_tokens = -1
+ new_unit_tokens = int(new_unit_tokens)
+
+ max_tgt_len = max([len(hypo[0]["tokens"]) for hypo in finalized_mt])
+ prev_output_tokens_mt = (
+ src_indices.new_zeros(src_indices.shape[0], max_tgt_len)
+ .fill_(mt_decoder.padding_idx)
+ .int()
+ ) # B x T
+
+ for i, hypo in enumerate(finalized_mt):
+ i_beam = 0
+ tmp = hypo[i_beam]["tokens"].int() # hyp + eos
+ prev_output_tokens_mt[i, 0] = self.generator_mt.eos
+ if tmp[-1] == self.generator_mt.eos:
+ tmp = tmp[:-1]
+ prev_output_tokens_mt[i, 1 : len(tmp) + 1] = tmp
+
+ text = "".join([self.generator_mt.tgt_dict[c] for c in tmp])
+ text = text.replace("_", " ")
+ text = text.replace("▁", " ")
+ text = text.replace("", " ")
+ text = text.replace("", "")
+ text = text.replace("", "")
+ if len(text) > 0 and text[0] == " ":
+ text = text[1:]
+ # print('text: ',text)
+
+ x = mt_decoder(
+ prev_output_tokens_mt,
+ encoder_out=self.encoder_outs[0],
+ features_only=True,
+ )[0].transpose(0, 1)
+
+ if getattr(single_model, "proj", None) is not None:
+ x = single_model.proj(x)
+
+ mt_decoder_padding_mask = None
+ if prev_output_tokens_mt.eq(mt_decoder.padding_idx).any():
+ mt_decoder_padding_mask = prev_output_tokens_mt.eq(mt_decoder.padding_idx)
+
+ # 2. T2U encoder
+ if getattr(single_model, "synthesizer_encoder", None) is not None:
+ t2u_encoder_out = single_model.synthesizer_encoder(
+ x,
+ mt_decoder_padding_mask,
+ )
+ else:
+ t2u_encoder_out = {
+ "encoder_out": [x], # T x B x C
+ "encoder_padding_mask": (
+ [mt_decoder_padding_mask]
+ if mt_decoder_padding_mask is not None
+ else []
+ ), # B x T
+ "encoder_embedding": [],
+ "encoder_states": [],
+ "src_tokens": [],
+ "src_lengths": [],
+ }
+
+ if getattr(single_model, "t2u_augmented_cross_attn", False):
+ encoder_outs_aug = [t2u_encoder_out]
+ else:
+ encoder_outs = [t2u_encoder_out]
+ encoder_outs_aug = None
+
+ # 3. T2U decoder
+ finalized = self.generator.generate_decoder(
+ encoder_outs,
+ src_indices,
+ src_lengths,
+ {
+ "id": 1,
+ "net_input": {"src_tokens": src_indices, "src_lengths": src_lengths},
+ },
+ self.tgt_units_indices,
+ None,
+ None,
+ encoder_outs_aug=encoder_outs_aug,
+ max_new_tokens=new_unit_tokens,
+ )
+
+ for i, hypo in enumerate(finalized):
+ i_beam = 0
+ tmp = hypo[i_beam]["tokens"].int() # hyp + eos
+ if tmp[-1] == self.generator.eos:
+ tmp = tmp[:-1]
+
+ unit = [int(self.generator.tgt_dict[c]) for c in tmp]
+ if len(unit) > 0 and unit[0] == " ":
+ unit = unit[1:]
+ cur_unit = unit if self.unit is None else unit[len(self.unit) :]
+ if len(unit) < 1 or len(cur_unit) < 1:
+ if not self.states.source_finished:
+ return ReadAction()
+ else:
+ return WriteAction(
+ SpeechSegment(
+ content=[],
+ sample_rate=SAMPLE_RATE,
+ finished=True,
+ ),
+ finished=True,
+ )
+
+ x = {
+ "code": torch.tensor(unit, dtype=torch.long, device=self.device).view(
+ 1, -1
+ ),
+ }
+ wav, dur = self.vocoder(x, self.dur_prediction)
+
+ cur_wav_length = dur[:, -len(cur_unit) :].sum() * 320
+ new_wav = wav[-cur_wav_length:]
+ self.wav = wav
+ self.unit = unit
+
+ if new_subword_tokens == -1 and new_unit_tokens == -1:
+ self.states.target_finished = True
+ self.reset()
+
+ return WriteAction(
+ SpeechSegment(
+ content=new_wav.tolist(),
+ sample_rate=SAMPLE_RATE,
+ finished=self.states.source_finished,
+ ),
+ finished=self.states.target_finished,
+ )
diff --git a/agent/speech_to_text.asr.streamspeech.agent.py b/agent/speech_to_text.asr.streamspeech.agent.py
new file mode 100644
index 0000000000000000000000000000000000000000..f9478adcace85913b2c30a1bc62da21a79cdaf2a
--- /dev/null
+++ b/agent/speech_to_text.asr.streamspeech.agent.py
@@ -0,0 +1,433 @@
+##########################################
+# Streaming ASR Agent for StreamSpeech
+#
+# StreamSpeech: Simultaneous Speech-to-Speech Translation with Multi-task Learning (ACL 2024)
+##########################################
+
+from simuleval.utils import entrypoint
+from simuleval.data.segments import SpeechSegment
+from simuleval.agents import SpeechToTextAgent
+from simuleval.agents.actions import WriteAction, ReadAction
+from fairseq.checkpoint_utils import load_model_ensemble_and_task
+from fairseq.models.text_to_speech.hub_interface import TTSHubInterface
+from pathlib import Path
+from typing import Any, Dict, Optional, Union
+from fairseq.data.audio.audio_utils import convert_waveform
+from examples.speech_to_text.data_utils import extract_fbank_features
+import ast
+import math
+import os
+import json
+import numpy as np
+import torch
+import torchaudio.compliance.kaldi as kaldi
+import yaml
+from fairseq import checkpoint_utils, tasks, utils, options
+from fairseq.file_io import PathManager
+from fairseq import search
+from fairseq.data.audio.feature_transforms import CompositeAudioFeatureTransform
+
+from examples.speech_to_speech.asr_bleu.utils import retrieve_asr_config, ASRGenerator
+
+
+SHIFT_SIZE = 10
+WINDOW_SIZE = 25
+ORG_SAMPLE_RATE = 48000
+SAMPLE_RATE = 16000
+FEATURE_DIM = 80
+BOW_PREFIX = "\u2581"
+DEFAULT_EOS = 2
+
+
+class OnlineFeatureExtractor:
+ """
+ Extract speech feature on the fly.
+ """
+
+ def __init__(self, args, cfg):
+ self.shift_size = args.shift_size
+ self.window_size = args.window_size
+ assert self.window_size >= self.shift_size
+
+ self.sample_rate = args.sample_rate
+ self.feature_dim = args.feature_dim
+ self.num_samples_per_shift = int(self.shift_size * self.sample_rate / 1000)
+ self.num_samples_per_window = int(self.window_size * self.sample_rate / 1000)
+ self.len_ms_to_samples = lambda x: x * self.sample_rate / 1000
+ self.previous_residual_samples = []
+ self.global_cmvn = args.global_cmvn
+ self.device = "cuda" if args.device == "gpu" else "cpu"
+ self.feature_transforms = CompositeAudioFeatureTransform.from_config_dict(
+ {"feature_transforms": ["utterance_cmvn"]}
+ )
+
+ def clear_cache(self):
+ self.previous_residual_samples = []
+
+ def __call__(self, new_samples, sr=ORG_SAMPLE_RATE):
+ samples = new_samples
+
+ # # num_frames is the number of frames from the new segment
+ num_frames = math.floor(
+ (len(samples) - self.len_ms_to_samples(self.window_size - self.shift_size))
+ / self.num_samples_per_shift
+ )
+
+ # # the number of frames used for feature extraction
+ # # including some part of thte previous segment
+ effective_num_samples = int(
+ num_frames * self.len_ms_to_samples(self.shift_size)
+ + self.len_ms_to_samples(self.window_size - self.shift_size)
+ )
+ samples = samples[:effective_num_samples]
+ waveform, sample_rate = convert_waveform(
+ torch.tensor([samples]), sr, to_mono=True, to_sample_rate=16000
+ )
+ output = extract_fbank_features(waveform, 16000)
+ output = self.transform(output)
+ return torch.tensor(output, device=self.device)
+
+ def transform(self, input):
+ if self.global_cmvn is None:
+ return input
+
+ mean = self.global_cmvn["mean"]
+ std = self.global_cmvn["std"]
+
+ x = np.subtract(input, mean)
+ x = np.divide(x, std)
+ return x
+
+
+@entrypoint
+class StreamSpeechASRAgent(SpeechToTextAgent):
+ """
+ Incrementally feed text to this offline Fastspeech2 TTS model,
+ with a minimum numbers of phonemes every chunk.
+ """
+
+ def __init__(self, args):
+ super().__init__(args)
+ self.eos = DEFAULT_EOS
+
+ self.gpu = self.args.device == "gpu"
+ self.device = "cuda" if args.device == "gpu" else "cpu"
+
+ self.args = args
+
+ self.load_model_vocab(args)
+
+ self.max_len = args.max_len
+
+ self.force_finish = args.force_finish
+
+ torch.set_grad_enabled(False)
+
+ tgt_dict_mt = self.dict[f"{self.models[0].mt_task_name}"]
+ tgt_dict = self.dict["tgt"]
+ tgt_dict_asr = self.dict["source_unigram"]
+ tgt_dict_st = self.dict["ctc_target_unigram"]
+ args.user_dir=args.agent_dir
+ utils.import_user_module(args)
+ from agent.sequence_generator import SequenceGenerator
+ from agent.ctc_generator import CTCSequenceGenerator
+ from agent.ctc_decoder import CTCDecoder
+ from agent.tts.vocoder import CodeHiFiGANVocoderWithDur
+
+ self.ctc_generator = CTCSequenceGenerator(
+ tgt_dict, self.models, use_incremental_states=True
+ )
+
+ self.asr_ctc_generator = CTCDecoder(tgt_dict_asr, self.models)
+ self.st_ctc_generator = CTCDecoder(tgt_dict_st, self.models)
+
+ self.generator = SequenceGenerator(
+ self.models,
+ tgt_dict,
+ beam_size=1,
+ max_len_a=1,
+ max_len_b=200,
+ max_len=0,
+ min_len=1,
+ normalize_scores=True,
+ len_penalty=1.0,
+ unk_penalty=0.0,
+ temperature=1.0,
+ match_source_len=False,
+ no_repeat_ngram_size=0,
+ search_strategy=search.BeamSearch(tgt_dict),
+ eos=tgt_dict.eos(),
+ symbols_to_strip_from_output=None,
+ )
+
+ self.generator_mt = SequenceGenerator(
+ self.models,
+ tgt_dict_mt,
+ beam_size=1,
+ max_len_a=1,
+ max_len_b=200,
+ max_len=0,
+ min_len=1,
+ normalize_scores=True,
+ len_penalty=1.0,
+ unk_penalty=0.0,
+ temperature=1.0,
+ match_source_len=False,
+ no_repeat_ngram_size=0,
+ search_strategy=search.BeamSearch(tgt_dict_mt),
+ eos=tgt_dict_mt.eos(),
+ symbols_to_strip_from_output=None,
+ use_incremental_states=True,
+ )
+ self.lagging_k1 = args.lagging_k1
+ self.lagging_k2 = args.lagging_k2
+ self.segment_size = args.segment_size
+ self.stride_n = args.stride_n
+ self.unit_per_subword = args.unit_per_subword
+ self.stride_n2 = args.stride_n2
+ if args.extra_output_dir is not None:
+ self.asr_file = Path(args.extra_output_dir + "/asr.txt")
+ self.st_file = Path(args.extra_output_dir + "/st.txt")
+ self.unit_file = Path(args.extra_output_dir + "/unit.txt")
+ # pass
+ self.quiet = False
+ else:
+ self.quiet = True
+
+ self.reset()
+
+ @staticmethod
+ def add_args(parser):
+ parser.add_argument(
+ "--model-path",
+ type=str,
+ required=True,
+ help="path to your pretrained model.",
+ )
+ parser.add_argument(
+ "--data-bin", type=str, required=True, help="Path of data binary"
+ )
+ parser.add_argument(
+ "--config-yaml", type=str, default=None, help="Path to config yaml file"
+ )
+ parser.add_argument(
+ "--multitask-config-yaml",
+ type=str,
+ default=None,
+ help="Path to config yaml file",
+ )
+ parser.add_argument(
+ "--global-stats",
+ type=str,
+ default=None,
+ help="Path to json file containing cmvn stats",
+ )
+ parser.add_argument(
+ "--tgt-splitter-type",
+ type=str,
+ default="SentencePiece",
+ help="Subword splitter type for target text",
+ )
+ parser.add_argument(
+ "--tgt-splitter-path",
+ type=str,
+ default=None,
+ help="Subword splitter model path for target text",
+ )
+ parser.add_argument(
+ "--user-dir",
+ type=str,
+ default="researches/ctc_unity",
+ help="User directory for model",
+ )
+ parser.add_argument(
+ "--agent-dir",
+ type=str,
+ default="agent",
+ help="User directory for agents",
+ )
+ parser.add_argument(
+ "--max-len", type=int, default=200, help="Max length of translation"
+ )
+ parser.add_argument(
+ "--force-finish",
+ default=False,
+ action="store_true",
+ help="Force the model to finish the hypothsis if the source is not finished",
+ )
+ parser.add_argument(
+ "--shift-size",
+ type=int,
+ default=SHIFT_SIZE,
+ help="Shift size of feature extraction window.",
+ )
+ parser.add_argument(
+ "--window-size",
+ type=int,
+ default=WINDOW_SIZE,
+ help="Window size of feature extraction window.",
+ )
+ parser.add_argument(
+ "--sample-rate", type=int, default=ORG_SAMPLE_RATE, help="Sample rate"
+ )
+ parser.add_argument(
+ "--feature-dim",
+ type=int,
+ default=FEATURE_DIM,
+ help="Acoustic feature dimension.",
+ )
+
+ parser.add_argument("--lagging-k1", type=int, default=0, help="lagging number")
+ parser.add_argument("--lagging-k2", type=int, default=0, help="lagging number")
+ parser.add_argument(
+ "--segment-size", type=int, default=320, help="segment-size"
+ )
+ parser.add_argument("--stride-n", type=int, default=1, help="lagging number")
+ parser.add_argument("--stride-n2", type=int, default=1, help="lagging number")
+ parser.add_argument(
+ "--unit-per-subword", type=int, default=15, help="lagging number"
+ )
+ parser.add_argument(
+ "--extra-output-dir", type=str, default=None, help="extra output dir"
+ )
+
+ def reset(self):
+ self.src_seg_num = 0
+ self.tgt_subwords_indices = None
+ self.src_ctc_indices = None
+ self.src_ctc_prefix_length = 0
+ self.tgt_ctc_prefix_length = 0
+ self.tgt_units_indices = None
+ self.prev_output_tokens_mt = None
+ self.tgt_text = ""
+ self.asr_text = ""
+ self.mt_decoder_out = None
+ self.unit = None
+ self.wav = []
+ self.post_transcription = ""
+ self.unfinished_wav = None
+ self.states.reset()
+ try:
+ self.generator_mt.reset_incremental_states()
+ self.ctc_generator.reset_incremental_states()
+ except:
+ pass
+
+ def to_device(self, tensor):
+ if self.gpu:
+ return tensor.cuda()
+ else:
+ return tensor.cpu()
+
+ def load_model_vocab(self, args):
+ filename = args.model_path
+ if not os.path.exists(filename):
+ raise IOError("Model file not found: {}".format(filename))
+
+ state = checkpoint_utils.load_checkpoint_to_cpu(filename)
+ state["cfg"].common['user_dir']=args.user_dir
+ utils.import_user_module(state["cfg"].common)
+
+ task_args = state["cfg"]["task"]
+ task_args.data = args.data_bin
+
+ args.global_cmvn = None
+ if args.config_yaml is not None:
+ task_args.config_yaml = args.config_yaml
+ with open(os.path.join(args.data_bin, args.config_yaml), "r") as f:
+ config = yaml.load(f, Loader=yaml.BaseLoader)
+
+ if "global_cmvn" in config:
+ args.global_cmvn = np.load(config["global_cmvn"]["stats_npz_path"])
+
+ self.feature_extractor = OnlineFeatureExtractor(args, config)
+
+ if args.multitask_config_yaml is not None:
+ task_args.multitask_config_yaml = args.multitask_config_yaml
+
+ task = tasks.setup_task(task_args)
+ self.task = task
+ overrides = ast.literal_eval(state["cfg"].common_eval.model_overrides)
+
+ models, saved_cfg = checkpoint_utils.load_model_ensemble(
+ utils.split_paths(filename),
+ arg_overrides=overrides,
+ task=task,
+ suffix=state["cfg"].checkpoint.checkpoint_suffix,
+ strict=(state["cfg"].checkpoint.checkpoint_shard_count == 1),
+ num_shards=state["cfg"].checkpoint.checkpoint_shard_count,
+ )
+
+ chunk_size = args.source_segment_size // 40
+
+ self.models = models
+
+ for model in self.models:
+ model.eval()
+ model.share_memory()
+ if self.gpu:
+ model.cuda()
+ model.encoder.chunk_size = chunk_size
+ chunk_size = min(chunk_size, 16)
+ for conv in model.encoder.subsample.conv_layers:
+ conv.chunk_size = chunk_size
+ for layer in model.encoder.conformer_layers:
+ layer.conv_module.depthwise_conv.chunk_size = chunk_size
+
+ # Set dictionary
+ self.dict = {}
+ self.dict["tgt"] = task.target_dictionary
+
+ for k, v in task.multitask_tasks.items():
+ self.dict[k] = v.tgt_dict
+
+ @torch.inference_mode()
+ def policy(self):
+
+ feature = self.feature_extractor(self.states.source)
+ if feature.size(0) == 0 and not self.states.source_finished:
+ return ReadAction()
+
+ src_indices = feature.unsqueeze(0)
+ src_lengths = torch.tensor([feature.size(0)], device=self.device).long()
+
+ self.encoder_outs = self.generator.model.forward_encoder(
+ {"src_tokens": src_indices, "src_lengths": src_lengths}
+ )
+
+ finalized_asr = self.asr_ctc_generator.generate(
+ self.encoder_outs[0], aux_task_name="source_unigram"
+ )
+ asr_probs = torch.exp(finalized_asr[0][0]["lprobs"])
+
+ for i, hypo in enumerate(finalized_asr):
+ i_beam = 0
+ tmp = hypo[i_beam]["tokens"].int() # hyp + eos
+ src_ctc_indices = tmp
+ src_ctc_index = hypo[i_beam]["index"]
+ tokens = [self.dict["source_unigram"][c] for c in tmp]
+ text = "".join([self.dict["source_unigram"][c] for c in tmp])
+ text = text.replace("_", " ")
+ text = text.replace("▁", " ")
+ text = text.replace("", " ")
+ text = text.replace("", "")
+ text = text.replace("", "")
+ if len(text) > 0 and text[0] == " ":
+ text = text[1:]
+ if self.states.source_finished and not self.quiet:
+ with open(self.asr_file, "a") as file:
+ print(text, file=file)
+
+ text = " ".join(tokens)
+ new_text = text[len(self.asr_text) :]
+
+ self.asr_text = text
+
+ if self.states.source_finished:
+ self.states.target_finished = True
+ self.reset()
+
+ return WriteAction(
+ new_text,
+ finished=self.states.target_finished,
+ )
diff --git a/agent/speech_to_text.s2tt.streamspeech.agent.py b/agent/speech_to_text.s2tt.streamspeech.agent.py
new file mode 100644
index 0000000000000000000000000000000000000000..6e089f88eabac18a318585ced91a5c659236bee4
--- /dev/null
+++ b/agent/speech_to_text.s2tt.streamspeech.agent.py
@@ -0,0 +1,545 @@
+##########################################
+# Simultaneous Speech-to-Text Translation Agent for StreamSpeech
+#
+# StreamSpeech: Simultaneous Speech-to-Speech Translation with Multi-task Learning (ACL 2024)
+##########################################
+
+from simuleval.utils import entrypoint
+from simuleval.data.segments import SpeechSegment
+from simuleval.agents import SpeechToTextAgent
+from simuleval.agents.actions import WriteAction, ReadAction
+from fairseq.checkpoint_utils import load_model_ensemble_and_task
+from fairseq.models.text_to_speech.hub_interface import TTSHubInterface
+from pathlib import Path
+from typing import Any, Dict, Optional, Union
+from fairseq.data.audio.audio_utils import convert_waveform
+from examples.speech_to_text.data_utils import extract_fbank_features
+import ast
+import math
+import os
+import json
+import numpy as np
+import torch
+import torchaudio.compliance.kaldi as kaldi
+import yaml
+from fairseq import checkpoint_utils, tasks, utils, options
+from fairseq.file_io import PathManager
+from fairseq import search
+from fairseq.data.audio.feature_transforms import CompositeAudioFeatureTransform
+
+from examples.speech_to_speech.asr_bleu.utils import retrieve_asr_config, ASRGenerator
+
+SHIFT_SIZE = 10
+WINDOW_SIZE = 25
+ORG_SAMPLE_RATE = 48000
+SAMPLE_RATE = 16000
+FEATURE_DIM = 80
+BOW_PREFIX = "\u2581"
+DEFAULT_EOS = 2
+
+
+class OnlineFeatureExtractor:
+ """
+ Extract speech feature on the fly.
+ """
+
+ def __init__(self, args, cfg):
+ self.shift_size = args.shift_size
+ self.window_size = args.window_size
+ assert self.window_size >= self.shift_size
+
+ self.sample_rate = args.sample_rate
+ self.feature_dim = args.feature_dim
+ self.num_samples_per_shift = int(self.shift_size * self.sample_rate / 1000)
+ self.num_samples_per_window = int(self.window_size * self.sample_rate / 1000)
+ self.len_ms_to_samples = lambda x: x * self.sample_rate / 1000
+ self.previous_residual_samples = []
+ self.global_cmvn = args.global_cmvn
+ self.device = "cuda" if args.device == "gpu" else "cpu"
+ self.feature_transforms = CompositeAudioFeatureTransform.from_config_dict(
+ {"feature_transforms": ["utterance_cmvn"]}
+ )
+
+ def clear_cache(self):
+ self.previous_residual_samples = []
+
+ def __call__(self, new_samples, sr=ORG_SAMPLE_RATE):
+ samples = new_samples
+ # # num_frames is the number of frames from the new segment
+ num_frames = math.floor(
+ (len(samples) - self.len_ms_to_samples(self.window_size - self.shift_size))
+ / self.num_samples_per_shift
+ )
+
+ # # the number of frames used for feature extraction
+ # # including some part of thte previous segment
+ effective_num_samples = int(
+ num_frames * self.len_ms_to_samples(self.shift_size)
+ + self.len_ms_to_samples(self.window_size - self.shift_size)
+ )
+ samples = samples[:effective_num_samples]
+ waveform, sample_rate = convert_waveform(
+ torch.tensor([samples]), sr, to_mono=True, to_sample_rate=16000
+ )
+ output = extract_fbank_features(waveform, 16000)
+ output = self.transform(output)
+ return torch.tensor(output, device=self.device)
+
+ def transform(self, input):
+ if self.global_cmvn is None:
+ return input
+
+ mean = self.global_cmvn["mean"]
+ std = self.global_cmvn["std"]
+
+ x = np.subtract(input, mean)
+ x = np.divide(x, std)
+ return x
+
+
+@entrypoint
+class StreamSpeechS2TTAgent(SpeechToTextAgent):
+ """
+ Incrementally feed text to this offline Fastspeech2 TTS model,
+ with a minimum numbers of phonemes every chunk.
+ """
+
+ def __init__(self, args):
+ super().__init__(args)
+ self.eos = DEFAULT_EOS
+
+ self.gpu = self.args.device == "gpu"
+ self.device = "cuda" if args.device == "gpu" else "cpu"
+
+ self.args = args
+
+ self.load_model_vocab(args)
+
+ self.max_len = args.max_len
+
+ self.force_finish = args.force_finish
+
+ torch.set_grad_enabled(False)
+
+ tgt_dict_mt = self.dict[f"{self.models[0].mt_task_name}"]
+ tgt_dict = self.dict["tgt"]
+ tgt_dict_asr = self.dict["source_unigram"]
+ tgt_dict_st = self.dict["ctc_target_unigram"]
+ args.user_dir=args.agent_dir
+ utils.import_user_module(args)
+ from agent.sequence_generator import SequenceGenerator
+ from agent.ctc_generator import CTCSequenceGenerator
+ from agent.ctc_decoder import CTCDecoder
+ from agent.tts.vocoder import CodeHiFiGANVocoderWithDur
+
+ self.ctc_generator = CTCSequenceGenerator(
+ tgt_dict, self.models, use_incremental_states=True
+ )
+
+ self.asr_ctc_generator = CTCDecoder(tgt_dict_asr, self.models)
+ self.st_ctc_generator = CTCDecoder(tgt_dict_st, self.models)
+
+ self.generator = SequenceGenerator(
+ self.models,
+ tgt_dict,
+ beam_size=1,
+ max_len_a=1,
+ max_len_b=200,
+ max_len=0,
+ min_len=1,
+ normalize_scores=True,
+ len_penalty=1.0,
+ unk_penalty=0.0,
+ temperature=1.0,
+ match_source_len=False,
+ no_repeat_ngram_size=0,
+ search_strategy=search.BeamSearch(tgt_dict),
+ eos=tgt_dict.eos(),
+ symbols_to_strip_from_output=None,
+ )
+
+ self.generator_mt = SequenceGenerator(
+ self.models,
+ tgt_dict_mt,
+ beam_size=1,
+ max_len_a=1,
+ max_len_b=200,
+ max_len=0,
+ min_len=1,
+ normalize_scores=True,
+ len_penalty=1.0,
+ unk_penalty=0.0,
+ temperature=1.0,
+ match_source_len=False,
+ no_repeat_ngram_size=0,
+ search_strategy=search.BeamSearch(tgt_dict_mt),
+ eos=tgt_dict_mt.eos(),
+ symbols_to_strip_from_output=None,
+ use_incremental_states=True,
+ )
+ self.lagging_k1 = args.lagging_k1
+ self.lagging_k2 = args.lagging_k2
+ self.segment_size = args.segment_size
+ self.stride_n = args.stride_n
+ self.unit_per_subword = args.unit_per_subword
+ self.stride_n2 = args.stride_n2
+ if args.extra_output_dir is not None:
+ self.asr_file = Path(args.extra_output_dir + "/asr.txt")
+ self.st_file = Path(args.extra_output_dir + "/st.txt")
+ self.unit_file = Path(args.extra_output_dir + "/unit.txt")
+ self.quiet = False
+ else:
+ self.quiet = True
+
+ self.reset()
+
+ @staticmethod
+ def add_args(parser):
+ parser.add_argument(
+ "--model-path",
+ type=str,
+ required=True,
+ help="path to your pretrained model.",
+ )
+ parser.add_argument(
+ "--data-bin", type=str, required=True, help="Path of data binary"
+ )
+ parser.add_argument(
+ "--config-yaml", type=str, default=None, help="Path to config yaml file"
+ )
+ parser.add_argument(
+ "--multitask-config-yaml",
+ type=str,
+ default=None,
+ help="Path to config yaml file",
+ )
+ parser.add_argument(
+ "--global-stats",
+ type=str,
+ default=None,
+ help="Path to json file containing cmvn stats",
+ )
+ parser.add_argument(
+ "--tgt-splitter-type",
+ type=str,
+ default="SentencePiece",
+ help="Subword splitter type for target text",
+ )
+ parser.add_argument(
+ "--tgt-splitter-path",
+ type=str,
+ default=None,
+ help="Subword splitter model path for target text",
+ )
+ parser.add_argument(
+ "--user-dir",
+ type=str,
+ default="researches/ctc_unity",
+ help="User directory for model",
+ )
+ parser.add_argument(
+ "--agent-dir",
+ type=str,
+ default="agent",
+ help="User directory for agents",
+ )
+ parser.add_argument(
+ "--max-len", type=int, default=200, help="Max length of translation"
+ )
+ parser.add_argument(
+ "--force-finish",
+ default=False,
+ action="store_true",
+ help="Force the model to finish the hypothsis if the source is not finished",
+ )
+ parser.add_argument(
+ "--shift-size",
+ type=int,
+ default=SHIFT_SIZE,
+ help="Shift size of feature extraction window.",
+ )
+ parser.add_argument(
+ "--window-size",
+ type=int,
+ default=WINDOW_SIZE,
+ help="Window size of feature extraction window.",
+ )
+ parser.add_argument(
+ "--sample-rate", type=int, default=ORG_SAMPLE_RATE, help="Sample rate"
+ )
+ parser.add_argument(
+ "--feature-dim",
+ type=int,
+ default=FEATURE_DIM,
+ help="Acoustic feature dimension.",
+ )
+ parser.add_argument("--lagging-k1", type=int, default=0, help="lagging number")
+ parser.add_argument("--lagging-k2", type=int, default=0, help="lagging number")
+ parser.add_argument(
+ "--segment-size", type=int, default=320, help="segment-size"
+ )
+ parser.add_argument("--stride-n", type=int, default=1, help="lagging number")
+ parser.add_argument("--stride-n2", type=int, default=1, help="lagging number")
+ parser.add_argument(
+ "--unit-per-subword", type=int, default=15, help="lagging number"
+ )
+ parser.add_argument(
+ "--extra-output-dir", type=str, default=None, help="extra output dir"
+ )
+
+ def reset(self):
+ self.src_seg_num = 0
+ self.tgt_subwords_indices = None
+ self.src_ctc_indices = None
+ self.src_ctc_prefix_length = 0
+ self.tgt_ctc_prefix_length = 0
+ self.tgt_units_indices = None
+ self.prev_output_tokens_mt = None
+ self.tgt_text = ""
+ self.mt_decoder_out = None
+ self.unit = None
+ self.wav = []
+ self.post_transcription = ""
+ self.unfinished_wav = None
+ self.states.reset()
+ try:
+ self.generator_mt.reset_incremental_states()
+ self.ctc_generator.reset_incremental_states()
+ except:
+ pass
+
+ def to_device(self, tensor):
+ if self.gpu:
+ return tensor.cuda()
+ else:
+ return tensor.cpu()
+
+ def load_model_vocab(self, args):
+ filename = args.model_path
+ if not os.path.exists(filename):
+ raise IOError("Model file not found: {}".format(filename))
+
+ state = checkpoint_utils.load_checkpoint_to_cpu(filename)
+ state["cfg"].common['user_dir']=args.user_dir
+ utils.import_user_module(state["cfg"].common)
+
+ task_args = state["cfg"]["task"]
+ task_args.data = args.data_bin
+
+ args.global_cmvn = None
+ if args.config_yaml is not None:
+ task_args.config_yaml = args.config_yaml
+ with open(os.path.join(args.data_bin, args.config_yaml), "r") as f:
+ config = yaml.load(f, Loader=yaml.BaseLoader)
+
+ if "global_cmvn" in config:
+ args.global_cmvn = np.load(config["global_cmvn"]["stats_npz_path"])
+
+ self.feature_extractor = OnlineFeatureExtractor(args, config)
+
+ if args.multitask_config_yaml is not None:
+ task_args.multitask_config_yaml = args.multitask_config_yaml
+
+ task = tasks.setup_task(task_args)
+ self.task = task
+
+ overrides = ast.literal_eval(state["cfg"].common_eval.model_overrides)
+
+ models, saved_cfg = checkpoint_utils.load_model_ensemble(
+ utils.split_paths(filename),
+ arg_overrides=overrides,
+ task=task,
+ suffix=state["cfg"].checkpoint.checkpoint_suffix,
+ strict=(state["cfg"].checkpoint.checkpoint_shard_count == 1),
+ num_shards=state["cfg"].checkpoint.checkpoint_shard_count,
+ )
+
+ chunk_size = args.source_segment_size // 40
+
+ self.models = models
+
+ for model in self.models:
+ model.eval()
+ model.share_memory()
+ if self.gpu:
+ model.cuda()
+ model.encoder.chunk_size = chunk_size
+ chunk_size = min(chunk_size, 16)
+ for conv in model.encoder.subsample.conv_layers:
+ conv.chunk_size = chunk_size
+ for layer in model.encoder.conformer_layers:
+ layer.conv_module.depthwise_conv.chunk_size = chunk_size
+
+ # Set dictionary
+ self.dict = {}
+ self.dict["tgt"] = task.target_dictionary
+
+ for k, v in task.multitask_tasks.items():
+ self.dict[k] = v.tgt_dict
+
+ @torch.inference_mode()
+ def policy(self):
+
+ feature = self.feature_extractor(self.states.source)
+ if feature.size(0) == 0 and not self.states.source_finished:
+ return ReadAction()
+
+ src_indices = feature.unsqueeze(0)
+ src_lengths = torch.tensor([feature.size(0)], device=self.device).long()
+ self.encoder_outs = self.generator.model.forward_encoder(
+ {"src_tokens": src_indices, "src_lengths": src_lengths}
+ )
+
+ finalized_asr = self.asr_ctc_generator.generate(
+ self.encoder_outs[0], aux_task_name="source_unigram"
+ )
+ asr_probs = torch.exp(finalized_asr[0][0]["lprobs"])
+
+ for i, hypo in enumerate(finalized_asr):
+ i_beam = 0
+ tmp = hypo[i_beam]["tokens"].int()
+ src_ctc_indices = tmp
+ src_ctc_index = hypo[i_beam]["index"]
+ text = "".join([self.dict["source_unigram"][c] for c in tmp])
+ text = text.replace("_", " ")
+ text = text.replace("▁", " ")
+ text = text.replace("", " ")
+ text = text.replace("", "")
+ text = text.replace("", "")
+ if len(text) > 0 and text[0] == " ":
+ text = text[1:]
+ if self.states.source_finished and not self.quiet:
+ with open(self.asr_file, "a") as file:
+ print(text, file=file)
+
+ finalized_st = self.st_ctc_generator.generate(
+ self.encoder_outs[0], aux_task_name="ctc_target_unigram"
+ )
+ st_probs = torch.exp(finalized_st[0][0]["lprobs"])
+
+ for i, hypo in enumerate(finalized_st):
+ i_beam = 0
+ tmp = hypo[i_beam]["tokens"].int()
+ tgt_ctc_indices = tmp
+ tgt_ctc_index = hypo[i_beam]["index"]
+
+ text = "".join([self.dict["ctc_target_unigram"][c] for c in tmp])
+ text = text.replace("_", " ")
+ text = text.replace("▁", " ")
+ text = text.replace("", " ")
+ text = text.replace("", "")
+ text = text.replace("", "")
+ if len(text) > 0 and text[0] == " ":
+ text = text[1:]
+
+ if not self.states.source_finished:
+ src_ctc_prefix_length = src_ctc_indices.size(-1)
+ tgt_ctc_prefix_length = tgt_ctc_indices.size(-1)
+
+ self.src_ctc_indices = src_ctc_indices
+ if (
+ src_ctc_prefix_length < self.src_ctc_prefix_length + self.stride_n
+ or tgt_ctc_prefix_length < self.tgt_ctc_prefix_length + self.stride_n
+ ):
+ return ReadAction()
+ self.src_ctc_prefix_length = max(
+ src_ctc_prefix_length, self.src_ctc_prefix_length
+ )
+ self.tgt_ctc_prefix_length = max(
+ tgt_ctc_prefix_length, self.tgt_ctc_prefix_length
+ )
+ subword_tokens = (
+ (tgt_ctc_prefix_length - self.lagging_k1) // self.stride_n
+ ) * self.stride_n
+ new_subword_tokens = (
+ (subword_tokens - self.tgt_subwords_indices.size(-1))
+ if self.tgt_subwords_indices is not None
+ else subword_tokens
+ )
+
+ if new_subword_tokens < 1:
+ return ReadAction()
+
+ else:
+ self.src_ctc_indices = src_ctc_indices
+ new_subword_tokens = -1
+
+ new_subword_tokens = int(new_subword_tokens)
+
+ single_model = self.generator.model.single_model
+ mt_decoder = getattr(single_model, f"{single_model.mt_task_name}_decoder")
+
+ # 1. MT decoder
+ finalized_mt = self.generator_mt.generate_decoder(
+ self.encoder_outs,
+ src_indices,
+ src_lengths,
+ {
+ "id": 1,
+ "net_input": {"src_tokens": src_indices, "src_lengths": src_lengths},
+ },
+ self.tgt_subwords_indices,
+ None,
+ None,
+ aux_task_name=single_model.mt_task_name,
+ max_new_tokens=new_subword_tokens,
+ )
+ if finalized_mt[0][0]["tokens"][-1] == 2:
+ tgt_subwords_indices = finalized_mt[0][0]["tokens"][:-1].unsqueeze(0)
+ else:
+ tgt_subwords_indices = finalized_mt[0][0]["tokens"].unsqueeze(0)
+
+ max_tgt_len = max([len(hypo[0]["tokens"]) for hypo in finalized_mt])
+ prev_output_tokens_mt = (
+ src_indices.new_zeros(src_indices.shape[0], max_tgt_len)
+ .fill_(mt_decoder.padding_idx)
+ .int()
+ ) # B x T
+
+ for i, hypo in enumerate(finalized_mt):
+ i_beam = 0
+ tmp = hypo[i_beam]["tokens"].int()
+ prev_output_tokens_mt[i, 0] = self.generator_mt.eos
+ if tmp[-1] == self.generator_mt.eos:
+ tmp = tmp[:-1]
+ prev_output_tokens_mt[i, 1 : len(tmp) + 1] = tmp
+
+ tokens = [self.generator_mt.tgt_dict[c] for c in tmp]
+
+ text = "".join(tokens)
+ text = text.replace("_", " ")
+ text = text.replace("▁", " ")
+ text = text.replace("", " ")
+ text = text.replace("", "")
+ text = text.replace("", "")
+ if len(text) > 0 and text[0] == " ":
+ text = text[1:]
+ if self.states.source_finished and not self.quiet:
+ with open(self.st_file, "a") as file:
+ print(text, file=file)
+
+ if self.tgt_subwords_indices is not None and torch.equal(
+ self.tgt_subwords_indices, tgt_subwords_indices
+ ):
+ if not self.states.source_finished:
+ return ReadAction()
+ else:
+ return WriteAction(
+ "",
+ finished=True,
+ )
+ self.tgt_subwords_indices = tgt_subwords_indices
+
+ text = " ".join(tokens)
+ new_text = text[len(self.tgt_text) :]
+
+ self.tgt_text = text
+
+ if self.states.source_finished and new_subword_tokens == -1:
+ self.states.target_finished = True
+ self.reset()
+
+ return WriteAction(
+ new_text,
+ finished=self.states.target_finished,
+ )
diff --git a/agent/speech_to_text.wait-k-stride-n.agent.py b/agent/speech_to_text.wait-k-stride-n.agent.py
new file mode 100644
index 0000000000000000000000000000000000000000..e8e14ad0edf9ac6b038f439e8f7567f492ce56de
--- /dev/null
+++ b/agent/speech_to_text.wait-k-stride-n.agent.py
@@ -0,0 +1,474 @@
+##########################################
+# Simultaneous Speech-to-Text Translation Agent for Wait-k Policy
+#
+# StreamSpeech: Simultaneous Speech-to-Speech Translation with Multi-task Learning (ACL 2024)
+##########################################
+
+from simuleval.utils import entrypoint
+from simuleval.data.segments import SpeechSegment
+from simuleval.agents import SpeechToTextAgent
+from simuleval.agents.actions import WriteAction, ReadAction
+from fairseq.checkpoint_utils import load_model_ensemble_and_task
+from fairseq.models.text_to_speech.hub_interface import TTSHubInterface
+from pathlib import Path
+from typing import Any, Dict, Optional, Union
+from fairseq.data.audio.audio_utils import convert_waveform
+from examples.speech_to_text.data_utils import extract_fbank_features
+import ast
+import math
+import os
+import json
+import numpy as np
+import torch
+import torchaudio.compliance.kaldi as kaldi
+import yaml
+from fairseq import checkpoint_utils, tasks, utils, options
+from fairseq.file_io import PathManager
+from fairseq import search
+from fairseq.data.audio.feature_transforms import CompositeAudioFeatureTransform
+from fairseq.models.text_to_speech.vocoder import CodeHiFiGANVocoder
+
+SHIFT_SIZE = 10
+WINDOW_SIZE = 25
+ORG_SAMPLE_RATE = 48000
+SAMPLE_RATE = 16000
+FEATURE_DIM = 80
+BOW_PREFIX = "\u2581"
+DEFAULT_EOS = 2
+
+
+class OnlineFeatureExtractor:
+ """
+ Extract speech feature on the fly.
+ """
+
+ def __init__(self, args, cfg):
+ self.shift_size = args.shift_size
+ self.window_size = args.window_size
+ assert self.window_size >= self.shift_size
+
+ self.sample_rate = args.sample_rate
+ self.feature_dim = args.feature_dim
+ self.num_samples_per_shift = int(self.shift_size * self.sample_rate / 1000)
+ self.num_samples_per_window = int(self.window_size * self.sample_rate / 1000)
+ self.len_ms_to_samples = lambda x: x * self.sample_rate / 1000
+ self.previous_residual_samples = []
+ self.global_cmvn = args.global_cmvn
+ self.device = "cuda" if args.device == "gpu" else "cpu"
+ self.feature_transforms = CompositeAudioFeatureTransform.from_config_dict(
+ {"feature_transforms": ["utterance_cmvn"]}
+ )
+
+ def clear_cache(self):
+ self.previous_residual_samples = []
+
+ def __call__(self, new_samples, sr=ORG_SAMPLE_RATE):
+ samples = new_samples
+ # # num_frames is the number of frames from the new segment
+ num_frames = math.floor(
+ (len(samples) - self.len_ms_to_samples(self.window_size - self.shift_size))
+ / self.num_samples_per_shift
+ )
+
+ # # the number of frames used for feature extraction
+ # # including some part of thte previous segment
+ effective_num_samples = int(
+ num_frames * self.len_ms_to_samples(self.shift_size)
+ + self.len_ms_to_samples(self.window_size - self.shift_size)
+ )
+ samples = samples[:effective_num_samples]
+ waveform, sample_rate = convert_waveform(
+ torch.tensor([samples]), sr, to_mono=True, to_sample_rate=16000
+ )
+ output = extract_fbank_features(waveform, 16000)
+ output = self.transform(output)
+ return torch.tensor(output, device=self.device)
+
+ def transform(self, input):
+ if self.global_cmvn is None:
+ return input
+
+ mean = self.global_cmvn["mean"]
+ std = self.global_cmvn["std"]
+
+ x = np.subtract(input, mean)
+ x = np.divide(x, std)
+ return x
+
+
+@entrypoint
+class EnglishSpeechCounter(SpeechToTextAgent):
+ """
+ Incrementally feed text to this offline Fastspeech2 TTS model,
+ with a minimum numbers of phonemes every chunk.
+ """
+
+ def __init__(self, args):
+ super().__init__(args)
+ self.eos = DEFAULT_EOS
+
+ self.gpu = self.args.device == "gpu"
+ self.device = "cuda" if args.device == "gpu" else "cpu"
+
+ self.args = args
+
+ self.load_model_vocab(args)
+
+ self.max_len = args.max_len
+
+ self.force_finish = args.force_finish
+
+ torch.set_grad_enabled(False)
+
+ tgt_dict_mt = self.dict[f"{self.models[0].mt_task_name}"]
+ tgt_dict = self.dict["tgt"]
+ args.user_dir=args.agent_dir
+ utils.import_user_module(args)
+ from agent.sequence_generator import SequenceGenerator
+
+ self.generator = SequenceGenerator(
+ self.models,
+ tgt_dict,
+ beam_size=1,
+ max_len_a=1,
+ max_len_b=200,
+ max_len=0,
+ min_len=1,
+ normalize_scores=True,
+ len_penalty=1.0,
+ unk_penalty=0.0,
+ temperature=1.0,
+ match_source_len=False,
+ no_repeat_ngram_size=0,
+ search_strategy=search.BeamSearch(tgt_dict),
+ eos=tgt_dict.eos(),
+ symbols_to_strip_from_output=None,
+ )
+
+ self.generator_mt = SequenceGenerator(
+ self.models,
+ tgt_dict_mt,
+ beam_size=1,
+ max_len_a=1,
+ max_len_b=200,
+ max_len=0,
+ min_len=1,
+ normalize_scores=True,
+ len_penalty=1.0,
+ unk_penalty=0.0,
+ temperature=1.0,
+ match_source_len=False,
+ no_repeat_ngram_size=0,
+ search_strategy=search.BeamSearch(tgt_dict_mt),
+ eos=tgt_dict_mt.eos(),
+ symbols_to_strip_from_output=None,
+ )
+
+ from agent.tts.vocoder import CodeHiFiGANVocoderWithDur
+
+ with open(args.vocoder_cfg) as f:
+ vocoder_cfg = json.load(f)
+ self.vocoder = CodeHiFiGANVocoderWithDur(args.vocoder, vocoder_cfg)
+ if self.device == "cuda":
+ self.vocoder = self.vocoder.cuda()
+ self.dur_prediction = args.dur_prediction
+
+ self.lagging_k1 = args.lagging_k1
+ self.lagging_k2 = args.lagging_k2
+ self.segment_size = args.segment_size
+ self.stride_n = args.stride_n
+ self.unit_per_subword = args.unit_per_subword
+ self.stride_n2 = args.stride_n2
+
+ self.reset()
+
+ @staticmethod
+ def add_args(parser):
+ parser.add_argument(
+ "--model-path",
+ type=str,
+ required=True,
+ help="path to your pretrained model.",
+ )
+ parser.add_argument(
+ "--data-bin", type=str, required=True, help="Path of data binary"
+ )
+ parser.add_argument(
+ "--config-yaml", type=str, default=None, help="Path to config yaml file"
+ )
+ parser.add_argument(
+ "--multitask-config-yaml",
+ type=str,
+ default=None,
+ help="Path to config yaml file",
+ )
+ parser.add_argument(
+ "--global-stats",
+ type=str,
+ default=None,
+ help="Path to json file containing cmvn stats",
+ )
+ parser.add_argument(
+ "--tgt-splitter-type",
+ type=str,
+ default="SentencePiece",
+ help="Subword splitter type for target text",
+ )
+ parser.add_argument(
+ "--tgt-splitter-path",
+ type=str,
+ default=None,
+ help="Subword splitter model path for target text",
+ )
+ parser.add_argument(
+ "--user-dir",
+ type=str,
+ default="researches/ctc_unity",
+ help="User directory for model",
+ )
+ parser.add_argument(
+ "--agent-dir",
+ type=str,
+ default="agent",
+ help="User directory for agents",
+ )
+ parser.add_argument(
+ "--max-len", type=int, default=200, help="Max length of translation"
+ )
+ parser.add_argument(
+ "--force-finish",
+ default=False,
+ action="store_true",
+ help="Force the model to finish the hypothsis if the source is not finished",
+ )
+ parser.add_argument(
+ "--shift-size",
+ type=int,
+ default=SHIFT_SIZE,
+ help="Shift size of feature extraction window.",
+ )
+ parser.add_argument(
+ "--window-size",
+ type=int,
+ default=WINDOW_SIZE,
+ help="Window size of feature extraction window.",
+ )
+ parser.add_argument(
+ "--sample-rate", type=int, default=SAMPLE_RATE, help="Sample rate"
+ )
+ parser.add_argument(
+ "--feature-dim",
+ type=int,
+ default=FEATURE_DIM,
+ help="Acoustic feature dimension.",
+ )
+ parser.add_argument(
+ "--vocoder", type=str, required=True, help="path to the CodeHiFiGAN vocoder"
+ )
+ parser.add_argument(
+ "--vocoder-cfg",
+ type=str,
+ required=True,
+ help="path to the CodeHiFiGAN vocoder config",
+ )
+ parser.add_argument(
+ "--dur-prediction",
+ action="store_true",
+ help="enable duration prediction (for reduced/unique code sequences)",
+ )
+ parser.add_argument("--lagging-k1", type=int, default=3, help="lagging number")
+ parser.add_argument("--lagging-k2", type=int, default=1, help="lagging number")
+ parser.add_argument(
+ "--segment-size", type=int, default=320, help="segment-size"
+ )
+ parser.add_argument("--stride-n", type=int, default=1, help="lagging number")
+ parser.add_argument("--stride-n2", type=int, default=1, help="lagging number")
+ parser.add_argument(
+ "--unit-per-subword", type=int, default=15, help="lagging number"
+ )
+
+ def reset(self):
+ self.src_seg_num = 0
+ self.tgt_subwords_indices = None
+ self.tgt_units_indices = None
+ self.tgt_text = ""
+ self.unit = None
+ self.wav = []
+ self.states.reset()
+
+ def to_device(self, tensor):
+ if self.gpu:
+ return tensor.cuda()
+ else:
+ return tensor.cpu()
+
+ def load_model_vocab(self, args):
+ filename = args.model_path
+ if not os.path.exists(filename):
+ raise IOError("Model file not found: {}".format(filename))
+
+ state = checkpoint_utils.load_checkpoint_to_cpu(filename)
+ state["cfg"].common['user_dir']=args.user_dir
+ utils.import_user_module(state["cfg"].common)
+
+ task_args = state["cfg"]["task"]
+ task_args.data = args.data_bin
+
+ args.global_cmvn = None
+ if args.config_yaml is not None:
+ task_args.config_yaml = args.config_yaml
+ with open(os.path.join(args.data_bin, args.config_yaml), "r") as f:
+ config = yaml.load(f, Loader=yaml.BaseLoader)
+
+ if "global_cmvn" in config:
+ args.global_cmvn = np.load(config["global_cmvn"]["stats_npz_path"])
+
+ self.feature_extractor = OnlineFeatureExtractor(args, config)
+
+ if args.multitask_config_yaml is not None:
+ task_args.multitask_config_yaml = args.multitask_config_yaml
+
+ task = tasks.setup_task(task_args)
+
+ overrides = ast.literal_eval(state["cfg"].common_eval.model_overrides)
+
+ models, saved_cfg = checkpoint_utils.load_model_ensemble(
+ utils.split_paths(filename),
+ arg_overrides=overrides,
+ task=task,
+ suffix=state["cfg"].checkpoint.checkpoint_suffix,
+ strict=(state["cfg"].checkpoint.checkpoint_shard_count == 1),
+ num_shards=state["cfg"].checkpoint.checkpoint_shard_count,
+ )
+ self.models = models
+
+ for model in self.models:
+ model.eval()
+ model.share_memory()
+ if self.gpu:
+ model.cuda()
+
+ # Set dictionary
+ self.dict = {}
+ self.dict["tgt"] = task.target_dictionary
+
+ for k, v in task.multitask_tasks.items():
+ self.dict[k] = v.tgt_dict
+
+ def policy(self):
+ src_seg_num = len(self.states.source) // (
+ self.segment_size * ORG_SAMPLE_RATE / 1000
+ )
+ cur_tgt_subword_tokens = (
+ self.tgt_subwords_indices.size(-1)
+ if self.tgt_subwords_indices is not None
+ else 0
+ )
+ cur_tgt_unit_tokens = (
+ self.tgt_units_indices.size(-1) if self.tgt_units_indices is not None else 0
+ )
+ if (
+ src_seg_num <= self.src_seg_num or src_seg_num < self.lagging_k1
+ ) and not self.states.source_finished:
+ return ReadAction()
+ else:
+ self.src_seg_num = src_seg_num
+
+ subword_tokens = (
+ (self.src_seg_num - self.lagging_k1) // self.stride_n
+ ) * self.stride_n
+ unit_tokens = (
+ ((subword_tokens - self.lagging_k2) // self.stride_n2)
+ * self.stride_n2
+ * self.unit_per_subword
+ )
+
+ new_subword_tokens = (
+ (subword_tokens - self.tgt_subwords_indices.size(-1))
+ if self.tgt_subwords_indices is not None
+ else subword_tokens
+ )
+ new_unit_tokens = (
+ (unit_tokens - self.tgt_units_indices.size(-1))
+ if self.tgt_units_indices is not None
+ else unit_tokens
+ )
+
+ if (
+ new_subword_tokens < 1 or new_unit_tokens < 1
+ ) and not self.states.source_finished:
+ return ReadAction()
+ if self.states.source_finished:
+ new_subword_tokens = -1
+ new_unit_tokens = -1
+ new_subword_tokens = int(new_subword_tokens)
+ new_unit_tokens = int(new_unit_tokens)
+
+ feature = self.feature_extractor(self.states.source)
+ if feature.size(0) == 0 and not self.states.source_finished:
+ return ReadAction()
+
+ src_indices = feature.unsqueeze(0)
+ src_lengths = torch.tensor([feature.size(0)], device=self.device).long()
+ self.encoder_outs = self.generator.model.forward_encoder(
+ {"src_tokens": src_indices, "src_lengths": src_lengths}
+ )
+
+ single_model = self.generator.model.single_model
+ mt_decoder = getattr(single_model, f"{single_model.mt_task_name}_decoder")
+
+ # 1. MT decoder
+ finalized_mt = self.generator_mt.generate_decoder(
+ self.encoder_outs,
+ src_indices,
+ src_lengths,
+ {
+ "id": 1,
+ "net_input": {"src_tokens": src_indices, "src_lengths": src_lengths},
+ },
+ self.tgt_subwords_indices,
+ None,
+ None,
+ aux_task_name=single_model.mt_task_name,
+ max_new_tokens=new_subword_tokens,
+ )
+ self.tgt_subwords_indices = finalized_mt[0][0]["tokens"][:-1].unsqueeze(0)
+
+ max_tgt_len = max([len(hypo[0]["tokens"]) for hypo in finalized_mt])
+ prev_output_tokens_mt = (
+ src_indices.new_zeros(src_indices.shape[0], max_tgt_len)
+ .fill_(mt_decoder.padding_idx)
+ .int()
+ )
+
+ for i, hypo in enumerate(finalized_mt):
+ i_beam = 0
+ tmp = hypo[i_beam]["tokens"].int()
+ prev_output_tokens_mt[i, 0] = self.generator_mt.eos
+ if tmp[-1] == self.generator_mt.eos:
+ tmp = tmp[:-1]
+ prev_output_tokens_mt[i, 1 : len(tmp) + 1] = tmp
+ tokens = [self.generator_mt.tgt_dict[c] for c in tmp]
+
+ text = "".join([self.generator_mt.tgt_dict[c] for c in tmp])
+ text = text.replace("_", " ")
+ text = text.replace("▁", " ")
+ text = text.replace("", " ")
+ text = text.replace("", "")
+ text = text.replace("", "")
+ if len(text) > 0 and text[0] == " ":
+ text = text[1:]
+
+ text = " ".join(tokens)
+ new_text = text[len(self.tgt_text) :]
+
+ self.tgt_text = text
+
+ if self.states.source_finished and new_subword_tokens == -1:
+ self.states.target_finished = True
+ self.reset()
+
+ return WriteAction(
+ new_text,
+ finished=self.states.target_finished,
+ )
diff --git a/agent/tts/__init__.py b/agent/tts/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..68dd42cdf90f52251a73ced8816979a6f48b289a
--- /dev/null
+++ b/agent/tts/__init__.py
@@ -0,0 +1,8 @@
+import os
+import importlib
+
+# automatically import any Python files in the criterions/ directory
+for file in os.listdir(os.path.dirname(__file__)):
+ if file.endswith(".py") and not file.startswith("_"):
+ file_name = file[: file.find(".py")]
+ importlib.import_module("agent.tts." + file_name)
diff --git a/agent/tts/codehifigan.py b/agent/tts/codehifigan.py
new file mode 100644
index 0000000000000000000000000000000000000000..19bb411b2ab1f20a93233aefde00f6c577f5752c
--- /dev/null
+++ b/agent/tts/codehifigan.py
@@ -0,0 +1,95 @@
+from argparse import Namespace
+import torch
+import torch.nn as nn
+
+from fairseq.models.text_to_speech.fastspeech2 import VariancePredictor
+from fairseq.models.text_to_speech.hifigan import Generator
+
+
+class CodeGenerator(Generator):
+ def __init__(self, cfg):
+ super().__init__(cfg)
+ self.dict = nn.Embedding(cfg["num_embeddings"], cfg["embedding_dim"])
+ self.multispkr = cfg.get("multispkr", None)
+ self.embedder = cfg.get("embedder_params", None)
+
+ if self.multispkr and not self.embedder:
+ self.spkr = nn.Embedding(cfg.get("num_speakers", 200), cfg["embedding_dim"])
+ elif self.embedder:
+ self.spkr = nn.Linear(cfg.get("embedder_dim", 256), cfg["embedding_dim"])
+
+ self.dur_predictor = None
+ if cfg.get("dur_predictor_params", None):
+ self.dur_predictor = VariancePredictor(
+ Namespace(**cfg["dur_predictor_params"])
+ )
+
+ self.f0 = cfg.get("f0", None)
+ n_f0_bin = cfg.get("f0_quant_num_bin", 0)
+ self.f0_quant_embed = (
+ None if n_f0_bin <= 0 else nn.Embedding(n_f0_bin, cfg["embedding_dim"])
+ )
+
+ @staticmethod
+ def _upsample(signal, max_frames):
+ if signal.dim() == 3:
+ bsz, channels, cond_length = signal.size()
+ elif signal.dim() == 2:
+ signal = signal.unsqueeze(2)
+ bsz, channels, cond_length = signal.size()
+ else:
+ signal = signal.view(-1, 1, 1)
+ bsz, channels, cond_length = signal.size()
+
+ signal = signal.unsqueeze(3).repeat(1, 1, 1, max_frames // cond_length)
+
+ # pad zeros as needed (if signal's shape does not divide completely with max_frames)
+ reminder = (max_frames - signal.shape[2] * signal.shape[3]) // signal.shape[3]
+ if reminder > 0:
+ raise NotImplementedError(
+ "Padding condition signal - misalignment between condition features."
+ )
+
+ signal = signal.view(bsz, channels, max_frames)
+ return signal
+
+ def forward(self, **kwargs):
+ x = self.dict(kwargs["code"]).transpose(1, 2)
+ dur_out = None
+ if self.dur_predictor and kwargs.get("dur_prediction", False):
+ assert x.size(0) == 1, "only support single sample"
+ log_dur_pred = self.dur_predictor(x.transpose(1, 2))
+ dur_out = torch.clamp(
+ torch.round((torch.exp(log_dur_pred) - 1)).long(), min=1
+ )
+ # B x C x T
+ x = torch.repeat_interleave(x, dur_out.view(-1), dim=2)
+
+ if self.f0:
+ if self.f0_quant_embed:
+ kwargs["f0"] = self.f0_quant_embed(kwargs["f0"].long()).transpose(1, 2)
+ else:
+ kwargs["f0"] = kwargs["f0"].unsqueeze(1)
+
+ if x.shape[-1] < kwargs["f0"].shape[-1]:
+ x = self._upsample(x, kwargs["f0"].shape[-1])
+ elif x.shape[-1] > kwargs["f0"].shape[-1]:
+ kwargs["f0"] = self._upsample(kwargs["f0"], x.shape[-1])
+ x = torch.cat([x, kwargs["f0"]], dim=1)
+
+ if self.multispkr:
+ assert (
+ "spkr" in kwargs
+ ), 'require "spkr" input for multispeaker CodeHiFiGAN vocoder'
+ spkr = self.spkr(kwargs["spkr"]).transpose(1, 2)
+ spkr = self._upsample(spkr, x.shape[-1])
+ x = torch.cat([x, spkr], dim=1)
+
+ for k, feat in kwargs.items():
+ if k in ["spkr", "code", "f0", "dur_prediction"]:
+ continue
+
+ feat = self._upsample(feat, x.shape[-1])
+ x = torch.cat([x, feat], dim=1)
+
+ return super().forward(x), dur_out
diff --git a/agent/tts/vocoder.py b/agent/tts/vocoder.py
new file mode 100644
index 0000000000000000000000000000000000000000..d4461b3cb332feb513e991e25f949cfaf33cbd68
--- /dev/null
+++ b/agent/tts/vocoder.py
@@ -0,0 +1,108 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import json
+import logging
+from typing import Dict
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from fairseq.data.audio.audio_utils import (
+ TTSSpectrogram,
+ get_fourier_basis,
+ get_mel_filters,
+ get_window,
+)
+from fairseq.data.audio.speech_to_text_dataset import S2TDataConfig
+from fairseq.models import BaseFairseqModel, register_model
+from agent.tts.codehifigan import CodeGenerator as CodeHiFiGANModel
+from fairseq.models.text_to_speech.hifigan import Generator as HiFiGANModel
+from fairseq.models.text_to_speech.hub_interface import VocoderHubInterface
+
+logger = logging.getLogger(__name__)
+
+
+@register_model("CodeHiFiGANVocoderWithDur")
+class CodeHiFiGANVocoderWithDur(BaseFairseqModel):
+ def __init__(
+ self, checkpoint_path: str, model_cfg: Dict[str, str], fp16: bool = False
+ ) -> None:
+ super().__init__()
+ self.model = CodeHiFiGANModel(model_cfg)
+ if torch.cuda.is_available():
+ state_dict = torch.load(checkpoint_path)
+ else:
+ state_dict = torch.load(checkpoint_path, map_location=torch.device("cpu"))
+ self.model.load_state_dict(state_dict["generator"])
+ self.model.eval()
+ if fp16:
+ self.model.half()
+ self.model.remove_weight_norm()
+ logger.info(f"loaded CodeHiFiGAN checkpoint from {checkpoint_path}")
+
+ def forward(self, x: Dict[str, torch.Tensor], dur_prediction=False) -> torch.Tensor:
+ assert "code" in x
+ x["dur_prediction"] = dur_prediction
+
+ # remove invalid code
+ mask = x["code"] >= 0
+ x["code"] = x["code"][mask].unsqueeze(dim=0)
+ if "f0" in x:
+ f0_up_ratio = x["f0"].size(1) // x["code"].size(1)
+ mask = mask.unsqueeze(2).repeat(1, 1, f0_up_ratio).view(-1, x["f0"].size(1))
+ x["f0"] = x["f0"][mask].unsqueeze(dim=0)
+ wav, dur = self.model(**x)
+ return wav.detach().squeeze(), dur
+
+ @classmethod
+ def from_data_cfg(cls, args, data_cfg):
+ vocoder_cfg = data_cfg.vocoder
+ assert vocoder_cfg is not None, "vocoder not specified in the data config"
+ with open(vocoder_cfg["config"]) as f:
+ model_cfg = json.load(f)
+ return cls(vocoder_cfg["checkpoint"], model_cfg, fp16=args.fp16)
+
+ @classmethod
+ def hub_models(cls):
+ base_url = "http://dl.fbaipublicfiles.com/fairseq/vocoder"
+ model_ids = [
+ "unit_hifigan_mhubert_vp_en_es_fr_it3_400k_layer11_km1000_lj_dur",
+ "unit_hifigan_mhubert_vp_en_es_fr_it3_400k_layer11_km1000_es_css10_dur",
+ "unit_hifigan_HK_layer12.km2500_frame_TAT-TTS",
+ ]
+ return {i: f"{base_url}/{i}.tar.gz" for i in model_ids}
+
+ @classmethod
+ def from_pretrained(
+ cls,
+ model_name_or_path,
+ checkpoint_file="model.pt",
+ data_name_or_path=".",
+ config="config.json",
+ fp16: bool = False,
+ **kwargs,
+ ):
+ from fairseq import hub_utils
+
+ x = hub_utils.from_pretrained(
+ model_name_or_path,
+ checkpoint_file,
+ data_name_or_path,
+ archive_map=cls.hub_models(),
+ config_yaml=config,
+ fp16=fp16,
+ is_vocoder=True,
+ **kwargs,
+ )
+
+ with open(f"{x['args']['data']}/{config}") as f:
+ vocoder_cfg = json.load(f)
+ assert len(x["args"]["model_path"]) == 1, "Too many vocoder models in the input"
+
+ vocoder = CodeHiFiGANVocoderWithDur(x["args"]["model_path"][0], vocoder_cfg)
+ return VocoderHubInterface(vocoder_cfg, vocoder)
diff --git a/asr_bleu/README.md b/asr_bleu/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..b808b136dc25018c9e43872e1377f12699758199
--- /dev/null
+++ b/asr_bleu/README.md
@@ -0,0 +1,34 @@
+# ASR-BLEU evaluation toolkit
+
+This toolkit provides a set of public ASR models used for evaluation of different speech-to-speech translation systems at Meta AI. It enables easier score comparisons between different system's outputs.
+
+The ASRGenerator wraps different CTC-based ASR models from HuggingFace and fairseq code bases. Torchaudio CTC decoder is built on top of it to decode given audio files.
+
+Please see `asr_model_cfgs.json` for a list of languages covered currently.
+
+The high-level pipeline is simple by design: given a lang tag, script loads the ASR model, transcribes model's predicted audio, and computes the BLEU score against provided reference translations using sacrebleu.
+
+# Dependencies
+
+Please see `requirements.txt`.
+
+# Usage examples
+
+This toolkit have been used with:
+
+* Speechmatrix project: https://github.com/facebookresearch/fairseq/tree/ust/examples/speech_matrix.
+
+* Hokkien speech-to-speech translation project: https://github.com/facebookresearch/fairseq/tree/ust/examples/hokkien.
+
+# Standalone run example
+
+High-level example, please substitute arguments per your case:
+
+```bash
+python compute_asr_bleu.py --lang \
+--audio_dirpath \
+--reference_path \
+--reference_format txt
+```
+
+For more details about arguments please see the script argparser help.
\ No newline at end of file
diff --git a/asr_bleu/__init__.py b/asr_bleu/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/asr_bleu/asr_model_cfgs.json b/asr_bleu/asr_model_cfgs.json
new file mode 100644
index 0000000000000000000000000000000000000000..e4613a7ad2b64f8cb6a71824494a74e006b2e775
--- /dev/null
+++ b/asr_bleu/asr_model_cfgs.json
@@ -0,0 +1,198 @@
+{
+ "en": {
+ "oct22": {
+ "desc": "Wav2Vec 2.0 Large (LV-60) + Self Training from https://github.com/facebookresearch/fairseq/tree/main/examples/wav2vec#pre-trained-models",
+ "ckpt_path": "https://dl.fbaipublicfiles.com/fairseq/wav2vec/wav2vec_vox_960h_pl.pt",
+ "dict_path": "https://dl.fbaipublicfiles.com/fairseq/wav2vec/dict.ltr.txt",
+ "model_type": "fairseq",
+ "lang": "en",
+ "post_process": "collapse"
+ }
+ },
+ "hok": {
+ "oct22": {
+ "desc": "Hokkien ASR model, for details check [TODO add paper link]",
+ "ckpt_path": "https://dl.fbaipublicfiles.com/ust_asr/hok/checkpoint_best.pt",
+ "dict_path": "https://dl.fbaipublicfiles.com/ust_asr/hok/dict.ltr.txt",
+ "model_type": "fairseq",
+ "lang": "hok",
+ "post_process": "none"
+ }
+ },
+ "es": {
+ "oct22": {
+ "model_path": "/data/zhangshaolei/pretrain_models/jonatasgrosman--wav2vec2-large-xlsr-53-spanish",
+ "model_type": "hf",
+ "lang": "es",
+ "post_process": "collapse"
+ }
+ },
+ "fr": {
+ "oct22": {
+ "model_path": "jonatasgrosman/wav2vec2-large-fr-voxpopuli-french",
+ "model_type": "hf",
+ "lang": "fr",
+ "post_process": "collapse"
+ }
+ },
+ "zh": {
+ "oct22": {
+ "model_path": "ydshieh/wav2vec2-large-xlsr-53-chinese-zh-cn-gpt",
+ "model_type": "hf",
+ "lang": "zh",
+ "post_process": "collapse"
+ }
+ },
+ "tr": {
+ "oct22": {
+ "model_path": "cahya/wav2vec2-large-xlsr-turkish-artificial-cv",
+ "model_type": "hf",
+ "lang": "tr",
+ "post_process": "collapse"
+ }
+ },
+ "ar": {
+ "oct22": {
+ "model_path": "jonatasgrosman/wav2vec2-large-xlsr-53-arabic",
+ "model_type": "hf",
+ "lang": "ar",
+ "post_process": "collapse"
+ }
+ },
+ "vi": {
+ "oct22": {
+ "model_path": "not-tanh/wav2vec2-large-xlsr-53-vietnamese",
+ "model_type": "hf",
+ "lang": "vi",
+ "post_process": "collapse"
+ }
+ },
+ "de": {
+ "oct22": {
+ "model_path": "jonatasgrosman/wav2vec2-xls-r-1b-german",
+ "model_type": "hf",
+ "lang": "de",
+ "post_process": "collapse"
+ }
+ },
+ "pl": {
+ "oct22": {
+ "model_path": "jonatasgrosman/wav2vec2-xls-r-1b-polish",
+ "model_type": "hf",
+ "lang": "pl",
+ "post_process": "collapse"
+ }
+ },
+ "it": {
+ "oct22": {
+ "model_path": "jonatasgrosman/wav2vec2-large-xlsr-53-italian",
+ "model_type": "hf",
+ "lang": "it",
+ "post_process": "collapse"
+ }
+ },
+ "pt": {
+ "oct22": {
+ "model_path": "jonatasgrosman/wav2vec2-xls-r-1b-portuguese",
+ "model_type": "hf",
+ "lang": "pt",
+ "post_process": "collapse"
+ }
+ },
+ "ro": {
+ "oct22": {
+ "model_path": "gigant/romanian-wav2vec2",
+ "model_type": "hf",
+ "lang": "ro",
+ "post_process": "collapse"
+ }
+ },
+ "cs": {
+ "oct22": {
+ "model_path": "comodoro/wav2vec2-xls-r-300m-cs-250",
+ "model_type": "hf",
+ "lang": "cs",
+ "post_process": "collapse"
+ }
+ },
+ "sk": {
+ "oct22": {
+ "model_path": "anuragshas/wav2vec2-xls-r-300m-sk-cv8-with-lm",
+ "model_type": "hf",
+ "lang": "sk",
+ "post_process": "collapse"
+ }
+ },
+ "sl": {
+ "oct22": {
+ "model_path": "anuragshas/wav2vec2-xls-r-300m-sl-cv8-with-lm",
+ "model_type": "hf",
+ "lang": "sl",
+ "post_process": "collapse"
+ }
+ },
+ "fi": {
+ "oct22": {
+ "model_path": "jonatasgrosman/wav2vec2-large-xlsr-53-finnish",
+ "model_type": "hf",
+ "lang": "fi",
+ "post_process": "collapse"
+ }
+ },
+ "hu": {
+ "oct22": {
+ "model_path": "jonatasgrosman/wav2vec2-large-xlsr-53-hungarian",
+ "model_type": "hf",
+ "lang": "hu",
+ "post_process": "collapse"
+ }
+ },
+ "et": {
+ "oct22": {
+ "model_path": "RASMUS/wav2vec2-xlsr-1b-et",
+ "model_type": "hf",
+ "lang": "et",
+ "post_process": "collapse"
+ }
+ },
+ "lt": {
+ "oct22": {
+ "model_path": "sammy786/wav2vec2-xlsr-lithuanian",
+ "model_type": "hf",
+ "lang": "lt",
+ "post_process": "collapse"
+ }
+ },
+ "nl": {
+ "oct22": {
+ "model_path": "jonatasgrosman/wav2vec2-xls-r-1b-dutch",
+ "model_type": "hf",
+ "lang": "nl",
+ "post_process": "collapse"
+ }
+ },
+ "lv": {
+ "oct22": {
+ "model_path": "reach-vb/wav2vec2-large-xls-r-1B-common_voice7-lv-ft",
+ "model_type": "hf",
+ "lang": "lv",
+ "post_process": "collapse"
+ }
+ },
+ "sv": {
+ "oct22": {
+ "model_path": "marinone94/xls-r-300m-sv-robust",
+ "model_type": "hf",
+ "lang": "sv",
+ "post_process": "collapse"
+ }
+ },
+ "hr": {
+ "oct22": {
+ "model_path": "classla/wav2vec2-xls-r-parlaspeech-hr",
+ "model_type": "hf",
+ "lang": "hr",
+ "post_process": "collapse"
+ }
+ }
+}
diff --git a/asr_bleu/compute_asr_bleu.py b/asr_bleu/compute_asr_bleu.py
new file mode 100644
index 0000000000000000000000000000000000000000..0f1dc2ff86158e82b82b85eaa6a7a27ad03e7469
--- /dev/null
+++ b/asr_bleu/compute_asr_bleu.py
@@ -0,0 +1,234 @@
+from typing import Dict, List
+import sacrebleu
+import pandas as pd
+from glob import glob
+from pathlib import Path
+from utils import retrieve_asr_config, ASRGenerator
+from tqdm import tqdm
+from argparse import ArgumentParser
+
+def merge_tailo_init_final(text):
+ """
+ Hokkien ASR hypothesis post-processing.
+ """
+ sps = text.strip().split()
+ results = []
+ last_syllable = ""
+ for sp in sps:
+ if sp == "NULLINIT" or sp == "nullinit":
+ continue
+ last_syllable += sp
+ if sp[-1].isnumeric():
+ results.append(last_syllable)
+ last_syllable = ""
+ if last_syllable != "":
+ results.append(last_syllable)
+ return " ".join(results)
+
+
+def remove_tone(text):
+ """
+ Used for tone-less evaluation of Hokkien
+ """
+ return " ".join([t[:-1] for t in text.split()])
+
+
+def extract_audio_for_eval(audio_dirpath: str, audio_format: str):
+ if audio_format == "n_pred.wav":
+ """
+ The assumption here is that pred_0.wav corresponds to the reference at line position 0 from the reference manifest
+ """
+ audio_list = []
+ audio_fp_list = glob((Path(audio_dirpath) / "*_pred.wav").as_posix())
+ for i in range(len(audio_fp_list)):
+ try:
+ audio_fp = (Path(audio_dirpath) / f"{i}_pred.wav").as_posix()
+ assert (
+ audio_fp in audio_fp_list
+ ), f"{Path(audio_fp).name} does not exist in {audio_dirpath}"
+ except AssertionError:
+ # check the audio with random speaker
+ audio_fp = Path(audio_dirpath) / f"{i}_spk*_pred.wav"
+ audio_fp = glob(
+ audio_fp.as_posix()
+ ) # resolve audio filepath with random speaker
+ assert len(audio_fp) == 1
+ audio_fp = audio_fp[0]
+
+ audio_list.append(audio_fp)
+ else:
+ raise NotImplementedError
+
+ return audio_list
+
+
+def extract_text_for_eval(
+ references_filepath: str, reference_format: str, reference_tsv_column: str = None
+):
+ if reference_format == "txt":
+ reference_sentences = open(references_filepath, "r").readlines()
+ reference_sentences = [l.strip() for l in reference_sentences]
+ elif reference_format == "tsv":
+ tsv_df = pd.read_csv(references_filepath, sep="\t", quoting=3)
+ reference_sentences = tsv_df[reference_tsv_column].to_list()
+ reference_sentences = [l.strip() for l in reference_sentences]
+ else:
+ raise NotImplementedError
+
+ return reference_sentences
+
+
+def compose_eval_data(
+ audio_dirpath: str,
+ audio_format: str,
+ references_filepath: str,
+ reference_format: str,
+ reference_tsv_column: str = None,
+ save_manifest_filepath=None,
+):
+ """
+ Speech matrix decoding pipeline produces audio with the following mask "N_pred.wav" where N is the order of the corresponding input sample
+ """
+
+ reference_sentences = extract_text_for_eval(
+ references_filepath, reference_format, reference_tsv_column
+ )
+ predicted_audio_fp_list = extract_audio_for_eval(audio_dirpath, audio_format)
+ assert len(predicted_audio_fp_list) == len(reference_sentences)
+
+ audio_text_pairs = [
+ (audio, reference)
+ for audio, reference in zip(predicted_audio_fp_list, reference_sentences)
+ ]
+
+ tsv_manifest = pd.DataFrame(audio_text_pairs, columns=["prediction", "reference"])
+
+ if save_manifest_filepath is not None:
+ tsv_manifest.to_csv(save_manifest_filepath, sep="\t", quoting=3)
+
+ return tsv_manifest
+
+
+def load_eval_data_from_tsv(eval_data_filepath: str):
+ """
+ We may load the result of `compose_eval_data` directly if needed
+ """
+ eval_df = pd.from_csv(eval_data_filepath, sep="\t")
+
+ return eval_df
+
+
+def run_asr_bleu(args):
+
+ asr_config = retrieve_asr_config(
+ args.lang, args.asr_version, json_path="./asr_model_cfgs.json"
+ )
+ asr_model = ASRGenerator(asr_config)
+
+ eval_manifest = compose_eval_data(
+ audio_dirpath=args.audio_dirpath,
+ audio_format=args.audio_format,
+ references_filepath=args.reference_path,
+ reference_format=args.reference_format,
+ reference_tsv_column=args.reference_tsv_column,
+ save_manifest_filepath=None,
+ )
+
+ prediction_transcripts = []
+ for _, eval_pair in tqdm(
+ eval_manifest.iterrows(),
+ desc="Transcribing predictions",
+ total=len(eval_manifest),
+ ):
+ try:
+ transcription = asr_model.transcribe_audiofile(eval_pair.prediction)
+ prediction_transcripts.append(transcription.lower())
+ except:
+ prediction_transcripts.append("")
+
+ if args.lang == "hok":
+ prediction_transcripts = [
+ merge_tailo_init_final(text) for text in prediction_transcripts
+ ]
+
+ with open(Path(args.audio_dirpath) / "pred.txt", "w") as f:
+ f.write("\n".join(prediction_transcripts))
+
+ references = eval_manifest["reference"].tolist()
+ bleu_score = sacrebleu.corpus_bleu(prediction_transcripts, [references])
+
+ print(bleu_score)
+
+ return bleu_score
+
+
+def main():
+ parser = ArgumentParser(
+ description="This script computes the ASR-BLEU metric between model's generated audio and the text reference sequences."
+ )
+
+ parser.add_argument(
+ "--lang",
+ help="The target language used to initialize ASR model, see asr_model_cfgs.json for available languages",
+ type=str,
+ )
+ parser.add_argument(
+ "--asr_version",
+ type=str,
+ default="oct22",
+ help="For future support we add and extra layer of asr versions. The current most recent version is oct22 meaning October 2022",
+ )
+ parser.add_argument(
+ "--audio_dirpath",
+ type=str,
+ help="Path to the directory containing the audio predictions from the translation model",
+ )
+ parser.add_argument(
+ "--reference_path",
+ type=str,
+ help="Path to the file containing reference translations in the form of normalized text (to be compared to ASR predictions",
+ )
+ parser.add_argument(
+ "--reference_format",
+ choices=["txt", "tsv"],
+ help="Format of reference file. Txt means plain text format where each line represents single reference sequence",
+ )
+ parser.add_argument(
+ "--reference_tsv_column",
+ default=None,
+ type=str,
+ help="If format is tsv, then specify the column name which contains reference sequence",
+ )
+ parser.add_argument(
+ "--audio_format",
+ default="n_pred.wav",
+ choices=["n_pred.wav"],
+ help="Audio format n_pred.wav corresponds to names like 94_pred.wav or 94_spk7_pred.wav where spk7 is the speaker id",
+ )
+ parser.add_argument(
+ "--results_dirpath",
+ default=None,
+ type=str,
+ help="If specified, the resulting BLEU score will be written to this file path as txt file",
+ )
+
+ args = parser.parse_args()
+
+ bleu_score = run_asr_bleu(args)
+ result_filename = f"{args.reference_format}_{args.lang}_bleu.txt"
+ if args.results_dirpath is not None:
+ if not Path(args.results_dirpath).exists():
+ Path(args.results_dirpath).mkdir(parents=True)
+ with open(Path(args.results_dirpath) / result_filename, "w") as f:
+ f.write(bleu_score.format(width=2))
+
+
+if __name__ == "__main__":
+ main()
+
+
+"""
+Example to load Sl audio and references, compute BLEU:
+
+export lang=fi; split=vp && python compute_asr_bleu.py --lang $lang --audio_dirpath /checkpoint/hygong/S2S/speech_matrix_release_ckpts/generated_waveform_release/en-$lang/test_$split/checkpoint.pt --audio_format n_pred.wav --reference_path /large_experiments/ust/hygong/S2S/SpeechEncoder/manifests/vp-vp/en-$lang/test_$split.$lang --reference_format txt --results_dirpath ./
+"""
diff --git a/asr_bleu/requirements.txt b/asr_bleu/requirements.txt
new file mode 100644
index 0000000000000000000000000000000000000000..cfa90f6aeffd879c1e309ae58a2f050c943e87b6
--- /dev/null
+++ b/asr_bleu/requirements.txt
@@ -0,0 +1,7 @@
+fairseq==0.12.2
+pandas==1.4.3
+sacrebleu==2.2.0
+torch==1.12.1
+torchaudio==0.12.1
+tqdm==4.64.0
+transformers==4.21.1
diff --git a/asr_bleu/utils.py b/asr_bleu/utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..0fed55a9b9032952e8cc20db36c127b8354f0478
--- /dev/null
+++ b/asr_bleu/utils.py
@@ -0,0 +1,306 @@
+import json
+import re
+import urllib.request
+from pathlib import Path
+
+import fairseq
+import torch
+from fairseq.data.data_utils import lengths_to_padding_mask
+from tqdm import tqdm
+
+try:
+ import torchaudio
+ from torchaudio.models.decoder import ctc_decoder
+except ImportError:
+ raise ImportError("Upgrade torchaudio to 0.12 to enable CTC decoding")
+
+
+class DownloadProgressBar(tqdm):
+ """A class to represent a download progress bar"""
+
+ def update_to(self, b=1, bsize=1, tsize=None) -> None:
+ """
+ Update the download progress
+ """
+ if tsize is not None:
+ self.total = tsize
+ self.update(b * bsize - self.n)
+
+
+def retrieve_asr_config(lang_key: str, asr_version: str, json_path: str) -> dict:
+ """
+ Retrieve the asr model configs
+
+ Args:
+ lang_key: the lanuage type as the key name
+ json_path: the path of the config json file
+
+ Returns:
+ Dict of all the configs in the json file
+ """
+
+ with open(json_path, "r") as f:
+ asr_model_cfgs = json.load(f)
+ return asr_model_cfgs[lang_key][asr_version]
+
+
+class ASRGenerator(object):
+ """A class to represent a ASR generator"""
+
+ def __init__(
+ self,
+ model_cfg: dict,
+ cache_dirpath: str = (Path.home() / ".cache" / "ust_asr").as_posix(),
+ ) -> None:
+ """
+ Construct all the necessary attributes of the ASRGenerator class
+
+ Args:
+ model_cfg: the dict of the asr model config
+ cache_dirpath: the default cache path is "Path.home()/.cache/ust_asr"
+ """
+
+ self.cache_dirpath = Path(cache_dirpath) / model_cfg["lang"]
+ self.model_cfg = model_cfg
+
+ self.use_cuda = torch.cuda.is_available()
+
+ torchaudio.set_audio_backend("sox_io")
+
+ if self.model_cfg["model_type"] == "hf":
+ self.prepare_hf_model(self.model_cfg)
+ elif self.model_cfg["model_type"] == "fairseq":
+ self.prepare_fairseq_model(self.model_cfg)
+ else:
+ raise NotImplementedError(
+ f"Model type {self.model_cfg['model_type']} is not supported"
+ )
+
+ if self.model_cfg["post_process"] == "collapse":
+ self.post_process_fn = lambda hypo: "".join(hypo).replace(
+ self.sil_token, " "
+ )
+ elif self.model_cfg["post_process"] == "none":
+ self.post_process_fn = lambda hypo: " ".join(hypo).replace(
+ self.sil_token, " "
+ )
+ else:
+ raise NotImplementedError
+
+ if self.use_cuda:
+ self.model.cuda()
+ self.model.eval()
+
+ self.decoder = ctc_decoder(
+ lexicon=None,
+ tokens=self.tokens,
+ lm=None,
+ nbest=1,
+ beam_size=1,
+ beam_size_token=None,
+ lm_weight=0.0,
+ word_score=0.0,
+ unk_score=float("-inf"),
+ sil_token=self.sil_token,
+ sil_score=0.0,
+ log_add=False,
+ blank_token=self.blank_token,
+ )
+
+ def prepare_hf_model(self, model_cfg: dict) -> None:
+ """
+ Prepare the huggingface asr model
+
+ Args:
+ model_cfg: dict with the relevant ASR config
+ """
+
+ def infer_silence_token(vocab: list):
+ """
+ Different HF checkpoints have different notion of silence token
+ such as | or " " (space)
+ Important: when adding new HF asr model in, check what silence token it uses
+ """
+ if "|" in vocab:
+ return "|"
+ elif " " in vocab:
+ return " "
+ else:
+ raise RuntimeError("Silence token is not found in the vocabulary")
+
+ try:
+ from transformers import (AutoFeatureExtractor, AutoTokenizer,
+ Wav2Vec2ForCTC, Wav2Vec2Processor)
+ except ImportError:
+ raise ImportError("Install transformers to load HF wav2vec model")
+
+ model_path = model_cfg["model_path"]
+ self.model = Wav2Vec2ForCTC.from_pretrained(model_path)
+ self.tokenizer = AutoTokenizer.from_pretrained(model_path)
+ self.preprocessor = AutoFeatureExtractor.from_pretrained(model_path)
+ self.processor = Wav2Vec2Processor.from_pretrained(model_path)
+
+ # extra unk tokens are there to make some models work e.g. Finnish ASR has some vocab issue
+ vocab_list = [
+ self.tokenizer.decoder.get(i, f"{self.tokenizer.unk_token}1")
+ for i in range(self.tokenizer.vocab_size)
+ ]
+
+ self.sampling_rate = self.preprocessor.sampling_rate
+ self.normalize_input = self.preprocessor.do_normalize
+ self.tokens = vocab_list
+ self.sil_token = infer_silence_token(vocab_list)
+ self.blank_token = self.tokenizer.pad_token
+
+ def prepare_fairseq_model(self, model_cfg: dict) -> None:
+ """
+ Prepare the fairseq asr model
+
+ Args:
+ model_cfg: the specific model config dict must have: (1) ckpt_path, (2) dict_path
+ """
+
+ def download_file(url: str, cache_dir: Path):
+ download_path = cache_dir / url.split("/")[-1]
+ if not (cache_dir / url.split("/")[-1]).exists():
+ with DownloadProgressBar(
+ unit="B", unit_scale=True, miniters=1, desc=url.split("/")[-1]
+ ) as t:
+ cache_dir.mkdir(parents=True, exist_ok=True)
+ urllib.request.urlretrieve(
+ url, filename=download_path.as_posix(), reporthook=t.update_to
+ )
+ else:
+ print(f"'{url}' exists in {cache_dir}")
+
+ return download_path.as_posix()
+
+ try:
+ ckpt_path = model_cfg["ckpt_path"]
+ dict_path = model_cfg["dict_path"]
+ except KeyError:
+ raise KeyError(
+ "Fairseq model cfg must provide (1) ckpt_path, (2) dict_path"
+ )
+
+ if re.search("^https", ckpt_path):
+ ckpt_path = download_file(ckpt_path, self.cache_dirpath)
+ if re.search("^https", dict_path):
+ dict_path = download_file(dict_path, self.cache_dirpath)
+
+ model, saved_cfg, _ = fairseq.checkpoint_utils.load_model_ensemble_and_task(
+ [ckpt_path],
+ arg_overrides={
+ "task": "audio_finetuning",
+ "data": self.cache_dirpath.as_posix(),
+ }, # data must have dict in it
+ )
+
+ dict_lines = open(dict_path, "r").readlines()
+ tokens = [l.split()[0] for l in dict_lines]
+ # adding default fairseq special tokens
+ tokens = ["", "", "", ""] + tokens
+
+ self.model = model[0]
+ self.tokens = tokens
+
+ if "|" in tokens:
+ self.sil_token = "|"
+ else:
+ self.sil_token = tokens[
+ 2
+ ] # use eos as silence token if | not presented e.g., Hok ASR model
+ print(f"Inferring silence token from the dict: {self.sil_token}")
+ self.blank_token = self.tokens[0]
+
+ self.sampling_rate = saved_cfg.task.sample_rate
+ self.normalize_input = saved_cfg.task.normalize
+
+ @torch.inference_mode()
+ def load_audiofile(self, audio_path: str) -> torch.Tensor:
+ """
+ Load the audio files and apply resampling and normalizaion
+
+ Args:
+ audio_path: the audio file path
+
+ Returns:
+ audio_waveform: the audio waveform as a torch.Tensor object
+ """
+
+ audio_waveform, sampling_rate = torchaudio.load(audio_path)
+ if audio_waveform.dim == 2:
+ audio_waveform = audio_waveform.mean(-1)
+ if self.sampling_rate != sampling_rate:
+ audio_waveform = torchaudio.functional.resample(
+ audio_waveform, sampling_rate, self.sampling_rate
+ )
+ if self.normalize_input:
+ # following fairseq raw audio dataset
+ audio_waveform = torch.nn.functional.layer_norm(
+ audio_waveform, audio_waveform.shape
+ )
+
+ return audio_waveform
+
+ @torch.inference_mode()
+ def compute_emissions(self, audio_input: torch.Tensor) -> torch.Tensor:
+ """
+ Compute the emissions for either fairseq or huggingface asr model
+
+ Args:
+ audio_path: the input audio waveform
+
+ Returns:
+ emissions: the logits of the encoded prediction.
+ """
+
+ if self.use_cuda:
+ audio_input = audio_input.to("cuda")
+ if isinstance(self.model, fairseq.models.wav2vec.wav2vec2_asr.Wav2VecCtc):
+ padding_mask = lengths_to_padding_mask(torch.tensor([audio_input.numel()]))
+ emissions = self.model.w2v_encoder(audio_input, padding_mask)[
+ "encoder_out"
+ ].transpose(0, 1)
+ else:
+ emissions = self.model(audio_input).logits
+
+ return emissions
+
+ def decode_emissions(self, emissions: torch.Tensor) -> str:
+ """
+ Decode the emissions and apply post process functions
+
+ Args:
+ emissions: the input Tensor object
+
+ Returns:
+ hypo: the str as the decoded transcriptions
+ """
+
+ emissions = emissions.cpu()
+ results = self.decoder(emissions)
+
+ # assuming the lexicon-free decoder and working with tokens
+ hypo = self.decoder.idxs_to_tokens(results[0][0].tokens)
+ hypo = self.post_process_fn(hypo)
+
+ return hypo
+
+ def transcribe_audiofile(self, audio_path: str, lower=True) -> str:
+ """
+ Transcribe the audio into string
+
+ Args:
+ audio_path: the input audio waveform
+ lower: the case of the transcriptions with lowercase as the default
+
+ Returns:
+ hypo: the transcription result
+ """
+
+ asr_input = self.load_audiofile(audio_path)
+ emissions = self.compute_emissions(asr_input)
+ hypo = self.decode_emissions(emissions)
+
+ return hypo.strip().lower() if lower else hypo.strip()
diff --git a/asr_bleu_rm_silence/README.md b/asr_bleu_rm_silence/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..b808b136dc25018c9e43872e1377f12699758199
--- /dev/null
+++ b/asr_bleu_rm_silence/README.md
@@ -0,0 +1,34 @@
+# ASR-BLEU evaluation toolkit
+
+This toolkit provides a set of public ASR models used for evaluation of different speech-to-speech translation systems at Meta AI. It enables easier score comparisons between different system's outputs.
+
+The ASRGenerator wraps different CTC-based ASR models from HuggingFace and fairseq code bases. Torchaudio CTC decoder is built on top of it to decode given audio files.
+
+Please see `asr_model_cfgs.json` for a list of languages covered currently.
+
+The high-level pipeline is simple by design: given a lang tag, script loads the ASR model, transcribes model's predicted audio, and computes the BLEU score against provided reference translations using sacrebleu.
+
+# Dependencies
+
+Please see `requirements.txt`.
+
+# Usage examples
+
+This toolkit have been used with:
+
+* Speechmatrix project: https://github.com/facebookresearch/fairseq/tree/ust/examples/speech_matrix.
+
+* Hokkien speech-to-speech translation project: https://github.com/facebookresearch/fairseq/tree/ust/examples/hokkien.
+
+# Standalone run example
+
+High-level example, please substitute arguments per your case:
+
+```bash
+python compute_asr_bleu.py --lang \
+--audio_dirpath \
+--reference_path \
+--reference_format txt
+```
+
+For more details about arguments please see the script argparser help.
\ No newline at end of file
diff --git a/asr_bleu_rm_silence/__init__.py b/asr_bleu_rm_silence/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/asr_bleu_rm_silence/asr_model_cfgs.json b/asr_bleu_rm_silence/asr_model_cfgs.json
new file mode 100644
index 0000000000000000000000000000000000000000..e4613a7ad2b64f8cb6a71824494a74e006b2e775
--- /dev/null
+++ b/asr_bleu_rm_silence/asr_model_cfgs.json
@@ -0,0 +1,198 @@
+{
+ "en": {
+ "oct22": {
+ "desc": "Wav2Vec 2.0 Large (LV-60) + Self Training from https://github.com/facebookresearch/fairseq/tree/main/examples/wav2vec#pre-trained-models",
+ "ckpt_path": "https://dl.fbaipublicfiles.com/fairseq/wav2vec/wav2vec_vox_960h_pl.pt",
+ "dict_path": "https://dl.fbaipublicfiles.com/fairseq/wav2vec/dict.ltr.txt",
+ "model_type": "fairseq",
+ "lang": "en",
+ "post_process": "collapse"
+ }
+ },
+ "hok": {
+ "oct22": {
+ "desc": "Hokkien ASR model, for details check [TODO add paper link]",
+ "ckpt_path": "https://dl.fbaipublicfiles.com/ust_asr/hok/checkpoint_best.pt",
+ "dict_path": "https://dl.fbaipublicfiles.com/ust_asr/hok/dict.ltr.txt",
+ "model_type": "fairseq",
+ "lang": "hok",
+ "post_process": "none"
+ }
+ },
+ "es": {
+ "oct22": {
+ "model_path": "/data/zhangshaolei/pretrain_models/jonatasgrosman--wav2vec2-large-xlsr-53-spanish",
+ "model_type": "hf",
+ "lang": "es",
+ "post_process": "collapse"
+ }
+ },
+ "fr": {
+ "oct22": {
+ "model_path": "jonatasgrosman/wav2vec2-large-fr-voxpopuli-french",
+ "model_type": "hf",
+ "lang": "fr",
+ "post_process": "collapse"
+ }
+ },
+ "zh": {
+ "oct22": {
+ "model_path": "ydshieh/wav2vec2-large-xlsr-53-chinese-zh-cn-gpt",
+ "model_type": "hf",
+ "lang": "zh",
+ "post_process": "collapse"
+ }
+ },
+ "tr": {
+ "oct22": {
+ "model_path": "cahya/wav2vec2-large-xlsr-turkish-artificial-cv",
+ "model_type": "hf",
+ "lang": "tr",
+ "post_process": "collapse"
+ }
+ },
+ "ar": {
+ "oct22": {
+ "model_path": "jonatasgrosman/wav2vec2-large-xlsr-53-arabic",
+ "model_type": "hf",
+ "lang": "ar",
+ "post_process": "collapse"
+ }
+ },
+ "vi": {
+ "oct22": {
+ "model_path": "not-tanh/wav2vec2-large-xlsr-53-vietnamese",
+ "model_type": "hf",
+ "lang": "vi",
+ "post_process": "collapse"
+ }
+ },
+ "de": {
+ "oct22": {
+ "model_path": "jonatasgrosman/wav2vec2-xls-r-1b-german",
+ "model_type": "hf",
+ "lang": "de",
+ "post_process": "collapse"
+ }
+ },
+ "pl": {
+ "oct22": {
+ "model_path": "jonatasgrosman/wav2vec2-xls-r-1b-polish",
+ "model_type": "hf",
+ "lang": "pl",
+ "post_process": "collapse"
+ }
+ },
+ "it": {
+ "oct22": {
+ "model_path": "jonatasgrosman/wav2vec2-large-xlsr-53-italian",
+ "model_type": "hf",
+ "lang": "it",
+ "post_process": "collapse"
+ }
+ },
+ "pt": {
+ "oct22": {
+ "model_path": "jonatasgrosman/wav2vec2-xls-r-1b-portuguese",
+ "model_type": "hf",
+ "lang": "pt",
+ "post_process": "collapse"
+ }
+ },
+ "ro": {
+ "oct22": {
+ "model_path": "gigant/romanian-wav2vec2",
+ "model_type": "hf",
+ "lang": "ro",
+ "post_process": "collapse"
+ }
+ },
+ "cs": {
+ "oct22": {
+ "model_path": "comodoro/wav2vec2-xls-r-300m-cs-250",
+ "model_type": "hf",
+ "lang": "cs",
+ "post_process": "collapse"
+ }
+ },
+ "sk": {
+ "oct22": {
+ "model_path": "anuragshas/wav2vec2-xls-r-300m-sk-cv8-with-lm",
+ "model_type": "hf",
+ "lang": "sk",
+ "post_process": "collapse"
+ }
+ },
+ "sl": {
+ "oct22": {
+ "model_path": "anuragshas/wav2vec2-xls-r-300m-sl-cv8-with-lm",
+ "model_type": "hf",
+ "lang": "sl",
+ "post_process": "collapse"
+ }
+ },
+ "fi": {
+ "oct22": {
+ "model_path": "jonatasgrosman/wav2vec2-large-xlsr-53-finnish",
+ "model_type": "hf",
+ "lang": "fi",
+ "post_process": "collapse"
+ }
+ },
+ "hu": {
+ "oct22": {
+ "model_path": "jonatasgrosman/wav2vec2-large-xlsr-53-hungarian",
+ "model_type": "hf",
+ "lang": "hu",
+ "post_process": "collapse"
+ }
+ },
+ "et": {
+ "oct22": {
+ "model_path": "RASMUS/wav2vec2-xlsr-1b-et",
+ "model_type": "hf",
+ "lang": "et",
+ "post_process": "collapse"
+ }
+ },
+ "lt": {
+ "oct22": {
+ "model_path": "sammy786/wav2vec2-xlsr-lithuanian",
+ "model_type": "hf",
+ "lang": "lt",
+ "post_process": "collapse"
+ }
+ },
+ "nl": {
+ "oct22": {
+ "model_path": "jonatasgrosman/wav2vec2-xls-r-1b-dutch",
+ "model_type": "hf",
+ "lang": "nl",
+ "post_process": "collapse"
+ }
+ },
+ "lv": {
+ "oct22": {
+ "model_path": "reach-vb/wav2vec2-large-xls-r-1B-common_voice7-lv-ft",
+ "model_type": "hf",
+ "lang": "lv",
+ "post_process": "collapse"
+ }
+ },
+ "sv": {
+ "oct22": {
+ "model_path": "marinone94/xls-r-300m-sv-robust",
+ "model_type": "hf",
+ "lang": "sv",
+ "post_process": "collapse"
+ }
+ },
+ "hr": {
+ "oct22": {
+ "model_path": "classla/wav2vec2-xls-r-parlaspeech-hr",
+ "model_type": "hf",
+ "lang": "hr",
+ "post_process": "collapse"
+ }
+ }
+}
diff --git a/asr_bleu_rm_silence/compute_asr_bleu.py b/asr_bleu_rm_silence/compute_asr_bleu.py
new file mode 100644
index 0000000000000000000000000000000000000000..9ed3680d10dea9efe1dda39d12d2a7b5b0e1fe58
--- /dev/null
+++ b/asr_bleu_rm_silence/compute_asr_bleu.py
@@ -0,0 +1,245 @@
+import os
+from typing import Dict, List
+import sacrebleu
+import pandas as pd
+from glob import glob
+from pathlib import Path
+from utils import retrieve_asr_config, ASRGenerator
+from tqdm import tqdm
+from argparse import ArgumentParser
+
+
+def merge_tailo_init_final(text):
+ """
+ Hokkien ASR hypothesis post-processing.
+ """
+ sps = text.strip().split()
+ results = []
+ last_syllable = ""
+ for sp in sps:
+ if sp == "NULLINIT" or sp == "nullinit":
+ continue
+ last_syllable += sp
+ if sp[-1].isnumeric():
+ results.append(last_syllable)
+ last_syllable = ""
+ if last_syllable != "":
+ results.append(last_syllable)
+ return " ".join(results)
+
+
+def remove_tone(text):
+ """
+ Used for tone-less evaluation of Hokkien
+ """
+ return " ".join([t[:-1] for t in text.split()])
+
+
+def extract_audio_for_eval(audio_dirpath: str, audio_format: str):
+ if audio_format == "n_pred.wav":
+ """
+ The assumption here is that 0_pred.wav corresponds to the reference at line position 0 from the reference manifest
+ """
+ audio_list = []
+ audio_fp_list = glob((Path(audio_dirpath) / "*_pred.wav").as_posix())
+ audio_fp_list = sorted(
+ audio_fp_list, key=lambda x: int(os.path.basename(x).split("_")[0])
+ )
+ for i in range(len(audio_fp_list)):
+ try:
+ audio_fp = (Path(audio_dirpath) / f"{i}_pred.wav").as_posix()
+ assert (
+ audio_fp in audio_fp_list
+ ), f"{Path(audio_fp).name} does not exist in {audio_dirpath}"
+ except AssertionError:
+ # check the audio with random speaker
+ audio_fp = Path(audio_dirpath) / f"{i}_spk*_pred.wav"
+ audio_fp = glob(
+ audio_fp.as_posix()
+ ) # resolve audio filepath with random speaker
+ assert len(audio_fp) == 1
+ audio_fp = audio_fp[0]
+
+ audio_list.append(audio_fp)
+ else:
+ raise NotImplementedError
+
+ return audio_list
+
+
+def extract_text_for_eval(
+ references_filepath: str, reference_format: str, reference_tsv_column: str = None
+):
+ if reference_format == "txt":
+ reference_sentences = open(references_filepath, "r").readlines()
+ reference_sentences = [l.strip() for l in reference_sentences]
+ elif reference_format == "tsv":
+ tsv_df = pd.read_csv(references_filepath, sep="\t", quoting=3)
+ reference_sentences = tsv_df[reference_tsv_column].to_list()
+ reference_sentences = [l.strip() for l in reference_sentences]
+ else:
+ raise NotImplementedError
+
+ return reference_sentences
+
+
+def compose_eval_data(
+ audio_dirpath: str,
+ audio_format: str,
+ references_filepath: str,
+ reference_format: str,
+ reference_tsv_column: str = None,
+ save_manifest_filepath=None,
+):
+ """
+ Speech matrix decoding pipeline produces audio with the following mask "N_pred.wav" where N is the order of the corresponding input sample
+ """
+
+ reference_sentences = extract_text_for_eval(
+ references_filepath, reference_format, reference_tsv_column
+ )
+ predicted_audio_fp_list = extract_audio_for_eval(audio_dirpath, audio_format)
+
+ assert len(predicted_audio_fp_list) == len(reference_sentences)
+
+ audio_text_pairs = [
+ (audio, reference)
+ for audio, reference in zip(predicted_audio_fp_list, reference_sentences)
+ ]
+
+ tsv_manifest = pd.DataFrame(audio_text_pairs, columns=["prediction", "reference"])
+
+ if save_manifest_filepath is not None:
+ tsv_manifest.to_csv(save_manifest_filepath, sep="\t", quoting=3)
+
+ return tsv_manifest
+
+
+def load_eval_data_from_tsv(eval_data_filepath: str):
+ """
+ We may load the result of `compose_eval_data` directly if needed
+ """
+ eval_df = pd.from_csv(eval_data_filepath, sep="\t")
+
+ return eval_df
+
+
+def run_asr_bleu(args):
+
+ asr_config = retrieve_asr_config(
+ args.lang, args.asr_version, json_path="./asr_model_cfgs.json"
+ )
+ asr_model = ASRGenerator(asr_config)
+
+ eval_manifest = compose_eval_data(
+ audio_dirpath=args.audio_dirpath,
+ audio_format=args.audio_format,
+ references_filepath=args.reference_path,
+ reference_format=args.reference_format,
+ reference_tsv_column=args.reference_tsv_column,
+ save_manifest_filepath=None,
+ )
+
+ prediction_transcripts = []
+ for _, eval_pair in tqdm(
+ eval_manifest.iterrows(),
+ desc="Transcribing predictions",
+ total=len(eval_manifest),
+ ):
+ transcription = asr_model.transcribe_audiofile(eval_pair.prediction)
+ prediction_transcripts.append(transcription.lower())
+
+ if args.lang == "hok":
+ prediction_transcripts = [
+ merge_tailo_init_final(text) for text in prediction_transcripts
+ ]
+
+ references = eval_manifest["reference"].tolist()
+ bleu_score = sacrebleu.corpus_bleu(prediction_transcripts, [references])
+
+ print(bleu_score)
+
+ return prediction_transcripts, bleu_score
+
+
+def main():
+ parser = ArgumentParser(
+ description="This script computes the ASR-BLEU metric between model's generated audio and the text reference sequences."
+ )
+
+ parser.add_argument(
+ "--lang",
+ help="The target language used to initialize ASR model, see asr_model_cfgs.json for available languages",
+ type=str,
+ )
+ parser.add_argument(
+ "--asr_version",
+ type=str,
+ default="oct22",
+ help="For future support we add and extra layer of asr versions. The current most recent version is oct22 meaning October 2022",
+ )
+ parser.add_argument(
+ "--audio_dirpath",
+ type=str,
+ help="Path to the directory containing the audio predictions from the translation model",
+ )
+ parser.add_argument(
+ "--reference_path",
+ type=str,
+ help="Path to the file containing reference translations in the form of normalized text (to be compared to ASR predictions",
+ )
+ parser.add_argument(
+ "--reference_format",
+ choices=["txt", "tsv"],
+ help="Format of reference file. Txt means plain text format where each line represents single reference sequence",
+ )
+ parser.add_argument(
+ "--reference_tsv_column",
+ default=None,
+ type=str,
+ help="If format is tsv, then specify the column name which contains reference sequence",
+ )
+ parser.add_argument(
+ "--audio_format",
+ default="n_pred.wav",
+ choices=["n_pred.wav"],
+ help="Audio format n_pred.wav corresponds to names like 94_pred.wav or 94_spk7_pred.wav where spk7 is the speaker id",
+ )
+ parser.add_argument(
+ "--results_dirpath",
+ default=None,
+ type=str,
+ help="If specified, the resulting BLEU score will be written to this file path as txt file",
+ )
+ parser.add_argument(
+ "--transcripts_path",
+ default=None,
+ type=str,
+ help="If specified, the predicted transcripts will be written to this path as a txt file.",
+ )
+
+ args = parser.parse_args()
+
+ prediction_transcripts, bleu_score = run_asr_bleu(args)
+ result_filename = f"{args.reference_format}_{args.lang}_bleu.txt"
+ if args.results_dirpath is not None:
+ if not Path(args.results_dirpath).exists():
+ Path(args.results_dirpath).mkdir(parents=True)
+ with open(Path(args.results_dirpath) / result_filename, "w") as f:
+ f.write(bleu_score.format(width=2))
+
+ if args.transcripts_path is not None:
+ with open(args.transcripts_path, "w") as f:
+ for transcript in prediction_transcripts:
+ f.write(transcript + "\n")
+
+
+if __name__ == "__main__":
+ main()
+
+
+"""
+Example to load Sl audio and references, compute BLEU:
+
+export lang=fi; split=vp && python compute_asr_bleu.py --lang $lang --audio_dirpath /checkpoint/hygong/S2S/speech_matrix_release_ckpts/generated_waveform_release/en-$lang/test_$split/checkpoint.pt --audio_format n_pred.wav --reference_path /large_experiments/ust/hygong/S2S/SpeechEncoder/manifests/vp-vp/en-$lang/test_$split.$lang --reference_format txt --results_dirpath ./
+"""
diff --git a/asr_bleu_rm_silence/requirements.txt b/asr_bleu_rm_silence/requirements.txt
new file mode 100644
index 0000000000000000000000000000000000000000..cfa90f6aeffd879c1e309ae58a2f050c943e87b6
--- /dev/null
+++ b/asr_bleu_rm_silence/requirements.txt
@@ -0,0 +1,7 @@
+fairseq==0.12.2
+pandas==1.4.3
+sacrebleu==2.2.0
+torch==1.12.1
+torchaudio==0.12.1
+tqdm==4.64.0
+transformers==4.21.1
diff --git a/asr_bleu_rm_silence/test.txt b/asr_bleu_rm_silence/test.txt
new file mode 100644
index 0000000000000000000000000000000000000000..1688efa3500f183c3de624b2ef49257ed3f14030
--- /dev/null
+++ b/asr_bleu_rm_silence/test.txt
@@ -0,0 +1,14759 @@
+a real interesting work will finally be brought to the subject
+a reform for the order is necessary to be taken to land
+not many of this
+an indomitial intermediary of the incapt was a few weeks
+the floor is now open to mister alan le mariere to support the twenty eighth amendment
+it is in all cases jupiclaria
+auvoirs are also a poet
+i have experienced the past years i tell myself later
+twelve minutes thirty
+it is an evidence of the french people's life
+you have excellent reading the newspapers and the other ones
+the first of them is that of the birth maturity
+the question of structural deficit is regularly revising
+would you have their budget from one point five bitin hugos
+this doesn't this doesn't this doesn't seem fair
+the artist the artist is in mister debun
+i hope you didn't leave me mister dimula
+but i will only give one example of this amendment
+it's right madam minister of state that differentiates us from
+nobody wants it to be less numbered in the classes
+the future wanted to ruin it
+finally we don't always know if the leader pays or not
+smalt lexham ennins is a member of the national team
+thank you for completing my dear colleague
+it's the crapanism of the carapalis
+nobody forgot the delusion to come in and get a little bit more
+the second remark is that all this comes back to the same
+this budget is virtuous
+applying it to the first uras can be a good solution in some cases
+we must evacuate this type of fear
+this article is central matam minister
+on the name of the government i am therefore not asking for the wisdom of the assembly
+but the minister installed a working group on the subject
+or else you simplify the socialist
+ph pco' niffer a different menter can be used in depensing any tool
+patiently your eyes were soon finished
+they are also very satisfied with measures that are taken into actions for housing
+unfavor
+wer were assure you on the contrary of the policy in ours
+through the companies the other companies their taking place in the same way
+it was my friend that was a great number of people
+what is the reason why being the s i e s p i l e
+the mayan civilization is very ancient and deserves more attention
+the work between all the pumentary groups was extremely instructive
+the floor is now open to mister denis chacket registered on article thirteen
+this is why
+the last date the wrapper current confirmed his incomprehensions
+mister minister listen to the company mister minister listen to the company's chief listen to the craftsmen listen to the employees
+it will be a progress that means for its workers
+let's get back to the kids it is to say mister laurent's amendment that it is necessary to vote for
+the floor is now open to mister mohammed lakila
+the setons want to be better known
+i said that part of the device passibility of two thousand and ten was repealed
+the tabelet aspect must be brilliant indicating that he has enough cold
+should i apply to the collectivities that take responsibility for the election
+the mid wods are competent to follow normal and patic pregnancy
+adieu mister de courson
+from then on it poses a real problem
+here 's a lesson of favings
+barrwing discourages and courages to the right of the country
+a word on the reform of retirements
+i remember the creation of the c e s here in this hemicycle
+i think that this is a major advice
+arte twelve is a good article
+the question of the leader is in a freely agreed control
+i invite you mister minister dear collits to support the whole of these amendments
+my glass is quickly ans i'm going to fill it before going on
+this will happen in a decrease in purchasing power of house holds
+i think she actually works the complexity of the texts that are offered to us
+at his age i was already brought to the wine in the lovers
+we found a difficult time to five point one
+i will grant me mister chirkey that of the germany is best known as ours
+this exact thing is that the most important
+politics has disastrous effects for modest households
+the prehistoric era is the beginning of computing and is considered uplicated
+no measure in the eyes is planed on these matters in this budget in two thousand and fourteen
+furthermore i do not remember the attempt to tax the b e
+it is the meaning of the proposition sevengny
+the generations of venicions did not undergo cathy the hostages of the echoes
+before contacting us we are looking at our frequently asked questions
+the floor is now open to mister patrick de vedian
+no sir doles
+your second jocky was seating in the brugel commission
+in a beautiful unanimous because this topicis as you say is sensitive
+he is taken off and the group will vote against the other identical amendments
+the change is now
+the order of the day calls for the government's questions
+it is this perfect road that we have to replace a certain virtue
+ten admeses you ten admeses you refusing to hear at all the constitutional council
+i'm running after
+i i i don't want any fundamental differences of approaching between these two requests
+these teams will arrive in the next few days
+at the level of configuration this needs the following lines to be added
+the specialty of rural areas
+the floor is now open to mister john jacques candelier
+your amendment is missus asino
+i would like to conclude with two or three comments
+i think that this text will cause a very strong break up of equality
+but sometimes on comparion
+in four letters five sup the morning
+independent of the measures taken to the article eleven and twelve of mister
+the floor is now open to missus barenger poletti
+so ho will find in this box the following tools
+ninety grams of soft butter
+unfavorable opinion same as for the funds i share your point
+come on
+the floor is now open to mister ilan fort to support the thirty eighth amendment
+i am very sincere
+he is forbidden
+i am before you were disappointed
+the floor is open to missus cristel dubos to support amendment number three hundred ninety nine
+what is you weady to keep mares
+give us chespins
+yesterday and to day a particular emotion was expressed
+is a contract that we are moving with our fellow citizens
+we are shoring
+we must watch out what they are generalized each time it's possible
+once the borders are no longer limited
+to day the tax revenues do not show any more
+the regulations call two minutes for the interventions on the articles
+the deficit of the state itself is from sixty three to seventy two billion uros
+mister president i will be the only one in tanet on this article
+w write to my dear collik to come back to their budgetary discussions
+if some insist on the lovely in advance others apply it in a pleasant unfair to a grateful nunity
+first counter truth no engagement would be resentment of public finances as elected
+we're not toting about the same thing
+it's not pretty game and to tell this one
+for example they work together on the subject
+le let's get down thirty two hours
+it is therefore the chronology of the line
+you got on it for ten years
+this article too is a regression that is a real challenge
+we have to move the balance of the introductory of the two parts of the team
+but to evidence you do but to evidence you do but to evidence you do simplify mister terris
+let's talk about the ensins you talk about it very little dear call its from the opposition
+idon't want to add but if
+bro
+in other words your plans we cannot be confused with difficulty
+its run its run in te tax increase frveles
+mister hemosh was at the same time in california the tispust of gold
+regarding the quotalizations we could call them on citizens occasions
+this provision is absolutely necessary in the case where the complications would occur
+the court of auditors wanted to say it
+it's wrong
+he had said it
+i suggest that you will reassure you and listen to mister hana's questions
+exactly at our colic demion bads
+we need to make the courage to
+the floor is now open to mister christian paul
+but on the lest when we are bringing new taxes to protect the most modest ones
+the device incriminated plans that the state of the legs one of its competencies
+his posts in fact the problem is a problem
+the government understood it
+these are two different conceptions of the same conception
+the floor is now open to mister shock lamblin to support the thirty eighth amendment
+on the contrary it is necessary to go to the country
+this will be done in the next week
+he has lost the boy
+it was a very negligible science at that time
+missus bato had said the time i was wrong
+square tucked at the plate him
+you know very briefly decided to go on it
+michael became an expert in pete he almost always hits the jack
+she couldn't see sir he doesn't deserve the converts
+i would like to say a few words from tudoratories
+it is a coherent amendment of the global device
+we have therefore not received anybody mister minister
+take notes
+after this advertising i've heard properly i can only take back my amendment
+this being said this academic date is of colossal importance
+recognize him at least
+at the back we all agree to adove such action
+it seemed to us that the text requires it
+here's the result of the work a we led to those associations
+in any case what we propose to not question the national solidarity
+come on let mister merton finish shaw
+i would like us to express our rejection of this inefficient measure
+this amendment in appeal is therefore quite simple
+these two actions were debated
+the floor is now open to missus president of the women's rihts
+the floor is now open to mister sachleer
+economic and culture health have been greatly influenced
+part of the dear collics john pollad
+you enjoy this contest but you don't make any difference
+such such reality is the only one who is a good one
+my son is the best player of scravele that is so good
+and this emergency was due to the deficit we had inherited
+it is desperate from the opposition
+they lost an average five hundred uras per year
+i'm now moving to the amendment number six hundred seventy four
+you were in the permanent constraint section without any coherence and economic logic
+manette was not really taller than me
+we will see if we generalized it
+here is a reality incontestable and not a secret
+missus masineau the amendment is maintained
+this is the reason why we are proposing this amendment which has already been rejected
+i greet and measured a lot of priests although the other one is
+it is their male that understands it is captivity of doing it
+we will have time to go and get to the cinema
+it is therefore a continuous sequence of reduces the nominal deficit
+she's appeared in two thousand and sixteen and is already an intermediary trajectorio
+so it is that basic stones were good
+yet a legend poses coming out of a difficult roman
+i was happy to have only the truth
+more she's not text which favors the high journals
+as a personal note i will vote this legislative proposal
+but it is necessary to resume the question and bring the answer
+he votes on a project a directive line of vision
+last night the world will title the french people judge the tax an unjust
+i am completely agreed with you i am completely agreed with you you you will need to go further
+toliddy this budget is that of lost ilutions
+to guarantee you the stellar equality we will offer two measures
+defenses based on legal functions does not part of it
+mister secretary of state for me the counts are not there
+i'm putting amendment number three to a vote
+it is the territorial jealousy
+the floor is now open to mister eric bline to support the seven hundred forty sixth amendment
+the floor is now open to mister patrick ollier
+you can you can you can't read these two objectives
+this will call them
+this is now finished
+a few words massan president
+we are not leaving just anything like this situation
+let us do so carefully use the words
+i withdraw it mister president
+read all night long ends up straining your eyes
+this article deals with more than a confusion and will organize a cohere
+it is therefore normal that we are asking for a more complete debt
+the united kingdom the united kingdoms one takes a longest work period
+my amendment number one hundredad fourteen is ended
+so only election victories then
+this manoeuvre has a single objective to revolve the reality of the structural deficit
+i still have to do your endments
+the sixty first contract was signed for a long time we must all play each other
+by nature the conducted operations were not planned in the budget of one thousand and seventeen
+you were creating a few in grate and thousands of sheep
+mister john frederick pusson may now speak
+we are toning about it very easy
+then we speak breton not much common
+the wet rod is particularly fired this morning
+especially among your friends
+in prefer the chaos is therefore not a question of amendment
+he yes it's not the same thing
+mister john lukelaront has the floor
+yes
+we put forward amendments in this direction in your direction
+and always put an ice rose in the bottle
+that is why he always succeeded in the reception period
+in reality after three hours of reflecting this majority
+this grove i wish to burdenward
+at dinner with the workers in france
+the question that we are proposing is to start this trend
+the hucker watches your network and knows where your security breeches are
+no it is not possible
+it's really a good friend
+the rider in the middle of the roman horse is in a new manner
+the other danger is the difficulty to create the growth is therefore employment
+the commission gave a negative opinion
+the floor is open to missus elizephen to support amendment number three hundred eighty
+good luck it is you who are you
+indeed this work from land elena's work is not used as a result
+i would like to eat that the english language is earned
+intellectually i understand the reform that you were proposing
+this makes me happy to hear it
+and you and you and you engraving the situation in two thousand fourteen
+what is the government's opinion
+he asks you to with trot otherwise he will have to issue an unfavorable opinion
+this one already has a lot of things i speak about an unfavorable budget
+we are working on it
+if only it was true
+mister minister you have a lot of steel i recognize him
+my question is to mister minister of education in the national education
+yes the commune is a major topic fundamental one
+i hope you measured it during our interventions
+the plain lands with more than one hour of delay
+i'm starting to work late in the night aspires me
+the bases were put in place and built on it
+but the growth is growing on the legs and a fewer people
+we are announcing a reform of taxation of companies
+we are not any lighter
+you were creating the confusion well
+stepfa is archivan
+this new model of development exists
+i think i see it again the emper again
+the commission's opinion is favorable
+but then these structures answer the request of a part of the users of professionals
+oh yes
+i take this budgetary date to ask for several other questions
+we wantto also the success of this conserveability
+i'm doing well to the security mission for two thousand and eighteen
+the abbey stat epis is that of rents cannot be convenient
+here's for the expenses
+french must know that thi your programme for the defence
+write a letter is assuming activity for me
+my parents were a bit more instructed
+he is mostly essential in the crisis period
+the numbers are cruel
+no i'm here
+satisfied with this amendment
+don't worry for us
+you must go to the end of your demonstration
+we didn' we didn't get disappointed
+to choose mister the rapotter to be uternia
+he was defending an amendment it is right
+taking the sector of teaching
+we have seen te ma lat and we worked with them
+the president the president's voices are penetrated
+bat you for complacing it
+especially the distribution of the effort is uniform
+then to the oven preheated to nine hundred degrees four to five minutes
+the households are smaller and clear
+the true figures will be known to night
+a small community of the left is starting to open the eyes
+france to day is the seventh union grand union in terms of mortality
+you were in total contradiction
+the twenty thirteenth budget has thus the twenty thirteenth budget has thus confirmed the orientations we had adopted
+the floor is now open to mister shock pellisard to support the amendment number three hundred one
+the marker and ten events saw the difficulty we had to hear ourselves
+some of them are sometimes penalyzed
+a summarized for two thousand and thirteen is already taking place twenty two thousand uros
+she continued with a seuser of witty of spending inefficiency
+heap unlike this and therefore a total adherence
+i agree with the idea of struerg
+of course you would correct some avarations
+madam minister is right to insist it is the measures very important
+the first article a b was studied in an extremely precise manner
+there will also be some savings
+all this is a omplex
+when you tuck out the metropolis very good
+the jobs help have been ignorantly throughout the year
+i like horses to go through the woods
+the government intends to observe it
+the situation is economic and complex and no one will say the opposite
+even the rapater general is in agreement with e general idea of the same opinion
+second example the formation is profinal
+yes the regime proportionately advantaged employees to get the highest income
+the right is general except that it is a break the right is general except that it is a break up of equality
+it is therefore a few additional hours
+this is why i said we eat our white bread
+in my opinion mister the general repatter has to rebuild his name
+the musculation enables to keep fitting into the lasting of the investigation
+an old man curved an old man curve on a table raised his head at our sight
+after several steps he is now waiting for four point five per cent
+i'd want to run in the morning and scream the nanette
+when the plates are moving on it can create earthquakes
+this is more than four half months that we discuss with them
+rebilitated taxes also make sure that public action is effective
+but we know how to do this difference between avert and media
+a for having paid tribute to our forces to the mali
+unsurprisingly i am very much in favour of this loody law proposal
+the trunk is good
+it's the one that will surely keep the pregnancy to weaker
+the floor is now open to mister olivia currews to support the
+this worries me because this shows how much the service does not know the business man's world
+the floor is now open to mister shopp pellisard to support the one hundred amendment
+he even tries to remove the drafts drawn to social meetings
+from the second part
+i've given two identical amendments numbers thirty and one hundred fourteen
+the commission's opinion is unfavorable to these two amendments
+the points are taken on the cout may be converted into a five year old age
+i've heard that it was a peal of amendment
+i understand that this amendment will disrupt the majority and the socialist group
+then looking at me with a little nice look
+the floor is now open to mister shock bumpered
+this is the meaning of the work that begins for the government
+this surleyer will therefore be very fast applied in a certain number of territories
+it's particularly true of the issues raised by the bill
+it is an evolution surely guilty of our finances
+this morning the assembly began listening to the speakers listed in the general discussion
+it is editorial amendment
+so let's not get our pleasure
+the third reason is that this proposition is a specific report
+we are announcing attack on the excellent part of exploitation before she disappears
+that is the cost
+we didn' we didn't we didn't need an assessment
+to day on the additional sceminaries the state has no way to direct
+on the subject
+many socialist deputies also know
+i would like to ask a question
+mister president i remind you of this funding on article fifty eight paragraph one
+i didn't say that i was perfect
+i understand the arguments from one and the other
+we can clearly see that the dches will have overload of work
+it would be too slow for the rest of the game
+the issue of this article is framed by the accumulation of retirement jobs
+i'm putting amendment number two thousand five hundred sixty seven to vote
+it is the last analysis but it seems to me th at the demonstration is capable
+yes we can deny it
+i would invite the withdrawal of the national representation to adopt them
+certainly the past mechanism was not restrictive enough
+this is not acceptable
+the spedges are used for the flin and brod to women
+it increases in the hour by the way of the first time
+i thank you for coming with them my dear college
+this article is very important because it is clarified
+those who will pay for me are not those who receive more
+the delicious potatoes cooked with oil and cumen slightly lemon flavored
+there were beautiful meetings
+that's what it would be a precise spirit
+indeed
+it will not be a soldier
+the train of the delay to day is still visible
+are turned to wear it
+it's what you're doing in permanently
+evidence have a lot of difficulties will be a post
+the national assembly is not a bark
+add vegetables with a bouquet garney and cover with water
+here is the explanations
+i am stunned by the remarks that the president of the socialist group
+the floor is now open to the minister of state
+my question is to the deputy minister in charge of the educational sector
+the climate rule is not the refraining
+nothing that
+i could be able to give you some examples in my own district
+you your economy is bad
+through the innovation
+he must however be a reason for you would be in two positions
+i think that it is a good vice which is about the authority
+i prefer to withdraw amendment number two hundred ninety two to his benefit
+call for the wisdom of the assembly and
+mister minister on this point keep you on this point
+the construction site was launched on the ground floor
+i hope for one day to use you to go to
+after the economic actors it is the contruters who've gone through
+the floor is now open to missus nataly aperay
+it is the instability of a decision that reinforces its fears
+a consultation to bring us thereby imposed
+it is now completely ignored
+department is not a case for the accounting
+six thousand one hundred an twenty two from the public health code
+their level of social protection is very high
+and there are several ways of transfer or two delegates of a skal
+this will lead first by less than growth
+indeed tell him and take responsibility of your choices
+you don't know what i am reading any more
+he didn't dare
+the significent may be signed up in the united states
+we won't we won't answer the government sort
+no document was written to arrangement
+it is a real subject
+yours are yours are more
+i still love him
+but doer callits we have to get involved in this gap
+i'm going to try to answer the previous one
+this consistency is already supported by tools
+to modernize finally the organization of our country
+it's a right
+i'm i'm assuming this adjumant in front of the national representation
+they expressed their support to this experimentation
+the french people have well noticed the reality of the tax bludge
+the first objective was
+we would have been so much that we have have taxed ability
+in definition we are witnessing the pursuit of public spending
+the floor is open to missus the deputy minister
+i've given two identical amendments number one hundred eighty six and seven hundred ninety four
+i would like to add a few words to remind and depend on the position of the law commission
+the amendment as defended by our colleague john liglerand is a member of parliament
+unfavorable opinion hence
+the university is closed every month of august and is a public education
+we didnt we didn't react to the laws before going to the end of the year
+the floor is now open to mister francois papony to support the three hundred sixty fourth amendment
+it prevents us to spring on as legal difficulties
+take responsibility for your choices she said to be a good job
+but new measures were heavy and planned
+cheese hot and you can see it in the middle of the world
+the four big ones are the only ones who are against the other
+why then
+i think he's capable of bearing a heavy burden
+this place is likely that of a high school
+not only one in these neighborhoods in the countrysides too
+the most beautiful castle i sought
+well find you
+you love it
+books are the most e
+forty one dogs
+the customs pointed towards the direction
+i saw it
+the store abiding it classes
+a bottle full of water
+which one is yours
+a second
+it's too big
+the time of dinner was ready to be in the winter
+the chair is broken
+twenty three black houses
+my father is a lawyer
+we will not let him down again by the winds in spring
+and he has his wife and his father
+why are you not going to go to the movies that you are not going to find
+we can we can see the player or so much
+to be dictious of the journalist's name
+so he opens the door and enters
+he came here all his own
+if we could send things to people to day it would be a good thing
+the dynamism of our demography is partly denomenon
+twenty three elephants
+twenty six yellow chairs
+applause from the benches of the socialist party is diverse and on the bench
+m i e n e s a i e s e s the one who is nowhere to the other
+this one is the only one who is in charge of the two thousand francs
+thirty dods were coming from the ground
+she retired to the national team
+i'm buying a lamp as well as the other one and the other
+you should come the fastest possible
+my brother and my sister called you often
+where is the house door
+everything is the daughter
+she put her shoes on and wer
+batrfr your answer
+three men
+nothing else
+she's loved
+twenty three black elephants
+difficult
+another book
+whereis yours
+we got up our hands before going to the table
+don't you come back
+he is reading ha news
+this is to whom is it
+you can believe what i am telling you
+salve also doesn't salve also doesn't come
+my courage to all of you
+this animal from australia is a major interest in the united states
+forget the past
+we were absolutely alone
+how old are they
+he moved his nose without a doubt
+what are you thinking about it
+come on
+i want to talk about the creation of the gasolene report
+this budget is only the translation of your stubbornness and incomprehension
+ali imaresca
+at the same time a specific seiling would be established in order to limit the effects of abers
+if i have to speak french more often
+i buy the newspaper at the store every day
+so why did you call me
+don't stop us from reme
+this is your brother
+fill it up please
+the area of the island is small and small
+what did you open the door for you
+very big
+he has a headache
+while going slowly you want while going slowly you won't be heard
+you should ease here
+he smokes
+i am very sad he is very strict
+their work
+thank you for your answers
+i am not assure
+she's a girl intelligent
+of course he is a member of the national assembly
+my phone is not breaking down the shadows
+i din i didn't go there
+we will read
+he told me that smiling to me
+what chop do you have to do with you currants
+more than one of you aren't walking fast enough to pay for the rest of your business
+that he became cutif it not true
+we tried to explain it clearly to you to explain it to me
+twenty one black books
+when reverse rehabilitated
+they are toning about it
+you you're doing something wrong
+can you ask him to come here
+it's over there
+i am surprised that i like it is not a question of principles
+how much time is the next train
+some of the thirty m b groups also offer it to the same amendment
+it is not an astrite budget and is all about touring towards the job
+you look like my sister and i will not go to the south
+there are some
+at thirsty i am
+he sick
+the oldest of elected representatives
+why are you bringing me without reason
+the best one you can
+the following year
+the little tiger
+the female meale
+what a strange thing
+im im i fear
+the dogs can obey the shepherds
+his is the one who is now in charge of the building
+some quickly they could run and it is too late
+i will take the same thing
+if she fishes
+i loved it
+boot there
+a more important question
+there was noise at home
+the machine works on the other hand are not used
+some prioities must guide our choices
+he was the conviction in this fight with two children
+it's possible you think
+he became without being invited
+it is a political choice
+twenty three is often dead
+a beautiful flower
+thirty two women
+such a
+you'll be able to bring the chart with it on your computer before the meeting
+and you have time to read the newspaper the morning
+itiis book
+i owe you to take him to cabaconam
+an old lady
+i didn't i didn't i didn't have decided yet
+the song of bread is made of two colors
+this does not have a final end if we have to invite everyone to do so
+the transportation shaal is also a small number of locals
+this tower even n the port
+let's be happy to fend you
+i find the proposition of francois paponi interesting
+a ticket to the west and the north of the city
+as i am telling you i am telling you that i am telling you
+the restoration which has never been produced is not a good thing
+alon is near the city of the same name
+no cheer is the same as the other
+maybe she will come who knows that he will come to the south
+our three sheets are made of a single paren
+nine o'clock and half were looking at the bottom of the room
+about thirty killas of bengales and half of the cars were found in the air
+he drinks from the past nine to five minutes
+where is it
+we will put either some wood or some stones
+you what
+these respectfuls of the environment are the only one who are against the two parts
+whatever
+is it possible to have the note please
+their fathers
+go away i haven't been looking for the rest of my life
+whatever that will happen and i will be there for you
+i want to give you a gift from the right time to the committee
+they read at last because they finance your c ic e you increase the bad weather
+the turret
+we will be numerous and more easy
+his parents are separated from the national team
+he soon arrived at the end of the season
+the beauty of the time is very hot
+we brought changes to the text
+why did you go tell them lies
+nobody else knows that
+so they can fish
+that's balanced it is true that is something else entirely
+i am listening to you
+the church is modern art
+everyone had come to see what happened to him
+the most expensive ones are not the most expensive
+there is no bus to day
+denser that the wines do not go on
+she's going very well in the middle of her life
+where am i
+there are sick people among them in which they are eighty thousand and two hundred and fifty five
+you will get with mister moyne and i am not going to play with you
+i've married to the americans who are not married
+thirty one on the other hand are not allowed to see the public
+my house is next to the bake and the sofa
+how do we do the business and the general we're going to do
+the freshest strawberries
+i i'm telling you yesterday i'm telling you yesterday it's a fate budget that you were presenting
+to morrow you will wear a warm coat
+the birth is the moment of hyper suspiciativeness of the french people
+a part of our story is our memory of the greater than the other story
+he was here
+a more important lesson
+it's not fair enough
+he wasn't he wasn't he wasn't as solid as i fought
+there is no book on the table
+he's going to kamper
+the best team
+who will find a job in the city
+his own
+is he working well
+the house is built in the middle of the town
+wait a minute if you're not going to wait for me i'm not sure
+this amendment corresponds to the sub amendment of the down payment and dissolved article number one thousand eght
+can you throw his art for me
+their hands are not very well known
+what time did they arrive
+around the square
+youhan
+she is too beautiful she is with a beautiful idea
+he will have to be somewhere else on the other hand and the other handar going to be a hard time
+there is a few books on the table and half of them
+there is no lot of people who aren't many of them
+twenty two
+we love it
+holled everyone at the same time
+nine o'clock in the morning to go to the south of the country
+troubles i am going to the south of the country
+four big chairs
+what buswle in ten score
+you think it his sister
+twenty four black chairs
+there was no time to go to the right of the house
+she went to see him early
+how much is your sister in school
+we'll stay reading as long as there are lights
+several measures are favorable to the households
+here some letters that we have learned
+my mother is very affectionate
+hope this is him that i live
+he is here that's all the way to get to the public
+he does nothing but eat because of his father's help
+pans that are weighed too tight
+you are coming to night i will not have any money
+this person over there
+this is not a question of very educated except that he knew how to express himself well
+he is there
+despite this he will be the one who will be able to findout
+happy new year to you
+believing himself as yours
+we told me to go to school and then to go to school
+finally his birth house presents economic and sanitary interests
+he is capable yes or no it is not a realisticated person
+she is calling her sister
+a beautiful story
+two white houses were built in the middle of the house
+tentations of functional are not leading the measurement on the other side
+why are you doing it my dear
+sa lot is a new name
+there is no time now
+i i have the luck it's not good
+there is no charm in the garden
+you were trying to make this forgotten
+you will have a lot of coffee and tea
+i said she is paying them in part i said she is paying them in part don't carrate the words
+she is fine and i will not have a nice chop
+it's wise it's the most important thing to do with the same thing
+you don't know him
+i would like to go to the house
+our family is a combined family
+what are they steadying
+where is the letter
+everything is full
+i eat paster often
+he showed himself like you and me
+my dear callis please follow un sensible there will only make him speak
+he put his shoes on his rambart
+in leader in front of the year it is also possible to have a large number of wang
+well they're things
+i would also like a glass of water please
+called to the police
+i'm often mistated when i'm in a hurry
+do you want your brother already in california
+i wrote a letter for him a letter and a letter
+here is exactly what we want to learn to learn to you about the future of the republic
+what are you doing on sunday
+part ukon
+it's for her to be a bit of work
+i would like to know when you be able to come with us
+twenty three tables
+the tower is made of a small stone
+they are asking for it
+i just gave chicago
+he doesn't he doesn't he doesn't want to expel the milk
+idn't want to eat anything
+does it seem to me to miss smith please
+you are sure we like it
+we are soon insured a tax on the advertising in the hemicycle
+recently i saw a man
+the parliamentarians from the lest front are not really understanding
+happy to hear this he said while working
+it is the holidays
+she will be with us with our fellow citizens
+he's the one i like the most important one
+more and more people are playing with a second one
+i have nothing heard about her to day since you rote me a month ago
+twenty five onde pallet street sixty one two hundred in saint jean
+he was pitched by the night
+two hundred and twenty three door leon street
+there is an obstacle against me and the purpose
+the weather is cooling down
+here she is
+at the library
+it's you hadn't it's you hadn't been here i went
+three by three
+here's the minutes of the region that we had the last week with harry
+it's passionate about it but it's not important
+there is not a lot of rain
+you are sure we like it
+six chairs were built in the middle of the building
+a fool as an ordinary skirt i would like to have a red hair
+it is the fault of kesfold and the french people
+fifty three birds are used as a result
+immaculate he is a foolish man
+i have two brothers and one sister are still in the middle of the world
+we didn't we didn't know you had arrived
+here you are something about it
+does he come to the office of the city
+i i'm asking you to be administered by those who want to speak
+this is the last one who is a parody
+it is also a question of gari
+a bottle of wine can be found in the middle of the garden
+he demands that he pays
+put your shirt down please
+there is a well but water and bracket
+in short i resigned
+the house is pretty but it is not very good
+what do you think about it
+fishing is done in the united states
+they were loved by the public authorities
+these concerns are not found kenser or in the brain of his wife
+the lakes are leaving for health without varying depending on the season
+more ancient
+travel was difficult the sea was bad and the wind was blowing strong
+i beliave
+he is not talking about making this as you were suggesting
+we will come all of us from our country to the south
+don't don't don't you have your ear
+i am extremely satisfied with the idea of the general duberg
+to fight against the two thousand five hundred fifty five
+no other chair is the same as the other
+we are therefore a little complex in your puzsy an economic puzzy
+what time did you wake up on the right with a few did you wake up
+let her fall
+i am not a girl
+but business does not want to do so
+he bought some beautiful clothes for us
+a dozen chicken eggs
+can you please arrange this for me
+mister general do you think it will re establish a little bit
+they loved it
+it's the curds that won in cobone
+no it is the first time that we have to go through the same way
+you were here and you are in the middle of the world
+they're running
+she loved it
+you were tired
+don't move
+automobilistic has not respected the code of the road
+it lives in rene and the city of paris
+a family is a very classic family
+on the top of the grandfather's top is not tired
+the term male tort is a bit of a kind
+it's yours
+he opens the door
+i am sure he would like
+i am interested by this book
+i am sure you liked it
+the first me is the first a
+never go any more he said to have a look at his father's
+here's not eating it these people that's not the same
+now she is fishing and she is fishing
+charles vinson is a member of the national assembly
+don't cry any more i'm not afraid of it
+very large and very big
+what are you thinking about
+i am sure that they liked it
+twenty four
+you read
+where is she from
+from a thousand eight of the bishop marks the end of the world were two hundred and twenty seven
+scydalus is a major figure
+you loved your sister
+the child in the arm is broken and the door opens up
+my name is roger and i am calling for an answer
+is it the costs of money at are not the only ones in the world
+which one is the most important
+you will have more than others
+proofs that he would have been a little better deal
+dtars are also a member of the national assembly
+encourage encourage growth to the public
+for them it would be a help to be honest and i'm not sure
+how long is the good in the country and the new yorkshire
+in your room the door is open to the public
+at home she is too tall and reading the lands on the right
+but sometimes on the comparison of the device they are not very well known
+he had learned a lot of son's being young and a little bit of his father
+the weather was at rainfall in the most of the time
+he works badly
+in briefs
+the commission gave a very favorable opinion
+how is he coming
+he must understand that he is not a poet
+in two thousand and thirteen four trimders equals seven thousand five hundred and forty four
+seven hundred one eighty five
+understanding a two of the species is also unconscious
+no i am not sure that i would like to have an egg of opinion
+i still have a lot of things to do
+it's like this it's like that
+mister philipp lorandferd will take the floor to support amendment number seventeen
+our amendments number four hundred sixty eight and one thousand five hundred ten
+montel is a dacir to build this extremely complicated weapon
+we understand why this name is not applied to the public
+the floor is now open to mister john jacques kurbor to support the thirty eighth amendment
+twenty one black books were written in the same way
+it's your opinion that we can see the same thing as the previous one
+the floor is now open to mister damien a bud to support the eighteenth amendment
+his roof was trying to be a good time everything would be pulled out on wheels
+the reform of the refraining is progressive
+he prejudices ecological and will not be a machine hide which will not be based
+philip offers him a chair
+mother you boo what are you unhappy
+he's aiming to introduce new misses about the new zealanders news
+let's see what did you go against me for a while
+and she will come
+i am sorry to learn that it does not work in your region
+it is the reality of the town of the same year
+i understand what logic i understand what logic you're going to have to set this imperative dead line
+i'm doing more i wanted something else
+most of the departments have no transport payment
+it's missus de cleffort who is the eldest
+he's standing up and goes up
+but there was an extreme in this movement and got up to him soon
+so i thought that he would learn english and not to make a single employment
+back and forth on the other hand the latter is not very expensive
+six led to him with his morever murder who promised to kill colony
+people who know the secrets disappeared from this world
+for her there was no other thing
+he is due to cangena in the eye
+you can be able to apply this to me to the public
+i need to see a doctor
+if it's more stronger then i do not trust him with him any more
+you didn't want to coffee or tea
+the floor is now open to mister secretary of state to support the ninety eighth amendment
+the floor is now open to mister mark dolles to support the four hundred fifty sixth amendment
+violent players were coming back to the national team
+dont don't don't be worried of him young man dear friend i'm not sure of him
+the floor is now open to missus chuntal berthelot to support them
+she lives mercelin ivanoi
+boys ordered trinning go by
+i didn't i didn't have lunch at all but i didn't have lunch
+some grains from the chapel were still showing silently
+later he had the general command of the cavalry
+mister dicharna is staying and sat down and sits down on the right
+without this we won without this we won't get it and we will not happen
+the commission's opinion is thus in favor
+it's my older sister who is my older brother and i'm not a poet
+he will kill him roxen and his father is a member of the house
+the north corner and the baron of newstitz but noon in germany
+what you have then been asked for it to be a good thing
+but while bombing he was drinking this man is still more than a low voice
+the french specificity in this area is based on exceptional assets
+he is not too far
+he leaves a significant squadron
+the catholics in such a few years jacques wanted to first
+fire in t place on the fire place and the five place
+many of them were even bigger they were the polthiosis of the furians
+the floor is now open to missus chantal gatette to support the two hundred twenty sixth amendment
+the floor is open to massa minister to support amendment number five hundred thirty three
+the c a s must must be retained and i am therefore not in favour of your amendment
+the german literature were for katherine
+our loss our loss is three men killed and fifty injured
+he took the presses it altogether like that
+we are not proposing here to return to it
+the seventy second one distinguished himself in this day and fifty seventh sepport his form reputation
+i support the amendment of mister laurent is very important
+mister sirgi cornato has the floor to support amendment seventy seven
+the influence of my father ceased this tour now in my mother's tol
+these criteria are to be even more securing the device
+at that time the entrance to the country had just been made against us
+a good number of prisoners said the jailer with a smile litigation
+colony signs at la rochelle a treaty with the nessau
+the objective mister benois is well generalization
+she was livid and had the lips to be seen in the middle of her breast
+we found in salsborg rather important stores
+it is our turn to france at the battle of cressy
+a second problem presents itself
+i would like to join a simple sum i would like to join a simple sum sis median and mister
+she was soaped in palace with her husband
+it is a problem that will be very popular with the results quickly
+he had to have the painful suit in their paper
+the prohibition in two thousand and eighteen is unreistic
+this interests me to participate in this work in the commission
+micky finally became a winner the clergy
+and you watch out attentively
+rigulus remained alone
+we need to respect this expert committee
+he had been planned by the government of legislatory
+however the group is acologist and insists on the role of the regional representatives
+mister secretary of state you have been really viricler
+citizen manager the army started walking on the do
+amendment number ninety five from mister giles sovery vapater is tetorio
+yet it run these conditions of the opposition exists
+he shouldn' he shouldn he should't have the simplification measure will bring us to this question
+the moon was rising in the middle of the brenches let the moon was rising in the middle of the brenches let's go beautiful and take your gun
+and then the king repaid herself in the middle of the war
+i am convinced that o commitment as in the state will contribute to this awareness
+it is therefore a rather definitive lock and we are getting on
+before she was the charity that took charge of the needs of health
+gabriel made the story of the murder and troubles that followed him
+jean let's go his neighbour and his neighbours are in the middle of the house
+i perfectly claim your pathetic approach
+she had seen the scene and spoke low in her eyes
+and would you like me to tell you something else
+this is correct and it is a step full of paper
+dasel street thirty two thousand och
+sete poviu street at number six
+the floor is now open to mister e blyne to support the one hundred thirty fifth amendment
+twenty six a general lamark street
+one hundred and fifty three a pinnaced path
+ducoto de bor street bribes suronsac
+i therefore invite you to the wisdom
+nine marie curi roundabout fifty seven thousand one hundred seventy shenesala
+what is the commission's opinion it is in favor
+four per four peros residence ninety seven three hundred saint quentin
+twenty three corp bo square sixty seven five hundred ninety in west huff
+one thousand one hundred eighty one voi commune one thousand one hundred eighty one voi commune dela grave eighty two tasand hundred moisad
+boi communal le press saint martin in parisee less forges
+g under motie at number eleven
+fourteen de bulets way forty three six hundred in saints
+the present amendment concerns the tax on the c a m
+castelburg road labor
+louis morsini street say no runs to murney
+swin road faverges de latour
+two hundred sixty six riceba street eighty eight thousand eight hundred bido
+the floor is now open to mister erick allows it to support the twenty third amendment
+one thousand seven hundred thirty six boulivargin teath bell
+three friend street zero nine one hundred eighty in saint jean
+twenty currink harris way twenty nine thousand four hundred seventy invalids
+one do fronted alley
+one hundred sixty one tornier alley forty five thousand seven hundred and seventy
+forty three leon avenue
+pal chendon avenue fifty one two hundred eperne
+no one else is going to have a hard time
+the floor is open to madam minister to support amendment number three hundred fifty nine
+eight stilo pathway thirty one five hundred seventy in prets
+ah good child a ah good child i'm in this nicety to say that you are in a hurry
+sat re de madu rode forty eight one hundredand forty le malveville
+four bombay thirty six two hundred in colino
+brind card brind card avenue zero six one hundred nd sixty umpes
+port saint leonard street at number two
+rime avenue forty eight tosand mendi
+the wood ear twenty five hundred less cullins
+charrier ben at number five
+de la sedre alley seventy six one hundred thirty monsed ina
+this is starting sidmarkov noticed
+as for miracles the reason was not surprised since they were unknown
+the floor is now open to missus francoise dumas to support the one hundred amendment
+from then on we must not hesitate to sanction them
+this amendment number three hundred eighty one is incompatible with amendment number two hundred ninety
+the roseberry roust is going to explore everything nights on the table
+there is only one way to get into service and the other is not to pay attention
+but we have the habit an idom but we have the habit and i don't even know why i am answering
+lucian cartey street bell ford
+one hundred thirty nine severs rode seventeen thousand one hundred seventy
+it is all his friends at the head of the country
+what's the matter with the same name as the other one
+old saint louis street at number three hundred and fifty five
+stock canon street fourteen thousand eight hundredand sixty bilen
+eight hundred ten rural road eighteen two hundred baldo eighty four thousand three hundred eighty four
+he needed
+i remind you that you have precised it even as we are here for a long time
+jasen de monteris boulevard twenty two hundred bastia
+he started by claiming a car for the second time
+a thousand of honour i didn't know any more i didn't hold it any more
+come on get lost my head and i am looking for it
+it's yours
+he's crazy about her he doesn't get her unhappy
+i prefer still better to go to the weddings
+excuse me i would have some for you
+twenty two annette street
+hey you're wrong i hay you're wrong i'm forgetting you change the social reasons
+the emperor was giving him the cross and the emperor was still unconscious
+cotel avenue sixty one zero zero zero alancon
+from then on he is useless to go further
+it's to morrow that he will happen and you soday's eclipse
+it's true adriana
+this amendment allows for the raise of these ambiguities
+thirteen a la rieged street eighty nine thousand five hundred thirty sat bres
+his majesty was very presenting the tabor county that he was now a coach
+a ford of the house a gateway of the petite borgun street in the middle of the house
+sir asked the doorman who saw them where the cabik asked
+the rapotter in summer missus barber remindin
+bardr street at number sixty seven
+dart way zero one seven hundred meron
+two hundred six malboro impass twenty six thousand one hundredn seventy in bule
+i'm stressing this and i persist it should be better to go back to the bear
+bogeron street in blaze
+lulodie street in saint martin glend
+bus bus in plinel street at number fifty one
+ninety nine a on minigan way ninety seven four hundred imbos terror
+doctor morry street at number thirty four
+three hundred ninety one kemen de bucher o six five hundred thirty
+chambagnele street number four hundredand twenty five
+two thousand sixty nine two thousand sixty nine p h d i n i e street zero one one hundred sixty
+sixth imma morezi street zero two one hundred san quentin
+gor montagne rode zero nine one hundred eighty in saint germain
+jury street innase surgeon
+isabel orasin lot number one hundredn twenty three
+de la fontaine de sharwa ebenes
+dukill street zero five zero zero zero gap
+tisn't alli in sedan
+two thousand five hundred eighty rudu colde brue
+what you want to do with this
+well said
+its the gothic legends became spectacle theologians are now only known for archaeologists
+eleven forge street thirty five zero zero zero in ren
+and you boring me
+this operation made especially with the management centers for ten years
+three hundred twelve bruno mercier street ninety seven one hundred forty one in ver fort
+six strong were the habits of monarchic habits
+delos had run away
+blanche takes the place in leonita and felix to blanche
+i think sir will do good to think
+the thing is executed
+blanche fort rode in the blanche fort rode in the ger
+we are frequently referring to in the length of the unpassed one meter away
+i would like to go to the shop
+twelve s faka street eighty thousand two hundred duay
+one hundred twenty four on dela defen street fifty seven thousand three hundred eighty five
+the mayors find themselves completely diminited
+the bees were ripped off
+you were compromising your success
+i know where these people are dead said the boy
+these figures regrettable are revealing a real crisis within our city
+she doesn't she doesn't she doesn't have this woman here
+i refer you to the species to my original argumentation
+this seems more reasonable to me and goes in the meaning of harmonization
+at the back of their antechamber door leading to the stairs
+i am particularly shocked by this kind of mendment
+let me know you will be our queen of all as you want
+from the gesture he reminded them of his rather than a hundred kilots
+several were taken out
+de campaign bore street in saint martin
+i am also maintaining
+three bertrand banlvard
+we will get back to it then
+duport de chassea road sevrer
+for ale de debrians thirty eight thousand four hundred sait quentin
+juli coyne monso
+these pretty ones are not allowed to see the other
+a square metr and quarter
+the floor is now open to mister alan le buff to support the four hundred fifteenth amendment
+come on the day i'll help you
+i was far from waiting for me to honour
+what is the government's opinion favorable to the sub amendment and the amendment
+on the right first shot a table covered in a green
+this produced a very unexpected theatre
+the priests tried to get a view the priests tried to get a view from the mother of guys
+doctor at n d street intelos
+doctor pierre bed street sixty three zero zero zero gap
+ah he can ah he can ah he can read and soft i said it still
+and yousi i'v asked for a few minutes
+eighteenth place in chateden in the middle of the thirty five
+the floor is now open to missus estelle grillier to support the eight hundred sixty fifth amendment
+young girl doesn't read when she's not read
+i would like to come back to the arcel on orbanization of the coastal municipalities
+i managed to do so far
+i ave given two amendments numbers two hundred fifty three and one hundred thirty nine
+he put it in his gusset
+when he was ahead of his team he opened the door all fears
+captain args the mars to the sun's cushions
+de lully in sack's place in x
+port oliver street bezes
+it is e ssential point and we hope that this amendment will be adopted
+our promise is that it is better for all french
+in front of this new placard a large counter to the accessory
+this proposal can this proposal can this proposal can i think that ou reporter
+the eco grande syners cylinders the arch a smoker with a space
+lagabe met last week
+ho but what am i then
+we will take the money from your count on the fifteenth of the month
+as strong as a bull
+the deterrence of nuclear is a political instrument
+i will not say anything at all before using it
+i i've done bad by he interrupting her
+i have received several amendments bearing additional articles after article five
+the amendment os the amendment doesn't talk about public institutions and corporation
+it was me who was taking the banks in a low voice
+undre planche avenue at number fourteen
+he is fairly number a
+in the course that in the course that's a matter of fact it will be greater
+i am sure they will like it
+i would like to make a shock like this
+two codes of the sector are not very well known
+four freres charles street fifty one thousand one hundred and twenty three
+some of the people n some of the people can run there is too late
+pilonier there are you hearing its coming
+chillinte street at number thirty eight
+de rolinville street ninety seven thousand two hundred fourt de france
+mister sebastin denaja has the floor to support the amendment
+passage to dasilo epinoi
+it is not tarted it's time for the pull
+this amendment uses the recommendations of the defender of rights
+de cote rose wrote eighty one hundred castes
+on the left in the foreground against the wall a small secret office chair
+really you re terrible
+adgree could impreise the definition of the determination of the determination
+since it's not raining we didn't go to the movies
+five hundred eighty six terrene street eleven seven hundred asers
+we are going to proceed in this logic
+for a gentleman is ringing in rights regulations
+chendall avenue forty four four hundred roseas
+livins rode at number one thousand eight hundred seventeen
+che publisher in paris
+it is tulia that is a very good thing to do with this
+the floor is open to missus isabel addered
+it is indeed a serious topic
+eleven plan dilatore road eighty three one hundred twenty sad x
+we must therefore take into account this article
+ah ah she is invited to us to come
+one thousand forty lechet rod zero seven one hundred seventy in sassart
+he doesn't have any laugh
+the floor is now open to mister john louis briqet to support the amendment
+madam rapater you told me earlier that we had worked well
+in all that he had twenty five miles
+he goes back to his cast at the back
+mank rod node longen
+nine hersa street twenty one thousand eight hundredand fifty saint apollinar
+you didn't see the figure he did mister de roshald
+while he was climbing the floor he was jinny declared his trip and the new wealth of the family
+bernes's dead end eighty two two hundred forty laville ox teames
+these means are enough and they are not useful to add
+the benel arsenal is all ready enough for the public
+it is an encouraged officer and deserved deserver
+he offers him ma chair
+the soul this man has been revealed so far it was stolen
+it is about anything
+however this number is stabilized since two thousand four
+the device predicts an abetment during sixty months
+at sixteen he had a fleet in front of spris of
+miss shellen is slandering us sir
+these two franks
+i am happy to learn that the government is not planning in the future
+if you can see in the subject of norman rotters i invite you to seize
+well yes alive said the captain
+eleven cret boulevard ninety one one hundred in corbe essence
+it was a first grief
+butbutdon't worry about it i am not concerned about it
+sixteen fornet street fifty four thousand one hundred and fifty four
+was it an ambition or emulation
+the derout looked the derout looked at the cout of all his eyes
+it is a house that is a house is not very good
+invited two ssecals to the public
+gevroch was a swarming turn
+i understood well my dear collick that you were scared of indulgence
+the servant is ringing on the right and leaves the other hand
+the floor is now open to the finance committee
+you were waiting for forty nine
+the mass of tickets is incurat one hundred million zeros
+you would like to know what is not a surprise
+the loss of a child came from conversing his life of tender and poor
+the lit is arrested
+noes had insisted on important movements
+the group deals with lurd i'd like a casser
+we were called him as gascon to nordon of these millets from the miller of barbis
+de la deverset impasse sa etienne
+the englo saxons that would listen to us were summarizing the date by top of an or bobanap
+the floor is now open to mister g nl carpenter to support the one hundredth amendment
+george clessino avenue number twenty five
+not to mention how happier you are here
+you were all with the exception of the sister no awful accused of that waited on the albert
+the group everything in a painting like to have it it will be easier to see clearly
+it is a very ingenious job and most pleasant nature
+all of a sudden she started sinking a weak voice like a suffle
+neither sir all is in the room
+the anxiety of his intonation was moving in alban with a secret anxiety
+finally let me sheare my hat ask me and let me
+his arms were found ready to grab his treasure
+eight on delagasell street thirty thousand nines
+we didn't say in front of us that woman we didn't say in front of us that woman's name was monstrous
+plan shuts impass egle ruct
+in duell he was hanging four children
+a chambermaid from the baroness is also a housewife
+we have created the sacrifice of our life
+eighteen july lefront
+de creek road allorn twenty nine thousand eight hundred thirty plutel mason
+he accepts the destiny to see her it hekel
+sir clay street in mages
+seven pot de lerook rod fifty thousand two hundred mantoan
+fourteen syrick street zero five zero zero zero gap
+their tenacity was such that he would have defied rather than the cowardice
+both of them started their turn and won the pinnerd district
+number one hundred twenty three on the palabear allotment
+these sats passions are not the only ones
+first of all y flora
+the permission of a smile
+i am arriving from corbva
+welcome
+other yellow animals are in the same direction
+madame champode taere takero entering from her carpent
+forty six chevalier geller street seventy seven thousand one hundred twenty seven lizon
+grampe wrote ninety seven thousand four hundred and twenty one saint louis
+the leaves have been pilled up leaves piled up in pillows
+my daughter is not here that is not a secret ballad
+mister the repatter participated in this adition and the commission of justice
+he is right mister christine pillard has the floor
+oh i have to go myself
+the question is really asked about the content of the issue
+five hundred sixty four adov stre street fifty nine six hundred ninety vicandi
+and saturday of a nine o'clock at eighteen o'clock thirty
+she is the only one who is a poet
+chi van bencover architect alone putting a friend
+twenty six unremassac avenue eighteen thousand four hundred sate flowr in serthe
+i've given the editorial amendment number six hundred eighty one from the government
+the more we walk and more tired
+he asks for the woman who push him back once again his wife and he is still in love with him
+once it isn't once it isn't once it isn't customary i share the opinion of missus minister
+l bangar street in bullineser marn
+the most as he can do without any doubt
+cone's boulevard manaske
+here we are in the middle of the building that is a great number of companies
+the address five drew streets is a major step forward
+lady since a year we are ddating soupto
+it's the one who is yours
+but se koje wrote twenty thousand one hundred and thirteen omedo
+the observation of mister corson is very important
+fourteen rujon jock ferrer sixty eight thousand four hundred sixty luter back
+what the devil are you doing here and here i'm going to do with you
+ah this is the time that you want to obtain me
+they show their regret to see their jeans abandon themselves
+the ones are clear the other boots are the same as the other ones
+i consider it is already the case
+i would like to submit this idea to the reflection of the national assembly
+he has nothing left for the sound of a manand even in a beach
+we must obey the dead
+thirty six
+they lost the time to occupy strong positions she had appointed colony
+one louis ogre street ten thousand one hundred twenty saint andre less vergers
+we wanted to force a bit of his head henry
+i thought i heard the mar de gratrops
+your amendment is fully sfied
+it is still quite useful to have frigated
+she still reasoned but inside
+then he remained silent
+sub amendment of ordination amendment
+eleven dfeld street fifty seven nine hundred twenty at ke sir kinner
+what i am in a chattering mood if you allow me
+alfred goria street chempley
+our commission modified the name of the rents used as a reference
+i was wipe and i was crossing the small watercourse
+this complaint of the defence that stopped the representation of his drama
+yes i'm not a thing i'm not a father no
+meanwhile he took four years to play in the united states
+they were the thinking of hopes for the most part
+the floor is now open to mister fulley vidger to support the five hundred eighth amendment
+these structures constitute a key factor of stabilization of the economic activity in the territories
+this requirement exists before the exam of the bill
+he is the campanian he's a bit the same rapiter
+we 're moving we're drinking some time now
+she will hit me out if she was a good step
+she changed colors and was looking for the first time
+the hearings have started late july
+along kemp street ninety five thousand three hundred carbs
+very well the floor is now open to mister mark fesno
+it is the same after all and his name is herf
+she mostly left the fatal competition of enrid she mostly left the fatal competition of enri danjou and enri de gaise
+the coast does not seem to be lower but the law was stending up high mountains
+to discourage the migration the ambassador of a f o took a police
+the floor is now open to mister denis boppin to support the six hundred eighty first amendment
+she does not represent
+let's be quiet there's not anyone that's not it
+lazini wrote surname
+he's not bad at all that time he's not a single
+he disappeared at the back door
+he would be in the ice on the right
+forty six avric boulevard zero nine one hundred eighty in saint jean
+here is the man who would make you happy and i will find it
+i thank you for your present amendment number two hundred and twenty seven
+the floor is now open to mister lionel tardy to support the one hundred thirty third amendment
+it is not our proposal
+de lawrence street in chaptell at sir mur
+say german who is talking in his cadence
+maybe this mister moroy could give him useful indications
+this came to me several times in a committee
+the first general farmers are in better than their successes
+there will be some sir
+it is the first time for two years that we think i disturb you
+it's run a large part of these rules are strictly national
+ah lady i've trembling your service
+twenty lignet street eighty thousand four hundred pvo
+they are commonly working or this won't work
+bvari street la saint sover
+it's a tool said injo ross
+eleven o'clock in the morning to this hour she is spoiling
+i know if he's going to come
+rue de la capatale in marseilles
+how many peopler
+the floor is now open to missus emily bon the floor is now open to missus emily bonevard to support the one hundred sixty sixth amendment
+he is already eight hours and a half i don't think he will come
+four chinese ali forty two thousand three hundred ninety villers
+thirteen laparel housing estate sixty one two hundred twenty in baler and lon
+we are going to do so in this way we are going to have a look at the same time
+bell rose led him to sisan and he was still a little bit of a sudden
+mister the rapotter i want to raise your skulls
+a note for you my friend
+but the midland had been wrong since i no doubt since i have noticed that
+but it is about the priority that a french work has been done
+clusi fights are being held in the battle of klose
+we do not stop reminding him from the well
+can you please give me a small service
+mister dillon was out
+it is time to reload this anomaly
+it is at least a trainer
+ofhow the denties of his speech were bo
+o my god o my god
+tos rode numbr fourteen
+you lost all your money
+charles du vickstreed forty two hundred thirty tus
+four co de aperels dead end forty eight thousand one hundred mendi
+trade from genoese for the thirteenth century
+to philip an boulevard thirteen thro zero four verse
+for these three reasons the buddy group will offer a support of this article
+it is not our erra that we have to admit that it is not a good time
+mister nicholas dewick will now talk to support the amendment number one hundredand forty three
+the device is well structured and the public
+victor gasford passage ninety seven one hundred bastere
+when he left the misfortune wanted him to fight against a chair
+armed the canoe the young people stepped forward
+the king of navarre the king of navarre took the king to go to the rein to make them lose them less
+the night is cold and lighting the cold
+third servant moved forward to me
+he painted his keys with his real colour that we invented anywhere the excess of his despair
+very well the floor is open to mister erikworth
+twenty three
+hi di cheto ale seventy three thousand two hundred and thirty sgendare
+wat still she said while becoming red
+there is no house
+aldoor conpetane see that she is a good man
+the first in the tone of fertility is the last
+eleven pol kernel dormano street twenty zero zero zero in a jaxio
+i did my university studies in washington
+well the number eight
+sir on very day leaf isn't
+see if mister pinglin may speak now
+pla street seventy four one hundred sixty beaumond
+potatoes in dressing rooms are made of white and black
+don't open the mouth of all this is in a hurry
+the bird does not sing from the ground it is not a smaller place
+i accept immediately to the amendment number one hundred and fifty five
+ten burnau street seventy one thousand six hundred seventy sen pierre de varence
+one hundred eighty one tiller street sixty three hundred twenty sade seer
+fontaine de pono street confulland
+and if he were going to the fire
+the floor is now open to mister germinel pero to support the one thousand thirty second amendment
+here we are still quiet and i can leave for one
+the floor is now open to missus sabine bes to support the fifty second amendment
+ah bruin without saying it if i'd want to have the courage to burn
+i am telling you again the trench of forty five per cent is not the only concern
+fifty four p blank street eighteen two hundred thirty in saint dilchard
+without a doubt the trend will be on the contrary to th amendments to analyze the savings
+yes is she easy for the employer no
+his house of chadelonne was taken and looted
+he has the last one you were absent during the marriage of mister his brother
+twenty four bouclet de bor postal code fifty seven one hundred tinville
+missus godan elier my child is reading
+we go out of it
+i think about the other hen owners of the signal now in great difficulties
+this depends on the minister
+of course there is a magnificent contradiction
+four murlett street twenty four six hundred and sixty in notre dame de seninac
+marinier's alley number thirty five
+all children were playing with the same desire
+inab street you had in paris
+it was actually as a partner between us and the other hand that was not very good
+however i did not want to use it
+there is a man who is here from bordeaux who is a man who is in charge of this
+bouet de la pereplace in letor
+rease osli street at number thirteen
+my friend if i
+lady if that's the beginning of love oh i'm the one who's me
+the floor is now open to mister francois esther to support the one hundred thirty eighth amendment
+marcel pagnell street at number twenty two
+what is the government's opinion for the same reasons
+commenda paraso avenue at number eight
+ah here's an idea
+it's very good
+and it's a pollution that has direct consequences on the health
+seventeen sak fiel road thirty two six hundred lile gordane
+one hundred thirty six de la pierre de fau pathway thirteen zero eighty in x n revence
+i have issued them in the press but i claimed
+thirty four women are also in the middle of the game
+etien nordy boulevard thirty seven three hundred twenty three
+seeing edgar at the back and the front are not very well
+the floor is now open to mister andrea chesan to support the eight hundred seventy fourth amendment
+what initials ar a
+de la fosce bernard street at number sixty seven
+we listened to the demonstration of her in particular
+two hundred sixty six p r i street
+it's mister careless dval travers who is asking for it
+sixty lane ion
+we have discussed it in case and the constitutional companies in two rounds
+but his joyful when everything was cytrocist
+this amendment is also an amendment of precision
+i perfectly recognize myself in the statements of pierre lanouche and francois paponi
+the floor is now open to mister andrea chesen to support the three hundred eighty ninth amendment
+finally he is the lost point he may have some semations of the science
+this terrible moment she said this death in the night with the knife full of us
+what is the commission's opinion favorable opinion
+certainly we will remove since we undergo a pair of revenues
+the chanet engaged in happiness to make her a
+this poem was analyzed in the history of literary history of france
+it is a good policy
+what is the government's opinion same opinion and favorable
+it is not a plate in the sorrow
+amoti flacort sephons
+monpaser avenue beaumont to paragore
+and hear that and her that's what she asked this poor lady
+all of a sudden missus joserin parrelled her maid
+in which the police had los tomeld his business because of engely certaily
+four hours for five minutes
+oh dear virgine i have all the sen in me bernard
+it must be an artist a employe
+one pal bona street seventy three one hundred x les baines
+bal profond pathway in tonners
+none of those are the ones who are not the only ones in the world
+what the hell are you planning to do with this amendment
+he kisses again
+in perlo modernist man is factored in china is composed of the oceanics
+seventeen belles maizon street thirty seven two hundred seventy in larqe
+fourteen t lier square thirty one eight hundred twenty pibre
+ah you father had needs to buy all these furniture in vain
+at most we'll be wed at the most
+without attacking mister zola is not her soul
+pa brame cooled the sack in notre dame de cinegli
+we come to the amendments number seven hundred and twelve
+who asked you that
+it's not it's not it's probably the young man
+but the father is strong to stand up and look at them with a depressing look
+tarran paragord street twenty four thousand parago
+my god i have nothing to do with you
+fifty two vigney street
+what did you care that you are crazy
+bend you for calling her back
+the baroness started me
+here lok this you will tell me if you know me
+while talking don't forget to throw him from time to time a little
+he feared he had matilda pante than vanity
+from now on i will not go with them
+mister hugo burnlizis has the floor to support amendment number two hundred sixty
+gaston metois street seventy nine thousand three hundred and forty menegut
+and he stayed in a hundred men to go out always going to drag themselves together
+it's alrealy done
+the floor is now open to mister giles lernen to support the forty two rectified amendment
+the lady from venice through paris
+twenty seven virginal boulevard eighty seven zero zero in the majes
+stiney a young farmer matan albert
+lucian gilla street ninety three thousand eight hundred epanas
+it's indeed a matter of extremely important
+the data is more complete that this time is now at dawn
+there is no doubt about this matter
+i'm taking off my hat and i'm going back to it
+it was the headquarters of the resistance
+sophie ermadebri
+the defender of laws is very careful
+darfoyle pathway in sad emand mantorin
+are you withdrawing the voice
+the compensation must be compulsory to take place on the territory
+however fantine was looking at the sky of her bed
+men played on the stock exchange
+well like my young man cried out the n man
+it would be a transparence of procedure to allow the citizens information in the modern language
+we are in a denial of marcy
+jennett street at number twenty eight
+the chief royalist knocked the chief royalist knock on the ground and closes the points with anger
+officers in support them were bigging them on their knees
+you were taking over a model that is still a former one
+the republic of the united states continued with success during the war against england
+ah that's such a guy
+missus godan stood up on her side and left me alone
+this was the last resolution that he took the garden after this pronounced two large hours
+a lawyer of the force the force escapes you
+leon chitnet street twenty three thousand gire
+i guess the countess said
+the chana shouted loudly screaming
+he died and i left them in a post of grape
+he was afraid of me
+the grin took the masses the real one gressed the individual
+jan what can you count im
+i will be far from you for sure free to the monsalis and beyond
+twenty one pole l fag avenue thirteen thousand seven hundred and sixty sa canat
+where are you put or the pass by
+gardias path a two thousand eight hundred negerles
+com de mart in yek street in sate part in isaac
+pray street fifty three thousand seven hundred and fifty four in corps and valet
+i'm asking you madame vernett to defend the amendment number three hundred and fifty five
+he does understand anything but he knows because he suffers
+do you think i can be happy with you
+i told you i had surprised the king in your house
+no my friend i did my time and i won't no my friend i did my time and i won't take back the services
+researchers you house interrupted a voice
+three village of la fontes eighty thousand two hundred eighty in the same way
+so you're taking me out
+the floor is now open to mister dominiqe tin to support the five hundred sixty seventh amendment
+there is one in the missus sasnesalon
+heap it on i am still listening to that sa caltos defeats it and take the king
+we must have a national view dear callits
+fourteen de cara street seventy two thousand four hundred eigten jempes
+i will keep ickout on the whole way through dscession
+de celestin boulevard amien
+the sister didn't the sister didn't want to talk to us about it
+our fellow citizens are more sensitive to this topic
+at carbonel his long paddled haired birds
+du marqi rode at number twenty seven
+two hundred ninety three dile strate thirty nine thousand five hundred seventy perine
+he says repeatit and swear with perseverance he was only one hundred and thirty lewis
+he goes out with the horse and the horse and the king of the castle
+but they will be poured at the public dates
+deserve that the government's did end of the heart
+i've done a bad full ill
+with pleasure to take no
+ho and your brother
+why would i dare it is here
+i am starting to pick up again
+there are people that have predicicit names
+the c i c of compensates rather than concerns and fairly
+there with his ideas he will be
+me henry something
+these are completely drunk
+mister derako sara do you maintain your amendment yes madam president
+we are prolonging this finge until the end of the text exam
+on the eighteenth of december began the death penalty
+this is the subject of my submendment
+my god you are so alive for men
+he mades a layer for every and waited
+boryou mister de courson who is yours
+courageous young girl strung as a boy named henry to his end
+don't fear anything for my health and my health
+i want to hear them sing and make them dance
+very well for the others its a individual choice like for any other xnation
+mister laurent rucker has the floor to support amendment one hundred and fifty seven
+twenty bis on re martin street fifty one two hundred twenty livre
+but mister de pomer already had seen himself between belle rose and mister de charner
+abbeys tengy street pontevence
+yes but it is a discaphimis dis
+i'm telling you coming home because at home
+his very discussions are not really discussable
+my dear collick my grand challon is not a great gentleman
+equals co equals co frence do equals paris
+you write to do something but i can't do
+the party was well ill and the one who was sick
+thirty two white houses were built
+what is the inside
+it is even here again a community obligation
+good evening everyone will be able to find out the public
+october had crossed the church and the church
+it's more recent
+oak flowers said markov as he saw his two
+it seems to have been ten o'clock
+saint urian street convent
+the great gentleman was waiting for you
+the floor is now open to mister favarini deficient to support the one hundredth amendment
+at the right time he was a poet
+the goods are you awakening me up
+meeting sa qualtes buffegrace ine
+know that the king of france is free in fact free
+he doesn't play any but also doesn't play
+palairme alley forty nine four hundred and sixty intre palerme alley forty nine four hundred and sixty intrejuner
+he will have spent some money and the money with the help of the people
+she's unfortunate that he can't be happy later
+the mercele thence it was my vacuum clena
+i have to get there to the right time i'll be able to reach it
+the general divisional sector was a horseman killed during the first division
+to morrow regret to have caused you teeth in him
+it was the procession that finally made its entry into the church
+idl maybe way for already
+seven martin street zero nine one hundred fifty in saint martin
+eleven girodel street eighty five six hundred ten in cugand
+the floor is now open to mister francis for kemmer registered on the arc
+what are you doing
+so she confessed herself
+albert at this place had to be borg and his eyes full of anger
+i would like to know the government's position on this question
+that is how it will get the government's position i will go to the opinion is very reserved
+sixteen kimende baouse eighty three thousand two hundred ten
+he was shrinking with thirsty
+outside of the hotel
+the geror hunters were talking the man ibas my boy
+ah but it ah but it's the only one who's going to kiss him
+eighteen rudu carpis seventy one thousand two hundred sixty saint abin
+hodvan street fifty four one hundred thirteen bolone
+the city is thus the city is thus in favour of both endments
+wade petite kemin street seventeen thousand three hundred ninety la tramblade
+exactly a man is stronger he will be able to give you a kiss
+the horses stopped in sensile
+he does nothing to play in kate since he came back
+he's less surprised that you don't he's less surprised that you don't see
+you were right to say olympe to consult it looks like a cemetery
+the spanish therefore has a lot of money in france
+morning it's no time to go there to night
+the floor is now open to mister jolger to support the forty fifth amendment
+i asked to give me the pleasure before my business
+how far is it from the south
+few wrote thirty seven one hundred and fifty a peny lesbor
+but indeed yes all of this is exposing the hubble on the nervous system
+seventy seven fulls lebois eighty eight zero zero zero in the majes
+to day it is sunday
+oldstone's alley ninety five thousand four hundred and twenty in oden
+delighted by singing the son
+it is totally false
+what as here is it
+it's my last wishes it's my last wishes it's my only desire don't it's my last wishes it's my only desire don't refuse the prayer of morin
+what said makes it something
+eleven cuscott street ninety seven three hundred fifty one in matori
+lermatash rise at number six bis
+mister is not the subject matter
+i can see you i'm going to think of him for a monoline
+we do not know each other
+the latter belongs to a terapeutic class in which he's care of heavy pathology
+i will not get back to the bottom of the measure put on points to clarify
+eleven de dots de savoy avenue seventy three thousand chambery
+i'm sure that we like it
+without mentioning this new bologni boulevard ninety seven thousand three hundred ten in kuru
+the amendment number four hundred and twelve is mister caver defended equally
+dusorber road champanysermur
+ladies do you want to accompany you on the other hand
+let's be gentlemen to day
+general gollard street in ball fort
+po narway at number three hundred nd seventy three
+to always have this pride said gion in front of me
+it is a serious and delicate question
+the ors forced to withdraw will not make him take part in the field
+dohay street at number twenty three
+terror terro bethered or houses
+she enters on the last with her two bottles
+mon plus had a very long time
+everything succeeded one another to his wishes
+bosele alley in furry sermice
+twelve on aradas residence sixty five zero zero zero in tarts
+one last time there is no brazilian
+prestila city is named after the city
+v j street in bagnell sir selly
+i say general that your threat is not faltering me
+the aid of fire fighters was not very popular
+sixteen corvegan street forty four zero zero zero in nod
+edward julian street eighty one zero zero zero alby
+you have opened the door for a while
+croherment boulevard thirty six zero zero zero chateru
+the second subject on which i would like to give sumerifications is the figure of the arstrio
+examining what happened after
+ie never been sick
+the closest star of the sun is at two hundred fifty thousand times
+here is mister the earl madam
+first no number of incrip first no number of incrips justify this suddenly change
+he joins him if you think it he joins him if you think it easy
+our two little ones were dressed like the country's children
+this amendment tends to establish a balance where everyone justifficial
+to george villiers street seventy seven five hundred eighty gerrard
+it is a beautiful thing
+are you withdrawing the amendment missus griff no i maintain it
+what you saying is that you're not going to watch it
+what do you think about it i have to do with you
+cosette's a band for me and the other one is a poet
+saliabel famin used for the most part in the time
+ah you would say it time to me to go to the right
+this throw as tall like a darker and a big
+its knockler the floor's now open to missus christine peers bug
+camille de os at number one thousand one hundred eighty four
+but the positive law already recognizes this power
+you have lunch with me
+in eight of four hundred and six
+the floor is open to mister damien abud to support amendment number two hundred eighty four
+one rue de la sino billet two three triple zero montana
+mister try benoi may now speak in support of amendment numbe twenty one
+ble bot street seventy nine four hundred thirty le chapelle
+missus godan very well
+dear marquise you are still very agitating
+you have a drink at the bottom of the meal
+the the they agree to get married
+this way i assure you these embarrassments
+everything is very clear but with complex explanations like our ball for example
+rays were rested they had to solve the orders
+twenty i obtain mucly
+he doesn't he doesn't he doesn't believe in science kilaban has a whole time
+fourteen vichpon street sixty two six hundred thirty
+well for the exclamation give me the hand vigorously old point
+call me please
+you were my work that opera is your work
+edward fortuet street in ridemac
+the du belles made him invite d melinton again
+i will take a soup to my lord and i will be a soup
+i would like to then all the intervenons to the people who are toning about
+eleven charles nogaser street thirty three one hundred and fifteen in lateste de buc
+there go to paris nowadays each binoi
+we should be explaining to me the government's policy
+he killed this is why i stumbled them
+wait a minute i would like to view on a cold
+the instruction at home at concerns more than seven thousand children working well
+mister chirol do you do the same in fact i withdraw amendment number one hundred thirty six
+ah sedilian woman said to have a nice hand
+the floor is now open to missus laurd de la rogiere to support the one hundred fifty second amendment
+a little bit too much sober
+jose did not tell you that a fear of delay
+this is why it is not possible to replace two per cent for five per cent
+but i am afraid that we are not having much knowledge
+but nearly three years later we are to jarna and well
+this gives the lens of course
+savage moy verlittle cough
+what is the government's opinion same opinion as perfectly exhibited and even favorable opinion
+we must say that we have to hurry up
+to rue de la nuel fifty six thousand six hundred sixty orvo
+in our remarks that i am having kneeled
+on the right first shot a partner overhang on the right first shot a partner overhang with a large ice cream
+she told him that she was getting dressed in a hurry
+delipable road in vir mustos
+the ortes eighty five three hundred frauds envalet
+who is coming in the company is a great number of companies
+we can buy the weapon and omer of families
+the government is in favor of this amendment
+the most beautiful one in our girls the purest must go to holofen
+we nevertheless wanted to bring a clarification which has a political sense
+he was a divorce living in the middle of the world
+the commission's opinion is't favorable to the amendment number one hundred and fifty two
+the floor is now open to mister patrick ketzell to support the one hundred sixty third amendment
+annette tried the battle
+he only revealed to jump by the whole road's mine
+foxvies is reluctant
+bould be shon at the seventeenth measure of the sun
+all the crew was silenced
+the floor is now open to mister dominik t n to support the one hundred seventy eighth amendment
+de la fosce de brevy road in chelon in champagne
+john admer avenue forty six thousand three hundred borden
+seventy four raymond rivier street twenty four zero zero zero in cahors
+would you be on the traces of our son
+the indians looked at you at the corner of a mocker and threatening look
+one day he thought he was going to do a great thing the girl's instructions
+so we must give them to them less and more respectfully
+it's her
+porse cloud avenue seventy five zero sixteen paris
+they were well punished
+he was paralical and took him into his bed
+souveri noted his head without answer ring
+forty five nache street fifty seven thousand one hundred and forty
+mister lionel tardy has the floor to support amendment number seven hundred twenty seven
+comedy she murmured with a smile
+terra he is warm and charming
+madame is coming to paris to see her coming in paris
+you were on the fifteenth of the week
+the poor girl often repeated him from the vernet did not love him
+torent apartment cuts me out to the romans
+at andys place it a great deal
+ro du coup forty one thousand and blois
+eight hundred nineteen charles street
+he is taking a file
+i was forgotten a moment he is true this man dazzled
+one who wants to listen to him isn't the one who wants to be able to escape
+babache you wantit
+jerome chart here reminded us that we are all in favour of it
+nineteenth december the perisaside
+that that's what i came to talk with you seriously
+alexandrn farexus of marseilles
+ten passage to shently forty nine zero zero zero angers
+two langmar road twenty seven three hundred seventy in foulkville
+if this project is implemented we will need to provide a resource
+fans road laterbl
+we cannot find that there
+have you afraid of you
+the amendment seems to be part of the wisdom
+our vanguard of monternstill that day bis the preparation of the austrian army
+ah what we need to warn these ladies
+this time it was a real battle rubby and soft
+but may be actions very noble
+a hope in snawing his heart
+i waited you sir and i was busy of you
+babche twelve kendles at the same time
+i was very long mister president i would more briefly
+i would like everyone to understand what this is about and that it is a pout
+don igmicio prior of benedictine
+and the big french groups contribute to the competitive french
+his son cheri is glorious charles nine erased
+we will therefore reject these amendments
+they have to work
+missus sanes and she told me about her visit to day
+eleventh of december the guwa
+personally i am in favour of it
+the floor is now open to mister john pierre dor to support the identical amendment number one hundred seventy five
+sa queltis still an ironical
+de tornel street at number three
+why do why do why don't you spend the room with the pen
+seventh bulletin of the great army
+moreover article l
+i called you to talk to you about it
+the bonuses of this terror were ver strong
+the priest went unto the little boys who were not very well known
+it was indeed my amendment to be withdrawn by the
+yet he does not have a required computer tools
+eleven fontaine de larti alley sixty four thousand one hundred ben
+twenty four on d shelan's lot
+seasal savash street in sa etienn
+three gimerit street fifty one thousand one hundredand twenty three
+and make you regrets and continue your cartridges
+he was also miss orel's lover
+here's madam i saw her
+at the right time it was a good time
+i want to bring some clarifications to the general duberg
+josephine asked the government to withdraw this amendment
+gartemp road nu two hundred twenty seven
+whyd whyde why don't you go to he asked where was
+grelvaria alley in saint german de bor
+i handed them back sir but i would have made him onoxious i would have died after paid off
+marie de laoir street forty five thousand orleans
+these are the admirable paintings of colors movement and life
+the first and foremost of his superiority
+forge in serventon street forty two zero zero zero sane etienne
+papiners pathway gessin in saint martin
+john marie pumark street ninety four thousand two hundredd ninety
+these venalities preparatory reasoning vaguely to her ears
+it is indeed a good argument
+never before thinking about nedlands that such animals exist
+there is no chairs and a lower back
+nineteen la croi de roulat forty three thousand six hundred saint x
+that he's stupid is it going to be a good thing
+the convention was reduced to the siege of the city
+the sensitive what is that
+nothing he said while getting angry
+with the help of god she will accomplish all the whole
+three kenta live
+odmuf has five mille lander ten thousand one hundred ninety six
+one hundred nineteen menton road forty thousand three hundred eighty enonard
+this extension will produce according to us the inverse effect of this research
+there is however a bracket madam the reporter general
+for three years our legislative work has been done these fruits
+when you have received this letter a two of us division will have already been accepted this army
+you know madame de marmoy you know madame de martmosy he asked
+jobet road seventy four seven hundred corbe
+what is the government's opinion same opinion as the same opinion
+what is the commission's opinion favorable to this decision
+since twelve years it's like it's like it's like that it's like that
+estimal street in the pamaria pemeris
+this worth one as the call stefano petper
+in fact the authorizations still contribute to the widening of existing operations
+i will not get back to it when the previous statements of la pater
+come on look at you there's your son
+sixth laton honore street fifty eight thousand three hundred sixty saint honore lesbanes
+twenty nine general butat street ninety one thousand two hundred in saint jean
+fifty five on demuthous street fifty seven two hundred thirty bar riento
+i would like to thank the government for its proposal which enables to move forward
+sounds were singing mysterious noises in the tower was still keeping the musical sheet
+general rosallan street zero five one hundred in briumkin
+white debourbon street eighty thousand amen
+once again we wish to know the situation ver exist
+martyrs of the twenty four thousand nine hundred forty four bis saint just
+servenly a mother
+you will wake me up when i will be done
+i'm an honest woman
+champio street twenty one thousand one hundred and twenty one d
+he has forty seven years old
+five lucian raestreed ninety five thousand three hundred and forty person
+it's not the opinion of happiness that is not the only one who is yours
+here goket my word it is blue
+i want to tell you thank you for two titles
+we will cause a case to one of the windows of the small rocket
+who knows what a norne thorin doc
+rocho street ninety seven thousand one hundred in bastere
+my father did not buy the newspaper
+annette knew how much that was happening in the house in paris
+alexandra talked about it you do not have any problem
+forty bis poit essenlot ninety seven two hundred thirty one
+one pierrle street forty one zero zero zero in montde marson
+even tho captain castele now was still a fighter
+a negotiation is coming to finish
+where the company is the company and the company is located in the south of the country
+his daughter came from her and went to see her
+the result are announced in the national assembly's enclosure
+when are we leaving for ferod
+my father went on wil pekling and pitifully i will have more strength
+she disappears on the right
+but there is a med but there is a med but there is a med it to drink it faster
+giving him the occupation from the occupation hero of the public order
+that this woman is romaneste for her good weight
+the floor is now open to missus baranger poletti to support the thirty ninth amendment
+on ren fornier alley fifty five thousand three hundred se mahio
+a train in welts engivin
+that way he has a surprise particularly double part this dear friend
+luckhead street iniciapel and juge
+on the left of the sofa a small side table
+the floor goes out of his room surrounded by bullets
+and bell's letter and bell's letter volume six page four hundred fourteen we will seta an example
+morey however lived in angishness
+two taxations are also provided for the income tax
+it doesn't matter to me that i'm not a question of principles
+crossing orto alle zero four seven hundred ore racin
+for pre to heyroad eighty eight one hundred and twenty in sapor
+oh it's so true we must in all my drawers
+he taked fom money from his pocket and left
+this amendment would lead to get out of this procedure
+the commission retained the removal of article thirty six bisk
+dson street sixty five one hundred and twenty lusete over
+ah he shouted they're finely
+twenty eight foley street thirty eight thousand six hundred ninety bevens
+chane cranier street ninety eight hundred beviliers
+one thousand seven hundred fifty five his business in the wind of lacking everything
+the poor animal is in my research
+the floor is now open to missus katherine votren to support the ninety second amendment
+the floor is now open to mister surge lechamy to support the one hundred thirty third amendment
+i'm waiting for a long time living with the greatest patience
+shes calls gaffran to the left and the king of the city
+barbara for twenty years had been endured all sorts of wealth
+the marquis had a look
+vulgar pustried solier tokes
+i know where his majesty came from
+its dupal branch is located in the south of the city
+a stable boy who was stinning the postilion was sitting on the couch of the quarry
+the floor is now open to missus marie gavne to support the one hundred sixty fourth amendment
+february vacation east on summer were vacant
+at the nigberstick he introduced her to his number
+john gerard bork street in borges surmer
+with what you were you diversing from you i was looking for it
+let's see if you don't want to wrint a letter
+do we need to keep going back the issue is asked for the global manor society
+pierre thomas avenue at number eleven
+i watch them in the lights
+etien now was at dowen
+if there is no time it's not good
+it is not legality
+and what did i see i'd asked you
+thepamentary practices are absolutely necessary
+seven rue de la capot
+i'l have twenty years in a month
+but where is the father yubu
+finally since the fifteen days into persecution
+it is moreover a half english dot full of shakespeare
+alas yes we are in the hotel garno
+jude rode in canteville
+you you do as honest women he said to you will go onominous
+the floor is now open to mister dominic tn to support the two hundred fifty seventh amendment
+there look at me my tenuma and i'm not asking you to come
+first steps were coming in front of this stranger unsubbled
+in addition this amendment is a consultant
+mister buvin it is not a matter of funding the studies by the municipalities
+the will of the government is to make you follow you
+deliner alley ubber
+john pierre veret street seventy three zero zero chambery
+six who got married others it's not a fire
+we are going up here some one is coming to see me the governor
+thirteen dilariscatel street ninety one four hundred ten in ordens
+frankly they were so beautiful that i did so yesterday
+sodor street thirty one hundred n thirty two casards
+driving in the drunkenness can be dangerous for you but also for others
+that's what we will see said the captain
+three tables
+alone mc no was hunted on the field
+and when obstacles the entrals are in the middle of the house
+the amendments that are trying to also ask for report
+since i am from your country i am from rambelas
+indicated in april nineteen hundred eighty five on thousand one hundred eighty five
+no it is only an interruption
+i suggest the withdrawal of these amendments to the amendment number one hundred and fifty two
+thus general amputation of the two legs to the under side the knee
+the rich having taken the check out was still the coward that the coward was dead
+disavell street barban lisbains
+he feared the impatience of the people who were talkingt
+thirty three faubourg dupont de juvens zero eight
+i would like to go home at home
+eleven la decel forty three one hundred twenty ministrol serleir
+renad llastret in genvilliers
+hi there it is no way to get used to the public
+de blese seventy three two hundred albert
+nine upper avenue zero four eight hundred sixty pierverd
+returning to his desk
+hureny richard seventy seven thousand five hundred ninety fontaine laporte
+ferdinand perier street in sa priest in sa priest in se riest
+ah sir here's a beautiful girl with a beautiful girl
+saint bade street thirty four thousand six hundred ninety february
+what you want to do with the french people's news
+patton street at number seven hundred and fifty four
+the color is good without being beautiful
+and you see me
+the floor is now open to mister petree's carblew to support the thirty sixth amendment
+doctor de jorto street eighty seven zo zero zero majes
+do you know who is here
+and when are you going to your father
+rue du maju eighty nine thousand two hundred ten billy
+de lilian road twenty six seven hundred ninety
+far away through paris srounded
+it is maybe the most important moment of our parlmaturdate
+plut street forty three thousand seven hundred saint german librate
+cam chaka rod in saint clod
+the current amendment is then made by the u s a
+du trotton road eighty four thousand two hundred fifty lithor
+i really have enough money to buy a car
+what we say i what we say i sleep on the last time she added esperance
+second and third act in paris
+i understand the feeling of humanist feeling that led you to write these amendments
+eighteen de la bas moti boulevard fifty six three hundred forty karni
+the buttom opens on a garden
+theend of te nose of the dog is rowing
+other races are less exposed
+the floor is now open to missus yale bron pivot to support the one hundred forty fifth amendment
+the compensation is indeed going to become a financial cost that we can buy
+in tennis the party continues to be a new one
+la boiso lot twelve three hundred and fifty linejes
+but look at that ten
+planned a lorg way in saint mark gemarde
+does it make a change
+the measure is by consequence to arm therefore to evaluate
+what is the hell of the world
+mister dominic tn has the floor to support amendment number three
+in great britain the venetian river
+laparier street twenty seven thousand nine hundred and forty leal sero
+the general outburst was recalled to this matter for a certain number of compensations
+while he is absent his wife died
+thirty seven tilly street ninety two seven hundred in commons
+is it for madame de blainville
+two hundred seventy four apass road forty two hundred fifty in
+be careful not to reassure the people who are right
+why for some time those weaknesses are about that
+the floor is now open to mister julian aubert to support the five hundred eighth amendment
+how many times will you stay with us
+boys here to see his children where his little children become a smoker
+nobody closed the doors here
+fifty three darmoten street eighty six thousand three hundred chiveny
+our hero is a member of the impatient but his recognition was not eroded
+the floor is now open to mister dominic tn to support the one hundred fifth amendment
+the floor is now open to mister patrick ketzell to support the one hundred fifty seventh amendment
+sate du pinson at number six
+the great countess is a revenue more richer than the better than you
+gilcourt street orleans
+externo at astronomer
+he repeated in spite of his imponation in call
+give me a setelio these future actresses who are in love with him
+as such you are not going to be thank you sir
+but this is not the subject of the date
+well the colonel sees me walked around and the doctor said
+twenty one rue de bergemats sixty six three hundred pantea
+fourteen brogl pathway thirty four four hundred ninety in merville lespezes
+i like my country to be a very good country
+no you boots you could have followed before my entry
+it's a beautiful girl cup that is a great frenchwoman
+guilo one thousand eight hundred thirty three in eight of three hundred sixty five
+i'm suring i'm lifting the ideas of the ideas of the general council
+morel futty dead end seventy six zero zero in rouan
+this proposal seems interesting to me
+louis cadore street at number eight
+the case of the shusse took the position
+he would have recognized the look of the generic as i have recognized his own
+the floor is open to madame minister to support amendment number seven hundred ninety one
+claudburgo street serene
+fifty four pierre de bordo street twenty four thousand one hundred ten brenta
+twenty one denot street sixty two one hundred thirteen sellat boars
+the floor is now open to mister arnold richard to support the one thousand one hundred eighty eighth amendment
+four borsed avenue seventy seven thousand four hundred lenesermarn
+but you resistance are already in you
+the floor is open to missus virgin duby miller to support amendment number nine
+he had felt all a true vocation by father
+mister crillen has often said to me in my presence
+ruinless embrans to peri in bologne billincord
+there is a place for all the uptions in the middle of the game
+how much is he
+my dear mister rastignac left a young man unable to ruin his master
+of course it is not a triumph in mavesina
+the crisis hit them hard
+mister has the communication
+three pierres nar street twenty two four hundred thirty herke
+the students having opened the bag fly to get rid of excommunicated
+twenty one port penny street
+he would have returned next week
+it is therefore that
+he had during his last words a safe savoy
+it seems to have lost its own device
+twelfe polse fort street ninety seven two hundred thirty one in abert
+furthermore she will guarantee a best payment for the dailies
+champ champ nova street in shelley
+are you sure that we would like to have some for a while
+at the moment she laughed when she was talking to him about low
+i want to know how you are in favour of mister secretary of state
+modila zero nine seven hundred satwar
+seven genetine alle twenty three two hundred fifty ponterid
+i would like to eat the restaurant of the fith street
+wat mare is a small problem because the point is left in parliament
+the floor is now open to mister gill mrive to support the one hundred eighty first amendment
+all parties were bred
+look shouted markov with anger the despair in my little soul's here
+madam minister i would like to call your attention on the medical file
+his adoption would be contrary to the employee's interests and it werin
+i just gave her to the door you julet and i've just opened it to the door
+that is to say leaving through your house while he was entering that is to say leaving through your house while he was entering theeger street
+yes yes covers the kiss dis charmand yes yes covers the kiss dis charmand posit
+we were loved
+bem clinton street in sa owner
+twenty three alphonse combe street seventy seven three hundred thirty in asbers la frier
+guy prenounced this fatal project the thought that she was a crime
+mudgal bow street ninety seven three hundred ten corew
+ga street at number seventeen f
+what did you say that woman at all nothing the time he continued as he got up
+forty six bourbon lentre road
+i write to the treasurer of his indecents that were driving with the flemish company
+second domestic servant stot
+you will be free to all of this scruple and i will get your bowls in your good order
+i'm counting on it
+he had been summoned in this fight with other children
+his eyes lived in size with all his length
+here they are these architects
+i am pleased that the proposition is maintained beyond this group
+oh you speak the opposite language but it is not a real question
+missus godan catting out awhile on the other hand is not there
+he's stretching out to go there and get a little bit of work
+at the rise of the curtain the scene remains empty for a moment
+as he was not coming he wasn't coming we didn't stay waiting
+frunch is in charge of animals
+casorn street seventy six thousand one hundred sixty sates ferser indra
+what is the commission's opinion on these endments being in common opinion
+here is wygors fenwick shun paul chantiga and myself have submitted amendments on this subject
+impossible to give me back to my opinion the government is not a question
+don't it be a foolish reflection on the other hand
+his goods were eating too much less than his father's sake
+creme road eighty six thousand two hundred and seventy la roche pos
+the commission is by consequence in favour of this amendment
+the great lord of foreigners mister the duke of chips
+you are a good boy i will not forget you somewhere
+the director a man who is very nice was not singing air
+when we have fixed rates still a time otherwise we meet in technical observation
+you ware sure that she liked it
+you need something else
+you know the rest madam
+he seemed ensurprised to me
+de belline or ricleser lois
+oh my friend micheline duria look at me
+the emperor wanted it to be as well and no one is now against the other
+pennsylvania dead end at number fourteen
+too he seemed so he had above him admiral and two vices
+she would have appeared unnecessy she would have appeared unnecessary salving in all her furious mouth
+george rebordy street sapre
+they are ranging three forty one thousand three hundred selbres
+it is still a question of it
+i like it
+don't forget anything i told you
+donfer le farte fressy street
+the first one that will be able to put the mercy
+albert cristo for boulevard at number twenty one
+he calls me often to lend him a strong hand
+a living room had a room for ball at the back doors
+the floor is now open to mister arno richard to support the one thousand one hundred eighty fifth amendment
+good lord you good lord you good lord you having a nice issue my good lord
+very closely in modiv remarks
+even though he was even an epigraph no resembles in the doctrines
+eight of bellory street seventeen three hundred ninety la trembla
+the floor is now open to mister denosnery to support the forty two
+mayon had promised to bring her to paris
+unfortunately my daughter will marry me she cried and thinks out herself
+three granchero street forty eight zero zero zero in baden
+i don't see anything that i can see any more
+albert gatches boulevard eighty one two hundred ossilian
+two thousand six hundred forty six kemin de cusnir ninety seven
+however the law was under way prevents recommendations on all the devices used on radios
+seventeen de la ruchambea street seventy seven six hundred eighty in royan
+you were in fact satisfied in all the way of the term
+this bothers me because such a writing is a little crazy
+the commission is against this amendment
+how much does it cost you a lot of money
+chantony dead end at number sixteen
+dentis amber street one thousand eight hundred seventy brisermarn
+col street in vite sir mur
+patiche as the memories were content with the rosary
+they're rich in aden and half of them
+me i loved your sister
+de frontie road rastpel in saint martin
+but he's gone
+in my quality of notary these mokels are the contracts
+it's you who curb the opening of mine to which i invite you too
+it's possible but you won't go out
+come on the'se decided we must get it over to day
+i said that the wood actors had according to me a lot of advantages on the others
+the problem is not here
+how do i go with that
+he was far from his provisions as far as the other ones arrived
+nevertheless she will not go far and right now it's right
+what had been done to me the brigand was a great deal
+what the hell my good friend do not be more royalist than the king
+ninth avenue fifty six thousand one hundred seventy coberon
+what is the government's opinion i share the same opinion as the repater
+the statements held by the minister of the budget is really need to conclude dear collics
+indeed fibrian is a real person
+mister dean luke lorent has the floor to support amendment number one
+petite sevenier street seventy six thousand three hundred eighty kantlo
+troneso street ninety five four hundred seventy safeits
+how is your sister a relative of the two sisters
+we have to be vizing in still a reacting measure
+the collection of number four is not known as the french
+delphine the men of six children and the father of the king
+i'm telling you that one's the one who's got it or greek
+the floor is now open to missus sandrin mizitia
+four hundred twenty seven grand pre de malanga seventy four two hundred forty in conic
+ask carrier a kind of valet who seemed to me to be divinely
+he is regrettable that we are keeping the accountof
+i will bring you one day to one day
+eleven made leviser street sixty thousand two hundred twenty budovan
+he turned away the key
+gaunt road eighty five two hundred seventy sant hilaire de res
+the enemy took advantage of all night to go straight
+the marquis here i want to speak
+we are saying that campaign with an intention
+jonatist lieutenant street in manus
+but where is we born i don't know
+gerret street twenty eight four hundred nogent lerotro
+but it's a very serious and very documentit on the police in patagonia
+stay the swiss who had nothing in fact and what made
+we say that some of the countries are beautiful than italy
+ernest cuto avenue fifty nine thousand two hundred thirty saint martin
+this noise eloquent recognition of henry fourd had to warn the poor people
+but how would you voice o but how would you voice don't you worry
+if that is your opinion i am ready to take back my amendment
+i like to watch teve
+he is doing it
+at this stage of our ddates i think this precision is necessary
+i share the feeling of our colleague la chow sentiment
+the gamoir and the bloods have denounced it
+the cry of the eagle rooted again
+coming back to the reality
+de la forte road zero three three hundred forty bass sir a layr
+and he refuses to receive brulish
+the floor is now open to mister gerrard chirpin to support the fourteenth amendment
+twenty six tublot lorraine avenue
+they are very few after the capture of chartries
+it is the window we hear
+and there concerned the improvement of the acquisition of knowledge
+mix pelliot street frech ser moselle
+roger pabrock street fifty nine thousand two hundred torcoin
+five hundred ninety eight one clisson road forty four thousand one hundred twenty bertut
+we know how to secure this article
+limo street at number forty
+the floor is now open to missus nataly upray to support the six hundred forty seventh amendment
+this is the reason why we are asking you to withdraw your amendment
+his hairs are yellow as the one of his bracelets
+salvetier street forty nine three hundred eighty notre dame de lanca
+thirty five de la dum pontac street thirty three thousand one hundred twenty ardichan
+if there is revolution i am sure that i am in the first role
+new frigor even in the corse of the boy
+some one's at the door he gets ot
+on ricneklas boulevard forty two thousand two hundred twenty bourg and ardena
+camille catlin street at number two
+a gunshot will never want a surface
+sixty one rue emile bertier ninety one thousand two hundredad forty saint michel sir orge
+sat water is imprisoned in shans
+it is the correction of anuble mister president
+the mayors want to see their children
+eleven gaston levee street forty six one hundred in fia
+doctor roger sano street at number fourteen
+nine merandon street fifty eight zero zero zero in nevers
+laden road and teps
+but this is done but this is done it's over
+the floor is now open to mister gherard backed to support the one hundred seventy eighth amendment
+the freemasons in general are not very much details on the pursuits of electropies
+to the hucker mister godash is not a poet
+two hundred and seventy five relief wrote zero one six hundredd thirty in saint jean de ginville
+sete at set l street sixteen thousand one hundredand twenty four
+missus godan with a satanic tone
+she wanted to know and she didn't dare question her
+sixty nine terrace street twenty six four hundred forty six zero zero cahores
+we knock we knock on the door
+i look for this portrait too
+this is the easy way the convention with the region
+but us we're not going through this stage it introduced a real good break up
+a terrible blow shook her body his eyes widening with anger's clever
+i receive my letter my dear actor on the battle of rivoli
+the beam pool number two
+eighteen bis geston morel street
+oh there is a long time
+the su the sh stimulates the intensive regulating the sugar in fat metabolism
+nine charles and blue square thirteen thousand one hundred and seventy cavres
+you'e not brought to anyone who knows
+i'm full of trying to let this text by stimulate the tudor
+you you not going to be mocked by animals any more
+the floor is now open to mister mark dowles to support the two hundred eighth amendment
+a bottle of white wine
+a certain amount of time now
+the federation of russia is in the united states
+lugel avenue satesten
+anasoral place and din inferiors
+the case on a hot iron made of iron and a great power
+dila kennetiere street bone table
+it is written the precise contents
+what is the commission's opinion entirely favorable
+they approach the ecclesiasts as they enjure me
+i have a few books about tea
+this seventeenth of art
+that is new
+cumbelan too ah if she wanted to believe me but no
+please mister juffe to repeated bordaco at the inspector
+your way to proceed is acceptable
+pinnard was suffocating a sigh of joy
+one thing again
+i don't know anything
+no i would have enough to get married
+the role of the deputies is not to be proud
+a beautiful villiers street at number twenty
+sixh napoleon troice avenue
+a red one
+i do not doubt my amendment would be aroused
+julian respected me
+the fight takes place on every front
+hullo everyone and it all
+if you allow me masan president i would help myself soon
+he did his career in ulis
+some time later she died and the young man is locked up in the funeral vault
+make wine by my friend landry
+michelin's father is witnessing me
+the objective was to carry on with akel sixty four as the sentence of condemning practices
+it's my daughter jestin repeated with clare
+what creppy man
+do you want to eat something
+we say they had cut him off
+rose who stopped at the full of one's interpolation
+she gradually growned and awesome
+come on to say pluton coming back to go back to the cafe
+to ansien farm deadend fourteen three hundred sixty in truvile sermer
+e lonpier's voice shouting
+amendments number twenty and two hundred seventy one are identical
+the general bernadat division is at laba capital of the carnel
+marius had the pistols in his gussets and his gussets
+desidec it deside it deside it's the wednesday that you chose to be another day
+reverend mother she will put an earth in the beer
+well let's well let's get us out of here
+to justic street fourteen thousand three hundred twenty meser or
+come on dear friend rend
+so i find a way
+the floor is now open to mister mayor a beb
+the cost of the measure does not have any problem the trajectory and the same
+paris augus offre one thousand eight hundred thirty two in aid of sixty three
+the men who will come to night
+we are getting into the state's constitutional technique
+it is the door of the room that i see here
+our daughter had blonde hair
+if he comes or if he doesn't come
+twenty six less elaso gardens twenty eight one hundred sixty
+twelve busquett avenue seventy five zero zero seven in paris
+they saw the system as well as they came down from the fortune
+ah would you like the murmured
+this testinosly turned towards agriculture where he thought it was as much as the people
+the one who were happy
+madame la clay are you in favour of the sub amendment yes
+this modern one was cut off to bear the swords and legs
+thirty five louis d pra street
+it is a kount of business agreement
+we must turn the final by the participate in the final
+it will be yourselves she said to eugene and you'll tell me how he goes
+general claperade
+the brother basil capin had achieved spanish troops for the lundoc
+the emperor will first insure the prosperity of the business to obtain good finances
+nevertheless your amendment presents a certain number of difficulties
+all these misfortunes were accused of the sin and called them to pay attention
+barche you will tell me a story
+and then by this marriage i escaped to chase the tropical widow
+george merkeley street in valdo
+we were getting closer she was dead
+how did you find me in your head
+given that the mother was bad the boats remained in the port
+amendment number eight hundred sixty four of the reporter is atorio
+that is to be known to be the most important ones in the middle of the world
+carso daughters and his daughter are you a girl
+it's over he said the musketeers from to day are not those of his majesty louise
+gastonero street in viltanus
+these mails are prepared there in his castle
+there is paradise and paradise
+oh no one's trougled
+he wasn't he wasn't he wasn't gone for two minutes that arrived
+ah sir you miss me
+often a poor woman was forgetting in the corner of the province
+fifty five point fort road ninety seven two hundred thirty three
+so many people hit them poorly
+yes he was down she had to be there so much that she would be precise
+the freedom of powerlessness is getting rid of it
+gabrouche clawsed on his hips
+me we are wringing my hands
+mister la chapelle is a friend of papa and
+he refers to the canoa on the last
+sixty three b paul dupont avenue ninety three one hundred ninety livrard
+because you stay in the middle of the war you are giving the arguments to the right
+i have already written eight letters
+looking around him
+salcidricht the deliberation of the community council
+all these people are novels
+i understood that the considerables were able to have the effect of reducing the principles
+or to day the vote is a vote for annual vote
+thirteen covi street
+in five individuals the roman estas continued this trip was only a person
+ah would we be sunny this ver well wait for
+it is a major demage satching unfortunately the loss it is a major demage satching unfortunately the loss of humans
+i think that this is what madam reporter has tried to do
+we have to be loved and i don't like anyway
+his daughter came five and fantacizes him in five cast nine der ninety
+please follow me mister de coursy
+the floor is now open to mister julian abert to support the one hundred thirty fourth amendment
+and he came
+you not taking anything
+for the rest he's doing very well to bother you both
+he's cold outside
+one hundred seventy three and the following one defines the pinol senctions in this case
+the prince of parma senala followed by loans and haras the horsemen of the barne
+we just wish to extend this amendment to all the reservations
+the thief was the same eighteen seventy three
+in favour of greeks he is a member of the royalist
+de kelly street carew
+he enthusiastically of the parisians was welcomed for their friends in spain
+carrying the box on the buffet
+for once i will be opposed to your amendment mister tardy
+it asked for something to be more to the subject
+i bought a new sofa
+a whole grim o beautiful recommendation had been done
+i have trouble to follow your reasoning to the point of view
+he loved it until the death of the death of the same year
+gevruch until this moment was distracted by his amusing thing hadn't even seen the same thing
+sant rose line avenue eighty three thousand nine hundred twenty lamoth
+he goes back up above the couch
+so we left without regret
+it's my case to night
+no reorganization no discipline is incupable of all public services
+what did i do
+andrea tbo street at number six
+your vision missus minister is gobl
+from this point of view no one can be against
+mister will not tar sun it
+showed him the salty piece of salt
+rue de la renander in saint quentin
+a fire fighter i a fire fighter it's just that
+and now i don't have you any more
+person dads what it means
+i withdraw this amendment
+very well that's it
+the floor is now open to mister marshall sedier to support the eighteenth amendment
+he is a lot used by the m e
+certainly melvina dominated this possession of the acestic men
+and how are you whenever you want to clean it
+lady i want to see me
+peacheburk street thirteen one hundred sixty in chateau neuvrenard
+nineteen l fertile street eighteen one hundred ninety
+he poured on nature of his lower voice
+his own
+new
+you were sure that she would have liked it
+to very
+follow my god my god i follow my god my god i am beautiful
+alas it alas it's good for mine
+she was there rewards of francs that god could make a move forward
+fifteen sa florentin road ninety one thousand one hundred in sate florentin
+the commissions were commissioned at the date and would have even been validated
+he thoughtit and sentenced with a jealousy to make an empty bottle
+the report of the political forces was the same
+at what school will he come with
+thirty five bis martyrs delie resistance avenue
+the count was complaining about seeing him
+i not at all agree with my colleagues from the opposition
+de la porte de trevou street ninety two one hundred forty clemard
+are you withdrawing your amendment mister gosselind yes
+i hun
+maybe we will have the time to talk about that at noon
+one onde tessaners rode eighty six one hundred forty sat janet dorliers
+the floor is now open to mister surgil cornado to support the fifteenth amendment
+he has to distinguish two things
+two hours had just run at the clock of the sorbon
+the floor is now open to mister sebastine denaja reporter
+bill e s na street thirty five thousand eight hundred dined
+place on the chapel elect
+the measure i suggest that could be the death of a mutualization
+two doors on the right two to the left
+the four winds
+medon road number one hundred and twenty
+martipa those who are writer these boys of the recipe
+arnoch road twenty six thousand six hundred tan lamatach
+de tera rod sixty six three hundred lupid
+you accept the two essay
+quin de chamess in calco
+three duchentrin road zero eight one hundred fifty rainway
+a little bitch saint maurice street in labury
+in truth sir is very good
+as for me who am a boy these one hundreds always crince
+gener generatere never drank
+the floor is now open to mister president of the special commission
+he sails to his writings himself
+this must leave for oni what a ministerian cabinet
+five page one hundred ninety seven
+and you are right my dear duke
+vancastl street berdun
+miss pemilla forgiveness to poor boy who is in love
+but the alcohol that poisons the ones grease the others
+i kan i kan i can't mention this matter without speaking about my reasion the banordy
+we can hear the overload of the morning
+the committee showed him at the right hand
+we can therefore only make one hundred kilometers
+it was the secretary of gersha
+the content of amecoda's work for a long time
+you put her delivery news in the morning
+johnsteinbeck street in fa
+but if by chance another pretended to be present
+addressed offers fifty four bases leo street
+twenty one crases nd street twenty nine thousand one hundred seventy
+in excess of their furs she already saw the emperor andamaged by the prussians
+jean at the state
+he threw like that
+for che im pass in macun
+i had never heard anything like this
+she contributes to my personal pride to vote for the entire budget
+i now want to put article twenty seven to vote
+doctor de la fontaine street ninety three two hundred sateny
+he seems to me that it is preferable to tarnet the municipality than the mayor
+curiously he seems that nobody is talking about organ harvesting in the process
+two edyor place six o'clock
+eighty nine de cap road
+you would have done a lot to see the history of france madam minister
+what is the government opinion favorable mister president
+a moment later we saw a prince with lying and the corpse was taken off from the entrance
+this time
+you see the head of the guy's face he came to touch his legacy
+they left it the same time and bought a port for fifty crowns
+ten small french street twenty one thousand three hundred forty nele
+all the sons pacic then browdry but not have a song to drink
+his father was called tesselin and his mother annette
+what colour is it
+you wouldn't have seen any chance the idea of buying gaftwer
+chasels was visiblely preoccupied
+hallo
+sixteen corells pathway fifty one thousand one hundred and fifty in saint german
+you won't see him but i will go join him he will niece
+how many did they catch like this
+thirty three etien de lay alley ninety three thousand one hundred ninety livery gartigan
+the two thousand and eighteen might make it difficult for the bears
+fifty seven day cremont street
+de la temch street in billems and hay
+this amendment proposes to remove this cancel in any case its legal proceeding
+one max loett street
+come near the mill mister the next speaker isn't come near the mill mister the next speaker isn't that the young man said
+it will be done
+and you didn't and you didn't see anything at the horse and nothing beats
+and then he was minister
+doctor brevot street at number thirty five
+long champ street in gynderon
+you propose that you do the same conclusion of this amendment
+the ceiling was let by the state council
+ricltesa
+prune's road eighty one hundred indevile
+you will still find a letter for me by waiting for me i have a way to faceus
+number twenty eight cotoval lot four three two four zero say just melnond
+sometimes these verces are applied on the living rooms window
+he has the vocation to become the law
+to day the situation has changed and the reform was not down
+he does a few steps in the parakuet and stops
+his grin fault was to be a flirtatious
+the next day freger general
+but to day we are discussing only through an article l
+per maville avenue number three hundred twenty
+i'm guilty
+boys ma'am is it your name
+i'm asking for the session suspension mister president
+i have a little surprise to do
+we hit him
+six gin webber street thirty nine eight hundred in polony
+it is the last amendment i will defend this article
+i know myself who made the war with him for ten years
+me sir i do not tell my life replied with a rather than a young widow
+the woman told him to comein and to be discreet
+shut madame de light is not here
+lowis leer street billins
+if i didn't if i didn't know where we are going nobody would have walked
+and you thought out the danger to which you were on the spot like your daugter
+unfavorable opinion from the commission
+he had one but nothing extraordinarily
+the idea was to fort saint martin
+i i'm really angry
+one hundred and fifty eight george de pech avenue forty seven zero zero zero investes
+clode mariners forty nine thousand six hundred eighty n
+a farmhouse boy note master
+from this point of view this law constituted a real probless
+don't go too fast
+he never owed money to you
+the great difficulty nowadays is analyzed quickly
+paldumer alle forty eight zero zero zero mendi
+do you want us
+we are masters of our lady of laret
+he is widow i don't want to marry the widow
+how is the party i tried to sell the people who wanted
+and pierre it's not too many of them
+where did you go
+the floor is now open to mister francois ruby to support the one hundred eleventh amendment
+she broke her hand in a desperate mean
+twenty on delarmary way thirty seven two hundred seventy in orus
+akia is the secretary of the order to hear his good intentions are good to accept this text
+fors in the united kingdom fale sixty two thousand one hundredand sixty six
+ch d programs that don't have changed much to accompanied
+you were coming to time
+e lucian's voice woke up the echo from the undergrouth
+he setting on his guards by squeaking while looting his head
+you did well your happiness
+it's my opinion on the other hand
+and the most beautiful in the case which made her own
+dressed without research or with a sterility
+and yet if the rague is convoked it is what he thought of your presence is necessary
+italy volume seven page two hundred fourteen
+it's not my fault it's michau who makes a strake
+yes sir since the day before yesterday
+here is this tale either you will open your parenthesis
+thirteen ferdinand gaffrino street
+as it was a literature for the immediately that was the pocket
+delphine that countess was not having a little over her
+delane rode twenty eight two hundred sixty agnet
+some guys roll on ther as they were broken down from the air
+of course the bucadentary prevention is absolutely vital
+larsh street eighty nine thousand four hundred and forty massenges
+he opened a large and fated to let them pass
+the bretons looked at each other and couldt help each other
+he had to work on his refusal in the escorts
+what am i want to see to see you happy
+he will be examined after articles sixty six or in thirty seven and thirty seven
+he heads towards his room then returns suddenly
+finding what i went back and make a death sent there
+toame street forty two zero zero saint atienne
+a hamlet had taken the flaws
+he'll love my mother when she's dead with me
+she ate some fruits for her dinner
+to morrow i will be at the same time
+she was unable to decide if enna had come to see her
+budwin wrote thirty six thousand four hundred lameni
+the country must be the gold fish of breakfast
+he could have them in hand at the time of his fall
+did we bring the treasure the ice cannons to the left
+at that time he was eroded with long runs
+the floor is now open to mister pierre henri dumont
+timse selwyn equals two equals three equals four
+eighteen degre shell street zero eight one hundred seventy in argne
+in its current state however the commission emits an unfavorable opinion
+what can i do for you
+the girastin was returned by plane who finally arrived in time
+second point all the data involved by this article belongs to the state
+about that i have good news to learn
+oh it's too expensive
+thirty seven orks rode seventy six thousand seven hundred in our fleur
+letters that possessed madame minister have been stolen from him
+the man you were marrying you probably to your child
+it is a coordination amendment matam president
+liddy who was very rich was that was delicately coffee
+you had met one of his traders i would have made him his business
+thirteen ductor cabria street twelve thousand five hundred issels
+oh i would have woken mister
+oh you must be beautiful with them
+avenue at number four
+five bruells with massive
+but what did you want to surprise me
+the floor is now open to mister giin paul de france
+the commission is as the commission is as ssured mister minister will be slightly blushed
+if she was born a denge of laughter we sold it well
+the memory of clotine was the memory of clotine was disounding the active
+mante o no selne
+oh time dies nothing to it because when we count
+missus grobois rubeuf jardine purloud
+rips lila sixty three thousand one hundred and seventy edes
+this text ceased to double the government's requests
+the floor is now open to mister john louis briquett to support the one hundred eighty eighth amendment
+twelfe benoi fret twelf benois fretto alley fifty thousand one hundred twenty equaterville at newville
+installed on the flourishing of their retirement home two men would be telling the story of the good old days
+when it arrived
+boodle drive
+by approaching the pariptal you would laugh as soon as they were hanging
+de mora wrote at number two thousand four hundred and twenty three
+max planned a place in crane
+he is however important to sensitize this issue
+ah yes in a street i see the shop
+here you go mister president the few clarifications that seem to have to be given
+lignes rode in catres
+the guter whispered masso looking like abusing
+matilda isarmed her eyes she had seen her letter from
+where and what time does it take
+he knows himself since fifty days in bame
+it was not the only man who was my tears
+he will have a unique trial
+we must be understood why i am holding the reference to a doctor
+rigulus asks his goods here that madame du bois is living in yes sir
+eleven da martin wrote sixty three hundred thirty avres
+the weather is awful the rain had fallen
+he was he was in twenty five leaks that the night was still snowing
+he didn't he didn't feel the situation deteriorate at home
+they had imagined to prohibit us from drinking in service
+he exists three cases of breaking
+you know maybe madam the minister it you know maybe madam the minister its a report
+all three of them settled down
+so you convinced that the flank was flashing
+i'm not feeling it
+the other harso was his exact name a small man with no stradaler
+weare not going to heaven affronts the social issue than the devil
+very well i put to vote amendment number two hundred and fifty nine
+cloted ormers at number six
+he was at the bodinen
+no i answered while getting back to my surprise it's not bill
+a simple kemra to learn the address to the order
+the distribution of things is made of a lot of people
+this is where the dates we had for a few moments
+my father is my father is my father is there in his office
+we can go to nepoch at the riot of venor
+the boy finds the way for the night to touch him in the shaft
+twelve pati no dubo street eighty seven zero zero mages
+lowis who lies with a candlestick
+in reality you organize the accumulation of some of them and your inter buying it to the others
+fifteen ro denuly sixty one thousand four hundred saint hilaire le chatel
+he confessed it himself
+sagilla avenue inville france de ruer
+i had a few sits of water before leaving
+give a kiss for am
+in the meantime he is more niture than himself
+the floor is now open to mister eli a bodd to support the one thousand one hundred sixteenth amendment
+five masher street zero eight three hundred an fifty in noyer pot medges
+the floor is now open to mister sherrard backed to support the thirty eighth amendment
+a exclaimed levid
+five labore restile five labor restile five labor restile was appointed brazil before the discovery of the nuria
+denis tried to calm the game and read it
+the choice of localities embarrassed him
+and then he likes his dependence
+all this is true but it is not everything
+five to milan cross street eleven four hundred in saint dani
+he didn't see me her's de digging
+i think ye will all come
+i'm making a bit the honest man
+weed valleys iner twenty seven thousand and thirty in ardencourt
+james bull soup is the first one who is a poet
+obrack wrote at number ten
+there too the example of transparency must be rules
+ah well you think about it again at that old story
+six hundred a d twenty eight morenzin road
+he had just noticed some sort of giant going down from a mine
+the file server does not answer the next day of eight p m
+perd uncle
+eight mervo street
+the latter word revealed mister dicharna as if he was in his direction
+this was satusterday
+listen he said to me about the minister
+which is you ancestor
+we don't know what we have to know and we know what to ignore
+no he will laugh maybe the minister but he won't go any gift
+i want to rent a car you do not want to rent a car
+the same one thousand five hundred three talked agiration where it is true
+all day on the twenty ninth was used to work
+she's giving these figures because they clean the reality better than the speech is
+sixty twopence to garni shack street
+it would be then necessary to get chased at the same time
+he was left handed over and had a very recognizable expression
+mubys te la cherand street chambin la fort
+the monopoly of wheat the monopoly of wheat the monopoly of wheat can harm less than he joins the authority
+descendants towards her husband
+the advantage was down
+missus burgham seems to fall into account
+national road descend in aleria
+the drafting proposed by the amendment seems better to correspond to the objective of the recharde objective
+the boy was a searchman in charge of the future
+it brings some technical correction to the affected tuxes
+emboss de lare street in bosso
+it would then be all the more state
+i will not accept it but i will never accept it
+you fault i decided to make an appointment with the operatself
+be careful to what do is not come in the house
+damile street in bastia
+bu is going to kill he will be use what can he use
+well i will just have nothing to do with it
+he doesn't he doesn't he doesn't work to day it's not a good thing to do so on the other hand
+certainly if you stay there i will stay there
+he knew perfectly the places of the places that were not very well known
+fifty seven outside demba street
+he exilated a sih by seeing a green sign
+it was a town hall officer and had distinguished himself in our first campaigns
+twenty nine clo de tonelers thirty one thousand one hundred twenty seven
+she comes back to antoine and is a member of the national assembly
+he's not opening the door
+she ate she ate some cats and were well off to take kilos
+oh i wanted to let himself leave by her
+i like to give the impression of refusing the date
+thirteen de bad thirteen de bad road twenty one thousand one hundred ninety corpos
+then they will be informed i will be a brown
+at f i hoson place at number eleven
+they are running
+thirty four elephants
+he is aiming to resist and resist
+this is why she would be more serene to reduce it to half way
+antoine marie colled street ninety four thousand four hundred vitri sercen
+damage your sister
+but no it's a tourist that started making fun of respect
+the old men must know how to retreat when they are doing
+come on give me that then you are ready to go to the right
+if he hadn't come to catch him he would fall
+i'm regretting it replied conseil
+me madam yes the meaning of my surprise is not escaped
+i like english and i don't know how to use the same thing
+it is the reason why we are asking for the removal of article first
+nothing of this is better
+clodo hotwer street twenty five six hundred fifty in lanney
+if id question him i aske him
+buchet's way in ba and bussy
+he said that he had to be hit too much dangerous
+i am at the window
+when will i see the end of the bush and thed
+i am a student at the university of saint andrew
+one hundred seventy one sa clair street zero one four hundred thirty mayett
+i went to paris so well that we would not bring it back to me
+artists who created the roles
+i write books
+duvo street fifty one hundred eighty ino
+albert is now saved she said firmly
+he waits for a moment a minute in the living room quickly before he came in
+the commissiner is very well known as the commission
+in short he took it to bury secretly
+it's not because of your dad's side because of his chest
+copious pathway ins
+amendments numbers seven hundred ninety two and seven hundred ninety three are offended
+he is very difficult to associate people
+his tomb is a very good image of the pusible and the sentence of the extreme mother
+here he's the right time of a good mood
+the floor is now open to mister francois roufhin
+reading manse what the hell is that
+as i told it before i had to recognize its mistakes
+the commission of the government does not have a negative opinion on a stoop
+when did you talk to your mother last
+yet very quickly there is a good thing for the public service
+i will send you a plate in the week
+monte de pagan thirty eight thousand seven hundred thirty cile
+also he pretended to play comedy
+tell me if you do it you like it i'm a little bit so much
+i now i'm not sure i'm not
+the curtains closed behind him
+the baron leray is a member of the national assembly
+we will succeed i hope
+the floor is now open to missus emily bonivard to support the two hundred fifty fifth amendment
+but by honour i will not discover anything
+this was generally cut short and t
+the floor is now open to mister pierre henri du mont to support the one thousand and twelve
+olivia la fever street eight empes
+the floor is now open to missus jenine dubier to support the three hundred ninety seventh amendment
+ten to vine street
+greet street seventy seven thousand two hundredand fifty venat les senmand
+for example the truck was closed with keys and did not find the feet
+what they do to create during the second time they took the prisoners on the outside
+chasenice was his sponter and one of our neighbors
+he will go to fetch a lot
+they are not married to my daughter
+bade badin go up their house at the beginning of the second round of the king
+also miss i'm goin to do the same thing
+pon jubert boulevard point cart
+once my master free will not get married to our hopes
+indeed we will be able to check that
+gabriel said that the adventure hopes no longer worth a less than a warning
+another big chair
+it was no longer the same year
+the floor is now open to mister loik prudam to support the one hundred seventy eighth amendment
+now this female heart didn't care about nervous prevention
+this argument does this argument doesn't matter to me and even not at all
+at our eyes the steps of letting these administrations are still too important
+seventeen de crene street seventy one thousand two hundred twenty four
+come on give me that
+oh he will read my hand
+ah it seems it was a rugulard
+same position as din o signary
+it bears the basket on a chair near the fireplace
+ernest madir street montavin
+she will continue to do it
+the figno then has spanish and italian
+twelfth chemin de furficd
+portonel deadend sat andre de rookland
+no more than the prisoners would have had a heart to eat
+i consider it the arrivals to a satisfactory situation
+hate impassin matory
+he thus he thus completed the amendment number six hundred and sixty three
+here again we have no clerfication
+far from this feeling everything is indifferent i din't like anything in the world
+secondly i understand the reasons developed by mister minister but this apparatus
+here's the mystery
+he met encircle the officers of these two armies and talked to them for a long time
+taking his pocket a small box woodrop
+the poor man has a lot of children the other one doesn't have any
+he had not thought enough
+i trained him furly defended him
+the floor is now open to mister lionel tardy to support the one hundred twenty ninth amendment
+foutel avenue ninety one thousand three hundred ninety more senser orge
+twenty one emile dutier street
+moratial dren avenue luxoy les bans
+in the middle there is a room waiting for me
+i was waiting for him when i saw his brother running
+thirty seven mariee boulevard fourteen thousand seven hundred eighty in leon sermer
+i perceive i perceive the priest of a few minutes
+brek red erlen la sutaran
+she had five ducks and then it had gone
+father fvent will do everything for the beer we will bring it in the ground
+i would have liked to have the owner
+certainly i will talk to him but not here
+you know what happened to him
+felix a nice mother
+what are you studding
+the crowd was listening bordinantly there was a sort of disappointment in cuffles
+jurincon leaves from the back
+the floor is now open to mister olivie veron to support them
+we kept on against each other
+they saw this bell that they held in his hand
+avenue de paris in lanister
+in the house the device would be difficult to be more than eih hundred year old
+what is the commission's opinion on these two amendments enfavorable
+he said the week he said the week it's normal that the parking lots are empty
+de rodeva impasse in chunge
+if i like you were a mayor of vitrie li ble in champagne
+for twenty lois i would like to draw your writings during the time that you want
+i would have liked that it would be done instead
+it's impossible it would be anormative and we refuse it
+twelve charles rucher street forty three zero zero zero le pui invalet
+that's what i thought to me
+but we must be much more specific because we have to go straight away
+may i get tired of you
+i wanted to exactly askhing if he is a professional had been consulted
+how long does it take to get
+there is therefore not
+ten lupi street eleven thousand five hundred ten in ka
+this is why we wish this amendment to be dobted
+ninety one a robert gayard street
+if the kind of exams you certified or don't
+a farundal of the gospel
+he was talking to a low voice but his mother was still a greater than him
+when he will come i'm ready
+very logically we cannot imagine the prescription of the order drafts
+he extended to right and left to the right and the large hins to push the nails
+god i'm so hot my dear
+the film is brown with black eyes
+so you can stay with the briks
+his last week was dying from the sixteen of france at the door of my permanently
+it is an artist and uses rage
+missus melinier sophie a basket under the wood
+rowell linctil street triterville
+dividsons boulevard fontane susbor
+i know what he is and who has it to day
+she goes away the paris fresh
+we wak up the guy from the intimate agency to prepare the time
+five alexis savestrid thirty one six hundred murit
+albert ba street zero five zero zero zero gap
+nineteen dillereunion street zero three one hundred montlecan
+tarte street zero one in rossen tarte street zero one in rossin
+twenty six martin pelle boulevard zero four hundred
+duke qualicy street zero two eight hundred chairs
+sixteen fontenets impasse zero one one hundred ninety blows
+rue de mendors at number nine hundred seventy six
+three hundred thirty one de la blenche rode thirty thousand five hundred saint victor de mar cap
+coradia rode twenty six hundred furo
+to homo de satan two eight one one zero jusse
+one rue du park twenty nine thousand five hundred twenty chateau nouf du fa
+thestables cousiere street saint salpis le garitoi
+five riquette street zero nine one hundred fifty in saint martin
+fourteen gen trades street zero three
+doctor carroll street at number one
+fifty six general quenit boulevard nineteen thousand one hundred brive la garde
+jesma street thirty one hundred thirty bulmas
+one of shock paraget street
+fleck three cities of kresburg twenty two five hundred in pampo
+eigd lair street
+this square onre is albert lur
+epenega and kinne he fears nothing more in the pinnace
+bernard manacer street at number eight bis
+less hotes to bellinor street fifty six five hundred ring bell
+twenty one lucian fromage street seventy six one hundred sixty in darne
+finally what can ye have
+freres le montre street at number one bis
+the kid he pirates where he commands the vedia
+the one who invented the famous candles in the burial
+the floor is now open to the minister to support the amendment
+there will not be anybody the girl who ran towards the interior of the building
+mother's joy is almost a young child
+the floor is now open to mister julian aubert to support the five hundred ninth amendment
+missus godan here are some painters poet a musician
+what would be the point of taking
+i beg you in front of him a child
+the legal regime is applicable to underground quarries result of ancient texts
+i've heard him sing with him
+seventy nine armond gobert street
+musta blay rode in brion fand
+he is a tribute to not seize this question
+seventy three a just in cateli street
+ninety four del faret place zero nine three hundred ten estans
+twenty eight kurdes pathway
+sad adele rode villa frence de erg
+thirty two rose street zero two three hundred twenty lizzie
+six marro road
+fourteen lesto lot zero nine one hundred la tor ducreo
+la catri republic five four two zero zero chemny
+claude duell in conde sir vescer
+five hundred sixty four chateau dufin street thirty eight thousand five hundred thirty la basier
+two hundred forty nine ruferan daarden sixty thousand seven hundred thirty nine
+for martin praise eighty eight four hundred thirty in german bay
+general du bessa street at numberone
+de dife street in se german le feu
+chusson street thirty one thousand two hundred to lose
+dimmon's alley at number fifty four
+enri de vernois street at number fifteen
+serminer street in sa paul de bor
+seventy six reo street seventy five zero zero zero nine paris
+ten lilaway eighty thousand two hundred fifty la bresse
+forty robert de bot street eighty one hundred thirty in arbe
+doberfield street ramsack leba
+eighteen fro julia cury street ninety three one hundred twenty in the corno
+jack kartane street seventy one thousand two hundred fifty clinney
+thirty seven pusheville road eighty one eight hundred pravesons
+du begune road ninety four thousand four hundred saint quentin
+corse de la croi depens in vermont
+sixteen marcel tennor street
+san bilie road die
+sixty nine prategat road
+ninevilla dubriers lot thirty two thousand eight hundred rivdeville
+de la sudwar dead end forty six thousand one hundred ten ferrick
+leal sir five at number eighty six
+amo de saint quentins rode forty three hundred ninety saint andre de san
+two de la fanadier street fifty four hundred thirty bredville sir a
+three bis maricial noel degenitive vo square forty three thousand one hundred twenty
+twenty nine ofen bano street fifty three two hundred forty in n da
+ah exclaimed medor i haven't heard you bow
+i ind accept you in'd accept you pretending me for the intentions i haven't been there
+he also had trouble he also had trouble the seven brothers in his satisfaction
+carriou square le chasse say victor
+very well i would like to send you
+comineery
+of course i am not sure
+the floor is now open to mister staphon trumpe to support the three hundred eighth amendment
+marici is in charge of writing
+however mesy
+at three months i have my name
+he spoke to be avoided but he was still hesitating
+the will is about encourage the will is about encourage their creative cannot collect my adversion
+a few people is very educated but it was well to express himself
+she is very beautiful
+it is therefore not what makes the problem
+rigulus answered this lady did not know me
+other words it will be the rules of article l
+if you think it is satisfied i withdraw this amendment
+burtle road eight one two nine zero villers les matanes
+i'd run i sauwh i heard
+the pack road luber
+what a wonderful expertes every night she joined colossus
+the sound of his planes is unbearable
+i'll get it for it
+ah yes madame the marquise the terranoy for the moment
+badieu sarah the programme tells me that the test is over
+she turned round her head and turned her head
+the same day the duke of montebello arrived in burghausen
+sir what does this conduct mean
+immediately after the wolves we will go to the baptized the patite
+however the president of the picardi region was eliminated in the habitat
+obedience was due to the chief of the family father no at the sea
+imagine if i had even forgotten my wallet
+faith combines this unfortunate life
+i believe it well i believe it well i'n soret
+some one has done a work or era
+she engaged herself
+think i will not be able to do it again
+your clothes are delaying our fever
+do you see your profile really retained my attention
+lucky lucky dichans allotment in a racin
+for rue gilbert rue avenue
+bullion avenue zero five one hundred ten in masses
+one rumont pleaiser o eight six o o give it
+gard neighborhood in ornacas at the banes
+fifty five tubito wa zero seven seven hundred in saint jestardesh
+fifteen a lewis sarrid avenue zero nine one hundred and twenty in verls
+lisde ostred in sa german sir lot
+one thousand two hundred eighty three chennery road
+six de lerotnor we zero seven two hundred abines
+lanois wrote zero eight zero ninety ornacord
+domam street at number one
+seventeen moratiel de latra to tassany street
+abina road zero four two hundred s
+ten catres travers street zero two under enambrim
+debro ske wrote zero six one six zero an
+da terror is road zero six one hundred sixty anties
+three de pasier place zero five five hundred saint bonnet and chablet
+avenue de la croi de gards zero six one hundred fifty can
+yes sat n demont street zero five zero zero in gap
+nine palod street zero nine one hundred eighty in less avails
+rivedill street at number one hundred and forty nine
+maurice bornion street at number fifteen
+five hundred and forty horses rode
+let is pathway in soen
+three masses guys square zero three one hundred forty in s quenton
+seventeen di la clare street zero five zero zero zero gap
+roll street at number eighteen
+du petice street chooses
+bracan yak rod ess
+fourteen gaston boyer avenue zero four zero zero zero in dine le spains
+three hundred thirty alon linglet street
+one geret street
+corva street zero eight hundredad twenty bne sermes
+roderson o two five o o rinne and tira
+montelevir rode zero seven
+nine bes isabel sandy road
+mene street zero nine one hundred sedan
+number twenty one on barron street zero six five hundred ten campl
+fifteen lestyards lot zero one nine hundred ad sixty in servu
+fernand pinnell street in saint john gorf
+sim street zero one hundreda fifty in limond
+lime road zero nine five hundred mulan
+nine jewitt street zero nine one hundred fifty in saint german
+bauvoir at belle roalen du monteglin
+eleven camille coro street
+de la victorihersin place zero nine one hundred in the middle of the street
+on re charpentr street at number five
+six patagistred zero nine six hundred lera de lane
+six breast street zero nine one hundred eighty in saint jean
+darbo street zero four eight hundred grelins lespans
+dickens wrote zero seven one hundred thirty saint roman de lerps
+one thousand two hundred thirty nine cornes of castella
+churincon road at number three
+do village villebois street villeboir
+twenty four dur milec pathway
+one hundred and fifty five grates way zero seven one hundred and seventy level de
+dumin street zero four one hundred and twenty castellin
+rene governor street in viviers a cas
+one alphonse dedt avenue zero four eight hundred sixty in pierre verd
+four to bujon street zero three one hundred fifty vavinser a lier
+one hundred and fifty five plante street zero one one hundred and twenty denu
+supe mort street gap
+de la magnary alley zero seven three hundred run and sero
+one hundred and fifteen te bad costrit boulevard zero two three hundred joine
+rue dela robert zero five two hundred embrin
+chavanas road gap
+parpetery street in pagne
+sipaninade road
+monseigneur avenue zero nine five hundred mans
+eight may avenue zero one nine hundred forty five zero zero one three hundred bellet
+six charles buffier street zero four seven hundred resin
+palison road zero seven six hundred ten sushon
+chattelin's road in lane
+five pasado street zero nine three hundred and fifty in cantaign sir eruse
+de la fondna street zero three one hundred forty chintels
+seven floor alley zero five zero zero zero gap
+grande de marme avenue zero two seven hundred erne
+one charles gond street zero one two hundred and fifty cisariot
+kenrone at number nine hundred and fifty five
+on re bilson street zero seven one hundred thirty s pere
+on cross avenue zero nine one hundred pemirs
+alphonse richard street delis baines
+duver milan courtyard at number twelve
+florence street in sedan
+detaline street in boses
+soil that way at number three d
+peti ju hot avenue inken
+ten champo de gendarme street zero four four hundred bar selonet
+gabriel massway street fifty one thousand two hundred and fifty three
+montaine de junis manask is a member of the national assembly
+twenty one bedty street zero eight two hundred in sedan
+fiftee delel fifte delelerote zero five zero zero gap
+eighty nine louis blank street zero two three hundred chomand
+fifty thousand five hundred chain say jock road
+two hundred sixty five on dave wizon street zero two pondered in billyser hazine
+four pariage avenue zero nine one hundred eighty in saint atienne
+four hundred fifty two on de erasance boulevard in the same way
+five arnaud de trent street zero four eight hundred in grilo lespains
+sixteen maxter and sardell street
+one hundred seven vanderod trints changed zero three one hundred ten in espinas
+thirteen juck forier street ninety one thousand one hundred twenty three
+l alley zero three six hundred mulicon
+to lose road at number ten
+constantine street at number sixty
+de la sere rode in humbo
+five hundred and sixty bar less asin boulevard zero six two hundred ten in siba
+fifty three malson street zero three seven
+three hundred and fifty three say joseph boulevard zero four one hundred manast
+champtono alley zero three seven hundred bell reeveser alir
+george maria street at number four
+in the vegeley street was talking five people
+we are looking back to it through a useless date the resolution is exhausted
+i hear from the most proficiential knight
+a voiceless face
+two black tables
+thirty one houses
+it is a non negotiational condition to envisage our collaboration
+we are standing up and we are in the middle of the world
+a servant opened a b
+he tried to focus his mind on the conduct
+i'd have a lot of water
+i got up and gave the arm to my husband
+i'm sure you like it
+he imagined all sorts of things
+the people in all sustenses is the only sovereign of which we recognize the authority
+they had never seen such in cards
+so he fell hiers
+his mother
+the good one
+maybe sarah you have trouble finding her decisions
+a year thats a year that's the age doesn't know her
+come when the weather will be coming
+all the brush did not give any reason to this family
+my uncle the good word that hear is
+the thing belonged to his family since seventeenth century
+he gave him a list of books in the reporter
+now she fishes and turns the right to the south
+to morrow she's fishing
+how are you going to paint my anguish during what moments
+ia
+she would like your sister to like it
+they put the holy spirit to the sacred heart
+if i can make a phone call if i can make a phone call i'll tell him immediately
+incredible show was opening his eyes
+she moved forward with joy
+he's hot he's going to play the game
+you know that of course you did a job for yourself
+the breakfast was completed
+ully in a desk car
+everyone he saw the widow of learning the scheme
+we had sold them one of our solutions in the hens
+one
+neighboring factory several hundreds of workers came up
+nine tables
+musis way as a neutral at the theatre
+my head is not very well known
+purse
+he's doing well
+he took his fingers
+i am still buying the red label
+we will read
+thirty one tables
+the conversation lost and went on
+she did not ask
+next to his chair in a jump called back the house civilized
+you will still be welcome at your nistonys
+another time assamir the royal family is going on the street the battalion
+their hands are not very well known
+i know it very well
+i will do some of the bridges on your candies
+you dont you don't eat enough
+seventeen eiht five
+i know very well
+we need to be brought in the burned
+we are here we are
+you were certainly going to go to the semenar
+iam a graduate and engineering
+model street thirty nine thousand one hundred crise
+she asked us for it to say how it was
+mister bernard coyer has the floor to support amendment eight hundred twenty seven
+i would like to stress three points
+de chevre square at number six
+yes i will come
+people known
+albert groman tried to mask his smile
+he went to bed after his death
+i'll explain this
+six six sections there is a specific device
+mister the rapotter said to have appointed a very important role of the financial issues
+when this will be ready to be done
+oh you know at this distance i only see the car
+this lack of pruder
+she saw me
+i hope we will succeed
+what is the government's opinion favorable
+two shots
+have you seen coffee
+it is of course a quantum field post mortal
+the floor is now open to mister francis for kemmer to support the nine hundred eighty third amendment
+we are still requesting interviews we will need to start again
+this night is the only one letter from her father
+she had written the name on a paper bottle
+you want
+they fish once
+in money
+she had seen herself at the end of not receiving
+maybe she is also legal
+he continued to prevent the policeman asking for a few days
+it is an expert
+she did not answer the question of the two thousand and fifteen
+the eighteenth one is the most important
+i remind you of the term of the arcl l
+it's your turn to play the role of the two parts
+he will be brought directly to pachalo to morrow
+forour fishes we fished with a lot of people
+we cannot say that about this
+they were even in spite of everything
+this mass over there
+it was one of the objections made by bolsman
+you can call me albert weare call its now
+they love their parents
+we are reading
+thought about me
+they will accept it solis since she is the biggest satser order it
+i am in the house on noon
+that is quite new
+we still have time
+we can see that the night was dark they came back on their steps
+i must absolutely go without fail
+on the thirteenth
+the work of the investigation showed a emergency state
+they had to see clearly show me what you prepared
+everyone was rubby everything
+into wine
+without a word the department are better equipped with others
+in my room
+in my head
+yes he used a technical gun and was garlitated with those words
+the good ones and bad ones
+he's tried to leave
+what's that
+three houses in woden
+she sighed this prayer in a smile of nevance and disappointing
+he has loved
+she quickly pais a decorts
+your chairs are not big
+you will love it
+forty horses brown
+the strong ones are the most important
+i'm going to skip all the more bread costs of the world
+the bridge had agreed to abortion the whitehall
+they had the eye on all night everyone else
+we made a popular consultation in fact a meeting
+in itself he was thinking about the means
+his son thought it
+he recalled he recalled that we were a several times the same time signgerer
+strange thing they didn't strange thing they didn't met throughout their profession
+i'm thinking about it without a misfortune to the same in the same act
+these intermediables will be pushed the hours of facility as they went by
+buttons are almost like the same
+a small piece of furniture is located in the south of the island
+she had not seen her again to sacrifice her
+since then
+you brother
+there are sick people among them
+with this the spirit of any is a great help of rest
+the shelly isso was very pleasant
+cores de boiso street in dubo
+she let her give her the notes to her under a new name
+the devices currently being enough to us
+without a doubt she will be effective
+it's dirty it's the same as the previous one
+you go to the business do what i said
+now they're fishing
+omer this let me not be worried about
+his college was waiting for a long time in an apartment
+he has a reversion on this ono
+i see how to put it differently
+he accelerated the pace without exaggerating
+and i am trying to follow them
+r day street at number forty five
+corbet rode at number five hundred and sixty one
+he was in the middle of his lifetime
+seven b chateau trumpet street fifty three seven b chateau trumpet street fifty three one hundred
+we are still happy to inherit from one uncle or a cousin
+but the enigma seemed unable to him
+we are waking and heavy
+you were thirsty
+an old thing
+but be certain we should use the most existing base
+i'm waiting for you to wait for you
+trustworthy of the press to save valery emily staying in herman
+thirty two for up street fifty four thousand inebre
+my brother is already arriving
+grellet street eighty five hundredand forty in chache
+the floor is open to missus michel de laney raputter to support amendment number fifty eight
+the cool of the pit cavil also shied in the first challenge
+and you and you will have the intention of giving you some day
+your father is tall and you are not very well known
+at its level i had got the strength of the english soldiers
+i am sure you would like it
+you assureme you assure we'll have it
+in union
+it also concerns the same neighborhoods as the privatary of the city's policy
+i'm sure you liked it
+don't come here
+his body is not there
+we are even going to fall in the regard the next speaker is mister damien a bud
+de montelevitleno tatton was a very good man
+madame ericourt was going to start her head held her shoulders
+twenty one tables
+he predicted that he couldn't he predicted that he couldn't he predicted that he couldn't have any trust in the following year
+she learned to be more than a deadly man that was the last ones who were filled with the bicycles
+my master remained a wanting to you
+what distance is salem from here
+you never come with me
+you have loved you
+these walls
+now she fishes
+this poenician tale which is another italian cout collected in lavorne
+ducresion road ameges
+the same ones madame bodeloche entering missus boudeloche
+please
+then this o then this osetter fegilitates the credibility of france to urion plans
+missus luegi are you maintaining the amendment mister tardy he is holding it
+renaud who was ull was still the house also
+in your house
+madame madame lenarquise i'm asking you to see
+the same gain was lifting her up in town
+the assembly agreed to it
+what is your father doing
+it's a letteros col to seven hundred and forty fasemer
+it's true that we can't be alone very slowly
+this house
+this amendment seeks to improve the arcle without a change of the substance
+ah that ah that's a pleasant view
+but elvir her seems to me to be lost
+because of which
+the efforts are not very well known
+you will love it
+de decas road at number seventy nine
+miss juliet poitier the scene takes place in galimard
+there is no left
+i came together with my family
+we should know how it is the majority that commands
+i first examined the outside my piece
+there god is with us
+but do you think i am shaving myself
+silence said markov shrugging the young man who was gone
+dupotlu street lerver
+but respect to all
+he consulted himself during the seconds that were running
+in his current traft this article mentions rural paths without any precision
+he's gong to play the game
+a little pleasure like this was not refusal
+my walls
+indeed the only risks of sanitary are those of a moment to take a moment
+i think this date is important for our regions and our populations
+us totally displeased that they gave it to them
+i'm the one who scares me
+lord montague that is in reality the spanish of italian
+i have his portrait said merceline
+for
+joseph dstreet ninety seven two hundred fourth france
+judith and alory fragmented in verse
+don't spend all the money
+good evening our masters
+we have already heard the discussion of a budgetary waste time
+he is also defended
+the boy is more than one of the news from him
+the scene was terrible
+one day they went to the mass and then returned
+nature seems more pleasant
+this bell tower i saw bot one under my windows
+i'm a nurse that's why they should helpful
+say sermon i don say sermon i dont see the use of your amendment
+i wanted to vote the phone
+i gave him four pillars of salt
+i would like to bring my testimony to the government's former members
+he had the costume of the dragons
+but when she learned that amanin o saint vec kelvy
+rawbites riders pickaxes i will not be old here
+come on come on we have to cut it
+we loved your sister
+the great man had his patterns
+what is your light of work
+the two old slender women
+residence dilass on erat and ron
+six joseph pseudo street
+his small actors are wood like gods
+it is a hmatonian system depending on the time
+why are we boared
+madame reporter will answer me very slowly
+he had to surrender to the hydens people who were toning about him
+men don't want to do anything
+i would like to have a table near the window
+me why would you be
+these tubla are far away
+can you please fear you'rl be abandoned
+i came to see you
+don't go straight away
+my mother is making me lose my mind
+it is not only to accumulate poor funds
+i even like to
+now you can fish and see you are fishing
+he's not with you
+how do you do to get him out when they happen
+its waters are shuddering
+silence and smiling to his habit he was looking around
+his hair took place and perfumes of jesmin brook him
+the projectyls threw two apprentices down far from their paper hats
+you asked me to write down to take a paper
+his finger pointed at the nest
+amencha stuff talking about a few automatic sugars
+his vibrated seas are also resuscitated in the big resuscitant rumans
+aner he will be vigorously his hand and the vel
+they want to hear from napoleon in the middle of the al capan office
+remembrances are very well known
+there is one of the most beautiful stimilions
+they hope that i have to go to toulon
+you have no idea
+a member of the municipal commission of the provisional government
+when he had told me you were out of the way
+sillet color was tearing with ferser
+the major wasn't the major wasn't listening to the resentments of inevitable
+the descent was lost by a precious allay
+women are in a hurry half happy between their gillet sleeves
+because of the breakdown flecting at the first sign of a coop
+the game from the day of races is the same as the french people
+this is therefore very strong
+do we really want to vote for this amendment but i don't really want to vote for
+the general sherard was born as a war minister and cassamer perrier at the finances
+it was too happy to meet him before the war
+a false savage was tearing through the air
+tu force she would do it
+i could i couldn't go up with my miss liddy pa what do you understand
+you you limiting you to the paston that's a torpedo
+the father noell he lives in you
+the associations will go on to quabeqir
+jundet street ada
+she will go to the end without duty
+rinaud kept a place
+the end had jumped against his canvas painting
+in worth giving them some diseases with our rocks
+twenty roue we still have our national guards at our disposal
+neither the investigation nor the major had their military costume
+we didn't hear each other any more
+you will sell the goods for a long time and the company will not get it
+and all of them were laughing and they took out their eyes out of the game
+the soldiers brushed their vegetables and relocated the jalers of the bonners
+he was addressing their powers in the middle of the sinister voices and dust
+if you were so young omer i would be asking you to join us
+outside the news is propagated
+and you saw how many times
+vine on meswoman to have attempted to win the tumult
+these two big men were whispering about walking away
+everywhere the people woke up
+he said that he sent us a madness
+yes he had shoes and it was electrical protinet
+what rospel approved in the huge pass of his frock coat
+thirty when a climate is the weather is nice
+we left by holidays in italy
+lafayette was coughing looking for his tissue in the folds of his frock coat to emble
+the great pity therefore does not provide happiness
+and the relation of lugubrious stopp the disorders of the liberal angas
+i will be there for you a little insignificantial board
+i think i will like in the runs
+the doors are only found in the north of the brazil
+i have i havena chose when we have the goal of this morning
+maybe the seventarians are in the middle of this century
+lionel juspin did not yet lionel juspin did not yet explode
+my computer was expensive
+lefite listened to the quarrel with her sofa
+the king tasted his weights shook around the hobby face
+the bustet marble of a six zero zero zero occupied the centre of the fireplace
+indeed edward supported him
+will he be able to give you the pope this or the rest of the time
+two people avenue ran to the stables the whackers and ciberis
+the lyrics were resounding outside of his lips trembling in the verdict
+one thousand memory of pathetic emotions was giving birth to the memory of these orators
+immediately the orders of the officers were propagated
+he gouod work
+replied one thousand voices and handed over the theatres
+the beam pole is the priest graniers in the police impatible gendarmes
+he dissipated these memories of the war that was still in there
+to come one day in day laughter when you were laughing
+what happened in the first place for the pull
+completed your happiness
+and he himself shuddered
+bruno saint lame this forvoy
+judith is leaving the corps
+the grand hare's feast ended at fulling
+nature merris her who is allowing her to get too much happy
+conclude the major
+she is a nicer i couldn't see her for her
+cati perrry will regulate the lovers by succeeding in a third party
+this is why this is why i am holding it against this half pay
+we are still not in their catalogue
+his sol kendic and radius faces him as a flower
+he still knipped he still knipped the civilians without supportable
+he gives you the same reasons to choose our duty
+i thought it was only a case for the project
+omer sent him to the devil
+laris didn't want to spek laris didn't want to spare from the beautiful way
+the monthly treatment was integrated into the planets
+his affair could not take a breath of delay
+then he tried the surviving hisbaks more he could have been chasing
+the name of poby is a new knowing
+he didn't see him
+piter's book is a snake in half of them
+almost all the adias preticisd and it is totally untangled
+she she grew up on her car inshore
+if she continues no bring or son
+i i want to tell you
+the fire will be slow down
+guy on the right side
+we are not going to print any more
+their bones shouted the joy of their pain
+the fallers bear their name
+here to each woman i wanted
+mercury is looking for more time to think of ethoclasus
+he was barely waiting for him or almost all
+the smell of the blues entered by the hots window
+the major gaslok blocked the belts ordered the potcher
+double a silence en girdled
+he caught the tourist o in tenia
+he would say god to life itself
+we are forced to refuse the validation
+she continues to be clearly on this captain
+he is a mystical and an inventory of the dibber of all the way to fight the king
+they have a government on the other hand
+fifty chests with time no stirred the ground
+jurian from others to business
+edward de ster
+if you wish i was a friend the chamber of the vicar general is accessible
+the government of the assassin you will succumb under the katakes
+indeed i had such a little bit so far
+to night we will go dancing to night
+it's ridiculous
+and i was far that was even more horrible for my side of the drama
+you want to say that you can't
+these detractors are very numerous
+she lived in a forestry of which yhe never found the fort
+hallo my old lady
+every morning this broad cast is a step
+the step the step is romantic with the colors agreement
+the arbans were piling in the garden of new burges used to be the address
+the windows are not opening the door with the windows
+i adore this movie
+he only had questions but none ansers
+violinists were playing with the lace of rossini in the sere
+the use of baits sometimes goes beyond the time
+lafite at omersam close to the enemy
+it is used by craquenes that she ate
+don't you break your weight
+you cannot be able to resume your resignation
+because general was right handed although the indications were out of the war
+or champs predicates that the cartridges show to the tanta soldiers
+nineteen scientifics are suffing from travelling to see your keyboard
+the bourgeois left
+the day of to morrow is going to be decisive for the rest of this project
+de loyola was de loola wash more than he converted
+scared olmers jimment was annoying
+she knew how to open her hair dress she knew how to open her hair dress into the sun's arrow
+she made noses follower
+you could have asked for lionel no
+it's a nipple it has changed a little
+they were cool and languorous
+the bordeaux was borne in glissd upon thely of the vehicle and ran to cavoir
+what do we do what do we do what do we do
+every day she would like the tree
+the pigeons are turning into the air
+these motes are the most deepest of europe
+it is a decision that is committing with you
+the conversation is made all politics
+i can i can i can't write any more sentences any more
+i think that dead lions attacked fifteen basil pacifists
+inside we were freed
+it's because she has to defend the bar
+he touched he touched the starr track of the bole at his neck
+suddenly he frowned his black eyebrows
+he seemed hesitant and embarrassed from his sabre
+the vender of the mill writes an ordinance
+nevertheless he had funa felt the band to go out of the shadow
+however he was persisting the arm to the sky in the mouth
+the crobuts of circus will renounce a dry hairy carnivorous
+all the windows were questioning the passers by
+the republic was returning to him it became a uniform of these heroes
+all their bloods were singing in their wolling veins
+a secretaryy said to the emberer's ten plateos
+don't fear another waterloo
+with them nothing has ever been certain
+a storm was formed at the horizon above the edges of the enkin
+a very sure thing than the success
+the snows of his hair spread on the windows
+proud limited the heroes of the selection the tenager was crampling there
+if you continue no ski weakend
+three white tables
+when he was crossing the road
+you you do this alone
+she said approaching now
+she was a lot of energy at this activity
+you were not complaicing anything to progress
+five times but it was huge
+olmer turned towards the corner where this path was
+she said murmuring
+a seductive doesn't a seductive doesn't want to let go of it again
+yet it would still be necessary to find them in the closets
+francis
+my little charles
+the decoration of the major is severe and made them impost
+skater's addiction o line and mills that are still a surprise
+not barely anybody is already loaded with it
+undertake and succeed
+for a long time omer was shivering on the humanity that paine was dead
+the admiration of general cout is a style to his own people
+where are you going
+a waker crashed down and spread out sitting on a spot the stools in the air
+some earthquakes are looking forward to give up the group
+a depal and pegan climbs on reasons
+i have two dads and i'm nonosexual for that much
+there were influences that she was at the door of a bakery
+tell me you tell me you're well aware of it
+the floor is now open to mister john pierre barber to support the one hundred seventeenth amendment
+here is the last grin master of the ceremony in france
+everywhere the red and the blue one the most illustrated were arbroted
+a scientific discussion brought them closer
+he is capable of making us lose
+his doesn't have anything
+you didn't want to talk to him about suffering but death
+at home there are three t b shows
+the pistols in citili were banging
+we were shot by the windows of the louvre
+first we are going to get the remix of the call for the experts
+astris were entertaining melipies
+in one hundred minutes and a montana will preside the ear of the reoceros
+the father and the daughter monter
+the dints are recognizing a power
+women in accompany them
+i am born in a thousand seventy three
+when ye're pleased to protect helo
+bridget a hubbled row
+sasan stirred the boots of the military
+teevening has fallen
+it was not the desire that i missed me was not to do it
+twenty one alphonse de noville avenue
+de la chamble street twenty four thousand two hundred sate flour
+ah finally my father who knows he was dead
+they were finishing off the stairs of the interior
+a colonel inspired by the alert
+looking for a lir looking for a litter to seduce her and she said to speak with a fever
+other animals
+the server will pour the coffee with smokes as a silence
+you have a censor
+medan mel medan meld comer live in the woods
+the fat married envert lead her in the pity of tragic
+having been addressed his nuff boat a whiteness looked for his cases
+at its garden a form would save the passage of the barriades improvised
+treneve
+so it was true in the baby
+curious as the inevil parish was limited to a church
+quickly the disappointed
+it is a return to the back far from all
+it's not raining at all the time was wet
+garfield is a shawl that speaks
+but we cannot see their promise to eat them again
+stop seeing you again in play
+since the soldiers were finishing with since the soldiers were finishing with ammunition it was the time of abruption
+the gasometer of the blood was spoiling his mouth
+he let the waist of the girl who fell down on
+he had the ambition to limit this courage
+i am grunter to the one who is a great man
+here is the report i am the deed of the ermatic
+it knew that the dead line tried to negotiate anything with banks
+only two of them spoke and not a little bit
+in such place the poor girl had come to take her desperately
+andtimiliterorists are imagining to succeed in this medium transport
+he was flopping at foots and literary flattery
+but one hundred and twenty god askis albert coin
+this ice is colouring without a curator and much better than the industrial ice
+the writer is a solitary master
+it is not capable of being an initiative
+the courtier will help the document before signing the client
+the world is too small to cosieri
+and it's in fact the case
+the storm arrived from blick to the place where the mails were in the middle of the room
+i think we agree
+he had never heard anything about his books
+it will be all
+i'm know if i i know if i'm going to put it back
+many colors to be swallowed to teachers good of unnamable priorities
+pity for them pity for us
+matilda waited that colossic strid
+you machine is assuring me to the right
+sebastian is not connected any more
+and pa the restaurants are really in a hurry
+and what is that a pintle of bread
+to is in the golden history
+the following this project is really desired
+where did i speak a pillar
+a page turned to sell last wenes
+let's go to persecute the church
+this attitude of politics consists of nothing of it since then
+by the agency of a united nation dedicated to standardization
+they finally get crazy and what is nfortunately very original
+i go down by running the stairs and make an out of gold
+pachiche had the weight of the kings
+bennan it was rather the opposite
+jilling gresky is mixing but
+she does nothing but complain all day long
+the eyes of the car screamed on the road with a joke and marie
+and juras had eleven pieces of clothing in all heads at the vestibule
+she got her wound
+i really believe such a thing to be possible
+jean it's judith
+the sure he likes it
+to day does not bring any more logically to morrow
+you you phone me if you need
+why wyded the lucifer's face had he feared the joyfulness
+he is buried
+forty seven forty seven liold forty seven licoldex way eighty eight four hundred in gerardmer
+adi was for jeserus for the king of victory
+he enters on the right second plan
+it is the first thought that came to me like that
+she listens to the door
+addressing the stage to mister brier's sus fler copist in the theatre
+this can't be seen as long as no one is o
+mister de la portere
+hit by accident he was not less staining a great laugh
+he was as much as you came back
+i accepted yes i declare that i accepted saint qualtas
+we will catch you one day or the other
+or a chance was questioning the right to drive them
+they joined the automobile in silence and their party
+aguspres avenue eighty two four hundred volens
+yes it is a woman a very beautiful woman
+the door you are painting
+in this case he is not able to shoot him
+he plays as a gift madam he may have a crime
+fourteenth march eighteen thirty three
+a big bellow of the velloor brown veals played the hangings in the air
+oh going back to the chimney
+exclamations that saluted the general republican generals became heavier
+there was a problem
+eighty six augusto avenue ninety four seven hundred in talents
+this need for my elvire you are dense
+what can he change a point
+he said while lowering the room
+the same noise was repeated four times weakened by the distance
+mosquitoes still red in this region where the city is still in this region
+but what do you think about it
+finally the ser finally the ser on finished and semi finished products that came from the zone
+on the right in the foreground there is a door leading to th emmett's room
+the prevention is always better than the repression
+paul ester has many sav megens favorite of the king
+indefinitely the count was talking about the way
+there are research engines to enterprise open sores
+one day of big reviews his wife recognized him
+you think the good saint martin still exists
+but you know i doubt
+i am defending it with correction
+is it important to take changes
+mister decharn lifted up in a wheel
+the music from marseilles by the guard of pluvican
+the bird flowed into the air
+regarding the surgeon of the law he is watered
+very well the floor is now open to mister shock lenblen
+what is that began
+if you were coming we would have lest
+the repatter's amendment will answer the opinions i am going to do
+unte ali in saint john de la montreal
+thirteen joseph thirteen joseph rine street nineteen three hundred egletons
+the unity is open
+bunistraf is the organ to be closed
+i thought he won't end up with those in america
+charles albert avenue seventy three two hundred ninety lamotte servelex
+our economy is based on everything
+say german missus luxish
+and here and here and here is coming a new world to my sense point
+he won't stay without any but he's a certainty
+six beckvell street sixty two hundred twenty one
+the firdust brother access to the torment early marches
+why is it not that there are several complete factors in this event
+i even went to run the novels
+twenty four monsagner turnell street fifty four thousand one hundred a twenty in nense
+the church is true but nothing decided to this respect
+he struggled and softly
+what is the government's opinion naturally favorable to numbe fourteen
+but he promised that he was not a poet
+cot cot dutton street at number ten
+here's a woman who is seeing herself in her work
+they all wear on the island and have been rejected in the commission
+punto dvesco housing estate twenty thousand one hundred an twenty eight were set upon
+on the contrary oh that he is leaving me he's bothering me
+what do you think you have on your back
+by the way all the romans departures attempted to celebrate
+the government will not doubt it
+ah well you were holding a lot of you not to lose your father and you
+he folded himself by the arches baying towards the alles
+oh sitting down and showed his plans
+we must conclude mister siadi and the minister of justice
+the floor is open to missus the rapater specialized in support of amendment number one hundred and fifty
+we have to be vigilant
+one hundred and twenty eight par avenue forty thousand one hundred and fifty sorts
+de mender square in bordeaux
+the floor is now open to missus maria chapdelaine to support the two hundred sixty third amendment
+he could be five in the afternoon
+they seem to me that it applies particularly well for this case
+thirty four pavoy street seventy two one hundred and fifty in le grande luse
+thereare other chairs here
+the floor is now open to mister d da martin to support the thirty eighth amendment
+she is going to progressive
+why did you let her run like that all of a sudden
+i did the work myself
+the waits ring the bell
+the communes can decide like this
+the floor is now open to mister john pierre dor to support the identical amendment number one hundred seventy seven
+wal the idea was mister taking this boat
+for eight days i am attached to you for an hour
+you didn't have the very regular movements but it will come
+how is this girl
+i had the happy hend
+in ectem there is no device of the diagnosis
+this species can be found in the heart of the world
+the fists were running the guns
+the kid was filling the same with the altogether
+the floor is now open to mister stapan claro to support the one hundred twenty first amendment
+ardoi street at number twenty one
+the dragon used to use more than fifty years
+with meat
+the floor is open to missus oddre lincoln held
+yes you are anarchists the terrorists and communists
+eleven bom street twenty one four hundred twenty bin
+two hundred twenty five john michil avenue ninety seven thousand three hundred fifty four
+so as i told you so as i told you i'm going to see gidard
+no way to go without raising the place to the clergy
+here are my dear calits elements i wanted to give you some day
+he is removed from the national assembly
+thirteen johan street forty eight one hundred marvehls
+here are some of the fostenses that you were finisting the event towards the story
+did we have the card please
+allotment on the benches of the socialist republican party
+we must not embarrass from the population
+twenty four wooden chairs
+and what are you doing before
+when did you see her last
+the floor is now open to missus n the floor is now open to missus m genevard to support the five hundred thirty eighth amendment
+another competency is not as another competency is not assuming economic development
+missus godan brings closer mister genard
+what do you have
+now you can go back to the kitchen to pick up the neck of your exploitation
+mullery sunnia street sixty four thousand five hundred saint john delos
+seventy however is not very distant
+don't i get dressed like others
+about the business in more than fifty employees we are getting further
+it is a real leadership with the conservation of these spaces
+i am not in favor of this amendment with the large legalization
+one du paty street seventy seven hundred in frednes chateau
+the benditing horse rose sadly in his stable by stirring his backs
+not to undergo a reason for the rest of the game
+i learn that the sential
+freres de red street at number fourteen
+in the neighborhood all houses are replaced by blue ensigns
+for a short time i would have predicted to at least impact and my acts on their environment
+besides maybe there a difference between them
+he's ten years old
+nine nine o'clock struck eight hundred thousand francs
+jerry is a hear r old man
+you don't know how to do it
+de wards road fifty one thousand two hundred twenty sa theory
+men in rode at la chapelle de lactour de lator
+a book
+this has been held on ten hours at five in the afternoon
+i have to work more and more
+alas i am suffering from the loaded people
+to morrow we will have had the convention
+ah bitter reason
+jok he at your toda
+it's cryin craps a good one for shows
+this amendment goes to take us a certain amount of time
+he can only be from time to time he can only be from time to time you need a break
+thinking about the future
+i have a large pain in my arm
+his natural habitat and his gensonest traditions carried him a lot further on m ps
+these dead base are protected as the property
+his father is without reproach
+however the first action has been considered to be the reduction of the paper
+i could i couldn't do anything with so little money
+same thing for amendment number eight hundred sixty nine
+the marquis in the moment of an unspeet killer
+twenty one another of his classes begins in san francisco sibils
+it is the subject i had a thought
+my ravit was a perlapse
+ther are identical elements painted in blue orse in the richelieu room
+adeline rises up and kiss marjory then florian
+i'm shouting
+this being said i am withdrawing the amendment
+sycoter in the spirits
+we are born in order of the order of the town of the same year
+yet he wast replied
+the long unpublished man threw you on your face
+the work consists in a good drafting of the order to determine that they will be the simplifications
+three hundred eighty one la chatri avenue thirty six zero zero chater
+he is complaining to school
+toke is this writing
+it was the case of january two thousand and eighteen
+when he was crossing we can see his truck the young prostitutes guards
+well what do you think about it i think it well what do you think about it i think it well what do you think about it i think it over the whole thing
+omer in this neighbourhood was walking a lot
+your suns are they dry
+he will come back to a minute
+ye say behead compassionate with the help of the people who were toning about
+the answer was provided by the roll of tartar
+forty six lucion du pi street twenty eight thousand five hundrede
+bigero street in kingeshime
+we are waiting for an inside the scene
+david tried to remer it but he came back
+a lady all the members doesnt a lady all the members doesn't know how to take the masses
+in this case it is a french issue will support seve
+you just did something else with me
+well do i am not the house of mister digin no doubt
+i could not say why but i had a bad feeling
+i would like to repeat what i have already told about this subject
+order to verbald is italian guide sent to brag
+lady i'd heard him
+we are not just thing at the future
+and now he is trustworthy
+fifteen claire basstreet seventy eight eig hundred ten in fectro
+erona street sixty thousand two hundred campain
+we have already brought a significant change to this article in the first reading
+i confirm my brothers are not the subject of public domain
+washington wanted to be a public republic
+don't attack me by you and i will not be able to attack
+the charles prince with the deboids of his armed forces extremely decoded fruits
+how much is this the problem with the help of a news of the people
+i'm stressing this is the only ones that are typical of the ones who are not the same
+there was however not a horrible rush in the house
+there is no heating in your bicok
+aldrup knows the inclinations in the conciliation of yours
+she was wondering for the idea how she had brought her closer to enna
+the apprenticeship requires an important sum of ungrateful attacks and up stream
+pierre tono is not far away
+maltinu trunkards in the peak
+but she has the reason to fear
+your friend does not change your mind i knew your consents
+an there the school children laughter and rush
+the commissioner of tisidon
+sport is not the only intention of any matter
+these are just there he only stained the engravings with a small old lady
+we will be proud here
+i can hear what mister torrain said
+this book is often studied in the colleges is easy to french literature
+i wonder if he didn't make a proper meeting
+twenty seven pierr dor street eighty six thousand pares
+everyone will hope to organize this race or not to mention the public
+we were fighting there
+but comony was freed and i didn't tell him
+it is then that i understood what she had left to do
+i in't know any one among all these passengers
+dunner riche is a very good man
+she loved
+this seems very useful to me
+the count of prexy blasons wasin the count of prexy blasons was in care of them
+commendant christian black avenue marmond
+there will then be a transmit to nue
+he didn't get involved but i was a kind of climer
+you were still hurting the back yard
+the pavilion bar bores its usual com
+but your amendments number twelve and eighty eight are identical
+deneberk street in mewiller
+mister john louis torraine is amendment number five hundred seventy four is withdrawn
+suddenly the silence is being done
+so the guys need a hand
+what is that o pastor soldier
+oh thank you very much
+in desert there will be abhorem
+miller street sixty six thousand two hundred and forty sad estive
+it's not exactly that
+people are selling to remove it
+mister ernest is not here
+they have too late
+a dog is that i would like to go through the window
+sorry it's in his bedroom
+endify andify em here strong enough it andify em here strong enough it andify em here strong enough it's all right to you
+we have to have about twenty years of differences we can hear well
+the department is not visited
+oh the man the wrecks of paris
+these two amendments allow to take into account all types of companies
+colory lot ninety seven thousand three hundred kaen
+ah well then you but what are you going to do
+alan had just taken from the chef
+come we're going to sit down i'm sending you back to life
+for them the simaris are not the same as you were killed
+seen to the same camille
+intention he is about a brand that we cannot recommend to everyone
+i have seized of two similar amendments
+eighteen au r m e street sixty eight two hundred forty in siglheim
+he does not prevent from it
+we made the tour of all the questions
+was he where no in the great now
+mothers at the end of the arms raised their children so they were great helping
+one thousand three hundred forty five b cister pathway eighty two thousand seven hundred ten barbes ss
+i will bet into a light on a few points
+how long will you stay here in the middle of the world
+coming back from the market you will bring your book to the next day
+forty nine louis delcip street ninety three one hundred in montroy
+there is nothing to do here
+this does not happen
+surge my son a famous and carpenter with the destiny
+it's too late
+it was a bad doctor who almost seated
+it's a sadness of sulitation in chubby faces
+he is also a key player in all the ministry of his conscience and his trust
+you saw this movie
+increase in him but what kind of situation is a
+so i saw the same corpse as before i found on the ground
+here is the one who begins
+it is not the most effective job but it is essential
+i'm leaving in your you admission
+the university is closed all the way of august
+i would like to know them
+it is in this way that article nine is a constitutional
+can you say how this hell is there that there will be thirty five at the end of the end of june
+it's an important moment for our palmed
+iam being out of the room without making a noise
+the husband will find the buddha and the woman spoiled
+it's good now
+she made him great smiles engaging
+i'm looking for the sun with my l i'm looking for the sun with my lick what you want
+we took a commitment
+twenty one chasalet street forty eight hundred invelford
+he also has shoulders and his face is closed
+lewis was dead so far before his child's birth
+i will ask you to be brief
+omer listened to him
+akl five quarter thus comes to the principle of necessity of delegats and punishments
+you could have hulked him out every night
+one thousand million bombs
+this is perfectly tolerated by all of them and the other ones are different
+this house over there is a few years old lady
+the scene is not held in paris at chaffone
+i heard a lot of complaints and
+oh but it's just that its oh but it's just that it's me
+we have already discussed it with it
+gm telli stret thirty one zero zero zero
+the movement is generically almost perudial or periodical
+semi escaped
+it was as easy as it had decided to make a profile
+at the same time he noticed malvina all occupied by him
+if you want my opinion they will read it no
+fifty four do douvil street seventy five thousand and eighteen paris
+my old man you have to be a bit of money to be on his knees
+louis philippe moved forward discovering the face of the crumpled part of the favorites
+i ask you to with tried
+we will see you to morrow
+the vices of the lock soon stopped
+lord i want to convince him and no more words
+i am too weak to jon the honour of the day
+we read that we are children
+the floor is open to missus edde the floor is open to missus edi conod
+we would say not that they are full of surprise to morrow morning i would more
+i idon't know any more but you are nicty that loves you
+denulet avenue in negoro
+basbasiles belles le campaigne cannot vote for the chases
+he dominated the ranks of the guard of the station
+she had already made a letter that had left without any answer
+a singer
+and he sat down in a suffa deep
+that's what brought us to ask for their withdrawal
+clodi was maneuvring me like a lion
+a coalstool asked the woman
+twelve doctor marcel her street sixty eight three hundred saint louis
+rin equent street in paris
+are they good for all britny no for the corny only
+ten to branie avenue seventy four zero zero zero ans
+tis kindergarten's apprehension without doubt the countess exaggerated him
+it seems to me that i have things to do with my best effort
+because i believe it will be very unhappy to make this amendment so much more unhappy
+the tribute is not to bigger
+the floor is now open to sir gil hornado to support the twenty ninth amendment
+look at what i am looking at
+he plays in the fayette a little bit marbled and a little bit
+i didnt i didn't want to have the effects
+the actor lost thirteen kilos for the movie it excavations
+the school of morley
+the floor is open to missus sandrin maseter to support amendment number six hundred forty four
+i like the jean but he doesn't love me
+his mother is once again in time
+is it a reason to help the good intentions like the betil
+it is difficult to support the kenik its difficult to support the kennik when we dont have climatizer
+here's a real subject of concerns
+it is still given a feeling im rish i am rich irish
+for the pretty calmness in this reply that he kissed her
+what you have
+near the door
+and what time will you go to work
+but he still has eight years to run in our way
+the house is under the control of the castle
+doctor terr de martel rode at number three
+preached in the oven to two hundred degrees
+sixteen camp del rostreet sixty six thousand four hundred fifty polisters
+he imagined this music of seventeen years old boyer at half
+we support this idea of regulation under hi
+here i am two times your editor
+twenty orby street
+they are passed here
+the fitimes passed by
+don don we say master do by credit
+for
+but he didn't but he didn't say no
+when are you going home
+the heat increased
+do you want to come with me with me
+your contract is for guarantees that you are currently being outdated
+send me a lease the hanging of it
+in a way to be heard by everyone
+we will measure the scope of the damages
+well and
+it's my duty i am a civil servant
+madame de sabra was madame desabray was stoodap and henriette
+an old man fighting took place in the breton mensal
+she takes a cap of the administration's cap i'm ready
+i also salute the progress of writing in this direction as well
+he is excused to save me i cannot follow her
+the gra preteris was moving and screamed down the major drums
+it was therefore a proposal to which i was expecting me to be upside
+once we used to fish
+what was the wrong time this is what was happening to us about the same thing
+croy quarry residence in litaly
+cors guret twenty two thousand two hundred ten plimo
+finishing he had taken advantage of a work accident to help himself invasive
+pinnard was walking straight to the cabinet of the presentative
+to find the ante room thinking justin
+a friend is waiting for you to take good decisions
+coming near the plat i turned the torch
+fifteen rue de reneles twenty eight thousand eight hundred monpasier
+to whom will you give your gift
+next to each other the two men climbed the steps without saying anything
+poisoning the labor inspection services yes
+there was no bravado in his case it was a natural door
+forty road street forty six zero ninety in flaw jack pujo
+the floor is open to masta minister to support amendment number two hundred nineteen
+he started by the selling of the jespets
+you don't want to come with me
+suddenly he's the hurricane
+the floor is now open to missus eledy jacker le force to support the six hundred ninety
+the old magistrate fell on the right
+here i am the one who carries out by reuniting all his forces to do something more
+nine nine o'clock in the morning has arrived
+his glassy eyes suddenly classified and made a bar in the back
+nevertheless it was only towards the enemy's approach that she reached a global level
+he discovered all the fear of the lover
+mister decharne was not here the fury gave him the idea
+i finally arranged our interview til tuesday eleven at the month
+missus valentine marta valentine and sumbele your voices
+i have orders i do know it
+he well he is the smallest tree in minno the future mister shenton
+he has his key
+it is the way to bring the pavilion
+other and big chairs are also used as a white one
+and he ate my duty
+prayer she was fishing yesterday and fishing
+the purpose of the final is marked by abdel at trebesse to the first mitet
+this wreck fire which the boats came from the ocean towards the silla port
+the kenton of castelle saracen is rebuilt by this degree
+ahchmer under the form of a trenhep and turning disk is on the bottom
+all the communities of communes are editor
+the tana mintohana a dorsh is a girl from a woman
+she is very solely in the water of the surveillances in angos
+we can find it in a bottle of water the small skyscrapers
+what she did in the educational and organizational zones
+the apartment to king hans blak
+she is also equipped with a large perimeter of protection
+one day in the middle of the battle cario tates fear and deserted
+the village of sad etien sers is surrounded by three communes of ruminant court poits and posersr
+it can be found near terpathia
+the translator tates a freedom with the latin text
+the suicides are made of a few sicidates
+the fifam got into marriage with the chatelon family by erd
+he is also the author of a few wants of society
+the cult room is located in the first floor
+when the fire took place the three cannons pulled out
+this fixed agreement will be the total of the paid work for the variable research
+each of the three chambers has several funeral or even cookums
+he is specialized in allowing a dolf fiddler to hold a speech
+despite the prejudices and only male accuracy
+to this date she merged with san erand de betan
+the crew's entire crew is captured
+contrary to a constant mathematic she implies directly to a great physical pain
+characters fix characters pitical and historical lecounter
+but the principality is invaded and the prince with joseph clement de bavir must exile
+thence to his interpretation this song is of great success
+she's registered at the inventory of the region of lorraine
+he has not been possible to find a single writing regarding his construction
+she is mistaken and runs the meeting trent's music in ren
+the county of dekolb is a county located in the american state of georgia
+it is a tributary of the polina itself a tributary of the kuskokin river
+a large construction programme must be answering to entangled
+michel richard started in france as a bakery leader
+the concept was defined by egippologist and mus burner trot
+he is often shortened in parallel
+two seasons later the club was promoted to second division
+economic str adjourning sport cultures
+father in garrisy canoma
+philips and john paul morel are knights of arts and letters
+o kan it mainly occupies the second ones
+the academy is made up of a majority of journalists and movie critics
+the island is known as an important area for the conservation of birds
+lindiana is considered a conservative state of republican tendency
+it bears this name after jencup van hartwe
+he is then administrator of hospice and judges at the angers court
+however this hypothesis has quite detractors
+as soon as they saw it is onsans falling apart from the world
+the project was then carried to the senate
+forty stoles of the renaissance presented by richmond's sculptured panels
+he is the final phase in the return matches
+on the right the front door is topped with a lintel
+the stylistic point of stylistics between both composers suggests a relationship
+he predicts this he predicts the associate reactions particularly badoz reactions
+the various works are exhibited from antiquity nowadays
+it was designed for a single railway
+i am going to be removed from the university of nice social in paris
+he is the father of the king of the same year in mayor and mot
+they immediately attracted attention of the spectators and specialists
+she is essentially bordered by individual houses
+this bee is found in europe asia and in the north of africa
+the activity of this stone is very old
+all are named after a german emperor
+egel never recognized that his bivan has been the author of this clanvest publication
+he is a member of the honour of the squad in lafayette
+from such a lovely man settle down little
+his parents are augusta martin tarr and schritzen and elizabeth joy
+in october he considered yup of engineers for the second consecutive year
+you don't want to know how to make his opponent twice to earn the game
+his grave was paid by their fretegos
+dominican chaplain of austria regent do france
+he remained in public life until the first storation
+the risks of psychiatrics are decreasing currently diminishing with the new lazarus
+he won the total of seven victories during this period
+he is the location of a pilgrimage
+de rote was named after edwin lekeim
+she then leaves to live in the united states and founded the she then leaves to live in the united states and founded the lagen wengipres
+the abilitation conducted researches to retrieve some functions of doctrine
+by definition the press drawing is a view of the news
+the sun starts as a ballad post in vericom
+on each of them the player is incombative with a new extra terrestrial ace
+probably greek he is rent and recommended in many italian provinces
+its western sides have a large number of outdoors
+to the east she climbs the charond limousine with the communes of the comunes of roxels
+the napiafia celebration each year
+it is located in the extension of theatre ospitill
+the biis is mostly written with a cerilic alphabet
+it meets towards rio kilima and utako between attitude
+currently the mult ethnic expression is developing in the art of cultures
+he forms the spakali to deny the force
+he continues to reside in athshala
+this species is named in honor of rashur or hagldki
+the municipality is located from sa manild and charlone and champagne
+the central part takes up the marian plan
+the doctors degree receive a nuance on royal sevention
+the equivalent is the republican of the rigisy is unterior
+the wife of angeli and their mother as their youngest child dies
+a street is named after him in paris and in boulogne
+the banvel geographically belongs to the silaments but is attached to the papacy
+it serves as a military hospital during the rest of the conflict
+a vice president is elected in his saint
+the most still in a state of flying
+this is how this street is named after the province of poito
+the realization of the street took place in two phases
+the site is used for climbing
+the soft ware for animation of the company linda
+over these different denominations the club changes as much as legative
+he is put back in retirement by bourbons after his hundred days
+he has eight pieces
+she registered at the zagi club the town where she was born
+the current governor is colson
+he reorganized the church he reorganized the church's children and helps to the pacification of the country
+the rules of the game were marie abreast now malcalo
+the bells are belled in austria
+the commander of this decoration is probably the lord of marpellas
+the city is almost michelline between lestgars and lowlan
+his back in the end nedabal fusha is promoted to his place
+his wine is conditioned in a hundred bottles
+chrissand bal ssentially divited scrotuously
+during during eruption a hardan can be shouted
+she got it as well as for her sister
+the star of david symbols from the jewish religion consists of two opposing triangles
+he approached the events of the current georipe
+similarly they were counting his industrial area zumb an we
+at eleven he began to make a bellow trair
+he lived near the coasts
+he was regularly a jittin and boydo whisky
+the left forst was produced in two different versions
+he filled up during several years the loreging in ireland
+the most famous and the most important belong to the sumtary and nika group
+the haftar forces also serve the city of moors tened by the tus
+chummens only took a final decision will be made
+the junta is the government of the ladas
+in particular it is kept in conserving the history of the slave
+there is usually no smell of the vengeance
+the local economy is based on agriculture and live stock
+this one is fed by several lakes and rivers into the kenons river
+access to the second gallery is done by three internal stairs
+a certain number of particularities is signalled
+he is the first filippino american and elected to the united states
+this text was significantly revealed and enriched
+sate the brother of willenfugel was born in dress in germany
+according to aggrabat du to pantony knew a question of questioning gegarin
+the case has great noise but the problem is still in the past
+the doubles event is won by sanataries and jessica castic
+she's married to jan shelton and two children
+the prince cypril is from her to save hr by strength
+during the execution between the two parties of loitorio placed in sermon
+the investigation will be able to get out of the experience and the workers while making the new ideas
+these elections were won by bizariel bacord
+later she puts the fire on houses and poisons the woods
+the whole of expressive music of postal characters
+this species is found in europe and ita central asia
+a petitiond the inhabitants of traders and then permitted them as well
+it is located north and the county of vixon
+the shape is very rare or in the high watershed
+the profile of these occupants is quivers and sebactors
+the evening francois died in a car accident
+in seven configurations the crochet described a cardinal dioed
+a rival strip then decides to steal the first and to free the second
+he succeeds urbain first
+they record several lines or not to the film that she turns on
+richi sanbora plays the guitar on his title
+it is found in the area polactic and e earctic area
+after having been chosen by the government dads form a team
+the c e c e gets into a local milu
+it is similar to the kriler baron of the same era
+ma madeles joined him for the convince of being kept er
+the night had come and roden returns to the cave ended up exploration
+cessars rendered young centuration of going to the pisine without any racious rival
+in the northwest was born an araqipa in the set sparrow
+in both cases we think it is a constraint with the topography
+the restoration full of the castle is committed to a master of passionate worker
+in the space godhede's workshop sa gallery in the rue carno at pontevi
+the album receives an excellent critical attack from the specialized press
+this position of principle suffers from exception especially in the middle of the war
+his father opened restaurant
+tarkin is a good statute wind in memory of this project
+in short the pope had no support to continue the hostilities
+added then the salt in the wall
+these names of tredo were also commander of the city of saint louis
+the most famous example is the position of svidder
+in addition he joined the jarville club where he remained seven years old
+the medal is for a long time with erring etiquettes
+the maurice de the maurice defr museum is a private museum
+he thus gave the first tunic of the leader of this race
+however it is not impossible for the station to be de ground
+the church was saved from revolution
+miltar changes when justin falls in love with the first time
+they are young and just all the horizons
+translations are carried out by pegasast
+he was one of the thirteen episodes to have been difficult to this period
+the arrangement of the crossing does not respond to any standard
+there is no victims that are the only ones who are in the middle of the world
+the church has a medical sponsorship in rural villages
+the winner is dismissed at the leal the winner is dismissed at the leal cycling state
+the application was not successful by micro soft
+it meets in chilago in the royal arka cave
+nine phon sbe died in a very large porty
+he is an officer of the order of the academic palms
+these elections are won by the whick party of the duke of grefton
+siminis in transaxania and sold turkish to the chinese
+it is dedicated to the egyptian collection of the vatican
+he has seven selections and a goal in spain
+the episode marks the beginning of the spy and rusela series
+no one took me away with the access to this claim
+installed from a few years to the meeting
+this species is found in tansna and cenia
+the deep valley of the bise the serene mountain range
+gilloric is not a single to have a very well known phenomenon
+each year it is the time when a sun reaches its emetus for this attitude
+the town hall is located in the south of the minami sgaya station
+he is in three terms of economic difficulties in france
+two ships never been completed
+a person with motor killed a man named a sick woman
+a renewal of the historic method happens
+his programming is mainly based on the reality of educational missions
+his government is marked by a support of the order to respect it
+some varied fragments on the marking carried by the detors
+eve at lebonn was the dory of french actresses
+the sensitivity has pain or teguments is exempted from the meadow
+we bring crucifixion practices as a torture in prison
+after literary studies camille ve deloque made three films
+grade mohans reside half grade mohans reside half season and replaced by steve broto
+and also plays gold concerts
+when he arrived at tuca he was nothing more than a hospital forest
+when summers are dry it is the stream and wetlands can be dry
+denimark reaches the quarter finals of competition
+his uncle mark advises him to pay attention to foreigners
+ten mills of the great mull
+the company is financed by the national assembly
+the swimming pool is made of heating even at the most cold in winter
+this capacity of acting is obviously depending on the resources of the provisions
+the double vent is won by ivera benezova and bitani
+she was little after published in french
+during the whild junior world championship he played four matches scoring two goals
+in addition the morpian is done in the color of the skin
+since the beginning of their fossil she fights at the staining line
+he never saw that and proved he never saw that and prove the need to confront their point of view
+it is based in the abbey of saint german de pras with her children and her husband
+he is the son of the former cyclist and journalist robert chaplin
+in the arme military traditions the sultan was still in the centre of the armies
+in this case it is solidary of the aculator
+globaly the critic of the album is positive
+let' let's give up to her husband set return he said
+he is the father of the actor massiester and the artist painter maria nator
+the station of super banner hosts several arrivals from the tour de france
+he is also linked to the independent actions as well as the adoption of the specialist milanesian socialist
+the council itself includes ten members elected by the assembly
+the software industry is in perpetuate evolution
+she polishes the sit before rending of f c a and mobile applications
+he worked as a civilian in demis but also in cairo
+the family home was in tep
+he violently love you
+their small bugs generally not have any more than a few minutes of water
+one of the other main characteristics of the game is the cards are general at the tournet
+he is buried in a convent in the minimum near calyx
+he may participate in the battle of mount fre
+we can see there many pieces in the history of post and assess
+technically performation he is able to perform a specific pass on both sides
+he was employed to reform all the abuses that he had seen it in his diocese
+apet's wife feared white feathers and scarlets olive golden olive
+a tower is built from the ground
+with dry ten at they are the only ministers to the extent
+it meets in the provinces of santiago cuba and ogim
+it allows access to the pant de large street in mos
+it is the reference to care and research and environment
+at an accident he lets a big shake to the face
+he then makes more conclusive effects
+it is the current high school of axpek
+when he is unjusted the handle is blocked
+the most popular among them is undoubtedly
+the teachers know each other on the viadic to correct the collections
+the committee therefore has twenty members
+as often john casabet showed two versions of the fire
+i sign a contract following the veparation of the feason
+from that moment on a milpa became a simple stisis
+these classifications encompasses the roofs of the building as well as its environment
+he is also popular among those who were interested by victoria and historical reconstitution
+he calls the zungo to tell him that mont mayon and the real demons
+its purpose is to reward the best young artists we are in the country
+fifty angles were added to the calculations
+he also made some offers in this area
+he returns to durbin in the club of the royal zalo
+l klance is worn on the two brothers of the offering is for dalke and hisnakes
+the municipality has its territory of the jean satone elementary school
+the tables show the dramas hid the spirit of the child
+she was the parish church of the milinaricourt and valcourt
+he was given the plans in life of the saragamites bridge
+eric syndex cull eric syndex cull is the current director of o v m s t u
+besides he is the french communist party supports african americans of emancipation
+in the middle ages it was found and bognia stops on its copy
+the fishing by mac is traditionally served on the current one
+they are called holidness
+then he discovers a bot the panazola club in the jackson para
+this species is endemic to the west street of this species is endemic to the west street of dilan in napa
+he is buried in the novre deci cemetery
+he even died of his measure in the following day
+the roman refers to institutions and therefore its first name
+it was the club were all young footballers of the bast dreams of playing
+the survivor of the function renel by cretine imposes itself
+it is located at a kilometer north of burke
+he is a goal keeper at the pono vorsa
+he later obtained the title of county denmark was a duse empire
+this one is now found at the national gallery of pag
+in the north an the senators are elected by the higher nating system
+its natural habitat was the saint lano rivers
+the temple is renovated under the corelian dynasty in shussen
+they are decorated with stars
+paul women who are threatening in their half different eras of japan
+here he is suddenly rich
+instead he was followed and transferred to the field of kowago
+this bridge of the history is located in two local museums in ades
+after the war his teaching touches several thousands of milk in the middle of the conservative catholics
+his daughter kentell received the city of simala
+the passage of the arqisti citar marks seven albums
+thi sickness is quite thi sickness is quite common to see a general sign
+the jesuits are by the first inlanders of the rock mountains
+digital an expert
+the city is secured in waiting for the river of the four children
+they listened to several hundreds of anig they listened to several hundreds of anigawards in gold and glue
+at the second round the title virgin laward succumbs to his temperament
+robert harr robert harr did not know much about his chrimeology
+the municipality was part of the town holl
+she's the lawyer of the colonian group
+fernando amoribieta is a historian al de cotora and ato
+he prevents being an encycle by the chronicle of the latin america
+it is a biforation station with a sulac line in vis consulas
+she still employed at the highest level
+he is also a co founder of the production stark hide
+all these recordings ar on disco columbia was made
+he did not have inesty
+he plays with the admirals of mibroki in the american hocki league
+the archeological pieces are the most important of the museum are presented
+see chapter
+after seven months of management several gods come out of the kmarbi belly
+the title is averagely marched in the united states
+gembotist costa thus gembotist costa thus became a branch of costa's traveller
+the group chose fifteen sons among fifty years old
+the devil's event is won by jenal jakvik and katharine a braveni
+it is part of the free masonry of the french and the french people who are in the same way
+this municipality is only a single and a metro station
+the novel of each set was even given with great resource
+he'd soaked with the chargs of my mind
+she depends on the municipality of salea of the department of ma severa
+the forestry government is in the woods of coldness
+he is considered as one of the fathers of soviet bames
+the album reaches the place of the german rankings
+waterbrig severs
+they agree that part of the non strateti industry must be left to the vote of the century
+he relies on a rocket of four pillars of the caste
+he was trained by the casaku river tributary of the river congo
+it is therefore a force of direct mobilization in case of war
+the visit of the last floor allows to have the panorama of shanga
+it borders in slove to the guests who are isabel rabu with tiriju
+it is also the case of the sirinim cook
+a large part of the managers declared itself on the edge of the sea crisis
+when he comes to the great depression is the emergence of the movies that he finally appeared
+i headquarters is currently owned by leoni cooper of the treve party
+jordan is going to be renamed in the season for the conference
+this species is named after the gerardobiger k this species is named after the gerardobiger kiola
+the zelig is used to provoke walls but sometimes on grounds
+the soils provide a multitude of functions
+bryanbass is a former soldier of the american special forces
+the noise prevents the knowledge
+boredy was born in chicago beno from aacian parents
+it is the key to fate with many resistance parties
+in his objective it is close to antiplisbagle and of lead and to torpe
+his name in the indian name tiro is hakopi
+he then takes the direction of the vadville theatre
+in saint louis they found him born the second day
+all t officials and soldiers were in service
+claire is a former football coach trinidadian now retired
+the documents recognized as well by tajidston are part of the internal law
+green economy activities such as the equatorism are organized there
+ramon and his exploitations serve as an example
+he left behind him a considerable testament in gane
+in france the series was published by sonic
+for a day a miracle happens and there's a symbol
+jorney garcia kendo was born in villereal in the province of castelon
+he played and presided over the racing clup delons
+carne and faroche run to the yellow dressing to the scriptorium
+fernando cruz spends the vast majority of his career in benfica lisbon
+he was built the first railway in the county of saint francois
+the neighbors of the place have been put in place where this wrestling is named
+the two others are located on the edge of the yellowsea
+these publications are regularly sitted in other scientific articles
+its headquarters are in leon currently capital of the rhonaps region
+the ski area is connected to the great domain of morions
+he begins his career by joining comete treman review
+she left a jena for a large sword
+he also described some species of plants
+in this respect there is a coffee
+the brothers calvany and late to arrive
+the liberal union is this fair instance from a sition
+the governor has its own flag the flag of the union with the armories of engilia
+he also filmed in advertising for money
+used for some additional character sa
+the locality of mop is a chief town
+this protein allows vers to infect this protein allows vers to infect a large drop of water
+bano is served by the line on the interbimobility network
+the same year he was elected member of the french academy and succeeded by the duke of alfrester
+its inhabitants are called fresneda
+coran was born in canju and thus koran was born in canju and thus his student
+at the university he studied psychology
+nevertheless some characteristics are getting closer to the majority of the companies
+grostark is a lost principality in the central yuk mountains
+geroche thinks that the roman empire was deeply influenced by the greek traditions
+the sovereignty refers to van act from the next day
+with more difficulty we can build the pathagon regular
+a sequence has been shot which started to death of nan
+his v is also a head office commissioner
+insurrection of november
+later the house was repaired after several times
+they can also boost as in the raid mode
+she crossed the renee panter street and rue charque chumand
+several laws remain in the manuscript
+it is available in the utupa uto
+his prefecture is handa
+richard docking you several times the example of the pasitism of couvere
+he is a general headquarters in the city of giana near the current gene
+the christmas festival is compromised following mister the christmas festival is compromised following mister's acting
+anabian was dancing the fashion
+some one was taking a look at the first gruomer's housekeeper
+but watana took refuge and beat him
+the north vienemy's troops are then at the sandsport
+after a crossing of the police she returned to cherselem
+he gathers about fifty militants
+she's a member of the arabian seaside of creation
+these military forces occupy the country these agents control the administration
+he is in addition a member of the presagery
+it vary according to nature and energy of the lion's runation
+this solid extends far away in venezuela and brazil
+in this case he is injured by iron shot
+he is also similar to age and deaths
+it is located in a few kilometers east of the city of benny
+the following year he is elected general adviser of the kenton of port leuis
+in the middle of the year he has curly trained the indian club in the brasaa
+i have the father of fevers cosete
+he then returned to teaching and research in the romans
+he was also a tiscian and a press director
+there are different types of counter attacks
+she is irrigated by a moisture
+she remained the only one not to be married
+the capsule the capsule is carried by a ped the capsule is carried by a pedisary and has a short upper gilletin beck
+she's the first and before last lady american to be born abroad
+he takes place every year on three or four in the winter
+chasario frequently used these two sisters as models for his paintings
+he started his bed
+the spodges in monstrous are sterilized before its first use
+max rudge entered the wooden cemetery in branx
+however the latter was not very popular during this period
+it can be found in the mountains
+finally the great number of augest is also an artist
+males on the other hand are fertile with february to november
+they will have seven children together
+she lived in the state when he became the first woman a lawyer kia
+used in this context he looks like a guest house arrest
+the diter is considered a specific form of interlerence
+for the next season she joined tarp
+most of these incomes are benefited from operational benefits
+she works since then anit independent and collaborates on the review of varia
+at that time senkiku kakigawa has a fortified village of the imagawa clan
+the state of hio reformed his sentence in perpetuity
+he is the youngest of the four children of samuel and sophia owitz
+this municipality is served by the departmental road and is crossed by rome
+the bell tower is topped by a bolbran
+she enters the nicaragua
+she's a national free mocrat
+his second wife was adrian enderson
+the feminine form corresponding to johanna
+he honored his first selection in the match against the team
+the pa the passier street is a public road located in paris
+none of these descendants brought the title of a rehabilitated stop
+the village is the most far westan navy
+cristo lexendel is also found in the same movement
+there he is hospitalized for superculosis and anemmonia
+finally he was director for television
+she's the daughter of james gord kingdeer and elizabeth wilson mitz
+he dies at the age of ninety two
+yes and beyond
+he is oriented in the current museums
+he prefers to modernize the vocabulary
+he was used in railway crossings and other general constructions
+he throws five rounds and is credited with the victory
+it can be a mistake
+she also participates in competitions of ju flun in burton
+on the wheel of his arms nobody is rising the dreams of the extraordinary personal killers
+it is a very big figorous rose that should not be cut too short
+the first episode an is a second one were reserved
+the owner of the wind then attacked the owner of the wind then attacked the commune of damage and interest
+he was supplanted by lawrence bacclini during the first parties in the evening
+the card o can follow is also valid to pay for the trips
+george pe was born in a villa in breston in the province of
+he was also a consultant doctor of viceroy and his family
+some repressions are taking hundreds of death
+like these structures the eves were established outside the city
+his work as a composer and interpreter was however devoted to harmonis co
+his wife who completed her last works wrote his mechology
+he was the companion of katherine grichen
+he succeeded in fleeing to italy at dri king of lombards
+the museums of marseilles are then exceptionally known
+he joined the former players of the bos non vendu and mar cinaski
+it is the adaptation of the new opotamus fair panovac which also sighed the senario
+in other words this posthumous crivicism of fragments would have been replaced
+he was the hurricane of the oldest one to be brought back up so far away in the north
+the distribution is ensured by schools
+but under a painting and a new start is accompanied by a separately
+she meets two men who will change her life
+he continues his secondary studies in kief
+the snow is on its part a particularly controversial activity
+it is the lithwanin runner of donesalu was the final ranking
+primary education is active in franc
+bello bella at thamesis chintz to the netherlands in waldwich
+after this very damaged source he succeeded hennedy
+marry twice only john has no ascendant
+the system allows a prefor station and a quick advantage in self construction
+the restoration cost the laws of his bishopric
+but she has a catholic church depending on the diocese of inan
+fatmasamor is married and has three children
+it allows to connect the valleys of the r and the russian valley
+universal buys the lessons and publishes the complete game
+no explanation was ever increasingly given to thisocation
+the nearest city is fording bridge located approximately one kilometers to the southwest
+they have two children and a nettle
+she was then equipped with a black eye
+the other companies are smaller
+missouri is a main character for foreign affairs
+sometimes only the paser offer to discretion
+they spend their childhood and their tenagers in gerand
+many strategies were scaffolded in the course of a story in fishery bearing
+gellen he tied a bouquet of flowers in his sofa
+the municipality of vadus extends over a part of the principal territory in issi
+since then he has been depending on paed of aster vell casti
+poitevant is made up of a large glass of wine
+it is part of many moderate democrats won by the
+we have seen people who invited on very varied actuality issues
+the documentary places are the sentm and marie rimmers dialect
+in the meantime it is entirely included in the district of grenade
+they were employed for communication and transportation purposes
+he is suspended six weeks
+frends and only finalist teams have won the title
+he plays as a filder in wooded timora
+his action as minister of defence is appreciated and he becomes popular
+the furniture has small eyes and membranous red or purple
+his funding caused a problem with him
+she regularly wrote to dun s
+she takes place at the louvre museum
+it is also a winter ski resort which is also a substantial ski resort
+all the inconditional models of the conservations was unacceptable
+this species is named in honor of fernando vidella
+german politics here is to countercart the french interests and especially british
+it also allows to bring trigon metric formulas
+the original design is miki sudina
+there is no municipal opposition and represented by six selected officials
+this incident led him to the secret of us
+he is replaced by radio and belgian television
+she's in relationship with te strempling station
+the caterpillar is green or brown with a brown spot
+james bagton is then played but with the first notes of trunk
+egypt has two madisin its name
+he visits in summer and then in the united states
+the first collections see the day
+he is also the author of several batteries
+they are common to prove in the presence of a historic tower
+the movie was shot with real and animal skulls
+in this club he is a player and a joint manager
+as associatin doctors
+is largely used in automatic learning
+he directed a doctorate at the nationally superior demins de paris
+the colleague and the members were inaugurated between samshi and the riko
+its presence is uncertain on the island of timor
+he stands out by his good mood
+most of the provinces especially ontario follow the american system
+this columbus fita cuba is the island of the youth
+she discovers floring a company a liben she discovers floring a company a liben and begins in france
+the gitarist is billism who will take the group after this album
+the articles are translated into account in german
+borne from the cross borne from the cross country together to sat sylvestr in the comedian company
+the lifetime style developed by the region is clearly mediterranean
+the roofs of mediaeval were restored by the city of metz
+he is entirely included in the district of mendi
+during the winter he went to the caribbean
+this work concerns the exterior as the building
+lack of success polydore mates the contract
+however the great italian hall is destroyed
+atiscoblane is to say this various functions within the cureroma
+different models of objectives will be produced
+the successor mister de brudoch influenced their decision
+cordulias however doubt serenely as i have to hide things
+he combines italian grimness to the spirit of mullin and humoristic realism
+the canada qualified for the final in the final
+not having a descendant he is the last brewer's named of trock
+he was the second man international to come from the terranean region
+matthew trilla was a nice man in a good voice
+the park has dunes of sand but also a sandy movement
+to day he in satiatric is located on the main place of verona to day he in satiatric is located on the main place of verona piazabra
+wrote de pre de locren for the uven union of paris
+finally the family sutchek accepts the proposition of the redwins
+it takes place in three stages in paris tunis and piley london
+he plays with different clubs in tonesia as the club of africa
+were therefore recurrent to import
+in the south in the southeast the con kitchen influenced chetcon beaches especially in stockfish
+this one will produce the first fill of how
+the second day the wor continues
+she does not participate in the rural championships in baton
+he is director and is the institute of marine mammals and university
+but it's already a small person modest and moves forward
+this species is endemic to the east of central china in papra
+it meets in the cive cave of molina in the valley of bu
+enough he asked him a reflecting six months
+the luckleys disappeared from this new party
+a group red by a lan patio purchases many courses
+father comes from getting down an not planning a higher education
+the manoeuvre of enclosion represented john shforlin the demolition company
+somiers talked about the economic concern of intervening nature
+one evening three men wanted to penetrate my chair and the five orchestras
+the change of enables the time is observed in various situations
+the current spiritual master of samar is camitipito
+the french revolution forced this moderately to this land and then to hide
+the existence of a seminol statue is very likely but it has not been found
+it is a discipline transversal letting people go together
+the preface votes them
+this one was regained
+the general assembly must be mandated the president of the session to sign this contract
+it is located in proximity of the way of the railway
+he is not what a season is the other's home
+hubet wen hubet widows and girmis unapproves the idea
+moreover he is announced that the orientity will be dismissed if he loses against himself
+director and photographer she is an independent
+these two forms of the capitalism are both extremely real
+asalim the castle was destroyed
+the prosonous phase is the internal brain
+it is placed under it is placed under the edges of the danish federation of tennis
+the media confirmed that they were going well
+the only camp in the vartegal was the extermination camp of the channel
+due to the political context many presidents and secretary of the commission are absent
+it is the first animated film that the artist won for the results
+the church is a former roman chapel of a benedict the church is a former roman chapel of a benedictine priory dependent on the abbey of clarac
+paris is at the heart of the community of brand
+he follows the causes of the academy of fine arts of balgade which he graduated from a graduate
+a total of theropist inup must be confirmed in the army of mexico
+lowis royer is the son of a tailor of stone
+these captors help navigation and mobile manipulation
+he was defeated and arrested
+in using all these data the xen applications can refer to any action
+species of this kind are found in arizona alifornia
+spiritually the commune was dependent on the deanery of vitel dices oto
+the end of the tensional of the way is bordering the town of the end of the tensional of the way is bordering the town of villiers lenence
+we are coming back to the initiation theme of the first time in the middle of the country
+he studied at the decorative arts of strusterg at the claude le point academy
+he obtained the seven selections in the french team
+he represented canada in the youth team
+as for all curves of circula thet developed it from the carnivoid
+he dispersed himself where he adects new commissioners
+his beautiful mother is also sick and the family returns to itaville
+during the first world war raymond martin was a nurse on the front
+located in croatia it peques an alvitude
+it is a friendly cal order in the newspaper
+it gives the possibility to regroup several versions and variants of the same story
+the other prisoners are considered as piratry and therefore forbidden
+the village hosts af football club
+it was associated with the oxford movement at the beginning of the movement
+we find these first successes surrounded by unprecedented titles
+it now gathers a break up of the boll
+its headquarters are located in the international airport steinfield de tuff and hensfield
+this region is named after him it is the municipality of chaco
+his gilt was in hell
+it is linked to the administratively to the municipality of hal mucha
+the marshal du retz was then appointed government of the nations
+they have also collaborated on some books
+he is also a professor of economics at the mohomedsand university of rabat
+frank is a masterpiece of extremely perceived and crepulous person who has the name of irvey gerrard
+she is appreciated for the justice of her interpretations
+his name was suggested and was shorteed by alessandre molbadelli
+she will soon be brought back to her sister
+the main crops are mainly rice palm oil and mania
+on the way we ere extended the dam black an mur cuber
+rivedvi adjoins the eastern region of mendoza
+this system also allows in best control of the rolling company
+the englishmen charged the scots but were rapidly cutt
+one of his first models is an abelle buffet
+from these two massive benches ran from the mountain streams
+the vases have provided are then becoming black by the carbon action
+italia he dedicated himself to research on the luttron palace
+furthermore he is part of the organization committee of corbio
+the shape and decorations vary during the time
+it can be organized around dem and concerned with a social group
+species of this kind are found in asia africa and australia
+ali was finally adopted in the circumstances never cleared
+pampa has a brother younger than ginevoloba who became a poet
+she was the first to sid rido at danish
+the captain ocasky street is a public road located in paris
+campbellton was one of the main production centres of whisky in scotland
+the summit is made up of two pikes connected by a colt of about a large
+several groups play towards terrarism
+hercule hercule furious as he was about to the doors of the city to get revenge
+the choir in the nave had to be built and two campaigns
+their potatoes are rather agrammatic and liepa
+ed esky was born in moscow
+the cheese gets its name from its faers abbey
+the adult has a green lateral line under its front wings
+the prehistorics believe that the gatsu the prehistorics believe that the gatsu live in several countries
+hence forsky is sometimes considered unfinished work
+his father was a ben player maxmillan and his mother was a son
+this release is followed by a new tour of promotions of per album
+there is a big discount at the court of versailles
+it is internationally four sides and leaders in russia
+he dies of a kencer of waters
+she marriag in butis charles leger de luce
+the competition ends with the finalilena osalki in the championship
+at various moments alexander pope rutson and johnson rented his work
+the chapel was built in bricks and stones in new roman style
+elmer is officially classified in the cidez village
+spasica symbals the san of his race appears on several shields
+the airport has one single terminal throughout the year
+active in resistance he animates the party of destiny
+she continues her studies in berlin in drama glasses
+the inquiry is retrieving the bottom of goroma that she had been elevated
+the field is very important in the american rituality
+we can we can we can find it in germany however in germany
+it is located in the department of charenton litium and der severs
+it is the families of fran mado and rodo
+he is sentenced to two years of imprisonment
+the nuns by reflection are often destined training
+he is named after the george tillerwood governor of exas
+it now houses the university building of the astro academy
+this title was associated with thorto impressionism
+a watch mission to observe the observatory
+his physical and size are worthy to play frequently
+he will later become the doctor in theology
+faced with this parliamentr faced with this parlamentrs codesac in the new elections are convosed
+at the same time it bears the national definition of the building and positive buildings
+ands to this technique we are avoiding to lower the beginning of a too brutal way
+she supports the algerian war of the penance
+he confirmed his donations by gimdar
+two children the three of this union john claude and christian
+pony of sports the canamera is ponival
+the locality of kitaco is the chief town of comune
+the next day four chitepills preceded the disarming of the inhabitants
+it is one of the most prominent and generous prizes in the world
+it first employed and put on order inactivity
+agriculture and industry are the sime to day main activities
+it supervises the same period several other works
+they are something to contribute because of their competence
+on the bolts the fungus forms a white pen which it appears
+it is located in the natural state in central america and north america
+the last two of the rankings are relegated in the second division
+he is the training place of the sport association of hain milile
+it is the municipality that has less populated in the region
+they staye in its castle are not visited
+there would then be financial loss
+his sola career has never been successful
+bibby also got a bachelor's degree in theology in the louisian val temena
+the cave is scrubbed and slightly yellow
+he is the last commander of the daoclin
+in this case the ex communicates should be absent by the bishop involved
+their trip is full of furry o mispotential adventures
+felix mendelssohn bartody then had close links with jean rayne family
+eport honestly lived with a baby that was waiting for him
+from this analysis he could propose a specific treatment
+it is located under the saint giles church it is located under the saint giles church is paved
+cutl is located in the southeast of miesk
+the failure of the charlestons agreement marks the end of the constitutional negotiations
+a big country organized by the committee of athletics is named after him
+since that time historians search for these pages
+erlet in gislan is in borin cob
+it remains fairly playd nowadays
+the convention was made up for example
+associations confederate and r independent of the energy work
+at midnight pincion is the floor to mister jean
+the drummer usually used to minist pling is also present
+everthin an be he was having fun to reproduce the portraits of all ends of the book
+the nations want the women single
+he scores the goal of the victory against the real saria
+they are equinatures in the earth and insectivorous
+the only one of the tribe of the wondo is replaced by king of henie
+the wit the witna cyndic force that allows for the lair of the high quarity mister gredlin
+they can be rented in their fully or individual replacement
+it is located in a family of active industrialists in the production of ob
+far is the son of isaac sard and deniels von eusen
+after studying in help he was pastor in controversy
+the dean of the chapter is well made of place to the sa
+the whole was then a form of horseback riding
+but one day their daughter emili found them in the others dams
+these ladies are often associated with conventional deposits
+the club of tanchi pullers is the title holder
+she's marry to the australian player of squash bratbardin
+non selected during the draft he is recruited by stilles
+it serves as a servant in the great areas
+originally from the end dulac he curtly lives in the ful ante campaign
+the grains despite their small sizes can provide a good flour
+elaspers made his studies in george herbeau a private school in the edempourg
+jerd ors are also a poet
+she's the miss caraga and the other one in the united states
+the idea is a lawyer of profession
+the evening of the ball arrived at the end of the season
+ducerfs ardunan du theatre and the french school
+and took note in the event but finished fourteen
+the species of the ladier chases soft and small scales
+number ten a boarding veldski to the bosby gesn
+for example in the difference of salary
+jon rakin was resolved from his beautiful family to perform this foundry
+but she meets peter and his father
+there is for them
+he is then a member of the cultural fair commission
+a ma albert is part of the national management of this new ganization
+in a moment these daily groups are sold copies
+its director was actually jacques de latelle
+the hovzml is generally tolerated
+a letter from casta was discovered in the house of shalteso
+noter was proposed by his predecessor perchard
+if they were living in his name the four kings and four parishes
+and the voids and works to form the mine
+if the world sesi camp himself who reveals the truth
+the chapel is located in luca dettangoration
+this species is named in the union of terdandi
+he hates jazin extends to an archduke
+but no one knows she sings with a big number
+the worse is not left but the second world war is in charge of
+the arimodalus welcomes its avliance to the cathir site
+on the right side of the table you will see more than a scarce of the harques that is a good mate
+the sun as symbolizes the south of africa and sport of new life
+issues a flitilating as he sails as a wind's traction
+she was born in pratunia and was married to the neser of marto in the har chepler
+and then the castle remained by women whose same familys
+she had continued appear under the compassion like the marquisis
+each member of parmen does not have a single floor as a high school
+from longer or thes de hots inroded on the square
+fisher street often visits trotsky in exile
+she's also a woman of grace for the very venerable order of saint john
+he enters the game to ivan slugano
+caken cakenan and calsiam are other than the same
+in the first part the choice of malvina appears inexplicable to m
+thus it allows its baby to experience the island of anepaten
+he has services and security equipment for the most recent weapons
+he separates the island of texel in the portuary city of elder
+touching the other cheek he left the match on a severe
+however it is also at this time that the first preservation efforts are undertaken
+they are called in french bulb
+by reward the camerman received the fill report that he wanted
+this type of non padiative transition can lead to the fake forest
+dunkan felt the need to establish a hospital where we could put
+the sound track of the movie uses many union standards
+the little village of vines is of good shape and good health
+the universe has put the rest in the bonus of the practices of the island
+but i can but i can but i can wait for a bit too long time
+he prefers to consider the person of christ in the resurrection light
+it is a sign that is a collection theme
+in winter he is used as chile for scatters
+numbers horse ridions also appear to a tri mestral rhydm
+this will allow us to new features and best performances
+see n
+hamilton returns to change in it hamilton returns to change in it hamilton returns to change in it only in the past
+the mexico market is now a significant butcher for french products
+this cindrum is secondary to a resistance to this latter
+a few years later the two young people got married
+the villages are planted at low by cross
+the social issues related to economic transformation have found different treatments
+the form spreads in the population and within the resistance groups
+krik krikin lives in an economic impact that is important
+these measures have as goal to grant the impartiality of the function
+he as oscar found himself in the presom with a group of malchars
+the priotary sectors are education health and human development
+the lighthouse is managed by marina milita
+consequently water is not a good sifi
+this same year he travelled to the united states and mexico
+this river flows southeast in the forests and mountainous regions
+he is the father of the cyclist in his hasen
+cassie runs after the bus in vain
+crommell also receives he embassy of the hepsburgs for the falter of his new position
+antoine garnier the master of the house is lost of there is a journalist of embers
+but his power collapsed in a few years
+his wife is nutha
+mariental is first of basilique notre dame de mariental and
+larus is located on the eastern side of the building and the type of
+however he became evident that the group's style would be different
+this withdrawal is the disappearance of the white russians in europe
+he has a new chance to fight
+he addresses a proletarian public who wants to light up conscience
+he teaches students like anbutist finne
+olin ducuni says he does not exist a ramsa
+he also had a son with sarah eddy
+these events mark the life of the young frenchwoman of baite
+the work of edish is mainly the greek epigraph and the indian history of alantic era
+he owes his name to the city of tourney in belgium
+the protesters were scattered by the gendermo nationaly who put in detention several people
+the singles were released from different dates over the five years
+it is the second edition that will be held in europe
+sheis created by kmiga otinga at darstat
+the viticulture with the application coto du loir
+the surface of the comune is then destined to repulse three steps
+agency to identify the genele several children per year
+its president is the snuhman
+many were also organized in the year
+cruiser formed locally by english on spitfair
+it isn' it doesn' it doesn't seem to be the head of the narration
+the same day at the alana rival at cdeloja kilja and alinta to
+it is mostly known for having created the tid seres for children
+two other villages are apart of the pietan and goterville sapality
+i is a great popular success
+the eg of the funerals
+his daughter the abbot feson is also an artist
+he is a professor of the anatome comparative royal corus and surgion
+borkingin a film ore inspired by the myth of russian
+the main school is the city of the city and the courts field school
+he supports the compilation for several years in italy and chimet
+he is buried in the notre dame dumont car mill church in vermont
+previously he studied first in order to dedicate to the republic of jibudi
+the history of sika had a deep impact on these years and it was there
+nowadays the order is registered as a cultural property
+johin de was born in grimland belgium
+helections were the elections were the elections were manage to get a majority
+i did not know how to answer you
+nobanc labored and gave him a position of semura
+if only he could have a second chince
+in one of them is the mausoleum of rail
+he participates in the battle of jackson in the valley campaign
+the entranc ecloido box was integrated in the middle of the hiss
+olligan olligan's opinion is a tax pagerd impotence of femble
+yet and algermons were presented in the favour of the algoma
+the srche in judes and he's a little bit so he stopped him
+however linda is known as tennia
+it is the time of the elt's expension
+the brucpabius mainly attacked the seeds of the plant that helperates
+the short movie ends with thousands of klons merchants outside the castle
+young men he studied with simarque with j young men he studied with simarque with joon in penton
+he is regularly the french voice of terror an state and james cribe
+the movement turns into armed insurrection then in a civil war
+he decorates the palace chamber of western art
+they are collectively exhibiting them in november in great palaces in paris
+these events put in question are still in snow board
+he has three children who didn't reign
+one pair of blows is also a red eyebrows
+he was loaned to him in particular to the faculty of transmitting the letter
+dense isolated she does not have a ball
+he is named after the spanish business in bolivia and peru
+pointer getting up threateningly
+l also had to meet another communist leader in shinha
+the aerapoit is not served by the eighth of the daylight
+o ous gave the death by accident to the same way to be executed
+its roots are created in the blues in the
+the bishop still determined to have his project to be carried out
+the menue marcel is also attacked in french state
+the communal school has a kindergarten and a deliveret place
+so little is the result the hypothesis so little is the result the hypothesis can be confirmed
+he is located in the east southeast of carter busel
+the scene intrusted with the animation of disney channel
+everywhere in france the federal insurrections are kate
+the locality of a detu is ta chief town of the commune
+there is also a series of roads alternating
+the activity is the economic activity of the country proves mainly dominated by agriculture and tourism
+the speed of a creature varies from very slow to very fast
+his aspect is yellowish the hitro and crimbly
+he is then successively a lecture
+he plays in simisargia
+he is known for his work in logic and foundation of the matocks
+each year a delegation will be made in moselle and then the news of the town
+the style force on prodigies the style force on prodigies is victory is home to rampint grandedor
+this group is cut by its cluk to free a place in effective place
+at the heart of abos in firteen he was also a t v production
+the title is made up of a hit by about fifty people
+the colloquium was directed by a specialized on the state
+the verso is identical with the nerves underlined by sales
+he was the sponsor of the luther tuzzy family of blien
+in particular they imply the conjecture of dilly current
+the first to receive are the supersidnats of the orient class
+it is found in u it is found in un it is found in yuna ston in the sensi in china
+regarding economic matter buck is a supporter of the economic liberalism
+it is dedicated to the male bernard de clarvaux
+he represents a lot of people and the city halltaland
+he is still a service to day
+the lower part of the body is marked with red spots
+the terrible year
+they agreed to hire their main photographers at enti
+these preferred spores or basca bolantinus
+it was located in the city of muscan in the state of michigan
+it is one of the three boleans still in the department of yon
+the leader of the government is the first bruno chippola
+only the catechism and part of the cult were made in astonian
+few things are known about his youth
+his daughter osa inherits the father of his temperament
+in bologna the properties of the tivlio became famous and the tivolo palace o destroyed
+this site contains four tombs of the dynasty a comminid and seven of the six hundreds
+it is the first documentary and the last feature film of directors
+as baseville is twin with buntrina in the county of dingle in ireland
+the village is located between grelin burg and adric
+thus sa pierre found his cultural vocation
+book has its own side made the comparative system with automatic boxes
+even the unions in sant etienne
+indigencies will retire
+the couple slowly in english
+the members are thin and strong
+aprestive usive of the dwarf is double
+the monographic pollution allows to observe the fuso phones especially during the second light
+thus formed the village and the monastery the small town of saint emond
+on march brook nevin and jimmy's son completes the distribution
+marian and gubby is the new president
+its unimaged pencroft asked
+the author wants to wide the canadian psycho analysis all the way of the conscience
+he meets ler bel laca in a city of another city in sitagus
+the williams will compete for the first time in the final of te grentulum
+the lower terraces invulnerable shelter the dependences
+alcove in the south wall is said to be unique
+she doesn't she doesn't clip the big mallion affair
+it is indeed the first class castle in the hatori of third groups
+the teeth of these megalocera serve the specification of the species
+he was born in neads mill in the kenton of saga
+the famous reef was found by john poljons
+the national theatre headquarters was located in the rat de pontador in corese
+the neighbouring church is a winter and heated church
+joa ferano also doubled him three times
+it is how he became royal notary in his village of nimne on the brack
+the youn must face certain difficulty on the path
+the other mammals indigenous to the island are mary
+friens has more than communes friends has more than communes as much as mayo's
+this square tower has two basin a bell tower
+his oragery is still intact
+the intersection theorem is also not a question of interest
+this animal is endemic california in the united states
+in photography two methods of correction of the distortion are available
+the cocket of the engines is reserved for being blinked against the draws from the ground
+the critical television is the type example
+his domain of research concerns syncratical religions
+this contest made some groups such as cane chatlin and fresh percy
+it covers the archipelago of the islands of gelpagos
+the municipality is managed by a mayor and six selected for four years
+it is also one of the biggest species of shahor in the world
+he is named after the honor of george washington
+united in the sacristy of port du
+riku explains his conception in so the verse
+they were progressively destroyed from the liege revolution
+wewe'll succeed by all
+and the defeat is prohibited
+two days later four french people started in the event of hart break
+she died five days after her entry into formal function
+when a school is limited to the ablissment happens
+the generation of revenue is a challenge for social media
+it is also in the great almost perfect dusk or white
+jean bauvoir speaks french as well as english
+the savour and his hermatic taste the pup is of red orange colour
+maybe it is a good old group that is related to the course of time
+his studies generally have a pirs
+most often it is a new construction after total evolution
+this tone was reduced then half months after three weeks and crazy
+gilline of less corbiers is a local commune in the ud and parisites
+der der antisco will begin his military career as a lieutenant
+its inhabitants are called the long champor
+he is the uncle of claude benett the biographer of runsed
+she notably supported the training of young aeronautics
+saint john sir resus is part of the bres
+he tried to show the importance d necessary for the christian religion
+the dinma the din mator is one of the three emblematic skies of shinh
+but several interruptions were provoked because of the internal conflicts
+he is the father of jean etien johani mesia painter
+she is managed by the city of nagahama
+he hibernates a caterpillar stage
+lands that are belonging to the benedictine abbey of saint german de praise
+for example the artist for example the artist's brother is relaxed by salvadorin musa passing
+the cathedral of sa honor de lavort bears this name
+the giants then made the woncats of mutton in the semi final
+nenan is located at apatapra in the savan region
+the white is bordered with blue in the cloth
+unfortunately he gave up before the end of the rilly
+he dies in an aircraft accident
+she then starts a journalist and a free linzer in london
+the name still refers to a city and the province that surrounded it
+this classification is made up of the participating teams in various rankings
+she formed the career of the existence of light of hope
+his leader was asainse
+the cities of athana denocrates later trought pompey just in park
+the artillery was free the staff could take refuge in the home traverses
+little parisian hus little parisian husin then left gorone in the morning of the chateau
+by the way the years he became one of the closest of the charley
+the governor lochward then left for the academy and died
+twelve years later san domes and the two lovers are all killed
+its population is estimated to a hundred individuals
+new this technique doesn't new this technique doesn't like any week in a hundred small living rooms
+he's doing well at the same time the cowards hairdressers did not go
+some of the craftsmen settled on the roadside
+both are collected by prits to be sold as algers
+and then she was expably degraded by the human exploitation inhabited by the greater speed
+he followed a season marked between malson's and the regions
+very short in his principle the text contains a long list of exception
+charlie's fezzer grew up in the countryside of the allis
+it allows for down loading by the intonete produced by partor studios
+his parents ind storekee his parents and storekeepers faced with difficult month ends
+he received his degree with distinction
+the building started in may with the roads of the road
+none of the two was released in french
+in four of his states the referendum has been held on two days
+on the spot they make the most recent discovery of the history of humanity
+churches in roman and orthodox church were not invited
+despite the sending of a butt despite the sending of a butt in several search hours he is not found
+the situation lasted for ten years
+from the island to the island is maintained in santa monica california
+they are therefore not understood the others they are therefore not understood the others rigt
+after his release he goes to portugal
+sh reproduction is ovibarous
+she realizes in the carnivan region
+the album is composed only by the retake
+if the brother twins were
+the future is toxic in the middle of the century
+he was the johan bernard
+the album will be recorded in california more precisely in hollywood
+selected by his federation he finished sixt in the race
+he was the friend and collaborator of gen renoir and renee claire
+i make sure to force the door to go back to spectacularly by the window
+sad sad is represented in general with a star above the head
+maxim is then on the side of his ambitions to marry elic
+it is located west of the burma mountains
+his spectrum is however close to the ladder
+he also defends his brother kermishe who had acquitted him in the ferrar affair
+this species is named after the nicola and mialori
+nicky is undoubtedly a daughter of her cousin siisoml
+he is one of the first acts of japanesism in belgium
+these weirdos you have the same time that shakes
+some developments are spectacular like those of the communes of fansords or satany
+he scapes three or four in paris and their daughter
+it is sometimes considered to be antimedent
+we can find it in pakistan india burma and north australia
+the teams are too late they use them because they are very popular
+the municipality is bordered by the rhone to the north
+the number of models characterizes the depth of the line
+these folls are on the list of sites an classified monuments
+if a flight is subject to several regulations we can generalize the most indicator
+among the me sexuals wernet did a distralization
+located at the va street rivenville in the country of c
+the population is numerous with the constructions and the austrians
+the napier family will remain in possession of the domain for five generations
+when was the main objective of the british forces to their landing
+the titles are mixed by maxton
+demlis was a marxist sociologist
+his sale price varies from two s more with seran
+leonard de vince is preparing the rank of being part of it
+the father of the essays created thereafter in the colluation a special branch
+he is under the modernest style
+ma mat eccense majou peres
+she's the most medinal municipality of the meenburg pomerania occident
+he goes under the command of his own nephew yorimakim of arcastad
+the law is raising negative reactions in quebec and abroad
+he is the brother of writers he is the brother of writers holdenhac
+on the left there is the royal army of keeping the pa transend
+the couple had two sons and a daughter
+the inhabitants of leech forma commune
+all are threatened with extinction
+the main industrial company of the city is a wooden eber factory
+this episode was sold in the united states as well as the united kingdom
+burgel has also been ambassador of several associations
+the mountain is part of the same time as vladimir circuvan penie or fidorco
+the new wall hits all communication between the two little ones and the border
+it is located ad it is located in the borough of the arango district
+but here and imaginary have the litical link to the attraction
+between the movies at the top is also shot by the band
+the fort had a museum and also welcomes historical institution
+according to the code was then preserved as the name of the soft wer
+we can access the fountain through a stairs at the bottom of the current street saint ellie
+he is named colonel in recognition of his states the service
+ah he became famous and met the greatest musicians of his time
+i'm not looking for replacing michael stratton
+three years later he resumed his studies
+this one chose in a tent from the king of bahia inwe
+the music is talking about itself
+founded in china
+these food are mainly made of edible fruits blackened seeds
+chilly is ren third in the competition
+he plays six matches during the olympic tournament organized in greece
+do you want to meet in several length the top ho will also indicate the place recommended by the relays
+it is one of the biggest and most naive hamlets
+a company was created in visual effects for fill makers
+the museum of larina organizes the international tryings of nineteen years
+several adaptations in the world are published by korancha and ichiga
+he will attack in brazil
+a year later that he is named archbishop of mevok in the wise conset
+he is a project that was launched and forced to be a huge contract
+it is the largest city in the district of rinca
+the second world war ended in the elenation of the todies
+and toimbers of the house are not used to sell masters
+ddimerd lake was partially located on its territory
+it's a digestive from a seaside to the amende
+septem piro continued john howard
+he is not impossible for to be an unprecedented day
+the quality of his work is still unknown
+the passage for pidden passes over the upper side
+she crosses a magnificent gorge
+he's in fact an active subject of the main ones the same one
+he is one of the most ancient century in pote
+he is a liberal party
+he was a judge at the very first village of commerce
+he is part of the poor timunicipality
+during the draft he was chosen by the sacramento fentisis
+they are often considered to be the most particular traces of the canadian english
+however the notes collected are very good and the music have been animous
+in any case this is the most reduced version that was kept
+two days later he signed up with the complaints of cleveland
+i washed my hands with such an extreme love and grew up and drew in er edleir
+the separation between the two and third one floor has not been made
+go on the series quickly becomes a commercial success and made of his master actors
+he follows the government to miserie
+these three previous albums are considered as references by his lands
+sheis buried in othon in the most strict entimaty
+it is also the other slope beyond a crete
+she represents two horses lent the other mare
+this one according to the legend but only the elevation of a new castle
+it is the first time that the supreme court provokes many interests
+never a move he smiled and barely seived his san
+he is opposed to verlin's anselica
+the manuscript is still placed in a category of turtland
+there were some mills there which were some equally in the middle of the world
+he is also equipped with he bedka transponder
+it also includes a planosphere that is the work of the sculptor vense
+religious franciscan he succeeded giome de la roche tandi following the ennes of vedin
+singing in singing in lyri e is already installed in the oliem delidaho
+it is the first time that these cities are represented in the first dutch division
+he is also active an actor
+several veterans such as mor vorman or jusa complete the team
+the white berries are opposed to the bakes who considerit as incarnation of this tube
+a young man he is interested in both ideology and catalism
+he was therefore often seen at this matter as a disabled person
+this allows him to become the fate of the time in the egyptian snima
+he is the director of the jennie civil section
+he is a member of the united party
+native of jamicca robert isame spent his childhood in canada
+this allows for example to define the expedentials
+the realization of dis covers is entrusted to difpel
+the action takes place in the county imaginary of kingdal in the west american midway
+he also collaborates with san severino and his father
+four years later she left political life
+the states may thus the states may thus be included in the erotic behavior
+it is the first italian victory against france of history
+we kept on with an indulgence in this world
+it is a large quaustic torture
+jus wain however had a hintage at the same time
+he offers masters he offers masters in doctrite programs
+the kenton of erva was organized around hervo in the district of parfene
+he finishes in the second place in the general ranking
+some parts of the plant particularly the roots could be used
+he is the elder brother of diminlenco from the flames of calgari
+the canvas is given by the school at the historical museum of the national museum
+the game was published by atne hise
+chuzzles made the majority of his career in la potele vive
+there are large world in groups
+the deputies are elected according to a proportional election by the province
+but the child of inansen informs that he used his own functions
+the massive is covered with a large glacier including pastures more than in austria
+the heart rarely intervenes plays that the heart rarely intervenes plays the acclamations of the crowd
+the street crosses the jacques fremen street the marketat en rue carpo
+the following week he purried in the village
+by default paqet uses pocket for authenticity of user
+he obtained his first selection against ramina
+his pedra is era in the form of shimatar teraverd
+raymond garrier followed his classmate at the seminary of quebec
+card flees from her husband
+she is a tributary of the mahada in the gangi basin
+the album was also certified as paknam in belgium and switzerland
+the mondo is also an obstacle to extent sessions
+several administrative corps had their official journal
+she's located bat the old town of antior
+hijuma is navigable from the mouth wash
+he increased the defences of the castle by adding protected from the entrance
+the cremater is made up of micro particles of mineral origin
+hull and nash went with their third teamat
+from now on it was verest and is mistaken with the natural slope of the field
+still at the tack the city is judged the first three special
+after the liberation he temporarily retreats from political life
+the engineer and the reporter were stood at
+the name of the company comes from the french indist concept
+to the south of mariagori is a member of the most populated part of the island
+cousin of tegilbe he played a key role in the turkish conquests
+this movie has been about the anticherist and the end of the world
+he made a secret mission in catalonia
+frederickson underlined that this sympiosis is an example of nicicologic
+in some places it is also called memor
+it presents the aspect of the facade to the ground
+the cacronic name of macaria is made up of a to macus phase
+a few matches team can be used for some extremely well known games
+this equiity is not verified if the padagon is not simple
+in particular fednos wanted to escape a forced murder in the district of bad air stream
+moreover his day he has no more played in professional hoke
+this extension was previously involved in legal entities
+the table below includes the moments of mercado and took place at interseason
+he then served in control and defence of the purtis pass
+he canons of the chapter sad etienne detroit
+ra is briefly associated with the prime palite movement
+his priory was fast and had the feathering of the feather starred
+this material has slightly over the use of which allows to be used
+pierre alphonse lorand street is the grandson of albert lorand a professor of mimi
+fastener armed in lobes
+the roman tour of b the roman tour of b selican was first built alone and isolated
+the duke of burgundy is a member of the order of all the quarrels
+a simple result of the album is published in the united kingdom by the swinger records label
+the romans will know two other adotations to the movie
+the study of his behavior allows to cycling the development of innovation project
+it is operated by the budapesti public company cos lacedes
+john waynes was born in new york and a second youngest of the middles
+the westminster church hallege is attached to the university of rider
+the garrison of the imperial army is tibet and are placed under their command
+the path then leads to the city of lacubro
+he is a founding member of the order of the dragon
+he also maintains a perfect exchange with rabin ben nori
+like this variable it is not directly absurd
+but philip tebert turned to play this role that was reserved to him
+rabiticle ask sissrins to not accept
+these staffs are approximately international in a thousand administrators
+he has patrick
+it was financed by greenmitch sportzer and built by vincey
+in double mixed pers bedsav and robert howitt took the title
+good lucky is part of the academy of toulouse
+the village is located on the banks of the tuponk river
+ther young children find themselves orphan and are vi familiar
+example each animal wears the specific brand of the owner
+its extremely eastern bank marks the limit of the ospetista
+he is equipped with hundred heads each one talking in a different language
+then his latest success was reduced and eighteen years in prison
+it is the version with the two of them of the hall
+this big form with old turchin did not have a right to say
+she's married to the soprano and elina barlin
+never hear i rice down e boll and on you
+this person this personnel would have smiled is enough
+its origins have each music hall
+he owes his name to isar terra
+drosacica is pie o the cremators conglomeration
+the confrontation was unavoidable
+this represents two daily back trobliations
+irony is often dead and still has been efficent
+we can distinguish several types of censorship
+warts of eric was married and has two sons
+the unline belongs to the new city of mar the unline belongs to the new city of mar ne vale called vila mabe
+confrontations are imputable
+he pulled out a hugerati and the two women have tramped
+towards the evening the ditches still dig into a master
+then he seemed to him that the voice of the secretary became harder
+he was going to wait for the wedding slowly
+we look like a reverend mother
+the pedagon stood up under a sudden
+he raised his name to the forty second degree of the attitude
+imprudence even required to listen to a lot of money
+i stood up and i stood up and observed i stood up and observe the canadian's opponent
+themina bordered the mine of the camilions pamorsky revoils the serzin revole
+in the north stretches the valois prielps and fry borges
+the marriage of savin is described as a happy one
+he is a professor and deserves at the university of princeton
+peter fogg pressed slightly with his fingers with his companion
+florence de fadro is patrician martines who is heading the house
+the origin of the name is still dedated
+a historian is a person who studies at the coconits on the history
+he should among other things go to more and more difficult and magnificent goading for example
+this city may be better known as the first husband of dam and clifford
+he is a member of the brazilian academy of letters
+the inchin class is built by the undian industry
+under the revolution the tribunal led by kladao made many victims
+it was also the first senator federal senator of lillinois with jesse thomas
+during the second world war she enrolls in the resistance
+she opens her private sectors for the plasm activity
+he is the father of randin he is the father of rendin he is the father of rendin
+they dedicated to butnist russian botinist ivan petrovitch kurle
+she takes her name from the city of kia
+esitila sargano studied at the poetic university of teran
+and till night the house was in air
+at least i will understand poetry and i will understand
+what a guard between the left side of the right
+it meets at about altitude on the shiro melaka
+only tell me how the colonel you were in the middle of the world
+but in front of the wet head but in front of the wet had stood up breeding under the covered street
+when we have to deal with a button
+four hundred ninety two three hundred fifty one
+you never told me about it you father to the hardship
+the dying woman was moving
+a door of the dining room held in a square cooked glass door
+the art is made up of a set of independent age values
+you see it we must accomplish his virtue as a powerlessness
+he continued his studies at the peristudent he continued his studies at the peristudent he continued his studies at the peristudent com with a college at brita
+all those who could hurt you my heart disavoid
+i am so thirsty to the wall and you
+its name is still santiago dermade rio negro
+they adapted from traditional music in a more modern style
+she claims independence from any political and non sexual party
+their native bridge is equipped with microphones in order to increase the area of the account
+all that was left his gazine equipment as well as an audiod god's audien
+a decorative title some chargners suggest a crative subject in the court
+it is one of the best known players of the history
+if there is a city is another known painter of this movement
+moreover a handufancher is named the worst defender of the double league
+doko was dubbed an original version by cojiada under his appearance
+he obtained his best results of eight places in canada and huntery
+winter football club completes the podeon
+a title which enables him to participate in the rorald championships in kunna
+sarmasu is located on the line of railroads of the adis bestrida
+however he took part in the second winter academy in the seven germans selection
+he starts the tour american by the master of canada
+the vigetalians quickly avoid this product since he is of animal origin
+he is built on an octagonal base with the same name as the octagonal base
+oh good mother she said to him to morrow i tell him that it's me
+utbut he's trying to share the second one
+this situation will not meet in er
+the regiment was entrusted to a lexispida baron of speld for the priest and reorganization
+dear smith was born in o in yorkshire
+this was five or six years that i was making him became the road
+ter is a group of the four communes trlelarspella and teshoder
+both older than are the trakbts compact gray basalts
+the decrease in the stability comes from the repulsion of columbia
+the two runners make their continue in the descent to the gaspe
+it is now devastate
+she did not have any texts to comment on her cancers
+the commune for his work on the eberumes calculation
+we know gats igilet and other crocodiles we have made them capitulated
+the regiment is attached to the orsgat
+the model is the sero count among the regional and less expensive products
+the final for its part is played in single games on neutral ground
+to day sentina is living in a retirement home pol heic cam wrote his story
+raymond heimes went back to his diocese of marseilles
+hellens also works in norway
+it is a graduate in art and design
+the kenton of pyrm was created for the first time after the french revolution
+his will to obtain the title of painter of mery is very quickly asserted
+it was recorded in the studios averad in london
+that's where the germans look like the fights
+cut off the connections hardening of the blockage and removal of the information about the troubles
+on the fishing point the corner is rend in second biscol cutory
+simus finney was then part of the school staff
+it is generalized in space by the elasticity of freso
+when antonin polos has been interred in eile singer is living in the kelli
+italy wins for the first time this ornament
+he forms as the surface area a real eco system
+they met of the tuin and sang mountains
+yes was very enthusiastic and gave his blessing to the ideas of parch
+their music blends clement ebrok
+during the period of french kingdoms thirteen shakens are leaving
+after the death of his grandfather men capali
+according to bernard debreer he exists two types of comparison
+by his friend the policeman stinisles hotz he learn secret
+later it hosts the ruanda national museum
+after dependance she took the name of laura
+the statue was made of high quality and probably untalistic
+his name is the teaching language or change many times during his history
+the dilegation of breck formed by three years does not win any medal
+she then settled down
+the village has a subway centre
+the producers can go through their products in the local market on a weakly basis
+the game was pure and butsoon
+the game is composed of several missions
+originally the lighthouse was equipped with an anti frenel of five orders
+he takes advantage of it to master the one with the remaining chains while trying to win
+the lower side of the young leaves is red brown in pubecant
+glean bat is the centre of the one shy weller team
+the other main summits are woshiban and rin
+this includes receptions marriage and burial
+it is also sensitive to humility
+he is located in indianapolis and is located in the united states
+he continues to fly with shure and ends his race against houses
+it is directed by cobus
+a third season is now announced
+right sans for the north sander offers two different advantages
+he is managed by the police of new york
+the trace of the child is rased by the policeman stopped in front of the house
+maxtow is the younger brother of the architect bernotat
+the idea of creating a range of clothes comes from tag
+it was in vain for the french entillies
+previously the square was occupied by the convents of jacobin and cordles
+all the partisans of the reversion is a carnival of the road in benipo
+the station was bombarded several times during the first world war
+apple scrip is designed as a script interface to control various programs
+the history of this lake is keys at the history of the dam
+she mainly lives in wet forests where teas are in the tuhis
+the stadium is the property of the university of minnesota where the colden goffer evolves
+they are however taken by the pelaton the favourites in the psarth tillmans
+the following year he received a schullarship from the popular rock foundation
+it made the rise of the islands which form the new archipelago
+the village is located between lhe verdum and snick pray
+desperate they end up assassinating his credit with his pal
+this species is endemic to the zambarian archipelago in the bobaen indanesia
+these muffins are sometimes made with full flour
+linda gople studied law at the university of charbrook
+the bull fight was defeated there during a hundred meters away
+his name comes from his sponsorship the beer cardula
+on the communal territory fishing is practised in the oves
+but the local rains can provoke significant overflows in its casing valley
+this association is the najurian branch of the earth's friends
+it is one of the greatest lakes in europe
+he becomes a priest of the parish of saint giles
+the special episode was also separated as a two parts
+she was done by a chalk of wood
+we can clearly check all this through calculation
+teinauguration church are particularly external
+all the plants are on good lords
+he was ambassador in the islamit republic of pepiston
+despite his celebrity despite his celebrity sa continues to go to this centre of new york
+this place should be semicircular
+unlike the other styles these and chancer practised in a strung and speed
+he gave up the acts of elicits
+but the bushman had no time to understand it
+the war did not prevent him from writing religious tragedy
+he bought the news rifle
+my delegate mother fulfils the functions of the civilian and judicial police offices
+demostins ends of resolving this peace
+mos indignant is still his trummer and the seven wounds were beating on the egypt
+i do not obtain ss
+she's then the first woman to manage autonymous community in spain
+he is taking part in competitions on trackan roats
+he is then the property of a noble family of grand duchy of athuee
+she divorces and re lunched her studies
+he is inspector and computer at the musical college at the creation
+the group continues to take place on stage
+te ter astetics of airqsan gasel sheft
+tn rabn also take several initiatives in the legislative field
+uma sahar studied psychology at the university of sentia in new
+as if for the rest of the series go operates and five companions
+captain stevens gisp captain stevens gisp was temporarily appointed as his replacement
+it is elected mayor of the lamenton at the end of the screw door
+she's president of the political edgar forward
+quickly after the company started to september
+he would have been able to cross this distance in twenty four hours
+he is buried in the collediate church of forcens
+at the end the children revolt and destroy school
+this has deafened the ditches between the chapter and the citizens
+he was discovered at the lin observatory by tincia and conse
+the totality of its biting is used for the irrigation of lazes
+with e start of the good coll to kelsi the last french followers had fallen
+he was discovered on in johannesburg by hendrick vandingden
+its headquarters are located in batterstey
+he participated in the battle of fort radia
+the song contains a lot of un tuk far in williams and ariana grande
+tumult of tumult of the future grooming war against the lords of arda
+but the jacksonians did not take the northeast
+the stems have made the absolute predominance
+members of the people of this one are questioned there
+torchav was born the administrative centre of the numina municipality
+this museum depends on the armort nature park
+the foundation supports the enterprise of the academy of sciences
+he was a member of the siling teams cuta gallery dot com seer nevada bainco
+moni intado spent the twelve years of his reign and reestablished the order
+avuron is a rural department in the same year
+the feature film was a hit at the maxs
+we can then look for some conclusions on this condition
+for these efforts agsburg and obtains the patent of colonel in the regular army
+jerry triche was first a researcher in the rumy martin house laboratory
+this line was once the main line to go from paris surreys
+the church has roman architecture in cradle is a former priory of mountains
+only her benevil was a literature student
+they had the absolute power of the madin super staff in remborg royal
+the church was built in the century
+in saint pierre de roups the battle took place in the curabat districts
+she is also a journalist for the manse letters
+he considers this condition as a sum he considers this condition as a summation of those professional and merital difficulties
+we are in charge of the space station is based on the earth
+it meets in africa
+his work was taken back and perfected by marie theresma
+paul mccartney is at the heart and at the base
+its headquarters are located chevert in veilance
+the movie is staring denish a the movie is staring denish a camotan and menisha
+the group is responsible for the construction of the megator to vary according to episodes
+he was the first secretary general of the united nations
+three actions were published from certain titles of the album
+he also was the mayor of leam by twenty three years
+it stops to function and falls in ruins
+she however established a national record
+i agree to change several times in the federation
+she bets the small part against the pole of the time
+theadis expanded its name in a gasa community
+the month of february was named in honor
+on the elephants there is a large column symbolized by a top of the pome
+the village is located at salin de provence
+he realizes the best result of the territory in front of the liberal erslago
+the staining can also be a light effect
+competitions are held in the first tires to the new competition
+they fly to cuba in mancofier
+the politician is a member of his parents
+brown is a prohibited scene in the united states whell his toured
+he is buried in the family vault of ormond
+this association of concerns linked to the black ones date back to the night of the time
+the araport is named after the former president bibokiba who is born in monastia
+dramis the life of american state of carvin denord
+he is closing in te infescant larva after eight days at several months
+located in anti anetis no longer the sexual character
+it finally goes to granislav marina it finally goes to granislav marinaski and frank
+gen savoy is accredited with the coaching of the earth
+the queen is mooting at love the house
+shortly after this eliminated is the club alexander separated from the devil
+he spent his childhood in terusa
+this species is named after harry harborg cerk
+to day the university continues the development of the garden as well as botanical census work
+it is also served by the internities connecting bordeaux say g to uso
+he then left on several points
+the material conditions are used to explain its departure
+if this version that appeared in france is not a french
+china immediately reacted by salving the progress made
+it's the case of puris centre in the
+it is generally found in conifer forests
+it has a station located on the railroad balgrate and jeva
+dominative also participated in various events of red bull crashed his
+the desise made by andres and dresa petitines
+by elsewhere the law also counts on the state intercommunic and syndicates
+the hermitage hotel is chosen as a cartinal area of the city
+she shared between sydney and los angeles
+their marsel shack their marsel shacks movies roger and fresmand rocher et cetera
+there is also the ilioplasts protoplasts
+the master is only taken during the liturgical ceremonies
+it was the goddess of underground waters and the patro of the wells
+bro bruble is mostly used in trains competitions
+he obtained a reduction in both thirds of the city state
+many players came to italy and switzerland especially land in the roses
+you want to murder it now that he is no longer on you
+hiding it to go
+the ship was passing by amid all the oneser only wrinkled
+this last study does not concern the medical use of malatonia
+catherine dihave won the gerard of the female despair for prformance
+he is married to hedwick who called him eddie he is the grandfather
+its inhabitants are called the arbasola
+lithographies in prints have also been produced
+his memory is about the groups of allegrabile trigonometers
+it is a defeat for jellidan who will not get up
+during this time marge and lisa tried find what to be busy
+but the poisonings are made an a baccasable to installed
+many systems anti air lines were destroyed during the time
+this union doesnt this union doesn't create the planned empire immediately
+the duchy is inherited by the prince of conti
+globaly his decline also marks the end of a period of frenchsan
+to finish the day the astronats deployed sadellites
+he is buried in the municipal cemetery of this city
+with the county of norton he is part of the del marba peninsula
+the city was then stored by the french
+heis also allowed to allow thery to access to the island
+the diocese of carthagine covers the autonomous camelity of the mercy region
+he is the damage telescope during a airtack
+during the first royal war paul did not wish to return to the front esit
+the diosper also participates in the development of the habitat
+she's emerging in mervis
+david hickins oversees the organization
+did he think of the escape
+sopia leaves the country in the middle of the night
+he is the most honorary lighthouse of brangberg and the north of the city
+he also desired lithographical works
+thence to this event the group wins a pity on the scene
+the goal is to make some paintings in the context of the festival
+nevertheless she is the work of alrich noviterite lament
+he is excavated to him
+a conservatory has to have to days for the promotion of this technique
+picknew never had a great success and his marketing was quickly abandoned
+he comes back to israel where he becomes the director of the firm iastutia
+a rip has been short since the corner of the eye above the pulamers
+he is elected deputy of the nobility of the seine echo
+his youngest son is tajawar political and informer player
+this castle is also interested in an architectural point
+he will also be a member of the capisins chapel of the saint jean convent
+he lives in the new york city
+he has never been said for the time
+tarlis and motivated his comrades were coming from the calvanized
+moyert is mostly david collins
+born in florence gicomo prestia studies sinking with sergio catoni
+finally marslus is murdered in strange circumstances before revolving a man
+the painting is in the style of the carriages school of trick
+the rehearse the preliminary the great final takes place at the national level
+this observation is particularly for the judgments carried on the inhabitants of a foreign country
+the primes the primes was considered as activity before anybody
+reconstruction of the seeds in mersion are still flying
+lieutenant of grand lodgery of france he is in charge of agriculture
+he points out the value of the plant
+specialist of criminal law he is a member of the law of university of lorraine
+at the end of the decade the regular connections towards lot were suspended
+it is represented by the yellow
+lither the mayor is coined decided to explain the capsule lither the mayor is coined decided to explain the capsule in the future
+they are the only affluent of the leon river
+the fight was going to the death of one of the two opponents
+each district was administered by a district commissioner named by the grand duke
+just the time to stay he is the beautiful guardian who will make a fortune
+the two members of the couple have been condemned to twelve years of criminal religion
+each region has its omontifs that are transmitting generations to generation
+in some cases the benefits of tickets are given to charity works
+you or currently the fort is used as traditional hostage
+used for their recitations of menters
+although rare a network is outside of the movie can cause him a door
+during a few months the band split up
+he obtained his second selection against the australian team in the second stage
+pastor in england tan amsterdam he was then a professor in alkosen
+they owe their name to their narrow leaves voqeto incurtivated
+the ceiling was made of wood
+the very threatened where out of seas are protected there
+they are cbora inusa mesme macablo
+he was also used as a coastal artillery
+the conflict between france and england had some cultural dimensions
+he is the first to have been elected during three consecutive mandits
+at the beginning of the war he joined the territorial army
+he is located in oric
+a large part of the east of lapeline is a natural reserve
+however he does not believe that liberalism is exempt from any relative turn
+the station is also available on internate
+and they're son of americangist he started steading in brake
+he was the director of the untilda die trick at pemer
+he is finally placed in the reserve
+he visits germany austria and switzerland
+the panteon is visible from the place
+both entered the office and the next april for a single master
+he is a founding member of the foundation of the new latino american
+as soon as the opposition between superviionn as soon as the opposition between supervision and intradition
+it is a classical species that is a plestogum
+the proportion of clay in the stone before alcination conditions disenhydelicity of the cat
+a porter a porter's method could be criticised for their deductive nature
+it his nephew marshal van derwad who succeeded him
+the earth became uninhabitable all the seals having exhausted
+the members of the family were active in the politic slip
+he follows the training of seamstrers at the bogly professional school
+i i'm footing you i'm not digging you up
+the dritang monastery is located insepid
+many of them have english speaking include a summary in the publication language
+this initiative is the sixth against the foreign population
+a significant part of the fake weapons is rejected to this stage of the treatment
+there is finally a siht that lost its design ation
+he even made a special for this fell
+for cultwright the elected one was you shew george
+then she withdrew to the bakes of the rhine to receive reinforcements
+the viewers can play without inscriptions prelab or not be able to apply
+he will be followed by his twin next year
+armies were astounded jusephin asked for secret to the king
+the historian is a member of the c e to a deep reflection
+she decides to enter with the rod davids and takes the lead of the scottish anti
+a reading committee set the program
+zero two condut
+he is known for his coal and the beach of herragley and the village of batma
+this increase is accompanied by the perspective of a new baptism
+we find the church oldtown hall and a small house
+author of civilization he notably intervenes in radio and television
+she can be confirmed to me who already has known her husband
+she's married to the actor giuseppi pompieri
+sir g d banks of the clermont ferret academy
+james claw maxwell is ludwick posman formed his mathematic expression
+he still remains very active in the life of the camp
+later the formila remains the same the file could take various forms
+they also make great petty animals fence to their calm character
+it was named for the virtue of compassion
+first name we would really believe this is a script writer
+it is the tenth championship title of berkine fasso in the history of the club
+it aims to represent the population nuvaja
+he is more than a french ambassador in paris an love
+far from shouting to triumph he thanked god for a certain amount of time
+the crypt has been restored to be inaugurated in the president of the republic
+during this time other ins during this time other inns are acquired
+it is part of the oversighting troubles of the development
+she's the first woman to win a title
+the filling rate was also weak
+the church is located in the french department of jan in the municipality of vinceles
+before all the vows are worth rearranged
+it is one of the oldest roads of communication on the current territory of belgium
+relatively to narcissism he was even defended as the report was entoed
+we were moving forward with passion and confidence
+it is possible to keep energy products through photo botanic installations for several days
+the insurrection is quickly imitated the conjuration of berry is dispersed but the duchess de berry is moved away
+it is the first part of tedipti dedicated to the victim of santa maria
+lena is a daughter of cerfs who becomes orphan after a war
+she gives access to the four rooms of audience
+at the end of the two merchants sheared the house
+he was a boarder of the show
+as well as paris catholic institute
+a shotat number in deposits essentially in an ice region
+sana is part of the tolos academy
+after the disappearance of shadows sunderland joined junson
+the waldomid el shabarby the waldomid el shabarby won his second title against the world firk
+however their interrevition lines became dangers to long and their own people
+seriously injured during the first wild war he lost the use of the right hand man
+this case is one of the favorites of the iflux of the extraterrestrial hypothesis
+patiche specialized in production and distribution of independent films
+indeed the inac indeed the inactivity loss indeed the inactivity loss of sleep and violence increase the subject
+the church is located in the municipality of coin in the ain department
+the success of the album popularizes the single from now on of platinum
+particularly floriverous and adapts virell to the culture in belonging
+she tries to improve her health with a handy pollover without success
+he is to day a coach of the swedish team of pederns
+she has common boarders with pentele stillin bornase
+these refugees come with their religions and their religions
+the castle belongs to the noble family of suria
+the pavilion of hook is the main building
+a group of students is gathering for the new year in a black manner
+its main cause is surgeon
+jup did not accept to be much due to the occupation of german troops
+this carbonis philosopher is among the usuals of which the voice is in the public space
+the skull alote was lost in the middle of the car
+several bus several bussy lines stop
+he is married to a moroccan and father of three children
+the actor erwood mc licker did not perform to be very much fired by this project
+he meets one of them twice a year
+he played as a back player in orlandon baon
+the ramparts are taking place to belt the castrom
+the producer pressus jones worked with several months as a salesman
+the only road that crosses it is the belle montrode bridge
+a buffet in a hydralisis installed in the station
+he left the island with pascal paly at the time of the english retirement
+assemblage which is necessity
+at the same time the memory at the same time the memors broadcast was regularly taking place
+water under the surface that goes on from the edge is
+at that time the seal of lord represents a porie at the reins of the dinner
+herman is ruined and loses the reason
+it was named after the elitus ascadian of the earth
+two mandular glands also tubular glands have amalcanuted color roll
+it was a large man of mindless spirit defender of freedom
+the association has also given its missions to rehearse some pedagoges texts that have become unpublished
+they are always in contrast as a well defined personality
+he's not chased but once captured as a company animal
+ab attended the service of meteoria has always been at the forefront of the technology
+we don't know when he decides to resume this work for completed
+founded by the magill university it is closely associated with this one
+she got a grader in high school votes to war
+he is sometimes chaste to be domesticated but mostly for his flesh
+by his writings amanual influences dramatically peter benois
+she plays as a defender for v f l wolsburg and the german team
+she founded the part of the refugee funin nationali dalaska
+she contains the best ximple in ireland
+he has amassd hers to gray in brazil in agriculture and in the county of asia
+favorable that she never had a public attention
+the multiplier mode is based on the attack or defense
+this sentence remained the motivation of roxen and then white
+they are russian in hyberd mountains
+it is broadcasted since on you to
+in the living room excitement is at its peke
+it is the beginning of a long collaboration with fraends while lore with plastic artists
+a chronicle of adventure and impressions that can be raised
+sante is geniton brown sante is geniton brown and their own catholics
+the self defense rate in terms of weight production is not inflating
+we have to say that joseph the conditions of damage between sran and vessels were particular
+he came back to a year to valencia of agent
+he is the younger brother of jan lesher
+bod crow spent his youth in belleville in new jersey
+she occupies the island of la chever
+the untenors are accumulating the cold an badly as well as the architect by the minister
+their leaves are persistent in the ocean and are more or less
+the latter entrusts the management of his axe to joseph
+thus mevascili their lives
+the following year the new maritime interrignal station is erected in burgovin
+the monks may be the founders of this village
+his father emile edward filler is dentist street de passy in paris
+he became the one for the lady in the oper district
+the cycli compounds can be satered or insituated
+besides the eldest does it besides the eldest does it besides the eldest doesn't say anything that it was taking into account
+the course of the chatanee river crosses the zetnort
+sporting club malicore is made up of two district football teams
+its bottom is flat and is flat
+in the context of a continuous training he studied at the agrian institute of almatia
+it is based in the military manoeuvre of kingston in ontario
+there were four ministers in the history of the country
+he is the eldest of seven children of the teachers in the province of gazo
+he is married to pete who is also mangaca
+he will have more than a thousand portraits of the feather or the sex
+she is also a cousin of henri cabulish and her granddaughter of thomas jefferson
+as often a strong opposition exists among the surrounding populations against this project
+he finished his career where he had started at the national monte vada
+founded in the middle ages by cestercian monks the city has several religious monuments
+he was also close to the engraver a goose lipper
+they are inventurated to historical monuments on an object
+as for the black trail as for the black trail or the centre it symbolizes independence in the future
+the buttresses are in gray ruds
+the company had about fifteen boats under control
+the most frequented tracts are those in the immediate vicinity of the pass
+the seras grandes house the main hydrographic reserve of the province of cordoba
+he is not the equivalent of his lemologist
+systems of investors are possible
+the first minister william and ward ledston is accused of having trained the expeditions
+this village is close to the village where the battle of malplay
+akra starts an addition in order to make new tricks
+most of them were from the great street in germany
+lowis barth avenue is a public road located in paris
+this building is built in the new classical sardan style
+the high the highskas thousands of water
+the whole is covered after grating others layers that will be compensated
+sarbais a subsidiary of the company ajornalin amusements
+he is considered as the first son of music arcor
+his sister juandu kasak vitch was also very connected to the tree
+this experience of conversion is a fundamental element of christianity
+he is a professor of an oriental studies at the university of columbia
+the hull is a key perimeter of the adjustment of a sail
+the laws were adopted and regulating banker activity and investments
+paul watson considers him as a unit around the roman siege
+he flena without heavy losss
+willbelcock is a legendary decision in west american
+however the main theme is love a female look towards society
+as a result relure with truth
+but the young man would turn out to be more sturdy than his manipulators imagined him
+how is this time a bit money tormed for television
+the castle is now in ruins
+shitikar is married and has two children
+the only handicraft was the savagery
+conded endelo in colini then starts talking with the court
+he is accepted in a ductred in his rigel university
+for each time the time of diplomacy and compromise seems to be come
+then he settles in a house on the corpus
+the members of the gun club did not know each other
+he is the father of emmanuel deson
+albertine morin lebreck began studying piano at the age of five
+he is the adviser of the mayor of shenking in china
+he is the first african to receive this prize
+in june he is transferred to the corfus prison where he was locked up in zacarius
+perjan is sentenced to struggling
+we can sete intereptricular tournament
+the verdict asn the verdict doesn't hold back the p m i c e
+there are other applications that like kits
+its greatest successes have been obtained with secretures in the way
+but the marriage had no time to do so a few months and was cancelled
+in studio he uses both a tulegister and offender star kister
+he was a specialist of combine
+the duchy of savoy sends a letter to tun during the period
+saint paul de juret is bordered by nine other municipalities
+the length however is restricted to one or two characters
+the synagogue is the second floor of the movie theatre and the other two are still present
+as soon as it was blenze
+their music is also black medal and pagan metal
+it is at this time that their official poets tried in the era
+ah she said simply without leaving those big black eyes
+it should be noted that this stop is not served by weakens
+nobody was arrested by the police
+he then became a medicine and guided by his father in fast progress
+his brother giuvanni is also a german painter of high level
+he has been governor of the buck province in afghanistan
+he became a knight of the order of the crescent
+he will finish at the kunafeca
+it is one of the examples of biologically of the sex war in sexual conflict
+we can also use this technique for helbert reol spaces
+at fourteen hamsi ansu in a bachelor of arts at twenty one year old
+gabriel gabriel vitch is also a student of levetisty
+the municipality is a former order of the order of saint jean jerusalem
+we will be exposed in the morning the sergeant answered
+he qualifies for the master junior where he reaches the semi finals
+everything is less and less political and more and more hydeological
+the nearest airdrom is that of vitri le francois ve clerc
+since its beginnings selva is considered a great hope of portuguese football
+by his mother he is the brother in law of arthur van denest
+the vir te viratel is usually banagous with child immuno competent
+in the same time he fights for the independence of the british empire
+and that they will be in the bundles in the bundles
+beyond his artistic work his work has an important and documentary importance
+the church ows a building in a single nave prolonged by a semicircular apse
+as far as it was his centre is the granal neighborhood
+at his face a pale greenish brown hair man with black spots
+the orders on the track are indeed place say pe
+the chair is white and steaming jutter with a sweet spot
+his coloration varies from a light brown to greenish brown
+after killing ethy godwin took the sea
+she emigrates with her mother in the whole castide eman
+the city of lenely has a long sport tradition
+radical wrex is a kind of mascot
+he takes advantage of it to form his own country as a cope
+for all thes consonatory logic is a new one
+most of the cresswe in mercedes are found in the norths
+he lives in rome nearly eight years old
+the locals of the stable are located in the city of nova italy
+they can be dragged they can be dragged to the back or lateral
+louis de chellons also left a legitimate child
+the turtle doves back the head
+it can be in some variations shown o dark border
+during these research she revealed little by a conspiracy fan
+this behavior is in ivory is not yet explained by the scientists
+he has a daughter ranja with his wife mariette
+she followed him by a quick tense the ward
+the show was going to continue
+excellent junior player he is excelled as a defensive centre
+the entrance is a rural sipality
+it has green and simple and basal leaves
+it is the radio of the public service he chose
+it is nevertheless only interesting
+it is listed on the complementary list of historical monuments
+chisen had a servant named okiku
+it is a stained work in the kikane company
+the groups brands offer lottery to the tonian of the economics
+john develops interest for blacken she was proclaimed in peckenstein
+the inhabitants of baruba are mainly descendants of german
+the contemporary historians believe that he has exaggerated this description
+some alane treatments are the only ones authorized
+five years later she moved to acadia near saint john de richelieu
+far t v found an orphan in the desert and her own son
+the village of teja is located in the resu valley
+the milady grantry is used as a selection process
+he fruits his small yellow berries around two diameter
+we are made up of the same basic elements he then becomes more and more abstract
+and is it by the way an unexpresstat
+irvin is the aimer and is appointed one of the match
+these ones generate positive flows an
+he was influenced by albert he was influenced by albert kissel calvigo jonesen and impressionists
+many people take part in the use of artificial intelligence techniques
+the municipality of lens is located on the banks of the cantilary sea
+the data is missing and the ray could be underestimated
+it can be heard on the sight
+he plays as a midfielder for an attacker at ternabolvana
+the city bella mountain has geological characteristics
+he currently lives in brussels
+he also published works in the field of physical sciences
+he was traded with the greeks installed in marseilles
+it would have fallen during the taking of pegitios bridge
+it was may be true that because it was not able to adect himself
+great open data depends on the academy of north of two
+in france it is known as the name of frico
+other phenomena such as the eclipse solar eclipse can also allow to detect the solar diameter
+he participates in the activities of the lake patronage and the pioneers of santwin
+the faculty recently announced that she was acow to extend her campus until malagia
+once arrived in karsburg will they come back to the authorities and demand their intervention
+he currently plays as a midfielder in drva county
+it is nevertheless worked on metals that we find it in the art
+the ensit was set up by the architect fan nime
+you were later on vacation besques as the city of lhe rue de arcades
+sitis on the benches of the relating to the north with a black beak
+systems marked can have a strait of temporal order
+the works are published under the collective pseudonym ran hinder
+colonel de la cro de colonel de la croy de legement is the duke of kent
+alas a drama will change the fate of sophy's fate
+dorian liish currently lives in leu conele in the brigney region
+the beginning of the quater are however
+harridan became the beginning of the place of princes of ash
+the series was cancelled immediately
+the lower cost is the hight of the chest and legabe
+he dies during this fight
+a romanofo is a roman speaker citizen of romania or not
+during the lympic tournament organized in the united kingdom he plays a match against bloris
+when he goes to the city he gets interested by two hunters
+david platt wrote connects vladimir david platt wrote connects vladimir jevic pitro vit and killson
+the wood of the tree is used in building and paper works
+willian has already worked as a record for radio
+one of the development factors is implementing a tax system
+he is with paul burnet the co director of thesis of julius
+the two windows spoiled on the floor of belfre are decorated with it
+one of the beautiful girls in the world of the buttend
+he's boring and therefore a cousin of her husband
+he was the father of the minister arol de cris
+he gets in the first place to a few kilometers from the rival line
+this expression was invented by charles peggy
+the visual beef is unchanged but the decoration was reduced to the mates
+he is the son in law of the architectis
+his invention will end up a wider dissemination of books which was impossible to
+furthermore paris hosts large bank offices as the zaro
+he is the father of paul pascal he is the father of paul pascal deputy
+it was a big work
+it is moreover a difficult marine
+she is opposed to those who are against the armis
+from that moment on the works were taken over by the state
+bellisluth won the trophy as a smith
+ernspers is the biggest producer of the prefert house in scandinavia
+they are functioning and totally different since they do not contain mercury
+the limits cannot be reached by the national assembly
+it is also a family
+the school airplane before the local authorities receive it
+he started at landscapes and then painted dead and animal issues
+his passion for the speed is already presoned the descent was his favorite discipline
+its headquarters are located in atlanta
+a column of water and debris damaged the plan
+the municipality is managed by a municipal council and a directly elected mayor
+his films were used for embroidery and decoration of clothes
+she has the status of microrea
+in return the same type of operations at the phoenix tenium university
+the placmetry is socially active between nero and the muscle cellulus
+the requests were examined and the administration was answering under the form of a postility
+duck a school is in college of stressford bears his name
+he must be capable of supporting a mutual and heavy animal
+he is lauriet of doctor
+in the common language the bulges say poll te fruit
+shees a member of the staff helps expert in realistic project
+a christian church is represented on the left and a pagan on the right
+recently the flexible moles and selicona diets
+he inspires an idea
+his name refers to samindger solem
+it is protected in tanisia
+three quarters of the medications are sold to the maximum one more than once more in the movie
+it is also the last episode to be written by stiren muffer
+it is a unique nave with side chapels and transet
+this rivalry grew up during the continental adding
+the church is located in the municipality of muntropus in the an department
+the legend says that he was not doing anything for dementing the rumor
+he plans to lose and decides to ignore the script
+the crown opens on the mouth and perhaps very colorful
+the commercial brand used it was zuki sentona
+the olympic title is then south karea jim junhu
+the particularity of the club is one of its reserves col chin azuri
+the vidia was shot in turty
+graduated in science's person he enrolled in the school of lionel gross
+meanwhile the little girl will understand the fraud of the meal portrait
+to day the sight of saint ludter in amiene is still a friezer for friends
+a game of dem is therefore necessary to dispagate the two teams
+then during the destroyed the river became a sealed and malson
+lucy repute is the father of five daughters and two sons
+two thirds of knife each opening a video given water
+he is found dead a little later by henry pop
+she hates her martinique mission
+nick to the sky was caring captains in the country
+she resides in berlin
+called for the tor muzzn called for the tour muzzin finally returned
+in addition to the canadian nationality he has the italian nationality
+it carved these mountains in the chamte len which he then translated into bronze
+the british cemetery has two soldiers and different countries
+other painters of nenda probably had to have secondary activities in the same joner
+he marries zan de bruk for montpleaser
+it crosses the longest station picor slightly towards northeast
+these furniture and objects are sold to the auction
+the comics trip edited by glena
+he still remains at the command of the troops in the valesa during the ceremony
+he also exists some sense complementary and mutual
+but a new technical problem appears with evan's device
+rumors pretend that it is a great deal
+his work from father richard full just before his death
+the walls were decorated with frescoes
+the united states won this first edition
+the craftsmen in trade were developing
+she's authoritatively in terms of the mod in toilet
+his trip is satellite of artery ulner
+after the ever festival it is said to be in fact a jackson and their group in the united states
+it is twen it is twint with a city of the division a commune of the finesder in brittany
+founder and leader of the stable techinger
+it advises her mother to register her in dense classes
+there they find filer who helped them
+the two brothers established their residence in olo
+the neighbouring communes are dengerben fisher and cosan haws
+maritime grills allow to see italy from the port of paters
+other exhibitions come to be blanding on the game
+austria is in povery
+this laboratory is now the fourth this laboratory is now the forefront of the imagery these works
+the list is currently recommending the members of the thamatical management
+this institution is at the top of the teaching and training system
+back to lozan he shares his career between racil and teaching
+some stations send more than one message by transmission
+while growing up would be interested in politics
+he will be precursor to dedicated exhibition to the painter baltis
+based in hamilton ontario it is the biggest steel producer in canada
+he is a quadruple world champion and triples the urion champion
+it is a shrub that can reach high to the kendulus leaves
+his fenley comes from picardy
+madame radform goes back home ignoring everything that comes from this past
+his action is international and disproportionate from early morning
+only the thirty three first three qualified are allowed for the race
+he plays three seasons with a team
+the rebels are pushed back will they threaten the capital with the french well
+it is a revision of the metaphysics and knowledge
+i tabernacle is classified as a historical monument
+the islands are also populated with lizards and serpents
+these symptoms are present under the new type of shilovena and organizing
+the movie is a reflection on the opposition between the evening and posian
+samuel john carter was born in sepa
+he is mostly used for sound effects or in the scratches
+she travelled she travelled with her box from jewel she disappeared
+he understood the cantons of gologne ciser de perikin and obermas
+the first season is composed of twelve episodes
+the villages are the most important in size their fight in the village of nempe kara
+albu's part of the taloos academy
+the lay out of the company has indeed been short lived in the hexagonal territory
+all this would participate in the fact that this canal is invested
+he specializes in metal and obtains a degree in art
+do you have a head to come
+it is his season and the pride to his career
+the flag itself has not survived
+the chapel is classified as a historical monument
+he becomes a major in the air force of general fredrick bran
+the game is indeed less than towards reflection
+the croy road painting is stepana durand
+a session lasts at two o'clock
+he measures long and in it greatest wit
+anna kovachi was born in a family of teachers
+jimil obtains the belt after herret of the referee to the third round
+the squadron then heads towards tripoli
+he acquired a lot of conpersations a great knowledge of the world
+during the month he signed te atlantic records label
+like kisa is located on the edges of the river tisa
+young man he studied at the academies of the fine arts in budapest and munich
+she will become a founding member the following year
+the country is eliminated
+the authors have inspired by itself for their character
+his logo a mouse moves towards the great speed expresses his ideas
+it is one of the first representatives of this family
+thus she opposes vegetablism and veganism
+he then married just and jidden
+other cities in the world use the p h d principle
+the ventilation at the end of the way is insured by a duck
+the vida is filled with gravel and silence
+it has about fifty members to day
+the women the women's league decides to make volunteers at the off house hospital
+on the beach of the dex river flows into the stream an aracatus
+these maternal grandparents are philipp danjou prince of tarante and catherine de valois
+claire whose name has been a daughter serena keller
+the conditions in which one' judges has reached power to different functions
+we got up from going to turn to this part of the world
+the athena is a ridding gem of freshwater hobby aloos in the man
+we thus we thus have many latins full and usable
+no one weighs as the producer the singer is often the camerman
+banges before studying in hemburg with arnold crodch
+leolinae was buried in these surroundings in baltido and the child was taken away
+moreover many repair works are helpful
+the ozel department is made up of a single community in thed
+five cristo and nibs are geographically part of the von islands
+thus tourism increased on line
+provencal mont friend of the novel of the bellademopanis
+he is next to the stock exchange
+both marriages are ending with a divorce
+his father died during his years of high school
+beautiful star is ripped and is brought back to the warehouse
+the set at a fixed decoration is thrown by a second magic round
+i headquarters are in poit aped a street felixbly
+the plans may be the work of gusta fell
+she's married to nicholas montero a former vizuela actor
+the provinces are divisions of the region
+george will have five brothers and sisters
+he planned to stay in town for summer
+these ones are not possible without limitation
+a transformer is not allowed to produce a polites system from monotasis
+there is some fotin death
+he is a professor at the university of shergok
+the translation propositions are released and integrated to the saver
+the controversy still continues to their current events
+the sealing is made of regulation and must be put according to standards
+he also has several records behind modo
+it is composed by the villages sesea and vinetari
+he is largely known for artistic sene
+i won't cry any more but answered the captain no
+all are modified outside
+she will then change radically or persh in a degraded biosphere
+she goes into her trouble in front of the stands before attacking the speeds ring
+his work was a geometric miser theory which was created the basis of this discipline
+it is named after mary wiserls
+he is married to the astronomer genetitia
+he was written and directed by a designer john cocoto
+this mushroom is not edible and slightly doxic
+she established in dragonn
+he commands the purse funtls through the magine line
+his body is found far from his position
+it is his only participation in the limpic games
+the specific name commemorates the geologist ferdinand hohstater
+alexander sellier is the son of catherine and marcer selier
+at the same time the bril are completely intubilent to death of writings and e igustine
+we could count these corpses amongst themselves
+the traction was built by the swiss company pantony
+he was then considered a priest in the united states
+it represents a large and powerful unit of armored and cavalry
+an expressionist rinin a new objectivity he is also a general
+the municipality is divided into a neighborhood which is spread across different names
+he then teaches mathematics at the university of pavia
+she played the second roles in majority
+they have two daughters linemina
+the vidio uses the stadium no longer in a few seeds
+following this renewal his majority has some changes in the world
+the action takes place in e castle in reniside in huntery
+the district between grandpar and rugin dark is affected
+jimmy wanjamson is in a rural region of mississippi
+there is a number of figures in the figure which is a simple sign
+the roof is kept in the old private hotel the current museum jacuma andre
+sick alexander kremmell goes back to frince and moves to paris
+from the exit of the album was released all titles are working on different sides
+the bellveder on the lower vider is the place was living prince eugene de savoy
+over there he met peter the most experienced boys and solutarry
+no entrepenor is therefore nedified
+in france the institute of currents is represented the permission
+all three of them keep the first in a double penalty
+the sounding adepts usually send corians to a large number of victims
+we're toning about this delicious interbending program
+it is the origin of the first territorial counterparts in the defense
+in addition the totality of the departmental concils will be renewed
+he studied at the ecclesiastical college of borgembres
+it is the second champion of indebsa in the history of the club
+at the end of february he ended up at the position of classical drum
+the architect as glutton defines the buildings and the frence leonore society and bucken is built
+he moved towards the very progressively
+the orders given to the two mattress by the fort lepin were executed with precision
+the sixth one can be called the district of costelion
+his contribution focuses on similarities between ordo and cowards
+she was surrounded by her family and stretches out peacefully without suffering
+the building is now used as a public community
+uttl uses a indirect way to approach the subjective
+the crooks are popular at the end of the summer
+den in a lito style called musso tinterio
+baden worked in several films
+it allows the control of the access and shares of intimate connection
+this kenton is organized around the city of at in the south
+initial installations welcome and welcome a new communal school
+you donkey is the first competition
+we must first be a draughtsman for fabrics
+the close meetings are arranged on two floors
+that's an engineer siencroft
+she grows in her distribution area in most wet soils
+we must knote we must know that the combined language or interpreted
+joseph allanet trained several solo clubs and checks
+he sometimes writes with his wife charlotte peckings
+each fury parish was born and absent from the rheum
+others have become part of canada
+men explored on earth the regions share the same dishes
+the clop starts at the finish of sixteen
+radars passive and to fertility that evolved and degraded
+their main goal is to prevent the digital cost of america films
+it meets at about altitude in the mokar mountains
+the religious wars are of the origin of the castle construction
+at the difference of the latter fsio is passionate about university life
+he never been defeated on a battlefield
+anita haksburg will follow her on weit for the same year
+covered by a forest a midigat is not permanently inhabited
+he has his chief town the village of babo
+the locations of the university of lamagez covers all the territory of the nouvelle academy
+well he appeals pmisioner
+the scientific studies on this practice did not have any more effectiveness
+jengillars announces that he will return to his house
+the team mainly played in the hosted
+this allows us to balance the scores between different levels
+the male event is held in warple wood wimbledon london and the canon of the chenora
+the pons are long wide
+nm twenty one rue tristle can be prepared by condensation of celdids and formality
+it is separated from paso by the close station of saint mesmen deville
+for budgetary reasons it cannot be a limited order of candidates
+smither used to use the pseudonym of tagar
+karlgsika the maternal grandfather of pole is a polish doctor in the prussian army
+the massive is interspersed with leaves
+the villages once again destroyed in the second world war
+she's the daughter of the script writer eve derells
+in the porto there is a large number of people in bronze
+he counts the branches of brard
+many evens do not contain any indication of the lore of derry
+he waits for hostages in trips yes of his victory
+this information usually takes the form of a penates towards the complexity of the model
+and keep all these sounds in their ears
+the movie has played four weeks in the rhome in montevidea
+duchen on villa is however re elected by boandre
+it is found in saba in malajia and in calament and inisa
+a possible follower a possible follower was said to be in the project
+catherine disuaded from sebastine de party promised to make everything he wanted to return
+it meets in the state of carba and the capital district of caracas
+in parallel he started writing
+the local church is located in the french department of butches de rome in the municipality of salant pronounce
+the climate of tankla is temperate continental with four seasons
+it dates back to the lower cretaceous and more precisely the utchen
+all the lyrics of the album are written by ben sodo unless otherwise indicated
+the siti is not used for the exam
+he was then appointed senior commander of the states of corps at asa
+however ashroft ignores and continued his way
+these different nos generated to finish and reaction notably politics and religious
+the airport is located in the municipality of kennes
+by october it is the one of our activities of the group
+after the war the majority of the first take care of the tunier
+the factory will be extended on this occasion to insure its production
+he approaches the design in architecture
+the sons are ranged by tetslo kachibuci
+this name is local and probably reference to the fast numbers that mark its course
+the church reformed in the village is listed as a cultural property of regional importance
+he plays in the professional team of his home town
+he shows his quality in heredivia
+they are the last fight
+the hamlet owes its name to the neighbors of boulingallier street
+the creature escaped from the third of cannons
+it is also a former colonel of the merasine army
+he was born in sheffield england
+the original sound track was composed by peter gabriel former member of the jennies group
+the double event is one savin paman and kim clicters
+it is the measure of a half way through the other
+paladitcola and pericola are sometimes assembled in the continents
+the neste canal crosses the village north of the comun
+he gave rise to unprecedented unitiousness
+the district extends over limiters
+this one commemorates the independence of the united states
+the sources on the history of vita revealed it the compound character
+she is the seat of the grenotivengno
+at his return to lorraine he settled as an entrepenor and paid coach and sara eld
+for her male domination is first sexual
+it's a color salt that is used as a metal poison
+he was a knight lord of covelville in picardy
+no source of inspiration is known for all ludo
+simultaneously joan began studying mathematics with the help of his brother jacques
+the church is located in the municipality of sageness and cre in the french department
+jack de park is a degree in literature in a doctor in law
+certainly some aspects seem to be similar but the purpose is not the same
+there they form a monstrous creature which proves to be the sjigo
+with five points she ended up the last place of the competition
+the guardian soldiers were fighting in vain from this land
+the withdrawal can also touch the withdrawal can also touch the flexion of mono
+the yamasar plan is part of the glorious fight for the establishment of the shugna
+but most of the time the children living on live are reduced in slavery
+he will end alcoholic
+we call vaso constrictors a substance capable of retreasing the light of a sicilitary bay
+this species is endemic to the northeast of the united states
+he relies on yong players that one he relies on yong players that once stooped the championship twice a row
+he was also a military in the navy
+he also takes private lessons from opposition with pol gilson
+charles rubier took the lead of a cabinet that will be quickly dominated by frere orbe
+he is a member of berlin should begin law at the university of zurich
+it includes most often paintings for claffin wrestling
+his objective is to encourage the creation and permitter the date
+he is named after the french explorer jules dumont durville
+in return she joins only to sing in favour of the racidia
+this extension of the urban area is made at the expense of the best agricultural lands of the plain
+but annie told him no again she said that she is not ready
+ivon begins his career in latlasa cominero
+he is part of the famous trio comic of unknowns
+he also was the target of pluck bentin in marson
+giant ventilators appeared and start to roars
+marko had no clash of war during the second world war
+his season recruited ended in july due to an injury in his shoulder
+with a increase in dis the manufactures can increase the production and hires
+indeed their reduce inspires blood and the flesh of his prey
+this time catalina himself would have ordered the cook
+the room is made up of twelve variations based on original theme
+he marries the daughter of a composer of atledmand
+wisky is black and proposes to match the skin
+in the province of elland it works an independent postal institution in london
+his name tells us that he was the son of a fa passes
+but if there is no other tools at the world's scale
+the doubl's event is won by angela mor tima andrea buxon
+some anthropologists could go out of water
+then he won the fifth stage in regulating a big round
+two main climb ways allow access to the summit of the char
+his patron is of romanian origin
+the wicilias edition in german is lodged simultaneously
+the panels of display have often been a transient for the acinereal activism
+in this production the rolls are held by sixteen commidians
+he lives in avignon and accelerates the end of the austrian occupation
+subsequently he continued to hold a rebri of cinema critics
+it is about now to transform it into a museum on the history of the city
+she became the first woman elected president of the republic veson
+the events relating to corresponding are therefore given on the same line
+however the textar still used to day have been written in appeal to the second world war
+it is in this episode that bitchell sees for the first time in bas terror
+the pigan of the armoric nature park was helping for forty years
+the party was opposed to the creation of the kingdom of c i c e
+he is the cousin of grimasti of jurusel nabed hemed al al
+this son became a symbol of the fight against the oppression of the soviet regime
+not a trace of maclus
+he is the son of doctor
+she actually has been ordered for australia
+further away she empeed herself in the dress near cormand
+in the academic style it was a professor at the royal institute of fine arts in naples
+he is sent to the western front as a soldier in a communication unit
+the money passed through people from the people in the jesonist network
+at his year of university each effort has a right handed pitcher
+the cause s a medium abundant river
+jenbutist operston indicates that baask means in her boat author
+this species is found in bercina fasso coastal engana
+the summer is then cooled and the food is diving there
+the one of the moridons is the most important
+the roser ranged in line with an alternatively on the stoop
+the castle will change several times in the owner
+with him the iron crowd is decorated as a whole of the nobility
+ensuing a derivative of the spanish term
+he found himself in the location of the castle of duck that we can see to day
+in france each university offers its own threat to specialty in option
+it is a tradition that comes from the roman empire
+thus ansun arrived on the last line must be known
+this disease extends along the australian cordier towards the australian alps
+it is a playin at friends of the canadian team
+on his grave is engraved his name as well as the code of his family
+ther len o genservile was then concerned in the first leader
+some artists such as ed van allen play instrumental titles composed of a single syllabe
+the sister of rose indermeier was a specialist in the descent
+responsible for this intrance ship and the sailors were made
+these informations always hoping for it
+the artisus covers the whole city of the grand dechir
+mark michel and lewis were dead for ten years
+however he failed in the fist placed in the finish
+the skeletons are now kept in the anthropology museum of tokio
+a museum is also dedicated to his town of suc
+full is a graduate of the university of cogate in history
+the drugeta is expanding to the east
+this clause does not apply to the person presently loaded of the maintenance
+he would have come to constantinople very young and decided to come a monk
+he is the son of rosa and l sherman an active in music hall
+mitch with malise invited milane to diner
+words or ardour sonared are synonymous
+it is then the elbum attributed to the group to seven sold
+it is administered by the northwood footbal club the clote of australian footbal club
+your commune of hafedi terva is in the head of the population
+the municipal museum of wilmington occupies an old tavern in the call
+this kenton will later be named after the town of vizl
+he had his successor de camel in sadini
+the water from espire a big number of flowers
+he speaks four of them often claimed as a polet
+he is the son of a crowd in ter cati
+the developers wish the developers wish to be satisfied by players as usual
+his partner has a ferventry on the ground floor
+the navigator is particularly present on the agin continent
+berrases are brown
+he is also a stocker in terms of a city close to the hob
+he has his main entrance to saint martin place and his secondary entrance to the saint maurice
+a footprint allows access to the dock and crossing the laws
+three years later he discovers the superliga
+the business man reacted all of a sudden to me
+it was classified as a historical monument of the antarctic
+the choice of the correction goes through deficiency
+he was born in sheffield yorkshire
+he is the grandson of a former football player of clapton orient
+a few elements make it possible to start questioning the absolute domination of human beings
+the region is located south the region is located southwest of padu
+he constitutes a continual and devoid of letters
+isn't it true mister patterson
+he is then appointed adviser to the environment of the consular government of maria antonemuda
+the first gold miniature tournaments also took place this year
+it is a species of rhoades found in mediterranean and in north america
+she takes the form of a monumental bust
+by a director general under the authority of a piloting committee
+replied ayrden with a tunke
+the montferrry according to the type of culture and variety
+he thus becomes rector of the polytechnic school of voro
+blyn would have used it to refer to the garden or a species
+its source is located in the mountains
+it can be several lawzers and on several occasions in a row
+osmopolita it is not only found in europe but also in quebec
+he was used at the beginning of their christian era
+it is also rewarded with the best players and defences of five
+he recruited dandozen a military experience
+the keka ku jib came the same year as one of the first national treasures
+the representative of the republic came out tasharp as entered the senate
+its habitat is the racipal zone on the summits or on slopes
+he is overturned by german matis poy and sloven roch dernitz
+the vell of the documentary went on
+the declaration gathers the tribe at the local level as a country
+the fleet of the dutch is no longer threatened because he got out of the gun
+evilin goria is michael kenna then obtained the rolls of the parents of the parents of thor
+its grafts can majestically be tiny
+nemazin contemporary of calpernus had the same quality and the same lack of funds
+the room of the number was close to the australians
+the city is mainly made up of wooden houses
+the academy offers work shops and civic education and political activities
+the title of the album is from a misin venture arrived at the bagota airport
+the coordinates are part of the cylindrical order
+gip and you for captains keats on ill dyes for the design of motors
+he is a professor of labor in law at the university of pire
+on april after obtaining the clemenzo river french on april after obtaining the flemenzo river friend to the order of evacuation
+gidoni has notably benefited from the collaboration of writers jean ruel and murney
+in the case of objects produced in seres some cavarets are used
+its regiment was dissolved very quickly
+the more resolution was taller the number of necessary bits is important
+the mich has a reference and is a rare stroller
+he follows teaching at the fine art school and then in paris
+other days of the fensive production line were also selected as walter jones
+at that moment robert falls to be victim of his imprussion
+he has appeared regularly in the role of doctor
+not to retain the glands in five or other plastic camps
+the roof is made of the square tower is still covered with malti plaer tiles
+it is also the year or sunebe and baslets ends from their corporation
+his counterpart can be a mita m
+books with epiduel pitcher is justine verlinder
+the raid inside the clin the eunich order
+write the name of god suffix is enough
+he is the father of actor arthur olden
+he has the leader andrew mc man
+the county becomes a neighborhood and the principality of leech
+alraga joined the armed forces of asbajan
+the republic was proclaimed in bullywick
+indeed the federal ministry of posts is restored after its port
+he escaped from the train that was carrying him but was again catched
+under a form of leveration are presented by articles on the globalism
+he around the movie teatr
+rudy morgan is the prime minister
+the blow techniques are all drawn from the menga and used as a speaker
+he decides to carry the body in dover in new hampshire
+cartley he is vice president of the union mathematics society
+he brought his douple and in the hands to the romans up to the abbey of the city
+at the french revolution the street is deterred by the revolutionary transporters
+he plays with the buffle color of preference
+the retirement houses were built by the park's room
+their attacks can provoke a delay of growth and affect the flowering
+chrinopus became the first friend the leader of jopis on the french market
+these four conflicts are happening in america as a consequence of urion war
+he was also involved in his simple career
+the responsibility of the warrings is under the community of municipalities
+the results therefore come in first time in the youth section
+we can find an ober's esther
+besides the sleep was soon to win the trips
+the municipality is merged with crinde lands
+chease a knight of the order of leopold
+he is edited by the tingo editions
+amatan is known for his resist against the amaden culture
+the death of the inca does not mean the disappearance of the people kashia
+it is won by estomat durgen shold
+for this the supporters must take a tomb or under their opponent
+he is interested in the saxafo and practises the tenor and the berton
+the university of helsake is the new role of forming the grand duchy of finland
+the practical beginning of astronautics are due to the invention of the reactor to the liquid propustion
+de latrede urions was a technician in hope he was going to landlord
+adel gord was a subject to her adomania
+what can need a new food table
+it is considered as part of the new animals in the country
+this species is found in casta rica in panama and columbia
+we are going to start a night club in a sun
+i' gon to be so preoccupied by the way
+before this operation the onlier before this operation the olier foot is cut and stranded
+there are several sanctuaries in japan
+therre also traps to hunt and can feed
+under the old regime the municipality belongs to the movat bar
+nevertheless the reaction of two different angles is perfect
+most of the trees in this forest are too gallipus
+would it be new instrumental model will be important
+the confrontations were determined by tierjosa
+adaptation of the work of george simonin
+the occupation of the village dates back to antiquity
+during the refrain he was accompanied by the electronic battery by joseph more
+to the north to the northwest they can be found in small crater let
+he was born in spain from an english father and a mother as are byjons
+she died in washington from apoplexy
+the romanisq and gothic walls were decorated with frescoes or scoges
+it is a common name in serbia rosha and bosna
+i'm following the story was born in zuric
+the director also plays the role of ha main actor in this tragedy
+he leaves the private sector and founds middle ages
+during the american war he served as lieutenant in the american army
+he abandoned his studies at the age of sixteen and becomes a railroad for some time
+the area now known as watter flup is a traditional navarro territory
+a shooting took place on the court occasionally for the panic
+the number of parentheses represent the number of creative tinels at this date
+his activities led him to travel a lot
+the lake matthew is surrounded by a certain protected land
+maria is yours all defects
+it is part of the important lancaster system found on lake
+michael gave a marriage to johannes one of his own small daughters
+this component is not compulsory the only one in a laborator plasma
+whipis concentrating on the blocks in the outside
+the movement is eparated from a shade with your own experience
+during the same period there were protests against the policeman
+we assume that i know the particular name
+theeffectively average of one of these promotions is twenty students
+this ticket arrived a few weeks later followed immediately by sub
+it is to be verse that the stream had to be too severe
+a few months later the legislative elections of the
+the control of adiovisual product is transferred to the ministry of culture
+this causes a kentinal election
+this bagonia is native of sabate
+it is during the last tours that he died in switzerland of a hard attack
+the coastal river is eva which flows into the adriatic sea
+in the small village of tungri andhe is the son of a doctor
+gerardin was addressing victor hugo alexander dumas frederick selier and ug
+are you the fifth child
+this often adds the authenticity to the rolls games
+oh yes i m b h is the biggest company in europe
+qualified for the final round and bold
+he lives in plain meadows
+the check version does not have any registration
+later this actors will start in the sun
+a long version is available on lines and is available on lines
+discouraged the chuns started to retreat
+this album was their first commercial success
+from the old base only exists a fortified gate
+having been there by the ader joins the merchant marine he engaged
+the totality of existing scenes represent a few twenty meters long
+one hundred nd thirty is located at about one kilometer away from another village
+she got a lot of regional prizes
+it is only found in the surroundings of tura
+he was declared excellent and even higher to the glands and festals of the pempa
+anten of greta is also installed in plasance
+he's part of the one section
+he supports the rights of sovereignty of france in metz
+traditional dance for example is goli
+he had married marguerite leverdeer with whom he had four children
+john bumont is a dead painter specialized in painting
+he bings attached to the series
+this season forms in change in the suppression of qualifying races
+the cloister looks like a true masterpiece of art menlin art
+however the government does not find a new government
+she also works as a free lencer contractal author
+heaping in occupation became difficult at the mafins
+it is a flower plant which is usually pink pale
+the morning in the evening some trains are inticat the morning in the evening some trains are inticat the morning in the evening some trains are intricat with the alban to touch
+the pieces are made of slate
+the public function has a hospitaality
+the principle of the hereditary empire is missing the collition
+the two races are considered as different
+the liner mogile is correctly marked to retrieve the astronads
+their provision is taken back to the identical for the entire series
+the success made a huge success without saying that he was not very good
+the yacht is the biggest private hut in the world
+the extraction of the sand is also available
+the victory was won by a pinal
+avignon dead ends
+opher small flexes would have been observed
+the seven first seven teams are qualified for this competition
+the classification is made by a commune according to the alphabetical order
+this is undeniably a benefit
+she answered the newspapers she answered the newspapers questions and was relatively relaxed
+strudden declared that he wasn't strudden declared that he wasn't strudden declared that he wasn't responsible for the pregnancy of his wife
+his name refers to nosika a princess of audite omer
+the limbins of sports in for sport are resistible
+he also use in chetria
+he has the merit of having largely popularized the reading of the senada gospels in spain
+at evigame the two teams will play a round trip to two legs
+she takes place in the etian in the tranceto district
+nepir has a port in airport
+the sa farian college is named after him
+in the opposite case we are toning out projection of power
+he decided several tables for namibia
+he is installed in the on terrace building of the futre scope
+he is one of the last leaders of the union of timbery
+five soldiers french were not allowed to fight for them
+a huge brand was also prisond there
+the historians know little about the young lady of edward fubred
+he is one of the first tenesian printers
+she's thus becoming the first she's thus becoming the first genis to play four consecutive olympic finals
+we can also replace the ambiguous by pseudonyms or elias
+one of this reference books is his biography of mozart
+she believes in time
+the videu surpassed the two millions of view on the platform
+she finished in second position
+the fie fighters are overflowed the number of renoceros increases in the city
+etiendion founded the priory of saint valeri
+his wife were married four years later with elizabeth de bosney
+he is selected with a du he is selected with a duo he is selected with a du of comelti
+he is also in blind of the chambodem
+he is separated from the agency the cherets gentelamere de fontaine clues
+it is eliminated in the resburg and wilbelden tournament
+the production is carried out in reolico and are reinforced while in paris o
+the floor is to mister emadicut has a different reactivity from the bazelle
+he is also the narrator of the japanese version of tom mijerry
+a sign of situations on the side of the mountain overlooking the city centre
+william remov is also not respondend to these accusations
+she formed the sub femily mono cupical sadeline
+we call ther bedelatamilanes
+the mill of herba property of chateau
+he is delivered to the field prize and the strawberries
+these climate conditions are great to harvest the recibroter for three times a year
+she smack is the second largest winter sport station in aran after ten years
+the colours of the team are red and white that of the bohybian flag
+the article has been greatly inspired by thelw
+during summer he is also possible to use this small interistic in brittany
+near the city the foundations dating back to the were discovered
+in response there did not seem to be a revolutionary contention of the general degr
+these books are translated into fifteen hundred countries
+he also translated works by taras he also translated works by tarasco maxgorke melan and bill
+initially everything that the player is sensitive and its name
+she manages the electrical network of its desert regions
+the dutch loses him and industrialists the project
+gregar went through his life in the nerick monastery
+we are waiting for the sequel of the procedure it is dismissed by his team
+the speech of the is one of the most important speeches of general de gal
+it is like his father a messenger of god
+the sale is confined in temperature conditions
+the history of dunho is one of his daughters provided theme of the daughters
+the global version is based in miami the united states
+the club managed to go back to the first division
+some representations take place on the bourget square in the city centre of juliet
+the population was spread across the four main characters
+he gets base classes
+cornward is located west of the province of fris
+mathematically the expedential corresponds to the function of distribution
+the institution is then representative of the style bell epoch
+one of the last coners to appear in fossils
+these performances allowed him to be a dutch sportsman of the year
+listen means here extraction of the signal the characters and his understanding
+she was considered as the best known painters of her husband
+it is with sotnow that he spent all his career in the lower bar
+this species is named after sis ruden
+he is considered a famous scholarship and a government official
+he is published in british label plant in records
+he was born in karp a town in the province of moden
+he is the lithuanian flag bearer during the olympic games in london
+this file will be clean to the object and follow him everywhere
+the department is based in atlanta georgia
+she aims at the human health to work and effectively
+a case of laying of tenue is probably the fact of two females
+he therefore looses this trophy for the second time in a personal capacity
+the hamlet has a communal school
+he starts by serving in the sa frontier regiment as a sub lieutenant
+it takes its name from the owner of the field
+the upper register has six levels and seven spins in him and not white painted
+the neighborhood is dominated by the insineration factory at the black devil lane
+a minotel is a passive terminal
+it meets in saba where sarawakan calamunta
+he plays as a three quarter wing
+each of the four sections of the tribunal is directed by a vice president
+the colonels separated the nives with other ionic orders
+it is a velvet road that cuts it
+the female can be hung on a year
+it is used in the united kingdom a medicine in alcor during the first world war
+she also has several private schools in london in c h
+erbaceous plants small and white flowers grouped in her bil
+males have two different pairs of monapas
+herse is considered to be of arab talent coming from polish and russian lines
+furthermore she was counting ver institutions
+the competativeness campaigns held each year under the pelets
+his first broadcast was un quitenant
+erson teaches at detroit and is on vord
+we can also use two seconds and second towers
+thi sale established a world record for a lover
+jim muffin is finly re elected with me a john
+it is a monoclinic mineral generally transparent when it is pure
+in fact the quadlobes remaining and the excels are home to six angels
+the church is located between this place and the new house of the same place
+he was taking care of him who was the sun in my voices like tonini
+misne and a few friends helped him join the pieces of them
+he studied at the royal he studied at the royal ic chool in the united states
+he goes to a rock to be a human and thus he goes to a rock to be a human and thus prevented american bombings
+lao bertrand a frigerd at the red and thausands of his council
+after a specialist of mitrix an hances this woman who is done to children
+yet she says that she is ugly the monster existing
+the color of suit is very well known for tourists
+the piece of annex engraves during the crisis
+the matthews are played in pennied on grass
+he wins his third gold medal in defensive
+the painting represents the design of the artist on the life cycle
+the jecolon is a bird often associated with convity
+originally from spain is selected from local powers in cestro
+during the previous eurovitian competition he had announced the points of seden
+he was very treasurer of the city of anvers antwerp the councellor of the councillor
+he dominates the lakes of lafray to the west
+the tors are still prolited from their effective charge
+but not in the production o renaissance for their rudagati
+the latter saw the dominiuko that his grandfather had locked up
+since then he has never played in nashe
+its inhabitants are called the nefica
+based on a station the line is almost on the right to the entrance of barkler
+however he was also known for his weight problems
+you invaded your mind to choose between his original and not adoption
+lendry is not a major trade
+the coastal stations must depend on this groups constantly work
+the work was fiercely criticised by florent and by the repetativeness of the composition
+loman also appears in psychological friars without reqend
+it is the first meeting of the german division
+he wins two titles of columbian champion with a team of junes
+the tower contained the machinery to lift the chain
+the leads of these families are opposite simple or pinny compounds
+he is a configuration based on the apemel engine
+the church is located in the french department of calvidos in the municipality of lesson
+the sons are classified by the album on the album
+he took already and held the same position
+giles and christi form a couple of badly teanagers in their skins
+at the moment the king arrived at the hospital he was already dead
+there is also charles ublin as well as different species of oaks
+the region of elmilla then enters a period of the powerful period of its history
+this a c is revealed satisfactory
+this building is located on the street of the selsma hospital
+the glook also has a commercial album by taking the different themes in the albums
+he died there the following year
+english wants to maintain the streets of lucas and manage to a specific solution
+the day goes on the elections violence extends in the pin
+the bredwin drawer was not considered enough to welcome thampionship
+he is two goals registered on penadi
+i never forget to talk about it about my career
+the pentepenum of prairies considerably over the church's activities
+the reputation deadley from the island has received different explanations
+he pulled a trainer on his career as a playwright but still wrote scrips
+he is only accessible by barbato and the other ones
+the buildings are located in the heart of an matamus park
+the trimetrp is emanation of the rhine pageants
+who loses consciousness then to my whim
+this transfer will make waves in the north of the country
+he was minister of riculture and numerous times during the second public
+it can always be visited
+one day he finds himself out of the day and was ruined to sell his house
+the building is the current headquarters of the estria vincenza
+it is a concept album focused on the pyramid of gus
+moreover the wind moreover the wind sickness is important
+no other canoe was wringing the network of the flowers
+after the medical training he joined the army for two years
+he moves into a minimul and a full charge
+the use of magic is limited
+she insisted in order to gibbert first warning to calvinism
+it could thus it could thus be used arbitrary for the repression of unsurciful people
+some species consume or some species consume or bacrats
+he was indeed a member of a black link of paintings
+he was the son of jose canido a spisito and also a beddler
+the bucco danosa received a defeat and in this form series
+he didn't he didn't go back to politics after the july revolution
+during his last season she was called gessel
+he participated in three urion championships and one world championship
+he then became an independent writer
+is one of the best between whaylander hullock and bruce
+the work was horrible and was in serving the public
+they have their collections to attract some patches on the right
+they take vengeance in the prostock of the council of anti
+the whole is covered by a large steel structure
+when the first world war broke out the mullins fought for the german empire
+i wanted to marry the question of the painting against the screen
+eight victims were executed then burnt in their cases
+this is seen as is seduced
+a the international level of bacco
+the same name ber serape will be more than one after the other
+what is that men
+the locality took the name of the sait den deme
+the use of this molecule has the reputical purposes
+she plays the actress with oscar lowya in a throw with te fruit
+the whole meetings took place at the stadium and the salacier in adi sliber
+for her tinnage added the country
+rene brafera considers him as one of the main english writers
+only only simels are cool
+the term refers to the semi territory of the pachart
+but the doctor said to have a look at the same time
+these proteins are engymes that are involved in the plegagy metabolicis
+the same day he obtained no longer sit at the councillor of the king
+this species is found in new zealand australia and tasmania
+fofor wiser the archive fillers at the columbus fill festival and waldy festival
+on the phone the last one to pass from the canstate
+it includes together the sun recorded by john alone with acoustic catar
+konenhiter proposes to write the discoveries in the life of olde
+char can be used as a name or a verb
+he is buried in the sa ge cemetery in toir reveres
+speto is located northwest of luhans
+as for those who are in terms of setuary they shoot him as a dilapid
+she doesn't she doesn't come alone but with sophia's mother fatima
+it is held in europe every year
+its nomination to the german ambassador
+these are the first traces of the rium an the city
+it has a laying teaching in the history of the city
+jon mark gencovency compares this dependency to an addiction
+this diocese is known as persecution of the moslim majority of the country
+bussy is the president of patagonic film crup
+before his election dor le fever was a real environmental movement
+its distribution area is located in the indian ocean and west of the pacific ocean
+association is created
+he created anociation for them
+a part of these objects is at the museum of lator de marciny
+after candy marshall leaves the cloth green becomes the first coal keeper
+letric is located in rome and in capris
+he becomes after exams first lieutenant
+coru saliade with the ribbl of espert to convince cassar in macre on alla
+suspecting caas she brings him to the sight of the guard
+a port was found on the side of the vivere located south of ario
+he also suffered a laugh at the twenty p h
+west terbaya is killed by funa the guard of the body used by the circus
+the verified the basic hypothesis is the one whose leader is now in the same way
+it extends from and includes two painted and glacier zones
+a few moments later he won two stages of the tour of maragante
+many legends circulate him on it
+he seated the largest collection of astronomical plates in the world
+he was formerly a drummer in das groups he had formed
+the wings are later on wearing several reddish tashes containing smaller ses
+the one of malins or the one of malins or the one of malins or the owner of his family
+he was probably the first refractory priest guillotined in brest
+in south africa he was known as the name andri peki
+the china of armilan was the main attraction
+she was also a independent state managed by delardka
+with the lock the canvas forms a human character and a man
+it is located in the west of montellu versu
+all these sons are built on the same tour and use the same mare
+finally the golden emancipated the ideal triple of rent and work
+claude wants to follow him and to do so far from money to his father
+the interior is closed with a glid
+it is found in walda in danda and pareva demsbrok
+their destruction would lead to the extinction of the human kind
+the movie is produced by david hill for the movies and the
+following the path of his wife he converted to catholics
+like his name suggests montignac and tupenery were originally independent parishes
+the milky box is moved but decided to go look for seeking for seats
+he is a friend of adolescence from francois holland
+he is particularly skilled to the gauls
+this is the norwegian race which gives the most heavy rainfalls to the battle
+the rivegan spreads to the east of the bruma mountains to the territory of wren
+zimberg wald was merging with esburg to form the new commune of wald
+bas going to go to almagera spain
+in a few hours thise supplies were made
+the name of the summer of this context led to four femilies
+they are mostly merabouts and stores they are mostly merabouts and stores cot for farmers
+he says he is impressed he says he is impressible and imaginative but courageous where the situations are required
+we must then take the work to join the london city centre
+he was dedicated to the wife of patrick muriano de mini zurfis
+at the exception high school bears his name in roman ser isir
+all these were naturally scandalous in the classical opinion
+it has twenty one episodes
+he's crying upon an average of about a year
+he gets a bachelor's degree in t heology of dogmatic at the pontifical gregury university
+it would be the wiser whisper the old times
+his work includes the photography vidi and the creation of public order
+delara belonged to the nobility of the robe
+the points or goals marked during the championship phase are retained
+after the war shak rose and went back to strasberg
+trio who lives from the expedition is killed by agents
+it is located east of franville and southeast of lieville
+he is true that their resemblance with the russian paintings of bovenin striking
+this model was visible only behind a window
+it is still there clenavan's waiting and roaling on the door
+it's a grim negative and positive one
+he sings the lyric repertoire of all periods
+after his convalescence herbe leaves the army because of his health is weak
+nevertheless others people and institutions have put in place in the madel's museum of art
+this pip is generally placed and vertically
+the old chiton church writes a long letter to his niece
+his father worked in the administration and his mother a doctor
+they are also the very first to self prigand with christian metal bands
+like his friends john russell and john montadu elfax were passionate about cricket
+a steel in a steep is also erected at the state of the mediterranean
+alkin reorganized the school structure
+it was the emblem of the tailor mc clinn
+only clubs were montenegrans are presented here
+he called her as such because of duf the snow color of the heart
+he also criticizes texts used by the american army in vienan
+yet all anthropologists are far in the correction
+its from this material includes the first quests
+most of these documents are kept it the national library of edinburgh
+besides he is a car parking and is set up near
+jonah is in old possession of the abbey of saint antoine nabe
+his function is the following
+these results allowed him to obtain his best ranking in the world
+he is considered the first one of modern discape
+finally he also joined mercida
+later the members of the united states emigrated to the united states
+the engine is of a
+the opinions received by the press often criticied towards the game
+this selection is to be standardized the inner color of the same box
+the card editor also allows to place events in the different zones
+and yet i didn't and yet i didn't feel in him with his compatriot
+zert or flotilla under the command of captain de corbet herbert fedrick
+antoine chessex is active internationally for more than fifteen years
+the victoria covered real storms in the air
+there is no one answered
+he spent his entire career in london
+the destruction of the canal is called visitome
+graduated from the title of engineers he travels to great britain and in the united states
+he studies with srit enmi and
+it one of the two main orthodox churches in the
+his nephew is curtly louis de la quinza
+she placed under the southern federation of the royal federation
+a poor word tartifies him in extraction
+the village of ever is located in the southern part of the parish
+but his valet persists and he regularly makes refuges
+vaya in a traditional style
+the civilization of ancient egypt has experienced one before the christian era
+he was part of the agriculture committee
+you would have paid them said severnon belfast
+polyn nigry holds the role of harowine
+he is mentioned and decorated with the crou degare
+my daughter grew up in peru indiana
+pascel lewis chocolate marie affrey was born a newly sursene
+then he formed a hard time with michel berrett
+he discovers the actor jeff strecker who became the new debt in these videos
+for a year he is their co ordinator of the offensive and minor league
+for a few days it had become a fixed idea that no one had been formulated
+even the best software translates are still many erus
+each seat can be moved forward or step back the sitting
+these concepts were more developed by arman tacteran
+she died in her eighty years
+this propaganda was rolled in most of the american media
+one of the historical leaders the professor
+the second world war is the annex drama will mark the spirit for a long time
+aligia a memory of my well liked wife
+the secondary school covers the neighborhood and the secondary school
+the director wants his brain and thinks his personality will be a considerable addition
+the story of a golf evidence verry dwarf is very hostile to his rival
+several application systems include information system
+the bowl could be in wood or stone
+the right in cardia disorders are a frequent complication
+it is built on sandstone material typical of the region
+the transept is the nave of the church and the church rebuilt in the gothic style
+he knows the life of his inhabitants conversions for most
+while talking about the mails i want to live in wine
+classical editors are metals carbon ind hydrogen
+in the medication of the blackbird he is a co rider and a fiancee in elizabeth wan
+mazameno studied political sciences at the university of athens
+he cassed about and was monstrous damages
+the north is a child born in eliza and jecky in the youth
+the latter accused the islamists of bokoheren islamists of being the authors of his massacres
+he spent the day of the year in cove in the same game in england
+he is also considered as one of the founding fathers of the vicaryof sanja
+his father is edward savo and his mother is demary savoy
+thus it is sometimes called malacles caxer
+he probably has as mission to organize the management of the imperial goats in this territory
+its most orest part is from teupon to the south
+daniel asked bob to organize the party
+desperate they declared himself to be giving a million to those who would solve his problem
+it is higher than i can run into relationship with them
+it is the most important scientific union charter for medicine
+the street is home to the regasine animation centr
+the plains were old and were in the time
+the ben the benhicene at saint alis lar bebardo becomes the first coach of the club
+his life is subsequently surrounded by ramparts
+many perimeters received by paul borden more specifically
+no not for everyone
+these funces facilitates the pribation of diseases between the rocks
+the implementation of a chamber of auditors will be put in place progressively
+he is part of the reunion national park
+he doesn't play against psycho slavakia but he holds against the yugu slavia
+his father was a worker and his mother was a housewife
+yesmen woud are graduated in classical literature in law and political science
+the use of intensive insectiity is not possible to reduce the number of funding
+these ties are critical of the new colonial society that will have a strong opposition to the bishops
+jacques clagier de borcells is also the first president of the academic society
+despite this failure several other attacks were led to the next
+the building is located at the top of the hill of anticiles
+paul in the minnesota is a american poet
+do you want five hundred twelve at five francs
+in general the more efficient is more efficient than the temptation of abusement
+simon john orflin of fourteen years has always been lived in the fahal castle
+one day costes is seriously injured
+he obtained medals and cords
+this medal was the first to win by his country cron
+those who confirm the creation of a fugative formmation
+this movie dedicated to jacques toner
+the route continues to get down to the herbers where the plane is waiting
+she had for the chief town vienna
+he was also one of the first australian and write their own titles
+he plays in greece with panylenos
+the cerie saw a lot of small and big screen
+state is e republican conservator
+after all it was a brave man despite his ridiculousness
+the vasi is rich in roman and precious metals in particularly the tropic
+it uses a flag to blot the two trembions
+ededilis pedlises were proscribed and exiled in england their castle of tantagon
+he is the son of charles adgre and jacques conag
+the prince er karin marries the japanese princess nach moto
+i have to go to the portal also sign the sala dico and the laragadeco
+gralin king of cordwall mentucua in norway at the head of a considerable fleet
+most of the inhabitants speak to english and pagano
+he lived in new york and taught a large part of his life
+she was broadcast on the sernflics in the united states
+the eldest is reneed the eldest is reneed by the patriarchy who will have an artifice
+he chose to seize the foreign affairs commission
+the work of the echo is mainly done in a national assembly
+isolation will be thermal
+cuppies lively circulative and great quickly
+this one asked denis to turn cripping a star life
+after the addition the festival became an
+she lives in fa defin
+we must believe that the coser di have to be overset
+in the absence of proof that we have been asked
+the inhabitants of white are descendants of culinasorian
+furthermore presentaju characters would have forgotten the script writers
+he also offers a new position taken from the cold and hot water towers
+ana discod was grated to a wis efet message which was more present
+during the middle ages the castle was attached to the lordship of sor
+the local museum is partly dedicated to him
+the mysographic space of bernard david brings together part of his collections in armand
+and rogers was born in the netherlands in indeneser
+it also produces comets the car caries and engravings
+to have more money in the end of a party
+but she is disappointed by thi xperience
+it has its defeat the following week on the elections
+the grains of hard wave are qualified as mideline
+the sultanate seems to be heading towards a constitutional monarchy
+the area is made up of hills covered with forests and vineyards
+she is the seat of the ladder
+returning to the castle the veteran uses a gem to develop a new dimension
+a large number of language is ethiopian
+she her mother are the only ones of the hardship
+finally the etruscans do not follow the upsets and any war started
+the lake owes its name to the neighboring village in lamia
+in the cinema as the television show he particularly plays in the western jonner
+joseph also talks about two other people in this dynasty
+its president is figrodo
+the refraclonology is the researcher professor lepan
+the milan was shouting a footbukle on his second title
+he is buried in the jewish cemetery of okopoa in worsa
+johan studied animation of sand an hollywood and new york
+some never been rebuilt and had been rebuilt
+during the time the anarchist unit was made
+bery young illuminating suffer the back of the petals is paler
+she collaborates with the highland ballet and composes music for different coreographers
+she succeeds the team to swalbur campanelo
+at the regional level olympic has three permissions in five seasons
+he then leaves for the wimblden club and takes charge of the reserve team
+shes not classified in the historical monuments registered in their additional inventory
+he is surrounded by a magnificent park
+the sight enabled the sight enabled to pass through the pass
+their motherly family is a family of french soldiers marked politically on the right
+in the past it was a high place of boxing
+his son william was a teacher in electronics and participating in the development of rida
+john marie leben sometimes stands to focus on writing
+he studied at the university of kazan in mongolia mangol and bud he studied at the university of kazan in mongolia mangel and buddhism
+the queen marguerite made several stays there
+the feathering is green
+she was drawn later by one of his disciples andre poginsky
+casting careers are exploited at the montcourt forges
+the couple settled in london
+hydenan in accessible from the cache erport in the land year
+he practises his missionary activity as wisi skan in papa
+after that he takes back studies history which interrupts him as well
+he plays as an attacking filder for skuslovan ric lava
+the most famous version was written by john uxen ford
+this effect is supposedly happened during the toba catastrophe
+despite his handicap he was in charge of a dangerous mission of moves
+his taste is a crea
+despite being a boarder with five other municipalities including three in the hots pyrenees department
+these bricks were often used for the chimney or window panel
+she did not reappear in any other book
+mfel on the stage o george gave birth to a girl
+his approach did not lead and corman never had the texts between his hands
+the castle is located immediately in the south of the hamlet of bois de la roche
+he is also a silver medal for the limp it gains the same here
+this species is endemed to the department of magdalena in the north of columbia
+the mission of these different types of ships is established in mine fields
+he is the author of two operas and ballets
+bernard de clargo was especially for his wing
+this instrument is mainly used for the study of references this hit
+in geneva he receves waters from art from the mountain range
+it can be valigated the realism of the details
+he dominates the pas de la barsa and the paris barsa
+it comes from the height and runs through the inner part of the trio gentripillar f
+a close relative of the two brothers turn out to be short duration
+in this position edward lourant was also a member of the pomet
+the municipality does not have any urban planning
+he hosts craftsmen workshops whose art work as well as artisan shop
+his style is known for his clarity his master's clarity
+sill is the capital of sath karea
+a village festival also took place every year during the first week in all
+he would have touched american at that time for about two hours of photographic session
+during its creation the kenton included twenty three entire towns
+the club's first club is part of the french federation of horse back riding
+the following year he agreed to work for the fevir section
+at the end of the team the orchestra disappears
+some forians is the common city of etraa
+born in sirs he entered the military school and quickly climbed the scales
+the park remains open to april to august
+the kemesita river entirely flows into a forest territory
+he served in the austrian army during the first wild war
+jim gordon lands on godam as inspector and take care of the badman affair
+the notion of epistological breaks applied to technical purbess
+the kaasaki frontals its title
+the reallest undre serta is with him
+this species also produces in e s this species also produces in e sels that have even been its capacity of ten seminations
+spain wins the eleventh tournament
+the municipality of tose is located in about fifteen kilometers north
+one hundred and twenty three silver dots
+his father hamenon was a footbuller of the national selection
+it feeds on fruits leaves and mushrooms found in the forest
+the palace of arts also hosts the central media library
+she left the house with a lot of fatle problems
+the majority of the line was parallel to that of northern pacific raywood
+he voted the death of his former weapon brother the marshalne
+she tried the dra to manage the memory of her old man
+it is fertile
+since he returned to ayres his painting is darker and painful
+this species is endemic to amele rumania and italy it meets in partigulla
+he is also one of the first promoters of bluck squadron
+the same year the president of the party is chosen by gnamada as his political duphin
+the number of rules and bodies are present in the base is variable
+restos mirnisky in the real name and hersto dmitros
+this trip that day is organized by one of the regular members ginde traveling
+to day shiversy is a village of artois for the cities of wesley johnston
+linega becker was born in zimmerway and studied medicine at the university of cape
+he finishes his political career as a general of south africa in he finishes his political career as a general of south africa in caca
+she joined a group of folk lorics in krakva called eberslick
+the formula is therefore known as the last known hydroplanes
+and they made trust relations
+she obtained a bachelor she obtained a bachelor's degree in law from the university of new york
+after the experience of his sentence he reflected his lifetime
+the leaves are of a yellowish or non intensible colour without any particular interest
+she is then forced to quit all her mendits
+one hundred and twenty five culminating points is at of altitude
+the tournament groups out of order of importance are the ranking tournaments
+this act puts the yeman this act puts the emans tent to be called revel
+my second one brought in burst of a second penalties
+he was buried in the north of nefina
+gresh sings and the main commentators
+they are published in the same period as commercial editions
+it contains about ter mados
+she works in humanitarian and especially in afghanistan
+these typically what we call a tuit river
+this engine is produced by the tiwins manufacture
+he is a professor ameridus at the university of edinburgh
+the first assault is pushed back because the whole city is under duty
+they are seaved with the boudoirs in a file on the wall
+and the only french speaking to have received the ward of alser
+the embryos will develop ther till their explosion
+two cosde tenis are located behind the stadion
+minina corona covers a circular surface of a pariameter
+the show is available in two parts on utube
+he found himself an opponent then defeated by doctor string
+the rio merinian crosses this gorge before reaching the grand amazonian basin
+the lateral valley of fusa is occupied by a ravine forest
+marina burneus was opposed to her future
+the horse is semitercular
+it is found in brune in indaniga and milagia
+we can potentially add wine
+ah and huck by the tatalin of capera
+the territories of albertinos then go to the leopold's indians
+the winner and skear wete the most fast time
+kaji and madzukais were not then never more lest
+the garrison shill used to prevent the castle's destruction
+he accepts and forgives the amazon
+he is buried in the racoscale cemetery in nandesta budda
+date of birth and death are unknown
+he is currently working as a physicist at the cornill university
+jack wes jack wesler studied in the primary and secondary school in the city of nedel
+many shows are also reached by projectiles
+this one is based in israel and directed by china player
+in england he is assigned to the maseneuf regiment
+to qualify this small region we speak temosus
+he is possible to watch out the game
+this puts an end to the reign of the farneys
+his name comes from the great confederation berbersenet
+burkett is a member of the british academy
+the man comes out of a cigar
+but the manuscripts are made but filed but the manuscripts are made but filed in paris to be printed
+let let's also manage the wee without sating coffman gimbrams and gersenfeld
+during his reign the situation already worse than the dynasty
+so that they're about thirty years old frintots about radio canada
+to day more prodigious sails on all the oceans
+he finishes in the french championship the same year
+it is in fact the only vestige of the batia claster
+only two prototypes were built
+a mere parret was born an armond paret a country's life in morocco
+it was developed by philips record system in vienna austria
+its traditional name was sir its traditional name was simbert
+at the difference of dutch he pronounced
+disimens bed allows to increase the seed proportion that sprouted
+it is found in siadl in the piano orchestra
+the second part is singular in the baccantatades
+played against the pas de la loire interactive teaching
+daniel leon oh was born in san lales figne
+the doman is classified as a national monument by the irish state
+he is buried in the wasing cemetery
+the dors from the west have lentons in france
+he defends the status of the conjut he defends the status of the conjunt of his employees and the right to retirement
+under the old regime lerin was part of the baily wook of darmay and bens
+this fortification is therefore comparable to its size to the one of the gold fighter
+he also includes national elections in the court he also includes national elections in the coortitative republics
+the important thing for him is the reference of the material with the sight particularly its history
+the decision is executed and applied by the policeman
+he had to be almost seventy five years to conquer all the island
+i try to achieve anti since a binery it is ab
+son of the historian merfend he lived in erat
+she also plays in theatres
+the number of green stadiums can vary according to the conditions of environment till conditions
+the latter takes a lead in the general ranking
+she said that during her time they were maybe she was a good one
+to sir how
+really empty it is located south of the river of s
+they fly in one generation in may to june with specimens
+a point is marked when a record stops in one of the score box
+born in aldershot in england brima
+laurentus ave preferred the text
+the masters don't the masters don't have cowards
+he was born in the hoe alps in the hote alps and the new yorkshire
+mebekas a residential neighborhood at the time of thomas
+she does not have this place any more
+originally from cristio in fertini the group was then moved to leon
+he did not see this incident in his memory
+located in the east of uncrez
+he rises at an altitude west of buzon
+both of the main deposits are the movy and coter deposits
+this construction was entrusted to the general company of satiny
+no leather alive in these solitudes that birds had even moved forward
+yes and you it's not the same
+since the troubles of the development result from a decrease in the dis
+this motief probably dates back to his salibe again
+however the definition of the conero numbus is not unanimous
+two years later the cherris averages were repaired
+he is also a graduate of public law
+we support many untra penors in the creation and development of their start up
+xorgen mabi continues to create a recut growth rate
+a certain number of verifications requires and access to all must be made regularly
+mauritania is sengales mauritania is sengales and his repatriot at that time
+at the opposite of stinterans
+followed exceptional years for the club was not very successful
+he early lives in new mexico and in the united states
+the hamlet is coloured in red between the nuts
+the matches played at warrner park de bastere
+austria is guilty of being more red austria is guilty of being more rhythmic but not invited in a party
+it is a subsidiary of the conglomert rmainder
+the roman gate of the church is one of the oldest in emusin
+the differences of a matrix and competer are an inter complex way
+the city was very damaged by both world wars
+he nevertheless finds his sister in austria
+they were established after the recunquency of its territory by the red army
+the movie is confined a series of pictures based on ecographic nature
+silbertry is also successful in jerry goldsmith the composer of the first movie
+the common states of commonwealth cover the territory on the five continents
+maco will have supported invitation by this christi
+your lords are as if it is generation this one has allowed the creation of modern
+he was also a model of oceanic objects
+despite the protests of grue admars this one never came back
+the dress is soes
+the videoclip is starving because johnson faced just in timberlick
+burt took refuge in asia where he continued his activity
+the facade signalled by their symmetry
+this bird bears a good hand to any artisan who sees him
+during the occupation he fled immediately
+susspuate island is uninhabited and houses a natural park
+some permanent residents live in the femine bay
+a final duel is in the rocks well put in the price bestry an bill masters
+artist was born in kempton california
+the group could have reconnected with his success as antenna
+it is the wife of mathias rippa in sweden
+she has charnish green or red leaves
+his wife ordenan de grtrocy was open there which he was reputeated
+study these rules with simplicity good sense
+his testimony brings historian's precious view of the people venetian people
+the following year a new station opens to the north marina
+the game is part of the series no bonagas ampechem
+due to the german progress he is sent to brittany
+he is the son of masudura aritio and lord of the takimasu domain
+the library is home to the two of the parent fices
+as a mitfilder his hum is precious in the middle of the field
+he stays as a lodge part of the season and does not usually serve
+profile the playing writers are their upper part
+batradon realizes his academic career in the university of sanjon
+the tristine has an antimocasic activity
+the centre of the look represents the list of ey
+if the last one filled from bryan higger wallace for the destroyer
+we call this the compate ability an electromanic
+eight of the groups albums are certified golden dists or more
+he is recognized as a specialist of baeton and diamonds
+it meets in the little demaqualan state cough
+the multiplication in unknown quantities are represented by revelations and lands of property
+it is the seventh championship title in the history of the club
+the division including general freger kettling was interned
+his first trip transport him to the penia and turkston
+this project is an alternative to metra vision plan
+the title of the movie is in hitler in hollywood
+this forced the inhabitants to move towards the continent
+the committee is dedicated to the nizus unanimously in the cinema
+during the wild cup he plays a match against germany
+subsequently jus benda formed the popular party
+ersin hardy started playing in the theatre in germany
+then of nature it can communicate with them and control it
+he travels a lot to venice and constantinople
+here you are on the floor said the lord
+an ajosman was one of the most active architects in bst
+there jake finds a childhood friend join always in love with her
+bors remained two years in prison and asks for two years
+she had completed her training at the louvre museum
+it was considered as the best work in galicia
+this one was firstly mosed by his interlocutor the soul friendly with her
+he presents a strong and yellow green in aroma
+located at abalvitude dingetina is part of the minicipality of chimny
+we can find it until several hundreds of ken grams per meter cubes
+the bishop of mirapat is in charge of philosophers
+alexey lifter insures the comments with arna madron then with marie bess
+metillation particularly protects genetic and mobile elements such as the bacteria offing
+they don't seem to have a nerve in other languages
+he is elected member of te parment for the west sa waltz district
+and so and so father of aurora does not manage to overcome the lost and so father of aurora does not manage to overcome the loss of his wife
+his research focuses on detection in a frash of colonizer inlit
+the origin of the crysofer is unknown
+he played as a hafer back lift
+the heat of red heat contrasts with t typographical rigor of the halveka police
+thedeconel were gray spain
+it can particularly imitate vinal platinums including the possibility of scratch
+he elaborated very well known syntesises
+the absence of any advertising proved progressively harmful
+he went to pos gardia as a japanese literature translator
+your work you search towards your daily dye is the aim
+even he walked to kolvick made several donations in wokus on the whole of the empire
+the church cannot tolerate another power than the one of the people
+it was a professor of anthropology and social science at the university of chicago
+this hand looked like the one of the child's hercules
+goring resigns from all his duties for health reasons
+most of the action is located in berlin
+he is afraid of typography to the ezilicophone and studied at the pentino school
+but the president bernard tupkin did not appreciate anything atall
+he will end on three wheels a bill from the rollo without a central palace
+after the match he continues to attack dika with a chair
+the laws concerning the offering pascal and are also mentioned
+married three children they were flint and russian
+he also confessed that his coach luka pavolshevik pushed him to his training
+she also translated her four first books in english
+commune located in the urbin unit of burgera
+the medical and psychological assistant programs
+this close station is characterized by a white ballet
+it is acou the government appointed by the president ad felix
+the peried water is one of the most famous seasons that were part of boats
+his name is derived from mount mango meaning to appure opur
+in particular the variables of real art have a probability density are different
+no novel was translated into french
+america is located on the eastern slopes of mount cosmeg
+the cathedral has undergone many monuments and partially broke
+to day a water source is exploited for the mineral mado house
+she was married to urgesa citatis with whom she had a daughter
+hamlin left the active service and returned to new york
+he will be assimilated to a feminine identity
+the commercial supports on a strong opposition between the jonners
+it is also the tenth event of the french cup on road
+he's incorporated that year two successes
+raphael remster
+his capital was the city of kora
+paristache he met cristo de chavons
+in this way his style reminds that of gregory hasaks or the crembers
+she's the daughter of actor leopold white
+it is found in the states of cola and jvisco
+it is about three months that we can see a certain amount of time and a trumpet
+he is the first tire of the netherlands to play for the rest
+he was based in tambo mores
+you didn't leave him for breeding
+the structure is classified as a cultural property of national importance
+he marries fanny keats the sister of john kitt
+in this part of the works an erro in this part of the works an engineer the sibyl near apullen
+the educations are two quations the five of the merchants of the gorge and the pamplor
+isn't isn't isn't it my dear colleague
+she opened her books to her race
+equestrian of this degree what is happening to them
+it is a public city with animals that is made up of humans
+some of these procedures were knocked at the level of state councils on education
+the capital is set up a few days later by portuguese
+replaced uta is not an easy road
+she is now inhabited
+the gallery custel is a member of the committee of galleries
+but the nature is still there and present
+the whites are quickly very popular because of their growth
+he is the younger brother of philosopher jonathan bart
+the neutrals model inspired all the austrian efforts
+the doman was built near the source of a stream
+schwartz was born in the suburb of building in the city of tilrap
+the jita farel is trishy evokes it to the under the name of
+the graphic style of characters depends on their age
+after a year passed in qebouts his family arrived in paris
+the street is journalist and bulky is in charge of the mail
+she first works as an interpreter in international organizations
+this species is endemic to the pritish state in india
+she then settles in pickwin and as news prematures meets a favorable success
+sigur won the victory
+he plays as an atfilder in panarola
+this brother is the jakel replaces him as president
+my lord a ma prince asks him to stop his mind
+mullet of the charm mullet of the charm must name that i was a numble continued madame
+however his presence in serenu in gayana and gana remaine confirmed
+his opinions are not limited to the scientific field
+lisa is sad omer did not have any interest in her
+the commissioner got to follow the kingdom on the island but it is imprisoned
+this section develops the works dedicated to the lynk of the seventeenth century
+the history of the galleries of casablanca will be created to that of lemages
+the plane is more often inspired on the biometism
+it is marked by the influence of leonid entry it is marked by the influence of leonid entrpdor on the band entry
+son of pompeys noble of tiraceum and his wife ariana
+its capital is value
+a point of view is made of a porographical point
+jenifer dodna grew up in the state of ry
+his rampage dedicated to snub is the longest of the world
+nobody so seeing the coma and never had the transportation in her bed
+the network is understanding the links of the region
+the mennar all the author of the most important summers of the psalms
+forty two countries were hosting the change tournament
+this species is endemic to mi in hawaii
+he then analyzes the relationship between development and culture
+the following year he became a path driver then president and a horseman
+for the first time in a trip to scientific the photography was then used
+these members order the capital of the witches to lead the german troops to the frank
+his name an contemporary is bleckia
+the church of levere de bay has rich furniture in large parts
+in hanover he discovers his culture in line with the kings of his production
+the ethian body is then dragged into the streets by the people of portoprince
+then the crookers can move away from water to winter
+after the ice absorbed of the dessert the convition passed by the living room
+finally the frost finally the frost stopped
+she riches from mecklburg stagger receive brandenburg schwitz street
+the orders of colonel everest were executed
+the work has three spins and one of the other in the same way
+i i've been revolved and significantly lessened from the revolutionary accent of the enemy lineage
+ifs at that time he decided to become a poet
+the keshot is the creation of marian shever
+it is rarely called in madrid where he teaches natural history in the semi finals
+stop he is recarned in england and spoken in a flater
+the lyrics improvised by the events of the community
+the luttes of winno
+dependencies on agricultural buildings are close the interior space
+david had these lords particularly but they left few of them
+furthermore a trip to foreigners to help the cause of machin
+alone vill no sir lot hits it spown from the game by two culotose effects
+i am giving you a rewarded by the grand pre de art and the automobile industry
+he is replaced in the second seat of his son
+this species lives in japan
+twelve thousand mister bune
+tina is the sister of sinecanopola
+this planner still has set up a slightly triangular head with two eyes
+suffolk is from connecticut
+but now i did but now i didn't forget anything
+the largest number of jobs is however available in public purvices
+this form of love is reminiscent and superficial love and sheltered from despair
+the lake fland pourses the cantons of ragot de saint marcel and paton
+he is wounded and spent some months in kiels
+in the first years the miners were living in fortune cabins
+the macullos arbreated
+in paralle ragnacius his passion for his ex wife of three point suspension points
+this species is found in tiwania japan in the nince archipelago
+he will study at the alger conservatory in the base
+he also creates sculptures for france waved
+he is then half lied to the street of alisia in paris
+in the civilian this device will renounce the vaulted in france
+eugene roland was also general adviser of the department de zal
+and never finds his capacities to hear after the war
+so we came back on the strike we asked him his opinion
+half of the south of the department of dips employs a similar dialect to burbendon
+santasia then turns up and is diluced from the pilotage
+this is manifest in particular by the lack of public equipment in particular
+alas the purpose of the formula he wants to be on the next page
+the game is played at two players
+it is in fact the hermes god who then collects poxe in his retirement
+he was a knight of the legion of honor and officer of academi
+this questioning is found up to his most recent works
+his name comes from butu the muse of music in greek mythology
+on the positive analogue layouts this effect is known under the name of kink pf
+this image evacuation knocks on people's imagination and was therefore associated with this strategy
+the sight has activity in the mechanical engineering near manufactores
+the korgons pigmy are also present in the lake katike in two kuta russia
+he is buried in somers in connecticut
+when the women consulted at a doctor they often set up a complete annex
+the mas life was structured by two calendars
+he will play at the victoria university college
+there is a local airport the adrone belbus located in saint gingo
+he arrived until the quarter he arrived until the quarter finals with his national selections
+he currently lives in london
+it is the largest and most populated in thi ropality
+it's this car was edited by mark lambert
+tal fil verbist and original nobles
+she was the aunt of queen emma and the netherlands
+the pontre is the most used in the village
+he is served by amyrtus
+clomister is brought back to blackgate
+he joined new influences cabelto art of mexico
+the domestic animals are cold clearances around the eyes and would be more sensitive
+after a training at the brod movie theatre
+a light a light artery semel to silat
+this region is one of the most popular in the country
+john has a twin sister bog
+the old brewery is rebuilt
+we started to be mone in moscow at that time
+states integrates the university of clanson and by the same occasion the team of university
+it can be associated with a strong reverse and wind rephaels
+this clip is rewarded by an appointment at the greenmea ward
+flennagan adopted the work in his revisionist approach to the history of the emeridins
+when the shippings would be filled or the bad one
+it is a theatre that was not known during the mado in a play's guitar
+shortly after a lok came closer to the virgin mary a bath
+cartley the president is insured by george and pasco and general coach is laurent beltier
+travellers pressed against each other had not been talking to each other
+he had the writer john france walmigni
+to increase the efficiency of this we need to change the responsibilities
+but there is not many of these types of production
+it will be the last vehicle with a new type of motorisation
+this offer was presented for the first inscriptions
+she joins england
+but sila but sila has not reassured any that he is not just a compassible
+the project is double
+thus the sneeze s related to the syndictality of osal
+it is found in the province of viga and santiago
+the king went on to pursue his activity within his company
+but the breeding could be seen more correctly it was used on the spot
+after the death of minhaf he became very solitary
+in the centre of the ps trus the statue of saint marie de laire
+several masons were being carpentered in cadion
+in the same time the movie theatres can witness retrospectives
+the terminal from passengers to the highest peak
+in analysis of sensitivity he introduced the sable indexes
+jacque parmo studied history in geography in paris
+she occupies the east coast of the country in florida
+restaurants patients will be executed
+it is the network that is the most important in america
+he started working in berlin for the house of editions in mass
+there is a bell tower in the north and of sacristy to the south
+bouvard will later on make sure that smith was the first to get to florda
+twenty eight celebrities are pie o the twelfth season
+he got on stage for the first time at the age of eight
+rothery is officially classified as a serbian village
+boligations has a savenna climate with heberse
+she's the third in the championship club in the club of the elite
+the first lords of the second bage had coated arms with debating ingles
+he owned a property on the bankes of the
+he owes his name to his inventor in alp fisher
+she met three carson and john parker rots
+by sasere submission to ecclesiastical authorities he gave up the public
+adult animal is deployed and therefore in a sense of equivalent to this respect
+he's the same for many of his protectors
+the germans flow the ship with two explosive loads
+the implantation is done under the mita by a very fast operation
+the county is crossed by the darling river and by the mitchell howay
+he will however be freed of his judge
+we can also find in some boats to fight against the maritime pipe
+it is a sirbot who lived on a territory that is now china
+two other forms the class renan
+first of all the entirety of term is defined in a recursive manner
+he occupies more than half of the island of seland
+the smells troubled the german dogs who didn't find it
+this species is strictly nocturnal
+it is an island of basalt where many glacial calots can be found
+at the christianization these places were dedicated to other saints including saint michel
+the song does not know much excess
+she notably posted this document on the friendshipsite
+the separation of the suberns was then by dicidulation
+he claims he had a happy childhood even though she was born in a modest environment
+hernando felt out of the current pater that his companion was the brother of the governor
+she gop in paris registers in florence and john paul bacare
+he has to face several litters filed against him
+this addition sees the entrance to the vidio assistance at the barracks
+he presents himself as a crystal red orange and salluble red in water
+the city is now replaced by other more mechanical processes
+here the man who has not excavated no statue have been found
+a movisan anmases in the abra
+it is a journal at the reading committee in the solitude is solid
+the village is located at forty kilometers southwest of gabs
+wav the points ito the country with progressive democrats
+it is just that it was not a good thing nabesard
+the four men are forced to study the judge is the idealogy of the regime
+he less the spirit up to join the paris diocese
+alfred verrarin is the cousin of the poet emil vern
+he also worked in moto logic topology and theop
+he was reserved by report to the neres tradition by raymond efin
+together they are going to find this tenager
+he was anomated by the king of spain
+the letter publicly announces the name after the leaders funerals
+the information may be the fourth counter power of real novecy
+his work is the most invative work in the final law
+the versions broadcasteef are produced in spain
+finally chapl moved to tacuma where he published the bal
+nowadays the rear peachers cross alone
+as a translator he adopted american more than seventy novels
+he lives with a master's degree and a young brother
+these various quel fires seem to be indicated that they belonged to the bourgeois road
+by seeing too far the sky scrapers of such
+manobluff aspira is antimolodia milica demerca
+this category verm indicates the aspect of the verb
+they are creating a new city there
+the mark on cos married katrina macbin in italy
+mik jones had two children with aexter alexander and anible dexter jones
+he plays as a winger at the northwest city
+the traditional lanetary is represented by the friend charles bourgois and christian
+this creates distinctive water layers with different physical properties
+once arrived on the spot they were all victims of deligion
+it is similar to the ones of the giant rocky
+after his mandate of chancellor constantine ferribach remains active
+from his action he discovers the creation of popular boxes in vienna
+on this location lived nobody
+that's clear that's clear
+the french engineer jean badello was the leader of the project
+on the first image on the left the sphere seen slide
+this body is now held by ronny oslivan
+it was inspired by the new brand new guinea of marseilles
+he was the rake of lapellation of the space jumping
+there is a cyprian coast of several selections for the youth team
+he was then grenvisser canon of sat mallow and doctor of sorbon
+his work is essentially located in maraquibo vezela
+his poetic work is placed between brok and lights
+it has a large beach but not a seaside dict
+arna miletty studied in paris
+if you don't want to go any more it's not you
+the two of arningen is also war against the count of savoy
+it is his third consecutive
+he lived in the family house of montaville house near jeppard
+we notably put the origins of caroga
+however the term refers to two different jobs on these two territories
+is there any appearance at the city bell music
+he plays curtly more in the club
+a large part of his works are currently at the museum of prato
+will he become an alien who is very volved
+the herbes are retrieved as a grop to help the sandy soils
+it is broadcasted on the international level on net flicks
+he then played in the cido as catedilasian league
+one day the driver of the roller compressor sir j took his defence
+the urban area extends on the same communes
+the municipal territory is crossed by the labra river
+then the term of uperters altendant replaces the pastors
+this album recosenated the fans with the news and the other ones
+she's the wife of the writer emmanuel berner
+he won most of his titles under the colors of russia
+toratiere are holding his way to turn off
+she is the daughter of george and martin baker a family of cuckos
+no this entertainment of the guests must dister the debut in the best region
+the emperor and his wife in particular because of the day they were found
+his mother amelia plays violin and his father had a symbolic
+besides the meals venomis studied for his pharmacological applications
+he is accompanied by victor campner and decnut vensilstet
+the work is played nineteen times
+a patient could call these new friends from a better voice
+glenardan did not stand up
+madras was a fill maker who explored all the possibilities without exploration
+he was made by two architects of urban planning anddetecture
+he was not elected
+it is the leader of a social movement for defence of women's rights
+the golden currency was made of coal and flora or cold cakes
+this device has a lightning phases following the gold ore poster
+later a postman visible in the cinema where he got alongside
+next to there is a pagoda
+the locality is populated under the instructions of brighaman
+at the time of the restaurant steve dyan starts to argue
+he campaigned in germany within this army
+besides tuberculation is reduced and the almost known plates
+the roof itself is dominated by two domes
+he is made up of sixteen arches in the open class
+the lauriate receives the financial award of a crown
+they are finally shot by the inspector redford
+he belonged to the lorraine house
+he teaches at the full time at the university of lebanes he teaches at the full time at the university of lebanes and in private universities
+a broken part in forced environments
+when he night gun gazuba stood up
+in six ethners was born in the family of offices
+the book is reading and the one to be put on its active
+he spent his night club in the charlotte's apartment which overlooks the arch of triumph
+but he has nothing more interesting
+among them were many musicians painters and singers
+it would also be posted intelligent and have a strong personality
+he helps the club aof the super league
+the patronal feast takes place for the sa norand agremanted by bell ful fer facto
+she also has the family in the united states
+it could be executed for instructions by a corps
+the text is written in the excursive of an expert hand
+the illustration provides details that did not give a coffin
+he arrived an apper in a tember
+general electric is still in this creno
+the message of the prostate can also be practised as sexual act
+the city of governor has a village called governor
+the song is presided by the senate president of the leon bochon
+it is a subt or lighding plant
+we are facing the situation in the middle of the arabians in morocco
+the immense cupal which was planned was never built
+the function is then broadcast in most western new classes
+the culture of the spanish is progressively transformed by the roman occupation
+this device had only one vague resemblance with the two previous ones
+he enters the composition of the gico salme which is mainly the mainly of extra sylemaric
+they are toxic as all the most marian leaves
+in france he met n born fermav also and his wife
+the municipality of sat etienne de eliers is located in the roman apartment
+it is in this art that he reveals his most personal style
+the tight shared between the abbey of base fontaine and saint pierre
+it works thence to a set on the broken down stream
+undre is born on in petesti romania
+he quickly hides to a tribute he quickly hides to a tribute housings to the water and lost the reputation
+dis my two arms have enough to you
+stand to pity for the man that nobody thinks he decides to help him
+the fact is that this technique is suffering from two faults
+he was a doctor in la and a doctor raton sciences
+it has ennex and a species that includes butterflies
+since its release the experiences with this album hase been very positive for the group
+the album is influenced by classical music and traditional union songs
+it was discovered on by the nio dernes chatel at an expedition
+in the middle of the decade we are introduced r r for interception
+the dinaanes of the atapaskan language were the first inhabitants of the region
+many are committed to a trade union in the cas marius army
+but cobat cogtin commissioner of the foreign affairs was in our opinion
+she does not take into account the present in the liquid or solitary
+he had to accomplish this little miracle by using his mose technique
+infections vlithic are actually possible and look frequent in these areas
+robert carson fought firmly with petif in the church
+the villa se charles is a public road located in the av paris
+because of these constraints the cleaning of the judge could not be justified
+in the small churches the sanctuary and the heart of the case
+this is not a year of aspension
+the bully is created in the teate dechans alise in paris
+he won several times empinch and the borderly
+it's the caterpillar that winters
+released during a christmas period the money was used for fire
+it is located between the general house of sa pal
+she participated in the championships trandan ara shancar
+he is a member of the buspet sout korean banton boys
+according to charles bell he was not chosen among the hills
+sudiets appear in the history during the gol war
+psychologist and social progress of the polish and wigers in the united states
+even though he was known
+edward vendell follows the career of the administration
+this species is found in theopen algeria
+he then became a teacher and a portrait teacher
+however the investigation is going to dispel any inter optations
+she started playing music from her age
+the initial training is made up of three cycles that can be done in alternation
+in parallel his missionary activity is very intense
+their musical style combines its psychoanalytics
+lesions are appearing as abliary as most of the eruption of the eruption
+many books indicate that hero and lordo vova
+doctor inre graduated from russian school
+we have also attributed him the construction of the first church of france
+now she's concentrating herself on boarding a picture in france
+for each of them there are many solutions in korea
+its back is very dark brown with dark spots or less attached
+however it is the precursor for the synthesis of the salmon gas nertoxic
+he was an actor in a few movies
+activity is made for the departure of an artistic career
+it takes place around the abijan municipality in the department of
+the village is close to a river that supplies the sworte
+the movee shot in the bats neighbourhood where he lives in pierre fitersen
+their ponsor to question williamson's car on these wheels without success
+charles gainsburg before this information had to play the role of the title
+he was immediately released
+he put an end to his career as a cyclist
+it is a referent spot for the overflows
+the city of tacama is located on the ba
+among others ther are productions for sir
+at don he felt his nose on his neighbor biantia
+to scold all th equaty of the employment as necessary
+he then established a chronology of gilustrated people in the united states
+the episodes of the ones were not yet broadcasted they could not be chosen
+this is the palace of the governor
+these plants are annoyed by war action
+there are other systems in mena among others
+the restaurant opens a panorama on the baltic sea and seden
+other stores on the same old fashion also exist in saint petersburg and kio
+on the same side of the same year his mother died
+his life in canada was dedicated to writing
+he also writes regularly for the finished publications
+alternations other ensines are used
+antonio and juse decided to give up to a prostitution
+he was a member of the progressive conservative party of nova scotia
+then he was helpful to the high school he turned to a tlby association
+after the elections mark donald becomes the second time primister
+the impatience made the three quarters of the clock
+one of them is not the other
+they are broadcasted in light rooms in order to see potentially disrapting elements
+alset klible develops a hydrogenic hybret engine for cars
+the price underlines exceptional reactions in mathematic science
+he became the great classic of the mountains
+they sell their duels from the government and scientific works
+she is sometimes gone by imagination of to menina
+she proud castid on average a movie american or other sub titled in french
+it is then as a man disappeared
+the urban unit of vernon is the urban pole of the urban area of vernon
+before the eir of professionalism he is an officer of the union of martin ramshere
+a sale is defined by the three base factors a b c
+max livian was the son of a german dealer in moscow ludwicklevian
+all models of the uro cardastrines and tracker are produced there
+in the championship alonzo keeps ten points ahead to five races
+these signatures are followed by the intangible figures or in order to figures
+only three meetings of this type wer organized
+alas we are all this when we are jealous of it
+he is named in guinea on the banks of the canoe and cammori
+his best performance in granslam remains a role in garris against otho
+the honey from bavaria is famous
+in both cases the meaning is a gathering a meeting
+it meets at altitude on mount meningoba and plateau
+the latter allows the commissioner to investigate in his special and luxurious apartment
+the first episode of ary tokon the fishe whip sigt
+his appearance is different from the one of the wriders in his own country
+the second part of the name is explained by the proximaty of the city of terison
+she's the daughter of the spanish immigrant david gondos and basilia cary usengos
+he was previously dropping in a smaller version
+this statement was announced and it was not the case for the following year
+she was inserted in the tournament system if w she was inserted in the tournament system f w esling was trained a supporter
+he will attain the rank of major
+after he entered libya he was wounded in a carpenter
+he was in the street mainly
+between these two inventable facts that are always a complete conflict
+all these conclusions are well preserved by the government
+only the seaside road remains
+so return to america the friends of albert's friends interrupted this marriage
+she also practised the obslig
+believing himself as the influence of a bad dream he went back to his cabin
+she must therefore be able to find the money that offers him
+interpealed a the m ps chamber the president of the council mustamasment
+you were called for this case with your first time in this case
+other impacts have incised more or less deeply the northern parish of the crater
+apprentice at a doctor he meet his future wife sarah elmy
+it is the case of the particularly of the breeding robert margor and the alberton
+his father worked in the university centre of walcuttos and his mother was a nurse
+contrary to the nilta the two medows are on several facts
+to be effective it is adapted to the system on which it is applied
+the toche was then one of the most equipped with south atrica
+a version of this manual is still authoritative at the time in ethiopia
+despite internal di despite internal dissensions the congress would go down in a fifth consecutive mended
+some months later he was appointed to the new france troops definitively
+it is composed of a mixed mineral and a gippon
+it belongs to the kenton of lulets
+he writes meetings in the parisian metro with a beautiful young man
+the chapels are more late
+the mosaic is located still in the southeast gallery
+the viticulture occupies part of the agricultural activity
+there would be confirmed if the construction was stopping on the day she died
+but the insatitude on the precision of the measurement of astronomy remains important
+the collegits went to bellnave and the high schools to saint poor saint ser
+the novel was well known at the byzantine era
+he is threatened by the loss he is threatened by the loss of his habitat
+she crosses the north of the department of kote she crosses the north of the department of kotor in the burgundy region
+it is the smallest appellation for the ruston language
+the course ends with its starting point by a gothic library
+he is formerly accessed by the municipality of forender
+this city has known im ien spain
+the plane was a monoplace less than a low cost
+the second designation is so happy
+the first edition is published by john claud letters
+under the square is the station of metro monisterati
+about ten primos were enough to insure their appearance
+the winner was attracted the gate himself by his lover
+the government of plussed the date of the emergency he said to open the policeman
+the church of litz chapel saint medard
+the selection intrance ship is made up of a six months training in the national assembly
+it is the general way the church in album a significant production of marimona
+among the churchs among the churchs paintings two of them were track
+he's going to die and believes that he can cheat
+macker macker stage is being implemented in the middle of the world
+it is accompanied by her dog harry a horn past with whom she communicates
+my satisfaction could listen to me about pious foundations in my old days
+but in this case we can voluntarily speak of a french direction
+trash can allow for the collection of the incounts being functioned at the crysock
+after taking off the bill at home kennon brought back anrica at her place
+it is at this time that he put an end to his career as a director
+the author himself will translate it in arabic
+it is classified as a result of these three franchises
+he is the brother of the former federal adviser semel schmit
+its administrative centre is hangda
+half of them form the district of rosemond
+the region is made up of several minority dialects from the usula
+it is two syllaets anthropological siuets one door in a bell
+the list was invited to the government of the collitian by david bengorgan
+a second generation is sometimes marked in the turanian countries
+aram worked with the architectn classically nicholas chantimony
+the lower part is decorated with small and vigital notivities
+for example at the pendes poit reached it thence to commuttors of compilation
+there saint pierre appeared in the engero and healed her wounds
+maybe we should walkin a little faster
+most of the prisoners came from france from the soviet union and poland
+these performances allow him to extend his contract with the club
+each of these trails is a number of kilometers through several uvion countries
+when the war broke out the tarvin rich tuffin was at the school of gendik
+it is the leader of the history of the country of gales
+a first album appears in the same year produced by gilbert castro
+it does not surprise me he left for that
+the following year verbert presented himself as a revisionist
+he discipulates himself under a train to cross the germaino french germaino
+travellers are admired the discipline translated into the statement
+the policy of the exterior policy of marge was servative
+since then the beginning of huminity it has been present
+several of his values have a sick in the paper
+entry is a cold and metal assassin
+he was a member of the freemasonry
+the four members and the writers compose and arrange their own songs
+he is one of the first fill makers to practise lithography in france
+the channel is classified first of media library penerade among the audiences
+thus in almost every country manages to win the case of the movie and to the west
+it could be consolation in a deep destruction of an enormous pieto
+he's a staggering retilum that concerns various insects he's a staggering retilum that concerns various insects and other atropoletes
+some allied reports indicated only one of the prisoners to take their destination
+in fragment a third could be rebuilt
+his magazine is mainly made up of iron cookles
+over time no dich is organized in other regions in medina
+but but it more and took redder and reveals the fro
+the newspaper strongly censored was brought the printing of authority
+it is therefore undipulated by the top
+that same year the career of jesson becker takes a dramatic turning point
+the creativity became the competency of the leadership to any organization
+in a hurry this capacity is a major element to the organization
+the first corps of the population is often of german or swiss
+he accompanies the new consul of nitendo during his urian release
+the fingers won a contest to participate in the top of the season
+thus the clubs owners decided to sell the franchise
+many associations student laughter the school's life
+married in father of two children he is a catholic
+its population requires its dependence from corrupt and arbitrary of the earth
+but on the derave's decision the reporter is not broadcasted
+this was designed for this was designed for another paralympic style
+some go directly to users others sell it only by the recordings
+piere lorda is the son of business men florist lora and salinmot
+drew died in the state hospital
+this proposition did not have a follow up
+expired by all this would gabriella and left
+the shans then went back to land
+he learned ba's monastic
+she will celebrate her seventy years of career
+the community of municipalities in the country is organized around saintes
+he married the united states a young woman from american superstitious
+since then the erudate in the quodum operation is challenged in con
+she marketed more than a hundred artisanal cheeses insix
+the supreme court within the collision was the capitus of the membcl
+similarly with two tickets on each side the last one is motes
+she had many evitars based on the byers and transformation of statues
+getting on holbert the young king made a stop running away with competition
+pascel molipier is a member of the son of francesco
+it allows for the practice of the air volt
+he founded the catholic college saint benito and the turkish base institute
+this one wants to be a professional and plans for injured players
+this does not prevent him from moving towards his twins
+women are particularly affected
+in this case he appeals to the judicial illitacy
+the muffew itello american carlos carlos marcela family comes from this commune
+the artists of these regions then turned towards the capital of the new empire
+besides the game is compatible with any buck specially designed for him
+many were volunteers not remunerated for their time and contributions
+this treaty finally took t with the beginning of the second world war
+he's a member of the integral part of the belgian history
+despite their opinions some were in charge of the financial administration of their fief
+some critics see the best writer's best novel
+after three years of marriage a woman discovers her husband is a faithful
+a poor animal had survived a few hours to his master
+he currently plays for athletics national at the forward position
+because finally he continued his canals have the title
+forty per cent of the titles are dedicated to the field of religious
+treman is concerned about the will of you and the progress
+adella during the afternoon the grini attack took place the austrian centre in finid
+it is kept in mers backer constafe and cotnard in switzerland
+the climate is semered hot and dry
+salbret is also a singer under the pseudonym francois fabruse
+the tramp is original music
+some corticoids can also be used
+the government regional by the resolution of numbers
+and why captain she asked
+the game is divided into a level series under the shape of simples
+the latter is carla where several tons were remixed
+both creators are played in position and especially destroyed
+the following season wither mal records the gold medal
+the system also adapts to the wooden metal flow that containing wood
+one day they rit up far from their squad and not reverse
+mock miller was born in montreal and grew up in wismouth
+the schooling is open to burundy and therefore not to date
+the landscapes alternate the wood and meadows
+his list wins two seats
+fears of attack are also at the order of the day
+each osteoderm has a central range pronounced and a reorganization
+in an elliptical shape elliptically lying to the apex big and pointless
+chane will then appear several studies on the medical interest of mineral sources
+a tundo of international boxing bears the name of tervick
+he owned money with the help of bidklas
+untill his death he tries to improve the situation of the political science
+tony brooks saved the british honor in german his own place
+however after the revolution it was changed by the lest
+each player are now a last fermal card
+this one also has a clearer browner than the human one
+he told himself that he was a political and was the author of many books
+this type of co havitation was common in some other regions
+weare in the same year she attacked wer in the same year she attacked the crimine conac
+two years later expositions collectively with erabazki
+one more day later he is married to the two races of his father
+the convention used in the calling of the gregorian calendar
+and he was named after astronomy metallic sin giovanni ciaparelli
+it's the timor mlsar charles the fear of the skull
+it is located northeast of the village on the road taking zube
+one day like this all of us are getting bored
+tryings invidious sculptures are done on the work of the rites of a very good friendship
+everywhere they look for seeing the sea and his fish
+to day the industry still plays an important role in economics with the air force
+the royal couple of england had given birth to nine children including them were born
+buzzle was erected in county and sold at the museum
+he became a member of the physical education teacher then he did not hesitate to be a little
+he is placed under t e jurisdiction of the city of leon
+marie marie cobart in london
+its presence is uncertain about mount ku
+at the end of the dynasty of cuni furina returned to the end of the story
+the victory won by brenda and frank can finly finish challenging
+he is the letter l di critical by a makron in a caro
+the introduction procedures and selections are identical
+comou located in the pas de chot
+the fire did not take a bit
+another example is kept in the perga museum in berlin
+there are two approaches of the snticis of the carina
+it is therefore with this technology that the first version was delivered
+he quickly becomes his main developer and in charge of maintenance
+the earth is therefore mortaled in microheda corresponding to the micro clima
+he owes a careful assistant to this one to favotize david
+he particularly distinguished himself during pantadas fights caster saint croche and
+she is the seat of the poet of acs
+and benefits of this priory goes to save martin duton
+it picks it jealousy
+junior is a young lizard
+its ventral side is white with a series of small stained black spots
+a list of names would e simply be too long
+he died in montrous
+lashman goes for africa where he works in circulating paths
+these results are considered to be insufficient and they return to take care of the training center
+he is the subject of an inscription to the general inventory of cultural heritage
+the americans wanted to be deported
+interpealed on the question each one rejected the accusations taken against her
+exces e door open to the police who found the automotives and the car hit him
+he sought after catherine decien in the middle of great festivities
+louis lebruke lived in london and south of france
+his older editor is barzak belonging to sophia and salmoja
+he is the father of alexander ramutu
+it is true in wist who takes the place of nelson
+this one indicates that the baby frindas already dead
+he went to the closest of his san invalet hok
+the latter is buried in the church which took his name
+he scores a total of five goals an half of his goals
+they are also firefages or cannibal
+he is even appointed captain of the tor
+then he had a real power and pill made diffusing in the regions andtil indens
+following the modernization of the syrian alphabet this letter written from no one
+in this episode we also learn that he only has one testacle
+the island mercer is located in its southern part
+hunting probably caused his extinction
+they're rather polivat rits
+following this episode louis m p d went back to try his son in europe
+this species is found in africa and in the maloi
+his population rises from the census in a large majority of native americans
+the spectral imit allowed to reveal the original scores
+this can therefore be introduced in gitties and ifformation exchanges
+ttheoretical research is in this area is important
+this interactivity can be distinguished from traditional larse
+she's a member of the whittlebock family
+denisen was skating like every speaker in the individual category
+leo sin fili is the hell
+he doesn't he doesn't call upon this decision and put an end to his career
+for him to be intermitted the finest must receive at least votes from the committee
+he was present during the bombing of inberg during the war of crimea
+the ceremony is similar to the grape vines
+esali and his friend are imprisoned on criminal accusations without basis
+roman road eper there is still the gallo roman era
+the exploitation so heavy would make the departure with completely automatic materials
+the record was to hold for fourteen years
+it also uses interruptions to commute the tests in multiplication systems
+the battle of qimar stops in the same year
+the bridge rests on five arches
+she has a younger brother alan fabian de louvre and a half brother antony doklin
+the attack goes through the air can be developed against the four payments
+the school was high before located in toulon
+the notion of seasonality is also taken
+he was the father of general brise
+she would like to become an actress later on
+in classic classification the super class of the astios as corresponds to the top of the list
+edward did not succeed in the countries published by his alive
+he didn't he didn't hesitate
+his fuel was housed between the blakes of heroes
+the castle is now a museum
+burst into a hot flush
+the marquis he stopped asked samuel
+from the jurisdiction of the navy
+she spent some theology in medical biolithic certificates
+the company dealers is hierarchiized in age class
+military famous for he settles in several places transmitted to his brother m
+it is the largest known villa in the region
+his career under his new colors started shaking
+file the limus didn't file the limus didn't file the limus didn't forget his discipline and found the office of the landriges
+on the religious plan this locality was an ennex of the parish of lesbo
+i said the customer bucket that arrives at the trains wednesday
+he then became a coach and sports manager
+he is named the elogy an philosophy the same year
+he practises ordinary jurisdiction on the priests who depended on
+a sealing fixed on a topic of the seasons is kept
+he is the one who is announcing him
+his population is essentially pectuna
+scrintus others are subitus and their hosts
+the university is named after julian the university is named after julian marchleski and seats in moscow
+although the game system was slower it seemed to be the original
+atania brings that dmitrios managed in luxury
+you were protected by luklahoma associete is open to the public
+chily is ren second in the tournament behind brazil
+he is however the first to establish a real chronology of the herrik period
+burburn then up protectort british indanesia and philippines
+to sandherd he finds himself forced to hear in express tenages
+this former piece of the ustracon was transformed into a farm
+jon darmania case is the left side of the already weakened in practice
+according to risstebri shever sent them a sexuals in the cam strabes
+this species is endemic to the island of rot in indanija
+on this one the group is accompanied by rlistard
+in another way he transports the name of the missisipile and the urins
+pittsburg pittsburg holds his name from william pett prime minister of the united kingdom
+and in this order does not exist
+the chapel was founded by thomas de villas
+the municipality although the shelter has no trip is part of the famous scholarship
+and these publications mainly caricatures
+this edition will be the best everst greek information mission
+he met this fact
+for these requirements she practises various jobs
+list ambassadors of france in barbado
+he was born in paris with the polish jewish parents both little artists
+to discover penton we believe in a little bit of surprise
+he became a plasance port
+the neromorphology is morpos which are both different
+a mosaic in the ground in the centre of this small room
+they can also take roots in the long history of a feudal thief
+frankinthal belongs to the reina carmaripolitan region
+the gat is of ogible shape like the gothic style
+the flour is used in colour particulaly for the manufacture of pins
+historically the mountain has always been part of the vallet
+mon magne died in the end of the first
+the film provokes critics lodgings and received many awards
+it's the return to the middle ages and the total method
+he is the father of pilet who is designed with anixibe
+he is buried in the fallenville cemetery
+the rock remembers the last two men to exchange the memories
+the whole is accessible by a large path of walks in terms of battle
+the interrupt system is placed above the blinkr of two verin pomatic and cylindrical drains
+the company employs our fifteen collaborators
+he gets the sleep by touching a person with his paws
+the largest part of the lower silica is covered in mosaic
+the regression model is the most famous recreation model
+he wrote many books on uperita for the composer claude terrac
+the porch was burned during the wars of religion
+it is one of the historical concepts of the united states
+he is thus he is thus frozen in winter
+then he served during a few times in the military division
+in parallel he continues studying to become a public education teacher
+he took his daughter into the garden of turis
+this society had been founded by two former employees of apple
+robespierre accused her of complicity with de maruse
+he is replaced by wadin porte
+this strategic position allows to cea siligr to knowa strong economic growth
+his composition teacher is leos genetic
+but he dies without being able to complete the last rehearsal
+he is based in the city of madang in the northern papo negane
+he was still a part of the left centre in the upper chamber
+he is the patron saint of the city of gap
+he was minister and president of the council of the kingdom of serbia
+the community of barchells is part of the first
+the driver of automobiles is having a best perception
+that pursued a musical career as a music career
+the association supports the divorce populations on a city of the scenes and the great
+when you are accusing of violence they are coming to his piver he is innocent by justice
+the first royal denish is part of the two thousand professional funds
+at that time the division was trained on the island of papovo to the canava
+these nicholas took him an oath
+he greeted her before leaving with his tourisn
+he went back to vishy
+pon pani coculapa is right in front of him before being clapped by tigo
+all the previous properties can be established in analytical way
+she will have to fill the mission to fight against the human forces
+that's where andelet stopped running through the next year
+she is also the mother of the football player in ningascaniden
+richard is going to succeed the contests of the outside and then in the intinite
+at that time menu was in glimts and piano villages
+the public messages of egypt were shod in three days
+he is among the remains of the vebboji the most famous
+the operator is a key element on the pill
+the male has a sibermal twist with blue sides
+he is restored and once inhabited by the successive owners
+russell is spraying the municipalities of gallards la monteu and embarros in the guard
+through the past as iftused a processor invented
+the last ones in the very plationg but are often composed of amantines
+what's the matter to do asked paganel to the padagon
+he produced them in a zone located between rubican and mercha
+they are still given and followed by a regional delegate
+she shows him the sun he appreciated
+en he practices gymnastics and judo
+she was the favorite singer of muster facimeltte
+the horse first accepted to be approached then accepted the school
+this technique is used for the characterization of the particular trinals
+they were named after the place of their discovery
+intermediate was crowned champion of italy for seventeenth time
+the most often currer fiend was the office of burgher
+his father tene ren foe cande was a spiritual leader
+he is in charge of the command of the coasts of the terranian
+therefore they're struller in five
+his career started as a coach for the first time
+doctor in law claud delmis was director of france and spain
+professor at the university of paris sarbonn he focused on the study of buddhism
+a french studio is still named julia prenny
+it was the most built moto
+at the national assembly loron bame sits on the finance committee
+these lists differentiates in a few notable points
+the taste was raised in sapierre
+the street operates a georgian law a few meters before the pt and sex street
+the new house is built for a moment of books
+the two sexes have the top metallic
+this latter represents the old hunt in his commercial relationships
+in addition to his career as a painter polly is the author of historical works
+era hiddle joined them the following year
+we know that he was we know that he was trained in more than fifty men
+she serves as a maintenance guide like the support of technical and economic negotiations
+still in paris he became the chief editor in the paris region
+angelis is inspired by music and ancient greece notably the storia
+the troubles of the independence of kameron will vibrate to finje and his population
+the temperature real is depending on the atmosphere that is still unknown
+other r j orpy have not a horn but a rostric as the fish
+hendrick freshford was born in edwes
+these species have only seen ore almost all of the heritage
+in france the ministers are members of the government
+it is a professional magazine destined to the nurses and nurses
+he will not be too late for the arts and especially the painting in order to increase his prestige
+in uption he can be equipped with electronic visor
+the first six copies of the new series were delivered to american airways
+about the quintettes with piano
+he is then in charge of finding the unreasonable habitat for the city of graen
+the cury was then insured by eric de paine ben rudenal charters
+josus well finish sevent
+he's one of the pioneers of lithuanian history
+their harden oviprous miseries
+the new orleans and museum is acquired by a letter that you're in the cradle
+duver fights on mazirs
+he is the brother of philip smuromin of three years his eldest
+six cente del six cente de la massage martile ribnon street
+indeed the kenton is bordered with witzerland
+father of five children he married the cludsite winner three years before his suicide
+this career was interrupted by the prematurely by the society
+the sword meant a war state and the book is not depleted
+qualified for the thing of cat you ere rakeson
+he is a secondary member of the supreme court of the apostolic signature
+military credits are motivated by the nationalis china
+the latter turns us into death of justice
+it is then destined to join the tropolitan
+as if he wouldn't have been married to go together
+this workshop continued itx tension
+he must stop to repair a wheel of his wagon that broke down
+chiperi furius was sentenced to exile
+of all this he was insulting that wittery
+he enters the same year with snyder
+hbbd merk completes the podium
+hydrolamine is a compound entirely explosive if it is heated
+located on the course of the river hidten
+kacing is a member of a group of the time called fremehen
+a single fin burg was born in bedura
+the intergenary sector therefore takes it retirement
+gisari is home to the sires tour which is the biggest tour of finland
+this pin will last long
+the first score is bellwinting on the same game
+only a few small sections are still visible hence this lock
+boramini was accused of having made an unstable building
+the director was sentenced to an almond
+it is the stellar closest to the southeast system
+it holds a bachelor it holds a bachelor's degree and a specialization in the university of ottawa
+the troll still lives under the bridge but only one has been there
+this first season is composed exclusively of independent episodes
+it is one of the greatest pictures in the world of lyra
+he also directed a few hundred sculptures of the private collection of the collection
+the last club of each pool is directly related to the division of honor
+this text is not signed by the r l ferrick brigade
+he is also the president of the sars association in ugonda
+couple searches a significant portion to play these three passionate football team
+he is a member of the graphic alliance
+she also lounged in painting on porcelain
+yoshkevitch was born in nodessa
+he grew up in lorraine where his father came to work in the citeology
+users are grasping a question of calculation demand
+married two children and the sons of the family of the two children
+the pers should then take assault and military fire
+well then i'm going to ask you a market
+a real stumbling in love with him
+yet fast chair affirm that the two entities were completely distinct
+except that we do except that we don't leave time to answer to the benjamin
+all these products are specific sources
+beas made him come to home and forced him to sign his resignation
+two elbums will be produced by the duo
+the material is not provided
+the crowd was heading towards the government's palace
+there is also a stage of nordicais
+the history of the lovers begins antiquity
+many of the little ideas are entirely artificial
+a major and richard has the opposition of monks
+sate urbane is the starting point of charlevois crossing
+he is trying to improve the design in engineering
+he brings them to marseilles and then in monte carlo
+he married marguerite de rosembes
+she was snowing and sometimes credited to the credit under the name richard enger
+she participated in three editions of the olympic games in the summer
+in parallel the kimmers constitute a market for more immersive experience than monocran
+the sacrice is funny only when i saw plain
+the documents of the aro muscle as a painter and in light
+the fossil remains discovered are not very partially
+the bancali he became a member of the orders of sa jen botes tov in nice
+he practises martial arts an clevrio
+woks by shelling are kept in the staite grafi sat non
+the three is located in engeled and until the revolution he was presumed
+his career as a professional career started in the carman mora company
+the sea was over but the sun acted against his own will
+it is the triumph of the etalan
+witcraft participated in the second world war as a banker
+for the different eres they are equipped with yellow or ends of the car
+it is an action that aims to reduce the prehension of children against the white blues
+the tower dates from and was restored many times
+this species is measured for males and females
+during his return to oberwell he became the captain of his team
+unfortunately attacked by a group of bandits he loses a arm
+a montife a montife cho morif is carried out through the clothing up to the substrait
+it also occurs at the operas of san francisco in chicago
+this club seems to have been promoted to the lower floors in
+he owned the canier castle in sensuns
+she worked as a consultant for the university
+montestri is a village of baarn located south of orthes
+it is on the other hand to use the technical atenance
+the vivre is integrated in the kingdom of france
+the spectators are not really waiting for the movies guard
+he plays as a goal keeper at the alon salima
+this system has a clean movement
+beyond saint martin street the street is extended by the anchor passage
+he continues his work on the alternative theory of sets
+they will then teach with the tor to promote this new obis
+he is the eldest son of charles etienne de tescin and marie therese de hascane
+the tor is divided into four parts including the last one
+like these predecessors immediately he maintains a lot of money
+he was discovered in bake by the observatory
+information about his past are unknown
+frendheim has won the title mister
+he will be distinguished from the croi de gare
+this amount varies as well as the pepper is transported or domestic
+do not pay attention to the child with de champes
+he is the son of joseph bastian and sculptor and marie louise desil de brec
+the municipality is not a station
+the sequel of his career will be less amusing
+on some of them a third non identified character appears among them
+he had representatives in gambia as well as his lover and french guiney
+the stars won a fifth consecutive championship and a third cup
+he became a suffragan of the borges metropolis
+on his grave there is different items attribute to his work
+this species is endemit to the department of esles de laba in honderus
+during this period he was selected in then italian football team for three matches
+a shepherd o
+he is the father of the comedian john paul french he is the father of the comedian john paul franko
+he is a member of the r strait were cursed
+he had animating the socialist notary
+they are now destroyed
+they also participated in many exhibitions of groups
+this album contained a set of sons mainly composed of a traditional style
+despite everything the competition is still organized for the lower divisions
+she is a mistress of novis and takes care of elderly and infirm
+indeed the martin border between indinita and panos is still disputed
+the museum is integrated in the visit of the opera garnier
+this society is highly patriotic and valued
+all the sentence will be objected in parallel
+this kenton was organized around muntre this kenton was organized around muntry in the district of vender
+why not asked the major
+bred in the living of the population
+several of the others had been volunteers in mobo
+the area of j is located on its western bank
+the majority are based in kachasa
+the official language of the city is the city the official language of the city is the city's french
+it is a sport station and summer
+the stadium hosts meetings of sporting club tlist
+she's not her daughter or her wife because redy luf declared herself
+the origin of the street is linked to the construction of the riu square
+it is the case of the topology produced in the category of popological spaces
+the title son is used as a musical theme in advertising
+large valleys and sheer lines characterize the northern part
+the united states team is replaced by the royal team gathering the best canadian players
+the man with the armed elevator is defeated by volverian diablau and colossus
+erth as guilty to that of the high school
+the rue de the rue de leveel is impressioned by the french writer marcel m a
+we know very little about antoine de gays
+it is bordered to the north by the duloc street
+it is a bird that goes on all sides of the coastal environments
+here is no personal objective which represents a terrifying role
+several years later a fountain in granite was installed
+she proclaimed a sad truth with glinded
+but the latter coldly assassinated a civilian who was taking her do in the surroundings
+the term wrynch is sometimes used
+jean milin did not forget his peace for peace
+but what a dozen pounds of cloisters of potis
+we don't know who originally compiled this list
+he is the first black model to perid forceno
+the turret is supported by three fins and half of them
+he is decorated with a war cross with pums
+avina was born in the united states
+on the other hand his age is born in birth
+he is the student of the sitaharo and publishes many books
+this one sells computers alongside the line service
+the choice is operated by a jury independent of three people
+rosea lucy succeeded him as the chamber of representatives
+this structure is not without any consequence
+sni also had a group of majority
+coming from a wealthy family from merseilles he is a student at the julie college
+the site on which it developed the municipality extends to the east of lonana
+the artistic work is based on the research of the blood in his image
+sella is one of the main inspirations of the french singer christine ecrans
+he had as an apprentice uberham boss
+this story is the sickle of codel pulling
+this species is found in australia greensland and figi
+the station was inaugurated on the line
+we are no longer used for local traffic
+in exchange of coet he agrees to attack the dana badrel in a branch
+the duke and the duchess the duke and the duchess of gira and go to the united kingdom
+like his name indicates the village is located on a hill
+gally panoe surfed with return soon after taking care of his country
+the latter constitutes one of the most important rocky parishes of the athis
+i i've heard the iron men calle the decision to the state council
+the rucklands were known as excellent cavaliers
+it has two solar energy stations and a forto
+the kimeroon participates and finished the competition without conserving any defeat
+under the current facade it remains three rows of stones of the old church
+he also belongs to the high council of integration
+and philip trecht meets cathnik and gil and teach them how to sign it
+he works with walter and gr he works with walter and grippis
+its belly is white and is yellow with dark spots
+he is comparatively smooth and gives some occasions to go out
+the pedagogical orientation also turned into account the design
+all the orthodox churches are found in the council of utomini cancil
+the time is to resumpt the economic resumption
+the municipality includes several villages and hamlets
+edward then adapts it to the western armies
+the phrygian seems to have been close to greek
+but tarses are rus garden
+his life was extremely accurate
+during the second world war the polynesian islands play a leading role
+his father is rande and his mother ucranian
+the couple received a grand annual of books poured by both fanas
+the position of high commissioner is abolished and the greek constitution adopted
+each one invested a unique port with the osifage light as well
+the superposition is said to have been a idea of cusven sad
+by working a bit more we show that it is in fact a generator
+he participated in the battle of tardeno
+firstly launched in imagination for paul with the autonomous cummer system
+he retired to the middle of his lens
+the s is decided to have the majority of those members
+the first attack to killon and the first attack to killin and toris judge
+ii'm in the musical context that is being considered to be free but noddy from the field
+minister of defense is at this same level asegation
+he remained a lot of monuments including ten of very ancient religious buildings
+the antigonist is the main antagonist
+francois beleau was a commander of the public health order
+the company manufactures the cardage water getegats and a c
+the first building is the work of the norwegian architect balfasar leny
+the probisisis regressed everything like the lyrics
+these values were chosen for her to be spaced
+the different announces that is gnely made place like the place
+oberliga sad west was dissolved during the creation of the bondelices
+the cabin was in a stet
+he first works as a photographer
+and who knows its longer delay would not understand the trip
+the ornithological circle of lozan is at the origin of this innovative project
+the elimination is mainly rinal
+after starting by refusing jemmy ends up accepting
+the ice wines are extraordinary from wealth and persistence
+a museum bears his name to point noar
+having been requested by a ghost perlet he did the main monuments for the first time
+it is located on the root of your innocange
+aan crianga is located on the road and raymond vullock
+the jewel which he existed four different types is the work of the engraver
+she puts the most precious objects in a suit case and decides to leave
+your sister never access to the price of rome
+he always supported the government throughout his pamentory career
+acoo station is a centre in the algier
+she eventually finished a mexte with a track
+it is translated and played in the entire world
+one day his mother died in a gravel pond
+this three sinister falters kawar and his companion at their moments
+the city is located in the northwest of the country and has about inhabitants
+the language is threatened because it is by the only one by the gibber
+he lives and works between paris and bordeaux
+etien look etian look at them trying to stay dinny
+giani started his professional career in uni hilleria
+these protectede people came into the other fashion of mediaeval cooking
+this plan shows a remarkable disposition of the sismiom
+a roman church named after him
+professional life is deprived of employees of a state society
+he is buried in the jewish cemetery of jerusalem
+eurostat also works with the organization of unit nations
+a county has a floting
+even checkers of h
+he designates the common device and there are several types
+the municipal elections organized the same day as the new report of force
+he is the son of tacitnet for local ador and daris
+the mibilation of the origons is also a office there
+he is attached to the development of fretoter
+she connects the cities of lekleas to jasper
+at that time the moslem community in the russian capital had eight members
+to the east the word neighbouringn was only a bastr
+i'm lying down on the fight without counting the depths of speech
+organ who knows a tropical climate is located on the seclor
+similarly the sight together
+under the test he sign a belief that etien under which it is arthur
+floring it floring its late in the end of the season
+cheese ulified by a good new voice
+the same month he finished fifth in the race of the rezints
+in this route the kudan does not cross the very important urban centre
+their properties are seized and assigned to the trainer of children
+it is double national for french
+his opponent was supposed to be cibito or sangren the tournament was interrupted
+indeed the testimonies describe three different ways although similar
+this accident marked the urian countries
+we were inflicted sometimes at the animal the italian law
+it's time for the first prayer of the morning the charge
+he is the father of the historian franco vetory and his father
+he turns away during his night at iscupe de menso
+the district is also a district of initial
+beware of the excess his life even if it is not liv
+the hunter of prems tracta group of bendits across the west america
+it is a monotypical species the only one of the la fortina type
+originally the street was only wide
+the works of these last three were not published
+it is directed from them by the independents kurennese
+however several slopes retain it locally
+branon was first and much to the court of brenan
+thus all the flowers are therefore thenfiguration on a total
+some species live in great death
+no mechanical process led to results
+the show is broadcasted every day at an unbely
+the competition takes place in three phases
+both titles are written by the french
+according to the portuty the aredes used the harassmenter charge
+in february he joined the israel team settling academy
+the next day rina dullish was released
+she collaborates with losland and she's from wolf interior designer
+he is opposed to the contrary that menace has a strong influence on the germans
+sometimes the places of the person were added
+backthen the day he was mainly playing in bars and coffee theatres
+canisus combined her difficulty the situation and authenticity of the nman
+it is the palmir who has an attack
+a meeting will change the course of his daily life
+three years later she follows a training in order to become a teacher
+adding omer he is in charge of the sustainable development of transports a stationing
+he had married fen livasor daughter of pierre emile leviser
+he is a tributary of the aglio in the pao basin
+initially the election of borges had a considerable fate
+his development was stopped after the renewal of the license
+ne trok operation was its headquarters in the city
+admirals are attributed to him after his death
+it is the weverbrand which was crowned champion of germany for the third time
+he is then allowed by china to trade din canti
+he plays as a defender in hoin united
+some describe them like the termites inahead spirit
+the vases are converted once again into port
+marie leo runs mouth oliac for twelve years
+he's gone from sapines and arrived at the bay batto
+denny and terrry reconcile at the end of the game
+that same year a fire had been shaving the city again
+since then it is a delegated municipality of the hot brander municipality
+the municipality of flutier is on the territory of the kenton of saint aran
+they are presented here by regon
+he was executed at the end of the great perch
+until the revolution was given by the balick of simor in the duchy of burgana
+theye also added characteristics to remove the network limitation
+he is then the largest pak he is then the largest pat in the world
+the church of new classical style has a rectangular shape and is crowned by a courtyard
+he is married to the director martin trener
+he is a member of the church and sees the day
+it is the only species of the genus a janodelps
+the clup nevertheless saved itself with several days of advance
+she connects china to burma and the south of the greater poland
+he marries general he marries general's daughter joseph de rufre de fors
+to this day the circumstances of the drama have not always been completely eluciated
+in the final she beats virginia waits and the french people
+during a time his president is perjacsiles
+the initial management of the committee was taken into account by emil the initial management of the committee was taken into account by emil franchi
+another role that he is incarnate he saved to select
+this species is found in mexico in salvado
+the artist carrier bellar sculptor artistic head
+philip kostage in massive woods was painted by the russian professor on dregodi
+the device takes place and the crew is forced to jump into perisht
+on the side many casmets were set up to store
+they ar other reknowing the counti they are other eknowing the counties and wens
+the narration is inspired by real experience
+this is therefore that juff this is therefore that juffy was also the sheriff of warler
+his artistic activity is therefore very reduced and unrevenable
+he was replaced by his younger brother georsan
+son of erst hess phillipsten and christian de schwartz
+he supports the limination of the fore band
+the advantage of red ars an andis maintained is quite pleased
+we don't know exactly why mubara passwayn his brothers
+pierre moray forms a government to which four communist ministers take part
+that means put the raid shag in the centre of the imperial policy
+he was elected to the onela county she was entrusted with the mayor frank fezzier
+it is his first album
+she did two passive traversies among other things
+he playss matches at the international cair stadium
+the ocean the ocean's angers
+a second clerk togilay participated in the confederation cup
+franchise is managed by a group led by bryan untersmick
+this latter was obtained thence to the verb and panel technique
+the pressure changes affect the wind
+it's actually the one who works aninch and a cornet to meet pierre
+this turnbile in vapor has only one record touring a tour
+walk is kept at the wenbad museum of amsterdam
+we will lift up some forty injured at the end of the day
+his little brother cristo is also a saker player
+he has a lot of big trees
+the company also manages many specialized services in honor the company also manages many specialized services in audience
+the winner is announced at half time in the t v show
+charat accused the representation of the various words of the jewish religion
+they were expelled from the nik towards the bicel or osper mountains
+he's not considered to be edible
+the minister in charge of the reform of the state has
+the colition will win several battles
+you were annoying shidy
+the first nation identified in the region was undoubtedly the one of the illyrians
+canalio then came the target of fansi canalio then came the target of fan ica sicards who managed to take away from dalia
+on the twentieth of these vineyards is good quarity
+several holicopters black hock are also affected by bog
+he is buried in the unaka cemetery in tokio
+the home work has only one generation per year
+the title is vacant after middlemore refuses to face his on brother
+we can find the biggest grenite quarry in canada
+aves had contributed to the organization of the resistance in the camps
+charlesburg became one of the subendments
+he took place on the molu circuits
+the stipria is the fountain became the source of a stream that fed the abbey
+we suggest a link with the bastard church of can in france
+one of the grandfathers dibert colored was a camp of marie
+the german name of the village was witsen
+the list of episodes of the adaptation a party for these reasons
+it is therefore antagonist a corisets femeral muscle
+fourteen she attends classes of marcel tussancourt in versailles
+according to the book of gemimes has prophetized during the esk station
+his horse rising in pereo enlowed especially under the main points of passage
+the american presence has taken
+however it survived the attack of the silence and reconciled with him
+tunjes trinite is a specialist in metal construction
+victor puddle started his career in the silet sinema then he continued
+julian cr suffers from melancholy
+the castle which would have belonged to the duke of burgundy to conserve its mediaeval appearance
+the date of this details lacks any article
+he is a knight of the national order of merit and officer of academic palms
+of them is brown and dark brown with a greenhouse
+on his study a true repeated on his study a true repeatedly
+od is part of the dan luttusy
+if there will be friendship with gorso marie mathus and henri came to her
+she didn't she didn't she didn't have much common senses in french
+almo saxual she directs a bully game which is associated with the micara finker
+he worked as a professional director in abija
+james carville is currently a political science teacher at the university of tuen
+the old saint martin church of the paintings dates from and is listed as a historical monument
+the movie was shot in nigger
+it is a blue star
+he was the first australian referee he was the first australian referee he was the first australian referee and the wild cup
+he is against the right collective and is in trase he is against the right collective and is in trasex to the groups involved in groups
+chair bork in the hart were preserved
+in certain cultures we can see that it is notebook
+the cartalosisrod are then imerged in carea
+his son inherited robert succeeded him
+the following month he won the denish on top
+the result of the french version to the clayet to the second round
+he crosses the aximo untrifana and enese regions
+he plays as an attacking midvilder for gustav tok in the first polish division
+after two consecutive defeats left the start
+he tates a source in a mar canpac in the michel mountains
+each month of seven women stands at fetive de fare invalet
+this ends in the title and finished the period of the duke's roegen
+this one is part of the cout oa bore before the publication of opus
+it is consumed generally in desserts sugar and canals
+it is a depot for fishermen located on a small island of dumfons land
+events when a vidior played on the game
+clepolin put him in the air and that one took refuge in the vase
+you will be nice sir eron you will be nice sir eronax to visit one of his sins
+the canadian careers of manituba are no longer used nowadays
+the decboran has no significant application even though it is made
+he was released and was wridten and composed by his friend pascal ofspe
+he has a great force and his very fast show on the ground
+bulletnt under muffins at ship bulletint under muffins at shore an otatan christian is a viceroy
+at the place as lis malay the friend of lus duc
+he spent all year in minor leaks
+he suggests that hector took ruins to his place
+in the district of barcelonet in the district of barcelonet it's not the case
+we will see sir but you think that this rese is known
+the elogensky has been offering a ticket to each game in the team
+the city of energy is divided into two
+adaptation also was planned at the movie theatre
+and i pushed him to scare me so much expected
+under the pressure the organizers celled
+the library is created in kofu
+colbet the monoplas colbet the monoplasar rather realistic behavior
+the whole is regularly involved in flanders festival in golona and los angeles
+out of the tatans she is a marshal of the french gendermarie national
+the game is provided with a hiter of scrips and cards
+his larva feeds on dry leaves of runs and clemati
+but this objective will not be agrade one of the fewest leaves
+in this sense he is one of the precursors of the rastofari culture
+commune located along the banks of veron between nedropolis and brunic
+a unitary state can be centralized to centralized barigials
+his father is from laurenzona and his mother was born in centafia
+she attended the saint louis college ten ta saint mark high school in leon
+he is dead in a car accident
+netta a young woman with kanelps is ready to marry
+it is won by the french pierre quonot
+rue valadon is a public road in paris
+both of the communities form this monistery
+brave jean he murmured the canadian
+tom paying a lot of money to unwrapping
+gradually their friendship turns into a deeper feeling
+to get around the problem created by the size of the combustion changer
+the drops of liqui in a fog for example defines a thermal system
+he rolls on the granular bed and resounded his arching lot of grains
+at the back three competments are present in the back
+he then participated in the campaigns of this army
+the nature impermeable of the field prevents the decomposition of the body
+this same year rosa acquired a wife to denbury in connecticut
+fearing for adam fearing for adam's life she flies to detroit
+it is during his stay in holland that linney took the knowledge of the clifford
+the site is the work of e architecture is a unique example
+the bridd is known as a very important colony of bats
+she won various prizes and has been defended against sexes
+the local groups were already inditted the pivot of the organism
+it is the ethnic group the most important in gascar
+he belongs to the movement of painters of poetic reality
+he is named in honor of jean cheorne also an astronomer
+a few private pilots chose to adept occur in a more classic
+he continues to accompany mary matchew in the entire world
+he is then the sixth in the tor diski
+it mocks against peru
+his opponent accuses him of being part of the leadership of a party
+junfrey is listed as a pomentary assistant of the uru deputy jean marie lep
+seven of the gura tukea pignals
+the line as aimed at its exploitation quickly stops
+the zone is located in baarn at the south door of po
+he studied at the university of kennezoa
+during the royal championship organized in bazil he plays three games
+it would be the bormandies
+i heard the gravele of th asperities of the back of the corelian front
+eric busses worked on many medical and sequencies
+he improved he improved his best ranking in the wild cuff
+this kenton was organized around krisan pave in the district of abevile
+he was a music critical and still have a beauty too
+there are many models of pummet
+he also had three eligible children
+it allows us to avoid to lose and reconcile the way despite a little accident
+tin tint was as the students of gransburg hih
+it is in the pressure of his family mariang and marries another man
+they finally obtained forgiveness against the loss of their goods
+he was an architect of the abbey of saint gendemin
+he is the son of former football player alvaro enrics
+circular are decorated with granitros and reverbers are of pediment
+ten pol yon was elected best player of the petition
+this reduction was the case of two seasons
+he is registered in the republican left group
+his arcides are found in the federal military arcades of freibourg and briscow
+he was elected by one of his feelings for which he missed the language
+subsequently other critics confirm the attribution
+the prose of sloff has been translated into sumurian languages
+built in the following years the metropole hotel is located in the boulevard de la mer
+the album is an eponomous album with a li toch
+they are due to the collapse of the initial cavity of the meteorite impact
+she participates in teaching programes in mirrad
+the painting will be inspired by a text by francoirne de chateau gond
+it can be found in the ethe hills
+all are not equivalent to ecclesiastical falability
+after his coach career he stays in the world of basketball
+the latter has however been kefermer due to the massive rival of turnet
+this species is endemic to the island of catendes in the philippines
+he is buried in glen
+he is the french
+in the province of ockland on the deserts of the bulsk
+denis for hoa and pol wima are considered as the first slimber
+in the same time smith is taking place from the cruel one
+she has notre dame de lasumption de la trese virgin
+the venette is the work of the graphic designer roger fund
+it is his first goals in prose
+the teacher ping has generalized the principle of starcastic maximm against sacastic optima
+the first threat studies is interested in a mutualization of the defeats of the territory
+she crosses a region mainly agricultural
+at the time the movie is being held on three months
+strake had asked that no asteroid would be named after him
+the building is curtly being built on the old lands
+the line of bordeaux will be entirely electric sized
+the game received ollive critics on the play
+see that he goes back to my work after
+deckends up saving examon exploding the fog
+we see the englishwoman at an evidence
+this is where sheil is the main town master
+his body is covered with wooded sheets and verythin yellow
+a virtual bravery enabled a user to manipulate these objects
+she sees the clok participating in her first finale in the french championship
+sheis the fourth daughter of mammer janicologist and beatrice
+the pilasters of the gable support a pediment aangular pediment bearing a gold star
+but ridi is also one of the wife of kufo
+chanoi is a public present in the middle of the book
+the two houses of production were barslo
+zastrag was appointed astrad in the middle ages of a large city in pagonia
+these sons are much about the love story
+its headquarters are located in the city of pelo saidso in the state of baher
+specialized in the video game lither educational is current
+later run he oriented them to education
+his presence is uncertain in gentina
+in october he won his first tournament that she was single in his men
+the species re groups a certain number of sub speces that once had the rank of ises
+it is a carnage an english one without appeal
+these formulas are also valid for dictorial values
+it grows south of the united states in argentina
+he manages the family printing and runsor presided several companies including tena comer
+despite the suspicions of the word and turskey are going to be trust in the octave
+he will beat his record by himself before the end of the season
+his colleagues are not able to explain how the intervention could go to the drama
+the path of the apushes crosses the dam
+kenwinder is the son of one of the biggest fossil reserves
+it is a costing treatment and requires a resounding over villans
+we could only take care of the suits could have his attempts
+then we made circulars around the wood
+several perimeters can be extracted
+as a traditionally the participants are presented by pier esis and fencing
+the community of municipalities is located in dine la spains
+nicol divis wins his fourth conconibant defeating retel greenham in the final
+several times use their roll in the second season
+it is found in a cave in bumtingham
+she mainly lives in griculture
+during the same time bart lethis room and a round
+the livy made it one of the five emperors
+he turned the call of the sporatom meeting there almost ten years ago
+the doctor in malice takes its leave in libya
+he walks up every day
+at the beginning of the disguisement was animal form reminding certain warrior customs
+but eric flord came with vince mc marin the owner of the company
+the casavoir is therefore a new relating object between researchers and interlocutors
+he spent a part of his childhood in the south
+this title is the singer of the compilation
+adinepark he qualifies in twenty third pysitian
+some even made the honor of leaving in depression
+it is crossed by the medern beck and will flow into the
+he then became a test driver for the fiot route pilot for altheramo
+then they invited the drummer morcell card and a basist
+it can be found at altitude on the south side of the mounts
+indeed the finals are taking place in the prince's park
+his first works pictorial were acquitted of characters on the shape of varicoloured landscapes
+he sees him as one of the best young hopes of the country
+marked by the celebrity of his grandfather john endr
+before the pilip haf is located in a building of the burgher hospital
+it is close to the mc carth or blen buddik
+since then rusham has been part of old bame
+they are part of the tent group
+he has three brothers and sisters eva avin and benjamin
+its specific name refers to the island of ambiona to day in danisia
+he is located in point north of the exel and beer island
+the medical doctor the medical doctor can't do anything for him any more
+there is even one who married a bested
+only the oncear reached to the point of view
+the two travellers got along the evening
+he was born in vaguis and educated by his father augustino latherini
+stepane is deberving to mary
+during his military service in anobl he meets fo caco
+the church is dedicated to the saint crowe that of jesus christ was crucifixed
+the two other semi finalists were florrie and marina kempo
+scott finally escapes from mangikin and finds his wife and friends
+philip pogin was born in velban in piedmont italy
+the light track allows the practice of the nocturnal skiing
+later he is general adviser of the aim
+he finds among others john claude man internationally in camero
+what a too heavy soul and more painful gise mercier is too important
+it is a variant of the olive the bucket the bucket
+the song is influenced by gaspel rhythms
+he became the head office of the mixed wildvivan union
+he left few years after spain for france before returning to bavaria
+only the last points concern the border with montenegro
+only is the least considered by critics
+martinez was following a few steps his guide jose
+he lives in sweden
+she regulates the training of microtypus in favor of their stabilization
+this series is oriented towards a young audience
+this species is found in peru and ecuador
+he presents works on swiss news problems
+indeed lere was breeding in a calm voice in montelac in the bordeaux region
+eight episodes were ordered
+he extends his contract with a year later
+queen elizabeth play ouse vancouver british olumbia
+one thousand francs and the brother of mershem another king of the magsett
+he was born in antiano near vince in the family house of nancy
+this hospital is located in kiselt and a specialty institution
+she's inspired by traditions of historical novels from the epope and the writing poem
+thetuckian is often son in a slem the tuckian is often son in a salemly or adorned way
+this doctrine is a direct consequence of the between russia and orga
+she will soon develop the feelings for super boy and will go out with him
+the six species of this kind are endemic to south africa
+the paride is asking for the inversion of the tenager a little bit
+the singer of the group was also a poet tonla and singer silva musman
+the mode refers to the assistance method issued by the winner
+he was a knight of the royal and military order of saint louis
+the passive squader in extrarient is laid in vladivostoc and port arthur
+an audobada engineer was a graduate of an archives at the age of mathematics
+a legend goes on that once gathered his stones will solve the secret of the universe
+after a war we find his name on the provincial policy
+eugene writt was born in paris in the small bourdeoiri and raised in tresberg
+he will be held at this decision and this piece is therefore his last
+the land belongs to the counts of geneva
+his work is focused on theory of games and biology
+the whole of the kenton the whole of the kentons communes is grouped together with the administrative destrictive gill
+his son hib was also a composer
+there is double traffic lanes
+the doctor trettier is part of the numerous national and international companies
+however several television channels are omitted in localities
+this word also refers to a mediaeval porter organ
+the municipality has industrial area
+such a fairy links them to bourgo
+he is fully devoted to his task throughout the brazilian razers
+the game was published by sufficient
+the vidio was directed by gible o animation studio
+the horses that survived the fight were said to have been covered with blood
+the old city is later rebuilt in broke style
+one of the complexes always crossed the seas to escape by persecution
+after christen she gives it just to the extent of the savoury
+he plays in the club and was the captain of the team
+visible from may to july is the one who is a victim of the country
+she feeds on insects particularly in spring
+it is received in the order by the marshal peton a regular at the grocier
+she's named after the bishopler and founded a mixed convent in the village
+and if and if he stood out the eye is a swift move on the sen
+bujotist he fights in the middle of the right and supporters of the algerian french
+here is what sers wanted to know and first
+the peninsula has a few cities such as tylerlan kursevine port mady and
+it is the largest suptil of water the biggest case in mesalalaci
+he is capable of reading a partitian he would learn by doing so what he hears
+pone less lens aserbe or the french turkish and english
+jose wanted to infiltrate the galatic police
+she was discovered in astation
+the french army withdrew and took a position on this dent and its surroundings
+he studied at the university of cambridge
+this manimla style evokes the japanese journals with the symmetric or symbolic compositions
+bacten sordo he was considered as the largest fortune in the country
+immediately the vice admiral shir decided to have a new operation
+it belongs to the parish branch of the mountains
+a feverish face the spirits in serecitation
+except the last words of resistance the battle of berlin was borne
+the same year he became a member of the juri of the abetna
+his wife worked at the doctor manuel calvillo
+the sum contains a cell to the whole effect
+it is composed of most often thirteen glasses acaslabic spread into three troughs
+however the next season is average and the course of europe
+bird is it
+before her wedding she was a director on the hans richer medal
+it was one of the biggest comacs in mokuta for the regency
+the central table represents the cline
+it meets in the makipuna reserve towards attitude in the western cordiera
+laguna is bordered by the gine of german vity near kloka
+the couple accepts the desired couple and canvoid irritage
+the otoman was carrolated in all directions without being disjunction
+from the right to the far right we accuse him of compromising the weapons
+his architects are called wallerson and max berhem of it
+he is the shrine at the university and various schools
+he comes from the name of the martyr christian of the hippolit
+the church is located in the french department of maedemoselle in the municipality of weville
+part of his works is kept in the museum of charteries
+emily realizes how much she missed her life
+the gos master is responsible for the training of these ectoplasms
+in these languages like moton and the foundation is therefore working as a committee
+he suggests that the enemy palace would be burned which was made
+the expedition was led by general benedict arnold
+sevinsis part of the tolos cademy
+on the other hand she mates the renewal energy and that on the other hand she mates the renewal energy and that's what is happening
+he was twice rased by the dignality of eran demin
+he is placed under the jurisdiction of the prefecture of ginga
+it's a weakness but i'm happy
+the rock ends with this under the mattress of the men who were unchanged
+he was the brother of the historian ban
+yes the two florins that she's listening to me
+it can be linner or thematic
+the programe was adapted in video training game for the national individuals
+we are acting as an alcohol content its
+she does rare concerts in a few days
+francine is a domesticated woman who is well known and imaginary
+keriunka rains third in the season of the season
+dubeson calmulated the records with the french team
+its coat of arms is still visible above the grange door
+his mother wilma is a nurse
+this bulletin is still at present nowadays
+its inhabitants are called the arasians
+this time his black hat had rolled among the rocks
+from his side meri son of kira salira to the mythic sappros to destroy kamari
+it is the headquarters of the pershe community
+engly you were archduke by poisoning this ervice
+if these elected officials are sometimes a bit green they have a rich medium
+he finishes the only conference bell
+petersburg is a capacity of seats
+the parents of the club remains totally burge
+he worked in germany especially in the bade wurtemberg
+three of the men were injured but the house three of the men were injured but the house was destroyed
+he sings in french in creel and english
+one day his father is coming with a young nadiada who married
+save bonnet te crest is part of the brina
+to the south of ark sir till a fountain also bears his name
+damiens teliver spends his entire career in worsestein
+it is at this moment that she had offered him to illustrate his book from the time
+this coloration extends on the side of the adomen
+in the middle ages races are organized by new market in england
+the twelve bester qualified for the final organize the day
+she presented herself at night under the name of verni
+quickly she became a home hospital for the entire region
+the sight is part of the lewis babel collegy reserve
+for this he is proscribed of the city
+the leaves are green and light brown on top
+the utiny won the keel and the operation of admiral sher is kenseled
+his execution will not take place
+utterance is also more distant from valderants than the weight
+there also a social condition that will not be a great use
+nacotin resumpes the role of zi
+it is the first mission to have a deal with a lunar reverse
+he currently plays at lach moscow
+the departure was set on the fourteenth of october the same year
+leonardo dradini participated in three editions of the world cup
+she is the seat of churchill scott in the county of the city of saint jean
+he doesn't he doesn't know any cap during this tour
+cases of rhinal failures have been described
+the samalia is a semitic language of the family of angucidic languages
+the rps irisby is transportable to man
+there are more pears but the waters are rich in fish
+three characters will use this identity
+his resistance is allowed to take away the retirement movement
+as a result the number of flats of the orlax stream rises three
+his mother princess ferni kicina meresp is rosabeld
+the network which is secret is called sivilian international services
+his funeral still was scotted by fritz eimano
+in total they will receive five cetaceans
+his kempus is located south of the city
+leaving the family home he lookes for small jobs in order to stay and float
+all three years she produced a beer to the new marketr
+the place gives a forest to be found in the roman era
+after his release we noted those events for disturbe
+in the group the orios starts as a singer
+drest metropols is also a member of the metropolitan country in brest
+the qualifying tournament for the wild cup of rubby
+the capital is britwood
+he was entrusted to the reflection of the cathedral of saint pierre and saint paul denant
+since then the party has been associated with its views remains the liberal party of quebec
+our bank is located in sydney and the south of the city
+he enters the sa carno high school in tinis before being received first in baccalorid
+the current bishop is johan bonni
+it is located on the ground floor of the madam pavilion of the sat jean campus
+he is currently director of sky buckl saff
+his father casia ademagd abbet is a political journalist and leader
+he is the brother of sculptor robert bach
+the bird is cooked naturally and does not contain any apetative
+he wrote produced and plays the role of nuvad in the t v show
+marguerite let her the fiftums of the hagen daisy
+it is one of the first fill makers cunkales after penance
+it is a recruitment sight that is used for job searching in quebec
+the latter now took a museum of pre story of the great prison
+you follow the old fence the captain nemo asked
+he was wearing the titles of count de gricky martin darsona
+the two titles live only on the bandel version
+the restoration of the railway system after the war lasts for many years
+it is democratic and does not depend on a religious or political party
+a pink spot marks the belly
+abstrated in concrete are dialectically linked
+unfortunately for him this piton contained some wine
+mar resides in the alfoi district of chaffar
+she's a result of the merging of brunel and marmond communes
+he plays as a mitfilder for the striker and then plays as a defender
+for example there is no aggregation between the ranks of the astrophes and apostrophe
+he is dubbed in the royal championships of tara lark
+the rhein charlotte basin is home to a particularly important and diversified form
+it is then entrusted to military police operations
+it was used as a camp in washington during the siege of yorktown
+the comparative element is almost completely rung
+but the two friends fall in love with the same woman
+a crisis of everything stopped him again
+the council has however rejected these requests
+it is also a concept album in the history as inspired by the character of caspeazar
+he goes to the federal capacities of the capacity
+the monument depicts the city is lucaspa
+tilpa's father was one of the king's friends in the tribe
+jen who carries the special feast is exposed interidient and dies
+they will have to take a rest week to get tired
+they are definitely dispersed during the french revolution
+it is now bred in different countries
+captain yorke lost life due to an incurable disease
+the station was inaugurated under the name of the station of gashi soher
+his flower stem is reaching height
+kanurka suggests his services to the individuals to the small companies
+three years later the production started
+he succeeds the archbishop sumerin
+more a courtyard with a knife but takes pity in front of jimme who doesn't like it
+the decisional hypothesis of defimin is linked to the legaric the district
+he is triggered by a signal extra senular
+he realized two short films
+he also plays in a second division with the motor shinbe lub
+the act was not published because of the unfounded test of the pope
+the director is currently the pharmacist refelsenia remirs
+two years later the construction of these buildings were completed
+two cols da tenny are also open to the public from the spring to the fog
+this is why we will now tuck out the writer to give into alcohol
+juben etin was his best supporter
+consequently a gas decreasing volumes by cooling down or increasing its pressure
+the first abbess is the sister of archbishop severin
+without doubt answered barbicane but the place was missed
+the minora is placed in front of the israeli parlment in jerusalem
+the group is then renamed nationally blimarine grindest
+the building is planned for the building is planned for judicial activities commercial and ceremonies
+he dominates the high korean plateaus in the south
+subsequently the williams family moves to san digo california
+land land gro is located in the north land gro is located in the northeast of the bar
+the wooden structure was replaced by a metalic pard
+the series will have been derived from the cinema
+the field is taken by jene fuchel and mac peddan
+during the second world war she opposed tunases
+the building was restored several times
+two singles are playing in fear and they will compact the album
+the personalized autonomy is not commutable with this income
+this type of youth had also been detected in porto rico
+guy then decided to leave for los angeles
+she's named after the sculptor of the belgian sculptor jaflembo
+the whole is accessible to people with reduced mobility
+the compilation of the results formed the poners
+robert cutis did not answer but the question is asked
+he didn't he didn't publish in newspapers
+the player is gone to take care of the aspetterd centre in kata
+he enters the senate of the university of bombay
+she depends on the arstises of the island
+the station was named in honor of ketton street
+the second space is conix the first one is not
+the first of each pool is qualified for the second phase of qualifying
+the lyrics suggested by the candidate display on the screen
+sometimes this house considered a line of the abrad dynasty
+the gardens jean maritel are ready to practise many sports
+an administrator was appointed
+in total they are thirteen selections that enroll this competition to their prime minister
+spinet and paulisles entered the petition
+the decusis implies genetic factors biologically pympology
+he has a brother jeff also an actor
+ah mister terius i understood exclaimed harbert
+the festivehe word is organized every year
+the english names remain more used
+once killing he reiberates and continues to the end of the game
+members of this forum are called ecoid
+for example there is nothing to do with no place or to a culland
+it is about sale made on buildings located in the visripiate good
+these actions ligage to be actively sought after by the guestapo
+the mental additions however continues their activity
+he will then marry her with a wise feeling even though she was a widow
+john and clara are like the day and the night
+the commission is not culling regarding the senction
+meanwhile he won the pasqua cycle as well as classical sent
+he succeeds raymond de mili de serval
+he is close to the famous one in decar claude de legande
+the name comes from the church sant feliciu near
+he launched himself in the production of war chickens
+they will have to face a very bad violence
+the art of the majeli is used with great delicacity as ornamental
+for the military it enables to gain time and avoid a loss of attention
+each one is painted and divisions thrice or in place
+the village has no hotel de vile
+this game will be coloured a lot of years to crown and will cause the indignation of the media
+wisla tis a melizon wisla tis a millizon vix died in bel grade
+the current part of the london subway is composed only of electric rafts
+rosada is located northwest of dessa and to the south of kiot
+he made cons and confirm the sax as a protestant principality
+the main religion is is lacit
+it has an experts service that establishes authenticity certificates
+she joins with alice drick founder of the sextys wooden company
+these buildings are characterized by their recleticism passing from the nogotic style
+the key for this sun was felled in tyland
+the cuppy is located in the downtown of the city of stubard
+his name refers to the rocks to which it was built the castle
+the effect is bound by the french moselle and rhind to the east
+the weak depth of waters prohibited from transportation ships
+the organization is called the identity network
+antari has a label of records is an example of this type of vantage
+the stelaning regime deported full of judged potentially de vol minority
+however the gate was still destined for the researches da detudens
+she looked at them with tread
+he keeps this function until his death
+franco was a sworn enemy of his contemporary geus polus who accused him of axe
+other paintings were not identified
+we can we can we can find any data on the sources consulted on a galaxy
+this species is named after grata beeford
+saint martin de lays is located northeast of the department of alar
+he is the ancestor of lewis martigay and gilbert he is the ancestor of lewis martiga and gilbert marchigay
+he stayed there four years before being named in benazere
+he is very widespread in german speaking countries
+linlyn also leads a career as a singer of cantuple and mandopop
+for water lipping water
+mariodesco was born in savacya
+he met jean charles de lastery his future associate
+used to explore of cases to use the process
+you have new isepic
+the canton of cressy surcier is not part of the cantons kept in the department
+the movement of retirement that is making up turns to panic
+his father is the painter adriano gajoni
+in practice this term is generally used to refer to improvised weapon
+when it rains the recreations take place in the class
+as a result we an then develops a more personal style in saxophone
+it is located northeast of varsavia the capital of poland
+because i protected scenes more funny than i am deficiency myself
+behind him he stands among the herbs a book an can
+he is known to be an honest man
+i will be like the frangs to see the eport
+sheis the niece of the letters and journalist paulette midd
+the insection is on a trail and moves from its individuals
+his dogmatic night is the same as the lespial
+it has eleven municipalities and half of the towns
+the team of cedar operates the engine's maximum with the artificial intelligence
+he also restored with the modern pece
+however he is not authored by reasionalist
+previously the postal folds were left for the good care of the vessels
+he finally comes back to his decision and enrols the first team
+with four successes billanven degarde holds a victory records in this event
+the system uses dead bases of the koot company
+after michintat after michintat in kidnappy par ferbiser
+he is useful to create a complete partial image
+the ant of deco is finally found tained to cercie in the garden
+apart from the music she manufactures birds and mechanical singers
+it is part of the municipality of broe and the kenton of ansiga
+the grand pri takes place exceptionally in the second fifteenth of april
+the husband as a member of the federal conservative party
+the use of these different terms more specific to the risk of infusion
+in memory one of his ancestors he took the name of
+most of them consecrated the duvi especially the philnitographic and television
+athlete is passionate about hunting during his friend and very prolific rider
+again duke mae has given it to the croaks
+they are from central america and north america
+she often makes me angry about important monuments
+he found it gray and soaped
+other words with caskall's have seen him more correctly than the other ones of the
+a large community of the game was developed on tarbler
+the municipality is a single persh it is to
+each of the two is made up of a title of a text and a visual text
+these contributions are in the department of mansion without numeros
+indeed at the end of the heart with a characteristic extension
+as many of the local policemen rennow stop at the same time andto clerical and violetly anti smying
+she thus becomes the fifth chinese champion
+he only has one night to go to me
+nevertheless critics are aiming for different aspects of the game
+the boss freed accuracy of the remorse that we had been deploring for the review
+three months a black servant and found himself in the hostile family
+the convention was in charge of goods to transport the scatterler of their marriages
+we say at the very low costs that the pasilus took the lead of the pasilus
+the next day he was taken back in the sector staff
+after the drying the rooms can be played
+phyllis she is the eldest child of ferry buller the french catalysm
+poit godwin is another unemployed notebook
+this species is named after alfred atu consultant in pennin
+they were making them together on an antimet network of the same time
+he is threatened by the destruction of his habita converted into agricultural land
+hence the formula above
+he represents among others nata siri and kai colotims
+his parents and men are his brothers and sisters albert and matilda rivers
+but it is where she is taking an environment well my dear child
+the arrangements of tet are blocked in the building of the oto
+she seems to be a daily use in the biology research laboratory
+the length of the plateau is adapted to weigh a road tractor with a semare lock
+it is a fracture of good positions
+among them seven out of ten are formally sub among them seven out of ten are formally subden to french administration
+the eldest joseph had only one
+a light indicator is left in the eight hours signaling the same functioning of this one
+the real tow is located on the brickbeck road in barnville carteret
+the two five are similar but the younger often whiter
+she's based on four engines led by b more than two supporters
+is it going to make him comein here
+it crosses a mainly wooded and helly area
+it is also the name of the village which is part of baso fuvet and banzo
+the rue den possen is a public road located in paris
+thus it is the ancestor of the current princely house of munico
+these radios are accessible in some areas of portuga
+the tower is located in the key damrne in nevers
+but with great skills it would be possible to obtain a good idea
+he rebuilt suses after killing his inhabitants rebel
+he signe his novels under his name and under a second pseudonym malcom douglas
+he continues his education at the kikod monastery and becomes a monk
+exits are organized by the parisom perlemologists
+the country is thus cut into several districts of one or several seats
+there is a father of marines
+he is considered as the best author in prose
+one hundred thirty four for suppliers from sa pine to the project
+we find a tiny small plantation in california
+he runs away from denmark and returnd with a troop
+in the final she beats rough in the final she beats rossakestle
+isabel is on the edge of a platian
+taragosar ze kenny mando jetty and ursula
+we are toning out the exis taxes on the cotton in the lexi
+it is the only story in the s torgema kills a man
+he replied to these rumours on his website
+he was a surgeon in honor of the bordeaux and honorary hospitals in the faculty
+and it is a high place of political pilgrimage
+the relic of the work forms horizontally to is made of a timil wood
+it is located on a street of sietine in the historical district of zagara
+walk is carried out by zodi
+the king of seden rejects the proclamation of independence and settles down an army
+she meets in a cave in the naon plateau
+only the saint mark's campaign is odor
+he is however brought back to the london tower the following year
+only the bell tower and the heart were inovated
+the bishops are waiting for the most important years before revoking the situation
+the nigerians followed the british example for the form and structure
+jews are living in vessel and has already been there
+however the day passed without any incident
+he was one of the first users of ex contic and diagnastic regions
+after his catcher the region was part of several sheep farding
+he was training in the boxing club west call
+he is also a director of simon france
+he is noticed from the public for his strong professional career
+alexander von ubelt was dismissed from del rio
+the orator is located in the french department of north on the municipality of relay
+elmer's island is the northermost island of canada
+the same goes according to literature passing by less than five symptoms
+he made a disturb he made a disturbed life the following spring asks to accept
+the gittle and my words take place during the market
+but they wak up hit the world that runs away
+crcrilais later elected coach of the season
+he and his accomplice had called de cuppertz a few hours earlier
+after a precise and rigorous selection she insures their training
+he married louise marcott dargantoy and a daughter of charles marcott
+agriculture and breeding is forestry
+the devile's event is won by marian bartholy and emile ded
+he decides to dedicate all his research to the artsonetic
+this erases the universe of superboy this erases the universe of superboy outside of the existence and modified the calendar
+this pavent this paved place is the animated centre of the court's neighborhood
+the deck correspondence of heroes of dry seceles is enrolled in false against this allegation
+it is the youngest of the six actors of this dramatic series to be a teenager
+next to that he wrote poems viral and delicious novels
+he is naturalized french at the age of one
+we are also in january
+the interior is in a degraded state
+these others are the subject of an inventory
+technically talking about the man is no more
+the objective of the work is to find the original structure of the palace
+one hundred and seven greek graphics are the point of the sight
+gekelburg is part of the administrative comelegi of lavora
+its elendic variant is gimro
+binger estimates the population to evidence
+and for that they're ready to polar boat
+trevis who couldn't trevis who couldn't have disappeared from their lives
+the director has however had to be held on the ebis during the film production
+parishit boulevard was in paris later built and was built
+they only play any games
+the public language is very dangerous to find a great deal
+he also received the japanese war medal
+shebetsan is located in the vast speaking linguistic area of nimar
+at the centre i find a well that fed at a fountain
+it collaborates an balance who is one of the founders
+she became the secretary of jen gimeling
+in this novel we find non trenches in writing
+do you have a happy news
+it is located in the monperige region
+it is then that the husband came back to the day after that he was revising
+she also makes a documentary film on archaeological sights and regions
+he is actually on at the sixth place in the world
+a little later emmanuel discovers that he was also involved in him
+before becoming king he is known as the name of
+comart is part of the pasnes
+he taught retoria at swasaen the beautiful letters in the school of tornan
+he was promoted by selevities such as the journalist resolute quintero
+terror is declared to be near the points of water
+uttren accepts the position of organist church of saint joseph
+he is also known for having around the spy in richard
+they are available in agencies
+the transept is presenting a more elaborated decoration
+he established a chapel and a redi he established a chapel and a rediers
+at that time only the western sides were part of the county
+a woodwork is located entirely in the municipality of boerde
+the main motif of the abbey is the same
+fair at mario was founded as a surgeon
+john paul shillinger received many distinctions for his commitment in assistance
+however it does not resume his account the plant cideas
+tis kenton was organized around zagivo in the mordel district
+they are supervised by real combat sports professionals
+abel accepted to take care of the son of joan
+it is the largest forced of the humid forest in the island
+the limit of the sector was carefully kept in the middle of the way
+his weapon master is lucy o'landy
+this latter marks the occasion of their hundredh caps
+some manors large farm or church were fortified
+it mocks his only goal with switzerland during his first international match
+the protocol of a three year period was regularly renewed
+most of the traditional monuments have pavo flowers symbolizes
+cartley there is the assistant secretary of the academy
+he is a piccola teacher at the royal conservatory of anvers
+face to the family pressure turdy abandons his studies and joins the public function
+it is located between the grade major and colonel
+she continues with her own definition
+this name will be withdrawn later after these different with the archbishop of paris
+two minutes on scorf were coming up at the end of the geographer jean roger
+this album includes published earlier on mixte
+she regularly attracts foreign teams including columbia
+the city welcomes a penitential institution
+it converges from a new zealander theory
+in this regard one of them tells a sinister adventure
+but their fight against each other cannot lead to a total victory
+she was realized to honor to james mac miller
+he was a descendant of the first aristocraticer
+the women the women's hands are in the tomb vary from one gospel to another
+the kambes tail is replaced by a masonary panel in the sauth of the terms
+but the match was going to bed to last minute fall
+thetwo species of this genus are endemic to the island of gasca
+their sons are freely down notable
+these relationships are qualified as excellent
+it is internationally known for its urbanism
+this error is generally made by dispensing not enough coats
+all like a community of populations are made of a total of five year rounds
+she sat down at jimle kanal for the choice of the illustrations
+andrea belkovitch is a catholic and a catholic
+he dies in place
+the little numbers mentioned in blue and like the atoms of the main chain
+i see this music like the destiny
+modor is egyptian born in rome stepana his best friend
+if the menter of the three
+the presentation of the official album was done throughout the beatrix hotel de cices
+he he cistan it
+the legalistic leak is each hero to choose his cab
+transpose the first runner of ramp in spain
+the number twenty five in berlin then went to his biographer the future beam of scotland
+every sunday there is the market to the opus
+these askers were like lucky and were experiences on humans
+some states of america and europe are known as elections
+wonce deville has aremero studied in validalid in sat chap de coposto anfi
+she starts playing with her time for the cochen velly club
+the mode of elections as a member of the state council as defined by the camps
+the ca early have their medals parnesses in half boxing
+the tropic works develop with their comparable and violence
+a dozen tons of tons were he ended up leaving but still alive
+however tatit was his work to complete the plan
+it was named after pilites irish fertility
+he killed corona covers a circular surface
+he then visits the youngster united states
+under this presence our thulls begin with other movements the cultural front
+the use of the laying from biological breeding is also progressive
+the last decade of the voice a new pup of the population
+general counsellor of the same he is the general repater of the budget
+later she was dependent on the abbey of billen
+he however tried to protect the innocence of his fights between clens
+especies of this kind are found in eastern africa and the pejorian
+the hucking is mostly dawn in the waters of the river
+the following station of balvian is located on the right bank of the line
+but the germans did not build them in great numbers
+it is dismissed by the club levalois sporting club
+the island has a dietary water supply device
+in these rights the individual is the main objective
+he has all the tiny oriental arkans
+the station will be for a long time in wooden buildings
+however she plays and a daughter confronted her head
+the poland will lose the match three goals to one
+he published thirteen books in french
+she was won by stavinian
+james plain wright answered that it is a tribute
+she meets the saren streets of traper and mont prior president
+the surgation is characterized by a permanent auditive deficit
+his distribution man is large
+it was the first title since about four decades
+m character is derived from the corresponding character of the erade
+we are looking for a new start and discovers the laso family
+bovestell fance is a self tagt person
+some grips can be radio controlled by many others
+he is considered to be the best colorist painter by francois haro
+it can nevertheless reveal some surprises
+his capital is sentto tekla
+extension of the line is planned to villers sat cro
+this separation marks the end of the planatic group
+a chemical reaction took place
+he is gaining close to the artere oxity to find himself in front of the artery
+this bird lives in small groups outside of the breeding season
+significant sights excels the associations oramatical sector
+she meets in han shu in ki she meets in han shu in kih and in shiku
+locally their name was given to the small village of saracens above the other bank
+and he is trying to prevent the wedding
+at the end of gray there are two stations in service and cleaning
+the following year he took part in the olympic games of barcelona
+the whole is made up of a corps of about a dozen musicians or all on ancient instruments
+she's the sister of actor cevendo
+lastly mugan gave up the photo journalism and focuses on art photography
+during the high middle ages the monks of the abbeys play the biggest role
+the apscestra is located on the right bank of the elb in the bohemian
+he realizes a good performance attracts many observers
+political demographics
+the legend says that roman dates have existed as place
+when he returned he was sa a celebration of his travels
+cal urst was probably bisexual
+he is killed by jorno
+she symbolizes the spirit of solidarity that plends in the foreign activity
+the doubles event is won by livian and renardo vacova
+atra he complained about his opponent
+the archides of berghard are located at the public library in the university of balvo
+this species is found in argentina and chile
+the butterflies are visible to mid august
+it is the first periodical period of portuguese
+for this purpose he was sentenced to be deported to guiana
+very tired he would serve by the trittle stop of the ustrip
+there is then a lower window and a window
+saint patron was santress
+the boat would have been torpedoed by the italians who were looking at their departure
+toughton studied at winnipeg
+it flows into the northeast of the autonomous districtov cantises
+butes is very proud of his brother who is doing his military service
+it is a niganist of the surtaneous receptors causing an arterial mass
+this species is named after francis petter in the united states
+this government will also commit on the project
+he was recovered by the propaganda of notse
+te algerian looks for wins the three stages following the sprint
+he will do everything to prove his courage and restitutes the white feathers
+frank corion stats career in saint martin de seno
+it can be found on the steward island
+at that time gustave jum is also a professor at the academy of ransom in paris
+the company was enrolled for three years where the war period is during the war
+britney is living there in inventori inventori
+he plays as a mitfilder with the argentinus club
+it is fed by lake mide located at proximately northeast of bay
+his grave surmounted by his tree bust is decorated with an inscription
+to day only the garage still exists
+the syima and the master's degree in process
+eight sant cogestried in flatter oxen
+they faced each other during two matches and each one one
+the name cato comes from the albanian word catand which means to the village
+he is also the founder of the royal geologists of corn
+we are expecting a real summary of the historical text
+the blood of colissal heads is not known
+pristle ends up in the new cheke slava state
+but he can also host a mutual or a company of assurance
+this plant lives in the clear zones of north america
+the station is coupled with a plumeter
+with this reconstruction the fortress acquired the structure that we can now see
+and till then some changes were intervened but without much change
+he is a former military pilot of the legion of aviation
+he became a free agent where the seasonens
+both are old rocky attacks
+letting himself go to the odulation of the blates
+he spent his career in the county cricket club
+in september he carried the troops from alexander ria malt
+this oil on canvas represents a part
+the labels show the tors of the abbey of terman in front of witche works
+in census of land is undertaken
+it is based in the mages and was put on the picture and its agglomeration
+his family is made up of poor peasants
+the inquisitors were rolling near her to obtain her return to catholic faith
+for more than half a century he taught histenic studies in chie
+the building has been arranged for the ice tands competition
+harbert loses the control and the frustrating of his car
+this incident to the consequences the departure of damimi at the end of the season
+each house has long had a symbol of its body above the door
+the summital glaciers of two of these buttresses were also covered with tolls
+here too the supporters are harassed
+a northwest direction of the southeast crosses the commune to the northeast of the village
+in france renee rousseau was the first dubbed by mone cugrons
+a particular characteristic of the vesse was the lower jaw
+these best results are three silver medals in the union championships
+nad stumping is a student in natural science at the imperial university of saint petersburg
+by default they will ask themselves to be abandoned by another species
+we can then get the body without bowing in the damage by a big rattle
+he has the first to have released the manufacture
+he sits on the right on the legitimist benches
+ph p coker is an oath giving in front of the departmental administration
+tax is westwest it is south of china that it is located
+seven however was not always the case
+she started her career in the theatre at the naissance
+she keeps the name of her exile to her death
+its headquarters are located in the majes
+it's a procussor movement in france for health pastorote
+he represented the roman republic in bern
+we speak about the textile of separation or festration
+he was recently shown that it has an ostrogen effect
+besides this is the one of the parents of prince george to cambridge
+he pays two to four kilos
+if a team wins the first two phases she is crown champion
+he grew up to live and thus creates the serf empire which he jumped for the emperor
+on repec firstly drums
+the type of resistant bakes is not indicated on pinsions of a certain size
+i am hermits constituted later on different congregations
+the airport is still named cranford on the tickets
+it is usually used to wake up every morning hence its name
+the monastery of erlin is the ton an courts manor are unfortunately saved
+first of all we have to support the costs of water
+immediately after he took a bowl in the head of the previous one hundred years
+the grener isbeds are prefect for the first time in these elections
+ten times i need to start leaning towards him
+others still are dependent
+this concerns in particular the electronics contained in the calculators
+he is part of the grust wicken series
+this species is endemic to northwest nippo
+later on he decided to publish his sons at the elbum
+is going to die in the procession and closes the winter
+it is decorated with natural and cross walls
+they were specialized in efimers and ses
+it grows in tropical africa from west and more particularly no
+in this mode one of the two players played o shear and the other castris
+his drawing is aimed at a large lizability
+at that time the team was not called an actor forvar
+it is the left side valley of vel de baden
+some of his works are found in the collections of modern art museum
+except in the environment of the italian coffee that takes advantage of it to whistle
+he is replaced as the following year by jules van dore
+she comes back to her family with her son
+his conception was supervised by the technical director of monitory
+the storage of data storage is of co ownership system
+she started working in radio cent coga
+to day the retirement home maintains the economic activity
+ten three to five per week
+the ulsis was then killed by anian mezans in coming to escape death
+rodr is his brother ifit kill their brother iwa
+she has four movements and an execution requires approximately forty five minutes
+the practice of internal forms is strongly criticized on the mental aspects of training
+it is after joining that he became arabic nationalist
+the history of the city is also marked by its activity
+this climb on the steps in a political tribunal at the cause he defends
+the ears are long and wore in front of the front
+the mulo poses a term of the peloponate is a hermophrodite pland that flourishes from june to july
+his grave is in the cliche cemetery in the village of clisse
+the pond of chale is classified as natural and ecological interest
+the carnival will be very brilliant
+legitimist he once several times visited on the chemboard court in an exile
+the whole is surmounted by a renda man crowned with a naval crown
+he inspired clearly from the style by vag in the united states
+she practises tilamark at the bottom of the leisure scheme
+he is the brother of gerard george clifford
+he undertakes a tour of conferences through the united states that he published
+it is then that he throws his guitar on the back and hits the wall
+he was incarcerated in the most part of freight prison
+these two leaves are never replaced a unique characteristic of their vegetable queen
+the special prices are announced on march the queen of the festival
+the word also refers to a traditional flemish caridum with the gida
+he obtained a reduction in the tests of import on the british killer
+he jumps and grams with te cat's elegiity
+he continues his activities in japan
+no fortifications are known as the ones of the most permanent fortification
+if the support is moved the desolation mode of the change pendulation
+the bans then went to the pumping by their average
+the cover of the album is also revealed
+he is trained as a lawyer by mention of censorship the previous one
+the small war table indicates the office pseudorimagus
+he gets his knees and baged before dying
+not one being able to talk about it demure
+socialist and socialist antiformist he struggles in his youth to refus
+stinis stepansky published a large number of compositions on the publisher robert cox
+it's a modern district mainly commercial and administrative district
+we can also get some red and white in chamore
+rajibe succeeds him in october and died in the same year
+it is endemet to new caledonia and grows in dry tropical forests
+he fell in love with her and spent his time in leisure on the cote de sure
+he started professional careers on the future circuit in spain
+one night the clients hit her come back to a week
+milite is a little natural satellite of jupiter
+three months later his mother died from a gabet crisis
+visually goes for ericanto
+the sound is turbulent and a sound is audible
+he was named posthumously as an exam
+carolin was a famous joyful and social man and in the white and the playment
+the community was holding a passionate young girl
+the runners can however be cut off by a speed change in the middle
+this addition also marked by the absence of the pays delex block soviets
+the series is successfully throughout the world
+his performance attracted his clip very hegono
+we are named in france oxitron in the south of the city
+the original school has been created in the context of the simmons the original school has been created in the context of the simons
+he participated in a single match in the tournament of the semifinals against chek slavakia
+he still remains linked to the imperial family
+she speaks english portuguese catle and castellan
+she's a defender then limited to defend the main fortress
+to day the city is living in the wood industry of farming and agriculture
+i am alone here
+she said the offer due to her pregnancy
+report to the religious a psychoanalysis and vegetation
+his imperation is therefore coined with the opening of the north rock
+timcan also belongs to the program
+as a principle anywhere as a principle anywhere can be practised by a melody
+original changes are success in france
+he plays as a striker for comifoot
+this canvas made during the cubist period of the painter and this canvas made during the cubis period of the painter and
+his sons were frequently censured and some forbidden
+he had his successor the blessed man francois joseph de roche vucold
+however there is no plan version for windows
+all this is a weig in the co operation between the old and the young reporter
+he underlines merely the singular character of the character
+for those my time to go to riddos it was the hospitals of the grand priory of saint gillar
+it is also the last election for six months apoleon per year
+he also worked on the celestial mechanics
+will hally was in charge of a few manoeuvres of his building he was trying to do it
+she attributes churlo and crushed her hat
+finally the commotion of the resolute is often the case
+this new love will last and resume the old one
+born in scotland he settled in bacunisa with his family during his youth
+the rest of the time is open to the public welfare
+his code name is chimo
+he was damaged during the second world war and then restored
+he plays at the house on jass records of his parents
+in france the series is broadcasted from the encarte network
+she breasts in the middle of the royal square in brussels
+clearing of the orlean forest
+he also produces more animal dryings
+it is author of the eira of the roman republic
+japan that curled his kiffin manchery carea
+the origin of the name verona is not clear
+tady will question robert at home
+it is composed of the barrre di critical by a karen
+water flows into the cafe by purlation under the effect of gravity
+it appears as a dock at the end of the chemist came in the last season
+water is produced by a jar cabulga
+he married the jonals daniel liman and his father
+he will then make them lose each game stage or everything is possible
+administratively the stiff survives depends on the sable stillon district
+zinbawe is the zambia plans the importance of the victorian falls
+the two oldest inhibitors of the phosphodastores are theophil theo
+it is therefore an open system far from a thermal balance
+the contract of adrian is also a suce also a regular basis
+he opens the door but does not find any anyone
+he paids also regularly the cities and the ports of brittany
+price changes depending on the periods of the year
+the roof of the church of saint sepulchre are another panoramic view
+his first operation in the finance is a tempradrom with many stars
+to remain single they are his nephews who collect his personal and military papers
+it is piloted by the norwegian pilot john hagland
+the threat of the war in this period has accelerated this new entrance
+the shooting lasts until the next day
+the chillin bruss is part of the guresian briss
+the main communities of the south are belabore and fair rock
+the press is double this engine like a big access
+its headquarters are in tarn and its president is philip bubet
+many hikers make the city of a starting point for excursions
+he is buried in saint sebustian
+the work consists of five parts
+the neve has englishan covered with a green vault
+at the death of arnold charles teneret took possession of the duchy of gi
+he was discovered in yacimo by a cunatory and tachec rother
+he was a priest alone in hysdia and priest of saint sabis in paris
+after a climbing up to the north it flows globily from northeast to north
+they made rollings and other pieces for the first automobiles
+only the one managed to crossit
+barsy protested with the british while maintaining this rumor
+the project is launched by buggl is the first plan for the printer j romcock
+in the amusement park zumblinum monsters are blooming
+it is not the only one to have found a few values
+aristocrat in exile he saw his separate from his frile and remained in naples
+he is sent to japan to study the translation of highly confidential documents
+it is involved in some level of general and ner degenerative diseases
+she turns the woad for the furniture industry and the algumen
+this period can be considered as a charmer in history and evolution of the city
+it is the time of the berlin blockade and the korean war
+the capital was the city of gisanos
+the spenking method can be replaced by ricus
+he dies while keeping a reputation as a prudence and spirituality
+a long time his role was occupied by the master
+he has the theme of the knights and the dinosaurs
+the second meeting is the opportunity to share the same opinion as doctors ind claire
+furthermore the deputies from kirzada in the united states represented the boriga
+he does many appearances on television
+willet is greeted a set of critics
+the company also manages the bietic system of paris and its suburbs
+as if he were expecting his application is not selected by the adolf commission
+the latter is tis that was recorded in tuxen in the sath of jemo mather
+in particular any vectorial typological space is uniform
+it is the largest cold mine in turte
+it isn't durand his life that we are arriving to the happiness
+both of them looked at each other to get their superior benefit
+he was borne in his task by the conservatoire uller
+to this day no fragment of uleric literature was found
+it is sometimes the case for review scientific
+this corresponds to the case of an application intea
+the central office is located in for chamba
+by his father he is a great grandson of the politician caralis niver victor pedera
+louis must leave for indachan while jean returns to paris with the tias
+he is endemic to the island of lucin in the philippines
+his political apparents although fluzzy seemed to remain the defense and questioning
+this is where she lived with her husband
+he was named after the chief of general catre in allier
+he doesn't he doesn't exist at the very common treatment
+his up it was a kind of artistic performance
+significant volumes of ballads are created to insure better health
+she is the sister of pascal french he is also a writer
+the main nave is bordered by two natural naves
+the duchy holds its name from the town of esclona in the province of tolid
+this defeat is the second against meria during the last three years
+the area is traditionally acquired on the right
+his son degobert succeeded him
+like npil the gloves and then decides to keep the cage
+on re tullerd as a democrat mother
+he is also a member of the academy of esperantor
+his first secretary general was his sundern prickish
+they seem that their alcoholism influence his children as well
+their life peralumed have been widely used
+to day a great part of the cubre is covered
+kusmanov and nav are launched on the first stage of the count
+the city of brokoponda borders the water plan
+the sounds of the sound were added later on between the pieces
+this species was endemic to rodrigus
+ladislus changes radically to politics led by his predecessor
+the herability should not be confused with the heroty which is a different concept
+it is his practically the five elements that have been subject to a decorative effort
+the following year shimaz created a floroscope system in the tel communication
+with the apple of data objects the museum is moving to the portal
+he is deeply jealous
+she is also a professor emeritus of english literature from the university of quebec
+moreover it is possible to rand directly on the spot
+adiaphosis or the winter is necessary during winter
+he would have participated in the battle of astens alongside giom la cimperand
+murt has this metamorphosis he also believes a budget
+it is on the piano and the same is the sand
+this interpretation saved the photographer
+four years left his fiv and blond
+all titles have been won with the austrian team
+his body was found in florence
+the female is characteristic by the shape of its adamen
+he is often considered as one of the senators most moderate
+consequently they were banished by holland under pandamore
+after his retirement homes were awarded the concero de sans and piano concero
+the science is the unity of fumer would be added during the ro
+loverier square is all the specific laws of the same
+like the frency or the creek fruits are fermented in the lamp
+each district is managed by a superintendent of police
+one of the representatives in france is francois de sinle
+the attempt a cyrus smith should succeed
+some species and varieties are spread across fifteen gardens
+he played four seasons with the asques of the technology university of michigan
+they are covered with a nipping of the technic sall
+he is easily visible from the road
+its inhabitants are called se lita
+his works are translated into numerous languages but not in french
+as a matter of information the dam is quite as a matter of information the dam is quite dried every ten years
+he watched he watched mars and was very reserved on the existence of canoes
+leon b leon body is the only owner of the newspaper shares and died the next year
+above the village the wood of ferry has a magnificent panurma
+the hall offers a panorama on the valley in the wall
+he was an american attensive poet
+the english one largely on the hans gols tonero
+he is facing the island of vinla to the east of the bay
+the garamot is made of red or pink
+in this regard he presides over the civil war of court
+the title of the album refers to the people hausen rocks
+its this meeting with point of view that is played in an exhibition
+she does not regret not the time shed as prostitute
+in other cases a battery of lviller and more common
+at that moment the models were obtained from japan
+she's the first woman to exercise professionally in the netherlands
+the building is accompanied by a land associated with a piece o
+this road is now permanently closed to the traffic
+he encourages the christian and moslem military in renowns to the weapons
+erasmus who got an important success in his last years
+he pretends not to have lost his rank
+the king louis rin attended the assembly
+you were created a small post house with a mounded post bearing a du
+desier took a bunch of cold hearts five thousand and twenty five
+the spaniards add a covered path and build up advanced funds
+upperham fortado is originally from a family of mins of lisbon
+the symptoms appear after a long period of time
+te advertisements are locally held by the telesovia the original signal
+the former weaving school is integrated in the martiner school
+he was the father of joseph poyle he was the father of joseph poyle de granges
+he does not mention the title of regular nobility for this family
+son of armenian pharmacist argilers was born in baoblu district
+the localities of the nearest arpettice the localities of the nearst ar petti cramar bully and cromoda
+the route to perform te least vernant and approximately in greenhouses
+i am waiting for you don't i am waiting for you don't appear like the dogmatic religion
+he married marguerite de la tour dupin
+thedithe debuchion alcohol were to say the cause of this misfortune
+she started to take off her own discoursion among the employees
+no new trade station is built on the path
+and the first two in the city are a hotel and a stable
+his population rises to habitants
+he turns regularly in london and the united states
+he depends on the orthodox patriarch of jerusalem
+he is a law student at seville when he bursts against france
+all the music is composed and reflected there is no improvisation
+the couple decided it will come to france
+this navia place is specially designed to use the combat system in iges
+he was however recruited by the witch and therpix
+to day the company has about fifty employees
+each man lasts the right to his equivalent to menda
+some white in symmetrical tremple on the edges of a thin and tight mouth
+his improvements also allowed to reduce toxic rock programs
+the jurisdiction goes on state and lord godbeck an blaro
+for example of cashat you feed the strong cheese
+the hair branches are small colonial animals connected by colloge study
+he spreads in several categories following their order to end up
+a spiral a loophole a modius modus a circle
+bu stubbs are established next to the car's party
+the combat law books a more positive message
+he leads to his entire career in the sion
+there are then many to settle in india and in balgalia catur
+the teacher of the smell the teacher of the smevil is vespa de loroski
+philipp looked at you
+sud arabia is eliminated in the round of sixteen by brazil
+a wrestler slavic a wrestler slavic wrestler approach is not accessible by a so called gelace
+the color changes indicates change
+supporters is located east of the aldio stock and northeast of necca
+we will never see him again
+it seems to have a good base in physics and in english
+this transmission is not available in europe
+they won the united states as allies that same year
+this kind is unknown in the old world
+the group is named after victor de logi
+a column cylindrical is based in the upper part of the vaspue centre
+to the north the districts of actala and neverine
+he is based on english with french and others
+after a few years he travels to besera in the current rack
+the right to pose lords was established in albager in favour of certain lords
+the movi pas became unstable to this otential
+the eyes were enclacing in a sculptor engraver with his strong tendency
+the concerto is composed of three movements
+the bill and the budget are still in favor
+the name originally of his hermitage was goro sinki carpo
+as the night falls the sticks are trying many manoeuvres but without success
+according to him the woman is the basis of society and must be educated
+he undertakes to weaken the cadios by limiting their powers
+transferred to se cristo philip ignored a pastoral approach that has been successful
+the direction is in paris
+she closed her eyes she joked
+he went out from the otdern high school and called the pravest of sa flower
+the characters are the main character of the series are part of this agency
+the governor general took out michoel adio therefore appoint somer as prime minister
+then their perpetual vows allowing the nuns to get its total end
+the author also has inspired by the professor the author also has inspired by the professor's career in the university of icago
+at the beginning of the second world war he was a member of the people who were imprisoned
+philip stultz mainly produces music video clips
+the colored museum in westport illustrates the history of the bluer district
+the mill is violent and tumultuous
+it is common that a derivative of necladic serves as a sugar loam
+a theoretical thought of the architecture developed perlo
+five statues represent the virgins of wis holding the lamps
+he ended up finding in the middle of a bunch of corps and decapitated
+its capital is cribby
+they use the spaniard among other things
+it meets in the osenisula
+he is considered as the most important architect of this period in spain
+he is seen in light between diving and passing through pedestil picas
+it coasts a hotel of the campiski channel
+this guy of the episodes follow the original diet
+the pregnant part that recreates the withos one is destroyed
+it's ten years old spends most of the living room in the living room
+he married and had six children
+marguerite lendre studied law at the faculty
+this is considered the best quality of the world
+it is found in h aringo and in cane region
+trinced on the edge of the rhone its path is flat
+the population of the city goes beyond the evidence
+it is the daughter of former president negarea
+it is the design of the engraving pierre alexander morland who will finally be adopted
+during the congress of shareholders the question was about the financing of dinland
+the disappearance will have consequences as unexpected as the min
+johann's jules marie tun is buried in heran in the cemetery of loa
+its renish or are filled in plabaly
+only rippens left his work in biology and antology
+the check republic is wisen saberin in the west of france and ugrain
+he dies of a long tempt to the leg in paris
+he notably won the tour of luxemburg
+the lirois marine too completes the podium
+the freezer decorative in objects
+the brewery is damaged by several bombings during the second world war
+the players have a spawning which lives in continuous
+the municipality is irrigated by the rimindo
+in france the young boy shouted
+the dust is in this case simply covered in case
+one day he finds himself in his taxi a bag containing francs
+the winner of these plaots is designated champion of turty
+the three other stages are climbing tests of difficulty
+the missil of a contest will not be able to visit it
+we can even find it in a little softer than it is possible to find it in the past
+the elaboration of the operational plan must precede the one of the budget
+then she started working in the valley which bears her name
+the hull becomes a mass in flying its fifth enemy plane
+they are made of stones mainly from court fontaine career
+it is the stop of the male descent
+the pope francis as well as the philosopher amilia pueti belongs to this movement
+he took the title of avoi of the holy sepulchre
+brebreer was born in california united states at the time of the darsindrum
+gordon was a specialist in the long distance races
+the following summer she crushes under the orders of prince albert the following summer she crushes under the orders of prince albert aser
+the cause of not too much dear eric is going to your father
+omer push the structure that flickers and they allow them to escape
+four years later ritasor in prison
+she is a great friend of mark le brec and jean renee de four
+the unit of the military genius it was distinguished during the first roal war
+on the way back he dies in acolais where he is buried
+i followed him and i went back more than a big sand
+this mission of counsel takes several forms of practical
+he maris province with alice tres marie aneville dumand
+it is a private building in a state that requires emergency restoration
+she takes her source on the heights of orsha in the bors channel
+he excels mainly in the art of the art of the medal
+the orchestra is found in the orchestra of the national opera of brach
+martin dorey is a researcher in acronomy
+he was related to charles peggy
+his new clouds feed on the island of the west southwest of the mayn village
+he has a degree from the technological university of heltke
+after his high school diploma he continued his studies and becomes a doctor in pharmacy
+to date the programe attracted a lot of uses
+among these municipalities only chasenard colanges and miliney is located in the allier
+the latter is prime minister of territorial authorities where in charge of energy issues
+he is kept at the paris debo arts of leo
+this is amused
+it is found on the municipality of batiner limanzo
+at his catalogue there is a hundred works including many of them are registered on ten
+it is located in dor angerson
+the urban unit of s treves is urban pole of the urban area of nimes
+he played as a winger
+a process that will see his achievement in the realization of the fullying brooker house
+the german won two other races in canada many
+according to portuguese interpretation the sovereignty is on meka
+he works as a nurse
+it is persuaded that it will help the right to obtain rights
+the captain paxton is an excellent sailor
+the gray through the exceptional frescoes
+the sights are meant to find a stranger emerging the painting
+it has two seats
+he leads to this title a group of intellectuals in the city
+this one makes three first models for the lewis but also renee
+paste paste maria who covers the southern part of the department
+it serves as a place for exhibitions
+the camapda of many monasteries in tubit and china
+its chief town is the city of beck
+priest of oro seventy one here periodically
+they are going to home in a lively area
+he insures nowadays the function of dialogue
+saint louis laurende will be recognizing him
+however these regiments are maintaining a tradition of a lead unit
+it is a two part contract
+after maringo millis was furious against sack
+their trees in the fruits are tasteless and sometimes with edible pulp
+the tower has two elevators to its summit is a platform of observations
+all of a sudden explosion was being made
+giving up positive criticisms park calaveriks once again
+the notion of numbers first is therefore extended
+we can consider the duff as a parody of the north american bet wiser
+a library is at the inhabitants
+a few hundred workers were working well and there would also be a construction site
+edward was born in resiga now his little brother still in holl country
+he then went through barbra brice and berlin
+i thought that francoise had seen us in a small puredor
+agnel did not ask much better
+during the trial phase the speedo was already attacked
+the sight is one of the stages of touristic andju
+many sports are also found in the university
+he is a member of the central society of architects
+he now occupies the position of coach
+his method uses the infinite cross
+he is also an organist and most founder of the association in africa
+this makes it possible to get rid of the legislative elections the following elections
+the provision in bab the provision in babel will favor the front
+the marriage must have oked the union between the spetel and the hestarmute
+it was named in honor of william stender former general inspector of nato
+he is during the whole period and financially defacted by his friend leuis
+this gerilla is one of the very rare to have a war navy
+paul hepson died shortly after
+the fountain was made by the beginning of
+it is located on the ronums top
+the aros financial districts of brittany are from the diocese
+he appears in eight times and scores two times during these two months
+we find in the erport
+most of these quotes are in most cases and in later version
+he is schooled in klong and pursued his studies in deseldorf and minich
+ancient buildings are almost nothing
+when the war struck the french move was not spid
+he was mounted on his several government
+we can still remain to the injective morphisn
+he specializes in the taxinmy of coocters
+other men of the diagram are completely empty of stars or very little populated
+a lot of big female hotels have a presense in the city
+however at my way you're saying i'm telling you
+it has seven reserves and is based in four tremillion
+however the action of the prince is his niece carrying his fruits
+this newspaper is one of the most ancient anti semitics of the time
+browned and redwin again
+the peristo is a central or slightly anterior oblique in french
+apparently the arrangements were similar but there were no plaged catore
+his rival yes to de vasarion in taking advantage of it
+the following year he won the tournament in diversom
+flat brigade the same month
+she was won by alexol
+it is also the city it is also the city's head office uround kloka especially in the pool
+i thought it was the doctor
+the japanese virgin is used as climbing plants
+he is the son of vinitcala who is also an actor in delis
+once the mobilized army the once the mobilized army the dictator leads his troops against the enemy collision
+his activities changed the good times
+jon milla polidy was the first secretary general
+betuy she was praying a little
+at this moment bluff peros was simply contemplating his power in the rock
+species of this kind were discovered in the united states and china
+it is the second abbey and champion title in the history of the club
+they have voluntarily used classical arches of the pantasy mediaeval
+he studied law inteles
+monteine took assistance and lived in the friends where he learned to be reinforced
+the treatment is justified if the ministiscod and sentomic
+she's directed by cherard peneroa and the french national team
+for example the number of people is also available
+this branch which exists to day is still orthodox
+the soft ware allowsis to prevent the installation of the espian soft wer
+also for finishing the movie before the appearance of ending credits
+the group chents in norwegian then acquire the interest of disk houses
+he was a member of the editorial board of the review of the world
+geographically srana is located between the sierra demanson and the mountains of pras
+turned into the coasting he worked in florida alabama and then texas
+at the adult he was a doctor of the months
+it was my turning the hardest on a plane of psychological and humans
+to this added a tribute to literary
+if all the creatures of the player are in it they speak the game
+he is immediately transferred to the trade he is immediately transferred to the trade i portland against the money
+the estate is then known for its vinyard
+this species is named after the little girl's eyes
+the cars seemed locally marketed under the fiastercl brend
+the language is threatened
+and pierre montesquieu flisanta is the ancestor of all mounts cu made by dartarians
+several anecdotes are won by members of his family and his students
+krausten liguf has been acquired an artist nathan semel
+news were going to worse for a little bit of the stadium
+moreover the weakens accumulate poiums that are shut down
+some grippings were ice creaming the place
+however this status appeared quickly and satisfied
+the other airport is located on the island of trinity the perre court international airport
+pine racento continues to dominate the season on clay
+all the tents of the summer had been so thirsty
+it is a district town placed under the jurisdiction of the prefecture city of linda
+she corresponds to rocks often under saturated in succa
+six dicors the middle of the facade is three of each side
+thomas tatton more discreetly showed on the right
+he is also famous for the first selector of the team
+the passengers are from the allies with their pistols insert to accumulate the points
+in another scene the king and queen are real mio
+the race and women the race and women's american did not attribute to medals that year
+since then the building travellers has been occupied by the technical services of the sapaloty
+the club president is dundl foodball club
+the administrative centre is located in calcuni
+philosopher theologian then doctor he turned toward psychonosis
+hendrik and fevland of the albus is one of the main representatives of the spanish school of vilin
+the bulsare for a long time constituted by human populations
+the canal is demont described by aridote and mentioned by cosid the canal is demont described by ardote and mentioned by cos
+this crystallality specialized in double colors offers a palet of eight colors
+friend bullen follows his studies at the university of syracuse
+tourism also tends to develop the important potential
+these stories are read in england and the french people
+charles michel male is a professor emerdis at the pier in marie curi university
+the dublin society let him win four prizes for his portraits and paintings
+some affirm that the dulter stories were prohibited in this place at that time
+other sources attribute his death to a bet of other sources attribute his death to a petificial crisis and in the trovelling circumstances
+he followed teachings of jean dillons and john clarius
+grill bond injured he is victim of the hollow and voter nation
+his activity as well as his exhibitions in colog and antivity
+the community is named after joseph sajen
+to day by tebe oba hollison let this converted
+one and th other is in fact a small meal
+a few times later all sacap escapes a tem to a sitina
+it is believed to renoard and grinnie
+the museum has a collection for which loves to build eirplaying models in miniature
+many acitic tales are told on their trip
+it ended up catching her as she was reading the last of the piece
+it is certainly in this time that the career has sick grains
+the same year he participated in the battle of fort baseju
+charisy is irrigated by the sense
+the date of may is retained
+its presence is therefore a sign of quality
+a large part of the barbium of smith is still kept in britishusium
+this ordinance is immediately translated into a circular destination of the conservators
+two degradations of arrow bridges were discovered near the church
+three other members of the family received a lot of life
+spi resort spi resort dskil is pi the sun das
+the municipality is located north of cell
+the lieutenant beason pulled but uncast the lieutenant beason pulled but uncastable and the man managed to escape
+his father michel bonaclo senior was the guardian of the stadium of narbone
+most often nothing is visible at the end of a period of these media parties
+his wife sonya and his son nicholas survived him
+finally he took participants manage to compromise
+midlovo is officially classified as one of serbio's villages
+the regiment then took part in the first battle of the marne
+even when a person died his force lasted
+the occupation is going to put an end to this phase by turning into adtic
+he is submissively the influence of the barbizon school
+it is located in the city of elfi
+present in this region they are celebrated in the tour
+after having done it she will denounce them to the police
+flurry street is a public road located in paris
+later we noticed that lalava is nabar were renamed very much
+conrad was born in minich bavaria
+in communications the admission is more commercial
+he is part of the arons district and located in the municipality of ermans
+women must children
+the bacteria can be defending anti bacterian agents
+in the morning springfield is covered with snow
+the french political classes condemn the act and brings its support to the affected military
+he notably took part in the meroni states
+since then the game has been still used to be in levoir
+it is won by kania paul herring
+a portal was added to the original facade
+the purso continues
+he was the host of bassa
+from this moment they are obliged to welcome the girl for some time
+she's a member of the french emerton administration board
+these research focuses on the instar physic of the high energies in acting
+eugene dubois discovered on the site of seden bruce
+it's the very first appearance of this club in a scottish football league
+many photos have vidadorke are compiled there
+the remarkable facade particularly round the central gate shows various oddities
+the lizers families will receive a compensation
+but his parents are eparated and their sather refuses she sees him
+its headquarters were previously located in saint dedier street
+being hit on the sand page one hundred twenty five mont replied vencroft
+she converted into a radio acting actress
+the neve has never been vaulted
+at this place another manswik in a very populated sector
+the astrate is actually a transporting vessel for several generations
+a petometer can also b an analye a petometer can also be an analye a petometer can also be an analo for the short pressures
+it's a village an it's a village in an located in the north of the village of kito
+this term is sometimes used to refer to micronations
+it is a fermentation product and a microtoxin
+the number of species described by willm n fert erson is particularly important
+the deportonation is a slation of this asalamic asalactu
+would you be appointed go more operation
+the pre seasin interancship took place as usual in lucan
+he is hanged to the prison of platanse
+he is a member of the patronage committee of gota
+it can be written according to different orthographers
+since then we can see him regularly in other series and teb movies
+the attack has several deaths
+there are other ways to measure the inflation rate
+the theme comes from the gospel
+it is the first agient agent a g
+splor stromhem is the roof and his friends are practising
+this strategy was van because this little son did not survive his successor
+the castle integrates the parish church of didies in saint jacquela major
+it is an ancient modern archaic style
+there is a real florism and it was monetery
+he designed several projects in the region of antwerp which he is difficult to documentaries
+later the same year one botis tasawins the presidential election
+it serves as a link between sydney and louisbourg
+in honour of the letter
+he wrote for the radio the publishing house a trampling reviewed from check
+closed the officers as well as a justice of the peace then according to politics
+it is located in central anuni but i is separated by the forest
+these closeer part of the ability of popular culture as the symbol of kitch
+the change of the players depends on the game
+this latter developd considerably in the decades that follow
+this species is named in honor of francis white's
+the pantetic speaks four times like adelaides destroyed
+she therefore took the fall of the textile production in camero
+she has a baliameter
+he is in prisoned for two years before being anisted
+this is based on the new york and unlocation of the big automobile grandre in germany
+the very strong growth of the parisian population nevertheless requires our job
+this increase in favor of the work in vertical position
+we will find it in among others at sedon and his neckels
+at home the players play at the tisso arena
+cartly the bis is fed up in july in august
+later he lost his career in the united states
+all these instances like the conference had their seat in geneva
+the war of wilecti is the second biggest island in the archipelago
+the law protected the twenty nature against the traffic inedutiv
+he puts in service on after competing tests
+moreover the professional and important cost
+the m p a l e are also distinguished by a modified landing gear
+we can see the movement of the bed to the north
+unfortunately for the south unfortunately for the southeast the dispine did his return in his pascal
+the boulevard owes its name to the church of la madeline tuprush
+he had finished the same time as the three others
+his coach frederi antonetti launched his half time replacing olivier vene
+they have no additional memories from their past nor their identity
+there was inaugurated the historical museum of altheramo settled in arise
+clars raks clars rak metallion attends the vocal production
+ionhero szuki or founder introduced to classical urian music in japan
+pelegrin pila was his student
+it has a spence the choir of benedictine
+he has his own preceptor victor cuzin
+here is what the point of doing my master
+the camel also used in persia
+a simple porch allows access to the street
+he was a member of the board of directors of the club of normalions in the company
+even slit failed to destroy her now that she was crazy
+later he became the architupe of the druid abusive to the star of henzer
+but she didn't yet took part in the university thought
+catin osman starts skating at the age of three
+this astrate is dedicated to stephany de belgian
+as fooker in balzac he has the ambition to take a total company
+the young girl then acted twelve years old then assured all the old cows off
+despite these events rukambolisk nuggledwill have greatly contributed to the game
+when he murdered his murder is taken a picture by the latter
+it is a tributary of the urga river in the pleda basin
+he now lives in tyrona
+she was also one of the first artists to appear on british television
+the latter were tolerated by the bench
+for rachel baldy the young lawyer who defends her this is a terrible failure
+the music was composed by ichiro shima kora
+the small breakfast allows the association to support the project one of its pipers
+finally she adds a floor to display more seats
+during the french revolution his goods were seized
+then strip sin the base on the season to consensus
+they were part of the happies
+his mother florence was born in weman was a music and a confirmed singer
+the explanation of the fact that the universe could be homogeneous and isotropic
+for ten years it is a former toxic mixture and a for ten years it is a former toxic mixture an acid
+copany was finally defeated at the battle of suli
+lack of proof he is released a week later
+it is then that the first expedition to astronomical belgian
+the railling substrait is suitable for some earthquakes
+black walls made as barriers of bulkan are on the southern edge
+their protection will be provided by this amendment
+it can be adapted to work by lifting the plateau
+the organization was divided strictly according to the principle of the leadership
+eased by the french charts he wanted to become a magician
+a universal championship is held by seven children who is the first year
+he knows a great quantity as well as on literary and politics
+in october she participates in the rural championships in rotterdam
+it is destined to the urban centre and the urban population
+no collution
+it is a newly tense of rome to these last ones
+it was part of the minero trangle
+he is the son of the tennis professional tennis player naguriantonimo
+tefford has carried out a great number of photographs
+he is a federal structure of the ministry of the higher education and research
+no one will get out of it here
+the spicel war is located in the cloak is not thorny
+it is suitable to distinguish the security of nuclear safety of nuclear superity
+mahi discovers that he is the daughter mahi discovers that he is the daughter and half sister of nabi
+this egg was meant for the commander square
+ande bus is located near the station of the station and the south of the city
+the city underwent important destruction during this period
+the elbum also released in the united states and is the first record of the compastor's label
+the collection of jewish millers is done at the importance of impressionists
+he is now administered by the roman ministry of culture
+they come in general in first
+when gregg appears on the side confederated when gregg appears on the side confederated dick stepped back
+he is named after the luseny oid
+in o broken leaves the pedlus refers to the axis of each fullium
+they are the most number of elves
+he would have been a residential advantage than a military
+he was vice minister president and minister of justice
+the city is narqell
+the fight against dobing is far from being won
+instead of a snake a church meets a fiance the king of the serpent
+he is executed three days later
+it was still at tartellet turned into the locks that a fighte
+there is also the balliard bata and the island of
+it is divided into a certain number of tetonic plates and also called plaks
+twelve people are officially candidad for the function
+it was the first attempt of this nature in north america
+he is buried in the church of luches in pelatinet
+these two other girls became religious
+with the club of fernsboro he plays four games in the cupa
+the mimushi station has two central platforms
+the specimen described by taylor and smith was therefore measuring the tail
+he also devoted himself to painting for a little while
+it belongs to the buclat aberegion
+is there also the department director of the history of the university of kerro
+ravin simvi then plays in japan
+the members were admitted to the base of their contributions
+rene se sir is buried in the chateau de n cemetery
+he works in his youth on the agricultural operation
+it is dedicated to frenc fernand and the orchestra of cleland
+ferryd ferry dindy received his first education at the classical venasium
+he finds his position but decides to run away in front of the mountains
+it is the expressions that are considered very light and remains absent
+the decoration in her is due to the school of domino and the goldsmith
+the word as dill as the conic
+the route was trodden by roland lebern
+this species is endemic to curtero mexico
+east is an old town in the moselle
+the normal voice goes through the camp of abico at altitude
+she grew up in westminford bullow in high school
+he also studied lands for redistributing local peasants
+gillnot has to reinforce the peace to the jesus christ now and forever
+pedal hove
+it is the beginning of the international war of achina m
+however you however you disembarking at one thousand forty in vain in the whales
+one or several an suggest interview of an introduction of a special noter
+a minature is announced in november in glasgow black pellin london
+they do use the same data systematic works
+after the second world war the coach became ucim de belly high school
+the case has reached the ears of the government and the rabbi is severedensed by the emperor
+the island is without car
+his mother jetin
+it belongs to the makak group of sulawasi
+he participates in the battle of england during the second world war
+this vast collection was soon kno international success
+the carpenters also use a mass for wood adjustment
+therefore it is totally unnamable as the archbishop of canterbury
+it is therefore important to insist on the country the price of the race
+during his adolescence grin was replaced by the master corvey
+maria straff daughter of a policeman followed the school of teacher
+the prescription of actomatic treatments could be done
+a cycle stretches over a few months
+he is employed in small quantities in the manufacturr he is employed in small quantities in the manufacturing of the glass and metallurgy
+the territory is covered in large parts by addressing
+six villages are found on the parks territory
+the street gets its name from a landowner
+he studies he studies zerology under the direction of professors pavel kocastav and bori
+then the latter are handed to her mother
+opinion is still in the middle ages
+he comes from the medico high school and studio
+to this election he puctuated the list of lists in the community of madrid
+he remains very few literary remains of the ancient persians
+we can play a cardn on two numbers on horse backs or hero
+the management of the bristles is located in north cupping
+traditional sensitivity he strongly supports triantine and syliba ecclesiastical ritual
+then they will insure rapids by mercelon in valencia and barcelones in
+after the first dall the humane rsurc is followed by the second group
+we can sometimes find it imprisoned in the shadows
+this one features doria and moves a boy who brings another one
+he played a few games with the flames companion of his brother
+there would be a stone hill on the rock
+acting as a marvellous variety of ability than the common and complication
+the roof is made of slabs and wallets
+after many explorations jack dunba went to rescue them
+his father a his father adar elserd is also a fiite religious in high rank
+it is in charge of the smokeless powder
+thinhabitants of fifteen streets frebourg and cologne are organized together
+it has a yellow and black sand with a very variable
+a show case was presented in core and gren de france
+another ceremony had once been held in the gerba community
+it bears a rack
+the album is well received by the specialized press and ranked in the albums
+the object bcame a divine or semi final
+if the heart of the varieties is good for example
+nevertheless in august he did a greater moment in paris
+he managed the sporting criste player of the defenser luma
+chees the daughter of actress mchees the daughter of actress mihin kan an actor jamgiba
+ures was one of the educators of the hotley and taught him the art
+he is trained only from spinclasection
+he studied in the centre of art and the technology of kalsru
+the oldest part has a very semi the oldest part has a very semientar axis
+montain top is addressing above sea level
+it is large and beautiful blue busko because we consider a terdi
+it is electrical to the valitre located in the south it is electrical to the valintre located in the southeast of millin and to the northeast of montfundien
+he also exhibited in galleries
+the compound is however stable in the water
+since then several prizes have been awarded in different categories of others
+first of all alone the unions represented lists
+he became director of the opera of
+this one rejects large quantities of the age which has no places under the flow
+this plan of water is located in the forest and mountainous area
+he plays in the same time in france belgium germany and the netherlands
+he made portraits
+his father nicholas lusky was the professor of philosophy in saint petersburg
+the old convent is transformed into the beginning of the office and apartment
+the painting of board has three floors and three lit
+the bishopric psychology would have a great influence on the whole christianity
+each indication is affected by a weight according to its importance
+alexander bufflege lies in the covalescor cemetery in saint petersburg
+after the bit after the betville in hemory federoski follows a training of a landscape gardener
+the cateracos hospitality strategy also creates important polemics
+at three minutes of the denier flag all the pilots are made of stone
+she was born in chicago indena
+his age favors the brewer
+the federal team of runts champagne is coached by sarkeste rebellion
+the tullisene street resumed immediately his first name
+everything is perfect hominy and ordering in high level album
+only the chits are limited to the unable to keep such frequency
+however a harmonization is still on the indian tax
+the chlorates are hydropilicate pilogens
+for the moment the port is considered to be unusable for at least one year
+he was president of the american po records science
+one day mohan was dead by a cobra
+for navicste
+the series has a particularity to set up in a large number of humours
+they feed from art provoids an art
+the larva the larva is not insects by arcoter or other animals
+ifirm how to work on his first album
+he held his positions at george unpido and valery until his death
+the territory of the commune includes the miami
+the constitution was completely deleted in basmia urzeguben
+there is no delineation agreement between both countries
+he is captured by the red army and is held by a kimp
+its inhabitants are built by a new name
+besides the water shed with reduced liqistosis
+the club plays at the waldst the club plays at the waldsting stadium at a capacity of seats
+a period of tests is thus planned during the season
+he is decorated with several occasions for his service
+its good quality waters make it a prolegated space for fishing
+in fact the young man will travel and signal the south america
+this pseudo class this pseudo class was mostly a good deal for both artists
+she has french and english origins
+princess oski is dismoying berlin princess oski is dismoying berlins on which she has based on her calculations
+these favored motifs are child portraits
+marie and sate are you a girl
+many elements of the game system have been changed
+coco comit coco comiter is a painter and sculptor
+it is mainly responsible for counter teroism
+the cogl is not navigable
+alter file micronitian is ray alter file micronitian is ray tenth an his event
+its headquarters are in the tribunal of estins
+the sit is used for transport purpoces
+they are all refused and new sectors have been hired
+the locality of prinley is the chief town of commune
+he is the last bishop of pi to be elected by the chapter
+the latter demens
+all greeted the jean's work from malaved
+after his return to japan no saka settled for the organization of the communists
+its inhabitants are called the seka
+the church covered with slate has a large square bell tower with the arrow
+he is however difficult to separate the effects of genetic environment
+the hales curators remaining on the french side and disused
+a block retates the scene of tatan on the savenna
+aragot is married since the tony perier
+however the application of this degree will not take place without results
+france waster wins the lady singles
+one pablod went to join the main distribution
+he is given two letters and is called napolier speaks of an exilliary
+during the french revolution he served in the army of alps
+the economy of the village is based in agriculture and live stock farming
+we can see that hunting for hortax systems
+you would become the leader of the congo republic of congo
+all the courtiers found there
+historically the name is given to the use
+she took over by the british authorities he dies executed by the greenhouse
+the brubbies reproduce in pures
+he is notably known for having taken the study of tesdomarch
+it is the third city that is the most populated city of catalonia
+each nation is coldin before depicting one mixed teams
+his command is entrusted to this occasion to captain of corbets and digish
+lack of means work stop there
+daniel bale is the son of joseph eugene but daniel bale is the son of joseph eugene bell and suzanne schiller
+a line of the utternimer group is implanted there
+he represented russia in the youth election
+threatens is crossed by the robian
+trivacos products are sold only to governments
+jokincok was born in rostock during the second world war
+he is also general adviser of the kenton of saveria
+determined which one of these treats or desirable it is the subject of dates and controversial
+commune located in the smaller lens in the country of marson
+the second and the police are in the middle of the war
+brocker studied in stratford hoppin before
+he came back to venice and dedicated himself to the art of the bar in travels a lot
+but these elections are a reverse point for radicals
+he is known for his contribution to the development of three rivers
+back in cruce he resumed his profession as a doctor who became freely aware of his poor
+it is facing merton's when a convocation and a belgian national team
+he married the daughter of the architect pierre prosper chebro
+she's the younger sister dowin right
+this species is endemic to bubble in the philippines
+a street in marseilles is named after him
+he was trying to hear well he was trying to hear a well known crilling of his carriage which did not respond
+its capital is lerec
+at the same time joswin is given a recurring role of his new series
+look at me the teacher said
+police sources indicate that barn is joseph accused of having taken him out from him
+in the park of the castle in the park of the castle of dfenburg lefelens also produces a factory
+william don studied at the king's school of burton
+it is located approximately one hour by the road between morelia and europe upon
+it even has a common algebra
+the name of asena is associated with incredible exploits
+at that beginning she was hoasted in a sociological throat
+the set in bad luck were thus the set in bad luck were thus rific in his career
+the group is for its performances supported by other artists
+during his studies he was yon christ commits for art an his story
+according to the guide of lacho the black and the most frequent
+during this time milder fuzzled a bas department and finds the keeper's apartment
+the roof lightly falls under the top of a parapet in the ancient style
+francis patery spent his childhood near the village of sonac uthian
+she published a demo on the internet of the m t editions
+the exterior aspect is deeply modified by making the drying of e villers negressible
+some are imprisoned without acting assessment
+she trained many bridges still remaining still
+it is a federation having its headquarters in berlin and re grouping all germany
+he played in the collections and quisite collections by naturalization
+the necropolis of louven
+paris sella has a degree in political sciences at the concordia university and the magill university
+however this world would have disappeared several centuries before
+blatshe e arty records the song in the studio
+speaker his marage bourgeoiri transformed a beautiful father
+thence to his hot water sources known for the antiquity bedel or at the end of the final
+under the authority of the abbey was ruined and evacuated
+finally they may still have the impression that there is a solved problem
+for him the poetry is both hands beyond social conformism
+she had to sing before talking
+however this one is dry up before setting up the laws again
+she was raised in the commentmor state
+he regularly taught nebehun at the academy of nevajo
+he funds his construction and proceeds to local subscriptions
+this dialect is often resented tas a very good one
+the island was inhabited there approximately in
+it is accessible to the very steep space that it owes its volcanic character
+the champion of a series can be directly to the upper level
+his father worked in a company of granite prenticing jablonka
+the first girl from paris is a subdle with melagia
+the college of jacques lenloi located in pan dela the college of jacques lenlois located in pant de lach was named after him
+it is the largest commercial success of the history of polish rap
+the pressis down by the echo of these sho cachiles
+fewer selibl in water
+sona geza is a subsidiary of sunder homer in malabo an equatorial guinea
+at vomera the stadium bears the name of the national championship and internationally
+the basest of the album is on the tour that follows lionel martnes
+he studied secondary education at the little seminary vesailles
+at that moment i am not a close fancy
+flashes of pinnats represent scenes of the city of the dominique order
+it is known for its t v series including hot water poy pennel
+on the roof terrace two beautiful villas were added to the agramation
+he declared the police that martin would look like a drudge look
+kurt kirt is one of the sights to think is more visited in the island
+crused by canoes it has a peerest character
+france centered in the first wild war with many villages and i
+for different reasons some of them prefer to use this term
+the little ones are therefore from the birth very fast
+he plays three games
+all the titles are by graams people unless stated otherwise
+the latter persuades him to be linked to italy and accompany him
+later he was reported and arrested as a resistance by the gestapo
+she is holding a vipper in her hand
+gord was an engineer of research on bridge
+but he believes really as stop's reincarnation of a japanese warrior
+he is secondary by the soven modernist memud cabeo
+these last germs after she passed
+this fact may be used to give elgerdam for the syntaxical analysis
+that can be simple or federated and more complex
+the badoir was disused and deleated since the second empire
+the exact composition of the orchestra depends on the workeecuted
+this place is located at eseler in belfeld
+there is a player mode in a multi player mode
+john was born in ten minutes before edward
+the cowing of the feet has important functions since antiquity
+this attribution is contested by some historians of art
+he owes his name to the senator daniel webster
+one of them has the capital new york the other london
+its chief town is the city of conception
+the road crosses the rupiere forier
+abbot castilre had expended the day before his prison
+it then enables the effort territory
+exciled his father decided to cut him off and live
+the album is richly illustrated and mixed with accent on the iconography
+those who really have dangers for environment
+the antion in charge of increase
+the history was also adapted for the comic strip
+many people are then forced to work for the public
+the neighborhood includes the village of warmilo
+this anancis enables him to distinguish the obstacles and the constraints
+the experience consists of using two similar biges of one as being led
+it's true it's true said robert grand
+hawai is not a single case
+he is then director of the rigal technical centre on ri gerin
+limva is officially classified as a serbian village
+the identity of the cotre city is belled with time
+a voice from paris the rue de bethas perpetuate the memory of this fight
+the province is rich in resources many as well as the nickel
+the manuscript belonged to the nju family and undoubtedly to rene danju
+you were also presided over the liberal cokes of the months
+they have in common but they share nothing
+it then serves in the british navy and becomes a spiran
+nevertheless he refused the offer to come back to england
+during this period the singer was de clade in the cotre circles
+it also serves as a nurse colled lemonade
+we will find the most important athematics related to social endtraveling in his works
+to concentrate you to find the normal world
+it is part of the municipality of wilson which is the main town
+of course he was in code and was a nica his home town in the centre of the nippo archipelago
+the facades of the boat are well developed at the level of the ditches
+he then becomes a real world reference
+we are getting back nine minutes later
+in order to commemorate everything in paris
+she filled her bell
+olking tenricua covers a circulous area of the pout diameter
+its capital is the city of butnan
+it is a singer it is a singer's poet and producer
+the say jen cluf is made by the bay of frendy at the new brunswick
+the kaisota would be located between california and oregon and the rain by the pacific ocean
+charles and teter leskine pronounced for
+he was influenced byth a the author of the humans of adenta
+sold by the mysterious air life is a higher intresis
+a council is in charge of defining macroconmal policy of six was instituted
+but located on the road of the invasions it was constantly piled and destroyed
+when he relates everyone the sound stops
+he then settled in paris
+the headquarters of the prefecture is in tropigon
+he is now used as a town hall in the municipality
+he then returns to france and leaves to paris a factory of crinkled porcelain
+this album made the tempjons at any new generation of women
+the best time around each car is registered on a broke background
+during a long time the place was dependent on the saint cloud
+but despite this he was eliminated during the beetles
+he belonged to the van copen family originally from strasen
+nearly one hundred recorded
+finally his stream casper allows the resion of being powered
+the species is endemic to the states of gro and ovasko in mexico
+despite the tests that receive the game they are globally negative
+it would be very well it would be very well known with german masts especially in hamburg
+a third lodgist became necessary
+she is also cultivated in the world as an ornamental shop
+more than a hundred types of pictures were designed by
+brooklyn is a member of the illinois and texas
+he learns how to make up his furniture and studies painting
+from there he was a shooting relationship for matrix two in the new base
+other meetings are planned in yp in burdenheim
+from that moment on the humorous sons are mixing other more ingigested
+she participates in the siege of metz and the paris
+he is controlled positive and diaretic after the race
+vermont was the regular army under the third wreck
+but what is going to the past
+the first district shows a density in a historical monument
+all use of these data is taken out the expliciture of the conductor
+the notion of platogen is taking importance
+john for which he createsprestoric character
+she was the organization of humanity international
+his president is virchang
+in the rolling models the generation is characterized by the madal number
+the emotionals created auto governance organs
+it includes a polygona tower perced on a rocky stool
+jacob otman in dernel kinves
+the southern part of the garden is entirely made up of perennial plants
+replied lieutenant hobson in a low voice
+with you contradiction mind
+this base does not contain any tracks
+at the beginning of the second world war listres went to the mediterranean
+all the fins are iniform white yellow wish white
+a forest road serves this hamlet
+he is probably the most famous idea is the subject of a christian relame
+elegant the toro of jane has a fine aesthetic stag in quality
+the epithet specifics of many vices par to him
+despite the popular belief it does not contribute to masting the smell
+she serves as a community house at the villagers community
+we find tigers mojes and gores among other things
+they live in luxembourg by several of them and speak many languages on te text
+this finding is made by a store at the topic of a drog
+he comes from the cross roads of isles surium
+more menonites live near the city of cortuma
+the beautiful season was running out
+she proposes three services to daily
+he comes from a gray family of grit
+during this time buffy managed to escape three men by the concil
+in the middle ages there is a certain mining industry in the surrounding area
+he is the nephew of sylvy varton
+the exterior walls are flenged with buttresses
+kan sends to the police
+it is a socialist and a pera convinced urion
+this makes an environment of victims among the jewish population
+it's argentina and entirely located in the national park tirdel fogo
+he is buried the next day at the silly busa cemetery
+this forms a medium intillation of two pieces of empiens following the corpolations
+new gaski with an other different countries and speak salon
+the prefix the prefix adds the idea of a sceptre of excellence
+during e winter period we have to go through the six thousand carots
+the majority of the cathedral is covered with slate
+the station is built in stone with two floors
+she is at the edge of the evelines and your eleir
+close to marie bonaparte she translates works by fraud
+this cuppy has survived and is currently in the saint marie de sandles bacilica
+the company also markets under its brand news of import
+his personal name was princer cabno his personal name was princeer kab noji
+during this period he got married and was the father of a daughter
+mercesar relotting for the final duel
+would you like to be honoured to dance with me
+the merle family seems to be ending up under the one of carbiners
+some terms are made of ambiguous
+he is placed under the jurisdiction of the prefecture city of anzan
+poposi the german population was expelled
+he is the half brother of the artist huron de frise
+the situation is economic and then catastrophical as well as the political situation
+for the legislature patrick godoy is a member of the defense commission
+the municipality of senon servien is located at the confluence of vienna and cle
+the keeper also register two goals under the vase of the floe
+he had several conflicts with pote who was ally of jon de gand
+the main sports equipment of the city it has a capacity of seats
+he received praise from the general governor of canada louis de bad
+this species is found in the waters of mexico to peru
+the referee the referee decides to stop the match after only one million games
+john philip syrianny studied in journalism at the university of quebec
+it can be found in the county of teladega
+the rank of his memory in eric is a provisional and compassion
+the territory of this kenton was very popular in the imblb and plain of clicky all over the chartr
+one of the general assemblies is also dedicated to the pribation of the years cats
+the book receives the female prize bacaresco and translate in fluvila
+he plays in extra liba the first lavacian
+his production includes among others sansin bourgena style
+barcelon barcelon won the spanish cup
+however she destroyed their offspring stock
+he loved the powers which reach high
+the little one was born the green eyes covered with the fury
+one night they feel like a friend and invite filiber to dinner at home
+it is part of the regional chamber of commerce and petucherant industry
+and that's how in the region the monks built a beautiful kenton chapel
+both groups are still going on with their journey
+the first one is the son of tutorial and is destined for the game
+the inhabitants of chishe are chitans
+the whole of the swamp was the work of architects paul tornan and olivier clement cacub
+and yet he didn't know what he said
+the night's charge the nights charge the bourbons of a few hours references will be made
+the city is the headquarters of the archbishop of kentor
+he then resumes his political activities but no longer held physician
+several hundreds of people were moved forward
+an idea is to be honest among them broke down in the murder and the director
+he tested his life on the iser in buthels in munich
+the party then decided to unite the party then decided to unite popular democrats
+in parallel the recipropates the help of the government by intering
+furthermore the station of arigico on the circular line
+it includes liberals of great sized bens
+he is a doctor of honorary school fine arts
+to this occasion gabriel gautier claims one a last victory with his group
+pierre lesker joined the show for paris at the beginning of manuel at the beginning
+passionate about theatre she had founded her dramatic art
+the procedures concerning his hardend f
+she did not know a great excess
+he adds that it is off
+cruckis then performed a union tour and participates in the sodon rock festival
+the belly is walled after the first will war
+these descendants continued to occupy important ranks for several generations
+the island as in vere populated agriculture is intensive
+after his sporting career he held responsibility within the ukranian federation
+the head of the east front is musa mohammed
+a big novel occurs
+the surface of the island is made of majority covered with tropical forests
+she dedicated herself to writing in plastic schools
+the jufre became a hector for making her more fresher than alterance
+he resides four months later
+alitus alidesant is also the northern wind of the north
+manfem has a town hall studied in narbau with robin denheim
+robert carre also signed that mass has proved races tendencies
+robert and i he added exactly
+the war council condemns him to death by contumes
+the sale serves as a result of their own resentment just a source of liquidity
+these districts are located in jakary pega anil frize
+the fena is located in the southeast of the village of esses
+this limestone is particularly pure
+he is also strongly involved in scottish policy
+the state of emergency is then declared as a gift fire in the grand tennis
+we can access to the portico by a part of four steps in morin
+he 's a member of the grand military of the initiative movement
+cards of genderico rio in the north of italy in prague
+he coordinated civil protection and security reactions
+buildings on the annex were collecting guardians
+one thousand one hundred minority he is out of the socialist lists for election
+since the node has been more able to see
+during the revolution the parish of se cristo de danzac forms the mnicipality of danzac
+he really does teem with samton prints and turns into a white car
+he is also estimated for these portraits and his caricatures
+in french all ways yuthis can also be found in the insertion of the upinic consinent
+among his particular clothes we were counting the guards of gion
+the portal still exists
+besides the tower is also exploited
+the second season starts again or the first oneis
+the name of the dutchman he generally became a key to determine an overload
+its population is expelled
+the selection represents the netherlands to the woral championships on the urian rodaspars
+critics mixing grims who loves his talent hit him with the parisians
+however zu chose to stay in the fortress
+she was aiming at the reprovision of the airport in normandy
+weger de la frontera belongs to the most beautiful villages of spain
+the bishops contained his chimino style and sa albert krite
+this species is named after vira cucha
+from that moment on she often lived in england
+spancrook also contributes to the remix of the son of benny blinco
+john gifford was the son of a worker
+administratively it is under the jurisdiction of the midsax district
+the velves are suspended
+it is located south of the mount vano
+both of them are caught by the rather longest
+the culture by this one is more delicate
+as for you you have a chair
+this animal is very thin and angeline
+the nearest post office is in norton
+john pierre galle was born in zunsa
+the inhabitants are called anter villas and the anger villas
+his career uneasily earned him great triumph but also great disappointment
+according to the first testimonies jocki would have been killed by a mortar
+the sound engineers on the sun are handy todd and jim hebis
+ayuguslavia is the nomologists ton and the similar circle
+the offensive of the alliers in italy inaugurated the campaign for the to vale
+these trips of successes are among the best watches of the arctic explorer
+the towers were also oversead and crowned with slate bullets
+the existence of the series of people who were also main characters
+hir studied literature and history in bag hir studied literature and history in bag frebourg and cologne
+alman redor studied medicine in marburg before the first wild war
+he gives a bachelor's degree in murdy
+in total the majority of articles censored are not leading to the military topic
+a hill in power has been planned
+mark calson wants to play the role of you on stage as a tourist
+the temperatures are not coming down from the top of the winter
+she took place in the hipparian theatre of the park
+the city includes a historical district
+the body is forbidden in a hole often filled with ground
+these spellings were built remared several times over the following entries
+he has more than a thousand inhabitants
+they get the government to exile in the profets of the city of constantinople
+he realizes a good season as a defender at scusto a team
+the venus decorated one of the many roman cities of senusa
+his brother shacks remi is also a professional football player
+let's consult to the frances eru he commented on the olympic games of snowball
+the composition only has five centuaries until fuffrey settled an independent
+it is installed at the highest point at the previous signal judesic signal
+it is not compoable to the real millanism of the black panter than this is rare
+this monoplane equal size was one of the main aircraft of the illist face
+it is part of the socialist group and citizen
+according to no according to nomics about which their house is due to the storm
+they can be explinted in a language or implicitly in practice
+the filiation with the lords of ame and hypothetic
+however he does not announce the arrival of randy tors
+he then wants to dedicate himself to god and write to his wife
+they then have three children two sons and one daughter
+it is essentially certain nocturnals
+he crosses the northeast of the republic of seca
+it is part of the gulf of pierre legrande
+these two civilizations are played against the main source of optating
+this stop is dependent on the railroad area of bordeaux
+i am also taking with your fierceness against p h d g a n i c which is often looking
+it is dissolved in fact the adjudicator of the first time
+the building is then a standard example of this company
+the nipper international airport now occupies the site
+his paintings are of mix of cold and warm colour
+searching home
+it is played in a direct illumination format
+this provision does not apply at the beginning of the war
+the works for the construction include the caliser spolia
+we can see where they were from their conversation and the dignified
+the best baltim led the proters team around the world
+the pill has a temperament in full of vivacity
+it is located in i it is located in midstenses in the city of gaza and rafa
+he is curtly headed by the languist claudin chamuro
+the double's event is won by lenelicoa and natalitosia
+the three tourists burning understood that they weren't alone
+the longevity of the averages vere present
+never a young man she added with a soft voice
+its capital is el pile
+all the permanent accounts were minimalized by the creation of demotion chambers
+he reined on a vast empire of central asia to the principality of benez
+the writing is in gothic character
+furthermore both stories left to hear the fight took place in the evening
+the monks were scrupulously noted the state of their domain
+however this dispute with paul sellin defress the chronicles
+his galaxy belld from a disease is located on the sand
+she is specialized in delinquent in columbli
+it is part of the park that you do no pack
+he is the same in the tibetan buddhism
+this is why the one of these depluations is assassinated
+at the c the dress is meticulously controlled
+the passers are thin long and hobvy
+the album is produced by the musician director
+he also reminds them that only one of the red state
+the city is served by the postplopinus line
+seven gordon is a skisevon who was guided by his mother in his childhood
+he is venerated like that and martyr by the catholic church
+she enters the school of fine art in toulon
+urban of meellen withdrew after this episode in valinsor they are appearing
+leon copvi is the son of baptista cobeli his date of birth
+they are used as astrilatory or persian in this way
+the legal protection will give him the opportunity to restructurate
+it is found north of central new york
+in the air force imbert wins the spread for the second place in bombobard
+she's integrated into the third five of the americans of creen mesin high school
+the municipality is located in the tempero oceanic temperate zone
+fennel is part of the academy of toulouse
+between fighters anetres is the way of a street fight
+this force is fundamental and the lowest of non gravitation interactions
+the castle remains in danjan and is still a tower of engeles
+a number of leviers is generally subject to limits notably for germans
+she was celebrating the first judge of the month of shower
+the same year he bought some interest in jelsey
+during the marcel during the marcel hassain was resistance fighter
+it is a thousand favien located on the moselle in the metz cave
+the management is to be made for several buying centers
+zep has two children including the mother was elin bruller
+he was particularly used in india from the colored like the miniatures
+in general the gemino dials are easily idrated to form a group
+andrew norman did not give me the same name in alley scotland
+no more mister
+it is part of the f a f i region
+the music resounded while dax arrives in the gepo
+the first book of villains contained in singing in poems
+it is the most south point of cambodia
+this species is endemed to the mankoches state of megaya and india
+there is no doctor and i put to live in brugier hospital
+barpelada is understanding the river pilots erected before the current foundations
+the urban unit of bozansan is the urban pole of the urban area of bozansa
+it is the latter who advises him to make baseball
+these ones however turn out to be a great
+during his retirement he studied at the north hollywood california
+it is the ideal place to start the rise of these mountains
+thousands of jewish children were escaping from the deportation
+finally helena left the young training before the release of the first century
+the stables were particularly transformed with adding several floors
+at many species it can be very complex
+it has urbin centre in the same name
+he will be ward his masters and dissuade all intrutured by his breakings
+the eliminated are marked by two incidents
+these different edifices separated from one another are connected by a courtyard
+he cand however be neutralized
+if the rest of the stending it means that the presages were favorable
+each chronic has a game of society
+since then the lighthouse of planiers signals the approach of the marseilles abbey
+it is the fourteenth lavinia championship title in the history of the club
+we also have to hili huge quantities in solid chains generated by hotels
+it was the first line of french defence during the battle of the orders
+the allot is a lot interacled on herself during the separation of her parents
+the secretary general of the union parliament is appointed by the union parliament office
+it covers the seven tenth of the terrestrial club
+in france the elfa is a fight against fraud in the insurance sector
+in many cases people start to avoid horses
+a precious horse escapes in the surrounding nature
+she established herself for ten years
+links are here with the russian chinese and korean agencies
+the film was released around a group of former batters in montreal
+he travels to tryst with his former student and friend gebaski
+the doctors of the time have encouraged the doctors of the time have encouraged the population to epo
+the stained glass of sage sprub in a broken arch of a gothic style
+it became one of the most common approaches of the first world war
+bosman is called back from hugoslavia to go back to shelnow
+she also acts in a prevention optical
+the line is opened by the stage
+in a vulgarization approach when the minister came out in the media
+traces are ritten at traces are ritten at tests from population from one thousand
+his writings are considered as the restaurant owner of latin letters
+these are all poets
+observer the sculptor is the narrator navig between death and love
+the beginning of the leone river is particularly limited
+he was sickly
+the accumulation is of semi the accumulation is of semiathomic type of coloring horizontal tail
+after the annexation by adolf hutley to the south asbistram and his mother
+he's aged thirteen when his mother died in moscow
+it is built before the death of president ofi philip johnson
+alan gine was born in the paris district
+the network rntary is notably put the contribution
+it is often compared to the muck group or calamero
+for me this church was not well oriented with his tournip ated in the west
+see on the surface
+the doubles event is won by mercy de spas and julie
+theexodus chibnes exodus of the artropods are obviously armors
+the eyes before the front were becoming a porter and they are smaller in the diameter
+at his death he was buried in the culamac
+the benefit is actually a part of the value added
+there was also a small interior chapel
+he spent a year at never before returning to normandy
+the first two of each group are qualified for the second round of qualifying
+the biography of rosita hawett is not well known
+this lamp is made up and sold by the flas company
+with annexation of atheist by turte it is taken
+he is remarkable that did not contain any definition im
+two barges are on the island during summer months
+patricia patrista wrote dozens of books about cooking
+it is the animation was an order of the department of molecular biology
+burgher who is the second one keeps this position and finishes behind the winner
+he forms the triplets line with nike de cutre and teller jones
+their degree in evolution is quite weak
+god is in a hurry
+in turtey there is a strong soviet territorial of vadication
+gandy approached gandy approached the subject but never practised it
+as an economicist morcel cote is specialized in regional and technological development
+but there wasn't but there wasn't enough to import
+she returns to cuba without him
+he is also a member of the royal academy
+the section connect rac to mescet was being studied
+his coach is married to this techna who is the coach of the city
+when she finds jess when she finds jess the old feelings are surfaced the time of a kiss
+kenkan differs films in cultural and religious shows
+the castle was initially lifted on a plan quadrangular
+they called them gozani
+sho is a double star
+its belly is white and is strongly dark blue with black
+just said the captain nimmel but he will not give up two days
+designed by bryan eddy and program designed by bryan eddy and program by leeman shitz
+he is part of the olympia ward complex insbra
+intercreasing between the crystals of durlet digenitus and cable
+he is the author of novels poems and micro fiction
+these two measures are carried on the syriac and aramenian crystal stinian communities
+it is the parish church of the sopality
+he insures the link with olympic and french sports committee
+with his death his friend line of outstanding purity disappears from the market
+she's defeated in the first round in rolangross in magdalened
+later she got interested in the history of national archives
+the cultivation of the cressant is usually catica
+marilla de ristavi also participates in the recivation of the georgian christianity
+a server brings to the table a dozen dishes in small quantities
+she winters in the chineer stadium
+he was a doctor in medicine at the faculty of angas
+the primus room is made up of two separate parts separated by two columns
+the estate is controlled from the ogoo castle
+there some authors consider as a reverse capital
+in the cardival warning the quality of the wines of the tv
+it is found in itabumer
+she finds herself for the fourth consecutive year in france at the start difference
+oberstende is located south of shimit
+in some games only the mixed strategies are upti
+the addition of these points enabled the addition of these points enabled to establish the classification of the triple crown
+it marks the appearance of terry begard a character appreciated by the players
+rosier is very de rosier is very ditocode and decorated with radial messages
+he played the role of the black hawks of chicago
+the movie is based on the tiblical theme of arcanale
+this beer made the reputation of the brewery
+it is equipped with two platforms with ubry entanglement
+a eternal prison for more
+this latter will become a priest and calls for the next year
+he was not allowed to photograph the emperor
+in the vessel glumberation more than forty kilometers of bicycle tracks are set up
+born in athens crish marver bejolly started painting as soon as he was a tenager
+george alas stops his career as a player and becomes a coach of chicago
+she has served as a garage for a long time
+it is in this context that its initiative is based on coffee
+after a master he returned to rent easters
+during his trip his brother who had accompanied him died
+he is the son of lowis de gears and claude de cecil
+the capital of the kaza is bidlus which also bears the name of spel
+his death raises a wave of indignation in the international unionist unionist circle
+bunnan pays only a glemeration with din whose love depends on the love
+the servant wears a knight is called jurifeter
+it meets toward san orentino
+this selection is far from certain unanimous
+jondey started to draw and children
+after that price the john pettis clark medal is rewarded with his own prize
+the suffix is a local suffolk indicating the abundance
+the cross were lost by the national sinity but kept the existence of their kingdom
+after the end of the war he plays in trimenes and morta
+the wild love of crystal is important
+initially designed for the professionals this meet quickly conquered the individuals
+from then on he decided to give himself to his lobac always
+spherical shape in his youth is colonized or cylindricical
+borgial barry is located in the mouth of the ns
+and since the evening they went back to the sofa
+there is no rule spelling
+the great majority of the population in frecarbian is of jamaican origin
+his research focuses on topography of romanticism
+their marriage is then cancelled and their mother is still unknown
+all the communications go through time
+these members of the time wrote the rules of this port
+this makes it possible a non commissioned origin to the gals war
+endlessly endlessly he's hiding his knees in the herbert dressing room
+the sessions took place in january and march
+stronger and hot occasional rains can occur during summer
+however a few months later the first wil war was about to start
+he sings in paris very savine
+a material goes in the semm islands obi the teacher remembers his childhood
+it was edited by martin ucus successor of the printer cristo plentin
+the two main materials used are the red brick and bluestone
+he voted with the majority but refuses to treat the sylic of
+he was also a satwor speaker and notably otered several funeral occasions
+the clin wish the clin wish to cross the railroads without crossing the station
+from the month of august he joined the confederates as inter
+species of this kind are found in telama in columbia
+the villa also includes necropolis antic
+the formation is made up of wilfried ribbon barringer e
+minority languages are known as albanian buskiin and cry
+the writer can be found in the french language at a micra novel
+it can also be useful to obtain sleep
+he was also general councillor of gerz
+we confined in particular to the state of beggar and the
+the president retakes his duties later
+later lisa and louis were crying together
+but olivia and bernard penner did not bother with his power
+he was four to five thousand of circumference
+it has a total of four members at the temples of the locality
+it is found on a spanish square
+a call was taken to asaka his aunt succeeded in running away
+the examples are mainly her motology
+the wards handed over every year at the autburg palace in diena
+she was founded by the king of venapiarisa
+his destination is san francisco that hopes to reach less than three months
+that same year he published a new catechism for his ices
+to day they are used to divers to day they are used to diverse techniques
+a museum distinctively survives in hagen the coral ersthusium
+there is also the iron mickel copper
+this psyca remains major non published jonistay
+he left school without qualifying but ended up doing something out his life
+the municipality of imart is located in the centre of the tislaer
+he participates in the more a plays of roles on the table
+the color is honestly all year round
+she starts by working for the architect wiva calio in al siki
+his construction is the subject of several delays
+the sovereignty had no litigation finally
+weapons of the crime were never used
+she's a general of ri she's a general of rectangular shape in the square
+moni a belgian came in paris to learn painting and became an artist
+pickern is the closest inhabited city after two to the west
+i represent clients established in the parisian cavarets of the red mill
+he is named in honor of the american astronomer anegilkan
+the double vent is won by nickol
+its end is located at the canal port lave near stewart
+he officially declared that he was neither on the left nor on the right
+we create a kind of blows in several ways to reach it to the h
+thus the length of the stops is a third of the long range of the original stops
+this type of speeches is greatly less objective than a direct speech
+he is the first mayor of democratically elected in the city
+a descendant of the lespian family creates a camp in the sixties
+she is also called tortony she is also called tortony or odenes
+that is to say the argument of london
+menino then joins the tresers club the chicago volls
+there is no information about these evidence anindit
+once past the shadley we can access the renaissance to the windows of mile
+its capital is the city of garza
+this reason is often used to provoke the dininution of the proteins
+he is directed by the foreign missions of paris
+vidias relating the collection of the treasure are very projected over the exhibition
+the ruins of tilis serve as a career in the surroundings
+he flies in one of two generations in april may and june to august
+after his birth he adopted the management of the southeast and then south
+it is the interpretation that denis levant previsitivity
+the municipality of makof is entirely included in the territory of blata
+he spent his entire school at the charlemagne high school
+the circumscription was located in the centre of a burga
+lady yes on eating me to day i be too beautiful
+he dies in new york ten years later
+application badi is ending to the queen of the dynasty
+he is known for his maritime landscapes and parliament
+fernando di herssterm is associated with him before becoming an assistant shareholder
+it is about the word of culture that have all been made
+he was a student at the shalloserseo then at the high school of barcalium
+it houses a biological study at the university of catarn
+he followed his studies at the sa claud college in ockland
+sotelences his sister fibrbax only available in some swiss cities
+itirrigates the communes of liqes bon and
+the roblest the roblester still divided into other units that vary according to the environment
+these trainings are capitulated in the fight against the segen and freizer
+as for cor setti he disappeared by winning the campaign
+it is the letter a diacritic of a circumflex and asidia
+we will then be able to tell acts that will not be
+in texas he was one of the first reconciler to a mixed recorder
+one or the markets are not introdicted and say george de cem
+she was then used as a powerless retroversary
+he plays as a goal keeper for the club morse well in scotland
+however the exchange is cancelled after that
+maurice de gandiec was also the first french translator of the french
+convinced sophie then had his investigation to discuss him
+despite this he ruled a seron between the two atories
+here is a stained glass pencils and she looks like an incri
+she is also often depicted at the mants horse and a water level
+by committee it sometimes employs it to designate the scottish groups
+genetic annalycis have shown the problem
+it is now managed by the glish and mitich
+crape is also a name of french family
+the territory of the islands forms legally with three states of the kingdom of the netherlands
+he is located in the jungle region in the centre of the country
+flaubert and gattier wrote articles on her
+his eligible and gay house are more often related to the digout
+it was the bottom of the charges and closes of this lost
+after the battle of france he won england
+it is one of the essences of the indian philosophy
+he is the painter of vinemon
+belly is the daughter of deploy in tussing their concern
+because of his tons rehmano couldn't take the play without any resources
+the prioress the prioress took care of me and i'm more at the same time a director than an architect
+she expresses mainly on american political and followers
+his son fredrick ode stvincobut was also a painter
+this mode is developed in the french department of footing
+the code name is operated in the red band
+he insures a transportation mission between the fleetship
+virtuues will go through france and touch virtuuses will go through france and touch mosaic organs in lebi and digi
+this proves the beauty of the eleven
+this one claims that the character created by conan dale is inspired
+the dovil's event sees lynsa devonport and marie joan
+the latter is imprisoned in smilins in russia after the battle of la burazina
+he organized the visit of robert lengells in turtey
+it is managed by the french stisis of
+she currently raises at fifteen meters above water levels
+promoted to the first division zorke still plays the time but mark
+juliet simon girard was born in paris
+tussaly alley was very text
+he had no facilitated face to watch the village
+his lecography of the rat is measured in the same size
+it remains episedically requested by friends
+the invention of tradition often focuses on places of memory
+we are having it
+inserzer avoided this closure
+the last ones are also available to the is of the prac
+iii no pay i am in a ret
+he showed you what to read in this way
+the plan on the other hand represents the piece as a more important quantity
+the facade is front of the ditch
+a first hypothesis is worth that this set was kept in the cathedral veris
+it is named after the trotter horse of the same name
+it is the first ones of the sidish revolution
+this latter gave him an appointment with the hotel space
+this accusation of heating was not proved
+the lady of merin is marketed as a concert for the two theories
+trumpeter at the beginning he is mainly known as a saxophone player
+the village is built as a landmark visible from everywhere
+leplesn incheka slavacia is the first union trithon
+he plays as a left back defender for vielsburg
+he was born at the home of his parents in the capeston street
+we also talk about new urbanism veryvug in the united states
+moreover they wished to make these seven parties a challenge by himself
+his inverse is a recent district of copevo and a city located in the southwest
+the counts were being built with he modest city
+a meeting between a young man and a man desperate
+the rounds one the women singles
+finally certain others parishes are not affiliated to any pacific organization
+she was used by the air force and the national navy
+no supplement nor sub titles present in the box
+it meets in the shian of linon in the cave
+the communities live in separate districts
+the operation succeeds with an incredible suffering from the
+he was discovered in the insulet by hansom ilchester
+the generalization of the teaching is that this meaning is disappears
+to c t e the scene confirms its acting role
+the realization relaunches the wave of production of films
+beat the minister of nucking to put his resignation again
+she's coming down with me as much as he is with the sex
+he also made some short films
+in china the starting age of the recipe is different as well as women
+however some fatois affirmed the opposite
+they are the first since the inspiration of a single part system
+the municipality is part of the community of municipalities of otery
+the team beats down and sirkof finishing the eighth place in europe
+the c e gave a rigid amount forming a detailed armor
+the patient crosses the three councils of salmina and sanrand dmars
+his delivery is black greenish pronotone and red
+it's a duncan snake on the earth's tide
+son of a commissioner of the french wars in switzerland
+german invaded the region before the blucks could not be used
+she's animated by devil as well as other hosts
+the pope is also a single group that is considered
+he needed a constant maintenance to work assessment
+he then participated in the italian and austria occupation
+after the death of asia per left norway
+buncho buncho is bordered by five other municipalities
+each of these two circles had written a literary prize rewarding a historical book
+these two dams are a hamlet of the city of milles
+it is the only ornithorinian in nostral africa
+the lower work is pressed against the beak of the instrument for him
+however a sis of suicide to cultivate transplants
+she extends through the street frederick peltier
+elizabeth melon then lived in aberin
+he had a first time experience to succeed in the music world
+this formation is offered to the court police academy
+the chaos is located just next to the desphere strio
+the battalion of rengers will lend the place expected
+their dressmaker is from east of north america
+following the route of the transformation of cotton the author had the ruraldization
+the prairie did not give up this destruction
+the joy of the union was still increasing when she arrived
+his editor in chief is crisp his editor in chief is crispar and ark his editor in chief is crispar an archaeology teacher at the university of derm
+interdisciplinarity and internationalization are quickly becoming the keys of the research
+the house is the chateau de montmorency where it is sheltered
+see also lucreht
+in africa the red and ucker colors are the initiation colors
+the capital ovara is located in the island of radagunga
+later he became a training ship
+the public school has a meeting with inter the public school has a meeting with intercainal logic between chersa and
+the upper pines of the port and the tail are black with blue rocks
+she comes back to europe after the end of the fight
+he existed public notary in the roman empire
+his story deserves to be counted
+the recording of the data is done on a wide band of two
+these committees are called special committees
+the national passage is difficult but serves as learning
+this property operates for example in the resolution of agibrace equation by radioco
+she draws her first characters in greek alphabet
+he went to the siege of this city
+edited a naive classic
+the nunots returned to the manor
+in the church is a gravel for several lords
+the lutour family will take the title of baron of sainvide
+the black and sart were getting involved in their side
+its decoration includes culums corinthian and the communes
+the weather is located near the west lake
+the grape is conical and ash
+finally progressively fell
+she ensures that the province of this money doesn't indicate the legal
+the station of china maguche is served by the shinksinsenian line
+sebastiannay was born in stepenstread near ottoban
+the brick was going to run into the bay
+the quebec has seventeen administrative regions
+the telephone connection is suspended
+the group of procenians is considered a rankof by the systematicians
+species of this genus are found in brazil and chile
+some of them know the inreparable mistakes
+she links near diesa to zida by masaka
+portugle is forth in this world
+they are in nature under the minealgy name of masconite
+ses blavich is the main town on the administrative plan
+he then played the socialist candidate john pierre bella
+marie montanova wins the lady singles
+there are many of them varying in topons in gaba and cameron
+he is certain that his name is in the breton armory an toponomy
+the main use of missilum is fungus crops
+in doing so the material increases its elastic energy
+the table below shows the calendar of the ten alpine trials
+burne castilka is a stop of crusades in remsel
+after studying at the bessu college he went to study in france and ren
+he excels there especially in the art of the reviger
+they are all impossible to reach
+this species is endemed to the marisian peru
+unfortunately it is known as the elements of the age in cass who are not secret
+if we are separated from the space so it can see the condition of unfinishable chains
+there contributed to the development of the city of x n provence on several aspects
+he is the only one to have beaten the olympic champion charles jacob
+the two public institutions have therefore the same general election
+however this is not achieved by the standards
+their origin is often well known and the malcumbry revolution
+cartley holder
+the history of montresor starts with the first pieces of the fif in the text
+he directed the frescoes of the roman church saint nicholas de lorand
+the controversy surrounds the two teams that are invited to the event
+she was inspired by the german tradition
+the taste of wings in the tail are black
+he is the brother of the cartonist george bastia and usonising bastia
+it is the last work of changeor music written by mozart
+the israel state recognizes the caribbean as jewish
+in fact a young independent woman he works as a photographer in the united states
+more than an average of the fight is high better elected
+sometimes anti semitism is the same year according to the period and the leaders of the people
+drogan was buried in vinosa in the abbey of trinity of venosa
+costading was popularized in the united states by the culture hapi
+he was notably skilled in the field of water management
+especially against localities of the world bank
+in borgno he is also possible to decline female names to the masculine
+it is found in cathligan national park
+on the other hand the situation sachon has the characteristic of being a climate
+peter robert was born in tapeca kansas
+in the region is only a hundred cassack riders in the region
+gruff illustrates graphics
+he belongs to the ware worm club
+the keys are converted to receive the necessary box tight for the new armidale engineer
+ah what a confession
+forasmuch as the people protected this article
+he also became friends with george seperis
+it declines in work workshops conferences and exhibitions
+at the beginning aherk tries to sagerbles which rejected his advances
+the start up is then an official launched by edward rudolph and edward pulles
+the latter is quickly replaced by berts knack
+his first confrontation played on mount chimico
+its superstructure its superstructure is an aluminum
+however their statistical data does not correspond to those of the stars
+he came to search for a tournament towards the history road
+he also participates in documentaries and has several to v series in time
+he is buried in various schools but did not get their pard
+the highest floor of this set has the red edge of the classic group
+the tarban
+during the war he resumed his activities as a teacher
+she appears from it
+his work was translated into japanese italian and arabic
+the night was fulling the latter accompanied by other members of sat se the night was fulling the latter accompanied by other members of sat setchers the massacre
+to morrow the other goods are being demolished
+the degree of the withdrawal of the french nationality to the imigrated judge
+as a record is located on the hen and the south of the island
+the county of arnston is located in the centre of the state of texas
+i am telling him that this was a white one
+one of the most common is the teno sanovite of de curvin
+the beginning of his rate is marked by a recreation of violence
+she's interested in the painter ernest mancoba and gerard
+the funzer four hundred and fiftee meters
+it was the first coproduction of french iprize
+all were sent by the italian navy
+the sun was a singlo vichats during the early morning at the beginning of the early stage
+the municipality is crossed by luncibo
+the images of the repression have done the world to
+he doubles the actors mark stroncu and giffer
+the capacity of the fuel tampe was increased
+it is the most expensive fill in the history of mexico
+he is noticed for his talents but also his body guards
+the detailed mechanism of this reaction has been quite complex
+it is normally transeced before being briefed or working
+these cauthers are divided between the other and branches of the state house
+there the mon sent on earth by the game
+these two sons robert and andre de tois were also architects
+between the two the majority of his court is located in sherindon
+he is common to hear it without managing too locally
+electric food is provided only by the sector
+they are widely used as private
+after the armistice everything is calming and we're confiscated
+pic te barret street is a public road located in paris
+in italy yes we can see the news of the country
+here is the list of the main recipients of the incinity in the notse party
+patrick meson analyses this as the triangulation but through a cruel strategy
+charles beneset was a forier and a nurse
+berthel berthelidy defends them with bravely all decided to nat go
+the legress was made in position to the left in the room
+i imagine how much it is
+denis paseta is the former drummer of bono aps and the mentor of the god
+the shame took place at the season of plutor may to november
+she continues nowadays to give a general jewish education
+due to the age of farmers the future of the race is very worrying
+but he doesn but he does't but he doesn't know anything about the farm's job and the chain of the gafts
+his mission is to be a catalyst for transformation into a lasting world
+such systems improve such systems improve such systems improve the ability of colling wenty planes climbed
+the house is not located in a sunny sand elly
+this species is endemic to mount carimicur innezuela
+this good results is due to the boil cosolitic buckets
+it will later be launched in televassion and on request
+its name refers to the region of pusse and the forter
+she then worked as an english teacher
+in the road the name changed for edward mutpat
+he was known for his role in the milngay group
+jenny ferkint was born in biskin
+at the ground floor the stores of alayin offices
+he wrote to his debut with albert merra his cary to the prefecture
+sadre has a private institution in the secretion school
+he allows to get down to several hundred meters
+dominique couvver is married to henri christine du viver
+he was revealed in the salon de janet
+the music is part of his life since childhood
+the sarvieu cassade is located at about
+the coasts of japan are in approximate northwest
+he doesn't he doesn't exist any more than thirty years depending on modern classification
+it's perhaps in the use of judas excelled it
+he founded the university of international law institute
+the birgeon airport was then inaugurated a second terminal at the end of the journey
+my vessel is repaired in one week
+son of a benist and he is the nephew of alexander cabedal who will be his student
+the goal is to be in the village and used as a field
+it is introduced in the rhine basin and the rhine
+the castle was built in the and the ph d from this period
+the priority signals are all reflected or potentially lighted
+he settled in leghes
+its president is entry papve
+can you let go of your houst it is so or to come help us
+he was a prosperant friend and advised by johan de wit
+doctor in theology the same year
+the fate of the founder of the medursa ends with the tragedy
+this species is the most common desert in porto rico
+the enemy precision draw is taken but is eliminated by its chinese amology
+he wrote three novels and two collections of short stories
+cesil relief on the steep slopes are uninhabited
+even if its account has increased quickly it is an element of the clop
+it is some popularity in greece and egypt and morocco
+menlin is part of the academy of toulouse
+he will become mayor of the chavelard chapel
+the intervention most commonly conducted is ministi receiving
+it is a key success for wi uterer
+it is apout writing a script riter of an action film but without action
+the reception of the critics is positive and borsasko made it to be praised
+when he returned to rome he was accused of having turned off money
+he concludes the belgian cycling season in october
+the volume is the fall the volume is the fall up of malby and software have evolved much
+he must therefore repel his project invasion in france
+brancopes died in deman
+each race would then be great experimental model of this basket
+the pressure applied to preaching and some tents of bars
+only the noise of the frustens in spurts slowly drew their calmness
+the municipality is located on the vexent the municipality is located on the vexen dorman plateau in the east of andiles
+all are prospered in theatre in laughs tired people
+their menda is also five years
+his mother is a gentleman of warhead
+and he worked as an illustrator of livid
+the buildings house a public institution for the public
+cromasny according to the end at pinal
+it serves during the entire war in a battalion of genius
+its headquarters are fet
+the architectural style of the monastery is at the cross roads of roman in gothic style
+he supports the creation of the hinisqa
+it was named after zivana in the slaves of life
+son of a baker he is a doctor in law and a lawyer
+he was in charge of two superimposed plates maded in beautiful effect
+he thinks he is of spanish origin
+the quality of the dra is very important to realize the inter opparability of complex systems
+the activities of this latter showed foren was enough for the officialization of this time
+farmers sell a lot of money on the market for the land
+during the last decade he extended the vegetative period of three weeks
+he wants to rent a night club
+born in new york untrained he was the son of a prosecutor
+she belongs to the company of east
+fernand street won many individuals in the course of his career
+it is the heart of the former sectuary
+its chief town is the city of hunch
+the abbey of batier grenville was born
+it seems to be education and almost always necessary to limit the handicapp
+this version has a metric pressure
+the details of his life are relatively unknown
+he is associate with the local cycling club
+she has a brother cobrady
+it is also a visited site by fishermen
+the rest of the tax takes place as a standard attack by returning to the library
+but it's a nice story so i started again where i fell down
+while the portuguese come into the game in sabo scores for gos
+the lady and her lover are banished
+the river is the sandman scor lake
+they have recently been for most of them sold on the rent market
+its headquarters are based in cerebrac germany
+t a of the elbum vinel is made up of rock titles
+oh my father really are you giving me
+this type of meeting is renewed o
+the french label sison of mist also occupied the distribution
+he also became french champion of de carin
+he obtained his patent of a sportive silling school teacher
+he measures approximately long
+he was also a consulter with the congregation for the eastern church
+two biseries were venerated by the javen as the god
+he was supposed to evolve and often in searts from delnot or imorim
+the defeat of petite is interpreted as a defeat of the government
+the stup overs were made at the port of basesles saint bet and fasse
+financially it is related to francars
+in the past they're cultivated in france
+she doesnt she doesn't manage to confess and make a nice difference in her way
+wickham is buried in the icorale cemetery near ashland virginia
+he gets the warnerly station an
+he retired to versailles where he did not practise medicine
+this area is the same as the one that produces the porto wines
+at the base a white light enstrolled the strands er darker
+however another option exists doc fright
+he enters the architecture academy
+now he is unable to express himself from verbally frederick to resolve himself to use the writing
+he is named by his ser the alger
+he is decorated with the resistance metal
+friends of the old regime as practised a lot of skins
+parisian architect gim trepsa was the student of voluntar
+he will lose the titles in april against jet carter in corgreens
+his tricky form circumvtions that allow him to create extremely strong
+debuty had a succession at charles jonlet's chanber
+he is a member of the order of benedictines in montcassen
+the rumor is now emaciated by tenserer who is hiding this street of an eye
+a few years later i am totting out the kashin who was clayed by the imperial university of tetu
+he was defeating all this screamed stanley quarter
+some sons have their own devices in publicity and in video games
+the principality had no right to hit the company
+but two men peter tesson and paul but two men peter tesson and paul nitson were not recovered
+the site is located at about a hundred meters away from the l
+their main rivals are the members of mexico mafia and their main rivals are the members of mexico mafia and the rin broser wood
+the bille river has its source in the south of enselle
+there is the blood you know
+don'tdon't he have to give up everything that looks don't he have to give up everything that look like a hope
+this set of symbols cans a set that encourages meditation
+to day this tribe is totally made up of a farm but it has retained its merds
+his companion pal dono co rote the film and we'll make him
+its parish is considerably thin and fragile than that of the ard
+this geir theorem came from the word brixton or brixon this gelwir theorem came from the word brixton or brikstat which means magic
+the city attracted mally inhabitants of bristol during the summer
+from his youth munter from his youth munter appeared as a very promising being a teacher
+this line attracts many opponents to its supporters
+he studied at the university of everhard in tubingen
+the head office is located in cumbras
+the son fird ran towards his father and told him that their dog is talking
+as a principle the university is open to the international level
+this presence in the act is the trace of a rather close relative
+the belgian runner gregfed avermat the belgian runner grigve avermat the belgian runner grigved avermat a member of the latter is a yellow jersey
+if you are on water it has the particularity of being close to germany belgium
+the second world war destroys some buildings
+this audacious text also marked the end of his presidency
+he is known for his research on antatebetan language
+alleron beckle started playing football at the votery mar
+the break up with textal adventures is permanently consumed
+the church is belt in choche
+the term is used by alpach during gedlins
+his specialty is seremic
+the church suffered from religious wars
+the meeting of these three individuals seemed unavoidable
+it can still be covered by zagari mountains
+ammunition promised by france ended up arriving through he greece
+it is bordered on its north sides and west by a wall
+the guests to put in the gendarmary of germany settle in the buildings of the agriculture academy
+the ammunitions used for the phosphorus or the termite
+the state council is eliminated in the fourth time in the united kingdom
+he received a special prize for his part won by fleur
+the series was broad castidon in the united states
+francisco nipoletono was the student of leonard de vici
+this technique is complementary to research of varing at the time of the transit
+to date he is from new york
+it is one of the biggest known in the university o it is one of the biggest known in the university of ran
+delemma has held in the sedish liberal party
+daniel omally received a team in the mediaeval times in ohali
+only one man has one life
+it is part of the french and medistance committee
+the parties played intimate the two manners
+saint marnel was born in vienna and was december in the second division
+he began studying in san posco in concord in the hunter place
+the ray system is based on radar and put by several antennas
+bonavi then joined the lino and the bath of his hand
+this is the opportunity for cury to dye his chintz at the level
+his administrative districts improve the traction on the glacier surfaces
+she passes from the north to two o'clock
+his presence is uncertain in tyland
+the pioneers go through the board undertime installed in the bartouet deaths
+his son is the trader john fledge
+we have more than five days
+in total were recorded on the different sites of the french super fade
+the dusage in leon is variable according to the resistance that we want to obtain
+we owe a first modern classification to the relegation of john francois chempelion
+the vineyard covers a large part of the valley of the casse up to grenpre
+nickbuton is appointed best player of the tournament
+their galleries in the press to become the spiker
+it is then in tion that he decides to continue his career
+he works at many times with evegask lora turse
+among his collavorators famous we can notice that or vocalist singer leo thomas
+the kuracu desert is located in the sat the kuracum desert is located in the southwest part of epricion
+maskady is parcel
+edward michael always wanted to stay in town
+the group recorded album for wax tracks
+species of this family are all venomous
+his second son is john maray from den mort
+octavian blowed even his team to five titles in the world successive and two olympic titles
+he was a student of gruvian and munich
+thon the first edition of the catartor
+he was seeing through a green village with many acology stands
+a case was reported in korea and in japan
+the prodido is finely grenity
+he lives in dussel dorf and paris
+the central part of satellite bears the name of malacetales or tazanova
+a single portal that was flanked by columns is surmounted by three windows
+he is currently operated by a sports association of the club of bourbon
+these errors are corresponding to a correct oral
+the original fund includes manuscripts of mark antois and john jacques crossoe
+before merging municipalities ruchesirs was part of the municipality of louvaner
+eler ment all of them are nine caps with the french team and scored one gall
+he exhibits in art galleries schools and many successes and good reviews
+thence to ticobrate from uracotus learns the danish astronomy
+during this time the reform gradually entered the fris
+the county seat is elizabeth tom
+he is the twin brother of heinrich lenvis also as hake player
+one day hunter's capacity to give up the possibility of order to the navy
+the human space is bound by a fence wall
+the weapons performances were better in a difficult field
+it is endemic to papo newguinea
+with the fliveland of south he forms the polde river
+the ido did not transmit the personal donations
+lichic latterly became the erandisement of western
+hollind returned to his initial training and sutmus to perform their concerts
+she now lives alone in the middle of her life
+in fact the nam bahman generally written bahman would have a french catholic root
+she's the mother of two sens
+the northwest car is occupied by a norman castle
+the jurisdiction on more is now abandoned
+it is located on the road that lisene in can
+it was a member of the legal team of oxipewastelet
+the rest of the building is marrying the denial of the earth
+he is the first sentenary to finish a race at the top
+one of them wear mke mask
+he played his first match with the ramanian team against the ukranian team
+cramme publishes his sons in
+the mayor is antonia salvain
+essentially active in theatre he articipated in many pieces
+i start playing the opens in your box
+the city has an important industrial heritage
+with the group pludy reason he will play scandinavia
+the municipality is well connected by proximity businesses
+some of these names move to me with some other things the differences
+in this first instance the vins capar limited
+a rumour said he is the father of the child of kulicax
+she works as an association with the police name
+roaldivis was elected as captain
+the urban area of saint naza is part of the urban space of no saint nazare
+he became friends with chegivar another member of the resistance
+these courts are about the effects of architecture on a bethrone
+several books wer pulling out of this fiance were published
+he is eight in the voar fire according to tradition
+he then i knew that the rents exist
+this species is found in panama an zila
+this piece is regularly played a bit everywhere in macedonia
+the german name suffice was also used
+the christian d more generally religious or peasant inspiration
+this one handed him prippus and made him executed
+the ruler determines what kind of players are starting to play
+his revenge is only moldian
+the physicians have then tried to expend this group
+an extension of the network is not planned in these conditions
+the building was entirely made to the museum's disposal
+his brother jason hilmer is also a professional hocky player
+one of the deserts were then expended
+the hunting and fishing are also allowed
+after the episode of a hundred days the general faon conserves his rank in the royal army
+the choirs are made by frido and phil
+alexander will stay at the mok outside the war campaigns
+the line was then shortened
+they both desire under the general denomination of the acadia
+the opposition criticized this operation is advised to refuse this ticket
+later he became the gitarist of the group after the arrival of simons stafford
+for centuries the river andrea vaselivinis landscape is tormented
+the priest is afire with the help of the inhabitants
+the canvas was part of the rare paintings of the early years in french
+the ceres tells the stories of cubus gross white and apero
+a plan of load is then carried out
+the same year becomes an associate with the study of humans
+some regions that are employed a lot of peppers an preparations
+we adyar to give him a little helpful taste
+he then worked in the trade in uperates of oil
+he is integrated into military engineering
+each convoy is accompanied by a doctor on board
+he also recorded all concerto in an ifor piano
+its frame is abeam with a wet and a wet tissue
+the life of clodette invented a narrator
+he teaches at the school in nebe de strasberg and in the universal esridian alliance
+indeed the device is isolated and difficult to access
+on board he met a beautiful young lady named anbege
+the dutty is also very widespread in the south india
+in this regard he can be considered as a protosylogical method
+many ampabiguous factories with variable trails
+jerry cole will take his sead to jury after three years of absence
+son of a captain of the sea he goes to the middlesburg school
+he trained at harmony and counter points with marbu
+all titles are two all titles are two ofesle bros and less stated
+the influences the style of his musical compositions essentially liturgical and choral
+it is located in the safe marmory district
+the gravy lake is located southeast of wid vincen
+list e happy see the cloud transformed di nice
+the authority of the executive will have a governor a vice governor and ten other members
+the cardinal trevolsa is a platante in vienna
+we're telling that rhetorick he mates worms
+but what does it mean
+cronos lives in a post pacitic world came back to prehistory
+he has only consulvative skills
+she's one of the five children of robert ignetus maloney and ester smith
+the fight fropo is then converted into a walk
+but the cliffs of poleroid seemed to be to say a future always in drama
+restored she continues her career
+in medicine the term decubis described again
+well bargell defends this late resignation for a few reasons
+many cadets are found without a land and without any future
+he composed many pieces in the majority of his instrument
+a square tower is housed in the main facade without abandoning it
+before exchanged in vienna he was born in a jewish family
+the belly sides and the tail are white
+it extends from the city of adura to the north of nicola to the south
+he spent all his career in the united states as a defender
+system mino extends to the southwest on the slopes of the valley of a
+the norway from western ends the majority of the after and use the minors
+notably the villa motors and scouters of these knowledge
+he has a degree from the royal academy of dramatic arts and he now belongs to it
+that same evolution is characterized sa pierre de ive
+he was possible to use the bibops of the company as interphones
+thus all hopes of the imperial house are placed in the prince ishito
+but nothing in his physiognomy didn't care that he understood him in history
+this section presents the list of norwegian championships in paris
+the collegiate has a cruciform plan
+the carias have several orgons of replication on each crimaso
+many members were exiled or imprisoned
+he enters a short time after the academy of fine arts and leech
+the general bucket of the municipal territory of mussy is double
+his son was born atwardo replaces him
+the teatre fronce is a street in paris
+naturally the marteri recounts the inhabitants
+a centre ethipical for childhood is open for the young public
+this plassans in the opposition to his party is probably that of the president obazam
+he is a student at the sa menot high school of saint barb van plessus in paris
+it is also a countryside service that is also a good one
+several dozens of thousands of russians migrated in western europe or america
+a pleasant moment to discover in family
+it is one of the most visited ardrems in france
+it is named after the countess bager
+the artist uses rooms to exhibit his paintings
+almost all series of are animated by ribradcast
+these quests are enormously non colored
+the first days of the walls are joking a teams perfect
+the chapel was dedicated to saint ja
+this one would have sold its name to urians
+she is going down and converging towards the rector in the coxey
+robert cutcher first cut his p h d in law and works on a law firm
+he is interested in the political speech in political psychology
+he was nicknamed sir
+he mainly attacks for the users files
+only the tourist trade remains between estivares and samvido
+the other roles are not created
+on the one hand these fronts of the valley to hevewell
+all the southern territories were therefore represented
+the department is not connected to the rest of the country by road
+it is a specialist of the date of ice
+but he did not have the doctor mayar's signature
+what can he want to do
+the dexterity is essential to achieve an activity of all day
+more than were touched by the storm and without uptoble sense
+artificial protection of woods of construction
+the altitude is understood between the north and the south
+he is at the same time appointed to the council of works
+the lawrence family is from bolswells in honesher in per
+all the villages are built on changes in the case
+she has also prepared a degree in crimology with sate provence
+the access to the lighthouse is only available as a bote
+he also published several mainly related to buttony
+it has four movements and an execution in front of aproximately half an hour
+davis will see his experience without finding no apiles
+she's married to custagoves
+here is a small calculation to fix the ideas
+the reason why the previous one is not known
+cooles was a student of herman berhard and bernard frederick albanus alled
+bu is a poor man who is a student of marks mat
+another is shot
+only then you the university of technology hid in china
+to day it symbolizes a recognition of the parish to its legals
+egypt is strengthened by this confrontation and finds influence in the region
+with this new architectural cross such a traditional process
+the actor is openly e sexual or in
+watercourse then
+the subject an compositions are inspired by traditional spanish music
+four generations of this family succeeded one another
+they were delivered in eighty waves
+he was then promoted to the rank of major
+from the great train stations are several junctions including the netsley manufacture
+the municipality is located at southwest of
+the speculation is therefore considered as essential by the wear of liquidity that it allows
+the protest will be dispersed brutally and will turn the riot
+this term can sometimes be opposed to the notion of producers
+the movie is presented as a requisitory against the death penalty
+the school has work shops as well as paintings and vidios
+he wrote sons by the way one of the sons of the felm
+the latter works are on the reports between anthropology
+he is indeed put in any ornamentation in any baroke style
+the inhabitants are called the glenday and the lemandes
+the japanese version of the album includes a path named after him
+and the royal bourbon is not a surrounder
+he has his colic andre vaffir and marcel gani
+he has a lot of destinies for the teacher
+she is bravely sneaky
+he is entirely included in the district of fresberg
+it was discovered independently by david pulitzer
+his sister michlippins is a famous stylist in australia
+they run by capitalists or italy of tywar
+and there's only kerpetson located in the clerch railway
+the cathedral is a country of sa pap currently a parish church
+his first goal is scored against cape in the african cup of nations
+silver plates could have existed in time of romans
+he killed smiles gathered in his memory
+he takes some freedom by a bo u f a
+towards the end of his life clive burn moves to the wheel chair
+the dab is the auxiliary of the conductor in his missions
+the houses are also larger in five rooms
+the house in columbie purest on the village square
+he studied at the sai john college of cambridge
+she became then became surgeon and also runs in a local division club
+he grew up with and backward the liberal political and progressive thought
+index becomes the subsidiary of segasani and managed by neaserme
+machell borges is the first wife of eclegran
+rinnie waits up at the hospital and went to the hospital
+the sipils form a calithian which gives to top
+the river doria flows through it from its neighbors
+he occupies a position of modest civil servant and gives particular courses
+he was then the youngest judge of the essex county
+she was following the old northwestern tumpoit in virginia as well as the portion of historic
+patrick bolognia has also one of the members of the political office of the presidential majority
+in catalunia he is called the republican party of escla and in galicia
+the blacks wan five difficulties the game
+the organ tribune is the chair of a sculptured oak
+he undergoes a long interrogation before being deported to berlin
+the boquet took off at fuss
+the northern part of the municipal entity is located in the wide valley of spircweat
+it is composed of he f of the man named after him
+a translator of several spanish arabic and arabic
+the municipality is composed of a beautiful but also a large number of villages
+back to koninsksburg he was in charge of his men
+the church built in the location of an old castle
+these paintings are made in workshops from skis made after the date
+born in france he spent part of his life in the united states
+this novel describes a similar afternoon in the terrestrial life
+guards were organized in the main cities of the country
+during the cultural revolution the church is vendalized and close to worship
+it is eaten boder disi in facini
+he is inspired by music classic an music
+the relations between the scots and truglo were president of the dominican republic
+the catadorf and their environment form a natural reserve
+he adds two kennons to and a modified battery
+he is rewarded by yal for his project
+the hydronums of rent called the same calculation as him
+immediately you don't immediately you don't want to attract attention too much
+according to the internal lucitive sector it would now be signed by this kitar sector
+he led the hospitals of nidium and durbo
+at last the messages passed by condat
+the dopercy the doperced at the jeffin as a civilian ridger
+they were in command of general pino
+i knew in preparing my article that cassander would be a headache
+according to the chinese beliefs the hair has a very long life xperience
+at his honour the village around the verry was named after bredville
+there are also almost atranian and prova train graphics
+beckleyan sees paris
+in the middle and the rays were deafening the amboo of a shear
+it is known for its fancery
+the efficiency of these mutations caused unrevelability is not the sea
+the school is reputed to be the first university institution of cameron
+it requires overhanding a higher temperature to develop
+this plan is generally served with rice
+ssti called this doctrine the integration
+originally from guinea by his mother is oreal in his father
+according to denis de lacarmis the city goes alongside porsina when he attacks rome
+the tenors are always a source of inspiring
+the way of the lake williswan goes through the north of lake widswan
+it is used in laboratory as a lendingr for isolated prodigies
+she took refuge in leon then in frizone
+this species is named in honor of ry moody
+he has the time to shoot an arrow that reaches the lady
+his ben eflic left for a while to make him
+you have sabbathy i'm driving it he accelerated
+agnes benser who is the wife of the french physicist david currey
+she is accompanied by northor berlin somerset and clifford
+some swords of the comedy gave by his family are exhibited there
+the municipality is located in the south of saint brioc the prefecture of the department
+these two companies form a very powerful manufactured
+a plaloon of mortar is also in charge of support
+he is currtly the main head of the association
+its headquarters are located in hera in the suburb of torrin ontario
+the rooting stables must form the pullman benefit from this brand
+he particularly defends frank at conor with the corp wine of cork in view
+it is the minimal order of the cateretus by the fact of the c
+the emperor sat on his royal seat was carried by alite soldiers
+she was the daughter of gion five bertrand count of provence and adelad de cove she was the daughter of gion five bertrad count of provence and adelad de
+he then joined the hospital of bliga in the context of his training in surgery
+type a year
+it is one of the oldest texts of the biographical jane known for the ancient empire
+nembezs bordering six other municipalities
+bretons wer parisians complete thefective
+the raon is located in the centre of the kemchat peninsula
+henri emile rogeroles was born in a modest family of artists
+the leaves are rich in proteins videmetc and
+the cover of the tiloherab is directed by mat breden
+quickly the group was made a name for its remarkable and explosive performances
+he lives in the first floor and his wife is the ground floor
+the abbey is covered by the ice from december to max
+he also determines road security actions at the national department
+also more trifier than ateletipism
+his appearance of young girls allows him to play in tenagers movies
+the swiss canton states of the confederation switzerland
+exceptionally the teams are made up of famous figures of rocco
+the ve place on the left is still standing
+the series is published in the form of the series is published in the form of dictacl seven interesting everyone to a generation
+to day the musgrave is the leader l b of the grose trade
+later the group focuses on the scene rather than recording
+the two fiftems will then be separated
+here again he has the opportunity to sell in excellent conditions
+the troderenberg is the most difficult sector of the race
+john marie perrier is the one of the most important company between jacqueline porel and enri savard
+comentel performs at the vath bath retl bormity and fendmity
+we can see the upperates in the quanta mile
+the attack was made up by the oniver and the ivory
+there is a number of texts given to drifler known in tenth
+for a part to reach the summit you must stop and adopt drads
+another one from the end of the ranking lasmeronaska finished largely behind matu
+once he was relaying the sajo street and the rue de larblette
+at the same time he was interested in archaeology
+the game is available only in taxes
+the soldiers came on the site were rewarded for the high school
+unfortunately the series was cancelled before
+with his young brother mustafa varzelli he fought the araki government
+this number could be specified and completed by
+the bridge of sa rablin is enhardy at the montoris population
+he was a teacher at the granatorium college
+is there is totally invisible by the tourism in the mass
+he runs in an artillery election organized by allan new
+he resumed his job as a journalist
+the platforms were not interpreted
+this book contains examples of the first variations of the history of music
+this distinction is found in bogolo where the power is practised in nobles
+the lies are long and they are waked by the head hence the name
+the lips are bycoprtins associated with the sugar part and a six animals
+after the war two back to scotland she teaches at the university of glasgow
+he proposes a significant exchange for establishment on all the privileged tax payers
+this version has never been officially published by athala records
+this libel sturdy has a gray coloration with black patterns
+we make a visit to lander who came back to the penish
+the vallu was once appreciated by the incas for its geographical features
+the numbers between parent the numbers between parentheses indicate the number of new records present in the building
+the victim of new health issues he has to support a god in heavy loading
+only since this morning she has been racking me because she is a hot weather
+the cows of clubs are found in the car
+trees were planted
+she initially affected the army of a central district
+the kenton corresponded to a part of the vale majun or the vallet
+here is anon here is a non crussive list of these people who grouped it by the activity
+there is also a remix of the capiter version of the radio disnesan
+one of the best females is male
+by boat fintrina is located west of m pasitre and east of
+the season was won by lidevic bourgeois
+he didn't leave the capital any more
+ju josmar abrazil in japan in south arabia and tyland
+she lived in saint metern
+he accepted to ritruct he accepted to retruck the post office
+she then teaches in a swedish school
+she attributed five seats to the assembly
+the only moment in these rules is the long lasting disease
+his trenches are opposite and deeper
+it is considered a spectacle
+on the other hand the plane was completely destroyed
+saben foot batim provides large world car manufactures
+rencary is the only columbian to have participated in the hals
+australia kindling on the sprint of the platine
+coville de lerbin de carcosan
+maria cassin was born in a family of braves
+it is associated with the banton's festival of the pivot de colentria
+several works and places of the maurice island have therefore been dedicated to engley cupin
+he keeps the particularity of having his own bakery and bakery
+all the individuals of this species present a large conic beak
+this victory was not attributed to other runners
+we are holding through the elbows we laughter
+the group will call for the donations to finance via kick starter their two albums
+once we chase we will wait as an old stet
+she interrupts the school at six or twenty eight of the neighbors
+it is the only species of the adathomis genus
+the two weapons were pulling out from the projectiles to penetrate the ben
+he used a ritter in a butterfly
+it is the second region of brazil the most populated
+it is designated best player in the final
+fortunately the wind of the man had to fight for a day
+sticono also participates in de busor collecte
+due to many air tacks the ship was not able to be repaired
+there are alps of pergo
+there is also petrice de bois who has doved two times
+the daily law has two seasons with nde patroner
+he created the company of marseilles
+he is also interested in vertebrates especially in crustaceans
+the means were undoubtedly related to the idea of being able to achieve it
+however his diploma did not give him a six months to play
+it was ordered by the association for the portine district
+thereafter he became a psychologist
+the realization of great domain is a result in the work of barb
+it is a voluntary product and unclaimable
+this custom is no longer current for several dozens of years
+it serves in the actonian relay in the entegrants
+so many days it so many days it's the last one i'll be revenged
+it is located at an atitude on the island of predicy
+he plays nine races and score points and eight points
+the park is almost completely located on the territory of north hended by the division
+he currently plays with the conock shill
+the palace has experienced several construction phases
+his head is pretty little with a black shading black
+he was named after the city of balda an ugrain
+the departure is given by mullin
+the count and the countess narbone pile died getined
+note with saka amoto as a central member
+it is a plant that can be seen invasive
+bubuget then prepares an album with more commercial taste than these resentments
+he had married a certain catherine of argentina who was also buried in the last abbey
+at the age of five he moves with his family for lay to texas
+the abbot of saint justin was the temporary and temporary lords of the city
+it is with that last civilian parish of billington and nengo
+he constitutes one of the only eastern churches of the city
+the situation became tense each camp has its supporters
+the lenes are reserved for the cruisers for the gufes
+after he passed the tests dominice demart with a new function
+this confluence is located in the northeast of the village of liovert
+the cafe is located in milandan
+sam is a key player of the catory of space is in itself
+it is located on ohan in hubai
+the sales are smashing in france then in the united kingdom and in the united states
+he later acquired the seal of grand riches
+the game is available curtly in two versions
+it is located in front of the hidtle bar in the stone's monument
+he finishes the show
+there is some realities and another in the san
+it was a major port for trade with india
+we can notice that oliver has its moustache for the occasion
+a double penalty or a double pink dok is necessary for the medal core
+the accumulated total is depressed and rises to francs
+there are various federations in national and international federations
+eight men are expected either as much as episodes for the sheet
+he then travels from the world o fubo
+his mother died when he was still a virian child
+the architectural style is based on that of the theadre of saint visi
+we find marshal cantero silipeste hugd morel and royin pemp
+the headquarters of the municipality is located in the city of bretestorp
+the forest occupies a plateau in adeled
+filled with the fort filled with the formoin and his taken over by jamaica troops
+this is how he progressed and developed during the century
+she has your preferential and insures its election to diet
+the windows allow the soul to see infinite light
+its administrative centre is the city of turiadere
+on the lest the ven didn't on the lest the ven didn't work from march to the end
+his drawing becomes more personal
+for his first professional season he compared with a games all competitions
+it is a countof wise strength
+she was located at the beginning of revenburger street
+george succeeded his brother at the head of the body and becomes marshal
+it is unprecedented and very connected on the internet on social networks
+before entering the federal council was mayor of the town of honor of dormo
+the restaurants of the two leaders carismatic leaders of the group
+only three copies of the map are also known in japan
+he is assigned to the army corps but without being indevised
+that's what i do in fifty years now
+president of many companies are also being elected throughout the year
+but he didn't but he didn't care about the voice the contract to abandon the oral teaching
+born in florida thomas grew up in the cleveland region
+he will often play and become a regular coach of the coach fully iron
+the reminders of clinking migrate the watersheds or ritrevus
+both of them got married and the princess hashkuboi thirteen you had killed
+he leads many letters from saint in poems
+boutmussy learns music with claridet at the age of twelve
+very quickly the persian acquires a great education
+the females are darker than the males
+the soil is made up of arable lands which are under the ground floor in sand
+the french residence is one of the most beautiful in apa
+the largest bottle in the world
+founded by jogy founded by jogly angry push presents the shape of a mandala
+we say that the name of his death we did not sleep with twelve years
+its capital is fernally
+the matches with the contract within a computer or in the game
+bring in the book of reasons of the family bring in the book of reasons of the family's book is living in settlement
+alice meville vacan received a seat in dimacoa
+another policeman also reveals how much it is a germanic to the company's business
+he received many prizes and rewards at the port of his career
+she knows that the philosophy professor has not accept the rule
+he says that his last test match with the new zealand team
+this period marks the real transition between age and killing
+he was a professor and dean at the academy of zagrip
+and the revel finally went to rori the true reason of their separations
+and in this village there is nicholas
+she suffers from the absence of her companion a soldier in the osniac army
+the skin is played at aribia capital of kenia
+the aglameration of the temoir south and sans community is located in the southwest of the department of s o
+many parks are also found in protected areas
+but no agent qualifies the investigation
+oh yes yes we have to put them back
+lin eugene he became junior world champion
+he is located in the southeast of the island a few kilometers from jabraza
+the chilin selection is therefore qualified for this competition
+she was crowned by frederick akinstene an archestracist of massa surtha
+the movement then takes place in completely heavy
+men my dear have appeared generally brcler
+she owes her name to the presons of a school on her site
+the game is now closed to the public
+meanwhile halusia talks and should get married without any way
+the problem of foundations that we gave up the eradication of the bell tower
+it is the local table that gives the cell where the last insertion is finished
+the script is written by andrew baldwin
+they are enimated by tele prescence and has a barshaltonomy
+he starts to publish his work on intnit
+rummond has a bascapa club the bait rummond has a bascapa club the baits basdit
+once the parish of purpeses was larger
+the con de con de bas the con de basing covers approximately from tyra
+from this union nine children are born in the same year
+his death has engelium on two plans
+the new teamer lewis to bourbon has vermandi
+france is cut into a defense area
+we estimated a thousand of state the area of we estimated a thousand of state the area of cosystems of assail
+both of them are held the position of captain of the team
+the matinel river mips is the main water plan of asia
+she sprays the lens apartment
+in addition there is no rain
+also the zerk stories were translated into hungarian and poetry
+the history of octavian is that of a difficult vocation to assume
+more generally gray four peseveral theatres and museums
+twelve on the other hand were modified
+one of the tributaries of the sagon river is the talpina river
+he didn't he didn't forget his mission for much
+the beasts threatening was also lord
+he is a passionate about poles and gif
+during the intubation the patient is a total nons
+retrospectively he was sometimes criticised for his light approach to his people
+it is described by its contemporaries as a pious and a great beauty
+the tea and coffee are melted
+he then took the decision to go to the university of volumbia
+one of the rare parliamentaries to oppose this decision is chukomeric
+bookser created by the old malabarian church will be burned
+he is the good of this abbey undl
+this one welcomed him thence to his scientific reputation
+it was introduced under vista
+the territory of this former commune was divided between chevili and sugi
+the stations molier with saint sir university will be re established
+the church is located in the municipality of rose saint alban in the an department
+the whole sets the autonymous district
+the atlantipites marging in fresh water generally rich in organic matters
+tierri le loron limits a lot of them which she has her own troubles at the beginning
+in the middle ages lur was considered as a magical pland
+it is found in the islands of brete
+the difficult relief will favor the mirnage of a few animals that sell this profession
+hungary ended third in the third place
+in quebec eveline jelina doubled it three times
+they will be able to put the new phenomenon a true popular historical culture in france
+the minister has general von krapper surrendered from saint petersburg in moscow
+oligermensu is a russian coach of the old school
+it is found in nageria guinea and ecuador
+the teenager falls and scumbed his wounds
+son of a italian immigrant and tradition baltic he became a journalist
+the results of the group is a major general
+he also works for tilvision with several popular series
+he dies of anymia in new york
+the group then appears in the festivals as sombis or partisan
+nicolinci is officially classified as one of the villages of serbia
+manly a player of doubles he was finest in dobl mixes in wimbleden
+the elder ones will die by a poptosis and their debirth will be eliminated by macrophagus
+some leaks between moscow and say petersburg are the same
+my stake is the fidel attito rio
+this model presents the end before in the form of w with five horizontal bars
+very short he reproduces during the villages festivals or at the pilgrimage
+there is also that these years in the army had learned of the authority
+she symbolizes being hunting and fishing
+the oppressents new type of destroyers
+it is located on the section under alboslindu
+the count is also chancellor of the protestant order of saint jean
+mumba is the architect of his associations who had similar status
+example reno motor juvis four
+it is the former wife of cater kirk and jiff
+he also plays for cricket
+henderson was also the director of important theatre in great britain
+square demle in meo
+the article is written by pup
+and why the art critics contribute to tucking standards
+his territory was divided between the commons of handelacor and pemard
+this species is endemic to central southeast asia
+it is by the british presst phone that the data arrives at bnny
+he organizes each year travels to thousands of scouts
+this fragment could be either the most important or the trums
+pierre lemote was born in paris in a family of orgins of music
+it is a project that is his friend friedo westburg it is a project that is his friend friedo westburg developed together
+she studies medallism of medicines on the same name
+he is to be noted that the main actress is the same as william sindron
+he sends his condolences to his successor to another anglo saxon
+the project can be exported to the windows foremat at the point of view or a two flashes
+however marlittle does not know much while leaving the first track of swalim
+others are equipped with the water jet in order not to use the toilet paper
+a normal takeover two strugens can cause a sexual version
+in return he gets a reduction of cotsations
+the portuguese did not leave very little marks in the senegales the portuguese did not leave very little marks in the senegales national team
+he was the assembly pulled out under the empire of a deep emotion
+this is also contained by a section in which it is a great interest in its part
+this bekram church now lives in a cultural centre
+the fall of the city is difficult to explain
+this tradition last until to day
+it meets in the province of the coast
+the church is located in the french department of main edlir on the palati of damera
+the question of the architect is author of these buildings is not resolved
+bshi is the famous enancines of sex
+it was only a big storm
+mit alfort was part of the seal's esteria
+he is passionate about sciences literature and music
+a bulgarian took the offensive and rushed into the terrible screams
+many retrievs of variable qualities were made during the last decades
+completely in general these submarines are toring their dezel engines per day
+its capital is binte
+only some facts decide for another destiny for him
+arind gebelli comes from an old local family
+the firearms were the firearms were drawn with the critical adjective agents
+the present is the time of sharing by reucelebrating
+the sales of good originals are offering a combination of known pieces
+but the night came he seemed to be the violence of the hurricane resolved
+the crossades looted and started to take place in the city without any pity
+it is composed of a hexagonal tail that supports a dome cove
+it will be trained to become a killer at the government's room
+they are even divided into sipalities
+one hundred gram six sharpenlich survil lesmers
+the fear is that no in the semi finals against england
+the divorce shortly after
+a restaurant opend from early october is located at the color
+this provision is common among the frontal guards and supports among others
+the guests barely arrive strange phenomeno occur
+some electronics are said natural
+it allows for the developer of experiments with multiple operations
+they are made employing different materials
+the surveillance of the tybrite is a clinical and economic examination
+it is located approximately four kilometers east of lake annvenan
+his daughter caroline married the m p a d m
+and one otiv was rased at the college de pandevis
+the double album collects various pieces until then none avoided and rare
+who also insures the presentation of the second part of this tour
+it marks the beginning of the manerel of tokio
+it was about resin people living in bidervilles in the cities
+the trainings are organized by the conomy faculty
+badr is bordered with the afgon province of kuna
+the movie was distributed by universal studios
+this castle is one of the oldest in gam
+the actor finds that his look is too lively
+he was even known as a major doctor of love him
+for the embarking he only takes two parts separated
+where is my father my father my father
+the arrival is still located in menico
+sir of annes asen is a also an actress klotil hesen and the youngest
+he doubled the gold disk
+later he has been learning frindes cobert
+it seems to have led several lives both brilliant
+shortly afterwards he will open a workshop in the street of nor
+these installations will be accessible throughout the year
+jacque cheva is not limited to the french maritime heritage
+the statue could have been found
+lismaner joins him and the guest statue of guests without salary
+the main actors are jimmy mystry cristo and niv kemple
+he was equipped with a classic lending gear with light legs
+he had started learning the typography
+it forms the substrat of the indo urian religions
+she meets at altitude in the regions of limor and vanuco
+he died in ow of the same year
+the latter will be available and closely close to the next one
+he left the upper part of the city of the victoria its gray bluish white
+the budget for this project is a reasonable gain around by pixon
+you you going to start the district of accident and come to the south
+the french is a mar comolial diurnal during the brantem
+he will be pushed back by the defenders of alexandri getro and retreat to the north of hate
+he was married to walt razer
+he plays his first season in the year after
+this law aims at depressing the impudicity
+gasse took a disproportioned life and worked in many union towns
+he then began his studies that he finished in san francisco art institute
+in spite of its cultural property the race is rare
+the hungarian instrument here is the revebl ancestor of the indian souf
+later the abbot received a decoration from prussia
+the abbey of moor the abbey of moran had goods
+these scales are bounded with lighter
+it is also a tite of the turn marshal
+lipinimo was optimized as bombardiers
+john pierre is strongly disturbed by the two aspects opposed to the same woman
+this species is named in honor of siminaba in cury
+the preparatory dryings are kept at the louvre museum
+this one has then accused of wanting to confiscate the seps to his profile
+his building is located on the post of goveslvi flag
+these symbols disappeared on the later editions
+we are called them treggles are sometimes reddish
+the hubbin is turning his fixation and breaks the propeller
+the co operative cave of buferic ends ectalus and ind
+the estate produced red wines and white and mosqueto
+in the capital the first tarnet is the american embassy
+however the product of two converging series is not always converging
+at that time this flower symbolizes the commitment and conjugal fidelity
+we say that every substantive is built by affixation
+the municipality is located in the shenden peninsula on the swe
+the abbey was first called save marie de cestre
+the launand visits the city of news mernabish
+he is then the hotel the highest in the world
+in fact according to your view orion is a camp
+he is notably mentioned in the white anse
+he then had to complu or galleries in citrat
+where the became one of the terminal stations in italy is most common
+it is also a player with a technical quarity
+the rebellion was but jubert succeeded to escape from the battle of the decision
+she cooked she cut on with courtiers to form the new commune of cheteliers
+the kemer is at every time moving to fell the chimelian
+this species is endemic to gomera in the canary islands
+i have never given up the right to suggest evening
+a player don martino is called a non stender
+we will give you a track
+during the middle ages the place is inhabited by slavic treboslavs
+the style remains very similar to the style
+she then took the name of sat faro
+furthermore the king's in torat is as much as jacobins
+the birth of a heir to the throne had a particular importance
+the first two of each pool qualify for the round of sixteen
+races take place in order to exhibit the last by cycle models
+the results are determined by a professional jury as well as the vote
+all the associations are evid to receive nations
+quatamela casta rica and panama were on their own
+the following objective is plizzing towards a more general failure
+the locality of trepona on the fechenga bears his name
+he is sometimes named gard glove
+he was a member of this trip he was observations on magnetism
+however adrian de witt gave him an sure taste of rhyme
+at the time of his completion the feste was the largest man in europe
+then began travel years
+the news spreads quickly in the village
+the trucks are made of aluminum cast iron
+the file is then transferred to a moon in
+there are three distinct parts during competition
+in any case there were no children since the lordship is falling to his brother
+yet all of them indicated that there is a segregation religious segregation
+one of the suns called rima is the third of the cube bedages
+this election marks the progressive return to the constitutional equality of this country
+he owned a sat e a popularity known as hermitage
+he will come back from this stay at the cell by gaspro eton
+poland is known this decision will never be applied
+she captures the village of macada which she destroyed and destroyed
+she's propagated to a vegetative way thence to a round man
+until then only one exposed projectile is not taken out by mortals
+this tone is revolved but the local economy is orbanization of to day at the quarter of the approach
+two days later he was sent to the official post of this edition
+santonio benefits from a sub tropical humid and typical climate of the states
+most of the investments are in the health sector
+in simitic the malicharn refers to other motifs
+she replaced the union she replaced the unions mission in the south
+these writings are not personally related to us
+the press is largely cotted by his heart nation with robert huson
+more than a hundred years screamed out of the games
+the geneter also meant to be significant
+the series is centered on these relationships often chaotical with ices
+she mainly crosses the city of boldluca in the serbian republic of moscaw
+it is now advised by its definition is engurged
+he is incorporated into the school of the people
+the supreme court of canada is pronounced in favor of the decament
+the horsemen are not generally seen from armouth and white
+each year the biggest ones to a server or a former trainer
+the battalion served in a roll of territorial defence in canada
+jasper hobson is his companion did not have a mistake
+after the french revolution he was sold to individuals
+several of his books were translated into several rinanasian languages
+the species considered indicerable also greatly recreased in rural areas
+nasea possible during the first time it can lead your honestment and dyar in some cases
+this species is endemic to southeast africa
+the competization is generalized even though the new techologies threaten to return to the solemn
+she's located in hase crowdley on the river
+lusan mullo was a great resistance fighter for colonization in hots kasmogi
+the teasetons are generally less stable than september
+the training in alternation is also a possibility
+the configuration of the cabin is o depend on the choice of the company
+he studies the role that plays violence between colonization an columbs
+he's inspired by similar short livations in other states as well as other regions
+norway participated in the chingin space in europe
+she participated in several salons in brussels in havra
+it's lightly big like melofre the delicious it's made up of glass
+in a personal way the ability of a mondel the classification is corresponding to its complicity
+the ivits of the secrets of the baptism the thirteenth of the black ones
+she seems to be managed by the austrian foundation of the bramsete
+inlocation symbol the gland is associated with the death of gude
+the chamber is bound by eleven orth stets
+he teaches literature at the vellefille college in saint amisi and monte
+of course paris at the end of the adventure was a lot of talk and the facts of becoming the people who got worse
+his square and germy rebels constitute the central defense
+the fruit is akins with several pointy scales to its summit
+he chose as a quecher his brother philip hubert
+the flanks are marked with two lateral lines
+biddle accepts the policeman arrested swedboh
+brigande of infantry is mechanized
+the quest is totally different
+he retired from is haki at the end of the season
+he continues his theology studies in private school while teaching ematics
+his life was documented by the book of his wife risesi
+it was dragged for about four years before reaching the group
+when the war of dependence is broke out he joins the english
+the life of pedan had a road
+the primary school was named after the city of yen arthur
+he gathers russian and foreign doctors in enda chronology at the bottom
+he was a japanese international twice for a goal registered
+the heritage of h romanticism is allso complex that the origins of the movement
+the captain shows a cup of silet still hot
+it can have several challenges at any time according to the population of the county
+the neighbouring hamlet whith marguerita has been completely devastated
+there is then named after the war of moins congo
+if the fishes are not aggressive they are getting place alot in community bars
+it is also a colored star of mercury
+three months later nnixenville did not see me at her place
+he was co ordinator at the poetry in writing
+it's through a lost child it's through a lost child's poster that this character is introduced
+he's associated with the legendary
+the dominant winds are found in the west
+this bird is found in gentina brazil and haraguay
+he is replaced by andre de bauvar as general director
+he says he will follow his heart and that's what she does
+he was produced by risfubert and mixed by gregrelly
+the arstrians have known difficult moments during the safe period
+it is the bowl of the history of the team
+exhibitions are temporary and dedicated to the public exhibitions are temporary and dedicated to the public and artists
+the album is pressed in limited addition and quickly cut off
+its head office was located in sanneville california
+he's also more difficult to establish a cost already through
+a surgeon must then be discussed
+he gave his name to the french bull fight street in paris
+he had done the poor travellers of the colony
+many bridges cross this river
+it is a walk way with a three third line
+she's with two planes one of the only deadly survived the lines
+at the evening of his life paul cullen madal dis coches of anitram
+this troop then attacks many in a tonado movement
+a chain can be closed or see its tips open a chain can be closed or see its tips sopen neus
+there are several houses between the settlers and the british owners
+in architecture a frontispic refers to an architectural element
+the royal street had become protectorate of the polish crown
+he was little by about five million evidence
+these tests are the most severe and ever designed for a regulatory weapon
+he was the main defender of revolt
+like cat like catgorized internative rocks or metallic rocks and postcrims
+the newsebm has exported its meat mostly in great britain
+there was no number in the street
+the city of ha the city of hasensad is a petrolary town in the centre of aldria
+a social forefront is therefore opposed to the principle of equalization
+hundreds of vehicles are stored in reserve
+it is considered as an invasive plant in the united states and in some pacific countries
+fukuda is then sent to the chinese frank
+lydden graham is based on the textchal data base
+this gave birth alean on her husband and her son
+i then becomes the main suspect
+at that same time he also attends the teachings of the loof school
+it is found against the wall and mederza it is found against the wall and mederza maglius
+a ground floor was added to him in the dark aces that can see to day
+this white life has a real structuring of the final composition
+the roof is made of slate at a long time
+the pail can also be used to experience air
+the lake cressy turns out an increase in the etawe river
+the name of this county comes from the former territorial governor of nebraska
+it's the juvidus its the juvidus acclamation of the dental inflammation of the dental
+she can receive slight and various successes
+some of this work are translated into french
+his missions are set by the heritage code
+he fights in the voca sector but he has no grob
+compared to the cardinal jean lucdisces
+he is known to day under the name of comsediuse
+as soon as it is the use of the continuous movement
+she has bernardy and her father take care of it
+the report was not followed by the facts
+he holds the language towards the front and towards the bottom
+vene is particularly marked by scandinavian architecture notably in nense
+yup of south and south america
+the flower of the imposture is a symbol of the arizona
+the national and national office has been functioned as police
+see the list of the pummies of the court of the canadian supreme court
+she notably serves sincenita
+it is located in the southern region between bondili and newfound
+seeing my uncle i cried out
+finally he is ordained bishop he would have been the first bishop
+baku very hot it always stands
+all titles are by jello biafra unless stated
+this microbia forms the largest microbiot tan
+he particularly plays the role of alexander cristo in the catucha team
+he served in the caribbean during the american war of independence
+this last list is obtained by the majority of votes and elections
+he will be named after the french attorney after the trial
+alfred dressed very close to the original
+this cabinet specializes in cos active ricos is an environment
+the group under the southwest is won by the gerundans of bordeaux
+they tried to destroy it in bi
+hide is also the sister of lisle also in ba
+he is the son of the duke edward and louise de sex
+it is a sub tributary of the goron by the pantagnic stream
+he said that a german princess would not be more serious than a princess
+the film is the work of edwin lutius and looks like weitford
+he then worked at the academy ofards
+he stayed two months before being replaced by hal shaken
+it is the smallest news of the sada
+he occupies until his retirement the chair of differential and integral limestone
+pierre de valois then moved to paris
+this remains a mystery
+his action is temporarily slowed by the accumulation of exposure
+he is invested the next day after resigning his predecessor
+the main offices are in middleberg and turnezan
+the oditole desert the oditole deesn't know how to be called for all
+the prince of arlesnoi is a village in the saint martin district
+she learned the ability to do instrumental manners
+it is the letter dia critic o dima and a carm
+it connects the tangin botshe station in raki cham in salte scade
+the different species of oran live to day and ex exclusively in indanita
+he represent austria during on petitions
+he couldn't he couldn't he couldn't tolerate in front of him his enagments of the emperor's praise
+the yunta the yunta shuttes when they're every summer
+a large part of these ships are old and not operation
+he is also influenced by aber and joseph bernard
+it is considered as one of the most important poles of this agricultural sector
+following the instruments of music are different
+the election of the quebec party led to th opponents to the east highway
+attacking is repeated by being able to play for any other club in france
+first it became the church of the jesuits
+since then thy swall has become a pilgrimage
+the city was then the two parishes of saint andre and saint coter
+the group one the group one the group one the group one i ali is from the city
+the hyberdation between the wolf and the dok gives a warm look
+when he died his son gerard also a painter undas
+his steam favours and that of a dark sea
+it is fairly common in the midday and in the west of france
+the dali lama was the first one to the lamba to exert intemporary power
+axeldema is a master's degree in law and a degree in philosophy
+bjotenis pa the telos academy
+the buildings of the monastery and its materia are sold as national assets
+i have the feeling that the wall is absorbed
+paris blenchard publishing house in paris
+his black body is exhibited at the entrance of the orsea museum
+this is the most important story of the women's history
+she became a medestri of espizin when he was ill at home with the nominee
+itis therefore a temperate climate hot without dry seasons
+arthur jetton was born in the tame
+he seems to me that he would be delivered the last appearances in the resistance
+it was obtained in the following month with cosler
+his homeni son was also a painter of simon martini
+a fill maker prepares its next movie he tries the actress who will play the role
+the character of miss henningan is absolutely out of order
+the second single of the album wilb
+the leaf is rectangularly visible on the tside
+he plays the works of jetto sussy and willam camer
+after his remeration in sydney he became a member of a group called jengo
+the municipality is full of a noble house of centuries
+those who declared in in serva in the new army were demobilized
+this clip was also produced by the zimica pex team
+finally the navy became sorbon as one of the most original beaches in the island
+he qualified his masters in my science
+he studied comedy at the heart in the cashman school in tramp
+the columns of the facade are made of greenwood x
+this last manoeuvre was experiencing the opposition that was organized for several months
+but but i'm crying didn't but i'm crying didn't accept the repression
+the castrolmat dad back to the middle age
+louis bellis was born in montreal in lundoc
+the soil is essentially limestone or cyli
+when he returned he settled in cossack of the sa honory coq in the hotel durtois
+in the album the device is present in near the third of the planks
+a municipality of the pyrenees located in the massiof plantero
+in agriculture a good health analysis is necessary to correct events
+a chapel is located on the island of saint cado
+it is based in proto in the province of proto in tescana
+then she started looking for a way to wars for airplanes
+the paper is then immediately rinped in water then dried
+the revolution interrupted his career and was imprisoned for the war
+he is replaced one month later by gedri sumpson
+according to wecker there would be no battle between charles battalions and british
+he is promoted captain of the selection after the dese bank
+she manages to eliminate freddie by turning her back
+this region has gone from caesar crater and salpitar
+these two lines are located under the exchange of george herrory near gesson
+she structures the development between the urion cities
+it is financial and a holiday as well as it does not renew
+the truck stretches in ietechnology is respecting the environment
+investments have included the large operations and development thats are natural
+it was spoken in the kingdom of versailles
+despite being practised of the iron age of body despite being practised of the iron age of body can consider it
+it is the main community of the network of the network and the main ones of the main building
+during the catastrophe this murderr has disappeared and his wounded
+there are ten me tribulations
+they risked the three years of prison and the ura of the world
+this sun has been the maus in sound serve as a pregeneric for the film
+the jeffre had disappeared jago is accepted to the palace
+one path connects the san clair door
+he dies of the defence and the discipline of his troops
+it blowms in south africa to september to october
+the capacity of musset is broad to the white proteins
+the main difference is out the number of applied players
+she's the language of a she's the language of anrgates of peset cumbo and fornish
+tribunal agencies can be independent or franchised
+the construction project of a line going from canata in the right bank
+formerly formerly rhin villers rhin willer or rinwiller
+and at the bottom of the reeds a old pe
+egel minor is really the function of the prince
+the king made the deaf of an ear
+the two regular straps are down on the right is built from a boss
+a flotter made up the floor from darendos
+i i there was his younger brother
+as we arrange the buildings like that now
+what you want is to advince
+cheers bunicon i even know what was he doing
+he didn't he didn't say anything
+yes the poor man cried for it
+he was there at his home
+let's make us eat some buffan
+they have only one thought to which one's engage the day insane
+no but you were wearing it we would like to go to the right time
+here is the currency that is certaily prepared to prepare a poor man
+the architecture of the lighthouse tower was amazing and very good at the end of the building
+that's all i can do for you to see what i am telling you
+you were run said a low voice
+up to this moment she had no thought about her face
+he was with you
+the male fetters capable are on the printing this upsile in the crime
+oh how he was hell
+he kisses him again
+always at the same goal
+the goren supplied her bread to the bones
+the doctor had arrived too late the priest had arrived too late
+he ran as a star
+during the time and wie
+we had to turn off anything or nothing
+you were strong and well important
+we could hear a warm coating than the cookies
+that day the man was a good man
+petersburg has his own
+she was paris she was peressed in the meantime that was a father
+duk for as to da can
+but this troll is out of our range
+this simulility is not a very rare thing
+moreover he has still time to go
+whatever thi storm said whatever thi storm sad ponsibilities
+but ireton answered while visiting the engineer to risk your life
+black as a swiss man he is the first time she said so
+the church was deserted again and fell back into his silence
+a large roof proprietor was hiding half of the skirt
+but here they are going to pursue you
+this one was old his hair style on his face was unrippable
+all that you have no idea how to suffocate my comrades
+he wanted horse and went back to the grin trot
+in his adoration she forgets even the commission that she was in charge of
+you were safe jen belton jin mathew
+nether me said the sailer
+that's how the english people are bad at the bottom
+you di you didn't recognize me he said
+evea courtier is an intendal
+i think so that it is like this one that we can say it
+but how can you see that
+she had a su she had a su she had a sete she had a sitan for that her pity rejoiced her hands
+listen before you there was no disty creature of this name
+hell was at the longest leaf of the greatest blessin
+others will stumble with a bargain
+he is often referred to invitations
+the bays of bigor the bays of bigor speak one of the three main dialects of the bay
+it strongly influences local activities and allows for example the creation of jobs
+maria camerini is first a member of augusta argenina his cousin
+this country participates in the winter olympic games for the second time
+never had he said if i had there if i tried to do so
+he covered her from kisses and camens
+the dream itself would fall back in front of such a climb
+you understand i demorate your crew
+yshidas won the battle that ended up with dnsdeth
+the movement was neither to be involved in his murder
+it is subordinate the superiors to see in him a strategist
+according to the two emperors he support him
+belle del belle de latoire andre la rouche francois champ prunes close to gringory
+he forgives him
+the municipality includes the districts of wesenburg and west hills
+he plays as a goal keeper at hapol hrani carmara
+the couple had two sons fasto and john beath
+with the dependence of slavania a new era starts for the club
+nipilla has been impass well to make a joker
+young women are forced to flee
+as rich as a ridge in the king
+he is ordered by the captain sera paris
+in summer hand backand i was raised
+the works of michel tork are sometimes described as a minimalist post
+by spurring as in summer the days ahead dare frequent
+in france igement is regularly doubbed by laura
+however sixteen smith did not answer
+he served as a junior adviser and went to help them in their lives
+thus the name of a shaba was adopted
+he is believed by the qlatter caliche
+the surrounding fabrics are filtered
+gas parna must be working on the theme of the penetration al verginal
+his interest for naturalist painting will expend
+every year in june takes place in the village festival
+this relief pitcher plays with the blue jay of toronto
+he keeps his strong god but finds god intuitively
+the russian was also known as the name of frencher
+he was in charge of one of the smallest city of the club
+in addition go also took another beneficent of the childhood
+the belaniples pass through the surrounding ships
+they are disseminated in most of the parishes of the city
+he enrolls in a latin cross in three naves
+it was the time when a door up was made whet the work was defeated
+the young woman was probably aged around twenty years at the time of her death
+joseph in westell raidford was extended until this location
+he is a member of frent's summary
+he played in thefinal with e sate
+a career is well known by the story of the contemporary historian in anien
+he was born an eskali five thousand near osier and fought on his territory
+in the longas senatis denounces the remarriage of herodins with erod
+under the bed of the youth are under the name of the bones
+this extension is unique
+the government does not take care of it any more
+she even recalls that he had skied these questions from a red background
+taking a position of the most strategic position he was totally restored
+but the draught is limited and only lasted four months
+the couple remains without descendants
+sixteen grandes and noble women were looking for their good story
+the arrangements in original music are right torcha
+it is the oldest part of the castle
+this technique allows to visualize the profile of the community
+its use was previously kept in the public so folk associations
+the vectorial species have a scaler product can be equipped with a standardize
+the song becomes internationally accurate
+he belongs to the jaminia he belongs to the gaminia of ovia to orhai
+its headquarters are located in montreal
+he was one of the most important representatives of their abstract in italy
+john arthur clos plays a seven games with the whales
+its inhabitants are called the colres
+in the time when he was called jore chalk was the first name
+he became friends with roman cartier who also joined the parnes troop
+augustin climbed on his arms polidand sleeping under his armies
+died in the city he was buried in limba
+he also took management lessons with arthur nickes
+he is attached to the field of kuscova property of the shirmidi family
+he then made several pictures of the state
+twenty five frants cross over their lips
+its headquarters is established in gincourt in the eve lines
+the congo brosevill can be excluded from the king burly reses
+five hundred eleven three hundred sixty
+he let himself be a fool
+she is a good mexico and was covered by the osaca museums of the palace
+all the benefits are from the peru and o n e p e r
+thethe clear of the cognia profitation in siled is not in between these two swords and tempats
+the recipe is a medium sized river
+the main language of the first nation is laudible
+notre dame de lac and say just to lack are also the municipalities of river miscaat
+in the same way for the ethiopian pyramids and murderi
+his first piano teacher was anast fanska a student of alexander michaelchi
+at home his wife was also killed
+his father sents him to study at the borges jesuits
+the totillas are the most typical of the kitchens of texts in mexico
+since then he is no longer a professional musician
+he finishes his whole bunch of sun and shots were once associated with the same name
+this block is known as the most important ones of the time
+he was the oldest clown in the world enversise
+it makes the end of this work brew and confused
+he is at the same time a good level rudby player
+the expedition will be able to cross the territory of delegates
+the neighborhood is located near the barret craft international airport
+it takes place in the beginning of july to august
+from there he served bergin batenburg alfrest and apple teron
+from now on i don't want to be modest any more
+it is maybe one of the most accessible chapters for the actor
+it serves the vellier station and its beeches
+together they travelled across the novile bazaar region
+he was previously managed by the french federation of sports de glass
+the television is on a brand that is a public
+yes replied these companions
+he could then carry on some ten meters
+from turkishm he keeps the importance of the social system
+in our hand there is also advantages to occur later
+this species is found in armenia attered by john and in turte
+its headquarters is large as city and lowe
+there was no mutation on the region
+only her parents left the children in the darkness
+she starts playing at the beskiba before choosing rudby
+thus in the tropical regions some served as glot for the
+once the tramway left the path is down underground
+he is co founder of squashbls web sight at the start of the player
+fifty three real estate sculptures were placed in the nests
+la artabrana wins the lady single singles
+the railway heritage is represented by the old passenger building
+with anclaude le brec he is one of the main memorialist film makers of quebec
+the federal state encompasses the communities and the regions
+typically the method uses a transformation of melon or place
+they would encrut they would encrust statues that are now disappeared to day
+he is used for the congrete fairs in meetings and other events
+during the battles of the objects of ginitindo falls on the field
+this species is endemic to madina in the philippines
+this company is quickly abandoned due to a real record that threatens algiers
+catherine de medicis ups for a boy cousin medicis grand duke of tuscane
+if it is a lady the honour chooses the color
+lurrell and hardy are selling fishermen
+it is made by best tools
+she became very quickly the means of the main commotion of the peasants
+in this case the fruit production is reduced in salt salt salto
+he is named in honor of ellen nuvake astronomer germany
+he is named voluntary to serve against th england during the american war of independence
+among the sentences to visit the engineer john antelic
+it is advised to change all two years
+buhan i buhan i'll see you at the right time to tell
+his games will allow us to reveal the poet and writer his games will allow us to reveal the poet and writer gepers
+now he says ervill while going back to his cabrulet to the other
+its brown one the inspiration
+when cicle is the son of lewis and maria bodeca
+he has increased spots with age and different depending on inhabitants
+firstly james played a lot
+the emonition is a part of the same vertical way as well as men
+their efficiency is dependent on the type of troops and armor of the adverse unit
+it is the first daily newspaper published photographs
+new trip then of belgium and netherlands
+but the student was born in tiblac sibilia in russia
+the theme is recorded and sent to the ratum where the term is finished
+his population lives in the dervo
+a frank shepherd named him to be created in new triplanes
+he studied urian college at the university of marchester
+exhibition by the kingdom of boulicer
+it is the largest birded clowing with green tiles
+this work of languetelane will still take him for several years
+in danchisam is the french sally of cairo and leoce
+the construction costs two starling books
+he was there and not pursued enough
+it was in fact another game ground
+after tolose he ended his career as a coach at the big club
+he is inaugurated by the president of the republic jules crivy
+we also find the influence of the hene ethnic ann ethni group
+this doesn't suffice him why he is not enough
+urson l is affiliated with a syrian army
+it was a professor at the conservatory of leon aninice
+he started in the fell career jason alexander
+tv films from the life of george sand's life until his death
+he also wrote other detective novels under the pseudonym of haory comical
+during the closer of two years the inhabitants were collected in a barn
+le grand duchy of moscow is exquisite a little bit his rivals
+two wagons were found on the bridges trev
+the nima terol and frick reopen on the line
+their construction was sometimes difficult
+dalin was very interested in english actor ropper and true
+in these few months the explanations on seeing afard in half anxiety
+he also undertakes a correspondence with eliot cri he also undertakes a correspondence with elliot crillin with whom he is studying
+gorban is the place of storage of nuclear chains chosen by the german government
+the children started their studies at the karmes desimer in uxoy
+there are many old underground quarries in tuffow on the municipal territory
+the situation became heavy and forced giom to compromise
+he is the father of bruno benny
+he understands a satuary amendapa a portic
+to day resources are judged economically and more ventable
+in a p college four of them got formed in clan
+the colleges and leses are located in naan
+very old discipline the pole shows a long time intr national structure
+in african hunters the recruitment is exclusively french
+the global form of the floral receptact is flat
+the movie must have a lot of dialogues by richand sons
+the site of villemardin is occupied since the prehistory without interruption
+two extensions are developed by micropus
+his breeding disappears after his death
+races are held in the united states are grouped in the united states
+its foundation marks the beginning of a reorientation phase by the left of the catalogue
+his chamber was the only person who had liked him
+a lot of characters from japan and make visited the place
+that is where he met father joseph marie crive
+she keeps her functional missionaries
+the city was once again regional capital under the sesenates
+however he continues to sleep in a slaughter house
+we find glander a medical firm a nurse in a pharmacy
+this one was called touchy or less flattering as far as catson
+his younger brother is the actor keveeran
+one of the towers was to be columbian the other one in the bailewick
+it is located between the sun and the peak of lagulla
+the kenton of stirring bendle is kept in and expanded
+the nephesis is now held in a cocanlash on the roadside
+we were running into the arms of the lefitte and the cayard's wings
+although these friends are sentenced to their behaviour does not have any intention of change
+he founded one of the oldest workers circles the patronage of notre dame dechants in enges
+she's among other things lighter performing and sober than the old model
+she signed coriography of about ten boles
+the sketches are open in front of the public
+the sark figures is a limestone car
+the concept of personal supervi the concept of personal supervision has not been the subject of interpreting the cord
+each white spot lost its wife its work and his accommodation
+this one is forced to recognize the setainty of margrave de brenburg
+the river crosses the commune of bulda berneser and san michel gusto
+she also received four honors from the order of this government
+it marks the beginning of the cume line
+the main activity is at the time the operation of a micomine
+the place is named the music hall adam mar
+it also notices capital electric dell intervenes an intrigue without being modified
+the males have only one vocal sack
+she only had one success in her mother's
+he receives lands from his brother walter indices of exeter
+without the robustness of its construction she certainly did not resist
+the minne boarders the bor ser goltono clattiers esco sla miados and riato
+the basest rico leaves the group who has to play without basi the basest rico leaves the group who has to play without basis
+carael de france is on sixteen in the championship
+it is part of the sermon sir le montania in ingastel
+one of his teachers encouraged him to view emanuel in order to learn sophagus
+the anglo saxon markets show interesting books often ancient and learned
+all tuesdays he had to last rendo de gornisi in saint malo
+his horses are recognized for their palivolence
+he was closely with the arithmetic combinator and the numbers of members
+but apart from these small pantasies they are people like everyone
+the municipality is presided over by a sendant
+he lives in the cultural centre of hunkan and the city of the county
+gabriel and sevend would come to take advantage of the sad man
+he started a master's degree in industrial relations that will not finish it
+he had of diogadi bernado the di puccio capana and octaviano de fegcov
+in india the water is hot by the official language
+from the civil society it is his first political tern
+prake prake was a romanist officer under diclation
+he dismissed javert with a sign of hens
+the designations of development are indicated between parenthesis
+in october she suffers from menjed
+the ritual ceremony takes place at the same time after specific days
+dusk or train for a long time with the oblasts government who falters himself
+ade composed them to serve as interlude during the formances of his oratorials
+they are among the most famous in the country
+the castle is entirely built in marne and is made of a small castle
+michel ebian played an important role in the management of the villason agriculture
+osnial brazi was born in the city of ama
+at the same time a pianist saxophonist patarist composer and arranger he is entirely an audodidact
+the troops quickly stretched out
+the name of the commune in breton is without any doubt
+he lives in a latent love
+this one o the cases cush in a false plank while the house was firing
+it is the property of capusines
+finally abetedor de brabec is becoming a greater bishop and the prior of the chapter
+the south of the municipality hosts the marshes of sacon around catimorn
+the municipality depends on the basin of education and saffin education of the amin academy
+the monument is now a museum
+the tradition attributes him for various acompanies
+he stopped his career as a worker at bernard russoe former collavorator of the corbise
+he will be a painter
+he joined the collition in quebec and his father's creation
+by marriage she is princess of hanover duchess do brunswick and lunburg
+heastena training creates the double agress of alexander minor korf and andre cacei
+the young woman in question but a book to young children
+the production starts in an old electrical city
+he is himself elected congressman
+he then takes the name of hart he then takes the name of harman dot
+in addition this year he became the rector of the quantifcal college in rome
+it is a shame that these leaders have not been made
+plan ind rugatively street its eastern arm has a chapel
+it should be noted that reinforcements are not always conscious phenomenons
+enari udvenco speaks french and polish
+this day they celebrate this last one
+he then plays with caster de charbgan will castes of bon
+the angle of advances is underlined by a de corps in bossage
+he is the father of charles budden and the brother in law of jane henrihison ritz
+the evenings still remain rare
+there is also a lot of them
+the new bridge over the rhone is a huge bridge
+he went through a levee bridge
+then became a community with his recognized objectives
+the group is very influenced by the whippy american pacifism and acology
+administratively the island is part of the province of benco
+his father is a small planter his mother is a housewife
+the alga of the fundan used in this association are called micbrod
+he was a widow of marie n mutter and lived for five children
+he is stirring how the soldier shouldn't be used
+a small population would have been restored
+a springer could however be abandoned
+niko will be his biographer and will succeed him
+the show receives the trip of various tribes less than two lengths
+this site is served by the metra station parrich
+she totals the casstantine
+so we have to start life
+all the lyrics are serious with you genero
+several houses are placed in the historical world
+she has a rounded trunk and round arches
+a will in reality
+the room is made of old by alectimar
+wiskrich was born in boxle wiskrich was born in boxla near swindam in the province of cape
+you have the black body to do in the line
+this school according to charles de lavecier had a great reputation
+a month was there for sure
+if you were on the side of a hill the fountain is not if you were on the side of a hill the fountain is not the place of a source
+he died after a fall fragnized by his liberal class
+in the middle ages tasana did not bring the city hall in the middle of the island
+it is without any doubt one of the greatest roles of the classical repertoire
+karlsol first takes part in the chinese that he translates into syria
+shortly after the cites are empty
+the compartment is flying in the media zone has a regular basis
+she meets at the surroundings of sa wacente
+he is the grand brother of bakarin and also a saker player
+as for the paintings he frescoes are not by crista maletti and christian sisin bach
+the commemorative clat reminds us of this spelling
+lafayette was threatening me with a few words and a little bit too
+his peak is located in the county of cusler
+the new launcher enables place and modest charge of
+then simplicates in ports ware the vessel is avoiding
+yes this carad had a singer's aspect
+building is arrested and placed in detention to be in the coach of the palmed
+he also proclaims the addition of spiritual and moral values
+the two following matches are played in the two world cup
+canterbury his captain michael and nice lost his second final in three years
+for eight years he embodies the gipsy of the tenna family
+he is since married to the conomist julian caged
+on the sanitary plan the dink is answering
+the building is then transformed into a primary school of young girls
+he's regretting the loss he's regretting the loss of a callavorator but makes it a place for hermeticism
+he takes a source in mahan hada in the east of the city of mento
+he develops a collector for the racers of chickens
+he studied medicine in orsa
+his current bishop is francis sco canal
+he owns a church in far he owns a church in far a port in a port
+alas the martpolami efton causes catastrophes in the island
+but he left the team in june to join the american team
+the female technics of western africa are changing
+he considers skaters victor han and kin dunson as his models
+he is part of the national park of dore sundo
+cantatenta is the wife of king godfrey
+these two men were fighting and contrasting a life to the agents
+our honourable minister has the floor to speak in a serious voice
+the side tables have a central pyre that can be above
+cheesen member of the country cheasen member of the country's family
+yes he is also very appreciated in jamecca or tahiti
+they include the languages at apuscan and the aa lantages
+he was proclaimed as a saint by the catholic church
+the choir is made of greenhouses with greener and white designs
+it is away and is facing the road and the hardwood
+said the wolf
+the company of versidewad is more mobid
+hugge is a part of sabatian
+among them the painter is a popular democrat and prive part of the buon
+peter decklake is the common selector of the movie
+the car is located in plant cota north of the city
+hence to the integrated living system it can work properly
+afraid of the storm i was still shivering
+he distinguishes himself by the great number of ado which he contains
+at the extreme end the daby mountains divide hubby from the inland
+the movie in the form of investigators tells this adventure
+the territory of two individuals can sometimes be overlapped
+the site is listed as a historical monument
+his first president was francis sobirus
+during that year he succeeded in enrolling students and different missionaries
+this contract was cancelled by belgium after one season
+but he dies after being fallen into water
+the couches currency is the concdor in paris with his equivalent
+roshen leaves the music to become a real estate trailer
+the two candidates remaining in the fight against each other
+this class is not necessarily trivial
+although his book is a success without prestige is not universal
+another one painted by avot gilm is kept at the montgison college
+on the other hand the immerse of rotation reduced facilitates acceleration
+leslie was renamed the gimbod of a yellowish bluck
+its capital is villenfeld
+he left ruzet for gibraltar to join the terranean
+our twin brother of the first has been mary desia lived with them
+munsa is a village with major rain fogs
+seven bridges on several days until the extinsion of the parasites
+he took part in the battle in normandy
+the inhabitants of sirce are called serisa
+come on start madeline go yes
+in the space of a few years of land were cultivated
+the well is then deep and the pit produced
+pom pompen murder for but they are going to reach a final file
+he is still to this day one of the national team's attackers
+he was named after the parisian coffee le prico
+his winner is therefore not systematically the winner of the tournament
+the characters play can thus the characters play can thus taste the joy of managing a garrison
+he specialized in electrotropy and eobathology
+in particular the members of the agency represent the state during the board of directors
+his successor was the bishop of anias
+he is the municipal councillor of torlavira and provincial deputy
+they are supported by the large army and the holders at the level of cass
+it served in the rain it served in the rain in flate and then the hemp flight
+john paul ragnet the french director is the jury president
+bear was locked but release due to his bad health
+its main markets are italy japan russia and the united states
+she opens a gallery leading to a patient surrounded by two levels
+his father claude giolo is a painter and paleontologist
+however this approach was completely impriccale
+we disappeared in the senses
+his discovery gave his name to riumpenia still in use
+madame such addressing asked herline
+he is endemic to the hawai archipelago
+the videous starts in of dark street were take several wiggarts
+the executions take place in conen in lidaho at the same time
+he founded the saint maurice fort in toire rivieres a little before his death
+these are the gods of the olympias
+it was also ambassader
+clemence jeffron is a member of the society of laws and citizenship
+the centre of broadcast is the main drafting is located in ternae
+mclhonco is buried in the valcovo cemetery in saint edersburg
+also tha championships are used as selection for the national team and the team
+the other inhabitants follow their example
+he was the only quatter member of the congress during his mandate
+it's the acting field that organizes controllers all parts of the winter
+the tibetan camp is made up of nine camps with two monasteries and one movement
+he collaborates on several projects in public circles
+other doors built at the end of the bay were found inside the port
+the soldiers who were stored in these surroundings had their own cousins to regulate their deffrence
+his mother married later the actor in castilder gimo
+in contrast he is a member of the masonic lodge with leipkin
+the vestiary of the air port is the centre of volevile de barselinet
+hector gilting is already a member of this sport committee named in this position
+under the government of manuel lefthia his political opponent he was brilliantly a breaker
+the god is in fiki the god is in fiki derico noma koto and vuia
+the whole of the cards mixing then spread out against the table
+he is married to bergesats
+if he is inhabited he can only be only by survival beings
+henri richard excerted the canadians in montreal signs the preface
+he ends the season with the bracoses of denver playing four games
+thus most elements are specific to the point of view
+he co ordinates the majority of the presidents to the same assembly
+it is the german runner tony martin who went in the head of the passes
+it also bears the traditional name ofa
+alenzo's father is from mexico and his mother is from england
+trilogy is the first time to be removed
+his name comes from the river close to tulga a tributary of the voga
+he grew up in the humble conditions in the benois district of opelstadt
+she was sold in four months
+each category of weight has a chimpin
+the game is based on the exploitation and discovery of the knill without any specific objective
+his father is bosnia and his mother is serbia
+the air force bombarded several times tuma during this period
+his activities as collector and ourt historian were then completely ignored by his students
+symbols alphabetical were found on money in some of them
+the eldest son afied conversations this process
+it is therefore the grandson of prince wifen and nephew of the last emperor pi
+if the child survives us we come to help him
+it is located directly at the door of aber villas
+the hermel city is a public road located in the of paris
+he also translated into english and poems of staphon milarme
+after their introduction the nucleus were adapted to the taste and local use
+he is the only australian representative in judo during these olympic games
+his code name is fenwick and references to the same name
+we can also find seros to frase in water
+clemence is raused and brought up the smallest way to the north
+it belongs to the kenton of ferney voltaire however in the most closest to gex
+a famous student of the eco normal superiora famous student of the eco normal superior digers she continued her high school in france
+even precise as the television only lesson is managed
+this one took it one of his threats ofvision
+hitto is the near the sea located dow the sea
+the access i open roon requires the use of boats
+it peaks at altitude at the ditch of tabanon
+it is in this context that the tutors interrupttal terro
+these grandsons took part in the revolt of emir abdel kender
+the patrimonial cities with the columbish to attract many visitors
+sphere of competition he also represented belgium in the competitions
+the veruga flows close to the city of cosmodamisk
+she will manage there for forty years to start working on the institution through her teaching
+the first is quickly completed and offers a view on the ocean
+she studied in primary school there
+the standards aim to improve the communication between set of works in the sector and the supplies
+he is thus he is thus one of the oldest french channels
+great enties
+in france we have the same way
+he seemed that he refused by saying it wouldn't look at me
+shes eno tercenace is a famous blaselo in the minamodo cla
+he also disappears the castles of insange in girmash
+originally they are simple swims of a pager
+he studied political and economic sciences
+the accomplishments were pulling from
+this species is named after the pater henrier martins
+at mexico she married to tihasten
+to day the project is still not carried out
+he houses the mount wote which is the highest mountain in the country
+erson is a son of the opereta
+the first proof of wetherburne is itself a source of correct action
+johnson depresed the chronic again by marrying lucille cameron a white woman
+the book is dedicated to his daughter enna
+he was also the godfather of athanius singer of the roboceli peppers
+she runs along the dial between old youth
+cables was the first ever sperentes and bloody star
+the inhabitants are still in major or malget in china
+the tour was announced by silina gimes itself
+he is one of the first personalities to focus on economy
+if the men were cut off she fell down from the bridge to my brother
+with t e success he came to pay attention undecided
+john marie lecok also has been a trainer in the young theatre and niderine theatre
+the action takes place at the imperial court of china in fabulus pond
+if the image is cre the passage of a value to the other is more gradually
+he came back to the mediterranean
+here are the directors of the successive publication of the review on these different groups
+she's held there with elizabeth laporte
+but in fact the lack of natural resources is against strong german efforts
+the family is over but the banker kisses on lavy's shoulders
+that is the condition of the pater but it is also his game
+henry benishd govil was born in months
+the term is in paris in the field of web applications
+i forgot him to complete his instruction
+she accepts the proposed duke
+the molds can be comeplex without any rise to the cost of manufacturing
+the kernels of the gate are destroyed
+the small semilindrical tower preceding the farm served as a defence for donu
+the police was hanting to have been to salu in his work
+the hospital escapes from bombing
+in this game the player embodies luck one is dead
+it is the first game of the president it is the first game of the president's team in the first swedish division
+the air control is managed by fernats
+she became patron of her life and corporations of podters and leaves
+it is then broadcasted in belgium in the top on the one
+but these ones are dispersed of any historical background
+she discovers the young alexander legribber who became a natal
+a nitter system allows to give more speed to its vehicle
+a neutral country is therefore not in belgium but the reciprocate is false
+amed elmi is married to actress egyptian mono zeki
+a little evolution only a good hygiene is necessary
+the song is well received and placed in head of several hitsbur
+she grew up at the type of bylane
+he is a member of the municipal council of midelin
+he went to the financial plague in the castle
+without succession to the position of swiss ismonsanier podos
+he was one of the students of pefgeni kosmersky
+i thought it was beetles were only in hot countries
+its situation made it the turning plate of transports to sichon
+they brought together the hard and vestival to a furian level
+the two ceremonies have been held in monica
+this room was the former library of the order
+this building is larger more solid and its roof is covered with tiles
+this magazine publishes scientific articles in psychology and health
+a large work planned to facilitate the preparation of the plates
+these descendants the budja and law live in the northern part of the bumba territory
+philip is a friend close to the other player who comes from the mosir carlpo
+from then his own power declined
+if the writing is sometimes apprehended it is not less severely fraternal
+he does not attract little supplies with the stones of the mountain
+the latter replaced by mac munza from ferman
+so what do we see asnichell
+vin promised two thousand roubles but wants the article to be written immediately
+most of these units come from the saxon army
+shortly afterwards donal discovers makin madley
+he was the second san of long
+the number of the painting and graphic are awarded in millions of books
+the kems were built in the district of albert
+mistigbl is bordering the village of achestals and nice bay
+the unions constitute of the institute of the diocese
+they were the main economic activities of the city
+this book is still a reference and has not been nor really contested
+it is first with the duke joan de lorraine and robert demart
+in panjavi the six use the term christ equivalent to grista
+the urban unit of vervis is the urban area of the urban area of vervis
+al had two elementary waters per day
+harvesting a yellocard he is sended for the return to sanziaro
+he wants to invite him to go out with him
+the airport is located in the city centre of san bremondo nevado
+the prince enters his youth in contact with six frads peter
+in the king's ear is in the king's ear isdamu is the charge of vizier became a military
+the clup realizes the douvils by winning redelin's cup
+one hundred sixty seven one hundred forty three
+sebastian jena realizes the coupling
+it headquarters is the village of abro which is located approximately east of verunto
+the system of the samidiated plans is made of light weight and age
+the avieve school was reinfected in the local school of the year
+after five months he is jemic he returns to his studies in the country
+however it is not the school that is compulsory but the instruction of children
+teancient catalogues attributed him to the audience of many perus
+she has spectacularly improved her athletes
+resistance he was not questioned
+the other ingredients and nights are then added
+he stopped talking in the instructor of this prince
+his capital is donaville
+most of the authors insist on the work of some plans
+this project has already allowed to form a hundred training courses
+the whole inland games generate many paris and attract crads
+in the middle ages the brecon spigins joined the massifs to the west
+he is orientable to desire by a vice
+he has a com
+we will have more than ten years to convince the international scientific community
+what a wonderful show was happing to the eyes of the travellers
+this building is located in the vuttery in ubehu
+he is the father of the bascaba player francesca ras he is the father of the bascaba player francesca rassi
+i travel to thirteen years to where he settled as a portrait painter
+he is a specialist in the history of ancient egypt
+both races are joining south of the city of mut
+and till then only god has a place of life
+for example
+it is found in the middle of the andes in a geographic area
+the materialism of jernin will not stop there
+most of the missionaries are imprisoned or expelled
+but i can close his eyes goodanapart
+it is one of the greatest sights of pilgrimage to the south of thibet
+his son denis also made a career as a cyclist
+he is named in english where to fit alpine
+saint andre church was located in saint andrea street and rue du cheve saint andre
+species of this type are endemic to minesera in brazil
+the sicilia leaves were boidable a visible nave are accumulated
+he's a complimentarty works with the sat enges museum
+despite his efforts the girl defeated by the director who holds a toxic power
+he has a very special relationship with the first of them roger nemer
+suzanne is a neighbor and the daughter of a wiger
+the channel was renamed south studio for a year
+pierre constant budden created the hospital of paris consultation of the feedings
+he is part of the arctic courtier
+montefello to day in museum is home to the rest in schools
+the main entripenor of the bunny family is called bartelene
+he is buried in the carkmtery
+he left his heart to ossan and the rest of his body in frontof ro
+probably the sigation of green bruno he was named archbishop of traves
+these privileges are extended to the colonial empires
+as the text does not exist any more we know very little about it
+like gasse and poet as well as painter
+he was murdered following conflict with the lator
+she brings the subject of the souls tribes and the two girls of girls
+but she had not moved a place
+daniel mckenn is a graduate in the science of education at the university of montreal
+he is the father of leon vender kinder and his father
+it has a roll for its stirring of a small sleeve on the right
+it is managed by macedanski laniski and fra structura and freight by massa dunsti
+the last fighters stopped the fight at the following year
+these animals are represented in the logo of the hotel
+it then flows into the beson vantahedo
+of french championship roman is made through fruits
+the visignacia is separated from the viburg city among the linets ari
+she sees the orthex rather than e care of life
+their father calle
+shortly afterwards geffrey was found dead
+he brought the first settlers to about thirty families
+she's the best of the savior was photographed by abe
+she's the second of the five sisters
+the presbytery is disused at that time
+smetier monumental and landscapes it has many statues and madele sculptures
+he is the most common in the mountains of north aberic
+at the beginning of the first world war he entered belgium
+the entire case is accessible to the dishimility persons
+his chief of body is dominic te bu his chief of body is dominic te brer
+a tower is covered in slate and the neve covered in tiels
+a destinality of the coherence was organized
+some sources indicate that the first castle is built in wood
+iser is in relationship with a young woman named carolina
+a lesson of humility
+after his death she goes to jean
+at the beginning the second world war place junou took part in many failures
+tesaya had relationships with lagmar in particular
+he became the prime minister of joan caraja and alexander sito
+the two condemned could choose between chapes and injections
+axicor is thus named governor of le
+the chief of the family is indicated in the registers
+this term represents the maximum obligation for the seller
+the companies of local interest are also concerned by requisition
+the regimental college of the federation of private education in quebec
+the petarch o poer wins his national
+it's also a caterina list
+the interveners between the couple and their inter the interveners between the couple and their interes in french as well as the senate is french
+the frescoes are kept with to day at the uskin museum
+he lived with t old fush levened the hut of the garden
+the columbian delegation wins two gold medals and three silver medals
+she refuses and finds herself in the city without a clock
+she's first known as a singer before starting a career as an actress
+a protestant school teacher was appointed
+he is then dismissed by radiose
+she died in beakhost the town of the suburb of washington
+this firm brought two major contributions to astronomy
+coret who is sincerely in love with michel is one of their first admirors
+the evon of its metallic section measures long and wide
+theyar fixed enneses located west of the brookes town
+pillister is a u tuber
+one of his first posts is officer of artillery in ranoble
+she took place in the following day the first day of rest
+the overy is glaborous
+the elogy of agust fang locarol on the site of the artillery liberal
+two sons paul and philip will be born from their union
+then he imposed during the following season
+it is the only one can be easily perceived
+it also criticizes to accept a lot of planks including some of them are unpublished
+he also makes the flights of demonstrations in the new york region
+sent up as a major figure
+he occupies a time of the cocuster de cors function
+she takes great inspiration from her spy
+many of them from the archipelago are corallian
+this way was a strong roman square
+three coffer street belongs to the northeast corner of the square
+schubert started to belgium about fifty poems
+chloe is his best known player
+jack lavy was born in a family related to the resistance
+it is during his college studies that she won miss andrews competition
+actors don't move to because of their too impressive costume
+with that the curtibe municipality is searching for more than a federal government
+its capital is located in mamboi in the city of the city
+brido is from silesia
+the success as a movie begin his film career as a movie of actors
+his father in law wanted to card him and send him to a pension
+she has a daughter helena and a son bastian calidor
+you still want to be a hard worker he was gitty up from the night
+he then lost his title in the favour of hippoma kanusi
+the medal is the work of the sculptor robert graham
+he travels to a traid station at another one
+his wealth allowed him to be a great patron of arts and letters
+at the end of the first season the caraveta california
+fakenburg is the headquarters of a county in the county of encante until the french revolution
+the final series is played in the best of the seven matches
+in the cinema he has toured in the films in netti and pan cabe
+the tripolet of the sized alphabet is the main character
+the capital is the city of dikebu
+merceille marked a goal on penlydus by joseph cupbard's intermediary
+she was in a child of spain
+one day a curious living room raises the cover and discovers a white snake
+he participated in the national repression of the sema revolt
+this species lives in a high altitude with a low vegetation
+almost nothing has been kept you the old man in the cafana
+it is one of the biggest lakes in the world
+in modelling he was stolento the vida on this occasion
+bushean is a relatively recent locality
+the team of meta bolonic anocis is simple and adomated or being under development
+the castle is located in the french department of maine edleoir on the municipality of ener
+their feathers are often dominant brownish
+you can see him touching him
+i've not the king and the doctor of charles and henri's autopsy of the calis
+let's observe but let's not insist on these five minutes
+it is located in belhem in the department of am
+the name is in reference to the swiss or swiss cemetery
+it is the story of a strike in the vineyards in alumia after fenkip
+when spuring comes the house of the fox is broke down
+however the productions under the name sky box are stopped
+both versions of the differ slightly one of the other
+it is found on the cheeks and soils
+the fa places are located on the municipality of ubins in the french department of everon
+several countries could participate in this edition
+seriously injured he was then putted with a leg
+it is bound to the south by cape and the koloski peninsula
+it is an evolution of the andro where couples have individuals
+the tower was blue and red as a distinctive sign of the day
+she's named in honour of the state of daniel webster
+she is the subject of research more and more deeper
+it is the only glacier accessible by road of the province
+but some royalist and royalist physicians remain
+this diet session is known as diit of partition
+hunts billness was the co founding of the hamburg administration
+in midway it consists of a dialectrical material that has a loading skin
+the noble armies of the benedictine are on the lapire of the door
+he realizes many portraits like a fiate
+this decision will make him a recognition of the angins
+the production will be interrupted an different disapprovals
+the writing of poetry has a certain number of novels
+other objects were actually of copered
+this species is endemic to the north of the kcen burma state
+this edition marks for the first time the victory of an italian
+a study suggests that consev effects are possible
+it connects the frenco district to the south of theb
+he is the son of stephan cristanski former national cheka slavakian of volibe
+this pook is now sought after by small but leophilic amateurs
+she's a fan of the wallets and the arms that were allowing them to payment in comtact
+it is the oldest of schumann brothers
+ammanuel also claimed that the dress could be commoded different forms
+three of the four soldiers were killed in the following fight
+his role was the support of the infantry during the attack of fortifie zon
+he was the father of the waste ornete laurente demasers
+he finishes a vouting line parallel to the axis of the vessel
+its territory was then divided between the district of bayon and saint laurende saint martin
+a doctorate programme is proposed in association with the free university of berlin
+the series has been hosted by critics
+to an event axis becomes interactor
+armand has headaches that goes all life of love is missing
+the machine has antibiotic and anti pungic properties
+gradually jacques rambeaud was his canvas in the middle of the war
+back in boston he continued his studies with george shadwick
+it presents its resignation later without having taken over these functions
+he is the subject of acquisition by the national art fund
+a surface area is under the public management or the current is strictly controlled
+the co neighborhoods try to reconstruct a negative plant of best known micro climes
+we put the problem to know what kind of effort could be calculated
+the grenite channel is a peduncal chain located in the centre of the village
+we are bringing his support to the abortion injuins
+his parents are jeffrey and jeffrey notable his grandfather will fredmont
+the failures have then been separated
+the columbian headed is generally considered as a lively orper
+it humtown dedicated him to avenue
+he works later in a frail and travelle bank
+the master's organ is connected to the place duveur market in the besives
+he had the lands of polish in western russia
+she is also the origin of the moon line in fanville
+christian gezel was awarded the legion of honor
+after hearing the form of a soul as a weapon that bernard is already married
+she notably crossed the edgarky street
+this simple and reputed to be the biggest ever sincere to the man
+the second one can be treated by a poison or also by customers
+according to malun it is completely according to malon it is completely according to malon it is completely according to malon it is completely sgish but its importance is of order
+he was pronounced finetically he condicene as a pajorative
+she's a medium part takes place on the stops and reliefs of the heads
+on the other hand the clip is broad casted only on the uninhabited night
+the village is located just south of ulivo
+the municipality is a rural territory mainly dedicated to the great
+in total they are very pricary which have been produced
+it can be used to create consul or graphic consul applications
+the club illustrates the new musical direction of larry
+senop serves as a base for ottoman's navy which has been built there
+he takes a knowledge of the annual valuation written by the office
+the texts on the subject are reaching emigrants and priests
+all the monuments of this cause are made of limestone or dilemite
+the composer lewis farnett who studied music with him was his son
+he is defeated a first time in murder then in france
+she's born and choe from paris on def
+the greek forces of the island are engaged alongside the pera military organizations
+he is mentioned as the imperial descent
+it bears a grayish brownish
+the addition attracted on the edges of the roads
+the tourism is a major local industry
+he has always played twelve matches for the columbia team
+she married the florentine dencer in gilovestrus younger brother of guato
+so i made a true art work
+you don't want to bring any sales
+she would have medicinal urses
+three trustworthy person is also a journalist and ecological activist in the conference
+she was then called in latin
+the plate can then be deposited in the acid
+there is only one road for the circulation
+the sabatach projects are mur thinking prepared
+his meda his medalin proved from the fact that she was disciple of alder represented by a faith
+it is the becendrum
+no details concerning the landing in municipality of the same commune
+junty ears is taking care of this
+it is free that eventually made him become a rider
+he particularly beats eight months in the monastery of lebren
+in january the family severely living
+the measure may be made fence to the sales price is supplies
+all these jones are part of the teno filladay family
+the song particularly produced by brefe is the third single of the album
+he was from the turkish family of urban origin of coprans
+the posters are the same but the plateaus are overturned
+she died in gudenburg in new jersey
+the clip was then shod in california
+gustav adolf duke of skene studied at the university of apsela and oslo
+he realizes the corpses of the gardusk and gives them life to the shape of a single
+his exploitation will stop two years later
+the coventia of first importance was taken into a freezing hand
+asetama gave a book from his library
+the summit is located in the perimeter of the regional natural park of the netherlands
+then we will start this exercise infinite series
+one day obi foet a magazine and discovers the identity of paul
+but above all they are his assassins silent incentuaries
+the hollow is moving on the peninsula of the ucatan in without developing
+it is found at about fifty meters from the river an apple tree
+during the last two years of his life he became a consultant
+she's the seat of the bishopric and isis of munpo
+the municipality has a post office and a fire station
+he hoped to concile the favours of the winners but did not apply to his mind
+after having studied medicine he became a surgeon
+however he has a possibility of liberate himself
+to day there is cultured and exported from the valley
+everyone at home the bomber at home the people at home
+all the teams are found on a round trip
+to whom so we are married to and has two children
+she is part of the municipality of zope
+the television offers an immediate gratification to the viewers
+the saint gabriel church is considered to be one of the most beautiful foundations of the prince
+these fortifications remain intact
+malet himself to panenius and leave his legs to command
+we owe him many works made for the churches of these regions
+lucian ven hump his color in the head
+poems and essays as well as their masters
+the object was discovered by the astronomer bedeni dicho
+he now rests in the communal cemetery
+maximum speeds different limits will be fixed on the road
+the adventure experimental takes over and over the longest
+the family then moved to say croy de la botonery
+do you see my son in law there is only one cellar
+the self paul
+rather than that sir the mayor
+after the fall of the work of libeonia the castle remains robust
+the earth quickly turned into a lake marsh
+to day the castle is hosted in the centre of rhodez and averon football
+later he became more and more prosperous
+a few years later john emmon joined the management of the label
+she is one of the first union building influenced by neogothic architecture in england
+in this movie franchise plays a policeman
+the dutamentation is available for the request authorities
+he also belongs to the j cera label in ferrari corn
+the tyland is one of the most important fruits and leaves
+he was also a diplomat
+from a jewish family he was poured into the hebrew and islamic writings
+a small museum school belonging to the rocorken museum is located in pra
+his father a sailor is not often there
+the boat was a cillo that was sturdy capable of resisting large waves
+he's not allowed to find several names of snichal for the same year
+milon was born in the middle of the time in marseilles
+he visits the communal school of fredy where he gets his primary education
+the church is located in the french department of an in the sipality of par fando
+duka isn't duka isn't starting to come to this point in arriving
+the locality of baakora is a capital of the municipality
+its east coast is part of the great edge of the great piece
+can be originally from tukia california
+the dixons then joins the bordr near dersens by crossing the vale de reminses
+it is often a holder and finishes the season with two goals in a championship
+they are named after the ereant islands in papo newganea
+the domans rode measures long and wide
+the valley is important and located several gilometers north the valley is important and located several gilometers northeast
+this species is found in muron and hakeston
+the series was published by netwark in great britain
+he imagine our nerones
+he went to paris between the odawena mountains
+this is once again compensated by the increase of the towers
+shortly after he went to mission nenita
+the same year he married aleta vencerest
+finally a poor man finally a poor man's destiny was a historical context
+access to this form is difficult and the site is closed
+the role of the catholics is poorly known
+on various occasions he designs and creates female architectures
+sullet's trive they succeeded by flimming a part of the old man
+the lips were in fact an actor who simulates trumpeters
+he is a doctorate in the university of amsterdam
+clubney is the founder of acustics
+some are subdivided in the communes while others constitute the administrative unit
+his father is a wool and a cat
+this painting represents the best goals in the history of the spanish championship
+portugal never won medals against the world
+we can we can't smile but we cannot deny despise nor to her
+his work is not composed of looting and unis
+he organized a save action
+copadicus collects high solution spectrums for hundreds of stars
+the population of the one jurisdiction is evidence
+when only were collected springer did an additional cost
+the headquarters of the community of communes is located in pont labe
+nevertheless they haven't common senses hitting the hands on the boots
+he is a member of the chamber of commerce in paris
+it was a student of philosopher paul ridger
+the children of orphan were among the indigents of the victorian era
+hiden hiden hide dare asking if he is a doctor's are also guilty
+te lentus is one of his dialects
+we are counting architectural views of allegory planks
+figerus is the seat of the house of lamporda
+his value made him attain a quick step forward
+it meets up near papusor
+my heart was the professor of the arts and the terrible
+these three brothers and sisters are also poorly conselend
+the birds of this species are found in polynesian islands
+it is the fifth and last single from this album
+he is one of the most active contributors in this reflection
+it's the wisdom and the knowledge of this
+it was a fury
+mosesla advises then omer to invite him for dinner to complain their relationship
+present in france especially in regions of altitude
+the evolution of his heritage is a sign of the recognition of his fellow citizens
+the polymorphiss are relatively frequent
+she succeeded the getter of a team with new electrical sponsors
+the block in the final is completed by herbelers
+the temple is built on the slopes of the disan
+she did her terminus at the relegina
+thethe dister then becomes a distress the dister then becomes a distress or sutopeat treaty and stabilized
+he hided and survived in white life and roots
+she under the pressure of sardar valba betel the union again
+the young house dwarf appeared for a long time to choose artistic career
+since the beginning of the short movie have been produced
+the baasist chrysalis joined the group
+together they tried to keep the scientists of the death of amian
+edmund edmund then goes to the mag of the greek cup
+john philipp cantier in a family in bourgeois and catholics with religious funds
+this one combines the tools of the type of different rolls
+antoine confessed his father in law with refugees and settles in clairmond
+a momentum the album was called crandon jorid specis
+the forest shelterd several treatenises
+its towers are the symbol of the epidrome of the derby
+after the war he moved to se armen in prise and conceives many members
+the island is the refuge of a wild pony breed called shincantin
+he is a playing as an attacking midfielder
+early on sirmat is also a show of two thousands of silence
+in his youth victoraba is a student at the university of paris
+there are also artisans on the state markets
+the brickhord that she was poisoned at the duchess of boian
+the movie was presented in the venice monster in the first world war
+it allows for a second delay to be used to plenify the leisure weakened
+the conflict became national and worn in front of the state council
+it is a restricted government in raad and aden
+he does not play any game during the world and pinshp
+we can show that the product scaler does not depend on ispesitation
+in china the worlds of junmian are part of the yal nationality
+niki alley was born and raised in a family of syclic confessions
+the situations encumbered and cocos are still in a way
+slaves and sovereigns wanted to obey because they could rule
+the area is the population can vary considerably from a sector to another
+he is actually made up of two distinct points almost critical
+bar kilfredo was standing up and heading against the duchessde all his secrets
+just after the war he was chief of orchestra for the casino de paris
+in the nucepil aromatic substitution the other reactive is
+ryan then signs with a record label capital records
+olivia prostitutes herself to hooto him
+ma n a m a n i alley in the northern districts of marseilles
+its chief town was verdo
+during this season rumors are short with the possible selection with algeria
+the same year the roman as the same year the roman as took place in the third of the year
+in the aggressive method the core of the deliberately overtaken with a hem
+the castle was then owned by the duke of va the castle was then owned by the duke of vasele
+he renews this experience as a leader of malayan expedition
+walk with numerous galleries in france
+it was a digging in the meant to save
+eugene blenff is the cousin of the never on the other
+he is rent as a historical monument
+e large part of his musical work is published by shop music
+well what do you think about it
+the specific mechanisms absorption of the cedmium are not really understood
+administratively he separates the communes of brehner le cla cersen and sasir
+another place was offered for a new cathedral
+this episode made it difficult to be mischievous of all forms of authority
+this animal also reminds the sa joseph's genealogy
+its administrative centre is the city of veniste
+this is probably the server smith replied
+these molecules are present in great quantities in the plentel islands
+john whites then released sulla projects and fissered projects
+a deep rabbi to the south and east pards the city's castle
+the river has its source in the cordlier of australia to the east of kingeroy
+no one understands nothing
+go therefore for florida and still tatarn is not a poet
+she owes her name to the village of anes which is located there
+this delicate nuance was the subject of many discussions between professionals
+the inacre seems to have been practised not even taught medicine in oxford
+another torture is to skate him bulbus in the fingers
+a total of a prisoners live in the camp
+among his students there is clarnetist of jaz brawnie cotman
+pierre gilman fact brought his support to robert forson
+the bishopric is in frontan of were the cathedral of the assumption
+this part of the documentary is ruined by intrigues in love with so ruppy
+he will get his diploma from the secondary school in ba he will get his diploma from the secondary school in basad h
+patrick ribad laid her ton at the young referees in the region of ases
+he tells his adventures to radio and south radio
+his north bage is a few kilometers south of ecuador
+he takes a source in the poet in purson and flows into the oves on the left bank
+saint malo le chatresard was this point
+the village is accessible from the highway bendo balo and
+he became a ballod in the tennis club
+the building is heated with the sillar panels
+italian uses the latin alphabet and the latin alphabet
+with the national team he occupies the left back position
+kakum looked at him coldly without answering him
+his sister is the writer katherine prinjokich
+he is trained by dave abas renkeims and george nichoson
+for a few hours the adventure of gwynplaine had been elaborated
+the company is founded by jenstadt one of the three founders of greenland
+the river is crossed by a game by back orbridge
+five louis different comes to neglect by this neighbour then a young monk
+his name comes from the bobasin region of new france
+my replication of the pant fars has several steps
+he had to make the transport of goods in the garrison of loers
+however the fabrian army remains the most fearful of the continent with that of tory
+it owned a camperlay antenna
+there is a large social mexicity
+the participation of catherine secho is announced in this production
+he beats together two days father
+this lordship goes through the centuries and then to the domain of
+he is part of the notary family
+some are nocturnal and grisvals others say
+he was available inversions of an four engines
+he gets a charge of archdeacon
+tom suppose of this situation and a certain amount of breaks off at minister
+a version of the encyclobelia is available on the lock
+because of this murder the king sent troops to kill she
+the body is separated from the lips and pitmot
+the price marier is the main economic region of the far east russian
+the alps form a barrier between the mediterranean and danube
+claud devancourt is named president of honor
+the plantation the old five hundred eight three hundred and fifty six
+the doubles event is won by jundry and anhops
+the ghloride of aluminum is a catust of the reaction
+ortgret was decorated with his work at the order of hester
+in france the enconbren or career as a more serious career
+such a link is called metalic links
+the awards list
+he is a lawyer in the president
+escor oesco ersko loby is the father of the swiss journalist helo
+the seragraph can reuse a screen for another work
+the plant reproduces very easily from its seeds
+the whole is placed at the fire until the grains are killed
+this team is general interest and represents major concerns
+he scores his first blue against this team during his fourth game
+pauline was the idea that he was depositing in all sorts could have printed the spirit
+with the event of the second empire river must take place on the application of this bag
+he discovers that he can he discovers that he can he discovers that he can mix his wine with water
+if this proportion is too important the bitten resulting and less solid
+the boy discovers how much he is difficult to grow
+his bedrooms were not limited to the public
+this process is generally used to produce automotive panels in mas
+the surface of the engel river is generally frozen from the beginning of the runs
+other factors in consciousness of the dangers link to the field
+it's a monster wollping a sticking and stumbling man
+he is probably from atempes in the morende district
+here is beginning to general donadeu a new career
+english procedes french
+this species is endemic to the cordera of the coasta in venezuela
+the line connects the bes station cogero to the her shema station
+but the group chedists will progressively resume the initiative
+his relationships are close and the public will have more easily changed
+his coming from the niden is comparing the best out of the country
+the champion ship is played by minitory machines at the end of the semi finals
+he gets to the batiqes of the uro cup nero who is presenting more than eleven children
+he built in memphis a temple aiste
+meanwhile the ima was made of different colours
+arranged for saver and cent her to the united states
+jun drive into his room with a strong smoke car
+mister baron will not hope any more he said the valet de chamber
+gespa was born in pelan alexander and grew up in tennessee
+she's about all aspects of research on by a diversity in se costumes
+the line is equipped with automatic lighting gears at the vessel
+this same wednesday a large butcher is treed on the square of the fruit
+the garangran estate which is also acquired for environments and meetings in the together
+there is then a guaranteed stability to say we found the solution to the problem
+we are toning ibout a high activity in theatre
+she has a double continental climate
+for copying something we pay four sub species
+a walk is estimated that he is oing to singular
+its capital is the city of vellaco tornova
+rio liversaaca captured by a game
+it is the oldest french alliance of russia
+azar de hausen he met claude bordey who accompanied him to bourdenville
+he participated in the training of the competition for the dollars
+well yes by coming we will have to cross the country with the previous one
+the christ life is not about your rest
+he then took part in the part in ircania
+this position is comfortable and brought some genoese territorial gemanese from the art of his neighbors
+he finished the general and seconds wist behind dide l couche
+the fief of latore moderned the duchy of chadlera
+his presidency is marked by the heron case and the financial rasis
+charles hendel was born in a house located on the place de fertesburg
+he is the brother of leonad burman
+tuncave is in charge of the negotiation that he was not long
+no one can answer with certainty to the question of vo pajula's origin
+this fund is made up of thousands of modest manufacturated and artisanal art
+in french this can be done adding a proposal an infinitive
+henri de lion chased the children of the duke of fire and took part in the county
+the current mayor of sanju is legion back
+see among others lims and messes
+they are general director and tiere deloros
+marius cut his way to see him immediately
+then a strategic pirership with sabre
+the community retains the church the community retains the church's enjoyment and the capitulary room
+him helps him learn something right and he studied caligraphy
+he taught physics and pilier
+the milon trial bears on manipulation courts per year
+the couple has a son pierre lewis and till destiny
+called him to brussels he was taught at the college tars in le retalicue
+this place would be closed by the attic of abundance and tree plantations
+both the first volumes appeared to adopt their novels from the long time
+he blew in a sportive and chubby manner
+the church was built in a car and the russian empire
+the law will only make it possible to get four clients of exclamatories
+yes we can find the veterinarian and a substantial formation of supervision on stage
+this species is native to china korea and japan
+all three medals can be attributed to the order of members
+seven hundred a twenty four from the burgundy
+the measure of the dance is an excellent way to define its estimate
+gilliot gilliot was the man of the dream
+the links that follow are in english
+she has a teaching of nadia blonger and lazare lavy
+his friends organize a party of god or kenico showed herself to kiss
+several reasons are pointed out of the law different depending on the historical period
+oh that but who is with her
+other famous actors in actress play secondary roles
+he thus he thus worked as industrial and commercial activities in latin america
+he is managed by the service of navigation from geisberg metericht
+the floors were improvised according to the circumstances
+during the march revolution he refused the liberal reforms
+in terms of simple terms it is the standardized exposure level
+he was thus the first president elected by the
+there is no neeto enguld god
+this philosophy has two levels of ausis
+the movement was put in honor in the tenth room
+he aims and works and bin he aims and works and binos also as a keeper of ruba
+this admission examination is based on matocks
+there were broken devices
+he had an automatic pilot
+in parallel he changes the athenian political system by opening the art to everyone
+but in the championship the borshan mancan klat the hague in the championship
+polypia is a heade roid family
+but it would be great cheos disappoints me to write this love
+however the latter laughed on the back road
+the united monastery the youth is threatened for us for a couple of times
+he lives and works in new york london and in ireland
+thus tapink is a big french electronic music
+without a doubt his childhood and without a doubt his childhood hadn't heard from him and that was left for him
+his works are most often painted in creek and very structured
+pelagel pelagelotino created after his marcel de marie
+at the same time when you were in the other hand
+there is only one way for an alex
+the following week he moves on the podium of slobing altabatia
+it can be down loaded for free on the associations web side
+the solitary cross is surmounted with the direct forms of dragons
+the head office is located in nenter france
+a general inbuilding of the surf
+the castle is rebuilt
+in the direction of the white sade intercells the white study on the ditch
+its headquarters are l emand
+shortly after this victory dil moved to vienna
+the decoration is made in ties
+ne inter ne inters have come from an external country in the urian union
+this great variety is found in the metablia environments
+the young manifest in the camp
+its surface s of its population of evidence
+it meets in the west of san diego county
+he is nevertheless elected from the first round and has to additionally a long time
+you can kiss the details of the confrontations between scotland and letander
+the one who are not reproduced here is astrate
+the only one in the vestiges of the pit and its domestic room
+several hundreds of alphaset are still being held in france
+this chili is a controversy
+jake and via managed to escape by allison now
+the province is located in the centre of the country on the lamani river
+there were several accidents involved in the river of the uranian
+it is used by media polymistics
+the star of the order was an octaginal star in silver
+boloskina stupid asked him for forgiveness
+he resulted a great period of political instability
+the first woman to receive her
+hiki tar hiki tara is located in the vessens of the arctic courtier
+the importance of these instincts is of need for the work that lasted several years
+he also actively in this general jurisdiction
+claude accepts freely to his two sons without business in set deni lespan
+populations in arabic pursue a migration to the next day of enturies
+the german name of the village was grapin
+he took a jersey and mark o calido's rock
+the first houses in columbich were built by the boilin method
+he represents the conservative party of the confederate
+we do negotiate the commitments of the standardized laboring to standardize things
+he imagined the most weresome who could have fantacizing
+he is in charge of insure the order after the riots
+it is located along the river cantu
+in germany azev works as a singer and a speculator in stock exchange
+the choir was not effected without any harbert would be dead
+finally he began worting as an actress for television and cinema
+the club will insure its maintenance and exercise
+he plays two matches in copolivo de dorels with a minions team
+during expelitione he was decorated by the three of the people
+the church has several references to the roman art and art
+this species is named in honor of julian cristop for mush
+antine marin visits but she died
+she was built in sanstone in a contemporary style
+fotty villiers is located in the bave district of the department of as
+this elbum was published in the dal court editions in the mirages collection
+these multiplic assaults urun the adverse and fragile troops
+he is the great sister named irene
+it is available on the purse site
+this species was described as a per
+the vote was compulsory for him
+augusta de cambridge was born in the country of montborough
+she lives in the lionane region
+una kojima is a model for the fashion magazine in habodi
+in a study lethnuxis ends the career as jean marie degler
+explained the indians and the canadians
+she also organizes training in new york and activates the musker movie
+but according to the door that was no consequence to my dear palma
+he costed a lot to the parish for supplies and equipment
+this disease affects the disease is therefore very subject to soils and by contamineous animals
+it's hard to be alone in fact i'm saying it
+however the countries signify a chance to issue reserves on certain articles
+the d thed took the name of mac elleric the redistration of their decrees
+the region will remain under the turkish domination and till the french colonization
+these species are similar to all over the tropical and temperate seas
+then he was a teacher in plastic and then a skeer in the radio
+this master this mask's painting is often found in the presence of the banjmars
+coursand all the following winds
+there is however a selen destined and organized by pierre cott
+both of the musicians jumped for the following hymn
+denounced jimmy is arrested and tortured
+the spent the spent the spent the spent the spencer they said to be delicious
+the skin is first wet and pulled to is the skin is first wet and pulled to assess his sis
+after the first wild war she was attached to poland
+she's not present in the habitats ave grated
+the finland indicas basta remta finishes first place in the world cup
+he is the younger brother of actor timaty bottoms
+he plays in north watch city
+right leech wii is the first female jaslig
+painter of history and landscapes he is the student of jean barden and john baptist
+she performs in australia but also in europe
+he started his career in the order of san augustine
+she married to an officer with whom she has a bad marriage
+at this time the company was only a small boy garage
+he is the brother of jean marie corlet a member of the loire atlantic
+its album reaches the albums of intenent
+its presence is swizerland
+atrino was made of all member of the groups of a bsatrino was made of all member of the groups of a bsezing and zung in the groups
+he confronts himself with the supporters of the reva
+during the war of dependence the united states the city was virstly a courser
+the first martyrs of luthan are in verse
+these bodies were replaced to day by a cover generally in hatred
+quine equatorial common
+it is not sure that the altitude of this mountain was correctly measured
+philvetic stadium as won the title by winning six matches
+the village of senu is one of the oldest majerks
+it is found everywhere on the northern hemisphere
+hubbert square study mathematics at blell college of oxford
+cinbiva studied in the field of economic sciences
+the locality of amin is a major part of the town
+in his cell to trip his boredom he threw himself into writing
+he separated and saw each of the adventures of incan literant
+this is why it was used in several attacks which costs its popularity
+the cub appears frequently in roman architecture of some french regions
+philip had the time that we built only one
+he also proposes reflections on the ideal country
+on the church in the presence of the vacluse no majority obtained
+cosette all screamed out of us moving
+the evaluation of precursed clearly the purpose of the job
+this is a fear that he maintains a very high height
+the sentence is attached and receives the sentence is attached and receives the not on the back
+manajumdi was a poet an actavist during the algerian revolution
+this one named the same george campad first minister
+the multicular areum is endemic to leba
+thethree times in the holders alongside john pierre pappin and eric cantoner
+after the absorption of the kingdom after the absorption of the kingdom of brok became the capital of the terra
+his sister marie terescurges is a fiction producer and documentary
+it could be compared to the literary side to the west in rhetoric
+she attends john fittle kennedy junior
+it is with mino mario that he won his first and only championship
+he travels across the territory including mississippi and the apaleches
+student of richford he smiled under the influence of the master nazarean art
+he waved at the nails in the middle of a simple rain
+we can find several types of trees used for their quality
+it is based in dumkir with the maritime force of the north
+just after the price he scores porto
+rather reveal that the director the date of the execution is in
+during these schools sir jabron practices sinning
+the architect in chief is ducating at the origin of lusen
+no vestige still exists in the two levels and primitive covers
+all the uniforms and insignes of great stayed the same
+the execution of the work lasted almost ten minutes
+he mostly showed incomparable organization
+according to a spokesperson of the minister of agriculture is a poster of interest
+its headquarters are in warm suckets
+that's why he protects himself against his enemies with his boat
+however the snoop has indicated that it could be consoled in a future king
+in fact he is interested in poetry and theatre
+the orientation of bonapartism of the society of discent
+blood is one of the slovian florids
+the seated were not inactive
+stances and the causes of the accident are not for them
+he won the competition three times since
+the inhabitants of sa pierre de cormers are called the columbians
+she then returned to italy
+the crippy is the eldest of a family of seven children
+the restricturation of the network was already brought up higher
+their antagonists said lucitris are in terms of efficiency comparable to the sentistes
+his mother n haley works as a servant
+tuberconscrablate ends in cereals on the two faces
+their differ and more important facilitates the release of highways
+the panorama is exhibited in a cylindrical room in the centre welcomes the spectators
+crate on the right
+the product is the most important point of marquetting in industrial environment
+michel vuve has his bachelor's degree in the school of beniso
+he is the son of sam and sophi cress
+the series was created by willisner shenicolas and bob pow
+this polamorphism is undoubtedly unique inis kind
+it is this time quickly dissipated in the south it was due to the medical situation
+on the second floor chumake went to the east and the east
+the presence is transformed into a thousand tufrins
+the the deblien has then divided in the sub tributary of the paris
+what is your line of primedican
+it is the case of the failures and all their versions for the press
+hisse berham then the first president of the club
+charles quint named him painter of the cup before his death
+jun shereer is interested in muscle work and is limited
+i i have anything and i asked for it for me and it askd me
+other advantages the club will easily access to the belgian counter
+quickly after his arrival karibujin was killed and looked for him for index
+a degree in technology of the origin she holds a degree in commercial branches
+intrase boots flottered there is under the breast underwater base in france
+in the north of france the term corresponds to
+during the bombing of the building in front of the school is touched by a bomb
+it is also a winter site for the torier in beno
+the uti will prove that merild hitter was under the influence of a subikan
+she looked at each other laughing at each other
+he participated in ten times in the roral championship with norway
+he can he can he can exercise more consecutive demands
+after secondary studies in cameroon he studied scientific and studies
+renee le forestir is a french tester and a frenchman
+the seed of benford settled in france in blumbian court
+he is armed he is armed by jeniver
+some sources confirm that treser were inhabited during the period of the age of the bronze age
+and i know my wife
+other yellower than in sweden and south africa
+after being the ssistant a mench but after being thesistant a mench batasharia he became a cinema doctor
+rid rimen is used as an ordaining officer
+the old village still has the history of gendo
+zimmerman ended up accepting this moment the things are comp
+the author shows that he held some information about the domestic rago
+each personality is trained at the martial art of the netherlands
+she's located after the pierre lay station in precet sade station
+the lieutenant general catcar succeed him
+these champs and you are very vi relating
+he is also called in the belgian hopes for many matches
+the beliqes are played by women
+i in fell is importable
+floor floor mant is part of the dillis for worth's agglomeration
+he is named rector general and deputy director of the financial administration
+she remained since very attached to french culture
+unlike a cross with this kind of misarine is not venomous
+purled runson hipson
+he often performs on the sites of lacoplasi
+the fate the first number is about to be removed
+five candidates and stores will soon have left for three years
+he spent all his life and my life with his family until his father
+it's a failure that was fatal to his health
+he is also one of the five doctors of the country
+on six replied pencroft
+he is the son of the architect of verbe
+it is then that he bursts out to persecute the danish mandarins against the christians
+he is the son of the actors antonio and rus rosa
+goffrey could not promise to take his breath either
+the cape possession is from the point of youth
+hunskins is educated in winsminster school then at the royal military academy of sanest
+these processes include activities
+this also applies for the fifteenth of brazil's networks according to your studies
+the loss the loss of this child led to the separation of the telesabet a philipp
+this temple is at the temple kozesin in his life of yuki
+he underlined the massive death of the poor medical staff
+the store is a robtive of slate and unknown but it is considered potentially active
+i was bored and i was teasing it since the day before yesterday
+history takes place before the training of justice league
+it serves as a reonal centre assigned to the legion of foreign legion
+the manager of the consultant warinko
+the body is made of silk
+speakers employed the chakalo in the castoo
+for this reason many laws approximate were elaborated
+he is the eldest son of thinkerland and marie niputer
+it is foundi abundantly in literature and cinema
+the abbey of sharle is owned by fishing on conflints
+the cross is set up on a square under the stone of a cut stone
+the media relayed this information but no measure is nor photographic
+scabille mc coy on the grill of hell
+the decoration in her is based on geometric forms in green and in the inchro
+the information has been announced under the tutor of nilpatrick harris
+its capital is san pablo
+he sees crickon ever became the first of this new era
+it should be noted that the extremely endowed grafts do not check the lower made
+he is made in the pois municipality in the cot dor in france
+the interior is totally new with the console central and slightly sloping
+arm conflict then burst between the local population on the left and the milish of the seven
+he is also a key actor in musical education in canada
+menes ebe proposes an explanation different from the first one
+guy ousing is located north of the lake
+the programme was abandoned
+it was a traditional fishing boat at the convet with the catendls of english cot
+on the hills the agles come to their nests
+cuba and his again confronted with the academy of kora
+the tournament takes place between the country of walls argentina and japan
+the pilot steaman coming from this sqall
+he spoiled he spoiled the playoths for the fourth season of a row
+the buddha is estimated for four millions
+the draught is carried out according to the places to rent associations
+during their trips they have met pope in rome
+he is then replaced in the crowd by pats parfrey
+for several months the front remains globaly stable and battles are not intentive
+citizen be the cannot
+this species is named after this species is named after the woltetic nix
+ellen and her mother start her career at the radio channel
+he lives in the wet forests in the middle of the tropical and southen he lives in the wet forests in the middle of the tropical and sub tropical regions
+the champion is for the first time qualified for the french championship
+isidor of these eight contributed to maintain a certain grade of knowledge
+my petrism is still hereditary since the foundation of the setuary
+he is the great grandson of edgar's clem
+the capital holumbator is managed separately as a federal district
+the group remains based in los angeles for this activity
+its chief town is the city of basta
+with the gleveration of danish engineer george hudson the project tates over
+part of these soldiers was transferred to the body of castus
+it is ther not neglecting what the sea is about
+as for you by the scopia
+individual applications can now be displayed on a single map
+this species is found in the united states mexico and initalies
+this mission is entrusted to the nephews of king baloveso and sicovesos
+after taking over various fortunes the family fled to egypt
+one day she has a clerk crisis and her hair style
+several hind levels of bil shinda were added
+his maliciousness is at home
+the part is also known as the one in the mazapors
+the kopo is also a way to insure the own people
+but he is currently the sports director of the saint george merdar team
+it is the last edition
+i am not for anything to do with this for nothing
+anneson company who was bought by warnebros
+the public life also occupied felixine
+power was wrong
+this policy is not about napoleon combine the religious and diplomatic questions
+his daughter is married to a man of mo
+it lives in the clear forests
+in these pastorals he was looking to give a kind of ancient modern form
+the current population of killer is no longer than one and continues to decline
+above this temperature the materials are permitted
+the two actors played for free in the movie
+the hor was attacking him we'll finish his tragically
+then suddenly shrinker disappears and leave his group
+he is then a singer
+danny henders was born in mullen bake sat jean
+we are understanding the rope
+these barons who uses the first to the objective in the enisti context
+only my swiss had to be financed by the geneva kenton
+the aerport is located in downtown of juna
+the latter put a museum all the carden of the studio until his wife
+the air company is perpetuated in a constant search
+pierre le masy said the abbot who was going to be the one who was about to say
+nihol and the reporter had thought about him
+we will thus we will thus go to anuse about our own air of water
+less littles often accepts a lacor
+the old man embraced her life with a young girl named sonya
+the empire does not attack even more about fifty years in hasena
+however this rell seems to be the teather of strange events
+jack jack linder pray was converting to judaism during his trip
+she's registered in the inventory of cultural heritage
+he is the vitors another man said the chief of the catholic church in france
+vilet to was his father san
+he is a total mess of money at most
+this rush took the lead to the french army
+but the voice came from a studio singer in an unknown
+run to do it as for you to discover the paintings that have been in touch with an accentuity
+why gilia gave all these sentences
+still junior she won four world titles
+the party has been anti nationalist deology developed quickly in the south east of the country
+it was for the historians to determine the king's responsibility
+missus camuser had gone a continent with the helping marking point
+the chemist crisis is already made his first victims
+stapmen produces the last recording of tarus
+he takes his retirement from the national team after the german world war
+an intendant to rebild a partial rebuild it is to the procedo san mari that i am being the bishop
+tourism is highly developed
+he published many publications on the herpetiphone of his country
+he is lord to have july and then in october
+this species is partnogenic
+made prisoners he escapes and enters the distance
+as a result ginebus separates from the universal and returns to the underground stage
+it is found in sa domingos
+she realizes chart er flights for upper turns
+imprisoned or as a previous fighter allowed him to be freed
+they are pronounced and occupy the of their respective syllabaries
+this commission was launched and directed by london in person
+the battle of moscow began
+he served at the top of a granitic eminence dominating the city
+the subject is the pundrock
+the name comes from the presence of tenory in the surrounding area
+the business man is console in his mistress the business man is console in his mistress's arms
+the war continued again on eleven years and organized many reverses
+ser suvian was married to the actress marie louise morvin
+it is one of the biggest churches in nives
+he is repatriated and resumed his job as a steefer
+petersky has taken on his side the seventeenth place
+the operation of bakes is at the same time consistency to borrow
+from now on his shoes are exhibited in the western sides
+married three times he was the father in law of actress olivia white
+he made the corps leader
+this highway is about two gilometers permitted to serve the east of montroe
+he's up curly wounded
+he takes office as a deputy director and then director and a general assistant
+trummer prolifics and regularly produces high scores
+she was born in burk naturally his support in the oman succession
+casiky is a complex character
+bo boset clupin
+contacts between players are still authorized on the rinkaki
+we couldn't see who was bringing her
+sophie de la porte is also interested in transposing his photographic universan film
+this system should enable the deposit of new corpse in the grave
+he is registered in a catholic college
+this decision does not make an ana mistake among the orthodox churches
+the sale in particular was completely revent
+at the century was founded a priory which would know the place
+he is indeed very popular with civilians but does not know the army
+colonel busman is a family of sirg
+inspired by pass the script
+its names of caar alphabetical logic
+richard wigner will marry that these two men in claire de goatsmood
+pierre de pacy was the usual judge of n j
+it is a new one of thirty meters crossing the territory of the reverse
+in a few cases the parasite is waiting for the digestive and conotes it
+the orchestra is not very well known for the time that the piano orchestra is still a co
+proposed by six sections the museum exhibits one hundred twenty vehicles but quite little bit of motors
+the royal city as well as other cypress citizens are under the name persian domination
+the black space is deserted before it is in front of it
+it meets il morris and in reunion
+tukase seto is melake cabera are free from their tract
+he studied a lecturo technic at tarm stat and en works atm
+he currently plays for the hon kong he currently plays for the hon konga club of san peasius
+at the beginning of his political application his vision was modest
+he makes his debut in the players by obtaining two games
+the music programme is adapted to the target group
+she was rebuilt in the roco style of the style
+ursus improvised here become visible in the finance of the religion
+come on all the church said ursula while getting af
+rizzling tunnel his music is escaped at any spirit of his system
+the thing is proclaimed the call that is published
+also i explain a position in the creation
+the generation of leets corresponds to the performance that allows to produce leets
+he then spent a year in his fight
+the race is won by the runner frenchman matthew purget
+george poyser plays as a defender as a left back defender
+if only with three atoms of endrogen and two carbon atoms
+the theme approached by john harvey is the pet hofaly
+it's the first quarter to a single parent
+his mother was the sister of the economist david lancardo
+the versions of the circulate still run under the shape of beclain
+nikola kopolof is a doctor
+the detached effort must adapt to the progress of the time
+the municipality is managed by the municipality of onberg
+face this paradox frener is questioning the real nature of love
+they concentrated in they concentrated in viviers and martial cultures
+with the maternity in other dits the child is not far
+it is also during these years that he wrote her first psalms
+the republic multiplied the number of school meeters
+this set of transformation thus this set of transformation thus gives it its current chrysiformal plan
+whiht lives in the united states
+engley is an international and identifying the recruitment of black troops in louisiana
+the family then moved to europe
+it is also a very fertile trace
+she's the founder of the dutch an rylands
+the male field looks like a reminder version of the merle of america
+the first objective of the general was attenpted then
+this conception refers to the notorites notorites conception
+he has two pavilions one in the other one with the other
+this conception is part of the ptolemian conception of geocentricism
+a geracist is absent to participate in a cycling race
+no enamel diodraxs is also being made
+the municipality is crossed by the bessin
+born in a directiona and the family of mirambits
+sepicurean
+lucid is the property of the conservatory of the toral council
+he succeeded one another until the second world war
+stevens weg is known to the romance rule in germany
+his departure will be effective in september
+he starts studying clarnet at the school of music in new york
+his parents had a me business
+with the years others develop permission in the intinite
+but despite the success o is destined ballake lost her own cinema
+the summers are dry enough
+the municipality is part of the ranz mountain and is crossed by the artra
+this species is found in mexicos and nicaragua
+pero nationally for hungarian and trader for serfs
+among other talents john speak french plays piano and sings
+she wants to give the dupe but after leaving school she became a model
+the breeches did not have the most expensive
+it was built at the naval shipyard of apple dor in the north of eaven in the united kingdom
+he stopped at the port dor led by prince chender beder
+he was the sixth of the fourteen children of leuis from dur marriages
+dicdell then appears in haslevensire and in a few movies
+she's the mother of skeer alban clericrisova
+one hundred first director was august bella
+in other words all anteges are not in minogenic
+he appears at tulerd under the name of athery
+in addition the damages caused by the conference hall
+the other tenants re moved to the dead or had been expelled
+four vigorous men stayed in the rays of the governor's double rad
+the place attracts many visitors notably at the time of the great religious festivals
+the analysis of plant enabled identifying the main little one the argentina
+it is the sixth novel of the district police station
+this glass is often used in the dialogues of comedy and tragedy
+this battle ends with a new georgian defeat
+the travler building has six rows with the floor
+he was seated in great number of times
+this sentence provokes a strong controversy and no ell written in the letter of exclusives
+he used to delegates to negotiate a hersus
+the article of his application is a greater silmar tup
+it has two current species and one fossil
+the family wile then turnd to learn
diff --git a/asr_bleu_rm_silence/text.txt b/asr_bleu_rm_silence/text.txt
new file mode 100644
index 0000000000000000000000000000000000000000..048c92cebe7c83ce40167da12e8af94e8c95517a
--- /dev/null
+++ b/asr_bleu_rm_silence/text.txt
@@ -0,0 +1,200 @@
+a real work is going to end up this subject
+a meeting with the necessary order
+the weather pass
+the weather is mist inter monetary wharf with the indicat
+the woman is talking about the floor too mister alad lalan to support the amendment number one hundred an floor
+it is that so much you like jupe
+the weather is
+i have therefore experienced the past years i will tell myself later
+twelve mill
+it is in evidence
+you have excellent reading
+the weather is the first of the safety of berth
+the question of question is structural deficit coming back too
+cove you receive there budget from one point five bitinuros
+this is no doubt not just
+the artist is then mister debonn
+i hope you not part of mister d
+but i will only give one example
+that's mit matan deputy minister who different from ump
+no one wants it to be less numerous in the classes
+i am being chase the future
+finally the child we child do not know if the leader pay
+thirty eight as ana
+tede for your accounting dear collits
+it set for venison
+no one forgot to forgot the delusion of a chimmer
+my secondry is that all this is this is doing itself
+this is the dest is virtuous
+im bing referred to the first roll can be a good solution in some cases
+the weather is necessary to evacuate
+this is article l a motion missus massin
+in the wam of the government i now moving unto the wisdom
+the ministers incess working a group on the subject
+or else you are simplified
+i was pet of sescio another engine betcan
+patiently your so is apes soon finished
+they also be also became satisfied with measures that are taken
+unfavorable opinion
+we asun we assume contrary to you politics
+speak to businesses about it
+it was my friend
+that is the reason to from the c i c e o w
+civilization is poory old and very ancient
+workis between all group parliamentary group has been between
+the floor is now open to mister denis jacket registered
+that's mit it's why it's why
+the last one date the reporter gerrin has confirmed his in
+mister the minister is listening to the heads of companies listen to the crafts
+the weather will be a great deal of progress for its wackers
+coming back to children it's to say the da of mister lorand that is necessary
+i am being chaste the floor is now open to mister mohammed lakila the special repater
+the people know how to be how better recognized
+i said that i part of the system penamelity of two thousand and ten was repealed
+the spirit of the tabelet must be shin while he was anaf
+here we go to the post office as well as the collectivities taking over the responsibility of co ordinating and coordinating
+the weathers the dry women are competent to follow progress and practically in delivery
+tell you mister de clo
+from here too it poses a pose
+heris a lesson of economic
+borrowing from from
+a word on the reform of retirements
+i remember the decreation of the c e s here in this heme
+i think i a major
+article l is a good article
+the action of the vaction is written in a framework a freely ensued
+he is the father of the city of you s a dear colleague to support it
+my father has to be empty i'll go phil before going on
+this translates herself into exile by a decrease in persia
+i am thinking she is actually improving the dye complexity of the text
+at his age i ge i was already brought
+we are having found a deficiency at five point one per cent
+you would give me mister chair that the german are better than our best
+this is exactly true
+it is political to have disastro effects for modest house holts
+the era prehistoric era is the day beginning of computing and is considered complicated
+no measureis in ward regard these matters in this two thousand fourteen
+furthermore i only only remember the scent to tax the bee
+chis the donning the propositions sevena
+here we are the generations of new generations two to not have to undergo four types
+before it this not up to us we were keep our time
+the floor is now open to
+no mister knowles
+i am bing you chased your second joe was in jockey office and in the box
+in a way beautiful anin because these topics are telling you as essential
+he is withtrawn and the withdrawn voted against it it's the other identical
+here we are change the change of
+the order of the voy calls for government to ask the government
+it is this unfad trashing that we have to replace a virtuous circle
+ten villani you refuse to hear it until the constitutional council
+i am being
+i donned see any difference in difference unde fundamental approach between these two request
+the one poos are arriving will arrive in the next few days
+at meve of the configuration this one passes by ad
+the weather is specially specialists
+the woman is toning about it is mister don jock candalier
+here we are your amendment missus masinea is therefore a question of question this new article
+i would lie to with two or three
+i therefore believe that that text will cause a very strong break
+but sometimes on the burry
+four letters to five letters coins
+a combination of the conversation of the members taken to the article eleven
+he has the floor to speak in order
+thus using is in this box and the following
+ninety ten gram
+unfavorable opinion even though it is not the by i share your opinion
+bubble
+the floor is now open to a lun for it
+i am in very unserved
+i am being
+it said before you
+the floor is now open to missus christ to support missus cristel dubos to support the three hundred eighty
+wat a dadter mind was mind keep the mothers
+tall and big but didn't chase it
+the weather is yesterday and to day a emotion
+the weather is a contract that we go with a different fellow citizens
+we are
+we have to watch what she is general each time that it is possible
+the weather was horrible to pass through no
+we can nowadays receiving tax revenues no longer come back
+the regulation plans for two minutes for the intervention
+the deficit of the state itself has passed by sixty three to seventy two bvinros
+here we are mister president i will mister president i will be the only one for this article
+the weather is going to have the same way my dear callits to get get to budgetary discussions
+if some artists are in favour of a favourable others opinion others apply in an unfat
+first against truth no one one of the engagements would have public finances was elected
+we are talking
+it said pretty and good to say that
+it is for it is working in the whole matter
+descendants in thirty
+it is thus the key
+you are here for ten years
+this article d is a regression
+we move the balance sheet
+batusi you are not not mister terris
+the one pos are talking about the young since you're very young that you dear call its from the position
+i am not want but yes yes
+i am being
+the weather is nice in other we have in days we haven't been informed of deputies
+its sat it's wrong it's a shame of tax pressure for families
+mister hamash was mister others in california in the hot water
+regarding the regard catilizations we could call them an citizen
+yes this device is absolutely new in the case or on the case
+the woman's court you
+the weather is
+the weather is you
+i suggest you too reassure you and listen to mister deputy minister's question
+exactly true at holmes callit damien
+it is necessary to steal
+the weather is now open to mister christian paul
+but on the bust when we left missing the taxes to protect the most modest ones
+the devices incriminated and plans that the delegate state of one of its competences
+his fests indeed cause
+the government is understand
+these are well here to design are different
+the floor is now open to mister shock lenlen
+i am being
+the weather is nice in the weather
+he was lost
+it was a science very neglected
+the street is named after him missus batho at the time
+carrier has been been
+the woman from bordeaux you really know how to bordeaux
+it is michael has the common expert in petank he almost hits the jack
+paris wasn't sir he doesn't deserve the paris
+the latter is i would like to say a few with two territories
+he is about an amendment of coherence and the system of global
+we therefore cut the got to receive people mister mister
+takes em
+after that date chief advertising that i really heard i didn't want to my amendment
+that being said it being said an academic date and practical indeed mister cullus
+recognize yours at least
+on the spack we back all agree to adopt such action
+we even telling us that the textile spirit requires it
+here is the result of the work we led with his
+in no case what we propose to recognize the solidarity tax
+come on let let's mister martyr finally
+i would wike i would like us our rejection of this in inefficient measure
+he is leaving the room for a while that is therefore quite
+the weather is to uption were a date
+the curofa is the floor to madam president to the right of women
+i am being chased the floor is now open to mister sass
+caltin hath is economic in culture stilicology acology a lot of essential structures will be affected
+the weather is part of the
+the one you enjoy this development but you don't know much
+such a true story is actually
+the weather is my son and the best player scrable kisa
+and this emergency was due to deficits we inherited
+chiis two desperate from the opposition
+he is nose middle aged without speaking
+on coming to the amendment number six hundred seventy four
+you are in the condition permanently good understanding is in the condition
+manette was mad taller than me
+we wall then if if we will
+herisa reality a question
+missus macinea is missus macineau the amendment maintained
+the woman from pordo is right for which one is offering this amendment which has two objectives
+the weather is nice and mixed many of the earthquake although next to it
+its said that's normal that we understand it's about fourteen years of doing it
+we wall ben have te time
+i am bing referred to a sequence of continuous decrease in decrease
diff --git a/asr_bleu_rm_silence/utils.py b/asr_bleu_rm_silence/utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..97a4c98963d872c22a8b5934980dee047a6a0ae7
--- /dev/null
+++ b/asr_bleu_rm_silence/utils.py
@@ -0,0 +1,318 @@
+import json
+import re
+import urllib.request
+from pathlib import Path
+
+import fairseq
+import torch
+from fairseq.data.data_utils import lengths_to_padding_mask
+from tqdm import tqdm
+
+try:
+ import torchaudio
+ from torchaudio.models.decoder import ctc_decoder
+except ImportError:
+ raise ImportError("Upgrade torchaudio to 0.12 to enable CTC decoding")
+
+
+class DownloadProgressBar(tqdm):
+ """A class to represent a download progress bar"""
+
+ def update_to(self, b=1, bsize=1, tsize=None) -> None:
+ """
+ Update the download progress
+ """
+ if tsize is not None:
+ self.total = tsize
+ self.update(b * bsize - self.n)
+
+
+def retrieve_asr_config(lang_key: str, asr_version: str, json_path: str) -> dict:
+ """
+ Retrieve the asr model configs
+
+ Args:
+ lang_key: the lanuage type as the key name
+ json_path: the path of the config json file
+
+ Returns:
+ Dict of all the configs in the json file
+ """
+
+ with open(json_path, "r") as f:
+ asr_model_cfgs = json.load(f)
+ return asr_model_cfgs[lang_key][asr_version]
+
+
+class ASRGenerator(object):
+ """A class to represent a ASR generator"""
+
+ def __init__(
+ self,
+ model_cfg: dict,
+ cache_dirpath: str = (Path.home() / ".cache" / "ust_asr").as_posix(),
+ ) -> None:
+ """
+ Construct all the necessary attributes of the ASRGenerator class
+
+ Args:
+ model_cfg: the dict of the asr model config
+ cache_dirpath: the default cache path is "Path.home()/.cache/ust_asr"
+ """
+
+ self.cache_dirpath = Path(cache_dirpath) / model_cfg["lang"]
+ self.model_cfg = model_cfg
+
+ self.use_cuda = torch.cuda.is_available()
+
+ torchaudio.set_audio_backend("sox_io")
+
+ if self.model_cfg["model_type"] == "hf":
+ self.prepare_hf_model(self.model_cfg)
+ elif self.model_cfg["model_type"] == "fairseq":
+ self.prepare_fairseq_model(self.model_cfg)
+ else:
+ raise NotImplementedError(
+ f"Model type {self.model_cfg['model_type']} is not supported"
+ )
+
+ if self.model_cfg["post_process"] == "collapse":
+ self.post_process_fn = lambda hypo: "".join(hypo).replace(
+ self.sil_token, " "
+ )
+ elif self.model_cfg["post_process"] == "none":
+ self.post_process_fn = lambda hypo: " ".join(hypo).replace(
+ self.sil_token, " "
+ )
+ else:
+ raise NotImplementedError
+
+ if self.use_cuda:
+ self.model.cuda()
+ self.model.eval()
+
+ self.decoder = ctc_decoder(
+ lexicon=None,
+ tokens=self.tokens,
+ lm=None,
+ nbest=1,
+ beam_size=1,
+ beam_size_token=None,
+ lm_weight=0.0,
+ word_score=0.0,
+ unk_score=float("-inf"),
+ sil_token=self.sil_token,
+ sil_score=0.0,
+ log_add=False,
+ blank_token=self.blank_token,
+ )
+
+ def prepare_hf_model(self, model_cfg: dict) -> None:
+ """
+ Prepare the huggingface asr model
+
+ Args:
+ model_cfg: dict with the relevant ASR config
+ """
+
+ def infer_silence_token(vocab: list):
+ """
+ Different HF checkpoints have different notion of silence token
+ such as | or " " (space)
+ Important: when adding new HF asr model in, check what silence token it uses
+ """
+ if "|" in vocab:
+ return "|"
+ elif " " in vocab:
+ return " "
+ else:
+ raise RuntimeError("Silence token is not found in the vocabulary")
+
+ try:
+ from transformers import (AutoFeatureExtractor, AutoTokenizer,
+ Wav2Vec2ForCTC, Wav2Vec2Processor)
+ except ImportError:
+ raise ImportError("Install transformers to load HF wav2vec model")
+
+ model_path = model_cfg["model_path"]
+ self.model = Wav2Vec2ForCTC.from_pretrained(model_path)
+ self.tokenizer = AutoTokenizer.from_pretrained(model_path)
+ self.preprocessor = AutoFeatureExtractor.from_pretrained(model_path)
+ self.processor = Wav2Vec2Processor.from_pretrained(model_path)
+
+ # extra unk tokens are there to make some models work e.g. Finnish ASR has some vocab issue
+ vocab_list = [
+ self.tokenizer.decoder.get(i, f"{self.tokenizer.unk_token}1")
+ for i in range(self.tokenizer.vocab_size)
+ ]
+
+ self.sampling_rate = self.preprocessor.sampling_rate
+ self.normalize_input = self.preprocessor.do_normalize
+ self.tokens = vocab_list
+ self.sil_token = infer_silence_token(vocab_list)
+ self.blank_token = self.tokenizer.pad_token
+
+ def prepare_fairseq_model(self, model_cfg: dict) -> None:
+ """
+ Prepare the fairseq asr model
+
+ Args:
+ model_cfg: the specific model config dict must have: (1) ckpt_path, (2) dict_path
+ """
+
+ def download_file(url: str, cache_dir: Path):
+ download_path = cache_dir / url.split("/")[-1]
+ if not (cache_dir / url.split("/")[-1]).exists():
+ with DownloadProgressBar(
+ unit="B", unit_scale=True, miniters=1, desc=url.split("/")[-1]
+ ) as t:
+ cache_dir.mkdir(parents=True, exist_ok=True)
+ urllib.request.urlretrieve(
+ url, filename=download_path.as_posix(), reporthook=t.update_to
+ )
+ else:
+ print(f"'{url}' exists in {cache_dir}")
+
+ return download_path.as_posix()
+
+ try:
+ ckpt_path = model_cfg["ckpt_path"]
+ dict_path = model_cfg["dict_path"]
+ except KeyError:
+ raise KeyError(
+ "Fairseq model cfg must provide (1) ckpt_path, (2) dict_path"
+ )
+
+ if re.search("^https", ckpt_path):
+ ckpt_path = download_file(ckpt_path, self.cache_dirpath)
+ if re.search("^https", dict_path):
+ dict_path = download_file(dict_path, self.cache_dirpath)
+
+ model, saved_cfg, _ = fairseq.checkpoint_utils.load_model_ensemble_and_task(
+ [ckpt_path],
+ arg_overrides={
+ "task": "audio_finetuning",
+ "data": self.cache_dirpath.as_posix(),
+ }, # data must have dict in it
+ )
+
+ dict_lines = open(dict_path, "r").readlines()
+ tokens = [l.split()[0] for l in dict_lines]
+ # adding default fairseq special tokens
+ tokens = ["", "", "", ""] + tokens
+
+ self.model = model[0]
+ self.tokens = tokens
+
+ if "|" in tokens:
+ self.sil_token = "|"
+ else:
+ self.sil_token = tokens[
+ 2
+ ] # use eos as silence token if | not presented e.g., Hok ASR model
+ print(f"Inferring silence token from the dict: {self.sil_token}")
+ self.blank_token = self.tokens[0]
+
+ self.sampling_rate = saved_cfg.task.sample_rate
+ self.normalize_input = saved_cfg.task.normalize
+
+ @torch.inference_mode()
+ def load_audiofile(self, audio_path: str) -> torch.Tensor:
+ """
+ Load the audio files and apply resampling and normalizaion
+
+ Args:
+ audio_path: the audio file path
+
+ Returns:
+ audio_waveform: the audio waveform as a torch.Tensor object
+ """
+
+ audio_waveform, sampling_rate = torchaudio.load(audio_path)
+
+ # remove simuleval silence with rules
+ left=(audio_waveform == 0).contiguous()
+ right=(audio_waveform == 0).contiguous()
+ for i in range(10):
+ left=left&(torch.roll(audio_waveform, shifts=-1*i, dims=-1) == 0)
+ right=right&(torch.roll(audio_waveform, shifts=i, dims=-1) == 0)
+
+ mask = (audio_waveform == 0) & (left| right)
+ audio_waveform=audio_waveform[:,~mask[0]]
+
+
+ if audio_waveform.dim == 2:
+ audio_waveform = audio_waveform.mean(-1)
+ if self.sampling_rate != sampling_rate:
+ audio_waveform = torchaudio.functional.resample(
+ audio_waveform, sampling_rate, self.sampling_rate
+ )
+ if self.normalize_input:
+ # following fairseq raw audio dataset
+ audio_waveform = torch.nn.functional.layer_norm(
+ audio_waveform, audio_waveform.shape
+ )
+
+ return audio_waveform
+
+ @torch.inference_mode()
+ def compute_emissions(self, audio_input: torch.Tensor) -> torch.Tensor:
+ """
+ Compute the emissions for either fairseq or huggingface asr model
+
+ Args:
+ audio_path: the input audio waveform
+
+ Returns:
+ emissions: the logits of the encoded prediction.
+ """
+
+ if self.use_cuda:
+ audio_input = audio_input.to("cuda")
+ if isinstance(self.model, fairseq.models.wav2vec.wav2vec2_asr.Wav2VecCtc):
+ padding_mask = lengths_to_padding_mask(torch.tensor([audio_input.numel()]))
+ emissions = self.model.w2v_encoder(audio_input, padding_mask)[
+ "encoder_out"
+ ].transpose(0, 1)
+ else:
+ emissions = self.model(audio_input).logits
+
+ return emissions
+
+ def decode_emissions(self, emissions: torch.Tensor) -> str:
+ """
+ Decode the emissions and apply post process functions
+
+ Args:
+ emissions: the input Tensor object
+
+ Returns:
+ hypo: the str as the decoded transcriptions
+ """
+
+ emissions = emissions.cpu()
+ results = self.decoder(emissions)
+
+ # assuming the lexicon-free decoder and working with tokens
+ hypo = self.decoder.idxs_to_tokens(results[0][0].tokens)
+ hypo = self.post_process_fn(hypo)
+
+ return hypo
+
+ def transcribe_audiofile(self, audio_path: str, lower=True) -> str:
+ """
+ Transcribe the audio into string
+
+ Args:
+ audio_path: the input audio waveform
+ lower: the case of the transcriptions with lowercase as the default
+
+ Returns:
+ hypo: the transcription result
+ """
+
+ asr_input = self.load_audiofile(audio_path)
+ emissions = self.compute_emissions(asr_input)
+ hypo = self.decode_emissions(emissions)
+
+ return hypo.strip().lower() if lower else hypo.strip()
diff --git a/assets/asr.png b/assets/asr.png
new file mode 100644
index 0000000000000000000000000000000000000000..63cf93bc8ee34e8299ae219777c646fc2a73b286
--- /dev/null
+++ b/assets/asr.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c0ae79dfcb777ab5b8348490d0f0ceeafe1b3e0f11805ff4494b03c83f0698e0
+size 112337
diff --git a/assets/model.png b/assets/model.png
new file mode 100644
index 0000000000000000000000000000000000000000..3c91e8abc388a86eb0effc97ffe7e0487a130884
--- /dev/null
+++ b/assets/model.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3b63c7cf2e7a8cf94493bc1a6aabf3d79001413aa901484ecd2f696f2da2ec9c
+size 224971
diff --git a/assets/offline_results.png b/assets/offline_results.png
new file mode 100644
index 0000000000000000000000000000000000000000..c2edd0eac195fea2cdda54b823b660b6ab146765
--- /dev/null
+++ b/assets/offline_results.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7956af71c0c84a4e55b45fdd103af59489d4d41c8653e02f5e7b3960918fc356
+size 147888
diff --git a/assets/s2tt.png b/assets/s2tt.png
new file mode 100644
index 0000000000000000000000000000000000000000..da6dec8efbf33692e1e6308cd955f7cb8d28d24f
--- /dev/null
+++ b/assets/s2tt.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:711c6da12a61b57ddc13b1613cec80c5e852b3c95fef624d0635d6f2736ac12e
+size 90410
diff --git a/assets/simultaneous_results.png b/assets/simultaneous_results.png
new file mode 100644
index 0000000000000000000000000000000000000000..f70fbf4f5b9e3881b94949d20d9b32f271910d83
--- /dev/null
+++ b/assets/simultaneous_results.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1b0a10d77e7efb809c4db44b957114dcb0aa1e303b56097707bf175423fdb3e6
+size 273466
diff --git a/assets/streamspeech.png b/assets/streamspeech.png
new file mode 100644
index 0000000000000000000000000000000000000000..885e343da954e6f8702a11559c9cf3ce317a7cea
--- /dev/null
+++ b/assets/streamspeech.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4c62bbac306488bcfd49d8c7c2ec312125be8a293318afbc57d28c81765bf16a
+size 138816
diff --git a/configs/de-en/config_gcmvn.yaml b/configs/de-en/config_gcmvn.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..32bd18166d7b4a65521bf3ce399a5b2e1779990f
--- /dev/null
+++ b/configs/de-en/config_gcmvn.yaml
@@ -0,0 +1,21 @@
+global_cmvn:
+ stats_npz_path: /data/zhangshaolei/StreamSpeech/configs/de-en/gcmvn.npz
+input_channels: 1
+input_feat_per_channel: 80
+specaugment:
+ freq_mask_F: 27
+ freq_mask_N: 1
+ time_mask_N: 1
+ time_mask_T: 100
+ time_mask_p: 1.0
+ time_wrap_W: 0
+transforms:
+ '*':
+ - global_cmvn
+ _train:
+ - global_cmvn
+ - specaugment
+vocoder:
+ checkpoint: /data/zhangshaolei/pretrain_models/unit-based_HiFi-GAN_vocoder/mHuBERT.layer11.km1000.en/g_00500000
+ config: /data/zhangshaolei/pretrain_models/unit-based_HiFi-GAN_vocoder/mHuBERT.layer11.km1000.en/config.json
+ type: code_hifigan
diff --git a/configs/de-en/config_mtl_asr_st_ctcst.yaml b/configs/de-en/config_mtl_asr_st_ctcst.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..56b68bbfde1c7fea82adf529ea8ffbd851c4316f
--- /dev/null
+++ b/configs/de-en/config_mtl_asr_st_ctcst.yaml
@@ -0,0 +1,36 @@
+target_unigram:
+ decoder_type: transformer
+ dict: /data/zhangshaolei/StreamSpeech/configs/de-en/tgt_unigram6000/spm_unigram_de.txt
+ data: /data/zhangshaolei/StreamSpeech/configs/de-en/tgt_unigram6000
+ loss_weight: 8.0
+ rdrop_alpha: 0.0
+ decoder_args:
+ decoder_layers: 4
+ decoder_embed_dim: 512
+ decoder_ffn_embed_dim: 2048
+ decoder_attention_heads: 8
+ label_smoothing: 0.1
+source_unigram:
+ decoder_type: ctc
+ dict: /data/zhangshaolei/StreamSpeech/configs/de-en/src_unigram6000/spm_unigram_de.txt
+ data: /data/zhangshaolei/StreamSpeech/configs/de-en/src_unigram6000
+ loss_weight: 4.0
+ rdrop_alpha: 0.0
+ decoder_args:
+ decoder_layers: 0
+ decoder_embed_dim: 512
+ decoder_ffn_embed_dim: 2048
+ decoder_attention_heads: 8
+ label_smoothing: 0.1
+ctc_target_unigram:
+ decoder_type: ctc
+ dict: /data/zhangshaolei/StreamSpeech/configs/de-en/tgt_unigram6000/spm_unigram_de.txt
+ data: /data/zhangshaolei/StreamSpeech/configs/de-en/tgt_unigram6000
+ loss_weight: 4.0
+ rdrop_alpha: 0.0
+ decoder_args:
+ decoder_layers: 0
+ decoder_embed_dim: 512
+ decoder_ffn_embed_dim: 2048
+ decoder_attention_heads: 8
+ label_smoothing: 0.1
\ No newline at end of file
diff --git a/configs/de-en/config_unity.yaml b/configs/de-en/config_unity.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..146bbdc275dcc146a30673799f902bc39717741c
--- /dev/null
+++ b/configs/de-en/config_unity.yaml
@@ -0,0 +1,12 @@
+target_unigram:
+ decoder_type: transformer
+ dict: /data/zhangshaolei/StreamSpeech/configs/de-en/tgt_unigram6000/spm_unigram_de.txt
+ data: /data/zhangshaolei/StreamSpeech/configs/de-en/tgt_unigram6000
+ loss_weight: 8.0
+ rdrop_alpha: 0.0
+ decoder_args:
+ decoder_layers: 4
+ decoder_embed_dim: 512
+ decoder_ffn_embed_dim: 2048
+ decoder_attention_heads: 8
+ label_smoothing: 0.1
\ No newline at end of file
diff --git a/configs/de-en/gcmvn.npz b/configs/de-en/gcmvn.npz
new file mode 100644
index 0000000000000000000000000000000000000000..830fa8b04c13827b954899b60d2678a393022b70
--- /dev/null
+++ b/configs/de-en/gcmvn.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e01eacdb51f6f1a461944c12c09d905347c1a4edab7d3cc53d462315b0b82e7c
+size 1140
diff --git a/configs/de-en/src_unigram6000/spm_unigram_de.model b/configs/de-en/src_unigram6000/spm_unigram_de.model
new file mode 100644
index 0000000000000000000000000000000000000000..2b212f12cb913ede72279477971c2e57170841c2
--- /dev/null
+++ b/configs/de-en/src_unigram6000/spm_unigram_de.model
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8c5dc18c8ff2989a95a645806c656355611b56f14afb7a8fdc808c086d3556d3
+size 342959
diff --git a/configs/de-en/src_unigram6000/spm_unigram_de.txt b/configs/de-en/src_unigram6000/spm_unigram_de.txt
new file mode 100644
index 0000000000000000000000000000000000000000..e9933df99bbfcc9b8feb657040ab3d675edc6050
--- /dev/null
+++ b/configs/de-en/src_unigram6000/spm_unigram_de.txt
@@ -0,0 +1,5996 @@
+en 1
+e 1
+s 1
+▁ 1
+n 1
+▁die 1
+▁der 1
+▁in 1
+t 1
+▁und 1
+▁ist 1
+er 1
+▁das 1
+▁er 1
+▁ein 1
+▁zu 1
+▁sie 1
+▁den 1
+▁auf 1
+r 1
+▁sich 1
+▁es 1
+▁nicht 1
+▁mit 1
+▁an 1
+▁von 1
+▁eine 1
+▁im 1
+▁aus 1
+▁als 1
+▁des 1
+m 1
+▁auch 1
+▁dem 1
+▁war 1
+▁für 1
+▁ich 1
+▁man 1
+▁nach 1
+▁sind 1
+▁wurde 1
+es 1
+te 1
+▁hat 1
+ung 1
+▁wird 1
+▁so 1
+▁werden 1
+st 1
+▁be 1
+▁vor 1
+▁ver 1
+▁noch 1
+▁du 1
+▁wie 1
+▁ge 1
+k 1
+d 1
+▁ab 1
+▁bei 1
+▁über 1
+in 1
+▁um 1
+▁seine 1
+▁durch 1
+▁kann 1
+▁einen 1
+▁wir 1
+▁diese 1
+p 1
+▁nur 1
+y 1
+ä 1
+▁sein 1
+h 1
+▁aber 1
+g 1
+▁am 1
+▁zum 1
+u 1
+▁ihr 1
+▁unter 1
+b 1
+o 1
+▁einer 1
+l 1
+▁oder 1
+▁was 1
+▁einem 1
+f 1
+a 1
+ge 1
+▁mal 1
+▁wenn 1
+z 1
+▁zwei 1
+de 1
+ig 1
+al 1
+▁haben 1
+ü 1
+▁wieder 1
+▁zur 1
+▁schon 1
+ch 1
+▁hier 1
+ten 1
+▁da 1
+el 1
+▁dass 1
+an 1
+sch 1
+le 1
+lich 1
+ck 1
+▁zeit 1
+▁dieser 1
+▁gibt 1
+▁mehr 1
+▁ihre 1
+ter 1
+▁muss 1
+be 1
+ö 1
+▁mir 1
+▁immer 1
+re 1
+ur 1
+▁alle 1
+ischen 1
+▁bis 1
+i 1
+▁sch 1
+is 1
+▁dann 1
+w 1
+et 1
+▁heute 1
+▁wurden 1
+▁keine 1
+▁können 1
+▁seiner 1
+▁k 1
+iert 1
+us 1
+dem 1
+▁doch 1
+den 1
+▁re 1
+der 1
+▁ja 1
+ische 1
+▁viel 1
+li 1
+▁stadt 1
+▁s 1
+▁drei 1
+▁dabei 1
+ll 1
+at 1
+▁jedoch 1
+▁hatte 1
+▁un 1
+or 1
+ver 1
+vom 1
+▁mich 1
+▁sehr 1
+▁später 1
+em 1
+chen 1
+zeit 1
+▁teil 1
+it 1
+▁ihm 1
+▁wo 1
+▁a 1
+ar 1
+▁jetzt 1
+ra 1
+▁dort 1
+▁sollte 1
+▁kein 1
+▁etwas 1
+▁liegt 1
+il 1
+▁gut 1
+ieren 1
+▁viele 1
+um 1
+▁einfach 1
+▁damit 1
+▁beim 1
+la 1
+on 1
+▁habe 1
+ungen 1
+▁selbst 1
+ie 1
+ri 1
+▁gegen 1
+▁geht 1
+ut 1
+▁waren 1
+▁vier 1
+▁la 1
+c 1
+isch 1
+▁bitte 1
+ler 1
+gel 1
+bau 1
+▁außer 1
+na 1
+▁kommt 1
+▁uns 1
+▁ihn 1
+land 1
+▁ma 1
+sten 1
+▁mein 1
+lo 1
+kt 1
+▁wer 1
+▁zurück 1
+gen 1
+zu 1
+he 1
+▁ohne 1
+ma 1
+bar 1
+▁bekannt 1
+▁ra 1
+▁ob 1
+▁kurz 1
+▁frei 1
+x 1
+di 1
+▁weiter 1
+▁soll 1
+au 1
+ro 1
+tra 1
+▁c 1
+ing 1
+bel 1
+▁de 1
+▁alles 1
+▁konnte 1
+halb 1
+▁andere 1
+▁le 1
+▁neu 1
+ation 1
+ka 1
+▁wasser 1
+▁ganz 1
+▁diesem 1
+▁will 1
+▁seinen 1
+mann 1
+▁anderen 1
+▁zusammen 1
+▁ersten 1
+▁b 1
+▁steht 1
+▁neben 1
+▁dir 1
+▁erst 1
+▁einige 1
+ul 1
+▁hin 1
+▁bau 1
+▁nichts 1
+▁ent 1
+▁ha 1
+▁erste 1
+▁gleich 1
+▁jahr 1
+me 1
+▁film 1
+▁nun 1
+▁weiß 1
+hl 1
+▁kinder 1
+▁he 1
+▁neue 1
+end 1
+ken 1
+▁macht 1
+as 1
+heit 1
+stand 1
+ien 1
+▁mehrere 1
+to 1
+▁fünf 1
+▁ro 1
+▁her 1
+cht 1
+▁zwischen 1
+▁dich 1
+▁na 1
+▁dazu 1
+berg 1
+▁land 1
+ta 1
+igen 1
+▁pro 1
+▁hoch 1
+ste 1
+▁li 1
+ben 1
+▁statt 1
+▁kam 1
+▁g 1
+▁ihrer 1
+vi 1
+ner 1
+▁oft 1
+▁bereits 1
+▁gehört 1
+da 1
+▁meine 1
+▁lassen 1
+▁weg 1
+pp 1
+▁befindet 1
+ha 1
+bi 1
+ab 1
+am 1
+▁fl 1
+▁weitere 1
+▁ober 1
+▁während 1
+ik 1
+go 1
+▁kon 1
+▁gemeinde 1
+▁jahre 1
+▁herr 1
+▁leben 1
+▁bo 1
+▁ta 1
+rat 1
+hr 1
+▁lange 1
+recht 1
+los 1
+ige 1
+▁frau 1
+▁ort 1
+che 1
+ff 1
+▁w 1
+▁menschen 1
+▁wäre 1
+▁lässt 1
+gt 1
+▁se 1
+▁welt 1
+pa 1
+▁seit 1
+▁schnell 1
+▁ka 1
+▁kommen 1
+▁danach 1
+▁z 1
+burg 1
+▁bin 1
+▁spiel 1
+▁würde 1
+▁gab 1
+ko 1
+zi 1
+▁nie 1
+un 1
+liche 1
+▁grund 1
+ant 1
+▁machen 1
+ster 1
+▁beiden 1
+▁also 1
+▁eigentlich 1
+ke 1
+keit 1
+lichen 1
+▁ende 1
+▁genau 1
+▁kannst 1
+werk 1
+▁gerade 1
+fall 1
+▁po 1
+ger 1
+gang 1
+▁jede 1
+▁gar 1
+id 1
+th 1
+co 1
+▁gr 1
+ber 1
+ga 1
+mit 1
+nd 1
+son 1
+form 1
+spiel 1
+▁welche 1
+▁h 1
+▁haupt 1
+▁jeder 1
+▁ins 1
+ton 1
+ol 1
+▁besonder 1
+▁euch 1
+▁acht 1
+▁gilt 1
+ich 1
+▁je 1
+v 1
+fer 1
+tag 1
+▁allem 1
+▁sp 1
+▁allerdings 1
+▁fest 1
+▁zahlreiche 1
+▁müssen 1
+wa 1
+▁hast 1
+▁fa 1
+▁stehen 1
+▁möglich 1
+▁deine 1
+und 1
+ad 1
+ei 1
+▁solche 1
+elt 1
+▁mittel 1
+schaft 1
+▁darf 1
+▁me 1
+ion 1
+nächst 1
+▁o 1
+▁ihnen 1
+tz 1
+▁einmal 1
+▁schwer 1
+▁ho 1
+▁al 1
+▁platz 1
+▁jahren 1
+stein 1
+▁sei 1
+▁p 1
+▁groß 1
+teil 1
+ein 1
+▁sehen 1
+▁blieb 1
+▁ne 1
+ti 1
+▁voll 1
+▁lang 1
+ia 1
+▁f 1
+▁meist 1
+▁musik 1
+▁sowie 1
+▁besser 1
+▁möchte 1
+▁beide 1
+▁hinter 1
+▁art 1
+haus 1
+bo 1
+▁natürlich 1
+mo 1
+▁sohn 1
+▁stark 1
+▁mar 1
+pe 1
+▁deutschland 1
+▁tag 1
+▁verschiedene 1
+▁könnte 1
+▁vater 1
+os 1
+▁ging 1
+▁namen 1
+▁hand 1
+▁schön 1
+▁berlin 1
+ling 1
+▁luft 1
+▁ri 1
+▁we 1
+ß 1
+▁davon 1
+ic 1
+ende 1
+▁familie 1
+ierung 1
+im 1
+fe 1
+stadt 1
+▁bi 1
+sa 1
+▁deutschen 1
+▁form 1
+tisch 1
+her 1
+▁große 1
+▁etwa 1
+▁mo 1
+▁wohl 1
+ier 1
+ium 1
+bl 1
+▁lo 1
+▁sa 1
+▁bundes 1
+▁recht 1
+zen 1
+tor 1
+▁dafür 1
+ierte 1
+kommen 1
+ir 1
+▁ausge 1
+ph 1
+▁sechs 1
+▁kr 1
+▁pa 1
+▁weit 1
+▁weil 1
+▁früh 1
+▁gehen 1
+▁meisten 1
+do 1
+▁studierte 1
+▁national 1
+▁universität 1
+▁rot 1
+▁jemand 1
+▁zweiten 1
+▁ar 1
+▁häufig 1
+▁haus 1
+zug 1
+platz 1
+acht 1
+va 1
+tion 1
+ru 1
+ft 1
+ib 1
+▁krieg 1
+ku 1
+gebiet 1
+wi 1
+ze 1
+lin 1
+enden 1
+lauf 1
+▁pi 1
+▁vi 1
+▁besteht 1
+▁di 1
+▁wegen 1
+▁ru 1
+weg 1
+▁su 1
+▁verwendet 1
+aus 1
+iv 1
+▁ebenfalls 1
+▁findet 1
+▁warum 1
+▁fast 1
+▁halb 1
+▁band 1
+▁ange 1
+▁wollte 1
+ot 1
+ment 1
+▁lass 1
+▁ca 1
+▁ba 1
+gehen 1
+▁unsere 1
+kel 1
+tal 1
+lang 1
+si 1
+▁wenig 1
+▁erhielt 1
+▁berg 1
+▁bezeichnet 1
+▁sitz 1
+hi 1
+▁mitglied 1
+▁schul 1
+▁geschichte 1
+▁sieht 1
+▁arbeit 1
+▁hu 1
+mer 1
+▁benannt 1
+▁klein 1
+ine 1
+▁darauf 1
+▁ste 1
+▁natur 1
+ell 1
+feld 1
+mal 1
+▁zehn 1
+▁kopf 1
+cher 1
+▁all 1
+▁lebt 1
+▁diesen 1
+▁gebiet 1
+bahn 1
+ja 1
+▁paar 1
+man 1
+▁landes 1
+auf 1
+po 1
+▁eigene 1
+iz 1
+so 1
+sel 1
+lage 1
+▁vielleicht 1
+▁erfolg 1
+▁daher 1
+▁bu 1
+▁ex 1
+ick 1
+▁wirklich 1
+▁st 1
+ist 1
+wert 1
+mittel 1
+fi 1
+ho 1
+grund 1
+▁verheiratet 1
+▁wort 1
+mi 1
+ce 1
+▁pe 1
+▁co 1
+ker 1
+▁könig 1
+▁letzte 1
+▁finden 1
+punkt 1
+▁deutsche 1
+ba 1
+einander 1
+▁arbeitete 1
+▁lebens 1
+op 1
+tel 1
+▁lu 1
+ität 1
+▁niemand 1
+▁the 1
+wohl 1
+halten 1
+▁stand 1
+heim 1
+▁eis 1
+weise 1
+▁licht 1
+▁bar 1
+ven 1
+▁neun 1
+gra 1
+▁alt 1
+bei 1
+▁abge 1
+▁weltkrieg 1
+▁hätte 1
+stoff 1
+▁burg 1
+per 1
+lu 1
+▁spielte 1
+▁mi 1
+▁werk 1
+haft 1
+om 1
+▁direkt 1
+raum 1
+iger 1
+▁kleine 1
+reich 1
+▁rolle 1
+▁morgen 1
+▁erhalten 1
+▁komm 1
+▁tätig 1
+▁richtig 1
+schen 1
+▁rechts 1
+rie 1
+▁alter 1
+▁ki 1
+isiert 1
+▁sonst 1
+▁verein 1
+geben 1
+▁nahm 1
+▁of 1
+▁klar 1
+ve 1
+▁sicher 1
+▁heißt 1
+pel 1
+▁neuen 1
+▁see 1
+▁wollen 1
+▁zeigt 1
+▁fall 1
+▁buch 1
+▁genannt 1
+▁anschließend 1
+▁unterschiedlich 1
+eck 1
+▁kunst 1
+▁liegen 1
+▁rein 1
+anz 1
+ak 1
+▁do 1
+ismus 1
+▁leute 1
+min 1
+ehm 1
+▁jo 1
+stellung 1
+▁versucht 1
+▁kirche 1
+ischer 1
+▁schl 1
+bach 1
+▁sonder 1
+▁dein 1
+▁kaum 1
+▁sieben 1
+▁l 1
+▁mag 1
+▁te 1
+▁auto 1
+▁begann 1
+▁versch 1
+▁wald 1
+▁ko 1
+▁süd 1
+▁folgende 1
+▁inner 1
+▁mutter 1
+▁einzige 1
+▁gesch 1
+ca 1
+▁darüber 1
+ppen 1
+sicht 1
+▁fr 1
+▁führt 1
+▁englisch 1
+mon 1
+▁heraus 1
+▁ursprünglich 1
+▁leicht 1
+▁sta 1
+arbeit 1
+▁nord 1
+bt 1
+▁stellen 1
+dessen 1
+▁europa 1
+▁fand 1
+lös 1
+reif 1
+▁privat 1
+▁alten 1
+führung 1
+stellen 1
+▁unternehmen 1
+qu 1
+zeichen 1
+nahme 1
+▁geld 1
+▁besten 1
+▁ehe 1
+▁musste 1
+frei 1
+▁öffentlich 1
+▁schlecht 1
+ow 1
+▁sport 1
+▁begriff 1
+▁echt 1
+▁ku 1
+▁fach 1
+ard 1
+▁qu 1
+▁halten 1
+▁geben 1
+▁laut 1
+mme 1
+▁kar 1
+▁dar 1
+▁tor 1
+mpf 1
+gesetzt 1
+▁früher 1
+▁essen 1
+erweise 1
+▁dorf 1
+fel 1
+▁größte 1
+▁wein 1
+▁bestimmt 1
+▁einge 1
+▁los 1
+sprech 1
+▁wissen 1
+▁bedeutung 1
+▁führte 1
+▁tun 1
+zahl 1
+▁pu 1
+schule 1
+wand 1
+▁erfolgreich 1
+▁kreis 1
+schrift 1
+▁si 1
+▁name 1
+ord 1
+stelle 1
+ok 1
+tsch 1
+▁nehmen 1
+▁stein 1
+▁bleibt 1
+dorf 1
+zeug 1
+meister 1
+▁anders 1
+▁eigenen 1
+▁bist 1
+▁schau 1
+▁wahl 1
+öl 1
+▁dadurch 1
+dienst 1
+▁west 1
+▁partei 1
+elle 1
+hin 1
+▁mark 1
+fa 1
+▁folge 1
+▁freund 1
+▁ziel 1
+▁va 1
+fahrt 1
+ack 1
+▁d 1
+räum 1
+▁besitz 1
+▁bildet 1
+vo 1
+▁stammt 1
+seite 1
+▁wert 1
+druck 1
+▁handelt 1
+sche 1
+▁el 1
+hör 1
+and 1
+▁beispiel 1
+▁plan 1
+io 1
+▁manchmal 1
+kon 1
+bild 1
+▁per 1
+rad 1
+▁allen 1
+ug 1
+▁französisch 1
+war 1
+▁gemacht 1
+ill 1
+▁text 1
+▁alte 1
+ang 1
+▁manche 1
+leben 1
+igkeit 1
+age 1
+straße 1
+▁mü 1
+▁titel 1
+ett 1
+▁gebäude 1
+halt 1
+▁pol 1
+art 1
+▁ad 1
+mmer 1
+▁großen 1
+▁fällt 1
+▁schwarz 1
+▁ließ 1
+▁for 1
+▁anderem 1
+▁schließlich 1
+▁rück 1
+▁gehörte 1
+▁tod 1
+▁braucht 1
+▁letzten 1
+▁augen 1
+kla 1
+▁eltern 1
+▁farb 1
+höhe 1
+▁hält 1
+▁insel 1
+▁straßen 1
+besondere 1
+gestellt 1
+sta 1
+see 1
+laden 1
+▁schule 1
+satz 1
+▁kom 1
+itz 1
+stift 1
+hol 1
+▁fort 1
+ziehen 1
+▁fi 1
+folge 1
+gelegt 1
+tan 1
+▁arm 1
+▁lehr 1
+kan 1
+▁schloss 1
+sam 1
+▁zie 1
+tre 1
+▁werde 1
+▁trat 1
+äl 1
+tzt 1
+▁kind 1
+▁regel 1
+gabe 1
+▁zog 1
+▁bevor 1
+▁bald 1
+▁tochter 1
+stieg 1
+ort 1
+▁ost 1
+▁damals 1
+▁gehören 1
+▁gesehen 1
+▁herum 1
+wer 1
+ide 1
+▁kra 1
+▁stellt 1
+▁arbeits 1
+handel 1
+▁fahr 1
+gebracht 1
+▁körper 1
+▁nieder 1
+▁jugend 1
+▁no 1
+▁schiff 1
+ator 1
+▁pf 1
+▁sagen 1
+ach 1
+▁gelang 1
+lt 1
+▁daneben 1
+▁wieso 1
+ring 1
+ah 1
+gi 1
+▁daten 1
+legen 1
+▁ir 1
+▁bas 1
+▁wa 1
+▁bereich 1
+pi 1
+vor 1
+▁vielen 1
+▁fernseh 1
+▁offen 1
+▁daraufhin 1
+▁park 1
+▁grab 1
+▁denn 1
+▁wann 1
+▁spielen 1
+ry 1
+▁sogar 1
+▁aktiv 1
+▁grün 1
+▁eigen 1
+▁gemeinsam 1
+▁gott 1
+▁langsam 1
+▁gegenüber 1
+▁jeden 1
+▁falsch 1
+port 1
+▁ehemalige 1
+▁gef 1
+we 1
+wo 1
+▁zweite 1
+▁geh 1
+▁preis 1
+schein 1
+▁seite 1
+▁nächsten 1
+▁ju 1
+▁gleichzeitig 1
+▁gen 1
+deck 1
+gie 1
+▁wenige 1
+▁überhaupt 1
+▁kreuz 1
+▁genutzt 1
+bruch 1
+lau 1
+▁fahren 1
+▁herz 1
+fahren 1
+▁to 1
+▁dritte 1
+▁weiterhin 1
+▁mannschaft 1
+▁ga 1
+scha 1
+▁gute 1
+verkehr 1
+▁irgendwie 1
+▁blick 1
+mar 1
+träger 1
+graf 1
+▁selten 1
+▁landwirtschaft 1
+▁weiteren 1
+▁wohn 1
+▁ganze 1
+▁sollten 1
+▁holz 1
+▁aufge 1
+▁boden 1
+setzen 1
+▁anfang 1
+tum 1
+voll 1
+▁lieber 1
+▁daran 1
+▁system 1
+pro 1
+▁frage 1
+igt 1
+▁nein 1
+bro 1
+▁tr 1
+eher 1
+▁spielt 1
+up 1
+the 1
+▁raum 1
+▁beste 1
+▁pass 1
+log 1
+▁bleiben 1
+▁meinen 1
+▁worden 1
+▁autor 1
+▁rest 1
+▁mil 1
+leiter 1
+▁roman 1
+amt 1
+▁san 1
+▁kloster 1
+▁not 1
+▁hatten 1
+▁flug 1
+geordnet 1
+▁sagt 1
+gebaut 1
+▁stets 1
+▁go 1
+▁betrieb 1
+▁minuten 1
+▁gewann 1
+▁gerne 1
+bereich 1
+▁größe 1
+▁ei 1
+netz 1
+▁fisch 1
+▁lag 1
+▁eben 1
+▁fluss 1
+▁dienst 1
+▁fe 1
+▁tief 1
+ina 1
+schlag 1
+▁theater 1
+▁einsatz 1
+▁eingesetzt 1
+▁bürger 1
+▁ben 1
+gleich 1
+ssen 1
+▁kl 1
+▁tür 1
+sion 1
+▁bra 1
+▁start 1
+▁sozial 1
+od 1
+▁hinaus 1
+▁schi 1
+▁mai 1
+isierung 1
+ähr 1
+▁reich 1
+▁gewinn 1
+▁blut 1
+▁arbeiten 1
+▁halt 1
+▁gold 1
+▁kleiner 1
+▁gesellschaft 1
+▁funktion 1
+▁mus 1
+▁leistung 1
+▁or 1
+mut 1
+▁entwicklung 1
+▁hundert 1
+▁feuer 1
+zer 1
+▁sprache 1
+▁länger 1
+▁kauf 1
+▁inter 1
+system 1
+tlich 1
+▁ach 1
+▁fri 1
+isten 1
+▁volks 1
+▁fertig 1
+▁straße 1
+▁beginn 1
+schrei 1
+ag 1
+▁wahrscheinlich 1
+▁gefunden 1
+▁rat 1
+▁endlich 1
+▁new 1
+▁raus 1
+▁projekt 1
+eiß 1
+▁müsste 1
+▁mensch 1
+▁karriere 1
+sprache 1
+scheid 1
+▁sätze 1
+▁punkt 1
+▁nächste 1
+▁liebe 1
+▁einzel 1
+▁reicht 1
+▁richtung 1
+weit 1
+▁nimmt 1
+ding 1
+kirche 1
+▁post 1
+▁tu 1
+▁bad 1
+station 1
+▁leider 1
+▁strom 1
+▁wesentlich 1
+▁städte 1
+▁sah 1
+▁aufgabe 1
+▁hauptstadt 1
+▁amt 1
+▁vorher 1
+geschichte 1
+genommen 1
+▁junge 1
+ki 1
+▁zusätzlich 1
+▁werke 1
+▁zer 1
+▁gruppe 1
+▁kirchen 1
+▁unbe 1
+▁weniger 1
+▁uhr 1
+▁bri 1
+▁eng 1
+▁meiner 1
+gesellschaft 1
+▁heiß 1
+schau 1
+▁tage 1
+▁team 1
+zig 1
+▁dr 1
+▁lied 1
+▁versuch 1
+▁veröffentlicht 1
+▁druck 1
+ud 1
+ron 1
+▁u 1
+▁hol 1
+▁polizei 1
+schiff 1
+▁nacht 1
+▁wi 1
+▁gi 1
+gegen 1
+▁bruder 1
+▁besuchte 1
+▁künstler 1
+▁komplett 1
+▁deutlich 1
+if 1
+▁klasse 1
+▁bisher 1
+▁hör 1
+hof 1
+plan 1
+film 1
+trieb 1
+▁befinden 1
+▁abend 1
+▁regional 1
+▁setzt 1
+▁übrig 1
+▁modern 1
+▁op 1
+▁regen 1
+▁kenn 1
+▁allein 1
+▁breit 1
+▁prä 1
+▁saison 1
+wirtschaft 1
+▁allgemein 1
+▁vergessen 1
+▁höher 1
+schlagen 1
+▁zug 1
+tik 1
+▁sprach 1
+gu 1
+▁seiten 1
+last 1
+▁rund 1
+xi 1
+▁wochen 1
+▁hamburg 1
+▁frauen 1
+▁selb 1
+treten 1
+▁keiner 1
+▁dreh 1
+▁sand 1
+▁neuer 1
+getragen 1
+▁wuchs 1
+▁geboren 1
+▁normal 1
+wegen 1
+tritt 1
+stä 1
+▁machte 1
+fl 1
+sitz 1
+▁gast 1
+▁wahr 1
+▁problem 1
+▁gesamte 1
+▁paul 1
+▁sollen 1
+weile 1
+▁stunden 1
+üss 1
+▁wichtige 1
+stände 1
+musik 1
+▁fragen 1
+▁blau 1
+▁gleichen 1
+▁entstanden 1
+wagen 1
+▁cha 1
+erei 1
+▁weise 1
+gemeinde 1
+▁fuß 1
+▁gas 1
+kreis 1
+▁hohe 1
+▁guten 1
+▁bild 1
+statt 1
+wehr 1
+▁nördlich 1
+▁ur 1
+les 1
+▁georg 1
+▁tut 1
+▁tiere 1
+fach 1
+▁zählt 1
+du 1
+▁dra 1
+len 1
+gegeben 1
+▁habt 1
+▁bestand 1
+j 1
+▁i 1
+▁lehrer 1
+▁hall 1
+▁stück 1
+oren 1
+▁kalt 1
+▁starb 1
+▁aller 1
+gelassen 1
+▁har 1
+fang 1
+▁schlaf 1
+führer 1
+▁region 1
+sehen 1
+stück 1
+▁seitdem 1
+▁fenster 1
+▁staats 1
+▁musst 1
+▁verbindung 1
+▁zentral 1
+▁kritik 1
+schloss 1
+▁bedeutet 1
+▁genug 1
+za 1
+▁teilweise 1
+lig 1
+▁innen 1
+▁bevölkerung 1
+länder 1
+stellt 1
+rich 1
+▁eure 1
+▁führen 1
+▁italien 1
+bu 1
+wasser 1
+bert 1
+▁hä 1
+▁vo 1
+ham 1
+▁finde 1
+pu 1
+▁hi 1
+▁bestehen 1
+oni 1
+▁letzter 1
+schutz 1
+▁kor 1
+▁hart 1
+▁arbeitet 1
+▁woche 1
+ständig 1
+nder 1
+▁bahn 1
+▁gu 1
+▁gerät 1
+▁rad 1
+▁miss 1
+▁kla 1
+▁wirkt 1
+wei 1
+▁seid 1
+▁pl 1
+unterricht 1
+▁errichtet 1
+▁bietet 1
+▁markt 1
+jahr 1
+ruf 1
+gg 1
+noch 1
+▁gewesen 1
+lassen 1
+▁ähnlich 1
+richtung 1
+▁elektro 1
+linie 1
+▁car 1
+teile 1
+▁kilometer 1
+▁reg 1
+tiert 1
+▁spieler 1
+▁heilig 1
+bank 1
+▁kleinen 1
+rin 1
+▁kur 1
+meisterschaft 1
+tür 1
+▁sommer 1
+▁unser 1
+▁außen 1
+▁älter 1
+▁winter 1
+▁bla 1
+▁stelle 1
+▁tier 1
+▁wind 1
+▁dom 1
+schicht 1
+▁namens 1
+zwischen 1
+▁anti 1
+▁mu 1
+▁blüten 1
+▁hans 1
+▁hinzu 1
+▁kultur 1
+▁th 1
+▁sal 1
+wesen 1
+▁legt 1
+▁insgesamt 1
+▁nutz 1
+▁str 1
+▁wechsel 1
+▁dürfen 1
+ika 1
+licht 1
+ris 1
+▁stimme 1
+▁orts 1
+setzung 1
+grün 1
+▁trotz 1
+ius 1
+▁kaiser 1
+▁trau 1
+▁strecke 1
+▁schuld 1
+▁trink 1
+▁schüler 1
+▁mittler 1
+ierten 1
+▁serie 1
+▁bahnhof 1
+innen 1
+▁tot 1
+steigen 1
+▁mitglieder 1
+gruppe 1
+▁glaub 1
+▁hauptsächlich 1
+leitung 1
+stätte 1
+▁internationale 1
+▁deck 1
+▁förder 1
+▁ni 1
+▁österreich 1
+▁familien 1
+▁still 1
+▁fin 1
+dauer 1
+▁schrieb 1
+▁offenbar 1
+▁fahrzeug 1
+je 1
+▁ausbildung 1
+▁kosten 1
+wissenschaftlich 1
+▁gering 1
+selben 1
+▁ti 1
+▁port 1
+▁tri 1
+▁pri 1
+schließlich 1
+▁baum 1
+▁glauben 1
+mat 1
+▁mor 1
+▁glas 1
+▁paris 1
+wald 1
+▁gl 1
+▁schrift 1
+fläche 1
+▁rede 1
+mus 1
+▁oh 1
+▁regelmäßig 1
+▁mund 1
+▁bezeichnung 1
+arzt 1
+▁motor 1
+ette 1
+▁angebot 1
+▁verkehrs 1
+▁völlig 1
+▁relativ 1
+▁internet 1
+rang 1
+▁heimat 1
+▁ruf 1
+▁provinz 1
+▁hände 1
+▁grenze 1
+▁unterschied 1
+richten 1
+nach 1
+▁fußball 1
+▁lernen 1
+▁ke 1
+▁sprechen 1
+▁feld 1
+karte 1
+▁kriegs 1
+▁glück 1
+▁einigen 1
+▁min 1
+▁zwölf 1
+by 1
+▁bre 1
+▁entsprechend 1
+▁fried 1
+▁dinge 1
+▁größere 1
+▁ph 1
+▁südlich 1
+▁mann 1
+▁hof 1
+▁chi 1
+old 1
+▁sofort 1
+effekt 1
+▁münchen 1
+führen 1
+▁schreiben 1
+▁mach 1
+arm 1
+▁stra 1
+▁zi 1
+▁zeug 1
+▁hervor 1
+▁stil 1
+▁konnten 1
+zog 1
+▁bauch 1
+prüf 1
+▁ton 1
+▁schmeckt 1
+▁geprägt 1
+▁vorhanden 1
+▁interesse 1
+ingen 1
+fahrer 1
+nfalls 1
+▁verschieden 1
+spieler 1
+buch 1
+turm 1
+▁bal 1
+▁ganzen 1
+blätter 1
+ologie 1
+▁derzeit 1
+▁westlich 1
+pass 1
+▁mon 1
+▁zieht 1
+▁fiel 1
+namen 1
+▁jedem 1
+let 1
+▁museum 1
+▁meister 1
+▁meistens 1
+führt 1
+▁übernahm 1
+▁vergleich 1
+▁trotzdem 1
+ath 1
+▁literatur 1
+▁back 1
+▁ball 1
+▁fern 1
+▁christ 1
+gericht 1
+▁bilden 1
+▁technisch 1
+▁links 1
+▁pflanzen 1
+ände 1
+witz 1
+▁kampf 1
+▁kennt 1
+▁zahl 1
+▁spricht 1
+▁jeweil 1
+▁eh 1
+▁gesagt 1
+val 1
+▁gewisse 1
+ly 1
+▁j 1
+▁programm 1
+▁mess 1
+schluss 1
+▁folgen 1
+schütz 1
+▁geschäfts 1
+familie 1
+verwaltung 1
+ück 1
+▁gesamt 1
+▁äußer 1
+▁general 1
+▁tro 1
+▁studium 1
+die 1
+▁bücher 1
+▁ber 1
+▁beginnt 1
+▁thema 1
+gebäude 1
+lose 1
+leit 1
+▁dient 1
+▁gewählt 1
+▁verbunden 1
+stärke 1
+brü 1
+▁erfolgt 1
+▁einfluss 1
+▁kindern 1
+▁lebte 1
+▁anderer 1
+▁sonne 1
+▁verstehen 1
+▁wetter 1
+anstalt 1
+▁wichtigste 1
+lief 1
+▁daraus 1
+ativ 1
+▁fährt 1
+▁steck 1
+▁tauch 1
+▁vertrag 1
+iell 1
+▁deutsch 1
+unter 1
+▁staat 1
+spruch 1
+▁enthält 1
+▁sinn 1
+▁tragen 1
+bogen 1
+▁dicht 1
+mannschaft 1
+▁europäische 1
+mlich 1
+kunde 1
+kal 1
+sätze 1
+▁modell 1
+förmig 1
+▁männer 1
+fest 1
+kurs 1
+ost 1
+▁gestern 1
+▁gesicht 1
+ac 1
+▁spar 1
+bra 1
+kauf 1
+▁linie 1
+▁anderes 1
+mäßig 1
+gemeinschaft 1
+sweise 1
+▁kennen 1
+▁finanz 1
+▁stern 1
+▁personen 1
+▁trägt 1
+▁nähe 1
+win 1
+hält 1
+hundert 1
+deutsch 1
+stet 1
+▁beschäftigt 1
+leiste 1
+wort 1
+off 1
+▁tal 1
+yl 1
+ock 1
+▁tages 1
+krieg 1
+▁stau 1
+▁gegeben 1
+▁hund 1
+band 1
+▁gib 1
+▁trug 1
+▁schreib 1
+▁runde 1
+▁dach 1
+▁nor 1
+richter 1
+▁sammel 1
+gegangen 1
+brechen 1
+wechsel 1
+▁international 1
+ay 1
+▁tatsächlich 1
+▁älteste 1
+▁hilfe 1
+▁heim 1
+▁galt 1
+▁automatisch 1
+▁end 1
+ica 1
+jährige 1
+▁temperatur 1
+▁fehler 1
+▁zuvor 1
+▁mari 1
+anlage 1
+▁setzte 1
+▁pre 1
+▁tat 1
+imm 1
+betrieb 1
+▁produktion 1
+▁beliebt 1
+stuhl 1
+▁besitzt 1
+▁besuch 1
+▁au 1
+▁schauspieler 1
+▁erster 1
+▁ziehen 1
+▁dick 1
+▁app 1
+▁politik 1
+körper 1
+▁gericht 1
+▁ve 1
+▁tre 1
+▁bloß 1
+dia 1
+änderung 1
+▁chor 1
+steuer 1
+▁fehlt 1
+▁zentrum 1
+▁stellte 1
+▁armee 1
+dy 1
+▁glaube 1
+▁folgte 1
+▁kamen 1
+▁politische 1
+▁com 1
+▁haar 1
+zentrum 1
+handlung 1
+▁streit 1
+▁rhein 1
+▁hell 1
+▁erneut 1
+▁verwenden 1
+▁bekommen 1
+wachs 1
+▁finger 1
+▁organ 1
+▁tanz 1
+▁steuer 1
+rot 1
+▁album 1
+▁peter 1
+pfen 1
+hilfe 1
+kehr 1
+▁liste 1
+▁schwierig 1
+▁frankreich 1
+▁hohen 1
+sprung 1
+▁überwiegend 1
+rand 1
+▁schwe 1
+▁jahres 1
+▁bett 1
+iker 1
+▁bank 1
+▁norden 1
+über 1
+fälle 1
+▁helfen 1
+kräfte 1
+▁probleme 1
+red 1
+▁silber 1
+▁ebenso 1
+▁wand 1
+▁folgten 1
+▁sam 1
+▁mö 1
+insel 1
+▁bringen 1
+▁stamm 1
+▁fliegen 1
+▁satz 1
+▁kraft 1
+▁schutz 1
+gezogen 1
+blick 1
+▁stell 1
+▁straf 1
+glieder 1
+sport 1
+▁erreicht 1
+ahn 1
+▁nahe 1
+▁fehl 1
+▁gruppen 1
+▁university 1
+▁ledig 1
+▁wappen 1
+▁gebaut 1
+▁verd 1
+▁süden 1
+stö 1
+▁gelb 1
+pper 1
+▁größten 1
+▁ernst 1
+▁meter 1
+stern 1
+▁lesen 1
+arten 1
+▁friedrich 1
+▁ergebnis 1
+▁ehren 1
+▁vollständig 1
+lager 1
+▁umfasst 1
+▁teile 1
+▁mer 1
+▁beruf 1
+▁industrie 1
+▁bekommt 1
+▁überall 1
+schließ 1
+▁schritt 1
+▁johann 1
+arch 1
+bus 1
+▁hören 1
+▁fo 1
+▁arten 1
+▁lauf 1
+▁verfügt 1
+stock 1
+technik 1
+prüfung 1
+▁militär 1
+preis 1
+▁reihe 1
+▁verwaltungs 1
+tät 1
+weiß 1
+▁karl 1
+▁garten 1
+▁typ 1
+▁dank 1
+▁lösung 1
+▁kamera 1
+▁treffen 1
+ian 1
+ther 1
+▁wechselte 1
+▁leitet 1
+▁wissenschaft 1
+▁red 1
+▁wild 1
+▁martin 1
+▁dagegen 1
+▁aufnahme 1
+▁ed 1
+▁bildung 1
+▁seh 1
+wind 1
+▁regierung 1
+gewiesen 1
+▁beziehung 1
+▁brenn 1
+▁sc 1
+▁verloren 1
+▁entstand 1
+hand 1
+▁monate 1
+org 1
+▁london 1
+▁köln 1
+▁treib 1
+▁vogel 1
+▁mehrfach 1
+▁sing 1
+ball 1
+▁schluss 1
+anlagen 1
+fü 1
+▁schweiz 1
+▁hört 1
+▁ag 1
+▁miteinander 1
+▁energie 1
+▁nachfolger 1
+▁brachte 1
+▁stich 1
+frage 1
+▁anzu 1
+▁westen 1
+▁real 1
+▁con 1
+ini 1
+▁müll 1
+▁politisch 1
+▁entscheidung 1
+▁befand 1
+▁gelten 1
+▁beendet 1
+flügel 1
+▁song 1
+▁institut 1
+hängen 1
+cho 1
+pflicht 1
+▁zwanzig 1
+▁haut 1
+schul 1
+tausch 1
+▁zweck 1
+▁präsident 1
+▁vorbei 1
+stehen 1
+▁hü 1
+▁egal 1
+▁unabhängig 1
+▁enthalten 1
+ara 1
+sicher 1
+▁zuerst 1
+ari 1
+stoffe 1
+▁studien 1
+▁feier 1
+▁schwester 1
+sort 1
+▁bringt 1
+hütte 1
+▁reis 1
+▁möglichkeit 1
+▁technik 1
+▁erklärt 1
+▁benutzt 1
+▁lokal 1
+▁zwar 1
+▁firma 1
+▁gegensatz 1
+maschine 1
+sieg 1
+▁radio 1
+▁suche 1
+▁frisch 1
+▁spät 1
+▁entscheide 1
+horn 1
+▁erstmals 1
+▁john 1
+▁grau 1
+▁trainer 1
+uß 1
+▁wagen 1
+gl 1
+fuhr 1
+treffen 1
+finale 1
+▁pal 1
+fo 1
+wende 1
+▁entfernt 1
+medaille 1
+▁erschien 1
+takt 1
+gänge 1
+▁pan 1
+▁schlacht 1
+▁verfahren 1
+hu 1
+▁cor 1
+runde 1
+kraft 1
+kette 1
+▁gesetz 1
+▁ring 1
+▁prozess 1
+▁leucht 1
+▁liga 1
+▁jung 1
+hard 1
+don 1
+▁zuletzt 1
+▁einwohner 1
+▁gespielt 1
+▁drauf 1
+▁wider 1
+fasst 1
+▁dauert 1
+▁zuständig 1
+kuchen 1
+leistung 1
+▁passiert 1
+▁sag 1
+äre 1
+▁antwort 1
+▁geb 1
+▁elf 1
+▁nimm 1
+▁verhalten 1
+geteilt 1
+▁wirtschafts 1
+herr 1
+▁zeigen 1
+well 1
+lü 1
+ino 1
+▁ständig 1
+▁frag 1
+▁unterstützt 1
+▁öffne 1
+▁östlich 1
+▁klassische 1
+▁überrasch 1
+▁scheint 1
+▁spring 1
+▁gö 1
+▁entstehen 1
+▁wohnung 1
+uf 1
+▁friedhof 1
+▁char 1
+titel 1
+kammer 1
+▁wächst 1
+▁praktisch 1
+▁erkennen 1
+▁hab 1
+▁wirtschaftlich 1
+wissenschaften 1
+▁schwerpunkt 1
+▁auszu 1
+▁lautet 1
+▁mä 1
+ktion 1
+▁treten 1
+▁stimmt 1
+hausen 1
+produkt 1
+▁existiert 1
+▁alternativ 1
+▁fallen 1
+▁siedlung 1
+gar 1
+▁hei 1
+abteilung 1
+▁milch 1
+▁euro 1
+▁verlassen 1
+▁ecke 1
+reise 1
+träge 1
+▁korrekt 1
+▁heutigen 1
+▁geworden 1
+▁kaffee 1
+▁aktuell 1
+▁rauch 1
+▁reichs 1
+▁passt 1
+▁platt 1
+▁tätigkeit 1
+▁eier 1
+▁lager 1
+ov 1
+▁rote 1
+▁erreichte 1
+▁sucht 1
+lag 1
+▁motiv 1
+fähig 1
+▁gebrauch 1
+maßnahmen 1
+▁baby 1
+▁entwickelte 1
+▁britische 1
+vier 1
+ya 1
+denken 1
+▁historische 1
+▁verläuft 1
+▁traum 1
+▁pla 1
+ober 1
+▁wal 1
+▁geschrieben 1
+▁meinung 1
+flug 1
+sucht 1
+▁spezial 1
+läuft 1
+cu 1
+▁bus 1
+alter 1
+▁konzert 1
+▁entwickelt 1
+▁beteiligt 1
+plätze 1
+▁erlaubt 1
+methode 1
+▁verbreitet 1
+▁wunder 1
+▁charakter 1
+▁air 1
+züge 1
+▁jen 1
+ial 1
+▁laufen 1
+ana 1
+ologische 1
+▁hause 1
+▁flugzeug 1
+▁abzu 1
+▁sogenannte 1
+▁mädchen 1
+▁extrem 1
+gespräch 1
+▁gebe 1
+gymnasium 1
+▁brand 1
+▁bern 1
+gro 1
+▁angriff 1
+▁rücken 1
+oli 1
+▁kasse 1
+grad 1
+wohn 1
+halle 1
+▁turm 1
+▁maschinen 1
+▁schach 1
+▁bier 1
+komm 1
+▁version 1
+arbeiten 1
+▁leiden 1
+▁wander 1
+▁zukunft 1
+▁york 1
+▁böse 1
+▁maschine 1
+▁regeln 1
+▁test 1
+▁bot 1
+▁einzelnen 1
+▁verband 1
+▁bewegung 1
+▁amerikanische 1
+▁diente 1
+▁forschungs 1
+ski 1
+▁hängt 1
+▁bio 1
+haltung 1
+▁schützen 1
+▁funktioniert 1
+▁juli 1
+▁transport 1
+▁tisch 1
+hob 1
+otten 1
+▁zeitung 1
+inger 1
+▁wichtig 1
+zeitig 1
+küste 1
+▁schriftsteller 1
+▁vermutlich 1
+▁gründung 1
+▁diverse 1
+▁beispielsweise 1
+▁fühlt 1
+theorie 1
+▁hielt 1
+▁erreichen 1
+▁kunden 1
+▁hit 1
+▁lieb 1
+▁landschaft 1
+fassen 1
+▁profi 1
+farbe 1
+stil 1
+wil 1
+▁reichen 1
+▁alb 1
+üll 1
+▁längst 1
+▁katholisch 1
+▁zerstört 1
+undzwanzig 1
+▁par 1
+auftrag 1
+▁pra 1
+▁anwendung 1
+▁erde 1
+ci 1
+▁würden 1
+schuss 1
+dach 1
+prä 1
+▁ze 1
+▁schwarze 1
+▁stabil 1
+▁produziert 1
+á 1
+▁üblich 1
+▁frankfurt 1
+▁fremd 1
+▁herrschaft 1
+▁monat 1
+▁schlag 1
+▁nachbarn 1
+▁weich 1
+▁maria 1
+▁einzelne 1
+ance 1
+▁absolvierte 1
+zel 1
+▁geburt 1
+▁veranstaltung 1
+▁küche 1
+▁para 1
+▁zahn 1
+▁betrieben 1
+liga 1
+▁reise 1
+▁bauern 1
+▁idee 1
+▁erzähl 1
+rett 1
+▁kommst 1
+▁cu 1
+▁original 1
+▁computer 1
+▁söhne 1
+▁wilhelm 1
+▁doktor 1
+▁durchgeführt 1
+▁mindestens 1
+▁flasche 1
+jo 1
+▁geschlossen 1
+▁franz 1
+verbot 1
+lenk 1
+▁rechte 1
+denkmal 1
+▁brauchen 1
+machen 1
+▁cla 1
+▁öl 1
+▁zugleich 1
+▁ruhe 1
+kreuz 1
+lehrer 1
+nummer 1
+▁schreibt 1
+▁amts 1
+▁rand 1
+▁gehalten 1
+▁eisenbahn 1
+▁auge 1
+rolle 1
+geschwindigkeit 1
+▁symbol 1
+fabrik 1
+▁verzicht 1
+▁null 1
+▁sprachen 1
+pol 1
+▁geschäft 1
+▁könnten 1
+▁entscheidend 1
+▁sonnen 1
+▁anlage 1
+nennt 1
+▁dü 1
+▁vereins 1
+▁erfolgte 1
+▁güter 1
+▁eu 1
+▁pat 1
+kirch 1
+markt 1
+geschoss 1
+▁typisch 1
+▁evangelisch 1
+▁grundsätzlich 1
+▁umgebung 1
+geschlecht 1
+▁brief 1
+▁versteht 1
+▁geführt 1
+▁hor 1
+schl 1
+▁gelände 1
+post 1
+▁nachbar 1
+liste 1
+bereit 1
+▁oper 1
+stall 1
+▁rang 1
+▁willst 1
+länge 1
+▁county 1
+▁entschied 1
+▁publikum 1
+▁jünger 1
+▁flughafen 1
+▁höchste 1
+▁wolf 1
+▁zucker 1
+▁danke 1
+▁zeitschrift 1
+rennen 1
+▁zeitpunkt 1
+▁osten 1
+▁fett 1
+ley 1
+▁doppel 1
+ordnung 1
+verein 1
+▁versuchen 1
+kämpfe 1
+pflanze 1
+▁digital 1
+▁situation 1
+▁weitgehend 1
+▁medien 1
+▁kontakt 1
+▁bibliothek 1
+▁mode 1
+akte 1
+▁china 1
+▁berufs 1
+liegen 1
+geht 1
+▁spitze 1
+▁lief 1
+▁spiegel 1
+▁maß 1
+bezirk 1
+häuser 1
+▁fleisch 1
+fried 1
+▁ski 1
+bombe 1
+▁artikel 1
+▁che 1
+▁kol 1
+schneid 1
+▁star 1
+▁benötigt 1
+▁plötzlich 1
+▁zoll 1
+▁geplant 1
+▁eingang 1
+fassung 1
+▁landkreis 1
+▁katze 1
+▁schnitt 1
+schalten 1
+▁republik 1
+▁aufgenommen 1
+▁verfasst 1
+▁bäume 1
+▁kontroll 1
+▁mitarbeiter 1
+▁video 1
+rauch 1
+▁beigesetzt 1
+▁bur 1
+bringen 1
+quelle 1
+▁wach 1
+geber 1
+▁prinzip 1
+bär 1
+▁bedeutende 1
+club 1
+▁mel 1
+▁dunkel 1
+▁darunter 1
+star 1
+▁themen 1
+▁denke 1
+▁bereit 1
+▁wem 1
+▁bisschen 1
+zutage 1
+klasse 1
+▁position 1
+▁rum 1
+▁roll 1
+bürger 1
+▁vorstellung 1
+▁quer 1
+waffe 1
+▁august 1
+▁gelegentlich 1
+höfe 1
+▁bezeichnen 1
+gäste 1
+▁zweifel 1
+sache 1
+▁gern 1
+tha 1
+kampf 1
+ndel 1
+gerechnet 1
+trop 1
+▁frank 1
+▁freiwillig 1
+▁knochen 1
+▁stadtteil 1
+▁erwachsene 1
+▁fahrrad 1
+papier 1
+▁mond 1
+▁hätten 1
+▁großer 1
+▁ausnahme 1
+▁zweimal 1
+▁play 1
+▁wärme 1
+▁religion 1
+▁gleichnamige 1
+▁soldat 1
+▁hafen 1
+meld 1
+▁lau 1
+▁toll 1
+▁trans 1
+▁beton 1
+pack 1
+reihe 1
+▁konzept 1
+▁professor 1
+büro 1
+ali 1
+wertung 1
+wie 1
+stufe 1
+▁unver 1
+▁koch 1
+▁tour 1
+▁instrument 1
+bücher 1
+▁handy 1
+▁vertreter 1
+▁stammen 1
+▁mittelalter 1
+istischen 1
+▁bü 1
+▁trag 1
+staat 1
+▁funk 1
+geleitet 1
+▁hunger 1
+▁zunehmend 1
+▁entgegen 1
+▁england 1
+rry 1
+▁handels 1
+▁linke 1
+lied 1
+fleck 1
+▁schade 1
+▁kron 1
+▁sänger 1
+programm 1
+▁darin 1
+stimmung 1
+bauten 1
+▁telefon 1
+▁zählen 1
+▁himmel 1
+▁signal 1
+▁schieß 1
+▁brücke 1
+▁jungen 1
+▁standort 1
+part 1
+▁vertreten 1
+fahr 1
+presse 1
+▁irgendw 1
+pft 1
+teilung 1
+▁schaffen 1
+▁bilder 1
+verband 1
+▁notwendig 1
+versorgung 1
+▁unbedingt 1
+▁michael 1
+ria 1
+▁italienische 1
+front 1
+stab 1
+spur 1
+olog 1
+▁besucher 1
+▁heut 1
+▁regierungs 1
+geist 1
+▁künstlich 1
+▁prozent 1
+▁entdeckt 1
+▁entspricht 1
+▁gestellt 1
+▁pferde 1
+▁vereinigt 1
+utz 1
+▁abgeschlossen 1
+▁stunde 1
+▁sache 1
+▁staaten 1
+blumen 1
+geld 1
+bb 1
+schnitt 1
+grenze 1
+arbeiter 1
+▁töchter 1
+▁heinrich 1
+▁belegt 1
+▁gedanken 1
+▁trocken 1
+▁römische 1
+streckt 1
+▁wende 1
+▁verwendung 1
+▁ähnliche 1
+▁gedicht 1
+▁blatt 1
+▁web 1
+gerichtet 1
+▁näher 1
+kapelle 1
+▁vorne 1
+▁fläche 1
+▁müsst 1
+wunde 1
+ula 1
+▁struktur 1
+mark 1
+▁fahrt 1
+siedlung 1
+▁dor 1
+fertig 1
+▁vermeiden 1
+▁griechisch 1
+▁speziell 1
+▁erheblich 1
+▁unmittelbar 1
+qui 1
+bald 1
+▁gezogen 1
+▁ausgezeichnet 1
+strich 1
+▁sinne 1
+hang 1
+▁ausstellung 1
+▁fließt 1
+▁kern 1
+packung 1
+gelegenheit 1
+streich 1
+weil 1
+▁züge 1
+▁verkehr 1
+▁füll 1
+▁rock 1
+▁zahlen 1
+reaktion 1
+har 1
+▁top 1
+▁organisation 1
+akademie 1
+▁erinnert 1
+verletzung 1
+▁bischof 1
+▁voraus 1
+▁klingt 1
+▁russisch 1
+bildung 1
+▁anwalt 1
+bund 1
+▁kehrte 1
+motor 1
+▁weltweit 1
+alität 1
+▁letzt 1
+▁boot 1
+tive 1
+beutel 1
+▁darum 1
+wüste 1
+▁wenigsten 1
+chi 1
+▁bekannteste 1
+way 1
+▁blätter 1
+▁braun 1
+artig 1
+▁figur 1
+▁nachrichten 1
+▁mont 1
+▁bezirk 1
+klappe 1
+▁wehr 1
+pri 1
+▁weh 1
+▁gefallen 1
+▁eingestellt 1
+würfe 1
+▁komponist 1
+▁schlüssel 1
+▁gewalt 1
+▁schlimm 1
+▁gegenseitig 1
+▁erzielt 1
+▁leer 1
+▁viertel 1
+▁statu 1
+▁person 1
+▁auftritt 1
+▁meer 1
+isse 1
+▁häuser 1
+▁sau 1
+▁vorder 1
+wissenschaft 1
+zweig 1
+zo 1
+▁bach 1
+spende 1
+wochenende 1
+istische 1
+dro 1
+wing 1
+▁anschluss 1
+▁gründen 1
+ani 1
+geschlossen 1
+▁moment 1
+▁getrennt 1
+▁hälfte 1
+▁betroffen 1
+▁college 1
+vögel 1
+▁pfarr 1
+▁klavier 1
+▁todes 1
+▁spaß 1
+▁krank 1
+▁zusammenhang 1
+forderung 1
+zimmer 1
+▁jahrhundert 1
+▁herzog 1
+▁verkauf 1
+deln 1
+▁brot 1
+▁tritt 1
+▁personal 1
+boden 1
+▁minister 1
+welt 1
+dian 1
+gruppen 1
+▁holt 1
+▁medizin 1
+kleid 1
+▁fühle 1
+▁unten 1
+▁dumm 1
+▁leichter 1
+▁volk 1
+ó 1
+▁widmet 1
+wörter 1
+studio 1
+▁südwest 1
+rechnung 1
+spannung 1
+fähigkeit 1
+▁aussage 1
+▁verboten 1
+brücke 1
+funde 1
+wick 1
+▁suchen 1
+▁software 1
+▁blu 1
+▁schwein 1
+▁gewicht 1
+▁rom 1
+▁könnt 1
+▁schlagen 1
+▁afrika 1
+▁entstammt 1
+▁nötig 1
+▁kreativ 1
+▁latein 1
+▁schafft 1
+▁gelesen 1
+spreng 1
+▁spanisch 1
+▁angst 1
+held 1
+▁hoffen 1
+text 1
+▁chef 1
+verbindung 1
+▁heutige 1
+▁absicht 1
+wiegend 1
+▁gefangen 1
+variante 1
+▁gegner 1
+stoß 1
+▁fotograf 1
+▁leg 1
+produktion 1
+▁heb 1
+▁gekommen 1
+▁helm 1
+▁jüdische 1
+▁spa 1
+▁beweis 1
+▁flach 1
+▁fein 1
+▁laufe 1
+▁schmal 1
+▁nase 1
+steiger 1
+issen 1
+zone 1
+ai 1
+säure 1
+▁grad 1
+städt 1
+▁militärische 1
+▁salz 1
+wärts 1
+▁aktion 1
+▁unbekannt 1
+▁verfügung 1
+▁besetzt 1
+▁ändert 1
+name 1
+▁bewohner 1
+mindest 1
+▁standard 1
+▁fluch 1
+▁robert 1
+▁schließen 1
+▁hochschule 1
+▁sicherheits 1
+▁urteil 1
+▁bor 1
+▁hinten 1
+▁ortschaft 1
+▁frucht 1
+stopp 1
+▁wohnt 1
+▁leid 1
+▁gefahr 1
+▁kochen 1
+▁kühl 1
+▁österreichische 1
+▁ohren 1
+schuhe 1
+gefallen 1
+▁verlor 1
+bot 1
+▁geheim 1
+▁strecken 1
+▁architekt 1
+▁ausgestattet 1
+▁speicher 1
+▁positiv 1
+▁termin 1
+▁quellen 1
+saft 1
+industrie 1
+projekt 1
+braun 1
+schlaf 1
+menge 1
+▁schalt 1
+▁woher 1
+▁mussten 1
+▁eva 1
+▁spi 1
+ix 1
+lichkeit 1
+▁welch 1
+▁ohr 1
+gelaufen 1
+▁tradition 1
+offizier 1
+▁journalist 1
+▁untersuchung 1
+▁hilft 1
+▁wünsche 1
+▁verlieh 1
+▁ernannt 1
+▁süß 1
+▁bahnstrecke 1
+gebrochen 1
+▁entsteht 1
+▁opfer 1
+präsident 1
+modell 1
+▁legen 1
+▁klaus 1
+gerät 1
+▁abschluss 1
+▁orden 1
+formation 1
+▁adel 1
+▁beamte 1
+▁kohl 1
+▁musiker 1
+▁bedarf 1
+▁laub 1
+aktion 1
+▁partner 1
+raub 1
+schrank 1
+í 1
+▁jüngst 1
+▁niedrig 1
+▁hersteller 1
+▁kletter 1
+▁verdient 1
+▁damalige 1
+▁reiß 1
+▁andre 1
+▁verstanden 1
+wirkung 1
+platte 1
+▁nerv 1
+▁hunde 1
+▁tim 1
+gefangen 1
+direktor 1
+▁material 1
+fluss 1
+▁schild 1
+▁max 1
+▁schwimm 1
+▁konflikt 1
+▁stuttgart 1
+▁generation 1
+▁qualität 1
+▁rahmen 1
+▁organisiert 1
+vorsitzende 1
+▁bürgermeister 1
+▁menge 1
+▁vorteil 1
+▁kostenlos 1
+theater 1
+schläge 1
+▁entwe 1
+▁rennen 1
+▁ordnung 1
+brach 1
+weist 1
+▁wu 1
+▁nochmal 1
+wal 1
+seitig 1
+▁gegründet 1
+▁gezeigt 1
+▁nämlich 1
+▁bronze 1
+speise 1
+schlüsse 1
+▁feuerwehr 1
+▁herrscht 1
+▁verwandt 1
+▁orange 1
+▁dargestellt 1
+ness 1
+▁minute 1
+mauer 1
+kleidung 1
+▁verkauft 1
+▁sitzt 1
+▁sachen 1
+▁tipp 1
+▁forschung 1
+daten 1
+▁leitung 1
+▁tau 1
+▁bericht 1
+gut 1
+▁super 1
+▁dachte 1
+▁basis 1
+▁fragt 1
+▁ludwig 1
+▁beschrieben 1
+▁leipzig 1
+▁verließ 1
+division 1
+▁voice 1
+▁ersetzt 1
+▁niemals 1
+▁protest 1
+▁ermöglicht 1
+▁waffen 1
+dichte 1
+▁sichtbar 1
+▁lebensmittel 1
+▁gestaltet 1
+▁mühl 1
+▁ewig 1
+partner 1
+▁schweizer 1
+lohn 1
+▁schwach 1
+▁nummer 1
+tiere 1
+▁kal 1
+gemäß 1
+▁fotos 1
+▁dokument 1
+▁abhängig 1
+▁sil 1
+▁mauer 1
+behörde 1
+▁persönlich 1
+▁goldene 1
+▁gesundheit 1
+▁informationen 1
+▁kindergarten 1
+▁herunter 1
+brett 1
+▁kino 1
+▁traditionell 1
+▁ausdruck 1
+politik 1
+universität 1
+funktion 1
+▁zunge 1
+gerufen 1
+elli 1
+▁bezirks 1
+▁ol 1
+▁blei 1
+▁fass 1
+▁lügen 1
+▁gesprochen 1
+versuch 1
+▁offizielle 1
+▁kämpft 1
+durch 1
+▁bayern 1
+▁engagiert 1
+▁parallel 1
+▁intensiv 1
+▁brust 1
+▁richtet 1
+▁zivil 1
+▁bedeutendste 1
+▁sendung 1
+▁musikalische 1
+▁sterben 1
+▁otto 1
+▁sachsen 1
+▁sicherheit 1
+▁ufer 1
+▁handlung 1
+säule 1
+▁diesmal 1
+▁geschenk 1
+flu 1
+nimmt 1
+▁ortsteil 1
+▁spitz 1
+▁sub 1
+▁kabel 1
+▁historisch 1
+▁jugendliche 1
+▁bundesliga 1
+▁krankheit 1
+▁erfahrung 1
+lock 1
+▁wirkung 1
+rasse 1
+▁verhindern 1
+öffnung 1
+▁täglich 1
+zähne 1
+▁übersetzt 1
+▁praxis 1
+▁gebildet 1
+▁okay 1
+▁kaputt 1
+▁brüder 1
+▁blind 1
+▁zimmer 1
+wurf 1
+▁umwelt 1
+▁lösen 1
+▁kant 1
+struktur 1
+▁versehen 1
+barkeit 1
+▁nah 1
+park 1
+▁gründe 1
+▁irgendwa 1
+▁grenz 1
+▁feind 1
+liegenden 1
+▁absolut 1
+▁erklären 1
+▁parlament 1
+▁interessiert 1
+▁wunsch 1
+▁chance 1
+▁erscheint 1
+▁truppen 1
+▁kostet 1
+käse 1
+▁hotel 1
+ford 1
+rät 1
+magnet 1
+▁hass 1
+agent 1
+▁gegend 1
+▁fred 1
+▁szene 1
+geschäft 1
+kultur 1
+▁single 1
+europa 1
+▁muster 1
+▁brems 1
+▁code 1
+einheit 1
+gesetz 1
+▁aufgestellt 1
+▁tausend 1
+vel 1
+▁masse 1
+▁fähr 1
+▁grenzt 1
+▁element 1
+gesprochen 1
+▁gal 1
+▁hilfs 1
+▁auffällig 1
+▁thomas 1
+▁student 1
+lassung 1
+rezept 1
+sturz 1
+▁schnee 1
+▁common 1
+▁angaben 1
+▁lebensjahr 1
+▁hals 1
+▁strand 1
+haupt 1
+▁fu 1
+▁polar 1
+▁hinein 1
+▁verfassung 1
+▁talent 1
+zweck 1
+saal 1
+▁ausgesprochen 1
+▁fund 1
+schreiben 1
+▁wartet 1
+reiter 1
+▁maxim 1
+▁schick 1
+▁brun 1
+▁anzahl 1
+▁nutzen 1
+▁nu 1
+▁col 1
+holz 1
+bindung 1
+▁jahrzehnt 1
+veröffentlichung 1
+▁diskussion 1
+▁hannover 1
+▁hierfür 1
+▁netzwerk 1
+festival 1
+geschmack 1
+gewerbe 1
+▁teuer 1
+▁investi 1
+▁klick 1
+▁darstellung 1
+▁sprecher 1
+▁fällen 1
+gerissen 1
+forscher 1
+▁folgt 1
+▁endet 1
+▁ortes 1
+▁davor 1
+▁passend 1
+gewicht 1
+▁foto 1
+fleisch 1
+museum 1
+spitze 1
+▁jeweilige 1
+olf 1
+ziehung 1
+▁kompliziert 1
+▁existieren 1
+▁joseph 1
+▁lieblings 1
+erkennung 1
+ström 1
+rekord 1
+▁nahezu 1
+▁ausgabe 1
+gleis 1
+zellen 1
+weltmeister 1
+▁regie 1
+▁wiener 1
+euer 1
+bewegung 1
+▁kata 1
+▁werfen 1
+▁kräftig 1
+operation 1
+▁wollt 1
+gramm 1
+kanal 1
+▁lind 1
+▁leb 1
+geworfen 1
+läufig 1
+▁vergangenheit 1
+▁ausgebildet 1
+kandidat 1
+▁online 1
+▁umstritten 1
+▁angeblich 1
+▁vorgänger 1
+gebirge 1
+▁verliert 1
+stadion 1
+▁bekam 1
+fisch 1
+▁camp 1
+▁aufgeführt 1
+legung 1
+▁kulturelle 1
+▁ländern 1
+▁hergestellt 1
+▁übertragen 1
+▁lernte 1
+▁guck 1
+▁my 1
+bad 1
+▁erzählt 1
+lösung 1
+▁gitarre 1
+▁wachsen 1
+dreh 1
+ano 1
+▁erhöht 1
+▁geeignet 1
+▁knapp 1
+▁roboter 1
+▁william 1
+▁fieber 1
+konstruktion 1
+▁sowas 1
+▁angesehen 1
+▁kanton 1
+▁zweitausend 1
+anzeige 1
+glück 1
+block 1
+▁angegeben 1
+psycholog 1
+▁cool 1
+ländische 1
+▁produkte 1
+▁witt 1
+▁schauspielerin 1
+▁müller 1
+▁messer 1
+▁sammlung 1
+▁verantwortlich 1
+material 1
+▁pause 1
+▁johannes 1
+ausschuss 1
+ash 1
+▁machst 1
+getriebe 1
+rseits 1
+▁verhältnis 1
+▁medizinische 1
+schild 1
+▁wettbewerb 1
+ehlt 1
+▁renn 1
+▁geblieben 1
+händler 1
+▁besaß 1
+verbrauch 1
+krebs 1
+stehende 1
+filter 1
+▁bewegt 1
+▁dreck 1
+▁hintergrund 1
+▁dünn 1
+▁japan 1
+zeichnung 1
+beeren 1
+tausend 1
+bahnhof 1
+läufer 1
+▁betreten 1
+▁überlegen 1
+schließt 1
+kenntnis 1
+ction 1
+▁freude 1
+▁vorbild 1
+kind 1
+▁april 1
+▁jährlich 1
+▁apfel 1
+▁gefährlich 1
+merkmal 1
+▁detail 1
+▁ursprung 1
+metall 1
+▁bieten 1
+▁stink 1
+schuppen 1
+▁straßenbahn 1
+▁verteilt 1
+▁merken 1
+▁stahl 1
+▁angenommen 1
+stimme 1
+ministerium 1
+dank 1
+▁beiträge 1
+▁schmi 1
+wahl 1
+▁einrichtung 1
+▁union 1
+▁angelegt 1
+möglich 1
+minister 1
+geschrieben 1
+prozess 1
+▁riech 1
+▁ferner 1
+▁studie 1
+hagen 1
+▁job 1
+▁heide 1
+▁bewusst 1
+disziplin 1
+therapie 1
+▁begrenzt 1
+▁berühmt 1
+▁solang 1
+▁verbrachte 1
+▁weiblich 1
+▁besatzung 1
+▁dunkle 1
+▁zumeist 1
+▁widerstand 1
+▁zell 1
+siebzig 1
+schwer 1
+▁sturm 1
+▁schlange 1
+▁gründete 1
+▁eth 1
+ziel 1
+rup 1
+▁beine 1
+forschung 1
+▁akt 1
+▁gedacht 1
+garten 1
+malerei 1
+▁schmerz 1
+▁geöffnet 1
+▁respekt 1
+▁verändert 1
+ingenieur 1
+▁reagiert 1
+training 1
+▁erweitert 1
+▁perfekt 1
+▁christian 1
+▁wertvolle 1
+▁sorgt 1
+kalender 1
+▁antrieb 1
+otter 1
+▁christlich 1
+kenntnisse 1
+position 1
+temperatur 1
+figur 1
+marsch 1
+verfahren 1
+treue 1
+low 1
+cent 1
+▁turnier 1
+▁engel 1
+▁melde 1
+ight 1
+▁turn 1
+▁philosophie 1
+▁gelegen 1
+▁vulkan 1
+kapitän 1
+▁herkunft 1
+▁ärger 1
+▁verbrannt 1
+physik 1
+wälder 1
+▁schreck 1
+▁besiedelt 1
+▁paket 1
+▁gabriel 1
+▁irgendein 1
+▁behandelt 1
+▁planeten 1
+▁panzer 1
+▁oberfläche 1
+▁knie 1
+▁passieren 1
+spricht 1
+▁angestellt 1
+sperre 1
+gehalt 1
+▁aussehen 1
+thal 1
+▁programmier 1
+siedel 1
+unddreißig 1
+▁entwickeln 1
+▁kommunikation 1
+▁regisseur 1
+▁ungefähr 1
+▁kapital 1
+▁erinnern 1
+▁erkannt 1
+▁verehrt 1
+▁werbung 1
+▁butter 1
+▁wusste 1
+▁ehrlich 1
+gearbeitet 1
+▁barock 1
+▁stör 1
+▁berichtet 1
+▁zusammenarbeit 1
+ama 1
+▁schale 1
+▁betrachtet 1
+▁schieb 1
+▁schauspiel 1
+karriere 1
+legende 1
+▁umgeben 1
+luft 1
+▁karten 1
+besitz 1
+▁gall 1
+griffen 1
+mehr 1
+▁japanische 1
+▁sach 1
+gefühl 1
+braten 1
+▁verlauf 1
+▁bestätigt 1
+▁eröffnet 1
+▁asteroid 1
+▁urlaub 1
+magazin 1
+sprüche 1
+▁männlich 1
+▁scharf 1
+▁ursache 1
+▁riesige 1
+▁tourismus 1
+versicherung 1
+▁jurist 1
+▁kämpfen 1
+▁autobahn 1
+▁tank 1
+verständnis 1
+lampe 1
+typ 1
+▁bestandteil 1
+▁freiburg 1
+▁heiz 1
+pokal 1
+▁frühstück 1
+lücke 1
+fällt 1
+strecke 1
+förderung 1
+▁eingerichtet 1
+empfind 1
+beruf 1
+ebene 1
+strahlung 1
+▁unterstützung 1
+▁pferd 1
+▁gesund 1
+gouverneur 1
+revolution 1
+▁beeinflusst 1
+geräusch 1
+schleunig 1
+▁dürfte 1
+▁geschwister 1
+▁gewöhn 1
+▁anonym 1
+▁löwe 1
+▁versteckt 1
+▁unklar 1
+▁bevorzugt 1
+jäger 1
+▁verlust 1
+▁rückwärts 1
+▁behörden 1
+▁kriegsende 1
+▁schlug 1
+▁schwanz 1
+fix 1
+leute 1
+▁trick 1
+▁med 1
+truppe 1
+tragen 1
+▁damen 1
+glocke 1
+▁doppelt 1
+röhre 1
+herausgeber 1
+▁wirken 1
+partei 1
+ulation 1
+freund 1
+▁live 1
+▁heil 1
+phon 1
+▁histor 1
+▁rö 1
+▁täusch 1
+▁dreißig 1
+▁demokratische 1
+▁dialekt 1
+▁empfehle 1
+▁gipfel 1
+trunken 1
+▁christoph 1
+▁hieß 1
+▁verlegt 1
+früchte 1
+▁lustig 1
+▁enorm 1
+▁pilot 1
+▁schall 1
+▁touristen 1
+▁ordentlich 1
+▁pfeif 1
+▁erscheinen 1
+▁rau 1
+▁leisten 1
+▁putz 1
+füge 1
+▁nachweis 1
+tasche 1
+ographie 1
+▁traf 1
+lanz 1
+▁auszeichnung 1
+▁bonn 1
+koch 1
+▁löst 1
+▁wichtiger 1
+maß 1
+▁künstlerische 1
+feuer 1
+breite 1
+kohl 1
+kosten 1
+ō 1
+▁gespannt 1
+▁übernommen 1
+▁eindeutig 1
+▁erforderlich 1
+▁negativ 1
+▁bedroh 1
+▁verpflichtet 1
+▁schatten 1
+▁abitur 1
+▁erwarten 1
+▁tunnel 1
+▁lecker 1
+▁keller 1
+▁chemische 1
+fressen 1
+▁überzeugen 1
+▁grundlage 1
+▁grand 1
+▁indien 1
+▁untersucht 1
+▁weltkrieges 1
+rmaßen 1
+▁dahin 1
+▁lob 1
+schritt 1
+ney 1
+▁nutzung 1
+▁tom 1
+▁jean 1
+regierung 1
+unternehmen 1
+werfen 1
+gemacht 1
+litz 1
+▁präg 1
+zeichnen 1
+▁höchst 1
+▁endgültig 1
+▁experiment 1
+▁israel 1
+▁olympische 1
+▁zugänglich 1
+schätzen 1
+vermögen 1
+▁richard 1
+▁wüsste 1
+▁david 1
+▁arabisch 1
+▁dialog 1
+▁nütz 1
+▁donau 1
+▁unterscheiden 1
+▁giftig 1
+▁hierzu 1
+▁sauber 1
+▁münz 1
+▁orgel 1
+▁anstrengend 1
+möglichkeiten 1
+▁abschnitt 1
+▁aufgebaut 1
+▁walter 1
+▁unfall 1
+▁kürz 1
+▁gemeint 1
+▁bord 1
+weich 1
+lach 1
+▁objekt 1
+ju 1
+▁wich 1
+▁drin 1
+sammlung 1
+aktiv 1
+produzent 1
+▁begonnen 1
+▁professionelle 1
+▁teufel 1
+▁gewähr 1
+▁drücken 1
+▁irgendwelche 1
+▁flagge 1
+▁klima 1
+▁anhänger 1
+▁konzern 1
+großvater 1
+▁mehrmals 1
+▁komposition 1
+▁brut 1
+achse 1
+tempo 1
+▁laufbahn 1
+archiv 1
+▁kollegen 1
+streifen 1
+▁mission 1
+▁beträgt 1
+▁wellen 1
+▁niederländisch 1
+▁schweden 1
+▁gemeindegebiet 1
+▁wall 1
+▁bislang 1
+▁interpret 1
+▁sender 1
+gift 1
+liegende 1
+▁elektrische 1
+▁kap 1
+graph 1
+sozialismus 1
+illa 1
+würdig 1
+initiative 1
+komplex 1
+▁konkurrenz 1
+▁tübingen 1
+▁verfolgt 1
+▁verurteilt 1
+▁gemüse 1
+▁langfristig 1
+▁scheiterte 1
+schätzung 1
+▁show 1
+▁froh 1
+▁jazz 1
+▁verlieren 1
+schlägt 1
+begründer 1
+▁atom 1
+▁ernsthaft 1
+weih 1
+▁begleitet 1
+▁werbe 1
+schirm 1
+▁gefällt 1
+schwäche 1
+miete 1
+▁bauwerk 1
+uzi 1
+▁kara 1
+▁zeige 1
+▁schuh 1
+region 1
+▁nehme 1
+▁mehrheit 1
+stunde 1
+weisung 1
+▁bemüh 1
+▁düsseldorf 1
+▁weibchen 1
+▁gattung 1
+▁geschehen 1
+▁ändern 1
+▁beherrsch 1
+▁wurzel 1
+▁bedient 1
+▁gelernt 1
+▁erlebt 1
+▁vorsitzender 1
+▁heiratete 1
+▁kirsch 1
+▁botschaft 1
+▁flori 1
+▁amtszeit 1
+▁realität 1
+uppe 1
+▁carl 1
+▁taschen 1
+rald 1
+gestaltung 1
+▁total 1
+▁kopie 1
+▁sperr 1
+wendung 1
+▁stimm 1
+wirkt 1
+ella 1
+lauch 1
+blatt 1
+friedhof 1
+tätigkeit 1
+▁bock 1
+▁senat 1
+seil 1
+▁gefärbt 1
+▁religiöse 1
+▁verteidigung 1
+▁orientiert 1
+▁philipp 1
+assistent 1
+▁rettungs 1
+▁scheinbar 1
+▁schatz 1
+▁billig 1
+▁jagd 1
+▁wäsch 1
+▁ideal 1
+▁kommentar 1
+▁aufmerksamkeit 1
+▁erwies 1
+▁kalk 1
+übung 1
+▁alarm 1
+▁streng 1
+▁kindheit 1
+▁herstellung 1
+▁geschie 1
+nicht 1
+▁analyse 1
+▁juden 1
+fassade 1
+▁übersetzung 1
+gärten 1
+zucht 1
+istik 1
+entwicklung 1
+▁sack 1
+versammlung 1
+hoff 1
+▁beruflich 1
+tief 1
+frequenz 1
+gegenwärtig 1
+▁restaurant 1
+▁empfang 1
+▁erwähnt 1
+▁verhindert 1
+▁weihnacht 1
+geschaltet 1
+▁russland 1
+▁sauerstoff 1
+konferenz 1
+▁ergibt 1
+▁audio 1
+▁trifft 1
+▁krankenhaus 1
+verleih 1
+▁brandenburg 1
+▁city 1
+▁spanien 1
+▁wiederholt 1
+▁schlagzeug 1
+schäden 1
+▁unsicher 1
+losigkeit 1
+▁vierzig 1
+▁führend 1
+zustellen 1
+▁auswahl 1
+knoten 1
+ladung 1
+steht 1
+ville 1
+ji 1
+▁gestalt 1
+zeile 1
+institut 1
+transport 1
+trainer 1
+▁blüte 1
+pause 1
+schädigt 1
+▁atmosphäre 1
+▁einkaufsliste 1
+▁geklärt 1
+▁großbritannien 1
+▁stattfinden 1
+gewässer 1
+▁experten 1
+▁preußische 1
+▁chinesische 1
+stütze 1
+▁gefertigt 1
+▁theologie 1
+▁gefühle 1
+▁verlag 1
+▁notfall 1
+▁öffentlichkeit 1
+▁vize 1
+▁extra 1
+mensch 1
+route 1
+▁münd 1
+▁montag 1
+▁mikro 1
+▁bedenken 1
+▁meeres 1
+▁türkis 1
+▁schulter 1
+▁rosen 1
+balken 1
+phase 1
+nteils 1
+vertrag 1
+▁eb 1
+▁benachbarte 1
+▁alexander 1
+▁bedeuten 1
+▁party 1
+▁toast 1
+▁fungiert 1
+▁wespe 1
+oxid 1
+▁probier 1
+darsteller 1
+▁erfunden 1
+▁locker 1
+▁amerikaner 1
+▁saint 1
+scheibe 1
+▁schied 1
+▁offiziell 1
+borg 1
+▁hauptort 1
+▁faul 1
+▁herein 1
+▁lungen 1
+fragt 1
+▁kuh 1
+▁mast 1
+furt 1
+▁att 1
+nationalpark 1
+▁disk 1
+▁steil 1
+▁brauche 1
+▁cap 1
+q 1
+version 1
+▁etablier 1
+▁beschränkt 1
+▁gewonnen 1
+▁erfordert 1
+▁aufgelöst 1
+▁tabelle 1
+▁gieß 1
+▁kühlschrank 1
+▁schwedische 1
+▁zuschauer 1
+▁unterteilt 1
+▁bisherige 1
+wirt 1
+▁rathaus 1
+▁hügel 1
+▁albert 1
+▁schneider 1
+▁fair 1
+▁funktionieren 1
+stiel 1
+wieso 1
+▁isa 1
+mitglieder 1
+▁klang 1
+iffer 1
+logisch 1
+▁opa 1
+rannt 1
+▁griff 1
+▁lehnte 1
+salat 1
+gekommen 1
+▁luther 1
+studium 1
+fehler 1
+album 1
+ässe 1
+begriff 1
+omi 1
+▁anerkannt 1
+▁identität 1
+▁komponiert 1
+▁sämtliche 1
+▁zerstören 1
+▁erinnerung 1
+▁islam 1
+▁vorgesehen 1
+▁autonom 1
+▁steigt 1
+▁angeles 1
+hüt 1
+▁architektur 1
+cia 1
+▁nachhaltig 1
+rechnet 1
+▁schuf 1
+▁handwerk 1
+▁villa 1
+kabel 1
+▁dreimal 1
+▁nick 1
+▁antike 1
+▁kiel 1
+organ 1
+zustimmen 1
+▁matt 1
+fahrzeug 1
+republik 1
+ruhe 1
+macher 1
+bühne 1
+▁augsburg 1
+▁böhm 1
+▁stirbt 1
+▁tastatur 1
+▁verfügbar 1
+▁dringend 1
+▁nahrung 1
+mühle 1
+▁deutschsprachig 1
+▁strömung 1
+▁bremen 1
+bemerkt 1
+▁droge 1
+▁dahinter 1
+durchschnittlich 1
+▁selbstständig 1
+▁hoffe 1
+▁franc 1
+▁alkohol 1
+gummi 1
+▁river 1
+▁moor 1
+▁kritisch 1
+soße 1
+▁unterlag 1
+▁empfi 1
+▁rosa 1
+hauen 1
+▁north 1
+firma 1
+▁menschliche 1
+▁beobachtet 1
+▁mathematik 1
+▁olympia 1
+▁südöstlich 1
+▁vierzehn 1
+▁batterie 1
+▁umgangssprach 1
+▁zusammengefasst 1
+▁erwartet 1
+lärm 1
+▁präsentier 1
+▁vermutet 1
+debüt 1
+▁aufnehmen 1
+▁email 1
+▁stürm 1
+▁onkel 1
+▁glatt 1
+▁gelingt 1
+▁stefan 1
+▁geriet 1
+▁aufpassen 1
+▁eingeführt 1
+▁königreich 1
+▁leiche 1
+▁sankt 1
+gebunden 1
+pflege 1
+▁ego 1
+rwerb 1
+▁marine 1
+patri 1
+▁kaufmann 1
+▁saure 1
+bett 1
+grä 1
+modul 1
+▁schwan 1
+kopf 1
+altar 1
+zylinder 1
+▁akzeptier 1
+trocknet 1
+▁schmecken 1
+▁toilette 1
+▁umgekehrt 1
+▁wählt 1
+gelehnt 1
+▁örtliche 1
+▁überliefert 1
+▁göttingen 1
+erlebnis 1
+▁sekunden 1
+▁schwierigkeiten 1
+▁kommunist 1
+▁jakob 1
+▁prinz 1
+▁schläf 1
+▁auseinandersetz 1
+angehörige 1
+▁runter 1
+▁golf 1
+▁ausgeführt 1
+geübt 1
+▁wahrheit 1
+▁schief 1
+▁jahreszeit 1
+▁saß 1
+elemente 1
+▁bundestag 1
+vereinigung 1
+▁spenden 1
+▁dazwischen 1
+▁hinweis 1
+▁nordamerika 1
+krise 1
+rlei 1
+kraut 1
+problem 1
+hafen 1
+bibliothek 1
+milch 1
+flossen 1
+▁sende 1
+▁akku 1
+bilität 1
+š 1
+▁achtzehn 1
+▁ansässig 1
+▁aufklär 1
+▁engagement 1
+▁interessant 1
+▁smartphone 1
+▁oktober 1
+▁umfangreiche 1
+▁ehrenamtlich 1
+▁nahrungsmittel 1
+▁wagner 1
+▁draußen 1
+▁wachstum 1
+▁erlitt 1
+räder 1
+▁wohnzimmer 1
+▁wolken 1
+▁emil 1
+▁spezialisiert 1
+▁puste 1
+▁persönlichkeit 1
+▁zwangs 1
+▁setze 1
+haufen 1
+▁zack 1
+▁hinsicht 1
+▁staatliche 1
+▁dortige 1
+▁dieselbe 1
+▁bohr 1
+aufnahmen 1
+▁russe 1
+▁chip 1
+▁erbaut 1
+▁schwä 1
+▁finanziert 1
+watt 1
+materie 1
+▁gesetzlich 1
+grill 1
+▁begleit 1
+▁bringe 1
+kommando 1
+erkrankung 1
+▁betreiben 1
+▁geschaffen 1
+▁reduziert 1
+▁völker 1
+▁würz 1
+▁zeichnet 1
+▁zuverlässig 1
+container 1
+▁literarische 1
+▁risiko 1
+▁rudolf 1
+▁gustav 1
+▁bayerische 1
+▁unterscheidet 1
+▁rätsel 1
+▁benutzen 1
+▁dresden 1
+▁verletzt 1
+▁geburtstag 1
+quartier 1
+▁eigenschaften 1
+▁verbrenn 1
+periode 1
+▁dreißigjährige 1
+▁bildschirm 1
+dynam 1
+▁ahnung 1
+▁ferien 1
+▁vieh 1
+▁vorlesen 1
+▁sinnvoll 1
+▁bleib 1
+winkel 1
+▁geschossen 1
+▁streu 1
+▁schraub 1
+kolleg 1
+▁jura 1
+▁bereitet 1
+▁mittelalterlich 1
+▁wohin 1
+▁schönste 1
+▁säu 1
+tafel 1
+strafe 1
+▁ruine 1
+blitz 1
+bedingung 1
+▁satelliten 1
+▁verstärkt 1
+▁geduld 1
+▁zürich 1
+▁fundament 1
+▁offensichtlich 1
+▁begraben 1
+▁insekten 1
+schaffung 1
+▁knall 1
+▁astronom 1
+▁nordosten 1
+▁schicksal 1
+▁gedreht 1
+▁analog 1
+▁ausreichend 1
+uddel 1
+kuppel 1
+▁besichtig 1
+zettel 1
+trennung 1
+▁island 1
+▁rückkehr 1
+▁klingen 1
+▁elektronische 1
+▁freundschaft 1
+▁partie 1
+▁momentan 1
+▁ausrede 1
+▁haushalt 1
+dreieck 1
+▁adler 1
+gehört 1
+bricht 1
+▁fürchte 1
+▁epi 1
+läufe 1
+▁satt 1
+verhältnisse 1
+bereitschaft 1
+schieben 1
+▁automat 1
+krone 1
+zusetzen 1
+reinigung 1
+▁rind 1
+▁getötet 1
+▁luxemburg 1
+▁nürnberg 1
+▁quadrat 1
+▁verursacht 1
+▁firmen 1
+▁besorgt 1
+▁justus 1
+▁betriebssystem 1
+▁überlegt 1
+▁spür 1
+▁freut 1
+konkret 1
+▁opposition 1
+▁inhaltlich 1
+▁lohnt 1
+▁bedingungen 1
+▁hauptrolle 1
+faktor 1
+▁speisen 1
+tonne 1
+▁aufhören 1
+▁mineral 1
+▁gleichmäßig 1
+▁niederlande 1
+▁human 1
+▁hinterher 1
+▁großteil 1
+▁sorte 1
+fenster 1
+taste 1
+schutzgebiet 1
+schmutz 1
+fühl 1
+▁schaf 1
+▁bunt 1
+ündig 1
+konzentration 1
+verschlüsselung 1
+▁beliebig 1
+▁fängt 1
+▁fünfzehn 1
+▁gedächtnis 1
+▁global 1
+▁maßgeblich 1
+▁verantwortung 1
+▁verliebt 1
+verfolgung 1
+verständlich 1
+▁begründet 1
+▁behindert 1
+▁stieß 1
+▁erfüllt 1
+▁garantiert 1
+▁behaart 1
+▁nachmittag 1
+▁gegenwart 1
+funk 1
+▁priester 1
+▁gesichert 1
+▁gekocht 1
+▁voraussetzung 1
+▁geistes 1
+verbände 1
+▁drehbuch 1
+bräu 1
+▁bühnen 1
+▁erhoben 1
+greift 1
+▁fell 1
+▁baustelle 1
+▁etliche 1
+jenige 1
+▁hessen 1
+boot 1
+▁mix 1
+▁mega 1
+füllen 1
+manager 1
+schub 1
+▁verläss 1
+▁entsprechen 1
+▁nordwest 1
+stange 1
+gezeichnet 1
+gedrückt 1
+▁aussieht 1
+▁kalifornien 1
+kompetenz 1
+▁populär 1
+beschreibung 1
+▁reagieren 1
+▁angesiedelt 1
+▁gewebe 1
+kommission 1
+▁neugier 1
+▁erhältlich 1
+▁verwechselt 1
+▁installiert 1
+▁aachen 1
+▁dutzend 1
+▁herrscher 1
+▁bedeckt 1
+▁kirchliche 1
+▁kanon 1
+plastik 1
+tomaten 1
+▁naturpark 1
+▁verkünd 1
+▁german 1
+üchtig 1
+▁stress 1
+▁hirn 1
+kämpft 1
+hochschule 1
+▁heirat 1
+▁flüsse 1
+▁alex 1
+▁schweigen 1
+▁reformation 1
+▁mündet 1
+▁luc 1
+▁papa 1
+spor 1
+korps 1
+fünfzig 1
+▁ereignisse 1
+▁langjährig 1
+▁kongress 1
+beitrag 1
+▁führ 1
+gasse 1
+▁profil 1
+ografie 1
+flüge 1
+▁nett 1
+▁unterlieg 1
+▁synchron 1
+▁entschuldige 1
+▁erklärung 1
+▁exemplar 1
+▁joachim 1
+▁lokomotiv 1
+▁records 1
+▁synonym 1
+▁verzweifelt 1
+bezieht 1
+gesperrt 1
+▁atlant 1
+▁australien 1
+▁geändert 1
+▁fliegt 1
+▁öfter 1
+▁behauptet 1
+▁erwartung 1
+▁hobby 1
+▁tempel 1
+▁award 1
+▁stolz 1
+▁überlebt 1
+▁royal 1
+▁pfanne 1
+▁nannte 1
+▁schnitz 1
+▁trennt 1
+klinik 1
+▁unterwegs 1
+▁bestritt 1
+▁fraktion 1
+böden 1
+▁bezahlt 1
+▁informations 1
+hänge 1
+sekretär 1
+▁einheimische 1
+frau 1
+▁fließend 1
+▁tja 1
+bezeichnung 1
+▁litt 1
+strom 1
+▁nordwestlich 1
+▁fruchtbar 1
+▁gräb 1
+▁qualifiziert 1
+damm 1
+shop 1
+röte 1
+frist 1
+▁amerika 1
+gängig 1
+konservativ 1
+perspektive 1
+▁integriert 1
+▁reparatur 1
+▁september 1
+▁vielfältig 1
+intelligenz 1
+▁befestigt 1
+▁großmutter 1
+▁kontinent 1
+▁reserve 1
+▁sächsisch 1
+▁begeistert 1
+▁diskutiert 1
+▁virtu 1
+▁weimar 1
+▁vorhaben 1
+substanz 1
+▁vergiss 1
+▁disco 1
+▁fortschritt 1
+▁teller 1
+▁flüssigkeit 1
+▁faust 1
+▁römer 1
+▁hype 1
+kapitel 1
+▁geistliche 1
+▁salzburg 1
+▁manuell 1
+▁andererseits 1
+▁wolfgang 1
+modus 1
+korrektur 1
+kunst 1
+▁liefert 1
+▁schützt 1
+wanderung 1
+▁acker 1
+säge 1
+organisation 1
+flugzeug 1
+▁junior 1
+ausbildung 1
+geräumt 1
+herrschaft 1
+ersdorf 1
+antwortet 1
+objektiv 1
+▁vergleichbar 1
+streit 1
+brille 1
+▁santa 1
+▁lernt 1
+analys 1
+zentrale 1
+▁thi 1
+regiment 1
+rgebnis 1
+syndrom 1
+▁abkürzung 1
+▁beobachten 1
+▁dreizehn 1
+▁gmbh 1
+▁publikation 1
+▁nachwuchs 1
+▁beschreibt 1
+▁patienten 1
+▁teppich 1
+▁ernährung 1
+▁kommunal 1
+▁kredit 1
+▁denkmalschutz 1
+▁josef 1
+▁moderator 1
+▁pianist 1
+▁magnetfeld 1
+▁übernehmen 1
+zügig 1
+mittwoch 1
+▁bequem 1
+sponsor 1
+▁wahrnehm 1
+▁school 1
+▁nebenwirkungen 1
+seriöse 1
+gefordert 1
+▁geweiht 1
+▁ausstattung 1
+▁rundfunk 1
+▁klischee 1
+pflaster 1
+▁graz 1
+▁erlangte 1
+▁vermittelt 1
+▁scheiben 1
+▁eigentum 1
+▁hauptsitz 1
+wettbewerb 1
+prinzip 1
+schwein 1
+viertel 1
+▁schaffe 1
+besuch 1
+aufgaben 1
+▁lehn 1
+▁falt 1
+▁kurve 1
+▁südamerika 1
+▁biologi 1
+molekül 1
+é 1
+▁aspekt 1
+▁gefördert 1
+▁hardware 1
+▁phänomen 1
+▁pseudonym 1
+▁definiert 1
+▁effizient 1
+▁fliehen 1
+▁günstig 1
+▁mecklenburg 1
+▁symptom 1
+träglich 1
+▁betätigt 1
+▁einzigartige 1
+▁dramatisch 1
+▁breslau 1
+▁sonntag 1
+▁deutet 1
+▁harmlos 1
+▁papst 1
+▁jemals 1
+▁durfte 1
+▁ehefrau 1
+▁porträt 1
+▁übt 1
+▁kleinkind 1
+▁verwaltungssitz 1
+▁bezahlen 1
+▁anzeigen 1
+▁volksschule 1
+mischung 1
+▁grundstück 1
+wolke 1
+cto 1
+pfad 1
+quote 1
+gewandt 1
+kapsel 1
+erfolg 1
+▁ortsname 1
+wärme 1
+▁betrachte 1
+▁eishockey 1
+▁nachricht 1
+▁zoo 1
+▁juni 1
+▁empfohlen 1
+▁facebook 1
+▁hinterließ 1
+▁niedersachsen 1
+▁stellvertretend 1
+▁verlängerung 1
+▁zeitgenössische 1
+▁kritisiert 1
+▁skulptur 1
+mittlung 1
+▁großeltern 1
+▁schräg 1
+▁karlsruhe 1
+▁verstopf 1
+▁versorgt 1
+▁zauber 1
+▁monument 1
+▁stockwerk 1
+▁grafschaft 1
+▁grundriss 1
+▁etappe 1
+▁verdanken 1
+pulver 1
+▁kolonie 1
+▁antrag 1
+▁bernhard 1
+▁high 1
+geflogen 1
+▁adolf 1
+▁gelöst 1
+▁arbeitgeber 1
+▁erfinde 1
+▁bundesstaat 1
+▁kult 1
+▁circ 1
+▁eiserne 1
+▁spinne 1
+▁ausgangs 1
+▁liest 1
+▁beraten 1
+galerie 1
+▁enttäusch 1
+patron 1
+work 1
+▁photo 1
+▁psycho 1
+▁basketball 1
+▁demokratie 1
+▁ergänzt 1
+▁eventuell 1
+▁herkömmliche 1
+▁inspiriert 1
+▁jürgen 1
+▁konstruiert 1
+▁lübeck 1
+▁moskau 1
+▁november 1
+▁tschechisch 1
+▁tödlich 1
+▁verdächtig 1
+▁zentimeter 1
+▁verlängert 1
+▁verschlüsselt 1
+▁adelsgeschlecht 1
+▁alfred 1
+▁ventil 1
+▁verblieb 1
+▁bäcker 1
+▁ticket 1
+▁terror 1
+▁radikal 1
+▁vorbereitung 1
+märkte 1
+▁kooperation 1
+▁herausforderung 1
+▁gelungen 1
+▁selbständig 1
+▁übel 1
+▁piep 1
+▁dienstag 1
+▁finanzierung 1
+▁städtische 1
+tronom 1
+▁touristisch 1
+abhängig 1
+kiste 1
+entsprechend 1
+konzert 1
+bringt 1
+erfahrung 1
+beute 1
+fusion 1
+▁philosoph 1
+liebe 1
+birne 1
+charts 1
+asiatische 1
+oskop 1
+distan 1
+▁feucht 1
+▁verhäng 1
+▁verbreitung 1
+▁sympathi 1
+ärztin 1
+▁ausgewählt 1
+▁browser 1
+▁diplomat 1
+▁intelligent 1
+▁konzentriert 1
+▁majestät 1
+▁medikament 1
+▁rechtsanwalt 1
+▁spaziergang 1
+▁territor 1
+▁vergnüg 1
+ifikation 1
+▁abseits 1
+▁drohne 1
+▁linux 1
+▁gelehrt 1
+▁rühr 1
+▁akzent 1
+▁spannend 1
+▁verstecken 1
+▁anbieten 1
+▁bestattet 1
+▁rechner 1
+▁goethe 1
+▁aktivitäten 1
+benutzung 1
+▁nervös 1
+▁hirsch 1
+▁mittelmeer 1
+▁schmidt 1
+▁schleim 1
+▁boiler 1
+management 1
+▁schiller 1
+▁erwarb 1
+▁behaupten 1
+▁abriss 1
+▁rüber 1
+▁passwort 1
+stedt 1
+pfalz 1
+▁janni 1
+aufenthalt 1
+sozialistische 1
+könig 1
+vertreter 1
+saison 1
+artikel 1
+messer 1
+töne 1
+ritten 1
+▁irgend 1
+tournee 1
+busse 1
+smitglied 1
+▁antoni 1
+sparen 1
+charakteristisch 1
+schießen 1
+wanderweg 1
+gräber 1
+bremse 1
+national 1
+zunehmen 1
+dynastie 1
+getränke 1
+verkürz 1
+▁beinhaltet 1
+▁bemerkenswert 1
+▁dimension 1
+▁dänemark 1
+▁eigentümer 1
+▁elisabeth 1
+▁gültig 1
+▁konsequent 1
+▁nirgend 1
+▁potsdam 1
+▁spaghetti 1
+▁verstorben 1
+kombination 1
+▁spontan 1
+▁world 1
+▁väter 1
+▁sparkasse 1
+▁kanada 1
+▁trainiert 1
+▁vorbereitet 1
+▁reihenfolge 1
+▁betreut 1
+▁aufgewachsen 1
+lädt 1
+▁chemie 1
+▁konstant 1
+▁annehmen 1
+▁multi 1
+▁klingel 1
+geliefert 1
+seminar 1
+▁labor 1
+▁nadel 1
+badenwürttemberg 1
+trommel 1
+▁daniel 1
+▁knack 1
+▁redaktion 1
+▁umständen 1
+▁aqua 1
+▁sachgebiet 1
+▁kleider 1
+▁gedruckt 1
+▁überschreite 1
+▁james 1
+flieger 1
+▁pirate 1
+soeben 1
+▁könne 1
+plattform 1
+kontrolle 1
+▁eiche 1
+entwurf 1
+▁aufmerksam 1
+▁flüssig 1
+schränkung 1
+zählig 1
+▁dörfer 1
+▁fantasie 1
+▁februar 1
+▁infrastruktur 1
+▁sowjetunion 1
+▁tshirt 1
+▁zwiebeln 1
+▁google 1
+▁gänz 1
+▁heidelberg 1
+▁herausragend 1
+▁qualifikation 1
+▁eduard 1
+▁pizza 1
+▁betrug 1
+▁hoffnung 1
+▁justiz 1
+schwangerschaft 1
+▁empfäng 1
+▁konsequenzen 1
+▁zitrone 1
+▁linear 1
+▁lorenz 1
+▁verstarb 1
+vertretung 1
+▁vorsichtig 1
+▁beleuchtet 1
+hydro 1
+▁heftig 1
+▁klappt 1
+▁humor 1
+▁unsichtbar 1
+▁würfel 1
+▁gezielt 1
+▁draht 1
+päck 1
+▁witzig 1
+▁erzeugt 1
+gewinde 1
+krankheit 1
+instrument 1
+fünf 1
+gesagt 1
+▁elektrisch 1
+oberseite 1
+▁fließ 1
+rahm 1
+▁tasse 1
+erfreu 1
+fänglich 1
+sehenswürdigkeit 1
+▁entstehung 1
+fernbedienung 1
+färbung 1
+schädigung 1
+▁akustisch 1
+▁brasilien 1
+▁ferdinand 1
+▁gewürz 1
+▁league 1
+▁nikolaus 1
+▁schokolade 1
+▁südafrika 1
+▁verbessert 1
+▁vernünftig 1
+▁konvention 1
+▁niveau 1
+▁scheitern 1
+▁rechtschreib 1
+▁erfordern 1
+▁margaret 1
+▁identisch 1
+▁waschmaschine 1
+▁buchstaben 1
+▁advent 1
+▁dauerhaft 1
+▁fernsehserie 1
+▁gegessen 1
+▁verhaftet 1
+▁peinlich 1
+▁geheimnis 1
+▁pilze 1
+▁wackel 1
+polizist 1
+▁ruhig 1
+▁geschnitten 1
+▁grafik 1
+knecht 1
+▁bewohnt 1
+pumpe 1
+mängel 1
+▁vortrag 1
+▁atem 1
+▁vertrat 1
+▁kürzeste 1
+wurst 1
+regnet 1
+politische 1
+wood 1
+angriff 1
+▁mächtig 1
+zähler 1
+wohngemeinschaft 1
+ärchen 1
+▁kicke 1
+sprachig 1
+synthese 1
+ömischkatholisch 1
+▁amazon 1
+▁automobil 1
+▁mathematisch 1
+▁orientierung 1
+▁ungeklärt 1
+▁ungewöhnlich 1
+▁unterschrieb 1
+▁zukünftig 1
+▁eignet 1
+▁schallplatten 1
+▁anerkennen 1
+▁reparier 1
+▁belgien 1
+▁pilger 1
+▁teilnehmen 1
+▁freuen 1
+▁bewahre 1
+▁verzehr 1
+▁kakao 1
+▁white 1
+▁märz 1
+▁nachfolgend 1
+▁unterstützen 1
+▁herbst 1
+▁motorräder 1
+▁angler 1
+▁kriminal 1
+verarbeitung 1
+breitungsgebiet 1
+▁studiert 1
+amerikanische 1
+dokumentation 1
+bestimmt 1
+huhn 1
+▁gefolg 1
+trink 1
+▁teure 1
+berater 1
+diskret 1
+geholt 1
+kragen 1
+ensemble 1
+rdrheinwestfalen 1
+▁akademische 1
+▁begegnung 1
+▁entfernung 1
+▁erledigt 1
+▁erleichtert 1
+▁fakultät 1
+▁griechenland 1
+▁kolonial 1
+▁kontinuierlich 1
+▁nordöstlich 1
+▁vegetarier 1
+▁verbleib 1
+▁mozilla 1
+▁abenteuer 1
+▁permanent 1
+▁stängel 1
+▁getreide 1
+▁kurzfristig 1
+▁website 1
+▁erweiterung 1
+▁genügt 1
+▁gelöscht 1
+▁mahlzeit 1
+matrix 1
+wählen 1
+▁liberal 1
+▁schnabel 1
+▁veranstaltet 1
+einflüsse 1
+▁zubereitet 1
+▁letztendlich 1
+▁germanistik 1
+▁louis 1
+▁generell 1
+▁beurteil 1
+treffer 1
+▁erdbeer 1
+▁asien 1
+geschichtliche 1
+▁schmied 1
+wurz 1
+▁mobil 1
+▁archäologische 1
+▁schwinge 1
+▁eingriff 1
+zange 1
+▁stilistisch 1
+kartoffeln 1
+freude 1
+güter 1
+▁lüge 1
+brunn 1
+ausweis 1
+▁kartoffel 1
+ropf 1
+▁röntgen 1
+▁illustr 1
+▁zigaretten 1
+▁grammatik 1
+▁attraktiv 1
+▁bayreuth 1
+▁beherbergt 1
+▁bewerbung 1
+▁bildhauer 1
+▁erwünscht 1
+▁gewidmet 1
+▁ländliche 1
+▁wikipedia 1
+▁wünscht 1
+▁zeigefinger 1
+erhöhung 1
+ometrische 1
+▁verabschied 1
+▁amateur 1
+▁oldenburg 1
+▁united 1
+▁rohstoffe 1
+▁appetit 1
+▁kläger 1
+▁börse 1
+▁erwischt 1
+▁unschuld 1
+▁quasi 1
+▁zeitweilig 1
+▁ämter 1
+▁adresse 1
+▁major 1
+▁uboote 1
+▁violin 1
+▁personalausweis 1
+▁jetzige 1
+▁grafikkarte 1
+▁verbiete 1
+▁verdammt 1
+▁leiser 1
+▁kälte 1
+▁mieter 1
+▁mutig 1
+gebrannt 1
+▁gemein 1
+lässlich 1
+bedarf 1
+künstler 1
+vorlesung 1
+▁geschützt 1
+sammeln 1
+schaltung 1
+▁begab 1
+▁muttersprach 1
+▁anlass 1
+schwimmer 1
+bänder 1
+▁aggressiv 1
+▁argentinien 1
+▁betreibt 1
+▁emotionale 1
+▁entworfen 1
+▁ergänzen 1
+▁evolution 1
+▁günther 1
+▁joghurt 1
+▁kathedrale 1
+▁kraftausdrücke 1
+▁offenkundig 1
+▁sebastian 1
+▁zwilling 1
+▁alphabet 1
+▁differenz 1
+▁verwaltet 1
+▁angeklagt 1
+horizont 1
+▁löcher 1
+▁review 1
+▁verbinden 1
+▁fulda 1
+▁kabinett 1
+▁überregional 1
+▁osmanische 1
+▁wiederaufbau 1
+▁ozean 1
+entwickler 1
+▁gewerkschaft 1
+▁umzugehen 1
+▁abgedeckt 1
+▁abgeschnitten 1
+▁folter 1
+▁windows 1
+▁tschüss 1
+▁smith 1
+▁dominiert 1
+▁erreichbar 1
+▁ungerecht 1
+▁moralisch 1
+▁servier 1
+▁samstag 1
+▁systematisch 1
+▁zitter 1
+▁behälter 1
+▁marmor 1
+▁heinz 1
+▁half 1
+▁neutral 1
+▁besiegt 1
+▁falk 1
+▁fahne 1
+sterblich 1
+▁träume 1
+▁lackier 1
+fuchs 1
+▁neffe 1
+eintrag 1
+sozialisten 1
+▁spitzname 1
+gestrichen 1
+korb 1
+kritik 1
+größte 1
+gedacht 1
+▁kippe 1
+▁hexe 1
+▁infanterie 1
+württemberg 1
+dimensional 1
+explosion 1
+rhythmus 1
+▁donnerstag 1
+▁entspannt 1
+▁fütter 1
+▁greifswald 1
+▁herauszufinden 1
+▁hollywood 1
+▁individuell 1
+▁relevant 1
+▁sechzehn 1
+▁verbesserung 1
+▁verbindet 1
+▁wohlhabend 1
+▁aufsichtsrat 1
+▁mexiko 1
+▁portugal 1
+westfälische 1
+▁klemmt 1
+▁beruht 1
+▁metropol 1
+▁brauchst 1
+▁wiesbaden 1
+▁gemischt 1
+▁bananen 1
+▁produktiv 1
+▁klammern 1
+▁töten 1
+pfeffer 1
+▁lastwagen 1
+▁schwäbisch 1
+▁tirol 1
+▁verarbeitet 1
+▁gehabt 1
+▁tennis 1
+▁folk 1
+▁geschickt 1
+gläser 1
+filiale 1
+▁bedienen 1
+käufe 1
+▁vereinbart 1
+▁dasselbe 1
+▁abstrakt 1
+wirksam 1
+▁verpasst 1
+rutscht 1
+nehmlich 1
+▁eckige 1
+feuchtigkeit 1
+▁gebiss 1
+trägt 1
+sänger 1
+▁stroh 1
+▁ereignis 1
+zitat 1
+konstrukt 1
+▁genoss 1
+entzündung 1
+gekühlt 1
+zyklus 1
+č 1
+▁aluminium 1
+▁argument 1
+▁arrondissement 1
+▁badewanne 1
+▁befördert 1
+▁betreuung 1
+▁columbia 1
+▁dezember 1
+▁erkrankt 1
+▁gestorben 1
+▁ignorieren 1
+▁publiziert 1
+▁resultat 1
+▁science 1
+▁skript 1
+▁sozusagen 1
+▁staatsanwalt 1
+▁webbrowser 1
+▁ägyptische 1
+effizienz 1
+▁begrüß 1
+▁beruhig 1
+▁hockey 1
+▁kombiniert 1
+▁kommissar 1
+▁schwör 1
+▁unzufrieden 1
+▁verzeih 1
+kampagne 1
+▁senkrecht 1
+▁kastanie 1
+▁lyrik 1
+▁forstwirtschaft 1
+▁herdplatte 1
+▁erdgas 1
+▁preußen 1
+▁verringert 1
+▁group 1
+▁tendenz 1
+▁vereinzelt 1
+▁praktik 1
+glühen 1
+sphäre 1
+▁jegliche 1
+▁terrasse 1
+stiefel 1
+▁vermehrt 1
+▁wieviel 1
+▁unternahm 1
+cross 1
+▁löt 1
+▁ingwer 1
+billy 1
+▁schäf 1
+unterhalt 1
+haube 1
+▁nachlass 1
+pedal 1
+aufzeichnungen 1
+▁klär 1
+spezialisten 1
+angebot 1
+frucht 1
+spiegel 1
+interpretiert 1
+nathan 1
+▁gymnas 1
+▁sekret 1
+▁beobachtung 1
+katastrophe 1
+▁beheimatet 1
+▁chicago 1
+▁eröffnen 1
+▁geschafft 1
+▁gewitter 1
+▁gewusst 1
+▁hubschrauber 1
+▁löffel 1
+▁münchner 1
+▁umgewandelt 1
+▁verteidiger 1
+▁wähler 1
+▁youtube 1
+▁ägypten 1
+chleswigholstein 1
+▁akkorde 1
+▁befestigung 1
+▁elegant 1
+▁erfindung 1
+▁illegal 1
+▁befindlich 1
+▁vanille 1
+▁bachelor 1
+▁prominente 1
+kompetent 1
+▁fußgänger 1
+▁minderheit 1
+genügend 1
+▁konstruktiv 1
+▁existenz 1
+▁unterhielt 1
+▁schimmel 1
+schmuck 1
+▁blonde 1
+▁moped 1
+▁verschwunden 1
+▁unheimlich 1
+▁touch 1
+▁formuliert 1
+▁mädel 1
+▁wofür 1
+▁simson 1
+▁zensur 1
+gestimmt 1
+▁angrenzende 1
+▁franzosen 1
+matratze 1
+▁heroin 1
+▁creme 1
+bruck 1
+schöpft 1
+▁stachel 1
+▁charlie 1
+▁vorbehalt 1
+▁vermiss 1
+▁einheitliche 1
+inhaber 1
+▁späteste 1
+▁exakt 1
+platin 1
+schwester 1
+▁betreff 1
+motiv 1
+ł 1
+schütteln 1
+spektrum 1
+sprünge 1
+▁abgeschafft 1
+▁ausgedehnte 1
+▁befürchte 1
+▁begeisterung 1
+▁bielefeld 1
+▁braunschweig 1
+▁gemütlich 1
+▁gespalten 1
+▁korruption 1
+▁kritisiere 1
+▁oxford 1
+▁scheinwerfer 1
+▁ungeduldig 1
+dämpfer 1
+▁angreifer 1
+▁aufwändig 1
+▁kompromiss 1
+▁konfession 1
+▁mechanisch 1
+▁skandinavi 1
+▁unregelmäßig 1
+▁feierabend 1
+▁passagier 1
+teleskop 1
+▁mozart 1
+▁nominiert 1
+▁unternehmer 1
+mechanik 1
+▁sandwich 1
+▁statistik 1
+▁türkei 1
+▁eklig 1
+▁behielt 1
+▁fleißig 1
+jörn 1
+▁installation 1
+zeichner 1
+▁klassik 1
+▁trikot 1
+▁bewaldet 1
+▁hübsch 1
+▁arnold 1
+▁gemarkung 1
+▁irritier 1
+▁norwegische 1
+▁deaktivier 1
+▁dozent 1
+▁weißwein 1
+▁floh 1
+turbine 1
+klopft 1
+▁earl 1
+▁ungarn 1
+koppel 1
+vektor 1
+zentren 1
+▁tapete 1
+▁längs 1
+schleife 1
+plateau 1
+▁jacob 1
+gerichtshof 1
+futter 1
+dünger 1
+zwischenzeitlich 1
+▁bahnsteig 1
+▁nudel 1
+kolben 1
+albrecht 1
+größe 1
+polizei 1
+turnier 1
+video 1
+kritische 1
+▁kindergärt 1
+zwang 1
+▁krimi 1
+gerührt 1
+orthodoxe 1
+▁außergewöhnlich 1
+▁bedeutsam 1
+▁beerdigung 1
+▁fälschlich 1
+▁gewünscht 1
+▁kennengelernt 1
+▁klimawandel 1
+▁kräuter 1
+▁merkwürdig 1
+▁mitbewohner 1
+▁pazifik 1
+▁produzieren 1
+▁qualitativ 1
+▁versöhn 1
+▁zurückgekehrt 1
+dürftig 1
+▁kardinal 1
+▁kümmert 1
+▁prototyp 1
+▁technologie 1
+▁vergrößert 1
+▁verschollen 1
+gürtel 1
+reflex 1
+▁theodor 1
+▁verrückt 1
+▁neuseeland 1
+▁abgebildet 1
+▁kündigung 1
+▁resonanz 1
+▁seltsam 1
+▁füße 1
+▁kanzler 1
+▁bräuchte 1
+▁demonstra 1
+express 1
+▁droht 1
+▁henrik 1
+▁semester 1
+prognose 1
+▁dänische 1
+▁krypt 1
+▁variiert 1
+▁entdeckung 1
+▁helfer 1
+▁clown 1
+▁manfred 1
+▁spanier 1
+▁fokus 1
+▁kannibal 1
+▁überarbeitet 1
+▁alexandr 1
+klopfen 1
+▁verkäufer 1
+▁chlor 1
+▁zwingen 1
+▁verpflichtung 1
+gemessen 1
+▁hessische 1
+▁entsprach 1
+brigade 1
+frachter 1
+zügl 1
+muschel 1
+screen 1
+▁kleist 1
+schmerzen 1
+pädagogik 1
+affin 1
+professor 1
+▁drübe 1
+beschäftigung 1
+feedback 1
+schränke 1
+undsechzig 1
+▁angestrebt 1
+▁berücksichtigt 1
+▁beschäftigen 1
+▁diagnose 1
+▁empirisch 1
+▁gebräuchlich 1
+▁gitarrist 1
+▁jesuiten 1
+▁kofferraum 1
+▁mercedes 1
+▁promoviert 1
+▁präsentation 1
+▁repertoire 1
+▁schildkröte 1
+▁schottland 1
+▁thailand 1
+▁vernunft 1
+▁versprochen 1
+▁vorübergehend 1
+fränkisch 1
+gebühren 1
+vorpommern 1
+▁ausleihe 1
+▁erfährt 1
+▁francisco 1
+▁kurfürst 1
+▁unterirdisch 1
+▁zeremonie 1
+geschliffen 1
+▁hähnchen 1
+▁skandal 1
+▁unfair 1
+▁barfuß 1
+▁grüße 1
+solvent 1
+▁brown 1
+▁wirkungsgrad 1
+▁kanadische 1
+▁frösche 1
+▁irrtum 1
+▁unterbrochen 1
+membran 1
+▁geklaut 1
+▁beidseitig 1
+pistole 1
+schleuder 1
+▁spieß 1
+föhn 1
+▁witwe 1
+▁brumm 1
+▁tugend 1
+reaktor 1
+▁zwerg 1
+▁diktier 1
+gelobt 1
+trüg 1
+wucher 1
+medizin 1
+bundesliga 1
+konzept 1
+historische 1
+schwarz 1
+▁hekti 1
+sequenz 1
+ı 1
+▁wahnsinn 1
+ohlenstoffdioxid 1
+▁abstammung 1
+▁autodidakt 1
+▁bestätigen 1
+▁chefredakteur 1
+▁defensiv 1
+▁dolmetscher 1
+▁entlehnt 1
+▁erforscht 1
+▁europäer 1
+▁football 1
+▁gärtner 1
+▁individuum 1
+▁integration 1
+▁koalition 1
+▁kohlenhydrat 1
+▁lächeln 1
+▁mythologie 1
+▁schädel 1
+▁staatsexamen 1
+▁update 1
+▁venedig 1
+▁victoria 1
+gewölbe 1
+▁ausdrücke 1
+▁betrachtung 1
+▁bewaffnet 1
+▁mexikanisch 1
+▁pinguin 1
+▁trainieren 1
+▁vietnam 1
+▁westfalen 1
+▁zeitgenosse 1
+wägung 1
+▁leopold 1
+▁straßburg 1
+▁stürzen 1
+▁darauffolgende 1
+▁oberschenkel 1
+▁separat 1
+ophyt 1
+▁gewaschen 1
+▁primär 1
+▁samuel 1
+gletscher 1
+▁episode 1
+gerasselt 1
+▁frosch 1
+▁angepasst 1
+geschoben 1
+▁flora 1
+pfeiler 1
+▁inszeniert 1
+quallen 1
+fizierung 1
+strumpf 1
+▁alliierte 1
+▁beschwerde 1
+▁morgendliche 1
+drüse 1
+binär 1
+schnecke 1
+▁skifahre 1
+▁hinterfrag 1
+kamera 1
+▁vorführ 1
+▁charakterisier 1
+▁archäologi 1
+denkmäler 1
+diagnosti 1
+kapazität 1
+theoretiker 1
+ungsgefahr 1
+▁abgetrennt 1
+▁amsterdam 1
+▁ausgiebig 1
+▁däumchen 1
+▁erzbistum 1
+▁garderobe 1
+▁gescheitert 1
+▁gestohlen 1
+▁gläubig 1
+▁höflich 1
+▁katharina 1
+▁koblenz 1
+▁kommerziell 1
+▁konzentrieren 1
+▁korrigieren 1
+▁madrid 1
+▁nächstgelegene 1
+▁plätzchen 1
+▁sigmund 1
+▁synagoge 1
+▁südkorea 1
+▁unternommen 1
+▁verfasser 1
+▁vernachlässigen 1
+▁zerstörung 1
+▁elternbeirat 1
+▁oscar 1
+▁valentin 1
+▁phantom 1
+empfunden 1
+▁tötet 1
+unterdrück 1
+▁charlotte 1
+wächter 1
+▁magdeburg 1
+▁brennessel 1
+▁nicole 1
+▁australische 1
+wasseroberfläche 1
+▁tränen 1
+▁vermarktet 1
+▁oskar 1
+ritual 1
+▁mikrowelle 1
+ergüsse 1
+abnehmer 1
+▁sabine 1
+▁drückt 1
+scanner 1
+▁morgenstund 1
+briefmarken 1
+plikat 1
+▁zünd 1
+kabine 1
+schwimmen 1
+leuchte 1
+gitarre 1
+bäume 1
+▁isländ 1
+▁jüdi 1
+▁differential 1
+▁rendsburg 1
+▁empör 1
+flüchtlinge 1
+weiterentwickelt 1
+▁akkuschrauber 1
+▁anspruchsvoll 1
+▁apotheker 1
+▁ausflugsziel 1
+▁bedürfnis 1
+▁boykott 1
+▁bündnis 1
+▁entspring 1
+▁erschrocken 1
+▁gebirgs 1
+▁gespeichert 1
+▁gleichgültig 1
+▁hiphop 1
+▁jerusalem 1
+▁jörg 1
+▁karosserie 1
+▁knopf 1
+▁konzipiert 1
+▁leichtathletik 1
+▁lügner 1
+▁napoleon 1
+▁rucksack 1
+▁schröder 1
+▁thüringen 1
+▁verlustbehaftet 1
+▁vermähl 1
+▁volumen 1
+▁werbespot 1
+▁whatsappgruppe 1
+donaudampfschiff 1
+getroffen 1
+pünktlich 1
+saxophon 1
+▁freimaurer 1
+▁physikalische 1
+▁volksarmee 1
+▁übersetzer 1
+▁erstaunlich 1
+▁gestört 1
+▁köstlich 1
+▁methodi 1
+▁nicola 1
+gehängt 1
+▁eberhard 1
+▁fiktive 1
+krümel 1
+▁apostel 1
+gewohnheiten 1
+▁torhüter 1
+▁monitor 1
+▁arrangiert 1
+▁ungesund 1
+▁satire 1
+unterkunft 1
+krabbelt 1
+▁gebüsch 1
+▁befreundet 1
+▁kündigte 1
+bezüglich 1
+rauszuholen 1
+marathon 1
+▁linguisti 1
+schwankungen 1
+▁epoche 1
+▁kulisse 1
+finsternis 1
+▁pflück 1
+gestreift 1
+einsatzkommando 1
+bewusstsein 1
+direktion 1
+genauigkeit 1
+höpf 1
+kräftig 1
+jugend 1
+europäische 1
+▁beeinträchtig 1
+▁panorama 1
+▁verwechsl 1
+observatorium 1
+undvierzig 1
+ř 1
+▁abkühlen 1
+▁anfällig 1
+▁angekündigt 1
+▁ausführlich 1
+▁außerordentlich 1
+▁begegnet 1
+▁breisgau 1
+▁dissertation 1
+▁ermordet 1
+▁fraktur 1
+▁gefällig 1
+▁genehmigt 1
+▁indirekt 1
+▁kontrahent 1
+▁kontrovers 1
+▁matthias 1
+▁mittagspause 1
+▁orientalische 1
+▁premierminister 1
+▁propaganda 1
+▁renommierte 1
+▁subventionier 1
+▁thronfolger 1
+▁ukraine 1
+▁variable 1
+▁vernachlässigt 1
+▁zerstörer 1
+kanzlei 1
+▁asphalt 1
+▁bestraft 1
+▁bewährung 1
+▁erledigen 1
+▁grammatikalisch 1
+▁kollektiv 1
+▁krokodil 1
+▁sachverhalt 1
+▁spekulation 1
+▁spionage 1
+▁edward 1
+▁sklaven 1
+▁texas 1
+▁woanders 1
+▁istanbul 1
+▁variieren 1
+▁gregor 1
+▁empor 1
+▁künste 1
+▁allianz 1
+▁schottische 1
+autoritäre 1
+▁papagei 1
+▁präzise 1
+murmel 1
+▁entnehme 1
+monarchie 1
+gefroren 1
+▁hülse 1
+▁vegane 1
+olivia 1
+späne 1
+marmelade 1
+rumpelt 1
+offensive 1
+▁aufflieg 1
+überwacht 1
+göttin 1
+supermarkt 1
+lutherische 1
+farbig 1
+▁vollzieh 1
+begabung 1
+zwanzig 1
+software 1
+literatur 1
+genossenschaft 1
+erscheinungsbild 1
+zugehörigkeit 1
+ā 1
+▁afghanistan 1
+▁beatrix 1
+▁begegnen 1
+▁beschuldig 1
+▁betäubung 1
+▁dokumentarfilm 1
+▁einkaufswagen 1
+▁elektrifizier 1
+▁entnommen 1
+▁explizit 1
+▁fortpflanzung 1
+▁fräulein 1
+▁fröhlich 1
+▁genehmigung 1
+▁gerüchte 1
+▁gesammelt 1
+▁heilbronn 1
+▁hervorragende 1
+▁identifizier 1
+▁katholiken 1
+▁kaufmännische 1
+▁ladungsträger 1
+▁laptop 1
+▁leichnam 1
+▁protagonist 1
+▁registriert 1
+▁rekonstruier 1
+▁resultierende 1
+▁schizophren 1
+ã 1
+ô 1
+â 1
+ø 1
+ū 1
+ú 1
+ý 1
+ș 1
+ë 1
+ş 1
+ʿ 1
+ī 1
+ě 1
+à 1
+î 1
+å 1
+æ 1
+ə 1
+ă 1
+ò 1
+đ 1
+ů 1
+ň 1
+û 1
+œ 1
+ť 1
+ḫ 1
+ź 1
+ğ 1
+ï 1
+ń 1
+ê 1
+ț 1
+ą 1
+ś 1
+ñ 1
+ž 1
+ď 1
+ð 1
+ç 1
+õ 1
+ù 1
+þ 1
+ć 1
+ė 1
+ħ 1
+ņ 1
+ő 1
+ż 1
+μ 1
+ṣ 1
+ắ 1
+ế 1
+カ 1
+临 1
+孙 1
+尣 1
+道 1
diff --git a/configs/de-en/src_unigram6000/spm_unigram_de.vocab b/configs/de-en/src_unigram6000/spm_unigram_de.vocab
new file mode 100644
index 0000000000000000000000000000000000000000..9e1c4636af47f785239b1d4bf832c377c6dd8571
--- /dev/null
+++ b/configs/de-en/src_unigram6000/spm_unigram_de.vocab
@@ -0,0 +1,6000 @@
+ 0
+ 0
+ 0
+ 0
+en -3.585
+e -3.75243
+s -3.77096
+▁ -3.79672
+n -3.83458
+▁die -3.88759
+▁der -3.90332
+▁in -4.43123
+t -4.46404
+▁und -4.5072
+▁ist -4.63531
+er -4.65076
+▁das -4.76096
+▁er -4.79871
+▁ein -4.88403
+▁zu -5.01818
+▁sie -5.04016
+▁den -5.04552
+▁auf -5.11434
+r -5.12161
+▁sich -5.16288
+▁es -5.1697
+▁nicht -5.17
+▁mit -5.18062
+▁an -5.18416
+▁von -5.22129
+▁eine -5.27066
+▁im -5.29875
+▁aus -5.43964
+▁als -5.45182
+▁des -5.52344
+m -5.54761
+▁auch -5.56566
+▁dem -5.5719
+▁war -5.57334
+▁für -5.62648
+▁ich -5.64096
+▁man -5.65263
+▁nach -5.72062
+▁sind -5.72659
+▁wurde -5.75963
+es -5.77551
+te -5.8034
+▁hat -5.83537
+ung -5.89111
+▁wird -5.90134
+▁so -5.91317
+▁werden -5.94001
+st -5.94989
+▁be -5.9729
+▁vor -6.00778
+▁ver -6.01299
+▁noch -6.02362
+▁du -6.07666
+▁wie -6.11221
+▁ge -6.15263
+k -6.15968
+d -6.1687
+▁ab -6.17855
+▁bei -6.17969
+▁über -6.1944
+in -6.21253
+▁um -6.2408
+▁seine -6.2702
+▁durch -6.2786
+▁kann -6.27905
+▁einen -6.28628
+▁wir -6.29351
+▁diese -6.31058
+p -6.31985
+▁nur -6.33553
+y -6.33897
+ä -6.35801
+▁sein -6.37929
+h -6.39597
+▁aber -6.40558
+g -6.41056
+▁am -6.41278
+▁zum -6.42846
+u -6.44151
+▁ihr -6.4502
+▁unter -6.4526
+b -6.46207
+o -6.48972
+▁einer -6.52153
+l -6.52426
+▁oder -6.52466
+▁was -6.56556
+▁einem -6.57363
+f -6.57761
+a -6.61126
+ge -6.63278
+▁mal -6.64275
+▁wenn -6.66542
+z -6.67506
+▁zwei -6.6791
+de -6.69245
+ig -6.69464
+al -6.69714
+▁haben -6.71039
+ü -6.72952
+▁wieder -6.7454
+▁zur -6.74762
+▁schon -6.74788
+ch -6.74999
+▁hier -6.75873
+ten -6.76765
+▁da -6.78465
+el -6.78507
+▁dass -6.78699
+an -6.7935
+sch -6.80721
+le -6.80733
+lich -6.81237
+ck -6.83571
+▁zeit -6.84373
+▁dieser -6.84481
+▁gibt -6.85145
+▁mehr -6.86782
+▁ihre -6.87234
+ter -6.87739
+▁muss -6.8933
+be -6.90701
+ö -6.9205
+▁mir -6.92818
+▁immer -6.94387
+re -6.95023
+ur -6.95079
+▁alle -6.95775
+ischen -6.96448
+▁bis -6.97819
+i -6.9786
+▁sch -6.9804
+is -7.00442
+▁dann -7.00676
+w -7.01292
+et -7.01503
+▁heute -7.0179
+▁wurden -7.02533
+▁keine -7.04031
+▁können -7.11509
+▁seiner -7.11609
+▁k -7.13494
+iert -7.13504
+us -7.14616
+dem -7.14815
+▁doch -7.15032
+den -7.17577
+▁re -7.17872
+der -7.1878
+▁ja -7.19784
+ische -7.22509
+▁viel -7.22848
+li -7.23008
+▁stadt -7.23589
+▁s -7.23951
+▁drei -7.24317
+▁dabei -7.26493
+ll -7.27005
+at -7.28315
+▁jedoch -7.28549
+▁hatte -7.29126
+▁un -7.29899
+or -7.29956
+ver -7.30156
+vom -7.30498
+▁mich -7.30825
+▁sehr -7.30939
+▁später -7.31046
+em -7.31054
+chen -7.31241
+zeit -7.33045
+▁teil -7.33106
+it -7.34046
+▁ihm -7.34635
+▁wo -7.35117
+▁a -7.35163
+ar -7.35325
+▁jetzt -7.35905
+ra -7.3655
+▁dort -7.36809
+▁sollte -7.3682
+▁kein -7.37818
+▁etwas -7.38741
+▁liegt -7.38753
+il -7.38785
+▁gut -7.39455
+ieren -7.39587
+▁viele -7.39679
+um -7.39743
+▁einfach -7.39919
+▁damit -7.40411
+▁beim -7.4114
+la -7.43744
+on -7.43811
+▁habe -7.44927
+ungen -7.45576
+▁selbst -7.45623
+ie -7.45727
+ri -7.45799
+▁gegen -7.46631
+▁geht -7.47135
+ut -7.4735
+▁waren -7.48145
+▁vier -7.48503
+▁la -7.49316
+c -7.49522
+isch -7.49812
+▁bitte -7.50227
+ler -7.50696
+gel -7.52342
+bau -7.53013
+▁außer -7.54722
+na -7.55344
+▁kommt -7.55905
+▁uns -7.56213
+▁ihn -7.56344
+land -7.56753
+▁ma -7.56995
+sten -7.57115
+▁mein -7.57258
+lo -7.57673
+kt -7.58211
+▁wer -7.58619
+▁zurück -7.58655
+gen -7.58869
+zu -7.58877
+he -7.59435
+▁ohne -7.59674
+ma -7.59755
+bar -7.5992
+▁bekannt -7.6183
+▁ra -7.62027
+▁ob -7.62676
+▁kurz -7.63141
+▁frei -7.63212
+x -7.63565
+di -7.63639
+▁weiter -7.6373
+▁soll -7.63853
+au -7.6447
+ro -7.65375
+tra -7.65672
+▁c -7.66361
+ing -7.66532
+bel -7.67765
+▁de -7.6819
+▁alles -7.68399
+▁konnte -7.68807
+halb -7.69129
+▁andere -7.69554
+▁le -7.69819
+▁neu -7.70179
+ation -7.70371
+ka -7.70708
+▁wasser -7.71059
+▁ganz -7.72112
+▁diesem -7.72134
+▁will -7.72353
+▁seinen -7.72471
+mann -7.72839
+▁anderen -7.72915
+▁zusammen -7.73045
+▁ersten -7.73264
+▁b -7.73337
+▁steht -7.73371
+▁neben -7.73499
+▁dir -7.7374
+▁erst -7.73992
+▁einige -7.74735
+ul -7.75035
+▁hin -7.75162
+▁bau -7.75514
+▁nichts -7.76138
+▁ent -7.76262
+▁ha -7.76566
+▁erste -7.76631
+▁gleich -7.7709
+▁jahr -7.77302
+me -7.77535
+▁film -7.77568
+▁nun -7.77789
+▁weiß -7.77799
+hl -7.77909
+▁kinder -7.78373
+▁he -7.79127
+▁neue -7.79476
+end -7.7963
+ken -7.79778
+▁macht -7.8012
+as -7.81061
+heit -7.81071
+stand -7.81117
+ien -7.81122
+▁mehrere -7.81345
+to -7.81527
+▁fünf -7.81531
+▁ro -7.81647
+▁her -7.81923
+cht -7.82705
+▁zwischen -7.82863
+▁dich -7.82889
+▁na -7.83147
+▁dazu -7.83649
+berg -7.83695
+▁land -7.84145
+ta -7.84245
+igen -7.85345
+▁pro -7.85586
+▁hoch -7.85639
+ste -7.86439
+▁li -7.86587
+ben -7.86612
+▁statt -7.87216
+▁kam -7.87349
+▁g -7.88034
+▁ihrer -7.88133
+vi -7.88323
+ner -7.88713
+▁oft -7.88848
+▁bereits -7.89096
+▁gehört -7.89221
+da -7.89237
+▁meine -7.89377
+▁lassen -7.90071
+▁weg -7.90224
+pp -7.90425
+▁befindet -7.90585
+ha -7.91258
+bi -7.91268
+ab -7.91653
+am -7.91693
+▁fl -7.917
+▁weitere -7.9171
+▁ober -7.91782
+▁während -7.92007
+ik -7.92314
+go -7.92338
+▁kon -7.93033
+▁gemeinde -7.93325
+▁jahre -7.93673
+▁herr -7.93832
+▁leben -7.93838
+▁bo -7.93871
+▁ta -7.93994
+rat -7.94056
+hr -7.948
+▁lange -7.94851
+recht -7.95036
+los -7.95294
+ige -7.95951
+▁frau -7.9641
+▁ort -7.96883
+che -7.9708
+ff -7.97327
+▁w -7.97375
+▁menschen -7.97713
+▁wäre -7.97913
+▁lässt -7.98075
+gt -7.98238
+▁se -7.98302
+▁welt -7.9834
+pa -7.98582
+▁seit -7.98891
+▁schnell -7.98926
+▁ka -7.98953
+▁kommen -7.99153
+▁danach -7.99298
+▁z -7.99505
+burg -8.00115
+▁bin -8.00251
+▁spiel -8.00286
+▁würde -8.00379
+▁gab -8.0051
+ko -8.00563
+zi -8.01063
+▁nie -8.01176
+un -8.01643
+liche -8.0191
+▁grund -8.03262
+ant -8.03281
+▁machen -8.0338
+ster -8.03691
+▁beiden -8.0437
+▁also -8.04429
+▁eigentlich -8.04541
+ke -8.05255
+keit -8.0526
+lichen -8.05432
+▁ende -8.06635
+▁genau -8.06878
+▁kannst -8.07326
+werk -8.07581
+▁gerade -8.07654
+fall -8.07668
+▁po -8.0772
+ger -8.07767
+gang -8.07776
+▁jede -8.07946
+▁gar -8.085
+id -8.08614
+th -8.09023
+co -8.09202
+▁gr -8.10136
+ber -8.10361
+ga -8.10536
+mit -8.10598
+nd -8.10701
+son -8.10717
+form -8.10868
+spiel -8.10916
+▁welche -8.10924
+▁h -8.11173
+▁haupt -8.11319
+▁jeder -8.11551
+▁ins -8.11656
+ton -8.11725
+ol -8.1177
+▁besonder -8.11785
+▁euch -8.12241
+▁acht -8.12297
+▁gilt -8.12418
+ich -8.13112
+▁je -8.13483
+v -8.1358
+fer -8.13891
+tag -8.13921
+▁allem -8.14065
+▁sp -8.14171
+▁allerdings -8.14182
+▁fest -8.14285
+▁zahlreiche -8.14384
+▁müssen -8.14588
+wa -8.14667
+▁hast -8.15001
+▁fa -8.15573
+▁stehen -8.159
+▁möglich -8.16066
+▁deine -8.1655
+und -8.16557
+ad -8.16558
+ei -8.16736
+▁solche -8.16813
+elt -8.1693
+▁mittel -8.17005
+schaft -8.17018
+▁darf -8.17292
+▁me -8.17437
+ion -8.17643
+nächst -8.17779
+▁o -8.17863
+▁ihnen -8.17891
+tz -8.18258
+▁einmal -8.18411
+▁schwer -8.18593
+▁ho -8.18718
+▁al -8.18936
+▁platz -8.19698
+▁jahren -8.19893
+stein -8.19952
+▁sei -8.20731
+▁p -8.20847
+▁groß -8.2099
+teil -8.21019
+ein -8.21106
+▁sehen -8.21117
+▁blieb -8.21175
+▁ne -8.21208
+ti -8.21661
+▁voll -8.21753
+▁lang -8.22097
+ia -8.22452
+▁f -8.22488
+▁meist -8.22554
+▁musik -8.22602
+▁sowie -8.22665
+▁besser -8.22681
+▁möchte -8.22895
+▁beide -8.22919
+▁hinter -8.22921
+▁art -8.22946
+haus -8.2328
+bo -8.23295
+▁natürlich -8.23327
+mo -8.23542
+▁sohn -8.238
+▁stark -8.23929
+▁mar -8.23932
+pe -8.24048
+▁deutschland -8.24219
+▁tag -8.24285
+▁verschiedene -8.24365
+▁könnte -8.24413
+▁vater -8.2493
+os -8.2499
+▁ging -8.25057
+▁namen -8.25545
+▁hand -8.25686
+▁schön -8.25924
+▁berlin -8.26022
+ling -8.26087
+▁luft -8.26097
+▁ri -8.2624
+▁we -8.26482
+ß -8.26527
+▁davon -8.26647
+ic -8.26672
+ende -8.26947
+▁familie -8.27265
+ierung -8.27318
+im -8.27402
+fe -8.27466
+stadt -8.276
+▁bi -8.27664
+sa -8.281
+▁deutschen -8.29088
+▁form -8.29768
+tisch -8.29869
+her -8.30056
+▁große -8.30096
+▁etwa -8.30111
+▁mo -8.30221
+▁wohl -8.30524
+ier -8.30548
+ium -8.30583
+bl -8.30701
+▁lo -8.30731
+▁sa -8.30799
+▁bundes -8.30974
+▁recht -8.31064
+zen -8.3124
+tor -8.31675
+▁dafür -8.31959
+ierte -8.32095
+kommen -8.32351
+ir -8.32406
+▁ausge -8.32487
+ph -8.32654
+▁sechs -8.32682
+▁kr -8.32807
+▁pa -8.32992
+▁weit -8.33133
+▁weil -8.3316
+▁früh -8.33361
+▁gehen -8.33461
+▁meisten -8.33484
+do -8.33908
+▁studierte -8.34323
+▁national -8.34429
+▁universität -8.34794
+▁rot -8.34803
+▁jemand -8.34848
+▁zweiten -8.34949
+▁ar -8.35334
+▁häufig -8.35337
+▁haus -8.35395
+zug -8.35537
+platz -8.35588
+acht -8.35736
+va -8.35947
+tion -8.36033
+ru -8.36208
+ft -8.36349
+ib -8.36366
+▁krieg -8.36721
+ku -8.36883
+gebiet -8.36992
+wi -8.37028
+ze -8.37197
+lin -8.37285
+enden -8.38352
+lauf -8.38501
+▁pi -8.38508
+▁vi -8.39068
+▁besteht -8.39125
+▁di -8.39284
+▁wegen -8.40105
+▁ru -8.40158
+weg -8.40285
+▁su -8.40762
+▁verwendet -8.40889
+aus -8.40964
+iv -8.41059
+▁ebenfalls -8.41407
+▁findet -8.41681
+▁warum -8.41695
+▁fast -8.41744
+▁halb -8.41824
+▁band -8.41827
+▁ange -8.41896
+▁wollte -8.42041
+ot -8.42125
+ment -8.42583
+▁lass -8.42757
+▁ca -8.43086
+▁ba -8.43729
+gehen -8.43734
+▁unsere -8.44089
+kel -8.44168
+tal -8.44526
+lang -8.44639
+si -8.44666
+▁wenig -8.44847
+▁erhielt -8.44856
+▁berg -8.45015
+▁bezeichnet -8.45126
+▁sitz -8.4523
+hi -8.45438
+▁mitglied -8.45509
+▁schul -8.45512
+▁geschichte -8.45552
+▁sieht -8.45726
+▁arbeit -8.45811
+▁hu -8.45861
+mer -8.46065
+▁benannt -8.46214
+▁klein -8.46332
+ine -8.46365
+▁darauf -8.46488
+▁ste -8.46581
+▁natur -8.46761
+ell -8.46799
+feld -8.46902
+mal -8.46924
+▁zehn -8.47593
+▁kopf -8.47899
+cher -8.47918
+▁all -8.47991
+▁lebt -8.48087
+▁diesen -8.48323
+▁gebiet -8.48395
+bahn -8.48462
+ja -8.4852
+▁paar -8.48785
+man -8.48845
+▁landes -8.48863
+auf -8.48873
+po -8.48923
+▁eigene -8.49051
+iz -8.49086
+so -8.49132
+sel -8.49169
+lage -8.49402
+▁vielleicht -8.49552
+▁erfolg -8.4983
+▁daher -8.49987
+▁bu -8.50321
+▁ex -8.50428
+ick -8.5043
+▁wirklich -8.5044
+▁st -8.5052
+ist -8.50569
+wert -8.50772
+mittel -8.50772
+fi -8.50984
+ho -8.51019
+grund -8.5115
+▁verheiratet -8.51264
+▁wort -8.51418
+mi -8.51491
+ce -8.51492
+▁pe -8.5199
+▁co -8.52434
+ker -8.52631
+▁könig -8.52641
+▁letzte -8.52706
+▁finden -8.5272
+punkt -8.52755
+▁deutsche -8.52961
+ba -8.53323
+einander -8.53342
+▁arbeitete -8.53359
+▁lebens -8.53689
+op -8.53691
+tel -8.53751
+▁lu -8.54144
+ität -8.54282
+▁niemand -8.54485
+▁the -8.54725
+wohl -8.5498
+halten -8.55073
+▁stand -8.55194
+heim -8.55339
+▁eis -8.5566
+weise -8.55692
+▁licht -8.55745
+▁bar -8.55952
+ven -8.56207
+▁neun -8.56212
+gra -8.56362
+▁alt -8.5638
+bei -8.56396
+▁abge -8.56439
+▁weltkrieg -8.56533
+▁hätte -8.56539
+stoff -8.56591
+▁burg -8.56737
+per -8.56759
+lu -8.5715
+▁spielte -8.57297
+▁mi -8.57345
+▁werk -8.5749
+haft -8.57495
+om -8.57618
+▁direkt -8.57658
+raum -8.57887
+iger -8.58057
+▁kleine -8.58137
+reich -8.58288
+▁rolle -8.58506
+▁morgen -8.58756
+▁erhalten -8.59007
+▁komm -8.59058
+▁tätig -8.59322
+▁richtig -8.59354
+schen -8.59366
+▁rechts -8.59582
+rie -8.59613
+▁alter -8.59667
+▁ki -8.59715
+isiert -8.59731
+▁sonst -8.59768
+▁verein -8.59825
+geben -8.59893
+▁nahm -8.5995
+▁of -8.59958
+▁klar -8.60092
+ve -8.60137
+▁sicher -8.6026
+▁heißt -8.60386
+pel -8.6046
+▁neuen -8.60614
+▁see -8.60854
+▁wollen -8.60932
+▁zeigt -8.60933
+▁fall -8.6101
+▁buch -8.61144
+▁genannt -8.61242
+▁anschließend -8.61243
+▁unterschiedlich -8.61373
+eck -8.61476
+▁kunst -8.61542
+▁liegen -8.61749
+▁rein -8.61799
+anz -8.61854
+ak -8.62079
+▁do -8.62296
+ismus -8.62317
+▁leute -8.62376
+min -8.62438
+ehm -8.62667
+▁jo -8.62687
+stellung -8.62985
+▁versucht -8.63056
+▁kirche -8.6318
+ischer -8.63244
+▁schl -8.63305
+bach -8.63385
+▁sonder -8.63684
+▁dein -8.63892
+▁kaum -8.63942
+▁sieben -8.64146
+▁l -8.64278
+▁mag -8.64448
+▁te -8.64694
+▁auto -8.64785
+▁begann -8.65133
+▁versch -8.65245
+▁wald -8.65334
+▁ko -8.65439
+▁süd -8.65474
+▁folgende -8.65496
+▁inner -8.65505
+▁mutter -8.65588
+▁einzige -8.6575
+▁gesch -8.65768
+ca -8.65785
+▁darüber -8.65804
+ppen -8.65878
+sicht -8.66031
+▁fr -8.66341
+▁führt -8.66409
+▁englisch -8.66466
+mon -8.66469
+▁heraus -8.66658
+▁ursprünglich -8.66799
+▁leicht -8.66822
+▁sta -8.66902
+arbeit -8.66946
+▁nord -8.67111
+bt -8.67161
+▁stellen -8.67233
+dessen -8.67558
+▁europa -8.67586
+▁fand -8.67595
+lös -8.67665
+reif -8.67679
+▁privat -8.67813
+▁alten -8.68151
+führung -8.68231
+stellen -8.68594
+▁unternehmen -8.68622
+qu -8.68655
+zeichen -8.68843
+nahme -8.68939
+▁geld -8.68955
+▁besten -8.6902
+▁ehe -8.6903
+▁musste -8.69069
+frei -8.69142
+▁öffentlich -8.69148
+▁schlecht -8.69182
+ow -8.69272
+▁sport -8.69341
+▁begriff -8.69389
+▁echt -8.69667
+▁ku -8.69859
+▁fach -8.69867
+ard -8.69962
+▁qu -8.70017
+▁halten -8.70025
+▁geben -8.70075
+▁laut -8.70095
+mme -8.70132
+▁kar -8.70191
+▁dar -8.70202
+▁tor -8.70258
+mpf -8.70266
+gesetzt -8.70553
+▁früher -8.70598
+▁essen -8.70779
+erweise -8.70799
+▁dorf -8.70813
+fel -8.71003
+▁größte -8.7102
+▁wein -8.71089
+▁bestimmt -8.71167
+▁einge -8.71372
+▁los -8.71483
+sprech -8.71533
+▁wissen -8.71612
+▁bedeutung -8.7162
+▁führte -8.71791
+▁tun -8.71799
+zahl -8.72032
+▁pu -8.72039
+schule -8.72192
+wand -8.72194
+▁erfolgreich -8.72351
+▁kreis -8.7246
+schrift -8.72492
+▁si -8.72711
+▁name -8.72745
+ord -8.73024
+stelle -8.73278
+ok -8.73309
+tsch -8.7333
+▁nehmen -8.73498
+▁stein -8.73571
+▁bleibt -8.73608
+dorf -8.73673
+zeug -8.73677
+meister -8.73718
+▁anders -8.73767
+▁eigenen -8.739
+▁bist -8.73948
+▁schau -8.73957
+▁wahl -8.74001
+öl -8.7401
+▁dadurch -8.74153
+dienst -8.74209
+▁west -8.7431
+▁partei -8.74331
+elle -8.7435
+hin -8.74697
+▁mark -8.74703
+fa -8.74749
+▁folge -8.74884
+▁freund -8.74972
+▁ziel -8.75043
+▁va -8.75065
+fahrt -8.75085
+ack -8.75211
+▁d -8.75442
+räum -8.75538
+▁besitz -8.7556
+▁bildet -8.7572
+vo -8.75723
+▁stammt -8.75826
+seite -8.7588
+▁wert -8.75961
+druck -8.76238
+▁handelt -8.76422
+sche -8.76442
+▁el -8.76519
+hör -8.76603
+and -8.76704
+▁beispiel -8.76706
+▁plan -8.76929
+io -8.76952
+▁manchmal -8.77056
+kon -8.77094
+bild -8.77381
+▁per -8.778
+rad -8.77945
+▁allen -8.78113
+ug -8.78145
+▁französisch -8.78179
+war -8.78581
+▁gemacht -8.78699
+ill -8.78871
+▁text -8.78933
+▁alte -8.78945
+ang -8.79115
+▁manche -8.792
+leben -8.79351
+igkeit -8.79489
+age -8.79545
+straße -8.79574
+▁mü -8.7958
+▁titel -8.79599
+ett -8.79808
+▁gebäude -8.79839
+halt -8.80013
+▁pol -8.80225
+art -8.803
+▁ad -8.80302
+mmer -8.80377
+▁großen -8.80614
+▁fällt -8.80637
+▁schwarz -8.80682
+▁ließ -8.80853
+▁for -8.80871
+▁anderem -8.80874
+▁schließlich -8.81108
+▁rück -8.81118
+▁gehörte -8.81159
+▁tod -8.81183
+▁braucht -8.81239
+▁letzten -8.81256
+▁augen -8.81415
+kla -8.81533
+▁eltern -8.81632
+▁farb -8.81737
+höhe -8.81992
+▁hält -8.82131
+▁insel -8.82195
+▁straßen -8.82253
+besondere -8.82496
+gestellt -8.82584
+sta -8.82794
+see -8.82798
+laden -8.82864
+▁schule -8.82917
+satz -8.83004
+▁kom -8.83078
+itz -8.83146
+stift -8.83618
+hol -8.83865
+▁fort -8.84014
+ziehen -8.84074
+▁fi -8.8419
+folge -8.84208
+gelegt -8.84242
+tan -8.84287
+▁arm -8.84312
+▁lehr -8.84326
+kan -8.84346
+▁schloss -8.84389
+sam -8.84545
+▁zie -8.84637
+tre -8.84666
+▁werde -8.8467
+▁trat -8.84679
+äl -8.84751
+tzt -8.84821
+▁kind -8.84857
+▁regel -8.85075
+gabe -8.85391
+▁zog -8.85402
+▁bevor -8.85449
+▁bald -8.85575
+▁tochter -8.85605
+stieg -8.85616
+ort -8.85871
+▁ost -8.85958
+▁damals -8.86015
+▁gehören -8.8602
+▁gesehen -8.86171
+▁herum -8.86193
+wer -8.86375
+ide -8.86404
+▁kra -8.86414
+▁stellt -8.8642
+▁arbeits -8.86489
+handel -8.86808
+▁fahr -8.86828
+gebracht -8.86852
+▁körper -8.86897
+▁nieder -8.86946
+▁jugend -8.87226
+▁no -8.87295
+▁schiff -8.87432
+ator -8.87516
+▁pf -8.87553
+▁sagen -8.8757
+ach -8.8757
+▁gelang -8.87571
+lt -8.87647
+▁daneben -8.87654
+▁wieso -8.87711
+ring -8.87844
+ah -8.87959
+gi -8.87967
+▁daten -8.88026
+legen -8.88106
+▁ir -8.8813
+▁bas -8.88227
+▁wa -8.88297
+▁bereich -8.88344
+pi -8.88379
+vor -8.88401
+▁vielen -8.88564
+▁fernseh -8.8858
+▁offen -8.88896
+▁daraufhin -8.8893
+▁park -8.88962
+▁grab -8.8899
+▁denn -8.89203
+▁wann -8.89379
+▁spielen -8.8943
+ry -8.89432
+▁sogar -8.89469
+▁aktiv -8.89472
+▁grün -8.89671
+▁eigen -8.89678
+▁gemeinsam -8.89748
+▁gott -8.89772
+▁langsam -8.89777
+▁gegenüber -8.8978
+▁jeden -8.8987
+▁falsch -8.90172
+port -8.9052
+▁ehemalige -8.90596
+▁gef -8.90718
+we -8.90751
+wo -8.90853
+▁zweite -8.91018
+▁geh -8.91155
+▁preis -8.91203
+schein -8.91281
+▁seite -8.91306
+▁nächsten -8.91342
+▁ju -8.91345
+▁gleichzeitig -8.91473
+▁gen -8.91517
+deck -8.9184
+gie -8.91853
+▁wenige -8.91945
+▁überhaupt -8.9195
+▁kreuz -8.92063
+▁genutzt -8.92316
+bruch -8.92317
+lau -8.92361
+▁fahren -8.92483
+▁herz -8.92555
+fahren -8.93014
+▁to -8.93052
+▁dritte -8.93239
+▁weiterhin -8.93249
+▁mannschaft -8.93361
+▁ga -8.93365
+scha -8.93385
+▁gute -8.93567
+verkehr -8.93643
+▁irgendwie -8.93653
+▁blick -8.93659
+mar -8.93695
+träger -8.94004
+graf -8.94038
+▁selten -8.94083
+▁landwirtschaft -8.94101
+▁weiteren -8.94174
+▁wohn -8.94236
+▁ganze -8.94411
+▁sollten -8.94461
+▁holz -8.94496
+▁aufge -8.94639
+▁boden -8.9476
+setzen -8.94854
+▁anfang -8.94971
+tum -8.94976
+voll -8.95038
+▁lieber -8.951
+▁daran -8.95115
+▁system -8.95247
+pro -8.95315
+▁frage -8.9532
+igt -8.95328
+▁nein -8.95492
+bro -8.9554
+▁tr -8.95551
+eher -8.95555
+▁spielt -8.95617
+up -8.9572
+the -8.9577
+▁raum -8.95937
+▁beste -8.96055
+▁pass -8.96119
+log -8.96195
+▁bleiben -8.96248
+▁meinen -8.96302
+▁worden -8.96302
+▁autor -8.96313
+▁rest -8.96378
+▁mil -8.96522
+leiter -8.96546
+▁roman -8.96576
+amt -8.96632
+▁san -8.96713
+▁kloster -8.96752
+▁not -8.96775
+▁hatten -8.96879
+▁flug -8.97066
+geordnet -8.97206
+▁sagt -8.97251
+gebaut -8.97276
+▁stets -8.973
+▁go -8.97407
+▁betrieb -8.97517
+▁minuten -8.97564
+▁gewann -8.97662
+▁gerne -8.97852
+bereich -8.97943
+▁größe -8.9797
+▁ei -8.97988
+netz -8.98026
+▁fisch -8.98063
+▁lag -8.9834
+▁eben -8.98357
+▁fluss -8.98532
+▁dienst -8.98543
+▁fe -8.98692
+▁tief -8.98717
+ina -8.98721
+schlag -8.98746
+▁theater -8.98977
+▁einsatz -8.98995
+▁eingesetzt -8.9913
+▁bürger -8.99153
+▁ben -8.99199
+gleich -8.99216
+ssen -8.99227
+▁kl -8.99262
+▁tür -8.9927
+sion -8.99317
+▁bra -8.99338
+▁start -8.99424
+▁sozial -8.9944
+od -8.9958
+▁hinaus -8.99593
+▁schi -8.99649
+▁mai -8.99796
+isierung -8.99821
+ähr -8.99938
+▁reich -9.00044
+▁gewinn -9.00053
+▁blut -9.00221
+▁arbeiten -9.00341
+▁halt -9.00348
+▁gold -9.0049
+▁kleiner -9.00617
+▁gesellschaft -9.00663
+▁funktion -9.00672
+▁mus -9.00701
+▁leistung -9.00805
+▁or -9.00971
+mut -9.00994
+▁entwicklung -9.01042
+▁hundert -9.01062
+▁feuer -9.01106
+zer -9.01251
+▁sprache -9.01358
+▁länger -9.01407
+▁kauf -9.01416
+▁inter -9.0153
+system -9.01552
+tlich -9.01671
+▁ach -9.01697
+▁fri -9.01924
+isten -9.01941
+▁volks -9.02336
+▁fertig -9.02396
+▁straße -9.02421
+▁beginn -9.0265
+schrei -9.02698
+ag -9.02729
+▁wahrscheinlich -9.02828
+▁gefunden -9.0283
+▁rat -9.02944
+▁endlich -9.0299
+▁new -9.03068
+▁raus -9.03271
+▁projekt -9.03272
+eiß -9.03375
+▁müsste -9.03464
+▁mensch -9.03514
+▁karriere -9.03536
+sprache -9.03697
+scheid -9.03917
+▁sätze -9.03944
+▁punkt -9.03996
+▁nächste -9.04024
+▁liebe -9.04064
+▁einzel -9.04095
+▁reicht -9.04146
+▁richtung -9.04235
+weit -9.04366
+▁nimmt -9.04368
+ding -9.04421
+kirche -9.0446
+▁post -9.0448
+▁tu -9.04556
+▁bad -9.04645
+station -9.047
+▁leider -9.04702
+▁strom -9.04763
+▁wesentlich -9.04775
+▁städte -9.04792
+▁sah -9.04803
+▁aufgabe -9.0482
+▁hauptstadt -9.04839
+▁amt -9.0501
+▁vorher -9.05023
+geschichte -9.05052
+genommen -9.05089
+▁junge -9.05157
+ki -9.05231
+▁zusätzlich -9.05267
+▁werke -9.05627
+▁zer -9.05656
+▁gruppe -9.05663
+▁kirchen -9.05701
+▁unbe -9.0571
+▁weniger -9.05756
+▁uhr -9.05811
+▁bri -9.05924
+▁eng -9.05992
+▁meiner -9.06002
+gesellschaft -9.06031
+▁heiß -9.06163
+schau -9.06231
+▁tage -9.06277
+▁team -9.06337
+zig -9.06339
+▁dr -9.06575
+▁lied -9.06605
+▁versuch -9.06617
+▁veröffentlicht -9.06759
+▁druck -9.06761
+ud -9.06845
+ron -9.06943
+▁u -9.06947
+▁hol -9.07027
+▁polizei -9.07039
+schiff -9.07058
+▁nacht -9.07148
+▁wi -9.07159
+▁gi -9.07176
+gegen -9.07195
+▁bruder -9.07262
+▁besuchte -9.07285
+▁künstler -9.07301
+▁komplett -9.07767
+▁deutlich -9.07768
+if -9.07774
+▁klasse -9.07823
+▁bisher -9.07975
+▁hör -9.07995
+hof -9.0802
+plan -9.08057
+film -9.08102
+trieb -9.08227
+▁befinden -9.08274
+▁abend -9.08343
+▁regional -9.08389
+▁setzt -9.08471
+▁übrig -9.08785
+▁modern -9.08796
+▁op -9.08905
+▁regen -9.09197
+▁kenn -9.09535
+▁allein -9.09537
+▁breit -9.09669
+▁prä -9.09681
+▁saison -9.09713
+wirtschaft -9.09779
+▁allgemein -9.09813
+▁vergessen -9.09813
+▁höher -9.09862
+schlagen -9.09888
+▁zug -9.09951
+tik -9.10009
+▁sprach -9.1009
+gu -9.10118
+▁seiten -9.10186
+last -9.10227
+▁rund -9.10244
+xi -9.10256
+▁wochen -9.10277
+▁hamburg -9.10332
+▁frauen -9.1034
+▁selb -9.10394
+treten -9.1045
+▁keiner -9.105
+▁dreh -9.10664
+▁sand -9.10719
+▁neuer -9.10737
+getragen -9.10746
+▁wuchs -9.10853
+▁geboren -9.1086
+▁normal -9.10869
+wegen -9.10889
+tritt -9.10911
+stä -9.10971
+▁machte -9.11133
+fl -9.11209
+sitz -9.11334
+▁gast -9.11364
+▁wahr -9.11383
+▁problem -9.11415
+▁gesamte -9.11423
+▁paul -9.11487
+▁sollen -9.11688
+weile -9.11811
+▁stunden -9.11966
+üss -9.12051
+▁wichtige -9.12286
+stände -9.12364
+musik -9.12539
+▁fragen -9.12624
+▁blau -9.12806
+▁gleichen -9.1282
+▁entstanden -9.12825
+wagen -9.13029
+▁cha -9.13085
+erei -9.13143
+▁weise -9.13225
+gemeinde -9.13421
+▁fuß -9.13521
+▁gas -9.13543
+kreis -9.13558
+▁hohe -9.13597
+▁guten -9.13627
+▁bild -9.13633
+statt -9.1366
+wehr -9.14011
+▁nördlich -9.14035
+▁ur -9.14065
+les -9.14138
+▁georg -9.1422
+▁tut -9.14394
+▁tiere -9.14458
+fach -9.14495
+▁zählt -9.14576
+du -9.14586
+▁dra -9.1467
+len -9.14736
+gegeben -9.14794
+▁habt -9.14862
+▁bestand -9.14899
+j -9.14906
+▁i -9.14983
+▁lehrer -9.15068
+▁hall -9.15143
+▁stück -9.15167
+oren -9.15168
+▁kalt -9.15285
+▁starb -9.15335
+▁aller -9.15406
+gelassen -9.15515
+▁har -9.15517
+fang -9.15566
+▁schlaf -9.15612
+führer -9.15644
+▁region -9.15738
+sehen -9.15759
+stück -9.15824
+▁seitdem -9.15924
+▁fenster -9.16224
+▁staats -9.16228
+▁musst -9.16274
+▁verbindung -9.16309
+▁zentral -9.16315
+▁kritik -9.16578
+schloss -9.16749
+▁bedeutet -9.16768
+▁genug -9.16799
+za -9.16833
+▁teilweise -9.16888
+lig -9.16935
+▁innen -9.1694
+▁bevölkerung -9.16976
+länder -9.17023
+stellt -9.17134
+rich -9.17174
+▁eure -9.17206
+▁führen -9.17218
+▁italien -9.17257
+bu -9.17268
+wasser -9.17336
+bert -9.17438
+▁hä -9.17465
+▁vo -9.17518
+ham -9.17536
+▁finde -9.17561
+pu -9.17586
+▁hi -9.17706
+▁bestehen -9.17716
+oni -9.17819
+▁letzter -9.17858
+schutz -9.17951
+▁kor -9.17991
+▁hart -9.18014
+▁arbeitet -9.18032
+▁woche -9.18216
+ständig -9.18274
+nder -9.18387
+▁bahn -9.18401
+▁gu -9.18457
+▁gerät -9.18481
+▁rad -9.18504
+▁miss -9.18559
+▁kla -9.18651
+▁wirkt -9.18689
+wei -9.18725
+▁seid -9.18784
+▁pl -9.18817
+unterricht -9.18825
+▁errichtet -9.19009
+▁bietet -9.19009
+▁markt -9.19039
+jahr -9.19125
+ruf -9.19167
+gg -9.19201
+noch -9.19208
+▁gewesen -9.19222
+lassen -9.19319
+▁ähnlich -9.19466
+richtung -9.19486
+▁elektro -9.19541
+linie -9.19542
+▁car -9.19558
+teile -9.19577
+▁kilometer -9.19577
+▁reg -9.19622
+tiert -9.19671
+▁spieler -9.19674
+▁heilig -9.19681
+bank -9.19684
+▁kleinen -9.19686
+rin -9.19729
+▁kur -9.19745
+meisterschaft -9.19887
+tür -9.19949
+▁sommer -9.19982
+▁unser -9.2007
+▁außen -9.2015
+▁älter -9.20152
+▁winter -9.20154
+▁bla -9.20176
+▁stelle -9.20189
+▁tier -9.2029
+▁wind -9.20331
+▁dom -9.20333
+schicht -9.2038
+▁namens -9.20509
+zwischen -9.20697
+▁anti -9.20722
+▁mu -9.20818
+▁blüten -9.20843
+▁hans -9.21053
+▁hinzu -9.21087
+▁kultur -9.21119
+▁th -9.21177
+▁sal -9.21203
+wesen -9.21262
+▁legt -9.21288
+▁insgesamt -9.21301
+▁nutz -9.21394
+▁str -9.21425
+▁wechsel -9.21452
+▁dürfen -9.21882
+ika -9.21911
+licht -9.22037
+ris -9.22082
+▁stimme -9.22088
+▁orts -9.22174
+setzung -9.22243
+grün -9.2229
+▁trotz -9.22415
+ius -9.22434
+▁kaiser -9.2247
+▁trau -9.22662
+▁strecke -9.22786
+▁schuld -9.229
+▁trink -9.22933
+▁schüler -9.23056
+▁mittler -9.23093
+ierten -9.23142
+▁serie -9.23394
+▁bahnhof -9.2341
+innen -9.23785
+▁tot -9.23876
+steigen -9.23958
+▁mitglieder -9.23973
+gruppe -9.24085
+▁glaub -9.24119
+▁hauptsächlich -9.24242
+leitung -9.24249
+stätte -9.24305
+▁internationale -9.24312
+▁deck -9.24336
+▁förder -9.24473
+▁ni -9.24586
+▁österreich -9.24618
+▁familien -9.24629
+▁still -9.24633
+▁fin -9.24677
+dauer -9.24713
+▁schrieb -9.24842
+▁offenbar -9.24914
+▁fahrzeug -9.24971
+je -9.25
+▁ausbildung -9.2504
+▁kosten -9.25133
+wissenschaftlich -9.25198
+▁gering -9.25241
+selben -9.25355
+▁ti -9.25369
+▁port -9.25384
+▁tri -9.25439
+▁pri -9.2547
+schließlich -9.25737
+▁baum -9.25809
+▁glauben -9.2581
+mat -9.25843
+▁mor -9.25854
+▁glas -9.25919
+▁paris -9.26146
+wald -9.26158
+▁gl -9.26184
+▁schrift -9.26191
+fläche -9.26228
+▁rede -9.26381
+mus -9.26385
+▁oh -9.26609
+▁regelmäßig -9.26672
+▁mund -9.26732
+▁bezeichnung -9.26735
+arzt -9.26758
+▁motor -9.26774
+ette -9.26957
+▁angebot -9.27053
+▁verkehrs -9.27117
+▁völlig -9.27273
+▁relativ -9.27273
+▁internet -9.27277
+rang -9.27395
+▁heimat -9.27425
+▁ruf -9.27487
+▁provinz -9.27518
+▁hände -9.27605
+▁grenze -9.27621
+▁unterschied -9.27636
+richten -9.27783
+nach -9.27897
+▁fußball -9.27899
+▁lernen -9.27908
+▁ke -9.27926
+▁sprechen -9.28016
+▁feld -9.28042
+karte -9.28118
+▁kriegs -9.28242
+▁glück -9.28295
+▁einigen -9.28314
+▁min -9.28453
+▁zwölf -9.28511
+by -9.28551
+▁bre -9.28566
+▁entsprechend -9.28595
+▁fried -9.28793
+▁dinge -9.28803
+▁größere -9.29067
+▁ph -9.29079
+▁südlich -9.29126
+▁mann -9.29191
+▁hof -9.29426
+▁chi -9.29544
+old -9.29741
+▁sofort -9.29766
+effekt -9.29766
+▁münchen -9.29768
+führen -9.29777
+▁schreiben -9.29779
+▁mach -9.29785
+arm -9.29804
+▁stra -9.29806
+▁zi -9.29813
+▁zeug -9.29827
+▁hervor -9.29828
+▁stil -9.29834
+▁konnten -9.29878
+zog -9.30205
+▁bauch -9.30333
+prüf -9.30371
+▁ton -9.30378
+▁schmeckt -9.30398
+▁geprägt -9.30398
+▁vorhanden -9.30404
+▁interesse -9.30414
+ingen -9.30429
+fahrer -9.30431
+nfalls -9.30529
+▁verschieden -9.30561
+spieler -9.30597
+buch -9.30628
+turm -9.30643
+▁bal -9.30709
+▁ganzen -9.30724
+blätter -9.30744
+ologie -9.30747
+▁derzeit -9.30759
+▁westlich -9.30818
+pass -9.30908
+▁mon -9.30928
+▁zieht -9.31038
+▁fiel -9.31234
+namen -9.31345
+▁jedem -9.31351
+let -9.3145
+▁museum -9.31466
+▁meister -9.31532
+▁meistens -9.31596
+führt -9.31686
+▁übernahm -9.31704
+▁vergleich -9.31707
+▁trotzdem -9.31764
+ath -9.3188
+▁literatur -9.31947
+▁back -9.32022
+▁ball -9.32041
+▁fern -9.32149
+▁christ -9.32184
+gericht -9.32224
+▁bilden -9.32249
+▁technisch -9.32321
+▁links -9.32322
+▁pflanzen -9.32354
+ände -9.32498
+witz -9.32499
+▁kampf -9.32561
+▁kennt -9.32589
+▁zahl -9.32606
+▁spricht -9.32639
+▁jeweil -9.32724
+▁eh -9.32796
+▁gesagt -9.32816
+val -9.32974
+▁gewisse -9.32975
+ly -9.33137
+▁j -9.33164
+▁programm -9.33216
+▁mess -9.33224
+schluss -9.33229
+▁folgen -9.33258
+schütz -9.33284
+▁geschäfts -9.33325
+familie -9.33368
+verwaltung -9.33476
+ück -9.33505
+▁gesamt -9.33583
+▁äußer -9.33624
+▁general -9.33629
+▁tro -9.3368
+▁studium -9.33736
+die -9.33766
+▁bücher -9.33782
+▁ber -9.33835
+▁beginnt -9.33872
+▁thema -9.33963
+gebäude -9.34038
+lose -9.34194
+leit -9.3422
+▁dient -9.34229
+▁gewählt -9.34282
+▁verbunden -9.34286
+stärke -9.34296
+brü -9.34442
+▁erfolgt -9.34528
+▁einfluss -9.34704
+▁kindern -9.34713
+▁lebte -9.3475
+▁anderer -9.34766
+▁sonne -9.34802
+▁verstehen -9.34806
+▁wetter -9.3495
+anstalt -9.34951
+▁wichtigste -9.34964
+lief -9.34992
+▁daraus -9.34992
+ativ -9.35128
+▁fährt -9.35207
+▁steck -9.35287
+▁tauch -9.35304
+▁vertrag -9.35359
+iell -9.35369
+▁deutsch -9.35391
+unter -9.35532
+▁staat -9.35592
+spruch -9.35612
+▁enthält -9.35654
+▁sinn -9.35665
+▁tragen -9.35718
+bogen -9.35732
+▁dicht -9.35784
+mannschaft -9.35843
+▁europäische -9.35904
+mlich -9.35944
+kunde -9.35946
+kal -9.36038
+sätze -9.36084
+▁modell -9.36265
+förmig -9.36282
+▁männer -9.36283
+fest -9.36329
+kurs -9.3644
+ost -9.36468
+▁gestern -9.36506
+▁gesicht -9.36723
+ac -9.36756
+▁spar -9.36796
+bra -9.36835
+kauf -9.36942
+▁linie -9.37004
+▁anderes -9.37047
+mäßig -9.37127
+gemeinschaft -9.37146
+sweise -9.37164
+▁kennen -9.37271
+▁finanz -9.37313
+▁stern -9.37326
+▁personen -9.37376
+▁trägt -9.37384
+▁nähe -9.37396
+win -9.374
+hält -9.3744
+hundert -9.37442
+deutsch -9.37486
+stet -9.37489
+▁beschäftigt -9.37638
+leiste -9.37641
+wort -9.377
+off -9.37726
+▁tal -9.3773
+yl -9.37813
+ock -9.37861
+▁tages -9.38089
+krieg -9.38149
+▁stau -9.38175
+▁gegeben -9.38192
+▁hund -9.38285
+band -9.38345
+▁gib -9.38347
+▁trug -9.38348
+▁schreib -9.38378
+▁runde -9.38452
+▁dach -9.38468
+▁nor -9.38581
+richter -9.38599
+▁sammel -9.38713
+gegangen -9.38724
+brechen -9.38818
+wechsel -9.38862
+▁international -9.38935
+ay -9.39
+▁tatsächlich -9.39013
+▁älteste -9.39013
+▁hilfe -9.39098
+▁heim -9.39158
+▁galt -9.39192
+▁automatisch -9.39261
+▁end -9.39267
+ica -9.39338
+jährige -9.39342
+▁temperatur -9.39367
+▁fehler -9.3937
+▁zuvor -9.39433
+▁mari -9.39487
+anlage -9.39494
+▁setzte -9.39506
+▁pre -9.39534
+▁tat -9.39608
+imm -9.39651
+betrieb -9.39655
+▁produktion -9.39683
+▁beliebt -9.39707
+stuhl -9.39711
+▁besitzt -9.39755
+▁besuch -9.39758
+▁au -9.3982
+▁schauspieler -9.39968
+▁erster -9.39997
+▁ziehen -9.40044
+▁dick -9.40069
+▁app -9.40198
+▁politik -9.40255
+körper -9.4029
+▁gericht -9.40322
+▁ve -9.40352
+▁tre -9.40383
+▁bloß -9.40407
+dia -9.40415
+änderung -9.40438
+▁chor -9.4045
+steuer -9.40474
+▁fehlt -9.40481
+▁zentrum -9.40509
+▁stellte -9.40528
+▁armee -9.40657
+dy -9.40687
+▁glaube -9.4083
+▁folgte -9.40865
+▁kamen -9.40896
+▁politische -9.40929
+▁com -9.40964
+▁haar -9.40996
+zentrum -9.41008
+handlung -9.41032
+▁streit -9.41066
+▁rhein -9.41111
+▁hell -9.41113
+▁erneut -9.41114
+▁verwenden -9.41116
+▁bekommen -9.41162
+wachs -9.41166
+▁finger -9.41178
+▁organ -9.41215
+▁tanz -9.4123
+▁steuer -9.41245
+rot -9.41258
+▁album -9.41282
+▁peter -9.41412
+pfen -9.41481
+hilfe -9.41734
+kehr -9.41747
+▁liste -9.41763
+▁schwierig -9.41819
+▁frankreich -9.41833
+▁hohen -9.41883
+sprung -9.41886
+▁überwiegend -9.41949
+rand -9.4206
+▁schwe -9.42086
+▁jahres -9.42128
+▁bett -9.42213
+iker -9.42229
+▁bank -9.4226
+▁norden -9.42394
+über -9.42474
+fälle -9.42489
+▁helfen -9.42534
+kräfte -9.42534
+▁probleme -9.4255
+red -9.42557
+▁silber -9.4257
+▁ebenso -9.42577
+▁wand -9.42671
+▁folgten -9.42683
+▁sam -9.42688
+▁mö -9.42716
+insel -9.42737
+▁bringen -9.42781
+▁stamm -9.4291
+▁fliegen -9.43059
+▁satz -9.4307
+▁kraft -9.43136
+▁schutz -9.43185
+gezogen -9.43233
+blick -9.43268
+▁stell -9.4336
+▁straf -9.43372
+glieder -9.43392
+sport -9.43446
+▁erreicht -9.43519
+ahn -9.43575
+▁nahe -9.4363
+▁fehl -9.43743
+▁gruppen -9.43951
+▁university -9.43978
+▁ledig -9.43981
+▁wappen -9.44012
+▁gebaut -9.44039
+▁verd -9.44072
+▁süden -9.44083
+stö -9.44196
+▁gelb -9.44199
+pper -9.44208
+▁größten -9.44233
+▁ernst -9.4424
+▁meter -9.44391
+stern -9.44597
+▁lesen -9.44659
+arten -9.44665
+▁friedrich -9.44721
+▁ergebnis -9.44724
+▁ehren -9.44739
+▁vollständig -9.44744
+lager -9.44797
+▁umfasst -9.44888
+▁teile -9.44985
+▁mer -9.45032
+▁beruf -9.45237
+▁industrie -9.45385
+▁bekommt -9.45446
+▁überall -9.4545
+schließ -9.45601
+▁schritt -9.45738
+▁johann -9.4581
+arch -9.45839
+bus -9.45965
+▁hören -9.46003
+▁fo -9.46004
+▁arten -9.46151
+▁lauf -9.46155
+▁verfügt -9.46184
+stock -9.46193
+technik -9.46193
+prüfung -9.46216
+▁militär -9.46219
+preis -9.46349
+▁reihe -9.46351
+▁verwaltungs -9.46361
+tät -9.46361
+weiß -9.46377
+▁karl -9.46394
+▁garten -9.46503
+▁typ -9.46545
+▁dank -9.46555
+▁lösung -9.46691
+▁kamera -9.4675
+▁treffen -9.46752
+ian -9.46762
+ther -9.4683
+▁wechselte -9.46896
+▁leitet -9.46937
+▁wissenschaft -9.46945
+▁red -9.46971
+▁wild -9.46985
+▁martin -9.46997
+▁dagegen -9.47103
+▁aufnahme -9.47282
+▁ed -9.47385
+▁bildung -9.4739
+▁seh -9.47427
+wind -9.47473
+▁regierung -9.47522
+gewiesen -9.47684
+▁beziehung -9.47827
+▁brenn -9.47921
+▁sc -9.47978
+▁verloren -9.48032
+▁entstand -9.48107
+hand -9.48348
+▁monate -9.48386
+org -9.48434
+▁london -9.48441
+▁köln -9.48446
+▁treib -9.4847
+▁vogel -9.48515
+▁mehrfach -9.48588
+▁sing -9.48594
+ball -9.48723
+▁schluss -9.48761
+anlagen -9.48839
+fü -9.49001
+▁schweiz -9.49011
+▁hört -9.4903
+▁ag -9.49088
+▁miteinander -9.49148
+▁energie -9.49205
+▁nachfolger -9.49209
+▁brachte -9.4924
+▁stich -9.49314
+frage -9.49367
+▁anzu -9.49514
+▁westen -9.49696
+▁real -9.49734
+▁con -9.4978
+ini -9.49872
+▁müll -9.49913
+▁politisch -9.49972
+▁entscheidung -9.49973
+▁befand -9.49973
+▁gelten -9.50045
+▁beendet -9.50054
+flügel -9.50055
+▁song -9.50114
+▁institut -9.50158
+hängen -9.50181
+cho -9.50285
+pflicht -9.50297
+▁zwanzig -9.50298
+▁haut -9.50322
+schul -9.50462
+tausch -9.50488
+▁zweck -9.5051
+▁präsident -9.50669
+▁vorbei -9.50695
+stehen -9.50705
+▁hü -9.50732
+▁egal -9.50749
+▁unabhängig -9.50768
+▁enthalten -9.50863
+ara -9.50885
+sicher -9.50955
+▁zuerst -9.50997
+ari -9.51016
+stoffe -9.51051
+▁studien -9.51154
+▁feier -9.5127
+▁schwester -9.51313
+sort -9.51341
+▁bringt -9.51398
+hütte -9.51424
+▁reis -9.51434
+▁möglichkeit -9.515
+▁technik -9.51519
+▁erklärt -9.51529
+▁benutzt -9.5153
+▁lokal -9.51568
+▁zwar -9.51606
+▁firma -9.51628
+▁gegensatz -9.51644
+maschine -9.51667
+sieg -9.51739
+▁radio -9.51788
+▁suche -9.51853
+▁frisch -9.51943
+▁spät -9.52057
+▁entscheide -9.5212
+horn -9.52303
+▁erstmals -9.52319
+▁john -9.52325
+▁grau -9.52424
+▁trainer -9.52506
+uß -9.52556
+▁wagen -9.52561
+gl -9.52602
+fuhr -9.52706
+treffen -9.5295
+finale -9.53013
+▁pal -9.5303
+fo -9.53073
+wende -9.53099
+▁entfernt -9.5311
+medaille -9.5311
+▁erschien -9.53116
+takt -9.53124
+gänge -9.53132
+▁pan -9.53165
+▁schlacht -9.53171
+▁verfahren -9.53173
+hu -9.53213
+▁cor -9.53232
+runde -9.53235
+kraft -9.53291
+kette -9.53366
+▁gesetz -9.53381
+▁ring -9.5349
+▁prozess -9.53556
+▁leucht -9.53578
+▁liga -9.53664
+▁jung -9.53704
+hard -9.53759
+don -9.53903
+▁zuletzt -9.53911
+▁einwohner -9.53917
+▁gespielt -9.53927
+▁drauf -9.53955
+▁wider -9.54006
+fasst -9.54011
+▁dauert -9.54041
+▁zuständig -9.54056
+kuchen -9.54103
+leistung -9.54115
+▁passiert -9.54131
+▁sag -9.54201
+äre -9.54374
+▁antwort -9.54408
+▁geb -9.54497
+▁elf -9.54538
+▁nimm -9.54577
+▁verhalten -9.54588
+geteilt -9.54661
+▁wirtschafts -9.54798
+herr -9.54888
+▁zeigen -9.54998
+well -9.5507
+lü -9.55254
+ino -9.5534
+▁ständig -9.55488
+▁frag -9.55527
+▁unterstützt -9.55529
+▁öffne -9.55529
+▁östlich -9.5553
+▁klassische -9.55532
+▁überrasch -9.55534
+▁scheint -9.55583
+▁spring -9.55596
+▁gö -9.55606
+▁entstehen -9.55606
+▁wohnung -9.55681
+uf -9.55682
+▁friedhof -9.55777
+▁char -9.55949
+titel -9.56081
+kammer -9.56119
+▁wächst -9.56349
+▁praktisch -9.5635
+▁erkennen -9.56353
+▁hab -9.56383
+▁wirtschaftlich -9.56404
+wissenschaften -9.56408
+▁schwerpunkt -9.56428
+▁auszu -9.56431
+▁lautet -9.56709
+▁mä -9.56725
+ktion -9.56776
+▁treten -9.56776
+▁stimmt -9.56865
+hausen -9.56887
+produkt -9.56961
+▁existiert -9.57176
+▁alternativ -9.57176
+▁fallen -9.57195
+▁siedlung -9.57261
+gar -9.57287
+▁hei -9.57288
+abteilung -9.574
+▁milch -9.57462
+▁euro -9.57536
+▁verlassen -9.57536
+▁ecke -9.57731
+reise -9.57909
+träge -9.5792
+▁korrekt -9.57963
+▁heutigen -9.57985
+▁geworden -9.58009
+▁kaffee -9.58009
+▁aktuell -9.58009
+▁rauch -9.5807
+▁reichs -9.58083
+▁passt -9.58147
+▁platt -9.5832
+▁tätigkeit -9.58334
+▁eier -9.58339
+▁lager -9.5847
+ov -9.58527
+▁rote -9.58547
+▁erreichte -9.58582
+▁sucht -9.58662
+lag -9.58731
+▁motiv -9.588
+fähig -9.58817
+▁gebrauch -9.58856
+maßnahmen -9.58857
+▁baby -9.5888
+▁entwickelte -9.58913
+▁britische -9.58931
+vier -9.58948
+ya -9.58957
+denken -9.59006
+▁historische -9.59041
+▁verläuft -9.59095
+▁traum -9.5918
+▁pla -9.59209
+ober -9.59217
+▁wal -9.5928
+▁geschrieben -9.59321
+▁meinung -9.5933
+flug -9.59335
+sucht -9.59384
+▁spezial -9.59428
+läuft -9.59449
+cu -9.59466
+▁bus -9.595
+alter -9.59539
+▁konzert -9.59557
+▁entwickelt -9.59633
+▁beteiligt -9.59697
+plätze -9.59697
+▁erlaubt -9.59697
+methode -9.59698
+▁verbreitet -9.59701
+▁wunder -9.59718
+▁charakter -9.59718
+▁air -9.59956
+züge -9.60046
+▁jen -9.60145
+ial -9.60159
+▁laufen -9.60264
+ana -9.60354
+ologische -9.60432
+▁hause -9.60445
+▁flugzeug -9.60514
+▁abzu -9.60542
+▁sogenannte -9.60551
+▁mädchen -9.60552
+▁extrem -9.60552
+gespräch -9.60552
+▁gebe -9.60594
+gymnasium -9.60705
+▁brand -9.6073
+▁bern -9.60734
+gro -9.60735
+▁angriff -9.60752
+▁rücken -9.60765
+oli -9.6083
+▁kasse -9.60858
+grad -9.60899
+wohn -9.609
+halle -9.60908
+▁turm -9.60979
+▁maschinen -9.61043
+▁schach -9.61096
+▁bier -9.611
+komm -9.61132
+▁version -9.61196
+arbeiten -9.61249
+▁leiden -9.61268
+▁wander -9.61299
+▁zukunft -9.61413
+▁york -9.61414
+▁böse -9.61419
+▁maschine -9.61629
+▁regeln -9.61761
+▁test -9.61779
+▁bot -9.61845
+▁einzelnen -9.61872
+▁verband -9.61873
+▁bewegung -9.61983
+▁amerikanische -9.62062
+▁diente -9.62211
+▁forschungs -9.62243
+ski -9.62256
+▁hängt -9.62284
+▁bio -9.62299
+haltung -9.62372
+▁schützen -9.6238
+▁funktioniert -9.62449
+▁juli -9.62459
+▁transport -9.62493
+▁tisch -9.62536
+hob -9.62658
+otten -9.62666
+▁zeitung -9.62702
+inger -9.62819
+▁wichtig -9.62937
+zeitig -9.63107
+küste -9.63153
+▁schriftsteller -9.63161
+▁vermutlich -9.63161
+▁gründung -9.63161
+▁diverse -9.63163
+▁beispielsweise -9.63182
+▁fühlt -9.63187
+theorie -9.6319
+▁hielt -9.63213
+▁erreichen -9.63235
+▁kunden -9.63241
+▁hit -9.63276
+▁lieb -9.63285
+▁landschaft -9.63337
+fassen -9.63428
+▁profi -9.63549
+farbe -9.63578
+stil -9.63601
+wil -9.63637
+▁reichen -9.63714
+▁alb -9.63927
+üll -9.64031
+▁längst -9.6404
+▁katholisch -9.64045
+▁zerstört -9.64045
+undzwanzig -9.6405
+▁par -9.64188
+auftrag -9.64215
+▁pra -9.64293
+▁anwendung -9.64338
+▁erde -9.64473
+ci -9.64512
+▁würden -9.64515
+schuss -9.64583
+dach -9.6464
+prä -9.64664
+▁ze -9.64668
+▁schwarze -9.64893
+▁stabil -9.64913
+▁produziert -9.64938
+á -9.64938
+▁üblich -9.64938
+▁frankfurt -9.64941
+▁fremd -9.64942
+▁herrschaft -9.65027
+▁monat -9.65125
+▁schlag -9.65205
+▁nachbarn -9.65286
+▁weich -9.65469
+▁maria -9.65587
+▁einzelne -9.65601
+ance -9.65643
+▁absolvierte -9.65803
+zel -9.65823
+▁geburt -9.6584
+▁veranstaltung -9.65843
+▁küche -9.65861
+▁para -9.65875
+▁zahn -9.65888
+▁betrieben -9.6593
+liga -9.6596
+▁reise -9.65981
+▁bauern -9.65986
+▁idee -9.6646
+▁erzähl -9.6666
+rett -9.6667
+▁kommst -9.66691
+▁cu -9.66719
+▁original -9.66748
+▁computer -9.66748
+▁söhne -9.66748
+▁wilhelm -9.66748
+▁doktor -9.66748
+▁durchgeführt -9.66748
+▁mindestens -9.66749
+▁flasche -9.66757
+jo -9.66769
+▁geschlossen -9.66791
+▁franz -9.66805
+verbot -9.66816
+lenk -9.66833
+▁rechte -9.66854
+denkmal -9.66865
+▁brauchen -9.669
+machen -9.6693
+▁cla -9.66997
+▁öl -9.6704
+▁zugleich -9.6713
+▁ruhe -9.67275
+kreuz -9.67281
+lehrer -9.67287
+nummer -9.67297
+▁schreibt -9.67321
+▁amts -9.67362
+▁rand -9.67413
+▁gehalten -9.67467
+▁eisenbahn -9.67491
+▁auge -9.67547
+rolle -9.67596
+geschwindigkeit -9.67665
+▁symbol -9.67665
+fabrik -9.67666
+▁verzicht -9.67666
+▁null -9.67715
+▁sprachen -9.6775
+pol -9.67806
+▁geschäft -9.67827
+▁könnten -9.6789
+▁entscheidend -9.67961
+▁sonnen -9.67981
+▁anlage -9.68021
+nennt -9.68034
+▁dü -9.68082
+▁vereins -9.68223
+▁erfolgte -9.68294
+▁güter -9.68308
+▁eu -9.68367
+▁pat -9.68412
+kirch -9.68447
+markt -9.68483
+geschoss -9.68544
+▁typisch -9.68576
+▁evangelisch -9.68591
+▁grundsätzlich -9.68591
+▁umgebung -9.68591
+geschlecht -9.68597
+▁brief -9.68602
+▁versteht -9.6872
+▁geführt -9.68906
+▁hor -9.68919
+schl -9.68942
+▁gelände -9.68994
+post -9.69022
+▁nachbar -9.69041
+liste -9.69067
+bereit -9.69173
+▁oper -9.69294
+stall -9.69316
+▁rang -9.69393
+▁willst -9.69441
+länge -9.69506
+▁county -9.69526
+▁entschied -9.69526
+▁publikum -9.69526
+▁jünger -9.69526
+▁flughafen -9.69532
+▁höchste -9.69571
+▁wolf -9.6961
+▁zucker -9.69639
+▁danke -9.69714
+▁zeitschrift -9.69813
+rennen -9.69829
+▁zeitpunkt -9.69872
+▁osten -9.69873
+▁fett -9.69876
+ley -9.69948
+▁doppel -9.70026
+ordnung -9.70156
+verein -9.70173
+▁versuchen -9.70183
+kämpfe -9.70418
+pflanze -9.70431
+▁digital -9.70469
+▁situation -9.70469
+▁weitgehend -9.7047
+▁medien -9.70505
+▁kontakt -9.70519
+▁bibliothek -9.70546
+▁mode -9.70554
+akte -9.70577
+▁china -9.70612
+▁berufs -9.70665
+liegen -9.7067
+geht -9.70776
+▁spitze -9.70902
+▁lief -9.70931
+▁spiegel -9.71044
+▁maß -9.71061
+bezirk -9.71103
+häuser -9.71104
+▁fleisch -9.71161
+fried -9.71203
+▁ski -9.71205
+bombe -9.71228
+▁artikel -9.71239
+▁che -9.71317
+▁kol -9.71354
+schneid -9.7136
+▁star -9.7139
+▁benötigt -9.71422
+▁plötzlich -9.71422
+▁zoll -9.71425
+▁geplant -9.71438
+▁eingang -9.71466
+fassung -9.71467
+▁landkreis -9.71494
+▁katze -9.71533
+▁schnitt -9.7155
+schalten -9.71554
+▁republik -9.71565
+▁aufgenommen -9.7163
+▁verfasst -9.71717
+▁bäume -9.7187
+▁kontroll -9.71875
+▁mitarbeiter -9.71953
+▁video -9.71959
+rauch -9.7196
+▁beigesetzt -9.71999
+▁bur -9.72031
+bringen -9.72101
+quelle -9.72166
+▁wach -9.72176
+geber -9.72213
+▁prinzip -9.72286
+bär -9.72331
+▁bedeutende -9.72376
+club -9.72392
+▁mel -9.72397
+▁dunkel -9.72401
+▁darunter -9.72409
+star -9.72423
+▁themen -9.72427
+▁denke -9.72439
+▁bereit -9.72481
+▁wem -9.72633
+▁bisschen -9.72764
+zutage -9.72771
+klasse -9.72849
+▁position -9.72881
+▁rum -9.72956
+▁roll -9.73083
+bürger -9.73094
+▁vorstellung -9.73125
+▁quer -9.73286
+waffe -9.73291
+▁august -9.73354
+▁gelegentlich -9.73363
+höfe -9.73434
+▁bezeichnen -9.73509
+gäste -9.73522
+▁zweifel -9.73713
+sache -9.73728
+▁gern -9.73825
+tha -9.74201
+kampf -9.74203
+ndel -9.74208
+gerechnet -9.74211
+trop -9.74312
+▁frank -9.74318
+▁freiwillig -9.74335
+▁knochen -9.74339
+▁stadtteil -9.74344
+▁erwachsene -9.74353
+▁fahrrad -9.74372
+papier -9.74399
+▁mond -9.7448
+▁hätten -9.74488
+▁großer -9.746
+▁ausnahme -9.74609
+▁zweimal -9.74803
+▁play -9.7487
+▁wärme -9.75193
+▁religion -9.75325
+▁gleichnamige -9.75325
+▁soldat -9.75334
+▁hafen -9.75438
+meld -9.75478
+▁lau -9.75501
+▁toll -9.75502
+▁trans -9.75625
+▁beton -9.75725
+pack -9.75749
+reihe -9.75752
+▁konzept -9.75836
+▁professor -9.75856
+büro -9.75884
+ali -9.75911
+wertung -9.75913
+wie -9.75922
+stufe -9.7597
+▁unver -9.76013
+▁koch -9.76066
+▁tour -9.76066
+▁instrument -9.76087
+bücher -9.76109
+▁handy -9.76129
+▁vertreter -9.76133
+▁stammen -9.76168
+▁mittelalter -9.76172
+istischen -9.76178
+▁bü -9.76188
+▁trag -9.7621
+staat -9.76289
+▁funk -9.76317
+geleitet -9.7634
+▁hunger -9.76343
+▁zunehmend -9.76376
+▁entgegen -9.76412
+▁england -9.76432
+rry -9.76445
+▁handels -9.76517
+▁linke -9.76614
+lied -9.76643
+fleck -9.76751
+▁schade -9.76809
+▁kron -9.76936
+▁sänger -9.76956
+programm -9.77148
+▁darin -9.77168
+stimmung -9.77329
+bauten -9.77333
+▁telefon -9.77335
+▁zählen -9.77335
+▁himmel -9.77335
+▁signal -9.77335
+▁schieß -9.77347
+▁brücke -9.77349
+▁jungen -9.77375
+▁standort -9.77397
+part -9.77507
+▁vertreten -9.77819
+fahr -9.77863
+presse -9.78036
+▁irgendw -9.78121
+pft -9.78211
+teilung -9.78215
+▁schaffen -9.7823
+▁bilder -9.78233
+verband -9.78282
+▁notwendig -9.78355
+versorgung -9.78356
+▁unbedingt -9.78356
+▁michael -9.78357
+ria -9.78371
+▁italienische -9.78477
+front -9.78507
+stab -9.78717
+spur -9.78737
+olog -9.78757
+▁besucher -9.78781
+▁heut -9.78974
+▁regierungs -9.79008
+geist -9.79256
+▁künstlich -9.79386
+▁prozent -9.79388
+▁entdeckt -9.79391
+▁entspricht -9.79416
+▁gestellt -9.7943
+▁pferde -9.79435
+▁vereinigt -9.79456
+utz -9.7955
+▁abgeschlossen -9.79586
+▁stunde -9.79595
+▁sache -9.79649
+▁staaten -9.79705
+blumen -9.79747
+geld -9.7979
+bb -9.79897
+schnitt -9.79988
+grenze -9.80305
+arbeiter -9.80419
+▁töchter -9.80428
+▁heinrich -9.80428
+▁belegt -9.8043
+▁gedanken -9.80432
+▁trocken -9.80433
+▁römische -9.80435
+streckt -9.8044
+▁wende -9.80481
+▁verwendung -9.80578
+▁ähnliche -9.80631
+▁gedicht -9.80737
+▁blatt -9.80761
+▁web -9.8078
+gerichtet -9.80788
+▁näher -9.80801
+kapelle -9.80846
+▁vorne -9.81076
+▁fläche -9.81126
+▁müsst -9.81181
+wunde -9.8119
+ula -9.81195
+▁struktur -9.81207
+mark -9.8129
+▁fahrt -9.81359
+siedlung -9.81371
+▁dor -9.81446
+fertig -9.81455
+▁vermeiden -9.8148
+▁griechisch -9.8148
+▁speziell -9.8148
+▁erheblich -9.81481
+▁unmittelbar -9.81481
+qui -9.81495
+bald -9.81504
+▁gezogen -9.81576
+▁ausgezeichnet -9.81683
+strich -9.81756
+▁sinne -9.81758
+hang -9.8184
+▁ausstellung -9.81853
+▁fließt -9.81893
+▁kern -9.82008
+packung -9.82046
+gelegenheit -9.82052
+streich -9.82141
+weil -9.82142
+▁züge -9.8215
+▁verkehr -9.82177
+▁füll -9.82266
+▁rock -9.82313
+▁zahlen -9.82346
+reaktion -9.82383
+har -9.82431
+▁top -9.82442
+▁organisation -9.82461
+akademie -9.82544
+▁erinnert -9.82544
+verletzung -9.82545
+▁bischof -9.82547
+▁voraus -9.82548
+▁klingt -9.8255
+▁russisch -9.8255
+bildung -9.82551
+▁anwalt -9.82552
+bund -9.82571
+▁kehrte -9.82588
+motor -9.82604
+▁weltweit -9.82622
+alität -9.82678
+▁letzt -9.82721
+▁boot -9.82725
+tive -9.82733
+beutel -9.82734
+▁darum -9.82747
+wüste -9.82772
+▁wenigsten -9.82797
+chi -9.82882
+▁bekannteste -9.82964
+way -9.82968
+▁blätter -9.83033
+▁braun -9.8306
+artig -9.83065
+▁figur -9.83097
+▁nachrichten -9.83215
+▁mont -9.83288
+▁bezirk -9.83404
+klappe -9.8348
+▁wehr -9.83491
+pri -9.83519
+▁weh -9.83564
+▁gefallen -9.83587
+▁eingestellt -9.83593
+würfe -9.83612
+▁komponist -9.83619
+▁schlüssel -9.83622
+▁gewalt -9.83623
+▁schlimm -9.83657
+▁gegenseitig -9.83671
+▁erzielt -9.83677
+▁leer -9.83724
+▁viertel -9.83741
+▁statu -9.83749
+▁person -9.83772
+▁auftritt -9.83799
+▁meer -9.83864
+isse -9.83908
+▁häuser -9.83981
+▁sau -9.84016
+▁vorder -9.84039
+wissenschaft -9.84048
+zweig -9.84148
+zo -9.8424
+▁bach -9.8425
+spende -9.84282
+wochenende -9.84351
+istische -9.84375
+dro -9.84427
+wing -9.84445
+▁anschluss -9.84565
+▁gründen -9.8459
+ani -9.84666
+geschlossen -9.84692
+▁moment -9.847
+▁getrennt -9.84706
+▁hälfte -9.84706
+▁betroffen -9.84706
+▁college -9.84706
+vögel -9.84707
+▁pfarr -9.84708
+▁klavier -9.84722
+▁todes -9.84728
+▁spaß -9.84732
+▁krank -9.84734
+▁zusammenhang -9.84752
+forderung -9.84758
+zimmer -9.84772
+▁jahrhundert -9.84774
+▁herzog -9.84779
+▁verkauf -9.84783
+deln -9.84877
+▁brot -9.84999
+▁tritt -9.85011
+▁personal -9.8508
+boden -9.85117
+▁minister -9.85129
+welt -9.85148
+dian -9.85153
+gruppen -9.85301
+▁holt -9.85391
+▁medizin -9.85413
+kleid -9.85506
+▁fühle -9.85533
+▁unten -9.85635
+▁dumm -9.85704
+▁leichter -9.85738
+▁volk -9.85792
+ó -9.85805
+▁widmet -9.85805
+wörter -9.85806
+studio -9.85806
+▁südwest -9.85806
+rechnung -9.85809
+spannung -9.85819
+fähigkeit -9.85847
+▁aussage -9.8587
+▁verboten -9.8589
+brücke -9.85914
+funde -9.85946
+wick -9.85953
+▁suchen -9.86142
+▁software -9.86272
+▁blu -9.86337
+▁schwein -9.86509
+▁gewicht -9.86553
+▁rom -9.86604
+▁könnt -9.86685
+▁schlagen -9.86706
+▁afrika -9.86916
+▁entstammt -9.86916
+▁nötig -9.86917
+▁kreativ -9.86917
+▁latein -9.8693
+▁schafft -9.86932
+▁gelesen -9.86936
+spreng -9.86957
+▁spanisch -9.86987
+▁angst -9.86993
+held -9.87093
+▁hoffen -9.87101
+text -9.87127
+▁chef -9.87193
+verbindung -9.87224
+▁heutige -9.87592
+▁absicht -9.87623
+wiegend -9.87835
+▁gefangen -9.87912
+variante -9.88043
+▁gegner -9.88044
+stoß -9.88051
+▁fotograf -9.88065
+▁leg -9.88076
+produktion -9.88118
+▁heb -9.88126
+▁gekommen -9.88169
+▁helm -9.88171
+▁jüdische -9.88189
+▁spa -9.88239
+▁beweis -9.88252
+▁flach -9.88327
+▁fein -9.88377
+▁laufe -9.88432
+▁schmal -9.88471
+▁nase -9.88473
+steiger -9.88499
+issen -9.88776
+zone -9.88822
+ai -9.88891
+säure -9.88913
+▁grad -9.891
+städt -9.89103
+▁militärische -9.89123
+▁salz -9.89139
+wärts -9.89161
+▁aktion -9.8917
+▁unbekannt -9.89176
+▁verfügung -9.89176
+▁besetzt -9.89177
+▁ändert -9.89178
+name -9.89179
+▁bewohner -9.8918
+mindest -9.89212
+▁standard -9.89245
+▁fluch -9.89257
+▁robert -9.89268
+▁schließen -9.89274
+▁hochschule -9.89295
+▁sicherheits -9.8932
+▁urteil -9.89327
+▁bor -9.89327
+▁hinten -9.89376
+▁ortschaft -9.89377
+▁frucht -9.89395
+stopp -9.89404
+▁wohnt -9.89407
+▁leid -9.89413
+▁gefahr -9.89418
+▁kochen -9.89512
+▁kühl -9.89521
+▁österreichische -9.89602
+▁ohren -9.89718
+schuhe -9.89728
+gefallen -9.8974
+▁verlor -9.89813
+bot -9.90001
+▁geheim -9.90015
+▁strecken -9.90218
+▁architekt -9.90228
+▁ausgestattet -9.90326
+▁speicher -9.90326
+▁positiv -9.90326
+▁termin -9.90336
+▁quellen -9.90387
+saft -9.90415
+industrie -9.90417
+projekt -9.9042
+braun -9.90479
+schlaf -9.90493
+menge -9.90589
+▁schalt -9.90624
+▁woher -9.90676
+▁mussten -9.90708
+▁eva -9.90723
+▁spi -9.90734
+ix -9.90946
+lichkeit -9.9099
+▁welch -9.91127
+▁ohr -9.91144
+gelaufen -9.9126
+▁tradition -9.91285
+offizier -9.91489
+▁journalist -9.91489
+▁untersuchung -9.91489
+▁hilft -9.91489
+▁wünsche -9.91489
+▁verlieh -9.91489
+▁ernannt -9.91489
+▁süß -9.91489
+▁bahnstrecke -9.91497
+gebrochen -9.91501
+▁entsteht -9.91517
+▁opfer -9.91562
+präsident -9.91608
+modell -9.91643
+▁legen -9.91706
+▁klaus -9.9174
+gerät -9.91788
+▁abschluss -9.91865
+▁orden -9.91916
+formation -9.91973
+▁adel -9.92034
+▁beamte -9.92053
+▁kohl -9.92125
+▁musiker -9.92205
+▁bedarf -9.92278
+▁laub -9.92392
+aktion -9.92419
+▁partner -9.92442
+raub -9.92558
+schrank -9.92664
+í -9.92665
+▁jüngst -9.92665
+▁niedrig -9.92665
+▁hersteller -9.92666
+▁kletter -9.92666
+▁verdient -9.92671
+▁damalige -9.92674
+▁reiß -9.92698
+▁andre -9.92703
+▁verstanden -9.92723
+wirkung -9.92807
+platte -9.92969
+▁nerv -9.92981
+▁hunde -9.93002
+▁tim -9.93008
+gefangen -9.93019
+direktor -9.93287
+▁material -9.93384
+fluss -9.93404
+▁schild -9.93421
+▁max -9.93526
+▁schwimm -9.93663
+▁konflikt -9.93856
+▁stuttgart -9.93856
+▁generation -9.93858
+▁qualität -9.93875
+▁rahmen -9.93881
+▁organisiert -9.93888
+vorsitzende -9.93891
+▁bürgermeister -9.93898
+▁menge -9.9395
+▁vorteil -9.93977
+▁kostenlos -9.93992
+theater -9.94038
+schläge -9.94059
+▁entwe -9.94176
+▁rennen -9.94271
+▁ordnung -9.94283
+brach -9.94519
+weist -9.94672
+▁wu -9.94743
+▁nochmal -9.94955
+wal -9.9496
+seitig -9.9503
+▁gegründet -9.9506
+▁gezeigt -9.95061
+▁nämlich -9.95061
+▁bronze -9.95061
+speise -9.95063
+schlüsse -9.95068
+▁feuerwehr -9.95088
+▁herrscht -9.95092
+▁verwandt -9.95098
+▁orange -9.95104
+▁dargestellt -9.95114
+ness -9.95258
+▁minute -9.95329
+mauer -9.9535
+kleidung -9.95393
+▁verkauft -9.95427
+▁sitzt -9.9545
+▁sachen -9.95461
+▁tipp -9.95554
+▁forschung -9.95757
+daten -9.95967
+▁leitung -9.95972
+▁tau -9.96043
+▁bericht -9.96118
+gut -9.96146
+▁super -9.96161
+▁dachte -9.96224
+▁basis -9.96229
+▁fragt -9.96278
+▁ludwig -9.9628
+▁beschrieben -9.9628
+▁leipzig -9.9628
+▁verließ -9.9628
+division -9.9628
+▁voice -9.96281
+▁ersetzt -9.96281
+▁niemals -9.96283
+▁protest -9.96301
+▁ermöglicht -9.96311
+▁waffen -9.96344
+dichte -9.96345
+▁sichtbar -9.96355
+▁lebensmittel -9.96368
+▁gestaltet -9.96411
+▁mühl -9.9642
+▁ewig -9.96424
+partner -9.96585
+▁schweizer -9.96625
+lohn -9.96676
+▁schwach -9.96739
+▁nummer -9.96816
+tiere -9.96909
+▁kal -9.96974
+gemäß -9.97101
+▁fotos -9.97184
+▁dokument -9.97201
+▁abhängig -9.97307
+▁sil -9.97323
+▁mauer -9.97324
+behörde -9.9748
+▁persönlich -9.97482
+▁goldene -9.97488
+▁gesundheit -9.97536
+▁informationen -9.9755
+▁kindergarten -9.97551
+▁herunter -9.97592
+brett -9.97644
+▁kino -9.9769
+▁traditionell -9.97731
+▁ausdruck -9.97742
+politik -9.97783
+universität -9.97787
+funktion -9.97798
+▁zunge -9.97818
+gerufen -9.97977
+elli -9.98137
+▁bezirks -9.98179
+▁ol -9.98186
+▁blei -9.98213
+▁fass -9.98309
+▁lügen -9.98376
+▁gesprochen -9.98394
+versuch -9.98413
+▁offizielle -9.98702
+▁kämpft -9.98725
+durch -9.98756
+▁bayern -9.98765
+▁engagiert -9.98765
+▁parallel -9.98765
+▁intensiv -9.98765
+▁brust -9.98766
+▁richtet -9.98766
+▁zivil -9.98766
+▁bedeutendste -9.98767
+▁sendung -9.98773
+▁musikalische -9.98778
+▁sterben -9.98779
+▁otto -9.98779
+▁sachsen -9.98821
+▁sicherheit -9.98822
+▁ufer -9.98905
+▁handlung -9.98913
+säule -9.98947
+▁diesmal -9.98989
+▁geschenk -9.98994
+flu -9.99007
+nimmt -9.99126
+▁ortsteil -9.99158
+▁spitz -9.99167
+▁sub -9.99271
+▁kabel -9.99329
+▁historisch -9.99342
+▁jugendliche -9.99369
+▁bundesliga -9.99448
+▁krankheit -9.99777
+▁erfahrung -9.99827
+lock -9.99874
+▁wirkung -9.99909
+rasse -9.99999
+▁verhindern -10.0003
+öffnung -10.0003
+▁täglich -10.0003
+zähne -10.0003
+▁übersetzt -10.0003
+▁praxis -10.0003
+▁gebildet -10.0003
+▁okay -10.0003
+▁kaputt -10.0003
+▁brüder -10.0003
+▁blind -10.0003
+▁zimmer -10.0008
+wurf -10.0018
+▁umwelt -10.0025
+▁lösen -10.0026
+▁kant -10.0028
+struktur -10.0036
+▁versehen -10.006
+barkeit -10.0066
+▁nah -10.0069
+park -10.0093
+▁gründe -10.0107
+▁irgendwa -10.0113
+▁grenz -10.0114
+▁feind -10.0114
+liegenden -10.012
+▁absolut -10.0131
+▁erklären -10.0131
+▁parlament -10.0131
+▁interessiert -10.0131
+▁wunsch -10.0132
+▁chance -10.0132
+▁erscheint -10.0134
+▁truppen -10.0135
+▁kostet -10.0135
+käse -10.014
+▁hotel -10.0146
+ford -10.0147
+rät -10.0147
+magnet -10.0149
+▁hass -10.0152
+agent -10.0154
+▁gegend -10.0155
+▁fred -10.0159
+▁szene -10.0165
+geschäft -10.0168
+kultur -10.017
+▁single -10.0172
+europa -10.0173
+▁muster -10.0177
+▁brems -10.0192
+▁code -10.0193
+einheit -10.0195
+gesetz -10.0198
+▁aufgestellt -10.0205
+▁tausend -10.0209
+vel -10.0209
+▁masse -10.0217
+▁fähr -10.0218
+▁grenzt -10.0227
+▁element -10.0246
+gesprochen -10.0256
+▁gal -10.0257
+▁hilfs -10.0261
+▁auffällig -10.0261
+▁thomas -10.0261
+▁student -10.0261
+lassung -10.0261
+rezept -10.0262
+sturz -10.0264
+▁schnee -10.0264
+▁common -10.0264
+▁angaben -10.0265
+▁lebensjahr -10.0266
+▁hals -10.0275
+▁strand -10.0278
+haupt -10.0282
+▁fu -10.029
+▁polar -10.0294
+▁hinein -10.0295
+▁verfassung -10.0295
+▁talent -10.0301
+zweck -10.0301
+saal -10.0304
+▁ausgesprochen -10.0305
+▁fund -10.031
+schreiben -10.0316
+▁wartet -10.0316
+reiter -10.0322
+▁maxim -10.033
+▁schick -10.0337
+▁brun -10.034
+▁anzahl -10.034
+▁nutzen -10.0341
+▁nu -10.0357
+▁col -10.0362
+holz -10.0375
+bindung -10.0378
+▁jahrzehnt -10.0393
+veröffentlichung -10.0393
+▁diskussion -10.0393
+▁hannover -10.0393
+▁hierfür -10.0393
+▁netzwerk -10.0393
+festival -10.0394
+geschmack -10.0394
+gewerbe -10.0395
+▁teuer -10.0396
+▁investi -10.0397
+▁klick -10.0399
+▁darstellung -10.04
+▁sprecher -10.0401
+▁fällen -10.0402
+gerissen -10.0404
+forscher -10.0409
+▁folgt -10.0411
+▁endet -10.0421
+▁ortes -10.0427
+▁davor -10.043
+▁passend -10.0431
+gewicht -10.0436
+▁foto -10.0436
+fleisch -10.0437
+museum -10.0437
+spitze -10.0437
+▁jeweilige -10.0447
+olf -10.0468
+ziehung -10.05
+▁kompliziert -10.0526
+▁existieren -10.0526
+▁joseph -10.0526
+▁lieblings -10.0526
+erkennung -10.0526
+ström -10.0527
+rekord -10.0528
+▁nahezu -10.0536
+▁ausgabe -10.054
+gleis -10.0541
+zellen -10.0549
+weltmeister -10.056
+▁regie -10.0563
+▁wiener -10.0568
+euer -10.0572
+bewegung -10.0574
+▁kata -10.058
+▁werfen -10.0585
+▁kräftig -10.0586
+operation -10.0596
+▁wollt -10.0606
+gramm -10.0622
+kanal -10.0623
+▁lind -10.0641
+▁leb -10.0653
+geworfen -10.0661
+läufig -10.0661
+▁vergangenheit -10.0661
+▁ausgebildet -10.0661
+kandidat -10.0661
+▁online -10.0661
+▁umstritten -10.0661
+▁angeblich -10.0661
+▁vorgänger -10.0661
+gebirge -10.0661
+▁verliert -10.0662
+stadion -10.0662
+▁bekam -10.0663
+fisch -10.0663
+▁camp -10.0664
+▁aufgeführt -10.0664
+legung -10.0666
+▁kulturelle -10.067
+▁ländern -10.0673
+▁hergestellt -10.0676
+▁übertragen -10.068
+▁lernte -10.0688
+▁guck -10.0688
+▁my -10.0699
+bad -10.07
+▁erzählt -10.0706
+lösung -10.0722
+▁gitarre -10.0725
+▁wachsen -10.0742
+dreh -10.0756
+ano -10.0767
+▁erhöht -10.0798
+▁geeignet -10.0798
+▁knapp -10.0798
+▁roboter -10.0798
+▁william -10.0798
+▁fieber -10.0798
+konstruktion -10.0799
+▁sowas -10.0803
+▁angesehen -10.0804
+▁kanton -10.0806
+▁zweitausend -10.0811
+anzeige -10.0811
+glück -10.0813
+block -10.0814
+▁angegeben -10.0816
+psycholog -10.0816
+▁cool -10.0817
+ländische -10.0827
+▁produkte -10.0829
+▁witt -10.0831
+▁schauspielerin -10.0833
+▁müller -10.0835
+▁messer -10.0841
+▁sammlung -10.0847
+▁verantwortlich -10.085
+material -10.0853
+▁pause -10.0866
+▁johannes -10.0868
+ausschuss -10.0878
+ash -10.089
+▁machst -10.0896
+getriebe -10.0907
+rseits -10.0911
+▁verhältnis -10.0917
+▁medizinische -10.0919
+schild -10.0919
+▁wettbewerb -10.0923
+ehlt -10.0929
+▁renn -10.0931
+▁geblieben -10.0937
+händler -10.0937
+▁besaß -10.0937
+verbrauch -10.0939
+krebs -10.0939
+stehende -10.0941
+filter -10.0942
+▁bewegt -10.0942
+▁dreck -10.0943
+▁hintergrund -10.095
+▁dünn -10.0951
+▁japan -10.0957
+zeichnung -10.0959
+beeren -10.0981
+tausend -10.0983
+bahnhof -10.1002
+läufer -10.1006
+▁betreten -10.1006
+▁überlegen -10.1007
+schließt -10.1015
+kenntnis -10.1027
+ction -10.1031
+▁freude -10.1035
+▁vorbild -10.1041
+kind -10.1052
+▁april -10.1078
+▁jährlich -10.1078
+▁apfel -10.1078
+▁gefährlich -10.1078
+merkmal -10.1078
+▁detail -10.1079
+▁ursprung -10.1081
+metall -10.1081
+▁bieten -10.1082
+▁stink -10.1082
+schuppen -10.1085
+▁straßenbahn -10.1086
+▁verteilt -10.1086
+▁merken -10.1088
+▁stahl -10.1096
+▁angenommen -10.1098
+stimme -10.1105
+ministerium -10.1105
+dank -10.111
+▁beiträge -10.1113
+▁schmi -10.1117
+wahl -10.1125
+▁einrichtung -10.1126
+▁union -10.1126
+▁angelegt -10.1137
+möglich -10.1138
+minister -10.1139
+geschrieben -10.1141
+prozess -10.1141
+▁riech -10.1149
+▁ferner -10.1153
+▁studie -10.1155
+hagen -10.1168
+▁job -10.1182
+▁heide -10.1194
+▁bewusst -10.1218
+disziplin -10.1221
+therapie -10.1221
+▁begrenzt -10.1221
+▁berühmt -10.1221
+▁solang -10.1221
+▁verbrachte -10.1221
+▁weiblich -10.1221
+▁besatzung -10.1221
+▁dunkle -10.1221
+▁zumeist -10.1223
+▁widerstand -10.1228
+▁zell -10.1228
+siebzig -10.123
+schwer -10.1231
+▁sturm -10.1249
+▁schlange -10.1255
+▁gründete -10.1256
+▁eth -10.1257
+ziel -10.1267
+rup -10.128
+▁beine -10.1283
+forschung -10.1289
+▁akt -10.1308
+▁gedacht -10.1315
+garten -10.1318
+malerei -10.1335
+▁schmerz -10.1353
+▁geöffnet -10.1366
+▁respekt -10.1366
+▁verändert -10.1366
+ingenieur -10.1366
+▁reagiert -10.1366
+training -10.1366
+▁erweitert -10.1366
+▁perfekt -10.1366
+▁christian -10.1368
+▁wertvolle -10.1368
+▁sorgt -10.1369
+kalender -10.1377
+▁antrieb -10.1383
+otter -10.139
+▁christlich -10.1395
+kenntnisse -10.1425
+position -10.1434
+temperatur -10.1437
+figur -10.1437
+marsch -10.144
+verfahren -10.146
+treue -10.1463
+low -10.1466
+cent -10.147
+▁turnier -10.1478
+▁engel -10.1487
+▁melde -10.1487
+ight -10.1488
+▁turn -10.1488
+▁philosophie -10.1497
+▁gelegen -10.151
+▁vulkan -10.1513
+kapitän -10.1513
+▁herkunft -10.1513
+▁ärger -10.1513
+▁verbrannt -10.1513
+physik -10.1513
+wälder -10.1513
+▁schreck -10.1513
+▁besiedelt -10.1513
+▁paket -10.1514
+▁gabriel -10.1514
+▁irgendein -10.1515
+▁behandelt -10.1515
+▁planeten -10.1515
+▁panzer -10.1516
+▁oberfläche -10.1522
+▁knie -10.1528
+▁passieren -10.1534
+spricht -10.1584
+▁angestellt -10.1586
+sperre -10.1611
+gehalt -10.1625
+▁aussehen -10.1631
+thal -10.1638
+▁programmier -10.1639
+siedel -10.166
+unddreißig -10.1662
+▁entwickeln -10.1662
+▁kommunikation -10.1662
+▁regisseur -10.1662
+▁ungefähr -10.1662
+▁kapital -10.1662
+▁erinnern -10.1662
+▁erkannt -10.1662
+▁verehrt -10.1662
+▁werbung -10.1663
+▁butter -10.1663
+▁wusste -10.1663
+▁ehrlich -10.1666
+gearbeitet -10.1666
+▁barock -10.1667
+▁stör -10.1667
+▁berichtet -10.1684
+▁zusammenarbeit -10.169
+ama -10.1702
+▁schale -10.1704
+▁betrachtet -10.1712
+▁schieb -10.1724
+▁schauspiel -10.1736
+karriere -10.1743
+legende -10.1746
+▁umgeben -10.1757
+luft -10.1759
+▁karten -10.1762
+besitz -10.1766
+▁gall -10.1773
+griffen -10.1788
+mehr -10.179
+▁japanische -10.1793
+▁sach -10.1794
+gefühl -10.1797
+braten -10.1804
+▁verlauf -10.181
+▁bestätigt -10.1813
+▁eröffnet -10.1813
+▁asteroid -10.1813
+▁urlaub -10.1813
+magazin -10.1814
+sprüche -10.1814
+▁männlich -10.1814
+▁scharf -10.1814
+▁ursache -10.1816
+▁riesige -10.1816
+▁tourismus -10.1816
+versicherung -10.1818
+▁jurist -10.1822
+▁kämpfen -10.1822
+▁autobahn -10.1823
+▁tank -10.1826
+verständnis -10.183
+lampe -10.1832
+typ -10.184
+▁bestandteil -10.1842
+▁freiburg -10.1857
+▁heiz -10.1864
+pokal -10.1868
+▁frühstück -10.1871
+lücke -10.1871
+fällt -10.1877
+strecke -10.1879
+förderung -10.1908
+▁eingerichtet -10.1925
+empfind -10.1932
+beruf -10.1933
+ebene -10.1956
+strahlung -10.196
+▁unterstützung -10.1961
+▁pferd -10.1962
+▁gesund -10.1967
+gouverneur -10.1967
+revolution -10.1967
+▁beeinflusst -10.1967
+geräusch -10.1967
+schleunig -10.1967
+▁dürfte -10.1967
+▁geschwister -10.1967
+▁gewöhn -10.1967
+▁anonym -10.1967
+▁löwe -10.1967
+▁versteckt -10.1967
+▁unklar -10.1968
+▁bevorzugt -10.1968
+jäger -10.1968
+▁verlust -10.1969
+▁rückwärts -10.1969
+▁behörden -10.1973
+▁kriegsende -10.1976
+▁schlug -10.1976
+▁schwanz -10.1978
+fix -10.1991
+leute -10.1994
+▁trick -10.2002
+▁med -10.2008
+truppe -10.2012
+tragen -10.2025
+▁damen -10.2025
+glocke -10.2027
+▁doppelt -10.204
+röhre -10.204
+herausgeber -10.2056
+▁wirken -10.206
+partei -10.2064
+ulation -10.2065
+freund -10.2069
+▁live -10.2093
+▁heil -10.2095
+phon -10.2095
+▁histor -10.2097
+▁rö -10.2103
+▁täusch -10.2104
+▁dreißig -10.2123
+▁demokratische -10.2124
+▁dialekt -10.2124
+▁empfehle -10.2124
+▁gipfel -10.2124
+trunken -10.2124
+▁christoph -10.2124
+▁hieß -10.2124
+▁verlegt -10.2124
+früchte -10.2124
+▁lustig -10.2124
+▁enorm -10.2125
+▁pilot -10.2125
+▁schall -10.2125
+▁touristen -10.2126
+▁ordentlich -10.2126
+▁pfeif -10.2138
+▁erscheinen -10.2142
+▁rau -10.215
+▁leisten -10.2152
+▁putz -10.2152
+füge -10.2157
+▁nachweis -10.216
+tasche -10.2161
+ographie -10.2162
+▁traf -10.2163
+lanz -10.2166
+▁auszeichnung -10.2173
+▁bonn -10.2183
+koch -10.2191
+▁löst -10.2203
+▁wichtiger -10.2205
+maß -10.2211
+▁künstlerische -10.2217
+feuer -10.2248
+breite -10.2259
+kohl -10.2263
+kosten -10.228
+ō -10.2282
+▁gespannt -10.2282
+▁übernommen -10.2282
+▁eindeutig -10.2282
+▁erforderlich -10.2282
+▁negativ -10.2282
+▁bedroh -10.2282
+▁verpflichtet -10.2282
+▁schatten -10.2283
+▁abitur -10.2283
+▁erwarten -10.2283
+▁tunnel -10.2284
+▁lecker -10.2284
+▁keller -10.2284
+▁chemische -10.2284
+fressen -10.2288
+▁überzeugen -10.229
+▁grundlage -10.2304
+▁grand -10.2309
+▁indien -10.2312
+▁untersucht -10.2312
+▁weltkrieges -10.2317
+rmaßen -10.2325
+▁dahin -10.2325
+▁lob -10.2331
+schritt -10.2335
+ney -10.236
+▁nutzung -10.2366
+▁tom -10.2367
+▁jean -10.2376
+regierung -10.2383
+unternehmen -10.2383
+werfen -10.2385
+gemacht -10.2392
+litz -10.2395
+▁präg -10.2403
+zeichnen -10.2418
+▁höchst -10.2436
+▁endgültig -10.2444
+▁experiment -10.2444
+▁israel -10.2444
+▁olympische -10.2444
+▁zugänglich -10.2444
+schätzen -10.2444
+vermögen -10.2444
+▁richard -10.2444
+▁wüsste -10.2444
+▁david -10.2444
+▁arabisch -10.2444
+▁dialog -10.2444
+▁nütz -10.2444
+▁donau -10.2445
+▁unterscheiden -10.2445
+▁giftig -10.2447
+▁hierzu -10.2448
+▁sauber -10.2449
+▁münz -10.2452
+▁orgel -10.2465
+▁anstrengend -10.2472
+möglichkeiten -10.2475
+▁abschnitt -10.2476
+▁aufgebaut -10.2491
+▁walter -10.2494
+▁unfall -10.251
+▁kürz -10.2514
+▁gemeint -10.2524
+▁bord -10.2528
+weich -10.253
+lach -10.2537
+▁objekt -10.2542
+ju -10.2543
+▁wich -10.2557
+▁drin -10.257
+sammlung -10.2576
+aktiv -10.2594
+produzent -10.2608
+▁begonnen -10.2608
+▁professionelle -10.2608
+▁teufel -10.2608
+▁gewähr -10.2608
+▁drücken -10.2608
+▁irgendwelche -10.2608
+▁flagge -10.2608
+▁klima -10.2608
+▁anhänger -10.2608
+▁konzern -10.2608
+großvater -10.2608
+▁mehrmals -10.2609
+▁komposition -10.2609
+▁brut -10.2609
+achse -10.2611
+tempo -10.2611
+▁laufbahn -10.2613
+archiv -10.2614
+▁kollegen -10.2615
+streifen -10.2615
+▁mission -10.2616
+▁beträgt -10.262
+▁wellen -10.2621
+▁niederländisch -10.2623
+▁schweden -10.2627
+▁gemeindegebiet -10.2632
+▁wall -10.2661
+▁bislang -10.2665
+▁interpret -10.2671
+▁sender -10.2677
+gift -10.2698
+liegende -10.2708
+▁elektrische -10.2712
+▁kap -10.272
+graph -10.2749
+sozialismus -10.2749
+illa -10.2771
+würdig -10.2774
+initiative -10.2774
+komplex -10.2774
+▁konkurrenz -10.2774
+▁tübingen -10.2774
+▁verfolgt -10.2774
+▁verurteilt -10.2774
+▁gemüse -10.2774
+▁langfristig -10.2774
+▁scheiterte -10.2774
+schätzung -10.2774
+▁show -10.2774
+▁froh -10.2774
+▁jazz -10.2775
+▁verlieren -10.2775
+schlägt -10.2775
+begründer -10.2776
+▁atom -10.2778
+▁ernsthaft -10.2779
+weih -10.2786
+▁begleitet -10.279
+▁werbe -10.2794
+schirm -10.2797
+▁gefällt -10.2798
+schwäche -10.2809
+miete -10.2813
+▁bauwerk -10.2815
+uzi -10.2829
+▁kara -10.2831
+▁zeige -10.2833
+▁schuh -10.2866
+region -10.2892
+▁nehme -10.2898
+▁mehrheit -10.2899
+stunde -10.2915
+weisung -10.2934
+▁bemüh -10.2944
+▁düsseldorf -10.2944
+▁weibchen -10.2944
+▁gattung -10.2944
+▁geschehen -10.2944
+▁ändern -10.2944
+▁beherrsch -10.2944
+▁wurzel -10.2944
+▁bedient -10.2944
+▁gelernt -10.2944
+▁erlebt -10.2944
+▁vorsitzender -10.2946
+▁heiratete -10.2946
+▁kirsch -10.2946
+▁botschaft -10.295
+▁flori -10.2956
+▁amtszeit -10.2957
+▁realität -10.2961
+uppe -10.2962
+▁carl -10.2963
+▁taschen -10.2975
+rald -10.2981
+gestaltung -10.2985
+▁total -10.2989
+▁kopie -10.301
+▁sperr -10.3011
+wendung -10.3039
+▁stimm -10.3053
+wirkt -10.3061
+ella -10.3062
+lauch -10.3063
+blatt -10.3066
+friedhof -10.3069
+tätigkeit -10.307
+▁bock -10.3073
+▁senat -10.3076
+seil -10.3082
+▁gefärbt -10.3116
+▁religiöse -10.3116
+▁verteidigung -10.3116
+▁orientiert -10.3116
+▁philipp -10.3116
+assistent -10.3116
+▁rettungs -10.3116
+▁scheinbar -10.3117
+▁schatz -10.3117
+▁billig -10.3117
+▁jagd -10.3117
+▁wäsch -10.3117
+▁ideal -10.3117
+▁kommentar -10.3118
+▁aufmerksamkeit -10.3118
+▁erwies -10.3122
+▁kalk -10.3122
+übung -10.3122
+▁alarm -10.3125
+▁streng -10.3129
+▁kindheit -10.3134
+▁herstellung -10.3138
+▁geschie -10.3148
+nicht -10.315
+▁analyse -10.3151
+▁juden -10.3151
+fassade -10.3157
+▁übersetzung -10.3178
+gärten -10.3181
+zucht -10.322
+istik -10.3229
+entwicklung -10.3246
+▁sack -10.3256
+versammlung -10.3262
+hoff -10.3281
+▁beruflich -10.3287
+tief -10.3288
+frequenz -10.3292
+gegenwärtig -10.3292
+▁restaurant -10.3292
+▁empfang -10.3292
+▁erwähnt -10.3292
+▁verhindert -10.3292
+▁weihnacht -10.3292
+geschaltet -10.3292
+▁russland -10.3292
+▁sauerstoff -10.3292
+konferenz -10.3292
+▁ergibt -10.3292
+▁audio -10.3292
+▁trifft -10.3292
+▁krankenhaus -10.3292
+verleih -10.3292
+▁brandenburg -10.3292
+▁city -10.3292
+▁spanien -10.3292
+▁wiederholt -10.3293
+▁schlagzeug -10.3295
+schäden -10.3304
+▁unsicher -10.3308
+losigkeit -10.3309
+▁vierzig -10.3313
+▁führend -10.3332
+zustellen -10.3332
+▁auswahl -10.3346
+knoten -10.3353
+ladung -10.336
+steht -10.338
+ville -10.3391
+ji -10.3401
+▁gestalt -10.3407
+zeile -10.341
+institut -10.3427
+transport -10.343
+trainer -10.343
+▁blüte -10.3433
+pause -10.3462
+schädigt -10.347
+▁atmosphäre -10.347
+▁einkaufsliste -10.347
+▁geklärt -10.347
+▁großbritannien -10.347
+▁stattfinden -10.347
+gewässer -10.347
+▁experten -10.347
+▁preußische -10.347
+▁chinesische -10.347
+stütze -10.3471
+▁gefertigt -10.3471
+▁theologie -10.3476
+▁gefühle -10.3477
+▁verlag -10.348
+▁notfall -10.3483
+▁öffentlichkeit -10.3487
+▁vize -10.3493
+▁extra -10.3501
+mensch -10.3505
+route -10.3506
+▁münd -10.3512
+▁montag -10.3516
+▁mikro -10.3517
+▁bedenken -10.3526
+▁meeres -10.3535
+▁türkis -10.354
+▁schulter -10.3546
+▁rosen -10.3554
+balken -10.3574
+phase -10.3594
+nteils -10.3598
+vertrag -10.3613
+▁eb -10.3619
+▁benachbarte -10.3652
+▁alexander -10.3652
+▁bedeuten -10.3653
+▁party -10.3653
+▁toast -10.3654
+▁fungiert -10.3654
+▁wespe -10.3654
+oxid -10.3655
+▁probier -10.3655
+darsteller -10.3655
+▁erfunden -10.3655
+▁locker -10.3656
+▁amerikaner -10.3656
+▁saint -10.3657
+scheibe -10.3658
+▁schied -10.3658
+▁offiziell -10.3661
+borg -10.3663
+▁hauptort -10.3667
+▁faul -10.3693
+▁herein -10.3696
+▁lungen -10.3699
+fragt -10.3711
+▁kuh -10.3718
+▁mast -10.3722
+furt -10.3723
+▁att -10.3724
+nationalpark -10.3748
+▁disk -10.3787
+▁steil -10.3789
+▁brauche -10.3807
+▁cap -10.3812
+q -10.3818
+version -10.383
+▁etablier -10.3837
+▁beschränkt -10.3837
+▁gewonnen -10.3837
+▁erfordert -10.3837
+▁aufgelöst -10.3837
+▁tabelle -10.3837
+▁gieß -10.3838
+▁kühlschrank -10.3838
+▁schwedische -10.3839
+▁zuschauer -10.384
+▁unterteilt -10.384
+▁bisherige -10.3842
+wirt -10.3844
+▁rathaus -10.3849
+▁hügel -10.385
+▁albert -10.3853
+▁schneider -10.3854
+▁fair -10.3861
+▁funktionieren -10.3863
+stiel -10.3863
+wieso -10.3865
+▁isa -10.3873
+mitglieder -10.3874
+▁klang -10.388
+iffer -10.3881
+logisch -10.3896
+▁opa -10.3914
+rannt -10.3914
+▁griff -10.3924
+▁lehnte -10.3945
+salat -10.3962
+gekommen -10.3965
+▁luther -10.3984
+studium -10.3993
+fehler -10.3994
+album -10.3994
+ässe -10.3999
+begriff -10.4011
+omi -10.4015
+▁anerkannt -10.4026
+▁identität -10.4026
+▁komponiert -10.4026
+▁sämtliche -10.4026
+▁zerstören -10.4026
+▁erinnerung -10.4026
+▁islam -10.4026
+▁vorgesehen -10.4026
+▁autonom -10.4026
+▁steigt -10.4037
+▁angeles -10.404
+hüt -10.4042
+▁architektur -10.4042
+cia -10.4044
+▁nachhaltig -10.4044
+rechnet -10.405
+▁schuf -10.4051
+▁handwerk -10.4053
+▁villa -10.4059
+kabel -10.4063
+▁dreimal -10.4077
+▁nick -10.4095
+▁antike -10.411
+▁kiel -10.4119
+organ -10.4127
+zustimmen -10.416
+▁matt -10.4168
+fahrzeug -10.4187
+republik -10.4189
+ruhe -10.4201
+macher -10.4202
+bühne -10.4214
+▁augsburg -10.4218
+▁böhm -10.4218
+▁stirbt -10.4218
+▁tastatur -10.4218
+▁verfügbar -10.4218
+▁dringend -10.4218
+▁nahrung -10.4218
+mühle -10.4219
+▁deutschsprachig -10.4219
+▁strömung -10.422
+▁bremen -10.4222
+bemerkt -10.4225
+▁droge -10.4227
+▁dahinter -10.4232
+durchschnittlich -10.4234
+▁selbstständig -10.4238
+▁hoffe -10.4245
+▁franc -10.4258
+▁alkohol -10.4274
+gummi -10.4278
+▁river -10.4279
+▁moor -10.4294
+▁kritisch -10.4297
+soße -10.4313
+▁unterlag -10.4314
+▁empfi -10.4321
+▁rosa -10.4355
+hauen -10.4363
+▁north -10.4369
+firma -10.439
+▁menschliche -10.4408
+▁beobachtet -10.4414
+▁mathematik -10.4414
+▁olympia -10.4414
+▁südöstlich -10.4414
+▁vierzehn -10.4414
+▁batterie -10.4414
+▁umgangssprach -10.4414
+▁zusammengefasst -10.4414
+▁erwartet -10.4414
+lärm -10.4414
+▁präsentier -10.4414
+▁vermutet -10.4414
+debüt -10.4414
+▁aufnehmen -10.4414
+▁email -10.4415
+▁stürm -10.4415
+▁onkel -10.4416
+▁glatt -10.4416
+▁gelingt -10.4416
+▁stefan -10.4418
+▁geriet -10.4418
+▁aufpassen -10.442
+▁eingeführt -10.442
+▁königreich -10.4427
+▁leiche -10.4447
+▁sankt -10.4452
+gebunden -10.4453
+pflege -10.4478
+▁ego -10.4482
+rwerb -10.449
+▁marine -10.4515
+patri -10.4524
+▁kaufmann -10.4545
+▁saure -10.455
+bett -10.4561
+grä -10.4569
+modul -10.4576
+▁schwan -10.4579
+kopf -10.4594
+altar -10.4606
+zylinder -10.4614
+▁akzeptier -10.4614
+trocknet -10.4614
+▁schmecken -10.4614
+▁toilette -10.4614
+▁umgekehrt -10.4614
+▁wählt -10.4614
+gelehnt -10.4614
+▁örtliche -10.4614
+▁überliefert -10.4614
+▁göttingen -10.4614
+erlebnis -10.4614
+▁sekunden -10.4614
+▁schwierigkeiten -10.4614
+▁kommunist -10.4615
+▁jakob -10.4615
+▁prinz -10.4616
+▁schläf -10.4616
+▁auseinandersetz -10.4619
+angehörige -10.462
+▁runter -10.462
+▁golf -10.4621
+▁ausgeführt -10.4622
+geübt -10.4623
+▁wahrheit -10.463
+▁schief -10.4635
+▁jahreszeit -10.4635
+▁saß -10.4637
+elemente -10.4639
+▁bundestag -10.464
+vereinigung -10.4645
+▁spenden -10.4651
+▁dazwischen -10.4655
+▁hinweis -10.4658
+▁nordamerika -10.4661
+krise -10.4671
+rlei -10.4691
+kraut -10.4732
+problem -10.4798
+hafen -10.4801
+bibliothek -10.4802
+milch -10.4803
+flossen -10.481
+▁sende -10.4812
+▁akku -10.4815
+bilität -10.4817
+š -10.4818
+▁achtzehn -10.4818
+▁ansässig -10.4818
+▁aufklär -10.4818
+▁engagement -10.4818
+▁interessant -10.4818
+▁smartphone -10.4818
+▁oktober -10.4818
+▁umfangreiche -10.4818
+▁ehrenamtlich -10.4818
+▁nahrungsmittel -10.4818
+▁wagner -10.4818
+▁draußen -10.4818
+▁wachstum -10.4818
+▁erlitt -10.4819
+räder -10.482
+▁wohnzimmer -10.4821
+▁wolken -10.4822
+▁emil -10.4823
+▁spezialisiert -10.4823
+▁puste -10.4823
+▁persönlichkeit -10.4824
+▁zwangs -10.4825
+▁setze -10.4835
+haufen -10.4835
+▁zack -10.484
+▁hinsicht -10.4846
+▁staatliche -10.4847
+▁dortige -10.486
+▁dieselbe -10.486
+▁bohr -10.4866
+aufnahmen -10.4881
+▁russe -10.4887
+▁chip -10.4888
+▁erbaut -10.4896
+▁schwä -10.49
+▁finanziert -10.4916
+watt -10.4943
+materie -10.4971
+▁gesetzlich -10.4976
+grill -10.4992
+▁begleit -10.5019
+▁bringe -10.5022
+kommando -10.5023
+erkrankung -10.5027
+▁betreiben -10.5027
+▁geschaffen -10.5027
+▁reduziert -10.5027
+▁völker -10.5027
+▁würz -10.5027
+▁zeichnet -10.5027
+▁zuverlässig -10.5027
+container -10.5027
+▁literarische -10.5027
+▁risiko -10.5027
+▁rudolf -10.5027
+▁gustav -10.5027
+▁bayerische -10.5027
+▁unterscheidet -10.5027
+▁rätsel -10.5027
+▁benutzen -10.5027
+▁dresden -10.5027
+▁verletzt -10.5027
+▁geburtstag -10.5027
+quartier -10.5027
+▁eigenschaften -10.5027
+▁verbrenn -10.5027
+periode -10.5028
+▁dreißigjährige -10.5028
+▁bildschirm -10.5028
+dynam -10.5028
+▁ahnung -10.5029
+▁ferien -10.5029
+▁vieh -10.503
+▁vorlesen -10.503
+▁sinnvoll -10.5031
+▁bleib -10.5035
+winkel -10.5038
+▁geschossen -10.5041
+▁streu -10.5043
+▁schraub -10.5044
+kolleg -10.5049
+▁jura -10.5058
+▁bereitet -10.5065
+▁mittelalterlich -10.5066
+▁wohin -10.5067
+▁schönste -10.5082
+▁säu -10.5087
+tafel -10.5092
+strafe -10.5135
+▁ruine -10.5152
+blitz -10.5215
+bedingung -10.5237
+▁satelliten -10.5239
+▁verstärkt -10.5239
+▁geduld -10.5239
+▁zürich -10.5239
+▁fundament -10.5239
+▁offensichtlich -10.5239
+▁begraben -10.5239
+▁insekten -10.5239
+schaffung -10.524
+▁knall -10.524
+▁astronom -10.524
+▁nordosten -10.524
+▁schicksal -10.524
+▁gedreht -10.524
+▁analog -10.524
+▁ausreichend -10.524
+uddel -10.524
+kuppel -10.5241
+▁besichtig -10.5242
+zettel -10.5242
+trennung -10.5243
+▁island -10.5244
+▁rückkehr -10.5244
+▁klingen -10.5248
+▁elektronische -10.5253
+▁freundschaft -10.5256
+▁partie -10.5257
+▁momentan -10.5262
+▁ausrede -10.5265
+▁haushalt -10.5271
+dreieck -10.5273
+▁adler -10.5274
+gehört -10.5278
+bricht -10.5279
+▁fürchte -10.5299
+▁epi -10.5345
+läufe -10.5349
+▁satt -10.5356
+verhältnisse -10.5362
+bereitschaft -10.5387
+schieben -10.5401
+▁automat -10.5409
+krone -10.5428
+zusetzen -10.5435
+reinigung -10.5441
+▁rind -10.5455
+▁getötet -10.5457
+▁luxemburg -10.5457
+▁nürnberg -10.5457
+▁quadrat -10.5457
+▁verursacht -10.5457
+▁firmen -10.5457
+▁besorgt -10.5457
+▁justus -10.5457
+▁betriebssystem -10.5457
+▁überlegt -10.5457
+▁spür -10.5458
+▁freut -10.5458
+konkret -10.5458
+▁opposition -10.5459
+▁inhaltlich -10.5461
+▁lohnt -10.5463
+▁bedingungen -10.5463
+▁hauptrolle -10.5464
+faktor -10.5465
+▁speisen -10.5466
+tonne -10.5466
+▁aufhören -10.5467
+▁mineral -10.5467
+▁gleichmäßig -10.5471
+▁niederlande -10.5475
+▁human -10.5477
+▁hinterher -10.5484
+▁großteil -10.5485
+▁sorte -10.5489
+fenster -10.5496
+taste -10.5499
+schutzgebiet -10.55
+schmutz -10.5507
+fühl -10.5535
+▁schaf -10.5538
+▁bunt -10.5556
+ündig -10.5575
+konzentration -10.5679
+verschlüsselung -10.5679
+▁beliebig -10.5679
+▁fängt -10.5679
+▁fünfzehn -10.5679
+▁gedächtnis -10.5679
+▁global -10.5679
+▁maßgeblich -10.5679
+▁verantwortung -10.5679
+▁verliebt -10.5679
+verfolgung -10.5679
+verständlich -10.5679
+▁begründet -10.5679
+▁behindert -10.5679
+▁stieß -10.5679
+▁erfüllt -10.5679
+▁garantiert -10.5679
+▁behaart -10.5679
+▁nachmittag -10.5679
+▁gegenwart -10.5679
+funk -10.568
+▁priester -10.568
+▁gesichert -10.568
+▁gekocht -10.568
+▁voraussetzung -10.5681
+▁geistes -10.5682
+verbände -10.5682
+▁drehbuch -10.5683
+bräu -10.5684
+▁bühnen -10.5685
+▁erhoben -10.5686
+greift -10.5689
+▁fell -10.5689
+▁baustelle -10.5695
+▁etliche -10.5697
+jenige -10.5702
+▁hessen -10.5705
+boot -10.5711
+▁mix -10.5716
+▁mega -10.5718
+füllen -10.5749
+manager -10.5793
+schub -10.5829
+▁verläss -10.5832
+▁entsprechen -10.5837
+▁nordwest -10.585
+stange -10.5858
+gezeichnet -10.5862
+gedrückt -10.5906
+▁aussieht -10.5906
+▁kalifornien -10.5906
+kompetenz -10.5906
+▁populär -10.5906
+beschreibung -10.5906
+▁reagieren -10.5906
+▁angesiedelt -10.5906
+▁gewebe -10.5906
+kommission -10.5907
+▁neugier -10.5907
+▁erhältlich -10.5907
+▁verwechselt -10.5907
+▁installiert -10.5907
+▁aachen -10.5908
+▁dutzend -10.5908
+▁herrscher -10.5909
+▁bedeckt -10.5909
+▁kirchliche -10.591
+▁kanon -10.591
+plastik -10.591
+tomaten -10.591
+▁naturpark -10.591
+▁verkünd -10.591
+▁german -10.5912
+üchtig -10.5912
+▁stress -10.5913
+▁hirn -10.5913
+kämpft -10.5913
+hochschule -10.5914
+▁heirat -10.5914
+▁flüsse -10.5914
+▁alex -10.5918
+▁schweigen -10.5921
+▁reformation -10.5922
+▁mündet -10.5924
+▁luc -10.593
+▁papa -10.5939
+spor -10.5941
+korps -10.5941
+fünfzig -10.5948
+▁ereignisse -10.5948
+▁langjährig -10.5981
+▁kongress -10.6006
+beitrag -10.6006
+▁führ -10.6011
+gasse -10.6029
+▁profil -10.6073
+ografie -10.609
+flüge -10.6123
+▁nett -10.6128
+▁unterlieg -10.6133
+▁synchron -10.6138
+▁entschuldige -10.6139
+▁erklärung -10.6139
+▁exemplar -10.6139
+▁joachim -10.6139
+▁lokomotiv -10.6139
+▁records -10.6139
+▁synonym -10.6139
+▁verzweifelt -10.6139
+bezieht -10.6139
+gesperrt -10.6139
+▁atlant -10.6139
+▁australien -10.6139
+▁geändert -10.6139
+▁fliegt -10.6139
+▁öfter -10.6139
+▁behauptet -10.6139
+▁erwartung -10.6139
+▁hobby -10.6139
+▁tempel -10.6139
+▁award -10.6139
+▁stolz -10.6139
+▁überlebt -10.6139
+▁royal -10.6139
+▁pfanne -10.6139
+▁nannte -10.614
+▁schnitz -10.614
+▁trennt -10.6141
+klinik -10.6141
+▁unterwegs -10.6142
+▁bestritt -10.6142
+▁fraktion -10.6142
+böden -10.6143
+▁bezahlt -10.6145
+▁informations -10.6146
+hänge -10.615
+sekretär -10.6151
+▁einheimische -10.6151
+frau -10.6152
+▁fließend -10.6154
+▁tja -10.616
+bezeichnung -10.617
+▁litt -10.6173
+strom -10.6175
+▁nordwestlich -10.6197
+▁fruchtbar -10.6236
+▁gräb -10.6257
+▁qualifiziert -10.6265
+damm -10.6267
+shop -10.6288
+röte -10.6299
+frist -10.6308
+▁amerika -10.6366
+gängig -10.6377
+konservativ -10.6377
+perspektive -10.6377
+▁integriert -10.6377
+▁reparatur -10.6377
+▁september -10.6377
+▁vielfältig -10.6377
+intelligenz -10.6377
+▁befestigt -10.6377
+▁großmutter -10.6377
+▁kontinent -10.6377
+▁reserve -10.6377
+▁sächsisch -10.6377
+▁begeistert -10.6377
+▁diskutiert -10.6377
+▁virtu -10.6377
+▁weimar -10.6377
+▁vorhaben -10.6377
+substanz -10.6377
+▁vergiss -10.6377
+▁disco -10.6378
+▁fortschritt -10.6379
+▁teller -10.6379
+▁flüssigkeit -10.638
+▁faust -10.638
+▁römer -10.6381
+▁hype -10.6382
+kapitel -10.6382
+▁geistliche -10.6382
+▁salzburg -10.6384
+▁manuell -10.6384
+▁andererseits -10.6385
+▁wolfgang -10.6385
+modus -10.6387
+korrektur -10.639
+kunst -10.6391
+▁liefert -10.6391
+▁schützt -10.6392
+wanderung -10.6394
+▁acker -10.6394
+säge -10.6394
+organisation -10.6396
+flugzeug -10.64
+▁junior -10.6401
+ausbildung -10.6402
+geräumt -10.641
+herrschaft -10.6415
+ersdorf -10.6472
+antwortet -10.6472
+objektiv -10.6473
+▁vergleichbar -10.6493
+streit -10.6501
+brille -10.6522
+▁santa -10.6548
+▁lernt -10.6573
+analys -10.6574
+zentrale -10.6576
+▁thi -10.6586
+regiment -10.6611
+rgebnis -10.6617
+syndrom -10.6621
+▁abkürzung -10.6621
+▁beobachten -10.6621
+▁dreizehn -10.6621
+▁gmbh -10.6621
+▁publikation -10.6621
+▁nachwuchs -10.6621
+▁beschreibt -10.6621
+▁patienten -10.6621
+▁teppich -10.6621
+▁ernährung -10.6621
+▁kommunal -10.6621
+▁kredit -10.6621
+▁denkmalschutz -10.6621
+▁josef -10.6621
+▁moderator -10.6621
+▁pianist -10.6621
+▁magnetfeld -10.6621
+▁übernehmen -10.6621
+zügig -10.6621
+mittwoch -10.6621
+▁bequem -10.6621
+sponsor -10.6621
+▁wahrnehm -10.6621
+▁school -10.6621
+▁nebenwirkungen -10.6621
+seriöse -10.6621
+gefordert -10.6621
+▁geweiht -10.6621
+▁ausstattung -10.6622
+▁rundfunk -10.6622
+▁klischee -10.6623
+pflaster -10.6623
+▁graz -10.6625
+▁erlangte -10.6626
+▁vermittelt -10.6627
+▁scheiben -10.6627
+▁eigentum -10.6629
+▁hauptsitz -10.6636
+wettbewerb -10.6645
+prinzip -10.6646
+schwein -10.6647
+viertel -10.6654
+▁schaffe -10.6668
+besuch -10.6689
+aufgaben -10.6729
+▁lehn -10.6735
+▁falt -10.6748
+▁kurve -10.6797
+▁südamerika -10.6838
+▁biologi -10.6847
+molekül -10.6871
+é -10.6871
+▁aspekt -10.6871
+▁gefördert -10.6871
+▁hardware -10.6871
+▁phänomen -10.6871
+▁pseudonym -10.6871
+▁definiert -10.6871
+▁effizient -10.6871
+▁fliehen -10.6871
+▁günstig -10.6871
+▁mecklenburg -10.6871
+▁symptom -10.6871
+träglich -10.6871
+▁betätigt -10.6871
+▁einzigartige -10.6871
+▁dramatisch -10.6871
+▁breslau -10.6871
+▁sonntag -10.6871
+▁deutet -10.6871
+▁harmlos -10.6871
+▁papst -10.6871
+▁jemals -10.6871
+▁durfte -10.6872
+▁ehefrau -10.6873
+▁porträt -10.6873
+▁übt -10.6874
+▁kleinkind -10.6875
+▁verwaltungssitz -10.6875
+▁bezahlen -10.6876
+▁anzeigen -10.688
+▁volksschule -10.688
+mischung -10.6885
+▁grundstück -10.6886
+wolke -10.6889
+cto -10.6934
+pfad -10.6972
+quote -10.6974
+gewandt -10.6981
+kapsel -10.6982
+erfolg -10.6989
+▁ortsname -10.6991
+wärme -10.7007
+▁betrachte -10.7041
+▁eishockey -10.7082
+▁nachricht -10.7105
+▁zoo -10.7106
+▁juni -10.7113
+▁empfohlen -10.7127
+▁facebook -10.7127
+▁hinterließ -10.7127
+▁niedersachsen -10.7127
+▁stellvertretend -10.7127
+▁verlängerung -10.7127
+▁zeitgenössische -10.7127
+▁kritisiert -10.7127
+▁skulptur -10.7127
+mittlung -10.7127
+▁großeltern -10.7127
+▁schräg -10.7127
+▁karlsruhe -10.7127
+▁verstopf -10.7127
+▁versorgt -10.7127
+▁zauber -10.7127
+▁monument -10.7127
+▁stockwerk -10.7127
+▁grafschaft -10.7128
+▁grundriss -10.7128
+▁etappe -10.7128
+▁verdanken -10.7129
+pulver -10.7129
+▁kolonie -10.7129
+▁antrag -10.7129
+▁bernhard -10.713
+▁high -10.713
+geflogen -10.713
+▁adolf -10.713
+▁gelöst -10.7131
+▁arbeitgeber -10.7133
+▁erfinde -10.7133
+▁bundesstaat -10.7134
+▁kult -10.7134
+▁circ -10.7134
+▁eiserne -10.7141
+▁spinne -10.7146
+▁ausgangs -10.7149
+▁liest -10.7149
+▁beraten -10.7156
+galerie -10.7188
+▁enttäusch -10.7196
+patron -10.7207
+work -10.7273
+▁photo -10.7332
+▁psycho -10.7356
+▁basketball -10.739
+▁demokratie -10.739
+▁ergänzt -10.739
+▁eventuell -10.739
+▁herkömmliche -10.739
+▁inspiriert -10.739
+▁jürgen -10.739
+▁konstruiert -10.739
+▁lübeck -10.739
+▁moskau -10.739
+▁november -10.739
+▁tschechisch -10.739
+▁tödlich -10.739
+▁verdächtig -10.739
+▁zentimeter -10.739
+▁verlängert -10.739
+▁verschlüsselt -10.739
+▁adelsgeschlecht -10.739
+▁alfred -10.739
+▁ventil -10.739
+▁verblieb -10.739
+▁bäcker -10.739
+▁ticket -10.739
+▁terror -10.7391
+▁radikal -10.7391
+▁vorbereitung -10.7391
+märkte -10.7391
+▁kooperation -10.7394
+▁herausforderung -10.7395
+▁gelungen -10.7395
+▁selbständig -10.7396
+▁übel -10.7396
+▁piep -10.7399
+▁dienstag -10.74
+▁finanzierung -10.7401
+▁städtische -10.7406
+tronom -10.7426
+▁touristisch -10.7427
+abhängig -10.7428
+kiste -10.7434
+entsprechend -10.7436
+konzert -10.7436
+bringt -10.7437
+erfahrung -10.7439
+beute -10.7441
+fusion -10.7442
+▁philosoph -10.7477
+liebe -10.7478
+birne -10.7518
+charts -10.7566
+asiatische -10.7588
+oskop -10.76
+distan -10.7609
+▁feucht -10.7627
+▁verhäng -10.7631
+▁verbreitung -10.7641
+▁sympathi -10.7661
+ärztin -10.7661
+▁ausgewählt -10.7661
+▁browser -10.7661
+▁diplomat -10.7661
+▁intelligent -10.7661
+▁konzentriert -10.7661
+▁majestät -10.7661
+▁medikament -10.7661
+▁rechtsanwalt -10.7661
+▁spaziergang -10.7661
+▁territor -10.7661
+▁vergnüg -10.7661
+ifikation -10.7661
+▁abseits -10.7661
+▁drohne -10.7661
+▁linux -10.7661
+▁gelehrt -10.7661
+▁rühr -10.7661
+▁akzent -10.7661
+▁spannend -10.7661
+▁verstecken -10.7661
+▁anbieten -10.7661
+▁bestattet -10.7661
+▁rechner -10.7661
+▁goethe -10.7661
+▁aktivitäten -10.7661
+benutzung -10.7661
+▁nervös -10.7661
+▁hirsch -10.7661
+▁mittelmeer -10.7661
+▁schmidt -10.7661
+▁schleim -10.7661
+▁boiler -10.7662
+management -10.7662
+▁schiller -10.7662
+▁erwarb -10.7662
+▁behaupten -10.7662
+▁abriss -10.7664
+▁rüber -10.7666
+▁passwort -10.7666
+stedt -10.7668
+pfalz -10.7668
+▁janni -10.7682
+aufenthalt -10.7691
+sozialistische -10.7698
+könig -10.7711
+vertreter -10.7713
+saison -10.7714
+artikel -10.7715
+messer -10.7716
+töne -10.772
+ritten -10.7732
+▁irgend -10.7743
+tournee -10.7752
+busse -10.7759
+smitglied -10.7767
+▁antoni -10.7771
+sparen -10.7772
+charakteristisch -10.7779
+schießen -10.7792
+wanderweg -10.7808
+gräber -10.7813
+bremse -10.7815
+national -10.782
+zunehmen -10.7925
+dynastie -10.7938
+getränke -10.7938
+verkürz -10.7938
+▁beinhaltet -10.7938
+▁bemerkenswert -10.7938
+▁dimension -10.7938
+▁dänemark -10.7938
+▁eigentümer -10.7938
+▁elisabeth -10.7938
+▁gültig -10.7938
+▁konsequent -10.7938
+▁nirgend -10.7938
+▁potsdam -10.7938
+▁spaghetti -10.7938
+▁verstorben -10.7938
+kombination -10.7938
+▁spontan -10.7938
+▁world -10.7938
+▁väter -10.7938
+▁sparkasse -10.7938
+▁kanada -10.7938
+▁trainiert -10.7938
+▁vorbereitet -10.7939
+▁reihenfolge -10.7939
+▁betreut -10.7939
+▁aufgewachsen -10.7939
+lädt -10.7939
+▁chemie -10.7939
+▁konstant -10.7939
+▁annehmen -10.7939
+▁multi -10.7939
+▁klingel -10.7939
+geliefert -10.7939
+seminar -10.7939
+▁labor -10.7939
+▁nadel -10.7939
+badenwürttemberg -10.794
+trommel -10.794
+▁daniel -10.794
+▁knack -10.794
+▁redaktion -10.794
+▁umständen -10.7941
+▁aqua -10.7942
+▁sachgebiet -10.7943
+▁kleider -10.7944
+▁gedruckt -10.7944
+▁überschreite -10.7945
+▁james -10.7952
+flieger -10.797
+▁pirate -10.7981
+soeben -10.799
+▁könne -10.8006
+plattform -10.8067
+kontrolle -10.809
+▁eiche -10.817
+entwurf -10.8199
+▁aufmerksam -10.8221
+▁flüssig -10.8221
+schränkung -10.8224
+zählig -10.8224
+▁dörfer -10.8224
+▁fantasie -10.8224
+▁februar -10.8224
+▁infrastruktur -10.8224
+▁sowjetunion -10.8224
+▁tshirt -10.8224
+▁zwiebeln -10.8224
+▁google -10.8224
+▁gänz -10.8224
+▁heidelberg -10.8224
+▁herausragend -10.8224
+▁qualifikation -10.8224
+▁eduard -10.8224
+▁pizza -10.8224
+▁betrug -10.8224
+▁hoffnung -10.8224
+▁justiz -10.8224
+schwangerschaft -10.8224
+▁empfäng -10.8224
+▁konsequenzen -10.8224
+▁zitrone -10.8224
+▁linear -10.8224
+▁lorenz -10.8224
+▁verstarb -10.8224
+vertretung -10.8225
+▁vorsichtig -10.8225
+▁beleuchtet -10.8225
+hydro -10.8225
+▁heftig -10.8225
+▁klappt -10.8225
+▁humor -10.8225
+▁unsichtbar -10.8226
+▁würfel -10.8226
+▁gezielt -10.8227
+▁draht -10.8228
+päck -10.8229
+▁witzig -10.823
+▁erzeugt -10.824
+gewinde -10.8293
+krankheit -10.8293
+instrument -10.8293
+fünf -10.8295
+gesagt -10.8295
+▁elektrisch -10.8332
+oberseite -10.8369
+▁fließ -10.8384
+rahm -10.8388
+▁tasse -10.844
+erfreu -10.8506
+fänglich -10.8518
+sehenswürdigkeit -10.8518
+▁entstehung -10.8518
+fernbedienung -10.8518
+färbung -10.8518
+schädigung -10.8518
+▁akustisch -10.8518
+▁brasilien -10.8518
+▁ferdinand -10.8518
+▁gewürz -10.8518
+▁league -10.8518
+▁nikolaus -10.8518
+▁schokolade -10.8518
+▁südafrika -10.8518
+▁verbessert -10.8518
+▁vernünftig -10.8518
+▁konvention -10.8518
+▁niveau -10.8518
+▁scheitern -10.8518
+▁rechtschreib -10.8518
+▁erfordern -10.8518
+▁margaret -10.8518
+▁identisch -10.8518
+▁waschmaschine -10.8518
+▁buchstaben -10.8518
+▁advent -10.8518
+▁dauerhaft -10.8518
+▁fernsehserie -10.8518
+▁gegessen -10.8518
+▁verhaftet -10.8519
+▁peinlich -10.8519
+▁geheimnis -10.8519
+▁pilze -10.8519
+▁wackel -10.8519
+polizist -10.8519
+▁ruhig -10.8519
+▁geschnitten -10.852
+▁grafik -10.852
+knecht -10.852
+▁bewohnt -10.8522
+pumpe -10.8524
+mängel -10.8526
+▁vortrag -10.8528
+▁atem -10.8532
+▁vertrat -10.8533
+▁kürzeste -10.8538
+wurst -10.8553
+regnet -10.8556
+politische -10.8597
+wood -10.8604
+angriff -10.8743
+▁mächtig -10.8746
+zähler -10.8751
+wohngemeinschaft -10.8772
+ärchen -10.8807
+▁kicke -10.8812
+sprachig -10.882
+synthese -10.8821
+ömischkatholisch -10.8821
+▁amazon -10.8821
+▁automobil -10.8821
+▁mathematisch -10.8821
+▁orientierung -10.8821
+▁ungeklärt -10.8821
+▁ungewöhnlich -10.8821
+▁unterschrieb -10.8821
+▁zukünftig -10.8821
+▁eignet -10.8821
+▁schallplatten -10.8821
+▁anerkennen -10.8821
+▁reparier -10.8821
+▁belgien -10.8821
+▁pilger -10.8821
+▁teilnehmen -10.8821
+▁freuen -10.8821
+▁bewahre -10.8821
+▁verzehr -10.8821
+▁kakao -10.8821
+▁white -10.8821
+▁märz -10.8822
+▁nachfolgend -10.8822
+▁unterstützen -10.8822
+▁herbst -10.8822
+▁motorräder -10.8823
+▁angler -10.8827
+▁kriminal -10.8828
+verarbeitung -10.8833
+breitungsgebiet -10.8844
+▁studiert -10.8869
+amerikanische -10.8875
+dokumentation -10.8899
+bestimmt -10.891
+huhn -10.8917
+▁gefolg -10.8928
+trink -10.893
+▁teure -10.8976
+berater -10.8987
+diskret -10.9011
+geholt -10.902
+kragen -10.9105
+ensemble -10.9134
+rdrheinwestfalen -10.9134
+▁akademische -10.9134
+▁begegnung -10.9134
+▁entfernung -10.9134
+▁erledigt -10.9134
+▁erleichtert -10.9134
+▁fakultät -10.9134
+▁griechenland -10.9134
+▁kolonial -10.9134
+▁kontinuierlich -10.9134
+▁nordöstlich -10.9134
+▁vegetarier -10.9134
+▁verbleib -10.9134
+▁mozilla -10.9134
+▁abenteuer -10.9134
+▁permanent -10.9134
+▁stängel -10.9134
+▁getreide -10.9134
+▁kurzfristig -10.9134
+▁website -10.9134
+▁erweiterung -10.9134
+▁genügt -10.9134
+▁gelöscht -10.9134
+▁mahlzeit -10.9134
+matrix -10.9134
+wählen -10.9134
+▁liberal -10.9134
+▁schnabel -10.9134
+▁veranstaltet -10.9134
+einflüsse -10.9134
+▁zubereitet -10.9134
+▁letztendlich -10.9134
+▁germanistik -10.9134
+▁louis -10.9134
+▁generell -10.9135
+▁beurteil -10.9139
+treffer -10.9139
+▁erdbeer -10.9141
+▁asien -10.9142
+geschichtliche -10.9148
+▁schmied -10.9149
+wurz -10.915
+▁mobil -10.9154
+▁archäologische -10.9156
+▁schwinge -10.9157
+▁eingriff -10.9158
+zange -10.916
+▁stilistisch -10.9177
+kartoffeln -10.9177
+freude -10.9231
+güter -10.9233
+▁lüge -10.9234
+brunn -10.9266
+ausweis -10.9392
+▁kartoffel -10.9412
+ropf -10.9421
+▁röntgen -10.9439
+▁illustr -10.9453
+▁zigaretten -10.9456
+▁grammatik -10.9456
+▁attraktiv -10.9456
+▁bayreuth -10.9456
+▁beherbergt -10.9456
+▁bewerbung -10.9456
+▁bildhauer -10.9456
+▁erwünscht -10.9456
+▁gewidmet -10.9456
+▁ländliche -10.9456
+▁wikipedia -10.9456
+▁wünscht -10.9456
+▁zeigefinger -10.9456
+erhöhung -10.9456
+ometrische -10.9456
+▁verabschied -10.9456
+▁amateur -10.9456
+▁oldenburg -10.9456
+▁united -10.9456
+▁rohstoffe -10.9456
+▁appetit -10.9456
+▁kläger -10.9456
+▁börse -10.9456
+▁erwischt -10.9456
+▁unschuld -10.9456
+▁quasi -10.9456
+▁zeitweilig -10.9456
+▁ämter -10.9456
+▁adresse -10.9456
+▁major -10.9457
+▁uboote -10.9457
+▁violin -10.9457
+▁personalausweis -10.9457
+▁jetzige -10.9457
+▁grafikkarte -10.9457
+▁verbiete -10.9458
+▁verdammt -10.9458
+▁leiser -10.9459
+▁kälte -10.9461
+▁mieter -10.9463
+▁mutig -10.9476
+gebrannt -10.9476
+▁gemein -10.951
+lässlich -10.9564
+bedarf -10.9564
+künstler -10.9565
+vorlesung -10.9569
+▁geschützt -10.9574
+sammeln -10.9601
+schaltung -10.9633
+▁begab -10.9654
+▁muttersprach -10.9667
+▁anlass -10.9721
+schwimmer -10.9723
+bänder -10.9788
+▁aggressiv -10.979
+▁argentinien -10.979
+▁betreibt -10.979
+▁emotionale -10.979
+▁entworfen -10.979
+▁ergänzen -10.979
+▁evolution -10.979
+▁günther -10.979
+▁joghurt -10.979
+▁kathedrale -10.979
+▁kraftausdrücke -10.979
+▁offenkundig -10.979
+▁sebastian -10.979
+▁zwilling -10.979
+▁alphabet -10.979
+▁differenz -10.979
+▁verwaltet -10.979
+▁angeklagt -10.979
+horizont -10.979
+▁löcher -10.979
+▁review -10.979
+▁verbinden -10.979
+▁fulda -10.979
+▁kabinett -10.979
+▁überregional -10.979
+▁osmanische -10.979
+▁wiederaufbau -10.979
+▁ozean -10.979
+entwickler -10.979
+▁gewerkschaft -10.979
+▁umzugehen -10.979
+▁abgedeckt -10.979
+▁abgeschnitten -10.979
+▁folter -10.979
+▁windows -10.979
+▁tschüss -10.979
+▁smith -10.979
+▁dominiert -10.979
+▁erreichbar -10.979
+▁ungerecht -10.979
+▁moralisch -10.979
+▁servier -10.979
+▁samstag -10.9791
+▁systematisch -10.9791
+▁zitter -10.9791
+▁behälter -10.9791
+▁marmor -10.9791
+▁heinz -10.9792
+▁half -10.9793
+▁neutral -10.9793
+▁besiegt -10.9794
+▁falk -10.9795
+▁fahne -10.9805
+sterblich -10.9807
+▁träume -10.9809
+▁lackier -10.9811
+fuchs -10.9813
+▁neffe -10.9815
+eintrag -10.9823
+sozialisten -10.9845
+▁spitzname -10.9856
+gestrichen -10.9876
+korb -10.9878
+kritik -10.9907
+größte -10.9911
+gedacht -10.9933
+▁kippe -10.9967
+▁hexe -10.9992
+▁infanterie -11.0119
+württemberg -11.0133
+dimensional -11.0135
+explosion -11.0135
+rhythmus -11.0135
+▁donnerstag -11.0135
+▁entspannt -11.0135
+▁fütter -11.0135
+▁greifswald -11.0135
+▁herauszufinden -11.0135
+▁hollywood -11.0135
+▁individuell -11.0135
+▁relevant -11.0135
+▁sechzehn -11.0135
+▁verbesserung -11.0135
+▁verbindet -11.0135
+▁wohlhabend -11.0135
+▁aufsichtsrat -11.0135
+▁mexiko -11.0135
+▁portugal -11.0135
+westfälische -11.0135
+▁klemmt -11.0135
+▁beruht -11.0135
+▁metropol -11.0135
+▁brauchst -11.0135
+▁wiesbaden -11.0135
+▁gemischt -11.0135
+▁bananen -11.0135
+▁produktiv -11.0135
+▁klammern -11.0135
+▁töten -11.0135
+pfeffer -11.0135
+▁lastwagen -11.0135
+▁schwäbisch -11.0135
+▁tirol -11.0135
+▁verarbeitet -11.0135
+▁gehabt -11.0135
+▁tennis -11.0135
+▁folk -11.0135
+▁geschickt -11.0135
+gläser -11.0135
+filiale -11.0135
+▁bedienen -11.0135
+käufe -11.0137
+▁vereinbart -11.0137
+▁dasselbe -11.0137
+▁abstrakt -11.0137
+wirksam -11.0146
+▁verpasst -11.0147
+rutscht -11.0148
+nehmlich -11.0152
+▁eckige -11.0163
+feuchtigkeit -11.0178
+▁gebiss -11.0201
+trägt -11.0241
+sänger -11.0267
+▁stroh -11.0368
+▁ereignis -11.0426
+zitat -11.0456
+konstrukt -11.0489
+▁genoss -11.0491
+entzündung -11.0492
+gekühlt -11.0492
+zyklus -11.0492
+č -11.0492
+▁aluminium -11.0492
+▁argument -11.0492
+▁arrondissement -11.0492
+▁badewanne -11.0492
+▁befördert -11.0492
+▁betreuung -11.0492
+▁columbia -11.0492
+▁dezember -11.0492
+▁erkrankt -11.0492
+▁gestorben -11.0492
+▁ignorieren -11.0492
+▁publiziert -11.0492
+▁resultat -11.0492
+▁science -11.0492
+▁skript -11.0492
+▁sozusagen -11.0492
+▁staatsanwalt -11.0492
+▁webbrowser -11.0492
+▁ägyptische -11.0492
+effizienz -11.0492
+▁begrüß -11.0492
+▁beruhig -11.0492
+▁hockey -11.0492
+▁kombiniert -11.0492
+▁kommissar -11.0492
+▁schwör -11.0492
+▁unzufrieden -11.0492
+▁verzeih -11.0492
+kampagne -11.0492
+▁senkrecht -11.0492
+▁kastanie -11.0492
+▁lyrik -11.0492
+▁forstwirtschaft -11.0492
+▁herdplatte -11.0492
+▁erdgas -11.0492
+▁preußen -11.0492
+▁verringert -11.0492
+▁group -11.0492
+▁tendenz -11.0492
+▁vereinzelt -11.0492
+▁praktik -11.0492
+glühen -11.0492
+sphäre -11.0492
+▁jegliche -11.0492
+▁terrasse -11.0492
+stiefel -11.0493
+▁vermehrt -11.0493
+▁wieviel -11.0494
+▁unternahm -11.0496
+cross -11.0496
+▁löt -11.0498
+▁ingwer -11.05
+billy -11.0514
+▁schäf -11.0515
+unterhalt -11.0518
+haube -11.052
+▁nachlass -11.0521
+pedal -11.0553
+aufzeichnungen -11.0592
+▁klär -11.0595
+spezialisten -11.0599
+angebot -11.0632
+frucht -11.0637
+spiegel -11.0637
+interpretiert -11.0718
+nathan -11.0744
+▁gymnas -11.0795
+▁sekret -11.0843
+▁beobachtung -11.086
+katastrophe -11.0862
+▁beheimatet -11.0862
+▁chicago -11.0862
+▁eröffnen -11.0862
+▁geschafft -11.0862
+▁gewitter -11.0862
+▁gewusst -11.0862
+▁hubschrauber -11.0862
+▁löffel -11.0862
+▁münchner -11.0862
+▁umgewandelt -11.0862
+▁verteidiger -11.0862
+▁wähler -11.0862
+▁youtube -11.0862
+▁ägypten -11.0862
+chleswigholstein -11.0862
+▁akkorde -11.0862
+▁befestigung -11.0862
+▁elegant -11.0862
+▁erfindung -11.0862
+▁illegal -11.0862
+▁befindlich -11.0862
+▁vanille -11.0862
+▁bachelor -11.0862
+▁prominente -11.0862
+kompetent -11.0862
+▁fußgänger -11.0862
+▁minderheit -11.0862
+genügend -11.0862
+▁konstruktiv -11.0862
+▁existenz -11.0862
+▁unterhielt -11.0862
+▁schimmel -11.0862
+schmuck -11.0862
+▁blonde -11.0862
+▁moped -11.0862
+▁verschwunden -11.0862
+▁unheimlich -11.0862
+▁touch -11.0862
+▁formuliert -11.0862
+▁mädel -11.0862
+▁wofür -11.0862
+▁simson -11.0862
+▁zensur -11.0862
+gestimmt -11.0862
+▁angrenzende -11.0862
+▁franzosen -11.0862
+matratze -11.0863
+▁heroin -11.0863
+▁creme -11.0863
+bruck -11.0863
+schöpft -11.0864
+▁stachel -11.0864
+▁charlie -11.0865
+▁vorbehalt -11.0867
+▁vermiss -11.087
+▁einheitliche -11.088
+inhaber -11.0889
+▁späteste -11.0905
+▁exakt -11.0941
+platin -11.0985
+schwester -11.1022
+▁betreff -11.1033
+motiv -11.1152
+ł -11.1247
+schütteln -11.1247
+spektrum -11.1247
+sprünge -11.1247
+▁abgeschafft -11.1247
+▁ausgedehnte -11.1247
+▁befürchte -11.1247
+▁begeisterung -11.1247
+▁bielefeld -11.1247
+▁braunschweig -11.1247
+▁gemütlich -11.1247
+▁gespalten -11.1247
+▁korruption -11.1247
+▁kritisiere -11.1247
+▁oxford -11.1247
+▁scheinwerfer -11.1247
+▁ungeduldig -11.1247
+dämpfer -11.1247
+▁angreifer -11.1247
+▁aufwändig -11.1247
+▁kompromiss -11.1247
+▁konfession -11.1247
+▁mechanisch -11.1247
+▁skandinavi -11.1247
+▁unregelmäßig -11.1247
+▁feierabend -11.1247
+▁passagier -11.1247
+teleskop -11.1247
+▁mozart -11.1247
+▁nominiert -11.1247
+▁unternehmer -11.1247
+mechanik -11.1247
+▁sandwich -11.1247
+▁statistik -11.1247
+▁türkei -11.1247
+▁eklig -11.1247
+▁behielt -11.1247
+▁fleißig -11.1247
+jörn -11.1247
+▁installation -11.1247
+zeichner -11.1247
+▁klassik -11.1247
+▁trikot -11.1247
+▁bewaldet -11.1247
+▁hübsch -11.1247
+▁arnold -11.1247
+▁gemarkung -11.1247
+▁irritier -11.1247
+▁norwegische -11.1247
+▁deaktivier -11.1247
+▁dozent -11.1247
+▁weißwein -11.1248
+▁floh -11.1248
+turbine -11.1248
+klopft -11.1248
+▁earl -11.1248
+▁ungarn -11.1248
+koppel -11.1248
+vektor -11.1248
+zentren -11.1248
+▁tapete -11.1248
+▁längs -11.1249
+schleife -11.125
+plateau -11.125
+▁jacob -11.1251
+gerichtshof -11.1251
+futter -11.1252
+dünger -11.1256
+zwischenzeitlich -11.1259
+▁bahnsteig -11.1267
+▁nudel -11.1302
+kolben -11.1401
+albrecht -11.1411
+größe -11.142
+polizei -11.1422
+turnier -11.1423
+video -11.1423
+kritische -11.1486
+▁kindergärt -11.1495
+zwang -11.1634
+▁krimi -11.1643
+gerührt -11.1647
+orthodoxe -11.1647
+▁außergewöhnlich -11.1647
+▁bedeutsam -11.1647
+▁beerdigung -11.1647
+▁fälschlich -11.1647
+▁gewünscht -11.1647
+▁kennengelernt -11.1647
+▁klimawandel -11.1647
+▁kräuter -11.1647
+▁merkwürdig -11.1647
+▁mitbewohner -11.1647
+▁pazifik -11.1647
+▁produzieren -11.1647
+▁qualitativ -11.1647
+▁versöhn -11.1647
+▁zurückgekehrt -11.1647
+dürftig -11.1647
+▁kardinal -11.1647
+▁kümmert -11.1647
+▁prototyp -11.1647
+▁technologie -11.1647
+▁vergrößert -11.1647
+▁verschollen -11.1647
+gürtel -11.1647
+reflex -11.1647
+▁theodor -11.1647
+▁verrückt -11.1647
+▁neuseeland -11.1647
+▁abgebildet -11.1647
+▁kündigung -11.1647
+▁resonanz -11.1647
+▁seltsam -11.1647
+▁füße -11.1647
+▁kanzler -11.1647
+▁bräuchte -11.1647
+▁demonstra -11.1647
+express -11.1647
+▁droht -11.1647
+▁henrik -11.1647
+▁semester -11.1647
+prognose -11.1647
+▁dänische -11.1647
+▁krypt -11.1647
+▁variiert -11.1647
+▁entdeckung -11.1647
+▁helfer -11.1647
+▁clown -11.1647
+▁manfred -11.1647
+▁spanier -11.1647
+▁fokus -11.1647
+▁kannibal -11.1647
+▁überarbeitet -11.1647
+▁alexandr -11.1647
+klopfen -11.1648
+▁verkäufer -11.1648
+▁chlor -11.1648
+▁zwingen -11.1648
+▁verpflichtung -11.1648
+gemessen -11.1651
+▁hessische -11.1652
+▁entsprach -11.1652
+brigade -11.1653
+frachter -11.1656
+zügl -11.1656
+muschel -11.1662
+screen -11.1669
+▁kleist -11.1675
+schmerzen -11.1682
+pädagogik -11.1709
+affin -11.1748
+professor -11.1839
+▁drübe -11.1891
+beschäftigung -11.2063
+feedback -11.2063
+schränke -11.2063
+undsechzig -11.2063
+▁angestrebt -11.2063
+▁berücksichtigt -11.2063
+▁beschäftigen -11.2063
+▁diagnose -11.2063
+▁empirisch -11.2063
+▁gebräuchlich -11.2063
+▁gitarrist -11.2063
+▁jesuiten -11.2063
+▁kofferraum -11.2063
+▁mercedes -11.2063
+▁promoviert -11.2063
+▁präsentation -11.2063
+▁repertoire -11.2063
+▁schildkröte -11.2063
+▁schottland -11.2063
+▁thailand -11.2063
+▁vernunft -11.2063
+▁versprochen -11.2063
+▁vorübergehend -11.2063
+fränkisch -11.2063
+gebühren -11.2063
+vorpommern -11.2063
+▁ausleihe -11.2063
+▁erfährt -11.2063
+▁francisco -11.2063
+▁kurfürst -11.2063
+▁unterirdisch -11.2063
+▁zeremonie -11.2063
+geschliffen -11.2063
+▁hähnchen -11.2063
+▁skandal -11.2063
+▁unfair -11.2063
+▁barfuß -11.2063
+▁grüße -11.2063
+solvent -11.2063
+▁brown -11.2063
+▁wirkungsgrad -11.2063
+▁kanadische -11.2063
+▁frösche -11.2063
+▁irrtum -11.2063
+▁unterbrochen -11.2063
+membran -11.2063
+▁geklaut -11.2063
+▁beidseitig -11.2064
+pistole -11.2064
+schleuder -11.2064
+▁spieß -11.2064
+föhn -11.2064
+▁witwe -11.2064
+▁brumm -11.2064
+▁tugend -11.2064
+reaktor -11.2064
+▁zwerg -11.2064
+▁diktier -11.2067
+gelobt -11.2069
+trüg -11.2083
+wucher -11.2216
+medizin -11.2273
+bundesliga -11.2273
+konzept -11.2273
+historische -11.2273
+schwarz -11.2274
+▁hekti -11.2284
+sequenz -11.2498
+ı -11.2498
+▁wahnsinn -11.2498
+ohlenstoffdioxid -11.2498
+▁abstammung -11.2498
+▁autodidakt -11.2498
+▁bestätigen -11.2498
+▁chefredakteur -11.2498
+▁defensiv -11.2498
+▁dolmetscher -11.2498
+▁entlehnt -11.2498
+▁erforscht -11.2498
+▁europäer -11.2498
+▁football -11.2498
+▁gärtner -11.2498
+▁individuum -11.2498
+▁integration -11.2498
+▁koalition -11.2498
+▁kohlenhydrat -11.2498
+▁lächeln -11.2498
+▁mythologie -11.2498
+▁schädel -11.2498
+▁staatsexamen -11.2498
+▁update -11.2498
+▁venedig -11.2498
+▁victoria -11.2498
+gewölbe -11.2498
+▁ausdrücke -11.2498
+▁betrachtung -11.2498
+▁bewaffnet -11.2498
+▁mexikanisch -11.2498
+▁pinguin -11.2498
+▁trainieren -11.2498
+▁vietnam -11.2498
+▁westfalen -11.2498
+▁zeitgenosse -11.2498
+wägung -11.2498
+▁leopold -11.2498
+▁straßburg -11.2498
+▁stürzen -11.2498
+▁darauffolgende -11.2498
+▁oberschenkel -11.2498
+▁separat -11.2498
+ophyt -11.2498
+▁gewaschen -11.2498
+▁primär -11.2498
+▁samuel -11.2498
+gletscher -11.2498
+▁episode -11.2498
+gerasselt -11.2498
+▁frosch -11.2499
+▁angepasst -11.2499
+geschoben -11.2499
+▁flora -11.2499
+pfeiler -11.2499
+▁inszeniert -11.2499
+quallen -11.2499
+fizierung -11.2499
+strumpf -11.2499
+▁alliierte -11.2499
+▁beschwerde -11.2501
+▁morgendliche -11.2502
+drüse -11.2524
+binär -11.2547
+schnecke -11.2571
+▁skifahre -11.2581
+▁hinterfrag -11.2727
+kamera -11.273
+▁vorführ -11.2741
+▁charakterisier -11.2915
+▁archäologi -11.292
+denkmäler -11.2953
+diagnosti -11.2953
+kapazität -11.2953
+theoretiker -11.2953
+ungsgefahr -11.2953
+▁abgetrennt -11.2953
+▁amsterdam -11.2953
+▁ausgiebig -11.2953
+▁däumchen -11.2953
+▁erzbistum -11.2953
+▁garderobe -11.2953
+▁gescheitert -11.2953
+▁gestohlen -11.2953
+▁gläubig -11.2953
+▁höflich -11.2953
+▁katharina -11.2953
+▁koblenz -11.2953
+▁kommerziell -11.2953
+▁konzentrieren -11.2953
+▁korrigieren -11.2953
+▁madrid -11.2953
+▁nächstgelegene -11.2953
+▁plätzchen -11.2953
+▁sigmund -11.2953
+▁synagoge -11.2953
+▁südkorea -11.2953
+▁unternommen -11.2953
+▁verfasser -11.2953
+▁vernachlässigen -11.2953
+▁zerstörung -11.2953
+▁elternbeirat -11.2953
+▁oscar -11.2953
+▁valentin -11.2953
+▁phantom -11.2953
+empfunden -11.2953
+▁tötet -11.2953
+unterdrück -11.2953
+▁charlotte -11.2953
+wächter -11.2953
+▁magdeburg -11.2953
+▁brennessel -11.2953
+▁nicole -11.2953
+▁australische -11.2953
+wasseroberfläche -11.2953
+▁tränen -11.2953
+▁vermarktet -11.2953
+▁oskar -11.2953
+ritual -11.2953
+▁mikrowelle -11.2953
+ergüsse -11.2953
+abnehmer -11.2953
+▁sabine -11.2953
+▁drückt -11.2953
+scanner -11.2954
+▁morgenstund -11.2954
+briefmarken -11.2955
+plikat -11.296
+▁zünd -11.2962
+kabine -11.3005
+schwimmen -11.3121
+leuchte -11.3147
+gitarre -11.3203
+bäume -11.3204
+▁isländ -11.3304
+▁jüdi -11.3365
+▁differential -11.3429
+▁rendsburg -11.3429
+▁empör -11.3429
+flüchtlinge -11.3429
+weiterentwickelt -11.3429
+▁akkuschrauber -11.3429
+▁anspruchsvoll -11.3429
+▁apotheker -11.3429
+▁ausflugsziel -11.3429
+▁bedürfnis -11.3429
+▁boykott -11.3429
+▁bündnis -11.3429
+▁entspring -11.3429
+▁erschrocken -11.3429
+▁gebirgs -11.3429
+▁gespeichert -11.3429
+▁gleichgültig -11.3429
+▁hiphop -11.3429
+▁jerusalem -11.3429
+▁jörg -11.3429
+▁karosserie -11.3429
+▁knopf -11.3429
+▁konzipiert -11.3429
+▁leichtathletik -11.3429
+▁lügner -11.3429
+▁napoleon -11.3429
+▁rucksack -11.3429
+▁schröder -11.3429
+▁thüringen -11.3429
+▁verlustbehaftet -11.3429
+▁vermähl -11.3429
+▁volumen -11.3429
+▁werbespot -11.3429
+▁whatsappgruppe -11.3429
+donaudampfschiff -11.3429
+getroffen -11.3429
+pünktlich -11.3429
+saxophon -11.3429
+▁freimaurer -11.3429
+▁physikalische -11.3429
+▁volksarmee -11.3429
+▁übersetzer -11.3429
+▁erstaunlich -11.3429
+▁gestört -11.3429
+▁köstlich -11.3429
+▁methodi -11.3429
+▁nicola -11.3429
+gehängt -11.3429
+▁eberhard -11.3429
+▁fiktive -11.3429
+krümel -11.3429
+▁apostel -11.3429
+gewohnheiten -11.3429
+▁torhüter -11.3429
+▁monitor -11.3429
+▁arrangiert -11.3429
+▁ungesund -11.3429
+▁satire -11.3429
+unterkunft -11.3429
+krabbelt -11.3429
+▁gebüsch -11.3429
+▁befreundet -11.3429
+▁kündigte -11.3429
+bezüglich -11.3429
+rauszuholen -11.3429
+marathon -11.3429
+▁linguisti -11.3429
+schwankungen -11.3429
+▁epoche -11.343
+▁kulisse -11.343
+finsternis -11.3431
+▁pflück -11.3431
+gestreift -11.3432
+einsatzkommando -11.3442
+bewusstsein -11.3448
+direktion -11.3514
+genauigkeit -11.3604
+höpf -11.3676
+kräftig -11.3703
+jugend -11.3704
+europäische -11.3841
+▁beeinträchtig -11.3929
+▁panorama -11.3929
+▁verwechsl -11.3929
+observatorium -11.3929
+undvierzig -11.3929
+ř -11.3929
+▁abkühlen -11.3929
+▁anfällig -11.3929
+▁angekündigt -11.3929
+▁ausführlich -11.3929
+▁außerordentlich -11.3929
+▁begegnet -11.3929
+▁breisgau -11.3929
+▁dissertation -11.3929
+▁ermordet -11.3929
+▁fraktur -11.3929
+▁gefällig -11.3929
+▁genehmigt -11.3929
+▁indirekt -11.3929
+▁kontrahent -11.3929
+▁kontrovers -11.3929
+▁matthias -11.3929
+▁mittagspause -11.3929
+▁orientalische -11.3929
+▁premierminister -11.3929
+▁propaganda -11.3929
+▁renommierte -11.3929
+▁subventionier -11.3929
+▁thronfolger -11.3929
+▁ukraine -11.3929
+▁variable -11.3929
+▁vernachlässigt -11.3929
+▁zerstörer -11.3929
+kanzlei -11.3929
+▁asphalt -11.3929
+▁bestraft -11.3929
+▁bewährung -11.3929
+▁erledigen -11.3929
+▁grammatikalisch -11.3929
+▁kollektiv -11.3929
+▁krokodil -11.3929
+▁sachverhalt -11.3929
+▁spekulation -11.3929
+▁spionage -11.3929
+▁edward -11.3929
+▁sklaven -11.3929
+▁texas -11.3929
+▁woanders -11.3929
+▁istanbul -11.3929
+▁variieren -11.3929
+▁gregor -11.3929
+▁empor -11.3929
+▁künste -11.3929
+▁allianz -11.3929
+▁schottische -11.3929
+autoritäre -11.3929
+▁papagei -11.3929
+▁präzise -11.3929
+murmel -11.3929
+▁entnehme -11.3929
+monarchie -11.393
+gefroren -11.393
+▁hülse -11.3931
+▁vegane -11.3931
+olivia -11.3931
+späne -11.3933
+marmelade -11.3933
+rumpelt -11.3933
+offensive -11.3935
+▁aufflieg -11.3938
+überwacht -11.3946
+göttin -11.3948
+supermarkt -11.4021
+lutherische -11.4058
+farbig -11.4098
+▁vollzieh -11.4102
+begabung -11.4141
+zwanzig -11.4227
+software -11.423
+literatur -11.423
+genossenschaft -11.4448
+erscheinungsbild -11.4455
+zugehörigkeit -11.4455
+ā -11.4455
+▁afghanistan -11.4455
+▁beatrix -11.4455
+▁begegnen -11.4455
+▁beschuldig -11.4455
+▁betäubung -11.4455
+▁dokumentarfilm -11.4455
+▁einkaufswagen -11.4455
+▁elektrifizier -11.4455
+▁entnommen -11.4455
+▁explizit -11.4455
+▁fortpflanzung -11.4455
+▁fräulein -11.4455
+▁fröhlich -11.4455
+▁genehmigung -11.4455
+▁gerüchte -11.4455
+▁gesammelt -11.4455
+▁heilbronn -11.4455
+▁hervorragende -11.4455
+▁identifizier -11.4455
+▁katholiken -11.4455
+▁kaufmännische -11.4455
+▁ladungsträger -11.4455
+▁laptop -11.4455
+▁leichnam -11.4455
+▁protagonist -11.4455
+▁registriert -11.4455
+▁rekonstruier -11.4455
+▁resultierende -11.4455
+▁schizophren -11.4455
+ã -11.5599
+ô -11.5599
+â -11.5599
+ø -11.6224
+ū -11.6224
+ú -11.6891
+ý -11.6891
+ș -11.7605
+ë -11.8374
+ş -11.8374
+ʿ -11.9207
+ī -11.9207
+ě -12.2228
+à -12.3478
+î -12.4906
+å -12.6573
+æ -12.8573
+ə -12.8573
+ă -13.1073
+ò -13.1073
+đ -13.1073
+ů -13.4406
+ň -13.9406
+û -13.9406
+œ -13.9406
+ť -13.9406
+ḫ -14.9395
+ź -14.9396
+ğ -14.9397
+ï -14.9398
+ń -14.9399
+ê -14.94
+ț -14.9401
+ą -14.9402
+ś -14.9403
+ñ -14.9404
+ž -14.9405
+ď -14.9406
+ð -14.9406
+ç -14.9406
+õ -14.9406
+ù -14.9406
+þ -14.9406
+ć -14.9406
+ė -14.9406
+ħ -14.9406
+ņ -14.9406
+ő -14.9406
+ż -14.9406
+μ -14.9406
+ṣ -14.9406
+ắ -14.9406
+ế -14.9406
+カ -14.9406
+临 -14.9406
+孙 -14.9406
+尣 -14.9406
+道 -14.9406
diff --git a/configs/de-en/tgt_unigram6000/spm_unigram_de.model b/configs/de-en/tgt_unigram6000/spm_unigram_de.model
new file mode 100644
index 0000000000000000000000000000000000000000..38e06f6866ed06d5043227bb67b2358e052867eb
--- /dev/null
+++ b/configs/de-en/tgt_unigram6000/spm_unigram_de.model
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:edbb7e7aac49ead78be0606edd7724901bc64cb6c077a37ecaf08e5ab13a0631
+size 344743
diff --git a/configs/de-en/tgt_unigram6000/spm_unigram_de.txt b/configs/de-en/tgt_unigram6000/spm_unigram_de.txt
new file mode 100644
index 0000000000000000000000000000000000000000..c99262c4a8a23af7fd6c2ec66f5f7b7b20334b9f
--- /dev/null
+++ b/configs/de-en/tgt_unigram6000/spm_unigram_de.txt
@@ -0,0 +1,5996 @@
+▁the 1
+s 1
+▁a 1
+▁of 1
+▁in 1
+▁to 1
+▁is 1
+▁and 1
+t 1
+▁was 1
+' 1
+▁it 1
+▁he 1
+ed 1
+▁be 1
+▁you 1
+▁for 1
+▁are 1
+ing 1
+n 1
+▁that 1
+▁on 1
+▁with 1
+▁this 1
+▁i 1
+d 1
+▁his 1
+▁as 1
+▁not 1
+ly 1
+y 1
+er 1
+▁can 1
+e 1
+▁from 1
+▁at 1
+▁have 1
+en 1
+▁there 1
+▁also 1
+▁ 1
+▁an 1
+▁by 1
+▁has 1
+▁one 1
+al 1
+▁do 1
+▁she 1
+r 1
+▁were 1
+▁we 1
+m 1
+▁time 1
+▁but 1
+▁all 1
+▁they 1
+▁her 1
+a 1
+▁only 1
+▁or 1
+▁no 1
+▁after 1
+▁will 1
+▁what 1
+o 1
+p 1
+es 1
+▁two 1
+▁me 1
+re 1
+c 1
+▁up 1
+▁out 1
+▁so 1
+▁first 1
+▁like 1
+▁had 1
+▁if 1
+▁your 1
+▁more 1
+▁does 1
+l 1
+▁other 1
+▁my 1
+▁into 1
+▁how 1
+▁when 1
+▁new 1
+▁now 1
+▁just 1
+ic 1
+ll 1
+▁about 1
+▁should 1
+▁their 1
+▁over 1
+▁here 1
+in 1
+▁would 1
+▁well 1
+g 1
+▁who 1
+▁get 1
+ion 1
+i 1
+▁don 1
+▁some 1
+▁still 1
+b 1
+▁did 1
+u 1
+le 1
+▁people 1
+▁which 1
+▁him 1
+k 1
+f 1
+▁used 1
+▁many 1
+▁however 1
+▁re 1
+▁go 1
+▁could 1
+▁un 1
+▁then 1
+ation 1
+▁again 1
+▁during 1
+▁part 1
+▁most 1
+▁place 1
+or 1
+▁these 1
+▁even 1
+▁work 1
+▁s 1
+el 1
+▁today 1
+▁already 1
+▁three 1
+te 1
+▁way 1
+us 1
+▁always 1
+la 1
+ri 1
+▁same 1
+▁through 1
+▁city 1
+▁name 1
+▁long 1
+▁later 1
+▁german 1
+ve 1
+▁right 1
+▁than 1
+ch 1
+▁year 1
+▁de 1
+▁war 1
+▁back 1
+▁e 1
+▁want 1
+z 1
+▁school 1
+▁world 1
+▁end 1
+▁take 1
+▁years 1
+ck 1
+an 1
+h 1
+li 1
+▁why 1
+cause 1
+▁where 1
+▁old 1
+ra 1
+▁made 1
+▁under 1
+▁too 1
+ers 1
+▁al 1
+▁know 1
+able 1
+ro 1
+▁make 1
+▁water 1
+▁c 1
+ce 1
+▁good 1
+▁look 1
+▁area 1
+ar 1
+▁without 1
+▁every 1
+▁co 1
+▁its 1
+▁please 1
+- 1
+▁off 1
+w 1
+on 1
+▁see 1
+th 1
+▁use 1
+▁o 1
+▁any 1
+▁became 1
+ter 1
+▁often 1
+▁last 1
+▁different 1
+▁g 1
+ur 1
+▁day 1
+▁both 1
+▁located 1
+▁four 1
+very 1
+▁car 1
+▁b 1
+▁think 1
+▁before 1
+▁sever 1
+▁let 1
+▁ver 1
+▁them 1
+▁am 1
+▁between 1
+▁second 1
+▁such 1
+ian 1
+it 1
+▁much 1
+de 1
+▁need 1
+▁district 1
+▁own 1
+ne 1
+▁put 1
+x 1
+il 1
+▁family 1
+is 1
+▁side 1
+est 1
+▁another 1
+▁high 1
+▁bo 1
+▁left 1
+▁state 1
+ia 1
+▁lot 1
+se 1
+▁found 1
+▁called 1
+▁important 1
+▁p 1
+▁something 1
+▁children 1
+▁town 1
+ble 1
+ma 1
+▁church 1
+▁our 1
+▁life 1
+▁t 1
+ies 1
+▁down 1
+▁main 1
+lo 1
+▁light 1
+▁small 1
+▁five 1
+ist 1
+▁free 1
+▁hand 1
+▁really 1
+ment 1
+▁being 1
+▁each 1
+▁next 1
+▁point 1
+▁must 1
+▁since 1
+▁took 1
+▁st 1
+▁further 1
+▁around 1
+▁man 1
+▁university 1
+ta 1
+ry 1
+▁building 1
+ity 1
+til 1
+▁addition 1
+▁come 1
+▁cannot 1
+▁won 1
+▁known 1
+▁against 1
+▁possible 1
+▁got 1
+na 1
+ul 1
+▁few 1
+um 1
+▁worked 1
+▁d 1
+▁company 1
+▁due 1
+ting 1
+▁head 1
+▁team 1
+▁son 1
+▁give 1
+▁us 1
+▁never 1
+▁mo 1
+▁national 1
+▁con 1
+ally 1
+ge 1
+▁course 1
+ol 1
+▁music 1
+▁great 1
+▁little 1
+▁mar 1
+▁going 1
+land 1
+▁village 1
+▁ro 1
+▁la 1
+▁station 1
+▁lead 1
+▁home 1
+▁public 1
+▁north 1
+▁film 1
+▁play 1
+▁open 1
+▁belong 1
+ru 1
+▁away 1
+▁v 1
+▁vo 1
+▁k 1
+▁best 1
+▁help 1
+▁short 1
+▁works 1
+▁ten 1
+▁far 1
+▁came 1
+ine 1
+▁li 1
+▁together 1
+▁himself 1
+▁system 1
+refore 1
+most 1
+▁air 1
+▁member 1
+ate 1
+▁everything 1
+ted 1
+ke 1
+▁large 1
+▁south 1
+co 1
+▁germany 1
+man 1
+▁center 1
+me 1
+▁di 1
+▁better 1
+▁among 1
+ary 1
+et 1
+▁find 1
+ive 1
+▁considered 1
+▁group 1
+▁law 1
+▁bad 1
+ut 1
+▁usual 1
+▁form 1
+ti 1
+go 1
+▁le 1
+▁office 1
+▁case 1
+▁house 1
+ant 1
+ler 1
+mp 1
+▁lo 1
+▁played 1
+▁result 1
+▁ha 1
+at 1
+vi 1
+▁father 1
+▁times 1
+to 1
+▁everyone 1
+▁berlin 1
+▁turn 1
+age 1
+▁show 1
+▁nothing 1
+▁once 1
+▁pre 1
+▁red 1
+ca 1
+ig 1
+▁order 1
+▁ra 1
+ther 1
+▁f 1
+ie 1
+id 1
+ner 1
+▁west 1
+▁pro 1
+da 1
+▁power 1
+v 1
+less 1
+▁room 1
+ci 1
+▁bar 1
+▁built 1
+▁number 1
+▁enough 1
+au 1
+▁sch 1
+ous 1
+▁close 1
+▁front 1
+▁per 1
+▁set 1
+▁bit 1
+▁become 1
+ton 1
+▁live 1
+▁land 1
+▁may 1
+▁sa 1
+▁construction 1
+ish 1
+▁real 1
+tic 1
+ling 1
+▁fire 1
+▁black 1
+ated 1
+▁studied 1
+▁service 1
+▁local 1
+▁line 1
+▁keep 1
+ism 1
+▁married 1
+▁middle 1
+▁completely 1
+bi 1
+▁named 1
+▁behind 1
+▁role 1
+▁sometimes 1
+▁season 1
+ff 1
+▁someone 1
+▁soon 1
+mi 1
+as 1
+▁stop 1
+▁party 1
+▁club 1
+▁things 1
+▁yet 1
+▁ma 1
+▁ta 1
+ial 1
+mann 1
+▁po 1
+▁band 1
+am 1
+▁following 1
+▁white 1
+berg 1
+▁call 1
+▁term 1
+▁say 1
+sh 1
+▁means 1
+▁age 1
+▁stay 1
+▁en 1
+▁six 1
+be 1
+▁history 1
+▁mainly 1
+▁castle 1
+ness 1
+▁longer 1
+▁ga 1
+▁round 1
+▁ti 1
+▁anything 1
+▁love 1
+ga 1
+▁went 1
+▁pa 1
+▁half 1
+▁region 1
+▁association 1
+ten 1
+▁language 1
+▁park 1
+▁period 1
+▁art 1
+▁probably 1
+▁wall 1
+▁community 1
+▁special 1
+▁eight 1
+▁river 1
+▁ho 1
+▁u 1
+▁big 1
+ber 1
+▁example 1
+▁position 1
+▁numerous 1
+▁started 1
+▁hu 1
+▁mi 1
+▁lu 1
+▁quickly 1
+▁change 1
+▁others 1
+ir 1
+ni 1
+▁whole 1
+rs 1
+▁international 1
+▁run 1
+▁according 1
+▁instead 1
+▁lives 1
+▁former 1
+▁street 1
+▁afterwards 1
+▁death 1
+▁train 1
+▁east 1
+bo 1
+▁mr 1
+▁yourself 1
+▁try 1
+▁comes 1
+▁hundred 1
+pe 1
+di 1
+▁else 1
+▁food 1
+▁mu 1
+▁received 1
+▁court 1
+▁federal 1
+▁various 1
+▁actually 1
+ha 1
+▁read 1
+▁lies 1
+▁ex 1
+▁test 1
+▁vi 1
+ot 1
+son 1
+▁based 1
+▁ju 1
+▁those 1
+ance 1
+ure 1
+bu 1
+ad 1
+▁career 1
+▁hard 1
+ty 1
+gg 1
+ec 1
+ka 1
+un 1
+▁clear 1
+▁bi 1
+▁seem 1
+▁early 1
+▁common 1
+▁post 1
+▁board 1
+▁dis 1
+▁research 1
+ical 1
+▁pi 1
+pp 1
+▁project 1
+▁na 1
+▁ka 1
+▁thing 1
+▁start 1
+▁sp 1
+▁followed 1
+▁general 1
+▁money 1
+▁capital 1
+tz 1
+▁reason 1
+▁english 1
+▁da 1
+▁wants 1
+ph 1
+▁process 1
+▁market 1
+hi 1
+ver 1
+▁va 1
+▁list 1
+▁production 1
+▁country 1
+▁support 1
+▁w 1
+▁ge 1
+▁ba 1
+▁night 1
+▁individual 1
+▁french 1
+▁believe 1
+▁oh 1
+om 1
+▁class 1
+▁title 1
+▁while 1
+▁road 1
+▁game 1
+mb 1
+▁remained 1
+▁em 1
+▁lower 1
+sch 1
+▁espe 1
+▁body 1
+▁human 1
+▁color 1
+▁hall 1
+▁te 1
+▁wood 1
+▁forest 1
+▁r 1
+▁young 1
+▁said 1
+cy 1
+▁lost 1
+▁model 1
+▁past 1
+▁person 1
+que 1
+▁transport 1
+▁bring 1
+do 1
+cially 1
+▁field 1
+op 1
+ster 1
+mo 1
+gen 1
+▁brother 1
+▁near 1
+▁working 1
+▁seen 1
+▁business 1
+▁remove 1
+ho 1
+▁active 1
+▁space 1
+▁data 1
+▁job 1
+▁h 1
+▁hold 1
+▁child 1
+▁top 1
+▁mostly 1
+▁program 1
+▁military 1
+▁level 1
+▁allowed 1
+▁sta 1
+▁book 1
+▁development 1
+▁cross 1
+▁mother 1
+▁least 1
+hr 1
+▁initially 1
+▁care 1
+▁author 1
+no 1
+▁se 1
+▁seven 1
+▁talk 1
+▁cause 1
+und 1
+▁pe 1
+▁voice 1
+▁consist 1
+ized 1
+▁ever 1
+▁education 1
+▁official 1
+▁thought 1
+▁strong 1
+▁railway 1
+▁natural 1
+▁drive 1
+▁currently 1
+▁dog 1
+▁makes 1
+▁ice 1
+▁seat 1
+▁interest 1
+▁beginning 1
+▁matter 1
+▁present 1
+▁finally 1
+▁single 1
+▁word 1
+▁purpose 1
+ger 1
+▁difficult 1
+▁route 1
+▁government 1
+lin 1
+▁similar 1
+▁teacher 1
+▁available 1
+▁ri 1
+▁europe 1
+▁subject 1
+▁cut 1
+ard 1
+▁sun 1
+les 1
+▁fall 1
+▁league 1
+▁series 1
+▁du 1
+▁european 1
+▁bank 1
+▁central 1
+▁ko 1
+▁blue 1
+▁break 1
+▁lake 1
+▁species 1
+▁kind 1
+▁mean 1
+▁ni 1
+▁follow 1
+▁tell 1
+▁plant 1
+tor 1
+▁pass 1
+▁men 1
+ba 1
+▁american 1
+▁green 1
+▁born 1
+he 1
+▁social 1
+▁ground 1
+▁sea 1
+▁died 1
+▁yes 1
+▁view 1
+▁tri 1
+▁leave 1
+▁hair 1
+▁island 1
+▁act 1
+▁jo 1
+▁nine 1
+▁door 1
+▁current 1
+led 1
+▁twenty 1
+▁thus 1
+os 1
+▁connection 1
+▁developed 1
+▁upper 1
+▁sound 1
+▁council 1
+we 1
+▁song 1
+▁wanted 1
+▁depend 1
+va 1
+▁roman 1
+ology 1
+▁learn 1
+min 1
+▁tower 1
+▁fact 1
+▁success 1
+▁maybe 1
+io 1
+▁created 1
+▁additional 1
+▁stand 1
+▁full 1
+▁idea 1
+▁members 1
+▁days 1
+▁fish 1
+▁count 1
+▁third 1
+ship 1
+vo 1
+▁taken 1
+▁final 1
+▁rock 1
+der 1
+▁garden 1
+▁text 1
+▁taste 1
+▁less 1
+▁themselves 1
+▁police 1
+▁week 1
+▁regional 1
+▁stone 1
+▁el 1
+▁museum 1
+ful 1
+▁rest 1
+▁hi 1
+▁ru 1
+▁moved 1
+ach 1
+▁machine 1
+▁album 1
+▁held 1
+▁animals 1
+▁itself 1
+▁needs 1
+▁published 1
+▁organ 1
+▁drink 1
+ization 1
+▁goal 1
+▁control 1
+▁economic 1
+▁low 1
+▁women 1
+▁anymore 1
+▁cup 1
+▁master 1
+mar 1
+▁takes 1
+▁successful 1
+▁director 1
+▁pay 1
+▁feel 1
+▁m 1
+ow 1
+▁su 1
+▁within 1
+▁design 1
+burg 1
+▁famous 1
+▁version 1
+▁written 1
+▁looks 1
+▁hit 1
+▁step 1
+▁heat 1
+▁areas 1
+im 1
+pa 1
+▁certain 1
+▁understand 1
+▁sleep 1
+▁information 1
+▁style 1
+▁responsible 1
+▁ca 1
+▁political 1
+▁traffic 1
+lu 1
+▁port 1
+▁cover 1
+dy 1
+den 1
+▁attack 1
+▁border 1
+▁mark 1
+▁supposed 1
+▁population 1
+mon 1
+▁je 1
+▁daughter 1
+▁wine 1
+▁th 1
+▁ste 1
+▁training 1
+nna 1
+▁track 1
+▁sub 1
+▁easy 1
+▁honor 1
+▁card 1
+▁along 1
+▁material 1
+▁department 1
+▁ch 1
+▁wa 1
+▁si 1
+▁ask 1
+▁pu 1
+▁structure 1
+▁gets 1
+▁direction 1
+▁floor 1
+▁done 1
+▁nature 1
+▁parents 1
+▁anyone 1
+▁rath 1
+wa 1
+▁king 1
+▁player 1
+▁stage 1
+▁gas 1
+▁either 1
+▁anti 1
+▁whe 1
+▁radio 1
+▁cool 1
+rk 1
+ron 1
+▁popular 1
+▁via 1
+▁cold 1
+▁influence 1
+▁moment 1
+▁gu 1
+▁original 1
+bach 1
+▁remains 1
+▁wrote 1
+▁ball 1
+▁army 1
+ite 1
+▁led 1
+▁towards 1
+wi 1
+▁professional 1
+up 1
+▁speak 1
+▁mountain 1
+if 1
+▁began 1
+▁modern 1
+▁walk 1
+▁bus 1
+▁sense 1
+▁watch 1
+▁nice 1
+▁exactly 1
+▁wrong 1
+▁anyway 1
+rought 1
+▁deal 1
+▁value 1
+red 1
+▁return 1
+▁die 1
+▁doing 1
+▁bra 1
+▁store 1
+▁using 1
+▁largest 1
+▁eat 1
+▁goes 1
+tu 1
+▁pan 1
+▁games 1
+▁mind 1
+▁eye 1
+gu 1
+ize 1
+more 1
+▁bu 1
+▁god 1
+▁focus 1
+▁performance 1
+▁above 1
+▁risk 1
+▁boy 1
+▁living 1
+▁southern 1
+ng 1
+▁dead 1
+▁ship 1
+mer 1
+▁studies 1
+▁entire 1
+▁grew 1
+▁pull 1
+nder 1
+tra 1
+▁ne 1
+▁minutes 1
+▁private 1
+▁plan 1
+▁non 1
+▁necessary 1
+▁self 1
+way 1
+▁lack 1
+▁question 1
+▁changed 1
+▁civil 1
+▁contract 1
+▁action 1
+▁roll 1
+▁er 1
+▁carrie 1
+▁technical 1
+▁outside 1
+▁ki 1
+▁bre 1
+▁location 1
+▁buried 1
+▁hope 1
+▁check 1
+▁playing 1
+▁hours 1
+▁cla 1
+per 1
+▁leaves 1
+▁experience 1
+▁story 1
+▁sand 1
+▁guy 1
+bra 1
+▁lived 1
+ze 1
+▁given 1
+▁award 1
+▁brown 1
+▁beside 1
+▁wife 1
+▁served 1
+▁prevent 1
+▁immediately 1
+▁approach 1
+ator 1
+▁task 1
+▁hot 1
+▁particular 1
+▁turned 1
+▁shows 1
+▁star 1
+▁network 1
+▁northern 1
+▁municipality 1
+▁base 1
+em 1
+▁race 1
+ak 1
+▁sign 1
+▁include 1
+▁temperature 1
+▁christian 1
+▁dark 1
+▁serve 1
+▁trade 1
+▁simpl 1
+▁buy 1
+▁administrative 1
+▁though 1
+▁table 1
+ists 1
+▁method 1
+▁basic 1
+▁coat 1
+▁protect 1
+▁youth 1
+▁clean 1
+▁study 1
+▁size 1
+▁neu 1
+▁paul 1
+▁president 1
+ium 1
+▁foundation 1
+▁parts 1
+pi 1
+▁operation 1
+▁face 1
+men 1
+▁win 1
+▁terms 1
+▁plays 1
+▁record 1
+line 1
+▁students 1
+▁particularly 1
+▁continued 1
+▁cat 1
+▁decision 1
+▁industry 1
+▁travel 1
+▁deep 1
+ki 1
+▁cell 1
+▁historical 1
+▁sure 1
+▁cap 1
+▁couple 1
+▁horse 1
+▁happen 1
+che 1
+▁produced 1
+▁gra 1
+▁pressure 1
+▁fe 1
+▁doctor 1
+heim 1
+▁function 1
+▁ke 1
+▁speed 1
+▁feature 1
+▁monastery 1
+▁thousand 1
+▁groups 1
+ac 1
+▁move 1
+▁tour 1
+▁oil 1
+▁arms 1
+▁unfortunately 1
+▁tree 1
+▁sports 1
+ric 1
+taking 1
+▁wide 1
+▁difference 1
+▁listen 1
+ius 1
+▁hands 1
+▁battle 1
+▁somehow 1
+▁actor 1
+▁date 1
+▁jump 1
+ud 1
+▁key 1
+▁type 1
+▁thirty 1
+▁nobody 1
+▁hamburg 1
+▁settlement 1
+▁ear 1
+▁closed 1
+▁buildings 1
+▁inter 1
+▁search 1
+▁despi 1
+▁bri 1
+▁legal 1
+▁regard 1
+▁site 1
+▁champion 1
+▁blood 1
+▁heart 1
+ide 1
+▁companies 1
+▁leg 1
+▁theater 1
+▁san 1
+▁organization 1
+▁fell 1
+▁grave 1
+▁gold 1
+▁extreme 1
+▁valley 1
+▁books 1
+▁covered 1
+▁secret 1
+▁section 1
+▁distance 1
+▁standard 1
+ier 1
+▁ran 1
+▁para 1
+▁magazine 1
+▁tomorrow 1
+▁bear 1
+▁protection 1
+itz 1
+▁unit 1
+▁catholic 1
+▁attention 1
+▁tu 1
+▁plants 1
+▁primari 1
+▁range 1
+▁glass 1
+▁share 1
+▁add 1
+▁county 1
+▁contact 1
+▁munich 1
+▁pick 1
+▁tra 1
+▁code 1
+▁degree 1
+▁bottle 1
+▁double 1
+▁origin 1
+▁otherwise 1
+▁movement 1
+▁provide 1
+iv 1
+▁manage 1
+▁across 1
+ors 1
+▁total 1
+fa 1
+ina 1
+▁knows 1
+▁province 1
+▁fight 1
+▁spot 1
+▁reached 1
+▁established 1
+▁scientific 1
+▁yellow 1
+▁originally 1
+▁cities 1
+▁apparently 1
+▁simple 1
+▁effect 1
+hl 1
+▁letter 1
+▁eyes 1
+▁directly 1
+▁wi 1
+▁save 1
+▁lay 1
+▁france 1
+▁phone 1
+the 1
+▁earth 1
+▁sentence 1
+▁friend 1
+▁cha 1
+▁true 1
+▁lie 1
+qui 1
+▁higher 1
+▁driver 1
+▁attempt 1
+▁cemetery 1
+▁cultural 1
+▁monument 1
+▁minister 1
+▁demand 1
+▁roof 1
+feld 1
+▁musical 1
+▁problem 1
+▁sentences 1
+▁ku 1
+▁report 1
+tan 1
+▁entrance 1
+▁marriage 1
+▁situation 1
+▁computer 1
+▁trains 1
+▁mrs 1
+▁shortly 1
+▁rules 1
+▁union 1
+▁happened 1
+▁create 1
+▁competition 1
+▁division 1
+ven 1
+▁nor 1
+▁multiple 1
+▁austria 1
+▁throw 1
+▁slowly 1
+▁hardly 1
+▁finished 1
+row 1
+▁ski 1
+▁mistake 1
+▁technology 1
+▁british 1
+▁vienna 1
+▁answer 1
+▁divided 1
+▁charge 1
+▁wind 1
+▁document 1
+▁hear 1
+▁happy 1
+▁painting 1
+▁completed 1
+lar 1
+bert 1
+▁apart 1
+▁surface 1
+▁figure 1
+▁weather 1
+▁formed 1
+▁print 1
+▁coming 1
+▁path 1
+▁founded 1
+▁host 1
+▁decided 1
+▁develop 1
+▁build 1
+▁major 1
+▁future 1
+ified 1
+▁hans 1
+▁camp 1
+▁words 1
+▁names 1
+▁shi 1
+▁peter 1
+▁republic 1
+▁fruit 1
+▁edge 1
+▁poor 1
+▁cost 1
+▁literature 1
+▁fair 1
+▁collection 1
+▁require 1
+▁ja 1
+▁independent 1
+▁months 1
+▁flo 1
+▁described 1
+ities 1
+side 1
+▁movie 1
+▁bridge 1
+▁wait 1
+▁iron 1
+▁saw 1
+▁force 1
+▁inside 1
+▁media 1
+▁stra 1
+▁send 1
+▁science 1
+▁energy 1
+▁l 1
+ich 1
+▁z 1
+tal 1
+▁appeared 1
+ever 1
+having 1
+▁officer 1
+ok 1
+▁chi 1
+▁fin 1
+ren 1
+▁solution 1
+▁saint 1
+▁planned 1
+▁worth 1
+▁meant 1
+▁dance 1
+▁mill 1
+▁suit 1
+▁trans 1
+▁visit 1
+▁notice 1
+▁paris 1
+▁remember 1
+▁scene 1
+▁avoid 1
+▁sister 1
+▁neighbor 1
+▁wear 1
+ov 1
+▁fan 1
+by 1
+▁involved 1
+▁female 1
+▁internet 1
+▁offer 1
+▁runs 1
+▁problems 1
+▁fast 1
+▁attended 1
+▁met 1
+▁forms 1
+▁easily 1
+▁late 1
+▁lock 1
+▁write 1
+▁older 1
+▁concert 1
+▁cream 1
+▁guest 1
+▁opened 1
+▁generally 1
+▁spa 1
+▁fla 1
+▁doubt 1
+▁newspaper 1
+▁writing 1
+hn 1
+▁comp 1
+fi 1
+▁tax 1
+▁direct 1
+▁dream 1
+▁silver 1
+▁cook 1
+▁flight 1
+▁sal 1
+▁increase 1
+▁farm 1
+▁rain 1
+▁measure 1
+▁har 1
+▁administration 1
+▁friedrich 1
+▁twelve 1
+▁driving 1
+▁morning 1
+▁countries 1
+gi 1
+▁price 1
+ward 1
+▁states 1
+▁significant 1
+log 1
+▁application 1
+▁festival 1
+▁principle 1
+▁content 1
+cient 1
+▁treat 1
+▁italian 1
+▁touch 1
+▁john 1
+▁imp 1
+ran 1
+con 1
+enburg 1
+uc 1
+▁awarded 1
+let 1
+▁released 1
+▁herself 1
+▁relationship 1
+▁foot 1
+▁kilometers 1
+▁receive 1
+▁changes 1
+▁trees 1
+▁animal 1
+▁miss 1
+▁professor 1
+▁kept 1
+▁fo 1
+▁exist 1
+graph 1
+▁agriculture 1
+▁nowadays 1
+▁included 1
+▁quit 1
+▁thereby 1
+▁leader 1
+▁returned 1
+▁television 1
+▁medal 1
+▁sold 1
+▁coach 1
+▁fun 1
+▁shot 1
+▁accident 1
+▁square 1
+▁fill 1
+▁fit 1
+▁tro 1
+▁rule 1
+▁sport 1
+▁switch 1
+val 1
+▁gi 1
+▁nevertheless 1
+▁society 1
+▁yesterday 1
+▁coast 1
+▁preserved 1
+▁sha 1
+▁cor 1
+▁commercial 1
+▁airport 1
+▁press 1
+ological 1
+time 1
+▁paper 1
+▁personal 1
+▁supported 1
+▁dam 1
+▁cable 1
+▁languages 1
+▁link 1
+▁novel 1
+▁appearance 1
+ug 1
+▁fi 1
+▁shape 1
+▁failed 1
+▁n 1
+▁thank 1
+▁performed 1
+▁wing 1
+▁volume 1
+▁spread 1
+▁cur 1
+▁wish 1
+cht 1
+▁fa 1
+▁starts 1
+▁ring 1
+▁meeting 1
+ose 1
+▁fu 1
+▁grand 1
+▁object 1
+▁connected 1
+▁circle 1
+▁coffee 1
+▁twice 1
+▁football 1
+▁flat 1
+▁contains 1
+▁occur 1
+▁college 1
+▁expert 1
+▁smoke 1
+▁shoot 1
+▁carry 1
+▁clearly 1
+▁shaped 1
+▁tip 1
+▁concept 1
+▁escape 1
+▁russian 1
+▁sur 1
+▁counter 1
+ries 1
+▁partner 1
+▁burn 1
+▁match 1
+▁municipal 1
+▁windows 1
+▁vehicle 1
+▁talent 1
+▁milk 1
+▁making 1
+▁display 1
+▁eastern 1
+▁ended 1
+▁pet 1
+▁recently 1
+▁karl 1
+▁speaking 1
+▁ya 1
+▁houses 1
+▁editor 1
+▁knowledge 1
+iz 1
+▁noise 1
+▁majority 1
+▁heard 1
+▁winter 1
+▁shop 1
+▁italy 1
+▁fine 1
+▁cra 1
+nne 1
+▁allow 1
+▁owner 1
+han 1
+▁electric 1
+▁flows 1
+▁access 1
+▁library 1
+▁running 1
+▁western 1
+▁pen 1
+▁fly 1
+hausen 1
+▁motor 1
+ik 1
+▁che 1
+▁results 1
+ai 1
+▁wheel 1
+▁events 1
+▁reasons 1
+▁der 1
+▁pri 1
+▁reach 1
+▁meaning 1
+▁behavior 1
+▁institute 1
+▁choir 1
+▁rich 1
+one 1
+▁statement 1
+▁ban 1
+▁mor 1
+▁friends 1
+▁safe 1
+▁amount 1
+▁procedure 1
+▁committee 1
+▁products 1
+enberg 1
+▁kar 1
+▁kill 1
+▁thanks 1
+▁mat 1
+▁trip 1
+▁comment 1
+▁broken 1
+▁arm 1
+▁beautiful 1
+▁switzerland 1
+▁opinion 1
+▁quarter 1
+▁straight 1
+▁lawyer 1
+▁complete 1
+chi 1
+▁y 1
+▁health 1
+▁dear 1
+▁designed 1
+▁instrument 1
+▁complain 1
+▁bishop 1
+▁election 1
+▁sort 1
+▁row 1
+▁stands 1
+stadt 1
+▁hour 1
+▁destroyed 1
+▁weight 1
+bel 1
+▁event 1
+▁tea 1
+tel 1
+▁arch 1
+▁cologne 1
+▁london 1
+▁video 1
+▁prison 1
+▁management 1
+ring 1
+▁character 1
+▁musician 1
+▁contain 1
+▁practice 1
+▁weak 1
+▁signal 1
+▁sounds 1
+▁offers 1
+▁asked 1
+▁limited 1
+▁equal 1
+ja 1
+▁baby 1
+▁release 1
+ek 1
+za 1
+▁separate 1
+▁article 1
+▁highly 1
+▁america 1
+▁offered 1
+▁starting 1
+▁fat 1
+istic 1
+▁highest 1
+▁below 1
+▁source 1
+▁joined 1
+▁cases 1
+▁grow 1
+▁graduat 1
+▁security 1
+▁block 1
+▁blind 1
+▁medical 1
+▁image 1
+▁woman 1
+▁bomb 1
+▁bike 1
+▁enter 1
+▁app 1
+▁writer 1
+fer 1
+▁johann 1
+cke 1
+eg 1
+▁connect 1
+mark 1
+▁quality 1
+▁increasing 1
+▁summer 1
+▁tourist 1
+old 1
+▁achieved 1
+▁picture 1
+▁mine 1
+▁dr 1
+▁cal 1
+over 1
+▁lose 1
+▁screen 1
+▁greek 1
+▁alternative 1
+▁ober 1
+▁wor 1
+cker 1
+▁weeks 1
+▁questions 1
+▁burg 1
+called 1
+▁removed 1
+▁absolute 1
+▁mouth 1
+▁prince 1
+▁strasse 1
+▁bread 1
+▁exhibition 1
+▁col 1
+▁parish 1
+▁par 1
+▁tail 1
+▁teams 1
+▁damage 1
+▁parliament 1
+▁hill 1
+work 1
+▁championship 1
+▁songs 1
+▁hospital 1
+▁culture 1
+▁reserve 1
+▁adult 1
+▁bag 1
+▁claim 1
+▁exception 1
+▁passed 1
+▁box 1
+▁boat 1
+ham 1
+pper 1
+▁might 1
+▁prize 1
+▁martin 1
+▁accept 1
+ative 1
+▁customer 1
+▁wo 1
+▁fle 1
+▁points 1
+▁normal 1
+▁artist 1
+▁pra 1
+▁engine 1
+▁permanent 1
+▁foreign 1
+▁crown 1
+▁royal 1
+▁balance 1
+▁super 1
+▁explain 1
+▁continue 1
+▁rank 1
+▁fields 1
+▁corner 1
+▁piece 1
+for 1
+▁euro 1
+▁heavy 1
+▁percent 1
+▁typical 1
+▁wilhelm 1
+▁guitar 1
+▁memory 1
+▁relatively 1
+▁waste 1
+quite 1
+▁costs 1
+▁ages 1
+▁successor 1
+▁gav 1
+▁moon 1
+▁correct 1
+gan 1
+▁records 1
+▁tur 1
+▁branch 1
+ctor 1
+▁wild 1
+▁planet 1
+▁appointed 1
+▁software 1
+▁overall 1
+▁forget 1
+▁scar 1
+out 1
+▁evening 1
+gel 1
+▁fund 1
+▁attract 1
+▁activities 1
+▁extensive 1
+▁broadcast 1
+▁background 1
+▁min 1
+▁wave 1
+▁pure 1
+▁x 1
+▁referr 1
+▁grown 1
+▁singer 1
+▁topic 1
+▁mass 1
+▁consider 1
+▁audience 1
+▁frankfurt 1
+▁orient 1
+▁judge 1
+▁thick 1
+▁authorities 1
+▁chance 1
+▁perform 1
+▁meet 1
+▁subsequently 1
+▁ros 1
+▁united 1
+▁bro 1
+▁pain 1
+▁significantly 1
+▁gun 1
+▁dry 1
+▁metal 1
+▁austrian 1
+▁communities 1
+▁vehicles 1
+▁york 1
+▁hol 1
+lon 1
+▁camera 1
+▁specialist 1
+▁pictures 1
+▁tradition 1
+rge 1
+▁says 1
+▁lines 1
+▁represented 1
+▁account 1
+▁institution 1
+▁teeth 1
+▁opposite 1
+▁missing 1
+▁appear 1
+▁limit 1
+sdorf 1
+▁spe 1
+▁regular 1
+▁beat 1
+▁meters 1
+▁product 1
+▁defense 1
+▁religious 1
+▁stretch 1
+▁user 1
+▁africa 1
+▁mountains 1
+▁colors 1
+▁warm 1
+▁rat 1
+del 1
+▁neighbors 1
+▁jewish 1
+▁landscape 1
+▁opponent 1
+▁situated 1
+▁fresh 1
+▁ze 1
+▁representative 1
+▁window 1
+▁cake 1
+▁rail 1
+▁male 1
+▁pieces 1
+▁mayor 1
+▁recognize 1
+▁shown 1
+▁trained 1
+▁required 1
+▁exchange 1
+▁stuff 1
+▁estate 1
+▁chair 1
+▁orange 1
+▁sugar 1
+▁sector 1
+▁beach 1
+▁apartment 1
+ile 1
+▁skin 1
+▁factory 1
+▁previously 1
+cher 1
+▁bor 1
+▁loud 1
+vent 1
+▁standing 1
+stein 1
+▁react 1
+▁financial 1
+▁spent 1
+▁movies 1
+▁push 1
+▁fuel 1
+▁expect 1
+▁reality 1
+▁flow 1
+▁replaced 1
+▁theory 1
+▁command 1
+mmer 1
+▁previous 1
+▁serious 1
+house 1
+▁advantage 1
+▁emergency 1
+▁interior 1
+▁stick 1
+▁funny 1
+▁hurt 1
+▁stream 1
+▁flag 1
+zer 1
+▁bone 1
+▁rate 1
+▁effects 1
+ue 1
+▁golden 1
+light 1
+aries 1
+▁artificial 1
+▁defend 1
+▁register 1
+▁captain 1
+▁complex 1
+▁weekend 1
+▁importance 1
+▁rid 1
+▁unknown 1
+▁pol 1
+▁speaker 1
+▁describe 1
+▁broke 1
+ette 1
+▁saying 1
+▁issue 1
+▁films 1
+▁exclusive 1
+▁olympic 1
+▁understood 1
+▁everybody 1
+▁brand 1
+definitely 1
+▁chairman 1
+▁winner 1
+▁painter 1
+je 1
+▁practical 1
+▁obviously 1
+▁van 1
+▁jan 1
+▁student 1
+▁listed 1
+ible 1
+▁tries 1
+▁fur 1
+▁birth 1
+▁august 1
+▁conflict 1
+▁effort 1
+▁passenger 1
+▁supply 1
+▁graphic 1
+▁shopping 1
+▁sorry 1
+▁desert 1
+▁spring 1
+▁reich 1
+▁bay 1
+▁customs 1
+▁fee 1
+▁bare 1
+ification 1
+▁fourth 1
+▁pack 1
+ley 1
+▁sit 1
+▁provided 1
+▁commission 1
+▁soccer 1
+▁latter 1
+▁latin 1
+▁spend 1
+▁automatically 1
+▁property 1
+▁indian 1
+▁industrial 1
+ition 1
+rew 1
+par 1
+com 1
+▁nu 1
+▁rarely 1
+▁ride 1
+car 1
+▁flowers 1
+van 1
+▁reflect 1
+▁applied 1
+▁snow 1
+▁wolf 1
+▁swiss 1
+▁zone 1
+▁bla 1
+▁creat 1
+ates 1
+▁climb 1
+▁mentioned 1
+▁resident 1
+▁conditions 1
+▁meat 1
+▁measures 1
+▁hero 1
+▁needed 1
+▁agricultural 1
+▁heinrich 1
+▁slightly 1
+▁composition 1
+▁pur 1
+▁drug 1
+▁meantime 1
+▁shock 1
+▁gear 1
+▁bern 1
+ability 1
+our 1
+▁discovered 1
+ang 1
+though 1
+▁acting 1
+▁respect 1
+▁specific 1
+▁disease 1
+▁equipped 1
+▁inhabitants 1
+▁rhine 1
+▁eleven 1
+▁click 1
+▁visible 1
+▁piano 1
+▁remain 1
+▁earlier 1
+▁killed 1
+▁brake 1
+▁digital 1
+▁finish 1
+▁bird 1
+▁pit 1
+▁march 1
+port 1
+▁pot 1
+▁evidence 1
+▁enjoy 1
+▁kitchen 1
+▁license 1
+▁spanish 1
+▁smell 1
+▁taught 1
+▁protestant 1
+▁package 1
+▁drum 1
+▁chain 1
+uck 1
+▁ger 1
+▁joke 1
+ria 1
+▁smaller 1
+can 1
+▁numbers 1
+wald 1
+▁diet 1
+▁sing 1
+▁collect 1
+▁admit 1
+▁contrast 1
+▁perfect 1
+▁somewhere 1
+▁favorite 1
+▁communication 1
+▁appears 1
+▁systems 1
+▁china 1
+▁represent 1
+ography 1
+▁max 1
+▁producer 1
+▁secondary 1
+▁forces 1
+▁regularly 1
+▁element 1
+▁quick 1
+▁drop 1
+▁tool 1
+lie 1
+▁qua 1
+▁restaurant 1
+▁characteristic 1
+▁headquarters 1
+fully 1
+ung 1
+▁aim 1
+▁correspond 1
+▁symbol 1
+▁era 1
+▁draw 1
+▁loss 1
+▁increased 1
+▁reform 1
+▁length 1
+▁examination 1
+▁empty 1
+▁england 1
+▁insurance 1
+▁stopped 1
+▁cloud 1
+▁forward 1
+▁frank 1
+▁trick 1
+▁skill 1
+illa 1
+hu 1
+▁elements 1
+▁blu 1
+▁maria 1
+▁dare 1
+▁stem 1
+hen 1
+chen 1
+▁photo 1
+▁lau 1
+▁month 1
+▁manager 1
+▁afraid 1
+▁announced 1
+▁conference 1
+▁experiment 1
+▁tournament 1
+▁sweet 1
+▁speech 1
+▁citizens 1
+▁acid 1
+▁strip 1
+▁whatever 1
+▁scale 1
+▁prove 1
+▁traditional 1
+▁plane 1
+▁birds 1
+▁recognized 1
+▁ideal 1
+▁opening 1
+▁gone 1
+▁und 1
+▁surrounded 1
+▁surrounding 1
+▁narrow 1
+▁troops 1
+▁discussion 1
+▁sight 1
+ino 1
+▁elected 1
+▁frequently 1
+▁compared 1
+▁layer 1
+▁berg 1
+▁severe 1
+▁championships 1
+▁happens 1
+▁joint 1
+dorf 1
+▁mal 1
+▁christ 1
+▁primar 1
+ular 1
+▁michael 1
+▁orchestra 1
+▁generation 1
+▁promise 1
+▁lessons 1
+▁staff 1
+▁op 1
+▁saxony 1
+▁nut 1
+▁rose 1
+▁composer 1
+▁luck 1
+▁dia 1
+▁players 1
+▁determined 1
+▁mis 1
+▁academy 1
+▁request 1
+▁edition 1
+▁stories 1
+▁criticism 1
+▁bright 1
+▁prefer 1
+▁electro 1
+▁machines 1
+▁mission 1
+lor 1
+▁visited 1
+▁concern 1
+▁char 1
+▁artists 1
+▁tre 1
+bility 1
+▁services 1
+▁pig 1
+▁including 1
+▁repair 1
+▁chamber 1
+▁adapt 1
+▁brief 1
+▁learned 1
+▁legs 1
+▁daily 1
+▁message 1
+▁sitting 1
+▁reaction 1
+▁soil 1
+ath 1
+▁neck 1
+▁feed 1
+known 1
+▁feeling 1
+▁ray 1
+▁schu 1
+▁flying 1
+▁classic 1
+▁distribution 1
+▁flash 1
+▁thin 1
+▁mode 1
+▁spin 1
+▁eggs 1
+▁larger 1
+▁disappeared 1
+▁added 1
+▁flee 1
+atic 1
+▁ahead 1
+▁ep 1
+▁sho 1
+▁busy 1
+▁plate 1
+pel 1
+▁concrete 1
+▁universities 1
+▁storm 1
+▁ticket 1
+▁visitors 1
+▁respective 1
+▁stock 1
+▁j 1
+bar 1
+▁presented 1
+form 1
+▁finds 1
+▁flu 1
+ese 1
+▁showed 1
+▁recent 1
+▁hole 1
+▁express 1
+▁mess 1
+▁spit 1
+▁anne 1
+▁device 1
+▁minute 1
+▁stomach 1
+▁knew 1
+▁knock 1
+▁stars 1
+▁chapel 1
+▁nei 1
+og 1
+▁waiting 1
+▁poem 1
+ker 1
+ku 1
+▁younger 1
+▁domin 1
+▁upon 1
+▁issues 1
+▁marked 1
+▁choice 1
+▁employed 1
+▁myself 1
+▁excuse 1
+▁basis 1
+▁detect 1
+▁operating 1
+▁organized 1
+▁option 1
+▁girl 1
+▁carrier 1
+▁types 1
+▁plans 1
+▁engines 1
+cal 1
+▁operated 1
+▁mont 1
+▁exists 1
+mate 1
+▁recommended 1
+▁solo 1
+clock 1
+lessly 1
+▁criminal 1
+▁leaf 1
+▁copy 1
+▁occasionally 1
+▁guard 1
+▁agent 1
+▁affected 1
+▁apple 1
+▁bur 1
+▁agree 1
+ches 1
+wing 1
+▁activity 1
+▁dialect 1
+▁load 1
+▁label 1
+▁ruin 1
+▁mail 1
+▁journalist 1
+her 1
+▁constant 1
+▁mixed 1
+▁cas 1
+▁del 1
+▁fingers 1
+▁bell 1
+▁cathedral 1
+▁kindergarten 1
+▁biggest 1
+▁ocean 1
+▁robert 1
+▁auto 1
+ple 1
+▁brick 1
+▁queen 1
+▁sales 1
+▁bal 1
+isch 1
+▁falls 1
+▁sources 1
+hof 1
+▁baden 1
+▁container 1
+▁trust 1
+▁politics 1
+▁positive 1
+▁okay 1
+▁laugh 1
+▁robot 1
+▁lord 1
+▁childhood 1
+▁crime 1
+▁frame 1
+▁status 1
+▁manufacturer 1
+▁steam 1
+▁throughout 1
+▁minor 1
+▁pool 1
+▁hotel 1
+▁bio 1
+▁reading 1
+▁ready 1
+▁rare 1
+▁recorded 1
+▁mix 1
+ike 1
+▁critic 1
+▁differently 1
+▁gen 1
+▁salt 1
+▁singing 1
+ita 1
+▁za 1
+▁architect 1
+▁candidate 1
+▁economy 1
+▁leaving 1
+▁meanwhile 1
+▁patient 1
+▁religion 1
+▁studio 1
+▁empire 1
+▁urban 1
+▁filter 1
+▁liquid 1
+▁policy 1
+▁tongue 1
+▁medium 1
+▁fought 1
+▁characterized 1
+▁blow 1
+▁zero 1
+▁occurs 1
+▁arranged 1
+▁teaching 1
+▁sin 1
+▁web 1
+▁et 1
+▁sto 1
+ey 1
+▁consequences 1
+▁functions 1
+▁daughters 1
+▁gar 1
+▁peace 1
+▁placed 1
+▁kl 1
+▁flower 1
+▁eva 1
+▁nose 1
+ya 1
+▁conduct 1
+▁cooking 1
+▁expensive 1
+▁medicine 1
+▁pattern 1
+▁decade 1
+▁franz 1
+▁japanese 1
+▁inspir 1
+▁warning 1
+▁dedicated 1
+▁factor 1
+▁gro 1
+▁million 1
+▁noble 1
+▁islands 1
+▁artistic 1
+top 1
+▁gate 1
+▁oldes 1
+▁originate 1
+▁refer 1
+▁boss 1
+▁eating 1
+▁georg 1
+ball 1
+▁employees 1
+▁suffered 1
+▁kra 1
+▁handle 1
+▁expected 1
+▁contribut 1
+▁investigat 1
+▁girls 1
+▁colleague 1
+▁facilities 1
+▁stuttgart 1
+▁essential 1
+▁chief 1
+▁actress 1
+▁shut 1
+▁average 1
+▁bought 1
+▁creative 1
+▁guide 1
+▁tram 1
+ft 1
+▁map 1
+▁root 1
+▁determine 1
+▁pair 1
+▁ships 1
+▁prepared 1
+▁inform 1
+ome 1
+▁fra 1
+▁global 1
+▁destination 1
+▁residence 1
+▁substance 1
+▁chemical 1
+▁dispute 1
+▁pretty 1
+▁shirt 1
+▁cheap 1
+▁juice 1
+▁profit 1
+▁delete 1
+▁tie 1
+▁trace 1
+▁pop 1
+ei 1
+▁cast 1
+▁duke 1
+▁promoted 1
+▁rejected 1
+▁stepp 1
+▁seriously 1
+tsch 1
+▁growing 1
+▁join 1
+▁equipment 1
+▁everywhere 1
+▁satellite 1
+▁territory 1
+▁bronze 1
+▁trouble 1
+▁appeal 1
+▁stress 1
+▁reveal 1
+▁breath 1
+▁polish 1
+▁easier 1
+▁kids 1
+▁hang 1
+▁useful 1
+see 1
+▁bill 1
+▁succeeded 1
+har 1
+▁rome 1
+▁spelling 1
+▁borders 1
+▁bavaria 1
+▁careful 1
+▁browser 1
+▁depict 1
+▁nucle 1
+▁stadium 1
+▁threat 1
+▁governor 1
+▁blame 1
+▁electricity 1
+▁washing 1
+▁nord 1
+▁presentation 1
+don 1
+▁gained 1
+▁walking 1
+meter 1
+▁fri 1
+▁gray 1
+▁lecture 1
+▁mur 1
+ula 1
+▁talking 1
+▁appropriate 1
+▁joseph 1
+▁leipzig 1
+▁sponsor 1
+▁stupid 1
+▁baroque 1
+▁translated 1
+▁target 1
+▁democratic 1
+▁agreement 1
+▁raised 1
+▁assumed 1
+▁aid 1
+▁steel 1
+▁regarding 1
+▁hat 1
+▁opera 1
+j 1
+▁participated 1
+▁sw 1
+▁detail 1
+horn 1
+▁grows 1
+room 1
+ah 1
+▁bun 1
+▁remaining 1
+▁slow 1
+ane 1
+▁harm 1
+kin 1
+rac 1
+▁pipe 1
+▁suggest 1
+▁yo 1
+▁owned 1
+▁maintenance 1
+▁scream 1
+▁chinese 1
+▁chosen 1
+▁husband 1
+▁alarm 1
+▁photograph 1
+▁hello 1
+▁ram 1
+nny 1
+▁ser 1
+▁huge 1
+▁recommend 1
+▁kiss 1
+head 1
+▁condition 1
+▁sum 1
+▁bass 1
+▁entered 1
+▁benefit 1
+▁perspective 1
+▁smooth 1
+▁volunteer 1
+▁aspect 1
+▁strike 1
+▁planning 1
+▁internal 1
+▁dress 1
+▁gift 1
+▁devices 1
+▁strange 1
+▁literar 1
+▁tourism 1
+▁suddenly 1
+▁prime 1
+▁surprised 1
+▁extended 1
+▁founder 1
+▁explained 1
+▁complicated 1
+▁contemporary 1
+▁cycle 1
+▁freight 1
+▁tunnel 1
+▁climate 1
+▁truth 1
+▁soviet 1
+▁split 1
+▁sciences 1
+▁william 1
+▁core 1
+▁keyboard 1
+▁native 1
+▁sole 1
+▁praise 1
+▁desired 1
+▁switched 1
+▁alone 1
+▁poet 1
+▁destroy 1
+▁shell 1
+▁formation 1
+▁trainer 1
+▁exact 1
+slav 1
+▁except 1
+▁otto 1
+é 1
+▁breakfast 1
+▁context 1
+▁philosophy 1
+▁thomas 1
+▁knight 1
+▁folk 1
+▁roast 1
+▁alright 1
+▁environment 1
+▁analysis 1
+▁script 1
+▁expression 1
+▁earn 1
+▁authority 1
+▁multi 1
+▁net 1
+▁clo 1
+▁spoke 1
+ben 1
+▁parties 1
+▁spell 1
+▁farmer 1
+▁gene 1
+▁rear 1
+▁pal 1
+▁favor 1
+▁bottom 1
+▁california 1
+▁campaign 1
+▁victim 1
+▁valuable 1
+▁eventually 1
+▁exhaust 1
+▁drain 1
+▁drill 1
+▁dough 1
+▁tank 1
+▁grant 1
+▁wonder 1
+▁introduced 1
+▁prepare 1
+▁myth 1
+not 1
+▁kur 1
+▁vice 1
+yard 1
+▁settled 1
+▁emerg 1
+▁classical 1
+▁kick 1
+▁reduce 1
+▁rod 1
+▁bake 1
+board 1
+▁ur 1
+nja 1
+▁danger 1
+▁battery 1
+▁ludwig 1
+▁hockey 1
+▁unlike 1
+▁chicken 1
+▁error 1
+▁partially 1
+▁housing 1
+stone 1
+▁associated 1
+▁loose 1
+▁disc 1
+▁lying 1
+▁installed 1
+endorf 1
+▁expanded 1
+▁lin 1
+▁composed 1
+berries 1
+▁rise 1
+▁recordings 1
+▁mad 1
+▁till 1
+▁apply 1
+▁residents 1
+▁cri 1
+tag 1
+▁characters 1
+▁ms 1
+▁shooting 1
+▁corn 1
+▁citizen 1
+▁suffer 1
+▁representatives 1
+▁interview 1
+▁negative 1
+▁possibility 1
+▁sufficient 1
+▁impact 1
+▁reference 1
+▁hidden 1
+▁address 1
+▁audio 1
+▁christmas 1
+▁strict 1
+ola 1
+▁semi 1
+▁attached 1
+▁gre 1
+▁fried 1
+▁matches 1
+▁alive 1
+▁den 1
+▁begins 1
+▁rust 1
+▁canton 1
+gar 1
+▁ham 1
+▁drama 1
+▁discover 1
+▁channel 1
+▁somewhat 1
+▁stroke 1
+▁crazy 1
+▁giving 1
+▁diving 1
+▁dutch 1
+▁agency 1
+▁rural 1
+▁launch 1
+▁racing 1
+▁aircraft 1
+▁weapons 1
+▁finance 1
+▁valid 1
+▁approved 1
+▁statu 1
+▁egg 1
+▁witt 1
+bed 1
+▁online 1
+▁laid 1
+mail 1
+▁lind 1
+▁chess 1
+▁magic 1
+▁fully 1
+▁photos 1
+▁thinking 1
+▁eli 1
+ule 1
+▁occasion 1
+▁descend 1
+▁australia 1
+▁comparison 1
+▁heavily 1
+▁hungry 1
+▁locomotive 1
+▁possession 1
+▁potential 1
+▁responsibility 1
+▁revolution 1
+▁assistant 1
+▁defeat 1
+▁beyond 1
+▁related 1
+▁pilot 1
+▁felt 1
+▁failure 1
+▁unclear 1
+▁safety 1
+▁massive 1
+▁burned 1
+▁knee 1
+▁clothes 1
+▁helped 1
+▁india 1
+▁converted 1
+▁sent 1
+▁dig 1
+press 1
+mel 1
+▁rent 1
+▁vote 1
+▁catch 1
+▁critical 1
+▁cry 1
+▁details 1
+ella 1
+▁holidays 1
+▁soldiers 1
+▁micro 1
+▁southeast 1
+▁bud 1
+▁publish 1
+▁advice 1
+▁bundesliga 1
+▁challenge 1
+▁component 1
+▁emperor 1
+▁imperial 1
+▁oblig 1
+▁applies 1
+zzle 1
+▁gym 1
+▁forgotten 1
+▁advertisement 1
+▁broad 1
+▁stuck 1
+▁decide 1
+▁rubb 1
+▁burst 1
+▁carbon 1
+▁derived 1
+▁jean 1
+▁hart 1
+▁arrived 1
+tory 1
+▁hau 1
+here 1
+▁conductor 1
+▁swed 1
+▁participants 1
+▁sell 1
+▁maintain 1
+▁lion 1
+▁commander 1
+▁actual 1
+▁constitution 1
+▁participate 1
+ani 1
+enbach 1
+▁historic 1
+▁mel 1
+▁tape 1
+▁accepted 1
+▁mole 1
+▁holy 1
+▁academic 1
+▁occupied 1
+▁regulation 1
+▁review 1
+▁sweat 1
+▁resistance 1
+▁caught 1
+▁cheese 1
+▁transferred 1
+▁comfortable 1
+▁shift 1
+▁corps 1
+▁architecture 1
+▁predominantly 1
+▁swim 1
+▁polar 1
+etic 1
+▁none 1
+▁hof 1
+▁reduced 1
+ily 1
+▁donation 1
+ana 1
+▁indicate 1
+▁ruler 1
+hagen 1
+▁zi 1
+▁rum 1
+▁interesting 1
+▁exhibit 1
+▁grab 1
+▁relevant 1
+▁tissue 1
+▁crowd 1
+▁electronic 1
+▁remind 1
+▁legend 1
+▁transition 1
+▁murder 1
+▁proof 1
+▁shoes 1
+▁brain 1
+▁told 1
+▁cleaning 1
+▁nest 1
+zo 1
+▁colored 1
+▁engineering 1
+▁attend 1
+▁alt 1
+▁neo 1
+▁manag 1
+▁que 1
+east 1
+▁fundamental 1
+▁gabriel 1
+▁asteroid 1
+▁collapse 1
+▁families 1
+▁truck 1
+▁fourteen 1
+▁transfer 1
+▁relativ 1
+▁alex 1
+▁spark 1
+▁familiar 1
+▁mag 1
+▁priest 1
+▁radi 1
+▁wald 1
+▁socialist 1
+▁grass 1
+▁westphalia 1
+▁bavarian 1
+▁pitch 1
+▁stable 1
+▁herb 1
+▁trail 1
+▁entirely 1
+▁sick 1
+▁formula 1
+▁initiative 1
+▁journey 1
+▁progress 1
+▁silence 1
+▁variety 1
+▁precise 1
+▁april 1
+▁intended 1
+▁rhein 1
+▁witness 1
+▁adjust 1
+▁flood 1
+▁consume 1
+▁sail 1
+▁wal 1
+▁finger 1
+▁hide 1
+▁grain 1
+▁survive 1
+▁bau 1
+▁clock 1
+▁constructed 1
+war 1
+day 1
+hood 1
+▁disco 1
+▁correctly 1
+▁asia 1
+▁dub 1
+ische 1
+▁faster 1
+▁trial 1
+▁operate 1
+▁survived 1
+▁jam 1
+lingen 1
+▁interrupt 1
+▁chocolate 1
+▁grandfather 1
+▁ministry 1
+▁yeah 1
+▁choose 1
+▁screw 1
+▁quiet 1
+▁storage 1
+▁annual 1
+▁annoy 1
+▁repeat 1
+▁justice 1
+▁sunday 1
+▁brun 1
+▁dangerous 1
+▁publications 1
+▁underground 1
+▁successfully 1
+▁growth 1
+▁lamp 1
+▁branches 1
+▁anton 1
+▁rice 1
+▁worse 1
+▁reported 1
+▁signed 1
+▁accordingly 1
+▁george 1
+▁strength 1
+▁calendar 1
+▁dimension 1
+▁extension 1
+▁vegetable 1
+▁silent 1
+▁kingdom 1
+▁decline 1
+▁sauce 1
+▁moving 1
+▁instance 1
+▁physical 1
+▁tonight 1
+▁bicycle 1
+▁freiburg 1
+off 1
+▁accessible 1
+▁pond 1
+▁investment 1
+▁tight 1
+▁intention 1
+▁peak 1
+▁zu 1
+▁crew 1
+▁ob 1
+▁fear 1
+▁wie 1
+▁genre 1
+▁directed 1
+▁impression 1
+▁lang 1
+ably 1
+▁bonn 1
+acti 1
+▁publisher 1
+▁cow 1
+▁emerged 1
+▁ordered 1
+▁processes 1
+▁server 1
+▁cro 1
+ura 1
+▁cord 1
+of 1
+▁relation 1
+▁affair 1
+▁alcohol 1
+▁brigade 1
+▁confused 1
+▁conversation 1
+▁cooperation 1
+▁diocese 1
+▁recognition 1
+▁changing 1
+▁enemy 1
+▁translation 1
+▁remote 1
+▁sixteen 1
+▁manual 1
+haps 1
+▁fifth 1
+▁workers 1
+▁approximately 1
+▁magnetic 1
+▁contribution 1
+▁fever 1
+▁hop 1
+▁toilet 1
+▁powerful 1
+▁sar 1
+▁grey 1
+▁produce 1
+▁sheet 1
+▁spoken 1
+▁belly 1
+▁normally 1
+▁hack 1
+▁protected 1
+▁maintained 1
+▁separated 1
+▁achieve 1
+▁scha 1
+▁hesse 1
+back 1
+▁hub 1
+▁calculate 1
+▁controversial 1
+▁identity 1
+▁simultaneously 1
+▁drunk 1
+▁jazz 1
+▁christoph 1
+▁column 1
+▁properties 1
+▁wedding 1
+▁stood 1
+▁honest 1
+▁requirements 1
+▁wealth 1
+▁politically 1
+▁alongside 1
+rald 1
+▁assembly 1
+▁fred 1
+▁prussian 1
+▁processing 1
+▁strongly 1
+▁manner 1
+▁pastor 1
+wise 1
+▁bull 1
+ux 1
+▁const 1
+▁promote 1
+▁settle 1
+jo 1
+▁subsequent 1
+▁saxon 1
+▁hind 1
+▁differ 1
+▁episode 1
+▁resources 1
+▁dresden 1
+▁profile 1
+▁labor 1
+▁swimming 1
+▁tubingen 1
+▁klaus 1
+▁forgot 1
+▁israel 1
+▁richard 1
+▁compete 1
+▁bremen 1
+▁walter 1
+▁cancer 1
+▁lung 1
+▁din 1
+▁highway 1
+▁curve 1
+▁defined 1
+▁spain 1
+▁breed 1
+▁obtained 1
+asse 1
+▁chest 1
+▁mary 1
+haus 1
+▁corresponding 1
+mmy 1
+▁suspect 1
+▁assigned 1
+▁conclusion 1
+▁decisive 1
+▁technique 1
+▁amazing 1
+▁relax 1
+▁freedom 1
+▁inherit 1
+▁mold 1
+vin 1
+▁sacr 1
+▁experienced 1
+▁realized 1
+▁lens 1
+▁shower 1
+▁organic 1
+▁whom 1
+▁sieg 1
+▁lip 1
+▁northwest 1
+▁pai 1
+▁engineer 1
+▁russia 1
+▁wil 1
+ima 1
+▁cancel 1
+▁stove 1
+▁dude 1
+▁gall 1
+▁closely 1
+▁cam 1
+▁surviv 1
+▁virtual 1
+▁circula 1
+▁description 1
+▁distributed 1
+▁oxygen 1
+▁platform 1
+▁predecessor 1
+▁supplies 1
+▁czech 1
+▁meadow 1
+▁resign 1
+▁david 1
+▁sharp 1
+▁advise 1
+▁appointment 1
+▁barrel 1
+fall 1
+▁income 1
+▁relief 1
+▁schon 1
+▁loop 1
+▁adopted 1
+fish 1
+▁crack 1
+bling 1
+▁villa 1
+▁discussed 1
+▁electrical 1
+▁chip 1
+post 1
+ju 1
+▁meter 1
+kirch 1
+▁raw 1
+▁refu 1
+▁environmental 1
+▁spirit 1
+▁dusseldorf 1
+▁facility 1
+▁incorrect 1
+▁instruction 1
+▁reception 1
+▁rhythm 1
+▁zurich 1
+▁junior 1
+▁deserve 1
+▁faith 1
+▁unique 1
+▁trigger 1
+▁crystal 1
+▁ernst 1
+▁phase 1
+▁konig 1
+▁sky 1
+▁brandenburg 1
+▁pedal 1
+▁debt 1
+▁poetry 1
+▁painted 1
+▁exam 1
+▁bow 1
+▁existed 1
+▁craft 1
+▁acquired 1
+▁kre 1
+▁gross 1
+▁constitutional 1
+▁employee 1
+▁bite 1
+▁proven 1
+▁independently 1
+▁proper 1
+▁button 1
+rious 1
+▁stein 1
+▁advertising 1
+▁alexander 1
+▁combination 1
+▁gymnasium 1
+▁selected 1
+▁plastic 1
+▁fridge 1
+▁idiot 1
+▁gothic 1
+▁committed 1
+▁tube 1
+▁installation 1
+▁logic 1
+▁theatre 1
+▁latest 1
+▁pump 1
+tional 1
+▁fixed 1
+well 1
+▁carefully 1
+▁tall 1
+▁damaged 1
+▁african 1
+▁eaten 1
+▁pole 1
+▁marie 1
+▁cloth 1
+▁cent 1
+▁uncle 1
+▁pap 1
+▁abandoned 1
+▁diploma 1
+▁memorial 1
+▁sculpture 1
+▁temporarily 1
+▁client 1
+▁credit 1
+▁cinema 1
+▁snake 1
+▁harbor 1
+▁munster 1
+▁website 1
+▁dropped 1
+▁nazi 1
+▁reign 1
+▁apprenticeship 1
+▁opposition 1
+▁solid 1
+astic 1
+▁workshop 1
+▁guess 1
+▁wound 1
+▁fold 1
+▁subsidi 1
+ologist 1
+▁string 1
+▁stamp 1
+▁lady 1
+▁shall 1
+▁sweden 1
+▁slipp 1
+▁cash 1
+▁hunting 1
+▁glasses 1
+▁file 1
+▁rad 1
+▁height 1
+▁begin 1
+▁carl 1
+▁drawing 1
+▁sad 1
+▁concerned 1
+ground 1
+▁initial 1
+▁cheat 1
+▁recording 1
+▁dro 1
+▁nation 1
+▁japan 1
+▁ultimate 1
+▁discipline 1
+▁eagle 1
+▁estimate 1
+▁identical 1
+▁identified 1
+▁incident 1
+▁occupation 1
+▁philipp 1
+▁phrase 1
+▁presence 1
+▁scratch 1
+▁smartphone 1
+▁britain 1
+▁victory 1
+▁slope 1
+▁bodies 1
+▁flies 1
+▁monitor 1
+▁integrated 1
+▁devote 1
+▁municipalities 1
+▁potatoes 1
+▁lift 1
+▁interpretation 1
+▁poly 1
+▁fake 1
+▁valve 1
+▁mister 1
+▁fighting 1
+▁sex 1
+▁solve 1
+▁enable 1
+▁passion 1
+▁format 1
+▁ol 1
+▁frau 1
+▁mining 1
+▁existing 1
+▁abbreviation 1
+▁ceremony 1
+▁executed 1
+▁maximum 1
+▁qualified 1
+▁railroad 1
+▁wurttemberg 1
+▁asleep 1
+▁parallel 1
+▁freelance 1
+▁fortress 1
+▁siblings 1
+▁pirate 1
+▁steep 1
+▁comprise 1
+▁armed 1
+▁korea 1
+▁dramatic 1
+▁hal 1
+▁hence 1
+▁equation 1
+ight 1
+▁sie 1
+▁conventional 1
+▁delivery 1
+▁seal 1
+▁declared 1
+▁mai 1
+poli 1
+▁lib 1
+▁complet 1
+itive 1
+▁weigh 1
+▁dan 1
+▁toast 1
+▁bath 1
+▁classes 1
+▁drawn 1
+▁bush 1
+▁swing 1
+▁earl 1
+▁practiced 1
+▁improve 1
+▁score 1
+▁cave 1
+▁forced 1
+▁consumption 1
+▁habitat 1
+▁response 1
+▁trunk 1
+▁garbage 1
+▁senior 1
+▁portrait 1
+▁thread 1
+▁regiment 1
+▁mount 1
+platz 1
+▁secretary 1
+▁encouraged 1
+▁boil 1
+▁representation 1
+▁oven 1
+enheim 1
+▁albert 1
+▁shore 1
+▁angle 1
+▁distinguished 1
+▁johannes 1
+▁southwest 1
+▁secured 1
+▁kro 1
+▁confirmed 1
+▁boiler 1
+▁accused 1
+jan 1
+low 1
+▁agreed 1
+mund 1
+▁interpret 1
+▁rap 1
+▁leon 1
+▁demo 1
+▁evaluat 1
+▁argument 1
+▁engaged 1
+▁imagine 1
+▁knife 1
+▁medieval 1
+▁unusual 1
+▁torture 1
+▁gottingen 1
+▁baking 1
+▁plot 1
+▁debut 1
+▁engel 1
+▁gain 1
+▁slid 1
+▁bigger 1
+▁atmosphere 1
+▁combined 1
+▁missed 1
+▁soft 1
+▁deco 1
+▁observed 1
+ville 1
+▁cul 1
+hand 1
+▁mental 1
+▁arrive 1
+▁surprise 1
+▁northeast 1
+▁fift 1
+▁advisor 1
+▁efficiency 1
+▁excellen 1
+▁exposed 1
+▁significance 1
+▁thereafter 1
+▁disturb 1
+▁wagner 1
+▁promotion 1
+▁afternoon 1
+▁stefan 1
+▁tough 1
+▁dishes 1
+▁backwards 1
+▁farming 1
+▁minim 1
+▁frei 1
+▁reader 1
+▁controlled 1
+▁effective 1
+▁extra 1
+▁hip 1
+▁reache 1
+uous 1
+▁scored 1
+▁indicated 1
+win 1
+▁dur 1
+▁profession 1
+▁journal 1
+▁reject 1
+▁nie 1
+▁climbing 1
+▁ji 1
+rising 1
+friend 1
+▁wire 1
+▁mono 1
+▁proceed 1
+▁absorb 1
+necessarily 1
+▁accompanied 1
+▁augsburg 1
+▁mediterranean 1
+▁populated 1
+▁injuries 1
+▁therapy 1
+▁plum 1
+▁illness 1
+▁birthday 1
+box 1
+▁dumb 1
+▁considerable 1
+▁jack 1
+▁stir 1
+▁prohibited 1
+▁dense 1
+▁constantly 1
+▁separately 1
+▁mention 1
+▁advance 1
+▁collected 1
+ault 1
+▁surely 1
+▁tear 1
+▁privately 1
+▁aggressive 1
+▁educated 1
+▁infantry 1
+▁neglect 1
+▁nervous 1
+▁pleasure 1
+▁vacuum 1
+▁widespread 1
+▁illegal 1
+▁brazil 1
+▁retired 1
+▁funeral 1
+▁behave 1
+▁fiction 1
+▁tomato 1
+▁lyric 1
+▁eighteen 1
+▁mineral 1
+term 1
+▁deutsche 1
+▁aware 1
+▁rapid 1
+▁leadership 1
+▁slim 1
+▁crash 1
+▁olive 1
+▁altar 1
+▁worst 1
+▁grape 1
+▁grandma 1
+▁seventy 1
+▁reformation 1
+▁navy 1
+▁arrested 1
+▁rein 1
+▁properly 1
+▁lecturer 1
+▁graz 1
+kow 1
+▁becoming 1
+▁val 1
+▁ego 1
+▁successes 1
+▁bowl 1
+▁businesses 1
+hard 1
+▁construct 1
+▁dar 1
+▁extend 1
+▁fail 1
+▁bond 1
+water 1
+▁autonomous 1
+▁celebrated 1
+▁confess 1
+▁congress 1
+▁delicious 1
+▁executive 1
+▁october 1
+▁password 1
+▁qualification 1
+▁telephone 1
+▁ignore 1
+▁welcome 1
+▁curious 1
+▁hardware 1
+▁literally 1
+▁vacation 1
+▁width 1
+▁shield 1
+▁nineteen 1
+▁saving 1
+▁violin 1
+▁amateur 1
+▁hiking 1
+▁muller 1
+▁moderate 1
+▁mood 1
+▁borrow 1
+▁operator 1
+▁fault 1
+▁admitted 1
+▁archive 1
+▁radiation 1
+▁treatment 1
+▁socialism 1
+▁versa 1
+▁hunt 1
+coming 1
+▁math 1
+▁delivered 1
+▁vary 1
+▁zo 1
+▁tub 1
+▁loudly 1
+▁suffering 1
+▁healthy 1
+▁alb 1
+▁emil 1
+bahn 1
+▁fighter 1
+▁recover 1
+▁sam 1
+break 1
+right 1
+▁zoo 1
+▁parent 1
+▁fate 1
+gau 1
+▁whisper 1
+▁approval 1
+▁certificate 1
+▁communist 1
+▁convince 1
+▁exercise 1
+▁expansion 1
+▁gustav 1
+▁hannover 1
+▁merchant 1
+▁netherlands 1
+▁purchase 1
+▁rudolf 1
+▁afford 1
+▁external 1
+▁monday 1
+▁visual 1
+▁palace 1
+▁plug 1
+▁false 1
+▁physics 1
+▁awake 1
+▁hollow 1
+▁classification 1
+▁occurred 1
+▁observation 1
+▁venu 1
+▁blade 1
+▁plai 1
+▁prior 1
+▁gang 1
+▁invited 1
+bell 1
+kirche 1
+▁interpreted 1
+rank 1
+▁bark 1
+▁hermann 1
+▁moor 1
+bro 1
+▁dispos 1
+▁gap 1
+▁weapon 1
+▁protest 1
+▁acoustic 1
+▁centuries 1
+▁cigarette 1
+▁circumstances 1
+▁domestic 1
+▁enormous 1
+▁mirror 1
+▁resolution 1
+▁subordinate 1
+▁syndrome 1
+▁terror 1
+▁yourselves 1
+▁athlete 1
+▁grammar 1
+▁spaghetti 1
+▁uniform 1
+▁intensive 1
+▁muscle 1
+▁cruise 1
+▁motivat 1
+▁evil 1
+▁wagon 1
+▁exceed 1
+▁permitted 1
+▁input 1
+▁remarkable 1
+band 1
+▁specialized 1
+▁fox 1
+▁angeles 1
+▁passage 1
+▁expl 1
+▁ferry 1
+flow 1
+▁dirty 1
+▁generated 1
+▁criticized 1
+▁examined 1
+▁setting 1
+▁strengthen 1
+neck 1
+▁taxi 1
+▁breeding 1
+phon 1
+witz 1
+▁parliamentar 1
+▁secure 1
+bank 1
+▁alter 1
+▁advertise 1
+▁category 1
+▁colloquial 1
+▁concentrate 1
+▁conservative 1
+▁cylinder 1
+▁decrease 1
+▁desperate 1
+▁elaborate 1
+▁employment 1
+▁encounter 1
+▁evangelical 1
+▁francis 1
+▁frequency 1
+▁gradually 1
+▁intelligence 1
+▁thieves 1
+▁voltage 1
+▁excited 1
+▁servant 1
+▁jurgen 1
+▁patron 1
+▁quant 1
+▁beauty 1
+▁firm 1
+▁pacific 1
+▁antenna 1
+▁nerve 1
+▁ensure 1
+▁elector 1
+▁burial 1
+▁slic 1
+▁preparation 1
+▁coins 1
+▁james 1
+▁shade 1
+▁kiel 1
+▁spur 1
+▁pose 1
+▁researchers 1
+▁rebel 1
+▁mast 1
+▁lent 1
+▁laz 1
+▁traumat 1
+▁angel 1
+▁buch 1
+▁observ 1
+▁establish 1
+lude 1
+▁arab 1
+▁equat 1
+▁impress 1
+▁consequent 1
+▁embarrass 1
+▁comprehensive 1
+▁deployed 1
+▁joachim 1
+▁politician 1
+▁possibly 1
+▁rebuilt 1
+▁scatter 1
+▁selection 1
+▁wolfgang 1
+▁abbey 1
+▁ceiling 1
+▁variant 1
+▁zealand 1
+▁liberal 1
+▁entry 1
+▁mouse 1
+▁golf 1
+▁happiness 1
+▁banned 1
+▁shaft 1
+▁tiny 1
+▁drag 1
+▁establishment 1
+▁ginger 1
+▁aachen 1
+▁grandpa 1
+▁charts 1
+▁manufactured 1
+▁rumor 1
+charging 1
+▁sixth 1
+▁calm 1
+▁belt 1
+▁implemented 1
+▁melt 1
+▁santa 1
+▁pf 1
+▁hunter 1
+▁shake 1
+▁affect 1
+▁sale 1
+enstein 1
+play 1
+▁mile 1
+pleasant 1
+▁aluminum 1
+▁anniversary 1
+▁concentration 1
+▁demonstrate 1
+▁ingredient 1
+▁psychology 1
+▁respond 1
+▁sacrific 1
+▁synonym 1
+▁thirteen 1
+▁fragment 1
+▁indicator 1
+▁advocate 1
+▁imagination 1
+▁shadow 1
+▁indication 1
+▁scholar 1
+▁pupil 1
+▁sword 1
+▁stink 1
+▁weimar 1
+▁crisis 1
+▁delay 1
+▁dried 1
+▁rescue 1
+▁crook 1
+▁index 1
+▁invented 1
+▁assess 1
+▁radical 1
+▁nearby 1
+▁hype 1
+▁rush 1
+▁investigation 1
+▁rocket 1
+bruck 1
+▁armor 1
+▁poland 1
+▁continuously 1
+▁allerg 1
+▁proportion 1
+▁recovery 1
+▁refused 1
+▁boxes 1
+▁trailer 1
+▁seeds 1
+▁dust 1
+lich 1
+▁graf 1
+▁lamb 1
+▁wu 1
+▁ost 1
+▁compos 1
+▁import 1
+▁centimeters 1
+▁circuit 1
+▁criteria 1
+▁danube 1
+▁deputy 1
+▁employer 1
+▁facade 1
+▁infection 1
+▁inhabited 1
+▁navigation 1
+▁portuguese 1
+▁substitute 1
+▁draft 1
+▁pursued 1
+▁recipe 1
+▁tyrol 1
+schaft 1
+▁cabinet 1
+▁century 1
+▁worldwide 1
+▁mobile 1
+▁existence 1
+▁pianist 1
+▁crucial 1
+▁kreuz 1
+▁harvest 1
+▁resort 1
+▁extent 1
+▁copies 1
+▁altitude 1
+▁crawl 1
+▁werner 1
+▁comic 1
+▁trend 1
+▁dozen 1
+▁adventure 1
+▁courtyard 1
+▁glad 1
+▁pepper 1
+▁vocal 1
+upon 1
+▁giant 1
+▁trash 1
+▁duty 1
+rose 1
+▁laser 1
+▁mason 1
+▁hugo 1
+▁doctorate 1
+▁roughly 1
+▁ripe 1
+▁arc 1
+▁alps 1
+meister 1
+▁tick 1
+▁bench 1
+lena 1
+▁brave 1
+▁accord 1
+▁continent 1
+iness 1
+▁extract 1
+▁dirt 1
+hop 1
+▁commit 1
+▁democracy 1
+▁dissolved 1
+▁eliminated 1
+▁enterprise 1
+▁jurisdiction 1
+▁konrad 1
+▁luxembourg 1
+▁mathematics 1
+▁mexico 1
+▁motorcycle 1
+▁preliminar 1
+▁salzburg 1
+▁september 1
+▁thoroughly 1
+▁compound 1
+▁explanation 1
+▁fashion 1
+▁somebody 1
+▁disgust 1
+▁throne 1
+▁frog 1
+▁injury 1
+▁sparse 1
+▁honey 1
+▁ninth 1
+▁carpet 1
+▁shipping 1
+▁barrier 1
+▁reverse 1
+▁canada 1
+▁tennis 1
+▁presum 1
+▁consul 1
+▁consol 1
+▁garage 1
+▁fluid 1
+▁reconstruction 1
+▁dragon 1
+▁arabic 1
+▁highlight 1
+▁charter 1
+▁gender 1
+▁airline 1
+▁romantic 1
+▁oak 1
+oxide 1
+▁runner 1
+maker 1
+▁sommer 1
+gard 1
+▁solved 1
+▁rub 1
+▁nich 1
+▁baron 1
+▁pope 1
+▁frequent 1
+point 1
+▁continu 1
+▁basketball 1
+▁capacity 1
+▁execution 1
+▁gymnastics 1
+▁illustration 1
+▁mecklenburg 1
+▁possibilities 1
+▁pseudonym 1
+▁reservoir 1
+▁scissors 1
+▁skeleton 1
+▁wednesday 1
+▁assistance 1
+▁diverse 1
+▁throat 1
+▁turkish 1
+▁combat 1
+▁ghost 1
+▁stimula 1
+▁chancellor 1
+▁airplane 1
+▁suburb 1
+▁gallery 1
+▁motif 1
+▁density 1
+▁senate 1
+▁denied 1
+▁clinic 1
+▁dependent 1
+▁confusion 1
+▁fellow 1
+▁extraordinary 1
+▁determin 1
+▁salad 1
+▁hammer 1
+▁tribe 1
+▁nave 1
+▁basement 1
+▁popularity 1
+▁attributed 1
+ladder 1
+▁findings 1
+plan 1
+▁palm 1
+▁holiday 1
+▁advanced 1
+▁soldier 1
+▁decid 1
+▁horn 1
+mination 1
+▁rie 1
+▁mechanical 1
+▁scal 1
+▁obvious 1
+▁digit 1
+stream 1
+▁advent 1
+▁bankrupt 1
+▁dynasty 1
+▁efficient 1
+▁goalkeeper 1
+▁pronounce 1
+▁supplement 1
+▁suspension 1
+▁thuringia 1
+▁comparable 1
+▁legislati 1
+▁chapter 1
+▁segment 1
+▁parcel 1
+▁architectural 1
+▁cattle 1
+▁repeatedly 1
+▁grim 1
+▁programming 1
+▁psycho 1
+▁vision 1
+▁orbit 1
+▁distinction 1
+▁baltic 1
+▁grun 1
+▁lutheran 1
+▁twin 1
+▁voluntarily 1
+▁fiber 1
+▁hohen 1
+▁suggestion 1
+▁decorated 1
+▁moral 1
+▁luther 1
+▁payment 1
+▁correction 1
+▁forestry 1
+▁accuse 1
+▁nail 1
+▁peel 1
+▁improved 1
+▁chile 1
+▁nau 1
+▁divid 1
+▁gravel 1
+mpression 1
+▁invite 1
+field 1
+▁magnet 1
+▁hydro 1
+▁introduct 1
+▁territori 1
+▁demolition 1
+▁elisabeth 1
+▁facebook 1
+▁interaction 1
+▁jersey 1
+▁karlsruhe 1
+▁moscow 1
+▁opportunities 1
+▁opportunity 1
+▁peninsula 1
+▁resemble 1
+▁sequence 1
+▁symptoms 1
+▁triangle 1
+▁unconscious 1
+▁voluntary 1
+▁regulate 1
+▁suicide 1
+▁underneath 1
+▁adolf 1
+▁islamic 1
+▁penalty 1
+▁rabbi 1
+▁temple 1
+▁frost 1
+▁summit 1
+▁pioneer 1
+▁spoon 1
+▁reliable 1
+▁canadian 1
+▁breast 1
+▁grill 1
+▁truly 1
+▁bundestag 1
+▁smart 1
+▁expectations 1
+▁ireland 1
+▁intense 1
+▁sample 1
+▁consistent 1
+▁bounc 1
+▁arrangement 1
+▁sustainable 1
+▁klein 1
+▁brut 1
+▁genetic 1
+▁pause 1
+▁glow 1
+▁collective 1
+▁ridge 1
+▁adj 1
+▁elder 1
+▁noticeable 1
+▁insert 1
+▁questionable 1
+▁update 1
+▁seventh 1
+▁float 1
+▁marine 1
+▁linear 1
+▁tang 1
+▁ferr 1
+▁cage 1
+▁pine 1
+▁scr 1
+▁continuous 1
+▁georgia 1
+craft 1
+▁rough 1
+▁mala 1
+▁accurate 1
+▁anonymous 1
+▁blossom 1
+▁cultivated 1
+▁definition 1
+▁dortmund 1
+▁grandparents 1
+▁intelligent 1
+▁norwegian 1
+▁november 1
+▁schmidt 1
+▁schneider 1
+▁signature 1
+▁supervisor 1
+▁neutral 1
+▁spiritual 1
+▁thumb 1
+▁alfred 1
+▁josef 1
+▁lubeck 1
+▁analyze 1
+▁louis 1
+▁vector 1
+▁dancing 1
+▁historian 1
+▁goethe 1
+▁weird 1
+▁particle 1
+▁rotation 1
+▁duct 1
+▁pink 1
+▁mask 1
+▁alternat 1
+violet 1
+▁sheep 1
+▁tension 1
+▁hook 1
+▁drone 1
+▁muhl 1
+▁items 1
+▁consideration 1
+▁cab 1
+▁losses 1
+▁june 1
+schedule 1
+▁grade 1
+scu 1
+▁deliver 1
+▁implement 1
+▁succeed 1
+▁disappear 1
+▁compare 1
+space 1
+▁encrypt 1
+▁abolish 1
+▁adolescent 1
+▁conservation 1
+▁helicopter 1
+▁jakob 1
+▁manufacturing 1
+▁mathematical 1
+▁nonsense 1
+▁revenue 1
+▁structural 1
+▁transmission 1
+▁violence 1
+▁swabia 1
+▁unfair 1
+▁withdraw 1
+▁curtain 1
+▁variable 1
+▁budget 1
+▁retreat 1
+▁pocket 1
+▁capsule 1
+▁optical 1
+▁hunger 1
+▁retain 1
+▁bundes 1
+▁lunch 1
+▁mystery 1
+▁pretend 1
+▁humor 1
+▁jelly 1
+▁linux 1
+▁loyal 1
+▁thermal 1
+▁referee 1
+▁preferred 1
+▁fritz 1
+▁linguistic 1
+▁merit 1
+▁harmless 1
+▁janni 1
+▁skat 1
+▁encryption 1
+▁volcano 1
+▁lemon 1
+▁marburg 1
+ji 1
+▁fries 1
+▁bibl 1
+stedt 1
+▁kant 1
+▁bang 1
+▁seiz 1
+leben 1
+▁replac 1
+▁fog 1
+uction 1
+▁measur 1
+born 1
+▁categories 1
+▁divorce 1
+▁forbidden 1
+▁fracture 1
+▁infrastructure 1
+▁lieutenant 1
+▁potsdam 1
+▁quote 1
+▁regret 1
+▁sneeze 1
+▁varieties 1
+▁hobby 1
+▁prevail 1
+▁slept 1
+▁dessert 1
+▁influential 1
+▁plaster 1
+▁suprem 1
+▁gesture 1
+▁random 1
+▁sexual 1
+▁daniel 1
+▁disabled 1
+▁alliance 1
+▁domain 1
+▁bypass 1
+▁turtle 1
+▁crypt 1
+▁denmark 1
+▁poisonous 1
+▁dialogue 1
+▁mild 1
+▁toxic 1
+▁seminar 1
+▁revis 1
+▁priority 1
+▁rattle 1
+▁deadline 1
+▁plateau 1
+▁ulm 1
+▁systematic 1
+▁flesh 1
+▁arise 1
+▁fool 1
+▁fame 1
+▁upright 1
+▁govern 1
+▁delight 1
+▁worm 1
+▁gaze 1
+alogue 1
+▁tide 1
+schule 1
+▁sang 1
+holz 1
+▁obtain 1
+▁hike 1
+▁examine 1
+wood 1
+power 1
+▁yu 1
+tally 1
+stand 1
+pole 1
+▁assumption 1
+▁belgium 1
+▁chemistry 1
+▁demolished 1
+▁elevator 1
+▁explosive 1
+▁february 1
+▁furniture 1
+▁guarantee 1
+▁heidelberg 1
+▁illustrate 1
+▁initiated 1
+▁manuscript 1
+▁negotiations 1
+▁renowned 1
+▁retirement 1
+▁schleswig 1
+▁struggle 1
+▁suspended 1
+▁temporary 1
+▁thursday 1
+▁vegetarian 1
+▁displac 1
+▁exciting 1
+▁intervention 1
+▁pizza 1
+▁download 1
+▁schiller 1
+▁discrete 1
+▁einstein 1
+▁google 1
+▁mixture 1
+▁abroad 1
+▁clown 1
+▁renew 1
+▁cherry 1
+▁clerk 1
+▁stolen 1
+▁milan 1
+▁blank 1
+▁edible 1
+▁friday 1
+▁pinch 1
+▁virus 1
+▁reduction 1
+▁mandate 1
+▁airfield 1
+▁basin 1
+▁burden 1
+▁egyptian 1
+▁mamm 1
+▁ownership 1
+▁friendship 1
+▁kohl 1
+▁warned 1
+▁heinz 1
+▁aunt 1
+▁mortal 1
+▁cookies 1
+▁invest 1
+▁sink 1
+▁erna 1
+▁chef 1
+nkel 1
+cast 1
+▁chop 1
+▁prepar 1
+▁curse 1
+ngest 1
+fusion 1
+▁lend 1
+▁fix 1
+▁generate 1
+▁sudden 1
+▁derive 1
+course 1
+▁egypt 1
+▁abstract 1
+▁algorithm 1
+▁amazon 1
+▁announcement 1
+▁curiosity 1
+▁excluded 1
+▁ferdinand 1
+▁grammatical 1
+▁interface 1
+▁metropoli 1
+▁narrator 1
+▁psychological 1
+▁sensitive 1
+▁florida 1
+▁helmet 1
+▁surrender 1
+▁margaret 1
+▁possesse 1
+▁prague 1
+▁translator 1
+▁fossil 1
+▁swamp 1
+▁lorenz 1
+▁subtitl 1
+▁widow 1
+▁greeting 1
+▁pillar 1
+▁portray 1
+▁terribl 1
+▁chalk 1
+▁donkey 1
+▁angry 1
+▁carrot 1
+▁queue 1
+▁irish 1
+▁trench 1
+▁shark 1
+▁export 1
+▁assert 1
+▁biological 1
+▁attractive 1
+▁salv 1
+▁wagg 1
+▁documentary 1
+▁genus 1
+gitte 1
+▁moni 1
+▁collector 1
+▁pace 1
+pell 1
+grid 1
+▁designat 1
+▁syl 1
+stairs 1
+▁deposit 1
+▁immediate 1
+▁refuse 1
+▁volcan 1
+▁squa 1
+▁accompany 1
+▁bachelor 1
+▁overwhelm 1
+▁symphon 1
+satisfied 1
+▁assignment 1
+▁columbia 1
+▁conscience 1
+▁engagement 1
+▁ensemble 1
+▁franconia 1
+▁hometown 1
+▁punished 1
+▁societies 1
+▁suspicious 1
+▁swallow 1
+▁wurzburg 1
+▁dwarf 1
+▁emotional 1
+▁fabric 1
+▁complement 1
+▁diamond 1
+▁matrix 1
+▁landlord 1
+▁turkey 1
+▁melody 1
+▁ultra 1
+▁leather 1
+▁heritage 1
+▁fifteen 1
+▁kaiser 1
+▁ladies 1
+▁personnel 1
+▁scrap 1
+▁bathroom 1
+▁strawberr 1
+▁quantities 1
+▁vertical 1
+▁soap 1
+▁banana 1
+▁resistant 1
+▁stalk 1
+▁insult 1
+▁tractor 1
+▁integration 1
+▁beforehand 1
+▁steak 1
+▁classified 1
+▁gunther 1
+licate 1
+▁insight 1
+▁evaluation 1
+▁implie 1
+▁stability 1
+enthal 1
+finite 1
+weiler 1
+▁symbolize 1
+▁mechanic 1
+zig 1
+▁consequence 1
+▁styl 1
+▁baker 1
+▁confirm 1
+▁declar 1
+▁expand 1
+▁poison 1
+▁physi 1
+▁distract 1
+▁tuesday 1
+▁acknowledge 1
+▁bayreuth 1
+▁calculation 1
+▁destruction 1
+▁fireplace 1
+▁horror 1
+▁independence 1
+▁membrane 1
+▁monarch 1
+▁orthodox 1
+▁pharmaceutic 1
+▁renovation 1
+▁spectator 1
+▁virgin 1
+▁contrary 1
+▁medication 1
+▁mozilla 1
+▁powder 1
+▁schwarz 1
+▁secular 1
+▁terrace 1
+nforcement 1
+▁miracle 1
+▁theoretical 1
+▁pencil 1
+▁feather 1
+▁cardinal 1
+▁premier 1
+▁spider 1
+▁saturday 1
+▁silesia 1
+▁stylistic 1
+▁contest 1
+▁trading 1
+▁essen 1
+▁holstein 1
+▁participation 1
+▁jacket 1
+▁automatic 1
+▁rival 1
+▁java 1
+▁simon 1
+▁clip 1
+▁eich 1
+▁teach 1
+▁chant 1
+▁relie 1
+▁hatch 1
+hipped 1
+cute 1
+▁neuro 1
+▁criticize 1
+itude 1
+acent 1
+▁integr 1
+▁accent 1
+▁gentle 1
+▁immun 1
+▁contradict 1
+▁principal 1
+▁argentina 1
+▁confident 1
+▁convincing 1
+▁erfurt 1
+▁evolution 1
+▁extinct 1
+▁grandmother 1
+▁impatient 1
+▁libraries 1
+▁majesty 1
+▁perception 1
+▁recreation 1
+▁relegation 1
+▁reputation 1
+▁scandal 1
+▁sebastian 1
+▁striking 1
+▁supplier 1
+▁terrain 1
+▁underestimate 1
+▁wikipedia 1
+▁colonial 1
+▁compensation 1
+▁cough 1
+▁innovation 1
+▁moderator 1
+▁observatory 1
+▁separation 1
+▁disorder 1
+▁revenge 1
+▁sneak 1
+▁perceive 1
+▁injured 1
+▁optim 1
+▁superior 1
+▁estonia 1
+▁shaking 1
+▁achievements 1
+▁frisian 1
+▁bernhard 1
+▁betray 1
+▁serbia 1
+▁purple 1
+▁axis 1
+▁kramer 1
+▁collaboration 1
+▁bohm 1
+▁textile 1
+▁athens 1
+▁nieder 1
+▁implementation 1
+▁massage 1
+▁specifically 1
+▁objective 1
+▁exile 1
+▁schult 1
+▁sailor 1
+position 1
+roscop 1
+hawk 1
+▁combine 1
+▁apparent 1
+worth 1
+florescence 1
+▁captured 1
+▁chimney 1
+▁compromise 1
+▁corporation 1
+▁corridor 1
+▁creativity 1
+▁disagree 1
+▁disappointed 1
+▁fertile 1
+▁fulfill 1
+▁hollywood 1
+▁hungarian 1
+▁nephew 1
+▁nickname 1
+▁renaissance 1
+▁reproduce 1
+▁restrictions 1
+▁slalom 1
+▁surprising 1
+▁uncertain 1
+▁defect 1
+▁enemies 1
+▁peasant 1
+▁spontaneous 1
+▁squadron 1
+▁tributar 1
+▁volks 1
+▁creek 1
+▁slogan 1
+▁module 1
+▁robber 1
+▁greece 1
+▁smith 1
+▁wreck 1
+▁retail 1
+▁inscription 1
+▁drift 1
+▁bulb 1
+▁couch 1
+▁duchy 1
+▁jacob 1
+▁plent 1
+▁witch 1
+▁alleged 1
+▁oscar 1
+▁competent 1
+▁scottish 1
+▁neighborhood 1
+▁tuning 1
+▁shy 1
+▁jones 1
+▁sought 1
+▁danish 1
+▁interval 1
+▁norway 1
+▁dominated 1
+▁ottoman 1
+▁weiss 1
+poo 1
+▁custom 1
+▁giga 1
+trogen 1
+abea 1
+▁prussia 1
+junction 1
+▁alphabet 1
+▁breakthrough 1
+▁chronicle 1
+▁confidence 1
+▁corruption 1
+▁fertilize 1
+▁frequencies 1
+▁identify 1
+▁irritate 1
+▁phenomenon 1
+▁provision 1
+▁renovated 1
+▁reservation 1
+▁screwdriver 1
+▁template 1
+▁unbelievable 1
+▁wiesbaden 1
+▁anywhere 1
+▁balcony 1
+▁muslim 1
+▁poverty 1
+▁habits 1
+▁laundry 1
+▁midfield 1
+▁portugal 1
+▁replacement 1
+▁nominated 1
+▁sensation 1
+▁fulda 1
+▁schwerin 1
+▁kassel 1
+▁colony 1
+▁outdoor 1
+▁attitude 1
+▁gerhard 1
+▁junk 1
+▁departure 1
+ybody 1
+▁moped 1
+statt 1
+▁jackson 1
+▁assessment 1
+▁vault 1
+▁partnership 1
+▁cannon 1
+▁upset 1
+▁font 1
+▁productive 1
+issa 1
+schau 1
+media 1
+▁chase 1
+rough 1
+▁discuss 1
+▁prohibit 1
+▁distinguish 1
+▁kilometer 1
+hoff 1
+▁techno 1
+▁define 1
+▁nav 1
+final 1
+▁reconstruct 1
+went 1
+▁attenti 1
+▁telescop 1
+▁wrestling 1
+▁archaeological 1
+▁champagne 1
+▁competitor 1
+▁conservatory 1
+▁cyclist 1
+▁darmstadt 1
+▁disadvantage 1
+▁eduard 1
+▁greifswald 1
+▁nuremberg 1
+▁occurrence 1
+▁pedestrian 1
+▁pomerania 1
+▁prosecutor 1
+▁qualify 1
+▁recognizable 1
+▁staircase 1
+▁supplied 1
+▁athletic 1
+▁bohemia 1
+▁circum 1
+▁conscious 1
+▁elegant 1
+▁helmut 1
+▁penguin 1
+▁wrapp 1
+▁database 1
+▁entertain 1
+▁scope 1
+▁cherries 1
+▁coalition 1
+▁balloon 1
+▁glove 1
+▁babies 1
+▁schnitz 1
+▁hanover 1
+▁victoria 1
+▁endanger 1
+▁spray 1
+▁senator 1
+▁croatia 1
+▁lodge 1
+▁motto 1
+▁convey 1
+▁refine 1
+▁restored 1
+▁offence 1
+▁michel 1
+▁fence 1
+▁casual 1
+▁advers 1
+▁appeti 1
+▁creme 1
+▁diesel 1
+▁moths 1
+▁trainee 1
+▁slate 1
+▁charles 1
+▁feud 1
+▁beef 1
+▁automat 1
+▁resist 1
+▁adam 1
+gonal 1
+▁peaceful 1
+▁commen 1
+patch 1
+▁analog 1
+borg 1
+▁wedg 1
+▁preserve 1
+fels 1
+founder 1
+bridge 1
+fortunately 1
+▁communicat 1
+▁carpenter 1
+▁dissertation 1
+▁ignorance 1
+▁mattress 1
+▁mexican 1
+▁packaging 1
+▁plaintiff 1
+▁psychiatr 1
+▁scotland 1
+▁sovereign 1
+▁statistics 1
+▁submarines 1
+▁upgrade 1
+▁utensils 1
+▁fantasy 1
+▁mozart 1
+▁sandwich 1
+▁unsuccessful 1
+▁typing 1
+▁decoration 1
+▁civic 1
+▁sausage 1
+▁oneself 1
+▁simson 1
+▁courage 1
+▁julius 1
+▁modified 1
+▁unlock 1
+▁twist 1
+▁monster 1
+▁atomic 1
+▁horst 1
+▁owl 1
+▁tackle 1
+▁motives 1
+▁relay 1
+▁edward 1
+▁motion 1
+▁pilgrimage 1
+▁debate 1
+▁manuel 1
+▁preced 1
+▁bernard 1
+▁inventor 1
+▁sabin 1
+▁horizontal 1
+▁faci 1
+holder 1
+▁jorg 1
+▁whip 1
+nathan 1
+▁attach 1
+▁attribute 1
+script 1
+fried 1
+burn 1
+▁potato 1
+▁depart 1
+▁oppos 1
+▁immigrat 1
+deficiency 1
+responsibilities 1
+▁albrecht 1
+▁bielefeld 1
+▁brilliant 1
+▁bulgaria 1
+▁cannibal 1
+▁commemorat 1
+▁competitive 1
+▁december 1
+▁endeavor 1
+▁essays 1
+▁explosion 1
+▁fountain 1
+▁infected 1
+▁jesuit 1
+▁lifestyle 1
+▁limestone 1
+▁oxford 1
+▁pediatric 1
+▁peripher 1
+▁thailand 1
+▁bacteria 1
+▁chicago 1
+▁headphones 1
+▁hungary 1
+▁rostock 1
+▁arnold 1
+▁commerce 1
+▁memories 1
+▁thirst 1
+▁colonel 1
+▁disaster 1
+▁garlic 1
+▁swear 1
+▁clever 1
+▁expertise 1
+▁outstanding 1
+▁survey 1
+▁sensor 1
+▁whale 1
+▁accusations 1
+▁spiral 1
+trustworthy 1
+▁patent 1
+▁vienne 1
+▁satur 1
+▁relocat 1
+▁shelter 1
+▁explod 1
+▁considerably 1
+▁distinctive 1
+▁jesus 1
+▁gauge 1
+ogram 1
+▁charlie 1
+▁proceedings 1
+▁eisen 1
+▁abuse 1
+▁wonderful 1
+ubb 1
+▁chris 1
+▁remov 1
+fumes 1
+▁arrest 1
+▁arrange 1
+▁assume 1
+▁headquarter 1
+tractable 1
+▁archbishop 1
+▁attorney 1
+▁batteries 1
+▁breslau 1
+▁comparative 1
+▁confront 1
+▁cultivation 1
+▁deceased 1
+▁diagnosis 1
+▁emphasize 1
+▁evident 1
+▁impossible 1
+▁palatinate 1
+▁proposal 1
+▁prototype 1
+▁punishment 1
+▁rupture 1
+▁scandinavia 1
+▁sculptor 1
+▁technician 1
+▁teenager 1
+▁timbered 1
+▁transformation 1
+▁transmitted 1
+▁violations 1
+▁virtue 1
+▁carinthia 1
+▁synchroni 1
+▁treasure 1
+▁vineyard 1
+▁privatiz 1
+▁spacious 1
+▁binding 1
+▁ritual 1
+▁manfred 1
+▁urgent 1
+▁judicial 1
+▁serving 1
+▁pigeon 1
+▁compact 1
+▁hurry 1
+▁bleach 1
+▁escalat 1
+currency 1
+▁syria 1
+▁smile 1
+▁henrik 1
+▁theodor 1
+▁tempt 1
+▁feedback 1
+▁gentlemen 1
+▁activat 1
+▁semiconductor 1
+▁fortification 1
+▁framework 1
+▁winning 1
+▁hydrogen 1
+▁erwin 1
+▁tibet 1
+▁plus 1
+▁exert 1
+▁differentiate 1
+▁concei 1
+▁celt 1
+▁pont 1
+undula 1
+lateral 1
+▁raise 1
+woman 1
+▁absence 1
+▁accumulat 1
+▁administered 1
+▁anonymity 1
+▁arrondissement 1
+▁braunschweig 1
+▁constituency 1
+▁diabetes 1
+▁discontinued 1
+▁discretion 1
+▁diversity 1
+▁enthusiastic 1
+▁explicitly 1
+▁financing 1
+▁hemisphere 1
+▁innocent 1
+▁intellectual 1
+▁irregular 1
+▁leopold 1
+▁mandatory 1
+▁maneuver 1
+▁mercedes 1
+▁permission 1
+▁postponed 1
+▁receipt 1
+▁refrain 1
+▁regensburg 1
+▁repertoire 1
+▁sweep 1
+▁transaction 1
+▁twitch 1
+▁vietnam 1
+▁inevitabl 1
+▁nikolaus 1
+▁renamed 1
+▁semester 1
+▁solidarity 1
+▁verdict 1
+▁chords 1
+▁rainbow 1
+▁nicola 1
+▁sociology 1
+▁ethnic 1
+▁cartoon 1
+▁replica 1
+▁tropical 1
+▁rigid 1
+▁magnus 1
+▁pliers 1
+▁reddish 1
+▁tactic 1
+▁margin 1
+▁robin 1
+▁incorporated 1
+▁reinhard 1
+▁stamens 1
+▁invention 1
+▁stopping 1
+▁interloc 1
+▁silk 1
+▁clemen 1
+▁niko 1
+▁spice 1
+▁deceiv 1
+motive 1
+▁alpin 1
+▁nitr 1
+western 1
+spiel 1
+moller 1
+▁install 1
+▁convention 1
+▁fluctuat 1
+▁acquaint 1
+▁appreciate 1
+▁arrival 1
+▁artillery 1
+▁carbohydrate 1
+▁categoriz 1
+▁celebration 1
+▁coincidence 1
+▁comrade 1
+▁crocodile 1
+▁defensive 1
+▁depriv 1
+▁diagram 1
+▁inflammation 1
+▁interference 1
+▁intersection 1
+▁migration 1
+▁millennium 1
+▁philosophic 1
+▁pregnant 1
+▁prehistoric 1
+▁prescription 1
+▁preservation 1
+▁prominent 1
+▁propaganda 1
+▁rehearsal 1
+▁smoking 1
+▁stabilize 1
+▁substantial 1
+▁suspicion 1
+▁ulrich 1
+▁universal 1
+▁unnoticed 1
+▁venice 1
+▁yield 1
+▁laboratory 1
+▁surname 1
+▁utilized 1
+▁valentin 1
+▁victories 1
+▁federation 1
+▁samuel 1
+▁thigh 1
+▁charlotte 1
+▁clarified 1
+▁texas 1
+▁thief 1
+▁livestock 1
+▁sprint 1
+▁guinea 1
+▁latvia 1
+▁cliff 1
+▁premise 1
+▁toddler 1
+▁nightmare 1
+▁leash 1
+▁session 1
+▁cheerful 1
+▁overcome 1
+▁sirens 1
+▁cement 1
+▁withdrew 1
+▁tasty 1
+▁forgive 1
+▁scrub 1
+▁squir 1
+▁trauma 1
+▁elect 1
+▁introduce 1
+▁easie 1
+▁permit 1
+▁aqua 1
+▁electromagnet 1
+▁reflex 1
+▁accommodation 1
+▁acquisition 1
+▁amsterdam 1
+▁applicable 1
+▁applicant 1
+▁boycott 1
+▁cloakroom 1
+▁demonstration 1
+▁electrified 1
+▁emphasis 1
+▁enthusiasm 1
+▁expense 1
+▁faculty 1
+▁figurative 1
+▁flirt 1
+▁geometric 1
+▁imbalance 1
+▁immigrant 1
+▁inconsistencies 1
+▁instructor 1
+▁katharina 1
+▁koblenz 1
+▁magdeburg 1
+▁mechanism 1
+▁mercury 1
+▁microwave 1
+▁monkey 1
+▁pharmacist 1
+▁relegated 1
+▁rendsburg 1
+▁ribbon 1
+▁simulate 1
+▁symmetric 1
+▁syrup 1
+▁vicinity 1
+▁affiliat 1
+▁autumn 1
+▁restricted 1
+▁scanner 1
+▁sigmund 1
+▁exploit 1
+▁surgeon 1
+▁colonies 1
+▁depression 1
+▁bloom 1
+▁indonesia 1
+▁bargain 1
+▁minuet 1
+▁trophy 1
+▁competence 1
+▁satire 1
+▁tamara 1
+▁thrill 1
+▁generous 1
+▁spoil 1
+▁bureau 1
+▁caesar 1
+▁madrid 1
+▁nicole 1
+▁probation 1
+▁styria 1
+▁rascal 1
+▁argue 1
+▁blond 1
+▁palestin 1
+▁quin 1
+▁plasm 1
+▁vogt 1
+▁beetle 1
+ophyt 1
+alysis 1
+▁remu 1
+▁snor 1
+ocated 1
+eastern 1
+▁invasi 1
+friedhof 1
+▁presid 1
+fluent 1
+piece 1
+heart 1
+▁horizon 1
+▁democrat 1
+▁pilgrim 1
+▁apprentice 1
+▁napoleon 1
+▁sympath 1
+▁abrupt 1
+▁accelerated 1
+▁antarctic 1
+▁applause 1
+▁astronomer 1
+▁barracks 1
+▁byzantine 1
+▁celebrities 1
+▁combustion 1
+▁conversion 1
+▁crucifix 1
+▁cuddle 1
+▁dinosaur 1
+▁eberhard 1
+▁entrepreneur 1
+▁envelop 1
+▁expedition 1
+▁indirect 1
+▁industries 1
+▁intermediate 1
+▁kennedy 1
+▁migrant 1
+▁neolithic 1
+▁outskirts 1
+▁patience 1
+▁provincial 1
+▁registration 1
+▁speculation 1
+▁synagogue 1
+▁textbook 1
+▁unhealthy 1
+▁unpopular 1
+▁vegetation 1
+▁violent 1
+▁viticulture 1
+▁wrinkl 1
+▁botanical 1
+▁hymn 1
+▁marathon 1
+▁physician 1
+▁hypo 1
+▁inflation 1
+▁schroder 1
+▁alsace 1
+▁knives 1
+▁brush 1
+▁lehmann 1
+▁sniff 1
+estimation 1
+▁backpack 1
+▁chestnut 1
+▁bleeding 1
+▁decent 1
+▁yolk 1
+▁bochum 1
+▁boredom 1
+▁froze 1
+▁uncover 1
+▁warehouse 1
+▁protein 1
+▁dentist 1
+▁spores 1
+▁parrot 1
+▁fischer 1
+▁seventeen 1
+▁depth 1
+▁petersburg 1
+▁output 1
+▁detach 1
+▁recommendations 1
+▁inverse 1
+▁falk 1
+cursor 1
+▁drip 1
+ipping 1
+rpor 1
+foot 1
+scribed 1
+▁sustain 1
+í 1
+▁exempt 1
+▁stubborn 1
+▁ambassador 1
+▁archdiocese 1
+▁aristocrat 1
+▁astronomy 1
+▁barricade 1
+▁breisgau 1
+▁captivity 1
+▁claustrophobi 1
+▁compulsory 1
+▁deliberat 1
+▁dismissed 1
+▁elephant 1
+▁heilbronn 1
+▁hydraulic 1
+▁innocence 1
+▁intercept 1
+▁istanbul 1
+▁mosquitoes 1
+▁nationwide 1
+▁phenomena 1
+▁photosynthesis 1
+▁prerequisite 1
+▁protagonist 1
+▁reinforced 1
+▁reluctant 1
+▁rhetoric 1
+▁slovakia 1
+▁strategy 1
+▁supermarket 1
+▁supervision 1
+▁buffet 1
+▁cabaret 1
+▁triumph 1
+▁asylum 1
+▁conquest 1
+▁january 1
+▁offensive 1
+▁rugby 1
+▁nectar 1
+▁sepals 1
+▁counterpart 1
+▁farewell 1
+▁incredibl 1
+▁pretense 1
+▁tariff 1
+▁romanesque 1
+▁barbara 1
+▁falcon 1
+▁soluble 1
+▁balkan 1
+▁glott 1
+▁restoration 1
+▁granite 1
+▁flavor 1
+believing 1
+▁reunion 1
+▁allotment 1
+▁sicily 1
+▁intact 1
+▁vegan 1
+▁abdom 1
+▁discriminat 1
+▁wolfsburg 1
+ccup 1
+▁scheme 1
+▁agnes 1
+▁cousin 1
+▁charming 1
+▁newcomer 1
+▁costume 1
+▁headlights 1
+▁unfit 1
+▁terminal 1
+▁deviate 1
+▁emulate 1
+deutsch 1
+▁bertol 1
+tropic 1
+mposed 1
+goblin 1
+imaginable 1
+▁devil 1
+believers 1
+meyer 1
+▁slip 1
+▁acquir 1
+▁adopt 1
+▁convert 1
+iddle 1
+▁comfort 1
+▁happi 1
+▁absurd 1
+▁accommodate 1
+▁adequate 1
+▁aesthetic 1
+▁afghanistan 1
+▁amphibi 1
+▁antiquity 1
+▁avalanche 1
+▁barbecue 1
+▁cavalry 1
+▁censorship 1
+▁clutch 1
+▁cockpit 1
+▁deactivate 1
+▁devastat 1
+▁dungeon 1
+▁eclipse 1
+▁enlightenment 1
+▁explanatory 1
+▁fragrance 1
+▁horrible 1
+▁imprisoned 1
+▁imprisonment 1
+▁indigenous 1
+▁instinct 1
+▁jerusalem 1
+▁matthias 1
+▁ornamental 1
+▁python 1
+▁rectangular 1
+▁recurr 1
+▁schizophreni 1
+▁scooter 1
+▁slaughter 1
+▁sporadically 1
+▁strategic 1
+▁unemployed 1
+▁vorpommern 1
+▁wisdom 1
+▁brackets 1
+▁grenade 1
+▁household 1
+▁reunification 1
+▁statistical 1
+▁suffocat 1
+▁almonds 1
+▁beatrix 1
+▁occupy 1
+▁sediment 1
+▁vitamin 1
+▁wreath 1
+▁sketch 1
+▁xaver 1
+▁provoke 1
+▁charcoal 1
+▁okres 1
+▁pathogen 1
+▁quarry 1
+▁plaque 1
+▁cloister 1
+▁multipli 1
+▁chassis 1
+▁premium 1
+▁relieved 1
+▁vague 1
+▁pylon 1
+▁trivial 1
+▁emblem 1
+▁fisherman 1
+▁sahara 1
+▁jerry 1
+▁platt 1
+▁ventilat 1
+▁rifle 1
+▁choler 1
+▁gregor 1
+▁abundan 1
+▁nurse 1
+▁duties 1
+▁warrior 1
+▁supervise 1
+▁frederik 1
+▁undertake 1
+▁merger 1
+▁slam 1
+▁martyr 1
+▁uneven 1
+▁guilt 1
+▁wobbl 1
+▁lichten 1
+ointment 1
+▁glue 1
+▁ammoni 1
+▁regin 1
+▁westphal 1
+woke 1
+profit 1
+▁distinct 1
+corporate 1
+ó 1
+▁administrator 1
+▁astronaut 1
+▁bamboo 1
+▁commuters 1
+▁cuisine 1
+▁cycling 1
+▁decreasing 1
+▁destiny 1
+▁detroit 1
+▁elevation 1
+▁elongated 1
+▁fascinat 1
+▁firefighters 1
+▁gigabyte 1
+▁godfather 1
+▁hereditary 1
+▁hurricane 1
+▁illusion 1
+▁kaiserstuhl 1
+▁ketchup 1
+▁maximilian 1
+▁minnesota 1
+▁monasteries 1
+▁nurnberg 1
+▁panorama 1
+▁peanuts 1
+▁platinum 1
+▁prefecture 1
+▁prejudices 1
+▁prosecution 1
+▁prosperity 1
+▁reproduction 1
+▁saarbrucken 1
+▁spectrum 1
+▁strategies 1
+▁subspecies 1
+▁undamaged 1
+ncreas 1
+▁creep 1
+▁florence 1
+▁pajama 1
+▁refract 1
+▁promising 1
+▁boulder 1
+▁illuminat 1
+▁landshut 1
+▁asphalt 1
+▁expelled 1
+▁gottfried 1
+▁healthcare 1
+▁transformer 1
+▁baptize 1
+▁bettina 1
+▁ancestr 1
+▁burgenland 1
+▁fauna 1
+▁ironically 1
+▁arthur 1
+▁ghana 1
+▁lounge 1
+▁schloss 1
+▁overseas 1
+▁dairy 1
+▁ignite 1
+▁kauf 1
+▁mutual 1
+fascist 1
+▁predict 1
+▁butterfl 1
+▁altona 1
+▁fortune 1
+▁cease 1
+▁slum 1
+▁flank 1
+▁pollute 1
+▁scam 1
+▁britta 1
+▁lexic 1
+▁blend 1
+▁audibl 1
+▁accus 1
+ctically 1
+choke 1
+▁decorate 1
+▁regime 1
+▁magnify 1
+▁abstain 1
+▁bizarre 1
+▁homosexual 1
+isexual 1
+philharmoni 1
+▁accuracy 1
+▁achilles 1
+▁agencies 1
+▁anatomical 1
+▁barcelona 1
+▁buddhist 1
+▁cliché 1
+▁confederation 1
+▁cooperative 1
+▁cosmonaut 1
+▁culprit 1
+▁denounce 1
+▁dictator 1
+▁dictionary 1
+▁earthquake 1
+▁educator 1
+▁hildesheim 1
+▁hybrid 1
+▁hydrochlori 1
+▁injustice 1
+▁insufficient 1
+▁keyword 1
+▁manhattan 1
+▁mysterious 1
+▁narrative 1
+▁navigable 1
+▁ordinary 1
+▁philosopher 1
+▁protocol 1
+▁punctuat 1
+▁recyclable 1
+▁referendum 1
+▁remscheid 1
+▁repetition 1
+▁scenario 1
+▁schreiber 1
+▁subscribe 1
+▁sulfur 1
+▁twitter 1
+▁elsewhere 1
+▁embryo 1
+▁nikolai 1
+▁utilities 1
+▁cornelia 1
+▁quantify 1
+▁reliably 1
+▁ridicule 1
+▁vessel 1
+▁absent 1
+▁imperative 1
+▁lithium 1
+▁submerge 1
+▁folklore 1
+▁leisure 1
+▁messenger 1
+▁opaque 1
+▁adjoin 1
+▁gasoline 1
+▁volker 1
+▁freak 1
+▁logistic 1
+▁taylor 1
+▁volley 1
+▁globe 1
+▁provocati 1
+▁sultan 1
+▁residu 1
+▁delegate 1
+▁gloat 1
+▁anhalt 1
+▁remix 1
+▁ecological 1
+▁upload 1
+selotte 1
+▁cliche 1
+▁underpass 1
+▁fierce 1
+▁reimann 1
+▁aerob 1
+▁verona 1
+blazon 1
+▁swap 1
+▁desire 1
+▁whil 1
+aught 1
+▁scott 1
+▁eponym 1
+▁pyramid 1
+comprehensible 1
+habilitation 1
+á 1
+▁acceleration 1
+▁accomplished 1
+▁admiral 1
+▁admission 1
+▁ambulance 1
+▁amplifie 1
+▁antelope 1
+▁anxiety 1
+▁aschaffenburg 1
+▁backdrop 1
+▁basilica 1
+▁brussels 1
+▁cinnamon 1
+▁cognitive 1
+▁conspicuous 1
+▁crescent 1
+▁crusade 1
+▁czechoslovakia 1
+▁delicacy 1
+▁deployment 1
+▁dismissal 1
+▁disregarded 1
+▁doctrine 1
+▁episcop 1
+▁exaggerate 1
+▁excursion 1
+▁hooligans 1
+▁immense 1
+▁infrequent 1
+▁larvae 1
+▁memorize 1
+▁mongolia 1
+▁mortgage 1
+▁olivia 1
+▁osnabruck 1
+▁overview 1
+▁posthumous 1
+▁prognos 1
+▁proximity 1
+▁quartet 1
+▁recursive 1
+▁reminiscen 1
+▁rhyme 1
+▁singapore 1
+▁sophisticated 1
+▁strasbourg 1
+▁stumbled 1
+▁thunderstorm 1
+▁tremendous 1
+▁undecided 1
+▁venezuela 1
+▁yugoslavia 1
+▁belgian 1
+▁constitute 1
+▁counsel 1
+▁deteriorat 1
+▁oblast 1
+▁patricia 1
+▁tendency 1
+▁wolves 1
+▁bikini 1
+▁carolina 1
+▁probability 1
+▁soundtrack 1
+▁intestine 1
+▁moisture 1
+▁offenders 1
+▁sniper 1
+▁utility 1
+▁introver 1
+â 1
+ʿ 1
+ə 1
+ú 1
+̇ 1
+ñ 1
+q 1
+ð 1
+ā 1
+ū 1
diff --git a/configs/de-en/tgt_unigram6000/spm_unigram_de.vocab b/configs/de-en/tgt_unigram6000/spm_unigram_de.vocab
new file mode 100644
index 0000000000000000000000000000000000000000..1d3914dbc37fbfd3796ef4695d1f862fc5469589
--- /dev/null
+++ b/configs/de-en/tgt_unigram6000/spm_unigram_de.vocab
@@ -0,0 +1,6000 @@
+ 0
+ 0
+ 0
+ 0
+▁the -2.7193
+s -3.12575
+▁a -3.83084
+▁of -3.86707
+▁in -3.97477
+▁to -4.04507
+▁is -4.0957
+▁and -4.36613
+t -4.76263
+▁was -4.77948
+' -4.789
+▁it -4.83293
+▁he -4.85124
+ed -4.89749
+▁be -4.93374
+▁you -4.94344
+▁for -4.98581
+▁are -5.02188
+ing -5.04093
+n -5.14027
+▁that -5.1721
+▁on -5.19672
+▁with -5.20348
+▁this -5.27671
+▁i -5.29256
+d -5.29352
+▁his -5.39929
+▁as -5.4009
+▁not -5.47158
+ly -5.4948
+y -5.53844
+er -5.5735
+▁can -5.57648
+e -5.59657
+▁from -5.61655
+▁at -5.61971
+▁have -5.67588
+en -5.73884
+▁there -5.74416
+▁also -5.74803
+▁ -5.78583
+▁an -5.79045
+▁by -5.8114
+▁has -5.8836
+▁one -5.88917
+al -5.96556
+▁do -5.98381
+▁she -5.9975
+r -6.06171
+▁were -6.07138
+▁we -6.09636
+m -6.18858
+▁time -6.21557
+▁but -6.22422
+▁all -6.23518
+▁they -6.27823
+▁her -6.31001
+a -6.33187
+▁only -6.3344
+▁or -6.3555
+▁no -6.37293
+▁after -6.37605
+▁will -6.39714
+▁what -6.39728
+o -6.42727
+p -6.44125
+es -6.44476
+▁two -6.45379
+▁me -6.46032
+re -6.48073
+c -6.48079
+▁up -6.48527
+▁out -6.50109
+▁so -6.51722
+▁first -6.51741
+▁like -6.53471
+▁had -6.56486
+▁if -6.57092
+▁your -6.5808
+▁more -6.64127
+▁does -6.64557
+l -6.64767
+▁other -6.65786
+▁my -6.6947
+▁into -6.73458
+▁how -6.75169
+▁when -6.75207
+▁new -6.75357
+▁now -6.77921
+▁just -6.77927
+ic -6.78852
+ll -6.79818
+▁about -6.80162
+▁should -6.80445
+▁their -6.81338
+▁over -6.81911
+▁here -6.82766
+in -6.84089
+▁would -6.84091
+▁well -6.85706
+g -6.85725
+▁who -6.86594
+▁get -6.86772
+ion -6.89071
+i -6.90187
+▁don -6.90187
+▁some -6.90364
+▁still -6.90669
+b -6.91505
+▁did -6.93603
+u -6.94503
+le -6.95048
+▁people -6.95374
+▁which -6.95704
+▁him -6.95808
+k -6.96199
+f -6.96438
+▁used -6.9705
+▁many -6.98285
+▁however -6.99083
+▁re -6.99998
+▁go -7.00095
+▁could -7.00103
+▁un -7.0162
+▁then -7.02262
+ation -7.03
+▁again -7.03658
+▁during -7.03865
+▁part -7.05254
+▁most -7.05309
+▁place -7.07448
+or -7.0903
+▁these -7.10714
+▁even -7.11582
+▁work -7.12522
+▁s -7.13218
+el -7.13725
+▁today -7.13972
+▁already -7.15558
+▁three -7.15558
+te -7.15593
+▁way -7.15848
+us -7.17068
+▁always -7.17269
+la -7.18004
+ri -7.18021
+▁same -7.18708
+▁through -7.1875
+▁city -7.19125
+▁name -7.19135
+▁long -7.2064
+▁later -7.20863
+▁german -7.22119
+ve -7.22495
+▁right -7.22596
+▁than -7.22683
+ch -7.23499
+▁year -7.23587
+▁de -7.23946
+▁war -7.24357
+▁back -7.24634
+▁e -7.25524
+▁want -7.25627
+z -7.26244
+▁school -7.26287
+▁world -7.26918
+▁end -7.27286
+▁take -7.27912
+▁years -7.28877
+ck -7.29933
+an -7.30844
+h -7.3162
+li -7.33164
+▁why -7.33944
+cause -7.34005
+▁where -7.3476
+▁old -7.35066
+ra -7.3671
+▁made -7.37175
+▁under -7.37467
+▁too -7.38371
+ers -7.38428
+▁al -7.38707
+▁know -7.38861
+able -7.39157
+ro -7.3988
+▁make -7.40089
+▁water -7.4018
+▁c -7.40618
+ce -7.40745
+▁good -7.4098
+▁look -7.41286
+▁area -7.41384
+ar -7.41753
+▁without -7.41885
+▁every -7.42421
+▁co -7.43261
+▁its -7.43293
+▁please -7.43829
+- -7.45123
+▁off -7.45262
+w -7.45459
+on -7.45733
+▁see -7.46621
+th -7.46821
+▁use -7.475
+▁o -7.47696
+▁any -7.4785
+▁became -7.47986
+ter -7.48296
+▁often -7.49114
+▁last -7.49138
+▁different -7.50054
+▁g -7.50225
+ur -7.50456
+▁day -7.51175
+▁both -7.51495
+▁located -7.52322
+▁four -7.52628
+very -7.53269
+▁car -7.53517
+▁b -7.54092
+▁think -7.54281
+▁before -7.54918
+▁sever -7.55256
+▁let -7.55582
+▁ver -7.55676
+▁them -7.56599
+▁am -7.56697
+▁between -7.57343
+▁second -7.57679
+▁such -7.58351
+ian -7.58509
+it -7.58815
+▁much -7.5886
+de -7.59202
+▁need -7.60011
+▁district -7.60204
+▁own -7.60769
+ne -7.60986
+▁put -7.6146
+x -7.62075
+il -7.62747
+▁family -7.6302
+is -7.63354
+▁side -7.64968
+est -7.65081
+▁another -7.65917
+▁high -7.66262
+▁bo -7.67066
+▁left -7.68355
+▁state -7.68536
+ia -7.68823
+▁lot -7.68872
+se -7.69297
+▁found -7.69358
+▁called -7.69384
+▁important -7.69592
+▁p -7.69623
+▁something -7.6973
+▁children -7.70291
+▁town -7.70299
+ble -7.71873
+ma -7.72193
+▁church -7.72402
+▁our -7.72808
+▁life -7.7283
+▁t -7.73025
+ies -7.73032
+▁down -7.73407
+▁main -7.73675
+lo -7.73786
+▁light -7.7389
+▁small -7.73892
+▁five -7.74456
+ist -7.75075
+▁free -7.75151
+▁hand -7.76262
+▁really -7.76676
+ment -7.76866
+▁being -7.7714
+▁each -7.77247
+▁next -7.7737
+▁point -7.77775
+▁must -7.78606
+▁since -7.78736
+▁took -7.78839
+▁st -7.78964
+▁further -7.79036
+▁around -7.79337
+▁man -7.79591
+▁university -7.79949
+ta -7.80641
+ry -7.80722
+▁building -7.80764
+ity -7.8111
+til -7.81152
+▁addition -7.81225
+▁come -7.81748
+▁cannot -7.82173
+▁won -7.82889
+▁known -7.82909
+▁against -7.83154
+▁possible -7.83231
+▁got -7.83517
+na -7.85872
+ul -7.85966
+▁few -7.8615
+um -7.86285
+▁worked -7.87043
+▁d -7.87125
+▁company -7.87283
+▁due -7.87391
+ting -7.87563
+▁head -7.87802
+▁team -7.88135
+▁son -7.88224
+▁give -7.88303
+▁us -7.88383
+▁never -7.88626
+▁mo -7.89136
+▁national -7.90005
+▁con -7.90377
+ally -7.90397
+ge -7.90463
+▁course -7.90496
+ol -7.90579
+▁music -7.90869
+▁great -7.90996
+▁little -7.91507
+▁mar -7.92315
+▁going -7.93841
+land -7.94233
+▁village -7.94489
+▁ro -7.94565
+▁la -7.94677
+▁station -7.94769
+▁lead -7.95104
+▁home -7.95168
+▁public -7.95373
+▁north -7.95697
+▁film -7.95772
+▁play -7.95798
+▁open -7.95836
+▁belong -7.95917
+ru -7.95965
+▁away -7.96039
+▁v -7.96309
+▁vo -7.96439
+▁k -7.96498
+▁best -7.96562
+▁help -7.96625
+▁short -7.96775
+▁works -7.96995
+▁ten -7.97523
+▁far -7.97597
+▁came -7.97928
+ine -7.98057
+▁li -7.98083
+▁together -7.98105
+▁himself -7.98289
+▁system -7.98823
+refore -7.9888
+most -7.98914
+▁air -7.9921
+▁member -7.99385
+ate -7.99561
+▁everything -7.99591
+ted -7.99679
+ke -7.9973
+▁large -7.99889
+▁south -8.00113
+co -8.00221
+▁germany -8.00415
+man -8.00745
+▁center -8.0096
+me -8.01265
+▁di -8.01636
+▁better -8.01681
+▁among -8.02054
+ary -8.02103
+et -8.02131
+▁find -8.02308
+ive -8.02599
+▁considered -8.03383
+▁group -8.03573
+▁law -8.03658
+▁bad -8.03683
+ut -8.03938
+▁usual -8.04586
+▁form -8.04719
+ti -8.04738
+go -8.04861
+▁le -8.05014
+▁office -8.05108
+▁case -8.05274
+▁house -8.05298
+ant -8.05326
+ler -8.05927
+mp -8.0593
+▁lo -8.06187
+▁played -8.07802
+▁result -8.07946
+▁ha -8.08185
+at -8.0842
+vi -8.08949
+▁father -8.0901
+▁times -8.09387
+to -8.09395
+▁everyone -8.09636
+▁berlin -8.10038
+▁turn -8.10116
+age -8.10145
+▁show -8.10194
+▁nothing -8.10249
+▁once -8.10521
+▁pre -8.10891
+▁red -8.11188
+ca -8.11319
+ig -8.11615
+▁order -8.11703
+▁ra -8.11732
+ther -8.1176
+▁f -8.11821
+ie -8.12245
+id -8.12393
+ner -8.12452
+▁west -8.12491
+▁pro -8.12554
+da -8.127
+▁power -8.13614
+v -8.13739
+less -8.13834
+▁room -8.14357
+ci -8.14635
+▁bar -8.14644
+▁built -8.1471
+▁number -8.14769
+▁enough -8.15146
+au -8.1571
+▁sch -8.1585
+ous -8.1617
+▁close -8.16198
+▁front -8.16245
+▁per -8.1631
+▁set -8.16318
+▁bit -8.16782
+▁become -8.16913
+ton -8.17058
+▁live -8.17548
+▁land -8.17775
+▁may -8.17804
+▁sa -8.18066
+▁construction -8.18265
+ish -8.18827
+▁real -8.19077
+tic -8.19609
+ling -8.19637
+▁fire -8.19747
+▁black -8.20091
+ated -8.20139
+▁studied -8.20534
+▁service -8.20625
+▁local -8.21005
+▁line -8.21051
+▁keep -8.21229
+ism -8.21476
+▁married -8.21695
+▁middle -8.21695
+▁completely -8.22057
+bi -8.22107
+▁named -8.2215
+▁behind -8.22162
+▁role -8.22171
+▁sometimes -8.22399
+▁season -8.22414
+ff -8.22848
+▁someone -8.22902
+▁soon -8.22974
+mi -8.22994
+as -8.23209
+▁stop -8.23516
+▁party -8.23626
+▁club -8.23817
+▁things -8.24247
+▁yet -8.24308
+▁ma -8.24629
+▁ta -8.24673
+ial -8.25084
+mann -8.2516
+▁po -8.25346
+▁band -8.2553
+am -8.25678
+▁following -8.25694
+▁white -8.2623
+berg -8.26414
+▁call -8.26501
+▁term -8.26719
+▁say -8.26826
+sh -8.26834
+▁means -8.27018
+▁age -8.27161
+▁stay -8.27246
+▁en -8.27395
+▁six -8.27429
+be -8.27474
+▁history -8.27788
+▁mainly -8.27918
+▁castle -8.2798
+ness -8.28074
+▁longer -8.28316
+▁ga -8.28656
+▁round -8.28979
+▁ti -8.29008
+▁anything -8.29457
+▁love -8.29488
+ga -8.29684
+▁went -8.29786
+▁pa -8.29837
+▁half -8.30237
+▁region -8.30421
+▁association -8.30725
+ten -8.31191
+▁language -8.3129
+▁park -8.31343
+▁period -8.31494
+▁art -8.31513
+▁probably -8.31752
+▁wall -8.31962
+▁community -8.32529
+▁special -8.32729
+▁eight -8.32798
+▁river -8.32806
+▁ho -8.32993
+▁u -8.33071
+▁big -8.33161
+ber -8.33291
+▁example -8.33313
+▁position -8.3334
+▁numerous -8.33575
+▁started -8.33739
+▁hu -8.33831
+▁mi -8.33967
+▁lu -8.33972
+▁quickly -8.34213
+▁change -8.34324
+▁others -8.34361
+ir -8.34422
+ni -8.34527
+▁whole -8.34579
+rs -8.34628
+▁international -8.34632
+▁run -8.34659
+▁according -8.34821
+▁instead -8.34898
+▁lives -8.3499
+▁former -8.35039
+▁street -8.35432
+▁afterwards -8.35703
+▁death -8.3572
+▁train -8.35757
+▁east -8.3576
+bo -8.35843
+▁mr -8.35904
+▁yourself -8.35969
+▁try -8.36322
+▁comes -8.36388
+▁hundred -8.36508
+pe -8.3675
+di -8.37246
+▁else -8.37646
+▁food -8.37871
+▁mu -8.38003
+▁received -8.38077
+▁court -8.38142
+▁federal -8.38421
+▁various -8.38424
+▁actually -8.385
+ha -8.38543
+▁read -8.38575
+▁lies -8.38618
+▁ex -8.3922
+▁test -8.39319
+▁vi -8.39396
+ot -8.39573
+son -8.39857
+▁based -8.40126
+▁ju -8.40205
+▁those -8.40659
+ance -8.41145
+ure -8.41471
+bu -8.41549
+ad -8.41671
+▁career -8.41858
+▁hard -8.4251
+ty -8.42708
+gg -8.43092
+ec -8.43156
+ka -8.43281
+un -8.43508
+▁clear -8.43683
+▁bi -8.43712
+▁seem -8.43774
+▁early -8.43799
+▁common -8.43805
+▁post -8.43911
+▁board -8.43915
+▁dis -8.44049
+▁research -8.44307
+ical -8.44451
+▁pi -8.44532
+pp -8.44642
+▁project -8.44682
+▁na -8.44705
+▁ka -8.44815
+▁thing -8.44942
+▁start -8.45125
+▁sp -8.45163
+▁followed -8.45297
+▁general -8.46047
+▁money -8.46162
+▁capital -8.4646
+tz -8.46678
+▁reason -8.46684
+▁english -8.46759
+▁da -8.46785
+▁wants -8.46791
+ph -8.46793
+▁process -8.46832
+▁market -8.4709
+hi -8.4747
+ver -8.47536
+▁va -8.47836
+▁list -8.47909
+▁production -8.48001
+▁country -8.4801
+▁support -8.48125
+▁w -8.48128
+▁ge -8.48135
+▁ba -8.48233
+▁night -8.48271
+▁individual -8.48575
+▁french -8.48881
+▁believe -8.48963
+▁oh -8.4907
+om -8.49272
+▁class -8.49288
+▁title -8.49807
+▁while -8.49821
+▁road -8.50152
+▁game -8.50156
+mb -8.50586
+▁remained -8.50667
+▁em -8.50846
+▁lower -8.5105
+sch -8.51221
+▁espe -8.51369
+▁body -8.51401
+▁human -8.51401
+▁color -8.5142
+▁hall -8.51516
+▁te -8.51584
+▁wood -8.51712
+▁forest -8.52092
+▁r -8.52603
+▁young -8.52604
+▁said -8.52667
+cy -8.52675
+▁lost -8.52684
+▁model -8.52746
+▁past -8.53112
+▁person -8.5318
+que -8.53254
+▁transport -8.53266
+▁bring -8.53299
+do -8.53413
+cially -8.53445
+▁field -8.53478
+op -8.53582
+ster -8.53613
+mo -8.53706
+gen -8.54157
+▁brother -8.54239
+▁near -8.5426
+▁working -8.54301
+▁seen -8.54312
+▁business -8.54376
+▁remove -8.54444
+ho -8.54483
+▁active -8.5458
+▁space -8.54621
+▁data -8.54638
+▁job -8.54641
+▁h -8.54711
+▁hold -8.54846
+▁child -8.54861
+▁top -8.54911
+▁mostly -8.55092
+▁program -8.552
+▁military -8.55202
+▁level -8.55202
+▁allowed -8.55213
+▁sta -8.55326
+▁book -8.55455
+▁development -8.55552
+▁cross -8.5557
+▁mother -8.55586
+▁least -8.55859
+hr -8.55864
+▁initially -8.55932
+▁care -8.56016
+▁author -8.56134
+no -8.56394
+▁se -8.56405
+▁seven -8.5649
+▁talk -8.56861
+▁cause -8.57338
+und -8.57403
+▁pe -8.57482
+▁voice -8.57519
+▁consist -8.57528
+ized -8.57573
+▁ever -8.57717
+▁education -8.58182
+▁official -8.58192
+▁thought -8.58425
+▁strong -8.58444
+▁railway -8.58524
+▁natural -8.58534
+▁drive -8.5854
+▁currently -8.58561
+▁dog -8.58586
+▁makes -8.58787
+▁ice -8.58807
+▁seat -8.58839
+▁interest -8.58919
+▁beginning -8.59196
+▁matter -8.59225
+▁present -8.59373
+▁finally -8.59584
+▁single -8.59932
+▁word -8.60211
+▁purpose -8.6022
+ger -8.60435
+▁difficult -8.60563
+▁route -8.60566
+▁government -8.60568
+lin -8.60793
+▁similar -8.60908
+▁teacher -8.60995
+▁available -8.61254
+▁ri -8.61394
+▁europe -8.61544
+▁subject -8.61601
+▁cut -8.61697
+ard -8.61729
+▁sun -8.61782
+les -8.6179
+▁fall -8.61946
+▁league -8.6195
+▁series -8.61973
+▁du -8.62191
+▁european -8.62367
+▁bank -8.62776
+▁central -8.63003
+▁ko -8.63283
+▁blue -8.63635
+▁break -8.63776
+▁lake -8.63811
+▁species -8.64066
+▁kind -8.64097
+▁mean -8.64186
+▁ni -8.64229
+▁follow -8.64464
+▁tell -8.64552
+▁plant -8.64563
+tor -8.64806
+▁pass -8.64907
+▁men -8.64916
+ba -8.64928
+▁american -8.65082
+▁green -8.6529
+▁born -8.65393
+he -8.65428
+▁social -8.6548
+▁ground -8.65574
+▁sea -8.65876
+▁died -8.65892
+▁yes -8.65962
+▁view -8.66229
+▁tri -8.66259
+▁leave -8.66272
+▁hair -8.66272
+▁island -8.66337
+▁act -8.66449
+▁jo -8.66529
+▁nine -8.66668
+▁door -8.6674
+▁current -8.66914
+led -8.67007
+▁twenty -8.67327
+▁thus -8.67401
+os -8.67432
+▁connection -8.67744
+▁developed -8.678
+▁upper -8.67834
+▁sound -8.67971
+▁council -8.68067
+we -8.68164
+▁song -8.68166
+▁wanted -8.68208
+▁depend -8.68436
+va -8.68456
+▁roman -8.68486
+ology -8.68491
+▁learn -8.68809
+min -8.68878
+▁tower -8.69227
+▁fact -8.69391
+▁success -8.69623
+▁maybe -8.69626
+io -8.697
+▁created -8.69806
+▁additional -8.69852
+▁stand -8.69908
+▁full -8.69996
+▁idea -8.70155
+▁members -8.70357
+▁days -8.70458
+▁fish -8.70464
+▁count -8.70498
+▁third -8.70699
+ship -8.70734
+vo -8.7074
+▁taken -8.70819
+▁final -8.70857
+▁rock -8.70917
+der -8.71068
+▁garden -8.71087
+▁text -8.71456
+▁taste -8.71495
+▁less -8.71824
+▁themselves -8.71848
+▁police -8.71848
+▁week -8.71885
+▁regional -8.71923
+▁stone -8.71996
+▁el -8.72
+▁museum -8.72234
+ful -8.72286
+▁rest -8.7249
+▁hi -8.72493
+▁ru -8.7255
+▁moved -8.72606
+ach -8.72925
+▁machine -8.72975
+▁album -8.73019
+▁held -8.73043
+▁animals -8.7321
+▁itself -8.73402
+▁needs -8.7349
+▁published -8.73607
+▁organ -8.7372
+▁drink -8.73794
+ization -8.73842
+▁goal -8.74092
+▁control -8.74169
+▁economic -8.74187
+▁low -8.74481
+▁women -8.74589
+▁anymore -8.74633
+▁cup -8.74641
+▁master -8.74702
+mar -8.74895
+▁takes -8.74938
+▁successful -8.74979
+▁director -8.75048
+▁pay -8.75101
+▁feel -8.75255
+▁m -8.75307
+ow -8.75546
+▁su -8.75855
+▁within -8.75882
+▁design -8.75977
+burg -8.76075
+▁famous -8.76179
+▁version -8.76191
+▁written -8.76197
+▁looks -8.76221
+▁hit -8.76313
+▁step -8.76379
+▁heat -8.76451
+▁areas -8.7646
+im -8.76503
+pa -8.76559
+▁certain -8.76582
+▁understand -8.76591
+▁sleep -8.76987
+▁information -8.77018
+▁style -8.77169
+▁responsible -8.77404
+▁ca -8.77491
+▁political -8.77796
+▁traffic -8.77802
+lu -8.77863
+▁port -8.77967
+▁cover -8.7799
+dy -8.78024
+den -8.78144
+▁attack -8.78214
+▁border -8.78426
+▁mark -8.7847
+▁supposed -8.78623
+▁population -8.78623
+mon -8.787
+▁je -8.78828
+▁daughter -8.78847
+▁wine -8.79062
+▁th -8.79132
+▁ste -8.79158
+▁training -8.79203
+nna -8.79468
+▁track -8.79521
+▁sub -8.79584
+▁easy -8.79869
+▁honor -8.79869
+▁card -8.80105
+▁along -8.80278
+▁material -8.80287
+▁department -8.80291
+▁ch -8.80319
+▁wa -8.80468
+▁si -8.80488
+▁ask -8.80641
+▁pu -8.81098
+▁structure -8.81129
+▁gets -8.81292
+▁direction -8.81618
+▁floor -8.8164
+▁done -8.81847
+▁nature -8.81978
+▁parents -8.82007
+▁anyone -8.82013
+▁rath -8.82035
+wa -8.82133
+▁king -8.82242
+▁player -8.82357
+▁stage -8.82377
+▁gas -8.82391
+▁either -8.82405
+▁anti -8.82457
+▁whe -8.82474
+▁radio -8.82482
+▁cool -8.82564
+rk -8.82574
+ron -8.82746
+▁popular -8.82795
+▁via -8.83024
+▁cold -8.83239
+▁influence -8.83284
+▁moment -8.83373
+▁gu -8.83381
+▁original -8.83486
+bach -8.83495
+▁remains -8.83497
+▁wrote -8.83699
+▁ball -8.83783
+▁army -8.83919
+ite -8.83972
+▁led -8.84115
+▁towards -8.84184
+wi -8.84305
+▁professional -8.84343
+up -8.84577
+▁speak -8.84623
+▁mountain -8.84789
+if -8.84822
+▁began -8.84965
+▁modern -8.85009
+▁walk -8.85292
+▁bus -8.85361
+▁sense -8.85451
+▁watch -8.85452
+▁nice -8.85555
+▁exactly -8.85589
+▁wrong -8.85892
+▁anyway -8.85947
+rought -8.86078
+▁deal -8.86322
+▁value -8.86345
+red -8.86386
+▁return -8.86475
+▁die -8.86544
+▁doing -8.86661
+▁bra -8.867
+▁store -8.86714
+▁using -8.86801
+▁largest -8.86871
+▁eat -8.86891
+▁goes -8.87029
+tu -8.87064
+▁pan -8.87101
+▁games -8.87309
+▁mind -8.87327
+▁eye -8.87528
+gu -8.87664
+ize -8.87786
+more -8.87975
+▁bu -8.88001
+▁god -8.88119
+▁focus -8.88134
+▁performance -8.88146
+▁above -8.88591
+▁risk -8.88594
+▁boy -8.88933
+▁living -8.89046
+▁southern -8.8905
+ng -8.89138
+▁dead -8.89152
+▁ship -8.89161
+mer -8.89427
+▁studies -8.89504
+▁entire -8.89549
+▁grew -8.89551
+▁pull -8.89631
+nder -8.89807
+tra -8.8981
+▁ne -8.89818
+▁minutes -8.89973
+▁private -8.90011
+▁plan -8.90332
+▁non -8.90354
+▁necessary -8.90427
+▁self -8.90429
+way -8.90547
+▁lack -8.90601
+▁question -8.90708
+▁changed -8.90835
+▁civil -8.90893
+▁contract -8.90893
+▁action -8.91022
+▁roll -8.91327
+▁er -8.91335
+▁carrie -8.91343
+▁technical -8.9136
+▁outside -8.91486
+▁ki -8.91517
+▁bre -8.91792
+▁location -8.9183
+▁buried -8.9183
+▁hope -8.92087
+▁check -8.92372
+▁playing -8.92568
+▁hours -8.92653
+▁cla -8.92788
+per -8.93101
+▁leaves -8.93195
+▁experience -8.93227
+▁story -8.93301
+▁sand -8.93583
+▁guy -8.93675
+bra -8.93819
+▁lived -8.93837
+ze -8.93915
+▁given -8.93963
+▁award -8.94089
+▁brown -8.94211
+▁beside -8.9434
+▁wife -8.94694
+▁served -8.947
+▁prevent -8.9471
+▁immediately -8.94759
+▁approach -8.95179
+ator -8.95183
+▁task -8.95187
+▁hot -8.95224
+▁particular -8.95253
+▁turned -8.95482
+▁shows -8.95529
+▁star -8.95539
+▁network -8.9567
+▁northern -8.95673
+▁municipality -8.95694
+▁base -8.95768
+em -8.95883
+▁race -8.95905
+ak -8.95914
+▁sign -8.96127
+▁include -8.9615
+▁temperature -8.96157
+▁christian -8.96182
+▁dark -8.9622
+▁serve -8.96362
+▁trade -8.9651
+▁simpl -8.9651
+▁buy -8.96554
+▁administrative -8.96649
+▁though -8.96958
+▁table -8.96969
+ists -8.97044
+▁method -8.97145
+▁basic -8.97157
+▁coat -8.97176
+▁protect -8.97282
+▁youth -8.97357
+▁clean -8.97489
+▁study -8.97642
+▁size -8.97764
+▁neu -8.97833
+▁paul -8.98067
+▁president -8.98142
+ium -8.98311
+▁foundation -8.98465
+▁parts -8.98498
+pi -8.98708
+▁operation -8.98713
+▁face -8.98713
+men -8.98715
+▁win -8.98817
+▁terms -8.98905
+▁plays -8.98976
+▁record -8.98977
+line -8.98994
+▁students -8.9902
+▁particularly -8.99072
+▁continued -8.99107
+▁cat -8.99413
+▁decision -8.99657
+▁industry -8.99657
+▁travel -8.99658
+▁deep -8.99673
+ki -8.99686
+▁cell -8.997
+▁historical -8.99705
+▁sure -8.99855
+▁cap -9.0007
+▁couple -9.00168
+▁horse -9.00169
+▁happen -9.00206
+che -9.00259
+▁produced -9.00329
+▁gra -9.00513
+▁pressure -9.00697
+▁fe -9.00746
+▁doctor -9.00768
+heim -9.00949
+▁function -9.01013
+▁ke -9.01353
+▁speed -9.01656
+▁feature -9.01714
+▁monastery -9.01714
+▁thousand -9.01714
+▁groups -9.01789
+ac -9.01908
+▁move -9.01915
+▁tour -9.01931
+▁oil -9.0203
+▁arms -9.02149
+▁unfortunately -9.02247
+▁tree -9.02267
+▁sports -9.02406
+ric -9.02656
+taking -9.02716
+▁wide -9.02744
+▁difference -9.02776
+▁listen -9.02802
+ius -9.031
+▁hands -9.03257
+▁battle -9.03285
+▁somehow -9.03285
+▁actor -9.03323
+▁date -9.03332
+▁jump -9.03342
+ud -9.03635
+▁key -9.03693
+▁type -9.037
+▁thirty -9.03814
+▁nobody -9.03814
+▁hamburg -9.0382
+▁settlement -9.03828
+▁ear -9.03843
+▁closed -9.03962
+▁buildings -9.04237
+▁inter -9.04267
+▁search -9.04347
+▁despi -9.04356
+▁bri -9.04421
+▁legal -9.04477
+▁regard -9.04787
+▁site -9.04863
+▁champion -9.04864
+▁blood -9.0488
+▁heart -9.05251
+ide -9.0533
+▁companies -9.05418
+▁leg -9.05483
+▁theater -9.05616
+▁san -9.05875
+▁organization -9.05981
+▁fell -9.06132
+▁grave -9.06182
+▁gold -9.06218
+▁extreme -9.06502
+▁valley -9.06504
+▁books -9.06666
+▁covered -9.06953
+▁secret -9.07015
+▁section -9.07048
+▁distance -9.07049
+▁standard -9.07075
+ier -9.07162
+▁ran -9.07295
+▁para -9.07581
+▁magazine -9.07598
+▁tomorrow -9.07598
+▁bear -9.07704
+▁protection -9.0771
+itz -9.07942
+▁unit -9.08103
+▁catholic -9.0815
+▁attention -9.0816
+▁tu -9.0821
+▁plants -9.08247
+▁primari -9.08308
+▁range -9.08469
+▁glass -9.08529
+▁share -9.08554
+▁add -9.08582
+▁county -9.08704
+▁contact -9.08706
+▁munich -9.08706
+▁pick -9.08847
+▁tra -9.09012
+▁code -9.09125
+▁degree -9.09265
+▁bottle -9.09265
+▁double -9.09266
+▁origin -9.09304
+▁otherwise -9.09309
+▁movement -9.0931
+▁provide -9.09327
+iv -9.09498
+▁manage -9.09558
+▁across -9.09826
+ors -9.09911
+▁total -9.10036
+fa -9.10066
+ina -9.1007
+▁knows -9.10332
+▁province -9.10391
+▁fight -9.10425
+▁spot -9.1044
+▁reached -9.1055
+▁established -9.10583
+▁scientific -9.1096
+▁yellow -9.1096
+▁originally -9.10972
+▁cities -9.11017
+▁apparently -9.11029
+▁simple -9.11156
+▁effect -9.11159
+hl -9.11254
+▁letter -9.11267
+▁eyes -9.11315
+▁directly -9.11321
+▁wi -9.11416
+▁save -9.11454
+▁lay -9.11463
+▁france -9.11531
+▁phone -9.11533
+the -9.11571
+▁earth -9.11593
+▁sentence -9.1179
+▁friend -9.11868
+▁cha -9.11944
+▁true -9.12106
+▁lie -9.12308
+qui -9.1254
+▁higher -9.12582
+▁driver -9.12674
+▁attempt -9.12684
+▁cemetery -9.12684
+▁cultural -9.12684
+▁monument -9.12684
+▁minister -9.12684
+▁demand -9.12685
+▁roof -9.12697
+feld -9.1278
+▁musical -9.12933
+▁problem -9.12999
+▁sentences -9.13003
+▁ku -9.131
+▁report -9.13146
+tan -9.13247
+▁entrance -9.13265
+▁marriage -9.13265
+▁situation -9.13265
+▁computer -9.13265
+▁trains -9.13321
+▁mrs -9.13356
+▁shortly -9.13422
+▁rules -9.13511
+▁union -9.13524
+▁happened -9.13537
+▁create -9.13574
+▁competition -9.1385
+▁division -9.1385
+ven -9.13977
+▁nor -9.14206
+▁multiple -9.14441
+▁austria -9.14447
+▁throw -9.14454
+▁slowly -9.14633
+▁hardly -9.14699
+▁finished -9.14902
+row -9.14977
+▁ski -9.14982
+▁mistake -9.1503
+▁technology -9.1503
+▁british -9.1503
+▁vienna -9.1503
+▁answer -9.1503
+▁divided -9.15203
+▁charge -9.15363
+▁wind -9.15559
+▁document -9.15591
+▁hear -9.15608
+▁happy -9.15625
+▁painting -9.15656
+▁completed -9.15668
+lar -9.15893
+bert -9.15968
+▁apart -9.16193
+▁surface -9.16224
+▁figure -9.16224
+▁weather -9.16226
+▁formed -9.1623
+▁print -9.16249
+▁coming -9.16329
+▁path -9.16384
+▁founded -9.16485
+▁host -9.16537
+▁decided -9.16618
+▁develop -9.16635
+▁build -9.16782
+▁major -9.16785
+▁future -9.16827
+ified -9.16839
+▁hans -9.16915
+▁camp -9.17021
+▁words -9.17222
+▁names -9.17223
+▁shi -9.17257
+▁peter -9.1743
+▁republic -9.17432
+▁fruit -9.17432
+▁edge -9.17444
+▁poor -9.17633
+▁cost -9.17779
+▁literature -9.18042
+▁fair -9.18065
+▁collection -9.18106
+▁require -9.18153
+▁ja -9.18221
+▁independent -9.1824
+▁months -9.18248
+▁flo -9.18265
+▁described -9.18388
+ities -9.18441
+side -9.18591
+▁movie -9.18625
+▁bridge -9.18709
+▁wait -9.18732
+▁iron -9.18764
+▁saw -9.18818
+▁force -9.18894
+▁inside -9.189
+▁media -9.18947
+▁stra -9.19147
+▁send -9.19189
+▁science -9.19271
+▁energy -9.19273
+▁l -9.19374
+ich -9.19398
+▁z -9.1953
+tal -9.19766
+▁appeared -9.19815
+ever -9.19913
+having -9.20065
+▁officer -9.20099
+ok -9.2025
+▁chi -9.20342
+▁fin -9.20406
+ren -9.20407
+▁solution -9.20519
+▁saint -9.20524
+▁planned -9.20525
+▁worth -9.20623
+▁meant -9.20629
+▁dance -9.20642
+▁mill -9.20744
+▁suit -9.20753
+▁trans -9.20884
+▁visit -9.21014
+▁notice -9.21093
+▁paris -9.21098
+▁remember -9.21148
+▁scene -9.21148
+▁avoid -9.21149
+▁sister -9.21185
+▁neighbor -9.21226
+▁wear -9.21253
+ov -9.21387
+▁fan -9.216
+by -9.21683
+▁involved -9.21781
+▁female -9.21781
+▁internet -9.21782
+▁offer -9.21823
+▁runs -9.21992
+▁problems -9.22072
+▁fast -9.22093
+▁attended -9.22107
+▁met -9.22356
+▁forms -9.22363
+▁easily -9.22418
+▁late -9.22539
+▁lock -9.22718
+▁write -9.22767
+▁older -9.22961
+▁concert -9.23059
+▁cream -9.23059
+▁guest -9.23131
+▁opened -9.23172
+▁generally -9.23308
+▁spa -9.23419
+▁fla -9.23641
+▁doubt -9.23704
+▁newspaper -9.23704
+▁writing -9.23704
+hn -9.23709
+▁comp -9.23993
+fi -9.24189
+▁tax -9.24232
+▁direct -9.24274
+▁dream -9.24354
+▁silver -9.24362
+▁cook -9.24371
+▁flight -9.24375
+▁sal -9.24398
+▁increase -9.24506
+▁farm -9.24636
+▁rain -9.24674
+▁measure -9.24674
+▁har -9.24853
+▁administration -9.25007
+▁friedrich -9.25007
+▁twelve -9.25007
+▁driving -9.25007
+▁morning -9.2501
+▁countries -9.25022
+gi -9.25034
+▁price -9.2507
+ward -9.25158
+▁states -9.25237
+▁significant -9.25251
+log -9.25486
+▁application -9.25665
+▁festival -9.25665
+▁principle -9.25665
+▁content -9.25666
+cient -9.25675
+▁treat -9.25676
+▁italian -9.25676
+▁touch -9.25752
+▁john -9.25835
+▁imp -9.25839
+ran -9.25896
+con -9.25896
+enburg -9.25922
+uc -9.25996
+▁awarded -9.26046
+let -9.26074
+▁released -9.26095
+▁herself -9.26327
+▁relationship -9.26334
+▁foot -9.26391
+▁kilometers -9.26464
+▁receive -9.26494
+▁changes -9.2652
+▁trees -9.2664
+▁animal -9.26654
+▁miss -9.26973
+▁professor -9.26994
+▁kept -9.26996
+▁fo -9.2739
+▁exist -9.274
+graph -9.27417
+▁agriculture -9.27665
+▁nowadays -9.27665
+▁included -9.27676
+▁quit -9.27725
+▁thereby -9.28033
+▁leader -9.28082
+▁returned -9.28129
+▁television -9.28341
+▁medal -9.28371
+▁sold -9.28436
+▁coach -9.28449
+▁fun -9.28565
+▁shot -9.28945
+▁accident -9.29021
+▁square -9.29055
+▁fill -9.29135
+▁fit -9.29226
+▁tro -9.29305
+▁rule -9.29396
+▁sport -9.29549
+▁switch -9.29638
+val -9.29642
+▁gi -9.2965
+▁nevertheless -9.29706
+▁society -9.29706
+▁yesterday -9.29706
+▁coast -9.29721
+▁preserved -9.29797
+▁sha -9.30227
+▁cor -9.30291
+▁commercial -9.30396
+▁airport -9.3042
+▁press -9.30474
+ological -9.30499
+time -9.30624
+▁paper -9.30639
+▁personal -9.30655
+▁supported -9.30725
+▁dam -9.30865
+▁cable -9.30924
+▁languages -9.30948
+▁link -9.31079
+▁novel -9.31096
+▁appearance -9.31108
+ug -9.3118
+▁fi -9.31265
+▁shape -9.31312
+▁failed -9.31347
+▁n -9.31495
+▁thank -9.31658
+▁performed -9.31713
+▁wing -9.31746
+▁volume -9.31789
+▁spread -9.31789
+▁cur -9.31856
+▁wish -9.31893
+cht -9.32209
+▁fa -9.32255
+▁starts -9.32297
+▁ring -9.32303
+▁meeting -9.32337
+ose -9.32402
+▁fu -9.3244
+▁grand -9.32445
+▁object -9.3246
+▁connected -9.32462
+▁circle -9.32493
+▁coffee -9.32493
+▁twice -9.32493
+▁football -9.32499
+▁flat -9.32892
+▁contains -9.32985
+▁occur -9.33113
+▁college -9.33203
+▁expert -9.33203
+▁smoke -9.33203
+▁shoot -9.33282
+▁carry -9.33452
+▁clearly -9.33502
+▁shaped -9.33714
+▁tip -9.33834
+▁concept -9.33917
+▁escape -9.33917
+▁russian -9.34167
+▁sur -9.34167
+▁counter -9.34204
+ries -9.34204
+▁partner -9.3424
+▁burn -9.34381
+▁match -9.34531
+▁municipal -9.34585
+▁windows -9.34616
+▁vehicle -9.34628
+▁talent -9.34636
+▁milk -9.34637
+▁making -9.34638
+▁display -9.34644
+▁eastern -9.34694
+▁ended -9.34846
+▁pet -9.34866
+▁recently -9.34914
+▁karl -9.34955
+▁speaking -9.34985
+▁ya -9.35109
+▁houses -9.35169
+▁editor -9.35361
+▁knowledge -9.35361
+iz -9.35364
+▁noise -9.35373
+▁majority -9.3541
+▁heard -9.35432
+▁winter -9.35449
+▁shop -9.35496
+▁italy -9.35532
+▁fine -9.35534
+▁cra -9.35593
+nne -9.35725
+▁allow -9.35742
+▁owner -9.35772
+han -9.35781
+▁electric -9.35896
+▁flows -9.36015
+▁access -9.36084
+▁library -9.36091
+▁running -9.36101
+▁western -9.36158
+▁pen -9.36165
+▁fly -9.36167
+hausen -9.36173
+▁motor -9.36174
+ik -9.36189
+▁che -9.36195
+▁results -9.36204
+ai -9.36208
+▁wheel -9.36225
+▁events -9.36274
+▁reasons -9.36276
+▁der -9.3636
+▁pri -9.36576
+▁reach -9.36615
+▁meaning -9.36694
+▁behavior -9.36826
+▁institute -9.36826
+▁choir -9.36826
+▁rich -9.36841
+one -9.36991
+▁statement -9.37055
+▁ban -9.37205
+▁mor -9.37208
+▁friends -9.37233
+▁safe -9.37551
+▁amount -9.37567
+▁procedure -9.37567
+▁committee -9.37567
+▁products -9.37584
+enberg -9.37597
+▁kar -9.37737
+▁kill -9.37739
+▁thanks -9.37754
+▁mat -9.3784
+▁trip -9.37952
+▁comment -9.37962
+▁broken -9.38043
+▁arm -9.38289
+▁beautiful -9.38313
+▁switzerland -9.38313
+▁opinion -9.38313
+▁quarter -9.38313
+▁straight -9.38317
+▁lawyer -9.3832
+▁complete -9.38344
+chi -9.38398
+▁y -9.38639
+▁health -9.38696
+▁dear -9.38829
+▁designed -9.38963
+▁instrument -9.39065
+▁complain -9.39065
+▁bishop -9.39065
+▁election -9.39082
+▁sort -9.39143
+▁row -9.39189
+▁stands -9.39404
+stadt -9.39406
+▁hour -9.39434
+▁destroyed -9.39504
+▁weight -9.39583
+bel -9.39636
+▁event -9.39703
+▁tea -9.39707
+tel -9.39779
+▁arch -9.39802
+▁cologne -9.39823
+▁london -9.39823
+▁video -9.39823
+▁prison -9.39849
+▁management -9.39879
+ring -9.39944
+▁character -9.40026
+▁musician -9.40051
+▁contain -9.40061
+▁practice -9.40083
+▁weak -9.40204
+▁signal -9.40224
+▁sounds -9.40352
+▁offers -9.40384
+▁asked -9.40481
+▁limited -9.40492
+▁equal -9.40586
+ja -9.40603
+▁baby -9.40611
+▁release -9.40854
+ek -9.40964
+za -9.41096
+▁separate -9.4121
+▁article -9.41356
+▁highly -9.41416
+▁america -9.41482
+▁offered -9.41535
+▁starting -9.41573
+▁fat -9.41576
+istic -9.41601
+▁highest -9.41633
+▁below -9.41659
+▁source -9.41661
+▁joined -9.41839
+▁cases -9.42045
+▁grow -9.42056
+▁graduat -9.4213
+▁security -9.4213
+▁block -9.42131
+▁blind -9.42131
+▁medical -9.42133
+▁image -9.42136
+▁woman -9.42224
+▁bomb -9.4226
+▁bike -9.42297
+▁enter -9.42358
+▁app -9.42369
+▁writer -9.42485
+fer -9.42617
+▁johann -9.42735
+cke -9.42754
+eg -9.42767
+▁connect -9.42846
+mark -9.42867
+▁quality -9.42912
+▁increasing -9.42912
+▁summer -9.42935
+▁tourist -9.42995
+old -9.43039
+▁achieved -9.43158
+▁picture -9.43293
+▁mine -9.43419
+▁dr -9.43567
+▁cal -9.43573
+over -9.43586
+▁lose -9.43686
+▁screen -9.43699
+▁greek -9.43705
+▁alternative -9.43709
+▁ober -9.43774
+▁wor -9.43787
+cker -9.43834
+▁weeks -9.43883
+▁questions -9.43908
+▁burg -9.44162
+called -9.44184
+▁removed -9.44277
+▁absolute -9.44493
+▁mouth -9.44493
+▁prince -9.44493
+▁strasse -9.44518
+▁bread -9.44532
+▁exhibition -9.44548
+▁col -9.44552
+▁parish -9.44649
+▁par -9.44653
+▁tail -9.44656
+▁teams -9.44676
+▁damage -9.44825
+▁parliament -9.44869
+▁hill -9.44878
+work -9.44955
+▁championship -9.44977
+▁songs -9.45047
+▁hospital -9.45293
+▁culture -9.45293
+▁reserve -9.45293
+▁adult -9.45294
+▁bag -9.45303
+▁claim -9.45319
+▁exception -9.45355
+▁passed -9.45375
+▁box -9.45535
+▁boat -9.45539
+ham -9.45736
+pper -9.46024
+▁might -9.46107
+▁prize -9.46116
+▁martin -9.46142
+▁accept -9.46154
+ative -9.46205
+▁customer -9.46361
+▁wo -9.46409
+▁fle -9.46409
+▁points -9.46457
+▁normal -9.46497
+▁artist -9.46547
+▁pra -9.46707
+▁engine -9.46747
+▁permanent -9.46912
+▁foreign -9.46912
+▁crown -9.46913
+▁royal -9.46917
+▁balance -9.46929
+▁super -9.46941
+▁explain -9.46974
+▁continue -9.47055
+▁rank -9.47285
+▁fields -9.47285
+▁corner -9.47313
+▁piece -9.47551
+for -9.47589
+▁euro -9.47712
+▁heavy -9.47732
+▁percent -9.47732
+▁typical -9.47732
+▁wilhelm -9.47732
+▁guitar -9.47732
+▁memory -9.47732
+▁relatively -9.47733
+▁waste -9.47771
+quite -9.47833
+▁costs -9.47841
+▁ages -9.47941
+▁successor -9.47985
+▁gav -9.48016
+▁moon -9.48041
+▁correct -9.48042
+gan -9.48077
+▁records -9.48199
+▁tur -9.48344
+▁branch -9.48382
+ctor -9.48433
+▁wild -9.48459
+▁planet -9.4852
+▁appointed -9.48558
+▁software -9.48558
+▁overall -9.48566
+▁forget -9.48589
+▁scar -9.48677
+out -9.48746
+▁evening -9.4884
+gel -9.49094
+▁fund -9.49127
+▁attract -9.49359
+▁activities -9.49392
+▁extensive -9.49392
+▁broadcast -9.49393
+▁background -9.49425
+▁min -9.49509
+▁wave -9.49545
+▁pure -9.49793
+▁x -9.49839
+▁referr -9.49993
+▁grown -9.50017
+▁singer -9.50065
+▁topic -9.50074
+▁mass -9.50161
+▁consider -9.50168
+▁audience -9.50232
+▁frankfurt -9.50232
+▁orient -9.50232
+▁judge -9.50233
+▁thick -9.50235
+▁authorities -9.50259
+▁chance -9.50276
+▁perform -9.50324
+▁meet -9.50335
+▁subsequently -9.50448
+▁ros -9.50456
+▁united -9.50524
+▁bro -9.50717
+▁pain -9.50755
+▁significantly -9.50764
+▁gun -9.50847
+▁dry -9.50875
+▁metal -9.50885
+▁austrian -9.51066
+▁communities -9.51079
+▁vehicles -9.5109
+▁york -9.5109
+▁hol -9.51203
+lon -9.51312
+▁camera -9.51395
+▁specialist -9.51438
+▁pictures -9.51519
+▁tradition -9.51522
+rge -9.51527
+▁says -9.51602
+▁lines -9.51713
+▁represented -9.51738
+▁account -9.51934
+▁institution -9.51934
+▁teeth -9.51935
+▁opposite -9.51937
+▁missing -9.51941
+▁appear -9.51972
+▁limit -9.52042
+sdorf -9.52174
+▁spe -9.52339
+▁regular -9.52382
+▁beat -9.52539
+▁meters -9.52621
+▁product -9.52642
+▁defense -9.52796
+▁religious -9.52796
+▁stretch -9.52796
+▁user -9.52939
+▁africa -9.53081
+▁mountains -9.5323
+▁colors -9.53332
+▁warm -9.53405
+▁rat -9.5352
+del -9.53542
+▁neighbors -9.53552
+▁jewish -9.53666
+▁landscape -9.53666
+▁opponent -9.53666
+▁situated -9.53666
+▁fresh -9.53666
+▁ze -9.53667
+▁representative -9.53695
+▁window -9.53715
+▁cake -9.53738
+▁rail -9.53845
+▁male -9.53903
+▁pieces -9.53919
+▁mayor -9.54023
+▁recognize -9.54041
+▁shown -9.54347
+▁trained -9.54361
+▁required -9.54381
+▁exchange -9.54543
+▁stuff -9.54543
+▁estate -9.54545
+▁chair -9.54545
+▁orange -9.54545
+▁sugar -9.54556
+▁sector -9.54564
+▁beach -9.54566
+▁apartment -9.54606
+ile -9.54621
+▁skin -9.54701
+▁factory -9.5483
+▁previously -9.54952
+cher -9.5497
+▁bor -9.54974
+▁loud -9.55068
+vent -9.55123
+▁standing -9.55172
+stein -9.55369
+▁react -9.55382
+▁financial -9.55428
+▁spent -9.55442
+▁movies -9.55473
+▁push -9.55497
+▁fuel -9.55524
+▁expect -9.55559
+▁reality -9.5559
+▁flow -9.55605
+▁replaced -9.55668
+▁theory -9.55777
+▁command -9.55813
+mmer -9.55851
+▁previous -9.55907
+▁serious -9.55916
+house -9.55999
+▁advantage -9.56321
+▁emergency -9.56321
+▁interior -9.56321
+▁stick -9.56322
+▁funny -9.5633
+▁hurt -9.56366
+▁stream -9.56437
+▁flag -9.56482
+zer -9.56535
+▁bone -9.56821
+▁rate -9.56837
+▁effects -9.56854
+ue -9.56859
+▁golden -9.56866
+light -9.56937
+aries -9.5717
+▁artificial -9.57222
+▁defend -9.57222
+▁register -9.57222
+▁captain -9.57222
+▁complex -9.57222
+▁weekend -9.57226
+▁importance -9.5723
+▁rid -9.5725
+▁unknown -9.57299
+▁pol -9.57569
+▁speaker -9.57595
+▁describe -9.5763
+▁broke -9.57767
+ette -9.57805
+▁saying -9.57807
+▁issue -9.57911
+▁films -9.57952
+▁exclusive -9.58131
+▁olympic -9.58131
+▁understood -9.58131
+▁everybody -9.58136
+▁brand -9.5814
+definitely -9.58156
+▁chairman -9.58166
+▁winner -9.58194
+▁painter -9.58197
+je -9.58218
+▁practical -9.58271
+▁obviously -9.58281
+▁van -9.58313
+▁jan -9.58335
+▁student -9.58364
+▁listed -9.58582
+ible -9.58606
+▁tries -9.58675
+▁fur -9.58748
+▁birth -9.59044
+▁august -9.59048
+▁conflict -9.59048
+▁effort -9.59048
+▁passenger -9.59048
+▁supply -9.59048
+▁graphic -9.5905
+▁shopping -9.5905
+▁sorry -9.59051
+▁desert -9.59052
+▁spring -9.59079
+▁reich -9.59199
+▁bay -9.59238
+▁customs -9.59305
+▁fee -9.59347
+▁bare -9.59353
+ification -9.5943
+▁fourth -9.595
+▁pack -9.5963
+ley -9.5965
+▁sit -9.59826
+▁provided -9.59873
+▁commission -9.59974
+▁soccer -9.59974
+▁latter -9.59977
+▁latin -9.59982
+▁spend -9.59983
+▁automatically -9.59985
+▁property -9.59988
+▁indian -9.6002
+▁industrial -9.60021
+ition -9.6012
+rew -9.60143
+par -9.60172
+com -9.60218
+▁nu -9.60262
+▁rarely -9.60313
+▁ride -9.60336
+car -9.60347
+▁flowers -9.60744
+van -9.60747
+▁reflect -9.60909
+▁applied -9.60909
+▁snow -9.6091
+▁wolf -9.60917
+▁swiss -9.60921
+▁zone -9.60948
+▁bla -9.61064
+▁creat -9.61074
+ates -9.61187
+▁climb -9.61198
+▁mentioned -9.61263
+▁resident -9.61273
+▁conditions -9.61316
+▁meat -9.6139
+▁measures -9.61467
+▁hero -9.61479
+▁needed -9.61818
+▁agricultural -9.61852
+▁heinrich -9.61852
+▁slightly -9.61853
+▁composition -9.61855
+▁pur -9.61856
+▁drug -9.61865
+▁meantime -9.61877
+▁shock -9.61913
+▁gear -9.62044
+▁bern -9.6209
+ability -9.62165
+our -9.62204
+▁discovered -9.62363
+ang -9.62379
+though -9.62477
+▁acting -9.62617
+▁respect -9.62772
+▁specific -9.62797
+▁disease -9.62805
+▁equipped -9.62805
+▁inhabitants -9.62805
+▁rhine -9.62805
+▁eleven -9.62805
+▁click -9.62805
+▁visible -9.62806
+▁piano -9.62806
+▁remain -9.62811
+▁earlier -9.62811
+▁killed -9.62819
+▁brake -9.62883
+▁digital -9.62997
+▁finish -9.63052
+▁bird -9.63269
+▁pit -9.63283
+▁march -9.63391
+port -9.63437
+▁pot -9.63444
+▁evidence -9.63766
+▁enjoy -9.63766
+▁kitchen -9.63766
+▁license -9.63766
+▁spanish -9.63766
+▁smell -9.63767
+▁taught -9.63773
+▁protestant -9.63781
+▁package -9.63798
+▁drum -9.6385
+▁chain -9.63943
+uck -9.63989
+▁ger -9.64079
+▁joke -9.64101
+ria -9.64169
+▁smaller -9.64367
+can -9.6447
+▁numbers -9.64472
+wald -9.64476
+▁diet -9.64544
+▁sing -9.6459
+▁collect -9.64664
+▁admit -9.64697
+▁contrast -9.64737
+▁perfect -9.64737
+▁somewhere -9.64737
+▁favorite -9.64748
+▁communication -9.64763
+▁appears -9.64793
+▁systems -9.64817
+▁china -9.64825
+▁represent -9.64863
+ography -9.64902
+▁max -9.64934
+▁producer -9.64993
+▁secondary -9.65002
+▁forces -9.65116
+▁regularly -9.65206
+▁element -9.65295
+▁quick -9.65306
+▁drop -9.65398
+▁tool -9.65492
+lie -9.65583
+▁qua -9.65712
+▁restaurant -9.65717
+▁characteristic -9.65728
+▁headquarters -9.65787
+fully -9.65881
+ung -9.65924
+▁aim -9.66009
+▁correspond -9.66055
+▁symbol -9.66271
+▁era -9.66286
+▁draw -9.66465
+▁loss -9.66471
+▁increased -9.66476
+▁reform -9.66704
+▁length -9.66707
+▁examination -9.66708
+▁empty -9.66708
+▁england -9.66708
+▁insurance -9.66708
+▁stopped -9.66713
+▁cloud -9.66717
+▁forward -9.66754
+▁frank -9.66791
+▁trick -9.66893
+▁skill -9.66897
+illa -9.66985
+hu -9.67127
+▁elements -9.67136
+▁blu -9.67156
+▁maria -9.67207
+▁dare -9.67292
+▁stem -9.67296
+hen -9.67312
+chen -9.67315
+▁photo -9.67353
+▁lau -9.67425
+▁month -9.67445
+▁manager -9.67669
+▁afraid -9.67707
+▁announced -9.67707
+▁conference -9.67707
+▁experiment -9.67707
+▁tournament -9.67707
+▁sweet -9.67708
+▁speech -9.67708
+▁citizens -9.67772
+▁acid -9.67772
+▁strip -9.6783
+▁whatever -9.67977
+▁scale -9.68092
+▁prove -9.68144
+▁traditional -9.68192
+▁plane -9.68194
+▁birds -9.68232
+▁recognized -9.68283
+▁ideal -9.68305
+▁opening -9.68315
+▁gone -9.68531
+▁und -9.68617
+▁surrounded -9.68717
+▁surrounding -9.68717
+▁narrow -9.68718
+▁troops -9.68719
+▁discussion -9.68746
+▁sight -9.68768
+ino -9.68871
+▁elected -9.68885
+▁frequently -9.68888
+▁compared -9.68905
+▁layer -9.68937
+▁berg -9.69001
+▁severe -9.6914
+▁championships -9.6915
+▁happens -9.69186
+▁joint -9.69475
+dorf -9.69494
+▁mal -9.69494
+▁christ -9.69498
+▁primar -9.695
+ular -9.69674
+▁michael -9.69738
+▁orchestra -9.69738
+▁generation -9.69738
+▁promise -9.69741
+▁lessons -9.69745
+▁staff -9.6982
+▁op -9.69823
+▁saxony -9.69979
+▁nut -9.69993
+▁rose -9.70012
+▁composer -9.70052
+▁luck -9.70112
+▁dia -9.7025
+▁players -9.70314
+▁determined -9.70486
+▁mis -9.70549
+▁academy -9.70769
+▁request -9.70769
+▁edition -9.70772
+▁stories -9.70778
+▁criticism -9.70793
+▁bright -9.70796
+▁prefer -9.70797
+▁electro -9.70815
+▁machines -9.70863
+▁mission -9.70933
+lor -9.7097
+▁visited -9.70973
+▁concern -9.71051
+▁char -9.71066
+▁artists -9.71178
+▁tre -9.71201
+bility -9.714
+▁services -9.71402
+▁pig -9.7144
+▁including -9.71811
+▁repair -9.71811
+▁chamber -9.71811
+▁adapt -9.71811
+▁brief -9.71811
+▁learned -9.71819
+▁legs -9.71819
+▁daily -9.71829
+▁message -9.71842
+▁sitting -9.71862
+▁reaction -9.7193
+▁soil -9.71971
+ath -9.71993
+▁neck -9.72127
+▁feed -9.72294
+known -9.72446
+▁feeling -9.72506
+▁ray -9.72527
+▁schu -9.72535
+▁flying -9.72758
+▁classic -9.72783
+▁distribution -9.72863
+▁flash -9.72911
+▁thin -9.72928
+▁mode -9.72983
+▁spin -9.7302
+▁eggs -9.73038
+▁larger -9.73131
+▁disappeared -9.73157
+▁added -9.7324
+▁flee -9.73389
+atic -9.73509
+▁ahead -9.73626
+▁ep -9.73698
+▁sho -9.73803
+▁busy -9.73823
+▁plate -9.739
+pel -9.73919
+▁concrete -9.73927
+▁universities -9.73927
+▁storm -9.73935
+▁ticket -9.73944
+▁visitors -9.73954
+▁respective -9.73964
+▁stock -9.74013
+▁j -9.7411
+bar -9.74126
+▁presented -9.74158
+form -9.74202
+▁finds -9.74276
+▁flu -9.74366
+ese -9.74438
+▁showed -9.74576
+▁recent -9.74588
+▁hole -9.74804
+▁express -9.74918
+▁mess -9.7495
+▁spit -9.74968
+▁anne -9.74976
+▁device -9.74976
+▁minute -9.74982
+▁stomach -9.75002
+▁knew -9.75002
+▁knock -9.75002
+▁stars -9.75006
+▁chapel -9.75014
+▁nei -9.75045
+og -9.75051
+▁waiting -9.75063
+▁poem -9.75064
+ker -9.75118
+ku -9.75194
+▁younger -9.75222
+▁domin -9.75223
+▁upon -9.75664
+▁issues -9.75846
+▁marked -9.75918
+▁choice -9.76089
+▁employed -9.76089
+▁myself -9.76089
+▁excuse -9.76089
+▁basis -9.76097
+▁detect -9.76099
+▁operating -9.76128
+▁organized -9.76142
+▁option -9.76183
+▁girl -9.76183
+▁carrier -9.7619
+▁types -9.76326
+▁plans -9.76442
+▁engines -9.76449
+cal -9.76462
+▁operated -9.76485
+▁mont -9.76642
+▁exists -9.76647
+mate -9.76736
+▁recommended -9.76744
+▁solo -9.76763
+clock -9.7686
+lessly -9.77187
+▁criminal -9.77188
+▁leaf -9.77189
+▁copy -9.77195
+▁occasionally -9.77218
+▁guard -9.77236
+▁agent -9.77275
+▁affected -9.77554
+▁apple -9.77623
+▁bur -9.77685
+▁agree -9.77753
+ches -9.77754
+wing -9.77808
+▁activity -9.78299
+▁dialect -9.78299
+▁load -9.78428
+▁label -9.78434
+▁ruin -9.78606
+▁mail -9.78646
+▁journalist -9.78755
+her -9.78805
+▁constant -9.79029
+▁mixed -9.79037
+▁cas -9.79074
+▁del -9.79166
+▁fingers -9.79302
+▁bell -9.79391
+▁cathedral -9.79423
+▁kindergarten -9.79423
+▁biggest -9.79423
+▁ocean -9.79423
+▁robert -9.7949
+▁auto -9.79538
+ple -9.79551
+▁brick -9.79617
+▁queen -9.79677
+▁sales -9.79761
+▁bal -9.79777
+isch -9.79831
+▁falls -9.79853
+▁sources -9.8011
+hof -9.80128
+▁baden -9.80253
+▁container -9.80552
+▁trust -9.80559
+▁politics -9.80559
+▁positive -9.8056
+▁okay -9.8056
+▁laugh -9.80562
+▁robot -9.80563
+▁lord -9.80569
+▁childhood -9.80574
+▁crime -9.80588
+▁frame -9.80602
+▁status -9.8061
+▁manufacturer -9.80622
+▁steam -9.80636
+▁throughout -9.80666
+▁minor -9.80714
+▁pool -9.8073
+▁hotel -9.80789
+▁bio -9.80882
+▁reading -9.80963
+▁ready -9.80991
+▁rare -9.81035
+▁recorded -9.81116
+▁mix -9.81161
+ike -9.81183
+▁critic -9.81232
+▁differently -9.81278
+▁gen -9.81333
+▁salt -9.81405
+▁singing -9.81412
+ita -9.81617
+▁za -9.81622
+▁architect -9.81686
+▁candidate -9.81709
+▁economy -9.81709
+▁leaving -9.81709
+▁meanwhile -9.81709
+▁patient -9.81709
+▁religion -9.81709
+▁studio -9.81709
+▁empire -9.81709
+▁urban -9.81709
+▁filter -9.81709
+▁liquid -9.81709
+▁policy -9.81709
+▁tongue -9.8171
+▁medium -9.81711
+▁fought -9.81713
+▁characterized -9.81738
+▁blow -9.81744
+▁zero -9.81794
+▁occurs -9.81828
+▁arranged -9.81851
+▁teaching -9.81947
+▁sin -9.81968
+▁web -9.81969
+▁et -9.82007
+▁sto -9.82076
+ey -9.82094
+▁consequences -9.82161
+▁functions -9.82167
+▁daughters -9.82241
+▁gar -9.82253
+▁peace -9.82417
+▁placed -9.82432
+▁kl -9.82498
+▁flower -9.82501
+▁eva -9.8256
+▁nose -9.82593
+ya -9.82653
+▁conduct -9.8272
+▁cooking -9.82869
+▁expensive -9.82871
+▁medicine -9.82871
+▁pattern -9.82871
+▁decade -9.82872
+▁franz -9.82872
+▁japanese -9.82876
+▁inspir -9.82882
+▁warning -9.82921
+▁dedicated -9.82945
+▁factor -9.83056
+▁gro -9.83057
+▁million -9.83061
+▁noble -9.83123
+▁islands -9.83188
+▁artistic -9.83233
+top -9.83325
+▁gate -9.83338
+▁oldes -9.8334
+▁originate -9.83341
+▁refer -9.83348
+▁boss -9.83385
+▁eating -9.83405
+▁georg -9.83452
+ball -9.83488
+▁employees -9.8349
+▁suffered -9.83523
+▁kra -9.83685
+▁handle -9.8375
+▁expected -9.83868
+▁contribut -9.83946
+▁investigat -9.83947
+▁girls -9.83947
+▁colleague -9.84048
+▁facilities -9.84048
+▁stuttgart -9.84048
+▁essential -9.84048
+▁chief -9.84049
+▁actress -9.84049
+▁shut -9.8405
+▁average -9.8405
+▁bought -9.84056
+▁creative -9.84089
+▁guide -9.84112
+▁tram -9.84145
+ft -9.84152
+▁map -9.84191
+▁root -9.8421
+▁determine -9.84339
+▁pair -9.84553
+▁ships -9.84644
+▁prepared -9.84804
+▁inform -9.8489
+ome -9.8498
+▁fra -9.85119
+▁global -9.85238
+▁destination -9.85238
+▁residence -9.85238
+▁substance -9.85238
+▁chemical -9.85238
+▁dispute -9.85238
+▁pretty -9.8524
+▁shirt -9.85242
+▁cheap -9.85243
+▁juice -9.85305
+▁profit -9.85308
+▁delete -9.85379
+▁tie -9.85396
+▁trace -9.85538
+▁pop -9.8554
+ei -9.85541
+▁cast -9.85553
+▁duke -9.85573
+▁promoted -9.8561
+▁rejected -9.85697
+▁stepp -9.85718
+▁seriously -9.85795
+tsch -9.86078
+▁growing -9.86248
+▁join -9.86353
+▁equipment -9.86443
+▁everywhere -9.86443
+▁satellite -9.86443
+▁territory -9.86443
+▁bronze -9.86443
+▁trouble -9.86443
+▁appeal -9.86445
+▁stress -9.86446
+▁reveal -9.86446
+▁breath -9.86463
+▁polish -9.86481
+▁easier -9.86505
+▁kids -9.86506
+▁hang -9.86559
+▁useful -9.86623
+see -9.86658
+▁bill -9.86679
+▁succeeded -9.8678
+har -9.86854
+▁rome -9.87044
+▁spelling -9.87044
+▁borders -9.87055
+▁bavaria -9.87078
+▁careful -9.87308
+▁browser -9.87663
+▁depict -9.87663
+▁nucle -9.87663
+▁stadium -9.87663
+▁threat -9.87663
+▁governor -9.87704
+▁blame -9.87721
+▁electricity -9.87731
+▁washing -9.87734
+▁nord -9.87771
+▁presentation -9.87985
+don -9.8814
+▁gained -9.88156
+▁walking -9.88247
+meter -9.88384
+▁fri -9.88551
+▁gray -9.88578
+▁lecture -9.88593
+▁mur -9.88681
+ula -9.88703
+▁talking -9.88847
+▁appropriate -9.88897
+▁joseph -9.88897
+▁leipzig -9.88897
+▁sponsor -9.88897
+▁stupid -9.88897
+▁baroque -9.88897
+▁translated -9.88898
+▁target -9.88898
+▁democratic -9.88929
+▁agreement -9.88945
+▁raised -9.89047
+▁assumed -9.89052
+▁aid -9.89069
+▁steel -9.89145
+▁regarding -9.89163
+▁hat -9.89219
+▁opera -9.89242
+j -9.89307
+▁participated -9.89452
+▁sw -9.89502
+▁detail -9.89544
+horn -9.8957
+▁grows -9.89629
+room -9.89642
+ah -9.8968
+▁bun -9.89725
+▁remaining -9.89741
+▁slow -9.89796
+ane -9.89883
+▁harm -9.89915
+kin -9.89944
+rac -9.9
+▁pipe -9.90019
+▁suggest -9.90045
+▁yo -9.90086
+▁owned -9.901
+▁maintenance -9.90147
+▁scream -9.90147
+▁chinese -9.90147
+▁chosen -9.90148
+▁husband -9.90148
+▁alarm -9.90159
+▁photograph -9.90159
+▁hello -9.90184
+▁ram -9.90268
+nny -9.90372
+▁ser -9.90373
+▁huge -9.9043
+▁recommend -9.90652
+▁kiss -9.90691
+head -9.9079
+▁condition -9.90863
+▁sum -9.91109
+▁bass -9.91223
+▁entered -9.91376
+▁benefit -9.91413
+▁perspective -9.91413
+▁smooth -9.91413
+▁volunteer -9.91413
+▁aspect -9.91413
+▁strike -9.91417
+▁planning -9.91423
+▁internal -9.91426
+▁dress -9.91429
+▁gift -9.91432
+▁devices -9.91444
+▁strange -9.91446
+▁literar -9.91449
+▁tourism -9.91486
+▁suddenly -9.91591
+▁prime -9.91724
+▁surprised -9.91744
+▁extended -9.91884
+▁founder -9.92325
+▁explained -9.92597
+▁complicated -9.92695
+▁contemporary -9.92695
+▁cycle -9.92695
+▁freight -9.92695
+▁tunnel -9.92695
+▁climate -9.92695
+▁truth -9.92695
+▁soviet -9.92695
+▁split -9.92698
+▁sciences -9.92699
+▁william -9.92699
+▁core -9.92704
+▁keyboard -9.92707
+▁native -9.92714
+▁sole -9.92725
+▁praise -9.92752
+▁desired -9.92796
+▁switched -9.92823
+▁alone -9.92837
+▁poet -9.92872
+▁destroy -9.93238
+▁shell -9.93373
+▁formation -9.93381
+▁trainer -9.93567
+▁exact -9.93581
+slav -9.93834
+▁except -9.93892
+▁otto -9.93964
+é -9.93994
+▁breakfast -9.93994
+▁context -9.93994
+▁philosophy -9.93994
+▁thomas -9.93994
+▁knight -9.93994
+▁folk -9.93995
+▁roast -9.9401
+▁alright -9.94035
+▁environment -9.94052
+▁analysis -9.94084
+▁script -9.94104
+▁expression -9.94117
+▁earn -9.94139
+▁authority -9.94157
+▁multi -9.9419
+▁net -9.94363
+▁clo -9.94519
+▁spoke -9.94727
+ben -9.94776
+▁parties -9.9482
+▁spell -9.94877
+▁farmer -9.94971
+▁gene -9.95057
+▁rear -9.95141
+▁pal -9.9517
+▁favor -9.95295
+▁bottom -9.9531
+▁california -9.9531
+▁campaign -9.9531
+▁victim -9.9531
+▁valuable -9.9531
+▁eventually -9.9531
+▁exhaust -9.9531
+▁drain -9.95311
+▁drill -9.95311
+▁dough -9.95312
+▁tank -9.95323
+▁grant -9.95369
+▁wonder -9.95428
+▁introduced -9.95453
+▁prepare -9.95477
+▁myth -9.95483
+not -9.95525
+▁kur -9.95635
+▁vice -9.95641
+yard -9.95694
+▁settled -9.95722
+▁emerg -9.95973
+▁classical -9.95998
+▁kick -9.96016
+▁reduce -9.96141
+▁rod -9.96148
+▁bake -9.96333
+board -9.96397
+▁ur -9.96427
+nja -9.96517
+▁danger -9.9663
+▁battery -9.96643
+▁ludwig -9.96643
+▁hockey -9.96643
+▁unlike -9.96643
+▁chicken -9.96644
+▁error -9.96645
+▁partially -9.96646
+▁housing -9.96646
+stone -9.96692
+▁associated -9.96714
+▁loose -9.9673
+▁disc -9.96817
+▁lying -9.96835
+▁installed -9.96873
+endorf -9.96947
+▁expanded -9.96965
+▁lin -9.97042
+▁composed -9.97095
+berries -9.97141
+▁rise -9.97148
+▁recordings -9.97296
+▁mad -9.97305
+▁till -9.97358
+▁apply -9.97402
+▁residents -9.97469
+▁cri -9.97478
+tag -9.97554
+▁characters -9.97584
+▁ms -9.97772
+▁shooting -9.97877
+▁corn -9.97906
+▁citizen -9.97907
+▁suffer -9.97947
+▁representatives -9.9796
+▁interview -9.97994
+▁negative -9.97994
+▁possibility -9.97994
+▁sufficient -9.97994
+▁impact -9.97995
+▁reference -9.97995
+▁hidden -9.97995
+▁address -9.97995
+▁audio -9.97995
+▁christmas -9.97997
+▁strict -9.97998
+ola -9.98172
+▁semi -9.9818
+▁attached -9.98259
+▁gre -9.98266
+▁fried -9.98324
+▁matches -9.98341
+▁alive -9.9848
+▁den -9.98574
+▁begins -9.98592
+▁rust -9.98823
+▁canton -9.9903
+gar -9.99045
+▁ham -9.99213
+▁drama -9.99335
+▁discover -9.99347
+▁channel -9.99364
+▁somewhat -9.99364
+▁stroke -9.99364
+▁crazy -9.99365
+▁giving -9.99365
+▁diving -9.99366
+▁dutch -9.99367
+▁agency -9.99367
+▁rural -9.99367
+▁launch -9.99372
+▁racing -9.99373
+▁aircraft -9.99382
+▁weapons -9.99404
+▁finance -9.99413
+▁valid -9.99419
+▁approved -9.99465
+▁statu -9.99468
+▁egg -9.99482
+▁witt -9.99507
+bed -9.99509
+▁online -9.99542
+▁laid -9.9959
+mail -9.99689
+▁lind -9.99717
+▁chess -9.99771
+▁magic -9.99789
+▁fully -9.9984
+▁photos -9.99841
+▁thinking -10
+▁eli -10.0025
+ule -10.0063
+▁occasion -10.0072
+▁descend -10.0075
+▁australia -10.0075
+▁comparison -10.0075
+▁heavily -10.0075
+▁hungry -10.0075
+▁locomotive -10.0075
+▁possession -10.0075
+▁potential -10.0075
+▁responsibility -10.0075
+▁revolution -10.0075
+▁assistant -10.0075
+▁defeat -10.0075
+▁beyond -10.0075
+▁related -10.0075
+▁pilot -10.0076
+▁felt -10.0076
+▁failure -10.0077
+▁unclear -10.008
+▁safety -10.008
+▁massive -10.0081
+▁burned -10.0083
+▁knee -10.0083
+▁clothes -10.0089
+▁helped -10.0092
+▁india -10.0094
+▁converted -10.0095
+▁sent -10.0096
+▁dig -10.0102
+press -10.0113
+mel -10.0118
+▁rent -10.0126
+▁vote -10.0127
+▁catch -10.0128
+▁critical -10.0129
+▁cry -10.0131
+▁details -10.0143
+ella -10.0171
+▁holidays -10.0182
+▁soldiers -10.0182
+▁micro -10.0187
+▁southeast -10.0213
+▁bud -10.0214
+▁publish -10.0214
+▁advice -10.0216
+▁bundesliga -10.0216
+▁challenge -10.0216
+▁component -10.0216
+▁emperor -10.0216
+▁imperial -10.0216
+▁oblig -10.0216
+▁applies -10.0216
+zzle -10.0217
+▁gym -10.0217
+▁forgotten -10.0219
+▁advertisement -10.0219
+▁broad -10.022
+▁stuck -10.0223
+▁decide -10.0225
+▁rubb -10.0226
+▁burst -10.0227
+▁carbon -10.023
+▁derived -10.024
+▁jean -10.0241
+▁hart -10.0252
+▁arrived -10.0253
+tory -10.0254
+▁hau -10.0255
+here -10.0265
+▁conductor -10.0266
+▁swed -10.0273
+▁participants -10.0278
+▁sell -10.0278
+▁maintain -10.0283
+▁lion -10.0289
+▁commander -10.0297
+▁actual -10.0318
+▁constitution -10.0319
+▁participate -10.0319
+ani -10.0333
+enbach -10.0337
+▁historic -10.0346
+▁mel -10.0347
+▁tape -10.0349
+▁accepted -10.0349
+▁mole -10.0351
+▁holy -10.0356
+▁academic -10.0359
+▁occupied -10.0359
+▁regulation -10.0359
+▁review -10.0359
+▁sweat -10.0359
+▁resistance -10.036
+▁caught -10.036
+▁cheese -10.036
+▁transferred -10.0361
+▁comfortable -10.0361
+▁shift -10.0361
+▁corps -10.0361
+▁architecture -10.0362
+▁predominantly -10.0366
+▁swim -10.0367
+▁polar -10.0376
+etic -10.038
+▁none -10.0388
+▁hof -10.0389
+▁reduced -10.0413
+ily -10.0422
+▁donation -10.0423
+ana -10.0434
+▁indicate -10.0454
+▁ruler -10.047
+hagen -10.0478
+▁zi -10.048
+▁rum -10.0488
+▁interesting -10.0493
+▁exhibit -10.0494
+▁grab -10.05
+▁relevant -10.0504
+▁tissue -10.0504
+▁crowd -10.0504
+▁electronic -10.0504
+▁remind -10.0505
+▁legend -10.0505
+▁transition -10.0506
+▁murder -10.0506
+▁proof -10.0507
+▁shoes -10.0528
+▁brain -10.0539
+▁told -10.054
+▁cleaning -10.0549
+▁nest -10.055
+zo -10.0557
+▁colored -10.056
+▁engineering -10.0562
+▁attend -10.0577
+▁alt -10.0588
+▁neo -10.0597
+▁manag -10.0599
+▁que -10.062
+east -10.0647
+▁fundamental -10.0651
+▁gabriel -10.0651
+▁asteroid -10.0651
+▁collapse -10.0651
+▁families -10.0651
+▁truck -10.0651
+▁fourteen -10.0651
+▁transfer -10.0652
+▁relativ -10.0653
+▁alex -10.0654
+▁spark -10.0654
+▁familiar -10.0657
+▁mag -10.0659
+▁priest -10.0661
+▁radi -10.0662
+▁wald -10.0689
+▁socialist -10.0696
+▁grass -10.0709
+▁westphalia -10.0714
+▁bavarian -10.0722
+▁pitch -10.0722
+▁stable -10.0732
+▁herb -10.0739
+▁trail -10.0768
+▁entirely -10.0785
+▁sick -10.0792
+▁formula -10.0793
+▁initiative -10.08
+▁journey -10.08
+▁progress -10.08
+▁silence -10.08
+▁variety -10.08
+▁precise -10.08
+▁april -10.08
+▁intended -10.08
+▁rhein -10.08
+▁witness -10.08
+▁adjust -10.08
+▁flood -10.08
+▁consume -10.0802
+▁sail -10.0813
+▁wal -10.082
+▁finger -10.0823
+▁hide -10.0827
+▁grain -10.0833
+▁survive -10.0841
+▁bau -10.0847
+▁clock -10.0853
+▁constructed -10.0856
+war -10.0857
+day -10.0858
+hood -10.0868
+▁disco -10.0869
+▁correctly -10.0876
+▁asia -10.0881
+▁dub -10.0881
+ische -10.0885
+▁faster -10.0895
+▁trial -10.0908
+▁operate -10.0913
+▁survived -10.0919
+▁jam -10.0935
+lingen -10.0936
+▁interrupt -10.0952
+▁chocolate -10.0952
+▁grandfather -10.0952
+▁ministry -10.0952
+▁yeah -10.0952
+▁choose -10.0952
+▁screw -10.0952
+▁quiet -10.0952
+▁storage -10.0952
+▁annual -10.0952
+▁annoy -10.0952
+▁repeat -10.0952
+▁justice -10.0953
+▁sunday -10.0954
+▁brun -10.0954
+▁dangerous -10.0955
+▁publications -10.0955
+▁underground -10.0957
+▁successfully -10.0961
+▁growth -10.0963
+▁lamp -10.0981
+▁branches -10.0984
+▁anton -10.0994
+▁rice -10.0995
+▁worse -10.0995
+▁reported -10.0997
+▁signed -10.1034
+▁accordingly -10.1037
+▁george -10.1051
+▁strength -10.1073
+▁calendar -10.1106
+▁dimension -10.1106
+▁extension -10.1106
+▁vegetable -10.1106
+▁silent -10.1106
+▁kingdom -10.1106
+▁decline -10.1106
+▁sauce -10.1106
+▁moving -10.1106
+▁instance -10.1106
+▁physical -10.1106
+▁tonight -10.1106
+▁bicycle -10.1107
+▁freiburg -10.1107
+off -10.1107
+▁accessible -10.1107
+▁pond -10.1108
+▁investment -10.1108
+▁tight -10.111
+▁intention -10.1111
+▁peak -10.1112
+▁zu -10.1116
+▁crew -10.1116
+▁ob -10.1121
+▁fear -10.1124
+▁wie -10.1126
+▁genre -10.1128
+▁directed -10.113
+▁impression -10.1143
+▁lang -10.1145
+ably -10.1176
+▁bonn -10.1178
+acti -10.118
+▁publisher -10.1181
+▁cow -10.1188
+▁emerged -10.1194
+▁ordered -10.1195
+▁processes -10.1197
+▁server -10.1208
+▁cro -10.1217
+ura -10.1232
+▁cord -10.1244
+of -10.125
+▁relation -10.1261
+▁affair -10.1262
+▁alcohol -10.1262
+▁brigade -10.1262
+▁confused -10.1262
+▁conversation -10.1262
+▁cooperation -10.1262
+▁diocese -10.1262
+▁recognition -10.1262
+▁changing -10.1262
+▁enemy -10.1262
+▁translation -10.1262
+▁remote -10.1262
+▁sixteen -10.1262
+▁manual -10.1266
+haps -10.1266
+▁fifth -10.1267
+▁workers -10.1268
+▁approximately -10.1269
+▁magnetic -10.1269
+▁contribution -10.1275
+▁fever -10.1276
+▁hop -10.1285
+▁toilet -10.1286
+▁powerful -10.129
+▁sar -10.1292
+▁grey -10.1307
+▁produce -10.1329
+▁sheet -10.133
+▁spoken -10.1332
+▁belly -10.1335
+▁normally -10.1343
+▁hack -10.1343
+▁protected -10.1344
+▁maintained -10.1346
+▁separated -10.1365
+▁achieve -10.137
+▁scha -10.1377
+▁hesse -10.1409
+back -10.1417
+▁hub -10.1419
+▁calculate -10.1421
+▁controversial -10.1421
+▁identity -10.1421
+▁simultaneously -10.1421
+▁drunk -10.1421
+▁jazz -10.1421
+▁christoph -10.1421
+▁column -10.1421
+▁properties -10.1421
+▁wedding -10.1421
+▁stood -10.1421
+▁honest -10.1421
+▁requirements -10.1421
+▁wealth -10.1421
+▁politically -10.1423
+▁alongside -10.1425
+rald -10.1425
+▁assembly -10.1429
+▁fred -10.1434
+▁prussian -10.1447
+▁processing -10.1448
+▁strongly -10.1457
+▁manner -10.1458
+▁pastor -10.1464
+wise -10.1469
+▁bull -10.1472
+ux -10.15
+▁const -10.1529
+▁promote -10.1532
+▁settle -10.1535
+jo -10.1539
+▁subsequent -10.154
+▁saxon -10.1544
+▁hind -10.1567
+▁differ -10.1582
+▁episode -10.1582
+▁resources -10.1582
+▁dresden -10.1582
+▁profile -10.1582
+▁labor -10.1582
+▁swimming -10.1582
+▁tubingen -10.1582
+▁klaus -10.1582
+▁forgot -10.1584
+▁israel -10.1586
+▁richard -10.1586
+▁compete -10.1587
+▁bremen -10.1587
+▁walter -10.1589
+▁cancer -10.1589
+▁lung -10.1593
+▁din -10.1593
+▁highway -10.1599
+▁curve -10.1601
+▁defined -10.1604
+▁spain -10.1609
+▁breed -10.1612
+▁obtained -10.1625
+asse -10.1638
+▁chest -10.1647
+▁mary -10.165
+haus -10.1669
+▁corresponding -10.1689
+mmy -10.1693
+▁suspect -10.1746
+▁assigned -10.1746
+▁conclusion -10.1746
+▁decisive -10.1746
+▁technique -10.1746
+▁amazing -10.1746
+▁relax -10.1746
+▁freedom -10.1746
+▁inherit -10.1746
+▁mold -10.1753
+vin -10.1754
+▁sacr -10.1756
+▁experienced -10.1759
+▁realized -10.176
+▁lens -10.1773
+▁shower -10.1785
+▁organic -10.1799
+▁whom -10.1802
+▁sieg -10.181
+▁lip -10.1813
+▁northwest -10.1821
+▁pai -10.1824
+▁engineer -10.1826
+▁russia -10.1854
+▁wil -10.1858
+ima -10.1859
+▁cancel -10.1862
+▁stove -10.187
+▁dude -10.1873
+▁gall -10.1878
+▁closely -10.189
+▁cam -10.1894
+▁surviv -10.1904
+▁virtual -10.1912
+▁circula -10.1912
+▁description -10.1912
+▁distributed -10.1912
+▁oxygen -10.1912
+▁platform -10.1912
+▁predecessor -10.1912
+▁supplies -10.1912
+▁czech -10.1912
+▁meadow -10.1912
+▁resign -10.1912
+▁david -10.1913
+▁sharp -10.1913
+▁advise -10.1913
+▁appointment -10.1913
+▁barrel -10.1913
+fall -10.1914
+▁income -10.1915
+▁relief -10.1917
+▁schon -10.1925
+▁loop -10.193
+▁adopted -10.1936
+fish -10.1938
+▁crack -10.1938
+bling -10.1946
+▁villa -10.1948
+▁discussed -10.1948
+▁electrical -10.1948
+▁chip -10.1961
+post -10.1991
+ju -10.202
+▁meter -10.2024
+kirch -10.2036
+▁raw -10.2064
+▁refu -10.2074
+▁environmental -10.2074
+▁spirit -10.2082
+▁dusseldorf -10.2082
+▁facility -10.2082
+▁incorrect -10.2082
+▁instruction -10.2082
+▁reception -10.2082
+▁rhythm -10.2082
+▁zurich -10.2082
+▁junior -10.2082
+▁deserve -10.2082
+▁faith -10.2082
+▁unique -10.2082
+▁trigger -10.2082
+▁crystal -10.2082
+▁ernst -10.2082
+▁phase -10.2082
+▁konig -10.2082
+▁sky -10.2083
+▁brandenburg -10.2084
+▁pedal -10.2085
+▁debt -10.2087
+▁poetry -10.2088
+▁painted -10.209
+▁exam -10.2092
+▁bow -10.2096
+▁existed -10.2103
+▁craft -10.2105
+▁acquired -10.2106
+▁kre -10.2111
+▁gross -10.2115
+▁constitutional -10.2129
+▁employee -10.2163
+▁bite -10.2184
+▁proven -10.219
+▁independently -10.2198
+▁proper -10.2222
+▁button -10.2239
+rious -10.2242
+▁stein -10.225
+▁advertising -10.2254
+▁alexander -10.2254
+▁combination -10.2254
+▁gymnasium -10.2254
+▁selected -10.2254
+▁plastic -10.2254
+▁fridge -10.2254
+▁idiot -10.2255
+▁gothic -10.2255
+▁committed -10.2257
+▁tube -10.2257
+▁installation -10.2258
+▁logic -10.2263
+▁theatre -10.2264
+▁latest -10.2265
+▁pump -10.2267
+tional -10.227
+▁fixed -10.2299
+well -10.2306
+▁carefully -10.2317
+▁tall -10.2352
+▁damaged -10.2364
+▁african -10.2372
+▁eaten -10.2383
+▁pole -10.2385
+▁marie -10.2412
+▁cloth -10.242
+▁cent -10.2423
+▁uncle -10.2424
+▁pap -10.2427
+▁abandoned -10.243
+▁diploma -10.243
+▁memorial -10.243
+▁sculpture -10.243
+▁temporarily -10.243
+▁client -10.243
+▁credit -10.243
+▁cinema -10.243
+▁snake -10.243
+▁harbor -10.243
+▁munster -10.243
+▁website -10.243
+▁dropped -10.243
+▁nazi -10.243
+▁reign -10.243
+▁apprenticeship -10.243
+▁opposition -10.243
+▁solid -10.2432
+astic -10.2432
+▁workshop -10.2433
+▁guess -10.2433
+▁wound -10.2434
+▁fold -10.2437
+▁subsidi -10.2437
+ologist -10.244
+▁string -10.2441
+▁stamp -10.2446
+▁lady -10.2447
+▁shall -10.2459
+▁sweden -10.2471
+▁slipp -10.2472
+▁cash -10.2483
+▁hunting -10.2486
+▁glasses -10.2486
+▁file -10.2487
+▁rad -10.25
+▁height -10.2506
+▁begin -10.2533
+▁carl -10.2541
+▁drawing -10.2547
+▁sad -10.2549
+▁concerned -10.256
+ground -10.2562
+▁initial -10.2568
+▁cheat -10.257
+▁recording -10.2573
+▁dro -10.2595
+▁nation -10.2598
+▁japan -10.2608
+▁ultimate -10.2608
+▁discipline -10.2608
+▁eagle -10.2608
+▁estimate -10.2608
+▁identical -10.2608
+▁identified -10.2608
+▁incident -10.2608
+▁occupation -10.2608
+▁philipp -10.2608
+▁phrase -10.2608
+▁presence -10.2608
+▁scratch -10.2608
+▁smartphone -10.2608
+▁britain -10.2608
+▁victory -10.2608
+▁slope -10.2608
+▁bodies -10.2609
+▁flies -10.2609
+▁monitor -10.261
+▁integrated -10.261
+▁devote -10.2611
+▁municipalities -10.2611
+▁potatoes -10.2617
+▁lift -10.2617
+▁interpretation -10.2617
+▁poly -10.2621
+▁fake -10.2626
+▁valve -10.263
+▁mister -10.2634
+▁fighting -10.2636
+▁sex -10.2636
+▁solve -10.2654
+▁enable -10.2656
+▁passion -10.266
+▁format -10.2671
+▁ol -10.2675
+▁frau -10.2721
+▁mining -10.2757
+▁existing -10.2784
+▁abbreviation -10.279
+▁ceremony -10.279
+▁executed -10.279
+▁maximum -10.279
+▁qualified -10.279
+▁railroad -10.279
+▁wurttemberg -10.279
+▁asleep -10.279
+▁parallel -10.279
+▁freelance -10.279
+▁fortress -10.279
+▁siblings -10.279
+▁pirate -10.279
+▁steep -10.279
+▁comprise -10.2791
+▁armed -10.2793
+▁korea -10.2794
+▁dramatic -10.2794
+▁hal -10.2794
+▁hence -10.2798
+▁equation -10.2799
+ight -10.2806
+▁sie -10.2807
+▁conventional -10.2808
+▁delivery -10.2816
+▁seal -10.2826
+▁declared -10.2834
+▁mai -10.2842
+poli -10.2843
+▁lib -10.285
+▁complet -10.2856
+itive -10.2869
+▁weigh -10.2871
+▁dan -10.2887
+▁toast -10.2891
+▁bath -10.2891
+▁classes -10.2897
+▁drawn -10.2899
+▁bush -10.2901
+▁swing -10.2905
+▁earl -10.2907
+▁practiced -10.2912
+▁improve -10.2925
+▁score -10.2925
+▁cave -10.2926
+▁forced -10.297
+▁consumption -10.2975
+▁habitat -10.2975
+▁response -10.2975
+▁trunk -10.2975
+▁garbage -10.2975
+▁senior -10.2975
+▁portrait -10.2975
+▁thread -10.2975
+▁regiment -10.2976
+▁mount -10.2976
+platz -10.2978
+▁secretary -10.2987
+▁encouraged -10.2987
+▁boil -10.2991
+▁representation -10.2993
+▁oven -10.2995
+enheim -10.2996
+▁albert -10.2997
+▁shore -10.3002
+▁angle -10.3012
+▁distinguished -10.3015
+▁johannes -10.3018
+▁southwest -10.3018
+▁secured -10.3018
+▁kro -10.3019
+▁confirmed -10.302
+▁boiler -10.3031
+▁accused -10.3037
+jan -10.3044
+low -10.3047
+▁agreed -10.306
+mund -10.3077
+▁interpret -10.3079
+▁rap -10.312
+▁leon -10.312
+▁demo -10.3125
+▁evaluat -10.3154
+▁argument -10.3164
+▁engaged -10.3164
+▁imagine -10.3164
+▁knife -10.3164
+▁medieval -10.3164
+▁unusual -10.3164
+▁torture -10.3164
+▁gottingen -10.3164
+▁baking -10.3165
+▁plot -10.3165
+▁debut -10.3165
+▁engel -10.3171
+▁gain -10.3175
+▁slid -10.3178
+▁bigger -10.3178
+▁atmosphere -10.3183
+▁combined -10.3193
+▁missed -10.3198
+▁soft -10.3198
+▁deco -10.3207
+▁observed -10.3229
+ville -10.3233
+▁cul -10.3249
+hand -10.3268
+▁mental -10.3291
+▁arrive -10.3306
+▁surprise -10.3306
+▁northeast -10.333
+▁fift -10.3354
+▁advisor -10.3356
+▁efficiency -10.3356
+▁excellen -10.3356
+▁exposed -10.3356
+▁significance -10.3356
+▁thereafter -10.3356
+▁disturb -10.3356
+▁wagner -10.3356
+▁promotion -10.3356
+▁afternoon -10.3356
+▁stefan -10.3357
+▁tough -10.3357
+▁dishes -10.3357
+▁backwards -10.3357
+▁farming -10.3358
+▁minim -10.336
+▁frei -10.3362
+▁reader -10.3365
+▁controlled -10.3365
+▁effective -10.3368
+▁extra -10.3402
+▁hip -10.3404
+▁reache -10.3405
+uous -10.3406
+▁scored -10.3409
+▁indicated -10.3424
+win -10.3428
+▁dur -10.3444
+▁profession -10.3458
+▁journal -10.3472
+▁reject -10.3477
+▁nie -10.3479
+▁climbing -10.3492
+▁ji -10.3495
+rising -10.35
+friend -10.352
+▁wire -10.3542
+▁mono -10.3544
+▁proceed -10.355
+▁absorb -10.3552
+necessarily -10.3552
+▁accompanied -10.3552
+▁augsburg -10.3552
+▁mediterranean -10.3552
+▁populated -10.3552
+▁injuries -10.3552
+▁therapy -10.3553
+▁plum -10.3553
+▁illness -10.3553
+▁birthday -10.3553
+box -10.3557
+▁dumb -10.3564
+▁considerable -10.3567
+▁jack -10.3568
+▁stir -10.3584
+▁prohibited -10.3594
+▁dense -10.3601
+▁constantly -10.3626
+▁separately -10.366
+▁mention -10.3678
+▁advance -10.3702
+▁collected -10.3707
+ault -10.3713
+▁surely -10.3717
+▁tear -10.3726
+▁privately -10.3732
+▁aggressive -10.3752
+▁educated -10.3752
+▁infantry -10.3752
+▁neglect -10.3752
+▁nervous -10.3752
+▁pleasure -10.3752
+▁vacuum -10.3752
+▁widespread -10.3752
+▁illegal -10.3752
+▁brazil -10.3752
+▁retired -10.3752
+▁funeral -10.3753
+▁behave -10.3753
+▁fiction -10.3753
+▁tomato -10.3753
+▁lyric -10.3753
+▁eighteen -10.3753
+▁mineral -10.3754
+term -10.3755
+▁deutsche -10.3755
+▁aware -10.3755
+▁rapid -10.3756
+▁leadership -10.3758
+▁slim -10.3758
+▁crash -10.3762
+▁olive -10.3762
+▁altar -10.3762
+▁worst -10.3763
+▁grape -10.3763
+▁grandma -10.3765
+▁seventy -10.3768
+▁reformation -10.3771
+▁navy -10.3775
+▁arrested -10.379
+▁rein -10.3799
+▁properly -10.3801
+▁lecturer -10.3802
+▁graz -10.3805
+kow -10.3809
+▁becoming -10.3825
+▁val -10.383
+▁ego -10.3832
+▁successes -10.3848
+▁bowl -10.3857
+▁businesses -10.3864
+hard -10.3865
+▁construct -10.3871
+▁dar -10.3879
+▁extend -10.3881
+▁fail -10.3892
+▁bond -10.394
+water -10.3954
+▁autonomous -10.3957
+▁celebrated -10.3957
+▁confess -10.3957
+▁congress -10.3957
+▁delicious -10.3957
+▁executive -10.3957
+▁october -10.3957
+▁password -10.3957
+▁qualification -10.3957
+▁telephone -10.3957
+▁ignore -10.3957
+▁welcome -10.3957
+▁curious -10.3957
+▁hardware -10.3957
+▁literally -10.3957
+▁vacation -10.3957
+▁width -10.3957
+▁shield -10.3957
+▁nineteen -10.3957
+▁saving -10.3957
+▁violin -10.3957
+▁amateur -10.3957
+▁hiking -10.3957
+▁muller -10.3957
+▁moderate -10.3957
+▁mood -10.3958
+▁borrow -10.3959
+▁operator -10.396
+▁fault -10.396
+▁admitted -10.3965
+▁archive -10.3966
+▁radiation -10.3969
+▁treatment -10.397
+▁socialism -10.3972
+▁versa -10.3987
+▁hunt -10.399
+coming -10.4002
+▁math -10.4013
+▁delivered -10.4014
+▁vary -10.4023
+▁zo -10.4026
+▁tub -10.4053
+▁loudly -10.4054
+▁suffering -10.4056
+▁healthy -10.4062
+▁alb -10.408
+▁emil -10.4081
+bahn -10.4099
+▁fighter -10.4124
+▁recover -10.4127
+▁sam -10.4127
+break -10.4132
+right -10.4133
+▁zoo -10.4154
+▁parent -10.4154
+▁fate -10.4158
+gau -10.4163
+▁whisper -10.4165
+▁approval -10.4165
+▁certificate -10.4165
+▁communist -10.4165
+▁convince -10.4165
+▁exercise -10.4165
+▁expansion -10.4165
+▁gustav -10.4165
+▁hannover -10.4165
+▁merchant -10.4165
+▁netherlands -10.4165
+▁purchase -10.4165
+▁rudolf -10.4165
+▁afford -10.4165
+▁external -10.4165
+▁monday -10.4165
+▁visual -10.4165
+▁palace -10.4165
+▁plug -10.4165
+▁false -10.4165
+▁physics -10.4165
+▁awake -10.4165
+▁hollow -10.4166
+▁classification -10.417
+▁occurred -10.417
+▁observation -10.4173
+▁venu -10.4174
+▁blade -10.4178
+▁plai -10.4182
+▁prior -10.4185
+▁gang -10.4189
+▁invited -10.4207
+bell -10.4212
+kirche -10.4245
+▁interpreted -10.425
+rank -10.4321
+▁bark -10.4323
+▁hermann -10.433
+▁moor -10.4341
+bro -10.4342
+▁dispos -10.4351
+▁gap -10.4361
+▁weapon -10.4371
+▁protest -10.4375
+▁acoustic -10.4378
+▁centuries -10.4378
+▁cigarette -10.4378
+▁circumstances -10.4378
+▁domestic -10.4378
+▁enormous -10.4378
+▁mirror -10.4378
+▁resolution -10.4378
+▁subordinate -10.4378
+▁syndrome -10.4378
+▁terror -10.4378
+▁yourselves -10.4378
+▁athlete -10.4378
+▁grammar -10.4378
+▁spaghetti -10.4378
+▁uniform -10.4378
+▁intensive -10.4378
+▁muscle -10.4378
+▁cruise -10.4378
+▁motivat -10.4378
+▁evil -10.4378
+▁wagon -10.4378
+▁exceed -10.4378
+▁permitted -10.438
+▁input -10.4381
+▁remarkable -10.4385
+band -10.4386
+▁specialized -10.4394
+▁fox -10.4394
+▁angeles -10.4395
+▁passage -10.4396
+▁expl -10.4407
+▁ferry -10.4409
+flow -10.441
+▁dirty -10.4412
+▁generated -10.4414
+▁criticized -10.4414
+▁examined -10.4415
+▁setting -10.4421
+▁strengthen -10.4423
+neck -10.4442
+▁taxi -10.4462
+▁breeding -10.4467
+phon -10.4483
+witz -10.4488
+▁parliamentar -10.4492
+▁secure -10.4545
+bank -10.4563
+▁alter -10.4565
+▁advertise -10.4591
+▁category -10.4595
+▁colloquial -10.4595
+▁concentrate -10.4595
+▁conservative -10.4595
+▁cylinder -10.4595
+▁decrease -10.4595
+▁desperate -10.4595
+▁elaborate -10.4595
+▁employment -10.4595
+▁encounter -10.4595
+▁evangelical -10.4595
+▁francis -10.4595
+▁frequency -10.4595
+▁gradually -10.4595
+▁intelligence -10.4595
+▁thieves -10.4595
+▁voltage -10.4595
+▁excited -10.4595
+▁servant -10.4595
+▁jurgen -10.4595
+▁patron -10.4595
+▁quant -10.4595
+▁beauty -10.4595
+▁firm -10.4595
+▁pacific -10.4595
+▁antenna -10.4595
+▁nerve -10.4595
+▁ensure -10.4597
+▁elector -10.4599
+▁burial -10.4601
+▁slic -10.4601
+▁preparation -10.4602
+▁coins -10.4605
+▁james -10.4622
+▁shade -10.4633
+▁kiel -10.4634
+▁spur -10.4638
+▁pose -10.4652
+▁researchers -10.4657
+▁rebel -10.4668
+▁mast -10.4687
+▁lent -10.4694
+▁laz -10.4696
+▁traumat -10.4711
+▁angel -10.4715
+▁buch -10.4724
+▁observ -10.4732
+▁establish -10.4737
+lude -10.4744
+▁arab -10.4766
+▁equat -10.4807
+▁impress -10.4807
+▁consequent -10.4817
+▁embarrass -10.4817
+▁comprehensive -10.4817
+▁deployed -10.4817
+▁joachim -10.4817
+▁politician -10.4817
+▁possibly -10.4817
+▁rebuilt -10.4817
+▁scatter -10.4817
+▁selection -10.4817
+▁wolfgang -10.4817
+▁abbey -10.4817
+▁ceiling -10.4817
+▁variant -10.4817
+▁zealand -10.4817
+▁liberal -10.4817
+▁entry -10.4818
+▁mouse -10.4818
+▁golf -10.4818
+▁happiness -10.4818
+▁banned -10.4819
+▁shaft -10.4819
+▁tiny -10.482
+▁drag -10.4821
+▁establishment -10.4822
+▁ginger -10.4822
+▁aachen -10.4822
+▁grandpa -10.4824
+▁charts -10.4825
+▁manufactured -10.4833
+▁rumor -10.4858
+charging -10.486
+▁sixth -10.487
+▁calm -10.4876
+▁belt -10.488
+▁implemented -10.488
+▁melt -10.4889
+▁santa -10.4904
+▁pf -10.4924
+▁hunter -10.495
+▁shake -10.4965
+▁affect -10.4969
+▁sale -10.4985
+enstein -10.5006
+play -10.5016
+▁mile -10.5024
+pleasant -10.5045
+▁aluminum -10.5045
+▁anniversary -10.5045
+▁concentration -10.5045
+▁demonstrate -10.5045
+▁ingredient -10.5045
+▁psychology -10.5045
+▁respond -10.5045
+▁sacrific -10.5045
+▁synonym -10.5045
+▁thirteen -10.5045
+▁fragment -10.5045
+▁indicator -10.5045
+▁advocate -10.5045
+▁imagination -10.5045
+▁shadow -10.5045
+▁indication -10.5045
+▁scholar -10.5045
+▁pupil -10.5045
+▁sword -10.5045
+▁stink -10.5045
+▁weimar -10.5045
+▁crisis -10.5045
+▁delay -10.5045
+▁dried -10.5046
+▁rescue -10.5046
+▁crook -10.5046
+▁index -10.5046
+▁invented -10.5049
+▁assess -10.5049
+▁radical -10.5049
+▁nearby -10.5052
+▁hype -10.5054
+▁rush -10.5063
+▁investigation -10.5064
+▁rocket -10.5067
+bruck -10.5073
+▁armor -10.5074
+▁poland -10.5078
+▁continuously -10.5081
+▁allerg -10.5083
+▁proportion -10.5085
+▁recovery -10.5087
+▁refused -10.5087
+▁boxes -10.5096
+▁trailer -10.5114
+▁seeds -10.5114
+▁dust -10.5133
+lich -10.5134
+▁graf -10.5152
+▁lamb -10.5156
+▁wu -10.5188
+▁ost -10.5208
+▁compos -10.5236
+▁import -10.5273
+▁centimeters -10.5277
+▁circuit -10.5277
+▁criteria -10.5277
+▁danube -10.5277
+▁deputy -10.5277
+▁employer -10.5277
+▁facade -10.5277
+▁infection -10.5277
+▁inhabited -10.5277
+▁navigation -10.5277
+▁portuguese -10.5277
+▁substitute -10.5277
+▁draft -10.5277
+▁pursued -10.5277
+▁recipe -10.5277
+▁tyrol -10.5277
+schaft -10.5277
+▁cabinet -10.5277
+▁century -10.5277
+▁worldwide -10.5277
+▁mobile -10.5277
+▁existence -10.5277
+▁pianist -10.5277
+▁crucial -10.5277
+▁kreuz -10.5277
+▁harvest -10.5277
+▁resort -10.5277
+▁extent -10.5277
+▁copies -10.5277
+▁altitude -10.5277
+▁crawl -10.5278
+▁werner -10.5278
+▁comic -10.5278
+▁trend -10.5278
+▁dozen -10.528
+▁adventure -10.528
+▁courtyard -10.5281
+▁glad -10.5282
+▁pepper -10.5283
+▁vocal -10.5286
+upon -10.5286
+▁giant -10.5289
+▁trash -10.5289
+▁duty -10.5294
+rose -10.5297
+▁laser -10.5301
+▁mason -10.5302
+▁hugo -10.5305
+▁doctorate -10.5308
+▁roughly -10.5314
+▁ripe -10.5316
+▁arc -10.5324
+▁alps -10.5326
+meister -10.5336
+▁tick -10.5338
+▁bench -10.534
+lena -10.5375
+▁brave -10.5402
+▁accord -10.545
+▁continent -10.5462
+iness -10.5464
+▁extract -10.5472
+▁dirt -10.548
+hop -10.5494
+▁commit -10.5511
+▁democracy -10.5515
+▁dissolved -10.5515
+▁eliminated -10.5515
+▁enterprise -10.5515
+▁jurisdiction -10.5515
+▁konrad -10.5515
+▁luxembourg -10.5515
+▁mathematics -10.5515
+▁mexico -10.5515
+▁motorcycle -10.5515
+▁preliminar -10.5515
+▁salzburg -10.5515
+▁september -10.5515
+▁thoroughly -10.5515
+▁compound -10.5515
+▁explanation -10.5515
+▁fashion -10.5515
+▁somebody -10.5515
+▁disgust -10.5515
+▁throne -10.5515
+▁frog -10.5515
+▁injury -10.5515
+▁sparse -10.5515
+▁honey -10.5515
+▁ninth -10.5515
+▁carpet -10.5516
+▁shipping -10.5516
+▁barrier -10.5516
+▁reverse -10.5516
+▁canada -10.5517
+▁tennis -10.5517
+▁presum -10.5517
+▁consul -10.5518
+▁consol -10.5519
+▁garage -10.5522
+▁fluid -10.5522
+▁reconstruction -10.5522
+▁dragon -10.5523
+▁arabic -10.5528
+▁highlight -10.5528
+▁charter -10.5529
+▁gender -10.5531
+▁airline -10.5533
+▁romantic -10.5533
+▁oak -10.5545
+oxide -10.5547
+▁runner -10.5548
+maker -10.5553
+▁sommer -10.5555
+gard -10.5558
+▁solved -10.5583
+▁rub -10.5595
+▁nich -10.5621
+▁baron -10.5668
+▁pope -10.5672
+▁frequent -10.5718
+point -10.5723
+▁continu -10.5734
+▁basketball -10.5759
+▁capacity -10.5759
+▁execution -10.5759
+▁gymnastics -10.5759
+▁illustration -10.5759
+▁mecklenburg -10.5759
+▁possibilities -10.5759
+▁pseudonym -10.5759
+▁reservoir -10.5759
+▁scissors -10.5759
+▁skeleton -10.5759
+▁wednesday -10.5759
+▁assistance -10.5759
+▁diverse -10.5759
+▁throat -10.5759
+▁turkish -10.5759
+▁combat -10.5759
+▁ghost -10.5759
+▁stimula -10.5759
+▁chancellor -10.5759
+▁airplane -10.5759
+▁suburb -10.5759
+▁gallery -10.5759
+▁motif -10.5759
+▁density -10.5759
+▁senate -10.576
+▁denied -10.576
+▁clinic -10.5761
+▁dependent -10.5761
+▁confusion -10.5762
+▁fellow -10.5767
+▁extraordinary -10.5767
+▁determin -10.5767
+▁salad -10.5769
+▁hammer -10.577
+▁tribe -10.5777
+▁nave -10.5777
+▁basement -10.5781
+▁popularity -10.5783
+▁attributed -10.5791
+ladder -10.5795
+▁findings -10.5796
+plan -10.5811
+▁palm -10.5812
+▁holiday -10.5819
+▁advanced -10.5821
+▁soldier -10.5824
+▁decid -10.583
+▁horn -10.5831
+mination -10.588
+▁rie -10.5883
+▁mechanical -10.5895
+▁scal -10.595
+▁obvious -10.5968
+▁digit -10.5968
+stream -10.5976
+▁advent -10.6009
+▁bankrupt -10.6009
+▁dynasty -10.6009
+▁efficient -10.6009
+▁goalkeeper -10.6009
+▁pronounce -10.6009
+▁supplement -10.6009
+▁suspension -10.6009
+▁thuringia -10.6009
+▁comparable -10.6009
+▁legislati -10.6009
+▁chapter -10.6009
+▁segment -10.6009
+▁parcel -10.6009
+▁architectural -10.6009
+▁cattle -10.6009
+▁repeatedly -10.601
+▁grim -10.601
+▁programming -10.601
+▁psycho -10.6011
+▁vision -10.6013
+▁orbit -10.6014
+▁distinction -10.6014
+▁baltic -10.6015
+▁grun -10.6016
+▁lutheran -10.6016
+▁twin -10.6016
+▁voluntarily -10.6016
+▁fiber -10.6018
+▁hohen -10.6028
+▁suggestion -10.603
+▁decorated -10.603
+▁moral -10.6031
+▁luther -10.6031
+▁payment -10.6037
+▁correction -10.604
+▁forestry -10.6043
+▁accuse -10.6045
+▁nail -10.6052
+▁peel -10.6075
+▁improved -10.6078
+▁chile -10.6123
+▁nau -10.6173
+▁divid -10.6192
+▁gravel -10.6206
+mpression -10.6215
+▁invite -10.6218
+field -10.6229
+▁magnet -10.6254
+▁hydro -10.6264
+▁introduct -10.6265
+▁territori -10.6265
+▁demolition -10.6265
+▁elisabeth -10.6265
+▁facebook -10.6265
+▁interaction -10.6265
+▁jersey -10.6265
+▁karlsruhe -10.6265
+▁moscow -10.6265
+▁opportunities -10.6265
+▁opportunity -10.6265
+▁peninsula -10.6265
+▁resemble -10.6265
+▁sequence -10.6265
+▁symptoms -10.6265
+▁triangle -10.6265
+▁unconscious -10.6265
+▁voluntary -10.6265
+▁regulate -10.6266
+▁suicide -10.6266
+▁underneath -10.6266
+▁adolf -10.6266
+▁islamic -10.6266
+▁penalty -10.6266
+▁rabbi -10.6266
+▁temple -10.6266
+▁frost -10.6266
+▁summit -10.6266
+▁pioneer -10.6266
+▁spoon -10.6266
+▁reliable -10.6266
+▁canadian -10.6266
+▁breast -10.6266
+▁grill -10.6266
+▁truly -10.6266
+▁bundestag -10.6266
+▁smart -10.6266
+▁expectations -10.6267
+▁ireland -10.6267
+▁intense -10.6267
+▁sample -10.6267
+▁consistent -10.6268
+▁bounc -10.6268
+▁arrangement -10.6268
+▁sustainable -10.6269
+▁klein -10.627
+▁brut -10.6272
+▁genetic -10.6272
+▁pause -10.6273
+▁glow -10.6274
+▁collective -10.6274
+▁ridge -10.6276
+▁adj -10.6277
+▁elder -10.628
+▁noticeable -10.6291
+▁insert -10.6299
+▁questionable -10.63
+▁update -10.6301
+▁seventh -10.6311
+▁float -10.6318
+▁marine -10.6318
+▁linear -10.6334
+▁tang -10.6338
+▁ferr -10.6355
+▁cage -10.6367
+▁pine -10.6382
+▁scr -10.6384
+▁continuous -10.649
+▁georgia -10.6491
+craft -10.6496
+▁rough -10.6511
+▁mala -10.6525
+▁accurate -10.6529
+▁anonymous -10.6529
+▁blossom -10.6529
+▁cultivated -10.6529
+▁definition -10.6529
+▁dortmund -10.6529
+▁grandparents -10.6529
+▁intelligent -10.6529
+▁norwegian -10.6529
+▁november -10.6529
+▁schmidt -10.6529
+▁schneider -10.6529
+▁signature -10.6529
+▁supervisor -10.6529
+▁neutral -10.6529
+▁spiritual -10.6529
+▁thumb -10.6529
+▁alfred -10.6529
+▁josef -10.6529
+▁lubeck -10.6529
+▁analyze -10.6529
+▁louis -10.6529
+▁vector -10.6529
+▁dancing -10.6529
+▁historian -10.6529
+▁goethe -10.6529
+▁weird -10.653
+▁particle -10.653
+▁rotation -10.653
+▁duct -10.653
+▁pink -10.653
+▁mask -10.6533
+▁alternat -10.6534
+violet -10.6534
+▁sheep -10.6535
+▁tension -10.6535
+▁hook -10.6539
+▁drone -10.6541
+▁muhl -10.6544
+▁items -10.655
+▁consideration -10.6557
+▁cab -10.6564
+▁losses -10.6577
+▁june -10.6579
+schedule -10.6602
+▁grade -10.6603
+scu -10.6666
+▁deliver -10.6685
+▁implement -10.6714
+▁succeed -10.6723
+▁disappear -10.6723
+▁compare -10.6749
+space -10.677
+▁encrypt -10.6789
+▁abolish -10.6799
+▁adolescent -10.6799
+▁conservation -10.6799
+▁helicopter -10.6799
+▁jakob -10.6799
+▁manufacturing -10.6799
+▁mathematical -10.6799
+▁nonsense -10.6799
+▁revenue -10.6799
+▁structural -10.6799
+▁transmission -10.6799
+▁violence -10.6799
+▁swabia -10.6799
+▁unfair -10.6799
+▁withdraw -10.6799
+▁curtain -10.6799
+▁variable -10.6799
+▁budget -10.6799
+▁retreat -10.6799
+▁pocket -10.6799
+▁capsule -10.6799
+▁optical -10.6799
+▁hunger -10.6799
+▁retain -10.6799
+▁bundes -10.6799
+▁lunch -10.6799
+▁mystery -10.6799
+▁pretend -10.6799
+▁humor -10.6799
+▁jelly -10.68
+▁linux -10.68
+▁loyal -10.6802
+▁thermal -10.6802
+▁referee -10.6803
+▁preferred -10.6803
+▁fritz -10.6804
+▁linguistic -10.6804
+▁merit -10.6804
+▁harmless -10.6806
+▁janni -10.6808
+▁skat -10.6808
+▁encryption -10.6809
+▁volcano -10.6816
+▁lemon -10.6821
+▁marburg -10.6824
+ji -10.6838
+▁fries -10.685
+▁bibl -10.6884
+stedt -10.6898
+▁kant -10.6915
+▁bang -10.6925
+▁seiz -10.6946
+leben -10.6976
+▁replac -10.7001
+▁fog -10.7023
+uction -10.7039
+▁measur -10.7053
+born -10.7064
+▁categories -10.7077
+▁divorce -10.7077
+▁forbidden -10.7077
+▁fracture -10.7077
+▁infrastructure -10.7077
+▁lieutenant -10.7077
+▁potsdam -10.7077
+▁quote -10.7077
+▁regret -10.7077
+▁sneeze -10.7077
+▁varieties -10.7077
+▁hobby -10.7077
+▁prevail -10.7077
+▁slept -10.7077
+▁dessert -10.7077
+▁influential -10.7077
+▁plaster -10.7077
+▁suprem -10.7077
+▁gesture -10.7077
+▁random -10.7077
+▁sexual -10.7077
+▁daniel -10.7077
+▁disabled -10.7077
+▁alliance -10.7077
+▁domain -10.7077
+▁bypass -10.7078
+▁turtle -10.7078
+▁crypt -10.7079
+▁denmark -10.7079
+▁poisonous -10.7079
+▁dialogue -10.708
+▁mild -10.708
+▁toxic -10.7081
+▁seminar -10.7082
+▁revis -10.7082
+▁priority -10.7082
+▁rattle -10.7082
+▁deadline -10.7085
+▁plateau -10.7085
+▁ulm -10.7085
+▁systematic -10.7087
+▁flesh -10.7087
+▁arise -10.7092
+▁fool -10.7093
+▁fame -10.7094
+▁upright -10.7098
+▁govern -10.7098
+▁delight -10.71
+▁worm -10.7147
+▁gaze -10.7151
+alogue -10.7158
+▁tide -10.7171
+schule -10.7215
+▁sang -10.7216
+holz -10.7232
+▁obtain -10.7286
+▁hike -10.7312
+▁examine -10.7326
+wood -10.733
+power -10.7331
+▁yu -10.7337
+tally -10.7342
+stand -10.7347
+pole -10.736
+▁assumption -10.7362
+▁belgium -10.7362
+▁chemistry -10.7362
+▁demolished -10.7362
+▁elevator -10.7362
+▁explosive -10.7362
+▁february -10.7362
+▁furniture -10.7362
+▁guarantee -10.7362
+▁heidelberg -10.7362
+▁illustrate -10.7362
+▁initiated -10.7362
+▁manuscript -10.7362
+▁negotiations -10.7362
+▁renowned -10.7362
+▁retirement -10.7362
+▁schleswig -10.7362
+▁struggle -10.7362
+▁suspended -10.7362
+▁temporary -10.7362
+▁thursday -10.7362
+▁vegetarian -10.7362
+▁displac -10.7362
+▁exciting -10.7362
+▁intervention -10.7362
+▁pizza -10.7362
+▁download -10.7362
+▁schiller -10.7362
+▁discrete -10.7362
+▁einstein -10.7362
+▁google -10.7362
+▁mixture -10.7362
+▁abroad -10.7362
+▁clown -10.7363
+▁renew -10.7363
+▁cherry -10.7363
+▁clerk -10.7363
+▁stolen -10.7363
+▁milan -10.7363
+▁blank -10.7363
+▁edible -10.7363
+▁friday -10.7363
+▁pinch -10.7364
+▁virus -10.7364
+▁reduction -10.7366
+▁mandate -10.7366
+▁airfield -10.7366
+▁basin -10.7367
+▁burden -10.7367
+▁egyptian -10.7367
+▁mamm -10.7369
+▁ownership -10.7369
+▁friendship -10.7371
+▁kohl -10.7373
+▁warned -10.7378
+▁heinz -10.7381
+▁aunt -10.7381
+▁mortal -10.7382
+▁cookies -10.7383
+▁invest -10.7385
+▁sink -10.7387
+▁erna -10.7388
+▁chef -10.7401
+nkel -10.7407
+cast -10.7413
+▁chop -10.7415
+▁prepar -10.7433
+▁curse -10.7464
+ngest -10.7495
+fusion -10.7537
+▁lend -10.759
+▁fix -10.7602
+▁generate -10.7606
+▁sudden -10.7615
+▁derive -10.7621
+course -10.7639
+▁egypt -10.7651
+▁abstract -10.7657
+▁algorithm -10.7657
+▁amazon -10.7657
+▁announcement -10.7657
+▁curiosity -10.7657
+▁excluded -10.7657
+▁ferdinand -10.7657
+▁grammatical -10.7657
+▁interface -10.7657
+▁metropoli -10.7657
+▁narrator -10.7657
+▁psychological -10.7657
+▁sensitive -10.7657
+▁florida -10.7657
+▁helmet -10.7657
+▁surrender -10.7657
+▁margaret -10.7657
+▁possesse -10.7657
+▁prague -10.7657
+▁translator -10.7657
+▁fossil -10.7657
+▁swamp -10.7657
+▁lorenz -10.7657
+▁subtitl -10.7657
+▁widow -10.7657
+▁greeting -10.7657
+▁pillar -10.7657
+▁portray -10.7657
+▁terribl -10.7657
+▁chalk -10.7657
+▁donkey -10.7657
+▁angry -10.7658
+▁carrot -10.7658
+▁queue -10.7658
+▁irish -10.7659
+▁trench -10.7659
+▁shark -10.7663
+▁export -10.7664
+▁assert -10.7664
+▁biological -10.7666
+▁attractive -10.7668
+▁salv -10.7671
+▁wagg -10.7672
+▁documentary -10.7673
+▁genus -10.7679
+gitte -10.7691
+▁moni -10.7701
+▁collector -10.7705
+▁pace -10.7747
+pell -10.7796
+grid -10.7833
+▁designat -10.7849
+▁syl -10.7849
+stairs -10.7864
+▁deposit -10.7918
+▁immediate -10.7918
+▁refuse -10.7938
+▁volcan -10.794
+▁squa -10.7944
+▁accompany -10.796
+▁bachelor -10.796
+▁overwhelm -10.796
+▁symphon -10.796
+satisfied -10.796
+▁assignment -10.796
+▁columbia -10.796
+▁conscience -10.796
+▁engagement -10.796
+▁ensemble -10.796
+▁franconia -10.796
+▁hometown -10.796
+▁punished -10.796
+▁societies -10.796
+▁suspicious -10.796
+▁swallow -10.796
+▁wurzburg -10.796
+▁dwarf -10.796
+▁emotional -10.796
+▁fabric -10.796
+▁complement -10.796
+▁diamond -10.796
+▁matrix -10.796
+▁landlord -10.796
+▁turkey -10.796
+▁melody -10.796
+▁ultra -10.796
+▁leather -10.796
+▁heritage -10.796
+▁fifteen -10.796
+▁kaiser -10.796
+▁ladies -10.796
+▁personnel -10.796
+▁scrap -10.796
+▁bathroom -10.796
+▁strawberr -10.7961
+▁quantities -10.7961
+▁vertical -10.7962
+▁soap -10.7962
+▁banana -10.7962
+▁resistant -10.7962
+▁stalk -10.7963
+▁insult -10.7963
+▁tractor -10.7964
+▁integration -10.7968
+▁beforehand -10.7968
+▁steak -10.797
+▁classified -10.797
+▁gunther -10.7971
+licate -10.7973
+▁insight -10.7974
+▁evaluation -10.7976
+▁implie -10.7978
+▁stability -10.7979
+enthal -10.8017
+finite -10.8032
+weiler -10.805
+▁symbolize -10.8096
+▁mechanic -10.81
+zig -10.8122
+▁consequence -10.8149
+▁styl -10.8154
+▁baker -10.8168
+▁confirm -10.8196
+▁declar -10.8196
+▁expand -10.8196
+▁poison -10.8269
+▁physi -10.8271
+▁distract -10.8272
+▁tuesday -10.8272
+▁acknowledge -10.8272
+▁bayreuth -10.8272
+▁calculation -10.8272
+▁destruction -10.8272
+▁fireplace -10.8272
+▁horror -10.8272
+▁independence -10.8272
+▁membrane -10.8272
+▁monarch -10.8272
+▁orthodox -10.8272
+▁pharmaceutic -10.8272
+▁renovation -10.8272
+▁spectator -10.8272
+▁virgin -10.8272
+▁contrary -10.8272
+▁medication -10.8272
+▁mozilla -10.8272
+▁powder -10.8272
+▁schwarz -10.8272
+▁secular -10.8272
+▁terrace -10.8272
+nforcement -10.8272
+▁miracle -10.8272
+▁theoretical -10.8272
+▁pencil -10.8272
+▁feather -10.8272
+▁cardinal -10.8272
+▁premier -10.8272
+▁spider -10.8272
+▁saturday -10.8272
+▁silesia -10.8273
+▁stylistic -10.8273
+▁contest -10.8273
+▁trading -10.8274
+▁essen -10.8275
+▁holstein -10.8275
+▁participation -10.8276
+▁jacket -10.8278
+▁automatic -10.8278
+▁rival -10.8281
+▁java -10.8281
+▁simon -10.8284
+▁clip -10.8289
+▁eich -10.831
+▁teach -10.8311
+▁chant -10.8316
+▁relie -10.8335
+▁hatch -10.8357
+hipped -10.8382
+cute -10.8412
+▁neuro -10.8522
+▁criticize -10.8548
+itude -10.855
+acent -10.8562
+▁integr -10.8583
+▁accent -10.859
+▁gentle -10.8594
+▁immun -10.8595
+▁contradict -10.8595
+▁principal -10.8595
+▁argentina -10.8595
+▁confident -10.8595
+▁convincing -10.8595
+▁erfurt -10.8595
+▁evolution -10.8595
+▁extinct -10.8595
+▁grandmother -10.8595
+▁impatient -10.8595
+▁libraries -10.8595
+▁majesty -10.8595
+▁perception -10.8595
+▁recreation -10.8595
+▁relegation -10.8595
+▁reputation -10.8595
+▁scandal -10.8595
+▁sebastian -10.8595
+▁striking -10.8595
+▁supplier -10.8595
+▁terrain -10.8595
+▁underestimate -10.8595
+▁wikipedia -10.8595
+▁colonial -10.8595
+▁compensation -10.8595
+▁cough -10.8595
+▁innovation -10.8595
+▁moderator -10.8595
+▁observatory -10.8595
+▁separation -10.8595
+▁disorder -10.8595
+▁revenge -10.8595
+▁sneak -10.8595
+▁perceive -10.8595
+▁injured -10.8595
+▁optim -10.8595
+▁superior -10.8595
+▁estonia -10.8595
+▁shaking -10.8595
+▁achievements -10.8595
+▁frisian -10.8595
+▁bernhard -10.8596
+▁betray -10.8597
+▁serbia -10.8597
+▁purple -10.8597
+▁axis -10.8598
+▁kramer -10.8599
+▁collaboration -10.8599
+▁bohm -10.8599
+▁textile -10.8602
+▁athens -10.8602
+▁nieder -10.8602
+▁implementation -10.8606
+▁massage -10.8607
+▁specifically -10.8608
+▁objective -10.861
+▁exile -10.8616
+▁schult -10.8619
+▁sailor -10.8643
+position -10.8646
+roscop -10.872
+hawk -10.8746
+▁combine -10.8878
+▁apparent -10.8887
+worth -10.8895
+florescence -10.8928
+▁captured -10.8928
+▁chimney -10.8928
+▁compromise -10.8928
+▁corporation -10.8928
+▁corridor -10.8928
+▁creativity -10.8928
+▁disagree -10.8928
+▁disappointed -10.8928
+▁fertile -10.8928
+▁fulfill -10.8928
+▁hollywood -10.8928
+▁hungarian -10.8928
+▁nephew -10.8928
+▁nickname -10.8928
+▁renaissance -10.8928
+▁reproduce -10.8928
+▁restrictions -10.8928
+▁slalom -10.8928
+▁surprising -10.8928
+▁uncertain -10.8928
+▁defect -10.8928
+▁enemies -10.8928
+▁peasant -10.8928
+▁spontaneous -10.8928
+▁squadron -10.8928
+▁tributar -10.8928
+▁volks -10.8928
+▁creek -10.8928
+▁slogan -10.8928
+▁module -10.8928
+▁robber -10.8928
+▁greece -10.8928
+▁smith -10.8928
+▁wreck -10.8928
+▁retail -10.8928
+▁inscription -10.8928
+▁drift -10.8928
+▁bulb -10.8928
+▁couch -10.8928
+▁duchy -10.8928
+▁jacob -10.8928
+▁plent -10.8928
+▁witch -10.8928
+▁alleged -10.8929
+▁oscar -10.8929
+▁competent -10.8929
+▁scottish -10.8929
+▁neighborhood -10.893
+▁tuning -10.8931
+▁shy -10.8931
+▁jones -10.8931
+▁sought -10.8932
+▁danish -10.8934
+▁interval -10.8935
+▁norway -10.8936
+▁dominated -10.8937
+▁ottoman -10.8937
+▁weiss -10.8959
+poo -10.8969
+▁custom -10.9064
+▁giga -10.9099
+trogen -10.9108
+abea -10.9145
+▁prussia -10.9214
+junction -10.9273
+▁alphabet -10.9273
+▁breakthrough -10.9273
+▁chronicle -10.9273
+▁confidence -10.9273
+▁corruption -10.9273
+▁fertilize -10.9273
+▁frequencies -10.9273
+▁identify -10.9273
+▁irritate -10.9273
+▁phenomenon -10.9273
+▁provision -10.9273
+▁renovated -10.9273
+▁reservation -10.9273
+▁screwdriver -10.9273
+▁template -10.9273
+▁unbelievable -10.9273
+▁wiesbaden -10.9273
+▁anywhere -10.9273
+▁balcony -10.9273
+▁muslim -10.9273
+▁poverty -10.9273
+▁habits -10.9273
+▁laundry -10.9273
+▁midfield -10.9273
+▁portugal -10.9273
+▁replacement -10.9273
+▁nominated -10.9273
+▁sensation -10.9273
+▁fulda -10.9273
+▁schwerin -10.9273
+▁kassel -10.9273
+▁colony -10.9273
+▁outdoor -10.9273
+▁attitude -10.9273
+▁gerhard -10.9274
+▁junk -10.9275
+▁departure -10.9275
+ybody -10.9275
+▁moped -10.9276
+statt -10.9276
+▁jackson -10.9277
+▁assessment -10.9279
+▁vault -10.928
+▁partnership -10.9281
+▁cannon -10.9283
+▁upset -10.929
+▁font -10.9292
+▁productive -10.9296
+issa -10.9346
+schau -10.9387
+media -10.9431
+▁chase -10.9509
+rough -10.9512
+▁discuss -10.9545
+▁prohibit -10.9554
+▁distinguish -10.9554
+▁kilometer -10.9555
+hoff -10.9565
+▁techno -10.9571
+▁define -10.958
+▁nav -10.9584
+final -10.96
+▁reconstruct -10.962
+went -10.9624
+▁attenti -10.9626
+▁telescop -10.963
+▁wrestling -10.963
+▁archaeological -10.963
+▁champagne -10.963
+▁competitor -10.963
+▁conservatory -10.963
+▁cyclist -10.963
+▁darmstadt -10.963
+▁disadvantage -10.963
+▁eduard -10.963
+▁greifswald -10.963
+▁nuremberg -10.963
+▁occurrence -10.963
+▁pedestrian -10.963
+▁pomerania -10.963
+▁prosecutor -10.963
+▁qualify -10.963
+▁recognizable -10.963
+▁staircase -10.963
+▁supplied -10.963
+▁athletic -10.963
+▁bohemia -10.963
+▁circum -10.963
+▁conscious -10.963
+▁elegant -10.963
+▁helmut -10.963
+▁penguin -10.963
+▁wrapp -10.963
+▁database -10.963
+▁entertain -10.963
+▁scope -10.963
+▁cherries -10.963
+▁coalition -10.963
+▁balloon -10.963
+▁glove -10.963
+▁babies -10.963
+▁schnitz -10.963
+▁hanover -10.963
+▁victoria -10.963
+▁endanger -10.963
+▁spray -10.963
+▁senator -10.963
+▁croatia -10.963
+▁lodge -10.963
+▁motto -10.9631
+▁convey -10.9631
+▁refine -10.9631
+▁restored -10.9631
+▁offence -10.9631
+▁michel -10.9631
+▁fence -10.9631
+▁casual -10.9631
+▁advers -10.9632
+▁appeti -10.9632
+▁creme -10.9633
+▁diesel -10.9633
+▁moths -10.9635
+▁trainee -10.9635
+▁slate -10.9637
+▁charles -10.9638
+▁feud -10.9639
+▁beef -10.9645
+▁automat -10.9652
+▁resist -10.9657
+▁adam -10.9686
+gonal -10.9688
+▁peaceful -10.9773
+▁commen -10.9803
+patch -10.984
+▁analog -10.989
+borg -10.993
+▁wedg -10.994
+▁preserve -10.995
+fels -10.996
+founder -10.9968
+bridge -10.9968
+fortunately -10.9991
+▁communicat -10.9993
+▁carpenter -11
+▁dissertation -11
+▁ignorance -11
+▁mattress -11
+▁mexican -11
+▁packaging -11
+▁plaintiff -11
+▁psychiatr -11
+▁scotland -11
+▁sovereign -11
+▁statistics -11
+▁submarines -11
+▁upgrade -11
+▁utensils -11
+▁fantasy -11
+▁mozart -11
+▁sandwich -11
+▁unsuccessful -11
+▁typing -11
+▁decoration -11
+▁civic -11
+▁sausage -11
+▁oneself -11
+▁simson -11
+▁courage -11
+▁julius -11
+▁modified -11
+▁unlock -11
+▁twist -11.0001
+▁monster -11.0001
+▁atomic -11.0001
+▁horst -11.0001
+▁owl -11.0001
+▁tackle -11.0001
+▁motives -11.0002
+▁relay -11.0002
+▁edward -11.0002
+▁motion -11.0002
+▁pilgrimage -11.0003
+▁debate -11.0003
+▁manuel -11.0004
+▁preced -11.0005
+▁bernard -11.0008
+▁inventor -11.0009
+▁sabin -11.0019
+▁horizontal -11.002
+▁faci -11.0048
+holder -11.0062
+▁jorg -11.0207
+▁whip -11.0212
+nathan -11.0262
+▁attach -11.0317
+▁attribute -11.0334
+script -11.0352
+fried -11.0355
+burn -11.0363
+▁potato -11.0367
+▁depart -11.038
+▁oppos -11.0383
+▁immigrat -11.0385
+deficiency -11.0385
+responsibilities -11.0385
+▁albrecht -11.0385
+▁bielefeld -11.0385
+▁brilliant -11.0385
+▁bulgaria -11.0385
+▁cannibal -11.0385
+▁commemorat -11.0385
+▁competitive -11.0385
+▁december -11.0385
+▁endeavor -11.0385
+▁essays -11.0385
+▁explosion -11.0385
+▁fountain -11.0385
+▁infected -11.0385
+▁jesuit -11.0385
+▁lifestyle -11.0385
+▁limestone -11.0385
+▁oxford -11.0385
+▁pediatric -11.0385
+▁peripher -11.0385
+▁thailand -11.0385
+▁bacteria -11.0385
+▁chicago -11.0385
+▁headphones -11.0385
+▁hungary -11.0385
+▁rostock -11.0385
+▁arnold -11.0385
+▁commerce -11.0385
+▁memories -11.0385
+▁thirst -11.0385
+▁colonel -11.0385
+▁disaster -11.0385
+▁garlic -11.0385
+▁swear -11.0385
+▁clever -11.0385
+▁expertise -11.0385
+▁outstanding -11.0385
+▁survey -11.0385
+▁sensor -11.0385
+▁whale -11.0385
+▁accusations -11.0385
+▁spiral -11.0385
+trustworthy -11.0385
+▁patent -11.0385
+▁vienne -11.0386
+▁satur -11.0386
+▁relocat -11.0386
+▁shelter -11.0386
+▁explod -11.0387
+▁considerably -11.0387
+▁distinctive -11.0387
+▁jesus -11.0388
+▁gauge -11.0389
+ogram -11.0389
+▁charlie -11.0389
+▁proceedings -11.0391
+▁eisen -11.0392
+▁abuse -11.0403
+▁wonderful -11.0407
+ubb -11.0429
+▁chris -11.047
+▁remov -11.0626
+fumes -11.0709
+▁arrest -11.0709
+▁arrange -11.073
+▁assume -11.0736
+▁headquarter -11.0756
+tractable -11.0785
+▁archbishop -11.0785
+▁attorney -11.0785
+▁batteries -11.0785
+▁breslau -11.0785
+▁comparative -11.0785
+▁confront -11.0785
+▁cultivation -11.0785
+▁deceased -11.0785
+▁diagnosis -11.0785
+▁emphasize -11.0785
+▁evident -11.0785
+▁impossible -11.0785
+▁palatinate -11.0785
+▁proposal -11.0785
+▁prototype -11.0785
+▁punishment -11.0785
+▁rupture -11.0785
+▁scandinavia -11.0785
+▁sculptor -11.0785
+▁technician -11.0785
+▁teenager -11.0785
+▁timbered -11.0785
+▁transformation -11.0785
+▁transmitted -11.0785
+▁violations -11.0785
+▁virtue -11.0785
+▁carinthia -11.0785
+▁synchroni -11.0785
+▁treasure -11.0785
+▁vineyard -11.0785
+▁privatiz -11.0785
+▁spacious -11.0785
+▁binding -11.0785
+▁ritual -11.0785
+▁manfred -11.0785
+▁urgent -11.0785
+▁judicial -11.0785
+▁serving -11.0785
+▁pigeon -11.0785
+▁compact -11.0785
+▁hurry -11.0785
+▁bleach -11.0785
+▁escalat -11.0785
+currency -11.0785
+▁syria -11.0785
+▁smile -11.0785
+▁henrik -11.0786
+▁theodor -11.0786
+▁tempt -11.0786
+▁feedback -11.0787
+▁gentlemen -11.0787
+▁activat -11.0787
+▁semiconductor -11.0787
+▁fortification -11.0788
+▁framework -11.0788
+▁winning -11.0788
+▁hydrogen -11.0788
+▁erwin -11.0788
+▁tibet -11.0788
+▁plus -11.0789
+▁exert -11.08
+▁differentiate -11.0803
+▁concei -11.0828
+▁celt -11.0849
+▁pont -11.0895
+undula -11.1067
+lateral -11.1088
+▁raise -11.1153
+woman -11.1169
+▁absence -11.1202
+▁accumulat -11.1202
+▁administered -11.1202
+▁anonymity -11.1202
+▁arrondissement -11.1202
+▁braunschweig -11.1202
+▁constituency -11.1202
+▁diabetes -11.1202
+▁discontinued -11.1202
+▁discretion -11.1202
+▁diversity -11.1202
+▁enthusiastic -11.1202
+▁explicitly -11.1202
+▁financing -11.1202
+▁hemisphere -11.1202
+▁innocent -11.1202
+▁intellectual -11.1202
+▁irregular -11.1202
+▁leopold -11.1202
+▁mandatory -11.1202
+▁maneuver -11.1202
+▁mercedes -11.1202
+▁permission -11.1202
+▁postponed -11.1202
+▁receipt -11.1202
+▁refrain -11.1202
+▁regensburg -11.1202
+▁repertoire -11.1202
+▁sweep -11.1202
+▁transaction -11.1202
+▁twitch -11.1202
+▁vietnam -11.1202
+▁inevitabl -11.1202
+▁nikolaus -11.1202
+▁renamed -11.1202
+▁semester -11.1202
+▁solidarity -11.1202
+▁verdict -11.1202
+▁chords -11.1202
+▁rainbow -11.1202
+▁nicola -11.1202
+▁sociology -11.1202
+▁ethnic -11.1202
+▁cartoon -11.1202
+▁replica -11.1202
+▁tropical -11.1202
+▁rigid -11.1202
+▁magnus -11.1202
+▁pliers -11.1202
+▁reddish -11.1202
+▁tactic -11.1202
+▁margin -11.1202
+▁robin -11.1202
+▁incorporated -11.1202
+▁reinhard -11.1202
+▁stamens -11.1202
+▁invention -11.1203
+▁stopping -11.1204
+▁interloc -11.1204
+▁silk -11.1205
+▁clemen -11.1207
+▁niko -11.1208
+▁spice -11.1216
+▁deceiv -11.1218
+motive -11.1239
+▁alpin -11.1264
+▁nitr -11.1276
+western -11.1307
+spiel -11.1335
+moller -11.1351
+▁install -11.1553
+▁convention -11.1595
+▁fluctuat -11.1636
+▁acquaint -11.1636
+▁appreciate -11.1636
+▁arrival -11.1636
+▁artillery -11.1636
+▁carbohydrate -11.1636
+▁categoriz -11.1636
+▁celebration -11.1636
+▁coincidence -11.1636
+▁comrade -11.1636
+▁crocodile -11.1636
+▁defensive -11.1636
+▁depriv -11.1636
+▁diagram -11.1636
+▁inflammation -11.1636
+▁interference -11.1636
+▁intersection -11.1636
+▁migration -11.1636
+▁millennium -11.1636
+▁philosophic -11.1636
+▁pregnant -11.1636
+▁prehistoric -11.1636
+▁prescription -11.1636
+▁preservation -11.1636
+▁prominent -11.1636
+▁propaganda -11.1636
+▁rehearsal -11.1636
+▁smoking -11.1636
+▁stabilize -11.1636
+▁substantial -11.1636
+▁suspicion -11.1636
+▁ulrich -11.1636
+▁universal -11.1636
+▁unnoticed -11.1636
+▁venice -11.1636
+▁yield -11.1636
+▁laboratory -11.1636
+▁surname -11.1636
+▁utilized -11.1636
+▁valentin -11.1636
+▁victories -11.1636
+▁federation -11.1636
+▁samuel -11.1636
+▁thigh -11.1636
+▁charlotte -11.1636
+▁clarified -11.1636
+▁texas -11.1636
+▁thief -11.1636
+▁livestock -11.1636
+▁sprint -11.1636
+▁guinea -11.1636
+▁latvia -11.1636
+▁cliff -11.1636
+▁premise -11.1636
+▁toddler -11.1636
+▁nightmare -11.1637
+▁leash -11.1637
+▁session -11.1637
+▁cheerful -11.1637
+▁overcome -11.1637
+▁sirens -11.1637
+▁cement -11.1638
+▁withdrew -11.1638
+▁tasty -11.1638
+▁forgive -11.1639
+▁scrub -11.1642
+▁squir -11.1649
+▁trauma -11.1852
+▁elect -11.1973
+▁introduce -11.2041
+▁easie -11.207
+▁permit -11.2088
+▁aqua -11.2091
+▁electromagnet -11.2091
+▁reflex -11.2091
+▁accommodation -11.2091
+▁acquisition -11.2091
+▁amsterdam -11.2091
+▁applicable -11.2091
+▁applicant -11.2091
+▁boycott -11.2091
+▁cloakroom -11.2091
+▁demonstration -11.2091
+▁electrified -11.2091
+▁emphasis -11.2091
+▁enthusiasm -11.2091
+▁expense -11.2091
+▁faculty -11.2091
+▁figurative -11.2091
+▁flirt -11.2091
+▁geometric -11.2091
+▁imbalance -11.2091
+▁immigrant -11.2091
+▁inconsistencies -11.2091
+▁instructor -11.2091
+▁katharina -11.2091
+▁koblenz -11.2091
+▁magdeburg -11.2091
+▁mechanism -11.2091
+▁mercury -11.2091
+▁microwave -11.2091
+▁monkey -11.2091
+▁pharmacist -11.2091
+▁relegated -11.2091
+▁rendsburg -11.2091
+▁ribbon -11.2091
+▁simulate -11.2091
+▁symmetric -11.2091
+▁syrup -11.2091
+▁vicinity -11.2091
+▁affiliat -11.2091
+▁autumn -11.2091
+▁restricted -11.2091
+▁scanner -11.2091
+▁sigmund -11.2091
+▁exploit -11.2091
+▁surgeon -11.2091
+▁colonies -11.2091
+▁depression -11.2091
+▁bloom -11.2091
+▁indonesia -11.2091
+▁bargain -11.2091
+▁minuet -11.2091
+▁trophy -11.2091
+▁competence -11.2091
+▁satire -11.2091
+▁tamara -11.2091
+▁thrill -11.2091
+▁generous -11.2091
+▁spoil -11.2091
+▁bureau -11.2091
+▁caesar -11.2091
+▁madrid -11.2091
+▁nicole -11.2091
+▁probation -11.2091
+▁styria -11.2091
+▁rascal -11.2092
+▁argue -11.2092
+▁blond -11.2092
+▁palestin -11.2092
+▁quin -11.2093
+▁plasm -11.2093
+▁vogt -11.2094
+▁beetle -11.2097
+ophyt -11.21
+alysis -11.2111
+▁remu -11.2116
+▁snor -11.2117
+ocated -11.2131
+eastern -11.2153
+▁invasi -11.2201
+friedhof -11.2305
+▁presid -11.232
+fluent -11.2397
+piece -11.2535
+heart -11.2537
+▁horizon -11.2542
+▁democrat -11.2555
+▁pilgrim -11.2564
+▁apprentice -11.2565
+▁napoleon -11.2567
+▁sympath -11.2567
+▁abrupt -11.2567
+▁accelerated -11.2567
+▁antarctic -11.2567
+▁applause -11.2567
+▁astronomer -11.2567
+▁barracks -11.2567
+▁byzantine -11.2567
+▁celebrities -11.2567
+▁combustion -11.2567
+▁conversion -11.2567
+▁crucifix -11.2567
+▁cuddle -11.2567
+▁dinosaur -11.2567
+▁eberhard -11.2567
+▁entrepreneur -11.2567
+▁envelop -11.2567
+▁expedition -11.2567
+▁indirect -11.2567
+▁industries -11.2567
+▁intermediate -11.2567
+▁kennedy -11.2567
+▁migrant -11.2567
+▁neolithic -11.2567
+▁outskirts -11.2567
+▁patience -11.2567
+▁provincial -11.2567
+▁registration -11.2567
+▁speculation -11.2567
+▁synagogue -11.2567
+▁textbook -11.2567
+▁unhealthy -11.2567
+▁unpopular -11.2567
+▁vegetation -11.2567
+▁violent -11.2567
+▁viticulture -11.2567
+▁wrinkl -11.2567
+▁botanical -11.2567
+▁hymn -11.2567
+▁marathon -11.2567
+▁physician -11.2567
+▁hypo -11.2567
+▁inflation -11.2567
+▁schroder -11.2567
+▁alsace -11.2567
+▁knives -11.2567
+▁brush -11.2567
+▁lehmann -11.2567
+▁sniff -11.2567
+estimation -11.2567
+▁backpack -11.2567
+▁chestnut -11.2567
+▁bleeding -11.2567
+▁decent -11.2567
+▁yolk -11.2567
+▁bochum -11.2567
+▁boredom -11.2567
+▁froze -11.2567
+▁uncover -11.2567
+▁warehouse -11.2567
+▁protein -11.2567
+▁dentist -11.2567
+▁spores -11.2567
+▁parrot -11.2567
+▁fischer -11.2568
+▁seventeen -11.2568
+▁depth -11.2568
+▁petersburg -11.2568
+▁output -11.2568
+▁detach -11.2568
+▁recommendations -11.2569
+▁inverse -11.2569
+▁falk -11.2571
+cursor -11.2591
+▁drip -11.2632
+ipping -11.2633
+rpor -11.2639
+foot -11.3035
+scribed -11.306
+▁sustain -11.3061
+í -11.3067
+▁exempt -11.3067
+▁stubborn -11.3067
+▁ambassador -11.3067
+▁archdiocese -11.3067
+▁aristocrat -11.3067
+▁astronomy -11.3067
+▁barricade -11.3067
+▁breisgau -11.3067
+▁captivity -11.3067
+▁claustrophobi -11.3067
+▁compulsory -11.3067
+▁deliberat -11.3067
+▁dismissed -11.3067
+▁elephant -11.3067
+▁heilbronn -11.3067
+▁hydraulic -11.3067
+▁innocence -11.3067
+▁intercept -11.3067
+▁istanbul -11.3067
+▁mosquitoes -11.3067
+▁nationwide -11.3067
+▁phenomena -11.3067
+▁photosynthesis -11.3067
+▁prerequisite -11.3067
+▁protagonist -11.3067
+▁reinforced -11.3067
+▁reluctant -11.3067
+▁rhetoric -11.3067
+▁slovakia -11.3067
+▁strategy -11.3067
+▁supermarket -11.3067
+▁supervision -11.3067
+▁buffet -11.3067
+▁cabaret -11.3067
+▁triumph -11.3067
+▁asylum -11.3067
+▁conquest -11.3067
+▁january -11.3067
+▁offensive -11.3067
+▁rugby -11.3067
+▁nectar -11.3067
+▁sepals -11.3067
+▁counterpart -11.3067
+▁farewell -11.3067
+▁incredibl -11.3067
+▁pretense -11.3067
+▁tariff -11.3067
+▁romanesque -11.3067
+▁barbara -11.3067
+▁falcon -11.3067
+▁soluble -11.3067
+▁balkan -11.3067
+▁glott -11.3067
+▁restoration -11.3067
+▁granite -11.3067
+▁flavor -11.3067
+believing -11.3067
+▁reunion -11.3067
+▁allotment -11.3067
+▁sicily -11.3067
+▁intact -11.3067
+▁vegan -11.3067
+▁abdom -11.3067
+▁discriminat -11.3067
+▁wolfsburg -11.3068
+ccup -11.3068
+▁scheme -11.3068
+▁agnes -11.3068
+▁cousin -11.3068
+▁charming -11.3068
+▁newcomer -11.3068
+▁costume -11.3068
+▁headlights -11.3068
+▁unfit -11.3068
+▁terminal -11.3069
+▁deviate -11.3069
+▁emulate -11.3069
+deutsch -11.3072
+▁bertol -11.3073
+tropic -11.3084
+mposed -11.3093
+goblin -11.3095
+imaginable -11.3294
+▁devil -11.3423
+believers -11.3452
+meyer -11.3466
+▁slip -11.3471
+▁acquir -11.3518
+▁adopt -11.3518
+▁convert -11.3519
+iddle -11.3543
+▁comfort -11.3587
+▁happi -11.3592
+▁absurd -11.3593
+▁accommodate -11.3593
+▁adequate -11.3593
+▁aesthetic -11.3593
+▁afghanistan -11.3593
+▁amphibi -11.3593
+▁antiquity -11.3593
+▁avalanche -11.3593
+▁barbecue -11.3593
+▁cavalry -11.3593
+▁censorship -11.3593
+▁clutch -11.3593
+▁cockpit -11.3593
+▁deactivate -11.3593
+▁devastat -11.3593
+▁dungeon -11.3593
+▁eclipse -11.3593
+▁enlightenment -11.3593
+▁explanatory -11.3593
+▁fragrance -11.3593
+▁horrible -11.3593
+▁imprisoned -11.3593
+▁imprisonment -11.3593
+▁indigenous -11.3593
+▁instinct -11.3593
+▁jerusalem -11.3593
+▁matthias -11.3593
+▁ornamental -11.3593
+▁python -11.3593
+▁rectangular -11.3593
+▁recurr -11.3593
+▁schizophreni -11.3593
+▁scooter -11.3593
+▁slaughter -11.3593
+▁sporadically -11.3593
+▁strategic -11.3593
+▁unemployed -11.3593
+▁vorpommern -11.3593
+▁wisdom -11.3593
+▁brackets -11.3593
+▁grenade -11.3593
+▁household -11.3593
+▁reunification -11.3593
+▁statistical -11.3593
+▁suffocat -11.3593
+▁almonds -11.3593
+▁beatrix -11.3593
+▁occupy -11.3593
+▁sediment -11.3593
+▁vitamin -11.3593
+▁wreath -11.3593
+▁sketch -11.3593
+▁xaver -11.3593
+▁provoke -11.3593
+▁charcoal -11.3593
+▁okres -11.3593
+▁pathogen -11.3593
+▁quarry -11.3593
+▁plaque -11.3593
+▁cloister -11.3594
+▁multipli -11.3594
+▁chassis -11.3594
+▁premium -11.3594
+▁relieved -11.3594
+▁vague -11.3594
+▁pylon -11.3594
+▁trivial -11.3594
+▁emblem -11.3594
+▁fisherman -11.3594
+▁sahara -11.3594
+▁jerry -11.3594
+▁platt -11.3594
+▁ventilat -11.3594
+▁rifle -11.3594
+▁choler -11.3594
+▁gregor -11.3594
+▁abundan -11.3594
+▁nurse -11.3594
+▁duties -11.3594
+▁warrior -11.3594
+▁supervise -11.3594
+▁frederik -11.3594
+▁undertake -11.3594
+▁merger -11.3595
+▁slam -11.3595
+▁martyr -11.3595
+▁uneven -11.3595
+▁guilt -11.3597
+▁wobbl -11.3599
+▁lichten -11.3602
+ointment -11.3608
+▁glue -11.3611
+▁ammoni -11.3615
+▁regin -11.3673
+▁westphal -11.3911
+woke -11.4025
+profit -11.4116
+▁distinct -11.4135
+corporate -11.4148
+ó -11.4149
+▁administrator -11.4149
+▁astronaut -11.4149
+▁bamboo -11.4149
+▁commuters -11.4149
+▁cuisine -11.4149
+▁cycling -11.4149
+▁decreasing -11.4149
+▁destiny -11.4149
+▁detroit -11.4149
+▁elevation -11.4149
+▁elongated -11.4149
+▁fascinat -11.4149
+▁firefighters -11.4149
+▁gigabyte -11.4149
+▁godfather -11.4149
+▁hereditary -11.4149
+▁hurricane -11.4149
+▁illusion -11.4149
+▁kaiserstuhl -11.4149
+▁ketchup -11.4149
+▁maximilian -11.4149
+▁minnesota -11.4149
+▁monasteries -11.4149
+▁nurnberg -11.4149
+▁panorama -11.4149
+▁peanuts -11.4149
+▁platinum -11.4149
+▁prefecture -11.4149
+▁prejudices -11.4149
+▁prosecution -11.4149
+▁prosperity -11.4149
+▁reproduction -11.4149
+▁saarbrucken -11.4149
+▁spectrum -11.4149
+▁strategies -11.4149
+▁subspecies -11.4149
+▁undamaged -11.4149
+ncreas -11.4149
+▁creep -11.4149
+▁florence -11.4149
+▁pajama -11.4149
+▁refract -11.4149
+▁promising -11.4149
+▁boulder -11.4149
+▁illuminat -11.4149
+▁landshut -11.4149
+▁asphalt -11.4149
+▁expelled -11.4149
+▁gottfried -11.4149
+▁healthcare -11.4149
+▁transformer -11.4149
+▁baptize -11.4149
+▁bettina -11.4149
+▁ancestr -11.4149
+▁burgenland -11.4149
+▁fauna -11.4149
+▁ironically -11.4149
+▁arthur -11.4149
+▁ghana -11.4149
+▁lounge -11.4149
+▁schloss -11.4149
+▁overseas -11.4149
+▁dairy -11.4149
+▁ignite -11.4149
+▁kauf -11.4149
+▁mutual -11.4149
+fascist -11.4149
+▁predict -11.4149
+▁butterfl -11.4149
+▁altona -11.4149
+▁fortune -11.4149
+▁cease -11.415
+▁slum -11.415
+▁flank -11.415
+▁pollute -11.415
+▁scam -11.4152
+▁britta -11.4156
+▁lexic -11.4157
+▁blend -11.4182
+▁audibl -11.4195
+▁accus -11.4457
+ctically -11.4647
+choke -11.465
+▁decorate -11.4687
+▁regime -11.4737
+▁magnify -11.4737
+▁abstain -11.4737
+▁bizarre -11.4737
+▁homosexual -11.4737
+isexual -11.4737
+philharmoni -11.4737
+▁accuracy -11.4737
+▁achilles -11.4737
+▁agencies -11.4737
+▁anatomical -11.4737
+▁barcelona -11.4737
+▁buddhist -11.4737
+▁cliché -11.4737
+▁confederation -11.4737
+▁cooperative -11.4737
+▁cosmonaut -11.4737
+▁culprit -11.4737
+▁denounce -11.4737
+▁dictator -11.4737
+▁dictionary -11.4737
+▁earthquake -11.4737
+▁educator -11.4737
+▁hildesheim -11.4737
+▁hybrid -11.4737
+▁hydrochlori -11.4737
+▁injustice -11.4737
+▁insufficient -11.4737
+▁keyword -11.4737
+▁manhattan -11.4737
+▁mysterious -11.4737
+▁narrative -11.4737
+▁navigable -11.4737
+▁ordinary -11.4737
+▁philosopher -11.4737
+▁protocol -11.4737
+▁punctuat -11.4737
+▁recyclable -11.4737
+▁referendum -11.4737
+▁remscheid -11.4737
+▁repetition -11.4737
+▁scenario -11.4737
+▁schreiber -11.4737
+▁subscribe -11.4737
+▁sulfur -11.4737
+▁twitter -11.4737
+▁elsewhere -11.4737
+▁embryo -11.4737
+▁nikolai -11.4737
+▁utilities -11.4737
+▁cornelia -11.4737
+▁quantify -11.4737
+▁reliably -11.4737
+▁ridicule -11.4737
+▁vessel -11.4737
+▁absent -11.4737
+▁imperative -11.4737
+▁lithium -11.4737
+▁submerge -11.4737
+▁folklore -11.4737
+▁leisure -11.4737
+▁messenger -11.4737
+▁opaque -11.4737
+▁adjoin -11.4737
+▁gasoline -11.4737
+▁volker -11.4737
+▁freak -11.4737
+▁logistic -11.4737
+▁taylor -11.4737
+▁volley -11.4737
+▁globe -11.4737
+▁provocati -11.4737
+▁sultan -11.4737
+▁residu -11.4737
+▁delegate -11.4737
+▁gloat -11.4737
+▁anhalt -11.4737
+▁remix -11.4737
+▁ecological -11.4737
+▁upload -11.4737
+selotte -11.4737
+▁cliche -11.4737
+▁underpass -11.4737
+▁fierce -11.4737
+▁reimann -11.4737
+▁aerob -11.4737
+▁verona -11.4738
+blazon -11.4746
+▁swap -11.4747
+▁desire -11.5312
+▁whil -11.5325
+aught -11.5354
+▁scott -11.5359
+▁eponym -11.5362
+▁pyramid -11.5362
+comprehensible -11.5362
+habilitation -11.5362
+á -11.5362
+▁acceleration -11.5362
+▁accomplished -11.5362
+▁admiral -11.5362
+▁admission -11.5362
+▁ambulance -11.5362
+▁amplifie -11.5362
+▁antelope -11.5362
+▁anxiety -11.5362
+▁aschaffenburg -11.5362
+▁backdrop -11.5362
+▁basilica -11.5362
+▁brussels -11.5362
+▁cinnamon -11.5362
+▁cognitive -11.5362
+▁conspicuous -11.5362
+▁crescent -11.5362
+▁crusade -11.5362
+▁czechoslovakia -11.5362
+▁delicacy -11.5362
+▁deployment -11.5362
+▁dismissal -11.5362
+▁disregarded -11.5362
+▁doctrine -11.5362
+▁episcop -11.5362
+▁exaggerate -11.5362
+▁excursion -11.5362
+▁hooligans -11.5362
+▁immense -11.5362
+▁infrequent -11.5362
+▁larvae -11.5362
+▁memorize -11.5362
+▁mongolia -11.5362
+▁mortgage -11.5362
+▁olivia -11.5362
+▁osnabruck -11.5362
+▁overview -11.5362
+▁posthumous -11.5362
+▁prognos -11.5362
+▁proximity -11.5362
+▁quartet -11.5362
+▁recursive -11.5362
+▁reminiscen -11.5362
+▁rhyme -11.5362
+▁singapore -11.5362
+▁sophisticated -11.5362
+▁strasbourg -11.5362
+▁stumbled -11.5362
+▁thunderstorm -11.5362
+▁tremendous -11.5362
+▁undecided -11.5362
+▁venezuela -11.5362
+▁yugoslavia -11.5362
+▁belgian -11.5362
+▁constitute -11.5362
+▁counsel -11.5362
+▁deteriorat -11.5362
+▁oblast -11.5362
+▁patricia -11.5362
+▁tendency -11.5362
+▁wolves -11.5362
+▁bikini -11.5362
+▁carolina -11.5362
+▁probability -11.5362
+▁soundtrack -11.5362
+▁intestine -11.5362
+▁moisture -11.5362
+▁offenders -11.5362
+▁sniper -11.5362
+▁utility -11.5362
+▁introver -11.5362
+â -12.0255
+ʿ -12.1366
+ə -13.0211
+ú -13.8545
+̇ -14.8543
+ñ -14.8544
+q -14.8545
+ð -14.8545
+ā -14.8545
+ū -14.8545
diff --git a/configs/es-en/config_gcmvn.yaml b/configs/es-en/config_gcmvn.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..6083d2ac6c4bf3a4276107fee747d3f47acc51b2
--- /dev/null
+++ b/configs/es-en/config_gcmvn.yaml
@@ -0,0 +1,21 @@
+global_cmvn:
+ stats_npz_path: /data/zhangshaolei/StreamSpeech/configs/es-en/gcmvn.npz
+input_channels: 1
+input_feat_per_channel: 80
+specaugment:
+ freq_mask_F: 27
+ freq_mask_N: 1
+ time_mask_N: 1
+ time_mask_T: 100
+ time_mask_p: 1.0
+ time_wrap_W: 0
+transforms:
+ '*':
+ - global_cmvn
+ _train:
+ - global_cmvn
+ - specaugment
+vocoder:
+ checkpoint: /data/zhangshaolei/pretrain_models/unit-based_HiFi-GAN_vocoder/mHuBERT.layer11.km1000.en/g_00500000
+ config: /data/zhangshaolei/pretrain_models/unit-based_HiFi-GAN_vocoder/mHuBERT.layer11.km1000.en/config.json
+ type: code_hifigan
diff --git a/configs/es-en/config_mtl_asr_st_ctcst.yaml b/configs/es-en/config_mtl_asr_st_ctcst.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..5dba702bdcf6d47d9da5ba326a18775878983b8c
--- /dev/null
+++ b/configs/es-en/config_mtl_asr_st_ctcst.yaml
@@ -0,0 +1,36 @@
+target_unigram:
+ decoder_type: transformer
+ dict: /data/zhangshaolei/StreamSpeech/configs/es-en/tgt_unigram6000/spm_unigram_es.txt
+ data: /data/zhangshaolei/StreamSpeech/configs/es-en/tgt_unigram6000
+ loss_weight: 8.0
+ rdrop_alpha: 0.0
+ decoder_args:
+ decoder_layers: 4
+ decoder_embed_dim: 512
+ decoder_ffn_embed_dim: 2048
+ decoder_attention_heads: 8
+ label_smoothing: 0.1
+source_unigram:
+ decoder_type: ctc
+ dict: /data/zhangshaolei/StreamSpeech/configs/es-en/src_unigram6000/spm_unigram_es.txt
+ data: /data/zhangshaolei/StreamSpeech/configs/es-en/src_unigram6000
+ loss_weight: 4.0
+ rdrop_alpha: 0.0
+ decoder_args:
+ decoder_layers: 0
+ decoder_embed_dim: 512
+ decoder_ffn_embed_dim: 2048
+ decoder_attention_heads: 8
+ label_smoothing: 0.1
+ctc_target_unigram:
+ decoder_type: ctc
+ dict: /data/zhangshaolei/StreamSpeech/configs/es-en/tgt_unigram6000/spm_unigram_es.txt
+ data: /data/zhangshaolei/StreamSpeech/configs/es-en/tgt_unigram6000
+ loss_weight: 4.0
+ rdrop_alpha: 0.0
+ decoder_args:
+ decoder_layers: 0
+ decoder_embed_dim: 512
+ decoder_ffn_embed_dim: 2048
+ decoder_attention_heads: 8
+ label_smoothing: 0.1
\ No newline at end of file
diff --git a/configs/es-en/config_unity.yaml b/configs/es-en/config_unity.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..b455d546d88668ba55895705c0ba9d1f685c792f
--- /dev/null
+++ b/configs/es-en/config_unity.yaml
@@ -0,0 +1,12 @@
+target_unigram:
+ decoder_type: transformer
+ dict: /data/zhangshaolei/StreamSpeech/configs/es-en/tgt_unigram6000/spm_unigram_es.txt
+ data: /data/zhangshaolei/StreamSpeech/configs/es-en/tgt_unigram6000
+ loss_weight: 8.0
+ rdrop_alpha: 0.0
+ decoder_args:
+ decoder_layers: 4
+ decoder_embed_dim: 512
+ decoder_ffn_embed_dim: 2048
+ decoder_attention_heads: 8
+ label_smoothing: 0.1
\ No newline at end of file
diff --git a/configs/es-en/gcmvn.npz b/configs/es-en/gcmvn.npz
new file mode 100644
index 0000000000000000000000000000000000000000..37e341a008edc2dadb1022282407e3217670a481
--- /dev/null
+++ b/configs/es-en/gcmvn.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3ffac0a9b3fae4f914a5a5c723ffa876d83ade3179d5e45b98974b7fb8add823
+size 1140
diff --git a/configs/es-en/src_unigram6000/spm_unigram_es.model b/configs/es-en/src_unigram6000/spm_unigram_es.model
new file mode 100644
index 0000000000000000000000000000000000000000..e4237f1fa5a6e82f5078f299f21592526ebb2f20
--- /dev/null
+++ b/configs/es-en/src_unigram6000/spm_unigram_es.model
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9a7a7b62a4eeb9e2d63c070864c031d10b64d5d42fd3f3c148fd9efe88a7343b
+size 346950
diff --git a/configs/es-en/src_unigram6000/spm_unigram_es.txt b/configs/es-en/src_unigram6000/spm_unigram_es.txt
new file mode 100644
index 0000000000000000000000000000000000000000..74e9bab6b302c8b176d3a0d7e8b7fb275da63c26
--- /dev/null
+++ b/configs/es-en/src_unigram6000/spm_unigram_es.txt
@@ -0,0 +1,5996 @@
+s 1
+▁de 1
+▁la 1
+▁en 1
+▁el 1
+▁y 1
+▁a 1
+▁se 1
+▁es 1
+n 1
+▁que 1
+▁los 1
+▁su 1
+r 1
+▁un 1
+▁del 1
+▁con 1
+▁ 1
+▁no 1
+▁por 1
+▁una 1
+a 1
+ó 1
+o 1
+e 1
+▁las 1
+▁al 1
+do 1
+▁fue 1
+es 1
+▁como 1
+▁para 1
+y 1
+l 1
+▁ha 1
+da 1
+▁lo 1
+▁más 1
+mente 1
+▁me 1
+▁son 1
+▁o 1
+▁esta 1
+▁sus 1
+c 1
+▁le 1
+▁este 1
+ta 1
+ción 1
+t 1
+▁está 1
+u 1
+▁dos 1
+▁mu 1
+ca 1
+la 1
+ba 1
+te 1
+▁tiene 1
+▁re 1
+ra 1
+to 1
+▁sin 1
+▁mi 1
+▁te 1
+ron 1
+▁era 1
+na 1
+▁puede 1
+d 1
+ndo 1
+ti 1
+▁si 1
+▁pero 1
+g 1
+i 1
+dos 1
+ado 1
+co 1
+▁hay 1
+le 1
+os 1
+▁ser 1
+ma 1
+ar 1
+▁entre 1
+▁e 1
+ro 1
+▁encuentra 1
+▁esto 1
+p 1
+▁durant 1
+f 1
+▁sobre 1
+lo 1
+ciones 1
+▁parte 1
+go 1
+de 1
+▁in 1
+me 1
+no 1
+v 1
+das 1
+ce 1
+▁todo 1
+re 1
+▁tres 1
+í 1
+á 1
+mos 1
+▁ciudad 1
+▁ca 1
+▁forma 1
+▁di 1
+ri 1
+▁desde 1
+▁fueron 1
+▁he 1
+ía 1
+ga 1
+se 1
+er 1
+li 1
+as 1
+m 1
+▁actualmente 1
+▁años 1
+an 1
+▁gran 1
+z 1
+▁ni 1
+▁cuando 1
+▁además 1
+▁ya 1
+nte 1
+▁primera 1
+▁nombre 1
+di 1
+▁año 1
+▁sido 1
+ada 1
+▁uno 1
+en 1
+rse 1
+▁bien 1
+mi 1
+é 1
+▁tiempo 1
+▁ma 1
+▁mismo 1
+tuvo 1
+mo 1
+▁yo 1
+▁ver 1
+▁san 1
+▁co 1
+al 1
+▁embargo 1
+cia 1
+▁vida 1
+▁final 1
+▁todos 1
+▁han 1
+▁mar 1
+ón 1
+▁cada 1
+▁da 1
+idad 1
+va 1
+ja 1
+za 1
+▁ese 1
+ci 1
+tu 1
+▁tu 1
+pa 1
+res 1
+ría 1
+▁primer 1
+b 1
+ne 1
+▁pi 1
+po 1
+donde 1
+▁general 1
+▁universidad 1
+▁va 1
+▁ella 1
+▁hace 1
+▁bo 1
+les 1
+▁otros 1
+▁ba 1
+▁nada 1
+▁ro 1
+ista 1
+bo 1
+▁estos 1
+▁vi 1
+▁mayor 1
+▁entonces 1
+qué 1
+hasta 1
+ve 1
+h 1
+▁historia 1
+ria 1
+▁están 1
+ban 1
+dor 1
+▁casa 1
+▁cuatro 1
+k 1
+▁luego 1
+▁color 1
+▁así 1
+je 1
+▁eso 1
+án 1
+▁mejor 1
+jo 1
+▁varios 1
+▁i 1
+gu 1
+▁había 1
+▁vez 1
+▁estaba 1
+▁nueva 1
+gi 1
+▁pre 1
+pe 1
+▁familia 1
+▁tras 1
+bi 1
+▁ta 1
+▁siempre 1
+que 1
+▁nacional 1
+▁mundo 1
+▁solo 1
+▁sa 1
+x 1
+▁pa 1
+is 1
+▁día 1
+▁vamos 1
+▁mo 1
+▁pe 1
+▁algunos 1
+nos 1
+▁cuenta 1
+▁tarde 1
+▁equipo 1
+fi 1
+▁centro 1
+▁vale 1
+▁tanto 1
+▁li 1
+▁lugar 1
+▁ahora 1
+▁padre 1
+▁tienen 1
+▁pro 1
+pi 1
+▁otro 1
+ver 1
+on 1
+▁poco 1
+ia 1
+sta 1
+qui 1
+▁car 1
+ina 1
+ur 1
+▁ex 1
+▁ra 1
+▁hijo 1
+ni 1
+▁guerra 1
+▁estado 1
+or 1
+▁cu 1
+rá 1
+▁hacer 1
+▁allí 1
+▁varias 1
+▁posteriormente 1
+▁muchos 1
+▁nunca 1
+ación 1
+▁bajo 1
+lla 1
+tas 1
+▁antes 1
+▁contra 1
+ng 1
+▁agua 1
+▁nació 1
+▁carrera 1
+zo 1
+▁mucho 1
+▁estas 1
+▁mucha 1
+ando 1
+ismo 1
+▁río 1
+io 1
+▁momento 1
+▁per 1
+ter 1
+▁serie 1
+vo 1
+nd 1
+ch 1
+▁obra 1
+▁p 1
+us 1
+▁tenía 1
+▁grupo 1
+ul 1
+▁c 1
+sólo 1
+mp 1
+▁otra 1
+▁ellos 1
+ero 1
+▁partido 1
+in 1
+ida 1
+▁club 1
+▁ve 1
+▁ti 1
+ido 1
+ita 1
+▁trabajo 1
+ll 1
+ck 1
+▁álbum 1
+▁escuela 1
+ica 1
+ú 1
+▁película 1
+w 1
+▁veces 1
+▁otras 1
+▁sur 1
+▁todas 1
+▁trata 1
+▁niña 1
+▁principal 1
+be 1
+▁presenta 1
+▁na 1
+▁ri 1
+man 1
+▁segunda 1
+▁ven 1
+▁nuevo 1
+▁hecho 1
+▁canción 1
+tos 1
+▁cinco 1
+ados 1
+ge 1
+▁españa 1
+era 1
+ntes 1
+▁banda 1
+▁noche 1
+▁algunas 1
+▁ho 1
+▁tan 1
+tó 1
+▁sistema 1
+▁sabe 1
+▁eran 1
+▁ambos 1
+▁debe 1
+▁música 1
+unque 1
+▁zona 1
+▁hoy 1
+▁sol 1
+▁iglesia 1
+▁estudió 1
+▁parece 1
+▁tra 1
+si 1
+▁país 1
+tra 1
+▁estudios 1
+cio 1
+ir 1
+▁toda 1
+nes 1
+▁lleva 1
+du 1
+il 1
+tivo 1
+▁man 1
+▁esa 1
+▁sede 1
+▁algo 1
+▁so 1
+▁siguiente 1
+gún 1
+▁importante 1
+▁barco 1
+eron 1
+▁favor 1
+▁norte 1
+▁porque 1
+▁fu 1
+▁grandes 1
+▁especial 1
+aba 1
+▁hu 1
+ble 1
+▁pu 1
+rio 1
+▁ce 1
+cha 1
+▁ja 1
+▁hizo 1
+▁b 1
+▁diferentes 1
+nt 1
+ente 1
+▁casi 1
+▁pueblo 1
+▁temporada 1
+▁mal 1
+ros 1
+▁madre 1
+▁fi 1
+▁tener 1
+▁ar 1
+▁provincia 1
+▁estados 1
+▁capital 1
+▁ojos 1
+▁ir 1
+ico 1
+ste 1
+▁hijos 1
+▁dentro 1
+▁pasa 1
+▁amor 1
+▁muerte 1
+▁aquí 1
+che 1
+cho 1
+▁tal 1
+▁quiero 1
+▁quien 1
+▁unidos 1
+▁deja 1
+▁orden 1
+▁largo 1
+▁junto 1
+▁tierra 1
+el 1
+▁obras 1
+▁edificio 1
+▁política 1
+▁par 1
+cos 1
+▁gusta 1
+▁último 1
+▁cabeza 1
+dores 1
+▁segundo 1
+▁hija 1
+▁igual 1
+las 1
+ones 1
+▁menos 1
+▁especie 1
+▁hombre 1
+▁principalmente 1
+son 1
+tro 1
+▁bu 1
+▁seis 1
+▁fe 1
+▁juan 1
+ya 1
+▁baja 1
+▁pan 1
+▁tengo 1
+▁siendo 1
+iendo 1
+▁misma 1
+▁habita 1
+▁gobierno 1
+▁the 1
+▁to 1
+▁suele 1
+▁comenzó 1
+▁medio 1
+▁internacional 1
+▁derecho 1
+▁posee 1
+▁cor 1
+men 1
+▁conocido 1
+▁isla 1
+▁queda 1
+▁pues 1
+▁jo 1
+▁existen 1
+▁sí 1
+▁municipio 1
+▁disco 1
+▁planta 1
+▁miembro 1
+▁mira 1
+▁central 1
+▁juego 1
+tar 1
+▁recibió 1
+ín 1
+▁or 1
+▁fuerte 1
+mb 1
+ito 1
+tor 1
+▁cha 1
+▁do 1
+▁ne 1
+sión 1
+remos 1
+car 1
+▁cal 1
+▁original 1
+▁fútbol 1
+▁sal 1
+▁presidente 1
+um 1
+▁llama 1
+▁alta 1
+▁título 1
+▁puedo 1
+▁sigue 1
+▁bueno 1
+fe 1
+▁región 1
+▁considera 1
+ven 1
+▁cuerpo 1
+ndi 1
+▁ga 1
+▁tipo 1
+▁be 1
+▁negro 1
+▁señor 1
+▁dice 1
+ton 1
+▁po 1
+▁estilo 1
+tado 1
+▁natural 1
+▁chole 1
+th 1
+encia 1
+▁fin 1
+▁decir 1
+▁estar 1
+▁representa 1
+endo 1
+miento 1
+▁blanco 1
+▁costa 1
+adas 1
+osa 1
+▁gu 1
+▁programa 1
+ig 1
+▁grande 1
+▁ru 1
+▁número 1
+entras 1
+▁gente 1
+mento 1
+▁media 1
+▁caso 1
+▁realiza 1
+ían 1
+▁verdad 1
+▁libro 1
+▁quiere 1
+▁arte 1
+▁ad 1
+▁poder 1
+▁usa 1
+▁argentina 1
+▁flores 1
+▁corre 1
+▁dar 1
+▁can 1
+vis 1
+▁maría 1
+▁buena 1
+ario 1
+▁fa 1
+▁proyecto 1
+▁puerta 1
+▁jug 1
+▁g 1
+aron 1
+▁cargo 1
+st 1
+▁les 1
+nia 1
+▁santa 1
+▁mayoría 1
+▁camino 1
+ales 1
+ct 1
+▁director 1
+▁go 1
+un 1
+et 1
+tiva 1
+ke 1
+▁conoce 1
+ieron 1
+▁origen 1
+rra 1
+ncia 1
+▁única 1
+▁único 1
+▁muestra 1
+▁mujer 1
+▁canciones 1
+▁llega 1
+▁bar 1
+j 1
+mer 1
+▁gal 1
+▁comp 1
+▁acaba 1
+▁hi 1
+▁toma 1
+▁lu 1
+▁honor 1
+iones 1
+it 1
+▁trabaj 1
+▁ju 1
+ue 1
+mu 1
+▁ka 1
+▁teatro 1
+▁aún 1
+▁bi 1
+▁bri 1
+▁real 1
+ing 1
+pu 1
+▁wa 1
+▁du 1
+▁madrid 1
+▁producción 1
+ante 1
+▁chile 1
+▁público 1
+emos 1
+bra 1
+▁manera 1
+▁ello 1
+▁ejemplo 1
+▁mundial 1
+▁cualquier 1
+▁campo 1
+▁ayuda 1
+▁hermano 1
+▁cambio 1
+▁tenemos 1
+ho 1
+eta 1
+ana 1
+▁hermana 1
+▁museo 1
+▁pat 1
+▁personas 1
+▁última 1
+▁estación 1
+▁destaca 1
+▁nadie 1
+▁cerca 1
+▁rec 1
+▁europa 1
+▁división 1
+▁dicho 1
+▁principales 1
+istas 1
+▁mano 1
+▁fo 1
+▁encuentran 1
+▁cri 1
+▁cer 1
+▁cine 1
+▁cosas 1
+ie 1
+▁ten 1
+▁eres 1
+▁base 1
+▁antiguo 1
+▁movimiento 1
+llo 1
+▁gan 1
+ces 1
+▁mon 1
+▁posible 1
+▁juega 1
+▁gracias 1
+▁actualidad 1
+▁días 1
+▁com 1
+gen 1
+▁realizó 1
+▁él 1
+▁sub 1
+ic 1
+▁azul 1
+tura 1
+▁antigua 1
+iz 1
+los 1
+nal 1
+▁militar 1
+▁distrito 1
+▁interior 1
+▁actual 1
+tal 1
+▁haber 1
+▁larga 1
+par 1
+▁k 1
+cer 1
+▁utiliza 1
+▁torre 1
+▁considerado 1
+tuvieron 1
+▁sitio 1
+ger 1
+▁don 1
+▁oficial 1
+bu 1
+▁línea 1
+▁ambas 1
+▁suelo 1
+▁uso 1
+nza 1
+per 1
+▁mor 1
+▁estamos 1
+▁pronto 1
+ut 1
+▁fuera 1
+▁for 1
+ió 1
+▁inter 1
+día 1
+▁primeros 1
+▁radio 1
+▁julia 1
+▁pesa 1
+▁col 1
+▁televisión 1
+▁causa 1
+lle 1
+ran 1
+▁población 1
+▁alto 1
+▁parque 1
+▁v 1
+▁habla 1
+ru 1
+mina 1
+▁d 1
+▁conocida 1
+▁t 1
+ber 1
+cal 1
+▁popular 1
+▁mí 1
+rlo 1
+▁figura 1
+ador 1
+▁china 1
+▁español 1
+ha 1
+▁debido 1
+ad 1
+▁mis 1
+▁total 1
+rme 1
+▁claro 1
+▁vive 1
+tic 1
+▁capítulo 1
+▁pertenece 1
+▁siete 1
+▁gi 1
+▁condado 1
+▁decidi 1
+▁u 1
+▁pasó 1
+▁convirti 1
+▁llegó 1
+▁principio 1
+ka 1
+▁super 1
+▁ocho 1
+▁república 1
+▁proceso 1
+▁barrio 1
+▁dios 1
+▁buen 1
+rre 1
+▁todavía 1
+▁estudio 1
+▁liga 1
+▁ab 1
+▁local 1
+▁estaban 1
+▁aquel 1
+▁cubierta 1
+▁autor 1
+tre 1
+▁falta 1
+ion 1
+▁profesor 1
+▁existe 1
+▁francia 1
+▁quedó 1
+▁actividad 1
+▁sería 1
+▁puesto 1
+▁gra 1
+▁joven 1
+▁celebra 1
+▁nave 1
+▁ante 1
+▁villa 1
+▁musical 1
+▁manos 1
+▁fra 1
+▁st 1
+port 1
+▁llamado 1
+tera 1
+tes 1
+▁modelo 1
+▁ocupa 1
+▁particip 1
+▁española 1
+▁meses 1
+▁hora 1
+ras 1
+▁je 1
+▁brasil 1
+▁propia 1
+▁civil 1
+▁edad 1
+tt 1
+▁corta 1
+▁cruz 1
+ner 1
+ología 1
+▁niño 1
+hi 1
+▁pie 1
+▁hábitat 1
+▁sencillo 1
+▁libre 1
+den 1
+▁obtuv 1
+tación 1
+▁marca 1
+▁oye 1
+▁rey 1
+▁punto 1
+land 1
+▁f 1
+▁territorio 1
+▁espacio 1
+ándose 1
+▁dio 1
+▁semana 1
+▁primero 1
+▁inclus 1
+▁desarrollo 1
+▁tema 1
+▁verde 1
+▁pequeña 1
+wa 1
+▁relación 1
+▁plan 1
+▁cantante 1
+▁amigos 1
+▁área 1
+▁puerto 1
+▁hojas 1
+▁bastante 1
+cciones 1
+▁mil 1
+▁profesional 1
+▁dura 1
+▁aparece 1
+▁pri 1
+▁valle 1
+rán 1
+▁castillo 1
+▁novela 1
+▁emp 1
+lin 1
+via 1
+tina 1
+▁pon 1
+▁estructura 1
+▁selección 1
+▁john 1
+▁nombrado 1
+▁tri 1
+▁interpreta 1
+▁jugador 1
+▁diferencia 1
+▁ninguna 1
+▁nosotros 1
+▁propio 1
+▁voz 1
+▁libros 1
+ano 1
+▁cr 1
+mar 1
+▁países 1
+ard 1
+ex 1
+▁será 1
+▁social 1
+▁soy 1
+▁luz 1
+ud 1
+▁numerosos 1
+▁político 1
+▁común 1
+▁pareja 1
+rro 1
+▁control 1
+▁versión 1
+▁sola 1
+lí 1
+▁tom 1
+▁van 1
+▁vas 1
+▁pasado 1
+▁fiesta 1
+▁dicha 1
+▁papel 1
+▁ci 1
+gar 1
+dad 1
+▁corte 1
+▁incluye 1
+▁voy 1
+▁horas 1
+▁unos 1
+im 1
+iente 1
+til 1
+▁australia 1
+▁india 1
+▁servicio 1
+iza 1
+ine 1
+▁calle 1
+▁partir 1
+▁alguien 1
+▁ocasiones 1
+▁dónde 1
+▁educación 1
+▁viene 1
+▁bro 1
+ura 1
+▁diversos 1
+▁batalla 1
+▁york 1
+imismo 1
+▁menor 1
+▁padres 1
+▁localidad 1
+▁reina 1
+▁trabaja 1
+▁logra 1
+▁mujeres 1
+▁áfrica 1
+▁contiene 1
+▁vuelve 1
+▁cómo 1
+▁torneo 1
+▁diez 1
+▁escrito 1
+▁pla 1
+▁normal 1
+toria 1
+▁ama 1
+▁compañía 1
+▁viaje 1
+▁piedra 1
+▁frente 1
+id 1
+▁fondo 1
+ino 1
+am 1
+▁dan 1
+▁nu 1
+▁tradicional 1
+▁diversas 1
+▁visto 1
+▁termina 1
+torio 1
+▁qui 1
+▁im 1
+▁revista 1
+▁bosques 1
+▁chi 1
+▁volver 1
+▁colombia 1
+tero 1
+▁espera 1
+▁clima 1
+▁especies 1
+ki 1
+▁nivel 1
+▁mata 1
+▁gira 1
+illa 1
+▁lucha 1
+▁oh 1
+▁cree 1
+▁resto 1
+▁departamento 1
+▁pequeños 1
+▁lista 1
+▁pequeño 1
+▁amigo 1
+ización 1
+▁matrimonio 1
+▁nueve 1
+▁acuerdo 1
+▁busca 1
+lina 1
+▁wi 1
+tino 1
+▁premio 1
+fer 1
+▁pas 1
+▁colegio 1
+▁come 1
+▁hermanos 1
+eño 1
+▁funciona 1
+▁palomares 1
+▁instituto 1
+▁dejó 1
+▁luis 1
+▁empresa 1
+der 1
+▁lengua 1
+▁influencia 1
+▁importantes 1
+▁sala 1
+gue 1
+▁podía 1
+▁desarrolla 1
+▁california 1
+▁práctica 1
+▁elegido 1
+▁ki 1
+dura 1
+ment 1
+▁and 1
+▁oro 1
+dio 1
+▁intenta 1
+ju 1
+▁niños 1
+ed 1
+rri 1
+▁resultado 1
+▁grupos 1
+▁ap 1
+eros 1
+▁juguete 1
+▁permite 1
+sh 1
+▁rosa 1
+▁creo 1
+▁información 1
+▁cual 1
+▁encontrar 1
+vin 1
+ster 1
+aban 1
+▁episodio 1
+▁pública 1
+▁escribi 1
+▁plaza 1
+bles 1
+▁partidos 1
+gan 1
+▁idea 1
+▁consta 1
+▁continua 1
+▁podemos 1
+stra 1
+▁rojo 1
+ji 1
+▁cultura 1
+illo 1
+▁construcción 1
+▁municipal 1
+▁nuestro 1
+▁murió 1
+▁cara 1
+▁zonas 1
+▁temas 1
+▁ref 1
+▁perro 1
+▁modo 1
+▁gana 1
+▁iba 1
+▁film 1
+▁santo 1
+▁observa 1
+▁arriba 1
+▁muerto 1
+▁volvió 1
+if 1
+▁miembros 1
+ega 1
+▁numerosas 1
+▁venga 1
+▁histórico 1
+ngún 1
+▁superior 1
+▁plantas 1
+▁juntos 1
+▁incluyen 1
+▁nuestra 1
+▁parís 1
+▁playa 1
+▁vino 1
+▁paso 1
+▁posición 1
+▁hacen 1
+ent 1
+ela 1
+▁compra 1
+dia 1
+▁persona 1
+▁categoría 1
+▁rep 1
+amente 1
+▁conserva 1
+▁visita 1
+▁toca 1
+▁tor 1
+▁arma 1
+▁tercera 1
+rica 1
+▁academia 1
+▁demasiado 1
+▁escucha 1
+▁carlos 1
+▁vivo 1
+▁restos 1
+▁mala 1
+bla 1
+ática 1
+▁hor 1
+▁actividades 1
+▁alemania 1
+▁memoria 1
+▁siento 1
+▁vivi 1
+ático 1
+▁sali 1
+eña 1
+aciones 1
+▁llev 1
+▁entra 1
+acu 1
+▁produce 1
+▁hospital 1
+▁viejo 1
+dero 1
+▁personaje 1
+▁fre 1
+▁escritor 1
+▁alrededor 1
+▁exterior 1
+▁francisco 1
+▁etapa 1
+▁z 1
+ff 1
+▁italia 1
+▁w 1
+▁aire 1
+▁tamaño 1
+▁tampoco 1
+▁victoria 1
+▁período 1
+▁consejo 1
+▁islas 1
+▁personal 1
+chi 1
+▁trabajar 1
+▁cabo 1
+▁ahí 1
+▁reino 1
+▁equipos 1
+▁gol 1
+▁crea 1
+▁fruto 1
+▁doce 1
+▁mari 1
+▁ge 1
+▁ant 1
+▁monta 1
+▁quieres 1
+▁económica 1
+▁puente 1
+▁hombres 1
+▁vía 1
+bilidad 1
+▁aspecto 1
+▁barcelona 1
+▁necesita 1
+▁economía 1
+▁indica 1
+rios 1
+▁cosa 1
+▁tú 1
+▁ke 1
+▁fecha 1
+▁vol 1
+cord 1
+▁salud 1
+▁películas 1
+▁actor 1
+▁anti 1
+▁boca 1
+▁viento 1
+▁escrita 1
+▁significa 1
+for 1
+cía 1
+▁juegos 1
+▁últimos 1
+▁inc 1
+▁llamada 1
+▁comercial 1
+▁alguna 1
+▁sociedad 1
+▁cuanto 1
+▁val 1
+▁oeste 1
+▁gar 1
+nder 1
+tada 1
+mediante 1
+wi 1
+▁levanta 1
+▁ben 1
+▁bur 1
+▁comienza 1
+▁conjunto 1
+▁amarillo 1
+▁obsta 1
+▁verano 1
+▁sea 1
+▁public 1
+▁buenos 1
+▁fundador 1
+▁luna 1
+▁llegar 1
+▁aquella 1
+lio 1
+▁of 1
+▁vo 1
+▁sta 1
+▁tro 1
+▁industria 1
+▁empieza 1
+▁objetivo 1
+▁cuarto 1
+▁situado 1
+▁canta 1
+▁indio 1
+▁cl 1
+▁formación 1
+udi 1
+▁acceso 1
+▁acepta 1
+▁pedro 1
+▁periodo 1
+▁ricardo 1
+▁inicia 1
+ly 1
+▁ban 1
+del 1
+ez 1
+▁za 1
+▁dr 1
+apareció 1
+▁agricultura 1
+▁dirección 1
+▁independiente 1
+▁japón 1
+▁presente 1
+▁fuerza 1
+bre 1
+▁duda 1
+ivo 1
+▁op 1
+ona 1
+cl 1
+ier 1
+▁occidental 1
+▁usted 1
+▁actriz 1
+▁situada 1
+▁estableci 1
+▁estadounidense 1
+▁entrada 1
+▁publicado 1
+▁eh 1
+▁unión 1
+▁campeonato 1
+▁invierno 1
+▁menudo 1
+▁templo 1
+▁pintura 1
+▁comenza 1
+tró 1
+▁form 1
+▁hubo 1
+▁anterior 1
+illas 1
+▁asia 1
+izado 1
+▁grado 1
+grafía 1
+▁difícil 1
+▁empezó 1
+▁golpe 1
+▁críticas 1
+▁lugares 1
+▁salir 1
+▁problemas 1
+▁ob 1
+▁alimenta 1
+▁recibe 1
+inició 1
+▁aeropuerto 1
+▁oscuro 1
+▁prueba 1
+▁londres 1
+▁casado 1
+▁compuesta 1
+▁seguro 1
+▁madera 1
+▁pequeñas 1
+▁marcha 1
+▁paul 1
+▁utilizado 1
+▁ley 1
+▁amplia 1
+▁h 1
+▁realizar 1
+bil 1
+▁texto 1
+▁palacio 1
+▁dedica 1
+▁valencia 1
+▁perú 1
+dra 1
+▁estrella 1
+▁blanca 1
+▁datos 1
+▁m 1
+▁cuidado 1
+▁monte 1
+puso 1
+▁hemo 1
+▁labor 1
+▁mañana 1
+▁santiago 1
+▁referencia 1
+▁corriente 1
+▁tenido 1
+▁sufri 1
+▁ganador 1
+▁integra 1
+▁sp 1
+▁también 1
+▁ofrece 1
+▁defini 1
+▁esposa 1
+▁luc 1
+▁podría 1
+▁cita 1
+▁trans 1
+▁roma 1
+▁tercer 1
+▁clase 1
+cula 1
+▁reyes 1
+▁red 1
+▁corto 1
+▁continuó 1
+▁enfermedad 1
+▁filosofía 1
+▁edición 1
+▁ingres 1
+▁valor 1
+▁auto 1
+ena 1
+▁corona 1
+▁opera 1
+▁ves 1
+▁vista 1
+▁ber 1
+▁campaña 1
+▁terminó 1
+▁logró 1
+▁maestro 1
+▁pobre 1
+▁diario 1
+▁doble 1
+gio 1
+▁pasar 1
+▁caja 1
+▁sale 1
+▁encontrado 1
+▁servicios 1
+▁localiza 1
+lógico 1
+▁variedad 1
+▁situación 1
+▁acerca 1
+▁acción 1
+ko 1
+▁diga 1
+▁diseño 1
+▁roja 1
+▁problema 1
+▁comunidad 1
+▁rica 1
+hu 1
+▁dedic 1
+▁ulises 1
+▁campeón 1
+▁camarote 1
+▁cementerio 1
+▁miedo 1
+▁sueño 1
+▁ciudades 1
+▁similar 1
+▁palabra 1
+▁jugar 1
+▁fla 1
+ión 1
+ería 1
+▁características 1
+▁negra 1
+ley 1
+sen 1
+▁decide 1
+▁completamente 1
+▁vio 1
+▁ger 1
+▁paz 1
+▁ganado 1
+▁aquello 1
+▁canal 1
+▁denomina 1
+▁artes 1
+▁accidente 1
+▁periódico 1
+▁determina 1
+▁alas 1
+▁ciencias 1
+▁sonido 1
+▁trabajos 1
+▁remo 1
+▁considerada 1
+▁che 1
+▁plata 1
+▁halla 1
+▁personajes 1
+▁propiedad 1
+▁medi 1
+▁echa 1
+▁simple 1
+▁tira 1
+▁silencio 1
+gráfico 1
+▁partes 1
+az 1
+▁casó 1
+▁enterrado 1
+ller 1
+val 1
+▁colores 1
+dieron 1
+icio 1
+▁fuente 1
+▁montaña 1
+▁cola 1
+▁elementos 1
+▁disputa 1
+▁jóvenes 1
+▁deportivo 1
+▁quién 1
+▁seguir 1
+▁animales 1
+▁basa 1
+▁pato 1
+▁ana 1
+▁comer 1
+▁artista 1
+▁doctor 1
+▁aguas 1
+▁tomar 1
+▁compañero 1
+▁cámara 1
+▁mezcla 1
+▁secundaria 1
+▁objeto 1
+▁cierto 1
+▁famoso 1
+▁capilla 1
+▁pecado 1
+▁investigación 1
+▁organización 1
+shi 1
+▁distintas 1
+ula 1
+▁dijo 1
+▁ubica 1
+dera 1
+▁ubicado 1
+▁literatura 1
+▁defensa 1
+▁festival 1
+▁adapta 1
+▁poeta 1
+▁motor 1
+▁alma 1
+▁acto 1
+▁gen 1
+▁cuba 1
+▁completa 1
+table 1
+▁relativa 1
+▁estudia 1
+▁mayores 1
+▁pens 1
+ctor 1
+▁mont 1
+▁rod 1
+▁función 1
+uda 1
+▁morir 1
+puesto 1
+ío 1
+▁web 1
+▁caballo 1
+▁fabrica 1
+▁ministro 1
+▁venezuela 1
+▁parecía 1
+▁fachada 1
+▁regresa 1
+▁video 1
+▁ataque 1
+▁abajo 1
+▁gui 1
+▁siglo 1
+▁hogar 1
+▁bru 1
+▁premios 1
+▁destino 1
+op 1
+▁funda 1
+▁hablar 1
+▁crítica 1
+▁semillas 1
+ü 1
+▁extremo 1
+amos 1
+▁dejar 1
+▁escena 1
+▁chica 1
+▁pocos 1
+▁cama 1
+▁casos 1
+▁estadio 1
+▁llevar 1
+▁carga 1
+▁mat 1
+▁fal 1
+▁agr 1
+ler 1
+▁recib 1
+▁uruguay 1
+▁antonio 1
+▁líder 1
+▁veinte 1
+▁relaciones 1
+▁permiti 1
+▁medicina 1
+▁poner 1
+▁debut 1
+▁caracteriza 1
+▁ciento 1
+▁notable 1
+▁delante 1
+▁obliga 1
+uri 1
+▁fan 1
+▁paga 1
+▁atrás 1
+▁reside 1
+▁abierto 1
+▁después 1
+▁experiencia 1
+▁mantiene 1
+▁seguridad 1
+▁poesía 1
+▁dulce 1
+▁versiones 1
+▁residencia 1
+▁natal 1
+▁fundación 1
+▁buenas 1
+at 1
+culo 1
+▁ellas 1
+▁exp 1
+▁aires 1
+tion 1
+▁tipos 1
+▁fuerzas 1
+idos 1
+▁católica 1
+▁clásico 1
+▁menciona 1
+▁necesito 1
+▁presencia 1
+▁policía 1
+▁distintos 1
+▁cel 1
+cip 1
+▁probablemente 1
+▁mercado 1
+▁mayo 1
+▁ventana 1
+▁valeria 1
+▁arena 1
+▁bella 1
+▁productor 1
+▁particular 1
+▁material 1
+▁torn 1
+lica 1
+ré 1
+▁tele 1
+▁arquitectura 1
+▁biblioteca 1
+▁inspira 1
+▁abogado 1
+▁bolivia 1
+▁cabecera 1
+▁terreno 1
+▁sentido 1
+▁frontera 1
+▁junta 1
+▁jefe 1
+▁ruta 1
+ben 1
+▁puntos 1
+▁vaya 1
+▁pos 1
+▁artistas 1
+▁curso 1
+▁bala 1
+▁pol 1
+ell 1
+▁traslad 1
+burgo 1
+▁carácter 1
+▁olvida 1
+▁perdido 1
+▁alemán 1
+▁guarda 1
+▁elimina 1
+▁deriva 1
+▁sirve 1
+▁oriental 1
+▁pal 1
+▁gas 1
+▁vuelo 1
+mba 1
+▁varia 1
+▁post 1
+▁hola 1
+▁permanece 1
+▁motivo 1
+▁copa 1
+▁ko 1
+arse 1
+▁realizado 1
+▁humanos 1
+cel 1
+▁capitán 1
+▁dirigido 1
+▁ninguno 1
+▁complejo 1
+▁contenido 1
+▁feliz 1
+▁haciendo 1
+▁industrial 1
+▁teoría 1
+rico 1
+▁tropicales 1
+▁coche 1
+▁resulta 1
+▁dije 1
+logía 1
+▁enseñ 1
+▁entrar 1
+▁abierta 1
+▁david 1
+▁corazón 1
+▁ambiente 1
+▁abarca 1
+▁importancia 1
+▁razón 1
+▁publica 1
+▁siguientes 1
+▁naturales 1
+▁jugaba 1
+▁gris 1
+▁pocas 1
+▁mostr 1
+▁j 1
+▁pone 1
+▁pica 1
+▁em 1
+▁nuevos 1
+▁aun 1
+▁regalo 1
+gre 1
+▁británico 1
+▁física 1
+▁inglaterra 1
+▁superficie 1
+▁treinta 1
+▁siguió 1
+▁virgen 1
+▁alcalde 1
+▁manuel 1
+▁vilma 1
+▁distribuye 1
+▁vivir 1
+▁imperio 1
+▁lago 1
+▁lejos 1
+▁triste 1
+▁coro 1
+▁consiste 1
+▁resultados 1
+mit 1
+▁escond 1
+▁libera 1
+▁fase 1
+ker 1
+iste 1
+▁pinta 1
+▁burbuja 1
+▁encuentro 1
+▁jardín 1
+▁miguel 1
+sistió 1
+▁combate 1
+▁perdió 1
+▁cadena 1
+▁llena 1
+▁muere 1
+▁cultural 1
+▁comunicación 1
+▁realidad 1
+▁internacionales 1
+▁continuación 1
+▁cura 1
+ular 1
+▁bosque 1
+bur 1
+▁arco 1
+▁bra 1
+▁medios 1
+▁anima 1
+▁locales 1
+▁bel 1
+ónica 1
+▁medida 1
+creció 1
+▁organiza 1
+▁carretera 1
+▁permanente 1
+▁cuadra 1
+▁entrenador 1
+▁soldado 1
+▁actuación 1
+▁planeta 1
+▁canto 1
+▁cambiar 1
+tima 1
+ty 1
+▁altura 1
+▁regional 1
+▁reserva 1
+▁abuelo 1
+▁distancia 1
+▁cráter 1
+▁escudo 1
+▁aumenta 1
+▁francesa 1
+▁alberga 1
+▁inmediatamente 1
+▁apena 1
+▁salvador 1
+▁tratamiento 1
+▁pena 1
+ólogo 1
+▁encontraba 1
+▁julio 1
+▁rock 1
+▁adelante 1
+▁árboles 1
+em 1
+▁prepara 1
+▁romano 1
+cida 1
+▁jugadores 1
+var 1
+▁genera 1
+▁arquitecto 1
+▁disponible 1
+▁ministerio 1
+▁amenaza 1
+▁tradición 1
+▁hembra 1
+▁fotografía 1
+▁asociación 1
+▁ruso 1
+▁carne 1
+▁ángeles 1
+▁mejores 1
+▁palabras 1
+om 1
+▁solamente 1
+▁pensar 1
+▁emplea 1
+bert 1
+▁esos 1
+▁ainhoa 1
+▁cuarenta 1
+▁músico 1
+▁idioma 1
+▁cantidad 1
+▁mitad 1
+▁depende 1
+▁voces 1
+▁zo 1
+▁actu 1
+▁creado 1
+up 1
+▁señal 1
+cri 1
+▁italiano 1
+▁metros 1
+▁clases 1
+cá 1
+▁pesca 1
+▁sector 1
+▁evitar 1
+▁cielo 1
+▁medalla 1
+▁error 1
+▁comercio 1
+▁compuesto 1
+ira 1
+▁administración 1
+▁desarroll 1
+▁venta 1
+▁ur 1
+▁ataca 1
+izada 1
+▁encontra 1
+▁dado 1
+▁sha 1
+ist 1
+osas 1
+▁frecuente 1
+▁españoles 1
+▁participa 1
+▁haga 1
+pen 1
+▁construye 1
+▁catedral 1
+▁continúa 1
+▁dirigió 1
+▁izquierda 1
+▁recuerda 1
+▁canadá 1
+▁guitarra 1
+▁hicieron 1
+▁usado 1
+▁áreas 1
+▁comida 1
+▁fuentes 1
+▁críticos 1
+rry 1
+▁diputado 1
+▁programas 1
+▁inferiores 1
+ham 1
+puesta 1
+▁chicos 1
+▁clara 1
+▁desempeña 1
+▁separa 1
+berg 1
+▁grabado 1
+▁dirigida 1
+▁juventud 1
+▁amistad 1
+▁combina 1
+▁agrícola 1
+▁mas 1
+▁delantero 1
+▁diseñado 1
+▁vuelta 1
+▁escala 1
+▁doctora 1
+▁debería 1
+▁pelo 1
+▁demostr 1
+▁fl 1
+▁ayudar 1
+▁adopt 1
+▁dur 1
+mán 1
+▁tierras 1
+▁forman 1
+tano 1
+▁bio 1
+stre 1
+▁energía 1
+▁frecuencia 1
+▁mexicano 1
+▁dispone 1
+▁noroeste 1
+▁obispo 1
+▁primaria 1
+▁europeo 1
+▁turismo 1
+▁libertad 1
+▁dinero 1
+▁rusia 1
+▁garcía 1
+▁eco 1
+▁decora 1
+▁letras 1
+▁inicio 1
+▁am 1
+▁cho 1
+▁anteriormente 1
+▁poli 1
+▁ple 1
+▁marina 1
+▁aves 1
+bri 1
+▁piel 1
+▁mad 1
+ense 1
+▁leche 1
+▁sociales 1
+▁metal 1
+visión 1
+▁caer 1
+▁acompaña 1
+q 1
+▁abre 1
+▁constituye 1
+▁corresponde 1
+▁fundamental 1
+▁huevos 1
+▁imposible 1
+▁vídeo 1
+▁monumento 1
+▁regresó 1
+▁titular 1
+▁sombra 1
+▁cultivo 1
+▁europea 1
+▁afecta 1
+▁borde 1
+▁firma 1
+▁derecha 1
+▁encima 1
+▁pintor 1
+▁carta 1
+iana 1
+▁colonia 1
+bel 1
+▁hered 1
+rien 1
+▁contrario 1
+▁completo 1
+▁nacionales 1
+▁barba 1
+▁estela 1
+▁activo 1
+▁posterior 1
+▁buscar 1
+▁pop 1
+▁clave 1
+▁aprendi 1
+▁investiga 1
+▁flor 1
+▁noble 1
+▁argentino 1
+▁compositor 1
+▁consiguió 1
+▁periodista 1
+▁página 1
+▁negocio 1
+▁brillante 1
+▁conferencia 1
+▁elabora 1
+▁agrupa 1
+▁herido 1
+cito 1
+▁saber 1
+▁llegado 1
+▁clasifica 1
+▁productos 1
+▁nace 1
+▁limita 1
+▁extra 1
+▁vela 1
+▁marc 1
+▁entrega 1
+▁pudo 1
+▁blancas 1
+▁musicales 1
+▁cabe 1
+iano 1
+▁pasos 1
+▁añad 1
+▁ejecuta 1
+▁oportunidad 1
+▁proporciona 1
+▁respuesta 1
+▁noticia 1
+▁dolor 1
+▁distribución 1
+kin 1
+bido 1
+▁habitantes 1
+ima 1
+▁señora 1
+▁basta 1
+▁pico 1
+▁pregunta 1
+▁ubicada 1
+▁unido 1
+plement 1
+▁flora 1
+▁ciencia 1
+▁practica 1
+▁velocidad 1
+▁william 1
+▁contacto 1
+▁poema 1
+▁junio 1
+▁th 1
+▁sh 1
+▁tocar 1
+▁participación 1
+▁crece 1
+rena 1
+▁new 1
+▁repar 1
+▁liberal 1
+▁león 1
+▁esperaba 1
+▁dedicada 1
+▁dra 1
+▁igualmente 1
+▁calles 1
+▁autores 1
+▁vue 1
+ui 1
+▁ayuntamiento 1
+▁colección 1
+▁económico 1
+▁introduc 1
+▁octubre 1
+▁gamboa 1
+▁identifica 1
+▁futuro 1
+▁vecino 1
+▁éste 1
+▁instrumento 1
+▁pueda 1
+▁esperanza 1
+▁producido 1
+▁grave 1
+▁rápidamente 1
+▁condena 1
+▁curs 1
+▁nuevamente 1
+▁guion 1
+▁caza 1
+bro 1
+▁inferior 1
+▁gre 1
+▁cap 1
+▁norma 1
+▁fernando 1
+▁pájaro 1
+▁religiosa 1
+▁capacidad 1
+▁secreto 1
+▁privada 1
+▁diferente 1
+▁lá 1
+▁pelea 1
+▁comisión 1
+▁breve 1
+▁mantener 1
+▁mesa 1
+▁inicialmente 1
+▁salta 1
+▁pista 1
+tru 1
+▁paisaje 1
+▁hacía 1
+▁we 1
+▁minutos 1
+▁aparecen 1
+▁x 1
+tario 1
+▁compone 1
+▁conoci 1
+▁matar 1
+▁colabora 1
+▁científico 1
+▁documento 1
+▁originario 1
+▁primavera 1
+▁dibujo 1
+▁símbolo 1
+▁columna 1
+▁oficina 1
+▁paraguay 1
+▁cuento 1
+▁dedicado 1
+▁piloto 1
+▁apoya 1
+▁cuá 1
+▁milita 1
+▁sierra 1
+▁corea 1
+▁derrota 1
+▁representado 1
+▁sé 1
+cepción 1
+▁camp 1
+▁hacerlo 1
+éis 1
+▁capa 1
+▁fuego 1
+▁representante 1
+caso 1
+▁falleció 1
+▁ferrocarril 1
+▁portugal 1
+▁construido 1
+▁navidad 1
+▁imagen 1
+▁cultiva 1
+▁federal 1
+▁guinea 1
+▁lanzamiento 1
+▁demás 1
+▁flota 1
+▁acusa 1
+▁infancia 1
+▁ganar 1
+▁humano 1
+▁caballos 1
+▁piano 1
+▁semanas 1
+und 1
+▁alcanz 1
+quí 1
+▁corri 1
+▁intento 1
+americana 1
+▁foto 1
+▁contar 1
+▁estrellas 1
+▁piezas 1
+▁salida 1
+▁rara 1
+▁pies 1
+▁adicional 1
+▁histórica 1
+▁insectos 1
+▁lluvia 1
+▁originaria 1
+▁requiere 1
+▁revolución 1
+▁orquesta 1
+▁profunda 1
+▁graduó 1
+▁hierro 1
+ible 1
+▁revela 1
+▁convento 1
+▁marrón 1
+▁tienda 1
+▁contigo 1
+▁punta 1
+▁trabajado 1
+▁presentan 1
+▁ramón 1
+com 1
+▁infl 1
+▁rural 1
+▁hubiera 1
+▁efecto 1
+▁calidad 1
+▁llevan 1
+▁volar 1
+▁ronda 1
+▁activa 1
+▁moderno 1
+▁basada 1
+▁apart 1
+▁niveles 1
+▁tren 1
+▁cien 1
+cina 1
+▁suficiente 1
+▁consecuencia 1
+▁justicia 1
+▁encontró 1
+▁griego 1
+▁comprende 1
+▁quería 1
+▁explica 1
+▁regla 1
+▁gobernador 1
+▁amiga 1
+▁borda 1
+▁comarca 1
+▁exposiciones 1
+▁fri 1
+▁formato 1
+▁ido 1
+▁rosario 1
+▁patio 1
+▁escenario 1
+▁hablando 1
+▁enfrenta 1
+rito 1
+▁revistas 1
+▁dicen 1
+▁fall 1
+▁rico 1
+▁bailar 1
+▁estudiantes 1
+▁salva 1
+fre 1
+▁estatal 1
+▁graba 1
+▁registro 1
+▁robert 1
+▁artículos 1
+zón 1
+▁armas 1
+▁empeza 1
+ang 1
+riendo 1
+▁fábrica 1
+▁resistencia 1
+▁regreso 1
+▁positiva 1
+▁describe 1
+▁nacido 1
+▁nieve 1
+▁proviene 1
+▁comuna 1
+▁fresco 1
+▁debajo 1
+▁reunión 1
+▁peces 1
+▁usar 1
+▁desempeñ 1
+▁evento 1
+▁debía 1
+▁reconocido 1
+▁quedan 1
+▁montañas 1
+▁posteriores 1
+▁masa 1
+▁mc 1
+▁formado 1
+▁grit 1
+▁restaura 1
+▁eventos 1
+▁usada 1
+▁reuni 1
+mó 1
+▁euro 1
+▁martín 1
+▁concurso 1
+▁cubierto 1
+▁mexicana 1
+▁plataforma 1
+▁princesa 1
+▁príncipe 1
+▁artística 1
+▁impuls 1
+▁literaria 1
+▁fruta 1
+▁afirma 1
+▁cierta 1
+▁conecta 1
+▁funciones 1
+▁reconocimiento 1
+▁conmigo 1
+▁enrique 1
+▁dientes 1
+ther 1
+▁ecuador 1
+▁selva 1
+▁arregl 1
+▁compara 1
+▁sangre 1
+▁sección 1
+▁tenían 1
+▁poblado 1
+▁empresas 1
+▁considerable 1
+▁cambia 1
+by 1
+▁brazos 1
+▁moderna 1
+▁enamorado 1
+▁materiales 1
+▁arre 1
+vil 1
+▁utilizan 1
+▁efectos 1
+▁traduc 1
+▁llam 1
+▁recupera 1
+▁comunica 1
+▁gobern 1
+▁candidato 1
+▁córdoba 1
+▁imágenes 1
+▁judicial 1
+▁núcleo 1
+▁posibilidad 1
+▁septiembre 1
+▁sitúa 1
+▁concepto 1
+▁ganadería 1
+▁milagro 1
+▁tormenta 1
+▁congreso 1
+▁naturaleza 1
+▁cuarta 1
+▁crecimiento 1
+▁avanza 1
+▁prensa 1
+▁decisión 1
+▁emigr 1
+▁sabor 1
+ure 1
+▁elecciones 1
+▁gato 1
+▁mud 1
+▁beso 1
+▁peter 1
+▁fac 1
+▁creación 1
+▁finales 1
+▁viven 1
+▁navega 1
+▁extendi 1
+▁estudiar 1
+▁bre 1
+▁ramas 1
+creta 1
+▁tru 1
+▁fila 1
+cada 1
+▁cable 1
+▁plato 1
+▁recursos 1
+▁habitual 1
+▁asesina 1
+▁volv 1
+▁incorpora 1
+ice 1
+▁ascenso 1
+▁construida 1
+▁facultad 1
+▁murmuró 1
+▁necesario 1
+▁víctima 1
+▁caballero 1
+▁regiones 1
+▁adulto 1
+▁dispara 1
+▁conquista 1
+▁aprob 1
+▁suerte 1
+▁modific 1
+▁ed 1
+▁recibido 1
+▁tabla 1
+▁pod 1
+▁pecho 1
+▁abrir 1
+▁húmedos 1
+▁banco 1
+▁et 1
+▁pel 1
+▁piso 1
+▁lope 1
+izar 1
+▁huel 1
+▁bon 1
+▁sor 1
+gráfica 1
+▁diócesis 1
+▁jamás 1
+▁límite 1
+▁monasterio 1
+▁respecto 1
+▁school 1
+▁vehículo 1
+▁implica 1
+▁gloria 1
+▁cuadro 1
+▁ramiro 1
+▁ocasión 1
+▁sonrisa 1
+▁rango 1
+▁establece 1
+▁cerrado 1
+▁cristal 1
+▁producir 1
+▁domingo 1
+▁vocal 1
+▁sacerdote 1
+▁raro 1
+▁metro 1
+▁mancha 1
+quiera 1
+▁factor 1
+▁directo 1
+▁cab 1
+▁desea 1
+▁rob 1
+▁emb 1
+▁descubri 1
+▁misión 1
+▁distin 1
+▁conciertos 1
+▁peligro 1
+▁escribir 1
+▁encarga 1
+▁result 1
+▁conflicto 1
+▁george 1
+▁temperatura 1
+▁usuario 1
+▁enemigo 1
+▁michael 1
+▁secretario 1
+▁jorge 1
+▁riesgo 1
+▁alegre 1
+▁composición 1
+▁tribu 1
+▁rompe 1
+▁coloca 1
+▁agrega 1
+▁editor 1
+▁cocina 1
+▁beber 1
+▁nativa 1
+▁abundante 1
+▁atención 1
+▁constante 1
+▁neu 1
+▁nega 1
+▁deporte 1
+han 1
+▁menores 1
+ight 1
+▁dirige 1
+▁mac 1
+▁you 1
+▁comienzo 1
+▁ingeniero 1
+▁judío 1
+▁respira 1
+▁circula 1
+▁rechaza 1
+▁florece 1
+▁retrato 1
+▁viaja 1
+▁sevilla 1
+▁permaneció 1
+▁bandera 1
+▁bloque 1
+▁conocimiento 1
+▁laguna 1
+▁prisión 1
+▁barra 1
+▁traje 1
+▁publicación 1
+▁dime 1
+▁patas 1
+▁tendrá 1
+▁localidades 1
+▁chico 1
+▁árbol 1
+▁ángel 1
+▁administra 1
+▁militares 1
+▁familiar 1
+▁descansa 1
+▁ceremonia 1
+▁competición 1
+▁contempla 1
+▁extranjero 1
+▁independencia 1
+▁juvenil 1
+▁periodismo 1
+▁protagonista 1
+▁universitario 1
+▁órgano 1
+▁jesús 1
+▁turístico 1
+▁vicente 1
+▁vacaciones 1
+▁privado 1
+▁adquiri 1
+ascendencia 1
+▁seguido 1
+▁avión 1
+▁animal 1
+▁mr 1
+▁cruc 1
+▁contrato 1
+▁operaciones 1
+ting 1
+▁aten 1
+▁river 1
+▁veo 1
+▁mando 1
+▁frecuentemente 1
+▁presentado 1
+▁restaurante 1
+▁guía 1
+ford 1
+▁ropa 1
+▁varía 1
+▁deportes 1
+▁enferm 1
+▁marino 1
+▁vende 1
+▁calma 1
+▁sello 1
+▁incorpor 1
+amiento 1
+▁viva 1
+▁jr 1
+ándome 1
+▁sabi 1
+pp 1
+▁má 1
+▁prof 1
+▁unidad 1
+▁portada 1
+▁presentó 1
+▁veracruz 1
+▁parlamento 1
+▁solicit 1
+▁hambre 1
+▁tropas 1
+▁alimento 1
+▁administrativa 1
+▁tallo 1
+▁recuerdo 1
+▁relacionado 1
+▁fil 1
+▁daño 1
+▁promet 1
+▁perder 1
+▁regular 1
+ómic 1
+▁coge 1
+▁prote 1
+▁papa 1
+▁cercana 1
+▁destac 1
+▁representación 1
+▁anteriores 1
+▁salvar 1
+▁comerciales 1
+▁interna 1
+▁provincial 1
+▁tim 1
+▁provoca 1
+▁lateral 1
+▁imagina 1
+▁oficiales 1
+▁anda 1
+▁aplica 1
+▁autoridad 1
+▁parti 1
+▁especializ 1
+▁colabor 1
+ney 1
+▁cincuenta 1
+▁documental 1
+▁esfuerzo 1
+▁ilustra 1
+▁ocurre 1
+▁secundario 1
+▁tecnología 1
+▁árabe 1
+▁asegur 1
+▁reforma 1
+▁cuenca 1
+▁procede 1
+▁vieja 1
+▁pedir 1
+▁explota 1
+▁condiciones 1
+▁merced 1
+▁existi 1
+▁segura 1
+▁tribunal 1
+▁james 1
+▁protagonizada 1
+▁west 1
+▁retira 1
+dujo 1
+▁herm 1
+▁pescado 1
+▁loco 1
+▁llevado 1
+▁abandona 1
+▁turb 1
+▁rap 1
+▁bill 1
+ville 1
+▁sch 1
+▁somo 1
+▁relaciona 1
+▁destacado 1
+▁alterna 1
+▁adolescente 1
+▁desaparece 1
+▁humilde 1
+▁manifesta 1
+▁sombrero 1
+▁viajó 1
+▁austria 1
+▁panamá 1
+▁guardia 1
+▁frío 1
+▁espacial 1
+▁comparti 1
+▁habría 1
+▁avenida 1
+▁anual 1
+▁diego 1
+▁cinta 1
+▁declara 1
+▁carece 1
+▁archivo 1
+▁hotel 1
+▁produjo 1
+▁recoge 1
+▁concentra 1
+▁angel 1
+▁producto 1
+▁verdadero 1
+▁camina 1
+trás 1
+▁cena 1
+▁ciudadano 1
+▁sentí 1
+bula 1
+▁ll 1
+▁desarrollado 1
+▁terminar 1
+▁creer 1
+▁merc 1
+▁cría 1
+▁casti 1
+▁malo 1
+tch 1
+▁haz 1
+▁brazo 1
+▁viviend 1
+▁crear 1
+▁pega 1
+▁marido 1
+▁búsqueda 1
+▁extensión 1
+▁incluido 1
+▁intensa 1
+▁próximo 1
+▁religioso 1
+▁vegetación 1
+▁acabó 1
+▁college 1
+▁retiró 1
+▁desconoce 1
+▁consigue 1
+▁ensayo 1
+▁internet 1
+cedió 1
+▁bat 1
+▁allá 1
+▁enorme 1
+▁dificultad 1
+▁interesante 1
+▁demanda 1
+bió 1
+▁recientemente 1
+▁socio 1
+▁invitado 1
+▁licenciado 1
+▁escultura 1
+▁tono 1
+▁interpretación 1
+▁capaz 1
+▁habilidad 1
+erra 1
+▁ideal 1
+▁permanec 1
+▁misa 1
+▁participado 1
+umb 1
+▁tiro 1
+▁integr 1
+▁llegan 1
+▁fácil 1
+imiento 1
+▁abadesa 1
+▁británica 1
+▁espectáculo 1
+▁excelente 1
+▁extraña 1
+▁irregular 1
+▁mediterráneo 1
+▁máximo 1
+▁organismo 1
+▁presidencia 1
+▁febrero 1
+▁pasajeros 1
+▁templado 1
+▁entiende 1
+▁suroeste 1
+▁digital 1
+▁volcán 1
+▁transporte 1
+▁escolar 1
+▁humana 1
+▁creó 1
+▁relato 1
+▁abril 1
+▁aroma 1
+▁múltiples 1
+▁aleja 1
+▁sonata 1
+▁abr 1
+▁ejerció 1
+▁detalle 1
+▁clasificación 1
+▁peso 1
+tracción 1
+▁partida 1
+▁parroquia 1
+▁llegaron 1
+▁especiales 1
+▁marco 1
+guer 1
+▁asociado 1
+▁tí 1
+▁experiment 1
+▁conservan 1
+▁llevaba 1
+▁hol 1
+▁franco 1
+▁tratado 1
+▁podi 1
+mbo 1
+oña 1
+▁conseguir 1
+▁construir 1
+▁extraño 1
+▁isabel 1
+▁obtiene 1
+▁responsable 1
+▁temporal 1
+▁temprana 1
+▁divide 1
+▁móvil 1
+▁convertido 1
+▁femenina 1
+▁limpia 1
+▁chris 1
+▁estrecha 1
+▁quinto 1
+▁vivía 1
+▁ruinas 1
+▁llora 1
+▁alumno 1
+▁escasa 1
+ticos 1
+▁publicaciones 1
+▁cala 1
+▁socialista 1
+ificación 1
+▁marzo 1
+▁recuperar 1
+▁podrá 1
+▁suiza 1
+▁cerro 1
+▁bot 1
+▁pala 1
+▁piensa 1
+▁cuales 1
+▁fiel 1
+▁deseo 1
+▁universidades 1
+iga 1
+misión 1
+▁cero 1
+▁inicial 1
+▁piens 1
+▁construyó 1
+▁extensa 1
+▁fantasma 1
+▁jornada 1
+▁oscura 1
+▁usual 1
+▁except 1
+▁profesión 1
+▁show 1
+▁dormir 1
+▁teatral 1
+▁tendencia 1
+▁explora 1
+▁profundo 1
+▁hermoso 1
+▁perenne 1
+▁quince 1
+▁pierna 1
+▁infantil 1
+▁cabello 1
+▁suave 1
+▁obtener 1
+▁gusto 1
+▁cercano 1
+▁palo 1
+▁taller 1
+▁mono 1
+▁grabación 1
+▁calor 1
+▁empezar 1
+▁fija 1
+▁poca 1
+▁ku 1
+▁captura 1
+nominada 1
+▁profesora 1
+▁culturales 1
+▁similares 1
+▁superiores 1
+ive 1
+arias 1
+▁plano 1
+▁clubes 1
+▁casada 1
+▁terminal 1
+ling 1
+▁católico 1
+▁convierte 1
+▁eduardo 1
+▁homenaje 1
+▁rodríguez 1
+▁rápido 1
+▁terrestre 1
+▁sobrino 1
+▁confirma 1
+▁enseñanza 1
+▁aproximadamente 1
+▁alemana 1
+▁aventura 1
+▁editorial 1
+▁daniel 1
+▁famosa 1
+▁batería 1
+▁tarea 1
+▁baile 1
+▁rumbo 1
+firmó 1
+▁nacimiento 1
+tío 1
+▁labios 1
+▁gracia 1
+idez 1
+▁investigador 1
+▁suceso 1
+▁lenguaje 1
+old 1
+▁palma 1
+▁flu 1
+▁generación 1
+▁impresion 1
+disc 1
+▁cop 1
+▁americano 1
+▁hecha 1
+▁escribe 1
+▁componente 1
+▁gri 1
+▁tenga 1
+▁edita 1
+▁moral 1
+▁manda 1
+▁letra 1
+▁populares 1
+▁licencia 1
+▁tendr 1
+▁domina 1
+▁destacan 1
+▁importa 1
+▁instru 1
+▁análisis 1
+▁evidencia 1
+▁ocupó 1
+▁ámbito 1
+▁literario 1
+▁razones 1
+▁refleja 1
+▁catorce 1
+▁arroyo 1
+▁apoyo 1
+▁comunes 1
+▁gigante 1
+▁agosto 1
+▁cueva 1
+▁talento 1
+iforme 1
+▁bronce 1
+▁láser 1
+▁caída 1
+▁psico 1
+▁tercero 1
+▁expuls 1
+▁city 1
+▁rayo 1
+▁caliente 1
+▁retorn 1
+▁sonora 1
+▁joder 1
+▁ayer 1
+▁participar 1
+▁verdadera 1
+▁ze 1
+▁originales 1
+▁levantó 1
+▁music 1
+▁alcanza 1
+▁park 1
+▁unidades 1
+gram 1
+ku 1
+▁ampliamente 1
+▁separado 1
+▁esperar 1
+tuve 1
+▁realizada 1
+▁coleccion 1
+▁duro 1
+▁hip 1
+▁enamora 1
+▁utilizada 1
+▁invit 1
+▁produc 1
+▁calla 1
+▁ferr 1
+▁reciente 1
+acional 1
+▁contribuy 1
+nunció 1
+▁abandonó 1
+▁asamblea 1
+▁cataluña 1
+▁clásica 1
+▁diciembre 1
+▁filipinas 1
+▁físico 1
+▁herramienta 1
+▁ilumina 1
+▁noviembre 1
+▁otoño 1
+▁reconoce 1
+▁vietnam 1
+▁artístico 1
+▁emperador 1
+▁texas 1
+▁entrevista 1
+▁cristiano 1
+▁griega 1
+▁lleno 1
+▁migra 1
+▁desperta 1
+▁instituciones 1
+▁enfrentamiento 1
+▁abrió 1
+▁procesa 1
+▁bonita 1
+▁característica 1
+▁justo 1
+▁protesta 1
+▁esencial 1
+▁chu 1
+▁narra 1
+▁subtropicales 1
+▁sexo 1
+▁huye 1
+▁monja 1
+▁hici 1
+▁repit 1
+▁estren 1
+▁registr 1
+▁autoridades 1
+▁teología 1
+▁she 1
+▁siente 1
+▁onda 1
+▁utilizar 1
+nzó 1
+▁besa 1
+▁realmente 1
+▁alcanzar 1
+▁tío 1
+▁tropical 1
+▁arm 1
+▁materia 1
+▁silla 1
+zco 1
+▁crec 1
+▁colaborador 1
+▁descubierto 1
+▁específica 1
+▁indígena 1
+▁inflorescencia 1
+▁inglesa 1
+▁mensaje 1
+▁religión 1
+▁republicano 1
+▁temprano 1
+▁zaragoza 1
+▁aparición 1
+▁deportiva 1
+▁máquina 1
+stituido 1
+▁marquesa 1
+▁pantalla 1
+▁precisa 1
+▁sirvió 1
+▁visual 1
+▁urbano 1
+▁ediciones 1
+▁asunto 1
+▁rodeado 1
+▁alberto 1
+▁habrá 1
+▁quinta 1
+▁producida 1
+▁rueda 1
+▁archi 1
+▁grecia 1
+▁steve 1
+▁peña 1
+▁prisa 1
+▁tumba 1
+▁lanzado 1
+▁criado 1
+▁king 1
+▁alarg 1
+▁impacto 1
+▁operación 1
+▁precio 1
+▁existencia 1
+▁saque 1
+▁invent 1
+▁perdona 1
+▁franc 1
+▁sam 1
+▁encarg 1
+gh 1
+▁tanta 1
+▁centra 1
+▁comedia 1
+▁wo 1
+grado 1
+▁quita 1
+▁enterra 1
+▁acompañado 1
+▁conocer 1
+▁finalizar 1
+▁exclusiva 1
+▁femenino 1
+▁identidad 1
+▁olímpico 1
+▁rectangular 1
+▁sudáfrica 1
+▁tailandia 1
+▁telenovela 1
+▁gestiona 1
+▁quizá 1
+▁hierba 1
+▁equipa 1
+▁servir 1
+▁desnudo 1
+▁bolsa 1
+▁pierde 1
+▁mixta 1
+▁rostro 1
+▁cierra 1
+▁alternativa 1
+▁equivalente 1
+▁herida 1
+▁sexual 1
+▁oído 1
+▁once 1
+▁cubre 1
+▁ligeramente 1
+▁mortal 1
+▁mirada 1
+lfa 1
+▁directamente 1
+▁concierto 1
+▁tramo 1
+▁participaron 1
+▁senador 1
+well 1
+▁fama 1
+▁particularmente 1
+▁celos 1
+▁amplio 1
+ístico 1
+▁descubr 1
+mont 1
+▁roberto 1
+▁vega 1
+▁lima 1
+▁pese 1
+▁presta 1
+▁pino 1
+▁chino 1
+ix 1
+▁papeles 1
+▁terminado 1
+▁adelant 1
+▁transforma 1
+▁interpret 1
+▁espíritu 1
+▁gonzález 1
+▁inmigrantes 1
+▁murmullo 1
+▁turquía 1
+▁divisiones 1
+▁alemanes 1
+▁ladrillo 1
+▁felipe 1
+▁riqueza 1
+▁ácido 1
+▁bahía 1
+▁pablo 1
+▁refiere 1
+vertirse 1
+▁presidi 1
+▁berlín 1
+▁vincula 1
+▁ficción 1
+▁asesor 1
+▁expresa 1
+▁martínez 1
+▁zapat 1
+siguieron 1
+siguiendo 1
+▁interrog 1
+▁dorado 1
+▁suicid 1
+▁copia 1
+▁aplicaciones 1
+▁significado 1
+▁max 1
+▁top 1
+▁variable 1
+▁legal 1
+▁colaboración 1
+▁romana 1
+▁elemento 1
+▁danza 1
+▁tomado 1
+▁oír 1
+▁provoc 1
+vió 1
+▁orilla 1
+▁experimenta 1
+▁asocia 1
+▁declar 1
+▁eje 1
+▁actores 1
+▁retirado 1
+▁profesionales 1
+▁impe 1
+▁abandonar 1
+posición 1
+▁ecua 1
+▁apariencia 1
+▁científica 1
+▁fenómeno 1
+▁guitarrista 1
+▁necesaria 1
+▁próxima 1
+▁alcance 1
+▁solitario 1
+▁transcur 1
+▁paciente 1
+▁gafas 1
+▁elección 1
+▁urbana 1
+▁leyenda 1
+▁crema 1
+▁ejerce 1
+▁conforma 1
+▁galería 1
+▁significativa 1
+▁suecia 1
+▁pariente 1
+▁favorito 1
+▁contaba 1
+▁duque 1
+▁primo 1
+▁fiera 1
+▁casco 1
+▁sonreí 1
+▁fácilmente 1
+▁publicada 1
+▁olor 1
+▁salón 1
+▁denominación 1
+rgió 1
+▁ruido 1
+▁trabajando 1
+▁pieza 1
+▁sacar 1
+▁concepción 1
+▁clo 1
+▁nombra 1
+▁cumpli 1
+▁rose 1
+▁basado 1
+▁pareció 1
+▁descendiente 1
+▁diagrama 1
+▁extiende 1
+▁kilómetro 1
+▁matemática 1
+▁pertenecía 1
+▁relieve 1
+▁renuncia 1
+▁arbusto 1
+▁aprecia 1
+▁escritura 1
+▁estrib 1
+▁apertura 1
+▁escalera 1
+▁coronel 1
+▁imperial 1
+▁cuerda 1
+▁ancla 1
+▁marinero 1
+▁ahoga 1
+▁encanta 1
+▁asusta 1
+▁soporta 1
+▁multi 1
+▁convoca 1
+▁aprender 1
+▁louis 1
+▁estatua 1
+▁micro 1
+▁cliente 1
+▁castellano 1
+▁ji 1
+▁interesa 1
+▁institución 1
+▁rival 1
+▁lic 1
+▁maria 1
+▁trabajadores 1
+▁corteza 1
+▁mandato 1
+▁pasión 1
+gel 1
+▁regal 1
+▁linda 1
+ini 1
+▁reales 1
+▁quedaba 1
+▁paseo 1
+▁radical 1
+▁armada 1
+▁chileno 1
+▁propiedades 1
+▁repent 1
+▁cruzar 1
+▁ayud 1
+▁informe 1
+▁frase 1
+▁vota 1
+▁pot 1
+▁derrotado 1
+▁artificial 1
+▁código 1
+▁fortaleza 1
+▁futbolista 1
+▁guerrero 1
+▁incidente 1
+▁federación 1
+▁multitud 1
+▁richard 1
+▁dibuja 1
+▁globo 1
+▁sabía 1
+▁recorre 1
+▁visible 1
+▁oreja 1
+▁electoral 1
+▁liceo 1
+▁ritmo 1
+▁trayectoria 1
+▁boston 1
+▁single 1
+▁propósito 1
+▁dominio 1
+▁florida 1
+▁bordo 1
+▁encaja 1
+▁reclam 1
+▁patrona 1
+▁recibir 1
+▁interven 1
+cimiento 1
+▁tesis 1
+▁nativo 1
+▁pensaba 1
+▁presión 1
+▁individuales 1
+▁reconocida 1
+wood 1
+▁especializado 1
+▁traza 1
+agua 1
+▁empleado 1
+▁artículo 1
+▁programación 1
+▁mora 1
+▁suyo 1
+▁instala 1
+mita 1
+▁redes 1
+▁inform 1
+▁organizado 1
+▁acontecimiento 1
+▁bóveda 1
+▁colectivo 1
+▁competencia 1
+▁defensor 1
+▁egipto 1
+▁ejemplar 1
+▁máxima 1
+▁prisionero 1
+▁voluntario 1
+▁washington 1
+▁desemboca 1
+▁volumen 1
+▁aceite 1
+▁rebelde 1
+▁hidro 1
+▁millones 1
+▁ermita 1
+▁cuello 1
+▁ciclo 1
+▁league 1
+▁fiscal 1
+▁repeti 1
+▁esposo 1
+▁poblaciones 1
+▁catalan 1
+▁defendi 1
+▁noreste 1
+▁juzg 1
+▁marqués 1
+▁cómic 1
+▁anillo 1
+▁granada 1
+▁dueño 1
+▁propuesta 1
+▁oriente 1
+lito 1
+▁barroco 1
+▁viena 1
+▁aldea 1
+▁novia 1
+▁popularidad 1
+▁rusa 1
+▁aguanta 1
+▁laterales 1
+rich 1
+▁sectores 1
+▁anot 1
+▁oficialmente 1
+▁informa 1
+▁darle 1
+▁mate 1
+▁astr 1
+▁eva 1
+▁ram 1
+▁chap 1
+▁obten 1
+▁manten 1
+▁pluma 1
+▁baloncesto 1
+▁camerún 1
+▁cerámica 1
+▁democrática 1
+▁guillermo 1
+▁hacienda 1
+▁individuo 1
+▁intercambi 1
+▁involucra 1
+▁pacífico 1
+▁plástico 1
+▁videojuego 1
+▁básica 1
+▁dinamarca 1
+▁necesidad 1
+▁química 1
+▁thomas 1
+▁cambió 1
+▁mecanismo 1
+▁típica 1
+▁violencia 1
+▁vuelva 1
+▁asistente 1
+▁temática 1
+▁ópera 1
+▁posiciones 1
+▁protección 1
+▁cosecha 1
+▁paredes 1
+▁irlanda 1
+▁almacena 1
+▁desierto 1
+▁fundó 1
+▁lectura 1
+▁dependien 1
+▁detenido 1
+▁culto 1
+▁promocion 1
+▁mayordomo 1
+▁amarilla 1
+▁condición 1
+▁rumor 1
+▁moneda 1
+▁teniente 1
+▁sentir 1
+▁ahorr 1
+▁seguía 1
+▁caga 1
+▁chicago 1
+▁constitución 1
+▁fría 1
+▁buque 1
+▁paloma 1
+▁dividi 1
+▁tendría 1
+▁asesinato 1
+▁tranquila 1
+▁hil 1
+▁opt 1
+▁juro 1
+▁placa 1
+▁vasco 1
+▁peor 1
+▁my 1
+▁roto 1
+▁preparado 1
+▁estancia 1
+▁adam 1
+▁hill 1
+marca 1
+▁escapar 1
+▁trasladado 1
+▁cantar 1
+▁marcado 1
+▁luch 1
+▁robo 1
+▁distribu 1
+▁asoma 1
+▁misterio 1
+▁actitud 1
+▁ambienta 1
+▁circuito 1
+▁comportamiento 1
+▁contexto 1
+▁controversia 1
+▁desaparecido 1
+▁desconocida 1
+▁embajador 1
+▁fernández 1
+▁hábito 1
+▁israel 1
+▁lópez 1
+▁órdenes 1
+▁alegría 1
+▁moscú 1
+▁oculta 1
+▁belleza 1
+▁cerebro 1
+▁navarra 1
+▁negocia 1
+▁ofreci 1
+▁describi 1
+▁vigila 1
+▁transmiti 1
+▁humor 1
+▁seminario 1
+▁curva 1
+▁poseía 1
+▁reloj 1
+▁cristiana 1
+▁costera 1
+▁emiti 1
+▁administrativo 1
+▁denominado 1
+▁sureste 1
+▁distribuido 1
+▁entidad 1
+▁comentario 1
+▁sucede 1
+▁front 1
+▁semifinales 1
+▁reparti 1
+▁muchacho 1
+▁venir 1
+movido 1
+▁hiper 1
+▁habitualmente 1
+▁retirada 1
+▁daba 1
+▁borr 1
+▁bello 1
+▁simplemente 1
+▁edifica 1
+▁aprovecha 1
+ugh 1
+▁expone 1
+▁manga 1
+▁comunidades 1
+▁estudiante 1
+años 1
+▁familiares 1
+▁vete 1
+▁martin 1
+▁intent 1
+cientos 1
+▁capturado 1
+▁trabajaba 1
+▁cuida 1
+▁preocupa 1
+▁cumple 1
+▁prestigio 1
+▁amarillento 1
+▁apasionad 1
+▁contemporánea 1
+▁círculo 1
+▁específico 1
+▁evolución 1
+▁gracioso 1
+▁húmeda 1
+▁incluida 1
+▁indonesia 1
+▁ingeniería 1
+▁montevideo 1
+▁parcialmente 1
+▁prefiere 1
+▁propietario 1
+▁rafael 1
+▁universo 1
+▁virtud 1
+▁felices 1
+▁rojizo 1
+▁gabriel 1
+▁típico 1
+▁medieval 1
+▁espalda 1
+▁primario 1
+▁alianza 1
+▁huerto 1
+scripción 1
+▁vuelto 1
+▁oferta 1
+▁murieron 1
+▁esquina 1
+▁jackson 1
+▁high 1
+▁fracaso 1
+▁reseña 1
+▁signo 1
+▁reemplazado 1
+▁asign 1
+▁misterioso 1
+▁juez 1
+▁iniciativa 1
+▁combati 1
+▁agente 1
+▁judía 1
+▁espino 1
+▁enviado 1
+ística 1
+▁mov 1
+▁encend 1
+▁oposición 1
+▁finaliza 1
+▁feria 1
+▁educado 1
+▁regularmente 1
+▁debate 1
+▁gener 1
+▁confia 1
+▁califica 1
+▁trenes 1
+▁blo 1
+▁presa 1
+cenas 1
+ington 1
+▁techo 1
+▁declarado 1
+▁mantuv 1
+▁clasificado 1
+▁parroquial 1
+▁llamaba 1
+▁jugado 1
+venció 1
+▁rápida 1
+▁pint 1
+▁brav 1
+▁aproxima 1
+ö 1
+▁aficionado 1
+▁antigüedad 1
+▁autónoma 1
+▁bradomín 1
+▁circunstancia 1
+▁comandante 1
+▁cordillera 1
+▁envidia 1
+▁georgia 1
+▁metropolitana 1
+▁orígenes 1
+▁retablo 1
+▁naranja 1
+▁positivo 1
+▁preferido 1
+▁testigo 1
+▁raíz 1
+▁fragata 1
+▁prohibi 1
+▁audiencia 1
+▁someti 1
+▁promoción 1
+▁asesino 1
+▁innova 1
+▁mudanza 1
+▁campana 1
+▁empleo 1
+▁ésta 1
+▁perfil 1
+▁embarc 1
+▁saint 1
+▁supone 1
+▁juicio 1
+▁nieto 1
+sentimiento 1
+▁salvaje 1
+▁retirarse 1
+▁instalaciones 1
+▁variada 1
+▁naval 1
+▁predominante 1
+▁cabal 1
+▁trae 1
+▁pase 1
+▁especializada 1
+▁arran 1
+▁tranquilo 1
+▁cobra 1
+▁planea 1
+▁recorrido 1
+▁separada 1
+▁querido 1
+▁encontraron 1
+▁mozo 1
+▁ubicación 1
+▁activ 1
+▁clav 1
+▁dividida 1
+▁cumplir 1
+▁cambiado 1
+▁potencia 1
+cillo 1
+field 1
+line 1
+▁rem 1
+▁promov 1
+▁integrante 1
+▁mark 1
+▁abandonado 1
+▁jord 1
+▁ech 1
+▁regula 1
+▁individual 1
+▁educ 1
+dista 1
+▁disfruta 1
+▁predomina 1
+▁grab 1
+incl 1
+▁perteneci 1
+▁fui 1
+▁apellido 1
+▁arqueológico 1
+▁automóvil 1
+▁bretaña 1
+▁característico 1
+▁constitucional 1
+▁cálido 1
+▁desconocido 1
+▁dimensiones 1
+▁dispositivo 1
+▁disputó 1
+▁ejecutivo 1
+▁embajada 1
+▁esclavo 1
+▁incluir 1
+▁intelectual 1
+▁longitud 1
+▁ocurrió 1
+▁patrimonio 1
+▁península 1
+▁pálido 1
+▁sorprende 1
+▁sorpresa 1
+▁utilizó 1
+▁configura 1
+▁orgullo 1
+▁refugio 1
+▁reacción 1
+▁exhibi 1
+▁ocasionalmente 1
+▁sencilla 1
+▁previamente 1
+▁lentamente 1
+▁recinto 1
+▁carolina 1
+▁abuela 1
+▁cierre 1
+▁brown 1
+▁excava 1
+▁otorga 1
+▁sustancia 1
+▁espejo 1
+▁novio 1
+▁obrero 1
+▁cuánto 1
+glo 1
+▁destinado 1
+▁cancela 1
+▁licenciatura 1
+▁paralela 1
+▁entorno 1
+▁pobl 1
+▁preocup 1
+vés 1
+▁magos 1
+▁obtenido 1
+▁significativo 1
+▁renov 1
+▁matorrales 1
+house 1
+▁verá 1
+▁toque 1
+▁actuar 1
+stein 1
+▁lector 1
+▁clar 1
+▁conservación 1
+▁variante 1
+▁lisa 1
+▁traslada 1
+▁beca 1
+▁amante 1
+▁cima 1
+▁dividido 1
+▁visitante 1
+▁miel 1
+▁enfermo 1
+jón 1
+▁rama 1
+▁bob 1
+▁subir 1
+▁prometi 1
+▁pudie 1
+▁perdon 1
+▁atletismo 1
+▁básico 1
+▁cañón 1
+▁destrucción 1
+▁evidente 1
+▁felicidad 1
+▁gestión 1
+▁izquierdo 1
+▁rechazó 1
+▁renacentista 1
+▁responsabilidad 1
+▁subespecie 1
+▁ucrania 1
+▁voluntad 1
+▁exacta 1
+▁geográfica 1
+▁guzmán 1
+▁sábado 1
+▁temblor 1
+▁atravesa 1
+▁continente 1
+▁dormido 1
+▁respectiva 1
+▁proclam 1
+▁recomend 1
+▁jazz 1
+▁hombro 1
+▁instante 1
+▁discoteca 1
+▁pastilla 1
+▁rubia 1
+▁charles 1
+▁maestra 1
+▁distingui 1
+▁vestido 1
+▁domin 1
+▁dispuesto 1
+▁intereses 1
+▁pensamiento 1
+▁crisis 1
+▁perteneciente 1
+▁tonto 1
+▁cristo 1
+▁castilla 1
+▁entiendo 1
+▁apaga 1
+▁hueso 1
+tropel 1
+▁ancho 1
+▁green 1
+▁receta 1
+▁padec 1
+▁jardines 1
+▁huir 1
+▁jack 1
+▁miró 1
+▁margen 1
+▁registrado 1
+andr 1
+▁agita 1
+▁docente 1
+▁italiana 1
+centr 1
+▁raza 1
+ndar 1
+nacida 1
+▁localizada 1
+▁plate 1
+▁pez 1
+▁corro 1
+▁realista 1
+▁conservador 1
+▁limitada 1
+▁vini 1
+▁exposi 1
+▁edit 1
+▁distinta 1
+▁estuvi 1
+▁recogi 1
+▁consider 1
+▁abundan 1
+▁colonial 1
+▁desespera 1
+▁caballerango 1
+▁centrocampista 1
+▁competiciones 1
+▁demócrata 1
+▁desembarc 1
+▁guatemala 1
+▁homónima 1
+▁incluía 1
+▁manuscrito 1
+▁territorial 1
+▁alejandro 1
+▁borneo 1
+▁cáncer 1
+▁interrumpi 1
+▁respondi 1
+▁lágrimas 1
+▁oxford 1
+▁zelanda 1
+▁inquieta 1
+▁august 1
+▁franceses 1
+▁cumpleaños 1
+▁muralla 1
+▁golfo 1
+▁chaleco 1
+▁frank 1
+▁alfredo 1
+▁negó 1
+ão 1
+▁concejo 1
+▁ejercicio 1
+▁manual 1
+▁protector 1
+▁actuaciones 1
+▁comunista 1
+▁remonta 1
+▁inunda 1
+▁tortura 1
+▁polvo 1
+▁tigre 1
+▁expedición 1
+▁ventaja 1
+zuel 1
+▁publicidad 1
+▁cobre 1
+▁personalidad 1
+▁definitivamente 1
+▁enlace 1
+▁raya 1
+▁pudiera 1
+▁alimentación 1
+▁nudos 1
+▁colga 1
+▁encarn 1
+▁asom 1
+▁trono 1
+▁mediano 1
+▁mapa 1
+▁descon 1
+ward 1
+▁potencial 1
+▁señala 1
+▁radic 1
+▁barro 1
+▁alguno 1
+▁recuper 1
+volver 1
+▁directa 1
+▁combustible 1
+▁contemporáneo 1
+▁continental 1
+▁cráneo 1
+▁diversidad 1
+▁ejecución 1
+▁financiero 1
+▁infraestructura 1
+▁sobrevivir 1
+▁vicepresidente 1
+▁acostumbra 1
+▁cohete 1
+▁colectiva 1
+▁expresión 1
+▁extinción 1
+▁salamanca 1
+▁santuario 1
+▁conducir 1
+▁anuncia 1
+▁bienvenido 1
+▁trofeo 1
+▁patrocin 1
+▁inestable 1
+▁protegido 1
+▁sudoeste 1
+▁estrecho 1
+▁manzana 1
+▁setenta 1
+▁descrito 1
+▁homosexual 1
+▁fauna 1
+▁jaime 1
+▁solución 1
+▁envenena 1
+▁descarga 1
+▁mierda 1
+▁entender 1
+▁infini 1
+▁sabana 1
+▁atrapa 1
+▁pulsa 1
+▁narrativa 1
+▁crónica 1
+▁consume 1
+▁patrici 1
+▁flujo 1
+▁cayó 1
+▁lanzada 1
+▁mostrar 1
+▁sujeta 1
+▁producciones 1
+▁firme 1
+▁perfectamente 1
+▁lento 1
+▁acero 1
+▁reduc 1
+▁ficha 1
+▁gem 1
+▁empresario 1
+▁carmen 1
+▁traducción 1
+▁goza 1
+▁modern 1
+▁toros 1
+▁solía 1
+▁baila 1
+▁iris 1
+▁mineral 1
+▁sabr 1
+▁juli 1
+▁island 1
+▁nombr 1
+▁crítico 1
+▁desplaza 1
+▁creativ 1
+▁enviar 1
+▁probable 1
+▁design 1
+▁andalucía 1
+▁arrecife 1
+▁asentamiento 1
+▁cadáver 1
+▁costumbre 1
+▁descendencia 1
+▁embarazada 1
+▁ernesto 1
+▁ingrediente 1
+▁maestría 1
+▁mamíferos 1
+▁nicaragua 1
+▁opinión 1
+▁púrpura 1
+▁sacristía 1
+▁videoclip 1
+▁acceder 1
+▁intenso 1
+▁javier 1
+▁maniobra 1
+▁prolonga 1
+▁resistente 1
+▁destruida 1
+▁tixul 1
+▁renacimiento 1
+▁sicilia 1
+▁veinti 1
+▁articula 1
+▁noruega 1
+▁aterriza 1
+▁mínimo 1
+▁melodía 1
+▁black 1
+▁financia 1
+▁arriesg 1
+▁propone 1
+▁mitología 1
+▁trece 1
+▁patrón 1
+▁elegir 1
+▁ausencia 1
+▁disney 1
+▁vencido 1
+▁ornamental 1
+▁vertical 1
+▁sentencia 1
+▁oliva 1
+▁emite 1
+▁hueco 1
+▁whit 1
+▁adiós 1
+▁afili 1
+▁descans 1
+▁latín 1
+▁enfada 1
+▁warne 1
+▁misiones 1
+▁predic 1
+▁grabaciones 1
+▁instalación 1
+▁salvo 1
+▁madura 1
+▁aplicación 1
+▁place 1
+▁decoración 1
+▁estándar 1
+nutri 1
+▁querer 1
+▁encontrarse 1
+▁hice 1
+pusieron 1
+▁funcionario 1
+▁sob 1
+▁escapa 1
+ducto 1
+▁aporta 1
+▁cristi 1
+▁interno 1
+▁hund 1
+▁grita 1
+▁definitiva 1
+▁culpa 1
+ógrafo 1
+▁aceptó 1
+▁aerolínea 1
+▁azúcar 1
+▁canadiense 1
+▁cápsula 1
+▁evolucion 1
+▁grijalba 1
+▁gómez 1
+▁homónimo 1
+▁ingleses 1
+▁inteligente 1
+▁nicolás 1
+▁nocturno 1
+▁sepultado 1
+▁sobrevivió 1
+▁triunfo 1
+▁universitaria 1
+▁aragón 1
+▁dominicana 1
+▁finlandia 1
+▁honduras 1
+▁monstruos 1
+▁sucedido 1
+▁violenta 1
+▁cervantes 1
+▁imparti 1
+scripciones 1
+▁laboratorio 1
+▁mayoritaria 1
+▁raíces 1
+▁bachillerato 1
+▁wilson 1
+▁gótico 1
+▁polonia 1
+▁smith 1
+▁forestal 1
+▁comerciante 1
+▁rescate 1
+▁regimiento 1
+▁soledad 1
+▁catalina 1
+▁fatiga 1
+▁comparte 1
+▁arrastra 1
+ástica 1
+▁arrestado 1
+▁introduj 1
+▁consumo 1
+▁piscina 1
+▁cercanía 1
+▁aviones 1
+▁phil 1
+▁sudeste 1
+▁transmisión 1
+▁sesenta 1
+▁rocosa 1
+▁cerrada 1
+lusiones 1
+▁oficio 1
+▁nominado 1
+▁culpable 1
+▁dependencia 1
+▁volante 1
+▁kenia 1
+▁llano 1
+▁campus 1
+▁reposa 1
+▁volvie 1
+town 1
+▁atenas 1
+▁dominante 1
+▁drama 1
+▁milán 1
+▁uh 1
+▁totalidad 1
+▁cuya 1
+▁crono 1
+▁trigo 1
+▁pleno 1
+▁levant 1
+▁ál 1
+▁milit 1
+▁marea 1
+luci 1
+▁húmedo 1
+▁american 1
+▁protagoniz 1
+▁disminu 1
+güenza 1
+▁accesible 1
+▁alexander 1
+▁aparecido 1
+▁atractivo 1
+▁atraviesa 1
+▁botánico 1
+▁caña 1
+▁chocolate 1
+▁cinematográfica 1
+▁dieciocho 1
+▁dinastía 1
+▁diplomático 1
+▁facilita 1
+▁fórmula 1
+▁fósiles 1
+▁inmediato 1
+▁inteligencia 1
+▁perspectiva 1
+▁pálida 1
+▁simboliza 1
+▁coordina 1
+▁diablo 1
+▁dispersa 1
+▁equivoca 1
+▁parecido 1
+▁rugby 1
+▁intención 1
+scendió 1
+▁bautista 1
+▁catalán 1
+▁preciso 1
+▁jinete 1
+▁textil 1
+▁criatura 1
+▁protegida 1
+▁existía 1
+▁vaquero 1
+▁serpiente 1
+▁útil 1
+▁maíz 1
+▁admira 1
+▁electro 1
+▁experto 1
+▁entrenamiento 1
+▁ribera 1
+▁grueso 1
+▁brigada 1
+▁compite 1
+▁curiosa 1
+▁arroja 1
+▁sucesión 1
+▁romance 1
+▁caudal 1
+▁tolera 1
+▁malasia 1
+▁tocó 1
+▁tejido 1
+▁vendido 1
+▁gravedad 1
+▁fresca 1
+▁descubrimiento 1
+▁dorada 1
+▁detener 1
+▁rasgo 1
+▁participantes 1
+▁fuimos 1
+▁silencioso 1
+▁peligroso 1
+▁diseñada 1
+▁organizaciones 1
+▁imita 1
+▁grasa 1
+▁uva 1
+▁listo 1
+▁north 1
+iéndome 1
+▁saldr 1
+eiro 1
+cultura 1
+lecciones 1
+▁ampli 1
+▁roll 1
+▁bast 1
+pósito 1
+▁absoluto 1
+▁bogotá 1
+▁cristianismo 1
+▁depresión 1
+▁descenso 1
+▁diplomática 1
+▁disfraz 1
+▁envuelto 1
+▁escultor 1
+▁estrategia 1
+▁horizontal 1
+▁horizonte 1
+▁infierno 1
+▁joseph 1
+▁oscuridad 1
+▁procedimiento 1
+▁protocolo 1
+▁sebastián 1
+▁síntoma 1
+▁universal 1
+▁veinticuatro 1
+▁órbita 1
+▁absoluta 1
+▁cenizas 1
+▁computador 1
+▁cuartel 1
+▁rebelión 1
+▁satisfac 1
+▁argumento 1
+▁compromet 1
+▁limpio 1
+▁madruga 1
+▁montañosa 1
+▁málaga 1
+▁repertorio 1
+▁campamento 1
+▁supongo 1
+▁conexión 1
+dolfo 1
+▁estadística 1
+▁responde 1
+▁turística 1
+▁cruel 1
+▁avance 1
+▁recordar 1
+▁nigeria 1
+▁aborda 1
+▁diciendo 1
+▁estremecimiento 1
+▁recopila 1
+▁millón 1
+onomía 1
+▁turistas 1
+▁defender 1
+▁gruesa 1
+▁hielo 1
+▁dictadura 1
+▁royal 1
+▁emergencia 1
+▁pretende 1
+▁esquema 1
+▁huyó 1
+▁reducida 1
+▁reducir 1
+▁robust 1
+▁plumaje 1
+▁cuyo 1
+▁tarda 1
+▁vecina 1
+▁chaval 1
+▁helado 1
+▁yacimiento 1
+▁crucero 1
+▁falsa 1
+▁complet 1
+▁designado 1
+▁traducido 1
+▁amito 1
+ish 1
+▁sorprendi 1
+▁seleccionado 1
+▁dirigente 1
+▁incluy 1
+▁minas 1
+▁seleccion 1
+▁diseñ 1
+▁cli 1
+▁comercializa 1
+vuelta 1
+▁recrea 1
+lusión 1
+▁autoriza 1
+▁bolívar 1
+▁botánica 1
+▁cambridge 1
+▁camiseta 1
+▁correspondiente 1
+▁electrónica 1
+▁electrónico 1
+▁extranjera 1
+▁grisáceo 1
+▁hermandad 1
+▁hollywood 1
+▁joaquín 1
+▁marruecos 1
+▁mediocampista 1
+▁médico 1
+▁nuclear 1
+▁párpados 1
+▁secuestra 1
+▁software 1
+▁tiburones 1
+▁valladolid 1
+▁yucatán 1
+▁ábside 1
+▁acredita 1
+▁altitud 1
+▁columbia 1
+▁conmemora 1
+▁consolida 1
+▁inaugura 1
+▁líquido 1
+▁national 1
+▁ochenta 1
+▁privilegi 1
+▁romántico 1
+▁triunfa 1
+▁campesino 1
+▁castaño 1
+▁cerveza 1
+▁humanidad 1
+▁exitosa 1
+▁solitaria 1
+▁disparo 1
+scendido 1
+▁terrible 1
+▁precedid 1
+▁apoyó 1
+▁henry 1
+▁pionero 1
+▁increment 1
+▁cumbre 1
+▁centenar 1
+sistieron 1
+▁deposita 1
+▁oveja 1
+▁mascota 1
+▁proporci 1
+▁congrega 1
+▁gemelo 1
+▁bebida 1
+▁divina 1
+▁toledo 1
+▁edward 1
+▁herencia 1
+▁conozco 1
+▁correcta 1
+▁arroz 1
+▁penetr 1
+▁rioja 1
+▁ajusta 1
+▁merece 1
+▁insult 1
+▁serbia 1
+▁rastro 1
+▁tomás 1
+▁conocía 1
+▁suelta 1
+▁escocia 1
+▁plateados 1
+▁resolver 1
+▁inclina 1
+▁conductor 1
+▁conocieron 1
+▁habana 1
+▁gordo 1
+▁venganza 1
+▁tony 1
+▁reporta 1
+▁funcionamiento 1
+▁gru 1
+▁michel 1
+▁plantea 1
+▁fibra 1
+▁mosca 1
+▁prat 1
+▁electo 1
+▁probar 1
+▁elige 1
+▁afro 1
+▁pregunt 1
+▁licenci 1
+▁confund 1
+poniendo 1
+▁afect 1
+▁afluente 1
+▁aparecieron 1
+▁arquitectónico 1
+▁biblia 1
+▁convertiría 1
+▁cuestión 1
+▁cádiz 1
+▁difusión 1
+▁discípulo 1
+▁financiera 1
+▁hazaña 1
+▁impreso 1
+▁japonesa 1
+▁legislatura 1
+▁madagascar 1
+▁magnífic 1
+▁modalidad 1
+▁muñecas 1
+▁nocturna 1
+▁octavo 1
+▁paralímpico 1
+▁prominente 1
+▁pronuncia 1
+▁taiwán 1
+▁virginia 1
+▁llanura 1
+▁lorenzo 1
+▁semejante 1
+▁acogida 1
+▁asumió 1
+▁garganta 1
+▁reflexi 1
+▁scott 1
+▁tanzania 1
+▁vientre 1
+▁concejal 1
+▁externa 1
+▁despacio 1
+▁puntua 1
+▁redactor 1
+▁reúne 1
+▁hoguera 1
+▁café 1
+▁teclado 1
+▁traición 1
+▁venecia 1
+▁simón 1
+▁relaja 1
+▁certifica 1
+terapia 1
+▁residi 1
+▁moreno 1
+▁magia 1
+▁pidió 1
+▁coloniza 1
+▁favorable 1
+▁meteor 1
+▁lineal 1
+▁cuesta 1
+▁suspendi 1
+ándola 1
+▁miami 1
+▁distingue 1
+▁campanario 1
+▁confundir 1
+▁honra 1
+▁reuniones 1
+▁conservatorio 1
+▁distintivo 1
+▁reducido 1
+▁compuso 1
+▁prado 1
+▁postura 1
+▁peligrosa 1
+▁impone 1
+▁alinea 1
+▁diaria 1
+▁motiv 1
+▁extremad 1
+▁reconoci 1
+▁perdie 1
+▁enfrent 1
+▁impresi 1
+cuerdo 1
+▁desgracia 1
+▁adquiere 1
+▁agustín 1
+▁ajedrez 1
+▁beneficio 1
+▁catálogo 1
+▁coincide 1
+▁confederación 1
+▁consejero 1
+▁crustáceos 1
+▁cárcel 1
+▁emisora 1
+▁espiritual 1
+▁hernández 1
+▁ilustre 1
+▁lingüística 1
+▁lógica 1
+▁mampostería 1
+▁masculino 1
+▁mecánica 1
+▁múnich 1
+▁ocurrir 1
+▁partícula 1
+▁profundidad 1
+▁reconstrui 1
+▁sospecha 1
+▁supremo 1
+▁tráfico 1
+▁university 1
+▁veinticinco 1
+▁afuera 1
+▁comestible 1
+▁infantería 1
+▁secuencia 1
+▁casual 1
+▁educativa 1
+▁educativo 1
+▁exitoso 1
+▁renombr 1
+▁reputación 1
+▁respeto 1
+▁suspiró 1
+▁arabia 1
+▁compleja 1
+▁fiebre 1
+▁simpson 1
+▁aparato 1
+▁youtube 1
+▁pastoral 1
+▁singular 1
+▁himno 1
+▁moderada 1
+▁pidiendo 1
+▁pizza 1
+▁vidrio 1
+▁soporte 1
+▁funeral 1
+▁botella 1
+▁delicioso 1
+▁contenía 1
+▁alonso 1
+▁leonor 1
+▁franja 1
+▁rodeada 1
+▁abrazo 1
+▁persiste 1
+▁tesoro 1
+▁uau 1
+▁terror 1
+▁seguida 1
+▁rescata 1
+▁túnel 1
+▁quieto 1
+▁bautiz 1
+▁comenta 1
+▁fusil 1
+▁incursi 1
+rgenes 1
+▁cerró 1
+▁unirse 1
+▁castro 1
+▁compró 1
+▁exporta 1
+▁caracol 1
+▁caído 1
+▁south 1
+▁percib 1
+▁compiti 1
+▁sagrado 1
+▁analiza 1
+▁claridad 1
+▁florencia 1
+▁turno 1
+▁muelle 1
+▁conversa 1
+▁dicta 1
+▁cartel 1
+▁remata 1
+▁semilla 1
+▁premia 1
+▁ranch 1
+▁baño 1
+▁beat 1
+▁delgado 1
+▁stone 1
+▁lanzar 1
+▁instal 1
+▁qu 1
+▁derram 1
+▁azu 1
+estrati 1
+▁sobrevive 1
+ábamos 1
+▁diputa 1
+▁escrib 1
+▁estrat 1
+▁recorri 1
+▁jardin 1
+▁expedi 1
+▁actúa 1
+▁agujero 1
+▁aluminio 1
+▁balneario 1
+▁compromiso 1
+▁construcciones 1
+▁convencional 1
+▁cortometraje 1
+▁disculpa 1
+▁estricta 1
+▁hipótesis 1
+▁ignacio 1
+▁jurídica 1
+▁legislativo 1
+▁marítimo 1
+▁multiplica 1
+▁mármol 1
+▁pacífica 1
+▁pontificia 1
+▁portuguesa 1
+▁sostiene 1
+▁tripulación 1
+▁acumula 1
+▁agradable 1
+▁bernardo 1
+▁inusual 1
+▁manipula 1
+▁tragedia 1
+▁ciclista 1
+▁contienda 1
+▁especifica 1
+▁eventual 1
+▁proveniente 1
+▁secuela 1
+▁alejó 1
+▁entendido 1
+▁caribe 1
+▁hungría 1
+▁jarocho 1
+▁acuático 1
+▁exilio 1
+▁fortuna 1
+▁estudiantil 1
+▁denuncia 1
+▁elegida 1
+▁negativas 1
+▁andrés 1
+▁examen 1
+▁asiático 1
+▁encanto 1
+▁conciencia 1
+▁maneja 1
+▁cicatri 1
+▁punk 1
+▁propaga 1
+▁tiende 1
+▁lejana 1
+▁derrotó 1
+▁cubrir 1
+▁perdida 1
+▁preveni 1
+▁muerta 1
+▁consola 1
+▁virus 1
+▁aumento 1
+▁creía 1
+▁canarias 1
+▁discográfica 1
+▁gentil 1
+▁delegado 1
+▁define 1
+▁logro 1
+▁mantenido 1
+▁infan 1
+▁invadi 1
+▁contesta 1
+▁acerta 1
+▁envía 1
+▁camil 1
+vivieron 1
+▁veneno 1
+▁correcto 1
+cogido 1
+▁celebr 1
+▁sacerdo 1
+ándole 1
+ō 1
+▁adolescencia 1
+▁afganistán 1
+▁africano 1
+▁arquidiócesis 1
+▁artillería 1
+▁athletic 1
+▁atribuye 1
+▁bermúdez 1
+▁bicicleta 1
+▁caníbales 1
+▁concluye 1
+▁croacia 1
+▁cúpula 1
+▁democracia 1
+▁depredador 1
+▁despedida 1
+▁despedido 1
+▁despierta 1
+▁ecosistema 1
+▁encarcelado 1
+▁enseguida 1
+▁espectador 1
+▁feminista 1
+▁fructifica 1
+▁garantiza 1
+▁global 1
+▁harvard 1
+▁herbácea 1
+▁jesuita 1
+▁magdalena 1
+▁massachusetts 1
+▁mecánico 1
+▁presupuesto 1
+▁romántica 1
+▁secuestro 1
+▁silvestre 1
+▁temblaban 1
+▁transparente 1
+▁violento 1
+▁diploma 1
+▁incendio 1
+▁orientada 1
+▁proteínas 1
+▁taylor 1
+▁caballería 1
+▁clínica 1
+▁escuche 1
+▁aliento 1
+▁asistir 1
+▁defensiv 1
+▁enfoque 1
+▁fracasó 1
+▁rendimiento 1
+▁peruano 1
+▁sugiere 1
+▁escolta 1
+▁puebla 1
+▁viernes 1
+▁ballet 1
+▁tokio 1
+▁transición 1
+▁atleta 1
+▁grifo 1
+▁pesadilla 1
+▁angola 1
+▁evalu 1
+▁jersey 1
+▁limón 1
+▁oxida 1
+▁toronto 1
+▁aparente 1
+▁rodaje 1
+▁cardenal 1
+▁fracasa 1
+▁llave 1
+▁ausente 1
+▁sensación 1
+▁duelo 1
+▁inverna 1
+▁excita 1
+▁sequía 1
+▁preparatoria 1
+▁caracas 1
+▁añadió 1
+▁asume 1
+▁paralelo 1
+▁suelto 1
+▁confesa 1
+▁sexta 1
+▁crecido 1
+▁mental 1
+▁carbon 1
+▁simon 1
+▁revisa 1
+▁transform 1
+▁reaccion 1
+▁tracto 1
+plazamiento 1
+sorte 1
+▁posesi 1
+▁east 1
+▁acompañ 1
+▁denomin 1
+▁actualiza 1
+siguiente 1
+▁subtropical 1
+▁perfecta 1
+▁acantilad 1
+▁almirante 1
+▁aventurero 1
+▁birmania 1
+▁caserío 1
+▁cocoteros 1
+▁consecutivo 1
+▁continuo 1
+▁custodia 1
+▁diminut 1
+▁discurso 1
+▁discuti 1
+▁efectiva 1
+▁egipcio 1
+▁elizabeth 1
+▁episcopal 1
+▁grammy 1
+▁guadalajara 1
+▁guardián 1
+▁hormigón 1
+▁ignora 1
+▁institucional 1
+▁invertebrados 1
+▁mandíbula 1
+▁mozambique 1
+▁ocurra 1
+▁proyecta 1
+▁receptor 1
+▁repercusi 1
+▁simultáneamente 1
+▁voluptuos 1
+▁volúmenes 1
+▁aprendizaje 1
+▁astrónomo 1
+▁excepcional 1
+▁expansión 1
+▁exótica 1
+▁gratuito 1
+▁oculto 1
+▁álvarez 1
+▁francisca 1
+▁margarita 1
+▁otomano 1
+▁recluta 1
+▁geografía 1
+▁paciencia 1
+▁pascua 1
+▁pendiente 1
+▁compagin 1
+▁consuelo 1
+▁obedec 1
+▁rodrigo 1
+▁apartamento 1
+▁asteroid 1
+▁textura 1
+▁peruana 1
+▁street 1
+▁carbón 1
+▁destruido 1
+▁produjer 1
+▁concurr 1
+▁blue 1
+▁ortiz 1
+▁monarca 1
+▁niebla 1
+▁piratas 1
+▁provenía 1
+▁ciudadanía 1
+▁reforz 1
+▁eligió 1
+▁supervivencia 1
+▁consulta 1
+▁sacud 1
+▁mantenimiento 1
+▁desnuda 1
+▁simula 1
+▁dispuesta 1
+▁redondeada 1
+▁polaco 1
+▁cascada 1
+▁vencedor 1
+scender 1
+▁cresta 1
+rtesanía 1
+▁detecta 1
+▁socorr 1
+▁patata 1
+▁díg 1
+▁tinta 1
+▁harina 1
+▁convence 1
+▁ocean 1
+▁turco 1
+▁anciano 1
+▁comuni 1
+▁sopl 1
+▁gesto 1
+presiones 1
+conduct 1
+ratón 1
+aúl 1
+plicó 1
+tividad 1
+tendida 1
+ísimo 1
+lítica 1
+▁emerge 1
+magnetismo 1
+▁adyacente 1
+▁anarquista 1
+▁andaluza 1
+▁arqueológica 1
+▁bacteria 1
+▁bulgaria 1
+▁candidatura 1
+▁cinematográfico 1
+▁cirugía 1
+▁compatible 1
+▁composiciones 1
+▁cooperativ 1
+▁cristóbal 1
+▁diagnóstico 1
+▁diálogo 1
+▁doctrina 1
+▁embarazo 1
+▁entertainment 1
+▁entretenimiento 1
+▁escandal 1
+▁ferroviario 1
+▁gabinete 1
+▁guadalupe 1
+▁intensidad 1
+▁locutor 1
+▁mallorca 1
+▁masculina 1
+▁mongolia 1
+▁muñoz 1
+▁nápoles 1
+▁orgánica 1
+▁pakistán 1
+▁parámetro 1
+▁prosiguió 1
+▁quiméric 1
+▁químico 1
+▁ramírez 1
+▁recomienda 1
+▁reproduce 1
+▁republicana 1
+▁sacrificio 1
+▁trujillo 1
+▁valparaíso 1
+▁algoritmo 1
+▁asturias 1
+▁bilbao 1
+▁cataloga 1
+▁charlie 1
+▁coalición 1
+▁dramática 1
+▁glaciar 1
+▁napoleón 1
+▁paviment 1
+▁sagrada 1
+▁sindicato 1
+▁veintidós 1
+▁víctor 1
+pareciera 1
+▁aeronave 1
+▁alcohol 1
+▁curiosidad 1
+▁destruye 1
+▁frustr 1
+▁impacien 1
+▁rompió 1
+▁censura 1
+▁mauricio 1
+▁ídolo 1
+▁ecuatorial 1
+▁flamenco 1
+▁concreto 1
+▁evacua 1
+▁geográfico 1
+▁ideología 1
+▁certamen 1
+▁procesión 1
+▁situaciones 1
+▁instancia 1
+▁permiso 1
+▁saudita 1
+▁graduado 1
+▁violín 1
+▁emmy 1
+▁herrera 1
+▁transmite 1
+▁gallego 1
+▁josef 1
+▁algodón 1
+▁virrey 1
+hidratado 1
+▁dúo 1
+▁paquete 1
+▁beach 1
+▁espinas 1
+▁trastorno 1
+▁nevada 1
+▁asalto 1
+▁engancha 1
+▁genial 1
+▁tambor 1
+▁precede 1
+▁abrigo 1
+▁criminal 1
+▁dorsal 1
+ulenta 1
+▁eterna 1
+▁subleva 1
+▁wolf 1
+▁conduce 1
+▁heladas 1
+▁hostil 1
+▁racional 1
+▁silva 1
+▁retiro 1
+▁congelado 1
+▁pedido 1
+▁ducado 1
+▁buey 1
+▁mueva 1
+▁icono 1
+▁impidi 1
+monio 1
+▁balear 1
+ávido 1
+▁acoge 1
+▁apodad 1
+óxido 1
+▁mueve 1
+tensiones 1
+▁llov 1
+rezca 1
+▁nazi 1
+traída 1
+▁parezc 1
+▁hispano 1
+▁aspir 1
+▁ligera 1
+▁hubie 1
+▁ocasiona 1
+▁aceitun 1
+▁adecuado 1
+▁agricultor 1
+▁alfonso 1
+▁audiovisual 1
+▁basílica 1
+▁claustro 1
+▁cuestiones 1
+▁cálculo 1
+▁especímenes 1
+▁etiopía 1
+▁guayana 1
+▁hermafrodita 1
+▁improvisa 1
+▁largometraje 1
+▁negrita 1
+▁occidente 1
+▁ofensivo 1
+▁olimpia 1
+▁ortodoxa 1
+▁peculiar 1
+▁perseguido 1
+▁proyección 1
+▁sensibilidad 1
+▁septentrional 1
+▁seudónimo 1
+▁subterránea 1
+▁superviviente 1
+▁sánchez 1
+▁terremoto 1
+▁trágica 1
+▁tuviera 1
+▁almohada 1
+▁basilio 1
+▁channel 1
+▁espectro 1
+▁exclam 1
+▁gregorio 1
+▁mercenario 1
+▁periódica 1
+▁relevante 1
+▁suprema 1
+▁suárez 1
+▁gratuita 1
+▁incierto 1
+▁sintetiza 1
+tíbulo 1
+▁progreso 1
+▁sólido 1
+▁conjunta 1
+▁acarici 1
+▁lisboa 1
+▁mosquito 1
+▁umbral 1
+▁ánimo 1
+▁armenio 1
+▁incendia 1
+▁emilio 1
+▁gusanos 1
+▁modesto 1
+▁festividad 1
+▁decisiones 1
+▁pianista 1
+▁racing 1
+▁galicia 1
+▁alquiler 1
+▁premier 1
+▁asunción 1
+▁anderson 1
+▁fluvial 1
+▁óscar 1
+▁armenia 1
+fección 1
+▁erigi 1
+▁barrera 1
+▁dorso 1
+▁good 1
+▁señas 1
+▁nepal 1
+▁tostada 1
+▁minería 1
+▁cuerno 1
+▁ríe 1
+▁molino 1
+▁duele 1
+▁filial 1
+ectividad 1
+▁racimo 1
+▁codifica 1
+script 1
+▁prosper 1
+▁consultor 1
+▁traidor 1
+▁servía 1
+▁vapor 1
+▁filtr 1
+▁ferm 1
+▁drog 1
+▁besarl 1
+cogí 1
+axila 1
+▁despid 1
+▁bald 1
+▁extrema 1
+▁cansad 1
+▁impact 1
+▁president 1
+▁castella 1
+▁estremec 1
+revolucionaria 1
+ä 1
+▁academy 1
+▁arzobispo 1
+▁atlántico 1
+▁cariño 1
+▁clasificó 1
+▁coincidi 1
+▁complutense 1
+▁corresponsal 1
+▁desprecia 1
+▁diecisiete 1
+▁directivo 1
+▁discurre 1
+▁discusión 1
+▁eficaz 1
+▁eficiente 1
+▁engaña 1
+▁espadaña 1
+▁evangelio 1
+▁franquicia 1
+▁gimnasia 1
+▁guarnición 1
+▁helicóptero 1
+▁inconsciente 1
+▁inmóvil 1
+▁international 1
+▁jerarquía 1
+▁jurisdicción 1
+▁legislativa 1
+▁lóbulo 1
+▁meridional 1
+▁metálico 1
+▁monarquía 1
+▁míchigan 1
+▁negruzca 1
+▁orquídea 1
+▁peregrino 1
+▁portavoz 1
+▁pretendía 1
+▁restringido 1
+▁resuelto 1
+▁revolote 1
+▁revolucionario 1
+▁sacramento 1
+▁sucesiva 1
+▁sábana 1
+▁televisivo 1
+▁vestimenta 1
+▁world 1
+▁zimbabue 1
+▁éxito 1
+â 1
+▁asemeja 1
+▁averigua 1
+▁bombarde 1
+▁boxeo 1
+▁desciende 1
+▁fecunda 1
+▁interfaz 1
+▁maravilla 1
+▁nómada 1
+▁unánime 1
+▁desigual 1
+▁erecto 1
+▁explosión 1
+▁mensual 1
+▁primitivo 1
+▁azteca 1
+▁cobertura 1
+▁ilegal 1
+▁montañoso 1
+▁topografía 1
+▁diamante 1
+▁dispares 1
+▁mártir 1
+▁christian 1
+▁mozart 1
+▁villaverde 1
+▁demasiada 1
+▁glabra 1
+▁vuelvo 1
+▁carnaval 1
+▁diversión 1
+▁sonríe 1
+▁campesina 1
+▁inversión 1
+▁vacío 1
+▁zamora 1
+▁flecha 1
+▁previo 1
+▁permitía 1
+▁progresa 1
+▁utah 1
+▁rechazo 1
+▁estambul 1
+▁iván 1
+▁remedio 1
+▁restantes 1
+▁contento 1
+▁sensible 1
+▁arquero 1
+▁especula 1
+▁vocación 1
+▁caduca 1
+▁criterio 1
+▁sky 1
+▁plomo 1
+▁detiene 1
+▁oliver 1
+▁escamas 1
+▁teórica 1
+▁cantón 1
+▁rubio 1
+▁fallecid 1
+▁deterior 1
+grima 1
+▁moderado 1
+▁alicante 1
+▁gallard 1
+▁calambr 1
+▁rocoso 1
+▁tronco 1
+▁sienta 1
+▁disminuye 1
+▁contradic 1
+ícola 1
+▁cintur 1
+▁lejano 1
+▁pastel 1
+▁jacob 1
+uestos 1
+▁francesc 1
+▁corral 1
+viene 1
+triz 1
+▁vagab 1
+▁mirándo 1
+illetes 1
+▁recurso 1
+▁relativ 1
+▁suplic 1
+▁aprend 1
+▁institu 1
+▁baviera 1
+imilitud 1
+lanceoladas 1
+▁amazonas 1
+▁aparcamiento 1
+▁atmósfera 1
+▁autodidacta 1
+▁birmingham 1
+▁blanquecino 1
+▁cabaña 1
+▁cálida 1
+▁cólera 1
+▁decisiva 1
+▁democrático 1
+▁departamental 1
+▁desaparición 1
+▁diccionario 1
+▁dinosaurio 1
+▁directiva 1
+▁disfrace 1
+▁disolvió 1
+▁electricidad 1
+▁enciclopedia 1
+▁equilibrio 1
+▁escándalo 1
+▁estupendo 1
+▁estúpido 1
+▁explosivo 1
+▁extenso 1
+▁extraordinaria 1
+▁extraordinario 1
+▁fascinado 1
+▁filósofo 1
+▁franquista 1
+▁gimnasio 1
+▁hidrógeno 1
+▁historieta 1
+▁illinois 1
+▁increíble 1
+▁independentista 1
+▁lawrence 1
+▁matemático 1
+▁mausoleo 1
+▁membrana 1
+▁musulmanes 1
+▁méxico 1
+▁network 1
+▁ofensiva 1
+▁olvide 1
+▁oscurec 1
+▁pensilvania 1
+▁periodística 1
+▁persigue 1
+▁petróleo 1
+▁pórtico 1
+▁reemplazó 1
+▁religiones 1
+▁románico 1
+▁subcampeón 1
+▁suministro 1
+▁tardío 1
+▁testimonio 1
+▁urbanización 1
+▁águila 1
+▁dignidad 1
+▁divulga 1
+▁esqueleto 1
+▁fallecimiento 1
+▁infecta 1
+▁longitudinales 1
+▁tarifa 1
+▁tarjeta 1
+▁zambia 1
+▁arizona 1
+▁atalaya 1
+▁competir 1
+▁elefante 1
+▁esparcido 1
+▁ginebra 1
+▁herejes 1
+▁primitiva 1
+▁asiento 1
+▁condecora 1
+▁cuchillo 1
+▁federico 1
+▁hitler 1
+▁plástica 1
+▁prusia 1
+▁veintiocho 1
+▁fractura 1
+▁autonomía 1
+▁mágico 1
+▁campeones 1
+▁difund 1
+▁referente 1
+▁visitó 1
+▁amateur 1
+▁promotor 1
+▁rocío 1
+▁cartagena 1
+▁chipre 1
+▁burdeos 1
+▁quinteto 1
+▁sobrina 1
+▁encerrado 1
+▁mercader 1
+▁pistola 1
+▁masiva 1
+▁ghana 1
+▁terraza 1
+▁retraso 1
+▁encomend 1
+▁mentira 1
+▁comproba 1
+▁ladrones 1
+▁secretaría 1
+▁studio 1
+▁siervo 1
+▁what 1
+▁soluciones 1
+▁descrita 1
+▁perdón 1
+▁colonos 1
+▁erosión 1
+▁cifra 1
+▁eterno 1
+▁andrea 1
+▁requeri 1
+ténga 1
+▁reposo 1
+▁blanqu 1
+▁baraja 1
+▁redonda 1
+▁ligero 1
+▁oprim 1
+▁ondula 1
+▁inspec 1
+▁corregir 1
+patía 1
+ponerse 1
+▁pulp 1
+▁conserv 1
+spiraba 1
+▁oblig 1
+▁minuto 1
+▁tranquil 1
+▁austral 1
+▁represent 1
+cionamiento 1
+▁ejerci 1
+▁matorral 1
+▁bachiller 1
+prefectura 1
+▁acústica 1
+▁albahaca 1
+▁alfabeto 1
+▁ansiedad 1
+▁antártida 1
+▁arequipa 1
+▁atlanta 1
+▁belgrano 1
+▁caracteres 1
+▁christopher 1
+▁cirujan 1
+▁concluyó 1
+▁contribuciones 1
+▁coruña 1
+▁cátedra 1
+▁desafortunada 1
+▁descubierta 1
+▁dieciséis 1
+▁difunto 1
+▁dormitorio 1
+▁douglas 1
+▁elíptica 1
+▁epidemia 1
+▁exámenes 1
+▁galardonado 1
+▁guayaquil 1
+▁hallazgos 1
+▁hidalgo 1
+▁ignición 1
+▁imprescindible 1
+▁insignia 1
+▁jerónimo 1
+▁legendario 1
+▁lluvioso 1
+▁logotipo 1
+▁maquinaria 1
+▁mediterránea 1
+▁mezquita 1
+▁microsoft 1
+▁molecular 1
+▁mountain 1
+▁módulo 1
+▁músculo 1
+▁namibia 1
+▁neoclásico 1
+▁olímpica 1
+▁oviedo 1
+▁pabellón 1
+▁palpitante 1
+▁pamplona 1
+▁pastizales 1
+▁playoff 1
+▁predecesor 1
+▁progresivamente 1
+▁reconstrucción 1
+▁resistir 1
+▁rígido 1
+▁science 1
+▁sobreviviente 1
+▁sufragánea 1
+▁suspirante 1
+▁televisiva 1
+▁tranvía 1
+ë 1
+▁absurdo 1
+▁argelia 1
+▁automático 1
+▁camboya 1
+▁dramático 1
+▁inocente 1
+▁preferencia 1
+▁progresista 1
+▁reconcilia 1
+▁alegoría 1
+▁aprieta 1
+▁calderón 1
+▁heavy 1
+▁octava 1
+▁preferente 1
+▁redacción 1
+▁súper 1
+▁automática 1
+▁exacto 1
+▁imprim 1
+▁lujoso 1
+▁orientado 1
+▁confusa 1
+▁sajonia 1
+▁beatriz 1
+▁neutral 1
+▁misionero 1
+▁montreal 1
+▁uganda 1
+▁capitel 1
+▁charlot 1
+▁complicada 1
+▁emblema 1
+▁gabón 1
+▁pizarra 1
+▁complicado 1
+▁desorden 1
+▁trío 1
+▁pionera 1
+▁surgieron 1
+▁anatom 1
+▁perezos 1
+▁tradiciones 1
+▁aborto 1
+▁vargas 1
+discreto 1
+▁sofía 1
+▁tenerife 1
+▁decreto 1
+▁muscular 1
+▁parodia 1
+▁abandono 1
+▁indecis 1
+▁camisa 1
+▁bodega 1
+▁aclara 1
+▁vigente 1
+▁corazon 1
+▁rally 1
+▁compensa 1
+▁repetí 1
+▁molesta 1
+▁materno 1
+▁policial 1
+▁capitan 1
+ministra 1
+worth 1
+gonzalo 1
+▁exige 1
+▁cautiv 1
+▁tímid 1
+partía 1
+▁entabl 1
+▁numerosa 1
+▁aprovech 1
+stitución 1
+oriental 1
+▁semifinal 1
+▁resolv 1
+queño 1
+▁propuls 1
+▁azerbai 1
+▁virtual 1
+dimensional 1
+▁acondicionado 1
+▁antecedente 1
+▁aparecía 1
+▁asciende 1
+▁autobuses 1
+▁beneficia 1
+▁berkeley 1
+▁brácteas 1
+▁cafetería 1
+▁calendario 1
+▁cilíndrica 1
+▁circunvala 1
+▁cocodrilo 1
+▁columnista 1
+▁continuidad 1
+▁corrupción 1
+▁cotidiana 1
+▁cretácico 1
+▁cuarteto 1
+▁desaparecieron 1
+▁difiere 1
+▁dinámica 1
+▁divertido 1
+▁emocionante 1
+▁encorvado 1
+▁espectacular 1
+▁exclusivo 1
+▁excursión 1
+▁facebook 1
+▁football 1
+▁fotógrafo 1
+▁galaxia 1
+▁gloriosa 1
+▁gradualmente 1
+▁graduarse 1
+▁indiferente 1
+▁japoneses 1
+▁jimmy 1
+▁juguetón 1
+▁lámpara 1
+▁madrileño 1
+▁magistrado 1
+▁minnesota 1
+▁monumental 1
+▁multijugador 1
+▁neerlandesa 1
+▁nobiliario 1
+▁núñez 1
+▁ocurrido 1
+▁oxígeno 1
+▁pictures 1
+▁predilecto 1
+▁pubescente 1
+▁purísima 1
+▁recolección 1
+▁recurrente 1
+▁reemplazada 1
+▁refuerzos 1
+▁reliquia 1
+▁restringida 1
+▁rincón 1
+▁sacrifica 1
+▁sobrevivido 1
+▁sombrío 1
+▁sospechoso 1
+▁stephen 1
+▁submarino 1
+▁surrealista 1
+▁topónimo 1
+▁toulouse 1
+▁tregua 1
+▁tácti 1
+▁tóxico 1
+▁vanguardia 1
+▁windows 1
+▁íntegra 1
+_ 1
+▁apetece 1
+▁deprisa 1
+▁lánguid 1
+▁cuestiona 1
+▁emprendi 1
+▁finanzas 1
+▁herradura 1
+▁oponente 1
+▁recolect 1
+▁caprichos 1
+▁follaje 1
+▁suplente 1
+▁andrew 1
+▁antecesor 1
+▁arthur 1
+▁canónico 1
+▁aqua 1
+▁frederic 1
+▁precisión 1
+▁navarro 1
+▁palidec 1
+▁preserva 1
+▁acomoda 1
+▁afrodita 1
+▁axilares 1
+▁cantautor 1
+▁embalse 1
+▁austin 1
+▁idiota 1
+▁potosí 1
+▁lienzo 1
+▁asumir 1
+▁superliga 1
+▁instrucción 1
+▁consumir 1
+▁radiación 1
+▁perfeccion 1
+▁plantilla 1
+▁comentarista 1
+▁exhibe 1
+▁morisco 1
+▁macizo 1
+▁philip 1
+▁triple 1
+▁barquero 1
+▁pasaporte 1
+iggins 1
+▁frijol 1
+▁santísim 1
+▁degrad 1
+▁vatican 1
+enclavad 1
+▁artesanal 1
+sgraciadamente 1
+▁intervi 1
+▁florenc 1
+▁exigi 1
+liciosa 1
+▁imagin 1
+corporación 1
+luís 1
+▁albacete 1
+▁clandestin 1
+▁clausura 1
+▁excursion 1
+▁perseguía 1
+apareciendo 1
+tífice 1
+▁acuerdas 1
+▁adecuada 1
+▁alfombra 1
+▁alguacil 1
+▁ametralladora 1
+▁amurallado 1
+▁aromática 1
+▁arquitectónica 1
+▁autobús 1
+▁broadway 1
+▁cilindro 1
+▁cofradía 1
+▁comisionado 1
+▁compañera 1
+▁conceptual 1
+▁coordenadas 1
+▁croquetas 1
+▁cuadrangular 1
+▁cáliz 1
+▁danubio 1
+▁diecinueve 1
+▁eclesiástico 1
+▁economista 1
+▁efectivo 1
+▁efectu 1
+▁ejecutiva 1
+▁evangelista 1
+▁farmacia 1
+▁ferroviaria 1
+▁ficticio 1
+▁filosófico 1
+ā 1
+ø 1
+š 1
+ū 1
+æ 1
+č 1
+ô 1
+е 1
+ê 1
+ʻ 1
+ð 1
+đ 1
+ş 1
+ý 1
+ă 1
+ė 1
+ī 1
+ł 1
+ř 1
+ː 1
+ś 1
+ゴ 1
+ã 1
+ё 1
+ß 1
+ć 1
+ï 1
+ì 1
+à 1
+þ 1
+ə 1
+î 1
+û 1
+ě 1
+ı 1
+ń 1
+œ 1
+ž 1
+ț 1
+ʿ 1
+ϙ 1
+ה 1
+נ 1
+ة 1
+ي 1
+ṃ 1
+の 1
+ミ 1
+ム 1
+箱 1
+鮓 1
+鮨 1
+ò 1
+し 1
+ñ 1
+י 1
+כ 1
+ש 1
+ب 1
+ذ 1
+ه 1
+消 1
diff --git a/configs/es-en/src_unigram6000/spm_unigram_es.vocab b/configs/es-en/src_unigram6000/spm_unigram_es.vocab
new file mode 100644
index 0000000000000000000000000000000000000000..930d07dc4a142e96ac4d20367f58fe8bd23b14e1
--- /dev/null
+++ b/configs/es-en/src_unigram6000/spm_unigram_es.vocab
@@ -0,0 +1,6000 @@
+ 0
+ 0
+ 0
+ 0
+s -3.05969
+▁de -3.11604
+▁la -3.47645
+▁en -3.75933
+▁el -3.83733
+▁y -4.05137
+▁a -4.06108
+▁se -4.29875
+▁es -4.51904
+n -4.54227
+▁que -4.61802
+▁los -4.64806
+▁su -4.65501
+r -4.74072
+▁un -4.81993
+▁del -4.8471
+▁con -4.85745
+▁ -4.86631
+▁no -4.90398
+▁por -4.93835
+▁una -5.01475
+a -5.03503
+ó -5.05936
+o -5.10599
+e -5.2692
+▁las -5.2874
+▁al -5.29029
+do -5.3067
+▁fue -5.31075
+es -5.4086
+▁como -5.51533
+▁para -5.53445
+y -5.57324
+l -5.62689
+▁ha -5.6724
+da -5.71108
+▁lo -5.75824
+▁más -5.7862
+mente -5.7954
+▁me -5.92168
+▁son -6.0009
+▁o -6.01499
+▁esta -6.01661
+▁sus -6.01987
+c -6.08332
+▁le -6.10865
+▁este -6.16131
+ta -6.18545
+ción -6.18762
+t -6.19236
+▁está -6.22439
+u -6.24624
+▁dos -6.31035
+▁mu -6.31661
+ca -6.33942
+la -6.35084
+ba -6.35569
+te -6.36931
+▁tiene -6.39109
+▁re -6.40181
+ra -6.42646
+to -6.44457
+▁sin -6.46127
+▁mi -6.48705
+▁te -6.49106
+ron -6.5043
+▁era -6.50592
+na -6.50908
+▁puede -6.52548
+d -6.53446
+ndo -6.5346
+ti -6.54985
+▁si -6.55471
+▁pero -6.55696
+g -6.63845
+i -6.64275
+dos -6.65663
+ado -6.68224
+co -6.68518
+▁hay -6.70089
+le -6.7101
+os -6.72642
+▁ser -6.73516
+ma -6.73745
+ar -6.74704
+▁entre -6.77
+▁e -6.7906
+ro -6.80687
+▁encuentra -6.80738
+▁esto -6.81791
+p -6.82913
+▁durant -6.83852
+f -6.87234
+▁sobre -6.87649
+lo -6.88037
+ciones -6.88585
+▁parte -6.88809
+go -6.89967
+de -6.9074
+▁in -6.91535
+me -6.91975
+no -6.93821
+v -6.94949
+das -6.9524
+ce -6.95242
+▁todo -6.95342
+re -6.95401
+▁tres -6.95431
+í -6.98678
+á -6.98903
+mos -6.99453
+▁ciudad -6.99761
+▁ca -6.99869
+▁forma -7.02844
+▁di -7.03511
+ri -7.03555
+▁desde -7.03769
+▁fueron -7.0407
+▁he -7.04136
+ía -7.0521
+ga -7.05226
+se -7.06896
+er -7.07294
+li -7.07685
+as -7.09102
+m -7.09698
+▁actualmente -7.11728
+▁años -7.11948
+an -7.13154
+▁gran -7.13341
+z -7.13392
+▁ni -7.13642
+▁cuando -7.13757
+▁además -7.1573
+▁ya -7.17354
+nte -7.17361
+▁primera -7.18342
+▁nombre -7.1897
+di -7.19659
+▁año -7.20088
+▁sido -7.20539
+ada -7.21087
+▁uno -7.21499
+en -7.21587
+rse -7.21834
+▁bien -7.22521
+mi -7.22595
+é -7.22908
+▁tiempo -7.23446
+▁ma -7.24602
+▁mismo -7.25444
+tuvo -7.26369
+mo -7.27458
+▁yo -7.2757
+▁ver -7.2838
+▁san -7.28405
+▁co -7.29222
+al -7.2928
+▁embargo -7.29471
+cia -7.3035
+▁vida -7.30384
+▁final -7.30788
+▁todos -7.31857
+▁han -7.3189
+▁mar -7.33394
+ón -7.34191
+▁cada -7.34779
+▁da -7.35345
+idad -7.35492
+va -7.3563
+ja -7.37588
+za -7.3786
+▁ese -7.38482
+ci -7.3966
+tu -7.40381
+▁tu -7.42134
+pa -7.42635
+res -7.4266
+ría -7.42975
+▁primer -7.44086
+b -7.44144
+ne -7.44472
+▁pi -7.44485
+po -7.45328
+donde -7.46357
+▁general -7.46413
+▁universidad -7.46641
+▁va -7.46738
+▁ella -7.46889
+▁hace -7.49763
+▁bo -7.50119
+les -7.50245
+▁otros -7.51216
+▁ba -7.51758
+▁nada -7.52282
+▁ro -7.52473
+ista -7.52882
+bo -7.52968
+▁estos -7.53123
+▁vi -7.54112
+▁mayor -7.54605
+▁entonces -7.54623
+qué -7.54811
+hasta -7.54991
+ve -7.55152
+h -7.55912
+▁historia -7.55931
+ria -7.56358
+▁están -7.56717
+ban -7.56722
+dor -7.57007
+▁casa -7.57716
+▁cuatro -7.57831
+k -7.58297
+▁luego -7.58408
+▁color -7.58537
+▁así -7.58567
+je -7.58616
+▁eso -7.58808
+án -7.58928
+▁mejor -7.59107
+jo -7.60419
+▁varios -7.6116
+▁i -7.61514
+gu -7.61731
+▁había -7.62143
+▁vez -7.62202
+▁estaba -7.62556
+▁nueva -7.62839
+gi -7.63466
+▁pre -7.63783
+pe -7.64257
+▁familia -7.64665
+▁tras -7.65105
+bi -7.65428
+▁ta -7.65561
+▁siempre -7.65603
+que -7.65692
+▁nacional -7.65757
+▁mundo -7.65979
+▁solo -7.65997
+▁sa -7.66115
+x -7.66415
+▁pa -7.68585
+is -7.68703
+▁día -7.68777
+▁vamos -7.68885
+▁mo -7.69412
+▁pe -7.69607
+▁algunos -7.70116
+nos -7.70251
+▁cuenta -7.70924
+▁tarde -7.71583
+▁equipo -7.7161
+fi -7.71829
+▁centro -7.72023
+▁vale -7.72446
+▁tanto -7.72617
+▁li -7.72967
+▁lugar -7.73568
+▁ahora -7.73583
+▁padre -7.73937
+▁tienen -7.74021
+▁pro -7.74435
+pi -7.74923
+▁otro -7.7501
+ver -7.75018
+on -7.75125
+▁poco -7.75203
+ia -7.7588
+sta -7.75911
+qui -7.75968
+▁car -7.75996
+ina -7.76299
+ur -7.7656
+▁ex -7.785
+▁ra -7.78691
+▁hijo -7.78783
+ni -7.78866
+▁guerra -7.78886
+▁estado -7.79059
+or -7.79383
+▁cu -7.79539
+rá -7.79804
+▁hacer -7.7985
+▁allí -7.80258
+▁varias -7.80474
+▁posteriormente -7.80845
+▁muchos -7.81154
+▁nunca -7.81279
+ación -7.81647
+▁bajo -7.81831
+lla -7.81995
+tas -7.8251
+▁antes -7.82668
+▁contra -7.83001
+ng -7.83225
+▁agua -7.83284
+▁nació -7.83759
+▁carrera -7.83981
+zo -7.84264
+▁mucho -7.84483
+▁estas -7.84793
+▁mucha -7.84841
+ando -7.85242
+ismo -7.85633
+▁río -7.85741
+io -7.85881
+▁momento -7.86019
+▁per -7.86202
+ter -7.8631
+▁serie -7.8689
+vo -7.87057
+nd -7.8723
+ch -7.87312
+▁obra -7.87727
+▁p -7.8862
+us -7.89024
+▁tenía -7.89129
+▁grupo -7.89213
+ul -7.89375
+▁c -7.89392
+sólo -7.8941
+mp -7.89426
+▁otra -7.90511
+▁ellos -7.91074
+ero -7.91276
+▁partido -7.91785
+in -7.91916
+ida -7.92078
+▁club -7.92085
+▁ve -7.92104
+▁ti -7.92248
+ido -7.93071
+ita -7.93364
+▁trabajo -7.93484
+ll -7.94398
+ck -7.94546
+▁álbum -7.94752
+▁escuela -7.9503
+ica -7.951
+ú -7.95432
+▁película -7.95747
+w -7.97474
+▁veces -7.9759
+▁otras -7.97685
+▁sur -7.97836
+▁todas -7.97843
+▁trata -7.98027
+▁niña -7.98138
+▁principal -7.98224
+be -7.98267
+▁presenta -7.98353
+▁na -7.98439
+▁ri -7.98464
+man -7.98495
+▁segunda -7.99002
+▁ven -7.9908
+▁nuevo -7.99647
+▁hecho -7.99654
+▁canción -7.99719
+tos -7.99816
+▁cinco -7.99875
+ados -8.00399
+ge -8.00501
+▁españa -8.00754
+era -8.00937
+ntes -8.01159
+▁banda -8.01247
+▁noche -8.01682
+▁algunas -8.01693
+▁ho -8.01724
+▁tan -8.01733
+tó -8.01849
+▁sistema -8.01939
+▁sabe -8.02334
+▁eran -8.02343
+▁ambos -8.03153
+▁debe -8.03188
+▁música -8.03441
+unque -8.03565
+▁zona -8.03737
+▁hoy -8.03839
+▁sol -8.03911
+▁iglesia -8.04048
+▁estudió -8.04965
+▁parece -8.05584
+▁tra -8.05686
+si -8.05957
+▁país -8.06361
+tra -8.06429
+▁estudios -8.07012
+cio -8.07114
+ir -8.07437
+▁toda -8.07503
+nes -8.07798
+▁lleva -8.08013
+du -8.08758
+il -8.09072
+tivo -8.09313
+▁man -8.09421
+▁esa -8.09677
+▁sede -8.10269
+▁algo -8.10383
+▁so -8.10683
+▁siguiente -8.1133
+gún -8.11559
+▁importante -8.1163
+▁barco -8.11763
+eron -8.11926
+▁favor -8.11938
+▁norte -8.12001
+▁porque -8.12112
+▁fu -8.12173
+▁grandes -8.12747
+▁especial -8.13061
+aba -8.13141
+▁hu -8.13212
+ble -8.13518
+▁pu -8.14186
+rio -8.14238
+▁ce -8.14368
+cha -8.14448
+▁ja -8.14453
+▁hizo -8.14662
+▁b -8.14917
+▁diferentes -8.1497
+nt -8.14975
+ente -8.15291
+▁casi -8.15623
+▁pueblo -8.15655
+▁temporada -8.15655
+▁mal -8.15995
+ros -8.16369
+▁madre -8.16716
+▁fi -8.16879
+▁tener -8.17191
+▁ar -8.17215
+▁provincia -8.17519
+▁estados -8.1789
+▁capital -8.18077
+▁ojos -8.18098
+▁ir -8.18263
+ico -8.1853
+ste -8.18623
+▁hijos -8.18965
+▁dentro -8.19182
+▁pasa -8.19204
+▁amor -8.19293
+▁muerte -8.19845
+▁aquí -8.20228
+che -8.20417
+cho -8.2048
+▁tal -8.20487
+▁quiero -8.20578
+▁quien -8.206
+▁unidos -8.2093
+▁deja -8.2111
+▁orden -8.21298
+▁largo -8.21379
+▁junto -8.21753
+▁tierra -8.22149
+el -8.22206
+▁obras -8.22339
+▁edificio -8.22371
+▁política -8.22373
+▁par -8.22645
+cos -8.22977
+▁gusta -8.23529
+▁último -8.23675
+▁cabeza -8.24233
+dores -8.24878
+▁segundo -8.25344
+▁hija -8.25389
+▁igual -8.2541
+las -8.25489
+ones -8.25534
+▁menos -8.26254
+▁especie -8.26428
+▁hombre -8.2643
+▁principalmente -8.265
+son -8.26558
+tro -8.2675
+▁bu -8.26781
+▁seis -8.27436
+▁fe -8.27617
+▁juan -8.27871
+ya -8.27883
+▁baja -8.28149
+▁pan -8.28513
+▁tengo -8.2887
+▁siendo -8.28958
+iendo -8.29238
+▁misma -8.29294
+▁habita -8.29547
+▁gobierno -8.29571
+▁the -8.30187
+▁to -8.30375
+▁suele -8.30813
+▁comenzó -8.31158
+▁medio -8.31295
+▁internacional -8.31485
+▁derecho -8.3156
+▁posee -8.31563
+▁cor -8.31719
+men -8.31787
+▁conocido -8.31837
+▁isla -8.3196
+▁queda -8.31999
+▁pues -8.32037
+▁jo -8.32298
+▁existen -8.32373
+▁sí -8.32513
+▁municipio -8.32765
+▁disco -8.328
+▁planta -8.32833
+▁miembro -8.33028
+▁mira -8.33056
+▁central -8.33397
+▁juego -8.33456
+tar -8.33459
+▁recibió -8.33586
+ín -8.33613
+▁or -8.33928
+▁fuerte -8.34034
+mb -8.34246
+ito -8.34276
+tor -8.3443
+▁cha -8.34488
+▁do -8.34526
+▁ne -8.34763
+sión -8.34842
+remos -8.34956
+car -8.34998
+▁cal -8.35107
+▁original -8.35172
+▁fútbol -8.35229
+▁sal -8.35323
+▁presidente -8.35685
+um -8.35828
+▁llama -8.36381
+▁alta -8.36429
+▁título -8.36485
+▁puedo -8.36488
+▁sigue -8.36569
+▁bueno -8.36924
+fe -8.37009
+▁región -8.37331
+▁considera -8.37659
+ven -8.37662
+▁cuerpo -8.37756
+ndi -8.37957
+▁ga -8.38316
+▁tipo -8.38382
+▁be -8.38417
+▁negro -8.38613
+▁señor -8.39396
+▁dice -8.39525
+ton -8.39638
+▁po -8.39725
+▁estilo -8.39927
+tado -8.40056
+▁natural -8.40314
+▁chole -8.40387
+th -8.40401
+encia -8.40554
+▁fin -8.40965
+▁decir -8.41364
+▁estar -8.41385
+▁representa -8.41884
+endo -8.41905
+miento -8.42043
+▁blanco -8.42114
+▁costa -8.42335
+adas -8.4241
+osa -8.42482
+▁gu -8.42552
+▁programa -8.42668
+ig -8.42697
+▁grande -8.42756
+▁ru -8.42941
+▁número -8.43009
+entras -8.43023
+▁gente -8.43167
+mento -8.43342
+▁media -8.43638
+▁caso -8.43734
+▁realiza -8.43854
+ían -8.44264
+▁verdad -8.44306
+▁libro -8.4431
+▁quiere -8.44383
+▁arte -8.44416
+▁ad -8.44492
+▁poder -8.44583
+▁usa -8.44781
+▁argentina -8.44823
+▁flores -8.44969
+▁corre -8.44982
+▁dar -8.45153
+▁can -8.45155
+vis -8.45665
+▁maría -8.45738
+▁buena -8.45762
+ario -8.45913
+▁fa -8.46111
+▁proyecto -8.46205
+▁puerta -8.46206
+▁jug -8.46242
+▁g -8.46401
+aron -8.46412
+▁cargo -8.46881
+st -8.46942
+▁les -8.47189
+nia -8.47222
+▁santa -8.47351
+▁mayoría -8.47384
+▁camino -8.47608
+ales -8.47667
+ct -8.47838
+▁director -8.4789
+▁go -8.48037
+un -8.48225
+et -8.4843
+tiva -8.48436
+ke -8.48646
+▁conoce -8.48651
+ieron -8.48707
+▁origen -8.49029
+rra -8.49439
+ncia -8.49477
+▁única -8.49508
+▁único -8.49508
+▁muestra -8.49509
+▁mujer -8.49521
+▁canciones -8.4962
+▁llega -8.50042
+▁bar -8.50278
+j -8.50353
+mer -8.50403
+▁gal -8.50405
+▁comp -8.50443
+▁acaba -8.50624
+▁hi -8.50651
+▁toma -8.50782
+▁lu -8.50834
+▁honor -8.51449
+iones -8.51524
+it -8.51674
+▁trabaj -8.51721
+▁ju -8.51875
+ue -8.5204
+mu -8.52104
+▁ka -8.52164
+▁teatro -8.52429
+▁aún -8.52466
+▁bi -8.52537
+▁bri -8.52922
+▁real -8.53041
+ing -8.53161
+pu -8.53329
+▁wa -8.53349
+▁du -8.53378
+▁madrid -8.5342
+▁producción -8.53463
+ante -8.53491
+▁chile -8.53805
+▁público -8.5392
+emos -8.54278
+bra -8.54402
+▁manera -8.54483
+▁ello -8.5464
+▁ejemplo -8.54928
+▁mundial -8.54929
+▁cualquier -8.55
+▁campo -8.55339
+▁ayuda -8.55367
+▁hermano -8.55393
+▁cambio -8.55435
+▁tenemos -8.55457
+ho -8.55668
+eta -8.55861
+ana -8.55984
+▁hermana -8.56464
+▁museo -8.56498
+▁pat -8.56626
+▁personas -8.56804
+▁última -8.56974
+▁estación -8.57387
+▁destaca -8.57437
+▁nadie -8.57493
+▁cerca -8.57501
+▁rec -8.57951
+▁europa -8.58031
+▁división -8.58044
+▁dicho -8.58139
+▁principales -8.58346
+istas -8.58362
+▁mano -8.58425
+▁fo -8.58431
+▁encuentran -8.58461
+▁cri -8.58468
+▁cer -8.58506
+▁cine -8.58624
+▁cosas -8.5871
+ie -8.58753
+▁ten -8.58929
+▁eres -8.58964
+▁base -8.59009
+▁antiguo -8.59063
+▁movimiento -8.59063
+llo -8.59169
+▁gan -8.59467
+ces -8.5947
+▁mon -8.5948
+▁posible -8.59595
+▁juega -8.59605
+▁gracias -8.6016
+▁actualidad -8.60186
+▁días -8.6037
+▁com -8.60465
+gen -8.60504
+▁realizó -8.61196
+▁él -8.61209
+▁sub -8.61274
+ic -8.61537
+▁azul -8.61561
+tura -8.61603
+▁antigua -8.61737
+iz -8.61878
+los -8.62006
+nal -8.6207
+▁militar -8.62131
+▁distrito -8.6228
+▁interior -8.62281
+▁actual -8.62339
+tal -8.62514
+▁haber -8.62631
+▁larga -8.62958
+par -8.6309
+▁k -8.63462
+cer -8.63582
+▁utiliza -8.6367
+▁torre -8.63725
+▁considerado -8.63876
+tuvieron -8.63896
+▁sitio -8.63941
+ger -8.63944
+▁don -8.6417
+▁oficial -8.64258
+bu -8.64307
+▁línea -8.64484
+▁ambas -8.64526
+▁suelo -8.6455
+▁uso -8.64552
+nza -8.64613
+per -8.64637
+▁mor -8.64683
+▁estamos -8.64703
+▁pronto -8.65048
+ut -8.65088
+▁fuera -8.65235
+▁for -8.65289
+ió -8.65522
+▁inter -8.6599
+día -8.66137
+▁primeros -8.66196
+▁radio -8.66203
+▁julia -8.66373
+▁pesa -8.66614
+▁col -8.6671
+▁televisión -8.66741
+▁causa -8.66758
+lle -8.67181
+ran -8.67278
+▁población -8.67316
+▁alto -8.67828
+▁parque -8.67961
+▁v -8.67979
+▁habla -8.68026
+ru -8.68089
+mina -8.68102
+▁d -8.68106
+▁conocida -8.68166
+▁t -8.68168
+ber -8.68219
+cal -8.68268
+▁popular -8.68507
+▁mí -8.68514
+rlo -8.68692
+▁figura -8.69043
+ador -8.69134
+▁china -8.6917
+▁español -8.69399
+ha -8.69704
+▁debido -8.69705
+ad -8.69726
+▁mis -8.69992
+▁total -8.70182
+rme -8.70193
+▁claro -8.70383
+▁vive -8.70418
+tic -8.70773
+▁capítulo -8.70808
+▁pertenece -8.70808
+▁siete -8.70816
+▁gi -8.70901
+▁condado -8.70915
+▁decidi -8.71046
+▁u -8.71366
+▁pasó -8.71385
+▁convirti -8.71403
+▁llegó -8.71403
+▁principio -8.71403
+ka -8.71699
+▁super -8.71934
+▁ocho -8.72416
+▁república -8.72604
+▁proceso -8.72605
+▁barrio -8.72705
+▁dios -8.73007
+▁buen -8.7308
+rre -8.73109
+▁todavía -8.73211
+▁estudio -8.73458
+▁liga -8.73493
+▁ab -8.73522
+▁local -8.73529
+▁estaban -8.73569
+▁aquel -8.7368
+▁cubierta -8.7382
+▁autor -8.73846
+tre -8.74002
+▁falta -8.74032
+ion -8.74113
+▁profesor -8.74164
+▁existe -8.74305
+▁francia -8.74467
+▁quedó -8.7449
+▁actividad -8.74591
+▁sería -8.74685
+▁puesto -8.74695
+▁gra -8.74972
+▁joven -8.75087
+▁celebra -8.75174
+▁nave -8.75682
+▁ante -8.75699
+▁villa -8.75834
+▁musical -8.75843
+▁manos -8.7607
+▁fra -8.76101
+▁st -8.76229
+port -8.76337
+▁llamado -8.7638
+tera -8.76401
+tes -8.768
+▁modelo -8.76927
+▁ocupa -8.76927
+▁particip -8.76992
+▁española -8.77296
+▁meses -8.77376
+▁hora -8.77528
+ras -8.77532
+▁je -8.77548
+▁brasil -8.7756
+▁propia -8.77561
+▁civil -8.77564
+▁edad -8.77677
+tt -8.77737
+▁corta -8.77756
+▁cruz -8.7788
+ner -8.77995
+ología -8.78
+▁niño -8.78033
+hi -8.78095
+▁pie -8.78104
+▁hábitat -8.78196
+▁sencillo -8.78199
+▁libre -8.78227
+den -8.78341
+▁obtuv -8.78394
+tación -8.78797
+▁marca -8.78922
+▁oye -8.79118
+▁rey -8.79212
+▁punto -8.79383
+land -8.79468
+▁f -8.7948
+▁territorio -8.79482
+▁espacio -8.79483
+ándose -8.79506
+▁dio -8.79671
+▁semana -8.79679
+▁primero -8.79703
+▁inclus -8.79764
+▁desarrollo -8.79843
+▁tema -8.79865
+▁verde -8.79948
+▁pequeña -8.80059
+wa -8.80091
+▁relación -8.80137
+▁plan -8.80223
+▁cantante -8.80264
+▁amigos -8.80351
+▁área -8.80744
+▁puerto -8.80785
+▁hojas -8.80791
+▁bastante -8.80844
+cciones -8.80866
+▁mil -8.80952
+▁profesional -8.80968
+▁dura -8.80988
+▁aparece -8.81039
+▁pri -8.8109
+▁valle -8.81144
+rán -8.8129
+▁castillo -8.81449
+▁novela -8.81454
+▁emp -8.8152
+lin -8.8163
+via -8.81986
+tina -8.82057
+▁pon -8.82075
+▁estructura -8.82105
+▁selección -8.82105
+▁john -8.82106
+▁nombrado -8.82269
+▁tri -8.82574
+▁interpreta -8.82647
+▁jugador -8.82701
+▁diferencia -8.82772
+▁ninguna -8.82772
+▁nosotros -8.82772
+▁propio -8.82773
+▁voz -8.82833
+▁libros -8.82874
+ano -8.82935
+▁cr -8.82941
+mar -8.8311
+▁países -8.83113
+ard -8.83139
+ex -8.83142
+▁será -8.83457
+▁social -8.83487
+▁soy -8.836
+▁luz -8.83625
+ud -8.83684
+▁numerosos -8.84119
+▁político -8.84119
+▁común -8.84119
+▁pareja -8.84119
+rro -8.84141
+▁control -8.84165
+▁versión -8.84234
+▁sola -8.84289
+lí -8.84362
+▁tom -8.84365
+▁van -8.84506
+▁vas -8.84708
+▁pasado -8.84734
+▁fiesta -8.84804
+▁dicha -8.84978
+▁papel -8.85009
+▁ci -8.85138
+gar -8.85222
+dad -8.85227
+▁corte -8.85329
+▁incluye -8.8544
+▁voy -8.85788
+▁horas -8.85821
+▁unos -8.85876
+im -8.85892
+iente -8.85905
+til -8.86006
+▁australia -8.86255
+▁india -8.86286
+▁servicio -8.86398
+iza -8.86476
+ine -8.86496
+▁calle -8.86702
+▁partir -8.86769
+▁alguien -8.86868
+▁ocasiones -8.86868
+▁dónde -8.86868
+▁educación -8.86877
+▁viene -8.86958
+▁bro -8.87389
+ura -8.87475
+▁diversos -8.87567
+▁batalla -8.87568
+▁york -8.87578
+imismo -8.87611
+▁menor -8.87823
+▁padres -8.87864
+▁localidad -8.87872
+▁reina -8.88113
+▁trabaja -8.88176
+▁logra -8.88193
+▁mujeres -8.88258
+▁áfrica -8.88271
+▁contiene -8.88271
+▁vuelve -8.88272
+▁cómo -8.88272
+▁torneo -8.88275
+▁diez -8.88344
+▁escrito -8.88404
+▁pla -8.88454
+▁normal -8.88661
+toria -8.88831
+▁ama -8.88945
+▁compañía -8.88981
+▁viaje -8.88982
+▁piedra -8.88993
+▁frente -8.89066
+id -8.89072
+▁fondo -8.892
+ino -8.89248
+am -8.89417
+▁dan -8.89443
+▁nu -8.89495
+▁tradicional -8.89695
+▁diversas -8.89695
+▁visto -8.89747
+▁termina -8.89763
+torio -8.89778
+▁qui -8.8995
+▁im -8.89962
+▁revista -8.901
+▁bosques -8.90142
+▁chi -8.90176
+▁volver -8.90371
+▁colombia -8.90414
+tero -8.90435
+▁espera -8.90438
+▁clima -8.90441
+▁especies -8.90506
+ki -8.90592
+▁nivel -8.90667
+▁mata -8.90689
+▁gira -8.90723
+illa -8.90839
+▁lucha -8.90942
+▁oh -8.90947
+▁cree -8.90961
+▁resto -8.90965
+▁departamento -8.91139
+▁pequeños -8.91384
+▁lista -8.91396
+▁pequeño -8.91627
+▁amigo -8.91632
+ización -8.91648
+▁matrimonio -8.91869
+▁nueve -8.91869
+▁acuerdo -8.92128
+▁busca -8.92205
+lina -8.92212
+▁wi -8.92261
+tino -8.92335
+▁premio -8.92392
+fer -8.92493
+▁pas -8.92508
+▁colegio -8.92604
+▁come -8.92608
+▁hermanos -8.92671
+eño -8.92941
+▁funciona -8.93145
+▁palomares -8.93362
+▁instituto -8.93365
+▁dejó -8.9339
+▁luis -8.9343
+▁empresa -8.93914
+der -8.9399
+▁lengua -8.9404
+▁influencia -8.94091
+▁importantes -8.94138
+▁sala -8.94229
+gue -8.94433
+▁podía -8.94443
+▁desarrolla -8.94836
+▁california -8.94843
+▁práctica -8.94843
+▁elegido -8.94843
+▁ki -8.94854
+dura -8.94991
+ment -8.95112
+▁and -8.9514
+▁oro -8.95509
+dio -8.95516
+▁intenta -8.95579
+ju -8.9571
+▁niños -8.95795
+ed -8.95824
+rri -8.95851
+▁resultado -8.95887
+▁grupos -8.96016
+▁ap -8.96077
+eros -8.96126
+▁juguete -8.96364
+▁permite -8.96365
+sh -8.96832
+▁rosa -8.96947
+▁creo -8.97139
+▁información -8.97202
+▁cual -8.97555
+▁encontrar -8.9767
+vin -8.97727
+ster -8.97743
+aban -8.97762
+▁episodio -8.97908
+▁pública -8.97908
+▁escribi -8.97913
+▁plaza -8.97977
+bles -8.98012
+▁partidos -8.98048
+gan -8.98264
+▁idea -8.98302
+▁consta -8.98356
+▁continua -8.98485
+▁podemos -8.98699
+stra -8.98871
+▁rojo -8.98901
+ji -8.99026
+▁cultura -8.99213
+illo -8.99434
+▁construcción -8.99477
+▁municipal -8.99477
+▁nuestro -8.99477
+▁murió -8.99487
+▁cara -8.99507
+▁zonas -8.99633
+▁temas -8.99661
+▁ref -8.99703
+▁perro -8.9973
+▁modo -8.99904
+▁gana -8.99986
+▁iba -9.00056
+▁film -9.00082
+▁santo -9.00153
+▁observa -9.00271
+▁arriba -9.00274
+▁muerto -9.00277
+▁volvió -9.00282
+if -9.00474
+▁miembros -9.00553
+ega -9.00566
+▁numerosas -9.00718
+▁venga -9.00791
+▁histórico -9.01071
+ngún -9.0126
+▁superior -9.01403
+▁plantas -9.0144
+▁juntos -9.01635
+▁incluyen -9.01802
+▁nuestra -9.01877
+▁parís -9.01887
+▁playa -9.0191
+▁vino -9.01966
+▁paso -9.02006
+▁posición -9.02105
+▁hacen -9.02114
+ent -9.02148
+ela -9.02348
+▁compra -9.02447
+dia -9.02449
+▁persona -9.0253
+▁categoría -9.0269
+▁rep -9.02736
+amente -9.02746
+▁conserva -9.03017
+▁visita -9.0302
+▁toca -9.03201
+▁tor -9.03221
+▁arma -9.03222
+▁tercera -9.03293
+rica -9.03455
+▁academia -9.0351
+▁demasiado -9.0351
+▁escucha -9.0351
+▁carlos -9.03604
+▁vivo -9.03701
+▁restos -9.03852
+▁mala -9.03947
+bla -9.04078
+ática -9.04082
+▁hor -9.04116
+▁actividades -9.04141
+▁alemania -9.04336
+▁memoria -9.04337
+▁siento -9.04339
+▁vivi -9.04382
+ático -9.04384
+▁sali -9.04572
+eña -9.04723
+aciones -9.04728
+▁llev -9.04804
+▁entra -9.04909
+acu -9.04965
+▁produce -9.05064
+▁hospital -9.0517
+▁viejo -9.05171
+dero -9.05263
+▁personaje -9.05628
+▁fre -9.05676
+▁escritor -9.05862
+▁alrededor -9.0601
+▁exterior -9.0601
+▁francisco -9.0601
+▁etapa -9.06012
+▁z -9.06223
+ff -9.0627
+▁italia -9.06344
+▁w -9.06382
+▁aire -9.06504
+▁tamaño -9.06857
+▁tampoco -9.06857
+▁victoria -9.06858
+▁período -9.06858
+▁consejo -9.06861
+▁islas -9.07164
+▁personal -9.07189
+chi -9.07591
+▁trabajar -9.07606
+▁cabo -9.07628
+▁ahí -9.07719
+▁reino -9.07935
+▁equipos -9.08462
+▁gol -9.08486
+▁crea -9.085
+▁fruto -9.08574
+▁doce -9.08703
+▁mari -9.08741
+▁ge -9.0891
+▁ant -9.08925
+▁monta -9.09156
+▁quieres -9.09415
+▁económica -9.09444
+▁puente -9.09522
+▁hombres -9.09549
+▁vía -9.09585
+bilidad -9.10039
+▁aspecto -9.10321
+▁barcelona -9.10321
+▁necesita -9.10321
+▁economía -9.10321
+▁indica -9.10325
+rios -9.10325
+▁cosa -9.10343
+▁tú -9.10377
+▁ke -9.10383
+▁fecha -9.10389
+▁vol -9.10428
+cord -9.10602
+▁salud -9.10667
+▁películas -9.10703
+▁actor -9.10987
+▁anti -9.11022
+▁boca -9.11186
+▁viento -9.11208
+▁escrita -9.1121
+▁significa -9.11286
+for -9.113
+cía -9.11449
+▁juegos -9.11476
+▁últimos -9.11612
+▁inc -9.11643
+▁llamada -9.11724
+▁comercial -9.11822
+▁alguna -9.11949
+▁sociedad -9.12099
+▁cuanto -9.121
+▁val -9.12101
+▁oeste -9.12121
+▁gar -9.1215
+nder -9.12197
+tada -9.12213
+mediante -9.12349
+wi -9.1248
+▁levanta -9.12552
+▁ben -9.12812
+▁bur -9.12961
+▁comienza -9.13
+▁conjunto -9.13
+▁amarillo -9.13
+▁obsta -9.13053
+▁verano -9.13111
+▁sea -9.13118
+▁public -9.13131
+▁buenos -9.13154
+▁fundador -9.13257
+▁luna -9.13314
+▁llegar -9.13433
+▁aquella -9.13454
+lio -9.13506
+▁of -9.1366
+▁vo -9.13682
+▁sta -9.13693
+▁tro -9.13693
+▁industria -9.13903
+▁empieza -9.13909
+▁objetivo -9.13909
+▁cuarto -9.1391
+▁situado -9.1391
+▁canta -9.1402
+▁indio -9.14038
+▁cl -9.14595
+▁formación -9.14714
+udi -9.1474
+▁acceso -9.14826
+▁acepta -9.14827
+▁pedro -9.14827
+▁periodo -9.14828
+▁ricardo -9.14831
+▁inicia -9.14841
+ly -9.14878
+▁ban -9.15279
+del -9.15367
+ez -9.1547
+▁za -9.15497
+▁dr -9.15533
+apareció -9.15752
+▁agricultura -9.15752
+▁dirección -9.15752
+▁independiente -9.15752
+▁japón -9.15752
+▁presente -9.15761
+▁fuerza -9.15793
+bre -9.16027
+▁duda -9.16031
+ivo -9.16077
+▁op -9.1629
+ona -9.16461
+cl -9.1648
+ier -9.16656
+▁occidental -9.16687
+▁usted -9.16687
+▁actriz -9.16687
+▁situada -9.16688
+▁estableci -9.16739
+▁estadounidense -9.16742
+▁entrada -9.16815
+▁publicado -9.1694
+▁eh -9.17248
+▁unión -9.17485
+▁campeonato -9.1763
+▁invierno -9.1763
+▁menudo -9.1763
+▁templo -9.17631
+▁pintura -9.17636
+▁comenza -9.17655
+tró -9.17658
+▁form -9.17683
+▁hubo -9.17787
+▁anterior -9.178
+illas -9.17954
+▁asia -9.18024
+izado -9.18116
+▁grado -9.18196
+grafía -9.18576
+▁difícil -9.18582
+▁empezó -9.18583
+▁golpe -9.18636
+▁críticas -9.18641
+▁lugares -9.18674
+▁salir -9.18695
+▁problemas -9.1906
+▁ob -9.19087
+▁alimenta -9.19337
+▁recibe -9.1951
+inició -9.19544
+▁aeropuerto -9.19544
+▁oscuro -9.19544
+▁prueba -9.19544
+▁londres -9.19545
+▁casado -9.19549
+▁compuesta -9.19557
+▁seguro -9.19559
+▁madera -9.19631
+▁pequeñas -9.19651
+▁marcha -9.19777
+▁paul -9.19926
+▁utilizado -9.19935
+▁ley -9.20121
+▁amplia -9.20184
+▁h -9.20187
+▁realizar -9.20205
+bil -9.20413
+▁texto -9.20518
+▁palacio -9.2053
+▁dedica -9.2061
+▁valencia -9.20634
+▁perú -9.20649
+dra -9.20744
+▁estrella -9.20764
+▁blanca -9.20765
+▁datos -9.2078
+▁m -9.20893
+▁cuidado -9.20901
+▁monte -9.21183
+puso -9.21275
+▁hemo -9.21277
+▁labor -9.21322
+▁mañana -9.21495
+▁santiago -9.21495
+▁referencia -9.21495
+▁corriente -9.21515
+▁tenido -9.21569
+▁sufri -9.21612
+▁ganador -9.22014
+▁integra -9.22213
+▁sp -9.22456
+▁también -9.22485
+▁ofrece -9.22486
+▁defini -9.2249
+▁esposa -9.22491
+▁luc -9.22502
+▁podría -9.22521
+▁cita -9.22662
+▁trans -9.22731
+▁roma -9.22737
+▁tercer -9.2274
+▁clase -9.2284
+cula -9.23048
+▁reyes -9.23253
+▁red -9.23347
+▁corto -9.23388
+▁continuó -9.23485
+▁enfermedad -9.23485
+▁filosofía -9.23485
+▁edición -9.23487
+▁ingres -9.23491
+▁valor -9.2358
+▁auto -9.23728
+ena -9.23901
+▁corona -9.23926
+▁opera -9.23954
+▁ves -9.23958
+▁vista -9.24058
+▁ber -9.2409
+▁campaña -9.24495
+▁terminó -9.24495
+▁logró -9.24496
+▁maestro -9.24496
+▁pobre -9.2452
+▁diario -9.24687
+▁doble -9.24741
+gio -9.24781
+▁pasar -9.24918
+▁caja -9.25094
+▁sale -9.25102
+▁encontrado -9.25154
+▁servicios -9.25184
+▁localiza -9.25276
+lógico -9.25511
+▁variedad -9.25516
+▁situación -9.25516
+▁acerca -9.25539
+▁acción -9.25599
+ko -9.25653
+▁diga -9.25715
+▁diseño -9.25754
+▁roja -9.25865
+▁problema -9.26033
+▁comunidad -9.26073
+▁rica -9.2613
+hu -9.26244
+▁dedic -9.26312
+▁ulises -9.26547
+▁campeón -9.26547
+▁camarote -9.26547
+▁cementerio -9.26547
+▁miedo -9.26592
+▁sueño -9.26834
+▁ciudades -9.2693
+▁similar -9.27113
+▁palabra -9.27126
+▁jugar -9.27257
+▁fla -9.27283
+ión -9.27497
+ería -9.27514
+▁características -9.27561
+▁negra -9.27589
+ley -9.27672
+sen -9.27713
+▁decide -9.27767
+▁completamente -9.27881
+▁vio -9.27921
+▁ger -9.28029
+▁paz -9.28254
+▁ganado -9.2837
+▁aquello -9.2839
+▁canal -9.28412
+▁denomina -9.28503
+▁artes -9.28571
+▁accidente -9.28641
+▁periódico -9.28641
+▁determina -9.28644
+▁alas -9.28711
+▁ciencias -9.28739
+▁sonido -9.28759
+▁trabajos -9.28783
+▁remo -9.28863
+▁considerada -9.28897
+▁che -9.28909
+▁plata -9.28942
+▁halla -9.29124
+▁personajes -9.29189
+▁propiedad -9.29229
+▁medi -9.29299
+▁echa -9.29345
+▁simple -9.29398
+▁tira -9.29554
+▁silencio -9.29666
+gráfico -9.29695
+▁partes -9.29696
+az -9.29821
+▁casó -9.30003
+▁enterrado -9.30037
+ller -9.30067
+val -9.30087
+▁colores -9.30169
+dieron -9.3021
+icio -9.30245
+▁fuente -9.30314
+▁montaña -9.30445
+▁cola -9.3047
+▁elementos -9.30594
+▁disputa -9.3078
+▁jóvenes -9.3078
+▁deportivo -9.30781
+▁quién -9.30781
+▁seguir -9.30783
+▁animales -9.31102
+▁basa -9.31156
+▁pato -9.31279
+▁ana -9.31408
+▁comer -9.31516
+▁artista -9.31587
+▁doctor -9.31624
+▁aguas -9.31638
+▁tomar -9.31643
+▁compañero -9.31867
+▁cámara -9.31867
+▁mezcla -9.31867
+▁secundaria -9.31867
+▁objeto -9.31867
+▁cierto -9.31868
+▁famoso -9.3187
+▁capilla -9.3193
+▁pecado -9.31933
+▁investigación -9.32023
+▁organización -9.32039
+shi -9.32117
+▁distintas -9.32268
+ula -9.32346
+▁dijo -9.32357
+▁ubica -9.32429
+dera -9.32454
+▁ubicado -9.32721
+▁literatura -9.32966
+▁defensa -9.32966
+▁festival -9.32966
+▁adapta -9.32969
+▁poeta -9.33015
+▁motor -9.33129
+▁alma -9.33411
+▁acto -9.33433
+▁gen -9.33436
+▁cuba -9.33443
+▁completa -9.33496
+table -9.33506
+▁relativa -9.33564
+▁estudia -9.33581
+▁mayores -9.33618
+▁pens -9.33657
+ctor -9.33767
+▁mont -9.33809
+▁rod -9.34074
+▁función -9.34085
+uda -9.34186
+▁morir -9.34218
+puesto -9.34343
+ío -9.34488
+▁web -9.34668
+▁caballo -9.35108
+▁fabrica -9.35201
+▁ministro -9.35201
+▁venezuela -9.35201
+▁parecía -9.35201
+▁fachada -9.35202
+▁regresa -9.35202
+▁video -9.35206
+▁ataque -9.3522
+▁abajo -9.3522
+▁gui -9.35252
+▁siglo -9.3526
+▁hogar -9.35264
+▁bru -9.35278
+▁premios -9.35529
+▁destino -9.3553
+op -9.35542
+▁funda -9.35946
+▁hablar -9.36233
+▁crítica -9.36267
+▁semillas -9.36283
+ü -9.36337
+▁extremo -9.36339
+amos -9.36415
+▁dejar -9.36468
+▁escena -9.36503
+▁chica -9.3652
+▁pocos -9.36531
+▁cama -9.36621
+▁casos -9.36679
+▁estadio -9.36716
+▁llevar -9.36743
+▁carga -9.36771
+▁mat -9.36997
+▁fal -9.37342
+▁agr -9.37347
+ler -9.3745
+▁recib -9.37454
+▁uruguay -9.37487
+▁antonio -9.37487
+▁líder -9.37487
+▁veinte -9.37487
+▁relaciones -9.37491
+▁permiti -9.37492
+▁medicina -9.37494
+▁poner -9.37851
+▁debut -9.37949
+▁caracteriza -9.38164
+▁ciento -9.38199
+▁notable -9.38252
+▁delante -9.38289
+▁obliga -9.38349
+uri -9.3845
+▁fan -9.38451
+▁paga -9.38482
+▁atrás -9.38483
+▁reside -9.3863
+▁abierto -9.38649
+▁después -9.38649
+▁experiencia -9.38649
+▁mantiene -9.38649
+▁seguridad -9.38649
+▁poesía -9.3865
+▁dulce -9.38651
+▁versiones -9.38657
+▁residencia -9.38671
+▁natal -9.38736
+▁fundación -9.3886
+▁buenas -9.38875
+at -9.3896
+culo -9.38965
+▁ellas -9.38997
+▁exp -9.39021
+▁aires -9.39243
+tion -9.39515
+▁tipos -9.39555
+▁fuerzas -9.39807
+idos -9.39811
+▁católica -9.39826
+▁clásico -9.39826
+▁menciona -9.39826
+▁necesito -9.39826
+▁presencia -9.39832
+▁policía -9.39835
+▁distintos -9.39852
+▁cel -9.39893
+cip -9.39919
+▁probablemente -9.39922
+▁mercado -9.39978
+▁mayo -9.39978
+▁ventana -9.39982
+▁valeria -9.40116
+▁arena -9.40261
+▁bella -9.4045
+▁productor -9.40674
+▁particular -9.40708
+▁material -9.40717
+▁torn -9.40769
+lica -9.40783
+ré -9.40915
+▁tele -9.40938
+▁arquitectura -9.41016
+▁biblioteca -9.41016
+▁inspira -9.41016
+▁abogado -9.41016
+▁bolivia -9.41016
+▁cabecera -9.41017
+▁terreno -9.41017
+▁sentido -9.41021
+▁frontera -9.41031
+▁junta -9.41043
+▁jefe -9.41063
+▁ruta -9.41092
+ben -9.41174
+▁puntos -9.41208
+▁vaya -9.4125
+▁pos -9.41261
+▁artistas -9.41325
+▁curso -9.41462
+▁bala -9.41643
+▁pol -9.42042
+ell -9.42097
+▁traslad -9.42116
+burgo -9.42162
+▁carácter -9.42221
+▁olvida -9.42221
+▁perdido -9.42224
+▁alemán -9.42225
+▁guarda -9.42226
+▁elimina -9.42229
+▁deriva -9.4226
+▁sirve -9.42302
+▁oriental -9.42331
+▁pal -9.42383
+▁gas -9.42411
+▁vuelo -9.42417
+mba -9.42519
+▁varia -9.42539
+▁post -9.42654
+▁hola -9.42729
+▁permanece -9.42786
+▁motivo -9.42911
+▁copa -9.43147
+▁ko -9.43157
+arse -9.43231
+▁realizado -9.43288
+▁humanos -9.43317
+cel -9.43439
+▁capitán -9.43441
+▁dirigido -9.43441
+▁ninguno -9.43441
+▁complejo -9.43441
+▁contenido -9.43441
+▁feliz -9.43441
+▁haciendo -9.43444
+▁industrial -9.43448
+▁teoría -9.43448
+rico -9.43626
+▁tropicales -9.43682
+▁coche -9.43685
+▁resulta -9.43963
+▁dije -9.4401
+logía -9.4404
+▁enseñ -9.44209
+▁entrar -9.44311
+▁abierta -9.44675
+▁david -9.44675
+▁corazón -9.44675
+▁ambiente -9.44676
+▁abarca -9.44677
+▁importancia -9.44687
+▁razón -9.44707
+▁publica -9.44712
+▁siguientes -9.44758
+▁naturales -9.44771
+▁jugaba -9.448
+▁gris -9.44809
+▁pocas -9.44837
+▁mostr -9.45079
+▁j -9.45311
+▁pone -9.45358
+▁pica -9.45437
+▁em -9.45465
+▁nuevos -9.4565
+▁aun -9.45676
+▁regalo -9.45682
+gre -9.45915
+▁británico -9.45925
+▁física -9.45925
+▁inglaterra -9.45925
+▁superficie -9.45925
+▁treinta -9.45925
+▁siguió -9.45925
+▁virgen -9.45926
+▁alcalde -9.45926
+▁manuel -9.45928
+▁vilma -9.45929
+▁distribuye -9.4593
+▁vivir -9.45932
+▁imperio -9.45942
+▁lago -9.45982
+▁lejos -9.45993
+▁triste -9.46003
+▁coro -9.46133
+▁consiste -9.46165
+▁resultados -9.46206
+mit -9.46247
+▁escond -9.46396
+▁libera -9.46435
+▁fase -9.46439
+ker -9.46692
+iste -9.47052
+▁pinta -9.47166
+▁burbuja -9.47191
+▁encuentro -9.47191
+▁jardín -9.47191
+▁miguel -9.47192
+sistió -9.47192
+▁combate -9.47192
+▁perdió -9.47194
+▁cadena -9.47195
+▁llena -9.47204
+▁muere -9.47232
+▁cultural -9.47278
+▁comunicación -9.47341
+▁realidad -9.47358
+▁internacionales -9.47416
+▁continuación -9.47525
+▁cura -9.47554
+ular -9.47589
+▁bosque -9.47689
+bur -9.47712
+▁arco -9.47915
+▁bra -9.48026
+▁medios -9.4809
+▁anima -9.48112
+▁locales -9.48198
+▁bel -9.48217
+ónica -9.48259
+▁medida -9.48343
+creció -9.48374
+▁organiza -9.48388
+▁carretera -9.48473
+▁permanente -9.48473
+▁cuadra -9.48474
+▁entrenador -9.48497
+▁soldado -9.48507
+▁actuación -9.48514
+▁planeta -9.48523
+▁canto -9.48671
+▁cambiar -9.49004
+tima -9.49038
+ty -9.49073
+▁altura -9.49706
+▁regional -9.49772
+▁reserva -9.49772
+▁abuelo -9.49772
+▁distancia -9.49772
+▁cráter -9.49772
+▁escudo -9.49772
+▁aumenta -9.49773
+▁francesa -9.49774
+▁alberga -9.49806
+▁inmediatamente -9.4981
+▁apena -9.49871
+▁salvador -9.49873
+▁tratamiento -9.49873
+▁pena -9.49893
+ólogo -9.49928
+▁encontraba -9.49978
+▁julio -9.50019
+▁rock -9.50028
+▁adelante -9.50041
+▁árboles -9.50098
+em -9.50268
+▁prepara -9.50409
+▁romano -9.50465
+cida -9.50643
+▁jugadores -9.50729
+var -9.50803
+▁genera -9.51017
+▁arquitecto -9.51088
+▁disponible -9.51088
+▁ministerio -9.51088
+▁amenaza -9.51088
+▁tradición -9.51088
+▁hembra -9.51089
+▁fotografía -9.51096
+▁asociación -9.51194
+▁ruso -9.51298
+▁carne -9.51398
+▁ángeles -9.51418
+▁mejores -9.51649
+▁palabras -9.51689
+om -9.51705
+▁solamente -9.51753
+▁pensar -9.51821
+▁emplea -9.52062
+bert -9.52254
+▁esos -9.523
+▁ainhoa -9.52421
+▁cuarenta -9.52421
+▁músico -9.52421
+▁idioma -9.52421
+▁cantidad -9.52422
+▁mitad -9.52429
+▁depende -9.5244
+▁voces -9.52458
+▁zo -9.52715
+▁actu -9.52724
+▁creado -9.52725
+up -9.52819
+▁señal -9.52904
+cri -9.52983
+▁italiano -9.52983
+▁metros -9.53257
+▁clases -9.53294
+cá -9.53346
+▁pesca -9.53561
+▁sector -9.53644
+▁evitar -9.53772
+▁cielo -9.53774
+▁medalla -9.53777
+▁error -9.53786
+▁comercio -9.53813
+▁compuesto -9.53833
+ira -9.53895
+▁administración -9.53919
+▁desarroll -9.53947
+▁venta -9.53958
+▁ur -9.54027
+▁ataca -9.5412
+izada -9.5417
+▁encontra -9.54188
+▁dado -9.54554
+▁sha -9.54568
+ist -9.54628
+osas -9.54712
+▁frecuente -9.54834
+▁españoles -9.54874
+▁participa -9.54921
+▁haga -9.54966
+pen -9.55093
+▁construye -9.55142
+▁catedral -9.55142
+▁continúa -9.55142
+▁dirigió -9.55142
+▁izquierda -9.55142
+▁recuerda -9.55142
+▁canadá -9.55142
+▁guitarra -9.55142
+▁hicieron -9.55161
+▁usado -9.55194
+▁áreas -9.55229
+▁comida -9.55242
+▁fuentes -9.55425
+▁críticos -9.55481
+rry -9.55497
+▁diputado -9.55499
+▁programas -9.55559
+▁inferiores -9.55565
+ham -9.5572
+puesta -9.55746
+▁chicos -9.55836
+▁clara -9.56111
+▁desempeña -9.56204
+▁separa -9.5628
+berg -9.56427
+▁grabado -9.56457
+▁dirigida -9.56531
+▁juventud -9.56531
+▁amistad -9.56531
+▁combina -9.56532
+▁agrícola -9.56533
+▁mas -9.56653
+▁delantero -9.56704
+▁diseñado -9.56725
+▁vuelta -9.5678
+▁escala -9.56821
+▁doctora -9.56876
+▁debería -9.56877
+▁pelo -9.57106
+▁demostr -9.57144
+▁fl -9.57155
+▁ayudar -9.57362
+▁adopt -9.57376
+▁dur -9.57377
+mán -9.5741
+▁tierras -9.57413
+▁forman -9.57437
+tano -9.5754
+▁bio -9.5755
+stre -9.5786
+▁energía -9.57939
+▁frecuencia -9.57939
+▁mexicano -9.57939
+▁dispone -9.5794
+▁noroeste -9.5794
+▁obispo -9.5794
+▁primaria -9.5794
+▁europeo -9.5794
+▁turismo -9.57942
+▁libertad -9.57944
+▁dinero -9.57947
+▁rusia -9.57947
+▁garcía -9.57958
+▁eco -9.57965
+▁decora -9.58047
+▁letras -9.58057
+▁inicio -9.58073
+▁am -9.58082
+▁cho -9.58115
+▁anteriormente -9.58399
+▁poli -9.58406
+▁ple -9.58465
+▁marina -9.58627
+▁aves -9.58657
+bri -9.58714
+▁piel -9.58729
+▁mad -9.5876
+ense -9.58811
+▁leche -9.58889
+▁sociales -9.58938
+▁metal -9.58972
+visión -9.58997
+▁caer -9.5905
+▁acompaña -9.59051
+q -9.59179
+▁abre -9.59262
+▁constituye -9.59368
+▁corresponde -9.59368
+▁fundamental -9.59368
+▁huevos -9.59368
+▁imposible -9.59368
+▁vídeo -9.59368
+▁monumento -9.59368
+▁regresó -9.59368
+▁titular -9.59368
+▁sombra -9.59368
+▁cultivo -9.59369
+▁europea -9.59369
+▁afecta -9.59376
+▁borde -9.59378
+▁firma -9.59379
+▁derecha -9.59387
+▁encima -9.59398
+▁pintor -9.59404
+▁carta -9.59444
+iana -9.59484
+▁colonia -9.59489
+bel -9.59503
+▁hered -9.59576
+rien -9.59606
+▁contrario -9.5961
+▁completo -9.59664
+▁nacionales -9.59747
+▁barba -9.59916
+▁estela -9.59926
+▁activo -9.59983
+▁posterior -9.60036
+▁buscar -9.60196
+▁pop -9.60277
+▁clave -9.60491
+▁aprendi -9.60498
+▁investiga -9.60531
+▁flor -9.60535
+▁noble -9.60807
+▁argentino -9.60817
+▁compositor -9.60817
+▁consiguió -9.60817
+▁periodista -9.60817
+▁página -9.60817
+▁negocio -9.60817
+▁brillante -9.60818
+▁conferencia -9.60818
+▁elabora -9.60818
+▁agrupa -9.60819
+▁herido -9.60851
+cito -9.60922
+▁saber -9.61091
+▁llegado -9.61137
+▁clasifica -9.61148
+▁productos -9.61288
+▁nace -9.61421
+▁limita -9.61442
+▁extra -9.61545
+▁vela -9.61758
+▁marc -9.61763
+▁entrega -9.61766
+▁pudo -9.61801
+▁blancas -9.6191
+▁musicales -9.61941
+▁cabe -9.62086
+iano -9.62224
+▁pasos -9.62226
+▁añad -9.6227
+▁ejecuta -9.62288
+▁oportunidad -9.62288
+▁proporciona -9.62295
+▁respuesta -9.62297
+▁noticia -9.623
+▁dolor -9.62323
+▁distribución -9.6239
+kin -9.62398
+bido -9.62417
+▁habitantes -9.6243
+ima -9.62443
+▁señora -9.62565
+▁basta -9.62751
+▁pico -9.62768
+▁pregunta -9.62851
+▁ubicada -9.63063
+▁unido -9.63095
+plement -9.6319
+▁flora -9.63601
+▁ciencia -9.63756
+▁practica -9.6378
+▁velocidad -9.6378
+▁william -9.6378
+▁contacto -9.6378
+▁poema -9.63783
+▁junio -9.63784
+▁th -9.63905
+▁sh -9.63944
+▁tocar -9.64043
+▁participación -9.64108
+▁crece -9.64108
+rena -9.64165
+▁new -9.64266
+▁repar -9.64385
+▁liberal -9.64703
+▁león -9.64753
+▁esperaba -9.64855
+▁dedicada -9.64869
+▁dra -9.64889
+▁igualmente -9.65013
+▁calles -9.65025
+▁autores -9.65061
+▁vue -9.65122
+ui -9.65124
+▁ayuntamiento -9.65296
+▁colección -9.65296
+▁económico -9.65296
+▁introduc -9.65296
+▁octubre -9.65296
+▁gamboa -9.65296
+▁identifica -9.65296
+▁futuro -9.65296
+▁vecino -9.65298
+▁éste -9.65299
+▁instrumento -9.6531
+▁pueda -9.65316
+▁esperanza -9.65335
+▁producido -9.65338
+▁grave -9.65418
+▁rápidamente -9.65444
+▁condena -9.65635
+▁curs -9.6574
+▁nuevamente -9.66097
+▁guion -9.66272
+▁caza -9.66285
+bro -9.66306
+▁inferior -9.66361
+▁gre -9.66536
+▁cap -9.66654
+▁norma -9.66825
+▁fernando -9.66834
+▁pájaro -9.66834
+▁religiosa -9.66834
+▁capacidad -9.66834
+▁secreto -9.66834
+▁privada -9.66836
+▁diferente -9.66853
+▁lá -9.66864
+▁pelea -9.66868
+▁comisión -9.66882
+▁breve -9.66883
+▁mantener -9.66908
+▁mesa -9.66909
+▁inicialmente -9.67054
+▁salta -9.67089
+▁pista -9.6709
+tru -9.67168
+▁paisaje -9.67286
+▁hacía -9.67554
+▁we -9.67702
+▁minutos -9.67786
+▁aparecen -9.67789
+▁x -9.67855
+tario -9.67942
+▁compone -9.67973
+▁conoci -9.6811
+▁matar -9.68304
+▁colabora -9.68354
+▁científico -9.68396
+▁documento -9.68396
+▁originario -9.68396
+▁primavera -9.68396
+▁dibujo -9.68397
+▁símbolo -9.68397
+▁columna -9.68397
+▁oficina -9.68397
+▁paraguay -9.68397
+▁cuento -9.68401
+▁dedicado -9.68401
+▁piloto -9.68401
+▁apoya -9.68412
+▁cuá -9.68466
+▁milita -9.6849
+▁sierra -9.68517
+▁corea -9.68528
+▁derrota -9.68586
+▁representado -9.68738
+▁sé -9.68819
+cepción -9.69206
+▁camp -9.69246
+▁hacerlo -9.69332
+éis -9.69486
+▁capa -9.69812
+▁fuego -9.6988
+▁representante -9.6989
+caso -9.69924
+▁falleció -9.69984
+▁ferrocarril -9.69984
+▁portugal -9.69984
+▁construido -9.69984
+▁navidad -9.69984
+▁imagen -9.69984
+▁cultiva -9.69984
+▁federal -9.69985
+▁guinea -9.69985
+▁lanzamiento -9.69986
+▁demás -9.69986
+▁flota -9.69987
+▁acusa -9.70002
+▁infancia -9.70005
+▁ganar -9.70083
+▁humano -9.70148
+▁caballos -9.70179
+▁piano -9.70186
+▁semanas -9.70222
+und -9.70251
+▁alcanz -9.70273
+quí -9.70325
+▁corri -9.70346
+▁intento -9.70679
+americana -9.71089
+▁foto -9.71131
+▁contar -9.71158
+▁estrellas -9.71184
+▁piezas -9.71245
+▁salida -9.71261
+▁rara -9.71514
+▁pies -9.71547
+▁adicional -9.71597
+▁histórica -9.71597
+▁insectos -9.71597
+▁lluvia -9.71597
+▁originaria -9.71597
+▁requiere -9.71597
+▁revolución -9.71597
+▁orquesta -9.71597
+▁profunda -9.71597
+▁graduó -9.71597
+▁hierro -9.71598
+ible -9.71598
+▁revela -9.716
+▁convento -9.71602
+▁marrón -9.71604
+▁tienda -9.71604
+▁contigo -9.71615
+▁punta -9.71617
+▁trabajado -9.7162
+▁presentan -9.71623
+▁ramón -9.71641
+com -9.71667
+▁infl -9.71667
+▁rural -9.71675
+▁hubiera -9.71678
+▁efecto -9.71833
+▁calidad -9.71852
+▁llevan -9.71889
+▁volar -9.71906
+▁ronda -9.72037
+▁activa -9.72199
+▁moderno -9.72229
+▁basada -9.72559
+▁apart -9.72621
+▁niveles -9.72664
+▁tren -9.72846
+▁cien -9.72906
+cina -9.7302
+▁suficiente -9.73236
+▁consecuencia -9.73236
+▁justicia -9.73236
+▁encontró -9.73236
+▁griego -9.73236
+▁comprende -9.73237
+▁quería -9.73242
+▁explica -9.73249
+▁regla -9.73252
+▁gobernador -9.73257
+▁amiga -9.73262
+▁borda -9.73264
+▁comarca -9.73275
+▁exposiciones -9.73289
+▁fri -9.7331
+▁formato -9.73328
+▁ido -9.73328
+▁rosario -9.73355
+▁patio -9.73375
+▁escenario -9.73387
+▁hablando -9.73675
+▁enfrenta -9.73732
+rito -9.73819
+▁revistas -9.73962
+▁dicen -9.73997
+▁fall -9.74001
+▁rico -9.7403
+▁bailar -9.74094
+▁estudiantes -9.74183
+▁salva -9.7419
+fre -9.7422
+▁estatal -9.74304
+▁graba -9.74316
+▁registro -9.74352
+▁robert -9.7436
+▁artículos -9.74376
+zón -9.74388
+▁armas -9.74405
+▁empeza -9.74605
+ang -9.74682
+riendo -9.74704
+▁fábrica -9.74903
+▁resistencia -9.74903
+▁regreso -9.74903
+▁positiva -9.74903
+▁describe -9.74903
+▁nacido -9.74906
+▁nieve -9.74907
+▁proviene -9.74913
+▁comuna -9.74917
+▁fresco -9.7493
+▁debajo -9.74969
+▁reunión -9.74974
+▁peces -9.75089
+▁usar -9.75283
+▁desempeñ -9.75298
+▁evento -9.75313
+▁debía -9.75342
+▁reconocido -9.75342
+▁quedan -9.7539
+▁montañas -9.75426
+▁posteriores -9.75453
+▁masa -9.75559
+▁mc -9.75644
+▁formado -9.75952
+▁grit -9.76116
+▁restaura -9.76121
+▁eventos -9.76183
+▁usada -9.76278
+▁reuni -9.76483
+mó -9.76501
+▁euro -9.76582
+▁martín -9.76587
+▁concurso -9.76598
+▁cubierto -9.76598
+▁mexicana -9.76598
+▁plataforma -9.76598
+▁princesa -9.76598
+▁príncipe -9.76598
+▁artística -9.76598
+▁impuls -9.76598
+▁literaria -9.76598
+▁fruta -9.76598
+▁afirma -9.76598
+▁cierta -9.766
+▁conecta -9.76602
+▁funciones -9.76603
+▁reconocimiento -9.76606
+▁conmigo -9.76607
+▁enrique -9.76613
+▁dientes -9.76624
+ther -9.76632
+▁ecuador -9.76635
+▁selva -9.76652
+▁arregl -9.76686
+▁compara -9.76686
+▁sangre -9.76698
+▁sección -9.76707
+▁tenían -9.76733
+▁poblado -9.76769
+▁empresas -9.76814
+▁considerable -9.76939
+▁cambia -9.7694
+by -9.77055
+▁brazos -9.77062
+▁moderna -9.7721
+▁enamorado -9.77312
+▁materiales -9.77323
+▁arre -9.77344
+vil -9.7758
+▁utilizan -9.77991
+▁efectos -9.78069
+▁traduc -9.78222
+▁llam -9.78235
+▁recupera -9.78236
+▁comunica -9.78287
+▁gobern -9.78302
+▁candidato -9.78322
+▁córdoba -9.78322
+▁imágenes -9.78322
+▁judicial -9.78322
+▁núcleo -9.78322
+▁posibilidad -9.78322
+▁septiembre -9.78322
+▁sitúa -9.78322
+▁concepto -9.78322
+▁ganadería -9.78322
+▁milagro -9.78322
+▁tormenta -9.78322
+▁congreso -9.78322
+▁naturaleza -9.78323
+▁cuarta -9.78324
+▁crecimiento -9.78325
+▁avanza -9.78326
+▁prensa -9.78331
+▁decisión -9.78333
+▁emigr -9.78336
+▁sabor -9.7834
+ure -9.78355
+▁elecciones -9.78405
+▁gato -9.78478
+▁mud -9.78512
+▁beso -9.78582
+▁peter -9.78645
+▁fac -9.78735
+▁creación -9.78737
+▁finales -9.78767
+▁viven -9.78808
+▁navega -9.78815
+▁extendi -9.78868
+▁estudiar -9.78923
+▁bre -9.7895
+▁ramas -9.79126
+creta -9.79144
+▁tru -9.79286
+▁fila -9.79319
+cada -9.79337
+▁cable -9.79363
+▁plato -9.79471
+▁recursos -9.79478
+▁habitual -9.79766
+▁asesina -9.79779
+▁volv -9.79951
+▁incorpora -9.79991
+ice -9.80006
+▁ascenso -9.80076
+▁construida -9.80076
+▁facultad -9.80076
+▁murmuró -9.80076
+▁necesario -9.80076
+▁víctima -9.80076
+▁caballero -9.80077
+▁regiones -9.80077
+▁adulto -9.80077
+▁dispara -9.80078
+▁conquista -9.80082
+▁aprob -9.80084
+▁suerte -9.80104
+▁modific -9.80104
+▁ed -9.80154
+▁recibido -9.80156
+▁tabla -9.80182
+▁pod -9.80203
+▁pecho -9.80301
+▁abrir -9.80317
+▁húmedos -9.80341
+▁banco -9.80596
+▁et -9.80626
+▁pel -9.80795
+▁piso -9.80999
+▁lope -9.81085
+izar -9.81355
+▁huel -9.81385
+▁bon -9.81623
+▁sor -9.81702
+gráfica -9.81764
+▁diócesis -9.81862
+▁jamás -9.81862
+▁límite -9.81862
+▁monasterio -9.81862
+▁respecto -9.81862
+▁school -9.81862
+▁vehículo -9.81862
+▁implica -9.81862
+▁gloria -9.81862
+▁cuadro -9.81862
+▁ramiro -9.81863
+▁ocasión -9.81864
+▁sonrisa -9.81864
+▁rango -9.81875
+▁establece -9.81882
+▁cerrado -9.81882
+▁cristal -9.81888
+▁producir -9.81905
+▁domingo -9.81915
+▁vocal -9.81918
+▁sacerdote -9.81956
+▁raro -9.82
+▁metro -9.82019
+▁mancha -9.8204
+quiera -9.82519
+▁factor -9.82558
+▁directo -9.82584
+▁cab -9.82714
+▁desea -9.82731
+▁rob -9.82802
+▁emb -9.82805
+▁descubri -9.83043
+▁misión -9.83208
+▁distin -9.83347
+▁conciertos -9.83463
+▁peligro -9.83526
+▁escribir -9.83554
+▁encarga -9.83592
+▁result -9.8364
+▁conflicto -9.8368
+▁george -9.8368
+▁temperatura -9.8368
+▁usuario -9.8368
+▁enemigo -9.8368
+▁michael -9.83681
+▁secretario -9.83681
+▁jorge -9.83683
+▁riesgo -9.83683
+▁alegre -9.83688
+▁composición -9.83693
+▁tribu -9.83701
+▁rompe -9.83709
+▁coloca -9.83711
+▁agrega -9.83717
+▁editor -9.83726
+▁cocina -9.83759
+▁beber -9.83774
+▁nativa -9.83818
+▁abundante -9.83854
+▁atención -9.83881
+▁constante -9.8397
+▁neu -9.83999
+▁nega -9.84403
+▁deporte -9.84641
+han -9.8468
+▁menores -9.84925
+ight -9.8506
+▁dirige -9.85334
+▁mac -9.85432
+▁you -9.85445
+▁comienzo -9.85532
+▁ingeniero -9.85532
+▁judío -9.85532
+▁respira -9.85532
+▁circula -9.85532
+▁rechaza -9.85533
+▁florece -9.85534
+▁retrato -9.85537
+▁viaja -9.85538
+▁sevilla -9.85538
+▁permaneció -9.85548
+▁bandera -9.85554
+▁bloque -9.85561
+▁conocimiento -9.85562
+▁laguna -9.85578
+▁prisión -9.856
+▁barra -9.85795
+▁traje -9.85841
+▁publicación -9.85934
+▁dime -9.86114
+▁patas -9.86391
+▁tendrá -9.86695
+▁localidades -9.86797
+▁chico -9.86873
+▁árbol -9.86945
+▁ángel -9.86946
+▁administra -9.87167
+▁militares -9.87252
+▁familiar -9.87349
+▁descansa -9.87404
+▁ceremonia -9.87419
+▁competición -9.87419
+▁contempla -9.87419
+▁extranjero -9.87419
+▁independencia -9.87419
+▁juvenil -9.87419
+▁periodismo -9.87419
+▁protagonista -9.87419
+▁universitario -9.87419
+▁órgano -9.87419
+▁jesús -9.87419
+▁turístico -9.87419
+▁vicente -9.8742
+▁vacaciones -9.87421
+▁privado -9.87421
+▁adquiri -9.87422
+ascendencia -9.87426
+▁seguido -9.87428
+▁avión -9.87436
+▁animal -9.87459
+▁mr -9.87466
+▁cruc -9.87481
+▁contrato -9.87523
+▁operaciones -9.87613
+ting -9.87659
+▁aten -9.87696
+▁river -9.87781
+▁veo -9.87814
+▁mando -9.87815
+▁frecuentemente -9.87846
+▁presentado -9.87867
+▁restaurante -9.87953
+▁guía -9.87961
+ford -9.87987
+▁ropa -9.88041
+▁varía -9.88189
+▁deportes -9.88466
+▁enferm -9.88466
+▁marino -9.8857
+▁vende -9.88576
+▁calma -9.88643
+▁sello -9.88774
+▁incorpor -9.88777
+amiento -9.88862
+▁viva -9.88877
+▁jr -9.88974
+ándome -9.88995
+▁sabi -9.89026
+pp -9.89057
+▁má -9.89125
+▁prof -9.8913
+▁unidad -9.89169
+▁portada -9.89272
+▁presentó -9.89342
+▁veracruz -9.89342
+▁parlamento -9.89342
+▁solicit -9.89342
+▁hambre -9.89344
+▁tropas -9.89348
+▁alimento -9.89351
+▁administrativa -9.89363
+▁tallo -9.89402
+▁recuerdo -9.89408
+▁relacionado -9.89472
+▁fil -9.89473
+▁daño -9.89495
+▁promet -9.89592
+▁perder -9.89592
+▁regular -9.89632
+ómic -9.89816
+▁coge -9.8988
+▁prote -9.8989
+▁papa -9.89905
+▁cercana -9.89954
+▁destac -9.90238
+▁representación -9.90254
+▁anteriores -9.90314
+▁salvar -9.90339
+▁comerciales -9.90376
+▁interna -9.90415
+▁provincial -9.90512
+▁tim -9.90711
+▁provoca -9.90712
+▁lateral -9.90876
+▁imagina -9.90884
+▁oficiales -9.91007
+▁anda -9.91017
+▁aplica -9.91019
+▁autoridad -9.91038
+▁parti -9.91166
+▁especializ -9.91174
+▁colabor -9.91183
+ney -9.91246
+▁cincuenta -9.91303
+▁documental -9.91303
+▁esfuerzo -9.91303
+▁ilustra -9.91303
+▁ocurre -9.91303
+▁secundario -9.91303
+▁tecnología -9.91303
+▁árabe -9.91303
+▁asegur -9.91304
+▁reforma -9.91304
+▁cuenca -9.91304
+▁procede -9.91306
+▁vieja -9.91306
+▁pedir -9.91309
+▁explota -9.9131
+▁condiciones -9.91315
+▁merced -9.91317
+▁existi -9.91325
+▁segura -9.91335
+▁tribunal -9.91335
+▁james -9.91338
+▁protagonizada -9.91345
+▁west -9.91576
+▁retira -9.9158
+dujo -9.91601
+▁herm -9.91749
+▁pescado -9.91805
+▁loco -9.91892
+▁llevado -9.92132
+▁abandona -9.92285
+▁turb -9.92342
+▁rap -9.92484
+▁bill -9.92595
+ville -9.92942
+▁sch -9.92945
+▁somo -9.93098
+▁relaciona -9.93173
+▁destacado -9.93252
+▁alterna -9.9329
+▁adolescente -9.93303
+▁desaparece -9.93303
+▁humilde -9.93303
+▁manifesta -9.93303
+▁sombrero -9.93303
+▁viajó -9.93303
+▁austria -9.93303
+▁panamá -9.93303
+▁guardia -9.93303
+▁frío -9.93303
+▁espacial -9.93303
+▁comparti -9.93304
+▁habría -9.93305
+▁avenida -9.93306
+▁anual -9.93311
+▁diego -9.93312
+▁cinta -9.93319
+▁declara -9.93323
+▁carece -9.93328
+▁archivo -9.93336
+▁hotel -9.93341
+▁produjo -9.93347
+▁recoge -9.93379
+▁concentra -9.93386
+▁angel -9.93402
+▁producto -9.93438
+▁verdadero -9.93622
+▁camina -9.9363
+trás -9.93689
+▁cena -9.937
+▁ciudadano -9.93703
+▁sentí -9.93744
+bula -9.93843
+▁ll -9.93858
+▁desarrollado -9.93958
+▁terminar -9.94042
+▁creer -9.94301
+▁merc -9.94397
+▁cría -9.94552
+▁casti -9.94682
+▁malo -9.94761
+tch -9.94772
+▁haz -9.94784
+▁brazo -9.94852
+▁viviend -9.95025
+▁crear -9.95205
+▁pega -9.95249
+▁marido -9.95318
+▁búsqueda -9.95343
+▁extensión -9.95343
+▁incluido -9.95343
+▁intensa -9.95343
+▁próximo -9.95343
+▁religioso -9.95343
+▁vegetación -9.95343
+▁acabó -9.95344
+▁college -9.95344
+▁retiró -9.95344
+▁desconoce -9.95344
+▁consigue -9.95344
+▁ensayo -9.95344
+▁internet -9.95345
+cedió -9.95368
+▁bat -9.95384
+▁allá -9.9539
+▁enorme -9.95417
+▁dificultad -9.95421
+▁interesante -9.95544
+▁demanda -9.95569
+bió -9.95578
+▁recientemente -9.95614
+▁socio -9.95718
+▁invitado -9.95809
+▁licenciado -9.95852
+▁escultura -9.95893
+▁tono -9.95912
+▁interpretación -9.96
+▁capaz -9.96111
+▁habilidad -9.9615
+erra -9.96323
+▁ideal -9.96401
+▁permanec -9.96435
+▁misa -9.96515
+▁participado -9.96516
+umb -9.96773
+▁tiro -9.96931
+▁integr -9.96966
+▁llegan -9.9708
+▁fácil -9.97118
+imiento -9.97379
+▁abadesa -9.97427
+▁británica -9.97427
+▁espectáculo -9.97427
+▁excelente -9.97427
+▁extraña -9.97427
+▁irregular -9.97427
+▁mediterráneo -9.97427
+▁máximo -9.97427
+▁organismo -9.97427
+▁presidencia -9.97427
+▁febrero -9.97427
+▁pasajeros -9.97427
+▁templado -9.97427
+▁entiende -9.97427
+▁suroeste -9.97427
+▁digital -9.97427
+▁volcán -9.97427
+▁transporte -9.97427
+▁escolar -9.9743
+▁humana -9.97431
+▁creó -9.97443
+▁relato -9.97465
+▁abril -9.97472
+▁aroma -9.97484
+▁múltiples -9.97492
+▁aleja -9.97494
+▁sonata -9.97542
+▁abr -9.97557
+▁ejerció -9.97597
+▁detalle -9.97617
+▁clasificación -9.97723
+▁peso -9.97894
+tracción -9.97941
+▁partida -9.98135
+▁parroquia -9.98166
+▁llegaron -9.98203
+▁especiales -9.98288
+▁marco -9.98348
+guer -9.98387
+▁asociado -9.98561
+▁tí -9.9876
+▁experiment -9.98781
+▁conservan -9.99008
+▁llevaba -9.99063
+▁hol -9.99074
+▁franco -9.99079
+▁tratado -9.99125
+▁podi -9.99341
+mbo -9.99543
+oña -9.99554
+▁conseguir -9.99554
+▁construir -9.99554
+▁extraño -9.99554
+▁isabel -9.99554
+▁obtiene -9.99554
+▁responsable -9.99554
+▁temporal -9.99554
+▁temprana -9.99554
+▁divide -9.99554
+▁móvil -9.99554
+▁convertido -9.99555
+▁femenina -9.99555
+▁limpia -9.99555
+▁chris -9.99555
+▁estrecha -9.99556
+▁quinto -9.9956
+▁vivía -9.99562
+▁ruinas -9.99564
+▁llora -9.99572
+▁alumno -9.99574
+▁escasa -9.99585
+ticos -9.99589
+▁publicaciones -9.99765
+▁cala -9.99878
+▁socialista -10.0006
+ificación -10.0009
+▁marzo -10.0013
+▁recuperar -10.0017
+▁podrá -10.0037
+▁suiza -10.0039
+▁cerro -10.004
+▁bot -10.0057
+▁pala -10.007
+▁piensa -10.0075
+▁cuales -10.0107
+▁fiel -10.0112
+▁deseo -10.0112
+▁universidades -10.0117
+iga -10.0124
+misión -10.0128
+▁cero -10.0129
+▁inicial -10.0133
+▁piens -10.0152
+▁construyó -10.0173
+▁extensa -10.0173
+▁fantasma -10.0173
+▁jornada -10.0173
+▁oscura -10.0173
+▁usual -10.0173
+▁except -10.0173
+▁profesión -10.0173
+▁show -10.0173
+▁dormir -10.0173
+▁teatral -10.0173
+▁tendencia -10.0173
+▁explora -10.0173
+▁profundo -10.0173
+▁hermoso -10.0173
+▁perenne -10.0173
+▁quince -10.0173
+▁pierna -10.0174
+▁infantil -10.0174
+▁cabello -10.0176
+▁suave -10.018
+▁obtener -10.0182
+▁gusto -10.0182
+▁cercano -10.0184
+▁palo -10.0186
+▁taller -10.0196
+▁mono -10.02
+▁grabación -10.0202
+▁calor -10.0207
+▁empezar -10.0213
+▁fija -10.0214
+▁poca -10.023
+▁ku -10.0237
+▁captura -10.0249
+nominada -10.0264
+▁profesora -10.027
+▁culturales -10.0273
+▁similares -10.0273
+▁superiores -10.0303
+ive -10.0307
+arias -10.0308
+▁plano -10.0315
+▁clubes -10.0338
+▁casada -10.034
+▁terminal -10.0342
+ling -10.0371
+▁católico -10.0395
+▁convierte -10.0395
+▁eduardo -10.0395
+▁homenaje -10.0395
+▁rodríguez -10.0395
+▁rápido -10.0395
+▁terrestre -10.0395
+▁sobrino -10.0395
+▁confirma -10.0395
+▁enseñanza -10.0395
+▁aproximadamente -10.0395
+▁alemana -10.0395
+▁aventura -10.0395
+▁editorial -10.0395
+▁daniel -10.0396
+▁famosa -10.0396
+▁batería -10.0396
+▁tarea -10.0396
+▁baile -10.0397
+▁rumbo -10.0398
+firmó -10.0398
+▁nacimiento -10.0398
+tío -10.04
+▁labios -10.04
+▁gracia -10.0405
+idez -10.0406
+▁investigador -10.0407
+▁suceso -10.0411
+▁lenguaje -10.0411
+old -10.042
+▁palma -10.0425
+▁flu -10.0431
+▁generación -10.0433
+▁impresion -10.0434
+disc -10.0442
+▁cop -10.0453
+▁americano -10.0454
+▁hecha -10.0455
+▁escribe -10.0464
+▁componente -10.047
+▁gri -10.0488
+▁tenga -10.0499
+▁edita -10.0506
+▁moral -10.0533
+▁manda -10.0535
+▁letra -10.0539
+▁populares -10.058
+▁licencia -10.0597
+▁tendr -10.0601
+▁domina -10.0608
+▁destacan -10.0608
+▁importa -10.0609
+▁instru -10.0621
+▁análisis -10.0622
+▁evidencia -10.0622
+▁ocupó -10.0622
+▁ámbito -10.0622
+▁literario -10.0622
+▁razones -10.0622
+▁refleja -10.0622
+▁catorce -10.0622
+▁arroyo -10.0622
+▁apoyo -10.0622
+▁comunes -10.0622
+▁gigante -10.0623
+▁agosto -10.0623
+▁cueva -10.0623
+▁talento -10.0623
+iforme -10.0623
+▁bronce -10.0623
+▁láser -10.0624
+▁caída -10.0624
+▁psico -10.0624
+▁tercero -10.0625
+▁expuls -10.0625
+▁city -10.0625
+▁rayo -10.0628
+▁caliente -10.063
+▁retorn -10.0631
+▁sonora -10.0633
+▁joder -10.0635
+▁ayer -10.0636
+▁participar -10.064
+▁verdadera -10.0641
+▁ze -10.0652
+▁originales -10.0654
+▁levantó -10.0661
+▁music -10.0665
+▁alcanza -10.0671
+▁park -10.0674
+▁unidades -10.0686
+gram -10.0693
+ku -10.0694
+▁ampliamente -10.0696
+▁separado -10.0709
+▁esperar -10.0728
+tuve -10.0731
+▁realizada -10.0732
+▁coleccion -10.0744
+▁duro -10.0752
+▁hip -10.0756
+▁enamora -10.0758
+▁utilizada -10.0785
+▁invit -10.0805
+▁produc -10.0817
+▁calla -10.0818
+▁ferr -10.0819
+▁reciente -10.0831
+acional -10.0833
+▁contribuy -10.0855
+nunció -10.0855
+▁abandonó -10.0855
+▁asamblea -10.0855
+▁cataluña -10.0855
+▁clásica -10.0855
+▁diciembre -10.0855
+▁filipinas -10.0855
+▁físico -10.0855
+▁herramienta -10.0855
+▁ilumina -10.0855
+▁noviembre -10.0855
+▁otoño -10.0855
+▁reconoce -10.0855
+▁vietnam -10.0855
+▁artístico -10.0855
+▁emperador -10.0855
+▁texas -10.0855
+▁entrevista -10.0856
+▁cristiano -10.0856
+▁griega -10.0856
+▁lleno -10.0856
+▁migra -10.0857
+▁desperta -10.0857
+▁instituciones -10.0858
+▁enfrentamiento -10.0858
+▁abrió -10.0859
+▁procesa -10.086
+▁bonita -10.0861
+▁característica -10.0861
+▁justo -10.0862
+▁protesta -10.0863
+▁esencial -10.087
+▁chu -10.0872
+▁narra -10.0873
+▁subtropicales -10.0875
+▁sexo -10.088
+▁huye -10.088
+▁monja -10.0884
+▁hici -10.0885
+▁repit -10.0899
+▁estren -10.09
+▁registr -10.0903
+▁autoridades -10.0912
+▁teología -10.0913
+▁she -10.0938
+▁siente -10.0972
+▁onda -10.0978
+▁utilizar -10.0986
+nzó -10.0995
+▁besa -10.0998
+▁realmente -10.1003
+▁alcanzar -10.1011
+▁tío -10.1026
+▁tropical -10.1046
+▁arm -10.1056
+▁materia -10.1056
+▁silla -10.106
+zco -10.1068
+▁crec -10.107
+▁colaborador -10.1077
+▁descubierto -10.1093
+▁específica -10.1093
+▁indígena -10.1093
+▁inflorescencia -10.1093
+▁inglesa -10.1093
+▁mensaje -10.1093
+▁religión -10.1093
+▁republicano -10.1093
+▁temprano -10.1093
+▁zaragoza -10.1093
+▁aparición -10.1093
+▁deportiva -10.1093
+▁máquina -10.1093
+stituido -10.1093
+▁marquesa -10.1093
+▁pantalla -10.1093
+▁precisa -10.1093
+▁sirvió -10.1093
+▁visual -10.1093
+▁urbano -10.1093
+▁ediciones -10.1093
+▁asunto -10.1093
+▁rodeado -10.1094
+▁alberto -10.1094
+▁habrá -10.1094
+▁quinta -10.1094
+▁producida -10.1097
+▁rueda -10.1097
+▁archi -10.1103
+▁grecia -10.1104
+▁steve -10.1104
+▁peña -10.1105
+▁prisa -10.1105
+▁tumba -10.1105
+▁lanzado -10.1107
+▁criado -10.1107
+▁king -10.1116
+▁alarg -10.1129
+▁impacto -10.114
+▁operación -10.1142
+▁precio -10.1144
+▁existencia -10.1144
+▁saque -10.1147
+▁invent -10.1151
+▁perdona -10.1154
+▁franc -10.1174
+▁sam -10.1185
+▁encarg -10.1191
+gh -10.1194
+▁tanta -10.1208
+▁centra -10.1209
+▁comedia -10.122
+▁wo -10.1238
+grado -10.1257
+▁quita -10.1262
+▁enterra -10.1272
+▁acompañado -10.1277
+▁conocer -10.1302
+▁finalizar -10.1302
+▁exclusiva -10.1337
+▁femenino -10.1337
+▁identidad -10.1337
+▁olímpico -10.1337
+▁rectangular -10.1337
+▁sudáfrica -10.1337
+▁tailandia -10.1337
+▁telenovela -10.1337
+▁gestiona -10.1337
+▁quizá -10.1337
+▁hierba -10.1337
+▁equipa -10.1337
+▁servir -10.1337
+▁desnudo -10.1337
+▁bolsa -10.1337
+▁pierde -10.1338
+▁mixta -10.1338
+▁rostro -10.1339
+▁cierra -10.1339
+▁alternativa -10.1339
+▁equivalente -10.134
+▁herida -10.1341
+▁sexual -10.1342
+▁oído -10.1343
+▁once -10.1344
+▁cubre -10.1348
+▁ligeramente -10.1349
+▁mortal -10.1351
+▁mirada -10.1352
+lfa -10.1354
+▁directamente -10.1358
+▁concierto -10.1366
+▁tramo -10.1372
+▁participaron -10.1379
+▁senador -10.1386
+well -10.1392
+▁fama -10.1399
+▁particularmente -10.1401
+▁celos -10.1402
+▁amplio -10.1404
+ístico -10.1418
+▁descubr -10.142
+mont -10.142
+▁roberto -10.1428
+▁vega -10.1438
+▁lima -10.147
+▁pese -10.1471
+▁presta -10.1471
+▁pino -10.1474
+▁chino -10.1498
+ix -10.1504
+▁papeles -10.151
+▁terminado -10.1532
+▁adelant -10.1535
+▁transforma -10.1552
+▁interpret -10.1554
+▁espíritu -10.1587
+▁gonzález -10.1587
+▁inmigrantes -10.1587
+▁murmullo -10.1587
+▁turquía -10.1587
+▁divisiones -10.1587
+▁alemanes -10.1587
+▁ladrillo -10.1587
+▁felipe -10.1587
+▁riqueza -10.1587
+▁ácido -10.1587
+▁bahía -10.1587
+▁pablo -10.1587
+▁refiere -10.1587
+vertirse -10.1587
+▁presidi -10.1587
+▁berlín -10.1587
+▁vincula -10.1587
+▁ficción -10.1587
+▁asesor -10.1588
+▁expresa -10.1589
+▁martínez -10.1589
+▁zapat -10.159
+siguieron -10.159
+siguiendo -10.159
+▁interrog -10.1591
+▁dorado -10.1592
+▁suicid -10.1594
+▁copia -10.1595
+▁aplicaciones -10.16
+▁significado -10.1609
+▁max -10.1611
+▁top -10.1613
+▁variable -10.1618
+▁legal -10.1619
+▁colaboración -10.1626
+▁romana -10.1628
+▁elemento -10.1637
+▁danza -10.1644
+▁tomado -10.1644
+▁oír -10.165
+▁provoc -10.1663
+vió -10.1666
+▁orilla -10.1677
+▁experimenta -10.1679
+▁asocia -10.1686
+▁declar -10.173
+▁eje -10.1737
+▁actores -10.1742
+▁retirado -10.1749
+▁profesionales -10.1771
+▁impe -10.1798
+▁abandonar -10.1827
+posición -10.1831
+▁ecua -10.1838
+▁apariencia -10.1843
+▁científica -10.1843
+▁fenómeno -10.1843
+▁guitarrista -10.1843
+▁necesaria -10.1843
+▁próxima -10.1843
+▁alcance -10.1843
+▁solitario -10.1843
+▁transcur -10.1843
+▁paciente -10.1843
+▁gafas -10.1843
+▁elección -10.1843
+▁urbana -10.1843
+▁leyenda -10.1844
+▁crema -10.1844
+▁ejerce -10.1844
+▁conforma -10.1846
+▁galería -10.1849
+▁significativa -10.1849
+▁suecia -10.1852
+▁pariente -10.1854
+▁favorito -10.1863
+▁contaba -10.1865
+▁duque -10.1868
+▁primo -10.1871
+▁fiera -10.1871
+▁casco -10.1873
+▁sonreí -10.1876
+▁fácilmente -10.1882
+▁publicada -10.1888
+▁olor -10.1889
+▁salón -10.1889
+▁denominación -10.1896
+rgió -10.1901
+▁ruido -10.1908
+▁trabajando -10.1922
+▁pieza -10.1927
+▁sacar -10.196
+▁concepción -10.1971
+▁clo -10.1981
+▁nombra -10.2033
+▁cumpli -10.2035
+▁rose -10.2055
+▁basado -10.2081
+▁pareció -10.2106
+▁descendiente -10.2106
+▁diagrama -10.2106
+▁extiende -10.2106
+▁kilómetro -10.2106
+▁matemática -10.2106
+▁pertenecía -10.2106
+▁relieve -10.2106
+▁renuncia -10.2106
+▁arbusto -10.2106
+▁aprecia -10.2106
+▁escritura -10.2106
+▁estrib -10.2107
+▁apertura -10.2107
+▁escalera -10.2107
+▁coronel -10.2107
+▁imperial -10.2107
+▁cuerda -10.2107
+▁ancla -10.2107
+▁marinero -10.2107
+▁ahoga -10.2107
+▁encanta -10.2108
+▁asusta -10.2108
+▁soporta -10.2108
+▁multi -10.2109
+▁convoca -10.211
+▁aprender -10.211
+▁louis -10.2111
+▁estatua -10.2111
+▁micro -10.2112
+▁cliente -10.2113
+▁castellano -10.2114
+▁ji -10.2115
+▁interesa -10.2116
+▁institución -10.2117
+▁rival -10.2118
+▁lic -10.2119
+▁maria -10.2128
+▁trabajadores -10.2128
+▁corteza -10.2131
+▁mandato -10.2133
+▁pasión -10.2139
+gel -10.2156
+▁regal -10.2172
+▁linda -10.2178
+ini -10.2188
+▁reales -10.2189
+▁quedaba -10.2192
+▁paseo -10.2194
+▁radical -10.2206
+▁armada -10.2212
+▁chileno -10.2215
+▁propiedades -10.2226
+▁repent -10.2238
+▁cruzar -10.224
+▁ayud -10.2261
+▁informe -10.2263
+▁frase -10.2272
+▁vota -10.2293
+▁pot -10.2363
+▁derrotado -10.2366
+▁artificial -10.2377
+▁código -10.2377
+▁fortaleza -10.2377
+▁futbolista -10.2377
+▁guerrero -10.2377
+▁incidente -10.2377
+▁federación -10.2377
+▁multitud -10.2377
+▁richard -10.2377
+▁dibuja -10.2377
+▁globo -10.2377
+▁sabía -10.2377
+▁recorre -10.2377
+▁visible -10.2377
+▁oreja -10.2377
+▁electoral -10.2378
+▁liceo -10.2378
+▁ritmo -10.2378
+▁trayectoria -10.2378
+▁boston -10.2379
+▁single -10.2379
+▁propósito -10.238
+▁dominio -10.238
+▁florida -10.2383
+▁bordo -10.2384
+▁encaja -10.2384
+▁reclam -10.2385
+▁patrona -10.2386
+▁recibir -10.2387
+▁interven -10.2392
+cimiento -10.2393
+▁tesis -10.2394
+▁nativo -10.2407
+▁pensaba -10.2409
+▁presión -10.2409
+▁individuales -10.2417
+▁reconocida -10.2419
+wood -10.2434
+▁especializado -10.2437
+▁traza -10.2445
+agua -10.2453
+▁empleado -10.2458
+▁artículo -10.2463
+▁programación -10.2503
+▁mora -10.2503
+▁suyo -10.2518
+▁instala -10.2519
+mita -10.2529
+▁redes -10.2536
+▁inform -10.2589
+▁organizado -10.2606
+▁acontecimiento -10.2654
+▁bóveda -10.2654
+▁colectivo -10.2654
+▁competencia -10.2654
+▁defensor -10.2654
+▁egipto -10.2654
+▁ejemplar -10.2654
+▁máxima -10.2654
+▁prisionero -10.2654
+▁voluntario -10.2654
+▁washington -10.2654
+▁desemboca -10.2655
+▁volumen -10.2655
+▁aceite -10.2655
+▁rebelde -10.2655
+▁hidro -10.2655
+▁millones -10.2655
+▁ermita -10.2655
+▁cuello -10.2655
+▁ciclo -10.2655
+▁league -10.2655
+▁fiscal -10.2655
+▁repeti -10.2655
+▁esposo -10.2655
+▁poblaciones -10.2656
+▁catalan -10.2656
+▁defendi -10.2656
+▁noreste -10.2658
+▁juzg -10.2659
+▁marqués -10.2659
+▁cómic -10.266
+▁anillo -10.2662
+▁granada -10.2662
+▁dueño -10.2662
+▁propuesta -10.2663
+▁oriente -10.2666
+lito -10.2668
+▁barroco -10.2668
+▁viena -10.2672
+▁aldea -10.2676
+▁novia -10.2683
+▁popularidad -10.2686
+▁rusa -10.2687
+▁aguanta -10.2692
+▁laterales -10.2724
+rich -10.2731
+▁sectores -10.2756
+▁anot -10.2811
+▁oficialmente -10.2812
+▁informa -10.2814
+▁darle -10.282
+▁mate -10.2833
+▁astr -10.2833
+▁eva -10.2861
+▁ram -10.2879
+▁chap -10.2881
+▁obten -10.292
+▁manten -10.2926
+▁pluma -10.2935
+▁baloncesto -10.294
+▁camerún -10.294
+▁cerámica -10.294
+▁democrática -10.294
+▁guillermo -10.294
+▁hacienda -10.294
+▁individuo -10.294
+▁intercambi -10.294
+▁involucra -10.294
+▁pacífico -10.294
+▁plástico -10.294
+▁videojuego -10.294
+▁básica -10.294
+▁dinamarca -10.294
+▁necesidad -10.294
+▁química -10.294
+▁thomas -10.294
+▁cambió -10.294
+▁mecanismo -10.294
+▁típica -10.294
+▁violencia -10.294
+▁vuelva -10.294
+▁asistente -10.294
+▁temática -10.294
+▁ópera -10.294
+▁posiciones -10.294
+▁protección -10.294
+▁cosecha -10.294
+▁paredes -10.294
+▁irlanda -10.294
+▁almacena -10.294
+▁desierto -10.294
+▁fundó -10.294
+▁lectura -10.2941
+▁dependien -10.2941
+▁detenido -10.2941
+▁culto -10.2941
+▁promocion -10.2941
+▁mayordomo -10.2942
+▁amarilla -10.2942
+▁condición -10.2944
+▁rumor -10.2944
+▁moneda -10.2946
+▁teniente -10.2949
+▁sentir -10.2949
+▁ahorr -10.2949
+▁seguía -10.295
+▁caga -10.2962
+▁chicago -10.2966
+▁constitución -10.2967
+▁fría -10.2969
+▁buque -10.2976
+▁paloma -10.2977
+▁dividi -10.2982
+▁tendría -10.2989
+▁asesinato -10.299
+▁tranquila -10.2998
+▁hil -10.3002
+▁opt -10.3004
+▁juro -10.3005
+▁placa -10.3012
+▁vasco -10.3039
+▁peor -10.3063
+▁my -10.3069
+▁roto -10.3089
+▁preparado -10.3092
+▁estancia -10.3099
+▁adam -10.3111
+▁hill -10.3128
+marca -10.3151
+▁escapar -10.3155
+▁trasladado -10.3171
+▁cantar -10.318
+▁marcado -10.3192
+▁luch -10.3192
+▁robo -10.3196
+▁distribu -10.3203
+▁asoma -10.3211
+▁misterio -10.323
+▁actitud -10.3234
+▁ambienta -10.3234
+▁circuito -10.3234
+▁comportamiento -10.3234
+▁contexto -10.3234
+▁controversia -10.3234
+▁desaparecido -10.3234
+▁desconocida -10.3234
+▁embajador -10.3234
+▁fernández -10.3234
+▁hábito -10.3234
+▁israel -10.3234
+▁lópez -10.3234
+▁órdenes -10.3234
+▁alegría -10.3234
+▁moscú -10.3234
+▁oculta -10.3234
+▁belleza -10.3234
+▁cerebro -10.3234
+▁navarra -10.3234
+▁negocia -10.3234
+▁ofreci -10.3234
+▁describi -10.3234
+▁vigila -10.3234
+▁transmiti -10.3235
+▁humor -10.3235
+▁seminario -10.3235
+▁curva -10.3235
+▁poseía -10.3235
+▁reloj -10.3235
+▁cristiana -10.3236
+▁costera -10.3237
+▁emiti -10.3238
+▁administrativo -10.3239
+▁denominado -10.3239
+▁sureste -10.324
+▁distribuido -10.3244
+▁entidad -10.3245
+▁comentario -10.3246
+▁sucede -10.3247
+▁front -10.3249
+▁semifinales -10.3251
+▁reparti -10.3267
+▁muchacho -10.3267
+▁venir -10.3269
+movido -10.3274
+▁hiper -10.3284
+▁habitualmente -10.3287
+▁retirada -10.3294
+▁daba -10.3298
+▁borr -10.331
+▁bello -10.3313
+▁simplemente -10.3321
+▁edifica -10.3329
+▁aprovecha -10.3335
+ugh -10.3341
+▁expone -10.3349
+▁manga -10.3357
+▁comunidades -10.3373
+▁estudiante -10.3395
+años -10.341
+▁familiares -10.3411
+▁vete -10.3432
+▁martin -10.3434
+▁intent -10.3434
+cientos -10.3435
+▁capturado -10.3436
+▁trabajaba -10.3467
+▁cuida -10.3472
+▁preocupa -10.3528
+▁cumple -10.3537
+▁prestigio -10.3537
+▁amarillento -10.3537
+▁apasionad -10.3537
+▁contemporánea -10.3537
+▁círculo -10.3537
+▁específico -10.3537
+▁evolución -10.3537
+▁gracioso -10.3537
+▁húmeda -10.3537
+▁incluida -10.3537
+▁indonesia -10.3537
+▁ingeniería -10.3537
+▁montevideo -10.3537
+▁parcialmente -10.3537
+▁prefiere -10.3537
+▁propietario -10.3537
+▁rafael -10.3537
+▁universo -10.3537
+▁virtud -10.3537
+▁felices -10.3537
+▁rojizo -10.3537
+▁gabriel -10.3537
+▁típico -10.3537
+▁medieval -10.3537
+▁espalda -10.3537
+▁primario -10.3537
+▁alianza -10.3538
+▁huerto -10.3538
+scripción -10.3538
+▁vuelto -10.3538
+▁oferta -10.3538
+▁murieron -10.3539
+▁esquina -10.354
+▁jackson -10.354
+▁high -10.354
+▁fracaso -10.354
+▁reseña -10.354
+▁signo -10.3541
+▁reemplazado -10.3541
+▁asign -10.3542
+▁misterioso -10.3542
+▁juez -10.3544
+▁iniciativa -10.3544
+▁combati -10.3545
+▁agente -10.3548
+▁judía -10.3548
+▁espino -10.3551
+▁enviado -10.3552
+ística -10.3554
+▁mov -10.3555
+▁encend -10.356
+▁oposición -10.3566
+▁finaliza -10.3566
+▁feria -10.3581
+▁educado -10.3584
+▁regularmente -10.3587
+▁debate -10.3601
+▁gener -10.3613
+▁confia -10.3622
+▁califica -10.3625
+▁trenes -10.3627
+▁blo -10.3633
+▁presa -10.364
+cenas -10.3642
+ington -10.3669
+▁techo -10.3671
+▁declarado -10.3673
+▁mantuv -10.3679
+▁clasificado -10.3734
+▁parroquial -10.3739
+▁llamaba -10.3762
+▁jugado -10.3783
+venció -10.3792
+▁rápida -10.3819
+▁pint -10.3819
+▁brav -10.3827
+▁aproxima -10.385
+ö -10.385
+▁aficionado -10.385
+▁antigüedad -10.385
+▁autónoma -10.385
+▁bradomín -10.385
+▁circunstancia -10.385
+▁comandante -10.385
+▁cordillera -10.385
+▁envidia -10.385
+▁georgia -10.385
+▁metropolitana -10.385
+▁orígenes -10.385
+▁retablo -10.385
+▁naranja -10.385
+▁positivo -10.385
+▁preferido -10.385
+▁testigo -10.385
+▁raíz -10.385
+▁fragata -10.385
+▁prohibi -10.385
+▁audiencia -10.385
+▁someti -10.385
+▁promoción -10.385
+▁asesino -10.385
+▁innova -10.385
+▁mudanza -10.385
+▁campana -10.385
+▁empleo -10.3851
+▁ésta -10.3851
+▁perfil -10.3852
+▁embarc -10.3852
+▁saint -10.3853
+▁supone -10.3855
+▁juicio -10.3856
+▁nieto -10.3856
+sentimiento -10.3857
+▁salvaje -10.386
+▁retirarse -10.3862
+▁instalaciones -10.3863
+▁variada -10.3863
+▁naval -10.3864
+▁predominante -10.3867
+▁cabal -10.3871
+▁trae -10.3878
+▁pase -10.3889
+▁especializada -10.389
+▁arran -10.3891
+▁tranquilo -10.3892
+▁cobra -10.3893
+▁planea -10.3896
+▁recorrido -10.3898
+▁separada -10.3899
+▁querido -10.39
+▁encontraron -10.3907
+▁mozo -10.3909
+▁ubicación -10.391
+▁activ -10.3918
+▁clav -10.3928
+▁dividida -10.3929
+▁cumplir -10.3935
+▁cambiado -10.3974
+▁potencia -10.3975
+cillo -10.3984
+field -10.3985
+line -10.3989
+▁rem -10.3995
+▁promov -10.4007
+▁integrante -10.4009
+▁mark -10.4031
+▁abandonado -10.4033
+▁jord -10.404
+▁ech -10.4059
+▁regula -10.4072
+▁individual -10.4125
+▁educ -10.4127
+dista -10.4139
+▁disfruta -10.4151
+▁predomina -10.4155
+▁grab -10.4156
+incl -10.4164
+▁perteneci -10.4169
+▁fui -10.4169
+▁apellido -10.4172
+▁arqueológico -10.4172
+▁automóvil -10.4172
+▁bretaña -10.4172
+▁característico -10.4172
+▁constitucional -10.4172
+▁cálido -10.4172
+▁desconocido -10.4172
+▁dimensiones -10.4172
+▁dispositivo -10.4172
+▁disputó -10.4172
+▁ejecutivo -10.4172
+▁embajada -10.4172
+▁esclavo -10.4172
+▁incluir -10.4172
+▁intelectual -10.4172
+▁longitud -10.4172
+▁ocurrió -10.4172
+▁patrimonio -10.4172
+▁península -10.4172
+▁pálido -10.4172
+▁sorprende -10.4172
+▁sorpresa -10.4172
+▁utilizó -10.4172
+▁configura -10.4172
+▁orgullo -10.4172
+▁refugio -10.4172
+▁reacción -10.4172
+▁exhibi -10.4172
+▁ocasionalmente -10.4172
+▁sencilla -10.4172
+▁previamente -10.4173
+▁lentamente -10.4173
+▁recinto -10.4173
+▁carolina -10.4173
+▁abuela -10.4173
+▁cierre -10.4173
+▁brown -10.4173
+▁excava -10.4173
+▁otorga -10.4173
+▁sustancia -10.4173
+▁espejo -10.4173
+▁novio -10.4173
+▁obrero -10.4173
+▁cuánto -10.4174
+glo -10.4174
+▁destinado -10.4174
+▁cancela -10.4174
+▁licenciatura -10.4175
+▁paralela -10.4177
+▁entorno -10.4178
+▁pobl -10.4179
+▁preocup -10.4182
+vés -10.4183
+▁magos -10.4183
+▁obtenido -10.4184
+▁significativo -10.4184
+▁renov -10.4192
+▁matorrales -10.4192
+house -10.4208
+▁verá -10.4212
+▁toque -10.4214
+▁actuar -10.4228
+stein -10.4232
+▁lector -10.4238
+▁clar -10.4244
+▁conservación -10.4257
+▁variante -10.4279
+▁lisa -10.4326
+▁traslada -10.433
+▁beca -10.4333
+▁amante -10.4337
+▁cima -10.4347
+▁dividido -10.4373
+▁visitante -10.4375
+▁miel -10.4383
+▁enfermo -10.4411
+jón -10.4442
+▁rama -10.445
+▁bob -10.4479
+▁subir -10.4493
+▁prometi -10.4493
+▁pudie -10.4499
+▁perdon -10.45
+▁atletismo -10.4506
+▁básico -10.4506
+▁cañón -10.4506
+▁destrucción -10.4506
+▁evidente -10.4506
+▁felicidad -10.4506
+▁gestión -10.4506
+▁izquierdo -10.4506
+▁rechazó -10.4506
+▁renacentista -10.4506
+▁responsabilidad -10.4506
+▁subespecie -10.4506
+▁ucrania -10.4506
+▁voluntad -10.4506
+▁exacta -10.4506
+▁geográfica -10.4506
+▁guzmán -10.4506
+▁sábado -10.4506
+▁temblor -10.4506
+▁atravesa -10.4506
+▁continente -10.4506
+▁dormido -10.4506
+▁respectiva -10.4506
+▁proclam -10.4506
+▁recomend -10.4506
+▁jazz -10.4506
+▁hombro -10.4506
+▁instante -10.4506
+▁discoteca -10.4506
+▁pastilla -10.4506
+▁rubia -10.4506
+▁charles -10.4506
+▁maestra -10.4506
+▁distingui -10.4506
+▁vestido -10.4507
+▁domin -10.4507
+▁dispuesto -10.4507
+▁intereses -10.4507
+▁pensamiento -10.4507
+▁crisis -10.4508
+▁perteneciente -10.4509
+▁tonto -10.4511
+▁cristo -10.4511
+▁castilla -10.4512
+▁entiendo -10.4513
+▁apaga -10.4514
+▁hueso -10.4515
+tropel -10.4515
+▁ancho -10.4524
+▁green -10.4524
+▁receta -10.4524
+▁padec -10.4536
+▁jardines -10.4538
+▁huir -10.454
+▁jack -10.454
+▁miró -10.4544
+▁margen -10.4548
+▁registrado -10.4549
+andr -10.4556
+▁agita -10.4565
+▁docente -10.4566
+▁italiana -10.4569
+centr -10.4571
+▁raza -10.4599
+ndar -10.4602
+nacida -10.4637
+▁localizada -10.4642
+▁plate -10.4643
+▁pez -10.4651
+▁corro -10.4662
+▁realista -10.4666
+▁conservador -10.4674
+▁limitada -10.4711
+▁vini -10.4722
+▁exposi -10.4724
+▁edit -10.4731
+▁distinta -10.477
+▁estuvi -10.4779
+▁recogi -10.4791
+▁consider -10.481
+▁abundan -10.4818
+▁colonial -10.4821
+▁desespera -10.4848
+▁caballerango -10.4851
+▁centrocampista -10.4851
+▁competiciones -10.4851
+▁demócrata -10.4851
+▁desembarc -10.4851
+▁guatemala -10.4851
+▁homónima -10.4851
+▁incluía -10.4851
+▁manuscrito -10.4851
+▁territorial -10.4851
+▁alejandro -10.4851
+▁borneo -10.4851
+▁cáncer -10.4851
+▁interrumpi -10.4851
+▁respondi -10.4851
+▁lágrimas -10.4851
+▁oxford -10.4851
+▁zelanda -10.4851
+▁inquieta -10.4851
+▁august -10.4851
+▁franceses -10.4851
+▁cumpleaños -10.4851
+▁muralla -10.4851
+▁golfo -10.4851
+▁chaleco -10.4851
+▁frank -10.4851
+▁alfredo -10.4851
+▁negó -10.4851
+ão -10.4852
+▁concejo -10.4852
+▁ejercicio -10.4852
+▁manual -10.4852
+▁protector -10.4852
+▁actuaciones -10.4854
+▁comunista -10.4854
+▁remonta -10.4854
+▁inunda -10.4855
+▁tortura -10.4859
+▁polvo -10.4862
+▁tigre -10.4862
+▁expedición -10.4865
+▁ventaja -10.4867
+zuel -10.4872
+▁publicidad -10.4876
+▁cobre -10.4877
+▁personalidad -10.4878
+▁definitivamente -10.4878
+▁enlace -10.4886
+▁raya -10.4909
+▁pudiera -10.492
+▁alimentación -10.4926
+▁nudos -10.4927
+▁colga -10.4936
+▁encarn -10.4938
+▁asom -10.4944
+▁trono -10.496
+▁mediano -10.4985
+▁mapa -10.4998
+▁descon -10.5
+ward -10.5008
+▁potencial -10.507
+▁señala -10.5079
+▁radic -10.5086
+▁barro -10.5091
+▁alguno -10.5105
+▁recuper -10.5122
+volver -10.513
+▁directa -10.5177
+▁combustible -10.5208
+▁contemporáneo -10.5208
+▁continental -10.5208
+▁cráneo -10.5208
+▁diversidad -10.5208
+▁ejecución -10.5208
+▁financiero -10.5208
+▁infraestructura -10.5208
+▁sobrevivir -10.5208
+▁vicepresidente -10.5208
+▁acostumbra -10.5208
+▁cohete -10.5208
+▁colectiva -10.5208
+▁expresión -10.5208
+▁extinción -10.5208
+▁salamanca -10.5208
+▁santuario -10.5208
+▁conducir -10.5208
+▁anuncia -10.5208
+▁bienvenido -10.5208
+▁trofeo -10.5208
+▁patrocin -10.5208
+▁inestable -10.5208
+▁protegido -10.5208
+▁sudoeste -10.5208
+▁estrecho -10.5208
+▁manzana -10.5208
+▁setenta -10.5208
+▁descrito -10.5208
+▁homosexual -10.5208
+▁fauna -10.5208
+▁jaime -10.5208
+▁solución -10.5208
+▁envenena -10.5208
+▁descarga -10.5209
+▁mierda -10.5209
+▁entender -10.521
+▁infini -10.521
+▁sabana -10.521
+▁atrapa -10.5211
+▁pulsa -10.5211
+▁narrativa -10.5211
+▁crónica -10.5212
+▁consume -10.5213
+▁patrici -10.5216
+▁flujo -10.5216
+▁cayó -10.5216
+▁lanzada -10.5218
+▁mostrar -10.5221
+▁sujeta -10.5222
+▁producciones -10.5224
+▁firme -10.5226
+▁perfectamente -10.5228
+▁lento -10.5228
+▁acero -10.5231
+▁reduc -10.5233
+▁ficha -10.5242
+▁gem -10.5243
+▁empresario -10.5248
+▁carmen -10.525
+▁traducción -10.5251
+▁goza -10.5264
+▁modern -10.5301
+▁toros -10.5312
+▁solía -10.5383
+▁baila -10.5393
+▁iris -10.541
+▁mineral -10.542
+▁sabr -10.5436
+▁juli -10.5444
+▁island -10.5463
+▁nombr -10.548
+▁crítico -10.5486
+▁desplaza -10.5496
+▁creativ -10.5509
+▁enviar -10.5546
+▁probable -10.5548
+▁design -10.5565
+▁andalucía -10.5578
+▁arrecife -10.5578
+▁asentamiento -10.5578
+▁cadáver -10.5578
+▁costumbre -10.5578
+▁descendencia -10.5578
+▁embarazada -10.5578
+▁ernesto -10.5578
+▁ingrediente -10.5578
+▁maestría -10.5578
+▁mamíferos -10.5578
+▁nicaragua -10.5578
+▁opinión -10.5578
+▁púrpura -10.5578
+▁sacristía -10.5578
+▁videoclip -10.5578
+▁acceder -10.5578
+▁intenso -10.5578
+▁javier -10.5578
+▁maniobra -10.5578
+▁prolonga -10.5578
+▁resistente -10.5578
+▁destruida -10.5578
+▁tixul -10.5578
+▁renacimiento -10.5578
+▁sicilia -10.5578
+▁veinti -10.5578
+▁articula -10.5578
+▁noruega -10.5578
+▁aterriza -10.5578
+▁mínimo -10.5578
+▁melodía -10.5578
+▁black -10.5578
+▁financia -10.5578
+▁arriesg -10.5578
+▁propone -10.5578
+▁mitología -10.5578
+▁trece -10.5579
+▁patrón -10.5579
+▁elegir -10.5579
+▁ausencia -10.5579
+▁disney -10.5579
+▁vencido -10.5579
+▁ornamental -10.5579
+▁vertical -10.5579
+▁sentencia -10.5579
+▁oliva -10.558
+▁emite -10.558
+▁hueco -10.558
+▁whit -10.558
+▁adiós -10.558
+▁afili -10.5581
+▁descans -10.5582
+▁latín -10.5584
+▁enfada -10.5585
+▁warne -10.5592
+▁misiones -10.5597
+▁predic -10.5601
+▁grabaciones -10.5603
+▁instalación -10.561
+▁salvo -10.5614
+▁madura -10.5618
+▁aplicación -10.5618
+▁place -10.5625
+▁decoración -10.5633
+▁estándar -10.5639
+nutri -10.5639
+▁querer -10.564
+▁encontrarse -10.5654
+▁hice -10.5656
+pusieron -10.5657
+▁funcionario -10.5669
+▁sob -10.5674
+▁escapa -10.5685
+ducto -10.5701
+▁aporta -10.5716
+▁cristi -10.5741
+▁interno -10.5843
+▁hund -10.5845
+▁grita -10.5884
+▁definitiva -10.594
+▁culpa -10.5957
+ógrafo -10.5963
+▁aceptó -10.5963
+▁aerolínea -10.5963
+▁azúcar -10.5963
+▁canadiense -10.5963
+▁cápsula -10.5963
+▁evolucion -10.5963
+▁grijalba -10.5963
+▁gómez -10.5963
+▁homónimo -10.5963
+▁ingleses -10.5963
+▁inteligente -10.5963
+▁nicolás -10.5963
+▁nocturno -10.5963
+▁sepultado -10.5963
+▁sobrevivió -10.5963
+▁triunfo -10.5963
+▁universitaria -10.5963
+▁aragón -10.5963
+▁dominicana -10.5963
+▁finlandia -10.5963
+▁honduras -10.5963
+▁monstruos -10.5963
+▁sucedido -10.5963
+▁violenta -10.5963
+▁cervantes -10.5963
+▁imparti -10.5963
+scripciones -10.5963
+▁laboratorio -10.5963
+▁mayoritaria -10.5963
+▁raíces -10.5963
+▁bachillerato -10.5963
+▁wilson -10.5963
+▁gótico -10.5963
+▁polonia -10.5963
+▁smith -10.5963
+▁forestal -10.5963
+▁comerciante -10.5963
+▁rescate -10.5963
+▁regimiento -10.5963
+▁soledad -10.5963
+▁catalina -10.5963
+▁fatiga -10.5963
+▁comparte -10.5963
+▁arrastra -10.5963
+ástica -10.5963
+▁arrestado -10.5963
+▁introduj -10.5963
+▁consumo -10.5963
+▁piscina -10.5963
+▁cercanía -10.5963
+▁aviones -10.5964
+▁phil -10.5964
+▁sudeste -10.5964
+▁transmisión -10.5965
+▁sesenta -10.5965
+▁rocosa -10.5965
+▁cerrada -10.5965
+lusiones -10.5966
+▁oficio -10.5967
+▁nominado -10.5969
+▁culpable -10.5969
+▁dependencia -10.5969
+▁volante -10.5972
+▁kenia -10.5972
+▁llano -10.5974
+▁campus -10.5975
+▁reposa -10.5983
+▁volvie -10.5987
+town -10.5987
+▁atenas -10.5988
+▁dominante -10.5993
+▁drama -10.5994
+▁milán -10.5996
+▁uh -10.6
+▁totalidad -10.6006
+▁cuya -10.6011
+▁crono -10.6021
+▁trigo -10.6022
+▁pleno -10.6065
+▁levant -10.6111
+▁ál -10.6144
+▁milit -10.6149
+▁marea -10.6202
+luci -10.6285
+▁húmedo -10.6302
+▁american -10.6349
+▁protagoniz -10.6354
+▁disminu -10.6359
+güenza -10.6363
+▁accesible -10.6363
+▁alexander -10.6363
+▁aparecido -10.6363
+▁atractivo -10.6363
+▁atraviesa -10.6363
+▁botánico -10.6363
+▁caña -10.6363
+▁chocolate -10.6363
+▁cinematográfica -10.6363
+▁dieciocho -10.6363
+▁dinastía -10.6363
+▁diplomático -10.6363
+▁facilita -10.6363
+▁fórmula -10.6363
+▁fósiles -10.6363
+▁inmediato -10.6363
+▁inteligencia -10.6363
+▁perspectiva -10.6363
+▁pálida -10.6363
+▁simboliza -10.6363
+▁coordina -10.6363
+▁diablo -10.6363
+▁dispersa -10.6363
+▁equivoca -10.6363
+▁parecido -10.6363
+▁rugby -10.6363
+▁intención -10.6363
+scendió -10.6363
+▁bautista -10.6363
+▁catalán -10.6363
+▁preciso -10.6363
+▁jinete -10.6363
+▁textil -10.6363
+▁criatura -10.6363
+▁protegida -10.6363
+▁existía -10.6363
+▁vaquero -10.6363
+▁serpiente -10.6363
+▁útil -10.6363
+▁maíz -10.6363
+▁admira -10.6363
+▁electro -10.6363
+▁experto -10.6363
+▁entrenamiento -10.6363
+▁ribera -10.6363
+▁grueso -10.6363
+▁brigada -10.6363
+▁compite -10.6364
+▁curiosa -10.6364
+▁arroja -10.6364
+▁sucesión -10.6364
+▁romance -10.6364
+▁caudal -10.6364
+▁tolera -10.6364
+▁malasia -10.6364
+▁tocó -10.6364
+▁tejido -10.6365
+▁vendido -10.6365
+▁gravedad -10.6367
+▁fresca -10.6367
+▁descubrimiento -10.6369
+▁dorada -10.6371
+▁detener -10.6371
+▁rasgo -10.6372
+▁participantes -10.6375
+▁fuimos -10.6375
+▁silencioso -10.6378
+▁peligroso -10.6379
+▁diseñada -10.6381
+▁organizaciones -10.6395
+▁imita -10.6398
+▁grasa -10.6411
+▁uva -10.6412
+▁listo -10.6418
+▁north -10.6429
+iéndome -10.6431
+▁saldr -10.6478
+eiro -10.6481
+cultura -10.6591
+lecciones -10.6599
+▁ampli -10.6625
+▁roll -10.666
+▁bast -10.676
+pósito -10.6775
+▁absoluto -10.6779
+▁bogotá -10.6779
+▁cristianismo -10.6779
+▁depresión -10.6779
+▁descenso -10.6779
+▁diplomática -10.6779
+▁disfraz -10.6779
+▁envuelto -10.6779
+▁escultor -10.6779
+▁estrategia -10.6779
+▁horizontal -10.6779
+▁horizonte -10.6779
+▁infierno -10.6779
+▁joseph -10.6779
+▁oscuridad -10.6779
+▁procedimiento -10.6779
+▁protocolo -10.6779
+▁sebastián -10.6779
+▁síntoma -10.6779
+▁universal -10.6779
+▁veinticuatro -10.6779
+▁órbita -10.6779
+▁absoluta -10.6779
+▁cenizas -10.6779
+▁computador -10.6779
+▁cuartel -10.6779
+▁rebelión -10.6779
+▁satisfac -10.6779
+▁argumento -10.6779
+▁compromet -10.6779
+▁limpio -10.6779
+▁madruga -10.6779
+▁montañosa -10.6779
+▁málaga -10.6779
+▁repertorio -10.6779
+▁campamento -10.6779
+▁supongo -10.6779
+▁conexión -10.6779
+dolfo -10.6779
+▁estadística -10.6779
+▁responde -10.6779
+▁turística -10.6779
+▁cruel -10.6779
+▁avance -10.6779
+▁recordar -10.678
+▁nigeria -10.678
+▁aborda -10.678
+▁diciendo -10.678
+▁estremecimiento -10.678
+▁recopila -10.678
+▁millón -10.678
+onomía -10.678
+▁turistas -10.678
+▁defender -10.678
+▁gruesa -10.678
+▁hielo -10.678
+▁dictadura -10.6781
+▁royal -10.6781
+▁emergencia -10.6781
+▁pretende -10.6781
+▁esquema -10.6782
+▁huyó -10.6783
+▁reducida -10.6784
+▁reducir -10.6784
+▁robust -10.6785
+▁plumaje -10.6787
+▁cuyo -10.6787
+▁tarda -10.679
+▁vecina -10.679
+▁chaval -10.6792
+▁helado -10.6793
+▁yacimiento -10.6794
+▁crucero -10.6799
+▁falsa -10.68
+▁complet -10.6803
+▁designado -10.6804
+▁traducido -10.6806
+▁amito -10.6812
+ish -10.6817
+▁sorprendi -10.682
+▁seleccionado -10.6826
+▁dirigente -10.6826
+▁incluy -10.6846
+▁minas -10.6861
+▁seleccion -10.7061
+▁diseñ -10.7071
+▁cli -10.711
+▁comercializa -10.7117
+vuelta -10.7136
+▁recrea -10.7157
+lusión -10.7169
+▁autoriza -10.72
+▁bolívar -10.7214
+▁botánica -10.7214
+▁cambridge -10.7214
+▁camiseta -10.7214
+▁correspondiente -10.7214
+▁electrónica -10.7214
+▁electrónico -10.7214
+▁extranjera -10.7214
+▁grisáceo -10.7214
+▁hermandad -10.7214
+▁hollywood -10.7214
+▁joaquín -10.7214
+▁marruecos -10.7214
+▁mediocampista -10.7214
+▁médico -10.7214
+▁nuclear -10.7214
+▁párpados -10.7214
+▁secuestra -10.7214
+▁software -10.7214
+▁tiburones -10.7214
+▁valladolid -10.7214
+▁yucatán -10.7214
+▁ábside -10.7214
+▁acredita -10.7214
+▁altitud -10.7214
+▁columbia -10.7214
+▁conmemora -10.7214
+▁consolida -10.7214
+▁inaugura -10.7214
+▁líquido -10.7214
+▁national -10.7214
+▁ochenta -10.7214
+▁privilegi -10.7214
+▁romántico -10.7214
+▁triunfa -10.7214
+▁campesino -10.7214
+▁castaño -10.7214
+▁cerveza -10.7214
+▁humanidad -10.7214
+▁exitosa -10.7214
+▁solitaria -10.7214
+▁disparo -10.7214
+scendido -10.7214
+▁terrible -10.7214
+▁precedid -10.7214
+▁apoyó -10.7214
+▁henry -10.7214
+▁pionero -10.7214
+▁increment -10.7214
+▁cumbre -10.7214
+▁centenar -10.7215
+sistieron -10.7215
+▁deposita -10.7215
+▁oveja -10.7215
+▁mascota -10.7215
+▁proporci -10.7215
+▁congrega -10.7215
+▁gemelo -10.7215
+▁bebida -10.7215
+▁divina -10.7215
+▁toledo -10.7215
+▁edward -10.7215
+▁herencia -10.7215
+▁conozco -10.7215
+▁correcta -10.7215
+▁arroz -10.7215
+▁penetr -10.7215
+▁rioja -10.7215
+▁ajusta -10.7216
+▁merece -10.7216
+▁insult -10.7216
+▁serbia -10.7216
+▁rastro -10.7217
+▁tomás -10.7217
+▁conocía -10.7217
+▁suelta -10.7217
+▁escocia -10.7219
+▁plateados -10.7219
+▁resolver -10.7221
+▁inclina -10.7222
+▁conductor -10.7223
+▁conocieron -10.7223
+▁habana -10.7224
+▁gordo -10.7225
+▁venganza -10.7225
+▁tony -10.7226
+▁reporta -10.7228
+▁funcionamiento -10.7228
+▁gru -10.7231
+▁michel -10.7233
+▁plantea -10.7236
+▁fibra -10.724
+▁mosca -10.7247
+▁prat -10.7248
+▁electo -10.7251
+▁probar -10.7269
+▁elige -10.7294
+▁afro -10.7424
+▁pregunt -10.7497
+▁licenci -10.7606
+▁confund -10.7662
+poniendo -10.7665
+▁afect -10.7667
+▁afluente -10.7669
+▁aparecieron -10.7669
+▁arquitectónico -10.7669
+▁biblia -10.7669
+▁convertiría -10.7669
+▁cuestión -10.7669
+▁cádiz -10.7669
+▁difusión -10.7669
+▁discípulo -10.7669
+▁financiera -10.7669
+▁hazaña -10.7669
+▁impreso -10.7669
+▁japonesa -10.7669
+▁legislatura -10.7669
+▁madagascar -10.7669
+▁magnífic -10.7669
+▁modalidad -10.7669
+▁muñecas -10.7669
+▁nocturna -10.7669
+▁octavo -10.7669
+▁paralímpico -10.7669
+▁prominente -10.7669
+▁pronuncia -10.7669
+▁taiwán -10.7669
+▁virginia -10.7669
+▁llanura -10.7669
+▁lorenzo -10.7669
+▁semejante -10.7669
+▁acogida -10.7669
+▁asumió -10.7669
+▁garganta -10.7669
+▁reflexi -10.7669
+▁scott -10.7669
+▁tanzania -10.7669
+▁vientre -10.7669
+▁concejal -10.7669
+▁externa -10.7669
+▁despacio -10.7669
+▁puntua -10.7669
+▁redactor -10.7669
+▁reúne -10.7669
+▁hoguera -10.7669
+▁café -10.7669
+▁teclado -10.7669
+▁traición -10.7669
+▁venecia -10.7669
+▁simón -10.767
+▁relaja -10.767
+▁certifica -10.767
+terapia -10.767
+▁residi -10.767
+▁moreno -10.767
+▁magia -10.767
+▁pidió -10.767
+▁coloniza -10.767
+▁favorable -10.7671
+▁meteor -10.7671
+▁lineal -10.7672
+▁cuesta -10.7672
+▁suspendi -10.7672
+ándola -10.7672
+▁miami -10.7672
+▁distingue -10.7673
+▁campanario -10.7675
+▁confundir -10.7676
+▁honra -10.7676
+▁reuniones -10.7677
+▁conservatorio -10.7677
+▁distintivo -10.7677
+▁reducido -10.7678
+▁compuso -10.7678
+▁prado -10.768
+▁postura -10.7684
+▁peligrosa -10.7689
+▁impone -10.7768
+▁alinea -10.7797
+▁diaria -10.7802
+▁motiv -10.7819
+▁extremad -10.791
+▁reconoci -10.7943
+▁perdie -10.8025
+▁enfrent -10.8041
+▁impresi -10.8082
+cuerdo -10.8089
+▁desgracia -10.8143
+▁adquiere -10.8145
+▁agustín -10.8145
+▁ajedrez -10.8145
+▁beneficio -10.8145
+▁catálogo -10.8145
+▁coincide -10.8145
+▁confederación -10.8145
+▁consejero -10.8145
+▁crustáceos -10.8145
+▁cárcel -10.8145
+▁emisora -10.8145
+▁espiritual -10.8145
+▁hernández -10.8145
+▁ilustre -10.8145
+▁lingüística -10.8145
+▁lógica -10.8145
+▁mampostería -10.8145
+▁masculino -10.8145
+▁mecánica -10.8145
+▁múnich -10.8145
+▁ocurrir -10.8145
+▁partícula -10.8145
+▁profundidad -10.8145
+▁reconstrui -10.8145
+▁sospecha -10.8145
+▁supremo -10.8145
+▁tráfico -10.8145
+▁university -10.8145
+▁veinticinco -10.8145
+▁afuera -10.8145
+▁comestible -10.8145
+▁infantería -10.8145
+▁secuencia -10.8145
+▁casual -10.8145
+▁educativa -10.8145
+▁educativo -10.8145
+▁exitoso -10.8145
+▁renombr -10.8145
+▁reputación -10.8145
+▁respeto -10.8145
+▁suspiró -10.8145
+▁arabia -10.8145
+▁compleja -10.8145
+▁fiebre -10.8145
+▁simpson -10.8145
+▁aparato -10.8145
+▁youtube -10.8145
+▁pastoral -10.8145
+▁singular -10.8145
+▁himno -10.8145
+▁moderada -10.8145
+▁pidiendo -10.8145
+▁pizza -10.8145
+▁vidrio -10.8145
+▁soporte -10.8145
+▁funeral -10.8145
+▁botella -10.8145
+▁delicioso -10.8145
+▁contenía -10.8145
+▁alonso -10.8145
+▁leonor -10.8145
+▁franja -10.8145
+▁rodeada -10.8146
+▁abrazo -10.8146
+▁persiste -10.8146
+▁tesoro -10.8146
+▁uau -10.8146
+▁terror -10.8146
+▁seguida -10.8146
+▁rescata -10.8147
+▁túnel -10.8147
+▁quieto -10.8147
+▁bautiz -10.8148
+▁comenta -10.8149
+▁fusil -10.8149
+▁incursi -10.8149
+rgenes -10.8149
+▁cerró -10.815
+▁unirse -10.815
+▁castro -10.8151
+▁compró -10.8151
+▁exporta -10.8151
+▁caracol -10.8152
+▁caído -10.8152
+▁south -10.8152
+▁percib -10.8153
+▁compiti -10.8154
+▁sagrado -10.8154
+▁analiza -10.8154
+▁claridad -10.8155
+▁florencia -10.8155
+▁turno -10.8155
+▁muelle -10.8156
+▁conversa -10.8158
+▁dicta -10.8163
+▁cartel -10.8164
+▁remata -10.8167
+▁semilla -10.8171
+▁premia -10.8179
+▁ranch -10.8182
+▁baño -10.8185
+▁beat -10.8188
+▁delgado -10.8208
+▁stone -10.8265
+▁lanzar -10.8286
+▁instal -10.8328
+▁qu -10.8388
+▁derram -10.8414
+▁azu -10.8461
+estrati -10.8465
+▁sobrevive -10.8466
+ábamos -10.8473
+▁diputa -10.8519
+▁escrib -10.855
+▁estrat -10.8571
+▁recorri -10.8571
+▁jardin -10.8597
+▁expedi -10.8625
+▁actúa -10.8645
+▁agujero -10.8645
+▁aluminio -10.8645
+▁balneario -10.8645
+▁compromiso -10.8645
+▁construcciones -10.8645
+▁convencional -10.8645
+▁cortometraje -10.8645
+▁disculpa -10.8645
+▁estricta -10.8645
+▁hipótesis -10.8645
+▁ignacio -10.8645
+▁jurídica -10.8645
+▁legislativo -10.8645
+▁marítimo -10.8645
+▁multiplica -10.8645
+▁mármol -10.8645
+▁pacífica -10.8645
+▁pontificia -10.8645
+▁portuguesa -10.8645
+▁sostiene -10.8645
+▁tripulación -10.8645
+▁acumula -10.8645
+▁agradable -10.8645
+▁bernardo -10.8645
+▁inusual -10.8645
+▁manipula -10.8645
+▁tragedia -10.8645
+▁ciclista -10.8645
+▁contienda -10.8645
+▁especifica -10.8645
+▁eventual -10.8645
+▁proveniente -10.8645
+▁secuela -10.8645
+▁alejó -10.8645
+▁entendido -10.8645
+▁caribe -10.8645
+▁hungría -10.8645
+▁jarocho -10.8645
+▁acuático -10.8645
+▁exilio -10.8645
+▁fortuna -10.8645
+▁estudiantil -10.8645
+▁denuncia -10.8645
+▁elegida -10.8645
+▁negativas -10.8645
+▁andrés -10.8645
+▁examen -10.8645
+▁asiático -10.8645
+▁encanto -10.8645
+▁conciencia -10.8645
+▁maneja -10.8646
+▁cicatri -10.8646
+▁punk -10.8646
+▁propaga -10.8646
+▁tiende -10.8646
+▁lejana -10.8646
+▁derrotó -10.8646
+▁cubrir -10.8647
+▁perdida -10.8647
+▁preveni -10.8648
+▁muerta -10.8648
+▁consola -10.8648
+▁virus -10.8648
+▁aumento -10.8649
+▁creía -10.865
+▁canarias -10.865
+▁discográfica -10.8652
+▁gentil -10.8653
+▁delegado -10.8654
+▁define -10.8656
+▁logro -10.8656
+▁mantenido -10.8657
+▁infan -10.8658
+▁invadi -10.866
+▁contesta -10.8671
+▁acerta -10.8672
+▁envía -10.8679
+▁camil -10.8729
+vivieron -10.8733
+▁veneno -10.8735
+▁correcto -10.875
+cogido -10.8766
+▁celebr -10.9065
+▁sacerdo -10.9143
+ándole -10.9153
+ō -10.9171
+▁adolescencia -10.9171
+▁afganistán -10.9171
+▁africano -10.9171
+▁arquidiócesis -10.9171
+▁artillería -10.9171
+▁athletic -10.9171
+▁atribuye -10.9171
+▁bermúdez -10.9171
+▁bicicleta -10.9171
+▁caníbales -10.9171
+▁concluye -10.9171
+▁croacia -10.9171
+▁cúpula -10.9171
+▁democracia -10.9171
+▁depredador -10.9171
+▁despedida -10.9171
+▁despedido -10.9171
+▁despierta -10.9171
+▁ecosistema -10.9171
+▁encarcelado -10.9171
+▁enseguida -10.9171
+▁espectador -10.9171
+▁feminista -10.9171
+▁fructifica -10.9171
+▁garantiza -10.9171
+▁global -10.9171
+▁harvard -10.9171
+▁herbácea -10.9171
+▁jesuita -10.9171
+▁magdalena -10.9171
+▁massachusetts -10.9171
+▁mecánico -10.9171
+▁presupuesto -10.9171
+▁romántica -10.9171
+▁secuestro -10.9171
+▁silvestre -10.9171
+▁temblaban -10.9171
+▁transparente -10.9171
+▁violento -10.9171
+▁diploma -10.9171
+▁incendio -10.9171
+▁orientada -10.9171
+▁proteínas -10.9171
+▁taylor -10.9171
+▁caballería -10.9171
+▁clínica -10.9171
+▁escuche -10.9171
+▁aliento -10.9171
+▁asistir -10.9171
+▁defensiv -10.9171
+▁enfoque -10.9171
+▁fracasó -10.9171
+▁rendimiento -10.9171
+▁peruano -10.9171
+▁sugiere -10.9171
+▁escolta -10.9171
+▁puebla -10.9171
+▁viernes -10.9171
+▁ballet -10.9171
+▁tokio -10.9171
+▁transición -10.9171
+▁atleta -10.9171
+▁grifo -10.9171
+▁pesadilla -10.9171
+▁angola -10.9171
+▁evalu -10.9171
+▁jersey -10.9171
+▁limón -10.9172
+▁oxida -10.9172
+▁toronto -10.9172
+▁aparente -10.9172
+▁rodaje -10.9172
+▁cardenal -10.9172
+▁fracasa -10.9172
+▁llave -10.9172
+▁ausente -10.9173
+▁sensación -10.9173
+▁duelo -10.9173
+▁inverna -10.9173
+▁excita -10.9174
+▁sequía -10.9176
+▁preparatoria -10.9177
+▁caracas -10.9177
+▁añadió -10.9178
+▁asume -10.9178
+▁paralelo -10.9178
+▁suelto -10.9179
+▁confesa -10.918
+▁sexta -10.918
+▁crecido -10.9182
+▁mental -10.9183
+▁carbon -10.9186
+▁simon -10.9195
+▁revisa -10.9195
+▁transform -10.9246
+▁reaccion -10.9271
+▁tracto -10.9274
+plazamiento -10.9291
+sorte -10.9319
+▁posesi -10.9339
+▁east -10.9377
+▁acompañ -10.9425
+▁denomin -10.9571
+▁actualiza -10.9619
+siguiente -10.9649
+▁subtropical -10.9679
+▁perfecta -10.9696
+▁acantilad -10.9727
+▁almirante -10.9727
+▁aventurero -10.9727
+▁birmania -10.9727
+▁caserío -10.9727
+▁cocoteros -10.9727
+▁consecutivo -10.9727
+▁continuo -10.9727
+▁custodia -10.9727
+▁diminut -10.9727
+▁discurso -10.9727
+▁discuti -10.9727
+▁efectiva -10.9727
+▁egipcio -10.9727
+▁elizabeth -10.9727
+▁episcopal -10.9727
+▁grammy -10.9727
+▁guadalajara -10.9727
+▁guardián -10.9727
+▁hormigón -10.9727
+▁ignora -10.9727
+▁institucional -10.9727
+▁invertebrados -10.9727
+▁mandíbula -10.9727
+▁mozambique -10.9727
+▁ocurra -10.9727
+▁proyecta -10.9727
+▁receptor -10.9727
+▁repercusi -10.9727
+▁simultáneamente -10.9727
+▁voluptuos -10.9727
+▁volúmenes -10.9727
+▁aprendizaje -10.9727
+▁astrónomo -10.9727
+▁excepcional -10.9727
+▁expansión -10.9727
+▁exótica -10.9727
+▁gratuito -10.9727
+▁oculto -10.9727
+▁álvarez -10.9727
+▁francisca -10.9727
+▁margarita -10.9727
+▁otomano -10.9727
+▁recluta -10.9727
+▁geografía -10.9727
+▁paciencia -10.9727
+▁pascua -10.9727
+▁pendiente -10.9727
+▁compagin -10.9727
+▁consuelo -10.9727
+▁obedec -10.9727
+▁rodrigo -10.9727
+▁apartamento -10.9727
+▁asteroid -10.9727
+▁textura -10.9727
+▁peruana -10.9727
+▁street -10.9727
+▁carbón -10.9727
+▁destruido -10.9727
+▁produjer -10.9727
+▁concurr -10.9727
+▁blue -10.9727
+▁ortiz -10.9727
+▁monarca -10.9727
+▁niebla -10.9727
+▁piratas -10.9727
+▁provenía -10.9727
+▁ciudadanía -10.9727
+▁reforz -10.9727
+▁eligió -10.9727
+▁supervivencia -10.9728
+▁consulta -10.9728
+▁sacud -10.9728
+▁mantenimiento -10.9728
+▁desnuda -10.9728
+▁simula -10.9728
+▁dispuesta -10.9729
+▁redondeada -10.9729
+▁polaco -10.9729
+▁cascada -10.9729
+▁vencedor -10.9729
+scender -10.973
+▁cresta -10.973
+rtesanía -10.973
+▁detecta -10.973
+▁socorr -10.9732
+▁patata -10.9733
+▁díg -10.9735
+▁tinta -10.9735
+▁harina -10.9737
+▁convence -10.9738
+▁ocean -10.9738
+▁turco -10.9741
+▁anciano -10.9742
+▁comuni -10.9755
+▁sopl -10.9756
+▁gesto -10.9758
+presiones -10.9768
+conduct -10.978
+ratón -10.986
+aúl -10.988
+plicó -10.9962
+tividad -11.0065
+tendida -11.0174
+ísimo -11.0312
+lítica -11.0313
+▁emerge -11.0314
+magnetismo -11.0315
+▁adyacente -11.0315
+▁anarquista -11.0315
+▁andaluza -11.0315
+▁arqueológica -11.0315
+▁bacteria -11.0315
+▁bulgaria -11.0315
+▁candidatura -11.0315
+▁cinematográfico -11.0315
+▁cirugía -11.0315
+▁compatible -11.0315
+▁composiciones -11.0315
+▁cooperativ -11.0315
+▁cristóbal -11.0315
+▁diagnóstico -11.0315
+▁diálogo -11.0315
+▁doctrina -11.0315
+▁embarazo -11.0315
+▁entertainment -11.0315
+▁entretenimiento -11.0315
+▁escandal -11.0315
+▁ferroviario -11.0315
+▁gabinete -11.0315
+▁guadalupe -11.0315
+▁intensidad -11.0315
+▁locutor -11.0315
+▁mallorca -11.0315
+▁masculina -11.0315
+▁mongolia -11.0315
+▁muñoz -11.0315
+▁nápoles -11.0315
+▁orgánica -11.0315
+▁pakistán -11.0315
+▁parámetro -11.0315
+▁prosiguió -11.0315
+▁quiméric -11.0315
+▁químico -11.0315
+▁ramírez -11.0315
+▁recomienda -11.0315
+▁reproduce -11.0315
+▁republicana -11.0315
+▁sacrificio -11.0315
+▁trujillo -11.0315
+▁valparaíso -11.0315
+▁algoritmo -11.0315
+▁asturias -11.0315
+▁bilbao -11.0315
+▁cataloga -11.0315
+▁charlie -11.0315
+▁coalición -11.0315
+▁dramática -11.0315
+▁glaciar -11.0315
+▁napoleón -11.0315
+▁paviment -11.0315
+▁sagrada -11.0315
+▁sindicato -11.0315
+▁veintidós -11.0315
+▁víctor -11.0315
+pareciera -11.0315
+▁aeronave -11.0315
+▁alcohol -11.0315
+▁curiosidad -11.0315
+▁destruye -11.0315
+▁frustr -11.0315
+▁impacien -11.0315
+▁rompió -11.0315
+▁censura -11.0315
+▁mauricio -11.0315
+▁ídolo -11.0315
+▁ecuatorial -11.0315
+▁flamenco -11.0315
+▁concreto -11.0315
+▁evacua -11.0315
+▁geográfico -11.0315
+▁ideología -11.0315
+▁certamen -11.0315
+▁procesión -11.0315
+▁situaciones -11.0315
+▁instancia -11.0315
+▁permiso -11.0315
+▁saudita -11.0315
+▁graduado -11.0315
+▁violín -11.0315
+▁emmy -11.0315
+▁herrera -11.0315
+▁transmite -11.0315
+▁gallego -11.0315
+▁josef -11.0315
+▁algodón -11.0315
+▁virrey -11.0315
+hidratado -11.0315
+▁dúo -11.0315
+▁paquete -11.0315
+▁beach -11.0315
+▁espinas -11.0315
+▁trastorno -11.0316
+▁nevada -11.0316
+▁asalto -11.0316
+▁engancha -11.0316
+▁genial -11.0316
+▁tambor -11.0316
+▁precede -11.0316
+▁abrigo -11.0316
+▁criminal -11.0316
+▁dorsal -11.0316
+ulenta -11.0316
+▁eterna -11.0317
+▁subleva -11.0317
+▁wolf -11.0317
+▁conduce -11.0317
+▁heladas -11.0317
+▁hostil -11.0317
+▁racional -11.0317
+▁silva -11.0317
+▁retiro -11.0317
+▁congelado -11.0317
+▁pedido -11.0317
+▁ducado -11.0318
+▁buey -11.0318
+▁mueva -11.0319
+▁icono -11.0319
+▁impidi -11.032
+monio -11.032
+▁balear -11.0322
+ávido -11.0322
+▁acoge -11.0333
+▁apodad -11.0334
+óxido -11.035
+▁mueve -11.0378
+tensiones -11.0387
+▁llov -11.0403
+rezca -11.0475
+▁nazi -11.0537
+traída -11.0659
+▁parezc -11.0815
+▁hispano -11.0885
+▁aspir -11.0897
+▁ligera -11.091
+▁hubie -11.091
+▁ocasiona -11.094
+▁aceitun -11.094
+▁adecuado -11.094
+▁agricultor -11.094
+▁alfonso -11.094
+▁audiovisual -11.094
+▁basílica -11.094
+▁claustro -11.094
+▁cuestiones -11.094
+▁cálculo -11.094
+▁especímenes -11.094
+▁etiopía -11.094
+▁guayana -11.094
+▁hermafrodita -11.094
+▁improvisa -11.094
+▁largometraje -11.094
+▁negrita -11.094
+▁occidente -11.094
+▁ofensivo -11.094
+▁olimpia -11.094
+▁ortodoxa -11.094
+▁peculiar -11.094
+▁perseguido -11.094
+▁proyección -11.094
+▁sensibilidad -11.094
+▁septentrional -11.094
+▁seudónimo -11.094
+▁subterránea -11.094
+▁superviviente -11.094
+▁sánchez -11.094
+▁terremoto -11.094
+▁trágica -11.094
+▁tuviera -11.094
+▁almohada -11.094
+▁basilio -11.094
+▁channel -11.094
+▁espectro -11.094
+▁exclam -11.094
+▁gregorio -11.094
+▁mercenario -11.094
+▁periódica -11.094
+▁relevante -11.094
+▁suprema -11.094
+▁suárez -11.094
+▁gratuita -11.094
+▁incierto -11.094
+▁sintetiza -11.094
+tíbulo -11.094
+▁progreso -11.094
+▁sólido -11.094
+▁conjunta -11.094
+▁acarici -11.094
+▁lisboa -11.094
+▁mosquito -11.094
+▁umbral -11.094
+▁ánimo -11.094
+▁armenio -11.094
+▁incendia -11.094
+▁emilio -11.094
+▁gusanos -11.094
+▁modesto -11.094
+▁festividad -11.094
+▁decisiones -11.094
+▁pianista -11.094
+▁racing -11.094
+▁galicia -11.094
+▁alquiler -11.094
+▁premier -11.094
+▁asunción -11.094
+▁anderson -11.094
+▁fluvial -11.094
+▁óscar -11.094
+▁armenia -11.0941
+fección -11.0941
+▁erigi -11.0941
+▁barrera -11.0941
+▁dorso -11.0941
+▁good -11.0941
+▁señas -11.0941
+▁nepal -11.0941
+▁tostada -11.0941
+▁minería -11.0942
+▁cuerno -11.0942
+▁ríe -11.0942
+▁molino -11.0942
+▁duele -11.0942
+▁filial -11.0942
+ectividad -11.0942
+▁racimo -11.0943
+▁codifica -11.0943
+script -11.0943
+▁prosper -11.0943
+▁consultor -11.0948
+▁traidor -11.0948
+▁servía -11.0948
+▁vapor -11.096
+▁filtr -11.0963
+▁ferm -11.0965
+▁drog -11.1008
+▁besarl -11.1038
+cogí -11.1072
+axila -11.1094
+▁despid -11.114
+▁bald -11.1151
+▁extrema -11.1259
+▁cansad -11.1468
+▁impact -11.1475
+▁president -11.1543
+▁castella -11.1586
+▁estremec -11.1606
+revolucionaria -11.1607
+ä -11.1607
+▁academy -11.1607
+▁arzobispo -11.1607
+▁atlántico -11.1607
+▁cariño -11.1607
+▁clasificó -11.1607
+▁coincidi -11.1607
+▁complutense -11.1607
+▁corresponsal -11.1607
+▁desprecia -11.1607
+▁diecisiete -11.1607
+▁directivo -11.1607
+▁discurre -11.1607
+▁discusión -11.1607
+▁eficaz -11.1607
+▁eficiente -11.1607
+▁engaña -11.1607
+▁espadaña -11.1607
+▁evangelio -11.1607
+▁franquicia -11.1607
+▁gimnasia -11.1607
+▁guarnición -11.1607
+▁helicóptero -11.1607
+▁inconsciente -11.1607
+▁inmóvil -11.1607
+▁international -11.1607
+▁jerarquía -11.1607
+▁jurisdicción -11.1607
+▁legislativa -11.1607
+▁lóbulo -11.1607
+▁meridional -11.1607
+▁metálico -11.1607
+▁monarquía -11.1607
+▁míchigan -11.1607
+▁negruzca -11.1607
+▁orquídea -11.1607
+▁peregrino -11.1607
+▁portavoz -11.1607
+▁pretendía -11.1607
+▁restringido -11.1607
+▁resuelto -11.1607
+▁revolote -11.1607
+▁revolucionario -11.1607
+▁sacramento -11.1607
+▁sucesiva -11.1607
+▁sábana -11.1607
+▁televisivo -11.1607
+▁vestimenta -11.1607
+▁world -11.1607
+▁zimbabue -11.1607
+▁éxito -11.1607
+â -11.1607
+▁asemeja -11.1607
+▁averigua -11.1607
+▁bombarde -11.1607
+▁boxeo -11.1607
+▁desciende -11.1607
+▁fecunda -11.1607
+▁interfaz -11.1607
+▁maravilla -11.1607
+▁nómada -11.1607
+▁unánime -11.1607
+▁desigual -11.1607
+▁erecto -11.1607
+▁explosión -11.1607
+▁mensual -11.1607
+▁primitivo -11.1607
+▁azteca -11.1607
+▁cobertura -11.1607
+▁ilegal -11.1607
+▁montañoso -11.1607
+▁topografía -11.1607
+▁diamante -11.1607
+▁dispares -11.1607
+▁mártir -11.1607
+▁christian -11.1607
+▁mozart -11.1607
+▁villaverde -11.1607
+▁demasiada -11.1607
+▁glabra -11.1607
+▁vuelvo -11.1607
+▁carnaval -11.1607
+▁diversión -11.1607
+▁sonríe -11.1607
+▁campesina -11.1607
+▁inversión -11.1607
+▁vacío -11.1607
+▁zamora -11.1607
+▁flecha -11.1607
+▁previo -11.1607
+▁permitía -11.1607
+▁progresa -11.1607
+▁utah -11.1607
+▁rechazo -11.1607
+▁estambul -11.1607
+▁iván -11.1607
+▁remedio -11.1607
+▁restantes -11.1607
+▁contento -11.1607
+▁sensible -11.1607
+▁arquero -11.1607
+▁especula -11.1607
+▁vocación -11.1607
+▁caduca -11.1607
+▁criterio -11.1607
+▁sky -11.1607
+▁plomo -11.1607
+▁detiene -11.1607
+▁oliver -11.1607
+▁escamas -11.1607
+▁teórica -11.1608
+▁cantón -11.1608
+▁rubio -11.1608
+▁fallecid -11.1608
+▁deterior -11.1608
+grima -11.1608
+▁moderado -11.1608
+▁alicante -11.1609
+▁gallard -11.1609
+▁calambr -11.1609
+▁rocoso -11.161
+▁tronco -11.161
+▁sienta -11.1611
+▁disminuye -11.1613
+▁contradic -11.1614
+ícola -11.1614
+▁cintur -11.1614
+▁lejano -11.1614
+▁pastel -11.1615
+▁jacob -11.1616
+uestos -11.1631
+▁francesc -11.1631
+▁corral -11.1649
+viene -11.1651
+triz -11.1705
+▁vagab -11.1717
+▁mirándo -11.1795
+illetes -11.182
+▁recurso -11.1843
+▁relativ -11.1991
+▁suplic -11.2044
+▁aprend -11.2102
+▁institu -11.227
+▁baviera -11.2321
+imilitud -11.2321
+lanceoladas -11.2321
+▁amazonas -11.2321
+▁aparcamiento -11.2321
+▁atmósfera -11.2321
+▁autodidacta -11.2321
+▁birmingham -11.2321
+▁blanquecino -11.2321
+▁cabaña -11.2321
+▁cálida -11.2321
+▁cólera -11.2321
+▁decisiva -11.2321
+▁democrático -11.2321
+▁departamental -11.2321
+▁desaparición -11.2321
+▁diccionario -11.2321
+▁dinosaurio -11.2321
+▁directiva -11.2321
+▁disfrace -11.2321
+▁disolvió -11.2321
+▁electricidad -11.2321
+▁enciclopedia -11.2321
+▁equilibrio -11.2321
+▁escándalo -11.2321
+▁estupendo -11.2321
+▁estúpido -11.2321
+▁explosivo -11.2321
+▁extenso -11.2321
+▁extraordinaria -11.2321
+▁extraordinario -11.2321
+▁fascinado -11.2321
+▁filósofo -11.2321
+▁franquista -11.2321
+▁gimnasio -11.2321
+▁hidrógeno -11.2321
+▁historieta -11.2321
+▁illinois -11.2321
+▁increíble -11.2321
+▁independentista -11.2321
+▁lawrence -11.2321
+▁matemático -11.2321
+▁mausoleo -11.2321
+▁membrana -11.2321
+▁musulmanes -11.2321
+▁méxico -11.2321
+▁network -11.2321
+▁ofensiva -11.2321
+▁olvide -11.2321
+▁oscurec -11.2321
+▁pensilvania -11.2321
+▁periodística -11.2321
+▁persigue -11.2321
+▁petróleo -11.2321
+▁pórtico -11.2321
+▁reemplazó -11.2321
+▁religiones -11.2321
+▁románico -11.2321
+▁subcampeón -11.2321
+▁suministro -11.2321
+▁tardío -11.2321
+▁testimonio -11.2321
+▁urbanización -11.2321
+▁águila -11.2321
+▁dignidad -11.2321
+▁divulga -11.2321
+▁esqueleto -11.2321
+▁fallecimiento -11.2321
+▁infecta -11.2321
+▁longitudinales -11.2321
+▁tarifa -11.2321
+▁tarjeta -11.2321
+▁zambia -11.2321
+▁arizona -11.2321
+▁atalaya -11.2321
+▁competir -11.2321
+▁elefante -11.2321
+▁esparcido -11.2321
+▁ginebra -11.2321
+▁herejes -11.2321
+▁primitiva -11.2321
+▁asiento -11.2321
+▁condecora -11.2321
+▁cuchillo -11.2321
+▁federico -11.2321
+▁hitler -11.2321
+▁plástica -11.2321
+▁prusia -11.2321
+▁veintiocho -11.2321
+▁fractura -11.2321
+▁autonomía -11.2321
+▁mágico -11.2321
+▁campeones -11.2321
+▁difund -11.2321
+▁referente -11.2321
+▁visitó -11.2321
+▁amateur -11.2321
+▁promotor -11.2321
+▁rocío -11.2321
+▁cartagena -11.2321
+▁chipre -11.2321
+▁burdeos -11.2321
+▁quinteto -11.2321
+▁sobrina -11.2321
+▁encerrado -11.2321
+▁mercader -11.2321
+▁pistola -11.2321
+▁masiva -11.2321
+▁ghana -11.2321
+▁terraza -11.2321
+▁retraso -11.2321
+▁encomend -11.2321
+▁mentira -11.2321
+▁comproba -11.2321
+▁ladrones -11.2321
+▁secretaría -11.2321
+▁studio -11.2321
+▁siervo -11.2321
+▁what -11.2321
+▁soluciones -11.2321
+▁descrita -11.2321
+▁perdón -11.2322
+▁colonos -11.2322
+▁erosión -11.2322
+▁cifra -11.2322
+▁eterno -11.2322
+▁andrea -11.2322
+▁requeri -11.2322
+ténga -11.2322
+▁reposo -11.2322
+▁blanqu -11.2322
+▁baraja -11.2322
+▁redonda -11.2322
+▁ligero -11.2323
+▁oprim -11.2323
+▁ondula -11.2324
+▁inspec -11.2325
+▁corregir -11.2336
+patía -11.2359
+ponerse -11.2365
+▁pulp -11.2389
+▁conserv -11.2393
+spiraba -11.2397
+▁oblig -11.2514
+▁minuto -11.2613
+▁tranquil -11.2828
+▁austral -11.2997
+▁represent -11.3003
+cionamiento -11.3018
+▁ejerci -11.3023
+▁matorral -11.3042
+▁bachiller -11.309
+prefectura -11.309
+▁acústica -11.309
+▁albahaca -11.309
+▁alfabeto -11.309
+▁ansiedad -11.309
+▁antártida -11.309
+▁arequipa -11.309
+▁atlanta -11.309
+▁belgrano -11.309
+▁caracteres -11.309
+▁christopher -11.309
+▁cirujan -11.309
+▁concluyó -11.309
+▁contribuciones -11.309
+▁coruña -11.309
+▁cátedra -11.309
+▁desafortunada -11.309
+▁descubierta -11.309
+▁dieciséis -11.309
+▁difunto -11.309
+▁dormitorio -11.309
+▁douglas -11.309
+▁elíptica -11.309
+▁epidemia -11.309
+▁exámenes -11.309
+▁galardonado -11.309
+▁guayaquil -11.309
+▁hallazgos -11.309
+▁hidalgo -11.309
+▁ignición -11.309
+▁imprescindible -11.309
+▁insignia -11.309
+▁jerónimo -11.309
+▁legendario -11.309
+▁lluvioso -11.309
+▁logotipo -11.309
+▁maquinaria -11.309
+▁mediterránea -11.309
+▁mezquita -11.309
+▁microsoft -11.309
+▁molecular -11.309
+▁mountain -11.309
+▁módulo -11.309
+▁músculo -11.309
+▁namibia -11.309
+▁neoclásico -11.309
+▁olímpica -11.309
+▁oviedo -11.309
+▁pabellón -11.309
+▁palpitante -11.309
+▁pamplona -11.309
+▁pastizales -11.309
+▁playoff -11.309
+▁predecesor -11.309
+▁progresivamente -11.309
+▁reconstrucción -11.309
+▁resistir -11.309
+▁rígido -11.309
+▁science -11.309
+▁sobreviviente -11.309
+▁sufragánea -11.309
+▁suspirante -11.309
+▁televisiva -11.309
+▁tranvía -11.309
+ë -11.309
+▁absurdo -11.309
+▁argelia -11.309
+▁automático -11.309
+▁camboya -11.309
+▁dramático -11.309
+▁inocente -11.309
+▁preferencia -11.309
+▁progresista -11.309
+▁reconcilia -11.309
+▁alegoría -11.309
+▁aprieta -11.309
+▁calderón -11.309
+▁heavy -11.309
+▁octava -11.309
+▁preferente -11.309
+▁redacción -11.309
+▁súper -11.309
+▁automática -11.309
+▁exacto -11.309
+▁imprim -11.309
+▁lujoso -11.309
+▁orientado -11.309
+▁confusa -11.309
+▁sajonia -11.309
+▁beatriz -11.309
+▁neutral -11.309
+▁misionero -11.309
+▁montreal -11.309
+▁uganda -11.309
+▁capitel -11.309
+▁charlot -11.309
+▁complicada -11.309
+▁emblema -11.309
+▁gabón -11.309
+▁pizarra -11.309
+▁complicado -11.309
+▁desorden -11.309
+▁trío -11.309
+▁pionera -11.309
+▁surgieron -11.309
+▁anatom -11.309
+▁perezos -11.309
+▁tradiciones -11.309
+▁aborto -11.309
+▁vargas -11.309
+discreto -11.309
+▁sofía -11.309
+▁tenerife -11.309
+▁decreto -11.309
+▁muscular -11.309
+▁parodia -11.309
+▁abandono -11.3091
+▁indecis -11.3091
+▁camisa -11.3091
+▁bodega -11.3091
+▁aclara -11.3091
+▁vigente -11.3091
+▁corazon -11.3091
+▁rally -11.3091
+▁compensa -11.3091
+▁repetí -11.3092
+▁molesta -11.3092
+▁materno -11.3092
+▁policial -11.3092
+▁capitan -11.3093
+ministra -11.3093
+worth -11.3094
+gonzalo -11.3095
+▁exige -11.3104
+▁cautiv -11.3104
+▁tímid -11.311
+partía -11.3137
+▁entabl -11.3381
+▁numerosa -11.3448
+▁aprovech -11.3636
+stitución -11.3834
+oriental -11.385
+▁semifinal -11.3876
+▁resolv -11.391
+queño -11.3919
+▁propuls -11.3921
+▁azerbai -11.3924
+▁virtual -11.3924
+dimensional -11.3924
+▁acondicionado -11.3924
+▁antecedente -11.3924
+▁aparecía -11.3924
+▁asciende -11.3924
+▁autobuses -11.3924
+▁beneficia -11.3924
+▁berkeley -11.3924
+▁brácteas -11.3924
+▁cafetería -11.3924
+▁calendario -11.3924
+▁cilíndrica -11.3924
+▁circunvala -11.3924
+▁cocodrilo -11.3924
+▁columnista -11.3924
+▁continuidad -11.3924
+▁corrupción -11.3924
+▁cotidiana -11.3924
+▁cretácico -11.3924
+▁cuarteto -11.3924
+▁desaparecieron -11.3924
+▁difiere -11.3924
+▁dinámica -11.3924
+▁divertido -11.3924
+▁emocionante -11.3924
+▁encorvado -11.3924
+▁espectacular -11.3924
+▁exclusivo -11.3924
+▁excursión -11.3924
+▁facebook -11.3924
+▁football -11.3924
+▁fotógrafo -11.3924
+▁galaxia -11.3924
+▁gloriosa -11.3924
+▁gradualmente -11.3924
+▁graduarse -11.3924
+▁indiferente -11.3924
+▁japoneses -11.3924
+▁jimmy -11.3924
+▁juguetón -11.3924
+▁lámpara -11.3924
+▁madrileño -11.3924
+▁magistrado -11.3924
+▁minnesota -11.3924
+▁monumental -11.3924
+▁multijugador -11.3924
+▁neerlandesa -11.3924
+▁nobiliario -11.3924
+▁núñez -11.3924
+▁ocurrido -11.3924
+▁oxígeno -11.3924
+▁pictures -11.3924
+▁predilecto -11.3924
+▁pubescente -11.3924
+▁purísima -11.3924
+▁recolección -11.3924
+▁recurrente -11.3924
+▁reemplazada -11.3924
+▁refuerzos -11.3924
+▁reliquia -11.3924
+▁restringida -11.3924
+▁rincón -11.3924
+▁sacrifica -11.3924
+▁sobrevivido -11.3924
+▁sombrío -11.3924
+▁sospechoso -11.3924
+▁stephen -11.3924
+▁submarino -11.3924
+▁surrealista -11.3924
+▁topónimo -11.3924
+▁toulouse -11.3924
+▁tregua -11.3924
+▁tácti -11.3924
+▁tóxico -11.3924
+▁vanguardia -11.3924
+▁windows -11.3924
+▁íntegra -11.3924
+_ -11.3924
+▁apetece -11.3924
+▁deprisa -11.3924
+▁lánguid -11.3924
+▁cuestiona -11.3924
+▁emprendi -11.3924
+▁finanzas -11.3924
+▁herradura -11.3924
+▁oponente -11.3924
+▁recolect -11.3924
+▁caprichos -11.3924
+▁follaje -11.3924
+▁suplente -11.3924
+▁andrew -11.3924
+▁antecesor -11.3924
+▁arthur -11.3924
+▁canónico -11.3924
+▁aqua -11.3924
+▁frederic -11.3924
+▁precisión -11.3924
+▁navarro -11.3924
+▁palidec -11.3924
+▁preserva -11.3924
+▁acomoda -11.3924
+▁afrodita -11.3924
+▁axilares -11.3924
+▁cantautor -11.3924
+▁embalse -11.3924
+▁austin -11.3924
+▁idiota -11.3924
+▁potosí -11.3924
+▁lienzo -11.3924
+▁asumir -11.3924
+▁superliga -11.3924
+▁instrucción -11.3924
+▁consumir -11.3924
+▁radiación -11.3924
+▁perfeccion -11.3924
+▁plantilla -11.3924
+▁comentarista -11.3924
+▁exhibe -11.3924
+▁morisco -11.3924
+▁macizo -11.3924
+▁philip -11.3924
+▁triple -11.3924
+▁barquero -11.3924
+▁pasaporte -11.3924
+iggins -11.3925
+▁frijol -11.3925
+▁santísim -11.3926
+▁degrad -11.3926
+▁vatican -11.3927
+enclavad -11.3927
+▁artesanal -11.3927
+sgraciadamente -11.3928
+▁intervi -11.3936
+▁florenc -11.3942
+▁exigi -11.3947
+liciosa -11.3967
+▁imagin -11.4109
+corporación -11.4224
+luís -11.4663
+▁albacete -11.4832
+▁clandestin -11.4833
+▁clausura -11.4833
+▁excursion -11.4833
+▁perseguía -11.4833
+apareciendo -11.4833
+tífice -11.4833
+▁acuerdas -11.4833
+▁adecuada -11.4833
+▁alfombra -11.4833
+▁alguacil -11.4833
+▁ametralladora -11.4833
+▁amurallado -11.4833
+▁aromática -11.4833
+▁arquitectónica -11.4833
+▁autobús -11.4833
+▁broadway -11.4833
+▁cilindro -11.4833
+▁cofradía -11.4833
+▁comisionado -11.4833
+▁compañera -11.4833
+▁conceptual -11.4833
+▁coordenadas -11.4833
+▁croquetas -11.4833
+▁cuadrangular -11.4833
+▁cáliz -11.4833
+▁danubio -11.4833
+▁diecinueve -11.4833
+▁eclesiástico -11.4833
+▁economista -11.4833
+▁efectivo -11.4833
+▁efectu -11.4833
+▁ejecutiva -11.4833
+▁evangelista -11.4833
+▁farmacia -11.4833
+▁ferroviaria -11.4833
+▁ficticio -11.4833
+▁filosófico -11.4833
+ā -11.6944
+ø -12.1289
+š -12.1289
+ū -12.1289
+æ -12.3289
+č -12.3289
+ô -12.3289
+е -12.5789
+ê -12.5789
+ʻ -12.9122
+ð -13.4122
+đ -13.4122
+ş -13.4122
+ý -13.4122
+ă -13.4122
+ė -13.4122
+ī -13.4122
+ł -13.4122
+ř -13.4122
+ː -13.4122
+ś -13.4122
+ゴ -13.4122
+ã -14.4065
+ё -14.4116
+ß -14.4117
+ć -14.4118
+ï -14.4119
+ì -14.412
+à -14.4121
+þ -14.4122
+ə -14.4122
+î -14.4122
+û -14.4122
+ě -14.4122
+ı -14.4122
+ń -14.4122
+œ -14.4122
+ž -14.4122
+ț -14.4122
+ʿ -14.4122
+ϙ -14.4122
+ה -14.4122
+נ -14.4122
+ة -14.4122
+ي -14.4122
+ṃ -14.4122
+の -14.4122
+ミ -14.4122
+ム -14.4122
+箱 -14.4122
+鮓 -14.4122
+鮨 -14.4122
+ò -14.4122
+し -14.4122
+ñ -14.4122
+י -14.4122
+כ -14.4122
+ש -14.4122
+ب -14.4122
+ذ -14.4122
+ه -14.4122
+消 -14.4122
diff --git a/configs/es-en/tgt_unigram6000/spm_unigram_es.model b/configs/es-en/tgt_unigram6000/spm_unigram_es.model
new file mode 100644
index 0000000000000000000000000000000000000000..9b478a4f6c46344c9ad2f084b235e295ab5a0e57
--- /dev/null
+++ b/configs/es-en/tgt_unigram6000/spm_unigram_es.model
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:333afbc60aaa4b05b0b92f0a39cbcc092c7238ab51a1785a871aa254d5ed0d9e
+size 344958
diff --git a/configs/es-en/tgt_unigram6000/spm_unigram_es.txt b/configs/es-en/tgt_unigram6000/spm_unigram_es.txt
new file mode 100644
index 0000000000000000000000000000000000000000..711c90e00557db7acd3d61d26ab1940f8e51da96
--- /dev/null
+++ b/configs/es-en/tgt_unigram6000/spm_unigram_es.txt
@@ -0,0 +1,5996 @@
+▁the 1
+s 1
+▁of 1
+▁a 1
+▁in 1
+▁and 1
+▁to 1
+▁is 1
+▁it 1
+▁he 1
+▁was 1
+' 1
+ed 1
+▁with 1
+t 1
+▁i 1
+▁are 1
+▁that 1
+ing 1
+▁they 1
+▁this 1
+▁for 1
+▁his 1
+n 1
+▁you 1
+y 1
+▁be 1
+d 1
+▁by 1
+▁on 1
+▁as 1
+l 1
+▁has 1
+e 1
+▁from 1
+ly 1
+▁have 1
+▁not 1
+▁were 1
+▁we 1
+▁at 1
+o 1
+▁there 1
+▁ 1
+▁an 1
+▁one 1
+a 1
+er 1
+▁can 1
+▁their 1
+▁she 1
+m 1
+al 1
+▁its 1
+r 1
+▁de 1
+c 1
+▁me 1
+es 1
+▁two 1
+▁first 1
+▁her 1
+▁or 1
+▁all 1
+▁but 1
+▁no 1
+▁had 1
+▁do 1
+▁will 1
+▁been 1
+ra 1
+▁up 1
+i 1
+▁like 1
+▁my 1
+▁other 1
+▁some 1
+p 1
+re 1
+g 1
+▁when 1
+▁time 1
+▁out 1
+▁them 1
+u 1
+▁go 1
+▁many 1
+▁during 1
+▁these 1
+▁new 1
+▁three 1
+▁don 1
+▁only 1
+▁la 1
+▁city 1
+▁most 1
+▁if 1
+ro 1
+▁what 1
+▁so 1
+▁more 1
+ll 1
+in 1
+▁part 1
+▁would 1
+▁get 1
+la 1
+ic 1
+▁him 1
+▁about 1
+▁located 1
+▁same 1
+▁after 1
+▁made 1
+▁used 1
+f 1
+ation 1
+▁did 1
+▁re 1
+▁al 1
+▁later 1
+▁o 1
+or 1
+▁year 1
+▁who 1
+very 1
+le 1
+▁work 1
+▁which 1
+ri 1
+▁however 1
+▁name 1
+▁severa 1
+ve 1
+▁both 1
+ta 1
+▁where 1
+▁well 1
+▁k 1
+ar 1
+▁look 1
+nt 1
+▁years 1
+▁currently 1
+en 1
+▁good 1
+▁call 1
+▁team 1
+▁found 1
+▁different 1
+an 1
+out 1
+ch 1
+▁place 1
+▁people 1
+na 1
+te 1
+▁come 1
+▁your 1
+b 1
+▁make 1
+ion 1
+k 1
+▁over 1
+▁last 1
+us 1
+▁s 1
+ma 1
+▁known 1
+▁school 1
+ist 1
+▁great 1
+h 1
+▁main 1
+▁see 1
+▁university 1
+▁because 1
+▁life 1
+▁national 1
+▁second 1
+▁than 1
+▁then 1
+▁back 1
+▁ver 1
+▁c 1
+▁way 1
+to 1
+lo 1
+z 1
+▁each 1
+▁know 1
+▁want 1
+▁let 1
+on 1
+▁man 1
+▁war 1
+▁into 1
+▁love 1
+▁bet 1
+▁does 1
+able 1
+▁just 1
+▁world 1
+w 1
+▁now 1
+▁born 1
+▁end 1
+ween 1
+▁us 1
+▁general 1
+▁four 1
+▁being 1
+co 1
+ca 1
+th 1
+ne 1
+ia 1
+▁could 1
+li 1
+▁san 1
+va 1
+▁since 1
+ok 1
+▁got 1
+ur 1
+ment 1
+▁take 1
+▁family 1
+▁any 1
+▁even 1
+▁p 1
+▁old 1
+ce 1
+▁played 1
+▁right 1
+▁el 1
+▁never 1
+▁high 1
+▁album 1
+▁son 1
+▁through 1
+il 1
+▁long 1
+▁am 1
+de 1
+▁small 1
+▁car 1
+do 1
+▁always 1
+▁song 1
+▁water 1
+▁under 1
+x 1
+▁considered 1
+▁co 1
+▁music 1
+▁started 1
+▁those 1
+▁important 1
+▁area 1
+▁still 1
+▁group 1
+▁before 1
+ers 1
+▁also 1
+▁river 1
+▁left 1
+▁father 1
+ting 1
+ter 1
+vi 1
+▁studied 1
+▁con 1
+▁club 1
+no 1
+▁another 1
+ive 1
+▁little 1
+▁center 1
+ity 1
+el 1
+▁day 1
+▁going 1
+▁works 1
+▁li 1
+ian 1
+▁spanish 1
+▁such 1
+▁worked 1
+▁may 1
+ism 1
+▁among 1
+▁due 1
+se 1
+▁town 1
+ck 1
+▁head 1
+▁children 1
+▁state 1
+▁un 1
+▁down 1
+▁think 1
+▁turn 1
+▁ship 1
+▁white 1
+da 1
+▁e 1
+ies 1
+ous 1
+▁black 1
+▁st 1
+▁how 1
+st 1
+▁girl 1
+ul 1
+▁five 1
+▁house 1
+me 1
+▁play 1
+▁again 1
+▁south 1
+▁species 1
+▁country 1
+▁band 1
+▁young 1
+▁together 1
+▁pro 1
+▁member 1
+▁order 1
+▁film 1
+it 1
+ti 1
+ga 1
+▁t 1
+▁region 1
+▁usually 1
+man 1
+▁ro 1
+est 1
+ry 1
+▁sea 1
+▁series 1
+▁find 1
+▁best 1
+came 1
+▁won 1
+▁use 1
+▁international 1
+▁career 1
+▁ca 1
+▁received 1
+▁need 1
+- 1
+▁short 1
+▁married 1
+os 1
+▁president 1
+ba 1
+▁night 1
+▁said 1
+▁better 1
+ary 1
+▁f 1
+▁united 1
+▁church 1
+▁unti 1
+is 1
+▁show 1
+▁must 1
+▁spain 1
+▁public 1
+▁di 1
+▁away 1
+▁next 1
+▁party 1
+▁big 1
+▁b 1
+▁following 1
+▁capital 1
+▁help 1
+▁say 1
+▁original 1
+um 1
+id 1
+▁death 1
+▁against 1
+mo 1
+▁single 1
+▁vi 1
+▁na 1
+▁system 1
+▁here 1
+▁home 1
+▁mo 1
+tic 1
+▁should 1
+▁bo 1
+▁del 1
+▁give 1
+▁history 1
+ton 1
+▁rest 1
+▁mother 1
+▁came 1
+▁mainly 1
+▁six 1
+▁live 1
+▁close 1
+▁finally 1
+▁lot 1
+ted 1
+▁red 1
+▁too 1
+▁tell 1
+▁pi 1
+▁our 1
+que 1
+▁belong 1
+▁movie 1
+▁natural 1
+▁moment 1
+▁states 1
+▁v 1
+▁while 1
+▁much 1
+▁hand 1
+▁ma 1
+go 1
+▁named 1
+▁game 1
+▁se 1
+▁y 1
+pe 1
+un 1
+▁north 1
+▁around 1
+▁nothing 1
+▁per 1
+▁lives 1
+▁d 1
+▁season 1
+▁third 1
+▁few 1
+ent 1
+ty 1
+tor 1
+▁built 1
+▁ta 1
+▁province 1
+mp 1
+▁building 1
+▁mar 1
+▁light 1
+land 1
+ge 1
+▁law 1
+▁side 1
+▁own 1
+po 1
+▁color 1
+▁already 1
+▁company 1
+▁saint 1
+di 1
+▁open 1
+▁something 1
+▁government 1
+▁published 1
+▁died 1
+▁bar 1
+mi 1
+▁include 1
+▁lo 1
+ness 1
+▁en 1
+▁period 1
+most 1
+et 1
+▁american 1
+▁seem 1
+▁mi 1
+▁today 1
+▁station 1
+ate 1
+ble 1
+▁written 1
+▁eyes 1
+▁every 1
+▁book 1
+▁daughter 1
+▁art 1
+▁player 1
+bi 1
+▁le 1
+ir 1
+▁plays 1
+▁po 1
+▁special 1
+▁local 1
+les 1
+▁direct 1
+▁pa 1
+▁everything 1
+▁set 1
+▁ba 1
+▁land 1
+▁leave 1
+▁political 1
+▁line 1
+▁times 1
+tion 1
+▁based 1
+▁district 1
+io 1
+ish 1
+as 1
+▁number 1
+▁te 1
+▁human 1
+▁production 1
+▁park 1
+▁division 1
+▁full 1
+ated 1
+▁island 1
+ine 1
+▁songs 1
+▁pre 1
+▁lost 1
+▁education 1
+son 1
+▁maria 1
+▁hu 1
+▁style 1
+▁nowadays 1
+▁similar 1
+▁studies 1
+▁eight 1
+▁fight 1
+▁chole 1
+▁dark 1
+gu 1
+▁director 1
+▁moved 1
+▁cor 1
+▁case 1
+sh 1
+▁ga 1
+▁honor 1
+▁current 1
+▁beside 1
+▁put 1
+▁strong 1
+▁please 1
+▁title 1
+▁common 1
+nd 1
+▁seen 1
+▁position 1
+ical 1
+▁become 1
+▁central 1
+▁fall 1
+▁social 1
+▁juan 1
+der 1
+▁military 1
+▁green 1
+▁bad 1
+▁off 1
+▁mu 1
+ru 1
+bo 1
+▁pu 1
+▁free 1
+ph 1
+▁brother 1
+▁person 1
+▁start 1
+▁final 1
+▁friend 1
+▁once 1
+ut 1
+▁body 1
+ff 1
+at 1
+▁lu 1
+▁degree 1
+▁league 1
+▁real 1
+▁popular 1
+▁madrid 1
+▁support 1
+za 1
+▁less 1
+be 1
+▁stop 1
+sa 1
+▁origin 1
+▁sister 1
+age 1
+ver 1
+▁present 1
+▁seven 1
+▁space 1
+▁ra 1
+vo 1
+ig 1
+▁half 1
+▁age 1
+▁ten 1
+▁sun 1
+▁men 1
+▁large 1
+▁eat 1
+ie 1
+▁process 1
+▁ex 1
+▁according 1
+▁front 1
+▁areas 1
+▁professional 1
+ant 1
+▁pe 1
+▁population 1
+ina 1
+ke 1
+▁door 1
+▁ho 1
+▁field 1
+v 1
+▁boy 1
+▁wh 1
+▁role 1
+im 1
+▁th 1
+▁ja 1
+▁power 1
+▁consist 1
+▁german 1
+less 1
+▁sometimes 1
+pa 1
+▁twenty 1
+▁inside 1
+▁project 1
+▁ri 1
+▁example 1
+▁story 1
+▁decided 1
+▁carrie 1
+▁museum 1
+▁flowers 1
+res 1
+▁began 1
+▁los 1
+ni 1
+▁blue 1
+▁native 1
+▁count 1
+▁sa 1
+▁leaves 1
+▁hard 1
+▁dr 1
+▁kind 1
+▁europe 1
+▁soon 1
+▁act 1
+▁charge 1
+▁participated 1
+▁council 1
+▁women 1
+▁plant 1
+▁really 1
+▁famous 1
+▁created 1
+▁julia 1
+▁himself 1
+▁things 1
+▁point 1
+▁ne 1
+▁east 1
+▁county 1
+▁shows 1
+ling 1
+▁ar 1
+▁mean 1
+▁economic 1
+▁job 1
+though 1
+▁video 1
+▁beginning 1
+▁room 1
+ard 1
+▁g 1
+▁face 1
+▁floor 1
+cy 1
+▁port 1
+▁talk 1
+▁u 1
+ria 1
+▁members 1
+▁mon 1
+▁bu 1
+ized 1
+▁africa 1
+▁continued 1
+▁change 1
+▁produced 1
+▁games 1
+▁included 1
+▁anything 1
+▁movement 1
+ial 1
+▁star 1
+▁ti 1
+▁form 1
+▁hundred 1
+▁singer 1
+▁result 1
+▁taken 1
+▁having 1
+▁everyone 1
+ny 1
+▁possible 1
+▁lived 1
+▁chapter 1
+▁takes 1
+▁author 1
+▁king 1
+▁friends 1
+▁republic 1
+▁fl 1
+▁hi 1
+▁cross 1
+▁roman 1
+▁battle 1
+▁sound 1
+▁civil 1
+▁fish 1
+ha 1
+ja 1
+▁da 1
+▁fact 1
+▁remains 1
+ho 1
+▁keep 1
+▁dedicat 1
+▁brown 1
+▁fe 1
+▁date 1
+▁comes 1
+▁gu 1
+▁feel 1
+▁move 1
+▁done 1
+▁others 1
+▁musical 1
+▁character 1
+▁theater 1
+▁sub 1
+▁tu 1
+▁return 1
+nes 1
+▁france 1
+▁br 1
+▁couple 1
+▁village 1
+▁office 1
+▁far 1
+▁professor 1
+▁buried 1
+▁football 1
+ally 1
+ber 1
+▁indian 1
+▁chile 1
+▁argentina 1
+▁mountain 1
+▁du 1
+▁rock 1
+▁official 1
+ful 1
+▁influence 1
+▁middle 1
+▁round 1
+▁study 1
+▁municipality 1
+▁opera 1
+bu 1
+▁west 1
+fi 1
+hi 1
+▁version 1
+▁wait 1
+ka 1
+▁language 1
+men 1
+▁john 1
+▁means 1
+▁covered 1
+▁beautiful 1
+op 1
+▁days 1
+▁base 1
+▁wind 1
+ac 1
+▁lead 1
+ana 1
+▁lower 1
+▁ru 1
+▁dis 1
+▁hands 1
+▁program 1
+▁court 1
+▁master 1
+▁female 1
+▁woman 1
+▁stone 1
+▁groups 1
+▁reach 1
+▁die 1
+▁thing 1
+▁goes 1
+▁cha 1
+▁newspaper 1
+▁held 1
+▁developed 1
+▁met 1
+▁try 1
+ology 1
+▁fire 1
+▁actor 1
+▁addition 1
+▁pri 1
+▁books 1
+▁street 1
+▁nine 1
+▁early 1
+▁past 1
+▁habitat 1
+▁radio 1
+▁kept 1
+▁information 1
+▁stage 1
+▁god 1
+▁cu 1
+▁bi 1
+▁m 1
+▁late 1
+▁coast 1
+▁garden 1
+▁development 1
+▁jo 1
+▁doctor 1
+qui 1
+con 1
+▁department 1
+▁modern 1
+▁television 1
+▁board 1
+▁level 1
+jo 1
+▁attack 1
+▁follow 1
+▁york 1
+▁teacher 1
+▁cover 1
+▁mat 1
+▁ha 1
+▁soccer 1
+▁service 1
+▁tra 1
+▁sir 1
+▁parents 1
+▁allow 1
+▁happy 1
+▁wife 1
+▁activity 1
+▁european 1
+▁oh 1
+▁yellow 1
+▁near 1
+▁r 1
+▁traditional 1
+▁voice 1
+▁relationship 1
+▁access 1
+ance 1
+▁su 1
+▁enter 1
+▁toward 1
+▁painting 1
+ell 1
+▁divided 1
+am 1
+▁plan 1
+▁type 1
+▁seat 1
+▁wrote 1
+▁wood 1
+▁headquarters 1
+ric 1
+▁comp 1
+▁food 1
+▁neighborhood 1
+▁tower 1
+▁road 1
+▁run 1
+mer 1
+▁em 1
+▁ground 1
+▁composed 1
+▁months 1
+ron 1
+▁magazine 1
+up 1
+▁summer 1
+▁alone 1
+▁china 1
+▁ones 1
+▁saw 1
+▁sp 1
+▁historical 1
+▁hours 1
+▁dog 1
+▁whole 1
+▁exist 1
+▁low 1
+ions 1
+▁child 1
+▁along 1
+▁released 1
+▁working 1
+fa 1
+▁structure 1
+▁tournament 1
+▁regard 1
+▁depend 1
+▁countries 1
+▁train 1
+▁writer 1
+▁told 1
+▁union 1
+▁inter 1
+▁commercial 1
+▁performance 1
+▁design 1
+lin 1
+▁playing 1
+▁completely 1
+ler 1
+▁arrived 1
+pi 1
+ino 1
+▁appeared 1
+▁paris 1
+▁week 1
+▁community 1
+▁research 1
+▁tour 1
+ley 1
+▁provide 1
+▁kill 1
+▁gi 1
+▁organization 1
+▁hair 1
+▁cap 1
+▁idea 1
+▁test 1
+▁air 1
+rre 1
+▁sure 1
+▁sports 1
+▁model 1
+▁finished 1
+▁ni 1
+▁british 1
+▁border 1
+▁kiss 1
+tes 1
+▁previous 1
+▁ended 1
+▁performed 1
+▁top 1
+▁california 1
+▁competition 1
+▁episode 1
+▁winner 1
+▁section 1
+▁sorry 1
+mon 1
+▁para 1
+▁control 1
+▁activities 1
+▁mexican 1
+▁palomares 1
+▁someone 1
+▁dead 1
+ries 1
+▁nor 1
+ster 1
+▁edition 1
+▁arm 1
+▁castle 1
+▁hall 1
+▁range 1
+▁often 1
+per 1
+▁chi 1
+ization 1
+▁goal 1
+ure 1
+ner 1
+▁academy 1
+▁ge 1
+ero 1
+▁given 1
+▁list 1
+gi 1
+▁represent 1
+▁catholic 1
+refore 1
+▁cultural 1
+▁distributed 1
+▁rain 1
+▁course 1
+▁served 1
+▁royal 1
+▁trip 1
+▁forests 1
+▁class 1
+▁believe 1
+▁shape 1
+▁event 1
+▁h 1
+▁luis 1
+▁managed 1
+▁returned 1
+▁size 1
+▁post 1
+▁tropical 1
+▁unknown 1
+▁culture 1
+▁winter 1
+▁sent 1
+▁fruit 1
+▁reason 1
+▁lake 1
+▁record 1
+cia 1
+▁business 1
+▁cannot 1
+▁caused 1
+▁ka 1
+▁thought 1
+▁dry 1
+ki 1
+▁private 1
+▁various 1
+▁stories 1
+▁recorded 1
+▁raised 1
+▁classic 1
+▁simple 1
+ow 1
+▁leader 1
+▁ce 1
+▁major 1
+▁sin 1
+▁territory 1
+▁enough 1
+▁relative 1
+au 1
+▁santa 1
+▁force 1
+▁dance 1
+▁hope 1
+ium 1
+▁consider 1
+▁beach 1
+▁queen 1
+▁related 1
+nce 1
+▁w 1
+▁wanted 1
+▁trees 1
+▁closed 1
+▁hold 1
+▁islands 1
+▁sta 1
+gue 1
+▁trans 1
+▁barcelona 1
+▁london 1
+▁square 1
+▁championship 1
+illa 1
+▁heart 1
+▁male 1
+rs 1
+ten 1
+▁sleep 1
+▁dream 1
+▁french 1
+▁pass 1
+▁thank 1
+▁chosen 1
+ction 1
+▁coming 1
+▁remained 1
+▁certain 1
+▁russian 1
+▁nu 1
+▁teams 1
+line 1
+▁institute 1
+ities 1
+▁bring 1
+▁hospital 1
+▁representative 1
+▁active 1
+▁care 1
+▁poor 1
+▁learn 1
+▁clear 1
+tu 1
+▁critics 1
+▁anti 1
+▁behind 1
+▁complex 1
+▁actress 1
+▁francisco 1
+▁western 1
+▁sign 1
+ng 1
+▁fourth 1
+▁gar 1
+▁australia 1
+▁brazil 1
+▁artist 1
+▁represented 1
+▁experience 1
+▁win 1
+▁plants 1
+▁asia 1
+▁mass 1
+gen 1
+▁forest 1
+▁gav 1
+▁carlos 1
+▁month 1
+▁race 1
+▁stay 1
+▁birds 1
+▁events 1
+▁germany 1
+ad 1
+▁match 1
+▁situation 1
+▁image 1
+▁bridge 1
+▁construction 1
+let 1
+▁tail 1
+▁calm 1
+▁joined 1
+ship 1
+▁understand 1
+▁ancient 1
+▁kingdom 1
+▁chapel 1
+▁health 1
+▁touch 1
+▁cities 1
+▁save 1
+▁true 1
+▁shot 1
+▁cur 1
+▁treat 1
+▁bell 1
+▁though 1
+ub 1
+▁italian 1
+▁travel 1
+▁che 1
+▁former 1
+▁remain 1
+▁har 1
+▁bra 1
+ther 1
+▁word 1
+▁cold 1
+▁starts 1
+▁fi 1
+▁champion 1
+▁accident 1
+▁venezuela 1
+▁camp 1
+▁graduated 1
+▁host 1
+▁italy 1
+▁pen 1
+▁table 1
+▁peru 1
+▁super 1
+▁youth 1
+▁cal 1
+▁wide 1
+▁est 1
+▁throw 1
+▁japan 1
+▁matter 1
+▁silver 1
+▁mayor 1
+ros 1
+he 1
+▁gold 1
+ious 1
+▁blood 1
+▁literature 1
+▁primar 1
+▁outside 1
+▁maybe 1
+▁wi 1
+▁bank 1
+▁science 1
+▁novel 1
+▁sand 1
+▁cup 1
+▁store 1
+▁characters 1
+▁established 1
+▁appears 1
+▁buen 1
+▁movies 1
+▁site 1
+▁religious 1
+▁airport 1
+▁villa 1
+▁earth 1
+▁bur 1
+▁priest 1
+▁fell 1
+▁cut 1
+▁cast 1
+▁meeting 1
+▁rule 1
+act 1
+▁association 1
+▁environment 1
+▁festival 1
+▁individual 1
+▁santiago 1
+▁greek 1
+▁temple 1
+▁appearance 1
+▁market 1
+▁feed 1
+▁tri 1
+▁difficult 1
+▁secret 1
+▁agriculture 1
+▁replaced 1
+▁break 1
+▁writing 1
+▁rome 1
+▁formed 1
+▁media 1
+▁passed 1
+▁making 1
+▁typical 1
+▁numer 1
+▁collection 1
+▁spen 1
+▁fine 1
+▁defeated 1
+▁obtained 1
+▁appear 1
+ence 1
+ang 1
+▁christian 1
+car 1
+▁college 1
+▁least 1
+▁grew 1
+▁zone 1
+▁visit 1
+▁owner 1
+▁sha 1
+▁sport 1
+▁unit 1
+▁arms 1
+the 1
+▁colombia 1
+▁imp 1
+▁stand 1
+▁municipal 1
+▁themselves 1
+▁valencia 1
+▁devot 1
+▁mission 1
+▁hill 1
+▁added 1
+▁bit 1
+▁horse 1
+▁hear 1
+▁pale 1
+▁favor 1
+▁especially 1
+▁literary 1
+▁location 1
+▁avoid 1
+▁stopped 1
+▁worth 1
+▁remember 1
+▁designed 1
+▁moon 1
+▁fan 1
+▁problem 1
+ica 1
+▁pan 1
+▁text 1
+▁available 1
+▁independent 1
+▁probabl 1
+▁philosophy 1
+▁rep 1
+led 1
+▁listen 1
+ue 1
+▁personal 1
+▁economy 1
+▁positive 1
+▁twelve 1
+▁marriage 1
+▁ricardo 1
+▁spring 1
+▁initial 1
+▁action 1
+▁necessary 1
+▁ci 1
+▁led 1
+▁includ 1
+▁places 1
+▁stra 1
+▁deep 1
+▁focus 1
+▁urban 1
+▁pedro 1
+▁silent 1
+▁police 1
+▁award 1
+▁rich 1
+▁walk 1
+▁brothers 1
+▁doing 1
+▁arch 1
+▁industry 1
+▁spread 1
+▁captain 1
+▁crown 1
+▁studio 1
+▁money 1
+▁poet 1
+▁immediately 1
+▁none 1
+▁gra 1
+wise 1
+▁services 1
+▁non 1
+▁si 1
+▁results 1
+spite 1
+▁founded 1
+cha 1
+▁cemetery 1
+▁thousand 1
+▁channel 1
+▁institution 1
+ak 1
+▁reached 1
+▁eu 1
+▁continue 1
+lan 1
+▁flow 1
+▁l 1
+▁practice 1
+▁fifth 1
+▁minister 1
+▁data 1
+▁organized 1
+▁total 1
+▁thanks 1
+▁followed 1
+des 1
+ite 1
+▁sal 1
+▁function 1
+▁happened 1
+▁presented 1
+▁fast 1
+▁guitar 1
+▁future 1
+▁favorite 1
+▁route 1
+▁living 1
+▁guide 1
+▁latin 1
+▁sold 1
+nia 1
+▁matches 1
+▁elected 1
+▁completed 1
+▁sad 1
+▁tradition 1
+▁par 1
+▁elements 1
+▁grow 1
+▁bishop 1
+▁subject 1
+▁answer 1
+▁watch 1
+▁felt 1
+▁lady 1
+▁receive 1
+▁wa 1
+▁recognized 1
+eless 1
+▁aires 1
+▁wor 1
+▁beat 1
+▁mostly 1
+▁com 1
+▁cause 1
+▁foundation 1
+▁deck 1
+▁cra 1
+▁changed 1
+▁prison 1
+▁key 1
+▁palace 1
+▁southern 1
+▁duck 1
+▁pet 1
+▁seeds 1
+▁allowed 1
+▁fly 1
+▁opened 1
+hu 1
+▁symbol 1
+▁officer 1
+way 1
+▁arts 1
+▁decide 1
+kin 1
+▁showed 1
+▁mountains 1
+▁campaign 1
+▁mouth 1
+▁double 1
+▁taking 1
+▁afterwards 1
+ita 1
+▁talking 1
+▁yet 1
+▁tree 1
+▁valley 1
+ea 1
+ern 1
+▁parts 1
+▁approach 1
+▁view 1
+▁meet 1
+▁figure 1
+▁mor 1
+▁artistic 1
+▁founder 1
+▁problems 1
+▁hit 1
+▁produce 1
+bra 1
+▁says 1
+par 1
+▁ulises 1
+▁companies 1
+▁cases 1
+▁pick 1
+era 1
+▁awards 1
+▁singing 1
+▁lack 1
+▁additional 1
+▁pieces 1
+▁ven 1
+▁bro 1
+mar 1
+lon 1
+▁entire 1
+▁reviews 1
+▁trade 1
+▁wings 1
+▁ear 1
+▁specific 1
+▁respect 1
+▁foreign 1
+▁purpose 1
+▁uruguay 1
+▁facade 1
+▁medicine 1
+▁cabin 1
+▁antonio 1
+▁poetry 1
+cio 1
+▁golden 1
+▁specialized 1
+▁material 1
+▁doubt 1
+▁path 1
+▁burn 1
+▁uses 1
+▁build 1
+▁attended 1
+by 1
+ena 1
+▁fa 1
+ile 1
+▁administrative 1
+▁appointed 1
+▁august 1
+▁strength 1
+▁lord 1
+▁brought 1
+▁careful 1
+▁frequently 1
+▁upper 1
+▁forward 1
+ab 1
+▁mixed 1
+▁mal 1
+▁yes 1
+▁je 1
+▁offer 1
+▁write 1
+wn 1
+▁paul 1
+▁ob 1
+ez 1
+▁husband 1
+▁industrial 1
+▁instrument 1
+▁valeria 1
+▁architecture 1
+▁agreement 1
+▁finish 1
+▁versions 1
+▁wave 1
+▁publication 1
+▁ski 1
+▁happen 1
+▁tro 1
+▁forces 1
+▁significant 1
+▁words 1
+▁neighbor 1
+▁bolivia 1
+▁climate 1
+▁decision 1
+▁weather 1
+▁reference 1
+▁niña 1
+▁edge 1
+▁track 1
+▁boat 1
+▁journalist 1
+▁wine 1
+▁described 1
+▁suffered 1
+chi 1
+▁kid 1
+▁easter 1
+▁mr 1
+sis 1
+ola 1
+▁concept 1
+▁empire 1
+▁furthermore 1
+▁measure 1
+▁distance 1
+▁minor 1
+▁hot 1
+ever 1
+▁features 1
+▁mari 1
+▁regional 1
+ado 1
+▁refer 1
+▁angel 1
+▁signed 1
+▁english 1
+▁nobody 1
+▁machine 1
+▁multiple 1
+▁stadium 1
+▁spot 1
+▁nature 1
+▁northern 1
+▁medal 1
+▁governor 1
+▁producer 1
+▁students 1
+▁oil 1
+▁previously 1
+gar 1
+▁por 1
+▁discovered 1
+▁ask 1
+▁training 1
+▁fit 1
+▁pay 1
+▁sail 1
+▁box 1
+▁agricultural 1
+▁entrance 1
+▁society 1
+▁disease 1
+▁leaving 1
+▁bread 1
+▁david 1
+▁risk 1
+▁acting 1
+cent 1
+▁perform 1
+fo 1
+▁animal 1
+▁bird 1
+▁value 1
+ide 1
+over 1
+ugh 1
+▁ab 1
+▁surface 1
+▁monument 1
+▁sector 1
+▁vilma 1
+▁content 1
+▁search 1
+▁iron 1
+▁suddenly 1
+▁march 1
+ought 1
+▁coach 1
+▁fu 1
+▁meters 1
+▁rank 1
+▁languages 1
+▁huge 1
+▁walls 1
+▁steps 1
+▁hat 1
+▁players 1
+▁leon 1
+▁engine 1
+▁smile 1
+▁costa 1
+▁colors 1
+▁contribut 1
+▁architect 1
+▁congress 1
+▁block 1
+▁bubble 1
+▁guard 1
+▁manuel 1
+▁quickly 1
+▁continues 1
+▁abandoned 1
+▁complete 1
+▁press 1
+ving 1
+ons 1
+▁pla 1
+▁category 1
+▁thirty 1
+▁england 1
+▁interior 1
+▁phase 1
+▁warm 1
+ture 1
+▁sick 1
+form 1
+▁surrounded 1
+▁involved 1
+▁semi 1
+▁ready 1
+ulate 1
+▁limited 1
+▁val 1
+mina 1
+▁bear 1
+▁ju 1
+dad 1
+ker 1
+▁india 1
+az 1
+▁miss 1
+▁occur 1
+▁cinema 1
+▁cathedral 1
+▁impossible 1
+▁property 1
+▁biggest 1
+▁clean 1
+▁houses 1
+▁mistake 1
+▁sky 1
+▁lawyer 1
+thus 1
+▁mark 1
+▁secondary 1
+▁develop 1
+▁report 1
+▁speed 1
+▁skin 1
+▁stands 1
+ger 1
+▁stars 1
+▁read 1
+▁placed 1
+ña 1
+▁extreme 1
+▁administration 1
+▁scientific 1
+▁politics 1
+▁strange 1
+▁flag 1
+▁crater 1
+▁preserved 1
+▁buildings 1
+stood 1
+▁quite 1
+ju 1
+▁mont 1
+▁rein 1
+▁accept 1
+▁alive 1
+▁soil 1
+▁ban 1
+▁opening 1
+▁animals 1
+▁student 1
+ors 1
+▁za 1
+ida 1
+▁relation 1
+▁ainhoa 1
+▁ministry 1
+▁question 1
+ches 1
+▁childhood 1
+▁dress 1
+▁gather 1
+▁artists 1
+▁serve 1
+▁killed 1
+▁angeles 1
+▁musician 1
+▁mag 1
+▁nest 1
+ko 1
+▁resid 1
+▁legs 1
+ize 1
+▁dan 1
+▁holy 1
+▁characteriz 1
+▁create 1
+ative 1
+é 1
+▁aspect 1
+▁commission 1
+▁landscape 1
+▁library 1
+▁physical 1
+▁variety 1
+▁platform 1
+▁memory 1
+▁storm 1
+▁thick 1
+▁wall 1
+▁ball 1
+▁farm 1
+▁types 1
+▁van 1
+▁shaped 1
+▁drink 1
+▁accepted 1
+▁planet 1
+▁command 1
+▁recording 1
+▁sciences 1
+▁contain 1
+ano 1
+▁fil 1
+▁j 1
+▁wear 1
+▁object 1
+pu 1
+▁parliament 1
+▁permanent 1
+▁birth 1
+▁breath 1
+▁network 1
+▁organ 1
+▁fresh 1
+▁communication 1
+▁compete 1
+▁term 1
+▁script 1
+▁generation 1
+▁luck 1
+▁rejected 1
+▁legal 1
+▁drive 1
+▁fun 1
+rew 1
+▁participate 1
+▁asked 1
+▁remove 1
+time 1
+▁except 1
+▁ran 1
+▁correspond 1
+▁defense 1
+▁christmas 1
+▁william 1
+▁jump 1
+▁classified 1
+▁secretary 1
+▁join 1
+ify 1
+▁branches 1
+▁chilean 1
+▁palm 1
+com 1
+gan 1
+ward 1
+han 1
+▁soul 1
+▁gas 1
+▁starr 1
+cross 1
+▁ring 1
+▁neither 1
+▁contact 1
+▁debut 1
+▁bright 1
+▁canada 1
+▁bull 1
+mes 1
+umb 1
+▁achieved 1
+▁x 1
+▁mentioned 1
+▁sat 1
+▁tre 1
+▁cat 1
+ua 1
+▁release 1
+▁rica 1
+▁step 1
+ral 1
+▁cy 1
+▁suggest 1
+▁bottom 1
+▁candidate 1
+▁column 1
+▁conference 1
+▁energy 1
+▁everybody 1
+▁miguel 1
+▁constant 1
+▁enjoy 1
+▁gift 1
+▁olive 1
+▁discover 1
+▁judge 1
+▁ago 1
+▁slight 1
+▁treatment 1
+▁tourism 1
+▁liberal 1
+▁pop 1
+▁partner 1
+den 1
+▁flat 1
+ier 1
+▁suit 1
+▁haven 1
+▁metal 1
+▁neverth 1
+ification 1
+▁mad 1
+▁limit 1
+▁heard 1
+▁bill 1
+▁horses 1
+▁records 1
+▁prince 1
+▁source 1
+▁mine 1
+rry 1
+▁ecuador 1
+▁sweet 1
+▁weapon 1
+▁either 1
+▁instead 1
+▁northwest 1
+▁translated 1
+▁anymore 1
+▁wrong 1
+▁transport 1
+▁might 1
+▁soft 1
+▁southeast 1
+▁rural 1
+▁points 1
+▁deal 1
+▁piti 1
+▁higher 1
+▁flower 1
+▁rose 1
+▁promoted 1
+je 1
+ological 1
+▁films 1
+▁chest 1
+▁draw 1
+▁lay 1
+▁cost 1
+▁slow 1
+▁awarded 1
+▁tried 1
+burg 1
+▁virgin 1
+▁condition 1
+▁essential 1
+▁salvador 1
+▁reserve 1
+▁presence 1
+▁foot 1
+▁prize 1
+▁creation 1
+▁basic 1
+ified 1
+▁fra 1
+fer 1
+▁inhabit 1
+▁vote 1
+▁bal 1
+▁happens 1
+▁broke 1
+while 1
+board 1
+▁medi 1
+▁declared 1
+▁gamboa 1
+▁october 1
+▁promotion 1
+▁audience 1
+▁chinese 1
+▁computer 1
+▁request 1
+▁smell 1
+▁composition 1
+▁fishing 1
+▁extra 1
+▁editor 1
+▁piano 1
+▁fuck 1
+▁surrounding 1
+▁send 1
+▁vice 1
+bre 1
+rin 1
+rous 1
+▁disappeared 1
+side 1
+▁require 1
+work 1
+▁positions 1
+▁disc 1
+▁monte 1
+▁midfielder 1
+▁cream 1
+▁federal 1
+▁southwest 1
+▁taught 1
+▁contest 1
+▁medium 1
+▁moving 1
+▁truth 1
+▁democratic 1
+▁characteristic 1
+▁lope 1
+▁qui 1
+▁highest 1
+▁product 1
+▁forms 1
+▁mary 1
+▁arrive 1
+▁direction 1
+▁ends 1
+▁ice 1
+▁rise 1
+▁clo 1
+▁bed 1
+▁escape 1
+▁card 1
+▁streets 1
+▁upon 1
+▁products 1
+▁fre 1
+▁avenue 1
+▁families 1
+▁financial 1
+▁heavy 1
+▁knowledge 1
+▁unique 1
+▁bodies 1
+▁morning 1
+▁below 1
+▁brand 1
+▁tourist 1
+▁chance 1
+▁pink 1
+▁contract 1
+▁grey 1
+▁hide 1
+▁prove 1
+ual 1
+▁launch 1
+ario 1
+▁reasons 1
+▁forty 1
+▁levels 1
+▁manage 1
+▁gre 1
+▁sil 1
+dor 1
+▁pain 1
+▁regular 1
+▁changes 1
+▁robert 1
+▁minutes 1
+▁revolution 1
+▁conflict 1
+▁jewish 1
+▁prevent 1
+▁guinea 1
+▁weird 1
+▁fac 1
+▁quiet 1
+▁bigger 1
+▁ramon 1
+▁glo 1
+▁princess 1
+▁gives 1
+▁introduced 1
+▁gal 1
+▁mode 1
+▁eh 1
+▁contains 1
+matic 1
+▁affected 1
+can 1
+▁fin 1
+▁cho 1
+▁marked 1
+▁ur 1
+▁russia 1
+▁buy 1
+▁mrs 1
+lla 1
+ito 1
+▁valle 1
+▁kings 1
+old 1
+ham 1
+▁peace 1
+▁account 1
+▁distribution 1
+▁michael 1
+▁participation 1
+▁twice 1
+▁welcome 1
+▁chief 1
+▁orange 1
+▁lips 1
+▁sense 1
+▁taste 1
+▁fernando 1
+▁older 1
+▁destroyed 1
+▁ride 1
+▁schools 1
+▁connect 1
+▁fo 1
+ade 1
+ji 1
+▁wing 1
+▁min 1
+▁christ 1
+▁format 1
+▁windows 1
+▁offered 1
+one 1
+▁stayed 1
+▁advance 1
+▁mediterranean 1
+▁detail 1
+▁roof 1
+▁running 1
+▁successful 1
+▁patron 1
+cho 1
+▁hurry 1
+▁olympic 1
+▁fair 1
+▁romantic 1
+▁george 1
+▁vas 1
+▁zo 1
+▁noble 1
+ev 1
+▁voices 1
+▁attend 1
+▁hey 1
+house 1
+▁issue 1
+▁directly 1
+▁convent 1
+▁argentinian 1
+▁independence 1
+▁prominent 1
+▁screen 1
+▁easy 1
+▁milk 1
+▁above 1
+▁objective 1
+▁management 1
+▁bay 1
+▁associated 1
+▁million 1
+▁reduced 1
+▁adventure 1
+▁occasions 1
+▁gun 1
+▁sing 1
+▁granted 1
+▁window 1
+▁linked 1
+ella 1
+▁resources 1
+wa 1
+▁char 1
+▁bus 1
+▁broadcast 1
+lio 1
+▁increase 1
+▁whom 1
+▁ana 1
+▁add 1
+▁saying 1
+▁characteristics 1
+▁citizen 1
+▁accompani 1
+▁contemporar 1
+▁experiment 1
+▁subtropical 1
+▁portrait 1
+▁troops 1
+▁soap 1
+▁acid 1
+ography 1
+▁anyone 1
+▁weeks 1
+▁systems 1
+▁parish 1
+▁arrested 1
+▁longer 1
+▁sex 1
+▁vo 1
+▁col 1
+▁marine 1
+▁speak 1
+▁parties 1
+▁feature 1
+▁mc 1
+dia 1
+▁recover 1
+▁dot 1
+▁promote 1
+▁particular 1
+▁plane 1
+▁glass 1
+▁feeling 1
+▁century 1
+▁inspired 1
+▁miracle 1
+▁restaurant 1
+▁september 1
+▁exchange 1
+▁hidden 1
+▁humid 1
+▁pretty 1
+▁root 1
+▁pilot 1
+▁flight 1
+▁paper 1
+▁pair 1
+▁reform 1
+ura 1
+▁compar 1
+▁chain 1
+▁victory 1
+▁seasons 1
+▁peter 1
+▁mind 1
+▁shin 1
+▁expected 1
+▁theory 1
+▁workers 1
+▁pictures 1
+▁republican 1
+▁gall 1
+▁obtain 1
+▁establish 1
+▁fruits 1
+▁believed 1
+▁nave 1
+▁kar 1
+▁bla 1
+▁kids 1
+▁describe 1
+▁normal 1
+▁insects 1
+berg 1
+▁marc 1
+▁orchestra 1
+▁portugal 1
+▁quality 1
+▁composer 1
+▁nun 1
+▁wild 1
+▁intense 1
+▁forget 1
+ño 1
+▁wish 1
+▁gen 1
+▁safe 1
+▁ever 1
+bar 1
+▁carry 1
+▁magazines 1
+▁july 1
+▁leg 1
+▁ko 1
+▁effect 1
+▁bag 1
+▁eye 1
+▁mill 1
+▁salt 1
+▁disappear 1
+▁bru 1
+▁mix 1
+ari 1
+ov 1
+▁sell 1
+▁sch 1
+▁perfect 1
+▁app 1
+▁poly 1
+▁lyrics 1
+▁fat 1
+▁sides 1
+▁waiting 1
+▁attempt 1
+▁grandfather 1
+▁religion 1
+▁vehicle 1
+▁alright 1
+▁livestock 1
+▁importance 1
+▁stem 1
+▁james 1
+▁wake 1
+▁occupied 1
+▁grave 1
+▁cry 1
+▁formation 1
+▁lands 1
+▁scar 1
+▁topic 1
+▁drawing 1
+rio 1
+▁plate 1
+▁asking 1
+▁murder 1
+▁grand 1
+▁catch 1
+ze 1
+rce 1
+▁authors 1
+▁hour 1
+▁ec 1
+▁novels 1
+▁sur 1
+▁rosa 1
+▁elections 1
+▁surviv 1
+▁apart 1
+▁alternative 1
+▁consequence 1
+▁effort 1
+▁monastery 1
+▁security 1
+▁custom 1
+▁louis 1
+▁ramiro 1
+▁volume 1
+▁combat 1
+▁naked 1
+▁forgive 1
+▁further 1
+▁temperature 1
+▁catalan 1
+▁martin 1
+▁prefer 1
+▁multi 1
+▁fiction 1
+▁hydro 1
+▁engineer 1
+▁sources 1
+▁thin 1
+▁defined 1
+lli 1
+▁representation 1
+▁fear 1
+▁painter 1
+▁scale 1
+▁shit 1
+▁extended 1
+▁widely 1
+▁gri 1
+▁sca 1
+▁earn 1
+▁scared 1
+▁scene 1
+▁piece 1
+▁flying 1
+▁effects 1
+▁issues 1
+▁mac 1
+graph 1
+▁figures 1
+▁argentine 1
+▁diocese 1
+▁rescue 1
+▁website 1
+▁itself 1
+▁narrow 1
+▁internal 1
+▁friendship 1
+▁programs 1
+▁rob 1
+▁plain 1
+▁subsequently 1
+▁presents 1
+▁recently 1
+▁page 1
+▁adopted 1
+▁thinking 1
+▁else 1
+▁spa 1
+▁bern 1
+▁file 1
+▁covers 1
+▁launched 1
+▁ham 1
+▁q 1
+q 1
+ick 1
+▁materials 1
+zi 1
+▁gene 1
+▁scor 1
+cular 1
+▁application 1
+▁decorated 1
+▁garcía 1
+▁knew 1
+▁interesting 1
+▁inhabitants 1
+▁self 1
+▁offers 1
+▁eventual 1
+▁feet 1
+▁retired 1
+▁epi 1
+▁collaborated 1
+▁realiz 1
+▁guy 1
+▁score 1
+▁han 1
+▁reading 1
+udi 1
+▁connected 1
+▁solo 1
+▁document 1
+▁telling 1
+▁visited 1
+▁express 1
+▁coat 1
+▁code 1
+▁politician 1
+▁veracruz 1
+▁victim 1
+▁fifty 1
+▁herself 1
+▁crime 1
+▁swimming 1
+▁pacific 1
+▁sculpture 1
+▁giving 1
+▁hotel 1
+▁franco 1
+▁mount 1
+▁check 1
+▁hello 1
+▁share 1
+▁esp 1
+▁exile 1
+cion 1
+▁max 1
+▁executed 1
+▁sam 1
+▁teaching 1
+▁amount 1
+▁drop 1
+▁protected 1
+▁sales 1
+▁projects 1
+▁cell 1
+illo 1
+▁ideal 1
+▁spe 1
+▁corn 1
+ada 1
+▁gate 1
+▁damage 1
+▁african 1
+▁universi 1
+▁afraid 1
+▁outstanding 1
+▁provincial 1
+▁nominated 1
+▁shadow 1
+▁legend 1
+▁puerto 1
+▁blow 1
+▁device 1
+▁okay 1
+▁snow 1
+▁impact 1
+▁journalism 1
+▁criticism 1
+port 1
+▁nice 1
+▁lock 1
+▁hang 1
+▁cave 1
+▁bat 1
+▁slowly 1
+▁failed 1
+▁publish 1
+▁runs 1
+▁shown 1
+▁articles 1
+▁height 1
+▁historic 1
+▁roads 1
+▁defeat 1
+▁suffer 1
+▁lose 1
+▁magic 1
+▁achieve 1
+▁cam 1
+▁climb 1
+▁digital 1
+▁evidence 1
+▁excellent 1
+▁renaissance 1
+▁quarter 1
+▁wedding 1
+▁shut 1
+▁panama 1
+▁correct 1
+▁assistant 1
+▁northeast 1
+▁dangerous 1
+▁quit 1
+▁factor 1
+▁authority 1
+▁photograph 1
+▁distinguished 1
+▁assembly 1
+▁row 1
+pped 1
+▁pot 1
+▁painted 1
+▁colored 1
+▁tend 1
+▁corner 1
+vision 1
+▁begins 1
+uro 1
+▁loved 1
+ach 1
+▁gone 1
+▁mention 1
+ility 1
+▁rio 1
+▁separate 1
+stan 1
+▁operation 1
+field 1
+▁advantage 1
+▁explain 1
+▁principle 1
+▁wound 1
+▁injured 1
+▁choose 1
+▁plastic 1
+▁railway 1
+▁loud 1
+▁desert 1
+▁tired 1
+▁settlement 1
+▁conduct 1
+▁diego 1
+▁transferr 1
+▁guest 1
+▁neck 1
+▁meaning 1
+▁acquired 1
+▁eggs 1
+▁shore 1
+▁manager 1
+phe 1
+▁bow 1
+▁regularly 1
+▁tries 1
+ville 1
+▁maintain 1
+▁hole 1
+▁n 1
+▁cuba 1
+▁concerts 1
+▁hug 1
+▁ray 1
+▁models 1
+▁poems 1
+ford 1
+▁interview 1
+▁rodriguez 1
+▁violent 1
+▁april 1
+▁vocal 1
+▁nation 1
+▁resistance 1
+▁silence 1
+▁proposed 1
+▁expression 1
+▁trained 1
+▁separated 1
+▁greatest 1
+▁cook 1
+lie 1
+▁smiled 1
+▁survive 1
+▁ze 1
+win 1
+▁walking 1
+emi 1
+▁courses 1
+▁adults 1
+▁fund 1
+▁bone 1
+▁occasion 1
+ku 1
+▁headquarter 1
+water 1
+▁branch 1
+▁sit 1
+▁responsib 1
+▁manufacture 1
+▁deputy 1
+▁crazy 1
+▁reddish 1
+▁teeth 1
+▁daily 1
+▁status 1
+▁sonata 1
+▁executive 1
+▁emerg 1
+▁task 1
+▁talent 1
+▁bought 1
+▁stick 1
+▁interpretation 1
+▁differences 1
+▁trust 1
+ight 1
+▁herb 1
+▁victor 1
+▁heat 1
+▁guys 1
+▁clearly 1
+▁exhibitions 1
+▁meat 1
+▁dies 1
+audi 1
+▁expect 1
+▁driv 1
+▁survived 1
+▁moral 1
+▁gua 1
+▁tru 1
+lot 1
+▁recent 1
+▁generat 1
+ivity 1
+á 1
+▁ceremony 1
+▁comfort 1
+▁constitute 1
+▁easily 1
+▁february 1
+▁ghost 1
+▁solution 1
+▁swear 1
+▁terminal 1
+▁superior 1
+▁afternoon 1
+▁jorge 1
+▁vegetation 1
+▁jesus 1
+▁straight 1
+▁spoke 1
+▁crossing 1
+▁demand 1
+▁ruins 1
+▁brick 1
+▁filled 1
+▁sho 1
+▁wheel 1
+▁conditions 1
+▁justice 1
+▁flora 1
+▁rebel 1
+▁majority 1
+▁investigation 1
+▁demo 1
+▁staff 1
+▁tele 1
+▁reality 1
+▁settled 1
+▁youngest 1
+▁hor 1
+▁particularly 1
+▁cab 1
+▁socialist 1
+▁sto 1
+▁practical 1
+▁roll 1
+▁larges 1
+▁grown 1
+ake 1
+ai 1
+ory 1
+▁yo 1
+aya 1
+▁formal 1
+agua 1
+▁reflect 1
+▁difference 1
+▁paraguay 1
+▁aim 1
+▁soldiers 1
+ologist 1
+▁dimension 1
+▁extension 1
+▁length 1
+▁yourself 1
+▁fifteen 1
+▁cruz 1
+▁electronic 1
+▁universities 1
+▁dinner 1
+▁comprise 1
+▁winning 1
+▁basque 1
+▁opposite 1
+▁broad 1
+▁richard 1
+▁standard 1
+▁load 1
+▁anyway 1
+▁auto 1
+▁web 1
+▁weight 1
+▁presentation 1
+▁grass 1
+ex 1
+▁trial 1
+▁diplomat 1
+▁commander 1
+istic 1
+off 1
+▁removed 1
+▁scenes 1
+▁defend 1
+▁actions 1
+▁severe 1
+▁lee 1
+▁bare 1
+▁shoot 1
+▁indicate 1
+▁laugh 1
+▁studying 1
+▁progress 1
+▁spirit 1
+▁behavior 1
+▁emperor 1
+▁expedition 1
+▁refuge 1
+▁nucle 1
+▁poison 1
+▁baroque 1
+▁sierra 1
+▁cattle 1
+▁historian 1
+▁internet 1
+▁evil 1
+▁daniel 1
+▁failure 1
+▁residence 1
+▁soldier 1
+▁isabel 1
+▁authorities 1
+▁protection 1
+▁crop 1
+▁comment 1
+▁growth 1
+▁causes 1
+▁frame 1
+onic 1
+▁june 1
+▁syn 1
+▁captured 1
+▁fixed 1
+▁normally 1
+▁attacked 1
+▁prior 1
+hen 1
+▁spend 1
+▁protect 1
+▁op 1
+iz 1
+▁prepar 1
+▁tall 1
+▁teach 1
+▁philosoph 1
+▁serious 1
+▁frequent 1
+ient 1
+▁pin 1
+▁govern 1
+▁background 1
+▁crowd 1
+▁equivalent 1
+▁extensive 1
+▁irregular 1
+▁negative 1
+▁nephew 1
+▁november 1
+▁remarkable 1
+▁analysis 1
+▁myself 1
+▁mysterious 1
+▁texas 1
+▁attention 1
+▁catalonia 1
+▁varied 1
+▁premier 1
+▁estela 1
+▁visual 1
+▁berlin 1
+▁attach 1
+▁tonight 1
+▁hurt 1
+▁rosario 1
+▁option 1
+▁bullet 1
+▁echo 1
+▁labor 1
+▁label 1
+▁shop 1
+▁passion 1
+▁sailor 1
+ably 1
+▁rarely 1
+▁edited 1
+▁factory 1
+▁operations 1
+▁stable 1
+▁slave 1
+▁camera 1
+▁disco 1
+▁recovered 1
+▁practiced 1
+▁observ 1
+▁pine 1
+▁holding 1
+cut 1
+▁bio 1
+ati 1
+▁needed 1
+▁mess 1
+▁qua 1
+▁comic 1
+ira 1
+▁hunt 1
+graphic 1
+▁danger 1
+▁descend 1
+▁protest 1
+back 1
+▁december 1
+▁judicial 1
+▁perennial 1
+▁purple 1
+▁vietnam 1
+▁zaragoza 1
+▁expand 1
+▁expert 1
+▁fourteen 1
+▁translation 1
+▁goodbye 1
+▁vicente 1
+▁enrique 1
+▁micro 1
+▁annual 1
+▁patient 1
+▁anchor 1
+▁latter 1
+▁wealth 1
+▁steve 1
+▁sexual 1
+▁altar 1
+▁destination 1
+▁colo 1
+▁debate 1
+▁whatever 1
+▁element 1
+▁concert 1
+▁rare 1
+ney 1
+▁election 1
+▁adapted 1
+▁signal 1
+▁increased 1
+▁classes 1
+clan 1
+▁owned 1
+▁psycho 1
+▁starting 1
+▁speaking 1
+▁net 1
+▁rev 1
+▁farming 1
+▁missing 1
+▁habit 1
+▁victoria 1
+▁reported 1
+▁sentence 1
+light 1
+▁pat 1
+▁celebrate 1
+▁martín 1
+▁asleep 1
+▁athletic 1
+▁eduardo 1
+▁enemy 1
+▁external 1
+▁diverse 1
+▁chamber 1
+▁senator 1
+▁fierce 1
+▁interested 1
+▁rival 1
+circular 1
+▁statue 1
+ice 1
+▁worry 1
+▁jack 1
+▁convention 1
+▁servants 1
+▁succeeded 1
+▁jr 1
+▁mau 1
+▁cuban 1
+▁equal 1
+▁gro 1
+▁mid 1
+▁whispered 1
+▁documents 1
+▁rat 1
+▁et 1
+▁letters 1
+▁illustrat 1
+▁ox 1
+▁photo 1
+▁mono 1
+called 1
+wood 1
+mbo 1
+▁traveled 1
+▁toys 1
+▁display 1
+▁gonzalez 1
+▁immigrant 1
+▁inflorescence 1
+▁journey 1
+▁metropolitan 1
+▁supposed 1
+▁technology 1
+▁thailand 1
+▁component 1
+▁license 1
+▁opinion 1
+▁botanic 1
+▁humble 1
+▁radical 1
+▁everyday 1
+▁classification 1
+▁basin 1
+▁collective 1
+▁decoration 1
+▁comedy 1
+▁celebration 1
+▁documentary 1
+▁exact 1
+bert 1
+▁grace 1
+▁connection 1
+mont 1
+▁glasses 1
+ache 1
+▁smaller 1
+▁colombian 1
+▁fr 1
+▁anger 1
+▁repeated 1
+▁chart 1
+▁celebrated 1
+▁publishing 1
+▁gray 1
+▁broadcasted 1
+▁officially 1
+▁esc 1
+▁standing 1
+▁dar 1
+▁pull 1
+▁hearing 1
+▁whisper 1
+▁link 1
+▁fingers 1
+well 1
+▁rent 1
+▁promise 1
+▁surprise 1
+▁remembered 1
+▁zi 1
+▁medic 1
+ense 1
+▁bachelor 1
+▁demonstrat 1
+▁fulfill 1
+▁incident 1
+▁lieutenant 1
+▁territories 1
+▁essay 1
+▁blind 1
+▁cheese 1
+▁charles 1
+▁reaction 1
+ette 1
+▁zero 1
+▁bushes 1
+▁ken 1
+▁theatre 1
+▁brain 1
+▁beard 1
+▁adam 1
+▁defender 1
+ball 1
+▁laser 1
+▁confirmed 1
+▁juli 1
+▁lima 1
+▁claimed 1
+gne 1
+▁simpl 1
+▁cable 1
+▁prime 1
+▁rib 1
+▁sar 1
+▁passengers 1
+▁vol 1
+▁ram 1
+▁egypt 1
+▁emigrat 1
+▁kilometer 1
+▁physics 1
+▁principal 1
+▁recognition 1
+▁rectangular 1
+▁rhythm 1
+▁thomas 1
+▁tomorrow 1
+▁turkey 1
+▁imperial 1
+▁suprem 1
+▁parallel 1
+▁shield 1
+▁deserve 1
+▁workshop 1
+▁freedom 1
+▁dispute 1
+▁admir 1
+▁uncle 1
+▁criticized 1
+▁descendant 1
+▁finance 1
+▁pressure 1
+▁semifinal 1
+▁funny 1
+▁dawn 1
+▁twin 1
+▁understood 1
+▁brazilian 1
+▁prisoners 1
+▁critic 1
+▁striker 1
+▁price 1
+▁graphic 1
+anda 1
+▁roberto 1
+▁shooting 1
+▁sab 1
+▁adult 1
+▁required 1
+▁review 1
+▁spite 1
+▁pala 1
+▁sum 1
+▁moor 1
+aria 1
+anza 1
+▁map 1
+mann 1
+▁notice 1
+▁extend 1
+▁abandon 1
+▁sang 1
+▁grant 1
+▁sever 1
+▁deposit 1
+elia 1
+▁capture 1
+ability 1
+▁vegetables 1
+▁usual 1
+j 1
+▁inherit 1
+▁abundant 1
+▁circle 1
+▁exercise 1
+▁florida 1
+▁holiday 1
+▁january 1
+▁nickname 1
+▁peninsula 1
+▁switzerland 1
+▁eleven 1
+▁federation 1
+▁peruvian 1
+▁concern 1
+▁resign 1
+▁identity 1
+▁drown 1
+▁applied 1
+▁ahead 1
+▁sample 1
+▁smoke 1
+▁solve 1
+▁boston 1
+▁frank 1
+▁collaboration 1
+▁pool 1
+▁loose 1
+▁sugar 1
+▁vega 1
+▁nearby 1
+▁plaza 1
+▁seminar 1
+▁percent 1
+▁effective 1
+▁duke 1
+iness 1
+▁crew 1
+▁alberto 1
+▁shoe 1
+itu 1
+▁sixth 1
+▁elect 1
+▁surprised 1
+▁critical 1
+▁butler 1
+ago 1
+▁dancer 1
+▁affairs 1
+▁bark 1
+▁prepared 1
+bor 1
+bel 1
+▁alter 1
+▁philippines 1
+▁dam 1
+▁tied 1
+▁inhabited 1
+pen 1
+▁claim 1
+▁affect 1
+▁nin 1
+▁hip 1
+▁korea 1
+oz 1
+▁investigat 1
+▁apparent 1
+▁coordinat 1
+▁dominican 1
+▁entertainment 1
+▁familiar 1
+▁proposal 1
+▁siblings 1
+▁smooth 1
+▁vault 1
+▁relief 1
+▁monarch 1
+▁harvest 1
+▁witness 1
+▁basketball 1
+▁fortress 1
+▁traffic 1
+▁address 1
+▁reign 1
+▁steel 1
+▁gothic 1
+▁angry 1
+▁crist 1
+▁flock 1
+▁bush 1
+▁choir 1
+▁apartment 1
+▁shi 1
+▁deliver 1
+▁exception 1
+▁ability 1
+▁registered 1
+▁spin 1
+▁entit 1
+grad 1
+dent 1
+▁sala 1
+▁originate 1
+▁tar 1
+▁engineering 1
+▁coin 1
+ung 1
+ille 1
+▁closer 1
+▁rub 1
+▁sco 1
+▁jud 1
+lice 1
+▁involve 1
+▁chose 1
+▁graduate 1
+▁quick 1
+▁collaborate 1
+▁precise 1
+▁accused 1
+▁ambassador 1
+▁autonomous 1
+▁bloom 1
+▁capacity 1
+▁context 1
+▁córdoba 1
+▁evolution 1
+▁fundamental 1
+▁merchant 1
+▁phenomenon 1
+▁potential 1
+▁specimen 1
+▁sponsor 1
+▁substitute 1
+▁washington 1
+▁yesterday 1
+▁bronze 1
+▁confess 1
+▁indonesia 1
+▁granada 1
+▁greece 1
+▁mechanic 1
+▁mobile 1
+▁beauty 1
+▁potato 1
+▁navy 1
+▁varies 1
+▁shirt 1
+▁agency 1
+▁gallery 1
+▁partially 1
+▁ocean 1
+▁japanese 1
+▁andalusia 1
+▁barrel 1
+▁gym 1
+▁alternate 1
+▁confrontation 1
+▁tank 1
+▁israel 1
+▁bomb 1
+▁peña 1
+▁henry 1
+▁unfortunately 1
+▁plot 1
+▁silk 1
+▁scent 1
+acion 1
+▁combined 1
+atory 1
+▁fleet 1
+▁congo 1
+ctually 1
+▁refused 1
+▁photography 1
+▁agree 1
+▁ranch 1
+▁pure 1
+▁beak 1
+▁poll 1
+eval 1
+▁ple 1
+▁bass 1
+▁joy 1
+▁rocky 1
+rich 1
+▁stated 1
+▁bite 1
+pan 1
+▁artificial 1
+▁flood 1
+▁hungry 1
+▁mechanism 1
+▁murmur 1
+▁properties 1
+▁spaniard 1
+▁underground 1
+▁variable 1
+▁chicago 1
+▁feather 1
+▁felipe 1
+▁pablo 1
+▁cameroon 1
+▁colonel 1
+▁denmark 1
+▁cloud 1
+▁spider 1
+▁shaking 1
+▁copies 1
+▁cousin 1
+▁expos 1
+▁ji 1
+▁occasionally 1
+▁municipalities 1
+▁approximately 1
+▁wow 1
+▁basis 1
+▁powerful 1
+▁yellowish 1
+▁scarce 1
+▁chair 1
+▁pad 1
+▁identified 1
+▁popularity 1
+east 1
+▁charm 1
+▁vest 1
+▁damn 1
+▁trace 1
+▁shared 1
+▁fur 1
+▁broken 1
+bound 1
+▁article 1
+otic 1
+hood 1
+▁churches 1
+▁letter 1
+▁profession 1
+▁counter 1
+▁possibl 1
+▁success 1
+ñ 1
+▁interpret 1
+▁gain 1
+▁francis 1
+▁integra 1
+▁oppos 1
+▁predominant 1
+▁attitude 1
+▁categories 1
+▁coffee 1
+▁criminal 1
+▁montevideo 1
+▁moscow 1
+▁sweden 1
+▁uncertain 1
+▁violence 1
+▁nacional 1
+▁prevail 1
+▁junior 1
+▁intended 1
+▁flash 1
+▁folk 1
+▁clothes 1
+▁intention 1
+▁firm 1
+▁patriot 1
+▁sharp 1
+▁invited 1
+▁vacation 1
+▁canon 1
+▁tears 1
+▁desire 1
+▁message 1
+▁noise 1
+▁existence 1
+▁travelled 1
+▁progressive 1
+▁glory 1
+▁tribe 1
+▁rough 1
+▁commune 1
+▁yu 1
+mark 1
+▁paid 1
+▁peak 1
+▁shall 1
+▁tang 1
+▁perfectly 1
+luc 1
+▁entirely 1
+stein 1
+▁drinking 1
+▁exhibition 1
+▁farmer 1
+▁horn 1
+▁lover 1
+▁swa 1
+war 1
+▁austria 1
+known 1
+▁drama 1
+▁challenge 1
+▁circuit 1
+▁continuous 1
+▁decrease 1
+▁facilities 1
+▁girlfriend 1
+▁happiness 1
+▁legislative 1
+▁nicaragua 1
+▁scratch 1
+▁suicide 1
+▁universe 1
+▁visible 1
+▁stupid 1
+▁atlantic 1
+▁possibility 1
+▁caught 1
+▁crystal 1
+▁monkey 1
+▁leaf 1
+▁moderate 1
+▁wrapp 1
+▁shopping 1
+▁fool 1
+▁ireland 1
+▁merced 1
+▁shelter 1
+▁convert 1
+▁dying 1
+▁educated 1
+▁descent 1
+▁honey 1
+▁jackson 1
+▁initiative 1
+▁passionate 1
+▁specifically 1
+▁resist 1
+▁airline 1
+▁deportivo 1
+▁selection 1
+▁clock 1
+▁sheets 1
+▁iran 1
+▁restored 1
+▁fever 1
+▁seriously 1
+▁via 1
+▁determined 1
+ista 1
+▁threatened 1
+▁hunting 1
+uda 1
+▁larger 1
+▁killer 1
+▁lean 1
+▁joint 1
+▁australian 1
+▁picture 1
+▁tape 1
+▁mom 1
+▁killing 1
+▁cata 1
+▁organize 1
+▁attract 1
+▁hate 1
+▁mario 1
+▁bradomin 1
+▁committed 1
+▁communities 1
+▁explore 1
+▁gabriel 1
+▁guatemala 1
+▁portugues 1
+▁rafael 1
+▁sacred 1
+▁salamanca 1
+▁volunteer 1
+▁archeological 1
+▁erect 1
+▁worldwide 1
+▁canadian 1
+▁profile 1
+▁knight 1
+▁drunk 1
+▁drug 1
+▁margin 1
+▁torture 1
+▁forgotten 1
+▁solid 1
+▁orbit 1
+▁dominate 1
+▁sym 1
+▁electora 1
+▁milan 1
+▁grape 1
+▁stream 1
+▁preparation 1
+▁praised 1
+room 1
+▁tax 1
+▁announced 1
+fici 1
+▁oak 1
+▁marquise 1
+licate 1
+▁constitutional 1
+▁shade 1
+▁assist 1
+▁promised 1
+▁agent 1
+▁highlighted 1
+▁revealed 1
+▁geographical 1
+▁georgia 1
+▁significantly 1
+▁bath 1
+▁maintained 1
+mine 1
+▁sandy 1
+▁invad 1
+▁growing 1
+▁valu 1
+▁begin 1
+▁pav 1
+▁learning 1
+▁ross 1
+▁destroy 1
+▁divid 1
+▁designer 1
+▁devil 1
+▁paint 1
+▁raise 1
+▁shoulder 1
+▁subsequent 1
+writer 1
+▁strike 1
+▁casa 1
+▁increasing 1
+▁administer 1
+▁alexander 1
+▁anybody 1
+▁bulgaria 1
+▁combination 1
+▁commerce 1
+▁hondura 1
+▁imprisoned 1
+▁indigenous 1
+▁interrupt 1
+▁mirror 1
+▁nicolas 1
+▁temporary 1
+▁thirteen 1
+▁trembling 1
+▁athlete 1
+▁reputation 1
+▁zealand 1
+▁chemical 1
+▁fossil 1
+▁dismiss 1
+▁schedule 1
+▁pioneer 1
+▁strateg 1
+▁cycle 1
+▁split 1
+▁occurred 1
+▁naval 1
+▁martínez 1
+▁engag 1
+▁swamp 1
+▁execution 1
+▁adaptation 1
+▁saving 1
+▁arise 1
+▁eliminated 1
+▁kick 1
+▁zu 1
+▁colonial 1
+▁alex 1
+▁operated 1
+▁floating 1
+▁volcano 1
+▁rice 1
+▁bond 1
+▁otherwise 1
+▁paz 1
+▁indicated 1
+ius 1
+▁referr 1
+▁beaches 1
+▁ger 1
+▁observed 1
+▁whe 1
+▁wan 1
+▁listening 1
+▁repeat 1
+logy 1
+▁fail 1
+▁stepp 1
+▁formula 1
+▁constitution 1
+▁arc 1
+▁narrat 1
+▁cooper 1
+▁swim 1
+▁assigned 1
+▁carolina 1
+▁compound 1
+▁definition 1
+▁destruction 1
+▁dissolv 1
+▁encourage 1
+▁fountain 1
+▁manuscript 1
+▁phrase 1
+▁recipe 1
+▁regulation 1
+▁residential 1
+▁saturday 1
+▁scientist 1
+▁somebody 1
+▁sword 1
+▁brief 1
+▁convince 1
+▁crisis 1
+▁guzman 1
+▁predator 1
+▁situated 1
+▁ticket 1
+▁communist 1
+▁consumed 1
+▁jazz 1
+▁memories 1
+▁divine 1
+▁repair 1
+▁vienna 1
+▁architectural 1
+▁snake 1
+▁conquer 1
+▁shark 1
+▁isolated 1
+▁throne 1
+▁rumor 1
+▁conservative 1
+▁preferred 1
+▁strict 1
+▁session 1
+▁programming 1
+▁incorporated 1
+▁racing 1
+▁creative 1
+▁logic 1
+▁simon 1
+▁exam 1
+▁cancer 1
+▁estate 1
+▁lineage 1
+▁attraction 1
+izing 1
+▁cake 1
+▁sheep 1
+▁exhibited 1
+▁selected 1
+▁vision 1
+▁sentenced 1
+▁jun 1
+▁marta 1
+▁damaged 1
+▁preced 1
+▁tie 1
+▁mining 1
+duct 1
+▁liber 1
+▁grande 1
+▁bou 1
+ong 1
+▁harm 1
+▁stripe 1
+▁lion 1
+▁brave 1
+▁reject 1
+▁journal 1
+▁jer 1
+▁verb 1
+▁faith 1
+▁exhaust 1
+▁advertising 1
+▁ceramic 1
+▁circumstance 1
+▁confused 1
+▁cultivated 1
+▁frigate 1
+▁gradually 1
+▁intellectual 1
+▁proclaim 1
+▁response 1
+▁soundtrack 1
+▁treasure 1
+▁triumph 1
+▁cruel 1
+▁domingo 1
+▁uniform 1
+▁unlike 1
+▁description 1
+▁navigat 1
+▁statistic 1
+▁birthday 1
+▁mythology 1
+▁serbia 1
+▁cliff 1
+▁flour 1
+▁funeral 1
+▁nurse 1
+▁curved 1
+▁gulf 1
+▁arabia 1
+▁lectur 1
+▁oxford 1
+▁bedroom 1
+▁diagram 1
+▁loyal 1
+▁clinic 1
+▁polish 1
+▁contrary 1
+▁speech 1
+▁retirement 1
+haps 1
+▁motiv 1
+▁baby 1
+▁giant 1
+▁divorced 1
+▁alfredo 1
+▁trick 1
+▁archive 1
+▁teenager 1
+▁dishes 1
+▁candle 1
+▁cult 1
+▁assume 1
+▁motor 1
+▁manner 1
+▁advisor 1
+▁vary 1
+▁sighed 1
+▁tomb 1
+arrow 1
+▁apple 1
+bbing 1
+▁mist 1
+▁laughing 1
+▁fig 1
+▁deeply 1
+▁wire 1
+▁defended 1
+▁strip 1
+▁hal 1
+▁pont 1
+▁bob 1
+▁pig 1
+▁sight 1
+▁murdered 1
+▁reader 1
+▁layer 1
+▁pope 1
+▁wr 1
+▁reveal 1
+▁breed 1
+▁emblem 1
+where 1
+▁dig 1
+▁tale 1
+▁abroad 1
+▁absolute 1
+▁aircraft 1
+▁benefit 1
+▁boyfriend 1
+▁centuries 1
+▁committee 1
+▁embassy 1
+▁encounter 1
+▁flavor 1
+▁infrastructure 1
+▁ingredient 1
+▁intelligence 1
+▁maximum 1
+▁pleasure 1
+▁pregnant 1
+▁rainforest 1
+▁taiwan 1
+▁amazon 1
+▁bottle 1
+▁britain 1
+▁croatia 1
+▁amateur 1
+▁pattern 1
+▁smith 1
+▁trouble 1
+▁hispanic 1
+▁turkis 1
+▁injuries 1
+▁rebellion 1
+▁sunday 1
+▁flesh 1
+▁duties 1
+▁chorus 1
+▁mercy 1
+▁craft 1
+▁alliance 1
+▁motto 1
+▁clay 1
+▁reef 1
+▁choice 1
+▁borneo 1
+▁concrete 1
+▁collaborator 1
+▁dramatic 1
+▁instant 1
+▁narrative 1
+▁attractive 1
+▁seville 1
+▁trend 1
+▁safety 1
+▁shock 1
+phone 1
+▁debt 1
+▁simultaneously 1
+▁fuel 1
+▁arrival 1
+▁guitarist 1
+▁touristic 1
+▁motion 1
+▁taxi 1
+▁temperat 1
+▁rico 1
+▁arriv 1
+▁complaint 1
+▁recommended 1
+▁sink 1
+▁korean 1
+▁eva 1
+▁arches 1
+▁escaped 1
+▁gained 1
+▁mora 1
+▁cool 1
+▁dish 1
+▁coll 1
+▁compli 1
+▁core 1
+▁handle 1
+pping 1
+phon 1
+eiro 1
+▁continent 1
+▁sara 1
+▁albert 1
+▁initiat 1
+▁employ 1
+▁abadesa 1
+▁aquatic 1
+▁cervantes 1
+▁chocolate 1
+▁colegio 1
+▁diversity 1
+▁grandmother 1
+▁guillermo 1
+▁nervous 1
+▁opportunity 1
+▁pilgrim 1
+▁sacrifice 1
+▁sanctuary 1
+▁scattered 1
+▁alcohol 1
+▁tixul 1
+▁blonde 1
+▁cordoba 1
+▁cowboy 1
+▁variation 1
+▁target 1
+▁shouted 1
+▁segment 1
+▁belief 1
+▁heritage 1
+▁withdraw 1
+▁cartoon 1
+actic 1
+▁costume 1
+▁gomez 1
+▁false 1
+▁javier 1
+▁opposition 1
+▁invasion 1
+▁decline 1
+▁average 1
+▁tragic 1
+▁vertical 1
+▁fauna 1
+▁instance 1
+▁catalog 1
+▁quantit 1
+▁breast 1
+▁resistant 1
+▁everywhere 1
+▁disney 1
+▁malaga 1
+▁wheat 1
+▁guess 1
+▁dense 1
+▁cannon 1
+▁stock 1
+▁detect 1
+▁rocket 1
+▁mike 1
+▁useful 1
+▁spur 1
+▁campus 1
+▁romance 1
+▁modest 1
+▁stain 1
+▁austrian 1
+▁breeding 1
+▁collected 1
+▁squ 1
+▁presenter 1
+▁sale 1
+▁scr 1
+▁reproduce 1
+awa 1
+▁provid 1
+▁jan 1
+▁supporter 1
+▁collect 1
+▁tube 1
+▁possess 1
+greed 1
+corporate 1
+▁conversation 1
+▁definite 1
+▁faculty 1
+▁imagine 1
+▁knife 1
+▁lópez 1
+▁maneuver 1
+▁orthodox 1
+▁protocol 1
+▁reception 1
+▁respond 1
+▁revenge 1
+▁software 1
+▁supply 1
+▁territorial 1
+▁terror 1
+▁varieties 1
+▁argument 1
+▁hometown 1
+▁rugby 1
+▁capsule 1
+▁courage 1
+▁scandal 1
+▁columbia 1
+▁error 1
+▁fragment 1
+▁retreat 1
+▁saxon 1
+▁symptom 1
+▁videogame 1
+▁wolf 1
+▁wilson 1
+▁protein 1
+▁global 1
+▁relax 1
+▁jaime 1
+▁liquid 1
+▁shift 1
+▁stairs 1
+▁tempt 1
+▁tennis 1
+▁creature 1
+▁homonymous 1
+▁insult 1
+▁infect 1
+▁humor 1
+▁riding 1
+▁smart 1
+▁planned 1
+▁textile 1
+▁improvement 1
+▁stateroom 1
+▁contributions 1
+▁finland 1
+▁fox 1
+▁efficient 1
+▁doctorate 1
+▁manual 1
+▁font 1
+▁grade 1
+▁logo 1
+▁rath 1
+▁exit 1
+▁carmen 1
+▁pack 1
+▁grain 1
+▁restricted 1
+▁locality 1
+▁bible 1
+▁cala 1
+▁worse 1
+▁crosses 1
+▁rate 1
+▁symbolize 1
+▁sao 1
+▁chan 1
+▁tim 1
+▁confirm 1
+▁mineral 1
+▁egg 1
+▁servant 1
+▁assembl 1
+▁rode 1
+▁volcan 1
+▁grab 1
+▁horizon 1
+▁rig 1
+▁recognize 1
+▁bolivar 1
+▁concentration 1
+▁concepcion 1
+▁consecutive 1
+▁consumption 1
+▁depression 1
+▁dispersed 1
+▁forehead 1
+▁goalkeeper 1
+▁grijalba 1
+▁homosexual 1
+▁mammals 1
+▁muslim 1
+▁opponent 1
+▁participant 1
+▁sacristy 1
+▁subspecies 1
+▁swedish 1
+▁universal 1
+▁estrella 1
+▁guilty 1
+▁joseph 1
+▁substance 1
+▁unusual 1
+▁malaysia 1
+▁leather 1
+▁transformed 1
+▁trophy 1
+▁beyond 1
+▁complement 1
+▁remind 1
+▁siege 1
+▁jungle 1
+▁client 1
+▁lagoon 1
+▁friar 1
+▁flores 1
+▁memoria 1
+▁pretend 1
+▁eighteen 1
+▁domain 1
+▁honest 1
+▁proud 1
+▁edward 1
+▁ladies 1
+▁denied 1
+▁policy 1
+▁hiding 1
+▁castillo 1
+▁keyboard 1
+▁swis 1
+▁toxic 1
+▁colony 1
+▁lift 1
+▁balance 1
+▁heavi 1
+▁equation 1
+▁egyptian 1
+▁mild 1
+▁approved 1
+▁highway 1
+▁determine 1
+▁galician 1
+▁exclusively 1
+▁improved 1
+▁dust 1
+▁presid 1
+▁procession 1
+▁paulo 1
+▁mail 1
+▁functional 1
+▁belt 1
+▁joan 1
+▁gru 1
+▁hop 1
+▁introduc 1
+▁struggle 1
+ock 1
+▁america 1
+rino 1
+▁settle 1
+▁improve 1
+▁reduce 1
+▁define 1
+eche 1
+▁hyp 1
+▁tap 1
+▁antarctic 1
+▁contemplat 1
+▁disciple 1
+▁distinctive 1
+▁dynasty 1
+▁earthquake 1
+▁equipment 1
+▁ernesto 1
+▁excited 1
+▁extraordinary 1
+▁fernández 1
+▁instruction 1
+▁intervention 1
+▁linguistic 1
+▁migration 1
+▁ornamental 1
+▁penetrat 1
+▁perspective 1
+▁suspect 1
+▁tanzania 1
+▁terrestrial 1
+▁terrible 1
+▁valladolid 1
+▁delicious 1
+▁disorder 1
+▁evolv 1
+▁tribute 1
+▁whitish 1
+▁configur 1
+▁proportion 1
+▁skull 1
+▁smiling 1
+▁empty 1
+▁nigeria 1
+▁cadiz 1
+▁stuff 1
+▁clothing 1
+▁nueva 1
+▁injury 1
+▁dropped 1
+▁dutch 1
+▁syria 1
+▁goddess 1
+▁pirate 1
+▁plus 1
+▁illness 1
+▁presidency 1
+▁spiritual 1
+▁cinematograph 1
+▁irish 1
+▁faithful 1
+▁annoy 1
+▁package 1
+▁athens 1
+▁forgot 1
+▁delay 1
+▁realistic 1
+▁curious 1
+▁sculptor 1
+▁animation 1
+▁contrast 1
+▁controlled 1
+▁sitting 1
+▁cabinet 1
+▁angle 1
+▁attributed 1
+▁successor 1
+▁chase 1
+▁theories 1
+▁destiny 1
+▁continental 1
+▁hyper 1
+▁rap 1
+▁finishes 1
+tima 1
+▁linear 1
+iza 1
+▁speaker 1
+▁calci 1
+oya 1
+▁refus 1
+▁adopt 1
+▁threaten 1
+▁highlight 1
+▁fix 1
+▁combine 1
+▁exclusive 1
+▁transfer 1
+▁commu 1
+▁punish 1
+▁arrange 1
+▁brigade 1
+▁conclude 1
+▁controversial 1
+▁dictatorship 1
+▁disturb 1
+▁emergency 1
+▁grandparents 1
+▁guarantee 1
+▁hollywood 1
+▁impression 1
+▁industries 1
+▁jurisdiction 1
+▁limestone 1
+▁navarra 1
+▁netherlands 1
+▁primitive 1
+▁purchase 1
+▁relevant 1
+▁slaughter 1
+▁variant 1
+▁virtue 1
+▁yucatan 1
+▁advice 1
+▁crush 1
+▁download 1
+▁easier 1
+▁enemies 1
+▁jordan 1
+▁munich 1
+▁populated 1
+▁prayer 1
+▁stretch 1
+▁bogota 1
+▁muscle 1
+▁perceive 1
+▁regiment 1
+▁romanesque 1
+▁transition 1
+▁earlier 1
+dolfo 1
+▁unstable 1
+▁sixteen 1
+▁renew 1
+▁cereal 1
+▁michel 1
+▁aragon 1
+▁congratulations 1
+▁oriented 1
+▁punk 1
+▁portray 1
+▁dozen 1
+▁phil 1
+▁scholarship 1
+▁comedian 1
+▁copy 1
+▁automatic 1
+▁aware 1
+▁oscar 1
+▁string 1
+▁joyful 1
+▁dynamic 1
+▁mask 1
+▁rapid 1
+▁boom 1
+▁visitors 1
+▁apse 1
+▁editorial 1
+▁castilla 1
+▁toledo 1
+▁revolutionary 1
+▁valid 1
+▁crash 1
+▁mortal 1
+▁horizontal 1
+▁clari 1
+▁collector 1
+▁designated 1
+▁stamp 1
+▁shack 1
+▁dizzy 1
+▁jealousy 1
+▁aunt 1
+▁jean 1
+▁plata 1
+▁mountainous 1
+▁poland 1
+▁marble 1
+▁organic 1
+▁soup 1
+▁boil 1
+▁boxes 1
+▁specialist 1
+quez 1
+▁tender 1
+skirt 1
+▁chess 1
+▁chu 1
+▁passes 1
+▁carv 1
+▁institut 1
+▁cow 1
+▁cerr 1
+▁mala 1
+tally 1
+maker 1
+▁typ 1
+▁complain 1
+▁acquire 1
+what 1
+▁communicat 1
+▁construct 1
+polis 1
+▁accurate 1
+▁evident 1
+▁activist 1
+▁alejandro 1
+▁altitude 1
+▁appreciat 1
+▁cambridge 1
+▁decisive 1
+▁elizabeth 1
+▁equipped 1
+▁expansion 1
+▁fantastic 1
+▁frequency 1
+▁incredibl 1
+▁installation 1
+▁lawsuit 1
+▁madagascar 1
+▁morocco 1
+▁offensive 1
+▁ottoman 1
+▁predecessor 1
+▁procedure 1
+▁shrub 1
+▁sustain 1
+▁terrain 1
+▁transmission 1
+▁coconut 1
+▁identify 1
+▁masonry 1
+▁sevilla 1
+▁dragon 1
+▁throat 1
+▁arabic 1
+▁compilation 1
+▁eternal 1
+▁modification 1
+▁helmet 1
+▁method 1
+▁violin 1
+▁analyze 1
+▁spark 1
+▁dirty 1
+▁miami 1
+▁viceroy 1
+▁discipline 1
+▁absence 1
+▁dancing 1
+▁mystery 1
+▁fried 1
+▁squad 1
+▁rename 1
+▁hired 1
+▁ensure 1
+▁gentleman 1
+▁rigid 1
+▁tongue 1
+▁carbon 1
+▁resume 1
+▁worship 1
+aceous 1
+▁idol 1
+▁integrated 1
+▁conservation 1
+▁tiger 1
+▁closing 1
+▁fiber 1
+▁islamic 1
+▁income 1
+▁succession 1
+ridge 1
+▁passage 1
+▁coloniz 1
+▁greet 1
+▁poem 1
+▁lemon 1
+▁virginia 1
+▁personality 1
+▁bitter 1
+▁armenian 1
+▁massive 1
+▁norman 1
+▁impos 1
+▁parent 1
+▁basilica 1
+itz 1
+▁phone 1
+mund 1
+▁wooden 1
+▁raw 1
+negro 1
+▁toast 1
+▁boot 1
+▁saga 1
+▁armor 1
+▁serv 1
+▁slip 1
+▁drove 1
+▁fog 1
+▁von 1
+▁joke 1
+▁edit 1
+vista 1
+▁advis 1
+▁threat 1
+▁educa 1
+▁altarpiece 1
+▁archdiocese 1
+▁budget 1
+▁chemistry 1
+▁confusing 1
+▁defensive 1
+▁emotional 1
+▁episcopa 1
+▁equator 1
+▁ethiopia 1
+▁expensive 1
+▁hernandez 1
+▁investment 1
+▁lorenzo 1
+▁michigan 1
+▁paralympic 1
+▁plumage 1
+▁repertoire 1
+▁structural 1
+▁swallow 1
+▁derived 1
+▁fashion 1
+▁florence 1
+▁friday 1
+▁minimal 1
+▁peasant 1
+▁qualify 1
+▁surgery 1
+▁pizza 1
+▁scream 1
+▁ciudad 1
+▁devastat 1
+▁nobility 1
+▁copper 1
+▁scotland 1
+▁dominant 1
+▁modified 1
+▁andrés 1
+▁monitor 1
+▁envy 1
+▁pastoral 1
+▁frog 1
+▁alonso 1
+▁outdoor 1
+▁burgos 1
+▁employee 1
+▁responsibility 1
+▁oval 1
+▁particle 1
+▁lucas 1
+▁consol 1
+▁infantry 1
+▁brotherhood 1
+▁lightning 1
+▁youtube 1
+▁export 1
+▁inspect 1
+atherine 1
+▁elementary 1
+▁adja 1
+▁burial 1
+▁factories 1
+▁loving 1
+▁roger 1
+▁pearl 1
+▁lucia 1
+▁jenn 1
+▁antique 1
+▁johnson 1
+▁plent 1
+▁basal 1
+▁duty 1
+▁leonor 1
+▁react 1
+▁wonder 1
+▁minute 1
+▁raul 1
+▁specie 1
+rbon 1
+▁vigo 1
+▁rail 1
+▁sap 1
+▁luna 1
+▁gau 1
+▁bench 1
+▁susan 1
+▁chris 1
+▁helen 1
+▁poli 1
+▁select 1
+▁associate 1
+▁specialize 1
+▁democrat 1
+script 1
+▁scott 1
+▁affirm 1
+▁airplane 1
+▁aluminum 1
+▁appointment 1
+▁archbishop 1
+▁artillery 1
+▁astronomer 1
+▁blossom 1
+▁breakfast 1
+▁caribbean 1
+▁coincide 1
+▁disappoint 1
+▁distracted 1
+▁draft 1
+▁díaz 1
+▁excavat 1
+▁expelled 1
+▁explode 1
+▁exterior 1
+▁guerrero 1
+▁hungaria 1
+▁ignacio 1
+▁illuminate 1
+▁inspiration 1
+▁jesuit 1
+▁kitchen 1
+▁legislature 1
+▁manufacturing 1
+▁mathematics 1
+▁negotiation 1
+▁pontifical 1
+▁prestigio 1
+▁simpson 1
+▁supplie 1
+▁surrender 1
+▁transparent 1
+▁hungary 1
+▁vessel 1
+▁edible 1
+▁excuse 1
+▁senior 1
+▁tissue 1
+▁caracas 1
+▁jersey 1
+▁conquest 1
+▁interaction 1
+▁courtyard 1
+▁regulate 1
+▁tokyo 1
+▁sauce 1
+▁cardinal 1
+▁stuck 1
+▁cauca 1
+▁aztec 1
+▁martyr 1
+▁secure 1
+▁corpse 1
+▁robust 1
+▁rector 1
+▁rational 1
+▁sociolog 1
+▁ugly 1
+▁nerve 1
+▁ascend 1
+▁warrior 1
+▁nightmare 1
+▁crest 1
+▁assets 1
+▁virus 1
+▁glad 1
+▁juice 1
+▁items 1
+▁extinction 1
+▁ballad 1
+▁venue 1
+▁feminist 1
+point 1
+▁turb 1
+▁subway 1
+▁brush 1
+▁runner 1
+▁creator 1
+▁bold 1
+▁harsh 1
+▁lung 1
+▁authorized 1
+▁guardian 1
+▁fake 1
+▁tiny 1
+▁kong 1
+▁walter 1
+▁fame 1
+▁nick 1
+▁islam 1
+▁cub 1
+▁nazi 1
+▁castro 1
+▁vegetable 1
+▁lyric 1
+▁insect 1
+▁mental 1
+▁update 1
+▁server 1
+shire 1
+qualified 1
+▁electro 1
+▁supervis 1
+gging 1
+▁recommend 1
+▁actual 1
+▁gentle 1
+▁femini 1
+río 1
+ó 1
+▁algorithm 1
+▁ancestry 1
+▁bermudez 1
+▁coalition 1
+▁confederation 1
+▁creole 1
+▁crustaceans 1
+▁democracy 1
+▁discharge 1
+▁esteban 1
+▁exciting 1
+▁guadalupe 1
+▁harvard 1
+▁hawaii 1
+▁impressive 1
+▁inferior 1
+▁magnificent 1
+▁massachusetts 1
+▁nocturnal 1
+▁occupation 1
+▁opportunities 1
+▁pakistan 1
+▁participating 1
+▁protagonist 1
+▁pseudonym 1
+▁railroad 1
+▁taylor 1
+▁theoretical 1
+▁tunnel 1
+▁ukraine 1
+▁ángel 1
+▁cancelled 1
+▁cuisine 1
+▁delegate 1
+▁impressed 1
+▁manifest 1
+▁resembl 1
+▁chicken 1
+▁complicated 1
+▁glorious 1
+▁migrate 1
+▁kidnapping 1
+▁remote 1
+▁orchard 1
+▁restoration 1
+▁scheme 1
+▁amazing 1
+▁mutual 1
+▁kenya 1
+▁extent 1
+▁optim 1
+metry 1
+▁angola 1
+▁annex 1
+▁toronto 1
+▁cannibals 1
+▁jews 1
+▁summit 1
+▁mouse 1
+▁metallic 1
+▁exposition 1
+▁flame 1
+▁admitted 1
+▁planning 1
+▁steal 1
+▁inheritance 1
+▁reproduction 1
+▁citizenship 1
+▁rioja 1
+▁peaceful 1
+▁punishment 1
+▁stability 1
+▁correspondent 1
+▁pagan 1
+▁shake 1
+▁countryside 1
+stratigraph 1
+▁ballet 1
+▁ashes 1
+▁difficulties 1
+▁christianity 1
+▁rush 1
+▁businessman 1
+▁protector 1
+▁pond 1
+▁differentiate 1
+▁legion 1
+▁herald 1
+▁jaw 1
+▁kei 1
+▁wipe 1
+▁chronicle 1
+▁creat 1
+▁qualifie 1
+▁ripe 1
+▁marquis 1
+▁immediate 1
+ador 1
+▁accusation 1
+▁activate 1
+▁afghanistan 1
+▁aforementioned 1
+▁alvarez 1
+▁apprentice 1
+▁autumn 1
+▁bilbao 1
+▁cavalry 1
+▁commitment 1
+▁consequently 1
+▁corporation 1
+▁cylindrical 1
+▁definitive 1
+▁delegation 1
+▁distinction 1
+▁ecosystem 1
+▁elaborate 1
+▁grammy 1
+▁guadalajara 1
+▁hermitage 1
+▁invertebrate 1
+▁joaquín 1
+▁meadows 1
+▁mongolia 1
+▁mozambique 1
+▁offspring 1
+▁outbreak 1
+▁peculiar 1
+▁privilege 1
+▁puebla 1
+▁reconstruct 1
+▁rodrigo 1
+▁steep 1
+▁suspended 1
+▁ultimate 1
+▁absorb 1
+▁dwarf 1
+▁eyelids 1
+▁fracture 1
+▁sketch 1
+▁cuadra 1
+▁fabric 1
+▁ignore 1
+▁abbess 1
+▁czech 1
+▁random 1
+▁catalina 1
+▁chrono 1
+▁remix 1
+▁reunite 1
+▁suburb 1
+▁colonies 1
+▁blanco 1
+▁laguna 1
+▁oriental 1
+▁companion 1
+▁knock 1
+▁regain 1
+▁gustav 1
+▁sensor 1
+▁pardon 1
+▁innocent 1
+itarian 1
+▁ortiz 1
+▁frost 1
+▁cotton 1
+bular 1
+▁drain 1
+▁courtesy 1
+▁prosper 1
+▁davis 1
+▁mosque 1
+▁solar 1
+▁savannah 1
+▁scottish 1
+▁turbine 1
+▁persian 1
+▁literal 1
+▁gerard 1
+▁grayish 1
+▁scope 1
+▁interfer 1
+▁patio 1
+▁norway 1
+▁texture 1
+▁alien 1
+▁riders 1
+▁lighthouse 1
+▁truck 1
+▁tight 1
+▁barrio 1
+▁harris 1
+▁colorado 1
+▁fold 1
+▁jury 1
+▁paraguayan 1
+▁dunes 1
+▁payment 1
+▁mold 1
+▁possession 1
+▁panel 1
+clave 1
+▁reflection 1
+▁stew 1
+▁finger 1
+intestin 1
+▁robot 1
+▁sigh 1
+estio 1
+▁diploma 1
+▁contra 1
+▁execut 1
+▁adapt 1
+▁arrest 1
+▁bibl 1
+▁teenage 1
+▁theatr 1
+▁animat 1
+▁desperate 1
+▁anarchist 1
+▁asteroid 1
+▁atmosphere 1
+▁automobile 1
+▁breeze 1
+▁brilliant 1
+▁contempt 1
+▁controversy 1
+▁curiosity 1
+▁expired 1
+▁festivities 1
+▁frustrat 1
+▁granddaughter 1
+▁hydra 1
+▁imagination 1
+▁mallorca 1
+▁muñoz 1
+▁napoleon 1
+▁neutral 1
+▁overwhelmed 1
+▁profound 1
+▁recruit 1
+▁renowned 1
+▁trujillo 1
+▁unanimous 1
+▁valparaiso 1
+▁voluptuous 1
+▁bacteria 1
+▁discussed 1
+▁downtown 1
+▁improvis 1
+▁jarocho 1
+▁maintenance 1
+▁patience 1
+▁policies 1
+▁asturian 1
+▁bulbs 1
+▁neoclassic 1
+▁philip 1
+▁pillars 1
+▁poverty 1
+▁sanchez 1
+▁accessible 1
+▁evangeli 1
+▁penalty 1
+▁sahara 1
+▁alleged 1
+▁surreal 1
+▁dolores 1
+▁archaeolog 1
+▁speculat 1
+▁changing 1
+▁postpone 1
+▁clip 1
+▁ceiling 1
+▁implied 1
+▁implies 1
+▁josé 1
+▁raising 1
+▁ideology 1
+▁lonely 1
+▁depict 1
+▁weekend 1
+▁fortune 1
+▁behave 1
+▁exceed 1
+▁hooked 1
+▁overcome 1
+▁inscription 1
+▁emmy 1
+▁cruci 1
+▁proceed 1
+▁horribl 1
+▁infant 1
+▁charlie 1
+ñez 1
+▁axis 1
+▁herrera 1
+▁slim 1
+▁introduction 1
+▁ruiz 1
+▁plateau 1
+▁foster 1
+atomic 1
+▁garcia 1
+▁odd 1
+▁innovation 1
+▁arose 1
+▁biology 1
+▁wales 1
+▁hong 1
+▁fence 1
+▁gang 1
+therapy 1
+physi 1
+▁sophi 1
+▁passenger 1
+racts 1
+▁produc 1
+andr 1
+▁pep 1
+▁geographic 1
+▁achiev 1
+▁exhibit 1
+▁preserve 1
+▁fernand 1
+loved 1
+▁sculpt 1
+▁innovati 1
+▁conserv 1
+▁consult 1
+▁alicante 1
+▁andalucia 1
+▁anthropolog 1
+▁asturias 1
+▁audiovisual 1
+▁autonomy 1
+▁baptist 1
+▁bonfire 1
+▁cantabria 1
+▁centaur 1
+▁compatible 1
+▁condemn 1
+▁cyclist 1
+▁denomination 1
+▁discussion 1
+▁explosion 1
+▁fascinat 1
+▁franchise 1
+▁hammock 1
+▁helicopter 1
+▁hypothesis 1
+▁implementation 1
+▁intelligent 1
+▁jesús 1
+▁juvenile 1
+▁larvae 1
+▁membrane 1
+▁mexico 1
+▁migratory 1
+▁mosquito 1
+▁norwegian 1
+▁pregnancy 1
+▁pursued 1
+▁sacrament 1
+▁sebastián 1
+▁solitary 1
+▁sovereign 1
+▁switch 1
+▁tuesday 1
+▁collapse 1
+▁disguise 1
+▁distress 1
+▁intermedia 1
+▁manchester 1
+▁melody 1
+▁miserable 1
+▁paternal 1
+▁prussia 1
+▁retiring 1
+▁sequence 1
+▁suarez 1
+▁comparison 1
+▁emilio 1
+▁wandering 1
+▁ashlar 1
+▁idiot 1
+▁mercenar 1
+▁asshole 1
+▁gospel 1
+▁mandatory 1
+▁essence 1
+▁patricia 1
+▁podium 1
+▁arthur 1
+▁snack 1
+▁claw 1
+▁stroke 1
+▁fault 1
+▁fiesta 1
+▁tremor 1
+▁escort 1
+▁centro 1
+▁competing 1
+▁nepal 1
+▁orchid 1
+▁multipli 1
+▁laboratory 1
+▁observations 1
+▁cooperative 1
+▁josef 1
+▁circulation 1
+▁advertisement 1
+▁kidney 1
+▁saddle 1
+▁diagnosis 1
+▁groom 1
+▁tourney 1
+▁clarity 1
+▁espino 1
+▁geography 1
+▁protestant 1
+▁bored 1
+▁topography 1
+▁johan 1
+▁examine 1
+▁drought 1
+▁entities 1
+▁tension 1
+▁anderson 1
+▁mural 1
+finity 1
+▁serg 1
+▁optic 1
+omy 1
+nibal 1
+▁illustr 1
+acea 1
+▁lamb 1
+▁register 1
+▁temper 1
+▁galicia 1
+▁announce 1
+worth 1
+bank 1
+direct 1
+omb 1
+▁differ 1
+▁scholar 1
+▁abstract 1
+▁alfonso 1
+▁anthologies 1
+▁calculation 1
+▁cartagena 1
+▁christopher 1
+▁concentrate 1
+▁confusion 1
+▁consecrat 1
+▁cricket 1
+▁cristobal 1
+▁dinosaur 1
+▁eagle 1
+▁elegant 1
+▁endangered 1
+▁entrepreneur 1
+▁escuela 1
+▁español 1
+▁estimated 1
+▁explanation 1
+▁explosive 1
+▁forbidden 1
+▁furniture 1
+▁glacier 1
+▁illegal 1
+▁kiwi 1
+▁marqués 1
+▁militant 1
+▁occident 1
+▁ourselves 1
+▁rebuilt 1
+▁recreation 1
+▁seaweed 1
+▁societies 1
+▁suspicion 1
+▁tenerife 1
+▁vineyard 1
+▁wool 1
+▁departure 1
+▁fantasy 1
+▁interface 1
+▁mozart 1
+▁obvious 1
+▁proximit 1
+▁terrace 1
+▁villaverde 1
+▁castilian 1
+▁mendoza 1
+▁navarre 1
+▁translator 1
+▁zamora 1
+▁nevada 1
+▁havana 1
+▁intensity 1
+▁kevin 1
+▁pillow 1
+▁conifer 1
+▁enhance 1
+▁radius 1
+▁utah 1
+▁nineteen 1
+▁widow 1
+▁disaster 1
+▁lawrence 1
+▁thirst 1
+▁seventeen 1
+▁corolla 1
+▁storage 1
+▁trunk 1
+▁eager 1
+▁dulce 1
+▁choreograph 1
+▁yelled 1
+▁resort 1
+▁duration 1
+▁replied 1
+▁intervene 1
+▁gesture 1
+▁await 1
+▁tierra 1
+▁mocking 1
+▁clara 1
+▁entry 1
+▁dragg 1
+▁carpet 1
+▁faint 1
+▁requirements 1
+▁pilar 1
+▁whistle 1
+▁vain 1
+▁fortified 1
+▁sequel 1
+▁pepper 1
+▁stephen 1
+▁gentlemen 1
+▁dorsal 1
+▁consultant 1
+▁basilio 1
+▁rapper 1
+▁philology 1
+▁furious 1
+▁golf 1
+▁roast 1
+▁venice 1
+▁harbor 1
+▁settlers 1
+▁considerably 1
+▁preference 1
+▁shiver 1
+▁convict 1
+▁boost 1
+▁distort 1
+▁kitt 1
+tropic 1
+▁abuse 1
+▁cleve 1
+▁equ 1
+rank 1
+▁analog 1
+mmit 1
+▁receiv 1
+uart 1
+▁chronic 1
+oxide 1
+▁surround 1
+▁flee 1
+▁basil 1
+▁retire 1
+▁destin 1
+▁occupie 1
+▁propose 1
+▁colonia 1
+▁franc 1
+scribed 1
+▁kidnap 1
+▁diamond 1
+▁mathematical 1
+▁advocat 1
+▁affiliate 1
+▁alexandria 1
+▁arizona 1
+▁assault 1
+▁birmingham 1
+▁carnival 1
+▁competitive 1
+▁complutense 1
+▁conspiracy 1
+▁cultivation 1
+▁currency 1
+▁deciduous 1
+▁demolished 1
+▁doctrine 1
+▁elephant 1
+▁encyclopedia 1
+▁exploitation 1
+▁illinois 1
+▁istanbul 1
+▁legislation 1
+▁longitudinal 1
+▁magdalena 1
+▁margarita 1
+▁mausoleum 1
+▁pennsylvania 1
+▁prosecutor 1
+▁pyramid 1
+▁registration 1
+▁relegated 1
+▁resolve 1
+▁sicily 1
+▁skeleton 1
+▁suffrag 1
+▁thorough 1
+▁unbearable 1
+▁zambia 1
+▁zimbabwe 1
+▁acceler 1
+▁acoustic 1
+▁betrayal 1
+▁obeyed 1
+▁oviedo 1
+▁quantum 1
+▁sumatra 1
+▁surgeon 1
+▁tendency 1
+▁trigger 1
+▁surname 1
+▁attorney 1
+▁bavaria 1
+▁lizard 1
+▁thief 1
+▁conceiv 1
+▁granite 1
+▁summarize 1
+▁murcia 1
+▁waterfall 1
+▁ghana 1
+▁flamenco 1
+▁flank 1
+▁naples 1
+▁vocation 1
+▁evaluat 1
+▁grief 1
+▁serrat 1
+▁dictat 1
+▁cloister 1
+▁atlanta 1
+▁width 1
+▁pianist 1
+▁pupil 1
+▁yelling 1
+▁stirr 1
+▁import 1
+▁devout 1
+▁ceased 1
+▁iraq 1
+▁arrangements 1
+▁conservatory 1
+▁serena 1
+▁worried 1
+▁imitate 1
+▁tigre 1
+▁gastro 1
+▁portico 1
+▁barrier 1
+▁donkey 1
+ogram 1
+▁hebrew 1
+neath 1
+▁jacob 1
+▁balear 1
+▁disdainful 1
+▁fetus 1
+▁gently 1
+▁harass 1
+▁cheap 1
+▁mortar 1
+▁fusion 1
+▁entre 1
+▁bosnia 1
+▁cried 1
+▁mixture 1
+cession 1
+▁spill 1
+▁vital 1
+▁fuente 1
+▁index 1
+▁molecul 1
+▁separati 1
+ldom 1
+▁francesc 1
+crete 1
+▁whip 1
+▁meteorologic 1
+▁dye 1
+mount 1
+▁succeed 1
+▁float 1
+▁jealous 1
+▁dizz 1
+▁sudden 1
+graduate 1
+▁confront 1
+▁walt 1
+▁abundance 1
+▁accomplish 1
+▁amsterdam 1
+▁anxiety 1
+▁approval 1
+▁arrogant 1
+▁axillary 1
+▁baptized 1
+▁bautista 1
+▁ceremonies 1
+▁circus 1
+▁colleague 1
+▁consolidated 1
+▁continuation 1
+▁creek 1
+▁cretaceous 1
+▁dialogue 1
+▁douglas 1
+▁ecclesiastic 1
+▁elliptic 1
+▁elongated 1
+▁embark 1
+▁engraving 1
+▁enrolled 1
+▁evacuate 1
+▁frozen 1
+▁gregorio 1
+▁guayaquil 1
+▁guerrilla 1
+▁hidalgo 1
+▁hybrid 1
+▁indefinite 1
+▁lithuania 1
+▁majesty 1
+▁manipulat 1
+▁microsoft 1
+▁namibia 1
+▁perception 1
+▁pleasant 1
+▁possibilities 1
+▁pronounc 1
+▁rehearsal 1
+▁satisfy 1
+▁sebastian 1
+▁slept 1
+▁slovenia 1
+▁sydney 1
+▁tattoo 1
+▁technologies 1
+▁tributary 1
+▁unconscious 1
+▁uncovered 1
+▁vargas 1
+▁aquariu 1
+▁batteries 1
+▁contamina 1
+▁counsel 1
+▁credit 1
+▁macedonia 1
+▁patrick 1
+▁posterior 1
+▁supplement 1
+▁uganda 1
+onymous 1
+▁anatomy 1
+▁arsenal 1
+▁coruña 1
+▁garrison 1
+▁qualities 1
+▁signature 1
+▁surpass 1
+▁clums 1
+▁monterrey 1
+▁slogan 1
+▁statute 1
+gnition 1
+▁radiation 1
+▁reduction 1
+▁tobacco 1
+▁albania 1
+▁dignity 1
+▁erosion 1
+▁futsal 1
+▁invention 1
+▁ancestor 1
+▁fluid 1
+▁tomás 1
+▁abbey 1
+▁curtain 1
+▁legacy 1
+▁homage 1
+▁screw 1
+gonzalo 1
+▁arequipa 1
+▁nelson 1
+▁snail 1
+▁niece 1
+▁dairy 1
+▁vergara 1
+▁uprising 1
+▁gabon 1
+▁reunion 1
+▁duchy 1
+▁decease 1
+▁almond 1
+▁contestants 1
+▁retain 1
+▁lodge 1
+▁potosi 1
+▁mandate 1
+▁blank 1
+▁depilat 1
+▁steam 1
+▁spinning 1
+▁valentin 1
+▁housing 1
+▁absent 1
+▁decree 1
+▁adobe 1
+▁algeria 1
+▁auction 1
+▁vein 1
+▁anglo 1
+▁audio 1
+▁adrian 1
+▁loaf 1
+ockey 1
+▁quadr 1
+▁knee 1
+▁jewel 1
+vinci 1
+alawi 1
+▁telephon 1
+▁corona 1
+▁biologic 1
+▁practic 1
+position 1
+apua 1
+▁meteor 1
+▁georg 1
+positor 1
+▁celebrat 1
+▁recogniz 1
+doubted 1
+▁distinguish 1
+▁restrict 1
+egui 1
+▁distribute 1
+woman 1
+▁alfred 1
+culture 1
+▁dispos 1
+▁diagnos 1
+tinent 1
+▁embroider 1
+▁ultra 1
+▁absurd 1
+▁agencies 1
+▁albacete 1
+▁belgrano 1
+▁berkeley 1
+▁bicycle 1
+▁byzantine 1
+▁cambodia 1
+▁chaplain 1
+▁cluster 1
+▁commemorate 1
+▁controversies 1
+▁creep 1
+▁crocodile 1
+▁crusade 1
+▁cylinder 1
+▁cyprus 1
+▁disembark 1
+▁emphasize 1
+▁enormous 1
+▁evergreen 1
+▁florenci 1
+▁frederick 1
+▁galaxy 1
+▁gratitude 1
+▁happily 1
+▁hierarchy 1
+▁indifferent 1
+▁influential 1
+▁interrogated 1
+▁jimmy 1
+▁libraries 1
+▁luxurious 1
+▁magistrate 1
+▁minnesota 1
+▁motorcycle 1
+▁multiplayer 1
+▁navarro 1
+▁pamplona 1
+▁pharmac 1
+▁platinum 1
+▁quartet 1
+▁ramírez 1
+▁reappear 1
+▁simón 1
+▁soluble 1
+▁spectrum 1
+▁spontaneous 1
+▁stablish 1
+▁subordinate 1
+▁suppress 1
+▁suspension 1
+▁sweetheart 1
+▁syndrome 1
+▁temporal 1
+▁twentie 1
+▁ulysses 1
+▁unexpected 1
+▁vancouver 1
+▁vatican 1
+▁wrestling 1
+▁alphabet 1
+▁corridor 1
+▁embodie 1
+▁foliage 1
+▁intersection 1
+▁jamaica 1
+▁midnight 1
+▁module 1
+▁nutritio 1
+▁quintet 1
+▁shepherd 1
+▁thread 1
+▁transformation 1
+▁calderon 1
+▁expense 1
+▁facebook 1
+▁lutheran 1
+▁singular 1
+▁impulse 1
+▁lisbon 1
+▁pocket 1
+▁cheerful 1
+▁debris 1
+▁domestic 1
+▁massacre 1
+▁rabbi 1
+▁sediment 1
+▁teatro 1
+▁buzz 1
+▁artisan 1
+▁leonardo 1
+▁soledad 1
+▁weakness 1
+▁adhere 1
+▁balkan 1
+▁galleries 1
+▁heaven 1
+▁austin 1
+▁kirk 1
+▁eurasia 1
+▁traitor 1
+▁hunger 1
+▁petersburg 1
+▁montreal 1
+▁somehow 1
+▁antenna 1
+▁atalaya 1
+▁fraud 1
+▁interval 1
+▁rifle 1
+▁tibet 1
+▁appeal 1
+▁seventies 1
+▁sofia 1
+▁mundo 1
+▁output 1
+▁sharing 1
+▁oath 1
+▁berries 1
+▁elevate 1
+▁carlist 1
+▁fresco 1
+immun 1
+cellular 1
+▁libertad 1
+▁anime 1
+▁señor 1
+▁yama 1
+▁addict 1
+▁titan 1
+▁vaca 1
+▁alarm 1
+▁restore 1
+▁alpin 1
+▁antoni 1
+angular 1
+▁philippine 1
+▁recall 1
+aval 1
+precedent 1
+▁remov 1
+▁armenia 1
+▁necessar 1
+stream 1
+american 1
+▁austral 1
+▁admit 1
+▁acknowledg 1
+▁exponent 1
+▁percussion 1
+▁psychiatr 1
+▁visigoth 1
+▁yugoslav 1
+nunciation 1
+▁agustín 1
+▁argentino 1
+▁auditorium 1
+▁azerbaijan 1
+▁benedict 1
+▁chromosom 1
+▁coherent 1
+▁compensation 1
+▁congregation 1
+▁constellation 1
+▁criteria 1
+▁croquettes 1
+▁detroit 1
+▁dioecious 1
+▁extremadura 1
+▁faucet 1
+▁gibraltar 1
+▁graduation 1
+▁hemisphere 1
+▁iglesia 1
+▁irrigation 1
+▁liquor 1
+▁lombard 1
+▁maldonado 1
+▁medellin 1
+▁militia 1
+▁minimum 1
+▁missouri 1
+▁mourning 1
+▁olimpia 1
+▁patagonia 1
+▁patriarch 1
+▁patricio 1
+▁permission 1
+▁preservation 1
+▁prestige 1
+▁prolific 1
+▁prosecuted 1
+▁prototype 1
+▁puzzle 1
+▁reggae 1
+▁reservoir 1
+▁resolution 1
+▁scandinavia 1
+▁serrano 1
+▁shipwreck 1
+▁sidewalk 1
+▁staircase 1
+▁submarine 1
+▁subsidiar 1
+▁surprising 1
+▁tarragona 1
+▁technological 1
+▁tragedies 1
+▁tuxtlan 1
+▁twitter 1
+▁ukrainian 1
+▁vicinity 1
+▁vulnerable 1
+▁widespread 1
+▁abdomen 1
+▁abilities 1
+▁behalf 1
+▁belgian 1
+▁coexist 1
+▁glued 1
+▁jupiter 1
+▁memoirs 1
+▁phillip 1
+▁specified 1
+▁tragedy 1
+▁glabro 1
+▁gordon 1
+▁grammar 1
+▁lifestyle 1
+▁oxygen 1
+▁simplicity 1
+▁betrayed 1
+▁ferran 1
+▁nomination 1
+▁perfume 1
+▁sausage 1
+▁taxonom 1
+▁uneven 1
+▁humbert 1
+▁scanner 1
+▁botany 1
+▁inventor 1
+▁russell 1
+▁houston 1
+▁lewis 1
+▁bucket 1
+▁orientation 1
+▁palermo 1
+▁sensation 1
+▁insurance 1
+▁virgen 1
+▁depriv 1
+▁mosaic 1
+▁sultan 1
+▁abraham 1
+▁dylan 1
+▁invented 1
+▁ritual 1
+▁kidnapped 1
+▁anguish 1
+▁blanca 1
+▁vincent 1
+▁density 1
+▁horseshoe 1
+▁santander 1
+▁abortion 1
+▁arbore 1
+▁trinity 1
+▁patrol 1
+▁pigeon 1
+▁clue 1
+▁binary 1
+▁sparta 1
+▁erotic 1
+▁senate 1
+▁alpha 1
+▁impatien 1
+▁hollow 1
+▁carrera 1
+▁polar 1
+▁deficien 1
+▁pulp 1
+▁drill 1
+▁libya 1
+ghett 1
+▁predict 1
+▁davie 1
+dmother 1
+▁leopol 1
+ventura 1
+▁flute 1
+wald 1
+▁niche 1
+arochi 1
+▁sensib 1
+agram 1
+▁adventur 1
+finals 1
+▁approve 1
+▁eliminate 1
+▁praise 1
+▁decorati 1
+ffluent 1
+▁absolve 1
+▁accuracy 1
+▁ambitious 1
+▁anniversary 1
+▁antioquia 1
+▁appropriate 1
+▁aragón 1
+▁aristocrat 1
+▁astronomy 1
+▁baptism 1
+▁benjamin 1
+▁bluish 1
+▁bordeaux 1
+▁boulevard 1
+▁brooklyn 1
+▁buffalo 1
+▁bundesliga 1
+▁casualties 1
+▁catolica 1
+▁censorship 1
+▁centimeters 1
+▁certified 1
+▁competitor 1
+▁compostela 1
+▁continuity 1
+▁conversion 1
+▁corrientes 1
+▁czechoslovakia 1
+▁declaration 1
+▁depressed 1
+▁derivative 1
+▁disadvantage 1
+▁eclipse 1
+▁edinburgh 1
+▁enthusiast 1
+▁epidemic 1
+▁ethnic 1
+▁excellenc 1
+▁excursion 1
+▁exploited 1
+▁eyebrow 1
+▁frighten 1
+▁grandchildren 1
+▁grotesque 1
+▁hurricane 1
+▁independiente 1
+▁indicator 1
+▁intermittent 1
+▁internacional 1
+▁lincoln 1
+▁liverpool 1
+▁metaboli 1
+▁modality 1
+▁nashville 1
+▁nicholas 1
+▁nutrients 1
+▁obstacle 1
+▁ordinary 1
+▁philadelphia 1
+▁phosphate 1
+▁precarious 1
+▁precipitation 1
+▁predominate 1
+▁promenade 1
+▁psychologist 1
+▁pubescent 1
+▁registry 1
+▁relentless 1
+▁repercussion 1
+▁resurrect 1
+▁segovia 1
+▁significance 1
+▁simulation 1
+▁slavic 1
+▁slovakia 1
+▁sociedad 1
+▁soprano 1
+▁spectacular 1
+▁spokesperson 1
+▁succulent 1
+▁telescope 1
+▁testimony 1
+▁toponym 1
+▁trajectory 1
+▁trinidad 1
+▁videoclip 1
+▁vitamin 1
+▁antofagasta 1
+▁colón 1
+▁despair 1
+▁guarani 1
+▁illusion 1
+▁lleida 1
+▁margaret 1
+▁monotonous 1
+▁premises 1
+▁regret 1
+▁solemn 1
+▁sprouts 1
+▁subdued 1
+▁wizard 1
+higgins 1
+▁acknowledgement 1
+▁assassin 1
+▁gravity 1
+▁hermaphrodite 1
+▁puddle 1
+▁respirat 1
+▁robinson 1
+▁spacious 1
+▁submitt 1
+▁vegetarian 1
+▁cruise 1
+▁kansas 1
+▁oblong 1
+▁rainbow 1
+▁raymond 1
+▁vowel 1
+▁census 1
+▁madonna 1
+▁thorns 1
+▁commentator 1
+▁powder 1
+▁recreate 1
+▁savage 1
+pelagic 1
+▁dessert 1
+▁pereira 1
+▁spike 1
+▁tunisia 1
+▁cartridge 1
+▁himalaya 1
+▁sphere 1
+▁urgent 1
+▁ethno 1
+▁upright 1
+▁smash 1
+▁survey 1
+▁propell 1
+▁franz 1
+▁ideological 1
+▁chimeric 1
+▁ebro 1
+▁overseas 1
+▁federic 1
+▁cairo 1
+▁centen 1
+▁bridle 1
+▁hawk 1
+▁fermi 1
+▁stubb 1
+▁civic 1
+discreet 1
+▁iconograph 1
+▁tacit 1
+athlon 1
+▁physiolog 1
+▁pulsati 1
+▁recurr 1
+existent 1
+▁significati 1
+▁resource 1
+▁determin 1
+priesthood 1
+▁identifi 1
+▁olympi 1
+▁extinct 1
+▁homonym 1
+▁disdain 1
+▁basket 1
+▁ambiguous 1
+▁convenient 1
+▁diagonal 1
+▁excessive 1
+ceptive 1
+elecommunication 1
+legitimate 1
+▁acclaim 1
+▁activism 1
+▁admission 1
+▁allegory 1
+▁anglican 1
+▁antagonist 1
+▁apostol 1
+▁aristotle 1
+▁armstrong 1
+▁arrogance 1
+▁articulated 1
+▁asunción 1
+â 1
+. 1
+ʻ 1
+ð 1
+ú 1
+ʿ 1
+! 1
+þ 1
+ā 1
+̇ 1
+í 1
+ə 1
diff --git a/configs/es-en/tgt_unigram6000/spm_unigram_es.vocab b/configs/es-en/tgt_unigram6000/spm_unigram_es.vocab
new file mode 100644
index 0000000000000000000000000000000000000000..5f22db03e2334e8151008a1c83aba07146742c70
--- /dev/null
+++ b/configs/es-en/tgt_unigram6000/spm_unigram_es.vocab
@@ -0,0 +1,6000 @@
+ 0
+ 0
+ 0
+ 0
+▁the -2.77341
+s -3.09541
+▁of -3.66098
+▁a -3.88633
+▁in -3.89982
+▁and -3.98116
+▁to -4.083
+▁is -4.15941
+▁it -4.27761
+▁he -4.50019
+▁was -4.50511
+' -4.7014
+ed -4.76793
+▁with -4.98895
+t -5.0103
+▁i -5.05722
+▁are -5.06178
+▁that -5.10158
+ing -5.14342
+▁they -5.22301
+▁this -5.23746
+▁for -5.23929
+▁his -5.26238
+n -5.34829
+▁you -5.35449
+y -5.42892
+▁be -5.44453
+d -5.44963
+▁by -5.45419
+▁on -5.48423
+▁as -5.54122
+l -5.54785
+▁has -5.59156
+e -5.61177
+▁from -5.65668
+ly -5.68809
+▁have -5.73187
+▁not -5.73944
+▁were -5.81289
+▁we -5.83065
+▁at -5.83973
+o -5.84099
+▁there -5.87693
+▁ -5.87708
+▁an -5.88351
+▁one -5.9345
+a -6.02718
+er -6.03194
+▁can -6.03517
+▁their -6.05261
+▁she -6.08362
+m -6.13491
+al -6.1421
+▁its -6.15104
+r -6.20872
+▁de -6.2346
+c -6.23779
+▁me -6.24144
+es -6.26791
+▁two -6.31222
+▁first -6.31279
+▁her -6.33811
+▁or -6.35597
+▁all -6.40679
+▁but -6.43627
+▁no -6.46792
+▁had -6.48395
+▁do -6.48691
+▁will -6.55129
+▁been -6.55421
+ra -6.57141
+▁up -6.57759
+i -6.6063
+▁like -6.62078
+▁my -6.63622
+▁other -6.65515
+▁some -6.66346
+p -6.69
+re -6.71963
+g -6.73072
+▁when -6.73958
+▁time -6.7417
+▁out -6.74658
+▁them -6.75395
+u -6.76823
+▁go -6.77046
+▁many -6.78469
+▁during -6.83054
+▁these -6.83631
+▁new -6.83848
+▁three -6.84806
+▁don -6.84849
+▁only -6.85457
+▁la -6.85924
+▁city -6.86119
+▁most -6.87453
+▁if -6.88147
+ro -6.89559
+▁what -6.90568
+▁so -6.90629
+▁more -6.90861
+ll -6.91454
+in -6.94419
+▁part -6.94507
+▁would -6.94597
+▁get -6.94812
+la -6.96748
+ic -6.97571
+▁him -6.9772
+▁about -6.98807
+▁located -6.99935
+▁same -7.00998
+▁after -7.01547
+▁made -7.02773
+▁used -7.04629
+f -7.04752
+ation -7.05237
+▁did -7.05629
+▁re -7.06367
+▁al -7.06464
+▁later -7.06544
+▁o -7.07278
+or -7.07475
+▁year -7.08427
+▁who -7.08821
+very -7.09332
+le -7.09522
+▁work -7.09828
+▁which -7.09895
+ri -7.11967
+▁however -7.13105
+▁name -7.13346
+▁severa -7.13463
+ve -7.1359
+▁both -7.15
+ta -7.15707
+▁where -7.17104
+▁well -7.17221
+▁k -7.18521
+ar -7.18571
+▁look -7.19578
+nt -7.20013
+▁years -7.20357
+▁currently -7.20644
+en -7.20971
+▁good -7.21667
+▁call -7.21764
+▁team -7.22158
+▁found -7.22399
+▁different -7.23375
+an -7.2399
+out -7.24351
+ch -7.25244
+▁place -7.25673
+▁people -7.26314
+na -7.26808
+te -7.26847
+▁come -7.26962
+▁your -7.27617
+b -7.27889
+▁make -7.28805
+ion -7.29081
+k -7.29579
+▁over -7.30169
+▁last -7.30546
+us -7.30928
+▁s -7.31238
+ma -7.31502
+▁known -7.32409
+▁school -7.32553
+ist -7.32714
+▁great -7.32887
+h -7.33379
+▁main -7.33764
+▁see -7.33809
+▁university -7.34687
+▁because -7.35333
+▁life -7.36151
+▁national -7.36338
+▁second -7.37604
+▁than -7.38601
+▁then -7.38638
+▁back -7.39356
+▁ver -7.39963
+▁c -7.41057
+▁way -7.41134
+to -7.41161
+lo -7.41256
+z -7.4129
+▁each -7.42767
+▁know -7.42822
+▁want -7.43381
+▁let -7.43441
+on -7.43554
+▁man -7.44263
+▁war -7.44673
+▁into -7.44688
+▁love -7.44903
+▁bet -7.45216
+▁does -7.45427
+able -7.45499
+▁just -7.45618
+▁world -7.45969
+w -7.4603
+▁now -7.46301
+▁born -7.46513
+▁end -7.46803
+ween -7.46879
+▁us -7.47025
+▁general -7.47613
+▁four -7.48672
+▁being -7.48738
+co -7.4879
+ca -7.50274
+th -7.50512
+ne -7.50803
+ia -7.51359
+▁could -7.51552
+li -7.52186
+▁san -7.52286
+va -7.54083
+▁since -7.54283
+ok -7.54641
+▁got -7.54914
+ur -7.5515
+ment -7.55537
+▁take -7.55778
+▁family -7.55854
+▁any -7.56009
+▁even -7.56188
+▁p -7.56776
+▁old -7.56859
+ce -7.57263
+▁played -7.5768
+▁right -7.57873
+▁el -7.58526
+▁never -7.58827
+▁high -7.60474
+▁album -7.60768
+▁son -7.60949
+▁through -7.61399
+il -7.61592
+▁long -7.61802
+▁am -7.61872
+de -7.6233
+▁small -7.62431
+▁car -7.62589
+do -7.62673
+▁always -7.63534
+▁song -7.64032
+▁water -7.64224
+▁under -7.64631
+x -7.65009
+▁considered -7.65721
+▁co -7.65755
+▁music -7.65966
+▁started -7.66794
+▁those -7.67271
+▁important -7.67271
+▁area -7.67352
+▁still -7.68171
+▁group -7.68512
+▁before -7.68623
+ers -7.68635
+▁also -7.69106
+▁river -7.69546
+▁left -7.69995
+▁father -7.69997
+ting -7.70205
+ter -7.7024
+vi -7.70269
+▁studied -7.70688
+▁con -7.71012
+▁club -7.72088
+no -7.73229
+▁another -7.7327
+ive -7.73643
+▁little -7.73747
+▁center -7.73747
+ity -7.74849
+el -7.75154
+▁day -7.758
+▁going -7.76268
+▁works -7.76431
+▁li -7.76497
+ian -7.76836
+▁spanish -7.77149
+▁such -7.77174
+▁worked -7.7807
+▁may -7.78077
+ism -7.78548
+▁among -7.789
+▁due -7.7906
+se -7.79405
+▁town -7.79514
+ck -7.7958
+▁head -7.79767
+▁children -7.80161
+▁state -7.80466
+▁un -7.80903
+▁down -7.80965
+▁think -7.82163
+▁turn -7.82221
+▁ship -7.82653
+▁white -7.83009
+da -7.83473
+▁e -7.83595
+ies -7.84043
+ous -7.84089
+▁black -7.84861
+▁st -7.85172
+▁how -7.85638
+st -7.86106
+▁girl -7.862
+ul -7.87168
+▁five -7.87349
+▁house -7.87889
+me -7.88119
+▁play -7.883
+▁again -7.88597
+▁south -7.88929
+▁species -7.89476
+▁country -7.89793
+▁band -7.8995
+▁young -7.90222
+▁together -7.90344
+▁pro -7.90382
+▁member -7.90724
+▁order -7.9074
+▁film -7.90807
+it -7.91661
+ti -7.92007
+ga -7.92075
+▁t -7.921
+▁region -7.92409
+▁usually -7.92952
+man -7.93049
+▁ro -7.93059
+est -7.93524
+ry -7.94062
+▁sea -7.96134
+▁series -7.96162
+▁find -7.96841
+▁best -7.97379
+came -7.97489
+▁won -7.97966
+▁use -7.98222
+▁international -7.98265
+▁career -7.98365
+▁ca -7.98459
+▁received -7.98729
+▁need -7.99106
+- -7.99186
+▁short -7.99186
+▁married -7.99186
+os -7.99485
+▁president -7.99497
+ba -7.99627
+▁night -7.99805
+▁said -7.99821
+▁better -7.9989
+ary -8.00069
+▁f -8.00093
+▁united -8.00379
+▁church -8.00531
+▁unti -8.00576
+is -8.00861
+▁show -8.0097
+▁must -8.01109
+▁spain -8.01383
+▁public -8.01594
+▁di -8.01687
+▁away -8.01947
+▁next -8.01999
+▁party -8.02255
+▁big -8.02385
+▁b -8.02936
+▁following -8.03231
+▁capital -8.03274
+▁help -8.03294
+▁say -8.03321
+▁original -8.04821
+um -8.05072
+id -8.05158
+▁death -8.05548
+▁against -8.05959
+mo -8.0604
+▁single -8.06239
+▁vi -8.06564
+▁na -8.06737
+▁system -8.06824
+▁here -8.06834
+▁home -8.06888
+▁mo -8.06921
+tic -8.07192
+▁should -8.07242
+▁bo -8.07344
+▁del -8.07358
+▁give -8.07409
+▁history -8.0754
+ton -8.07544
+▁rest -8.07843
+▁mother -8.07884
+▁came -8.07905
+▁mainly -8.07919
+▁six -8.08169
+▁live -8.08217
+▁close -8.08311
+▁finally -8.08459
+▁lot -8.08837
+ted -8.09021
+▁red -8.09113
+▁too -8.09748
+▁tell -8.09804
+▁pi -8.10711
+▁our -8.10748
+que -8.10751
+▁belong -8.10945
+▁movie -8.11026
+▁natural -8.11292
+▁moment -8.1135
+▁states -8.11754
+▁v -8.12016
+▁while -8.12069
+▁much -8.12069
+▁hand -8.12328
+▁ma -8.12423
+go -8.12784
+▁named -8.13152
+▁game -8.13262
+▁se -8.13512
+▁y -8.13668
+pe -8.13753
+un -8.13759
+▁north -8.13777
+▁around -8.14114
+▁nothing -8.1412
+▁per -8.14736
+▁lives -8.14915
+▁d -8.14916
+▁season -8.15152
+▁third -8.15193
+▁few -8.15227
+ent -8.15436
+ty -8.15582
+tor -8.15904
+▁built -8.15919
+▁ta -8.16208
+▁province -8.16284
+mp -8.16404
+▁building -8.16434
+▁mar -8.16528
+▁light -8.1669
+land -8.16737
+ge -8.16839
+▁law -8.17022
+▁side -8.17965
+▁own -8.17979
+po -8.1805
+▁color -8.18118
+▁already -8.18129
+▁company -8.18129
+▁saint -8.1813
+di -8.1818
+▁open -8.18206
+▁something -8.18504
+▁government -8.18511
+▁published -8.18664
+▁died -8.19018
+▁bar -8.19053
+mi -8.19407
+▁include -8.19478
+▁lo -8.19487
+ness -8.19608
+▁en -8.19816
+▁period -8.20009
+most -8.20133
+et -8.20353
+▁american -8.20458
+▁seem -8.2054
+▁mi -8.20718
+▁today -8.21164
+▁station -8.21293
+ate -8.21332
+ble -8.21393
+▁written -8.21539
+▁eyes -8.21747
+▁every -8.21772
+▁book -8.21991
+▁daughter -8.22312
+▁art -8.22476
+▁player -8.22489
+bi -8.22791
+▁le -8.22828
+ir -8.22897
+▁plays -8.23265
+▁po -8.23652
+▁special -8.23784
+▁local -8.23835
+les -8.23918
+▁direct -8.24124
+▁pa -8.24405
+▁everything -8.2467
+▁set -8.24698
+▁ba -8.24716
+▁land -8.24992
+▁leave -8.25849
+▁political -8.2587
+▁line -8.26675
+▁times -8.2698
+tion -8.27377
+▁based -8.27404
+▁district -8.27903
+io -8.28012
+ish -8.2821
+as -8.2823
+▁number -8.28727
+▁te -8.28752
+▁human -8.28762
+▁production -8.29174
+▁park -8.29655
+▁division -8.30405
+▁full -8.30892
+ated -8.30994
+▁island -8.31112
+ine -8.31438
+▁songs -8.31844
+▁pre -8.31851
+▁lost -8.32732
+▁education -8.32958
+son -8.33204
+▁maria -8.33685
+▁hu -8.34146
+▁style -8.34262
+▁nowadays -8.34699
+▁similar -8.34699
+▁studies -8.34699
+▁eight -8.34707
+▁fight -8.35147
+▁chole -8.35163
+▁dark -8.35193
+gu -8.35207
+▁director -8.35255
+▁moved -8.35461
+▁cor -8.35701
+▁case -8.35728
+sh -8.35824
+▁ga -8.35874
+▁honor -8.36028
+▁current -8.3614
+▁beside -8.36162
+▁put -8.36241
+▁strong -8.36922
+▁please -8.36922
+▁title -8.36924
+▁common -8.36927
+nd -8.36936
+▁seen -8.36982
+▁position -8.37006
+ical -8.37061
+▁become -8.37375
+▁central -8.37824
+▁fall -8.38332
+▁social -8.3861
+▁juan -8.38921
+der -8.39443
+▁military -8.39655
+▁green -8.39677
+▁bad -8.39758
+▁off -8.39796
+▁mu -8.39925
+ru -8.39941
+bo -8.40006
+▁pu -8.40059
+▁free -8.40244
+ph -8.40463
+▁brother -8.40817
+▁person -8.41112
+▁start -8.41218
+▁final -8.41275
+▁friend -8.41284
+▁once -8.41586
+ut -8.41761
+▁body -8.41994
+ff -8.42097
+at -8.42172
+▁lu -8.4241
+▁degree -8.42466
+▁league -8.42466
+▁real -8.4275
+▁popular -8.42898
+▁madrid -8.42942
+▁support -8.43118
+za -8.4316
+▁less -8.43189
+be -8.43234
+▁stop -8.43313
+sa -8.43371
+▁origin -8.43382
+▁sister -8.43429
+age -8.43455
+ver -8.44154
+▁present -8.44237
+▁seven -8.44384
+▁space -8.444
+▁ra -8.44493
+vo -8.44553
+ig -8.44856
+▁half -8.44883
+▁age -8.44948
+▁ten -8.44976
+▁sun -8.45231
+▁men -8.45238
+▁large -8.45445
+▁eat -8.45466
+ie -8.45467
+▁process -8.45738
+▁ex -8.4582
+▁according -8.45848
+▁front -8.45848
+▁areas -8.45905
+▁professional -8.45927
+ant -8.46019
+▁pe -8.46174
+▁population -8.4634
+ina -8.46462
+ke -8.46606
+▁door -8.46801
+▁ho -8.46838
+▁field -8.46911
+v -8.47881
+▁boy -8.4795
+▁wh -8.47991
+▁role -8.48176
+im -8.48454
+▁th -8.48605
+▁ja -8.49266
+▁power -8.49331
+▁consist -8.49355
+▁german -8.49451
+less -8.49796
+▁sometimes -8.4986
+pa -8.49874
+▁twenty -8.50371
+▁inside -8.50584
+▁project -8.50639
+▁ri -8.50712
+▁example -8.50886
+▁story -8.50907
+▁decided -8.51087
+▁carrie -8.51342
+▁museum -8.51405
+▁flowers -8.51438
+res -8.5153
+▁began -8.51572
+▁los -8.51752
+ni -8.51914
+▁blue -8.51926
+▁native -8.51929
+▁count -8.52069
+▁sa -8.52392
+▁leaves -8.52477
+▁hard -8.52645
+▁dr -8.52922
+▁kind -8.5298
+▁europe -8.53433
+▁soon -8.53825
+▁act -8.53853
+▁charge -8.54064
+▁participated -8.54202
+▁council -8.54571
+▁women -8.54571
+▁plant -8.54647
+▁really -8.55044
+▁famous -8.55109
+▁created -8.55363
+▁julia -8.55537
+▁himself -8.55649
+▁things -8.55667
+▁point -8.55672
+▁ne -8.55878
+▁east -8.55898
+▁county -8.56075
+▁shows -8.56247
+ling -8.56288
+▁ar -8.56376
+▁mean -8.56502
+▁economic -8.56739
+▁job -8.56795
+though -8.56979
+▁video -8.57289
+▁beginning -8.57289
+▁room -8.57343
+ard -8.57381
+▁g -8.575
+▁face -8.58003
+▁floor -8.58397
+cy -8.58499
+▁port -8.58847
+▁talk -8.58929
+▁u -8.59034
+ria -8.59061
+▁members -8.59322
+▁mon -8.59429
+▁bu -8.59681
+ized -8.60106
+▁africa -8.60178
+▁continued -8.60311
+▁change -8.60328
+▁produced -8.60383
+▁games -8.60391
+▁included -8.60631
+▁anything -8.60651
+▁movement -8.60701
+ial -8.60718
+▁star -8.61264
+▁ti -8.61366
+▁form -8.61473
+▁hundred -8.61796
+▁singer -8.61948
+▁result -8.61967
+▁taken -8.6199
+▁having -8.62381
+▁everyone -8.62385
+ny -8.62416
+▁possible -8.62445
+▁lived -8.62535
+▁chapter -8.62956
+▁takes -8.63026
+▁author -8.63114
+▁king -8.63199
+▁friends -8.63234
+▁republic -8.63266
+▁fl -8.63404
+▁hi -8.63453
+▁cross -8.63458
+▁roman -8.63529
+▁battle -8.63541
+▁sound -8.63541
+▁civil -8.63559
+▁fish -8.63605
+ha -8.63838
+ja -8.63849
+▁da -8.63876
+▁fact -8.64014
+▁remains -8.64054
+ho -8.64068
+▁keep -8.64129
+▁dedicat -8.6413
+▁brown -8.64134
+▁fe -8.64456
+▁date -8.64464
+▁comes -8.64505
+▁gu -8.64593
+▁feel -8.6477
+▁move -8.6497
+▁done -8.65032
+▁others -8.65119
+▁musical -8.65127
+▁character -8.65181
+▁theater -8.65402
+▁sub -8.65617
+▁tu -8.65645
+▁return -8.65663
+nes -8.65705
+▁france -8.65834
+▁br -8.65834
+▁couple -8.65915
+▁village -8.65955
+▁office -8.6625
+▁far -8.66452
+▁professor -8.66517
+▁buried -8.66519
+▁football -8.66519
+ally -8.66573
+ber -8.66632
+▁indian -8.66814
+▁chile -8.67118
+▁argentina -8.67123
+▁mountain -8.67162
+▁du -8.67299
+▁rock -8.67374
+▁official -8.67388
+ful -8.67612
+▁influence -8.67733
+▁middle -8.67733
+▁round -8.67733
+▁study -8.67755
+▁municipality -8.6776
+▁opera -8.67841
+bu -8.67993
+▁west -8.67994
+fi -8.68022
+hi -8.68057
+▁version -8.68206
+▁wait -8.68409
+ka -8.68476
+▁language -8.68844
+men -8.68844
+▁john -8.68937
+▁means -8.69105
+▁covered -8.69121
+▁beautiful -8.69585
+op -8.69696
+▁days -8.69697
+▁base -8.69915
+▁wind -8.6993
+ac -8.70055
+▁lead -8.70347
+ana -8.70373
+▁lower -8.70422
+▁ru -8.70436
+▁dis -8.70573
+▁hands -8.70647
+▁program -8.70808
+▁court -8.7084
+▁master -8.70866
+▁female -8.71472
+▁woman -8.71494
+▁stone -8.7157
+▁groups -8.71786
+▁reach -8.72009
+▁die -8.72193
+▁thing -8.72195
+▁goes -8.72274
+▁cha -8.72289
+▁newspaper -8.7275
+▁held -8.72881
+▁developed -8.73168
+▁met -8.7318
+▁try -8.73484
+ology -8.73535
+▁fire -8.73554
+▁actor -8.73574
+▁addition -8.73759
+▁pri -8.7396
+▁books -8.74037
+▁street -8.74161
+▁nine -8.74222
+▁early -8.74238
+▁past -8.74308
+▁habitat -8.74391
+▁radio -8.74695
+▁kept -8.74698
+▁information -8.74699
+▁stage -8.74776
+▁god -8.74937
+▁cu -8.75002
+▁bi -8.75124
+▁m -8.75673
+▁late -8.75852
+▁coast -8.76019
+▁garden -8.76024
+▁development -8.76044
+▁jo -8.76499
+▁doctor -8.76657
+qui -8.7666
+con -8.76682
+▁department -8.76684
+▁modern -8.76686
+▁television -8.76686
+▁board -8.76859
+▁level -8.76929
+jo -8.77114
+▁attack -8.77177
+▁follow -8.77275
+▁york -8.77356
+▁teacher -8.77476
+▁cover -8.7761
+▁mat -8.77774
+▁ha -8.78024
+▁soccer -8.78031
+▁service -8.78339
+▁tra -8.78518
+▁sir -8.78574
+▁parents -8.78685
+▁allow -8.78693
+▁happy -8.78712
+▁wife -8.78712
+▁activity -8.78718
+▁european -8.78803
+▁oh -8.79302
+▁yellow -8.79371
+▁near -8.79482
+▁r -8.79636
+▁traditional -8.797
+▁voice -8.79777
+▁relationship -8.80092
+▁access -8.80094
+ance -8.80102
+▁su -8.8015
+▁enter -8.80177
+▁toward -8.80215
+▁painting -8.80233
+ell -8.80266
+▁divided -8.80267
+am -8.80412
+▁plan -8.80497
+▁type -8.80722
+▁seat -8.80756
+▁wrote -8.80781
+▁wood -8.80791
+▁headquarters -8.80871
+ric -8.81209
+▁comp -8.81426
+▁food -8.81481
+▁neighborhood -8.81482
+▁tower -8.81509
+▁road -8.81833
+▁run -8.81877
+mer -8.82014
+▁em -8.82122
+▁ground -8.82184
+▁composed -8.82185
+▁months -8.822
+ron -8.82383
+▁magazine -8.82704
+up -8.82781
+▁summer -8.82898
+▁alone -8.82946
+▁china -8.82956
+▁ones -8.82961
+▁saw -8.83001
+▁sp -8.83102
+▁historical -8.83224
+▁hours -8.83253
+▁dog -8.83398
+▁whole -8.83404
+▁exist -8.83444
+▁low -8.83481
+ions -8.83585
+▁child -8.83605
+▁along -8.83632
+▁released -8.83824
+▁working -8.84049
+fa -8.8418
+▁structure -8.84327
+▁tournament -8.84327
+▁regard -8.84327
+▁depend -8.84328
+▁countries -8.84346
+▁train -8.84494
+▁writer -8.84551
+▁told -8.84556
+▁union -8.85344
+▁inter -8.85637
+▁commercial -8.85782
+▁performance -8.85791
+▁design -8.85809
+lin -8.85844
+▁playing -8.85858
+▁completely -8.85978
+ler -8.86054
+▁arrived -8.86159
+pi -8.86221
+ino -8.86348
+▁appeared -8.86417
+▁paris -8.86468
+▁week -8.86495
+▁community -8.86517
+▁research -8.86517
+▁tour -8.86781
+ley -8.86851
+▁provide -8.86908
+▁kill -8.86992
+▁gi -8.87133
+▁organization -8.87265
+▁hair -8.87288
+▁cap -8.87506
+▁idea -8.87617
+▁test -8.87796
+▁air -8.88017
+rre -8.88265
+▁sure -8.88378
+▁sports -8.88386
+▁model -8.88458
+▁finished -8.88542
+▁ni -8.88554
+▁british -8.88756
+▁border -8.88756
+▁kiss -8.88767
+tes -8.89127
+▁previous -8.89169
+▁ended -8.89234
+▁performed -8.89274
+▁top -8.89329
+▁california -8.89513
+▁competition -8.89513
+▁episode -8.89513
+▁winner -8.89528
+▁section -8.8954
+▁sorry -8.89566
+mon -8.89883
+▁para -8.89994
+▁control -8.90262
+▁activities -8.90277
+▁mexican -8.90277
+▁palomares -8.90278
+▁someone -8.9034
+▁dead -8.9045
+ries -8.90485
+▁nor -8.90808
+ster -8.91063
+▁edition -8.91073
+▁arm -8.91104
+▁castle -8.91174
+▁hall -8.91198
+▁range -8.91208
+▁often -8.9125
+per -8.91252
+▁chi -8.91265
+ization -8.91277
+▁goal -8.9135
+ure -8.91534
+ner -8.91707
+▁academy -8.91821
+▁ge -8.92025
+ero -8.92094
+▁given -8.9213
+▁list -8.92135
+gi -8.92207
+▁represent -8.92437
+▁catholic -8.92602
+refore -8.92607
+▁cultural -8.92618
+▁distributed -8.92642
+▁rain -8.92767
+▁course -8.92866
+▁served -8.92998
+▁royal -8.93392
+▁trip -8.93445
+▁forests -8.93494
+▁class -8.93741
+▁believe -8.93753
+▁shape -8.93774
+▁event -8.9381
+▁h -8.93844
+▁luis -8.94256
+▁managed -8.94404
+▁returned -8.94518
+▁size -8.94755
+▁post -8.94953
+▁tropical -8.94984
+▁unknown -8.94994
+▁culture -8.95016
+▁winter -8.95028
+▁sent -8.95349
+▁fruit -8.95364
+▁reason -8.95364
+▁lake -8.95458
+▁record -8.95559
+cia -8.95745
+▁business -8.95755
+▁cannot -8.95795
+▁caused -8.96109
+▁ka -8.96154
+▁thought -8.96398
+▁dry -8.96486
+ki -8.96585
+▁private -8.96603
+▁various -8.96603
+▁stories -8.96609
+▁recorded -8.96644
+▁raised -8.96712
+▁classic -8.9683
+▁simple -8.9708
+ow -8.97231
+▁leader -8.97268
+▁ce -8.9729
+▁major -8.97379
+▁sin -8.97381
+▁territory -8.97423
+▁enough -8.97423
+▁relative -8.97423
+au -8.97479
+▁santa -8.97816
+▁force -8.97961
+▁dance -8.98085
+▁hope -8.98089
+ium -8.98117
+▁consider -8.98223
+▁beach -8.98243
+▁queen -8.9825
+▁related -8.9825
+nce -8.98618
+▁w -8.9891
+▁wanted -8.9893
+▁trees -8.99193
+▁closed -8.99336
+▁hold -8.99348
+▁islands -8.99655
+▁sta -8.99661
+gue -8.99704
+▁trans -8.9992
+▁barcelona -8.99923
+▁london -8.99923
+▁square -8.99924
+▁championship -8.99933
+illa -9.00026
+▁heart -9.00066
+▁male -9.00281
+rs -9.00438
+ten -9.00505
+▁sleep -9.0077
+▁dream -9.0077
+▁french -9.0077
+▁pass -9.00858
+▁thank -9.009
+▁chosen -9.00915
+ction -9.01117
+▁coming -9.01221
+▁remained -9.01563
+▁certain -9.01625
+▁russian -9.01879
+▁nu -9.01984
+▁teams -9.02122
+line -9.02186
+▁institute -9.02217
+ities -9.02282
+▁bring -9.02479
+▁hospital -9.02487
+▁representative -9.02496
+▁active -9.02559
+▁care -9.02636
+▁poor -9.02671
+▁learn -9.02766
+▁clear -9.03011
+tu -9.03034
+▁critics -9.03091
+▁anti -9.03189
+▁behind -9.03357
+▁complex -9.03357
+▁actress -9.03358
+▁francisco -9.03373
+▁western -9.03374
+▁sign -9.0352
+ng -9.0356
+▁fourth -9.03619
+▁gar -9.03644
+▁australia -9.03764
+▁brazil -9.04191
+▁artist -9.0421
+▁represented -9.04313
+▁experience -9.04321
+▁win -9.04535
+▁plants -9.04549
+▁asia -9.04604
+▁mass -9.0477
+gen -9.04885
+▁forest -9.04942
+▁gav -9.05086
+▁carlos -9.0513
+▁month -9.05234
+▁race -9.05251
+▁stay -9.05596
+▁birds -9.05633
+▁events -9.05816
+▁germany -9.0584
+ad -9.05871
+▁match -9.059
+▁situation -9.06012
+▁image -9.06015
+▁bridge -9.06018
+▁construction -9.06031
+let -9.0604
+▁tail -9.06132
+▁calm -9.06236
+▁joined -9.06561
+ship -9.06562
+▁understand -9.06913
+▁ancient -9.06913
+▁kingdom -9.06913
+▁chapel -9.06913
+▁health -9.06924
+▁touch -9.06963
+▁cities -9.06977
+▁save -9.07061
+▁true -9.07067
+▁shot -9.07297
+▁cur -9.07317
+▁treat -9.07403
+▁bell -9.07545
+▁though -9.07665
+ub -9.07777
+▁italian -9.07833
+▁travel -9.07863
+▁che -9.07897
+▁former -9.07993
+▁remain -9.08004
+▁har -9.08051
+▁bra -9.08183
+ther -9.08228
+▁word -9.08241
+▁cold -9.08248
+▁starts -9.083
+▁fi -9.08628
+▁champion -9.08727
+▁accident -9.08739
+▁venezuela -9.08739
+▁camp -9.0884
+▁graduated -9.08881
+▁host -9.08958
+▁italy -9.08965
+▁pen -9.0901
+▁table -9.09013
+▁peru -9.09075
+▁super -9.09192
+▁youth -9.09215
+▁cal -9.09233
+▁wide -9.09327
+▁est -9.0933
+▁throw -9.09665
+▁japan -9.09671
+▁matter -9.09733
+▁silver -9.09799
+▁mayor -9.10019
+ros -9.10349
+he -9.10521
+▁gold -9.10555
+ious -9.10564
+▁blood -9.10599
+▁literature -9.10599
+▁primar -9.10648
+▁outside -9.10677
+▁maybe -9.10683
+▁wi -9.1071
+▁bank -9.10803
+▁science -9.10843
+▁novel -9.10851
+▁sand -9.10958
+▁cup -9.10979
+▁store -9.10981
+▁characters -9.11041
+▁established -9.11064
+▁appears -9.11112
+▁buen -9.11285
+▁movies -9.11322
+▁site -9.11509
+▁religious -9.11543
+▁airport -9.1155
+▁villa -9.11553
+▁earth -9.11589
+▁bur -9.11664
+▁priest -9.1167
+▁fell -9.1175
+▁cut -9.11779
+▁cast -9.12016
+▁meeting -9.12045
+▁rule -9.12175
+act -9.12482
+▁association -9.12495
+▁environment -9.12495
+▁festival -9.12495
+▁individual -9.12495
+▁santiago -9.12495
+▁greek -9.12495
+▁temple -9.12495
+▁appearance -9.1251
+▁market -9.12521
+▁feed -9.12894
+▁tri -9.13403
+▁difficult -9.13414
+▁secret -9.13432
+▁agriculture -9.13457
+▁replaced -9.13457
+▁break -9.13457
+▁writing -9.13457
+▁rome -9.13621
+▁formed -9.14028
+▁media -9.14356
+▁passed -9.14429
+▁making -9.14429
+▁typical -9.14433
+▁numer -9.14452
+▁collection -9.14462
+▁spen -9.14555
+▁fine -9.14634
+▁defeated -9.14853
+▁obtained -9.14912
+▁appear -9.1499
+ence -9.15109
+ang -9.15173
+▁christian -9.15389
+car -9.15401
+▁college -9.15408
+▁least -9.15413
+▁grew -9.15457
+▁zone -9.15503
+▁visit -9.1553
+▁owner -9.15582
+▁sha -9.15913
+▁sport -9.15936
+▁unit -9.15976
+▁arms -9.16215
+the -9.16226
+▁colombia -9.16259
+▁imp -9.16308
+▁stand -9.16326
+▁municipal -9.16342
+▁themselves -9.16398
+▁valencia -9.16398
+▁devot -9.16448
+▁mission -9.16508
+▁hill -9.16656
+▁added -9.16913
+▁bit -9.1693
+▁horse -9.16964
+▁hear -9.1708
+▁pale -9.17387
+▁favor -9.17388
+▁especially -9.17398
+▁literary -9.17398
+▁location -9.17399
+▁avoid -9.17399
+▁stopped -9.17403
+▁worth -9.17495
+▁remember -9.17545
+▁designed -9.17563
+▁moon -9.17658
+▁fan -9.1767
+▁problem -9.18172
+ica -9.18232
+▁pan -9.18262
+▁text -9.18384
+▁available -9.18408
+▁independent -9.18408
+▁probabl -9.18413
+▁philosophy -9.18591
+▁rep -9.18599
+led -9.18767
+▁listen -9.18937
+ue -9.19265
+▁personal -9.19331
+▁economy -9.19429
+▁positive -9.19429
+▁twelve -9.19429
+▁marriage -9.19429
+▁ricardo -9.19429
+▁spring -9.1943
+▁initial -9.19431
+▁action -9.19462
+▁necessary -9.19473
+▁ci -9.19721
+▁led -9.19772
+▁includ -9.19877
+▁places -9.19934
+▁stra -9.19936
+▁deep -9.20145
+▁focus -9.2046
+▁urban -9.2046
+▁pedro -9.20465
+▁silent -9.20479
+▁police -9.20481
+▁award -9.20697
+▁rich -9.20852
+▁walk -9.20918
+▁brothers -9.20982
+▁doing -9.20984
+▁arch -9.21007
+▁industry -9.21501
+▁spread -9.21501
+▁captain -9.21501
+▁crown -9.21501
+▁studio -9.21501
+▁money -9.21519
+▁poet -9.21549
+▁immediately -9.21564
+▁none -9.21706
+▁gra -9.21958
+wise -9.21967
+▁services -9.22078
+▁non -9.22119
+▁si -9.22234
+▁results -9.22242
+spite -9.22253
+▁founded -9.2249
+cha -9.2251
+▁cemetery -9.22554
+▁thousand -9.22554
+▁channel -9.22554
+▁institution -9.22605
+ak -9.22621
+▁reached -9.22909
+▁eu -9.22939
+▁continue -9.22974
+lan -9.23144
+▁flow -9.2335
+▁l -9.23377
+▁practice -9.23385
+▁fifth -9.23618
+▁minister -9.23618
+▁data -9.23754
+▁organized -9.23779
+▁total -9.2386
+▁thanks -9.23889
+▁followed -9.23893
+des -9.24034
+ite -9.24104
+▁sal -9.24174
+▁function -9.24313
+▁happened -9.24384
+▁presented -9.24429
+▁fast -9.24534
+▁guitar -9.24626
+▁future -9.247
+▁favorite -9.24704
+▁route -9.24705
+▁living -9.24728
+▁guide -9.24739
+▁latin -9.24745
+▁sold -9.24763
+nia -9.24766
+▁matches -9.24956
+▁elected -9.25032
+▁completed -9.25059
+▁sad -9.25163
+▁tradition -9.253
+▁par -9.25375
+▁elements -9.25669
+▁grow -9.25777
+▁bishop -9.2578
+▁subject -9.2578
+▁answer -9.2578
+▁watch -9.25782
+▁felt -9.2579
+▁lady -9.258
+▁receive -9.2591
+▁wa -9.25997
+▁recognized -9.26005
+eless -9.26031
+▁aires -9.26104
+▁wor -9.26142
+▁beat -9.26175
+▁mostly -9.26225
+▁com -9.26347
+▁cause -9.26379
+▁foundation -9.26447
+▁deck -9.26606
+▁cra -9.26608
+▁changed -9.26612
+▁prison -9.26867
+▁key -9.26886
+▁palace -9.26901
+▁southern -9.26921
+▁duck -9.26952
+▁pet -9.26993
+▁seeds -9.27132
+▁allowed -9.2716
+▁fly -9.27175
+▁opened -9.27199
+hu -9.27315
+▁symbol -9.27382
+▁officer -9.2739
+way -9.27391
+▁arts -9.27436
+▁decide -9.27558
+kin -9.27627
+▁showed -9.27655
+▁mountains -9.27845
+▁campaign -9.2799
+▁mouth -9.2799
+▁double -9.27991
+▁taking -9.27991
+▁afterwards -9.27993
+ita -9.28029
+▁talking -9.28043
+▁yet -9.2809
+▁tree -9.28275
+▁valley -9.28292
+ea -9.28489
+ern -9.28718
+▁parts -9.28815
+▁approach -9.29114
+▁view -9.29114
+▁meet -9.29151
+▁figure -9.29174
+▁mor -9.29311
+▁artistic -9.29319
+▁founder -9.29458
+▁problems -9.29577
+▁hit -9.29748
+▁produce -9.29941
+bra -9.29994
+▁says -9.30017
+par -9.30081
+▁ulises -9.3025
+▁companies -9.3025
+▁cases -9.30277
+▁pick -9.3042
+era -9.30581
+▁awards -9.30681
+▁singing -9.30694
+▁lack -9.30712
+▁additional -9.30751
+▁pieces -9.30776
+▁ven -9.30803
+▁bro -9.30884
+mar -9.3104
+lon -9.31052
+▁entire -9.31052
+▁reviews -9.31066
+▁trade -9.31088
+▁wings -9.31183
+▁ear -9.31185
+▁specific -9.31382
+▁respect -9.31399
+▁foreign -9.31399
+▁purpose -9.31399
+▁uruguay -9.31399
+▁facade -9.31405
+▁medicine -9.31411
+▁cabin -9.31426
+▁antonio -9.31438
+▁poetry -9.31439
+cio -9.31445
+▁golden -9.31493
+▁specialized -9.31552
+▁material -9.31561
+▁doubt -9.31617
+▁path -9.31636
+▁burn -9.31853
+▁uses -9.31881
+▁build -9.31974
+▁attended -9.32004
+by -9.32044
+ena -9.32324
+▁fa -9.32386
+ile -9.32416
+▁administrative -9.32562
+▁appointed -9.32562
+▁august -9.32562
+▁strength -9.32562
+▁lord -9.32569
+▁brought -9.32579
+▁careful -9.32586
+▁frequently -9.32738
+▁upper -9.3278
+▁forward -9.33001
+ab -9.33121
+▁mixed -9.33134
+▁mal -9.33167
+▁yes -9.33264
+▁je -9.33433
+▁offer -9.33446
+▁write -9.33488
+wn -9.33527
+▁paul -9.33578
+▁ob -9.33616
+ez -9.33638
+▁husband -9.33739
+▁industrial -9.33739
+▁instrument -9.33739
+▁valeria -9.33739
+▁architecture -9.33752
+▁agreement -9.3376
+▁finish -9.33885
+▁versions -9.34027
+▁wave -9.34054
+▁publication -9.34067
+▁ski -9.34323
+▁happen -9.34377
+▁tro -9.34498
+▁forces -9.34503
+▁significant -9.34582
+▁words -9.34855
+▁neighbor -9.34926
+▁bolivia -9.34929
+▁climate -9.34929
+▁decision -9.34929
+▁weather -9.3493
+▁reference -9.34939
+▁niña -9.34941
+▁edge -9.34957
+▁track -9.35
+▁boat -9.35015
+▁journalist -9.35173
+▁wine -9.35228
+▁described -9.35261
+▁suffered -9.3545
+chi -9.35457
+▁kid -9.35525
+▁easter -9.3571
+▁mr -9.35713
+sis -9.36032
+ola -9.36072
+▁concept -9.36134
+▁empire -9.36134
+▁furthermore -9.36134
+▁measure -9.36134
+▁distance -9.36134
+▁minor -9.36206
+▁hot -9.3661
+ever -9.36831
+▁features -9.36849
+▁mari -9.36958
+▁regional -9.37051
+ado -9.37158
+▁refer -9.37173
+▁angel -9.37212
+▁signed -9.37317
+▁english -9.37353
+▁nobody -9.37353
+▁machine -9.37354
+▁multiple -9.37354
+▁stadium -9.37354
+▁spot -9.37358
+▁nature -9.37383
+▁northern -9.3739
+▁medal -9.37395
+▁governor -9.37405
+▁producer -9.37706
+▁students -9.37779
+▁oil -9.37862
+▁previously -9.37911
+gar -9.3803
+▁por -9.38062
+▁discovered -9.3808
+▁ask -9.38178
+▁training -9.38351
+▁fit -9.38435
+▁pay -9.38515
+▁sail -9.38539
+▁box -9.38552
+▁agricultural -9.38588
+▁entrance -9.38588
+▁society -9.38588
+▁disease -9.38588
+▁leaving -9.38588
+▁bread -9.38588
+▁david -9.38588
+▁risk -9.38595
+▁acting -9.38737
+cent -9.38899
+▁perform -9.38994
+fo -9.39074
+▁animal -9.39114
+▁bird -9.39117
+▁value -9.39203
+ide -9.39356
+over -9.3945
+ugh -9.39639
+▁ab -9.39768
+▁surface -9.39838
+▁monument -9.39838
+▁sector -9.39838
+▁vilma -9.39839
+▁content -9.39841
+▁search -9.39842
+▁iron -9.39879
+▁suddenly -9.39892
+▁march -9.39903
+ought -9.39932
+▁coach -9.39974
+▁fu -9.4
+▁meters -9.40007
+▁rank -9.40058
+▁languages -9.40081
+▁huge -9.40146
+▁walls -9.40201
+▁steps -9.40297
+▁hat -9.40398
+▁players -9.40571
+▁leon -9.40592
+▁engine -9.40623
+▁smile -9.40669
+▁costa -9.40725
+▁colors -9.40959
+▁contribut -9.41085
+▁architect -9.41089
+▁congress -9.41104
+▁block -9.41104
+▁bubble -9.41104
+▁guard -9.41104
+▁manuel -9.41106
+▁quickly -9.4126
+▁continues -9.41358
+▁abandoned -9.41517
+▁complete -9.41608
+▁press -9.41702
+ving -9.41927
+ons -9.4219
+▁pla -9.42306
+▁category -9.42386
+▁thirty -9.42386
+▁england -9.42386
+▁interior -9.42386
+▁phase -9.42387
+▁warm -9.42404
+ture -9.42405
+▁sick -9.42498
+form -9.42499
+▁surrounded -9.42537
+▁involved -9.42585
+▁semi -9.42648
+▁ready -9.4279
+ulate -9.43112
+▁limited -9.43147
+▁val -9.43162
+mina -9.43199
+▁bear -9.43407
+▁ju -9.43472
+dad -9.43485
+ker -9.43499
+▁india -9.43571
+az -9.43582
+▁miss -9.43655
+▁occur -9.43682
+▁cinema -9.43684
+▁cathedral -9.43685
+▁impossible -9.43685
+▁property -9.43685
+▁biggest -9.43685
+▁clean -9.43685
+▁houses -9.43687
+▁mistake -9.43687
+▁sky -9.43687
+▁lawyer -9.43689
+thus -9.43694
+▁mark -9.43831
+▁secondary -9.43956
+▁develop -9.44091
+▁report -9.44206
+▁speed -9.44274
+▁skin -9.44291
+▁stands -9.44332
+ger -9.4444
+▁stars -9.44762
+▁read -9.44785
+▁placed -9.44875
+ña -9.44948
+▁extreme -9.45
+▁administration -9.45
+▁scientific -9.45
+▁politics -9.45001
+▁strange -9.45004
+▁flag -9.45007
+▁crater -9.4506
+▁preserved -9.4509
+▁buildings -9.45128
+stood -9.45197
+▁quite -9.4532
+ju -9.45362
+▁mont -9.45369
+▁rein -9.45375
+▁accept -9.45478
+▁alive -9.45544
+▁soil -9.45634
+▁ban -9.45763
+▁opening -9.45764
+▁animals -9.4577
+▁student -9.45871
+ors -9.45923
+▁za -9.46021
+ida -9.46059
+▁relation -9.46326
+▁ainhoa -9.46334
+▁ministry -9.46334
+▁question -9.46334
+ches -9.46334
+▁childhood -9.4634
+▁dress -9.46362
+▁gather -9.46367
+▁artists -9.46431
+▁serve -9.46457
+▁killed -9.46538
+▁angeles -9.46561
+▁musician -9.46591
+▁mag -9.46753
+▁nest -9.46815
+ko -9.46845
+▁resid -9.46916
+▁legs -9.47097
+ize -9.47182
+▁dan -9.47228
+▁holy -9.47252
+▁characteriz -9.47343
+▁create -9.47633
+ative -9.47674
+é -9.47684
+▁aspect -9.47685
+▁commission -9.47685
+▁landscape -9.47685
+▁library -9.47685
+▁physical -9.47685
+▁variety -9.47685
+▁platform -9.47685
+▁memory -9.47686
+▁storm -9.47687
+▁thick -9.47702
+▁wall -9.47883
+▁ball -9.48186
+▁farm -9.48239
+▁types -9.48281
+▁van -9.484
+▁shaped -9.48432
+▁drink -9.48461
+▁accepted -9.4856
+▁planet -9.48574
+▁command -9.48611
+▁recording -9.48674
+▁sciences -9.48698
+▁contain -9.48705
+ano -9.48844
+▁fil -9.48844
+▁j -9.48892
+▁wear -9.48909
+▁object -9.49011
+pu -9.49047
+▁parliament -9.49055
+▁permanent -9.49055
+▁birth -9.49055
+▁breath -9.49055
+▁network -9.49059
+▁organ -9.49073
+▁fresh -9.49078
+▁communication -9.49092
+▁compete -9.49111
+▁term -9.49115
+▁script -9.49138
+▁generation -9.49159
+▁luck -9.49171
+▁rejected -9.49393
+▁legal -9.49488
+▁drive -9.49691
+▁fun -9.4972
+rew -9.49918
+▁participate -9.50011
+▁asked -9.50063
+▁remove -9.50089
+time -9.5028
+▁except -9.50292
+▁ran -9.50317
+▁correspond -9.50415
+▁defense -9.50444
+▁christmas -9.50446
+▁william -9.50446
+▁jump -9.50475
+▁classified -9.5049
+▁secretary -9.50494
+▁join -9.50518
+ify -9.50554
+▁branches -9.50582
+▁chilean -9.50608
+▁palm -9.5063
+com -9.50726
+gan -9.50849
+ward -9.50857
+han -9.50945
+▁soul -9.50963
+▁gas -9.51261
+▁starr -9.51572
+cross -9.5159
+▁ring -9.51794
+▁neither -9.51852
+▁contact -9.51853
+▁debut -9.51857
+▁bright -9.51864
+▁canada -9.52051
+▁bull -9.52079
+mes -9.52182
+umb -9.52317
+▁achieved -9.52355
+▁x -9.52403
+▁mentioned -9.52458
+▁sat -9.52567
+▁tre -9.52694
+▁cat -9.52698
+ua -9.52765
+▁release -9.52848
+▁rica -9.529
+▁step -9.53021
+ral -9.53079
+▁cy -9.53259
+▁suggest -9.53281
+▁bottom -9.53281
+▁candidate -9.53281
+▁column -9.53281
+▁conference -9.53281
+▁energy -9.53281
+▁everybody -9.53281
+▁miguel -9.53281
+▁constant -9.53281
+▁enjoy -9.53281
+▁gift -9.53283
+▁olive -9.53296
+▁discover -9.53306
+▁judge -9.53312
+▁ago -9.53313
+▁slight -9.53327
+▁treatment -9.53364
+▁tourism -9.53369
+▁liberal -9.53372
+▁pop -9.53407
+▁partner -9.53505
+den -9.53533
+▁flat -9.53585
+ier -9.53598
+▁suit -9.53681
+▁haven -9.53711
+▁metal -9.53829
+▁neverth -9.5383
+ification -9.53847
+▁mad -9.53859
+▁limit -9.53878
+▁heard -9.53904
+▁bill -9.53947
+▁horses -9.53955
+▁records -9.54019
+▁prince -9.54547
+▁source -9.54629
+▁mine -9.54649
+rry -9.54682
+▁ecuador -9.5473
+▁sweet -9.5473
+▁weapon -9.5473
+▁either -9.5473
+▁instead -9.5473
+▁northwest -9.5473
+▁translated -9.5473
+▁anymore -9.5473
+▁wrong -9.54732
+▁transport -9.54739
+▁might -9.54742
+▁soft -9.54748
+▁southeast -9.54749
+▁rural -9.5475
+▁points -9.54765
+▁deal -9.54854
+▁piti -9.54878
+▁higher -9.5496
+▁flower -9.55039
+▁rose -9.55096
+▁promoted -9.55102
+je -9.55157
+ological -9.55225
+▁films -9.55499
+▁chest -9.55556
+▁draw -9.5562
+▁lay -9.55667
+▁cost -9.55763
+▁slow -9.55856
+▁awarded -9.55957
+▁tried -9.56036
+burg -9.5607
+▁virgin -9.56144
+▁condition -9.56164
+▁essential -9.56201
+▁salvador -9.56201
+▁reserve -9.56201
+▁presence -9.56202
+▁foot -9.56205
+▁prize -9.56217
+▁creation -9.56229
+▁basic -9.56236
+ified -9.56274
+▁fra -9.56769
+fer -9.56788
+▁inhabit -9.56828
+▁vote -9.57104
+▁bal -9.57306
+▁happens -9.57316
+▁broke -9.57321
+while -9.57385
+board -9.57436
+▁medi -9.57582
+▁declared -9.57693
+▁gamboa -9.57693
+▁october -9.57693
+▁promotion -9.57693
+▁audience -9.57693
+▁chinese -9.57693
+▁computer -9.57693
+▁request -9.57694
+▁smell -9.57694
+▁composition -9.57695
+▁fishing -9.57697
+▁extra -9.57699
+▁editor -9.57708
+▁piano -9.57732
+▁fuck -9.57742
+▁surrounding -9.57814
+▁send -9.57816
+▁vice -9.57999
+bre -9.58113
+rin -9.58147
+rous -9.58278
+▁disappeared -9.5843
+side -9.58684
+▁require -9.58772
+work -9.58901
+▁positions -9.58978
+▁disc -9.59144
+▁monte -9.59185
+▁midfielder -9.59208
+▁cream -9.59208
+▁federal -9.59208
+▁southwest -9.59208
+▁taught -9.59208
+▁contest -9.5921
+▁medium -9.59241
+▁moving -9.59246
+▁truth -9.59248
+▁democratic -9.59279
+▁characteristic -9.5933
+▁lope -9.59339
+▁qui -9.59443
+▁highest -9.59485
+▁product -9.59497
+▁forms -9.59573
+▁mary -9.59595
+▁arrive -9.59644
+▁direction -9.59651
+▁ends -9.59716
+▁ice -9.59982
+▁rise -9.60137
+▁clo -9.60175
+▁bed -9.6023
+▁escape -9.60308
+▁card -9.60314
+▁streets -9.60469
+▁upon -9.60485
+▁products -9.60554
+▁fre -9.60738
+▁avenue -9.60747
+▁families -9.60747
+▁financial -9.60747
+▁heavy -9.60747
+▁knowledge -9.60747
+▁unique -9.60747
+▁bodies -9.60748
+▁morning -9.60751
+▁below -9.60752
+▁brand -9.60798
+▁tourist -9.60798
+▁chance -9.60802
+▁pink -9.60868
+▁contract -9.6098
+▁grey -9.61153
+▁hide -9.61207
+▁prove -9.61408
+ual -9.61414
+▁launch -9.61451
+ario -9.61517
+▁reasons -9.61569
+▁forty -9.61733
+▁levels -9.61738
+▁manage -9.61744
+▁gre -9.61754
+▁sil -9.61755
+dor -9.61784
+▁pain -9.6189
+▁regular -9.61964
+▁changes -9.6202
+▁robert -9.62045
+▁minutes -9.62178
+▁revolution -9.62274
+▁conflict -9.62309
+▁jewish -9.62309
+▁prevent -9.6231
+▁guinea -9.6231
+▁weird -9.62315
+▁fac -9.62343
+▁quiet -9.62343
+▁bigger -9.62351
+▁ramon -9.62397
+▁glo -9.62509
+▁princess -9.62539
+▁gives -9.62585
+▁introduced -9.62627
+▁gal -9.62629
+▁mode -9.62687
+▁eh -9.62701
+▁contains -9.62711
+matic -9.62798
+▁affected -9.62805
+can -9.62902
+▁fin -9.63016
+▁cho -9.63108
+▁marked -9.63353
+▁ur -9.63368
+▁russia -9.63426
+▁buy -9.63469
+▁mrs -9.6349
+lla -9.63542
+ito -9.63569
+▁valle -9.63611
+▁kings -9.63636
+old -9.63672
+ham -9.63718
+▁peace -9.6388
+▁account -9.63897
+▁distribution -9.63897
+▁michael -9.63897
+▁participation -9.63897
+▁twice -9.63897
+▁welcome -9.63897
+▁chief -9.63897
+▁orange -9.63899
+▁lips -9.63932
+▁sense -9.63957
+▁taste -9.63966
+▁fernando -9.6397
+▁older -9.6422
+▁destroyed -9.64319
+▁ride -9.64365
+▁schools -9.64418
+▁connect -9.64478
+▁fo -9.64554
+ade -9.64573
+ji -9.64686
+▁wing -9.64883
+▁min -9.64902
+▁christ -9.64968
+▁format -9.64988
+▁windows -9.65016
+▁offered -9.65097
+one -9.65199
+▁stayed -9.65451
+▁advance -9.6551
+▁mediterranean -9.6551
+▁detail -9.65511
+▁roof -9.65511
+▁running -9.65516
+▁successful -9.65519
+▁patron -9.65526
+cho -9.65528
+▁hurry -9.65536
+▁olympic -9.65536
+▁fair -9.65549
+▁romantic -9.656
+▁george -9.65635
+▁vas -9.65934
+▁zo -9.66122
+▁noble -9.66179
+ev -9.66185
+▁voices -9.66246
+▁attend -9.66297
+▁hey -9.66449
+house -9.66823
+▁issue -9.66931
+▁directly -9.66935
+▁convent -9.67007
+▁argentinian -9.67149
+▁independence -9.67149
+▁prominent -9.67149
+▁screen -9.67149
+▁easy -9.6715
+▁milk -9.6715
+▁above -9.67155
+▁objective -9.67202
+▁management -9.67205
+▁bay -9.67273
+▁associated -9.67289
+▁million -9.67304
+▁reduced -9.67318
+▁adventure -9.67343
+▁occasions -9.67373
+▁gun -9.67597
+▁sing -9.67652
+▁granted -9.67687
+▁window -9.67691
+▁linked -9.67715
+ella -9.67773
+▁resources -9.67813
+wa -9.67903
+▁char -9.67906
+▁bus -9.67943
+▁broadcast -9.67952
+lio -9.68283
+▁increase -9.68379
+▁whom -9.68408
+▁ana -9.68438
+▁add -9.68523
+▁saying -9.68527
+▁characteristics -9.68696
+▁citizen -9.68804
+▁accompani -9.68816
+▁contemporar -9.68816
+▁experiment -9.68816
+▁subtropical -9.68816
+▁portrait -9.68816
+▁troops -9.68817
+▁soap -9.6882
+▁acid -9.68835
+ography -9.68837
+▁anyone -9.68874
+▁weeks -9.68876
+▁systems -9.69048
+▁parish -9.69053
+▁arrested -9.69056
+▁longer -9.69057
+▁sex -9.69098
+▁vo -9.69173
+▁col -9.69223
+▁marine -9.69228
+▁speak -9.6929
+▁parties -9.69464
+▁feature -9.6951
+▁mc -9.69561
+dia -9.69583
+▁recover -9.69692
+▁dot -9.69724
+▁promote -9.7009
+▁particular -9.70165
+▁plane -9.70253
+▁glass -9.70303
+▁feeling -9.70496
+▁century -9.7051
+▁inspired -9.7051
+▁miracle -9.7051
+▁restaurant -9.7051
+▁september -9.7051
+▁exchange -9.7051
+▁hidden -9.7051
+▁humid -9.70511
+▁pretty -9.70511
+▁root -9.70537
+▁pilot -9.70538
+▁flight -9.70549
+▁paper -9.70569
+▁pair -9.70623
+▁reform -9.70625
+ura -9.70632
+▁compar -9.7073
+▁chain -9.70779
+▁victory -9.70842
+▁seasons -9.7086
+▁peter -9.71088
+▁mind -9.71102
+▁shin -9.71157
+▁expected -9.71198
+▁theory -9.71251
+▁workers -9.71286
+▁pictures -9.71288
+▁republican -9.71315
+▁gall -9.71349
+▁obtain -9.71377
+▁establish -9.7138
+▁fruits -9.71416
+▁believed -9.71449
+▁nave -9.71527
+▁kar -9.7172
+▁bla -9.71742
+▁kids -9.71758
+▁describe -9.71801
+▁normal -9.71891
+▁insects -9.71988
+berg -9.72115
+▁marc -9.7223
+▁orchestra -9.72235
+▁portugal -9.72235
+▁quality -9.72235
+▁composer -9.72235
+▁nun -9.72237
+▁wild -9.72243
+▁intense -9.72247
+▁forget -9.72254
+ño -9.72321
+▁wish -9.72343
+▁gen -9.72345
+▁safe -9.72353
+▁ever -9.72408
+bar -9.72464
+▁carry -9.72571
+▁magazines -9.72698
+▁july -9.72704
+▁leg -9.7275
+▁ko -9.72834
+▁effect -9.72963
+▁bag -9.72965
+▁eye -9.73054
+▁mill -9.73069
+▁salt -9.7308
+▁disappear -9.73129
+▁bru -9.73269
+▁mix -9.73326
+ari -9.73383
+ov -9.73469
+▁sell -9.73528
+▁sch -9.73609
+▁perfect -9.73642
+▁app -9.73689
+▁poly -9.73729
+▁lyrics -9.7376
+▁fat -9.73789
+▁sides -9.73797
+▁waiting -9.73901
+▁attempt -9.73989
+▁grandfather -9.73989
+▁religion -9.73989
+▁vehicle -9.73989
+▁alright -9.73989
+▁livestock -9.73989
+▁importance -9.73993
+▁stem -9.73996
+▁james -9.74024
+▁wake -9.74038
+▁occupied -9.74093
+▁grave -9.74123
+▁cry -9.74363
+▁formation -9.74418
+▁lands -9.74556
+▁scar -9.74641
+▁topic -9.74692
+▁drawing -9.74695
+rio -9.74739
+▁plate -9.74818
+▁asking -9.74874
+▁murder -9.74913
+▁grand -9.74916
+▁catch -9.74951
+ze -9.75027
+rce -9.75051
+▁authors -9.75057
+▁hour -9.75059
+▁ec -9.75145
+▁novels -9.75309
+▁sur -9.75357
+▁rosa -9.75558
+▁elections -9.7561
+▁surviv -9.75624
+▁apart -9.75744
+▁alternative -9.75775
+▁consequence -9.75775
+▁effort -9.75775
+▁monastery -9.75775
+▁security -9.75775
+▁custom -9.75775
+▁louis -9.75775
+▁ramiro -9.75775
+▁volume -9.75775
+▁combat -9.75776
+▁naked -9.75778
+▁forgive -9.7578
+▁further -9.75782
+▁temperature -9.75782
+▁catalan -9.75784
+▁martin -9.75793
+▁prefer -9.75795
+▁multi -9.75796
+▁fiction -9.75798
+▁hydro -9.75837
+▁engineer -9.75862
+▁sources -9.759
+▁thin -9.75917
+▁defined -9.75959
+lli -9.75966
+▁representation -9.75976
+▁fear -9.76004
+▁painter -9.76058
+▁scale -9.76133
+▁shit -9.76233
+▁extended -9.76361
+▁widely -9.7645
+▁gri -9.7654
+▁sca -9.7656
+▁earn -9.76594
+▁scared -9.76653
+▁scene -9.76743
+▁piece -9.76755
+▁flying -9.76762
+▁effects -9.76784
+▁issues -9.76969
+▁mac -9.77017
+graph -9.77451
+▁figures -9.77507
+▁argentine -9.77593
+▁diocese -9.77593
+▁rescue -9.77593
+▁website -9.77593
+▁itself -9.77595
+▁narrow -9.77597
+▁internal -9.776
+▁friendship -9.77638
+▁programs -9.77679
+▁rob -9.77722
+▁plain -9.77736
+▁subsequently -9.77787
+▁presents -9.77853
+▁recently -9.77889
+▁page -9.7793
+▁adopted -9.77951
+▁thinking -9.78076
+▁else -9.78088
+▁spa -9.78147
+▁bern -9.78343
+▁file -9.78449
+▁covers -9.78549
+▁launched -9.78606
+▁ham -9.78624
+▁q -9.78817
+q -9.78972
+ick -9.79101
+▁materials -9.79184
+zi -9.7922
+▁gene -9.79294
+▁scor -9.79368
+cular -9.79436
+▁application -9.79445
+▁decorated -9.79445
+▁garcía -9.79445
+▁knew -9.79445
+▁interesting -9.79446
+▁inhabitants -9.79446
+▁self -9.79447
+▁offers -9.79474
+▁eventual -9.7948
+▁feet -9.79538
+▁retired -9.79555
+▁epi -9.79692
+▁collaborated -9.79809
+▁realiz -9.79999
+▁guy -9.80063
+▁score -9.80152
+▁han -9.80201
+▁reading -9.80282
+udi -9.80445
+▁connected -9.80446
+▁solo -9.80507
+▁document -9.80587
+▁telling -9.808
+▁visited -9.81082
+▁express -9.81179
+▁coat -9.8119
+▁code -9.81313
+▁politician -9.81332
+▁veracruz -9.81332
+▁victim -9.81332
+▁fifty -9.81332
+▁herself -9.81332
+▁crime -9.81332
+▁swimming -9.81332
+▁pacific -9.81335
+▁sculpture -9.81337
+▁giving -9.81352
+▁hotel -9.81405
+▁franco -9.81408
+▁mount -9.81411
+▁check -9.81417
+▁hello -9.81438
+▁share -9.81445
+▁esp -9.81481
+▁exile -9.81496
+cion -9.81561
+▁max -9.81579
+▁executed -9.81604
+▁sam -9.81812
+▁teaching -9.81825
+▁amount -9.81836
+▁drop -9.82048
+▁protected -9.82052
+▁sales -9.82201
+▁projects -9.82248
+▁cell -9.82304
+illo -9.82348
+▁ideal -9.82372
+▁spe -9.82373
+▁corn -9.82401
+ada -9.82434
+▁gate -9.8271
+▁damage -9.82833
+▁african -9.82927
+▁universi -9.83222
+▁afraid -9.83255
+▁outstanding -9.83255
+▁provincial -9.83255
+▁nominated -9.83255
+▁shadow -9.83255
+▁legend -9.83255
+▁puerto -9.83255
+▁blow -9.83255
+▁device -9.83255
+▁okay -9.83256
+▁snow -9.83256
+▁impact -9.83277
+▁journalism -9.83278
+▁criticism -9.83286
+port -9.83333
+▁nice -9.83392
+▁lock -9.83469
+▁hang -9.83471
+▁cave -9.836
+▁bat -9.83659
+▁slowly -9.83709
+▁failed -9.83752
+▁publish -9.83778
+▁runs -9.8398
+▁shown -9.84218
+▁articles -9.84286
+▁height -9.84287
+▁historic -9.84305
+▁roads -9.84306
+▁defeat -9.84357
+▁suffer -9.84361
+▁lose -9.84761
+▁magic -9.84784
+▁achieve -9.84973
+▁cam -9.85022
+▁climb -9.85215
+▁digital -9.85215
+▁evidence -9.85215
+▁excellent -9.85215
+▁renaissance -9.85215
+▁quarter -9.85215
+▁wedding -9.85216
+▁shut -9.85217
+▁panama -9.85219
+▁correct -9.85228
+▁assistant -9.85229
+▁northeast -9.85235
+▁dangerous -9.8525
+▁quit -9.85303
+▁factor -9.85351
+▁authority -9.85364
+▁photograph -9.8537
+▁distinguished -9.85413
+▁assembly -9.85417
+▁row -9.85649
+pped -9.85692
+▁pot -9.85752
+▁painted -9.85786
+▁colored -9.85787
+▁tend -9.85827
+▁corner -9.85862
+vision -9.85869
+▁begins -9.85986
+uro -9.86015
+▁loved -9.8613
+ach -9.86368
+▁gone -9.86477
+▁mention -9.86485
+ility -9.86488
+▁rio -9.86543
+▁separate -9.8678
+stan -9.86796
+▁operation -9.86829
+field -9.86908
+▁advantage -9.87215
+▁explain -9.87215
+▁principle -9.87215
+▁wound -9.87215
+▁injured -9.87215
+▁choose -9.87216
+▁plastic -9.87218
+▁railway -9.87219
+▁loud -9.87224
+▁desert -9.87224
+▁tired -9.87225
+▁settlement -9.87241
+▁conduct -9.87242
+▁diego -9.8731
+▁transferr -9.87333
+▁guest -9.87345
+▁neck -9.87362
+▁meaning -9.87385
+▁acquired -9.87405
+▁eggs -9.87431
+▁shore -9.87494
+▁manager -9.8757
+phe -9.87594
+▁bow -9.8764
+▁regularly -9.87661
+▁tries -9.87765
+ville -9.8839
+▁maintain -9.8839
+▁hole -9.88562
+▁n -9.88768
+▁cuba -9.8899
+▁concerts -9.89058
+▁hug -9.89077
+▁ray -9.89087
+▁models -9.89117
+▁poems -9.89186
+ford -9.89222
+▁interview -9.89256
+▁rodriguez -9.89256
+▁violent -9.89256
+▁april -9.89256
+▁vocal -9.89257
+▁nation -9.89262
+▁resistance -9.89266
+▁silence -9.8927
+▁proposed -9.89378
+▁expression -9.89421
+▁trained -9.89628
+▁separated -9.89703
+▁greatest -9.89751
+▁cook -9.89778
+lie -9.89822
+▁smiled -9.89965
+▁survive -9.9009
+▁ze -9.90219
+win -9.90271
+▁walking -9.90422
+emi -9.90488
+▁courses -9.90636
+▁adults -9.90745
+▁fund -9.90772
+▁bone -9.90868
+▁occasion -9.91037
+ku -9.91042
+▁headquarter -9.91067
+water -9.91099
+▁branch -9.91133
+▁sit -9.91196
+▁responsib -9.91332
+▁manufacture -9.91339
+▁deputy -9.9134
+▁crazy -9.9134
+▁reddish -9.9134
+▁teeth -9.91341
+▁daily -9.91342
+▁status -9.91342
+▁sonata -9.91343
+▁executive -9.91355
+▁emerg -9.91356
+▁task -9.91357
+▁talent -9.91389
+▁bought -9.91394
+▁stick -9.91396
+▁interpretation -9.91401
+▁differences -9.91433
+▁trust -9.91479
+ight -9.91517
+▁herb -9.91566
+▁victor -9.92038
+▁heat -9.92045
+▁guys -9.92133
+▁clearly -9.92178
+▁exhibitions -9.92366
+▁meat -9.92444
+▁dies -9.92511
+audi -9.92602
+▁expect -9.92609
+▁driv -9.92678
+▁survived -9.92784
+▁moral -9.92919
+▁gua -9.93044
+▁tru -9.93064
+lot -9.93068
+▁recent -9.93275
+▁generat -9.93341
+ivity -9.93456
+á -9.93467
+▁ceremony -9.93467
+▁comfort -9.93467
+▁constitute -9.93467
+▁easily -9.93467
+▁february -9.93467
+▁ghost -9.93467
+▁solution -9.93467
+▁swear -9.93467
+▁terminal -9.93468
+▁superior -9.93468
+▁afternoon -9.93468
+▁jorge -9.93468
+▁vegetation -9.93471
+▁jesus -9.93473
+▁straight -9.93478
+▁spoke -9.93478
+▁crossing -9.93482
+▁demand -9.93482
+▁ruins -9.93483
+▁brick -9.93491
+▁filled -9.93492
+▁sho -9.93493
+▁wheel -9.93498
+▁conditions -9.9352
+▁justice -9.9353
+▁flora -9.93532
+▁rebel -9.93548
+▁majority -9.93582
+▁investigation -9.9359
+▁demo -9.93648
+▁staff -9.93656
+▁tele -9.9367
+▁reality -9.93672
+▁settled -9.93689
+▁youngest -9.93758
+▁hor -9.93771
+▁particularly -9.93904
+▁cab -9.94037
+▁socialist -9.94062
+▁sto -9.94229
+▁practical -9.94241
+▁roll -9.94258
+▁larges -9.94364
+▁grown -9.94409
+ake -9.94544
+ai -9.9472
+ory -9.94766
+▁yo -9.94866
+aya -9.94995
+▁formal -9.95062
+agua -9.95298
+▁reflect -9.95494
+▁difference -9.95547
+▁paraguay -9.95568
+▁aim -9.95605
+▁soldiers -9.95637
+ologist -9.95641
+▁dimension -9.95641
+▁extension -9.95641
+▁length -9.95641
+▁yourself -9.95641
+▁fifteen -9.95641
+▁cruz -9.95641
+▁electronic -9.95641
+▁universities -9.95641
+▁dinner -9.95642
+▁comprise -9.95647
+▁winning -9.95648
+▁basque -9.95648
+▁opposite -9.95649
+▁broad -9.95673
+▁richard -9.95681
+▁standard -9.95683
+▁load -9.95718
+▁anyway -9.95749
+▁auto -9.9578
+▁web -9.95895
+▁weight -9.9595
+▁presentation -9.96038
+▁grass -9.96068
+ex -9.96089
+▁trial -9.96142
+▁diplomat -9.96325
+▁commander -9.96354
+istic -9.96381
+off -9.96414
+▁removed -9.96531
+▁scenes -9.9667
+▁defend -9.96671
+▁actions -9.96796
+▁severe -9.96919
+▁lee -9.97161
+▁bare -9.97201
+▁shoot -9.97275
+▁indicate -9.97425
+▁laugh -9.97457
+▁studying -9.97783
+▁progress -9.97819
+▁spirit -9.97856
+▁behavior -9.97863
+▁emperor -9.97863
+▁expedition -9.97863
+▁refuge -9.97863
+▁nucle -9.97864
+▁poison -9.97864
+▁baroque -9.97864
+▁sierra -9.97864
+▁cattle -9.97864
+▁historian -9.97865
+▁internet -9.97865
+▁evil -9.97865
+▁daniel -9.97865
+▁failure -9.97873
+▁residence -9.97881
+▁soldier -9.97885
+▁isabel -9.97889
+▁authorities -9.9791
+▁protection -9.97939
+▁crop -9.97971
+▁comment -9.97971
+▁growth -9.97979
+▁causes -9.98006
+▁frame -9.981
+onic -9.98139
+▁june -9.98205
+▁syn -9.9823
+▁captured -9.98232
+▁fixed -9.98303
+▁normally -9.98312
+▁attacked -9.98463
+▁prior -9.98474
+hen -9.98777
+▁spend -9.98931
+▁protect -9.9908
+▁op -9.99142
+iz -9.99225
+▁prepar -9.99253
+▁tall -9.99272
+▁teach -9.99341
+▁philosoph -9.99723
+▁serious -9.99789
+▁frequent -9.99792
+ient -9.99813
+▁pin -9.99863
+▁govern -10.0003
+▁background -10.0014
+▁crowd -10.0014
+▁equivalent -10.0014
+▁extensive -10.0014
+▁irregular -10.0014
+▁negative -10.0014
+▁nephew -10.0014
+▁november -10.0014
+▁remarkable -10.0014
+▁analysis -10.0014
+▁myself -10.0014
+▁mysterious -10.0014
+▁texas -10.0014
+▁attention -10.0014
+▁catalonia -10.0014
+▁varied -10.0014
+▁premier -10.0014
+▁estela -10.0014
+▁visual -10.0014
+▁berlin -10.0014
+▁attach -10.0014
+▁tonight -10.0014
+▁hurt -10.0015
+▁rosario -10.0015
+▁option -10.0016
+▁bullet -10.0019
+▁echo -10.0022
+▁labor -10.0024
+▁label -10.0024
+▁shop -10.0027
+▁passion -10.004
+▁sailor -10.0045
+ably -10.0047
+▁rarely -10.0048
+▁edited -10.0055
+▁factory -10.0064
+▁operations -10.0068
+▁stable -10.0073
+▁slave -10.0073
+▁camera -10.0109
+▁disco -10.0125
+▁recovered -10.0131
+▁practiced -10.0132
+▁observ -10.016
+▁pine -10.0164
+▁holding -10.0177
+cut -10.0195
+▁bio -10.0196
+ati -10.0197
+▁needed -10.0198
+▁mess -10.0204
+▁qua -10.0214
+▁comic -10.0227
+ira -10.0237
+▁hunt -10.024
+graphic -10.0243
+▁danger -10.0244
+▁descend -10.0244
+▁protest -10.0244
+back -10.0246
+▁december -10.0246
+▁judicial -10.0246
+▁perennial -10.0246
+▁purple -10.0246
+▁vietnam -10.0246
+▁zaragoza -10.0246
+▁expand -10.0246
+▁expert -10.0246
+▁fourteen -10.0246
+▁translation -10.0246
+▁goodbye -10.0246
+▁vicente -10.0246
+▁enrique -10.0246
+▁micro -10.0246
+▁annual -10.0246
+▁patient -10.0247
+▁anchor -10.0247
+▁latter -10.0247
+▁wealth -10.0247
+▁steve -10.0247
+▁sexual -10.0247
+▁altar -10.0248
+▁destination -10.0249
+▁colo -10.0249
+▁debate -10.025
+▁whatever -10.0266
+▁element -10.0267
+▁concert -10.0269
+▁rare -10.027
+ney -10.0274
+▁election -10.0275
+▁adapted -10.028
+▁signal -10.0305
+▁increased -10.0308
+▁classes -10.0308
+clan -10.0312
+▁owned -10.0317
+▁psycho -10.0319
+▁starting -10.0326
+▁speaking -10.0329
+▁net -10.0332
+▁rev -10.0338
+▁farming -10.0348
+▁missing -10.0354
+▁habit -10.0357
+▁victoria -10.0359
+▁reported -10.04
+▁sentence -10.0444
+light -10.0462
+▁pat -10.0473
+▁celebrate -10.0474
+▁martín -10.0483
+▁asleep -10.0484
+▁athletic -10.0484
+▁eduardo -10.0484
+▁enemy -10.0484
+▁external -10.0484
+▁diverse -10.0484
+▁chamber -10.0484
+▁senator -10.0484
+▁fierce -10.0484
+▁interested -10.0484
+▁rival -10.0484
+circular -10.0485
+▁statue -10.0485
+ice -10.0485
+▁worry -10.0492
+▁jack -10.05
+▁convention -10.0505
+▁servants -10.051
+▁succeeded -10.051
+▁jr -10.0523
+▁mau -10.0535
+▁cuban -10.0543
+▁equal -10.0547
+▁gro -10.0549
+▁mid -10.0566
+▁whispered -10.0567
+▁documents -10.0575
+▁rat -10.0581
+▁et -10.0602
+▁letters -10.0614
+▁illustrat -10.0632
+▁ox -10.0633
+▁photo -10.0658
+▁mono -10.0661
+called -10.0674
+wood -10.0697
+mbo -10.071
+▁traveled -10.0713
+▁toys -10.072
+▁display -10.0728
+▁gonzalez -10.0728
+▁immigrant -10.0728
+▁inflorescence -10.0728
+▁journey -10.0728
+▁metropolitan -10.0728
+▁supposed -10.0728
+▁technology -10.0728
+▁thailand -10.0728
+▁component -10.0728
+▁license -10.0728
+▁opinion -10.0728
+▁botanic -10.0728
+▁humble -10.0728
+▁radical -10.0728
+▁everyday -10.0728
+▁classification -10.073
+▁basin -10.0731
+▁collective -10.0731
+▁decoration -10.0731
+▁comedy -10.0732
+▁celebration -10.0733
+▁documentary -10.0733
+▁exact -10.0734
+bert -10.0738
+▁grace -10.074
+▁connection -10.0754
+mont -10.0756
+▁glasses -10.0758
+ache -10.0762
+▁smaller -10.0762
+▁colombian -10.0763
+▁fr -10.0784
+▁anger -10.0792
+▁repeated -10.0792
+▁chart -10.0792
+▁celebrated -10.0798
+▁publishing -10.0802
+▁gray -10.0823
+▁broadcasted -10.0856
+▁officially -10.0869
+▁esc -10.0872
+▁standing -10.0875
+▁dar -10.0878
+▁pull -10.0878
+▁hearing -10.0889
+▁whisper -10.0892
+▁link -10.0892
+▁fingers -10.0923
+well -10.0927
+▁rent -10.0927
+▁promise -10.0934
+▁surprise -10.0934
+▁remembered -10.0941
+▁zi -10.0958
+▁medic -10.0965
+ense -10.0971
+▁bachelor -10.0978
+▁demonstrat -10.0978
+▁fulfill -10.0978
+▁incident -10.0978
+▁lieutenant -10.0978
+▁territories -10.0978
+▁essay -10.0978
+▁blind -10.0978
+▁cheese -10.0978
+▁charles -10.0983
+▁reaction -10.0986
+ette -10.0992
+▁zero -10.0993
+▁bushes -10.0998
+▁ken -10.0999
+▁theatre -10.1003
+▁brain -10.1006
+▁beard -10.1008
+▁adam -10.1013
+▁defender -10.1015
+ball -10.1024
+▁laser -10.1027
+▁confirmed -10.1032
+▁juli -10.1041
+▁lima -10.1046
+▁claimed -10.1059
+gne -10.107
+▁simpl -10.1089
+▁cable -10.111
+▁prime -10.1145
+▁rib -10.1146
+▁sar -10.1153
+▁passengers -10.1165
+▁vol -10.1169
+▁ram -10.1172
+▁egypt -10.123
+▁emigrat -10.1235
+▁kilometer -10.1235
+▁physics -10.1235
+▁principal -10.1235
+▁recognition -10.1235
+▁rectangular -10.1235
+▁rhythm -10.1235
+▁thomas -10.1235
+▁tomorrow -10.1235
+▁turkey -10.1235
+▁imperial -10.1235
+▁suprem -10.1235
+▁parallel -10.1235
+▁shield -10.1235
+▁deserve -10.1235
+▁workshop -10.1235
+▁freedom -10.1235
+▁dispute -10.1235
+▁admir -10.1235
+▁uncle -10.1235
+▁criticized -10.1236
+▁descendant -10.1237
+▁finance -10.1237
+▁pressure -10.1238
+▁semifinal -10.1238
+▁funny -10.1238
+▁dawn -10.1238
+▁twin -10.124
+▁understood -10.1244
+▁brazilian -10.1247
+▁prisoners -10.1248
+▁critic -10.1249
+▁striker -10.1251
+▁price -10.1256
+▁graphic -10.1269
+anda -10.1274
+▁roberto -10.1284
+▁shooting -10.1304
+▁sab -10.1307
+▁adult -10.1309
+▁required -10.1309
+▁review -10.131
+▁spite -10.1315
+▁pala -10.1328
+▁sum -10.1333
+▁moor -10.1334
+aria -10.1349
+anza -10.1357
+▁map -10.1366
+mann -10.1377
+▁notice -10.1378
+▁extend -10.1412
+▁abandon -10.1412
+▁sang -10.142
+▁grant -10.1433
+▁sever -10.1439
+▁deposit -10.1439
+elia -10.1454
+▁capture -10.1457
+ability -10.1464
+▁vegetables -10.1464
+▁usual -10.1474
+j -10.1493
+▁inherit -10.1496
+▁abundant -10.1498
+▁circle -10.1498
+▁exercise -10.1498
+▁florida -10.1498
+▁holiday -10.1498
+▁january -10.1498
+▁nickname -10.1498
+▁peninsula -10.1498
+▁switzerland -10.1498
+▁eleven -10.1498
+▁federation -10.1498
+▁peruvian -10.1498
+▁concern -10.1498
+▁resign -10.1498
+▁identity -10.1498
+▁drown -10.1498
+▁applied -10.1498
+▁ahead -10.1498
+▁sample -10.1498
+▁smoke -10.1498
+▁solve -10.1499
+▁boston -10.1499
+▁frank -10.1499
+▁collaboration -10.1499
+▁pool -10.1501
+▁loose -10.1501
+▁sugar -10.1501
+▁vega -10.1501
+▁nearby -10.1501
+▁plaza -10.1502
+▁seminar -10.1504
+▁percent -10.1504
+▁effective -10.1504
+▁duke -10.1507
+iness -10.1509
+▁crew -10.151
+▁alberto -10.1519
+▁shoe -10.1527
+itu -10.1541
+▁sixth -10.1542
+▁elect -10.1543
+▁surprised -10.1544
+▁critical -10.1558
+▁butler -10.1563
+ago -10.1583
+▁dancer -10.1591
+▁affairs -10.1593
+▁bark -10.1598
+▁prepared -10.1598
+bor -10.1613
+bel -10.1613
+▁alter -10.1614
+▁philippines -10.1631
+▁dam -10.1641
+▁tied -10.1644
+▁inhabited -10.1653
+pen -10.1677
+▁claim -10.1682
+▁affect -10.1683
+▁nin -10.1693
+▁hip -10.1698
+▁korea -10.172
+oz -10.1742
+▁investigat -10.1752
+▁apparent -10.1768
+▁coordinat -10.1768
+▁dominican -10.1768
+▁entertainment -10.1768
+▁familiar -10.1768
+▁proposal -10.1768
+▁siblings -10.1768
+▁smooth -10.1768
+▁vault -10.1768
+▁relief -10.1768
+▁monarch -10.1768
+▁harvest -10.1768
+▁witness -10.1768
+▁basketball -10.1768
+▁fortress -10.1768
+▁traffic -10.1768
+▁address -10.1768
+▁reign -10.1768
+▁steel -10.1768
+▁gothic -10.1768
+▁angry -10.1769
+▁crist -10.1769
+▁flock -10.177
+▁bush -10.1771
+▁choir -10.1773
+▁apartment -10.1776
+▁shi -10.1793
+▁deliver -10.1796
+▁exception -10.1798
+▁ability -10.1802
+▁registered -10.1803
+▁spin -10.181
+▁entit -10.1811
+grad -10.1839
+dent -10.185
+▁sala -10.185
+▁originate -10.1851
+▁tar -10.1858
+▁engineering -10.1859
+▁coin -10.1898
+ung -10.191
+ille -10.1927
+▁closer -10.1933
+▁rub -10.1943
+▁sco -10.1955
+▁jud -10.1955
+lice -10.1961
+▁involve -10.2002
+▁chose -10.2008
+▁graduate -10.2013
+▁quick -10.2019
+▁collaborate -10.2032
+▁precise -10.2046
+▁accused -10.2046
+▁ambassador -10.2046
+▁autonomous -10.2046
+▁bloom -10.2046
+▁capacity -10.2046
+▁context -10.2046
+▁córdoba -10.2046
+▁evolution -10.2046
+▁fundamental -10.2046
+▁merchant -10.2046
+▁phenomenon -10.2046
+▁potential -10.2046
+▁specimen -10.2046
+▁sponsor -10.2046
+▁substitute -10.2046
+▁washington -10.2046
+▁yesterday -10.2046
+▁bronze -10.2046
+▁confess -10.2046
+▁indonesia -10.2046
+▁granada -10.2046
+▁greece -10.2046
+▁mechanic -10.2046
+▁mobile -10.2046
+▁beauty -10.2046
+▁potato -10.2046
+▁navy -10.2046
+▁varies -10.2046
+▁shirt -10.2046
+▁agency -10.2046
+▁gallery -10.2046
+▁partially -10.2046
+▁ocean -10.2046
+▁japanese -10.2046
+▁andalusia -10.2046
+▁barrel -10.2046
+▁gym -10.2046
+▁alternate -10.2047
+▁confrontation -10.2049
+▁tank -10.205
+▁israel -10.205
+▁bomb -10.205
+▁peña -10.2051
+▁henry -10.2051
+▁unfortunately -10.2052
+▁plot -10.2054
+▁silk -10.2058
+▁scent -10.2064
+acion -10.2071
+▁combined -10.2074
+atory -10.2077
+▁fleet -10.208
+▁congo -10.2083
+ctually -10.2084
+▁refused -10.2101
+▁photography -10.2105
+▁agree -10.2114
+▁ranch -10.2143
+▁pure -10.2148
+▁beak -10.2165
+▁poll -10.2171
+eval -10.2176
+▁ple -10.2181
+▁bass -10.2222
+▁joy -10.2224
+▁rocky -10.2239
+rich -10.2264
+▁stated -10.2279
+▁bite -10.2307
+pan -10.2327
+▁artificial -10.2331
+▁flood -10.2331
+▁hungry -10.2331
+▁mechanism -10.2331
+▁murmur -10.2331
+▁properties -10.2331
+▁spaniard -10.2331
+▁underground -10.2331
+▁variable -10.2331
+▁chicago -10.2331
+▁feather -10.2331
+▁felipe -10.2331
+▁pablo -10.2332
+▁cameroon -10.2332
+▁colonel -10.2332
+▁denmark -10.2332
+▁cloud -10.2332
+▁spider -10.2332
+▁shaking -10.2332
+▁copies -10.2332
+▁cousin -10.2332
+▁expos -10.2332
+▁ji -10.2333
+▁occasionally -10.2334
+▁municipalities -10.2335
+▁approximately -10.2336
+▁wow -10.2337
+▁basis -10.2339
+▁powerful -10.2341
+▁yellowish -10.2342
+▁scarce -10.2343
+▁chair -10.2344
+▁pad -10.2352
+▁identified -10.2353
+▁popularity -10.2358
+east -10.2365
+▁charm -10.237
+▁vest -10.237
+▁damn -10.2389
+▁trace -10.2396
+▁shared -10.2399
+▁fur -10.2417
+▁broken -10.2424
+bound -10.2457
+▁article -10.2469
+otic -10.2481
+hood -10.2489
+▁churches -10.2526
+▁letter -10.2548
+▁profession -10.2577
+▁counter -10.2581
+▁possibl -10.2589
+▁success -10.2607
+ñ -10.2611
+▁interpret -10.2617
+▁gain -10.2621
+▁francis -10.2622
+▁integra -10.2622
+▁oppos -10.2625
+▁predominant -10.2626
+▁attitude -10.2626
+▁categories -10.2626
+▁coffee -10.2626
+▁criminal -10.2626
+▁montevideo -10.2626
+▁moscow -10.2626
+▁sweden -10.2626
+▁uncertain -10.2626
+▁violence -10.2626
+▁nacional -10.2626
+▁prevail -10.2626
+▁junior -10.2626
+▁intended -10.2626
+▁flash -10.2626
+▁folk -10.2626
+▁clothes -10.2626
+▁intention -10.2626
+▁firm -10.2626
+▁patriot -10.2626
+▁sharp -10.2626
+▁invited -10.2627
+▁vacation -10.2627
+▁canon -10.2627
+▁tears -10.2628
+▁desire -10.2629
+▁message -10.2629
+▁noise -10.2629
+▁existence -10.263
+▁travelled -10.263
+▁progressive -10.2632
+▁glory -10.2633
+▁tribe -10.2633
+▁rough -10.2634
+▁commune -10.2635
+▁yu -10.2644
+mark -10.2646
+▁paid -10.2648
+▁peak -10.2657
+▁shall -10.2657
+▁tang -10.2673
+▁perfectly -10.2684
+luc -10.2716
+▁entirely -10.2716
+stein -10.2723
+▁drinking -10.2755
+▁exhibition -10.2785
+▁farmer -10.2793
+▁horn -10.2838
+▁lover -10.2862
+▁swa -10.2863
+war -10.2874
+▁austria -10.2881
+known -10.2894
+▁drama -10.2926
+▁challenge -10.2929
+▁circuit -10.2929
+▁continuous -10.2929
+▁decrease -10.2929
+▁facilities -10.2929
+▁girlfriend -10.2929
+▁happiness -10.2929
+▁legislative -10.2929
+▁nicaragua -10.2929
+▁scratch -10.2929
+▁suicide -10.2929
+▁universe -10.2929
+▁visible -10.2929
+▁stupid -10.2929
+▁atlantic -10.2929
+▁possibility -10.2929
+▁caught -10.2929
+▁crystal -10.2929
+▁monkey -10.2929
+▁leaf -10.2929
+▁moderate -10.2929
+▁wrapp -10.2929
+▁shopping -10.2929
+▁fool -10.2929
+▁ireland -10.2929
+▁merced -10.293
+▁shelter -10.293
+▁convert -10.2931
+▁dying -10.2931
+▁educated -10.2931
+▁descent -10.2931
+▁honey -10.2932
+▁jackson -10.2932
+▁initiative -10.2932
+▁passionate -10.2932
+▁specifically -10.2933
+▁resist -10.2934
+▁airline -10.2934
+▁deportivo -10.2935
+▁selection -10.2937
+▁clock -10.2937
+▁sheets -10.2938
+▁iran -10.2944
+▁restored -10.2945
+▁fever -10.2948
+▁seriously -10.2975
+▁via -10.2983
+▁determined -10.2985
+ista -10.2987
+▁threatened -10.2989
+▁hunting -10.3007
+uda -10.3018
+▁larger -10.3055
+▁killer -10.3056
+▁lean -10.307
+▁joint -10.3078
+▁australian -10.3098
+▁picture -10.3098
+▁tape -10.3111
+▁mom -10.3154
+▁killing -10.3177
+▁cata -10.3196
+▁organize -10.32
+▁attract -10.3215
+▁hate -10.3219
+▁mario -10.3229
+▁bradomin -10.3241
+▁committed -10.3241
+▁communities -10.3241
+▁explore -10.3241
+▁gabriel -10.3241
+▁guatemala -10.3241
+▁portugues -10.3241
+▁rafael -10.3241
+▁sacred -10.3241
+▁salamanca -10.3241
+▁volunteer -10.3241
+▁archeological -10.3241
+▁erect -10.3241
+▁worldwide -10.3241
+▁canadian -10.3241
+▁profile -10.3241
+▁knight -10.3241
+▁drunk -10.3241
+▁drug -10.3241
+▁margin -10.3242
+▁torture -10.3242
+▁forgotten -10.3242
+▁solid -10.3242
+▁orbit -10.3242
+▁dominate -10.3243
+▁sym -10.3245
+▁electora -10.3247
+▁milan -10.3251
+▁grape -10.3251
+▁stream -10.3252
+▁preparation -10.3253
+▁praised -10.3254
+room -10.3255
+▁tax -10.3257
+▁announced -10.3261
+fici -10.3264
+▁oak -10.3268
+▁marquise -10.3269
+licate -10.3274
+▁constitutional -10.3276
+▁shade -10.3278
+▁assist -10.3288
+▁promised -10.3296
+▁agent -10.3302
+▁highlighted -10.3304
+▁revealed -10.3317
+▁geographical -10.3322
+▁georgia -10.3327
+▁significantly -10.3334
+▁bath -10.3366
+▁maintained -10.3376
+mine -10.3378
+▁sandy -10.3387
+▁invad -10.3388
+▁growing -10.3415
+▁valu -10.3435
+▁begin -10.344
+▁pav -10.345
+▁learning -10.346
+▁ross -10.3475
+▁destroy -10.3478
+▁divid -10.3482
+▁designer -10.3486
+▁devil -10.3497
+▁paint -10.3515
+▁raise -10.3521
+▁shoulder -10.3525
+▁subsequent -10.3529
+writer -10.3533
+▁strike -10.3543
+▁casa -10.3547
+▁increasing -10.3564
+▁administer -10.3564
+▁alexander -10.3564
+▁anybody -10.3564
+▁bulgaria -10.3564
+▁combination -10.3564
+▁commerce -10.3564
+▁hondura -10.3564
+▁imprisoned -10.3564
+▁indigenous -10.3564
+▁interrupt -10.3564
+▁mirror -10.3564
+▁nicolas -10.3564
+▁temporary -10.3564
+▁thirteen -10.3564
+▁trembling -10.3564
+▁athlete -10.3564
+▁reputation -10.3564
+▁zealand -10.3564
+▁chemical -10.3564
+▁fossil -10.3564
+▁dismiss -10.3564
+▁schedule -10.3564
+▁pioneer -10.3564
+▁strateg -10.3564
+▁cycle -10.3564
+▁split -10.3564
+▁occurred -10.3564
+▁naval -10.3564
+▁martínez -10.3565
+▁engag -10.3566
+▁swamp -10.3567
+▁execution -10.3567
+▁adaptation -10.3569
+▁saving -10.3569
+▁arise -10.3575
+▁eliminated -10.3577
+▁kick -10.3579
+▁zu -10.358
+▁colonial -10.3581
+▁alex -10.3583
+▁operated -10.3583
+▁floating -10.3588
+▁volcano -10.3592
+▁rice -10.3599
+▁bond -10.3599
+▁otherwise -10.3605
+▁paz -10.3624
+▁indicated -10.3628
+ius -10.3629
+▁referr -10.3631
+▁beaches -10.3649
+▁ger -10.3652
+▁observed -10.3685
+▁whe -10.3743
+▁wan -10.3751
+▁listening -10.3756
+▁repeat -10.3811
+logy -10.3811
+▁fail -10.3824
+▁stepp -10.3837
+▁formula -10.3846
+▁constitution -10.386
+▁arc -10.3883
+▁narrat -10.3893
+▁cooper -10.3896
+▁swim -10.3897
+▁assigned -10.3897
+▁carolina -10.3897
+▁compound -10.3897
+▁definition -10.3897
+▁destruction -10.3897
+▁dissolv -10.3897
+▁encourage -10.3897
+▁fountain -10.3897
+▁manuscript -10.3897
+▁phrase -10.3897
+▁recipe -10.3897
+▁regulation -10.3897
+▁residential -10.3897
+▁saturday -10.3897
+▁scientist -10.3897
+▁somebody -10.3897
+▁sword -10.3897
+▁brief -10.3897
+▁convince -10.3897
+▁crisis -10.3897
+▁guzman -10.3897
+▁predator -10.3897
+▁situated -10.3897
+▁ticket -10.3897
+▁communist -10.3897
+▁consumed -10.3897
+▁jazz -10.3897
+▁memories -10.3897
+▁divine -10.3897
+▁repair -10.3897
+▁vienna -10.3897
+▁architectural -10.3897
+▁snake -10.3897
+▁conquer -10.3897
+▁shark -10.3897
+▁isolated -10.3897
+▁throne -10.3897
+▁rumor -10.3897
+▁conservative -10.3897
+▁preferred -10.3897
+▁strict -10.3898
+▁session -10.3898
+▁programming -10.3898
+▁incorporated -10.3898
+▁racing -10.3898
+▁creative -10.39
+▁logic -10.3902
+▁simon -10.3902
+▁exam -10.3904
+▁cancer -10.3905
+▁estate -10.3907
+▁lineage -10.3916
+▁attraction -10.392
+izing -10.3926
+▁cake -10.3929
+▁sheep -10.3931
+▁exhibited -10.3943
+▁selected -10.3955
+▁vision -10.3956
+▁sentenced -10.3959
+▁jun -10.3965
+▁marta -10.3967
+▁damaged -10.3974
+▁preced -10.3992
+▁tie -10.4002
+▁mining -10.4031
+duct -10.404
+▁liber -10.4047
+▁grande -10.4069
+▁bou -10.4088
+ong -10.41
+▁harm -10.41
+▁stripe -10.4134
+▁lion -10.4154
+▁brave -10.4156
+▁reject -10.4158
+▁journal -10.4167
+▁jer -10.4172
+▁verb -10.422
+▁faith -10.424
+▁exhaust -10.4242
+▁advertising -10.4242
+▁ceramic -10.4242
+▁circumstance -10.4242
+▁confused -10.4242
+▁cultivated -10.4242
+▁frigate -10.4242
+▁gradually -10.4242
+▁intellectual -10.4242
+▁proclaim -10.4242
+▁response -10.4242
+▁soundtrack -10.4242
+▁treasure -10.4242
+▁triumph -10.4242
+▁cruel -10.4242
+▁domingo -10.4242
+▁uniform -10.4242
+▁unlike -10.4242
+▁description -10.4242
+▁navigat -10.4242
+▁statistic -10.4242
+▁birthday -10.4242
+▁mythology -10.4242
+▁serbia -10.4242
+▁cliff -10.4242
+▁flour -10.4242
+▁funeral -10.4242
+▁nurse -10.4242
+▁curved -10.4242
+▁gulf -10.4242
+▁arabia -10.4242
+▁lectur -10.4243
+▁oxford -10.4243
+▁bedroom -10.4243
+▁diagram -10.4243
+▁loyal -10.4243
+▁clinic -10.4244
+▁polish -10.4244
+▁contrary -10.4244
+▁speech -10.4244
+▁retirement -10.4244
+haps -10.4245
+▁motiv -10.425
+▁baby -10.425
+▁giant -10.4253
+▁divorced -10.4253
+▁alfredo -10.4254
+▁trick -10.4257
+▁archive -10.4257
+▁teenager -10.426
+▁dishes -10.4262
+▁candle -10.4265
+▁cult -10.4265
+▁assume -10.4269
+▁motor -10.4273
+▁manner -10.4275
+▁advisor -10.4275
+▁vary -10.4278
+▁sighed -10.4294
+▁tomb -10.4304
+arrow -10.4324
+▁apple -10.433
+bbing -10.4331
+▁mist -10.4334
+▁laughing -10.4335
+▁fig -10.4338
+▁deeply -10.4361
+▁wire -10.4361
+▁defended -10.4379
+▁strip -10.4385
+▁hal -10.4391
+▁pont -10.4393
+▁bob -10.44
+▁pig -10.4403
+▁sight -10.4411
+▁murdered -10.4412
+▁reader -10.4422
+▁layer -10.4422
+▁pope -10.4425
+▁wr -10.4475
+▁reveal -10.4513
+▁breed -10.4543
+▁emblem -10.4559
+where -10.457
+▁dig -10.4591
+▁tale -10.4594
+▁abroad -10.4599
+▁absolute -10.4599
+▁aircraft -10.4599
+▁benefit -10.4599
+▁boyfriend -10.4599
+▁centuries -10.4599
+▁committee -10.4599
+▁embassy -10.4599
+▁encounter -10.4599
+▁flavor -10.4599
+▁infrastructure -10.4599
+▁ingredient -10.4599
+▁intelligence -10.4599
+▁maximum -10.4599
+▁pleasure -10.4599
+▁pregnant -10.4599
+▁rainforest -10.4599
+▁taiwan -10.4599
+▁amazon -10.4599
+▁bottle -10.4599
+▁britain -10.4599
+▁croatia -10.4599
+▁amateur -10.4599
+▁pattern -10.4599
+▁smith -10.4599
+▁trouble -10.4599
+▁hispanic -10.4599
+▁turkis -10.4599
+▁injuries -10.4599
+▁rebellion -10.4599
+▁sunday -10.4599
+▁flesh -10.4599
+▁duties -10.4599
+▁chorus -10.4599
+▁mercy -10.4599
+▁craft -10.4599
+▁alliance -10.4599
+▁motto -10.4599
+▁clay -10.46
+▁reef -10.46
+▁choice -10.46
+▁borneo -10.4601
+▁concrete -10.4601
+▁collaborator -10.4602
+▁dramatic -10.4603
+▁instant -10.4603
+▁narrative -10.4604
+▁attractive -10.4604
+▁seville -10.4604
+▁trend -10.4605
+▁safety -10.4606
+▁shock -10.4606
+phone -10.4607
+▁debt -10.4608
+▁simultaneously -10.4609
+▁fuel -10.4612
+▁arrival -10.4621
+▁guitarist -10.4622
+▁touristic -10.4625
+▁motion -10.4627
+▁taxi -10.4628
+▁temperat -10.4636
+▁rico -10.464
+▁arriv -10.4651
+▁complaint -10.4652
+▁recommended -10.4658
+▁sink -10.4661
+▁korean -10.4663
+▁eva -10.4678
+▁arches -10.4679
+▁escaped -10.4703
+▁gained -10.4706
+▁mora -10.4718
+▁cool -10.4727
+▁dish -10.4751
+▁coll -10.4759
+▁compli -10.4777
+▁core -10.4806
+▁handle -10.4831
+pping -10.4871
+phon -10.4894
+eiro -10.4912
+▁continent -10.4949
+▁sara -10.4949
+▁albert -10.4958
+▁initiat -10.4966
+▁employ -10.4969
+▁abadesa -10.4969
+▁aquatic -10.4969
+▁cervantes -10.4969
+▁chocolate -10.4969
+▁colegio -10.4969
+▁diversity -10.4969
+▁grandmother -10.4969
+▁guillermo -10.4969
+▁nervous -10.4969
+▁opportunity -10.4969
+▁pilgrim -10.4969
+▁sacrifice -10.4969
+▁sanctuary -10.4969
+▁scattered -10.4969
+▁alcohol -10.4969
+▁tixul -10.4969
+▁blonde -10.4969
+▁cordoba -10.4969
+▁cowboy -10.4969
+▁variation -10.4969
+▁target -10.4969
+▁shouted -10.4969
+▁segment -10.4969
+▁belief -10.4969
+▁heritage -10.4969
+▁withdraw -10.497
+▁cartoon -10.497
+actic -10.497
+▁costume -10.497
+▁gomez -10.497
+▁false -10.497
+▁javier -10.497
+▁opposition -10.497
+▁invasion -10.497
+▁decline -10.497
+▁average -10.497
+▁tragic -10.497
+▁vertical -10.497
+▁fauna -10.497
+▁instance -10.497
+▁catalog -10.4971
+▁quantit -10.4971
+▁breast -10.4972
+▁resistant -10.4972
+▁everywhere -10.4972
+▁disney -10.4972
+▁malaga -10.4973
+▁wheat -10.4973
+▁guess -10.4976
+▁dense -10.4976
+▁cannon -10.4977
+▁stock -10.4981
+▁detect -10.4987
+▁rocket -10.4988
+▁mike -10.4989
+▁useful -10.4991
+▁spur -10.4998
+▁campus -10.5
+▁romance -10.5006
+▁modest -10.5012
+▁stain -10.5017
+▁austrian -10.5028
+▁breeding -10.5032
+▁collected -10.5054
+▁squ -10.5055
+▁presenter -10.5056
+▁sale -10.5083
+▁scr -10.5084
+▁reproduce -10.5103
+awa -10.514
+▁provid -10.5149
+▁jan -10.5164
+▁supporter -10.5211
+▁collect -10.5236
+▁tube -10.5325
+▁possess -10.5333
+greed -10.5334
+corporate -10.5353
+▁conversation -10.5354
+▁definite -10.5354
+▁faculty -10.5354
+▁imagine -10.5354
+▁knife -10.5354
+▁lópez -10.5354
+▁maneuver -10.5354
+▁orthodox -10.5354
+▁protocol -10.5354
+▁reception -10.5354
+▁respond -10.5354
+▁revenge -10.5354
+▁software -10.5354
+▁supply -10.5354
+▁territorial -10.5354
+▁terror -10.5354
+▁varieties -10.5354
+▁argument -10.5354
+▁hometown -10.5354
+▁rugby -10.5354
+▁capsule -10.5354
+▁courage -10.5354
+▁scandal -10.5354
+▁columbia -10.5354
+▁error -10.5354
+▁fragment -10.5354
+▁retreat -10.5354
+▁saxon -10.5354
+▁symptom -10.5354
+▁videogame -10.5354
+▁wolf -10.5354
+▁wilson -10.5354
+▁protein -10.5354
+▁global -10.5354
+▁relax -10.5354
+▁jaime -10.5354
+▁liquid -10.5354
+▁shift -10.5354
+▁stairs -10.5355
+▁tempt -10.5355
+▁tennis -10.5355
+▁creature -10.5355
+▁homonymous -10.5355
+▁insult -10.5355
+▁infect -10.5356
+▁humor -10.5356
+▁riding -10.5356
+▁smart -10.5356
+▁planned -10.5356
+▁textile -10.5358
+▁improvement -10.5359
+▁stateroom -10.5359
+▁contributions -10.536
+▁finland -10.5361
+▁fox -10.5366
+▁efficient -10.5369
+▁doctorate -10.537
+▁manual -10.5372
+▁font -10.5382
+▁grade -10.5383
+▁logo -10.5385
+▁rath -10.5386
+▁exit -10.5386
+▁carmen -10.5387
+▁pack -10.5391
+▁grain -10.5392
+▁restricted -10.5394
+▁locality -10.5398
+▁bible -10.5405
+▁cala -10.541
+▁worse -10.5439
+▁crosses -10.5495
+▁rate -10.5497
+▁symbolize -10.557
+▁sao -10.5577
+▁chan -10.5606
+▁tim -10.5658
+▁confirm -10.5668
+▁mineral -10.5676
+▁egg -10.5711
+▁servant -10.5712
+▁assembl -10.5712
+▁rode -10.5719
+▁volcan -10.572
+▁grab -10.5726
+▁horizon -10.5741
+▁rig -10.5747
+▁recognize -10.5752
+▁bolivar -10.5754
+▁concentration -10.5754
+▁concepcion -10.5754
+▁consecutive -10.5754
+▁consumption -10.5754
+▁depression -10.5754
+▁dispersed -10.5754
+▁forehead -10.5754
+▁goalkeeper -10.5754
+▁grijalba -10.5754
+▁homosexual -10.5754
+▁mammals -10.5754
+▁muslim -10.5754
+▁opponent -10.5754
+▁participant -10.5754
+▁sacristy -10.5754
+▁subspecies -10.5754
+▁swedish -10.5754
+▁universal -10.5754
+▁estrella -10.5754
+▁guilty -10.5754
+▁joseph -10.5754
+▁substance -10.5754
+▁unusual -10.5754
+▁malaysia -10.5754
+▁leather -10.5754
+▁transformed -10.5754
+▁trophy -10.5754
+▁beyond -10.5754
+▁complement -10.5754
+▁remind -10.5754
+▁siege -10.5754
+▁jungle -10.5754
+▁client -10.5754
+▁lagoon -10.5754
+▁friar -10.5754
+▁flores -10.5754
+▁memoria -10.5754
+▁pretend -10.5754
+▁eighteen -10.5754
+▁domain -10.5754
+▁honest -10.5755
+▁proud -10.5755
+▁edward -10.5755
+▁ladies -10.5755
+▁denied -10.5755
+▁policy -10.5756
+▁hiding -10.5756
+▁castillo -10.5756
+▁keyboard -10.5757
+▁swis -10.5757
+▁toxic -10.5757
+▁colony -10.5758
+▁lift -10.5758
+▁balance -10.5758
+▁heavi -10.5759
+▁equation -10.5759
+▁egyptian -10.5761
+▁mild -10.5766
+▁approved -10.5771
+▁highway -10.5773
+▁determine -10.5778
+▁galician -10.5782
+▁exclusively -10.5786
+▁improved -10.5796
+▁dust -10.5829
+▁presid -10.584
+▁procession -10.5848
+▁paulo -10.5858
+▁mail -10.5891
+▁functional -10.5902
+▁belt -10.5904
+▁joan -10.5917
+▁gru -10.5957
+▁hop -10.5967
+▁introduc -10.6083
+▁struggle -10.6086
+ock -10.6106
+▁america -10.6108
+rino -10.6119
+▁settle -10.612
+▁improve -10.6121
+▁reduce -10.6127
+▁define -10.6128
+eche -10.6131
+▁hyp -10.6141
+▁tap -10.6152
+▁antarctic -10.6171
+▁contemplat -10.6171
+▁disciple -10.6171
+▁distinctive -10.6171
+▁dynasty -10.6171
+▁earthquake -10.6171
+▁equipment -10.6171
+▁ernesto -10.6171
+▁excited -10.6171
+▁extraordinary -10.6171
+▁fernández -10.6171
+▁instruction -10.6171
+▁intervention -10.6171
+▁linguistic -10.6171
+▁migration -10.6171
+▁ornamental -10.6171
+▁penetrat -10.6171
+▁perspective -10.6171
+▁suspect -10.6171
+▁tanzania -10.6171
+▁terrestrial -10.6171
+▁terrible -10.6171
+▁valladolid -10.6171
+▁delicious -10.6171
+▁disorder -10.6171
+▁evolv -10.6171
+▁tribute -10.6171
+▁whitish -10.6171
+▁configur -10.6171
+▁proportion -10.6171
+▁skull -10.6171
+▁smiling -10.6171
+▁empty -10.6171
+▁nigeria -10.6171
+▁cadiz -10.6171
+▁stuff -10.6171
+▁clothing -10.6171
+▁nueva -10.6171
+▁injury -10.6171
+▁dropped -10.6171
+▁dutch -10.6171
+▁syria -10.6171
+▁goddess -10.6171
+▁pirate -10.6171
+▁plus -10.6171
+▁illness -10.6171
+▁presidency -10.6172
+▁spiritual -10.6172
+▁cinematograph -10.6172
+▁irish -10.6173
+▁faithful -10.6173
+▁annoy -10.6173
+▁package -10.6173
+▁athens -10.6174
+▁forgot -10.6174
+▁delay -10.6175
+▁realistic -10.6175
+▁curious -10.6176
+▁sculptor -10.6179
+▁animation -10.6179
+▁contrast -10.6179
+▁controlled -10.6179
+▁sitting -10.6181
+▁cabinet -10.6182
+▁angle -10.6184
+▁attributed -10.6185
+▁successor -10.6185
+▁chase -10.6186
+▁theories -10.619
+▁destiny -10.6195
+▁continental -10.6196
+▁hyper -10.6213
+▁rap -10.6216
+▁finishes -10.6246
+tima -10.6248
+▁linear -10.6252
+iza -10.6277
+▁speaker -10.6331
+▁calci -10.643
+oya -10.6454
+▁refus -10.6519
+▁adopt -10.652
+▁threaten -10.6526
+▁highlight -10.6528
+▁fix -10.6542
+▁combine -10.6562
+▁exclusive -10.6571
+▁transfer -10.6584
+▁commu -10.6595
+▁punish -10.66
+▁arrange -10.6605
+▁brigade -10.6605
+▁conclude -10.6605
+▁controversial -10.6605
+▁dictatorship -10.6605
+▁disturb -10.6605
+▁emergency -10.6605
+▁grandparents -10.6605
+▁guarantee -10.6605
+▁hollywood -10.6605
+▁impression -10.6605
+▁industries -10.6605
+▁jurisdiction -10.6605
+▁limestone -10.6605
+▁navarra -10.6605
+▁netherlands -10.6605
+▁primitive -10.6605
+▁purchase -10.6605
+▁relevant -10.6605
+▁slaughter -10.6605
+▁variant -10.6605
+▁virtue -10.6605
+▁yucatan -10.6605
+▁advice -10.6605
+▁crush -10.6605
+▁download -10.6605
+▁easier -10.6605
+▁enemies -10.6605
+▁jordan -10.6605
+▁munich -10.6605
+▁populated -10.6605
+▁prayer -10.6605
+▁stretch -10.6605
+▁bogota -10.6605
+▁muscle -10.6605
+▁perceive -10.6605
+▁regiment -10.6605
+▁romanesque -10.6606
+▁transition -10.6606
+▁earlier -10.6606
+dolfo -10.6606
+▁unstable -10.6606
+▁sixteen -10.6606
+▁renew -10.6606
+▁cereal -10.6606
+▁michel -10.6606
+▁aragon -10.6606
+▁congratulations -10.6606
+▁oriented -10.6606
+▁punk -10.6606
+▁portray -10.6606
+▁dozen -10.6606
+▁phil -10.6606
+▁scholarship -10.6606
+▁comedian -10.6606
+▁copy -10.6607
+▁automatic -10.6607
+▁aware -10.6607
+▁oscar -10.6607
+▁string -10.6608
+▁joyful -10.6608
+▁dynamic -10.6609
+▁mask -10.6609
+▁rapid -10.6609
+▁boom -10.6609
+▁visitors -10.6611
+▁apse -10.6611
+▁editorial -10.6611
+▁castilla -10.6611
+▁toledo -10.6611
+▁revolutionary -10.6615
+▁valid -10.6617
+▁crash -10.6617
+▁mortal -10.6618
+▁horizontal -10.6619
+▁clari -10.662
+▁collector -10.662
+▁designated -10.6621
+▁stamp -10.6622
+▁shack -10.6624
+▁dizzy -10.6629
+▁jealousy -10.6629
+▁aunt -10.663
+▁jean -10.6631
+▁plata -10.6632
+▁mountainous -10.6635
+▁poland -10.6638
+▁marble -10.6639
+▁organic -10.6644
+▁soup -10.6673
+▁boil -10.668
+▁boxes -10.6681
+▁specialist -10.6682
+quez -10.6691
+▁tender -10.6694
+skirt -10.6704
+▁chess -10.671
+▁chu -10.6715
+▁passes -10.6715
+▁carv -10.6723
+▁institut -10.6724
+▁cow -10.6781
+▁cerr -10.6791
+▁mala -10.6795
+tally -10.6887
+maker -10.6918
+▁typ -10.6936
+▁complain -10.6992
+▁acquire -10.7016
+what -10.703
+▁communicat -10.7048
+▁construct -10.705
+polis -10.706
+▁accurate -10.706
+▁evident -10.706
+▁activist -10.706
+▁alejandro -10.706
+▁altitude -10.706
+▁appreciat -10.706
+▁cambridge -10.706
+▁decisive -10.706
+▁elizabeth -10.706
+▁equipped -10.706
+▁expansion -10.706
+▁fantastic -10.706
+▁frequency -10.706
+▁incredibl -10.706
+▁installation -10.706
+▁lawsuit -10.706
+▁madagascar -10.706
+▁morocco -10.706
+▁offensive -10.706
+▁ottoman -10.706
+▁predecessor -10.706
+▁procedure -10.706
+▁shrub -10.706
+▁sustain -10.706
+▁terrain -10.706
+▁transmission -10.706
+▁coconut -10.706
+▁identify -10.706
+▁masonry -10.706
+▁sevilla -10.706
+▁dragon -10.706
+▁throat -10.706
+▁arabic -10.706
+▁compilation -10.706
+▁eternal -10.706
+▁modification -10.706
+▁helmet -10.706
+▁method -10.706
+▁violin -10.706
+▁analyze -10.706
+▁spark -10.706
+▁dirty -10.706
+▁miami -10.706
+▁viceroy -10.706
+▁discipline -10.706
+▁absence -10.706
+▁dancing -10.706
+▁mystery -10.706
+▁fried -10.7061
+▁squad -10.7061
+▁rename -10.7062
+▁hired -10.7062
+▁ensure -10.7062
+▁gentleman -10.7063
+▁rigid -10.7064
+▁tongue -10.7065
+▁carbon -10.7065
+▁resume -10.7065
+▁worship -10.7065
+aceous -10.7065
+▁idol -10.7065
+▁integrated -10.7066
+▁conservation -10.7066
+▁tiger -10.7067
+▁closing -10.7068
+▁fiber -10.7069
+▁islamic -10.7069
+▁income -10.707
+▁succession -10.7071
+ridge -10.7071
+▁passage -10.7072
+▁coloniz -10.7072
+▁greet -10.7075
+▁poem -10.7077
+▁lemon -10.7078
+▁virginia -10.7078
+▁personality -10.708
+▁bitter -10.7083
+▁armenian -10.7084
+▁massive -10.7084
+▁norman -10.7084
+▁impos -10.7088
+▁parent -10.7093
+▁basilica -10.7095
+itz -10.7099
+▁phone -10.7101
+mund -10.7102
+▁wooden -10.7111
+▁raw -10.7115
+negro -10.7117
+▁toast -10.712
+▁boot -10.7128
+▁saga -10.7149
+▁armor -10.7153
+▁serv -10.7166
+▁slip -10.7241
+▁drove -10.7252
+▁fog -10.7258
+▁von -10.7274
+▁joke -10.7386
+▁edit -10.7443
+vista -10.7465
+▁advis -10.7491
+▁threat -10.7529
+▁educa -10.753
+▁altarpiece -10.7536
+▁archdiocese -10.7536
+▁budget -10.7536
+▁chemistry -10.7536
+▁confusing -10.7536
+▁defensive -10.7536
+▁emotional -10.7536
+▁episcopa -10.7536
+▁equator -10.7536
+▁ethiopia -10.7536
+▁expensive -10.7536
+▁hernandez -10.7536
+▁investment -10.7536
+▁lorenzo -10.7536
+▁michigan -10.7536
+▁paralympic -10.7536
+▁plumage -10.7536
+▁repertoire -10.7536
+▁structural -10.7536
+▁swallow -10.7536
+▁derived -10.7536
+▁fashion -10.7536
+▁florence -10.7536
+▁friday -10.7536
+▁minimal -10.7536
+▁peasant -10.7536
+▁qualify -10.7536
+▁surgery -10.7536
+▁pizza -10.7536
+▁scream -10.7536
+▁ciudad -10.7536
+▁devastat -10.7536
+▁nobility -10.7536
+▁copper -10.7536
+▁scotland -10.7536
+▁dominant -10.7536
+▁modified -10.7536
+▁andrés -10.7536
+▁monitor -10.7536
+▁envy -10.7536
+▁pastoral -10.7536
+▁frog -10.7536
+▁alonso -10.7536
+▁outdoor -10.7537
+▁burgos -10.7537
+▁employee -10.7537
+▁responsibility -10.7537
+▁oval -10.7537
+▁particle -10.7538
+▁lucas -10.7539
+▁consol -10.7539
+▁infantry -10.7539
+▁brotherhood -10.754
+▁lightning -10.754
+▁youtube -10.754
+▁export -10.7542
+▁inspect -10.7542
+atherine -10.7543
+▁elementary -10.7543
+▁adja -10.7544
+▁burial -10.7547
+▁factories -10.7547
+▁loving -10.7547
+▁roger -10.7551
+▁pearl -10.7552
+▁lucia -10.7552
+▁jenn -10.7553
+▁antique -10.7554
+▁johnson -10.7558
+▁plent -10.7559
+▁basal -10.7561
+▁duty -10.7563
+▁leonor -10.7573
+▁react -10.7576
+▁wonder -10.7576
+▁minute -10.7577
+▁raul -10.7578
+▁specie -10.7579
+rbon -10.758
+▁vigo -10.7585
+▁rail -10.7587
+▁sap -10.763
+▁luna -10.7652
+▁gau -10.7654
+▁bench -10.7665
+▁susan -10.768
+▁chris -10.7687
+▁helen -10.7701
+▁poli -10.7854
+▁select -10.7936
+▁associate -10.7993
+▁specialize -10.8003
+▁democrat -10.8013
+script -10.8017
+▁scott -10.8035
+▁affirm -10.8036
+▁airplane -10.8036
+▁aluminum -10.8036
+▁appointment -10.8036
+▁archbishop -10.8036
+▁artillery -10.8036
+▁astronomer -10.8036
+▁blossom -10.8036
+▁breakfast -10.8036
+▁caribbean -10.8036
+▁coincide -10.8036
+▁disappoint -10.8036
+▁distracted -10.8036
+▁draft -10.8036
+▁díaz -10.8036
+▁excavat -10.8036
+▁expelled -10.8036
+▁explode -10.8036
+▁exterior -10.8036
+▁guerrero -10.8036
+▁hungaria -10.8036
+▁ignacio -10.8036
+▁illuminate -10.8036
+▁inspiration -10.8036
+▁jesuit -10.8036
+▁kitchen -10.8036
+▁legislature -10.8036
+▁manufacturing -10.8036
+▁mathematics -10.8036
+▁negotiation -10.8036
+▁pontifical -10.8036
+▁prestigio -10.8036
+▁simpson -10.8036
+▁supplie -10.8036
+▁surrender -10.8036
+▁transparent -10.8036
+▁hungary -10.8036
+▁vessel -10.8036
+▁edible -10.8036
+▁excuse -10.8036
+▁senior -10.8036
+▁tissue -10.8036
+▁caracas -10.8036
+▁jersey -10.8036
+▁conquest -10.8036
+▁interaction -10.8036
+▁courtyard -10.8036
+▁regulate -10.8036
+▁tokyo -10.8036
+▁sauce -10.8036
+▁cardinal -10.8036
+▁stuck -10.8036
+▁cauca -10.8036
+▁aztec -10.8036
+▁martyr -10.8036
+▁secure -10.8036
+▁corpse -10.8036
+▁robust -10.8036
+▁rector -10.8036
+▁rational -10.8036
+▁sociolog -10.8037
+▁ugly -10.8037
+▁nerve -10.8037
+▁ascend -10.8037
+▁warrior -10.8037
+▁nightmare -10.8037
+▁crest -10.8037
+▁assets -10.8038
+▁virus -10.8038
+▁glad -10.8038
+▁juice -10.8038
+▁items -10.8039
+▁extinction -10.8039
+▁ballad -10.8041
+▁venue -10.8042
+▁feminist -10.8043
+point -10.8043
+▁turb -10.8044
+▁subway -10.8044
+▁brush -10.8044
+▁runner -10.8049
+▁creator -10.8049
+▁bold -10.8049
+▁harsh -10.8051
+▁lung -10.8053
+▁authorized -10.8053
+▁guardian -10.8054
+▁fake -10.8054
+▁tiny -10.8055
+▁kong -10.8055
+▁walter -10.8056
+▁fame -10.8064
+▁nick -10.8066
+▁islam -10.8072
+▁cub -10.8085
+▁nazi -10.8086
+▁castro -10.809
+▁vegetable -10.8102
+▁lyric -10.8104
+▁insect -10.811
+▁mental -10.8124
+▁update -10.8127
+▁server -10.8152
+shire -10.8191
+qualified -10.8241
+▁electro -10.8277
+▁supervis -10.8312
+gging -10.8338
+▁recommend -10.8476
+▁actual -10.8498
+▁gentle -10.8557
+▁femini -10.856
+río -10.8563
+ó -10.8563
+▁algorithm -10.8563
+▁ancestry -10.8563
+▁bermudez -10.8563
+▁coalition -10.8563
+▁confederation -10.8563
+▁creole -10.8563
+▁crustaceans -10.8563
+▁democracy -10.8563
+▁discharge -10.8563
+▁esteban -10.8563
+▁exciting -10.8563
+▁guadalupe -10.8563
+▁harvard -10.8563
+▁hawaii -10.8563
+▁impressive -10.8563
+▁inferior -10.8563
+▁magnificent -10.8563
+▁massachusetts -10.8563
+▁nocturnal -10.8563
+▁occupation -10.8563
+▁opportunities -10.8563
+▁pakistan -10.8563
+▁participating -10.8563
+▁protagonist -10.8563
+▁pseudonym -10.8563
+▁railroad -10.8563
+▁taylor -10.8563
+▁theoretical -10.8563
+▁tunnel -10.8563
+▁ukraine -10.8563
+▁ángel -10.8563
+▁cancelled -10.8563
+▁cuisine -10.8563
+▁delegate -10.8563
+▁impressed -10.8563
+▁manifest -10.8563
+▁resembl -10.8563
+▁chicken -10.8563
+▁complicated -10.8563
+▁glorious -10.8563
+▁migrate -10.8563
+▁kidnapping -10.8563
+▁remote -10.8563
+▁orchard -10.8563
+▁restoration -10.8563
+▁scheme -10.8563
+▁amazing -10.8563
+▁mutual -10.8563
+▁kenya -10.8563
+▁extent -10.8563
+▁optim -10.8563
+metry -10.8563
+▁angola -10.8563
+▁annex -10.8563
+▁toronto -10.8563
+▁cannibals -10.8563
+▁jews -10.8564
+▁summit -10.8564
+▁mouse -10.8564
+▁metallic -10.8564
+▁exposition -10.8564
+▁flame -10.8565
+▁admitted -10.8565
+▁planning -10.8565
+▁steal -10.8565
+▁inheritance -10.8566
+▁reproduction -10.8566
+▁citizenship -10.8566
+▁rioja -10.8568
+▁peaceful -10.8568
+▁punishment -10.8569
+▁stability -10.8571
+▁correspondent -10.8574
+▁pagan -10.8574
+▁shake -10.8576
+▁countryside -10.8577
+stratigraph -10.8577
+▁ballet -10.858
+▁ashes -10.8585
+▁difficulties -10.8586
+▁christianity -10.8587
+▁rush -10.8587
+▁businessman -10.8588
+▁protector -10.859
+▁pond -10.8597
+▁differentiate -10.8598
+▁legion -10.8601
+▁herald -10.8616
+▁jaw -10.8622
+▁kei -10.8653
+▁wipe -10.8689
+▁chronicle -10.8729
+▁creat -10.8845
+▁qualifie -10.8896
+▁ripe -10.8975
+▁marquis -10.9069
+▁immediate -10.9084
+ador -10.911
+▁accusation -10.9118
+▁activate -10.9118
+▁afghanistan -10.9118
+▁aforementioned -10.9118
+▁alvarez -10.9118
+▁apprentice -10.9118
+▁autumn -10.9118
+▁bilbao -10.9118
+▁cavalry -10.9118
+▁commitment -10.9118
+▁consequently -10.9118
+▁corporation -10.9118
+▁cylindrical -10.9118
+▁definitive -10.9118
+▁delegation -10.9118
+▁distinction -10.9118
+▁ecosystem -10.9118
+▁elaborate -10.9118
+▁grammy -10.9118
+▁guadalajara -10.9118
+▁hermitage -10.9118
+▁invertebrate -10.9118
+▁joaquín -10.9118
+▁meadows -10.9118
+▁mongolia -10.9118
+▁mozambique -10.9118
+▁offspring -10.9118
+▁outbreak -10.9118
+▁peculiar -10.9118
+▁privilege -10.9118
+▁puebla -10.9118
+▁reconstruct -10.9118
+▁rodrigo -10.9118
+▁steep -10.9118
+▁suspended -10.9118
+▁ultimate -10.9118
+▁absorb -10.9118
+▁dwarf -10.9118
+▁eyelids -10.9118
+▁fracture -10.9118
+▁sketch -10.9118
+▁cuadra -10.9118
+▁fabric -10.9118
+▁ignore -10.9118
+▁abbess -10.9118
+▁czech -10.9118
+▁random -10.9118
+▁catalina -10.9118
+▁chrono -10.9118
+▁remix -10.9118
+▁reunite -10.9118
+▁suburb -10.9118
+▁colonies -10.9118
+▁blanco -10.9118
+▁laguna -10.9118
+▁oriental -10.9118
+▁companion -10.9118
+▁knock -10.9118
+▁regain -10.9118
+▁gustav -10.9118
+▁sensor -10.9118
+▁pardon -10.9118
+▁innocent -10.9118
+itarian -10.9119
+▁ortiz -10.9119
+▁frost -10.9119
+▁cotton -10.9119
+bular -10.9119
+▁drain -10.9119
+▁courtesy -10.9119
+▁prosper -10.912
+▁davis -10.912
+▁mosque -10.912
+▁solar -10.9121
+▁savannah -10.9121
+▁scottish -10.9121
+▁turbine -10.9121
+▁persian -10.9121
+▁literal -10.9122
+▁gerard -10.9124
+▁grayish -10.9124
+▁scope -10.9124
+▁interfer -10.9125
+▁patio -10.9125
+▁norway -10.9125
+▁texture -10.9126
+▁alien -10.9127
+▁riders -10.9128
+▁lighthouse -10.9128
+▁truck -10.9129
+▁tight -10.9129
+▁barrio -10.9129
+▁harris -10.913
+▁colorado -10.9132
+▁fold -10.9133
+▁jury -10.9135
+▁paraguayan -10.9137
+▁dunes -10.9138
+▁payment -10.9143
+▁mold -10.9144
+▁possession -10.9148
+▁panel -10.9152
+clave -10.9153
+▁reflection -10.9156
+▁stew -10.923
+▁finger -10.9251
+intestin -10.9315
+▁robot -10.9317
+▁sigh -10.9356
+estio -10.9486
+▁diploma -10.952
+▁contra -10.9606
+▁execut -10.9609
+▁adapt -10.9611
+▁arrest -10.9622
+▁bibl -10.9661
+▁teenage -10.9676
+▁theatr -10.9688
+▁animat -10.9705
+▁desperate -10.9706
+▁anarchist -10.9706
+▁asteroid -10.9706
+▁atmosphere -10.9706
+▁automobile -10.9706
+▁breeze -10.9706
+▁brilliant -10.9706
+▁contempt -10.9706
+▁controversy -10.9706
+▁curiosity -10.9706
+▁expired -10.9706
+▁festivities -10.9706
+▁frustrat -10.9706
+▁granddaughter -10.9706
+▁hydra -10.9706
+▁imagination -10.9706
+▁mallorca -10.9706
+▁muñoz -10.9706
+▁napoleon -10.9706
+▁neutral -10.9706
+▁overwhelmed -10.9706
+▁profound -10.9706
+▁recruit -10.9706
+▁renowned -10.9706
+▁trujillo -10.9706
+▁unanimous -10.9706
+▁valparaiso -10.9706
+▁voluptuous -10.9706
+▁bacteria -10.9706
+▁discussed -10.9706
+▁downtown -10.9706
+▁improvis -10.9706
+▁jarocho -10.9706
+▁maintenance -10.9706
+▁patience -10.9706
+▁policies -10.9706
+▁asturian -10.9706
+▁bulbs -10.9706
+▁neoclassic -10.9706
+▁philip -10.9706
+▁pillars -10.9706
+▁poverty -10.9706
+▁sanchez -10.9706
+▁accessible -10.9706
+▁evangeli -10.9706
+▁penalty -10.9706
+▁sahara -10.9706
+▁alleged -10.9706
+▁surreal -10.9706
+▁dolores -10.9706
+▁archaeolog -10.9706
+▁speculat -10.9706
+▁changing -10.9706
+▁postpone -10.9707
+▁clip -10.9707
+▁ceiling -10.9707
+▁implied -10.9707
+▁implies -10.9707
+▁josé -10.9707
+▁raising -10.9707
+▁ideology -10.9707
+▁lonely -10.9707
+▁depict -10.9707
+▁weekend -10.9707
+▁fortune -10.9707
+▁behave -10.9707
+▁exceed -10.9707
+▁hooked -10.9707
+▁overcome -10.9707
+▁inscription -10.9707
+▁emmy -10.9707
+▁cruci -10.9707
+▁proceed -10.9707
+▁horribl -10.9708
+▁infant -10.9708
+▁charlie -10.9708
+ñez -10.9709
+▁axis -10.9709
+▁herrera -10.971
+▁slim -10.971
+▁introduction -10.971
+▁ruiz -10.9711
+▁plateau -10.9711
+▁foster -10.9713
+atomic -10.9713
+▁garcia -10.9716
+▁odd -10.9716
+▁innovation -10.9717
+▁arose -10.9718
+▁biology -10.9722
+▁wales -10.9723
+▁hong -10.9729
+▁fence -10.9741
+▁gang -10.9743
+therapy -10.9749
+physi -10.9772
+▁sophi -10.9869
+▁passenger -10.987
+racts -10.9871
+▁produc -10.9914
+andr -10.9955
+▁pep -11.0056
+▁geographic -11.017
+▁achiev -11.0184
+▁exhibit -11.0236
+▁preserve -11.0288
+▁fernand -11.0302
+loved -11.0304
+▁sculpt -11.0317
+▁innovati -11.0321
+▁conserv -11.0324
+▁consult -11.0329
+▁alicante -11.0331
+▁andalucia -11.0331
+▁anthropolog -11.0331
+▁asturias -11.0331
+▁audiovisual -11.0331
+▁autonomy -11.0331
+▁baptist -11.0331
+▁bonfire -11.0331
+▁cantabria -11.0331
+▁centaur -11.0331
+▁compatible -11.0331
+▁condemn -11.0331
+▁cyclist -11.0331
+▁denomination -11.0331
+▁discussion -11.0331
+▁explosion -11.0331
+▁fascinat -11.0331
+▁franchise -11.0331
+▁hammock -11.0331
+▁helicopter -11.0331
+▁hypothesis -11.0331
+▁implementation -11.0331
+▁intelligent -11.0331
+▁jesús -11.0331
+▁juvenile -11.0331
+▁larvae -11.0331
+▁membrane -11.0331
+▁mexico -11.0331
+▁migratory -11.0331
+▁mosquito -11.0331
+▁norwegian -11.0331
+▁pregnancy -11.0331
+▁pursued -11.0331
+▁sacrament -11.0331
+▁sebastián -11.0331
+▁solitary -11.0331
+▁sovereign -11.0331
+▁switch -11.0331
+▁tuesday -11.0331
+▁collapse -11.0331
+▁disguise -11.0331
+▁distress -11.0331
+▁intermedia -11.0331
+▁manchester -11.0331
+▁melody -11.0331
+▁miserable -11.0331
+▁paternal -11.0331
+▁prussia -11.0331
+▁retiring -11.0331
+▁sequence -11.0331
+▁suarez -11.0331
+▁comparison -11.0331
+▁emilio -11.0331
+▁wandering -11.0331
+▁ashlar -11.0331
+▁idiot -11.0331
+▁mercenar -11.0331
+▁asshole -11.0331
+▁gospel -11.0331
+▁mandatory -11.0331
+▁essence -11.0331
+▁patricia -11.0331
+▁podium -11.0331
+▁arthur -11.0331
+▁snack -11.0331
+▁claw -11.0331
+▁stroke -11.0332
+▁fault -11.0332
+▁fiesta -11.0332
+▁tremor -11.0332
+▁escort -11.0332
+▁centro -11.0332
+▁competing -11.0332
+▁nepal -11.0332
+▁orchid -11.0332
+▁multipli -11.0332
+▁laboratory -11.0332
+▁observations -11.0332
+▁cooperative -11.0333
+▁josef -11.0334
+▁circulation -11.0334
+▁advertisement -11.0334
+▁kidney -11.0334
+▁saddle -11.0334
+▁diagnosis -11.0334
+▁groom -11.0336
+▁tourney -11.0336
+▁clarity -11.0336
+▁espino -11.0336
+▁geography -11.0336
+▁protestant -11.0338
+▁bored -11.0338
+▁topography -11.0338
+▁johan -11.0339
+▁examine -11.0341
+▁drought -11.0341
+▁entities -11.0342
+▁tension -11.0342
+▁anderson -11.0342
+▁mural -11.0344
+finity -11.0349
+▁serg -11.035
+▁optic -11.0356
+omy -11.049
+nibal -11.0504
+▁illustr -11.0587
+acea -11.0706
+▁lamb -11.0709
+▁register -11.0911
+▁temper -11.0926
+▁galicia -11.095
+▁announce -11.0954
+worth -11.0967
+bank -11.0969
+direct -11.097
+omb -11.0989
+▁differ -11.0994
+▁scholar -11.0997
+▁abstract -11.0998
+▁alfonso -11.0998
+▁anthologies -11.0998
+▁calculation -11.0998
+▁cartagena -11.0998
+▁christopher -11.0998
+▁concentrate -11.0998
+▁confusion -11.0998
+▁consecrat -11.0998
+▁cricket -11.0998
+▁cristobal -11.0998
+▁dinosaur -11.0998
+▁eagle -11.0998
+▁elegant -11.0998
+▁endangered -11.0998
+▁entrepreneur -11.0998
+▁escuela -11.0998
+▁español -11.0998
+▁estimated -11.0998
+▁explanation -11.0998
+▁explosive -11.0998
+▁forbidden -11.0998
+▁furniture -11.0998
+▁glacier -11.0998
+▁illegal -11.0998
+▁kiwi -11.0998
+▁marqués -11.0998
+▁militant -11.0998
+▁occident -11.0998
+▁ourselves -11.0998
+▁rebuilt -11.0998
+▁recreation -11.0998
+▁seaweed -11.0998
+▁societies -11.0998
+▁suspicion -11.0998
+▁tenerife -11.0998
+▁vineyard -11.0998
+▁wool -11.0998
+▁departure -11.0998
+▁fantasy -11.0998
+▁interface -11.0998
+▁mozart -11.0998
+▁obvious -11.0998
+▁proximit -11.0998
+▁terrace -11.0998
+▁villaverde -11.0998
+▁castilian -11.0998
+▁mendoza -11.0998
+▁navarre -11.0998
+▁translator -11.0998
+▁zamora -11.0998
+▁nevada -11.0998
+▁havana -11.0998
+▁intensity -11.0998
+▁kevin -11.0998
+▁pillow -11.0998
+▁conifer -11.0998
+▁enhance -11.0998
+▁radius -11.0998
+▁utah -11.0998
+▁nineteen -11.0998
+▁widow -11.0998
+▁disaster -11.0998
+▁lawrence -11.0998
+▁thirst -11.0998
+▁seventeen -11.0998
+▁corolla -11.0998
+▁storage -11.0998
+▁trunk -11.0998
+▁eager -11.0998
+▁dulce -11.0998
+▁choreograph -11.0998
+▁yelled -11.0998
+▁resort -11.0998
+▁duration -11.0998
+▁replied -11.0998
+▁intervene -11.0998
+▁gesture -11.0998
+▁await -11.0998
+▁tierra -11.0998
+▁mocking -11.0998
+▁clara -11.0998
+▁entry -11.0999
+▁dragg -11.0999
+▁carpet -11.0999
+▁faint -11.0999
+▁requirements -11.0999
+▁pilar -11.0999
+▁whistle -11.0999
+▁vain -11.1
+▁fortified -11.1
+▁sequel -11.1
+▁pepper -11.1
+▁stephen -11.1
+▁gentlemen -11.1
+▁dorsal -11.1
+▁consultant -11.1
+▁basilio -11.1
+▁rapper -11.1
+▁philology -11.1001
+▁furious -11.1001
+▁golf -11.1001
+▁roast -11.1001
+▁venice -11.1001
+▁harbor -11.1001
+▁settlers -11.1002
+▁considerably -11.1002
+▁preference -11.1002
+▁shiver -11.1004
+▁convict -11.1004
+▁boost -11.1004
+▁distort -11.1005
+▁kitt -11.1012
+tropic -11.1015
+▁abuse -11.1022
+▁cleve -11.1043
+▁equ -11.1044
+rank -11.1138
+▁analog -11.1142
+mmit -11.1216
+▁receiv -11.1254
+uart -11.1301
+▁chronic -11.1489
+oxide -11.1545
+▁surround -11.1567
+▁flee -11.1638
+▁basil -11.1658
+▁retire -11.1664
+▁destin -11.1668
+▁occupie -11.1669
+▁propose -11.1669
+▁colonia -11.1676
+▁franc -11.1679
+scribed -11.1693
+▁kidnap -11.1712
+▁diamond -11.1712
+▁mathematical -11.1712
+▁advocat -11.1712
+▁affiliate -11.1712
+▁alexandria -11.1712
+▁arizona -11.1712
+▁assault -11.1712
+▁birmingham -11.1712
+▁carnival -11.1712
+▁competitive -11.1712
+▁complutense -11.1712
+▁conspiracy -11.1712
+▁cultivation -11.1712
+▁currency -11.1712
+▁deciduous -11.1712
+▁demolished -11.1712
+▁doctrine -11.1712
+▁elephant -11.1712
+▁encyclopedia -11.1712
+▁exploitation -11.1712
+▁illinois -11.1712
+▁istanbul -11.1712
+▁legislation -11.1712
+▁longitudinal -11.1712
+▁magdalena -11.1712
+▁margarita -11.1712
+▁mausoleum -11.1712
+▁pennsylvania -11.1712
+▁prosecutor -11.1712
+▁pyramid -11.1712
+▁registration -11.1712
+▁relegated -11.1712
+▁resolve -11.1712
+▁sicily -11.1712
+▁skeleton -11.1712
+▁suffrag -11.1712
+▁thorough -11.1712
+▁unbearable -11.1712
+▁zambia -11.1712
+▁zimbabwe -11.1712
+▁acceler -11.1712
+▁acoustic -11.1712
+▁betrayal -11.1712
+▁obeyed -11.1712
+▁oviedo -11.1712
+▁quantum -11.1712
+▁sumatra -11.1712
+▁surgeon -11.1712
+▁tendency -11.1712
+▁trigger -11.1712
+▁surname -11.1712
+▁attorney -11.1712
+▁bavaria -11.1712
+▁lizard -11.1712
+▁thief -11.1712
+▁conceiv -11.1712
+▁granite -11.1712
+▁summarize -11.1712
+▁murcia -11.1712
+▁waterfall -11.1712
+▁ghana -11.1712
+▁flamenco -11.1712
+▁flank -11.1712
+▁naples -11.1712
+▁vocation -11.1712
+▁evaluat -11.1712
+▁grief -11.1712
+▁serrat -11.1712
+▁dictat -11.1712
+▁cloister -11.1712
+▁atlanta -11.1712
+▁width -11.1712
+▁pianist -11.1712
+▁pupil -11.1712
+▁yelling -11.1712
+▁stirr -11.1712
+▁import -11.1713
+▁devout -11.1713
+▁ceased -11.1713
+▁iraq -11.1713
+▁arrangements -11.1713
+▁conservatory -11.1713
+▁serena -11.1713
+▁worried -11.1713
+▁imitate -11.1713
+▁tigre -11.1713
+▁gastro -11.1713
+▁portico -11.1713
+▁barrier -11.1713
+▁donkey -11.1713
+ogram -11.1713
+▁hebrew -11.1713
+neath -11.1713
+▁jacob -11.1713
+▁balear -11.1713
+▁disdainful -11.1713
+▁fetus -11.1714
+▁gently -11.1714
+▁harass -11.1714
+▁cheap -11.1714
+▁mortar -11.1715
+▁fusion -11.1715
+▁entre -11.1715
+▁bosnia -11.1715
+▁cried -11.1715
+▁mixture -11.1716
+cession -11.1716
+▁spill -11.1716
+▁vital -11.1716
+▁fuente -11.1717
+▁index -11.1719
+▁molecul -11.1725
+▁separati -11.1733
+ldom -11.1814
+▁francesc -11.1874
+crete -11.1893
+▁whip -11.1908
+▁meteorologic -11.1987
+▁dye -11.2153
+mount -11.2246
+▁succeed -11.2395
+▁float -11.2423
+▁jealous -11.2439
+▁dizz -11.244
+▁sudden -11.2447
+graduate -11.245
+▁confront -11.2473
+▁walt -11.2474
+▁abundance -11.2481
+▁accomplish -11.2481
+▁amsterdam -11.2481
+▁anxiety -11.2481
+▁approval -11.2481
+▁arrogant -11.2481
+▁axillary -11.2481
+▁baptized -11.2481
+▁bautista -11.2481
+▁ceremonies -11.2481
+▁circus -11.2481
+▁colleague -11.2481
+▁consolidated -11.2481
+▁continuation -11.2481
+▁creek -11.2481
+▁cretaceous -11.2481
+▁dialogue -11.2481
+▁douglas -11.2481
+▁ecclesiastic -11.2481
+▁elliptic -11.2481
+▁elongated -11.2481
+▁embark -11.2481
+▁engraving -11.2481
+▁enrolled -11.2481
+▁evacuate -11.2481
+▁frozen -11.2481
+▁gregorio -11.2481
+▁guayaquil -11.2481
+▁guerrilla -11.2481
+▁hidalgo -11.2481
+▁hybrid -11.2481
+▁indefinite -11.2481
+▁lithuania -11.2481
+▁majesty -11.2481
+▁manipulat -11.2481
+▁microsoft -11.2481
+▁namibia -11.2481
+▁perception -11.2481
+▁pleasant -11.2481
+▁possibilities -11.2481
+▁pronounc -11.2481
+▁rehearsal -11.2481
+▁satisfy -11.2481
+▁sebastian -11.2481
+▁slept -11.2481
+▁slovenia -11.2481
+▁sydney -11.2481
+▁tattoo -11.2481
+▁technologies -11.2481
+▁tributary -11.2481
+▁unconscious -11.2481
+▁uncovered -11.2481
+▁vargas -11.2481
+▁aquariu -11.2481
+▁batteries -11.2481
+▁contamina -11.2481
+▁counsel -11.2481
+▁credit -11.2481
+▁macedonia -11.2481
+▁patrick -11.2481
+▁posterior -11.2481
+▁supplement -11.2481
+▁uganda -11.2481
+onymous -11.2481
+▁anatomy -11.2481
+▁arsenal -11.2481
+▁coruña -11.2481
+▁garrison -11.2481
+▁qualities -11.2481
+▁signature -11.2481
+▁surpass -11.2481
+▁clums -11.2482
+▁monterrey -11.2482
+▁slogan -11.2482
+▁statute -11.2482
+gnition -11.2482
+▁radiation -11.2482
+▁reduction -11.2482
+▁tobacco -11.2482
+▁albania -11.2482
+▁dignity -11.2482
+▁erosion -11.2482
+▁futsal -11.2482
+▁invention -11.2482
+▁ancestor -11.2482
+▁fluid -11.2482
+▁tomás -11.2482
+▁abbey -11.2482
+▁curtain -11.2482
+▁legacy -11.2482
+▁homage -11.2482
+▁screw -11.2482
+gonzalo -11.2482
+▁arequipa -11.2482
+▁nelson -11.2482
+▁snail -11.2482
+▁niece -11.2482
+▁dairy -11.2482
+▁vergara -11.2482
+▁uprising -11.2482
+▁gabon -11.2482
+▁reunion -11.2482
+▁duchy -11.2482
+▁decease -11.2482
+▁almond -11.2482
+▁contestants -11.2482
+▁retain -11.2482
+▁lodge -11.2482
+▁potosi -11.2482
+▁mandate -11.2482
+▁blank -11.2482
+▁depilat -11.2482
+▁steam -11.2482
+▁spinning -11.2482
+▁valentin -11.2482
+▁housing -11.2482
+▁absent -11.2483
+▁decree -11.2483
+▁adobe -11.2483
+▁algeria -11.2483
+▁auction -11.2484
+▁vein -11.2484
+▁anglo -11.2484
+▁audio -11.2485
+▁adrian -11.2485
+▁loaf -11.2487
+ockey -11.2491
+▁quadr -11.2494
+▁knee -11.2499
+▁jewel -11.2532
+vinci -11.2544
+alawi -11.2566
+▁telephon -11.2647
+▁corona -11.2685
+▁biologic -11.2725
+▁practic -11.2764
+position -11.289
+apua -11.296
+▁meteor -11.3004
+▁georg -11.3021
+positor -11.3061
+▁celebrat -11.3095
+▁recogniz -11.3141
+doubted -11.3154
+▁distinguish -11.3229
+▁restrict -11.324
+egui -11.3246
+▁distribute -11.3272
+woman -11.3284
+▁alfred -11.3285
+culture -11.3289
+▁dispos -11.3297
+▁diagnos -11.3311
+tinent -11.3314
+▁embroider -11.3315
+▁ultra -11.3315
+▁absurd -11.3315
+▁agencies -11.3315
+▁albacete -11.3315
+▁belgrano -11.3315
+▁berkeley -11.3315
+▁bicycle -11.3315
+▁byzantine -11.3315
+▁cambodia -11.3315
+▁chaplain -11.3315
+▁cluster -11.3315
+▁commemorate -11.3315
+▁controversies -11.3315
+▁creep -11.3315
+▁crocodile -11.3315
+▁crusade -11.3315
+▁cylinder -11.3315
+▁cyprus -11.3315
+▁disembark -11.3315
+▁emphasize -11.3315
+▁enormous -11.3315
+▁evergreen -11.3315
+▁florenci -11.3315
+▁frederick -11.3315
+▁galaxy -11.3315
+▁gratitude -11.3315
+▁happily -11.3315
+▁hierarchy -11.3315
+▁indifferent -11.3315
+▁influential -11.3315
+▁interrogated -11.3315
+▁jimmy -11.3315
+▁libraries -11.3315
+▁luxurious -11.3315
+▁magistrate -11.3315
+▁minnesota -11.3315
+▁motorcycle -11.3315
+▁multiplayer -11.3315
+▁navarro -11.3315
+▁pamplona -11.3315
+▁pharmac -11.3315
+▁platinum -11.3315
+▁quartet -11.3315
+▁ramírez -11.3315
+▁reappear -11.3315
+▁simón -11.3315
+▁soluble -11.3315
+▁spectrum -11.3315
+▁spontaneous -11.3315
+▁stablish -11.3315
+▁subordinate -11.3315
+▁suppress -11.3315
+▁suspension -11.3315
+▁sweetheart -11.3315
+▁syndrome -11.3315
+▁temporal -11.3315
+▁twentie -11.3315
+▁ulysses -11.3315
+▁unexpected -11.3315
+▁vancouver -11.3315
+▁vatican -11.3315
+▁wrestling -11.3315
+▁alphabet -11.3315
+▁corridor -11.3315
+▁embodie -11.3315
+▁foliage -11.3315
+▁intersection -11.3315
+▁jamaica -11.3315
+▁midnight -11.3315
+▁module -11.3315
+▁nutritio -11.3315
+▁quintet -11.3315
+▁shepherd -11.3315
+▁thread -11.3315
+▁transformation -11.3315
+▁calderon -11.3315
+▁expense -11.3315
+▁facebook -11.3315
+▁lutheran -11.3315
+▁singular -11.3315
+▁impulse -11.3315
+▁lisbon -11.3315
+▁pocket -11.3315
+▁cheerful -11.3315
+▁debris -11.3315
+▁domestic -11.3315
+▁massacre -11.3315
+▁rabbi -11.3315
+▁sediment -11.3315
+▁teatro -11.3315
+▁buzz -11.3315
+▁artisan -11.3315
+▁leonardo -11.3315
+▁soledad -11.3315
+▁weakness -11.3315
+▁adhere -11.3315
+▁balkan -11.3315
+▁galleries -11.3315
+▁heaven -11.3315
+▁austin -11.3315
+▁kirk -11.3315
+▁eurasia -11.3315
+▁traitor -11.3315
+▁hunger -11.3315
+▁petersburg -11.3315
+▁montreal -11.3315
+▁somehow -11.3315
+▁antenna -11.3315
+▁atalaya -11.3315
+▁fraud -11.3315
+▁interval -11.3315
+▁rifle -11.3315
+▁tibet -11.3315
+▁appeal -11.3315
+▁seventies -11.3315
+▁sofia -11.3316
+▁mundo -11.3316
+▁output -11.3316
+▁sharing -11.3316
+▁oath -11.3316
+▁berries -11.3316
+▁elevate -11.3316
+▁carlist -11.3316
+▁fresco -11.3316
+immun -11.3317
+cellular -11.3318
+▁libertad -11.3318
+▁anime -11.3319
+▁señor -11.3321
+▁yama -11.3322
+▁addict -11.3322
+▁titan -11.3326
+▁vaca -11.3327
+▁alarm -11.3331
+▁restore -11.3371
+▁alpin -11.3394
+▁antoni -11.3422
+angular -11.3679
+▁philippine -11.3762
+▁recall -11.382
+aval -11.3837
+precedent -11.3969
+▁remov -11.4085
+▁armenia -11.4176
+▁necessar -11.4183
+stream -11.4193
+american -11.4194
+▁austral -11.4214
+▁admit -11.4221
+▁acknowledg -11.4224
+▁exponent -11.4224
+▁percussion -11.4224
+▁psychiatr -11.4224
+▁visigoth -11.4224
+▁yugoslav -11.4224
+nunciation -11.4224
+▁agustín -11.4224
+▁argentino -11.4224
+▁auditorium -11.4224
+▁azerbaijan -11.4224
+▁benedict -11.4224
+▁chromosom -11.4224
+▁coherent -11.4224
+▁compensation -11.4224
+▁congregation -11.4224
+▁constellation -11.4224
+▁criteria -11.4224
+▁croquettes -11.4224
+▁detroit -11.4224
+▁dioecious -11.4224
+▁extremadura -11.4224
+▁faucet -11.4224
+▁gibraltar -11.4224
+▁graduation -11.4224
+▁hemisphere -11.4224
+▁iglesia -11.4224
+▁irrigation -11.4224
+▁liquor -11.4224
+▁lombard -11.4224
+▁maldonado -11.4224
+▁medellin -11.4224
+▁militia -11.4224
+▁minimum -11.4224
+▁missouri -11.4224
+▁mourning -11.4224
+▁olimpia -11.4224
+▁patagonia -11.4224
+▁patriarch -11.4224
+▁patricio -11.4224
+▁permission -11.4224
+▁preservation -11.4224
+▁prestige -11.4224
+▁prolific -11.4224
+▁prosecuted -11.4224
+▁prototype -11.4224
+▁puzzle -11.4224
+▁reggae -11.4224
+▁reservoir -11.4224
+▁resolution -11.4224
+▁scandinavia -11.4224
+▁serrano -11.4224
+▁shipwreck -11.4224
+▁sidewalk -11.4224
+▁staircase -11.4224
+▁submarine -11.4224
+▁subsidiar -11.4224
+▁surprising -11.4224
+▁tarragona -11.4224
+▁technological -11.4224
+▁tragedies -11.4224
+▁tuxtlan -11.4224
+▁twitter -11.4224
+▁ukrainian -11.4224
+▁vicinity -11.4224
+▁vulnerable -11.4224
+▁widespread -11.4224
+▁abdomen -11.4224
+▁abilities -11.4224
+▁behalf -11.4224
+▁belgian -11.4224
+▁coexist -11.4224
+▁glued -11.4224
+▁jupiter -11.4224
+▁memoirs -11.4224
+▁phillip -11.4224
+▁specified -11.4224
+▁tragedy -11.4224
+▁glabro -11.4224
+▁gordon -11.4224
+▁grammar -11.4224
+▁lifestyle -11.4224
+▁oxygen -11.4224
+▁simplicity -11.4224
+▁betrayed -11.4224
+▁ferran -11.4224
+▁nomination -11.4224
+▁perfume -11.4224
+▁sausage -11.4224
+▁taxonom -11.4224
+▁uneven -11.4224
+▁humbert -11.4224
+▁scanner -11.4224
+▁botany -11.4224
+▁inventor -11.4224
+▁russell -11.4224
+▁houston -11.4224
+▁lewis -11.4224
+▁bucket -11.4224
+▁orientation -11.4224
+▁palermo -11.4224
+▁sensation -11.4224
+▁insurance -11.4224
+▁virgen -11.4224
+▁depriv -11.4224
+▁mosaic -11.4224
+▁sultan -11.4224
+▁abraham -11.4224
+▁dylan -11.4224
+▁invented -11.4224
+▁ritual -11.4224
+▁kidnapped -11.4224
+▁anguish -11.4224
+▁blanca -11.4224
+▁vincent -11.4224
+▁density -11.4224
+▁horseshoe -11.4224
+▁santander -11.4224
+▁abortion -11.4224
+▁arbore -11.4224
+▁trinity -11.4224
+▁patrol -11.4224
+▁pigeon -11.4224
+▁clue -11.4224
+▁binary -11.4224
+▁sparta -11.4224
+▁erotic -11.4224
+▁senate -11.4225
+▁alpha -11.4225
+▁impatien -11.4225
+▁hollow -11.4225
+▁carrera -11.4225
+▁polar -11.4225
+▁deficien -11.4226
+▁pulp -11.4226
+▁drill -11.4226
+▁libya -11.4226
+ghett -11.4227
+▁predict -11.4227
+▁davie -11.4227
+dmother -11.4229
+▁leopol -11.4232
+ventura -11.4233
+▁flute -11.4233
+wald -11.4271
+▁niche -11.4286
+arochi -11.4362
+▁sensib -11.4404
+agram -11.4789
+▁adventur -11.5101
+finals -11.5139
+▁approve -11.518
+▁eliminate -11.518
+▁praise -11.518
+▁decorati -11.5212
+ffluent -11.5224
+▁absolve -11.5224
+▁accuracy -11.5224
+▁ambitious -11.5224
+▁anniversary -11.5224
+▁antioquia -11.5224
+▁appropriate -11.5224
+▁aragón -11.5224
+▁aristocrat -11.5224
+▁astronomy -11.5224
+▁baptism -11.5224
+▁benjamin -11.5224
+▁bluish -11.5224
+▁bordeaux -11.5224
+▁boulevard -11.5224
+▁brooklyn -11.5224
+▁buffalo -11.5224
+▁bundesliga -11.5224
+▁casualties -11.5224
+▁catolica -11.5224
+▁censorship -11.5224
+▁centimeters -11.5224
+▁certified -11.5224
+▁competitor -11.5224
+▁compostela -11.5224
+▁continuity -11.5224
+▁conversion -11.5224
+▁corrientes -11.5224
+▁czechoslovakia -11.5224
+▁declaration -11.5224
+▁depressed -11.5224
+▁derivative -11.5224
+▁disadvantage -11.5224
+▁eclipse -11.5224
+▁edinburgh -11.5224
+▁enthusiast -11.5224
+▁epidemic -11.5224
+▁ethnic -11.5224
+▁excellenc -11.5224
+▁excursion -11.5224
+▁exploited -11.5224
+▁eyebrow -11.5224
+▁frighten -11.5224
+▁grandchildren -11.5224
+▁grotesque -11.5224
+▁hurricane -11.5224
+▁independiente -11.5224
+▁indicator -11.5224
+▁intermittent -11.5224
+▁internacional -11.5224
+▁lincoln -11.5224
+▁liverpool -11.5224
+▁metaboli -11.5224
+▁modality -11.5224
+▁nashville -11.5224
+▁nicholas -11.5224
+▁nutrients -11.5224
+▁obstacle -11.5224
+▁ordinary -11.5224
+▁philadelphia -11.5224
+▁phosphate -11.5224
+▁precarious -11.5224
+▁precipitation -11.5224
+▁predominate -11.5224
+▁promenade -11.5224
+▁psychologist -11.5224
+▁pubescent -11.5224
+▁registry -11.5224
+▁relentless -11.5224
+▁repercussion -11.5224
+▁resurrect -11.5224
+▁segovia -11.5224
+▁significance -11.5224
+▁simulation -11.5224
+▁slavic -11.5224
+▁slovakia -11.5224
+▁sociedad -11.5224
+▁soprano -11.5224
+▁spectacular -11.5224
+▁spokesperson -11.5224
+▁succulent -11.5224
+▁telescope -11.5224
+▁testimony -11.5224
+▁toponym -11.5224
+▁trajectory -11.5224
+▁trinidad -11.5224
+▁videoclip -11.5224
+▁vitamin -11.5224
+▁antofagasta -11.5224
+▁colón -11.5224
+▁despair -11.5224
+▁guarani -11.5224
+▁illusion -11.5224
+▁lleida -11.5224
+▁margaret -11.5224
+▁monotonous -11.5224
+▁premises -11.5224
+▁regret -11.5224
+▁solemn -11.5224
+▁sprouts -11.5224
+▁subdued -11.5224
+▁wizard -11.5224
+higgins -11.5224
+▁acknowledgement -11.5224
+▁assassin -11.5224
+▁gravity -11.5224
+▁hermaphrodite -11.5224
+▁puddle -11.5224
+▁respirat -11.5224
+▁robinson -11.5224
+▁spacious -11.5224
+▁submitt -11.5224
+▁vegetarian -11.5224
+▁cruise -11.5224
+▁kansas -11.5224
+▁oblong -11.5224
+▁rainbow -11.5224
+▁raymond -11.5224
+▁vowel -11.5224
+▁census -11.5224
+▁madonna -11.5224
+▁thorns -11.5224
+▁commentator -11.5224
+▁powder -11.5224
+▁recreate -11.5224
+▁savage -11.5224
+pelagic -11.5224
+▁dessert -11.5224
+▁pereira -11.5224
+▁spike -11.5224
+▁tunisia -11.5224
+▁cartridge -11.5224
+▁himalaya -11.5224
+▁sphere -11.5224
+▁urgent -11.5224
+▁ethno -11.5224
+▁upright -11.5224
+▁smash -11.5224
+▁survey -11.5224
+▁propell -11.5224
+▁franz -11.5224
+▁ideological -11.5224
+▁chimeric -11.5224
+▁ebro -11.5224
+▁overseas -11.5224
+▁federic -11.5224
+▁cairo -11.5225
+▁centen -11.5225
+▁bridle -11.5225
+▁hawk -11.5225
+▁fermi -11.5225
+▁stubb -11.5226
+▁civic -11.5227
+discreet -11.5228
+▁iconograph -11.5229
+▁tacit -11.523
+athlon -11.5231
+▁physiolog -11.5233
+▁pulsati -11.5233
+▁recurr -11.5347
+existent -11.5601
+▁significati -11.5871
+▁resource -11.5875
+▁determin -11.6056
+priesthood -11.6248
+▁identifi -11.6248
+▁olympi -11.6316
+▁extinct -11.6328
+▁homonym -11.6331
+▁disdain -11.6333
+▁basket -11.6335
+▁ambiguous -11.6335
+▁convenient -11.6335
+▁diagonal -11.6335
+▁excessive -11.6335
+ceptive -11.6335
+elecommunication -11.6335
+legitimate -11.6335
+▁acclaim -11.6335
+▁activism -11.6335
+▁admission -11.6335
+▁allegory -11.6335
+▁anglican -11.6335
+▁antagonist -11.6335
+▁apostol -11.6335
+▁aristotle -11.6335
+▁armstrong -11.6335
+▁arrogance -11.6335
+▁articulated -11.6335
+▁asunción -11.6335
+â -12.518
+. -12.518
+ʻ -12.8514
+ð -13.3514
+ú -13.3514
+ʿ -14.3514
+! -14.3514
+þ -14.3514
+ā -14.3514
+̇ -14.3514
+í -14.3514
+ə -14.3514
diff --git a/configs/fr-en/config_gcmvn.yaml b/configs/fr-en/config_gcmvn.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..9f71b517cbb1f564da016a9a6d5a12dd3744aa33
--- /dev/null
+++ b/configs/fr-en/config_gcmvn.yaml
@@ -0,0 +1,21 @@
+global_cmvn:
+ stats_npz_path: /data/zhangshaolei/StreamSpeech/configs/fr-en/gcmvn.npz
+input_channels: 1
+input_feat_per_channel: 80
+specaugment:
+ freq_mask_F: 27
+ freq_mask_N: 1
+ time_mask_N: 1
+ time_mask_T: 100
+ time_mask_p: 1.0
+ time_wrap_W: 0
+transforms:
+ '*':
+ - global_cmvn
+ _train:
+ - global_cmvn
+ - specaugment
+vocoder:
+ checkpoint: /data/zhangshaolei/pretrain_models/unit-based_HiFi-GAN_vocoder/mHuBERT.layer11.km1000.en/g_00500000
+ config: /data/zhangshaolei/pretrain_models/unit-based_HiFi-GAN_vocoder/mHuBERT.layer11.km1000.en/config.json
+ type: code_hifigan
diff --git a/configs/fr-en/config_mtl_asr_st_ctcst.yaml b/configs/fr-en/config_mtl_asr_st_ctcst.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..e6eb374a2ae77a8943958abe1a0c792716098a74
--- /dev/null
+++ b/configs/fr-en/config_mtl_asr_st_ctcst.yaml
@@ -0,0 +1,36 @@
+target_unigram:
+ decoder_type: transformer
+ dict: /data/zhangshaolei/StreamSpeech/configs/fr-en/tgt_unigram6000/spm_unigram_fr.txt
+ data: /data/zhangshaolei/StreamSpeech/configs/fr-en/tgt_unigram6000
+ loss_weight: 8.0
+ rdrop_alpha: 0.0
+ decoder_args:
+ decoder_layers: 4
+ decoder_embed_dim: 512
+ decoder_ffn_embed_dim: 2048
+ decoder_attention_heads: 8
+ label_smoothing: 0.1
+source_unigram:
+ decoder_type: ctc
+ dict: /data/zhangshaolei/StreamSpeech/configs/fr-en/src_unigram6000/spm_unigram_fr.txt
+ data: /data/zhangshaolei/StreamSpeech/configs/fr-en/src_unigram6000
+ loss_weight: 4.0
+ rdrop_alpha: 0.0
+ decoder_args:
+ decoder_layers: 0
+ decoder_embed_dim: 512
+ decoder_ffn_embed_dim: 2048
+ decoder_attention_heads: 8
+ label_smoothing: 0.1
+ctc_target_unigram:
+ decoder_type: ctc
+ dict: /data/zhangshaolei/StreamSpeech/configs/fr-en/tgt_unigram6000/spm_unigram_fr.txt
+ data: /data/zhangshaolei/StreamSpeech/configs/fr-en/tgt_unigram6000
+ loss_weight: 4.0
+ rdrop_alpha: 0.0
+ decoder_args:
+ decoder_layers: 0
+ decoder_embed_dim: 512
+ decoder_ffn_embed_dim: 2048
+ decoder_attention_heads: 8
+ label_smoothing: 0.1
\ No newline at end of file
diff --git a/configs/fr-en/config_unity.yaml b/configs/fr-en/config_unity.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..3ba193b4bd163d53ec4d3df51e9b5deff19cf6c2
--- /dev/null
+++ b/configs/fr-en/config_unity.yaml
@@ -0,0 +1,12 @@
+target_unigram:
+ decoder_type: transformer
+ dict: /data/zhangshaolei/StreamSpeech/configs/fr-en/tgt_unigram6000/spm_unigram_fr.txt
+ data: /data/zhangshaolei/StreamSpeech/configs/fr-en/tgt_unigram6000
+ loss_weight: 8.0
+ rdrop_alpha: 0.0
+ decoder_args:
+ decoder_layers: 4
+ decoder_embed_dim: 512
+ decoder_ffn_embed_dim: 2048
+ decoder_attention_heads: 8
+ label_smoothing: 0.1
\ No newline at end of file
diff --git a/configs/fr-en/gcmvn.npz b/configs/fr-en/gcmvn.npz
new file mode 100644
index 0000000000000000000000000000000000000000..0308fc4ffc056e29226d26aa54de58fa4e8f8d50
--- /dev/null
+++ b/configs/fr-en/gcmvn.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ea70c4e0e2465ce81e6dd9f9a7e580ed6d3c0d9f089567d1aea7a57ff4203566
+size 1140
diff --git a/configs/fr-en/src_unigram6000/spm_unigram_fr.model b/configs/fr-en/src_unigram6000/spm_unigram_fr.model
new file mode 100644
index 0000000000000000000000000000000000000000..09825e0e461226d6c8b94936ecaf54d53349ca9b
--- /dev/null
+++ b/configs/fr-en/src_unigram6000/spm_unigram_fr.model
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7917b5d10dac2aefc3119c42b3043b909afc6fd96ff05d05daae191f85d520e5
+size 344305
diff --git a/configs/fr-en/src_unigram6000/spm_unigram_fr.txt b/configs/fr-en/src_unigram6000/spm_unigram_fr.txt
new file mode 100644
index 0000000000000000000000000000000000000000..fdc2587aba95e36d97161df88d8007ec06e316ae
--- /dev/null
+++ b/configs/fr-en/src_unigram6000/spm_unigram_fr.txt
@@ -0,0 +1,5996 @@
+s 1
+▁de 1
+▁la 1
+▁le 1
+▁à 1
+e 1
+▁est 1
+▁il 1
+▁et 1
+▁les 1
+r 1
+▁en 1
+▁du 1
+▁a 1
+▁un 1
+▁des 1
+t 1
+▁au 1
+▁une 1
+▁dans 1
+▁je 1
+▁pour 1
+▁se 1
+▁pas 1
+▁elle 1
+▁que 1
+▁cent 1
+▁par 1
+▁ce 1
+▁ne 1
+▁rue 1
+▁sa 1
+▁son 1
+▁ 1
+▁cest 1
+a 1
+▁vous 1
+nt 1
+é 1
+n 1
+▁sur 1
+▁deux 1
+▁sont 1
+er 1
+▁plus 1
+es 1
+y 1
+c 1
+ment 1
+▁qui 1
+▁cette 1
+l 1
+u 1
+▁avec 1
+▁nous 1
+ait 1
+▁monsieur 1
+i 1
+ra 1
+▁ma 1
+▁numéro 1
+▁mille 1
+▁on 1
+ant 1
+▁trois 1
+d 1
+le 1
+▁mais 1
+z 1
+▁ses 1
+▁y 1
+re 1
+▁été 1
+o 1
+▁quatre 1
+m 1
+▁fait 1
+▁me 1
+▁parole 1
+▁comme 1
+▁lui 1
+▁bien 1
+▁zéro 1
+p 1
+▁tu 1
+ne 1
+it 1
+▁dé 1
+te 1
+▁tout 1
+▁re 1
+▁aux 1
+in 1
+▁était 1
+la 1
+g 1
+il 1
+▁ces 1
+x 1
+ée 1
+▁également 1
+f 1
+▁dun 1
+de 1
+▁lamendement 1
+▁si 1
+ri 1
+▁aussi 1
+ur 1
+ent 1
+lo 1
+▁quil 1
+▁mon 1
+▁donc 1
+me 1
+is 1
+que 1
+▁ou 1
+▁soutenir 1
+di 1
+▁da 1
+▁madame 1
+li 1
+on 1
+▁alors 1
+▁dune 1
+en 1
+▁même 1
+sse 1
+▁cinq 1
+▁va 1
+▁très 1
+▁ont 1
+ch 1
+té 1
+▁in 1
+▁na 1
+▁ça 1
+▁é 1
+▁jai 1
+▁fut 1
+b 1
+ation 1
+▁route 1
+▁leur 1
+v 1
+aient 1
+ti 1
+ro 1
+k 1
+▁six 1
+▁nest 1
+na 1
+▁cet 1
+▁suis 1
+▁ta 1
+ma 1
+▁être 1
+▁cela 1
+ta 1
+▁ra 1
+▁saint 1
+ff 1
+▁ils 1
+▁avait 1
+che 1
+▁l 1
+▁cents 1
+▁cha 1
+▁tous 1
+▁ville 1
+▁temps 1
+ci 1
+▁peu 1
+h 1
+▁chemin 1
+▁p 1
+▁vingt 1
+ce 1
+▁après 1
+▁plusieurs 1
+▁con 1
+▁d 1
+▁nom 1
+▁entre 1
+▁ca 1
+▁huit 1
+▁sans 1
+▁sept 1
+ge 1
+▁faire 1
+ie 1
+al 1
+▁dix 1
+▁votre 1
+▁lé 1
+▁peut 1
+▁ré 1
+nd 1
+vi 1
+▁pro 1
+▁c 1
+▁porte 1
+▁moi 1
+▁co 1
+ni 1
+mi 1
+ons 1
+lé 1
+▁dit 1
+▁sous 1
+ve 1
+an 1
+da 1
+va 1
+▁ba 1
+▁amendement 1
+ine 1
+▁partie 1
+▁di 1
+▁quelle 1
+▁car 1
+▁encore 1
+▁trente 1
+tte 1
+▁premier 1
+age 1
+at 1
+co 1
+▁te 1
+mp 1
+ité 1
+▁mar 1
+ton 1
+▁pa 1
+us 1
+ier 1
+▁père 1
+▁là 1
+rai 1
+pe 1
+ba 1
+▁première 1
+▁grand 1
+▁s 1
+èrent 1
+au 1
+▁non 1
+▁où 1
+ement 1
+ré 1
+▁li 1
+▁e 1
+ion 1
+▁ainsi 1
+▁bo 1
+▁avenue 1
+▁place 1
+▁bon 1
+▁groupe 1
+or 1
+ca 1
+ter 1
+ga 1
+▁ro 1
+▁g 1
+ées 1
+tre 1
+th 1
+ll 1
+be 1
+▁ni 1
+les 1
+ez 1
+ais 1
+▁pi 1
+lu 1
+▁po 1
+▁quand 1
+▁tour 1
+▁ja 1
+▁i 1
+lle 1
+▁neuf 1
+▁o 1
+▁toujours 1
+▁jean 1
+aire 1
+▁quelques 1
+el 1
+▁puis 1
+bi 1
+ard 1
+no 1
+ance 1
+▁paris 1
+go 1
+▁pré 1
+▁or 1
+ille 1
+▁famille 1
+▁vi 1
+tu 1
+▁dis 1
+▁trouve 1
+▁commune 1
+aux 1
+▁ar 1
+▁faut 1
+▁fois 1
+▁quarante 1
+▁ensuite 1
+ar 1
+▁cinquante 1
+▁fils 1
+▁vers 1
+ique 1
+eur 1
+ir 1
+et 1
+to 1
+▁mont 1
+rie 1
+▁aujourdhui 1
+▁lo 1
+▁contre 1
+▁france 1
+és 1
+gue 1
+▁leurs 1
+ac 1
+▁ga 1
+▁grande 1
+ng 1
+▁fa 1
+vient 1
+▁sé 1
+able 1
+ck 1
+né 1
+▁rien 1
+▁eu 1
+▁reste 1
+elle 1
+um 1
+▁do 1
+▁depuis 1
+▁sagit 1
+ry 1
+▁soixante 1
+▁lin 1
+▁mi 1
+▁pendant 1
+▁fin 1
+▁ans 1
+do 1
+mé 1
+▁cependant 1
+gi 1
+ou 1
+rent 1
+▁mal 1
+▁pays 1
+▁sera 1
+▁ah 1
+iste 1
+▁sou 1
+vo 1
+ul 1
+ot 1
+▁ri 1
+▁quel 1
+pp 1
+ure 1
+▁monde 1
+▁b 1
+▁ex 1
+ron 1
+▁bi 1
+▁mort 1
+ville 1
+du 1
+▁mé 1
+po 1
+tra 1
+▁beaucoup 1
+▁al 1
+▁jamais 1
+▁point 1
+tion 1
+▁oui 1
+▁vie 1
+▁gouvernement 1
+▁ju 1
+▁maison 1
+▁guerre 1
+▁sil 1
+▁comment 1
+mo 1
+▁moins 1
+▁avant 1
+▁tard 1
+▁es 1
+ver 1
+▁cas 1
+pa 1
+▁femme 1
+ê 1
+▁petit 1
+▁fille 1
+▁fais 1
+▁rencontre 1
+▁fi 1
+sur 1
+w 1
+gu 1
+ction 1
+ux 1
+▁sûr 1
+▁ny 1
+▁trop 1
+ph 1
+▁k 1
+man 1
+▁mère 1
+▁lu 1
+▁général 1
+rait 1
+im 1
+▁pu 1
+▁per 1
+▁doit 1
+▁ha 1
+▁avoir 1
+▁ici 1
+if 1
+ia 1
+▁cor 1
+▁jour 1
+▁bou 1
+ing 1
+▁forme 1
+▁m 1
+▁quatrevingtdix 1
+ha 1
+▁notre 1
+▁travail 1
+eux 1
+til 1
+bo 1
+dé 1
+▁lors 1
+▁ministre 1
+ette 1
+▁homme 1
+▁toute 1
+▁su 1
+▁lun 1
+▁compte 1
+▁chose 1
+ère 1
+▁autres 1
+ru 1
+▁pierre 1
+▁né 1
+▁dautres 1
+▁français 1
+▁située 1
+▁mes 1
+▁toutes 1
+▁dor 1
+▁bonne 1
+▁commission 1
+▁voir 1
+▁port 1
+▁voilà 1
+ï 1
+▁avez 1
+▁village 1
+ale 1
+▁étaient 1
+use 1
+▁cétait 1
+▁club 1
+▁jeune 1
+▁autre 1
+▁soixantedix 1
+▁part 1
+isme 1
+▁cou 1
+▁souvent 1
+▁fond 1
+▁coup 1
+▁fer 1
+ront 1
+gé 1
+▁sest 1
+ière 1
+don 1
+▁be 1
+▁elles 1
+qu 1
+min 1
+ka 1
+▁assez 1
+▁joue 1
+▁espèce 1
+▁droit 1
+▁onze 1
+rou 1
+son 1
+▁ton 1
+▁jours 1
+▁so 1
+cu 1
+▁gra 1
+▁situé 1
+▁mo 1
+▁chez 1
+ut 1
+▁nos 1
+▁questce 1
+ise 1
+eau 1
+▁loi 1
+pé 1
+▁demande 1
+▁pourquoi 1
+qui 1
+cha 1
+▁men 1
+tique 1
+▁sud 1
+▁vit 1
+mar 1
+▁donne 1
+▁ru 1
+bu 1
+▁suite 1
+ateur 1
+▁nord 1
+▁dire 1
+▁impasse 1
+ran 1
+▁comp 1
+▁lieu 1
+▁quelque 1
+▁enfants 1
+▁bar 1
+▁pe 1
+▁léglise 1
+▁sœur 1
+un 1
+▁nouveau 1
+▁seul 1
+▁plan 1
+▁toi 1
+▁ho 1
+▁président 1
+▁frère 1
+▁beau 1
+▁train 1
+▁char 1
+▁dont 1
+▁prend 1
+▁vo 1
+▁tête 1
+▁titre 1
+mont 1
+▁fort 1
+▁douze 1
+▁cour 1
+▁no 1
+▁région 1
+ol 1
+▁moment 1
+▁jo 1
+▁nombreux 1
+▁déjà 1
+▁château 1
+▁val 1
+▁notamment 1
+▁politique 1
+vé 1
+▁ver 1
+chi 1
+▁allée 1
+tin 1
+lan 1
+▁mot 1
+con 1
+▁cours 1
+urs 1
+là 1
+je 1
+▁parti 1
+▁sen 1
+▁dernier 1
+ence 1
+cé 1
+▁mari 1
+▁quinze 1
+op 1
+ions 1
+▁an 1
+mer 1
+▁quon 1
+▁petite 1
+▁effet 1
+▁bois 1
+nes 1
+ger 1
+▁avons 1
+om 1
+▁gi 1
+rons 1
+ens 1
+▁luniversité 1
+▁mou 1
+▁dar 1
+ssa 1
+ive 1
+▁bri 1
+▁connu 1
+tes 1
+▁maintenant 1
+ché 1
+▁sais 1
+▁avis 1
+▁go 1
+▁heures 1
+land 1
+ki 1
+▁siège 1
+▁présente 1
+hi 1
+è 1
+▁amendements 1
+▁st 1
+▁carrière 1
+ble 1
+mu 1
+ire 1
+▁gu 1
+▁fit 1
+ienne 1
+▁mois 1
+▁soit 1
+bre 1
+illa 1
+▁commence 1
+▁ti 1
+▁quoi 1
+tif 1
+ner 1
+▁année 1
+▁saison 1
+▁vu 1
+os 1
+▁ai 1
+â 1
+▁tant 1
+▁passe 1
+▁travaille 1
+des 1
+▁terre 1
+▁vais 1
+ide 1
+ja 1
+▁long 1
+par 1
+▁bas 1
+▁cro 1
+▁crois 1
+▁centre 1
+▁serait 1
+▁certains 1
+ante 1
+▁vol 1
+▁ka 1
+▁personne 1
+mes 1
+▁nai 1
+▁ami 1
+▁ch 1
+▁larticle 1
+nc 1
+▁estce 1
+▁durant 1
+▁nouvelle 1
+▁près 1
+▁cher 1
+▁peuvent 1
+▁défavorable 1
+▁cri 1
+▁quatorze 1
+ations 1
+euse 1
+▁vrai 1
+za 1
+▁bu 1
+▁roi 1
+ze 1
+▁cu 1
+▁seule 1
+▁col 1
+bl 1
+▁semble 1
+▁mor 1
+▁film 1
+▁louis 1
+▁prendre 1
+ois 1
+▁aller 1
+▁raison 1
+▁important 1
+▁conseil 1
+let 1
+and 1
+tait 1
+▁lavis 1
+gra 1
+▁pont 1
+ig 1
+▁présent 1
+▁seize 1
+dre 1
+▁t 1
+▁côté 1
+▁projet 1
+▁gens 1
+▁ci 1
+▁début 1
+▁deuxième 1
+vent 1
+▁lautre 1
+▁mer 1
+▁fo 1
+rez 1
+lli 1
+▁dernière 1
+ès 1
+▁oh 1
+val 1
+cou 1
+▁acc 1
+▁celle 1
+▁soir 1
+dou 1
+▁aimé 1
+▁ligne 1
+▁chambre 1
+▁poste 1
+fi 1
+▁cla 1
+▁permet 1
+▁man 1
+▁haut 1
+▁font 1
+ient 1
+▁mieux 1
+cul 1
+▁question 1
+▁nombre 1
+▁devant 1
+▁enfin 1
+gne 1
+▁voix 1
+▁inter 1
+▁lécole 1
+▁furent 1
+▁léquipe 1
+▁aime 1
+lie 1
+pi 1
+vingtdixsept 1
+▁voie 1
+ite 1
+▁main 1
+▁écrit 1
+▁fu 1
+▁problème 1
+▁table 1
+ven 1
+▁série 1
+teur 1
+▁sy 1
+lon 1
+▁gauche 1
+▁jusquà 1
+ô 1
+▁produit 1
+▁sujet 1
+▁nombreuses 1
+▁combien 1
+ezvous 1
+qua 1
+tation 1
+ler 1
+▁mur 1
+▁din 1
+▁he 1
+▁dabord 1
+▁hu 1
+▁treize 1
+▁droite 1
+vers 1
+▁selon 1
+ris 1
+▁dan 1
+▁jeu 1
+▁met 1
+ic 1
+ct 1
+ya 1
+▁département 1
+tta 1
+▁membre 1
+▁mesure 1
+ue 1
+iers 1
+▁venu 1
+▁études 1
+▁aucun 1
+ai 1
+vin 1
+ho 1
+aud 1
+▁mis 1
+▁vent 1
+onne 1
+▁garde 1
+▁paul 1
+lin 1
+▁bé 1
+▁tra 1
+▁mu 1
+▁face 1
+▁êtes 1
+▁dêtre 1
+▁comprend 1
+▁sens 1
+▁fou 1
+▁possède 1
+bé 1
+▁existe 1
+▁quatrevingts 1
+▁lhistoire 1
+▁dal 1
+▁sommes 1
+▁regard 1
+▁bis 1
+▁lac 1
+▁gar 1
+▁dixhuit 1
+ix 1
+▁ve 1
+▁vos 1
+fa 1
+▁lou 1
+aires 1
+▁pé 1
+▁roman 1
+▁champ 1
+▁eux 1
+▁gé 1
+▁matin 1
+▁ten 1
+▁livre 1
+▁ber 1
+ttes 1
+▁pense 1
+▁eh 1
+gan 1
+▁peutêtre 1
+ud 1
+cher 1
+▁tri 1
+▁dieu 1
+car 1
+ad 1
+▁dès 1
+▁qua 1
+▁yeux 1
+▁lor 1
+istes 1
+ko 1
+▁seconde 1
+▁parfois 1
+▁surtout 1
+gre 1
+▁française 1
+▁bra 1
+tier 1
+isation 1
+ang 1
+mon 1
+▁nuit 1
+▁celui 1
+▁territoire 1
+tri 1
+▁boulevard 1
+▁système 1
+▁retire 1
+▁sol 1
+▁bord 1
+▁dou 1
+illes 1
+▁rouge 1
+▁musique 1
+▁succès 1
+ins 1
+ke 1
+isé 1
+▁tel 1
+▁années 1
+the 1
+sta 1
+▁participe 1
+erie 1
+ture 1
+▁utilisé 1
+▁marie 1
+men 1
+elles 1
+der 1
+▁chi 1
+nne 1
+▁belle 1
+zvous 1
+nu 1
+▁loin 1
+▁dép 1
+wa 1
+▁étatsunis 1
+▁nationale 1
+nie 1
+▁rou 1
+▁aurait 1
+▁nen 1
+▁rapport 1
+vre 1
+▁avaient 1
+bert 1
+illon 1
+nce 1
+▁sort 1
+▁parle 1
+▁nommé 1
+▁scène 1
+▁rôle 1
+▁blanc 1
+▁vient 1
+vais 1
+cy 1
+▁choses 1
+oir 1
+▁national 1
+▁veux 1
+▁ren 1
+▁charles 1
+▁hi 1
+ndre 1
+▁heure 1
+▁propose 1
+▁voiture 1
+igne 1
+port 1
+▁dixsept 1
+▁travers 1
+▁j 1
+▁quant 1
+bri 1
+▁létat 1
+▁attend 1
+▁jen 1
+lot 1
+▁lim 1
+▁pouvoir 1
+▁tombe 1
+▁public 1
+▁société 1
+▁maître 1
+▁joueur 1
+▁actuellement 1
+ain 1
+▁docteur 1
+▁rapidement 1
+▁pla 1
+▁toutefois 1
+▁quartier 1
+▁noir 1
+▁bâtiment 1
+▁har 1
+ore 1
+▁gre 1
+tal 1
+jo 1
+ques 1
+▁personnes 1
+▁f 1
+▁veut 1
+▁cause 1
+▁rend 1
+▁lon 1
+▁professeur 1
+▁lîle 1
+▁rivière 1
+▁permett 1
+née 1
+▁troisième 1
+lier 1
+▁juste 1
+▁prix 1
+▁pris 1
+fort 1
+ième 1
+▁ad 1
+▁dixneuf 1
+bou 1
+ologie 1
+pin 1
+van 1
+▁pied 1
+▁françois 1
+▁partir 1
+▁vingttrois 1
+▁chef 1
+▁u 1
+▁but 1
+▁dailleurs 1
+▁ab 1
+▁bal 1
+▁ko 1
+▁imp 1
+ien 1
+▁lis 1
+▁nu 1
+▁pri 1
+▁sainte 1
+▁aucune 1
+▁site 1
+resse 1
+▁mouvement 1
+▁milieu 1
+fer 1
+▁enfant 1
+ole 1
+gon 1
+▁longtemps 1
+▁certaines 1
+▁sait 1
+▁hommes 1
+moi 1
+▁montre 1
+voir 1
+▁dau 1
+ste 1
+com 1
+bourg 1
+▁lannée 1
+▁r 1
+▁autour 1
+▁commun 1
+jou 1
+id 1
+▁couleur 1
+▁quun 1
+▁marque 1
+▁appelé 1
+▁genre 1
+▁z 1
+î 1
+▁donné 1
+eurs 1
+▁article 1
+▁cap 1
+▁celuici 1
+pu 1
+▁vingtdeux 1
+▁environ 1
+▁corps 1
+▁lune 1
+▁parmi 1
+▁fe 1
+▁luimême 1
+▁simple 1
+▁sch 1
+enne 1
+▁besoin 1
+▁anglais 1
+▁cal 1
+▁combat 1
+ly 1
+▁chien 1
+▁femmes 1
+rd 1
+▁exemple 1
+▁doute 1
+nier 1
+▁pan 1
+ç 1
+▁vingtcinq 1
+▁pourtant 1
+▁can 1
+▁grands 1
+▁ki 1
+▁texte 1
+nou 1
+▁arrive 1
+▁vite 1
+tour 1
+▁rapporteur 1
+▁pratique 1
+▁meurt 1
+aine 1
+ges 1
+▁bureau 1
+am 1
+▁chacun 1
+▁vieux 1
+▁importante 1
+▁moyen 1
+ières 1
+▁langue 1
+▁fini 1
+▁type 1
+▁nétait 1
+he 1
+▁continue 1
+▁peine 1
+rin 1
+ju 1
+▁niveau 1
+▁canton 1
+▁débat 1
+▁omer 1
+▁suivante 1
+▁mondiale 1
+▁plait 1
+▁demanda 1
+▁parler 1
+sie 1
+▁possible 1
+▁trans 1
+lect 1
+dor 1
+▁vert 1
+▁pauvre 1
+ban 1
+▁demain 1
+▁parc 1
+▁perdu 1
+▁seulement 1
+▁musée 1
+▁bru 1
+▁zone 1
+qué 1
+▁fonction 1
+ral 1
+▁publique 1
+▁figure 1
+▁john 1
+ber 1
+ves 1
+ène 1
+cal 1
+▁période 1
+▁service 1
+iques 1
+▁bras 1
+▁semaine 1
+▁propre 1
+▁chant 1
+▁bout 1
+ieux 1
+alité 1
+▁jardin 1
+▁vin 1
+▁décide 1
+j 1
+▁certain 1
+▁eut 1
+▁date 1
+▁finalement 1
+ât 1
+stu 1
+▁recherche 1
+▁habitants 1
+▁allez 1
+▁min 1
+▁saisi 1
+ssé 1
+▁sta 1
+▁bre 1
+fe 1
+▁devenu 1
+▁capitaine 1
+▁riche 1
+▁longue 1
+▁dessin 1
+cent 1
+▁remporte 1
+▁laisse 1
+ster 1
+▁w 1
+▁to 1
+▁double 1
+lla 1
+▁militaire 1
+naire 1
+▁froid 1
+raient 1
+▁domaine 1
+▁san 1
+▁formation 1
+▁voit 1
+éré 1
+▁vingtsix 1
+▁principale 1
+▁cré 1
+▁ge 1
+bon 1
+▁quune 1
+▁reçu 1
+▁camp 1
+▁fl 1
+io 1
+▁principalement 1
+▁jaune 1
+▁ferme 1
+▁vingtquatre 1
+bel 1
+▁cheflieu 1
+ines 1
+▁passage 1
+▁cœur 1
+gent 1
+▁passer 1
+tel 1
+ny 1
+▁deau 1
+▁difficile 1
+▁province 1
+▁app 1
+▁for 1
+mme 1
+▁propos 1
+▁nommée 1
+form 1
+▁op 1
+▁ceux 1
+▁mission 1
+▁n 1
+▁thé 1
+leur 1
+▁v 1
+▁perd 1
+ay 1
+ule 1
+▁henri 1
+▁juge 1
+▁équipe 1
+▁lh 1
+tan 1
+▁généralement 1
+▁représente 1
+▁vingtsept 1
+non 1
+▁batt 1
+illé 1
+▁fre 1
+▁théâtre 1
+▁venir 1
+▁fr 1
+▁second 1
+nay 1
+mie 1
+▁style 1
+▁situe 1
+▁vé 1
+▁sal 1
+▁voyage 1
+▁championnat 1
+▁hier 1
+▁roche 1
+▁révolution 1
+▁gare 1
+▁feu 1
+teurs 1
+▁retrouve 1
+▁situation 1
+▁ailleurs 1
+▁plupart 1
+nelle 1
+▁faites 1
+ret 1
+▁lhonneur 1
+▁étudie 1
+▁sub 1
+ssent 1
+▁sc 1
+▁sp 1
+ade 1
+▁gou 1
+▁meilleur 1
+gn 1
+▁habit 1
+spir 1
+gh 1
+ura 1
+▁comporte 1
+▁new 1
+mor 1
+▁parents 1
+▁identiques 1
+rée 1
+vou 1
+court 1
+▁choisi 1
+▁lettre 1
+▁passé 1
+ric 1
+▁hé 1
+▁gui 1
+▁trentetrois 1
+ji 1
+saint 1
+▁entend 1
+▁vingthuit 1
+cor 1
+▁proche 1
+quelle 1
+▁larmée 1
+telle 1
+▁presque 1
+▁donner 1
+▁fla 1
+ras 1
+▁capitale 1
+up 1
+▁peux 1
+▁ensemble 1
+▁album 1
+▁mettre 1
+▁terme 1
+ît 1
+vis 1
+▁vingtneuf 1
+zi 1
+▁ob 1
+xi 1
+▁ph 1
+▁lotissement 1
+▁côte 1
+▁jou 1
+▁peur 1
+wi 1
+ages 1
+▁favorable 1
+mou 1
+▁lair 1
+▁navait 1
+▁journal 1
+eaux 1
+▁nécessaire 1
+▁pouvait 1
+▁merci 1
+▁chanson 1
+quet 1
+▁telle 1
+▁ancienne 1
+▁population 1
+▁large 1
+lais 1
+▁salle 1
+▁membres 1
+iens 1
+▁am 1
+▁historique 1
+▁lépoque 1
+▁libre 1
+▁match 1
+our 1
+▁all 1
+▁nouvelles 1
+yer 1
+ches 1
+▁marche 1
+ty 1
+▁lensemble 1
+char 1
+étant 1
+▁court 1
+▁vraiment 1
+▁organisé 1
+▁livres 1
+rre 1
+▁bel 1
+vons 1
+vint 1
+▁différents 1
+▁façon 1
+▁sac 1
+▁philippe 1
+duc 1
+mit 1
+ères 1
+col 1
+ales 1
+▁néanmoins 1
+▁retour 1
+oire 1
+din 1
+can 1
+ros 1
+▁lheure 1
+▁cho 1
+▁aura 1
+▁haute 1
+▁déco 1
+▁vois 1
+lis 1
+gie 1
+tions 1
+▁fenêtre 1
+▁plutôt 1
+▁trentequatre 1
+▁trentedeux 1
+▁député 1
+▁croix 1
+iser 1
+▁ancien 1
+▁michel 1
+▁manière 1
+sson 1
+▁signe 1
+▁cimetière 1
+▁super 1
+▁anti 1
+▁lex 1
+ice 1
+▁source 1
+▁trentecinq 1
+▁petits 1
+▁tiens 1
+▁qu 1
+ani 1
+▁jeunes 1
+▁principal 1
+▁communes 1
+▁len 1
+▁ter 1
+den 1
+berg 1
+venir 1
+▁trouver 1
+▁endémique 1
+▁visite 1
+▁dame 1
+▁déc 1
+▁charge 1
+▁savoir 1
+hon 1
+▁position 1
+▁état 1
+tti 1
+▁appel 1
+▁h 1
+▁différentes 1
+▁originaire 1
+▁travaux 1
+res 1
+▁premiers 1
+art 1
+▁rejoint 1
+▁production 1
+▁bleu 1
+▁etait 1
+gen 1
+▁ouvert 1
+tz 1
+bes 1
+▁mains 1
+▁force 1
+▁cru 1
+▁allons 1
+èle 1
+▁pal 1
+han 1
+▁idée 1
+hé 1
+net 1
+▁termine 1
+▁division 1
+igné 1
+▁créé 1
+▁peuple 1
+▁breton 1
+▁forte 1
+▁lit 1
+▁connue 1
+▁comte 1
+ria 1
+▁bro 1
+▁œuvres 1
+▁her 1
+▁voisin 1
+gène 1
+spect 1
+près 1
+▁station 1
+nant 1
+▁compétition 1
+▁art 1
+▁im 1
+▁lai 1
+▁dispose 1
+▁dorigine 1
+ica 1
+▁leau 1
+▁malgré 1
+mba 1
+lement 1
+í 1
+tiques 1
+▁évolue 1
+▁robert 1
+▁version 1
+▁jacques 1
+iv 1
+vie 1
+▁vend 1
+▁suivi 1
+▁fau 1
+tor 1
+▁construit 1
+▁pen 1
+▁fil 1
+▁désormais 1
+▁élu 1
+▁cé 1
+▁celleci 1
+▁maisons 1
+▁vue 1
+▁inf 1
+▁che 1
+▁four 1
+▁martin 1
+▁inscrit 1
+roi 1
+▁gros 1
+▁victoire 1
+▁directeur 1
+logique 1
+ète 1
+hu 1
+▁ram 1
+▁vont 1
+bit 1
+pt 1
+position 1
+▁quarantecinq 1
+▁détat 1
+▁dépend 1
+▁malade 1
+▁péri 1
+▁descend 1
+ü 1
+▁cinéma 1
+▁construction 1
+▁dh 1
+▁soi 1
+▁don 1
+▁pêche 1
+▁issu 1
+▁chapelle 1
+▁base 1
+▁lordre 1
+nel 1
+igny 1
+▁frères 1
+ina 1
+▁pres 1
+dit 1
+▁wa 1
+▁clair 1
+▁marché 1
+▁particulier 1
+▁contraire 1
+▁rare 1
+▁naît 1
+▁considéré 1
+▁bataille 1
+range 1
+▁mai 1
+▁chaud 1
+chant 1
+gence 1
+ab 1
+▁carte 1
+▁allemand 1
+▁jaime 1
+tho 1
+▁jeux 1
+▁histoire 1
+▁publié 1
+▁blanche 1
+▁plante 1
+▁finale 1
+▁avezvous 1
+▁direction 1
+▁arrivé 1
+▁mariage 1
+we 1
+tive 1
+érie 1
+▁communauté 1
+iez 1
+▁critique 1
+▁suivant 1
+▁bande 1
+▁ban 1
+▁rose 1
+▁particulière 1
+▁classe 1
+▁sein 1
+▁dentre 1
+dy 1
+▁reprend 1
+▁louest 1
+▁disposition 1
+pri 1
+▁my 1
+▁manque 1
+▁rep 1
+del 1
+▁ben 1
+▁liste 1
+▁utilisée 1
+▁médecin 1
+▁mange 1
+▁forêt 1
+udi 1
+▁marc 1
+cro 1
+ley 1
+rac 1
+▁grandes 1
+▁bla 1
+▁blo 1
+▁campagne 1
+▁utilise 1
+▁garçon 1
+ça 1
+isse 1
+issant 1
+▁petites 1
+ables 1
+fin 1
+▁correspond 1
+▁grâce 1
+gar 1
+▁ke 1
+ié 1
+ayant 1
+ève 1
+ult 1
+tic 1
+▁van 1
+▁gris 1
+▁rec 1
+nière 1
+▁lancien 1
+▁lalbum 1
+▁église 1
+▁voici 1
+▁quarantehuit 1
+▁transport 1
+à 1
+▁éc 1
+isée 1
+▁cherche 1
+▁cheval 1
+▁dr 1
+▁cinquantetrois 1
+▁reçoit 1
+▁américain 1
+▁vis 1
+▁lorigine 1
+▁époque 1
+▁collection 1
+▁voulu 1
+▁rendre 1
+▁prince 1
+▁vieille 1
+▁marquis 1
+▁école 1
+uel 1
+▁outre 1
+cour 1
+▁ditil 1
+▁souhaite 1
+▁proposition 1
+▁champion 1
+▁cre 1
+▁plein 1
+▁constitue 1
+▁république 1
+▁trentesept 1
+▁fête 1
+▁commande 1
+▁joseph 1
+nais 1
+▁sang 1
+▁suisse 1
+▁sor 1
+▁raisons 1
+▁répond 1
+▁làbas 1
+▁clos 1
+▁noire 1
+vel 1
+aille 1
+▁secrétaire 1
+sol 1
+▁rest 1
+oul 1
+▁divers 1
+▁économique 1
+▁régulièrement 1
+▁viens 1
+▁tor 1
+zo 1
+▁mise 1
+lev 1
+œ 1
+▁lhomme 1
+▁pose 1
+▁pouss 1
+▁seront 1
+▁devait 1
+á 1
+ède 1
+▁victor 1
+▁connais 1
+▁albert 1
+lam 1
+ham 1
+chet 1
+by 1
+▁dy 1
+▁entendu 1
+▁cinquantequatre 1
+▁faible 1
+▁note 1
+▁trentehuit 1
+▁andré 1
+▁comté 1
+▁sétait 1
+▁jouer 1
+▁rem 1
+▁sin 1
+▁fé 1
+▁front 1
+▁catholique 1
+▁cinquantecinq 1
+▁lauteur 1
+▁peintre 1
+▁respect 1
+vet 1
+▁aide 1
+▁content 1
+▁ven 1
+▁élevé 1
+ût 1
+gna 1
+my 1
+▁œuvre 1
+▁prie 1
+▁danger 1
+▁mau 1
+▁pou 1
+▁extrême 1
+▁remplacé 1
+▁élève 1
+▁référence 1
+▁modèle 1
+▁régime 1
+cation 1
+éra 1
+▁maire 1
+▁obtien 1
+quà 1
+od 1
+▁trenteneuf 1
+▁fonctionne 1
+▁cité 1
+▁développement 1
+▁wi 1
+▁gagne 1
+▁programme 1
+▁quarantequatre 1
+▁york 1
+oc 1
+▁couple 1
+shi 1
+▁course 1
+dan 1
+▁relation 1
+▁reven 1
+▁rang 1
+▁autant 1
+vit 1
+▁lorsque 1
+▁composé 1
+▁jétais 1
+▁mademoiselle 1
+▁quelquun 1
+▁soutien 1
+▁monte 1
+euses 1
+▁cinquantedeux 1
+▁dirige 1
+▁réalité 1
+▁tre 1
+ane 1
+▁taille 1
+▁allait 1
+▁épouse 1
+▁trouvé 1
+▁dac 1
+▁coupe 1
+▁international 1
+▁radio 1
+▁technique 1
+▁majorité 1
+▁limite 1
+▁regarde 1
+anne 1
+▁rester 1
+tori 1
+pro 1
+▁jy 1
+▁montagne 1
+▁servi 1
+▁espèces 1
+▁suit 1
+▁paraît 1
+▁quarantesix 1
+▁présence 1
+lès 1
+bra 1
+nois 1
+gno 1
+▁peinture 1
+heim 1
+gard 1
+bar 1
+▁inc 1
+▁serai 1
+▁bientôt 1
+▁compagnie 1
+▁réponse 1
+▁terrain 1
+▁silence 1
+▁emp 1
+▁arriv 1
+▁tou 1
+▁the 1
+▁travailler 1
+▁réussi 1
+▁déroule 1
+▁enterré 1
+pen 1
+ian 1
+gal 1
+rique 1
+▁prise 1
+▁contribu 1
+▁lintérieur 1
+▁largent 1
+▁auto 1
+▁pat 1
+▁ceci 1
+▁appartient 1
+▁lacadémie 1
+▁voulez 1
+▁gri 1
+▁observ 1
+ition 1
+▁départ 1
+▁savez 1
+▁minutes 1
+▁profit 1
+▁quitte 1
+ification 1
+▁davoir 1
+▁développe 1
+▁salon 1
+▁tir 1
+▁vaut 1
+ow 1
+▁soixantedouze 1
+▁édition 1
+▁londres 1
+▁soixantedixhuit 1
+▁dha 1
+▁fri 1
+antes 1
+▁gla 1
+▁amis 1
+▁fran 1
+▁assis 1
+▁occupe 1
+▁el 1
+▁cinquantesept 1
+▁quarantesept 1
+▁soixantecinq 1
+▁âge 1
+▁remporté 1
+▁manger 1
+dra 1
+lay 1
+▁répondit 1
+nal 1
+table 1
+▁affaires 1
+ulaire 1
+cle 1
+▁siècle 1
+▁trou 1
+▁contrôle 1
+riez 1
+nor 1
+thé 1
+▁santé 1
+▁retourne 1
+▁partage 1
+▁sert 1
+ub 1
+▁précédent 1
+▁doivent 1
+▁jusquau 1
+▁trentesix 1
+▁pre 1
+▁professionnel 1
+▁rome 1
+▁traverse 1
+▁commercial 1
+▁total 1
+▁placé 1
+▁parce 1
+gnon 1
+▁emb 1
+▁réseau 1
+▁dargent 1
+▁réalisé 1
+▁troupe 1
+tine 1
+itude 1
+▁grave 1
+tar 1
+▁bour 1
+▁tro 1
+▁we 1
+gère 1
+ov 1
+nage 1
+▁basse 1
+▁comb 1
+▁sorte 1
+▁activité 1
+▁presse 1
+▁café 1
+▁soleil 1
+q 1
+▁déli 1
+▁pièces 1
+ssou 1
+▁rive 1
+▁vise 1
+▁derrière 1
+▁structure 1
+▁défend 1
+▁ordre 1
+▁dos 1
+▁far 1
+▁prêt 1
+▁joué 1
+▁points 1
+tré 1
+▁georges 1
+▁défendu 1
+▁êtesvous 1
+nom 1
+▁hor 1
+▁com 1
+▁han 1
+ures 1
+▁rentre 1
+▁pin 1
+pit 1
+▁apporte 1
+▁lance 1
+▁air 1
+▁tend 1
+▁confi 1
+▁ste 1
+▁soixantequinze 1
+▁reconnu 1
+▁instant 1
+▁quai 1
+▁pouvez 1
+▁verre 1
+▁films 1
+▁russe 1
+▁tient 1
+▁quart 1
+quer 1
+▁établi 1
+▁dispositif 1
+▁lentreprise 1
+▁téléphone 1
+▁lobjet 1
+▁document 1
+▁résidence 1
+▁choix 1
+▁entièrement 1
+viennent 1
+▁lyon 1
+▁sélection 1
+ème 1
+met 1
+▁nont 1
+ssant 1
+nard 1
+▁gro 1
+▁brun 1
+irent 1
+illy 1
+▁revenu 1
+éro 1
+▁domin 1
+▁apprend 1
+▁composée 1
+▁lendemain 1
+ault 1
+▁marié 1
+▁réalise 1
+▁viendra 1
+▁représentant 1
+cho 1
+line 1
+▁collègues 1
+ivité 1
+ok 1
+▁pourrait 1
+▁page 1
+ak 1
+▁fra 1
+▁nature 1
+▁tourne 1
+▁japon 1
+▁développé 1
+▁populaire 1
+▁lettres 1
+ange 1
+▁dart 1
+▁télé 1
+▁daller 1
+▁nor 1
+▁laide 1
+agne 1
+▁chance 1
+ssait 1
+cep 1
+▁difficulté 1
+▁objet 1
+▁commerce 1
+▁os 1
+▁conte 1
+▁java 1
+gou 1
+▁essay 1
+▁dites 1
+ö 1
+▁sinstalle 1
+vaient 1
+▁étudiant 1
+▁ed 1
+▁création 1
+▁personnage 1
+▁entra 1
+▁pra 1
+▁larrondissement 1
+▁lempereur 1
+ana 1
+▁monument 1
+▁rappel 1
+vait 1
+▁social 1
+▁fallait 1
+▁internationale 1
+▁anne 1
+ncer 1
+ari 1
+▁mit 1
+oli 1
+gues 1
+▁plu 1
+▁équipes 1
+▁refus 1
+▁retirer 1
+▁blessé 1
+▁rené 1
+▁plat 1
+oise 1
+▁tom 1
+▁fermé 1
+▁reprises 1
+▁culture 1
+ative 1
+père 1
+mier 1
+ó 1
+▁constitué 1
+▁véritable 1
+▁maladie 1
+▁réserve 1
+▁labbé 1
+▁cinquantehuit 1
+mat 1
+▁probablement 1
+▁gen 1
+▁pain 1
+bois 1
+▁connaît 1
+mporte 1
+▁code 1
+▁célèbre 1
+▁principe 1
+▁quaranteneuf 1
+▁derniers 1
+▁villa 1
+▁refuse 1
+▁stade 1
+ray 1
+▁remercie 1
+▁science 1
+▁contrat 1
+▁accord 1
+ngage 1
+▁reprit 1
+ini 1
+▁présenté 1
+▁mauvais 1
+▁consul 1
+lia 1
+▁pièce 1
+vezvous 1
+▁prison 1
+▁nav 1
+▁district 1
+▁parallèle 1
+▁soixantetrois 1
+▁concert 1
+▁défini 1
+▁vote 1
+▁football 1
+lly 1
+▁sortir 1
+▁fleur 1
+▁utilis 1
+▁centrale 1
+▁ouverte 1
+ute 1
+▁quatrevingtcinq 1
+▁décision 1
+▁quarantedeux 1
+chon 1
+▁ange 1
+▁fourni 1
+issait 1
+▁quarantetrois 1
+▁quatrième 1
+▁atteint 1
+▁poisson 1
+▁classique 1
+▁vérité 1
+▁vallée 1
+▁questions 1
+▁risque 1
+▁fur 1
+▁douce 1
+▁dam 1
+▁filles 1
+▁britannique 1
+▁daccord 1
+▁soixantesept 1
+▁lenfant 1
+▁enseigne 1
+trice 1
+nge 1
+▁luc 1
+▁locale 1
+âtre 1
+▁espagnol 1
+▁spectacle 1
+▁urbain 1
+▁remplace 1
+▁soixantesix 1
+▁métier 1
+▁té 1
+▁finit 1
+▁chasse 1
+sible 1
+▁change 1
+assi 1
+ker 1
+▁claire 1
+oud 1
+▁magasin 1
+laient 1
+▁dominique 1
+▁associé 1
+ë 1
+▁bruit 1
+▁mill 1
+▁européen 1
+▁chargé 1
+▁disparu 1
+▁bancs 1
+▁demeure 1
+▁quen 1
+▁tenu 1
+▁représenté 1
+▁moulin 1
+nique 1
+▁cli 1
+▁bouche 1
+▁auteur 1
+uf 1
+bonne 1
+▁origine 1
+▁conserve 1
+▁wal 1
+▁conseiller 1
+▁marcel 1
+▁arrête 1
+▁tué 1
+dia 1
+▁présidente 1
+▁boul 1
+▁soixantequatre 1
+▁étrange 1
+▁cycl 1
+crit 1
+▁pot 1
+▁dautre 1
+▁cau 1
+▁poly 1
+▁lien 1
+zy 1
+▁soixantedeux 1
+▁papier 1
+▁clé 1
+▁soixantedixneuf 1
+▁sociale 1
+▁sport 1
+mir 1
+▁objectif 1
+▁occasion 1
+▁satisfait 1
+▁daltitude 1
+▁argent 1
+▁revient 1
+▁hum 1
+fo 1
+▁marin 1
+ographie 1
+▁acte 1
+marie 1
+▁soixanteseize 1
+▁méthode 1
+▁lycée 1
+issement 1
+▁police 1
+▁patron 1
+▁lourd 1
+▁seuls 1
+rêt 1
+dal 1
+▁basé 1
+▁caractère 1
+▁quatrevingtneuf 1
+▁quatrevingthuit 1
+▁canada 1
+▁dehors 1
+▁soixantehuit 1
+▁navire 1
+▁bernard 1
+▁courant 1
+▁bus 1
+tha 1
+▁bourg 1
+▁tournée 1
+▁ciel 1
+▁retrait 1
+atrice 1
+▁tente 1
+▁tar 1
+ach 1
+▁gars 1
+donne 1
+▁resta 1
+▁dû 1
+▁profond 1
+pon 1
+soixantequatorze 1
+▁lœuvre 1
+▁voulait 1
+▁saut 1
+▁resté 1
+▁réuni 1
+▁dure 1
+vert 1
+roy 1
+cide 1
+▁événement 1
+▁relative 1
+▁cinquanteneuf 1
+▁îles 1
+▁post 1
+▁sha 1
+▁pouv 1
+mètre 1
+▁collège 1
+▁nicolas 1
+▁david 1
+▁armée 1
+▁interdit 1
+èvre 1
+▁entreprises 1
+bro 1
+ama 1
+vingtquatre 1
+▁langues 1
+▁laisser 1
+bat 1
+sous 1
+▁fatigué 1
+▁propriétaire 1
+▁propriété 1
+▁lhôtel 1
+▁solution 1
+▁branche 1
+▁feuilles 1
+▁localité 1
+▁journée 1
+▁tru 1
+▁visage 1
+toi 1
+▁actuel 1
+▁seigneur 1
+▁coll 1
+▁proximité 1
+quatrevingttrois 1
+▁emploi 1
+▁vivant 1
+▁form 1
+▁gal 1
+▁épi 1
+nat 1
+yo 1
+▁hy 1
+▁nouveaux 1
+▁faudra 1
+vu 1
+▁ger 1
+▁mod 1
+▁régional 1
+▁qualité 1
+▁partout 1
+▁classé 1
+ezmoi 1
+pression 1
+▁concerne 1
+▁ouvre 1
+▁foi 1
+ling 1
+▁repose 1
+▁collègue 1
+▁fera 1
+▁audessus 1
+▁façade 1
+▁conduit 1
+▁connaissance 1
+▁ly 1
+▁noble 1
+▁vienne 1
+▁ham 1
+▁eff 1
+▁matchs 1
+▁souhait 1
+▁ty 1
+▁grosse 1
+gée 1
+▁quatrevingtdeux 1
+▁europe 1
+▁dirigé 1
+▁jaimerais 1
+▁quatrevingtun 1
+▁demandé 1
+▁prochain 1
+▁résultats 1
+▁conservé 1
+▁comité 1
+▁mine 1
+▁supprimer 1
+▁chapeau 1
+▁somme 1
+▁hameau 1
+▁cadre 1
+tou 1
+▁dia 1
+ō 1
+▁responsable 1
+▁scientifique 1
+▁support 1
+▁officier 1
+▁preuve 1
+▁liberté 1
+lation 1
+▁poursuit 1
+▁chat 1
+▁soixanteneuf 1
+▁moderne 1
+▁coût 1
+▁passion 1
+▁travaillé 1
+▁savait 1
+▁continu 1
+▁cathédrale 1
+▁exécut 1
+▁préfér 1
+▁laissé 1
+thique 1
+▁devenir 1
+▁accepte 1
+▁mil 1
+cre 1
+▁kar 1
+euil 1
+▁désir 1
+▁and 1
+zon 1
+▁physique 1
+▁quatrevingtonze 1
+▁record 1
+▁inhumé 1
+▁télévision 1
+▁essentiellement 1
+▁cinquantesix 1
+▁sombre 1
+▁daniel 1
+▁viennent 1
+rich 1
+▁portrait 1
+▁chaises 1
+old 1
+▁proposé 1
+▁test 1
+▁changement 1
+▁compren 1
+mment 1
+▁diverses 1
+▁lutte 1
+▁générale 1
+▁rh 1
+▁sommet 1
+▁animaux 1
+▁suppression 1
+▁favori 1
+▁richard 1
+atoire 1
+▁parlé 1
+mal 1
+▁frontière 1
+▁prit 1
+▁lart 1
+part 1
+▁formé 1
+▁repris 1
+▁pol 1
+▁tort 1
+▁flam 1
+vingtquatorze 1
+▁festival 1
+▁inconnu 1
+▁lâge 1
+▁municipalité 1
+▁daprès 1
+▁soldats 1
+ano 1
+▁x 1
+▁flo 1
+▁moyenne 1
+▁rêve 1
+▁anciens 1
+▁appelle 1
+▁mode 1
+▁arrêté 1
+▁soixantetreize 1
+▁surface 1
+▁conditions 1
+▁secteur 1
+matique 1
+▁tournoi 1
+▁venait 1
+▁lavait 1
+▁chansons 1
+pha 1
+ssement 1
+gui 1
+▁vendu 1
+▁services 1
+▁conçu 1
+▁approche 1
+▁prévu 1
+▁hou 1
+▁cache 1
+▁glace 1
+▁guéri 1
+▁rappelle 1
+▁ouvrages 1
+▁baron 1
+▁règle 1
+out 1
+▁disponible 1
+▁dispute 1
+▁statue 1
+▁julien 1
+▁inclu 1
+▁za 1
+▁bio 1
+her 1
+stro 1
+len 1
+ignac 1
+▁décrit 1
+▁fusil 1
+▁citoyen 1
+▁sousamendement 1
+vingtquinze 1
+▁belgique 1
+quatrevingtdouze 1
+▁conclu 1
+▁sétend 1
+▁machine 1
+▁remonte 1
+▁fontaine 1
+dire 1
+oux 1
+▁construite 1
+▁direct 1
+fu 1
+▁baie 1
+▁comprendre 1
+▁outil 1
+▁guillaume 1
+▁installé 1
+▁gabriel 1
+▁offre 1
+▁étage 1
+▁voulezvous 1
+▁piano 1
+▁assure 1
+▁puisque 1
+roux 1
+▁tableau 1
+particulièrement 1
+ération 1
+▁romain 1
+▁disque 1
+▁lancienne 1
+▁forces 1
+▁personnages 1
+▁accept 1
+▁faute 1
+▁dois 1
+▁canadien 1
+▁attaque 1
+▁cloche 1
+▁davantage 1
+▁mono 1
+▁visible 1
+iel 1
+kh 1
+gin 1
+▁séance 1
+▁rendu 1
+▁facile 1
+▁sortie 1
+▁artistes 1
+▁paroisse 1
+▁retiré 1
+▁repondit 1
+▁quatrevingtsix 1
+▁frédéric 1
+▁compris 1
+▁berlin 1
+▁personnel 1
+▁pleine 1
+▁cousin 1
+▁retraite 1
+▁banque 1
+▁agricole 1
+▁détruit 1
+▁contient 1
+▁marchand 1
+▁immédiatement 1
+igni 1
+llier 1
+▁jolie 1
+▁fondateur 1
+▁laurent 1
+▁royal 1
+▁fleurs 1
+▁agr 1
+hen 1
+▁employé 1
+▁disait 1
+▁local 1
+▁million 1
+▁largement 1
+▁bau 1
+vol 1
+▁née 1
+▁parlement 1
+lard 1
+▁indépendant 1
+▁composition 1
+▁trompe 1
+▁italien 1
+▁major 1
+▁découverte 1
+iège 1
+mbre 1
+ignan 1
+ologue 1
+▁grandi 1
+▁comm 1
+fon 1
+▁québec 1
+▁thomas 1
+▁victime 1
+▁théorie 1
+▁supérieur 1
+▁climat 1
+zmoi 1
+▁prêtre 1
+▁joli 1
+▁troubl 1
+▁labbaye 1
+▁principaux 1
+▁chiffre 1
+▁activités 1
+▁maintien 1
+▁allé 1
+▁japonais 1
+▁découvre 1
+▁ceuxci 1
+▁francis 1
+roche 1
+ku 1
+long 1
+abilité 1
+▁lassociation 1
+▁lépreuve 1
+▁simpose 1
+▁plaisir 1
+▁naissance 1
+▁raconte 1
+▁parfaitement 1
+▁olympique 1
+▁rejet 1
+▁marqué 1
+▁détail 1
+▁palais 1
+▁calme 1
+▁thi 1
+▁lire 1
+▁peint 1
+lette 1
+▁donna 1
+▁lucien 1
+▁unique 1
+▁diplômé 1
+▁sécria 1
+▁invité 1
+▁lidée 1
+▁abrite 1
+▁pilote 1
+▁fixe 1
+▁considérée 1
+dol 1
+▁européenne 1
+▁revenir 1
+usement 1
+▁couche 1
+▁pensa 1
+géné 1
+▁apparaît 1
+▁protestant 1
+▁dhonneur 1
+▁moitié 1
+▁voudrais 1
+▁rendezvous 1
+▁accueille 1
+▁heureux 1
+▁chaise 1
+▁linstitut 1
+être 1
+▁linter 1
+▁mauvaise 1
+▁franc 1
+lav 1
+mond 1
+▁valeur 1
+▁bibliothèque 1
+▁discussion 1
+▁lassemblée 1
+▁lumière 1
+▁goût 1
+▁calcul 1
+▁monta 1
+▁voter 1
+▁lamour 1
+▁hors 1
+▁atil 1
+▁meilleure 1
+ney 1
+▁rapporteure 1
+▁moteur 1
+▁rapide 1
+cré 1
+▁réside 1
+▁brûl 1
+▁justice 1
+▁évidemment 1
+▁domestique 1
+▁revue 1
+oï 1
+▁prisonnier 1
+▁normal 1
+▁emm 1
+▁données 1
+▁éch 1
+▁futur 1
+▁passa 1
+▁puissant 1
+vingttreize 1
+▁endroit 1
+▁étudié 1
+▁quatrevingtsept 1
+▁christ 1
+▁remarque 1
+▁sudouest 1
+▁engagé 1
+▁hauteur 1
+▁opération 1
+▁nation 1
+ap 1
+▁tian 1
+▁destiné 1
+zan 1
+▁central 1
+▁naturel 1
+▁couvert 1
+champ 1
+▁allemagne 1
+▁bénéficie 1
+▁soixantedixsept 1
+▁consacré 1
+▁vacances 1
+▁studio 1
+▁marseille 1
+▁chaîne 1
+▁attention 1
+▁industriel 1
+▁thème 1
+▁médaille 1
+▁dépense 1
+▁gliss 1
+▁bapt 1
+ther 1
+ford 1
+▁courage 1
+▁essai 1
+▁classement 1
+neau 1
+▁excellent 1
+▁religieuse 1
+▁autrefois 1
+▁crayon 1
+▁trembl 1
+▁féminin 1
+▁explique 1
+▁acheté 1
+▁considérable 1
+▁confiance 1
+▁ubu 1
+▁affluent 1
+▁dédié 1
+▁tenir 1
+▁claude 1
+▁américaine 1
+▁cheminée 1
+▁th 1
+▁francs 1
+▁hon 1
+▁éléments 1
+▁complète 1
+▁musicien 1
+▁défense 1
+▁ligue 1
+▁bur 1
+▁coin 1
+nau 1
+gny 1
+où 1
+▁simplement 1
+▁nourri 1
+▁fortement 1
+▁différence 1
+▁original 1
+▁phénomène 1
+▁revanche 1
+▁bonheur 1
+▁saintjean 1
+▁dom 1
+▁leçon 1
+▁nou 1
+▁ker 1
+▁conten 1
+▁mini 1
+▁royaume 1
+▁californie 1
+▁conséquence 1
+▁traduit 1
+▁éléphants 1
+▁arbres 1
+▁matière 1
+▁réalisée 1
+▁mol 1
+▁belge 1
+▁relie 1
+▁chercher 1
+▁trouva 1
+▁papa 1
+▁réunion 1
+▁muni 1
+▁demander 1
+▁aimera 1
+▁déf 1
+▁défi 1
+▁chine 1
+tention 1
+▁tranquille 1
+▁étape 1
+▁constat 1
+▁jespère 1
+▁aima 1
+▁léon 1
+▁média 1
+▁traité 1
+▁piste 1
+▁attendre 1
+▁logique 1
+▁fruits 1
+▁instrument 1
+▁lindustrie 1
+▁préoccup 1
+▁obtenu 1
+▁envoyé 1
+▁tôt 1
+▁auguste 1
+▁sappelle 1
+▁succède 1
+▁fiscal 1
+▁lecture 1
+▁chinois 1
+▁vide 1
+▁poète 1
+▁zo 1
+▁affaire 1
+▁vivre 1
+▁changer 1
+▁touche 1
+▁journaliste 1
+▁sorti 1
+▁lieux 1
+▁cir 1
+▁via 1
+▁envie 1
+▁occupé 1
+▁mich 1
+▁simon 1
+▁tradition 1
+▁candidat 1
+▁résistance 1
+▁manifest 1
+▁dimanche 1
+▁séparé 1
+▁exact 1
+▁colonel 1
+▁prénom 1
+▁vécu 1
+étaient 1
+▁antoine 1
+ppo 1
+▁uniquement 1
+▁ku 1
+▁avions 1
+▁poule 1
+▁parlez 1
+terre 1
+▁écout 1
+▁remis 1
+▁coûte 1
+▁caractéristique 1
+▁supplémentaire 1
+▁distingue 1
+▁restaurant 1
+▁langlais 1
+▁terminé 1
+▁commissaire 1
+▁réforme 1
+▁retard 1
+▁secret 1
+▁nordouest 1
+hérit 1
+▁directement 1
+▁précision 1
+mbra 1
+nni 1
+▁trouvait 1
+▁cur 1
+▁démon 1
+cueil 1
+▁tai 1
+▁carré 1
+▁trait 1
+▁midi 1
+▁littérature 1
+▁électrique 1
+▁cabinet 1
+▁matériel 1
+▁diminu 1
+▁âme 1
+▁épaule 1
+▁photo 1
+▁reine 1
+▁neveu 1
+nçon 1
+bb 1
+▁pren 1
+▁italie 1
+▁servir 1
+▁diocèse 1
+▁exploit 1
+▁loccasion 1
+▁détermin 1
+▁défaut 1
+▁maurice 1
+▁consomm 1
+▁lempire 1
+▁récit 1
+▁aliment 1
+▁section 1
+géni 1
+▁guy 1
+▁chevalier 1
+▁chante 1
+gli 1
+▁orient 1
+▁ann 1
+▁éli 1
+ctor 1
+▁ruelle 1
+▁ministère 1
+▁épisode 1
+▁voul 1
+▁tâche 1
+▁rédactionnel 1
+▁brésil 1
+▁chauss 1
+▁totalement 1
+▁march 1
+▁client 1
+▁complètement 1
+▁lunion 1
+▁privée 1
+▁privé 1
+▁tal 1
+▁poche 1
+wood 1
+▁flori 1
+▁sécurité 1
+▁auparavant 1
+▁cheveux 1
+▁australie 1
+obtenir 1
+▁contact 1
+▁information 1
+gg 1
+quel 1
+▁pouvons 1
+▁quitter 1
+▁trace 1
+prime 1
+▁réalis 1
+▁trouvent 1
+▁frais 1
+▁abandonné 1
+▁intérêt 1
+▁malheureusement 1
+▁doigt 1
+▁prétend 1
+▁actif 1
+▁argument 1
+▁faudrait 1
+▁folle 1
+▁manche 1
+▁tire 1
+▁rédaction 1
+▁monastère 1
+▁témoign 1
+▁phase 1
+▁hasard 1
+▁lentrée 1
+▁préfère 1
+lage 1
+▁bay 1
+▁saintpierre 1
+▁absolument 1
+▁souvenir 1
+▁berger 1
+▁cata 1
+▁barre 1
+▁secondaire 1
+voyait 1
+▁vaste 1
+▁jaurais 1
+▁inde 1
+▁jon 1
+▁venez 1
+▁vive 1
+▁continua 1
+zin 1
+▁chen 1
+▁reconnais 1
+ä 1
+▁chrétien 1
+▁légèrement 1
+▁contraint 1
+▁roul 1
+▁regret 1
+▁loup 1
+▁payer 1
+▁guère 1
+▁médecine 1
+▁parlementaire 1
+envoie 1
+▁lequel 1
+▁qualifi 1
+▁centaine 1
+lio 1
+▁juridique 1
+▁logiciel 1
+▁indique 1
+cade 1
+▁désigne 1
+▁écrivain 1
+▁pensée 1
+▁régl 1
+▁naturelle 1
+▁connaître 1
+▁concern 1
+▁compose 1
+▁précise 1
+▁poursuiv 1
+▁découvert 1
+applaudissement 1
+▁augmente 1
+▁desservi 1
+icité 1
+▁bassin 1
+▁publication 1
+▁désert 1
+▁rouss 1
+▁colonne 1
+hum 1
+▁jouait 1
+▁entreprise 1
+▁nomme 1
+▁lob 1
+otte 1
+▁durée 1
+mann 1
+▁put 1
+▁apparten 1
+▁fleuve 1
+▁impossible 1
+▁philosophie 1
+▁règne 1
+▁soccupe 1
+▁cuisine 1
+▁volonté 1
+▁dizaine 1
+▁mourir 1
+▁erreur 1
+▁celuilà 1
+▁acheter 1
+▁annonce 1
+▁crise 1
+▁tim 1
+▁grandit 1
+▁barbe 1
+▁hall 1
+beau 1
+▁robe 1
+▁vidéo 1
+▁rectifié 1
+▁rejoindre 1
+▁budget 1
+▁voyez 1
+▁tré 1
+▁sudest 1
+▁mandat 1
+▁tombé 1
+▁convention 1
+▁éclair 1
+▁boutique 1
+▁solide 1
+▁pomme 1
+▁wo 1
+▁colère 1
+▁grotte 1
+▁ira 1
+▁retrouver 1
+▁tomba 1
+cq 1
+▁armes 1
+▁posé 1
+▁casse 1
+▁portant 1
+▁baisse 1
+▁participa 1
+▁william 1
+▁avocat 1
+▁élection 1
+▁administrative 1
+▁laffaire 1
+▁exposé 1
+▁humain 1
+▁neige 1
+▁gilles 1
+▁sentiment 1
+▁influence 1
+paul 1
+chang 1
+▁montée 1
+▁joie 1
+▁mène 1
+mporta 1
+▁répondre 1
+▁remet 1
+▁color 1
+▁jeunesse 1
+▁léconomie 1
+▁sauvage 1
+▁mémoire 1
+▁phrase 1
+▁conserva 1
+▁nexiste 1
+▁signal 1
+▁ajoute 1
+▁dici 1
+▁manqu 1
+▁traditionnelle 1
+▁gaz 1
+▁plateau 1
+▁larmes 1
+▁devrait 1
+▁paire 1
+endra 1
+▁spéciale 1
+▁décor 1
+▁cria 1
+û 1
+▁paysan 1
+▁taux 1
+▁allemande 1
+▁complexe 1
+▁intérieur 1
+▁procédure 1
+▁reconnaît 1
+▁équipé 1
+▁rapproch 1
+▁affirm 1
+▁animal 1
+▁cultiv 1
+▁réduit 1
+▁sourire 1
+▁discut 1
+▁fauteuil 1
+▁écrire 1
+▁surprise 1
+▁eût 1
+▁entouré 1
+pol 1
+▁récent 1
+quin 1
+▁phil 1
+▁poli 1
+michel 1
+▁compos 1
+▁év 1
+▁prévoi 1
+▁évolu 1
+neuve 1
+mère 1
+▁résultat 1
+▁conséquent 1
+▁quotidien 1
+▁effectue 1
+▁producteur 1
+▁triomph 1
+▁espace 1
+▁raymond 1
+▁ouvrir 1
+▁souverain 1
+▁ignor 1
+▁exerce 1
+▁changé 1
+▁esprit 1
+▁amour 1
+▁navez 1
+▁association 1
+▁recul 1
+▁foule 1
+▁repos 1
+▁donnant 1
+▁exprim 1
+▁royale 1
+hou 1
+huit 1
+▁senti 1
+▁rire 1
+▁catégorie 1
+▁hommage 1
+▁lagriculture 1
+▁lédifice 1
+▁messieurs 1
+▁terrible 1
+▁consacre 1
+▁félicit 1
+▁voyons 1
+▁temple 1
+pparition 1
+▁finances 1
+▁puissance 1
+▁saintmartin 1
+illot 1
+▁varie 1
+▁organise 1
+▁indien 1
+▁roger 1
+▁flott 1
+▁maintenu 1
+▁perte 1
+▁brave 1
+mène 1
+▁pié 1
+▁dessus 1
+▁suffit 1
+▁créer 1
+▁humaine 1
+▁municipal 1
+▁elvire 1
+▁titulaire 1
+▁notredame 1
+▁square 1
+▁cinquième 1
+▁bête 1
+▁massif 1
+▁constant 1
+▁nordest 1
+gnol 1
+▁noy 1
+▁penser 1
+▁fréquente 1
+▁traite 1
+issent 1
+▁retrouvé 1
+▁morceau 1
+▁conservée 1
+▁courte 1
+▁républicain 1
+▁étroit 1
+▁gérard 1
+▁réparti 1
+▁admir 1
+▁acteurs 1
+▁dim 1
+▁paix 1
+bach 1
+▁méd 1
+ward 1
+▁affect 1
+▁kh 1
+▁guer 1
+▁malheur 1
+▁aussitôt 1
+▁rétabli 1
+▁réfugi 1
+▁patrick 1
+▁adulte 1
+▁souffle 1
+▁commencé 1
+▁simpl 1
+prit 1
+▁gorge 1
+▁représentation 1
+▁croire 1
+▁aj 1
+▁veille 1
+▁fondation 1
+▁considér 1
+▁jules 1
+▁lhé 1
+▁rappeler 1
+▁plage 1
+▁abord 1
+▁prenez 1
+ben 1
+mbo 1
+▁lieutenant 1
+▁assuré 1
+bul 1
+▁ménage 1
+▁entendre 1
+▁indic 1
+▁supprimé 1
+▁alex 1
+▁certainement 1
+▁vigne 1
+▁étions 1
+▁mariée 1
+▁déten 1
+▁bâti 1
+neuf 1
+▁accompagné 1
+▁contemporain 1
+▁enregistre 1
+▁fernand 1
+▁prennent 1
+ibilité 1
+▁boîte 1
+▁gratuit 1
+▁murmura 1
+▁enfance 1
+▁étend 1
+▁tol 1
+▁triste 1
+▁couvre 1
+▁juin 1
+▁pratiqu 1
+▁colon 1
+pres 1
+▁dirai 1
+▁loc 1
+▁longueur 1
+orge 1
+▁grim 1
+▁appren 1
+▁arrivée 1
+▁appartement 1
+▁kilomètres 1
+▁lorganisation 1
+▁religion 1
+▁commandant 1
+▁assassin 1
+▁laéroport 1
+▁faveur 1
+▁gardien 1
+▁fixé 1
+▁modeste 1
+▁avance 1
+▁geste 1
+▁mangé 1
+▁jan 1
+▁signé 1
+▁damien 1
+▁narr 1
+▁liv 1
+▁fréquent 1
+▁convient 1
+▁cave 1
+▁actuelle 1
+olog 1
+▁parut 1
+▁rencontrent 1
+▁réussit 1
+▁circonscription 1
+▁température 1
+▁salarié 1
+▁statut 1
+▁menace 1
+▁ennemi 1
+▁toilette 1
+▁concernant 1
+▁mention 1
+▁réalisation 1
+▁ouest 1
+▁dol 1
+▁italienne 1
+▁vitesse 1
+▁sentier 1
+▁bateau 1
+▁drapeau 1
+▁régul 1
+▁prête 1
+š 1
+▁larrivée 1
+▁possibilité 1
+▁compagnon 1
+▁compositeur 1
+▁continent 1
+▁seraient 1
+▁russie 1
+▁chevaux 1
+▁faubourg 1
+▁humide 1
+▁crée 1
+▁lentement 1
+▁alain 1
+▁lord 1
+▁amant 1
+cession 1
+▁alexandre 1
+▁ski 1
+▁rock 1
+bord 1
+américain 1
+▁livr 1
+viez 1
+▁marine 1
+▁sengage 1
+▁battu 1
+▁canon 1
+▁supérieure 1
+▁compétence 1
+▁inutile 1
+▁performance 1
+▁remarquable 1
+▁responsabilité 1
+▁sintéresse 1
+▁universitaire 1
+▁accident 1
+▁frappé 1
+▁cercle 1
+▁rempli 1
+▁frappe 1
+▁crédit 1
+▁brou 1
+▁psycho 1
+▁hauss 1
+▁anna 1
+▁george 1
+istan 1
+▁sacré 1
+▁traitement 1
+gur 1
+▁war 1
+▁regroupe 1
+▁vivent 1
+▁sénat 1
+stein 1
+▁chamb 1
+▁initiale 1
+▁toulouse 1
+▁vainqueur 1
+▁relève 1
+▁protection 1
+▁soyez 1
+▁décidé 1
+▁of 1
+▁publie 1
+▁inspiré 1
+▁christian 1
+▁gouverneur 1
+▁invite 1
+▁locaux 1
+▁sembl 1
+▁dormir 1
+▁balle 1
+▁zi 1
+▁spécialisé 1
+▁lève 1
+kov 1
+card 1
+▁pourcent 1
+▁sauf 1
+▁plume 1
+burg 1
+onsnous 1
+▁qualifié 1
+▁imagin 1
+reuil 1
+▁ellemême 1
+▁extérieur 1
+▁olivier 1
+▁deurope 1
+▁démarche 1
+▁afrique 1
+▁évoque 1
+accio 1
+▁distance 1
+▁arrière 1
+▁jeanpierre 1
+▁justifi 1
+▁west 1
+▁gagn 1
+▁she 1
+▁devra 1
+▁tienne 1
+▁oublié 1
+▁chanteur 1
+▁limp 1
+▁romaine 1
+▁fidèle 1
+▁restauration 1
+▁artistique 1
+▁commença 1
+▁phare 1
+▁poids 1
+▁éclat 1
+▁échange 1
+▁civile 1
+▁loncle 1
+▁sportif 1
+franc 1
+plan 1
+▁égal 1
+vaux 1
+▁shi 1
+▁faim 1
+▁facilement 1
+▁condamné 1
+▁adopt 1
+▁mineur 1
+▁queue 1
+urban 1
+▁maman 1
+accompagne 1
+clair 1
+▁influencé 1
+▁religieux 1
+▁réflexion 1
+▁accessible 1
+▁génération 1
+▁rassembl 1
+▁énorme 1
+▁femelle 1
+▁sagissait 1
+▁angleterre 1
+ordinateur 1
+▁hein 1
+▁fan 1
+▁sensible 1
+▁lesprit 1
+▁franchi 1
+▁accompagne 1
+▁mainten 1
+▁mette 1
+▁lundi 1
+▁top 1
+▁brune 1
+▁professionnelle 1
+dge 1
+▁exception 1
+▁catherine 1
+▁inférieur 1
+▁jattend 1
+installation 1
+▁considère 1
+▁amérique 1
+▁précisément 1
+▁effort 1
+▁affich 1
+▁chauff 1
+▁utile 1
+▁ruisseau 1
+▁lespace 1
+▁recevoir 1
+▁établissement 1
+▁édité 1
+▁sépare 1
+▁déplace 1
+▁nulle 1
+▁majeure 1
+▁logement 1
+▁capable 1
+▁écoute 1
+vision 1
+▁gentil 1
+▁peter 1
+▁déclin 1
+▁bill 1
+▁perdre 1
+▁canal 1
+▁univers 1
+▁honte 1
+▁démission 1
+▁lenseignement 1
+▁littéraire 1
+▁mécanisme 1
+▁médicament 1
+▁bruxelles 1
+▁primaire 1
+▁septembre 1
+▁brillant 1
+▁légion 1
+▁single 1
+▁réel 1
+▁galerie 1
+▁décret 1
+▁justin 1
+▁héros 1
+▁picard 1
+▁variété 1
+▁allezvous 1
+▁appris 1
+▁recette 1
+▁dessert 1
+ouche 1
+▁accueilli 1
+▁mark 1
+▁billet 1
+échappe 1
+▁succ 1
+▁mien 1
+surmer 1
+▁tableaux 1
+▁occup 1
+▁image 1
+▁mc 1
+▁castel 1
+▁born 1
+▁spi 1
+▁concept 1
+▁effective 1
+▁fédération 1
+▁réclam 1
+▁protégé 1
+▁amoureux 1
+▁déposé 1
+▁lhabit 1
+▁semblait 1
+interroge 1
+▁marius 1
+▁vice 1
+▁soirée 1
+▁verra 1
+▁majeur 1
+▁distribu 1
+▁civil 1
+▁profession 1
+▁mavez 1
+▁disparaît 1
+▁effectué 1
+▁léducation 1
+▁arnaud 1
+atteindre 1
+▁croissance 1
+▁fallu 1
+▁divisé 1
+▁mètres 1
+▁internet 1
+▁ajouta 1
+▁pape 1
+▁debout 1
+▁christophe 1
+▁naturellement 1
+▁tourisme 1
+scendant 1
+ifiée 1
+▁eaux 1
+alisation 1
+wan 1
+▁bah 1
+▁apprécié 1
+▁réservé 1
+▁lutilisation 1
+▁noël 1
+▁smith 1
+▁mexique 1
+▁concours 1
+▁leffet 1
+▁usage 1
+▁puits 1
+▁spécialiste 1
+▁conception 1
+▁sinscri 1
+▁chanteuse 1
+▁poursuivi 1
+fond 1
+▁défenseur 1
+▁james 1
+▁lappel 1
+▁soient 1
+▁final 1
+dieu 1
+▁dispos 1
+▁faux 1
+▁gré 1
+▁habite 1
+▁larchitecte 1
+▁ouvrier 1
+▁ressemble 1
+▁similaire 1
+▁épreuve 1
+▁évêque 1
+▁diffusé 1
+▁dimension 1
+▁lopposition 1
+▁polonais 1
+▁veulent 1
+▁célébr 1
+▁suivre 1
+▁honneur 1
+▁loge 1
+▁active 1
+▁plaît 1
+▁administr 1
+▁navon 1
+▁rural 1
+posé 1
+▁règlement 1
+▁oncle 1
+▁musicale 1
+án 1
+bru 1
+vière 1
+▁syn 1
+françois 1
+▁entraîne 1
+époux 1
+▁retourna 1
+grand 1
+▁condition 1
+▁déjeuner 1
+▁limitrophe 1
+▁royaumeuni 1
+▁cellule 1
+▁jentend 1
+▁montréal 1
+▁lionel 1
+▁juillet 1
+▁clef 1
+▁mérite 1
+▁bouteille 1
+▁gestion 1
+▁chaussures 1
+▁collaboration 1
+▁charmant 1
+▁université 1
+▁feuille 1
+▁sauve 1
+▁wil 1
+▁dho 1
+poli 1
+chard 1
+▁coeur 1
+▁achev 1
+▁tirer 1
+▁dhu 1
+▁blessure 1
+▁conflit 1
+▁conférence 1
+▁manuscrit 1
+▁agréable 1
+▁interromp 1
+▁mâle 1
+▁séjour 1
+▁mélange 1
+▁normand 1
+▁bourgeois 1
+▁détude 1
+▁tribu 1
+▁ronde 1
+▁tenez 1
+script 1
+▁stra 1
+▁motif 1
+▁extr 1
+▁sourd 1
+chère 1
+▁navais 1
+▁chère 1
+▁chercha 1
+▁arrêt 1
+▁ditelle 1
+▁conscience 1
+▁processus 1
+▁spécifique 1
+▁véhicule 1
+▁étrangère 1
+▁récemment 1
+▁bretagne 1
+▁sarrête 1
+▁larrière 1
+▁éliminé 1
+▁gêne 1
+▁tropical 1
+▁mairie 1
+▁beaux 1
+▁comtesse 1
+▁mondial 1
+▁fondé 1
+▁lancé 1
+▁ravi 1
+▁saintgermain 1
+▁attribué 1
+▁fortdefrance 1
+▁larchitecture 1
+▁monseigneur 1
+▁rencontré 1
+▁récompense 1
+▁signifie 1
+▁xavier 1
+▁participé 1
+▁réaction 1
+▁défaite 1
+▁drame 1
+▁nétaient 1
+▁termina 1
+vènement 1
+▁permis 1
+▁personnelle 1
+▁entrée 1
+▁forcé 1
+▁identique 1
+▁aigu 1
+▁précis 1
+▁éric 1
+▁vivement 1
+▁vélo 1
+▁réelle 1
+▁notion 1
+▁plaine 1
+▁pareille 1
+▁serre 1
+▁von 1
+▁serge 1
+bré 1
+▁lamb 1
+▁oiseau 1
+▁cérémonie 1
+▁diplôme 1
+▁lhôpital 1
+▁organisation 1
+▁laccès 1
+▁cardinal 1
+▁légende 1
+▁strasbourg 1
+▁dheure 1
+▁abouti 1
+▁fusion 1
+▁certes 1
+chou 1
+▁jusque 1
+▁frappa 1
+▁confond 1
+▁micro 1
+▁négociation 1
+▁grec 1
+▁toile 1
+▁bloc 1
+▁croyez 1
+▁pli 1
+▁porc 1
+ł 1
+▁domicile 1
+▁familial 1
+▁introduit 1
+▁mathématique 1
+▁pêchai 1
+▁renommé 1
+▁provoque 1
+▁achète 1
+▁germain 1
+▁fortune 1
+▁sportive 1
+▁tandi 1
+▁nimporte 1
+▁socialiste 1
+▁lacteur 1
+▁suffi 1
+▁pousse 1
+▁corne 1
+▁guide 1
+▁commis 1
+▁apport 1
+▁liée 1
+▁sauver 1
+idité 1
+▁quau 1
+▁efficace 1
+▁enregistré 1
+▁novembre 1
+▁ernest 1
+▁transformé 1
+▁frémi 1
+▁leurope 1
+▁modification 1
+▁reproche 1
+▁nourriture 1
+▁diable 1
+▁litalie 1
+▁institution 1
+▁comédien 1
+▁boucle 1
+▁angeles 1
+ical 1
+▁heureusement 1
+▁officiellement 1
+▁repas 1
+▁moindre 1
+▁lille 1
+mini 1
+▁puisse 1
+▁oubli 1
+▁individu 1
+▁bern 1
+▁entière 1
+rupt 1
+▁réfléchi 1
+▁expérience 1
+▁patrimoine 1
+▁promenade 1
+▁soviétique 1
+▁circuit 1
+▁espagne 1
+▁michael 1
+▁matériau 1
+▁marcof 1
+▁œil 1
+▁bonjour 1
+▁aventure 1
+▁lennemi 1
+▁hôtel 1
+▁aîné 1
+▁volume 1
+▁soupir 1
+▁parcours 1
+▁notable 1
+▁voulais 1
+▁émis 1
+▁attire 1
+▁salut 1
+▁talent 1
+investi 1
+▁délai 1
+▁rayon 1
+▁courir 1
+▁élément 1
+▁hiver 1
+ône 1
+▁léger 1
+▁métro 1
+▁chou 1
+▁subit 1
+enberg 1
+▁recueil 1
+▁regardait 1
+▁vêtements 1
+▁faculté 1
+▁lorsquil 1
+▁pleut 1
+▁lactivité 1
+▁touristique 1
+▁viande 1
+▁sagesse 1
+▁réalisateur 1
+▁valérie 1
+▁parlait 1
+▁tapis 1
+▁officielle 1
+▁gall 1
+▁soins 1
+▁bourse 1
+▁abandonne 1
+▁tiré 1
+illage 1
+▁confié 1
+▁colomb 1
+▁quasi 1
+▁chasseur 1
+▁romp 1
+▁culturel 1
+▁tendre 1
+▁écriv 1
+brivelagaillarde 1
+▁administratif 1
+▁capacité 1
+▁dîner 1
+▁ingénieur 1
+▁majoritaire 1
+▁procès 1
+▁photographie 1
+▁loreille 1
+▁lâche 1
+▁vierge 1
+▁atelier 1
+▁lactrice 1
+haut 1
+▁violon 1
+▁bourbon 1
+▁oiseaux 1
+▁cadet 1
+▁gamin 1
+▁multi 1
+▁soif 1
+châtel 1
+▁constitution 1
+▁explication 1
+▁sapplique 1
+▁suffisamment 1
+▁symbole 1
+▁adapté 1
+▁dialogue 1
+▁dinformation 1
+▁pantalon 1
+▁soutenu 1
+xtrémité 1
+▁létude 1
+▁attaché 1
+▁journaux 1
+▁déclare 1
+▁avril 1
+▁chute 1
+▁aimait 1
+▁corde 1
+villiers 1
+▁exige 1
+▁colline 1
+▁herb 1
+▁prenant 1
+▁soldat 1
+vienne 1
+▁sinon 1
+▁essayé 1
+offic 1
+iforme 1
+▁vision 1
+influence 1
+▁nouvel 1
+▁poul 1
+▁contenu 1
+▁plaque 1
+▁alfred 1
+▁construire 1
+▁intégré 1
+▁lévêque 1
+▁marguerite 1
+▁octobre 1
+▁suspension 1
+▁viceprésident 1
+▁avancée 1
+▁impérial 1
+▁intéressant 1
+▁lécriture 1
+▁précipita 1
+▁standard 1
+▁communication 1
+▁keinec 1
+▁maxime 1
+▁terrasse 1
+▁poésie 1
+▁complément 1
+▁soumis 1
+▁douleur 1
+▁arrange 1
+artisan 1
+▁paquet 1
+loup 1
+▁tiers 1
+▁pressé 1
+▁niveaux 1
+▁avion 1
+▁officiel 1
+aël 1
+▁cassé 1
+▁god 1
+▁délégué 1
+▁maréchal 1
+▁méditerranée 1
+▁scénario 1
+▁conservatoire 1
+▁bronze 1
+▁lavocat 1
+▁réputé 1
+▁janvier 1
+▁conversation 1
+▁sarrêta 1
+▁électronique 1
+▁reconnaissance 1
+ographe 1
+▁étais 1
+▁genoux 1
+▁prolonge 1
+▁limpression 1
+▁reten 1
+▁annuel 1
+▁message 1
+▁enseignant 1
+▁confirme 1
+▁entraîneur 1
+▁garanti 1
+▁béni 1
+▁tourné 1
+nny 1
+▁culturelle 1
+▁adopté 1
+blanc 1
+▁dirigeant 1
+▁comédie 1
+▁exemplaire 1
+▁immense 1
+▁sévère 1
+▁lhiver 1
+▁produis 1
+▁discipline 1
+▁modifie 1
+▁secria 1
+▁impression 1
+▁colonie 1
+▁violent 1
+▁modifié 1
+▁caisse 1
+▁veuxtu 1
+▁sentait 1
+asso 1
+▁navaient 1
+▁boire 1
+▁avantage 1
+▁délivr 1
+▁parisien 1
+▁naurai 1
+crire 1
+améliorer 1
+▁sculpture 1
+▁soupçonn 1
+▁stéphan 1
+▁libéral 1
+▁rattaché 1
+▁traduction 1
+▁pologne 1
+▁fatigue 1
+plomb 1
+▁août 1
+▁pénétr 1
+▁lissue 1
+▁subsiste 1
+▁culte 1
+▁assistant 1
+▁facteur 1
+▁dossier 1
+▁paysbas 1
+▁lavez 1
+▁lunettes 1
+▁seigneurie 1
+▁découvrir 1
+▁rennes 1
+▁éviter 1
+▁examin 1
+▁yann 1
+▁proposons 1
+▁rival 1
+▁repart 1
+▁sixième 1
+▁aimerait 1
+▁retenu 1
+▁colombie 1
+▁souci 1
+▁serra 1
+ñ 1
+▁dhistoire 1
+▁lobjectif 1
+▁néerlandais 1
+▁souterrain 1
+interprète 1
+▁espoir 1
+▁dommage 1
+▁chemise 1
+surloire 1
+▁empêche 1
+▁thèse 1
+▁veuve 1
+▁policier 1
+▁exclu 1
+▁jirai 1
+▁sympa 1
+▁opposé 1
+▁craign 1
+▁souffre 1
+▁violence 1
+▁bordeaux 1
+▁sucre 1
+▁amateur 1
+issaient 1
+▁pluie 1
+▁footballeur 1
+coup 1
+▁dense 1
+▁orateur 1
+martin 1
+▁renard 1
+ntendre 1
+▁communiste 1
+ctuelle 1
+▁seuil 1
+▁régulière 1
+▁modern 1
+▁ferai 1
+▁financier 1
+▁lespèce 1
+▁secrète 1
+▁successeur 1
+▁vôtre 1
+▁chassaigne 1
+▁réputation 1
+▁indiqué 1
+▁trafic 1
+▁lâme 1
+surseine 1
+▁mappelle 1
+▁notaire 1
+▁florence 1
+▁win 1
+▁latéral 1
+augustin 1
+▁touché 1
+▁nantes 1
+▁lentend 1
+▁taxe 1
+▁trahi 1
+▁web 1
+▁viol 1
+▁gagné 1
+▁chienne 1
+▁soeur 1
+▁pétr 1
+▁sto 1
+▁oct 1
+▁sienne 1
+well 1
+ć 1
+▁additionnel 1
+▁fabrication 1
+▁lexistence 1
+▁gaillard 1
+▁évoqué 1
+▁décembre 1
+▁posséda 1
+▁costume 1
+▁conduis 1
+▁croyais 1
+▁envoyer 1
+▁consiste 1
+▁amoureuse 1
+▁lartiste 1
+▁tendance 1
+▁grandpère 1
+▁maroc 1
+▁princesse 1
+▁jim 1
+▁dinter 1
+▁émet 1
+train 1
+▁planche 1
+▁iron 1
+pierre 1
+jacques 1
+avant 1
+▁architecte 1
+▁circonstance 1
+▁circulaire 1
+▁importance 1
+▁montpellier 1
+▁plateforme 1
+▁progressivement 1
+▁robiliard 1
+▁libération 1
+▁fondamental 1
+▁portugal 1
+▁révèle 1
+▁limoges 1
+▁ordinaire 1
+▁tambour 1
+▁dragon 1
+▁henry 1
+▁souffrir 1
+▁émile 1
+▁mourut 1
+▁despagne 1
+▁occidentale 1
+▁envahi 1
+▁dannées 1
+▁partenaire 1
+identité 1
+▁moscou 1
+▁automatique 1
+▁plastique 1
+▁vouloir 1
+▁robinet 1
+▁ditalie 1
+▁vincent 1
+▁foyer 1
+▁chair 1
+▁venons 1
+▁manuel 1
+▁morceaux 1
+▁annie 1
+▁publicité 1
+fou 1
+trop 1
+▁music 1
+▁tha 1
+▁bré 1
+▁décoré 1
+▁surpris 1
+▁essentiel 1
+▁application 1
+▁congrès 1
+▁financière 1
+▁généraux 1
+▁poitrine 1
+▁spirituel 1
+▁denviron 1
+▁fédéral 1
+▁recouvert 1
+▁sociaux 1
+▁engagement 1
+▁guitare 1
+▁libéré 1
+▁apparu 1
+▁deviendra 1
+▁opéra 1
+▁ajouté 1
+▁discours 1
+▁émission 1
+▁embrass 1
+▁damour 1
+▁milliers 1
+▁toulon 1
+▁voile 1
+▁hydro 1
+▁carrier 1
+▁sénateur 1
+▁parlons 1
+▁surmonté 1
+▁juive 1
+▁soulev 1
+▁clip 1
+antoine 1
+▁traditionnel 1
+▁géré 1
+▁apporté 1
+č 1
+▁indispensable 1
+▁refroidi 1
+▁saintquentin 1
+▁sculpteur 1
+▁appareil 1
+▁nôtre 1
+▁suédois 1
+▁typique 1
+▁disputé 1
+▁radical 1
+▁déficit 1
+▁plaisant 1
+xtension 1
+pparence 1
+▁voyant 1
+▁renonce 1
+▁déclara 1
+▁aimable 1
+▁fût 1
+▁répéta 1
+annonce 1
+▁arabe 1
+field 1
+▁conven 1
+provence 1
+▁ramass 1
+▁évolution 1
+▁orientale 1
+▁pente 1
+▁gramm 1
+▁vaud 1
+▁heureuse 1
+▁cidessous 1
+▁collabore 1
+▁enceinte 1
+▁honnête 1
+▁magazine 1
+▁renforce 1
+▁camarade 1
+▁delphine 1
+▁jémet 1
+▁obligé 1
+▁remarqué 1
+lesbains 1
+▁mécanique 1
+▁escalier 1
+▁sèche 1
+▁profondément 1
+▁environnement 1
+▁âgé 1
+▁pâle 1
+▁dhomme 1
+▁bouge 1
+▁partagé 1
+▁douvrir 1
+▁stock 1
+▁malheureux 1
+▁coureur 1
+▁maigre 1
+▁secours 1
+▁villeneuve 1
+▁préfecture 1
+▁cite 1
+▁promet 1
+▁décoration 1
+▁rideau 1
+▁rougi 1
+hypothèse 1
+▁gendarme 1
+▁informatique 1
+▁lallemagne 1
+▁létranger 1
+▁pacifique 1
+▁territorial 1
+▁œufs 1
+▁brusquement 1
+▁héricourt 1
+▁chicago 1
+▁enquête 1
+▁lemploi 1
+▁dispers 1
+▁miracle 1
+▁chanoine 1
+▁liaison 1
+▁audelà 1
+▁césar 1
+▁tribunal 1
+▁irré 1
+▁lointain 1
+▁larrêt 1
+▁plaint 1
+▁collaborateur 1
+▁vocation 1
+▁fruit 1
+▁minute 1
+▁demeura 1
+▁répandu 1
+▁spécial 1
+▁affecté 1
+▁grain 1
+mbassadeur 1
+ayez 1
+ançai 1
+▁sérieuse 1
+▁sho 1
+▁herman 1
+▁délégation 1
+▁numérique 1
+▁pistolet 1
+▁saintlaurent 1
+▁suprême 1
+▁vieillard 1
+▁drôle 1
+▁déménage 1
+▁emprunt 1
+▁promotion 1
+▁laccord 1
+▁excusezmoi 1
+▁trésor 1
+▁étienne 1
+▁armand 1
+▁pitié 1
+▁lescalier 1
+▁gâteau 1
+▁échec 1
+▁parfum 1
+▁champagne 1
+▁quitté 1
+▁existence 1
+▁agit 1
+▁format 1
+▁prouve 1
+▁profondeur 1
+▁couverture 1
+▁couronne 1
+▁teint 1
+▁couteau 1
+▁repli 1
+▁sélectionné 1
+rhône 1
+▁lattend 1
+▁franco 1
+claude 1
+ouvrit 1
+▁fuir 1
+▁accompli 1
+▁espérance 1
+▁létablissement 1
+▁récupér 1
+▁emprisonn 1
+▁lépouse 1
+▁intervention 1
+▁linstant 1
+▁virginie 1
+▁diamètre 1
+▁autorisé 1
+▁décès 1
+▁deuros 1
+▁monstre 1
+▁véron 1
+▁complet 1
+pouvant 1
+▁allié 1
+hôte 1
+▁artiste 1
+amine 1
+▁releva 1
+ight 1
+▁condamn 1
+▁thu 1
+ntraîneur 1
+▁basketball 1
+▁cohérence 1
+▁jeanbaptiste 1
+▁judiciaire 1
+▁ladministration 1
+▁lautorité 1
+▁suspendu 1
+▁incertain 1
+▁chœur 1
+▁trentaine 1
+▁éteint 1
+▁témoin 1
+▁fabrique 1
+▁apparaissent 1
+▁fraysse 1
+équipage 1
+▁acquis 1
+▁suivent 1
+▁algérie 1
+▁expose 1
+▁promesse 1
+▁doublé 1
+▁variante 1
+▁variable 1
+▁lèvres 1
+▁sousmarin 1
+▁sébastien 1
+▁connait 1
+▁gaston 1
+▁isolé 1
+▁reims 1
+▁degré 1
+acteur 1
+▁aidé 1
+▁passionné 1
+▁morale 1
+▁molle 1
+▁lorient 1
+▁archéologique 1
+▁irlandais 1
+▁napoléon 1
+▁saintétienne 1
+▁sanctuaire 1
+▁scrutin 1
+intermédiaire 1
+▁circulation 1
+▁historien 1
+▁expert 1
+▁abondant 1
+▁félix 1
+▁lessentiel 1
+▁locéan 1
+▁autrichien 1
+▁tentative 1
+▁confirmé 1
+▁arthur 1
+▁résolu 1
+▁nentend 1
+▁fantas 1
+▁jazz 1
+▁préparé 1
+▁samedi 1
+▁pension 1
+▁admis 1
+▁désigné 1
+▁copie 1
+▁patient 1
+▁constitutionnel 1
+▁balcon 1
+▁fuite 1
+▁pérou 1
+▁médor 1
+▁recueilli 1
+▁musical 1
+▁adam 1
+▁santa 1
+▁pareil 1
+▁africain 1
+▁aménagé 1
+▁artificiel 1
+▁grandebretagne 1
+▁renaissance 1
+▁technologie 1
+▁potentiel 1
+▁registre 1
+▁scandale 1
+▁excuse 1
+▁lœil 1
+▁daffaires 1
+▁beauxarts 1
+▁gaudin 1
+▁fâch 1
+▁lespagne 1
+▁animé 1
+▁lévolution 1
+▁logis 1
+▁lattaque 1
+▁usine 1
+▁lombre 1
+▁flore 1
+▁voila 1
+▁lesquels 1
+▁partisan 1
+▁voyageurs 1
+▁décrire 1
+▁cellelà 1
+graphie 1
+▁filière 1
+anluc 1
+▁dappel 1
+▁négoci 1
+▁référ 1
+▁hockey 1
+▁hongrois 1
+▁laboratoire 1
+▁luxembourg 1
+▁lécrivain 1
+▁lémission 1
+▁personnalité 1
+▁économie 1
+▁analyse 1
+▁rythme 1
+▁ultérieur 1
+▁appliqué 1
+▁rupture 1
+▁mensonge 1
+▁gloire 1
+▁serbie 1
+▁liquide 1
+▁licence 1
+▁déclaré 1
+▁foncé 1
+▁parfait 1
+▁barrage 1
+▁soudain 1
+▁échoue 1
+▁connaissent 1
+▁curieux 1
+▁hélas 1
+chonnet 1
+▁rentré 1
+▁leader 1
+▁grecque 1
+▁renvoy 1
+▁sérieux 1
+▁parrain 1
+ouillet 1
+crate 1
+▁iso 1
+▁philip 1
+▁robin 1
+▁benjamin 1
+▁convaincu 1
+▁coordination 1
+▁doctorat 1
+▁géographique 1
+▁jacquelin 1
+▁mathilde 1
+▁obligatoire 1
+▁photographe 1
+▁printemps 1
+▁saintgeorges 1
+▁satisfaction 1
+▁superficie 1
+▁nêtes 1
+▁accoyer 1
+▁alinéa 1
+▁procédé 1
+▁adversaire 1
+▁thierry 1
+▁aérienne 1
+▁fragile 1
+▁scolaire 1
+▁chêne 1
+▁labsence 1
+quatrevingtseize 1
+▁inquiétude 1
+▁éditeur 1
+▁baroque 1
+▁meuble 1
+▁décida 1
+▁plafond 1
+▁grandmère 1
+▁planète 1
+mbition 1
+▁varié 1
+▁étude 1
+iculture 1
+▁distinct 1
+▁immédiat 1
+▁demifinale 1
+▁exclusivement 1
+▁fréquemment 1
+▁guitariste 1
+▁jeanlouis 1
+▁maîtrise 1
+▁nîmes 1
+▁scénariste 1
+▁cavalerie 1
+▁entreprend 1
+▁explor 1
+▁explos 1
+▁préparation 1
+▁merveille 1
+▁saintdenis 1
+▁diffuse 1
+▁vendredi 1
+▁vigueur 1
+▁obscur 1
+▁vaincu 1
+▁encourage 1
+▁dérivé 1
+obstacle 1
+▁clavier 1
+▁davance 1
+▁sexuel 1
+▁approuv 1
+atlantique 1
+▁lopération 1
+▁menacé 1
+▁lusage 1
+▁saurai 1
+▁participation 1
+métrique 1
+▁suède 1
+▁grille 1
+▁lavion 1
+▁angle 1
+autant 1
+▁will 1
+▁régulier 1
+ouest 1
+▁employ 1
+▁mouill 1
+allemand 1
+▁dadministration 1
+▁ferroviaire 1
+▁francophone 1
+▁lexpérience 1
+▁louvrage 1
+▁moimême 1
+▁cratère 1
+▁pupponi 1
+▁tressaill 1
+▁annoncé 1
+▁compliqué 1
+▁trône 1
+▁pavillon 1
+▁mèche 1
+▁couramment 1
+▁panneaux 1
+▁souligne 1
+▁caroline 1
+▁canapé 1
+▁préparer 1
+▁bellerose 1
+▁demoiselle 1
+▁étoile 1
+▁hâte 1
+▁multipli 1
+▁pôle 1
+▁habituelle 1
+▁suppli 1
+▁opinion 1
+▁troyes 1
+▁cerveau 1
+▁quimper 1
+▁genève 1
+▁paradis 1
+▁formule 1
+▁décède 1
+▁boston 1
+▁distribution 1
+▁remettre 1
+▁abbé 1
+▁définition 1
+▁expression 1
+town 1
+▁turc 1
+▁retrouv 1
+patri 1
+▁ouvrage 1
+acheter 1
+▁voudra 1
+▁lunivers 1
+▁agriculteur 1
+▁banlieue 1
+▁chocolat 1
+▁divorce 1
+▁dorléans 1
+▁lappareil 1
+▁linguistique 1
+▁maximum 1
+▁nucléaire 1
+▁nécessite 1
+▁pharmacie 1
+▁syndicat 1
+▁vestiges 1
+▁éponyme 1
+▁bombard 1
+▁hospital 1
+▁vengeance 1
+▁médicale 1
+▁viendront 1
+▁inscription 1
+thlète 1
+▁capturé 1
+▁contexte 1
+▁lélève 1
+▁masculin 1
+▁oppose 1
+▁existant 1
+▁maritime 1
+▁gueule 1
+▁prononce 1
+métrie 1
+▁natale 1
+▁terrestre 1
+▁cellesci 1
+▁otto 1
+▁tunisie 1
+▁mattend 1
+▁venise 1
+▁distinction 1
+▁semblable 1
+▁pénal 1
+▁savant 1
+▁pensé 1
+thétique 1
+▁gouvern 1
+utorise 1
+▁bizarre 1
+▁enjolras 1
+▁interprété 1
+▁laustralie 1
+▁madeleine 1
+▁médiéval 1
+▁nécessité 1
+▁repoussé 1
+▁saintlouis 1
+▁bourgogne 1
+▁définitivement 1
+▁transféré 1
+▁bertrand 1
+▁linformation 1
+▁réduire 1
+▁cadavre 1
+▁dramatique 1
+surmarne 1
+▁puisquil 1
+▁étiez 1
+▁college 1
+▁cadio 1
+▁appuy 1
+▁impose 1
+▁dégrad 1
+▁spécialité 1
+▁théologie 1
+▁dolez 1
+▁individuelle 1
+▁gold 1
+rgio 1
+▁thor 1
+▁averti 1
+dauphin 1
+▁err 1
+▁associ 1
+▁vaisseau 1
+occidental 1
+▁cavrois 1
+▁extraordinaire 1
+▁février 1
+▁intéressé 1
+▁isabelle 1
+▁langleterre 1
+▁lextérieur 1
+▁positive 1
+▁répertoire 1
+▁tempête 1
+▁édifice 1
+▁lintérêt 1
+▁lénergie 1
+▁statistique 1
+▁étouff 1
+▁texas 1
+▁dépôt 1
+▁dhiver 1
+▁cavalier 1
+▁actrice 1
+▁converti 1
+▁jacob 1
+▁lexpression 1
+▁maintenir 1
+▁fonctionnaire 1
+▁dentrée 1
+▁poudre 1
+▁déesse 1
+▁écart 1
+▁virg 1
+▁manteau 1
+▁mentionné 1
+▁regroup 1
+willer 1
+_ 1
+xpérimentation 1
+▁architectural 1
+▁concrète 1
+▁disparition 1
+▁musulman 1
+▁reconstruit 1
+▁épousé 1
+▁grenoble 1
+▁horrible 1
+▁lorchestre 1
+▁louverture 1
+▁mercredi 1
+▁perpignan 1
+▁gosselin 1
+▁priorité 1
+▁résistant 1
+▁dollars 1
+▁suicide 1
+▁soppose 1
+▁saintandré 1
+▁spectateur 1
+▁sélève 1
+▁vaccin 1
+▁éloigné 1
+▁impôt 1
+▁chapitre 1
+▁révolte 1
+▁apparent 1
+▁offert 1
+▁examen 1
+▁loffice 1
+▁inédit 1
+▁nancy 1
+▁critère 1
+▁tunnel 1
+▁borgo 1
+▁racine 1
+ncêtre 1
+▁emil 1
+▁provenan 1
+ppellation 1
+ignait 1
+▁jumel 1
+▁connaissez 1
+paraissait 1
+▁séchapp 1
+bject 1
+château 1
+achève 1
+rchevêque 1
+▁adolescent 1
+▁danemark 1
+▁dinfanterie 1
+▁possession 1
+▁seffectue 1
+▁stratégique 1
+ā 1
+▁centralisateur 1
+▁rugby 1
+▁dactivité 1
+▁nachète 1
+obligation 1
+▁répète 1
+▁suppose 1
+▁quelconque 1
+▁illustra 1
+▁rejoign 1
+▁brève 1
+▁dijon 1
+▁avignon 1
+▁gravure 1
+▁monnaie 1
+▁fichier 1
+▁télévisé 1
+▁lépaule 1
+▁parking 1
+▁trophée 1
+▁travaillait 1
+▁retomb 1
+▁lattention 1
+▁dartistes 1
+▁institu 1
+▁parlant 1
+▁cadeau 1
+quiert 1
+place 1
+joly 1
+▁enjeu 1
+structur 1
+xplique 1
+▁lopéra 1
+ú 1
+▁comparaison 1
+▁incroyable 1
+▁lapplication 1
+▁attitude 1
+▁multiples 1
+▁substitu 1
+▁petitfils 1
+▁conduire 1
+▁montauban 1
+▁turquie 1
+▁jésus 1
+▁cinieri 1
+▁débarras 1
+▁transfert 1
+▁permanent 1
+▁fragment 1
+▁directrice 1
+▁affronte 1
+évaluation 1
+▁enseigné 1
+▁rédigé 1
+ttraction 1
+▁amitié 1
+▁néglige 1
+▁vapeur 1
+▁martial 1
+▁contempl 1
+▁lambert 1
+▁rétro 1
+▁dancien 1
+▁savère 1
+sphère 1
+▁douvr 1
+▁drap 1
+▁correct 1
+▁entretien 1
+feld 1
+▁accueil 1
+▁caractérise 1
+▁disciple 1
+▁dynamique 1
+▁identifié 1
+▁juridiction 1
+▁làdessus 1
+▁nautilus 1
+▁portugais 1
+▁progrès 1
+▁provincial 1
+▁recommanda 1
+▁réservoir 1
+▁senfuit 1
+▁archives 1
+▁catalogue 1
+▁fromage 1
+▁succession 1
+▁sophie 1
+▁confortable 1
+▁souviens 1
+▁implanté 1
+▁papillon 1
+▁gilbert 1
+▁serpent 1
+▁décédé 1
+▁précieux 1
+▁reviens 1
+▁aprèsmidi 1
+▁verdun 1
+perçoi 1
+▁recrute 1
+▁sinspire 1
+▁lavenue 1
+▁johann 1
+▁sauvé 1
+▁martyr 1
+▁lydie 1
+▁latelier 1
+▁avanç 1
+▁absent 1
+▁cosette 1
+arctique 1
+doption 1
+▁supprim 1
+frédéric 1
+ū 1
+▁anniversaire 1
+▁clermontferrand 1
+▁dartagnan 1
+▁différenci 1
+▁désespéré 1
+▁internationaux 1
+▁librairie 1
+▁magnifique 1
+▁nouvellezélande 1
+▁systématique 1
+▁barricade 1
+▁confusion 1
+▁dintérêt 1
+▁emmanuel 1
+▁hetzel 1
+▁limportance 1
+▁décennie 1
+▁pomereux 1
+▁épuisé 1
+▁jahoua 1
+▁étonnant 1
+▁pionnier 1
+▁réduction 1
+▁chronique 1
+▁solitaire 1
+▁réparation 1
+▁brown 1
+▁ceinture 1
+▁sempare 1
+▁idéal 1
+▁niort 1
+▁fumée 1
+▁servait 1
+▁plou 1
+▁propag 1
+▁renseig 1
+▁sabot 1
+▁recev 1
+paraissent 1
+▁design 1
+▁chevr 1
+éviter 1
+▁hurl 1
+agnac 1
+grange 1
+▁île 1
+égyptien 1
+▁dangleterre 1
+▁diffusion 1
+▁dynastie 1
+▁intelligent 1
+▁jérôme 1
+▁madrid 1
+▁refuge 1
+▁junior 1
+▁survécu 1
+▁sexprime 1
+▁militant 1
+▁lélection 1
+▁commandé 1
+▁carfor 1
+▁calcaire 1
+▁frissonn 1
+▁naples 1
+▁filiale 1
+▁diamant 1
+▁mixte 1
+▁rumeur 1
+▁pâte 1
+▁office 1
+▁inquiét 1
+▁maçon 1
+▁fief 1
+yvonne 1
+▁contract 1
+▁désigna 1
+▁lactuel 1
+▁détien 1
+▁représent 1
+dopter 1
+major 1
+▁quatrevingt 1
+▁aperçu 1
+▁automobile 1
+▁complexité 1
+▁concurrence 1
+▁despèce 1
+▁immeuble 1
+▁lemplacement 1
+▁lexemple 1
+▁lextrême 1
+▁loccupation 1
+▁reproduction 1
+▁solidarité 1
+▁tchèque 1
+▁washington 1
+▁élégant 1
+▁ressort 1
+▁réveiller 1
+▁absorb 1
+▁school 1
+▁plaindre 1
+▁présidé 1
+▁chagrin 1
+▁instance 1
+▁illustre 1
+▁patience 1
+▁veuille 1
+▁dialecte 1
+▁suggère 1
+▁autonome 1
+▁lédition 1
+▁chaumont 1
+▁savary 1
+▁suscit 1
+▁travaillant 1
+▁leffort 1
+▁dépasse 1
+▁carross 1
+crypt 1
+▁suivez 1
+▁suffisant 1
+▁heurt 1
+▁rigol 1
+▁apais 1
+▁olymp 1
+ø 1
+▁adolphe 1
+▁chirurgie 1
+▁compromis 1
+▁geneviève 1
+▁grèce 1
+▁larchipel 1
+▁lunanimité 1
+▁méridional 1
+▁métallique 1
+▁particularité 1
+▁temporaire 1
+▁collectif 1
+▁linspecteur 1
+▁organisme 1
+▁prononcé 1
+▁provisoire 1
+▁bulletin 1
+▁élabor 1
+▁dafrique 1
+▁lalinéa 1
+▁quelquefois 1
+▁bouvard 1
+▁dimportant 1
+▁débarque 1
+▁jaloux 1
+▁augmenté 1
+▁camille 1
+▁partiellement 1
+▁habitude 1
+▁accusé 1
+▁décriv 1
+▁chauve 1
+apporte 1
+▁granit 1
+accuse 1
+▁transmis 1
+griffe 1
+aije 1
+▁express 1
+▁aveugle 1
+▁envelopp 1
+▁hollande 1
+▁incapable 1
+▁jalousie 1
+▁joséphine 1
+▁latmosphère 1
+▁lenvironnement 1
+▁louwagie 1
+▁lévénement 1
+▁molécule 1
+▁norvège 1
+▁problématique 1
+▁théorème 1
+▁écossais 1
+▁delaunay 1
+▁dieudonné 1
+▁lendroit 1
+▁massonneau 1
+▁survivant 1
+▁volontaire 1
+▁sénégal 1
+▁daccès 1
+▁damérique 1
+▁massacre 1
+▁produire 1
+▁rencontra 1
+▁absurde 1
+▁délicate 1
+▁hongrie 1
+▁graphique 1
+▁interrogati 1
+▁daoût 1
+▁attaqué 1
+▁raisonnable 1
+▁dœuvre 1
+▁alliance 1
+▁bailli 1
+▁concentre 1
+▁mépris 1
+▁lafrique 1
+▁dachat 1
+▁description 1
+▁porthos 1
+▁lécoute 1
+▁destination 1
+▁adrien 1
+▁schéma 1
+▁arrach 1
+▁amené 1
+transmission 1
+▁entame 1
+▁insist 1
+▁amérindien 1
+▁culminant 1
+▁démocratique 1
+▁genevard 1
+▁licencié 1
+▁légitime 1
+▁pécuchet 1
+▁stratégie 1
+▁tokyo 1
+▁équivalent 1
+▁apprécie 1
+▁floraison 1
+▁saddier 1
+▁illustré 1
+▁vercamer 1
+▁géorgie 1
+▁évident 1
+▁rajouter 1
+▁botanique 1
+▁cabaret 1
+▁combattants 1
+▁noublie 1
+▁proportion 1
+thèque 1
+▁ordonné 1
+▁crains 1
+▁minuit 1
+▁soyons 1
+oeuvre 1
+▁embarqu 1
+ugmentation 1
+▁lanimal 1
+▁létang 1
+▁versailles 1
+▁désolé 1
+▁camion 1
+▁wolf 1
+accéder 1
+▁compare 1
+▁gamb 1
+▁trott 1
+▁contrôl 1
+▁industrie 1
+▁affilié 1
+▁byzantin 1
+▁forestière 1
+▁geoffroy 1
+▁indépendance 1
+▁lantiquité 1
+▁lenquête 1
+▁linformatique 1
+▁madagascar 1
+▁paroissiale 1
+▁quarantaine 1
+▁remplaça 1
+▁rezdechaussée 1
+▁satellite 1
+▁transparence 1
+▁vestibule 1
+▁cameroun 1
+▁perpétue 1
+▁redoutable 1
+▁cohérent 1
+▁encadré 1
+▁indonésie 1
+▁nièce 1
+▁vaincre 1
+▁fiancé 1
+▁maltenu 1
+▁révélé 1
+▁altitude 1
+▁joël 1
+▁abattu 1
+▁andrew 1
+▁lanterne 1
+▁érigé 1
+▁sprint 1
+▁parbleu 1
+▁boulanger 1
+▁transforme 1
+▁coffre 1
+▁atteign 1
+▁joint 1
+▁convoi 1
+▁project 1
+limard 1
+▁agrandi 1
+▁ouvra 1
+▁détest 1
+▁capital 1
+joseph 1
+philippe 1
+sursaône 1
+▁alternative 1
+▁chômage 1
+▁coronado 1
+▁crâne 1
+▁demiheure 1
+▁dignelesbains 1
+▁distingué 1
+▁ecclésiastique 1
+▁korigane 1
+▁lindépendance 1
+▁magistrat 1
+▁philosophe 1
+▁pseudonyme 1
+▁saintnicolas 1
+▁facilite 1
+▁impliqué 1
+▁paupières 1
+▁perspective 1
+▁raphaël 1
+▁vêtu 1
+▁remiremont 1
+▁réveillé 1
+scenseur 1
+▁clarifie 1
+▁littoral 1
+▁aéroport 1
+▁manœuvre 1
+▁complété 1
+▁aménage 1
+▁cayenne 1
+▁traîna 1
+▁barbara 1
+▁terroris 1
+▁copain 1
+▁arrondissement 1
+▁mobile 1
+▁coiffe 1
+▁relever 1
+▁doyen 1
+▁allumé 1
+▁scott 1
+échelle 1
+▁rêv 1
+vantage 1
+bandonner 1
+▁probable 1
+éléphant 1
+▁angoulême 1
+▁commerçant 1
+▁controversé 1
+▁convaincre 1
+▁endommagé 1
+▁interprétation 1
+▁norvégien 1
+▁prétexte 1
+▁recommande 1
+▁récompensé 1
+▁significative 1
+▁successivement 1
+▁symbolis 1
+▁weekend 1
+▁électoral 1
+▁bénédict 1
+▁denseignement 1
+▁démocrate 1
+▁ferdinand 1
+▁lartillerie 1
+▁mystère 1
+▁pardessus 1
+▁rattrap 1
+▁procéder 1
+▁honoraire 1
+▁roumanie 1
+▁piscine 1
+▁rabault 1
+▁imprimé 1
+▁poitiers 1
+▁brigade 1
+▁émotion 1
+▁demploi 1
+▁grève 1
+phérique 1
+▁pompier 1
+▁trompé 1
+▁tabac 1
+quérir 1
+▁mâch 1
+▁halte 1
+rèze 1
+▁recense 1
+▁valentin 1
+▁plonge 1
+▁cumul 1
+▁parvenu 1
+boeuf 1
+▁diffère 1
+▁censé 1
+stadt 1
+utilise 1
+▁alexandr 1
+▁tissu 1
+saintpétersbourg 1
+▁apparemment 1
+▁besançon 1
+▁collective 1
+▁constructeur 1
+▁construis 1
+▁controverse 1
+▁cravate 1
+▁curiosité 1
+▁dimpôt 1
+▁dissimul 1
+▁effectif 1
+▁lalphabet 1
+▁lefebvre 1
+▁mammifère 1
+▁misérable 1
+▁présidence 1
+▁pèlerinage 1
+▁sylvain 1
+▁énormément 1
+▁balbutia 1
+▁confronté 1
+▁conjoint 1
+▁prairies 1
+▁syndical 1
+▁entraîné 1
+▁fabriqué 1
+▁fraîche 1
+▁transformation 1
+▁brigitte 1
+▁compensation 1
+▁kirsch 1
+▁cuivre 1
+▁océan 1
+▁podium 1
+▁séminaire 1
+▁cartouche 1
+▁répéter 1
+▁incident 1
+▁sultan 1
+▁implique 1
+▁relatif 1
+▁incarne 1
+▁oblige 1
+▁métonne 1
+▁connect 1
+instinct 1
+▁portail 1
+▁mitigé 1
+▁ultra 1
+▁confident 1
+▁hésita 1
+▁électro 1
+▁vibr 1
+oriental 1
+▁jaurai 1
+▁décroch 1
+▁plaign 1
+▁découvr 1
+saintsauveur 1
+▁adaptation 1
+▁concurrent 1
+▁darchitecture 1
+▁dinquiétude 1
+▁flaviana 1
+▁insectes 1
+▁jésuite 1
+▁lexpédition 1
+▁lhistorien 1
+▁mathieu 1
+▁montaudoin 1
+▁pleuvoir 1
+▁synthèse 1
+▁sépulture 1
+▁jurassien 1
+▁perfection 1
+▁signature 1
+▁antonio 1
+▁générique 1
+▁patrouille 1
+▁revêt 1
+▁arménien 1
+▁cardiaque 1
+▁laccident 1
+▁réception 1
+▁séparation 1
+▁corriger 1
+▁proclam 1
+nouvelleguinée 1
+▁toronto 1
+▁essuy 1
+▁casquette 1
+▁massive 1
+▁orchestr 1
+▁cigare 1
+▁restauré 1
+▁gregor 1
+▁sylvi 1
+▁georg 1
+▁prun 1
+▁marteau 1
+chapelle 1
+▁étudia 1
+▁cambr 1
+▁arrondi 1
+▁barcelone 1
+▁biodiversité 1
+▁congrégation 1
+▁cylindr 1
+▁cyprien 1
+▁dolorès 1
+▁déconomie 1
+▁facilité 1
+▁lépisode 1
+▁protocole 1
+▁précédemment 1
+▁pédagogique 1
+▁recommandé 1
+▁reconstruction 1
+▁reçoivent 1
+▁récurrent 1
+▁symbolique 1
+▁dartillerie 1
+▁foudroy 1
+▁israël 1
+giovanni 1
+xpansion 1
+▁garnison 1
+▁hélène 1
+▁conducteur 1
+▁recherché 1
+▁suzanne 1
+▁léditeur 1
+▁dépassé 1
+▁précoce 1
+▁tranché 1
+▁installe 1
+▁virtuel 1
+▁couloir 1
+▁benoit 1
+▁linfini 1
+▁comptoir 1
+▁matthi 1
+▁barbier 1
+▁oraison 1
+▁furieuse 1
+▁inégal 1
+▁héberge 1
+▁nymph 1
+▁assume 1
+▁habitué 1
+▁persist 1
+▁métai 1
+organiser 1
+cœur 1
+volution 1
+▁anonyme 1
+▁attentif 1
+▁bathilde 1
+▁charpentier 1
+▁concentration 1
+▁consécutive 1
+▁douloureuse 1
+▁immobile 1
+▁insupportable 1
+▁laprèsmidi 1
+▁lobservatoire 1
+▁portefeuille 1
+▁programmé 1
+▁rectangulaire 1
+▁sacrifice 1
+▁souffert 1
+▁vérifier 1
+▁éventuelle 1
+▁bonhomme 1
+▁désordre 1
+▁prononça 1
+▁critiqué 1
+▁grandchose 1
+▁porteparole 1
+▁sétabli 1
+▁violemment 1
+ntrepreneur 1
+▁métropole 1
+▁brièvement 1
+▁cahors 1
+▁légumes 1
+▁bilingue 1
+▁cinématographi 1
+▁savança 1
+▁étiezvous 1
+▁mennuie 1
+▁nettoy 1
+▁consolid 1
+▁prouvé 1
+▁chandelle 1
+▁préfet 1
+▁mutualis 1
+▁prolongé 1
+▁habile 1
+▁consenti 1
+▁éternel 1
+▁suivai 1
+christophe 1
+affluent 1
+œuvr 1
+▁appartienne 1
+▁architecture 1
+▁communautaire 1
+▁conquête 1
+▁détruire 1
+▁euxmêmes 1
+▁fatigant 1
+▁hiérarchie 1
+▁infrastructure 1
+▁lenregistrement 1
+▁lexécution 1
+▁lintelligence 1
+▁marvejol 1
+▁procureur 1
+▁péninsule 1
+▁renseignements 1
+▁résoudre 1
+▁révérend 1
+▁sapprocha 1
+▁élargi 1
+▁mobilier 1
+▁rédacteur 1
+▁lélevage 1
+▁mosquée 1
+▁inspecteur 1
+▁goriot 1
+▁manosque 1
+▁limpact 1
+▁pianiste 1
+▁tonnerre 1
+▁réagir 1
+▁senfuir 1
+▁complice 1
+▁trajet 1
+▁voyezvous 1
+▁avanthier 1
+▁general 1
+▁étonné 1
+▁cognac 1
+▁dépouill 1
+▁reprena 1
+▁criminel 1
+▁repond 1
+▁emmenez 1
+▁jumeau 1
+ncontre 1
+▁nallai 1
+observer 1
+▁répand 1
+▁saintjulien 1
+ı 1
+▁calendrier 1
+▁catastrophe 1
+▁diagnostic 1
+▁exercice 1
+▁fauchelevent 1
+▁fréquence 1
+▁handicap 1
+▁investissements 1
+▁jusquici 1
+▁lexploitation 1
+▁linitiative 1
+▁québécois 1
+▁recommenc 1
+▁singulier 1
+▁souffrance 1
+▁surveillance 1
+▁théodore 1
+ã 1
+▁clandestin 1
+▁davril 1
+▁immigr 1
+▁orphelin 1
+▁pencroff 1
+▁dautriche 1
+▁dentreprise 1
+▁lombard 1
+▁classification 1
+▁emporté 1
+▁récepteur 1
+▁écologique 1
+▁lituanie 1
+▁létoile 1
+▁nerveuse 1
+▁seniors 1
+▁buisson 1
+▁succéda 1
+stronome 1
+▁saintjust 1
+▁basilique 1
+▁entretenu 1
+▁parasite 1
+▁génétique 1
+▁tablier 1
+▁tragique 1
+▁bavard 1
+▁surprend 1
+▁verticale 1
+percevoir 1
+▁chimique 1
+évidence 1
+▁attente 1
+▁facial 1
+▁malaisie 1
+▁ralenti 1
+▁paternel 1
+dministrateur 1
+préfecture 1
+▁absolu 1
+å 1
+▁compatriote 1
+▁confession 1
+▁doctrine 1
+▁galaxie 1
+▁gwynplaine 1
+▁hôpital 1
+▁inacceptable 1
+▁lappartement 1
+▁législative 1
+▁petitdéjeuner 1
+▁popularité 1
+▁reconverti 1
+▁renforcé 1
+▁saintdizier 1
+▁sarkozy 1
+▁simultanément 1
+▁thaïlande 1
+identifier 1
+▁improvis 1
+▁innovant 1
+▁lalcool 1
+▁orléans 1
+▁substance 1
+▁inaugur 1
+▁linventaire 1
+▁flèche 1
+▁limpôt 1
+ă 1
+ń 1
+æ 1
+ß 1
+ş 1
+ș 1
+ż 1
+ÿ 1
+ž 1
+ý 1
+ð 1
+ř 1
+ą 1
+ț 1
+ğ 1
+õ 1
+đ 1
+ő 1
+ò 1
+ē 1
+ī 1
+ě 1
+ņ 1
+ù 1
+ì 1
+ʻ 1
+ṣ 1
+ė 1
+1 1
+2 1
+ň 1
+ź 1
+ʾ 1
+ҫ 1
+⁄ 1
+ʉ 1
+π 1
+ṇ 1
+ả 1
+星 1
+ṯ 1
+τ 1
+κ 1
+δ 1
+þ 1
+R 1
+σ 1
+β 1
+α 1
+ŏ 1
+ʿ 1
+ś 1
+ġ 1
+ơ 1
+ư 1
+ṭ 1
+E 1
+N 1
+Z 1
+ħ 1
+ļ 1
+ť 1
+ų 1
+ǃ 1
+ǫ 1
+ǹ 1
+ʼ 1
+ː 1
+ζ 1
+ν 1
+υ 1
+ω 1
+а 1
+г 1
+е 1
+з 1
+и 1
+м 1
+н 1
+ደ 1
+ḍ 1
+ạ 1
+ầ 1
+ậ 1
+ễ 1
+ệ 1
+ị 1
+ồ 1
+ộ 1
+ợ 1
+ử 1
+ỳ 1
+い 1
+た 1
+つ 1
+乃 1
+扬 1
+术 1
+杜 1
+美 1
+馆 1
+ṅ 1
+ę 1
+京 1
+北 1
diff --git a/configs/fr-en/src_unigram6000/spm_unigram_fr.vocab b/configs/fr-en/src_unigram6000/spm_unigram_fr.vocab
new file mode 100644
index 0000000000000000000000000000000000000000..76067ee836869ee94094b27321a803ea20be5266
--- /dev/null
+++ b/configs/fr-en/src_unigram6000/spm_unigram_fr.vocab
@@ -0,0 +1,6000 @@
+ 0
+ 0
+ 0
+ 0
+s -3.02455
+▁de -3.33627
+▁la -3.71659
+▁le -3.97336
+▁à -4.22734
+e -4.31458
+▁est -4.33442
+▁il -4.35967
+▁et -4.51032
+▁les -4.60248
+r -4.7652
+▁en -4.80674
+▁du -4.81168
+▁a -4.8597
+▁un -4.87695
+▁des -4.96075
+t -5.09159
+▁au -5.21961
+▁une -5.23023
+▁dans -5.29303
+▁je -5.32391
+▁pour -5.33254
+▁se -5.37397
+▁pas -5.37937
+▁elle -5.39943
+▁que -5.43154
+▁cent -5.44697
+▁par -5.47833
+▁ce -5.52963
+▁ne -5.5334
+▁rue -5.56854
+▁sa -5.59106
+▁son -5.59443
+▁ -5.61195
+▁cest -5.6979
+a -5.70042
+▁vous -5.70357
+nt -5.73115
+é -5.74044
+n -5.76196
+▁sur -5.76324
+▁deux -5.77698
+▁sont -5.80446
+er -5.83495
+▁plus -5.8925
+es -5.89327
+y -5.89465
+c -5.95446
+ment -5.98226
+▁qui -5.98334
+▁cette -5.99554
+l -6.06036
+u -6.10984
+▁avec -6.11243
+▁nous -6.13131
+ait -6.17332
+▁monsieur -6.17972
+i -6.24017
+ra -6.25754
+▁ma -6.26691
+▁numéro -6.26903
+▁mille -6.27259
+▁on -6.27854
+ant -6.28218
+▁trois -6.35053
+d -6.39141
+le -6.4052
+▁mais -6.42164
+z -6.42196
+▁ses -6.42381
+▁y -6.42978
+re -6.4306
+▁été -6.44144
+o -6.44751
+▁quatre -6.47141
+m -6.47144
+▁fait -6.48847
+▁me -6.50447
+▁parole -6.50547
+▁comme -6.50731
+▁lui -6.53996
+▁bien -6.54703
+▁zéro -6.55635
+p -6.5703
+▁tu -6.58317
+ne -6.59314
+it -6.59996
+▁dé -6.6293
+te -6.63032
+▁tout -6.63152
+▁re -6.63904
+▁aux -6.64947
+in -6.67099
+▁était -6.67317
+la -6.67726
+g -6.67861
+il -6.68915
+▁ces -6.70993
+x -6.72412
+ée -6.73024
+▁également -6.74605
+f -6.76707
+▁dun -6.76832
+de -6.77288
+▁lamendement -6.776
+▁si -6.78281
+ri -6.78908
+▁aussi -6.78963
+ur -6.78971
+ent -6.78994
+lo -6.79344
+▁quil -6.79364
+▁mon -6.8
+▁donc -6.84161
+me -6.88104
+is -6.88345
+que -6.89706
+▁ou -6.90654
+▁soutenir -6.92433
+di -6.9342
+▁da -6.94693
+▁madame -6.95256
+li -6.95668
+on -6.95951
+▁alors -6.96066
+▁dune -6.96167
+en -6.96551
+▁même -6.9684
+sse -6.97225
+▁cinq -6.97251
+▁va -6.9799
+▁très -6.99665
+▁ont -6.99955
+ch -7.00218
+té -7.00787
+▁in -7.02055
+▁na -7.02591
+▁ça -7.02653
+▁é -7.02831
+▁jai -7.03364
+▁fut -7.03967
+b -7.04142
+ation -7.04529
+▁route -7.05731
+▁leur -7.06119
+v -7.09881
+aient -7.09882
+ti -7.10013
+ro -7.10528
+k -7.11594
+▁six -7.12149
+▁nest -7.12149
+na -7.12412
+▁cet -7.1287
+▁suis -7.14573
+▁ta -7.14626
+ma -7.15737
+▁être -7.15742
+▁cela -7.16609
+ta -7.17758
+▁ra -7.18573
+▁saint -7.19012
+ff -7.19372
+▁ils -7.19414
+▁avait -7.19417
+che -7.2097
+▁l -7.21003
+▁cents -7.21248
+▁cha -7.2218
+▁tous -7.22787
+▁ville -7.23489
+▁temps -7.24127
+ci -7.24747
+▁peu -7.25685
+h -7.26252
+▁chemin -7.26284
+▁p -7.26423
+▁vingt -7.27454
+ce -7.2749
+▁après -7.2841
+▁plusieurs -7.28581
+▁con -7.28769
+▁d -7.29501
+▁nom -7.29667
+▁entre -7.29851
+▁ca -7.30221
+▁huit -7.30285
+▁sans -7.30633
+▁sept -7.31973
+ge -7.32525
+▁faire -7.32682
+ie -7.32694
+al -7.33703
+▁dix -7.34134
+▁votre -7.34211
+▁lé -7.34483
+▁peut -7.34498
+▁ré -7.35285
+nd -7.35584
+vi -7.35781
+▁pro -7.35855
+▁c -7.36697
+▁porte -7.36908
+▁moi -7.36993
+▁co -7.37655
+ni -7.37832
+mi -7.3901
+ons -7.39355
+lé -7.39788
+▁dit -7.40347
+▁sous -7.40568
+ve -7.41519
+an -7.41657
+da -7.42228
+va -7.42834
+▁ba -7.43114
+▁amendement -7.43471
+ine -7.43493
+▁partie -7.43667
+▁di -7.44063
+▁quelle -7.4455
+▁car -7.44669
+▁encore -7.44994
+▁trente -7.45048
+tte -7.45078
+▁premier -7.45699
+age -7.4782
+at -7.50932
+co -7.52342
+▁te -7.52487
+mp -7.52611
+ité -7.53468
+▁mar -7.53524
+ton -7.53855
+▁pa -7.54149
+us -7.54236
+ier -7.54516
+▁père -7.54532
+▁là -7.55278
+rai -7.55321
+pe -7.55437
+ba -7.55883
+▁première -7.56202
+▁grand -7.56212
+▁s -7.56404
+èrent -7.56492
+au -7.57241
+▁non -7.5735
+▁où -7.57732
+ement -7.5904
+ré -7.59069
+▁li -7.59184
+▁e -7.59323
+ion -7.59358
+▁ainsi -7.59904
+▁bo -7.60108
+▁avenue -7.60212
+▁place -7.60568
+▁bon -7.60975
+▁groupe -7.61689
+or -7.61744
+ca -7.61939
+ter -7.62223
+ga -7.62551
+▁ro -7.6271
+▁g -7.63929
+ées -7.64386
+tre -7.65603
+th -7.65768
+ll -7.6581
+be -7.66371
+▁ni -7.66825
+les -7.6691
+ez -7.67313
+ais -7.67571
+▁pi -7.68243
+lu -7.68639
+▁po -7.68808
+▁quand -7.7013
+▁tour -7.70261
+▁ja -7.70476
+▁i -7.70504
+lle -7.70833
+▁neuf -7.71219
+▁o -7.71464
+▁toujours -7.71821
+▁jean -7.71941
+aire -7.71959
+▁quelques -7.72149
+el -7.72165
+▁puis -7.72327
+bi -7.72357
+ard -7.72885
+no -7.73102
+ance -7.73466
+▁paris -7.73565
+go -7.74342
+▁pré -7.74452
+▁or -7.74511
+ille -7.74942
+▁famille -7.74992
+▁vi -7.7502
+tu -7.75353
+▁dis -7.75402
+▁trouve -7.75587
+▁commune -7.76179
+aux -7.76659
+▁ar -7.77155
+▁faut -7.77371
+▁fois -7.77452
+▁quarante -7.78073
+▁ensuite -7.78174
+ar -7.78355
+▁cinquante -7.78443
+▁fils -7.78652
+▁vers -7.78899
+ique -7.7891
+eur -7.78939
+ir -7.78976
+et -7.79033
+to -7.79748
+▁mont -7.79757
+rie -7.79933
+▁aujourdhui -7.80039
+▁lo -7.8025
+▁contre -7.80499
+▁france -7.81683
+és -7.81912
+gue -7.82092
+▁leurs -7.82121
+ac -7.82166
+▁ga -7.82291
+▁grande -7.82612
+ng -7.82931
+▁fa -7.83018
+vient -7.83218
+▁sé -7.8322
+able -7.83376
+ck -7.83491
+né -7.8353
+▁rien -7.83989
+▁eu -7.84016
+▁reste -7.84629
+elle -7.84708
+um -7.84773
+▁do -7.85074
+▁depuis -7.85455
+▁sagit -7.85559
+ry -7.8573
+▁soixante -7.86143
+▁lin -7.86353
+▁mi -7.86799
+▁pendant -7.87056
+▁fin -7.87165
+▁ans -7.87356
+do -7.87778
+mé -7.88389
+▁cependant -7.8848
+gi -7.88583
+ou -7.88737
+rent -7.89469
+▁mal -7.89594
+▁pays -7.89653
+▁sera -7.89685
+▁ah -7.89961
+iste -7.9053
+▁sou -7.90553
+vo -7.91053
+ul -7.91095
+ot -7.91181
+▁ri -7.91432
+▁quel -7.91797
+pp -7.9189
+ure -7.91892
+▁monde -7.92399
+▁b -7.92625
+▁ex -7.92631
+ron -7.92892
+▁bi -7.93124
+▁mort -7.93192
+ville -7.93521
+du -7.93559
+▁mé -7.93588
+po -7.93905
+tra -7.94093
+▁beaucoup -7.9417
+▁al -7.94314
+▁jamais -7.94386
+▁point -7.94449
+tion -7.94581
+▁oui -7.94742
+▁vie -7.94763
+▁gouvernement -7.9559
+▁ju -7.95859
+▁maison -7.96438
+▁guerre -7.96492
+▁sil -7.96922
+▁comment -7.97085
+mo -7.97589
+▁moins -7.97632
+▁avant -7.97693
+▁tard -7.97842
+▁es -7.98003
+ver -7.98424
+▁cas -7.98692
+pa -7.99085
+▁femme -7.99117
+ê -7.99189
+▁petit -7.99258
+▁fille -7.99262
+▁fais -7.99325
+▁rencontre -7.99414
+▁fi -7.99491
+sur -7.99576
+w -7.9981
+gu -7.99874
+ction -8.00237
+ux -8.00657
+▁sûr -8.00894
+▁ny -8.01729
+▁trop -8.01806
+ph -8.01882
+▁k -8.01891
+man -8.02113
+▁mère -8.02496
+▁lu -8.02813
+▁général -8.03046
+rait -8.0423
+im -8.04412
+▁pu -8.04707
+▁per -8.04935
+▁doit -8.05328
+▁ha -8.05522
+▁avoir -8.0574
+▁ici -8.0599
+if -8.06003
+ia -8.06358
+▁cor -8.06931
+▁jour -8.07329
+▁bou -8.07391
+ing -8.07437
+▁forme -8.07563
+▁m -8.0802
+▁quatrevingtdix -8.081
+ha -8.08196
+▁notre -8.08268
+▁travail -8.0833
+eux -8.08333
+til -8.08355
+bo -8.08796
+dé -8.09005
+▁lors -8.09433
+▁ministre -8.09737
+ette -8.09855
+▁homme -8.10238
+▁toute -8.10353
+▁su -8.10395
+▁lun -8.10697
+▁compte -8.10983
+▁chose -8.1134
+ère -8.11403
+▁autres -8.11603
+ru -8.11868
+▁pierre -8.12206
+▁né -8.12283
+▁dautres -8.12533
+▁français -8.12537
+▁située -8.12638
+▁mes -8.12638
+▁toutes -8.12671
+▁dor -8.12683
+▁bonne -8.12771
+▁commission -8.12955
+▁voir -8.1302
+▁port -8.13212
+▁voilà -8.13348
+ï -8.13478
+▁avez -8.13483
+▁village -8.139
+ale -8.13941
+▁étaient -8.142
+use -8.1426
+▁cétait -8.14411
+▁club -8.14537
+▁jeune -8.14617
+▁autre -8.15106
+▁soixantedix -8.15196
+▁part -8.15318
+isme -8.16421
+▁cou -8.16435
+▁souvent -8.16576
+▁fond -8.1658
+▁coup -8.16674
+▁fer -8.16986
+ront -8.17176
+gé -8.17716
+▁sest -8.17999
+ière -8.18188
+don -8.18753
+▁be -8.18772
+▁elles -8.18783
+qu -8.18936
+min -8.19522
+ka -8.19659
+▁assez -8.19726
+▁joue -8.1979
+▁espèce -8.20073
+▁droit -8.2013
+▁onze -8.20155
+rou -8.20274
+son -8.20296
+▁ton -8.20323
+▁jours -8.2063
+▁so -8.20936
+cu -8.21064
+▁gra -8.21065
+▁situé -8.21474
+▁mo -8.2187
+▁chez -8.21939
+ut -8.21947
+▁nos -8.22055
+▁questce -8.22127
+ise -8.22206
+eau -8.22387
+▁loi -8.22446
+pé -8.22549
+▁demande -8.22745
+▁pourquoi -8.22991
+qui -8.23039
+cha -8.2375
+▁men -8.23801
+tique -8.2391
+▁sud -8.24157
+▁vit -8.243
+mar -8.24301
+▁donne -8.24418
+▁ru -8.24475
+bu -8.24581
+▁suite -8.24929
+ateur -8.25187
+▁nord -8.25405
+▁dire -8.25422
+▁impasse -8.2548
+ran -8.25839
+▁comp -8.25883
+▁lieu -8.25964
+▁quelque -8.2598
+▁enfants -8.26021
+▁bar -8.26253
+▁pe -8.26503
+▁léglise -8.26523
+▁sœur -8.26523
+un -8.26712
+▁nouveau -8.26853
+▁seul -8.26863
+▁plan -8.27059
+▁toi -8.27628
+▁ho -8.27735
+▁président -8.27907
+▁frère -8.28072
+▁beau -8.28175
+▁train -8.28213
+▁char -8.28215
+▁dont -8.28334
+▁prend -8.28381
+▁vo -8.28454
+▁tête -8.2849
+▁titre -8.28533
+mont -8.28622
+▁fort -8.28751
+▁douze -8.28964
+▁cour -8.28975
+▁no -8.29245
+▁région -8.29258
+ol -8.29275
+▁moment -8.29839
+▁jo -8.29977
+▁nombreux -8.30367
+▁déjà -8.30496
+▁château -8.30836
+▁val -8.30861
+▁notamment -8.31122
+▁politique -8.31291
+vé -8.31352
+▁ver -8.3162
+chi -8.32214
+▁allée -8.32298
+tin -8.32488
+lan -8.32702
+▁mot -8.3278
+con -8.32806
+▁cours -8.32948
+urs -8.33037
+là -8.33107
+je -8.33418
+▁parti -8.33488
+▁sen -8.33803
+▁dernier -8.33833
+ence -8.33835
+cé -8.33979
+▁mari -8.34016
+▁quinze -8.34472
+op -8.34527
+ions -8.34586
+▁an -8.34873
+mer -8.35079
+▁quon -8.35163
+▁petite -8.35234
+▁effet -8.35298
+▁bois -8.3555
+nes -8.35614
+ger -8.35633
+▁avons -8.35697
+om -8.35977
+▁gi -8.36073
+rons -8.36077
+ens -8.36245
+▁luniversité -8.36276
+▁mou -8.36345
+▁dar -8.36704
+ssa -8.36907
+ive -8.36922
+▁bri -8.36999
+▁connu -8.37193
+tes -8.37328
+▁maintenant -8.37463
+ché -8.37755
+▁sais -8.37816
+▁avis -8.37909
+▁go -8.38094
+▁heures -8.38158
+land -8.38194
+ki -8.38626
+▁siège -8.38796
+▁présente -8.39257
+hi -8.39286
+è -8.3937
+▁amendements -8.39645
+▁st -8.39654
+▁carrière -8.39802
+ble -8.39985
+mu -8.40057
+ire -8.40058
+▁gu -8.40088
+▁fit -8.40225
+ienne -8.40294
+▁mois -8.40355
+▁soit -8.40676
+bre -8.40759
+illa -8.4097
+▁commence -8.41022
+▁ti -8.41351
+▁quoi -8.41354
+tif -8.41375
+ner -8.41812
+▁année -8.42707
+▁saison -8.42766
+▁vu -8.42776
+os -8.42838
+▁ai -8.42961
+â -8.43039
+▁tant -8.43174
+▁passe -8.4331
+▁travaille -8.43403
+des -8.43452
+▁terre -8.43535
+▁vais -8.43656
+ide -8.4366
+ja -8.43781
+▁long -8.43801
+par -8.43813
+▁bas -8.44076
+▁cro -8.44106
+▁crois -8.44358
+▁centre -8.44507
+▁serait -8.44667
+▁certains -8.44767
+ante -8.44776
+▁vol -8.44856
+▁ka -8.44999
+▁personne -8.4523
+mes -8.45353
+▁nai -8.45912
+▁ami -8.46513
+▁ch -8.46519
+▁larticle -8.46524
+nc -8.4664
+▁estce -8.46801
+▁durant -8.47167
+▁nouvelle -8.47242
+▁près -8.47248
+▁cher -8.47492
+▁peuvent -8.47509
+▁défavorable -8.47628
+▁cri -8.47818
+▁quatorze -8.47998
+ations -8.48098
+euse -8.48187
+▁vrai -8.48337
+za -8.48344
+▁bu -8.48358
+▁roi -8.48548
+ze -8.48863
+▁cu -8.49157
+▁seule -8.49227
+▁col -8.49409
+bl -8.49425
+▁semble -8.49429
+▁mor -8.49556
+▁film -8.49611
+▁louis -8.49723
+▁prendre -8.49755
+ois -8.49811
+▁aller -8.49909
+▁raison -8.50155
+▁important -8.50229
+▁conseil -8.50426
+let -8.50529
+and -8.50573
+tait -8.5095
+▁lavis -8.5099
+gra -8.51145
+▁pont -8.51252
+ig -8.51283
+▁présent -8.51405
+▁seize -8.51591
+dre -8.51812
+▁t -8.52136
+▁côté -8.52363
+▁projet -8.52365
+▁gens -8.5252
+▁ci -8.52522
+▁début -8.52565
+▁deuxième -8.5259
+vent -8.52619
+▁lautre -8.52771
+▁mer -8.52775
+▁fo -8.5285
+rez -8.52853
+lli -8.5309
+▁dernière -8.53548
+ès -8.53573
+▁oh -8.53597
+val -8.53729
+cou -8.53974
+▁acc -8.54143
+▁celle -8.54156
+▁soir -8.54243
+dou -8.54586
+▁aimé -8.54787
+▁ligne -8.5483
+▁chambre -8.54974
+▁poste -8.55024
+fi -8.55102
+▁cla -8.55204
+▁permet -8.55426
+▁man -8.55531
+▁haut -8.55611
+▁font -8.55641
+ient -8.55903
+▁mieux -8.56001
+cul -8.5614
+▁question -8.56358
+▁nombre -8.56372
+▁devant -8.56414
+▁enfin -8.56707
+gne -8.5692
+▁voix -8.5695
+▁inter -8.57254
+▁lécole -8.57335
+▁furent -8.57413
+▁léquipe -8.57539
+▁aime -8.57604
+lie -8.57613
+pi -8.5774
+vingtdixsept -8.57955
+▁voie -8.57996
+ite -8.58053
+▁main -8.58089
+▁écrit -8.58187
+▁fu -8.58224
+▁problème -8.58359
+▁table -8.58378
+ven -8.58396
+▁série -8.58733
+teur -8.58761
+▁sy -8.58809
+lon -8.5882
+▁gauche -8.5898
+▁jusquà -8.5898
+ô -8.59174
+▁produit -8.59395
+▁sujet -8.59604
+▁nombreuses -8.59604
+▁combien -8.59609
+ezvous -8.59858
+qua -8.59947
+tation -8.60092
+ler -8.60096
+▁mur -8.60101
+▁din -8.60215
+▁he -8.60233
+▁dabord -8.6046
+▁hu -8.60545
+▁treize -8.61073
+▁droite -8.6109
+vers -8.61126
+▁selon -8.61157
+ris -8.61273
+▁dan -8.61297
+▁jeu -8.61413
+▁met -8.61439
+ic -8.61666
+ct -8.6177
+ya -8.61825
+▁département -8.61946
+tta -8.62104
+▁membre -8.62259
+▁mesure -8.62301
+ue -8.62649
+iers -8.62785
+▁venu -8.62863
+▁études -8.62904
+▁aucun -8.62906
+ai -8.6298
+vin -8.63127
+ho -8.63141
+aud -8.63187
+▁mis -8.63193
+▁vent -8.63358
+onne -8.63522
+▁garde -8.63541
+▁paul -8.63688
+lin -8.64238
+▁bé -8.64318
+▁tra -8.64396
+▁mu -8.64585
+▁face -8.64685
+▁êtes -8.64737
+▁dêtre -8.64758
+▁comprend -8.64773
+▁sens -8.64831
+▁fou -8.64901
+▁possède -8.64957
+bé -8.65195
+▁existe -8.65239
+▁quatrevingts -8.65383
+▁lhistoire -8.65398
+▁dal -8.65572
+▁sommes -8.65623
+▁regard -8.65649
+▁bis -8.65715
+▁lac -8.65891
+▁gar -8.65901
+▁dixhuit -8.66078
+ix -8.66244
+▁ve -8.66329
+▁vos -8.66395
+fa -8.66406
+▁lou -8.66434
+aires -8.66771
+▁pé -8.66817
+▁roman -8.66842
+▁champ -8.66961
+▁eux -8.6711
+▁gé -8.67495
+▁matin -8.67938
+▁ten -8.67964
+▁livre -8.67977
+▁ber -8.68002
+ttes -8.68117
+▁pense -8.6829
+▁eh -8.68421
+gan -8.68432
+▁peutêtre -8.68789
+ud -8.6922
+cher -8.69236
+▁tri -8.69283
+▁dieu -8.69482
+car -8.69525
+ad -8.69676
+▁dès -8.69775
+▁qua -8.69933
+▁yeux -8.6999
+▁lor -8.701
+istes -8.70112
+ko -8.70329
+▁seconde -8.70363
+▁parfois -8.70383
+▁surtout -8.70384
+gre -8.70585
+▁française -8.70662
+▁bra -8.70738
+tier -8.709
+isation -8.7091
+ang -8.71089
+mon -8.71329
+▁nuit -8.71434
+▁celui -8.71456
+▁territoire -8.71551
+tri -8.71587
+▁boulevard -8.71786
+▁système -8.72022
+▁retire -8.72048
+▁sol -8.72368
+▁bord -8.7244
+▁dou -8.72481
+illes -8.72602
+▁rouge -8.72633
+▁musique -8.72736
+▁succès -8.72737
+ins -8.72876
+ke -8.72902
+isé -8.72994
+▁tel -8.73035
+▁années -8.73105
+the -8.73113
+sta -8.73144
+▁participe -8.7321
+erie -8.7339
+ture -8.73464
+▁utilisé -8.73467
+▁marie -8.73477
+men -8.73517
+elles -8.73634
+der -8.74142
+▁chi -8.74453
+nne -8.74498
+▁belle -8.74519
+zvous -8.74554
+nu -8.74753
+▁loin -8.74792
+▁dép -8.74816
+wa -8.74918
+▁étatsunis -8.7514
+▁nationale -8.75328
+nie -8.75461
+▁rou -8.75672
+▁aurait -8.75737
+▁nen -8.75814
+▁rapport -8.75844
+vre -8.75915
+▁avaient -8.75969
+bert -8.76251
+illon -8.76263
+nce -8.76437
+▁sort -8.76561
+▁parle -8.76664
+▁nommé -8.76732
+▁scène -8.76869
+▁rôle -8.77108
+▁blanc -8.77245
+▁vient -8.77429
+vais -8.77441
+cy -8.77505
+▁choses -8.77521
+oir -8.7757
+▁national -8.7785
+▁veux -8.77874
+▁ren -8.77908
+▁charles -8.77948
+▁hi -8.78121
+ndre -8.78182
+▁heure -8.78264
+▁propose -8.78368
+▁voiture -8.7841
+igne -8.78428
+port -8.78502
+▁dixsept -8.78609
+▁travers -8.78658
+▁j -8.78806
+▁quant -8.78842
+bri -8.78971
+▁létat -8.79119
+▁attend -8.79239
+▁jen -8.7932
+lot -8.79333
+▁lim -8.7936
+▁pouvoir -8.79383
+▁tombe -8.79803
+▁public -8.79808
+▁société -8.79879
+▁maître -8.79881
+▁joueur -8.79989
+▁actuellement -8.802
+ain -8.80281
+▁docteur -8.80391
+▁rapidement -8.80474
+▁pla -8.80629
+▁toutefois -8.80648
+▁quartier -8.80703
+▁noir -8.80711
+▁bâtiment -8.80711
+▁har -8.8075
+ore -8.80763
+▁gre -8.81431
+tal -8.81577
+jo -8.8165
+ques -8.81765
+▁personnes -8.81893
+▁f -8.8229
+▁veut -8.82503
+▁cause -8.82777
+▁rend -8.82798
+▁lon -8.82985
+▁professeur -8.82992
+▁lîle -8.82992
+▁rivière -8.83
+▁permett -8.83127
+née -8.83162
+▁troisième -8.83164
+lier -8.83293
+▁juste -8.83308
+▁prix -8.83374
+▁pris -8.83502
+fort -8.83533
+ième -8.83837
+▁ad -8.83983
+▁dixneuf -8.84335
+bou -8.8436
+ologie -8.84484
+pin -8.84578
+van -8.84581
+▁pied -8.84601
+▁françois -8.84668
+▁partir -8.84782
+▁vingttrois -8.84853
+▁chef -8.85067
+▁u -8.8527
+▁but -8.85357
+▁dailleurs -8.85392
+▁ab -8.85418
+▁bal -8.85435
+▁ko -8.8544
+▁imp -8.85516
+ien -8.85527
+▁lis -8.85715
+▁nu -8.85852
+▁pri -8.85861
+▁sainte -8.86171
+▁aucune -8.86327
+▁site -8.86411
+resse -8.86559
+▁mouvement -8.86751
+▁milieu -8.86999
+fer -8.87052
+▁enfant -8.8714
+ole -8.87181
+gon -8.87195
+▁longtemps -8.87299
+▁certaines -8.87325
+▁sait -8.8733
+▁hommes -8.87335
+moi -8.87407
+▁montre -8.87506
+voir -8.8751
+▁dau -8.87514
+ste -8.87536
+com -8.87598
+bourg -8.87628
+▁lannée -8.8786
+▁r -8.87873
+▁autour -8.87877
+▁commun -8.88046
+jou -8.88073
+id -8.8813
+▁couleur -8.88432
+▁quun -8.88476
+▁marque -8.88574
+▁appelé -8.88648
+▁genre -8.89058
+▁z -8.89098
+î -8.89241
+▁donné -8.89327
+eurs -8.89435
+▁article -8.89525
+▁cap -8.89585
+▁celuici -8.89611
+pu -8.89836
+▁vingtdeux -8.90089
+▁environ -8.90092
+▁corps -8.90105
+▁lune -8.90198
+▁parmi -8.90492
+▁fe -8.90606
+▁luimême -8.90656
+▁simple -8.90674
+▁sch -8.90722
+enne -8.90846
+▁besoin -8.90941
+▁anglais -8.90941
+▁cal -8.91118
+▁combat -8.91128
+ly -8.91173
+▁chien -8.91304
+▁femmes -8.9137
+rd -8.91383
+▁exemple -8.91513
+▁doute -8.91707
+nier -8.91721
+▁pan -8.91737
+ç -8.91789
+▁vingtcinq -8.92089
+▁pourtant -8.92142
+▁can -8.92154
+▁grands -8.92267
+▁ki -8.92304
+▁texte -8.92379
+nou -8.92413
+▁arrive -8.92415
+▁vite -8.92462
+tour -8.92508
+▁rapporteur -8.9254
+▁pratique -8.92655
+▁meurt -8.92764
+aine -8.9279
+ges -8.92964
+▁bureau -8.9298
+am -8.93071
+▁chacun -8.93254
+▁vieux -8.93453
+▁importante -8.93578
+▁moyen -8.93608
+ières -8.93682
+▁langue -8.93703
+▁fini -8.93826
+▁type -8.9386
+▁nétait -8.93883
+he -8.93941
+▁continue -8.94262
+▁peine -8.94277
+rin -8.94327
+ju -8.94346
+▁niveau -8.946
+▁canton -8.94786
+▁débat -8.94787
+▁omer -8.94808
+▁suivante -8.94813
+▁mondiale -8.94859
+▁plait -8.94881
+▁demanda -8.95055
+▁parler -8.95183
+sie -8.95271
+▁possible -8.95316
+▁trans -8.95765
+lect -8.95838
+dor -8.9598
+▁vert -8.96146
+▁pauvre -8.96213
+ban -8.96215
+▁demain -8.96247
+▁parc -8.96335
+▁perdu -8.96363
+▁seulement -8.96428
+▁musée -8.96528
+▁bru -8.96622
+▁zone -8.96634
+qué -8.9669
+▁fonction -8.96782
+ral -8.9681
+▁publique -8.96816
+▁figure -8.96819
+▁john -8.97122
+ber -8.97238
+ves -8.97238
+ène -8.97289
+cal -8.97305
+▁période -8.97424
+▁service -8.97488
+iques -8.97579
+▁bras -8.97584
+▁semaine -8.97734
+▁propre -8.97824
+▁chant -8.98114
+▁bout -8.98121
+ieux -8.98463
+alité -8.98774
+▁jardin -8.98958
+▁vin -8.98962
+▁décide -8.99045
+j -8.99065
+▁certain -8.99087
+▁eut -8.99089
+▁date -8.99123
+▁finalement -8.99155
+ât -8.99204
+stu -8.9923
+▁recherche -8.99267
+▁habitants -8.99275
+▁allez -8.99318
+▁min -8.99319
+▁saisi -8.99321
+ssé -8.99329
+▁sta -8.99403
+▁bre -8.99474
+fe -8.99547
+▁devenu -8.99617
+▁capitaine -8.99889
+▁riche -9.00027
+▁longue -9.00196
+▁dessin -9.00234
+cent -9.00256
+▁remporte -9.00272
+▁laisse -9.00367
+ster -9.00411
+▁w -9.0042
+▁to -9.00528
+▁double -9.00545
+lla -9.00749
+▁militaire -9.00829
+naire -9.01113
+▁froid -9.01146
+raient -9.01147
+▁domaine -9.01151
+▁san -9.01193
+▁formation -9.01243
+▁voit -9.01355
+éré -9.01383
+▁vingtsix -9.01466
+▁principale -9.01594
+▁cré -9.01788
+▁ge -9.01883
+bon -9.019
+▁quune -9.02038
+▁reçu -9.02099
+▁camp -9.02417
+▁fl -9.0243
+io -9.02644
+▁principalement -9.02717
+▁jaune -9.02744
+▁ferme -9.02849
+▁vingtquatre -9.02922
+bel -9.03049
+▁cheflieu -9.03059
+ines -9.03124
+▁passage -9.03408
+▁cœur -9.0342
+gent -9.03584
+▁passer -9.03626
+tel -9.03658
+ny -9.03691
+▁deau -9.03744
+▁difficile -9.0403
+▁province -9.0403
+▁app -9.04079
+▁for -9.04084
+mme -9.04156
+▁propos -9.04174
+▁nommée -9.04418
+form -9.04433
+▁op -9.04455
+▁ceux -9.04609
+▁mission -9.047
+▁n -9.04831
+▁thé -9.04872
+leur -9.04933
+▁v -9.04976
+▁perd -9.04989
+ay -9.05167
+ule -9.05199
+▁henri -9.0534
+▁juge -9.05358
+▁équipe -9.05437
+▁lh -9.05516
+tan -9.0553
+▁généralement -9.0561
+▁représente -9.05617
+▁vingtsept -9.05669
+non -9.05811
+▁batt -9.05938
+illé -9.05953
+▁fre -9.05984
+▁théâtre -9.06007
+▁venir -9.06025
+▁fr -9.06166
+▁second -9.06187
+nay -9.06218
+mie -9.06271
+▁style -9.06334
+▁situe -9.06353
+▁vé -9.06373
+▁sal -9.0654
+▁voyage -9.06636
+▁championnat -9.0667
+▁hier -9.0679
+▁roche -9.0688
+▁révolution -9.07007
+▁gare -9.07265
+▁feu -9.0739
+teurs -9.07466
+▁retrouve -9.07527
+▁situation -9.07673
+▁ailleurs -9.07673
+▁plupart -9.07676
+nelle -9.08071
+▁faites -9.08226
+ret -9.08255
+▁lhonneur -9.08349
+▁étudie -9.08349
+▁sub -9.08637
+ssent -9.0864
+▁sc -9.08658
+▁sp -9.0866
+ade -9.08663
+▁gou -9.08747
+▁meilleur -9.08831
+gn -9.09048
+▁habit -9.09069
+spir -9.09105
+gh -9.09347
+ura -9.09389
+▁comporte -9.09412
+▁new -9.09503
+mor -9.09541
+▁parents -9.09592
+▁identiques -9.09664
+rée -9.09762
+vou -9.09811
+court -9.0984
+▁choisi -9.10079
+▁lettre -9.10097
+▁passé -9.10128
+ric -9.10143
+▁hé -9.10211
+▁gui -9.10359
+▁trentetrois -9.10408
+ji -9.10468
+saint -9.10778
+▁entend -9.10779
+▁vingthuit -9.1078
+cor -9.10801
+▁proche -9.10835
+quelle -9.10876
+▁larmée -9.11105
+telle -9.1119
+▁presque -9.1119
+▁donner -9.1137
+▁fla -9.11603
+ras -9.11619
+▁capitale -9.11643
+up -9.11736
+▁peux -9.11796
+▁ensemble -9.11804
+▁album -9.11804
+▁mettre -9.11891
+▁terme -9.12034
+ît -9.1204
+vis -9.1216
+▁vingtneuf -9.1218
+zi -9.12328
+▁ob -9.1258
+xi -9.12597
+▁ph -9.12649
+▁lotissement -9.12864
+▁côte -9.12864
+▁jou -9.12877
+▁peur -9.13001
+wi -9.13043
+ages -9.13153
+▁favorable -9.1322
+mou -9.13232
+▁lair -9.13317
+▁navait -9.13406
+▁journal -9.13484
+eaux -9.13498
+▁nécessaire -9.13577
+▁pouvait -9.13698
+▁merci -9.13729
+▁chanson -9.13742
+quet -9.138
+▁telle -9.13822
+▁ancienne -9.13866
+▁population -9.13935
+▁large -9.14092
+lais -9.1439
+▁salle -9.14488
+▁membres -9.14523
+iens -9.14541
+▁am -9.14606
+▁historique -9.14656
+▁lépoque -9.14656
+▁libre -9.14763
+▁match -9.14942
+our -9.14956
+▁all -9.1497
+▁nouvelles -9.15135
+yer -9.1543
+ches -9.15623
+▁marche -9.15669
+ty -9.15677
+▁lensemble -9.15747
+char -9.15774
+étant -9.15958
+▁court -9.16213
+▁vraiment -9.16245
+▁organisé -9.16481
+▁livres -9.16547
+rre -9.16561
+▁bel -9.16603
+vons -9.16705
+vint -9.16723
+▁différents -9.1685
+▁façon -9.1685
+▁sac -9.16896
+▁philippe -9.16911
+duc -9.17022
+mit -9.17119
+ères -9.17135
+col -9.17148
+ales -9.17173
+▁néanmoins -9.1722
+▁retour -9.17224
+oire -9.17316
+din -9.1738
+can -9.17489
+ros -9.17512
+▁lheure -9.17593
+▁cho -9.17658
+▁aura -9.17942
+▁haute -9.17993
+▁déco -9.18029
+▁vois -9.18047
+lis -9.18225
+gie -9.18256
+tions -9.18329
+▁fenêtre -9.1834
+▁plutôt -9.1834
+▁trentequatre -9.1834
+▁trentedeux -9.1834
+▁député -9.18341
+▁croix -9.18376
+iser -9.18386
+▁ancien -9.18444
+▁michel -9.18531
+▁manière -9.18603
+sson -9.18658
+▁signe -9.18673
+▁cimetière -9.18716
+▁super -9.18721
+▁anti -9.18972
+▁lex -9.19039
+ice -9.19084
+▁source -9.19096
+▁trentecinq -9.19472
+▁petits -9.19485
+▁tiens -9.19559
+▁qu -9.19561
+ani -9.19602
+▁jeunes -9.19611
+▁principal -9.19715
+▁communes -9.19719
+▁len -9.19798
+▁ter -9.19882
+den -9.19883
+berg -9.19895
+venir -9.20097
+▁trouver -9.20196
+▁endémique -9.20234
+▁visite -9.20299
+▁dame -9.20321
+▁déc -9.20329
+▁charge -9.20347
+▁savoir -9.20367
+hon -9.20419
+▁position -9.20582
+▁état -9.20626
+tti -9.2085
+▁appel -9.20907
+▁h -9.20942
+▁différentes -9.21001
+▁originaire -9.21001
+▁travaux -9.21008
+res -9.21104
+▁premiers -9.21138
+art -9.2131
+▁rejoint -9.21387
+▁production -9.21388
+▁bleu -9.21401
+▁etait -9.21498
+gen -9.21626
+▁ouvert -9.21645
+tz -9.2174
+bes -9.21792
+▁mains -9.22073
+▁force -9.22123
+▁cru -9.22149
+▁allons -9.22243
+èle -9.22443
+▁pal -9.22511
+han -9.22561
+▁idée -9.22567
+hé -9.22641
+net -9.22674
+▁termine -9.2295
+▁division -9.22967
+igné -9.22973
+▁créé -9.23003
+▁peuple -9.23347
+▁breton -9.23427
+▁forte -9.23515
+▁lit -9.23565
+▁connue -9.23642
+▁comte -9.23778
+ria -9.23782
+▁bro -9.23805
+▁œuvres -9.23822
+▁her -9.23871
+▁voisin -9.23955
+gène -9.2397
+spect -9.24047
+près -9.2408
+▁station -9.24356
+nant -9.24494
+▁compétition -9.24531
+▁art -9.24547
+▁im -9.24731
+▁lai -9.24754
+▁dispose -9.24907
+▁dorigine -9.24931
+ica -9.25254
+▁leau -9.25282
+▁malgré -9.25333
+mba -9.25549
+lement -9.25618
+í -9.25736
+tiques -9.2577
+▁évolue -9.25799
+▁robert -9.25824
+▁version -9.25826
+▁jacques -9.25834
+iv -9.25903
+vie -9.25919
+▁vend -9.26086
+▁suivi -9.26146
+▁fau -9.26206
+tor -9.26209
+▁construit -9.26437
+▁pen -9.26488
+▁fil -9.26543
+▁désormais -9.26547
+▁élu -9.26588
+▁cé -9.26632
+▁celleci -9.26698
+▁maisons -9.26714
+▁vue -9.26714
+▁inf -9.26787
+▁che -9.26884
+▁four -9.2691
+▁martin -9.26943
+▁inscrit -9.26958
+roi -9.27695
+▁gros -9.27754
+▁victoire -9.27777
+▁directeur -9.27813
+logique -9.2786
+ète -9.27887
+hu -9.27947
+▁ram -9.28012
+▁vont -9.281
+bit -9.2811
+pt -9.28127
+position -9.2816
+▁quarantecinq -9.2819
+▁détat -9.28202
+▁dépend -9.28216
+▁malade -9.28247
+▁péri -9.28536
+▁descend -9.28539
+ü -9.28605
+▁cinéma -9.28605
+▁construction -9.28605
+▁dh -9.28712
+▁soi -9.28854
+▁don -9.28886
+▁pêche -9.29022
+▁issu -9.29058
+▁chapelle -9.29074
+▁base -9.29115
+▁lordre -9.29127
+nel -9.2916
+igny -9.29258
+▁frères -9.29383
+ina -9.29432
+▁pres -9.29476
+dit -9.29616
+▁wa -9.2978
+▁clair -9.29782
+▁marché -9.29794
+▁particulier -9.2986
+▁contraire -9.2986
+▁rare -9.29937
+▁naît -9.29969
+▁considéré -9.30062
+▁bataille -9.30283
+range -9.30423
+▁mai -9.30436
+▁chaud -9.30509
+chant -9.30582
+gence -9.30613
+ab -9.30636
+▁carte -9.30663
+▁allemand -9.30809
+▁jaime -9.30878
+tho -9.31052
+▁jeux -9.31127
+▁histoire -9.31131
+▁publié -9.31131
+▁blanche -9.31161
+▁plante -9.31234
+▁finale -9.31248
+▁avezvous -9.31264
+▁direction -9.31284
+▁arrivé -9.31333
+▁mariage -9.31353
+we -9.31372
+tive -9.3141
+érie -9.31465
+▁communauté -9.31559
+iez -9.31761
+▁critique -9.31828
+▁suivant -9.31854
+▁bande -9.31862
+▁ban -9.31867
+▁rose -9.31918
+▁particulière -9.31928
+▁classe -9.32017
+▁sein -9.32043
+▁dentre -9.32056
+dy -9.32086
+▁reprend -9.32424
+▁louest -9.3244
+▁disposition -9.3247
+pri -9.32502
+▁my -9.32514
+▁manque -9.32625
+▁rep -9.32721
+del -9.32768
+▁ben -9.32777
+▁liste -9.32777
+▁utilisée -9.331
+▁médecin -9.33204
+▁mange -9.33283
+▁forêt -9.333
+udi -9.33425
+▁marc -9.33506
+cro -9.33508
+ley -9.33553
+rac -9.33575
+▁grandes -9.33603
+▁bla -9.33607
+▁blo -9.33662
+▁campagne -9.33733
+▁utilise -9.3388
+▁garçon -9.34162
+ça -9.34187
+isse -9.34203
+issant -9.3425
+▁petites -9.3435
+ables -9.345
+fin -9.34564
+▁correspond -9.34603
+▁grâce -9.34603
+gar -9.34618
+▁ke -9.34775
+ié -9.34788
+ayant -9.34829
+ève -9.34833
+ult -9.34989
+tic -9.35003
+▁van -9.35069
+▁gris -9.35166
+▁rec -9.35171
+nière -9.35289
+▁lancien -9.35349
+▁lalbum -9.35489
+▁église -9.3549
+▁voici -9.3549
+▁quarantehuit -9.35509
+▁transport -9.35513
+à -9.35529
+▁éc -9.35558
+isée -9.35567
+▁cherche -9.35588
+▁cheval -9.3582
+▁dr -9.35919
+▁cinquantetrois -9.35936
+▁reçoit -9.35936
+▁américain -9.3596
+▁vis -9.36249
+▁lorigine -9.36384
+▁époque -9.36385
+▁collection -9.36385
+▁voulu -9.36407
+▁rendre -9.36733
+▁prince -9.36763
+▁vieille -9.36842
+▁marquis -9.36844
+▁école -9.36856
+uel -9.36927
+▁outre -9.36955
+cour -9.36964
+▁ditil -9.37043
+▁souhaite -9.37087
+▁proposition -9.37368
+▁champion -9.37505
+▁cre -9.37543
+▁plein -9.37577
+▁constitue -9.37742
+▁république -9.37742
+▁trentesept -9.37742
+▁fête -9.37742
+▁commande -9.37744
+▁joseph -9.37805
+nais -9.37806
+▁sang -9.37828
+▁suisse -9.37847
+▁sor -9.37962
+▁raisons -9.37984
+▁répond -9.38156
+▁làbas -9.38214
+▁clos -9.38241
+▁noire -9.38336
+vel -9.38378
+aille -9.38462
+▁secrétaire -9.38657
+sol -9.38843
+▁rest -9.38851
+oul -9.38856
+▁divers -9.38913
+▁économique -9.39118
+▁régulièrement -9.39196
+▁viens -9.3935
+▁tor -9.39392
+zo -9.39393
+▁mise -9.39429
+lev -9.39555
+œ -9.39576
+▁lhomme -9.39581
+▁pose -9.39663
+▁pouss -9.39754
+▁seront -9.39803
+▁devait -9.39818
+á -9.39854
+ède -9.39924
+▁victor -9.40072
+▁connais -9.40114
+▁albert -9.4012
+lam -9.40154
+ham -9.40205
+chet -9.40284
+by -9.40367
+▁dy -9.40447
+▁entendu -9.40475
+▁cinquantequatre -9.40513
+▁faible -9.40514
+▁note -9.4052
+▁trentehuit -9.40542
+▁andré -9.40561
+▁comté -9.40597
+▁sétait -9.40616
+▁jouer -9.40682
+▁rem -9.40742
+▁sin -9.40801
+▁fé -9.4095
+▁front -9.40965
+▁catholique -9.40983
+▁cinquantecinq -9.40983
+▁lauteur -9.41004
+▁peintre -9.41105
+▁respect -9.41138
+vet -9.41141
+▁aide -9.41272
+▁content -9.4143
+▁ven -9.41453
+▁élevé -9.41459
+ût -9.41461
+gna -9.41486
+my -9.4161
+▁œuvre -9.41827
+▁prie -9.41925
+▁danger -9.4199
+▁mau -9.42143
+▁pou -9.42382
+▁extrême -9.42405
+▁remplacé -9.42405
+▁élève -9.42405
+▁référence -9.42411
+▁modèle -9.42412
+▁régime -9.42474
+cation -9.42631
+éra -9.4266
+▁maire -9.42685
+▁obtien -9.42693
+quà -9.42787
+od -9.42903
+▁trenteneuf -9.42911
+▁fonctionne -9.42917
+▁cité -9.43069
+▁développement -9.43117
+▁wi -9.43193
+▁gagne -9.4324
+▁programme -9.43364
+▁quarantequatre -9.43364
+▁york -9.43365
+oc -9.43465
+▁couple -9.43613
+shi -9.43749
+▁course -9.43794
+dan -9.43943
+▁relation -9.43954
+▁reven -9.43991
+▁rang -9.4402
+▁autant -9.44052
+vit -9.4422
+▁lorsque -9.44275
+▁composé -9.44313
+▁jétais -9.44332
+▁mademoiselle -9.44332
+▁quelquun -9.44332
+▁soutien -9.44401
+▁monte -9.44444
+euses -9.44801
+▁cinquantedeux -9.44821
+▁dirige -9.44904
+▁réalité -9.44922
+▁tre -9.44958
+ane -9.44968
+▁taille -9.45039
+▁allait -9.45123
+▁épouse -9.4531
+▁trouvé -9.45335
+▁dac -9.45433
+▁coupe -9.4546
+▁international -9.45628
+▁radio -9.45685
+▁technique -9.45803
+▁majorité -9.45839
+▁limite -9.45878
+▁regarde -9.45902
+anne -9.46123
+▁rester -9.46198
+tori -9.46279
+pro -9.46325
+▁jy -9.46337
+▁montagne -9.46351
+▁servi -9.46436
+▁espèces -9.46533
+▁suit -9.46774
+▁paraît -9.46798
+▁quarantesix -9.468
+▁présence -9.46802
+lès -9.46845
+bra -9.46888
+nois -9.47174
+gno -9.4723
+▁peinture -9.47355
+heim -9.47418
+gard -9.47474
+bar -9.47551
+▁inc -9.47552
+▁serai -9.47669
+▁bientôt -9.47798
+▁compagnie -9.47798
+▁réponse -9.47798
+▁terrain -9.47799
+▁silence -9.4791
+▁emp -9.47979
+▁arriv -9.48148
+▁tou -9.48187
+▁the -9.48228
+▁travailler -9.483
+▁réussi -9.48306
+▁déroule -9.48307
+▁enterré -9.48315
+pen -9.48352
+ian -9.48404
+gal -9.48573
+rique -9.48597
+▁prise -9.48755
+▁contribu -9.48811
+▁lintérieur -9.48811
+▁largent -9.48811
+▁auto -9.4891
+▁pat -9.48946
+▁ceci -9.49066
+▁appartient -9.49321
+▁lacadémie -9.49321
+▁voulez -9.4935
+▁gri -9.49507
+▁observ -9.49568
+ition -9.49576
+▁départ -9.49905
+▁savez -9.50305
+▁minutes -9.50315
+▁profit -9.50385
+▁quitte -9.5041
+ification -9.50484
+▁davoir -9.5053
+▁développe -9.50614
+▁salon -9.50639
+▁tir -9.50812
+▁vaut -9.50813
+ow -9.50863
+▁soixantedouze -9.50867
+▁édition -9.5087
+▁londres -9.50874
+▁soixantedixhuit -9.50883
+▁dha -9.50922
+▁fri -9.50939
+antes -9.51011
+▁gla -9.51077
+▁amis -9.51095
+▁fran -9.51221
+▁assis -9.51224
+▁occupe -9.51286
+▁el -9.51315
+▁cinquantesept -9.51388
+▁quarantesept -9.51388
+▁soixantecinq -9.51388
+▁âge -9.51389
+▁remporté -9.51389
+▁manger -9.51443
+dra -9.51455
+lay -9.51481
+▁répondit -9.51556
+nal -9.51805
+table -9.51819
+▁affaires -9.51829
+ulaire -9.51831
+cle -9.51869
+▁siècle -9.51912
+▁trou -9.51966
+▁contrôle -9.51995
+riez -9.52001
+nor -9.52069
+thé -9.52118
+▁santé -9.52132
+▁retourne -9.52162
+▁partage -9.52211
+▁sert -9.5226
+ub -9.52433
+▁précédent -9.52438
+▁doivent -9.52439
+▁jusquau -9.52439
+▁trentesix -9.52445
+▁pre -9.52445
+▁professionnel -9.52536
+▁rome -9.52654
+▁traverse -9.52913
+▁commercial -9.52967
+▁total -9.52977
+▁placé -9.53042
+▁parce -9.53106
+gnon -9.53292
+▁emb -9.53383
+▁réseau -9.53384
+▁dargent -9.5354
+▁réalisé -9.53542
+▁troupe -9.53566
+tine -9.53619
+itude -9.53729
+▁grave -9.5373
+tar -9.53754
+▁bour -9.53785
+▁tro -9.53943
+▁we -9.53973
+gère -9.5398
+ov -9.5405
+nage -9.54093
+▁basse -9.54126
+▁comb -9.54222
+▁sorte -9.5453
+▁activité -9.54568
+▁presse -9.54568
+▁café -9.54576
+▁soleil -9.54578
+q -9.54678
+▁déli -9.54682
+▁pièces -9.54698
+ssou -9.54811
+▁rive -9.54951
+▁vise -9.54999
+▁derrière -9.55113
+▁structure -9.55151
+▁défend -9.55165
+▁ordre -9.55242
+▁dos -9.55259
+▁far -9.55282
+▁prêt -9.55304
+▁joué -9.55311
+▁points -9.55361
+tré -9.55459
+▁georges -9.5546
+▁défendu -9.55633
+▁êtesvous -9.55657
+nom -9.55675
+▁hor -9.55775
+▁com -9.55806
+▁han -9.5581
+ures -9.55826
+▁rentre -9.55844
+▁pin -9.55894
+pit -9.56072
+▁apporte -9.56084
+▁lance -9.56088
+▁air -9.56144
+▁tend -9.56188
+▁confi -9.56188
+▁ste -9.56197
+▁soixantequinze -9.56202
+▁reconnu -9.56202
+▁instant -9.56204
+▁quai -9.56213
+▁pouvez -9.56357
+▁verre -9.56361
+▁films -9.56583
+▁russe -9.56584
+▁tient -9.56637
+▁quart -9.56674
+quer -9.56686
+▁établi -9.56746
+▁dispositif -9.56751
+▁lentreprise -9.56751
+▁téléphone -9.56751
+▁lobjet -9.56752
+▁document -9.56752
+▁résidence -9.56763
+▁choix -9.56777
+▁entièrement -9.56857
+viennent -9.57061
+▁lyon -9.5712
+▁sélection -9.5712
+ème -9.57154
+met -9.57159
+▁nont -9.57191
+ssant -9.57231
+nard -9.57239
+▁gro -9.57335
+▁brun -9.57358
+irent -9.57367
+illy -9.57457
+▁revenu -9.57578
+éro -9.5763
+▁domin -9.57694
+▁apprend -9.57712
+▁composée -9.57744
+▁lendemain -9.57859
+ault -9.57889
+▁marié -9.57891
+▁réalise -9.57898
+▁viendra -9.57905
+▁représentant -9.57935
+cho -9.5795
+line -9.57976
+▁collègues -9.58009
+ivité -9.58023
+ok -9.58073
+▁pourrait -9.58112
+▁page -9.58164
+ak -9.58305
+▁fra -9.58316
+▁nature -9.58353
+▁tourne -9.5836
+▁japon -9.58407
+▁développé -9.58418
+▁populaire -9.58418
+▁lettres -9.58443
+ange -9.58443
+▁dart -9.58471
+▁télé -9.58475
+▁daller -9.58479
+▁nor -9.58556
+▁laide -9.5866
+agne -9.58729
+▁chance -9.58742
+ssait -9.58744
+cep -9.58933
+▁difficulté -9.5898
+▁objet -9.5898
+▁commerce -9.58994
+▁os -9.59124
+▁conte -9.59228
+▁java -9.59232
+gou -9.59262
+▁essay -9.59497
+▁dites -9.59511
+ö -9.59545
+▁sinstalle -9.59545
+vaient -9.59604
+▁étudiant -9.59607
+▁ed -9.59616
+▁création -9.59779
+▁personnage -9.59802
+▁entra -9.59811
+▁pra -9.59851
+▁larrondissement -9.60113
+▁lempereur -9.60113
+ana -9.60116
+▁monument -9.60123
+▁rappel -9.60142
+vait -9.60358
+▁social -9.60544
+▁fallait -9.60685
+▁internationale -9.60888
+▁anne -9.60924
+ncer -9.60933
+ari -9.60973
+▁mit -9.60984
+oli -9.61022
+gues -9.61039
+▁plu -9.61061
+▁équipes -9.61088
+▁refus -9.6115
+▁retirer -9.6123
+▁blessé -9.6126
+▁rené -9.61349
+▁plat -9.61411
+oise -9.61426
+▁tom -9.61435
+▁fermé -9.61452
+▁reprises -9.61545
+▁culture -9.61556
+ative -9.61587
+père -9.61636
+mier -9.61742
+ó -9.61837
+▁constitué -9.61837
+▁véritable -9.61837
+▁maladie -9.61837
+▁réserve -9.61838
+▁labbé -9.61849
+▁cinquantehuit -9.61862
+mat -9.61865
+▁probablement -9.61889
+▁gen -9.61907
+▁pain -9.62118
+bois -9.62217
+▁connaît -9.62249
+mporte -9.62334
+▁code -9.62356
+▁célèbre -9.62419
+▁principe -9.62419
+▁quaranteneuf -9.62443
+▁derniers -9.62453
+▁villa -9.62527
+▁refuse -9.62537
+▁stade -9.62745
+ray -9.62757
+▁remercie -9.63004
+▁science -9.63004
+▁contrat -9.63006
+▁accord -9.63008
+ngage -9.63087
+▁reprit -9.63104
+ini -9.63129
+▁présenté -9.63134
+▁mauvais -9.63234
+▁consul -9.63283
+lia -9.63299
+▁pièce -9.63453
+vezvous -9.63475
+▁prison -9.6351
+▁nav -9.63516
+▁district -9.63592
+▁parallèle -9.63592
+▁soixantetrois -9.63592
+▁concert -9.63596
+▁défini -9.63703
+▁vote -9.63762
+▁football -9.64098
+lly -9.64118
+▁sortir -9.64126
+▁fleur -9.64184
+▁utilis -9.64268
+▁centrale -9.64463
+▁ouverte -9.64504
+ute -9.64578
+▁quatrevingtcinq -9.64779
+▁décision -9.64779
+▁quarantedeux -9.64779
+chon -9.6486
+▁ange -9.65223
+▁fourni -9.65243
+issait -9.65253
+▁quarantetrois -9.65377
+▁quatrième -9.65377
+▁atteint -9.65378
+▁poisson -9.65378
+▁classique -9.65378
+▁vérité -9.65378
+▁vallée -9.65385
+▁questions -9.65421
+▁risque -9.65438
+▁fur -9.65601
+▁douce -9.65665
+▁dam -9.65686
+▁filles -9.65766
+▁britannique -9.6598
+▁daccord -9.6598
+▁soixantesept -9.6598
+▁lenfant -9.65981
+▁enseigne -9.65981
+trice -9.66064
+nge -9.66094
+▁luc -9.66486
+▁locale -9.66538
+âtre -9.66559
+▁espagnol -9.66586
+▁spectacle -9.66586
+▁urbain -9.66586
+▁remplace -9.66588
+▁soixantesix -9.66591
+▁métier -9.66686
+▁té -9.66711
+▁finit -9.66878
+▁chasse -9.66886
+sible -9.66912
+▁change -9.66916
+assi -9.67107
+ker -9.6712
+▁claire -9.6712
+oud -9.67121
+▁magasin -9.67196
+laient -9.6721
+▁dominique -9.67243
+▁associé -9.67263
+ë -9.67326
+▁bruit -9.67472
+▁mill -9.67483
+▁européen -9.67561
+▁chargé -9.67765
+▁disparu -9.67811
+▁bancs -9.67837
+▁demeure -9.67845
+▁quen -9.67872
+▁tenu -9.67905
+▁représenté -9.6791
+▁moulin -9.67919
+nique -9.67993
+▁cli -9.6807
+▁bouche -9.68129
+▁auteur -9.68144
+uf -9.68212
+bonne -9.68329
+▁origine -9.68427
+▁conserve -9.68429
+▁wal -9.68433
+▁conseiller -9.68495
+▁marcel -9.68513
+▁arrête -9.68713
+▁tué -9.68803
+dia -9.68923
+▁présidente -9.68939
+▁boul -9.69004
+▁soixantequatre -9.69048
+▁étrange -9.69048
+▁cycl -9.69055
+crit -9.69098
+▁pot -9.69169
+▁dautre -9.69301
+▁cau -9.69407
+▁poly -9.69426
+▁lien -9.69462
+zy -9.69589
+▁soixantedeux -9.69673
+▁papier -9.69675
+▁clé -9.6968
+▁soixantedixneuf -9.6969
+▁sociale -9.69723
+▁sport -9.69795
+mir -9.6986
+▁objectif -9.70302
+▁occasion -9.70302
+▁satisfait -9.70302
+▁daltitude -9.70302
+▁argent -9.70404
+▁revient -9.70521
+▁hum -9.70535
+fo -9.70627
+▁marin -9.70702
+ographie -9.70807
+▁acte -9.70847
+marie -9.70863
+▁soixanteseize -9.70934
+▁méthode -9.70935
+▁lycée -9.70937
+issement -9.70942
+▁police -9.70994
+▁patron -9.71021
+▁lourd -9.71101
+▁seuls -9.71209
+rêt -9.71275
+dal -9.71288
+▁basé -9.7156
+▁caractère -9.71571
+▁quatrevingtneuf -9.71572
+▁quatrevingthuit -9.71572
+▁canada -9.71573
+▁dehors -9.71577
+▁soixantehuit -9.71597
+▁navire -9.716
+▁bernard -9.71615
+▁courant -9.71662
+▁bus -9.71689
+tha -9.71731
+▁bourg -9.71739
+▁tournée -9.71742
+▁ciel -9.71813
+▁retrait -9.71831
+atrice -9.71849
+▁tente -9.71851
+▁tar -9.71861
+ach -9.71948
+▁gars -9.72044
+donne -9.72056
+▁resta -9.72071
+▁dû -9.72161
+▁profond -9.72182
+pon -9.72205
+soixantequatorze -9.72212
+▁lœuvre -9.72213
+▁voulait -9.72371
+▁saut -9.7243
+▁resté -9.7254
+▁réuni -9.72581
+▁dure -9.72603
+vert -9.72653
+roy -9.72772
+cide -9.72791
+▁événement -9.72858
+▁relative -9.72858
+▁cinquanteneuf -9.72884
+▁îles -9.72901
+▁post -9.7292
+▁sha -9.73143
+▁pouv -9.73306
+mètre -9.73495
+▁collège -9.73507
+▁nicolas -9.73507
+▁david -9.7351
+▁armée -9.73524
+▁interdit -9.7355
+èvre -9.73565
+▁entreprises -9.7369
+bro -9.73741
+ama -9.73752
+vingtquatre -9.73787
+▁langues -9.73868
+▁laisser -9.73889
+bat -9.73939
+sous -9.74031
+▁fatigué -9.7416
+▁propriétaire -9.7416
+▁propriété -9.7416
+▁lhôtel -9.74161
+▁solution -9.74161
+▁branche -9.74162
+▁feuilles -9.74197
+▁localité -9.74218
+▁journée -9.74249
+▁tru -9.74307
+▁visage -9.74345
+toi -9.7448
+▁actuel -9.745
+▁seigneur -9.74751
+▁coll -9.74797
+▁proximité -9.74818
+quatrevingttrois -9.74819
+▁emploi -9.74822
+▁vivant -9.74825
+▁form -9.74865
+▁gal -9.74966
+▁épi -9.74991
+nat -9.75047
+yo -9.75063
+▁hy -9.75306
+▁nouveaux -9.75349
+▁faudra -9.75354
+vu -9.75369
+▁ger -9.75412
+▁mod -9.7542
+▁régional -9.7548
+▁qualité -9.75503
+▁partout -9.75516
+▁classé -9.75551
+ezmoi -9.75706
+pression -9.75707
+▁concerne -9.75749
+▁ouvre -9.75799
+▁foi -9.75855
+ling -9.75874
+▁repose -9.75897
+▁collègue -9.75969
+▁fera -9.75993
+▁audessus -9.76147
+▁façade -9.76147
+▁conduit -9.7615
+▁connaissance -9.7615
+▁ly -9.76211
+▁noble -9.76246
+▁vienne -9.76252
+▁ham -9.76361
+▁eff -9.76374
+▁matchs -9.76402
+▁souhait -9.76451
+▁ty -9.76519
+▁grosse -9.76746
+gée -9.7676
+▁quatrevingtdeux -9.76818
+▁europe -9.76819
+▁dirigé -9.7682
+▁jaimerais -9.76824
+▁quatrevingtun -9.76826
+▁demandé -9.76826
+▁prochain -9.7683
+▁résultats -9.76837
+▁conservé -9.7688
+▁comité -9.76891
+▁mine -9.76896
+▁supprimer -9.76913
+▁chapeau -9.76935
+▁somme -9.76938
+▁hameau -9.76956
+▁cadre -9.77063
+tou -9.77079
+▁dia -9.7708
+ō -9.77494
+▁responsable -9.77494
+▁scientifique -9.77494
+▁support -9.77494
+▁officier -9.77494
+▁preuve -9.77495
+▁liberté -9.77495
+lation -9.77495
+▁poursuit -9.77498
+▁chat -9.775
+▁soixanteneuf -9.7752
+▁moderne -9.77566
+▁coût -9.77598
+▁passion -9.7774
+▁travaillé -9.77837
+▁savait -9.77958
+▁continu -9.77975
+▁cathédrale -9.78174
+▁exécut -9.78175
+▁préfér -9.78186
+▁laissé -9.78266
+thique -9.78281
+▁devenir -9.7842
+▁accepte -9.78481
+▁mil -9.78516
+cre -9.78567
+▁kar -9.78569
+euil -9.7857
+▁désir -9.7869
+▁and -9.78742
+zon -9.788
+▁physique -9.78859
+▁quatrevingtonze -9.78859
+▁record -9.7886
+▁inhumé -9.78861
+▁télévision -9.78864
+▁essentiellement -9.78864
+▁cinquantesix -9.78866
+▁sombre -9.78885
+▁daniel -9.78956
+▁viennent -9.7898
+rich -9.79021
+▁portrait -9.79039
+▁chaises -9.79074
+old -9.79102
+▁proposé -9.79113
+▁test -9.79128
+▁changement -9.79145
+▁compren -9.79208
+mment -9.7937
+▁diverses -9.79375
+▁lutte -9.79377
+▁générale -9.7942
+▁rh -9.79452
+▁sommet -9.79476
+▁animaux -9.79549
+▁suppression -9.79549
+▁favori -9.79551
+▁richard -9.79571
+atoire -9.79571
+▁parlé -9.79615
+mal -9.7968
+▁frontière -9.7969
+▁prit -9.7971
+▁lart -9.79753
+part -9.7979
+▁formé -9.79913
+▁repris -9.79962
+▁pol -9.79983
+▁tort -9.80019
+▁flam -9.80133
+vingtquatorze -9.80243
+▁festival -9.80243
+▁inconnu -9.80244
+▁lâge -9.80244
+▁municipalité -9.80282
+▁daprès -9.80414
+▁soldats -9.8059
+ano -9.80646
+▁x -9.80678
+▁flo -9.8072
+▁moyenne -9.80788
+▁rêve -9.80875
+▁anciens -9.80885
+▁appelle -9.809
+▁mode -9.80903
+▁arrêté -9.80909
+▁soixantetreize -9.80943
+▁surface -9.80944
+▁conditions -9.80948
+▁secteur -9.80956
+matique -9.81115
+▁tournoi -9.81155
+▁venait -9.8125
+▁lavait -9.81264
+▁chansons -9.81322
+pha -9.81352
+ssement -9.81356
+gui -9.81401
+▁vendu -9.81456
+▁services -9.81601
+▁conçu -9.81648
+▁approche -9.81658
+▁prévu -9.81698
+▁hou -9.81709
+▁cache -9.81735
+▁glace -9.81758
+▁guéri -9.8176
+▁rappelle -9.81939
+▁ouvrages -9.81988
+▁baron -9.82098
+▁règle -9.82102
+out -9.82255
+▁disponible -9.82356
+▁dispute -9.82357
+▁statue -9.82363
+▁julien -9.82423
+▁inclu -9.8245
+▁za -9.82455
+▁bio -9.82457
+her -9.82498
+stro -9.82522
+len -9.82526
+ignac -9.82531
+▁décrit -9.82534
+▁fusil -9.83052
+▁citoyen -9.8307
+▁sousamendement -9.8307
+vingtquinze -9.83071
+▁belgique -9.83071
+quatrevingtdouze -9.83071
+▁conclu -9.83073
+▁sétend -9.83074
+▁machine -9.83078
+▁remonte -9.83157
+▁fontaine -9.83193
+dire -9.83219
+oux -9.83258
+▁construite -9.83265
+▁direct -9.83284
+fu -9.83403
+▁baie -9.8343
+▁comprendre -9.83434
+▁outil -9.83724
+▁guillaume -9.8379
+▁installé -9.8379
+▁gabriel -9.83795
+▁offre -9.83805
+▁étage -9.83806
+▁voulezvous -9.83806
+▁piano -9.83842
+▁assure -9.83846
+▁puisque -9.83853
+roux -9.83854
+▁tableau -9.83856
+particulièrement -9.83891
+ération -9.83909
+▁romain -9.8391
+▁disque -9.8398
+▁lancienne -9.84021
+▁forces -9.84131
+▁personnages -9.84186
+▁accept -9.84191
+▁faute -9.84396
+▁dois -9.84469
+▁canadien -9.84517
+▁attaque -9.84522
+▁cloche -9.84528
+▁davantage -9.8454
+▁mono -9.84555
+▁visible -9.84566
+iel -9.84597
+kh -9.84619
+gin -9.84775
+▁séance -9.84847
+▁rendu -9.84922
+▁facile -9.84992
+▁sortie -9.85032
+▁artistes -9.85156
+▁paroisse -9.85248
+▁retiré -9.85248
+▁repondit -9.85276
+▁quatrevingtsix -9.85303
+▁frédéric -9.85365
+▁compris -9.8537
+▁berlin -9.85442
+▁personnel -9.85442
+▁pleine -9.85515
+▁cousin -9.85593
+▁retraite -9.85698
+▁banque -9.85721
+▁agricole -9.8598
+▁détruit -9.8598
+▁contient -9.85983
+▁marchand -9.86
+▁immédiatement -9.86001
+igni -9.86014
+llier -9.86046
+▁jolie -9.86062
+▁fondateur -9.86123
+▁laurent -9.86137
+▁royal -9.8614
+▁fleurs -9.86152
+▁agr -9.86186
+hen -9.8623
+▁employé -9.86245
+▁disait -9.86377
+▁local -9.86443
+▁million -9.86454
+▁largement -9.86502
+▁bau -9.86517
+vol -9.86522
+▁née -9.86629
+▁parlement -9.86648
+lard -9.86711
+▁indépendant -9.8672
+▁composition -9.8674
+▁trompe -9.86744
+▁italien -9.86751
+▁major -9.86791
+▁découverte -9.86802
+iège -9.86819
+mbre -9.86908
+ignan -9.86995
+ologue -9.87075
+▁grandi -9.87078
+▁comm -9.87137
+fon -9.87206
+▁québec -9.87467
+▁thomas -9.87467
+▁victime -9.87467
+▁théorie -9.87467
+▁supérieur -9.87503
+▁climat -9.87639
+zmoi -9.87709
+▁prêtre -9.88039
+▁joli -9.88138
+▁troubl -9.882
+▁labbaye -9.88219
+▁principaux -9.88219
+▁chiffre -9.8822
+▁activités -9.88224
+▁maintien -9.88255
+▁allé -9.88282
+▁japonais -9.88286
+▁découvre -9.88309
+▁ceuxci -9.88388
+▁francis -9.88599
+roche -9.88658
+ku -9.88713
+long -9.88817
+abilité -9.88972
+▁lassociation -9.88976
+▁lépreuve -9.88976
+▁simpose -9.88976
+▁plaisir -9.88976
+▁naissance -9.88977
+▁raconte -9.88984
+▁parfaitement -9.88999
+▁olympique -9.89035
+▁rejet -9.89086
+▁marqué -9.8913
+▁détail -9.89142
+▁palais -9.8918
+▁calme -9.89205
+▁thi -9.89261
+▁lire -9.89312
+▁peint -9.89474
+lette -9.89569
+▁donna -9.89579
+▁lucien -9.89637
+▁unique -9.89717
+▁diplômé -9.8974
+▁sécria -9.8974
+▁invité -9.89746
+▁lidée -9.89746
+▁abrite -9.89752
+▁pilote -9.89753
+▁fixe -9.89848
+▁considérée -9.89875
+dol -9.9001
+▁européenne -9.9005
+▁revenir -9.90072
+usement -9.90114
+▁couche -9.90124
+▁pensa -9.90184
+géné -9.90327
+▁apparaît -9.90509
+▁protestant -9.90509
+▁dhonneur -9.90509
+▁moitié -9.90513
+▁voudrais -9.90553
+▁rendezvous -9.90564
+▁accueille -9.9057
+▁heureux -9.90614
+▁chaise -9.90686
+▁linstitut -9.90724
+être -9.90753
+▁linter -9.90883
+▁mauvaise -9.91005
+▁franc -9.91055
+lav -9.9107
+mond -9.91087
+▁valeur -9.91258
+▁bibliothèque -9.91284
+▁discussion -9.91284
+▁lassemblée -9.91284
+▁lumière -9.91285
+▁goût -9.91322
+▁calcul -9.9134
+▁monta -9.91367
+▁voter -9.91374
+▁lamour -9.91441
+▁hors -9.91482
+▁atil -9.91653
+▁meilleure -9.91745
+ney -9.91753
+▁rapporteure -9.91779
+▁moteur -9.9179
+▁rapide -9.91813
+cré -9.92038
+▁réside -9.92065
+▁brûl -9.92065
+▁justice -9.92065
+▁évidemment -9.92065
+▁domestique -9.92066
+▁revue -9.9207
+oï -9.92074
+▁prisonnier -9.92085
+▁normal -9.92086
+▁emm -9.92124
+▁données -9.92245
+▁éch -9.9238
+▁futur -9.92391
+▁passa -9.92573
+▁puissant -9.92577
+vingttreize -9.92853
+▁endroit -9.92853
+▁étudié -9.92853
+▁quatrevingtsept -9.92853
+▁christ -9.92855
+▁remarque -9.92856
+▁sudouest -9.92864
+▁engagé -9.92865
+▁hauteur -9.92883
+▁opération -9.92933
+▁nation -9.93103
+ap -9.93133
+▁tian -9.93166
+▁destiné -9.93211
+zan -9.93507
+▁central -9.93531
+▁naturel -9.93599
+▁couvert -9.93601
+champ -9.93604
+▁allemagne -9.93646
+▁bénéficie -9.93646
+▁soixantedixsept -9.93646
+▁consacré -9.93646
+▁vacances -9.93646
+▁studio -9.93646
+▁marseille -9.93647
+▁chaîne -9.93647
+▁attention -9.93651
+▁industriel -9.93652
+▁thème -9.93654
+▁médaille -9.93654
+▁dépense -9.93656
+▁gliss -9.93732
+▁bapt -9.93781
+ther -9.93917
+ford -9.93985
+▁courage -9.93994
+▁essai -9.94134
+▁classement -9.94141
+neau -9.94394
+▁excellent -9.94446
+▁religieuse -9.94446
+▁autrefois -9.94446
+▁crayon -9.94447
+▁trembl -9.94447
+▁féminin -9.94448
+▁explique -9.94466
+▁acheté -9.94479
+▁considérable -9.945
+▁confiance -9.94511
+▁ubu -9.94528
+▁affluent -9.94539
+▁dédié -9.94597
+▁tenir -9.94606
+▁claude -9.94628
+▁américaine -9.94686
+▁cheminée -9.94751
+▁th -9.94771
+▁francs -9.94776
+▁hon -9.94919
+▁éléments -9.94971
+▁complète -9.94999
+▁musicien -9.95008
+▁défense -9.9506
+▁ligue -9.95084
+▁bur -9.95128
+▁coin -9.95144
+nau -9.95149
+gny -9.95164
+où -9.95175
+▁simplement -9.95208
+▁nourri -9.95242
+▁fortement -9.95246
+▁différence -9.95253
+▁original -9.95253
+▁phénomène -9.95253
+▁revanche -9.95253
+▁bonheur -9.95253
+▁saintjean -9.95254
+▁dom -9.9528
+▁leçon -9.95293
+▁nou -9.95308
+▁ker -9.954
+▁conten -9.95721
+▁mini -9.9596
+▁royaume -9.96066
+▁californie -9.96066
+▁conséquence -9.96066
+▁traduit -9.96066
+▁éléphants -9.96071
+▁arbres -9.96077
+▁matière -9.96084
+▁réalisée -9.9614
+▁mol -9.96195
+▁belge -9.96342
+▁relie -9.9638
+▁chercher -9.96397
+▁trouva -9.9646
+▁papa -9.9647
+▁réunion -9.96473
+▁muni -9.96559
+▁demander -9.96687
+▁aimera -9.96719
+▁déf -9.9674
+▁défi -9.96754
+▁chine -9.96817
+tention -9.96831
+▁tranquille -9.96885
+▁étape -9.96886
+▁constat -9.96914
+▁jespère -9.96919
+▁aima -9.97075
+▁léon -9.97159
+▁média -9.97182
+▁traité -9.97277
+▁piste -9.97312
+▁attendre -9.97315
+▁logique -9.97626
+▁fruits -9.9766
+▁instrument -9.97712
+▁lindustrie -9.97712
+▁préoccup -9.97712
+▁obtenu -9.97712
+▁envoyé -9.97712
+▁tôt -9.97712
+▁auguste -9.97713
+▁sappelle -9.97713
+▁succède -9.97718
+▁fiscal -9.97723
+▁lecture -9.97733
+▁chinois -9.97743
+▁vide -9.97787
+▁poète -9.97805
+▁zo -9.97811
+▁affaire -9.97851
+▁vivre -9.97859
+▁changer -9.97958
+▁touche -9.97965
+▁journaliste -9.97984
+▁sorti -9.97994
+▁lieux -9.98094
+▁cir -9.98158
+▁via -9.9819
+▁envie -9.98263
+▁occupé -9.98327
+▁mich -9.98337
+▁simon -9.98494
+▁tradition -9.98524
+▁candidat -9.98545
+▁résistance -9.98545
+▁manifest -9.98545
+▁dimanche -9.98546
+▁séparé -9.98553
+▁exact -9.98559
+▁colonel -9.98585
+▁prénom -9.98617
+▁vécu -9.98625
+étaient -9.98706
+▁antoine -9.98742
+ppo -9.98784
+▁uniquement -9.98841
+▁ku -9.98848
+▁avions -9.9893
+▁poule -9.99028
+▁parlez -9.99038
+terre -9.99187
+▁écout -9.99212
+▁remis -9.99234
+▁coûte -9.99375
+▁caractéristique -9.99386
+▁supplémentaire -9.99386
+▁distingue -9.99386
+▁restaurant -9.99386
+▁langlais -9.99388
+▁terminé -9.99389
+▁commissaire -9.99389
+▁réforme -9.99395
+▁retard -9.99396
+▁secret -9.99396
+▁nordouest -9.99397
+hérit -9.99412
+▁directement -9.99449
+▁précision -9.99462
+mbra -9.99523
+nni -9.99525
+▁trouvait -9.99539
+▁cur -9.99614
+▁démon -9.99614
+cueil -9.99687
+▁tai -9.99791
+▁carré -9.99893
+▁trait -10.0006
+▁midi -10.0021
+▁littérature -10.0023
+▁électrique -10.0023
+▁cabinet -10.0023
+▁matériel -10.0023
+▁diminu -10.0024
+▁âme -10.0024
+▁épaule -10.0024
+▁photo -10.0025
+▁reine -10.005
+▁neveu -10.0055
+nçon -10.0057
+bb -10.006
+▁pren -10.0076
+▁italie -10.0076
+▁servir -10.0099
+▁diocèse -10.0109
+▁exploit -10.0109
+▁loccasion -10.0109
+▁détermin -10.0109
+▁défaut -10.0109
+▁maurice -10.0109
+▁consomm -10.0109
+▁lempire -10.0109
+▁récit -10.0113
+▁aliment -10.0113
+▁section -10.0113
+géni -10.0124
+▁guy -10.0134
+▁chevalier -10.0136
+▁chante -10.0139
+gli -10.0145
+▁orient -10.0154
+▁ann -10.0155
+▁éli -10.0169
+ctor -10.0181
+▁ruelle -10.019
+▁ministère -10.0195
+▁épisode -10.0195
+▁voul -10.0195
+▁tâche -10.0195
+▁rédactionnel -10.0195
+▁brésil -10.0195
+▁chauss -10.0204
+▁totalement -10.0204
+▁march -10.0205
+▁client -10.0212
+▁complètement -10.0222
+▁lunion -10.023
+▁privée -10.0241
+▁privé -10.0244
+▁tal -10.025
+▁poche -10.0258
+wood -10.0273
+▁flori -10.0281
+▁sécurité -10.0282
+▁auparavant -10.0282
+▁cheveux -10.0282
+▁australie -10.0282
+obtenir -10.0282
+▁contact -10.0282
+▁information -10.0282
+gg -10.0295
+quel -10.0299
+▁pouvons -10.03
+▁quitter -10.0302
+▁trace -10.0328
+prime -10.0341
+▁réalis -10.0342
+▁trouvent -10.0344
+▁frais -10.0366
+▁abandonné -10.037
+▁intérêt -10.037
+▁malheureusement -10.037
+▁doigt -10.037
+▁prétend -10.0371
+▁actif -10.0371
+▁argument -10.0382
+▁faudrait -10.0388
+▁folle -10.0392
+▁manche -10.0431
+▁tire -10.0452
+▁rédaction -10.0458
+▁monastère -10.0458
+▁témoign -10.0458
+▁phase -10.0458
+▁hasard -10.046
+▁lentrée -10.046
+▁préfère -10.046
+lage -10.0461
+▁bay -10.0462
+▁saintpierre -10.0463
+▁absolument -10.0465
+▁souvenir -10.0467
+▁berger -10.0482
+▁cata -10.0492
+▁barre -10.0502
+▁secondaire -10.0507
+voyait -10.0514
+▁vaste -10.0516
+▁jaurais -10.0517
+▁inde -10.0527
+▁jon -10.053
+▁venez -10.053
+▁vive -10.0533
+▁continua -10.0534
+zin -10.0539
+▁chen -10.0547
+▁reconnais -10.0547
+ä -10.0547
+▁chrétien -10.0547
+▁légèrement -10.0547
+▁contraint -10.0548
+▁roul -10.0556
+▁regret -10.0558
+▁loup -10.0561
+▁payer -10.0563
+▁guère -10.0563
+▁médecine -10.0564
+▁parlementaire -10.0568
+envoie -10.0571
+▁lequel -10.0574
+▁qualifi -10.0593
+▁centaine -10.0607
+lio -10.0637
+▁juridique -10.0638
+▁logiciel -10.0638
+▁indique -10.0638
+cade -10.064
+▁désigne -10.064
+▁écrivain -10.0644
+▁pensée -10.0651
+▁régl -10.066
+▁naturelle -10.0662
+▁connaître -10.0664
+▁concern -10.0673
+▁compose -10.0701
+▁précise -10.071
+▁poursuiv -10.0713
+▁découvert -10.0718
+applaudissement -10.0728
+▁augmente -10.0729
+▁desservi -10.073
+icité -10.0732
+▁bassin -10.0734
+▁publication -10.0734
+▁désert -10.0734
+▁rouss -10.0737
+▁colonne -10.0777
+hum -10.0778
+▁jouait -10.0783
+▁entreprise -10.0794
+▁nomme -10.0795
+▁lob -10.0805
+otte -10.0815
+▁durée -10.0816
+mann -10.0816
+▁put -10.0818
+▁apparten -10.082
+▁fleuve -10.082
+▁impossible -10.082
+▁philosophie -10.082
+▁règne -10.082
+▁soccupe -10.082
+▁cuisine -10.082
+▁volonté -10.082
+▁dizaine -10.082
+▁mourir -10.0821
+▁erreur -10.0826
+▁celuilà -10.083
+▁acheter -10.0839
+▁annonce -10.0842
+▁crise -10.087
+▁tim -10.0875
+▁grandit -10.0882
+▁barbe -10.0883
+▁hall -10.0896
+beau -10.0897
+▁robe -10.0897
+▁vidéo -10.0912
+▁rectifié -10.0913
+▁rejoindre -10.0913
+▁budget -10.0913
+▁voyez -10.0914
+▁tré -10.0914
+▁sudest -10.0914
+▁mandat -10.0914
+▁tombé -10.0916
+▁convention -10.0917
+▁éclair -10.092
+▁boutique -10.0921
+▁solide -10.0922
+▁pomme -10.0927
+▁wo -10.0935
+▁colère -10.0936
+▁grotte -10.0939
+▁ira -10.094
+▁retrouver -10.0941
+▁tomba -10.0942
+cq -10.0943
+▁armes -10.0944
+▁posé -10.0944
+▁casse -10.0949
+▁portant -10.0966
+▁baisse -10.0993
+▁participa -10.1001
+▁william -10.1006
+▁avocat -10.1006
+▁élection -10.1006
+▁administrative -10.1007
+▁laffaire -10.1008
+▁exposé -10.1009
+▁humain -10.1019
+▁neige -10.1022
+▁gilles -10.1029
+▁sentiment -10.1031
+▁influence -10.1032
+paul -10.1036
+chang -10.1037
+▁montée -10.1045
+▁joie -10.1048
+▁mène -10.1052
+mporta -10.1056
+▁répondre -10.1079
+▁remet -10.1083
+▁color -10.1091
+▁jeunesse -10.1092
+▁léconomie -10.1101
+▁sauvage -10.1101
+▁mémoire -10.1101
+▁phrase -10.1101
+▁conserva -10.1101
+▁nexiste -10.1101
+▁signal -10.1103
+▁ajoute -10.1107
+▁dici -10.111
+▁manqu -10.111
+▁traditionnelle -10.1111
+▁gaz -10.1112
+▁plateau -10.1113
+▁larmes -10.1114
+▁devrait -10.1115
+▁paire -10.1119
+endra -10.114
+▁spéciale -10.1144
+▁décor -10.1146
+▁cria -10.1167
+û -10.1169
+▁paysan -10.1175
+▁taux -10.1191
+▁allemande -10.1191
+▁complexe -10.1196
+▁intérieur -10.1196
+▁procédure -10.1196
+▁reconnaît -10.1196
+▁équipé -10.1196
+▁rapproch -10.1196
+▁affirm -10.1196
+▁animal -10.1196
+▁cultiv -10.1196
+▁réduit -10.1196
+▁sourire -10.1196
+▁discut -10.12
+▁fauteuil -10.1202
+▁écrire -10.1202
+▁surprise -10.1204
+▁eût -10.1207
+▁entouré -10.1207
+pol -10.1214
+▁récent -10.1216
+quin -10.1225
+▁phil -10.1231
+▁poli -10.1251
+michel -10.1255
+▁compos -10.1268
+▁év -10.1269
+▁prévoi -10.1269
+▁évolu -10.1271
+neuve -10.1272
+mère -10.1285
+▁résultat -10.1289
+▁conséquent -10.1292
+▁quotidien -10.1292
+▁effectue -10.1292
+▁producteur -10.1292
+▁triomph -10.1292
+▁espace -10.1292
+▁raymond -10.1292
+▁ouvrir -10.1292
+▁souverain -10.1292
+▁ignor -10.1293
+▁exerce -10.1294
+▁changé -10.1295
+▁esprit -10.1296
+▁amour -10.1297
+▁navez -10.1299
+▁association -10.1306
+▁recul -10.1312
+▁foule -10.1328
+▁repos -10.1334
+▁donnant -10.1336
+▁exprim -10.1355
+▁royale -10.1369
+hou -10.1374
+huit -10.1376
+▁senti -10.1377
+▁rire -10.1381
+▁catégorie -10.1389
+▁hommage -10.1389
+▁lagriculture -10.1389
+▁lédifice -10.1389
+▁messieurs -10.1389
+▁terrible -10.1389
+▁consacre -10.1389
+▁félicit -10.1389
+▁voyons -10.1389
+▁temple -10.1389
+pparition -10.1389
+▁finances -10.1391
+▁puissance -10.1391
+▁saintmartin -10.1394
+illot -10.1402
+▁varie -10.1404
+▁organise -10.1411
+▁indien -10.1414
+▁roger -10.1419
+▁flott -10.1422
+▁maintenu -10.1427
+▁perte -10.1427
+▁brave -10.1439
+mène -10.1453
+▁pié -10.1457
+▁dessus -10.1459
+▁suffit -10.1467
+▁créer -10.1475
+▁humaine -10.1479
+▁municipal -10.1482
+▁elvire -10.1487
+▁titulaire -10.1487
+▁notredame -10.1487
+▁square -10.1487
+▁cinquième -10.1487
+▁bête -10.1487
+▁massif -10.1487
+▁constant -10.1488
+▁nordest -10.1489
+gnol -10.1489
+▁noy -10.1497
+▁penser -10.1503
+▁fréquente -10.1541
+▁traite -10.1547
+issent -10.155
+▁retrouvé -10.1566
+▁morceau -10.1574
+▁conservée -10.1577
+▁courte -10.1582
+▁républicain -10.1586
+▁étroit -10.1586
+▁gérard -10.1586
+▁réparti -10.159
+▁admir -10.1591
+▁acteurs -10.1593
+▁dim -10.1605
+▁paix -10.161
+bach -10.163
+▁méd -10.1633
+ward -10.1657
+▁affect -10.1658
+▁kh -10.1668
+▁guer -10.1671
+▁malheur -10.1683
+▁aussitôt -10.1686
+▁rétabli -10.1686
+▁réfugi -10.1686
+▁patrick -10.1686
+▁adulte -10.1686
+▁souffle -10.1687
+▁commencé -10.1687
+▁simpl -10.1688
+prit -10.1689
+▁gorge -10.1692
+▁représentation -10.1695
+▁croire -10.1699
+▁aj -10.1703
+▁veille -10.1713
+▁fondation -10.1714
+▁considér -10.1714
+▁jules -10.1737
+▁lhé -10.1737
+▁rappeler -10.174
+▁plage -10.1751
+▁abord -10.1761
+▁prenez -10.1768
+ben -10.1769
+mbo -10.177
+▁lieutenant -10.1787
+▁assuré -10.1789
+bul -10.1791
+▁ménage -10.1795
+▁entendre -10.1798
+▁indic -10.1799
+▁supprimé -10.1799
+▁alex -10.1802
+▁certainement -10.1805
+▁vigne -10.1816
+▁étions -10.1816
+▁mariée -10.182
+▁déten -10.1848
+▁bâti -10.1864
+neuf -10.1873
+▁accompagné -10.1889
+▁contemporain -10.1889
+▁enregistre -10.1889
+▁fernand -10.1889
+▁prennent -10.1889
+ibilité -10.1889
+▁boîte -10.1889
+▁gratuit -10.189
+▁murmura -10.189
+▁enfance -10.189
+▁étend -10.1893
+▁tol -10.1896
+▁triste -10.19
+▁couvre -10.1903
+▁juin -10.1904
+▁pratiqu -10.1905
+▁colon -10.1908
+pres -10.1923
+▁dirai -10.1924
+▁loc -10.1932
+▁longueur -10.1939
+orge -10.1967
+▁grim -10.1968
+▁appren -10.1973
+▁arrivée -10.1976
+▁appartement -10.1992
+▁kilomètres -10.1992
+▁lorganisation -10.1992
+▁religion -10.1992
+▁commandant -10.1992
+▁assassin -10.1992
+▁laéroport -10.1992
+▁faveur -10.1992
+▁gardien -10.1995
+▁fixé -10.1996
+▁modeste -10.1997
+▁avance -10.1998
+▁geste -10.2006
+▁mangé -10.2011
+▁jan -10.2012
+▁signé -10.2013
+▁damien -10.2017
+▁narr -10.2019
+▁liv -10.202
+▁fréquent -10.2039
+▁convient -10.2066
+▁cave -10.2067
+▁actuelle -10.207
+olog -10.2071
+▁parut -10.2084
+▁rencontrent -10.2085
+▁réussit -10.2096
+▁circonscription -10.2096
+▁température -10.2096
+▁salarié -10.2096
+▁statut -10.2097
+▁menace -10.21
+▁ennemi -10.2103
+▁toilette -10.2117
+▁concernant -10.2118
+▁mention -10.2123
+▁réalisation -10.2123
+▁ouest -10.2126
+▁dol -10.2128
+▁italienne -10.2132
+▁vitesse -10.2133
+▁sentier -10.2146
+▁bateau -10.2158
+▁drapeau -10.2175
+▁régul -10.2177
+▁prête -10.2201
+š -10.2202
+▁larrivée -10.2202
+▁possibilité -10.2202
+▁compagnon -10.2202
+▁compositeur -10.2202
+▁continent -10.2202
+▁seraient -10.2202
+▁russie -10.2202
+▁chevaux -10.2203
+▁faubourg -10.2205
+▁humide -10.2206
+▁crée -10.2207
+▁lentement -10.2211
+▁alain -10.2222
+▁lord -10.2239
+▁amant -10.2242
+cession -10.2252
+▁alexandre -10.2254
+▁ski -10.2254
+▁rock -10.2254
+bord -10.2267
+américain -10.2271
+▁livr -10.2272
+viez -10.2273
+▁marine -10.228
+▁sengage -10.2287
+▁battu -10.2291
+▁canon -10.2294
+▁supérieure -10.2303
+▁compétence -10.2308
+▁inutile -10.2308
+▁performance -10.2308
+▁remarquable -10.2308
+▁responsabilité -10.2308
+▁sintéresse -10.2308
+▁universitaire -10.2308
+▁accident -10.2308
+▁frappé -10.2308
+▁cercle -10.2309
+▁rempli -10.2309
+▁frappe -10.231
+▁crédit -10.2314
+▁brou -10.2315
+▁psycho -10.2317
+▁hauss -10.2321
+▁anna -10.2334
+▁george -10.2336
+istan -10.2336
+▁sacré -10.2339
+▁traitement -10.2342
+gur -10.2349
+▁war -10.2359
+▁regroupe -10.2365
+▁vivent -10.2371
+▁sénat -10.2375
+stein -10.2391
+▁chamb -10.2407
+▁initiale -10.2416
+▁toulouse -10.2416
+▁vainqueur -10.2416
+▁relève -10.2416
+▁protection -10.2416
+▁soyez -10.2416
+▁décidé -10.2416
+▁of -10.2416
+▁publie -10.2416
+▁inspiré -10.2416
+▁christian -10.2417
+▁gouverneur -10.2418
+▁invite -10.2419
+▁locaux -10.242
+▁sembl -10.2421
+▁dormir -10.2422
+▁balle -10.2424
+▁zi -10.2429
+▁spécialisé -10.2434
+▁lève -10.2435
+kov -10.2466
+card -10.247
+▁pourcent -10.2483
+▁sauf -10.2484
+▁plume -10.2491
+burg -10.2496
+onsnous -10.2499
+▁qualifié -10.2501
+▁imagin -10.2507
+reuil -10.2508
+▁ellemême -10.2524
+▁extérieur -10.2524
+▁olivier -10.2524
+▁deurope -10.2524
+▁démarche -10.2524
+▁afrique -10.2524
+▁évoque -10.2524
+accio -10.2525
+▁distance -10.2525
+▁arrière -10.2525
+▁jeanpierre -10.2527
+▁justifi -10.2528
+▁west -10.2531
+▁gagn -10.2533
+▁she -10.2538
+▁devra -10.2545
+▁tienne -10.2548
+▁oublié -10.2551
+▁chanteur -10.2579
+▁limp -10.2604
+▁romaine -10.2616
+▁fidèle -10.2634
+▁restauration -10.2634
+▁artistique -10.2634
+▁commença -10.2634
+▁phare -10.2635
+▁poids -10.2635
+▁éclat -10.2637
+▁échange -10.2637
+▁civile -10.2637
+▁loncle -10.2638
+▁sportif -10.264
+franc -10.2653
+plan -10.2659
+▁égal -10.2661
+vaux -10.2666
+▁shi -10.2671
+▁faim -10.2671
+▁facilement -10.2673
+▁condamné -10.2676
+▁adopt -10.2677
+▁mineur -10.2678
+▁queue -10.2679
+urban -10.2692
+▁maman -10.2707
+accompagne -10.2708
+clair -10.2723
+▁influencé -10.2745
+▁religieux -10.2745
+▁réflexion -10.2745
+▁accessible -10.2745
+▁génération -10.2745
+▁rassembl -10.2745
+▁énorme -10.2745
+▁femelle -10.2745
+▁sagissait -10.2745
+▁angleterre -10.2746
+ordinateur -10.2748
+▁hein -10.2748
+▁fan -10.2749
+▁sensible -10.275
+▁lesprit -10.2763
+▁franchi -10.2764
+▁accompagne -10.2783
+▁mainten -10.2792
+▁mette -10.2827
+▁lundi -10.2832
+▁top -10.2838
+▁brune -10.2845
+▁professionnelle -10.2845
+dge -10.2851
+▁exception -10.2858
+▁catherine -10.2858
+▁inférieur -10.2858
+▁jattend -10.2858
+installation -10.2858
+▁considère -10.2858
+▁amérique -10.2858
+▁précisément -10.2858
+▁effort -10.2858
+▁affich -10.2858
+▁chauff -10.2858
+▁utile -10.2858
+▁ruisseau -10.2858
+▁lespace -10.2858
+▁recevoir -10.2858
+▁établissement -10.2859
+▁édité -10.2859
+▁sépare -10.2863
+▁déplace -10.2865
+▁nulle -10.2879
+▁majeure -10.288
+▁logement -10.2882
+▁capable -10.2885
+▁écoute -10.289
+vision -10.2894
+▁gentil -10.2895
+▁peter -10.2897
+▁déclin -10.2905
+▁bill -10.2905
+▁perdre -10.2946
+▁canal -10.2956
+▁univers -10.2967
+▁honte -10.2967
+▁démission -10.2971
+▁lenseignement -10.2971
+▁littéraire -10.2971
+▁mécanisme -10.2971
+▁médicament -10.2971
+▁bruxelles -10.2971
+▁primaire -10.2971
+▁septembre -10.2971
+▁brillant -10.2972
+▁légion -10.2972
+▁single -10.2972
+▁réel -10.2973
+▁galerie -10.2976
+▁décret -10.2976
+▁justin -10.2976
+▁héros -10.2976
+▁picard -10.2977
+▁variété -10.2977
+▁allezvous -10.2978
+▁appris -10.298
+▁recette -10.2981
+▁dessert -10.2982
+ouche -10.2987
+▁accueilli -10.2988
+▁mark -10.2989
+▁billet -10.2992
+échappe -10.3003
+▁succ -10.3005
+▁mien -10.3007
+surmer -10.3009
+▁tableaux -10.3014
+▁occup -10.3024
+▁image -10.3024
+▁mc -10.3037
+▁castel -10.3046
+▁born -10.3074
+▁spi -10.3075
+▁concept -10.3081
+▁effective -10.3086
+▁fédération -10.3086
+▁réclam -10.3086
+▁protégé -10.3087
+▁amoureux -10.309
+▁déposé -10.3099
+▁lhabit -10.3101
+▁semblait -10.3109
+interroge -10.312
+▁marius -10.3133
+▁vice -10.3136
+▁soirée -10.3144
+▁verra -10.316
+▁majeur -10.318
+▁distribu -10.3196
+▁civil -10.32
+▁profession -10.32
+▁mavez -10.3201
+▁disparaît -10.3202
+▁effectué -10.3202
+▁léducation -10.3203
+▁arnaud -10.3203
+atteindre -10.3203
+▁croissance -10.3204
+▁fallu -10.3204
+▁divisé -10.3204
+▁mètres -10.3206
+▁internet -10.3209
+▁ajouta -10.3211
+▁pape -10.3212
+▁debout -10.3214
+▁christophe -10.3216
+▁naturellement -10.3237
+▁tourisme -10.3242
+scendant -10.3259
+ifiée -10.3265
+▁eaux -10.3281
+alisation -10.3294
+wan -10.3294
+▁bah -10.33
+▁apprécié -10.332
+▁réservé -10.332
+▁lutilisation -10.332
+▁noël -10.332
+▁smith -10.332
+▁mexique -10.332
+▁concours -10.3321
+▁leffet -10.3322
+▁usage -10.3323
+▁puits -10.3327
+▁spécialiste -10.3329
+▁conception -10.333
+▁sinscri -10.3331
+▁chanteuse -10.334
+▁poursuivi -10.3341
+fond -10.3343
+▁défenseur -10.3348
+▁james -10.3357
+▁lappel -10.3363
+▁soient -10.337
+▁final -10.3398
+dieu -10.341
+▁dispos -10.3412
+▁faux -10.3418
+▁gré -10.3424
+▁habite -10.3435
+▁larchitecte -10.3439
+▁ouvrier -10.3439
+▁ressemble -10.3439
+▁similaire -10.3439
+▁épreuve -10.3439
+▁évêque -10.3439
+▁diffusé -10.3439
+▁dimension -10.3439
+▁lopposition -10.3439
+▁polonais -10.3439
+▁veulent -10.3439
+▁célébr -10.3439
+▁suivre -10.3439
+▁honneur -10.3439
+▁loge -10.344
+▁active -10.344
+▁plaît -10.3444
+▁administr -10.3454
+▁navon -10.3471
+▁rural -10.3474
+posé -10.348
+▁règlement -10.3482
+▁oncle -10.3487
+▁musicale -10.3488
+án -10.3489
+bru -10.3503
+vière -10.352
+▁syn -10.3521
+françois -10.3522
+▁entraîne -10.3545
+époux -10.3553
+▁retourna -10.3555
+grand -10.3556
+▁condition -10.3559
+▁déjeuner -10.356
+▁limitrophe -10.356
+▁royaumeuni -10.356
+▁cellule -10.356
+▁jentend -10.356
+▁montréal -10.356
+▁lionel -10.356
+▁juillet -10.356
+▁clef -10.356
+▁mérite -10.356
+▁bouteille -10.356
+▁gestion -10.3561
+▁chaussures -10.3561
+▁collaboration -10.3561
+▁charmant -10.3563
+▁université -10.3566
+▁feuille -10.3575
+▁sauve -10.361
+▁wil -10.3611
+▁dho -10.3635
+poli -10.3635
+chard -10.3642
+▁coeur -10.3646
+▁achev -10.3646
+▁tirer -10.3651
+▁dhu -10.3666
+▁blessure -10.3682
+▁conflit -10.3682
+▁conférence -10.3682
+▁manuscrit -10.3682
+▁agréable -10.3682
+▁interromp -10.3682
+▁mâle -10.3682
+▁séjour -10.3682
+▁mélange -10.3682
+▁normand -10.3682
+▁bourgeois -10.3682
+▁détude -10.3682
+▁tribu -10.3696
+▁ronde -10.3704
+▁tenez -10.371
+script -10.3711
+▁stra -10.3724
+▁motif -10.3726
+▁extr -10.3727
+▁sourd -10.3732
+chère -10.3733
+▁navais -10.3735
+▁chère -10.3749
+▁chercha -10.3758
+▁arrêt -10.3763
+▁ditelle -10.3782
+▁conscience -10.3805
+▁processus -10.3805
+▁spécifique -10.3805
+▁véhicule -10.3805
+▁étrangère -10.3805
+▁récemment -10.3805
+▁bretagne -10.3805
+▁sarrête -10.3805
+▁larrière -10.3805
+▁éliminé -10.3805
+▁gêne -10.3805
+▁tropical -10.381
+▁mairie -10.3819
+▁beaux -10.3832
+▁comtesse -10.3836
+▁mondial -10.3873
+▁fondé -10.3874
+▁lancé -10.3902
+▁ravi -10.3917
+▁saintgermain -10.393
+▁attribué -10.393
+▁fortdefrance -10.393
+▁larchitecture -10.393
+▁monseigneur -10.393
+▁rencontré -10.393
+▁récompense -10.393
+▁signifie -10.393
+▁xavier -10.393
+▁participé -10.393
+▁réaction -10.393
+▁défaite -10.3931
+▁drame -10.3931
+▁nétaient -10.3933
+▁termina -10.3933
+vènement -10.3933
+▁permis -10.3933
+▁personnelle -10.3936
+▁entrée -10.3937
+▁forcé -10.3941
+▁identique -10.3949
+▁aigu -10.3952
+▁précis -10.3954
+▁éric -10.3965
+▁vivement -10.3966
+▁vélo -10.3973
+▁réelle -10.3976
+▁notion -10.3978
+▁plaine -10.3982
+▁pareille -10.399
+▁serre -10.4009
+▁von -10.4012
+▁serge -10.4014
+bré -10.4025
+▁lamb -10.4044
+▁oiseau -10.4045
+▁cérémonie -10.4057
+▁diplôme -10.4057
+▁lhôpital -10.4057
+▁organisation -10.4057
+▁laccès -10.4057
+▁cardinal -10.4057
+▁légende -10.4057
+▁strasbourg -10.4057
+▁dheure -10.4057
+▁abouti -10.4058
+▁fusion -10.4059
+▁certes -10.406
+chou -10.4062
+▁jusque -10.4063
+▁frappa -10.4063
+▁confond -10.4065
+▁micro -10.4069
+▁négociation -10.4077
+▁grec -10.4089
+▁toile -10.4095
+▁bloc -10.4113
+▁croyez -10.4122
+▁pli -10.4123
+▁porc -10.4165
+ł -10.4185
+▁domicile -10.4185
+▁familial -10.4185
+▁introduit -10.4185
+▁mathématique -10.4185
+▁pêchai -10.4185
+▁renommé -10.4185
+▁provoque -10.4185
+▁achète -10.4185
+▁germain -10.4185
+▁fortune -10.4186
+▁sportive -10.419
+▁tandi -10.4191
+▁nimporte -10.4195
+▁socialiste -10.4206
+▁lacteur -10.4207
+▁suffi -10.4208
+▁pousse -10.4214
+▁corne -10.4223
+▁guide -10.4226
+▁commis -10.4226
+▁apport -10.4239
+▁liée -10.4242
+▁sauver -10.4283
+idité -10.429
+▁quau -10.4295
+▁efficace -10.4315
+▁enregistré -10.4315
+▁novembre -10.4315
+▁ernest -10.4315
+▁transformé -10.4315
+▁frémi -10.4316
+▁leurope -10.4316
+▁modification -10.4316
+▁reproche -10.4317
+▁nourriture -10.4317
+▁diable -10.4319
+▁litalie -10.4319
+▁institution -10.4322
+▁comédien -10.4323
+▁boucle -10.4325
+▁angeles -10.4327
+ical -10.4329
+▁heureusement -10.4338
+▁officiellement -10.434
+▁repas -10.4346
+▁moindre -10.4348
+▁lille -10.441
+mini -10.4411
+▁puisse -10.4424
+▁oubli -10.4426
+▁individu -10.4429
+▁bern -10.4433
+▁entière -10.4438
+rupt -10.4446
+▁réfléchi -10.4446
+▁expérience -10.4446
+▁patrimoine -10.4446
+▁promenade -10.4446
+▁soviétique -10.4446
+▁circuit -10.4446
+▁espagne -10.4446
+▁michael -10.4446
+▁matériau -10.4446
+▁marcof -10.4446
+▁œil -10.4447
+▁bonjour -10.4447
+▁aventure -10.4447
+▁lennemi -10.4447
+▁hôtel -10.4447
+▁aîné -10.4447
+▁volume -10.4448
+▁soupir -10.4451
+▁parcours -10.4453
+▁notable -10.4453
+▁voulais -10.4454
+▁émis -10.4454
+▁attire -10.4455
+▁salut -10.4465
+▁talent -10.4468
+investi -10.4471
+▁délai -10.4472
+▁rayon -10.4473
+▁courir -10.4486
+▁élément -10.4493
+▁hiver -10.4498
+ône -10.4499
+▁léger -10.4502
+▁métro -10.4505
+▁chou -10.4516
+▁subit -10.4534
+enberg -10.4539
+▁recueil -10.4572
+▁regardait -10.4572
+▁vêtements -10.458
+▁faculté -10.458
+▁lorsquil -10.458
+▁pleut -10.458
+▁lactivité -10.458
+▁touristique -10.458
+▁viande -10.4581
+▁sagesse -10.4581
+▁réalisateur -10.4584
+▁valérie -10.4588
+▁parlait -10.4589
+▁tapis -10.4593
+▁officielle -10.4594
+▁gall -10.4599
+▁soins -10.4606
+▁bourse -10.4609
+▁abandonne -10.4612
+▁tiré -10.4615
+illage -10.4621
+▁confié -10.4621
+▁colomb -10.4622
+▁quasi -10.4623
+▁chasseur -10.4661
+▁romp -10.4677
+▁culturel -10.4697
+▁tendre -10.4701
+▁écriv -10.4707
+brivelagaillarde -10.4715
+▁administratif -10.4715
+▁capacité -10.4715
+▁dîner -10.4715
+▁ingénieur -10.4715
+▁majoritaire -10.4715
+▁procès -10.4715
+▁photographie -10.4715
+▁loreille -10.4715
+▁lâche -10.4715
+▁vierge -10.4716
+▁atelier -10.4716
+▁lactrice -10.472
+haut -10.4725
+▁violon -10.4726
+▁bourbon -10.4726
+▁oiseaux -10.4728
+▁cadet -10.476
+▁gamin -10.4766
+▁multi -10.4771
+▁soif -10.4784
+châtel -10.4791
+▁constitution -10.4845
+▁explication -10.4852
+▁sapplique -10.4852
+▁suffisamment -10.4852
+▁symbole -10.4852
+▁adapté -10.4852
+▁dialogue -10.4852
+▁dinformation -10.4852
+▁pantalon -10.4852
+▁soutenu -10.4852
+xtrémité -10.4852
+▁létude -10.4852
+▁attaché -10.4852
+▁journaux -10.4852
+▁déclare -10.4852
+▁avril -10.4852
+▁chute -10.4858
+▁aimait -10.4877
+▁corde -10.4879
+villiers -10.4885
+▁exige -10.489
+▁colline -10.4899
+▁herb -10.4902
+▁prenant -10.4903
+▁soldat -10.4922
+vienne -10.4922
+▁sinon -10.4929
+▁essayé -10.4931
+offic -10.4933
+iforme -10.4941
+▁vision -10.495
+influence -10.4953
+▁nouvel -10.4959
+▁poul -10.496
+▁contenu -10.4974
+▁plaque -10.4984
+▁alfred -10.4991
+▁construire -10.4991
+▁intégré -10.4991
+▁lévêque -10.4991
+▁marguerite -10.4991
+▁octobre -10.4991
+▁suspension -10.4991
+▁viceprésident -10.4991
+▁avancée -10.4991
+▁impérial -10.4991
+▁intéressant -10.4991
+▁lécriture -10.4991
+▁précipita -10.4991
+▁standard -10.4991
+▁communication -10.4991
+▁keinec -10.4991
+▁maxime -10.4991
+▁terrasse -10.4991
+▁poésie -10.4991
+▁complément -10.4991
+▁soumis -10.4996
+▁douleur -10.4998
+▁arrange -10.5006
+artisan -10.5014
+▁paquet -10.5014
+loup -10.5022
+▁tiers -10.5031
+▁pressé -10.5036
+▁niveaux -10.5048
+▁avion -10.5073
+▁officiel -10.5089
+aël -10.5101
+▁cassé -10.5111
+▁god -10.5128
+▁délégué -10.5132
+▁maréchal -10.5132
+▁méditerranée -10.5132
+▁scénario -10.5132
+▁conservatoire -10.5132
+▁bronze -10.5132
+▁lavocat -10.5132
+▁réputé -10.5132
+▁janvier -10.5132
+▁conversation -10.5132
+▁sarrêta -10.5132
+▁électronique -10.5132
+▁reconnaissance -10.5132
+ographe -10.5132
+▁étais -10.5133
+▁genoux -10.5133
+▁prolonge -10.5134
+▁limpression -10.5135
+▁reten -10.5138
+▁annuel -10.5139
+▁message -10.514
+▁enseignant -10.5142
+▁confirme -10.5152
+▁entraîneur -10.5153
+▁garanti -10.5158
+▁béni -10.517
+▁tourné -10.519
+nny -10.5205
+▁culturelle -10.522
+▁adopté -10.5229
+blanc -10.523
+▁dirigeant -10.5251
+▁comédie -10.5266
+▁exemplaire -10.5274
+▁immense -10.5274
+▁sévère -10.5274
+▁lhiver -10.5275
+▁produis -10.5275
+▁discipline -10.5275
+▁modifie -10.5275
+▁secria -10.5275
+▁impression -10.5277
+▁colonie -10.5285
+▁violent -10.5286
+▁modifié -10.5286
+▁caisse -10.5298
+▁veuxtu -10.53
+▁sentait -10.5307
+asso -10.5312
+▁navaient -10.5318
+▁boire -10.5331
+▁avantage -10.537
+▁délivr -10.5383
+▁parisien -10.5388
+▁naurai -10.5398
+crire -10.5399
+améliorer -10.5419
+▁sculpture -10.5419
+▁soupçonn -10.5419
+▁stéphan -10.5419
+▁libéral -10.5419
+▁rattaché -10.5419
+▁traduction -10.5419
+▁pologne -10.5419
+▁fatigue -10.5419
+plomb -10.5419
+▁août -10.542
+▁pénétr -10.542
+▁lissue -10.542
+▁subsiste -10.542
+▁culte -10.5421
+▁assistant -10.5421
+▁facteur -10.5421
+▁dossier -10.5421
+▁paysbas -10.5423
+▁lavez -10.5432
+▁lunettes -10.5432
+▁seigneurie -10.5435
+▁découvrir -10.5439
+▁rennes -10.5442
+▁éviter -10.5444
+▁examin -10.5445
+▁yann -10.5447
+▁proposons -10.5449
+▁rival -10.5455
+▁repart -10.5458
+▁sixième -10.5464
+▁aimerait -10.5469
+▁retenu -10.5512
+▁colombie -10.5523
+▁souci -10.5541
+▁serra -10.5543
+ñ -10.5566
+▁dhistoire -10.5566
+▁lobjectif -10.5566
+▁néerlandais -10.5566
+▁souterrain -10.5566
+interprète -10.5566
+▁espoir -10.5566
+▁dommage -10.5566
+▁chemise -10.5566
+surloire -10.5566
+▁empêche -10.5566
+▁thèse -10.5566
+▁veuve -10.5566
+▁policier -10.5566
+▁exclu -10.5567
+▁jirai -10.5567
+▁sympa -10.5567
+▁opposé -10.5568
+▁craign -10.5568
+▁souffre -10.5568
+▁violence -10.5569
+▁bordeaux -10.5574
+▁sucre -10.5574
+▁amateur -10.5582
+issaient -10.5584
+▁pluie -10.5585
+▁footballeur -10.5588
+coup -10.56
+▁dense -10.5606
+▁orateur -10.5613
+martin -10.5628
+▁renard -10.5629
+ntendre -10.5635
+▁communiste -10.5643
+ctuelle -10.5651
+▁seuil -10.5658
+▁régulière -10.57
+▁modern -10.5701
+▁ferai -10.5704
+▁financier -10.5716
+▁lespèce -10.5716
+▁secrète -10.5716
+▁successeur -10.5716
+▁vôtre -10.5716
+▁chassaigne -10.5716
+▁réputation -10.5716
+▁indiqué -10.5716
+▁trafic -10.5716
+▁lâme -10.5716
+surseine -10.5716
+▁mappelle -10.5716
+▁notaire -10.5716
+▁florence -10.5717
+▁win -10.572
+▁latéral -10.5721
+augustin -10.5722
+▁touché -10.5724
+▁nantes -10.5725
+▁lentend -10.5745
+▁taxe -10.5752
+▁trahi -10.5771
+▁web -10.578
+▁viol -10.5784
+▁gagné -10.5784
+▁chienne -10.5808
+▁soeur -10.5815
+▁pétr -10.582
+▁sto -10.5831
+▁oct -10.5835
+▁sienne -10.5842
+well -10.5847
+ć -10.5867
+▁additionnel -10.5867
+▁fabrication -10.5867
+▁lexistence -10.5867
+▁gaillard -10.5867
+▁évoqué -10.5867
+▁décembre -10.5867
+▁posséda -10.5867
+▁costume -10.5867
+▁conduis -10.5867
+▁croyais -10.5867
+▁envoyer -10.5867
+▁consiste -10.5872
+▁amoureuse -10.5874
+▁lartiste -10.5876
+▁tendance -10.588
+▁grandpère -10.5881
+▁maroc -10.589
+▁princesse -10.5901
+▁jim -10.5903
+▁dinter -10.5904
+▁émet -10.5906
+train -10.5913
+▁planche -10.5915
+▁iron -10.5959
+pierre -10.5974
+jacques -10.5983
+avant -10.6007
+▁architecte -10.6021
+▁circonstance -10.6021
+▁circulaire -10.6021
+▁importance -10.6021
+▁montpellier -10.6021
+▁plateforme -10.6021
+▁progressivement -10.6021
+▁robiliard -10.6021
+▁libération -10.6021
+▁fondamental -10.6021
+▁portugal -10.6021
+▁révèle -10.6021
+▁limoges -10.6021
+▁ordinaire -10.6021
+▁tambour -10.6021
+▁dragon -10.6021
+▁henry -10.6021
+▁souffrir -10.6021
+▁émile -10.6021
+▁mourut -10.6021
+▁despagne -10.6022
+▁occidentale -10.6022
+▁envahi -10.6022
+▁dannées -10.6022
+▁partenaire -10.6022
+identité -10.6022
+▁moscou -10.6022
+▁automatique -10.6023
+▁plastique -10.6023
+▁vouloir -10.6024
+▁robinet -10.6025
+▁ditalie -10.6026
+▁vincent -10.6028
+▁foyer -10.603
+▁chair -10.6033
+▁venons -10.6034
+▁manuel -10.6037
+▁morceaux -10.604
+▁annie -10.6053
+▁publicité -10.6055
+fou -10.6094
+trop -10.6096
+▁music -10.6114
+▁tha -10.6143
+▁bré -10.6145
+▁décoré -10.6164
+▁surpris -10.6169
+▁essentiel -10.6176
+▁application -10.6177
+▁congrès -10.6177
+▁financière -10.6177
+▁généraux -10.6177
+▁poitrine -10.6177
+▁spirituel -10.6177
+▁denviron -10.6177
+▁fédéral -10.6177
+▁recouvert -10.6177
+▁sociaux -10.6177
+▁engagement -10.6177
+▁guitare -10.6177
+▁libéré -10.6177
+▁apparu -10.6177
+▁deviendra -10.6178
+▁opéra -10.6178
+▁ajouté -10.6178
+▁discours -10.6178
+▁émission -10.6179
+▁embrass -10.618
+▁damour -10.6181
+▁milliers -10.6183
+▁toulon -10.6183
+▁voile -10.6183
+▁hydro -10.6184
+▁carrier -10.6189
+▁sénateur -10.6195
+▁parlons -10.6196
+▁surmonté -10.6198
+▁juive -10.6211
+▁soulev -10.6213
+▁clip -10.6293
+antoine -10.6321
+▁traditionnel -10.6322
+▁géré -10.6328
+▁apporté -10.6331
+č -10.6336
+▁indispensable -10.6336
+▁refroidi -10.6336
+▁saintquentin -10.6336
+▁sculpteur -10.6336
+▁appareil -10.6336
+▁nôtre -10.6336
+▁suédois -10.6336
+▁typique -10.6336
+▁disputé -10.6336
+▁radical -10.6336
+▁déficit -10.6336
+▁plaisant -10.6336
+xtension -10.6336
+pparence -10.6336
+▁voyant -10.6337
+▁renonce -10.6338
+▁déclara -10.6338
+▁aimable -10.6341
+▁fût -10.6341
+▁répéta -10.6342
+annonce -10.6343
+▁arabe -10.6345
+field -10.6346
+▁conven -10.6347
+provence -10.6349
+▁ramass -10.6349
+▁évolution -10.635
+▁orientale -10.6357
+▁pente -10.6372
+▁gramm -10.6418
+▁vaud -10.6453
+▁heureuse -10.6491
+▁cidessous -10.6497
+▁collabore -10.6497
+▁enceinte -10.6497
+▁honnête -10.6497
+▁magazine -10.6497
+▁renforce -10.6497
+▁camarade -10.6497
+▁delphine -10.6497
+▁jémet -10.6497
+▁obligé -10.6497
+▁remarqué -10.6497
+lesbains -10.6497
+▁mécanique -10.6497
+▁escalier -10.6497
+▁sèche -10.6497
+▁profondément -10.6497
+▁environnement -10.6498
+▁âgé -10.6498
+▁pâle -10.6498
+▁dhomme -10.6499
+▁bouge -10.65
+▁partagé -10.65
+▁douvrir -10.65
+▁stock -10.6502
+▁malheureux -10.6502
+▁coureur -10.6502
+▁maigre -10.6505
+▁secours -10.6506
+▁villeneuve -10.651
+▁préfecture -10.6515
+▁cite -10.6524
+▁promet -10.6528
+▁décoration -10.6537
+▁rideau -10.6549
+▁rougi -10.6552
+hypothèse -10.6661
+▁gendarme -10.6661
+▁informatique -10.6661
+▁lallemagne -10.6661
+▁létranger -10.6661
+▁pacifique -10.6661
+▁territorial -10.6661
+▁œufs -10.6661
+▁brusquement -10.6661
+▁héricourt -10.6661
+▁chicago -10.6661
+▁enquête -10.6661
+▁lemploi -10.6661
+▁dispers -10.6661
+▁miracle -10.6661
+▁chanoine -10.6661
+▁liaison -10.6662
+▁audelà -10.6662
+▁césar -10.6662
+▁tribunal -10.6662
+▁irré -10.6662
+▁lointain -10.6666
+▁larrêt -10.6667
+▁plaint -10.6667
+▁collaborateur -10.6667
+▁vocation -10.667
+▁fruit -10.6672
+▁minute -10.6676
+▁demeura -10.6702
+▁répandu -10.6703
+▁spécial -10.6704
+▁affecté -10.6714
+▁grain -10.6725
+mbassadeur -10.6728
+ayez -10.6742
+ançai -10.6765
+▁sérieuse -10.6771
+▁sho -10.6777
+▁herman -10.6824
+▁délégation -10.6828
+▁numérique -10.6828
+▁pistolet -10.6828
+▁saintlaurent -10.6828
+▁suprême -10.6828
+▁vieillard -10.6828
+▁drôle -10.6828
+▁déménage -10.6828
+▁emprunt -10.6828
+▁promotion -10.6828
+▁laccord -10.6828
+▁excusezmoi -10.6828
+▁trésor -10.6828
+▁étienne -10.6828
+▁armand -10.6828
+▁pitié -10.6829
+▁lescalier -10.6829
+▁gâteau -10.6829
+▁échec -10.6829
+▁parfum -10.6832
+▁champagne -10.6833
+▁quitté -10.6834
+▁existence -10.684
+▁agit -10.6841
+▁format -10.6843
+▁prouve -10.6845
+▁profondeur -10.6845
+▁couverture -10.6846
+▁couronne -10.6848
+▁teint -10.6857
+▁couteau -10.6867
+▁repli -10.6882
+▁sélectionné -10.6884
+rhône -10.6895
+▁lattend -10.6915
+▁franco -10.6927
+claude -10.696
+ouvrit -10.6964
+▁fuir -10.6983
+▁accompli -10.6997
+▁espérance -10.6997
+▁létablissement -10.6997
+▁récupér -10.6997
+▁emprisonn -10.6997
+▁lépouse -10.6997
+▁intervention -10.6997
+▁linstant -10.6998
+▁virginie -10.6998
+▁diamètre -10.6999
+▁autorisé -10.6999
+▁décès -10.7007
+▁deuros -10.7008
+▁monstre -10.7009
+▁véron -10.7024
+▁complet -10.7025
+pouvant -10.703
+▁allié -10.703
+hôte -10.7041
+▁artiste -10.7043
+amine -10.7046
+▁releva -10.7048
+ight -10.7061
+▁condamn -10.7104
+▁thu -10.7122
+ntraîneur -10.7164
+▁basketball -10.717
+▁cohérence -10.717
+▁jeanbaptiste -10.717
+▁judiciaire -10.717
+▁ladministration -10.717
+▁lautorité -10.717
+▁suspendu -10.717
+▁incertain -10.717
+▁chœur -10.717
+▁trentaine -10.717
+▁éteint -10.717
+▁témoin -10.717
+▁fabrique -10.717
+▁apparaissent -10.717
+▁fraysse -10.717
+équipage -10.717
+▁acquis -10.717
+▁suivent -10.717
+▁algérie -10.717
+▁expose -10.717
+▁promesse -10.717
+▁doublé -10.717
+▁variante -10.717
+▁variable -10.717
+▁lèvres -10.7171
+▁sousmarin -10.7171
+▁sébastien -10.7172
+▁connait -10.7172
+▁gaston -10.7174
+▁isolé -10.7178
+▁reims -10.7179
+▁degré -10.7181
+acteur -10.7192
+▁aidé -10.721
+▁passionné -10.723
+▁morale -10.7234
+▁molle -10.7254
+▁lorient -10.7258
+▁archéologique -10.7345
+▁irlandais -10.7345
+▁napoléon -10.7345
+▁saintétienne -10.7345
+▁sanctuaire -10.7345
+▁scrutin -10.7345
+intermédiaire -10.7345
+▁circulation -10.7345
+▁historien -10.7345
+▁expert -10.7345
+▁abondant -10.7345
+▁félix -10.7345
+▁lessentiel -10.7345
+▁locéan -10.7345
+▁autrichien -10.7345
+▁tentative -10.7345
+▁confirmé -10.7345
+▁arthur -10.7345
+▁résolu -10.7345
+▁nentend -10.7345
+▁fantas -10.7346
+▁jazz -10.7346
+▁préparé -10.7347
+▁samedi -10.7347
+▁pension -10.7348
+▁admis -10.7348
+▁désigné -10.7348
+▁copie -10.735
+▁patient -10.7353
+▁constitutionnel -10.7353
+▁balcon -10.7362
+▁fuite -10.7365
+▁pérou -10.7373
+▁médor -10.7377
+▁recueilli -10.7384
+▁musical -10.7403
+▁adam -10.7415
+▁santa -10.7492
+▁pareil -10.7501
+▁africain -10.7524
+▁aménagé -10.7524
+▁artificiel -10.7524
+▁grandebretagne -10.7524
+▁renaissance -10.7524
+▁technologie -10.7524
+▁potentiel -10.7524
+▁registre -10.7524
+▁scandale -10.7524
+▁excuse -10.7524
+▁lœil -10.7524
+▁daffaires -10.7524
+▁beauxarts -10.7524
+▁gaudin -10.7524
+▁fâch -10.7525
+▁lespagne -10.7525
+▁animé -10.7527
+▁lévolution -10.7527
+▁logis -10.7528
+▁lattaque -10.7528
+▁usine -10.7531
+▁lombre -10.7534
+▁flore -10.7534
+▁voila -10.7534
+▁lesquels -10.7535
+▁partisan -10.7537
+▁voyageurs -10.754
+▁décrire -10.754
+▁cellelà -10.7546
+graphie -10.756
+▁filière -10.7565
+anluc -10.7574
+▁dappel -10.767
+▁négoci -10.7677
+▁référ -10.7703
+▁hockey -10.7706
+▁hongrois -10.7706
+▁laboratoire -10.7706
+▁luxembourg -10.7706
+▁lécrivain -10.7706
+▁lémission -10.7706
+▁personnalité -10.7706
+▁économie -10.7706
+▁analyse -10.7706
+▁rythme -10.7706
+▁ultérieur -10.7706
+▁appliqué -10.7706
+▁rupture -10.7706
+▁mensonge -10.7706
+▁gloire -10.7706
+▁serbie -10.7706
+▁liquide -10.7706
+▁licence -10.7706
+▁déclaré -10.7706
+▁foncé -10.7706
+▁parfait -10.7707
+▁barrage -10.7707
+▁soudain -10.7707
+▁échoue -10.7708
+▁connaissent -10.7708
+▁curieux -10.7709
+▁hélas -10.7709
+chonnet -10.771
+▁rentré -10.7716
+▁leader -10.7724
+▁grecque -10.7729
+▁renvoy -10.7731
+▁sérieux -10.7736
+▁parrain -10.7752
+ouillet -10.7765
+crate -10.7816
+▁iso -10.7826
+▁philip -10.7888
+▁robin -10.7889
+▁benjamin -10.7891
+▁convaincu -10.7891
+▁coordination -10.7891
+▁doctorat -10.7891
+▁géographique -10.7891
+▁jacquelin -10.7891
+▁mathilde -10.7891
+▁obligatoire -10.7891
+▁photographe -10.7891
+▁printemps -10.7891
+▁saintgeorges -10.7891
+▁satisfaction -10.7891
+▁superficie -10.7891
+▁nêtes -10.7891
+▁accoyer -10.7891
+▁alinéa -10.7891
+▁procédé -10.7891
+▁adversaire -10.7891
+▁thierry -10.7891
+▁aérienne -10.7891
+▁fragile -10.7891
+▁scolaire -10.7891
+▁chêne -10.7891
+▁labsence -10.7891
+quatrevingtseize -10.7891
+▁inquiétude -10.7891
+▁éditeur -10.7891
+▁baroque -10.7891
+▁meuble -10.7892
+▁décida -10.7893
+▁plafond -10.7893
+▁grandmère -10.7901
+▁planète -10.7903
+mbition -10.7913
+▁varié -10.7914
+▁étude -10.7971
+iculture -10.7985
+▁distinct -10.8063
+▁immédiat -10.8074
+▁demifinale -10.8079
+▁exclusivement -10.8079
+▁fréquemment -10.8079
+▁guitariste -10.8079
+▁jeanlouis -10.8079
+▁maîtrise -10.8079
+▁nîmes -10.8079
+▁scénariste -10.8079
+▁cavalerie -10.8079
+▁entreprend -10.8079
+▁explor -10.8079
+▁explos -10.8079
+▁préparation -10.8079
+▁merveille -10.8079
+▁saintdenis -10.8079
+▁diffuse -10.808
+▁vendredi -10.808
+▁vigueur -10.808
+▁obscur -10.808
+▁vaincu -10.808
+▁encourage -10.808
+▁dérivé -10.808
+obstacle -10.808
+▁clavier -10.808
+▁davance -10.808
+▁sexuel -10.808
+▁approuv -10.8081
+atlantique -10.8082
+▁lopération -10.8083
+▁menacé -10.8083
+▁lusage -10.8084
+▁saurai -10.8085
+▁participation -10.8091
+métrique -10.8091
+▁suède -10.8092
+▁grille -10.8099
+▁lavion -10.81
+▁angle -10.8103
+autant -10.8107
+▁will -10.8117
+▁régulier -10.8117
+ouest -10.8186
+▁employ -10.8203
+▁mouill -10.8205
+allemand -10.824
+▁dadministration -10.8272
+▁ferroviaire -10.8272
+▁francophone -10.8272
+▁lexpérience -10.8272
+▁louvrage -10.8272
+▁moimême -10.8272
+▁cratère -10.8272
+▁pupponi -10.8272
+▁tressaill -10.8272
+▁annoncé -10.8272
+▁compliqué -10.8272
+▁trône -10.8272
+▁pavillon -10.8272
+▁mèche -10.8272
+▁couramment -10.8272
+▁panneaux -10.8272
+▁souligne -10.8272
+▁caroline -10.8272
+▁canapé -10.8272
+▁préparer -10.8272
+▁bellerose -10.8272
+▁demoiselle -10.8272
+▁étoile -10.8272
+▁hâte -10.8272
+▁multipli -10.8272
+▁pôle -10.8272
+▁habituelle -10.8272
+▁suppli -10.8272
+▁opinion -10.8272
+▁troyes -10.8272
+▁cerveau -10.8272
+▁quimper -10.8274
+▁genève -10.8275
+▁paradis -10.8275
+▁formule -10.8276
+▁décède -10.8276
+▁boston -10.8278
+▁distribution -10.8282
+▁remettre -10.8283
+▁abbé -10.8286
+▁définition -10.8294
+▁expression -10.8302
+town -10.8305
+▁turc -10.8306
+▁retrouv -10.8335
+patri -10.8366
+▁ouvrage -10.8379
+acheter -10.842
+▁voudra -10.846
+▁lunivers -10.8462
+▁agriculteur -10.8468
+▁banlieue -10.8468
+▁chocolat -10.8468
+▁divorce -10.8468
+▁dorléans -10.8468
+▁lappareil -10.8468
+▁linguistique -10.8468
+▁maximum -10.8468
+▁nucléaire -10.8468
+▁nécessite -10.8468
+▁pharmacie -10.8468
+▁syndicat -10.8468
+▁vestiges -10.8468
+▁éponyme -10.8468
+▁bombard -10.8468
+▁hospital -10.8468
+▁vengeance -10.8468
+▁médicale -10.8468
+▁viendront -10.8468
+▁inscription -10.8468
+thlète -10.8468
+▁capturé -10.8468
+▁contexte -10.8468
+▁lélève -10.8468
+▁masculin -10.8468
+▁oppose -10.8468
+▁existant -10.8468
+▁maritime -10.8468
+▁gueule -10.8468
+▁prononce -10.8469
+métrie -10.8469
+▁natale -10.847
+▁terrestre -10.847
+▁cellesci -10.8471
+▁otto -10.8476
+▁tunisie -10.8477
+▁mattend -10.8478
+▁venise -10.8479
+▁distinction -10.8485
+▁semblable -10.849
+▁pénal -10.8491
+▁savant -10.8497
+▁pensé -10.8653
+thétique -10.8658
+▁gouvern -10.8659
+utorise -10.866
+▁bizarre -10.8668
+▁enjolras -10.8668
+▁interprété -10.8668
+▁laustralie -10.8668
+▁madeleine -10.8668
+▁médiéval -10.8668
+▁nécessité -10.8668
+▁repoussé -10.8668
+▁saintlouis -10.8668
+▁bourgogne -10.8668
+▁définitivement -10.8668
+▁transféré -10.8668
+▁bertrand -10.8668
+▁linformation -10.8668
+▁réduire -10.8668
+▁cadavre -10.8668
+▁dramatique -10.8668
+surmarne -10.8668
+▁puisquil -10.8668
+▁étiez -10.8668
+▁college -10.8668
+▁cadio -10.8668
+▁appuy -10.867
+▁impose -10.8671
+▁dégrad -10.8674
+▁spécialité -10.8675
+▁théologie -10.8677
+▁dolez -10.8678
+▁individuelle -10.8694
+▁gold -10.8696
+rgio -10.8712
+▁thor -10.8728
+▁averti -10.8736
+dauphin -10.8748
+▁err -10.8763
+▁associ -10.8819
+▁vaisseau -10.8861
+occidental -10.8871
+▁cavrois -10.8872
+▁extraordinaire -10.8872
+▁février -10.8872
+▁intéressé -10.8872
+▁isabelle -10.8872
+▁langleterre -10.8872
+▁lextérieur -10.8872
+▁positive -10.8872
+▁répertoire -10.8872
+▁tempête -10.8872
+▁édifice -10.8872
+▁lintérêt -10.8872
+▁lénergie -10.8872
+▁statistique -10.8872
+▁étouff -10.8872
+▁texas -10.8872
+▁dépôt -10.8872
+▁dhiver -10.8872
+▁cavalier -10.8872
+▁actrice -10.8873
+▁converti -10.8874
+▁jacob -10.8875
+▁lexpression -10.8876
+▁maintenir -10.888
+▁fonctionnaire -10.8881
+▁dentrée -10.8881
+▁poudre -10.8881
+▁déesse -10.8883
+▁écart -10.8888
+▁virg -10.8913
+▁manteau -10.8929
+▁mentionné -10.8942
+▁regroup -10.897
+willer -10.8976
+_ -10.908
+xpérimentation -10.908
+▁architectural -10.908
+▁concrète -10.908
+▁disparition -10.908
+▁musulman -10.908
+▁reconstruit -10.908
+▁épousé -10.908
+▁grenoble -10.908
+▁horrible -10.908
+▁lorchestre -10.908
+▁louverture -10.908
+▁mercredi -10.908
+▁perpignan -10.908
+▁gosselin -10.908
+▁priorité -10.908
+▁résistant -10.908
+▁dollars -10.908
+▁suicide -10.908
+▁soppose -10.908
+▁saintandré -10.908
+▁spectateur -10.908
+▁sélève -10.908
+▁vaccin -10.908
+▁éloigné -10.908
+▁impôt -10.908
+▁chapitre -10.908
+▁révolte -10.908
+▁apparent -10.908
+▁offert -10.908
+▁examen -10.908
+▁loffice -10.908
+▁inédit -10.908
+▁nancy -10.9081
+▁critère -10.9081
+▁tunnel -10.9082
+▁borgo -10.9083
+▁racine -10.9084
+ncêtre -10.9088
+▁emil -10.9101
+▁provenan -10.9112
+ppellation -10.9116
+ignait -10.912
+▁jumel -10.9131
+▁connaissez -10.9138
+paraissait -10.9222
+▁séchapp -10.9243
+bject -10.9253
+château -10.9256
+achève -10.9286
+rchevêque -10.9293
+▁adolescent -10.9293
+▁danemark -10.9293
+▁dinfanterie -10.9293
+▁possession -10.9293
+▁seffectue -10.9293
+▁stratégique -10.9293
+ā -10.9293
+▁centralisateur -10.9293
+▁rugby -10.9293
+▁dactivité -10.9293
+▁nachète -10.9293
+obligation -10.9293
+▁répète -10.9293
+▁suppose -10.9293
+▁quelconque -10.9293
+▁illustra -10.9293
+▁rejoign -10.9293
+▁brève -10.9293
+▁dijon -10.9293
+▁avignon -10.9293
+▁gravure -10.9293
+▁monnaie -10.9293
+▁fichier -10.9294
+▁télévisé -10.9294
+▁lépaule -10.9294
+▁parking -10.9294
+▁trophée -10.9294
+▁travaillait -10.9294
+▁retomb -10.9296
+▁lattention -10.9297
+▁dartistes -10.93
+▁institu -10.9324
+▁parlant -10.9325
+▁cadeau -10.9355
+quiert -10.9389
+place -10.9398
+joly -10.9413
+▁enjeu -10.9428
+structur -10.9495
+xplique -10.9505
+▁lopéra -10.9507
+ú -10.951
+▁comparaison -10.951
+▁incroyable -10.951
+▁lapplication -10.951
+▁attitude -10.951
+▁multiples -10.951
+▁substitu -10.951
+▁petitfils -10.951
+▁conduire -10.951
+▁montauban -10.951
+▁turquie -10.951
+▁jésus -10.951
+▁cinieri -10.951
+▁débarras -10.951
+▁transfert -10.951
+▁permanent -10.951
+▁fragment -10.9511
+▁directrice -10.9511
+▁affronte -10.9511
+évaluation -10.9511
+▁enseigné -10.9511
+▁rédigé -10.9511
+ttraction -10.9511
+▁amitié -10.9511
+▁néglige -10.9511
+▁vapeur -10.9512
+▁martial -10.9512
+▁contempl -10.9513
+▁lambert -10.9514
+▁rétro -10.952
+▁dancien -10.9522
+▁savère -10.9523
+sphère -10.9532
+▁douvr -10.9556
+▁drap -10.9579
+▁correct -10.96
+▁entretien -10.9602
+feld -10.9612
+▁accueil -10.9683
+▁caractérise -10.9733
+▁disciple -10.9733
+▁dynamique -10.9733
+▁identifié -10.9733
+▁juridiction -10.9733
+▁làdessus -10.9733
+▁nautilus -10.9733
+▁portugais -10.9733
+▁progrès -10.9733
+▁provincial -10.9733
+▁recommanda -10.9733
+▁réservoir -10.9733
+▁senfuit -10.9733
+▁archives -10.9733
+▁catalogue -10.9733
+▁fromage -10.9733
+▁succession -10.9733
+▁sophie -10.9733
+▁confortable -10.9733
+▁souviens -10.9733
+▁implanté -10.9733
+▁papillon -10.9733
+▁gilbert -10.9733
+▁serpent -10.9733
+▁décédé -10.9733
+▁précieux -10.9733
+▁reviens -10.9733
+▁aprèsmidi -10.9733
+▁verdun -10.9733
+perçoi -10.9733
+▁recrute -10.9733
+▁sinspire -10.9733
+▁lavenue -10.9733
+▁johann -10.9734
+▁sauvé -10.9734
+▁martyr -10.9734
+▁lydie -10.9736
+▁latelier -10.9737
+▁avanç -10.9737
+▁absent -10.9738
+▁cosette -10.9738
+arctique -10.9783
+doption -10.9798
+▁supprim -10.9901
+frédéric -10.9922
+ū -10.996
+▁anniversaire -10.996
+▁clermontferrand -10.996
+▁dartagnan -10.996
+▁différenci -10.996
+▁désespéré -10.996
+▁internationaux -10.996
+▁librairie -10.996
+▁magnifique -10.996
+▁nouvellezélande -10.996
+▁systématique -10.996
+▁barricade -10.996
+▁confusion -10.996
+▁dintérêt -10.996
+▁emmanuel -10.996
+▁hetzel -10.996
+▁limportance -10.996
+▁décennie -10.996
+▁pomereux -10.996
+▁épuisé -10.996
+▁jahoua -10.996
+▁étonnant -10.996
+▁pionnier -10.996
+▁réduction -10.996
+▁chronique -10.996
+▁solitaire -10.996
+▁réparation -10.996
+▁brown -10.996
+▁ceinture -10.996
+▁sempare -10.996
+▁idéal -10.9961
+▁niort -10.9961
+▁fumée -10.9961
+▁servait -10.9962
+▁plou -10.9965
+▁propag -10.9973
+▁renseig -10.9976
+▁sabot -10.9982
+▁recev -10.999
+paraissent -11.0011
+▁design -11.0021
+▁chevr -11.0037
+éviter -11.0089
+▁hurl -11.011
+agnac -11.0129
+grange -11.0169
+▁île -11.0177
+égyptien -11.0192
+▁dangleterre -11.0192
+▁diffusion -11.0192
+▁dynastie -11.0192
+▁intelligent -11.0192
+▁jérôme -11.0192
+▁madrid -11.0192
+▁refuge -11.0192
+▁junior -11.0193
+▁survécu -11.0193
+▁sexprime -11.0193
+▁militant -11.0193
+▁lélection -11.0193
+▁commandé -11.0193
+▁carfor -11.0193
+▁calcaire -11.0193
+▁frissonn -11.0193
+▁naples -11.0193
+▁filiale -11.0193
+▁diamant -11.0193
+▁mixte -11.0193
+▁rumeur -11.0193
+▁pâte -11.0194
+▁office -11.0195
+▁inquiét -11.0196
+▁maçon -11.0197
+▁fief -11.0205
+yvonne -11.0217
+▁contract -11.0291
+▁désigna -11.0312
+▁lactuel -11.0333
+▁détien -11.0341
+▁représent -11.0375
+dopter -11.0377
+major -11.0394
+▁quatrevingt -11.0423
+▁aperçu -11.0431
+▁automobile -11.0431
+▁complexité -11.0431
+▁concurrence -11.0431
+▁despèce -11.0431
+▁immeuble -11.0431
+▁lemplacement -11.0431
+▁lexemple -11.0431
+▁lextrême -11.0431
+▁loccupation -11.0431
+▁reproduction -11.0431
+▁solidarité -11.0431
+▁tchèque -11.0431
+▁washington -11.0431
+▁élégant -11.0431
+▁ressort -11.0431
+▁réveiller -11.0431
+▁absorb -11.0431
+▁school -11.0431
+▁plaindre -11.0431
+▁présidé -11.0431
+▁chagrin -11.0431
+▁instance -11.0431
+▁illustre -11.0431
+▁patience -11.0431
+▁veuille -11.0431
+▁dialecte -11.0431
+▁suggère -11.0431
+▁autonome -11.0431
+▁lédition -11.0431
+▁chaumont -11.0431
+▁savary -11.0431
+▁suscit -11.0432
+▁travaillant -11.0432
+▁leffort -11.0433
+▁dépasse -11.0436
+▁carross -11.044
+crypt -11.044
+▁suivez -11.0441
+▁suffisant -11.0442
+▁heurt -11.0443
+▁rigol -11.0471
+▁apais -11.0479
+▁olymp -11.0655
+ø -11.0674
+▁adolphe -11.0674
+▁chirurgie -11.0674
+▁compromis -11.0674
+▁geneviève -11.0674
+▁grèce -11.0674
+▁larchipel -11.0674
+▁lunanimité -11.0674
+▁méridional -11.0674
+▁métallique -11.0674
+▁particularité -11.0674
+▁temporaire -11.0674
+▁collectif -11.0674
+▁linspecteur -11.0674
+▁organisme -11.0674
+▁prononcé -11.0674
+▁provisoire -11.0674
+▁bulletin -11.0674
+▁élabor -11.0674
+▁dafrique -11.0674
+▁lalinéa -11.0674
+▁quelquefois -11.0674
+▁bouvard -11.0675
+▁dimportant -11.0675
+▁débarque -11.0675
+▁jaloux -11.0675
+▁augmenté -11.0675
+▁camille -11.0675
+▁partiellement -11.0675
+▁habitude -11.0676
+▁accusé -11.0677
+▁décriv -11.0678
+▁chauve -11.0679
+apporte -11.0692
+▁granit -11.0696
+accuse -11.0719
+▁transmis -11.0863
+griffe -11.0887
+aije -11.0899
+▁express -11.09
+▁aveugle -11.0924
+▁envelopp -11.0924
+▁hollande -11.0924
+▁incapable -11.0924
+▁jalousie -11.0924
+▁joséphine -11.0924
+▁latmosphère -11.0924
+▁lenvironnement -11.0924
+▁louwagie -11.0924
+▁lévénement -11.0924
+▁molécule -11.0924
+▁norvège -11.0924
+▁problématique -11.0924
+▁théorème -11.0924
+▁écossais -11.0924
+▁delaunay -11.0924
+▁dieudonné -11.0924
+▁lendroit -11.0924
+▁massonneau -11.0924
+▁survivant -11.0924
+▁volontaire -11.0924
+▁sénégal -11.0924
+▁daccès -11.0924
+▁damérique -11.0924
+▁massacre -11.0924
+▁produire -11.0924
+▁rencontra -11.0924
+▁absurde -11.0924
+▁délicate -11.0924
+▁hongrie -11.0924
+▁graphique -11.0925
+▁interrogati -11.0925
+▁daoût -11.0925
+▁attaqué -11.0925
+▁raisonnable -11.0925
+▁dœuvre -11.0925
+▁alliance -11.0925
+▁bailli -11.0925
+▁concentre -11.0926
+▁mépris -11.0926
+▁lafrique -11.0926
+▁dachat -11.0926
+▁description -11.0926
+▁porthos -11.0926
+▁lécoute -11.0927
+▁destination -11.0927
+▁adrien -11.0929
+▁schéma -11.0931
+▁arrach -11.0949
+▁amené -11.0958
+transmission -11.099
+▁entame -11.1062
+▁insist -11.1128
+▁amérindien -11.1181
+▁culminant -11.1181
+▁démocratique -11.1181
+▁genevard -11.1181
+▁licencié -11.1181
+▁légitime -11.1181
+▁pécuchet -11.1181
+▁stratégie -11.1181
+▁tokyo -11.1181
+▁équivalent -11.1181
+▁apprécie -11.1181
+▁floraison -11.1181
+▁saddier -11.1181
+▁illustré -11.1181
+▁vercamer -11.1181
+▁géorgie -11.1181
+▁évident -11.1181
+▁rajouter -11.1181
+▁botanique -11.1181
+▁cabaret -11.1181
+▁combattants -11.1181
+▁noublie -11.1181
+▁proportion -11.1181
+thèque -11.1181
+▁ordonné -11.1181
+▁crains -11.1181
+▁minuit -11.1181
+▁soyons -11.1181
+oeuvre -11.1181
+▁embarqu -11.1181
+ugmentation -11.1181
+▁lanimal -11.1181
+▁létang -11.1182
+▁versailles -11.1182
+▁désolé -11.1183
+▁camion -11.1184
+▁wolf -11.119
+accéder -11.1193
+▁compare -11.122
+▁gamb -11.1227
+▁trott -11.1355
+▁contrôl -11.1402
+▁industrie -11.1442
+▁affilié -11.1444
+▁byzantin -11.1444
+▁forestière -11.1444
+▁geoffroy -11.1444
+▁indépendance -11.1444
+▁lantiquité -11.1444
+▁lenquête -11.1444
+▁linformatique -11.1444
+▁madagascar -11.1444
+▁paroissiale -11.1444
+▁quarantaine -11.1444
+▁remplaça -11.1444
+▁rezdechaussée -11.1444
+▁satellite -11.1444
+▁transparence -11.1444
+▁vestibule -11.1444
+▁cameroun -11.1444
+▁perpétue -11.1444
+▁redoutable -11.1444
+▁cohérent -11.1444
+▁encadré -11.1444
+▁indonésie -11.1444
+▁nièce -11.1444
+▁vaincre -11.1444
+▁fiancé -11.1444
+▁maltenu -11.1444
+▁révélé -11.1444
+▁altitude -11.1444
+▁joël -11.1444
+▁abattu -11.1444
+▁andrew -11.1444
+▁lanterne -11.1444
+▁érigé -11.1445
+▁sprint -11.1445
+▁parbleu -11.1445
+▁boulanger -11.1445
+▁transforme -11.1446
+▁coffre -11.1446
+▁atteign -11.1447
+▁joint -11.1448
+▁convoi -11.1461
+▁project -11.1462
+limard -11.1465
+▁agrandi -11.1473
+▁ouvra -11.1476
+▁détest -11.1489
+▁capital -11.1572
+joseph -11.1677
+philippe -11.1677
+sursaône -11.1714
+▁alternative -11.1714
+▁chômage -11.1714
+▁coronado -11.1714
+▁crâne -11.1714
+▁demiheure -11.1714
+▁dignelesbains -11.1714
+▁distingué -11.1714
+▁ecclésiastique -11.1714
+▁korigane -11.1714
+▁lindépendance -11.1714
+▁magistrat -11.1714
+▁philosophe -11.1714
+▁pseudonyme -11.1714
+▁saintnicolas -11.1714
+▁facilite -11.1714
+▁impliqué -11.1714
+▁paupières -11.1714
+▁perspective -11.1714
+▁raphaël -11.1714
+▁vêtu -11.1714
+▁remiremont -11.1714
+▁réveillé -11.1714
+scenseur -11.1714
+▁clarifie -11.1714
+▁littoral -11.1714
+▁aéroport -11.1714
+▁manœuvre -11.1714
+▁complété -11.1714
+▁aménage -11.1714
+▁cayenne -11.1714
+▁traîna -11.1715
+▁barbara -11.1715
+▁terroris -11.1715
+▁copain -11.1715
+▁arrondissement -11.1715
+▁mobile -11.1715
+▁coiffe -11.1715
+▁relever -11.1715
+▁doyen -11.1715
+▁allumé -11.1716
+▁scott -11.1716
+échelle -11.1717
+▁rêv -11.1745
+vantage -11.188
+bandonner -11.1925
+▁probable -11.1967
+éléphant -11.199
+▁angoulême -11.1992
+▁commerçant -11.1992
+▁controversé -11.1992
+▁convaincre -11.1992
+▁endommagé -11.1992
+▁interprétation -11.1992
+▁norvégien -11.1992
+▁prétexte -11.1992
+▁recommande -11.1992
+▁récompensé -11.1992
+▁significative -11.1992
+▁successivement -11.1992
+▁symbolis -11.1992
+▁weekend -11.1992
+▁électoral -11.1992
+▁bénédict -11.1992
+▁denseignement -11.1992
+▁démocrate -11.1992
+▁ferdinand -11.1992
+▁lartillerie -11.1992
+▁mystère -11.1992
+▁pardessus -11.1992
+▁rattrap -11.1992
+▁procéder -11.1992
+▁honoraire -11.1992
+▁roumanie -11.1992
+▁piscine -11.1992
+▁rabault -11.1992
+▁imprimé -11.1992
+▁poitiers -11.1992
+▁brigade -11.1992
+▁émotion -11.1992
+▁demploi -11.1992
+▁grève -11.1992
+phérique -11.1992
+▁pompier -11.1993
+▁trompé -11.1993
+▁tabac -11.1993
+quérir -11.1993
+▁mâch -11.1994
+▁halte -11.1994
+rèze -11.1994
+▁recense -11.1995
+▁valentin -11.1995
+▁plonge -11.1996
+▁cumul -11.1996
+▁parvenu -11.1997
+boeuf -11.2002
+▁diffère -11.2009
+▁censé -11.2011
+stadt -11.202
+utilise -11.2034
+▁alexandr -11.2139
+▁tissu -11.219
+saintpétersbourg -11.2278
+▁apparemment -11.2278
+▁besançon -11.2278
+▁collective -11.2278
+▁constructeur -11.2278
+▁construis -11.2278
+▁controverse -11.2278
+▁cravate -11.2278
+▁curiosité -11.2278
+▁dimpôt -11.2278
+▁dissimul -11.2278
+▁effectif -11.2278
+▁lalphabet -11.2278
+▁lefebvre -11.2278
+▁mammifère -11.2278
+▁misérable -11.2278
+▁présidence -11.2278
+▁pèlerinage -11.2278
+▁sylvain -11.2278
+▁énormément -11.2278
+▁balbutia -11.2278
+▁confronté -11.2278
+▁conjoint -11.2278
+▁prairies -11.2278
+▁syndical -11.2278
+▁entraîné -11.2278
+▁fabriqué -11.2278
+▁fraîche -11.2278
+▁transformation -11.2278
+▁brigitte -11.2278
+▁compensation -11.2278
+▁kirsch -11.2278
+▁cuivre -11.2278
+▁océan -11.2278
+▁podium -11.2278
+▁séminaire -11.2278
+▁cartouche -11.2278
+▁répéter -11.2278
+▁incident -11.2278
+▁sultan -11.2278
+▁implique -11.2278
+▁relatif -11.2278
+▁incarne -11.2278
+▁oblige -11.2278
+▁métonne -11.2278
+▁connect -11.2278
+instinct -11.2279
+▁portail -11.2279
+▁mitigé -11.2279
+▁ultra -11.2279
+▁confident -11.2279
+▁hésita -11.2288
+▁électro -11.2292
+▁vibr -11.2331
+oriental -11.2334
+▁jaurai -11.2377
+▁décroch -11.2443
+▁plaign -11.2474
+▁découvr -11.2559
+saintsauveur -11.2572
+▁adaptation -11.2572
+▁concurrent -11.2572
+▁darchitecture -11.2572
+▁dinquiétude -11.2572
+▁flaviana -11.2572
+▁insectes -11.2572
+▁jésuite -11.2572
+▁lexpédition -11.2572
+▁lhistorien -11.2572
+▁mathieu -11.2572
+▁montaudoin -11.2572
+▁pleuvoir -11.2572
+▁synthèse -11.2572
+▁sépulture -11.2572
+▁jurassien -11.2572
+▁perfection -11.2572
+▁signature -11.2572
+▁antonio -11.2572
+▁générique -11.2572
+▁patrouille -11.2572
+▁revêt -11.2572
+▁arménien -11.2572
+▁cardiaque -11.2572
+▁laccident -11.2572
+▁réception -11.2572
+▁séparation -11.2572
+▁corriger -11.2572
+▁proclam -11.2572
+nouvelleguinée -11.2572
+▁toronto -11.2572
+▁essuy -11.2573
+▁casquette -11.2573
+▁massive -11.2573
+▁orchestr -11.2573
+▁cigare -11.2575
+▁restauré -11.2575
+▁gregor -11.2605
+▁sylvi -11.2605
+▁georg -11.2609
+▁prun -11.2611
+▁marteau -11.2636
+chapelle -11.2838
+▁étudia -11.2842
+▁cambr -11.2845
+▁arrondi -11.2874
+▁barcelone -11.2875
+▁biodiversité -11.2875
+▁congrégation -11.2875
+▁cylindr -11.2875
+▁cyprien -11.2875
+▁dolorès -11.2875
+▁déconomie -11.2875
+▁facilité -11.2875
+▁lépisode -11.2875
+▁protocole -11.2875
+▁précédemment -11.2875
+▁pédagogique -11.2875
+▁recommandé -11.2875
+▁reconstruction -11.2875
+▁reçoivent -11.2875
+▁récurrent -11.2875
+▁symbolique -11.2875
+▁dartillerie -11.2875
+▁foudroy -11.2875
+▁israël -11.2875
+giovanni -11.2875
+xpansion -11.2875
+▁garnison -11.2875
+▁hélène -11.2875
+▁conducteur -11.2875
+▁recherché -11.2875
+▁suzanne -11.2875
+▁léditeur -11.2875
+▁dépassé -11.2875
+▁précoce -11.2875
+▁tranché -11.2875
+▁installe -11.2875
+▁virtuel -11.2875
+▁couloir -11.2875
+▁benoit -11.2875
+▁linfini -11.2875
+▁comptoir -11.2875
+▁matthi -11.2875
+▁barbier -11.2875
+▁oraison -11.2875
+▁furieuse -11.2875
+▁inégal -11.2875
+▁héberge -11.2876
+▁nymph -11.2876
+▁assume -11.2877
+▁habitué -11.288
+▁persist -11.2982
+▁métai -11.3024
+organiser -11.312
+cœur -11.3151
+volution -11.3165
+▁anonyme -11.3187
+▁attentif -11.3187
+▁bathilde -11.3187
+▁charpentier -11.3187
+▁concentration -11.3187
+▁consécutive -11.3187
+▁douloureuse -11.3187
+▁immobile -11.3187
+▁insupportable -11.3187
+▁laprèsmidi -11.3187
+▁lobservatoire -11.3187
+▁portefeuille -11.3187
+▁programmé -11.3187
+▁rectangulaire -11.3187
+▁sacrifice -11.3187
+▁souffert -11.3187
+▁vérifier -11.3187
+▁éventuelle -11.3187
+▁bonhomme -11.3187
+▁désordre -11.3187
+▁prononça -11.3187
+▁critiqué -11.3187
+▁grandchose -11.3187
+▁porteparole -11.3187
+▁sétabli -11.3187
+▁violemment -11.3187
+ntrepreneur -11.3187
+▁métropole -11.3187
+▁brièvement -11.3187
+▁cahors -11.3187
+▁légumes -11.3187
+▁bilingue -11.3187
+▁cinématographi -11.3188
+▁savança -11.3188
+▁étiezvous -11.3188
+▁mennuie -11.3188
+▁nettoy -11.3188
+▁consolid -11.3188
+▁prouvé -11.3188
+▁chandelle -11.3188
+▁préfet -11.3188
+▁mutualis -11.3189
+▁prolongé -11.3189
+▁habile -11.3189
+▁consenti -11.319
+▁éternel -11.3198
+▁suivai -11.3208
+christophe -11.3472
+affluent -11.3473
+œuvr -11.351
+▁appartienne -11.351
+▁architecture -11.351
+▁communautaire -11.351
+▁conquête -11.351
+▁détruire -11.351
+▁euxmêmes -11.351
+▁fatigant -11.351
+▁hiérarchie -11.351
+▁infrastructure -11.351
+▁lenregistrement -11.351
+▁lexécution -11.351
+▁lintelligence -11.351
+▁marvejol -11.351
+▁procureur -11.351
+▁péninsule -11.351
+▁renseignements -11.351
+▁résoudre -11.351
+▁révérend -11.351
+▁sapprocha -11.351
+▁élargi -11.351
+▁mobilier -11.351
+▁rédacteur -11.351
+▁lélevage -11.351
+▁mosquée -11.351
+▁inspecteur -11.351
+▁goriot -11.351
+▁manosque -11.351
+▁limpact -11.351
+▁pianiste -11.351
+▁tonnerre -11.351
+▁réagir -11.351
+▁senfuir -11.351
+▁complice -11.351
+▁trajet -11.351
+▁voyezvous -11.351
+▁avanthier -11.351
+▁general -11.351
+▁étonné -11.3511
+▁cognac -11.3512
+▁dépouill -11.3512
+▁reprena -11.3514
+▁criminel -11.3519
+▁repond -11.3519
+▁emmenez -11.3521
+▁jumeau -11.3581
+ncontre -11.3588
+▁nallai -11.3653
+observer -11.3682
+▁répand -11.3762
+▁saintjulien -11.3843
+ı -11.3843
+▁calendrier -11.3843
+▁catastrophe -11.3843
+▁diagnostic -11.3843
+▁exercice -11.3843
+▁fauchelevent -11.3843
+▁fréquence -11.3843
+▁handicap -11.3843
+▁investissements -11.3843
+▁jusquici -11.3843
+▁lexploitation -11.3843
+▁linitiative -11.3843
+▁québécois -11.3843
+▁recommenc -11.3843
+▁singulier -11.3843
+▁souffrance -11.3843
+▁surveillance -11.3843
+▁théodore -11.3843
+ã -11.3843
+▁clandestin -11.3843
+▁davril -11.3843
+▁immigr -11.3843
+▁orphelin -11.3843
+▁pencroff -11.3843
+▁dautriche -11.3843
+▁dentreprise -11.3843
+▁lombard -11.3843
+▁classification -11.3843
+▁emporté -11.3843
+▁récepteur -11.3843
+▁écologique -11.3843
+▁lituanie -11.3843
+▁létoile -11.3843
+▁nerveuse -11.3843
+▁seniors -11.3843
+▁buisson -11.3843
+▁succéda -11.3843
+stronome -11.3843
+▁saintjust -11.3843
+▁basilique -11.3843
+▁entretenu -11.3844
+▁parasite -11.3844
+▁génétique -11.3844
+▁tablier -11.3844
+▁tragique -11.3844
+▁bavard -11.3844
+▁surprend -11.3844
+▁verticale -11.3844
+percevoir -11.3844
+▁chimique -11.3844
+évidence -11.3844
+▁attente -11.3844
+▁facial -11.3844
+▁malaisie -11.3845
+▁ralenti -11.3845
+▁paternel -11.3856
+dministrateur -11.4143
+préfecture -11.415
+▁absolu -11.4163
+å -11.4188
+▁compatriote -11.4188
+▁confession -11.4188
+▁doctrine -11.4188
+▁galaxie -11.4188
+▁gwynplaine -11.4188
+▁hôpital -11.4188
+▁inacceptable -11.4188
+▁lappartement -11.4188
+▁législative -11.4188
+▁petitdéjeuner -11.4188
+▁popularité -11.4188
+▁reconverti -11.4188
+▁renforcé -11.4188
+▁saintdizier -11.4188
+▁sarkozy -11.4188
+▁simultanément -11.4188
+▁thaïlande -11.4188
+identifier -11.4188
+▁improvis -11.4188
+▁innovant -11.4188
+▁lalcool -11.4188
+▁orléans -11.4188
+▁substance -11.4188
+▁inaugur -11.4188
+▁linventaire -11.4188
+▁flèche -11.4188
+▁limpôt -11.4188
+ă -11.4545
+ń -11.6117
+æ -11.6117
+ß -11.6552
+ş -11.7982
+ș -11.8509
+ż -12.0278
+ÿ -12.0944
+ž -12.0944
+ý -12.2428
+ð -12.6281
+ř -12.6281
+ą -12.6281
+ț -12.6281
+ğ -12.896
+õ -12.896
+đ -12.896
+ő -12.896
+ò -13.0627
+ē -13.0627
+ī -13.0627
+ě -13.0627
+ņ -13.0627
+ù -13.184
+ì -13.2627
+ʻ -13.2627
+ṣ -13.2627
+ė -13.2627
+1 -13.846
+2 -13.846
+ň -13.846
+ź -13.846
+ʾ -13.846
+ҫ -13.846
+⁄ -13.846
+ʉ -14.346
+π -14.346
+ṇ -14.346
+ả -14.346
+星 -14.346
+ṯ -15.3448
+τ -15.3449
+κ -15.345
+δ -15.3451
+þ -15.3452
+R -15.3453
+σ -15.3454
+β -15.3455
+α -15.3456
+ŏ -15.3457
+ʿ -15.3458
+ś -15.3459
+ġ -15.346
+ơ -15.346
+ư -15.346
+ṭ -15.346
+E -15.346
+N -15.346
+Z -15.346
+ħ -15.346
+ļ -15.346
+ť -15.346
+ų -15.346
+ǃ -15.346
+ǫ -15.346
+ǹ -15.346
+ʼ -15.346
+ː -15.346
+ζ -15.346
+ν -15.346
+υ -15.346
+ω -15.346
+а -15.346
+г -15.346
+е -15.346
+з -15.346
+и -15.346
+м -15.346
+н -15.346
+ደ -15.346
+ḍ -15.346
+ạ -15.346
+ầ -15.346
+ậ -15.346
+ễ -15.346
+ệ -15.346
+ị -15.346
+ồ -15.346
+ộ -15.346
+ợ -15.346
+ử -15.346
+ỳ -15.346
+い -15.346
+た -15.346
+つ -15.346
+乃 -15.346
+扬 -15.346
+术 -15.346
+杜 -15.346
+美 -15.346
+馆 -15.346
+ṅ -15.346
+ę -15.346
+京 -15.346
+北 -15.346
diff --git a/configs/fr-en/tgt_unigram6000/spm_unigram_fr.model b/configs/fr-en/tgt_unigram6000/spm_unigram_fr.model
new file mode 100644
index 0000000000000000000000000000000000000000..b84841f0fe401483b263a47d76797c4360c75ec3
--- /dev/null
+++ b/configs/fr-en/tgt_unigram6000/spm_unigram_fr.model
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9722dc6f2890302585a268c272172237d8361636d6cfa2074c2135a35a23dba8
+size 342752
diff --git a/configs/fr-en/tgt_unigram6000/spm_unigram_fr.txt b/configs/fr-en/tgt_unigram6000/spm_unigram_fr.txt
new file mode 100644
index 0000000000000000000000000000000000000000..8e714b2e505c8d6b7bab09c5c7b9aa17bd6865a7
--- /dev/null
+++ b/configs/fr-en/tgt_unigram6000/spm_unigram_fr.txt
@@ -0,0 +1,5996 @@
+▁the 1
+s 1
+▁of 1
+▁a 1
+▁to 1
+▁is 1
+▁in 1
+▁and 1
+▁he 1
+' 1
+▁it 1
+▁i 1
+t 1
+▁was 1
+ed 1
+ing 1
+▁this 1
+▁you 1
+▁are 1
+n 1
+▁that 1
+▁one 1
+▁his 1
+▁for 1
+▁hundred 1
+▁on 1
+d 1
+e 1
+▁be 1
+▁with 1
+▁at 1
+▁we 1
+▁two 1
+▁street 1
+▁by 1
+▁she 1
+▁not 1
+y 1
+▁de 1
+▁will 1
+m 1
+ly 1
+▁from 1
+▁have 1
+▁an 1
+▁three 1
+▁they 1
+▁as 1
+▁has 1
+re 1
+er 1
+al 1
+▁number 1
+▁me 1
+▁also 1
+▁there 1
+▁him 1
+▁her 1
+▁were 1
+▁amendment 1
+▁four 1
+▁my 1
+▁thousand 1
+▁do 1
+▁what 1
+▁zero 1
+es 1
+▁five 1
+▁but 1
+▁can 1
+c 1
+▁twenty 1
+▁had 1
+▁all 1
+▁la 1
+▁their 1
+o 1
+r 1
+▁your 1
+▁no 1
+▁seven 1
+▁six 1
+g 1
+▁first 1
+▁would 1
+▁up 1
+▁floor 1
+▁saint 1
+a 1
+▁eight 1
+▁time 1
+▁or 1
+▁thirty 1
+▁like 1
+p 1
+▁these 1
+▁so 1
+▁mr 1
+i 1
+in 1
+▁then 1
+ll 1
+▁support 1
+▁who 1
+▁did 1
+▁its 1
+u 1
+▁road 1
+▁about 1
+▁more 1
+▁le 1
+▁du 1
+▁very 1
+▁nine 1
+▁been 1
+▁forty 1
+▁them 1
+▁only 1
+▁go 1
+▁back 1
+▁after 1
+▁am 1
+le 1
+▁fifty 1
+on 1
+▁some 1
+▁other 1
+▁well 1
+▁re 1
+ation 1
+▁now 1
+k 1
+▁out 1
+▁s 1
+▁know 1
+l 1
+en 1
+▁o 1
+ro 1
+el 1
+ch 1
+▁don 1
+f 1
+▁seventy 1
+▁when 1
+▁d 1
+▁new 1
+▁see 1
+▁if 1
+▁sixty 1
+▁dur 1
+ra 1
+▁here 1
+▁part 1
+an 1
+▁ninety 1
+▁ten 1
+or 1
+ion 1
+th 1
+▁e 1
+ic 1
+st 1
+ar 1
+▁man 1
+▁which 1
+▁how 1
+ve 1
+▁come 1
+▁name 1
+▁good 1
+▁does 1
+▁made 1
+▁people 1
+▁located 1
+▁same 1
+z 1
+il 1
+▁work 1
+is 1
+b 1
+ne 1
+é 1
+▁eighty 1
+▁place 1
+ri 1
+ia 1
+▁mister 1
+de 1
+▁however 1
+▁several 1
+▁said 1
+ers 1
+▁our 1
+▁us 1
+▁right 1
+▁way 1
+▁take 1
+it 1
+▁many 1
+▁could 1
+▁still 1
+et 1
+▁where 1
+li 1
+at 1
+▁jean 1
+▁l 1
+w 1
+▁end 1
+ur 1
+▁co 1
+▁going 1
+lo 1
+▁car 1
+▁father 1
+▁over 1
+▁avenue 1
+▁p 1
+se 1
+ta 1
+ian 1
+▁open 1
+▁french 1
+ry 1
+▁into 1
+ity 1
+▁found 1
+▁most 1
+te 1
+ma 1
+▁sur 1
+us 1
+▁year 1
+▁under 1
+▁than 1
+▁give 1
+▁city 1
+na 1
+▁long 1
+▁get 1
+ck 1
+▁ 1
+▁second 1
+v 1
+▁want 1
+ce 1
+▁last 1
+la 1
+▁g 1
+▁must 1
+▁used 1
+ter 1
+▁day 1
+▁ca 1
+ine 1
+▁left 1
+▁let 1
+ment 1
+▁mrs 1
+▁main 1
+▁each 1
+▁years 1
+▁son 1
+▁any 1
+▁group 1
+▁war 1
+▁ro 1
+▁f 1
+▁con 1
+▁through 1
+▁ra 1
+▁paris 1
+ul 1
+h 1
+▁without 1
+▁old 1
+▁school 1
+▁des 1
+▁between 1
+▁un 1
+▁mar 1
+▁little 1
+ou 1
+▁make 1
+▁france 1
+▁order 1
+ist 1
+▁en 1
+ard 1
+▁down 1
+▁speak 1
+co 1
+ol 1
+▁even 1
+▁family 1
+▁opinion 1
+▁later 1
+able 1
+vi 1
+▁love 1
+▁before 1
+▁mo 1
+ant 1
+▁b 1
+ie 1
+▁al 1
+▁t 1
+▁won 1
+▁small 1
+▁look 1
+▁door 1
+ier 1
+came 1
+que 1
+▁world 1
+▁therefore 1
+▁too 1
+▁since 1
+▁just 1
+▁much 1
+▁against 1
+▁put 1
+be 1
+▁think 1
+ac 1
+▁came 1
+ir 1
+▁government 1
+▁yes 1
+- 1
+▁may 1
+▁young 1
+ge 1
+▁great 1
+▁took 1
+ca 1
+▁town 1
+▁head 1
+▁born 1
+▁madam 1
+▁never 1
+▁general 1
+to 1
+▁find 1
+▁again 1
+▁mother 1
+▁being 1
+x 1
+au 1
+di 1
+▁few 1
+▁those 1
+▁r 1
+▁m 1
+ate 1
+▁house 1
+ot 1
+man 1
+ba 1
+▁lot 1
+ni 1
+▁went 1
+me 1
+▁article 1
+▁per 1
+▁species 1
+▁se 1
+▁church 1
+▁team 1
+▁st 1
+▁water 1
+▁such 1
+▁need 1
+▁rue 1
+▁today 1
+▁oh 1
+ent 1
+▁become 1
+▁national 1
+▁why 1
+▁k 1
+▁c 1
+ies 1
+▁himself 1
+▁village 1
+▁sure 1
+ton 1
+▁vi 1
+▁law 1
+ga 1
+▁got 1
+▁life 1
+▁president 1
+▁minister 1
+▁hand 1
+▁because 1
+po 1
+ti 1
+▁while 1
+ive 1
+ff 1
+sh 1
+▁bo 1
+▁say 1
+id 1
+▁li 1
+▁known 1
+▁pa 1
+go 1
+▁ga 1
+▁bar 1
+ting 1
+▁ri 1
+▁di 1
+▁important 1
+▁pi 1
+che 1
+▁away 1
+ous 1
+▁na 1
+ville 1
+▁mi 1
+▁big 1
+▁club 1
+▁show 1
+▁post 1
+▁state 1
+▁should 1
+▁named 1
+ted 1
+ary 1
+▁around 1
+▁sister 1
+age 1
+▁thu 1
+pe 1
+▁les 1
+▁large 1
+do 1
+▁use 1
+▁mont 1
+▁pierre 1
+ron 1
+ci 1
+da 1
+▁called 1
+▁often 1
+▁room 1
+▁south 1
+▁university 1
+um 1
+▁public 1
+▁talk 1
+son 1
+▁side 1
+▁both 1
+ated 1
+bi 1
+tic 1
+ha 1
+ke 1
+▁works 1
+▁brother 1
+un 1
+▁turn 1
+▁goes 1
+▁high 1
+▁ta 1
+va 1
+▁home 1
+▁please 1
+▁men 1
+▁ar 1
+▁ma 1
+▁tell 1
+▁children 1
+▁bi 1
+▁different 1
+▁cha 1
+▁point 1
+▁n 1
+ness 1
+tte 1
+▁always 1
+ism 1
+▁north 1
+▁fifteen 1
+▁case 1
+les 1
+▁every 1
+▁river 1
+▁another 1
+am 1
+bo 1
+▁following 1
+▁really 1
+▁country 1
+▁form 1
+ka 1
+ish 1
+▁twelve 1
+▁everything 1
+▁ah 1
+▁sir 1
+▁already 1
+▁next 1
+▁u 1
+ig 1
+no 1
+▁land 1
+▁eleven 1
+▁jo 1
+▁off 1
+▁gra 1
+▁own 1
+▁pro 1
+▁sub 1
+▁member 1
+gu 1
+▁ex 1
+▁table 1
+cy 1
+▁sp 1
+ure 1
+▁code 1
+▁toward 1
+▁ju 1
+▁particular 1
+▁cor 1
+▁red 1
+ille 1
+▁close 1
+▁district 1
+▁finally 1
+as 1
+▁leave 1
+▁help 1
+▁season 1
+land 1
+▁ha 1
+▁black 1
+▁pre 1
+▁front 1
+▁game 1
+▁lo 1
+▁da 1
+▁center 1
+▁v 1
+▁night 1
+▁ho 1
+om 1
+▁nothing 1
+▁station 1
+▁among 1
+▁far 1
+tion 1
+▁company 1
+▁enough 1
+▁dis 1
+▁bu 1
+▁po 1
+▁paul 1
+▁commission 1
+▁ba 1
+▁building 1
+▁king 1
+▁coming 1
+▁loved 1
+▁career 1
+▁thought 1
+ze 1
+bu 1
+▁live 1
+ling 1
+▁something 1
+▁louis 1
+▁system 1
+▁fa 1
+▁god 1
+mo 1
+gue 1
+▁office 1
+mi 1
+▁light 1
+op 1
+im 1
+▁ka 1
+▁fourteen 1
+▁friend 1
+▁daughter 1
+▁white 1
+let 1
+ru 1
+▁music 1
+▁album 1
+▁book 1
+ver 1
+▁once 1
+▁mu 1
+▁region 1
+▁lu 1
+▁best 1
+▁sixteen 1
+▁comes 1
+▁h 1
+▁plays 1
+▁asked 1
+▁eighteen 1
+▁pe 1
+▁quit 1
+▁present 1
+▁val 1
+▁married 1
+▁dead 1
+▁played 1
+pa 1
+▁things 1
+▁free 1
+▁built 1
+▁takes 1
+▁face 1
+▁real 1
+▁given 1
+▁near 1
+▁area 1
+▁third 1
+▁understand 1
+ang 1
+▁child 1
+ther 1
+▁anything 1
+▁women 1
+ical 1
+▁project 1
+▁quickly 1
+▁question 1
+▁doctor 1
+▁play 1
+▁woman 1
+vo 1
+▁certain 1
+nes 1
+▁english 1
+hi 1
+ance 1
+▁hu 1
+▁ti 1
+▁honor 1
+▁party 1
+ng 1
+▁ve 1
+▁bit 1
+▁tour 1
+▁sa 1
+▁fact 1
+▁ver 1
+▁sometimes 1
+▁thirteen 1
+ut 1
+▁less 1
+▁mer 1
+▁done 1
+▁line 1
+▁title 1
+▁american 1
+eau 1
+▁amendments 1
+▁seem 1
+▁color 1
+▁united 1
+▁call 1
+▁art 1
+▁municipality 1
+▁ne 1
+▁according 1
+lin 1
+▁lives 1
+▁bou 1
+ence 1
+▁yet 1
+▁stay 1
+▁department 1
+▁film 1
+▁history 1
+▁stop 1
+▁alley 1
+con 1
+▁lost 1
+▁death 1
+▁days 1
+▁short 1
+▁started 1
+▁grand 1
+est 1
+▁wife 1
+▁count 1
+▁har 1
+▁set 1
+▁bad 1
+▁miss 1
+▁keep 1
+▁park 1
+▁position 1
+▁council 1
+▁bring 1
+▁island 1
+▁author 1
+▁capital 1
+men 1
+mer 1
+ph 1
+ak 1
+ful 1
+▁é 1
+▁moment 1
+▁ja 1
+▁having 1
+▁believe 1
+▁power 1
+ix 1
+▁va 1
+▁wanted 1
+▁marie 1
+▁francois 1
+▁mor 1
+▁voice 1
+▁indeed 1
+▁inter 1
+▁soon 1
+▁poor 1
+▁period 1
+▁international 1
+▁char 1
+▁german 1
+he 1
+ized 1
+ad 1
+▁gave 1
+▁taking 1
+▁morning 1
+▁nice 1
+▁meet 1
+line 1
+▁local 1
+▁read 1
+▁eyes 1
+ty 1
+▁cla 1
+▁san 1
+▁course 1
+min 1
+▁taken 1
+▁half 1
+sse 1
+▁word 1
+▁better 1
+▁air 1
+▁square 1
+▁seems 1
+▁change 1
+▁charles 1
+▁thing 1
+▁shall 1
+ric 1
+▁hear 1
+ran 1
+lle 1
+ki 1
+▁full 1
+▁pay 1
+▁states 1
+▁sea 1
+tin 1
+▁seen 1
+▁role 1
+▁boulevard 1
+▁studies 1
+▁considered 1
+▁ask 1
+▁ste 1
+lu 1
+▁died 1
+▁happy 1
+za 1
+os 1
+▁level 1
+▁true 1
+▁sha 1
+▁plan 1
+▁girl 1
+▁arm 1
+▁thank 1
+▁withdraw 1
+▁kind 1
+▁master 1
+▁castle 1
+less 1
+▁former 1
+▁port 1
+▁success 1
+▁gu 1
+der 1
+▁ru 1
+▁roman 1
+▁almo 1
+pi 1
+▁possible 1
+▁fire 1
+mont 1
+▁subject 1
+▁created 1
+▁player 1
+▁saw 1
+ny 1
+▁et 1
+ise 1
+ois 1
+▁vote 1
+▁age 1
+sa 1
+▁received 1
+▁seventeen 1
+▁act 1
+▁west 1
+▁series 1
+▁martin 1
+▁idea 1
+▁omer 1
+▁bra 1
+▁heart 1
+▁together 1
+▁ni 1
+ber 1
+ial 1
+▁site 1
+▁beautiful 1
+tra 1
+▁makes 1
+▁au 1
+io 1
+▁times 1
+▁che 1
+▁buy 1
+▁money 1
+elle 1
+▁hard 1
+ring 1
+▁political 1
+ten 1
+▁care 1
+▁until 1
+▁commune 1
+▁ge 1
+▁worked 1
+▁gar 1
+▁start 1
+ster 1
+▁husband 1
+▁date 1
+▁strong 1
+▁job 1
+▁person 1
+▁buried 1
+▁measure 1
+ho 1
+▁z 1
+ja 1
+▁include 1
+▁tu 1
+▁nineteen 1
+▁text 1
+▁wood 1
+▁el 1
+▁rest 1
+the 1
+▁guy 1
+▁army 1
+▁join 1
+ette 1
+ger 1
+▁john 1
+ology 1
+▁problem 1
+▁path 1
+▁matter 1
+▁bri 1
+▁fe 1
+▁em 1
+▁report 1
+▁route 1
+▁getting 1
+▁dear 1
+▁horse 1
+▁doing 1
+▁hope 1
+▁looking 1
+▁example 1
+▁fall 1
+▁chi 1
+▁cold 1
+▁regard 1
+▁ci 1
+▁director 1
+▁fight 1
+▁books 1
+▁social 1
+▁bon 1
+▁imp 1
+▁wi 1
+ile 1
+▁michel 1
+▁tri 1
+▁european 1
+and 1
+▁wait 1
+▁museum 1
+▁studied 1
+▁double 1
+▁attack 1
+mar 1
+dy 1
+▁major 1
+gra 1
+▁visit 1
+▁hi 1
+▁final 1
+▁court 1
+led 1
+▁run 1
+▁original 1
+▁sc 1
+▁gre 1
+ize 1
+▁bre 1
+▁fi 1
+▁single 1
+▁committee 1
+▁type 1
+▁band 1
+▁boy 1
+▁various 1
+▁impasse 1
+▁remains 1
+mon 1
+▁east 1
+▁waiting 1
+j 1
+▁championship 1
+▁based 1
+iere 1
+▁th 1
+▁story 1
+qui 1
+▁origin 1
+▁everyone 1
+▁cap 1
+▁anymore 1
+ina 1
+▁movie 1
+up 1
+▁train 1
+tor 1
+ko 1
+▁vo 1
+▁pathway 1
+use 1
+▁competition 1
+▁style 1
+▁further 1
+▁talking 1
+▁sign 1
+rs 1
+▁able 1
+▁charge 1
+▁language 1
+▁currently 1
+▁tomorrow 1
+▁je 1
+way 1
+lon 1
+▁offer 1
+▁playing 1
+▁hit 1
+lic 1
+▁song 1
+▁written 1
+▁wall 1
+▁return 1
+▁w 1
+▁ki 1
+ner 1
+▁unfavorable 1
+ler 1
+val 1
+▁living 1
+▁brought 1
+▁latter 1
+▁told 1
+▁bank 1
+lé 1
+▁issue 1
+▁middle 1
+▁cannot 1
+▁pont 1
+▁met 1
+▁pan 1
+ens 1
+out 1
+▁ko 1
+em 1
+iv 1
+fa 1
+▁answer 1
+▁yellow 1
+per 1
+▁identical 1
+ization 1
+ey 1
+▁working 1
+gi 1
+▁common 1
+▁captain 1
+▁whole 1
+▁qu 1
+▁words 1
+cha 1
+▁longer 1
+▁cost 1
+▁tra 1
+ien 1
+▁mountain 1
+▁te 1
+ré 1
+our 1
+tu 1
+▁central 1
+▁current 1
+▁sal 1
+rew 1
+▁sta 1
+▁walk 1
+▁cut 1
+▁cause 1
+▁bis 1
+▁hold 1
+▁mean 1
+▁seat 1
+▁royal 1
+▁months 1
+▁county 1
+cu 1
+▁published 1
+▁result 1
+▁military 1
+▁garden 1
+▁along 1
+▁mon 1
+tan 1
+chi 1
+▁fish 1
+▁members 1
+aud 1
+▁fear 1
+ain 1
+▁kept 1
+▁hat 1
+▁human 1
+▁sun 1
+▁rock 1
+▁myself 1
+▁battle 1
+fi 1
+ities 1
+▁movement 1
+▁camp 1
+if 1
+▁held 1
+nce 1
+ay 1
+han 1
+▁die 1
+▁writing 1
+▁games 1
+ud 1
+ov 1
+▁research 1
+▁depend 1
+▁bell 1
+ward 1
+bert 1
+▁move 1
+▁green 1
+▁debate 1
+▁professional 1
+▁others 1
+▁teacher 1
+den 1
+ez 1
+▁sch 1
+ions 1
+ven 1
+▁houses 1
+▁situation 1
+und 1
+cher 1
+▁lord 1
+▁italian 1
+▁round 1
+▁cat 1
+▁forest 1
+▁practice 1
+▁write 1
+▁bas 1
+▁reason 1
+▁favor 1
+ite 1
+▁nor 1
+▁territory 1
+▁usual 1
+ai 1
+▁tax 1
+ya 1
+tre 1
+▁past 1
+▁follow 1
+side 1
+▁dark 1
+▁jacques 1
+▁body 1
+▁bois 1
+▁par 1
+▁late 1
+▁anti 1
+gon 1
+▁arrived 1
+don 1
+▁turned 1
+ves 1
+▁parents 1
+▁dog 1
+▁mine 1
+▁heard 1
+▁cross 1
+▁meeting 1
+▁border 1
+▁allow 1
+zi 1
+▁above 1
+▁field 1
+▁least 1
+▁cre 1
+▁wa 1
+▁sainte 1
+▁province 1
+▁lake 1
+▁mission 1
+▁blue 1
+ab 1
+ow 1
+▁rich 1
+▁host 1
+▁beginning 1
+▁ser 1
+▁hall 1
+ory 1
+▁service 1
+▁del 1
+▁alone 1
+▁plant 1
+▁division 1
+▁information 1
+ally 1
+▁mal 1
+▁su 1
+lan 1
+▁asking 1
+onne 1
+tel 1
+wa 1
+▁stage 1
+▁philippe 1
+▁letter 1
+▁business 1
+▁champ 1
+▁van 1
+▁population 1
+▁pass 1
+▁andré 1
+ec 1
+▁organized 1
+▁education 1
+▁future 1
+▁hands 1
+▁reading 1
+▁followed 1
+▁si 1
+▁eat 1
+▁marc 1
+▁event 1
+▁friends 1
+▁yesterday 1
+ins 1
+ble 1
+ator 1
+▁fre 1
+▁approach 1
+▁program 1
+▁stopped 1
+▁y 1
+▁scene 1
+▁fo 1
+▁gi 1
+▁community 1
+▁inhabitants 1
+▁herself 1
+▁necessary 1
+ans 1
+▁simple 1
+▁goal 1
+▁luc 1
+▁difficult 1
+▁clear 1
+▁bro 1
+gan 1
+ast 1
+▁ab 1
+▁liked 1
+▁cra 1
+ux 1
+ction 1
+▁joseph 1
+▁itself 1
+▁star 1
+▁wish 1
+▁ber 1
+▁comp 1
+▁famous 1
+▁police 1
+▁window 1
+▁gri 1
+▁watch 1
+ée 1
+gen 1
+▁cast 1
+▁action 1
+ges 1
+▁revolution 1
+▁pu 1
+▁list 1
+mes 1
+▁bal 1
+▁leg 1
+▁europe 1
+▁leaves 1
+dic 1
+▁training 1
+▁decided 1
+▁mind 1
+▁ce 1
+▁ever 1
+▁version 1
+▁shows 1
+rin 1
+▁low 1
+▁test 1
+▁col 1
+▁race 1
+▁special 1
+▁leaving 1
+▁cro 1
+▁collection 1
+▁model 1
+▁mat 1
+▁due 1
+court 1
+dor 1
+▁moved 1
+▁making 1
+▁production 1
+wi 1
+▁beau 1
+▁wind 1
+▁serve 1
+ious 1
+▁henri 1
+▁rose 1
+vin 1
+▁cu 1
+pp 1
+▁private 1
+art 1
+▁completely 1
+▁behind 1
+▁rapporteur 1
+av 1
+her 1
+▁newspaper 1
+▁female 1
+▁defend 1
+▁week 1
+▁natural 1
+▁remained 1
+▁generally 1
+▁breton 1
+▁break 1
+▁hour 1
+▁especially 1
+▁numerous 1
+▁travel 1
+▁tall 1
+▁immediately 1
+▁store 1
+▁served 1
+▁hair 1
+▁pin 1
+▁cup 1
+▁fla 1
+oc 1
+ap 1
+▁cal 1
+▁view 1
+▁soul 1
+▁fu 1
+▁continued 1
+▁mou 1
+▁step 1
+net 1
+▁sent 1
+▁doubt 1
+▁albert 1
+rie 1
+hu 1
+lie 1
+▁source 1
+▁couple 1
+▁robert 1
+over 1
+▁fell 1
+ide 1
+▁killed 1
+▁personal 1
+qu 1
+▁figure 1
+▁ground 1
+▁gold 1
+▁pra 1
+▁within 1
+▁tower 1
+▁nevertheless 1
+▁someone 1
+▁development 1
+▁officer 1
+▁maybe 1
+▁actor 1
+ists 1
+▁para 1
+mit 1
+ship 1
+▁effect 1
+las 1
+▁progress 1
+ade 1
+lay 1
+▁interest 1
+▁control 1
+▁chief 1
+▁trans 1
+gne 1
+▁reasons 1
+▁novel 1
+▁opening 1
+▁j 1
+▁replaced 1
+▁top 1
+▁cemetery 1
+▁economic 1
+ative 1
+▁dr 1
+▁feel 1
+▁lady 1
+▁popular 1
+du 1
+▁brown 1
+▁joined 1
+▁prince 1
+▁neu 1
+▁association 1
+▁hot 1
+▁cur 1
+ole 1
+▁similar 1
+▁themselves 1
+▁space 1
+mor 1
+▁accept 1
+▁extreme 1
+▁knew 1
+ps 1
+tz 1
+▁access 1
+▁catholic 1
+ana 1
+▁pla 1
+igny 1
+▁agree 1
+▁respect 1
+▁outside 1
+▁mé 1
+▁fourth 1
+▁process 1
+▁critic 1
+▁headquarters 1
+vers 1
+▁design 1
+bb 1
+▁lead 1
+rou 1
+▁majority 1
+ches 1
+▁ville 1
+nay 1
+▁operation 1
+▁listen 1
+▁suggest 1
+▁judge 1
+▁union 1
+▁forward 1
+▁try 1
+des 1
+▁elected 1
+▁character 1
+▁mass 1
+▁produced 1
+ip 1
+▁remain 1
+▁los 1
+▁deal 1
+▁early 1
+ane 1
+▁neighbor 1
+▁covered 1
+▁guard 1
+euil 1
+▁key 1
+▁wrong 1
+▁wrote 1
+jo 1
+▁month 1
+▁anne 1
+ham 1
+▁closed 1
+by 1
+▁gro 1
+my 1
+▁leader 1
+▁bridge 1
+▁belong 1
+▁hours 1
+ria 1
+van 1
+▁allows 1
+▁chair 1
+▁send 1
+▁dedicated 1
+▁study 1
+▁ch 1
+▁administrative 1
+▁york 1
+tri 1
+▁consider 1
+▁nu 1
+ors 1
+ari 1
+▁fer 1
+▁bird 1
+▁wine 1
+▁denis 1
+▁priest 1
+▁stone 1
+▁kill 1
+fort 1
+▁bay 1
+car 1
+▁minutes 1
+non 1
+▁weather 1
+vent 1
+bra 1
+▁add 1
+▁glass 1
+▁gui 1
+▁object 1
+▁account 1
+▁probably 1
+▁contract 1
+pped 1
+▁term 1
+lier 1
+▁georges 1
+▁fra 1
+▁request 1
+▁running 1
+▁bel 1
+▁bru 1
+▁sau 1
+▁academy 1
+▁league 1
+▁gives 1
+ok 1
+ley 1
+▁drink 1
+▁organization 1
+▁mostly 1
+▁rep 1
+▁base 1
+▁inside 1
+▁students 1
+▁coach 1
+▁official 1
+▁card 1
+▁ship 1
+▁lived 1
+▁theater 1
+ev 1
+▁activity 1
+▁mayor 1
+ral 1
+▁actually 1
+▁rain 1
+▁composed 1
+▁deep 1
+illon 1
+▁bur 1
+▁bishop 1
+▁neighborhood 1
+▁remove 1
+▁construction 1
+▁fishing 1
+▁create 1
+▁legal 1
+▁managed 1
+▁bat 1
+▁available 1
+▁urban 1
+▁market 1
+té 1
+wn 1
+▁republic 1
+ieres 1
+▁london 1
+▁british 1
+▁comm 1
+▁sentence 1
+▁coast 1
+▁began 1
+▁sick 1
+▁franc 1
+▁wins 1
+▁thanks 1
+gar 1
+▁pen 1
+enne 1
+▁discover 1
+▁added 1
+▁secret 1
+▁beat 1
+▁ob 1
+▁painter 1
+▁historical 1
+oux 1
+▁bé 1
+▁cour 1
+▁health 1
+▁discovered 1
+▁led 1
+ome 1
+▁endemic 1
+▁science 1
+cle 1
+▁sound 1
+▁lou 1
+▁rou 1
+▁gen 1
+▁sw 1
+▁dream 1
+▁sold 1
+▁answered 1
+▁reach 1
+▁board 1
+▁share 1
+▁starts 1
+▁bla 1
+▁hotel 1
+▁developed 1
+ugh 1
+▁brothers 1
+▁evening 1
+ates 1
+▁canton 1
+▁tro 1
+▁rank 1
+▁method 1
+pin 1
+▁regional 1
+▁entire 1
+▁professor 1
+▁remind 1
+▁countries 1
+▁christian 1
+ified 1
+ub 1
+▁ph 1
+▁dance 1
+ought 1
+▁student 1
+▁edition 1
+▁march 1
+▁ready 1
+▁germany 1
+rat 1
+▁trip 1
+▁sell 1
+fe 1
+▁welcome 1
+▁degree 1
+▁increase 1
+▁stand 1
+▁secretary 1
+illa 1
+▁class 1
+▁opened 1
+▁sleep 1
+▁writer 1
+▁lower 1
+▁exist 1
+▁cho 1
+▁hospital 1
+▁fruit 1
+▁includ 1
+▁tired 1
+▁domini 1
+ex 1
+▁app 1
+▁win 1
+▁ran 1
+▁civil 1
+▁spanish 1
+red 1
+▁clos 1
+▁previous 1
+▁fine 1
+▁chapel 1
+▁decision 1
+▁situated 1
+▁kingdom 1
+ault 1
+▁engineer 1
+▁sl 1
+tier 1
+din 1
+▁network 1
+▁despite 1
+▁prove 1
+▁plane 1
+▁happen 1
+▁addition 1
+▁rule 1
+▁develop 1
+▁emperor 1
+▁finished 1
+bel 1
+▁comb 1
+▁link 1
+ban 1
+▁laurent 1
+▁function 1
+▁uncle 1
+▁track 1
+heim 1
+▁complete 1
+▁maintain 1
+▁piece 1
+ries 1
+aire 1
+▁experience 1
+▁arch 1
+▁foot 1
+fer 1
+▁painting 1
+▁calm 1
+▁occasion 1
+▁felt 1
+▁modern 1
+▁worth 1
+▁mol 1
+▁though 1
+▁nature 1
+▁western 1
+ais 1
+▁sat 1
+▁winter 1
+▁radio 1
+▁non 1
+can 1
+▁super 1
+▁sar 1
+▁stayed 1
+▁italy 1
+▁bill 1
+▁era 1
+▁singer 1
+▁wear 1
+▁refer 1
+▁pol 1
+▁carried 1
+▁football 1
+▁industry 1
+▁spread 1
+▁replied 1
+ines 1
+▁prefer 1
+▁songs 1
+▁lose 1
+ify 1
+▁saying 1
+▁wing 1
+del 1
+▁sh 1
+▁bed 1
+▁structure 1
+▁material 1
+▁marquis 1
+▁transport 1
+▁direction 1
+tive 1
+vert 1
+▁tried 1
+▁sav 1
+▁duke 1
+igne 1
+uc 1
+▁religious 1
+▁complex 1
+▁related 1
+▁influence 1
+▁parish 1
+▁victor 1
+erie 1
+▁contain 1
+we 1
+▁grow 1
+▁colleagues 1
+▁tonight 1
+▁ring 1
+quet 1
+lot 1
+▁reached 1
+pen 1
+▁gets 1
+clock 1
+▁document 1
+▁giving 1
+▁moving 1
+▁representative 1
+▁shi 1
+▁section 1
+ah 1
+▁page 1
+▁belongs 1
+▁pretty 1
+▁points 1
+▁active 1
+▁lie 1
+berg 1
+▁shot 1
+af 1
+ular 1
+▁tar 1
+▁er 1
+lor 1
+▁artist 1
+▁lawyer 1
+▁villa 1
+▁manage 1
+xi 1
+rupt 1
+ach 1
+▁ancient 1
+▁says 1
+▁plu 1
+▁placed 1
+▁rome 1
+▁yourself 1
+▁claude 1
+▁admir 1
+vel 1
+▁sand 1
+▁finish 1
+aux 1
+▁events 1
+▁pit 1
+▁lyon 1
+ft 1
+▁bor 1
+▁pale 1
+▁cou 1
+▁fair 1
+lar 1
+ique 1
+form 1
+▁press 1
+lia 1
+▁nicolas 1
+▁scream 1
+▁straight 1
+▁risk 1
+▁attract 1
+▁eye 1
+▁lack 1
+▁raised 1
+▁fan 1
+ius 1
+rac 1
+▁disc 1
+ag 1
+ji 1
+▁altitude 1
+▁england 1
+▁results 1
+▁angel 1
+▁lac 1
+▁brand 1
+pu 1
+▁mill 1
+ological 1
+bourg 1
+tes 1
+▁japan 1
+▁companies 1
+▁antoine 1
+▁valley 1
+▁election 1
+▁ya 1
+▁spot 1
+▁russian 1
+▁equal 1
+▁servant 1
+▁machine 1
+▁assembly 1
+▁ball 1
+▁else 1
+▁box 1
+▁rené 1
+▁belle 1
+▁bus 1
+▁silence 1
+▁pré 1
+▁trade 1
+▁relationship 1
+▁southern 1
+more 1
+▁paper 1
+▁agreement 1
+▁participate 1
+▁ter 1
+▁touch 1
+je 1
+ray 1
+zo 1
+▁activities 1
+▁advantage 1
+▁escape 1
+▁festival 1
+▁individual 1
+▁promise 1
+▁truth 1
+▁trees 1
+▁shop 1
+▁carry 1
+▁america 1
+▁device 1
+▁choice 1
+lam 1
+▁x 1
+▁hill 1
+▁traditional 1
+▁appointed 1
+▁frédéric 1
+▁ear 1
+▁regularly 1
+▁ben 1
+▁means 1
+▁college 1
+▁ré 1
+▁cell 1
+▁parliament 1
+▁fifth 1
+▁unknown 1
+▁match 1
+▁besides 1
+cent 1
+▁sad 1
+▁pick 1
+▁victory 1
+bar 1
+▁hy 1
+▁specific 1
+▁nowadays 1
+▁smile 1
+▁phone 1
+▁spain 1
+▁creation 1
+▁ke 1
+▁command 1
+▁cri 1
+ini 1
+▁environment 1
+▁presence 1
+rk 1
+▁included 1
+bre 1
+▁returned 1
+▁ban 1
+▁tre 1
+▁mad 1
+▁entered 1
+nne 1
+nier 1
+▁marriage 1
+▁reporter 1
+▁climb 1
+▁sou 1
+▁champion 1
+hen 1
+▁fur 1
+par 1
+burg 1
+ible 1
+umb 1
+▁normal 1
+▁bernard 1
+▁owner 1
+zz 1
+▁significant 1
+▁mas 1
+▁total 1
+▁established 1
+▁tur 1
+▁summer 1
+▁lé 1
+▁surviv 1
+▁limited 1
+▁unit 1
+▁require 1
+▁players 1
+▁concern 1
+uff 1
+▁animal 1
+▁ham 1
+ura 1
+▁address 1
+▁consist 1
+▁claim 1
+zy 1
+▁canada 1
+▁laugh 1
+▁spent 1
+▁anger 1
+▁questions 1
+▁shape 1
+▁services 1
+▁mil 1
+▁explain 1
+az 1
+▁david 1
+▁organ 1
+ik 1
+lat 1
+▁sé 1
+▁learn 1
+ene 1
+▁roll 1
+▁empire 1
+▁monument 1
+▁daniel 1
+▁changed 1
+▁arms 1
+▁buildings 1
+▁reference 1
+▁africa 1
+duc 1
+▁propose 1
+▁oldest 1
+▁upon 1
+▁artists 1
+▁sit 1
+▁shouted 1
+rand 1
+▁sports 1
+▁male 1
+▁additional 1
+▁counter 1
+yl 1
+▁sil 1
+▁passed 1
+▁enjoy 1
+▁easy 1
+▁stra 1
+▁partner 1
+▁soldiers 1
+▁speaking 1
+▁ram 1
+▁presented 1
+▁leads 1
+▁independent 1
+▁shoulder 1
+▁soft 1
+▁est 1
+▁represents 1
+▁op 1
+▁gene 1
+▁han 1
+istic 1
+dou 1
+▁avoid 1
+▁northern 1
+cho 1
+▁cover 1
+ail 1
+▁august 1
+▁sam 1
+▁feeling 1
+▁force 1
+▁correspond 1
+elles 1
+▁loire 1
+▁languages 1
+▁blo 1
+▁aim 1
+▁destroyed 1
+▁protect 1
+▁julien 1
+▁fri 1
+▁improve 1
+▁draw 1
+▁farm 1
+▁chairs 1
+▁surprised 1
+▁record 1
+▁lane 1
+cio 1
+▁remember 1
+▁mistake 1
+▁solution 1
+▁birth 1
+▁blood 1
+▁bread 1
+▁matches 1
+▁conditions 1
+▁released 1
+▁fort 1
+▁sara 1
+▁ended 1
+▁relative 1
+▁germain 1
+▁food 1
+▁cinema 1
+▁forget 1
+▁receive 1
+ino 1
+fo 1
+▁lion 1
+▁dam 1
+▁cli 1
+▁architect 1
+▁commercial 1
+▁deliver 1
+▁drawing 1
+▁abbey 1
+▁kiss 1
+ual 1
+▁involved 1
+▁petit 1
+▁characters 1
+ept 1
+▁equally 1
+▁appears 1
+▁whom 1
+▁express 1
+▁broke 1
+▁benefit 1
+▁cultural 1
+▁afterwards 1
+▁earth 1
+▁medal 1
+▁bear 1
+▁size 1
+rd 1
+▁happened 1
+▁movies 1
+eur 1
+▁kid 1
+▁noticed 1
+ography 1
+▁flat 1
+▁television 1
+▁affair 1
+▁limit 1
+▁mark 1
+agne 1
+▁francis 1
+▁reality 1
+▁wo 1
+▁pull 1
+▁animals 1
+▁using 1
+oche 1
+▁spend 1
+▁tree 1
+▁prevent 1
+▁poet 1
+pre 1
+▁dan 1
+▁rem 1
+▁succeed 1
+▁cathedral 1
+▁inspired 1
+▁computer 1
+▁foreign 1
+▁japanese 1
+▁heavy 1
+▁municipalities 1
+▁chemin 1
+▁prior 1
+ium 1
+▁gather 1
+▁consul 1
+▁ice 1
+▁colleague 1
+ctor 1
+▁fast 1
+▁cab 1
+▁sab 1
+▁thinking 1
+▁interested 1
+▁collect 1
+▁aud 1
+▁deputy 1
+▁teams 1
+▁photograph 1
+oul 1
+▁farmer 1
+▁telling 1
+▁mount 1
+▁literature 1
+▁quebec 1
+▁society 1
+▁complain 1
+▁sky 1
+▁argument 1
+▁croix 1
+▁apartment 1
+▁reform 1
+▁wal 1
+▁mac 1
+▁iron 1
+▁prison 1
+▁trying 1
+aries 1
+▁compar 1
+ange 1
+▁financial 1
+▁pot 1
+▁performance 1
+▁quarter 1
+▁sport 1
+ker 1
+▁walls 1
+▁surrounded 1
+▁volume 1
+▁represented 1
+oli 1
+▁suddenly 1
+▁cy 1
+ire 1
+ieu 1
+cul 1
+▁min 1
+▁wants 1
+▁broadcast 1
+▁expensive 1
+▁families 1
+▁instrument 1
+▁symbol 1
+▁jump 1
+▁prize 1
+▁estate 1
+▁slowly 1
+▁represent 1
+▁sing 1
+▁discuss 1
+▁arts 1
+▁château 1
+▁impact 1
+▁gabriel 1
+▁electric 1
+▁price 1
+▁objective 1
+▁native 1
+▁cook 1
+▁richard 1
+illy 1
+ability 1
+▁flag 1
+▁bourg 1
+▁letters 1
+▁serious 1
+▁grave 1
+▁drive 1
+▁via 1
+▁episode 1
+▁concept 1
+▁maurice 1
+▁flor 1
+▁cave 1
+▁youth 1
+▁lover 1
+cor 1
+▁budget 1
+▁slightly 1
+▁breath 1
+▁understood 1
+▁standard 1
+▁direct 1
+ama 1
+▁divided 1
+ib 1
+eux 1
+▁forces 1
+▁leading 1
+▁designed 1
+▁disappeared 1
+coming 1
+▁wh 1
+▁ker 1
+▁equipped 1
+▁exchange 1
+▁marseille 1
+▁canadian 1
+▁rob 1
+▁firm 1
+▁perform 1
+ek 1
+▁holding 1
+▁huge 1
+▁passion 1
+▁vau 1
+▁settle 1
+▁linked 1
+▁brave 1
+▁lock 1
+eaux 1
+▁institution 1
+rel 1
+▁media 1
+▁belgium 1
+▁purpose 1
+▁tournament 1
+▁otherwise 1
+▁initial 1
+▁needed 1
+▁fau 1
+ache 1
+▁abandoned 1
+▁data 1
+▁entirely 1
+▁recognized 1
+▁aux 1
+▁provision 1
+▁attention 1
+▁either 1
+▁weapon 1
+▁actress 1
+▁library 1
+▁edge 1
+▁sector 1
+▁continue 1
+▁rate 1
+bou 1
+▁million 1
+▁dies 1
+▁rise 1
+▁flow 1
+▁dar 1
+▁lesson 1
+▁sac 1
+▁appeared 1
+▁eating 1
+▁mau 1
+ification 1
+▁amount 1
+▁concert 1
+▁policy 1
+▁allowed 1
+▁thomas 1
+▁hurt 1
+▁airport 1
+▁finance 1
+▁management 1
+▁simon 1
+har 1
+bro 1
+▁dame 1
+▁enable 1
+▁certainly 1
+▁regular 1
+▁whe 1
+▁gone 1
+▁replace 1
+▁flowers 1
+▁negative 1
+▁succeeded 1
+▁coll 1
+▁cul 1
+▁pair 1
+▁offered 1
+▁cher 1
+▁tail 1
+▁bare 1
+▁bag 1
+▁formed 1
+▁semi 1
+mé 1
+▁spirit 1
+▁industrial 1
+▁institute 1
+▁colonel 1
+▁block 1
+▁chart 1
+▁rare 1
+uz 1
+▁cru 1
+▁contains 1
+cia 1
+▁registered 1
+▁aims 1
+▁bau 1
+▁thi 1
+▁indian 1
+▁ou 1
+▁gun 1
+▁clearly 1
+▁assistant 1
+▁statue 1
+▁housing 1
+though 1
+▁dress 1
+▁queen 1
+▁favorable 1
+▁chance 1
+▁chat 1
+▁hide 1
+▁prisoner 1
+▁video 1
+ica 1
+ara 1
+▁noise 1
+▁pieces 1
+▁kids 1
+▁impossible 1
+▁effort 1
+▁century 1
+▁snow 1
+▁memory 1
+▁editorial 1
+▁cousin 1
+▁piano 1
+▁sense 1
+▁quiet 1
+▁successful 1
+▁cities 1
+▁unfortunately 1
+▁lucien 1
+▁ros 1
+gor 1
+▁scar 1
+old 1
+▁mentioned 1
+▁coat 1
+▁windows 1
+▁restaurant 1
+▁across 1
+▁surface 1
+▁medicine 1
+▁enemy 1
+▁focus 1
+▁crazy 1
+▁produce 1
+▁notably 1
+▁girls 1
+▁surprise 1
+▁listening 1
+▁musical 1
+▁proposed 1
+▁mid 1
+ani 1
+▁terms 1
+▁difference 1
+▁entrance 1
+▁principle 1
+▁recommend 1
+▁search 1
+▁appearance 1
+▁palace 1
+lais 1
+▁signed 1
+ute 1
+▁ze 1
+▁organize 1
+▁save 1
+▁catch 1
+▁dry 1
+▁coffee 1
+▁lieutenant 1
+▁residence 1
+▁nobody 1
+▁studio 1
+▁length 1
+ever 1
+▁clothes 1
+▁feet 1
+▁poly 1
+▁tea 1
+zar 1
+▁christ 1
+▁za 1
+▁ur 1
+▁vol 1
+▁completed 1
+▁everywhere 1
+▁excellent 1
+▁proposal 1
+▁worry 1
+▁clean 1
+▁treat 1
+▁producer 1
+▁boat 1
+uri 1
+▁flo 1
+▁pet 1
+▁ahead 1
+▁patric 1
+▁adult 1
+▁greek 1
+▁architecture 1
+▁dor 1
+pper 1
+part 1
+▁ren 1
+▁boss 1
+▁jan 1
+▁mono 1
+▁vice 1
+▁topic 1
+▁sy 1
+▁islands 1
+▁specializ 1
+▁display 1
+▁chinese 1
+▁medical 1
+▁louise 1
+▁murder 1
+zan 1
+rop 1
+▁lan 1
+▁introduced 1
+▁range 1
+row 1
+▁guillaume 1
+▁culture 1
+▁lines 1
+▁balance 1
+▁metal 1
+▁berlin 1
+▁throw 1
+▁regions 1
+▁pain 1
+▁george 1
+▁rat 1
+▁holiday 1
+▁olympic 1
+▁easily 1
+▁belgian 1
+▁taught 1
+▁psych 1
+▁honest 1
+▁crossed 1
+▁image 1
+gues 1
+▁discussion 1
+oy 1
+▁mouth 1
+▁specie 1
+▁pé 1
+▁higher 1
+▁defended 1
+▁obtained 1
+gh 1
+ku 1
+▁bottom 1
+▁california 1
+▁accident 1
+▁speech 1
+▁portrait 1
+▁visible 1
+▁demand 1
+▁benches 1
+▁lit 1
+▁bour 1
+▁dogs 1
+yr 1
+▁appear 1
+com 1
+▁problems 1
+▁filled 1
+▁refused 1
+▁forced 1
+▁cote 1
+▁tian 1
+▁gou 1
+▁fund 1
+▁responsible 1
+▁scientific 1
+▁neither 1
+▁sunday 1
+▁constant 1
+▁caught 1
+▁opposition 1
+▁shoot 1
+▁ruin 1
+▁defeat 1
+▁grey 1
+▁zone 1
+▁roche 1
+vier 1
+▁loud 1
+▁arrive 1
+▁dare 1
+▁clo 1
+▁corner 1
+▁elephants 1
+▁tourist 1
+▁reign 1
+▁wide 1
+hy 1
+▁docteur 1
+▁procedure 1
+▁victim 1
+▁resistance 1
+▁pur 1
+ella 1
+▁passage 1
+▁mur 1
+▁china 1
+▁mention 1
+▁fro 1
+dé 1
+▁apart 1
+▁paintings 1
+▁dad 1
+▁auguste 1
+bon 1
+▁campaign 1
+▁picture 1
+▁credit 1
+▁attend 1
+▁joy 1
+▁careful 1
+▁absolutely 1
+qua 1
+▁multi 1
+▁simply 1
+▁seine 1
+gny 1
+▁refuse 1
+▁peter 1
+back 1
+▁areas 1
+▁pri 1
+▁phil 1
+▁application 1
+▁diocese 1
+▁energy 1
+▁channel 1
+▁abbot 1
+▁worse 1
+▁property 1
+▁height 1
+▁learned 1
+phe 1
+▁fun 1
+▁standing 1
+▁noble 1
+▁marcel 1
+▁marry 1
+▁feed 1
+port 1
+▁adventure 1
+▁stories 1
+▁closer 1
+▁knee 1
+▁records 1
+▁fault 1
+▁spring 1
+east 1
+▁ubu 1
+▁expression 1
+illard 1
+▁realize 1
+▁invited 1
+▁condition 1
+▁settled 1
+▁starting 1
+chy 1
+▁tam 1
+▁peace 1
+▁advance 1
+▁agriculture 1
+▁disease 1
+▁nephew 1
+▁mini 1
+▁defense 1
+▁stream 1
+▁freedom 1
+▁swiss 1
+▁william 1
+▁favorite 1
+▁journalist 1
+▁build 1
+▁workers 1
+▁musician 1
+▁label 1
+▁ger 1
+▁delete 1
+▁sum 1
+▁ven 1
+ax 1
+▁triple 1
+▁gain 1
+▁sixth 1
+▁equ 1
+▁fou 1
+▁excuse 1
+▁ocean 1
+▁software 1
+▁conduct 1
+▁founder 1
+▁helped 1
+▁drama 1
+▁expect 1
+▁tin 1
+▁eli 1
+▁covers 1
+▁fat 1
+▁cent 1
+ignac 1
+▁spa 1
+phi 1
+war 1
+▁branch 1
+fu 1
+▁suit 1
+▁prime 1
+▁liber 1
+stood 1
+▁crowd 1
+▁distance 1
+▁self 1
+▁protestant 1
+ju 1
+▁rush 1
+▁luck 1
+▁chau 1
+▁reserve 1
+▁engine 1
+hon 1
+▁crossing 1
+▁stadium 1
+▁elvire 1
+▁positive 1
+▁troops 1
+▁session 1
+▁tears 1
+▁minor 1
+▁meaning 1
+▁exam 1
+▁léon 1
+▁bl 1
+▁shortly 1
+▁file 1
+▁cer 1
+▁ago 1
+▁eu 1
+▁strange 1
+▁participated 1
+▁pav 1
+▁employ 1
+▁circle 1
+▁contact 1
+▁brazil 1
+▁christophe 1
+▁gilles 1
+▁pilot 1
+▁finger 1
+▁skill 1
+▁follows 1
+▁exists 1
+▁extend 1
+▁bears 1
+▁moulin 1
+▁foundation 1
+▁burn 1
+gna 1
+▁chosen 1
+▁concerns 1
+▁eva 1
+▁arc 1
+ken 1
+▁accepted 1
+▁lar 1
+jou 1
+▁mai 1
+▁recognize 1
+▁baron 1
+▁anna 1
+▁guide 1
+ologist 1
+▁empty 1
+▁location 1
+▁shut 1
+▁recorded 1
+▁concerned 1
+▁trust 1
+▁obtain 1
+▁paid 1
+▁continues 1
+▁euros 1
+▁blu 1
+matic 1
+▁print 1
+▁publication 1
+▁ski 1
+ograph 1
+▁gé 1
+▁latin 1
+▁pray 1
+▁economy 1
+▁opportunity 1
+▁étienne 1
+▁facade 1
+▁interior 1
+▁childhood 1
+▁removed 1
+▁speed 1
+▁eugen 1
+▁sole 1
+▁lift 1
+▁protected 1
+▁founded 1
+▁mary 1
+▁suffer 1
+▁score 1
+▁plants 1
+▁kar 1
+▁review 1
+▁composition 1
+▁workshop 1
+▁reveal 1
+▁vin 1
+▁powerful 1
+▁esp 1
+niere 1
+▁desire 1
+ieux 1
+▁dro 1
+▁studying 1
+▁mixed 1
+bé 1
+▁fin 1
+▁hou 1
+▁essential 1
+▁enters 1
+lk 1
+▁wave 1
+▁afraid 1
+▁agricultural 1
+▁contrary 1
+▁draft 1
+▁roof 1
+▁eastern 1
+▁shock 1
+▁fool 1
+▁staff 1
+▁pal 1
+▁value 1
+mann 1
+▁billion 1
+▁australia 1
+rier 1
+▁separated 1
+▁cle 1
+▁cry 1
+▁decides 1
+ula 1
+▁bother 1
+q 1
+▁weight 1
+ple 1
+▁weak 1
+▁talked 1
+▁tru 1
+▁hurry 1
+▁regret 1
+▁phase 1
+▁justice 1
+▁awarded 1
+▁elements 1
+kin 1
+ima 1
+▁flows 1
+▁taste 1
+▁associated 1
+▁dé 1
+▁check 1
+▁weeks 1
+▁pat 1
+▁tradition 1
+▁recent 1
+▁var 1
+▁administration 1
+▁instead 1
+▁anyone 1
+▁sailor 1
+▁golden 1
+▁calling 1
+▁manager 1
+▁gal 1
+▁row 1
+▁vent 1
+etic 1
+▁begins 1
+west 1
+▁marks 1
+▁mess 1
+▁puy 1
+▁challenge 1
+▁smell 1
+▁lunch 1
+▁funny 1
+▁percent 1
+▁carrier 1
+▁provide 1
+▁highly 1
+▁steps 1
+▁bains 1
+▁temple 1
+▁prepar 1
+▁chamber 1
+▁directly 1
+▁dia 1
+▁flower 1
+▁trou 1
+▁haut 1
+▁oil 1
+▁shin 1
+▁performed 1
+▁attempt 1
+▁opposite 1
+▁twice 1
+cou 1
+▁flee 1
+▁convention 1
+▁authorities 1
+▁drum 1
+▁minute 1
+▁enter 1
+▁hor 1
+▁owe 1
+▁maintained 1
+▁trained 1
+▁wild 1
+▁frame 1
+▁syn 1
+▁india 1
+emp 1
+▁receives 1
+▁accompanied 1
+▁forbidden 1
+▁magazine 1
+▁switzerland 1
+▁stretch 1
+▁target 1
+▁ensure 1
+▁marked 1
+▁ku 1
+▁exhibition 1
+▁roger 1
+isse 1
+▁occupied 1
+aine 1
+▁bomb 1
+▁fixed 1
+▁pou 1
+▁features 1
+▁tools 1
+▁bigge 1
+▁danger 1
+▁october 1
+▁attached 1
+▁diploma 1
+▁refuge 1
+▁climate 1
+▁achieve 1
+▁average 1
+▁dinner 1
+▁winner 1
+▁rural 1
+▁terrible 1
+▁yo 1
+▁highest 1
+▁blow 1
+▁adopt 1
+▁lor 1
+lip 1
+▁hence 1
+▁upper 1
+▁operate 1
+▁plate 1
+▁commander 1
+▁provided 1
+▁lang 1
+▁beach 1
+▁provides 1
+▁tomb 1
+▁hamlet 1
+▁polish 1
+▁comment 1
+▁yu 1
+cq 1
+▁translated 1
+▁walking 1
+▁onto 1
+▁gap 1
+▁skin 1
+▁zo 1
+wer 1
+ola 1
+▁occur 1
+mber 1
+▁injured 1
+▁quality 1
+▁raymond 1
+▁sorry 1
+▁egg 1
+▁ladies 1
+▁crown 1
+▁correct 1
+▁wake 1
+▁believed 1
+not 1
+▁preserved 1
+ano 1
+▁perfectly 1
+▁sail 1
+▁teaching 1
+olo 1
+var 1
+▁adopted 1
+▁sin 1
+▁african 1
+▁withdrawal 1
+▁municipal 1
+▁virgin 1
+▁happiness 1
+▁monastery 1
+▁arrival 1
+▁mexico 1
+▁traffic 1
+▁choose 1
+▁legend 1
+▁contest 1
+▁daily 1
+▁gate 1
+▁selection 1
+▁henry 1
+▁effective 1
+▁corn 1
+bach 1
+▁moral 1
+▁quin 1
+▁chan 1
+▁perfect 1
+▁bedroom 1
+▁custom 1
+▁primary 1
+▁governor 1
+▁peak 1
+▁gray 1
+▁alain 1
+▁grass 1
+▁flu 1
+▁chest 1
+ick 1
+▁fill 1
+▁older 1
+▁repeated 1
+▁tho 1
+bri 1
+kh 1
+work 1
+gno 1
+▁candidate 1
+▁bottle 1
+gging 1
+▁magn 1
+▁bert 1
+▁max 1
+▁exactly 1
+▁yve 1
+▁orange 1
+▁referr 1
+▁qui 1
+▁release 1
+▁advis 1
+▁ty 1
+izing 1
+hor 1
+gli 1
+rich 1
+▁principal 1
+▁april 1
+▁knowledge 1
+▁wolf 1
+▁fontaine 1
+▁damage 1
+▁domain 1
+▁laughing 1
+▁younger 1
+▁republican 1
+off 1
+▁relie 1
+▁citizens 1
+▁notice 1
+▁painted 1
+▁hum 1
+▁suffered 1
+ova 1
+▁maria 1
+▁independence 1
+▁ourselves 1
+▁philosophy 1
+▁jewish 1
+▁pencil 1
+▁pleasure 1
+▁physical 1
+▁witness 1
+▁trouble 1
+▁severe 1
+▁sort 1
+▁connect 1
+▁showing 1
+▁manner 1
+▁maur 1
+▁plain 1
+▁invite 1
+ford 1
+▁chant 1
+ignan 1
+▁tries 1
+▁scared 1
+atic 1
+chu 1
+▁paragraph 1
+▁phenomenon 1
+▁satisfied 1
+▁knock 1
+▁historian 1
+▁resign 1
+▁importance 1
+▁vincent 1
+▁cloud 1
+▁below 1
+▁might 1
+▁silver 1
+▁kilometers 1
+▁sales 1
+▁hello 1
+ua 1
+▁authority 1
+▁announc 1
+▁alive 1
+▁drop 1
+▁ferr 1
+▁ken 1
+▁fighting 1
+▁por 1
+▁warm 1
+mond 1
+ppe 1
+▁raise 1
+▁haute 1
+ami 1
+▁increased 1
+▁celebrate 1
+▁toulouse 1
+▁railway 1
+▁beyond 1
+▁knight 1
+▁fernand 1
+▁afternoon 1
+▁comprise 1
+▁fresh 1
+▁jule 1
+▁desert 1
+▁boul 1
+▁obviously 1
+bal 1
+▁admit 1
+▁temperature 1
+▁gérard 1
+▁peasant 1
+▁suspect 1
+▁violent 1
+▁platform 1
+▁generation 1
+▁status 1
+▁strip 1
+▁establishment 1
+▁courage 1
+▁sim 1
+▁jack 1
+▁opposed 1
+▁rent 1
+▁selected 1
+▁mag 1
+▁repeat 1
+▁composer 1
+▁terre 1
+▁nation 1
+▁ul 1
+▁damn 1
+▁philip 1
+▁vern 1
+▁technical 1
+▁vehicle 1
+▁algeria 1
+▁obey 1
+▁stock 1
+▁existence 1
+▁earlier 1
+lio 1
+▁recording 1
+▁sous 1
+▁commissioner 1
+▁greatest 1
+▁thé 1
+▁products 1
+▁classes 1
+▁rouss 1
+▁glasses 1
+▁failed 1
+▁vic 1
+lant 1
+▁slow 1
+▁except 1
+▁cass 1
+▁worried 1
+▁patient 1
+▁expert 1
+▁sweet 1
+▁renew 1
+▁smoke 1
+▁labor 1
+▁picard 1
+▁bright 1
+▁internet 1
+▁james 1
+▁silent 1
+▁treatment 1
+▁font 1
+anne 1
+▁wedding 1
+elli 1
+▁mom 1
+▁pun 1
+▁trial 1
+ight 1
+▁saved 1
+▁threat 1
+▁aspect 1
+▁experiment 1
+▁technique 1
+▁tributary 1
+▁suppli 1
+▁everyday 1
+ock 1
+▁hydro 1
+▁shoes 1
+▁meters 1
+cker 1
+▁june 1
+une 1
+asse 1
+▁carbon 1
+▁mel 1
+▁details 1
+▁typical 1
+▁drunk 1
+▁pretend 1
+▁facing 1
+▁failure 1
+▁income 1
+▁task 1
+▁learning 1
+▁reg 1
+▁disappear 1
+▁secondary 1
+▁interesting 1
+▁parties 1
+▁july 1
+▁element 1
+▁employees 1
+▁quick 1
+▁faith 1
+▁atmosphere 1
+▁decrease 1
+▁september 1
+▁voted 1
+▁qualified 1
+▁catherine 1
+▁decorated 1
+▁quentin 1
+▁decree 1
+▁annoy 1
+▁characteristic 1
+▁classified 1
+▁flight 1
+right 1
+▁epi 1
+▁protection 1
+▁motor 1
+▁declared 1
+▁hol 1
+kov 1
+▁sho 1
+▁religion 1
+▁emotion 1
+▁senate 1
+▁retain 1
+▁security 1
+▁wheel 1
+time 1
+▁useful 1
+▁mc 1
+▁liberal 1
+▁arrested 1
+▁establish 1
+▁hang 1
+▁breed 1
+▁extra 1
+▁pushed 1
+▁salt 1
+▁temp 1
+▁broken 1
+▁strength 1
+▁potential 1
+▁brussels 1
+▁explanation 1
+▁global 1
+▁olivier 1
+▁evolution 1
+▁intend 1
+▁korea 1
+▁blanc 1
+▁dig 1
+▁baby 1
+▁rap 1
+▁compete 1
+▁jou 1
+▁confirmed 1
+▁russia 1
+▁affected 1
+▁influenced 1
+while 1
+▁encourage 1
+▁mechanism 1
+▁reinforce 1
+▁recruit 1
+▁ticket 1
+▁instruct 1
+▁ernest 1
+▁absence 1
+▁appeal 1
+▁supposed 1
+▁deriv 1
+▁missing 1
+▁opera 1
+▁sleeping 1
+▁stable 1
+▁politic 1
+▁separate 1
+▁sheet 1
+▁bol 1
+ida 1
+▁monte 1
+itz 1
+▁watching 1
+▁journal 1
+▁dressed 1
+field 1
+▁entrust 1
+▁landscape 1
+▁observation 1
+▁sacrifice 1
+▁alfred 1
+▁federation 1
+▁stupid 1
+▁monday 1
+▁screen 1
+▁storm 1
+▁ajac 1
+▁dutch 1
+▁growth 1
+▁alexandre 1
+▁happening 1
+▁bow 1
+nois 1
+▁none 1
+▁factor 1
+isle 1
+rine 1
+▁nave 1
+▁signal 1
+▁constitute 1
+▁furniture 1
+▁orchestra 1
+▁implement 1
+▁jealous 1
+▁heritage 1
+▁brief 1
+▁proceed 1
+▁angry 1
+▁dangerous 1
+house 1
+▁reaction 1
+▁soil 1
+▁tourism 1
+bois 1
+▁driver 1
+▁ranking 1
+▁que 1
+▁paint 1
+▁pink 1
+▁stu 1
+▁classic 1
+▁simpl 1
+▁mix 1
+light 1
+▁liv 1
+▁directed 1
+▁applause 1
+▁conflict 1
+▁xavier 1
+▁eventually 1
+gé 1
+▁arnaud 1
+▁shelter 1
+▁content 1
+▁smith 1
+▁publish 1
+roux 1
+▁marius 1
+▁strongly 1
+▁bringing 1
+▁meat 1
+▁legion 1
+▁justin 1
+▁loo 1
+▁magic 1
+▁bab 1
+▁von 1
+▁tale 1
+mark 1
+▁passes 1
+▁bern 1
+▁sudden 1
+▁fix 1
+▁exhibit 1
+nais 1
+▁defeated 1
+▁stan 1
+▁resident 1
+▁conversation 1
+▁sword 1
+▁stairs 1
+▁planned 1
+▁tardy 1
+▁stick 1
+▁relation 1
+▁expressed 1
+▁hearing 1
+▁unique 1
+▁precisely 1
+hol 1
+▁withdrawn 1
+▁cé 1
+▁chap 1
+oire 1
+▁announced 1
+▁marne 1
+uge 1
+pha 1
+▁cool 1
+▁perhaps 1
+▁responsibility 1
+▁cardinal 1
+▁talent 1
+▁sharp 1
+▁argent 1
+▁clu 1
+▁modest 1
+▁gor 1
+▁cried 1
+stein 1
+▁honorable 1
+ivity 1
+evre 1
+▁reward 1
+▁begin 1
+ché 1
+▁grant 1
+bonne 1
+▁sul 1
+▁gil 1
+▁pér 1
+▁extended 1
+▁demo 1
+▁ji 1
+▁sala 1
+▁consequence 1
+▁executive 1
+▁bordeaux 1
+▁michael 1
+▁regulation 1
+▁terror 1
+▁ray 1
+▁sugar 1
+▁israel 1
+▁detect 1
+▁crime 1
+▁shame 1
+▁thin 1
+▁vert 1
+▁recently 1
+▁brive 1
+▁classical 1
+▁accus 1
+▁crosses 1
+▁lille 1
+▁pour 1
+▁listed 1
+▁monk 1
+▁possess 1
+▁regime 1
+▁capacity 1
+▁soviet 1
+▁repair 1
+▁slave 1
+baptiste 1
+▁serge 1
+▁aunt 1
+▁corps 1
+▁dominat 1
+▁specialist 1
+▁emil 1
+▁factory 1
+▁vé 1
+range 1
+▁citizen 1
+▁basic 1
+▁milk 1
+▁push 1
+▁spoke 1
+▁agent 1
+▁neo 1
+▁connected 1
+▁sight 1
+▁trail 1
+▁consequent 1
+▁gradually 1
+▁opponent 1
+▁column 1
+▁proof 1
+▁blind 1
+▁solid 1
+gent 1
+▁background 1
+▁gift 1
+▁editor 1
+▁bio 1
+▁marine 1
+▁trace 1
+▁bored 1
+▁reject 1
+▁successor 1
+▁aware 1
+▁eighth 1
+▁solve 1
+▁laval 1
+▁ol 1
+▁edit 1
+▁approximately 1
+▁conference 1
+▁equipment 1
+▁negotiation 1
+▁occupation 1
+▁prohibit 1
+▁ministry 1
+▁tropical 1
+▁incident 1
+▁pocket 1
+▁courtyard 1
+▁decline 1
+▁slope 1
+▁cyr 1
+▁fellow 1
+▁twin 1
+▁investigation 1
+▁proper 1
+hr 1
+▁agreed 1
+▁lips 1
+▁wonder 1
+▁artistic 1
+▁dubourg 1
+▁bull 1
+board 1
+▁greet 1
+▁din 1
+▁described 1
+▁falling 1
+▁rejected 1
+▁previously 1
+▁officially 1
+▁parisian 1
+▁zi 1
+▁entering 1
+▁naturally 1
+▁gall 1
+▁seventh 1
+etti 1
+▁construct 1
+▁introduce 1
+▁basketball 1
+▁contemporary 1
+▁netherlands 1
+▁variety 1
+▁adjust 1
+▁departure 1
+▁continent 1
+▁decoration 1
+▁gentlemen 1
+▁vign 1
+▁hole 1
+▁chateau 1
+▁nearby 1
+▁émil 1
+▁literar 1
+▁poem 1
+▁lamp 1
+▁decide 1
+▁representation 1
+▁lev 1
+well 1
+▁growing 1
+▁nail 1
+▁deny 1
+▁hunter 1
+▁register 1
+▁basin 1
+▁raining 1
+▁required 1
+ware 1
+▁delay 1
+▁fail 1
+▁tie 1
+▁disturb 1
+▁manuscript 1
+▁surrounding 1
+▁triumph 1
+▁november 1
+▁december 1
+▁driving 1
+▁shadow 1
+▁copies 1
+▁rifle 1
+▁frank 1
+▁pursue 1
+▁sculpture 1
+▁message 1
+▁countess 1
+▁guer 1
+▁elder 1
+▁thick 1
+dding 1
+▁hunting 1
+itive 1
+▁nest 1
+▁affect 1
+▁austria 1
+▁suffering 1
+▁bourbon 1
+▁bapt 1
+▁passing 1
+▁carrie 1
+bil 1
+▁surprising 1
+▁apparently 1
+▁extension 1
+▁feather 1
+▁passenger 1
+▁junior 1
+▁properties 1
+▁gaillarde 1
+▁tank 1
+▁micro 1
+▁internal 1
+▁wealth 1
+▁fleur 1
+▁guess 1
+▁parliamentary 1
+▁grande 1
+▁drug 1
+▁claire 1
+▁restored 1
+▁angeles 1
+▁logic 1
+aye 1
+▁pel 1
+▁taxes 1
+▁indicate 1
+▁bonne 1
+atory 1
+▁pope 1
+▁prepared 1
+▁barbe 1
+▁greatly 1
+▁originat 1
+▁drawn 1
+▁attribut 1
+▁installed 1
+▁nonetheless 1
+▁clarification 1
+▁committed 1
+▁parallel 1
+▁pascal 1
+▁marcof 1
+▁prefecture 1
+▁descendant 1
+▁retirement 1
+▁difficulties 1
+▁ye 1
+▁safe 1
+▁whispered 1
+▁essentially 1
+▁contributed 1
+▁spoken 1
+▁launched 1
+zé 1
+▁dome 1
+▁compos 1
+stone 1
+▁soldier 1
+▁chose 1
+▁chamb 1
+▁shaped 1
+▁strike 1
+▁cape 1
+▁rust 1
+▁guitar 1
+▁congress 1
+▁communities 1
+▁général 1
+▁assassin 1
+▁strasbourg 1
+▁britain 1
+phone 1
+▁scholar 1
+enberg 1
+▁serbia 1
+▁accessible 1
+▁holy 1
+▁mail 1
+▁olive 1
+▁bound 1
+▁sale 1
+▁stepp 1
+▁fled 1
+▁basis 1
+▁socialist 1
+plan 1
+gged 1
+▁pool 1
+▁apple 1
+▁associate 1
+▁pure 1
+▁euro 1
+▁radi 1
+▁initiat 1
+▁divorce 1
+▁territories 1
+▁reputation 1
+▁valérie 1
+▁federal 1
+▁gesture 1
+▁imperial 1
+▁transformed 1
+▁modified 1
+▁electronic 1
+▁recall 1
+▁crisis 1
+▁comic 1
+▁damien 1
+card 1
+▁option 1
+▁subsequently 1
+graphic 1
+▁poetry 1
+▁fold 1
+▁reflection 1
+▁poland 1
+▁defender 1
+▁imagin 1
+▁rio 1
+▁neck 1
+▁till 1
+▁seriously 1
+▁root 1
+fall 1
+hou 1
+▁australian 1
+▁product 1
+▁contribute 1
+▁teache 1
+▁photo 1
+▁fit 1
+▁difficulty 1
+▁pig 1
+▁dialogue 1
+▁evolve 1
+▁resources 1
+▁sensitive 1
+▁remarkable 1
+▁maxime 1
+▁widow 1
+▁nantes 1
+▁forgotten 1
+▁transfer 1
+villiers 1
+▁furious 1
+▁inspector 1
+▁insult 1
+zac 1
+▁patron 1
+▁proud 1
+fen 1
+▁fought 1
+▁eliminated 1
+▁useless 1
+▁colomb 1
+▁branches 1
+▁graduated 1
+lette 1
+▁wise 1
+▁confirm 1
+reux 1
+▁reserved 1
+▁emergenc 1
+▁reduce 1
+▁churches 1
+▁southwest 1
+▁conclude 1
+▁identified 1
+▁criminal 1
+▁keinec 1
+▁expand 1
+▁companion 1
+▁dirty 1
+▁devoted 1
+▁applied 1
+▁vessel 1
+▁spiritual 1
+▁champagne 1
+▁passionate 1
+tally 1
+▁resist 1
+▁singing 1
+▁load 1
+▁wil 1
+▁steel 1
+▁austrian 1
+▁hop 1
+chard 1
+▁earned 1
+▁genre 1
+▁fate 1
+▁earn 1
+▁deposit 1
+▁solo 1
+▁alcohol 1
+▁exercise 1
+▁exploit 1
+▁intelligence 1
+▁journey 1
+▁marguerite 1
+▁soccer 1
+▁sovereign 1
+▁website 1
+▁narrow 1
+▁struck 1
+▁plastic 1
+▁charming 1
+▁preced 1
+▁shared 1
+▁rennes 1
+▁client 1
+▁pond 1
+▁flour 1
+▁angle 1
+▁theatre 1
+▁exile 1
+lation 1
+▁temper 1
+▁delight 1
+▁porte 1
+▁significantly 1
+▁launch 1
+▁jar 1
+lav 1
+▁scor 1
+▁listened 1
+▁busy 1
+▁combat 1
+▁headquarter 1
+▁interpret 1
+wood 1
+▁michele 1
+▁advice 1
+▁fulfill 1
+▁spokesperson 1
+▁sponsor 1
+▁permanent 1
+▁bronze 1
+hard 1
+▁irish 1
+▁romain 1
+▁nucle 1
+▁florence 1
+▁pop 1
+▁scale 1
+▁sitting 1
+▁berthe 1
+▁berger 1
+▁quar 1
+▁chain 1
+▁existing 1
+▁insist 1
+long 1
+vaux 1
+▁historic 1
+▁scho 1
+▁obvious 1
+▁expected 1
+▁continuous 1
+▁inspiration 1
+▁communist 1
+▁uniform 1
+▁swear 1
+▁gallery 1
+▁split 1
+▁proposition 1
+▁underground 1
+▁mah 1
+▁dragon 1
+▁quay 1
+▁appreciated 1
+▁convinced 1
+▁bigger 1
+▁crew 1
+▁duty 1
+ogen 1
+▁graduate 1
+aison 1
+win 1
+▁map 1
+▁profession 1
+▁sud 1
+▁treated 1
+▁mode 1
+▁dimension 1
+▁jurisdiction 1
+▁accompany 1
+▁encounter 1
+▁mediterranean 1
+▁pioneer 1
+▁respond 1
+▁thirsty 1
+▁removal 1
+▁revenge 1
+▁vienna 1
+▁mutual 1
+▁annual 1
+▁suburb 1
+▁instance 1
+▁hiding 1
+▁collaboration 1
+▁spending 1
+▁habit 1
+▁bullet 1
+▁guest 1
+▁bré 1
+▁distinguished 1
+▁pack 1
+▁rarely 1
+▁allée 1
+▁log 1
+▁mask 1
+▁promote 1
+▁thrown 1
+▁giro 1
+▁arrang 1
+▁rav 1
+zon 1
+▁centuries 1
+▁chassaigne 1
+▁execution 1
+▁grandfather 1
+▁limoges 1
+▁artillery 1
+▁january 1
+▁tribute 1
+▁acquire 1
+▁arthur 1
+▁senior 1
+▁pleasant 1
+▁teeth 1
+▁shirt 1
+▁yann 1
+▁employee 1
+▁detail 1
+▁explained 1
+hl 1
+▁praise 1
+ada 1
+▁gaze 1
+oir 1
+▁promoted 1
+▁cub 1
+▁mende 1
+▁accused 1
+▁nord 1
+▁cout 1
+▁extract 1
+▁vast 1
+▁egypt 1
+▁protest 1
+▁communication 1
+▁context 1
+▁impression 1
+▁substitute 1
+▁technology 1
+▁merchant 1
+▁integrate 1
+▁interrupted 1
+▁cultiva 1
+▁dismiss 1
+▁cabinet 1
+▁senator 1
+▁basket 1
+lemish 1
+▁false 1
+▁radical 1
+▁sacred 1
+▁toulon 1
+▁impose 1
+▁strengthen 1
+▁princess 1
+▁resumed 1
+hel 1
+usse 1
+▁cake 1
+▁laid 1
+▁mala 1
+▁kra 1
+▁poul 1
+▁cere 1
+▁burned 1
+ably 1
+▁breakfast 1
+▁deputies 1
+▁enjolras 1
+▁fireplace 1
+▁portugal 1
+▁uncertain 1
+▁audience 1
+▁arrangement 1
+▁digital 1
+▁birthday 1
+▁restoration 1
+▁lectur 1
+▁wander 1
+▁strict 1
+▁agency 1
+▁shaking 1
+▁heaven 1
+▁restrict 1
+▁appointment 1
+▁friendship 1
+▁broad 1
+lene 1
+▁highway 1
+▁constitutional 1
+▁vain 1
+▁setting 1
+▁locality 1
+bru 1
+▁adam 1
+▁adapted 1
+▁wounded 1
+▁describe 1
+▁reduced 1
+▁finishes 1
+▁statement 1
+▁vitr 1
+▁gru 1
+liard 1
+▁shout 1
+▁unfortunate 1
+▁reflect 1
+known 1
+▁react 1
+▁behavior 1
+▁category 1
+▁concrete 1
+▁expedition 1
+▁hungarian 1
+▁infantry 1
+▁sébastien 1
+▁clermont 1
+▁identity 1
+▁ordinary 1
+▁rhythm 1
+▁component 1
+▁zealand 1
+▁explore 1
+▁chicken 1
+▁ninth 1
+▁alliance 1
+▁isolated 1
+▁copy 1
+▁transferred 1
+▁shook 1
+▁ireland 1
+▁household 1
+▁highlight 1
+▁democratic 1
+▁geneva 1
+maker 1
+▁vienne 1
+▁preparation 1
+▁plot 1
+▁norman 1
+▁rival 1
+▁plateau 1
+▁sigh 1
+▁oak 1
+▁emerge 1
+▁advisor 1
+▁bark 1
+rick 1
+▁deeply 1
+▁determine 1
+▁imagine 1
+omb 1
+▁petite 1
+itude 1
+dres 1
+▁aircraft 1
+▁ambassador 1
+▁congratulate 1
+▁guarantee 1
+▁increasing 1
+▁moscow 1
+▁reassure 1
+▁struggle 1
+▁ceremony 1
+▁concentrat 1
+▁investment 1
+▁smooth 1
+▁architectural 1
+▁pupil 1
+▁forehead 1
+▁guinea 1
+▁dollar 1
+▁translation 1
+▁dozen 1
+▁summit 1
+▁hesitate 1
+▁consum 1
+▁lighthouse 1
+▁reims 1
+▁otto 1
+▁mandate 1
+▁beauty 1
+▁candle 1
+▁armand 1
+▁countryside 1
+▁manufacturer 1
+▁anyway 1
+▁coins 1
+vé 1
+assi 1
+▁eaten 1
+▁determined 1
+lau 1
+▁widely 1
+▁exit 1
+gie 1
+▁pala 1
+▁augustin 1
+▁destroy 1
+▁handle 1
+▁lamb 1
+▁maker 1
+▁cloth 1
+▁pharmac 1
+▁academic 1
+▁delphine 1
+▁medieval 1
+▁practical 1
+▁renaissance 1
+▁surrender 1
+▁universe 1
+▁villeneuve 1
+▁turkish 1
+▁violence 1
+▁bourgeois 1
+▁avignon 1
+▁curtain 1
+▁physi 1
+▁gothic 1
+▁martial 1
+▁altar 1
+▁depict 1
+▁cycle 1
+▁sculptor 1
+▁cheek 1
+▁hungry 1
+▁provence 1
+▁vision 1
+▁monster 1
+▁lying 1
+▁cré 1
+▁suspended 1
+hood 1
+▁slid 1
+▁characteriz 1
+▁fake 1
+▁auto 1
+▁baker 1
+▁maz 1
+▁blame 1
+▁alter 1
+▁format 1
+▁granted 1
+▁apply 1
+▁convince 1
+▁kh 1
+▁comfort 1
+▁analysis 1
+▁confused 1
+▁exclaimed 1
+▁illustrate 1
+▁montpellier 1
+▁chicago 1
+▁sustain 1
+▁delegate 1
+▁foreground 1
+▁error 1
+▁rhone 1
+▁estimate 1
+▁dancing 1
+▁montreal 1
+▁benoit 1
+▁etienne 1
+▁confront 1
+▁amiens 1
+▁informed 1
+hold 1
+▁ferrand 1
+▁hired 1
+champ 1
+▁collective 1
+▁fond 1
+▁arab 1
+▁odd 1
+▁complicated 1
+▁hél 1
+▁tré 1
+▁truck 1
+▁clock 1
+▁defined 1
+▁jim 1
+▁northwest 1
+▁bob 1
+▁faster 1
+▁pole 1
+éné 1
+▁gh 1
+▁constitution 1
+▁brittany 1
+▁chimney 1
+▁confidence 1
+▁midfielder 1
+▁vegetable 1
+▁intervention 1
+▁scandal 1
+▁throat 1
+▁recipe 1
+▁sweat 1
+▁violin 1
+▁intention 1
+▁ump 1
+▁retreat 1
+▁multiple 1
+▁marshal 1
+▁predict 1
+▁classification 1
+▁decade 1
+▁gaston 1
+▁dying 1
+▁pressure 1
+▁debt 1
+▁trois 1
+▁formula 1
+▁shore 1
+▁vend 1
+▁chen 1
+▁chapelle 1
+▁truly 1
+▁feature 1
+▁indicated 1
+▁palm 1
+▁jeanne 1
+▁barre 1
+▁santa 1
+▁joke 1
+ñ 1
+▁arriving 1
+▁circular 1
+▁discipline 1
+▁distribution 1
+▁engaged 1
+▁evidence 1
+▁fundamental 1
+▁pacific 1
+▁treasure 1
+▁volunteer 1
+▁dialect 1
+▁interview 1
+▁indonesia 1
+▁poisson 1
+▁swedish 1
+▁allotment 1
+▁bodies 1
+▁stamp 1
+▁regiment 1
+▁keyboard 1
+▁clerk 1
+▁jesus 1
+▁specifically 1
+▁scored 1
+▁initiative 1
+▁tight 1
+▁crush 1
+▁seized 1
+▁valid 1
+▁brick 1
+▁reache 1
+▁burst 1
+▁salon 1
+▁earl 1
+▁q 1
+▁printer 1
+▁horn 1
+▁occupie 1
+▁riv 1
+▁colombia 1
+▁precise 1
+shire 1
+▁math 1
+▁confident 1
+▁archbishop 1
+▁february 1
+▁héricourt 1
+▁installation 1
+▁suicide 1
+▁compromis 1
+▁quimper 1
+▁sexual 1
+▁expense 1
+▁fraysse 1
+▁random 1
+▁superior 1
+▁narrat 1
+▁edward 1
+▁remarks 1
+▁forgive 1
+▁script 1
+▁curious 1
+▁alban 1
+▁basse 1
+▁retired 1
+▁captured 1
+asi 1
+▁crack 1
+▁curve 1
+▁fog 1
+▁wishes 1
+▁businesses 1
+▁undergo 1
+illac 1
+▁define 1
+▁trap 1
+▁kon 1
+▁punish 1
+▁circuit 1
+▁contribution 1
+▁definitely 1
+▁everybody 1
+▁possibility 1
+▁provincial 1
+▁saturday 1
+▁hockey 1
+▁purchase 1
+▁communal 1
+▁attitude 1
+▁gaudin 1
+▁christmas 1
+▁planning 1
+▁instant 1
+▁médor 1
+▁tip 1
+▁jazz 1
+▁assume 1
+▁jacket 1
+▁sheep 1
+▁vieux 1
+▁fox 1
+▁approved 1
+▁pants 1
+▁exhausted 1
+▁observed 1
+▁whatever 1
+▁bitter 1
+▁comedy 1
+▁metro 1
+▁trick 1
+▁grace 1
+laine 1
+icle 1
+▁franco 1
+▁exceptional 1
+▁electro 1
+▁juan 1
+▁frequent 1
+▁gué 1
+▁marrie 1
+▁teach 1
+▁apprentice 1
+▁recognition 1
+▁rectified 1
+▁tunnel 1
+▁ancestor 1
+▁cavalry 1
+▁thierry 1
+▁collapse 1
+▁torture 1
+▁moderate 1
+▁nancy 1
+▁accent 1
+▁controlled 1
+▁fiction 1
+▁runner 1
+borough 1
+▁crater 1
+▁grandson 1
+▁robin 1
+▁acid 1
+aternal 1
+▁brain 1
+▁abandon 1
+▁arg 1
+▁vary 1
+▁recover 1
+▁kur 1
+▁putt 1
+igné 1
+▁efficient 1
+▁kitchen 1
+▁mathematics 1
+▁napoleon 1
+▁nervous 1
+▁renowned 1
+▁teenager 1
+▁rapidly 1
+▁miracle 1
+▁tunisia 1
+▁amateur 1
+▁baroque 1
+▁isabelle 1
+▁prosper 1
+▁undertake 1
+▁bertrand 1
+▁ignore 1
+▁faubourg 1
+▁argentina 1
+▁clair 1
+▁nimes 1
+▁chairman 1
+▁rub 1
+ienne 1
+▁invade 1
+▁youngest 1
+troph 1
+▁asset 1
+▁tele 1
+▁publishe 1
+▁eric 1
+▁robi 1
+▁ideal 1
+▁burg 1
+▁spin 1
+▁involve 1
+▁hau 1
+▁assembl 1
+▁actual 1
+▁libr 1
+▁northeast 1
+▁accoyer 1
+▁armchair 1
+▁diameter 1
+▁legislative 1
+▁mathilde 1
+▁satisfaction 1
+▁stéphan 1
+▁wisdom 1
+▁counsel 1
+▁ambition 1
+▁license 1
+▁turkey 1
+▁exclusively 1
+▁troyes 1
+▁ugly 1
+▁archives 1
+▁distant 1
+▁jacquelin 1
+smith 1
+▁mood 1
+▁generate 1
+▁claudi 1
+▁bunch 1
+▁humid 1
+▁robinet 1
+▁prim 1
+▁distributed 1
+▁spite 1
+▁stake 1
+▁gorge 1
+▁connection 1
+▁rider 1
+▁rouen 1
+éré 1
+▁romania 1
+▁prose 1
+ignon 1
+▁distinguish 1
+▁abd 1
+▁rang 1
+▁observe 1
+▁exact 1
+▁enormous 1
+▁geographical 1
+í 1
+▁chocolate 1
+▁deficit 1
+▁definition 1
+▁delegation 1
+▁equivalent 1
+▁horrible 1
+▁threshold 1
+▁unanimous 1
+▁undoubtedly 1
+▁jersey 1
+▁friday 1
+▁easier 1
+▁eldest 1
+▁nurse 1
+▁navy 1
+▁inherit 1
+▁jewel 1
+▁boston 1
+▁josé 1
+▁véron 1
+▁henriette 1
+▁grabb 1
+▁interpretation 1
+▁massive 1
+▁tiger 1
+▁choir 1
+▁arabic 1
+▁pride 1
+▁varies 1
+▁frequently 1
+phor 1
+▁fuel 1
+▁motion 1
+▁clip 1
+▁cruel 1
+▁chu 1
+▁dino 1
+▁pipe 1
+▁jail 1
+pont 1
+fortunately 1
+▁prefect 1
+▁artificial 1
+▁luxembourg 1
+▁obligation 1
+▁pupponi 1
+▁alphabet 1
+▁innovation 1
+▁promotion 1
+▁proposing 1
+▁diagnos 1
+▁insurance 1
+▁statistic 1
+▁athlete 1
+▁prestig 1
+▁disaster 1
+▁squad 1
+▁funeral 1
+▁glory 1
+▁scope 1
+▁blessed 1
+▁sweden 1
+▁medium 1
+▁goddess 1
+▁rabbi 1
+▁distract 1
+▁depth 1
+▁gentleman 1
+▁meantime 1
+▁convict 1
+▁cancer 1
+▁graphic 1
+▁stuck 1
+▁chapter 1
+▁suitable 1
+▁ease 1
+▁employment 1
+▁kha 1
+▁regent 1
+▁hervé 1
+▁recovery 1
+▁tender 1
+▁marcelin 1
+▁brian 1
+▁mechanical 1
+▁publisher 1
+▁chab 1
+▁manufacture 1
+finals 1
+▁confess 1
+▁circumstances 1
+▁executed 1
+▁faculty 1
+▁félix 1
+▁imprisoned 1
+▁jérome 1
+▁simultaneously 1
+▁vietnam 1
+▁wednesday 1
+▁scotland 1
+▁texas 1
+▁hidden 1
+▁dijon 1
+▁threatening 1
+▁threatened 1
+▁flood 1
+▁domestic 1
+▁creating 1
+▁okay 1
+▁diamond 1
+▁overall 1
+▁winning 1
+▁baroness 1
+▁fortune 1
+▁preferred 1
+▁package 1
+▁adaptation 1
+▁riviere 1
+▁collaborator 1
+▁vault 1
+▁bruno 1
+▁plou 1
+▁authorized 1
+▁genus 1
+water 1
+▁woke 1
+▁rubb 1
+bridge 1
+▁absorb 1
+▁aggressive 1
+▁linguistic 1
+▁response 1
+▁peninsula 1
+▁smiling 1
+▁balcony 1
+▁czech 1
+▁diverse 1
+▁liquid 1
+▁madeleine 1
+▁throne 1
+▁applies 1
+▁sophie 1
+▁siege 1
+▁laboratory 1
+▁dolez 1
+▁vachon 1
+▁quarrel 1
+▁regain 1
+▁francais 1
+▁polite 1
+▁persist 1
+▁egyptian 1
+▁confession 1
+▁toilet 1
+▁revolutionary 1
+▁bench 1
+▁collaborate 1
+uvre 1
+▁glu 1
+▁stranger 1
+▁hunt 1
+▁loop 1
+▁arrest 1
+▁logical 1
+rnac 1
+▁bush 1
+▁cavrois 1
+▁conservatory 1
+▁emmanuel 1
+▁obstacle 1
+▁subdivision 1
+▁universal 1
+▁édouard 1
+▁alternative 1
+▁desperate 1
+▁muslim 1
+▁supply 1
+▁tobacco 1
+▁switch 1
+▁relax 1
+ibility 1
+▁stress 1
+▁loyal 1
+▁illness 1
+▁rescue 1
+▁folk 1
+▁comfortable 1
+▁admitted 1
+▁darkness 1
+▁romantic 1
+▁rough 1
+▁giant 1
+▁archeolog 1
+▁greg 1
+▁milan 1
+down 1
+amba 1
+▁habitat 1
+piece 1
+▁angouleme 1
+▁barricade 1
+▁external 1
+▁genevieve 1
+▁identify 1
+▁primarily 1
+▁reception 1
+▁shudder 1
+▁supreme 1
+▁clarify 1
+▁condemn 1
+▁perpignan 1
+▁enlighten 1
+▁gosselin 1
+▁mysterious 1
+▁powder 1
+▁assigned 1
+▁atlantic 1
+▁fierce 1
+▁resort 1
+▁worst 1
+▁twist 1
+▁costume 1
+▁saone 1
+▁fortress 1
+▁jacob 1
+▁intense 1
+▁lambert 1
+▁nerve 1
+▁johann 1
+▁betray 1
+▁abbé 1
+▁hammer 1
+▁clause 1
+▁victories 1
+▁charny 1
+▁irre 1
+▁borgo 1
+▁fever 1
+▁marthe 1
+▁spell 1
+▁chor 1
+▁bak 1
+▁julie 1
+▁exception 1
+▁alexandr 1
+▁christin 1
+metric 1
+▁argu 1
+quote 1
+worth 1
+▁adjacent 1
+▁calculation 1
+▁elegant 1
+▁elsewhere 1
+▁incredible 1
+▁scattered 1
+▁suspension 1
+▁compensation 1
+▁montauban 1
+▁pillars 1
+▁revolt 1
+▁rugby 1
+▁provoke 1
+▁flesh 1
+▁partially 1
+▁cadio 1
+▁harvest 1
+▁automatic 1
+▁cuba 1
+▁prix 1
+▁cheap 1
+▁consistent 1
+▁alex 1
+▁cream 1
+▁brush 1
+▁considerable 1
+▁femini 1
+▁sink 1
+▁reasonable 1
+▁humor 1
+▁elite 1
+position 1
+ouji 1
+▁bowl 1
+▁refus 1
+▁southeast 1
+▁georgia 1
+▁immediate 1
+▁benjamin 1
+▁conservative 1
+▁extensive 1
+▁fountain 1
+▁grenoble 1
+▁inhabited 1
+▁intelligent 1
+▁populated 1
+▁portuguese 1
+▁sanctuary 1
+▁stomach 1
+▁temporary 1
+▁ultimate 1
+▁cinieri 1
+▁compound 1
+▁corridor 1
+▁knife 1
+▁commitment 1
+▁shepherd 1
+▁neutral 1
+▁abroad 1
+▁schedule 1
+▁crystal 1
+▁ornament 1
+▁dispute 1
+▁adolph 1
+▁greece 1
+▁ceiling 1
+▁vocal 1
+▁comedian 1
+▁cheese 1
+▁gilbert 1
+▁boots 1
+▁migrat 1
+▁secur 1
+▁township 1
+▁striker 1
+▁parent 1
+▁hind 1
+▁sergi 1
+▁devil 1
+▁cheat 1
+▁wound 1
+▁restore 1
+▁baz 1
+▁investigat 1
+▁affiliat 1
+▁contemplat 1
+▁dynasty 1
+▁reconstruct 1
+▁revenue 1
+▁somewhere 1
+▁spectator 1
+▁virtual 1
+▁enclos 1
+▁enemies 1
+▁maintenance 1
+▁qualify 1
+▁fashion 1
+▁profile 1
+▁overseas 1
+▁martyr 1
+▁monarch 1
+▁mystery 1
+▁outfit 1
+▁awful 1
+▁conquer 1
+▁deserve 1
+▁sample 1
+▁scottish 1
+▁operator 1
+otype 1
+▁racing 1
+▁victorious 1
+▁punishment 1
+▁strategic 1
+▁export 1
+▁abbe 1
+▁convent 1
+▁hér 1
+▁dragg 1
+▁hung 1
+▁gab 1
+feld 1
+▁porch 1
+▁distinct 1
+willer 1
+▁attorney 1
+▁categories 1
+▁dynamic 1
+▁excessive 1
+▁exclud 1
+▁jahoua 1
+▁pomereux 1
+▁submitted 1
+▁universities 1
+▁contradict 1
+▁fragile 1
+▁penalty 1
+▁underscore 1
+▁worship 1
+▁activist 1
+artagnan 1
+▁communicate 1
+▁unlike 1
+corporate 1
+▁fossil 1
+▁delicate 1
+▁vocation 1
+▁terminal 1
+▁dread 1
+▁hetzel 1
+▁sprint 1
+▁niort 1
+▁absent 1
+▁flies 1
+fold 1
+▁duties 1
+▁cosette 1
+▁safety 1
+▁confusion 1
+▁blanche 1
+▁handsome 1
+▁lordship 1
+▁operating 1
+▁rumor 1
+▁politician 1
+▁criticism 1
+▁guitarist 1
+▁liege 1
+▁possession 1
+cultur 1
+ainville 1
+stitution 1
+▁couch 1
+cute 1
+asso 1
+▁oppose 1
+▁archipelago 1
+▁celebration 1
+▁controversial 1
+▁controversy 1
+▁eponym 1
+▁legislation 1
+▁misfortune 1
+▁norwegian 1
+▁pilgrim 1
+▁satisfy 1
+▁yvonne 1
+▁heavily 1
+▁exposed 1
+▁guilty 1
+▁memorial 1
+▁analyze 1
+▁munich 1
+▁nominated 1
+▁commerce 1
+▁modification 1
+▁saxon 1
+▁contrast 1
+▁carriage 1
+▁antonio 1
+▁lantern 1
+▁madrid 1
+▁raphael 1
+▁shield 1
+▁poison 1
+▁generous 1
+▁carfor 1
+▁honey 1
+▁visual 1
+▁ashamed 1
+▁dishes 1
+▁borrow 1
+▁wonderful 1
+▁chaumont 1
+▁renaud 1
+▁holland 1
+▁coordination 1
+▁elect 1
+▁alexis 1
+▁canvas 1
+▁raoul 1
+▁montjo 1
+▁gallo 1
+lithic 1
+▁vaux 1
+▁nazi 1
+head 1
+took 1
+▁wage 1
+▁karl 1
+▁hameau 1
+▁select 1
+▁preserve 1
+▁advocate 1
+▁explosion 1
+▁maximum 1
+▁monsieur 1
+▁satellite 1
+▁scenario 1
+▁smoking 1
+▁sufficient 1
+▁territorial 1
+▁twentieth 1
+▁unhappy 1
+▁midnight 1
+▁washington 1
+▁maritime 1
+▁overlook 1
+▁electoral 1
+▁fragment 1
+▁comparison 1
+▁precious 1
+▁tokyo 1
+▁columbia 1
+▁modify 1
+▁variation 1
+▁glaci 1
+▁facial 1
+▁camille 1
+▁compensate 1
+▁profit 1
+▁smart 1
+▁lodge 1
+▁educate 1
+▁hungary 1
+▁entertainment 1
+▁rochelle 1
+▁massif 1
+▁seminar 1
+▁savings 1
+▁savary 1
+▁weird 1
+▁alert 1
+▁tolerate 1
+▁maison 1
+explo 1
+▁mechanic 1
+▁declare 1
+▁evil 1
+▁consecutive 1
+▁engagement 1
+▁enthusiasm 1
+▁extraordinary 1
+▁nautilus 1
+▁neglect 1
+▁participation 1
+▁plenty 1
+▁reappear 1
+▁unexpected 1
+▁versailles 1
+▁virtue 1
+▁volcano 1
+▁morocco 1
+▁orléans 1
+▁complement 1
+▁despair 1
+▁bouvard 1
+▁horror 1
+▁sorrow 1
+▁harbor 1
+▁perfume 1
+▁regulate 1
+▁stopping 1
+▁divine 1
+▁entry 1
+▁louvre 1
+▁craft 1
+▁bonneton 1
+▁assault 1
+▁forgot 1
+▁burial 1
+▁virginie 1
+▁drag 1
+▁socks 1
+▁quote 1
+▁recogniz 1
+▁urge 1
+▁adapt 1
+▁jacqu 1
+agnac 1
+▁acknowledg 1
+▁affluent 1
+▁algorithm 1
+▁arrondissement 1
+▁aurélie 1
+▁demonstrate 1
+▁dieudonné 1
+▁dissolve 1
+▁engraving 1
+▁espérance 1
+▁hypothesis 1
+▁louwagie 1
+▁massonneau 1
+▁molecule 1
+▁monseigneur 1
+▁railroad 1
+▁rebuilt 1
+▁suppress 1
+▁delaunay 1
+▁sleeve 1
+▁solidarity 1
+▁unusual 1
+▁worries 1
+▁pattern 1
+▁stabiliz 1
+▁fabric 1
+boeuf 1
+▁verdun 1
+▁injury 1
+▁charente 1
+▁denmark 1
+▁repress 1
+▁telephone 1
+▁adrien 1
+▁naples 1
+▁porthos 1
+▁behave 1
+▁introduction 1
+▁florida 1
+▁aubin 1
+▁normandy 1
+▁trend 1
+▁jardin 1
+▁intervene 1
+▁fortified 1
+▁painful 1
+▁tennis 1
+▁distinction 1
+pierre 1
+roving 1
+▁abundant 1
+▁appropriate 1
+▁bicycle 1
+▁bilingual 1
+▁coincide 1
+▁erected 1
+▁facilities 1
+▁illustration 1
+▁lydie 1
+▁mirror 1
+▁philosopher 1
+▁pregnant 1
+▁reverend 1
+▁senegal 1
+▁struggling 1
+▁sympath 1
+▁creature 1
+▁download 1
+▁giraud 1
+▁inaugura 1
+▁description 1
+▁destination 1
+▁vercamer 1
+▁orleans 1
+▁saddier 1
+▁massacre 1
+▁amazon 1
+▁thunder 1
+▁vacation 1
+▁serving 1
+▁skirt 1
+▁vosges 1
+▁spouse 1
+▁comte 1
+▁physics 1
+▁genevard 1
+▁carpet 1
+▁valentin 1
+▁jews 1
+▁spark 1
+▁remire 1
+▁anglo 1
+▁cluster 1
+▁faithful 1
+▁vertical 1
+▁vital 1
+▁peaceful 1
+▁tyran 1
+▁drill 1
+▁weakness 1
+▁creator 1
+▁cancel 1
+▁alice 1
+▁theories 1
+▁reviv 1
+▁brunet 1
+▁elephant 1
+uchet 1
+blanc 1
+▁capture 1
+▁resume 1
+▁scott 1
+▁accumulat 1
+▁affirm 1
+▁bachelor 1
+▁brigitte 1
+▁byzantine 1
+▁cyprien 1
+▁defensive 1
+▁ferdinand 1
+▁folleville 1
+▁geoffroy 1
+▁henceforth 1
+▁limestone 1
+▁transparency 1
+▁accelerat 1
+▁grandmother 1
+▁thereafter 1
+▁lorient 1
+▁renamed 1
+▁monitor 1
+▁école 1
+▁navigat 1
+▁laurence 1
+▁assisted 1
+▁inscription 1
+▁orbit 1
+▁theorem 1
+▁disappearance 1
+▁brewer 1
+▁aisne 1
+▁cease 1
+▁prun 1
+front 1
+▁gamb 1
+▁imitat 1
+▁retire 1
+▁appreciate 1
+▁divide 1
+▁plea 1
+▁harmoni 1
+stadt 1
+script 1
+▁entertain 1
+▁apologize 1
+▁autonomous 1
+▁bulgaria 1
+▁coronado 1
+▁embassy 1
+▁fantasy 1
+▁hurried 1
+▁immigrant 1
+▁madagascar 1
+▁mistress 1
+▁precision 1
+▁suspicious 1
+▁penetrat 1
+▁tragedy 1
+▁applaud 1
+▁clément 1
+▁embrace 1
+▁hilaire 1
+▁proportion 1
+▁disciple 1
+▁substance 1
+▁alsace 1
+▁transition 1
+▁trophy 1
+▁anthrop 1
+▁cayenne 1
+▁animated 1
+▁mobile 1
+▁stolen 1
+beuf 1
+▁belief 1
+▁caroline 1
+▁bavaria 1
+▁colony 1
+▁supervise 1
+▁varied 1
+▁inviting 1
+▁glove 1
+▁noisy 1
+▁occurred 1
+▁toxic 1
+▁horizontal 1
+▁infectio 1
+camp 1
+▁accord 1
+▁aubr 1
+▁absolute 1
+▁descend 1
+▁import 1
+▁illuminat 1
+responsibilities 1
+▁accomplish 1
+▁alexander 1
+▁aristocrat 1
+▁brilliant 1
+▁corporation 1
+▁expansion 1
+▁gloomy 1
+▁intellectual 1
+▁mammal 1
+▁phenomena 1
+▁poitiers 1
+▁privilege 1
+▁pécuchet 1
+▁régulus 1
+▁striking 1
+▁asleep 1
+▁fontenay 1
+▁trigger 1
+▁criteria 1
+▁offensive 1
+▁rabault 1
+▁prevail 1
+▁cheerful 1
+▁samuel 1
+▁censor 1
+▁mixture 1
+▁memories 1
+▁ghost 1
+▁patrol 1
+▁venice 1
+▁aurillac 1
+▁calculate 1
+▁cameroon 1
+▁managing 1
+▁dolores 1
+▁spoon 1
+▁terrace 1
+▁framework 1
+▁purple 1
+▁prud 1
+▁crypt 1
+▁edme 1
+▁nasie 1
+communica 1
+born 1
+▁mineral 1
+▁wrap 1
+▁necessar 1
+▁coordinat 1
+▁assumption 1
+▁athletic 1
+▁bathroom 1
+▁conceived 1
+▁financing 1
+▁lefebvre 1
+▁legitimate 1
+▁observatory 1
+▁suzanne 1
+▁sylvain 1
+▁transmission 1
+▁disagree 1
+▁disposal 1
+▁lorraine 1
+▁kirsch 1
+▁inevitabl 1
+▁meadow 1
+▁patriot 1
+▁squeez 1
+▁brigade 1
+▁malaysia 1
+▁snake 1
+▁cartoon 1
+▁gaulle 1
+▁granite 1
+▁veteran 1
+▁weekend 1
+▁niece 1
+▁spider 1
+▁corsica 1
+▁gently 1
+▁moselle 1
+▁blank 1
+▁disposition 1
+▁sultan 1
+▁juliette 1
+▁consent 1
+▁faint 1
+owski 1
+limard 1
+▁ollier 1
+fusion 1
+▁update 1
+▁translate 1
+▁govern 1
+▁swallow 1
+▁accommodation 1
+▁archaeological 1
+▁biodiversity 1
+▁congregation 1
+▁constituen 1
+▁efficiency 1
+▁familiar 1
+▁flaviana 1
+▁jesuit 1
+▁maneuver 1
+▁montaudoin 1
+▁potatoes 1
+▁predominant 1
+▁scientist 1
+▁shrugged 1
+▁societies 1
+▁swimming 1
+▁unbearable 1
+▁decorative 1
+▁garrison 1
+▁instinct 1
+▁besancon 1
+▁cycling 1
+▁maltenu 1
+▁shrub 1
+▁sauveur 1
+giovanni 1
+▁gospel 1
+▁chemical 1
+▁producing 1
+▁shivering 1
+▁devotion 1
+▁petersburg 1
+▁lurton 1
+▁flew 1
+▁variant 1
+▁gulf 1
+▁compact 1
+▁surgery 1
+▁aramis 1
+▁quadr 1
+▁outcome 1
+neath 1
+▁cotton 1
+▁indication 1
+▁toronto 1
+▁considerably 1
+▁foresee 1
+▁bump 1
+▁dropped 1
+▁renault 1
+▁slept 1
+▁pension 1
+▁scheme 1
+▁thief 1
+▁blond 1
+▁assert 1
+▁provid 1
+▁weigh 1
+▁measur 1
+establish 1
+woman 1
+á 1
+▁barcelona 1
+▁calendar 1
+▁consecrat 1
+▁innovative 1
+▁occupy 1
+▁perspective 1
+▁supplement 1
+▁thursday 1
+▁vineyard 1
+▁yield 1
+▁abolish 1
+▁consumption 1
+▁eternal 1
+▁evaluation 1
+▁oxford 1
+▁uttered 1
+▁diversity 1
+▁sequence 1
+▁submarine 1
+▁podium 1
+▁internship 1
+▁kourou 1
+▁paradise 1
+▁glance 1
+▁essay 1
+▁pistol 1
+▁strangle 1
+▁overcome 1
+▁stroll 1
+▁shopping 1
+▁alternate 1
+▁changing 1
+▁deported 1
+▁mongol 1
+▁peyre 1
+▁sorgue 1
+▁automatically 1
+▁cancelled 1
+▁hastily 1
+▁guérin 1
+judicial 1
+▁glad 1
+▁fisc 1
+▁pouill 1
+▁chris 1
+▁whisper 1
+▁bathilde 1
+▁clergy 1
+▁conclusion 1
+▁cooperative 1
+▁cyclist 1
+▁demonstration 1
+▁ecuador 1
+▁elizabeth 1
+▁enforce 1
+▁exterior 1
+▁hierarchy 1
+▁hippolyte 1
+▁inferior 1
+▁infinite 1
+▁mézieres 1
+▁nickname 1
+▁participating 1
+▁rebellion 1
+▁rectangular 1
+▁regroup 1
+▁antiquity 1
+▁burgundy 1
+▁embark 1
+▁homosexual 1
+▁excited 1
+▁invisible 1
+▁ribbon 1
+▁invitation 1
+▁adorned 1
+▁variable 1
+▁ecological 1
+▁stroke 1
+▁misery 1
+▁yelled 1
+▁orphan 1
+▁cahors 1
+▁césar 1
+▁pounds 1
+▁mathieu 1
+▁cognac 1
+▁withdrew 1
+▁banned 1
+▁extinct 1
+▁carved 1
+▁alarm 1
+▁tibet 1
+▁awake 1
+▁irrita 1
+fleur 1
+terrain 1
+eutic 1
+ergic 1
+▁exhaust 1
+▁gentle 1
+▁profound 1
+▁slaughter 1
+▁tremendous 1
+▁uruguay 1
+▁advertising 1
+▁butterfly 1
+▁cartridge 1
+▁circum 1
+▁conscience 1
+▁conspiracy 1
+▁depression 1
+▁discourage 1
+▁infrastructure 1
+▁maréchal 1
+▁orthodox 1
+▁reservation 1
+▁strategy 1
+▁absurd 1
+▁theoretical 1
+▁valjean 1
+▁reproach 1
+▁transaction 1
+▁disorder 1
+▁crucial 1
+▁reduction 1
+▁skull 1
+▁manosque 1
+▁bretagne 1
+▁sketch 1
+▁combination 1
+▁harmony 1
+▁syrian 1
+▁drank 1
+▁goriot 1
+▁vallée 1
+▁tragic 1
+▁donkey 1
+▁extent 1
+▁tackle 1
+▁refine 1
+▁thermal 1
+▁diving 1
+▁fresco 1
+▁urgent 1
+▁croatia 1
+itarian 1
+▁agitat 1
+▁pomp 1
+flex 1
+▁brutal 1
+▁democrat 1
+▁accommodate 1
+▁acquisition 1
+▁aluminum 1
+▁comrade 1
+▁curiosity 1
+▁destruction 1
+▁disappointed 1
+▁doctrine 1
+▁exaggerate 1
+▁fauchelevent 1
+▁goalkeeper 1
+▁irregular 1
+▁lithuania 1
+▁metropolitan 1
+▁philadelphia 1
+▁residency 1
+▁structural 1
+▁théâtre 1
+▁tuesday 1
+▁ukraine 1
+▁unacceptable 1
+▁venezuela 1
+▁envelope 1
+▁fascinat 1
+▁miserable 1
+▁protocol 1
+▁tissue 1
+▁edouard 1
+▁manipulat 1
+▁narbonne 1
+▁translator 1
+▁disguise 1
+▁worldwide 1
+▁respectively 1
+▁predator 1
+▁filter 1
+▁protein 1
+▁barbier 1
+▁pianist 1
+▁vieille 1
+▁guéret 1
+▁chronic 1
+▁dessert 1
+▁amazing 1
+▁nightmare 1
+▁exceed 1
+agnol 1
+▁chateauroux 1
+▁flank 1
+▁oath 1
+▁survey 1
+▁deletion 1
+▁emblem 1
+▁collaborat 1
+▁appoint 1
+▁eliminate 1
+western 1
+mouth 1
+▁agglomeration 1
+▁alphonse 1
+▁cooperation 1
+▁devastated 1
+▁ethiopia 1
+▁footsteps 1
+▁félici 1
+▁hollywood 1
+▁meticulous 1
+▁overwhelm 1
+▁pencroff 1
+▁possibilities 1
+▁pseudonym 1
+▁pyramid 1
+▁ridiculous 1
+▁sarkozy 1
+▁suspicion 1
+▁interface 1
+▁mexican 1
+▁postponed 1
+▁prominent 1
+▁psychiatri 1
+▁richelieu 1
+▁cécile 1
+▁kennedy 1
+▁lombard 1
+▁colonial 1
+▁tremble 1
+▁triangle 1
+▁slipped 1
+▁cabaret 1
+▁laundry 1
+▁airplane 1
+▁depriv 1
+▁discreet 1
+▁amended 1
+▁dizier 1
+▁dwelling 1
+▁innocent 1
+▁reunion 1
+▁breeze 1
+▁loving 1
+▁await 1
+▁invention 1
+▁avenge 1
+homme 1
+logy 1
+▁moderniz 1
+speaking 1
+▁georg 1
+▁administrator 1
+▁adolescen 1
+▁anonymous 1
+▁applicable 1
+▁bailiff 1
+▁basilica 1
+▁bonaparte 1
+▁centimeters 1
+▁coalition 1
+▁entrepreneur 1
+▁fluctuat 1
+▁gwynplaine 1
+▁immense 1
+▁magistrate 1
+▁marvejol 1
+▁mistreat 1
+▁necessity 1
+▁olympe 1
+▁opportunities 1
+▁outstanding 1
+▁patriarch 1
+▁prehistoric 1
+▁repertoire 1
+▁reproduction 1
+▁synthesis 1
+▁thailand 1
+▁thirtieth 1
+▁théodore 1
+▁viscount 1
+▁vowels 1
+▁coherence 1
+▁fiancé 1
+▁nymph 1
+▁prussia 1
+▁nobility 1
+▁prompt 1
+▁separation 1
+▁signature 1
+▁arsenal 1
+▁disgrace 1
+▁patience 1
+▁abstract 1
+▁puppet 1
+▁secular 1
+▁resembl 1
+▁frozen 1
+▁marmande 1
+▁celtic 1
+▁dignity 1
+▁disgust 1
+▁eagle 1
+▁junction 1
+▁trunk 1
+▁conseil 1
+▁reliable 1
+▁surname 1
+▁goodbye 1
+▁playoffs 1
+▁warrior 1
+▁ariege 1
+▁nicole 1
+▁bernadet 1
+▁stefan 1
+▁retrac 1
+▁seduce 1
+▁automat 1
+▁chevr 1
+franc 1
+uangan 1
+ignol 1
+▁evident 1
+▁amsterdam 1
+▁aquitaine 1
+▁archdiocese 1
+▁attachment 1
+▁automobile 1
+▁carcassonne 1
+▁emphasize 1
+▁episcopal 1
+▁hesitating 1
+▁industries 1
+▁jerusalem 1
+▁promenade 1
+▁pyrenees 1
+▁reproduce 1
+▁staircase 1
+▁stubborn 1
+▁subsidiary 1
+▁temporarily 1
+▁twelfth 1
+▁upstream 1
+▁volcanic 1
+▁consolidat 1
+▁frighten 1
+▁godfather 1
+▁humiliat 1
+▁illusion 1
+▁opposing 1
+▁abdomen 1
+▁fluently 1
+▁swamp 1
+▁examination 1
+▁charlotte 1
+▁pigeon 1
+▁ritual 1
+▁spiral 1
+▁café 1
+▁iraq 1
+▁margaret 1
+▁whistle 1
+▁worn 1
+▁condom 1
+▁lounge 1
+▁ethic 1
+▁analog 1
+▁replac 1
+eastern 1
+mposing 1
+▁dedicate 1
+▁advertise 1
+▁authentic 1
+▁brunswick 1
+▁configuration 1
+▁consciousness 1
+▁dhuicq 1
+▁eloquence 1
+▁frequency 1
+▁frémilly 1
+▁google 1
+▁granddaughter 1
+▁impatience 1
+▁joséphine 1
+▁languedoc 1
+▁prosecutor 1
+▁slippers 1
+▁supermarket 1
+▁widespread 1
+ó 1
+▁combelaine 1
+▁eyelids 1
+▁pakistan 1
+▁terrasse 1
+▁torrent 1
+▁crossroad 1
+▁premises 1
+▁qualification 1
+▁compliment 1
+▁bolivia 1
+▁chestnut 1
+▁désert 1
+▁sylvi 1
+▁charleville 1
+▁parliamentarian 1
+▁warsaw 1
+▁ghison 1
+▁warehouse 1
+▁fossé 1
+▁dubié 1
+▁defect 1
+▁width 1
+▁eager 1
+▁rotten 1
+▁nigeria 1
+▁turtle 1
+▁scrap 1
+▁approve 1
+▁kilometer 1
+▁horizon 1
+▁catalog 1
+▁exclusiv 1
+▁apostolic 1
+▁ascension 1
+▁aubenas 1
+▁auvergne 1
+▁bibichon 1
+▁caterpillar 1
+▁certificate 1
+▁chateauneuf 1
+▁compatible 1
+▁compulsory 1
+▁deceased 1
+▁demolished 1
+▁douarnenez 1
+▁identification 1
+▁impressive 1
+▁josserand 1
+▁laffitte 1
+▁nathalie 1
+▁octave 1
+▁pedestrian 1
+▁pollution 1
+▁precaution 1
+▁presidency 1
+▁relevant 1
+▁screenwriter 1
+▁stammered 1
+▁synagogue 1
+▁synonym 1
+▁testimony 1
+▁transmitted 1
+▁vaccination 1
+▁villefranche 1
+▁renovation 1
+▁slender 1
+▁soundtrack 1
+▁wolves 1
+▁marathon 1
+▁toponym 1
+▁apostle 1
+▁carhaix 1
+▁drowned 1
+▁illegal 1
+▁policies 1
+▁shareholder 1
+▁wharf 1
+▁allocated 1
+▁sociology 1
+▁tramway 1
+▁bacteri 1
+▁colonies 1
+▁finnish 1
+▁felix 1
+▁jordan 1
+▁dominant 1
+▁musket 1
+▁propell 1
+▁computing 1
+▁flavor 1
+▁sphere 1
+▁gwen 1
+▁pearl 1
+▁terribl 1
+ground 1
+misunderstanding 1
+▁compiegne 1
+▁dauphiné 1
+▁definitive 1
+▁democracy 1
+▁escaping 1
+▁extinguish 1
+▁glenarvan 1
+▁influential 1
+▁keraotred 1
+▁lincoln 1
+▁massachusetts 1
+▁metropolis 1
+▁obliged 1
+▁oxygen 1
+▁paraguay 1
+▁pyrénées 1
+▁subscribe 1
+▁suffocating 1
+▁systematically 1
+▁thorough 1
+▁transparent 1
+▁unpleasant 1
+▁unsuccessful 1
+▁catalonia 1
+▁geometric 1
+▁supervisor 1
+▁adjoin 1
+▁decisive 1
+▁stunned 1
+▁escort 1
+▁gratitude 1
+▁injuries 1
+▁compilation 1
+▁rendered 1
+▁poverty 1
+▁sentiment 1
+▁lobby 1
+▁mandatory 1
+▁fayette 1
+▁wreck 1
+▁gossip 1
+▁auroi 1
+▁default 1
+▁conquest 1
+▁grumbl 1
+▁invasion 1
+▁remote 1
+▁edible 1
+ð 1
+ú 1
+ʾ 1
+. 1
+ā 1
+ʉ 1
+ʻ 1
+þ 1
+̇ 1
+~ 1
+ʿ 1
+! 1
+â 1
+ˋ 1
+ː 1
+̀ 1
+̠ 1
diff --git a/configs/fr-en/tgt_unigram6000/spm_unigram_fr.vocab b/configs/fr-en/tgt_unigram6000/spm_unigram_fr.vocab
new file mode 100644
index 0000000000000000000000000000000000000000..5562767d550ca1ae8906c4329a6ca2f3fb10efcf
--- /dev/null
+++ b/configs/fr-en/tgt_unigram6000/spm_unigram_fr.vocab
@@ -0,0 +1,6000 @@
+ 0
+ 0
+ 0
+ 0
+▁the -2.91814
+s -3.10752
+▁of -3.91695
+▁a -3.9883
+▁to -4.04795
+▁is -4.08027
+▁in -4.09322
+▁and -4.28415
+▁he -4.51887
+' -4.53226
+▁it -4.54676
+▁i -4.77223
+t -4.87671
+▁was -4.95325
+ed -4.98132
+ing -4.98791
+▁this -5.07064
+▁you -5.08862
+▁are -5.1969
+n -5.20731
+▁that -5.23725
+▁one -5.28638
+▁his -5.30104
+▁for -5.32756
+▁hundred -5.33011
+▁on -5.33846
+d -5.36433
+e -5.37676
+▁be -5.48498
+▁with -5.53855
+▁at -5.54279
+▁we -5.57746
+▁two -5.58708
+▁street -5.59683
+▁by -5.65074
+▁she -5.65582
+▁not -5.69203
+y -5.69627
+▁de -5.79706
+▁will -5.80066
+m -5.81877
+ly -5.82274
+▁from -5.83364
+▁have -5.84285
+▁an -5.91028
+▁three -5.95706
+▁they -5.97513
+▁as -6.03094
+▁has -6.03161
+re -6.05524
+er -6.05707
+al -6.10023
+▁number -6.11325
+▁me -6.1438
+▁also -6.18796
+▁there -6.19762
+▁him -6.21481
+▁her -6.2219
+▁were -6.23262
+▁amendment -6.24265
+▁four -6.2465
+▁my -6.24804
+▁thousand -6.26997
+▁do -6.27419
+▁what -6.29529
+▁zero -6.31188
+es -6.34034
+▁five -6.34845
+▁but -6.34852
+▁can -6.36085
+c -6.37418
+▁twenty -6.38129
+▁had -6.38369
+▁all -6.38591
+▁la -6.39014
+▁their -6.39376
+o -6.39729
+r -6.47048
+▁your -6.47626
+▁no -6.51584
+▁seven -6.52383
+▁six -6.5287
+g -6.53252
+▁first -6.57094
+▁would -6.57387
+▁up -6.58809
+▁floor -6.60781
+▁saint -6.60903
+a -6.60971
+▁eight -6.62323
+▁time -6.64027
+▁or -6.64734
+▁thirty -6.65329
+▁like -6.65758
+p -6.66262
+▁these -6.67568
+▁so -6.68027
+▁mr -6.68118
+i -6.68743
+in -6.69191
+▁then -6.69947
+ll -6.70719
+▁support -6.71703
+▁who -6.73696
+▁did -6.74709
+▁its -6.75546
+u -6.75589
+▁road -6.76057
+▁about -6.78515
+▁more -6.81916
+▁le -6.82779
+▁du -6.83695
+▁very -6.84849
+▁nine -6.85628
+▁been -6.86332
+▁forty -6.86716
+▁them -6.87004
+▁only -6.87305
+▁go -6.87321
+▁back -6.88152
+▁after -6.88219
+▁am -6.89766
+le -6.90277
+▁fifty -6.91046
+on -6.915
+▁some -6.91909
+▁other -6.93794
+▁well -6.94316
+▁re -6.95204
+ation -6.96209
+▁now -6.98449
+k -6.99115
+▁out -6.99637
+▁s -6.99843
+▁know -7.0121
+l -7.0299
+en -7.03535
+▁o -7.03819
+ro -7.0465
+el -7.04816
+ch -7.05483
+▁don -7.08109
+f -7.0814
+▁seventy -7.08339
+▁when -7.09526
+▁d -7.10804
+▁new -7.11026
+▁see -7.11075
+▁if -7.11797
+▁sixty -7.11857
+▁dur -7.12117
+ra -7.12223
+▁here -7.12258
+▁part -7.14819
+an -7.1525
+▁ninety -7.15337
+▁ten -7.15777
+or -7.16632
+ion -7.17256
+th -7.17343
+▁e -7.18744
+ic -7.18833
+st -7.18837
+ar -7.1947
+▁man -7.20143
+▁which -7.21389
+▁how -7.22966
+ve -7.24093
+▁come -7.24215
+▁name -7.2462
+▁good -7.24661
+▁does -7.25244
+▁made -7.27013
+▁people -7.28101
+▁located -7.28517
+▁same -7.28573
+z -7.28984
+il -7.29175
+▁work -7.29547
+is -7.31747
+b -7.31875
+ne -7.31966
+é -7.32482
+▁eighty -7.32831
+▁place -7.32839
+ri -7.33443
+ia -7.34467
+▁mister -7.35176
+de -7.35478
+▁however -7.35557
+▁several -7.35999
+▁said -7.36138
+ers -7.36296
+▁our -7.36573
+▁us -7.37159
+▁right -7.37557
+▁way -7.37892
+▁take -7.3891
+it -7.39095
+▁many -7.39144
+▁could -7.39461
+▁still -7.39727
+et -7.39919
+▁where -7.40619
+li -7.40627
+at -7.41095
+▁jean -7.41127
+▁l -7.41198
+w -7.42338
+▁end -7.43025
+ur -7.43147
+▁co -7.43185
+▁going -7.43344
+lo -7.43345
+▁car -7.44641
+▁father -7.44723
+▁over -7.44781
+▁avenue -7.45629
+▁p -7.45736
+se -7.46643
+ta -7.46908
+ian -7.47345
+▁open -7.48166
+▁french -7.48425
+ry -7.48469
+▁into -7.487
+ity -7.48829
+▁found -7.49105
+▁most -7.49161
+te -7.50157
+ma -7.50563
+▁sur -7.51036
+us -7.5112
+▁year -7.51187
+▁under -7.51263
+▁than -7.51541
+▁give -7.51614
+▁city -7.53233
+na -7.53368
+▁long -7.53396
+▁get -7.54169
+ck -7.54528
+▁ -7.55118
+▁second -7.55864
+v -7.56743
+▁want -7.5704
+ce -7.57336
+▁last -7.57501
+la -7.57515
+▁g -7.57515
+▁must -7.57521
+▁used -7.58372
+ter -7.58758
+▁day -7.59254
+▁ca -7.60238
+ine -7.60344
+▁left -7.60891
+▁let -7.61661
+ment -7.61969
+▁mrs -7.62264
+▁main -7.63087
+▁each -7.6381
+▁years -7.64102
+▁son -7.64153
+▁any -7.64216
+▁group -7.64652
+▁war -7.65233
+▁ro -7.65662
+▁f -7.66775
+▁con -7.66814
+▁through -7.66985
+▁ra -7.66996
+▁paris -7.67027
+ul -7.67114
+h -7.67841
+▁without -7.68424
+▁old -7.68807
+▁school -7.68843
+▁des -7.6886
+▁between -7.69108
+▁un -7.69406
+▁mar -7.69672
+▁little -7.69916
+ou -7.70055
+▁make -7.70298
+▁france -7.70848
+▁order -7.71698
+ist -7.71823
+▁en -7.72599
+ard -7.72608
+▁down -7.7279
+▁speak -7.73276
+co -7.73488
+ol -7.73528
+▁even -7.73651
+▁family -7.73681
+▁opinion -7.73775
+▁later -7.74924
+able -7.75253
+vi -7.75835
+▁love -7.76125
+▁before -7.76241
+▁mo -7.76499
+ant -7.77249
+▁b -7.7758
+ie -7.78312
+▁al -7.7885
+▁t -7.79693
+▁won -7.79762
+▁small -7.80057
+▁look -7.80996
+▁door -7.81037
+ier -7.81303
+came -7.81319
+que -7.81415
+▁world -7.81463
+▁therefore -7.81565
+▁too -7.81593
+▁since -7.81777
+▁just -7.82003
+▁much -7.83038
+▁against -7.83406
+▁put -7.83562
+be -7.8405
+▁think -7.84286
+ac -7.84296
+▁came -7.85402
+ir -7.85551
+▁government -7.86337
+▁yes -7.86427
+- -7.86548
+▁may -7.86682
+▁young -7.8703
+ge -7.87032
+▁great -7.87106
+▁took -7.87176
+ca -7.87367
+▁town -7.87478
+▁head -7.87761
+▁born -7.87903
+▁madam -7.88163
+▁never -7.88176
+▁general -7.88403
+to -7.88423
+▁find -7.88543
+▁again -7.88701
+▁mother -7.89158
+▁being -7.89383
+x -7.8963
+au -7.89863
+di -7.90088
+▁few -7.90211
+▁those -7.90353
+▁r -7.90609
+▁m -7.90729
+ate -7.90869
+▁house -7.91402
+ot -7.91459
+man -7.91603
+ba -7.91609
+▁lot -7.91839
+ni -7.91951
+▁went -7.92073
+me -7.9292
+▁article -7.93035
+▁per -7.93131
+▁species -7.93223
+▁se -7.9323
+▁church -7.93289
+▁team -7.93307
+▁st -7.93999
+▁water -7.94082
+▁such -7.94109
+▁need -7.94327
+▁rue -7.9461
+▁today -7.94875
+▁oh -7.95293
+ent -7.95317
+▁become -7.95333
+▁national -7.95359
+▁why -7.95389
+▁k -7.95503
+▁c -7.95731
+ies -7.9589
+▁himself -7.95914
+▁village -7.96057
+▁sure -7.96203
+ton -7.96768
+▁vi -7.96892
+▁law -7.97116
+ga -7.97421
+▁got -7.97593
+▁life -7.99129
+▁president -7.99847
+▁minister -7.9985
+▁hand -7.99926
+▁because -7.99969
+po -7.99999
+ti -8.00035
+▁while -8.00099
+ive -8.004
+ff -8.00525
+sh -8.0055
+▁bo -8.00799
+▁say -8.01227
+id -8.01583
+▁li -8.01626
+▁known -8.01696
+▁pa -8.01736
+go -8.03036
+▁ga -8.0355
+▁bar -8.03633
+ting -8.04396
+▁ri -8.04441
+▁di -8.04543
+▁important -8.04707
+▁pi -8.05278
+che -8.0533
+▁away -8.05509
+ous -8.05723
+▁na -8.05843
+ville -8.0585
+▁mi -8.06003
+▁big -8.06101
+▁club -8.06778
+▁show -8.0702
+▁post -8.07286
+▁state -8.07854
+▁should -8.0792
+▁named -8.08104
+ted -8.08772
+ary -8.09091
+▁around -8.09142
+▁sister -8.0955
+age -8.101
+▁thu -8.10149
+pe -8.1037
+▁les -8.10417
+▁large -8.10718
+do -8.10795
+▁use -8.1087
+▁mont -8.10892
+▁pierre -8.1185
+ron -8.1191
+ci -8.12076
+da -8.12241
+▁called -8.12447
+▁often -8.12698
+▁room -8.12986
+▁south -8.13056
+▁university -8.13088
+um -8.13398
+▁public -8.1341
+▁talk -8.13532
+son -8.13739
+▁side -8.13894
+▁both -8.14102
+ated -8.14313
+bi -8.14501
+tic -8.1489
+ha -8.14893
+ke -8.15271
+▁works -8.15317
+▁brother -8.15732
+un -8.15884
+▁turn -8.15944
+▁goes -8.16193
+▁high -8.16194
+▁ta -8.1623
+va -8.17083
+▁home -8.17231
+▁please -8.17375
+▁men -8.17449
+▁ar -8.18119
+▁ma -8.18174
+▁tell -8.18181
+▁children -8.18507
+▁bi -8.19135
+▁different -8.19387
+▁cha -8.19503
+▁point -8.19526
+▁n -8.19783
+ness -8.20142
+tte -8.20376
+▁always -8.2043
+ism -8.20724
+▁north -8.21028
+▁fifteen -8.21028
+▁case -8.21061
+les -8.22248
+▁every -8.23136
+▁river -8.23397
+▁another -8.23463
+am -8.24308
+bo -8.24539
+▁following -8.24547
+▁really -8.2467
+▁country -8.25042
+▁form -8.2525
+ka -8.26388
+ish -8.26442
+▁twelve -8.26908
+▁everything -8.26908
+▁ah -8.27664
+▁sir -8.28247
+▁already -8.28514
+▁next -8.28839
+▁u -8.29155
+ig -8.29238
+no -8.29247
+▁land -8.29664
+▁eleven -8.29819
+▁jo -8.30054
+▁off -8.30209
+▁gra -8.30289
+▁own -8.30312
+▁pro -8.30689
+▁sub -8.30715
+▁member -8.30869
+gu -8.30971
+▁ex -8.30986
+▁table -8.3106
+cy -8.31081
+▁sp -8.31106
+ure -8.31194
+▁code -8.32
+▁toward -8.32192
+▁ju -8.32282
+▁particular -8.3231
+▁cor -8.32382
+▁red -8.32465
+ille -8.32482
+▁close -8.32489
+▁district -8.32816
+▁finally -8.33168
+as -8.33275
+▁leave -8.33286
+▁help -8.33493
+▁season -8.33689
+land -8.33742
+▁ha -8.33805
+▁black -8.33855
+▁pre -8.33887
+▁front -8.34013
+▁game -8.34049
+▁lo -8.34251
+▁da -8.344
+▁center -8.3477
+▁v -8.34906
+▁night -8.35216
+▁ho -8.3538
+om -8.35759
+▁nothing -8.36089
+▁station -8.36245
+▁among -8.36609
+▁far -8.36693
+tion -8.3708
+▁company -8.37135
+▁enough -8.37666
+▁dis -8.37689
+▁bu -8.3769
+▁po -8.379
+▁paul -8.3797
+▁commission -8.37995
+▁ba -8.38126
+▁building -8.38164
+▁king -8.38556
+▁coming -8.38577
+▁loved -8.38716
+▁career -8.39134
+▁thought -8.39189
+ze -8.39279
+bu -8.39743
+▁live -8.39889
+ling -8.39983
+▁something -8.40184
+▁louis -8.40214
+▁system -8.40364
+▁fa -8.40467
+▁god -8.41196
+mo -8.41199
+gue -8.41325
+▁office -8.41402
+mi -8.41434
+▁light -8.41664
+op -8.41855
+im -8.42209
+▁ka -8.42551
+▁fourteen -8.42576
+▁friend -8.42779
+▁daughter -8.42949
+▁white -8.42953
+let -8.43191
+ru -8.43476
+▁music -8.43511
+▁album -8.437
+▁book -8.43798
+ver -8.43952
+▁once -8.43958
+▁mu -8.44285
+▁region -8.44344
+▁lu -8.44413
+▁best -8.44465
+▁sixteen -8.44838
+▁comes -8.45119
+▁h -8.45223
+▁plays -8.45471
+▁asked -8.45592
+▁eighteen -8.45606
+▁pe -8.45897
+▁quit -8.46266
+▁present -8.46658
+▁val -8.46991
+▁married -8.47009
+▁dead -8.47026
+▁played -8.47041
+pa -8.47349
+▁things -8.47386
+▁free -8.47438
+▁built -8.47544
+▁takes -8.47799
+▁face -8.47848
+▁real -8.47935
+▁given -8.47995
+▁near -8.47998
+▁area -8.48121
+▁third -8.48135
+▁understand -8.4833
+ang -8.48857
+▁child -8.48921
+ther -8.49063
+▁anything -8.49124
+▁women -8.49328
+ical -8.49681
+▁project -8.49722
+▁quickly -8.49776
+▁question -8.49876
+▁doctor -8.49979
+▁play -8.50008
+▁woman -8.50138
+vo -8.50207
+▁certain -8.50429
+nes -8.50437
+▁english -8.50526
+hi -8.50555
+ance -8.50571
+▁hu -8.50714
+▁ti -8.50718
+▁honor -8.50888
+▁party -8.51031
+ng -8.51123
+▁ve -8.51141
+▁bit -8.51162
+▁tour -8.51295
+▁sa -8.51418
+▁fact -8.51479
+▁ver -8.51525
+▁sometimes -8.51541
+▁thirteen -8.51744
+ut -8.519
+▁less -8.51924
+▁mer -8.52027
+▁done -8.52407
+▁line -8.52416
+▁title -8.52773
+▁american -8.52785
+eau -8.52841
+▁amendments -8.53046
+▁seem -8.53073
+▁color -8.53075
+▁united -8.53216
+▁call -8.53319
+▁art -8.53361
+▁municipality -8.53404
+▁ne -8.5359
+▁according -8.54068
+lin -8.54124
+▁lives -8.54252
+▁bou -8.54328
+ence -8.54395
+▁yet -8.54593
+▁stay -8.54624
+▁department -8.54646
+▁film -8.54646
+▁history -8.5465
+▁stop -8.54688
+▁alley -8.5523
+con -8.55534
+▁lost -8.5599
+▁death -8.56131
+▁days -8.56136
+▁short -8.5626
+▁started -8.56282
+▁grand -8.56319
+est -8.56391
+▁wife -8.56563
+▁count -8.56854
+▁har -8.56891
+▁set -8.57152
+▁bad -8.57469
+▁miss -8.57801
+▁keep -8.57852
+▁park -8.57992
+▁position -8.58074
+▁council -8.58286
+▁bring -8.58307
+▁island -8.58581
+▁author -8.58636
+▁capital -8.58724
+men -8.58789
+mer -8.58928
+ph -8.5911
+ak -8.5911
+ful -8.59164
+▁é -8.59201
+▁moment -8.59281
+▁ja -8.59298
+▁having -8.59604
+▁believe -8.59802
+▁power -8.59812
+ix -8.59891
+▁va -8.59981
+▁wanted -8.60196
+▁marie -8.60549
+▁francois -8.60783
+▁mor -8.60949
+▁voice -8.61389
+▁indeed -8.61429
+▁inter -8.61614
+▁soon -8.61616
+▁poor -8.61711
+▁period -8.61839
+▁international -8.62064
+▁char -8.62185
+▁german -8.62193
+he -8.62198
+ized -8.62358
+ad -8.62381
+▁gave -8.6242
+▁taking -8.62748
+▁morning -8.6275
+▁nice -8.62808
+▁meet -8.6284
+line -8.63003
+▁local -8.63377
+▁read -8.63432
+▁eyes -8.63446
+ty -8.63503
+▁cla -8.63545
+▁san -8.63606
+▁course -8.63745
+min -8.63802
+▁taken -8.63959
+▁half -8.64125
+sse -8.64162
+▁word -8.64302
+▁better -8.64367
+▁air -8.64416
+▁square -8.64588
+▁seems -8.64778
+▁change -8.64848
+▁charles -8.64853
+▁thing -8.65135
+▁shall -8.65171
+ric -8.65193
+▁hear -8.6528
+ran -8.65435
+lle -8.65619
+ki -8.6566
+▁full -8.65874
+▁pay -8.66084
+▁states -8.66449
+▁sea -8.66658
+tin -8.66952
+▁seen -8.67099
+▁role -8.67105
+▁boulevard -8.67415
+▁studies -8.67415
+▁considered -8.67549
+▁ask -8.67572
+▁ste -8.67587
+lu -8.67593
+▁died -8.6762
+▁happy -8.67654
+za -8.67666
+os -8.68139
+▁level -8.682
+▁true -8.68534
+▁sha -8.68709
+▁plan -8.68817
+▁girl -8.69012
+▁arm -8.69228
+▁thank -8.69316
+▁withdraw -8.69549
+▁kind -8.69592
+▁master -8.69642
+▁castle -8.69675
+less -8.69765
+▁former -8.69844
+▁port -8.69881
+▁success -8.70193
+▁gu -8.70314
+der -8.70364
+▁ru -8.70666
+▁roman -8.7082
+▁almo -8.70854
+pi -8.70898
+▁possible -8.71079
+▁fire -8.71307
+mont -8.71456
+▁subject -8.71562
+▁created -8.71586
+▁player -8.71618
+▁saw -8.71661
+ny -8.71907
+▁et -8.72025
+ise -8.7204
+ois -8.72175
+▁vote -8.72368
+▁age -8.72468
+sa -8.72744
+▁received -8.72747
+▁seventeen -8.72815
+▁act -8.72885
+▁west -8.72909
+▁series -8.73116
+▁martin -8.7315
+▁idea -8.73228
+▁omer -8.73427
+▁bra -8.74172
+▁heart -8.74305
+▁together -8.7434
+▁ni -8.74697
+ber -8.74878
+ial -8.74992
+▁site -8.75078
+▁beautiful -8.75111
+tra -8.75216
+▁makes -8.75433
+▁au -8.75484
+io -8.7554
+▁times -8.75543
+▁che -8.75612
+▁buy -8.75622
+▁money -8.75642
+elle -8.75712
+▁hard -8.75928
+ring -8.76039
+▁political -8.76246
+ten -8.76277
+▁care -8.76308
+▁until -8.76673
+▁commune -8.76674
+▁ge -8.76736
+▁worked -8.76779
+▁gar -8.76828
+▁start -8.76893
+ster -8.76928
+▁husband -8.77197
+▁date -8.77363
+▁strong -8.77427
+▁job -8.77633
+▁person -8.77654
+▁buried -8.77727
+▁measure -8.77756
+ho -8.78005
+▁z -8.78149
+ja -8.78156
+▁include -8.78206
+▁tu -8.78249
+▁nineteen -8.78257
+▁text -8.78524
+▁wood -8.78536
+▁el -8.78537
+▁rest -8.78784
+the -8.78895
+▁guy -8.78896
+▁army -8.78906
+▁join -8.78978
+ette -8.78999
+ger -8.79045
+▁john -8.79059
+ology -8.79127
+▁problem -8.79167
+▁path -8.79235
+▁matter -8.79367
+▁bri -8.79421
+▁fe -8.79519
+▁em -8.79557
+▁report -8.79877
+▁route -8.79937
+▁getting -8.79999
+▁dear -8.80034
+▁horse -8.8018
+▁doing -8.80277
+▁hope -8.80284
+▁looking -8.80346
+▁example -8.80412
+▁fall -8.80764
+▁chi -8.80856
+▁cold -8.80878
+▁regard -8.80968
+▁ci -8.8118
+▁director -8.81316
+▁fight -8.8138
+▁books -8.81382
+▁social -8.81439
+▁bon -8.81543
+▁imp -8.81916
+▁wi -8.81932
+ile -8.81947
+▁michel -8.82068
+▁tri -8.82286
+▁european -8.82398
+and -8.82443
+▁wait -8.82548
+▁museum -8.82611
+▁studied -8.82611
+▁double -8.82612
+▁attack -8.82613
+mar -8.82712
+dy -8.82773
+▁major -8.83112
+gra -8.83113
+▁visit -8.83176
+▁hi -8.83208
+▁final -8.83245
+▁court -8.83332
+led -8.83533
+▁run -8.83642
+▁original -8.83849
+▁sc -8.83876
+▁gre -8.83959
+ize -8.83992
+▁bre -8.8408
+▁fi -8.84096
+▁single -8.84123
+▁committee -8.84294
+▁type -8.84304
+▁band -8.8449
+▁boy -8.84518
+▁various -8.84581
+▁impasse -8.84581
+▁remains -8.8468
+mon -8.84751
+▁east -8.84942
+▁waiting -8.84992
+j -8.85264
+▁championship -8.85437
+▁based -8.8551
+iere -8.85608
+▁th -8.85862
+▁story -8.86038
+qui -8.86196
+▁origin -8.8622
+▁everyone -8.86295
+▁cap -8.86473
+▁anymore -8.86609
+ina -8.86673
+▁movie -8.86761
+up -8.8688
+▁train -8.86913
+tor -8.86944
+ko -8.87014
+▁vo -8.87107
+▁pathway -8.87179
+use -8.87394
+▁competition -8.87455
+▁style -8.87458
+▁further -8.87465
+▁talking -8.87719
+▁sign -8.88061
+rs -8.88146
+▁able -8.88306
+▁charge -8.88352
+▁language -8.88599
+▁currently -8.88614
+▁tomorrow -8.8863
+▁je -8.88749
+way -8.88962
+lon -8.8899
+▁offer -8.89014
+▁playing -8.89046
+▁hit -8.89251
+lic -8.8927
+▁song -8.89301
+▁written -8.8952
+▁wall -8.89981
+▁return -8.90028
+▁w -8.90034
+▁ki -8.90072
+ner -8.90384
+▁unfavorable -8.90418
+ler -8.9047
+val -8.90505
+▁living -8.90596
+▁brought -8.9072
+▁latter -8.91027
+▁told -8.91093
+▁bank -8.91183
+lé -8.91252
+▁issue -8.91483
+▁middle -8.91629
+▁cannot -8.91689
+▁pont -8.91724
+▁met -8.91928
+▁pan -8.92039
+ens -8.92112
+out -8.92203
+▁ko -8.92222
+em -8.92242
+iv -8.92252
+fa -8.92426
+▁answer -8.92476
+▁yellow -8.92546
+per -8.92632
+▁identical -8.92854
+ization -8.92909
+ey -8.92939
+▁working -8.92958
+gi -8.93141
+▁common -8.93227
+▁captain -8.93472
+▁whole -8.93592
+▁qu -8.9379
+▁words -8.93858
+cha -8.93872
+▁longer -8.94118
+▁cost -8.94123
+▁tra -8.94306
+ien -8.9432
+▁mountain -8.94421
+▁te -8.94454
+ré -8.94461
+our -8.94552
+tu -8.94678
+▁central -8.94724
+▁current -8.94737
+▁sal -8.9476
+rew -8.94782
+▁sta -8.94849
+▁walk -8.95082
+▁cut -8.95109
+▁cause -8.95195
+▁bis -8.95273
+▁hold -8.95344
+▁mean -8.9535
+▁seat -8.95598
+▁royal -8.9567
+▁months -8.95763
+▁county -8.95843
+cu -8.95865
+▁published -8.95937
+▁result -8.95974
+▁military -8.95984
+▁garden -8.96001
+▁along -8.96078
+▁mon -8.96205
+tan -8.96239
+chi -8.96398
+▁fish -8.9642
+▁members -8.96576
+aud -8.9666
+▁fear -8.96715
+ain -8.96783
+▁kept -8.96961
+▁hat -8.96978
+▁human -8.97052
+▁sun -8.97104
+▁rock -8.97153
+▁myself -8.97264
+▁battle -8.97265
+fi -8.974
+ities -8.97774
+▁movement -8.98283
+▁camp -8.98589
+if -8.98901
+▁held -8.98926
+nce -8.99049
+ay -8.99134
+han -8.99217
+▁die -8.99294
+▁writing -8.99544
+▁games -8.99588
+ud -8.99613
+ov -8.99852
+▁research -8.99874
+▁depend -8.99876
+▁bell -8.99953
+ward -8.99988
+bert -9.00083
+▁move -9.00205
+▁green -9.00208
+▁debate -9.00223
+▁professional -9.00258
+▁others -9.0029
+▁teacher -9.00375
+den -9.00397
+ez -9.00504
+▁sch -9.00575
+ions -9.0068
+ven -9.007
+▁houses -9.00823
+▁situation -9.00871
+und -9.01065
+cher -9.01118
+▁lord -9.01121
+▁italian -9.01217
+▁round -9.01255
+▁cat -9.01315
+▁forest -9.01446
+▁practice -9.01541
+▁write -9.01654
+▁bas -9.01717
+▁reason -9.01754
+▁favor -9.01822
+ite -9.01939
+▁nor -9.02
+▁territory -9.02215
+▁usual -9.02267
+ai -9.02411
+▁tax -9.0245
+ya -9.02459
+tre -9.02466
+▁past -9.02486
+▁follow -9.0249
+side -9.02596
+▁dark -9.02766
+▁jacques -9.02921
+▁body -9.02939
+▁bois -9.02948
+▁par -9.02959
+▁late -9.03002
+▁anti -9.03053
+gon -9.03388
+▁arrived -9.03436
+don -9.03471
+▁turned -9.03519
+ves -9.03525
+▁parents -9.03588
+▁dog -9.03807
+▁mine -9.03815
+▁heard -9.03898
+▁cross -9.03903
+▁meeting -9.03931
+▁border -9.03935
+▁allow -9.04119
+zi -9.04174
+▁above -9.0427
+▁field -9.04284
+▁least -9.04388
+▁cre -9.04431
+▁wa -9.04514
+▁sainte -9.0457
+▁province -9.04613
+▁lake -9.04657
+▁mission -9.0488
+▁blue -9.04922
+ab -9.04944
+ow -9.04953
+▁rich -9.05127
+▁host -9.05292
+▁beginning -9.05309
+▁ser -9.05541
+▁hall -9.05607
+ory -9.05653
+▁service -9.05754
+▁del -9.0602
+▁alone -9.06028
+▁plant -9.06307
+▁division -9.06361
+▁information -9.06362
+ally -9.06543
+▁mal -9.06546
+▁su -9.06564
+lan -9.06672
+▁asking -9.06682
+onne -9.06705
+tel -9.0681
+wa -9.06894
+▁stage -9.06915
+▁philippe -9.07079
+▁letter -9.07092
+▁business -9.07225
+▁champ -9.07328
+▁van -9.07336
+▁population -9.07425
+▁pass -9.07536
+▁andré -9.076
+ec -9.07616
+▁organized -9.07658
+▁education -9.07782
+▁future -9.07782
+▁hands -9.07799
+▁reading -9.08134
+▁followed -9.0815
+▁si -9.08397
+▁eat -9.08497
+▁marc -9.08592
+▁event -9.08749
+▁friends -9.08809
+▁yesterday -9.08861
+ins -9.08874
+ble -9.09028
+ator -9.09058
+▁fre -9.09126
+▁approach -9.09224
+▁program -9.09224
+▁stopped -9.09225
+▁y -9.09228
+▁scene -9.09236
+▁fo -9.09368
+▁gi -9.09488
+▁community -9.09587
+▁inhabitants -9.09587
+▁herself -9.09587
+▁necessary -9.09995
+ans -9.10113
+▁simple -9.10283
+▁goal -9.10567
+▁luc -9.10606
+▁difficult -9.1068
+▁clear -9.10788
+▁bro -9.10799
+gan -9.10818
+ast -9.1099
+▁ab -9.1115
+▁liked -9.11318
+▁cra -9.11531
+ux -9.11624
+ction -9.11781
+▁joseph -9.11797
+▁itself -9.11798
+▁star -9.11911
+▁wish -9.12055
+▁ber -9.12094
+▁comp -9.1214
+▁famous -9.12171
+▁police -9.12171
+▁window -9.12205
+▁gri -9.12308
+▁watch -9.12613
+ée -9.12686
+gen -9.12724
+▁cast -9.12748
+▁action -9.12758
+ges -9.12771
+▁revolution -9.12896
+▁pu -9.13028
+▁list -9.13037
+mes -9.13039
+▁bal -9.13085
+▁leg -9.13124
+▁europe -9.13218
+▁leaves -9.13385
+dic -9.13386
+▁training -9.13434
+▁decided -9.13437
+▁mind -9.13462
+▁ce -9.1353
+▁ever -9.13613
+▁version -9.13691
+▁shows -9.13707
+rin -9.13715
+▁low -9.13767
+▁test -9.13839
+▁col -9.13904
+▁race -9.13966
+▁special -9.14019
+▁leaving -9.14058
+▁cro -9.14143
+▁collection -9.14144
+▁model -9.14205
+▁mat -9.14331
+▁due -9.14355
+court -9.14448
+dor -9.14567
+▁moved -9.146
+▁making -9.14824
+▁production -9.14855
+wi -9.14858
+▁beau -9.14898
+▁wind -9.14999
+▁serve -9.15064
+ious -9.15144
+▁henri -9.15233
+▁rose -9.15322
+vin -9.15333
+▁cu -9.15341
+pp -9.15535
+▁private -9.15593
+art -9.15806
+▁completely -9.15821
+▁behind -9.1598
+▁rapporteur -9.1598
+av -9.16148
+her -9.1624
+▁newspaper -9.16369
+▁female -9.1637
+▁defend -9.16393
+▁week -9.16413
+▁natural -9.16461
+▁remained -9.16603
+▁generally -9.16659
+▁breton -9.16814
+▁break -9.17179
+▁hour -9.17238
+▁especially -9.17546
+▁numerous -9.17546
+▁travel -9.17557
+▁tall -9.17587
+▁immediately -9.17599
+▁store -9.17614
+▁served -9.17632
+▁hair -9.18037
+▁pin -9.1806
+▁cup -9.18208
+▁fla -9.1825
+oc -9.18261
+ap -9.18311
+▁cal -9.18328
+▁view -9.1834
+▁soul -9.18477
+▁fu -9.18549
+▁continued -9.186
+▁mou -9.18602
+▁step -9.18605
+net -9.18691
+▁sent -9.18712
+▁doubt -9.18738
+▁albert -9.1879
+rie -9.19084
+hu -9.19087
+lie -9.19101
+▁source -9.19137
+▁couple -9.19141
+▁robert -9.19198
+over -9.19297
+▁fell -9.19317
+ide -9.19337
+▁killed -9.1936
+▁personal -9.19383
+qu -9.1942
+▁figure -9.19538
+▁ground -9.19555
+▁gold -9.19887
+▁pra -9.19918
+▁within -9.19928
+▁tower -9.19932
+▁nevertheless -9.19941
+▁someone -9.19947
+▁development -9.19981
+▁officer -9.20081
+▁maybe -9.20092
+▁actor -9.20112
+ists -9.20138
+▁para -9.20143
+mit -9.20218
+ship -9.20297
+▁effect -9.20312
+las -9.20332
+▁progress -9.20346
+ade -9.20369
+lay -9.20429
+▁interest -9.20637
+▁control -9.20737
+▁chief -9.20753
+▁trans -9.20772
+gne -9.20833
+▁reasons -9.2085
+▁novel -9.20864
+▁opening -9.20927
+▁j -9.2097
+▁replaced -9.21091
+▁top -9.21134
+▁cemetery -9.21161
+▁economic -9.21161
+ative -9.21171
+▁dr -9.21311
+▁feel -9.21432
+▁lady -9.21432
+▁popular -9.21573
+du -9.21577
+▁brown -9.21581
+▁joined -9.21807
+▁prince -9.21925
+▁neu -9.21933
+▁association -9.21982
+▁hot -9.22043
+▁cur -9.22124
+ole -9.22264
+▁similar -9.22395
+▁themselves -9.22395
+▁space -9.22429
+mor -9.22469
+▁accept -9.22534
+▁extreme -9.2281
+▁knew -9.2281
+ps -9.22848
+tz -9.23101
+▁access -9.2322
+▁catholic -9.23227
+ana -9.23394
+▁pla -9.23456
+igny -9.23542
+▁agree -9.23545
+▁respect -9.23645
+▁outside -9.23763
+▁mé -9.23807
+▁fourth -9.23915
+▁process -9.24066
+▁critic -9.24068
+▁headquarters -9.24111
+vers -9.24175
+▁design -9.24225
+bb -9.24312
+▁lead -9.24337
+rou -9.24491
+▁majority -9.24572
+ches -9.24585
+▁ville -9.24649
+nay -9.24752
+▁operation -9.24921
+▁listen -9.24938
+▁suggest -9.25337
+▁judge -9.25338
+▁union -9.25434
+▁forward -9.2554
+▁try -9.25754
+des -9.25822
+▁elected -9.25895
+▁character -9.25931
+▁mass -9.25986
+▁produced -9.26034
+ip -9.26146
+▁remain -9.2622
+▁los -9.26464
+▁deal -9.26468
+▁early -9.26501
+ane -9.2651
+▁neighbor -9.26622
+▁covered -9.2667
+▁guard -9.26703
+euil -9.26709
+▁key -9.26925
+▁wrong -9.27057
+▁wrote -9.27057
+jo -9.27099
+▁month -9.27121
+▁anne -9.27126
+ham -9.27135
+▁closed -9.27151
+by -9.27159
+▁gro -9.27186
+my -9.27195
+▁leader -9.27326
+▁bridge -9.27504
+▁belong -9.27509
+▁hours -9.27661
+ria -9.27844
+van -9.2789
+▁allows -9.27907
+▁chair -9.27914
+▁send -9.27945
+▁dedicated -9.27981
+▁study -9.28126
+▁ch -9.28157
+▁administrative -9.28367
+▁york -9.28371
+tri -9.28403
+▁consider -9.28514
+▁nu -9.28679
+ors -9.2868
+ari -9.28704
+▁fer -9.28824
+▁bird -9.2883
+▁wine -9.28894
+▁denis -9.2909
+▁priest -9.29262
+▁stone -9.29293
+▁kill -9.29319
+fort -9.29384
+▁bay -9.29481
+car -9.29614
+▁minutes -9.29621
+non -9.29623
+▁weather -9.29696
+vent -9.29708
+bra -9.29903
+▁add -9.29944
+▁glass -9.29945
+▁gui -9.3006
+▁object -9.30107
+▁account -9.30141
+▁probably -9.30141
+▁contract -9.30142
+pped -9.30184
+▁term -9.30207
+lier -9.30304
+▁georges -9.30484
+▁fra -9.30557
+▁request -9.3059
+▁running -9.30595
+▁bel -9.31154
+▁bru -9.31361
+▁sau -9.31404
+▁academy -9.31492
+▁league -9.31493
+▁gives -9.31592
+ok -9.31832
+ley -9.31854
+▁drink -9.31947
+▁organization -9.31957
+▁mostly -9.31977
+▁rep -9.32008
+▁base -9.32015
+▁inside -9.32023
+▁students -9.32028
+▁coach -9.32068
+▁official -9.32105
+▁card -9.32156
+▁ship -9.32228
+▁lived -9.32486
+▁theater -9.32624
+ev -9.32691
+▁activity -9.32869
+▁mayor -9.32889
+ral -9.32925
+▁actually -9.32936
+▁rain -9.32977
+▁composed -9.331
+▁deep -9.33167
+illon -9.33216
+▁bur -9.33245
+▁bishop -9.33323
+▁neighborhood -9.33326
+▁remove -9.33338
+▁construction -9.3336
+▁fishing -9.3339
+▁create -9.33446
+▁legal -9.33612
+▁managed -9.33725
+▁bat -9.33748
+▁available -9.33786
+▁urban -9.33791
+▁market -9.33836
+té -9.33924
+wn -9.34045
+▁republic -9.34106
+ieres -9.34106
+▁london -9.34251
+▁british -9.34252
+▁comm -9.34262
+▁sentence -9.34275
+▁coast -9.34323
+▁began -9.3433
+▁sick -9.34368
+▁franc -9.34387
+▁wins -9.34387
+▁thanks -9.34421
+gar -9.34433
+▁pen -9.34446
+enne -9.34497
+▁discover -9.34506
+▁added -9.34509
+▁secret -9.34688
+▁beat -9.34774
+▁ob -9.34878
+▁painter -9.34905
+▁historical -9.3493
+oux -9.35032
+▁bé -9.3504
+▁cour -9.35108
+▁health -9.3521
+▁discovered -9.35411
+▁led -9.35519
+ome -9.35619
+▁endemic -9.3566
+▁science -9.3566
+cle -9.35759
+▁sound -9.35843
+▁lou -9.35862
+▁rou -9.35949
+▁gen -9.35952
+▁sw -9.3611
+▁dream -9.36134
+▁sold -9.36166
+▁answered -9.3625
+▁reach -9.36285
+▁board -9.36328
+▁share -9.36601
+▁starts -9.3664
+▁bla -9.36694
+▁hotel -9.36749
+▁developed -9.3679
+ugh -9.36899
+▁brothers -9.37209
+▁evening -9.37238
+ates -9.37336
+▁canton -9.37343
+▁tro -9.37486
+▁rank -9.37522
+▁method -9.37569
+pin -9.37607
+▁regional -9.37655
+▁entire -9.37755
+▁professor -9.38052
+▁remind -9.38057
+▁countries -9.38072
+▁christian -9.38089
+ified -9.38114
+ub -9.38275
+▁ph -9.38279
+▁dance -9.38373
+ought -9.38449
+▁student -9.38452
+▁edition -9.38606
+▁march -9.38702
+▁ready -9.38762
+▁germany -9.3877
+rat -9.38936
+▁trip -9.38951
+▁sell -9.38958
+fe -9.39018
+▁welcome -9.39025
+▁degree -9.39026
+▁increase -9.39044
+▁stand -9.39054
+▁secretary -9.39058
+illa -9.39126
+▁class -9.39223
+▁opened -9.39264
+▁sleep -9.39319
+▁writer -9.39351
+▁lower -9.39371
+▁exist -9.3938
+▁cho -9.39416
+▁hospital -9.39516
+▁fruit -9.39516
+▁includ -9.39529
+▁tired -9.39541
+▁domini -9.39589
+ex -9.39806
+▁app -9.39834
+▁win -9.39852
+▁ran -9.39922
+▁civil -9.40008
+▁spanish -9.40008
+red -9.40168
+▁clos -9.40193
+▁previous -9.40204
+▁fine -9.40218
+▁chapel -9.404
+▁decision -9.40503
+▁situated -9.40503
+▁kingdom -9.40504
+ault -9.40543
+▁engineer -9.40742
+▁sl -9.40819
+tier -9.40879
+din -9.4096
+▁network -9.41001
+▁despite -9.41001
+▁prove -9.41032
+▁plane -9.41044
+▁happen -9.41123
+▁addition -9.4114
+▁rule -9.41255
+▁develop -9.41262
+▁emperor -9.41501
+▁finished -9.41629
+bel -9.41704
+▁comb -9.41813
+▁link -9.41832
+ban -9.41979
+▁laurent -9.42004
+▁function -9.4201
+▁uncle -9.42052
+▁track -9.42116
+heim -9.42168
+▁complete -9.42247
+▁maintain -9.42324
+▁piece -9.42618
+ries -9.42735
+aire -9.42863
+▁experience -9.42881
+▁arch -9.43011
+▁foot -9.43071
+fer -9.43249
+▁painting -9.43459
+▁calm -9.43525
+▁occasion -9.43526
+▁felt -9.43531
+▁modern -9.43537
+▁worth -9.43539
+▁mol -9.43707
+▁though -9.43832
+▁nature -9.4404
+▁western -9.44055
+ais -9.44077
+▁sat -9.44083
+▁winter -9.44096
+▁radio -9.44133
+▁non -9.44147
+can -9.44562
+▁super -9.44573
+▁sar -9.44596
+▁stayed -9.44702
+▁italy -9.44727
+▁bill -9.4477
+▁era -9.44893
+▁singer -9.44914
+▁wear -9.44973
+▁refer -9.45005
+▁pol -9.45228
+▁carried -9.45263
+▁football -9.45593
+▁industry -9.45593
+▁spread -9.45593
+▁replied -9.45594
+ines -9.45609
+▁prefer -9.45609
+▁songs -9.45652
+▁lose -9.45656
+ify -9.45823
+▁saying -9.45886
+▁wing -9.46012
+del -9.46022
+▁sh -9.46045
+▁bed -9.4609
+▁structure -9.46117
+▁material -9.46117
+▁marquis -9.46122
+▁transport -9.46123
+▁direction -9.46206
+tive -9.46256
+vert -9.4631
+▁tried -9.46374
+▁sav -9.46434
+▁duke -9.46538
+igne -9.46597
+uc -9.46597
+▁religious -9.46643
+▁complex -9.46643
+▁related -9.46644
+▁influence -9.46701
+▁parish -9.46747
+▁victor -9.46824
+erie -9.46849
+▁contain -9.46942
+we -9.46984
+▁grow -9.4704
+▁colleagues -9.47041
+▁tonight -9.47177
+▁ring -9.4718
+quet -9.4729
+lot -9.47371
+▁reached -9.47557
+pen -9.47566
+▁gets -9.47612
+clock -9.4765
+▁document -9.47704
+▁giving -9.47705
+▁moving -9.47707
+▁representative -9.47712
+▁shi -9.47733
+▁section -9.47756
+ah -9.47795
+▁page -9.47795
+▁belongs -9.47839
+▁pretty -9.4824
+▁points -9.48285
+▁active -9.48314
+▁lie -9.48334
+berg -9.4872
+▁shot -9.48768
+af -9.48797
+ular -9.48831
+▁tar -9.48943
+▁er -9.48993
+lor -9.49206
+▁artist -9.49217
+▁lawyer -9.49323
+▁villa -9.49403
+▁manage -9.49422
+xi -9.49678
+rupt -9.49801
+ach -9.49822
+▁ancient -9.49861
+▁says -9.49964
+▁plu -9.50018
+▁placed -9.50152
+▁rome -9.50168
+▁yourself -9.50407
+▁claude -9.5041
+▁admir -9.50412
+vel -9.50428
+▁sand -9.50626
+▁finish -9.50633
+aux -9.50733
+▁events -9.50806
+▁pit -9.50907
+▁lyon -9.50989
+ft -9.5105
+▁bor -9.51074
+▁pale -9.5109
+▁cou -9.51091
+▁fair -9.51102
+lar -9.51368
+ique -9.51428
+form -9.51441
+▁press -9.51464
+lia -9.51489
+▁nicolas -9.51509
+▁scream -9.51509
+▁straight -9.51513
+▁risk -9.51527
+▁attract -9.51535
+▁eye -9.51557
+▁lack -9.51716
+▁raised -9.51767
+▁fan -9.51773
+ius -9.51824
+rac -9.51848
+▁disc -9.5185
+ag -9.51971
+ji -9.52047
+▁altitude -9.52065
+▁england -9.52065
+▁results -9.52082
+▁angel -9.52095
+▁lac -9.52113
+▁brand -9.52154
+pu -9.5217
+▁mill -9.52204
+ological -9.52285
+bourg -9.52478
+tes -9.52482
+▁japan -9.52623
+▁companies -9.52623
+▁antoine -9.52624
+▁valley -9.52651
+▁election -9.5267
+▁ya -9.52689
+▁spot -9.52756
+▁russian -9.52897
+▁equal -9.53038
+▁servant -9.53185
+▁machine -9.53186
+▁assembly -9.53287
+▁ball -9.53311
+▁else -9.53348
+▁box -9.53356
+▁rené -9.53401
+▁belle -9.53422
+▁bus -9.53476
+▁silence -9.53769
+▁pré -9.53859
+▁trade -9.53968
+▁relationship -9.54323
+▁southern -9.54324
+more -9.54352
+▁paper -9.54382
+▁agreement -9.54385
+▁participate -9.54419
+▁ter -9.54469
+▁touch -9.54507
+je -9.54631
+ray -9.54772
+zo -9.54871
+▁activities -9.5489
+▁advantage -9.5489
+▁escape -9.5489
+▁festival -9.5489
+▁individual -9.5489
+▁promise -9.54892
+▁truth -9.54939
+▁trees -9.54975
+▁shop -9.55056
+▁carry -9.55355
+▁america -9.55424
+▁device -9.55465
+▁choice -9.55466
+lam -9.5554
+▁x -9.55561
+▁hill -9.55565
+▁traditional -9.55573
+▁appointed -9.55579
+▁frédéric -9.55639
+▁ear -9.55672
+▁regularly -9.55703
+▁ben -9.558
+▁means -9.55988
+▁college -9.56043
+▁ré -9.56508
+▁cell -9.56524
+▁parliament -9.56593
+▁fifth -9.56624
+▁unknown -9.56641
+▁match -9.56642
+▁besides -9.56657
+cent -9.56723
+▁sad -9.56798
+▁pick -9.569
+▁victory -9.56994
+bar -9.57
+▁hy -9.57158
+▁specific -9.57197
+▁nowadays -9.57209
+▁smile -9.57213
+▁phone -9.57242
+▁spain -9.57344
+▁creation -9.57371
+▁ke -9.5738
+▁command -9.57446
+▁cri -9.57569
+ini -9.57594
+▁environment -9.57797
+▁presence -9.578
+rk -9.57872
+▁included -9.57893
+bre -9.57951
+▁returned -9.57974
+▁ban -9.58129
+▁tre -9.5816
+▁mad -9.58184
+▁entered -9.5824
+nne -9.5838
+nier -9.58386
+▁marriage -9.58389
+▁reporter -9.58437
+▁climb -9.58459
+▁sou -9.58531
+▁champion -9.58565
+hen -9.58758
+▁fur -9.58775
+par -9.58819
+burg -9.58963
+ible -9.58969
+umb -9.58988
+▁normal -9.58992
+▁bernard -9.59008
+▁owner -9.59097
+zz -9.59251
+▁significant -9.59283
+▁mas -9.59317
+▁total -9.59351
+▁established -9.59356
+▁tur -9.5937
+▁summer -9.59597
+▁lé -9.59603
+▁surviv -9.59606
+▁limited -9.59618
+▁unit -9.59709
+▁require -9.59714
+▁players -9.59742
+▁concern -9.59845
+uff -9.59888
+▁animal -9.59941
+▁ham -9.6012
+ura -9.60184
+▁address -9.60186
+▁consist -9.60215
+▁claim -9.60244
+zy -9.60256
+▁canada -9.60259
+▁laugh -9.60316
+▁spent -9.60316
+▁anger -9.60318
+▁questions -9.60324
+▁shape -9.60397
+▁services -9.60626
+▁mil -9.60656
+▁explain -9.60706
+az -9.60737
+▁david -9.60792
+▁organ -9.6097
+ik -9.60974
+lat -9.61183
+▁sé -9.61252
+▁learn -9.61254
+ene -9.61299
+▁roll -9.61348
+▁empire -9.61401
+▁monument -9.61401
+▁daniel -9.61409
+▁changed -9.61411
+▁arms -9.61412
+▁buildings -9.61413
+▁reference -9.61424
+▁africa -9.61464
+duc -9.61507
+▁propose -9.61544
+▁oldest -9.61546
+▁upon -9.61576
+▁artists -9.61588
+▁sit -9.6163
+▁shouted -9.61671
+rand -9.61714
+▁sports -9.61777
+▁male -9.6178
+▁additional -9.61844
+▁counter -9.61853
+yl -9.6186
+▁sil -9.6191
+▁passed -9.61982
+▁enjoy -9.62015
+▁easy -9.62016
+▁stra -9.62199
+▁partner -9.62222
+▁soldiers -9.62226
+▁speaking -9.62235
+▁ram -9.62239
+▁presented -9.62352
+▁leads -9.62382
+▁independent -9.62632
+▁shoulder -9.62764
+▁soft -9.62818
+▁est -9.62922
+▁represents -9.62958
+▁op -9.62975
+▁gene -9.6299
+▁han -9.62995
+istic -9.63037
+dou -9.63229
+▁avoid -9.63255
+▁northern -9.63259
+cho -9.63414
+▁cover -9.63489
+ail -9.63617
+▁august -9.6363
+▁sam -9.63661
+▁feeling -9.63673
+▁force -9.6387
+▁correspond -9.63878
+elles -9.63927
+▁loire -9.63934
+▁languages -9.63944
+▁blo -9.64081
+▁aim -9.64102
+▁destroyed -9.64146
+▁protect -9.64369
+▁julien -9.64408
+▁fri -9.64482
+▁improve -9.64507
+▁draw -9.64539
+▁farm -9.64597
+▁chairs -9.64643
+▁surprised -9.64905
+▁record -9.65005
+▁lane -9.65034
+cio -9.65071
+▁remember -9.6514
+▁mistake -9.6514
+▁solution -9.6514
+▁birth -9.6514
+▁blood -9.65141
+▁bread -9.65181
+▁matches -9.65376
+▁conditions -9.65394
+▁released -9.65441
+▁fort -9.6569
+▁sara -9.65745
+▁ended -9.65747
+▁relative -9.65777
+▁germain -9.65777
+▁food -9.65778
+▁cinema -9.65779
+▁forget -9.65806
+▁receive -9.6596
+ino -9.65972
+fo -9.66015
+▁lion -9.66027
+▁dam -9.66114
+▁cli -9.6614
+▁architect -9.66392
+▁commercial -9.66418
+▁deliver -9.66422
+▁drawing -9.66472
+▁abbey -9.66514
+▁kiss -9.66524
+ual -9.66574
+▁involved -9.6658
+▁petit -9.66627
+▁characters -9.66688
+ept -9.66723
+▁equally -9.66768
+▁appears -9.66821
+▁whom -9.66823
+▁express -9.66872
+▁broke -9.66989
+▁benefit -9.67063
+▁cultural -9.67063
+▁afterwards -9.67072
+▁earth -9.67157
+▁medal -9.67159
+▁bear -9.67227
+▁size -9.67281
+rd -9.67301
+▁happened -9.67309
+▁movies -9.67314
+eur -9.6732
+▁kid -9.67359
+▁noticed -9.674
+ography -9.67516
+▁flat -9.67592
+▁television -9.67712
+▁affair -9.67713
+▁limit -9.6773
+▁mark -9.67797
+agne -9.67841
+▁francis -9.67884
+▁reality -9.67901
+▁wo -9.67934
+▁pull -9.67947
+▁animals -9.67979
+▁using -9.6819
+oche -9.68223
+▁spend -9.68337
+▁tree -9.684
+▁prevent -9.68404
+▁poet -9.68549
+pre -9.68786
+▁dan -9.68862
+▁rem -9.68871
+▁succeed -9.68994
+▁cathedral -9.69024
+▁inspired -9.69024
+▁computer -9.69024
+▁foreign -9.69024
+▁japanese -9.69025
+▁heavy -9.6903
+▁municipalities -9.69032
+▁chemin -9.69071
+▁prior -9.69094
+ium -9.69126
+▁gather -9.69132
+▁consul -9.69148
+▁ice -9.69166
+▁colleague -9.69187
+ctor -9.69279
+▁fast -9.69319
+▁cab -9.69385
+▁sab -9.69391
+▁thinking -9.69407
+▁interested -9.69484
+▁collect -9.69511
+▁aud -9.69563
+▁deputy -9.69686
+▁teams -9.69732
+▁photograph -9.69876
+oul -9.69933
+▁farmer -9.70032
+▁telling -9.70146
+▁mount -9.70342
+▁literature -9.70352
+▁quebec -9.70352
+▁society -9.70352
+▁complain -9.70353
+▁sky -9.70358
+▁argument -9.70368
+▁croix -9.70385
+▁apartment -9.70394
+▁reform -9.70468
+▁wal -9.7047
+▁mac -9.70554
+▁iron -9.70655
+▁prison -9.70744
+▁trying -9.70757
+aries -9.708
+▁compar -9.70826
+ange -9.70895
+▁financial -9.71024
+▁pot -9.7104
+▁performance -9.71043
+▁quarter -9.71049
+▁sport -9.71373
+ker -9.71533
+▁walls -9.71651
+▁surrounded -9.71699
+▁volume -9.717
+▁represented -9.71704
+oli -9.71851
+▁suddenly -9.71871
+▁cy -9.71962
+ire -9.71984
+ieu -9.72023
+cul -9.72119
+▁min -9.72337
+▁wants -9.72341
+▁broadcast -9.7238
+▁expensive -9.7238
+▁families -9.7238
+▁instrument -9.7238
+▁symbol -9.7238
+▁jump -9.72384
+▁prize -9.72402
+▁estate -9.72439
+▁slowly -9.72572
+▁represent -9.72593
+▁sing -9.72952
+▁discuss -9.72952
+▁arts -9.72967
+▁château -9.73065
+▁impact -9.73066
+▁gabriel -9.73066
+▁electric -9.73077
+▁price -9.73115
+▁objective -9.73116
+▁native -9.73155
+▁cook -9.73158
+▁richard -9.73169
+illy -9.73245
+ability -9.73282
+▁flag -9.73408
+▁bourg -9.73467
+▁letters -9.73476
+▁serious -9.73477
+▁grave -9.73529
+▁drive -9.73586
+▁via -9.7368
+▁episode -9.73754
+▁concept -9.73754
+▁maurice -9.73755
+▁flor -9.73814
+▁cave -9.74126
+▁youth -9.74156
+▁lover -9.74233
+cor -9.74418
+▁budget -9.74449
+▁slightly -9.74449
+▁breath -9.74451
+▁understood -9.74494
+▁standard -9.74512
+▁direct -9.74516
+ama -9.74562
+▁divided -9.74569
+ib -9.7464
+eux -9.74714
+▁forces -9.74762
+▁leading -9.74885
+▁designed -9.74893
+▁disappeared -9.74893
+coming -9.75022
+▁wh -9.75085
+▁ker -9.75116
+▁equipped -9.75148
+▁exchange -9.75148
+▁marseille -9.75148
+▁canadian -9.75148
+▁rob -9.75161
+▁firm -9.75166
+▁perform -9.75185
+ek -9.75186
+▁holding -9.75268
+▁huge -9.75297
+▁passion -9.75333
+▁vau -9.754
+▁settle -9.75403
+▁linked -9.75409
+▁brave -9.75427
+▁lock -9.75546
+eaux -9.75624
+▁institution -9.75659
+rel -9.75733
+▁media -9.75746
+▁belgium -9.75852
+▁purpose -9.75852
+▁tournament -9.75852
+▁otherwise -9.75852
+▁initial -9.75856
+▁needed -9.75859
+▁fau -9.75865
+ache -9.76105
+▁abandoned -9.7612
+▁data -9.76178
+▁entirely -9.76289
+▁recognized -9.76433
+▁aux -9.7645
+▁provision -9.76561
+▁attention -9.76562
+▁either -9.76562
+▁weapon -9.76562
+▁actress -9.76563
+▁library -9.76573
+▁edge -9.76575
+▁sector -9.766
+▁continue -9.76608
+▁rate -9.76628
+bou -9.76711
+▁million -9.76748
+▁dies -9.76763
+▁rise -9.76874
+▁flow -9.76943
+▁dar -9.76946
+▁lesson -9.77111
+▁sac -9.77114
+▁appeared -9.77181
+▁eating -9.77211
+▁mau -9.77254
+ification -9.77267
+▁amount -9.77276
+▁concert -9.77276
+▁policy -9.77276
+▁allowed -9.77276
+▁thomas -9.77278
+▁hurt -9.77287
+▁airport -9.77292
+▁finance -9.77293
+▁management -9.77341
+▁simon -9.7739
+har -9.77495
+bro -9.77501
+▁dame -9.77606
+▁enable -9.77616
+▁certainly -9.77618
+▁regular -9.77702
+▁whe -9.77813
+▁gone -9.77907
+▁replace -9.77911
+▁flowers -9.77979
+▁negative -9.77995
+▁succeeded -9.78028
+▁coll -9.78114
+▁cul -9.78212
+▁pair -9.78226
+▁offered -9.7823
+▁cher -9.78253
+▁tail -9.78291
+▁bare -9.78316
+▁bag -9.78341
+▁formed -9.78379
+▁semi -9.78499
+mé -9.78642
+▁spirit -9.78713
+▁industrial -9.7872
+▁institute -9.7872
+▁colonel -9.7872
+▁block -9.78827
+▁chart -9.78836
+▁rare -9.78892
+uz -9.78925
+▁cru -9.78971
+▁contains -9.79038
+cia -9.79095
+▁registered -9.79112
+▁aims -9.79144
+▁bau -9.79198
+▁thi -9.79255
+▁indian -9.79272
+▁ou -9.79285
+▁gun -9.7929
+▁clearly -9.79309
+▁assistant -9.7945
+▁statue -9.79452
+▁housing -9.79478
+though -9.79492
+▁dress -9.7952
+▁queen -9.79542
+▁favorable -9.79544
+▁chance -9.79592
+▁chat -9.79668
+▁hide -9.79768
+▁prisoner -9.79786
+▁video -9.79825
+ica -9.79869
+ara -9.7989
+▁noise -9.79904
+▁pieces -9.80047
+▁kids -9.80175
+▁impossible -9.80185
+▁effort -9.80185
+▁century -9.80185
+▁snow -9.80186
+▁memory -9.80188
+▁editorial -9.80195
+▁cousin -9.80197
+▁piano -9.80211
+▁sense -9.80217
+▁quiet -9.80238
+▁successful -9.80241
+▁cities -9.80269
+▁unfortunately -9.80342
+▁lucien -9.80449
+▁ros -9.80497
+gor -9.80498
+▁scar -9.805
+old -9.8067
+▁mentioned -9.8082
+▁coat -9.80862
+▁windows -9.80881
+▁restaurant -9.80926
+▁across -9.80926
+▁surface -9.80926
+▁medicine -9.80926
+▁enemy -9.80927
+▁focus -9.80928
+▁crazy -9.8094
+▁produce -9.80976
+▁notably -9.81091
+▁girls -9.81201
+▁surprise -9.81202
+▁listening -9.81383
+▁musical -9.81468
+▁proposed -9.81497
+▁mid -9.81648
+ani -9.81651
+▁terms -9.81665
+▁difference -9.81672
+▁entrance -9.81672
+▁principle -9.81672
+▁recommend -9.81672
+▁search -9.81677
+▁appearance -9.81691
+▁palace -9.81697
+lais -9.81772
+▁signed -9.81933
+ute -9.81935
+▁ze -9.81955
+▁organize -9.81972
+▁save -9.82192
+▁catch -9.82229
+▁dry -9.824
+▁coffee -9.82424
+▁lieutenant -9.82424
+▁residence -9.82424
+▁nobody -9.82424
+▁studio -9.82424
+▁length -9.82424
+ever -9.82456
+▁clothes -9.82517
+▁feet -9.82612
+▁poly -9.82671
+▁tea -9.82686
+zar -9.82786
+▁christ -9.82794
+▁za -9.82912
+▁ur -9.82935
+▁vol -9.83092
+▁completed -9.83128
+▁everywhere -9.83181
+▁excellent -9.83181
+▁proposal -9.83181
+▁worry -9.83182
+▁clean -9.83265
+▁treat -9.83285
+▁producer -9.83412
+▁boat -9.83565
+uri -9.83598
+▁flo -9.8361
+▁pet -9.83611
+▁ahead -9.83781
+▁patric -9.83826
+▁adult -9.83947
+▁greek -9.83962
+▁architecture -9.83975
+▁dor -9.83988
+pper -9.84155
+part -9.84233
+▁ren -9.84257
+▁boss -9.84306
+▁jan -9.84321
+▁mono -9.84357
+▁vice -9.84401
+▁topic -9.84405
+▁sy -9.84569
+▁islands -9.8459
+▁specializ -9.84591
+▁display -9.84714
+▁chinese -9.84715
+▁medical -9.8472
+▁louise -9.84721
+▁murder -9.84733
+zan -9.84803
+rop -9.84804
+▁lan -9.8502
+▁introduced -9.85111
+▁range -9.85231
+row -9.85349
+▁guillaume -9.85489
+▁culture -9.85491
+▁lines -9.85518
+▁balance -9.85532
+▁metal -9.85576
+▁berlin -9.85647
+▁throw -9.85735
+▁regions -9.85881
+▁pain -9.8592
+▁george -9.85948
+▁rat -9.86132
+▁holiday -9.8627
+▁olympic -9.8627
+▁easily -9.86271
+▁belgian -9.86271
+▁taught -9.86271
+▁psych -9.86272
+▁honest -9.86273
+▁crossed -9.86277
+▁image -9.86298
+gues -9.86402
+▁discussion -9.86408
+oy -9.86445
+▁mouth -9.86487
+▁specie -9.86531
+▁pé -9.86739
+▁higher -9.86779
+▁defended -9.86875
+▁obtained -9.86896
+gh -9.869
+ku -9.86979
+▁bottom -9.87058
+▁california -9.87058
+▁accident -9.87058
+▁speech -9.87058
+▁portrait -9.87058
+▁visible -9.8706
+▁demand -9.87075
+▁benches -9.87157
+▁lit -9.87204
+▁bour -9.87289
+▁dogs -9.87303
+yr -9.87383
+▁appear -9.87439
+com -9.87499
+▁problems -9.87535
+▁filled -9.8762
+▁refused -9.87768
+▁forced -9.87785
+▁cote -9.8781
+▁tian -9.87833
+▁gou -9.87848
+▁fund -9.87849
+▁responsible -9.87851
+▁scientific -9.87851
+▁neither -9.87852
+▁sunday -9.87852
+▁constant -9.87852
+▁caught -9.87852
+▁opposition -9.87854
+▁shoot -9.87883
+▁ruin -9.87934
+▁defeat -9.87958
+▁grey -9.87986
+▁zone -9.88043
+▁roche -9.88054
+vier -9.881
+▁loud -9.88163
+▁arrive -9.88183
+▁dare -9.88234
+▁clo -9.8827
+▁corner -9.88335
+▁elephants -9.88418
+▁tourist -9.88435
+▁reign -9.88573
+▁wide -9.88577
+hy -9.88624
+▁docteur -9.88651
+▁procedure -9.88651
+▁victim -9.88652
+▁resistance -9.88663
+▁pur -9.88678
+ella -9.88698
+▁passage -9.88721
+▁mur -9.88724
+▁china -9.88838
+▁mention -9.88911
+▁fro -9.88937
+dé -9.88993
+▁apart -9.89088
+▁paintings -9.8927
+▁dad -9.89337
+▁auguste -9.89387
+bon -9.89456
+▁campaign -9.89458
+▁picture -9.89458
+▁credit -9.89461
+▁attend -9.89466
+▁joy -9.89502
+▁careful -9.89518
+▁absolutely -9.89544
+qua -9.89572
+▁multi -9.8967
+▁simply -9.89692
+▁seine -9.89841
+gny -9.89844
+▁refuse -9.90011
+▁peter -9.90042
+back -9.90125
+▁areas -9.90129
+▁pri -9.90132
+▁phil -9.90244
+▁application -9.90271
+▁diocese -9.90271
+▁energy -9.90271
+▁channel -9.90272
+▁abbot -9.90273
+▁worse -9.90274
+▁property -9.90284
+▁height -9.90291
+▁learned -9.90377
+phe -9.90401
+▁fun -9.90488
+▁standing -9.90602
+▁noble -9.90606
+▁marcel -9.90757
+▁marry -9.90788
+▁feed -9.90789
+port -9.90931
+▁adventure -9.91091
+▁stories -9.91091
+▁closer -9.91099
+▁knee -9.91107
+▁records -9.91127
+▁fault -9.91169
+▁spring -9.91172
+east -9.9119
+▁ubu -9.91206
+▁expression -9.91276
+illard -9.91334
+▁realize -9.91421
+▁invited -9.91479
+▁condition -9.91586
+▁settled -9.91642
+▁starting -9.91675
+chy -9.91775
+▁tam -9.91849
+▁peace -9.91899
+▁advance -9.91917
+▁agriculture -9.91917
+▁disease -9.91917
+▁nephew -9.91917
+▁mini -9.91917
+▁defense -9.91917
+▁stream -9.91918
+▁freedom -9.91918
+▁swiss -9.91923
+▁william -9.91931
+▁favorite -9.91949
+▁journalist -9.92036
+▁build -9.92068
+▁workers -9.92113
+▁musician -9.92171
+▁label -9.92214
+▁ger -9.92301
+▁delete -9.92314
+▁sum -9.92334
+▁ven -9.92446
+ax -9.92452
+▁triple -9.92455
+▁gain -9.92476
+▁sixth -9.92543
+▁equ -9.92561
+▁fou -9.92642
+▁excuse -9.92751
+▁ocean -9.92752
+▁software -9.92755
+▁conduct -9.92757
+▁founder -9.92779
+▁helped -9.92793
+▁drama -9.92823
+▁expect -9.92896
+▁tin -9.9291
+▁eli -9.92914
+▁covers -9.93013
+▁fat -9.93023
+▁cent -9.9312
+ignac -9.93196
+▁spa -9.93247
+phi -9.93263
+war -9.93346
+▁branch -9.93392
+fu -9.93469
+▁suit -9.93538
+▁prime -9.93542
+▁liber -9.93547
+stood -9.93563
+▁crowd -9.93592
+▁distance -9.93592
+▁self -9.93593
+▁protestant -9.93615
+ju -9.93644
+▁rush -9.93714
+▁luck -9.93864
+▁chau -9.9393
+▁reserve -9.93966
+▁engine -9.94032
+hon -9.94107
+▁crossing -9.94434
+▁stadium -9.94438
+▁elvire -9.94438
+▁positive -9.94439
+▁troops -9.9444
+▁session -9.94441
+▁tears -9.94472
+▁minor -9.94584
+▁meaning -9.94607
+▁exam -9.94677
+▁léon -9.94695
+▁bl -9.94752
+▁shortly -9.94774
+▁file -9.94815
+▁cer -9.94861
+▁ago -9.94901
+▁eu -9.95028
+▁strange -9.95085
+▁participated -9.95141
+▁pav -9.9524
+▁employ -9.9524
+▁circle -9.95293
+▁contact -9.95293
+▁brazil -9.95293
+▁christophe -9.95293
+▁gilles -9.9533
+▁pilot -9.95352
+▁finger -9.95359
+▁skill -9.95417
+▁follows -9.95423
+▁exists -9.95443
+▁extend -9.95462
+▁bears -9.95496
+▁moulin -9.95503
+▁foundation -9.9555
+▁burn -9.9558
+gna -9.95637
+▁chosen -9.95653
+▁concerns -9.95653
+▁eva -9.95704
+▁arc -9.95737
+ken -9.95796
+▁accepted -9.95865
+▁lar -9.95866
+jou -9.95913
+▁mai -9.96059
+▁recognize -9.96071
+▁baron -9.96107
+▁anna -9.96107
+▁guide -9.9613
+ologist -9.96139
+▁empty -9.96155
+▁location -9.96156
+▁shut -9.9621
+▁recorded -9.96231
+▁concerned -9.96281
+▁trust -9.96296
+▁obtain -9.96332
+▁paid -9.96385
+▁continues -9.96394
+▁euros -9.9641
+▁blu -9.96431
+matic -9.96478
+▁print -9.96506
+▁publication -9.96749
+▁ski -9.96754
+ograph -9.96783
+▁gé -9.96809
+▁latin -9.9688
+▁pray -9.96968
+▁economy -9.97025
+▁opportunity -9.97025
+▁étienne -9.97025
+▁facade -9.97025
+▁interior -9.97025
+▁childhood -9.97036
+▁removed -9.97039
+▁speed -9.97045
+▁eugen -9.9705
+▁sole -9.97053
+▁lift -9.97105
+▁protected -9.97108
+▁founded -9.97374
+▁mary -9.97377
+▁suffer -9.97401
+▁score -9.97721
+▁plants -9.97758
+▁kar -9.97886
+▁review -9.97902
+▁composition -9.97902
+▁workshop -9.97902
+▁reveal -9.97905
+▁vin -9.97943
+▁powerful -9.97976
+▁esp -9.98012
+niere -9.98039
+▁desire -9.98141
+ieux -9.98223
+▁dro -9.98302
+▁studying -9.98399
+▁mixed -9.98427
+bé -9.98475
+▁fin -9.98483
+▁hou -9.98485
+▁essential -9.98488
+▁enters -9.98672
+lk -9.98726
+▁wave -9.98781
+▁afraid -9.98787
+▁agricultural -9.98787
+▁contrary -9.98787
+▁draft -9.98787
+▁roof -9.9879
+▁eastern -9.98803
+▁shock -9.98837
+▁fool -9.98905
+▁staff -9.98909
+▁pal -9.98917
+▁value -9.98918
+mann -9.99005
+▁billion -9.99039
+▁australia -9.99129
+rier -9.99146
+▁separated -9.99154
+▁cle -9.992
+▁cry -9.99227
+▁decides -9.99236
+ula -9.99251
+▁bother -9.99253
+q -9.99307
+▁weight -9.99413
+ple -9.99529
+▁weak -9.99603
+▁talked -9.99629
+▁tru -9.99647
+▁hurry -9.9968
+▁regret -9.99681
+▁phase -9.99682
+▁justice -9.99683
+▁awarded -9.99714
+▁elements -9.99766
+kin -9.9984
+ima -9.99863
+▁flows -9.99914
+▁taste -9.99922
+▁associated -9.99991
+▁dé -10.0003
+▁check -10.0007
+▁weeks -10.0011
+▁pat -10.0026
+▁tradition -10.0041
+▁recent -10.0043
+▁var -10.0056
+▁administration -10.0058
+▁instead -10.0058
+▁anyone -10.0059
+▁sailor -10.0075
+▁golden -10.0078
+▁calling -10.0092
+▁manager -10.0096
+▁gal -10.0098
+▁row -10.0101
+▁vent -10.0106
+etic -10.0119
+▁begins -10.0124
+west -10.0141
+▁marks -10.0141
+▁mess -10.0145
+▁puy -10.0145
+▁challenge -10.0149
+▁smell -10.0149
+▁lunch -10.015
+▁funny -10.0151
+▁percent -10.0155
+▁carrier -10.0155
+▁provide -10.0156
+▁highly -10.0156
+▁steps -10.0157
+▁bains -10.0158
+▁temple -10.0168
+▁prepar -10.0168
+▁chamber -10.0182
+▁directly -10.0188
+▁dia -10.019
+▁flower -10.019
+▁trou -10.0197
+▁haut -10.0204
+▁oil -10.0206
+▁shin -10.0238
+▁performed -10.0239
+▁attempt -10.0241
+▁opposite -10.0241
+▁twice -10.0241
+cou -10.0242
+▁flee -10.0245
+▁convention -10.0245
+▁authorities -10.0246
+▁drum -10.0248
+▁minute -10.0257
+▁enter -10.0261
+▁hor -10.0262
+▁owe -10.0274
+▁maintained -10.0274
+▁trained -10.0276
+▁wild -10.0281
+▁frame -10.0281
+▁syn -10.0295
+▁india -10.0296
+emp -10.0308
+▁receives -10.0332
+▁accompanied -10.0333
+▁forbidden -10.0333
+▁magazine -10.0333
+▁switzerland -10.0333
+▁stretch -10.0333
+▁target -10.0333
+▁ensure -10.0335
+▁marked -10.034
+▁ku -10.034
+▁exhibition -10.035
+▁roger -10.035
+isse -10.0353
+▁occupied -10.0359
+aine -10.0361
+▁bomb -10.0378
+▁fixed -10.0388
+▁pou -10.0389
+▁features -10.0394
+▁tools -10.0415
+▁bigge -10.0421
+▁danger -10.0426
+▁october -10.0427
+▁attached -10.0427
+▁diploma -10.0427
+▁refuge -10.0427
+▁climate -10.0427
+▁achieve -10.0427
+▁average -10.0427
+▁dinner -10.0428
+▁winner -10.043
+▁rural -10.043
+▁terrible -10.0437
+▁yo -10.0438
+▁highest -10.0441
+▁blow -10.0449
+▁adopt -10.0453
+▁lor -10.046
+lip -10.0461
+▁hence -10.047
+▁upper -10.0474
+▁operate -10.0485
+▁plate -10.0486
+▁commander -10.0486
+▁provided -10.0497
+▁lang -10.0498
+▁beach -10.0506
+▁provides -10.0516
+▁tomb -10.0521
+▁hamlet -10.0525
+▁polish -10.0526
+▁comment -10.0529
+▁yu -10.0532
+cq -10.0535
+▁translated -10.0536
+▁walking -10.0538
+▁onto -10.0543
+▁gap -10.0565
+▁skin -10.0592
+▁zo -10.0594
+wer -10.0596
+ola -10.0612
+▁occur -10.0615
+mber -10.0616
+▁injured -10.0616
+▁quality -10.0616
+▁raymond -10.0617
+▁sorry -10.0617
+▁egg -10.0617
+▁ladies -10.0617
+▁crown -10.0619
+▁correct -10.062
+▁wake -10.0625
+▁believed -10.0628
+not -10.063
+▁preserved -10.0635
+ano -10.0639
+▁perfectly -10.0646
+▁sail -10.0652
+▁teaching -10.0653
+olo -10.066
+var -10.0661
+▁adopted -10.0687
+▁sin -10.0702
+▁african -10.0703
+▁withdrawal -10.0705
+▁municipal -10.0706
+▁virgin -10.0708
+▁happiness -10.0712
+▁monastery -10.0712
+▁arrival -10.0712
+▁mexico -10.0712
+▁traffic -10.0713
+▁choose -10.0713
+▁legend -10.0713
+▁contest -10.0713
+▁daily -10.0713
+▁gate -10.0717
+▁selection -10.0718
+▁henry -10.072
+▁effective -10.0721
+▁corn -10.0735
+bach -10.0741
+▁moral -10.0746
+▁quin -10.0747
+▁chan -10.0768
+▁perfect -10.078
+▁bedroom -10.081
+▁custom -10.081
+▁primary -10.0811
+▁governor -10.0812
+▁peak -10.0819
+▁gray -10.082
+▁alain -10.0826
+▁grass -10.0827
+▁flu -10.0833
+▁chest -10.0853
+ick -10.0854
+▁fill -10.0872
+▁older -10.0874
+▁repeated -10.0875
+▁tho -10.0888
+bri -10.0891
+kh -10.0895
+work -10.0902
+gno -10.0907
+▁candidate -10.0908
+▁bottle -10.0908
+gging -10.0909
+▁magn -10.091
+▁bert -10.0914
+▁max -10.0917
+▁exactly -10.0923
+▁yve -10.093
+▁orange -10.0934
+▁referr -10.0938
+▁qui -10.094
+▁release -10.096
+▁advis -10.0964
+▁ty -10.0975
+izing -10.098
+hor -10.0987
+gli -10.0992
+rich -10.1001
+▁principal -10.1007
+▁april -10.1007
+▁knowledge -10.1007
+▁wolf -10.1007
+▁fontaine -10.1008
+▁damage -10.1012
+▁domain -10.1013
+▁laughing -10.102
+▁younger -10.1029
+▁republican -10.1038
+off -10.1056
+▁relie -10.1057
+▁citizens -10.1061
+▁notice -10.1068
+▁painted -10.1071
+▁hum -10.1082
+▁suffered -10.1085
+ova -10.1098
+▁maria -10.1106
+▁independence -10.1107
+▁ourselves -10.1107
+▁philosophy -10.1107
+▁jewish -10.1107
+▁pencil -10.1107
+▁pleasure -10.1107
+▁physical -10.1107
+▁witness -10.1107
+▁trouble -10.1108
+▁severe -10.111
+▁sort -10.1116
+▁connect -10.1122
+▁showing -10.1131
+▁manner -10.1138
+▁maur -10.1142
+▁plain -10.1147
+▁invite -10.1162
+ford -10.1183
+▁chant -10.1183
+ignan -10.119
+▁tries -10.1195
+▁scared -10.1201
+atic -10.1202
+chu -10.1207
+▁paragraph -10.1208
+▁phenomenon -10.1208
+▁satisfied -10.1208
+▁knock -10.1208
+▁historian -10.1208
+▁resign -10.1208
+▁importance -10.1208
+▁vincent -10.1208
+▁cloud -10.1209
+▁below -10.1211
+▁might -10.1211
+▁silver -10.1211
+▁kilometers -10.1218
+▁sales -10.1221
+▁hello -10.1226
+ua -10.1233
+▁authority -10.1234
+▁announc -10.124
+▁alive -10.1246
+▁drop -10.1255
+▁ferr -10.1256
+▁ken -10.1256
+▁fighting -10.1258
+▁por -10.1266
+▁warm -10.1267
+mond -10.1269
+ppe -10.1271
+▁raise -10.1284
+▁haute -10.1286
+ami -10.1292
+▁increased -10.1306
+▁celebrate -10.131
+▁toulouse -10.131
+▁railway -10.131
+▁beyond -10.131
+▁knight -10.131
+▁fernand -10.131
+▁afternoon -10.131
+▁comprise -10.131
+▁fresh -10.132
+▁jule -10.1321
+▁desert -10.1325
+▁boul -10.1328
+▁obviously -10.1331
+bal -10.1389
+▁admit -10.1411
+▁temperature -10.1413
+▁gérard -10.1413
+▁peasant -10.1413
+▁suspect -10.1413
+▁violent -10.1413
+▁platform -10.1413
+▁generation -10.1413
+▁status -10.1413
+▁strip -10.1413
+▁establishment -10.1417
+▁courage -10.142
+▁sim -10.1427
+▁jack -10.1431
+▁opposed -10.1434
+▁rent -10.1442
+▁selected -10.1445
+▁mag -10.1446
+▁repeat -10.1448
+▁composer -10.1453
+▁terre -10.1459
+▁nation -10.1476
+▁ul -10.1497
+▁damn -10.1507
+▁philip -10.1516
+▁vern -10.1516
+▁technical -10.1517
+▁vehicle -10.1517
+▁algeria -10.1518
+▁obey -10.1521
+▁stock -10.1521
+▁existence -10.1521
+▁earlier -10.1521
+lio -10.1522
+▁recording -10.1525
+▁sous -10.1533
+▁commissioner -10.1535
+▁greatest -10.1538
+▁thé -10.1541
+▁products -10.1544
+▁classes -10.156
+▁rouss -10.1564
+▁glasses -10.1564
+▁failed -10.1574
+▁vic -10.1578
+lant -10.158
+▁slow -10.1598
+▁except -10.1609
+▁cass -10.1618
+▁worried -10.1622
+▁patient -10.1622
+▁expert -10.1622
+▁sweet -10.1622
+▁renew -10.1622
+▁smoke -10.1622
+▁labor -10.1623
+▁picard -10.1625
+▁bright -10.1626
+▁internet -10.1628
+▁james -10.1628
+▁silent -10.1628
+▁treatment -10.1629
+▁font -10.1632
+anne -10.1645
+▁wedding -10.1653
+elli -10.1654
+▁mom -10.1691
+▁pun -10.1692
+▁trial -10.1708
+ight -10.1721
+▁saved -10.1722
+▁threat -10.1729
+▁aspect -10.1729
+▁experiment -10.1729
+▁technique -10.1729
+▁tributary -10.1729
+▁suppli -10.1729
+▁everyday -10.1729
+ock -10.1736
+▁hydro -10.1736
+▁shoes -10.1748
+▁meters -10.1771
+cker -10.1771
+▁june -10.1773
+une -10.18
+asse -10.1808
+▁carbon -10.181
+▁mel -10.1824
+▁details -10.1833
+▁typical -10.1836
+▁drunk -10.1836
+▁pretend -10.1836
+▁facing -10.1837
+▁failure -10.1838
+▁income -10.184
+▁task -10.1842
+▁learning -10.1849
+▁reg -10.1867
+▁disappear -10.1873
+▁secondary -10.1879
+▁interesting -10.1911
+▁parties -10.1915
+▁july -10.1918
+▁element -10.1934
+▁employees -10.1942
+▁quick -10.1942
+▁faith -10.1943
+▁atmosphere -10.1945
+▁decrease -10.1945
+▁september -10.1945
+▁voted -10.1945
+▁qualified -10.1945
+▁catherine -10.1945
+▁decorated -10.1945
+▁quentin -10.1945
+▁decree -10.1945
+▁annoy -10.1946
+▁characteristic -10.1946
+▁classified -10.1947
+▁flight -10.1949
+right -10.1952
+▁epi -10.1956
+▁protection -10.1959
+▁motor -10.1962
+▁declared -10.1966
+▁hol -10.2034
+kov -10.204
+▁sho -10.2049
+▁religion -10.2055
+▁emotion -10.2055
+▁senate -10.2055
+▁retain -10.2055
+▁security -10.2058
+▁wheel -10.2068
+time -10.2069
+▁useful -10.207
+▁mc -10.2083
+▁liberal -10.2084
+▁arrested -10.2092
+▁establish -10.2094
+▁hang -10.2103
+▁breed -10.2105
+▁extra -10.2107
+▁pushed -10.2117
+▁salt -10.212
+▁temp -10.2121
+▁broken -10.215
+▁strength -10.2153
+▁potential -10.2166
+▁brussels -10.2166
+▁explanation -10.2166
+▁global -10.2166
+▁olivier -10.2166
+▁evolution -10.2167
+▁intend -10.217
+▁korea -10.2175
+▁blanc -10.2195
+▁dig -10.2205
+▁baby -10.2205
+▁rap -10.2212
+▁compete -10.2218
+▁jou -10.2221
+▁confirmed -10.2222
+▁russia -10.2223
+▁affected -10.2225
+▁influenced -10.2266
+while -10.2273
+▁encourage -10.2278
+▁mechanism -10.2278
+▁reinforce -10.2278
+▁recruit -10.2278
+▁ticket -10.2278
+▁instruct -10.2278
+▁ernest -10.2278
+▁absence -10.2278
+▁appeal -10.2279
+▁supposed -10.229
+▁deriv -10.2291
+▁missing -10.2298
+▁opera -10.2308
+▁sleeping -10.2323
+▁stable -10.2327
+▁politic -10.2344
+▁separate -10.2345
+▁sheet -10.2347
+▁bol -10.2351
+ida -10.2352
+▁monte -10.2353
+itz -10.2369
+▁watching -10.2372
+▁journal -10.2375
+▁dressed -10.2385
+field -10.2387
+▁entrust -10.2392
+▁landscape -10.2392
+▁observation -10.2392
+▁sacrifice -10.2392
+▁alfred -10.2392
+▁federation -10.2392
+▁stupid -10.2392
+▁monday -10.2392
+▁screen -10.2392
+▁storm -10.2393
+▁ajac -10.2393
+▁dutch -10.2394
+▁growth -10.2408
+▁alexandre -10.2417
+▁happening -10.2435
+▁bow -10.244
+nois -10.2441
+▁none -10.2443
+▁factor -10.2447
+isle -10.2452
+rine -10.2467
+▁nave -10.2477
+▁signal -10.248
+▁constitute -10.2507
+▁furniture -10.2507
+▁orchestra -10.2507
+▁implement -10.2507
+▁jealous -10.2507
+▁heritage -10.2507
+▁brief -10.2507
+▁proceed -10.2507
+▁angry -10.2508
+▁dangerous -10.2511
+house -10.2516
+▁reaction -10.2517
+▁soil -10.2521
+▁tourism -10.2523
+bois -10.2528
+▁driver -10.2546
+▁ranking -10.2555
+▁que -10.256
+▁paint -10.2561
+▁pink -10.2569
+▁stu -10.258
+▁classic -10.2601
+▁simpl -10.2602
+▁mix -10.2607
+light -10.2618
+▁liv -10.2618
+▁directed -10.2622
+▁applause -10.2623
+▁conflict -10.2623
+▁xavier -10.2623
+▁eventually -10.2623
+gé -10.2623
+▁arnaud -10.2623
+▁shelter -10.2624
+▁content -10.2624
+▁smith -10.2627
+▁publish -10.2629
+roux -10.263
+▁marius -10.2633
+▁strongly -10.2638
+▁bringing -10.264
+▁meat -10.265
+▁legion -10.2653
+▁justin -10.2665
+▁loo -10.2671
+▁magic -10.2672
+▁bab -10.2676
+▁von -10.2676
+▁tale -10.268
+mark -10.2681
+▁passes -10.269
+▁bern -10.2691
+▁sudden -10.2712
+▁fix -10.2712
+▁exhibit -10.2719
+nais -10.2725
+▁defeated -10.2725
+▁stan -10.2725
+▁resident -10.2741
+▁conversation -10.2741
+▁sword -10.2741
+▁stairs -10.2741
+▁planned -10.2743
+▁tardy -10.2744
+▁stick -10.2746
+▁relation -10.2748
+▁expressed -10.2749
+▁hearing -10.275
+▁unique -10.2751
+▁precisely -10.2761
+hol -10.2765
+▁withdrawn -10.277
+▁cé -10.2787
+▁chap -10.2793
+oire -10.2819
+▁announced -10.2821
+▁marne -10.2831
+uge -10.2833
+pha -10.2837
+▁cool -10.2857
+▁perhaps -10.286
+▁responsibility -10.286
+▁cardinal -10.286
+▁talent -10.2861
+▁sharp -10.2861
+▁argent -10.2865
+▁clu -10.2871
+▁modest -10.2872
+▁gor -10.2874
+▁cried -10.288
+stein -10.2884
+▁honorable -10.2885
+ivity -10.2891
+evre -10.2893
+▁reward -10.2895
+▁begin -10.2896
+ché -10.2908
+▁grant -10.2916
+bonne -10.2918
+▁sul -10.2928
+▁gil -10.2941
+▁pér -10.2951
+▁extended -10.2956
+▁demo -10.2957
+▁ji -10.2958
+▁sala -10.2971
+▁consequence -10.298
+▁executive -10.298
+▁bordeaux -10.298
+▁michael -10.298
+▁regulation -10.298
+▁terror -10.298
+▁ray -10.2983
+▁sugar -10.2983
+▁israel -10.2987
+▁detect -10.2991
+▁crime -10.2993
+▁shame -10.2999
+▁thin -10.3004
+▁vert -10.3009
+▁recently -10.302
+▁brive -10.3022
+▁classical -10.3028
+▁accus -10.3033
+▁crosses -10.3051
+▁lille -10.3053
+▁pour -10.306
+▁listed -10.3084
+▁monk -10.3087
+▁possess -10.3089
+▁regime -10.3102
+▁capacity -10.3102
+▁soviet -10.3102
+▁repair -10.3102
+▁slave -10.3104
+baptiste -10.3112
+▁serge -10.3116
+▁aunt -10.3117
+▁corps -10.3118
+▁dominat -10.3123
+▁specialist -10.3134
+▁emil -10.3137
+▁factory -10.3145
+▁vé -10.3149
+range -10.3157
+▁citizen -10.3158
+▁basic -10.3167
+▁milk -10.3169
+▁push -10.317
+▁spoke -10.3171
+▁agent -10.3175
+▁neo -10.3176
+▁connected -10.3187
+▁sight -10.3199
+▁trail -10.3219
+▁consequent -10.3226
+▁gradually -10.3226
+▁opponent -10.3226
+▁column -10.3226
+▁proof -10.3226
+▁blind -10.3226
+▁solid -10.3228
+gent -10.3228
+▁background -10.3229
+▁gift -10.3229
+▁editor -10.3236
+▁bio -10.3239
+▁marine -10.3242
+▁trace -10.3255
+▁bored -10.3269
+▁reject -10.3282
+▁successor -10.3282
+▁aware -10.3284
+▁eighth -10.3292
+▁solve -10.3297
+▁laval -10.33
+▁ol -10.331
+▁edit -10.3338
+▁approximately -10.3351
+▁conference -10.3351
+▁equipment -10.3351
+▁negotiation -10.3351
+▁occupation -10.3351
+▁prohibit -10.3351
+▁ministry -10.3351
+▁tropical -10.3351
+▁incident -10.3351
+▁pocket -10.3351
+▁courtyard -10.3351
+▁decline -10.3351
+▁slope -10.3351
+▁cyr -10.3353
+▁fellow -10.3355
+▁twin -10.3355
+▁investigation -10.3356
+▁proper -10.3359
+hr -10.3366
+▁agreed -10.3366
+▁lips -10.3368
+▁wonder -10.3374
+▁artistic -10.3374
+▁dubourg -10.3375
+▁bull -10.3377
+board -10.3381
+▁greet -10.3385
+▁din -10.3395
+▁described -10.3408
+▁falling -10.3415
+▁rejected -10.342
+▁previously -10.3427
+▁officially -10.3433
+▁parisian -10.3434
+▁zi -10.3445
+▁entering -10.3445
+▁naturally -10.3447
+▁gall -10.3447
+▁seventh -10.3448
+etti -10.3461
+▁construct -10.3467
+▁introduce -10.3475
+▁basketball -10.3477
+▁contemporary -10.3477
+▁netherlands -10.3477
+▁variety -10.3477
+▁adjust -10.3477
+▁departure -10.3477
+▁continent -10.3477
+▁decoration -10.3477
+▁gentlemen -10.3478
+▁vign -10.3479
+▁hole -10.348
+▁chateau -10.3481
+▁nearby -10.3483
+▁émil -10.3484
+▁literar -10.3486
+▁poem -10.3488
+▁lamp -10.3494
+▁decide -10.3494
+▁representation -10.3495
+▁lev -10.3503
+well -10.3507
+▁growing -10.3515
+▁nail -10.3524
+▁deny -10.3525
+▁hunter -10.3533
+▁register -10.3536
+▁basin -10.3541
+▁raining -10.3542
+▁required -10.3577
+ware -10.3594
+▁delay -10.3595
+▁fail -10.3596
+▁tie -10.3602
+▁disturb -10.3605
+▁manuscript -10.3605
+▁surrounding -10.3605
+▁triumph -10.3605
+▁november -10.3605
+▁december -10.3605
+▁driving -10.3605
+▁shadow -10.3605
+▁copies -10.3606
+▁rifle -10.3606
+▁frank -10.3606
+▁pursue -10.3606
+▁sculpture -10.3606
+▁message -10.361
+▁countess -10.3611
+▁guer -10.3611
+▁elder -10.3615
+▁thick -10.3628
+dding -10.3645
+▁hunting -10.3649
+itive -10.3661
+▁nest -10.3663
+▁affect -10.3666
+▁austria -10.368
+▁suffering -10.3707
+▁bourbon -10.3716
+▁bapt -10.3725
+▁passing -10.3727
+▁carrie -10.373
+bil -10.3733
+▁surprising -10.3735
+▁apparently -10.3735
+▁extension -10.3735
+▁feather -10.3735
+▁passenger -10.3735
+▁junior -10.3735
+▁properties -10.3735
+▁gaillarde -10.3735
+▁tank -10.3736
+▁micro -10.3736
+▁internal -10.3737
+▁wealth -10.3737
+▁fleur -10.3738
+▁guess -10.3741
+▁parliamentary -10.3742
+▁grande -10.3743
+▁drug -10.3745
+▁claire -10.3757
+▁restored -10.3763
+▁angeles -10.3782
+▁logic -10.3784
+aye -10.3798
+▁pel -10.3809
+▁taxes -10.3813
+▁indicate -10.3821
+▁bonne -10.3824
+atory -10.3826
+▁pope -10.3831
+▁prepared -10.3835
+▁barbe -10.3837
+▁greatly -10.3839
+▁originat -10.3845
+▁drawn -10.385
+▁attribut -10.3867
+▁installed -10.3867
+▁nonetheless -10.3867
+▁clarification -10.3867
+▁committed -10.3867
+▁parallel -10.3867
+▁pascal -10.3867
+▁marcof -10.3867
+▁prefecture -10.3869
+▁descendant -10.3869
+▁retirement -10.3869
+▁difficulties -10.388
+▁ye -10.3883
+▁safe -10.3884
+▁whispered -10.3896
+▁essentially -10.3912
+▁contributed -10.3912
+▁spoken -10.3926
+▁launched -10.3931
+zé -10.3938
+▁dome -10.394
+▁compos -10.3951
+stone -10.3958
+▁soldier -10.3966
+▁chose -10.3969
+▁chamb -10.397
+▁shaped -10.3971
+▁strike -10.3983
+▁cape -10.3987
+▁rust -10.3987
+▁guitar -10.3991
+▁congress -10.4
+▁communities -10.4
+▁général -10.4
+▁assassin -10.4
+▁strasbourg -10.4
+▁britain -10.4
+phone -10.4
+▁scholar -10.4001
+enberg -10.4002
+▁serbia -10.4002
+▁accessible -10.4002
+▁holy -10.4004
+▁mail -10.4009
+▁olive -10.4012
+▁bound -10.4016
+▁sale -10.4018
+▁stepp -10.402
+▁fled -10.4024
+▁basis -10.4032
+▁socialist -10.4033
+plan -10.4038
+gged -10.4041
+▁pool -10.4055
+▁apple -10.4086
+▁associate -10.4088
+▁pure -10.4101
+▁euro -10.4128
+▁radi -10.4128
+▁initiat -10.4133
+▁divorce -10.4135
+▁territories -10.4135
+▁reputation -10.4135
+▁valérie -10.4135
+▁federal -10.4135
+▁gesture -10.4135
+▁imperial -10.4135
+▁transformed -10.4135
+▁modified -10.4135
+▁electronic -10.4135
+▁recall -10.4135
+▁crisis -10.4136
+▁comic -10.4138
+▁damien -10.4138
+card -10.4139
+▁option -10.4141
+▁subsequently -10.4142
+graphic -10.4146
+▁poetry -10.4147
+▁fold -10.4148
+▁reflection -10.415
+▁poland -10.4158
+▁defender -10.4159
+▁imagin -10.4168
+▁rio -10.4182
+▁neck -10.4183
+▁till -10.4189
+▁seriously -10.4194
+▁root -10.4198
+fall -10.42
+hou -10.4214
+▁australian -10.4219
+▁product -10.4225
+▁contribute -10.4225
+▁teache -10.4228
+▁photo -10.4233
+▁fit -10.4258
+▁difficulty -10.426
+▁pig -10.4265
+▁dialogue -10.4272
+▁evolve -10.4272
+▁resources -10.4272
+▁sensitive -10.4272
+▁remarkable -10.4272
+▁maxime -10.4272
+▁widow -10.4272
+▁nantes -10.4272
+▁forgotten -10.4273
+▁transfer -10.4274
+villiers -10.4274
+▁furious -10.4275
+▁inspector -10.4275
+▁insult -10.4276
+zac -10.4277
+▁patron -10.428
+▁proud -10.4283
+fen -10.4284
+▁fought -10.4285
+▁eliminated -10.4291
+▁useless -10.4291
+▁colomb -10.4305
+▁branches -10.4305
+▁graduated -10.4316
+lette -10.4321
+▁wise -10.4331
+▁confirm -10.4342
+reux -10.4345
+▁reserved -10.4349
+▁emergenc -10.4352
+▁reduce -10.4365
+▁churches -10.4377
+▁southwest -10.4378
+▁conclude -10.4411
+▁identified -10.4411
+▁criminal -10.4411
+▁keinec -10.4411
+▁expand -10.4411
+▁companion -10.4411
+▁dirty -10.4411
+▁devoted -10.4411
+▁applied -10.4411
+▁vessel -10.4411
+▁spiritual -10.4413
+▁champagne -10.4414
+▁passionate -10.4419
+tally -10.442
+▁resist -10.4421
+▁singing -10.4427
+▁load -10.4428
+▁wil -10.446
+▁steel -10.4462
+▁austrian -10.447
+▁hop -10.4474
+chard -10.4477
+▁earned -10.4481
+▁genre -10.4482
+▁fate -10.4495
+▁earn -10.4498
+▁deposit -10.4524
+▁solo -10.454
+▁alcohol -10.4552
+▁exercise -10.4552
+▁exploit -10.4552
+▁intelligence -10.4552
+▁journey -10.4552
+▁marguerite -10.4552
+▁soccer -10.4552
+▁sovereign -10.4552
+▁website -10.4552
+▁narrow -10.4552
+▁struck -10.4552
+▁plastic -10.4552
+▁charming -10.4553
+▁preced -10.4554
+▁shared -10.4556
+▁rennes -10.4559
+▁client -10.456
+▁pond -10.4562
+▁flour -10.4564
+▁angle -10.4579
+▁theatre -10.4582
+▁exile -10.4583
+lation -10.4588
+▁temper -10.459
+▁delight -10.4591
+▁porte -10.4613
+▁significantly -10.4623
+▁launch -10.4628
+▁jar -10.4636
+lav -10.4636
+▁scor -10.4639
+▁listened -10.4642
+▁busy -10.465
+▁combat -10.4677
+▁headquarter -10.468
+▁interpret -10.4686
+wood -10.4689
+▁michele -10.4691
+▁advice -10.4695
+▁fulfill -10.4695
+▁spokesperson -10.4695
+▁sponsor -10.4695
+▁permanent -10.4695
+▁bronze -10.4695
+hard -10.4696
+▁irish -10.4697
+▁romain -10.4698
+▁nucle -10.4698
+▁florence -10.4699
+▁pop -10.4699
+▁scale -10.47
+▁sitting -10.4704
+▁berthe -10.4705
+▁berger -10.4706
+▁quar -10.4715
+▁chain -10.4724
+▁existing -10.4753
+▁insist -10.4754
+long -10.4761
+vaux -10.4773
+▁historic -10.4774
+▁scho -10.4798
+▁obvious -10.481
+▁expected -10.4814
+▁continuous -10.484
+▁inspiration -10.484
+▁communist -10.484
+▁uniform -10.484
+▁swear -10.484
+▁gallery -10.484
+▁split -10.484
+▁proposition -10.4841
+▁underground -10.4842
+▁mah -10.4845
+▁dragon -10.4846
+▁quay -10.4846
+▁appreciated -10.4865
+▁convinced -10.4882
+▁bigger -10.4882
+▁crew -10.4882
+▁duty -10.4919
+ogen -10.4939
+▁graduate -10.494
+aison -10.4941
+win -10.4956
+▁map -10.4959
+▁profession -10.4963
+▁sud -10.498
+▁treated -10.4981
+▁mode -10.4986
+▁dimension -10.4987
+▁jurisdiction -10.4987
+▁accompany -10.4987
+▁encounter -10.4987
+▁mediterranean -10.4987
+▁pioneer -10.4987
+▁respond -10.4987
+▁thirsty -10.4987
+▁removal -10.4987
+▁revenge -10.4987
+▁vienna -10.4987
+▁mutual -10.4987
+▁annual -10.4987
+▁suburb -10.4988
+▁instance -10.4988
+▁hiding -10.4989
+▁collaboration -10.4991
+▁spending -10.4994
+▁habit -10.4994
+▁bullet -10.4997
+▁guest -10.5
+▁bré -10.5008
+▁distinguished -10.5041
+▁pack -10.5046
+▁rarely -10.5049
+▁allée -10.5058
+▁log -10.5066
+▁mask -10.5075
+▁promote -10.5089
+▁thrown -10.5092
+▁giro -10.5127
+▁arrang -10.5129
+▁rav -10.513
+zon -10.5136
+▁centuries -10.5136
+▁chassaigne -10.5136
+▁execution -10.5136
+▁grandfather -10.5136
+▁limoges -10.5136
+▁artillery -10.5136
+▁january -10.5136
+▁tribute -10.5136
+▁acquire -10.5136
+▁arthur -10.5136
+▁senior -10.5136
+▁pleasant -10.5136
+▁teeth -10.5136
+▁shirt -10.5137
+▁yann -10.5137
+▁employee -10.514
+▁detail -10.5141
+▁explained -10.5157
+hl -10.5157
+▁praise -10.5159
+ada -10.5167
+▁gaze -10.5171
+oir -10.5182
+▁promoted -10.5184
+▁cub -10.5202
+▁mende -10.5202
+▁accused -10.5222
+▁nord -10.5225
+▁cout -10.5227
+▁extract -10.5231
+▁vast -10.5273
+▁egypt -10.5282
+▁protest -10.5284
+▁communication -10.5288
+▁context -10.5288
+▁impression -10.5288
+▁substitute -10.5288
+▁technology -10.5288
+▁merchant -10.5288
+▁integrate -10.5288
+▁interrupted -10.5288
+▁cultiva -10.5288
+▁dismiss -10.5288
+▁cabinet -10.5288
+▁senator -10.5288
+▁basket -10.5288
+lemish -10.5288
+▁false -10.5288
+▁radical -10.5288
+▁sacred -10.529
+▁toulon -10.5292
+▁impose -10.53
+▁strengthen -10.5305
+▁princess -10.5311
+▁resumed -10.5315
+hel -10.5319
+usse -10.5338
+▁cake -10.5345
+▁laid -10.5347
+▁mala -10.5348
+▁kra -10.5377
+▁poul -10.5381
+▁cere -10.5384
+▁burned -10.5413
+ably -10.5427
+▁breakfast -10.5441
+▁deputies -10.5441
+▁enjolras -10.5441
+▁fireplace -10.5441
+▁portugal -10.5441
+▁uncertain -10.5441
+▁audience -10.5441
+▁arrangement -10.5442
+▁digital -10.5442
+▁birthday -10.5442
+▁restoration -10.5442
+▁lectur -10.5442
+▁wander -10.5442
+▁strict -10.5442
+▁agency -10.5442
+▁shaking -10.5442
+▁heaven -10.5443
+▁restrict -10.5443
+▁appointment -10.5444
+▁friendship -10.545
+▁broad -10.5452
+lene -10.5456
+▁highway -10.5457
+▁constitutional -10.5468
+▁vain -10.5468
+▁setting -10.5469
+▁locality -10.548
+bru -10.548
+▁adam -10.5482
+▁adapted -10.5484
+▁wounded -10.549
+▁describe -10.5493
+▁reduced -10.5494
+▁finishes -10.5498
+▁statement -10.55
+▁vitr -10.5518
+▁gru -10.552
+liard -10.5534
+▁shout -10.5537
+▁unfortunate -10.5568
+▁reflect -10.5581
+known -10.5588
+▁react -10.5594
+▁behavior -10.5598
+▁category -10.5598
+▁concrete -10.5598
+▁expedition -10.5598
+▁hungarian -10.5598
+▁infantry -10.5598
+▁sébastien -10.5598
+▁clermont -10.5598
+▁identity -10.5598
+▁ordinary -10.5598
+▁rhythm -10.5598
+▁component -10.5598
+▁zealand -10.5598
+▁explore -10.5598
+▁chicken -10.5598
+▁ninth -10.5598
+▁alliance -10.5598
+▁isolated -10.5599
+▁copy -10.5599
+▁transferred -10.5599
+▁shook -10.5599
+▁ireland -10.56
+▁household -10.5601
+▁highlight -10.5602
+▁democratic -10.5602
+▁geneva -10.5604
+maker -10.5609
+▁vienne -10.5612
+▁preparation -10.5614
+▁plot -10.5616
+▁norman -10.5621
+▁rival -10.5635
+▁plateau -10.5644
+▁sigh -10.5649
+▁oak -10.5662
+▁emerge -10.5665
+▁advisor -10.5666
+▁bark -10.5695
+rick -10.57
+▁deeply -10.57
+▁determine -10.5709
+▁imagine -10.5719
+omb -10.572
+▁petite -10.5722
+itude -10.5725
+dres -10.5754
+▁aircraft -10.5756
+▁ambassador -10.5756
+▁congratulate -10.5756
+▁guarantee -10.5756
+▁increasing -10.5756
+▁moscow -10.5756
+▁reassure -10.5756
+▁struggle -10.5756
+▁ceremony -10.5756
+▁concentrat -10.5756
+▁investment -10.5756
+▁smooth -10.5756
+▁architectural -10.5756
+▁pupil -10.5757
+▁forehead -10.5757
+▁guinea -10.5757
+▁dollar -10.5757
+▁translation -10.5758
+▁dozen -10.5758
+▁summit -10.5759
+▁hesitate -10.5759
+▁consum -10.5759
+▁lighthouse -10.576
+▁reims -10.576
+▁otto -10.5763
+▁mandate -10.5765
+▁beauty -10.5765
+▁candle -10.5766
+▁armand -10.5767
+▁countryside -10.5769
+▁manufacturer -10.5779
+▁anyway -10.5783
+▁coins -10.5795
+vé -10.5796
+assi -10.5797
+▁eaten -10.5798
+▁determined -10.5804
+lau -10.5814
+▁widely -10.5816
+▁exit -10.5819
+gie -10.5821
+▁pala -10.5834
+▁augustin -10.5834
+▁destroy -10.5849
+▁handle -10.585
+▁lamb -10.5881
+▁maker -10.5886
+▁cloth -10.5914
+▁pharmac -10.5918
+▁academic -10.5918
+▁delphine -10.5918
+▁medieval -10.5918
+▁practical -10.5918
+▁renaissance -10.5918
+▁surrender -10.5918
+▁universe -10.5918
+▁villeneuve -10.5918
+▁turkish -10.5918
+▁violence -10.5918
+▁bourgeois -10.5918
+▁avignon -10.5918
+▁curtain -10.5918
+▁physi -10.5918
+▁gothic -10.5918
+▁martial -10.5918
+▁altar -10.5919
+▁depict -10.592
+▁cycle -10.592
+▁sculptor -10.5922
+▁cheek -10.5922
+▁hungry -10.5922
+▁provence -10.5922
+▁vision -10.5925
+▁monster -10.5928
+▁lying -10.5928
+▁cré -10.5937
+▁suspended -10.5939
+hood -10.5941
+▁slid -10.5944
+▁characteriz -10.5946
+▁fake -10.5953
+▁auto -10.5962
+▁baker -10.5963
+▁maz -10.5991
+▁blame -10.6002
+▁alter -10.6009
+▁format -10.6011
+▁granted -10.6012
+▁apply -10.6019
+▁convince -10.6034
+▁kh -10.6037
+▁comfort -10.6077
+▁analysis -10.6082
+▁confused -10.6082
+▁exclaimed -10.6082
+▁illustrate -10.6082
+▁montpellier -10.6082
+▁chicago -10.6082
+▁sustain -10.6082
+▁delegate -10.6082
+▁foreground -10.6082
+▁error -10.6082
+▁rhone -10.6082
+▁estimate -10.6082
+▁dancing -10.6082
+▁montreal -10.6082
+▁benoit -10.6083
+▁etienne -10.6083
+▁confront -10.6085
+▁amiens -10.6085
+▁informed -10.6085
+hold -10.6086
+▁ferrand -10.6087
+▁hired -10.6088
+champ -10.6088
+▁collective -10.609
+▁fond -10.6093
+▁arab -10.6095
+▁odd -10.6098
+▁complicated -10.6099
+▁hél -10.6101
+▁tré -10.6104
+▁truck -10.6106
+▁clock -10.611
+▁defined -10.6126
+▁jim -10.6127
+▁northwest -10.6148
+▁bob -10.6155
+▁faster -10.6158
+▁pole -10.6186
+éné -10.6203
+▁gh -10.6204
+▁constitution -10.6223
+▁brittany -10.6248
+▁chimney -10.6248
+▁confidence -10.6248
+▁midfielder -10.6248
+▁vegetable -10.6248
+▁intervention -10.6248
+▁scandal -10.6248
+▁throat -10.6248
+▁recipe -10.6248
+▁sweat -10.6248
+▁violin -10.6249
+▁intention -10.6249
+▁ump -10.6249
+▁retreat -10.6249
+▁multiple -10.6249
+▁marshal -10.6249
+▁predict -10.625
+▁classification -10.625
+▁decade -10.6251
+▁gaston -10.6251
+▁dying -10.6256
+▁pressure -10.6257
+▁debt -10.626
+▁trois -10.6276
+▁formula -10.628
+▁shore -10.6286
+▁vend -10.6288
+▁chen -10.6297
+▁chapelle -10.63
+▁truly -10.6305
+▁feature -10.6308
+▁indicated -10.6309
+▁palm -10.6318
+▁jeanne -10.6323
+▁barre -10.6357
+▁santa -10.6381
+▁joke -10.639
+ñ -10.6418
+▁arriving -10.6418
+▁circular -10.6418
+▁discipline -10.6418
+▁distribution -10.6418
+▁engaged -10.6418
+▁evidence -10.6418
+▁fundamental -10.6418
+▁pacific -10.6418
+▁treasure -10.6418
+▁volunteer -10.6418
+▁dialect -10.6418
+▁interview -10.6418
+▁indonesia -10.6418
+▁poisson -10.6418
+▁swedish -10.6418
+▁allotment -10.6418
+▁bodies -10.6418
+▁stamp -10.6418
+▁regiment -10.6419
+▁keyboard -10.6419
+▁clerk -10.6419
+▁jesus -10.6419
+▁specifically -10.6421
+▁scored -10.6422
+▁initiative -10.6422
+▁tight -10.6425
+▁crush -10.6428
+▁seized -10.6445
+▁valid -10.6449
+▁brick -10.6457
+▁reache -10.6467
+▁burst -10.6468
+▁salon -10.6485
+▁earl -10.651
+▁q -10.6516
+▁printer -10.6526
+▁horn -10.6534
+▁occupie -10.6543
+▁riv -10.655
+▁colombia -10.6552
+▁precise -10.6561
+shire -10.6568
+▁math -10.658
+▁confident -10.659
+▁archbishop -10.659
+▁february -10.659
+▁héricourt -10.659
+▁installation -10.659
+▁suicide -10.659
+▁compromis -10.659
+▁quimper -10.659
+▁sexual -10.659
+▁expense -10.659
+▁fraysse -10.659
+▁random -10.659
+▁superior -10.659
+▁narrat -10.659
+▁edward -10.6591
+▁remarks -10.6591
+▁forgive -10.6592
+▁script -10.6594
+▁curious -10.6595
+▁alban -10.6608
+▁basse -10.6612
+▁retired -10.662
+▁captured -10.6621
+asi -10.6627
+▁crack -10.6629
+▁curve -10.6667
+▁fog -10.6677
+▁wishes -10.6684
+▁businesses -10.6689
+▁undergo -10.6701
+illac -10.6704
+▁define -10.672
+▁trap -10.6732
+▁kon -10.6752
+▁punish -10.6765
+▁circuit -10.6766
+▁contribution -10.6766
+▁definitely -10.6766
+▁everybody -10.6766
+▁possibility -10.6766
+▁provincial -10.6766
+▁saturday -10.6766
+▁hockey -10.6766
+▁purchase -10.6766
+▁communal -10.6766
+▁attitude -10.6766
+▁gaudin -10.6766
+▁christmas -10.6766
+▁planning -10.6768
+▁instant -10.677
+▁médor -10.677
+▁tip -10.677
+▁jazz -10.6771
+▁assume -10.6773
+▁jacket -10.6777
+▁sheep -10.6777
+▁vieux -10.6779
+▁fox -10.6784
+▁approved -10.6786
+▁pants -10.6791
+▁exhausted -10.6803
+▁observed -10.681
+▁whatever -10.6811
+▁bitter -10.6813
+▁comedy -10.6816
+▁metro -10.6817
+▁trick -10.682
+▁grace -10.6822
+laine -10.6825
+icle -10.6831
+▁franco -10.6841
+▁exceptional -10.6863
+▁electro -10.6881
+▁juan -10.6885
+▁frequent -10.6914
+▁gué -10.6927
+▁marrie -10.6928
+▁teach -10.6933
+▁apprentice -10.6944
+▁recognition -10.6944
+▁rectified -10.6944
+▁tunnel -10.6944
+▁ancestor -10.6944
+▁cavalry -10.6944
+▁thierry -10.6944
+▁collapse -10.6944
+▁torture -10.6944
+▁moderate -10.6944
+▁nancy -10.6944
+▁accent -10.6945
+▁controlled -10.6951
+▁fiction -10.6951
+▁runner -10.6956
+borough -10.6964
+▁crater -10.6971
+▁grandson -10.6973
+▁robin -10.6995
+▁acid -10.7005
+aternal -10.7038
+▁brain -10.7041
+▁abandon -10.7057
+▁arg -10.7086
+▁vary -10.7092
+▁recover -10.7092
+▁kur -10.7098
+▁putt -10.7106
+igné -10.7121
+▁efficient -10.7126
+▁kitchen -10.7126
+▁mathematics -10.7126
+▁napoleon -10.7126
+▁nervous -10.7126
+▁renowned -10.7126
+▁teenager -10.7126
+▁rapidly -10.7126
+▁miracle -10.7126
+▁tunisia -10.7126
+▁amateur -10.7126
+▁baroque -10.7126
+▁isabelle -10.7126
+▁prosper -10.7127
+▁undertake -10.7127
+▁bertrand -10.7127
+▁ignore -10.7128
+▁faubourg -10.7128
+▁argentina -10.7128
+▁clair -10.7134
+▁nimes -10.7136
+▁chairman -10.7142
+▁rub -10.7144
+ienne -10.7145
+▁invade -10.7159
+▁youngest -10.7163
+troph -10.7182
+▁asset -10.7187
+▁tele -10.7201
+▁publishe -10.7204
+▁eric -10.7213
+▁robi -10.7213
+▁ideal -10.7222
+▁burg -10.7255
+▁spin -10.7261
+▁involve -10.7264
+▁hau -10.7273
+▁assembl -10.7278
+▁actual -10.7286
+▁libr -10.7307
+▁northeast -10.7307
+▁accoyer -10.7311
+▁armchair -10.7311
+▁diameter -10.7311
+▁legislative -10.7311
+▁mathilde -10.7311
+▁satisfaction -10.7311
+▁stéphan -10.7311
+▁wisdom -10.7311
+▁counsel -10.7311
+▁ambition -10.7311
+▁license -10.7311
+▁turkey -10.7311
+▁exclusively -10.7311
+▁troyes -10.7312
+▁ugly -10.7312
+▁archives -10.7312
+▁distant -10.7312
+▁jacquelin -10.7313
+smith -10.7314
+▁mood -10.7314
+▁generate -10.7316
+▁claudi -10.7316
+▁bunch -10.7317
+▁humid -10.7318
+▁robinet -10.7319
+▁prim -10.7322
+▁distributed -10.7322
+▁spite -10.7326
+▁stake -10.733
+▁gorge -10.7334
+▁connection -10.7341
+▁rider -10.7343
+▁rouen -10.7348
+éré -10.7356
+▁romania -10.7386
+▁prose -10.7405
+ignon -10.7428
+▁distinguish -10.743
+▁abd -10.7451
+▁rang -10.7453
+▁observe -10.7453
+▁exact -10.7471
+▁enormous -10.75
+▁geographical -10.75
+í -10.75
+▁chocolate -10.75
+▁deficit -10.75
+▁definition -10.75
+▁delegation -10.75
+▁equivalent -10.75
+▁horrible -10.75
+▁threshold -10.75
+▁unanimous -10.75
+▁undoubtedly -10.75
+▁jersey -10.75
+▁friday -10.75
+▁easier -10.75
+▁eldest -10.75
+▁nurse -10.75
+▁navy -10.75
+▁inherit -10.7501
+▁jewel -10.7501
+▁boston -10.7503
+▁josé -10.7503
+▁véron -10.7505
+▁henriette -10.7507
+▁grabb -10.7511
+▁interpretation -10.7513
+▁massive -10.7516
+▁tiger -10.7516
+▁choir -10.7517
+▁arabic -10.7518
+▁pride -10.7521
+▁varies -10.7522
+▁frequently -10.7532
+phor -10.7538
+▁fuel -10.7543
+▁motion -10.7548
+▁clip -10.7553
+▁cruel -10.7563
+▁chu -10.7569
+▁dino -10.7627
+▁pipe -10.7664
+▁jail -10.767
+pont -10.7672
+fortunately -10.7688
+▁prefect -10.769
+▁artificial -10.7692
+▁luxembourg -10.7692
+▁obligation -10.7692
+▁pupponi -10.7692
+▁alphabet -10.7692
+▁innovation -10.7692
+▁promotion -10.7692
+▁proposing -10.7692
+▁diagnos -10.7692
+▁insurance -10.7692
+▁statistic -10.7692
+▁athlete -10.7692
+▁prestig -10.7692
+▁disaster -10.7692
+▁squad -10.7692
+▁funeral -10.7692
+▁glory -10.7692
+▁scope -10.7692
+▁blessed -10.7693
+▁sweden -10.7693
+▁medium -10.7693
+▁goddess -10.7693
+▁rabbi -10.7693
+▁distract -10.7694
+▁depth -10.7694
+▁gentleman -10.7694
+▁meantime -10.7695
+▁convict -10.7697
+▁cancer -10.7698
+▁graphic -10.77
+▁stuck -10.7701
+▁chapter -10.7701
+▁suitable -10.7702
+▁ease -10.7706
+▁employment -10.7706
+▁kha -10.7706
+▁regent -10.7715
+▁hervé -10.7717
+▁recovery -10.7728
+▁tender -10.7762
+▁marcelin -10.7789
+▁brian -10.7794
+▁mechanical -10.7799
+▁publisher -10.7826
+▁chab -10.7836
+▁manufacture -10.786
+finals -10.7875
+▁confess -10.7875
+▁circumstances -10.7888
+▁executed -10.7888
+▁faculty -10.7888
+▁félix -10.7888
+▁imprisoned -10.7888
+▁jérome -10.7888
+▁simultaneously -10.7888
+▁vietnam -10.7888
+▁wednesday -10.7888
+▁scotland -10.7888
+▁texas -10.7888
+▁hidden -10.7888
+▁dijon -10.7888
+▁threatening -10.7888
+▁threatened -10.7888
+▁flood -10.7889
+▁domestic -10.7889
+▁creating -10.7889
+▁okay -10.7889
+▁diamond -10.7889
+▁overall -10.7889
+▁winning -10.789
+▁baroness -10.7891
+▁fortune -10.7891
+▁preferred -10.7892
+▁package -10.7893
+▁adaptation -10.7896
+▁riviere -10.7896
+▁collaborator -10.7899
+▁vault -10.7899
+▁bruno -10.79
+▁plou -10.79
+▁authorized -10.7905
+▁genus -10.7914
+water -10.7915
+▁woke -10.7934
+▁rubb -10.8004
+bridge -10.8047
+▁absorb -10.8088
+▁aggressive -10.8088
+▁linguistic -10.8088
+▁response -10.8088
+▁peninsula -10.8088
+▁smiling -10.8088
+▁balcony -10.8088
+▁czech -10.8088
+▁diverse -10.8088
+▁liquid -10.8088
+▁madeleine -10.8089
+▁throne -10.8089
+▁applies -10.8089
+▁sophie -10.8089
+▁siege -10.8089
+▁laboratory -10.8089
+▁dolez -10.8089
+▁vachon -10.8089
+▁quarrel -10.8089
+▁regain -10.8092
+▁francais -10.8093
+▁polite -10.8093
+▁persist -10.8094
+▁egyptian -10.8096
+▁confession -10.8102
+▁toilet -10.8105
+▁revolutionary -10.8105
+▁bench -10.811
+▁collaborate -10.8117
+uvre -10.8134
+▁glu -10.8135
+▁stranger -10.8144
+▁hunt -10.8149
+▁loop -10.8206
+▁arrest -10.8224
+▁logical -10.823
+rnac -10.8259
+▁bush -10.829
+▁cavrois -10.8292
+▁conservatory -10.8292
+▁emmanuel -10.8292
+▁obstacle -10.8292
+▁subdivision -10.8292
+▁universal -10.8292
+▁édouard -10.8292
+▁alternative -10.8292
+▁desperate -10.8292
+▁muslim -10.8292
+▁supply -10.8292
+▁tobacco -10.8292
+▁switch -10.8293
+▁relax -10.8293
+ibility -10.8293
+▁stress -10.8294
+▁loyal -10.8294
+▁illness -10.8294
+▁rescue -10.8294
+▁folk -10.8295
+▁comfortable -10.8298
+▁admitted -10.83
+▁darkness -10.831
+▁romantic -10.8317
+▁rough -10.8317
+▁giant -10.8318
+▁archeolog -10.8342
+▁greg -10.8343
+▁milan -10.8368
+down -10.8436
+amba -10.844
+▁habitat -10.8491
+piece -10.8496
+▁angouleme -10.8501
+▁barricade -10.8501
+▁external -10.8501
+▁genevieve -10.8501
+▁identify -10.8501
+▁primarily -10.8501
+▁reception -10.8501
+▁shudder -10.8501
+▁supreme -10.8501
+▁clarify -10.8501
+▁condemn -10.8501
+▁perpignan -10.8501
+▁enlighten -10.8501
+▁gosselin -10.8501
+▁mysterious -10.8501
+▁powder -10.8501
+▁assigned -10.8501
+▁atlantic -10.8501
+▁fierce -10.8501
+▁resort -10.8501
+▁worst -10.8501
+▁twist -10.8501
+▁costume -10.8501
+▁saone -10.8501
+▁fortress -10.8502
+▁jacob -10.8502
+▁intense -10.8502
+▁lambert -10.8502
+▁nerve -10.8502
+▁johann -10.8503
+▁betray -10.8503
+▁abbé -10.8504
+▁hammer -10.8508
+▁clause -10.8514
+▁victories -10.8516
+▁charny -10.8518
+▁irre -10.8521
+▁borgo -10.8526
+▁fever -10.8535
+▁marthe -10.8538
+▁spell -10.8571
+▁chor -10.8573
+▁bak -10.8577
+▁julie -10.8591
+▁exception -10.8624
+▁alexandr -10.8667
+▁christin -10.8682
+metric -10.8698
+▁argu -10.8707
+quote -10.8708
+worth -10.8708
+▁adjacent -10.8714
+▁calculation -10.8714
+▁elegant -10.8714
+▁elsewhere -10.8714
+▁incredible -10.8714
+▁scattered -10.8714
+▁suspension -10.8714
+▁compensation -10.8714
+▁montauban -10.8714
+▁pillars -10.8714
+▁revolt -10.8714
+▁rugby -10.8714
+▁provoke -10.8714
+▁flesh -10.8714
+▁partially -10.8714
+▁cadio -10.8714
+▁harvest -10.8714
+▁automatic -10.8717
+▁cuba -10.8723
+▁prix -10.8724
+▁cheap -10.8726
+▁consistent -10.8726
+▁alex -10.8729
+▁cream -10.8731
+▁brush -10.8731
+▁considerable -10.8735
+▁femini -10.8736
+▁sink -10.8739
+▁reasonable -10.8741
+▁humor -10.8742
+▁elite -10.8743
+position -10.8745
+ouji -10.8796
+▁bowl -10.8803
+▁refus -10.8806
+▁southeast -10.8846
+▁georgia -10.8858
+▁immediate -10.8901
+▁benjamin -10.8931
+▁conservative -10.8931
+▁extensive -10.8931
+▁fountain -10.8931
+▁grenoble -10.8931
+▁inhabited -10.8931
+▁intelligent -10.8931
+▁populated -10.8931
+▁portuguese -10.8931
+▁sanctuary -10.8931
+▁stomach -10.8931
+▁temporary -10.8931
+▁ultimate -10.8931
+▁cinieri -10.8931
+▁compound -10.8931
+▁corridor -10.8931
+▁knife -10.8931
+▁commitment -10.8931
+▁shepherd -10.8931
+▁neutral -10.8931
+▁abroad -10.8931
+▁schedule -10.8931
+▁crystal -10.8931
+▁ornament -10.8931
+▁dispute -10.8931
+▁adolph -10.8931
+▁greece -10.8931
+▁ceiling -10.8931
+▁vocal -10.8932
+▁comedian -10.8932
+▁cheese -10.8932
+▁gilbert -10.8933
+▁boots -10.8934
+▁migrat -10.8934
+▁secur -10.8939
+▁township -10.8952
+▁striker -10.8959
+▁parent -10.8961
+▁hind -10.8969
+▁sergi -10.9009
+▁devil -10.9027
+▁cheat -10.9029
+▁wound -10.9088
+▁restore -10.9109
+▁baz -10.9116
+▁investigat -10.9143
+▁affiliat -10.9153
+▁contemplat -10.9153
+▁dynasty -10.9153
+▁reconstruct -10.9153
+▁revenue -10.9153
+▁somewhere -10.9153
+▁spectator -10.9153
+▁virtual -10.9153
+▁enclos -10.9153
+▁enemies -10.9153
+▁maintenance -10.9153
+▁qualify -10.9153
+▁fashion -10.9153
+▁profile -10.9153
+▁overseas -10.9153
+▁martyr -10.9153
+▁monarch -10.9154
+▁mystery -10.9154
+▁outfit -10.9154
+▁awful -10.9154
+▁conquer -10.9154
+▁deserve -10.9155
+▁sample -10.9155
+▁scottish -10.9155
+▁operator -10.9156
+otype -10.9156
+▁racing -10.9157
+▁victorious -10.9158
+▁punishment -10.9159
+▁strategic -10.916
+▁export -10.916
+▁abbe -10.9161
+▁convent -10.9166
+▁hér -10.9168
+▁dragg -10.917
+▁hung -10.9176
+▁gab -10.9183
+feld -10.9277
+▁porch -10.9324
+▁distinct -10.9368
+willer -10.9379
+▁attorney -10.938
+▁categories -10.938
+▁dynamic -10.938
+▁excessive -10.938
+▁exclud -10.938
+▁jahoua -10.938
+▁pomereux -10.938
+▁submitted -10.938
+▁universities -10.938
+▁contradict -10.938
+▁fragile -10.938
+▁penalty -10.938
+▁underscore -10.938
+▁worship -10.938
+▁activist -10.938
+artagnan -10.938
+▁communicate -10.938
+▁unlike -10.938
+corporate -10.938
+▁fossil -10.938
+▁delicate -10.9381
+▁vocation -10.9381
+▁terminal -10.9381
+▁dread -10.9381
+▁hetzel -10.9381
+▁sprint -10.9381
+▁niort -10.9381
+▁absent -10.9382
+▁flies -10.9382
+fold -10.9382
+▁duties -10.9383
+▁cosette -10.9383
+▁safety -10.9384
+▁confusion -10.9384
+▁blanche -10.9386
+▁handsome -10.9387
+▁lordship -10.9387
+▁operating -10.9388
+▁rumor -10.939
+▁politician -10.9392
+▁criticism -10.9396
+▁guitarist -10.9396
+▁liege -10.9398
+▁possession -10.9405
+cultur -10.9418
+ainville -10.9419
+stitution -10.9439
+▁couch -10.9463
+cute -10.9467
+asso -10.9503
+▁oppose -10.9566
+▁archipelago -10.9613
+▁celebration -10.9613
+▁controversial -10.9613
+▁controversy -10.9613
+▁eponym -10.9613
+▁legislation -10.9613
+▁misfortune -10.9613
+▁norwegian -10.9613
+▁pilgrim -10.9613
+▁satisfy -10.9613
+▁yvonne -10.9613
+▁heavily -10.9613
+▁exposed -10.9613
+▁guilty -10.9613
+▁memorial -10.9613
+▁analyze -10.9613
+▁munich -10.9613
+▁nominated -10.9613
+▁commerce -10.9613
+▁modification -10.9613
+▁saxon -10.9613
+▁contrast -10.9613
+▁carriage -10.9613
+▁antonio -10.9613
+▁lantern -10.9613
+▁madrid -10.9613
+▁raphael -10.9613
+▁shield -10.9613
+▁poison -10.9613
+▁generous -10.9613
+▁carfor -10.9613
+▁honey -10.9613
+▁visual -10.9614
+▁ashamed -10.9614
+▁dishes -10.9615
+▁borrow -10.9615
+▁wonderful -10.9616
+▁chaumont -10.9618
+▁renaud -10.9618
+▁holland -10.9618
+▁coordination -10.9619
+▁elect -10.962
+▁alexis -10.9621
+▁canvas -10.9626
+▁raoul -10.963
+▁montjo -10.9634
+▁gallo -10.9644
+lithic -10.9649
+▁vaux -10.9657
+▁nazi -10.968
+head -10.9684
+took -10.9693
+▁wage -10.9693
+▁karl -10.9738
+▁hameau -10.9751
+▁select -10.9769
+▁preserve -10.9804
+▁advocate -10.9851
+▁explosion -10.9851
+▁maximum -10.9851
+▁monsieur -10.9851
+▁satellite -10.9851
+▁scenario -10.9851
+▁smoking -10.9851
+▁sufficient -10.9851
+▁territorial -10.9851
+▁twentieth -10.9851
+▁unhappy -10.9851
+▁midnight -10.9851
+▁washington -10.9851
+▁maritime -10.9851
+▁overlook -10.9851
+▁electoral -10.9851
+▁fragment -10.9851
+▁comparison -10.9851
+▁precious -10.9851
+▁tokyo -10.9851
+▁columbia -10.9851
+▁modify -10.9851
+▁variation -10.9851
+▁glaci -10.9851
+▁facial -10.9851
+▁camille -10.9851
+▁compensate -10.9851
+▁profit -10.9851
+▁smart -10.9852
+▁lodge -10.9852
+▁educate -10.9852
+▁hungary -10.9854
+▁entertainment -10.9855
+▁rochelle -10.9857
+▁massif -10.9859
+▁seminar -10.9862
+▁savings -10.9865
+▁savary -10.9865
+▁weird -10.9871
+▁alert -10.988
+▁tolerate -10.9885
+▁maison -10.9895
+explo -10.9935
+▁mechanic -10.9961
+▁declare -11.0048
+▁evil -11.0079
+▁consecutive -11.0095
+▁engagement -11.0095
+▁enthusiasm -11.0095
+▁extraordinary -11.0095
+▁nautilus -11.0095
+▁neglect -11.0095
+▁participation -11.0095
+▁plenty -11.0095
+▁reappear -11.0095
+▁unexpected -11.0095
+▁versailles -11.0095
+▁virtue -11.0095
+▁volcano -11.0095
+▁morocco -11.0095
+▁orléans -11.0095
+▁complement -11.0095
+▁despair -11.0095
+▁bouvard -11.0095
+▁horror -11.0095
+▁sorrow -11.0095
+▁harbor -11.0095
+▁perfume -11.0095
+▁regulate -11.0095
+▁stopping -11.0096
+▁divine -11.0096
+▁entry -11.0096
+▁louvre -11.0098
+▁craft -11.0101
+▁bonneton -11.0102
+▁assault -11.0102
+▁forgot -11.0102
+▁burial -11.0104
+▁virginie -11.0106
+▁drag -11.0111
+▁socks -11.0114
+▁quote -11.0128
+▁recogniz -11.0164
+▁urge -11.024
+▁adapt -11.0266
+▁jacqu -11.0327
+agnac -11.0329
+▁acknowledg -11.0345
+▁affluent -11.0345
+▁algorithm -11.0345
+▁arrondissement -11.0345
+▁aurélie -11.0345
+▁demonstrate -11.0345
+▁dieudonné -11.0345
+▁dissolve -11.0345
+▁engraving -11.0345
+▁espérance -11.0345
+▁hypothesis -11.0345
+▁louwagie -11.0345
+▁massonneau -11.0345
+▁molecule -11.0345
+▁monseigneur -11.0345
+▁railroad -11.0345
+▁rebuilt -11.0345
+▁suppress -11.0345
+▁delaunay -11.0345
+▁sleeve -11.0345
+▁solidarity -11.0345
+▁unusual -11.0345
+▁worries -11.0345
+▁pattern -11.0345
+▁stabiliz -11.0345
+▁fabric -11.0345
+boeuf -11.0345
+▁verdun -11.0345
+▁injury -11.0345
+▁charente -11.0345
+▁denmark -11.0345
+▁repress -11.0345
+▁telephone -11.0345
+▁adrien -11.0345
+▁naples -11.0345
+▁porthos -11.0346
+▁behave -11.0346
+▁introduction -11.0346
+▁florida -11.0346
+▁aubin -11.0346
+▁normandy -11.0348
+▁trend -11.0348
+▁jardin -11.0348
+▁intervene -11.0349
+▁fortified -11.035
+▁painful -11.0351
+▁tennis -11.0355
+▁distinction -11.0359
+pierre -11.036
+roving -11.0408
+▁abundant -11.0601
+▁appropriate -11.0601
+▁bicycle -11.0601
+▁bilingual -11.0601
+▁coincide -11.0601
+▁erected -11.0601
+▁facilities -11.0601
+▁illustration -11.0601
+▁lydie -11.0601
+▁mirror -11.0601
+▁philosopher -11.0601
+▁pregnant -11.0601
+▁reverend -11.0601
+▁senegal -11.0601
+▁struggling -11.0601
+▁sympath -11.0601
+▁creature -11.0601
+▁download -11.0601
+▁giraud -11.0601
+▁inaugura -11.0601
+▁description -11.0601
+▁destination -11.0601
+▁vercamer -11.0601
+▁orleans -11.0601
+▁saddier -11.0601
+▁massacre -11.0602
+▁amazon -11.0602
+▁thunder -11.0602
+▁vacation -11.0602
+▁serving -11.0602
+▁skirt -11.0602
+▁vosges -11.0602
+▁spouse -11.0602
+▁comte -11.0602
+▁physics -11.0603
+▁genevard -11.0603
+▁carpet -11.0603
+▁valentin -11.0603
+▁jews -11.0604
+▁spark -11.0604
+▁remire -11.0604
+▁anglo -11.0605
+▁cluster -11.0606
+▁faithful -11.0606
+▁vertical -11.0607
+▁vital -11.0607
+▁peaceful -11.0607
+▁tyran -11.0607
+▁drill -11.0608
+▁weakness -11.061
+▁creator -11.061
+▁cancel -11.0611
+▁alice -11.0614
+▁theories -11.0616
+▁reviv -11.062
+▁brunet -11.0624
+▁elephant -11.0677
+uchet -11.0711
+blanc -11.0772
+▁capture -11.0817
+▁resume -11.0817
+▁scott -11.0864
+▁accumulat -11.0865
+▁affirm -11.0865
+▁bachelor -11.0865
+▁brigitte -11.0865
+▁byzantine -11.0865
+▁cyprien -11.0865
+▁defensive -11.0865
+▁ferdinand -11.0865
+▁folleville -11.0865
+▁geoffroy -11.0865
+▁henceforth -11.0865
+▁limestone -11.0865
+▁transparency -11.0865
+▁accelerat -11.0865
+▁grandmother -11.0865
+▁thereafter -11.0865
+▁lorient -11.0865
+▁renamed -11.0865
+▁monitor -11.0865
+▁école -11.0865
+▁navigat -11.0865
+▁laurence -11.0865
+▁assisted -11.0865
+▁inscription -11.0865
+▁orbit -11.0867
+▁theorem -11.0869
+▁disappearance -11.087
+▁brewer -11.087
+▁aisne -11.087
+▁cease -11.0876
+▁prun -11.0877
+front -11.0901
+▁gamb -11.0967
+▁imitat -11.0981
+▁retire -11.1083
+▁appreciate -11.1088
+▁divide -11.1089
+▁plea -11.1095
+▁harmoni -11.1101
+stadt -11.1127
+script -11.1129
+▁entertain -11.113
+▁apologize -11.1135
+▁autonomous -11.1135
+▁bulgaria -11.1135
+▁coronado -11.1135
+▁embassy -11.1135
+▁fantasy -11.1135
+▁hurried -11.1135
+▁immigrant -11.1135
+▁madagascar -11.1135
+▁mistress -11.1135
+▁precision -11.1135
+▁suspicious -11.1135
+▁penetrat -11.1135
+▁tragedy -11.1135
+▁applaud -11.1135
+▁clément -11.1135
+▁embrace -11.1135
+▁hilaire -11.1135
+▁proportion -11.1135
+▁disciple -11.1135
+▁substance -11.1135
+▁alsace -11.1135
+▁transition -11.1135
+▁trophy -11.1135
+▁anthrop -11.1135
+▁cayenne -11.1135
+▁animated -11.1135
+▁mobile -11.1135
+▁stolen -11.1135
+beuf -11.1135
+▁belief -11.1135
+▁caroline -11.1135
+▁bavaria -11.1135
+▁colony -11.1135
+▁supervise -11.1136
+▁varied -11.1136
+▁inviting -11.1137
+▁glove -11.1137
+▁noisy -11.1137
+▁occurred -11.1138
+▁toxic -11.114
+▁horizontal -11.1146
+▁infectio -11.115
+camp -11.1184
+▁accord -11.1344
+▁aubr -11.1353
+▁absolute -11.1383
+▁descend -11.1408
+▁import -11.141
+▁illuminat -11.1413
+responsibilities -11.1413
+▁accomplish -11.1413
+▁alexander -11.1413
+▁aristocrat -11.1413
+▁brilliant -11.1413
+▁corporation -11.1413
+▁expansion -11.1413
+▁gloomy -11.1413
+▁intellectual -11.1413
+▁mammal -11.1413
+▁phenomena -11.1413
+▁poitiers -11.1413
+▁privilege -11.1413
+▁pécuchet -11.1413
+▁régulus -11.1413
+▁striking -11.1413
+▁asleep -11.1413
+▁fontenay -11.1413
+▁trigger -11.1413
+▁criteria -11.1413
+▁offensive -11.1413
+▁rabault -11.1413
+▁prevail -11.1413
+▁cheerful -11.1413
+▁samuel -11.1413
+▁censor -11.1413
+▁mixture -11.1413
+▁memories -11.1413
+▁ghost -11.1413
+▁patrol -11.1413
+▁venice -11.1413
+▁aurillac -11.1413
+▁calculate -11.1413
+▁cameroon -11.1413
+▁managing -11.1413
+▁dolores -11.1413
+▁spoon -11.1413
+▁terrace -11.1413
+▁framework -11.1414
+▁purple -11.1414
+▁prud -11.1414
+▁crypt -11.1414
+▁edme -11.1414
+▁nasie -11.1417
+communica -11.1423
+born -11.1438
+▁mineral -11.1548
+▁wrap -11.1641
+▁necessar -11.1665
+▁coordinat -11.1691
+▁assumption -11.1698
+▁athletic -11.1698
+▁bathroom -11.1698
+▁conceived -11.1698
+▁financing -11.1698
+▁lefebvre -11.1698
+▁legitimate -11.1698
+▁observatory -11.1698
+▁suzanne -11.1698
+▁sylvain -11.1698
+▁transmission -11.1698
+▁disagree -11.1698
+▁disposal -11.1698
+▁lorraine -11.1698
+▁kirsch -11.1698
+▁inevitabl -11.1698
+▁meadow -11.1698
+▁patriot -11.1698
+▁squeez -11.1698
+▁brigade -11.1698
+▁malaysia -11.1698
+▁snake -11.1698
+▁cartoon -11.1698
+▁gaulle -11.1698
+▁granite -11.1698
+▁veteran -11.1698
+▁weekend -11.1699
+▁niece -11.1699
+▁spider -11.1699
+▁corsica -11.1699
+▁gently -11.17
+▁moselle -11.17
+▁blank -11.1701
+▁disposition -11.1702
+▁sultan -11.1702
+▁juliette -11.1702
+▁consent -11.1704
+▁faint -11.1708
+owski -11.171
+limard -11.1716
+▁ollier -11.1717
+fusion -11.1718
+▁update -11.1853
+▁translate -11.1946
+▁govern -11.198
+▁swallow -11.1992
+▁accommodation -11.1992
+▁archaeological -11.1992
+▁biodiversity -11.1992
+▁congregation -11.1992
+▁constituen -11.1992
+▁efficiency -11.1992
+▁familiar -11.1992
+▁flaviana -11.1992
+▁jesuit -11.1992
+▁maneuver -11.1992
+▁montaudoin -11.1992
+▁potatoes -11.1992
+▁predominant -11.1992
+▁scientist -11.1992
+▁shrugged -11.1992
+▁societies -11.1992
+▁swimming -11.1992
+▁unbearable -11.1992
+▁decorative -11.1992
+▁garrison -11.1992
+▁instinct -11.1992
+▁besancon -11.1992
+▁cycling -11.1992
+▁maltenu -11.1992
+▁shrub -11.1992
+▁sauveur -11.1992
+giovanni -11.1992
+▁gospel -11.1992
+▁chemical -11.1992
+▁producing -11.1992
+▁shivering -11.1993
+▁devotion -11.1993
+▁petersburg -11.1993
+▁lurton -11.1993
+▁flew -11.1993
+▁variant -11.1993
+▁gulf -11.1993
+▁compact -11.1993
+▁surgery -11.1993
+▁aramis -11.1993
+▁quadr -11.1993
+▁outcome -11.1993
+neath -11.1993
+▁cotton -11.1994
+▁indication -11.1994
+▁toronto -11.1994
+▁considerably -11.1994
+▁foresee -11.1994
+▁bump -11.1995
+▁dropped -11.1996
+▁renault -11.1996
+▁slept -11.1996
+▁pension -11.1997
+▁scheme -11.1999
+▁thief -11.2004
+▁blond -11.2011
+▁assert -11.2015
+▁provid -11.2064
+▁weigh -11.219
+▁measur -11.226
+establish -11.229
+woman -11.229
+á -11.2295
+▁barcelona -11.2295
+▁calendar -11.2295
+▁consecrat -11.2295
+▁innovative -11.2295
+▁occupy -11.2295
+▁perspective -11.2295
+▁supplement -11.2295
+▁thursday -11.2295
+▁vineyard -11.2295
+▁yield -11.2295
+▁abolish -11.2295
+▁consumption -11.2295
+▁eternal -11.2295
+▁evaluation -11.2295
+▁oxford -11.2295
+▁uttered -11.2295
+▁diversity -11.2295
+▁sequence -11.2295
+▁submarine -11.2295
+▁podium -11.2296
+▁internship -11.2296
+▁kourou -11.2296
+▁paradise -11.2296
+▁glance -11.2296
+▁essay -11.2296
+▁pistol -11.2296
+▁strangle -11.2296
+▁overcome -11.2296
+▁stroll -11.2296
+▁shopping -11.2296
+▁alternate -11.2296
+▁changing -11.2296
+▁deported -11.2296
+▁mongol -11.2297
+▁peyre -11.2297
+▁sorgue -11.2297
+▁automatically -11.2297
+▁cancelled -11.2297
+▁hastily -11.2297
+▁guérin -11.2298
+judicial -11.23
+▁glad -11.2303
+▁fisc -11.2348
+▁pouill -11.2463
+▁chris -11.2478
+▁whisper -11.2538
+▁bathilde -11.2608
+▁clergy -11.2608
+▁conclusion -11.2608
+▁cooperative -11.2608
+▁cyclist -11.2608
+▁demonstration -11.2608
+▁ecuador -11.2608
+▁elizabeth -11.2608
+▁enforce -11.2608
+▁exterior -11.2608
+▁hierarchy -11.2608
+▁hippolyte -11.2608
+▁inferior -11.2608
+▁infinite -11.2608
+▁mézieres -11.2608
+▁nickname -11.2608
+▁participating -11.2608
+▁rebellion -11.2608
+▁rectangular -11.2608
+▁regroup -11.2608
+▁antiquity -11.2608
+▁burgundy -11.2608
+▁embark -11.2608
+▁homosexual -11.2608
+▁excited -11.2608
+▁invisible -11.2608
+▁ribbon -11.2608
+▁invitation -11.2608
+▁adorned -11.2608
+▁variable -11.2608
+▁ecological -11.2608
+▁stroke -11.2608
+▁misery -11.2608
+▁yelled -11.2608
+▁orphan -11.2609
+▁cahors -11.2609
+▁césar -11.2609
+▁pounds -11.2609
+▁mathieu -11.2609
+▁cognac -11.2609
+▁withdrew -11.261
+▁banned -11.261
+▁extinct -11.261
+▁carved -11.2611
+▁alarm -11.2612
+▁tibet -11.2614
+▁awake -11.2615
+▁irrita -11.2626
+fleur -11.266
+terrain -11.2687
+eutic -11.2734
+ergic -11.2844
+▁exhaust -11.2861
+▁gentle -11.2926
+▁profound -11.2931
+▁slaughter -11.2931
+▁tremendous -11.2931
+▁uruguay -11.2931
+▁advertising -11.2931
+▁butterfly -11.2931
+▁cartridge -11.2931
+▁circum -11.2931
+▁conscience -11.2931
+▁conspiracy -11.2931
+▁depression -11.2931
+▁discourage -11.2931
+▁infrastructure -11.2931
+▁maréchal -11.2931
+▁orthodox -11.2931
+▁reservation -11.2931
+▁strategy -11.2931
+▁absurd -11.2931
+▁theoretical -11.2931
+▁valjean -11.2931
+▁reproach -11.2931
+▁transaction -11.2931
+▁disorder -11.2931
+▁crucial -11.2931
+▁reduction -11.2931
+▁skull -11.2931
+▁manosque -11.2931
+▁bretagne -11.2931
+▁sketch -11.2931
+▁combination -11.2931
+▁harmony -11.2931
+▁syrian -11.2931
+▁drank -11.2931
+▁goriot -11.2931
+▁vallée -11.2931
+▁tragic -11.2931
+▁donkey -11.2931
+▁extent -11.2931
+▁tackle -11.2931
+▁refine -11.2932
+▁thermal -11.2932
+▁diving -11.2932
+▁fresco -11.2932
+▁urgent -11.2933
+▁croatia -11.2934
+itarian -11.2935
+▁agitat -11.2942
+▁pomp -11.2952
+flex -11.3101
+▁brutal -11.3115
+▁democrat -11.3255
+▁accommodate -11.3264
+▁acquisition -11.3264
+▁aluminum -11.3264
+▁comrade -11.3264
+▁curiosity -11.3264
+▁destruction -11.3264
+▁disappointed -11.3264
+▁doctrine -11.3264
+▁exaggerate -11.3264
+▁fauchelevent -11.3264
+▁goalkeeper -11.3264
+▁irregular -11.3264
+▁lithuania -11.3264
+▁metropolitan -11.3264
+▁philadelphia -11.3264
+▁residency -11.3264
+▁structural -11.3264
+▁théâtre -11.3264
+▁tuesday -11.3264
+▁ukraine -11.3264
+▁unacceptable -11.3264
+▁venezuela -11.3264
+▁envelope -11.3264
+▁fascinat -11.3264
+▁miserable -11.3264
+▁protocol -11.3264
+▁tissue -11.3264
+▁edouard -11.3264
+▁manipulat -11.3264
+▁narbonne -11.3264
+▁translator -11.3264
+▁disguise -11.3264
+▁worldwide -11.3264
+▁respectively -11.3264
+▁predator -11.3264
+▁filter -11.3264
+▁protein -11.3264
+▁barbier -11.3264
+▁pianist -11.3264
+▁vieille -11.3264
+▁guéret -11.3264
+▁chronic -11.3264
+▁dessert -11.3264
+▁amazing -11.3264
+▁nightmare -11.3265
+▁exceed -11.3265
+agnol -11.3265
+▁chateauroux -11.3265
+▁flank -11.3265
+▁oath -11.3265
+▁survey -11.3265
+▁deletion -11.3265
+▁emblem -11.3269
+▁collaborat -11.3528
+▁appoint -11.3534
+▁eliminate -11.3561
+western -11.3569
+mouth -11.3604
+▁agglomeration -11.3609
+▁alphonse -11.3609
+▁cooperation -11.3609
+▁devastated -11.3609
+▁ethiopia -11.3609
+▁footsteps -11.3609
+▁félici -11.3609
+▁hollywood -11.3609
+▁meticulous -11.3609
+▁overwhelm -11.3609
+▁pencroff -11.3609
+▁possibilities -11.3609
+▁pseudonym -11.3609
+▁pyramid -11.3609
+▁ridiculous -11.3609
+▁sarkozy -11.3609
+▁suspicion -11.3609
+▁interface -11.3609
+▁mexican -11.3609
+▁postponed -11.3609
+▁prominent -11.3609
+▁psychiatri -11.3609
+▁richelieu -11.3609
+▁cécile -11.3609
+▁kennedy -11.3609
+▁lombard -11.3609
+▁colonial -11.3609
+▁tremble -11.3609
+▁triangle -11.3609
+▁slipped -11.3609
+▁cabaret -11.3609
+▁laundry -11.3609
+▁airplane -11.3609
+▁depriv -11.3609
+▁discreet -11.3609
+▁amended -11.3609
+▁dizier -11.3609
+▁dwelling -11.3609
+▁innocent -11.3609
+▁reunion -11.3609
+▁breeze -11.3609
+▁loving -11.361
+▁await -11.361
+▁invention -11.3611
+▁avenge -11.365
+homme -11.3661
+logy -11.3707
+▁moderniz -11.3959
+speaking -11.396
+▁georg -11.3961
+▁administrator -11.3966
+▁adolescen -11.3966
+▁anonymous -11.3966
+▁applicable -11.3966
+▁bailiff -11.3966
+▁basilica -11.3966
+▁bonaparte -11.3966
+▁centimeters -11.3966
+▁coalition -11.3966
+▁entrepreneur -11.3966
+▁fluctuat -11.3966
+▁gwynplaine -11.3966
+▁immense -11.3966
+▁magistrate -11.3966
+▁marvejol -11.3966
+▁mistreat -11.3966
+▁necessity -11.3966
+▁olympe -11.3966
+▁opportunities -11.3966
+▁outstanding -11.3966
+▁patriarch -11.3966
+▁prehistoric -11.3966
+▁repertoire -11.3966
+▁reproduction -11.3966
+▁synthesis -11.3966
+▁thailand -11.3966
+▁thirtieth -11.3966
+▁théodore -11.3966
+▁viscount -11.3966
+▁vowels -11.3966
+▁coherence -11.3966
+▁fiancé -11.3966
+▁nymph -11.3966
+▁prussia -11.3966
+▁nobility -11.3966
+▁prompt -11.3966
+▁separation -11.3966
+▁signature -11.3966
+▁arsenal -11.3966
+▁disgrace -11.3966
+▁patience -11.3966
+▁abstract -11.3966
+▁puppet -11.3966
+▁secular -11.3966
+▁resembl -11.3966
+▁frozen -11.3966
+▁marmande -11.3966
+▁celtic -11.3966
+▁dignity -11.3966
+▁disgust -11.3966
+▁eagle -11.3966
+▁junction -11.3966
+▁trunk -11.3966
+▁conseil -11.3966
+▁reliable -11.3966
+▁surname -11.3966
+▁goodbye -11.3966
+▁playoffs -11.3966
+▁warrior -11.3966
+▁ariege -11.3967
+▁nicole -11.3967
+▁bernadet -11.3968
+▁stefan -11.3969
+▁retrac -11.3969
+▁seduce -11.3981
+▁automat -11.3986
+▁chevr -11.404
+franc -11.4113
+uangan -11.4139
+ignol -11.4162
+▁evident -11.4336
+▁amsterdam -11.4336
+▁aquitaine -11.4336
+▁archdiocese -11.4336
+▁attachment -11.4336
+▁automobile -11.4336
+▁carcassonne -11.4336
+▁emphasize -11.4336
+▁episcopal -11.4336
+▁hesitating -11.4336
+▁industries -11.4336
+▁jerusalem -11.4336
+▁promenade -11.4336
+▁pyrenees -11.4336
+▁reproduce -11.4336
+▁staircase -11.4336
+▁stubborn -11.4336
+▁subsidiary -11.4336
+▁temporarily -11.4336
+▁twelfth -11.4336
+▁upstream -11.4336
+▁volcanic -11.4336
+▁consolidat -11.4336
+▁frighten -11.4336
+▁godfather -11.4336
+▁humiliat -11.4336
+▁illusion -11.4336
+▁opposing -11.4336
+▁abdomen -11.4336
+▁fluently -11.4336
+▁swamp -11.4336
+▁examination -11.4336
+▁charlotte -11.4336
+▁pigeon -11.4336
+▁ritual -11.4336
+▁spiral -11.4336
+▁café -11.4336
+▁iraq -11.4336
+▁margaret -11.4337
+▁whistle -11.4337
+▁worn -11.4337
+▁condom -11.4338
+▁lounge -11.4339
+▁ethic -11.4366
+▁analog -11.4443
+▁replac -11.4445
+eastern -11.4515
+mposing -11.4569
+▁dedicate -11.4674
+▁advertise -11.4721
+▁authentic -11.4721
+▁brunswick -11.4721
+▁configuration -11.4721
+▁consciousness -11.4721
+▁dhuicq -11.4721
+▁eloquence -11.4721
+▁frequency -11.4721
+▁frémilly -11.4721
+▁google -11.4721
+▁granddaughter -11.4721
+▁impatience -11.4721
+▁joséphine -11.4721
+▁languedoc -11.4721
+▁prosecutor -11.4721
+▁slippers -11.4721
+▁supermarket -11.4721
+▁widespread -11.4721
+ó -11.4721
+▁combelaine -11.4721
+▁eyelids -11.4721
+▁pakistan -11.4721
+▁terrasse -11.4721
+▁torrent -11.4721
+▁crossroad -11.4721
+▁premises -11.4721
+▁qualification -11.4721
+▁compliment -11.4721
+▁bolivia -11.4721
+▁chestnut -11.4721
+▁désert -11.4721
+▁sylvi -11.4721
+▁charleville -11.4721
+▁parliamentarian -11.4721
+▁warsaw -11.4721
+▁ghison -11.4721
+▁warehouse -11.4721
+▁fossé -11.4721
+▁dubié -11.4721
+▁defect -11.4721
+▁width -11.4721
+▁eager -11.4721
+▁rotten -11.4722
+▁nigeria -11.4722
+▁turtle -11.4723
+▁scrap -11.4834
+▁approve -11.5074
+▁kilometer -11.5079
+▁horizon -11.5104
+▁catalog -11.512
+▁exclusiv -11.5121
+▁apostolic -11.5121
+▁ascension -11.5121
+▁aubenas -11.5121
+▁auvergne -11.5121
+▁bibichon -11.5121
+▁caterpillar -11.5121
+▁certificate -11.5121
+▁chateauneuf -11.5121
+▁compatible -11.5121
+▁compulsory -11.5121
+▁deceased -11.5121
+▁demolished -11.5121
+▁douarnenez -11.5121
+▁identification -11.5121
+▁impressive -11.5121
+▁josserand -11.5121
+▁laffitte -11.5121
+▁nathalie -11.5121
+▁octave -11.5121
+▁pedestrian -11.5121
+▁pollution -11.5121
+▁precaution -11.5121
+▁presidency -11.5121
+▁relevant -11.5121
+▁screenwriter -11.5121
+▁stammered -11.5121
+▁synagogue -11.5121
+▁synonym -11.5121
+▁testimony -11.5121
+▁transmitted -11.5121
+▁vaccination -11.5121
+▁villefranche -11.5121
+▁renovation -11.5121
+▁slender -11.5121
+▁soundtrack -11.5121
+▁wolves -11.5121
+▁marathon -11.5121
+▁toponym -11.5121
+▁apostle -11.5121
+▁carhaix -11.5121
+▁drowned -11.5121
+▁illegal -11.5121
+▁policies -11.5121
+▁shareholder -11.5121
+▁wharf -11.5121
+▁allocated -11.5121
+▁sociology -11.5121
+▁tramway -11.5121
+▁bacteri -11.5121
+▁colonies -11.5121
+▁finnish -11.5121
+▁felix -11.5121
+▁jordan -11.5121
+▁dominant -11.5121
+▁musket -11.5121
+▁propell -11.5121
+▁computing -11.5121
+▁flavor -11.5121
+▁sphere -11.5122
+▁gwen -11.5124
+▁pearl -11.5127
+▁terribl -11.5491
+ground -11.5516
+misunderstanding -11.5537
+▁compiegne -11.5537
+▁dauphiné -11.5537
+▁definitive -11.5537
+▁democracy -11.5537
+▁escaping -11.5537
+▁extinguish -11.5537
+▁glenarvan -11.5537
+▁influential -11.5537
+▁keraotred -11.5537
+▁lincoln -11.5537
+▁massachusetts -11.5537
+▁metropolis -11.5537
+▁obliged -11.5537
+▁oxygen -11.5537
+▁paraguay -11.5537
+▁pyrénées -11.5537
+▁subscribe -11.5537
+▁suffocating -11.5537
+▁systematically -11.5537
+▁thorough -11.5537
+▁transparent -11.5537
+▁unpleasant -11.5537
+▁unsuccessful -11.5537
+▁catalonia -11.5537
+▁geometric -11.5537
+▁supervisor -11.5537
+▁adjoin -11.5537
+▁decisive -11.5537
+▁stunned -11.5537
+▁escort -11.5538
+▁gratitude -11.5538
+▁injuries -11.5538
+▁compilation -11.5538
+▁rendered -11.5538
+▁poverty -11.5538
+▁sentiment -11.5538
+▁lobby -11.5538
+▁mandatory -11.5538
+▁fayette -11.5538
+▁wreck -11.5538
+▁gossip -11.5538
+▁auroi -11.5538
+▁default -11.5538
+▁conquest -11.5538
+▁grumbl -11.5538
+▁invasion -11.5538
+▁remote -11.5538
+▁edible -11.5538
+ð -12.838
+ú -13.788
+ʾ -13.788
+. -14.288
+ā -14.288
+ʉ -14.288
+ʻ -14.288
+þ -15.2876
+̇ -15.2877
+~ -15.2878
+ʿ -15.2879
+! -15.288
+â -15.288
+ˋ -15.288
+ː -15.288
+̀ -15.288
+̠ -15.288
diff --git a/demo/README.md b/demo/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..e7fcf01ed881e1e7cc17f45fbe6d13c9e33d2a56
--- /dev/null
+++ b/demo/README.md
@@ -0,0 +1,31 @@
+# GUI Demo for StreamSpeech
+
+## One Speech Inference
+```shell
+export CUDA_VISIBLE_DEVICES=0
+
+ROOT=/data/zhangshaolei/StreamSpeech # path to StreamSpeech repo
+PRETRAIN_ROOT=/data/zhangshaolei/pretrain_models
+VOCODER_CKPT=$PRETRAIN_ROOT/unit-based_HiFi-GAN_vocoder/mHuBERT.layer11.km1000.en/g_00500000 # path to downloaded Unit-based HiFi-GAN Vocoder
+VOCODER_CFG=$PRETRAIN_ROOT/unit-based_HiFi-GAN_vocoder/mHuBERT.layer11.km1000.en/config.json # path to downloaded Unit-based HiFi-GAN Vocoder
+
+LANG=fr
+file=/streamspeech.simultaneous.${LANG}-en.pt # path to downloaded StreamSpeech model
+output_dir=$ROOT/res/streamspeech.simultaneous.${LANG}-en/simul-s2st
+
+chunk_size=320 #ms
+PYTHONPATH=$ROOT/fairseq python infer.py --data-bin ${ROOT}/configs/${LANG}-en \
+ --user-dir ${ROOT}/researches/ctc_unity --agent-dir ${ROOT}/agent \
+ --model-path $file \
+ --config-yaml config_gcmvn.yaml --multitask-config-yaml config_mtl_asr_st_ctcst.yaml \
+ --segment-size $chunk_size \
+ --vocoder $VOCODER_CKPT --vocoder-cfg $VOCODER_CFG --dur-prediction
+```
+
+## GUI Demo
+1. Modify the config file [config.json](./config.json)
+2. Run GUI Inference
+```shell
+python app.py
+```
+The Web UI will be displayed at http://0.0.0.0:7860/ and can be accessed using a browser.
\ No newline at end of file
diff --git a/demo/app.py b/demo/app.py
new file mode 100644
index 0000000000000000000000000000000000000000..2e279345163e928bbf6cdeb823e204248f34ab60
--- /dev/null
+++ b/demo/app.py
@@ -0,0 +1,967 @@
+
+##########################################
+# Simultaneous Speech-to-Speech Translation Agent for StreamSpeech
+#
+# StreamSpeech: Simultaneous Speech-to-Speech Translation with Multi-task Learning (ACL 2024)
+##########################################
+from flask import Flask, request, jsonify, render_template, send_from_directory,url_for
+import os
+import json
+import pdb
+import argparse
+from pydub import AudioSegment
+import math
+import numpy as np
+import shutil
+
+from simuleval.utils import entrypoint
+from simuleval.data.segments import SpeechSegment
+from simuleval.agents import SpeechToSpeechAgent
+from simuleval.agents.actions import WriteAction, ReadAction
+from fairseq.checkpoint_utils import load_model_ensemble_and_task
+from fairseq.models.text_to_speech.hub_interface import TTSHubInterface
+from pathlib import Path
+from typing import Any, Dict, Optional, Union
+from fairseq.data.audio.audio_utils import convert_waveform
+from examples.speech_to_text.data_utils import extract_fbank_features
+import ast
+import math
+import os
+import json
+import numpy as np
+from copy import deepcopy
+import torch
+import torchaudio.compliance.kaldi as kaldi
+import yaml
+from fairseq import checkpoint_utils, tasks, utils, options
+from fairseq.file_io import PathManager
+from fairseq import search
+from fairseq.data.audio.feature_transforms import CompositeAudioFeatureTransform
+import soundfile
+import argparse
+
+SHIFT_SIZE = 10
+WINDOW_SIZE = 25
+ORG_SAMPLE_RATE = 48000
+SAMPLE_RATE = 16000
+FEATURE_DIM = 80
+BOW_PREFIX = "\u2581"
+DEFAULT_EOS = 2
+OFFSET_MS=-1
+Finished=False
+
+ASR={}
+
+S2TT={}
+
+S2ST=[]
+
+class OnlineFeatureExtractor:
+ """
+ Extract speech feature on the fly.
+ """
+
+ def __init__(self, args, cfg):
+ self.shift_size = args.shift_size
+ self.window_size = args.window_size
+ assert self.window_size >= self.shift_size
+
+ self.sample_rate = args.sample_rate
+ self.feature_dim = args.feature_dim
+ self.num_samples_per_shift = int(self.shift_size * self.sample_rate / 1000)
+ self.num_samples_per_window = int(self.window_size * self.sample_rate / 1000)
+ self.len_ms_to_samples = lambda x: x * self.sample_rate / 1000
+ self.previous_residual_samples = []
+ self.global_cmvn = args.global_cmvn
+ self.device = "cuda" if torch.cuda.is_available() else "cpu"
+ self.feature_transforms = CompositeAudioFeatureTransform.from_config_dict(
+ {"feature_transforms": ["utterance_cmvn"]}
+ )
+
+ def clear_cache(self):
+ self.previous_residual_samples = []
+
+ def __call__(self, new_samples, sr=ORG_SAMPLE_RATE):
+ samples = new_samples
+
+ # # num_frames is the number of frames from the new segment
+ num_frames = math.floor(
+ (len(samples) - self.len_ms_to_samples(self.window_size - self.shift_size))
+ / self.num_samples_per_shift
+ )
+
+ # # the number of frames used for feature extraction
+ # # including some part of thte previous segment
+ effective_num_samples = int(
+ num_frames * self.len_ms_to_samples(self.shift_size)
+ + self.len_ms_to_samples(self.window_size - self.shift_size)
+ )
+ samples = samples[:effective_num_samples]
+ waveform, sample_rate = convert_waveform(
+ torch.tensor([samples]), sr, to_mono=True, to_sample_rate=16000
+ )
+ output = extract_fbank_features(waveform, 16000)
+ output = self.transform(output)
+ return torch.tensor(output, device=self.device)
+
+ def transform(self, input):
+ if self.global_cmvn is None:
+ return input
+
+ mean = self.global_cmvn["mean"]
+ std = self.global_cmvn["std"]
+
+ x = np.subtract(input, mean)
+ x = np.divide(x, std)
+ return x
+
+class StreamSpeechS2STAgent(SpeechToSpeechAgent):
+ """
+ Incrementally feed text to this offline Fastspeech2 TTS model,
+ with a minimum numbers of phonemes every chunk.
+ """
+
+ def __init__(self, args):
+ super().__init__(args)
+ self.eos = DEFAULT_EOS
+
+ self.gpu = torch.cuda.is_available()
+ self.device = "cuda" if torch.cuda.is_available() else "cpu"
+
+ self.args = args
+
+ self.load_model_vocab(args)
+
+ self.max_len = args.max_len
+
+ self.force_finish = args.force_finish
+
+ torch.set_grad_enabled(False)
+
+ tgt_dict_mt = self.dict[f"{self.models[0].mt_task_name}"]
+ tgt_dict = self.dict["tgt"]
+ tgt_dict_asr = self.dict["source_unigram"]
+ tgt_dict_st = self.dict["ctc_target_unigram"]
+ args.user_dir=args.agent_dir
+ utils.import_user_module(args)
+ from agent.sequence_generator import SequenceGenerator
+ from agent.ctc_generator import CTCSequenceGenerator
+ from agent.ctc_decoder import CTCDecoder
+ from agent.tts.vocoder import CodeHiFiGANVocoderWithDur
+
+ self.ctc_generator = CTCSequenceGenerator(
+ tgt_dict, self.models, use_incremental_states=False
+ )
+
+ self.asr_ctc_generator = CTCDecoder(tgt_dict_asr, self.models)
+ self.st_ctc_generator = CTCDecoder(tgt_dict_st, self.models)
+
+ self.generator = SequenceGenerator(
+ self.models,
+ tgt_dict,
+ beam_size=1,
+ max_len_a=1,
+ max_len_b=200,
+ max_len=0,
+ min_len=1,
+ normalize_scores=True,
+ len_penalty=1.0,
+ unk_penalty=0.0,
+ temperature=1.0,
+ match_source_len=False,
+ no_repeat_ngram_size=0,
+ search_strategy=search.BeamSearch(tgt_dict),
+ eos=tgt_dict.eos(),
+ symbols_to_strip_from_output=None,
+ )
+
+ self.generator_mt = SequenceGenerator(
+ self.models,
+ tgt_dict_mt,
+ beam_size=1,
+ max_len_a=0,
+ max_len_b=100,
+ max_len=0,
+ min_len=1,
+ normalize_scores=True,
+ len_penalty=1.0,
+ unk_penalty=0.0,
+ temperature=1.0,
+ match_source_len=False,
+ no_repeat_ngram_size=0,
+ search_strategy=search.BeamSearch(tgt_dict_mt),
+ eos=tgt_dict_mt.eos(),
+ symbols_to_strip_from_output=None,
+ use_incremental_states=False,
+ )
+
+ with open(args.vocoder_cfg) as f:
+ vocoder_cfg = json.load(f)
+ self.vocoder = CodeHiFiGANVocoderWithDur(args.vocoder, vocoder_cfg)
+ if self.device == "cuda":
+ self.vocoder = self.vocoder.cuda()
+ self.dur_prediction = args.dur_prediction
+
+ self.lagging_k1 = args.lagging_k1
+ self.lagging_k2 = args.lagging_k2
+ self.segment_size = args.segment_size
+ self.stride_n = args.stride_n
+
+ self.unit_per_subword = args.unit_per_subword
+ self.stride_n2 = args.stride_n2
+
+ if args.extra_output_dir is not None:
+ self.asr_file = Path(args.extra_output_dir + "/asr.txt")
+ self.st_file = Path(args.extra_output_dir + "/st.txt")
+ self.unit_file = Path(args.extra_output_dir + "/unit.txt")
+ self.quiet = False
+ else:
+ self.quiet = True
+
+ self.output_asr_translation = args.output_asr_translation
+
+ self.segment_size=args.segment_size
+
+ if args.segment_size >= 640:
+ self.whole_word = True
+ else:
+ self.whole_word = False
+
+ self.states = self.build_states()
+ self.reset()
+
+ @staticmethod
+ def add_args(parser):
+ parser.add_argument(
+ "--model-path",
+ type=str,
+ required=True,
+ help="path to your pretrained model.",
+ )
+ parser.add_argument(
+ "--data-bin", type=str, required=True, help="Path of data binary"
+ )
+ parser.add_argument(
+ "--config-yaml", type=str, default=None, help="Path to config yaml file"
+ )
+ parser.add_argument(
+ "--multitask-config-yaml",
+ type=str,
+ default=None,
+ help="Path to config yaml file",
+ )
+ parser.add_argument(
+ "--global-stats",
+ type=str,
+ default=None,
+ help="Path to json file containing cmvn stats",
+ )
+ parser.add_argument(
+ "--tgt-splitter-type",
+ type=str,
+ default="SentencePiece",
+ help="Subword splitter type for target text",
+ )
+ parser.add_argument(
+ "--tgt-splitter-path",
+ type=str,
+ default=None,
+ help="Subword splitter model path for target text",
+ )
+ parser.add_argument(
+ "--user-dir",
+ type=str,
+ default="researches/ctc_unity",
+ help="User directory for model",
+ )
+ parser.add_argument(
+ "--agent-dir",
+ type=str,
+ default="agent",
+ help="User directory for agents",
+ )
+ parser.add_argument(
+ "--max-len", type=int, default=200, help="Max length of translation"
+ )
+ parser.add_argument(
+ "--force-finish",
+ default=False,
+ action="store_true",
+ help="Force the model to finish the hypothsis if the source is not finished",
+ )
+ parser.add_argument(
+ "--shift-size",
+ type=int,
+ default=SHIFT_SIZE,
+ help="Shift size of feature extraction window.",
+ )
+ parser.add_argument(
+ "--window-size",
+ type=int,
+ default=WINDOW_SIZE,
+ help="Window size of feature extraction window.",
+ )
+ parser.add_argument(
+ "--sample-rate", type=int, default=ORG_SAMPLE_RATE, help="Sample rate"
+ )
+ parser.add_argument(
+ "--feature-dim",
+ type=int,
+ default=FEATURE_DIM,
+ help="Acoustic feature dimension.",
+ )
+ parser.add_argument(
+ "--vocoder", type=str, required=True, help="path to the CodeHiFiGAN vocoder"
+ )
+ parser.add_argument(
+ "--vocoder-cfg",
+ type=str,
+ required=True,
+ help="path to the CodeHiFiGAN vocoder config",
+ )
+ parser.add_argument(
+ "--dur-prediction",
+ action="store_true",
+ help="enable duration prediction (for reduced/unique code sequences)",
+ )
+ parser.add_argument("--lagging-k1", type=int, default=0, help="lagging number")
+ parser.add_argument("--lagging-k2", type=int, default=0, help="lagging number")
+ parser.add_argument(
+ "--segment-size", type=int, default=320, help="segment-size"
+ )
+ parser.add_argument("--stride-n", type=int, default=1, help="lagging number")
+ parser.add_argument("--stride-n2", type=int, default=1, help="lagging number")
+ parser.add_argument(
+ "--unit-per-subword", type=int, default=15, help="lagging number"
+ )
+ parser.add_argument(
+ "--extra-output-dir", type=str, default=None, help="extra output dir"
+ )
+ parser.add_argument(
+ "--output-asr-translation",
+ type=bool,
+ default=False,
+ help="extra output dir",
+ )
+
+ def reset(self):
+ self.src_seg_num = 0
+ self.tgt_subwords_indices = None
+ self.src_ctc_indices = None
+ self.src_ctc_prefix_length = 0
+ self.tgt_ctc_prefix_length = 0
+ self.tgt_units_indices = None
+ self.prev_output_tokens_mt = None
+ self.tgt_text = []
+ self.mt_decoder_out = None
+ self.unit = None
+ self.wav = []
+ self.post_transcription = ""
+ self.unfinished_wav = None
+ self.states.reset()
+ try:
+ self.generator_mt.reset_incremental_states()
+ self.ctc_generator.reset_incremental_states()
+ except:
+ pass
+
+ def set_chunk_size(self,segment_size):
+ # print(segment_size)
+ self.segment_size=segment_size
+ chunk_size = segment_size // 40
+
+
+ for model in self.models:
+ model.encoder.chunk_size = chunk_size
+
+ if chunk_size >= 16:
+ chunk_size = 16
+ else:
+ chunk_size = 8
+ for conv in model.encoder.subsample.conv_layers:
+ conv.chunk_size = chunk_size
+ for layer in model.encoder.conformer_layers:
+ layer.conv_module.depthwise_conv.chunk_size = chunk_size
+
+ if segment_size >= 640:
+ self.whole_word = True
+ else:
+ self.whole_word = False
+
+
+ def to_device(self, tensor):
+ if self.gpu:
+ return tensor.cuda()
+ else:
+ return tensor.cpu()
+
+ def load_model_vocab(self, args):
+ filename = args.model_path
+ if not os.path.exists(filename):
+ raise IOError("Model file not found: {}".format(filename))
+
+ state = checkpoint_utils.load_checkpoint_to_cpu(filename)
+ state["cfg"].common['user_dir']=args.user_dir
+ utils.import_user_module(state["cfg"].common)
+
+ task_args = state["cfg"]["task"]
+ task_args.data = args.data_bin
+
+ args.global_cmvn = None
+ if args.config_yaml is not None:
+ task_args.config_yaml = args.config_yaml
+ with open(os.path.join(args.data_bin, args.config_yaml), "r") as f:
+ config = yaml.load(f, Loader=yaml.BaseLoader)
+
+ if "global_cmvn" in config:
+ args.global_cmvn = np.load(config["global_cmvn"]["stats_npz_path"])
+
+ self.feature_extractor = OnlineFeatureExtractor(args, config)
+
+ if args.multitask_config_yaml is not None:
+ task_args.multitask_config_yaml = args.multitask_config_yaml
+
+ task = tasks.setup_task(task_args)
+ self.task = task
+
+ overrides = ast.literal_eval(state["cfg"].common_eval.model_overrides)
+
+ models, saved_cfg = checkpoint_utils.load_model_ensemble(
+ utils.split_paths(filename),
+ arg_overrides=overrides,
+ task=task,
+ suffix=state["cfg"].checkpoint.checkpoint_suffix,
+ strict=(state["cfg"].checkpoint.checkpoint_shard_count == 1),
+ num_shards=state["cfg"].checkpoint.checkpoint_shard_count,
+ )
+
+ chunk_size = args.segment_size // 40
+
+ self.models = models
+
+ for model in self.models:
+ model.eval()
+ model.share_memory()
+ if self.gpu:
+ model.cuda()
+ model.encoder.chunk_size = chunk_size
+
+ if chunk_size >= 16:
+ chunk_size = 16
+ else:
+ chunk_size = 8
+ for conv in model.encoder.subsample.conv_layers:
+ conv.chunk_size = chunk_size
+ for layer in model.encoder.conformer_layers:
+ layer.conv_module.depthwise_conv.chunk_size = chunk_size
+
+ # Set dictionary
+ self.dict = {}
+ self.dict["tgt"] = task.target_dictionary
+
+ for k, v in task.multitask_tasks.items():
+ self.dict[k] = v.tgt_dict
+
+ @torch.inference_mode()
+ def policy(self):
+ # print(self.states.source)
+ feature = self.feature_extractor(self.states.source)
+
+ if feature.size(0) == 0 and not self.states.source_finished:
+ return ReadAction()
+
+ src_indices = feature.unsqueeze(0)
+ src_lengths = torch.tensor([feature.size(0)], device=self.device).long()
+
+ self.encoder_outs = self.generator.model.forward_encoder(
+ {"src_tokens": src_indices, "src_lengths": src_lengths}
+ )
+
+ finalized_asr = self.asr_ctc_generator.generate(
+ self.encoder_outs[0], aux_task_name="source_unigram"
+ )
+ asr_probs = torch.exp(finalized_asr[0][0]["lprobs"])
+
+ for i, hypo in enumerate(finalized_asr):
+ i_beam = 0
+ tmp = hypo[i_beam]["tokens"].int()
+ src_ctc_indices = tmp
+ src_ctc_index = hypo[i_beam]["index"]
+ text = "".join([self.dict["source_unigram"][c] for c in tmp])
+ text = text.replace("_", " ")
+ text = text.replace("▁", " ")
+ text = text.replace("", " ")
+ text = text.replace("", "")
+ text = text.replace("", "")
+ if len(text) > 0 and text[0] == " ":
+ text = text[1:]
+ if self.states.source_finished and not self.quiet:
+ with open(self.asr_file, "a") as file:
+ print(text, file=file)
+ if self.output_asr_translation:
+ print("Streaming ASR:", text)
+
+ ASR[len(self.states.source)]=text
+
+
+ finalized_st = self.st_ctc_generator.generate(
+ self.encoder_outs[0], aux_task_name="ctc_target_unigram"
+ )
+ st_probs = torch.exp(finalized_st[0][0]["lprobs"])
+
+ for i, hypo in enumerate(finalized_st):
+ i_beam = 0
+ tmp = hypo[i_beam]["tokens"].int()
+ tgt_ctc_indices = tmp
+ tgt_ctc_index = hypo[i_beam]["index"]
+ text = "".join([self.dict["ctc_target_unigram"][c] for c in tmp])
+ text = text.replace("_", " ")
+ text = text.replace("▁", " ")
+ text = text.replace("", " ")
+ text = text.replace("", "")
+ text = text.replace("", "")
+ if len(text) > 0 and text[0] == " ":
+ text = text[1:]
+
+ if not self.states.source_finished:
+ src_ctc_prefix_length = src_ctc_indices.size(-1)
+ tgt_ctc_prefix_length = tgt_ctc_indices.size(-1)
+
+ self.src_ctc_indices = src_ctc_indices
+ if (
+ src_ctc_prefix_length < self.src_ctc_prefix_length + self.stride_n
+ or tgt_ctc_prefix_length < self.tgt_ctc_prefix_length + self.stride_n
+ ):
+ return ReadAction()
+ self.src_ctc_prefix_length = max(
+ src_ctc_prefix_length, self.src_ctc_prefix_length
+ )
+ self.tgt_ctc_prefix_length = max(
+ tgt_ctc_prefix_length, self.tgt_ctc_prefix_length
+ )
+ subword_tokens = (
+ (tgt_ctc_prefix_length - self.lagging_k1) // self.stride_n
+ ) * self.stride_n
+
+ if self.whole_word:
+ subword_tokens += 1
+ new_subword_tokens = (
+ (subword_tokens - self.tgt_subwords_indices.size(-1))
+ if self.tgt_subwords_indices is not None
+ else subword_tokens
+ )
+
+ if new_subword_tokens < 1:
+ return ReadAction()
+ else:
+ self.src_ctc_indices = src_ctc_indices
+ new_subword_tokens = -1
+
+ new_subword_tokens = int(new_subword_tokens)
+
+ single_model = self.generator.model.single_model
+ mt_decoder = getattr(single_model, f"{single_model.mt_task_name}_decoder")
+
+ # 1. MT decoder
+ finalized_mt = self.generator_mt.generate_decoder(
+ self.encoder_outs,
+ src_indices,
+ src_lengths,
+ {
+ "id": 1,
+ "net_input": {"src_tokens": src_indices, "src_lengths": src_lengths},
+ },
+ self.tgt_subwords_indices,
+ None,
+ None,
+ aux_task_name=single_model.mt_task_name,
+ max_new_tokens=new_subword_tokens,
+ )
+
+ if finalized_mt[0][0]["tokens"][-1] == 2:
+ tgt_subwords_indices = finalized_mt[0][0]["tokens"][:-1].unsqueeze(0)
+ else:
+ tgt_subwords_indices = finalized_mt[0][0]["tokens"].unsqueeze(0)
+
+ if self.whole_word:
+ j = 999999
+ if not self.states.source_finished:
+ for j in range(tgt_subwords_indices.size(-1) - 1, -1, -1):
+ if self.generator_mt.tgt_dict[
+ tgt_subwords_indices[0][j]
+ ].startswith("▁"):
+ break
+ tgt_subwords_indices = tgt_subwords_indices[:, :j]
+ finalized_mt[0][0]["tokens"] = finalized_mt[0][0]["tokens"][:j]
+
+ if j == 0:
+ return ReadAction()
+
+ new_incremental_states = [{}]
+ if (
+ self.generator_mt.incremental_states is not None
+ and self.generator_mt.use_incremental_states
+ ):
+ for k, v in self.generator_mt.incremental_states[0].items():
+ if v["prev_key"].size(2) == v["prev_value"].size(2):
+ new_incremental_states[0][k] = {
+ "prev_key": v["prev_key"][:, :, :j, :].contiguous(),
+ "prev_value": v["prev_value"][:, :, :j, :].contiguous(),
+ "prev_key_padding_mask": None,
+ }
+ else:
+ new_incremental_states[0][k] = {
+ "prev_key": v["prev_key"],
+ "prev_value": v["prev_value"][:, :, :j, :].contiguous(),
+ "prev_key_padding_mask": None,
+ }
+ self.generator_mt.incremental_states = deepcopy(
+ new_incremental_states
+ )
+
+ max_tgt_len = max([len(hypo[0]["tokens"]) for hypo in finalized_mt])
+ if self.whole_word:
+ max_tgt_len += 1
+ prev_output_tokens_mt = (
+ src_indices.new_zeros(src_indices.shape[0], max_tgt_len)
+ .fill_(mt_decoder.padding_idx)
+ .int()
+ )
+
+ for i, hypo in enumerate(finalized_mt):
+ i_beam = 0
+ tmp = hypo[i_beam]["tokens"].int()
+ prev_output_tokens_mt[i, 0] = self.generator_mt.eos
+ if tmp[-1] == self.generator_mt.eos:
+ tmp = tmp[:-1]
+ prev_output_tokens_mt[i, 1 : len(tmp) + 1] = tmp
+
+ tokens = [self.generator_mt.tgt_dict[c] for c in tmp]
+
+ text = "".join(tokens)
+ text = text.replace("_", " ")
+ text = text.replace("▁", " ")
+ text = text.replace("", " ")
+ text = text.replace("", "")
+ text = text.replace("", "")
+ if len(text) > 0 and text[0] == " ":
+ text = text[1:]
+ if self.states.source_finished and not self.quiet:
+ with open(self.st_file, "a") as file:
+ print(text, file=file)
+ if self.output_asr_translation:
+ print("Simultaneous translation:", text)
+
+ S2TT[len(self.states.source)]=text
+
+ if self.tgt_subwords_indices is not None and torch.equal(
+ self.tgt_subwords_indices, tgt_subwords_indices
+ ):
+ if not self.states.source_finished:
+ return ReadAction()
+ else:
+ return WriteAction(
+ SpeechSegment(
+ content=(
+ self.unfinished_wav.tolist()
+ if self.unfinished_wav is not None
+ else []
+ ),
+ sample_rate=SAMPLE_RATE,
+ finished=True,
+ ),
+ finished=True,
+ )
+ self.tgt_subwords_indices = tgt_subwords_indices
+
+ if not self.states.source_finished:
+ if self.prev_output_tokens_mt is not None:
+ if torch.equal(
+ self.prev_output_tokens_mt, prev_output_tokens_mt
+ ) or prev_output_tokens_mt.size(-1) <= self.prev_output_tokens_mt.size(
+ -1
+ ):
+ return ReadAction()
+ self.prev_output_tokens_mt = prev_output_tokens_mt
+ mt_decoder_out = mt_decoder(
+ prev_output_tokens_mt,
+ encoder_out=self.encoder_outs[0],
+ features_only=True,
+ )[0].transpose(0, 1)
+
+ if self.mt_decoder_out is None:
+ self.mt_decoder_out = mt_decoder_out
+ else:
+ self.mt_decoder_out = torch.cat(
+ (self.mt_decoder_out, mt_decoder_out[self.mt_decoder_out.size(0) :]),
+ dim=0,
+ )
+ self.mt_decoder_out = mt_decoder_out
+ x = self.mt_decoder_out
+
+ if getattr(single_model, "proj", None) is not None:
+ x = single_model.proj(x)
+
+ mt_decoder_padding_mask = None
+ if prev_output_tokens_mt.eq(mt_decoder.padding_idx).any():
+ mt_decoder_padding_mask = prev_output_tokens_mt.eq(mt_decoder.padding_idx)
+
+ # 2. T2U encoder
+ if getattr(single_model, "synthesizer_encoder", None) is not None:
+ t2u_encoder_out = single_model.synthesizer_encoder(
+ x,
+ mt_decoder_padding_mask,
+ )
+ else:
+ t2u_encoder_out = {
+ "encoder_out": [x], # T x B x C
+ "encoder_padding_mask": (
+ [mt_decoder_padding_mask]
+ if mt_decoder_padding_mask is not None
+ else []
+ ), # B x T
+ "encoder_embedding": [],
+ "encoder_states": [],
+ "src_tokens": [],
+ "src_lengths": [],
+ }
+
+ if getattr(single_model, "t2u_augmented_cross_attn", False):
+ encoder_outs_aug = [t2u_encoder_out]
+ else:
+ encoder_outs = [t2u_encoder_out]
+ encoder_outs_aug = None
+ finalized = self.ctc_generator.generate(
+ encoder_outs[0],
+ prefix=self.tgt_units_indices,
+ )
+
+ if len(finalized[0][0]["tokens"]) == 0:
+ if not self.states.source_finished:
+ return ReadAction()
+ else:
+ return WriteAction(
+ SpeechSegment(
+ content=(
+ self.unfinished_wav.tolist()
+ if self.unfinished_wav is not None
+ else []
+ ),
+ sample_rate=SAMPLE_RATE,
+ finished=True,
+ ),
+ finished=True,
+ )
+
+ for i, hypo in enumerate(finalized):
+ i_beam = 0
+ tmp = hypo[i_beam]["tokens"].int() # hyp + eos
+ if tmp[-1] == self.generator.eos:
+ tmp = tmp[:-1]
+ unit = []
+ for c in tmp:
+ u = self.generator.tgt_dict[c].replace("", "").replace("", "")
+ if u != "":
+ unit.append(int(u))
+
+ if len(unit) > 0 and unit[0] == " ":
+ unit = unit[1:]
+ text = " ".join([str(_) for _ in unit])
+ if self.states.source_finished and not self.quiet:
+ with open(self.unit_file, "a") as file:
+ print(text, file=file)
+ cur_unit = unit if self.unit is None else unit[len(self.unit) :]
+ if len(unit) < 1 or len(cur_unit) < 1:
+ if not self.states.source_finished:
+ return ReadAction()
+ else:
+ return WriteAction(
+ SpeechSegment(
+ content=(
+ self.unfinished_wav.tolist()
+ if self.unfinished_wav is not None
+ else []
+ ),
+ sample_rate=SAMPLE_RATE,
+ finished=True,
+ ),
+ finished=True,
+ )
+
+ x = {
+ "code": torch.tensor(unit, dtype=torch.long, device=self.device).view(
+ 1, -1
+ ),
+ }
+ wav, dur = self.vocoder(x, self.dur_prediction)
+
+ cur_wav_length = dur[:, -len(cur_unit) :].sum() * 320
+ new_wav = wav[-cur_wav_length:]
+ if self.unfinished_wav is not None and len(self.unfinished_wav) > 0:
+ new_wav = torch.cat((self.unfinished_wav, new_wav), dim=0)
+
+ self.wav = wav
+ self.unit = unit
+
+ # A SpeechSegment has to be returned for speech-to-speech translation system
+ if self.states.source_finished and new_subword_tokens == -1:
+ self.states.target_finished = True
+ # self.reset()
+
+ S2ST.extend(new_wav.tolist())
+ global OFFSET_MS
+ if OFFSET_MS==-1:
+ OFFSET_MS=1000*len(self.states.source)/ORG_SAMPLE_RATE
+
+ return WriteAction(
+ SpeechSegment(
+ content=new_wav.tolist(),
+ sample_rate=SAMPLE_RATE,
+ finished=self.states.source_finished,
+ ),
+ finished=self.states.target_finished,
+ )
+
+def run(source):
+ # if len(S2ST)!=0: return
+ samples, _ = soundfile.read(source, dtype="float32")
+ agent.reset()
+
+ interval=int(agent.segment_size*(ORG_SAMPLE_RATE/1000))
+ cur_idx=0
+ while not agent.states.target_finished:
+ cur_idx+=interval
+ agent.states.source=samples[:cur_idx]
+ agent.states.source_finished=cur_idx>len(samples)
+ action=agent.policy()
+ # print("ASR_RESULT",ASR)
+ # print("S2ST_RESULT",S2ST)
+
+def reset():
+ global OFFSET_MS
+ OFFSET_MS=-1
+ global ASR
+ ASR={}
+ global S2TT
+ S2TT={}
+ global S2ST
+ S2ST=[]
+
+
+def find_largest_key_value(dictionary, N):
+ keys = [key for key in dictionary.keys() if key < N]
+ if not keys:
+ return ""
+ largest_key = max(keys)
+ return dictionary[largest_key]
+
+def merge_audio(left_audio_path, right_audio_path, offset_ms):
+ # 读取左右声道音频文件
+ left_audio = AudioSegment.from_file(left_audio_path)
+ right_audio = AudioSegment.from_file(right_audio_path)
+
+ right_audio=AudioSegment.silent(duration=offset_ms)+right_audio
+
+
+ # 确保两个音频文件具有相同的长度
+ if len(left_audio) > len(right_audio):
+ right_audio += AudioSegment.silent(duration=len(left_audio) - len(right_audio))
+ elif len(left_audio) < len(right_audio):
+ left_audio += AudioSegment.silent(duration=len(right_audio) - len(left_audio))
+
+ # # 将左右声道音频合并
+ # merged_audio = left_audio.overlay(right_audio.pan(1))
+ # # 保存合并后的音频文件
+ # merged_audio.export(output_file, format="wav")
+
+ return left_audio,right_audio
+
+app = Flask(__name__)
+app.config['UPLOAD_FOLDER'] = 'uploads'
+os.makedirs(app.config['UPLOAD_FOLDER'], exist_ok=True)
+
+
+@app.route('/')
+def index():
+ return render_template('index.html')
+
+@app.route('/upload', methods=['POST'])
+def upload():
+ if 'file' not in request.files:
+ return 'No file part', 400
+ file = request.files['file']
+ if file.filename == '':
+ return 'No selected file', 400
+ if file:
+ filepath = os.path.join(app.config['UPLOAD_FOLDER'], file.filename)
+ file.save(filepath)
+ return filepath
+
+@app.route('/uploads/')
+def uploaded_file(filename):
+ latency = request.args.get('latency', default=320, type=int)
+ agent.set_chunk_size(latency)
+
+ path=app.config['UPLOAD_FOLDER']+'/'+filename
+ # pdb.set_trace()
+ # if len(S2ST)==0:
+ reset()
+ run(path)
+ soundfile.write('/'.join(path.split('/')[:-1])+'/output.'+path.split('/')[-1],S2ST,SAMPLE_RATE)
+ left,right=merge_audio(path, '/'.join(path.split('/')[:-1])+'/output.'+path.split('/')[-1], OFFSET_MS)
+ left.export('/'.join(path.split('/')[:-1])+'/input.'+path.split('/')[-1], format="wav")
+ right.export('/'.join(path.split('/')[:-1])+'/output.'+path.split('/')[-1], format="wav")
+ # left=left.split_to_mono()[0]
+ # right=right.split_to_mono()[1]
+ # pdb.set_trace()
+ return send_from_directory(app.config['UPLOAD_FOLDER'], 'input.'+filename)
+
+@app.route('/uploads/output/')
+def uploaded_output_file(filename):
+
+ return send_from_directory(app.config['UPLOAD_FOLDER'], 'output.'+filename)
+
+
+@app.route('/asr/', methods=['GET'])
+def asr(current_time):
+ # asr_result = f"ABCD... {int(current_time * 1000)}"
+ N = current_time*ORG_SAMPLE_RATE
+
+ asr_result=find_largest_key_value(ASR, N)
+ return jsonify(result=asr_result)
+
+@app.route('/translation/', methods=['GET'])
+def translation(current_time):
+ N = current_time*ORG_SAMPLE_RATE
+
+ translation_result=find_largest_key_value(S2TT, N)
+ # translation_result = f"1234... {int(current_time * 1000)}"
+ return jsonify(result=translation_result)
+
+with open('/data/zhangshaolei/StreamSpeech/demo/config.json', 'r') as f:
+ args_dict = json.load(f)
+
+# Initialize agent
+parser = argparse.ArgumentParser()
+StreamSpeechS2STAgent.add_args(parser)
+
+# Create the list of arguments from args_dict
+args_list = []
+# pdb.set_trace()
+for key, value in args_dict.items():
+ if isinstance(value, bool):
+ if value:
+ args_list.append(f'--{key}')
+ else:
+ args_list.append(f'--{key}')
+ args_list.append(str(value))
+
+args = parser.parse_args(args_list)
+
+agent = StreamSpeechS2STAgent(args)
+
+
+
+
+if __name__ == '__main__':
+ app.run(host='0.0.0.0', port=7860, debug=True)
diff --git a/demo/config.json b/demo/config.json
new file mode 100644
index 0000000000000000000000000000000000000000..8d9b4accfc12d3a2030313fea5e4df3fd2da09b9
--- /dev/null
+++ b/demo/config.json
@@ -0,0 +1,12 @@
+{
+ "data-bin": "/data/zhangshaolei/StreamSpeech/configs/fr-en",
+ "user-dir": "/data/zhangshaolei/StreamSpeech/researches/ctc_unity",
+ "agent-dir": "/data/zhangshaolei/StreamSpeech/agent",
+ "model-path": "/data/zhangshaolei/StreamSpeech_model/streamspeech.simultaneous.fr-en.pt",
+ "config-yaml": "config_gcmvn.yaml",
+ "multitask-config-yaml": "config_mtl_asr_st_ctcst.yaml",
+ "segment-size": 320,
+ "vocoder": "/data/zhangshaolei/pretrain_models/unit-based_HiFi-GAN_vocoder/mHuBERT.layer11.km1000.en/g_00500000",
+ "vocoder-cfg": "/data/zhangshaolei/pretrain_models/unit-based_HiFi-GAN_vocoder/mHuBERT.layer11.km1000.en/config.json",
+ "dur-prediction": true
+}
diff --git a/demo/infer.py b/demo/infer.py
new file mode 100644
index 0000000000000000000000000000000000000000..3be1f15a21fdd7ef097a6e67299c8031d87960ce
--- /dev/null
+++ b/demo/infer.py
@@ -0,0 +1,816 @@
+##########################################
+# Simultaneous Speech-to-Speech Translation Agent for StreamSpeech
+#
+# StreamSpeech: Simultaneous Speech-to-Speech Translation with Multi-task Learning (ACL 2024)
+##########################################
+
+from simuleval.utils import entrypoint
+from simuleval.data.segments import SpeechSegment
+from simuleval.agents import SpeechToSpeechAgent
+from simuleval.agents.actions import WriteAction, ReadAction
+from fairseq.checkpoint_utils import load_model_ensemble_and_task
+from fairseq.models.text_to_speech.hub_interface import TTSHubInterface
+from pathlib import Path
+from typing import Any, Dict, Optional, Union
+from fairseq.data.audio.audio_utils import convert_waveform
+from examples.speech_to_text.data_utils import extract_fbank_features
+import ast
+import math
+import os
+import json
+import numpy as np
+from copy import deepcopy
+import torch
+import torchaudio.compliance.kaldi as kaldi
+import yaml
+from fairseq import checkpoint_utils, tasks, utils, options
+from fairseq.file_io import PathManager
+from fairseq import search
+from fairseq.data.audio.feature_transforms import CompositeAudioFeatureTransform
+import soundfile
+
+SHIFT_SIZE = 10
+WINDOW_SIZE = 25
+ORG_SAMPLE_RATE = 48000
+SAMPLE_RATE = 16000
+FEATURE_DIM = 80
+BOW_PREFIX = "\u2581"
+DEFAULT_EOS = 2
+
+
+global ASR
+ASR=''
+
+global S2TT
+S2TT=''
+
+global S2ST
+S2ST=[]
+
+class OnlineFeatureExtractor:
+ """
+ Extract speech feature on the fly.
+ """
+
+ def __init__(self, args, cfg):
+ self.shift_size = args.shift_size
+ self.window_size = args.window_size
+ assert self.window_size >= self.shift_size
+
+ self.sample_rate = args.sample_rate
+ self.feature_dim = args.feature_dim
+ self.num_samples_per_shift = int(self.shift_size * self.sample_rate / 1000)
+ self.num_samples_per_window = int(self.window_size * self.sample_rate / 1000)
+ self.len_ms_to_samples = lambda x: x * self.sample_rate / 1000
+ self.previous_residual_samples = []
+ self.global_cmvn = args.global_cmvn
+ self.device = "cuda" if torch.cuda.is_available() else "cpu"
+ self.feature_transforms = CompositeAudioFeatureTransform.from_config_dict(
+ {"feature_transforms": ["utterance_cmvn"]}
+ )
+
+ def clear_cache(self):
+ self.previous_residual_samples = []
+
+ def __call__(self, new_samples, sr=ORG_SAMPLE_RATE):
+ samples = new_samples
+
+ # # num_frames is the number of frames from the new segment
+ num_frames = math.floor(
+ (len(samples) - self.len_ms_to_samples(self.window_size - self.shift_size))
+ / self.num_samples_per_shift
+ )
+
+ # # the number of frames used for feature extraction
+ # # including some part of thte previous segment
+ effective_num_samples = int(
+ num_frames * self.len_ms_to_samples(self.shift_size)
+ + self.len_ms_to_samples(self.window_size - self.shift_size)
+ )
+ samples = samples[:effective_num_samples]
+ waveform, sample_rate = convert_waveform(
+ torch.tensor([samples]), sr, to_mono=True, to_sample_rate=16000
+ )
+ output = extract_fbank_features(waveform, 16000)
+ output = self.transform(output)
+ return torch.tensor(output, device=self.device)
+
+ def transform(self, input):
+ if self.global_cmvn is None:
+ return input
+
+ mean = self.global_cmvn["mean"]
+ std = self.global_cmvn["std"]
+
+ x = np.subtract(input, mean)
+ x = np.divide(x, std)
+ return x
+
+class StreamSpeechS2STAgent(SpeechToSpeechAgent):
+ """
+ Incrementally feed text to this offline Fastspeech2 TTS model,
+ with a minimum numbers of phonemes every chunk.
+ """
+
+ def __init__(self, args):
+ super().__init__(args)
+ self.eos = DEFAULT_EOS
+
+ self.gpu = torch.cuda.is_available()
+ self.device = "cuda" if torch.cuda.is_available() else "cpu"
+
+ self.args = args
+
+ self.load_model_vocab(args)
+
+ self.max_len = args.max_len
+
+ self.force_finish = args.force_finish
+
+ torch.set_grad_enabled(False)
+
+ tgt_dict_mt = self.dict[f"{self.models[0].mt_task_name}"]
+ tgt_dict = self.dict["tgt"]
+ tgt_dict_asr = self.dict["source_unigram"]
+ tgt_dict_st = self.dict["ctc_target_unigram"]
+ args.user_dir=args.agent_dir
+ utils.import_user_module(args)
+ from agent.sequence_generator import SequenceGenerator
+ from agent.ctc_generator import CTCSequenceGenerator
+ from agent.ctc_decoder import CTCDecoder
+ from agent.tts.vocoder import CodeHiFiGANVocoderWithDur
+
+ self.ctc_generator = CTCSequenceGenerator(
+ tgt_dict, self.models, use_incremental_states=False
+ )
+
+ self.asr_ctc_generator = CTCDecoder(tgt_dict_asr, self.models)
+ self.st_ctc_generator = CTCDecoder(tgt_dict_st, self.models)
+
+ self.generator = SequenceGenerator(
+ self.models,
+ tgt_dict,
+ beam_size=1,
+ max_len_a=1,
+ max_len_b=200,
+ max_len=0,
+ min_len=1,
+ normalize_scores=True,
+ len_penalty=1.0,
+ unk_penalty=0.0,
+ temperature=1.0,
+ match_source_len=False,
+ no_repeat_ngram_size=0,
+ search_strategy=search.BeamSearch(tgt_dict),
+ eos=tgt_dict.eos(),
+ symbols_to_strip_from_output=None,
+ )
+
+ self.generator_mt = SequenceGenerator(
+ self.models,
+ tgt_dict_mt,
+ beam_size=1,
+ max_len_a=0,
+ max_len_b=100,
+ max_len=0,
+ min_len=1,
+ normalize_scores=True,
+ len_penalty=1.0,
+ unk_penalty=0.0,
+ temperature=1.0,
+ match_source_len=False,
+ no_repeat_ngram_size=0,
+ search_strategy=search.BeamSearch(tgt_dict_mt),
+ eos=tgt_dict_mt.eos(),
+ symbols_to_strip_from_output=None,
+ use_incremental_states=False,
+ )
+
+ with open(args.vocoder_cfg) as f:
+ vocoder_cfg = json.load(f)
+ self.vocoder = CodeHiFiGANVocoderWithDur(args.vocoder, vocoder_cfg)
+ if self.device == "cuda":
+ self.vocoder = self.vocoder.cuda()
+ self.dur_prediction = args.dur_prediction
+
+ self.lagging_k1 = args.lagging_k1
+ self.lagging_k2 = args.lagging_k2
+ self.segment_size = args.segment_size
+ self.stride_n = args.stride_n
+
+ self.unit_per_subword = args.unit_per_subword
+ self.stride_n2 = args.stride_n2
+
+ if args.extra_output_dir is not None:
+ self.asr_file = Path(args.extra_output_dir + "/asr.txt")
+ self.st_file = Path(args.extra_output_dir + "/st.txt")
+ self.unit_file = Path(args.extra_output_dir + "/unit.txt")
+ self.quiet = False
+ else:
+ self.quiet = True
+
+ self.output_asr_translation = args.output_asr_translation
+
+ if args.segment_size >= 640:
+ self.whole_word = True
+ else:
+ self.whole_word = False
+
+ self.states = self.build_states()
+ self.reset()
+
+ @staticmethod
+ def add_args(parser):
+ parser.add_argument(
+ "--model-path",
+ type=str,
+ required=True,
+ help="path to your pretrained model.",
+ )
+ parser.add_argument(
+ "--data-bin", type=str, required=True, help="Path of data binary"
+ )
+ parser.add_argument(
+ "--config-yaml", type=str, default=None, help="Path to config yaml file"
+ )
+ parser.add_argument(
+ "--multitask-config-yaml",
+ type=str,
+ default=None,
+ help="Path to config yaml file",
+ )
+ parser.add_argument(
+ "--global-stats",
+ type=str,
+ default=None,
+ help="Path to json file containing cmvn stats",
+ )
+ parser.add_argument(
+ "--tgt-splitter-type",
+ type=str,
+ default="SentencePiece",
+ help="Subword splitter type for target text",
+ )
+ parser.add_argument(
+ "--tgt-splitter-path",
+ type=str,
+ default=None,
+ help="Subword splitter model path for target text",
+ )
+ parser.add_argument(
+ "--user-dir",
+ type=str,
+ default="researches/ctc_unity",
+ help="User directory for model",
+ )
+ parser.add_argument(
+ "--agent-dir",
+ type=str,
+ default="agent",
+ help="User directory for agents",
+ )
+ parser.add_argument(
+ "--max-len", type=int, default=200, help="Max length of translation"
+ )
+ parser.add_argument(
+ "--force-finish",
+ default=False,
+ action="store_true",
+ help="Force the model to finish the hypothsis if the source is not finished",
+ )
+ parser.add_argument(
+ "--shift-size",
+ type=int,
+ default=SHIFT_SIZE,
+ help="Shift size of feature extraction window.",
+ )
+ parser.add_argument(
+ "--window-size",
+ type=int,
+ default=WINDOW_SIZE,
+ help="Window size of feature extraction window.",
+ )
+ parser.add_argument(
+ "--sample-rate", type=int, default=ORG_SAMPLE_RATE, help="Sample rate"
+ )
+ parser.add_argument(
+ "--feature-dim",
+ type=int,
+ default=FEATURE_DIM,
+ help="Acoustic feature dimension.",
+ )
+ parser.add_argument(
+ "--vocoder", type=str, required=True, help="path to the CodeHiFiGAN vocoder"
+ )
+ parser.add_argument(
+ "--vocoder-cfg",
+ type=str,
+ required=True,
+ help="path to the CodeHiFiGAN vocoder config",
+ )
+ parser.add_argument(
+ "--dur-prediction",
+ action="store_true",
+ help="enable duration prediction (for reduced/unique code sequences)",
+ )
+ parser.add_argument("--lagging-k1", type=int, default=0, help="lagging number")
+ parser.add_argument("--lagging-k2", type=int, default=0, help="lagging number")
+ parser.add_argument(
+ "--segment-size", type=int, default=320, help="segment-size"
+ )
+ parser.add_argument("--stride-n", type=int, default=1, help="lagging number")
+ parser.add_argument("--stride-n2", type=int, default=1, help="lagging number")
+ parser.add_argument(
+ "--unit-per-subword", type=int, default=15, help="lagging number"
+ )
+ parser.add_argument(
+ "--extra-output-dir", type=str, default=None, help="extra output dir"
+ )
+ parser.add_argument(
+ "--output-asr-translation",
+ type=bool,
+ default=False,
+ help="extra output dir",
+ )
+
+ def reset(self):
+ self.src_seg_num = 0
+ self.tgt_subwords_indices = None
+ self.src_ctc_indices = None
+ self.src_ctc_prefix_length = 0
+ self.tgt_ctc_prefix_length = 0
+ self.tgt_units_indices = None
+ self.prev_output_tokens_mt = None
+ self.tgt_text = []
+ self.mt_decoder_out = None
+ self.unit = None
+ self.wav = []
+ self.post_transcription = ""
+ self.unfinished_wav = None
+ self.states.reset()
+ try:
+ self.generator_mt.reset_incremental_states()
+ self.ctc_generator.reset_incremental_states()
+ except:
+ pass
+
+ def to_device(self, tensor):
+ if self.gpu:
+ return tensor.cuda()
+ else:
+ return tensor.cpu()
+
+ def load_model_vocab(self, args):
+ filename = args.model_path
+ if not os.path.exists(filename):
+ raise IOError("Model file not found: {}".format(filename))
+
+ state = checkpoint_utils.load_checkpoint_to_cpu(filename)
+ state["cfg"].common['user_dir']=args.user_dir
+ utils.import_user_module(state["cfg"].common)
+
+ task_args = state["cfg"]["task"]
+ task_args.data = args.data_bin
+
+ args.global_cmvn = None
+ if args.config_yaml is not None:
+ task_args.config_yaml = args.config_yaml
+ with open(os.path.join(args.data_bin, args.config_yaml), "r") as f:
+ config = yaml.load(f, Loader=yaml.BaseLoader)
+
+ if "global_cmvn" in config:
+ args.global_cmvn = np.load(config["global_cmvn"]["stats_npz_path"])
+
+ self.feature_extractor = OnlineFeatureExtractor(args, config)
+
+ if args.multitask_config_yaml is not None:
+ task_args.multitask_config_yaml = args.multitask_config_yaml
+
+ task = tasks.setup_task(task_args)
+ self.task = task
+
+ overrides = ast.literal_eval(state["cfg"].common_eval.model_overrides)
+
+ models, saved_cfg = checkpoint_utils.load_model_ensemble(
+ utils.split_paths(filename),
+ arg_overrides=overrides,
+ task=task,
+ suffix=state["cfg"].checkpoint.checkpoint_suffix,
+ strict=(state["cfg"].checkpoint.checkpoint_shard_count == 1),
+ num_shards=state["cfg"].checkpoint.checkpoint_shard_count,
+ )
+
+ chunk_size = args.segment_size // 40
+
+ self.models = models
+
+ for model in self.models:
+ model.eval()
+ model.share_memory()
+ if self.gpu:
+ model.cuda()
+ model.encoder.chunk_size = chunk_size
+
+ if chunk_size >= 16:
+ chunk_size = 16
+ else:
+ chunk_size = 8
+ for conv in model.encoder.subsample.conv_layers:
+ conv.chunk_size = chunk_size
+ for layer in model.encoder.conformer_layers:
+ layer.conv_module.depthwise_conv.chunk_size = chunk_size
+
+ # Set dictionary
+ self.dict = {}
+ self.dict["tgt"] = task.target_dictionary
+
+ for k, v in task.multitask_tasks.items():
+ self.dict[k] = v.tgt_dict
+
+ @torch.inference_mode()
+ def policy(self):
+ # print(self.states.source)
+ feature = self.feature_extractor(self.states.source)
+
+ if feature.size(0) == 0 and not self.states.source_finished:
+ return ReadAction()
+
+ src_indices = feature.unsqueeze(0)
+ src_lengths = torch.tensor([feature.size(0)], device=self.device).long()
+
+ self.encoder_outs = self.generator.model.forward_encoder(
+ {"src_tokens": src_indices, "src_lengths": src_lengths}
+ )
+
+ finalized_asr = self.asr_ctc_generator.generate(
+ self.encoder_outs[0], aux_task_name="source_unigram"
+ )
+ asr_probs = torch.exp(finalized_asr[0][0]["lprobs"])
+
+ for i, hypo in enumerate(finalized_asr):
+ i_beam = 0
+ tmp = hypo[i_beam]["tokens"].int()
+ src_ctc_indices = tmp
+ src_ctc_index = hypo[i_beam]["index"]
+ text = "".join([self.dict["source_unigram"][c] for c in tmp])
+ text = text.replace("_", " ")
+ text = text.replace("▁", " ")
+ text = text.replace("", " ")
+ text = text.replace("", "")
+ text = text.replace("", "")
+ if len(text) > 0 and text[0] == " ":
+ text = text[1:]
+ if self.states.source_finished and not self.quiet:
+ with open(self.asr_file, "a") as file:
+ print(text, file=file)
+ if self.output_asr_translation:
+ print("Streaming ASR:", text)
+
+ global ASR
+ ASR=text
+
+ finalized_st = self.st_ctc_generator.generate(
+ self.encoder_outs[0], aux_task_name="ctc_target_unigram"
+ )
+ st_probs = torch.exp(finalized_st[0][0]["lprobs"])
+
+ for i, hypo in enumerate(finalized_st):
+ i_beam = 0
+ tmp = hypo[i_beam]["tokens"].int()
+ tgt_ctc_indices = tmp
+ tgt_ctc_index = hypo[i_beam]["index"]
+ text = "".join([self.dict["ctc_target_unigram"][c] for c in tmp])
+ text = text.replace("_", " ")
+ text = text.replace("▁", " ")
+ text = text.replace("", " ")
+ text = text.replace("", "")
+ text = text.replace("", "")
+ if len(text) > 0 and text[0] == " ":
+ text = text[1:]
+
+ if not self.states.source_finished:
+ src_ctc_prefix_length = src_ctc_indices.size(-1)
+ tgt_ctc_prefix_length = tgt_ctc_indices.size(-1)
+
+ self.src_ctc_indices = src_ctc_indices
+ if (
+ src_ctc_prefix_length < self.src_ctc_prefix_length + self.stride_n
+ or tgt_ctc_prefix_length < self.tgt_ctc_prefix_length + self.stride_n
+ ):
+ return ReadAction()
+ self.src_ctc_prefix_length = max(
+ src_ctc_prefix_length, self.src_ctc_prefix_length
+ )
+ self.tgt_ctc_prefix_length = max(
+ tgt_ctc_prefix_length, self.tgt_ctc_prefix_length
+ )
+ subword_tokens = (
+ (tgt_ctc_prefix_length - self.lagging_k1) // self.stride_n
+ ) * self.stride_n
+
+ if self.whole_word:
+ subword_tokens += 1
+ new_subword_tokens = (
+ (subword_tokens - self.tgt_subwords_indices.size(-1))
+ if self.tgt_subwords_indices is not None
+ else subword_tokens
+ )
+
+ if new_subword_tokens < 1:
+ return ReadAction()
+ else:
+ self.src_ctc_indices = src_ctc_indices
+ new_subword_tokens = -1
+
+ new_subword_tokens = int(new_subword_tokens)
+
+ single_model = self.generator.model.single_model
+ mt_decoder = getattr(single_model, f"{single_model.mt_task_name}_decoder")
+
+ # 1. MT decoder
+ finalized_mt = self.generator_mt.generate_decoder(
+ self.encoder_outs,
+ src_indices,
+ src_lengths,
+ {
+ "id": 1,
+ "net_input": {"src_tokens": src_indices, "src_lengths": src_lengths},
+ },
+ self.tgt_subwords_indices,
+ None,
+ None,
+ aux_task_name=single_model.mt_task_name,
+ max_new_tokens=new_subword_tokens,
+ )
+
+ if finalized_mt[0][0]["tokens"][-1] == 2:
+ tgt_subwords_indices = finalized_mt[0][0]["tokens"][:-1].unsqueeze(0)
+ else:
+ tgt_subwords_indices = finalized_mt[0][0]["tokens"].unsqueeze(0)
+
+ if self.whole_word:
+ j = 999999
+ if not self.states.source_finished:
+ for j in range(tgt_subwords_indices.size(-1) - 1, -1, -1):
+ if self.generator_mt.tgt_dict[
+ tgt_subwords_indices[0][j]
+ ].startswith("▁"):
+ break
+ tgt_subwords_indices = tgt_subwords_indices[:, :j]
+ finalized_mt[0][0]["tokens"] = finalized_mt[0][0]["tokens"][:j]
+
+ if j == 0:
+ return ReadAction()
+
+ new_incremental_states = [{}]
+ if (
+ self.generator_mt.incremental_states is not None
+ and self.generator_mt.use_incremental_states
+ ):
+ for k, v in self.generator_mt.incremental_states[0].items():
+ if v["prev_key"].size(2) == v["prev_value"].size(2):
+ new_incremental_states[0][k] = {
+ "prev_key": v["prev_key"][:, :, :j, :].contiguous(),
+ "prev_value": v["prev_value"][:, :, :j, :].contiguous(),
+ "prev_key_padding_mask": None,
+ }
+ else:
+ new_incremental_states[0][k] = {
+ "prev_key": v["prev_key"],
+ "prev_value": v["prev_value"][:, :, :j, :].contiguous(),
+ "prev_key_padding_mask": None,
+ }
+ self.generator_mt.incremental_states = deepcopy(
+ new_incremental_states
+ )
+
+ max_tgt_len = max([len(hypo[0]["tokens"]) for hypo in finalized_mt])
+ if self.whole_word:
+ max_tgt_len += 1
+ prev_output_tokens_mt = (
+ src_indices.new_zeros(src_indices.shape[0], max_tgt_len)
+ .fill_(mt_decoder.padding_idx)
+ .int()
+ )
+
+ for i, hypo in enumerate(finalized_mt):
+ i_beam = 0
+ tmp = hypo[i_beam]["tokens"].int()
+ prev_output_tokens_mt[i, 0] = self.generator_mt.eos
+ if tmp[-1] == self.generator_mt.eos:
+ tmp = tmp[:-1]
+ prev_output_tokens_mt[i, 1 : len(tmp) + 1] = tmp
+
+ tokens = [self.generator_mt.tgt_dict[c] for c in tmp]
+
+ text = "".join(tokens)
+ text = text.replace("_", " ")
+ text = text.replace("▁", " ")
+ text = text.replace("", " ")
+ text = text.replace("", "")
+ text = text.replace("", "")
+ if len(text) > 0 and text[0] == " ":
+ text = text[1:]
+ if self.states.source_finished and not self.quiet:
+ with open(self.st_file, "a") as file:
+ print(text, file=file)
+ if self.output_asr_translation:
+ print("Simultaneous translation:", text)
+
+ global S2TT
+ S2TT=text
+
+ if self.tgt_subwords_indices is not None and torch.equal(
+ self.tgt_subwords_indices, tgt_subwords_indices
+ ):
+ if not self.states.source_finished:
+ return ReadAction()
+ else:
+ return WriteAction(
+ SpeechSegment(
+ content=(
+ self.unfinished_wav.tolist()
+ if self.unfinished_wav is not None
+ else []
+ ),
+ sample_rate=SAMPLE_RATE,
+ finished=True,
+ ),
+ finished=True,
+ )
+ self.tgt_subwords_indices = tgt_subwords_indices
+
+ if not self.states.source_finished:
+ if self.prev_output_tokens_mt is not None:
+ if torch.equal(
+ self.prev_output_tokens_mt, prev_output_tokens_mt
+ ) or prev_output_tokens_mt.size(-1) <= self.prev_output_tokens_mt.size(
+ -1
+ ):
+ return ReadAction()
+ self.prev_output_tokens_mt = prev_output_tokens_mt
+ mt_decoder_out = mt_decoder(
+ prev_output_tokens_mt,
+ encoder_out=self.encoder_outs[0],
+ features_only=True,
+ )[0].transpose(0, 1)
+
+ if self.mt_decoder_out is None:
+ self.mt_decoder_out = mt_decoder_out
+ else:
+ self.mt_decoder_out = torch.cat(
+ (self.mt_decoder_out, mt_decoder_out[self.mt_decoder_out.size(0) :]),
+ dim=0,
+ )
+ self.mt_decoder_out = mt_decoder_out
+ x = self.mt_decoder_out
+
+ if getattr(single_model, "proj", None) is not None:
+ x = single_model.proj(x)
+
+ mt_decoder_padding_mask = None
+ if prev_output_tokens_mt.eq(mt_decoder.padding_idx).any():
+ mt_decoder_padding_mask = prev_output_tokens_mt.eq(mt_decoder.padding_idx)
+
+ # 2. T2U encoder
+ if getattr(single_model, "synthesizer_encoder", None) is not None:
+ t2u_encoder_out = single_model.synthesizer_encoder(
+ x,
+ mt_decoder_padding_mask,
+ )
+ else:
+ t2u_encoder_out = {
+ "encoder_out": [x], # T x B x C
+ "encoder_padding_mask": (
+ [mt_decoder_padding_mask]
+ if mt_decoder_padding_mask is not None
+ else []
+ ), # B x T
+ "encoder_embedding": [],
+ "encoder_states": [],
+ "src_tokens": [],
+ "src_lengths": [],
+ }
+
+ if getattr(single_model, "t2u_augmented_cross_attn", False):
+ encoder_outs_aug = [t2u_encoder_out]
+ else:
+ encoder_outs = [t2u_encoder_out]
+ encoder_outs_aug = None
+ finalized = self.ctc_generator.generate(
+ encoder_outs[0],
+ prefix=self.tgt_units_indices,
+ )
+
+ if len(finalized[0][0]["tokens"]) == 0:
+ if not self.states.source_finished:
+ return ReadAction()
+ else:
+ return WriteAction(
+ SpeechSegment(
+ content=(
+ self.unfinished_wav.tolist()
+ if self.unfinished_wav is not None
+ else []
+ ),
+ sample_rate=SAMPLE_RATE,
+ finished=True,
+ ),
+ finished=True,
+ )
+
+ for i, hypo in enumerate(finalized):
+ i_beam = 0
+ tmp = hypo[i_beam]["tokens"].int() # hyp + eos
+ if tmp[-1] == self.generator.eos:
+ tmp = tmp[:-1]
+ unit = []
+ for c in tmp:
+ u = self.generator.tgt_dict[c].replace("", "").replace("", "")
+ if u != "":
+ unit.append(int(u))
+
+ if len(unit) > 0 and unit[0] == " ":
+ unit = unit[1:]
+ text = " ".join([str(_) for _ in unit])
+ if self.states.source_finished and not self.quiet:
+ with open(self.unit_file, "a") as file:
+ print(text, file=file)
+ cur_unit = unit if self.unit is None else unit[len(self.unit) :]
+ if len(unit) < 1 or len(cur_unit) < 1:
+ if not self.states.source_finished:
+ return ReadAction()
+ else:
+ return WriteAction(
+ SpeechSegment(
+ content=(
+ self.unfinished_wav.tolist()
+ if self.unfinished_wav is not None
+ else []
+ ),
+ sample_rate=SAMPLE_RATE,
+ finished=True,
+ ),
+ finished=True,
+ )
+
+ x = {
+ "code": torch.tensor(unit, dtype=torch.long, device=self.device).view(
+ 1, -1
+ ),
+ }
+ wav, dur = self.vocoder(x, self.dur_prediction)
+
+ cur_wav_length = dur[:, -len(cur_unit) :].sum() * 320
+ new_wav = wav[-cur_wav_length:]
+ if self.unfinished_wav is not None and len(self.unfinished_wav) > 0:
+ new_wav = torch.cat((self.unfinished_wav, new_wav), dim=0)
+
+ self.wav = wav
+ self.unit = unit
+
+ # A SpeechSegment has to be returned for speech-to-speech translation system
+ if self.states.source_finished and new_subword_tokens == -1:
+ self.states.target_finished = True
+ # self.reset()
+
+ global S2ST
+ S2ST_RESULT=S2ST.extend(new_wav.tolist())
+
+ return WriteAction(
+ SpeechSegment(
+ content=new_wav.tolist(),
+ sample_rate=SAMPLE_RATE,
+ finished=self.states.source_finished,
+ ),
+ finished=self.states.target_finished,
+ )
+
+
+
+
+def run(source):
+ samples, _ = soundfile.read(source, dtype="float32")
+
+ interval=int(args.segment_size*(ORG_SAMPLE_RATE/1000))
+ cur_idx=0
+ while not agent.states.target_finished:
+ cur_idx+=interval
+ agent.states.source=samples[:cur_idx]
+ agent.states.source_finished=cur_idx>len(samples)
+ action=agent.policy()
+ print("ASR_RESULT",ASR)
+ print("S2ST_RESULT",S2ST)
+
+
+if __name__ == "__main__":
+ import argparse
+
+ parser = argparse.ArgumentParser()
+ StreamSpeechS2STAgent.add_args(parser)
+ args = parser.parse_args()
+
+ agent = StreamSpeechS2STAgent(args)
+
+ source="StreamSpeech/example/wavs/common_voice_fr_17301936.mp3"
+ run(source)
\ No newline at end of file
diff --git a/demo/run_one_audio.sh b/demo/run_one_audio.sh
new file mode 100644
index 0000000000000000000000000000000000000000..a418fcbb7e43674a1d36a5ed22888bc6d4c9cf9d
--- /dev/null
+++ b/demo/run_one_audio.sh
@@ -0,0 +1,18 @@
+export CUDA_VISIBLE_DEVICES=0
+
+ROOT=/data/zhangshaolei/StreamSpeech # path to StreamSpeech repo
+PRETRAIN_ROOT=/data/zhangshaolei/pretrain_models
+VOCODER_CKPT=$PRETRAIN_ROOT/unit-based_HiFi-GAN_vocoder/mHuBERT.layer11.km1000.en/g_00500000 # path to downloaded Unit-based HiFi-GAN Vocoder
+VOCODER_CFG=$PRETRAIN_ROOT/unit-based_HiFi-GAN_vocoder/mHuBERT.layer11.km1000.en/config.json # path to downloaded Unit-based HiFi-GAN Vocoder
+
+LANG=fr
+file=/streamspeech.simultaneous.${LANG}-en.pt # path to downloaded StreamSpeech model
+output_dir=$ROOT/res/streamspeech.simultaneous.${LANG}-en/simul-s2st
+
+chunk_size=320 #ms
+PYTHONPATH=$ROOT/fairseq python infer.py --data-bin ${ROOT}/configs/${LANG}-en \
+ --user-dir ${ROOT}/researches/ctc_unity --agent-dir ${ROOT}/agent \
+ --model-path $file \
+ --config-yaml config_gcmvn.yaml --multitask-config-yaml config_mtl_asr_st_ctcst.yaml \
+ --segment-size $chunk_size \
+ --vocoder $VOCODER_CKPT --vocoder-cfg $VOCODER_CFG --dur-prediction
\ No newline at end of file
diff --git a/demo/templates/index.html b/demo/templates/index.html
new file mode 100644
index 0000000000000000000000000000000000000000..be314791d0ea0aa8774aea017f3e073a0c6da5ce
--- /dev/null
+++ b/demo/templates/index.html
@@ -0,0 +1,364 @@
+
+
+
+
+
+ StreamSpeech
+
+
+
+
+
+
+
+
ACL 2024
+
StreamSpeech: Simultaneous Speech-to-Speech Translation with Multi-task Learning
+
+
Authors: Shaolei Zhang, Qingkai Fang, Shoutao Guo, Zhengrui Ma, Min Zhang, Yang Feng*
+
💡StreamSpeech is an "All in One" seamless model for offline and simultaneous speech recognition, speech translation and speech synthesis under any latency.
+
+
+
+
+
+
+
+
+
Streaming Inputs
+
+
+
+
+
+
Streaming Speech Recognition
+
+
+
+
Simultaneous Speech-to-Text Translation
+
+
+
+
Simultaneous Speech-to-Speech Translation
+
+
+
+
+
+
+
diff --git a/example/outputs/offline_s2st.streamspeech.common_voice_fr_17301936.merge.mov b/example/outputs/offline_s2st.streamspeech.common_voice_fr_17301936.merge.mov
new file mode 100644
index 0000000000000000000000000000000000000000..83dadde687bc42d455fa287a12407dcc93161a1c
--- /dev/null
+++ b/example/outputs/offline_s2st.streamspeech.common_voice_fr_17301936.merge.mov
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0e6c83d4af8b7b0957d6e5337948fd72dfcecf48f516fcb7f3a152875b6cd090
+size 136600
diff --git a/example/outputs/offline_tts.streamspeech.common_voice_fr_17301936.mov b/example/outputs/offline_tts.streamspeech.common_voice_fr_17301936.mov
new file mode 100644
index 0000000000000000000000000000000000000000..4569500558d58a2f1bce2219ac9dc6ed8faecd66
--- /dev/null
+++ b/example/outputs/offline_tts.streamspeech.common_voice_fr_17301936.mov
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b3b7ec5123c26705e4d232a1d7ab87e33d4d95cd7dc857ad84741efbb0caded6
+size 46946
diff --git a/example/outputs/simul_s2st.streamspeech.common_voice_fr_17301936.merge.mov b/example/outputs/simul_s2st.streamspeech.common_voice_fr_17301936.merge.mov
new file mode 100644
index 0000000000000000000000000000000000000000..03cbb8cb45f3e8b0bdfe23ba1a0ff65b18c39dcd
--- /dev/null
+++ b/example/outputs/simul_s2st.streamspeech.common_voice_fr_17301936.merge.mov
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:70c3119c11d3439931f2dc2cef65db467289a0eb495b8b7f6b6268853cf493ec
+size 98573
diff --git a/example/outputs/simul_tts.streamspeech.common_voice_fr_17301936.mov b/example/outputs/simul_tts.streamspeech.common_voice_fr_17301936.mov
new file mode 100644
index 0000000000000000000000000000000000000000..ba46c63c66a848cbc01518663cc25a5b1133944e
--- /dev/null
+++ b/example/outputs/simul_tts.streamspeech.common_voice_fr_17301936.mov
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8651e3481a0a4c8de8a63a8d0561c462b82ca1f8e37fe709ed78cec31d0ad258
+size 48476
diff --git a/example/source.txt b/example/source.txt
new file mode 100644
index 0000000000000000000000000000000000000000..4363741d39e9a4c494ed0bd639383a556b9ce7f2
--- /dev/null
+++ b/example/source.txt
@@ -0,0 +1,2 @@
+je voulais soumettre cette idée à la réflexion de lassemblée nationale
+jai donc lexpérience des années passées jen dirai un mot tout à lheure
diff --git a/example/target.txt b/example/target.txt
new file mode 100644
index 0000000000000000000000000000000000000000..66154487d782d954c8a49d69f74b79368f6426b6
--- /dev/null
+++ b/example/target.txt
@@ -0,0 +1,2 @@
+i wanted to submit this idea for the national assembly to think about it
+i therefore have the experience of the passed years i'll say a few words about that later
diff --git a/example/wav_list.txt b/example/wav_list.txt
new file mode 100644
index 0000000000000000000000000000000000000000..8c3d4470717ae13d1c903886ef324919b33466b6
--- /dev/null
+++ b/example/wav_list.txt
@@ -0,0 +1,2 @@
+example/wavs/common_voice_fr_17767732.mp3
+example/wavs/common_voice_fr_17301936.mp3
diff --git a/example/wavs/common_voice_fr_17301936.mp3 b/example/wavs/common_voice_fr_17301936.mp3
new file mode 100644
index 0000000000000000000000000000000000000000..61b6c432480c48a813e20956d335080fbc7f112a
--- /dev/null
+++ b/example/wavs/common_voice_fr_17301936.mp3
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5c694c7b965c3ab01c8d32f2bb18038221fe3ffe16e6933843c153ef4f6d4f7f
+size 34797
diff --git a/example/wavs/common_voice_fr_17767732.mp3 b/example/wavs/common_voice_fr_17767732.mp3
new file mode 100644
index 0000000000000000000000000000000000000000..3d290ba826616d51b52e041fc3ad9e46df1f5680
--- /dev/null
+++ b/example/wavs/common_voice_fr_17767732.mp3
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bc28897d9173963fa4e0ebf7a98b3edfdb808fb0501d9b3de7e827471a2dec89
+size 31917
diff --git a/fairseq/.circleci/config.yml b/fairseq/.circleci/config.yml
new file mode 100644
index 0000000000000000000000000000000000000000..7710c78c2fb2508ee824932971b4ec2a6caccab2
--- /dev/null
+++ b/fairseq/.circleci/config.yml
@@ -0,0 +1,128 @@
+# Use 2.1 for orbs
+version: 2.1
+
+# -------------------------------------------------------------------------------------
+# Environments to run the jobs in
+# -------------------------------------------------------------------------------------
+gpu: &gpu
+ environment:
+ CUDA_VERSION: "11.2"
+ machine:
+ image: ubuntu-2004-cuda-11.2:202103-01
+ resource_class: gpu.nvidia.medium.multi
+
+
+# -------------------------------------------------------------------------------------
+# Re-usable commands
+# -------------------------------------------------------------------------------------
+cache_key: &cache_key cache-key-{{ .Environment.CIRCLE_JOB }}-{{ checksum ".circleci/config.yml" }}-{{ checksum "setup.py"}}
+
+install_dep_pt1_10: &install_dep_pt1_10
+ - run:
+ name: Install Pytorch Dependencies
+ command: |
+ source activate fairseq
+ pip install --upgrade setuptools
+ pip install torch==1.10.1+cu111 torchaudio==0.10.1+cu111 -f https://download.pytorch.org/whl/torch_stable.html
+ python -c 'import torch; print("Torch version:", torch.__version__)'
+
+install_dep_pt1_12: &install_dep_pt1_12
+ - run:
+ name: Install Pytorch Dependencies
+ command: |
+ source activate fairseq
+ pip install --upgrade setuptools
+ pip install torch==1.12.1+cu116 torchaudio==0.12.1+cu116 -f https://download.pytorch.org/whl/torch_stable.html
+ python -c 'import torch; print("Torch version:", torch.__version__)'
+
+install_repo: &install_repo
+ - run:
+ name: Install Repository
+ command: |
+ source activate fairseq
+ python -m pip install fairscale
+ python -m pip install -e '.[dev,docs]'
+ python -c 'import torch; print("Torch version:", torch.__version__)'
+
+run_unittests: &run_unittests
+ - run:
+ name: Run Unit Tests
+ command: |
+ source activate fairseq
+ pytest tests/gpu/test_binaries_gpu.py
+
+check_nvidia_driver: &check_nvidia_driver
+ - run:
+ name: Check NVIDIA Driver
+ working_directory: ~/
+ command: |
+ pyenv versions
+ nvidia-smi
+
+create_conda_env: &create_conda_env
+ - run:
+ name: Install and Create Conda Environment
+ command: |
+ curl -o ~/miniconda.sh -O https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
+ chmod +x ~/miniconda.sh
+ bash ~/miniconda.sh -b -p $HOME/miniconda
+ rm ~/miniconda.sh
+ echo 'export PATH=$HOME/miniconda/bin:$PATH' >> $BASH_ENV
+ source $BASH_ENV
+ if [ ! -d ~/miniconda/envs/fairseq ]
+ then
+ conda create -y -n fairseq python=3.8
+ fi
+ source activate fairseq
+ python --version
+ pip install --upgrade pip
+# -------------------------------------------------------------------------------------
+# Jobs to run
+# -------------------------------------------------------------------------------------
+
+jobs:
+
+ gpu_tests_pt1_10:
+ <<: *gpu
+
+ working_directory: ~/fairseq-py
+
+ steps:
+ - checkout
+ - <<: *check_nvidia_driver
+ - <<: *create_conda_env
+ - restore_cache:
+ key: *cache_key
+ - <<: *install_dep_pt1_10
+ - save_cache:
+ paths:
+ - ~/miniconda/
+ key: *cache_key
+ - <<: *install_repo
+ - <<: *run_unittests
+
+ gpu_tests_pt1_12:
+ <<: *gpu
+
+ working_directory: ~/fairseq-py
+
+ steps:
+ - checkout
+ - <<: *check_nvidia_driver
+ - <<: *create_conda_env
+ - restore_cache:
+ key: *cache_key
+ - <<: *install_dep_pt1_12
+ - save_cache:
+ paths:
+ - ~/miniconda/
+ key: *cache_key
+ - <<: *install_repo
+ - <<: *run_unittests
+
+workflows:
+ version: 2
+ build:
+ jobs:
+ - gpu_tests_pt1_12
+ - gpu_tests_pt1_10
diff --git a/fairseq/.gitignore b/fairseq/.gitignore
new file mode 100644
index 0000000000000000000000000000000000000000..4be13638de637fe994e601f3de8e215109b30322
--- /dev/null
+++ b/fairseq/.gitignore
@@ -0,0 +1,141 @@
+# JetBrains PyCharm IDE
+.idea/
+
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# macOS dir files
+.DS_Store
+
+# Distribution / packaging
+.Python
+env/
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+
+# Checkpoints
+checkpoints
+
+# PyInstaller
+# Usually these files are written by a python script from a template
+# before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+.hypothesis/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# pyenv
+.python-version
+
+# celery beat schedule file
+celerybeat-schedule
+
+# SageMath parsed files
+*.sage.py
+
+# dotenv
+.env
+
+# virtualenv
+.venv
+venv/
+ENV/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+
+# Generated files
+/fairseq/temporal_convolution_tbc
+/fairseq/modules/*_layer/*_forward.cu
+/fairseq/modules/*_layer/*_backward.cu
+/fairseq/version.py
+
+# data
+data-bin/
+
+# reranking
+/examples/reranking/rerank_data
+
+# Cython-generated C++ source files
+/fairseq/data/data_utils_fast.cpp
+/fairseq/data/token_block_utils_fast.cpp
+
+# VSCODE
+.vscode/ftp-sync.json
+.vscode/settings.json
+
+# Experimental Folder
+experimental/*
+
+# Weights and Biases logs
+wandb/
+
+# Hydra artifacts
+nohup.out
+multirun
+outputs
diff --git a/fairseq/.gitmodules b/fairseq/.gitmodules
new file mode 100644
index 0000000000000000000000000000000000000000..07a55d45d4f0bed755dbfc1f440f214ed43d206a
--- /dev/null
+++ b/fairseq/.gitmodules
@@ -0,0 +1,4 @@
+[submodule "fairseq/model_parallel/megatron"]
+ path = fairseq/model_parallel/megatron
+ url = https://github.com/ngoyal2707/Megatron-LM
+ branch = fairseq
diff --git a/fairseq/.pre-commit-config.yaml b/fairseq/.pre-commit-config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..6b1d6aed8c7a4eb2d0c9b4eba7104ed9dc41b591
--- /dev/null
+++ b/fairseq/.pre-commit-config.yaml
@@ -0,0 +1,40 @@
+exclude: 'build|stubs'
+
+default_language_version:
+ python: python3
+
+repos:
+- repo: https://github.com/pre-commit/pre-commit-hooks
+ rev: v4.1.0
+ hooks:
+ - id: trailing-whitespace
+ - id: check-ast
+ - id: check-merge-conflict
+ - id: no-commit-to-branch
+ args: ['--branch=master']
+ - id: check-added-large-files
+ args: ['--maxkb=500']
+ - id: end-of-file-fixer
+
+- repo: https://github.com/ambv/black
+ rev: 22.3.0
+ hooks:
+ - id: black
+ language_version: python3.8
+
+- repo: https://gitlab.com/pycqa/flake8
+ rev: 3.9.2
+ hooks:
+ - id: flake8
+ args: [
+ # only error for syntax errors and undefined names
+ "--select=E9,F63,F7,F82",
+ ]
+
+- repo: https://github.com/pycqa/isort
+ rev: 5.10.1
+ hooks:
+ - id: isort
+ exclude: README.md
+ additional_dependencies: [toml]
+ args: ["--profile", "black"]
diff --git a/fairseq/CODE_OF_CONDUCT.md b/fairseq/CODE_OF_CONDUCT.md
new file mode 100644
index 0000000000000000000000000000000000000000..a0cbeaab7650bf08267fbdbc9bb54e845c88f392
--- /dev/null
+++ b/fairseq/CODE_OF_CONDUCT.md
@@ -0,0 +1,77 @@
+# Code of Conduct
+
+## Our Pledge
+
+In the interest of fostering an open and welcoming environment, we as
+contributors and maintainers pledge to make participation in our project and
+our community a harassment-free experience for everyone, regardless of age, body
+size, disability, ethnicity, sex characteristics, gender identity and expression,
+level of experience, education, socio-economic status, nationality, personal
+appearance, race, religion, or sexual identity and orientation.
+
+## Our Standards
+
+Examples of behavior that contributes to creating a positive environment
+include:
+
+* Using welcoming and inclusive language
+* Being respectful of differing viewpoints and experiences
+* Gracefully accepting constructive criticism
+* Focusing on what is best for the community
+* Showing empathy towards other community members
+
+Examples of unacceptable behavior by participants include:
+
+* The use of sexualized language or imagery and unwelcome sexual attention or
+ advances
+* Trolling, insulting/derogatory comments, and personal or political attacks
+* Public or private harassment
+* Publishing others' private information, such as a physical or electronic
+ address, without explicit permission
+* Other conduct which could reasonably be considered inappropriate in a
+ professional setting
+
+## Our Responsibilities
+
+Project maintainers are responsible for clarifying the standards of acceptable
+behavior and are expected to take appropriate and fair corrective action in
+response to any instances of unacceptable behavior.
+
+Project maintainers have the right and responsibility to remove, edit, or
+reject comments, commits, code, wiki edits, issues, and other contributions
+that are not aligned to this Code of Conduct, or to ban temporarily or
+permanently any contributor for other behaviors that they deem inappropriate,
+threatening, offensive, or harmful.
+
+## Scope
+
+This Code of Conduct applies within all project spaces, and it also applies when
+an individual is representing the project or its community in public spaces.
+Examples of representing a project or community include using an official
+project e-mail address, posting via an official social media account, or acting
+as an appointed representative at an online or offline event. Representation of
+a project may be further defined and clarified by project maintainers.
+
+## Enforcement
+
+Instances of abusive, harassing, or otherwise unacceptable behavior may be
+reported by contacting the project team at . All
+complaints will be reviewed and investigated and will result in a response that
+is deemed necessary and appropriate to the circumstances. The project team is
+obligated to maintain confidentiality with regard to the reporter of an incident.
+Further details of specific enforcement policies may be posted separately.
+
+Project maintainers who do not follow or enforce the Code of Conduct in good
+faith may face temporary or permanent repercussions as determined by other
+members of the project's leadership.
+
+## Attribution
+
+This Code of Conduct is adapted from the [Contributor Covenant][homepage], version 1.4,
+available at https://www.contributor-covenant.org/version/1/4/code-of-conduct.html
+
+[homepage]: https://www.contributor-covenant.org
+
+For answers to common questions about this code of conduct, see
+https://www.contributor-covenant.org/faq
+
diff --git a/fairseq/CONTRIBUTING.md b/fairseq/CONTRIBUTING.md
new file mode 100644
index 0000000000000000000000000000000000000000..60e90258877423bb458fafbc9d35781484dbe9c6
--- /dev/null
+++ b/fairseq/CONTRIBUTING.md
@@ -0,0 +1,82 @@
+# Contributing to Facebook AI Research Sequence-to-Sequence Toolkit (fairseq)
+We want to make contributing to this project as easy and transparent as
+possible.
+
+## Pull Requests
+We actively welcome your pull requests.
+
+1. Fork the repo and create your branch from `main`.
+2. If you've added code that should be tested, add tests.
+3. If you've changed APIs, update the documentation.
+4. Ensure the test suite passes.
+5. Make sure your code lints.
+6. If you haven't already, complete the Contributor License Agreement ("CLA").
+
+## Contributor License Agreement ("CLA")
+In order to accept your pull request, we need you to submit a CLA. You only need
+to do this once to work on any of Facebook's open source projects.
+
+Complete your CLA here:
+
+## Issues
+We use GitHub issues to track public bugs. Please ensure your description is
+clear and has sufficient instructions to be able to reproduce the issue.
+
+## License
+By contributing to Facebook AI Research Sequence-to-Sequence Toolkit (fairseq),
+you agree that your contributions will be licensed under the LICENSE file in
+the root directory of this source tree.
+
+## Pre-commit hooks
+In order to ensure your code lints, there are pre-commit hooks configured in the repository which you can install.
+After installation, they will automatically run each time you commit.
+An abbreviated guide is given below; for more information, refer to [the offical pre-commit documentation](https://pre-commit.com/).
+
+### Installation
+```
+pip install pre-commit
+pre-commit install
+```
+
+### Usage
+Just commit your changes:
+```
+git commit -m "My informative commit message"
+```
+
+If there was a failure, you will get feedback
+```
+[INFO] Initializing environment for https://github.com/PyCQA/flake8.
+[INFO] Installing environment for https://github.com/pre-commit/pre-commit-hooks.
+[INFO] Once installed this environment will be reused.
+[INFO] This may take a few minutes...
+[INFO] Installing environment for https://github.com/PyCQA/flake8.
+[INFO] Once installed this environment will be reused.
+[INFO] This may take a few minutes...
+Trim Trailing Whitespace.................................................Failed
+- hook id: trailing-whitespace
+- exit code: 1
+- files were modified by this hook
+Fixing examples/nllb/modeling/wmt15_benchmark/eval_langs2.sh
+Fix End of Files.........................................................Failed
+- hook id: end-of-file-fixer
+- exit code: 1
+- files were modified by this hook
+Fixing examples/few_shot/scripts/schedule_jobs_few_shot.py
+flake8...................................................................Passed
+```
+
+Certain hooks modify your files to comply.
+To include these modifications, you will need to add them (i.e. `git add ...`) and commit again.
+
+If all is well, you should see something like:
+```
+Trim Trailing Whitespace.................................................Passed
+Fix End of Files.........................................................Passed
+flake8...................................................................Passed
+[gshard-fix-ci 8698644e1] Fix lint, add pre-commit hooks
+ 10 files changed, 148 insertions(+), 110 deletions(-)
+ create mode 100644 .flake8
+ create mode 100644 .pre-commit-config.yaml
+ rename examples/nllb/modeling/wmt15_benchmark/{eval_langs2.py => eval_langs2.sh} (99%)
+ ```
diff --git a/fairseq/LICENSE b/fairseq/LICENSE
new file mode 100644
index 0000000000000000000000000000000000000000..b96dcb0480a0b0be0727976e5202a1e7b23edc3f
--- /dev/null
+++ b/fairseq/LICENSE
@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) Facebook, Inc. and its affiliates.
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
diff --git a/fairseq/MANIFEST.in b/fairseq/MANIFEST.in
new file mode 100644
index 0000000000000000000000000000000000000000..4f719da85c737a216ad4f373837c0e96c647ec36
--- /dev/null
+++ b/fairseq/MANIFEST.in
@@ -0,0 +1 @@
+include fairseq/version.txt
diff --git a/fairseq/README.md b/fairseq/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..1150c66cbeaa92aaa3a062b91369914a8a191306
--- /dev/null
+++ b/fairseq/README.md
@@ -0,0 +1,242 @@
+
+
+
+
+
+
+
+
+
+
+
+
+--------------------------------------------------------------------------------
+
+Fairseq(-py) is a sequence modeling toolkit that allows researchers and
+developers to train custom models for translation, summarization, language
+modeling and other text generation tasks.
+
+We provide reference implementations of various sequence modeling papers:
+
+List of implemented papers
+
+* **Convolutional Neural Networks (CNN)**
+ + [Language Modeling with Gated Convolutional Networks (Dauphin et al., 2017)](examples/language_model/conv_lm/README.md)
+ + [Convolutional Sequence to Sequence Learning (Gehring et al., 2017)](examples/conv_seq2seq/README.md)
+ + [Classical Structured Prediction Losses for Sequence to Sequence Learning (Edunov et al., 2018)](https://github.com/pytorch/fairseq/tree/classic_seqlevel)
+ + [Hierarchical Neural Story Generation (Fan et al., 2018)](examples/stories/README.md)
+ + [wav2vec: Unsupervised Pre-training for Speech Recognition (Schneider et al., 2019)](examples/wav2vec/README.md)
+* **LightConv and DynamicConv models**
+ + [Pay Less Attention with Lightweight and Dynamic Convolutions (Wu et al., 2019)](examples/pay_less_attention_paper/README.md)
+* **Long Short-Term Memory (LSTM) networks**
+ + Effective Approaches to Attention-based Neural Machine Translation (Luong et al., 2015)
+* **Transformer (self-attention) networks**
+ + Attention Is All You Need (Vaswani et al., 2017)
+ + [Scaling Neural Machine Translation (Ott et al., 2018)](examples/scaling_nmt/README.md)
+ + [Understanding Back-Translation at Scale (Edunov et al., 2018)](examples/backtranslation/README.md)
+ + [Adaptive Input Representations for Neural Language Modeling (Baevski and Auli, 2018)](examples/language_model/README.adaptive_inputs.md)
+ + [Lexically constrained decoding with dynamic beam allocation (Post & Vilar, 2018)](examples/constrained_decoding/README.md)
+ + [Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context (Dai et al., 2019)](examples/truncated_bptt/README.md)
+ + [Adaptive Attention Span in Transformers (Sukhbaatar et al., 2019)](examples/adaptive_span/README.md)
+ + [Mixture Models for Diverse Machine Translation: Tricks of the Trade (Shen et al., 2019)](examples/translation_moe/README.md)
+ + [RoBERTa: A Robustly Optimized BERT Pretraining Approach (Liu et al., 2019)](examples/roberta/README.md)
+ + [Facebook FAIR's WMT19 News Translation Task Submission (Ng et al., 2019)](examples/wmt19/README.md)
+ + [Jointly Learning to Align and Translate with Transformer Models (Garg et al., 2019)](examples/joint_alignment_translation/README.md )
+ + [Multilingual Denoising Pre-training for Neural Machine Translation (Liu et at., 2020)](examples/mbart/README.md)
+ + [Neural Machine Translation with Byte-Level Subwords (Wang et al., 2020)](examples/byte_level_bpe/README.md)
+ + [Unsupervised Quality Estimation for Neural Machine Translation (Fomicheva et al., 2020)](examples/unsupervised_quality_estimation/README.md)
+ + [wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations (Baevski et al., 2020)](examples/wav2vec/README.md)
+ + [Generating Medical Reports from Patient-Doctor Conversations Using Sequence-to-Sequence Models (Enarvi et al., 2020)](examples/pointer_generator/README.md)
+ + [Linformer: Self-Attention with Linear Complexity (Wang et al., 2020)](examples/linformer/README.md)
+ + [Cross-lingual Retrieval for Iterative Self-Supervised Training (Tran et al., 2020)](examples/criss/README.md)
+ + [Deep Transformers with Latent Depth (Li et al., 2020)](examples/latent_depth/README.md)
+ + [Unsupervised Cross-lingual Representation Learning for Speech Recognition (Conneau et al., 2020)](https://arxiv.org/abs/2006.13979)
+ + [Self-training and Pre-training are Complementary for Speech Recognition (Xu et al., 2020)](https://arxiv.org/abs/2010.11430)
+ + [Robust wav2vec 2.0: Analyzing Domain Shift in Self-Supervised Pre-Training (Hsu, et al., 2021)](https://arxiv.org/abs/2104.01027)
+ + [Unsupervised Speech Recognition (Baevski, et al., 2021)](https://arxiv.org/abs/2105.11084)
+ + [Simple and Effective Zero-shot Cross-lingual Phoneme Recognition (Xu et al., 2021)](https://arxiv.org/abs/2109.11680)
+ + [VideoCLIP: Contrastive Pre-training for Zero-shot Video-Text Understanding (Xu et. al., 2021)](https://arxiv.org/pdf/2109.14084.pdf)
+ + [VLM: Task-agnostic Video-Language Model Pre-training for Video Understanding (Xu et. al., 2021)](https://aclanthology.org/2021.findings-acl.370.pdf)
+ + [NormFormer: Improved Transformer Pretraining with Extra Normalization (Shleifer et. al, 2021)](examples/normformer/README.md)
+* **Non-autoregressive Transformers**
+ + Non-Autoregressive Neural Machine Translation (Gu et al., 2017)
+ + Deterministic Non-Autoregressive Neural Sequence Modeling by Iterative Refinement (Lee et al. 2018)
+ + Insertion Transformer: Flexible Sequence Generation via Insertion Operations (Stern et al. 2019)
+ + Mask-Predict: Parallel Decoding of Conditional Masked Language Models (Ghazvininejad et al., 2019)
+ + [Levenshtein Transformer (Gu et al., 2019)](examples/nonautoregressive_translation/README.md)
+* **Finetuning**
+ + [Better Fine-Tuning by Reducing Representational Collapse (Aghajanyan et al. 2020)](examples/rxf/README.md)
+
+
+
+### What's New:
+* May 2023 [Released models for Scaling Speech Technology to 1,000+ Languages (Pratap, et al., 2023)](examples/mms/README.md)
+* June 2022 [Released code for wav2vec-U 2.0 from Towards End-to-end Unsupervised Speech Recognition (Liu, et al., 2022)](examples/wav2vec/unsupervised/README.md)
+* May 2022 [Integration with xFormers](https://github.com/facebookresearch/xformers)
+* December 2021 [Released Direct speech-to-speech translation code](examples/speech_to_speech/README.md)
+* October 2021 [Released VideoCLIP and VLM models](examples/MMPT/README.md)
+* October 2021 [Released multilingual finetuned XLSR-53 model](examples/wav2vec/README.md)
+* September 2021 [`master` branch renamed to `main`](https://github.com/github/renaming).
+* July 2021 [Released DrNMT code](examples/discriminative_reranking_nmt/README.md)
+* July 2021 [Released Robust wav2vec 2.0 model](examples/wav2vec/README.md)
+* June 2021 [Released XLMR-XL and XLMR-XXL models](examples/xlmr/README.md)
+* May 2021 [Released Unsupervised Speech Recognition code](examples/wav2vec/unsupervised/README.md)
+* March 2021 [Added full parameter and optimizer state sharding + CPU offloading](examples/fully_sharded_data_parallel/README.md)
+* February 2021 [Added LASER training code](examples/laser/README.md)
+* December 2020: [Added Adaptive Attention Span code](examples/adaptive_span/README.md)
+* December 2020: [GottBERT model and code released](examples/gottbert/README.md)
+* November 2020: Adopted the [Hydra](https://github.com/facebookresearch/hydra) configuration framework
+ * [see documentation explaining how to use it for new and existing projects](docs/hydra_integration.md)
+* November 2020: [fairseq 0.10.0 released](https://github.com/pytorch/fairseq/releases/tag/v0.10.0)
+* October 2020: [Added R3F/R4F (Better Fine-Tuning) code](examples/rxf/README.md)
+* October 2020: [Deep Transformer with Latent Depth code released](examples/latent_depth/README.md)
+* October 2020: [Added CRISS models and code](examples/criss/README.md)
+
+Previous updates
+
+* September 2020: [Added Linformer code](examples/linformer/README.md)
+* September 2020: [Added pointer-generator networks](examples/pointer_generator/README.md)
+* August 2020: [Added lexically constrained decoding](examples/constrained_decoding/README.md)
+* August 2020: [wav2vec2 models and code released](examples/wav2vec/README.md)
+* July 2020: [Unsupervised Quality Estimation code released](examples/unsupervised_quality_estimation/README.md)
+* May 2020: [Follow fairseq on Twitter](https://twitter.com/fairseq)
+* April 2020: [Monotonic Multihead Attention code released](examples/simultaneous_translation/README.md)
+* April 2020: [Quant-Noise code released](examples/quant_noise/README.md)
+* April 2020: [Initial model parallel support and 11B parameters unidirectional LM released](examples/megatron_11b/README.md)
+* March 2020: [Byte-level BPE code released](examples/byte_level_bpe/README.md)
+* February 2020: [mBART model and code released](examples/mbart/README.md)
+* February 2020: [Added tutorial for back-translation](https://github.com/pytorch/fairseq/tree/main/examples/backtranslation#training-your-own-model-wmt18-english-german)
+* December 2019: [fairseq 0.9.0 released](https://github.com/pytorch/fairseq/releases/tag/v0.9.0)
+* November 2019: [VizSeq released (a visual analysis toolkit for evaluating fairseq models)](https://facebookresearch.github.io/vizseq/docs/getting_started/fairseq_example)
+* November 2019: [CamemBERT model and code released](examples/camembert/README.md)
+* November 2019: [BART model and code released](examples/bart/README.md)
+* November 2019: [XLM-R models and code released](examples/xlmr/README.md)
+* September 2019: [Nonautoregressive translation code released](examples/nonautoregressive_translation/README.md)
+* August 2019: [WMT'19 models released](examples/wmt19/README.md)
+* July 2019: fairseq relicensed under MIT license
+* July 2019: [RoBERTa models and code released](examples/roberta/README.md)
+* June 2019: [wav2vec models and code released](examples/wav2vec/README.md)
+
+
+
+### Features:
+
+* multi-GPU training on one machine or across multiple machines (data and model parallel)
+* fast generation on both CPU and GPU with multiple search algorithms implemented:
+ + beam search
+ + Diverse Beam Search ([Vijayakumar et al., 2016](https://arxiv.org/abs/1610.02424))
+ + sampling (unconstrained, top-k and top-p/nucleus)
+ + [lexically constrained decoding](examples/constrained_decoding/README.md) (Post & Vilar, 2018)
+* [gradient accumulation](https://fairseq.readthedocs.io/en/latest/getting_started.html#large-mini-batch-training-with-delayed-updates) enables training with large mini-batches even on a single GPU
+* [mixed precision training](https://fairseq.readthedocs.io/en/latest/getting_started.html#training-with-half-precision-floating-point-fp16) (trains faster with less GPU memory on [NVIDIA tensor cores](https://developer.nvidia.com/tensor-cores))
+* [extensible](https://fairseq.readthedocs.io/en/latest/overview.html): easily register new models, criterions, tasks, optimizers and learning rate schedulers
+* [flexible configuration](docs/hydra_integration.md) based on [Hydra](https://github.com/facebookresearch/hydra) allowing a combination of code, command-line and file based configuration
+* [full parameter and optimizer state sharding](examples/fully_sharded_data_parallel/README.md)
+* [offloading parameters to CPU](examples/fully_sharded_data_parallel/README.md)
+
+We also provide [pre-trained models for translation and language modeling](#pre-trained-models-and-examples)
+with a convenient `torch.hub` interface:
+
+``` python
+en2de = torch.hub.load('pytorch/fairseq', 'transformer.wmt19.en-de.single_model')
+en2de.translate('Hello world', beam=5)
+# 'Hallo Welt'
+```
+
+See the PyTorch Hub tutorials for [translation](https://pytorch.org/hub/pytorch_fairseq_translation/)
+and [RoBERTa](https://pytorch.org/hub/pytorch_fairseq_roberta/) for more examples.
+
+# Requirements and Installation
+
+* [PyTorch](http://pytorch.org/) version >= 1.10.0
+* Python version >= 3.8
+* For training new models, you'll also need an NVIDIA GPU and [NCCL](https://github.com/NVIDIA/nccl)
+* **To install fairseq** and develop locally:
+
+``` bash
+git clone https://github.com/pytorch/fairseq
+cd fairseq
+pip install --editable ./
+
+# on MacOS:
+# CFLAGS="-stdlib=libc++" pip install --editable ./
+
+# to install the latest stable release (0.10.x)
+# pip install fairseq
+```
+
+* **For faster training** install NVIDIA's [apex](https://github.com/NVIDIA/apex) library:
+
+``` bash
+git clone https://github.com/NVIDIA/apex
+cd apex
+pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" \
+ --global-option="--deprecated_fused_adam" --global-option="--xentropy" \
+ --global-option="--fast_multihead_attn" ./
+```
+
+* **For large datasets** install [PyArrow](https://arrow.apache.org/docs/python/install.html#using-pip): `pip install pyarrow`
+* If you use Docker make sure to increase the shared memory size either with `--ipc=host` or `--shm-size`
+ as command line options to `nvidia-docker run` .
+
+# Getting Started
+
+The [full documentation](https://fairseq.readthedocs.io/) contains instructions
+for getting started, training new models and extending fairseq with new model
+types and tasks.
+
+# Pre-trained models and examples
+
+We provide pre-trained models and pre-processed, binarized test sets for several tasks listed below,
+as well as example training and evaluation commands.
+
+* [Translation](examples/translation/README.md): convolutional and transformer models are available
+* [Language Modeling](examples/language_model/README.md): convolutional and transformer models are available
+
+We also have more detailed READMEs to reproduce results from specific papers:
+
+* [XLS-R: Self-supervised Cross-lingual Speech Representation Learning at Scale (Babu et al., 2021)](examples/wav2vec/xlsr/README.md)
+* [Cross-lingual Retrieval for Iterative Self-Supervised Training (Tran et al., 2020)](examples/criss/README.md)
+* [wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations (Baevski et al., 2020)](examples/wav2vec/README.md)
+* [Unsupervised Quality Estimation for Neural Machine Translation (Fomicheva et al., 2020)](examples/unsupervised_quality_estimation/README.md)
+* [Training with Quantization Noise for Extreme Model Compression ({Fan*, Stock*} et al., 2020)](examples/quant_noise/README.md)
+* [Neural Machine Translation with Byte-Level Subwords (Wang et al., 2020)](examples/byte_level_bpe/README.md)
+* [Multilingual Denoising Pre-training for Neural Machine Translation (Liu et at., 2020)](examples/mbart/README.md)
+* [Reducing Transformer Depth on Demand with Structured Dropout (Fan et al., 2019)](examples/layerdrop/README.md)
+* [Jointly Learning to Align and Translate with Transformer Models (Garg et al., 2019)](examples/joint_alignment_translation/README.md)
+* [Levenshtein Transformer (Gu et al., 2019)](examples/nonautoregressive_translation/README.md)
+* [Facebook FAIR's WMT19 News Translation Task Submission (Ng et al., 2019)](examples/wmt19/README.md)
+* [RoBERTa: A Robustly Optimized BERT Pretraining Approach (Liu et al., 2019)](examples/roberta/README.md)
+* [wav2vec: Unsupervised Pre-training for Speech Recognition (Schneider et al., 2019)](examples/wav2vec/README.md)
+* [Mixture Models for Diverse Machine Translation: Tricks of the Trade (Shen et al., 2019)](examples/translation_moe/README.md)
+* [Pay Less Attention with Lightweight and Dynamic Convolutions (Wu et al., 2019)](examples/pay_less_attention_paper/README.md)
+* [Understanding Back-Translation at Scale (Edunov et al., 2018)](examples/backtranslation/README.md)
+* [Classical Structured Prediction Losses for Sequence to Sequence Learning (Edunov et al., 2018)](https://github.com/pytorch/fairseq/tree/classic_seqlevel)
+* [Hierarchical Neural Story Generation (Fan et al., 2018)](examples/stories/README.md)
+* [Scaling Neural Machine Translation (Ott et al., 2018)](examples/scaling_nmt/README.md)
+* [Convolutional Sequence to Sequence Learning (Gehring et al., 2017)](examples/conv_seq2seq/README.md)
+* [Language Modeling with Gated Convolutional Networks (Dauphin et al., 2017)](examples/language_model/README.conv.md)
+
+# Join the fairseq community
+
+* Twitter: https://twitter.com/fairseq
+* Facebook page: https://www.facebook.com/groups/fairseq.users
+* Google group: https://groups.google.com/forum/#!forum/fairseq-users
+
+# License
+
+fairseq(-py) is MIT-licensed.
+The license applies to the pre-trained models as well.
+
+# Citation
+
+Please cite as:
+
+``` bibtex
+@inproceedings{ott2019fairseq,
+ title = {fairseq: A Fast, Extensible Toolkit for Sequence Modeling},
+ author = {Myle Ott and Sergey Edunov and Alexei Baevski and Angela Fan and Sam Gross and Nathan Ng and David Grangier and Michael Auli},
+ booktitle = {Proceedings of NAACL-HLT 2019: Demonstrations},
+ year = {2019},
+}
+```
diff --git a/fairseq/RELEASE.md b/fairseq/RELEASE.md
new file mode 100644
index 0000000000000000000000000000000000000000..79480a11c58a91178e9fa189a8a8e0ceea13efc6
--- /dev/null
+++ b/fairseq/RELEASE.md
@@ -0,0 +1,13 @@
+# Creating a New Release
+
+In order to create a new release:
+
+1. Navigate to the [Fairseq Workflows](https://github.com/facebookresearch/fairseq/actions) and find the one named _Fairseq Release_.
+
+2. Under _Run Workflow_ choose the branch `main` and for _Release Type_ enter either `major`, `minor`, or `patch`.
+
+3. A branch named `$new_version-release` will be created where the `version.txt` file is updated. Merge those changes into `main`.
+
+4. Make sure that a [new PYPI package](https://pypi.org/project/fairseq/) has been uploaded.
+
+5. Make sure that a [new github release](https://github.com/facebookresearch/fairseq/releases) has been created.
diff --git a/fairseq/docs/Makefile b/fairseq/docs/Makefile
new file mode 100644
index 0000000000000000000000000000000000000000..c2f5b1a89cfc9e02d1bb09027d9e1e520ba53d53
--- /dev/null
+++ b/fairseq/docs/Makefile
@@ -0,0 +1,20 @@
+# Minimal makefile for Sphinx documentation
+#
+
+# You can set these variables from the command line.
+SPHINXOPTS =
+SPHINXBUILD = python -msphinx
+SPHINXPROJ = fairseq
+SOURCEDIR = .
+BUILDDIR = _build
+
+# Put it first so that "make" without argument is like "make help".
+help:
+ @$(SPHINXBUILD) -M help "$(SOURCEDIR)" "$(BUILDDIR)" $(SPHINXOPTS) $(O)
+
+.PHONY: help Makefile
+
+# Catch-all target: route all unknown targets to Sphinx using the new
+# "make mode" option. $(O) is meant as a shortcut for $(SPHINXOPTS).
+%: Makefile
+ @$(SPHINXBUILD) -M $@ "$(SOURCEDIR)" "$(BUILDDIR)" $(SPHINXOPTS) $(O)
\ No newline at end of file
diff --git a/fairseq/docs/command_line_tools.rst b/fairseq/docs/command_line_tools.rst
new file mode 100644
index 0000000000000000000000000000000000000000..c16300ff5cd42d9a6c0070c2d9bec3a802eacfad
--- /dev/null
+++ b/fairseq/docs/command_line_tools.rst
@@ -0,0 +1,85 @@
+.. _Command-line Tools:
+
+Command-line Tools
+==================
+
+Fairseq provides several command-line tools for training and evaluating models:
+
+- :ref:`fairseq-preprocess`: Data pre-processing: build vocabularies and binarize training data
+- :ref:`fairseq-train`: Train a new model on one or multiple GPUs
+- :ref:`fairseq-generate`: Translate pre-processed data with a trained model
+- :ref:`fairseq-interactive`: Translate raw text with a trained model
+- :ref:`fairseq-score`: BLEU scoring of generated translations against reference translations
+- :ref:`fairseq-eval-lm`: Language model evaluation
+
+
+.. _fairseq-preprocess:
+
+fairseq-preprocess
+~~~~~~~~~~~~~~~~~~
+.. automodule:: fairseq_cli.preprocess
+
+ .. argparse::
+ :module: fairseq.options
+ :func: get_preprocessing_parser
+ :prog: fairseq-preprocess
+
+
+.. _fairseq-train:
+
+fairseq-train
+~~~~~~~~~~~~~
+.. automodule:: fairseq_cli.train
+
+ .. argparse::
+ :module: fairseq.options
+ :func: get_training_parser
+ :prog: fairseq-train
+
+
+.. _fairseq-generate:
+
+fairseq-generate
+~~~~~~~~~~~~~~~~
+.. automodule:: fairseq_cli.generate
+
+ .. argparse::
+ :module: fairseq.options
+ :func: get_generation_parser
+ :prog: fairseq-generate
+
+
+.. _fairseq-interactive:
+
+fairseq-interactive
+~~~~~~~~~~~~~~~~~~~
+.. automodule:: fairseq_cli.interactive
+
+ .. argparse::
+ :module: fairseq.options
+ :func: get_interactive_generation_parser
+ :prog: fairseq-interactive
+
+
+.. _fairseq-score:
+
+fairseq-score
+~~~~~~~~~~~~~
+.. automodule:: fairseq_cli.score
+
+ .. argparse::
+ :module: fairseq_cli.score
+ :func: get_parser
+ :prog: fairseq-score
+
+
+.. _fairseq-eval-lm:
+
+fairseq-eval-lm
+~~~~~~~~~~~~~~~
+.. automodule:: fairseq_cli.eval_lm
+
+ .. argparse::
+ :module: fairseq.options
+ :func: get_eval_lm_parser
+ :prog: fairseq-eval-lm
diff --git a/fairseq/docs/conf.py b/fairseq/docs/conf.py
new file mode 100644
index 0000000000000000000000000000000000000000..0bc049f8028bb3a4d28059f03e59cd368c5b95d4
--- /dev/null
+++ b/fairseq/docs/conf.py
@@ -0,0 +1,98 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+#
+# fairseq documentation build configuration file, created by
+# sphinx-quickstart on Fri Aug 17 21:45:30 2018.
+#
+# This file is execfile()d with the current directory set to its
+# containing dir.
+#
+# Note that not all possible configuration values are present in this
+# autogenerated file.
+#
+# All configuration values have a default; values that are commented out
+# serve to show the default.
+
+# If extensions (or modules to document with autodoc) are in another directory,
+# add these directories to sys.path here. If the directory is relative to the
+# documentation root, use os.path.abspath to make it absolute, like shown here.
+
+import os
+import sys
+from fairseq import __version__
+
+
+# source code directory, relative to this file, for sphinx-autobuild
+sys.path.insert(0, os.path.abspath(".."))
+
+source_suffix = [".rst"]
+
+# -- General configuration ------------------------------------------------
+
+# If your documentation needs a minimal Sphinx version, state it here.
+#
+# needs_sphinx = '1.0'
+
+# Add any Sphinx extension module names here, as strings. They can be
+# extensions coming with Sphinx (named 'sphinx.ext.*') or your custom
+# ones.
+extensions = [
+ "sphinx.ext.autodoc",
+ "sphinx.ext.intersphinx",
+ "sphinx.ext.viewcode",
+ "sphinx.ext.napoleon",
+ "sphinxarg.ext",
+]
+
+# Add any paths that contain templates here, relative to this directory.
+templates_path = ["_templates"]
+
+# The master toctree document.
+master_doc = "index"
+
+# General information about the project.
+project = "fairseq"
+copyright = "Facebook AI Research (FAIR)"
+author = "Facebook AI Research (FAIR)"
+
+github_doc_root = "https://github.com/pytorch/fairseq/tree/main/docs/"
+
+# The version info for the project you're documenting, acts as replacement for
+# |version| and |release|, also used in various other places throughout the
+# built documents.
+#
+# The short X.Y version.
+version = __version__
+# The full version, including alpha/beta/rc tags.
+release = __version__
+
+# The language for content autogenerated by Sphinx. Refer to documentation
+# for a list of supported languages.
+#
+# This is also used if you do content translation via gettext catalogs.
+# Usually you set "language" from the command line for these cases.
+language = None
+
+# List of patterns, relative to source directory, that match files and
+# directories to ignore when looking for source files.
+# This patterns also effect to html_static_path and html_extra_path
+exclude_patterns = ["_build", "Thumbs.db", ".DS_Store"]
+
+# The name of the Pygments (syntax highlighting) style to use.
+pygments_style = "sphinx"
+highlight_language = "python"
+
+# If true, `todo` and `todoList` produce output, else they produce nothing.
+todo_include_todos = False
+
+
+# -- Options for HTML output ----------------------------------------------
+
+html_theme = "classic"
+
+# Example configuration for intersphinx: refer to the Python standard library.
+intersphinx_mapping = {
+ "numpy": ("http://docs.scipy.org/doc/numpy/", None),
+ "python": ("https://docs.python.org/", None),
+ "torch": ("https://pytorch.org/docs/master/", None),
+}
diff --git a/fairseq/docs/criterions.rst b/fairseq/docs/criterions.rst
new file mode 100644
index 0000000000000000000000000000000000000000..d6b8ca6b671a32d0da4aca7b18626e0df58a7258
--- /dev/null
+++ b/fairseq/docs/criterions.rst
@@ -0,0 +1,31 @@
+.. role:: hidden
+ :class: hidden-section
+
+.. _Criterions:
+
+Criterions
+==========
+
+Criterions compute the loss function given the model and batch, roughly::
+
+ loss = criterion(model, batch)
+
+.. automodule:: fairseq.criterions
+ :members:
+
+.. autoclass:: fairseq.criterions.FairseqCriterion
+ :members:
+ :undoc-members:
+
+.. autoclass:: fairseq.criterions.adaptive_loss.AdaptiveLoss
+ :members:
+ :undoc-members:
+.. autoclass:: fairseq.criterions.composite_loss.CompositeLoss
+ :members:
+ :undoc-members:
+.. autoclass:: fairseq.criterions.cross_entropy.CrossEntropyCriterion
+ :members:
+ :undoc-members:
+.. autoclass:: fairseq.criterions.label_smoothed_cross_entropy.LabelSmoothedCrossEntropyCriterion
+ :members:
+ :undoc-members:
diff --git a/fairseq/docs/data.rst b/fairseq/docs/data.rst
new file mode 100644
index 0000000000000000000000000000000000000000..6a390cb336ab3c5fb28edec7448abc35a8e22bbb
--- /dev/null
+++ b/fairseq/docs/data.rst
@@ -0,0 +1,58 @@
+.. role:: hidden
+ :class: hidden-section
+
+.. module:: fairseq.data
+
+Data Loading and Utilities
+==========================
+
+.. _datasets:
+
+Datasets
+--------
+
+**Datasets** define the data format and provide helpers for creating
+mini-batches.
+
+.. autoclass:: fairseq.data.FairseqDataset
+ :members:
+.. autoclass:: fairseq.data.LanguagePairDataset
+ :members:
+.. autoclass:: fairseq.data.MonolingualDataset
+ :members:
+
+**Helper Datasets**
+
+These datasets wrap other :class:`fairseq.data.FairseqDataset` instances and
+provide additional functionality:
+
+.. autoclass:: fairseq.data.BacktranslationDataset
+ :members:
+.. autoclass:: fairseq.data.ConcatDataset
+ :members:
+.. autoclass:: fairseq.data.ResamplingDataset
+ :members:
+.. autoclass:: fairseq.data.RoundRobinZipDatasets
+ :members:
+.. autoclass:: fairseq.data.TransformEosDataset
+ :members:
+
+
+Dictionary
+----------
+
+.. autoclass:: fairseq.data.Dictionary
+ :members:
+
+
+Iterators
+---------
+
+.. autoclass:: fairseq.data.CountingIterator
+ :members:
+.. autoclass:: fairseq.data.EpochBatchIterator
+ :members:
+.. autoclass:: fairseq.data.GroupedIterator
+ :members:
+.. autoclass:: fairseq.data.ShardedIterator
+ :members:
diff --git a/fairseq/docs/docutils.conf b/fairseq/docs/docutils.conf
new file mode 100644
index 0000000000000000000000000000000000000000..526acffd32d16217160aee917db2b120354f20f0
--- /dev/null
+++ b/fairseq/docs/docutils.conf
@@ -0,0 +1,2 @@
+[writers]
+option-limit=0
diff --git a/fairseq/docs/fairseq.gif b/fairseq/docs/fairseq.gif
new file mode 100644
index 0000000000000000000000000000000000000000..8701763fa799e18f99ecf255f9cadb058724ab0a
--- /dev/null
+++ b/fairseq/docs/fairseq.gif
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b7551e9682c816fca1fa00458f3b657177c8d90d2e87db31e42197cb3ae80fca
+size 2664833
diff --git a/fairseq/docs/fairseq_logo.png b/fairseq/docs/fairseq_logo.png
new file mode 100644
index 0000000000000000000000000000000000000000..1830151c109c00329e4b9327fb5410999bcfdb6e
--- /dev/null
+++ b/fairseq/docs/fairseq_logo.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:850abbe321cffb628d19cb5c20585ae87659ad265ec8635c57fd91ffcd81d742
+size 73036
diff --git a/fairseq/docs/getting_started.rst b/fairseq/docs/getting_started.rst
new file mode 100644
index 0000000000000000000000000000000000000000..745ad7763cee67a8dec25bdd7ba7b79cbe0b7754
--- /dev/null
+++ b/fairseq/docs/getting_started.rst
@@ -0,0 +1,216 @@
+Evaluating Pre-trained Models
+=============================
+
+First, download a pre-trained model along with its vocabularies:
+
+.. code-block:: console
+
+ > curl https://dl.fbaipublicfiles.com/fairseq/models/wmt14.v2.en-fr.fconv-py.tar.bz2 | tar xvjf -
+
+This model uses a `Byte Pair Encoding (BPE)
+vocabulary `__, so we'll have to apply
+the encoding to the source text before it can be translated. This can be
+done with the
+`apply\_bpe.py `__
+script using the ``wmt14.en-fr.fconv-cuda/bpecodes`` file. ``@@`` is
+used as a continuation marker and the original text can be easily
+recovered with e.g. ``sed s/@@ //g`` or by passing the ``--remove-bpe``
+flag to :ref:`fairseq-generate`. Prior to BPE, input text needs to be tokenized
+using ``tokenizer.perl`` from
+`mosesdecoder `__.
+
+Let's use :ref:`fairseq-interactive` to generate translations interactively.
+Here, we use a beam size of 5 and preprocess the input with the Moses
+tokenizer and the given Byte-Pair Encoding vocabulary. It will automatically
+remove the BPE continuation markers and detokenize the output.
+
+.. code-block:: console
+
+ > MODEL_DIR=wmt14.en-fr.fconv-py
+ > fairseq-interactive \
+ --path $MODEL_DIR/model.pt $MODEL_DIR \
+ --beam 5 --source-lang en --target-lang fr \
+ --tokenizer moses \
+ --bpe subword_nmt --bpe-codes $MODEL_DIR/bpecodes
+ | loading model(s) from wmt14.en-fr.fconv-py/model.pt
+ | [en] dictionary: 44206 types
+ | [fr] dictionary: 44463 types
+ | Type the input sentence and press return:
+ Why is it rare to discover new marine mammal species?
+ S-0 Why is it rare to discover new marine mam@@ mal species ?
+ H-0 -0.0643349438905716 Pourquoi est-il rare de découvrir de nouvelles espèces de mammifères marins?
+ P-0 -0.0763 -0.1849 -0.0956 -0.0946 -0.0735 -0.1150 -0.1301 -0.0042 -0.0321 -0.0171 -0.0052 -0.0062 -0.0015
+
+This generation script produces three types of outputs: a line prefixed
+with *O* is a copy of the original source sentence; *H* is the
+hypothesis along with an average log-likelihood; and *P* is the
+positional score per token position, including the
+end-of-sentence marker which is omitted from the text.
+
+Other types of output lines you might see are *D*, the detokenized hypothesis,
+*T*, the reference target, *A*, alignment info, *E* the history of generation steps.
+
+See the `README `__ for a
+full list of pre-trained models available.
+
+Training a New Model
+====================
+
+The following tutorial is for machine translation. For an example of how
+to use Fairseq for other tasks, such as :ref:`language modeling`, please see the
+``examples/`` directory.
+
+Data Pre-processing
+-------------------
+
+Fairseq contains example pre-processing scripts for several translation
+datasets: IWSLT 2014 (German-English), WMT 2014 (English-French) and WMT
+2014 (English-German). To pre-process and binarize the IWSLT dataset:
+
+.. code-block:: console
+
+ > cd examples/translation/
+ > bash prepare-iwslt14.sh
+ > cd ../..
+ > TEXT=examples/translation/iwslt14.tokenized.de-en
+ > fairseq-preprocess --source-lang de --target-lang en \
+ --trainpref $TEXT/train --validpref $TEXT/valid --testpref $TEXT/test \
+ --destdir data-bin/iwslt14.tokenized.de-en
+
+This will write binarized data that can be used for model training to
+``data-bin/iwslt14.tokenized.de-en``.
+
+Training
+--------
+
+Use :ref:`fairseq-train` to train a new model. Here a few example settings that work
+well for the IWSLT 2014 dataset:
+
+.. code-block:: console
+
+ > mkdir -p checkpoints/fconv
+ > CUDA_VISIBLE_DEVICES=0 fairseq-train data-bin/iwslt14.tokenized.de-en \
+ --optimizer nag --lr 0.25 --clip-norm 0.1 --dropout 0.2 --max-tokens 4000 \
+ --arch fconv_iwslt_de_en --save-dir checkpoints/fconv
+
+By default, :ref:`fairseq-train` will use all available GPUs on your machine. Use the
+``CUDA_VISIBLE_DEVICES`` environment variable to select specific GPUs and/or to
+change the number of GPU devices that will be used.
+
+Also note that the batch size is specified in terms of the maximum
+number of tokens per batch (``--max-tokens``). You may need to use a
+smaller value depending on the available GPU memory on your system.
+
+Generation
+----------
+
+Once your model is trained, you can generate translations using
+:ref:`fairseq-generate` **(for binarized data)** or
+:ref:`fairseq-interactive` **(for raw text)**:
+
+.. code-block:: console
+
+ > fairseq-generate data-bin/iwslt14.tokenized.de-en \
+ --path checkpoints/fconv/checkpoint_best.pt \
+ --batch-size 128 --beam 5
+ | [de] dictionary: 35475 types
+ | [en] dictionary: 24739 types
+ | data-bin/iwslt14.tokenized.de-en test 6750 examples
+ | model fconv
+ | loaded checkpoint trainings/fconv/checkpoint_best.pt
+ S-721 danke .
+ T-721 thank you .
+ ...
+
+To generate translations with only a CPU, use the ``--cpu`` flag. BPE
+continuation markers can be removed with the ``--remove-bpe`` flag.
+
+Advanced Training Options
+=========================
+
+Large mini-batch training with delayed updates
+----------------------------------------------
+
+The ``--update-freq`` option can be used to accumulate gradients from
+multiple mini-batches and delay updating, creating a larger effective
+batch size. Delayed updates can also improve training speed by reducing
+inter-GPU communication costs and by saving idle time caused by variance
+in workload across GPUs. See `Ott et al.
+(2018) `__ for more details.
+
+To train on a single GPU with an effective batch size that is equivalent
+to training on 8 GPUs:
+
+.. code-block:: console
+
+ > CUDA_VISIBLE_DEVICES=0 fairseq-train --update-freq 8 (...)
+
+Training with half precision floating point (FP16)
+--------------------------------------------------
+
+.. note::
+
+ FP16 training requires a Volta GPU and CUDA 9.1 or greater
+
+Recent GPUs enable efficient half precision floating point computation,
+e.g., using `Nvidia Tensor Cores
+`__.
+Fairseq supports FP16 training with the ``--fp16`` flag:
+
+.. code-block:: console
+
+ > fairseq-train --fp16 (...)
+
+Distributed training
+--------------------
+
+Distributed training in fairseq is implemented on top of ``torch.distributed``.
+The easiest way to launch jobs is with the `torch.distributed.launch
+`__ tool.
+
+For example, to train a large English-German Transformer model on 2 nodes each
+with 8 GPUs (in total 16 GPUs), run the following command on each node,
+replacing ``node_rank=0`` with ``node_rank=1`` on the second node and making
+sure to update ``--master_addr`` to the IP address of the first node:
+
+.. code-block:: console
+
+ > python -m torch.distributed.launch --nproc_per_node=8 \
+ --nnodes=2 --node_rank=0 --master_addr="192.168.1.1" \
+ --master_port=12345 \
+ $(which fairseq-train) data-bin/wmt16_en_de_bpe32k \
+ --arch transformer_vaswani_wmt_en_de_big --share-all-embeddings \
+ --optimizer adam --adam-betas '(0.9, 0.98)' --clip-norm 0.0 \
+ --lr-scheduler inverse_sqrt --warmup-init-lr 1e-07 --warmup-updates 4000 \
+ --lr 0.0005 \
+ --dropout 0.3 --weight-decay 0.0 --criterion label_smoothed_cross_entropy --label-smoothing 0.1 \
+ --max-tokens 3584 \
+ --max-epoch 70 \
+ --fp16
+
+On SLURM clusters, fairseq will automatically detect the number of nodes and
+GPUs, but a port number must be provided:
+
+.. code-block:: console
+
+ > salloc --gpus=16 --nodes 2 (...)
+ > srun fairseq-train --distributed-port 12345 (...).
+
+Sharding very large datasets
+----------------------------
+
+It can be challenging to train over very large datasets, particularly if your
+machine does not have much system RAM. Most tasks in fairseq support training
+over "sharded" datasets, in which the original dataset has been preprocessed
+into non-overlapping chunks (or "shards").
+
+For example, instead of preprocessing all your data into a single "data-bin"
+directory, you can split the data and create "data-bin1", "data-bin2", etc.
+Then you can adapt your training command like so:
+
+.. code-block:: console
+
+ > fairseq-train data-bin1:data-bin2:data-bin3 (...)
+
+Training will now iterate over each shard, one by one, with each shard
+corresponding to an "epoch", thus reducing system memory usage.
diff --git a/fairseq/docs/hydra_integration.md b/fairseq/docs/hydra_integration.md
new file mode 100644
index 0000000000000000000000000000000000000000..6a15298382a6a16dfc4c5a4a812ea1cd0477ed52
--- /dev/null
+++ b/fairseq/docs/hydra_integration.md
@@ -0,0 +1,284 @@
+## Hydra
+
+[Hydra](https://github.com/facebookresearch/hydra) is an open-source Python
+framework that simplifies the development of research and other complex
+applications. The key feature is the ability to dynamically create a
+hierarchical configuration by composition and override it through config files
+and the command line. The name Hydra comes from its ability to run multiple
+similar jobs - much like a Hydra with multiple heads.
+
+## Motivation
+
+Until recently, all components in fairseq were configured through a shared
+`args` namespace that was created at application startup. Components declared
+their own `add_args` method to update the argparse parser, hoping that the names
+would not clash with arguments from other components. While this model works for
+smaller applications, as fairseq grew and became integrated into other
+applications, this became problematic. In order to determine how to configure
+each component, one needed to a) examine what args were added by this component,
+and b) read the code to figure out what shared arguments it is using that were
+added in other places. Reproducing models involved sharing commands that often
+contained dozens of command line switches.
+
+The model described above is still supported by fairseq for backward
+compatibility, but will be deprecated some time in the future.
+
+New components in fairseq should now create a dataclass that encapsulates all
+parameters required to configure this component. The dataclass is registered
+along with the component, and fairseq takes care of constructing and providing
+this configuration object to the component's constructor. Note that sharing
+parameters can optionally still work, but one has to explicitly point to the
+"source of truth" (see inheritance example below). These changes make components
+in fairseq more independent and re-usable by other applications: all that is
+needed to create a component is to initialize its dataclass and overwrite some
+of the defaults.
+
+While configuring fairseq through command line (using either the legacy argparse
+based or the new Hydra based entry points) is still fully supported, you can now
+take advantage of configuring fairseq completely or piece-by-piece through
+hierarchical YAML configuration files. These files can also be shipped as
+examples that others can use to run an identically configured job.
+
+Additionally, Hydra has a rich and growing [library of
+plugins](https://github.com/facebookresearch/hydra/tree/master/plugins) that
+provide functionality such as hyperparameter sweeping (including using bayesian
+optimization through the [Ax](https://github.com/facebook/Ax) library), job
+launching across various platforms, and more.
+
+## Creating or migrating components
+
+In general, each new (or updated) component should provide a companion
+[dataclass](https://www.python.org/dev/peps/pep-0557/). These dataclass are
+typically located in the same file as the component and are passed as arguments
+to the `register_*()` functions. Top-level configs that should be present in
+every fairseq application are placed in the
+[global](fairseq/dataclass/configs.py) config file and added to the
+`FairseqConfig` object.
+
+Each dataclass is a plain-old-data object, similar to a `NamedTuple`. These
+classes are decorated with a `@dataclass` decorator, and typically inherit from
+`FairseqDataclass` (which adds some functionality for backward compatibility).
+Each field must have a type, and generally has metadata (such as a help string)
+and a default value. Only primitive types or other config objects are allowed as
+data types for each field.
+
+#### Example:
+
+```python
+from dataclasses import dataclass, field
+from fairseq.dataclass import FairseqDataclass
+
+@dataclass
+class InteractiveConfig(FairseqDataclass):
+ buffer_size: int = field(
+ default=0,
+ metadata={
+ "help": "read this many sentences into a buffer before processing them"
+ },
+ )
+ input: str = field(
+ default="-",
+ metadata={"help": "file to read from; use - for stdin"},
+ )
+```
+
+### Inherting values
+
+Some components require sharing a value. For example, a learning rate scheduler
+and an optimizer may both need to know the initial learning rate value. One can
+declare a field that, by default, will inherit its value from another config
+node in the same hierarchy:
+
+```python
+@dataclass
+FairseqAdamConfig(FairseqDataclass):
+ ...
+ lr: List[float] = II("optimization.lr")
+ ...
+```
+
+`II("optimization.lr")` is syntactic sugar for `"${optimization.lr}"`, which is
+the value one can use in a YAML config file or through command line to achieve
+the same effect. Note that this assumes that there is an "optimization" config
+object in the root config and it has a field called "lr".
+
+### Tasks and Models
+
+Creating Tasks and Models works same as before, except that legacy
+implementations now inherit from `LegacyFairseq*` base classes, while new
+components inherit from `FairseqTask` and `FairseqModel` and provide a dataclass
+to the `register_*()` functions.
+
+#### Task example:
+
+```python
+@dataclass
+class LanguageModelingConfig(FairseqDataclass):
+ data: Optional[str] = field(
+ default=None, metadata={"help": "path to data directory"}
+ )
+ ...
+
+@register_task("language_modeling", dataclass=LanguageModelingConfig)
+class LanguageModelingTask(FairseqTask):
+ ...
+ @classmethod
+ def setup_task(cls, cfg: LanguageModelingConfig):
+ ...
+```
+
+#### Model example:
+
+```python
+@dataclass
+class TransformerLanguageModelConfig(FairseqDataclass):
+ activation_fn: ChoiceEnum(utils.get_available_activation_fns()) = field(
+ default="relu", metadata={"help": "activation function to use"}
+ )
+ dropout: float = field(default=0.1, metadata={"help": "dropout probability"})
+ ...
+
+@register_model("transformer_lm", dataclass=TransformerLanguageModelConfig)
+class TransformerLanguageModel(FairseqLanguageModel):
+ ...
+ @classmethod
+ def build_model(cls, cfg: TransformerLanguageModelConfig, task: FairseqTask):
+ ...
+```
+
+### Other components
+
+Other components work as before, but they now take their configuration dataclass
+as the only constructor argument:
+
+```python
+@dataclass
+class MosesTokenizerConfig(FairseqDataclass):
+ source_lang: str = field(default="en", metadata={"help": "source language"})
+ ...
+
+@register_tokenizer("moses", dataclass=MosesTokenizerConfig)
+class MosesTokenizer(object):
+ def __init__(self, cfg: MosesTokenizerConfig):
+ ...
+```
+
+Note that if you are adding a new registry for a new set of components, you need
+to add it to the `FairseqConfig` object in `fairseq/dataclass/configs.py`:
+
+```python
+@dataclass
+class FairseqConfig(object):
+ ...
+ my_new_registry: Any = None
+```
+
+## Training with `fairseq-hydra-train`
+
+To fully take advantage of configuration flexibility offered by Hydra, you may
+want to train new models using the `fairseq-hydra-train` entry point. Legacy CLI
+tools such as `fairseq-train` will remain supported for the foreseeable future
+but will be deprecated eventually.
+
+On startup, Hydra will create a configuration object that contains a hierarchy
+of all the necessary dataclasses populated with their default values in the
+code. The default values are overwritten by values found in YAML files in
+`fairseq/config` directory (which currently sets minimal defaults) and then
+further overwritten by values provided through command line arguments.
+
+Some of the most common use cases are shown below:
+
+### 1. Override default values through command line:
+
+```shell script
+$ fairseq-hydra-train \
+ distributed_training.distributed_world_size=1 \
+ dataset.batch_size=2 \
+ task.data=data-bin \
+ model=transformer_lm/transformer_lm_gpt \
+ task=language_modeling \
+ optimization.max_update=5000
+```
+
+Note that along with explicitly providing values for parameters such as
+`dataset.batch_size`, this also tells Hydra to overlay configuration found in
+`fairseq/config/model/transformer_lm/transformer_lm_gpt.yaml` over the default
+values in the dataclass. If you want to train a model without specifying a
+particular architecture you can simply specify `model=transformer_lm`. This only
+works for migrated tasks and models.
+
+### 2. Replace bundled configs with an external config:
+
+```shell script
+$ fairseq-hydra-train \
+ --config-dir /path/to/external/configs \
+ --config-name wiki103
+```
+
+where `/path/to/external/configs/wiki103.yaml` contains:
+
+```yaml
+# @package _group_
+
+model:
+ _name: transformer_lm
+distributed_training:
+ distributed_world_size: 1
+dataset:
+ batch_size: 2
+task:
+ _name: language_modeling
+ data: /path/to/data
+ add_bos_token: false
+ max_target_positions: 1024
+optimization:
+ max_update: 50000
+ lr: [ 0.25 ]
+criterion: cross_entropy
+optimizer: adam
+lr_scheduler:
+ _name: cosine
+```
+
+Note that here bundled configs from `fairseq/config` directory are not used,
+however the defaults from each dataclass will still be used (unless overwritten
+by your external config).
+
+Additionally you can choose to break up your configs by creating a directory
+structure in the same location as your main config file, with the names of the
+top-level fields (such as "model", "dataset", etc), and placing config files
+with meaningful names that would populate that specific section of your
+top-level config file (for example, you might have
+`model/small_transformer_lm.yaml`, `model/big_transformer_lm.yaml`, etc). You
+can then specify the correct configuration via command line, defaults in the
+main config, or even launch all of them as a sweep (see Hydra documentation on
+how to do this).
+
+### 3. Add an external config directory to Hydra search path:
+
+This allows combining default configuration (including using any bundled config
+files), while specifying your own config files for some parts of the
+configuration.
+
+```shell script
+$ fairseq-hydra-train \
+ distributed_training.distributed_world_size=1 \
+ dataset.batch_size=2 \
+ task.data=/path/to/data/ \
+ model=transformer_lm/2_layers \
+ task=language_modeling \
+ optimization.max_update=5000 \
+ --config-dir /path/to/external/configs
+```
+
+where `/path/to/external/configs` has the following structure:
+```
+.
++-- model
+| +-- transformer_lm
+| | +-- 2_layers.yaml
+```
+
+and `2_layers.yaml` contains a copy of `transformer_lm_gpt.yaml` but with
+`decoder_layers` set to 2. You can add other configs to configure other
+components as well.
diff --git a/fairseq/docs/index.rst b/fairseq/docs/index.rst
new file mode 100644
index 0000000000000000000000000000000000000000..591db86cdf49e6f0a7a6686df2150f11418e90d0
--- /dev/null
+++ b/fairseq/docs/index.rst
@@ -0,0 +1,49 @@
+.. fairseq documentation master file, created by
+ sphinx-quickstart on Fri Aug 17 21:45:30 2018.
+ You can adapt this file completely to your liking, but it should at least
+ contain the root `toctree` directive.
+
+:github_url: https://github.com/pytorch/fairseq
+
+
+fairseq documentation
+=====================
+
+Fairseq is a sequence modeling toolkit written in `PyTorch
+`_ that allows researchers and developers to
+train custom models for translation, summarization, language modeling and other
+text generation tasks.
+
+.. toctree::
+ :maxdepth: 1
+ :caption: Getting Started
+
+ getting_started
+ command_line_tools
+
+.. toctree::
+ :maxdepth: 1
+ :caption: Extending Fairseq
+
+ overview
+ tutorial_simple_lstm
+ tutorial_classifying_names
+
+.. toctree::
+ :maxdepth: 2
+ :caption: Library Reference
+
+ tasks
+ models
+ criterions
+ optim
+ lr_scheduler
+ data
+ modules
+
+
+Indices and tables
+==================
+
+* :ref:`genindex`
+* :ref:`search`
diff --git a/fairseq/docs/lr_scheduler.rst b/fairseq/docs/lr_scheduler.rst
new file mode 100644
index 0000000000000000000000000000000000000000..bbc09dc22e6a7ac05137954e0b9c80ca030f62f4
--- /dev/null
+++ b/fairseq/docs/lr_scheduler.rst
@@ -0,0 +1,34 @@
+.. role:: hidden
+ :class: hidden-section
+
+.. _Learning Rate Schedulers:
+
+Learning Rate Schedulers
+========================
+
+Learning Rate Schedulers update the learning rate over the course of training.
+Learning rates can be updated after each update via :func:`step_update` or at
+epoch boundaries via :func:`step`.
+
+.. automodule:: fairseq.optim.lr_scheduler
+ :members:
+
+.. autoclass:: fairseq.optim.lr_scheduler.FairseqLRScheduler
+ :members:
+ :undoc-members:
+
+.. autoclass:: fairseq.optim.lr_scheduler.cosine_lr_scheduler.CosineSchedule
+ :members:
+ :undoc-members:
+.. autoclass:: fairseq.optim.lr_scheduler.fixed_schedule.FixedSchedule
+ :members:
+ :undoc-members:
+.. autoclass:: fairseq.optim.lr_scheduler.inverse_square_root_schedule.InverseSquareRootSchedule
+ :members:
+ :undoc-members:
+.. autoclass:: fairseq.optim.lr_scheduler.reduce_lr_on_plateau.ReduceLROnPlateau
+ :members:
+ :undoc-members:
+.. autoclass:: fairseq.optim.lr_scheduler.triangular_lr_scheduler.TriangularSchedule
+ :members:
+ :undoc-members:
diff --git a/fairseq/docs/make.bat b/fairseq/docs/make.bat
new file mode 100644
index 0000000000000000000000000000000000000000..baa9d02a79266ed17e0841f08a83931d46583393
--- /dev/null
+++ b/fairseq/docs/make.bat
@@ -0,0 +1,36 @@
+@ECHO OFF
+
+pushd %~dp0
+
+REM Command file for Sphinx documentation
+
+if "%SPHINXBUILD%" == "" (
+ set SPHINXBUILD=python -msphinx
+)
+set SOURCEDIR=.
+set BUILDDIR=_build
+set SPHINXPROJ=fairseq
+
+if "%1" == "" goto help
+
+%SPHINXBUILD% >NUL 2>NUL
+if errorlevel 9009 (
+ echo.
+ echo.The Sphinx module was not found. Make sure you have Sphinx installed,
+ echo.then set the SPHINXBUILD environment variable to point to the full
+ echo.path of the 'sphinx-build' executable. Alternatively you may add the
+ echo.Sphinx directory to PATH.
+ echo.
+ echo.If you don't have Sphinx installed, grab it from
+ echo.http://sphinx-doc.org/
+ exit /b 1
+)
+
+%SPHINXBUILD% -M %1 %SOURCEDIR% %BUILDDIR% %SPHINXOPTS%
+goto end
+
+:help
+%SPHINXBUILD% -M help %SOURCEDIR% %BUILDDIR% %SPHINXOPTS%
+
+:end
+popd
diff --git a/fairseq/docs/models.rst b/fairseq/docs/models.rst
new file mode 100644
index 0000000000000000000000000000000000000000..054622d587c3b7f01f17f442919140755acd8f9e
--- /dev/null
+++ b/fairseq/docs/models.rst
@@ -0,0 +1,104 @@
+.. role:: hidden
+ :class: hidden-section
+
+.. module:: fairseq.models
+
+.. _Models:
+
+Models
+======
+
+A Model defines the neural network's ``forward()`` method and encapsulates all
+of the learnable parameters in the network. Each model also provides a set of
+named *architectures* that define the precise network configuration (e.g.,
+embedding dimension, number of layers, etc.).
+
+Both the model type and architecture are selected via the ``--arch``
+command-line argument. Once selected, a model may expose additional command-line
+arguments for further configuration.
+
+.. note::
+
+ All fairseq Models extend :class:`BaseFairseqModel`, which in turn extends
+ :class:`torch.nn.Module`. Thus any fairseq Model can be used as a
+ stand-alone Module in other PyTorch code.
+
+
+Convolutional Neural Networks (CNN)
+-----------------------------------
+
+.. module:: fairseq.models.fconv
+.. autoclass:: fairseq.models.fconv.FConvModel
+ :members:
+.. autoclass:: fairseq.models.fconv.FConvEncoder
+ :members:
+ :undoc-members:
+.. autoclass:: fairseq.models.fconv.FConvDecoder
+ :members:
+
+
+Long Short-Term Memory (LSTM) networks
+--------------------------------------
+
+.. module:: fairseq.models.lstm
+.. autoclass:: fairseq.models.lstm.LSTMModel
+ :members:
+.. autoclass:: fairseq.models.lstm.LSTMEncoder
+ :members:
+.. autoclass:: fairseq.models.lstm.LSTMDecoder
+ :members:
+
+
+Transformer (self-attention) networks
+-------------------------------------
+
+.. module:: fairseq.models.transformer
+.. autoclass:: fairseq.models.transformer.TransformerModel
+ :members:
+.. autoclass:: fairseq.models.transformer.TransformerEncoder
+ :members:
+.. autoclass:: fairseq.models.transformer.TransformerEncoderLayer
+ :members:
+.. autoclass:: fairseq.models.transformer.TransformerDecoder
+ :members:
+.. autoclass:: fairseq.models.transformer.TransformerDecoderLayer
+ :members:
+
+
+Adding new models
+-----------------
+
+.. currentmodule:: fairseq.models
+.. autofunction:: fairseq.models.register_model
+.. autofunction:: fairseq.models.register_model_architecture
+.. autoclass:: fairseq.models.BaseFairseqModel
+ :members:
+ :undoc-members:
+.. autoclass:: fairseq.models.FairseqEncoderDecoderModel
+ :members:
+ :undoc-members:
+.. autoclass:: fairseq.models.FairseqEncoderModel
+ :members:
+ :undoc-members:
+.. autoclass:: fairseq.models.FairseqLanguageModel
+ :members:
+ :undoc-members:
+.. autoclass:: fairseq.models.FairseqMultiModel
+ :members:
+ :undoc-members:
+.. autoclass:: fairseq.models.FairseqEncoder
+ :members:
+.. autoclass:: fairseq.models.CompositeEncoder
+ :members:
+.. autoclass:: fairseq.models.FairseqDecoder
+ :members:
+
+
+.. _Incremental decoding:
+
+Incremental decoding
+--------------------
+
+.. autoclass:: fairseq.models.FairseqIncrementalDecoder
+ :members:
+ :undoc-members:
diff --git a/fairseq/docs/modules.rst b/fairseq/docs/modules.rst
new file mode 100644
index 0000000000000000000000000000000000000000..9631c93d4682286e1cea1ddd961d3f6ab06f2589
--- /dev/null
+++ b/fairseq/docs/modules.rst
@@ -0,0 +1,9 @@
+Modules
+=======
+
+Fairseq provides several stand-alone :class:`torch.nn.Module` classes that may
+be helpful when implementing a new :class:`~fairseq.models.BaseFairseqModel`.
+
+.. automodule:: fairseq.modules
+ :members:
+ :undoc-members:
diff --git a/fairseq/docs/optim.rst b/fairseq/docs/optim.rst
new file mode 100644
index 0000000000000000000000000000000000000000..c3326456bd9291a1d05bd3316bef5c9fb25c6c49
--- /dev/null
+++ b/fairseq/docs/optim.rst
@@ -0,0 +1,38 @@
+.. role:: hidden
+ :class: hidden-section
+
+.. _optimizers:
+
+Optimizers
+==========
+
+Optimizers update the Model parameters based on the gradients.
+
+.. automodule:: fairseq.optim
+ :members:
+
+.. autoclass:: fairseq.optim.FairseqOptimizer
+ :members:
+ :undoc-members:
+
+.. autoclass:: fairseq.optim.adadelta.Adadelta
+ :members:
+ :undoc-members:
+.. autoclass:: fairseq.optim.adagrad.Adagrad
+ :members:
+ :undoc-members:
+.. autoclass:: fairseq.optim.adafactor.FairseqAdafactor
+ :members:
+ :undoc-members:
+.. autoclass:: fairseq.optim.adam.FairseqAdam
+ :members:
+ :undoc-members:
+.. autoclass:: fairseq.optim.fp16_optimizer.FP16Optimizer
+ :members:
+ :undoc-members:
+.. autoclass:: fairseq.optim.nag.FairseqNAG
+ :members:
+ :undoc-members:
+.. autoclass:: fairseq.optim.sgd.SGD
+ :members:
+ :undoc-members:
diff --git a/fairseq/docs/overview.rst b/fairseq/docs/overview.rst
new file mode 100644
index 0000000000000000000000000000000000000000..026b3b5c7b21d071d8b8a3405898977c760d05b8
--- /dev/null
+++ b/fairseq/docs/overview.rst
@@ -0,0 +1,74 @@
+Overview
+========
+
+Fairseq can be extended through user-supplied `plug-ins
+`_. We support five kinds of
+plug-ins:
+
+- :ref:`Models` define the neural network architecture and encapsulate all of the
+ learnable parameters.
+- :ref:`Criterions` compute the loss function given the model outputs and targets.
+- :ref:`Tasks` store dictionaries and provide helpers for loading/iterating over
+ Datasets, initializing the Model/Criterion and calculating the loss.
+- :ref:`Optimizers` update the Model parameters based on the gradients.
+- :ref:`Learning Rate Schedulers` update the learning rate over the course of
+ training.
+
+**Training Flow**
+
+Given a ``model``, ``criterion``, ``task``, ``optimizer`` and ``lr_scheduler``,
+fairseq implements the following high-level training flow::
+
+ for epoch in range(num_epochs):
+ itr = task.get_batch_iterator(task.dataset('train'))
+ for num_updates, batch in enumerate(itr):
+ task.train_step(batch, model, criterion, optimizer)
+ average_and_clip_gradients()
+ optimizer.step()
+ lr_scheduler.step_update(num_updates)
+ lr_scheduler.step(epoch)
+
+where the default implementation for ``task.train_step`` is roughly::
+
+ def train_step(self, batch, model, criterion, optimizer, **unused):
+ loss = criterion(model, batch)
+ optimizer.backward(loss)
+ return loss
+
+**Registering new plug-ins**
+
+New plug-ins are *registered* through a set of ``@register`` function
+decorators, for example::
+
+ @register_model('my_lstm')
+ class MyLSTM(FairseqEncoderDecoderModel):
+ (...)
+
+Once registered, new plug-ins can be used with the existing :ref:`Command-line
+Tools`. See the Tutorial sections for more detailed walkthroughs of how to add
+new plug-ins.
+
+**Loading plug-ins from another directory**
+
+New plug-ins can be defined in a custom module stored in the user system. In
+order to import the module, and make the plugin available to *fairseq*, the
+command line supports the ``--user-dir`` flag that can be used to specify a
+custom location for additional modules to load into *fairseq*.
+
+For example, assuming this directory tree::
+
+ /home/user/my-module/
+ └── __init__.py
+
+with ``__init__.py``::
+
+ from fairseq.models import register_model_architecture
+ from fairseq.models.transformer import transformer_vaswani_wmt_en_de_big
+
+ @register_model_architecture('transformer', 'my_transformer')
+ def transformer_mmt_big(args):
+ transformer_vaswani_wmt_en_de_big(args)
+
+it is possible to invoke the :ref:`fairseq-train` script with the new architecture with::
+
+ fairseq-train ... --user-dir /home/user/my-module -a my_transformer --task translation
diff --git a/fairseq/docs/tasks.rst b/fairseq/docs/tasks.rst
new file mode 100644
index 0000000000000000000000000000000000000000..5f65c3c866865e50332d8e6ca012a4a81e7bea74
--- /dev/null
+++ b/fairseq/docs/tasks.rst
@@ -0,0 +1,61 @@
+.. role:: hidden
+ :class: hidden-section
+
+.. module:: fairseq.tasks
+
+.. _Tasks:
+
+Tasks
+=====
+
+Tasks store dictionaries and provide helpers for loading/iterating over
+Datasets, initializing the Model/Criterion and calculating the loss.
+
+Tasks can be selected via the ``--task`` command-line argument. Once selected, a
+task may expose additional command-line arguments for further configuration.
+
+Example usage::
+
+ # setup the task (e.g., load dictionaries)
+ task = fairseq.tasks.setup_task(args)
+
+ # build model and criterion
+ model = task.build_model(args)
+ criterion = task.build_criterion(args)
+
+ # load datasets
+ task.load_dataset('train')
+ task.load_dataset('valid')
+
+ # iterate over mini-batches of data
+ batch_itr = task.get_batch_iterator(
+ task.dataset('train'), max_tokens=4096,
+ )
+ for batch in batch_itr:
+ # compute the loss
+ loss, sample_size, logging_output = task.get_loss(
+ model, criterion, batch,
+ )
+ loss.backward()
+
+
+Translation
+-----------
+
+.. autoclass:: fairseq.tasks.translation.TranslationTask
+
+.. _language modeling:
+
+Language Modeling
+-----------------
+
+.. autoclass:: fairseq.tasks.language_modeling.LanguageModelingTask
+
+
+Adding new tasks
+----------------
+
+.. autofunction:: fairseq.tasks.register_task
+.. autoclass:: fairseq.tasks.FairseqTask
+ :members:
+ :undoc-members:
diff --git a/fairseq/docs/tutorial_classifying_names.rst b/fairseq/docs/tutorial_classifying_names.rst
new file mode 100644
index 0000000000000000000000000000000000000000..de099f08f548d4fb92829922337a4b7f4ceacc9b
--- /dev/null
+++ b/fairseq/docs/tutorial_classifying_names.rst
@@ -0,0 +1,415 @@
+Tutorial: Classifying Names with a Character-Level RNN
+======================================================
+
+In this tutorial we will extend fairseq to support *classification* tasks. In
+particular we will re-implement the PyTorch tutorial for `Classifying Names with
+a Character-Level RNN `_
+in fairseq. It is recommended to quickly skim that tutorial before beginning
+this one.
+
+This tutorial covers:
+
+1. **Preprocessing the data** to create dictionaries.
+2. **Registering a new Model** that encodes an input sentence with a simple RNN
+ and predicts the output label.
+3. **Registering a new Task** that loads our dictionaries and dataset.
+4. **Training the Model** using the existing command-line tools.
+5. **Writing an evaluation script** that imports fairseq and allows us to
+ interactively evaluate our model on new inputs.
+
+
+1. Preprocessing the data
+-------------------------
+
+The original tutorial provides raw data, but we'll work with a modified version
+of the data that is already tokenized into characters and split into separate
+train, valid and test sets.
+
+Download and extract the data from here:
+`tutorial_names.tar.gz `_
+
+Once extracted, let's preprocess the data using the :ref:`fairseq-preprocess`
+command-line tool to create the dictionaries. While this tool is primarily
+intended for sequence-to-sequence problems, we're able to reuse it here by
+treating the label as a "target" sequence of length 1. We'll also output the
+preprocessed files in "raw" format using the ``--dataset-impl`` option to
+enhance readability:
+
+.. code-block:: console
+
+ > fairseq-preprocess \
+ --trainpref names/train --validpref names/valid --testpref names/test \
+ --source-lang input --target-lang label \
+ --destdir names-bin --dataset-impl raw
+
+After running the above command you should see a new directory,
+:file:`names-bin/`, containing the dictionaries for *inputs* and *labels*.
+
+
+2. Registering a new Model
+--------------------------
+
+Next we'll register a new model in fairseq that will encode an input sentence
+with a simple RNN and predict the output label. Compared to the original PyTorch
+tutorial, our version will also work with batches of data and GPU Tensors.
+
+First let's copy the simple RNN module implemented in the `PyTorch tutorial
+`_.
+Create a new file named :file:`fairseq/models/rnn_classifier.py` with the
+following contents::
+
+ import torch
+ import torch.nn as nn
+
+ class RNN(nn.Module):
+
+ def __init__(self, input_size, hidden_size, output_size):
+ super(RNN, self).__init__()
+
+ self.hidden_size = hidden_size
+
+ self.i2h = nn.Linear(input_size + hidden_size, hidden_size)
+ self.i2o = nn.Linear(input_size + hidden_size, output_size)
+ self.softmax = nn.LogSoftmax(dim=1)
+
+ def forward(self, input, hidden):
+ combined = torch.cat((input, hidden), 1)
+ hidden = self.i2h(combined)
+ output = self.i2o(combined)
+ output = self.softmax(output)
+ return output, hidden
+
+ def initHidden(self):
+ return torch.zeros(1, self.hidden_size)
+
+We must also *register* this model with fairseq using the
+:func:`~fairseq.models.register_model` function decorator. Once the model is
+registered we'll be able to use it with the existing :ref:`Command-line Tools`.
+
+All registered models must implement the :class:`~fairseq.models.BaseFairseqModel`
+interface, so we'll create a small wrapper class in the same file and register
+it in fairseq with the name ``'rnn_classifier'``::
+
+ from fairseq.models import BaseFairseqModel, register_model
+
+ # Note: the register_model "decorator" should immediately precede the
+ # definition of the Model class.
+
+ @register_model('rnn_classifier')
+ class FairseqRNNClassifier(BaseFairseqModel):
+
+ @staticmethod
+ def add_args(parser):
+ # Models can override this method to add new command-line arguments.
+ # Here we'll add a new command-line argument to configure the
+ # dimensionality of the hidden state.
+ parser.add_argument(
+ '--hidden-dim', type=int, metavar='N',
+ help='dimensionality of the hidden state',
+ )
+
+ @classmethod
+ def build_model(cls, args, task):
+ # Fairseq initializes models by calling the ``build_model()``
+ # function. This provides more flexibility, since the returned model
+ # instance can be of a different type than the one that was called.
+ # In this case we'll just return a FairseqRNNClassifier instance.
+
+ # Initialize our RNN module
+ rnn = RNN(
+ # We'll define the Task in the next section, but for now just
+ # notice that the task holds the dictionaries for the "source"
+ # (i.e., the input sentence) and "target" (i.e., the label).
+ input_size=len(task.source_dictionary),
+ hidden_size=args.hidden_dim,
+ output_size=len(task.target_dictionary),
+ )
+
+ # Return the wrapped version of the module
+ return FairseqRNNClassifier(
+ rnn=rnn,
+ input_vocab=task.source_dictionary,
+ )
+
+ def __init__(self, rnn, input_vocab):
+ super(FairseqRNNClassifier, self).__init__()
+
+ self.rnn = rnn
+ self.input_vocab = input_vocab
+
+ # The RNN module in the tutorial expects one-hot inputs, so we can
+ # precompute the identity matrix to help convert from indices to
+ # one-hot vectors. We register it as a buffer so that it is moved to
+ # the GPU when ``cuda()`` is called.
+ self.register_buffer('one_hot_inputs', torch.eye(len(input_vocab)))
+
+ def forward(self, src_tokens, src_lengths):
+ # The inputs to the ``forward()`` function are determined by the
+ # Task, and in particular the ``'net_input'`` key in each
+ # mini-batch. We'll define the Task in the next section, but for
+ # now just know that *src_tokens* has shape `(batch, src_len)` and
+ # *src_lengths* has shape `(batch)`.
+ bsz, max_src_len = src_tokens.size()
+
+ # Initialize the RNN hidden state. Compared to the original PyTorch
+ # tutorial we'll also handle batched inputs and work on the GPU.
+ hidden = self.rnn.initHidden()
+ hidden = hidden.repeat(bsz, 1) # expand for batched inputs
+ hidden = hidden.to(src_tokens.device) # move to GPU
+
+ for i in range(max_src_len):
+ # WARNING: The inputs have padding, so we should mask those
+ # elements here so that padding doesn't affect the results.
+ # This is left as an exercise for the reader. The padding symbol
+ # is given by ``self.input_vocab.pad()`` and the unpadded length
+ # of each input is given by *src_lengths*.
+
+ # One-hot encode a batch of input characters.
+ input = self.one_hot_inputs[src_tokens[:, i].long()]
+
+ # Feed the input to our RNN.
+ output, hidden = self.rnn(input, hidden)
+
+ # Return the final output state for making a prediction
+ return output
+
+Finally let's define a *named architecture* with the configuration for our
+model. This is done with the :func:`~fairseq.models.register_model_architecture`
+function decorator. Thereafter this named architecture can be used with the
+``--arch`` command-line argument, e.g., ``--arch pytorch_tutorial_rnn``::
+
+ from fairseq.models import register_model_architecture
+
+ # The first argument to ``register_model_architecture()`` should be the name
+ # of the model we registered above (i.e., 'rnn_classifier'). The function we
+ # register here should take a single argument *args* and modify it in-place
+ # to match the desired architecture.
+
+ @register_model_architecture('rnn_classifier', 'pytorch_tutorial_rnn')
+ def pytorch_tutorial_rnn(args):
+ # We use ``getattr()`` to prioritize arguments that are explicitly given
+ # on the command-line, so that the defaults defined below are only used
+ # when no other value has been specified.
+ args.hidden_dim = getattr(args, 'hidden_dim', 128)
+
+
+3. Registering a new Task
+-------------------------
+
+Now we'll register a new :class:`~fairseq.tasks.FairseqTask` that will load our
+dictionaries and dataset. Tasks can also control how the data is batched into
+mini-batches, but in this tutorial we'll reuse the batching provided by
+:class:`fairseq.data.LanguagePairDataset`.
+
+Create a new file named :file:`fairseq/tasks/simple_classification.py` with the
+following contents::
+
+ import os
+ import torch
+
+ from fairseq.data import Dictionary, LanguagePairDataset
+ from fairseq.tasks import LegacyFairseqTask, register_task
+
+
+ @register_task('simple_classification')
+ class SimpleClassificationTask(LegacyFairseqTask):
+
+ @staticmethod
+ def add_args(parser):
+ # Add some command-line arguments for specifying where the data is
+ # located and the maximum supported input length.
+ parser.add_argument('data', metavar='FILE',
+ help='file prefix for data')
+ parser.add_argument('--max-positions', default=1024, type=int,
+ help='max input length')
+
+ @classmethod
+ def setup_task(cls, args, **kwargs):
+ # Here we can perform any setup required for the task. This may include
+ # loading Dictionaries, initializing shared Embedding layers, etc.
+ # In this case we'll just load the Dictionaries.
+ input_vocab = Dictionary.load(os.path.join(args.data, 'dict.input.txt'))
+ label_vocab = Dictionary.load(os.path.join(args.data, 'dict.label.txt'))
+ print('| [input] dictionary: {} types'.format(len(input_vocab)))
+ print('| [label] dictionary: {} types'.format(len(label_vocab)))
+
+ return SimpleClassificationTask(args, input_vocab, label_vocab)
+
+ def __init__(self, args, input_vocab, label_vocab):
+ super().__init__(args)
+ self.input_vocab = input_vocab
+ self.label_vocab = label_vocab
+
+ def load_dataset(self, split, **kwargs):
+ """Load a given dataset split (e.g., train, valid, test)."""
+
+ prefix = os.path.join(self.args.data, '{}.input-label'.format(split))
+
+ # Read input sentences.
+ sentences, lengths = [], []
+ with open(prefix + '.input', encoding='utf-8') as file:
+ for line in file:
+ sentence = line.strip()
+
+ # Tokenize the sentence, splitting on spaces
+ tokens = self.input_vocab.encode_line(
+ sentence, add_if_not_exist=False,
+ )
+
+ sentences.append(tokens)
+ lengths.append(tokens.numel())
+
+ # Read labels.
+ labels = []
+ with open(prefix + '.label', encoding='utf-8') as file:
+ for line in file:
+ label = line.strip()
+ labels.append(
+ # Convert label to a numeric ID.
+ torch.LongTensor([self.label_vocab.add_symbol(label)])
+ )
+
+ assert len(sentences) == len(labels)
+ print('| {} {} {} examples'.format(self.args.data, split, len(sentences)))
+
+ # We reuse LanguagePairDataset since classification can be modeled as a
+ # sequence-to-sequence task where the target sequence has length 1.
+ self.datasets[split] = LanguagePairDataset(
+ src=sentences,
+ src_sizes=lengths,
+ src_dict=self.input_vocab,
+ tgt=labels,
+ tgt_sizes=torch.ones(len(labels)), # targets have length 1
+ tgt_dict=self.label_vocab,
+ left_pad_source=False,
+ # Since our target is a single class label, there's no need for
+ # teacher forcing. If we set this to ``True`` then our Model's
+ # ``forward()`` method would receive an additional argument called
+ # *prev_output_tokens* that would contain a shifted version of the
+ # target sequence.
+ input_feeding=False,
+ )
+
+ def max_positions(self):
+ """Return the max input length allowed by the task."""
+ # The source should be less than *args.max_positions* and the "target"
+ # has max length 1.
+ return (self.args.max_positions, 1)
+
+ @property
+ def source_dictionary(self):
+ """Return the source :class:`~fairseq.data.Dictionary`."""
+ return self.input_vocab
+
+ @property
+ def target_dictionary(self):
+ """Return the target :class:`~fairseq.data.Dictionary`."""
+ return self.label_vocab
+
+ # We could override this method if we wanted more control over how batches
+ # are constructed, but it's not necessary for this tutorial since we can
+ # reuse the batching provided by LanguagePairDataset.
+ #
+ # def get_batch_iterator(
+ # self, dataset, max_tokens=None, max_sentences=None, max_positions=None,
+ # ignore_invalid_inputs=False, required_batch_size_multiple=1,
+ # seed=1, num_shards=1, shard_id=0, num_workers=0, epoch=1,
+ # data_buffer_size=0, disable_iterator_cache=False,
+ # ):
+ # (...)
+
+
+4. Training the Model
+---------------------
+
+Now we're ready to train the model. We can use the existing :ref:`fairseq-train`
+command-line tool for this, making sure to specify our new Task (``--task
+simple_classification``) and Model architecture (``--arch
+pytorch_tutorial_rnn``):
+
+.. note::
+
+ You can also configure the dimensionality of the hidden state by passing the
+ ``--hidden-dim`` argument to :ref:`fairseq-train`.
+
+.. code-block:: console
+
+ > fairseq-train names-bin \
+ --task simple_classification \
+ --arch pytorch_tutorial_rnn \
+ --optimizer adam --lr 0.001 --lr-shrink 0.5 \
+ --max-tokens 1000
+ (...)
+ | epoch 027 | loss 1.200 | ppl 2.30 | wps 15728 | ups 119.4 | wpb 116 | bsz 116 | num_updates 3726 | lr 1.5625e-05 | gnorm 1.290 | clip 0% | oom 0 | wall 32 | train_wall 21
+ | epoch 027 | valid on 'valid' subset | valid_loss 1.41304 | valid_ppl 2.66 | num_updates 3726 | best 1.41208
+ | done training in 31.6 seconds
+
+The model files should appear in the :file:`checkpoints/` directory.
+
+
+5. Writing an evaluation script
+-------------------------------
+
+Finally we can write a short script to evaluate our model on new inputs. Create
+a new file named :file:`eval_classifier.py` with the following contents::
+
+ from fairseq import checkpoint_utils, data, options, tasks
+
+ # Parse command-line arguments for generation
+ parser = options.get_generation_parser(default_task='simple_classification')
+ args = options.parse_args_and_arch(parser)
+
+ # Setup task
+ task = tasks.setup_task(args)
+
+ # Load model
+ print('| loading model from {}'.format(args.path))
+ models, _model_args = checkpoint_utils.load_model_ensemble([args.path], task=task)
+ model = models[0]
+
+ while True:
+ sentence = input('\nInput: ')
+
+ # Tokenize into characters
+ chars = ' '.join(list(sentence.strip()))
+ tokens = task.source_dictionary.encode_line(
+ chars, add_if_not_exist=False,
+ )
+
+ # Build mini-batch to feed to the model
+ batch = data.language_pair_dataset.collate(
+ samples=[{'id': -1, 'source': tokens}], # bsz = 1
+ pad_idx=task.source_dictionary.pad(),
+ eos_idx=task.source_dictionary.eos(),
+ left_pad_source=False,
+ input_feeding=False,
+ )
+
+ # Feed batch to the model and get predictions
+ preds = model(**batch['net_input'])
+
+ # Print top 3 predictions and their log-probabilities
+ top_scores, top_labels = preds[0].topk(k=3)
+ for score, label_idx in zip(top_scores, top_labels):
+ label_name = task.target_dictionary.string([label_idx])
+ print('({:.2f})\t{}'.format(score, label_name))
+
+Now we can evaluate our model interactively. Note that we have included the
+original data path (:file:`names-bin/`) so that the dictionaries can be loaded:
+
+.. code-block:: console
+
+ > python eval_classifier.py names-bin --path checkpoints/checkpoint_best.pt
+ | [input] dictionary: 64 types
+ | [label] dictionary: 24 types
+ | loading model from checkpoints/checkpoint_best.pt
+
+ Input: Satoshi
+ (-0.61) Japanese
+ (-1.20) Arabic
+ (-2.86) Italian
+
+ Input: Sinbad
+ (-0.30) Arabic
+ (-1.76) English
+ (-4.08) Russian
diff --git a/fairseq/docs/tutorial_simple_lstm.rst b/fairseq/docs/tutorial_simple_lstm.rst
new file mode 100644
index 0000000000000000000000000000000000000000..f52988507c5da5125668e143bd2bfe4df117b41c
--- /dev/null
+++ b/fairseq/docs/tutorial_simple_lstm.rst
@@ -0,0 +1,518 @@
+Tutorial: Simple LSTM
+=====================
+
+In this tutorial we will extend fairseq by adding a new
+:class:`~fairseq.models.FairseqEncoderDecoderModel` that encodes a source
+sentence with an LSTM and then passes the final hidden state to a second LSTM
+that decodes the target sentence (without attention).
+
+This tutorial covers:
+
+1. **Writing an Encoder and Decoder** to encode/decode the source/target
+ sentence, respectively.
+2. **Registering a new Model** so that it can be used with the existing
+ :ref:`Command-line tools`.
+3. **Training the Model** using the existing command-line tools.
+4. **Making generation faster** by modifying the Decoder to use
+ :ref:`Incremental decoding`.
+
+
+1. Building an Encoder and Decoder
+----------------------------------
+
+In this section we'll define a simple LSTM Encoder and Decoder. All Encoders
+should implement the :class:`~fairseq.models.FairseqEncoder` interface and
+Decoders should implement the :class:`~fairseq.models.FairseqDecoder` interface.
+These interfaces themselves extend :class:`torch.nn.Module`, so FairseqEncoders
+and FairseqDecoders can be written and used in the same ways as ordinary PyTorch
+Modules.
+
+
+Encoder
+~~~~~~~
+
+Our Encoder will embed the tokens in the source sentence, feed them to a
+:class:`torch.nn.LSTM` and return the final hidden state. To create our encoder
+save the following in a new file named :file:`fairseq/models/simple_lstm.py`::
+
+ import torch.nn as nn
+ from fairseq import utils
+ from fairseq.models import FairseqEncoder
+
+ class SimpleLSTMEncoder(FairseqEncoder):
+
+ def __init__(
+ self, args, dictionary, embed_dim=128, hidden_dim=128, dropout=0.1,
+ ):
+ super().__init__(dictionary)
+ self.args = args
+
+ # Our encoder will embed the inputs before feeding them to the LSTM.
+ self.embed_tokens = nn.Embedding(
+ num_embeddings=len(dictionary),
+ embedding_dim=embed_dim,
+ padding_idx=dictionary.pad(),
+ )
+ self.dropout = nn.Dropout(p=dropout)
+
+ # We'll use a single-layer, unidirectional LSTM for simplicity.
+ self.lstm = nn.LSTM(
+ input_size=embed_dim,
+ hidden_size=hidden_dim,
+ num_layers=1,
+ bidirectional=False,
+ batch_first=True,
+ )
+
+ def forward(self, src_tokens, src_lengths):
+ # The inputs to the ``forward()`` function are determined by the
+ # Task, and in particular the ``'net_input'`` key in each
+ # mini-batch. We discuss Tasks in the next tutorial, but for now just
+ # know that *src_tokens* has shape `(batch, src_len)` and *src_lengths*
+ # has shape `(batch)`.
+
+ # Note that the source is typically padded on the left. This can be
+ # configured by adding the `--left-pad-source "False"` command-line
+ # argument, but here we'll make the Encoder handle either kind of
+ # padding by converting everything to be right-padded.
+ if self.args.left_pad_source:
+ # Convert left-padding to right-padding.
+ src_tokens = utils.convert_padding_direction(
+ src_tokens,
+ padding_idx=self.dictionary.pad(),
+ left_to_right=True
+ )
+
+ # Embed the source.
+ x = self.embed_tokens(src_tokens)
+
+ # Apply dropout.
+ x = self.dropout(x)
+
+ # Pack the sequence into a PackedSequence object to feed to the LSTM.
+ x = nn.utils.rnn.pack_padded_sequence(x, src_lengths, batch_first=True)
+
+ # Get the output from the LSTM.
+ _outputs, (final_hidden, _final_cell) = self.lstm(x)
+
+ # Return the Encoder's output. This can be any object and will be
+ # passed directly to the Decoder.
+ return {
+ # this will have shape `(bsz, hidden_dim)`
+ 'final_hidden': final_hidden.squeeze(0),
+ }
+
+ # Encoders are required to implement this method so that we can rearrange
+ # the order of the batch elements during inference (e.g., beam search).
+ def reorder_encoder_out(self, encoder_out, new_order):
+ """
+ Reorder encoder output according to `new_order`.
+
+ Args:
+ encoder_out: output from the ``forward()`` method
+ new_order (LongTensor): desired order
+
+ Returns:
+ `encoder_out` rearranged according to `new_order`
+ """
+ final_hidden = encoder_out['final_hidden']
+ return {
+ 'final_hidden': final_hidden.index_select(0, new_order),
+ }
+
+
+Decoder
+~~~~~~~
+
+Our Decoder will predict the next word, conditioned on the Encoder's final
+hidden state and an embedded representation of the previous target word -- which
+is sometimes called *teacher forcing*. More specifically, we'll use a
+:class:`torch.nn.LSTM` to produce a sequence of hidden states that we'll project
+to the size of the output vocabulary to predict each target word.
+
+::
+
+ import torch
+ from fairseq.models import FairseqDecoder
+
+ class SimpleLSTMDecoder(FairseqDecoder):
+
+ def __init__(
+ self, dictionary, encoder_hidden_dim=128, embed_dim=128, hidden_dim=128,
+ dropout=0.1,
+ ):
+ super().__init__(dictionary)
+
+ # Our decoder will embed the inputs before feeding them to the LSTM.
+ self.embed_tokens = nn.Embedding(
+ num_embeddings=len(dictionary),
+ embedding_dim=embed_dim,
+ padding_idx=dictionary.pad(),
+ )
+ self.dropout = nn.Dropout(p=dropout)
+
+ # We'll use a single-layer, unidirectional LSTM for simplicity.
+ self.lstm = nn.LSTM(
+ # For the first layer we'll concatenate the Encoder's final hidden
+ # state with the embedded target tokens.
+ input_size=encoder_hidden_dim + embed_dim,
+ hidden_size=hidden_dim,
+ num_layers=1,
+ bidirectional=False,
+ )
+
+ # Define the output projection.
+ self.output_projection = nn.Linear(hidden_dim, len(dictionary))
+
+ # During training Decoders are expected to take the entire target sequence
+ # (shifted right by one position) and produce logits over the vocabulary.
+ # The *prev_output_tokens* tensor begins with the end-of-sentence symbol,
+ # ``dictionary.eos()``, followed by the target sequence.
+ def forward(self, prev_output_tokens, encoder_out):
+ """
+ Args:
+ prev_output_tokens (LongTensor): previous decoder outputs of shape
+ `(batch, tgt_len)`, for teacher forcing
+ encoder_out (Tensor, optional): output from the encoder, used for
+ encoder-side attention
+
+ Returns:
+ tuple:
+ - the last decoder layer's output of shape
+ `(batch, tgt_len, vocab)`
+ - the last decoder layer's attention weights of shape
+ `(batch, tgt_len, src_len)`
+ """
+ bsz, tgt_len = prev_output_tokens.size()
+
+ # Extract the final hidden state from the Encoder.
+ final_encoder_hidden = encoder_out['final_hidden']
+
+ # Embed the target sequence, which has been shifted right by one
+ # position and now starts with the end-of-sentence symbol.
+ x = self.embed_tokens(prev_output_tokens)
+
+ # Apply dropout.
+ x = self.dropout(x)
+
+ # Concatenate the Encoder's final hidden state to *every* embedded
+ # target token.
+ x = torch.cat(
+ [x, final_encoder_hidden.unsqueeze(1).expand(bsz, tgt_len, -1)],
+ dim=2,
+ )
+
+ # Using PackedSequence objects in the Decoder is harder than in the
+ # Encoder, since the targets are not sorted in descending length order,
+ # which is a requirement of ``pack_padded_sequence()``. Instead we'll
+ # feed nn.LSTM directly.
+ initial_state = (
+ final_encoder_hidden.unsqueeze(0), # hidden
+ torch.zeros_like(final_encoder_hidden).unsqueeze(0), # cell
+ )
+ output, _ = self.lstm(
+ x.transpose(0, 1), # convert to shape `(tgt_len, bsz, dim)`
+ initial_state,
+ )
+ x = output.transpose(0, 1) # convert to shape `(bsz, tgt_len, hidden)`
+
+ # Project the outputs to the size of the vocabulary.
+ x = self.output_projection(x)
+
+ # Return the logits and ``None`` for the attention weights
+ return x, None
+
+
+2. Registering the Model
+------------------------
+
+Now that we've defined our Encoder and Decoder we must *register* our model with
+fairseq using the :func:`~fairseq.models.register_model` function decorator.
+Once the model is registered we'll be able to use it with the existing
+:ref:`Command-line Tools`.
+
+All registered models must implement the
+:class:`~fairseq.models.BaseFairseqModel` interface. For sequence-to-sequence
+models (i.e., any model with a single Encoder and Decoder), we can instead
+implement the :class:`~fairseq.models.FairseqEncoderDecoderModel` interface.
+
+Create a small wrapper class in the same file and register it in fairseq with
+the name ``'simple_lstm'``::
+
+ from fairseq.models import FairseqEncoderDecoderModel, register_model
+
+ # Note: the register_model "decorator" should immediately precede the
+ # definition of the Model class.
+
+ @register_model('simple_lstm')
+ class SimpleLSTMModel(FairseqEncoderDecoderModel):
+
+ @staticmethod
+ def add_args(parser):
+ # Models can override this method to add new command-line arguments.
+ # Here we'll add some new command-line arguments to configure dropout
+ # and the dimensionality of the embeddings and hidden states.
+ parser.add_argument(
+ '--encoder-embed-dim', type=int, metavar='N',
+ help='dimensionality of the encoder embeddings',
+ )
+ parser.add_argument(
+ '--encoder-hidden-dim', type=int, metavar='N',
+ help='dimensionality of the encoder hidden state',
+ )
+ parser.add_argument(
+ '--encoder-dropout', type=float, default=0.1,
+ help='encoder dropout probability',
+ )
+ parser.add_argument(
+ '--decoder-embed-dim', type=int, metavar='N',
+ help='dimensionality of the decoder embeddings',
+ )
+ parser.add_argument(
+ '--decoder-hidden-dim', type=int, metavar='N',
+ help='dimensionality of the decoder hidden state',
+ )
+ parser.add_argument(
+ '--decoder-dropout', type=float, default=0.1,
+ help='decoder dropout probability',
+ )
+
+ @classmethod
+ def build_model(cls, args, task):
+ # Fairseq initializes models by calling the ``build_model()``
+ # function. This provides more flexibility, since the returned model
+ # instance can be of a different type than the one that was called.
+ # In this case we'll just return a SimpleLSTMModel instance.
+
+ # Initialize our Encoder and Decoder.
+ encoder = SimpleLSTMEncoder(
+ args=args,
+ dictionary=task.source_dictionary,
+ embed_dim=args.encoder_embed_dim,
+ hidden_dim=args.encoder_hidden_dim,
+ dropout=args.encoder_dropout,
+ )
+ decoder = SimpleLSTMDecoder(
+ dictionary=task.target_dictionary,
+ encoder_hidden_dim=args.encoder_hidden_dim,
+ embed_dim=args.decoder_embed_dim,
+ hidden_dim=args.decoder_hidden_dim,
+ dropout=args.decoder_dropout,
+ )
+ model = SimpleLSTMModel(encoder, decoder)
+
+ # Print the model architecture.
+ print(model)
+
+ return model
+
+ # We could override the ``forward()`` if we wanted more control over how
+ # the encoder and decoder interact, but it's not necessary for this
+ # tutorial since we can inherit the default implementation provided by
+ # the FairseqEncoderDecoderModel base class, which looks like:
+ #
+ # def forward(self, src_tokens, src_lengths, prev_output_tokens):
+ # encoder_out = self.encoder(src_tokens, src_lengths)
+ # decoder_out = self.decoder(prev_output_tokens, encoder_out)
+ # return decoder_out
+
+Finally let's define a *named architecture* with the configuration for our
+model. This is done with the :func:`~fairseq.models.register_model_architecture`
+function decorator. Thereafter this named architecture can be used with the
+``--arch`` command-line argument, e.g., ``--arch tutorial_simple_lstm``::
+
+ from fairseq.models import register_model_architecture
+
+ # The first argument to ``register_model_architecture()`` should be the name
+ # of the model we registered above (i.e., 'simple_lstm'). The function we
+ # register here should take a single argument *args* and modify it in-place
+ # to match the desired architecture.
+
+ @register_model_architecture('simple_lstm', 'tutorial_simple_lstm')
+ def tutorial_simple_lstm(args):
+ # We use ``getattr()`` to prioritize arguments that are explicitly given
+ # on the command-line, so that the defaults defined below are only used
+ # when no other value has been specified.
+ args.encoder_embed_dim = getattr(args, 'encoder_embed_dim', 256)
+ args.encoder_hidden_dim = getattr(args, 'encoder_hidden_dim', 256)
+ args.decoder_embed_dim = getattr(args, 'decoder_embed_dim', 256)
+ args.decoder_hidden_dim = getattr(args, 'decoder_hidden_dim', 256)
+
+
+3. Training the Model
+---------------------
+
+Now we're ready to train the model. We can use the existing :ref:`fairseq-train`
+command-line tool for this, making sure to specify our new Model architecture
+(``--arch tutorial_simple_lstm``).
+
+.. note::
+
+ Make sure you've already preprocessed the data from the IWSLT example in the
+ :file:`examples/translation/` directory.
+
+.. code-block:: console
+
+ > fairseq-train data-bin/iwslt14.tokenized.de-en \
+ --arch tutorial_simple_lstm \
+ --encoder-dropout 0.2 --decoder-dropout 0.2 \
+ --optimizer adam --lr 0.005 --lr-shrink 0.5 \
+ --max-tokens 12000
+ (...)
+ | epoch 052 | loss 4.027 | ppl 16.30 | wps 420805 | ups 39.7 | wpb 9841 | bsz 400 | num_updates 20852 | lr 1.95313e-05 | gnorm 0.218 | clip 0% | oom 0 | wall 529 | train_wall 396
+ | epoch 052 | valid on 'valid' subset | valid_loss 4.74989 | valid_ppl 26.91 | num_updates 20852 | best 4.74954
+
+The model files should appear in the :file:`checkpoints/` directory. While this
+model architecture is not very good, we can use the :ref:`fairseq-generate` script to
+generate translations and compute our BLEU score over the test set:
+
+.. code-block:: console
+
+ > fairseq-generate data-bin/iwslt14.tokenized.de-en \
+ --path checkpoints/checkpoint_best.pt \
+ --beam 5 \
+ --remove-bpe
+ (...)
+ | Translated 6750 sentences (153132 tokens) in 17.3s (389.12 sentences/s, 8827.68 tokens/s)
+ | Generate test with beam=5: BLEU4 = 8.18, 38.8/12.1/4.7/2.0 (BP=1.000, ratio=1.066, syslen=139865, reflen=131146)
+
+
+4. Making generation faster
+---------------------------
+
+While autoregressive generation from sequence-to-sequence models is inherently
+slow, our implementation above is especially slow because it recomputes the
+entire sequence of Decoder hidden states for every output token (i.e., it is
+``O(n^2)``). We can make this significantly faster by instead caching the
+previous hidden states.
+
+In fairseq this is called :ref:`Incremental decoding`. Incremental decoding is a
+special mode at inference time where the Model only receives a single timestep
+of input corresponding to the immediately previous output token (for teacher
+forcing) and must produce the next output incrementally. Thus the model must
+cache any long-term state that is needed about the sequence, e.g., hidden
+states, convolutional states, etc.
+
+To implement incremental decoding we will modify our model to implement the
+:class:`~fairseq.models.FairseqIncrementalDecoder` interface. Compared to the
+standard :class:`~fairseq.models.FairseqDecoder` interface, the incremental
+decoder interface allows ``forward()`` methods to take an extra keyword argument
+(*incremental_state*) that can be used to cache state across time-steps.
+
+Let's replace our ``SimpleLSTMDecoder`` with an incremental one::
+
+ import torch
+ from fairseq.models import FairseqIncrementalDecoder
+
+ class SimpleLSTMDecoder(FairseqIncrementalDecoder):
+
+ def __init__(
+ self, dictionary, encoder_hidden_dim=128, embed_dim=128, hidden_dim=128,
+ dropout=0.1,
+ ):
+ # This remains the same as before.
+ super().__init__(dictionary)
+ self.embed_tokens = nn.Embedding(
+ num_embeddings=len(dictionary),
+ embedding_dim=embed_dim,
+ padding_idx=dictionary.pad(),
+ )
+ self.dropout = nn.Dropout(p=dropout)
+ self.lstm = nn.LSTM(
+ input_size=encoder_hidden_dim + embed_dim,
+ hidden_size=hidden_dim,
+ num_layers=1,
+ bidirectional=False,
+ )
+ self.output_projection = nn.Linear(hidden_dim, len(dictionary))
+
+ # We now take an additional kwarg (*incremental_state*) for caching the
+ # previous hidden and cell states.
+ def forward(self, prev_output_tokens, encoder_out, incremental_state=None):
+ if incremental_state is not None:
+ # If the *incremental_state* argument is not ``None`` then we are
+ # in incremental inference mode. While *prev_output_tokens* will
+ # still contain the entire decoded prefix, we will only use the
+ # last step and assume that the rest of the state is cached.
+ prev_output_tokens = prev_output_tokens[:, -1:]
+
+ # This remains the same as before.
+ bsz, tgt_len = prev_output_tokens.size()
+ final_encoder_hidden = encoder_out['final_hidden']
+ x = self.embed_tokens(prev_output_tokens)
+ x = self.dropout(x)
+ x = torch.cat(
+ [x, final_encoder_hidden.unsqueeze(1).expand(bsz, tgt_len, -1)],
+ dim=2,
+ )
+
+ # We will now check the cache and load the cached previous hidden and
+ # cell states, if they exist, otherwise we will initialize them to
+ # zeros (as before). We will use the ``utils.get_incremental_state()``
+ # and ``utils.set_incremental_state()`` helpers.
+ initial_state = utils.get_incremental_state(
+ self, incremental_state, 'prev_state',
+ )
+ if initial_state is None:
+ # first time initialization, same as the original version
+ initial_state = (
+ final_encoder_hidden.unsqueeze(0), # hidden
+ torch.zeros_like(final_encoder_hidden).unsqueeze(0), # cell
+ )
+
+ # Run one step of our LSTM.
+ output, latest_state = self.lstm(x.transpose(0, 1), initial_state)
+
+ # Update the cache with the latest hidden and cell states.
+ utils.set_incremental_state(
+ self, incremental_state, 'prev_state', latest_state,
+ )
+
+ # This remains the same as before
+ x = output.transpose(0, 1)
+ x = self.output_projection(x)
+ return x, None
+
+ # The ``FairseqIncrementalDecoder`` interface also requires implementing a
+ # ``reorder_incremental_state()`` method, which is used during beam search
+ # to select and reorder the incremental state.
+ def reorder_incremental_state(self, incremental_state, new_order):
+ # Load the cached state.
+ prev_state = utils.get_incremental_state(
+ self, incremental_state, 'prev_state',
+ )
+
+ # Reorder batches according to *new_order*.
+ reordered_state = (
+ prev_state[0].index_select(1, new_order), # hidden
+ prev_state[1].index_select(1, new_order), # cell
+ )
+
+ # Update the cached state.
+ utils.set_incremental_state(
+ self, incremental_state, 'prev_state', reordered_state,
+ )
+
+Finally, we can rerun generation and observe the speedup:
+
+.. code-block:: console
+
+ # Before
+
+ > fairseq-generate data-bin/iwslt14.tokenized.de-en \
+ --path checkpoints/checkpoint_best.pt \
+ --beam 5 \
+ --remove-bpe
+ (...)
+ | Translated 6750 sentences (153132 tokens) in 17.3s (389.12 sentences/s, 8827.68 tokens/s)
+ | Generate test with beam=5: BLEU4 = 8.18, 38.8/12.1/4.7/2.0 (BP=1.000, ratio=1.066, syslen=139865, reflen=131146)
+
+ # After
+
+ > fairseq-generate data-bin/iwslt14.tokenized.de-en \
+ --path checkpoints/checkpoint_best.pt \
+ --beam 5 \
+ --remove-bpe
+ (...)
+ | Translated 6750 sentences (153132 tokens) in 5.5s (1225.54 sentences/s, 27802.94 tokens/s)
+ | Generate test with beam=5: BLEU4 = 8.18, 38.8/12.1/4.7/2.0 (BP=1.000, ratio=1.066, syslen=139865, reflen=131146)
diff --git a/fairseq/examples/.gitignore b/fairseq/examples/.gitignore
new file mode 100644
index 0000000000000000000000000000000000000000..1ef816f2cd7b4a9aa7adf8bd5635a644834738f1
--- /dev/null
+++ b/fairseq/examples/.gitignore
@@ -0,0 +1,2 @@
+!*/*.sh
+!*/*.md
diff --git a/fairseq/examples/MMPT/.gitignore b/fairseq/examples/MMPT/.gitignore
new file mode 100644
index 0000000000000000000000000000000000000000..70a255dc918bf4a4242dd7ea5ded63e6848264db
--- /dev/null
+++ b/fairseq/examples/MMPT/.gitignore
@@ -0,0 +1,139 @@
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+pip-wheel-metadata/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+# Usually these files are written by a python script from a template
+# before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+.python-version
+
+# pipenv
+# According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+# However, in case of collaboration, if having platform-specific dependencies or dependencies
+# having no cross-platform support, pipenv may install dependencies that don't work, or not
+# install all needed dependencies.
+#Pipfile.lock
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+runs
+data
+pretrained_models
+projects/mmfusion_*
+log_test
+third-party
+python_log
+slurm_snapshot_code
+lightning_logs
+demos
diff --git a/fairseq/examples/MMPT/CONFIG.md b/fairseq/examples/MMPT/CONFIG.md
new file mode 100644
index 0000000000000000000000000000000000000000..bbd1403dfafc79ab581c691281ca49de28a4656f
--- /dev/null
+++ b/fairseq/examples/MMPT/CONFIG.md
@@ -0,0 +1,41 @@
+### Config Files Explained
+
+Taking `projects/mfmmlm.yaml` for example, which run pretraining using masked frame model (MFM) and masked language model (MLM) on a single BERT:
+
+```yaml
+project_dir: mfmmlm # specify the project dir for this baseline.
+run_task:
+ - how2.yaml # run pretraining on how2 when launching `projects/taskmfmmlm.yaml`
+ - [vtt.yaml, vttcap.yaml, vttqa.yaml, youcook.yaml, youcookcap.yaml, crosstask.yaml, coin.yaml] # run fine-tuning tasks.
+base_dir: task # a global template folder to specify each training task.
+task_group:
+ pretrain: # section for pretraining. Most baselines differs in this section.
+ task_list:
+ - how2.yaml # reconfig `projects/task/how2.yaml`
+ dataset:
+ aligner: MFMMLMAligner # overwrite the aligner for MFMMLM training task.
+ model:
+ model_cls: MMFusionMFMMLM # overwrite the model, which constructs negative examples for MFM on-the-fly.
+ loss:
+ loss_cls: MFMMLM # overwrite the loss as MFMMLM, which combines MFM and MLM together.
+ fairseq: # all fairseq args can be expecified under this name.
+ dataset:
+ batch_size: 128
+ finetune: # section for fine-tuning tasks, we don't need to change anything here mostly since we want to see how pretraining can contribute to finetuning.
+ task_list: # specify the list of downstream tasks, e.g., copy `projects/task/vtt.yaml` to `projects/mfmmlm`.
+ - vtt.yaml
+ - vttqa.yaml
+ - youcook.yaml
+ - youcookcap.yaml
+ - crosstask.yaml
+ - coin.yaml
+ test: # section for testing.
+ task_list:
+ - test_vtt.yaml
+ - test_vttqa.yaml
+ - test_youcook.yaml
+ - test_youcookcap.yaml
+ - test_crosstask.yaml
+ - test_crosstask_zs.yaml
+ - test_coin.yaml
+```
diff --git a/fairseq/examples/MMPT/DATASET.md b/fairseq/examples/MMPT/DATASET.md
new file mode 100644
index 0000000000000000000000000000000000000000..930403eb3651a14dfcd015d094383faeb052c8ac
--- /dev/null
+++ b/fairseq/examples/MMPT/DATASET.md
@@ -0,0 +1,34 @@
+# Dataset
+
+We understand video data are challenging to download and process. For videos, we provide our preprocessing scripts under `scripts/video_feature_extractor` (deeply adapted from `https://github.com/antoine77340/video_feature_extractor`); for text, we pre-tokenizing scripts under `scripts/text_token_extractor`.
+
+### S3D Feature Extraction
+We use pre-trained [S3D](https://github.com/antoine77340/S3D_HowTo100M) for video feature extraction. Please place the models as `pretrained_models/s3d_dict.npy` and `pretrained_models/s3d_howto100m.pth`.
+
+We implement a `PathBuilder` to automatically track video ids, source video paths to their feature locations (you may need `conda install -c anaconda pandas`). Decoding may need `pip install ffmpeg-python`.
+
+### Howto100M
+[Howto100M](https://www.di.ens.fr/willow/research/howto100m/) is a large-scale video pre-training datasets. You may download videos by yourself and run preprocessing of our scripts.
+
+Several key differences of our preprocessing from existing papers: (1) we use `raw_caption.json` instead of `caption.json` to have pure self-supervision on text (`caption.json` has manual removal of stop words); (2) we remove partially duplicated texts that are originally designed for real-time readability (see `mmpt/processors/dedupprocessor.py`); (3) then we shard video/text features using `SharedTensor` in `mmpt/utils/shardedtensor.py` for fast loading during training (faster than `h5py`).
+
+#### Steps
+##### video
+To extract video features: edit and run `bash scripts/video_feature_extractor/how2/s3d.sh`. (consider to run this on multiple machines; by default, we store features in fp16 to save space and also for faster training).
+
+Split available video ids as `data/how2/how2_s3d_train.lst` and `data/how2/how2_s3d_val.lst`.
+
+Lastly, pack video features into `ShardedTensor` using `python scripts/video_feature_extractor/shard_feature.py`.
+
+##### text
+Clean captions using `python -m mmpt.processors.dedupprocessor`.
+
+Tokenize dedupped captions `data/how2/raw_caption_dedup.pkl` into sharded numpy arrays:
+```
+python scripts/text_token_extractor/pretokenization.py scripts/text_token_extractor/configs/bert-base-uncased.yaml
+```
+
+### Youcook, MSRVTT etc.
+We use the version of Youcook and MSRVTT come with Howto100M and MILNCE. Please download the data to `data/youcook` and `data/msrvtt` accordingly, you can also check `projects/task/youcook.yaml` and `projects/task/vtt.yaml` etc. in details.
+We extract features for Youcook, MSRVTT similar to the first step of Howto100M but we read text from meta data directly and perform on-the-fly tokenization.
+
diff --git a/fairseq/examples/MMPT/README.md b/fairseq/examples/MMPT/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..4a84819d9ddc992170296958a9431ff0538a85be
--- /dev/null
+++ b/fairseq/examples/MMPT/README.md
@@ -0,0 +1,166 @@
+# VideoCLIP and VLM
+
+You just find this toolkit for multimodal video understanding! It contains implementation of two recent multi-modal video understanding papers [VideoCLIP](https://arxiv.org/pdf/2109.14084.pdf) (EMNLP, 2021) and [VLM](https://aclanthology.org/2021.findings-acl.370.pdf) (ACL Findings, 2021), along with high-performance toolkits that are typically lacking in existing codebase. The toolkit is desigend to contain generic performance-tuned components that can be potentially adapted to other frameworks (we initially use fairseq).
+
+VideoCLIP is a contrastive learning model for zero-shot transfer to retrieval/classification/sequence labeling style tasks.
+
+
+
+VLM is a masked language model style pre-training using only one encoder with masked modality model (MMM) for retrieval/generation/sequence labeling style tasks.
+
+
+
+### News
+[Oct. 2021] Initial release of implementation for the following papers:
+[VideoCLIP: Contrastive Pre-training for Zero-shot Video-Text Understanding](https://arxiv.org/pdf/2109.14084.pdf) (Xu et. al., EMNLP 2021)
+[VLM: Task-agnostic Video-Language Model Pre-training for Video Understanding](https://aclanthology.org/2021.findings-acl.370.pdf) (Xu et. al., ACL Findings 2021)
+
+
+### Installation
+We aim to minimize the dependency of this repo on other packages.
+We use fairseq as the main trainer (no models/datasets dependency on fairseq. We will support other trainer in future):
+```
+git clone https://github.com/pytorch/fairseq
+cd fairseq
+pip install -e . # also optionally follow fairseq README for apex installation for fp16 training.
+export MKL_THREADING_LAYER=GNU # fairseq may need this for numpy.
+```
+
+Then install this toolkit:
+```
+cd examples/MMPT # MMPT can be in any folder, not necessarily under fairseq/examples.
+pip install -e .
+```
+
+The code is developed under Python=3.8.8, Pytorch=1.8, cuda=11.0 with fairseq=1.0.0a0+af0389f and tested under Python=3.8.8 pytorch=1.9 cuda=11.0 fairseq=1.0.0a0+8e7bc73 during code release.
+Most models require `transformers==3.4` for API compatibility `pip install transformers==3.4`.
+In addition, some downstream tasks may need `conda install pandas`.
+
+
+### Usage
+#### Download Checkpoints
+We use pre-trained [S3D](https://github.com/antoine77340/S3D_HowTo100M) for video feature extraction. Please place the models as `pretrained_models/s3d_dict.npy` and `pretrained_models/s3d_howto100m.pth`.
+
+Download VideoCLIP checkpoint `https://dl.fbaipublicfiles.com/MMPT/retri/videoclip/checkpoint_best.pt` to `runs/retri/videoclip` or VLM checkpoint `https://dl.fbaipublicfiles.com/MMPT/mtm/vlm/checkpoint_best.pt` to `runs/mtm/vlm`.
+
+#### Demo of Inference
+run `python locallaunch.py projects/retri/videoclip.yaml --dryrun` to get all `.yaml`s for VideoCLIP.
+
+```python
+import torch
+
+from mmpt.models import MMPTModel
+
+
+model, tokenizer, aligner = MMPTModel.from_pretrained(
+ "projects/retri/videoclip/how2.yaml")
+
+model.eval()
+
+
+# B, T, FPS, H, W, C (VideoCLIP is trained on 30 fps of s3d)
+video_frames = torch.randn(1, 2, 30, 224, 224, 3)
+caps, cmasks = aligner._build_text_seq(
+ tokenizer("some text", add_special_tokens=False)["input_ids"]
+)
+
+caps, cmasks = caps[None, :], cmasks[None, :] # bsz=1
+
+with torch.no_grad():
+ output = model(video_frames, caps, cmasks, return_score=True)
+print(output["score"]) # dot-product
+```
+
+#### Data Preparation
+See [dataset](DATASET.md) for each dataset.
+
+#### Global Config for Training Pipeline
+We organize a global config file for a training/testing pipeline under projects (see a detailed [explanation](CONFIG.md)). For example, VideoCLIP in `projects/retri/videoclip.yaml` and VLM is in `projects/mtm/vlm.yaml`.
+
+We wrap all cmds into `locallaunch.py` and `mmpt_cli/localjob.py`. You can check concrete cmds by `--dryrun` and then drop it for actual run.
+
+First, run `python locallaunch.py projects/retri/videoclip.yaml --dryrun` will generate configs for all configs of pre-training, zero-shot evaluation, fine-tuning and testing, for VideoCLIP under `projects/retri/videoclip`.
+
+Then each (either training or evaluation) process will be configed by a concrete config file (we save all complex arguments into the concrete config file for reproducibility, including fairseq args). For example, run zero-shot evaluation on youcook,
+```
+python locallaunch.py projects/retri/videoclip/test_youcook_zs.yaml --jobtype local_predict # zero-shot evaluation.
+python locallaunch.py projects/retri/videoclip/youcook_videoclip.yaml --jobtype local_single --dryrun # fine-tuning: use --dryrun to check cmds and drop it to make an actual run; local_small will run on two gpus (as in paper).
+python locallaunch.py projects/retri/videoclip/test_youcook_videoclip.yaml --jobtype local_predict # testing on fine-tuned model.
+```
+
+Pretraining can be run as:
+```
+python locallaunch.py projects/retri/videoclip/how2.yaml --jobtype local_single --dryrun # check then drop dryrun; paper is ran on local_big as 8 gpus.
+```
+You may need to change `--jobtype`, check/extend `LocalJob` in `mmpt_cli/localjob.py` for multi-gpu/multi-node pre-training.
+
+The detailed instructions of pretraining and fine-tuning can be found at [pretraining instruction](pretraining.md) and [finetuning instruction](endtask.md).
+
+
+### Development
+Several components of this toolkit can be re-used for future research (and also our ongoing research).
+
+#### Framework Wrapper
+We currently only support fairseq, but most components can be easily fit into other frameworks like huggingface. This repo is a `--user-dir` of fairseq with fairseq wrapper. For example, `mmpt/tasks` includes a `FairseqMMTTask`, which manages `mmpt/datasets` with `FairseqDataset`, `mmpt/models` with `FairseqModel`, `mmpt/losses` with `FairseqCriterion`.
+
+#### Processors
+**Multi**modal research introduces the complexity on modality alignment from different input sources to losses. Inspired by [MMF](https://github.com/facebookresearch/mmf), this toolkit leverages `mmpt/processors` to handle various needs of data preprocessing and loading, **alleviating** the needs of multiple `torch.data.utils.Dataset` (that can be tricky for ablation study).
+Processors can also be decoupled from `torch.data.utils.Dataset` for offline preprocessing instead of on-the-fly data preprocessing.
+
+We decouple a `mmpt.MMDataset` as 3 types of processors: `MetaProcessor`, `VideoProcessor`, `TextProcessor` and `Aligner`. They can be configed in `dataset` field of a config file (e.g., see `projects/task/how2.yaml`).
+`MetaProcessor` is used to load the meta data about a dataset, aka, all video_ids of how2 dataset.
+`VideoProcessor` is used to load the video features about a dataset. For example, S3D features for each second of a video.
+`TextProcessor` is used to load the text (feature). For example, BERT pre-tokenized text clips for how2 dataset (with `start`s, `end`s of timestamps and `cap` for `token_ids`).
+`Aligner` is the core class for different baselines that prepares the training data. For example, sampling a clip, masking tokens for MLM, etc.
+
+#### Performance-tuned Components
+To speed up pre-training, this toolkit uses sharded features stored in mmaped numpy, backed by `ShardedTensor` in `mmpt/utils/shardedtensor.py` (adopted from MARGE paper). This reduces the loads of IO for multi-GPU training without loading all features for a video into the memory each time and `ShardedTensor` ensure features are stored in continuous disk space for near random access. This is used for both How2 video features and texts in `mmpt/processors/how2processor.py`.
+
+
+### Citation
+If this codebase is useful for your work, please cite the following papers:
+
+```BibTeX
+@inproceedings{xu-etal-2021-videoclip,
+ title = "{VideoCLIP}: Contrastive Pre-training for\\Zero-shot Video-Text Understanding",
+ author = "Xu, Hu and
+ Ghosh, Gargi and
+ Huang, Po-Yao and
+ Okhonko, Dmytro and
+ Aghajanyan, Armen and
+ Metze, Florian and
+ Zettlemoyer, Luke and
+ Feichtenhofer, Christoph",
+ booktitle = "Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
+ month = nov,
+ year = "2021",
+ address = "Online",
+ publisher = "Association for Computational Linguistics",
+}
+
+@inproceedings{xu-etal-2021-vlm,
+ title = "{VLM}: Task-agnostic Video-Language Model Pre-training for Video Understanding",
+ author = "Xu, Hu and
+ Ghosh, Gargi and
+ Huang, Po-Yao and
+ Arora, Prahal and
+ Aminzadeh, Masoumeh and
+ Feichtenhofer, Christoph and
+ Metze, Florian and
+ Zettlemoyer, Luke",
+ booktitle = "Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021",
+ month = aug,
+ year = "2021",
+ address = "Online",
+ publisher = "Association for Computational Linguistics",
+ url = "https://aclanthology.org/2021.findings-acl.370",
+ doi = "10.18653/v1/2021.findings-acl.370",
+ pages = "4227--4239",
+}
+```
+
+### Bug Reports
+This repo is in its initial stage, welcome bug reports to huxu@fb.com
+
+### Copyright
+The majority of Multimodal Pre-training (MMPT) is licensed under CC-BY-NC, however portions of the project are available under separate license terms: Evaluation Codes/Models: Howto100M and HuggingFace Transformers are licensed under the Apache2.0 license; COIN and NLG-eval are licensed under the MIT license; CrossTask is licensed under the BSD-3; DiDeMo is licensed under the BSD-2 license.
diff --git a/fairseq/examples/MMPT/endtask.md b/fairseq/examples/MMPT/endtask.md
new file mode 100644
index 0000000000000000000000000000000000000000..7690955327283ded4b37857a4a7b78463e0eb0f8
--- /dev/null
+++ b/fairseq/examples/MMPT/endtask.md
@@ -0,0 +1,41 @@
+# Zero-shot Transfer and Finetuning
+
+(If you are new to the ideas of `mmpt.processors`, see [README](README.md) first.)
+All finetuning datasets (specifically `processors`) are defined in `mmpt.processors.dsprocessor`.
+Given the complexity of different types of finetuning tasks, each task may have their own meta/video/text/aligner processors and `mmpt/evaluators/{Predictor,Metric}`.
+
+### Tasks
+
+Currently, we support 5 end datasets: `MSRVTT`, `Youcook`, `COIN`, `Crosstask` and `DiDeMo` with the following tasks:
+text-video retrieval: `MSRVTT`, `Youcook`, `DiDeMo`;
+video captioning: `Youcook`;
+Video Question and Answering: `MSRVTT-QA`.
+
+To add your own dataset, you can specify the corresponding processors and config them in the `dataset` field of a config file, such as `projects/task/vtt.yaml`.
+
+### Zero-shot Transfer (no Training)
+Zero-shot transfer will run the pre-trained model (e.g., VideoCLIP) directly on testing data. Configs with pattern: `projects/task/*_zs_*.yaml` are dedicated for zero-shot transfer.
+
+### Fine-tuning
+
+The training of a downstream task is similar to pretraining, execept you may need to specify the `restore_file` in `fairseq.checkpoint` and reset optimizers, see `projects/task/ft.yaml` that is included by `projects/task/vtt.yaml`.
+
+We typically do finetuning on 2 gpus (`local_small`).
+
+### Testing
+For each finetuning dataset, you may need to specify a testing config, similar to `projects/task/test_vtt.yaml`.
+
+We define `mmpt.evaluators.Predictor` for different types of prediction. For example, `MSRVTT` and `Youcook` are video-retrieval tasks and expecting to use `RetrievalPredictor`. You may need to define your new type of predictors and specify that in `predictor` field of a testing config.
+
+Each task may also have their own metric for evaluation. This can be created in `mmpt.evaluators.Metric` and specified in the `metric` field of a testing config.
+
+Launching a testing is as simple as training by specifying the path of a testing config:
+```python locallaunch.py projects/mfmmlm/test_vtt.yaml```
+Testing will be launched locally by default since prediction is computationally less expensive.
+
+### Third-party Libraries
+We list the following finetuning tasks that require third-party libraries.
+
+Youcook captioning: `https://github.com/Maluuba/nlg-eval`
+
+CrossTask: `https://github.com/DmZhukov/CrossTask`'s `dp` under `third-party/CrossTask` (`python setup.py build_ext --inplace`)
diff --git a/fairseq/examples/MMPT/locallaunch.py b/fairseq/examples/MMPT/locallaunch.py
new file mode 100644
index 0000000000000000000000000000000000000000..e20fd816fa3bed8b1af8f6a4d1a07ccb69a1fffa
--- /dev/null
+++ b/fairseq/examples/MMPT/locallaunch.py
@@ -0,0 +1,148 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+import argparse
+import os
+
+from omegaconf import OmegaConf
+
+from mmpt.utils import recursive_config, overwrite_dir
+from mmpt_cli.localjob import LocalJob
+
+
+class JobLauncher(object):
+ JOB_CONFIG = {
+ "local": LocalJob,
+ }
+
+ def __init__(self, yaml_file):
+ self.yaml_file = yaml_file
+ job_key = "local"
+
+ if yaml_file.endswith(".yaml"):
+ config = recursive_config(yaml_file)
+ if config.task_type is not None:
+ job_key = config.task_type.split("_")[0]
+ else:
+ raise ValueError("unknown extension of job file:", yaml_file)
+ self.job_key = job_key
+
+ def __call__(self, job_type=None, dryrun=False):
+ if job_type is not None:
+ self.job_key = job_type.split("_")[0]
+ print("[JobLauncher] job_key", self.job_key)
+ job = JobLauncher.JOB_CONFIG[self.job_key](
+ self.yaml_file, job_type=job_type, dryrun=dryrun)
+ return job.submit()
+
+
+class Pipeline(object):
+ """a job that loads yaml config."""
+
+ def __init__(self, fn):
+ """
+ load a yaml config of a job and save generated configs as yaml for each task.
+ return: a list of files to run as specified by `run_task`.
+ """
+ if fn.endswith(".py"):
+ # a python command.
+ self.backend = "python"
+ self.run_yamls = [fn]
+ return
+
+ job_config = recursive_config(fn)
+ if job_config.base_dir is None: # single file job config.
+ self.run_yamls = [fn]
+ return
+
+ self.project_dir = os.path.join("projects", job_config.project_dir)
+ self.run_dir = os.path.join("runs", job_config.project_dir)
+
+ if job_config.run_task is not None:
+ run_yamls = []
+ for stage in job_config.run_task:
+ # each stage can have multiple tasks running in parallel.
+ if OmegaConf.is_list(stage):
+ stage_yamls = []
+ for task_file in stage:
+ stage_yamls.append(
+ os.path.join(self.project_dir, task_file))
+ run_yamls.append(stage_yamls)
+ else:
+ run_yamls.append(os.path.join(self.project_dir, stage))
+ self.run_yamls = run_yamls
+ configs_to_save = self._overwrite_task(job_config)
+ self._save_configs(configs_to_save)
+
+ def __getitem__(self, idx):
+ yaml_files = self.run_yamls[idx]
+ if isinstance(yaml_files, list):
+ return [JobLauncher(yaml_file) for yaml_file in yaml_files]
+ return [JobLauncher(yaml_files)]
+
+ def __len__(self):
+ return len(self.run_yamls)
+
+ def _save_configs(self, configs_to_save: dict):
+ # save
+ os.makedirs(self.project_dir, exist_ok=True)
+ for config_file in configs_to_save:
+ config = configs_to_save[config_file]
+ print("saving", config_file)
+ OmegaConf.save(config=config, f=config_file)
+
+ def _overwrite_task(self, job_config):
+ configs_to_save = {}
+ self.base_project_dir = os.path.join("projects", job_config.base_dir)
+ self.base_run_dir = os.path.join("runs", job_config.base_dir)
+
+ for config_sets in job_config.task_group:
+ overwrite_config = job_config.task_group[config_sets]
+ if (
+ overwrite_config.task_list is None
+ or len(overwrite_config.task_list) == 0
+ ):
+ print(
+ "[warning]",
+ job_config.task_group,
+ "has no task_list specified.")
+ # we don't want this added to a final config.
+ task_list = overwrite_config.pop("task_list", None)
+ for config_file in task_list:
+ config_file_path = os.path.join(
+ self.base_project_dir, config_file)
+ config = recursive_config(config_file_path)
+ # overwrite it.
+ if overwrite_config:
+ config = OmegaConf.merge(config, overwrite_config)
+ overwrite_dir(config, self.run_dir, basedir=self.base_run_dir)
+ save_file_path = os.path.join(self.project_dir, config_file)
+ configs_to_save[save_file_path] = config
+ return configs_to_save
+
+
+def main(args):
+ job_type = args.jobtype if args.jobtype else None
+ # parse multiple pipelines.
+ pipelines = [Pipeline(fn) for fn in args.yamls.split(",")]
+
+ for pipe_id, pipeline in enumerate(pipelines):
+ if not hasattr(pipeline, "project_dir"):
+ for job in pipeline[0]:
+ job(job_type=job_type, dryrun=args.dryrun)
+
+
+if __name__ == "__main__":
+ parser = argparse.ArgumentParser()
+ parser.add_argument("yamls", type=str)
+ parser.add_argument(
+ "--dryrun",
+ action="store_true",
+ help="run config and prepare to submit without launch the job.",
+ )
+ parser.add_argument(
+ "--jobtype", type=str, default="",
+ help="force to run jobs as specified.")
+ args = parser.parse_args()
+ main(args)
diff --git a/fairseq/examples/MMPT/mmpt/__init__.py b/fairseq/examples/MMPT/mmpt/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..6ff86ddd5ce0c454281e6568c628bd3c49ea5024
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/__init__.py
@@ -0,0 +1,12 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+try:
+ # fairseq user dir
+ from .datasets import FairseqMMDataset
+ from .losses import FairseqCriterion
+ from .models import FairseqMMModel
+ from .tasks import FairseqMMTask
+except ImportError:
+ pass
diff --git a/fairseq/examples/MMPT/mmpt/datasets/__init__.py b/fairseq/examples/MMPT/mmpt/datasets/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..2578235e1771fdc7e6fcfb66a519cbe891d7e254
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/datasets/__init__.py
@@ -0,0 +1,10 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+from .mmdataset import *
+
+try:
+ from .fairseqmmdataset import *
+except ImportError:
+ pass
diff --git a/fairseq/examples/MMPT/mmpt/datasets/fairseqmmdataset.py b/fairseq/examples/MMPT/mmpt/datasets/fairseqmmdataset.py
new file mode 100644
index 0000000000000000000000000000000000000000..02c49141db69c44663bd438b947c268d06f8aa2b
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/datasets/fairseqmmdataset.py
@@ -0,0 +1,57 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+"""
+TODO (huxu): fairseq wrapper class for all dataset you defined: mostly MMDataset.
+"""
+
+from collections import OrderedDict
+
+from torch.utils.data import Dataset
+from torch.utils.data.dataloader import default_collate
+from fairseq.data import FairseqDataset, data_utils
+
+
+class FairseqMMDataset(FairseqDataset):
+ """
+ A wrapper class for MMDataset for fairseq.
+ """
+
+ def __init__(self, mmdataset):
+ if not isinstance(mmdataset, Dataset):
+ raise TypeError("mmdataset must be of type `torch.utils.data.dataset`.")
+ self.mmdataset = mmdataset
+
+ def set_epoch(self, epoch, **unused):
+ super().set_epoch(epoch)
+ self.epoch = epoch
+
+ def __getitem__(self, idx):
+ with data_utils.numpy_seed(43211, self.epoch, idx):
+ return self.mmdataset[idx]
+
+ def __len__(self):
+ return len(self.mmdataset)
+
+ def collater(self, samples):
+ if hasattr(self.mmdataset, "collator"):
+ return self.mmdataset.collator(samples)
+ if len(samples) == 0:
+ return {}
+ if isinstance(samples[0], dict):
+ batch = OrderedDict()
+ for key in samples[0]:
+ if samples[0][key] is not None:
+ batch[key] = default_collate([sample[key] for sample in samples])
+ return batch
+ else:
+ return default_collate(samples)
+
+ def size(self, index):
+ """dummy implementation: we don't use --max-tokens"""
+ return 1
+
+ def num_tokens(self, index):
+ """dummy implementation: we don't use --max-tokens"""
+ return 1
diff --git a/fairseq/examples/MMPT/mmpt/datasets/mmdataset.py b/fairseq/examples/MMPT/mmpt/datasets/mmdataset.py
new file mode 100644
index 0000000000000000000000000000000000000000..3d07283f917a430a8d9b1226c8fa6ab71450e8a9
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/datasets/mmdataset.py
@@ -0,0 +1,111 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+
+from collections import OrderedDict
+
+from torch.utils.data import Dataset
+from torch.utils.data.dataloader import default_collate
+
+from ..utils import set_seed
+
+
+class MMDataset(Dataset):
+ """
+ A generic multi-modal dataset.
+ Args:
+ `meta_processor`: a meta processor,
+ handling loading meta data and return video_id and text_id.
+ `video_processor`: a video processor,
+ handling e.g., decoding, loading .np files.
+ `text_processor`: a text processor,
+ handling e.g., tokenization.
+ `aligner`: combine the video and text feature
+ as one training example.
+ """
+
+ def __init__(
+ self,
+ meta_processor,
+ video_processor,
+ text_processor,
+ align_processor,
+ ):
+ self.split = meta_processor.split
+ self.meta_processor = meta_processor
+ self.video_processor = video_processor
+ self.text_processor = text_processor
+ self.align_processor = align_processor
+
+ def __len__(self):
+ return len(self.meta_processor)
+
+ def __getitem__(self, idx):
+ if self.split == "test":
+ set_seed(idx)
+ video_id, text_id = self.meta_processor[idx]
+ video_feature = self.video_processor(video_id)
+ text_feature = self.text_processor(text_id)
+ output = self.align_processor(video_id, video_feature, text_feature)
+ # TODO (huxu): the following is for debug purpose.
+ output.update({"idx": idx})
+ return output
+
+ def collater(self, samples):
+ """This collator is deprecated.
+ set self.collator = MMDataset.collater.
+ see collator in FairseqMMDataset.
+ """
+
+ if len(samples) == 0:
+ return {}
+ if isinstance(samples[0], dict):
+ batch = OrderedDict()
+ for key in samples[0]:
+ if samples[0][key] is not None:
+ batch[key] = default_collate(
+ [sample[key] for sample in samples])
+ # if torch.is_tensor(batch[key]):
+ # print(key, batch[key].size())
+ # else:
+ # print(key, len(batch[key]))
+ return batch
+ else:
+ return default_collate(samples)
+
+ def print_example(self, output):
+ print("[one example]", output["video_id"])
+ if (
+ hasattr(self.align_processor, "subsampling")
+ and self.align_processor.subsampling is not None
+ and self.align_processor.subsampling > 1
+ ):
+ for key in output:
+ if torch.is_tensor(output[key]):
+ output[key] = output[key][0]
+
+ # search tokenizer to translate ids back.
+ tokenizer = None
+ if hasattr(self.text_processor, "tokenizer"):
+ tokenizer = self.text_processor.tokenizer
+ elif hasattr(self.align_processor, "tokenizer"):
+ tokenizer = self.align_processor.tokenizer
+ if tokenizer is not None:
+ caps = output["caps"].tolist()
+ if isinstance(caps[0], list):
+ caps = caps[0]
+ print("caps", tokenizer.decode(caps))
+ print("caps", tokenizer.convert_ids_to_tokens(caps))
+
+ for key, value in output.items():
+ if torch.is_tensor(value):
+ if len(value.size()) >= 3: # attention_mask.
+ print(key, value.size())
+ print(key, "first", value[0, :, :])
+ print(key, "last", value[-1, :, :])
+ else:
+ print(key, value)
+ print("[end of one example]")
diff --git a/fairseq/examples/MMPT/mmpt/evaluators/__init__.py b/fairseq/examples/MMPT/mmpt/evaluators/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..2d06b9d7974db251143025124468f48cd230e89a
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/evaluators/__init__.py
@@ -0,0 +1,13 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+from .metric import *
+from .evaluator import *
+
+
+# experimental.
+try:
+ from .expmetric import *
+except ImportError:
+ pass
diff --git a/fairseq/examples/MMPT/mmpt/evaluators/evaluator.py b/fairseq/examples/MMPT/mmpt/evaluators/evaluator.py
new file mode 100644
index 0000000000000000000000000000000000000000..94d9c5ec9a6e84434dbced8d5647754c5a571570
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/evaluators/evaluator.py
@@ -0,0 +1,54 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+import os
+import glob
+import numpy as np
+
+from . import metric as metric_path
+from . import predictor as predictor_path
+
+
+class Evaluator(object):
+ """
+ perform evaluation on a single (downstream) task.
+ make this both offline and online.
+ TODO(huxu) saving evaluation results.
+ """
+
+ def __init__(self, config, eval_dataloader=None):
+ if config.metric is None:
+ raise ValueError("config.metric is", config.metric)
+ metric_cls = getattr(metric_path, config.metric)
+ self.metric = metric_cls(config)
+ if config.predictor is None:
+ raise ValueError("config.predictor is", config.predictor)
+ predictor_cls = getattr(predictor_path, config.predictor)
+ self.predictor = predictor_cls(config)
+ self.eval_dataloader = eval_dataloader
+
+ def __call__(self):
+ try:
+ print(self.predictor.pred_dir)
+ for pred_file in glob.glob(
+ self.predictor.pred_dir + "/*_merged.npy"):
+ outputs = np.load(pred_file)
+ results = self.metric.compute_metrics(outputs)
+ self.metric.print_computed_metrics(results)
+
+ outputs = np.load(os.path.join(
+ self.predictor.pred_dir, "merged.npy"))
+ results = self.metric.compute_metrics(outputs)
+ return {"results": results, "metric": self.metric}
+ except FileNotFoundError:
+ print("\n[missing]", self.predictor.pred_dir)
+ return {}
+
+ def evaluate(self, model, eval_dataloader=None, output_file="merged"):
+ if eval_dataloader is None:
+ eval_dataloader = self.eval_dataloader
+ outputs = self.predictor.predict_loop(
+ model, eval_dataloader, output_file)
+ results = self.metric.compute_metrics(**outputs)
+ return results
diff --git a/fairseq/examples/MMPT/mmpt/evaluators/metric.py b/fairseq/examples/MMPT/mmpt/evaluators/metric.py
new file mode 100644
index 0000000000000000000000000000000000000000..163724bb250cb1b7057b3fa4d75a9fafa9c181f5
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/evaluators/metric.py
@@ -0,0 +1,313 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import numpy as np
+import json
+
+
+class Metric(object):
+ def __init__(self, config, metric_names):
+ self.metric_names = metric_names
+
+ def best_metric(self, metric):
+ return metric[self.metric_names[0]]
+
+ def save_metrics(self, fn, metrics):
+ with open(fn, "w") as fw:
+ json.dump(fw, metrics)
+
+ def print_computed_metrics(self, metrics):
+ raise NotImplementedError
+
+
+class RetrievalMetric(Metric):
+ """
+ this is modified from `howto100m/metrics.py`.
+ History of changes:
+ refactor as a class.
+ add metric_key in __init__
+ """
+
+ def __init__(self, config, metric_names=["R1", "R5", "R10", "MR"]):
+ super().__init__(config, metric_names)
+ self.error = False # TODO(huxu): add to config to print error.
+
+ def compute_metrics(self, outputs, texts, **kwargs):
+ x = outputs
+ sx = np.sort(-x, axis=1)
+ d = np.diag(-x)
+ d = d[:, np.newaxis]
+ ind = sx - d
+ ind = np.where(ind == 0)
+ ind = ind[1]
+ metrics = {}
+ metrics["R1"] = float(np.sum(ind == 0)) / len(ind)
+ metrics["R5"] = float(np.sum(ind < 5)) / len(ind)
+ metrics["R10"] = float(np.sum(ind < 10)) / len(ind)
+ metrics["MR"] = np.median(ind) + 1
+
+ max_idx = np.argmax(outputs, axis=1)
+ if self.error:
+ # print top-20 errors.
+ error = []
+ for ex_idx in range(20):
+ error.append((texts[ex_idx], texts[max_idx[ex_idx]]))
+ metrics["error"] = error
+ return metrics
+
+ def print_computed_metrics(self, metrics):
+ r1 = metrics["R1"]
+ r5 = metrics["R5"]
+ r10 = metrics["R10"]
+ mr = metrics["MR"]
+ print(
+ "R@1: {:.4f} - R@5: {:.4f} - R@10: {:.4f} - Median R: {}".format(
+ r1, r5, r10, mr
+ )
+ )
+ if "error" in metrics:
+ print(metrics["error"])
+
+
+class DiDeMoMetric(Metric):
+ """
+ History of changes:
+ python 2.x to python 3.x.
+ merge utils.py into eval to save one file.
+ reference: https://github.com/LisaAnne/LocalizingMoments/blob/master/utils/eval.py
+ Code to evaluate your results on the DiDeMo dataset.
+ """
+ def __init__(self, config, metric_names=["rank1", "rank5", "miou"]):
+ super().__init__(config, metric_names)
+
+ def compute_metrics(self, outputs, targets, **kwargs):
+ assert len(outputs) == len(targets)
+ rank1, rank5, miou = self._eval_predictions(outputs, targets)
+ metrics = {
+ "rank1": rank1,
+ "rank5": rank5,
+ "miou": miou
+ }
+ return metrics
+
+ def print_computed_metrics(self, metrics):
+ rank1 = metrics["rank1"]
+ rank5 = metrics["rank5"]
+ miou = metrics["miou"]
+ # print("Average rank@1: %f" % rank1)
+ # print("Average rank@5: %f" % rank5)
+ # print("Average iou: %f" % miou)
+
+ print(
+ "Average rank@1: {:.4f} Average rank@5: {:.4f} Average iou: {:.4f}".format(
+ rank1, rank5, miou
+ )
+ )
+
+ def _iou(self, pred, gt):
+ intersection = max(0, min(pred[1], gt[1]) + 1 - max(pred[0], gt[0]))
+ union = max(pred[1], gt[1]) + 1 - min(pred[0], gt[0])
+ return float(intersection)/union
+
+ def _rank(self, pred, gt):
+ return pred.index(tuple(gt)) + 1
+
+ def _eval_predictions(self, segments, data):
+ '''
+ Inputs:
+ segments: For each item in the ground truth data, rank possible video segments given the description and video.
+ In DiDeMo, there are 21 posible moments extracted for each video so the list of video segments will be of length 21.
+ The first video segment should be the video segment that best corresponds to the text query.
+ There are 4180 sentence in the validation data, so when evaluating a model on the val dataset,
+ segments should be a list of lenght 4180, and each item in segments should be a list of length 21.
+ data: ground truth data
+ '''
+ average_ranks = []
+ average_iou = []
+ for s, d in zip(segments, data):
+ pred = s[0]
+ ious = [self._iou(pred, t) for t in d['times']]
+ average_iou.append(np.mean(np.sort(ious)[-3:]))
+ ranks = [self._rank(s, t) for t in d['times'] if tuple(t) in s] # if t in s] is added for s, e not in prediction.
+ average_ranks.append(np.mean(np.sort(ranks)[:3]))
+ rank1 = np.sum(np.array(average_ranks) <= 1)/float(len(average_ranks))
+ rank5 = np.sum(np.array(average_ranks) <= 5)/float(len(average_ranks))
+ miou = np.mean(average_iou)
+
+ # print("Average rank@1: %f" % rank1)
+ # print("Average rank@5: %f" % rank5)
+ # print("Average iou: %f" % miou)
+ return rank1, rank5, miou
+
+
+class NLGMetric(Metric):
+ def __init__(
+ self,
+ config,
+ metric_names=[
+ "Bleu_1", "Bleu_2", "Bleu_3", "Bleu_4",
+ "METEOR", "ROUGE_L", "CIDEr"
+ ]
+ ):
+ super().__init__(config, metric_names)
+ # please install NLGEval from `https://github.com/Maluuba/nlg-eval`
+ from nlgeval import NLGEval
+ self.nlg = NLGEval()
+
+ def compute_metrics(self, outputs, targets, **kwargs):
+ return self.nlg.compute_metrics(
+ hyp_list=outputs, ref_list=targets)
+
+ def print_computed_metrics(self, metrics):
+ Bleu_1 = metrics["Bleu_1"]
+ Bleu_2 = metrics["Bleu_2"]
+ Bleu_3 = metrics["Bleu_3"]
+ Bleu_4 = metrics["Bleu_4"]
+ METEOR = metrics["METEOR"]
+ ROUGE_L = metrics["ROUGE_L"]
+ CIDEr = metrics["CIDEr"]
+
+ print(
+ "Bleu_1: {:.4f} - Bleu_2: {:.4f} - Bleu_3: {:.4f} - Bleu_4: {:.4f} - METEOR: {:.4f} - ROUGE_L: {:.4f} - CIDEr: {:.4f}".format(
+ Bleu_1, Bleu_2, Bleu_3, Bleu_4, METEOR, ROUGE_L, CIDEr
+ )
+ )
+
+
+class QAMetric(Metric):
+ def __init__(
+ self,
+ config,
+ metric_names=["acc"]
+ ):
+ super().__init__(config, metric_names)
+
+ def compute_metrics(self, outputs, targets, **kwargs):
+ from sklearn.metrics import accuracy_score
+ return {"acc": accuracy_score(targets, outputs)}
+
+ def print_computed_metrics(self, metrics):
+ print("acc: {:.4f}".format(metrics["acc"]))
+
+
+class COINActionSegmentationMetric(Metric):
+ """
+ COIN dataset listed 3 repos for Action Segmentation.
+ Action Sets, NeuralNetwork-Viterbi, TCFPN-ISBA.
+ The first and second are the same.
+ https://github.com/alexanderrichard/action-sets/blob/master/eval.py
+
+ Future reference for the third:
+ `https://github.com/Zephyr-D/TCFPN-ISBA/blob/master/utils/metrics.py`
+ """
+ def __init__(self, config, metric_name=["frame_acc"]):
+ super().__init__(config, metric_name)
+
+ def compute_metrics(self, outputs, targets):
+ n_frames = 0
+ n_errors = 0
+ n_errors = sum(outputs != targets)
+ n_frames = len(targets)
+ return {"frame_acc": 1.0 - float(n_errors) / n_frames}
+
+ def print_computed_metrics(self, metrics):
+ fa = metrics["frame_acc"]
+ print("frame accuracy:", fa)
+
+
+class CrossTaskMetric(Metric):
+ def __init__(self, config, metric_names=["recall"]):
+ super().__init__(config, metric_names)
+
+ def compute_metrics(self, outputs, targets, **kwargs):
+ """refactored from line 166:
+ https://github.com/DmZhukov/CrossTask/blob/master/train.py"""
+
+ recalls = self._get_recalls(Y_true=targets, Y_pred=outputs)
+ results = {}
+ for task, rec in recalls.items():
+ results[str(task)] = rec
+
+ avg_recall = np.mean(list(recalls.values()))
+ results["recall"] = avg_recall
+ return results
+
+ def print_computed_metrics(self, metrics):
+ print('Recall: {0:0.3f}'.format(metrics["recall"]))
+ for task in metrics:
+ if task != "recall":
+ print('Task {0}. Recall = {1:0.3f}'.format(
+ task, metrics[task]))
+
+ def _get_recalls(self, Y_true, Y_pred):
+ """refactored from
+ https://github.com/DmZhukov/CrossTask/blob/master/train.py"""
+
+ step_match = {task: 0 for task in Y_true.keys()}
+ step_total = {task: 0 for task in Y_true.keys()}
+ for task, ys_true in Y_true.items():
+ ys_pred = Y_pred[task]
+ for vid in set(ys_pred.keys()).intersection(set(ys_true.keys())):
+ y_true = ys_true[vid]
+ y_pred = ys_pred[vid]
+ step_total[task] += (y_true.sum(axis=0) > 0).sum()
+ step_match[task] += (y_true*y_pred).sum()
+ recalls = {
+ task: step_match[task] / n for task, n in step_total.items()}
+ return recalls
+
+
+class ActionRecognitionMetric(Metric):
+ def __init__(
+ self,
+ config,
+ metric_names=["acc", "acc_splits", "r1_splits", "r5_splits", "r10_splits"]
+ ):
+ super().__init__(config, metric_names)
+
+ def compute_metrics(self, outputs, targets, splits, **kwargs):
+ all_video_embd = outputs
+ labels = targets
+ split1, split2, split3 = splits
+ accs = []
+ r1s = []
+ r5s = []
+ r10s = []
+ for split in range(3):
+ if split == 0:
+ s = split1
+ elif split == 1:
+ s = split2
+ else:
+ s = split3
+
+ X_pred = all_video_embd[np.where(s == 2)[0]]
+ label_test = labels[np.where(s == 2)[0]]
+ logits = X_pred
+ X_pred = np.argmax(X_pred, axis=1)
+ acc = np.sum(X_pred == label_test) / float(len(X_pred))
+ accs.append(acc)
+ # compute recall.
+ sorted_pred = (-logits).argsort(axis=-1)
+ label_test_sp = label_test.reshape(-1, 1)
+
+ r1 = np.mean((sorted_pred[:, :1] == label_test_sp).sum(axis=1), axis=0)
+ r5 = np.mean((sorted_pred[:, :5] == label_test_sp).sum(axis=1), axis=0)
+ r10 = np.mean((sorted_pred[:, :10] == label_test_sp).sum(axis=1), axis=0)
+ r1s.append(r1)
+ r5s.append(r5)
+ r10s.append(r10)
+
+ return {"acc": accs[0], "acc_splits": accs, "r1_splits": r1s, "r5_splits": r5s, "r10_splits": r10s}
+
+ def print_computed_metrics(self, metrics):
+ for split, acc in enumerate(metrics["acc_splits"]):
+ print("Top 1 accuracy on split {}: {}; r1 {}; r5 {}; r10 {}".format(
+ split + 1, acc,
+ metrics["r1_splits"][split],
+ metrics["r5_splits"][split],
+ metrics["r10_splits"][split],
+ )
+ )
diff --git a/fairseq/examples/MMPT/mmpt/evaluators/predictor.py b/fairseq/examples/MMPT/mmpt/evaluators/predictor.py
new file mode 100644
index 0000000000000000000000000000000000000000..2ffef6ab474a7a275d90784250d04222bd6dc70f
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/evaluators/predictor.py
@@ -0,0 +1,595 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+import os
+import random
+import json
+import numpy as np
+import torch
+import pickle
+import math
+
+from tqdm import tqdm
+
+
+class Predictor(object):
+ """this base class is used to save predictions to disk
+ (and being called by a evaluator later).
+ Predictor has minimum support of single gpu prediction.
+ """
+ def __init__(self, config):
+ self.pred_dir = None # on-the-fly eval does not save the results.
+ if hasattr(config, "eval") and config.eval is not None:
+ self.pred_dir = config.eval.save_path
+ os.makedirs(self.pred_dir, exist_ok=True)
+
+ def __call__(self, outputs):
+ """extract the prediction and save it."""
+ raise NotImplementedError
+
+ def predict_loop(self, model, eval_dataloader, output_file=None):
+ """on-the-fly prediction on a single gpu."""
+ self.full_scores = []
+ model.eval()
+ model = model.to(0)
+ with torch.no_grad():
+ for data in eval_dataloader:
+ data = self.to_ctx(data)
+ outputs = model(**data)
+ outputs.update(data)
+ self(outputs)
+ return self.finalize(output_file)
+
+ def finalize(self, output_file):
+ pass
+
+ def to_ctx(self, data, ctx=0, dtype=None):
+ if isinstance(data, dict):
+ for key in data:
+ if torch.is_tensor(data[key]):
+ if dtype is not None and data[key].dtype == torch.float32:
+ data[key] = data[key].to(dtype)
+ data[key] = data[key].to(ctx)
+ return data
+ else:
+ raise ValueError("non-dict type of batch is not supported yet.")
+
+
+class NLGPredictor(Predictor):
+ """Predicting Text from MMFusion models."""
+ """TODO: make a context."""
+ def __init__(self, config):
+ super().__init__(config)
+ from transformers import AutoTokenizer
+
+ self.tokenizer = AutoTokenizer.from_pretrained(
+ config.dataset.bert_name,
+ bos_token="[CLS]", eos_token="[SEP]")
+ self.bos_token_id = self.tokenizer.bos_token_id
+ self.eos_token_id = self.tokenizer.eos_token_id
+
+ def predict_loop(self, model, eval_dataloader, output_file=None):
+ """TODO: refactor base classes."""
+ ctx = 0
+ outputs = {"outputs": [], "targets": [[]]}
+ model.eval()
+ model = model.to(ctx)
+ with torch.no_grad():
+ for data in tqdm(eval_dataloader):
+ data = self.to_ctx(data, ctx)
+ self(data, model, outputs)
+ return self.finalize(outputs, output_file)
+
+ def __call__(self, data, model, outputs):
+ data.update({
+ "bos_token_id": self.bos_token_id,
+ "eos_token_id": self.eos_token_id
+ })
+
+ output = model.generate(**data)
+ assert len(output) == len(data["ref"])
+ for idx, _output in enumerate(output):
+ generated_text = self.tokenizer.decode(
+ _output, skip_special_tokens=True)
+ if generated_text == "":
+ generated_text = "none"
+ outputs["outputs"].append(generated_text)
+ outputs["targets"][0].append(data["ref"][idx])
+ if random.random() < 0.001:
+ print("_output", _output)
+ print("generated_text", generated_text)
+ print("ref", data["ref"][idx])
+
+ def finalize(self, outputs, output_file=None):
+ if output_file is not None:
+ with open(os.path.join(
+ self.pred_dir, output_file + ".json"), "w") as fw:
+ json.dump(outputs, fw, indent=4)
+ return outputs
+
+
+class RetrievalPredictor(Predictor):
+ """generated `pooled_video` and `pooled_text`."""
+ def __init__(self, config):
+ super().__init__(config)
+ from transformers import AutoTokenizer
+ self.tokenizer = AutoTokenizer.from_pretrained(
+ config.dataset.bert_name)
+
+ def predict_loop(
+ self,
+ model,
+ eval_dataloader,
+ output_file="retrieval.npy"
+ ):
+ """on-the-fly prediction on a single gpu."""
+ full_scores = []
+ texts = []
+ model.eval()
+ model = model.cuda()
+ with torch.no_grad():
+ for data in eval_dataloader:
+ # convert to dict.
+ if not isinstance(data, dict):
+ data = {
+ "caps": data[0],
+ "cmasks": data[1],
+ "vfeats": data[2],
+ "vmasks": data[3],
+ "video_id": data[4]
+ }
+ data = self.to_ctx(data)
+ outputs = model(**data)
+ outputs.update(data)
+ self(outputs, full_scores)
+ for _cap in data["caps"]:
+ texts.append(
+ self.tokenizer.decode(_cap, skip_special_tokens=True)
+ )
+
+ return self.finalize(full_scores, texts, output_file)
+
+ def __call__(self, sample, full_scores):
+ scores = self._get_pooled_outputs(sample)
+ self._append_scores(scores, full_scores)
+
+ def finalize(self, full_scores, texts, output_file=None):
+ outputs = self._aggregate_scores(full_scores)
+ if output_file is not None:
+ np.save(os.path.join(self.pred_dir, output_file + ".npy"), outputs)
+ return {"outputs": outputs, "texts": texts}
+
+ def _get_pooled_outputs(self, outputs):
+ if "pooled_video" in outputs:
+ return outputs["pooled_video"], outputs["pooled_text"]
+ else:
+ raise ValueError("unknown format of outputs.")
+
+ def _append_scores(self, scores, full_scores):
+ assert len(scores) == 2
+ if len(full_scores) == 0:
+ full_scores.append([])
+ full_scores.append([])
+ full_scores[0].append(scores[0].cpu().detach().numpy())
+ full_scores[1].append(scores[1].cpu().detach().numpy())
+
+ def _aggregate_scores(self, scores):
+ assert len(scores) == 2
+ video_hidden = np.concatenate(scores[0], axis=0)
+ text_hidden = np.concatenate(scores[1], axis=0)
+ # clear up.
+ self.full_scores = []
+ return np.matmul(text_hidden, video_hidden.T)
+
+
+class QAPredictor(Predictor):
+ """generated `pooled_video` and `pooled_text`."""
+ def __init__(self, config):
+ super().__init__(config)
+ """predictor maintains scores and aggregate them."""
+
+ def predict_loop(self, model, eval_dataloader, output_file="qa.npy"):
+ """on-the-fly prediction on a single gpu."""
+ self.full_scores = []
+ model.eval()
+ model = model.cuda()
+ with torch.no_grad():
+ for data in eval_dataloader:
+ # reshape ans and dup video 5 times.
+ v_len = data["vfeats"].size(1)
+ hidden_size = data["vfeats"].size(2)
+ data["vfeats"] = data["vfeats"].unsqueeze(1).repeat(1, 5, 1, 1).view(-1, v_len, hidden_size)
+ data["vmasks"] = data["vmasks"].unsqueeze(1).repeat(1, 5, 1).view(-1, v_len)
+
+ t_len = data["caps"].size(-1)
+ data["caps"] = data["caps"].view(-1, t_len)
+ data["cmasks"] = data["cmasks"].view(-1, t_len)
+
+ data = self.to_ctx(data)
+ outputs = model(**data)
+ outputs.update(data)
+ self(outputs)
+ return self.finalize(output_file)
+
+ def __call__(self, sample):
+ hidden_size = sample["pooled_video"].size(-1)
+ pooled_video = sample["pooled_video"].view(-1, 5, hidden_size)
+ pooled_text = sample["pooled_text"].view(-1, 5, hidden_size)
+ scores = torch.bmm(pooled_video, pooled_text.transpose(2, 1))
+ scores = scores.argmax(-1)
+ self._append_scores(scores[:, 0], sample["answers"], self.full_scores)
+
+ def finalize(self, output_file=None):
+ outputs, targets = self._aggregate_scores(self.full_scores)
+ if output_file is not None:
+ np.save(os.path.join(self.pred_dir, output_file + ".npy"), outputs)
+ return {"outputs": outputs, "targets": targets}
+
+ def _append_scores(self, scores, answers, full_scores):
+ if len(full_scores) == 0:
+ full_scores.append([])
+ full_scores.append([])
+ full_scores[0].append(scores.cpu().detach().numpy())
+ full_scores[1].append(answers.cpu().detach().numpy())
+
+ def _aggregate_scores(self, scores):
+ assert len(scores) == 2
+ outputs = np.concatenate(scores[0], axis=0)
+ targets = np.concatenate(scores[1], axis=0)
+ # clear up.
+ self.full_scores = []
+ return outputs, targets
+
+
+class CrossTaskPredictor(Predictor):
+ """
+ CrossTaskPredictor needs to compute the average of logits
+ for overlapped sliding-window.
+ """
+ def __init__(self, config):
+ super().__init__(config)
+ self.lsm = torch.nn.LogSoftmax(dim=1)
+ self.max_video_len = config.dataset.max_video_len
+ self.sliding_window = config.dataset.sliding_window
+ self.sliding_window_size = config.dataset.sliding_window_size
+ self.annotation_path = config.dataset.annotation_path
+
+ def predict_loop(self, model, eval_dataloader, output_file="result.pkl"):
+ """refactored from line 144:
+ https://github.com/DmZhukov/CrossTask/blob/master/train.py
+ """
+ ctx = 0
+ model.eval()
+ model = model.to(ctx)
+ # this is not a loss but just compute neg_log_prob.
+ Y_pred = {}
+ Y_true = {}
+ with torch.no_grad():
+ for batch in eval_dataloader:
+ self(batch, model, Y_pred, Y_true)
+ return self.finalize(Y_pred, Y_true, output_file)
+
+ def __call__(self, sample, model, Y_pred, Y_true):
+ # please install dp from `https://github.com/DmZhukov/CrossTask`
+ from dp import dp
+ vid, task = sample['video_id'][0], sample['task'][0]
+ sample = self.to_ctx(sample)
+ # compute the average logits over sliding windows.
+ output = model(**sample)
+ batch_logits = output["logits"].cpu()
+
+ video_len = sample["video_len"][0]
+
+ # the following version is slow.
+ logits = torch.zeros((video_len, batch_logits.size(1)))
+ logits_counts = torch.zeros((video_len, 1), dtype=torch.long)
+ # use the same loop as aligner to recover.
+ batch_logit_idx = 0
+ for window_start in range(0, video_len, self.sliding_window):
+ video_end = min(video_len - window_start, self.sliding_window_size)
+ logits[window_start: window_start + video_end] += batch_logits[
+ batch_logit_idx: batch_logit_idx + video_end]
+ batch_logit_idx += video_end
+ logits_counts[window_start: window_start + video_end] += torch.ones((video_end, 1), dtype=torch.long)
+
+ if (video_len - window_start) <= self.sliding_window_size:
+ break
+
+ logits /= logits_counts
+ assert logits.size() == (video_len, batch_logits.size(1)), "{}, {}".format(logits.size(), video_len)
+
+ O = self.lsm(logits)
+ y = np.zeros(O.size(), dtype=np.float32)
+ dp(y, -O.detach().cpu().numpy())
+ if task not in Y_pred:
+ Y_pred[task] = {}
+ Y_pred[task][vid] = y
+ annot_path = os.path.join(
+ self.annotation_path, task+'_'+vid+'.csv')
+ if os.path.exists(annot_path):
+ if task not in Y_true:
+ Y_true[task] = {}
+ Y_true[task][vid] = self._read_assignment(
+ *y.shape, annot_path)
+
+ def finalize(self, Y_pred, Y_true, output_file=None):
+ if output_file is not None:
+ with open(
+ os.path.join(self.pred_dir, output_file + ".pkl"),
+ "wb") as fw:
+ pickle.dump(
+ {"Y_pred": Y_pred, "Y_true": Y_true}, fw,
+ protocol=pickle.HIGHEST_PROTOCOL)
+ return {"outputs": Y_pred, "targets": Y_true}
+
+ def _read_assignment(self, T, K, path):
+ """
+ refactored from https://github.com/DmZhukov/CrossTask/blob/master/data.py
+ Howto interpret contraints on loss that is going to be minimized:
+ lambd is a big number;
+ self.lambd * C is a big number for all valid position (csv stores invalids)
+
+ def forward(self, O, Y, C):
+ return (Y*(self.lambd * C - self.lsm(O))).mean(dim=0).sum()
+
+ This will load the csv file and fill-in the step col from start to end rows.
+ """
+
+ Y = np.zeros([T, K], dtype=np.uint8)
+ with open(path, 'r') as f:
+ for line in f:
+ step, start, end = line.strip().split(',')
+ start = int(math.floor(float(start)))
+ end = int(math.ceil(float(end)))
+ step = int(step) - 1
+ Y[start:end, step] = 1
+ return Y
+
+
+class COINPredictor(Predictor):
+ """
+ COINPredictor is similar to CrossTask on sliding windows.
+ """
+ def __init__(self, config):
+ super().__init__(config)
+ self.max_video_len = config.dataset.max_video_len
+ self.sliding_window = config.dataset.sliding_window
+ self.sliding_window_size = config.dataset.sliding_window_size
+
+ def predict_loop(self, model, eval_dataloader, output_file="result.pkl"):
+ """refactored from line 144:
+ https://github.com/DmZhukov/CrossTask/blob/master/train.py
+ """
+ ctx = 0
+ model.eval()
+ model = model.to(ctx)
+ # this is not a loss but just compute neg_log_prob.
+ Y_pred = []
+ Y_true = []
+ with torch.no_grad():
+ for batch in eval_dataloader:
+ self(batch, model, Y_pred, Y_true)
+ return self.finalize(Y_pred, Y_true, output_file)
+
+ def __call__(self, sample, model, Y_pred, Y_true):
+ sample = self.to_ctx(sample)
+ # compute the average logits over sliding windows.
+ output = model(**sample)
+ logits = self._merge_windows(sample, output)
+ Y_pred.append(logits.argmax(dim=1))
+ Y_true.append(sample["video_targets"].squeeze(0).cpu())
+
+ def _merge_windows(self, sample, output):
+ targets = sample["targets"].reshape(-1).cpu()
+ valid_mask = targets != -100
+ targets = targets[valid_mask]
+ batch_logits = output["logits"].cpu()
+ batch_logits = batch_logits.reshape(-1, batch_logits.size(-1))
+ batch_logits = batch_logits[valid_mask]
+
+ video_len = sample["video_len"][0]
+
+ # the following version is slow.
+ logits = torch.zeros((video_len, batch_logits.size(1)))
+ logits_counts = torch.zeros((video_len, 1), dtype=torch.long)
+ # use the same loop as aligner to recover.
+ batch_logit_idx = 0
+ for window_start in range(0, video_len, self.sliding_window):
+ video_end = min(video_len - window_start, self.sliding_window_size)
+ logits[window_start: window_start + video_end] += batch_logits[
+ batch_logit_idx: batch_logit_idx + video_end]
+ batch_logit_idx += video_end
+ logits_counts[window_start: window_start + video_end] += torch.ones((video_end, 1), dtype=torch.long)
+ if (video_len - window_start) <= self.sliding_window_size:
+ break
+ logits /= logits_counts
+ assert logits.size() == (video_len, batch_logits.size(1)), "{}, {}".format(logits.size(), video_len)
+ return logits
+
+ def finalize(self, Y_pred, Y_true, output_file=None):
+ Y_pred = torch.cat(Y_pred, dim=0).numpy()
+ Y_true = torch.cat(Y_true, dim=0).numpy()
+ assert len(Y_pred) == len(Y_true)
+
+ error_mask = Y_pred != Y_true
+ print("sample error", Y_pred[error_mask][:10], Y_true[error_mask][:10])
+ print("sample error", Y_pred[error_mask][10:20], Y_true[error_mask][10:20])
+
+ if output_file is not None:
+ with open(
+ os.path.join(self.pred_dir, output_file + ".pkl"),
+ "wb") as fw:
+ pickle.dump(
+ {"Y_pred": Y_pred, "Y_true": Y_true}, fw,
+ protocol=pickle.HIGHEST_PROTOCOL)
+ return {"outputs": Y_pred, "targets": Y_true}
+
+
+class COINZSPredictor(COINPredictor):
+ """
+ COINZSPredictor for COIN zero-shot prediction.
+ """
+
+ def __init__(self, config):
+ super().__init__(config)
+ self.dataset_config = config.dataset
+
+ def predict_loop(self, model, eval_dataloader, output_file="result.pkl"):
+ """refactored from line 144:
+ https://github.com/DmZhukov/CrossTask/blob/master/train.py
+ """
+ ctx = 0
+ model.eval()
+ model = model.to(ctx)
+
+ with torch.no_grad():
+ outputs = eval_dataloader.dataset.meta_processor.meta_text_labels(
+ self.dataset_config)
+ outputs = self.to_ctx(outputs, ctx)
+ label_hidden_states = model.forward_text(**outputs).cpu()
+ label_sim = label_hidden_states @ label_hidden_states.t()
+ num_labels = label_sim.size(0)
+ eye_mask = ~torch.eye(num_labels, dtype=torch.bool)
+ label_sim = label_sim.masked_select(eye_mask).view(num_labels, num_labels - 1)
+ lbd = label_sim.max()
+
+ # this is not a loss but just compute neg_log_prob.
+ Y_pred = []
+ Y_true = []
+ with torch.no_grad():
+ for batch in eval_dataloader:
+ self(batch, label_hidden_states, model, lbd, Y_pred, Y_true)
+ return self.finalize(Y_pred, Y_true, output_file)
+
+ def reshape_subsample(self, sample):
+ for key in sample:
+ if torch.is_tensor(sample[key]):
+ sample[key] = self.flat_subsample(sample[key])
+ return sample
+
+ def flat_subsample(self, tensor):
+ if len(tensor.size()) > 1 and tensor.size(0) == 1:
+ tensor = tensor.squeeze(0)
+ return tensor
+
+ def __call__(self, sample, label_hidden_states, model, lbd, Y_pred, Y_true):
+ sample = self.reshape_subsample(sample)
+ sample = self.to_ctx(sample)
+ # compute the average logits over sliding windows.
+ sample["output_hidden_states"] = True
+ video_outputs = model.forward_video(**sample).cpu()
+ output = {"logits": video_outputs[:, 1:sample["vmasks"].size(1)+1] @ label_hidden_states.t()}
+ logits = self._merge_windows(sample, output)
+ # logic of zero-shot for sequence labeling.
+ logits_argmax = logits.argmax(dim=1) + 1 # 0 is "O" label.
+ logits_max = logits.max(dim=1)[0]
+
+ pred = torch.zeros_like(logits_argmax)
+ label_select = logits_max > lbd # 73 or 74
+ pred[label_select] = logits_argmax[label_select]
+
+ Y_pred.append(pred)
+ Y_true.append(sample["video_targets"].squeeze(0).cpu())
+
+ def finalize(self, Y_pred, Y_true, output_file=None):
+ Y_pred = torch.cat(Y_pred, dim=0).numpy()
+ Y_true = torch.cat(Y_true, dim=0).numpy()
+ assert len(Y_pred) == len(Y_true)
+
+ error_mask = Y_pred != Y_true
+ print("sample error", Y_pred[error_mask][:10], Y_true[error_mask][:10])
+ print("sample error", Y_pred[error_mask][10:20], Y_true[error_mask][10:20])
+
+ if output_file is not None:
+ with open(
+ os.path.join(self.pred_dir, output_file + ".pkl"),
+ "wb") as fw:
+ pickle.dump(
+ {"Y_pred": Y_pred, "Y_true": Y_true}, fw,
+ protocol=pickle.HIGHEST_PROTOCOL)
+ return {"outputs": Y_pred, "targets": Y_true}
+
+
+class DiDeMoPredictor(Predictor):
+ """reference: https://github.com/LisaAnne/LocalizingMoments/blob/master/utils/eval.py
+ https://github.com/LisaAnne/LocalizingMoments/blob/master/utils/data_processing.py
+ """
+ def __init__(self, config):
+ super().__init__(config)
+ # load targets.
+ with open(config.dataset.test_path) as data_file:
+ self.test_data = json.load(data_file)
+
+ def predict_loop(self, model, eval_dataloader, output_file="didemo.npy"):
+ """
+ TODO: two solutions here.
+ """
+ import itertools
+ # 21 chunks.
+ self.possible_segments = [(0,0), (1,1), (2,2), (3,3), (4,4), (5,5)]
+ for i in itertools.combinations(range(6), 2):
+ self.possible_segments.append(i)
+ # pick segments from a video.
+
+ """on-the-fly prediction on a single gpu."""
+ self.full_scores = []
+ model.eval()
+ model = model.cuda()
+ with torch.no_grad():
+ for data in eval_dataloader:
+ # TODO special forwarding logic here.
+ data = self.to_ctx(data)
+ data["output_hidden_states"] = True
+ hidden_video = model.forward_video(**data)
+ data["output_hidden_states"] = False
+ pooled_text = model.forward_text(**data)
+ outputs = {
+ "hidden_video": hidden_video,
+ "pooled_text": pooled_text
+ }
+ outputs.update(data)
+ self(outputs)
+ return self.finalize(output_file)
+
+ def __call__(self, sample):
+ # TODO: make an index select from self.possible_segments.
+ hidden_video = sample["hidden_video"]
+ pooled_text = sample["pooled_text"]
+ vmasks = sample["vmasks"]
+ # probably maintain valid results here.
+
+ hidden_video = hidden_video[:, 1:-1, :]
+ # probably maintain valid results here.
+ pooled_video = []
+ for s, e in self.possible_segments:
+ pooled_video.append(
+ torch.mean(
+ hidden_video[:, int(s*5):int((e+1)*5), :],
+ dim=1, keepdim=True)
+ )
+ pooled_video = torch.cat(pooled_video, dim=1)
+ scores = torch.bmm(
+ pooled_video, pooled_text.unsqueeze(-1)).squeeze(-1).cpu()
+
+ ranks = scores.argsort(dim=-1, descending=True)
+
+ for batch_idx, rank in enumerate(ranks):
+ rank_of_moment = []
+ for m_idx, moment in enumerate(rank):
+ s, e = self.possible_segments[moment.item()]
+ if torch.any(
+ vmasks[batch_idx, int(s*5):int((e+1)*5)]
+ ):
+ rank_of_moment.append((s, e))
+ self.full_scores.append(rank_of_moment)
+
+ def finalize(self, output_file=None):
+ outputs = self._aggregate_scores(self.full_scores)
+ if output_file is not None:
+ np.save(os.path.join(self.pred_dir, output_file + ".npy"), outputs)
+ return {"outputs": outputs, "targets": self.test_data}
+
+ def _aggregate_scores(self, scores):
+ self.full_scores = []
+ return scores
diff --git a/fairseq/examples/MMPT/mmpt/losses/__init__.py b/fairseq/examples/MMPT/mmpt/losses/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..8dc32c96d2d8aed25c59e69e9d9a2ff24a9a2a47
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/losses/__init__.py
@@ -0,0 +1,16 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+from .loss import *
+from .nce import *
+
+try:
+ from .fairseqmmloss import *
+except ImportError:
+ pass
+
+try:
+ from .expnce import *
+except ImportError:
+ pass
diff --git a/fairseq/examples/MMPT/mmpt/losses/fairseqmmloss.py b/fairseq/examples/MMPT/mmpt/losses/fairseqmmloss.py
new file mode 100644
index 0000000000000000000000000000000000000000..a95e5ecf45d90098c1719487bb9a11c36be7c507
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/losses/fairseqmmloss.py
@@ -0,0 +1,63 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""
+TODO (huxu): a general fairseq criterion for all your pre-defined losses.
+"""
+
+from fairseq.criterions import FairseqCriterion, register_criterion
+from fairseq.logging import metrics
+
+
+@register_criterion("mmloss")
+class MMCriterion(FairseqCriterion):
+ def __init__(self, task):
+ super().__init__(task)
+ # TODO (huxu): wrap forward call of loss_fn and eval_fn into task.
+ self.mmtask = task.mmtask
+
+ def forward(self, model, sample):
+ """Compute the loss for the given sample.
+ Returns a tuple with three elements:
+ 1) the loss
+ 2) the sample size, which is used as the denominator for the gradient
+ 3) logging outputs to display while training
+ """
+ outputs = self.mmtask(model, sample)
+
+ loss, loss_scalar, max_len, batch_size, sample_size = (
+ outputs["loss"],
+ outputs["loss_scalar"],
+ outputs["max_len"],
+ outputs["batch_size"],
+ outputs["sample_size"],
+ )
+
+ logging_output = {
+ "loss": loss_scalar,
+ "ntokens": max_len * batch_size, # dummy report.
+ "nsentences": batch_size, # dummy report.
+ "sample_size": sample_size,
+ }
+
+ return loss, 1, logging_output
+
+ @staticmethod
+ def reduce_metrics(logging_outputs) -> None:
+ """Aggregate logging outputs from data parallel training."""
+ """since we use NCE, our actual batch_size is 1 per GPU.
+ Then we take the mean of each worker."""
+ loss_sum = sum(log.get("loss", 0.0) for log in logging_outputs)
+ sample_size = sum(log.get("sample_size", 0) for log in logging_outputs)
+ metrics.log_scalar("loss", loss_sum / sample_size, round=3)
+
+ @staticmethod
+ def logging_outputs_can_be_summed() -> bool:
+ """
+ Whether the logging outputs returned by `forward` can be summed
+ across workers prior to calling `reduce_metrics`. Setting this
+ to True will improves distributed training speed.
+ """
+ return True
diff --git a/fairseq/examples/MMPT/mmpt/losses/loss.py b/fairseq/examples/MMPT/mmpt/losses/loss.py
new file mode 100644
index 0000000000000000000000000000000000000000..99c05d067edac220f9e53080f09f0b40d7dc1e8d
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/losses/loss.py
@@ -0,0 +1,87 @@
+# Copyright (c) Facebook, Inc. All Rights Reserved
+
+import torch
+
+from torch import nn
+
+
+class Loss(object):
+ def __call__(self, *args, **kwargs):
+ raise NotImplementedError
+
+
+# Dummy Loss for testing.
+class DummyLoss(Loss):
+ def __init__(self):
+ self.loss = nn.CrossEntropyLoss()
+
+ def __call__(self, logits, targets, **kwargs):
+ return self.loss(logits, targets)
+
+
+class DummyK400Loss(Loss):
+ """dummy k400 loss for MViT."""
+ def __init__(self):
+ self.loss = nn.CrossEntropyLoss()
+
+ def __call__(self, logits, targets, **kwargs):
+ return self.loss(
+ logits, torch.randint(0, 400, (logits.size(0),), device=logits.device))
+
+
+class CrossEntropy(Loss):
+ def __init__(self):
+ self.loss = nn.CrossEntropyLoss()
+
+ def __call__(self, logits, targets, **kwargs):
+ return self.loss(logits.reshape(-1, logits.size(-1)), targets.reshape(-1))
+
+
+class ArgmaxCrossEntropy(Loss):
+ def __init__(self):
+ self.loss = nn.CrossEntropyLoss()
+
+ def __call__(self, logits, targets, **kwargs):
+ return self.loss(logits, targets.argmax(dim=1))
+
+
+class BCE(Loss):
+ def __init__(self):
+ self.loss = nn.BCEWithLogitsLoss()
+
+ def __call__(self, logits, targets, **kwargs):
+ targets = targets.squeeze(0)
+ return self.loss(logits, targets)
+
+
+class NLGLoss(Loss):
+ def __init__(self):
+ self.loss = nn.CrossEntropyLoss()
+
+ def __call__(self, logits, text_label, **kwargs):
+ targets = text_label[text_label != -100]
+ return self.loss(logits, targets)
+
+
+class MSE(Loss):
+ def __init__(self):
+ self.loss = nn.MSELoss()
+
+ def __call__(self, logits, targets, **kwargs):
+ return self.loss(logits, targets)
+
+
+class L1(Loss):
+ def __init__(self):
+ self.loss = nn.L1Loss()
+
+ def __call__(self, logits, targets, **kwargs):
+ return self.loss(logits, targets)
+
+
+class SmoothL1(Loss):
+ def __init__(self):
+ self.loss = nn.SmoothL1Loss()
+
+ def __call__(self, logits, targets, **kwargs):
+ return self.loss(logits, targets)
diff --git a/fairseq/examples/MMPT/mmpt/losses/nce.py b/fairseq/examples/MMPT/mmpt/losses/nce.py
new file mode 100644
index 0000000000000000000000000000000000000000..ed7be8d372e371bb0e0d6166f76e01d3466d2306
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/losses/nce.py
@@ -0,0 +1,156 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""
+softmax-based NCE loss, used by this project.
+"""
+
+import torch
+
+from torch import nn
+
+from .loss import Loss
+
+
+class NCE(Loss):
+ def __init__(self):
+ # TODO (huxu): define temperature.
+ self.loss = nn.CrossEntropyLoss()
+
+ def __call__(self, align_scores, **kargs):
+ # note: we reuse the same shape as cls head in BERT (batch_size, 2)
+ # but NCE only needs one logits.
+ # (so we drop all weights in the second neg logits.)
+ align_scores = align_scores[:, :1]
+ # duplicate negative examples
+ batch_size = align_scores.size(0) // 2
+ pos_scores = align_scores[:batch_size]
+ neg_scores = align_scores[batch_size:].view(1, batch_size).repeat(
+ batch_size, 1)
+ scores = torch.cat([pos_scores, neg_scores], dim=1)
+ return self.loss(
+ scores,
+ torch.zeros(
+ (batch_size,),
+ dtype=torch.long,
+ device=align_scores.device),
+ )
+
+
+class T2VContraLoss(Loss):
+ """NCE for MM joint space, on softmax text2video matrix.
+ """
+ def __init__(self):
+ # TODO (huxu): define temperature.
+ self.loss = nn.CrossEntropyLoss()
+
+ def __call__(self, pooled_video, pooled_text, **kargs):
+ batch_size = pooled_video.size(0)
+ logits = torch.mm(pooled_text, pooled_video.transpose(1, 0))
+ targets = torch.arange(
+ batch_size,
+ dtype=torch.long,
+ device=pooled_video.device)
+ return self.loss(logits, targets)
+
+
+class V2TContraLoss(Loss):
+ """NCE for MM joint space, with softmax on video2text matrix."""
+
+ def __init__(self):
+ # TODO (huxu): define temperature.
+ self.loss = nn.CrossEntropyLoss()
+
+ def __call__(self, pooled_video, pooled_text, **kargs):
+ batch_size = pooled_video.size(0)
+ logits = torch.mm(pooled_video, pooled_text.transpose(1, 0))
+ targets = torch.arange(
+ batch_size,
+ dtype=torch.long,
+ device=pooled_video.device)
+ return self.loss(logits, targets)
+
+
+class MMContraLoss(Loss):
+ def __init__(self):
+ self.loss = nn.CrossEntropyLoss()
+
+ def __call__(self, pooled_video, pooled_text, **kwargs):
+ logits_per_video = pooled_video @ pooled_text.t()
+ logits_per_text = pooled_text @ pooled_video.t()
+
+ targets = torch.arange(
+ pooled_video.size(0),
+ dtype=torch.long,
+ device=pooled_video.device)
+ loss_video = self.loss(logits_per_video, targets)
+ loss_text = self.loss(logits_per_text, targets)
+ return loss_video + loss_text
+
+
+class MTM(Loss):
+ """Combination of MFM and MLM."""
+
+ def __init__(self):
+ self.loss = nn.CrossEntropyLoss()
+
+ def __call__(
+ self,
+ video_logits,
+ text_logits,
+ video_label,
+ text_label,
+ **kwargs
+ ):
+ text_logits = torch.cat([
+ text_logits,
+ torch.zeros(
+ (text_logits.size(0), 1), device=text_logits.device)
+ ], dim=1)
+ vt_logits = torch.cat([video_logits, text_logits], dim=0)
+ # loss for video.
+ video_label = torch.zeros(
+ (video_logits.size(0),),
+ dtype=torch.long,
+ device=video_logits.device
+ )
+
+ # loss for text.
+ text_label = text_label.reshape(-1)
+ labels_mask = text_label != -100
+ selected_text_label = text_label[labels_mask]
+
+ vt_label = torch.cat([video_label, selected_text_label], dim=0)
+ return self.loss(vt_logits, vt_label)
+
+
+class MFMMLM(Loss):
+ """Combination of MFM and MLM."""
+
+ def __init__(self):
+ self.loss = nn.CrossEntropyLoss()
+
+ def __call__(
+ self,
+ video_logits,
+ text_logits,
+ video_label,
+ text_label,
+ **kwargs
+ ):
+ # loss for video.
+ video_label = torch.zeros(
+ (video_logits.size(0),),
+ dtype=torch.long,
+ device=video_logits.device
+ )
+ masked_frame_loss = self.loss(video_logits, video_label)
+
+ # loss for text.
+ text_label = text_label.reshape(-1)
+ labels_mask = text_label != -100
+ selected_text_label = text_label[labels_mask]
+ masked_lm_loss = self.loss(text_logits, selected_text_label)
+ return masked_frame_loss + masked_lm_loss
diff --git a/fairseq/examples/MMPT/mmpt/models/__init__.py b/fairseq/examples/MMPT/mmpt/models/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..825250cd007f5e072b6c9d2376445b955a7aa71e
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/models/__init__.py
@@ -0,0 +1,17 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+from .mmfusion import *
+from .transformermodel import *
+from .mmfusionnlg import *
+
+try:
+ from .fairseqmmmodel import *
+except ImportError:
+ pass
+
+try:
+ from .expmmfusion import *
+except ImportError:
+ pass
diff --git a/fairseq/examples/MMPT/mmpt/models/fairseqmmmodel.py b/fairseq/examples/MMPT/mmpt/models/fairseqmmmodel.py
new file mode 100644
index 0000000000000000000000000000000000000000..b7dd643693dee8cfc20ca77d6cea798d07eaf15a
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/models/fairseqmmmodel.py
@@ -0,0 +1,51 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+from fairseq.models import (
+ BaseFairseqModel,
+ register_model,
+ register_model_architecture
+)
+
+
+@register_model("mmmodel")
+class FairseqMMModel(BaseFairseqModel):
+ """a fairseq wrapper of model built by `task`."""
+
+ @classmethod
+ def build_model(cls, args, task):
+ return FairseqMMModel(task.mmtask.model)
+
+ def __init__(self, mmmodel):
+ super().__init__()
+ self.mmmodel = mmmodel
+
+ def forward(self, *args, **kwargs):
+ return self.mmmodel(*args, **kwargs)
+
+ def upgrade_state_dict_named(self, state_dict, name):
+
+ super().upgrade_state_dict_named(state_dict, name)
+
+ keys_to_delete = []
+
+ for key in state_dict:
+ if key not in self.state_dict():
+ keys_to_delete.append(key)
+ for key in keys_to_delete:
+ print("[INFO]", key, "not used anymore.")
+ del state_dict[key]
+
+ # copy any newly defined parameters.
+ for key in self.state_dict():
+ if key not in state_dict:
+ print("[INFO] adding", key)
+ state_dict[key] = self.state_dict()[key]
+
+
+# a dummy arch, we config the model.
+@register_model_architecture("mmmodel", "mmarch")
+def mmarch(args):
+ pass
diff --git a/fairseq/examples/MMPT/mmpt/models/mmfusion.py b/fairseq/examples/MMPT/mmpt/models/mmfusion.py
new file mode 100644
index 0000000000000000000000000000000000000000..2509e26b67b467c3b18c76630881e66cf334a350
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/models/mmfusion.py
@@ -0,0 +1,926 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# Copyright (c) Facebook, Inc. All Rights Reserved
+
+
+import torch
+
+from torch import nn
+
+try:
+ from transformers import AutoConfig, AutoTokenizer
+except ImportError:
+ pass
+
+from . import transformermodel
+
+
+class MMPTModel(nn.Module):
+ """An e2e wrapper of inference model.
+ """
+ @classmethod
+ def from_pretrained(cls, config, checkpoint="checkpoint_best.pt"):
+ import os
+ from ..utils import recursive_config
+ from ..tasks import Task
+ config = recursive_config(config)
+ mmtask = Task.config_task(config)
+ checkpoint_path = os.path.join(config.eval.save_path, checkpoint)
+ mmtask.build_model(checkpoint=checkpoint_path)
+ # TODO(huxu): make the video encoder configurable.
+ from ..processors.models.s3dg import S3D
+ video_encoder = S3D('pretrained_models/s3d_dict.npy', 512)
+ video_encoder.load_state_dict(
+ torch.load('pretrained_models/s3d_howto100m.pth'))
+ from transformers import AutoTokenizer
+ tokenizer = AutoTokenizer.from_pretrained(
+ config.dataset.bert_name, use_fast=config.dataset.use_fast
+ )
+ from ..processors import Aligner
+ aligner = Aligner(config.dataset)
+ return (
+ MMPTModel(config, mmtask.model, video_encoder),
+ tokenizer,
+ aligner
+ )
+
+ def __init__(self, config, model, video_encoder, **kwargs):
+ super().__init__()
+ self.max_video_len = config.dataset.max_video_len
+ self.video_encoder = video_encoder
+ self.model = model
+
+ def forward(self, video_frames, caps, cmasks, return_score=False):
+ bsz = video_frames.size(0)
+ assert bsz == 1, "only bsz=1 is supported now."
+ seq_len = video_frames.size(1)
+ video_frames = video_frames.view(-1, *video_frames.size()[2:])
+ vfeats = self.video_encoder(video_frames.permute(0, 4, 1, 2, 3))
+ vfeats = vfeats['video_embedding']
+ vfeats = vfeats.view(bsz, seq_len, vfeats.size(-1))
+ padding = torch.zeros(
+ bsz, self.max_video_len - seq_len, vfeats.size(-1))
+ vfeats = torch.cat([vfeats, padding], dim=1)
+ vmasks = torch.cat([
+ torch.ones((bsz, seq_len), dtype=torch.bool),
+ torch.zeros((bsz, self.max_video_len - seq_len), dtype=torch.bool)
+ ],
+ dim=1
+ )
+ output = self.model(caps, cmasks, vfeats, vmasks)
+ if return_score:
+ output = {"score": torch.bmm(
+ output["pooled_video"][:, None, :],
+ output["pooled_text"][:, :, None]
+ ).squeeze(-1).squeeze(-1)}
+ return output
+
+
+class MMFusion(nn.Module):
+ """a MMPT wrapper class for MMBert style models.
+ TODO: move isolated mask to a subclass.
+ """
+ def __init__(self, config, **kwargs):
+ super().__init__()
+ transformer_config = AutoConfig.from_pretrained(
+ config.dataset.bert_name)
+ self.hidden_size = transformer_config.hidden_size
+ self.is_train = False
+ if config.dataset.train_path is not None:
+ self.is_train = True
+ # 0 means no iso; 1-12 means iso up to that layer.
+ self.num_hidden_layers = transformer_config.num_hidden_layers
+ self.last_iso_layer = 0
+ if config.dataset.num_iso_layer is not None:
+ self.last_iso_layer = config.dataset.num_iso_layer - 1 + 1
+
+ if config.model.mm_encoder_cls is not None:
+ mm_encoder_cls = getattr(transformermodel, config.model.mm_encoder_cls)
+ model_config = AutoConfig.from_pretrained(config.dataset.bert_name)
+ model_config.max_video_len = config.dataset.max_video_len
+ # TODO: a general way to add parameter for a model.
+ model_config.use_seg_emb = config.model.use_seg_emb
+ self.mm_encoder = mm_encoder_cls.from_pretrained(
+ config.dataset.bert_name, config=model_config)
+ elif config.model.video_encoder_cls is not None\
+ and config.model.text_encoder_cls is not None:
+ video_encoder_cls = getattr(transformermodel, config.model.video_encoder_cls)
+ model_config = AutoConfig.from_pretrained(config.dataset.bert_name)
+ model_config.max_video_len = config.dataset.max_video_len
+ # TODO: make each model a set of config class.
+ if hasattr(model_config, "num_layers"):
+ model_config.num_layers = config.model.num_hidden_video_layers
+ else:
+ model_config.num_hidden_layers = config.model.num_hidden_video_layers
+ self.video_encoder = video_encoder_cls.from_pretrained(
+ config.dataset.bert_name, config=model_config)
+ # exact same NLP model from Huggingface.
+ text_encoder_cls = getattr(transformermodel, config.model.text_encoder_cls)
+ self.text_encoder = text_encoder_cls.from_pretrained(
+ config.dataset.bert_name)
+ else:
+ raise ValueError("the encoder must be either MM or two backbones.")
+
+ def forward(
+ self,
+ caps,
+ cmasks,
+ vfeats,
+ vmasks,
+ **kwargs
+ ):
+ raise NotImplementedError(
+ "Please derive MMFusion module."
+ )
+
+ def _mm_on_the_fly(
+ self,
+ cmasks,
+ vmasks,
+ attention_mask
+ ):
+ """helper function for mask, seg_ids and token_type_ids."""
+ if attention_mask is None:
+ attention_mask = self._mm_attention_mask(cmasks, vmasks)
+
+ """
+ 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1
+ | first sequence | second sequence |
+ """
+ token_type_ids = torch.cat(
+ [
+ torch.zeros(
+ (vmasks.size(0), vmasks.size(1) + 2),
+ dtype=torch.long,
+ device=vmasks.device,
+ ),
+ torch.ones(
+ (cmasks.size(0), cmasks.size(1) - 2),
+ dtype=torch.long,
+ device=cmasks.device,
+ ),
+ ],
+ dim=1,
+ )
+ return attention_mask, token_type_ids
+
+ def _mm_attention_mask(self, cmasks, vmasks):
+ assert cmasks.size(0) == vmasks.size(0), "{}, {}, {}, {}".format(
+ str(cmasks.size()),
+ str(vmasks.size()),
+ str(cmasks.size(0)),
+ str(vmasks.size(0)),
+ )
+
+ mm_mask = torch.cat([cmasks[:, :1], vmasks, cmasks[:, 1:]], dim=1)
+ if self.last_iso_layer == 0:
+ # hard attention mask.
+ return mm_mask
+ else:
+ # a gpu iso mask; 0 : num_iso_layer is isolated;
+ # num_iso_layer: are MM-fused.
+ # make an iso layer
+ batch_size = cmasks.size(0)
+ iso_mask = self._make_iso_mask(batch_size, cmasks, vmasks)
+ mm_mask = mm_mask[:, None, :].repeat(1, mm_mask.size(-1), 1)
+ iso_mm_masks = []
+ # hard attention mask.
+ iso_mask = iso_mask[:, None, :, :].repeat(
+ 1, self.last_iso_layer, 1, 1)
+ iso_mm_masks.append(iso_mask)
+ if self.last_iso_layer < self.num_hidden_layers:
+ mm_mask = mm_mask[:, None, :, :].repeat(
+ 1, self.num_hidden_layers - self.last_iso_layer, 1, 1
+ )
+ iso_mm_masks.append(mm_mask)
+ iso_mm_masks = torch.cat(iso_mm_masks, dim=1)
+ return iso_mm_masks
+
+ def _make_iso_mask(self, batch_size, cmasks, vmasks):
+ cls_self_mask = torch.cat(
+ [
+ torch.ones(
+ (batch_size, 1), dtype=torch.bool, device=cmasks.device),
+ torch.zeros(
+ (batch_size, cmasks.size(1) + vmasks.size(1) - 1),
+ dtype=torch.bool, device=cmasks.device)
+ ], dim=1)
+
+ iso_video_mask = torch.cat(
+ [
+ # [CLS] is not used.
+ torch.zeros(
+ (batch_size, 1), dtype=torch.bool, device=cmasks.device
+ ),
+ vmasks,
+ # assume to be 1.
+ cmasks[:, 1:2],
+ # 2 means [CLS] + [SEP]
+ torch.zeros(
+ (batch_size, cmasks.size(1) - 2),
+ dtype=torch.bool,
+ device=cmasks.device,
+ ),
+ ],
+ dim=1,
+ )
+ iso_text_mask = torch.cat(
+ [
+ torch.zeros(
+ (batch_size, 2 + vmasks.size(1)),
+ dtype=torch.bool,
+ device=cmasks.device,
+ ), # [CLS] is not used.
+ cmasks[:, 2:], # assume to be 1.
+ ],
+ dim=1,
+ )
+ cls_self_mask = cls_self_mask[:, None, :]
+ iso_video_mask = iso_video_mask[:, None, :].repeat(
+ 1, vmasks.size(1) + 1, 1)
+ iso_text_mask = iso_text_mask[:, None, :].repeat(
+ 1, cmasks.size(1) - 2, 1)
+ return torch.cat([cls_self_mask, iso_video_mask, iso_text_mask], dim=1)
+
+ def _pooling_vt_layer(
+ self,
+ layered_sequence_output,
+ cmasks,
+ vmasks
+ ):
+ layer_idx = self.last_iso_layer \
+ if self.last_iso_layer > 0 else self.num_hidden_layers
+ hidden_state = layered_sequence_output[layer_idx]
+ # also output pooled_video and pooled_text.
+ batch_size = cmasks.size(0)
+ # pool the modality.
+ text_offset = vmasks.size(1) + 2 # [CLS] + [SEP]
+ # video tokens + [SEP]
+ video_outputs = hidden_state[:, 1:text_offset]
+ video_attention_mask = torch.cat(
+ [
+ vmasks,
+ torch.ones(
+ (batch_size, 1), dtype=torch.bool, device=vmasks.device),
+ ],
+ dim=1,
+ )
+ assert video_outputs.size(1) == video_attention_mask.size(1)
+ pooled_video = torch.sum(
+ video_outputs * video_attention_mask.unsqueeze(-1), dim=1
+ ) / video_attention_mask.sum(1, keepdim=True)
+ # pooled_video = torch.mean(video_outputs[0], dim=1)
+
+ # text tokens + [SEP]
+ text_attention_mask = cmasks[:, 2:]
+ text_outputs = hidden_state[:, text_offset:]
+ assert text_outputs.size(1) == text_attention_mask.size(1)
+ pooled_text = torch.sum(
+ text_outputs * text_attention_mask.unsqueeze(-1), dim=1
+ ) / text_attention_mask.sum(1, keepdim=True)
+ return pooled_video, pooled_text
+
+
+class MMFusionMFMMLM(MMFusion):
+ """forward function for MFM and MLM."""
+ def forward(
+ self,
+ caps,
+ cmasks,
+ vfeats,
+ vmasks,
+ attention_mask=None,
+ video_label=None,
+ text_label=None,
+ **kwargs
+ ):
+ output_hidden_states = False if self.is_train else True
+
+ target_vfeats, non_masked_frame_mask = None, None
+ if video_label is not None:
+ target_vfeats = vfeats.masked_select(
+ video_label.unsqueeze(-1)).view(
+ -1, vfeats.size(-1)
+ )
+ # mask video token.
+ vfeats[video_label] = 0.0
+ non_masked_frame_mask = vmasks.clone()
+ non_masked_frame_mask[video_label] = False
+
+ attention_mask, token_type_ids = self._mm_on_the_fly(
+ cmasks, vmasks, attention_mask)
+
+ outputs = self.mm_encoder(
+ input_ids=caps,
+ input_video_embeds=vfeats,
+ attention_mask=attention_mask,
+ token_type_ids=token_type_ids,
+ masked_frame_labels=video_label,
+ target_video_hidden_states=target_vfeats,
+ non_masked_frame_mask=non_masked_frame_mask,
+ masked_lm_labels=text_label,
+ output_hidden_states=output_hidden_states,
+ )
+
+ video_logits, text_logits = outputs[0], outputs[1]
+
+ if self.is_train: # return earlier for training.
+ return {
+ "video_logits": video_logits,
+ "text_logits": text_logits,
+ }
+
+ pooled_video, pooled_text = self._pooling_vt_layer(
+ outputs[2], cmasks, vmasks)
+ return {"pooled_video": pooled_video, "pooled_text": pooled_text}
+
+
+class MMFusionMTM(MMFusionMFMMLM):
+ def __init__(self, config, **kwargs):
+ super().__init__(config)
+ """
+ For reproducibility:
+ self.mm_encoder will be initialized then discarded.
+ """
+ from .transformermodel import MMBertForMTM
+ model_config = AutoConfig.from_pretrained(config.dataset.bert_name)
+ model_config.max_video_len = config.dataset.max_video_len
+ model_config.use_seg_emb = config.model.use_seg_emb
+ self.mm_encoder = MMBertForMTM.from_pretrained(
+ config.dataset.bert_name, config=model_config)
+
+
+class MMFusionShare(MMFusion):
+ """A retrival wrapper using mm_encoder as both video/text backbone.
+ TODO: move formally.
+ """
+ def forward(
+ self,
+ caps,
+ cmasks,
+ vfeats,
+ vmasks,
+ attention_mask=None,
+ video_label=None,
+ text_label=None,
+ output_hidden_states=False,
+ **kwargs
+ ):
+ pooled_video = self.forward_video(
+ vfeats,
+ vmasks,
+ caps,
+ cmasks,
+ output_hidden_states
+ )
+
+ pooled_text = self.forward_text(
+ caps,
+ cmasks,
+ output_hidden_states
+ )
+
+ return {"pooled_video": pooled_video, "pooled_text": pooled_text}
+
+ def forward_video(
+ self,
+ vfeats,
+ vmasks,
+ caps,
+ cmasks,
+ output_hidden_states=False,
+ **kwargs
+ ):
+ input_ids = caps[:, :2]
+
+ attention_mask = torch.cat([
+ cmasks[:, :1],
+ vmasks,
+ cmasks[:, 1:2]
+ ], dim=1)
+
+ token_type_ids = torch.zeros(
+ (vmasks.size(0), vmasks.size(1) + 2),
+ dtype=torch.long,
+ device=vmasks.device)
+
+ outputs = self.mm_encoder(
+ input_ids=input_ids,
+ input_video_embeds=vfeats,
+ attention_mask=attention_mask,
+ token_type_ids=token_type_ids,
+ output_hidden_states=True
+ )
+ video_outputs = outputs[0]
+
+ if output_hidden_states:
+ return video_outputs
+
+ batch_size = cmasks.size(0)
+
+ video_attention_mask = torch.cat(
+ [
+ torch.zeros(
+ (batch_size, 1), dtype=torch.bool, device=vmasks.device),
+ vmasks,
+ torch.ones(
+ (batch_size, 1), dtype=torch.bool, device=vmasks.device),
+ ],
+ dim=1,
+ )
+ assert video_outputs.size(1) == video_attention_mask.size(1)
+
+ video_attention_mask = video_attention_mask.type(video_outputs.dtype) \
+ / video_attention_mask.sum(1, keepdim=True)
+
+ pooled_video = torch.bmm(
+ video_outputs.transpose(2, 1),
+ video_attention_mask.unsqueeze(2)
+ ).squeeze(-1)
+ return pooled_video # video_outputs
+
+ def forward_text(
+ self,
+ caps,
+ cmasks,
+ output_hidden_states=False,
+ **kwargs
+ ):
+ input_ids = torch.cat([
+ caps[:, :1], caps[:, 2:],
+ ], dim=1)
+
+ attention_mask = torch.cat([
+ cmasks[:, :1],
+ cmasks[:, 2:]
+ ], dim=1)
+
+ token_type_ids = torch.cat([
+ torch.zeros(
+ (cmasks.size(0), 1),
+ dtype=torch.long,
+ device=cmasks.device),
+ torch.ones(
+ (cmasks.size(0), cmasks.size(1) - 2),
+ dtype=torch.long,
+ device=cmasks.device)
+ ], dim=1)
+
+ outputs = self.mm_encoder(
+ input_ids=input_ids,
+ input_video_embeds=None,
+ attention_mask=attention_mask,
+ token_type_ids=token_type_ids,
+ output_hidden_states=True
+ )
+ text_outputs = outputs[0]
+
+ if output_hidden_states:
+ return text_outputs
+
+ batch_size = caps.size(0)
+ # text tokens + [SEP]
+ text_attention_mask = torch.cat([
+ torch.zeros(
+ (batch_size, 1), dtype=torch.bool, device=cmasks.device),
+ cmasks[:, 2:]
+ ], dim=1)
+
+ assert text_outputs.size(1) == text_attention_mask.size(1)
+
+ text_attention_mask = text_attention_mask.type(text_outputs.dtype) \
+ / text_attention_mask.sum(1, keepdim=True)
+
+ pooled_text = torch.bmm(
+ text_outputs.transpose(2, 1),
+ text_attention_mask.unsqueeze(2)
+ ).squeeze(-1)
+ return pooled_text # text_outputs
+
+
+class MMFusionSeparate(MMFusionShare):
+ def forward_video(
+ self,
+ vfeats,
+ vmasks,
+ caps,
+ cmasks,
+ output_hidden_states=False,
+ **kwargs
+ ):
+ input_ids = caps[:, :2]
+
+ attention_mask = torch.cat([
+ cmasks[:, :1],
+ vmasks,
+ cmasks[:, 1:2]
+ ], dim=1)
+
+ token_type_ids = torch.zeros(
+ (vmasks.size(0), vmasks.size(1) + 2),
+ dtype=torch.long,
+ device=vmasks.device)
+
+ outputs = self.video_encoder(
+ input_ids=input_ids,
+ input_video_embeds=vfeats,
+ attention_mask=attention_mask,
+ token_type_ids=token_type_ids,
+ output_hidden_states=True
+ )
+ video_outputs = outputs[0]
+
+ if output_hidden_states:
+ return video_outputs
+
+ batch_size = cmasks.size(0)
+
+ video_attention_mask = torch.cat(
+ [
+ torch.zeros(
+ (batch_size, 1), dtype=torch.bool, device=vmasks.device),
+ vmasks,
+ torch.ones(
+ (batch_size, 1), dtype=torch.bool, device=vmasks.device),
+ ],
+ dim=1,
+ )
+ assert video_outputs.size(1) == video_attention_mask.size(1)
+
+ video_attention_mask = video_attention_mask.type(video_outputs.dtype) \
+ / video_attention_mask.sum(1, keepdim=True)
+
+ pooled_video = torch.bmm(
+ video_outputs.transpose(2, 1),
+ video_attention_mask.unsqueeze(2)
+ ).squeeze(-1)
+ return pooled_video # video_outputs
+
+ def forward_text(
+ self,
+ caps,
+ cmasks,
+ output_hidden_states=False,
+ **kwargs
+ ):
+ input_ids = torch.cat([
+ caps[:, :1], caps[:, 2:],
+ ], dim=1)
+
+ attention_mask = torch.cat([
+ cmasks[:, :1],
+ cmasks[:, 2:]
+ ], dim=1)
+ # different from sharing, we use all-0 type.
+ token_type_ids = torch.zeros(
+ (cmasks.size(0), cmasks.size(1) - 1),
+ dtype=torch.long,
+ device=cmasks.device)
+
+ outputs = self.text_encoder(
+ input_ids=input_ids,
+ attention_mask=attention_mask,
+ token_type_ids=token_type_ids,
+ output_hidden_states=True
+ )
+ text_outputs = outputs[0]
+
+ if output_hidden_states:
+ return text_outputs
+
+ batch_size = caps.size(0)
+ # text tokens + [SEP]
+ text_attention_mask = torch.cat([
+ torch.zeros(
+ (batch_size, 1), dtype=torch.bool, device=cmasks.device),
+ cmasks[:, 2:]
+ ], dim=1)
+
+ assert text_outputs.size(1) == text_attention_mask.size(1)
+
+ text_attention_mask = text_attention_mask.type(text_outputs.dtype) \
+ / text_attention_mask.sum(1, keepdim=True)
+
+ pooled_text = torch.bmm(
+ text_outputs.transpose(2, 1),
+ text_attention_mask.unsqueeze(2)
+ ).squeeze(-1)
+ return pooled_text # text_outputs
+
+
+class MMFusionJoint(MMFusion):
+ """fine-tuning wrapper for retrival task."""
+
+ def forward(
+ self,
+ caps,
+ cmasks,
+ vfeats,
+ vmasks,
+ attention_mask=None,
+ video_label=None,
+ text_label=None,
+ **kwargs
+ ):
+ # TODO (huxu): other ways to do negative examples; move the following
+ # into your criterion forward.
+ output_hidden_states = True
+
+ attention_mask, token_type_ids = self._mm_on_the_fly(
+ cmasks, vmasks, attention_mask)
+
+ separate_forward_split = (
+ None if self.is_train else vmasks.size(1) + 2
+ ) # [CLS] + [SEP]
+
+ outputs = self.mm_encoder(
+ input_ids=caps,
+ input_video_embeds=vfeats,
+ attention_mask=attention_mask,
+ token_type_ids=token_type_ids,
+ output_hidden_states=output_hidden_states,
+ separate_forward_split=separate_forward_split,
+ )
+
+ pooled_video, pooled_text = self._pooling_vt_layer(
+ outputs[2], cmasks, vmasks)
+ return {"pooled_video": pooled_video, "pooled_text": pooled_text}
+
+
+class MMFusionActionSegmentation(MMFusion):
+ """Fine-tuning wrapper for action segmentation.
+ TODO: rename this for VLM.
+ """
+ def forward(
+ self,
+ caps,
+ cmasks,
+ vfeats,
+ vmasks,
+ attention_mask=None,
+ **kwargs
+ ):
+ # ActionLocalization assume of batch_size=1, squeeze it.
+ caps = caps.view(-1, caps.size(-1))
+ cmasks = cmasks.view(-1, cmasks.size(-1))
+ vfeats = vfeats.view(-1, vfeats.size(2), vfeats.size(3))
+ vmasks = vmasks.view(-1, vmasks.size(-1))
+
+ # this may not cover all shapes of attention_mask.
+ attention_mask = attention_mask.view(
+ -1, attention_mask.size(2), attention_mask.size(3)) \
+ if attention_mask is not None else None
+
+ # TODO (huxu): other ways to do negative examples; move the following
+ # into your criterion forward.
+ output_hidden_states = True
+
+ # video forwarding, text is dummy; never use attention_mask.
+ attention_mask, token_type_ids = self._mm_on_the_fly(
+ cmasks, vmasks, attention_mask)
+
+ logits = self.mm_encoder(
+ input_ids=caps,
+ input_video_embeds=vfeats,
+ attention_mask=attention_mask,
+ token_type_ids=token_type_ids,
+ output_hidden_states=output_hidden_states,
+ )
+ return {"logits": logits[0][:, 1:vmasks.size(1)+1]}
+
+
+class MMFusionActionLocalization(MMFusion):
+ """fine-tuning model for retrival task."""
+
+ def __init__(self, config, **kwargs):
+ super().__init__(config)
+ tokenizer = AutoTokenizer.from_pretrained(
+ config.dataset.bert_name)
+ self.cls_token_id = tokenizer.cls_token_id
+ self.sep_token_id = tokenizer.sep_token_id
+ self.pad_token_id = tokenizer.pad_token_id
+
+ def forward(
+ self,
+ caps,
+ cmasks,
+ vfeats,
+ vmasks,
+ attention_mask=None,
+ **kwargs
+ ):
+ # ActionLocalization assume of batch_size=1, squeeze it.
+ caps = caps.squeeze(0)
+ cmasks = cmasks.squeeze(0)
+ vfeats = vfeats.squeeze(0)
+ vmasks = vmasks.squeeze(0)
+ attention_mask = attention_mask.squeeze(0) if attention_mask is not None else None
+
+ # TODO (huxu): other ways to do negative examples; move the following
+ # into your criterion forward.
+ output_hidden_states = True
+
+ # a len1 dummy video token.
+ dummy_vfeats = torch.zeros(
+ (caps.size(0), 1, vfeats.size(-1)), device=vfeats.device, dtype=vfeats.dtype)
+ dummy_vmasks = torch.ones(
+ (caps.size(0), 1), dtype=torch.bool,
+ device=vfeats.device)
+
+ dummy_caps = torch.LongTensor(
+ [[self.cls_token_id, self.sep_token_id,
+ self.pad_token_id, self.sep_token_id]],
+ ).to(caps.device).repeat(vfeats.size(0), 1)
+ dummy_cmasks = torch.BoolTensor(
+ [[0, 1, 0, 1]] # pad are valid for attention.
+ ).to(caps.device).repeat(vfeats.size(0), 1)
+
+ # video forwarding, text is dummy; never use attention_mask.
+ attention_mask, token_type_ids = self._mm_on_the_fly(
+ dummy_cmasks, vmasks, None)
+
+ outputs = self.mm_encoder(
+ input_ids=dummy_caps,
+ input_video_embeds=vfeats,
+ attention_mask=attention_mask,
+ token_type_ids=token_type_ids,
+ output_hidden_states=output_hidden_states,
+ )
+
+ layer_idx = self.last_iso_layer \
+ if self.last_iso_layer > 0 else self.num_hidden_layers
+
+ video_seq = outputs[2][layer_idx][:, 1:vmasks.size(1)+1].masked_select(
+ vmasks.unsqueeze(-1)
+ ).view(-1, self.hidden_size)
+
+ # text forwarding, video is dummy
+ attention_mask, token_type_ids = self._mm_on_the_fly(
+ cmasks, dummy_vmasks, None)
+
+ outputs = self.mm_encoder(
+ input_ids=caps,
+ input_video_embeds=dummy_vfeats,
+ attention_mask=attention_mask,
+ token_type_ids=token_type_ids,
+ output_hidden_states=output_hidden_states,
+ )
+
+ _, pooled_text = self._pooling_vt_layer(
+ outputs[2], cmasks, dummy_vmasks)
+ # this line is not right.
+ logits = torch.mm(video_seq, pooled_text.transpose(1, 0))
+ return {"logits": logits}
+
+
+# --------------- MMFusionSeparate for end tasks ---------------
+
+class MMFusionSeparateActionSegmentation(MMFusionSeparate):
+ """Fine-tuning wrapper for action segmentation."""
+ def forward(
+ self,
+ caps,
+ cmasks,
+ vfeats,
+ vmasks,
+ attention_mask=None,
+ **kwargs
+ ):
+ # ActionLocalization assume of batch_size=1, squeeze it.
+ caps = caps.view(-1, caps.size(-1))
+ cmasks = cmasks.view(-1, cmasks.size(-1))
+ vfeats = vfeats.view(-1, vfeats.size(2), vfeats.size(3))
+ vmasks = vmasks.view(-1, vmasks.size(-1))
+ logits = self.forward_video(
+ vfeats,
+ vmasks,
+ caps,
+ cmasks,
+ output_hidden_states=True
+ )
+ return {"logits": logits[:, 1:vmasks.size(1)+1]}
+
+
+class MMFusionSeparateActionLocalization(MMFusionSeparate):
+ def __init__(self, config, **kwargs):
+ super().__init__(config)
+ tokenizer = AutoTokenizer.from_pretrained(
+ config.dataset.bert_name)
+ self.cls_token_id = tokenizer.cls_token_id
+ self.sep_token_id = tokenizer.sep_token_id
+ self.pad_token_id = tokenizer.pad_token_id
+
+ def forward(
+ self,
+ caps,
+ cmasks,
+ vfeats,
+ vmasks,
+ **kwargs
+ ):
+ # ActionLocalization assume of batch_size=1, squeeze it.
+ caps = caps.squeeze(0)
+ cmasks = cmasks.squeeze(0)
+ vfeats = vfeats.squeeze(0)
+ vmasks = vmasks.squeeze(0)
+
+ # TODO (huxu): other ways to do negative examples; move the following
+ # into your criterion forward.
+ dummy_caps = torch.LongTensor(
+ [[self.cls_token_id, self.sep_token_id,
+ self.pad_token_id, self.sep_token_id]],
+ ).to(caps.device).repeat(vfeats.size(0), 1)
+ dummy_cmasks = torch.BoolTensor(
+ [[0, 1, 0, 1]] # pad are valid for attention.
+ ).to(caps.device).repeat(vfeats.size(0), 1)
+
+ outputs = self.forward_video(
+ vfeats,
+ vmasks,
+ dummy_caps,
+ dummy_cmasks,
+ output_hidden_states=True
+ )
+
+ video_seq = outputs[:, 1:vmasks.size(1)+1].masked_select(
+ vmasks.unsqueeze(-1)
+ ).view(-1, self.hidden_size)
+
+ pooled_text = self.forward_text(
+ caps,
+ cmasks,
+ output_hidden_states=False
+ )
+
+ # this line is not right.
+ logits = torch.mm(video_seq, pooled_text.transpose(1, 0))
+ return {"logits": logits}
+
+
+class MMFusionShareActionLocalization(MMFusionShare):
+ def __init__(self, config, **kwargs):
+ super().__init__(config)
+ tokenizer = AutoTokenizer.from_pretrained(
+ config.dataset.bert_name)
+ self.cls_token_id = tokenizer.cls_token_id
+ self.sep_token_id = tokenizer.sep_token_id
+ self.pad_token_id = tokenizer.pad_token_id
+
+ def forward(
+ self,
+ caps,
+ cmasks,
+ vfeats,
+ vmasks,
+ **kwargs
+ ):
+ # ActionLocalization assume of batch_size=1, squeeze it.
+ caps = caps.squeeze(0)
+ cmasks = cmasks.squeeze(0)
+ vfeats = vfeats.squeeze(0)
+ vmasks = vmasks.squeeze(0)
+
+ # TODO (huxu): other ways to do negative examples; move the following
+ # into your criterion forward.
+ dummy_caps = torch.LongTensor(
+ [[self.cls_token_id, self.sep_token_id,
+ self.pad_token_id, self.sep_token_id]],
+ ).to(caps.device).repeat(vfeats.size(0), 1)
+ dummy_cmasks = torch.BoolTensor(
+ [[0, 1, 0, 1]] # pad are valid for attention.
+ ).to(caps.device).repeat(vfeats.size(0), 1)
+
+ outputs = self.forward_video(
+ vfeats,
+ vmasks,
+ dummy_caps,
+ dummy_cmasks,
+ output_hidden_states=True
+ )
+
+ video_seq = outputs[:, 1:vmasks.size(1)+1].masked_select(
+ vmasks.unsqueeze(-1)
+ ).view(-1, self.hidden_size)
+
+ pooled_text = self.forward_text(
+ caps,
+ cmasks,
+ output_hidden_states=False
+ )
+
+ # this line is not right.
+ logits = torch.mm(video_seq, pooled_text.transpose(1, 0))
+ return {"logits": logits}
diff --git a/fairseq/examples/MMPT/mmpt/models/mmfusionnlg.py b/fairseq/examples/MMPT/mmpt/models/mmfusionnlg.py
new file mode 100644
index 0000000000000000000000000000000000000000..9207e77dab3025d7a26efcce0795183de1d34fc7
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/models/mmfusionnlg.py
@@ -0,0 +1,999 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors, Facebook AI Research authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# Copyright (c) Facebook, Inc. All Rights Reserved
+
+
+import torch
+
+from torch.nn import functional as F
+
+from typing import Optional, Iterable
+
+try:
+ from transformers import BertPreTrainedModel
+ from transformers.modeling_bert import BertOnlyMLMHead
+
+ from transformers.file_utils import ModelOutput
+ from transformers.modeling_outputs import CausalLMOutput
+ from transformers.generation_utils import (
+ BeamHypotheses,
+ top_k_top_p_filtering
+ )
+except ImportError:
+ pass
+
+from .mmfusion import MMFusion
+from .transformermodel import MMBertModel
+from ..modules import VideoTokenMLP
+
+
+class MMFusionNLG(MMFusion):
+ def __init__(self, config, **kwargs):
+ super().__init__(config)
+ if config.model.max_decode_length is not None:
+ self.max_length = min(
+ config.model.max_decode_length,
+ config.dataset.max_len - config.dataset.max_video_len - 3
+ )
+ else:
+ self.max_length = \
+ config.dataset.max_len - config.dataset.max_video_len - 3
+ self.gen_param = config.gen_param if config.gen_param is not None \
+ else {}
+
+ def forward(
+ self,
+ caps,
+ cmasks,
+ vfeats,
+ vmasks,
+ attention_mask,
+ video_label=None,
+ text_label=None,
+ **kwargs
+ ):
+ """use pre-trained LM header for generation."""
+ attention_mask, token_type_ids = self._mm_on_the_fly(
+ cmasks, vmasks, attention_mask)
+
+ outputs = self.mm_encoder(
+ input_ids=caps,
+ input_video_embeds=vfeats,
+ attention_mask=attention_mask,
+ token_type_ids=token_type_ids,
+ masked_lm_labels=text_label,
+ )
+ return {"logits": outputs[0]}
+
+ @torch.no_grad()
+ def generate(
+ self,
+ caps, cmasks, vfeats, vmasks,
+ attention_mask=None,
+ bos_token_id=None,
+ eos_token_id=None,
+ **kwargs
+ ):
+ # a simplified interface from
+ # https://huggingface.co/transformers/v3.4.0/_modules/transformers/generation_utils.html#GenerationMixin.generate
+
+ # caps now only have
+ # [CLS], [SEP] (for video) and [CLS] (as bos_token)
+ assert caps.size(1) == 3
+
+ attention_mask, token_type_ids = self._mm_on_the_fly(
+ cmasks, vmasks, attention_mask)
+
+ output = self.mm_encoder.generate(
+ input_ids=caps,
+ input_video_embeds=vfeats,
+ attention_mask=attention_mask,
+ token_type_ids=token_type_ids,
+ bos_token_id=bos_token_id,
+ eos_token_id=eos_token_id,
+ max_length=self.max_length,
+ **self.gen_param
+ )
+ return output
+
+
+class MMBertForNLG(BertPreTrainedModel):
+ def __init__(self, config):
+ super().__init__(config)
+ self.bert = MMBertModel(config)
+ self.videomlp = VideoTokenMLP(config)
+ # we do not use `BertGenerationOnlyLMHead`
+ # because we can reuse pretraining.
+ self.cls = BertOnlyMLMHead(config)
+ self.hidden_size = config.hidden_size
+ self.init_weights()
+
+ def get_output_embeddings(self):
+ return self.cls.predictions.decoder
+
+ def forward(
+ self,
+ input_ids=None,
+ input_video_embeds=None,
+ attention_mask=None,
+ token_type_ids=None,
+ position_ids=None,
+ head_mask=None,
+ inputs_embeds=None,
+ masked_lm_labels=None,
+ output_attentions=None,
+ output_hidden_states=None,
+ return_dict=None,
+ ):
+ # similar to MMBertForMFMMLM without MFM.
+ video_tokens = self.videomlp(input_video_embeds)
+ outputs = self.bert(
+ input_ids,
+ video_tokens,
+ attention_mask=attention_mask,
+ token_type_ids=token_type_ids,
+ position_ids=position_ids,
+ head_mask=head_mask,
+ inputs_embeds=inputs_embeds,
+ output_attentions=output_attentions,
+ output_hidden_states=output_hidden_states,
+ return_dict=return_dict,
+ )
+
+ sequence_output = outputs[0]
+
+ prediction_scores = None
+ if masked_lm_labels is not None:
+ text_offset = input_video_embeds.size(1) + 1 # [CLS]
+ # recover caps format: [CLS] [SEP] text [SEP]
+ text_sequence_output = torch.cat(
+ [sequence_output[:, :1], sequence_output[:, text_offset:]],
+ dim=1
+ )
+
+ # only compute select tokens to training to speed up.
+ hidden_size = text_sequence_output.size(-1)
+ # masked_lm_labels = masked_lm_labels.reshape(-1)
+ labels_mask = masked_lm_labels != -100
+
+ selected_text_output = text_sequence_output.masked_select(
+ labels_mask.unsqueeze(-1)
+ ).view(-1, hidden_size)
+ prediction_scores = self.cls(selected_text_output)
+
+ if not return_dict:
+ output = (
+ prediction_scores,
+ ) + outputs[2:]
+ return output
+
+ # for generation.
+ text_offset = input_video_embeds.size(1) + 2 # [CLS]
+ text_sequence_output = sequence_output[:, text_offset:]
+ prediction_scores = self.cls(text_sequence_output)
+ return CausalLMOutput(
+ loss=None,
+ logits=prediction_scores,
+ )
+
+ def prepare_inputs_for_generation(
+ self,
+ input_ids,
+ input_video_embeds,
+ attention_mask=None,
+ token_type_ids=None,
+ **model_kwargs
+ ):
+ # must return a dictionary.
+ seq_len = input_ids.size(1) + input_video_embeds.size(1)
+ if attention_mask is not None:
+ if len(attention_mask.size()) == 4:
+ attention_mask = attention_mask[:, :, :seq_len, :seq_len]
+ elif len(attention_mask.size()) == 3:
+ attention_mask = attention_mask[:, :seq_len, :seq_len]
+ else:
+ attention_mask = attention_mask[:, :seq_len]
+ if token_type_ids is not None:
+ token_type_ids = token_type_ids[:, :seq_len]
+
+ return {
+ "input_ids": input_ids,
+ "input_video_embeds": input_video_embeds,
+ "attention_mask": attention_mask,
+ "token_type_ids": token_type_ids,
+ }
+
+ @torch.no_grad()
+ def generate(
+ self,
+ input_ids: Optional[torch.LongTensor] = None,
+ decoder_input_ids: Optional[torch.LongTensor] = None,
+ max_length: Optional[int] = None,
+ min_length: Optional[int] = None,
+ do_sample: Optional[bool] = None,
+ early_stopping: Optional[bool] = None,
+ num_beams: Optional[int] = None,
+ temperature: Optional[float] = None,
+ top_k: Optional[int] = None,
+ top_p: Optional[float] = None,
+ repetition_penalty: Optional[float] = None,
+ bad_words_ids: Optional[Iterable[int]] = None,
+ bos_token_id: Optional[int] = None,
+ pad_token_id: Optional[int] = None,
+ eos_token_id: Optional[int] = None,
+ length_penalty: Optional[float] = None,
+ no_repeat_ngram_size: Optional[int] = None,
+ num_return_sequences: Optional[int] = None,
+ attention_mask: Optional[torch.LongTensor] = None,
+ decoder_start_token_id: Optional[int] = None,
+ use_cache: Optional[bool] = None,
+ **model_kwargs
+ ) -> torch.LongTensor:
+ r"""
+ Generates sequences for models with a language modeling head. The method currently supports greedy decoding,
+ beam-search decoding, sampling with temperature, sampling with top-k or nucleus sampling.
+ Adapted in part from `Facebook's XLM beam search code
+ `__.
+ Apart from :obj:`input_ids` and :obj:`attention_mask`, all the arguments below will default to the value of the
+ attribute of the same name inside the :class:`~transformers.PretrainedConfig` of the model. The default values
+ indicated are the default values of those config.
+ Most of these parameters are explained in more detail in `this blog post
+ `__.
+ Parameters:
+ input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+ The sequence used as a prompt for the generation. If :obj:`None` the method initializes
+ it as an empty :obj:`torch.LongTensor` of shape :obj:`(1,)`.
+ decoder_input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+ initial input_ids for the decoder of encoder-decoder type models. If :obj:`None` then only
+ decoder_start_token_id is passed as the first token to the decoder.
+ max_length (:obj:`int`, `optional`, defaults to 20):
+ The maximum length of the sequence to be generated.
+ min_length (:obj:`int`, `optional`, defaults to 10):
+ The minimum length of the sequence to be generated.
+ do_sample (:obj:`bool`, `optional`, defaults to :obj:`False`):
+ Whether or not to use sampling ; use greedy decoding otherwise.
+ early_stopping (:obj:`bool`, `optional`, defaults to :obj:`False`):
+ Whether to stop the beam search when at least ``num_beams`` sentences are finished per batch or not.
+ num_beams (:obj:`int`, `optional`, defaults to 1):
+ Number of beams for beam search. 1 means no beam search.
+ temperature (:obj:`float`, `optional`, defaults tp 1.0):
+ The value used to module the next token probabilities.
+ top_k (:obj:`int`, `optional`, defaults to 50):
+ The number of highest probability vocabulary tokens to keep for top-k-filtering.
+ top_p (:obj:`float`, `optional`, defaults to 1.0):
+ If set to float < 1, only the most probable tokens with probabilities that add up to ``top_p`` or
+ higher are kept for generation.
+ repetition_penalty (:obj:`float`, `optional`, defaults to 1.0):
+ The parameter for repetition penalty. 1.0 means no penalty. See `this paper
+ `__ for more details.
+ pad_token_id (:obj:`int`, `optional`):
+ The id of the `padding` token.
+ bos_token_id (:obj:`int`, `optional`):
+ The id of the `beginning-of-sequence` token.
+ eos_token_id (:obj:`int`, `optional`):
+ The id of the `end-of-sequence` token.
+ length_penalty (:obj:`float`, `optional`, defaults to 1.0):
+ Exponential penalty to the length. 1.0 means no penalty.
+ Set to values < 1.0 in order to encourage the model to generate shorter sequences, to a value > 1.0 in
+ order to encourage the model to produce longer sequences.
+ no_repeat_ngram_size (:obj:`int`, `optional`, defaults to 0):
+ If set to int > 0, all ngrams of that size can only occur once.
+ bad_words_ids(:obj:`List[int]`, `optional`):
+ List of token ids that are not allowed to be generated. In order to get the tokens of the words that
+ should not appear in the generated text, use :obj:`tokenizer.encode(bad_word, add_prefix_space=True)`.
+ num_return_sequences(:obj:`int`, `optional`, defaults to 1):
+ The number of independently computed returned sequences for each element in the batch.
+ attention_mask (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+ Mask to avoid performing attention on padding token indices. Mask values are in ``[0, 1]``, 1 for
+ tokens that are not masked, and 0 for masked tokens.
+ If not provided, will default to a tensor the same shape as :obj:`input_ids` that masks the pad token.
+ `What are attention masks? <../glossary.html#attention-mask>`__
+ decoder_start_token_id (:obj:`int`, `optional`):
+ If an encoder-decoder model starts decoding with a different token than `bos`, the id of that token.
+ use_cache: (:obj:`bool`, `optional`, defaults to :obj:`True`):
+ Whether or not the model should use the past last key/values attentions (if applicable to the model) to
+ speed up decoding.
+ model_kwargs:
+ Additional model specific kwargs will be forwarded to the :obj:`forward` function of the model.
+ Return:
+ :obj:`torch.LongTensor` of shape :obj:`(batch_size * num_return_sequences, sequence_length)`:
+ The generated sequences. The second dimension (sequence_length) is either equal to :obj:`max_length` or
+ shorter if all batches finished early due to the :obj:`eos_token_id`.
+ Examples::
+ tokenizer = AutoTokenizer.from_pretrained('distilgpt2') # Initialize tokenizer
+ model = AutoModelWithLMHead.from_pretrained('distilgpt2') # Download model and configuration from S3 and cache.
+ outputs = model.generate(max_length=40) # do greedy decoding
+ print('Generated: {}'.format(tokenizer.decode(outputs[0], skip_special_tokens=True)))
+ tokenizer = AutoTokenizer.from_pretrained('openai-gpt') # Initialize tokenizer
+ model = AutoModelWithLMHead.from_pretrained('openai-gpt') # Download model and configuration from S3 and cache.
+ input_context = 'The dog'
+ input_ids = tokenizer.encode(input_context, return_tensors='pt') # encode input context
+ outputs = model.generate(input_ids=input_ids, num_beams=5, num_return_sequences=3, temperature=1.5) # generate 3 independent sequences using beam search decoding (5 beams) with sampling from initial context 'The dog'
+ for i in range(3): # 3 output sequences were generated
+ print('Generated {}: {}'.format(i, tokenizer.decode(outputs[i], skip_special_tokens=True)))
+ tokenizer = AutoTokenizer.from_pretrained('distilgpt2') # Initialize tokenizer
+ model = AutoModelWithLMHead.from_pretrained('distilgpt2') # Download model and configuration from S3 and cache.
+ input_context = 'The dog'
+ input_ids = tokenizer.encode(input_context, return_tensors='pt') # encode input context
+ outputs = model.generate(input_ids=input_ids, max_length=40, temperature=0.7, num_return_sequences=3, do_sample=True) # generate 3 candidates using sampling
+ for i in range(3): # 3 output sequences were generated
+ print('Generated {}: {}'.format(i, tokenizer.decode(outputs[i], skip_special_tokens=True)))
+ tokenizer = AutoTokenizer.from_pretrained('ctrl') # Initialize tokenizer
+ model = AutoModelWithLMHead.from_pretrained('ctrl') # Download model and configuration from S3 and cache.
+ input_context = 'Legal My neighbor is' # "Legal" is one of the control codes for ctrl
+ input_ids = tokenizer.encode(input_context, return_tensors='pt') # encode input context
+ outputs = model.generate(input_ids=input_ids, max_length=50, temperature=0.7, repetition_penalty=1.2) # generate sequences
+ print('Generated: {}'.format(tokenizer.decode(outputs[0], skip_special_tokens=True)))
+ tokenizer = AutoTokenizer.from_pretrained('gpt2') # Initialize tokenizer
+ model = AutoModelWithLMHead.from_pretrained('gpt2') # Download model and configuration from S3 and cache.
+ input_context = 'My cute dog' # "Legal" is one of the control codes for ctrl
+ bad_words_ids = [tokenizer.encode(bad_word, add_prefix_space=True) for bad_word in ['idiot', 'stupid', 'shut up']]
+ input_ids = tokenizer.encode(input_context, return_tensors='pt') # encode input context
+ outputs = model.generate(input_ids=input_ids, max_length=100, do_sample=True, bad_words_ids=bad_words_ids) # generate sequences without allowing bad_words to be generated
+ """
+
+ # We cannot generate if the model does not have a LM head
+ if self.get_output_embeddings() is None:
+ raise AttributeError(
+ "You tried to generate sequences with a model that does not have a LM Head."
+ "Please use another model class (e.g. `OpenAIGPTLMHeadModel`, `XLNetLMHeadModel`, `GPT2LMHeadModel`, `CTRLLMHeadModel`, `T5WithLMHeadModel`, `TransfoXLLMHeadModel`, `XLMWithLMHeadModel`, `BartForConditionalGeneration` )"
+ )
+
+ max_length = max_length if max_length is not None else self.config.max_length
+ min_length = min_length if min_length is not None else self.config.min_length
+ do_sample = do_sample if do_sample is not None else self.config.do_sample
+ early_stopping = early_stopping if early_stopping is not None else self.config.early_stopping
+ use_cache = use_cache if use_cache is not None else self.config.use_cache
+ num_beams = num_beams if num_beams is not None else self.config.num_beams
+ temperature = temperature if temperature is not None else self.config.temperature
+ top_k = top_k if top_k is not None else self.config.top_k
+ top_p = top_p if top_p is not None else self.config.top_p
+ repetition_penalty = repetition_penalty if repetition_penalty is not None else self.config.repetition_penalty
+ bos_token_id = bos_token_id if bos_token_id is not None else self.config.bos_token_id
+ pad_token_id = pad_token_id if pad_token_id is not None else self.config.pad_token_id
+ eos_token_id = eos_token_id if eos_token_id is not None else self.config.eos_token_id
+ length_penalty = length_penalty if length_penalty is not None else self.config.length_penalty
+ no_repeat_ngram_size = (
+ no_repeat_ngram_size if no_repeat_ngram_size is not None else self.config.no_repeat_ngram_size
+ )
+ bad_words_ids = bad_words_ids if bad_words_ids is not None else self.config.bad_words_ids
+ num_return_sequences = (
+ num_return_sequences if num_return_sequences is not None else self.config.num_return_sequences
+ )
+ decoder_start_token_id = (
+ decoder_start_token_id if decoder_start_token_id is not None else self.config.decoder_start_token_id
+ )
+
+ if input_ids is not None:
+ batch_size = input_ids.shape[0] # overriden by the input batch_size
+ else:
+ batch_size = 1
+
+ assert isinstance(max_length, int) and max_length > 0, "`max_length` should be a strictly positive integer."
+ assert isinstance(min_length, int) and min_length >= 0, "`min_length` should be a positive integer."
+ assert isinstance(do_sample, bool), "`do_sample` should be a boolean."
+ assert isinstance(early_stopping, bool), "`early_stopping` should be a boolean."
+ assert isinstance(use_cache, bool), "`use_cache` should be a boolean."
+ assert isinstance(num_beams, int) and num_beams > 0, "`num_beams` should be a strictly positive integer."
+ assert temperature > 0, "`temperature` should be strictly positive."
+ assert isinstance(top_k, int) and top_k >= 0, "`top_k` should be a positive integer."
+ assert 0 <= top_p <= 1, "`top_p` should be between 0 and 1."
+ assert repetition_penalty >= 1.0, "`repetition_penalty` should be >= 1."
+ assert input_ids is not None or (
+ isinstance(bos_token_id, int) and bos_token_id >= 0
+ ), "If input_ids is not defined, `bos_token_id` should be a positive integer."
+ assert pad_token_id is None or (
+ isinstance(pad_token_id, int) and (pad_token_id >= 0)
+ ), "`pad_token_id` should be a positive integer."
+ assert (eos_token_id is None) or (
+ isinstance(eos_token_id, int) and (eos_token_id >= 0)
+ ), "`eos_token_id` should be a positive integer."
+ assert length_penalty > 0, "`length_penalty` should be strictly positive."
+ assert (
+ isinstance(no_repeat_ngram_size, int) and no_repeat_ngram_size >= 0
+ ), "`no_repeat_ngram_size` should be a positive integer."
+ assert (
+ isinstance(num_return_sequences, int) and num_return_sequences > 0
+ ), "`num_return_sequences` should be a strictly positive integer."
+ assert (
+ bad_words_ids is None or isinstance(bad_words_ids, list) and isinstance(bad_words_ids[0], list)
+ ), "`bad_words_ids` is either `None` or a list of lists of tokens that should not be generated"
+
+ if input_ids is None:
+ assert isinstance(bos_token_id, int) and bos_token_id >= 0, (
+ "you should either supply a context to complete as `input_ids` input "
+ "or a `bos_token_id` (integer >= 0) as a first token to start the generation."
+ )
+ input_ids = torch.full(
+ (batch_size, 1),
+ bos_token_id,
+ dtype=torch.long,
+ device=next(self.parameters()).device,
+ )
+ else:
+ assert input_ids.dim() == 2, "Input prompt should be of shape (batch_size, sequence length)."
+
+ # not allow to duplicate outputs when greedy decoding
+ if do_sample is False:
+ if num_beams == 1:
+ # no_beam_search greedy generation conditions
+ assert (
+ num_return_sequences == 1
+ ), "Greedy decoding will always produce the same output for num_beams == 1 and num_return_sequences > 1. Please set num_return_sequences = 1"
+
+ else:
+ # beam_search greedy generation conditions
+ assert (
+ num_beams >= num_return_sequences
+ ), "Greedy beam search decoding cannot return more sequences than it has beams. Please set num_beams >= num_return_sequences"
+
+ # create attention mask if necessary
+ # TODO (PVP): this should later be handled by the forward fn() in each model in the future see PR 3140
+ if (attention_mask is None) and (pad_token_id is not None) and (pad_token_id in input_ids):
+ attention_mask = input_ids.ne(pad_token_id).long()
+ elif attention_mask is None:
+ attention_mask = input_ids.new_ones(input_ids.shape)
+
+ # set pad_token_id to eos_token_id if not set. Important that this is done after
+ # attention_mask is created
+ if pad_token_id is None and eos_token_id is not None:
+ print(
+ "Setting `pad_token_id` to {} (first `eos_token_id`) to generate sequence".format(eos_token_id)
+ )
+ pad_token_id = eos_token_id
+
+ # vocab size
+ if hasattr(self.config, "vocab_size"):
+ vocab_size = self.config.vocab_size
+ elif (
+ self.config.is_encoder_decoder
+ and hasattr(self.config, "decoder")
+ and hasattr(self.config.decoder, "vocab_size")
+ ):
+ vocab_size = self.config.decoder.vocab_size
+ else:
+ raise ValueError("either self.config.vocab_size or self.config.decoder.vocab_size needs to be defined")
+
+ # set effective batch size and effective batch multiplier according to do_sample
+ if do_sample:
+ effective_batch_size = batch_size * num_return_sequences
+ effective_batch_mult = num_return_sequences
+ else:
+ effective_batch_size = batch_size
+ effective_batch_mult = 1
+
+ if self.config.is_encoder_decoder:
+ if decoder_start_token_id is None:
+ # see if BOS token can be used for decoder_start_token_id
+ if bos_token_id is not None:
+ decoder_start_token_id = bos_token_id
+ elif (
+ hasattr(self.config, "decoder")
+ and hasattr(self.config.decoder, "bos_token_id")
+ and self.config.decoder.bos_token_id is not None
+ ):
+ decoder_start_token_id = self.config.decoder.bos_token_id
+ else:
+ raise ValueError(
+ "decoder_start_token_id or bos_token_id has to be defined for encoder-decoder generation"
+ )
+
+ assert hasattr(self, "get_encoder"), "{} should have a 'get_encoder' function defined".format(self)
+ assert callable(self.get_encoder), "{} should be a method".format(self.get_encoder)
+
+ # get encoder and store encoder outputs
+ encoder = self.get_encoder()
+ encoder_outputs: ModelOutput = encoder(input_ids, attention_mask=attention_mask, return_dict=True)
+
+ # Expand input ids if num_beams > 1 or num_return_sequences > 1
+ if num_return_sequences > 1 or num_beams > 1:
+ # TODO: make this a call-back function.
+ # input_ids=caps,
+ # input_video_embeds=vfeats,
+ # attention_mask=attention_mask,
+ # token_type_ids=token_type_ids,
+ input_video_embeds = model_kwargs.pop("input_video_embeds", None)
+ token_type_ids = model_kwargs.pop("token_type_ids", None)
+
+ input_ids_len = input_ids.shape[-1]
+ input_ids = input_ids.unsqueeze(1).expand(
+ batch_size, effective_batch_mult * num_beams, input_ids_len)
+
+ input_video_embeds_len, input_video_embeds_hidden = input_video_embeds.size(1), input_video_embeds.size(2)
+ input_video_embeds = input_video_embeds.unsqueeze(1).expand(
+ batch_size, effective_batch_mult * num_beams, input_video_embeds_len, input_video_embeds_hidden)
+
+ attention_mask_from_len, attention_mask_to_len = attention_mask.size(1), attention_mask.size(2)
+ attention_mask = attention_mask.unsqueeze(1).expand(
+ batch_size, effective_batch_mult * num_beams, attention_mask_from_len, attention_mask_to_len
+ )
+
+ token_type_ids_len = token_type_ids.size(1)
+ token_type_ids = token_type_ids.unsqueeze(1).expand(
+ batch_size, effective_batch_mult * num_beams, token_type_ids_len
+ )
+
+ # contiguous ...
+ input_ids = input_ids.contiguous().view(
+ effective_batch_size * num_beams, input_ids_len
+ ) # shape: (batch_size * num_return_sequences * num_beams, cur_len)
+
+ input_video_embeds = input_video_embeds.contiguous().view(
+ effective_batch_size * num_beams, input_video_embeds_len, input_video_embeds_hidden)
+
+ attention_mask = attention_mask.contiguous().view(
+ effective_batch_size * num_beams, attention_mask_from_len, attention_mask_to_len
+ ) # shape: (batch_size * num_return_sequences * num_beams, cur_len)
+
+ token_type_ids = token_type_ids.contiguous().view(
+ effective_batch_size * num_beams, token_type_ids_len
+ )
+
+ model_kwargs["input_video_embeds"] = input_video_embeds
+ model_kwargs["token_type_ids"] = token_type_ids
+
+ if self.config.is_encoder_decoder:
+ device = next(self.parameters()).device
+ if decoder_input_ids is not None:
+ # give initial decoder input ids
+ input_ids = decoder_input_ids.repeat(effective_batch_size * num_beams, 1).to(device)
+ else:
+ # create empty decoder input_ids
+ input_ids = torch.full(
+ (effective_batch_size * num_beams, 1),
+ decoder_start_token_id,
+ dtype=torch.long,
+ device=device,
+ )
+ cur_len = input_ids.shape[-1]
+
+ assert (
+ batch_size == encoder_outputs.last_hidden_state.shape[0]
+ ), f"expected encoder_outputs.last_hidden_state to have 1st dimension bs={batch_size}, got {encoder_outputs.last_hidden_state.shape[0]} "
+
+ # expand batch_idx to assign correct encoder output for expanded input_ids (due to num_beams > 1 and num_return_sequences > 1)
+ expanded_batch_idxs = (
+ torch.arange(batch_size)
+ .view(-1, 1)
+ .repeat(1, num_beams * effective_batch_mult)
+ .view(-1)
+ .to(input_ids.device)
+ )
+
+ # expand encoder_outputs
+ encoder_outputs["last_hidden_state"] = encoder_outputs.last_hidden_state.index_select(
+ 0, expanded_batch_idxs
+ )
+
+ # save encoder_outputs in `model_kwargs`
+ model_kwargs["encoder_outputs"] = encoder_outputs
+
+ else:
+ cur_len = input_ids.shape[-1]
+
+ assert (
+ cur_len < max_length
+ ), f"The context has {cur_len} number of tokens, but `max_length` is only {max_length}. Please make sure that `max_length` is bigger than the number of tokens, by setting either `generate(max_length=...,...)` or `config.max_length = ...`"
+
+ if num_beams > 1:
+ output = self._generate_beam_search(
+ input_ids,
+ cur_len=cur_len,
+ max_length=max_length,
+ min_length=min_length,
+ do_sample=do_sample,
+ early_stopping=early_stopping,
+ temperature=temperature,
+ top_k=top_k,
+ top_p=top_p,
+ repetition_penalty=repetition_penalty,
+ no_repeat_ngram_size=no_repeat_ngram_size,
+ bad_words_ids=bad_words_ids,
+ pad_token_id=pad_token_id,
+ eos_token_id=eos_token_id,
+ batch_size=effective_batch_size,
+ num_return_sequences=num_return_sequences,
+ length_penalty=length_penalty,
+ num_beams=num_beams,
+ vocab_size=vocab_size,
+ attention_mask=attention_mask,
+ use_cache=use_cache,
+ model_kwargs=model_kwargs,
+ )
+ else:
+ output = self._generate_no_beam_search(
+ input_ids,
+ cur_len=cur_len,
+ max_length=max_length,
+ min_length=min_length,
+ do_sample=do_sample,
+ temperature=temperature,
+ top_k=top_k,
+ top_p=top_p,
+ repetition_penalty=repetition_penalty,
+ no_repeat_ngram_size=no_repeat_ngram_size,
+ bad_words_ids=bad_words_ids,
+ pad_token_id=pad_token_id,
+ eos_token_id=eos_token_id,
+ batch_size=effective_batch_size,
+ attention_mask=attention_mask,
+ use_cache=use_cache,
+ model_kwargs=model_kwargs,
+ )
+
+ return output
+
+ def _generate_beam_search(
+ self,
+ input_ids,
+ cur_len,
+ max_length,
+ min_length,
+ do_sample,
+ early_stopping,
+ temperature,
+ top_k,
+ top_p,
+ repetition_penalty,
+ no_repeat_ngram_size,
+ bad_words_ids,
+ pad_token_id,
+ eos_token_id,
+ batch_size,
+ num_return_sequences,
+ length_penalty,
+ num_beams,
+ vocab_size,
+ attention_mask,
+ use_cache,
+ model_kwargs,
+ ):
+ """Generate sequences for each example with beam search."""
+
+ # generated hypotheses
+ generated_hyps = [
+ BeamHypotheses(num_beams, max_length, length_penalty, early_stopping=early_stopping)
+ for _ in range(batch_size)
+ ]
+
+ # scores for each sentence in the beam
+ beam_scores = torch.zeros((batch_size, num_beams), dtype=torch.float, device=input_ids.device)
+
+ # for greedy decoding it is made sure that only tokens of the first beam are considered to avoid sampling the exact same tokens three times
+ if do_sample is False:
+ beam_scores[:, 1:] = -1e9
+ beam_scores = beam_scores.view(-1) # shape (batch_size * num_beams,)
+
+ # cache compute states
+ past = None
+
+ # done sentences
+ done = [False for _ in range(batch_size)]
+
+ while cur_len < max_length:
+ model_inputs = self.prepare_inputs_for_generation(
+ input_ids, past=past, attention_mask=attention_mask, use_cache=use_cache, **model_kwargs
+ )
+ outputs = self(**model_inputs, return_dict=True) # (batch_size * num_beams, cur_len, vocab_size)
+ next_token_logits = outputs.logits[:, -1, :] # (batch_size * num_beams, vocab_size)
+
+ # if model has past, then set the past variable to speed up decoding
+ if "past_key_values" in outputs:
+ past = outputs.past_key_values
+ elif "mems" in outputs:
+ past = outputs.mems
+
+ if self.config.is_encoder_decoder and do_sample is False:
+ # TODO (PVP) still a bit hacky here - there might be a better solution
+ next_token_logits = self.adjust_logits_during_generation(
+ next_token_logits, cur_len=cur_len, max_length=max_length
+ )
+
+ scores = F.log_softmax(next_token_logits, dim=-1) # (batch_size * num_beams, vocab_size)
+
+ scores = self.postprocess_next_token_scores(
+ scores=scores,
+ input_ids=input_ids,
+ no_repeat_ngram_size=no_repeat_ngram_size,
+ bad_words_ids=bad_words_ids,
+ cur_len=cur_len,
+ min_length=min_length,
+ max_length=max_length,
+ eos_token_id=eos_token_id,
+ repetition_penalty=repetition_penalty,
+ batch_size=batch_size,
+ num_beams=num_beams,
+ )
+
+ assert scores.shape == (batch_size * num_beams, vocab_size), "Shapes of scores: {} != {}".format(
+ scores.shape, (batch_size * num_beams, vocab_size)
+ )
+
+ if do_sample:
+ _scores = scores + beam_scores[:, None].expand_as(scores) # (batch_size * num_beams, vocab_size)
+ # Temperature
+ if temperature != 1.0:
+ _scores = _scores / temperature
+ # Top-p/top-k filtering
+ _scores = top_k_top_p_filtering(
+ _scores, top_k=top_k, top_p=top_p, min_tokens_to_keep=2
+ ) # (batch_size * num_beams, vocab_size)
+ # re-organize to group the beam together to sample from all beam_idxs
+ _scores = _scores.contiguous().view(
+ batch_size, num_beams * vocab_size
+ ) # (batch_size, num_beams * vocab_size)
+
+ # Sample 2 next tokens for each beam (so we have some spare tokens and match output of greedy beam search)
+ probs = F.softmax(_scores, dim=-1)
+ next_tokens = torch.multinomial(probs, num_samples=2 * num_beams) # (batch_size, num_beams * 2)
+ # Compute next scores
+ next_scores = torch.gather(_scores, -1, next_tokens) # (batch_size, num_beams * 2)
+ # sort the sampled vector to make sure that the first num_beams samples are the best
+ next_scores, next_scores_indices = torch.sort(next_scores, descending=True, dim=1)
+ next_tokens = torch.gather(next_tokens, -1, next_scores_indices) # (batch_size, num_beams * 2)
+
+ else:
+ next_scores = scores + beam_scores[:, None].expand_as(scores) # (batch_size * num_beams, vocab_size)
+
+ # re-organize to group the beam together (we are keeping top hypothesis accross beams)
+ next_scores = next_scores.view(
+ batch_size, num_beams * vocab_size
+ ) # (batch_size, num_beams * vocab_size)
+
+ next_scores, next_tokens = torch.topk(next_scores, 2 * num_beams, dim=1, largest=True, sorted=True)
+
+ assert next_scores.size() == next_tokens.size() == (batch_size, 2 * num_beams)
+
+ # next batch beam content
+ next_batch_beam = []
+
+ # for each sentence
+ for batch_idx in range(batch_size):
+
+ # if we are done with this sentence, add a pad token
+ if done[batch_idx]:
+ assert (
+ len(generated_hyps[batch_idx]) >= num_beams
+ ), "Batch can only be done if at least {} beams have been generated".format(num_beams)
+ assert (
+ eos_token_id is not None and pad_token_id is not None
+ ), "generated beams >= num_beams -> eos_token_id and pad_token have to be defined"
+ next_batch_beam.extend([(0, pad_token_id, 0)] * num_beams) # pad the batch
+ continue
+
+ # next sentence beam content, this will get added to next_batch_beam
+ next_sent_beam = []
+
+ # next tokens for this sentence
+ for beam_token_rank, (beam_token_id, beam_token_score) in enumerate(
+ zip(next_tokens[batch_idx], next_scores[batch_idx])
+ ):
+ # get beam and token IDs
+ beam_id = beam_token_id // vocab_size
+ token_id = beam_token_id % vocab_size
+
+ effective_beam_id = batch_idx * num_beams + beam_id
+ # add to generated hypotheses if end of sentence
+ if (eos_token_id is not None) and (token_id.item() == eos_token_id):
+ # if beam_token does not belong to top num_beams tokens, it should not be added
+ is_beam_token_worse_than_top_num_beams = beam_token_rank >= num_beams
+ if is_beam_token_worse_than_top_num_beams:
+ continue
+ generated_hyps[batch_idx].add(
+ input_ids[effective_beam_id].clone(),
+ beam_token_score.item(),
+ )
+ else:
+ # add next predicted token since it is not eos_token
+ next_sent_beam.append((beam_token_score, token_id, effective_beam_id))
+
+ # once the beam for next step is full, don't add more tokens to it.
+ if len(next_sent_beam) == num_beams:
+ break
+
+ # Check if we are done so that we can save a pad step if all(done)
+ done[batch_idx] = done[batch_idx] or generated_hyps[batch_idx].is_done(
+ next_scores[batch_idx].max().item(), cur_len
+ )
+
+ # update next beam content
+ assert len(next_sent_beam) == num_beams, "Beam should always be full"
+ next_batch_beam.extend(next_sent_beam)
+ assert len(next_batch_beam) == num_beams * (batch_idx + 1), "We should have added num_beams each step"
+
+ # stop when we are done with each sentence
+ if all(done):
+ break
+
+ # sanity check / prepare next batch
+ assert len(next_batch_beam) == batch_size * num_beams
+ beam_scores = beam_scores.new([x[0] for x in next_batch_beam])
+ beam_tokens = input_ids.new([x[1] for x in next_batch_beam])
+ beam_idx = input_ids.new([x[2] for x in next_batch_beam])
+
+ # re-order batch and update current length
+ input_ids = input_ids[beam_idx, :]
+ input_ids = torch.cat([input_ids, beam_tokens.unsqueeze(1)], dim=-1)
+ cur_len = cur_len + 1
+
+ # re-order internal states
+ if past is not None:
+ past = self._reorder_cache(past, beam_idx)
+
+ # extend attention_mask for new generated input if only decoder
+ # (huxu): move out since we trim attention_mask by ourselves.
+ # if self.config.is_encoder_decoder is False:
+ # attention_mask = torch.cat(
+ # [attention_mask, attention_mask.new_ones((attention_mask.shape[0], 1))], dim=-1
+ # )
+
+ # finalize all open beam hypotheses and add to generated hypotheses
+ for batch_idx in range(batch_size):
+ if done[batch_idx]:
+ continue
+
+ # test that beam scores match previously calculated scores if not eos and batch_idx not done
+ if eos_token_id is not None and all(
+ (token_id % vocab_size).item() != eos_token_id for token_id in next_tokens[batch_idx]
+ ):
+ assert torch.all(
+ next_scores[batch_idx, :num_beams] == beam_scores.view(batch_size, num_beams)[batch_idx]
+ ), "If batch_idx is not done, final next scores: {} have to equal to accumulated beam_scores: {}".format(
+ next_scores[:, :num_beams][batch_idx],
+ beam_scores.view(batch_size, num_beams)[batch_idx],
+ )
+
+ # need to add best num_beams hypotheses to generated hyps
+ for beam_id in range(num_beams):
+ effective_beam_id = batch_idx * num_beams + beam_id
+ final_score = beam_scores[effective_beam_id].item()
+ final_tokens = input_ids[effective_beam_id]
+ generated_hyps[batch_idx].add(final_tokens, final_score)
+
+ # depending on whether greedy generation is wanted or not define different output_batch_size and output_num_return_sequences_per_batch
+ output_batch_size = batch_size if do_sample else batch_size * num_return_sequences
+ output_num_return_sequences_per_batch = 1 if do_sample else num_return_sequences
+
+ # select the best hypotheses
+ sent_lengths = input_ids.new(output_batch_size)
+ best = []
+
+ # retrieve best hypotheses
+ for i, hypotheses in enumerate(generated_hyps):
+ sorted_hyps = sorted(hypotheses.beams, key=lambda x: x[0])
+ for j in range(output_num_return_sequences_per_batch):
+ effective_batch_idx = output_num_return_sequences_per_batch * i + j
+ best_hyp = sorted_hyps.pop()[1]
+ sent_lengths[effective_batch_idx] = len(best_hyp)
+ best.append(best_hyp)
+
+ # prepare for adding eos
+ sent_max_len = min(sent_lengths.max().item() + 1, max_length)
+ decoded = input_ids.new(output_batch_size, sent_max_len)
+ # shorter batches are padded if needed
+ if sent_lengths.min().item() != sent_lengths.max().item():
+ assert pad_token_id is not None, "`pad_token_id` has to be defined"
+ decoded.fill_(pad_token_id)
+
+ # fill with hypotheses and eos_token_id if the latter fits in
+ for i, hypo in enumerate(best):
+ decoded[i, : sent_lengths[i]] = hypo
+ if sent_lengths[i] < max_length:
+ decoded[i, sent_lengths[i]] = eos_token_id
+
+ return decoded
+
+ def _generate_no_beam_search(
+ self,
+ input_ids,
+ cur_len,
+ max_length,
+ min_length,
+ do_sample,
+ temperature,
+ top_k,
+ top_p,
+ repetition_penalty,
+ no_repeat_ngram_size,
+ bad_words_ids,
+ pad_token_id,
+ eos_token_id,
+ batch_size,
+ attention_mask,
+ use_cache,
+ model_kwargs,
+ ):
+ """Generate sequences for each example without beam search (num_beams == 1).
+ All returned sequence are generated independantly.
+ """
+ # length of generated sentences / unfinished sentences
+ unfinished_sents = input_ids.new(batch_size).fill_(1)
+ sent_lengths = input_ids.new(batch_size).fill_(max_length)
+
+ past = None
+ while cur_len < max_length:
+ model_inputs = self.prepare_inputs_for_generation(
+ input_ids, past=past, attention_mask=attention_mask, use_cache=use_cache, **model_kwargs
+ )
+
+ outputs = self(**model_inputs, return_dict=True)
+ next_token_logits = outputs.logits[:, -1, :]
+ scores = self.postprocess_next_token_scores(
+ scores=next_token_logits,
+ input_ids=input_ids,
+ no_repeat_ngram_size=no_repeat_ngram_size,
+ bad_words_ids=bad_words_ids,
+ cur_len=cur_len,
+ min_length=min_length,
+ max_length=max_length,
+ eos_token_id=eos_token_id,
+ repetition_penalty=repetition_penalty,
+ batch_size=batch_size,
+ num_beams=1,
+ )
+
+ # if model has past, then set the past variable to speed up decoding
+ if "past_key_values" in outputs:
+ past = outputs.past_key_values
+ elif "mems" in outputs:
+ past = outputs.mems
+
+ if do_sample:
+ # Temperature (higher temperature => more likely to sample low probability tokens)
+ if temperature != 1.0:
+ scores = scores / temperature
+ # Top-p/top-k filtering
+ next_token_logscores = top_k_top_p_filtering(scores, top_k=top_k, top_p=top_p)
+ # Sample
+ probs = F.softmax(next_token_logscores, dim=-1)
+ next_token = torch.multinomial(probs, num_samples=1).squeeze(1)
+ else:
+ # Greedy decoding
+ next_token = torch.argmax(next_token_logits, dim=-1)
+
+ # print(next_token_logits[0,next_token[0]], next_token_logits[0,eos_token_id])
+
+ # update generations and finished sentences
+ if eos_token_id is not None:
+ # pad finished sentences if eos_token_id exist
+ tokens_to_add = next_token * unfinished_sents + (pad_token_id) * (1 - unfinished_sents)
+ else:
+ tokens_to_add = next_token
+
+ # add token and increase length by one
+ input_ids = torch.cat([input_ids, tokens_to_add.unsqueeze(-1)], dim=-1)
+ cur_len = cur_len + 1
+
+ if eos_token_id is not None:
+ eos_in_sents = tokens_to_add == eos_token_id
+ # if sentence is unfinished and the token to add is eos, sent_lengths is filled with current length
+ is_sents_unfinished_and_token_to_add_is_eos = unfinished_sents.mul(eos_in_sents.long()).bool()
+ sent_lengths.masked_fill_(is_sents_unfinished_and_token_to_add_is_eos, cur_len)
+ # unfinished_sents is set to zero if eos in sentence
+ unfinished_sents.mul_((~eos_in_sents).long())
+
+ # stop when there is a in each sentence, or if we exceed the maximul length
+ if unfinished_sents.max() == 0:
+ break
+
+
+ # extend attention_mask for new generated input if only decoder
+ # if self.config.is_encoder_decoder is False:
+ # attention_mask = torch.cat(
+ # [attention_mask, attention_mask.new_ones((attention_mask.shape[0], 1))], dim=-1
+ # )
+
+ return input_ids
diff --git a/fairseq/examples/MMPT/mmpt/models/transformermodel.py b/fairseq/examples/MMPT/mmpt/models/transformermodel.py
new file mode 100644
index 0000000000000000000000000000000000000000..6acc419f09edbd5c9007c4d33d517d59b2f79b77
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/models/transformermodel.py
@@ -0,0 +1,734 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# Copyright (c) Facebook, Inc. All Rights Reserved
+
+import torch
+
+from torch import nn
+
+try:
+ from transformers.modeling_bert import (
+ BertPreTrainedModel,
+ BertModel,
+ BertEncoder,
+ BertPredictionHeadTransform,
+ )
+except ImportError:
+ pass
+
+from ..modules import VideoTokenMLP, MMBertEmbeddings
+
+
+# --------------- fine-tuning models ---------------
+class MMBertForJoint(BertPreTrainedModel):
+ """A BertModel with isolated attention mask to separate modality."""
+
+ def __init__(self, config):
+ super().__init__(config)
+ self.videomlp = VideoTokenMLP(config)
+ self.bert = MMBertModel(config)
+ self.init_weights()
+
+ def forward(
+ self,
+ input_ids=None,
+ input_video_embeds=None,
+ attention_mask=None,
+ token_type_ids=None,
+ position_ids=None,
+ head_mask=None,
+ inputs_embeds=None,
+ next_sentence_label=None,
+ output_attentions=None,
+ output_hidden_states=None,
+ return_dict=None,
+ separate_forward_split=None,
+ ):
+ return_dict = (
+ return_dict if return_dict is not None
+ else self.config.use_return_dict
+ )
+ video_tokens = self.videomlp(input_video_embeds)
+
+ outputs = self.bert(
+ input_ids,
+ video_tokens,
+ attention_mask=attention_mask,
+ token_type_ids=token_type_ids,
+ position_ids=position_ids,
+ head_mask=head_mask,
+ inputs_embeds=inputs_embeds,
+ output_attentions=output_attentions,
+ output_hidden_states=output_hidden_states,
+ return_dict=return_dict,
+ separate_forward_split=separate_forward_split,
+ )
+
+ return outputs
+
+
+class MMBertForTokenClassification(BertPreTrainedModel):
+ """A BertModel similar to MMJointUni, with extra wrapper layer
+ to be fine-tuned from other pretrained MMFusion model."""
+
+ def __init__(self, config):
+ super().__init__(config)
+ self.videomlp = VideoTokenMLP(config)
+ self.bert = MMBertModel(config)
+ self.dropout = nn.Dropout(config.hidden_dropout_prob)
+ # TODO(huxu): 779 is the number of classes for COIN: move to config?
+ self.classifier = nn.Linear(config.hidden_size, 779)
+ self.init_weights()
+
+ def forward(
+ self,
+ input_ids=None,
+ input_video_embeds=None,
+ attention_mask=None,
+ token_type_ids=None,
+ position_ids=None,
+ head_mask=None,
+ inputs_embeds=None,
+ next_sentence_label=None,
+ output_attentions=None,
+ output_hidden_states=None,
+ return_dict=None,
+ separate_forward_split=None,
+ ):
+ return_dict = (
+ return_dict if return_dict is not None
+ else self.config.use_return_dict
+ )
+
+ video_tokens = self.videomlp(input_video_embeds)
+ outputs = self.bert(
+ input_ids,
+ video_tokens,
+ attention_mask=attention_mask,
+ token_type_ids=token_type_ids,
+ position_ids=position_ids,
+ head_mask=head_mask,
+ inputs_embeds=inputs_embeds,
+ output_attentions=output_attentions,
+ output_hidden_states=output_hidden_states,
+ return_dict=return_dict,
+ separate_forward_split=separate_forward_split,
+ )
+
+ return (self.classifier(outputs[0]),)
+
+
+# ------------ pre-training models ----------------
+
+class MMBertForEncoder(BertPreTrainedModel):
+ """A BertModel for Contrastive Learning."""
+ def __init__(self, config):
+ super().__init__(config)
+ self.videomlp = VideoTokenMLP(config)
+ self.bert = MMBertModel(config)
+ self.init_weights()
+
+ def forward(
+ self,
+ input_ids=None,
+ input_video_embeds=None,
+ attention_mask=None,
+ token_type_ids=None,
+ position_ids=None,
+ head_mask=None,
+ inputs_embeds=None,
+ output_attentions=None,
+ output_hidden_states=None,
+ return_dict=None,
+ ):
+ return_dict = (
+ return_dict if return_dict is not None
+ else self.config.use_return_dict
+ )
+ if input_video_embeds is not None:
+ video_tokens = self.videomlp(input_video_embeds)
+ else:
+ video_tokens = None
+
+ outputs = self.bert(
+ input_ids,
+ video_tokens,
+ attention_mask=attention_mask,
+ token_type_ids=token_type_ids,
+ position_ids=position_ids,
+ head_mask=head_mask,
+ inputs_embeds=inputs_embeds,
+ output_attentions=output_attentions,
+ output_hidden_states=output_hidden_states,
+ return_dict=return_dict,
+ )
+ return outputs
+
+
+class MMBertForMFMMLM(BertPreTrainedModel):
+ """A BertModel with shared prediction head on MFM-MLM."""
+ def __init__(self, config):
+ super().__init__(config)
+ self.videomlp = VideoTokenMLP(config)
+ self.bert = MMBertModel(config)
+ self.cls = MFMMLMHead(config)
+ self.hidden_size = config.hidden_size
+ self.init_weights()
+
+ def get_output_embeddings(self):
+ return self.cls.predictions.decoder
+
+ def forward(
+ self,
+ input_ids=None,
+ input_video_embeds=None,
+ attention_mask=None,
+ token_type_ids=None,
+ position_ids=None,
+ head_mask=None,
+ inputs_embeds=None,
+ masked_frame_labels=None,
+ target_video_hidden_states=None,
+ non_masked_frame_mask=None,
+ masked_lm_labels=None,
+ output_attentions=None,
+ output_hidden_states=None,
+ return_dict=None,
+ ):
+ return_dict = (
+ return_dict if return_dict is not None
+ else self.config.use_return_dict
+ )
+ if input_video_embeds is not None:
+ video_tokens = self.videomlp(input_video_embeds)
+ else:
+ video_tokens = None
+
+ if target_video_hidden_states is not None:
+ target_video_hidden_states = self.videomlp(
+ target_video_hidden_states)
+
+ non_masked_frame_hidden_states = video_tokens.masked_select(
+ non_masked_frame_mask.unsqueeze(-1)
+ ).view(-1, self.hidden_size)
+
+ outputs = self.bert(
+ input_ids,
+ video_tokens,
+ attention_mask=attention_mask,
+ token_type_ids=token_type_ids,
+ position_ids=position_ids,
+ head_mask=head_mask,
+ inputs_embeds=inputs_embeds,
+ output_attentions=output_attentions,
+ output_hidden_states=output_hidden_states,
+ return_dict=return_dict,
+ )
+
+ sequence_output = outputs[0]
+
+ mfm_scores, prediction_scores = None, None
+ if masked_frame_labels is not None and masked_lm_labels is not None:
+ # split the sequence.
+ text_offset = masked_frame_labels.size(1) + 1 # [CLS]
+ video_sequence_output = sequence_output[
+ :, 1:text_offset
+ ] # remove [SEP] as not in video_label.
+ text_sequence_output = torch.cat(
+ [sequence_output[:, :1], sequence_output[:, text_offset:]],
+ dim=1
+ )
+
+ hidden_size = video_sequence_output.size(-1)
+ selected_video_output = video_sequence_output.masked_select(
+ masked_frame_labels.unsqueeze(-1)
+ ).view(-1, hidden_size)
+
+ # only compute select tokens to training to speed up.
+ hidden_size = text_sequence_output.size(-1)
+ # masked_lm_labels = masked_lm_labels.reshape(-1)
+ labels_mask = masked_lm_labels != -100
+
+ selected_text_output = text_sequence_output.masked_select(
+ labels_mask.unsqueeze(-1)
+ ).view(-1, hidden_size)
+ mfm_scores, prediction_scores = self.cls(
+ selected_video_output,
+ target_video_hidden_states,
+ non_masked_frame_hidden_states,
+ selected_text_output,
+ )
+
+ output = (
+ mfm_scores,
+ prediction_scores,
+ ) + outputs
+ return output
+
+
+class BertMFMMLMPredictionHead(nn.Module):
+ def __init__(self, config):
+ super().__init__()
+ self.transform = BertPredictionHeadTransform(config)
+ # The output weights are the same as the input embeddings, but there is
+ # an output-only bias for each token.
+ self.decoder = nn.Linear(
+ config.hidden_size, config.vocab_size, bias=False)
+
+ self.bias = nn.Parameter(torch.zeros(config.vocab_size))
+
+ # Need a link between the two variables so that the bias is correctly
+ # resized with `resize_token_embeddings`
+ self.decoder.bias = self.bias
+
+ def forward(
+ self,
+ video_hidden_states=None,
+ target_video_hidden_states=None,
+ non_masked_frame_hidden_states=None,
+ text_hidden_states=None,
+ ):
+ video_logits, text_logits = None, None
+ if video_hidden_states is not None:
+ video_hidden_states = self.transform(video_hidden_states)
+ non_masked_frame_logits = torch.mm(
+ video_hidden_states,
+ non_masked_frame_hidden_states.transpose(1, 0)
+ )
+ masked_frame_logits = torch.bmm(
+ video_hidden_states.unsqueeze(1),
+ target_video_hidden_states.unsqueeze(-1),
+ ).squeeze(-1)
+ video_logits = torch.cat(
+ [masked_frame_logits, non_masked_frame_logits], dim=1
+ )
+
+ if text_hidden_states is not None:
+ text_hidden_states = self.transform(text_hidden_states)
+ text_logits = self.decoder(text_hidden_states)
+ return video_logits, text_logits
+
+
+class MFMMLMHead(nn.Module):
+ def __init__(self, config):
+ super().__init__()
+ self.predictions = BertMFMMLMPredictionHead(config)
+
+ def forward(
+ self,
+ video_hidden_states=None,
+ target_video_hidden_states=None,
+ non_masked_frame_hidden_states=None,
+ text_hidden_states=None,
+ ):
+ video_logits, text_logits = self.predictions(
+ video_hidden_states,
+ target_video_hidden_states,
+ non_masked_frame_hidden_states,
+ text_hidden_states,
+ )
+ return video_logits, text_logits
+
+
+class MMBertForMTM(MMBertForMFMMLM):
+ def __init__(self, config):
+ BertPreTrainedModel.__init__(self, config)
+ self.videomlp = VideoTokenMLP(config)
+ self.bert = MMBertModel(config)
+ self.cls = MTMHead(config)
+ self.hidden_size = config.hidden_size
+ self.init_weights()
+
+
+class BertMTMPredictionHead(nn.Module):
+ def __init__(self, config):
+ super().__init__()
+ self.transform = BertPredictionHeadTransform(config)
+ self.decoder = nn.Linear(
+ config.hidden_size, config.vocab_size, bias=False)
+
+ def forward(
+ self,
+ video_hidden_states=None,
+ target_video_hidden_states=None,
+ non_masked_frame_hidden_states=None,
+ text_hidden_states=None,
+ ):
+ non_masked_frame_hidden_states = non_masked_frame_hidden_states.transpose(1, 0)
+ video_logits, text_logits = None, None
+ if video_hidden_states is not None:
+ video_hidden_states = self.transform(video_hidden_states)
+
+ masked_frame_logits = torch.bmm(
+ video_hidden_states.unsqueeze(1),
+ target_video_hidden_states.unsqueeze(-1),
+ ).squeeze(-1)
+
+ non_masked_frame_logits = torch.mm(
+ video_hidden_states,
+ non_masked_frame_hidden_states
+ )
+ video_on_vocab_logits = self.decoder(video_hidden_states)
+ video_logits = torch.cat([
+ masked_frame_logits,
+ non_masked_frame_logits,
+ video_on_vocab_logits], dim=1)
+
+ if text_hidden_states is not None:
+ text_hidden_states = self.transform(text_hidden_states)
+ # text first so label does not need to be shifted.
+ text_on_vocab_logits = self.decoder(text_hidden_states)
+ text_on_video_logits = torch.mm(
+ text_hidden_states,
+ non_masked_frame_hidden_states
+ )
+ text_logits = torch.cat([
+ text_on_vocab_logits,
+ text_on_video_logits
+ ], dim=1)
+
+ return video_logits, text_logits
+
+
+class MTMHead(nn.Module):
+ def __init__(self, config):
+ super().__init__()
+ self.predictions = BertMTMPredictionHead(config)
+
+ def forward(
+ self,
+ video_hidden_states=None,
+ target_video_hidden_states=None,
+ non_masked_frame_hidden_states=None,
+ text_hidden_states=None,
+ ):
+ video_logits, text_logits = self.predictions(
+ video_hidden_states,
+ target_video_hidden_states,
+ non_masked_frame_hidden_states,
+ text_hidden_states,
+ )
+ return video_logits, text_logits
+
+
+class MMBertModel(BertModel):
+ """MMBertModel has MMBertEmbedding to support video tokens."""
+
+ def __init__(self, config, add_pooling_layer=True):
+ super().__init__(config)
+ # overwrite embedding
+ self.embeddings = MMBertEmbeddings(config)
+ self.encoder = MultiLayerAttentionMaskBertEncoder(config)
+ self.init_weights()
+
+ def forward(
+ self,
+ input_ids=None,
+ input_video_embeds=None,
+ attention_mask=None,
+ token_type_ids=None,
+ position_ids=None,
+ head_mask=None,
+ inputs_embeds=None,
+ encoder_hidden_states=None,
+ encoder_attention_mask=None,
+ output_attentions=None,
+ output_hidden_states=None,
+ return_dict=None,
+ separate_forward_split=None,
+ ):
+ output_attentions = (
+ output_attentions
+ if output_attentions is not None
+ else self.config.output_attentions
+ )
+ output_hidden_states = (
+ output_hidden_states
+ if output_hidden_states is not None
+ else self.config.output_hidden_states
+ )
+ return_dict = (
+ return_dict if return_dict is not None
+ else self.config.use_return_dict
+ )
+
+ if input_ids is not None and inputs_embeds is not None:
+ raise ValueError(
+ "You cannot specify both input_ids "
+ "and inputs_embeds at the same time"
+ )
+ elif input_ids is not None:
+ if input_video_embeds is not None:
+ input_shape = (
+ input_ids.size(0),
+ input_ids.size(1) + input_video_embeds.size(1),
+ )
+ else:
+ input_shape = (
+ input_ids.size(0),
+ input_ids.size(1),
+ )
+ elif inputs_embeds is not None:
+ if input_video_embeds is not None:
+ input_shape = (
+ inputs_embeds.size(0),
+ inputs_embeds.size(1) + input_video_embeds.size(1),
+ )
+ else:
+ input_shape = (
+ input_ids.size(0),
+ input_ids.size(1),
+ )
+ else:
+ raise ValueError(
+ "You have to specify either input_ids or inputs_embeds")
+
+ device = input_ids.device if input_ids is not None \
+ else inputs_embeds.device
+
+ if attention_mask is None:
+ attention_mask = torch.ones(input_shape, device=device)
+ if token_type_ids is None:
+ token_type_ids = torch.zeros(
+ input_shape, dtype=torch.long, device=device)
+
+ # We can provide a self-attention mask of dimensions
+ # [batch_size, from_seq_length, to_seq_length]
+ # ourselves in which case
+ # we just need to make it broadcastable to all heads.
+ extended_attention_mask: torch.Tensor = \
+ self.get_extended_attention_mask(
+ attention_mask, input_shape, device)
+
+ # If a 2D or 3D attention mask is provided for the cross-attention
+ # we need to make broadcastable to
+ # [batch_size, num_heads, seq_length, seq_length]
+ if self.config.is_decoder and encoder_hidden_states is not None:
+ (
+ encoder_batch_size,
+ encoder_sequence_length,
+ _,
+ ) = encoder_hidden_states.size()
+ encoder_hidden_shape = (
+ encoder_batch_size, encoder_sequence_length)
+ if encoder_attention_mask is None:
+ encoder_attention_mask = torch.ones(
+ encoder_hidden_shape, device=device)
+ encoder_extended_attention_mask = self.invert_attention_mask(
+ encoder_attention_mask
+ )
+ else:
+ encoder_extended_attention_mask = None
+
+ # Prepare head mask if needed
+ # 1.0 in head_mask indicate we keep the head
+ # attention_probs has shape bsz x n_heads x N x N
+ # input head_mask has shape [num_heads] or
+ # [num_hidden_layers x num_heads]
+ # and head_mask is converted to shape
+ # [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+
+ head_mask = self.get_head_mask(
+ head_mask, self.config.num_hidden_layers)
+
+ embedding_output = self.embeddings(
+ input_ids,
+ input_video_embeds,
+ position_ids=position_ids,
+ token_type_ids=token_type_ids,
+ inputs_embeds=inputs_embeds,
+ )
+
+ if separate_forward_split is not None:
+ split_embedding_output = \
+ embedding_output[:, :separate_forward_split]
+ split_extended_attention_mask = extended_attention_mask[
+ :, :, :, :separate_forward_split, :separate_forward_split
+ ]
+ split_encoder_outputs = self.encoder(
+ split_embedding_output,
+ attention_mask=split_extended_attention_mask,
+ head_mask=head_mask,
+ encoder_hidden_states=encoder_hidden_states,
+ encoder_attention_mask=encoder_extended_attention_mask,
+ output_attentions=output_attentions,
+ output_hidden_states=output_hidden_states,
+ return_dict=return_dict,
+ )
+ assert (
+ len(split_encoder_outputs) <= 2
+ ), "we do not support merge on attention for now."
+ encoder_outputs = []
+ encoder_outputs.append([split_encoder_outputs[0]])
+ if len(split_encoder_outputs) == 2:
+ encoder_outputs.append([])
+ for _all_hidden_states in split_encoder_outputs[1]:
+ encoder_outputs[-1].append([_all_hidden_states])
+
+ split_embedding_output = \
+ embedding_output[:, separate_forward_split:]
+ split_extended_attention_mask = extended_attention_mask[
+ :, :, :, separate_forward_split:, separate_forward_split:
+ ]
+
+ split_encoder_outputs = self.encoder(
+ split_embedding_output,
+ attention_mask=split_extended_attention_mask,
+ head_mask=head_mask,
+ encoder_hidden_states=encoder_hidden_states,
+ encoder_attention_mask=encoder_extended_attention_mask,
+ output_attentions=output_attentions,
+ output_hidden_states=output_hidden_states,
+ return_dict=return_dict,
+ )
+
+ assert (
+ len(split_encoder_outputs) <= 2
+ ), "we do not support merge on attention for now."
+ encoder_outputs[0].append(split_encoder_outputs[0])
+ encoder_outputs[0] = torch.cat(encoder_outputs[0], dim=1)
+ if len(split_encoder_outputs) == 2:
+ for layer_idx, _all_hidden_states in enumerate(
+ split_encoder_outputs[1]
+ ):
+ encoder_outputs[1][layer_idx].append(_all_hidden_states)
+ encoder_outputs[1][layer_idx] = torch.cat(
+ encoder_outputs[1][layer_idx], dim=1
+ )
+ encoder_outputs = tuple(encoder_outputs)
+ else:
+ encoder_outputs = self.encoder(
+ embedding_output,
+ attention_mask=extended_attention_mask,
+ head_mask=head_mask,
+ encoder_hidden_states=encoder_hidden_states,
+ encoder_attention_mask=encoder_extended_attention_mask,
+ output_attentions=output_attentions,
+ output_hidden_states=output_hidden_states,
+ return_dict=return_dict,
+ )
+
+ sequence_output = encoder_outputs[0]
+ pooled_output = (
+ self.pooler(sequence_output) if self.pooler is not None else None
+ )
+
+ return (sequence_output, pooled_output) + encoder_outputs[1:]
+
+ def get_extended_attention_mask(self, attention_mask, input_shape, device):
+ """This is borrowed from `modeling_utils.py` with the support of
+ multi-layer attention masks.
+ The second dim is expected to be number of layers.
+ See `MMAttentionMaskProcessor`.
+ Makes broadcastable attention and causal masks so that future
+ and masked tokens are ignored.
+
+ Arguments:
+ attention_mask (:obj:`torch.Tensor`):
+ Mask with ones indicating tokens to attend to,
+ zeros for tokens to ignore.
+ input_shape (:obj:`Tuple[int]`):
+ The shape of the input to the model.
+ device: (:obj:`torch.device`):
+ The device of the input to the model.
+
+ Returns:
+ :obj:`torch.Tensor` The extended attention mask, \
+ with a the same dtype as :obj:`attention_mask.dtype`.
+ """
+ # We can provide a self-attention mask of dimensions
+ # [batch_size, from_seq_length, to_seq_length]
+ # ourselves in which case we just need to make it broadcastable
+ # to all heads.
+ if attention_mask.dim() == 4:
+ extended_attention_mask = attention_mask[:, :, None, :, :]
+ extended_attention_mask = extended_attention_mask.to(
+ dtype=self.dtype
+ ) # fp16 compatibility
+ extended_attention_mask = (1.0 - extended_attention_mask) \
+ * -10000.0
+ return extended_attention_mask
+ else:
+ return super().get_extended_attention_mask(
+ attention_mask, input_shape, device
+ )
+
+
+class MultiLayerAttentionMaskBertEncoder(BertEncoder):
+ """extend BertEncoder with the capability of
+ multiple layers of attention mask."""
+
+ def forward(
+ self,
+ hidden_states,
+ attention_mask=None,
+ head_mask=None,
+ encoder_hidden_states=None,
+ encoder_attention_mask=None,
+ output_attentions=False,
+ output_hidden_states=False,
+ return_dict=False,
+ ):
+ all_hidden_states = () if output_hidden_states else None
+ all_attentions = () if output_attentions else None
+ for i, layer_module in enumerate(self.layer):
+ if output_hidden_states:
+ all_hidden_states = all_hidden_states + (hidden_states,)
+ layer_head_mask = head_mask[i] if head_mask is not None else None
+
+ layer_attention_mask = (
+ attention_mask[:, i, :, :, :]
+ if attention_mask.dim() == 5
+ else attention_mask
+ )
+
+ if getattr(self.config, "gradient_checkpointing", False):
+
+ def create_custom_forward(module):
+ def custom_forward(*inputs):
+ return module(*inputs, output_attentions)
+
+ return custom_forward
+
+ layer_outputs = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(layer_module),
+ hidden_states,
+ layer_attention_mask,
+ layer_head_mask,
+ encoder_hidden_states,
+ encoder_attention_mask,
+ )
+ else:
+ layer_outputs = layer_module(
+ hidden_states,
+ layer_attention_mask,
+ layer_head_mask,
+ encoder_hidden_states,
+ encoder_attention_mask,
+ output_attentions,
+ )
+ hidden_states = layer_outputs[0]
+ if output_attentions:
+ all_attentions = all_attentions + (layer_outputs[1],)
+
+ if output_hidden_states:
+ all_hidden_states = all_hidden_states + (hidden_states,)
+
+ return tuple(
+ v
+ for v in [hidden_states, all_hidden_states, all_attentions]
+ if v is not None
+ )
diff --git a/fairseq/examples/MMPT/mmpt/modules/__init__.py b/fairseq/examples/MMPT/mmpt/modules/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..4c78594c21d7ccdcb2ab11c918dd211adce27c14
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/modules/__init__.py
@@ -0,0 +1,10 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+from .mm import *
+
+try:
+ from .expmm import *
+except ImportError:
+ pass
diff --git a/fairseq/examples/MMPT/mmpt/modules/mm.py b/fairseq/examples/MMPT/mmpt/modules/mm.py
new file mode 100644
index 0000000000000000000000000000000000000000..5d9777371a5e7afc235a8b4e6c164b2a35258eb6
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/modules/mm.py
@@ -0,0 +1,145 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# Copyright (c) Facebook, Inc. All Rights Reserved
+
+
+import torch
+
+from torch import nn
+
+try:
+ from transformers.modeling_bert import (
+ BertEmbeddings,
+ ACT2FN,
+ )
+except ImportError:
+ pass
+
+
+class VideoTokenMLP(nn.Module):
+ def __init__(self, config):
+ super().__init__()
+ input_dim = config.input_dim if hasattr(config, "input_dim") else 512
+ self.linear1 = nn.Linear(input_dim, config.hidden_size)
+ self.LayerNorm = nn.LayerNorm(config.hidden_size)
+ self.activation = ACT2FN[config.hidden_act]
+ self.linear2 = nn.Linear(config.hidden_size, config.hidden_size)
+
+ def forward(self, hidden_states):
+ hidden_states = self.linear1(hidden_states)
+ hidden_states = self.activation(hidden_states)
+ hidden_states = self.LayerNorm(hidden_states)
+ hidden_states = self.linear2(hidden_states)
+ return hidden_states
+
+
+class MMBertEmbeddings(BertEmbeddings):
+ def __init__(self, config):
+ super().__init__(config)
+ self.max_video_len = config.max_video_len
+ if hasattr(config, "use_seg_emb") and config.use_seg_emb:
+ """the original VLM paper uses seg_embeddings for temporal space.
+ although not used it changed the randomness of initialization.
+ we keep it for reproducibility.
+ """
+ self.seg_embeddings = nn.Embedding(256, config.hidden_size)
+
+ def forward(
+ self,
+ input_ids,
+ input_video_embeds,
+ token_type_ids=None,
+ position_ids=None,
+ inputs_embeds=None,
+ ):
+ input_tensor = input_ids if input_ids is not None else inputs_embeds
+ if input_video_embeds is not None:
+ input_shape = (
+ input_tensor.size(0),
+ input_tensor.size(1) + input_video_embeds.size(1),
+ )
+ else:
+ input_shape = (input_tensor.size(0), input_tensor.size(1))
+
+ if position_ids is None:
+ """
+ Auto skip position embeddings for text only case.
+ use cases:
+ (1) action localization and segmentation:
+ feed in len-1 dummy video token needs text part to
+ skip input_video_embeds.size(1) for the right
+ position_ids for video [SEP] and rest text tokens.
+ (2) MMFusionShare for two forward passings:
+ in `forward_text`: input_video_embeds is None.
+ need to skip video [SEP] token.
+
+ # video_len + 1: [CLS] + video_embed
+ # self.max_video_len + 1: [SEP] for video.
+ # self.max_video_len + 2: [SEP] for video.
+ # self.max_video_len + input_ids.size(1): rest for text.
+ """
+ if input_video_embeds is not None:
+ video_len = input_video_embeds.size(1)
+ starting_offset = self.max_video_len + 1 # video [SEP]
+ ending_offset = self.max_video_len + input_ids.size(1)
+ else:
+ video_len = 0
+ starting_offset = self.max_video_len + 2 # first text token.
+ ending_offset = self.max_video_len + input_ids.size(1) + 1
+ position_ids = torch.cat([
+ self.position_ids[:, :video_len + 1],
+ self.position_ids[:, starting_offset:ending_offset]
+ ], dim=1)
+
+ if token_type_ids is None:
+ token_type_ids = torch.zeros(
+ input_shape, dtype=torch.long, device=self.position_ids.device
+ )
+
+ """
+ the format of input_ids is [CLS] [SEP] caption [SEP] padding.
+ the goal is to build [CLS] video tokens [SEP] caption [SEP] .
+ """
+ if inputs_embeds is None:
+ inputs_embeds = self.word_embeddings(input_ids)
+ if input_video_embeds is not None:
+ inputs_mm_embeds = torch.cat([
+ inputs_embeds[:, :1], input_video_embeds, inputs_embeds[:, 1:]
+ ], dim=1)
+ else:
+ # text only for `MMFusionShare`.
+ inputs_mm_embeds = inputs_embeds
+
+ position_embeddings = self.position_embeddings(position_ids)
+ token_type_embeddings = self.token_type_embeddings(token_type_ids)
+ embeddings = inputs_mm_embeds + position_embeddings
+ embeddings += token_type_embeddings
+
+ embeddings = self.LayerNorm(embeddings)
+ embeddings = self.dropout(embeddings)
+ return embeddings
+
+
+class AlignHead(nn.Module):
+ """this will load pre-trained weights for NSP, which is desirable."""
+
+ def __init__(self, config):
+ super().__init__()
+ self.seq_relationship = nn.Linear(config.hidden_size, 2)
+
+ def forward(self, dropout_pooled_output):
+ logits = self.seq_relationship(dropout_pooled_output)
+ return logits
diff --git a/fairseq/examples/MMPT/mmpt/modules/retri.py b/fairseq/examples/MMPT/mmpt/modules/retri.py
new file mode 100644
index 0000000000000000000000000000000000000000..d1b288f8e52308eeaefcb4411cb3d23e543f37d0
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/modules/retri.py
@@ -0,0 +1,429 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+import os
+import numpy as np
+import pickle
+import time
+
+try:
+ import faiss
+except ImportError:
+ pass
+
+from collections import defaultdict
+
+from ..utils import get_local_rank, print_on_rank0
+
+
+class VectorRetriever(object):
+ """
+ How2 Video Retriver.
+ Reference usage of FAISS:
+ https://github.com/fairinternal/fairseq-py/blob/paraphrase_pretraining/fairseq/data/multilingual_faiss_dataset.py
+ """
+
+ def __init__(self, hidden_size, cent, db_type, examples_per_cent_to_train):
+ if db_type == "flatl2":
+ quantizer = faiss.IndexFlatL2(hidden_size) # the other index
+ self.db = faiss.IndexIVFFlat(
+ quantizer, hidden_size, cent, faiss.METRIC_L2)
+ elif db_type == "pq":
+ self.db = faiss.index_factory(
+ hidden_size, f"IVF{cent}_HNSW32,PQ32"
+ )
+ else:
+ raise ValueError("unknown type of db", db_type)
+ self.train_thres = cent * examples_per_cent_to_train
+ self.train_cache = []
+ self.train_len = 0
+ self.videoid_to_vectoridx = {}
+ self.vectoridx_to_videoid = None
+ self.make_direct_maps_done = False
+
+ def make_direct_maps(self):
+ faiss.downcast_index(self.db).make_direct_map()
+
+ def __len__(self):
+ return self.db.ntotal
+
+ def save(self, out_dir):
+ faiss.write_index(
+ self.db,
+ os.path.join(out_dir, "faiss_idx")
+ )
+ with open(
+ os.path.join(
+ out_dir, "videoid_to_vectoridx.pkl"),
+ "wb") as fw:
+ pickle.dump(
+ self.videoid_to_vectoridx, fw,
+ protocol=pickle.HIGHEST_PROTOCOL
+ )
+
+ def load(self, out_dir):
+ fn = os.path.join(out_dir, "faiss_idx")
+ self.db = faiss.read_index(fn)
+ with open(
+ os.path.join(out_dir, "videoid_to_vectoridx.pkl"), "rb") as fr:
+ self.videoid_to_vectoridx = pickle.load(fr)
+
+ def add(self, hidden_states, video_ids, last=False):
+ assert len(hidden_states) == len(video_ids), "{}, {}".format(
+ str(len(hidden_states)), str(len(video_ids)))
+ assert len(hidden_states.shape) == 2
+ assert hidden_states.dtype == np.float32
+
+ valid_idx = []
+ for idx, video_id in enumerate(video_ids):
+ if video_id not in self.videoid_to_vectoridx:
+ valid_idx.append(idx)
+ self.videoid_to_vectoridx[video_id] = \
+ len(self.videoid_to_vectoridx)
+
+ hidden_states = hidden_states[valid_idx]
+ if not self.db.is_trained:
+ self.train_cache.append(hidden_states)
+ self.train_len += hidden_states.shape[0]
+ if self.train_len < self.train_thres:
+ return
+ self.finalize_training()
+ else:
+ self.db.add(hidden_states)
+
+ def finalize_training(self):
+ hidden_states = np.concatenate(self.train_cache, axis=0)
+ del self.train_cache
+ local_rank = get_local_rank()
+ if local_rank == 0:
+ start = time.time()
+ print("training db on", self.train_thres, "/", self.train_len)
+ self.db.train(hidden_states[:self.train_thres])
+ if local_rank == 0:
+ print("training db for", time.time() - start)
+ self.db.add(hidden_states)
+
+ def search(
+ self,
+ query_hidden_states,
+ orig_dist,
+ ):
+ if len(self.videoid_to_vectoridx) != self.db.ntotal:
+ raise ValueError(
+ "cannot search: size mismatch in-between index and db",
+ len(self.videoid_to_vectoridx),
+ self.db.ntotal
+ )
+
+ if self.vectoridx_to_videoid is None:
+ self.vectoridx_to_videoid = {
+ self.videoid_to_vectoridx[videoid]: videoid
+ for videoid in self.videoid_to_vectoridx
+ }
+ assert len(self.vectoridx_to_videoid) \
+ == len(self.videoid_to_vectoridx)
+
+ # MultilingualFaissDataset uses the following; not sure the purpose.
+ # faiss.ParameterSpace().set_index_parameter(self.db, "nprobe", 10)
+ queried_dist, index = self.db.search(query_hidden_states, 1)
+ queried_dist, index = queried_dist[:, 0], index[:, 0]
+
+ outputs = np.array(
+ [self.vectoridx_to_videoid[_index]
+ if _index != -1 else (-1, -1, -1) for _index in index],
+ dtype=np.int32)
+ outputs[queried_dist <= orig_dist] = -1
+ return outputs
+
+ def search_by_video_ids(
+ self,
+ video_ids,
+ retri_factor
+ ):
+ if len(self.videoid_to_vectoridx) != self.db.ntotal:
+ raise ValueError(
+ len(self.videoid_to_vectoridx),
+ self.db.ntotal
+ )
+
+ if not self.make_direct_maps_done:
+ self.make_direct_maps()
+
+ if self.vectoridx_to_videoid is None:
+ self.vectoridx_to_videoid = {
+ self.videoid_to_vectoridx[videoid]: videoid
+ for videoid in self.videoid_to_vectoridx
+ }
+ assert len(self.vectoridx_to_videoid) \
+ == len(self.videoid_to_vectoridx)
+
+ query_hidden_states = []
+ vector_ids = []
+ for video_id in video_ids:
+ vector_id = self.videoid_to_vectoridx[video_id]
+ vector_ids.append(vector_id)
+ query_hidden_state = self.db.reconstruct(vector_id)
+ query_hidden_states.append(query_hidden_state)
+ query_hidden_states = np.stack(query_hidden_states)
+
+ # MultilingualFaissDataset uses the following; not sure the reason.
+ # faiss.ParameterSpace().set_index_parameter(self.db, "nprobe", 10)
+ _, index = self.db.search(query_hidden_states, retri_factor)
+ outputs = []
+ for sample_idx, sample in enumerate(index):
+ # the first video_id is always the video itself.
+ cands = [video_ids[sample_idx]]
+ for vector_idx in sample:
+ if vector_idx >= 0 \
+ and vector_ids[sample_idx] != vector_idx:
+ cands.append(
+ self.vectoridx_to_videoid[vector_idx]
+ )
+ outputs.append(cands)
+ return outputs
+
+
+class VectorRetrieverDM(VectorRetriever):
+ """
+ with direct map.
+ How2 Video Retriver.
+ Reference usage of FAISS:
+ https://github.com/fairinternal/fairseq-py/blob/paraphrase_pretraining/fairseq/data/multilingual_faiss_dataset.py
+ """
+
+ def __init__(
+ self,
+ hidden_size,
+ cent,
+ db_type,
+ examples_per_cent_to_train
+ ):
+ super().__init__(
+ hidden_size, cent, db_type, examples_per_cent_to_train)
+ self.make_direct_maps_done = False
+
+ def make_direct_maps(self):
+ faiss.downcast_index(self.db).make_direct_map()
+ self.make_direct_maps_done = True
+
+ def search(
+ self,
+ query_hidden_states,
+ orig_dist,
+ ):
+ if len(self.videoid_to_vectoridx) != self.db.ntotal:
+ raise ValueError(
+ len(self.videoid_to_vectoridx),
+ self.db.ntotal
+ )
+
+ if not self.make_direct_maps_done:
+ self.make_direct_maps()
+ if self.vectoridx_to_videoid is None:
+ self.vectoridx_to_videoid = {
+ self.videoid_to_vectoridx[videoid]: videoid
+ for videoid in self.videoid_to_vectoridx
+ }
+ assert len(self.vectoridx_to_videoid) \
+ == len(self.videoid_to_vectoridx)
+
+ # MultilingualFaissDataset uses the following; not sure the reason.
+ # faiss.ParameterSpace().set_index_parameter(self.db, "nprobe", 10)
+ queried_dist, index = self.db.search(query_hidden_states, 1)
+ outputs = []
+ for sample_idx, sample in enumerate(index):
+ # and queried_dist[sample_idx] < thres \
+ if sample >= 0 \
+ and queried_dist[sample_idx] < orig_dist[sample_idx]:
+ outputs.append(self.vectoridx_to_videoid[sample])
+ else:
+ outputs.append(None)
+ return outputs
+
+ def search_by_video_ids(
+ self,
+ video_ids,
+ retri_factor=8
+ ):
+ if len(self.videoid_to_vectoridx) != self.db.ntotal:
+ raise ValueError(
+ len(self.videoid_to_vectoridx),
+ self.db.ntotal
+ )
+
+ if not self.make_direct_maps_done:
+ self.make_direct_maps()
+ if self.vectoridx_to_videoid is None:
+ self.vectoridx_to_videoid = {
+ self.videoid_to_vectoridx[videoid]: videoid
+ for videoid in self.videoid_to_vectoridx
+ }
+ assert len(self.vectoridx_to_videoid) \
+ == len(self.videoid_to_vectoridx)
+
+ query_hidden_states = []
+ vector_ids = []
+ for video_id in video_ids:
+ vector_id = self.videoid_to_vectoridx[video_id]
+ vector_ids.append(vector_id)
+ query_hidden_state = self.db.reconstruct(vector_id)
+ query_hidden_states.append(query_hidden_state)
+ query_hidden_states = np.stack(query_hidden_states)
+
+ # MultilingualFaissDataset uses the following; not sure the reason.
+ # faiss.ParameterSpace().set_index_parameter(self.db, "nprobe", 10)
+ _, index = self.db.search(query_hidden_states, retri_factor)
+ outputs = []
+ for sample_idx, sample in enumerate(index):
+ # the first video_id is always the video itself.
+ cands = [video_ids[sample_idx]]
+ for vector_idx in sample:
+ if vector_idx >= 0 \
+ and vector_ids[sample_idx] != vector_idx:
+ cands.append(
+ self.vectoridx_to_videoid[vector_idx]
+ )
+ outputs.append(cands)
+ return outputs
+
+
+class MMVectorRetriever(VectorRetrieverDM):
+ """
+ multimodal vector retriver:
+ text retrieve video or video retrieve text.
+ """
+
+ def __init__(self, hidden_size, cent, db_type, examples_per_cent_to_train):
+ super().__init__(
+ hidden_size, cent, db_type, examples_per_cent_to_train)
+ video_db = self.db
+ super().__init__(
+ hidden_size, cent, db_type, examples_per_cent_to_train)
+ text_db = self.db
+ self.db = {"video": video_db, "text": text_db}
+ self.video_to_videoid = defaultdict(list)
+
+ def __len__(self):
+ assert self.db["video"].ntotal == self.db["text"].ntotal
+ return self.db["video"].ntotal
+
+ def make_direct_maps(self):
+ faiss.downcast_index(self.db["video"]).make_direct_map()
+ faiss.downcast_index(self.db["text"]).make_direct_map()
+
+ def save(self, out_dir):
+ faiss.write_index(
+ self.db["video"],
+ os.path.join(out_dir, "video_faiss_idx")
+ )
+ faiss.write_index(
+ self.db["text"],
+ os.path.join(out_dir, "text_faiss_idx")
+ )
+
+ with open(
+ os.path.join(
+ out_dir, "videoid_to_vectoridx.pkl"),
+ "wb") as fw:
+ pickle.dump(
+ self.videoid_to_vectoridx, fw,
+ protocol=pickle.HIGHEST_PROTOCOL
+ )
+
+ def load(self, out_dir):
+ fn = os.path.join(out_dir, "video_faiss_idx")
+ video_db = faiss.read_index(fn)
+ fn = os.path.join(out_dir, "text_faiss_idx")
+ text_db = faiss.read_index(fn)
+ self.db = {"video": video_db, "text": text_db}
+ with open(
+ os.path.join(out_dir, "videoid_to_vectoridx.pkl"), "rb") as fr:
+ self.videoid_to_vectoridx = pickle.load(fr)
+ self.video_to_videoid = defaultdict(list)
+
+ def add(self, hidden_states, video_ids):
+ """hidden_states is a pair `(video, text)`"""
+ assert len(hidden_states) == len(video_ids), "{}, {}".format(
+ str(len(hidden_states)), str(len(video_ids)))
+ assert len(hidden_states.shape) == 3
+ assert len(self.video_to_videoid) == 0
+
+ valid_idx = []
+ for idx, video_id in enumerate(video_ids):
+ if video_id not in self.videoid_to_vectoridx:
+ valid_idx.append(idx)
+ self.videoid_to_vectoridx[video_id] = \
+ len(self.videoid_to_vectoridx)
+
+ batch_size = hidden_states.shape[0]
+ hidden_states = hidden_states[valid_idx]
+
+ hidden_states = np.transpose(hidden_states, (1, 0, 2)).copy()
+ if not self.db["video"].is_trained:
+ self.train_cache.append(hidden_states)
+ train_len = batch_size * len(self.train_cache)
+ if train_len < self.train_thres:
+ return
+
+ hidden_states = np.concatenate(self.train_cache, axis=1)
+ del self.train_cache
+ self.db["video"].train(hidden_states[0, :self.train_thres])
+ self.db["text"].train(hidden_states[1, :self.train_thres])
+ self.db["video"].add(hidden_states[0])
+ self.db["text"].add(hidden_states[1])
+
+ def get_clips_by_video_id(self, video_id):
+ if not self.video_to_videoid:
+ for video_id, video_clip, text_clip in self.videoid_to_vectoridx:
+ self.video_to_videoid[video_id].append(
+ (video_id, video_clip, text_clip))
+ return self.video_to_videoid[video_id]
+
+ def search(
+ self,
+ video_ids,
+ target_modality,
+ retri_factor=8
+ ):
+ if len(self.videoid_to_vectoridx) != len(self):
+ raise ValueError(
+ len(self.videoid_to_vectoridx),
+ len(self)
+ )
+
+ if not self.make_direct_maps_done:
+ self.make_direct_maps()
+ if self.vectoridx_to_videoid is None:
+ self.vectoridx_to_videoid = {
+ self.videoid_to_vectoridx[videoid]: videoid
+ for videoid in self.videoid_to_vectoridx
+ }
+ assert len(self.vectoridx_to_videoid) \
+ == len(self.videoid_to_vectoridx)
+
+ src_modality = "text" if target_modality == "video" else "video"
+
+ query_hidden_states = []
+ vector_ids = []
+ for video_id in video_ids:
+ vector_id = self.videoid_to_vectoridx[video_id]
+ vector_ids.append(vector_id)
+ query_hidden_state = self.db[src_modality].reconstruct(vector_id)
+ query_hidden_states.append(query_hidden_state)
+ query_hidden_states = np.stack(query_hidden_states)
+
+ # MultilingualFaissDataset uses the following; not sure the reason.
+ # faiss.ParameterSpace().set_index_parameter(self.db, "nprobe", 10)
+ _, index = self.db[target_modality].search(
+ query_hidden_states, retri_factor)
+ outputs = []
+ for sample_idx, sample in enumerate(index):
+ cands = []
+ for vector_idx in sample:
+ if vector_idx >= 0:
+ cands.append(
+ self.vectoridx_to_videoid[vector_idx]
+ )
+ outputs.append(cands)
+ return outputs
diff --git a/fairseq/examples/MMPT/mmpt/modules/vectorpool.py b/fairseq/examples/MMPT/mmpt/modules/vectorpool.py
new file mode 100644
index 0000000000000000000000000000000000000000..d2b23d2da888c714d7605d508c558fbad34e709d
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/modules/vectorpool.py
@@ -0,0 +1,246 @@
+# Copyright (c) Facebook, Inc. All Rights Reserved
+
+import torch
+import os
+import numpy as np
+import pickle
+
+from . import retri
+from ..utils import get_local_rank
+
+
+class VectorPool(object):
+ """
+ Base class of retrieval space.
+ """
+
+ def __init__(self, config):
+ from transformers import AutoConfig
+ self.hidden_size = AutoConfig.from_pretrained(
+ config.dataset.bert_name).hidden_size
+ self.retriever_cls = getattr(retri, config.retriever_cls)
+
+ def __call__(self, sample, **kwargs):
+ raise NotImplementedError
+
+ def build_retriver(
+ self,
+ retriever_cls=None,
+ hidden_size=None,
+ centroids=512,
+ db_type="flatl2",
+ examples_per_cent_to_train=48
+ ):
+
+ """merge results from multiple gpus and return a retriver.."""
+ self.retriver = retriever_cls(
+ hidden_size, centroids, db_type, examples_per_cent_to_train)
+ return self.retriver
+
+ def __repr__(self):
+ if hasattr(self, "retriver"):
+ retriver_name = str(len(self.retriver))
+ else:
+ retriver_name = "no retriver field yet"
+ return self.__class__.__name__ \
+ + "(" + retriver_name + ")"
+
+
+class VideoVectorPool(VectorPool):
+ """
+ average clips of a video as video representation.
+ """
+ def __init__(self, config):
+ super().__init__(config)
+ self.build_retriver(self.retriever_cls, self.hidden_size)
+
+ def __call__(self, sample, subsampling, **kwargs):
+ hidden_states = (
+ sample["pooled_video"] + sample["pooled_text"]) / 2.
+ hidden_states = hidden_states.view(
+ -1, subsampling,
+ hidden_states.size(-1))
+ hidden_states = torch.mean(hidden_states, dim=1)
+ hidden_states = hidden_states.cpu().detach().numpy()
+ video_ids = []
+ for offset_idx, video_id in enumerate(sample["video_id"]):
+ if isinstance(video_id, tuple) and len(video_id) == 3:
+ # a sharded video_id.
+ video_id = video_id[0]
+ video_ids.append(video_id)
+ assert len(video_ids) == len(hidden_states)
+ self.retriver.add(
+ hidden_states.astype("float32"),
+ video_ids
+ )
+
+
+class DistributedVectorPool(VectorPool):
+ """
+ support sync of multiple gpus/nodes.
+ """
+ def __init__(self, config):
+ super().__init__(config)
+ self.out_dir = os.path.join(
+ config.fairseq.checkpoint.save_dir,
+ "retri")
+ os.makedirs(self.out_dir, exist_ok=True)
+ self.hidden_states = []
+ self.video_ids = []
+
+ def build_retriver(
+ self,
+ retriever_cls=None,
+ hidden_size=None,
+ centroids=4096,
+ db_type="flatl2",
+ examples_per_cent_to_train=48
+ ):
+ if retriever_cls is None:
+ retriever_cls = self.retriever_cls
+ if hidden_size is None:
+ hidden_size = self.hidden_size
+ """merge results from multiple gpus and return a retriver.."""
+ if torch.distributed.is_initialized():
+ self.save()
+ # sync saving.
+ torch.distributed.barrier()
+ world_size = torch.distributed.get_world_size()
+ else:
+ world_size = 1
+ self.retriver = retriever_cls(
+ hidden_size, centroids, db_type, examples_per_cent_to_train)
+ # each gpu process has its own retriever.
+ for local_rank in range(world_size):
+ if get_local_rank() == 0:
+ print("load local_rank", local_rank)
+ hidden_states, video_ids = self.load(local_rank)
+ hidden_states = hidden_states.astype("float32")
+ self.retriver.add(hidden_states, video_ids)
+ return self.retriver
+
+ def load(self, local_rank):
+ hidden_states = np.load(
+ os.path.join(
+ self.out_dir,
+ "hidden_state" + str(local_rank) + ".npy"
+ )
+ )
+
+ with open(
+ os.path.join(
+ self.out_dir, "video_id" + str(local_rank) + ".pkl"),
+ "rb") as fr:
+ video_ids = pickle.load(fr)
+ return hidden_states, video_ids
+
+ def save(self):
+ hidden_states = np.vstack(self.hidden_states)
+ assert len(hidden_states) == len(self.video_ids), "{}, {}".format(
+ len(hidden_states),
+ len(self.video_ids)
+ )
+ local_rank = torch.distributed.get_rank() \
+ if torch.distributed.is_initialized() else 0
+
+ np.save(
+ os.path.join(
+ self.out_dir,
+ "hidden_state" + str(local_rank) + ".npy"),
+ hidden_states)
+
+ with open(
+ os.path.join(
+ self.out_dir,
+ "video_id" + str(local_rank) + ".pkl"),
+ "wb") as fw:
+ pickle.dump(
+ self.video_ids,
+ fw,
+ protocol=pickle.HIGHEST_PROTOCOL
+ )
+
+
+class DistributedVideoVectorPool(DistributedVectorPool):
+ """
+ average clips of a video as video representation.
+ """
+ def __call__(self, sample, subsampling, **kwargs):
+ hidden_states = (
+ sample["pooled_video"] + sample["pooled_text"]) / 2.
+ hidden_states = hidden_states.view(
+ -1, subsampling,
+ hidden_states.size(-1))
+ hidden_states = torch.mean(hidden_states, dim=1)
+ hidden_states = hidden_states.cpu().detach().numpy()
+ video_ids = []
+ for offset_idx, video_id in enumerate(sample["video_id"]):
+ if isinstance(video_id, tuple) and len(video_id) == 3:
+ # a sharded video_id.
+ video_id = video_id[0]
+ video_ids.append(video_id)
+ assert len(video_ids) == len(hidden_states)
+ self.hidden_states.append(hidden_states)
+ self.video_ids.extend(video_ids)
+
+
+# ------------ the following are deprecated --------------
+
+class TextClipVectorPool(VectorPool):
+ def __init__(self, config):
+ from transformers import AutoConfig
+ hidden_size = AutoConfig.from_pretrained(
+ config.dataset.bert_name).hidden_size
+ retriever_cls = getattr(retri, config.retriever_cls)
+ self.build_retriver(retriever_cls, hidden_size)
+
+ def __call__(self, sample, **kwargs):
+ clip_meta = sample["clip_meta"].cpu()
+ assert torch.all(torch.le(clip_meta[:, 4], clip_meta[:, 5]))
+ text_meta = [tuple(item.tolist()) for item in clip_meta[:, 3:]]
+
+ if hasattr(self, "retriver"):
+ # build_retriver is called.
+ self.retriver.add(
+ sample["pooled_text"].cpu().numpy().astype("float32"),
+ text_meta
+ )
+ else:
+ raise NotImplementedError
+
+
+class MMClipVectorPool(VectorPool):
+ """
+ Multimodal Clip-level vector pool.
+ """
+ def __init__(self, out_dir):
+ """use hidden_states to store `(video, text)`."""
+ """use video_ids to store `(video_id, start, end)`."""
+ super().__init__(out_dir)
+
+ def __call__(self, sample, **kwargs):
+ pooled_video = sample["pooled_video"].cpu().unsqueeze(1).numpy()
+ pooled_text = sample["pooled_text"].cpu().unsqueeze(1).numpy()
+
+ self.hidden_states.append(
+ np.concatenate([pooled_video, pooled_text], axis=1)
+ )
+
+ video_starts = sample["video_start"].cpu()
+ video_ends = sample["video_end"].cpu()
+ assert torch.all(torch.le(video_starts, video_ends))
+
+ text_starts = sample["text_start"].cpu()
+ text_ends = sample["text_end"].cpu()
+ assert torch.all(torch.le(text_starts, text_ends))
+ subsample_size = sample["pooled_video"].size(0) // len(sample["video_id"])
+ video_ids = [video_id for video_id in sample["video_id"]
+ for _ in range(subsample_size)
+ ]
+ for video_id, video_start, video_end, text_start, text_end in zip(
+ video_ids, video_starts, video_ends, text_starts, text_ends):
+ self.video_ids.append((
+ video_id,
+ (int(video_start), int(video_end)),
+ (int(text_start), int(text_end))
+ ))
diff --git a/fairseq/examples/MMPT/mmpt/processors/__init__.py b/fairseq/examples/MMPT/mmpt/processors/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..434d1d92b95846b0cd87625834c5d9bed279a44e
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/processors/__init__.py
@@ -0,0 +1,23 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+from .processor import *
+
+from .how2processor import *
+from .how2retriprocessor import *
+
+from .dsprocessor import *
+
+try:
+ from .rawvideoprocessor import *
+ from .codecprocessor import *
+ from .webvidprocessor import *
+ from .expprocessor import *
+ from .exphow2processor import *
+ from .exphow2retriprocessor import *
+ from .expcodecprocessor import *
+ from .expfeatureencoder import *
+ from .expdsprocessor import *
+except ImportError:
+ pass
diff --git a/fairseq/examples/MMPT/mmpt/processors/dedupprocessor.py b/fairseq/examples/MMPT/mmpt/processors/dedupprocessor.py
new file mode 100644
index 0000000000000000000000000000000000000000..8a1ad402cd890516457c188910ec265a47fc6e8e
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/processors/dedupprocessor.py
@@ -0,0 +1,242 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import random
+import json
+import pickle
+from tqdm import tqdm
+import os
+import numpy as np
+
+
+class CaptionDedupProcessor(object):
+ """remove overlapping of caption sentences(clip).
+ Some statistics:
+ caption:
+ {'t_clip_len': 246.6448431320854,
+ 'video_len': 281.09174795676245,
+ 'clip_tps': 0.8841283727427481,
+ 'video_tps': 0.7821156477732097,
+ 'min_clip_len': 0.0,
+ 'max_clip_len': 398.3,
+ 'mean_clip_len': 3.196580003006861,
+ 'num_clip': 77.15897706301081}
+
+ raw_caption:
+ {'t_clip_len': 238.95908778424115,
+ 'video_len': 267.5914859862507,
+ 'clip_tps': 2.4941363624267963,
+ 'video_tps': 2.258989769647173,
+ 'min_clip_len': 0.0,
+ 'max_clip_len': 398.3,
+ 'mean_clip_len': 3.0537954186814265,
+ 'num_clip': 78.24986779481756}
+ """
+
+ def __init__(self, pkl_file):
+ with open(pkl_file, "rb") as fd:
+ self.data = pickle.load(fd)
+ self.stat = {
+ "t_clip_len": [],
+ "video_len": [],
+ "clip_tps": [],
+ "video_tps": [],
+ "clip_len": [],
+ }
+
+ def __call__(self):
+ for idx, video_id in enumerate(tqdm(self.data)):
+ caption = json.loads(self.data[video_id])
+ caption = self._dedup(caption)
+ if idx < 4096: # for the first 4096 examples, compute the statistics.
+ self.save_stat(video_id, caption)
+ self.data[video_id] = json.dumps(caption)
+ self.print_stat()
+
+ def single(self, video_id):
+ caption = json.loads(self.data[video_id])
+ for clip_idx, (start, end, text) in enumerate(
+ zip(caption["start"], caption["end"], caption["text"])
+ ):
+ print(start, end, text)
+ print("@" * 100)
+ caption = self._dedup(caption)
+ for clip_idx, (start, end, text) in enumerate(
+ zip(caption["start"], caption["end"], caption["text"])
+ ):
+ print(start, end, text)
+ print("#" * 100)
+ self.save_stat(video_id, caption)
+ self.print_stat()
+
+ def finalize(self, tgt_fn):
+ with open(tgt_fn, "wb") as fw:
+ pickle.dump(self.data, fw, pickle.HIGHEST_PROTOCOL)
+
+ def save_stat(self, video_id, caption):
+ video_fn = os.path.join(
+ "data/feat/feat_how2_s3d", video_id + ".npy"
+ )
+ if os.path.isfile(video_fn):
+ with open(video_fn, "rb", 1) as fr: # 24 is the buffer size. buffered
+ version = np.lib.format.read_magic(fr)
+ shape, fortran, dtype = np.lib.format._read_array_header(fr, version)
+ video_len = shape[0]
+
+ t_clip_len = 0.0
+ t_tokens = 0
+ for idx, (start, end, text) in enumerate(
+ zip(caption["start"], caption["end"], caption["text"])
+ ):
+ clip_len = (
+ (end - max(caption["end"][idx - 1], start))
+ if idx > 0
+ else end - start
+ )
+ t_clip_len += clip_len
+ t_tokens += len(text.split(" "))
+ self.stat["clip_len"].append(clip_len)
+ self.stat["t_clip_len"].append(t_clip_len)
+ self.stat["video_len"].append(video_len)
+ self.stat["clip_tps"].append(t_tokens / t_clip_len)
+ self.stat["video_tps"].append(t_tokens / video_len)
+
+ def print_stat(self):
+ result = {
+ "t_clip_len": np.mean(self.stat["t_clip_len"]),
+ "video_len": np.mean(self.stat["video_len"]),
+ "clip_tps": np.mean(self.stat["clip_tps"]),
+ "video_tps": np.mean(self.stat["video_tps"]),
+ "min_clip_len": min(self.stat["clip_len"]),
+ "max_clip_len": max(self.stat["clip_len"]),
+ "mean_clip_len": np.mean(self.stat["clip_len"]),
+ "num_clip": len(self.stat["clip_len"]) / len(self.stat["video_tps"]),
+ }
+ print(result)
+
+ def _dedup(self, caption):
+ def random_merge(end_idx, start, end, text, starts, ends, texts):
+ if random.random() > 0.5:
+ # print(clip_idx, "[PARTIAL INTO PREV]", end_idx)
+ # overlapped part goes to the end of previous.
+ ends[-1] = max(ends[-1], start) # ?
+ rest_text = text[end_idx:].strip()
+ if rest_text:
+ starts.append(max(ends[-1], start))
+ ends.append(max(end, starts[-1]))
+ texts.append(rest_text)
+ else: # goes to the beginning of the current.
+ # strip the previous.
+ left_text = texts[-1][:-end_idx].strip()
+ if left_text:
+ # print(clip_idx, "[PREV PARTIAL INTO CUR]", end_idx)
+ ends[-1] = min(ends[-1], start)
+ texts[-1] = left_text
+ else:
+ # print(clip_idx, "[PREV LEFT NOTHING ALL INTO CUR]", end_idx)
+ starts.pop(-1)
+ ends.pop(-1)
+ texts.pop(-1)
+ starts.append(start)
+ ends.append(end)
+ texts.append(text)
+
+ starts, ends, texts = [], [], []
+ for clip_idx, (start, end, text) in enumerate(
+ zip(caption["start"], caption["end"], caption["text"])
+ ):
+ if not isinstance(text, str):
+ continue
+ text = text.replace("\n", " ").strip()
+ if len(text) == 0:
+ continue
+ starts.append(start)
+ ends.append(end)
+ texts.append(text)
+ break
+
+ for clip_idx, (start, end, text) in enumerate(
+ zip(
+ caption["start"][clip_idx + 1:],
+ caption["end"][clip_idx + 1:],
+ caption["text"][clip_idx + 1:],
+ )
+ ):
+ if not isinstance(text, str):
+ continue
+ text = text.replace("\n", " ").strip()
+ if len(text) == 0:
+ continue
+
+ # print(clip_idx, texts[-5:])
+ # print(clip_idx, start, end, text)
+ if texts[-1].endswith(text): # subset of prev caption -> merge
+ # print(clip_idx, "[MERGE INTO PREV]")
+ ends[-1] = max(ends[-1], end)
+ elif text.startswith(texts[-1]): # superset of prev caption -> merge
+ # print(clip_idx, "[PREV MERGE INTO CUR]")
+ texts[-1] = text
+ starts[-1] = min(starts[-1], start)
+ ends[-1] = max(ends[-1], end)
+ else: # overlapping or non-overlapping.
+ for end_idx in range(1, len(text) + 1):
+ if texts[-1].endswith(text[:end_idx]):
+ random_merge(end_idx, start, end, text, starts, ends, texts)
+ break
+ else:
+ starts.append(start)
+ ends.append(end)
+ texts.append(text)
+
+ assert (ends[-1] + 0.001) >= starts[-1] and len(
+ texts[-1]
+ ) > 0, "{} {} {} <- {} {} {}, {} {} {}".format(
+ str(starts[-1]),
+ str(ends[-1]),
+ texts[-1],
+ caption["start"][clip_idx - 1],
+ caption["end"][clip_idx - 1],
+ caption["text"][clip_idx - 1],
+ str(start),
+ str(end),
+ text,
+ )
+
+ return {"start": starts, "end": ends, "text": texts}
+
+
+if __name__ == "__main__":
+ import argparse
+
+ parser = argparse.ArgumentParser(description="dedup how2 caption")
+ parser.add_argument('--how2dir', default="data/how2")
+ args = parser.parse_args()
+
+ raw_caption_json = os.path.join(args.how2dir, "raw_caption.json")
+ raw_caption_pickle = os.path.join(args.how2dir, "raw_caption.pkl")
+ raw_caption_dedup_pickle = os.path.join(args.how2dir, "raw_caption_dedup.pkl")
+
+ def convert_to_pickle(src_fn, tgt_fn):
+ with open(src_fn) as fd:
+ captions = json.load(fd)
+
+ for video_id in captions:
+ captions[video_id] = json.dumps(captions[video_id])
+
+ with open(tgt_fn, "wb") as fw:
+ pickle.dump(captions, fw, pickle.HIGHEST_PROTOCOL)
+
+ if not os.path.isfile(raw_caption_pickle):
+ convert_to_pickle(raw_caption_json, raw_caption_pickle)
+
+ deduper = CaptionDedupProcessor(raw_caption_pickle)
+ deduper()
+ deduper.finalize(raw_caption_dedup_pickle)
+
+ """
+ # demo
+ deduper = CaptionDedupProcessor("data/how2/raw_caption.pkl")
+ deduper.single("HfIeQ9pzL5U")
+ """
diff --git a/fairseq/examples/MMPT/mmpt/processors/dsprocessor.py b/fairseq/examples/MMPT/mmpt/processors/dsprocessor.py
new file mode 100644
index 0000000000000000000000000000000000000000..ecebf0eea5c57b7846a2bd46ddffbdbf55bd83ab
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/processors/dsprocessor.py
@@ -0,0 +1,848 @@
+# Copyright (c) Facebook, Inc. All Rights Reserved
+
+"""
+Processors for all downstream (ds) tasks.
+"""
+
+import json
+import os
+import pickle
+import random
+import math
+import numpy as np
+import torch
+
+from collections import defaultdict
+
+from .processor import (
+ MetaProcessor,
+ VideoProcessor,
+ TextProcessor,
+ Aligner,
+ MMAttentionMask2DProcessor,
+)
+
+from .how2processor import TextGenerationProcessor
+
+
+# ------------- A General Aligner for all downstream tasks-----------------
+
+
+class DSAligner(Aligner):
+ """
+ Downstream (DS) aligner shared by all datasets.
+ """
+
+ def __call__(self, video_id, video_feature, text_feature, wps=0.7):
+ # random sample a starting sec for video.
+ video_start = 0
+ video_end = min(len(video_feature), self.max_video_len)
+ # the whole sequence is a single clip.
+ video_clips = {"start": [video_start], "end": [video_end]}
+
+ text_feature = {
+ "cap": [text_feature],
+ "start": [video_start],
+ "end": [len(text_feature) / wps],
+ }
+ text_clip_indexs = [0]
+
+ vfeats, vmasks = self._build_video_seq(
+ video_feature, video_clips
+ )
+ caps, cmasks = self._build_text_seq(
+ text_feature, text_clip_indexs
+ )
+
+ return {
+ "caps": caps,
+ "cmasks": cmasks,
+ "vfeats": vfeats,
+ "vmasks": vmasks,
+ "video_id": video_id,
+ }
+
+
+class NLGTextProcessor(TextProcessor):
+ """
+ Also return the original text as ref.
+ """
+ def __call__(self, text_id):
+ return super().__call__(text_id), text_id
+
+
+class DSNLGAligner(DSAligner):
+ """extend with the capability of 2d mask for generation."""
+ def __init__(self, config):
+ super().__init__(config)
+ self.attnmasker = MMAttentionMask2DProcessor()
+ from transformers import AutoTokenizer
+ tokenizer = AutoTokenizer.from_pretrained(
+ self.bert_name, use_fast=self.use_fast,
+ bos_token="[CLS]", eos_token="[SEP]"
+ )
+ self.tokenizer = tokenizer
+ self.bos_token_id = tokenizer.bos_token_id
+ self.eos_token_id = tokenizer.eos_token_id
+ self.textgen = TextGenerationProcessor(tokenizer)
+
+ def __call__(self, video_id, video_feature, text_feature):
+ output = super().__call__(video_id, video_feature, text_feature[0])
+ if self.split == "test":
+ # output.update({"ref": text_feature[1]})
+ output.update({"ref": self.tokenizer.decode(
+ output["caps"], skip_special_tokens=True)})
+ text_label = output["caps"]
+ cmasks = torch.BoolTensor([1] * text_label.size(0))
+ caps = torch.LongTensor([
+ self.cls_token_id,
+ self.sep_token_id,
+ self.bos_token_id])
+ else:
+ caps, text_label = self.textgen(output["caps"])
+ cmasks = output["cmasks"]
+
+ attention_mask = self.attnmasker(
+ output["vmasks"], cmasks, "textgen")
+
+ output.update({
+ "caps": caps,
+ "cmasks": cmasks,
+ "text_label": text_label,
+ "attention_mask": attention_mask,
+ })
+ return output
+
+
+# -------------------- MSRVTT ------------------------
+
+
+class MSRVTTMetaProcessor(MetaProcessor):
+ """MSRVTT dataset.
+ reference: `howto100m/msrvtt_dataloader.py`
+ """
+
+ def __init__(self, config):
+ super().__init__(config)
+ import pandas as pd
+ data = pd.read_csv(self._get_split_path(config))
+ # TODO: add a text1ka flag.
+ if config.split == "train" \
+ and config.full_test_path is not None \
+ and config.jsfusion_path is not None:
+ # add testing videos from full_test_path not used by jfusion.
+ additional_data = pd.read_csv(config.full_test_path)
+ jsfusion_data = pd.read_csv(config.jsfusion_path)
+
+ for video_id in additional_data["video_id"]:
+ if video_id not in jsfusion_data["video_id"].values:
+ data = data.append(
+ {"video_id": video_id}, ignore_index=True)
+
+ if config.dup is not None and config.split == "train":
+ data = data.append([data] * (config.dup - 1), ignore_index=True)
+ self.data = data
+
+ def __len__(self):
+ return len(self.data)
+
+ def __getitem__(self, idx):
+ """slightly modify with if condition to combine train/test."""
+ vid, sentence = None, None
+ vid = self.data["video_id"].values[idx]
+ if "sentence" in self.data: # for testing.
+ sentence = self.data["sentence"].values[idx]
+ else: # for training.
+ sentence = vid
+ return vid, sentence
+
+
+class MSRVTTTextProcessor(TextProcessor):
+ """MSRVTT dataset.
+ reference: `msrvtt_dataloader.py` `MSRVTT_TrainDataLoader`.
+ TODO (huxu): add max_words.
+ """
+
+ def __init__(self, config):
+ super().__init__(config)
+ self.sentences = None
+ if config.json_path is not None and config.split == "train":
+ with open(config.json_path) as fd:
+ self.data = json.load(fd)
+ self.sentences = defaultdict(list)
+ for s in self.data["sentences"]:
+ self.sentences[s["video_id"]].append(s["caption"])
+
+ def __call__(self, text_id):
+ if self.sentences is not None:
+ rind = random.randint(0, len(self.sentences[text_id]) - 1)
+ sentence = self.sentences[text_id][rind]
+ else:
+ sentence = text_id
+ caption = self.tokenizer(sentence, add_special_tokens=False)
+ return caption["input_ids"]
+
+
+class MSRVTTNLGTextProcessor(MSRVTTTextProcessor):
+ """TODO: change dsaligner and merge to avoid any NLG text processor."""
+ def __call__(self, text_id):
+ if self.sentences is not None:
+ rind = random.randint(0, len(self.sentences[text_id]) - 1)
+ sentence = self.sentences[text_id][rind]
+ else:
+ sentence = text_id
+ caption = self.tokenizer(sentence, add_special_tokens=False)
+ return caption["input_ids"], sentence
+
+
+class MSRVTTQAMetaProcessor(MetaProcessor):
+ """MSRVTT-QA: retrieval-based multi-choice QA from JSFusion dataset.
+ For simplicity, we use the train retrieval model.
+ reference: `https://github.com/yj-yu/lsmdc`
+ """
+
+ def __init__(self, config):
+ super().__init__(config)
+ import pandas as pd
+ csv_data = pd.read_csv(self._get_split_path(config), sep="\t")
+ data = []
+ for video_id, a1, a2, a3, a4, a5, answer in zip(
+ csv_data["vid_key"].values,
+ csv_data["a1"].values,
+ csv_data["a2"].values,
+ csv_data["a3"].values,
+ csv_data["a4"].values,
+ csv_data["a5"].values,
+ csv_data["answer"].values):
+ video_id = video_id.replace("msr", "video")
+ data.append((video_id, (answer, [a1, a2, a3, a4, a5])))
+ self.data = data
+
+ def __len__(self):
+ return len(self.data)
+
+ def __getitem__(self, idx):
+ return self.data[idx]
+
+
+class MSRVTTQATextProcessor(TextProcessor):
+ """MSRVTT-QA dataset.
+ text_ans is of format `(answer, [a1, a2, a3, a4, a5])`.
+ """
+
+ def __call__(self, text_ans):
+ for ans_idx, ans in enumerate(text_ans[1]):
+ if isinstance(ans, str):
+ text_ans[1][ans_idx] = self.tokenizer(ans, add_special_tokens=False)["input_ids"]
+ return text_ans
+
+
+class MSRVTTQAAligner(DSAligner):
+ """MSRVTT dataset.
+ similar to sample in how2.
+ we call __call__ multiple times.
+ """
+
+ def __call__(self, video_id, video_feature, text_feature, wps=0.7):
+ caps = []
+ cmasks = []
+ answer = text_feature[0]
+ for ans_idx, _text_feature in enumerate(text_feature[1]):
+ output = super().__call__(
+ video_id, video_feature, _text_feature, wps)
+ caps.append(output["caps"])
+ cmasks.append(output["cmasks"])
+ output.update({
+ "caps": torch.stack(caps),
+ "cmasks": torch.stack(cmasks),
+ "answers": torch.LongTensor([answer]),
+ })
+ return output
+
+
+# -------------------- Youcook -----------------------
+
+
+class YoucookMetaProcessor(MetaProcessor):
+ """Youcook dataset.
+ reference: `howto100m/youcook_dataloader.py`
+ note that the data can be different as the
+ (1) some videos already in Howto100m are removed.
+ (2) stop words are removed from caption
+ TODO (huxu): make a flag to load the original caption.
+ (see youcookii_annotations_trainval.json).
+
+ The max_video_len can be 264 and text can be 64 tokens.
+ In reality we may not need that long. see projects/task/youcook.yaml
+ """
+
+ def __init__(self, config):
+ super().__init__(config)
+ vfeat_dir = config.vfeat_dir
+ print(self._get_split_path(config))
+ with open(self._get_split_path(config), "rb") as fd:
+ data = pickle.load(fd)
+ all_valid_video_ids = set(
+ [os.path.splitext(fn)[0] for fn in os.listdir(vfeat_dir)]
+ )
+ recs = []
+ video_ids = set()
+ valid_video_ids = set()
+ for rec in data: # filter videos not available.
+ udl_idx = rec["id"].rindex("_")
+ video_id = rec["id"][:udl_idx]
+ video_ids.add(video_id)
+ if video_id in all_valid_video_ids:
+ valid_video_ids.add(video_id)
+ recs.append(rec)
+ print("total video_ids in .pkl", len(video_ids))
+ print("valid video_ids in .pkl", len(valid_video_ids))
+ print("please verify {train,val}_list.txt")
+ data = recs
+ self.data = data
+
+ with open(config.trainval_annotation) as fd:
+ self.youcook_annotation = json.load(fd)["database"]
+ if config.use_annotation_text is True:
+ print("using text in annotation.")
+ self.use_annotation_caption = True
+ else:
+ self.use_annotation_caption = False
+
+ def __getitem__(self, idx):
+ def _get_video_and_caption(rec):
+ vid = rec["id"]
+ udl_idx = vid.rindex("_")
+ video_id, clip_id = vid[:udl_idx], int(vid[udl_idx + 1:])
+ clip = self.youcook_annotation[video_id]["annotations"][clip_id]
+ start, end = clip["segment"]
+ if self.use_annotation_caption:
+ caption = clip["sentence"]
+ else:
+ caption = rec["caption"]
+ return (video_id, start, end), caption
+
+ rec = self.data[idx]
+ video_info, text_info = _get_video_and_caption(rec)
+ return video_info, text_info
+
+
+class YoucookVideoProcessor(VideoProcessor):
+ """video_fn is a tuple of (video_id, start, end) now."""
+
+ def __call__(self, video_fn):
+ video_id, start, end = video_fn
+ feat = np.load(os.path.join(self.vfeat_dir, video_id + ".npy"))
+ return feat[start:end]
+
+
+class YoucookNLGMetaProcessor(MetaProcessor):
+ """NLG uses the original split:
+ `train_list.txt` and `val_list.txt`
+ """
+
+ def __init__(self, config):
+ super().__init__(config)
+ vfeat_dir = config.vfeat_dir
+ print(self._get_split_path(config))
+ with open(self._get_split_path(config)) as fd:
+ video_ids = [
+ line.strip().split("/")[1] for line in fd.readlines()]
+ print("total video_ids in train/val_list.txt", len(video_ids))
+
+ all_valid_video_ids = set(
+ [os.path.splitext(fn)[0] for fn in os.listdir(vfeat_dir)]
+ )
+ video_ids = [
+ video_id for video_id in video_ids
+ if video_id in all_valid_video_ids]
+
+ print("valid video_ids in train/val_list.txt", len(video_ids))
+ with open(config.trainval_annotation) as fd:
+ self.youcook_annotation = json.load(fd)["database"]
+
+ data = []
+ for video_id in video_ids:
+ for clip in self.youcook_annotation[video_id]["annotations"]:
+ start, end = clip["segment"]
+ caption = clip["sentence"]
+ data.append(((video_id, start, end), caption))
+ self.data = data
+
+ def __getitem__(self, idx):
+ return self.data[idx]
+
+
+# --------------------- CrossTask -------------------------
+
+class CrossTaskMetaProcessor(MetaProcessor):
+ def __init__(self, config):
+ super().__init__(config)
+ np.random.seed(0) # deterministic random split.
+ task_vids = self._get_vids(
+ config.train_csv_path,
+ config.vfeat_dir,
+ config.annotation_path)
+
+ val_vids = self._get_vids(
+ config.val_csv_path,
+ config.vfeat_dir,
+ config.annotation_path)
+
+ # filter out those task and vids appear in val_vids.
+ task_vids = {
+ task: [
+ vid for vid in vids
+ if task not in val_vids or vid not in val_vids[task]]
+ for task, vids in task_vids.items()}
+
+ primary_info = self._read_task_info(config.primary_path)
+ test_tasks = set(primary_info['steps'].keys())
+
+ # if args.use_related:
+ related_info = self._read_task_info(config.related_path)
+ task_steps = {**primary_info['steps'], **related_info['steps']}
+ n_steps = {**primary_info['n_steps'], **related_info['n_steps']}
+ # else:
+ # task_steps = primary_info['steps']
+ # n_steps = primary_info['n_steps']
+ all_tasks = set(n_steps.keys())
+ # filter and keep task in primary or related.
+ task_vids = {
+ task: vids for task, vids in task_vids.items()
+ if task in all_tasks}
+ # vocab-by-step matrix (A) and vocab (M)
+ # (huxu): we do not use BoW.
+ # A, M = self._get_A(task_steps, share="words")
+
+ train_vids, test_vids = self._random_split(
+ task_vids, test_tasks, config.n_train)
+ print("train_num_videos", sum(len(vids) for vids in train_vids.values()))
+ print("test_num_videos", sum(len(vids) for vids in test_vids.values()))
+ # added by huxu to automatically determine the split.
+ split_map = {
+ "train": train_vids,
+ "valid": test_vids,
+ "test": test_vids
+ }
+ task_vids = split_map[config.split]
+
+ self.vids = []
+ for task, vids in task_vids.items():
+ self.vids.extend([(task, vid) for vid in vids])
+ self.task_steps = task_steps
+ self.n_steps = n_steps
+
+ def __getitem__(self, idx):
+ task, vid = self.vids[idx]
+ n_steps = self.n_steps[task]
+ steps = self.task_steps[task]
+ assert len(steps) == n_steps
+ return (task, vid, steps, n_steps), (task, vid, steps, n_steps)
+
+ def __len__(self):
+ return len(self.vids)
+
+ def _random_split(self, task_vids, test_tasks, n_train):
+ train_vids = {}
+ test_vids = {}
+ for task, vids in task_vids.items():
+ if task in test_tasks and len(vids) > n_train:
+ train_vids[task] = np.random.choice(
+ vids, n_train, replace=False).tolist()
+ test_vids[task] = [
+ vid for vid in vids if vid not in train_vids[task]]
+ else:
+ train_vids[task] = vids
+ return train_vids, test_vids
+
+ def _get_vids(self, path, vfeat_dir, annotation_path):
+ """refactored from
+ https://github.com/DmZhukov/CrossTask/blob/master/data.py
+ changes: add `vfeat_dir` to check if the video is available.
+ add `annotation_path` to check if the video is available.
+ """
+
+ task_vids = {}
+ with open(path, 'r') as f:
+ for line in f:
+ task, vid, url = line.strip().split(',')
+ # double check the video is available.
+ if not os.path.exists(
+ os.path.join(vfeat_dir, vid + ".npy")):
+ continue
+ # double check the annotation is available.
+ if not os.path.exists(os.path.join(
+ annotation_path,
+ task + "_" + vid + ".csv")):
+ continue
+ if task not in task_vids:
+ task_vids[task] = []
+ task_vids[task].append(vid)
+ return task_vids
+
+ def _read_task_info(self, path):
+ titles = {}
+ urls = {}
+ n_steps = {}
+ steps = {}
+ with open(path, 'r') as f:
+ idx = f.readline()
+ while idx != '':
+ idx = idx.strip()
+ titles[idx] = f.readline().strip()
+ urls[idx] = f.readline().strip()
+ n_steps[idx] = int(f.readline().strip())
+ steps[idx] = f.readline().strip().split(',')
+ next(f)
+ idx = f.readline()
+ return {
+ 'title': titles,
+ 'url': urls,
+ 'n_steps': n_steps,
+ 'steps': steps
+ }
+
+ def _get_A(self, task_steps, share="words"):
+ raise ValueError("running get_A is not allowed for BERT.")
+ """Step-to-component matrices."""
+ if share == 'words':
+ # share words
+ task_step_comps = {
+ task: [step.split(' ') for step in steps]
+ for task, steps in task_steps.items()}
+ elif share == 'task_words':
+ # share words within same task
+ task_step_comps = {
+ task: [[task+'_'+tok for tok in step.split(' ')] for step in steps]
+ for task, steps in task_steps.items()}
+ elif share == 'steps':
+ # share whole step descriptions
+ task_step_comps = {
+ task: [[step] for step in steps] for task, steps in task_steps.items()}
+ else:
+ # no sharing
+ task_step_comps = {
+ task: [[task+'_'+step] for step in steps]
+ for task, steps in task_steps.items()}
+ # BERT tokenizer here?
+ vocab = []
+ for task, steps in task_step_comps.items():
+ for step in steps:
+ vocab.extend(step)
+ vocab = {comp: m for m, comp in enumerate(set(vocab))}
+ M = len(vocab)
+ A = {}
+ for task, steps in task_step_comps.items():
+ K = len(steps)
+ a = torch.zeros(M, K)
+ for k, step in enumerate(steps):
+ a[[vocab[comp] for comp in step], k] = 1
+ a /= a.sum(dim=0)
+ A[task] = a
+ return A, M
+
+
+class CrossTaskVideoProcessor(VideoProcessor):
+ def __call__(self, video_fn):
+ task, vid, steps, n_steps = video_fn
+ video_fn = os.path.join(self.vfeat_dir, vid + ".npy")
+ feat = np.load(video_fn)
+ return feat
+
+
+class CrossTaskTextProcessor(TextProcessor):
+ def __call__(self, text_id):
+ task, vid, steps, n_steps = text_id
+ step_ids = []
+ for step_str in steps:
+ step_ids.append(
+ self.tokenizer(step_str, add_special_tokens=False)["input_ids"]
+ )
+ return step_ids
+
+
+class CrossTaskAligner(Aligner):
+ """
+ TODO: it's not clear yet the formulation of the task; finish this later.
+ """
+ def __init__(self, config):
+ super().__init__(config)
+ self.annotation_path = config.annotation_path
+ self.sliding_window = config.sliding_window
+ self.sliding_window_size = config.sliding_window_size
+
+ def __call__(self, video_id, video_feature, text_feature):
+ task, vid, steps, n_steps = video_id
+ annot_path = os.path.join(
+ self.annotation_path, task + '_' + vid + '.csv')
+ video_len = len(video_feature)
+
+ labels = torch.from_numpy(self._read_assignment(
+ video_len, n_steps, annot_path)).float()
+
+ vfeats, vmasks, targets = [], [], []
+ # sliding window on video features and targets.
+ for window_start in range(0, video_len, self.sliding_window):
+ video_start = 0
+ video_end = min(video_len - window_start, self.sliding_window_size)
+ video_clip = {"start": [video_start], "end": [video_end]}
+
+ vfeat, vmask = self._build_video_seq(
+ video_feature[window_start: window_start + video_end],
+ video_clip
+ )
+
+ target = labels[window_start: window_start + video_end]
+ assert len(vfeat) >= len(target), "{},{}".format(len(vfeat), len(target))
+ # TODO: randomly drop all zero targets for training ?
+ # if self.split == "train" and target.sum() == 0:
+ # continue
+ vfeats.append(vfeat)
+ vmasks.append(vmask)
+ targets.append(target)
+
+ if (video_len - window_start) <= self.sliding_window_size:
+ break
+
+ vfeats = torch.stack(vfeats)
+ vmasks = torch.stack(vmasks)
+ targets = torch.cat(targets, dim=0)
+
+ caps, cmasks = [], []
+ for step in text_feature:
+ step_text_feature = {"start": [0], "end": [1], "cap": [step]}
+ step_text_clip_index = [0]
+ cap, cmask = self._build_text_seq(
+ step_text_feature, step_text_clip_index
+ )
+ caps.append(cap)
+ cmasks.append(cmask)
+ caps = torch.stack(caps)
+ cmasks = torch.stack(cmasks)
+
+ return {
+ "caps": caps,
+ "cmasks": cmasks,
+ "vfeats": vfeats, # X for original code.
+ "vmasks": vmasks,
+ "targets": targets,
+ "video_id": vid,
+ "task": task,
+ "video_len": video_len # for later checking.
+ }
+
+ def _read_assignment(self, T, K, path):
+ """
+ refactored from https://github.com/DmZhukov/CrossTask/blob/master/data.py
+ Howto interpret contraints on loss that is going to be minimized:
+ lambd is a big number;
+ self.lambd * C is a big number for all valid position (csv stores invalids)
+
+ def forward(self, O, Y, C):
+ return (Y*(self.lambd * C - self.lsm(O))).mean(dim=0).sum()
+
+ This will load the csv file and fill-in the step col from start to end rows.
+ """
+
+ Y = np.zeros([T, K], dtype=np.uint8)
+ with open(path, 'r') as f:
+ for line in f:
+ step, start, end = line.strip().split(',')
+ start = int(math.floor(float(start)))
+ end = int(math.ceil(float(end)))
+ step = int(step) - 1
+ Y[start:end, step] = 1
+ return Y
+
+
+# --------------------- COIN -------------------------
+
+class MetaTextBinarizer(Aligner):
+ def __call__(self, text_feature):
+ text_feature = {
+ "cap": [text_feature],
+ "start": [0.],
+ "end": [100.],
+ }
+ text_clip_indexs = [0]
+
+ caps, cmasks = self._build_text_seq(
+ text_feature, text_clip_indexs
+ )
+ return {"caps": caps, "cmasks": cmasks}
+
+
+class COINActionSegmentationMetaProcessor(MetaProcessor):
+ split_map = {
+ "train": "training",
+ "valid": "testing",
+ "test": "testing",
+ }
+
+ def __init__(self, config):
+ super().__init__(config)
+ with open(self._get_split_path(config)) as fr:
+ database = json.load(fr)["database"]
+ id2label = {}
+ data = []
+ # filter the data by split.
+ for video_id, rec in database.items():
+ # always use testing to determine label_set
+ if rec["subset"] == "testing":
+ for segment in rec["annotation"]:
+ id2label[int(segment["id"])] = segment["label"]
+ # text_labels is used for ZS setting
+ self.text_labels = ["none"] * len(id2label)
+ for label_id in id2label:
+ self.text_labels[label_id-1] = id2label[label_id]
+
+ id2label[0] = "O"
+ print("num of labels", len(id2label))
+
+ for video_id, rec in database.items():
+ if not os.path.isfile(os.path.join(config.vfeat_dir, video_id + ".npy")):
+ continue
+ if rec["subset"] == COINActionSegmentationMetaProcessor.split_map[self.split]:
+ starts, ends, labels = [], [], []
+ for segment in rec["annotation"]:
+ start, end = segment["segment"]
+ label = int(segment["id"])
+ starts.append(start)
+ ends.append(end)
+ labels.append(label)
+ data.append(
+ (video_id, {"start": starts, "end": ends, "label": labels}))
+ self.data = data
+
+ def meta_text_labels(self, config):
+ from transformers import default_data_collator
+ from ..utils import get_local_rank
+
+ text_processor = TextProcessor(config)
+ binarizer = MetaTextBinarizer(config)
+ # TODO: add prompts to .yaml.
+ text_labels = [label for label in self.text_labels]
+
+ if get_local_rank() == 0:
+ print(text_labels)
+
+ outputs = []
+ for text_label in text_labels:
+ text_feature = text_processor(text_label)
+ outputs.append(binarizer(text_feature))
+ return default_data_collator(outputs)
+
+ def __getitem__(self, idx):
+ return self.data[idx]
+
+
+class COINActionSegmentationTextProcessor(TextProcessor):
+ def __call__(self, text_label):
+ return text_label
+
+
+class COINActionSegmentationAligner(Aligner):
+ def __init__(self, config):
+ super().__init__(config)
+ self.sliding_window = config.sliding_window
+ self.sliding_window_size = config.sliding_window_size
+
+ def __call__(self, video_id, video_feature, text_feature):
+ starts, ends, label_ids = text_feature["start"], text_feature["end"], text_feature["label"]
+ # sliding window.
+ video_len = len(video_feature)
+
+ vfeats, vmasks, targets = [], [], []
+ # sliding window on video features and targets.
+ for window_start in range(0, video_len, self.sliding_window):
+ video_start = 0
+ video_end = min(video_len - window_start, self.sliding_window_size)
+ video_clip = {"start": [video_start], "end": [video_end]}
+ vfeat, vmask = self._build_video_seq(
+ video_feature[window_start: window_start + video_end],
+ video_clip
+ )
+ # covers video length only.
+ target = torch.full_like(vmask, -100, dtype=torch.long)
+ target[vmask] = 0
+ for start, end, label_id in zip(starts, ends, label_ids):
+ if (window_start < end) and (start < (window_start + video_end)):
+ start_offset = max(0, math.floor(start) - window_start)
+ end_offset = min(video_end, math.ceil(end) - window_start)
+ target[start_offset:end_offset] = label_id
+ vfeats.append(vfeat)
+ vmasks.append(vmask)
+ targets.append(target)
+ if (video_len - window_start) <= self.sliding_window_size:
+ break
+
+ vfeats = torch.stack(vfeats)
+ vmasks = torch.stack(vmasks)
+ targets = torch.stack(targets)
+ video_targets = torch.full((video_len,), 0)
+ for start, end, label_id in zip(starts, ends, label_ids):
+ start_offset = max(0, math.floor(start))
+ end_offset = min(video_len, math.ceil(end))
+ video_targets[start_offset:end_offset] = label_id
+
+ caps = torch.LongTensor(
+ [[self.cls_token_id, self.sep_token_id,
+ self.pad_token_id, self.sep_token_id]],
+ ).repeat(vfeats.size(0), 1)
+ cmasks = torch.BoolTensor(
+ [[0, 1, 0, 1]] # pad are valid for attention.
+ ).repeat(vfeats.size(0), 1)
+ return {
+ "caps": caps,
+ "cmasks": cmasks,
+ "vfeats": vfeats, # X for original code.
+ "vmasks": vmasks,
+ "targets": targets,
+ "video_id": video_id,
+ "video_len": video_len, # for later checking.
+ "video_targets": video_targets
+ }
+
+
+class DiDeMoMetaProcessor(MetaProcessor):
+ """reference: https://github.com/LisaAnne/LocalizingMoments/blob/master/utils/eval.py
+ https://github.com/LisaAnne/LocalizingMoments/blob/master/utils/data_processing.py
+ """
+ def __init__(self, config):
+ super().__init__(config)
+
+ assert "test" in self._get_split_path(config), "DiDeMo only supports zero-shot testing for now."
+
+ with open(self._get_split_path(config)) as data_file:
+ json_data = json.load(data_file)
+
+ data = []
+ for record in json_data:
+ data.append((record["video"], record["description"]))
+ self.data = data
+
+ def __len__(self):
+ return len(self.data)
+
+ def __getitem__(self, idx):
+ return self.data[idx]
+
+
+class DiDeMoTextProcessor(TextProcessor):
+ """reference: https://github.com/LisaAnne/LocalizingMoments/blob/master/utils/eval.py
+ https://github.com/LisaAnne/LocalizingMoments/blob/master/utils/data_processing.py
+ """
+
+ def __call__(self, text):
+ return self.tokenizer(text, add_special_tokens=False)["input_ids"]
+
+
+class DiDeMoAligner(DSAligner):
+ """
+ check video length.
+ """
+
+ def __call__(self, video_id, video_feature, text_feature):
+ # print(video_feature.shape[0])
+ return super().__call__(video_id, video_feature, text_feature)
diff --git a/fairseq/examples/MMPT/mmpt/processors/how2processor.py b/fairseq/examples/MMPT/mmpt/processors/how2processor.py
new file mode 100644
index 0000000000000000000000000000000000000000..bed2168b1df28babc7a12e81b6bc31d36d73bc99
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/processors/how2processor.py
@@ -0,0 +1,887 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# Copyright (c) Facebook, Inc. All Rights Reserved
+
+
+import torch
+import math
+import pickle
+import random
+import os
+import numpy as np
+
+from collections import deque
+from typing import Optional, Tuple, List
+from .processor import (
+ Processor,
+ MetaProcessor,
+ TextProcessor,
+ Aligner,
+ MMAttentionMask2DProcessor
+)
+
+from ..utils import ShardedTensor
+
+
+class How2MetaProcessor(MetaProcessor):
+ def __init__(self, config):
+ super().__init__(config)
+ path = self._get_split_path(config)
+ with open(path) as fd:
+ self.data = [line.strip() for line in fd]
+
+ def __getitem__(self, idx):
+ video_id = self.data[idx]
+ return video_id, video_id
+
+
+class ShardedHow2MetaProcessor(How2MetaProcessor):
+ def __init__(self, config):
+ super().__init__(config)
+ self.split = str(config.split)
+ self.vfeat_dir = config.vfeat_dir
+ self._init_shard()
+
+ def _init_shard(self):
+ if self.split == "train":
+ meta_fn = os.path.join(self.vfeat_dir, "train" + "_meta.pkl")
+ with open(meta_fn, "rb") as fr:
+ meta = pickle.load(fr)
+ elif self.split == "valid":
+ meta_fn = os.path.join(self.vfeat_dir, "val" + "_meta.pkl")
+ with open(meta_fn, "rb") as fr:
+ meta = pickle.load(fr)
+ elif self.split == "test":
+ print("use how2 val as test.")
+ meta_fn = os.path.join(self.vfeat_dir, "val" + "_meta.pkl")
+ with open(meta_fn, "rb") as fr:
+ meta = pickle.load(fr)
+ else:
+ raise ValueError("unsupported for MetaProcessor:", self.split)
+ video_id_to_shard = {}
+ for shard_id in meta:
+ for video_idx, video_id in enumerate(meta[shard_id]):
+ video_id_to_shard[video_id] = (shard_id, video_idx)
+ self.video_id_to_shard = video_id_to_shard
+
+ def __getitem__(self, idx):
+ video_id, video_id = super().__getitem__(idx)
+ shard_id, shard_idx = self.video_id_to_shard[video_id]
+ meta = (video_id, idx, shard_id, shard_idx)
+ return meta, meta
+
+
+class ShardedVideoProcessor(Processor):
+ """
+ mmaped shards of numpy video features.
+ """
+
+ def __init__(self, config):
+ self.split = str(config.split)
+ self.vfeat_dir = config.vfeat_dir
+
+ def __call__(self, video_id):
+ _, _, shard_id, video_idx = video_id
+ if self.split == "train":
+ shard = ShardedTensor.load(
+ os.path.join(self.vfeat_dir, "train" + "_" + str(shard_id)),
+ "r"
+ )
+ elif self.split == "valid":
+ shard = ShardedTensor.load(
+ os.path.join(self.vfeat_dir, "val" + "_" + str(shard_id)),
+ "r"
+ )
+ elif self.split == "test":
+ shard = ShardedTensor.load(
+ os.path.join(self.vfeat_dir, "val" + "_" + str(shard_id)),
+ "r"
+ )
+ else:
+ raise ValueError("unknown split", self.split)
+ feat = shard[video_idx]
+ return feat
+
+
+class ShardedTextProcessor(Processor):
+ def __init__(self, config):
+ self.tfeat_dir = str(config.tfeat_dir)
+ self.split = str(config.split)
+
+ def __call__(self, video_id):
+ _, _, shard_id, shard_idx = video_id
+ if self.split == "train":
+ target_path = self.tfeat_dir + "train" + "_" + str(shard_id)
+ elif self.split == "valid":
+ target_path = self.tfeat_dir + "val" + "_" + str(shard_id)
+ elif self.split == "test":
+ target_path = self.tfeat_dir + "val" + "_" + str(shard_id)
+ else:
+ raise ValueError("unknown split", self.split)
+
+ startend = ShardedTensor.load(
+ target_path + ".startends", "r")[shard_idx]
+ cap_ids = ShardedTensor.load(
+ target_path + ".caps_ids", "r")[shard_idx]
+ cap = []
+ for clip_idx in range(len(cap_ids)):
+ clip = cap_ids[clip_idx]
+ cap.append(clip[clip != -1].tolist())
+ start, end = startend[:, 0].tolist(), startend[:, 1].tolist()
+ return {"start": start, "end": end, "cap": cap}
+
+
+class FixedLenAligner(Aligner):
+ """
+ In the model we assume text is on the left (closer to BERT formulation)
+ and video is on the right.
+ We fix the total length of text + video.
+ max_video_len is in number of secs.
+ max_text_len is in number of tokens.
+
+ special tokens formats:
+ we use the format [CLS] [SEP] text tokens [SEP] [PAD] ...
+ [CLS] will be splitted out into:
+ [CLS] video tokens [SEP] text tokens [SEP] [PAD] ...
+ token_type_ids will be generated by the model (for now).
+ 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1
+ | first sequence | second sequence |
+ so each sequence owns a [SEP] token for no-ops.
+ """
+
+ def __init__(self, config):
+ super().__init__(config)
+ self.text_clip_sampler = TextClipSamplingProcessor(
+ self.max_len - self.max_video_len - 3
+ )
+ """
+ decide subsampling:
+ `config.subsampling` will change batch_size in trainer.
+ `config.clip_per_video` (used by RetriTask) doesn't
+ change batch_size in trainer.
+ """
+ subsampling = config.subsampling \
+ if config.subsampling is not None else None
+ if config.clip_per_video is not None:
+ subsampling = config.clip_per_video
+ self.subsampling = subsampling
+
+ def _get_text_maxlen(self):
+ # use max text len
+ return self.text_clip_sampler.max_text_len
+
+ def __call__(self, video_id, video_feature, text_feature):
+ from transformers import default_data_collator
+ video_idx = video_id[1]
+ if self.subsampling is not None and self.subsampling >= 1:
+ batch = []
+ for _ in range(self.subsampling):
+ centerclip_idx = random.randint(
+ 0, len(text_feature["start"]) - 1)
+ batch.append(
+ self.sampling(
+ video_idx,
+ video_feature,
+ text_feature,
+ centerclip_idx,
+ self._get_text_maxlen()
+ ))
+ batch = self.batch_post_processing(batch, video_feature)
+ batch = default_data_collator(batch)
+ else:
+ raise ValueError(
+ "dataset.subsampling must be >= 1 for efficient video loading.")
+ batch = self.sampling(video_idx, video_feature, text_feature)
+ batch = self.batch_post_processing(batch, video_feature)
+
+ batch["video_id"] = video_id if isinstance(video_id, str) \
+ else video_id[0]
+ # e2e: make sure frame ids is into tensor.
+ assert torch.is_tensor(batch["vfeats"])
+ return batch
+
+ def sampling(
+ self,
+ video_idx,
+ video_feature,
+ text_feature,
+ centerclip_idx=None,
+ sampled_max_text_len=None,
+ ):
+ text_clip_indexs = self.text_clip_sampler(
+ text_feature, centerclip_idx,
+ sampled_max_text_len
+ )
+ if isinstance(video_feature, np.ndarray):
+ video_len = len(video_feature)
+ else:
+ video_len = math.ceil(text_feature["end"][-1])
+
+ video_end = min(
+ math.ceil(text_feature["end"][text_clip_indexs[-1]]),
+ video_len
+ )
+ video_start = max(
+ min(
+ math.floor(text_feature["start"][text_clip_indexs[0]]),
+ video_end),
+ 0
+ )
+
+ video_clips = {"start": [video_start], "end": [video_end]}
+
+ # tensorize.
+ vfeats, vmasks = self._build_video_seq(
+ video_feature, video_clips
+ )
+ caps, cmasks = self._build_text_seq(
+ text_feature, text_clip_indexs
+ )
+
+ text_start = text_clip_indexs[0]
+ text_end = text_clip_indexs[-1] + 1
+
+ return {
+ "caps": caps,
+ "cmasks": cmasks,
+ "vfeats": vfeats,
+ "vmasks": vmasks,
+ "video_start": video_start,
+ "video_end": video_end,
+ "text_start": text_start,
+ "text_end": text_end,
+ }
+
+
+class VariedLenAligner(FixedLenAligner):
+ def __init__(self, config):
+ super().__init__(config)
+ self.sampled_min_len = config.sampled_min_len
+ self.sampled_max_len = config.sampled_max_len
+
+ def _get_text_maxlen(self):
+ return random.randint(self.sampled_min_len, self.sampled_max_len)
+
+
+class StartClipAligner(VariedLenAligner):
+ def sampling(
+ self,
+ video_idx,
+ video_feature,
+ text_feature,
+ centerclip_idx=None,
+ sampled_max_text_len=None,
+ ):
+ return super().sampling(
+ video_idx, video_feature, text_feature, 0)
+
+
+class OverlappedAligner(VariedLenAligner):
+ """video clip and text clip has overlappings
+ but may not be the same start/end."""
+ def __init__(self, config):
+ super().__init__(config)
+ self.sampled_video_min_len = config.sampled_video_min_len
+ self.sampled_video_max_len = config.sampled_video_max_len
+
+ self.video_clip_sampler = VideoClipSamplingProcessor()
+
+ def _get_video_maxlen(self):
+ return random.randint(
+ self.sampled_video_min_len, self.sampled_video_max_len)
+
+ def sampling(
+ self,
+ video_idx,
+ video_feature,
+ text_feature,
+ centerclip_idx=None,
+ sampled_max_text_len=None,
+ ):
+ text_clip_indexs = self.text_clip_sampler(
+ text_feature, centerclip_idx,
+ sampled_max_text_len
+ )
+ if isinstance(video_feature, np.ndarray):
+ video_len = len(video_feature)
+ else:
+ video_len = math.ceil(text_feature["end"][-1])
+ low = math.floor(text_feature["start"][text_clip_indexs[0]])
+ high = math.ceil(text_feature["end"][text_clip_indexs[-1]])
+ if low < high:
+ center = random.randint(low, high)
+ else:
+ center = int((low + high) // 2)
+ center = max(0, min(video_feature.shape[0] - 1, center))
+
+ assert 0 <= center < video_feature.shape[0]
+
+ video_clips = self.video_clip_sampler(
+ video_len, self._get_video_maxlen(), center
+ )
+ video_start = video_clips["start"][0]
+ video_end = video_clips["end"][0]
+
+ # tensorize.
+ vfeats, vmasks = self._build_video_seq(
+ video_feature, video_clips
+ )
+ caps, cmasks = self._build_text_seq(
+ text_feature, text_clip_indexs
+ )
+
+ text_start = text_clip_indexs[0]
+ text_end = text_clip_indexs[-1] + 1
+
+ return {
+ "caps": caps,
+ "cmasks": cmasks,
+ "vfeats": vfeats,
+ "vmasks": vmasks,
+ "video_start": video_start,
+ "video_end": video_end,
+ "text_start": text_start,
+ "text_end": text_end,
+ }
+
+
+class MFMMLMAligner(FixedLenAligner):
+ """
+ `FixedLenAligner` with Masked Language Model and Masked Frame Model.
+ """
+
+ def __init__(self, config):
+ super().__init__(config)
+ keep_prob = config.keep_prob if config.keep_prob is not None else 1.0
+ self.text_clip_sampler = TextClipSamplingProcessor(
+ self.max_len - self.max_video_len - 3, keep_prob
+ )
+ self.sampled_min_len = config.sampled_min_len
+ self.sampled_max_len = config.sampled_max_len
+ self.masked_token_sampler = TextMaskingProcessor(config)
+ self.mm_type = config.mm_type \
+ if config.mm_type is not None else "full"
+ self.attnmasker = MMAttentionMask2DProcessor() \
+ if self.mm_type == "textgen" else None
+ self.masked_frame_sampler = FrameMaskingProcessor(config)
+ self.lazy_vfeat_mask = (
+ False if config.lazy_vfeat_mask is None else config.lazy_vfeat_mask
+ )
+ self.mm_prob = config.mm_prob if config.mm_prob is not None else 0.
+
+ def __call__(self, video_id, video_feature, text_feature):
+ from transformers import default_data_collator
+ if self.subsampling is not None and self.subsampling > 1:
+ batch = []
+ for _ in range(self.subsampling):
+ centerclip_idx = random.randint(
+ 0, len(text_feature["start"]) - 1)
+ sampled_max_text_len = random.randint(
+ self.sampled_min_len, self.sampled_max_len
+ )
+ batch.append(
+ self.sampling(
+ video_id,
+ video_feature,
+ text_feature,
+ centerclip_idx,
+ sampled_max_text_len,
+ )
+ )
+ batch = self.batch_post_processing(batch, video_feature)
+ batch = default_data_collator(batch)
+ else:
+ batch = self.sampling(video_id, video_feature, text_feature)
+ batch = self.batch_post_processing(batch, video_feature)
+ batch["video_id"] = video_id if isinstance(video_id, str) \
+ else video_id[0]
+ return batch
+
+ def sampling(
+ self,
+ video_id,
+ video_feature,
+ text_feature,
+ centerclip_idx=None,
+ sampled_max_text_len=None,
+ ):
+ output = FixedLenAligner.sampling(self,
+ video_id, video_feature, text_feature,
+ centerclip_idx, sampled_max_text_len)
+
+ masking_text, masking_video = None, None
+ if random.random() < self.mm_prob:
+ if random.random() > 0.5:
+ masking_text, masking_video = self.mm_type, "no"
+ else:
+ masking_text, masking_video = "no", "full"
+ video_feats = output["vfeats"] if not self.lazy_vfeat_mask else None
+ video_label = self.masked_frame_sampler(
+ output["vmasks"], masking_video, vfeats=video_feats)
+ caps, text_label = self.masked_token_sampler(
+ output["caps"], masking_text)
+
+ output.update({
+ "caps": caps,
+ "video_label": video_label,
+ "text_label": text_label,
+ })
+
+ if self.attnmasker is not None:
+ attention_mask = self.attnmasker(
+ output["vmasks"], output["cmasks"], masking_text)
+ output.update({
+ "attention_mask": attention_mask
+ })
+ return output
+
+
+class FrameMaskingProcessor(Processor):
+ def __init__(self, config):
+ self.mfm_probability = 0.15
+ if config.mfm_probability is not None:
+ self.mfm_probability = config.mfm_probability
+
+ def __call__(self, vmasks, modality_masking=None, vfeats=None):
+ """
+ We perform lazy masking to save data transfer time.
+ It only generates video_labels by default and MFM model
+ will do actualy masking.
+ Return: `video_label` is a binary mask.
+ """
+ video_label = vmasks.clone()
+ if modality_masking is not None:
+ if modality_masking == "full":
+ probability_matrix = torch.full(video_label.shape, 1.)
+ elif modality_masking == "no":
+ probability_matrix = torch.full(video_label.shape, 0.)
+ elif modality_masking == "inverse":
+ probability_matrix = torch.full(
+ video_label.shape, 1. - self.mfm_probability)
+ else:
+ raise ValueError("unknown modality masking.", modality_masking)
+ else:
+ probability_matrix = torch.full(
+ video_label.shape, self.mfm_probability)
+ masked_indices = torch.bernoulli(probability_matrix).bool()
+ # We only compute loss on masked tokens
+ video_label[~masked_indices] = 0
+ if vfeats is not None:
+ vfeats[video_label, :] = 0.0
+ return video_label
+
+
+class TextGenerationProcessor(Processor):
+ def __init__(self, tokenizer):
+ self.bos_token_id = tokenizer.bos_token_id
+ self.pad_token_id = tokenizer.pad_token_id
+
+ def __call__(self, inputs):
+ labels = inputs.clone()
+ # [CLS] [SEP] for video
+ labels[:2] = -100
+ # keep [SEP] for text.
+ pad_mask = labels == self.pad_token_id
+ labels[pad_mask] = -100
+ inputs[2:] = torch.cat([
+ torch.LongTensor([self.bos_token_id]),
+ inputs[2:-1]])
+ inputs[pad_mask] = self.pad_token_id
+ assert len(inputs) == len(labels)
+ return inputs, labels
+
+
+class TextMaskingProcessor(Processor):
+ def __init__(self, config):
+ """this function is borrowed from
+ `transformers/data/data_collator.DataCollatorForLanguageModeling`"""
+ self.mlm_probability = 0.15
+ if config.mlm_probability is not None:
+ self.mlm_probability = config.mlm_probability
+ self.bert_name = config.bert_name
+ # [CLS] is used as bos_token and [SEP] is used as eos_token.
+ # https://huggingface.co/transformers/master/model_doc/bertgeneration.html
+ from transformers import AutoTokenizer
+ self.tokenizer = AutoTokenizer.from_pretrained(
+ self.bert_name, bos_token="[CLS]", eos_token="[SEP]")
+ self.textgen = TextGenerationProcessor(self.tokenizer)
+
+ def __call__(
+ self, inputs: torch.Tensor,
+ modality_masking=None,
+ special_tokens_mask: Optional[torch.Tensor] = None
+ ) -> Tuple[torch.Tensor, torch.Tensor]:
+ """
+ expand modality_masking into
+ None: traditional bert masking.
+ "no": no masking.
+ "full": all [MASK] token for generation.
+ "gen": autoregressive generation.
+ """
+ """
+ Prepare masked tokens inputs/labels for masked language modeling:
+ 80% MASK, 10% random, 10% original.
+ """
+ labels = inputs.clone()
+ # We sample a few tokens in each sequence for MLM training
+ # (with probability `self.mlm_probability`)
+ if modality_masking is not None:
+ if modality_masking == "full":
+ probability_matrix = torch.full(labels.shape, 1.)
+ elif modality_masking == "no":
+ probability_matrix = torch.full(labels.shape, 0.)
+ elif modality_masking.startswith("textgen"):
+ # [CLS] [SEP] ...
+ inputs, labels = self.textgen(inputs)
+ if "mask" not in modality_masking:
+ return inputs, labels
+ inputs = self.mask_input(inputs, special_tokens_mask)
+ return inputs, labels
+ elif modality_masking == "mask":
+ inputs = self.mask_input(inputs, special_tokens_mask)
+ labels = torch.full(inputs.shape, -100)
+ return inputs, labels
+ elif modality_masking == "inverse":
+ probability_matrix = torch.full(labels.shape, 1. - self.mlm_probability)
+ else:
+ raise ValueError("unknown modality masking.", modality_masking)
+ else:
+ probability_matrix = torch.full(labels.shape, self.mlm_probability)
+
+ if special_tokens_mask is None:
+ special_tokens_mask = self.get_special_tokens_mask(
+ labels.tolist(), already_has_special_tokens=True
+ )
+ special_tokens_mask = torch.tensor(
+ special_tokens_mask, dtype=torch.bool)
+ else:
+ special_tokens_mask = special_tokens_mask.bool()
+
+ probability_matrix.masked_fill_(special_tokens_mask, value=0.0)
+ masked_indices = torch.bernoulli(probability_matrix).bool()
+ labels[~masked_indices] = -100 # We only compute loss on masked tokens
+
+ # 80% of the time,
+ # we replace masked input tokens with tokenizer.mask_token ([MASK])
+ indices_replaced = (
+ torch.bernoulli(
+ torch.full(labels.shape, 0.8)).bool() & masked_indices
+ )
+ inputs[indices_replaced] = self.tokenizer.convert_tokens_to_ids(
+ self.tokenizer.mask_token
+ )
+
+ # 10% of the time, we replace masked input tokens with random word
+ indices_random = (
+ torch.bernoulli(torch.full(labels.shape, 0.5)).bool()
+ & masked_indices
+ & ~indices_replaced
+ )
+ random_words = torch.randint(
+ len(self.tokenizer), labels.shape, dtype=torch.long
+ )
+ inputs[indices_random] = random_words[indices_random]
+
+ # The rest of the time (10% of the time) we keep the masked input
+ # tokens unchanged
+ return inputs, labels
+
+ def mask_input(self, inputs, special_tokens_mask=None):
+ # the following is new with masked autoregressive.
+ probability_matrix = torch.full(
+ inputs.shape, self.mlm_probability)
+ if special_tokens_mask is None:
+ special_tokens_mask = self.get_special_tokens_mask(
+ inputs.tolist(), already_has_special_tokens=True
+ )
+ special_tokens_mask = torch.tensor(
+ special_tokens_mask, dtype=torch.bool)
+ else:
+ special_tokens_mask = special_tokens_mask.bool()
+ probability_matrix.masked_fill_(special_tokens_mask, value=0.0)
+ masked_indices = torch.bernoulli(probability_matrix).bool()
+ indices_replaced = (
+ torch.bernoulli(
+ torch.full(inputs.shape, 0.8)).bool() & masked_indices
+ )
+ inputs[indices_replaced] = self.tokenizer.convert_tokens_to_ids(
+ self.tokenizer.mask_token
+ )
+
+ # 10% of the time, we replace masked input tokens with random word
+ indices_random = (
+ torch.bernoulli(torch.full(inputs.shape, 0.5)).bool()
+ & masked_indices
+ & ~indices_replaced
+ )
+ random_words = torch.randint(
+ len(self.tokenizer), inputs.shape, dtype=torch.long
+ )
+ inputs[indices_random] = random_words[indices_random]
+ return inputs
+
+ def get_special_tokens_mask(
+ self, token_ids_0: List[int],
+ token_ids_1: Optional[List[int]] = None,
+ already_has_special_tokens: bool = False
+ ) -> List[int]:
+ """
+ Note: the version from transformers do not consider pad
+ as special tokens.
+ """
+
+ if already_has_special_tokens:
+ if token_ids_1 is not None:
+ raise ValueError(
+ "You should not supply a second sequence if"
+ "the provided sequence of "
+ "ids is already formated with special tokens "
+ "for the model."
+ )
+ return list(map(lambda x: 1 if x in [
+ self.tokenizer.sep_token_id,
+ self.tokenizer.cls_token_id,
+ self.tokenizer.pad_token_id] else 0, token_ids_0))
+
+ if token_ids_1 is not None:
+ return [1] + ([0] * len(token_ids_0)) + [1] + ([0] * len(token_ids_1)) + [1]
+ return [1] + ([0] * len(token_ids_0)) + [1]
+
+
+class TextClipSamplingProcessor(Processor):
+ def __init__(self, max_text_len, keep_prob=1.0):
+ self.max_text_len = max_text_len
+ self.max_video_len = 256 # always hold.
+ self.keep_prob = keep_prob
+
+ def __call__(
+ self,
+ text_feature,
+ centerclip_idx=None,
+ sampled_max_text_len=None,
+ sampled_max_video_len=None,
+ ):
+ # Let's use all caps for now and see if 256 can cover all of them.
+ if sampled_max_text_len is not None:
+ max_text_len = sampled_max_text_len
+ else:
+ max_text_len = self.max_text_len
+ if sampled_max_video_len is not None:
+ max_video_len = sampled_max_video_len
+ else:
+ max_video_len = self.max_video_len
+
+ t_num_clips = len(text_feature["start"])
+
+ if centerclip_idx is None:
+ centerclip_idx = random.randint(0, t_num_clips - 1)
+
+ start_idx, end_idx = centerclip_idx, centerclip_idx + 1
+ text_clip_indexs = deque()
+ text_clip_indexs.append(start_idx)
+ text_len = len(text_feature["cap"][start_idx])
+
+ video_len = max(
+ 0,
+ text_feature["end"][start_idx]
+ - text_feature["start"][start_idx],
+ )
+
+ while (
+ (start_idx > 0 or end_idx < t_num_clips)
+ and text_len < max_text_len
+ and video_len < max_video_len
+ ):
+ if random.random() > 0.5 and end_idx < t_num_clips:
+ # skip the next one?
+ if random.random() > self.keep_prob and (end_idx + 1) < t_num_clips:
+ end_idx = end_idx + 1
+ text_clip_indexs.append(end_idx)
+ text_len += len(text_feature["cap"][end_idx])
+ end_idx += 1
+ elif start_idx > 0:
+ if random.random() > self.keep_prob and (start_idx - 1) > 0:
+ start_idx = start_idx - 1
+ start_idx -= 1
+ text_clip_indexs.insert(0, start_idx)
+ text_len += len(text_feature["cap"][start_idx])
+ else:
+ if end_idx < t_num_clips:
+ if random.random() > self.keep_prob and (end_idx + 1) < t_num_clips:
+ end_idx = end_idx + 1
+ text_clip_indexs.append(end_idx)
+ text_len += len(text_feature["cap"][end_idx])
+ end_idx += 1
+ else:
+ return text_clip_indexs
+ video_len = max(
+ 0,
+ text_feature["end"][text_clip_indexs[-1]]
+ - text_feature["start"][text_clip_indexs[0]],
+ )
+ return text_clip_indexs
+
+
+class VideoClipSamplingProcessor(Processor):
+ def __call__(self, video_len, max_video_len, center):
+ """
+ `video_len`: length of the video.
+ `max_video_len`: maximum video tokens allowd in a sequence.
+ `center`: initial starting index.
+ """
+ assert center >= 0 and center < video_len
+ t_clip_len = 0
+ start, end = center, center
+ while (start > 0 or end < video_len) and t_clip_len < max_video_len:
+ # decide the direction to grow.
+ if start <= 0:
+ end += 1
+ elif end >= video_len:
+ start -= 1
+ elif random.random() > 0.5:
+ end += 1
+ else:
+ start -= 1
+ t_clip_len += 1
+ return {"start": [start], "end": [end]}
+
+
+class How2MILNCEAligner(FixedLenAligner):
+ """reference: `antoine77340/MIL-NCE_HowTo100M/video_loader.py`"""
+
+ def __init__(self, config):
+ super().__init__(config)
+ self.num_candidates = 4
+ self.min_time = 5.0
+ self.num_sec = 3.2
+ # self.num_sec = self.num_frames / float(self.fps) num_frames=16 / fps = 5
+ # self.num_frames = 16
+
+ def sampling(
+ self,
+ video_id,
+ video_feature,
+ text_feature,
+ centerclip_idx=None, # will be ignored.
+ sampled_max_text_len=None # will be ignored.
+ ):
+ text, start, end = self._get_text(text_feature)
+ video = self._get_video(video_feature, start, end)
+
+ vfeats = torch.zeros((self.max_video_len, video_feature.shape[1]))
+ vmasks = torch.zeros((self.max_video_len,), dtype=torch.bool)
+ vfeats[: video.shape[0]] = torch.from_numpy(np.array(video))
+ vmasks[: video.shape[0]] = 1
+
+ caps, cmasks = [], []
+ for words in text:
+ cap, cmask = self._build_text_seq(text_feature, words)
+ caps.append(cap)
+ cmasks.append(cmask)
+ caps = torch.stack(caps)
+ cmasks = torch.stack(cmasks)
+ # video of shape: (video_len)
+ # text of shape (num_candidates, max_text_len)
+
+ return {
+ "caps": caps,
+ "cmasks": cmasks,
+ "vfeats": vfeats,
+ "vmasks": vmasks,
+ # "video_id": video_id,
+ }
+
+ def _get_video(self, video_feature, start, end):
+ start_seek = random.randint(start, int(max(start, end - self.num_sec)))
+ # duration = self.num_sec + 0.1
+ return video_feature[start_seek : int(start_seek + self.num_sec)]
+
+ def _get_text(self, cap):
+ ind = random.randint(0, len(cap["start"]) - 1)
+ if self.num_candidates == 1:
+ words = [ind]
+ else:
+ words = []
+ cap_start = self._find_nearest_candidates(cap, ind)
+ for i in range(self.num_candidates):
+ words.append([max(0, min(len(cap["cap"]) - 1, cap_start + i))])
+
+ start, end = cap["start"][ind], cap["end"][ind]
+ # TODO: May need to be improved for edge cases.
+ # expand the min time.
+ if end - start < self.min_time:
+ diff = self.min_time - end + start
+ start = max(0, start - diff / 2)
+ end = start + self.min_time
+ return words, int(start), int(end)
+
+ def _find_nearest_candidates(self, caption, ind):
+ """find the range of the clips."""
+ start, end = ind, ind
+ #diff = caption["end"][end] - caption["start"][start]
+ n_candidate = 1
+ while n_candidate < self.num_candidates:
+ # the first clip
+ if start == 0:
+ return 0
+ # we add () in the following condition to fix the bug.
+ elif end == (len(caption["start"]) - 1):
+ return start - (self.num_candidates - n_candidate)
+ elif (caption["end"][end] - caption["start"][start - 1]) < (
+ caption["end"][end + 1] - caption["start"][start]
+ ):
+ start -= 1
+ else:
+ end += 1
+ n_candidate += 1
+ return start
+
+
+class PKLJSONStrTextProcessor(TextProcessor):
+ """`caption.json` from howto100m are preprocessed as a
+ dict `[video_id, json_str]`.
+ Json parsing tokenization are conducted on-the-fly and cached into dict.
+ """
+
+ def __init__(self, config, max_clip_text_len=96):
+ print("[Warning] PKLJSONStrTextProcessor is slow for num_workers > 0.")
+ self.caption_pkl_path = str(config.caption_pkl_path)
+ with open(self.caption_pkl_path, "rb") as fd:
+ self.data = pickle.load(fd)
+ self.max_clip_text_len = max_clip_text_len
+ from transformers import AutoTokenizer
+ self.tokenizer = AutoTokenizer.from_pretrained(
+ str(config.bert_name), use_fast=config.use_fast
+ )
+
+ def __call__(self, video_id):
+ caption = self.data[video_id]
+ if isinstance(caption, str):
+ import json
+ caption = json.loads(caption)
+ cap = []
+ for clip_idx, text_clip in enumerate(caption["text"]):
+ clip_ids = []
+ if isinstance(text_clip, str):
+ clip_ids = self.tokenizer(
+ text_clip[: self.max_clip_text_len],
+ add_special_tokens=False
+ )["input_ids"]
+ cap.append(clip_ids)
+ caption["cap"] = cap
+ caption.pop("text") # save space.
+ self.data[video_id] = caption
+ return caption
diff --git a/fairseq/examples/MMPT/mmpt/processors/how2retriprocessor.py b/fairseq/examples/MMPT/mmpt/processors/how2retriprocessor.py
new file mode 100644
index 0000000000000000000000000000000000000000..b5a7730ec0bbe91d9997564214fffb10d0aef519
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/processors/how2retriprocessor.py
@@ -0,0 +1,100 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+from .how2processor import (
+ ShardedHow2MetaProcessor,
+ ShardedVideoProcessor,
+ ShardedTextProcessor,
+ VariedLenAligner,
+ OverlappedAligner
+)
+
+
+class ShardedHow2VideoRetriMetaProcessor(ShardedHow2MetaProcessor):
+ def __init__(self, config):
+ super().__init__(config)
+ self.num_video_per_batch = config.num_video_per_batch
+ self.cands = [
+ self.data[batch_offset:batch_offset + self.num_video_per_batch]
+ for batch_offset in
+ range(0, (len(self.data) // (8 * self.num_video_per_batch)) * 8 * self.num_video_per_batch, self.num_video_per_batch)]
+
+ def __len__(self):
+ return len(self.cands)
+
+ def set_candidates(self, cands):
+ # no changes on num of batches.
+ print(len(self.cands), "->", len(cands))
+ # assert len(self.cands) == len(cands)
+ self.cands = cands
+
+ def __getitem__(self, idx):
+ video_ids = self.cands[idx]
+ assert isinstance(video_ids, list)
+ sharded_video_idxs = []
+ for video_id in video_ids:
+ shard_id, video_idx = self.video_id_to_shard[video_id]
+ sharded_video_idxs.append((video_id, -1, shard_id, video_idx))
+ return sharded_video_idxs, sharded_video_idxs
+
+
+class ShardedVideoRetriVideoProcessor(ShardedVideoProcessor):
+ """In retrival case the video_id
+ is a list of tuples: `(shard_id, video_idx)` ."""
+
+ def __call__(self, sharded_video_idxs):
+ assert isinstance(sharded_video_idxs, list)
+ cand_feats = []
+ for shared_video_idx in sharded_video_idxs:
+ feat = super().__call__(shared_video_idx)
+ cand_feats.append(feat)
+ return cand_feats
+
+
+class ShardedVideoRetriTextProcessor(ShardedTextProcessor):
+ """In retrival case the video_id
+ is a list of tuples: `(shard_id, video_idx)` ."""
+
+ def __call__(self, sharded_video_idxs):
+ assert isinstance(sharded_video_idxs, list)
+ cand_caps = []
+ for shared_video_idx in sharded_video_idxs:
+ caps = super().__call__(shared_video_idx)
+ cand_caps.append(caps)
+ return cand_caps
+
+
+class VideoRetriAligner(VariedLenAligner):
+ # Retritask will trim dim-0.
+ def __call__(self, sharded_video_idxs, video_features, text_features):
+ from transformers import default_data_collator
+ batch, video_ids = [], []
+ for video_id, video_feature, text_feature in \
+ zip(sharded_video_idxs, video_features, text_features):
+ sub_batch = super().__call__(video_id, video_feature, text_feature)
+ batch.append(sub_batch)
+ if isinstance(video_id, tuple):
+ video_id = video_id[0]
+ video_ids.append(video_id)
+ batch = default_data_collator(batch)
+ batch["video_id"] = video_ids
+ return batch
+
+
+class VideoRetriOverlappedAligner(OverlappedAligner):
+ # Retritask will trim dim-0.
+ def __call__(self, sharded_video_idxs, video_features, text_features):
+ from transformers import default_data_collator
+ batch, video_ids = [], []
+ for video_id, video_feature, text_feature in \
+ zip(sharded_video_idxs, video_features, text_features):
+ sub_batch = super().__call__(video_id, video_feature, text_feature)
+ batch.append(sub_batch)
+ if isinstance(video_id, tuple):
+ video_id = video_id[0]
+ video_ids.append(video_id)
+ batch = default_data_collator(batch)
+ batch["video_id"] = video_ids
+ return batch
diff --git a/fairseq/examples/MMPT/mmpt/processors/models/s3dg.py b/fairseq/examples/MMPT/mmpt/processors/models/s3dg.py
new file mode 100644
index 0000000000000000000000000000000000000000..6c7a691e33551807ef258a54282884a409408691
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/processors/models/s3dg.py
@@ -0,0 +1,336 @@
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""Contains a PyTorch definition for Gated Separable 3D network (S3D-G)
+with a text module for computing joint text-video embedding from raw text
+and video input. The following code will enable you to load the HowTo100M
+pretrained S3D Text-Video model from:
+ A. Miech, J.-B. Alayrac, L. Smaira, I. Laptev, J. Sivic and A. Zisserman,
+ End-to-End Learning of Visual Representations from Uncurated Instructional Videos.
+ https://arxiv.org/abs/1912.06430.
+
+S3D-G was proposed by:
+ S. Xie, C. Sun, J. Huang, Z. Tu and K. Murphy,
+ Rethinking Spatiotemporal Feature Learning For Video Understanding.
+ https://arxiv.org/abs/1712.04851.
+ Tensorflow code: https://github.com/tensorflow/models/blob/master/research/slim/nets/s3dg.py
+
+The S3D architecture was slightly modified with a space to depth trick for TPU
+optimization.
+"""
+
+import torch as th
+import torch.nn.functional as F
+import torch.nn as nn
+import os
+import numpy as np
+import re
+
+
+class InceptionBlock(nn.Module):
+ def __init__(
+ self,
+ input_dim,
+ num_outputs_0_0a,
+ num_outputs_1_0a,
+ num_outputs_1_0b,
+ num_outputs_2_0a,
+ num_outputs_2_0b,
+ num_outputs_3_0b,
+ gating=True,
+ ):
+ super(InceptionBlock, self).__init__()
+ self.conv_b0 = STConv3D(input_dim, num_outputs_0_0a, [1, 1, 1])
+ self.conv_b1_a = STConv3D(input_dim, num_outputs_1_0a, [1, 1, 1])
+ self.conv_b1_b = STConv3D(
+ num_outputs_1_0a, num_outputs_1_0b, [3, 3, 3], padding=1, separable=True
+ )
+ self.conv_b2_a = STConv3D(input_dim, num_outputs_2_0a, [1, 1, 1])
+ self.conv_b2_b = STConv3D(
+ num_outputs_2_0a, num_outputs_2_0b, [3, 3, 3], padding=1, separable=True
+ )
+ self.maxpool_b3 = th.nn.MaxPool3d((3, 3, 3), stride=1, padding=1)
+ self.conv_b3_b = STConv3D(input_dim, num_outputs_3_0b, [1, 1, 1])
+ self.gating = gating
+ self.output_dim = (
+ num_outputs_0_0a + num_outputs_1_0b + num_outputs_2_0b + num_outputs_3_0b
+ )
+ if gating:
+ self.gating_b0 = SelfGating(num_outputs_0_0a)
+ self.gating_b1 = SelfGating(num_outputs_1_0b)
+ self.gating_b2 = SelfGating(num_outputs_2_0b)
+ self.gating_b3 = SelfGating(num_outputs_3_0b)
+
+ def forward(self, input):
+ """Inception block
+ """
+ b0 = self.conv_b0(input)
+ b1 = self.conv_b1_a(input)
+ b1 = self.conv_b1_b(b1)
+ b2 = self.conv_b2_a(input)
+ b2 = self.conv_b2_b(b2)
+ b3 = self.maxpool_b3(input)
+ b3 = self.conv_b3_b(b3)
+ if self.gating:
+ b0 = self.gating_b0(b0)
+ b1 = self.gating_b1(b1)
+ b2 = self.gating_b2(b2)
+ b3 = self.gating_b3(b3)
+ return th.cat((b0, b1, b2, b3), dim=1)
+
+
+class SelfGating(nn.Module):
+ def __init__(self, input_dim):
+ super(SelfGating, self).__init__()
+ self.fc = nn.Linear(input_dim, input_dim)
+
+ def forward(self, input_tensor):
+ """Feature gating as used in S3D-G.
+ """
+ spatiotemporal_average = th.mean(input_tensor, dim=[2, 3, 4])
+ weights = self.fc(spatiotemporal_average)
+ weights = th.sigmoid(weights)
+ return weights[:, :, None, None, None] * input_tensor
+
+
+class STConv3D(nn.Module):
+ def __init__(
+ self, input_dim, output_dim, kernel_size, stride=1, padding=0, separable=False
+ ):
+ super(STConv3D, self).__init__()
+ self.separable = separable
+ self.relu = nn.ReLU(inplace=True)
+ assert len(kernel_size) == 3
+ if separable and kernel_size[0] != 1:
+ spatial_kernel_size = [1, kernel_size[1], kernel_size[2]]
+ temporal_kernel_size = [kernel_size[0], 1, 1]
+ if isinstance(stride, list) and len(stride) == 3:
+ spatial_stride = [1, stride[1], stride[2]]
+ temporal_stride = [stride[0], 1, 1]
+ else:
+ spatial_stride = [1, stride, stride]
+ temporal_stride = [stride, 1, 1]
+ if isinstance(padding, list) and len(padding) == 3:
+ spatial_padding = [0, padding[1], padding[2]]
+ temporal_padding = [padding[0], 0, 0]
+ else:
+ spatial_padding = [0, padding, padding]
+ temporal_padding = [padding, 0, 0]
+ if separable:
+ self.conv1 = nn.Conv3d(
+ input_dim,
+ output_dim,
+ kernel_size=spatial_kernel_size,
+ stride=spatial_stride,
+ padding=spatial_padding,
+ bias=False,
+ )
+ self.bn1 = nn.BatchNorm3d(output_dim)
+ self.conv2 = nn.Conv3d(
+ output_dim,
+ output_dim,
+ kernel_size=temporal_kernel_size,
+ stride=temporal_stride,
+ padding=temporal_padding,
+ bias=False,
+ )
+ self.bn2 = nn.BatchNorm3d(output_dim)
+ else:
+ self.conv1 = nn.Conv3d(
+ input_dim,
+ output_dim,
+ kernel_size=kernel_size,
+ stride=stride,
+ padding=padding,
+ bias=False,
+ )
+ self.bn1 = nn.BatchNorm3d(output_dim)
+
+ def forward(self, input):
+ out = self.relu(self.bn1(self.conv1(input)))
+ if self.separable:
+ out = self.relu(self.bn2(self.conv2(out)))
+ return out
+
+
+class MaxPool3dTFPadding(th.nn.Module):
+ def __init__(self, kernel_size, stride=None, padding="SAME"):
+ super(MaxPool3dTFPadding, self).__init__()
+ if padding == "SAME":
+ padding_shape = self._get_padding_shape(kernel_size, stride)
+ self.padding_shape = padding_shape
+ self.pad = th.nn.ConstantPad3d(padding_shape, 0)
+ self.pool = th.nn.MaxPool3d(kernel_size, stride, ceil_mode=True)
+
+ def _get_padding_shape(self, filter_shape, stride):
+ def _pad_top_bottom(filter_dim, stride_val):
+ pad_along = max(filter_dim - stride_val, 0)
+ pad_top = pad_along // 2
+ pad_bottom = pad_along - pad_top
+ return pad_top, pad_bottom
+
+ padding_shape = []
+ for filter_dim, stride_val in zip(filter_shape, stride):
+ pad_top, pad_bottom = _pad_top_bottom(filter_dim, stride_val)
+ padding_shape.append(pad_top)
+ padding_shape.append(pad_bottom)
+ depth_top = padding_shape.pop(0)
+ depth_bottom = padding_shape.pop(0)
+ padding_shape.append(depth_top)
+ padding_shape.append(depth_bottom)
+ return tuple(padding_shape)
+
+ def forward(self, inp):
+ inp = self.pad(inp)
+ out = self.pool(inp)
+ return out
+
+
+class Sentence_Embedding(nn.Module):
+ def __init__(
+ self,
+ embd_dim,
+ num_embeddings=66250,
+ word_embedding_dim=300,
+ token_to_word_path="dict.npy",
+ max_words=16,
+ output_dim=2048,
+ ):
+ super(Sentence_Embedding, self).__init__()
+ self.word_embd = nn.Embedding(num_embeddings, word_embedding_dim)
+ self.fc1 = nn.Linear(word_embedding_dim, output_dim)
+ self.fc2 = nn.Linear(output_dim, embd_dim)
+ self.word_to_token = {}
+ self.max_words = max_words
+ token_to_word = np.load(token_to_word_path)
+ for i, t in enumerate(token_to_word):
+ self.word_to_token[t] = i + 1
+
+ def _zero_pad_tensor_token(self, tensor, size):
+ if len(tensor) >= size:
+ return tensor[:size]
+ else:
+ zero = th.zeros(size - len(tensor)).long()
+ return th.cat((tensor, zero), dim=0)
+
+ def _split_text(self, sentence):
+ w = re.findall(r"[\w']+", str(sentence))
+ return w
+
+ def _words_to_token(self, words):
+ words = [
+ self.word_to_token[word] for word in words if word in self.word_to_token
+ ]
+ if words:
+ we = self._zero_pad_tensor_token(th.LongTensor(words), self.max_words)
+ return we
+ else:
+ return th.zeros(self.max_words).long()
+
+ def _words_to_ids(self, x):
+ split_x = [self._words_to_token(self._split_text(sent.lower())) for sent in x]
+ return th.stack(split_x, dim=0)
+
+ def forward(self, x):
+ x = self._words_to_ids(x)
+ x = self.word_embd(x)
+ x = F.relu(self.fc1(x))
+ x = th.max(x, dim=1)[0]
+ x = self.fc2(x)
+ return {'text_embedding': x}
+
+
+class S3D(nn.Module):
+ def __init__(self, dict_path, num_classes=512, gating=True, space_to_depth=True):
+ super(S3D, self).__init__()
+ self.num_classes = num_classes
+ self.gating = gating
+ self.space_to_depth = space_to_depth
+ if space_to_depth:
+ self.conv1 = STConv3D(
+ 24, 64, [2, 4, 4], stride=1, padding=(1, 2, 2), separable=False
+ )
+ else:
+ self.conv1 = STConv3D(
+ 3, 64, [3, 7, 7], stride=2, padding=(1, 3, 3), separable=False
+ )
+ self.conv_2b = STConv3D(64, 64, [1, 1, 1], separable=False)
+ self.conv_2c = STConv3D(64, 192, [3, 3, 3], padding=1, separable=True)
+ self.gating = SelfGating(192)
+ self.maxpool_2a = MaxPool3dTFPadding(
+ kernel_size=(1, 3, 3), stride=(1, 2, 2), padding="SAME"
+ )
+ self.maxpool_3a = MaxPool3dTFPadding(
+ kernel_size=(1, 3, 3), stride=(1, 2, 2), padding="SAME"
+ )
+ self.mixed_3b = InceptionBlock(192, 64, 96, 128, 16, 32, 32)
+ self.mixed_3c = InceptionBlock(
+ self.mixed_3b.output_dim, 128, 128, 192, 32, 96, 64
+ )
+ self.maxpool_4a = MaxPool3dTFPadding(
+ kernel_size=(3, 3, 3), stride=(2, 2, 2), padding="SAME"
+ )
+ self.mixed_4b = InceptionBlock(
+ self.mixed_3c.output_dim, 192, 96, 208, 16, 48, 64
+ )
+ self.mixed_4c = InceptionBlock(
+ self.mixed_4b.output_dim, 160, 112, 224, 24, 64, 64
+ )
+ self.mixed_4d = InceptionBlock(
+ self.mixed_4c.output_dim, 128, 128, 256, 24, 64, 64
+ )
+ self.mixed_4e = InceptionBlock(
+ self.mixed_4d.output_dim, 112, 144, 288, 32, 64, 64
+ )
+ self.mixed_4f = InceptionBlock(
+ self.mixed_4e.output_dim, 256, 160, 320, 32, 128, 128
+ )
+ self.maxpool_5a = self.maxPool3d_5a_2x2 = MaxPool3dTFPadding(
+ kernel_size=(2, 2, 2), stride=(2, 2, 2), padding="SAME"
+ )
+ self.mixed_5b = InceptionBlock(
+ self.mixed_4f.output_dim, 256, 160, 320, 32, 128, 128
+ )
+ self.mixed_5c = InceptionBlock(
+ self.mixed_5b.output_dim, 384, 192, 384, 48, 128, 128
+ )
+ self.fc = nn.Linear(self.mixed_5c.output_dim, num_classes)
+ self.text_module = Sentence_Embedding(num_classes,
+ token_to_word_path=dict_path)
+
+ def _space_to_depth(self, input):
+ """3D space to depth trick for TPU optimization.
+ """
+ B, C, T, H, W = input.shape
+ input = input.view(B, C, T // 2, 2, H // 2, 2, W // 2, 2)
+ input = input.permute(0, 3, 5, 7, 1, 2, 4, 6)
+ input = input.contiguous().view(B, 8 * C, T // 2, H // 2, W // 2)
+ return input
+
+ def forward(self, inputs):
+ """Defines the S3DG base architecture."""
+ if self.space_to_depth:
+ inputs = self._space_to_depth(inputs)
+ net = self.conv1(inputs)
+ if self.space_to_depth:
+ # we need to replicate 'SAME' tensorflow padding
+ net = net[:, :, 1:, 1:, 1:]
+ net = self.maxpool_2a(net)
+ net = self.conv_2b(net)
+ net = self.conv_2c(net)
+ if self.gating:
+ net = self.gating(net)
+ net = self.maxpool_3a(net)
+ net = self.mixed_3b(net)
+ net = self.mixed_3c(net)
+ net = self.maxpool_4a(net)
+ net = self.mixed_4b(net)
+ net = self.mixed_4c(net)
+ net = self.mixed_4d(net)
+ net = self.mixed_4e(net)
+ net = self.mixed_4f(net)
+ net = self.maxpool_5a(net)
+ net = self.mixed_5b(net)
+ net = self.mixed_5c(net)
+ net = th.mean(net, dim=[2, 3, 4])
+ return {'video_embedding': self.fc(net), 'mixed_5c': net}
diff --git a/fairseq/examples/MMPT/mmpt/processors/processor.py b/fairseq/examples/MMPT/mmpt/processors/processor.py
new file mode 100644
index 0000000000000000000000000000000000000000..98edb051f16efef81fba98b0b2f6befbad09f2d4
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/processors/processor.py
@@ -0,0 +1,274 @@
+# Copyright (c) Facebook, Inc. All Rights Reserved
+
+import numpy as np
+import os
+import torch
+
+
+class Processor(object):
+ """
+ A generic processor for video (codec, feature etc.) and text.
+ """
+
+ def __call__(self, **kwargs):
+ raise NotImplementedError
+
+
+class MetaProcessor(Processor):
+ """
+ A meta processor is expected to load the metadata of a dataset:
+ (e.g., video_ids, or captions).
+ You must implement the `__getitem__` (meta datasets are rather diverse.).
+ """
+
+ def __init__(self, config):
+ self.split = config.split
+
+ def __len__(self):
+ return len(self.data)
+
+ def __getitem__(self, idx):
+ raise NotImplementedError
+
+ def _get_split_path(self, config):
+ splits = {
+ "train": config.train_path,
+ "valid": config.val_path,
+ "test": config.test_path,
+ }
+ if config.split is not None:
+ return splits[config.split]
+ return config.train_path
+
+
+class TextProcessor(Processor):
+ """
+ A generic Text processor: rename this as `withTokenizer`.
+ tokenize a string of text on-the-fly.
+ Warning: mostly used for end tasks.
+ (on-the-fly tokenization is slow for how2.)
+ TODO(huxu): move this class as a subclass.
+ """
+
+ def __init__(self, config):
+ self.bert_name = str(config.bert_name)
+ self.use_fast = config.use_fast
+ from transformers import AutoTokenizer
+ self.tokenizer = AutoTokenizer.from_pretrained(
+ self.bert_name, use_fast=self.use_fast
+ )
+
+ def __call__(self, text_id):
+ caption = self.tokenizer(text_id, add_special_tokens=False)
+ return caption["input_ids"]
+
+
+class VideoProcessor(Processor):
+ """
+ A generic video processor: load a numpy video tokens by default.
+ """
+
+ def __init__(self, config):
+ self.vfeat_dir = config.vfeat_dir
+
+ def __call__(self, video_fn):
+ if isinstance(video_fn, tuple):
+ video_fn = video_fn[0]
+ assert isinstance(video_fn, str)
+ video_fn = os.path.join(self.vfeat_dir, video_fn + ".npy")
+ feat = np.load(video_fn)
+ return feat
+
+
+class Aligner(object):
+ """
+ An alignprocessor align video and text and output a dict of tensors (for a model).
+ """
+ def __init__(self, config):
+ """__init__ needs to be light weight for more workers/threads."""
+ self.split = config.split
+ self.max_video_len = config.max_video_len
+ self.max_len = config.max_len
+ from transformers import AutoTokenizer
+ tokenizer = AutoTokenizer.from_pretrained(
+ str(config.bert_name), use_fast=config.use_fast
+ )
+ self.cls_token_id = tokenizer.cls_token_id
+ self.sep_token_id = tokenizer.sep_token_id
+ self.pad_token_id = tokenizer.pad_token_id
+ self.mask_token_id = tokenizer.mask_token_id
+
+ def __call__(self, video_id, video_feature, text_feature):
+ raise NotImplementedError
+
+ def _build_video_seq(self, video_feature, video_clips=None):
+ """
+ `video_feature`: available video tokens.
+ `video_clips`: video clip sequence to build.
+ """
+ if not isinstance(video_feature, np.ndarray):
+ raise ValueError(
+ "unsupported type of video_feature", type(video_feature)
+ )
+
+ if video_clips is None:
+ # this is borrowed from DSAligner
+ video_start = 0
+ video_end = min(len(video_feature), self.max_video_len)
+ # the whole sequence is a single clip.
+ video_clips = {"start": [video_start], "end": [video_end]}
+
+ vfeats = np.zeros(
+ (self.max_video_len, video_feature.shape[1]), dtype=np.float32
+ )
+ vmasks = torch.zeros((self.max_video_len,), dtype=torch.bool)
+ video_len = 0
+ for start, end in zip(video_clips["start"], video_clips["end"]):
+ clip_len = min(self.max_video_len - video_len, (end - start))
+ if clip_len > 0:
+ vfeats[video_len: video_len + clip_len] = video_feature[
+ start: start + clip_len
+ ]
+ vmasks[video_len: video_len + clip_len] = 1
+ video_len += clip_len
+ vfeats = torch.from_numpy(vfeats)
+
+ return vfeats, vmasks
+
+ def _build_text_seq(self, text_feature, text_clip_indexs=None):
+ """
+ `text_feature`: all available clips.
+ `text_clip_indexes`: clip sequence to build.
+ """
+ if text_clip_indexs is None:
+ text_clip_indexs = [0]
+
+ full_caps = []
+ if isinstance(text_feature, dict):
+ for clip_idx in text_clip_indexs:
+ full_caps.extend(text_feature["cap"][clip_idx])
+ else:
+ full_caps = text_feature
+ max_text_len = self.max_len - self.max_video_len - 3
+ full_caps = full_caps[:max_text_len]
+ full_caps = (
+ [self.cls_token_id, self.sep_token_id] + full_caps + [self.sep_token_id]
+ )
+ text_pad_len = self.max_len - len(full_caps) - self.max_video_len
+ padded_full_caps = full_caps + [self.pad_token_id] * text_pad_len
+ caps = torch.LongTensor(padded_full_caps)
+ cmasks = torch.zeros((len(padded_full_caps),), dtype=torch.bool)
+ cmasks[: len(full_caps)] = 1
+
+ return caps, cmasks
+
+ def batch_post_processing(self, batch, video_feature):
+ return batch
+
+
+class MMAttentionMask2DProcessor(Processor):
+ """text generation requires 2d mask
+ that is harder to generate by GPU at this stage."""
+
+ def __call__(self, vmask, cmask, mtype):
+ if mtype == "textgen":
+ return self._build_textgeneration_mask(vmask, cmask)
+ elif mtype == "videogen":
+ return self._build_videogeneration_mask(vmask, cmask)
+ else:
+ return self._build_mm_mask(vmask, cmask)
+
+ def _build_mm_mask(self, vmask, cmask):
+ mask_1d = torch.cat([cmask[:1], vmask, cmask[1:]], dim=0)
+ return mask_1d[None, :].repeat(mask_1d.size(0), 1)
+
+ def _build_videogeneration_mask(self, vmask, cmask):
+ # cls_mask is only about text otherwise it will leak generation.
+ cls_text_mask = torch.cat([
+ # [CLS]
+ torch.ones(
+ (1,), dtype=torch.bool, device=cmask.device),
+ # video tokens and [SEP] for video.
+ torch.zeros(
+ (vmask.size(0) + 1,), dtype=torch.bool, device=cmask.device),
+ cmask[2:]
+ ], dim=0)
+
+ # concat horizontially.
+ video_len = int(vmask.sum())
+ video_masks = torch.cat([
+ # [CLS]
+ torch.ones(
+ (video_len, 1), dtype=torch.bool, device=cmask.device
+ ),
+ torch.tril(
+ torch.ones(
+ (video_len, video_len),
+ dtype=torch.bool, device=cmask.device)),
+ # video_padding
+ torch.zeros(
+ (video_len, vmask.size(0) - video_len),
+ dtype=torch.bool, device=cmask.device
+ ),
+ # [SEP] for video (unused).
+ torch.zeros(
+ (video_len, 1), dtype=torch.bool, device=cmask.device
+ ),
+ cmask[2:].unsqueeze(0).repeat(video_len, 1)
+ ], dim=1)
+
+ text_masks = cls_text_mask[None, :].repeat(
+ cmask.size(0) - 2, 1)
+ video_padding_masks = cls_text_mask[None, :].repeat(
+ vmask.size(0) - video_len, 1)
+
+ return torch.cat([
+ cls_text_mask[None, :],
+ video_masks,
+ video_padding_masks,
+ torch.cat([cmask[:1], vmask, cmask[1:]], dim=0)[None,:],
+ text_masks
+ ], dim=0)
+
+ def _build_textgeneration_mask(self, vmask, cmask):
+ # cls_mask is only about video otherwise it will leak generation.
+ cls_video_mask = torch.cat([
+ # [CLS]
+ torch.ones(
+ (1,), dtype=torch.bool, device=cmask.device),
+ vmask,
+ # [SEP]
+ torch.ones((1,), dtype=torch.bool, device=cmask.device),
+ torch.zeros(
+ (cmask.size(0)-2,), dtype=torch.bool, device=cmask.device)
+ ], dim=0)
+
+ # concat horizontially.
+ text_len = int(cmask[2:].sum())
+ text_masks = torch.cat([
+ # [CLS]
+ torch.ones(
+ (text_len, 1), dtype=torch.bool, device=cmask.device
+ ),
+ vmask.unsqueeze(0).repeat(text_len, 1),
+ # [SEP] for video.
+ torch.ones(
+ (text_len, 1), dtype=torch.bool, device=cmask.device
+ ),
+ torch.tril(
+ torch.ones(
+ (text_len, text_len),
+ dtype=torch.bool, device=cmask.device)),
+ # padding.
+ torch.zeros(
+ (text_len, cmask.size(0) - text_len - 2),
+ dtype=torch.bool, device=cmask.device
+ )
+ ], dim=1)
+
+ cls_video_masks = cls_video_mask[None, :].repeat(
+ vmask.size(0) + 2, 1)
+ text_padding_masks = cls_video_mask[None, :].repeat(
+ cmask.size(0) - text_len - 2, 1)
+ return torch.cat([
+ cls_video_masks, text_masks, text_padding_masks], dim=0)
diff --git a/fairseq/examples/MMPT/mmpt/tasks/__init__.py b/fairseq/examples/MMPT/mmpt/tasks/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..e2e9323a530672ef9daecd793ef645a3c1d0f3e6
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/tasks/__init__.py
@@ -0,0 +1,22 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+from .task import *
+from .vlmtask import *
+from .retritask import *
+
+try:
+ from .fairseqmmtask import *
+except ImportError:
+ pass
+
+try:
+ from .milncetask import *
+except ImportError:
+ pass
+
+try:
+ from .expretritask import *
+except ImportError:
+ pass
diff --git a/fairseq/examples/MMPT/mmpt/tasks/fairseqmmtask.py b/fairseq/examples/MMPT/mmpt/tasks/fairseqmmtask.py
new file mode 100644
index 0000000000000000000000000000000000000000..f6b6115a39b2342be5513edd53016187ab91eb01
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/tasks/fairseqmmtask.py
@@ -0,0 +1,104 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+"""
+make a general fairseq task for MM pretraining.
+"""
+
+import random
+
+from fairseq.tasks import LegacyFairseqTask, register_task
+
+from .task import Task
+from .retritask import RetriTask
+from ..datasets import FairseqMMDataset
+from .. import utils
+
+
+@register_task("mmtask")
+class FairseqMMTask(LegacyFairseqTask):
+ @staticmethod
+ def add_args(parser):
+ # Add some command-line arguments for specifying where the data is
+ # located and the maximum supported input length.
+ parser.add_argument(
+ "taskconfig",
+ metavar="FILE",
+ help=("taskconfig to load all configurations" "outside fairseq parser."),
+ )
+
+ @classmethod
+ def setup_task(cls, args, **kwargs):
+ return FairseqMMTask(args)
+
+ def __init__(self, args):
+ super().__init__(args)
+ config = utils.load_config(args)
+ self.mmtask = Task.config_task(config)
+ self.mmtask.build_dataset()
+ self.mmtask.build_model()
+ self.mmtask.build_loss()
+
+ def load_dataset(self, split, **kwargs):
+ split_map = {
+ "train": self.mmtask.train_data,
+ "valid": self.mmtask.val_data,
+ "test": self.mmtask.test_data,
+ }
+ if split not in split_map:
+ raise ValueError("unknown split type.")
+ if split_map[split] is not None:
+ self.datasets[split] = FairseqMMDataset(split_map[split])
+
+ def get_batch_iterator(
+ self,
+ dataset,
+ max_tokens=None,
+ max_sentences=None,
+ max_positions=None,
+ ignore_invalid_inputs=False,
+ required_batch_size_multiple=1,
+ seed=1,
+ num_shards=1,
+ shard_id=0,
+ num_workers=0,
+ epoch=1,
+ data_buffer_size=0,
+ disable_iterator_cache=False,
+ skip_remainder_batch=False,
+ grouped_shuffling=False,
+ update_epoch_batch_itr=False,
+ ):
+ random.seed(epoch)
+ if dataset.mmdataset.split == "train" and isinstance(self.mmtask, RetriTask):
+ if epoch >= self.mmtask.config.retri_epoch:
+ if not hasattr(self.mmtask, "retri_dataloader"):
+ self.mmtask.build_dataloader()
+ self.mmtask.retrive_candidates(epoch)
+
+ return super().get_batch_iterator(
+ dataset,
+ max_tokens,
+ max_sentences,
+ max_positions,
+ ignore_invalid_inputs,
+ required_batch_size_multiple,
+ seed,
+ num_shards,
+ shard_id,
+ num_workers,
+ epoch,
+ data_buffer_size,
+ disable_iterator_cache,
+ grouped_shuffling,
+ update_epoch_batch_itr,
+ )
+
+ @property
+ def source_dictionary(self):
+ return None
+
+ @property
+ def target_dictionary(self):
+ return None
diff --git a/fairseq/examples/MMPT/mmpt/tasks/milncetask.py b/fairseq/examples/MMPT/mmpt/tasks/milncetask.py
new file mode 100644
index 0000000000000000000000000000000000000000..61b6ab0597f9f3a78bbcf1474613630b20c5a874
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/tasks/milncetask.py
@@ -0,0 +1,27 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+
+from .task import Task
+
+
+class MILNCETask(Task):
+ def reshape_subsample(self, sample):
+ if (
+ hasattr(self.config.dataset, "subsampling")
+ and self.config.dataset.subsampling is not None
+ and self.config.dataset.subsampling > 1
+ ):
+ for key in sample:
+ if torch.is_tensor(sample[key]):
+ tensor = self.flat_subsample(sample[key])
+ if key in ["caps", "cmasks"]:
+ size = tensor.size()
+ batch_size = size[0] * size[1]
+ expanded_size = (batch_size,) + size[2:]
+ tensor = tensor.view(expanded_size)
+ sample[key] = tensor
+ return sample
diff --git a/fairseq/examples/MMPT/mmpt/tasks/retritask.py b/fairseq/examples/MMPT/mmpt/tasks/retritask.py
new file mode 100644
index 0000000000000000000000000000000000000000..b43f20fddb31f29b210b1c726e5f6ccaad04bcf0
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/tasks/retritask.py
@@ -0,0 +1,253 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+import os
+import torch
+import pickle
+import random
+
+from tqdm import tqdm
+from torch.utils.data import DataLoader
+from torch.utils.data.distributed import DistributedSampler
+
+from ..processors import (
+ ShardedHow2MetaProcessor,
+ ShardedVideoProcessor,
+ ShardedTextProcessor,
+ VariedLenAligner,
+)
+
+from ..datasets import MMDataset
+from .task import Task
+from ..modules import vectorpool
+from ..evaluators.predictor import Predictor
+from ..utils import set_seed, get_local_rank, get_world_size
+
+
+class RetriTask(Task):
+ """abstract class for task with retrival."""
+
+ def reshape_subsample(self, sample):
+ for key in sample:
+ if torch.is_tensor(sample[key]):
+ sample[key] = self.flat_subsample(sample[key])
+ return sample
+
+ def flat_subsample(self, tensor):
+ if tensor.size(0) == 1:
+ tensor = tensor.squeeze(0)
+ return tensor
+
+ def build_dataloader(self):
+ """called by `get_batch_iterator` in fairseqmmtask. """
+ # TODO: hard-code dataloader for retri for now and configurable in .yaml.
+ # reuse the `train.lst`.
+ self.config.dataset.split = "train"
+ meta_processor = ShardedHow2MetaProcessor(self.config.dataset)
+ video_processor = ShardedVideoProcessor(self.config.dataset)
+ text_processor = ShardedTextProcessor(self.config.dataset)
+
+ aligner = VariedLenAligner(self.config.dataset)
+ aligner.subsampling = self.config.dataset.clip_per_video
+
+ self.retri_data = MMDataset(
+ meta_processor, video_processor, text_processor, aligner
+ )
+
+ retri_sampler = DistributedSampler(self.retri_data)
+ infer_scale = 16
+ batch_size = self.config.dataset.num_video_per_batch \
+ * infer_scale
+
+ self.retri_dataloader = DataLoader(
+ self.retri_data,
+ collate_fn=self.retri_data.collater,
+ batch_size=batch_size,
+ shuffle=False,
+ sampler=retri_sampler,
+ num_workers=self.config.fairseq.dataset.num_workers
+ )
+ return self.retri_dataloader
+
+ def retrive_candidates(self, epoch, dataloader=None):
+ if get_local_rank() == 0:
+ print("running retrieval model.")
+ out_dir = os.path.join(
+ self.config.fairseq.checkpoint.save_dir, "retri")
+ os.makedirs(out_dir, exist_ok=True)
+
+ if not os.path.isfile(
+ os.path.join(
+ out_dir, "batched_e" + str(epoch) + "_videos0.pkl")
+ ):
+ if dataloader is None:
+ dataloader = self.retri_dataloader
+
+ self.model.eval()
+ self.model.is_train = False
+
+ assert self.retri_data.meta_processor.data == \
+ self.train_data.meta_processor.data # video_ids not mutated.
+
+ self._retri_predict(epoch, dataloader)
+
+ self.model.train()
+ self.model.is_train = True
+
+ torch.distributed.barrier()
+ output = self._retri_sync(epoch, out_dir)
+ torch.distributed.barrier()
+ self.train_data.meta_processor.set_candidates(output)
+ return output
+
+
+class VideoRetriTask(RetriTask):
+ """RetriTask on video level."""
+
+ def reshape_subsample(self, sample):
+ if (
+ hasattr(self.config.dataset, "clip_per_video")
+ and self.config.dataset.clip_per_video is not None
+ and self.config.dataset.clip_per_video > 1
+ ):
+ for key in sample:
+ if torch.is_tensor(sample[key]):
+ sample[key] = self.flat_subsample(sample[key])
+ return sample
+
+ def flat_subsample(self, tensor):
+ if tensor.size(0) == 1:
+ tensor = tensor.squeeze(0)
+ return Task.flat_subsample(self, tensor)
+
+ def _retri_predict(self, epoch, dataloader):
+ set_seed(epoch)
+ # save for retrival.
+ predictor = VideoPredictor(self.config)
+ predictor.predict_loop(
+ self.model, dataloader)
+ set_seed(epoch) # get the same text clips.
+ # retrival.
+ retri_predictor = VideoRetriPredictor(
+ self.config)
+ retri_predictor.predict_loop(
+ self.model, predictor.vecpool.retriver, epoch)
+ del predictor
+ del retri_predictor
+
+ def _retri_sync(self, epoch, out_dir):
+ # gpu do the same merge.
+ batched_videos = []
+ for local_rank in range(get_world_size()):
+ fn = os.path.join(
+ out_dir,
+ "batched_e" + str(epoch) + "_videos" + str(local_rank) + ".pkl")
+ with open(fn, "rb") as fr:
+ batched_videos.extend(pickle.load(fr))
+ print(
+ "[INFO] batched_videos",
+ len(batched_videos), len(batched_videos[0]))
+ return batched_videos
+
+
+class VideoPredictor(Predictor):
+ def __init__(self, config):
+ vectorpool_cls = getattr(vectorpool, config.vectorpool_cls)
+ self.vecpool = vectorpool_cls(config)
+
+ def predict_loop(
+ self,
+ model,
+ dataloader,
+ early_stop=-1,
+ ):
+ with torch.no_grad():
+ if get_local_rank() == 0:
+ dataloader = tqdm(dataloader)
+ for batch_idx, batch in enumerate(dataloader):
+ if batch_idx == early_stop:
+ break
+ self(batch, model)
+ return self.finalize()
+
+ def __call__(self, sample, model, **kwargs):
+ param = next(model.parameters())
+ dtype = param.dtype
+ device = param.device
+ subsample = sample["vfeats"].size(1)
+ sample = self.to_ctx(sample, device, dtype)
+ for key in sample:
+ if torch.is_tensor(sample[key]):
+ size = sample[key].size()
+ if len(size) >= 2:
+ batch_size = size[0] * size[1]
+ expanded_size = (
+ (batch_size,) + size[2:] if len(size) > 2
+ else (batch_size,)
+ )
+ sample[key] = sample[key].view(expanded_size)
+
+ outputs = model(**sample)
+ sample.update(outputs)
+ self.vecpool(sample, subsample)
+
+ def finalize(self):
+ print("[INFO]", self.vecpool)
+ if not self.vecpool.retriver.db.is_trained:
+ self.vecpool.retriver.finalize_training()
+ return self.vecpool.retriver
+
+
+class VideoRetriPredictor(Predictor):
+ """
+ Online Retrieval Predictor for Clips (used by RetriTask).
+ TODO: merge this with VisPredictor?
+ """
+
+ def __init__(self, config):
+ self.pred_dir = os.path.join(
+ config.fairseq.checkpoint.save_dir,
+ "retri")
+ self.num_cands = config.num_cands
+ self.num_video_per_batch = config.dataset.num_video_per_batch
+
+ def predict_loop(
+ self,
+ model,
+ retriver,
+ epoch,
+ early_stop=-1
+ ):
+ # a fake loop that only try to recover video vector
+ # from video_id.
+ batched_videos = []
+ # obtain available video_ids.
+ video_ids = list(retriver.videoid_to_vectoridx.keys())
+
+ dataloader = random.sample(
+ video_ids,
+ len(video_ids) // self.num_video_per_batch
+ )
+
+ if get_local_rank() == 0:
+ dataloader = tqdm(dataloader)
+ for batch_idx, batch in enumerate(dataloader):
+ # batch is one video id.
+ if batch_idx == early_stop:
+ break
+ video_ids = retriver.search_by_video_ids(
+ [batch], self.num_cands)[0]
+ if len(video_ids) > self.num_video_per_batch:
+ # we moved the center to make cluster robust.
+ video_ids = random.sample(video_ids, self.num_video_per_batch)
+ batched_videos.append(video_ids)
+ return self.finalize(batched_videos, epoch)
+
+ def finalize(self, batched_videos, epoch):
+ fn = os.path.join(
+ self.pred_dir,
+ "batched_e" + str(epoch) + "_videos" + str(get_local_rank()) + ".pkl")
+ with open(fn, "wb") as fw:
+ pickle.dump(batched_videos, fw, pickle.HIGHEST_PROTOCOL)
+ return batched_videos
diff --git a/fairseq/examples/MMPT/mmpt/tasks/task.py b/fairseq/examples/MMPT/mmpt/tasks/task.py
new file mode 100644
index 0000000000000000000000000000000000000000..8bb50f24df49878d5b430f9df5ba1ffb1cf30e32
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/tasks/task.py
@@ -0,0 +1,184 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+import torch
+
+from .. import tasks
+from .. import models
+from .. import losses
+from ..datasets import MMDataset
+from .. import processors
+
+
+class Task(object):
+ """
+ A task refers to one generic training task (e.g., training one model).
+ """
+
+ @classmethod
+ def config_task(cls, config):
+ """
+ determine whether to load a hard-coded task or config from a generic one.
+ via if a task string is available in config.
+ """
+ if config.task is not None:
+ # TODO (huxu): expand the search scope.
+ task_cls = getattr(tasks, config.task)
+ return task_cls(config)
+ else:
+ return Task(config)
+
+ def __init__(self, config):
+ self.config = config
+ self.train_data = None
+ self.val_data = None
+ self.test_data = None
+
+ self.model = None
+ self.loss_fn = None
+ self.eval_fn = None
+
+ def build_dataset(self):
+ """TODO (huxu): move processor breakdown to MMDataset."""
+ """fill-in `self.train_data`, `self.val_data` and `self.test_data`."""
+
+ meta_processor_cls = getattr(
+ processors, self.config.dataset.meta_processor)
+ video_processor_cls = getattr(
+ processors, self.config.dataset.video_processor)
+ text_processor_cls = getattr(
+ processors, self.config.dataset.text_processor)
+ aligner_cls = getattr(
+ processors, self.config.dataset.aligner)
+
+ if self.config.dataset.train_path is not None:
+ self.config.dataset.split = "train"
+ # may be used by meta processor.
+ # meta_processor controls different dataset.
+ meta_processor = meta_processor_cls(self.config.dataset)
+ video_processor = video_processor_cls(self.config.dataset)
+ text_processor = text_processor_cls(self.config.dataset)
+ aligner = aligner_cls(self.config.dataset)
+ self.train_data = MMDataset(
+ meta_processor, video_processor, text_processor, aligner
+ )
+ print("train_len", len(self.train_data))
+ output = self.train_data[0]
+ self.train_data.print_example(output)
+ if self.config.dataset.val_path is not None:
+ self.config.dataset.split = "valid"
+ # may be used by meta processor.
+ meta_processor = meta_processor_cls(self.config.dataset)
+ video_processor = video_processor_cls(self.config.dataset)
+ text_processor = text_processor_cls(self.config.dataset)
+ aligner = aligner_cls(self.config.dataset)
+ self.val_data = MMDataset(
+ meta_processor, video_processor, text_processor, aligner
+ )
+ print("val_len", len(self.val_data))
+ output = self.val_data[0]
+ self.val_data.print_example(output)
+
+ if self.config.dataset.split == "test":
+ # the following is run via lauching fairseq-validate.
+ meta_processor = meta_processor_cls(self.config.dataset)
+ video_processor = video_processor_cls(self.config.dataset)
+ text_processor = text_processor_cls(self.config.dataset)
+
+ self.test_data = MMDataset(
+ meta_processor, video_processor, text_processor, aligner
+ )
+ print("test_len", len(self.test_data))
+ output = self.test_data[0]
+ self.test_data.print_example(output)
+
+ def build_model(self, checkpoint=None):
+ if self.model is None:
+ model_cls = getattr(models, self.config.model.model_cls)
+ self.model = model_cls(self.config)
+ if checkpoint is not None:
+ self.load_checkpoint(checkpoint)
+ return self.model
+
+ def load_checkpoint(self, checkpoint):
+ if self.model is None:
+ raise ValueError("model is not initialized.")
+ state_dict = torch.load(checkpoint)
+ state_dict = self._trim_state_dict(state_dict)
+ self.model.load_state_dict(state_dict, strict=False)
+ # if it's a fp16 model, turn it back.
+ if next(self.model.parameters()).dtype == torch.float16:
+ self.model = self.model.float()
+ return self.model
+
+ def _trim_state_dict(self, state_dict):
+ from collections import OrderedDict
+
+ if "state_dict" in state_dict:
+ state_dict = state_dict["state_dict"]
+ if "model" in state_dict: # fairseq checkpoint format.
+ state_dict = state_dict["model"]
+ ret_state_dict = OrderedDict()
+ for (
+ key,
+ value,
+ ) in state_dict.items():
+ # remove fairseq wrapper since this is a task.
+ if key.startswith("mmmodel"):
+ key = key[len("mmmodel."):]
+ ret_state_dict[key] = value
+ return ret_state_dict
+
+ def build_loss(self):
+ if self.loss_fn is None and self.config.loss is not None:
+ loss_cls = getattr(losses, self.config.loss.loss_cls)
+ self.loss_fn = loss_cls()
+ return self.loss_fn
+
+ def flat_subsample(self, tensor):
+ size = tensor.size()
+ if len(size) >= 2:
+ batch_size = size[0] * size[1]
+ expanded_size = (
+ (batch_size,) + size[2:] if len(size) > 2
+ else (batch_size,)
+ )
+ tensor = tensor.view(expanded_size)
+ return tensor
+
+ def reshape_subsample(self, sample):
+ if (
+ hasattr(self.config.dataset, "subsampling")
+ and self.config.dataset.subsampling is not None
+ and self.config.dataset.subsampling > 1
+ ):
+ for key in sample:
+ if torch.is_tensor(sample[key]):
+ sample[key] = self.flat_subsample(sample[key])
+ return sample
+
+ def __call__(self, model, sample):
+ loss = None
+ loss_scalar = float("inf")
+
+ sample = self.reshape_subsample(sample)
+ outputs = self.model(**sample)
+ sample.update(outputs)
+ if self.loss_fn is not None:
+ loss = self.loss_fn(**sample)
+ loss_scalar = loss.item()
+
+ batch_size = sample["caps"].size(0)
+ sample_size = 1
+ return {
+ "loss": loss,
+ "loss_scalar": loss_scalar,
+ "max_len": self.config.dataset.max_len,
+ "batch_size": batch_size,
+ "sample_size": sample_size,
+ }
+
+ def build_dataloader(self):
+ """only used for trainer that lacks building loaders."""
+ raise NotImplementedError
diff --git a/fairseq/examples/MMPT/mmpt/tasks/vlmtask.py b/fairseq/examples/MMPT/mmpt/tasks/vlmtask.py
new file mode 100644
index 0000000000000000000000000000000000000000..57dc4c91705fdb1292f2f2accbb42acb993eb6aa
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/tasks/vlmtask.py
@@ -0,0 +1,27 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+import torch
+
+from .task import Task
+
+
+class VLMTask(Task):
+ """A VLM task for reproducibility.
+ the collator split subsamples into two sub-batches.
+ This has should have no logic changes.
+ but changed the randomness in frame masking.
+ """
+
+ def flat_subsample(self, tensor):
+ size = tensor.size()
+ if len(size) >= 2:
+ batch_size = size[0] * (size[1] // 2)
+ expanded_size = (
+ (batch_size, 2) + size[2:] if len(size) > 2
+ else (batch_size, 2)
+ )
+ tensor = tensor.view(expanded_size)
+ tensor = torch.cat([tensor[:, 0], tensor[:, 1]], dim=0)
+ return tensor
diff --git a/fairseq/examples/MMPT/mmpt/utils/__init__.py b/fairseq/examples/MMPT/mmpt/utils/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..2429ee3757353e768f71b27d129eb3ca3bcbec73
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/utils/__init__.py
@@ -0,0 +1,68 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+import random
+import numpy as np
+import torch
+
+from .shardedtensor import *
+from .load_config import *
+
+
+def set_seed(seed=43211):
+ random.seed(seed)
+ np.random.seed(seed)
+ torch.manual_seed(seed)
+ torch.cuda.manual_seed_all(seed)
+ if torch.backends.cudnn.enabled:
+ torch.backends.cudnn.benchmark = False
+ torch.backends.cudnn.deterministic = True
+
+
+def get_world_size():
+ if torch.distributed.is_initialized():
+ world_size = torch.distributed.get_world_size()
+ else:
+ world_size = 1
+ return world_size
+
+
+def get_local_rank():
+ return torch.distributed.get_rank() \
+ if torch.distributed.is_initialized() else 0
+
+
+def print_on_rank0(func):
+ local_rank = get_local_rank()
+ if local_rank == 0:
+ print("[INFO]", func)
+
+
+class RetriMeter(object):
+ """
+ Statistics on whether retrieval yields a better pair.
+ """
+ def __init__(self, freq=1024):
+ self.freq = freq
+ self.total = 0
+ self.replace = 0
+ self.updates = 0
+
+ def __call__(self, data):
+ if isinstance(data, np.ndarray):
+ self.replace += data.shape[0] - int((data[:, 0] == -1).sum())
+ self.total += data.shape[0]
+ elif torch.is_tensor(data):
+ self.replace += int(data.sum())
+ self.total += data.size(0)
+ else:
+ raise ValueError("unsupported RetriMeter data type.", type(data))
+
+ self.updates += 1
+ if get_local_rank() == 0 and self.updates % self.freq == 0:
+ print("[INFO]", self)
+
+ def __repr__(self):
+ return "RetriMeter (" + str(self.replace / self.total) \
+ + "/" + str(self.replace) + "/" + str(self.total) + ")"
diff --git a/fairseq/examples/MMPT/mmpt/utils/load_config.py b/fairseq/examples/MMPT/mmpt/utils/load_config.py
new file mode 100644
index 0000000000000000000000000000000000000000..ede4f94117b197e5bdb551c5140604c6e47c91fb
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/utils/load_config.py
@@ -0,0 +1,81 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+import os
+import omegaconf
+from omegaconf import OmegaConf
+
+
+def load_config(args=None, config_file=None, overwrite_fairseq=False):
+ """TODO (huxu): move fairseq overwrite to another function."""
+ if args is not None:
+ config_file = args.taskconfig
+ config = recursive_config(config_file)
+
+ if config.dataset.subsampling is not None:
+ batch_size = config.fairseq.dataset.batch_size // config.dataset.subsampling
+ print(
+ "adjusting batch_size to {} due to subsampling {}.".format(
+ batch_size, config.dataset.subsampling
+ )
+ )
+ config.fairseq.dataset.batch_size = batch_size
+
+ is_test = config.dataset.split is not None and config.dataset.split == "test"
+ if not is_test:
+ if (
+ config.fairseq.checkpoint is None
+ or config.fairseq.checkpoint.save_dir is None
+ ):
+ raise ValueError("fairseq save_dir or save_path must be specified.")
+
+ save_dir = config.fairseq.checkpoint.save_dir
+ os.makedirs(save_dir, exist_ok=True)
+ if config.fairseq.common.tensorboard_logdir is not None:
+ tb_run_dir = suffix_rundir(
+ save_dir, config.fairseq.common.tensorboard_logdir
+ )
+ config.fairseq.common.tensorboard_logdir = tb_run_dir
+ print(
+ "update tensorboard_logdir as", config.fairseq.common.tensorboard_logdir
+ )
+ os.makedirs(save_dir, exist_ok=True)
+ OmegaConf.save(config=config, f=os.path.join(save_dir, "config.yaml"))
+
+ if overwrite_fairseq and config.fairseq is not None and args is not None:
+ # flatten fields.
+ for group in config.fairseq:
+ for field in config.fairseq[group]:
+ print("overwrite args." + field, "as", config.fairseq[group][field])
+ setattr(args, field, config.fairseq[group][field])
+ return config
+
+
+def recursive_config(config_path):
+ """allows for stacking of configs in any depth."""
+ config = OmegaConf.load(config_path)
+ if config.includes is not None:
+ includes = config.includes
+ config.pop("includes")
+ base_config = recursive_config(includes)
+ config = OmegaConf.merge(base_config, config)
+ return config
+
+
+def suffix_rundir(save_dir, run_dir):
+ max_id = -1
+ for search_dir in os.listdir(save_dir):
+ if search_dir.startswith(run_dir):
+ splits = search_dir.split("_")
+ cur_id = int(splits[1]) if len(splits) > 1 else 0
+ max_id = max(max_id, cur_id)
+ return os.path.join(save_dir, run_dir + "_" + str(max_id + 1))
+
+
+def overwrite_dir(config, replace, basedir):
+ for key in config:
+ if isinstance(config[key], str) and config[key].startswith(basedir):
+ config[key] = config[key].replace(basedir, replace)
+ if isinstance(config[key], omegaconf.dictconfig.DictConfig):
+ overwrite_dir(config[key], replace, basedir)
diff --git a/fairseq/examples/MMPT/mmpt/utils/shardedtensor.py b/fairseq/examples/MMPT/mmpt/utils/shardedtensor.py
new file mode 100644
index 0000000000000000000000000000000000000000..2424f360ef2a93f0f2783e38f0bd966086e92fd4
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/utils/shardedtensor.py
@@ -0,0 +1,46 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+import os
+import pickle
+import numpy as np
+
+
+class ShardedTensor(object):
+ def __init__(self, data, starts):
+ self.data = data
+ self.starts = starts
+ assert self.starts[0] == 0
+ assert self.starts[-1] == len(self.data)
+ assert (self.starts[1:] >= self.starts[:-1]).all()
+ assert (self.starts > -1).all()
+
+ @staticmethod
+ def from_list(xs):
+ starts = np.full((len(xs) + 1,), -1, dtype=np.long)
+ data = np.concatenate(xs, axis=0)
+ starts[0] = 0
+ for i, x in enumerate(xs):
+ starts[i + 1] = starts[i] + x.shape[0]
+ assert (starts > -1).all()
+ return ShardedTensor(data, starts)
+
+ def __getitem__(self, i):
+ return self.data[self.starts[i] : self.starts[i + 1]]
+
+ def __len__(self):
+ return len(self.starts) - 1
+
+ def lengths(self):
+ return self.starts[1:] - self.starts[:-1]
+
+ def save(self, path):
+ np.save(path + "_starts", self.starts)
+ np.save(path + "_data", self.data)
+
+ @staticmethod
+ def load(path, mmap_mode=None):
+ starts = np.load(path + "_starts.npy", mmap_mode)
+ data = np.load(path + "_data.npy", mmap_mode)
+ return ShardedTensor(data, starts)
diff --git a/fairseq/examples/MMPT/mmpt_cli/localjob.py b/fairseq/examples/MMPT/mmpt_cli/localjob.py
new file mode 100644
index 0000000000000000000000000000000000000000..2675d3511a9ca700185d2d7b853dc56ad70c638c
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt_cli/localjob.py
@@ -0,0 +1,117 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+import os
+
+from mmpt.utils import recursive_config
+
+
+class BaseJob(object):
+ def __init__(self, yaml_file, dryrun=False):
+ self.yaml_file = yaml_file
+ self.config = recursive_config(yaml_file)
+ self.dryrun = dryrun
+
+ def submit(self, **kwargs):
+ raise NotImplementedError
+
+ def _normalize_cmd(self, cmd_list):
+ cmd_list = list(cmd_list)
+ yaml_index = cmd_list.index("[yaml]")
+ cmd_list[yaml_index] = self.yaml_file
+ return cmd_list
+
+
+class LocalJob(BaseJob):
+
+ CMD_CONFIG = {
+ "local_single": [
+ "fairseq-train", "[yaml]", "--user-dir", "mmpt",
+ "--task", "mmtask", "--arch", "mmarch",
+ "--criterion", "mmloss",
+ ],
+ "local_small": [
+ "fairseq-train", "[yaml]", "--user-dir", "mmpt",
+ "--task", "mmtask", "--arch", "mmarch",
+ "--criterion", "mmloss",
+ "--distributed-world-size", "2"
+ ],
+ "local_big": [
+ "fairseq-train", "[yaml]", "--user-dir", "mmpt",
+ "--task", "mmtask", "--arch", "mmarch",
+ "--criterion", "mmloss",
+ "--distributed-world-size", "8"
+ ],
+ "local_predict": ["python", "mmpt_cli/predict.py", "[yaml]"],
+ }
+
+ def __init__(self, yaml_file, job_type=None, dryrun=False):
+ super().__init__(yaml_file, dryrun)
+ if job_type is None:
+ self.job_type = "local_single"
+ if self.config.task_type is not None:
+ self.job_type = self.config.task_type
+ else:
+ self.job_type = job_type
+ if self.job_type in ["local_single", "local_small"]:
+ if self.config.fairseq.dataset.batch_size > 32:
+ print("decreasing batch_size to 32 for local testing?")
+
+ def submit(self):
+ cmd_list = self._normalize_cmd(LocalJob.CMD_CONFIG[self.job_type])
+ if "predict" not in self.job_type:
+ # append fairseq args.
+ from mmpt.utils import load_config
+
+ config = load_config(config_file=self.yaml_file)
+ for field in config.fairseq:
+ for key in config.fairseq[field]:
+ if key in ["fp16", "reset_optimizer", "reset_dataloader", "reset_meters"]: # a list of binary flag.
+ param = ["--" + key.replace("_", "-")]
+ else:
+ if key == "lr":
+ value = str(config.fairseq[field][key][0])
+ elif key == "adam_betas":
+ value = "'"+str(config.fairseq[field][key])+"'"
+ else:
+ value = str(config.fairseq[field][key])
+ param = [
+ "--" + key.replace("_", "-"),
+ value
+ ]
+ cmd_list.extend(param)
+
+ print("launching", " ".join(cmd_list))
+ if not self.dryrun:
+ os.system(" ".join(cmd_list))
+ return JobStatus("12345678")
+
+
+class JobStatus(object):
+ def __init__(self, job_id):
+ self.job_id = job_id
+
+ def __repr__(self):
+ return self.job_id
+
+ def __str__(self):
+ return self.job_id
+
+ def done(self):
+ return False
+
+ def running(self):
+ return False
+
+ def result(self):
+ if self.done():
+ return "{} is done.".format(self.job_id)
+ else:
+ return "{} is running.".format(self.job_id)
+
+ def stderr(self):
+ return self.result()
+
+ def stdout(self):
+ return self.result()
diff --git a/fairseq/examples/MMPT/mmpt_cli/predict.py b/fairseq/examples/MMPT/mmpt_cli/predict.py
new file mode 100644
index 0000000000000000000000000000000000000000..4071e196d211f7b11170db2e7e35b716d3deeb69
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt_cli/predict.py
@@ -0,0 +1,113 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+import os
+import glob
+import argparse
+import pprint
+import omegaconf
+
+from omegaconf import OmegaConf
+from torch.utils.data import DataLoader
+
+from mmpt.utils import load_config, set_seed
+from mmpt.evaluators import Evaluator
+from mmpt.evaluators import predictor as predictor_path
+from mmpt.tasks import Task
+from mmpt import processors
+from mmpt.datasets import MMDataset
+
+
+def get_dataloader(config):
+ meta_processor_cls = getattr(processors, config.dataset.meta_processor)
+ video_processor_cls = getattr(processors, config.dataset.video_processor)
+ text_processor_cls = getattr(processors, config.dataset.text_processor)
+ aligner_cls = getattr(processors, config.dataset.aligner)
+
+ meta_processor = meta_processor_cls(config.dataset)
+ video_processor = video_processor_cls(config.dataset)
+ text_processor = text_processor_cls(config.dataset)
+ aligner = aligner_cls(config.dataset)
+
+ test_data = MMDataset(
+ meta_processor,
+ video_processor,
+ text_processor,
+ aligner,
+ )
+ print("test_len", len(test_data))
+ output = test_data[0]
+ test_data.print_example(output)
+
+ test_dataloader = DataLoader(
+ test_data,
+ batch_size=config.fairseq.dataset.batch_size,
+ shuffle=False,
+ num_workers=6,
+ collate_fn=test_data.collater,
+ )
+ return test_dataloader
+
+
+def main(args):
+ config = load_config(args)
+
+ if isinstance(config, omegaconf.dictconfig.DictConfig):
+ print(OmegaConf.to_yaml(config))
+ else:
+ pp = pprint.PrettyPrinter(indent=4)
+ pp.print(config)
+
+ mmtask = Task.config_task(config)
+ mmtask.build_model()
+
+ test_dataloader = get_dataloader(config)
+ checkpoint_search_path = os.path.dirname(config.eval.save_path)
+ results = []
+
+ prefix = os.path.basename(args.taskconfig)
+ if prefix.startswith("test"):
+ # loop all checkpoint for datasets without validation set.
+ if "best" not in config.fairseq.common_eval.path:
+ print("eval each epoch.")
+ for checkpoint in glob.glob(checkpoint_search_path + "/checkpoint*"):
+ model = mmtask.load_checkpoint(checkpoint)
+ ckpt = os.path.basename(checkpoint)
+ evaluator = Evaluator(config)
+ output = evaluator.evaluate(
+ model, test_dataloader, ckpt + "_merged")
+ results.append((checkpoint, output))
+ # use the one specified by the config lastly.
+ model = mmtask.load_checkpoint(config.fairseq.common_eval.path)
+ evaluator = Evaluator(config)
+ output = evaluator.evaluate(model, test_dataloader)
+ results.append((config.fairseq.common_eval.path, output))
+
+ best_result = None
+ best_metric = 0.
+ for checkpoint, result in results:
+ print(checkpoint)
+ evaluator.metric.print_computed_metrics(result)
+ best_score = evaluator.metric.best_metric(result)
+ if best_score > best_metric:
+ best_result = (checkpoint, result)
+ best_metric = best_score
+ print("best results:")
+ print(best_result[0])
+ evaluator.metric.print_computed_metrics(best_result[1])
+
+ elif prefix.startswith("vis"):
+ model = mmtask.load_checkpoint(config.fairseq.common_eval.path)
+ predictor_cls = getattr(predictor_path, config.predictor)
+ predictor = predictor_cls(config)
+ predictor.predict_loop(model, test_dataloader, mmtask, None)
+ else:
+ raise ValueError("unknown prefix of the config file", args.taskconfig)
+
+
+if __name__ == "__main__":
+ parser = argparse.ArgumentParser()
+ parser.add_argument("taskconfig", type=str)
+ args = parser.parse_args()
+ main(args)
diff --git a/fairseq/examples/MMPT/pretraining.md b/fairseq/examples/MMPT/pretraining.md
new file mode 100644
index 0000000000000000000000000000000000000000..8f8e6d0facaa47141342294d7eb26c4232a9677b
--- /dev/null
+++ b/fairseq/examples/MMPT/pretraining.md
@@ -0,0 +1,29 @@
+# Pretraining
+
+(If you are new to the ideas of `mmpt.processors`, see [README](README.md) first.)
+We mostly use [howto100M](https://github.com/antoine77340/howto100m) dataset for pretraining (other datasets are coming). So you are less likely to write a new `MetaProcessor`, `VideoProcessor` or `TextProcessor` but only working on a new `Aligner`, a new model and loss.
+
+### Data Sharding
+Pretraining on Howto100M is heavy on IO since we have millions of videos or captions on the hard disk that cannot be fit into the memory.
+It is desirable to have an optimized preprocessing step before the actual dataloading.
+
+We support data sharding to pack multiple videos into a shards of training data for both videos and captions. (see [dataset](DATASET.md) for preprocessing).
+These shards will be mapped into memory to reduce the frequency of IO access on millions of files. See (processors starting with `Sharded*`).
+This will be the default config for a how2 dataset `projects/task/how2.yaml`.
+
+Great thanks to Dmytro Okhonko for sharing the code from MARGE project.
+
+### Training
+Pretraining on Howto100m is expected on one or multiple nodes, where each node has 8 GPUS with 32 GB mem.
+launching a pretraing on MFM+MLM can be done, via:
+```python locallaunch.py projects/mfmmlm/how2.yaml```
+
+### Pre-training with a Retrieval Model (VideoCLIP)
+This projects now support alternatively run a retrieval model and pre-training.
+We implement a basic retrieval model that is built on the hidden states of a video and faiss.
+
+You may need to install faiss via `conda install faiss-cpu -c pytorch`.
+
+Right now, the hidden states of a video is computed as the average of 8 clips of their pooled visual/text hidden states.
+See `mmpt/tasks/retritask.py` for more details.
+The `.yaml` config for running pre-training with a retrieval model can be found at `projects/retri/videoretri.yaml`.
diff --git a/fairseq/examples/MMPT/projects/mfmmlm.yaml b/fairseq/examples/MMPT/projects/mfmmlm.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..0f3450a1e091517de19422b0c09cea6a01a02729
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/mfmmlm.yaml
@@ -0,0 +1,59 @@
+project_dir: mfmmlm
+run_task:
+ - how2.yaml
+ - [vtt.yaml, vttcap.yaml, vttqa.yaml, youcook.yaml, youcookcap.yaml, crosstask.yaml, coin.yaml]
+base_dir: task
+task_group:
+ pretrain:
+ task_list:
+ - how2.yaml
+ dataset:
+ subsampling: 32
+ sampled_min_len: 10
+ sampled_max_len: 64
+ max_video_len: 32
+ max_len: 96
+ aligner: MFMMLMAligner
+ lazy_vfeat_mask: True
+ mfm_probability: 0.15
+ mlm_probability: 0.15
+ mm_prob: 0.5
+ model:
+ model_cls: MMFusionMFMMLM
+ mm_encoder_cls: MMFusionForMFMMLM
+ loss:
+ loss_cls: MFMMLM
+ fairseq:
+ common:
+ fp16: true
+ dataset:
+ batch_size: 256
+ optimization:
+ max_epoch: 15
+ finetune:
+ task_list:
+ - vtt.yaml
+ - vttqa.yaml
+ - youcook.yaml
+ - youcookcap.yaml
+ - crosstask.yaml
+ - coin.yaml
+ dataset:
+ max_video_len: 32
+ max_len: 96
+ fairseq:
+ common:
+ fp16: true
+ # do not write any model or loss here (they are expected to be fixed in mmfusion).
+ test:
+ task_list:
+ - test_vtt.yaml
+ - test_vttqa.yaml
+ - test_youcook.yaml
+ - test_youcookcap.yaml
+ - test_crosstask.yaml
+ - test_crosstask_zs.yaml
+ - test_coin.yaml
+ dataset:
+ max_video_len: 32
+ max_len: 96
diff --git a/fairseq/examples/MMPT/projects/mtm/mmfusionmtm.yaml b/fairseq/examples/MMPT/projects/mtm/mmfusionmtm.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..337d66a2aadcff86da033d6656bdc570a20c32e6
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/mtm/mmfusionmtm.yaml
@@ -0,0 +1,19 @@
+includes: projects/mfmmlm.yaml
+project_dir: mtm/mmfusionmtm
+task_group:
+ pretrain:
+ task: VLMTask # reproducible
+ dataset:
+ aligner: MFMMLMAligner
+ model:
+ use_seg_emb: True # reproducible
+ model_cls: MMFusionMTM
+ mm_encoder_cls: MMBertForMFMMLM
+ loss:
+ loss_cls: MTM
+ finetune:
+ model:
+ use_seg_emb: True # reproducible
+ test:
+ model:
+ use_seg_emb: True # reproducible
diff --git a/fairseq/examples/MMPT/projects/mtm/vlm.yaml b/fairseq/examples/MMPT/projects/mtm/vlm.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..022a2623c5fdee1da51f211bef25a0dfebd7cbb0
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/mtm/vlm.yaml
@@ -0,0 +1,8 @@
+includes: projects/mtm/mmfusionmtm.yaml
+project_dir: mtm/vlm
+task_group:
+ pretrain:
+ dataset:
+ sampled_min_len: 8
+ loss:
+ loss_cls: MTM
diff --git a/fairseq/examples/MMPT/projects/mtm/vlm/coin.yaml b/fairseq/examples/MMPT/projects/mtm/vlm/coin.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..48fd64a5f4996a03b412c418e83c4672b7252e8d
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/mtm/vlm/coin.yaml
@@ -0,0 +1,47 @@
+dataset:
+ video_processor: VideoProcessor
+ bert_name: bert-base-uncased
+ meta_processor: COINActionSegmentationMetaProcessor
+ train_path: data/coin/COIN.json
+ val_path: data/coin/COIN.json
+ vfeat_dir: data/feat/feat_coin_s3d
+ text_processor: COINActionSegmentationTextProcessor
+ aligner: COINActionSegmentationAligner
+ num_iso_layer: 12
+ sliding_window: 8
+ sliding_window_size: 32
+ max_video_len: 32
+ max_len: 96
+fairseq:
+ common:
+ tensorboard_logdir: run
+ log_interval: 1000
+ fp16: true
+ dataset:
+ num_workers: 4
+ batch_size: 1
+ optimization:
+ lr:
+ - 5.0e-05
+ clip_norm: 2.0
+ optimizer: adam
+ adam_betas: (0.9, 0.98)
+ lr_scheduler: polynomial_decay
+ total_num_update: 1000000
+ warmup_updates: 122
+ weight_decay: 0.0
+ ddp_backend: no_c10d
+ max_epoch: 8
+ checkpoint:
+ restore_file: runs/mtm/vlm/checkpoint_best.pt
+ reset_optimizer: true
+ reset_dataloader: true
+ reset_meters: true
+ save_dir: runs/mtm/vlm/coin
+task_type: sweep_big
+model:
+ model_cls: MMFusionActionSegmentation
+ mm_encoder_cls: MMBertForTokenClassification
+ use_seg_emb: true
+loss:
+ loss_cls: CrossEntropy
diff --git a/fairseq/examples/MMPT/projects/mtm/vlm/crosstask.yaml b/fairseq/examples/MMPT/projects/mtm/vlm/crosstask.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..4e706b549e6085c9823049afaff3285aa2b88f4f
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/mtm/vlm/crosstask.yaml
@@ -0,0 +1,53 @@
+dataset:
+ video_processor: CrossTaskVideoProcessor
+ bert_name: bert-base-uncased
+ meta_processor: CrossTaskMetaProcessor
+ train_path: data/crosstask/crosstask_release/videos.csv
+ train_csv_path: data/crosstask/crosstask_release/videos.csv
+ val_path: data/crosstask/crosstask_release/videos_val.csv
+ val_csv_path: data/crosstask/crosstask_release/videos_val.csv
+ primary_path: data/crosstask/crosstask_release/tasks_primary.txt
+ related_path: data/crosstask/crosstask_release/tasks_related.txt
+ vfeat_dir: data/feat/feat_crosstask_s3d
+ annotation_path: data/crosstask/crosstask_release/annotations
+ n_train: 30
+ text_processor: CrossTaskTextProcessor
+ aligner: CrossTaskAligner
+ num_iso_layer: 12
+ sliding_window: 16
+ sliding_window_size: 32
+ max_video_len: 32
+ max_len: 96
+fairseq:
+ common:
+ tensorboard_logdir: run
+ log_interval: 1000
+ fp16: true
+ dataset:
+ num_workers: 4
+ batch_size: 1
+ optimization:
+ lr:
+ - 5.0e-05
+ clip_norm: 2.0
+ optimizer: adam
+ adam_betas: (0.9, 0.98)
+ lr_scheduler: polynomial_decay
+ total_num_update: 1000000
+ warmup_updates: 122
+ weight_decay: 0.0
+ ddp_backend: no_c10d
+ max_epoch: 5
+ checkpoint:
+ restore_file: runs/mtm/vlm/checkpoint11.pt
+ reset_optimizer: true
+ reset_dataloader: true
+ reset_meters: true
+ save_dir: runs/mtm/vlm/crosstask
+task_type: sweep_small
+model:
+ model_cls: MMFusionActionLocalization
+ mm_encoder_cls: MMBertForJoint
+ use_seg_emb: true
+loss:
+ loss_cls: BCE
diff --git a/fairseq/examples/MMPT/projects/mtm/vlm/how2.yaml b/fairseq/examples/MMPT/projects/mtm/vlm/how2.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..7ca40ad815d9393bc9dc4d248cf7ec8810f1011f
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/mtm/vlm/how2.yaml
@@ -0,0 +1,55 @@
+dataset:
+ video_processor: ShardedVideoProcessor
+ bert_name: bert-base-uncased
+ meta_processor: ShardedHow2MetaProcessor
+ train_path: data/how2/how2_s3d_train.lst
+ val_path: data/how2/how2_s3d_val.lst
+ vfeat_dir: data/feat/feat_how2_s3d_shard_small
+ text_processor: ShardedTextProcessor
+ tfeat_dir: data/feat/feat_how2_s3d_shard_small/raw_caption_dedup.bert-base-uncased.
+ aligner: MFMMLMAligner
+ subsampling: 32
+ sampled_min_len: 8
+ sampled_max_len: 64
+ max_video_len: 32
+ max_len: 96
+ lazy_vfeat_mask: true
+ mfm_probability: 0.15
+ mlm_probability: 0.15
+ mm_prob: 0.5
+fairseq:
+ common:
+ tensorboard_logdir: run
+ log_interval: 1000
+ fp16: true
+ dataset:
+ num_workers: 4
+ batch_size: 256
+ optimization:
+ lr:
+ - 5.0e-05
+ clip_norm: 2.0
+ optimizer: adam
+ adam_betas: (0.9, 0.98)
+ lr_scheduler: polynomial_decay
+ total_num_update: 1000000
+ warmup_updates: 1000
+ weight_decay: 0.0
+ ddp_backend: no_c10d
+ max_epoch: 15
+ checkpoint:
+ save_dir: runs/mtm/vlm
+ save_interval_updates: 1024
+ keep_interval_updates: 2
+ keep_last_epochs: 30
+task_type: sweep_big
+slurm_config: big
+eval:
+ save_path: runs/mtm/vlm
+model:
+ model_cls: MMFusionMTM
+ mm_encoder_cls: MMBertForMFMMLM
+ use_seg_emb: true
+loss:
+ loss_cls: MTM
+task: VLMTask
diff --git a/fairseq/examples/MMPT/projects/mtm/vlm/test_coin.yaml b/fairseq/examples/MMPT/projects/mtm/vlm/test_coin.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..8df2e66ad175529e0e86f69ade1e392849507036
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/mtm/vlm/test_coin.yaml
@@ -0,0 +1,31 @@
+slurm_config: big
+task_type: local_predict
+dataset:
+ split: test
+ video_processor: VideoProcessor
+ aligner: COINActionSegmentationAligner
+ bert_name: bert-base-uncased
+ test_path: data/coin/COIN.json
+ meta_processor: COINActionSegmentationMetaProcessor
+ vfeat_dir: data/feat/feat_coin_s3d
+ text_processor: COINActionSegmentationTextProcessor
+ num_iso_layer: 12
+ sliding_window: 16
+ sliding_window_size: 32
+ max_video_len: 32
+ max_len: 96
+fairseq:
+ dataset:
+ batch_size: 1
+ valid_subset: test
+ num_workers: 2
+ common_eval:
+ path: runs/mtm/vlm/coin/checkpoint_best.pt
+model:
+ model_cls: MMFusionActionSegmentation
+ mm_encoder_cls: MMBertForTokenClassification
+ use_seg_emb: true
+eval:
+ save_path: runs/mtm/vlm/coin/eval
+metric: COINActionSegmentationMetric
+predictor: COINPredictor
diff --git a/fairseq/examples/MMPT/projects/mtm/vlm/test_crosstask.yaml b/fairseq/examples/MMPT/projects/mtm/vlm/test_crosstask.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d159847875f74050fbae5f0c223fa6d56bbd3d38
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/mtm/vlm/test_crosstask.yaml
@@ -0,0 +1,38 @@
+slurm_config: big
+task_type: local_predict
+dataset:
+ split: test
+ video_processor: CrossTaskVideoProcessor
+ aligner: CrossTaskAligner
+ bert_name: bert-base-uncased
+ meta_processor: CrossTaskMetaProcessor
+ test_path: data/crosstask/crosstask_release/videos_val.csv
+ train_csv_path: data/crosstask/crosstask_release/videos.csv
+ val_path: data/crosstask/crosstask_release/videos_val.csv
+ val_csv_path: data/crosstask/crosstask_release/videos_val.csv
+ primary_path: data/crosstask/crosstask_release/tasks_primary.txt
+ related_path: data/crosstask/crosstask_release/tasks_related.txt
+ vfeat_dir: data/feat/feat_crosstask_s3d
+ annotation_path: data/crosstask/crosstask_release/annotations
+ n_train: 30
+ text_processor: CrossTaskTextProcessor
+ num_iso_layer: 12
+ sliding_window: 16
+ sliding_window_size: 32
+ max_video_len: 32
+ max_len: 96
+fairseq:
+ dataset:
+ batch_size: 1
+ valid_subset: test
+ num_workers: 2
+ common_eval:
+ path: runs/mtm/vlm/crosstask/checkpoint_best.pt
+model:
+ model_cls: MMFusionActionLocalization
+ mm_encoder_cls: MMBertForJoint
+ use_seg_emb: true
+eval:
+ save_path: runs/mtm/vlm/crosstask/eval
+metric: CrossTaskMetric
+predictor: CrossTaskPredictor
diff --git a/fairseq/examples/MMPT/projects/mtm/vlm/test_crosstask_zs.yaml b/fairseq/examples/MMPT/projects/mtm/vlm/test_crosstask_zs.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..59833c55406ce53225d704b46eb0c141ad4d5e9d
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/mtm/vlm/test_crosstask_zs.yaml
@@ -0,0 +1,38 @@
+slurm_config: big
+task_type: local_predict
+dataset:
+ split: test
+ video_processor: CrossTaskVideoProcessor
+ aligner: CrossTaskAligner
+ bert_name: bert-base-uncased
+ meta_processor: CrossTaskMetaProcessor
+ test_path: data/crosstask/crosstask_release/videos_val.csv
+ train_csv_path: data/crosstask/crosstask_release/videos.csv
+ val_path: data/crosstask/crosstask_release/videos_val.csv
+ val_csv_path: data/crosstask/crosstask_release/videos_val.csv
+ primary_path: data/crosstask/crosstask_release/tasks_primary.txt
+ related_path: data/crosstask/crosstask_release/tasks_related.txt
+ vfeat_dir: data/feat/feat_crosstask_s3d
+ annotation_path: data/crosstask/crosstask_release/annotations
+ n_train: 30
+ text_processor: CrossTaskTextProcessor
+ num_iso_layer: 12
+ sliding_window: 16
+ sliding_window_size: 32
+ max_video_len: 32
+ max_len: 96
+fairseq:
+ dataset:
+ batch_size: 1
+ valid_subset: test
+ num_workers: 2
+ common_eval:
+ path: runs/mtm/vlm/checkpoint_best.pt
+model:
+ model_cls: MMFusionActionLocalization
+ mm_encoder_cls: MMBertForJoint
+ use_seg_emb: true
+eval:
+ save_path: runs/mtm/vlm/crosstask_zs/eval
+metric: CrossTaskMetric
+predictor: CrossTaskPredictor
diff --git a/fairseq/examples/MMPT/projects/mtm/vlm/test_vtt.yaml b/fairseq/examples/MMPT/projects/mtm/vlm/test_vtt.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..a41557df6af496c755b0b32345a736b1cdc01ce4
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/mtm/vlm/test_vtt.yaml
@@ -0,0 +1,29 @@
+slurm_config: big
+task_type: local_predict
+dataset:
+ split: test
+ video_processor: VideoProcessor
+ aligner: DSAligner
+ bert_name: bert-base-uncased
+ meta_processor: MSRVTTMetaProcessor
+ test_path: data/msrvtt/MSRVTT_JSFUSION_test.csv
+ vfeat_dir: data/feat/feat_vtt_s3d
+ text_processor: MSRVTTTextProcessor
+ num_iso_layer: 12
+ max_video_len: 32
+ max_len: 96
+fairseq:
+ dataset:
+ batch_size: 256
+ valid_subset: test
+ num_workers: 2
+ common_eval:
+ path: runs/mtm/vlm/vtt/checkpoint_last.pt
+model:
+ model_cls: MMFusionJoint
+ mm_encoder_cls: MMBertForJoint
+ use_seg_emb: true
+eval:
+ save_path: runs/mtm/vlm/vtt/eval
+metric: RetrievalMetric
+predictor: RetrievalPredictor
diff --git a/fairseq/examples/MMPT/projects/mtm/vlm/test_vttqa.yaml b/fairseq/examples/MMPT/projects/mtm/vlm/test_vttqa.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..abf3309f7072c88a6b7c72e2540d4950d0e55575
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/mtm/vlm/test_vttqa.yaml
@@ -0,0 +1,29 @@
+slurm_config: big
+task_type: local_predict
+dataset:
+ split: test
+ video_processor: VideoProcessor
+ aligner: MSRVTTQAAligner
+ bert_name: bert-base-uncased
+ meta_processor: MSRVTTQAMetaProcessor
+ test_path: data/msrvtt-qa/MSR_MC_test.csv
+ vfeat_dir: data/feat/feat_vtt_s3d
+ text_processor: MSRVTTQATextProcessor
+ num_iso_layer: 12
+ max_video_len: 32
+ max_len: 96
+fairseq:
+ dataset:
+ batch_size: 256
+ valid_subset: test
+ num_workers: 2
+ common_eval:
+ path: runs/mtm/vlm/vttqa/checkpoint_last.pt
+model:
+ model_cls: MMFusionJoint
+ mm_encoder_cls: MMBertForJoint
+ use_seg_emb: true
+eval:
+ save_path: runs/mtm/vlm/vttqa/eval
+metric: QAMetric
+predictor: QAPredictor
diff --git a/fairseq/examples/MMPT/projects/mtm/vlm/test_youcook.yaml b/fairseq/examples/MMPT/projects/mtm/vlm/test_youcook.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..3a57d25c240758cf989ad0ec538562d046446232
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/mtm/vlm/test_youcook.yaml
@@ -0,0 +1,31 @@
+slurm_config: big
+task_type: local_predict
+dataset:
+ split: test
+ video_processor: YoucookVideoProcessor
+ aligner: DSAligner
+ bert_name: bert-base-uncased
+ meta_processor: YoucookMetaProcessor
+ test_path: data/youcook/youcook_val.pkl
+ trainval_annotation: data/youcook/youcookii_annotations_trainval.json
+ use_annotation_text: true
+ vfeat_dir: data/feat/feat_youcook_s3d
+ text_processor: TextProcessor
+ num_iso_layer: 12
+ max_video_len: 32
+ max_len: 96
+fairseq:
+ dataset:
+ batch_size: 256
+ valid_subset: test
+ num_workers: 2
+ common_eval:
+ path: runs/mtm/vlm/youcook/checkpoint_last.pt
+model:
+ model_cls: MMFusionJoint
+ mm_encoder_cls: MMBertForJoint
+ use_seg_emb: true
+eval:
+ save_path: runs/mtm/vlm/youcook/eval
+metric: RetrievalMetric
+predictor: RetrievalPredictor
diff --git a/fairseq/examples/MMPT/projects/mtm/vlm/test_youcookcap.yaml b/fairseq/examples/MMPT/projects/mtm/vlm/test_youcookcap.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..b2595d7c3c633d0faac5e4891457067659f7fafc
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/mtm/vlm/test_youcookcap.yaml
@@ -0,0 +1,32 @@
+slurm_config: big
+task_type: local_predict
+dataset:
+ split: test
+ video_processor: YoucookVideoProcessor
+ aligner: DSNLGAligner
+ bert_name: bert-base-uncased
+ meta_processor: YoucookNLGMetaProcessor
+ test_path: data/youcook/val_list.txt
+ trainval_annotation: data/youcook/youcookii_annotations_trainval.json
+ vfeat_dir: data/feat/feat_youcook_s3d
+ text_processor: NLGTextProcessor
+ max_video_len: 32
+ max_len: 96
+fairseq:
+ dataset:
+ batch_size: 256
+ valid_subset: test
+ num_workers: 2
+ common_eval:
+ path: runs/mtm/vlm/youcookcap/checkpoint_best.pt
+model:
+ model_cls: MMFusionNLG
+ mm_encoder_cls: MMBertForNLG
+ max_decode_length: 24
+ use_seg_emb: true
+eval:
+ save_path: runs/mtm/vlm/youcookcap/eval
+metric: NLGMetric
+predictor: NLGPredictor
+gen_param:
+ num_beams: 5
diff --git a/fairseq/examples/MMPT/projects/mtm/vlm/vtt.yaml b/fairseq/examples/MMPT/projects/mtm/vlm/vtt.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..c6c5b1ab40cbd2256919b42ea6b6a5fa52fe41a1
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/mtm/vlm/vtt.yaml
@@ -0,0 +1,49 @@
+dataset:
+ video_processor: VideoProcessor
+ bert_name: bert-base-uncased
+ meta_processor: MSRVTTMetaProcessor
+ train_path: data/msrvtt/MSRVTT_train.csv
+ jsfusion_path: data/msrvtt/MSRVTT_JSFUSION_test.csv
+ full_test_path: data/msrvtt/MSRVTT_FULL_test.csv
+ dup: 20
+ val_path: data/msrvtt/MSRVTT_JSFUSION_test.csv
+ vfeat_dir: data/feat/feat_vtt_s3d
+ text_processor: MSRVTTTextProcessor
+ json_path: data/msrvtt/MSRVTT_data.json
+ aligner: DSAligner
+ num_iso_layer: 12
+ max_video_len: 32
+ max_len: 96
+fairseq:
+ common:
+ tensorboard_logdir: run
+ log_interval: 1000
+ fp16: true
+ dataset:
+ num_workers: 4
+ batch_size: 256
+ optimization:
+ lr:
+ - 5.0e-05
+ clip_norm: 2.0
+ optimizer: adam
+ adam_betas: (0.9, 0.98)
+ lr_scheduler: polynomial_decay
+ total_num_update: 1000000
+ warmup_updates: 122
+ weight_decay: 0.0
+ ddp_backend: no_c10d
+ max_epoch: 10
+ checkpoint:
+ restore_file: runs/mtm/vlm/checkpoint_best.pt
+ reset_optimizer: true
+ reset_dataloader: true
+ reset_meters: true
+ save_dir: runs/mtm/vlm/vtt
+task_type: sweep_small
+model:
+ model_cls: MMFusionJoint
+ mm_encoder_cls: MMBertForJoint
+ use_seg_emb: true
+loss:
+ loss_cls: T2VContraLoss
diff --git a/fairseq/examples/MMPT/projects/mtm/vlm/vttqa.yaml b/fairseq/examples/MMPT/projects/mtm/vlm/vttqa.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..0a440c7dd2f5b324d57367456d22ca5aab8f397f
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/mtm/vlm/vttqa.yaml
@@ -0,0 +1,47 @@
+dataset:
+ video_processor: VideoProcessor
+ bert_name: bert-base-uncased
+ meta_processor: MSRVTTMetaProcessor
+ train_path: data/msrvtt/MSRVTT_train.csv
+ dup: 20
+ val_path: data/msrvtt/MSRVTT_JSFUSION_test.csv
+ vfeat_dir: data/feat/feat_vtt_s3d
+ text_processor: MSRVTTTextProcessor
+ json_path: data/msrvtt/MSRVTT_data.json
+ aligner: DSAligner
+ num_iso_layer: 12
+ max_video_len: 32
+ max_len: 96
+fairseq:
+ common:
+ tensorboard_logdir: run
+ log_interval: 1000
+ fp16: true
+ dataset:
+ num_workers: 4
+ batch_size: 128
+ optimization:
+ lr:
+ - 5.0e-05
+ clip_norm: 2.0
+ optimizer: adam
+ adam_betas: (0.9, 0.98)
+ lr_scheduler: polynomial_decay
+ total_num_update: 1000000
+ warmup_updates: 122
+ weight_decay: 0.0
+ ddp_backend: no_c10d
+ max_epoch: 5
+ checkpoint:
+ restore_file: runs/mtm/vlm/checkpoint_best.pt
+ reset_optimizer: true
+ reset_dataloader: true
+ reset_meters: true
+ save_dir: runs/mtm/vlm/vttqa
+task_type: sweep_small
+model:
+ model_cls: MMFusionJoint
+ mm_encoder_cls: MMBertForJoint
+ use_seg_emb: true
+loss:
+ loss_cls: V2TContraLoss
diff --git a/fairseq/examples/MMPT/projects/mtm/vlm/youcook.yaml b/fairseq/examples/MMPT/projects/mtm/vlm/youcook.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..9ee82b81b8b14a9f689e73bac22f21b780640984
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/mtm/vlm/youcook.yaml
@@ -0,0 +1,47 @@
+dataset:
+ video_processor: YoucookVideoProcessor
+ bert_name: bert-base-uncased
+ meta_processor: YoucookMetaProcessor
+ train_path: data/youcook/youcook_train.pkl
+ val_path: data/youcook/youcook_val.pkl
+ trainval_annotation: data/youcook/youcookii_annotations_trainval.json
+ use_annotation_text: true
+ vfeat_dir: data/feat/feat_youcook_s3d
+ text_processor: TextProcessor
+ aligner: DSAligner
+ num_iso_layer: 12
+ max_video_len: 32
+ max_len: 96
+fairseq:
+ common:
+ tensorboard_logdir: run
+ log_interval: 1000
+ fp16: true
+ dataset:
+ num_workers: 4
+ batch_size: 128
+ optimization:
+ lr:
+ - 5.0e-05
+ clip_norm: 2.0
+ optimizer: adam
+ adam_betas: (0.9, 0.98)
+ lr_scheduler: polynomial_decay
+ total_num_update: 1000000
+ warmup_updates: 122
+ weight_decay: 0.0
+ ddp_backend: no_c10d
+ max_epoch: 10
+ checkpoint:
+ restore_file: runs/mtm/vlm/checkpoint_best.pt
+ reset_optimizer: true
+ reset_dataloader: true
+ reset_meters: true
+ save_dir: runs/mtm/vlm/youcook
+task_type: sweep_small
+model:
+ model_cls: MMFusionJoint
+ mm_encoder_cls: MMBertForJoint
+ use_seg_emb: true
+loss:
+ loss_cls: T2VContraLoss
diff --git a/fairseq/examples/MMPT/projects/mtm/vlm/youcookcap.yaml b/fairseq/examples/MMPT/projects/mtm/vlm/youcookcap.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d29dfad5cd861cc1e8f587c730938ed491d60289
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/mtm/vlm/youcookcap.yaml
@@ -0,0 +1,45 @@
+dataset:
+ video_processor: YoucookVideoProcessor
+ bert_name: bert-base-uncased
+ meta_processor: YoucookNLGMetaProcessor
+ train_path: data/youcook/train_list.txt
+ val_path: data/youcook/val_list.txt
+ trainval_annotation: data/youcook/youcookii_annotations_trainval.json
+ vfeat_dir: data/feat/feat_youcook_s3d
+ text_processor: NLGTextProcessor
+ aligner: DSNLGAligner
+ max_video_len: 32
+ max_len: 96
+fairseq:
+ common:
+ tensorboard_logdir: run
+ log_interval: 1000
+ fp16: true
+ dataset:
+ num_workers: 4
+ batch_size: 128
+ optimization:
+ lr:
+ - 5.0e-05
+ clip_norm: 2.0
+ optimizer: adam
+ adam_betas: (0.9, 0.98)
+ lr_scheduler: polynomial_decay
+ total_num_update: 1000000
+ warmup_updates: 122
+ weight_decay: 0.0
+ ddp_backend: no_c10d
+ max_epoch: 10
+ checkpoint:
+ restore_file: runs/mtm/vlm/checkpoint_best.pt
+ reset_optimizer: true
+ reset_dataloader: true
+ reset_meters: true
+ save_dir: runs/mtm/vlm/youcookcap
+task_type: sweep_small
+model:
+ model_cls: MMFusionNLG
+ mm_encoder_cls: MMBertForNLG
+ use_seg_emb: true
+loss:
+ loss_cls: NLGLoss
diff --git a/fairseq/examples/MMPT/projects/retri/videoclip.yaml b/fairseq/examples/MMPT/projects/retri/videoclip.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..afd040ab050554aa0ca9c65ab4cf36cf6f3155dd
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/retri/videoclip.yaml
@@ -0,0 +1,10 @@
+includes: projects/retri/videoretri.yaml
+project_dir: retri/videoclip
+task_group:
+ pretrain:
+ model:
+ model_cls: MMFusionSeparate
+ mm_encoder_cls:
+ video_encoder_cls: MMBertForEncoder
+ text_encoder_cls: BertModel
+ num_hidden_video_layers: 6
diff --git a/fairseq/examples/MMPT/projects/retri/videoclip/coin_videoclip.yaml b/fairseq/examples/MMPT/projects/retri/videoclip/coin_videoclip.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..aaed5e47f62d421e5c434a2f539437831ae387db
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/retri/videoclip/coin_videoclip.yaml
@@ -0,0 +1,49 @@
+dataset:
+ video_processor: VideoProcessor
+ bert_name: bert-base-uncased
+ meta_processor: COINActionSegmentationMetaProcessor
+ train_path: data/coin/COIN.json
+ val_path: data/coin/COIN.json
+ vfeat_dir: data/feat/feat_coin_s3d
+ text_processor: COINActionSegmentationTextProcessor
+ aligner: COINActionSegmentationAligner
+ num_iso_layer: 12
+ sliding_window: 8
+ sliding_window_size: 32
+ max_video_len: 32
+ max_len: 96
+fairseq:
+ common:
+ tensorboard_logdir: run
+ log_interval: 1000
+ fp16: true
+ dataset:
+ num_workers: 4
+ batch_size: 1
+ optimization:
+ lr:
+ - 5.0e-05
+ clip_norm: 2.0
+ optimizer: adam
+ adam_betas: (0.9, 0.98)
+ lr_scheduler: polynomial_decay
+ total_num_update: 1000000
+ warmup_updates: 122
+ weight_decay: 0.0
+ ddp_backend: no_c10d
+ max_epoch: 8
+ checkpoint:
+ restore_file: runs/retri/videoclip/checkpoint_best.pt
+ reset_optimizer: true
+ reset_dataloader: true
+ reset_meters: true
+ save_dir: runs/retri/videoclip/coin
+task_type: sweep_big
+model:
+ model_cls: MMFusionSeparateActionSegmentation
+ mm_encoder_cls: null
+ video_encoder_cls: MMBertForTokenClassification
+ text_encoder_cls: BertModel
+ num_hidden_video_layers: 6
+loss:
+ loss_cls: CrossEntropy
diff --git a/fairseq/examples/MMPT/projects/retri/videoclip/crosstask_videoclip.yaml b/fairseq/examples/MMPT/projects/retri/videoclip/crosstask_videoclip.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..758601e3593aa3960fee996dcfdfd71972f9f068
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/retri/videoclip/crosstask_videoclip.yaml
@@ -0,0 +1,55 @@
+dataset:
+ video_processor: CrossTaskVideoProcessor
+ bert_name: bert-base-uncased
+ meta_processor: CrossTaskMetaProcessor
+ train_path: data/crosstask/crosstask_release/videos.csv
+ train_csv_path: data/crosstask/crosstask_release/videos.csv
+ val_path: data/crosstask/crosstask_release/videos_val.csv
+ val_csv_path: data/crosstask/crosstask_release/videos_val.csv
+ primary_path: data/crosstask/crosstask_release/tasks_primary.txt
+ related_path: data/crosstask/crosstask_release/tasks_related.txt
+ vfeat_dir: data/feat/feat_crosstask_s3d
+ annotation_path: data/crosstask/crosstask_release/annotations
+ n_train: 30
+ text_processor: CrossTaskTextProcessor
+ aligner: CrossTaskAligner
+ num_iso_layer: 12
+ sliding_window: 16
+ sliding_window_size: 32
+ max_video_len: 32
+ max_len: 96
+fairseq:
+ common:
+ tensorboard_logdir: run
+ log_interval: 1000
+ fp16: true
+ dataset:
+ num_workers: 4
+ batch_size: 1
+ optimization:
+ lr:
+ - 5.0e-05
+ clip_norm: 2.0
+ optimizer: adam
+ adam_betas: (0.9, 0.98)
+ lr_scheduler: polynomial_decay
+ total_num_update: 1000000
+ warmup_updates: 122
+ weight_decay: 0.0
+ ddp_backend: no_c10d
+ max_epoch: 5
+ checkpoint:
+ restore_file: runs/retri/videoclip/checkpoint_best.pt
+ reset_optimizer: true
+ reset_dataloader: true
+ reset_meters: true
+ save_dir: runs/retri/videoclip/crosstask
+task_type: sweep_small
+model:
+ model_cls: MMFusionSeparateActionLocalization
+ mm_encoder_cls: null
+ video_encoder_cls: MMBertForEncoder
+ text_encoder_cls: BertModel
+ num_hidden_video_layers: 6
+loss:
+ loss_cls: BCE
diff --git a/fairseq/examples/MMPT/projects/retri/videoclip/how2.yaml b/fairseq/examples/MMPT/projects/retri/videoclip/how2.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..b49581e878390cd8861926988a58a892b5a38606
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/retri/videoclip/how2.yaml
@@ -0,0 +1,65 @@
+dataset:
+ video_processor: ShardedVideoRetriVideoProcessor
+ bert_name: bert-base-uncased
+ meta_processor: ShardedHow2VideoRetriMetaProcessor
+ train_path: data/how2/how2_s3d_train.lst
+ val_path: data/how2/how2_s3d_val.lst
+ vfeat_dir: data/feat/feat_how2_s3d_shard_small
+ text_processor: ShardedVideoRetriTextProcessor
+ tfeat_dir: data/feat/feat_how2_s3d_shard_small/raw_caption_dedup.bert-base-uncased.
+ aligner: VideoRetriOverlappedAligner
+ subsampling: 1
+ sampled_min_len: 8
+ sampled_max_len: 64
+ max_video_len: 32
+ max_len: 96
+ lazy_vfeat_mask: true
+ mfm_probability: 0.15
+ mlm_probability: 0.15
+ mm_prob: 0.5
+ sampled_video_min_len: 3
+ sampled_video_max_len: 32
+ num_video_per_batch: 32
+ clip_per_video: 16
+fairseq:
+ common:
+ tensorboard_logdir: run
+ log_interval: 1000
+ fp16: true
+ dataset:
+ num_workers: 4
+ batch_size: 1
+ optimization:
+ lr:
+ - 5.0e-05
+ clip_norm: 2.0
+ optimizer: adam
+ adam_betas: (0.9, 0.98)
+ lr_scheduler: polynomial_decay
+ total_num_update: 1000000
+ warmup_updates: 1000
+ weight_decay: 0.0
+ ddp_backend: no_c10d
+ max_epoch: 25
+ checkpoint:
+ save_dir: runs/retri/videoclip
+ save_interval_updates: 1024
+ keep_interval_updates: 2
+ keep_last_epochs: 30
+task_type: sweep_big
+slurm_config: big
+eval:
+ save_path: runs/retri/videoclip
+model:
+ model_cls: MMFusionSeparate
+ mm_encoder_cls: null
+ video_encoder_cls: MMBertForEncoder
+ text_encoder_cls: BertModel
+ num_hidden_video_layers: 6
+loss:
+ loss_cls: MMContraLoss
+task: VideoRetriTask
+retri_epoch: 1
+vectorpool_cls: VideoVectorPool
+retriever_cls: VectorRetriever
+num_cands: 64
diff --git a/fairseq/examples/MMPT/projects/retri/videoclip/test_coin_videoclip.yaml b/fairseq/examples/MMPT/projects/retri/videoclip/test_coin_videoclip.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..409906203c22d670c361d011d1035874961d4b5b
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/retri/videoclip/test_coin_videoclip.yaml
@@ -0,0 +1,33 @@
+slurm_config: big
+task_type: local_predict
+dataset:
+ split: test
+ video_processor: VideoProcessor
+ aligner: COINActionSegmentationAligner
+ bert_name: bert-base-uncased
+ test_path: data/coin/COIN.json
+ meta_processor: COINActionSegmentationMetaProcessor
+ vfeat_dir: data/feat/feat_coin_s3d
+ text_processor: COINActionSegmentationTextProcessor
+ num_iso_layer: 12
+ sliding_window: 16
+ sliding_window_size: 32
+ max_video_len: 32
+ max_len: 96
+fairseq:
+ dataset:
+ batch_size: 1
+ valid_subset: test
+ num_workers: 2
+ common_eval:
+ path: runs/retri/videoclip/coin/checkpoint_best.pt
+model:
+ model_cls: MMFusionSeparateActionSegmentation
+ mm_encoder_cls: null
+ video_encoder_cls: MMBertForTokenClassification
+ text_encoder_cls: BertModel
+ num_hidden_video_layers: 6
+eval:
+ save_path: runs/retri/videoclip/coin/eval
+metric: COINActionSegmentationMetric
+predictor: COINPredictor
diff --git a/fairseq/examples/MMPT/projects/retri/videoclip/test_coin_zs.yaml b/fairseq/examples/MMPT/projects/retri/videoclip/test_coin_zs.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..b33739c7b6d7c4d935c0e4d78354a02f79e21d91
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/retri/videoclip/test_coin_zs.yaml
@@ -0,0 +1,33 @@
+slurm_config: big
+task_type: local_predict
+dataset:
+ split: test
+ video_processor: VideoProcessor
+ aligner: COINActionSegmentationAligner
+ bert_name: bert-base-uncased
+ test_path: data/coin/COIN.json
+ meta_processor: COINActionSegmentationMetaProcessor
+ vfeat_dir: data/feat/feat_coin_s3d
+ text_processor: COINActionSegmentationTextProcessor
+ num_iso_layer: 12
+ sliding_window: 16
+ sliding_window_size: 32
+ max_video_len: 32
+ max_len: 96
+fairseq:
+ dataset:
+ batch_size: 1
+ valid_subset: test
+ num_workers: 2
+ common_eval:
+ path: runs/retri/videoclip/checkpoint_best.pt
+model:
+ model_cls: MMFusionSeparate
+ mm_encoder_cls: null
+ video_encoder_cls: MMBertForEncoder
+ text_encoder_cls: BertModel
+ num_hidden_video_layers: 6
+eval:
+ save_path: runs/retri/videoclip/coin_zs/eval
+metric: COINActionSegmentationMetric
+predictor: COINZSPredictor
diff --git a/fairseq/examples/MMPT/projects/retri/videoclip/test_crosstask_videoclip.yaml b/fairseq/examples/MMPT/projects/retri/videoclip/test_crosstask_videoclip.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..e82f54fbe52defa233f1b561415737924dd6ba9c
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/retri/videoclip/test_crosstask_videoclip.yaml
@@ -0,0 +1,40 @@
+slurm_config: big
+task_type: local_predict
+dataset:
+ split: test
+ video_processor: CrossTaskVideoProcessor
+ aligner: CrossTaskAligner
+ bert_name: bert-base-uncased
+ meta_processor: CrossTaskMetaProcessor
+ test_path: data/crosstask/crosstask_release/videos_val.csv
+ train_csv_path: data/crosstask/crosstask_release/videos.csv
+ val_path: data/crosstask/crosstask_release/videos_val.csv
+ val_csv_path: data/crosstask/crosstask_release/videos_val.csv
+ primary_path: data/crosstask/crosstask_release/tasks_primary.txt
+ related_path: data/crosstask/crosstask_release/tasks_related.txt
+ vfeat_dir: data/feat/feat_crosstask_s3d
+ annotation_path: data/crosstask/crosstask_release/annotations
+ n_train: 30
+ text_processor: CrossTaskTextProcessor
+ num_iso_layer: 12
+ sliding_window: 16
+ sliding_window_size: 32
+ max_video_len: 32
+ max_len: 96
+fairseq:
+ dataset:
+ batch_size: 1
+ valid_subset: test
+ num_workers: 2
+ common_eval:
+ path: runs/retri/videoclip/crosstask/checkpoint_best.pt
+model:
+ model_cls: MMFusionSeparateActionLocalization
+ mm_encoder_cls: null
+ video_encoder_cls: MMBertForEncoder
+ text_encoder_cls: BertModel
+ num_hidden_video_layers: 6
+eval:
+ save_path: runs/retri/videoclip/crosstask/eval
+metric: CrossTaskMetric
+predictor: CrossTaskPredictor
diff --git a/fairseq/examples/MMPT/projects/retri/videoclip/test_crosstask_zs_videoclip.yaml b/fairseq/examples/MMPT/projects/retri/videoclip/test_crosstask_zs_videoclip.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..6fc357cc1f5285229a508a95dc31c2aae17c977d
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/retri/videoclip/test_crosstask_zs_videoclip.yaml
@@ -0,0 +1,40 @@
+slurm_config: big
+task_type: local_predict
+dataset:
+ split: test
+ video_processor: CrossTaskVideoProcessor
+ aligner: CrossTaskAligner
+ bert_name: bert-base-uncased
+ meta_processor: CrossTaskMetaProcessor
+ test_path: data/crosstask/crosstask_release/videos_val.csv
+ train_csv_path: data/crosstask/crosstask_release/videos.csv
+ val_path: data/crosstask/crosstask_release/videos_val.csv
+ val_csv_path: data/crosstask/crosstask_release/videos_val.csv
+ primary_path: data/crosstask/crosstask_release/tasks_primary.txt
+ related_path: data/crosstask/crosstask_release/tasks_related.txt
+ vfeat_dir: data/feat/feat_crosstask_s3d
+ annotation_path: data/crosstask/crosstask_release/annotations
+ n_train: 30
+ text_processor: CrossTaskTextProcessor
+ num_iso_layer: 12
+ sliding_window: 16
+ sliding_window_size: 32
+ max_video_len: 32
+ max_len: 96
+fairseq:
+ dataset:
+ batch_size: 1
+ valid_subset: test
+ num_workers: 2
+ common_eval:
+ path: runs/retri/videoclip/checkpoint_best.pt
+model:
+ model_cls: MMFusionSeparateActionLocalization
+ mm_encoder_cls: null
+ video_encoder_cls: MMBertForEncoder
+ text_encoder_cls: BertModel
+ num_hidden_video_layers: 6
+eval:
+ save_path: runs/retri/videoclip/crosstask_zs/eval
+metric: CrossTaskMetric
+predictor: CrossTaskPredictor
diff --git a/fairseq/examples/MMPT/projects/retri/videoclip/test_didemo_zs.yaml b/fairseq/examples/MMPT/projects/retri/videoclip/test_didemo_zs.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..8dc716815d8d4812237c814fc17a90dd7ac3cf1e
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/retri/videoclip/test_didemo_zs.yaml
@@ -0,0 +1,31 @@
+slurm_config: big
+task_type: local_predict
+dataset:
+ split: test
+ video_processor: VideoProcessor
+ aligner: DiDeMoAligner
+ bert_name: bert-base-uncased
+ meta_processor: DiDeMoMetaProcessor
+ test_path: data/didemo/test_data.json
+ vfeat_dir: data/feat/feat_didemo_s3d
+ text_processor: DiDeMoTextProcessor
+ num_iso_layer: 12
+ max_video_len: 32
+ max_len: 96
+fairseq:
+ dataset:
+ batch_size: 256
+ valid_subset: test
+ num_workers: 2
+ common_eval:
+ path: runs/retri/videoclip/checkpoint_best.pt
+model:
+ model_cls: MMFusionSeparate
+ mm_encoder_cls: null
+ video_encoder_cls: MMBertForEncoder
+ text_encoder_cls: BertModel
+ num_hidden_video_layers: 6
+eval:
+ save_path: runs/retri/videoclip/didemo_zs/eval
+metric: DiDeMoMetric
+predictor: DiDeMoPredictor
diff --git a/fairseq/examples/MMPT/projects/retri/videoclip/test_vtt_videoclip.yaml b/fairseq/examples/MMPT/projects/retri/videoclip/test_vtt_videoclip.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..19321ad5f489121381b1b78506b38ba3705b37e4
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/retri/videoclip/test_vtt_videoclip.yaml
@@ -0,0 +1,31 @@
+slurm_config: big
+task_type: local_predict
+dataset:
+ split: test
+ video_processor: VideoProcessor
+ aligner: DSAligner
+ bert_name: bert-base-uncased
+ meta_processor: MSRVTTMetaProcessor
+ test_path: data/msrvtt/MSRVTT_JSFUSION_test.csv
+ vfeat_dir: data/feat/feat_vtt_s3d
+ text_processor: MSRVTTTextProcessor
+ num_iso_layer: 12
+ max_video_len: 32
+ max_len: 96
+fairseq:
+ dataset:
+ batch_size: 256
+ valid_subset: test
+ num_workers: 2
+ common_eval:
+ path: runs/retri/videoclip/vtt/checkpoint_last.pt
+model:
+ model_cls: MMFusionSeparate
+ mm_encoder_cls: null
+ video_encoder_cls: MMBertForEncoder
+ text_encoder_cls: BertModel
+ num_hidden_video_layers: 6
+eval:
+ save_path: runs/retri/videoclip/vtt/eval
+metric: RetrievalMetric
+predictor: RetrievalPredictor
diff --git a/fairseq/examples/MMPT/projects/retri/videoclip/test_vtt_zs.yaml b/fairseq/examples/MMPT/projects/retri/videoclip/test_vtt_zs.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d149fa3960294f4d9dcac9c4f4f18631eb622435
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/retri/videoclip/test_vtt_zs.yaml
@@ -0,0 +1,31 @@
+slurm_config: big
+task_type: local_predict
+dataset:
+ split: test
+ video_processor: VideoProcessor
+ aligner: DSAligner
+ bert_name: bert-base-uncased
+ meta_processor: MSRVTTMetaProcessor
+ test_path: data/msrvtt/MSRVTT_JSFUSION_test.csv
+ vfeat_dir: data/feat/feat_vtt_s3d
+ text_processor: MSRVTTTextProcessor
+ num_iso_layer: 12
+ max_video_len: 32
+ max_len: 96
+fairseq:
+ dataset:
+ batch_size: 256
+ valid_subset: test
+ num_workers: 2
+ common_eval:
+ path: runs/retri/videoclip/checkpoint_best.pt
+model:
+ model_cls: MMFusionSeparate
+ mm_encoder_cls: null
+ video_encoder_cls: MMBertForEncoder
+ text_encoder_cls: BertModel
+ num_hidden_video_layers: 6
+eval:
+ save_path: runs/retri/videoclip/vtt_zs/eval
+metric: RetrievalMetric
+predictor: RetrievalPredictor
diff --git a/fairseq/examples/MMPT/projects/retri/videoclip/test_vttqa_videoclip.yaml b/fairseq/examples/MMPT/projects/retri/videoclip/test_vttqa_videoclip.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..295aeedbb0f29a28ad9065da0a18e82ac98b6683
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/retri/videoclip/test_vttqa_videoclip.yaml
@@ -0,0 +1,31 @@
+slurm_config: big
+task_type: local_predict
+dataset:
+ split: test
+ video_processor: VideoProcessor
+ aligner: MSRVTTQAAligner
+ bert_name: bert-base-uncased
+ meta_processor: MSRVTTQAMetaProcessor
+ test_path: data/msrvtt-qa/MSR_MC_test.csv
+ vfeat_dir: data/feat/feat_vtt_s3d
+ text_processor: MSRVTTQATextProcessor
+ num_iso_layer: 12
+ max_video_len: 32
+ max_len: 96
+fairseq:
+ dataset:
+ batch_size: 256
+ valid_subset: test
+ num_workers: 2
+ common_eval:
+ path: runs/retri/videoclip/vttqa/checkpoint_last.pt
+model:
+ model_cls: MMFusionSeparate
+ mm_encoder_cls: null
+ video_encoder_cls: MMBertForEncoder
+ text_encoder_cls: BertModel
+ num_hidden_video_layers: 6
+eval:
+ save_path: runs/retri/videoclip/vttqa/eval
+metric: QAMetric
+predictor: QAPredictor
diff --git a/fairseq/examples/MMPT/projects/retri/videoclip/test_vttqa_zs.yaml b/fairseq/examples/MMPT/projects/retri/videoclip/test_vttqa_zs.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..7a876c822ae38cadc03c69544e64025d516e8cca
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/retri/videoclip/test_vttqa_zs.yaml
@@ -0,0 +1,31 @@
+slurm_config: big
+task_type: local_predict
+dataset:
+ split: test
+ video_processor: VideoProcessor
+ aligner: MSRVTTQAAligner
+ bert_name: bert-base-uncased
+ meta_processor: MSRVTTQAMetaProcessor
+ test_path: data/msrvtt-qa/MSR_MC_test.csv
+ vfeat_dir: data/feat/feat_vtt_s3d
+ text_processor: MSRVTTQATextProcessor
+ num_iso_layer: 12
+ max_video_len: 32
+ max_len: 96
+fairseq:
+ dataset:
+ batch_size: 256
+ valid_subset: test
+ num_workers: 2
+ common_eval:
+ path: runs/retri/videoclip/checkpoint_best.pt
+model:
+ model_cls: MMFusionSeparate
+ mm_encoder_cls: null
+ video_encoder_cls: MMBertForEncoder
+ text_encoder_cls: BertModel
+ num_hidden_video_layers: 6
+eval:
+ save_path: runs/retri/videoclip/vttqa_zs/eval
+metric: QAMetric
+predictor: QAPredictor
diff --git a/fairseq/examples/MMPT/projects/retri/videoclip/test_youcook_videoclip.yaml b/fairseq/examples/MMPT/projects/retri/videoclip/test_youcook_videoclip.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..86a4ab203e097ec156ce02d88ce9c1b2374cec9f
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/retri/videoclip/test_youcook_videoclip.yaml
@@ -0,0 +1,33 @@
+slurm_config: big
+task_type: local_predict
+dataset:
+ split: test
+ video_processor: YoucookVideoProcessor
+ aligner: DSAligner
+ bert_name: bert-base-uncased
+ meta_processor: YoucookMetaProcessor
+ test_path: data/youcook/youcook_val.pkl
+ trainval_annotation: data/youcook/youcookii_annotations_trainval.json
+ use_annotation_text: true
+ vfeat_dir: data/feat/feat_youcook_s3d
+ text_processor: TextProcessor
+ num_iso_layer: 12
+ max_video_len: 32
+ max_len: 96
+fairseq:
+ dataset:
+ batch_size: 256
+ valid_subset: test
+ num_workers: 2
+ common_eval:
+ path: runs/retri/videoclip/youcook/checkpoint_last.pt
+model:
+ model_cls: MMFusionSeparate
+ mm_encoder_cls: null
+ video_encoder_cls: MMBertForEncoder
+ text_encoder_cls: BertModel
+ num_hidden_video_layers: 6
+eval:
+ save_path: runs/retri/videoclip/youcook/eval
+metric: RetrievalMetric
+predictor: RetrievalPredictor
diff --git a/fairseq/examples/MMPT/projects/retri/videoclip/test_youcook_zs.yaml b/fairseq/examples/MMPT/projects/retri/videoclip/test_youcook_zs.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fd2941708be0baacf9e12ca757e0de40676e570d
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/retri/videoclip/test_youcook_zs.yaml
@@ -0,0 +1,33 @@
+slurm_config: big
+task_type: local_predict
+dataset:
+ split: test
+ video_processor: YoucookVideoProcessor
+ aligner: DSAligner
+ bert_name: bert-base-uncased
+ meta_processor: YoucookMetaProcessor
+ test_path: data/youcook/youcook_val.pkl
+ trainval_annotation: data/youcook/youcookii_annotations_trainval.json
+ use_annotation_text: true
+ vfeat_dir: data/feat/feat_youcook_s3d
+ text_processor: TextProcessor
+ num_iso_layer: 12
+ max_video_len: 32
+ max_len: 96
+fairseq:
+ dataset:
+ batch_size: 256
+ valid_subset: test
+ num_workers: 2
+ common_eval:
+ path: runs/retri/videoclip/checkpoint_best.pt
+model:
+ model_cls: MMFusionSeparate
+ mm_encoder_cls: null
+ video_encoder_cls: MMBertForEncoder
+ text_encoder_cls: BertModel
+ num_hidden_video_layers: 6
+eval:
+ save_path: runs/retri/videoclip/youcook_zs/eval
+metric: RetrievalMetric
+predictor: RetrievalPredictor
diff --git a/fairseq/examples/MMPT/projects/retri/videoclip/vtt_videoclip.yaml b/fairseq/examples/MMPT/projects/retri/videoclip/vtt_videoclip.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d8b4079ac233af17ee16cfed7f3e2316e0e236c5
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/retri/videoclip/vtt_videoclip.yaml
@@ -0,0 +1,51 @@
+dataset:
+ video_processor: VideoProcessor
+ bert_name: bert-base-uncased
+ meta_processor: MSRVTTMetaProcessor
+ train_path: data/msrvtt/MSRVTT_train.csv
+ jsfusion_path: data/msrvtt/MSRVTT_JSFUSION_test.csv
+ full_test_path: data/msrvtt/MSRVTT_FULL_test.csv
+ dup: 20
+ val_path: data/msrvtt/MSRVTT_JSFUSION_test.csv
+ vfeat_dir: data/feat/feat_vtt_s3d
+ text_processor: MSRVTTTextProcessor
+ json_path: data/msrvtt/MSRVTT_data.json
+ aligner: DSAligner
+ num_iso_layer: 12
+ max_video_len: 32
+ max_len: 96
+fairseq:
+ common:
+ tensorboard_logdir: run
+ log_interval: 1000
+ fp16: true
+ dataset:
+ num_workers: 4
+ batch_size: 224
+ optimization:
+ lr:
+ - 5.0e-05
+ clip_norm: 2.0
+ optimizer: adam
+ adam_betas: (0.9, 0.98)
+ lr_scheduler: polynomial_decay
+ total_num_update: 1000000
+ warmup_updates: 122
+ weight_decay: 0.0
+ ddp_backend: no_c10d
+ max_epoch: 10
+ checkpoint:
+ restore_file: runs/retri/videoclip/checkpoint_best.pt
+ reset_optimizer: true
+ reset_dataloader: true
+ reset_meters: true
+ save_dir: runs/retri/videoclip/vtt
+task_type: sweep_small
+model:
+ model_cls: MMFusionSeparate
+ mm_encoder_cls: null
+ video_encoder_cls: MMBertForEncoder
+ text_encoder_cls: BertModel
+ num_hidden_video_layers: 6
+loss:
+ loss_cls: T2VContraLoss
diff --git a/fairseq/examples/MMPT/projects/retri/videoclip/vttqa_videoclip.yaml b/fairseq/examples/MMPT/projects/retri/videoclip/vttqa_videoclip.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..f0566d784ab53e0159a1db6b82222ad11ad12add
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/retri/videoclip/vttqa_videoclip.yaml
@@ -0,0 +1,49 @@
+dataset:
+ video_processor: VideoProcessor
+ bert_name: bert-base-uncased
+ meta_processor: MSRVTTMetaProcessor
+ train_path: data/msrvtt/MSRVTT_train.csv
+ dup: 20
+ val_path: data/msrvtt/MSRVTT_JSFUSION_test.csv
+ vfeat_dir: data/feat/feat_vtt_s3d
+ text_processor: MSRVTTTextProcessor
+ json_path: data/msrvtt/MSRVTT_data.json
+ aligner: DSAligner
+ num_iso_layer: 12
+ max_video_len: 32
+ max_len: 96
+fairseq:
+ common:
+ tensorboard_logdir: run
+ log_interval: 1000
+ fp16: true
+ dataset:
+ num_workers: 4
+ batch_size: 128
+ optimization:
+ lr:
+ - 5.0e-05
+ clip_norm: 2.0
+ optimizer: adam
+ adam_betas: (0.9, 0.98)
+ lr_scheduler: polynomial_decay
+ total_num_update: 1000000
+ warmup_updates: 122
+ weight_decay: 0.0
+ ddp_backend: no_c10d
+ max_epoch: 5
+ checkpoint:
+ restore_file: runs/retri/videoclip/checkpoint_best.pt
+ reset_optimizer: true
+ reset_dataloader: true
+ reset_meters: true
+ save_dir: runs/retri/videoclip/vttqa
+task_type: sweep_small
+model:
+ model_cls: MMFusionSeparate
+ mm_encoder_cls: null
+ video_encoder_cls: MMBertForEncoder
+ text_encoder_cls: BertModel
+ num_hidden_video_layers: 6
+loss:
+ loss_cls: V2TContraLoss
diff --git a/fairseq/examples/MMPT/projects/retri/videoclip/youcook_videoclip.yaml b/fairseq/examples/MMPT/projects/retri/videoclip/youcook_videoclip.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..c2b13e5519fe0839c6b3bc681f7ee97b522077d8
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/retri/videoclip/youcook_videoclip.yaml
@@ -0,0 +1,49 @@
+dataset:
+ video_processor: YoucookVideoProcessor
+ bert_name: bert-base-uncased
+ meta_processor: YoucookMetaProcessor
+ train_path: data/youcook/youcook_train.pkl
+ val_path: data/youcook/youcook_val.pkl
+ trainval_annotation: data/youcook/youcookii_annotations_trainval.json
+ use_annotation_text: true
+ vfeat_dir: data/feat/feat_youcook_s3d
+ text_processor: TextProcessor
+ aligner: DSAligner
+ num_iso_layer: 12
+ max_video_len: 32
+ max_len: 96
+fairseq:
+ common:
+ tensorboard_logdir: run
+ log_interval: 1000
+ fp16: true
+ dataset:
+ num_workers: 4
+ batch_size: 128
+ optimization:
+ lr:
+ - 5.0e-05
+ clip_norm: 2.0
+ optimizer: adam
+ adam_betas: (0.9, 0.98)
+ lr_scheduler: polynomial_decay
+ total_num_update: 1000000
+ warmup_updates: 122
+ weight_decay: 0.0
+ ddp_backend: no_c10d
+ max_epoch: 10
+ checkpoint:
+ restore_file: runs/retri/videoclip/checkpoint_best.pt
+ reset_optimizer: true
+ reset_dataloader: true
+ reset_meters: true
+ save_dir: runs/retri/videoclip/youcook
+task_type: sweep_small
+model:
+ model_cls: MMFusionSeparate
+ mm_encoder_cls: null
+ video_encoder_cls: MMBertForEncoder
+ text_encoder_cls: BertModel
+ num_hidden_video_layers: 6
+loss:
+ loss_cls: T2VContraLoss
diff --git a/fairseq/examples/MMPT/projects/retri/videoretri.yaml b/fairseq/examples/MMPT/projects/retri/videoretri.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..969e1fb2793d5381e52188c4ad78aea240698969
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/retri/videoretri.yaml
@@ -0,0 +1,51 @@
+includes: projects/mfmmlm.yaml
+project_dir: retri/videoretri
+run_task:
+ - how2.yaml
+task_group:
+ pretrain:
+ task: VideoRetriTask
+ retri_epoch: 1
+ vectorpool_cls: VideoVectorPool
+ retriever_cls: VectorRetriever
+ num_cands: 64
+ dataset:
+ train_path: data/how2/how2_s3d_train.lst
+ meta_processor: ShardedHow2VideoRetriMetaProcessor
+ video_processor: ShardedVideoRetriVideoProcessor
+ text_processor: ShardedVideoRetriTextProcessor
+ aligner: VideoRetriOverlappedAligner
+ sampled_video_min_len: 3
+ sampled_video_max_len: 32
+ sampled_min_len: 8
+ sampled_max_len: 64
+ num_video_per_batch: 32
+ # do not use subsampling as it changes fairseq batch_size.
+ subsampling: 1 # disable subsampling
+ clip_per_video: 16
+ fairseq:
+ dataset:
+ batch_size: 1
+ optimization:
+ max_epoch: 25
+ model:
+ model_cls: MMFusionShare
+ mm_encoder_cls: MMBertForEncoder
+ loss:
+ loss_cls: MMContraLoss
+ finetune:
+ task_list: [vtt_videoclip.yaml, youcook_videoclip.yaml, vttqa_videoclip.yaml, crosstask_videoclip.yaml, coin_videoclip.yaml]
+ test:
+ task_list:
+ - test_youcook_zs.yaml
+ - test_vtt_zs.yaml
+ - test_vttqa_zs.yaml
+ - test_crosstask_zs_videoclip.yaml
+ - test_coin_zs.yaml
+ - test_didemo_zs.yaml
+ - test_youcook_videoclip.yaml
+ - test_vtt_videoclip.yaml
+ - test_vttqa_videoclip.yaml
+ - test_crosstask_videoclip.yaml
+ - test_coin_videoclip.yaml
+
diff --git a/fairseq/examples/MMPT/projects/task/coin.yaml b/fairseq/examples/MMPT/projects/task/coin.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..e7772486e166348deff1f7c9deceb68c7e1bb443
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/task/coin.yaml
@@ -0,0 +1,25 @@
+includes: projects/task/ft.yaml
+task_type: sweep_big
+dataset:
+ meta_processor: COINActionSegmentationMetaProcessor
+ train_path: data/coin/COIN.json
+ val_path: data/coin/COIN.json
+ vfeat_dir: data/feat/feat_coin_s3d
+ video_processor: VideoProcessor
+ text_processor: COINActionSegmentationTextProcessor
+ aligner: COINActionSegmentationAligner
+ num_iso_layer: 12
+ sliding_window: 8
+ sliding_window_size: 32
+model:
+ model_cls: MMFusionActionSegmentation
+ mm_encoder_cls: MMBertForTokenClassification
+loss:
+ loss_cls: CrossEntropy
+fairseq:
+ dataset:
+ batch_size: 1
+ optimization:
+ max_epoch: 8
+ checkpoint:
+ save_dir: runs/task/coin
diff --git a/fairseq/examples/MMPT/projects/task/coin_videoclip.yaml b/fairseq/examples/MMPT/projects/task/coin_videoclip.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..69988bc18a03a8f9bf594266f3bf06aa54f11b6a
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/task/coin_videoclip.yaml
@@ -0,0 +1,7 @@
+includes: projects/task/coin.yaml
+model:
+ model_cls: MMFusionSeparateActionSegmentation
+ mm_encoder_cls:
+ video_encoder_cls: MMBertForTokenClassification
+ text_encoder_cls: BertModel # dummy, not used.
+ num_hidden_video_layers: 6
diff --git a/fairseq/examples/MMPT/projects/task/crosstask.yaml b/fairseq/examples/MMPT/projects/task/crosstask.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..cb4dbb0cb4a07020d1fbba9a810eb30fd81ae96d
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/task/crosstask.yaml
@@ -0,0 +1,31 @@
+includes: projects/task/ft.yaml
+dataset:
+ meta_processor: CrossTaskMetaProcessor
+ train_path: data/crosstask/crosstask_release/videos.csv # dummy
+ train_csv_path: data/crosstask/crosstask_release/videos.csv
+ val_path: data/crosstask/crosstask_release/videos_val.csv # dummy
+ val_csv_path: data/crosstask/crosstask_release/videos_val.csv
+ primary_path: data/crosstask/crosstask_release/tasks_primary.txt
+ related_path: data/crosstask/crosstask_release/tasks_related.txt
+ vfeat_dir: data/feat/feat_crosstask_s3d
+ annotation_path: data/crosstask/crosstask_release/annotations
+ n_train: 30
+ video_processor: CrossTaskVideoProcessor
+ text_processor: CrossTaskTextProcessor
+ aligner: CrossTaskAligner
+ num_iso_layer: 12
+ sliding_window: 16
+ sliding_window_size: 32
+model:
+ model_cls: MMFusionActionLocalization
+ mm_encoder_cls: MMBertForJoint
+loss:
+ loss_cls: BCE
+fairseq:
+ dataset:
+ batch_size: 1
+ optimization:
+ max_epoch: 5
+ checkpoint:
+ save_dir: runs/task/crosstask
+ restore_file: runs/task/checkpoint11.pt # for VLM
diff --git a/fairseq/examples/MMPT/projects/task/crosstask_videoclip.yaml b/fairseq/examples/MMPT/projects/task/crosstask_videoclip.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..6ec613c07fcacfa6e6d3c5fc78aaefcb2b33eff5
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/task/crosstask_videoclip.yaml
@@ -0,0 +1,10 @@
+includes: projects/task/crosstask.yaml
+model:
+ model_cls: MMFusionSeparateActionLocalization
+ mm_encoder_cls:
+ video_encoder_cls: MMBertForEncoder
+ text_encoder_cls: BertModel # dummy, not used.
+ num_hidden_video_layers: 6
+fairseq:
+ checkpoint:
+ restore_file: runs/task/checkpoint_best.pt # overwrite the default of VLM.
diff --git a/fairseq/examples/MMPT/projects/task/default.yaml b/fairseq/examples/MMPT/projects/task/default.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..087fef71a4b94d81b74c5e9ef647db189626ff36
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/task/default.yaml
@@ -0,0 +1,20 @@
+# this yaml cannot be run alone. you must use `how2.yaml`, `vtt.yaml` etc for training.
+dataset:
+ video_processor: VideoProcessor
+ bert_name: bert-base-uncased
+fairseq:
+ common:
+ tensorboard_logdir: run
+ log_interval: 1000
+ dataset:
+ num_workers: 4
+ optimization:
+ lr: [ 0.00005 ]
+ clip_norm: 2.0
+ optimizer: adam
+ adam_betas: (0.9, 0.98)
+ lr_scheduler: polynomial_decay
+ total_num_update: 1000000 # backward compatible on fairseq 1.0.0a0+af0389f for reproducibility.
+ warmup_updates: 1000
+ weight_decay: 0.0
+ ddp_backend: no_c10d
diff --git a/fairseq/examples/MMPT/projects/task/ft.yaml b/fairseq/examples/MMPT/projects/task/ft.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..c93b8a73ea93e00da8a0ee76751c9e1bd34ab8a2
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/task/ft.yaml
@@ -0,0 +1,13 @@
+includes: projects/task/default.yaml
+# all derived config will be run by fairseq-train.
+task_type: sweep_small
+fairseq:
+ optimization:
+ warmup_updates: 122 # copied from roberta glue: https://github.com/pytorch/fairseq/blob/master/examples/roberta/README.glue.md
+ checkpoint:
+ # save_interval_updates: 512
+ # borrowed from Roberta script.
+ restore_file: runs/task/checkpoint_best.pt
+ reset_optimizer: True
+ reset_dataloader: True
+ reset_meters: True
diff --git a/fairseq/examples/MMPT/projects/task/how2.yaml b/fairseq/examples/MMPT/projects/task/how2.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..094dd04bfcfdb96699ac80066f874fca954dc56a
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/task/how2.yaml
@@ -0,0 +1,22 @@
+includes: projects/task/default.yaml
+task_type: sweep_big
+slurm_config: big
+dataset:
+ meta_processor: ShardedHow2MetaProcessor
+ train_path: data/how2/how2_s3d_train.lst
+ val_path: data/how2/how2_s3d_val.lst
+ video_processor: ShardedVideoProcessor
+ vfeat_dir: data/feat/feat_how2_s3d_shard_small
+ text_processor: ShardedTextProcessor
+ tfeat_dir: data/feat/feat_how2_s3d_shard_small/raw_caption_dedup.bert-base-uncased.
+ aligner: FixedLenAligner
+# disable direct running of this yaml
+eval:
+ save_path: runs/task
+fairseq:
+ checkpoint:
+ save_dir: runs/task
+ save_interval_updates: 1024
+ keep_interval_updates: 2
+ keep_last_epochs: 30
+
diff --git a/fairseq/examples/MMPT/projects/task/test.yaml b/fairseq/examples/MMPT/projects/task/test.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..0a98445241a17bbfda84a3b3be0687d31e49d12f
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/task/test.yaml
@@ -0,0 +1,13 @@
+# this yaml cannot be run alone: implement a test_${dataset}.yaml
+slurm_config: big
+task_type: local_predict
+dataset:
+ split: test
+ video_processor: VideoProcessor
+ aligner: DSAligner
+ bert_name: bert-base-uncased
+fairseq:
+ dataset:
+ batch_size: 256
+ valid_subset: test
+ num_workers: 2
diff --git a/fairseq/examples/MMPT/projects/task/test_coin.yaml b/fairseq/examples/MMPT/projects/task/test_coin.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..6d919df7c2a69de33debd03a6d9b1a513da39618
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/task/test_coin.yaml
@@ -0,0 +1,24 @@
+includes: projects/task/test.yaml
+dataset:
+ split: test
+ test_path: data/coin/COIN.json
+ meta_processor: COINActionSegmentationMetaProcessor
+ vfeat_dir: data/feat/feat_coin_s3d
+ video_processor: VideoProcessor
+ text_processor: COINActionSegmentationTextProcessor
+ aligner: COINActionSegmentationAligner
+ num_iso_layer: 12
+ sliding_window: 16
+ sliding_window_size: 32
+model:
+ model_cls: MMFusionActionSegmentation
+ mm_encoder_cls: MMBertForTokenClassification
+eval:
+ save_path: runs/task/coin/eval
+fairseq:
+ dataset:
+ batch_size: 1
+ common_eval:
+ path: runs/task/coin/checkpoint_best.pt
+metric: COINActionSegmentationMetric
+predictor: COINPredictor
diff --git a/fairseq/examples/MMPT/projects/task/test_coin_videoclip.yaml b/fairseq/examples/MMPT/projects/task/test_coin_videoclip.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..b41f5bc4890405993217859c22653fa77a0dd4c3
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/task/test_coin_videoclip.yaml
@@ -0,0 +1,7 @@
+includes: projects/task/test_coin.yaml
+model:
+ model_cls: MMFusionSeparateActionSegmentation
+ mm_encoder_cls:
+ video_encoder_cls: MMBertForTokenClassification
+ text_encoder_cls: BertModel # dummy, not used.
+ num_hidden_video_layers: 6
diff --git a/fairseq/examples/MMPT/projects/task/test_coin_zs.yaml b/fairseq/examples/MMPT/projects/task/test_coin_zs.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..5d19b09f1dee1922ffddf5bf030135cb46016dc1
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/task/test_coin_zs.yaml
@@ -0,0 +1,13 @@
+includes: projects/task/test_coin.yaml
+model:
+ model_cls: MMFusionSeparate
+ mm_encoder_cls:
+ video_encoder_cls: MMBertForEncoder
+ text_encoder_cls: BertModel
+ num_hidden_video_layers: 6
+eval:
+ save_path: runs/task/coin_zs/eval
+fairseq:
+ common_eval:
+ path: runs/task/checkpoint_best.pt
+predictor: COINZSPredictor
diff --git a/fairseq/examples/MMPT/projects/task/test_crosstask.yaml b/fairseq/examples/MMPT/projects/task/test_crosstask.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..6dd778e30be543a51997f16e326aa29f6e49e973
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/task/test_crosstask.yaml
@@ -0,0 +1,32 @@
+includes: projects/task/test.yaml
+dataset:
+ split: test
+ meta_processor: CrossTaskMetaProcessor
+ test_path: data/crosstask/crosstask_release/videos_val.csv
+ train_csv_path: data/crosstask/crosstask_release/videos.csv
+ val_path: data/crosstask/crosstask_release/videos_val.csv # dummy
+ val_csv_path: data/crosstask/crosstask_release/videos_val.csv
+ primary_path: data/crosstask/crosstask_release/tasks_primary.txt
+ related_path: data/crosstask/crosstask_release/tasks_related.txt
+ vfeat_dir: data/feat/feat_crosstask_s3d
+ annotation_path: data/crosstask/crosstask_release/annotations
+ n_train: 30
+ video_processor: CrossTaskVideoProcessor
+ text_processor: CrossTaskTextProcessor
+ aligner: CrossTaskAligner
+ num_iso_layer: 12
+ sliding_window: 16
+ sliding_window_size: 32
+model:
+ model_cls: MMFusionActionLocalization
+ mm_encoder_cls: MMBertForJoint
+eval:
+ save_path: runs/task/crosstask/eval
+fairseq:
+ # read code and find what is the checkpoint arg.
+ dataset:
+ batch_size: 1
+ common_eval:
+ path: runs/task/crosstask/checkpoint_best.pt
+metric: CrossTaskMetric
+predictor: CrossTaskPredictor
diff --git a/fairseq/examples/MMPT/projects/task/test_crosstask_videoclip.yaml b/fairseq/examples/MMPT/projects/task/test_crosstask_videoclip.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..df12535d231ffe6d1759f77f161fe48d0833f4e4
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/task/test_crosstask_videoclip.yaml
@@ -0,0 +1,7 @@
+includes: projects/task/test_crosstask.yaml
+model:
+ model_cls: MMFusionSeparateActionLocalization
+ mm_encoder_cls:
+ video_encoder_cls: MMBertForEncoder
+ text_encoder_cls: BertModel # dummy, not used.
+ num_hidden_video_layers: 6
diff --git a/fairseq/examples/MMPT/projects/task/test_crosstask_zs.yaml b/fairseq/examples/MMPT/projects/task/test_crosstask_zs.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..19386e495b37ab2570cc1d763277193b28309814
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/task/test_crosstask_zs.yaml
@@ -0,0 +1,32 @@
+includes: projects/task/test.yaml
+dataset:
+ split: test
+ meta_processor: CrossTaskMetaProcessor
+ test_path: data/crosstask/crosstask_release/videos_val.csv
+ train_csv_path: data/crosstask/crosstask_release/videos.csv
+ val_path: data/crosstask/crosstask_release/videos_val.csv # dummy
+ val_csv_path: data/crosstask/crosstask_release/videos_val.csv
+ primary_path: data/crosstask/crosstask_release/tasks_primary.txt
+ related_path: data/crosstask/crosstask_release/tasks_related.txt
+ vfeat_dir: data/feat/feat_crosstask_s3d
+ annotation_path: data/crosstask/crosstask_release/annotations
+ n_train: 30
+ video_processor: CrossTaskVideoProcessor
+ text_processor: CrossTaskTextProcessor
+ aligner: CrossTaskAligner
+ num_iso_layer: 12
+ sliding_window: 16
+ sliding_window_size: 32
+model:
+ model_cls: MMFusionActionLocalization
+ mm_encoder_cls: MMBertForJoint
+eval:
+ save_path: runs/task/crosstask_zs/eval
+fairseq:
+ # read code and find what is the checkpoint arg.
+ dataset:
+ batch_size: 1
+ common_eval:
+ path: runs/task/checkpoint_best.pt # load the best from how2 on ACL submission: runs/task/checkpoint11.pt
+metric: CrossTaskMetric
+predictor: CrossTaskPredictor
diff --git a/fairseq/examples/MMPT/projects/task/test_crosstask_zs_videoclip.yaml b/fairseq/examples/MMPT/projects/task/test_crosstask_zs_videoclip.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..7f0198276f74f592413e99fb10672431543a8f67
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/task/test_crosstask_zs_videoclip.yaml
@@ -0,0 +1,7 @@
+includes: projects/task/test_crosstask_zs.yaml
+model:
+ model_cls: MMFusionSeparateActionLocalization
+ mm_encoder_cls:
+ video_encoder_cls: MMBertForEncoder
+ text_encoder_cls: BertModel # dummy, not used.
+ num_hidden_video_layers: 6
diff --git a/fairseq/examples/MMPT/projects/task/test_didemo_zs.yaml b/fairseq/examples/MMPT/projects/task/test_didemo_zs.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..4b53dca71e4a0e77793483ee971379f0476366f3
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/task/test_didemo_zs.yaml
@@ -0,0 +1,23 @@
+includes: projects/task/test.yaml
+dataset:
+ meta_processor: DiDeMoMetaProcessor
+ test_path: data/didemo/test_data.json
+ video_processor: VideoProcessor
+ vfeat_dir: data/feat/feat_didemo_s3d
+ text_processor: DiDeMoTextProcessor
+ aligner: DiDeMoAligner
+ num_iso_layer: 12
+model:
+ model_cls: MMFusionSeparate
+ mm_encoder_cls:
+ video_encoder_cls: MMBertForEncoder
+ text_encoder_cls: BertModel
+ num_hidden_video_layers: 6
+eval:
+ save_path: runs/task/didemo_zs/eval
+fairseq:
+ # read code and find what is the checkpoint arg.
+ common_eval:
+ path: runs/task/checkpoint_best.pt
+metric: DiDeMoMetric
+predictor: DiDeMoPredictor
diff --git a/fairseq/examples/MMPT/projects/task/test_vtt.yaml b/fairseq/examples/MMPT/projects/task/test_vtt.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..2f809b306d5d2b4d50d83ca3f9fa2db3f64530a5
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/task/test_vtt.yaml
@@ -0,0 +1,19 @@
+includes: projects/task/test.yaml
+dataset:
+ meta_processor: MSRVTTMetaProcessor
+ test_path: data/msrvtt/MSRVTT_JSFUSION_test.csv
+ video_processor: VideoProcessor
+ vfeat_dir: data/feat/feat_vtt_s3d
+ text_processor: MSRVTTTextProcessor
+ num_iso_layer: 12
+model:
+ model_cls: MMFusionJoint
+ mm_encoder_cls: MMBertForJoint
+eval:
+ save_path: runs/task/vtt/eval
+fairseq:
+ # read code and find what is the checkpoint arg.
+ common_eval:
+ path: runs/task/vtt/checkpoint_last.pt
+metric: RetrievalMetric
+predictor: RetrievalPredictor
diff --git a/fairseq/examples/MMPT/projects/task/test_vtt_videoclip.yaml b/fairseq/examples/MMPT/projects/task/test_vtt_videoclip.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..cb6564394c8de5edebe41d9a90e8233d2ef90edf
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/task/test_vtt_videoclip.yaml
@@ -0,0 +1,8 @@
+includes: projects/task/test_vtt.yaml
+model:
+ model_cls: MMFusionSeparate
+ mm_encoder_cls:
+ video_encoder_cls: MMBertForEncoder
+ text_encoder_cls: BertModel
+ num_hidden_video_layers: 6
+
diff --git a/fairseq/examples/MMPT/projects/task/test_vtt_zs.yaml b/fairseq/examples/MMPT/projects/task/test_vtt_zs.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..57340924b434338815743d0d2c326737ccfe3818
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/task/test_vtt_zs.yaml
@@ -0,0 +1,13 @@
+includes: projects/task/test_vtt.yaml
+model:
+ model_cls: MMFusionSeparate
+ mm_encoder_cls:
+ video_encoder_cls: MMBertForEncoder
+ text_encoder_cls: BertModel
+ num_hidden_video_layers: 6
+eval:
+ save_path: runs/task/vtt_zs/eval
+fairseq:
+ # read code and find what is the checkpoint arg.
+ common_eval:
+ path: runs/task/checkpoint_best.pt
diff --git a/fairseq/examples/MMPT/projects/task/test_vttqa.yaml b/fairseq/examples/MMPT/projects/task/test_vttqa.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..ddf813c535c0ecd6e6fd81ba9013a88f8ec4296b
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/task/test_vttqa.yaml
@@ -0,0 +1,20 @@
+includes: projects/task/test.yaml
+dataset:
+ meta_processor: MSRVTTQAMetaProcessor
+ test_path: data/msrvtt-qa/MSR_MC_test.csv
+ video_processor: VideoProcessor
+ vfeat_dir: data/feat/feat_vtt_s3d
+ text_processor: MSRVTTQATextProcessor
+ aligner: MSRVTTQAAligner
+ num_iso_layer: 12
+model:
+ model_cls: MMFusionJoint
+ mm_encoder_cls: MMBertForJoint
+eval:
+ save_path: runs/task/vttqa/eval
+fairseq:
+ # read code and find what is the checkpoint arg.
+ common_eval:
+ path: runs/task/vttqa/checkpoint_last.pt
+metric: QAMetric
+predictor: QAPredictor
diff --git a/fairseq/examples/MMPT/projects/task/test_vttqa_videoclip.yaml b/fairseq/examples/MMPT/projects/task/test_vttqa_videoclip.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..32a41e861c2e71642d4e79f8cf9dc314a5a3f621
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/task/test_vttqa_videoclip.yaml
@@ -0,0 +1,8 @@
+includes: projects/task/test_vttqa.yaml
+model:
+ model_cls: MMFusionSeparate
+ mm_encoder_cls:
+ video_encoder_cls: MMBertForEncoder
+ text_encoder_cls: BertModel
+ num_hidden_video_layers: 6
+
diff --git a/fairseq/examples/MMPT/projects/task/test_vttqa_zs.yaml b/fairseq/examples/MMPT/projects/task/test_vttqa_zs.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..5e0e29d20729445258f5d07348fd1799f7feecf4
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/task/test_vttqa_zs.yaml
@@ -0,0 +1,13 @@
+includes: projects/task/test_vttqa.yaml
+model:
+ model_cls: MMFusionSeparate
+ mm_encoder_cls:
+ video_encoder_cls: MMBertForEncoder
+ text_encoder_cls: BertModel
+ num_hidden_video_layers: 6
+eval:
+ save_path: runs/task/vttqa_zs/eval
+fairseq:
+ # read code and find what is the checkpoint arg.
+ common_eval:
+ path: runs/task/checkpoint_best.pt
diff --git a/fairseq/examples/MMPT/projects/task/test_youcook.yaml b/fairseq/examples/MMPT/projects/task/test_youcook.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..092b680fa6400458c447abe4f5cc5e452ffdb54f
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/task/test_youcook.yaml
@@ -0,0 +1,22 @@
+includes: projects/task/test.yaml
+dataset:
+ meta_processor: YoucookMetaProcessor
+ test_path: data/youcook/youcook_val.pkl
+ trainval_annotation: data/youcook/youcookii_annotations_trainval.json
+ use_annotation_text: True
+ video_processor: YoucookVideoProcessor
+ vfeat_dir: data/feat/feat_youcook_s3d # /checkpoint/huxu/feat/youcook_vmz # /checkpoint/prarora/berniehuang/feat_youcook_vmz
+ text_processor: TextProcessor
+ aligner: DSAligner
+ num_iso_layer: 12
+model:
+ model_cls: MMFusionJoint
+ mm_encoder_cls: MMBertForJoint
+eval:
+ save_path: runs/task/youcook/eval
+fairseq:
+ # read code and find what is the checkpoint arg.
+ common_eval:
+ path: runs/task/youcook/checkpoint_last.pt
+metric: RetrievalMetric
+predictor: RetrievalPredictor
diff --git a/fairseq/examples/MMPT/projects/task/test_youcook_videoclip.yaml b/fairseq/examples/MMPT/projects/task/test_youcook_videoclip.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..b85ea434749009f50eea057146403691bb48c4a9
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/task/test_youcook_videoclip.yaml
@@ -0,0 +1,8 @@
+includes: projects/task/test_youcook.yaml
+model:
+ model_cls: MMFusionSeparate
+ mm_encoder_cls:
+ video_encoder_cls: MMBertForEncoder
+ text_encoder_cls: BertModel
+ num_hidden_video_layers: 6
+
diff --git a/fairseq/examples/MMPT/projects/task/test_youcook_zs.yaml b/fairseq/examples/MMPT/projects/task/test_youcook_zs.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..0a5875bea4eaca1e46c65dfdc50b586550cc6f36
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/task/test_youcook_zs.yaml
@@ -0,0 +1,13 @@
+includes: projects/task/test_youcook.yaml
+model:
+ model_cls: MMFusionSeparate
+ mm_encoder_cls:
+ video_encoder_cls: MMBertForEncoder
+ text_encoder_cls: BertModel
+ num_hidden_video_layers: 6
+eval:
+ save_path: runs/task/youcook_zs/eval
+fairseq:
+ # read code and find what is the checkpoint arg.
+ common_eval:
+ path: runs/task/checkpoint_best.pt
diff --git a/fairseq/examples/MMPT/projects/task/test_youcookcap.yaml b/fairseq/examples/MMPT/projects/task/test_youcookcap.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..24f6518b7b4e210289f5200edafdb436d569adf7
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/task/test_youcookcap.yaml
@@ -0,0 +1,23 @@
+includes: projects/task/test.yaml
+dataset:
+ meta_processor: YoucookNLGMetaProcessor
+ test_path: data/youcook/val_list.txt
+ trainval_annotation: data/youcook/youcookii_annotations_trainval.json
+ video_processor: YoucookVideoProcessor
+ vfeat_dir: data/feat/feat_youcook_s3d
+ text_processor: NLGTextProcessor
+ aligner: DSNLGAligner
+model:
+ model_cls: MMFusionNLG
+ mm_encoder_cls: MMBertForNLG
+ max_decode_length: 24
+eval:
+ save_path: runs/task/youcookcap/eval
+fairseq:
+ # read code and find what is the checkpoint arg.
+ common_eval:
+ path: runs/task/youcookcap/checkpoint_best.pt
+metric: NLGMetric
+predictor: NLGPredictor
+gen_param:
+ num_beams: 5
diff --git a/fairseq/examples/MMPT/projects/task/vtt.yaml b/fairseq/examples/MMPT/projects/task/vtt.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..395e2ee6fe96899f9c47e39ad33b558358431554
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/task/vtt.yaml
@@ -0,0 +1,25 @@
+includes: projects/task/ft.yaml
+dataset:
+ meta_processor: MSRVTTMetaProcessor
+ train_path: data/msrvtt/MSRVTT_train.csv
+ jsfusion_path: data/msrvtt/MSRVTT_JSFUSION_test.csv
+ full_test_path: data/msrvtt/MSRVTT_FULL_test.csv
+ dup: 20
+ val_path: data/msrvtt/MSRVTT_JSFUSION_test.csv
+ vfeat_dir: data/feat/feat_vtt_s3d
+ text_processor: MSRVTTTextProcessor
+ json_path: data/msrvtt/MSRVTT_data.json
+ aligner: DSAligner
+ num_iso_layer: 12
+model:
+ model_cls: MMFusionJoint
+ mm_encoder_cls: MMBertForJoint
+loss:
+ loss_cls: T2VContraLoss
+fairseq:
+ dataset:
+ batch_size: 256
+ optimization:
+ max_epoch: 10
+ checkpoint:
+ save_dir: runs/task/vtt
diff --git a/fairseq/examples/MMPT/projects/task/vtt_videoclip.yaml b/fairseq/examples/MMPT/projects/task/vtt_videoclip.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..a9892cab01b810175dcef2ee91f0e64e96bd36d9
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/task/vtt_videoclip.yaml
@@ -0,0 +1,12 @@
+includes: projects/task/vtt.yaml
+model:
+ model_cls: MMFusionSeparate
+ mm_encoder_cls:
+ video_encoder_cls: MMBertForEncoder
+ text_encoder_cls: BertModel
+ num_hidden_video_layers: 6
+fairseq:
+ dataset:
+ batch_size: 224
+# model_cls: MMFusionShare
+# mm_encoder_cls: MMBertForEncoder
diff --git a/fairseq/examples/MMPT/projects/task/vttqa.yaml b/fairseq/examples/MMPT/projects/task/vttqa.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..56d578eff0104e5dcb4cea8720303100a9964db9
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/task/vttqa.yaml
@@ -0,0 +1,23 @@
+includes: projects/task/ft.yaml
+dataset:
+ meta_processor: MSRVTTMetaProcessor
+ train_path: data/msrvtt/MSRVTT_train.csv
+ dup: 20
+ val_path: data/msrvtt/MSRVTT_JSFUSION_test.csv
+ vfeat_dir: data/feat/feat_vtt_s3d
+ text_processor: MSRVTTTextProcessor
+ json_path: data/msrvtt/MSRVTT_data.json
+ aligner: DSAligner
+ num_iso_layer: 12
+model:
+ model_cls: MMFusionJoint
+ mm_encoder_cls: MMBertForJoint
+loss:
+ loss_cls: V2TContraLoss
+fairseq:
+ dataset:
+ batch_size: 128
+ optimization:
+ max_epoch: 5
+ checkpoint:
+ save_dir: runs/task/vttqa
diff --git a/fairseq/examples/MMPT/projects/task/vttqa_videoclip.yaml b/fairseq/examples/MMPT/projects/task/vttqa_videoclip.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..2d484ca8a5de0e96172b49b746f93d052f50dea7
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/task/vttqa_videoclip.yaml
@@ -0,0 +1,10 @@
+includes: projects/task/vttqa.yaml
+model:
+ model_cls: MMFusionSeparate
+ mm_encoder_cls:
+ video_encoder_cls: MMBertForEncoder
+ text_encoder_cls: BertModel
+ num_hidden_video_layers: 6
+
+# model_cls: MMFusionShare
+# mm_encoder_cls: MMBertForEncoder
diff --git a/fairseq/examples/MMPT/projects/task/youcook.yaml b/fairseq/examples/MMPT/projects/task/youcook.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..e0cd841747a5cfa4c77a94d2478bfcb029517678
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/task/youcook.yaml
@@ -0,0 +1,25 @@
+includes: projects/task/ft.yaml
+dataset:
+ meta_processor: YoucookMetaProcessor
+ train_path: data/youcook/youcook_train.pkl
+ val_path: data/youcook/youcook_val.pkl
+ trainval_annotation: data/youcook/youcookii_annotations_trainval.json
+ use_annotation_text: True
+ video_processor: YoucookVideoProcessor
+ vfeat_dir: data/feat/feat_youcook_s3d # /checkpoint/huxu/feat/youcook_vmz # /checkpoint/prarora/berniehuang/feat_youcook_vmz
+ text_processor: TextProcessor
+ aligner: DSAligner
+ num_iso_layer: 12
+model:
+ model_cls: MMFusionJoint
+ mm_encoder_cls: MMBertForJoint
+loss:
+ loss_cls: T2VContraLoss
+fairseq:
+ dataset:
+ batch_size: 128
+ optimization:
+ max_epoch: 10
+ checkpoint:
+ save_dir: runs/task/youcook
+
diff --git a/fairseq/examples/MMPT/projects/task/youcook_videoclip.yaml b/fairseq/examples/MMPT/projects/task/youcook_videoclip.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..e3e901c30ca8ef0e858b9fb0a317aea0b55f6e34
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/task/youcook_videoclip.yaml
@@ -0,0 +1,9 @@
+includes: projects/task/youcook.yaml
+model:
+ model_cls: MMFusionSeparate
+ mm_encoder_cls:
+ video_encoder_cls: MMBertForEncoder
+ text_encoder_cls: BertModel
+ num_hidden_video_layers: 6
+ # model_cls: MMFusionShare
+ # mm_encoder_cls: MMBertForEncoder
diff --git a/fairseq/examples/MMPT/projects/task/youcookcap.yaml b/fairseq/examples/MMPT/projects/task/youcookcap.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..047735f21716777354101ebaa230772244ac79e5
--- /dev/null
+++ b/fairseq/examples/MMPT/projects/task/youcookcap.yaml
@@ -0,0 +1,23 @@
+# finetuning for youcook captioning.
+includes: projects/task/ft.yaml
+dataset:
+ meta_processor: YoucookNLGMetaProcessor
+ train_path: data/youcook/train_list.txt
+ val_path: data/youcook/val_list.txt
+ trainval_annotation: data/youcook/youcookii_annotations_trainval.json
+ video_processor: YoucookVideoProcessor
+ vfeat_dir: data/feat/feat_youcook_s3d
+ text_processor: NLGTextProcessor
+ aligner: DSNLGAligner
+model:
+ model_cls: MMFusionNLG
+ mm_encoder_cls: MMBertForNLG
+loss:
+ loss_cls: NLGLoss
+fairseq:
+ dataset:
+ batch_size: 128
+ optimization:
+ max_epoch: 10
+ checkpoint:
+ save_dir: runs/task/youcookcap
diff --git a/fairseq/examples/MMPT/scripts/text_token_extractor/configs/bert-base-uncased.yaml b/fairseq/examples/MMPT/scripts/text_token_extractor/configs/bert-base-uncased.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..473dd9b45b78d6d4443921fb5f355356db4f5531
--- /dev/null
+++ b/fairseq/examples/MMPT/scripts/text_token_extractor/configs/bert-base-uncased.yaml
@@ -0,0 +1,5 @@
+dataset:
+ bert_name: bert-base-uncased
+ caption_pkl_path: data/how2/raw_caption_dedup.pkl
+ use_fast: true
+ target_dir: data/feat/feat_how2_s3d_shard_small
diff --git a/fairseq/examples/MMPT/scripts/text_token_extractor/pretokenization.py b/fairseq/examples/MMPT/scripts/text_token_extractor/pretokenization.py
new file mode 100644
index 0000000000000000000000000000000000000000..29ae5dc151085cc59de03e4e8bfc6e5341defb8b
--- /dev/null
+++ b/fairseq/examples/MMPT/scripts/text_token_extractor/pretokenization.py
@@ -0,0 +1,106 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import pickle
+import os
+import argparse
+import numpy as np
+
+from torch.utils.data import Dataset, DataLoader
+from mmpt.processors import PKLJSONStrTextProcessor
+from mmpt.utils import ShardedTensor, recursive_config
+
+
+class TokenizerDataset(Dataset):
+ def __init__(self, config):
+ self.text_processor = PKLJSONStrTextProcessor(config)
+ self.video_ids = list(self.text_processor.data.keys())
+
+ def __getitem__(self, idx):
+ video_id = self.video_ids[idx]
+ return video_id, self.text_processor(video_id)
+
+ def __len__(self):
+ return len(self.video_ids)
+
+
+def numpify(shard_idx, video_ids, captions, target_dir, split, prefix, max_cap_len=32):
+ startends = []
+ caps_ids = []
+ for video_id in video_ids:
+ caption = captions[video_id]
+ startend = []
+ cap_ids = []
+ for start, end, cap in zip(
+ caption["start"], caption["end"], caption["cap"]):
+ startend.append(np.array([start, end]).astype("float32"))
+ cap_id = np.full((max_cap_len,), -1, dtype=np.int32)
+ cap = cap[:max_cap_len]
+ cap_id[:len(cap)] = cap
+ cap_ids.append(cap_id)
+ startends.append(np.stack(startend))
+ caps_ids.append(np.stack(cap_ids))
+
+ startends = ShardedTensor.from_list(startends)
+ target_path = os.path.join(
+ target_dir,
+ prefix + split + "_" + str(shard_idx)
+ )
+ print("save to", target_path)
+ startends.save(target_path + ".startends")
+ caps_ids = ShardedTensor.from_list(caps_ids)
+ caps_ids.save(target_path + ".caps_ids")
+
+
+def sharding(config, out_file):
+ with open(out_file, "rb") as fr:
+ captions = pickle.load(fr)
+ target_dir = config.target_dir
+ prefix = os.path.basename(
+ os.path.splitext(config.caption_pkl_path)[0]
+ ) + "." + config.bert_name + "."
+ for split in ["train", "val"]:
+ target_path = os.path.join(target_dir, split + "_meta")
+ with open(target_path + ".pkl", "rb") as fr:
+ meta = pickle.load(fr)
+ print("load meta", target_path, len(meta))
+ for shard_id in meta:
+ numpify(
+ shard_id, meta[shard_id], captions,
+ target_dir, split, prefix
+ )
+
+
+def tokenize(config, out_file):
+ def collator(samples):
+ return samples
+ dataset = TokenizerDataset(config)
+ data = {}
+ for idx, batch in enumerate(
+ DataLoader(dataset, collate_fn=collator, num_workers=16)):
+ for video_id, caption in batch:
+ data[video_id] = caption
+ if idx % 5000 == 0:
+ print(idx)
+ with open(out_file, "wb") as fw:
+ pickle.dump(data, fw, pickle.HIGHEST_PROTOCOL)
+
+
+def main(args):
+ config = recursive_config(args.config).dataset
+
+ out_file = os.path.splitext(config.caption_pkl_path)[0] \
+ + "." + config.bert_name + ".pkl"
+ if not os.path.isfile(out_file):
+ tokenize(config, out_file)
+ sharding(config, out_file)
+
+
+if __name__ == "__main__":
+ parser = argparse.ArgumentParser(
+ description="pretokenize (raw_)caption.json into pkl.")
+ parser.add_argument('config', type=str)
+ args = parser.parse_args()
+ main(args)
diff --git a/fairseq/examples/MMPT/scripts/video_feature_extractor/extract.py b/fairseq/examples/MMPT/scripts/video_feature_extractor/extract.py
new file mode 100644
index 0000000000000000000000000000000000000000..b5ee7b778890994aa87d463b896f653792c073f2
--- /dev/null
+++ b/fairseq/examples/MMPT/scripts/video_feature_extractor/extract.py
@@ -0,0 +1,157 @@
+# Copyright Howto100M authors.
+# Copyright (c) Facebook, Inc. All Rights Reserved
+
+import torch as th
+import torch.nn.functional as F
+import math
+import numpy as np
+import argparse
+
+from torch.utils.data import DataLoader
+from model import get_model
+from preprocessing import Preprocessing
+from random_sequence_shuffler import RandomSequenceSampler
+
+from tqdm import tqdm
+from pathbuilder import PathBuilder
+from videoreader import VideoLoader
+
+
+parser = argparse.ArgumentParser(description='Easy video feature extractor')
+
+parser.add_argument('--vdir', type=str)
+parser.add_argument('--fdir', type=str)
+parser.add_argument('--hflip', type=int, default=0)
+
+parser.add_argument('--batch_size', type=int, default=64,
+ help='batch size')
+parser.add_argument('--type', type=str, default='2d',
+ help='CNN type')
+parser.add_argument('--half_precision', type=int, default=0,
+ help='output half precision float')
+parser.add_argument('--num_decoding_thread', type=int, default=4,
+ help='Num parallel thread for video decoding')
+parser.add_argument('--l2_normalize', type=int, default=1,
+ help='l2 normalize feature')
+parser.add_argument('--resnext101_model_path', type=str, default='model/resnext101.pth',
+ help='Resnext model path')
+parser.add_argument('--vmz_model_path', type=str, default='model/r2plus1d_34_clip8_ig65m_from_scratch-9bae36ae.pth',
+ help='vmz model path')
+
+args = parser.parse_args()
+
+
+# TODO: refactor all args into config. (current code is from different people.)
+CONFIGS = {
+ "2d": {
+ "fps": 1,
+ "size": 224,
+ "centercrop": False,
+ "shards": 0,
+ },
+ "3d": {
+ "fps": 24,
+ "size": 112,
+ "centercrop": True,
+ "shards": 0,
+ },
+ "s3d": {
+ "fps": 30,
+ "size": 224,
+ "centercrop": True,
+ "shards": 0,
+ },
+ "vmz": {
+ "fps": 24,
+ "size": 112,
+ "centercrop": True,
+ "shards": 0,
+ },
+ "vae": {
+ "fps": 2,
+ "size": 256,
+ "centercrop": True,
+ "shards": 100,
+ }
+}
+
+config = CONFIGS[args.type]
+
+
+video_dirs = args.vdir
+feature_dir = args.fdir
+
+video_dict = PathBuilder.build(video_dirs, feature_dir, ".npy", config["shards"])
+
+dataset = VideoLoader(
+ video_dict=video_dict,
+ framerate=config["fps"],
+ size=config["size"],
+ centercrop=config["centercrop"],
+ hflip=args.hflip
+)
+n_dataset = len(dataset)
+sampler = RandomSequenceSampler(n_dataset, 10)
+loader = DataLoader(
+ dataset,
+ batch_size=1,
+ shuffle=False,
+ num_workers=args.num_decoding_thread,
+ sampler=sampler if n_dataset > 10 else None,
+)
+preprocess = Preprocessing(args.type)
+model = get_model(args)
+
+with th.no_grad():
+ for k, data in tqdm(enumerate(loader), total=loader.__len__(), ascii=True):
+ input_file = data['input'][0]
+ output_file = data['output'][0]
+ if len(data['video'].shape) > 3:
+ video = data['video'].squeeze()
+ if len(video.shape) == 4:
+ video = preprocess(video)
+ n_chunk = len(video)
+ if args.type == 'vmz':
+ n_chunk = math.ceil(n_chunk/float(3))
+ features = th.cuda.FloatTensor(n_chunk, 512).fill_(0)
+ elif args.type == 's3d':
+ features = th.cuda.FloatTensor(n_chunk, 512).fill_(0)
+ elif args.type == "vae":
+ features = th.cuda.LongTensor(n_chunk, 1024).fill_(0)
+ else:
+ features = th.cuda.FloatTensor(n_chunk, 2048).fill_(0)
+ n_iter = int(math.ceil(n_chunk / float(args.batch_size)))
+ for i in range(n_iter):
+ factor = 1
+ if args.type == 'vmz':
+ factor = 3
+ min_ind = factor * i * args.batch_size
+ max_ind = factor * (i + 1) * args.batch_size
+ video_batch = video[min_ind:max_ind:factor].cuda()
+ if args.type == '2d':
+ batch_features = model(video_batch) # (51, 487), (51, 512)
+ elif args.type == 's3d':
+ batch_features = model(video_batch)
+ batch_features = batch_features['video_embedding']
+ elif args.type == "vae":
+ # image_code.
+ batch_features = model(video_batch)
+ else:
+ batch_pred, batch_features = model(video_batch) # (51, 487), (51, 512)
+ if args.l2_normalize:
+ batch_features = F.normalize(batch_features, dim=1)
+ features[i*args.batch_size:(i+1)*args.batch_size] = batch_features
+ features = features.cpu().numpy()
+ if args.half_precision:
+ if args.type == "vae":
+ features = features.astype(np.int16)
+ else:
+ features = features.astype('float16')
+ else:
+ if args.type == "vae":
+ features = features.astype(np.int32)
+ else:
+ features = features.astype('float32')
+ np.save(output_file, features)
+ else:
+ print('Video {} error.'.format(input_file))
diff --git a/fairseq/examples/MMPT/scripts/video_feature_extractor/how2/s3d.sh b/fairseq/examples/MMPT/scripts/video_feature_extractor/how2/s3d.sh
new file mode 100644
index 0000000000000000000000000000000000000000..90102c89fbb4fb413951dfd049b30b8dfe7cbe99
--- /dev/null
+++ b/fairseq/examples/MMPT/scripts/video_feature_extractor/how2/s3d.sh
@@ -0,0 +1,8 @@
+#!/bin/bash
+
+
+python scripts/video_feature_extractor/extract.py \
+ --vdir \
+ --fdir data/feat/feat_how2_s3d \
+ --type=s3d --num_decoding_thread=4 \
+ --batch_size 32 --half_precision 1
diff --git a/fairseq/examples/MMPT/scripts/video_feature_extractor/model.py b/fairseq/examples/MMPT/scripts/video_feature_extractor/model.py
new file mode 100644
index 0000000000000000000000000000000000000000..ac266e844c86246bbfce02b9e6a2999353661df9
--- /dev/null
+++ b/fairseq/examples/MMPT/scripts/video_feature_extractor/model.py
@@ -0,0 +1,58 @@
+# Copyright (c) Howto100M authors and Facebook, Inc. All Rights Reserved
+
+import torch as th
+
+from torch import nn
+
+
+class GlobalAvgPool(nn.Module):
+ def __init__(self):
+ super(GlobalAvgPool, self).__init__()
+
+ def forward(self, x):
+ return th.mean(x, dim=[-2, -1])
+
+
+def get_model(args):
+ assert args.type in ['2d', '3d', 'vmz', 's3d', 'vae']
+ if args.type == '2d':
+ print('Loading 2D-ResNet-152 ...')
+ import torchvision.models as models
+ model = models.resnet152(pretrained=True)
+ model = nn.Sequential(*list(model.children())[:-2], GlobalAvgPool())
+ model = model.cuda()
+ elif args.type == 'vmz':
+ print('Loading VMZ ...')
+ from vmz34 import r2plus1d_34
+ model = r2plus1d_34(pretrained_path=args.vmz_model_path, pretrained_num_classes=487)
+ model = model.cuda()
+ elif args.type == 's3d':
+ # we use one copy of s3d instead of dup another one for feature extraction.
+ from mmpt.processors.models.s3dg import S3D
+ model = S3D('pretrained_models/s3d_dict.npy', 512)
+ model.load_state_dict(th.load('pretrained_models/s3d_howto100m.pth'))
+ model = model.cuda()
+
+ elif args.type == '3d':
+ print('Loading 3D-ResneXt-101 ...')
+ from videocnn.models import resnext
+ model = resnext.resnet101(
+ num_classes=400,
+ shortcut_type='B',
+ cardinality=32,
+ sample_size=112,
+ sample_duration=16,
+ last_fc=False)
+ model = model.cuda()
+ model_data = th.load(args.resnext101_model_path)
+ model.load_state_dict(model_data)
+ elif args.type == 'vae':
+ from openaivae import OpenAIParallelDiscreteVAE
+ model = OpenAIParallelDiscreteVAE()
+ model = model.cuda()
+ else:
+ raise ValueError("model not supported yet.")
+
+ model.eval()
+ print('loaded')
+ return model
diff --git a/fairseq/examples/MMPT/scripts/video_feature_extractor/pathbuilder.py b/fairseq/examples/MMPT/scripts/video_feature_extractor/pathbuilder.py
new file mode 100644
index 0000000000000000000000000000000000000000..2392d6d63bc293fb24068e33f707ee4ce0e74a6f
--- /dev/null
+++ b/fairseq/examples/MMPT/scripts/video_feature_extractor/pathbuilder.py
@@ -0,0 +1,89 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+import os
+import urllib.parse
+import json
+import pandas as pd
+
+from tqdm import tqdm
+
+
+# TODO: extending to other datasets.
+supported_formats = {}
+
+
+class PathBuilder(object):
+ @classmethod
+ def build(cls, video_dirs, feature_dir, ext, shards=0, split=None):
+ meta_fn = os.path.join(feature_dir, "meta_plan.json")
+ os.makedirs(feature_dir, exist_ok=True)
+ if os.path.isfile(meta_fn):
+ with open(meta_fn) as fr:
+ meta = json.load(fr)
+ return meta
+ print("searching videos...")
+
+ video_id_to_path = {}
+ for video_dir in video_dirs.split(","):
+ # TODO: add supports of recursive listdir.
+ if video_dir in supported_formats:
+ supported_formats[video_dir].load(video_dir, video_id_to_path)
+ else:
+ for idx, fn in enumerate(tqdm(os.listdir(video_dir))):
+ video_fn = os.path.join(video_dir, fn)
+ if os.path.isfile(video_fn):
+ video_id = os.path.splitext(fn)[0]
+ video_id_to_path[video_id] = video_fn
+ elif os.path.isdir(video_fn):
+ # shards of folders.
+ shard_dir = video_fn
+ for idx, fn in enumerate(os.listdir(shard_dir)):
+ video_fn = os.path.join(shard_dir, fn)
+ if os.path.isfile(video_fn):
+ video_id = os.path.splitext(fn)[0]
+ video_id_to_path[video_id] = video_fn
+
+ video_path, feature_path = [], []
+ valid_ext = set()
+ for idx, video_id in enumerate(video_id_to_path):
+ video_path.append(video_id_to_path[video_id])
+ if ext is None:
+ # use original file ext for format compatibility.
+ video_id_to_path[video_id]
+ path = urllib.parse.urlparse(video_id_to_path[video_id]).path
+ ext = os.path.splitext(path)[1]
+ if ext not in valid_ext:
+ valid_ext.add(ext)
+ print("adding", ext)
+ if shards:
+ shard_id = str(idx % shards)
+ feature_fn = os.path.join(
+ feature_dir, shard_id, video_id + ext)
+ else:
+ feature_fn = os.path.join(
+ feature_dir, video_id + ext)
+ feature_path.append(feature_fn)
+
+ print("targeting", len(feature_path), "videos")
+ meta = {
+ "video_path": video_path, "feature_path": feature_path}
+ with open(meta_fn, "w") as fw:
+ json.dump(meta, fw)
+
+ if split is not None:
+ splits = split.split("/")
+ assert len(splits) == 2
+ cur, total = int(splits[0]), int(splits[1])
+ assert cur < total
+ import math
+ chunk = math.ceil(len(meta["video_path"]) / total)
+ start = cur * chunk
+ end = (cur + 1) * chunk
+ meta = {
+ "video_path": meta["video_path"][start:end],
+ "feature_path": meta["feature_path"][start:end]
+ }
+
+ return meta
diff --git a/fairseq/examples/MMPT/scripts/video_feature_extractor/preprocessing.py b/fairseq/examples/MMPT/scripts/video_feature_extractor/preprocessing.py
new file mode 100644
index 0000000000000000000000000000000000000000..fa0cec3a76b4665fae57d1feadd5a52469234226
--- /dev/null
+++ b/fairseq/examples/MMPT/scripts/video_feature_extractor/preprocessing.py
@@ -0,0 +1,57 @@
+# Copyright Howto100m authors.
+# Copyright (c) Facebook, Inc. All Rights Reserved
+
+import torch as th
+
+class Normalize(object):
+
+ def __init__(self, mean, std):
+ self.mean = th.FloatTensor(mean).view(1, 3, 1, 1)
+ self.std = th.FloatTensor(std).view(1, 3, 1, 1)
+
+ def __call__(self, tensor):
+ tensor = (tensor - self.mean) / (self.std + 1e-8)
+ return tensor
+
+class Preprocessing(object):
+
+ def __init__(self, type):
+ self.type = type
+ if type == '2d':
+ self.norm = Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
+ elif type == '3d':
+ self.norm = Normalize(mean=[110.6, 103.2, 96.3], std=[1.0, 1.0, 1.0])
+ elif type == 'vmz':
+ self.norm = Normalize(mean=[110.201, 100.64, 95.997], std=[58.1489, 56.4701, 55.3324])
+
+ def _zero_pad(self, tensor, size):
+ n = size - len(tensor) % size
+ if n == size:
+ return tensor
+ else:
+ z = th.zeros(n, tensor.shape[1], tensor.shape[2], tensor.shape[3])
+ return th.cat((tensor, z), 0)
+
+ def __call__(self, tensor):
+ if self.type == '2d':
+ tensor = tensor / 255.0
+ tensor = self.norm(tensor)
+ elif self.type == 'vmz':
+ #tensor = self._zero_pad(tensor, 8)
+ tensor = self._zero_pad(tensor, 10)
+ tensor = self.norm(tensor)
+ #tensor = tensor.view(-1, 8, 3, 112, 112)
+ tensor = tensor.view(-1, 10, 3, 112, 112)
+ tensor = tensor.transpose(1, 2)
+ elif self.type == '3d':
+ tensor = self._zero_pad(tensor, 16)
+ tensor = self.norm(tensor)
+ tensor = tensor.view(-1, 16, 3, 112, 112)
+ tensor = tensor.transpose(1, 2)
+ elif self.type == 's3d':
+ tensor = tensor / 255.0
+ tensor = self._zero_pad(tensor, 30)
+ tensor = tensor.view(-1, 30, 3, 224, 224) # N x 30 x 3 x H x W
+ tensor = tensor.transpose(1, 2) # N x 3 x 30 x H x W
+ # for vae do nothing
+ return tensor
diff --git a/fairseq/examples/MMPT/scripts/video_feature_extractor/random_sequence_shuffler.py b/fairseq/examples/MMPT/scripts/video_feature_extractor/random_sequence_shuffler.py
new file mode 100644
index 0000000000000000000000000000000000000000..1f3e4aceaa8f930c9eeac6283b15924e7eed5dc3
--- /dev/null
+++ b/fairseq/examples/MMPT/scripts/video_feature_extractor/random_sequence_shuffler.py
@@ -0,0 +1,29 @@
+# Copyright (c) Facebook, Inc. All Rights Reserved
+
+import numpy as np
+
+from torch.utils.data.sampler import Sampler
+
+
+class RandomSequenceSampler(Sampler):
+
+ def __init__(self, n_sample, seq_len):
+ self.n_sample = n_sample
+ self.seq_len = seq_len
+
+ def _pad_ind(self, ind):
+ zeros = np.zeros(self.seq_len - self.n_sample % self.seq_len)
+ ind = np.concatenate((ind, zeros))
+ return ind
+
+ def __iter__(self):
+ idx = np.arange(self.n_sample)
+ if self.n_sample % self.seq_len != 0:
+ idx = self._pad_ind(idx)
+ idx = np.reshape(idx, (-1, self.seq_len))
+ np.random.shuffle(idx)
+ idx = np.reshape(idx, (-1))
+ return iter(idx.astype(int))
+
+ def __len__(self):
+ return self.n_sample + (self.seq_len - self.n_sample % self.seq_len)
diff --git a/fairseq/examples/MMPT/scripts/video_feature_extractor/shard_feature.py b/fairseq/examples/MMPT/scripts/video_feature_extractor/shard_feature.py
new file mode 100644
index 0000000000000000000000000000000000000000..f75e1dfae558fa366d15fa19e89d353003e728b3
--- /dev/null
+++ b/fairseq/examples/MMPT/scripts/video_feature_extractor/shard_feature.py
@@ -0,0 +1,64 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+import numpy as np
+import os
+import pickle
+
+from mmpt.utils import ShardedTensor
+
+
+class Shard(object):
+ def __init__(
+ self,
+ vfeat_dir,
+ tfeat_dir,
+ target_dir,
+ file_paths,
+ shard_size=4096
+ ):
+ self.vfeat_dir = vfeat_dir
+ self.tfeat_dir = tfeat_dir
+ self.target_dir = target_dir
+ self.video_ids = {}
+ for split, file_path in zip(["train", "val"], file_paths):
+ with open(file_path) as fr:
+ self.video_ids[split] = [
+ line.strip() for line in fr.readlines()]
+ self.shard_size = shard_size
+
+ def __call__(self, split="train"):
+ for split in ["train", "val"]:
+ meta = {}
+ for shard_idx, shard_offset in enumerate(
+ range(0, len(self.video_ids[split]), self.shard_size)
+ ):
+ print(shard_idx)
+ meta_shard = []
+ video_shard = []
+ for video_id in self.video_ids[split][shard_offset:shard_offset+self.shard_size]:
+ meta_shard.append(video_id)
+ npy_file = os.path.join(self.vfeat_dir, video_id + ".npy")
+ video_shard.append(np.load(npy_file))
+
+ meta[shard_idx] = meta_shard
+ video_shard = ShardedTensor.from_list(video_shard)
+ target_path = os.path.join(
+ self.target_dir, split + "_" + str(shard_idx))
+ video_shard.save(target_path)
+
+ target_path = os.path.join(self.target_dir, split + "_meta")
+ with open(target_path + ".pkl", "wb") as fw:
+ pickle.dump(meta, fw, pickle.HIGHEST_PROTOCOL)
+
+
+if __name__ == "__main__":
+ shard = Shard(
+ "data/feat/feat_how2_s3d",
+ "data/how2/raw_caption_dedup.bert-base-uncased",
+ "data/feat/feat_how2_s3d_shard_small",
+ ["data/how2/how2_s3d_train.lst", "data/how2/how2_s3d_val.lst"]
+ )
+
+ shard()
diff --git a/fairseq/examples/MMPT/scripts/video_feature_extractor/videoreader.py b/fairseq/examples/MMPT/scripts/video_feature_extractor/videoreader.py
new file mode 100644
index 0000000000000000000000000000000000000000..429e05f8bc8667408b8c2057578b1e8c0e98638c
--- /dev/null
+++ b/fairseq/examples/MMPT/scripts/video_feature_extractor/videoreader.py
@@ -0,0 +1,242 @@
+# Copyright Howto100M authors.
+# Copyright (c) Facebook, Inc. All Rights Reserved
+
+import torch as th
+import pandas as pd
+import os
+import numpy as np
+import ffmpeg
+import random
+
+from torch.utils.data import Dataset
+
+
+class VideoLoader(Dataset):
+ """modified from how2's video_feature_extractor."""
+ def __init__(
+ self,
+ csv=None,
+ video_dict=None,
+ framerate=1,
+ size=112,
+ centercrop=False,
+ hflip=False,
+ **kwargs
+ ):
+ if csv is None and video_dict is None:
+ raise ValueError("csv and video_dict cannot be both None.")
+ if csv is not None:
+ self.csv = pd.read_csv(csv)
+ if video_dict is not None:
+ self.csv = pd.DataFrame.from_dict(video_dict)
+
+ self.centercrop = centercrop
+ self.size = size
+ self.framerate = framerate
+ self.hflip = hflip
+
+ def __len__(self):
+ return len(self.csv)
+
+ def _get_video_dim(self, video_path):
+ probe = ffmpeg.probe(video_path)
+ video_stream = next((stream for stream in probe['streams']
+ if stream['codec_type'] == 'video'), None)
+ width = int(video_stream['width'])
+ height = int(video_stream['height'])
+ return height, width
+
+ def _get_video_info(self, video_path):
+ probe = ffmpeg.probe(video_path)
+ video_stream = next((stream for stream in probe['streams']
+ if stream['codec_type'] == 'video'), None)
+ return video_stream
+
+ def _get_output_dim(self, h, w):
+ if isinstance(self.size, tuple) and len(self.size) == 2:
+ return self.size
+ elif h >= w:
+ return int(h * self.size / w), self.size
+ else:
+ return self.size, int(w * self.size / h)
+
+ def __getitem__(self, idx):
+ video_path = self.csv['video_path'].values[idx]
+ output_file = self.csv['feature_path'].values[idx]
+ return self._decode(output_file, video_path)
+
+ def _decode(self, output_file, video_path):
+ if not(os.path.isfile(output_file)) and os.path.isfile(video_path):
+ try:
+ h, w = self._get_video_dim(video_path)
+ except Exception:
+ print('ffprobe failed at: {}'.format(video_path))
+ return {'video': th.zeros(1), 'input': video_path,
+ 'output': output_file}
+ try:
+ os.makedirs(os.path.dirname(output_file), exist_ok=True)
+ height, width = self._get_output_dim(h, w)
+
+ cmd = (
+ ffmpeg
+ .input(video_path)
+ .filter('fps', fps=self.framerate)
+ .filter('scale', width, height)
+ )
+ if self.hflip:
+ cmd = cmd.filter('hflip')
+
+ if self.centercrop:
+ x = int((width - self.size) / 2.0)
+ y = int((height - self.size) / 2.0)
+ cmd = cmd.crop(x, y, self.size, self.size)
+ video = self._run(cmd, output_file)
+ except Exception:
+ video = th.zeros(1)
+ else:
+ video = th.zeros(1)
+
+ return {'video': video, 'input': video_path, 'output': output_file}
+
+ def _run(self, cmd, output_file):
+ out, _ = (
+ cmd.output('pipe:', format='rawvideo', pix_fmt='rgb24')
+ .run(capture_stdout=True, quiet=True)
+ )
+ if self.centercrop and isinstance(self.size, int):
+ height, width = self.size, self.size
+ video = np.frombuffer(out, np.uint8).reshape([-1, height, width, 3])
+ video = th.from_numpy(video.astype('float32'))
+ return video.permute(0, 3, 1, 2)
+
+
+class VideoVerifier(VideoLoader):
+ def __getitem__(self, idx):
+ video_path = self.csv['video_path'].values[idx]
+ try:
+ return self._get_video_info(video_path)
+ except Exception:
+ # print('ffprobe failed at: {}'.format(video_path))
+ return None
+
+
+class VideoCompressor(VideoLoader):
+ def __init__(
+ self,
+ csv=None,
+ video_dict=None,
+ framerate=1,
+ size=112,
+ centercrop=False,
+ hflip=False,
+ crf=32,
+ **kwargs
+ ):
+ super().__init__(
+ csv,
+ video_dict,
+ framerate,
+ size,
+ centercrop,
+ hflip
+ )
+ self.crf = crf
+
+ def _run(self, cmd, output_file):
+ out, _ = (
+ cmd.output(filename=output_file, crf=self.crf)
+ .run(quiet=True)
+ )
+ video = None
+ return video
+
+
+class VideoDownloader(VideoCompressor):
+ """download"""
+ def __getitem__(self, idx):
+ video_path = self.csv['video_path'].values[idx]
+ output_file = self.csv['feature_path'].values[idx]
+ if not(os.path.isfile(output_file)):
+ os.makedirs(os.path.dirname(output_file), exist_ok=True)
+ cmd = "wget -O" + output_file + " " + video_path
+ # import subprocess
+ # subprocess.check_output(
+ # cmd,
+ # stderr=subprocess.STDOUT, shell=True)
+ os.system(cmd)
+ return {'video': None, 'input': video_path, 'output': output_file}
+
+
+class AvKeyframeVideoCompressor(VideoLoader):
+ """extract keyframes from a video and save it as jpg.
+ TODO: consider to merge with `CodecProcessor`.
+ """
+ def __init__(
+ self,
+ csv=None,
+ video_dict=None,
+ framerate=1,
+ size=112,
+ centercrop=False,
+ max_num_frames=5,
+ **kwargs
+ ):
+ super().__init__(csv, video_dict, framerate, size, centercrop)
+ self.max_num_frames = max_num_frames
+
+ def _get_video_dim(self, video_fn):
+ """decord cannot probe the size of a video, we use pyav instead."""
+ import av
+ with av.open(video_fn) as container:
+ height = container.streams.video[0].codec_context.height
+ width = container.streams.video[0].codec_context.width
+ return height, width
+
+ def _get_output_dim(self, height, width):
+ """
+ keep the shorter side be `self.size`, strech the other.
+ """
+ if height >= width:
+ return int(height * self.size / width), self.size
+ else:
+ return self.size, int(width * self.size / height)
+
+ def __getitem__(self, idx):
+ import av
+ video_path = self.csv['video_path'].values[idx]
+ output_file = self.csv['feature_path'].values[idx]
+ if not(os.path.isdir(output_file)) and os.path.isfile(video_path):
+ try:
+ h, w = self._get_video_dim(video_path)
+ except Exception:
+ print('probe failed at: {}'.format(video_path))
+ return {'video': th.zeros(1), 'input': video_path,
+ 'output': output_file}
+
+ try:
+ height, width = self._get_output_dim(h, w)
+
+ # new for av.
+ with av.open(video_path) as container:
+ container.streams.video[0].thread_type = "AUTO"
+ container.streams.video[0].codec_context.height = height
+ container.streams.video[0].codec_context.width = width
+ if self.framerate == 0: # keyframe.
+ container.streams.video[0].codec_context.skip_frame = 'NONKEY'
+ frames = []
+ for frame in container.decode(video=0):
+ frames.append(frame)
+ frames = random.sample(frames, self.max_num_frames)
+
+ os.makedirs(output_file, exist_ok=True)
+ for frame in frames:
+ frame.to_image().save(
+ os.path.join(
+ output_file,
+ "%04d.jpg" % frame.index))
+ except Exception:
+ print('extract failed at: {}'.format(video_path))
+ return {'video': th.zeros(1), 'input': video_path,
+ 'output': output_file}
+ video = th.zeros(1)
+ return {'video': video, 'input': video_path, 'output': output_file}
diff --git a/fairseq/examples/MMPT/setup.py b/fairseq/examples/MMPT/setup.py
new file mode 100644
index 0000000000000000000000000000000000000000..a9a82296ea9c3a53f760c29c34020b5a90091a89
--- /dev/null
+++ b/fairseq/examples/MMPT/setup.py
@@ -0,0 +1,24 @@
+import setuptools
+
+with open("README.md", "r") as fh:
+ long_description = fh.read()
+
+setuptools.setup(
+ name="mmpt",
+ version="0.0.1",
+ author="Hu Xu, Po-yao Huang",
+ author_email="huxu@fb.com",
+ description="A package for multimodal pretraining.",
+ long_description=long_description,
+ long_description_content_type="text/markdown",
+ url="https://github.com/pytorch/fairseq/examples/MMPT",
+ packages=setuptools.find_packages(),
+ install_requires=[
+ ],
+ classifiers=[
+ "Programming Language :: Python :: 3",
+ "License :: CC-BY-NC",
+ "Operating System :: OS Independent",
+ ],
+ python_requires='>=3.6',
+)
diff --git a/fairseq/examples/MMPT/videoclip.png b/fairseq/examples/MMPT/videoclip.png
new file mode 100644
index 0000000000000000000000000000000000000000..2b6f8d959fcfe3ce229df16a52c5c8230950b582
--- /dev/null
+++ b/fairseq/examples/MMPT/videoclip.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1d54fe18d1259ade9332e78fdb74f834fdfbdb0b0486517e6a7cd48956b30663
+size 385871
diff --git a/fairseq/examples/MMPT/vlm.png b/fairseq/examples/MMPT/vlm.png
new file mode 100644
index 0000000000000000000000000000000000000000..c702b32b12d7d630c25af96a98ded4f83378bcb4
--- /dev/null
+++ b/fairseq/examples/MMPT/vlm.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:722852ed6258ac9f7ffd3e3913fa1a370702c4d989ef6d881847432d59ade4e5
+size 418405
diff --git a/fairseq/examples/__init__.py b/fairseq/examples/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..44bb24ae614941f23fea29c56d60167650c39bcb
--- /dev/null
+++ b/fairseq/examples/__init__.py
@@ -0,0 +1,9 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+try:
+ from fairseq.version import __version__ # noqa
+except ImportError:
+ pass
diff --git a/fairseq/examples/adaptive_span/README.md b/fairseq/examples/adaptive_span/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..d5224fb2894606a2a8027e01e224be190776ecfe
--- /dev/null
+++ b/fairseq/examples/adaptive_span/README.md
@@ -0,0 +1,90 @@
+# Adaptive Span
+
+Adaptive Span is a novel self-attention mechanism that can learn its optimal
+attention span. This allows us to extend significantly the maximum context size
+used in Transformer, while maintaining control over their memory footprint
+and computational time. It uses the Truncated BPTT technique for training,
+as in [transformerXL](https://github.com/pytorch/fairseq/blob/main/examples/truncated_bptt/README.md).
+
+Adaptive Span was introduced by paper:
+[Adaptive Attention Span in Transformers](https://arxiv.org/abs/1905.07799),
+which achieved state-of-the-art language modeling results at the time of publication.
+
+We manage to reproduce their result in fairseq and keep most of the
+[original implementation](https://github.com/facebookresearch/adaptive-span) untouched.
+You can refer to the their sweep file as well if any combination of hyperparameter is not clear.
+
+##### 0. Setup
+
+First you need to process the Enwik8 dataset, we use the pre-tokenized dataset
+from [adaptive span paper](https://github.com/facebookresearch/adaptive-span/blob/master/get_data.sh).
+You can download the dataset, and then run:
+```bash
+fairseq-preprocess --only-source --trainpref ~/data/enwik8/train.txt \
+ --validpref ~/data/enwik8/valid.txt --testpref ~/data/enwik8/test.txt \
+ --destdir ~/data/enwik8/data-bin/ --joined-dictionary --workers 20
+```
+
+##### 1. Train a Adaptive Span model on Enwik8
+
+We will train a 12-layer Adaptive Span model following the [hyperparameters
+used in the original
+paper](https://github.com/facebookresearch/adaptive-span/blob/master/experiments/enwik8.sh).
+
+The following command assumes 4 GPUs, so that the total batch size is 64
+sequences (4 x 16). Training should take 2-3 days on 4 V100 GPUs:
+```bash
+CUDA_VISIBLE_DEVICES=0,1,2,3 fairseq-train \
+ --user-dir examples/adaptive_span \
+ --data ~/data/enwik8/data-bin/ \
+ --fp16 --fp16-no-flatten-grads --max-update 600000 \
+ --task truncated_bptt_lm --tokens-per-sample 512 --arch adaptive_span \
+ --n-layer 12 --d-model 512 --n-head 8 --d-inner 2048 --dropout 0.3 \
+ --attn-span 8192 --optimizer adagrad_with_grad_clip --adagrad-clip 0.03 \
+ --validate-interval-updates 1000 \
+ --lr-scheduler fixed --warmup-updates 32000 --batch-size-valid 32 \
+ --lr 0.07 --criterion adaptive_span_loss --batch-size 16 --update-freq 1 \
+ --seed 2 --log-format json --log-interval 25 --aux-loss-scaler 5e-07
+```
+This should land around 1.05 on validation, 1.03 on test. You can lower the
+--aux-loss-scaler for better performance (longer span). It gives ~0.03 bpc
+improvement to the transformerXL baseline here.
+If training on a single GPU, set `--update-freq=4` to accumulate 4x gradients
+and simulate training on 4 GPUs.
+You can also reproduce the transformerXL result on enwik8 using this code base.
+It should land around 1.06 on test,matching the [original paper](https://github.com/kimiyoung/transformer-xl/blob/master/pytorch/run_enwik8_base.sh).
+You can try by
+```bash
+CUDA_VISIBLE_DEVICES=0,1,2,3 fairseq-train \
+ --user-dir examples/truncated_bptt \
+ ~/data/enwik8/data-bin/ \
+ --task truncated_bptt_lm --fp16 --max-update 400000 \
+ --tokens-per-sample 512 --arch transformer_xl --n-layer 12 \
+ --d-model 512 --n-head 8 --d-head 64 --d-inner 2048 --dropout 0.1 \
+ --dropatt 0.0 --mem-len 512 --optimizer adam --clip-norm 0.25 \
+ --lr-scheduler cosine --warmup-updates 0 \
+ --lr 0.0 --lr 0.00025 --batch-size 15 \
+ --update-freq 1 --seed 2 --log-format json --log-interval 25 \
+ --fp16
+```
+
+##### 2. Evaluate
+For Adaptive Span:
+```bash
+fairseq-eval-lm ~/data/enwik8/data-bin/ --path model/checkpoint_best.pt \
+ --user-dir examples/adaptive_span \
+ --task truncated_bptt_lm --batch-size 8 --tokens-per-sample 512 --gen-subset test
+```
+For Transformer-XL evaluation:
+```bash
+fairseq-eval-lm ~/data/enwik8/data-bin/ --path model/checkpoint_best.pt \
+ --user-dir examples/truncated_bptt/ --task truncated_bptt_lm --batch-size 8 \
+ --tokens-per-sample 80 \
+ --model-overrides '{"mem_len":2100,"clamp_len":820,"same_length":True}' \
+ --gen-subset valid
+```
+
+*Note:* During training the model saw 512 tokens of context
+(``--tokens-per-sample=512``), with batch size 8. These settings match the evaluation
+settings from [the original
+paper](https://github.com/facebookresearch/adaptive-span/blob/master/experiments/enwik8.sh).
diff --git a/fairseq/examples/adaptive_span/__init__.py b/fairseq/examples/adaptive_span/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..e0a142a769360e1140bf814c532eaf841f1d52d8
--- /dev/null
+++ b/fairseq/examples/adaptive_span/__init__.py
@@ -0,0 +1,19 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import importlib
+import os
+
+# automatically import any Python files in the current directory
+cur_dir = os.path.dirname(__file__)
+for file in os.listdir(cur_dir):
+ path = os.path.join(cur_dir, file)
+ if (
+ not file.startswith("_")
+ and not file.startswith(".")
+ and (file.endswith(".py") or os.path.isdir(path))
+ ):
+ mod_name = file[: file.find(".py")] if file.endswith(".py") else file
+ module = importlib.import_module(__name__ + "." + mod_name)
diff --git a/fairseq/examples/adaptive_span/adagrad_with_grad_clip.py b/fairseq/examples/adaptive_span/adagrad_with_grad_clip.py
new file mode 100644
index 0000000000000000000000000000000000000000..585ce184ab2d6bbde0d2f7fcafd6536fa8f6d8b6
--- /dev/null
+++ b/fairseq/examples/adaptive_span/adagrad_with_grad_clip.py
@@ -0,0 +1,128 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+from torch.optim import Adagrad
+
+from fairseq.optim import LegacyFairseqOptimizer, register_optimizer
+
+
+@register_optimizer("adagrad_with_grad_clip")
+class FairseqAdagradWithGradClip(LegacyFairseqOptimizer):
+ def __init__(self, args, params):
+ super().__init__(args)
+ self._optimizer = AdagradWithGradClip(params, **self.optimizer_config)
+
+ @staticmethod
+ def add_args(parser):
+ """Add optimizer-specific arguments to the parser."""
+ # fmt: off
+ parser.add_argument('--weight-decay', '--wd', default=0.0, type=float, metavar='WD',
+ help='weight decay')
+ parser.add_argument('--adagrad-clip', default=0.0, type=float, metavar='D',
+ help='internal grad clip')
+ # fmt: on
+
+ @property
+ def optimizer_config(self):
+ """
+ Return a kwarg dictionary that will be used to override optimizer
+ args stored in checkpoints. This allows us to load a checkpoint and
+ resume training using a different set of optimizer args, e.g., with a
+ different learning rate.
+ """
+ return {
+ "lr": self.args.lr[0],
+ "weight_decay": self.args.weight_decay,
+ "grad_clip": self.args.adagrad_clip,
+ }
+
+ @property
+ def supports_flat_params(self):
+ return False
+
+
+def _clip_grad(clr, grad, group_grad_clip):
+ if group_grad_clip > 0:
+ norm = grad.norm(2).item()
+ if norm > group_grad_clip:
+ clr *= group_grad_clip / (norm + 1e-10)
+ return clr
+
+
+class AdagradWithGradClip(Adagrad):
+ """Adagrad algorithm with custom gradient clipping"""
+
+ def __init__(
+ self,
+ params,
+ lr=1e-2,
+ lr_decay=0,
+ weight_decay=0,
+ initial_accumulator_value=0,
+ grad_clip=0,
+ ):
+ Adagrad.__init__(
+ self,
+ params,
+ lr=lr,
+ lr_decay=lr_decay,
+ weight_decay=weight_decay,
+ initial_accumulator_value=initial_accumulator_value,
+ )
+ self.defaults["grad_clip"] = grad_clip
+ self.param_groups[0].setdefault("grad_clip", grad_clip)
+
+ def step(self, closure=None):
+ loss = None
+ if closure is not None:
+ loss = closure()
+
+ for group in self.param_groups:
+ for p in group["params"]:
+ if p.grad is None:
+ continue
+
+ grad = p.grad.data
+ state = self.state[p]
+
+ state["step"] += 1
+
+ if group["weight_decay"] != 0:
+ if p.grad.data.is_sparse:
+ raise RuntimeError(
+ "weight_decay option is "
+ "not compatible with sparse "
+ "gradients"
+ )
+ grad = grad.add(group["weight_decay"], p.data)
+
+ clr = group["lr"] / (1 + (state["step"] - 1) * group["lr_decay"])
+
+ # clip
+ clr = _clip_grad(clr=clr, grad=grad, group_grad_clip=group["grad_clip"])
+
+ if grad.is_sparse:
+ # the update is non-linear so indices must be unique
+ grad = grad.coalesce()
+ grad_indices = grad._indices()
+ grad_values = grad._values()
+ size = grad.size()
+
+ def make_sparse(values):
+ constructor = grad.new
+ if grad_indices.dim() == 0 or values.dim() == 0:
+ return constructor().resize_as_(grad)
+ return constructor(grad_indices, values, size)
+
+ state["sum"].add_(make_sparse(grad_values.pow(2)))
+ std = state["sum"]._sparse_mask(grad)
+ std_values = std._values().sqrt_().add_(1e-10)
+ p.data.add_(-clr, make_sparse(grad_values / std_values))
+ else:
+ state["sum"].addcmul_(1, grad, grad)
+ std = state["sum"].sqrt().add_(1e-10)
+ p.data.addcdiv_(-clr, grad, std)
+
+ return loss
diff --git a/fairseq/examples/adaptive_span/adaptive_span_attention.py b/fairseq/examples/adaptive_span/adaptive_span_attention.py
new file mode 100644
index 0000000000000000000000000000000000000000..07f757bb8e1a8a67b1124175ee338c8735aa8d65
--- /dev/null
+++ b/fairseq/examples/adaptive_span/adaptive_span_attention.py
@@ -0,0 +1,160 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+import math
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class AdaptiveMask(nn.Module):
+ """Soft masking function for adaptive size.
+ It masks out the last K values of an input. The masking value
+ goes from 1 to 0 gradually, so K can be learned with
+ back-propagation.
+ Args:
+ max_size: maximum size (i.e. input dimension)
+ ramp_size: size of the ramp going from 0 to 1
+ init_val: initial size proportion not to be masked out
+ shape: learn multiple sizes independent of each other
+ """
+
+ def __init__(self, max_size, ramp_size, init_val=0, shape=(1,)):
+ nn.Module.__init__(self)
+ self._max_size = max_size
+ self._ramp_size = ramp_size
+ self.current_val = nn.Parameter(torch.zeros(*shape) + init_val)
+ mask_template = torch.linspace(1 - max_size, 0, steps=max_size)
+ self.register_buffer("mask_template", mask_template)
+
+ def forward(self, x):
+ mask = self.mask_template.float() + self.current_val.float() * self._max_size
+ mask = mask / self._ramp_size + 1
+ mask = mask.clamp(0, 1)
+ if x.size(-1) < self._max_size:
+ # the input could have been trimmed beforehand to save computation
+ mask = mask.narrow(-1, self._max_size - x.size(-1), x.size(-1))
+ x = (x * mask).type_as(x)
+ return x
+
+ def get_current_max_size(self, include_ramp=True):
+ current_size = math.ceil(self.current_val.max().item() * self._max_size)
+ if include_ramp:
+ current_size += self._ramp_size
+ current_size = max(0, min(self._max_size, current_size))
+ return current_size
+
+ def get_current_avg_size(self, include_ramp=True):
+ current_size = math.ceil(
+ self.current_val.float().mean().item() * self._max_size
+ )
+ if include_ramp:
+ current_size += self._ramp_size
+ current_size = max(0, min(self._max_size, current_size))
+ return current_size
+
+ def clamp_param(self):
+ """this need to be called after each update"""
+ self.current_val.data.clamp_(0, 1)
+
+
+class AdaptiveSpan(nn.Module):
+ """Adaptive attention span for Transformerself.
+ This module learns an attention span length from data for each
+ self-attention head.
+ Args:
+ attn_span: maximum attention span
+ adapt_span_loss: loss coefficient for the span length
+ adapt_span_ramp: length of the masking ramp
+ adapt_span_init: initial size ratio
+ adapt_span_cache: adapt cache size to reduce memory usage
+ """
+
+ def __init__(
+ self,
+ attn_span,
+ adapt_span_ramp,
+ adapt_span_init,
+ n_head,
+ adapt_span_layer,
+ **kargs
+ ):
+ nn.Module.__init__(self)
+ self._max_span = attn_span
+ self._n_head = n_head
+ self._adapt_span_layer = adapt_span_layer
+ if self._adapt_span_layer:
+ self._mask = AdaptiveMask(
+ max_size=self._max_span,
+ ramp_size=adapt_span_ramp,
+ init_val=adapt_span_init,
+ )
+ else:
+ self._mask = AdaptiveMask(
+ max_size=self._max_span,
+ ramp_size=adapt_span_ramp,
+ init_val=adapt_span_init,
+ shape=(n_head, 1, 1),
+ )
+
+ def forward(self, attn, normalize=True):
+ """mask attention with the right span"""
+ # batch and head dimensions are merged together, so separate them first
+ self.clamp_param()
+ if self._adapt_span_layer:
+ attn = self._mask(attn)
+ else:
+ B = attn.size(0) # batch size
+ M = attn.size(1) # block size
+ attn = attn.reshape(B // self._n_head, self._n_head, M, -1)
+ attn = self._mask(attn)
+ attn = attn.view(B, M, -1)
+ return attn
+
+ def get_trim_len(self):
+ """how much of memory can be trimmed to reduce computation"""
+ L = self._max_span
+ trim_len = min(L - 1, L - self._mask.get_current_max_size())
+ # too fine granularity might be bad for the memory management
+ trim_len = math.floor(trim_len / 64) * 64
+ return trim_len
+
+ def trim_memory(self, query, key, value, key_pe):
+ """trim out unnecessary memory beforehand to reduce computation"""
+ trim_len = self.get_trim_len()
+ cache_size = key.size(1) - query.size(1)
+ trim_len_cache = trim_len - (self._max_span - cache_size)
+ if trim_len_cache > 0:
+ key = key[:, trim_len_cache:, :]
+ value = value[:, trim_len_cache:, :]
+ elif trim_len_cache < 0:
+ # cache is too short! this happens when validation resumes
+ # after a lot of updates.
+ key = F.pad(key, [0, 0, -trim_len_cache, 0])
+ value = F.pad(value, [0, 0, -trim_len_cache, 0])
+ if trim_len > 0:
+ if key_pe is not None:
+ key_pe = key_pe[:, :, trim_len:]
+ return key, value, key_pe
+
+ def get_cache_size(self):
+ """determine how long the cache should be"""
+ trim_len = self.get_trim_len()
+ # give a buffer of 64 steps since a span might increase
+ # in future updates
+ return min(self._max_span, self._max_span - trim_len + 64)
+
+ def get_loss(self):
+ """a loss term for regularizing the span length"""
+ return self._max_span * self._mask.current_val.float().mean()
+
+ def get_current_max_span(self):
+ return self._mask.get_current_max_size()
+
+ def get_current_avg_span(self):
+ return self._mask.get_current_avg_size()
+
+ def clamp_param(self):
+ self._mask.clamp_param()
diff --git a/fairseq/examples/adaptive_span/adaptive_span_loss.py b/fairseq/examples/adaptive_span/adaptive_span_loss.py
new file mode 100644
index 0000000000000000000000000000000000000000..fe95b0d949250087eb4712172f7e3e04559de547
--- /dev/null
+++ b/fairseq/examples/adaptive_span/adaptive_span_loss.py
@@ -0,0 +1,107 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import math
+from dataclasses import dataclass
+
+import torch.nn.functional as F
+from fairseq import utils
+from fairseq.logging import metrics
+from fairseq.criterions import register_criterion
+from fairseq.criterions.cross_entropy import CrossEntropyCriterion
+from fairseq.dataclass import FairseqDataclass
+from omegaconf import II
+
+
+@dataclass
+class AdaptiveSpanCriterionConfig(FairseqDataclass):
+ sentence_avg: bool = II("optimization.sentence_avg")
+
+
+@register_criterion("adaptive_span_loss", dataclass=AdaptiveSpanCriterionConfig)
+class AdaptiveSpanCriterion(CrossEntropyCriterion):
+ def __init__(self, task, sentence_avg):
+ super().__init__(task, sentence_avg)
+
+ def forward(self, model, sample, reduce=True):
+ """Compute the loss for the given sample.
+
+ Returns a tuple with three elements:
+ 1) the loss here is summed, different from the adaptive span code
+ 2) the sample size, which is used as the denominator for the gradient
+ 3) logging outputs to display while training
+ """
+ net_output = model(**sample["net_input"])
+ loss, aux_loss, avg_span, max_span = self.compute_loss(
+ model, net_output, sample, reduce=reduce
+ )
+ sample_size = (
+ sample["target"].size(0) if self.sentence_avg else sample["ntokens"]
+ )
+ loss /= sample_size
+ total_loss = loss + aux_loss
+ sample_size = 1
+
+ logging_output = {
+ "loss": loss.data,
+ "ntokens": sample["ntokens"],
+ "nsentences": sample["target"].size(0),
+ "sample_size": sample_size,
+ "total_loss": total_loss.data,
+ "avg_span": avg_span * sample_size,
+ "max_span": max_span * sample_size,
+ }
+ return total_loss, sample_size, logging_output
+
+ def compute_loss(self, model, net_output, sample, reduce=True):
+ loss, _ = super().compute_loss(model, net_output, sample, reduce)
+ aux_loss = model.get_aux_loss()
+ avg_span = model.get_current_avg_span()
+ max_span = model.get_current_max_span()
+ return loss, aux_loss, avg_span, max_span
+
+ @staticmethod
+ def reduce_metrics(logging_outputs) -> None:
+ """Aggregate logging outputs from data parallel training."""
+ loss_sum = sum(log.get("loss", 0) for log in logging_outputs)
+ ntokens = sum(log.get("ntokens", 0) for log in logging_outputs)
+ sample_size = sum(log.get("sample_size", 0) for log in logging_outputs)
+ total_loss_sum = sum(log.get("total_loss", 0) for log in logging_outputs)
+ avg_span_sum = sum(log.get("avg_span", 0) for log in logging_outputs)
+ max_span_sum = sum(log.get("max_span", 0) for log in logging_outputs)
+
+ # we divide by log(2) to convert the loss from base e to base 2
+ metrics.log_scalar(
+ "loss", loss_sum / sample_size / math.log(2), sample_size, round=3
+ )
+ metrics.log_scalar("avg_span", avg_span_sum / sample_size, sample_size, round=3)
+ metrics.log_scalar("max_span", max_span_sum / sample_size, sample_size, round=3)
+ # total loss contains the L1 norm on adaptive-span
+ metrics.log_scalar(
+ "total_loss",
+ total_loss_sum / sample_size / math.log(2),
+ sample_size,
+ round=3,
+ )
+ if sample_size != ntokens:
+ metrics.log_scalar(
+ "nll_loss", loss_sum / ntokens / math.log(2), ntokens, round=3
+ )
+ metrics.log_derived(
+ "ppl", lambda meters: utils.get_perplexity(meters["nll_loss"].avg)
+ )
+ else:
+ metrics.log_derived(
+ "ppl", lambda meters: utils.get_perplexity(meters["loss"].avg)
+ )
+
+ @staticmethod
+ def logging_outputs_can_be_summed() -> bool:
+ """
+ Whether the logging outputs returned by `forward` can be summed
+ across workers prior to calling `reduce_metrics`. Setting this
+ to True will improves distributed training speed.
+ """
+ return True
diff --git a/fairseq/examples/adaptive_span/adaptive_span_model.py b/fairseq/examples/adaptive_span/adaptive_span_model.py
new file mode 100644
index 0000000000000000000000000000000000000000..d96c95b85dbcf29e9384cc6d8d9630d2489991b2
--- /dev/null
+++ b/fairseq/examples/adaptive_span/adaptive_span_model.py
@@ -0,0 +1,263 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import math
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from fairseq.modules.layer_norm import LayerNorm
+
+from .adaptive_span_attention import AdaptiveSpan
+
+# Size notations:
+# B = batch_size, H = d_model, M = block_size, L = attn_span
+
+
+def _skew(X, pad_value):
+ """shift every row 1 step to right"""
+ # X = B x M x L
+ B, M, L = X.size()
+ X = F.pad(X, (0, M + 1), value=pad_value) # B x M x (L+M+1)
+ X = X.view(B, -1) # B x ML+MM+M
+ X = X[:, :-M] # B x ML+MM
+ X = X.view(B, M, M + L) # B x M x L+M
+ return X
+
+
+def _unskew(X):
+ """reverse _skew operation"""
+ # X = B x M x L+M
+ B, M, L = X.size()
+ L -= M
+ X = X.view(B, -1) # B x ML+MM
+ X = F.pad(X, (0, M)) # B x ML+MM+M
+ X = X.view(B, M, M + L + 1) # B x M x L+M+1
+ X = X[:, :, :L] # B x M x L
+ return X
+
+
+class SeqAttention(nn.Module):
+ """Sequential self-attention layer.
+ Each token will attend to its previous fixed number of steps.
+ Note that attention doesn't include the current step itself.
+ """
+
+ def __init__(self, d_model, n_head, attn_span, dropout, adapt_span_layer, **kargs):
+ nn.Module.__init__(self)
+ self.dropout = nn.Dropout(dropout)
+ self.d_model = d_model # size of a single head
+ self.attn_span = attn_span
+ self.adaptive_span = AdaptiveSpan(
+ attn_span=attn_span,
+ n_head=n_head,
+ adapt_span_layer=adapt_span_layer,
+ **kargs
+ )
+
+ def forward(self, query, key, value, key_pe):
+ # query size = B x M x H
+ # key, value sizes = B x (M+L) x H
+
+ key, value, key_pe = self.adaptive_span.trim_memory(query, key, value, key_pe)
+
+ # compute attention from context
+ # B x M (dest) x (M+L) (src)
+ attn_cont = torch.matmul(query, key.transpose(-1, -2))
+ attn_cont = _unskew(attn_cont) # B x M x L
+
+ # compute the effect of position embedding
+ attn_pos = torch.matmul(query, key_pe) # B x M x L_pos
+ attn = attn_cont + attn_pos
+
+ attn = attn / math.sqrt(self.d_model) # B x M X L_pos
+
+ attn = F.softmax(attn.float(), dim=-1).type_as(attn)
+
+ # trim attention lengths according to the learned span
+ attn = self.adaptive_span(attn)
+
+ attn = self.dropout(attn) # B x M X L_pos
+
+ attn_cont = _skew(attn, 0) # B x M X (L+M)
+ out = torch.matmul(attn_cont, value) # B x M x H
+ return out
+
+ def get_cache_size(self):
+ return self.adaptive_span.get_cache_size()
+
+
+class MultiHeadSeqAttention(nn.Module):
+ def __init__(self, d_model, n_head, **kargs):
+ nn.Module.__init__(self)
+ assert d_model % n_head == 0
+ self.n_head = n_head
+ self.head_dim = d_model // n_head
+ self.attn = SeqAttention(d_model=self.head_dim, n_head=n_head, **kargs)
+ self.proj_query = nn.Linear(d_model, d_model, bias=False)
+ nn.init.xavier_normal_(self.proj_query.weight)
+ self.proj_out = nn.Linear(d_model, d_model, bias=False)
+ nn.init.xavier_normal_(self.proj_out.weight)
+ self.proj_val = nn.Linear(d_model, d_model, bias=False)
+ nn.init.xavier_normal_(self.proj_val.weight)
+ self.proj_key = nn.Linear(d_model, d_model, bias=False)
+ nn.init.xavier_normal_(self.proj_key.weight)
+
+ def head_reshape(self, x):
+ K = self.n_head
+ D = self.head_dim
+ x = x.view(x.size()[:-1] + (K, D)) # B x (M+L) x K x D
+ x = x.transpose(1, 2).contiguous() # B x K x (M+L) x D
+ x = x.view(-1, x.size(-2), x.size(-1)) # B_K x (M+L) x D
+ return x
+
+ def forward(self, query, key, value, key_pe):
+ B = query.size(0)
+ K = self.n_head
+ D = self.head_dim
+ M = query.size(1)
+
+ query = self.proj_query(query)
+ query = self.head_reshape(query)
+ value = self.proj_val(value)
+ value = self.head_reshape(value)
+ key = self.proj_key(key)
+ key = self.head_reshape(key)
+
+ out = self.attn(query, key, value, key_pe) # B_K x M x D
+ out = out.view(B, K, M, D) # B x K x M x D
+ out = out.transpose(1, 2).contiguous() # B x M x K x D
+ out = out.view(B, M, -1) # B x M x K_D
+ out = self.proj_out(out)
+ return out
+
+
+class FeedForwardLayer(nn.Module):
+ def __init__(self, d_model, d_inner, dropout, **kargs):
+ nn.Module.__init__(self)
+ self.fc1 = nn.Linear(d_model, d_inner)
+ self.fc2 = nn.Linear(d_inner, d_model)
+ nn.init.xavier_uniform_(self.fc1.weight)
+ nn.init.xavier_uniform_(self.fc2.weight)
+ self.dropout = nn.Dropout(dropout)
+
+ def forward(self, h):
+ h1 = F.relu(self.fc1(h))
+ h1 = self.dropout(h1)
+ h2 = self.fc2(h1)
+ return h2
+
+
+class TransformerSeqLayer(nn.Module):
+ def __init__(self, d_model, **kargs):
+ nn.Module.__init__(self)
+ self.attn = MultiHeadSeqAttention(d_model=d_model, **kargs)
+ self.norm1 = LayerNorm(d_model)
+ self.ff = FeedForwardLayer(d_model=d_model, **kargs)
+ self.norm2 = LayerNorm(d_model)
+
+ def forward(self, h, h_cache, key_pe):
+ # h = B x M x H
+ # h_cache = B x L x H
+ h_all = torch.cat([h_cache, h], dim=1) # B x (M+L) x H
+ attn_out = self.attn(h, h_all, h_all, key_pe)
+ h = self.norm1(h + attn_out) # B x M x H
+ if self.ff is not None:
+ ff_out = self.ff(h)
+ out = self.norm2(h + ff_out) # B x M x H
+ else:
+ out = h
+ return out
+
+ def get_cache_size(self):
+ return self.attn.attn.get_cache_size()
+
+
+class TransformerSeq(nn.Module):
+ def __init__(
+ self,
+ vocab_size,
+ d_model,
+ n_head,
+ n_layer,
+ attn_span,
+ emb_dropout,
+ aux_loss_scaler,
+ adapt_span_layer,
+ **kargs
+ ):
+ nn.Module.__init__(self)
+ # token embeddings
+ self.in_emb = nn.Embedding(vocab_size, d_model)
+ nn.init.normal_(self.in_emb.weight, mean=0, std=d_model ** -0.5)
+ self.out_emb = nn.Linear(d_model, vocab_size)
+ self.aux_loss_scaler = aux_loss_scaler
+ if emb_dropout > 0:
+ self.emb_dropout = nn.Dropout(emb_dropout)
+ else:
+ self.emb_dropout = None
+ # position embeddings
+ self.key_pe = nn.Parameter(torch.randn(1, d_model // n_head, attn_span))
+
+ self.layers = nn.ModuleList()
+ self.layers.extend(
+ TransformerSeqLayer(
+ d_model=d_model,
+ n_head=n_head,
+ attn_span=attn_span,
+ adapt_span_layer=adapt_span_layer,
+ **kargs
+ )
+ for _ in range(n_layer)
+ )
+
+ def forward(self, x, h_cache, target=None):
+ # x size = B x M
+ block_size = x.size(1)
+ h = self.in_emb(x) # B x M x H
+ if self.emb_dropout is not None:
+ h = self.emb_dropout(h)
+
+ h_cache_next = []
+ for l, layer in enumerate(self.layers):
+ cache_size = layer.attn.attn.get_cache_size()
+ if cache_size > block_size:
+ h_cache_next_l = torch.cat(
+ [h_cache[l][:, -cache_size + block_size :, :], h], dim=1
+ ).detach()
+ else:
+ h_cache_next_l = h[:, -cache_size:, :].detach()
+ h_cache_next.append(h_cache_next_l)
+ h = layer(h, h_cache[l], self.key_pe) # B x M x H
+
+ if self.emb_dropout is not None:
+ h = self.emb_dropout(h)
+
+ out = F.log_softmax(self.out_emb(h).float(), dim=-1).type_as(h)
+ dummy_loss = None
+
+ return out, h_cache_next, dummy_loss
+
+ def get_aux_loss(self):
+ loss = 0.0
+ for layer in self.layers:
+ loss += layer.attn.attn.adaptive_span.get_loss()
+ return self.aux_loss_scaler * loss
+
+ def get_current_max_span(self):
+ max_span = 0.0
+ for layer in self.layers:
+ max_span = max(
+ max_span, layer.attn.attn.adaptive_span.get_current_max_span()
+ )
+ return max_span
+
+ def get_current_avg_span(self):
+ avg_span = 0.0
+ for layer in self.layers:
+ avg_span += layer.attn.attn.adaptive_span.get_current_avg_span()
+ return avg_span / len(self.layers)
diff --git a/fairseq/examples/adaptive_span/adaptive_span_model_wrapper.py b/fairseq/examples/adaptive_span/adaptive_span_model_wrapper.py
new file mode 100644
index 0000000000000000000000000000000000000000..5b147fe11f9d730438d036321a2d4a5d776efaa2
--- /dev/null
+++ b/fairseq/examples/adaptive_span/adaptive_span_model_wrapper.py
@@ -0,0 +1,145 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import logging
+from dataclasses import dataclass
+from typing import Dict, List, Optional
+
+import torch
+from fairseq.dataclass import FairseqDataclass
+from fairseq.models import (
+ FairseqIncrementalDecoder,
+ FairseqLanguageModel,
+ register_model,
+)
+from .adaptive_span_model import TransformerSeq as AdaptiveSpanTransformerModel
+
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class AdaptiveSpanSmallConfig(FairseqDataclass):
+ # defaults come from https://github.com/facebookresearch/adaptive-span/blob/master/experiments/enwik8_small.sh
+ vocab_size: int = 50
+ d_model: int = 256
+ n_head: int = 4
+ d_inner: int = 1024
+ n_layer: int = 8
+ attn_span: int = 1024
+ dropout: float = 0.0
+ emb_dropout: float = 0.0
+ adapt_span_ramp: int = 32
+ adapt_span_init: float = 0.0
+ aux_loss_scaler: float = 0.000002
+ adapt_span_layer: bool = False
+
+
+@register_model("adaptive_span", dataclass=AdaptiveSpanSmallConfig)
+class AdaptiveSpanTransformer(FairseqLanguageModel):
+ @classmethod
+ def build_model(cls, cfg: AdaptiveSpanSmallConfig, task):
+ return cls(AdaptiveSpanDecoder(cfg, task))
+
+ def get_aux_loss(self):
+ return self.decoder.get_aux_loss()
+
+ def get_current_max_span(self):
+ return self.decoder.get_current_max_span()
+
+ def get_current_avg_span(self):
+ return self.decoder.get_current_avg_span()
+
+
+class AdaptiveSpanDecoder(FairseqIncrementalDecoder):
+ def __init__(self, cfg, task):
+
+ super().__init__(task.target_dictionary)
+
+ self.config = cfg
+ config = AdaptiveSpanSmallConfig(
+ vocab_size=len(task.target_dictionary),
+ d_model=cfg.d_model,
+ n_head=cfg.n_head,
+ d_inner=cfg.d_inner,
+ n_layer=cfg.n_layer,
+ attn_span=cfg.attn_span,
+ dropout=cfg.dropout,
+ emb_dropout=cfg.emb_dropout,
+ adapt_span_ramp=cfg.adapt_span_ramp,
+ adapt_span_init=cfg.adapt_span_init,
+ aux_loss_scaler=cfg.aux_loss_scaler,
+ adapt_span_layer=cfg.adapt_span_layer,
+ )
+ logger.info(config)
+ self.model = AdaptiveSpanTransformerModel(**config.__dict__)
+
+ self._mems = None
+
+ def forward(
+ self,
+ src_tokens,
+ incremental_state: Optional[Dict[str, List[torch.Tensor]]] = None,
+ encoder_out=None,
+ ):
+ bsz = src_tokens.size(0)
+ if incremental_state is not None: # used during inference
+ mems = self.get_incremental_state("mems")
+ src_tokens = src_tokens[:, -1:] # only keep the most recent token
+ else:
+ mems = self._mems
+
+ if mems is None:
+ # first time init
+ mems = self.init_hid_cache(bsz)
+ output = self.model(x=src_tokens, h_cache=mems,)
+ if incremental_state is not None:
+ self.set_incremental_state(incremental_state, "mems", output[1])
+ else:
+ self._mems = output[1]
+ return (output[0],)
+
+ def max_positions(self):
+ return self.config.attn_span
+
+ def init_hid_cache(self, batch_sz):
+ hid = []
+ for layer in self.model.layers:
+ param = next(self.model.parameters())
+ h = torch.zeros(
+ batch_sz,
+ layer.get_cache_size(),
+ self.config.d_model,
+ dtype=param.dtype,
+ device=param.device,
+ )
+ hid.append(h)
+ return hid
+
+ def get_aux_loss(self):
+ return self.model.get_aux_loss()
+
+ def get_current_max_span(self):
+ return self.model.get_current_max_span()
+
+ def get_current_avg_span(self):
+ return self.model.get_current_avg_span()
+
+ def reorder_incremental_state(
+ self,
+ incremental_state: Dict[str, Dict[str, Optional[torch.Tensor]]],
+ new_order: torch.Tensor,
+ ):
+ """Reorder incremental state.
+
+ This will be called when the order of the input has changed from the
+ previous time step. A typical use case is beam search, where the input
+ order changes between time steps based on the selection of beams.
+ """
+ raise NotImplementedError("This is required for generation/beam search")
+ # mems = self.get_incremental_state(incremental_state, "mems")
+ # if mems is not None:
+ # new_mems = [mems_i.index_select(1, new_order) for mems_i in mems]
+ # self.set_incremental_state(incremental_state, "mems", new_mems)
diff --git a/fairseq/examples/adaptive_span/truncated_bptt_lm_task.py b/fairseq/examples/adaptive_span/truncated_bptt_lm_task.py
new file mode 100644
index 0000000000000000000000000000000000000000..9978481b6d95134ab609499586c609913aee35df
--- /dev/null
+++ b/fairseq/examples/adaptive_span/truncated_bptt_lm_task.py
@@ -0,0 +1,285 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import logging
+import os
+from dataclasses import dataclass, field
+from typing import List, Optional, Tuple
+
+import torch
+from fairseq import utils
+from fairseq.data import (
+ Dictionary,
+ TokenBlockDataset,
+ data_utils,
+ iterators,
+)
+from fairseq.dataclass import FairseqDataclass
+from fairseq.distributed import utils as dist_utils
+from fairseq.tasks import FairseqTask, register_task
+from omegaconf import II
+
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class TruncatedBPTTLMConfig(FairseqDataclass):
+ data: str = field(default="???", metadata={"help": "path to data directory"})
+ tokens_per_sample: int = field(
+ default=1024, metadata={"help": "max number of tokens per sequence"},
+ )
+ batch_size: int = II("dataset.batch_size")
+ # Some models use *max_target_positions* to know how many positional
+ # embeddings to learn. We use II(...) to make it default to
+ # *tokens_per_sample*, but in principle there could be more positional
+ # embeddings than tokens in a single batch. This may also be irrelevant for
+ # custom model implementations.
+ max_target_positions: int = II("task.tokens_per_sample")
+ # these will be populated automatically if not provided
+ data_parallel_rank: Optional[int] = None
+ data_parallel_size: Optional[int] = None
+
+
+@register_task("truncated_bptt_lm", dataclass=TruncatedBPTTLMConfig)
+class TruncatedBPTTLMTask(FairseqTask):
+ def __init__(self, cfg: TruncatedBPTTLMConfig):
+ super().__init__(cfg)
+
+ if cfg.data_parallel_rank is None or cfg.data_parallel_size is None:
+ if torch.distributed.is_initialized():
+ cfg.data_parallel_rank = dist_utils.get_data_parallel_rank()
+ cfg.data_parallel_size = dist_utils.get_data_parallel_world_size()
+ else:
+ cfg.data_parallel_rank = 0
+ cfg.data_parallel_size = 1
+
+ # load the dictionary
+ paths = utils.split_paths(cfg.data)
+ assert len(paths) > 0
+ self.dictionary = Dictionary.load(os.path.join(paths[0], "dict.txt"))
+ logger.info("dictionary: {} types".format(len(self.dictionary)))
+
+ def load_dataset(self, split, epoch=1, combine=False, **kwargs):
+ """Load a given dataset split (e.g., train, valid, test)"""
+
+ # support sharded datasets
+ paths = utils.split_paths(self.cfg.data)
+ assert len(paths) > 0
+ data_path = paths[(epoch - 1) % len(paths)]
+ split_path = os.path.join(data_path, split)
+
+ # each element of *data* will be a tensorized line from the original
+ # text dataset, similar to ``open(split_path).readlines()``
+ data = data_utils.load_indexed_dataset(
+ split_path, self.dictionary, combine=combine
+ )
+ if data is None:
+ raise FileNotFoundError(
+ "Dataset not found: {} ({})".format(split, split_path)
+ )
+
+ # this is similar to ``data.view(-1).split(tokens_per_sample)``
+ data = TokenBlockDataset(
+ data,
+ data.sizes,
+ block_size=self.cfg.tokens_per_sample,
+ pad=None, # unused
+ eos=None, # unused
+ break_mode="none",
+ )
+
+ self.datasets[split] = TruncatedBPTTDataset(
+ data=data,
+ bsz_per_shard=self.cfg.batch_size,
+ shard_id=self.cfg.data_parallel_rank,
+ num_shards=self.cfg.data_parallel_size,
+ )
+
+ def dataset(self, split):
+ return self.datasets[split]
+
+ def get_batch_iterator(
+ self,
+ dataset,
+ num_workers=0,
+ epoch=1,
+ data_buffer_size=0,
+ skip_remainder_batch=False,
+ **kwargs
+ ):
+ return iterators.EpochBatchIterator(
+ dataset=dataset,
+ collate_fn=self._collate_fn,
+ num_workers=num_workers,
+ epoch=epoch,
+ buffer_size=data_buffer_size,
+ # we don't use the batching functionality from EpochBatchIterator;
+ # instead every item in *dataset* is a whole batch
+ batch_sampler=[[i] for i in range(len(dataset))],
+ disable_shuffling=True,
+ skip_remainder_batch=skip_remainder_batch,
+ )
+
+ def _collate_fn(self, items: List[List[torch.Tensor]]):
+ # we don't use fairseq's batching functionality, so we expect a single
+ # Tensor of type List[torch.Tensor]
+ assert len(items) == 1
+
+ # item will have shape B x T (the last batch may have length < T)
+ id, item = items[0]
+ item = data_utils.collate_tokens(item, pad_idx=self.source_dictionary.pad())
+ B, T = item.size()
+
+ # shift item one position over and append a padding token for the target
+ target = torch.nn.functional.pad(
+ item[:, 1:], (0, 1, 0, 0), value=self.target_dictionary.pad()
+ )
+
+ # fairseq expects batches to have the following structure
+ return {
+ "id": torch.tensor([id] * item.size(0)),
+ "net_input": {"src_tokens": item,},
+ "target": target,
+ "nsentences": item.size(0),
+ "ntokens": item.numel(),
+ }
+
+ def build_dataset_for_inference(
+ self, src_tokens: List[torch.Tensor], src_lengths: List[int], **kwargs
+ ) -> torch.utils.data.Dataset:
+ eos = self.source_dictionary.eos()
+ dataset = TokenBlockDataset(
+ src_tokens,
+ src_lengths,
+ block_size=None, # ignored for "eos" break mode
+ pad=self.source_dictionary.pad(),
+ eos=eos,
+ break_mode="eos",
+ )
+
+ class Dataset(torch.utils.data.Dataset):
+ def __getitem__(self, i):
+ item = dataset[i]
+ if item[-1] == eos:
+ # remove eos to support generating with a prefix
+ item = item[:-1]
+ return (i, [item])
+
+ def __len__(self):
+ return len(dataset)
+
+ return Dataset()
+
+ def inference_step(
+ self, generator, models, sample, prefix_tokens=None, constraints=None
+ ):
+ with torch.no_grad():
+ if constraints is not None:
+ raise NotImplementedError
+
+ # SequenceGenerator doesn't use *src_tokens* directly, we need to
+ # pass the *prefix_tokens* argument instead.
+ if prefix_tokens is None and sample["net_input"]["src_tokens"].nelement():
+ prefix_tokens = sample["net_input"]["src_tokens"]
+
+ # begin generation with the end-of-sentence token
+ bos_token = self.source_dictionary.eos()
+
+ return generator.generate(
+ models, sample, prefix_tokens=prefix_tokens, bos_token=bos_token
+ )
+
+ def eval_lm_dataloader(
+ self,
+ dataset,
+ max_tokens: Optional[int] = 36000,
+ batch_size: Optional[int] = None,
+ max_positions: Optional[int] = None,
+ num_shards: int = 1,
+ shard_id: int = 0,
+ num_workers: int = 1,
+ data_buffer_size: int = 10,
+ context_window: int = 0,
+ ):
+ if context_window > 0:
+ raise NotImplementedError(
+ "Transformer-XL doesn't need --context-window, try "
+ "--model-overrides '{\"mem_len\":42}' instead "
+ )
+ return self.get_batch_iterator(
+ dataset=dataset,
+ max_tokens=max_tokens,
+ max_sentences=batch_size,
+ max_positions=max_positions,
+ ignore_invalid_inputs=True,
+ num_shards=num_shards,
+ shard_id=shard_id,
+ num_workers=num_workers,
+ data_buffer_size=data_buffer_size,
+ ).next_epoch_itr(shuffle=False)
+
+ @property
+ def source_dictionary(self):
+ return self.dictionary
+
+ @property
+ def target_dictionary(self):
+ return self.dictionary
+
+
+class TruncatedBPTTDataset(torch.utils.data.Dataset):
+ def __init__(
+ self,
+ data: List[torch.Tensor], # ordered list of items
+ bsz_per_shard, # number of items processed per GPUs per forward
+ shard_id, # current GPU ID
+ num_shards, # number of GPUs
+ ):
+ super().__init__()
+ self.data = data
+
+ def batchify(data, bsz):
+ # Work out how cleanly we can divide the dataset into bsz parts.
+ nbatch = data.size(0) // bsz
+ # Trim off any extra elements that wouldn't cleanly fit (remainders).
+ data = data.narrow(0, 0, nbatch * bsz)
+ # Evenly divide the data across the bsz batches.
+ data = data.view(bsz, -1).contiguous()
+ return data
+
+ # total number of sequences processed by all GPUs in each forward pass
+ global_batch_size = bsz_per_shard * num_shards
+
+ """
+ With a 16 item dataset, bsz_per_shard=2 and num_shards=3,
+ *indices* might look like:
+
+ indices = [[0, 1],
+ [2, 3],
+ [4, 5],
+ [6, 7],
+ [8, 9],
+ [10, 11]]
+
+ The size of the TruncatedBPTTDataset instance will be 2,
+ and shard 1 will see items:
+
+ [(0, [data[4], data[6]]),
+ (1, [data[5], data[7]])]
+ """
+ indices = batchify(torch.arange(len(data)), global_batch_size)
+ assert indices.size(0) == global_batch_size
+
+ self.my_indices = indices[
+ shard_id * bsz_per_shard : (shard_id + 1) * bsz_per_shard
+ ]
+ assert self.my_indices.size(0) == bsz_per_shard
+
+ def __len__(self):
+ return self.my_indices.size(1)
+
+ def __getitem__(self, i) -> Tuple[int, List[torch.Tensor]]:
+ return (i, [self.data[idx] for idx in self.my_indices[:, i]])
diff --git a/fairseq/examples/attention_head_selection/README.md b/fairseq/examples/attention_head_selection/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..2434f1fb212fa075335f97507dc89508f29402ec
--- /dev/null
+++ b/fairseq/examples/attention_head_selection/README.md
@@ -0,0 +1,161 @@
+# Pay Better Attention to Attention: Head Selection in Multilingual and Multi-Domain Sequence Modeling (Gong et al., 2021)
+
+[https://arxiv.org/pdf/2106.10840.pdf](https://arxiv.org/pdf/2106.10840.pdf)
+
+## Introduction
+
+We present attention head selection strategies in multilingual and multi-domain sequence modeling including text translation, speech recognition and speech translation tasks.
+
+Below is an example of training multilingual/multi-domain speech recognition models.
+
+## Data Preparation
+Prepare mTEDx data as in [mTEDx example](https://github.com/fairinternal/fairseq-py/blob/0d9c5851e6fac40f9e366b3633ccd615c2901788/examples/speech_to_text/docs/mtedx_example.md) and CoVoST data as in [CoVoST example](https://github.com/fairinternal/fairseq-py/blob/0d9c5851e6fac40f9e366b3633ccd615c2901788/examples/speech_to_text/docs/covost_example.md). Similarly prepare EuroParl data.
+
+
+## Training a multilingual ASR model with attention head selection
+
+```bash
+data_dir=
+train_subset="train_ar_ar_tedx,train_de_de_tedx,train_el_el_tedx,train_es_es_tedx,train_fr_fr_tedx,train_it_it_tedx,train_pt_pt_tedx,train_ru_ru_tedx"
+valid_subset="valid_ar_ar_tedx,valid_de_de_tedx,valid_el_el_tedx,valid_es_es_tedx,valid_fr_fr_tedx,valid_it_it_tedx,valid_pt_pt_tedx,valid_ru_ru_tedx"
+strateg=
+
+fairseq-train ${data_dir} \
+ --user-dir examples/attention_head_selection/src \
+ --train-subset "${train_subset}" \
+ --valid-subset "${valid_subset}" \
+ --config-yaml 'config_asr.yaml' \
+ --arch 'head_selection_s2t_transformer_s' \
+ --task 'speech_to_text_head_selection' \
+ --criterion label_smoothed_cross_entropy --label-smoothing 0.1 \
+ --lr-scheduler 'inverse_sqrt' --stop-min-lr -1.0 --warmup-updates 10000 \
+ --lr 5e-4 \
+ --clip-norm 10.0 \
+ --seed 1 \
+ --max-epoch 400 \
+ --max-tokens 32000 \
+ --ignore-prefix-size 1 \
+ --dropout 0.3 \
+ --optimizer adam --adam-eps 1e-06 --adam-betas '(0.9, 0.98)' \
+ --skip-invalid-size-inputs-valid-test \
+ --encoder-attn-head-select \
+ --total-encoder-attention-heads 8 \
+ --decoder-self-attn-head-select \
+ --total-decoder-attention-heads 8 \
+ --attn-head-select-strategy ${strategy} \
+ --task-type lang \
+```
+
+## Training a multi-domain ASR model with attention head selection
+
+```bash
+data_dir=
+train_subset="train_es_es_tedx,train_fr_fr_tedx,train_pt_pt_tedx,train_it_it_tedx,train_ru_ru_tedx,train_el_el_tedx,train_ar_ar_tedx,train_de_de_tedx,train_ar_ar_cv,train_de_de_cv,train_es_es_cv,train_fr_fr_cv,train_it_it_cv,train_pt_pt_cv,train_ru_ru_cv,train_de_de_ep,train_es_es_ep,train_fr_fr_ep,train_it_it_ep,train_pt_pt_ep"
+valid_subset="dev_es_es_tedx,dev_fr_fr_tedx,dev_pt_pt_tedx,dev_it_it_tedx,dev_ru_ru_tedx,dev_el_el_tedx,dev_ar_ar_tedx,dev_de_de_tedx,dev_ar_ar_cv,dev_de_de_cv,dev_es_es_cv,dev_fr_fr_cv,dev_it_it_cv,dev_pt_pt_cv,dev_ru_ru_cv,dev_de_de_ep,dev_es_es_ep,dev_fr_fr_ep,dev_it_it_ep,dev_pt_pt_ep"
+strateg=
+
+fairseq-train ${data_dir} \
+ --user-dir examples/attention_head_selection/src \
+ --train-subset "${train_subset}" \
+ --valid-subset "${valid_subset}" \
+ --config-yaml 'config_asr.yaml' \
+ --arch head_selection_s2t_transformer_s \
+ --task speech_to_text_head_selection \
+ --criterion label_smoothed_cross_entropy --label-smoothing 0.1 \
+ --lr-scheduler 'inverse_sqrt' --stop-min-lr -1.0 --warmup-updates 10000 \
+ --lr 5e-4 \
+ --clip-norm 10.0 \
+ --seed 1 \
+ --max-epoch 400 \
+ --max-tokens 32000 \
+ --ignore-prefix-size 1 \
+ --dropout 0.3 \
+ --optimizer adam --adam-eps 1e-06 --adam-betas '(0.9, 0.98)' \
+ --skip-invalid-size-inputs-valid-test \
+ --encoder-attn-head-select \
+ --total-encoder-attention-heads 8 \
+ --decoder-self-attn-head-select \
+ --total-decoder-attention-heads 8 \
+ --attn-head-select-strategy ${strategy} \
+ --task-type domain
+```
+
+## Inference in multilingual setting
+
+```bash
+MODEL_DIR=
+data_dir=
+gen_subset=
+train_subset="train_ar_ar_tedx,train_de_de_tedx,train_el_el_tedx,train_es_es_tedx,train_fr_fr_tedx,train_it_it_tedx,train_pt_pt_tedx,train_ru_ru_tedx"
+last_n=10
+CHECKPOINT_FILENAME="avg_last_${last_n}_checkpoint.pt"
+CHECKPOINT="_avg"
+RESULTS="${MODEL_DIR}/ckpt${CHECKPOINT}"
+if [ ! -d $RESULTS ]; then
+ mkdir -p $RESULTS
+fi;
+
+python scripts/average_checkpoints.py \
+ --inputs ${MODEL_DIR} --num-epoch-checkpoints ${last_n} \
+ --output "${MODEL_DIR}/${CHECKPOINT_FILENAME}"
+
+fairseq-generate ${data_dir} \
+ --user-dir examples/attention_head_selection/src \
+ --arch 'head_selection_s2t_transformer_s' \
+ --task 'speech_to_text_head_selection' \
+ --train-subset ${train_subset} \
+ --gen-subset ${gen_subset} \
+ --path "${MODEL_DIR}/${CHECKPOINT_FILENAME}" \
+ --config-yaml 'config_asr.yaml' \
+ --prefix-size 1 \
+ --max-tokens 40000 --beam 5 \
+ --skip-invalid-size-inputs-valid-test \
+ --results-path ${RESULTS} \
+ --scoring wer --wer-tokenizer 13a \
+ --wer-lowercase --wer-remove-punct --remove-bpe
+```
+
+## Inference in multi-domain setting
+
+```bash
+MODEL_DIR=
+data_dir=
+gen_subset=
+train_subset="train_es_es_tedx,train_fr_fr_tedx,train_pt_pt_tedx,train_it_it_tedx,train_ru_ru_tedx,train_el_el_tedx,train_ar_ar_tedx,train_de_de_tedx,train_ar_ar_cv,train_de_de_cv,train_es_es_cv,train_fr_fr_cv,train_it_it_cv,train_pt_pt_cv,train_ru_ru_cv,train_de_de_ep,train_es_es_ep,train_fr_fr_ep,train_it_it_ep,train_pt_pt_ep"
+last_n=10
+CHECKPOINT_FILENAME="avg_last_${last_n}_checkpoint.pt"
+CHECKPOINT="_avg"
+RESULTS="${MODEL_DIR}/ckpt${CHECKPOINT}"
+if [ ! -d $RESULTS ]; then
+ mkdir -p $RESULTS
+fi;
+
+python scripts/average_checkpoints.py \
+ --inputs ${MODEL_DIR} --num-epoch-checkpoints ${last_n} \
+ --output "${MODEL_DIR}/${CHECKPOINT_FILENAME}"
+
+fairseq-generate ${data_dir} \
+ --user-dir examples/attention_head_selection/src \
+ --arch 'head_selection_s2t_transformer_s' \
+ --task 'speech_to_text_head_selection' \
+ --train-subset ${train_subset} \
+ --gen-subset ${gen_subset} \
+ --path "${MODEL_DIR}/${CHECKPOINT_FILENAME}" \
+ --config-yaml 'config_asr.yaml' \
+ --prefix-size 1 \
+ --max-tokens 40000 --beam 5 \
+ --skip-invalid-size-inputs-valid-test \
+ --results-path ${RESULTS} \
+ --scoring wer --wer-tokenizer 13a \
+ --wer-lowercase --wer-remove-punct --remove-bpe
+```
+
+## Citation
+```bibtex
+@article{gong2021pay,
+ title={Pay Better Attention to Attention: Head Selection in Multilingual and Multi-Domain Sequence Modeling},
+ author={Gong, Hongyu and Tang, Yun and Pino, Juan and Li, Xian},
+ journal={arXiv preprint arXiv:2106.10840},
+ year={2021}
+}
+'''
diff --git a/fairseq/examples/attention_head_selection/src/__init__.py b/fairseq/examples/attention_head_selection/src/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/fairseq/examples/attention_head_selection/src/data/__init__.py b/fairseq/examples/attention_head_selection/src/data/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/fairseq/examples/attention_head_selection/src/data/speech_to_text_dataset_with_domain.py b/fairseq/examples/attention_head_selection/src/data/speech_to_text_dataset_with_domain.py
new file mode 100644
index 0000000000000000000000000000000000000000..1f1823a7ac7c92c7a0e93fade59162b247f5db36
--- /dev/null
+++ b/fairseq/examples/attention_head_selection/src/data/speech_to_text_dataset_with_domain.py
@@ -0,0 +1,242 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import logging
+from pathlib import Path
+from typing import Dict, List, Optional
+from dataclasses import dataclass
+
+import torch
+from fairseq.data import (
+ ConcatDataset,
+ Dictionary,
+ FairseqDataset,
+ ResamplingDataset
+)
+from fairseq.data.audio.data_cfg import S2TDataConfig
+from fairseq.data.audio.speech_to_text_dataset import (
+ SpeechToTextDatasetItem,
+ SpeechToTextDataset,
+ SpeechToTextDatasetCreator
+)
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class SpeechToTextDatasetItemWithDomain(SpeechToTextDatasetItem):
+ src_lang_id: Optional[torch.Tensor] = None
+ tgt_lang_id: Optional[torch.Tensor] = None
+ domain_id: Optional[torch.Tensor] = None
+
+
+class SpeechToTextDatasetWithDomain(SpeechToTextDataset):
+
+ def __init__(
+ self,
+ split: str,
+ is_train_split: bool,
+ cfg: S2TDataConfig,
+ audio_paths: List[str],
+ n_frames: List[int],
+ src_texts: Optional[List[str]] = None,
+ tgt_texts: Optional[List[str]] = None,
+ speakers: Optional[List[str]] = None,
+ src_langs: Optional[List[str]] = None,
+ tgt_langs: Optional[List[str]] = None,
+ ids: Optional[List[str]] = None,
+ tgt_dict: Optional[Dictionary] = None,
+ pre_tokenizer=None,
+ bpe_tokenizer=None,
+ n_frames_per_step=1,
+ speaker_to_id=None,
+ src_lang_ids: Optional[List[int]] = None,
+ tgt_lang_ids: Optional[List[int]] = None,
+ domain_ids: Optional[List[int]] = None
+ ):
+ super().__init__(
+ split, is_train_split, cfg, audio_paths, n_frames,
+ src_texts, tgt_texts, speakers, src_langs, tgt_langs,
+ ids, tgt_dict, pre_tokenizer, bpe_tokenizer,
+ n_frames_per_step, speaker_to_id
+ )
+ assert src_lang_ids is None or len(src_lang_ids) == self.n_samples
+ assert tgt_lang_ids is None or len(tgt_lang_ids) == self.n_samples
+ assert domain_ids is None or len(domain_ids) == self.n_samples
+
+ self.src_lang_ids = src_lang_ids
+ self.tgt_lang_ids = tgt_lang_ids
+ self.domain_ids = domain_ids
+
+ def __getitem__(self, index: int) -> SpeechToTextDatasetItemWithDomain:
+ item = super().__getitem__(index)
+ src_lang_id = self.src_lang_ids[index]
+ tgt_lang_id = self.tgt_lang_ids[index]
+ domain_id = self.domain_ids[index]
+ return SpeechToTextDatasetItemWithDomain(
+ index=item.index, source=item.source,
+ target=item.target, speaker_id=item.speaker_id,
+ src_lang_id=src_lang_id,
+ tgt_lang_id=tgt_lang_id,
+ domain_id=domain_id
+ )
+
+ def collater(
+ self, samples: List[SpeechToTextDatasetItem], return_order: bool = False
+ ) -> Dict:
+ if len(samples) == 0:
+ return {}
+ out = super().collater(samples, return_order=True)
+ order = out["order"]
+ src_lang_ids = torch.tensor([x.src_lang_id for x in samples], dtype=torch.long).index_select(0, order)
+ tgt_lang_ids = torch.tensor([x.tgt_lang_id for x in samples], dtype=torch.long).index_select(0, order)
+ domain_ids = torch.tensor([x.domain_id for x in samples], dtype=torch.long).index_select(0, order)
+
+ out["src_lang_ids"] = src_lang_ids
+ out["tgt_lang_ids"] = tgt_lang_ids
+ out["domain_ids"] = domain_ids
+ if not return_order:
+ del out["order"]
+ return out
+
+
+class SpeechToTextDatasetCreatorWithDomain(SpeechToTextDatasetCreator):
+ KEY_SRC_LANG_ID, KEY_TGT_LANG_ID = "src_lang_id", "tgt_lang_id"
+ KEY_DOMAIN_ID = "domain_id"
+ # default values
+ DEFAULT_SRC_LANG_ID, DEFAULT_TGT_LANG_ID, DEFAULT_DOMAIN_ID = 0, 0, 0
+
+ @classmethod
+ def _from_list(
+ cls,
+ split_name: str,
+ is_train_split,
+ samples: List[Dict],
+ cfg: S2TDataConfig,
+ tgt_dict,
+ pre_tokenizer,
+ bpe_tokenizer,
+ n_frames_per_step,
+ speaker_to_id
+ ) -> SpeechToTextDatasetWithDomain:
+ audio_root = Path(cfg.audio_root)
+ ids = [s[cls.KEY_ID] for s in samples]
+ audio_paths = [(audio_root / s[cls.KEY_AUDIO]).as_posix() for s in samples]
+ n_frames = [int(s[cls.KEY_N_FRAMES]) for s in samples]
+ tgt_texts = [s[cls.KEY_TGT_TEXT] for s in samples]
+ src_texts = [s.get(cls.KEY_SRC_TEXT, cls.DEFAULT_SRC_TEXT) for s in samples]
+ speakers = [s.get(cls.KEY_SPEAKER, cls.DEFAULT_SPEAKER) for s in samples]
+ src_langs = [s.get(cls.KEY_SRC_LANG, cls.DEFAULT_LANG) for s in samples]
+ tgt_langs = [s.get(cls.KEY_TGT_LANG, cls.DEFAULT_LANG) for s in samples]
+ src_lang_ids = [s.get(cls.KEY_SRC_LANG_ID, cls.DEFAULT_SRC_LANG_ID) for s in samples]
+ tgt_lang_ids = [s.get(cls.KEY_TGT_LANG_ID, cls.DEFAULT_TGT_LANG_ID) for s in samples]
+ domain_ids = [s.get(cls.KEY_DOMAIN_ID, cls.DEFAULT_DOMAIN_ID) for s in samples]
+ return SpeechToTextDatasetWithDomain(
+ split_name,
+ is_train_split,
+ cfg,
+ audio_paths,
+ n_frames,
+ src_texts=src_texts,
+ tgt_texts=tgt_texts,
+ speakers=speakers,
+ src_langs=src_langs,
+ tgt_langs=tgt_langs,
+ ids=ids,
+ tgt_dict=tgt_dict,
+ pre_tokenizer=pre_tokenizer,
+ bpe_tokenizer=bpe_tokenizer,
+ n_frames_per_step=n_frames_per_step,
+ speaker_to_id=speaker_to_id,
+ src_lang_ids=src_lang_ids,
+ tgt_lang_ids=tgt_lang_ids,
+ domain_ids=domain_ids
+ )
+
+ @classmethod
+ def _load_samples_from_tsv(
+ cls,
+ root: str,
+ split: str,
+ src_lang_map,
+ tgt_lang_map,
+ domain_map
+ ):
+ # metadata from split
+ _, src_lang, tgt_lang, domain = split.split("_")
+ src_lang_id = src_lang_map[src_lang]
+ tgt_lang_id = tgt_lang_map[tgt_lang]
+ domain_id = domain_map[domain]
+
+ samples = SpeechToTextDatasetCreator._load_samples_from_tsv(root, split)
+ for s in samples:
+ s.update({
+ cls.KEY_SRC_LANG_ID: src_lang_id,
+ cls.KEY_TGT_LANG_ID: tgt_lang_id,
+ cls.KEY_DOMAIN_ID: domain_id
+ })
+ return samples
+
+ @classmethod
+ def _from_tsv(
+ cls,
+ root: str,
+ cfg: S2TDataConfig,
+ split: str,
+ tgt_dict,
+ is_train_split: bool,
+ pre_tokenizer,
+ bpe_tokenizer,
+ n_frames_per_step,
+ speaker_to_id,
+ src_lang_map: Dict[str, int],
+ tgt_lang_map: Dict[str, int],
+ domain_map: Dict[str, int]
+ ) -> SpeechToTextDatasetItemWithDomain:
+ samples = cls._load_samples_from_tsv(
+ root, split, src_lang_map,
+ tgt_lang_map, domain_map
+ )
+ return cls._from_list(
+ split, is_train_split, samples, cfg, tgt_dict, pre_tokenizer,
+ bpe_tokenizer, n_frames_per_step, speaker_to_id
+ )
+
+ @classmethod
+ def from_tsv(
+ cls,
+ root: str,
+ cfg: S2TDataConfig,
+ splits: str,
+ tgt_dict,
+ pre_tokenizer,
+ bpe_tokenizer,
+ is_train_split: bool,
+ epoch: int,
+ seed: int,
+ src_lang_map: Dict[str, int],
+ tgt_lang_map: Dict[str, int],
+ domain_map: Dict[str, int],
+ n_frames_per_step: int = 1,
+ speaker_to_id=None
+ ) -> SpeechToTextDatasetWithDomain:
+ datasets = [
+ cls._from_tsv(
+ root, cfg, split, tgt_dict, is_train_split, pre_tokenizer, bpe_tokenizer, n_frames_per_step, speaker_to_id, src_lang_map, tgt_lang_map, domain_map
+ )
+ for split in splits.split(",")
+ ]
+
+ if is_train_split and len(datasets) > 1 and cfg.sampling_alpha != 1.0:
+ # temperature-based sampling
+ size_ratios = cls.get_size_ratios(datasets, alpha=cfg.sampling_alpha)
+ datasets = [
+ ResamplingDataset(
+ d, size_ratio=r, seed=seed, epoch=epoch, replace=(r >= 1.0)
+ )
+ for r, d in zip(size_ratios, datasets)
+ ]
+
+ return ConcatDataset(datasets) if len(datasets) > 1 else datasets[0]
diff --git a/fairseq/examples/attention_head_selection/src/loss/__init__.py b/fairseq/examples/attention_head_selection/src/loss/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/fairseq/examples/attention_head_selection/src/loss/attention_head_selection.py b/fairseq/examples/attention_head_selection/src/loss/attention_head_selection.py
new file mode 100644
index 0000000000000000000000000000000000000000..4ba33954d0171572111eca94ef39e0e9a683e0ed
--- /dev/null
+++ b/fairseq/examples/attention_head_selection/src/loss/attention_head_selection.py
@@ -0,0 +1,27 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import math
+
+import torch
+from torch.nn.modules.loss import _Loss
+
+
+class HeadSelectionLoss(_Loss):
+
+ def __init__(self, args):
+ super().__init__()
+ self.args = args
+ self.kl_weight = getattr(args, "kl_weight", 0.0)
+
+ def forward(self, head_samples, sample_sizes, prior=0.5, eps=1e-7):
+ """
+ head_scores: (num_tasks, num_layers, num_heads)
+ sample_sizes: (num_tasks, )
+ """
+ kl_loss = (head_samples * (torch.log(head_samples + eps) - math.log(prior))).sum(-1).sum(-1)
+ kl_loss /= (torch.numel(head_samples) / head_samples.size(0))
+ kl_loss = self.kl_weight * torch.matmul(kl_loss, sample_sizes)
+ return kl_loss
diff --git a/fairseq/examples/attention_head_selection/src/models/__init__.py b/fairseq/examples/attention_head_selection/src/models/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/fairseq/examples/attention_head_selection/src/models/head_selection_s2t_transformer.py b/fairseq/examples/attention_head_selection/src/models/head_selection_s2t_transformer.py
new file mode 100644
index 0000000000000000000000000000000000000000..2c7ed89e89d2a902bd8419a735676941ece125f1
--- /dev/null
+++ b/fairseq/examples/attention_head_selection/src/models/head_selection_s2t_transformer.py
@@ -0,0 +1,170 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import logging
+from typing import Dict, List, Optional
+from pathlib import Path
+import torch.nn as nn
+from torch import Tensor
+from fairseq import checkpoint_utils
+
+from fairseq.models import register_model, register_model_architecture
+from fairseq.utils import safe_hasattr
+from fairseq.models.speech_to_text.s2t_transformer import (
+ S2TTransformerModel,
+ S2TTransformerEncoder,
+ TransformerDecoderScriptable
+)
+from fairseq.models.speech_to_text.s2t_transformer import base_architecture as s2t_base_architecture
+
+from ..modules.attn_head_selector import AttnHeadSelector
+from ..modules.head_selection_transformer_layer import HeadSelectionTransformerEncoderLayer
+from .head_selection_transformer import HeadSelectionTransformerDecoder
+
+
+logger = logging.getLogger(__name__)
+
+
+@register_model("head_selection_s2t_transformer")
+class HeadSelectionS2TTransformerModel(S2TTransformerModel):
+ """
+ Head selection implemented in S2TTransformer
+ """
+ def __init__(self, encoder, decoder):
+ super().__init__(encoder, decoder)
+
+ @staticmethod
+ def add_args(parser):
+ S2TTransformerModel.add_args(parser)
+ # encoder head selection
+ parser.add_argument(
+ "--encoder-attn-head-select",
+ action="store_true",
+ default=False,
+ help="encoder head selection"
+ )
+ parser.add_argument(
+ "--total-encoder-attention-heads",
+ type=int,
+ help="total number of encoder attention heads"
+ )
+ # decoder self attention selection
+ parser.add_argument(
+ "--decoder-self-attn-head-select",
+ action="store_true",
+ default=False,
+ help="decoder self-attention head selection"
+ )
+ # decoder-encoder attention selection
+ parser.add_argument(
+ "--dec-enc-attn-head-select",
+ action="store_true",
+ default=False,
+ help="decoder-encoder attention head selection"
+ )
+ parser.add_argument(
+ "--total-decoder-attention-heads",
+ type=int,
+ help="total number of decoder attention heads"
+ )
+ # selection strategy
+ parser.add_argument(
+ "--attn-head-select-strategy",
+ type=str,
+ help="attention head selection strategy, subset or group"
+ )
+
+ @classmethod
+ def build_encoder(cls, args):
+ if safe_hasattr(args, "encoder_attn_head_select") and args.encoder_attn_head_select:
+ encoder = HeadSelectionS2TTransformerEncoder(args)
+ else:
+ encoder = S2TTransformerEncoder(args)
+ pretraining_path = getattr(args, "load_pretrained_encoder_from", None)
+ if pretraining_path is not None:
+ if not Path(pretraining_path).exists():
+ logger.warning(
+ f"skipped pretraining because {pretraining_path} does not exist"
+ )
+ else:
+ encoder = checkpoint_utils.load_pretrained_component_from_model(
+ component=encoder, checkpoint=pretraining_path
+ )
+ logger.info(f"loaded pretrained encoder from: {pretraining_path}")
+ return encoder
+
+ @classmethod
+ def build_decoder(cls, args, task, embed_tokens):
+ if (safe_hasattr(args, "decoder_self_attn_head_select") and args.decoder_self_attn_head_select) or (safe_hasattr(args, "dec_enc_attn_head_select") and args.dec_enc_attn_head_select):
+ return HeadSelectionTransformerDecoderScriptable(args, task.target_dictionary, embed_tokens)
+ else:
+ return TransformerDecoderScriptable(args, task.target_dictionary, embed_tokens)
+
+
+class HeadSelectionS2TTransformerEncoder(S2TTransformerEncoder):
+
+ def __init__(self, args):
+ super().__init__(args)
+ self.attn_head_selector = AttnHeadSelector(
+ args.encoder_tasks,
+ args.encoder_layers,
+ args.total_encoder_attention_heads,
+ args.encoder_attention_heads,
+ args.attn_head_select_strategy,
+ )
+ self.task_ids = None
+ self.transformer_layers = nn.ModuleList([
+ HeadSelectionTransformerEncoderLayer(args, layer_idx, attn_head_selector=self.attn_head_selector) for layer_idx in range(args.encoder_layers)
+ ])
+
+ def set_task_ids(self, task_ids):
+ self.task_ids = task_ids
+
+ def _forward(self, src_tokens, src_lengths, return_all_hiddens=False):
+ self.attn_head_selector.head_select(self.task_ids)
+ return super()._forward(src_tokens, src_lengths, return_all_hiddens)
+
+
+class HeadSelectionTransformerDecoderScriptable(HeadSelectionTransformerDecoder):
+ def extract_features(
+ self,
+ prev_output_tokens,
+ encoder_out: Optional[Dict[str, List[Tensor]]] = None,
+ incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None,
+ full_context_alignment: bool = False,
+ alignment_layer: Optional[int] = None,
+ alignment_heads: Optional[int] = None,
+ ):
+ # call scriptable method from parent class
+ x, _ = self.extract_features_scriptable(
+ prev_output_tokens,
+ encoder_out,
+ incremental_state,
+ full_context_alignment,
+ alignment_layer,
+ alignment_heads,
+ )
+ return x, None
+
+
+@register_model_architecture(model_name="head_selection_s2t_transformer", arch_name="head_selection_s2t_transformer")
+def base_architecture(args):
+ s2t_base_architecture(args)
+ args.encoder_attn_head_select = getattr(args, "encoder_attn_head_select", False)
+ args.decoder_self_attn_head_select = getattr(args, "decoder_self_attn_head_select", False)
+ args.dec_enc_attn_head_select = getattr(args, "dec_enc_attn_head_select", False)
+ args.total_encoder_attention_heads = getattr(args, "total_encoder_attention_heads", 8)
+ args.total_decoder_attention_heads = getattr(args, "total_decoder_attention_heads", 8)
+ args.attn_head_select_strategy = getattr(args, "attn_head_select_strategy", "group")
+
+
+@register_model_architecture("head_selection_s2t_transformer", "head_selection_s2t_transformer_s")
+def head_selection_s2t_transformer_s(args):
+ args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 256)
+ args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 256 * 8)
+ args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 4)
+ args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 4)
+ args.dropout = getattr(args, "dropout", 0.1)
+ base_architecture(args)
diff --git a/fairseq/examples/attention_head_selection/src/models/head_selection_transformer.py b/fairseq/examples/attention_head_selection/src/models/head_selection_transformer.py
new file mode 100644
index 0000000000000000000000000000000000000000..b9d595699db62b247ad3dd82edac366f53e324bb
--- /dev/null
+++ b/fairseq/examples/attention_head_selection/src/models/head_selection_transformer.py
@@ -0,0 +1,215 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+from typing import Any, List, Dict, Optional
+import torch
+import torch.nn as nn
+from torch import Tensor
+
+from fairseq.utils import safe_hasattr
+from fairseq.models.transformer import (
+ TransformerModel,
+ TransformerEncoder,
+ TransformerDecoder
+)
+
+from ..modules.attn_head_selector import AttnHeadSelector
+from ..modules.head_selection_transformer_layer import (
+ HeadSelectionTransformerEncoderLayer,
+ HeadSelectionTransformerDecoderLayer
+)
+
+
+class HeadSelectionTransformerModel(TransformerModel):
+ def __init__(self, args, encoder, decoder):
+ super().__init__(args, encoder, decoder)
+
+ @staticmethod
+ def add_args(parser):
+ TransformerModel.add_args(parser)
+ # encoder head selection
+ parser.add_argument(
+ "--encoder-attn-head-select",
+ action="store_true",
+ default=False,
+ help="encoder head selection"
+ )
+ parser.add_argument(
+ "--total-encoder-attention-heads",
+ type=int,
+ help="total number of encoder attention heads"
+ )
+ # decoder self attention
+ parser.add_argument(
+ "--decoder-self-attn-head-select",
+ action="store_true",
+ default=False,
+ help="decoder self-attention head selection"
+ )
+ # decoder-encoder attention
+ parser.add_argument(
+ "--dec-enc-attn-head-select",
+ action="store_true",
+ default=False,
+ help="decoder-encoder attention head selection"
+ )
+ parser.add_argument(
+ "--total-decoder-attention-heads",
+ type=int,
+ help="total number of decoder attention heads"
+ )
+ # selection strategy
+ parser.add_argument(
+ "--attn-head-select-strategy",
+ type=str,
+ help="attention head selection strategy, subset or group"
+ )
+
+ @classmethod
+ def build_encoder(cls, args, src_dict, embed_tokens):
+ if safe_hasattr(args, "encoder_attn_head_select") and args.encoder_attn_head_select:
+ return HeadSelectionTransformerEncoder(
+ args, src_dict, embed_tokens
+ )
+ else:
+ return TransformerEncoder(args, src_dict, embed_tokens)
+
+ @classmethod
+ def build_decoder(cls, args, tgt_dict, embed_tokens):
+ if (safe_hasattr(args, "decoder_self_attn_head_select") and args.decoder_self_attn_head_select) or (safe_hasattr(args, "dec_enc_attn_head_select") and args.dec_enc_attn_head_select):
+ return HeadSelectionTransformerDecoder(
+ args, tgt_dict, embed_tokens
+ )
+ else:
+ return TransformerDecoder(args, tgt_dict, embed_tokens)
+
+
+class HeadSelectionTransformerEncoder(TransformerEncoder):
+
+ def __init__(self, args, dictionary, embed_tokens):
+ self.num_tasks = args.encoder_tasks
+ self.num_layers = args.encoder_layers
+ self.total_num_heads = args.total_encoder_attention_heads
+ self.num_heads = args.encoder_attention_heads
+ self.select_strategy = args.attn_head_select_strategy
+
+ super().__init__(args, dictionary, embed_tokens)
+ self.attn_head_selector = AttnHeadSelector(
+ self.num_tasks,
+ self.num_layers,
+ self.total_num_heads,
+ self.num_heads,
+ self.select_strategy
+ )
+ self.task_ids = None
+ self.layers = nn.ModuleList(
+ [self.build_encoder_layer(args, i) for i in range(args.encoder_layers)]
+ )
+
+ def set_task_ids(self, task_ids):
+ self.task_ids = task_ids
+
+ def build_encoder_layer(self, args, layer_idx=None):
+ return HeadSelectionTransformerEncoderLayer(
+ args,
+ layer_idx,
+ attn_head_selector=self.attn_head_selector
+ )
+
+ def forward(
+ self,
+ src_tokens,
+ src_lengths: Optional[torch.Tensor] = None,
+ return_all_hiddens: bool = False,
+ token_embeddings: Optional[torch.Tensor] = None,
+ ):
+ self.attn_head_selector.head_select(self.task_ids)
+ return super().forward(src_tokens, src_lengths, return_all_hiddens, token_embeddings)
+
+
+class HeadSelectionTransformerDecoder(TransformerDecoder):
+
+ def __init__(
+ self,
+ args,
+ dictionary,
+ embed_tokens,
+ no_encoder_attn=False,
+ output_projection=None,
+ ):
+ self.num_tasks = args.decoder_tasks
+ self.num_layers = args.decoder_layers
+ self.total_num_heads = args.total_decoder_attention_heads
+ self.num_heads = args.decoder_attention_heads
+ self.select_strategy = args.attn_head_select_strategy
+ super().__init__(
+ args, dictionary, embed_tokens,
+ no_encoder_attn=no_encoder_attn,
+ output_projection=output_projection
+ )
+ self.self_attn_head_selector = None
+ self.enc_attn_head_selector = None
+ if safe_hasattr(args, "decoder_self_attn_head_select") and args.decoder_self_attn_head_select:
+ self.self_attn_head_selector = AttnHeadSelector(
+ self.num_tasks,
+ self.num_layers,
+ self.total_num_heads,
+ self.num_heads,
+ self.select_strategy
+ )
+ if safe_hasattr(args, "dec_enc_attn_head_select") and args.dec_enc_attn_head_select:
+ self.enc_attn_head_selector = AttnHeadSelector(
+ self.num_tasks,
+ self.num_layers,
+ self.total_num_heads,
+ self.num_heads,
+ self.select_strategy
+ )
+ self.task_ids = None
+ self.layers = nn.ModuleList(
+ [
+ self.build_head_selection_decoder_layer(args, no_encoder_attn, idx) for idx in range(args.decoder_layers)
+ ]
+ )
+
+ def set_task_ids(self, task_ids):
+ self.task_ids = task_ids
+
+ def build_head_selection_decoder_layer(self, args, no_encoder_attn=False, layer_idx=None):
+ return HeadSelectionTransformerDecoderLayer(
+ args,
+ layer_idx,
+ self.self_attn_head_selector,
+ self.enc_attn_head_selector,
+ no_encoder_attn=no_encoder_attn
+ )
+
+ def forward(
+ self,
+ prev_output_tokens,
+ encoder_out: Optional[Dict[str, List[Tensor]]] = None,
+ incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None,
+ features_only: bool = False,
+ full_context_alignment: bool = False,
+ alignment_layer: Optional[int] = None,
+ alignment_heads: Optional[int] = None,
+ src_lengths: Optional[Any] = None,
+ return_all_hiddens: bool = False,
+ ):
+ if self.self_attn_head_selector is not None:
+ self.self_attn_head_selector.head_select(self.task_ids)
+ if self.enc_attn_head_selector is not None:
+ self.enc_attn_head_selector.head_select(self.task_ids)
+ return super().forward(
+ prev_output_tokens=prev_output_tokens,
+ encoder_out=encoder_out,
+ incremental_state=incremental_state,
+ features_only=features_only,
+ full_context_alignment=full_context_alignment,
+ alignment_layer=alignment_layer,
+ alignment_heads=alignment_heads,
+ src_lengths=src_lengths,
+ return_all_hiddens=return_all_hiddens
+ )
diff --git a/fairseq/examples/attention_head_selection/src/modules/__init__.py b/fairseq/examples/attention_head_selection/src/modules/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/fairseq/examples/attention_head_selection/src/modules/attn_head_selector.py b/fairseq/examples/attention_head_selection/src/modules/attn_head_selector.py
new file mode 100644
index 0000000000000000000000000000000000000000..346fc623089989e36e707f0ff80d68e4d35c3ed7
--- /dev/null
+++ b/fairseq/examples/attention_head_selection/src/modules/attn_head_selector.py
@@ -0,0 +1,81 @@
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+import torch.nn as nn
+import math
+
+
+class AttnHeadSelector(nn.Module):
+ """
+ Latent variable modeling of attention head selection
+ """
+ def __init__(
+ self, num_tasks, num_layers,
+ total_num_heads, num_heads,
+ select_strategy="group",
+ head_select_temp=5.0
+ ):
+ super(AttnHeadSelector, self).__init__()
+ self.num_tasks = num_tasks
+ self.num_layers = num_layers
+ self.total_num_heads = total_num_heads
+ self.num_heads = num_heads
+ self.select_strategy = select_strategy
+ self.temp = head_select_temp
+
+ self.head_logits = torch.nn.Parameter(
+ torch.Tensor(self.num_tasks, self.num_layers, total_num_heads),
+ requires_grad=True
+ )
+ nn.init.uniform_(
+ self.head_logits, a=math.log(0.01),
+ b=math.log(1.0)
+ )
+
+ def gumbel_sample(self, logits, tau=1.0):
+ gumbels1 = -torch.empty_like(logits, memory_format=torch.legacy_contiguous_format).exponential_().log()
+ gumbels2 = -torch.empty_like(logits, memory_format=torch.legacy_contiguous_format).exponential_().log()
+ gumbels1 = (logits + gumbels1 - gumbels2) / tau
+ y_soft = gumbels1.sigmoid()
+ return y_soft
+
+ def subset_select(self, y_soft, topk, dim=-1):
+ top_values, top_inds = torch.topk(y_soft, k=topk, dim=dim)
+ top_ret = 1.0 - top_values.detach() + top_values
+ return top_inds.detach(), top_ret
+
+ def group_selet(self, y_soft, topk, dim=-1):
+ # top_values: (num_tasks, num_layers, topk)
+ top_values, top_inds = torch.max(
+ y_soft.view(self.num_tasks, self.num_layers, -1, topk), dim=2
+ )
+ top_inds = top_inds * topk + torch.arange(topk, device=top_inds.device).unsqueeze(0).unsqueeze(1)
+ top_ret = 1.0 - top_values.detach() + top_values
+ return top_inds.detach(), top_ret
+
+ def head_select(self, task_ids=None):
+ # gumbel_sample
+ self.head_samples = self.gumbel_sample(self.head_logits, tau=self.temp)
+ # head select
+ if self.select_strategy == "subset":
+ self.subset_heads, self.subset_weights = self.subset_select(
+ self.head_samples,
+ topk=self.num_heads,
+ )
+ elif self.select_strategy == "group":
+ self.subset_heads, self.subset_weights = self.group_selet(
+ self.head_samples,
+ topk=self.num_heads,
+ )
+ else:
+ raise ValueError("{} is not supported".format(self.select_strategy))
+
+ self.batch_subset = self.subset_heads[task_ids, :, :]
+ self.batch_weights = self.subset_weights[task_ids, :, :]
+
+ def forward(self, layer_idx):
+ assert layer_idx is not None
+ batch_subset = self.batch_subset[:, layer_idx, :]
+ batch_weights = self.batch_weights[:, layer_idx, :]
+ return batch_subset, batch_weights
diff --git a/fairseq/examples/attention_head_selection/src/modules/head_selection_transformer_layer.py b/fairseq/examples/attention_head_selection/src/modules/head_selection_transformer_layer.py
new file mode 100644
index 0000000000000000000000000000000000000000..c7921435035666a27ba78cb9193cd555cf54a9d8
--- /dev/null
+++ b/fairseq/examples/attention_head_selection/src/modules/head_selection_transformer_layer.py
@@ -0,0 +1,92 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+from fairseq.utils import safe_getattr
+from fairseq.modules import TransformerEncoderLayer, TransformerDecoderLayer
+from ..modules.multihead_attention_selection import MultiheadAttentionSelection
+
+
+class HeadSelectionTransformerEncoderLayer(TransformerEncoderLayer):
+
+ def __init__(self, args, layer_idx, attn_head_selector=None):
+ super().__init__(args)
+ self.layer_idx = layer_idx
+ self.self_attn = self.build_self_attention_selection(
+ self.embed_dim, args, attn_head_selector
+ )
+
+ def build_self_attention_selection(self, embed_dim, args, attn_head_selector=None):
+ return MultiheadAttentionSelection(
+ embed_dim,
+ args.total_encoder_attention_heads,
+ args.encoder_attention_heads,
+ dropout=args.attention_dropout,
+ self_attention=True,
+ q_noise=self.quant_noise,
+ qn_block_size=self.quant_noise_block_size,
+ layer_idx=self.layer_idx,
+ attn_head_selector=attn_head_selector
+ )
+
+
+class HeadSelectionTransformerDecoderLayer(TransformerDecoderLayer):
+
+ def __init__(
+ self,
+ args,
+ layer_idx,
+ self_attn_head_selector=None,
+ enc_attn_head_selector=None,
+ no_encoder_attn=False,
+ add_bias_kv=False,
+ add_zero_attn=False,
+ ):
+ self.layer_idx = layer_idx
+ super().__init__(args, no_encoder_attn, add_bias_kv, add_zero_attn)
+ if self_attn_head_selector is not None:
+ self.self_attn = self.build_self_attention_selection(
+ self.embed_dim, args,
+ self_attn_head_selector=self_attn_head_selector,
+ add_bias_kv=add_bias_kv,
+ add_zero_attn=add_zero_attn
+ )
+ if enc_attn_head_selector is not None:
+ self.encoder_attn = self.build_encoder_attention_selection(
+ self.embed_dim, args,
+ enc_attn_head_selector=enc_attn_head_selector
+ )
+
+ def build_self_attention_selection(
+ self, embed_dim, args, self_attn_head_selector=None,
+ add_bias_kv=False, add_zero_attn=False
+ ):
+ return MultiheadAttentionSelection(
+ embed_dim,
+ args.total_decoder_attention_heads,
+ args.decoder_attention_heads,
+ dropout=args.attention_dropout,
+ add_bias_kv=add_bias_kv,
+ add_zero_attn=add_zero_attn,
+ self_attention=not safe_getattr(args, "cross_self_attention"),
+ q_noise=self.quant_noise,
+ qn_block_size=self.quant_noise_block_size,
+ layer_idx=self.layer_idx,
+ attn_head_selector=self_attn_head_selector,
+ )
+
+ def build_encoder_attention_selection(self, embed_dim, args, enc_attn_head_selector=None):
+ return MultiheadAttentionSelection(
+ embed_dim,
+ args.total_decoder_attention_heads,
+ args.decoder_attention_heads,
+ kdim=args.encoder_embed_dim,
+ vdim=args.encoder_embed_dim,
+ dropout=args.attention_dropout,
+ encoder_decoder_attention=True,
+ q_noise=self.quant_noise,
+ qn_block_size=self.quant_noise_block_size,
+ layer_idx=self.layer_idx,
+ attn_head_selector=enc_attn_head_selector,
+ )
diff --git a/fairseq/examples/attention_head_selection/src/modules/multihead_attention_selection.py b/fairseq/examples/attention_head_selection/src/modules/multihead_attention_selection.py
new file mode 100644
index 0000000000000000000000000000000000000000..566ad822ac4b1bdac573f3b419430d33300c076d
--- /dev/null
+++ b/fairseq/examples/attention_head_selection/src/modules/multihead_attention_selection.py
@@ -0,0 +1,355 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+from typing import Dict, Optional, Tuple
+import torch
+from fairseq import utils
+from fairseq.modules.quant_noise import quant_noise
+from torch import Tensor, nn
+from torch.nn import Parameter
+
+from fairseq.modules.multihead_attention import MultiheadAttention
+from ..modules.multihead_functional import multi_head_attention_forward
+
+
+class MultiheadAttentionSelection(MultiheadAttention):
+
+ def __init__(
+ self,
+ embed_dim,
+ total_num_heads,
+ num_heads,
+ kdim=None,
+ vdim=None,
+ dropout=0.0,
+ bias=True,
+ add_bias_kv=False,
+ add_zero_attn=False,
+ self_attention=False,
+ encoder_decoder_attention=False,
+ q_noise=0.0,
+ qn_block_size=8,
+ layer_idx=0,
+ attn_head_selector=None
+ ):
+ super().__init__(
+ embed_dim,
+ num_heads,
+ kdim=kdim,
+ vdim=vdim,
+ dropout=dropout,
+ bias=bias,
+ add_bias_kv=add_bias_kv,
+ add_zero_attn=add_zero_attn,
+ self_attention=self_attention,
+ encoder_decoder_attention=encoder_decoder_attention,
+ q_noise=q_noise,
+ qn_block_size=qn_block_size,
+ )
+ self.layer_idx = layer_idx
+ self.attn_head_selector = attn_head_selector
+ self.total_num_heads = total_num_heads
+ self.total_embed_dim = self.head_dim * total_num_heads
+ self.k_proj = quant_noise(
+ nn.Linear(self.kdim, self.total_embed_dim, bias=bias), q_noise, qn_block_size
+ )
+ self.v_proj = quant_noise(
+ nn.Linear(self.vdim, self.total_embed_dim, bias=bias), q_noise, qn_block_size
+ )
+ self.q_proj = quant_noise(
+ nn.Linear(embed_dim, self.total_embed_dim, bias=bias), q_noise, qn_block_size
+ )
+ if add_bias_kv:
+ self.bias_k = Parameter(torch.Tensor(1, 1, self.total_embed_dim))
+ self.bias_v = Parameter(torch.Tensor(1, 1, self.total_embed_dim))
+ else:
+ self.bias_k = self.bias_v = None
+ self.reset_parameters()
+
+ def forward(
+ self,
+ query,
+ key: Optional[Tensor],
+ value: Optional[Tensor],
+ key_padding_mask: Optional[Tensor] = None,
+ incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None,
+ need_weights: bool = True,
+ static_kv: bool = False,
+ attn_mask: Optional[Tensor] = None,
+ before_softmax: bool = False,
+ need_head_weights: bool = False,
+ # subset_heads: Optional[Tensor] = None,
+ # subset_weights: Optional[Tensor] = None
+ ) -> Tuple[Tensor, Optional[Tensor]]:
+ if need_head_weights:
+ need_weights = True
+
+ is_tpu = query.device.type == "xla"
+
+ subset_heads, subset_weights = self.attn_head_selector(self.layer_idx)
+
+ tgt_len, bsz, embed_dim = query.size()
+ src_len = tgt_len
+ assert list(query.size()) == [tgt_len, bsz, self.embed_dim]
+ if key is not None:
+ src_len, key_bsz, _ = key.size()
+ if not torch.jit.is_scripting():
+ assert key_bsz == bsz
+ assert value is not None
+ assert src_len, bsz == value.shape[:2]
+
+ if (
+ not self.onnx_trace
+ and not is_tpu # don't use PyTorch version on TPUs
+ and incremental_state is None
+ and not static_kv
+ # A workaround for quantization to work. Otherwise JIT compilation
+ # treats bias in linear module as method.
+ and not torch.jit.is_scripting()
+ ):
+ assert key is not None and value is not None
+ return multi_head_attention_forward(
+ query,
+ key,
+ value,
+ self.embed_dim,
+ self.total_num_heads,
+ self.num_heads,
+ torch.empty([0]),
+ torch.cat((self.q_proj.bias, self.k_proj.bias, self.v_proj.bias)),
+ self.bias_k,
+ self.bias_v,
+ self.add_zero_attn,
+ self.dropout_module.p,
+ self.out_proj.weight,
+ self.out_proj.bias,
+ self.training or self.dropout_module.apply_during_inference,
+ key_padding_mask,
+ need_weights,
+ attn_mask,
+ use_separate_proj_weight=True,
+ q_proj_weight=self.q_proj.weight,
+ k_proj_weight=self.k_proj.weight,
+ v_proj_weight=self.v_proj.weight,
+ subset_heads=subset_heads,
+ subset_weights=subset_weights
+ )
+
+ if incremental_state is not None:
+ saved_state = self._get_input_buffer(incremental_state)
+ if saved_state is not None and "prev_key" in saved_state:
+ # previous time steps are cached - no need to recompute
+ # key and value if they are static
+ if static_kv:
+ assert self.encoder_decoder_attention and not self.self_attention
+ key = value = None
+ else:
+ saved_state = None
+
+ if self.self_attention:
+ q = self.q_proj(query)
+ k = self.k_proj(query)
+ v = self.v_proj(query)
+ elif self.encoder_decoder_attention:
+ # encoder-decoder attention
+ q = self.q_proj(query)
+ if key is None:
+ assert value is None
+ k = v = None
+ else:
+ k = self.k_proj(key)
+ v = self.v_proj(key)
+
+ else:
+ assert key is not None and value is not None
+ q = self.q_proj(query)
+ k = self.k_proj(key)
+ v = self.v_proj(value)
+ q *= self.scaling
+
+ if self.bias_k is not None:
+ assert self.bias_v is not None
+ k = torch.cat([k, self.bias_k.repeat(1, bsz, 1)])
+ v = torch.cat([v, self.bias_v.repeat(1, bsz, 1)])
+ if attn_mask is not None:
+ attn_mask = torch.cat(
+ [attn_mask, attn_mask.new_zeros(attn_mask.size(0), 1)], dim=1
+ )
+ if key_padding_mask is not None:
+ key_padding_mask = torch.cat(
+ [
+ key_padding_mask,
+ key_padding_mask.new_zeros(key_padding_mask.size(0), 1),
+ ],
+ dim=1,
+ )
+
+ q = (
+ q.contiguous()
+ .view(tgt_len, bsz * self.total_num_heads, self.head_dim)
+ .transpose(0, 1)
+ )
+ if k is not None:
+ k = (
+ k.contiguous()
+ .view(-1, bsz * self.total_num_heads, self.head_dim)
+ .transpose(0, 1)
+ )
+ if v is not None:
+ v = (
+ v.contiguous()
+ .view(-1, bsz * self.total_num_heads, self.head_dim)
+ .transpose(0, 1)
+ )
+
+ if saved_state is not None:
+ # saved states are stored with shape (bsz, num_heads, seq_len, head_dim)
+ if "prev_key" in saved_state:
+ _prev_key = saved_state["prev_key"]
+ assert _prev_key is not None
+ prev_key = _prev_key.view(bsz * self.total_num_heads, -1, self.head_dim)
+ if static_kv:
+ k = prev_key
+ else:
+ assert k is not None
+ k = torch.cat([prev_key, k], dim=1)
+ src_len = k.size(1)
+ if "prev_value" in saved_state:
+ _prev_value = saved_state["prev_value"]
+ assert _prev_value is not None
+ prev_value = _prev_value.view(bsz * self.total_num_heads, -1, self.head_dim)
+ if static_kv:
+ v = prev_value
+ else:
+ assert v is not None
+ v = torch.cat([prev_value, v], dim=1)
+ prev_key_padding_mask: Optional[Tensor] = None
+ if "prev_key_padding_mask" in saved_state:
+ prev_key_padding_mask = saved_state["prev_key_padding_mask"]
+ assert k is not None and v is not None
+ key_padding_mask = MultiheadAttention._append_prev_key_padding_mask(
+ key_padding_mask=key_padding_mask,
+ prev_key_padding_mask=prev_key_padding_mask,
+ batch_size=bsz,
+ src_len=k.size(1),
+ static_kv=static_kv,
+ )
+
+ saved_state["prev_key"] = k.view(bsz, self.total_num_heads, -1, self.head_dim)
+ saved_state["prev_value"] = v.view(bsz, self.total_num_heads, -1, self.head_dim)
+ saved_state["prev_key_padding_mask"] = key_padding_mask
+ # In this branch incremental_state is never None
+ assert incremental_state is not None
+ incremental_state = self._set_input_buffer(incremental_state, saved_state)
+ assert k is not None
+ assert k.size(1) == src_len
+
+ # This is part of a workaround to get around fork/join parallelism
+ # not supporting Optional types.
+ if key_padding_mask is not None and key_padding_mask.dim() == 0:
+ key_padding_mask = None
+
+ if key_padding_mask is not None:
+ assert key_padding_mask.size(0) == bsz
+ assert key_padding_mask.size(1) == src_len
+
+ if self.add_zero_attn:
+ assert v is not None
+ src_len += 1
+ k = torch.cat([k, k.new_zeros((k.size(0), 1) + k.size()[2:])], dim=1)
+ v = torch.cat([v, v.new_zeros((v.size(0), 1) + v.size()[2:])], dim=1)
+ if attn_mask is not None:
+ attn_mask = torch.cat(
+ [attn_mask, attn_mask.new_zeros(attn_mask.size(0), 1)], dim=1
+ )
+ if key_padding_mask is not None:
+ key_padding_mask = torch.cat(
+ [
+ key_padding_mask,
+ torch.zeros(key_padding_mask.size(0), 1).type_as(
+ key_padding_mask
+ ),
+ ],
+ dim=1,
+ )
+
+ attn_weights = torch.bmm(q, k.transpose(1, 2))
+ attn_weights = self.apply_sparse_mask(attn_weights, tgt_len, src_len, bsz)
+
+ assert list(attn_weights.size()) == [bsz * self.total_num_heads, tgt_len, src_len]
+
+ if attn_mask is not None:
+ attn_mask = attn_mask.unsqueeze(0)
+ if self.onnx_trace:
+ attn_mask = attn_mask.repeat(attn_weights.size(0), 1, 1)
+ attn_weights += attn_mask
+
+ if key_padding_mask is not None:
+ # don't attend to padding symbols
+ attn_weights = attn_weights.view(bsz, self.total_num_heads, tgt_len, src_len)
+ if not is_tpu:
+ attn_weights = attn_weights.masked_fill(
+ key_padding_mask.unsqueeze(1).unsqueeze(2).to(torch.bool),
+ float("-inf"),
+ )
+ else:
+ attn_weights = attn_weights.transpose(0, 2)
+ attn_weights = attn_weights.masked_fill(key_padding_mask, float("-inf"))
+ attn_weights = attn_weights.transpose(0, 2)
+ attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
+
+ if before_softmax:
+ return attn_weights, v
+
+ attn_weights_float = utils.softmax(
+ attn_weights, dim=-1, onnx_trace=self.onnx_trace
+ )
+ attn_weights = attn_weights_float.type_as(attn_weights)
+ attn_probs = self.dropout_module(attn_weights)
+
+ assert v is not None
+
+ # evaluation
+ if subset_heads is not None and subset_heads.numel() == 1:
+ subset_heads = subset_heads.repeat(bsz)
+ subset_weights = subset_weights.repeat(bsz)
+
+ if subset_heads is None:
+ attn = torch.bmm(attn_probs, v)
+ else:
+ # training with head selection
+ mixed_attn = torch.bmm(attn_probs, v).contiguous().view(bsz, self.total_num_heads, tgt_len, self.head_dim)
+ attn = torch.stack(
+ [mixed_attn[torch.arange(bsz), subset_heads[:, col], :, :] for col in range(subset_heads.size(1))], dim=1
+ )
+ attn = attn * subset_weights.unsqueeze(2).unsqueeze(3)
+ attn = attn.contiguous().view(bsz * self.num_heads, tgt_len, self.head_dim)
+
+ assert list(attn.size()) == [bsz * self.num_heads, tgt_len, self.head_dim]
+ if self.onnx_trace and attn.size(1) == 1:
+ # when ONNX tracing a single decoder step (sequence length == 1)
+ # the transpose is a no-op copy before view, thus unnecessary
+ attn = attn.contiguous().view(tgt_len, bsz, embed_dim)
+ else:
+ attn = attn.transpose(0, 1).contiguous().view(tgt_len, bsz, embed_dim)
+ attn = self.out_proj(attn)
+ attn_weights: Optional[Tensor] = None
+ if need_weights:
+ if subset_heads is None:
+ attn_weights = attn_weights_float.view(
+ bsz, self.num_heads, tgt_len, src_len
+ ).transpose(1, 0)
+ else:
+ mixed_attn_weights = attn_weights_float.view(
+ bsz, self.total_num_heads, tgt_len, src_len
+ )
+ attn_weights = torch.stack(
+ [mixed_attn_weights[torch.arange(bsz), subset_heads[:, col], :, :] for col in range(subset_heads.size(1))], dim=1
+ ).transpose(1, 0)
+ if not need_head_weights:
+ # average attention weights over heads
+ attn_weights = attn_weights.mean(dim=0)
+
+ return attn, attn_weights
diff --git a/fairseq/examples/attention_head_selection/src/modules/multihead_functional.py b/fairseq/examples/attention_head_selection/src/modules/multihead_functional.py
new file mode 100644
index 0000000000000000000000000000000000000000..d5edc777e364f625986f402326aba9c4276bb75a
--- /dev/null
+++ b/fairseq/examples/attention_head_selection/src/modules/multihead_functional.py
@@ -0,0 +1,278 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+from typing import Optional, Tuple
+import torch
+from torch import Tensor
+from torch.nn.functional import (
+ linear, softmax, dropout, pad,
+ has_torch_function,
+ handle_torch_function,
+ _in_projection_packed,
+)
+import math
+import warnings
+
+
+def _scaled_dot_product_attention(
+ q: Tensor,
+ k: Tensor,
+ v: Tensor,
+ attn_mask: Optional[Tensor] = None,
+ dropout_p: float = 0.0,
+ bsz: int = 1,
+ subset_heads: Optional[Tensor] = None,
+ subset_weights: Optional[Tensor] = None,
+) -> Tuple[Tensor, Tensor]:
+ B, Nt, E = q.shape
+ q = q / math.sqrt(E)
+ # B: bsz * total_num_heads
+ # (B, Nt, E) x (B, E, Ns) -> (B, Nt, Ns)
+ attn = torch.bmm(q, k.transpose(-2, -1))
+ if attn_mask is not None:
+ attn += attn_mask
+ attn = softmax(attn, dim=-1)
+ if dropout_p > 0.0:
+ attn = dropout(attn, p=dropout_p)
+ if subset_heads is None:
+ # (B, Nt, Ns) x (B, Ns, E) -> (B, Nt, E)
+ output = torch.bmm(attn, v)
+ else:
+ mixed_output = torch.bmm(attn, v).contiguous().view(bsz, -1, Nt, E)
+ output = torch.stack(
+ [mixed_output[torch.arange(bsz), subset_heads[:, col], :, :] for col in range(subset_heads.size(1))],
+ dim=1
+ )
+ output = output * subset_weights.unsqueeze(2).unsqueeze(3)
+ output = output.contiguous().view(-1, Nt, E)
+ if subset_heads is not None:
+ _, Nt, Ns = attn.size()
+ mixed_attn = attn.view(bsz, -1, Nt, Ns)
+ attn = torch.stack(
+ [mixed_attn[torch.arange(bsz), subset_heads[:, col], :, :] for col in range(subset_heads.size(1))], dim=1
+ )
+ return output, attn
+
+
+def _in_projection(
+ q: Tensor,
+ k: Tensor,
+ v: Tensor,
+ w_q: Tensor,
+ w_k: Tensor,
+ w_v: Tensor,
+ b_q: Optional[Tensor] = None,
+ b_k: Optional[Tensor] = None,
+ b_v: Optional[Tensor] = None,
+) -> Tuple[Tensor, Tensor, Tensor]:
+ return linear(q, w_q, b_q), linear(k, w_k, b_k), linear(v, w_v, b_v)
+
+
+def multi_head_attention_forward(
+ query: Tensor,
+ key: Tensor,
+ value: Tensor,
+ embed_dim_to_check: int,
+ total_num_heads: int,
+ num_heads: int,
+ in_proj_weight: Tensor,
+ in_proj_bias: Optional[Tensor],
+ bias_k: Optional[Tensor],
+ bias_v: Optional[Tensor],
+ add_zero_attn: bool,
+ dropout_p: float,
+ out_proj_weight: Tensor,
+ out_proj_bias: Optional[Tensor],
+ training: bool = True,
+ key_padding_mask: Optional[Tensor] = None,
+ need_weights: bool = True,
+ attn_mask: Optional[Tensor] = None,
+ use_separate_proj_weight: bool = False,
+ q_proj_weight: Optional[Tensor] = None,
+ k_proj_weight: Optional[Tensor] = None,
+ v_proj_weight: Optional[Tensor] = None,
+ static_k: Optional[Tensor] = None,
+ static_v: Optional[Tensor] = None,
+ subset_heads: Optional[Tensor] = None,
+ subset_weights: Optional[Tensor] = None,
+):
+ tens_ops = (query, key, value, in_proj_weight, in_proj_bias, bias_k, bias_v, out_proj_weight, out_proj_bias)
+ if has_torch_function(tens_ops):
+ return handle_torch_function(
+ multi_head_attention_forward,
+ tens_ops,
+ query,
+ key,
+ value,
+ embed_dim_to_check,
+ total_num_heads,
+ num_heads,
+ in_proj_weight,
+ in_proj_bias,
+ bias_k,
+ bias_v,
+ add_zero_attn,
+ dropout_p,
+ out_proj_weight,
+ out_proj_bias,
+ training=training,
+ key_padding_mask=key_padding_mask,
+ need_weights=need_weights,
+ attn_mask=attn_mask,
+ use_separate_proj_weight=use_separate_proj_weight,
+ q_proj_weight=q_proj_weight,
+ k_proj_weight=k_proj_weight,
+ v_proj_weight=v_proj_weight,
+ static_k=static_k,
+ static_v=static_v,
+ subset_heads=subset_heads,
+ subset_weights=subset_weights
+ )
+
+ # set up shape vars
+ tgt_len, bsz, embed_dim = query.shape
+ src_len, _, _ = key.shape
+ assert embed_dim == embed_dim_to_check, \
+ f"was expecting embedding dimension of {embed_dim_to_check}, but got {embed_dim}"
+ if isinstance(embed_dim, torch.Tensor):
+ # embed_dim can be a tensor when JIT tracing
+ head_dim = embed_dim.div(num_heads, rounding_mode='trunc')
+ else:
+ head_dim = embed_dim // num_heads
+ assert head_dim * num_heads == embed_dim, f"embed_dim {embed_dim} not divisible by num_heads {num_heads}"
+ if use_separate_proj_weight:
+ # allow MHA to have different embedding dimensions when separate projection weights are used
+ assert key.shape[:2] == value.shape[:2], \
+ f"key's sequence and batch dims {key.shape[:2]} do not match value's {value.shape[:2]}"
+ else:
+ assert key.shape == value.shape, f"key shape {key.shape} does not match value shape {value.shape}"
+
+ #
+ # compute in-projection
+ #
+ if not use_separate_proj_weight:
+ q, k, v = _in_projection_packed(query, key, value, in_proj_weight, in_proj_bias)
+ else:
+ assert q_proj_weight is not None, "use_separate_proj_weight is True but q_proj_weight is None"
+ assert k_proj_weight is not None, "use_separate_proj_weight is True but k_proj_weight is None"
+ assert v_proj_weight is not None, "use_separate_proj_weight is True but v_proj_weight is None"
+ if in_proj_bias is None:
+ b_q = b_k = b_v = None
+ else:
+ b_q, b_k, b_v = in_proj_bias.chunk(3)
+ q, k, v = _in_projection(query, key, value, q_proj_weight, k_proj_weight, v_proj_weight, b_q, b_k, b_v)
+
+ # prep attention mask
+ if attn_mask is not None:
+ if attn_mask.dtype == torch.uint8:
+ warnings.warn("Byte tensor for attn_mask in nn.MultiheadAttention is deprecated. Use bool tensor instead.")
+ attn_mask = attn_mask.to(torch.bool)
+ else:
+ assert attn_mask.is_floating_point() or attn_mask.dtype == torch.bool, \
+ f"Only float, byte, and bool types are supported for attn_mask, not {attn_mask.dtype}"
+ # ensure attn_mask's dim is 3
+ if attn_mask.dim() == 2:
+ correct_2d_size = (tgt_len, src_len)
+ if attn_mask.shape != correct_2d_size:
+ raise RuntimeError(f"The shape of the 2D attn_mask is {attn_mask.shape}, but should be {correct_2d_size}.")
+ attn_mask = attn_mask.unsqueeze(0)
+ elif attn_mask.dim() == 3:
+ correct_3d_size = (bsz * total_num_heads, tgt_len, src_len)
+ if attn_mask.shape != correct_3d_size:
+ raise RuntimeError(f"The shape of the 3D attn_mask is {attn_mask.shape}, but should be {correct_3d_size}.")
+ else:
+ raise RuntimeError(f"attn_mask's dimension {attn_mask.dim()} is not supported")
+
+ # prep key padding mask
+ if key_padding_mask is not None and key_padding_mask.dtype == torch.uint8:
+ warnings.warn("Byte tensor for key_padding_mask in nn.MultiheadAttention is deprecated. Use bool tensor instead.")
+ key_padding_mask = key_padding_mask.to(torch.bool)
+
+ # add bias along batch dimension (currently second)
+ if bias_k is not None and bias_v is not None:
+ assert static_k is None, "bias cannot be added to static key."
+ assert static_v is None, "bias cannot be added to static value."
+ k = torch.cat([k, bias_k.repeat(1, bsz, 1)])
+ v = torch.cat([v, bias_v.repeat(1, bsz, 1)])
+ if attn_mask is not None:
+ attn_mask = pad(attn_mask, (0, 1))
+ if key_padding_mask is not None:
+ key_padding_mask = pad(key_padding_mask, (0, 1))
+ else:
+ assert bias_k is None
+ assert bias_v is None
+
+ #
+ # reshape q, k, v for multihead attention and make em batch first
+ #
+ q = q.contiguous().view(tgt_len, bsz * total_num_heads, head_dim).transpose(0, 1)
+ if static_k is None:
+ k = k.contiguous().view(k.shape[0], bsz * total_num_heads, head_dim).transpose(0, 1)
+ else:
+ # TODO finish disentangling control flow so we don't do in-projections when statics are passed
+ assert static_k.size(0) == bsz * total_num_heads, \
+ f"expecting static_k.size(0) of {bsz * total_num_heads}, but got {static_k.size(0)}"
+ assert static_k.size(2) == head_dim, \
+ f"expecting static_k.size(2) of {head_dim}, but got {static_k.size(2)}"
+ k = static_k
+ if static_v is None:
+ v = v.contiguous().view(v.shape[0], bsz * total_num_heads, head_dim).transpose(0, 1)
+ else:
+ # TODO finish disentangling control flow so we don't do in-projections when statics are passed
+ assert static_v.size(0) == bsz * total_num_heads, \
+ f"expecting static_v.size(0) of {bsz * total_num_heads}, but got {static_v.size(0)}"
+ assert static_v.size(2) == head_dim, \
+ f"expecting static_v.size(2) of {head_dim}, but got {static_v.size(2)}"
+ v = static_v
+
+ # add zero attention along batch dimension (now first)
+ if add_zero_attn:
+ zero_attn_shape = (bsz * total_num_heads, 1, head_dim)
+ k = torch.cat([k, torch.zeros(zero_attn_shape, dtype=k.dtype, device=k.device)], dim=1)
+ v = torch.cat([v, torch.zeros(zero_attn_shape, dtype=v.dtype, device=v.device)], dim=1)
+ if attn_mask is not None:
+ attn_mask = pad(attn_mask, (0, 1))
+ if key_padding_mask is not None:
+ key_padding_mask = pad(key_padding_mask, (0, 1))
+
+ # update source sequence length after adjustments
+ src_len = k.size(1)
+
+ # merge key padding and attention masks
+ if key_padding_mask is not None:
+ assert key_padding_mask.shape == (bsz, src_len), \
+ f"expecting key_padding_mask shape of {(bsz, src_len)}, but got {key_padding_mask.shape}"
+ key_padding_mask = key_padding_mask.view(bsz, 1, 1, src_len). \
+ expand(-1, total_num_heads, -1, -1).reshape(bsz * total_num_heads, 1, src_len)
+ if attn_mask is None:
+ attn_mask = key_padding_mask
+ elif attn_mask.dtype == torch.bool:
+ attn_mask = attn_mask.logical_or(key_padding_mask)
+ else:
+ attn_mask = attn_mask.masked_fill(key_padding_mask, float("-inf"))
+
+ # convert mask to float
+ if attn_mask is not None and attn_mask.dtype == torch.bool:
+ new_attn_mask = torch.zeros_like(attn_mask, dtype=torch.float)
+ new_attn_mask.masked_fill_(attn_mask, float("-inf"))
+ attn_mask = new_attn_mask
+
+ # adjust dropout probability
+ if not training:
+ dropout_p = 0.0
+
+ #
+ # (deep breath) calculate attention and out projection
+ #
+ attn_output, attn_output_weights = _scaled_dot_product_attention(q, k, v, attn_mask, dropout_p, bsz, subset_heads, subset_weights)
+ attn_output = attn_output.transpose(0, 1).contiguous().view(tgt_len, bsz, embed_dim)
+ attn_output = linear(attn_output, out_proj_weight, out_proj_bias)
+
+ if need_weights:
+ # average attention weights over heads
+ attn_output_weights = attn_output_weights.view(bsz, num_heads, tgt_len, src_len)
+ return attn_output, attn_output_weights.sum(dim=1) / num_heads
+ else:
+ return attn_output, None
diff --git a/fairseq/examples/attention_head_selection/src/speech_to_text_head_selection.py b/fairseq/examples/attention_head_selection/src/speech_to_text_head_selection.py
new file mode 100644
index 0000000000000000000000000000000000000000..6e0ce11d6307493b30da865ba23adf23d0015b7c
--- /dev/null
+++ b/fairseq/examples/attention_head_selection/src/speech_to_text_head_selection.py
@@ -0,0 +1,180 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+from fairseq.optim.amp_optimizer import AMPOptimizer
+from fairseq.tasks import register_task
+from fairseq.tasks.speech_to_text import SpeechToTextTask
+
+from .data.speech_to_text_dataset_with_domain import SpeechToTextDatasetCreatorWithDomain
+from .loss.attention_head_selection import HeadSelectionLoss
+
+
+@register_task("speech_to_text_head_selection")
+class SpeechToTextHeadSelectionTask(SpeechToTextTask):
+
+ @classmethod
+ def add_args(cls, parser):
+ SpeechToTextTask.add_args(parser)
+ parser.add_argument(
+ "--task-type",
+ type=str,
+ default="lang",
+ help="task type for head selection, lang or domain"
+ )
+ parser.add_argument(
+ "--kl-weight",
+ type=float,
+ default=0.0,
+ help="the weight of KL loss"
+ )
+
+ def __init__(self, args, tgt_dict):
+ super().__init__(args, tgt_dict)
+ self.task_type = args.task_type
+ assert self.task_type in ["lang", "domain"], "invalid task_type: {}, should be either lang or domain".format(self.task_type)
+ self.map_task_to_id(args.train_subset)
+ self.encoder_head_prior = float(args.decoder_attention_heads) / args.total_decoder_attention_heads
+ self.decoder_head_prior = float(args.encoder_attention_heads) / args.total_encoder_attention_heads
+ self.kl_loss = HeadSelectionLoss(args)
+
+ def map_task_to_id(self, train_subset):
+ src_lang_set, tgt_lang_set, domain_set = set(), set(), set()
+ for split in train_subset.split(","):
+ seq = split.split("_")
+ assert len(seq) == 4, "subset {} should be in the format of train_src_tgt_domain".format(split)
+ _, src_lang, tgt_lang, domain = seq
+ src_lang_set.add(src_lang)
+ tgt_lang_set.add(tgt_lang)
+ domain_set.add(domain)
+ src_langs = sorted(src_lang_set)
+ tgt_langs = sorted(tgt_lang_set)
+ domains = sorted(domain_set)
+ self.src_lang_map = {src_lang: i for (i, src_lang) in enumerate(src_langs)}
+ self.tgt_lang_map = {tgt_lang: i for (i, tgt_lang) in enumerate(tgt_langs)}
+ self.domain_map = {domain: i for (i, domain) in enumerate(domains)}
+ if self.task_type == "lang":
+ self.encoder_tasks = len(self.src_lang_map)
+ self.decoder_tasks = len(self.tgt_lang_map)
+ elif self.task_type == "domain":
+ self.encoder_tasks = len(self.domain_map)
+ self.decoder_tasks = len(self.domain_map)
+
+ def load_dataset(self, split, epoch=1, combine=False, **kwargs):
+ is_train_split = split.startswith("train")
+ pre_tokenizer = self.build_tokenizer(self.args)
+ bpe_tokenizer = self.build_bpe(self.args)
+ self.datasets[split] = SpeechToTextDatasetCreatorWithDomain.from_tsv(
+ self.args.data,
+ self.data_cfg,
+ split,
+ self.tgt_dict,
+ pre_tokenizer,
+ bpe_tokenizer,
+ is_train_split=is_train_split,
+ epoch=epoch,
+ seed=self.args.seed,
+ src_lang_map=self.src_lang_map,
+ tgt_lang_map=self.tgt_lang_map,
+ domain_map=self.domain_map,
+ speaker_to_id=self.speaker_to_id
+ )
+
+ def build_model(self, args):
+ args.encoder_tasks = self.encoder_tasks
+ args.decoder_tasks = self.decoder_tasks
+ return super(SpeechToTextHeadSelectionTask, self).build_model(args)
+
+ def get_sample_sizes(self, sample, task_ids, num_tasks):
+ """
+ task_ids: (bsz,)
+ get sample sizes for each task
+ """
+ bsz = task_ids.size(0)
+ mat = torch.zeros((num_tasks, bsz), device=task_ids.device)
+ mat[task_ids, torch.arange(bsz)] = 1.0
+ ntokens = torch.sum(sample['target'] != 1, dim=-1)
+ sample_sizes = torch.matmul(mat, ntokens.float())
+ return sample_sizes
+
+ def train_step(
+ self, sample, model, criterion, optimizer, update_num, ignore_grad=False
+ ):
+ model.train()
+ model.set_num_updates(update_num)
+ # task ids
+ if self.task_type == "lang":
+ encoder_task_ids = sample["src_lang_ids"]
+ decoder_task_ids = sample["tgt_lang_ids"]
+ elif self.task_type == "domain":
+ encoder_task_ids = sample["domain_ids"]
+ decoder_task_ids = sample["domain_ids"]
+ model.encoder.set_task_ids(encoder_task_ids)
+ model.decoder.set_task_ids(decoder_task_ids)
+
+ with torch.autograd.profiler.record_function("forward"):
+ with torch.cuda.amp.autocast(enabled=(isinstance(optimizer, AMPOptimizer))):
+ loss, sample_size, logging_output = criterion(model, sample)
+ # KL loss
+ if self.args.encoder_attn_head_select:
+ sample_sizes = self.get_sample_sizes(sample, encoder_task_ids, self.encoder_tasks)
+ loss += self.kl_loss(
+ model.encoder.attn_head_selector.head_samples,
+ sample_sizes,
+ self.encoder_head_prior
+ )
+ if self.args.decoder_self_attn_head_select:
+ sample_sizes = self.get_sample_sizes(sample, decoder_task_ids, self.decoder_tasks)
+ loss += self.kl_loss(
+ model.decoder.self_attn_head_selector.head_samples,
+ sample_sizes,
+ self.decoder_head_prior
+ )
+ if self.args.dec_enc_attn_head_select:
+ sample_sizes = self.get_sample_sizes(sample, decoder_task_ids, self.decoder_tasks)
+ loss += self.kl_loss(
+ model.decoder.enc_attn_head_selector.head_sampes,
+ sample_sizes,
+ self.decoder_head_prior
+ )
+
+ if ignore_grad:
+ loss *= 0
+ with torch.autograd.profiler.record_function("backward"):
+ optimizer.backward(loss)
+ return loss, sample_size, logging_output
+
+ def valid_step(self, sample, model, criterion):
+ model.eval()
+ # task ids
+ if self.task_type == "lang":
+ encoder_task_ids = sample["src_lang_ids"]
+ decoder_task_ids = sample["tgt_lang_ids"]
+ elif self.task_type == "domain":
+ encoder_task_ids = sample["domain_ids"]
+ decoder_task_ids = sample["domain_ids"]
+ model.encoder.set_task_ids(encoder_task_ids)
+ model.decoder.set_task_ids(decoder_task_ids)
+ with torch.no_grad():
+ loss, sample_size, logging_output = criterion(model, sample)
+ return loss, sample_size, logging_output
+
+ def inference_step(
+ self, generator, models, sample, prefix_tokens=None, constraints=None
+ ):
+ with torch.no_grad():
+ # task ids
+ if self.task_type == "lang":
+ encoder_task_ids = sample["src_lang_ids"][:1]
+ decoder_task_ids = sample["tgt_lang_ids"][:1]
+ elif self.task_type == "domain":
+ encoder_task_ids = sample["domain_ids"][:1]
+ decoder_task_ids = sample["domain_ids"][:1]
+ for model in models:
+ model.encoder.set_task_ids(encoder_task_ids)
+ model.decoder.set_task_ids(decoder_task_ids)
+ return generator.generate(
+ models, sample, prefix_tokens=prefix_tokens, constraints=constraints
+ )
diff --git a/fairseq/examples/audio_nlp/nlu/README.md b/fairseq/examples/audio_nlp/nlu/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..a11b3f306584c81a74aa7bf768d5fd264f177cd9
--- /dev/null
+++ b/fairseq/examples/audio_nlp/nlu/README.md
@@ -0,0 +1,53 @@
+# End-to-end NLU
+
+End-to-end spoken language understanding (SLU) predicts intent directly from audio using a single model. It promises to improve the performance of assistant systems by leveraging acoustic information lost in the intermediate textual representation and preventing cascading errors from Automatic Speech Recognition (ASR). Further, having one unified model has efficiency advantages when deploying assistant systems on-device.
+
+This page releases the code for reproducing the results in [STOP: A dataset for Spoken Task Oriented Semantic Parsing](https://arxiv.org/abs/2207.10643)
+
+The dataset can be downloaded here: [download link](https://dl.fbaipublicfiles.com/stop/stop.tar.gz)
+
+The low-resource splits can be downloaded here: [download link](http://dl.fbaipublicfiles.com/stop/low_resource_splits.tar.gz)
+
+## Pretrained models end-to-end NLU Models
+
+| Speech Pretraining | ASR Pretraining | Test EM Accuracy | Tesst EM-Tree Accuracy | Link |
+| ----------- | ----------- |----------|----------|----------|
+| None | None | 36.54 | 57.01 | [link](https://dl.fbaipublicfiles.com/stop/end-to-end-nlu-none-none.pt) |
+| Wav2Vec | None | 68.05 | 82.53 | [link](https://dl.fbaipublicfiles.com/stop/end-to-end-nlu-wav2vec-none.pt) |
+| HuBERT | None | 68.40 | 82.85 | [link](https://dl.fbaipublicfiles.com/stop/end-to-end-nlu-hubert-none.pt) |
+| Wav2Vec | STOP | 68.70 | 82.78 | [link](https://dl.fbaipublicfiles.com/stop/end-to-end-nlu-wav2vec-stop.pt) |
+| HuBERT | STOP | 69.23 | 82.87 | [link](https://dl.fbaipublicfiles.com/stop/end-to-end-nlu-hubert-stop.pt) |
+| Wav2Vec | Librispeech | 68.47 | 82.49 | [link](https://dl.fbaipublicfiles.com/stop/end-to-end-nlu-wav2vec-ls.pt) |
+| HuBERT | Librispeech | 68.70 | 82.78 | [link](https://dl.fbaipublicfiles.com/stop/end-to-end-nlu-hubert-ls.pt) |
+
+## Pretrained models ASR Models
+| Speech Pre-training | ASR Dataset | STOP Eval WER | STOP Test WER | dev\_other WER | dev\_clean WER | test\_clean WER | test\_other WER | Link |
+| ----------- | ----------- | ----------- | ----------- | ----------- | ----------- | ----------- | ----------- | ----------- |
+| HuBERT | Librispeech | 8.47 | 2.99 | 3.25 | 8.06 | 25.68 | 26.19 | [link](https://dl.fbaipublicfiles.com/stop/ctc-asr-hubert-ls.pt) |
+| Wav2Vec | Librispeech | 9.215 | 3.204 | 3.334 | 9.006 | 27.257 | 27.588 | [link](https://dl.fbaipublicfiles.com/stop/ctc-asr-wav2vec-ls.pt) |
+| HuBERT | STOP | 46.31 | 31.30 | 31.52 | 47.16 | 4.29 | 4.26 | [link](https://dl.fbaipublicfiles.com/stop/ctc-asr-hubert-stop.pt) |
+| Wav2Vec | STOP | 43.103 | 27.833 | 28.479 | 28.479 | 4.679 | 4.667 | [link](https://dl.fbaipublicfiles.com/stop/ctc-asr-wav2vec-stop.pt) |
+| HuBERT | Librispeech + STOP | 9.015 | 3.211 | 3.372 | 8.635 | 5.133 | 5.056 | [link](https://dl.fbaipublicfiles.com/stop/ctc-asr-hubert-ls-stop.pt) |
+| Wav2Vec | Librispeech + STOP | 9.549 | 3.537 | 3.625 | 9.514 | 5.59 | 5.562 | [link](https://dl.fbaipublicfiles.com/stop/ctc-asr-wav2vec-ls-stop.pt) |
+
+## Creating the fairseq datasets from STOP
+
+First, create the audio file manifests and label files:
+
+```
+python examples/audio_nlp/nlu/generate_manifests.py --stop_root $STOP_DOWNLOAD_DIR/stop --output $FAIRSEQ_DATASET_OUTPUT/
+```
+
+
+Run `./examples/audio_nlp/nlu/create_dict_stop.sh $FAIRSEQ_DATASET_OUTPUT` to generate the fairseq dictionaries.
+
+
+## Training an End-to-end NLU Model
+
+
+Download a wav2vec or hubert model from [link](https://github.com/facebookresearch/fairseq/tree/main/examples/hubert) or [link](https://github.com/facebookresearch/fairseq/tree/main/examples/wav2vec)
+
+
+```
+python fairseq_cli/hydra-train --config-dir examples/audio_nlp/nlu/configs/ --config-name nlu_finetuning task.data=$FAIRSEQ_DATA_OUTPUT model.w2v_path=$PRETRAINED_MODEL_PATH
+```
diff --git a/fairseq/examples/audio_nlp/nlu/configs/nlu_finetuning.yaml b/fairseq/examples/audio_nlp/nlu/configs/nlu_finetuning.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..bb90f45a307bd36040ba579a012bac8db911ec5c
--- /dev/null
+++ b/fairseq/examples/audio_nlp/nlu/configs/nlu_finetuning.yaml
@@ -0,0 +1,59 @@
+# @package _group_
+
+common:
+ fp16: true
+ log_format: json
+ log_interval: 10
+ tensorboard_logdir: tb
+
+checkpoint:
+ no_epoch_checkpoints: true
+ best_checkpoint_metric: em_error
+ save_interval: 10
+
+task:
+ _name: nlu_finetuning
+ data: ???
+ labels: parse
+ eval_wer_parse: true
+ autoregressive: true
+
+dataset:
+ num_workers: 6
+ max_tokens: 1600000
+ skip_invalid_size_inputs_valid_test: true
+ valid_subset: eval,test
+ train_subset: train
+ validate_interval: 10
+
+criterion:
+ _name: label_smoothed_cross_entropy
+
+optimization:
+ max_update: 320000
+ lr: [0.0001]
+ sentence_avg: true
+ update_freq: [1]
+
+optimizer:
+ _name: adam
+ adam_betas: (0.9,0.98)
+ adam_eps: 1e-08
+
+lr_scheduler:
+ _name: tri_stage
+ phase_ratio: [0.1, 0.4, 0.5]
+ final_lr_scale: 0.05
+
+model:
+ _name: wav2vec_seq2seq
+ w2v_path: ???
+ autoregressive: true
+ apply_mask: true
+ mask_prob: 0.5
+ mask_channel_prob: 0.5
+ mask_channel_length: 64
+ layerdrop: 0.1
+ activation_dropout: 0.1
+ feature_grad_mult: 0.0
+ freeze_finetune_updates: 0
diff --git a/fairseq/examples/audio_nlp/nlu/create_dict_stop.sh b/fairseq/examples/audio_nlp/nlu/create_dict_stop.sh
new file mode 100644
index 0000000000000000000000000000000000000000..753393284de3703247db31ab224ebf11fab0242b
--- /dev/null
+++ b/fairseq/examples/audio_nlp/nlu/create_dict_stop.sh
@@ -0,0 +1,38 @@
+#!/bin/bash
+
+### Script handling creation of data binaries
+### for model training within fairseq
+
+
+fairseq_root="."
+
+data_root=$1
+train_prefix="${data_root}/train"
+valid_prefix="${data_root}/eval"
+test_prefix="${data_root}/test"
+
+dest_dir="$data_root/"
+
+#echo "src dict: $src_dict" > "$dest_dir/src_dict.txt"
+#echo "trg dict: $tgt_dict" > "$dest_dir/tgt_dict.txt"
+
+ #--tgtdict $tgt_dict \
+PYTHONPATH=$fairseq_root \
+ python $fairseq_root/fairseq_cli/preprocess.py \
+ --source-lang "parse" \
+ --trainpref "$train_prefix" \
+ --validpref "$valid_prefix" \
+ --destdir "$dest_dir" \
+ --only-source \
+ --dict-only \
+ --workers 60;
+
+PYTHONPATH=$fairseq_root \
+ python $fairseq_root/fairseq_cli/preprocess.py \
+ --source-lang "ltr" \
+ --trainpref "$train_prefix" \
+ --validpref "$valid_prefix" \
+ --destdir "$dest_dir" \
+ --only-source \
+ --dict-only \
+ --workers 60;
diff --git a/fairseq/examples/audio_nlp/nlu/generate_manifests.py b/fairseq/examples/audio_nlp/nlu/generate_manifests.py
new file mode 100644
index 0000000000000000000000000000000000000000..e2176099cbd3993d3488ccc85b60f5a0da45d4df
--- /dev/null
+++ b/fairseq/examples/audio_nlp/nlu/generate_manifests.py
@@ -0,0 +1,83 @@
+import argparse
+from pathlib import Path
+import soundfile
+
+def get_insl_frame(parse):
+ out = []
+ def is_ont_token(tok):
+ return tok[0] in ["[", "]"];
+
+ res = []
+ x = []
+ for tok in parse.split():
+ if is_ont_token(tok):
+ res.extend('_'.join(x))
+ x = []
+ res.append(tok.upper())
+ else:
+ x.append(tok.upper())
+
+ return " ".join(res) + ' | '
+
+def sequencify_utterance(utterance):
+ utterance = utterance.upper()
+ utterance = utterance.replace(' ', '|') + '|'
+ utterance = list(utterance)
+ utterance = ' '.join(utterance)
+ return utterance
+
+
+def generate_fairseq_manifests(manifest, output_path, audio_root=None):
+
+ with open(manifest, 'r') as i:
+ parses = []
+ utterances = []
+ filepaths = []
+ keys = None
+ for (idx, line) in enumerate(i):
+ if idx == 0: keys = line.strip().split('\t')
+ else:
+ data = { k: v for (k, v) in zip(keys, line.split('\t'))}
+ parses.append(get_insl_frame(data['decoupled_normalized_seqlogical']))
+ utterances.append(sequencify_utterance(data['normalized_utterance']))
+ filepaths.append(data['file_id'])
+
+ parses_fp = output_path.with_suffix('.parse')
+ with open(str(parses_fp), 'w') as o:
+ for p in parses:
+ o.write(p + '\n')
+
+ utterances_fp = output_path.with_suffix('.ltr')
+ with open(str(utterances_fp), 'w') as o:
+ for u in utterances:
+ o.write(u + '\n')
+
+ filepaths_fp = output_path.with_suffix('.tsv')
+ with open(str(filepaths_fp), 'w') as o:
+ o.write(str(audio_root) + '\n')
+ for f in filepaths:
+ fullpath = audio_root / f
+ assert fullpath.exists(), f'{fullpath}'
+ frames = soundfile.info(fullpath).frames
+ o.write(f'{f}\t{frames}\n')
+
+def main(args):
+
+ splits = ['train', 'eval', 'test']
+ root = Path(args.stop_root)
+ output_root = Path(args.output)
+
+ for split in splits:
+ stop_manifest_path = root / 'manifests' / (split + '.tsv')
+ output_path = output_root / (split)
+
+ generate_fairseq_manifests(stop_manifest_path, output_path, root)
+
+if __name__ == '__main__':
+ parser = argparse.ArgumentParser(description='Process some integers.')
+ parser.add_argument('--stop_root', type=str,
+ help='path to stop root directory')
+ parser.add_argument('--output', type=str,
+ help='output directory')
+ args = parser.parse_args()
+ main(args)
diff --git a/fairseq/examples/backtranslation/README.md b/fairseq/examples/backtranslation/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..73675f1125d80f58aa824db67d8970504d4d6b2a
--- /dev/null
+++ b/fairseq/examples/backtranslation/README.md
@@ -0,0 +1,297 @@
+# Understanding Back-Translation at Scale (Edunov et al., 2018)
+
+This page includes pre-trained models from the paper [Understanding Back-Translation at Scale (Edunov et al., 2018)](https://arxiv.org/abs/1808.09381).
+
+## Pre-trained models
+
+Model | Description | Dataset | Download
+---|---|---|---
+`transformer.wmt18.en-de` | Transformer ([Edunov et al., 2018](https://arxiv.org/abs/1808.09381)) WMT'18 winner | [WMT'18 English-German](http://www.statmt.org/wmt18/translation-task.html) | [download (.tar.gz)](https://dl.fbaipublicfiles.com/fairseq/models/wmt18.en-de.ensemble.tar.gz) See NOTE in the archive
+
+## Example usage (torch.hub)
+
+We require a few additional Python dependencies for preprocessing:
+```bash
+pip install subword_nmt sacremoses
+```
+
+Then to generate translations from the full model ensemble:
+```python
+import torch
+
+# List available models
+torch.hub.list('pytorch/fairseq') # [..., 'transformer.wmt18.en-de', ... ]
+
+# Load the WMT'18 En-De ensemble
+en2de_ensemble = torch.hub.load(
+ 'pytorch/fairseq', 'transformer.wmt18.en-de',
+ checkpoint_file='wmt18.model1.pt:wmt18.model2.pt:wmt18.model3.pt:wmt18.model4.pt:wmt18.model5.pt',
+ tokenizer='moses', bpe='subword_nmt')
+
+# The ensemble contains 5 models
+len(en2de_ensemble.models)
+# 5
+
+# Translate
+en2de_ensemble.translate('Hello world!')
+# 'Hallo Welt!'
+```
+
+## Training your own model (WMT'18 English-German)
+
+The following instructions can be adapted to reproduce the models from the paper.
+
+
+#### Step 1. Prepare parallel data and optionally train a baseline (English-German) model
+
+First download and preprocess the data:
+```bash
+# Download and prepare the data
+cd examples/backtranslation/
+bash prepare-wmt18en2de.sh
+cd ../..
+
+# Binarize the data
+TEXT=examples/backtranslation/wmt18_en_de
+fairseq-preprocess \
+ --joined-dictionary \
+ --source-lang en --target-lang de \
+ --trainpref $TEXT/train --validpref $TEXT/valid --testpref $TEXT/test \
+ --destdir data-bin/wmt18_en_de --thresholdtgt 0 --thresholdsrc 0 \
+ --workers 20
+
+# Copy the BPE code into the data-bin directory for future use
+cp examples/backtranslation/wmt18_en_de/code data-bin/wmt18_en_de/code
+```
+
+(Optionally) Train a baseline model (English-German) using just the parallel data:
+```bash
+CHECKPOINT_DIR=checkpoints_en_de_parallel
+fairseq-train --fp16 \
+ data-bin/wmt18_en_de \
+ --source-lang en --target-lang de \
+ --arch transformer_wmt_en_de_big --share-all-embeddings \
+ --dropout 0.3 --weight-decay 0.0 \
+ --criterion label_smoothed_cross_entropy --label-smoothing 0.1 \
+ --optimizer adam --adam-betas '(0.9, 0.98)' --clip-norm 0.0 \
+ --lr 0.001 --lr-scheduler inverse_sqrt --warmup-updates 4000 \
+ --max-tokens 3584 --update-freq 16 \
+ --max-update 30000 \
+ --save-dir $CHECKPOINT_DIR
+# Note: the above command assumes 8 GPUs. Adjust `--update-freq` if you have a
+# different number of GPUs.
+```
+
+Average the last 10 checkpoints:
+```bash
+python scripts/average_checkpoints.py \
+ --inputs $CHECKPOINT_DIR \
+ --num-epoch-checkpoints 10 \
+ --output $CHECKPOINT_DIR/checkpoint.avg10.pt
+```
+
+Evaluate BLEU:
+```bash
+# tokenized BLEU on newstest2017:
+bash examples/backtranslation/tokenized_bleu.sh \
+ wmt17 \
+ en-de \
+ data-bin/wmt18_en_de \
+ data-bin/wmt18_en_de/code \
+ $CHECKPOINT_DIR/checkpoint.avg10.pt
+# BLEU4 = 29.57, 60.9/35.4/22.9/15.5 (BP=1.000, ratio=1.014, syslen=63049, reflen=62152)
+# compare to 29.46 in Table 1, which is also for tokenized BLEU
+
+# generally it's better to report (detokenized) sacrebleu though:
+bash examples/backtranslation/sacrebleu.sh \
+ wmt17 \
+ en-de \
+ data-bin/wmt18_en_de \
+ data-bin/wmt18_en_de/code \
+ $CHECKPOINT_DIR/checkpoint.avg10.pt
+# BLEU+case.mixed+lang.en-de+numrefs.1+smooth.exp+test.wmt17+tok.13a+version.1.4.3 = 29.0 60.6/34.7/22.4/14.9 (BP = 1.000 ratio = 1.013 hyp_len = 62099 ref_len = 61287)
+```
+
+
+#### Step 2. Back-translate monolingual German data
+
+Train a reverse model (German-English) to do the back-translation:
+```bash
+CHECKPOINT_DIR=checkpoints_de_en_parallel
+fairseq-train --fp16 \
+ data-bin/wmt18_en_de \
+ --source-lang de --target-lang en \
+ --arch transformer_wmt_en_de_big --share-all-embeddings \
+ --dropout 0.3 --weight-decay 0.0 \
+ --criterion label_smoothed_cross_entropy --label-smoothing 0.1 \
+ --optimizer adam --adam-betas '(0.9, 0.98)' --clip-norm 0.0 \
+ --lr 0.001 --lr-scheduler inverse_sqrt --warmup-updates 4000 \
+ --max-tokens 3584 --update-freq 16 \
+ --max-update 30000 \
+ --save-dir $CHECKPOINT_DIR
+# Note: the above command assumes 8 GPUs. Adjust `--update-freq` if you have a
+# different number of GPUs.
+```
+
+Let's evaluate the back-translation (BT) model to make sure it is well trained:
+```bash
+bash examples/backtranslation/sacrebleu.sh \
+ wmt17 \
+ de-en \
+ data-bin/wmt18_en_de \
+ data-bin/wmt18_en_de/code \
+ $CHECKPOINT_DIR/checkpoint_best.py
+# BLEU+case.mixed+lang.de-en+numrefs.1+smooth.exp+test.wmt17+tok.13a+version.1.4.3 = 34.9 66.9/41.8/28.5/19.9 (BP = 0.983 ratio = 0.984 hyp_len = 63342 ref_len = 64399)
+# compare to the best system from WMT'17 which scored 35.1: http://matrix.statmt.org/matrix/systems_list/1868
+```
+
+Next prepare the monolingual data:
+```bash
+# Download and prepare the monolingual data
+# By default the script samples 25M monolingual sentences, which after
+# deduplication should be just over 24M sentences. These are split into 25
+# shards, each with 1M sentences (except for the last shard).
+cd examples/backtranslation/
+bash prepare-de-monolingual.sh
+cd ../..
+
+# Binarize each shard of the monolingual data
+TEXT=examples/backtranslation/wmt18_de_mono
+for SHARD in $(seq -f "%02g" 0 24); do \
+ fairseq-preprocess \
+ --only-source \
+ --source-lang de --target-lang en \
+ --joined-dictionary \
+ --srcdict data-bin/wmt18_en_de/dict.de.txt \
+ --testpref $TEXT/bpe.monolingual.dedup.${SHARD} \
+ --destdir data-bin/wmt18_de_mono/shard${SHARD} \
+ --workers 20; \
+ cp data-bin/wmt18_en_de/dict.en.txt data-bin/wmt18_de_mono/shard${SHARD}/; \
+done
+```
+
+Now we're ready to perform back-translation over the monolingual data. The
+following command generates via sampling, but it's possible to use greedy
+decoding (`--beam 1`), beam search (`--beam 5`),
+top-k sampling (`--sampling --beam 1 --sampling-topk 10`), etc.:
+```bash
+mkdir backtranslation_output
+for SHARD in $(seq -f "%02g" 0 24); do \
+ fairseq-generate --fp16 \
+ data-bin/wmt18_de_mono/shard${SHARD} \
+ --path $CHECKPOINT_DIR/checkpoint_best.pt \
+ --skip-invalid-size-inputs-valid-test \
+ --max-tokens 4096 \
+ --sampling --beam 1 \
+ > backtranslation_output/sampling.shard${SHARD}.out; \
+done
+```
+
+After BT, use the `extract_bt_data.py` script to re-combine the shards, extract
+the back-translations and apply length ratio filters:
+```bash
+python examples/backtranslation/extract_bt_data.py \
+ --minlen 1 --maxlen 250 --ratio 1.5 \
+ --output backtranslation_output/bt_data --srclang en --tgtlang de \
+ backtranslation_output/sampling.shard*.out
+
+# Ensure lengths are the same:
+# wc -l backtranslation_output/bt_data.{en,de}
+# 21795614 backtranslation_output/bt_data.en
+# 21795614 backtranslation_output/bt_data.de
+# 43591228 total
+```
+
+Binarize the filtered BT data and combine it with the parallel data:
+```bash
+TEXT=backtranslation_output
+fairseq-preprocess \
+ --source-lang en --target-lang de \
+ --joined-dictionary \
+ --srcdict data-bin/wmt18_en_de/dict.en.txt \
+ --trainpref $TEXT/bt_data \
+ --destdir data-bin/wmt18_en_de_bt \
+ --workers 20
+
+# We want to train on the combined data, so we'll symlink the parallel + BT data
+# in the wmt18_en_de_para_plus_bt directory. We link the parallel data as "train"
+# and the BT data as "train1", so that fairseq will combine them automatically
+# and so that we can use the `--upsample-primary` option to upsample the
+# parallel data (if desired).
+PARA_DATA=$(readlink -f data-bin/wmt18_en_de)
+BT_DATA=$(readlink -f data-bin/wmt18_en_de_bt)
+COMB_DATA=data-bin/wmt18_en_de_para_plus_bt
+mkdir -p $COMB_DATA
+for LANG in en de; do \
+ ln -s ${PARA_DATA}/dict.$LANG.txt ${COMB_DATA}/dict.$LANG.txt; \
+ for EXT in bin idx; do \
+ ln -s ${PARA_DATA}/train.en-de.$LANG.$EXT ${COMB_DATA}/train.en-de.$LANG.$EXT; \
+ ln -s ${BT_DATA}/train.en-de.$LANG.$EXT ${COMB_DATA}/train1.en-de.$LANG.$EXT; \
+ ln -s ${PARA_DATA}/valid.en-de.$LANG.$EXT ${COMB_DATA}/valid.en-de.$LANG.$EXT; \
+ ln -s ${PARA_DATA}/test.en-de.$LANG.$EXT ${COMB_DATA}/test.en-de.$LANG.$EXT; \
+ done; \
+done
+```
+
+
+#### 3. Train an English-German model over the combined parallel + BT data
+
+Finally we can train a model over the parallel + BT data:
+```bash
+CHECKPOINT_DIR=checkpoints_en_de_parallel_plus_bt
+fairseq-train --fp16 \
+ data-bin/wmt18_en_de_para_plus_bt \
+ --upsample-primary 16 \
+ --source-lang en --target-lang de \
+ --arch transformer_wmt_en_de_big --share-all-embeddings \
+ --dropout 0.3 --weight-decay 0.0 \
+ --criterion label_smoothed_cross_entropy --label-smoothing 0.1 \
+ --optimizer adam --adam-betas '(0.9, 0.98)' --clip-norm 0.0 \
+ --lr 0.0007 --lr-scheduler inverse_sqrt --warmup-updates 4000 \
+ --max-tokens 3584 --update-freq 16 \
+ --max-update 100000 \
+ --save-dir $CHECKPOINT_DIR
+# Note: the above command assumes 8 GPUs. Adjust `--update-freq` if you have a
+# different number of GPUs.
+```
+
+Average the last 10 checkpoints:
+```bash
+python scripts/average_checkpoints.py \
+ --inputs $CHECKPOINT_DIR \
+ --num-epoch-checkpoints 10 \
+ --output $CHECKPOINT_DIR/checkpoint.avg10.pt
+```
+
+Evaluate BLEU:
+```bash
+# tokenized BLEU on newstest2017:
+bash examples/backtranslation/tokenized_bleu.sh \
+ wmt17 \
+ en-de \
+ data-bin/wmt18_en_de \
+ data-bin/wmt18_en_de/code \
+ $CHECKPOINT_DIR/checkpoint.avg10.pt
+# BLEU4 = 32.35, 64.4/38.9/26.2/18.3 (BP=0.977, ratio=0.977, syslen=60729, reflen=62152)
+# compare to 32.35 in Table 1, which is also for tokenized BLEU
+
+# generally it's better to report (detokenized) sacrebleu:
+bash examples/backtranslation/sacrebleu.sh \
+ wmt17 \
+ en-de \
+ data-bin/wmt18_en_de \
+ data-bin/wmt18_en_de/code \
+ $CHECKPOINT_DIR/checkpoint.avg10.pt
+# BLEU+case.mixed+lang.en-de+numrefs.1+smooth.exp+test.wmt17+tok.13a+version.1.4.3 = 31.5 64.3/38.2/25.6/17.6 (BP = 0.971 ratio = 0.971 hyp_len = 59515 ref_len = 61287)
+```
+
+
+## Citation
+```bibtex
+@inproceedings{edunov2018backtranslation,
+ title = {Understanding Back-Translation at Scale},
+ author = {Edunov, Sergey and Ott, Myle and Auli, Michael and Grangier, David},
+ booktitle = {Conference of the Association for Computational Linguistics (ACL)},
+ year = 2018,
+}
+```
diff --git a/fairseq/examples/backtranslation/deduplicate_lines.py b/fairseq/examples/backtranslation/deduplicate_lines.py
new file mode 100644
index 0000000000000000000000000000000000000000..50e458328c80b71c42a66d473381ca7e98d294da
--- /dev/null
+++ b/fairseq/examples/backtranslation/deduplicate_lines.py
@@ -0,0 +1,41 @@
+#!/usr/bin/python3
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import argparse
+import fileinput
+import hashlib
+import sys
+from multiprocessing import Pool
+
+
+def get_hashes_and_lines(raw_line):
+ hash = hashlib.md5(raw_line).hexdigest()
+ return hash, raw_line
+
+
+def main():
+ parser = argparse.ArgumentParser()
+ parser.add_argument("--workers", type=int, default=10)
+ parser.add_argument("files", nargs="*", help="input files")
+ args = parser.parse_args()
+
+ seen = set()
+ with fileinput.input(args.files, mode="rb") as h:
+ pool = Pool(args.workers)
+ results = pool.imap_unordered(get_hashes_and_lines, h, 1000)
+ for i, (hash, raw_line) in enumerate(results):
+ if hash not in seen:
+ seen.add(hash)
+ sys.stdout.buffer.write(raw_line)
+ if i % 1000000 == 0:
+ print(i, file=sys.stderr, end="", flush=True)
+ elif i % 100000 == 0:
+ print(".", file=sys.stderr, end="", flush=True)
+ print(file=sys.stderr, flush=True)
+
+
+if __name__ == "__main__":
+ main()
diff --git a/fairseq/examples/backtranslation/extract_bt_data.py b/fairseq/examples/backtranslation/extract_bt_data.py
new file mode 100644
index 0000000000000000000000000000000000000000..e766391e873d0d9a9561d67d5864934b2fad0681
--- /dev/null
+++ b/fairseq/examples/backtranslation/extract_bt_data.py
@@ -0,0 +1,72 @@
+#!/usr/bin/env python
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import argparse
+import fileinput
+
+from tqdm import tqdm
+
+
+def main():
+ parser = argparse.ArgumentParser(
+ description=(
+ "Extract back-translations from the stdout of fairseq-generate. "
+ "If there are multiply hypotheses for a source, we only keep the first one. "
+ )
+ )
+ parser.add_argument("--output", required=True, help="output prefix")
+ parser.add_argument(
+ "--srclang", required=True, help="source language (extracted from H-* lines)"
+ )
+ parser.add_argument(
+ "--tgtlang", required=True, help="target language (extracted from S-* lines)"
+ )
+ parser.add_argument("--minlen", type=int, help="min length filter")
+ parser.add_argument("--maxlen", type=int, help="max length filter")
+ parser.add_argument("--ratio", type=float, help="ratio filter")
+ parser.add_argument("files", nargs="*", help="input files")
+ args = parser.parse_args()
+
+ def validate(src, tgt):
+ srclen = len(src.split(" ")) if src != "" else 0
+ tgtlen = len(tgt.split(" ")) if tgt != "" else 0
+ if (
+ (args.minlen is not None and (srclen < args.minlen or tgtlen < args.minlen))
+ or (
+ args.maxlen is not None
+ and (srclen > args.maxlen or tgtlen > args.maxlen)
+ )
+ or (
+ args.ratio is not None
+ and (max(srclen, tgtlen) / float(min(srclen, tgtlen)) > args.ratio)
+ )
+ ):
+ return False
+ return True
+
+ def safe_index(toks, index, default):
+ try:
+ return toks[index]
+ except IndexError:
+ return default
+
+ with open(args.output + "." + args.srclang, "w") as src_h, open(
+ args.output + "." + args.tgtlang, "w"
+ ) as tgt_h:
+ for line in tqdm(fileinput.input(args.files)):
+ if line.startswith("S-"):
+ tgt = safe_index(line.rstrip().split("\t"), 1, "")
+ elif line.startswith("H-"):
+ if tgt is not None:
+ src = safe_index(line.rstrip().split("\t"), 2, "")
+ if validate(src, tgt):
+ print(src, file=src_h)
+ print(tgt, file=tgt_h)
+ tgt = None
+
+
+if __name__ == "__main__":
+ main()
diff --git a/fairseq/examples/backtranslation/prepare-de-monolingual.sh b/fairseq/examples/backtranslation/prepare-de-monolingual.sh
new file mode 100644
index 0000000000000000000000000000000000000000..5e67b2b3bcf27d3436031453e796e58a0ae79ec4
--- /dev/null
+++ b/fairseq/examples/backtranslation/prepare-de-monolingual.sh
@@ -0,0 +1,98 @@
+#!/bin/bash
+
+SCRIPTS=mosesdecoder/scripts
+TOKENIZER=$SCRIPTS/tokenizer/tokenizer.perl
+NORM_PUNC=$SCRIPTS/tokenizer/normalize-punctuation.perl
+REM_NON_PRINT_CHAR=$SCRIPTS/tokenizer/remove-non-printing-char.perl
+BPEROOT=subword-nmt/subword_nmt
+
+
+BPE_CODE=wmt18_en_de/code
+SUBSAMPLE_SIZE=25000000
+LANG=de
+
+
+OUTDIR=wmt18_${LANG}_mono
+orig=orig
+tmp=$OUTDIR/tmp
+mkdir -p $OUTDIR $tmp
+
+
+URLS=(
+ "http://www.statmt.org/wmt14/training-monolingual-news-crawl/news.2007.de.shuffled.gz"
+ "http://www.statmt.org/wmt14/training-monolingual-news-crawl/news.2008.de.shuffled.gz"
+ "http://www.statmt.org/wmt14/training-monolingual-news-crawl/news.2009.de.shuffled.gz"
+ "http://www.statmt.org/wmt14/training-monolingual-news-crawl/news.2010.de.shuffled.gz"
+ "http://www.statmt.org/wmt14/training-monolingual-news-crawl/news.2011.de.shuffled.gz"
+ "http://www.statmt.org/wmt14/training-monolingual-news-crawl/news.2012.de.shuffled.gz"
+ "http://www.statmt.org/wmt14/training-monolingual-news-crawl/news.2013.de.shuffled.gz"
+ "http://www.statmt.org/wmt15/training-monolingual-news-crawl-v2/news.2014.de.shuffled.v2.gz"
+ "http://data.statmt.org/wmt16/translation-task/news.2015.de.shuffled.gz"
+ "http://data.statmt.org/wmt17/translation-task/news.2016.de.shuffled.gz"
+ "http://data.statmt.org/wmt18/translation-task/news.2017.de.shuffled.deduped.gz"
+)
+FILES=(
+ "news.2007.de.shuffled.gz"
+ "news.2008.de.shuffled.gz"
+ "news.2009.de.shuffled.gz"
+ "news.2010.de.shuffled.gz"
+ "news.2011.de.shuffled.gz"
+ "news.2012.de.shuffled.gz"
+ "news.2013.de.shuffled.gz"
+ "news.2014.de.shuffled.v2.gz"
+ "news.2015.de.shuffled.gz"
+ "news.2016.de.shuffled.gz"
+ "news.2017.de.shuffled.deduped.gz"
+)
+
+
+cd $orig
+for ((i=0;i<${#URLS[@]};++i)); do
+ file=${FILES[i]}
+ if [ -f $file ]; then
+ echo "$file already exists, skipping download"
+ else
+ url=${URLS[i]}
+ wget "$url"
+ fi
+done
+cd ..
+
+
+if [ -f $tmp/monolingual.${SUBSAMPLE_SIZE}.${LANG} ]; then
+ echo "found monolingual sample, skipping shuffle/sample/tokenize"
+else
+ gzip -c -d -k $(for FILE in "${FILES[@]}"; do echo $orig/$FILE; done) \
+ | shuf -n $SUBSAMPLE_SIZE \
+ | perl $NORM_PUNC $LANG \
+ | perl $REM_NON_PRINT_CHAR \
+ | perl $TOKENIZER -threads 8 -a -l $LANG \
+ > $tmp/monolingual.${SUBSAMPLE_SIZE}.${LANG}
+fi
+
+
+if [ -f $tmp/bpe.monolingual.${SUBSAMPLE_SIZE}.${LANG} ]; then
+ echo "found BPE monolingual sample, skipping BPE step"
+else
+ python $BPEROOT/apply_bpe.py -c $BPE_CODE \
+ < $tmp/monolingual.${SUBSAMPLE_SIZE}.${LANG} \
+ > $tmp/bpe.monolingual.${SUBSAMPLE_SIZE}.${LANG}
+fi
+
+
+if [ -f $tmp/bpe.monolingual.dedup.${SUBSAMPLE_SIZE}.${LANG} ]; then
+ echo "found deduplicated monolingual sample, skipping deduplication step"
+else
+ python deduplicate_lines.py $tmp/bpe.monolingual.${SUBSAMPLE_SIZE}.${LANG} \
+ > $tmp/bpe.monolingual.dedup.${SUBSAMPLE_SIZE}.${LANG}
+fi
+
+
+if [ -f $OUTDIR/bpe.monolingual.dedup.00.de ]; then
+ echo "found sharded data, skipping sharding step"
+else
+ split --lines 1000000 --numeric-suffixes \
+ --additional-suffix .${LANG} \
+ $tmp/bpe.monolingual.dedup.${SUBSAMPLE_SIZE}.${LANG} \
+ $OUTDIR/bpe.monolingual.dedup.
+fi
diff --git a/fairseq/examples/backtranslation/prepare-wmt18en2de.sh b/fairseq/examples/backtranslation/prepare-wmt18en2de.sh
new file mode 100644
index 0000000000000000000000000000000000000000..f6fd275307db50ca84c299440ae02dce49064030
--- /dev/null
+++ b/fairseq/examples/backtranslation/prepare-wmt18en2de.sh
@@ -0,0 +1,135 @@
+#!/bin/bash
+# Adapted from https://github.com/facebookresearch/MIXER/blob/master/prepareData.sh
+
+echo 'Cloning Moses github repository (for tokenization scripts)...'
+git clone https://github.com/moses-smt/mosesdecoder.git
+
+echo 'Cloning Subword NMT repository (for BPE pre-processing)...'
+git clone https://github.com/rsennrich/subword-nmt.git
+
+SCRIPTS=mosesdecoder/scripts
+TOKENIZER=$SCRIPTS/tokenizer/tokenizer.perl
+CLEAN=$SCRIPTS/training/clean-corpus-n.perl
+NORM_PUNC=$SCRIPTS/tokenizer/normalize-punctuation.perl
+REM_NON_PRINT_CHAR=$SCRIPTS/tokenizer/remove-non-printing-char.perl
+BPEROOT=subword-nmt/subword_nmt
+BPE_TOKENS=32000
+
+URLS=(
+ "http://statmt.org/wmt13/training-parallel-europarl-v7.tgz"
+ "http://statmt.org/wmt13/training-parallel-commoncrawl.tgz"
+ "http://data.statmt.org/wmt18/translation-task/training-parallel-nc-v13.tgz"
+ "http://data.statmt.org/wmt18/translation-task/rapid2016.tgz"
+ "http://data.statmt.org/wmt17/translation-task/dev.tgz"
+ "http://statmt.org/wmt14/test-full.tgz"
+)
+FILES=(
+ "training-parallel-europarl-v7.tgz"
+ "training-parallel-commoncrawl.tgz"
+ "training-parallel-nc-v13.tgz"
+ "rapid2016.tgz"
+ "dev.tgz"
+ "test-full.tgz"
+)
+CORPORA=(
+ "training/europarl-v7.de-en"
+ "commoncrawl.de-en"
+ "training-parallel-nc-v13/news-commentary-v13.de-en"
+ "rapid2016.de-en"
+)
+
+if [ ! -d "$SCRIPTS" ]; then
+ echo "Please set SCRIPTS variable correctly to point to Moses scripts."
+ exit 1
+fi
+
+OUTDIR=wmt18_en_de
+
+src=en
+tgt=de
+lang=en-de
+prep=$OUTDIR
+tmp=$prep/tmp
+orig=orig
+
+mkdir -p $orig $tmp $prep
+
+cd $orig
+
+for ((i=0;i<${#URLS[@]};++i)); do
+ file=${FILES[i]}
+ if [ -f $file ]; then
+ echo "$file already exists, skipping download"
+ else
+ url=${URLS[i]}
+ wget "$url"
+ if [ -f $file ]; then
+ echo "$url successfully downloaded."
+ else
+ echo "$url not successfully downloaded."
+ exit 1
+ fi
+ if [ ${file: -4} == ".tgz" ]; then
+ tar zxvf $file
+ elif [ ${file: -4} == ".tar" ]; then
+ tar xvf $file
+ fi
+ fi
+done
+cd ..
+
+echo "pre-processing train data..."
+for l in $src $tgt; do
+ rm $tmp/train.tags.$lang.tok.$l
+ for f in "${CORPORA[@]}"; do
+ cat $orig/$f.$l | \
+ perl $NORM_PUNC $l | \
+ perl $REM_NON_PRINT_CHAR | \
+ perl $TOKENIZER -threads 8 -a -l $l >> $tmp/train.tags.$lang.tok.$l
+ done
+done
+
+echo "pre-processing test data..."
+for l in $src $tgt; do
+ if [ "$l" == "$src" ]; then
+ t="src"
+ else
+ t="ref"
+ fi
+ grep '\s*//g' | \
+ sed -e 's/\s*<\/seg>\s*//g' | \
+ sed -e "s/\’/\'/g" | \
+ perl $TOKENIZER -threads 8 -a -l $l > $tmp/test.$l
+ echo ""
+done
+
+echo "splitting train and valid..."
+for l in $src $tgt; do
+ awk '{if (NR%100 == 0) print $0; }' $tmp/train.tags.$lang.tok.$l > $tmp/valid.$l
+ awk '{if (NR%100 != 0) print $0; }' $tmp/train.tags.$lang.tok.$l > $tmp/train.$l
+done
+
+TRAIN=$tmp/train.de-en
+BPE_CODE=$prep/code
+rm -f $TRAIN
+for l in $src $tgt; do
+ cat $tmp/train.$l >> $TRAIN
+done
+
+echo "learn_bpe.py on ${TRAIN}..."
+python $BPEROOT/learn_bpe.py -s $BPE_TOKENS < $TRAIN > $BPE_CODE
+
+for L in $src $tgt; do
+ for f in train.$L valid.$L test.$L; do
+ echo "apply_bpe.py to ${f}..."
+ python $BPEROOT/apply_bpe.py -c $BPE_CODE < $tmp/$f > $tmp/bpe.$f
+ done
+done
+
+perl $CLEAN -ratio 1.5 $tmp/bpe.train $src $tgt $prep/train 1 250
+perl $CLEAN -ratio 1.5 $tmp/bpe.valid $src $tgt $prep/valid 1 250
+
+for L in $src $tgt; do
+ cp $tmp/bpe.test.$L $prep/test.$L
+done
diff --git a/fairseq/examples/backtranslation/sacrebleu.sh b/fairseq/examples/backtranslation/sacrebleu.sh
new file mode 100644
index 0000000000000000000000000000000000000000..a70da23f48e2699297799611412783d4560dc45a
--- /dev/null
+++ b/fairseq/examples/backtranslation/sacrebleu.sh
@@ -0,0 +1,37 @@
+#!/bin/bash
+
+if [ $# -ne 5 ]; then
+ echo "usage: $0 [dataset=wmt14/full] [langpair=en-de] [databin] [bpecode] [model]"
+ exit
+fi
+
+
+DATASET=$1
+LANGPAIR=$2
+DATABIN=$3
+BPECODE=$4
+MODEL=$5
+
+SRCLANG=$(echo $LANGPAIR | cut -d '-' -f 1)
+TGTLANG=$(echo $LANGPAIR | cut -d '-' -f 2)
+
+
+BPEROOT=examples/backtranslation/subword-nmt/subword_nmt
+if [ ! -e $BPEROOT ]; then
+ BPEROOT=subword-nmt/subword_nmt
+ if [ ! -e $BPEROOT ]; then
+ echo 'Cloning Subword NMT repository (for BPE pre-processing)...'
+ git clone https://github.com/rsennrich/subword-nmt.git
+ fi
+fi
+
+
+sacrebleu -t $DATASET -l $LANGPAIR --echo src \
+| sacremoses tokenize -a -l $SRCLANG -q \
+| python $BPEROOT/apply_bpe.py -c $BPECODE \
+| fairseq-interactive $DATABIN --path $MODEL \
+ -s $SRCLANG -t $TGTLANG \
+ --beam 5 --remove-bpe --buffer-size 1024 --max-tokens 8000 \
+| grep ^H- | cut -f 3- \
+| sacremoses detokenize -l $TGTLANG -q \
+| sacrebleu -t $DATASET -l $LANGPAIR
diff --git a/fairseq/examples/backtranslation/tokenized_bleu.sh b/fairseq/examples/backtranslation/tokenized_bleu.sh
new file mode 100644
index 0000000000000000000000000000000000000000..c6d6aaa193f6059299bc98909324fe4b9b060372
--- /dev/null
+++ b/fairseq/examples/backtranslation/tokenized_bleu.sh
@@ -0,0 +1,46 @@
+#!/bin/bash
+
+if [ $# -ne 5 ]; then
+ echo "usage: $0 [dataset=wmt14/full] [langpair=en-de] [databin] [bpecode] [model]"
+ exit
+fi
+
+
+DATASET=$1
+LANGPAIR=$2
+DATABIN=$3
+BPECODE=$4
+MODEL=$5
+
+SRCLANG=$(echo $LANGPAIR | cut -d '-' -f 1)
+TGTLANG=$(echo $LANGPAIR | cut -d '-' -f 2)
+
+
+BPEROOT=examples/backtranslation/subword-nmt/subword_nmt
+if [ ! -e $BPEROOT ]; then
+ BPEROOT=subword-nmt/subword_nmt
+ if [ ! -e $BPEROOT ]; then
+ echo 'Cloning Subword NMT repository (for BPE pre-processing)...'
+ git clone https://github.com/rsennrich/subword-nmt.git
+ fi
+fi
+
+
+TMP_REF=$(mktemp)
+
+sacrebleu -t $DATASET -l $LANGPAIR --echo ref -q \
+| sacremoses normalize -l $TGTLANG -q \
+| sacremoses tokenize -a -l $TGTLANG -q \
+> $TMP_REF
+
+sacrebleu -t $DATASET -l $LANGPAIR --echo src -q \
+| sacremoses normalize -l $SRCLANG -q \
+| sacremoses tokenize -a -l $SRCLANG -q \
+| python $BPEROOT/apply_bpe.py -c $BPECODE \
+| fairseq-interactive $DATABIN --path $MODEL \
+ -s $SRCLANG -t $TGTLANG \
+ --beam 5 --remove-bpe --buffer-size 1024 --max-tokens 8000 \
+| grep ^H- | cut -f 3- \
+| fairseq-score --ref $TMP_REF
+
+rm -f $TMP_REF
diff --git a/fairseq/examples/bart/README.glue.md b/fairseq/examples/bart/README.glue.md
new file mode 100644
index 0000000000000000000000000000000000000000..a010934e1e6dec491eb1c704ec02ba7405760510
--- /dev/null
+++ b/fairseq/examples/bart/README.glue.md
@@ -0,0 +1,99 @@
+# Fine-tuning BART on GLUE tasks
+
+### 1) Download the data from GLUE website (https://gluebenchmark.com/tasks) using following commands:
+```bash
+wget https://gist.githubusercontent.com/W4ngatang/60c2bdb54d156a41194446737ce03e2e/raw/17b8dd0d724281ed7c3b2aeeda662b92809aadd5/download_glue_data.py
+python download_glue_data.py --data_dir glue_data --tasks all
+```
+
+### 2) Preprocess GLUE task data (same as RoBERTa):
+```bash
+./examples/roberta/preprocess_GLUE_tasks.sh glue_data
+```
+`glue_task_name` is one of the following:
+`{ALL, QQP, MNLI, QNLI, MRPC, RTE, STS-B, SST-2, CoLA}`
+Use `ALL` for preprocessing all the glue tasks.
+
+### 3) Fine-tuning on GLUE task:
+Example fine-tuning cmd for `RTE` task
+```bash
+TOTAL_NUM_UPDATES=2036 # 10 epochs through RTE for bsz 16
+WARMUP_UPDATES=61 # 6 percent of the number of updates
+LR=1e-05 # Peak LR for polynomial LR scheduler.
+NUM_CLASSES=2
+MAX_SENTENCES=16 # Batch size.
+BART_PATH=/path/to/bart/model.pt
+
+CUDA_VISIBLE_DEVICES=0,1 fairseq-train RTE-bin/ \
+ --restore-file $BART_PATH \
+ --batch-size $MAX_SENTENCES \
+ --max-tokens 4400 \
+ --task sentence_prediction \
+ --add-prev-output-tokens \
+ --layernorm-embedding \
+ --share-all-embeddings \
+ --share-decoder-input-output-embed \
+ --reset-optimizer --reset-dataloader --reset-meters \
+ --required-batch-size-multiple 1 \
+ --init-token 0 \
+ --arch bart_large \
+ --criterion sentence_prediction \
+ --num-classes $NUM_CLASSES \
+ --dropout 0.1 --attention-dropout 0.1 \
+ --weight-decay 0.01 --optimizer adam --adam-betas "(0.9, 0.98)" --adam-eps 1e-08 \
+ --clip-norm 0.0 \
+ --lr-scheduler polynomial_decay --lr $LR --total-num-update $TOTAL_NUM_UPDATES --warmup-updates $WARMUP_UPDATES \
+ --fp16 --fp16-init-scale 4 --threshold-loss-scale 1 --fp16-scale-window 128 \
+ --max-epoch 10 \
+ --find-unused-parameters \
+ --best-checkpoint-metric accuracy --maximize-best-checkpoint-metric;
+```
+
+For each of the GLUE task, you will need to use following cmd-line arguments:
+
+Model | MNLI | QNLI | QQP | RTE | SST-2 | MRPC | CoLA | STS-B
+---|---|---|---|---|---|---|---|---
+`--num-classes` | 3 | 2 | 2 | 2 | 2 | 2 | 2 | 1
+`--lr` | 5e-6 | 1e-5 | 1e-5 | 1e-5 | 5e-6 | 2e-5 | 2e-5 | 2e-5
+`bsz` | 128 | 32 | 32 | 32 | 128 | 64 | 64 | 32
+`--total-num-update` | 30968 | 33112 | 113272 | 1018 | 5233 | 1148 | 1334 | 1799
+`--warmup-updates` | 1858 | 1986 | 6796 | 61 | 314 | 68 | 80 | 107
+
+For `STS-B` additionally add `--regression-target --best-checkpoint-metric loss` and remove `--maximize-best-checkpoint-metric`.
+
+**Note:**
+
+a) `--total-num-updates` is used by `--polynomial_decay` scheduler and is calculated for `--max-epoch=10` and `--batch-size=32/64/128` depending on the task.
+
+b) Above cmd-args and hyperparams are tested on Nvidia `V100` GPU with `32gb` of memory for each task. Depending on the GPU memory resources available to you, you can use increase `--update-freq` and reduce `--batch-size`.
+
+### Inference on GLUE task
+After training the model as mentioned in previous step, you can perform inference with checkpoints in `checkpoints/` directory using following python code snippet:
+
+```python
+from fairseq.models.bart import BARTModel
+
+bart = BARTModel.from_pretrained(
+ 'checkpoints/',
+ checkpoint_file='checkpoint_best.pt',
+ data_name_or_path='RTE-bin'
+)
+
+label_fn = lambda label: bart.task.label_dictionary.string(
+ [label + bart.task.label_dictionary.nspecial]
+)
+ncorrect, nsamples = 0, 0
+bart.cuda()
+bart.eval()
+with open('glue_data/RTE/dev.tsv') as fin:
+ fin.readline()
+ for index, line in enumerate(fin):
+ tokens = line.strip().split('\t')
+ sent1, sent2, target = tokens[1], tokens[2], tokens[3]
+ tokens = bart.encode(sent1, sent2)
+ prediction = bart.predict('sentence_classification_head', tokens).argmax().item()
+ prediction_label = label_fn(prediction)
+ ncorrect += int(prediction_label == target)
+ nsamples += 1
+print('| Accuracy: ', float(ncorrect)/float(nsamples))
+```
diff --git a/fairseq/examples/bart/README.md b/fairseq/examples/bart/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..4050a724ee6a2f20c9998a95df48c58b64764ab1
--- /dev/null
+++ b/fairseq/examples/bart/README.md
@@ -0,0 +1,228 @@
+# BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension
+
+[https://arxiv.org/abs/1910.13461](https://arxiv.org/abs/1910.13461)
+
+## Introduction
+
+BART is sequence-to-sequence model trained with denoising as pretraining objective. We show that this pretraining objective is more generic and show that we can match [RoBERTa](../roberta) results on SQuAD and GLUE and gain state-of-the-art results on summarization (XSum, CNN dataset), long form generative question answering (ELI5) and dialog response genration (ConvAI2). See the associated paper for more details.
+
+## Pre-trained models
+
+Model | Description | # params | Download
+---|---|---|---
+`bart.base` | BART model with 6 encoder and decoder layers | 140M | [bart.base.tar.gz](https://dl.fbaipublicfiles.com/fairseq/models/bart.base.tar.gz)
+`bart.large` | BART model with 12 encoder and decoder layers | 400M | [bart.large.tar.gz](https://dl.fbaipublicfiles.com/fairseq/models/bart.large.tar.gz)
+`bart.large.mnli` | `bart.large` finetuned on `MNLI` | 400M | [bart.large.mnli.tar.gz](https://dl.fbaipublicfiles.com/fairseq/models/bart.large.mnli.tar.gz)
+`bart.large.cnn` | `bart.large` finetuned on `CNN-DM` | 400M | [bart.large.cnn.tar.gz](https://dl.fbaipublicfiles.com/fairseq/models/bart.large.cnn.tar.gz)
+`bart.large.xsum` | `bart.large` finetuned on `Xsum` | 400M | [bart.large.xsum.tar.gz](https://dl.fbaipublicfiles.com/fairseq/models/bart.large.xsum.tar.gz)
+
+## Results
+
+**[GLUE (Wang et al., 2019)](https://gluebenchmark.com/)**
+_(dev set, single model, single-task finetuning)_
+
+Model | MNLI | QNLI | QQP | RTE | SST-2 | MRPC | CoLA | STS-B
+---|---|---|---|---|---|---|---|---
+`roberta.large` | 90.2 | 94.7 | 92.2 | 86.6 | 96.4 | 90.9 | 68.0 | 92.4
+`bart.large` | 89.9 | 94.9 | 92.5 | 87.0 | 96.6 | 90.4 | 62.8 | 91.2
+
+**[SQuAD (Rajpurkar et al., 2018)](https://rajpurkar.github.io/SQuAD-explorer/)**
+_(dev set, no additional data used)_
+
+Model | SQuAD 1.1 EM/F1 | SQuAD 2.0 EM/F1
+---|---|---
+`roberta.large` | 88.9/94.6 | 86.5/89.4
+`bart.large` | 88.8/94.6 | 86.1/89.2
+
+**[CNN/Daily Mail](http://nlpprogress.com/english/summarization.html)**
+_(test set, no additional data used)_
+
+Model | R1 | R2 | RL
+---|---|---|---
+`BERTSUMEXTABS` | 42.13 | 19.60 | 39.18
+`bart.large` | 44.16 | 21.28 | 40.90
+
+## Example usage
+
+##### Load BART from torch.hub (PyTorch >= 1.1):
+```python
+import torch
+bart = torch.hub.load('pytorch/fairseq', 'bart.large')
+bart.eval() # disable dropout (or leave in train mode to finetune)
+```
+
+##### Load BART (for PyTorch 1.0 or custom models):
+```python
+# Download bart.large model
+wget https://dl.fbaipublicfiles.com/fairseq/models/bart.large.tar.gz
+tar -xzvf bart.large.tar.gz
+
+# Load the model in fairseq
+from fairseq.models.bart import BARTModel
+bart = BARTModel.from_pretrained('/path/to/bart.large', checkpoint_file='model.pt')
+bart.eval() # disable dropout (or leave in train mode to finetune)
+```
+
+##### Apply Byte-Pair Encoding (BPE) to input text:
+```python
+tokens = bart.encode('Hello world!')
+assert tokens.tolist() == [0, 31414, 232, 328, 2]
+bart.decode(tokens) # 'Hello world!'
+```
+
+##### Extract features from BART:
+```python
+# Extract the last layer's features
+last_layer_features = bart.extract_features(tokens)
+assert last_layer_features.size() == torch.Size([1, 5, 1024])
+
+# Extract all layer's features from decoder (layer 0 is the embedding layer)
+all_layers = bart.extract_features(tokens, return_all_hiddens=True)
+assert len(all_layers) == 13
+assert torch.all(all_layers[-1] == last_layer_features)
+```
+
+##### Use BART for sentence-pair classification tasks:
+```python
+# Download BART already finetuned for MNLI
+bart = torch.hub.load('pytorch/fairseq', 'bart.large.mnli')
+bart.eval() # disable dropout for evaluation
+
+# Encode a pair of sentences and make a prediction
+tokens = bart.encode('BART is a seq2seq model.', 'BART is not sequence to sequence.')
+bart.predict('mnli', tokens).argmax() # 0: contradiction
+
+# Encode another pair of sentences
+tokens = bart.encode('BART is denoising autoencoder.', 'BART is version of autoencoder.')
+bart.predict('mnli', tokens).argmax() # 2: entailment
+```
+
+##### Register a new (randomly initialized) classification head:
+```python
+bart.register_classification_head('new_task', num_classes=3)
+logprobs = bart.predict('new_task', tokens)
+```
+
+##### Batched prediction:
+```python
+import torch
+from fairseq.data.data_utils import collate_tokens
+
+bart = torch.hub.load('pytorch/fairseq', 'bart.large.mnli')
+bart.eval()
+
+batch_of_pairs = [
+ ['BART is a seq2seq model.', 'BART is not sequence to sequence.'],
+ ['BART is denoising autoencoder.', 'BART is version of autoencoder.'],
+]
+
+batch = collate_tokens(
+ [bart.encode(pair[0], pair[1]) for pair in batch_of_pairs], pad_idx=1
+)
+
+logprobs = bart.predict('mnli', batch)
+print(logprobs.argmax(dim=1))
+# tensor([0, 2])
+```
+
+##### Using the GPU:
+```python
+bart.cuda()
+bart.predict('new_task', tokens)
+```
+
+#### Filling masks:
+
+BART can be used to fill multiple `` tokens in the input.
+```python
+bart = torch.hub.load('pytorch/fairseq', 'bart.base')
+bart.eval()
+bart.fill_mask(['The cat on the .'], topk=3, beam=10)
+# [[('The cat was on the ground.', tensor(-0.6183)), ('The cat was on the floor.', tensor(-0.6798)), ('The cat sleeps on the couch.', tensor(-0.6830))]]
+```
+
+Note that by default we enforce the output length to match the input length.
+This can be disabled by setting ``match_source_len=False``:
+```
+bart.fill_mask(['The cat on the .'], topk=3, beam=10, match_source_len=False)
+# [[('The cat was on the ground.', tensor(-0.6185)), ('The cat was asleep on the couch.', tensor(-0.6276)), ('The cat was on the floor.', tensor(-0.6800))]]
+```
+
+Example code to fill masks for a batch of sentences using GPU
+```
+bart.cuda()
+bart.fill_mask(['The cat on the .', 'The dog on the .'], topk=3, beam=10)
+# [[('The cat was on the ground.', tensor(-0.6183)), ('The cat was on the floor.', tensor(-0.6798)), ('The cat sleeps on the couch.', tensor(-0.6830))], [('The dog was on the ground.', tensor(-0.6190)), ('The dog lay on the ground.', tensor(-0.6711)),
+('The dog was asleep on the couch', tensor(-0.6796))]]
+```
+
+#### Evaluating the `bart.large.mnli` model:
+
+Example python code snippet to evaluate accuracy on the MNLI `dev_matched` set.
+```python
+label_map = {0: 'contradiction', 1: 'neutral', 2: 'entailment'}
+ncorrect, nsamples = 0, 0
+bart.cuda()
+bart.eval()
+with open('glue_data/MNLI/dev_matched.tsv') as fin:
+ fin.readline()
+ for index, line in enumerate(fin):
+ tokens = line.strip().split('\t')
+ sent1, sent2, target = tokens[8], tokens[9], tokens[-1]
+ tokens = bart.encode(sent1, sent2)
+ prediction = bart.predict('mnli', tokens).argmax().item()
+ prediction_label = label_map[prediction]
+ ncorrect += int(prediction_label == target)
+ nsamples += 1
+ print('| Accuracy: ', float(ncorrect)/float(nsamples))
+# Expected output: 0.9010
+```
+
+#### Evaluating the `bart.large.cnn` model:
+- Follow instructions [here](https://github.com/abisee/cnn-dailymail) to download and process into data-files such that `test.source` and `test.target` has one line for each non-tokenized sample.
+- For simpler preprocessing, you can also `wget https://cdn-datasets.huggingface.co/summarization/cnn_dm_v2.tgz`, although there is no guarantee of identical scores
+- `huggingface/transformers` has a simpler interface that supports [single-gpu](https://github.com/huggingface/transformers/blob/master/examples/legacy/seq2seq/run_eval.py) and [multi-gpu](https://github.com/huggingface/transformers/blob/master/examples/legacy/seq2seq/run_distributed_eval.py) beam search.
+ In `huggingface/transformers`, the BART models' paths are `facebook/bart-large-cnn` and `facebook/bart-large-xsum`.
+
+In `fairseq`, summaries can be generated using:
+
+```bash
+cp data-bin/cnn_dm/dict.source.txt checkpoints/
+python examples/bart/summarize.py \
+ --model-dir pytorch/fairseq \
+ --model-file bart.large.cnn \
+ --src cnn_dm/test.source \
+ --out cnn_dm/test.hypo
+```
+
+For calculating rouge, install `files2rouge` from [here](https://github.com/pltrdy/files2rouge).
+
+```bash
+export CLASSPATH=/path/to/stanford-corenlp-full-2016-10-31/stanford-corenlp-3.7.0.jar
+
+# Tokenize hypothesis and target files.
+cat test.hypo | java edu.stanford.nlp.process.PTBTokenizer -ioFileList -preserveLines > test.hypo.tokenized
+cat test.target | java edu.stanford.nlp.process.PTBTokenizer -ioFileList -preserveLines > test.hypo.target
+files2rouge test.hypo.tokenized test.hypo.target
+# Expected output: (ROUGE-2 Average_F: 0.21238)
+```
+
+
+## Finetuning
+
+- [Finetuning on GLUE](README.glue.md)
+- [Finetuning on CNN-DM](README.summarization.md)
+
+## Citation
+
+```bibtex
+@article{lewis2019bart,
+ title = {BART: Denoising Sequence-to-Sequence Pre-training for Natural
+Language Generation, Translation, and Comprehension},
+ author = {Mike Lewis and Yinhan Liu and Naman Goyal and Marjan Ghazvininejad and
+ Abdelrahman Mohamed and Omer Levy and Veselin Stoyanov
+ and Luke Zettlemoyer },
+ journal={arXiv preprint arXiv:1910.13461},
+ year = {2019},
+}
+```
diff --git a/fairseq/examples/bart/README.summarization.md b/fairseq/examples/bart/README.summarization.md
new file mode 100644
index 0000000000000000000000000000000000000000..8727584f2b2bdd880c6cd3abbf39b75dfbf4a67c
--- /dev/null
+++ b/fairseq/examples/bart/README.summarization.md
@@ -0,0 +1,102 @@
+# Fine-tuning BART on CNN-Dailymail summarization task
+
+### 1) Download the CNN and Daily Mail data and preprocess it into data files with non-tokenized cased samples.
+
+Follow the instructions [here](https://github.com/abisee/cnn-dailymail) to download the original CNN and Daily Mail datasets. To preprocess the data, refer to the pointers in [this issue](https://github.com/pytorch/fairseq/issues/1391) or check out the code [here](https://github.com/artmatsak/cnn-dailymail).
+
+Follow the instructions [here](https://github.com/EdinburghNLP/XSum) to download the original Extreme Summarization datasets, or check out the code [here](https://github.com/EdinburghNLP/XSum/tree/master/XSum-Dataset), Please keep the raw dataset and make sure no tokenization nor BPE on the dataset.
+
+### 2) BPE preprocess:
+
+```bash
+wget -N 'https://dl.fbaipublicfiles.com/fairseq/gpt2_bpe/encoder.json'
+wget -N 'https://dl.fbaipublicfiles.com/fairseq/gpt2_bpe/vocab.bpe'
+wget -N 'https://dl.fbaipublicfiles.com/fairseq/gpt2_bpe/dict.txt'
+
+TASK=cnn_dm
+for SPLIT in train val
+do
+ for LANG in source target
+ do
+ python -m examples.roberta.multiprocessing_bpe_encoder \
+ --encoder-json encoder.json \
+ --vocab-bpe vocab.bpe \
+ --inputs "$TASK/$SPLIT.$LANG" \
+ --outputs "$TASK/$SPLIT.bpe.$LANG" \
+ --workers 60 \
+ --keep-empty;
+ done
+done
+```
+
+### 3) Binarize dataset:
+```bash
+fairseq-preprocess \
+ --source-lang "source" \
+ --target-lang "target" \
+ --trainpref "${TASK}/train.bpe" \
+ --validpref "${TASK}/val.bpe" \
+ --destdir "${TASK}-bin/" \
+ --workers 60 \
+ --srcdict dict.txt \
+ --tgtdict dict.txt;
+```
+
+### 4) Fine-tuning on CNN-DM summarization task:
+Example fine-tuning CNN-DM
+```bash
+TOTAL_NUM_UPDATES=20000
+WARMUP_UPDATES=500
+LR=3e-05
+MAX_TOKENS=2048
+UPDATE_FREQ=4
+BART_PATH=/path/to/bart/model.pt
+
+CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 fairseq-train cnn_dm-bin \
+ --restore-file $BART_PATH \
+ --max-tokens $MAX_TOKENS \
+ --task translation \
+ --source-lang source --target-lang target \
+ --truncate-source \
+ --layernorm-embedding \
+ --share-all-embeddings \
+ --share-decoder-input-output-embed \
+ --reset-optimizer --reset-dataloader --reset-meters \
+ --required-batch-size-multiple 1 \
+ --arch bart_large \
+ --criterion label_smoothed_cross_entropy \
+ --label-smoothing 0.1 \
+ --dropout 0.1 --attention-dropout 0.1 \
+ --weight-decay 0.01 --optimizer adam --adam-betas "(0.9, 0.999)" --adam-eps 1e-08 \
+ --clip-norm 0.1 \
+ --lr-scheduler polynomial_decay --lr $LR --total-num-update $TOTAL_NUM_UPDATES --warmup-updates $WARMUP_UPDATES \
+ --fp16 --update-freq $UPDATE_FREQ \
+ --skip-invalid-size-inputs-valid-test \
+ --find-unused-parameters;
+```
+Above is expected to run on `1` node with `8 32gb-V100`.
+Expected training time is about `5 hours`. Training time can be reduced with distributed training on `4` nodes and `--update-freq 1`.
+
+Use TOTAL_NUM_UPDATES=15000 UPDATE_FREQ=2 for Xsum task
+
+### Inference for CNN-DM test data using above trained checkpoint.
+After training the model as mentioned in previous step, you can perform inference with checkpoints in `checkpoints/` directory using `eval_cnn.py`, for example
+
+```bash
+cp data-bin/cnn_dm/dict.source.txt checkpoints/
+python examples/bart/summarize.py \
+ --model-dir checkpoints \
+ --model-file checkpoint_best.pt \
+ --src cnn_dm/test.source \
+ --out cnn_dm/test.hypo
+```
+For XSUM, which uses beam=6, lenpen=1.0, max_len_b=60, min_len=10:
+```bash
+cp data-bin/cnn_dm/dict.source.txt checkpoints/
+python examples/bart/summarize.py \
+ --model-dir checkpoints \
+ --model-file checkpoint_best.pt \
+ --src cnn_dm/test.source \
+ --out cnn_dm/test.hypo \
+ --xsum-kwargs
+```
diff --git a/fairseq/examples/bart/summarize.py b/fairseq/examples/bart/summarize.py
new file mode 100644
index 0000000000000000000000000000000000000000..04435f80e39c2d9d894696dae7cba5b381e13da9
--- /dev/null
+++ b/fairseq/examples/bart/summarize.py
@@ -0,0 +1,100 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+from fairseq.models.bart import BARTModel
+import argparse
+
+XSUM_KWARGS = dict(beam=6, lenpen=1.0, max_len_b=60, min_len=10, no_repeat_ngram_size=3)
+CNN_KWARGS = dict(beam=4, lenpen=2.0, max_len_b=140, min_len=55, no_repeat_ngram_size=3)
+
+
+@torch.no_grad()
+def generate(bart, infile, outfile="bart_hypo.txt", bsz=32, n_obs=None, **eval_kwargs):
+ count = 1
+
+ # if n_obs is not None: bsz = min(bsz, n_obs)
+
+ with open(infile) as source, open(outfile, "w") as fout:
+ sline = source.readline().strip()
+ slines = [sline]
+ for sline in source:
+ if n_obs is not None and count > n_obs:
+ break
+ if count % bsz == 0:
+ hypotheses_batch = bart.sample(slines, **eval_kwargs)
+ for hypothesis in hypotheses_batch:
+ fout.write(hypothesis + "\n")
+ fout.flush()
+ slines = []
+
+ slines.append(sline.strip())
+ count += 1
+
+ if slines != []:
+ hypotheses_batch = bart.sample(slines, **eval_kwargs)
+ for hypothesis in hypotheses_batch:
+ fout.write(hypothesis + "\n")
+ fout.flush()
+
+
+def main():
+ """
+ Usage::
+
+ python examples/bart/summarize.py \
+ --model-dir $HOME/bart.large.cnn \
+ --model-file model.pt \
+ --src $HOME/data-bin/cnn_dm/test.source
+ """
+ parser = argparse.ArgumentParser()
+ parser.add_argument(
+ "--model-dir",
+ required=True,
+ type=str,
+ default="bart.large.cnn/",
+ help="path containing model file and src_dict.txt",
+ )
+ parser.add_argument(
+ "--model-file",
+ default="checkpoint_best.pt",
+ help="where in model_dir are weights saved",
+ )
+ parser.add_argument(
+ "--src", default="test.source", help="text to summarize", type=str
+ )
+ parser.add_argument(
+ "--out", default="test.hypo", help="where to save summaries", type=str
+ )
+ parser.add_argument("--bsz", default=32, help="where to save summaries", type=int)
+ parser.add_argument(
+ "--n", default=None, help="how many examples to summarize", type=int
+ )
+ parser.add_argument(
+ "--xsum-kwargs",
+ action="store_true",
+ default=False,
+ help="if true use XSUM_KWARGS else CNN_KWARGS",
+ )
+ args = parser.parse_args()
+ eval_kwargs = XSUM_KWARGS if args.xsum_kwargs else CNN_KWARGS
+ if args.model_dir == "pytorch/fairseq":
+ bart = torch.hub.load("pytorch/fairseq", args.model_file)
+ else:
+ bart = BARTModel.from_pretrained(
+ args.model_dir,
+ checkpoint_file=args.model_file,
+ data_name_or_path=args.model_dir,
+ )
+ bart = bart.eval()
+ if torch.cuda.is_available():
+ bart = bart.cuda().half()
+ generate(
+ bart, args.src, bsz=args.bsz, n_obs=args.n, outfile=args.out, **eval_kwargs
+ )
+
+
+if __name__ == "__main__":
+ main()
diff --git a/fairseq/examples/byte_level_bpe/README.md b/fairseq/examples/byte_level_bpe/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..657092660eae42d20f67647417623b8b8cb7b66c
--- /dev/null
+++ b/fairseq/examples/byte_level_bpe/README.md
@@ -0,0 +1,88 @@
+# Neural Machine Translation with Byte-Level Subwords
+
+https://arxiv.org/abs/1909.03341
+
+We provide an implementation of byte-level byte-pair encoding (BBPE), taking IWSLT 2017 Fr-En translation as
+example.
+
+## Data
+Get data and generate fairseq binary dataset:
+```bash
+bash ./get_data.sh
+```
+
+## Model Training
+Train Transformer model with Bi-GRU embedding contextualization (implemented in `gru_transformer.py`):
+```bash
+# VOCAB=bytes
+# VOCAB=chars
+VOCAB=bbpe2048
+# VOCAB=bpe2048
+# VOCAB=bbpe4096
+# VOCAB=bpe4096
+# VOCAB=bpe16384
+```
+```bash
+fairseq-train "data/bin_${VOCAB}" --task translation --user-dir examples/byte_level_bpe/gru_transformer \
+ --arch gru_transformer --encoder-layers 2 --decoder-layers 2 --dropout 0.3 --share-all-embeddings \
+ --optimizer adam --adam-betas '(0.9, 0.98)' \
+ --lr 5e-4 --lr-scheduler inverse_sqrt --warmup-updates 4000 \
+ --criterion label_smoothed_cross_entropy --label-smoothing 0.1 \
+ --log-format 'simple' --log-interval 100 --save-dir "checkpoints/${VOCAB}" \
+ --batch-size 100 --max-update 100000 --update-freq 2
+```
+
+## Generation
+`fairseq-generate` requires bytes (BBPE) decoder to convert byte-level representation back to characters:
+```bash
+# BPE=--bpe bytes
+# BPE=--bpe characters
+BPE=--bpe byte_bpe --sentencepiece-model-path data/spm_bbpe2048.model
+# BPE=--bpe sentencepiece --sentencepiece-model data/spm_bpe2048.model
+# BPE=--bpe byte_bpe --sentencepiece-model-path data/spm_bbpe4096.model
+# BPE=--bpe sentencepiece --sentencepiece-model data/spm_bpe4096.model
+# BPE=--bpe sentencepiece --sentencepiece-model data/spm_bpe16384.model
+```
+
+```bash
+fairseq-generate "data/bin_${VOCAB}" --task translation --user-dir examples/byte_level_bpe/gru_transformer \
+ --source-lang fr --gen-subset test --sacrebleu --path "checkpoints/${VOCAB}/checkpoint_last.pt" \
+ --tokenizer moses --moses-target-lang en ${BPE}
+```
+When using `fairseq-interactive`, bytes (BBPE) encoder/decoder is required to tokenize input data and detokenize model predictions:
+```bash
+fairseq-interactive "data/bin_${VOCAB}" --task translation --user-dir examples/byte_level_bpe/gru_transformer \
+ --path "checkpoints/${VOCAB}/checkpoint_last.pt" --input data/test.fr --tokenizer moses --moses-source-lang fr \
+ --moses-target-lang en ${BPE} --buffer-size 1000 --max-tokens 10000
+```
+
+## Results
+| Vocabulary | Model | BLEU |
+|:-------------:|:-------------:|:-------------:|
+| Joint BPE 16k ([Kudo, 2018](https://arxiv.org/abs/1804.10959)) | 512d LSTM 2+2 | 33.81 |
+| Joint BPE 16k | Transformer base 2+2 (w/ GRU) | 36.64 (36.72) |
+| Joint BPE 4k | Transformer base 2+2 (w/ GRU) | 35.49 (36.10) |
+| Joint BBPE 4k | Transformer base 2+2 (w/ GRU) | 35.61 (35.82) |
+| Joint BPE 2k | Transformer base 2+2 (w/ GRU) | 34.87 (36.13) |
+| Joint BBPE 2k | Transformer base 2+2 (w/ GRU) | 34.98 (35.43) |
+| Characters | Transformer base 2+2 (w/ GRU) | 31.78 (33.30) |
+| Bytes | Transformer base 2+2 (w/ GRU) | 31.57 (33.62) |
+
+
+## Citation
+```
+@misc{wang2019neural,
+ title={Neural Machine Translation with Byte-Level Subwords},
+ author={Changhan Wang and Kyunghyun Cho and Jiatao Gu},
+ year={2019},
+ eprint={1909.03341},
+ archivePrefix={arXiv},
+ primaryClass={cs.CL}
+}
+```
+
+
+## Contact
+Changhan Wang ([changhan@fb.com](mailto:changhan@fb.com)),
+Kyunghyun Cho ([kyunghyuncho@fb.com](mailto:kyunghyuncho@fb.com)),
+Jiatao Gu ([jgu@fb.com](mailto:jgu@fb.com))
diff --git a/fairseq/examples/byte_level_bpe/get_bitext.py b/fairseq/examples/byte_level_bpe/get_bitext.py
new file mode 100644
index 0000000000000000000000000000000000000000..6ac1eeec1e6167ec6bafd76b37173ee6987cae7e
--- /dev/null
+++ b/fairseq/examples/byte_level_bpe/get_bitext.py
@@ -0,0 +1,254 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+
+import argparse
+import os
+import os.path as op
+from collections import namedtuple
+from multiprocessing import cpu_count
+from typing import List, Optional
+
+import sentencepiece as sp
+from fairseq.data.encoders.byte_bpe import ByteBPE
+from fairseq.data.encoders.byte_utils import byte_encode
+from fairseq.data.encoders.bytes import Bytes
+from fairseq.data.encoders.characters import Characters
+from fairseq.data.encoders.moses_tokenizer import MosesTokenizer
+from fairseq.data.encoders.sentencepiece_bpe import SentencepieceBPE
+
+
+SPLITS = ["train", "valid", "test"]
+
+
+def _convert_xml(in_path: str, out_path: str):
+ with open(in_path) as f, open(out_path, "w") as f_o:
+ for s in f:
+ ss = s.strip()
+ if not ss.startswith("", "").split('">')
+ assert len(ss) == 2
+ f_o.write(ss[1].strip() + "\n")
+
+
+def _convert_train(in_path: str, out_path: str):
+ with open(in_path) as f, open(out_path, "w") as f_o:
+ for s in f:
+ ss = s.strip()
+ if ss.startswith("<"):
+ continue
+ f_o.write(ss.strip() + "\n")
+
+
+def _get_bytes(in_path: str, out_path: str):
+ with open(in_path) as f, open(out_path, "w") as f_o:
+ for s in f:
+ f_o.write(Bytes.encode(s.strip()) + "\n")
+
+
+def _get_chars(in_path: str, out_path: str):
+ with open(in_path) as f, open(out_path, "w") as f_o:
+ for s in f:
+ f_o.write(Characters.encode(s.strip()) + "\n")
+
+
+def pretokenize(in_path: str, out_path: str, src: str, tgt: str):
+ Args = namedtuple(
+ "Args",
+ [
+ "moses_source_lang",
+ "moses_target_lang",
+ "moses_no_dash_splits",
+ "moses_no_escape",
+ ],
+ )
+ args = Args(
+ moses_source_lang=src,
+ moses_target_lang=tgt,
+ moses_no_dash_splits=False,
+ moses_no_escape=False,
+ )
+ pretokenizer = MosesTokenizer(args)
+ with open(in_path) as f, open(out_path, "w") as f_o:
+ for s in f:
+ f_o.write(pretokenizer.encode(s.strip()) + "\n")
+
+
+def _convert_to_bchar(in_path_prefix: str, src: str, tgt: str, out_path: str):
+ with open(out_path, "w") as f_o:
+ for lang in [src, tgt]:
+ with open(f"{in_path_prefix}.{lang}") as f:
+ for s in f:
+ f_o.write(byte_encode(s.strip()) + "\n")
+
+
+def _get_bpe(in_path: str, model_prefix: str, vocab_size: int):
+ arguments = [
+ f"--input={in_path}",
+ f"--model_prefix={model_prefix}",
+ f"--model_type=bpe",
+ f"--vocab_size={vocab_size}",
+ "--character_coverage=1.0",
+ "--normalization_rule_name=identity",
+ f"--num_threads={cpu_count()}",
+ ]
+ sp.SentencePieceTrainer.Train(" ".join(arguments))
+
+
+def _apply_bbpe(model_path: str, in_path: str, out_path: str):
+ Args = namedtuple("Args", ["sentencepiece_model_path"])
+ args = Args(sentencepiece_model_path=model_path)
+ tokenizer = ByteBPE(args)
+ with open(in_path) as f, open(out_path, "w") as f_o:
+ for s in f:
+ f_o.write(tokenizer.encode(s.strip()) + "\n")
+
+
+def _apply_bpe(model_path: str, in_path: str, out_path: str):
+ Args = namedtuple("Args", ["sentencepiece_model"])
+ args = Args(sentencepiece_model=model_path)
+ tokenizer = SentencepieceBPE(args)
+ with open(in_path) as f, open(out_path, "w") as f_o:
+ for s in f:
+ f_o.write(tokenizer.encode(s.strip()) + "\n")
+
+
+def _concat_files(in_paths: List[str], out_path: str):
+ with open(out_path, "w") as f_o:
+ for p in in_paths:
+ with open(p) as f:
+ for r in f:
+ f_o.write(r)
+
+
+def preprocess_iwslt17(
+ root: str,
+ src: str,
+ tgt: str,
+ bpe_size: Optional[int],
+ need_chars: bool,
+ bbpe_size: Optional[int],
+ need_bytes: bool,
+):
+ # extract bitext
+ in_root = op.join(root, f"{src}-{tgt}")
+ for lang in [src, tgt]:
+ _convert_train(
+ op.join(in_root, f"train.tags.{src}-{tgt}.{lang}"),
+ op.join(root, f"train.{lang}"),
+ )
+ _convert_xml(
+ op.join(in_root, f"IWSLT17.TED.dev2010.{src}-{tgt}.{lang}.xml"),
+ op.join(root, f"valid.{lang}"),
+ )
+ _convert_xml(
+ op.join(in_root, f"IWSLT17.TED.tst2015.{src}-{tgt}.{lang}.xml"),
+ op.join(root, f"test.{lang}"),
+ )
+ # pre-tokenize
+ for lang in [src, tgt]:
+ for split in SPLITS:
+ pretokenize(
+ op.join(root, f"{split}.{lang}"),
+ op.join(root, f"{split}.moses.{lang}"),
+ src,
+ tgt,
+ )
+ # tokenize with BPE vocabulary
+ if bpe_size is not None:
+ # learn vocabulary
+ concated_train_path = op.join(root, "train.all")
+ _concat_files(
+ [op.join(root, "train.moses.fr"), op.join(root, "train.moses.en")],
+ concated_train_path,
+ )
+ bpe_model_prefix = op.join(root, f"spm_bpe{bpe_size}")
+ _get_bpe(concated_train_path, bpe_model_prefix, bpe_size)
+ os.remove(concated_train_path)
+ # apply
+ for lang in [src, tgt]:
+ for split in SPLITS:
+ _apply_bpe(
+ bpe_model_prefix + ".model",
+ op.join(root, f"{split}.moses.{lang}"),
+ op.join(root, f"{split}.moses.bpe{bpe_size}.{lang}"),
+ )
+ # tokenize with bytes vocabulary
+ if need_bytes:
+ for lang in [src, tgt]:
+ for split in SPLITS:
+ _get_bytes(
+ op.join(root, f"{split}.moses.{lang}"),
+ op.join(root, f"{split}.moses.bytes.{lang}"),
+ )
+ # tokenize with characters vocabulary
+ if need_chars:
+ for lang in [src, tgt]:
+ for split in SPLITS:
+ _get_chars(
+ op.join(root, f"{split}.moses.{lang}"),
+ op.join(root, f"{split}.moses.chars.{lang}"),
+ )
+ # tokenize with byte-level BPE vocabulary
+ if bbpe_size is not None:
+ # learn vocabulary
+ bchar_path = op.join(root, "train.bchar")
+ _convert_to_bchar(op.join(root, "train.moses"), src, tgt, bchar_path)
+ bbpe_model_prefix = op.join(root, f"spm_bbpe{bbpe_size}")
+ _get_bpe(bchar_path, bbpe_model_prefix, bbpe_size)
+ os.remove(bchar_path)
+ # apply
+ for lang in [src, tgt]:
+ for split in SPLITS:
+ _apply_bbpe(
+ bbpe_model_prefix + ".model",
+ op.join(root, f"{split}.moses.{lang}"),
+ op.join(root, f"{split}.moses.bbpe{bbpe_size}.{lang}"),
+ )
+
+
+def main():
+ parser = argparse.ArgumentParser()
+ parser.add_argument("--root", type=str, default="data")
+ parser.add_argument(
+ "--bpe-vocab",
+ default=None,
+ type=int,
+ help="Generate tokenized bitext with BPE of size K."
+ "Default to None (disabled).",
+ )
+ parser.add_argument(
+ "--bbpe-vocab",
+ default=None,
+ type=int,
+ help="Generate tokenized bitext with BBPE of size K."
+ "Default to None (disabled).",
+ )
+ parser.add_argument(
+ "--byte-vocab",
+ action="store_true",
+ help="Generate tokenized bitext with bytes vocabulary",
+ )
+ parser.add_argument(
+ "--char-vocab",
+ action="store_true",
+ help="Generate tokenized bitext with chars vocabulary",
+ )
+ args = parser.parse_args()
+
+ preprocess_iwslt17(
+ args.root,
+ "fr",
+ "en",
+ args.bpe_vocab,
+ args.char_vocab,
+ args.bbpe_vocab,
+ args.byte_vocab,
+ )
+
+
+if __name__ == "__main__":
+ main()
diff --git a/fairseq/examples/byte_level_bpe/get_data.sh b/fairseq/examples/byte_level_bpe/get_data.sh
new file mode 100644
index 0000000000000000000000000000000000000000..c3d55d4925a6e6e23d12d293f093c1ae14acf76e
--- /dev/null
+++ b/fairseq/examples/byte_level_bpe/get_data.sh
@@ -0,0 +1,47 @@
+#!/bin/bash
+
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+PY_BIN_ROOT=
+
+# PyPI dependency
+${PY_BIN_ROOT}pip install sentencepiece sacremoses
+
+# Get data
+if [ ! -d "data" ]; then
+ mkdir data
+fi
+
+if [ ! -f "data/fr-en.tgz" ]; then
+ wget https://wit3.fbk.eu/archive/2017-01-trnted/texts/fr/en/fr-en.tgz -P data
+ tar xvf data/fr-en.tgz -C data
+fi
+${PY_BIN_ROOT}python get_bitext.py --bpe-vocab 16384 --byte-vocab --char-vocab
+for VOCAB_SIZE in 2048 4096; do
+ ${PY_BIN_ROOT}python get_bitext.py --bpe-vocab ${VOCAB_SIZE} --bbpe-vocab ${VOCAB_SIZE}
+done
+rm -r data/fr-en data/fr-en.tgz
+
+# Generate binary dataset
+${PY_BIN_ROOT}/fairseq-preprocess --source-lang fr --target-lang en --destdir data/bin_bpe16384 --joined-dictionary \
+ --workers "$(nproc)" --trainpref data/train.moses.bpe16384 --validpref data/valid.moses.bpe16384 \
+ --testpref data/test.moses.bpe16384
+
+${PY_BIN_ROOT}/fairseq-preprocess --source-lang fr --target-lang en --destdir data/bin_bytes --joined-dictionary \
+ --workers "$(nproc)" --trainpref data/train.moses.bytes --validpref data/valid.moses.bytes \
+ --testpref data/test.moses.bytes
+
+${PY_BIN_ROOT}/fairseq-preprocess --source-lang fr --target-lang en --destdir data/bin_chars --joined-dictionary \
+ --workers "$(nproc)" --trainpref data/train.moses.chars --validpref data/valid.moses.chars \
+ --testpref data/test.moses.chars
+
+for VOCAB_SIZE in 2048 4096; do
+ for TYPE in bbpe bpe; do
+ ${PY_BIN_ROOT}/fairseq-preprocess --source-lang fr --target-lang en --destdir "data/bin_${TYPE}${VOCAB_SIZE}" \
+ --joined-dictionary --workers "$(nproc)" --trainpref "data/train.moses.${TYPE}${VOCAB_SIZE}" \
+ --validpref "data/valid.moses.${TYPE}${VOCAB_SIZE}" --testpref "data/test.moses.${TYPE}${VOCAB_SIZE}"
+ done
+done
diff --git a/fairseq/examples/byte_level_bpe/gru_transformer.py b/fairseq/examples/byte_level_bpe/gru_transformer.py
new file mode 100644
index 0000000000000000000000000000000000000000..d4efa93a4d75da71c78e786d7f62101ef3266af4
--- /dev/null
+++ b/fairseq/examples/byte_level_bpe/gru_transformer.py
@@ -0,0 +1,107 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch.nn as nn
+import torch.nn.functional as F
+from fairseq.models import register_model, register_model_architecture
+from fairseq.models.transformer import TransformerEncoder, TransformerModel
+
+
+@register_model("gru_transformer")
+class GRUTransformerModel(TransformerModel):
+ @classmethod
+ def build_encoder(cls, args, src_dict, embed_tokens):
+ return GRUTransformerEncoder(args, src_dict, embed_tokens)
+
+
+class GRUTransformerEncoder(TransformerEncoder):
+ def __init__(self, args, dictionary, embed_tokens):
+ super().__init__(args, dictionary, embed_tokens)
+ self.emb_ctx = nn.GRU(
+ input_size=embed_tokens.embedding_dim,
+ hidden_size=embed_tokens.embedding_dim // 2,
+ num_layers=1,
+ bidirectional=True,
+ )
+
+ def forward_embedding(self, src_tokens):
+ # embed tokens and positions
+ x = embed = self.embed_scale * self.embed_tokens(src_tokens)
+ if self.embed_positions is not None:
+ x = embed + self.embed_positions(src_tokens)
+
+ # contextualize embeddings
+ x = x.transpose(0, 1)
+ x = self.dropout_module(x)
+ x, _ = self.emb_ctx.forward(x)
+ x = x.transpose(0, 1)
+
+ if self.layernorm_embedding is not None:
+ x = self.layernorm_embedding(x)
+ x = self.dropout_module(x)
+ return x, embed
+
+
+@register_model_architecture("gru_transformer", "gru_transformer")
+def gru_transformer_base_architecture(args):
+ args.encoder_embed_path = getattr(args, "encoder_embed_path", None)
+ args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512)
+ args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 2048)
+ args.encoder_layers = getattr(args, "encoder_layers", 6)
+ args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 8)
+ args.encoder_normalize_before = getattr(args, "encoder_normalize_before", False)
+ args.encoder_learned_pos = getattr(args, "encoder_learned_pos", False)
+ args.decoder_embed_path = getattr(args, "decoder_embed_path", None)
+ args.decoder_embed_dim = getattr(args, "decoder_embed_dim", args.encoder_embed_dim)
+ args.decoder_ffn_embed_dim = getattr(
+ args, "decoder_ffn_embed_dim", args.encoder_ffn_embed_dim
+ )
+ args.decoder_layers = getattr(args, "decoder_layers", 6)
+ args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 8)
+ args.decoder_normalize_before = getattr(args, "decoder_normalize_before", False)
+ args.decoder_learned_pos = getattr(args, "decoder_learned_pos", False)
+ args.attention_dropout = getattr(args, "attention_dropout", 0.0)
+ args.activation_dropout = getattr(args, "activation_dropout", 0.0)
+ args.activation_fn = getattr(args, "activation_fn", "relu")
+ args.dropout = getattr(args, "dropout", 0.1)
+ args.adaptive_softmax_cutoff = getattr(args, "adaptive_softmax_cutoff", None)
+ args.adaptive_softmax_dropout = getattr(args, "adaptive_softmax_dropout", 0)
+ args.share_decoder_input_output_embed = getattr(
+ args, "share_decoder_input_output_embed", False
+ )
+ args.share_all_embeddings = getattr(args, "share_all_embeddings", False)
+ args.no_token_positional_embeddings = getattr(
+ args, "no_token_positional_embeddings", False
+ )
+ args.adaptive_input = getattr(args, "adaptive_input", False)
+ args.no_cross_attention = getattr(args, "no_cross_attention", False)
+ args.cross_self_attention = getattr(args, "cross_self_attention", False)
+ args.layer_wise_attention = getattr(args, "layer_wise_attention", False)
+
+ args.decoder_output_dim = getattr(
+ args, "decoder_output_dim", args.decoder_embed_dim
+ )
+ args.decoder_input_dim = getattr(args, "decoder_input_dim", args.decoder_embed_dim)
+
+ args.no_scale_embedding = getattr(args, "no_scale_embedding", False)
+ args.layernorm_embedding = getattr(args, "layernorm_embedding", False)
+
+
+@register_model_architecture("gru_transformer", "gru_transformer_big")
+def gru_transformer_big(args):
+ args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 1024)
+ args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 4096)
+ args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 16)
+ args.encoder_normalize_before = getattr(args, "encoder_normalize_before", False)
+ args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 1024)
+ args.decoder_ffn_embed_dim = getattr(args, "decoder_ffn_embed_dim", 4096)
+ args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 16)
+ args.dropout = getattr(args, "dropout", 0.3)
+ gru_transformer_base_architecture(args)
diff --git a/fairseq/examples/camembert/README.md b/fairseq/examples/camembert/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..5ef4fe3f151bb468712f3be935ea5bb1b1360bf7
--- /dev/null
+++ b/fairseq/examples/camembert/README.md
@@ -0,0 +1,75 @@
+# CamemBERT: a Tasty French Language Model
+
+## Introduction
+
+[CamemBERT](https://arxiv.org/abs/1911.03894) is a pretrained language model trained on 138GB of French text based on RoBERTa.
+
+Also available in [github.com/huggingface/transformers](https://github.com/huggingface/transformers/).
+
+## Pre-trained models
+
+| Model | #params | Download | Arch. | Training data |
+|--------------------------------|---------|--------------------------------------------------------------------------------------------------------------------------|-------|-----------------------------------|
+| `camembert` / `camembert-base` | 110M | [camembert-base.tar.gz](https://dl.fbaipublicfiles.com/fairseq/models/camembert-base.tar.gz) | Base | OSCAR (138 GB of text) |
+| `camembert-large` | 335M | [camembert-large.tar.gz](https://dl.fbaipublicfiles.com/fairseq/models/camembert-large.tar.gz) | Large | CCNet (135 GB of text) |
+| `camembert-base-ccnet` | 110M | [camembert-base-ccnet.tar.gz](https://dl.fbaipublicfiles.com/fairseq/models/camembert-base-ccnet.tar.gz) | Base | CCNet (135 GB of text) |
+| `camembert-base-wikipedia-4gb` | 110M | [camembert-base-wikipedia-4gb.tar.gz](https://dl.fbaipublicfiles.com/fairseq/models/camembert-base-wikipedia-4gb.tar.gz) | Base | Wikipedia (4 GB of text) |
+| `camembert-base-oscar-4gb` | 110M | [camembert-base-oscar-4gb.tar.gz](https://dl.fbaipublicfiles.com/fairseq/models/camembert-base-oscar-4gb.tar.gz) | Base | Subsample of OSCAR (4 GB of text) |
+| `camembert-base-ccnet-4gb` | 110M | [camembert-base-ccnet-4gb.tar.gz](https://dl.fbaipublicfiles.com/fairseq/models/camembert-base-ccnet-4gb.tar.gz) | Base | Subsample of CCNet (4 GB of text) |
+
+## Example usage
+
+### fairseq
+##### Load CamemBERT from torch.hub (PyTorch >= 1.1):
+```python
+import torch
+camembert = torch.hub.load('pytorch/fairseq', 'camembert')
+camembert.eval() # disable dropout (or leave in train mode to finetune)
+```
+
+##### Load CamemBERT (for PyTorch 1.0 or custom models):
+```python
+# Download camembert model
+wget https://dl.fbaipublicfiles.com/fairseq/models/camembert-base.tar.gz
+tar -xzvf camembert.tar.gz
+
+# Load the model in fairseq
+from fairseq.models.roberta import CamembertModel
+camembert = CamembertModel.from_pretrained('/path/to/camembert')
+camembert.eval() # disable dropout (or leave in train mode to finetune)
+```
+
+##### Filling masks:
+```python
+masked_line = 'Le camembert est :)'
+camembert.fill_mask(masked_line, topk=3)
+# [('Le camembert est délicieux :)', 0.4909118115901947, ' délicieux'),
+# ('Le camembert est excellent :)', 0.10556942224502563, ' excellent'),
+# ('Le camembert est succulent :)', 0.03453322499990463, ' succulent')]
+```
+
+##### Extract features from Camembert:
+```python
+# Extract the last layer's features
+line = "J'aime le camembert !"
+tokens = camembert.encode(line)
+last_layer_features = camembert.extract_features(tokens)
+assert last_layer_features.size() == torch.Size([1, 10, 768])
+
+# Extract all layer's features (layer 0 is the embedding layer)
+all_layers = camembert.extract_features(tokens, return_all_hiddens=True)
+assert len(all_layers) == 13
+assert torch.all(all_layers[-1] == last_layer_features)
+```
+
+## Citation
+If you use our work, please cite:
+
+```bibtex
+@inproceedings{martin2020camembert,
+ title={CamemBERT: a Tasty French Language Model},
+ author={Martin, Louis and Muller, Benjamin and Su{\'a}rez, Pedro Javier Ortiz and Dupont, Yoann and Romary, Laurent and de la Clergerie, {\'E}ric Villemonte and Seddah, Djam{\'e} and Sagot, Beno{\^\i}t},
+ booktitle={Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics},
+ year={2020}
+}
+```
diff --git a/fairseq/examples/constrained_decoding/README.md b/fairseq/examples/constrained_decoding/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..e04b8b6a018214c8233fa87fd91d46a6dd1519d4
--- /dev/null
+++ b/fairseq/examples/constrained_decoding/README.md
@@ -0,0 +1,123 @@
+# (Vectorized) Lexically constrained decoding with dynamic beam allocation
+
+This page provides instructions for how to use lexically constrained decoding in Fairseq.
+Fairseq implements the code described in the following papers:
+
+* [Fast Lexically Constrained Decoding With Dynamic Beam Allocation](https://www.aclweb.org/anthology/N18-1119/) (Post & Vilar, 2018)
+* [Improved Lexically Constrained Decoding for Translation and Monolingual Rewriting](https://www.aclweb.org/anthology/N19-1090/) (Hu et al., 2019)
+
+## Quick start
+
+Constrained search is enabled by adding the command-line argument `--constraints` to `fairseq-interactive`.
+Constraints are appended to each line of input, separated by tabs. Each constraint (one or more tokens)
+is a separate field.
+
+The following command, using [Fairseq's WMT19 German--English model](https://github.com/pytorch/fairseq/blob/main/examples/wmt19/README.md),
+translates the sentence *Die maschinelle Übersetzung ist schwer zu kontrollieren.* with the constraints
+"hard" and "to influence".
+
+ echo -e "Die maschinelle Übersetzung ist schwer zu kontrollieren.\thard\ttoinfluence" \
+ | normalize.py | tok.py \
+ | fairseq-interactive /path/to/model \
+ --path /path/to/model/model1.pt \
+ --bpe fastbpe \
+ --bpe-codes /path/to/model/bpecodes \
+ --constraints \
+ -s de -t en \
+ --beam 10
+
+(tok.py and normalize.py can be found in the same directory as this README; they are just shortcuts around Fairseq's WMT19 preprocessing).
+This will generate the following output:
+
+ [snip]
+ S-0 Die masch@@ in@@ elle Über@@ setzung ist schwer zu kontrollieren .
+ W-0 1.844 seconds
+ C-0 hard
+ C-0 influence
+ H-0 -1.5333266258239746 Mach@@ ine trans@@ lation is hard to influence .
+ D-0 -1.5333266258239746 Machine translation is hard to influence .
+ P-0 -0.5434 -0.1423 -0.1930 -0.1415 -0.2346 -1.8031 -0.1701 -11.7727 -0.1815 -0.1511
+
+By default, constraints are generated in the order supplied, with any number (zero or more) of tokens generated
+between constraints. If you wish for the decoder to order the constraints, then use `--constraints unordered`.
+Note that you may want to use a larger beam.
+
+## Implementation details
+
+The heart of the implementation is in `fairseq/search.py`, which adds a `LexicallyConstrainedBeamSearch` instance.
+This instance of beam search tracks the progress of each hypothesis in the beam through the set of constraints
+provided for each input sentence. It does this using one of two classes, both found in `fairseq/token_generation_contstraints.py`:
+
+* OrderedConstraintState: assumes the `C` input constraints will be generated in the provided order
+* UnorderedConstraintState: tries to apply `C` (phrasal) constraints in all `C!` orders
+
+## Differences from Sockeye
+
+There are a number of [differences from Sockeye's implementation](https://awslabs.github.io/sockeye/inference.html#lexical-constraints).
+
+* Generating constraints in the order supplied (the default option here) is not available in Sockeye.
+* Due to an improved beam allocation method, there is no need to prune the beam.
+* Again due to better allocation, beam sizes as low as 10 or even 5 are often sufficient.
+* [The vector extensions described in Hu et al.](https://github.com/edwardjhu/sockeye/tree/trie_constraints) (NAACL 2019) were never merged
+ into the main Sockeye branch.
+
+## Citation
+
+The paper first describing lexical constraints for seq2seq decoding is:
+
+```bibtex
+@inproceedings{hokamp-liu-2017-lexically,
+ title = "Lexically Constrained Decoding for Sequence Generation Using Grid Beam Search",
+ author = "Hokamp, Chris and
+ Liu, Qun",
+ booktitle = "Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
+ month = jul,
+ year = "2017",
+ address = "Vancouver, Canada",
+ publisher = "Association for Computational Linguistics",
+ url = "https://www.aclweb.org/anthology/P17-1141",
+ doi = "10.18653/v1/P17-1141",
+ pages = "1535--1546",
+}
+```
+
+The fairseq implementation uses the extensions described in
+
+```bibtex
+@inproceedings{post-vilar-2018-fast,
+ title = "Fast Lexically Constrained Decoding with Dynamic Beam Allocation for Neural Machine Translation",
+ author = "Post, Matt and
+ Vilar, David",
+ booktitle = "Proceedings of the 2018 Conference of the North {A}merican Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers)",
+ month = jun,
+ year = "2018",
+ address = "New Orleans, Louisiana",
+ publisher = "Association for Computational Linguistics",
+ url = "https://www.aclweb.org/anthology/N18-1119",
+ doi = "10.18653/v1/N18-1119",
+ pages = "1314--1324",
+}
+```
+
+and
+
+```bibtex
+@inproceedings{hu-etal-2019-improved,
+ title = "Improved Lexically Constrained Decoding for Translation and Monolingual Rewriting",
+ author = "Hu, J. Edward and
+ Khayrallah, Huda and
+ Culkin, Ryan and
+ Xia, Patrick and
+ Chen, Tongfei and
+ Post, Matt and
+ Van Durme, Benjamin",
+ booktitle = "Proceedings of the 2019 Conference of the North {A}merican Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers)",
+ month = jun,
+ year = "2019",
+ address = "Minneapolis, Minnesota",
+ publisher = "Association for Computational Linguistics",
+ url = "https://www.aclweb.org/anthology/N19-1090",
+ doi = "10.18653/v1/N19-1090",
+ pages = "839--850",
+}
+```
diff --git a/fairseq/examples/constrained_decoding/normalize.py b/fairseq/examples/constrained_decoding/normalize.py
new file mode 100644
index 0000000000000000000000000000000000000000..4ae2b5111ba025acb9e1613865c92fdc339a58d5
--- /dev/null
+++ b/fairseq/examples/constrained_decoding/normalize.py
@@ -0,0 +1,27 @@
+#!/usr/bin/env python3
+#
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import sys
+
+from sacremoses.normalize import MosesPunctNormalizer
+
+
+def main(args):
+ normalizer = MosesPunctNormalizer(lang=args.lang, penn=args.penn)
+ for line in sys.stdin:
+ print(normalizer.normalize(line.rstrip()), flush=True)
+
+
+if __name__ == "__main__":
+ import argparse
+
+ parser = argparse.ArgumentParser()
+ parser.add_argument("--lang", "-l", default="en")
+ parser.add_argument("--penn", "-p", action="store_true")
+ args = parser.parse_args()
+
+ main(args)
diff --git a/fairseq/examples/constrained_decoding/tok.py b/fairseq/examples/constrained_decoding/tok.py
new file mode 100644
index 0000000000000000000000000000000000000000..b1f888a8c0d1b8ec7174859476cc3222456e0d2c
--- /dev/null
+++ b/fairseq/examples/constrained_decoding/tok.py
@@ -0,0 +1,34 @@
+#!/usr/bin/env python3
+#
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import sys
+
+import sacremoses
+
+
+def main(args):
+ """Tokenizes, preserving tabs"""
+ mt = sacremoses.MosesTokenizer(lang=args.lang)
+
+ def tok(s):
+ return mt.tokenize(s, return_str=True)
+
+ for line in sys.stdin:
+ parts = list(map(tok, line.split("\t")))
+ print(*parts, sep="\t", flush=True)
+
+
+if __name__ == "__main__":
+ import argparse
+
+ parser = argparse.ArgumentParser()
+ parser.add_argument("--lang", "-l", default="en")
+ parser.add_argument("--penn", "-p", action="store_true")
+ parser.add_argument("--fields", "-f", help="fields to tokenize")
+ args = parser.parse_args()
+
+ main(args)
diff --git a/fairseq/examples/conv_seq2seq/README.md b/fairseq/examples/conv_seq2seq/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..95fe7e7909a77ee0e50fe31d4b8be38daa8f3be7
--- /dev/null
+++ b/fairseq/examples/conv_seq2seq/README.md
@@ -0,0 +1,25 @@
+# Convolutional Sequence to Sequence Learning (Gehring et al., 2017)
+
+## Pre-trained models
+
+Description | Dataset | Model | Test set(s)
+---|---|---|---
+Convolutional ([Gehring et al., 2017](https://arxiv.org/abs/1705.03122)) | [WMT14 English-French](http://statmt.org/wmt14/translation-task.html#Download) | [download (.tar.bz2)](https://dl.fbaipublicfiles.com/fairseq/models/wmt14.v2.en-fr.fconv-py.tar.bz2) | newstest2014: [download (.tar.bz2)](https://dl.fbaipublicfiles.com/fairseq/data/wmt14.v2.en-fr.newstest2014.tar.bz2) newstest2012/2013: [download (.tar.bz2)](https://dl.fbaipublicfiles.com/fairseq/data/wmt14.v2.en-fr.ntst1213.tar.bz2)
+Convolutional ([Gehring et al., 2017](https://arxiv.org/abs/1705.03122)) | [WMT14 English-German](http://statmt.org/wmt14/translation-task.html#Download) | [download (.tar.bz2)](https://dl.fbaipublicfiles.com/fairseq/models/wmt14.en-de.fconv-py.tar.bz2) | newstest2014: [download (.tar.bz2)](https://dl.fbaipublicfiles.com/fairseq/data/wmt14.en-de.newstest2014.tar.bz2)
+Convolutional ([Gehring et al., 2017](https://arxiv.org/abs/1705.03122)) | [WMT17 English-German](http://statmt.org/wmt17/translation-task.html#Download) | [download (.tar.bz2)](https://dl.fbaipublicfiles.com/fairseq/models/wmt17.v2.en-de.fconv-py.tar.bz2) | newstest2014: [download (.tar.bz2)](https://dl.fbaipublicfiles.com/fairseq/data/wmt17.v2.en-de.newstest2014.tar.bz2)
+
+## Example usage
+
+See the [translation README](../translation/README.md) for instructions on reproducing results for WMT'14 En-De and
+WMT'14 En-Fr using the `fconv_wmt_en_de` and `fconv_wmt_en_fr` model architectures.
+
+## Citation
+
+```bibtex
+@inproceedings{gehring2017convs2s,
+ title = {Convolutional Sequence to Sequence Learning},
+ author = {Gehring, Jonas, and Auli, Michael and Grangier, David and Yarats, Denis and Dauphin, Yann N},
+ booktitle = {Proc. of ICML},
+ year = 2017,
+}
+```
diff --git a/fairseq/examples/criss/README.md b/fairseq/examples/criss/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..4689ed7c10497a5100b28fe6d6801a7c089da569
--- /dev/null
+++ b/fairseq/examples/criss/README.md
@@ -0,0 +1,61 @@
+# Cross-lingual Retrieval for Iterative Self-Supervised Training
+
+https://arxiv.org/pdf/2006.09526.pdf
+
+## Introduction
+
+CRISS is a multilingual sequence-to-sequnce pretraining method where mining and training processes are applied iteratively, improving cross-lingual alignment and translation ability at the same time.
+
+## Requirements:
+
+* faiss: https://github.com/facebookresearch/faiss
+* mosesdecoder: https://github.com/moses-smt/mosesdecoder
+* flores: https://github.com/facebookresearch/flores
+* LASER: https://github.com/facebookresearch/LASER
+
+## Unsupervised Machine Translation
+##### 1. Download and decompress CRISS checkpoints
+```
+cd examples/criss
+wget https://dl.fbaipublicfiles.com/criss/criss_3rd_checkpoints.tar.gz
+tar -xf criss_checkpoints.tar.gz
+```
+##### 2. Download and preprocess Flores test dataset
+Make sure to run all scripts from examples/criss directory
+```
+bash download_and_preprocess_flores_test.sh
+```
+
+##### 3. Run Evaluation on Sinhala-English
+```
+bash unsupervised_mt/eval.sh
+```
+
+## Sentence Retrieval
+##### 1. Download and preprocess Tatoeba dataset
+```
+bash download_and_preprocess_tatoeba.sh
+```
+
+##### 2. Run Sentence Retrieval on Tatoeba Kazakh-English
+```
+bash sentence_retrieval/sentence_retrieval_tatoeba.sh
+```
+
+## Mining
+##### 1. Install faiss
+Follow instructions on https://github.com/facebookresearch/faiss/blob/master/INSTALL.md
+##### 2. Mine pseudo-parallel data between Kazakh and English
+```
+bash mining/mine_example.sh
+```
+
+## Citation
+```bibtex
+@article{tran2020cross,
+ title={Cross-lingual retrieval for iterative self-supervised training},
+ author={Tran, Chau and Tang, Yuqing and Li, Xian and Gu, Jiatao},
+ journal={arXiv preprint arXiv:2006.09526},
+ year={2020}
+}
+```
diff --git a/fairseq/examples/criss/download_and_preprocess_flores_test.sh b/fairseq/examples/criss/download_and_preprocess_flores_test.sh
new file mode 100644
index 0000000000000000000000000000000000000000..ed4b390fbdee3991efeb298050e12065d7fe605b
--- /dev/null
+++ b/fairseq/examples/criss/download_and_preprocess_flores_test.sh
@@ -0,0 +1,64 @@
+#!/bin/bash
+# Copyright (c) Facebook, Inc. and its affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+SPM_ENCODE=flores/scripts/spm_encode.py
+DATA=data_tmp
+SPM_MODEL=criss_checkpoints/sentence.bpe.model
+DICT=criss_checkpoints/dict.txt
+
+download_data() {
+ CORPORA=$1
+ URL=$2
+
+ if [ -f $CORPORA ]; then
+ echo "$CORPORA already exists, skipping download"
+ else
+ echo "Downloading $URL"
+ wget $URL -O $CORPORA --no-check-certificate || rm -f $CORPORA
+ if [ -f $CORPORA ]; then
+ echo "$URL successfully downloaded."
+ else
+ echo "$URL not successfully downloaded."
+ rm -f $CORPORA
+ fi
+ fi
+}
+
+if [[ -f flores ]]; then
+ echo "flores already cloned"
+else
+ git clone https://github.com/facebookresearch/flores
+fi
+
+mkdir -p $DATA
+download_data $DATA/wikipedia_en_ne_si_test_sets.tgz "https://github.com/facebookresearch/flores/raw/master/data/wikipedia_en_ne_si_test_sets.tgz"
+pushd $DATA
+pwd
+tar -vxf wikipedia_en_ne_si_test_sets.tgz
+popd
+
+
+for lang in ne_NP si_LK; do
+ datadir=$DATA/${lang}-en_XX-flores
+ rm -rf $datadir
+ mkdir -p $datadir
+ TEST_PREFIX=$DATA/wikipedia_en_ne_si_test_sets/wikipedia.test
+ python $SPM_ENCODE \
+ --model ${SPM_MODEL} \
+ --output_format=piece \
+ --inputs ${TEST_PREFIX}.${lang:0:2}-en.${lang:0:2} ${TEST_PREFIX}.${lang:0:2}-en.en \
+ --outputs $datadir/test.bpe.${lang}-en_XX.${lang} $datadir/test.bpe.${lang}-en_XX.en_XX
+
+ # binarize data
+ fairseq-preprocess \
+ --source-lang ${lang} --target-lang en_XX \
+ --testpref $datadir/test.bpe.${lang}-en_XX \
+ --destdir $datadir \
+ --srcdict ${DICT} \
+ --joined-dictionary \
+ --workers 4
+done
diff --git a/fairseq/examples/criss/download_and_preprocess_tatoeba.sh b/fairseq/examples/criss/download_and_preprocess_tatoeba.sh
new file mode 100644
index 0000000000000000000000000000000000000000..7ed64f017d5e62695ba73745c840507b994abc0f
--- /dev/null
+++ b/fairseq/examples/criss/download_and_preprocess_tatoeba.sh
@@ -0,0 +1,46 @@
+#!/bin/bash
+# Copyright (c) Facebook, Inc. and its affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+SPM_ENCODE=flores/scripts/spm_encode.py
+DATA=data_tmp
+SPM_MODEL=criss_checkpoints/sentence.bpe.model
+DICT=criss_checkpoints/dict.txt
+
+if [[ -f flores ]]; then
+ echo "flores already cloned"
+else
+ git clone https://github.com/facebookresearch/flores
+fi
+if [[ -f LASER ]]; then
+ echo "LASER already cloned"
+else
+ git clone https://github.com/facebookresearch/LASER
+fi
+mkdir -p data_tmp
+declare -A lang_tatoeba_map=( ["ar_AR"]="ara" ["de_DE"]="deu" ["es_XX"]="spa" ["et_EE"]="est" ["fi_FI"]="fin" ["fr_XX"]="fra" ["hi_IN"]="hin" ["it_IT"]="ita" ["ja_XX"]="jpn" ["ko_KR"]="kor" ["kk_KZ"]="kaz" ["nl_XX"]="nld" ["ru_RU"]="rus" ["tr_TR"]="tur" ["vi_VN"]="vie" ["zh_CN"]="cmn")
+for lang in ar_AR de_DE es_XX et_EE fi_FI fr_XX hi_IN it_IT ja_XX kk_KZ ko_KR nl_XX ru_RU tr_TR vi_VN zh_CN; do
+ lang_tatoeba=${lang_tatoeba_map[$lang]}
+ echo $lang_tatoeba
+ datadir=$DATA/${lang}-en_XX-tatoeba
+ rm -rf $datadir
+ mkdir -p $datadir
+ TEST_PREFIX=LASER/data/tatoeba/v1/tatoeba
+ python $SPM_ENCODE \
+ --model ${SPM_MODEL} \
+ --output_format=piece \
+ --inputs ${TEST_PREFIX}.${lang_tatoeba}-eng.${lang_tatoeba} ${TEST_PREFIX}.${lang_tatoeba}-eng.eng \
+ --outputs $datadir/test.bpe.${lang}-en_XX.${lang} $datadir/test.bpe.${lang}-en_XX.en_XX
+
+ # binarize data
+ fairseq-preprocess \
+ --source-lang ${lang} --target-lang en_XX \
+ --testpref $datadir/test.bpe.${lang}-en_XX \
+ --destdir $datadir \
+ --srcdict ${DICT} \
+ --joined-dictionary \
+ --workers 4
+done
diff --git a/fairseq/examples/criss/mining/mine.py b/fairseq/examples/criss/mining/mine.py
new file mode 100644
index 0000000000000000000000000000000000000000..c872da196fe0df776622365748ad7963fee1f0a0
--- /dev/null
+++ b/fairseq/examples/criss/mining/mine.py
@@ -0,0 +1,240 @@
+#!/usr/bin/env python3 -u
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+import argparse
+import glob
+from subprocess import check_call
+
+try:
+ import faiss
+
+ has_faiss = True
+except ImportError:
+ has_faiss = False
+import numpy as np
+
+
+GB = 1024 * 1024 * 1024
+
+
+def call(cmd):
+ print(cmd)
+ check_call(cmd, shell=True)
+
+
+def get_batches(directory, lang, prefix="all_avg_pool"):
+ print(f"Finding in {directory}/{prefix}.{lang}*")
+ files = glob.glob(f"{directory}/{prefix}.{lang}*")
+ emb_files = []
+ txt_files = []
+ for emb_fi in files:
+ emb_files.append(emb_fi)
+ txt_fi = emb_fi.replace(prefix, "sentences")
+ txt_files.append(txt_fi)
+ return emb_files, txt_files
+
+
+def load_batch(emb_file, dim):
+ embeddings = np.fromfile(emb_file, dtype=np.float32)
+ num_rows = int(embeddings.shape[0] / dim)
+ embeddings = embeddings.reshape((num_rows, dim))
+ faiss.normalize_L2(embeddings)
+ return embeddings
+
+
+def knnGPU_sharded(x_batches_f, y_batches_f, dim, k, direction="x2y"):
+ if not has_faiss:
+ raise ImportError("Please install Faiss")
+ sims = []
+ inds = []
+ xfrom = 0
+ xto = 0
+ for x_batch_f in x_batches_f:
+ yfrom = 0
+ yto = 0
+ x_batch = load_batch(x_batch_f, dim)
+ xto = xfrom + x_batch.shape[0]
+ bsims, binds = [], []
+ for y_batch_f in y_batches_f:
+ y_batch = load_batch(y_batch_f, dim)
+ neighbor_size = min(k, y_batch.shape[0])
+ yto = yfrom + y_batch.shape[0]
+ print("{}-{} -> {}-{}".format(xfrom, xto, yfrom, yto))
+ idx = faiss.IndexFlatIP(dim)
+ idx = faiss.index_cpu_to_all_gpus(idx)
+ idx.add(y_batch)
+ bsim, bind = idx.search(x_batch, neighbor_size)
+
+ bsims.append(bsim)
+ binds.append(bind + yfrom)
+ yfrom += y_batch.shape[0]
+ del idx
+ del y_batch
+ bsims = np.concatenate(bsims, axis=1)
+ binds = np.concatenate(binds, axis=1)
+ aux = np.argsort(-bsims, axis=1)
+ sim_batch = np.zeros((x_batch.shape[0], k), dtype=np.float32)
+ ind_batch = np.zeros((x_batch.shape[0], k), dtype=np.int64)
+ for i in range(x_batch.shape[0]):
+ for j in range(k):
+ sim_batch[i, j] = bsims[i, aux[i, j]]
+ ind_batch[i, j] = binds[i, aux[i, j]]
+ sims.append(sim_batch)
+ inds.append(ind_batch)
+ xfrom += x_batch.shape[0]
+ del x_batch
+ sim = np.concatenate(sims, axis=0)
+ ind = np.concatenate(inds, axis=0)
+ return sim, ind
+
+
+def score(sim, fwd_mean, bwd_mean, margin):
+ return margin(sim, (fwd_mean + bwd_mean) / 2)
+
+
+def score_candidates(
+ sim_mat, candidate_inds, fwd_mean, bwd_mean, margin, verbose=False
+):
+ print(" - scoring {:d} candidates".format(sim_mat.shape[0]))
+ scores = np.zeros(candidate_inds.shape)
+ for i in range(scores.shape[0]):
+ for j in range(scores.shape[1]):
+ k = int(candidate_inds[i, j])
+ scores[i, j] = score(sim_mat[i, j], fwd_mean[i], bwd_mean[k], margin)
+ return scores
+
+
+def load_text(files):
+ all_sentences = []
+ for fi in files:
+ with open(fi) as sentence_fi:
+ for line in sentence_fi:
+ all_sentences.append(line.strip())
+ print(f"Read {len(all_sentences)} sentences")
+ return all_sentences
+
+
+if __name__ == "__main__":
+ parser = argparse.ArgumentParser(description="Mine bitext")
+ parser.add_argument("--src-lang", help="Source language")
+ parser.add_argument("--tgt-lang", help="Target language")
+ parser.add_argument(
+ "--dict-path", help="Path to dictionary file", default="dict.txt"
+ )
+ parser.add_argument(
+ "--spm-path", help="Path to SPM model file", default="sentence.bpe.model"
+ )
+ parser.add_argument("--dim", type=int, default=1024, help="Embedding dimension")
+ parser.add_argument("--mem", type=int, default=5, help="Memory in GB")
+ parser.add_argument("--src-dir", help="Source directory")
+ parser.add_argument("--tgt-dir", help="Target directory")
+ parser.add_argument("--output", help="Output path")
+ parser.add_argument(
+ "--neighborhood", type=int, default=4, help="Embedding dimension"
+ )
+ parser.add_argument(
+ "--threshold", type=float, default=1.06, help="Threshold on mined bitext"
+ )
+ parser.add_argument(
+ "--valid-size",
+ type=int,
+ default=2000,
+ help="Number of sentences used for validation set",
+ )
+ parser.add_argument(
+ "--min-count",
+ type=int,
+ default=50000,
+ help="Min num sentences used for each language",
+ )
+ args = parser.parse_args()
+
+ x_batches_f, x_sents_f = get_batches(args.src_dir, args.src_lang)
+ y_batches_f, y_sents_f = get_batches(args.tgt_dir, args.tgt_lang)
+ margin = lambda a, b: a / b
+ y2x_sim, y2x_ind = knnGPU_sharded(
+ y_batches_f, x_batches_f, args.dim, args.neighborhood, direction="y2x"
+ )
+ x2y_sim, x2y_ind = knnGPU_sharded(
+ x_batches_f, y_batches_f, args.dim, args.neighborhood, direction="x2y"
+ )
+
+ x2y_mean = x2y_sim.mean(axis=1)
+ y2x_mean = y2x_sim.mean(axis=1)
+ fwd_scores = score_candidates(x2y_sim, x2y_ind, x2y_mean, y2x_mean, margin)
+ bwd_scores = score_candidates(y2x_sim, y2x_ind, y2x_mean, x2y_mean, margin)
+ fwd_best = x2y_ind[np.arange(x2y_sim.shape[0]), fwd_scores.argmax(axis=1)]
+ bwd_best = y2x_ind[np.arange(y2x_sim.shape[0]), bwd_scores.argmax(axis=1)]
+ indices = np.stack(
+ (
+ np.concatenate((np.arange(x2y_ind.shape[0]), bwd_best)),
+ np.concatenate((fwd_best, np.arange(y2x_ind.shape[0]))),
+ ),
+ axis=1,
+ )
+ scores = np.concatenate((fwd_scores.max(axis=1), bwd_scores.max(axis=1)))
+
+ x_sentences = load_text(x_sents_f)
+ y_sentences = load_text(y_sents_f)
+
+ threshold = args.threshold
+ min_count = args.min_count
+ seen_src, seen_trg = set(), set()
+ directory = args.output
+ call(f"mkdir -p {directory}")
+ src_out = open(
+ f"{directory}/all.{args.src_lang}",
+ mode="w",
+ encoding="utf-8",
+ errors="surrogateescape",
+ )
+ tgt_out = open(
+ f"{directory}/all.{args.tgt_lang}",
+ mode="w",
+ encoding="utf-8",
+ errors="surrogateescape",
+ )
+ scores_out = open(
+ f"{directory}/all.scores", mode="w", encoding="utf-8", errors="surrogateescape"
+ )
+ count = 0
+ for i in np.argsort(-scores):
+ src_ind, trg_ind = indices[i]
+ if src_ind not in seen_src and trg_ind not in seen_trg:
+ seen_src.add(src_ind)
+ seen_trg.add(trg_ind)
+ if scores[i] > threshold or count < min_count:
+ if x_sentences[src_ind]:
+ print(scores[i], file=scores_out)
+ print(x_sentences[src_ind], file=src_out)
+ print(y_sentences[trg_ind], file=tgt_out)
+ count += 1
+ else:
+ print(f"Ignoring sentence: {x_sentences[src_ind]}")
+ src_out.close()
+ tgt_out.close()
+ scores_out.close()
+
+ print(f"Found {count} pairs for threshold={threshold}")
+ with open(f"{directory}/all.{args.src_lang}") as all_s, open(
+ f"{directory}/all.{args.tgt_lang}"
+ ) as all_t, open(f"{directory}/valid.{args.src_lang}", "w") as valid_s, open(
+ f"{directory}/valid.{args.tgt_lang}", "w"
+ ) as valid_t, open(
+ f"{directory}/train.{args.src_lang}", "w"
+ ) as train_s, open(
+ f"{directory}/train.{args.tgt_lang}", "w"
+ ) as train_t:
+ count = 0
+ for s_line, t_line in zip(all_s, all_t):
+ s_line = s_line.split("\t")[1]
+ t_line = t_line.split("\t")[1]
+ if count >= args.valid_size:
+ train_s.write(s_line)
+ train_t.write(t_line)
+ else:
+ valid_s.write(s_line)
+ valid_t.write(t_line)
+ count += 1
diff --git a/fairseq/examples/criss/mining/mine_example.sh b/fairseq/examples/criss/mining/mine_example.sh
new file mode 100644
index 0000000000000000000000000000000000000000..ace995ac44665f99d904b6a89d7fbbce24103afe
--- /dev/null
+++ b/fairseq/examples/criss/mining/mine_example.sh
@@ -0,0 +1,103 @@
+#!/bin/bash
+# Copyright (c) Facebook, Inc. and its affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+#
+source_lang=kk_KZ
+target_lang=en_XX
+MODEL=criss_checkpoints/criss.3rd.pt
+SPM=criss_checkpoints/sentence.bpe.model
+SPLIT=test
+LANG_DICT=criss_checkpoints/lang_dict.txt
+SPM_ENCODE=flores/scripts/spm_encode.py
+SAVE_ENCODER=save_encoder.py
+ENCODER_SAVE_ROOT=sentence_embeddings/$MODEL
+DICT=criss_checkpoints/dict.txt
+THRESHOLD=1.02
+MIN_COUNT=500
+
+DATA_DIR=data_tmp
+SAVE_DIR=mining/${source_lang}_${target_lang}_mined
+ENCODER_SAVE_DIR=${ENCODER_SAVE_ROOT}/${source_lang}-${target_lang}
+INPUT_DIR=$DATA_DIR/${source_lang}-${target_lang}-tatoeba
+
+mkdir -p $ENCODER_SAVE_DIR/${target_lang}
+mkdir -p $ENCODER_SAVE_DIR/${source_lang}
+mkdir -p $SAVE_DIR
+
+## Save encoder outputs
+
+# Save encoder outputs for source sentences
+python $SAVE_ENCODER \
+ ${INPUT_DIR} \
+ --path ${MODEL} \
+ --task translation_multi_simple_epoch \
+ --lang-pairs ${source_lang}-${target_lang} \
+ --lang-dict ${LANG_DICT} \
+ --gen-subset ${SPLIT} \
+ --bpe 'sentencepiece' \
+ -s ${source_lang} -t ${target_lang} \
+ --sentencepiece-model ${SPM} \
+ --remove-bpe 'sentencepiece' \
+ --beam 1 \
+ --lang-tok-style mbart \
+ --encoder-save-dir ${ENCODER_SAVE_DIR}/${source_lang}
+
+## Save encoder outputs for target sentences
+python $SAVE_ENCODER \
+ ${INPUT_DIR} \
+ --path ${MODEL} \
+ --lang-pairs ${source_lang}-${target_lang} \
+ --lang-dict ${LANG_DICT} \
+ --task translation_multi_simple_epoch \
+ --gen-subset ${SPLIT} \
+ --bpe 'sentencepiece' \
+ -t ${source_lang} -s ${target_lang} \
+ --sentencepiece-model ${SPM} \
+ --remove-bpe 'sentencepiece' \
+ --beam 1 \
+ --lang-tok-style mbart \
+ --encoder-save-dir ${ENCODER_SAVE_DIR}/${target_lang}
+
+## Mining
+python mining/mine.py \
+ --src-lang ${source_lang} \
+ --tgt-lang ${target_lang} \
+ --dim 1024 \
+ --mem 10 \
+ --neighborhood 4 \
+ --src-dir ${ENCODER_SAVE_DIR}/${source_lang} \
+ --tgt-dir ${ENCODER_SAVE_DIR}/${target_lang} \
+ --output $SAVE_DIR \
+ --threshold ${THRESHOLD} \
+ --min-count ${MIN_COUNT} \
+ --valid-size 100 \
+ --dict-path ${DICT} \
+ --spm-path ${SPM} \
+
+
+## Process and binarize mined data
+python $SPM_ENCODE \
+ --model ${SPM} \
+ --output_format=piece \
+ --inputs mining/${source_lang}_${target_lang}_mined/train.${source_lang} mining/${source_lang}_${target_lang}_mined/train.${target_lang} \
+ --outputs mining/${source_lang}_${target_lang}_mined/train.bpe.${source_lang} mining/${source_lang}_${target_lang}_mined/train.bpe.${target_lang}
+
+python $SPM_ENCODE \
+ --model ${SPM} \
+ --output_format=piece \
+ --inputs mining/${source_lang}_${target_lang}_mined/valid.${source_lang} mining/${source_lang}_${target_lang}_mined/valid.${target_lang} \
+ --outputs mining/${source_lang}_${target_lang}_mined/valid.bpe.${source_lang} mining/${source_lang}_${target_lang}_mined/valid.bpe.${target_lang}
+
+
+fairseq-preprocess \
+ --source-lang ${source_lang} \
+ --target-lang ${target_lang} \
+ --trainpref mining/${source_lang}_${target_lang}_mined/train.bpe \
+ --validpref mining/${source_lang}_${target_lang}_mined/valid.bpe \
+ --destdir mining/${source_lang}_${target_lang}_mined \
+ --srcdict ${DICT} \
+ --joined-dictionary \
+ --workers 8
diff --git a/fairseq/examples/criss/save_encoder.py b/fairseq/examples/criss/save_encoder.py
new file mode 100644
index 0000000000000000000000000000000000000000..24a842e4092663c79c92a299fa85747b7c0bed64
--- /dev/null
+++ b/fairseq/examples/criss/save_encoder.py
@@ -0,0 +1,214 @@
+#!/usr/bin/env python3 -u
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+"""
+Translate pre-processed data with a trained model.
+"""
+
+import numpy as np
+import torch
+from fairseq import checkpoint_utils, options, progress_bar, tasks, utils
+from fairseq.sequence_generator import EnsembleModel
+from fairseq.utils import safe_hasattr
+
+
+def get_avg_pool(
+ models, sample, prefix_tokens, src_dict, remove_bpe, has_langtok=False
+):
+ model = EnsembleModel(models)
+
+ # model.forward normally channels prev_output_tokens into the decoder
+ # separately, but SequenceGenerator directly calls model.encoder
+ encoder_input = {
+ k: v for k, v in sample["net_input"].items() if k != "prev_output_tokens"
+ }
+
+ # compute the encoder output for each beam
+ encoder_outs = model.forward_encoder(encoder_input)
+ np_encoder_outs = encoder_outs[0].encoder_out.cpu().numpy().astype(np.float32)
+ encoder_mask = 1 - encoder_outs[0].encoder_padding_mask.cpu().numpy().astype(
+ np.float32
+ )
+ encoder_mask = np.expand_dims(encoder_mask.T, axis=2)
+ if has_langtok:
+ encoder_mask = encoder_mask[1:, :, :]
+ np_encoder_outs = np_encoder_outs[1, :, :]
+ masked_encoder_outs = encoder_mask * np_encoder_outs
+ avg_pool = (masked_encoder_outs / encoder_mask.sum(axis=0)).sum(axis=0)
+ return avg_pool
+
+
+def main(args):
+ assert args.path is not None, "--path required for generation!"
+ assert (
+ not args.sampling or args.nbest == args.beam
+ ), "--sampling requires --nbest to be equal to --beam"
+ assert (
+ args.replace_unk is None or args.raw_text
+ ), "--replace-unk requires a raw text dataset (--raw-text)"
+
+ args.beam = 1
+ utils.import_user_module(args)
+
+ if args.max_tokens is None:
+ args.max_tokens = 12000
+ print(args)
+ use_cuda = torch.cuda.is_available() and not args.cpu
+
+ # Load dataset splits
+ task = tasks.setup_task(args)
+ task.load_dataset(args.gen_subset)
+
+ # Set dictionaries
+ try:
+ src_dict = getattr(task, "source_dictionary", None)
+ except NotImplementedError:
+ src_dict = None
+ tgt_dict = task.target_dictionary
+
+ # Load ensemble
+ print("| loading model(s) from {}".format(args.path))
+ models, _model_args = checkpoint_utils.load_model_ensemble(
+ args.path.split(":"),
+ arg_overrides=eval(args.model_overrides),
+ task=task,
+ )
+
+ # Optimize ensemble for generation
+ for model in models:
+ model.make_generation_fast_(
+ beamable_mm_beam_size=None if args.no_beamable_mm else args.beam,
+ need_attn=args.print_alignment,
+ )
+ if args.fp16:
+ model.half()
+ if use_cuda:
+ model.cuda()
+
+ # Load alignment dictionary for unknown word replacement
+ # (None if no unknown word replacement, empty if no path to align dictionary)
+ align_dict = utils.load_align_dict(args.replace_unk)
+
+ # Load dataset (possibly sharded)
+ itr = task.get_batch_iterator(
+ dataset=task.dataset(args.gen_subset),
+ max_tokens=args.max_tokens,
+ max_positions=utils.resolve_max_positions(
+ task.max_positions(),
+ ),
+ ignore_invalid_inputs=args.skip_invalid_size_inputs_valid_test,
+ required_batch_size_multiple=args.required_batch_size_multiple,
+ num_shards=args.num_shards,
+ shard_id=args.shard_id,
+ num_workers=args.num_workers,
+ ).next_epoch_itr(shuffle=False)
+
+ num_sentences = 0
+ source_sentences = []
+ shard_id = 0
+ all_avg_pool = None
+ encoder_has_langtok = (
+ safe_hasattr(task.args, "encoder_langtok")
+ and task.args.encoder_langtok is not None
+ and safe_hasattr(task.args, "lang_tok_replacing_bos_eos")
+ and not task.args.lang_tok_replacing_bos_eos
+ )
+ with progress_bar.build_progress_bar(args, itr) as t:
+ for sample in t:
+ if sample is None:
+ print("Skipping None")
+ continue
+ sample = utils.move_to_cuda(sample) if use_cuda else sample
+ if "net_input" not in sample:
+ continue
+
+ prefix_tokens = None
+ if args.prefix_size > 0:
+ prefix_tokens = sample["target"][:, : args.prefix_size]
+
+ with torch.no_grad():
+ avg_pool = get_avg_pool(
+ models,
+ sample,
+ prefix_tokens,
+ src_dict,
+ args.post_process,
+ has_langtok=encoder_has_langtok,
+ )
+ if all_avg_pool is not None:
+ all_avg_pool = np.concatenate((all_avg_pool, avg_pool))
+ else:
+ all_avg_pool = avg_pool
+
+ if not isinstance(sample["id"], list):
+ sample_ids = sample["id"].tolist()
+ else:
+ sample_ids = sample["id"]
+ for i, sample_id in enumerate(sample_ids):
+ # Remove padding
+ src_tokens = utils.strip_pad(
+ sample["net_input"]["src_tokens"][i, :], tgt_dict.pad()
+ )
+
+ # Either retrieve the original sentences or regenerate them from tokens.
+ if align_dict is not None:
+ src_str = task.dataset(args.gen_subset).src.get_original_text(
+ sample_id
+ )
+ else:
+ if src_dict is not None:
+ src_str = src_dict.string(src_tokens, args.post_process)
+ else:
+ src_str = ""
+
+ if not args.quiet:
+ if src_dict is not None:
+ print("S-{}\t{}".format(sample_id, src_str))
+
+ source_sentences.append(f"{sample_id}\t{src_str}")
+
+ num_sentences += sample["nsentences"]
+ if all_avg_pool.shape[0] >= 1000000:
+ with open(
+ f"{args.encoder_save_dir}/all_avg_pool.{args.source_lang}.{shard_id}",
+ "w",
+ ) as avg_pool_file:
+ all_avg_pool.tofile(avg_pool_file)
+ with open(
+ f"{args.encoder_save_dir}/sentences.{args.source_lang}.{shard_id}",
+ "w",
+ ) as sentence_file:
+ sentence_file.writelines(f"{line}\n" for line in source_sentences)
+ all_avg_pool = None
+ source_sentences = []
+ shard_id += 1
+
+ if all_avg_pool is not None:
+ with open(
+ f"{args.encoder_save_dir}/all_avg_pool.{args.source_lang}.{shard_id}", "w"
+ ) as avg_pool_file:
+ all_avg_pool.tofile(avg_pool_file)
+ with open(
+ f"{args.encoder_save_dir}/sentences.{args.source_lang}.{shard_id}", "w"
+ ) as sentence_file:
+ sentence_file.writelines(f"{line}\n" for line in source_sentences)
+ return None
+
+
+def cli_main():
+ parser = options.get_generation_parser()
+ parser.add_argument(
+ "--encoder-save-dir",
+ default="",
+ type=str,
+ metavar="N",
+ help="directory to save encoder outputs",
+ )
+ args = options.parse_args_and_arch(parser)
+ main(args)
+
+
+if __name__ == "__main__":
+ cli_main()
diff --git a/fairseq/examples/criss/sentence_retrieval/encoder_analysis.py b/fairseq/examples/criss/sentence_retrieval/encoder_analysis.py
new file mode 100644
index 0000000000000000000000000000000000000000..b41bfbe38789ba14e6a5ea938c75d761424c00ab
--- /dev/null
+++ b/fairseq/examples/criss/sentence_retrieval/encoder_analysis.py
@@ -0,0 +1,92 @@
+#!/usr/bin/env python3 -u
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+import argparse
+import glob
+
+import numpy as np
+
+
+DIM = 1024
+
+
+def compute_dist(source_embs, target_embs, k=5, return_sim_mat=False):
+ target_ids = [tid for tid in target_embs]
+ source_mat = np.stack(source_embs.values(), axis=0)
+ normalized_source_mat = source_mat / np.linalg.norm(
+ source_mat, axis=1, keepdims=True
+ )
+ target_mat = np.stack(target_embs.values(), axis=0)
+ normalized_target_mat = target_mat / np.linalg.norm(
+ target_mat, axis=1, keepdims=True
+ )
+ sim_mat = normalized_source_mat.dot(normalized_target_mat.T)
+ if return_sim_mat:
+ return sim_mat
+ neighbors_map = {}
+ for i, sentence_id in enumerate(source_embs):
+ idx = np.argsort(sim_mat[i, :])[::-1][:k]
+ neighbors_map[sentence_id] = [target_ids[tid] for tid in idx]
+ return neighbors_map
+
+
+def load_embeddings(directory, LANGS):
+ sentence_embeddings = {}
+ sentence_texts = {}
+ for lang in LANGS:
+ sentence_embeddings[lang] = {}
+ sentence_texts[lang] = {}
+ lang_dir = f"{directory}/{lang}"
+ embedding_files = glob.glob(f"{lang_dir}/all_avg_pool.{lang}.*")
+ for embed_file in embedding_files:
+ shard_id = embed_file.split(".")[-1]
+ embeddings = np.fromfile(embed_file, dtype=np.float32)
+ num_rows = embeddings.shape[0] // DIM
+ embeddings = embeddings.reshape((num_rows, DIM))
+
+ with open(f"{lang_dir}/sentences.{lang}.{shard_id}") as sentence_file:
+ for idx, line in enumerate(sentence_file):
+ sentence_id, sentence = line.strip().split("\t")
+ sentence_texts[lang][sentence_id] = sentence
+ sentence_embeddings[lang][sentence_id] = embeddings[idx, :]
+
+ return sentence_embeddings, sentence_texts
+
+
+def compute_accuracy(directory, LANGS):
+ sentence_embeddings, sentence_texts = load_embeddings(directory, LANGS)
+
+ top_1_accuracy = {}
+
+ top1_str = " ".join(LANGS) + "\n"
+ for source_lang in LANGS:
+ top_1_accuracy[source_lang] = {}
+ top1_str += f"{source_lang} "
+ for target_lang in LANGS:
+ top1 = 0
+ top5 = 0
+ neighbors_map = compute_dist(
+ sentence_embeddings[source_lang], sentence_embeddings[target_lang]
+ )
+ for sentence_id, neighbors in neighbors_map.items():
+ if sentence_id == neighbors[0]:
+ top1 += 1
+ if sentence_id in neighbors[:5]:
+ top5 += 1
+ n = len(sentence_embeddings[target_lang])
+ top1_str += f"{top1/n} "
+ top1_str += "\n"
+
+ print(top1_str)
+ print(top1_str, file=open(f"{directory}/accuracy", "w"))
+
+
+if __name__ == "__main__":
+ parser = argparse.ArgumentParser(description="Analyze encoder outputs")
+ parser.add_argument("directory", help="Source language corpus")
+ parser.add_argument("--langs", help="List of langs")
+ args = parser.parse_args()
+ langs = args.langs.split(",")
+ compute_accuracy(args.directory, langs)
diff --git a/fairseq/examples/criss/sentence_retrieval/sentence_retrieval_tatoeba.sh b/fairseq/examples/criss/sentence_retrieval/sentence_retrieval_tatoeba.sh
new file mode 100644
index 0000000000000000000000000000000000000000..0428d8bef9d426ac3e664cd281ce0b688f5f580f
--- /dev/null
+++ b/fairseq/examples/criss/sentence_retrieval/sentence_retrieval_tatoeba.sh
@@ -0,0 +1,59 @@
+#!/bin/bash
+# Copyright (c) Facebook, Inc. and its affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+#
+source_lang=kk_KZ
+target_lang=en_XX
+MODEL=criss_checkpoints/criss.3rd.pt
+SPM=criss_checkpoints/sentence.bpe.model
+SPLIT=test
+LANG_DICT=criss_checkpoints/lang_dict.txt
+ENCODER_ANALYSIS=sentence_retrieval/encoder_analysis.py
+SAVE_ENCODER=save_encoder.py
+ENCODER_SAVE_ROOT=sentence_embeddings/$MODEL
+
+
+
+DATA_DIR=data_tmp
+INPUT_DIR=$DATA_DIR/${source_lang}-${target_lang}-tatoeba
+ENCODER_SAVE_DIR=${ENCODER_SAVE_ROOT}/${source_lang}-${target_lang}
+mkdir -p $ENCODER_SAVE_DIR/${target_lang}
+mkdir -p $ENCODER_SAVE_DIR/${source_lang}
+
+# Save encoder outputs for source sentences
+python $SAVE_ENCODER \
+ ${INPUT_DIR} \
+ --path ${MODEL} \
+ --task translation_multi_simple_epoch \
+ --lang-dict ${LANG_DICT} \
+ --gen-subset ${SPLIT} \
+ --bpe 'sentencepiece' \
+ --lang-pairs ${source_lang}-${target_lang} \
+ -s ${source_lang} -t ${target_lang} \
+ --sentencepiece-model ${SPM} \
+ --remove-bpe 'sentencepiece' \
+ --beam 1 \
+ --lang-tok-style mbart \
+ --encoder-save-dir ${ENCODER_SAVE_DIR}/${source_lang}
+
+# Save encoder outputs for target sentences
+python $SAVE_ENCODER \
+ ${INPUT_DIR} \
+ --path ${MODEL} \
+ --lang-dict ${LANG_DICT} \
+ --task translation_multi_simple_epoch \
+ --gen-subset ${SPLIT} \
+ --bpe 'sentencepiece' \
+ --lang-pairs ${target_lang}-${source_lang} \
+ -t ${source_lang} -s ${target_lang} \
+ --sentencepiece-model ${SPM} \
+ --remove-bpe 'sentencepiece' \
+ --beam 1 \
+ --lang-tok-style mbart \
+ --encoder-save-dir ${ENCODER_SAVE_DIR}/${target_lang}
+
+# Analyze sentence retrieval accuracy
+python $ENCODER_ANALYSIS --langs "${source_lang},${target_lang}" ${ENCODER_SAVE_DIR}
diff --git a/fairseq/examples/criss/unsupervised_mt/eval.sh b/fairseq/examples/criss/unsupervised_mt/eval.sh
new file mode 100644
index 0000000000000000000000000000000000000000..03b773ed5a522eb82186fea8ffbb6c557e14b6d3
--- /dev/null
+++ b/fairseq/examples/criss/unsupervised_mt/eval.sh
@@ -0,0 +1,37 @@
+#!/bin/bash
+# Copyright (c) Facebook, Inc. and its affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+#
+SRC=si_LK
+TGT=en_XX
+MODEL=criss_checkpoints/criss.3rd.pt
+
+MULTIBLEU=mosesdecoder/scripts/generic/multi-bleu.perl
+MOSES=mosesdecoder
+REPLACE_UNICODE_PUNCT=$MOSES/scripts/tokenizer/replace-unicode-punctuation.perl
+NORM_PUNC=$MOSES/scripts/tokenizer/normalize-punctuation.perl
+REM_NON_PRINT_CHAR=$MOSES/scripts/tokenizer/remove-non-printing-char.perl
+TOKENIZER=$MOSES/scripts/tokenizer/tokenizer.perl
+GEN_TMP_DIR=gen_tmp
+LANG_DICT=criss_checkpoints/lang_dict.txt
+
+if [ ! -d "mosesdecoder" ]; then
+ git clone https://github.com/moses-smt/mosesdecoder
+fi
+mkdir -p $GEN_TMP_DIR
+fairseq-generate data_tmp/${SRC}-${TGT}-flores \
+ --task translation_multi_simple_epoch \
+ --max-tokens 2000 \
+ --path ${MODEL} \
+ --skip-invalid-size-inputs-valid-test \
+ --beam 5 --lenpen 1.0 --gen-subset test \
+ --remove-bpe=sentencepiece \
+ --source-lang ${SRC} --target-lang ${TGT} \
+ --decoder-langtok --lang-pairs 'en_XX-ar_AR,en_XX-de_DE,en_XX-es_XX,en_XX-fr_XX,en_XX-hi_IN,en_XX-it_IT,en_XX-ja_XX,en_XX-ko_KR,en_XX-nl_XX,en_XX-ru_RU,en_XX-zh_CN,en_XX-tr_TR,en_XX-vi_VN,en_XX-ro_RO,en_XX-my_MM,en_XX-ne_NP,en_XX-si_LK,en_XX-cs_CZ,en_XX-lt_LT,en_XX-kk_KZ,en_XX-gu_IN,en_XX-fi_FI,en_XX-et_EE,en_XX-lv_LV,ar_AR-en_XX,cs_CZ-en_XX,de_DE-en_XX,es_XX-en_XX,et_EE-en_XX,fi_FI-en_XX,fr_XX-en_XX,gu_IN-en_XX,hi_IN-en_XX,it_IT-en_XX,ja_XX-en_XX,kk_KZ-en_XX,ko_KR-en_XX,lt_LT-en_XX,lv_LV-en_XX,my_MM-en_XX,ne_NP-en_XX,nl_XX-en_XX,ro_RO-en_XX,ru_RU-en_XX,si_LK-en_XX,tr_TR-en_XX,vi_VN-en_XX,zh_CN-en_XX,ar_AR-es_XX,es_XX-ar_AR,ar_AR-hi_IN,hi_IN-ar_AR,ar_AR-zh_CN,zh_CN-ar_AR,cs_CZ-es_XX,es_XX-cs_CZ,cs_CZ-hi_IN,hi_IN-cs_CZ,cs_CZ-zh_CN,zh_CN-cs_CZ,de_DE-es_XX,es_XX-de_DE,de_DE-hi_IN,hi_IN-de_DE,de_DE-zh_CN,zh_CN-de_DE,es_XX-hi_IN,hi_IN-es_XX,es_XX-zh_CN,zh_CN-es_XX,et_EE-es_XX,es_XX-et_EE,et_EE-hi_IN,hi_IN-et_EE,et_EE-zh_CN,zh_CN-et_EE,fi_FI-es_XX,es_XX-fi_FI,fi_FI-hi_IN,hi_IN-fi_FI,fi_FI-zh_CN,zh_CN-fi_FI,fr_XX-es_XX,es_XX-fr_XX,fr_XX-hi_IN,hi_IN-fr_XX,fr_XX-zh_CN,zh_CN-fr_XX,gu_IN-es_XX,es_XX-gu_IN,gu_IN-hi_IN,hi_IN-gu_IN,gu_IN-zh_CN,zh_CN-gu_IN,hi_IN-zh_CN,zh_CN-hi_IN,it_IT-es_XX,es_XX-it_IT,it_IT-hi_IN,hi_IN-it_IT,it_IT-zh_CN,zh_CN-it_IT,ja_XX-es_XX,es_XX-ja_XX,ja_XX-hi_IN,hi_IN-ja_XX,ja_XX-zh_CN,zh_CN-ja_XX,kk_KZ-es_XX,es_XX-kk_KZ,kk_KZ-hi_IN,hi_IN-kk_KZ,kk_KZ-zh_CN,zh_CN-kk_KZ,ko_KR-es_XX,es_XX-ko_KR,ko_KR-hi_IN,hi_IN-ko_KR,ko_KR-zh_CN,zh_CN-ko_KR,lt_LT-es_XX,es_XX-lt_LT,lt_LT-hi_IN,hi_IN-lt_LT,lt_LT-zh_CN,zh_CN-lt_LT,lv_LV-es_XX,es_XX-lv_LV,lv_LV-hi_IN,hi_IN-lv_LV,lv_LV-zh_CN,zh_CN-lv_LV,my_MM-es_XX,es_XX-my_MM,my_MM-hi_IN,hi_IN-my_MM,my_MM-zh_CN,zh_CN-my_MM,ne_NP-es_XX,es_XX-ne_NP,ne_NP-hi_IN,hi_IN-ne_NP,ne_NP-zh_CN,zh_CN-ne_NP,nl_XX-es_XX,es_XX-nl_XX,nl_XX-hi_IN,hi_IN-nl_XX,nl_XX-zh_CN,zh_CN-nl_XX,ro_RO-es_XX,es_XX-ro_RO,ro_RO-hi_IN,hi_IN-ro_RO,ro_RO-zh_CN,zh_CN-ro_RO,ru_RU-es_XX,es_XX-ru_RU,ru_RU-hi_IN,hi_IN-ru_RU,ru_RU-zh_CN,zh_CN-ru_RU,si_LK-es_XX,es_XX-si_LK,si_LK-hi_IN,hi_IN-si_LK,si_LK-zh_CN,zh_CN-si_LK,tr_TR-es_XX,es_XX-tr_TR,tr_TR-hi_IN,hi_IN-tr_TR,tr_TR-zh_CN,zh_CN-tr_TR,vi_VN-es_XX,es_XX-vi_VN,vi_VN-hi_IN,hi_IN-vi_VN,vi_VN-zh_CN,zh_CN-vi_VN' \
+ --lang-dict ${LANG_DICT} --lang-tok-style 'mbart' --sampling-method 'temperature' --sampling-temperature '1.0' > $GEN_TMP_DIR/${SRC}_${TGT}.gen
+cat $GEN_TMP_DIR/${SRC}_${TGT}.gen | grep -P "^T-" | cut -f2 | $REPLACE_UNICODE_PUNCT | $NORM_PUNC -l ${TGT:0:2} | $REM_NON_PRINT_CHAR | $TOKENIZER -no-escape ${TGT:0:2} > $GEN_TMP_DIR/${SRC}_${TGT}.hyp
+cat $GEN_TMP_DIR/${SRC}_${TGT}.gen | grep -P "^H-" | cut -f3 | $REPLACE_UNICODE_PUNCT | $NORM_PUNC -l ${TGT:0:2} | $REM_NON_PRINT_CHAR | $TOKENIZER -no-escape ${TGT:0:2} > $GEN_TMP_DIR/${SRC}_${TGT}.ref
+${MULTIBLEU} $GEN_TMP_DIR/${SRC}_${TGT}.ref < $GEN_TMP_DIR/${SRC}_${TGT}.hyp
diff --git a/fairseq/examples/cross_lingual_language_model/README.md b/fairseq/examples/cross_lingual_language_model/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..af9128e39e5925e9411d162c2f24a19e4532d618
--- /dev/null
+++ b/fairseq/examples/cross_lingual_language_model/README.md
@@ -0,0 +1,77 @@
+# Cross-Lingual Language Model Pre-training
+
+Below are some details for training Cross-Lingual Language Models (XLM) - similar to the ones presented in [Lample & Conneau, 2019](https://arxiv.org/pdf/1901.07291.pdf) - in Fairseq. The current implementation only supports the Masked Language Model (MLM) from the paper above.
+
+## Downloading and Tokenizing Monolingual Data
+
+Pointers to the monolingual data from wikipedia, used for training the XLM-style MLM model as well as details on processing (tokenization and BPE) it can be found in the [XLM Github Repository](https://github.com/facebookresearch/XLM#download--preprocess-monolingual-data).
+
+Let's assume the following for the code snippets in later sections to work
+- Processed data is in the folder: monolingual_data/processed
+- Each language has 3 files for train, test and validation. For example we have the following files for English:
+ train.en, valid.en
+- We are training a model for 5 languages: Arabic (ar), German (de), English (en), Hindi (hi) and French (fr)
+- The vocabulary file is monolingual_data/processed/vocab_mlm
+
+
+## Fairseq Pre-processing and Binarization
+
+Pre-process and binarize the data with the MaskedLMDictionary and cross_lingual_lm task
+
+```bash
+# Ensure the output directory exists
+DATA_DIR=monolingual_data/fairseq_processed
+mkdir -p "$DATA_DIR"
+
+for lg in ar de en hi fr
+do
+
+ fairseq-preprocess \
+ --task cross_lingual_lm \
+ --srcdict monolingual_data/processed/vocab_mlm \
+ --only-source \
+ --trainpref monolingual_data/processed/train \
+ --validpref monolingual_data/processed/valid \
+ --testpref monolingual_data/processed/test \
+ --destdir monolingual_data/fairseq_processed \
+ --workers 20 \
+ --source-lang $lg
+
+ # Since we only have a source language, the output file has a None for the
+ # target language. Remove this
+
+ for stage in train test valid
+
+ sudo mv "$DATA_DIR/$stage.$lg-None.$lg.bin" "$stage.$lg.bin"
+ sudo mv "$DATA_DIR/$stage.$lg-None.$lg.idx" "$stage.$lg.idx"
+
+ done
+
+done
+```
+
+## Train a Cross-lingual Language Model similar to the XLM MLM model
+
+Use the following command to train the model on 5 languages.
+
+```
+fairseq-train \
+--task cross_lingual_lm monolingual_data/fairseq_processed \
+--save-dir checkpoints/mlm \
+--max-update 2400000 --save-interval 1 --no-epoch-checkpoints \
+--arch xlm_base \
+--optimizer adam --lr-scheduler reduce_lr_on_plateau \
+--lr-shrink 0.5 --lr 0.0001 --stop-min-lr 1e-09 \
+--dropout 0.1 \
+--criterion legacy_masked_lm_loss \
+--max-tokens 2048 --tokens-per-sample 256 --attention-dropout 0.1 \
+--dataset-impl lazy --seed 0 \
+--masked-lm-only \
+--monolingual-langs 'ar,de,en,hi,fr' --num-segment 5 \
+--ddp-backend=legacy_ddp
+```
+
+Some Notes:
+- Using tokens_per_sample greater than 256 can cause OOM (out-of-memory) issues. Usually since MLM packs in streams of text, this parameter doesn't need much tuning.
+- The Evaluation workflow for computing MLM Perplexity on test data is in progress.
+- Finetuning this model on a downstream task is something which is not currently available.
diff --git a/fairseq/examples/data2vec/README.md b/fairseq/examples/data2vec/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..a0ff21b82ac420259f0abdc1d80479d7a6edd47c
--- /dev/null
+++ b/fairseq/examples/data2vec/README.md
@@ -0,0 +1,261 @@
+# data2vec 2.0
+
+data2vec 2.0 improves the training efficiency of the original data2vec algorithm. We make the following improvements for efficiency considerations - we forward only the unmasked timesteps through the encoder, we use convolutional decoder and we use multimasking to amortize the compute overhead of the teacher model. You can find details in the paper [Efficient Self-supervised Learning with Contextualized Target Representations for Vision, Speech and Language](https://arxiv.org/abs/2212.07525) and our [blog post](https://ai.facebook.com/blog/ai-self-supervised-learning-data2vec/).
+
+## Pretrained and finetuned models
+### Vision
+| Model | Finetuning split | Link
+|---|---|---
+data2vec ViT-B | No fine-tuning | [download](https://dl.fbaipublicfiles.com/fairseq/data2vec2/base_imagenet.pt)
+data2vec ViT-B | Imagenet-1K | [download](https://dl.fbaipublicfiles.com/fairseq/data2vec2/base_imagenet_ft.pt)
+data2vec ViT-L | No fine-tuning | [download](https://dl.fbaipublicfiles.com/fairseq/data2vec2/large_imagenet.pt)
+data2vec ViT-L | Imagenet-1K | [download](https://dl.fbaipublicfiles.com/fairseq/data2vec2/large_imagenet_ft.pt)
+data2vec ViT-H | No fine-tuning | [download](https://dl.fbaipublicfiles.com/fairseq/data2vec2/huge_imagenet.pt)
+data2vec ViT-H | Imagenet-1K | [download](https://dl.fbaipublicfiles.com/fairseq/data2vec2/huge_imagenet_ft.pt)
+
+Vision models only are license under CC-BY-NC.
+### Speech
+
+| Model | Finetuning split | Dataset | Link
+|---|---|---|---
+data2vec Base | No fine-tuning | [Librispeech](http://www.openslr.org/12) | [download](https://dl.fbaipublicfiles.com/fairseq/data2vec2/base_libri.pt)
+data2vec Base | 960 hours | [Librispeech](http://www.openslr.org/12) | [download](https://dl.fbaipublicfiles.com/fairseq/data2vec2/base_libri_960h.pt)
+data2vec Large | No fine-tuning | [Libri-light](https://github.com/facebookresearch/libri-light) | [download](https://dl.fbaipublicfiles.com/fairseq/data2vec2/large_vox.pt)
+data2vec Large | 960 hours | [Libri-light](https://github.com/facebookresearch/libri-light) | [download](https://dl.fbaipublicfiles.com/fairseq/data2vec2/large_vox_960h.pt)
+
+### NLP
+
+| Model | Fine-tuning data | Dataset | Link | Dict | BPE
+|---|---|---|---|---|---
+data2vec Base | No fine-tuning | Books + Wiki | [download](https://dl.fbaipublicfiles.com/fairseq/data2vec2/nlp_base.pt) | [dict](https://dl.fbaipublicfiles.com/fairseq/data2vec2/dict.txt) | [encoder](https://dl.fbaipublicfiles.com/fairseq/data2vec2/encoder.json) / [vocab](https://dl.fbaipublicfiles.com/fairseq/data2vec2/vocab.bpe)
+
+[//]: # (## Data Preparation)
+
+[//]: # ()
+[//]: # (### Vision)
+
+[//]: # (add details)
+
+[//]: # (### Speech)
+
+[//]: # (add details)
+
+[//]: # ()
+[//]: # (### NLP)
+
+[//]: # (add details)
+
+
+## Commands to train different models using data2vec 2.0
+
+### Vision
+
+Commands to pretrain different model configurations
+```shell script
+$ python fairseq_cli/hydra_train.py -m --config-dir examples/data2vec/config/v2 \
+--config-name base_images_only_task task.data=/path/to/dir
+```
+
+```shell script
+$ python fairseq_cli/hydra_train.py -m --config-dir examples/data2vec/config/v2 \
+--config-name large_images_only_task task.data=/path/to/dir
+```
+
+```shell script
+$ python fairseq_cli/hydra_train.py -m --config-dir examples/data2vec/config/v2 \
+--config-name huge_images14_only_task task.data=/path/to/dir
+```
+
+Commands to finetune different model configurations
+
+```shell script
+$ python fairseq_cli/hydra_train.py -m --config-dir examples/data2vec/config/vision/finetuning \
+--config-name mae_imagenet_clean task.data=/path/to/dir model.model_path=/path/to/pretrained/model
+```
+
+```shell script
+$ python fairseq_cli/hydra_train.py -m --config-dir examples/data2vec/config/vision/finetuning \
+--config-name mae_imagenet_large_clean task.data=/path/to/dir model.model_path=/path/to/pretrained/model
+```
+
+```shell script
+$ python fairseq_cli/hydra_train.py -m --config-dir examples/data2vec/config/vision/finetuning \
+--config-name mae_imagenet_huge_clean task.data=/path/to/dir model.model_path=/path/to/pretrained/model
+```
+
+### Speech
+
+```shell script
+$ python fairseq_cli/hydra_train.py -m --config-dir examples/data2vec/config/v2 \
+--config-name base_audio_only_task task.data=/path/to/manifests
+```
+
+```shell script
+$ python fairseq_cli/hydra_train.py -m --config-dir examples/data2vec/config/v2 \
+--config-name large_audio_only_task task.data=/path/to/manifests
+```
+
+Finetuning:
+
+```shell script
+$ python fairseq_cli/hydra_train.py -m --config-dir examples/wav2vec/config/finetuning --config-name vox_10h \
+task.data=/path/to/manifests model.w2v_path=/path/to/pretrained/model common.user_dir=examples/data2vec
+```
+
+Replace vox_10h with the right config depending on your model and fine-tuning split.
+See examples/wav2vec/config/finetuning for all available configs.
+
+### NLP
+
+Commands to pretrain
+```shell script
+$ python fairseq_cli/hydra_train.py -m --config-dir examples/data2vec/config/v2 \
+--config-name base_text_only_task task.data=/path/to/file
+```
+
+Commands to fine-tune all GLUE tasks
+```shell script
+$ task=cola # choose from [cola|qnli|mrpc|rte|sst_2|mnli|qqp|sts_b]
+$ lr=1e-5 # sweep [1e-5|2e-5|4e-5|6e-5] for each task
+$ python fairseq_cli/hydra_train.py -m --config-dir examples/data2vec/config/v2/text_finetuning \
+--config-name $task task.data=/path/to/file model.model_path=/path/to/pretrained/model "optimization.lr=[${lr}]"
+```
+
+# data2vec
+
+data2vec is a framework for self-supervised representation learning for images, speech, and text as described in [data2vec: A General Framework for Self-supervised Learning in Speech, Vision and Language (Baevski et al., 2022)](https://ai.facebook.com/research/data2vec-a-general-framework-for-self-supervised-learning-in-speech-vision-and-language). The algorithm uses the same learning mechanism for different modalities.
+
+
+## Pre-trained models
+
+### Vision
+
+Code and pre-trained models for data2vec visions can be found [here](https://github.com/facebookresearch/data2vec_vision/tree/main/beit).
+
+### Speech
+
+| Model | Finetuning split | Dataset | Link
+|---|---|---|---
+data2vec Base | No fine-tuning | [Librispeech](http://www.openslr.org/12) | [download](https://dl.fbaipublicfiles.com/fairseq/data2vec/audio_base_ls.pt)
+data2vec Base | 10 minutes | [Librispeech](http://www.openslr.org/12) | [download](https://dl.fbaipublicfiles.com/fairseq/data2vec/audio_base_ls_10m.pt)
+data2vec Base | 100 hours | [Librispeech](http://www.openslr.org/12) | [download](https://dl.fbaipublicfiles.com/fairseq/data2vec/audio_base_ls_100h.pt)
+data2vec Base | 960 hours | [Librispeech](http://www.openslr.org/12) | [download](https://dl.fbaipublicfiles.com/fairseq/data2vec/audio_base_ls_960h.pt)
+data2vec Large | No fine-tuning | [Libri-light](https://github.com/facebookresearch/libri-light) | [download](https://dl.fbaipublicfiles.com/fairseq/data2vec/vox_pretrained.pt)
+data2vec Large | 10 minutes | [Libri-light](https://github.com/facebookresearch/libri-light) | [download](https://dl.fbaipublicfiles.com/fairseq/data2vec/vox_10m.pt)
+data2vec Large | 100 hours | [Libri-light](https://github.com/facebookresearch/libri-light) | [download](https://dl.fbaipublicfiles.com/fairseq/data2vec/vox_100h.pt)
+data2vec Large | 960 hours | [Libri-light](https://github.com/facebookresearch/libri-light) | [download](https://dl.fbaipublicfiles.com/fairseq/data2vec/vox_960h.pt)
+---
+
+### NLP
+
+Model | Fine-tuning data | Dataset | Link
+|---|---|---|---|
+data2vec Base | No fine-tuning | Books + Wiki | [download](https://dl.fbaipublicfiles.com/fairseq/data2vec/nlp_base.pt)
+
+## Training a new speech model with the CLI tools
+
+Given a directory containing wav files to be used for pretraining (we recommend splitting each file into separate file 10 to 30 seconds in length)
+
+### Prepare training data manifest:
+
+First, install the `soundfile` library:
+```shell script
+pip install soundfile
+```
+
+Next, run:
+
+```shell script
+$ python examples/wav2vec/wav2vec_manifest.py /path/to/waves --dest /manifest/path --ext $ext --valid-percent $valid
+```
+
+$ext should be set to flac, wav, or whatever format your dataset happens to use that soundfile can read.
+
+$valid should be set to some reasonable percentage (like 0.01) of training data to use for validation.
+To use a pre-defined validation set (like dev-other from librispeech), set to it 0 and then overwrite valid.tsv with a
+separately pre-processed manifest file.
+
+### Train a data2vec Base model:
+
+This configuration was used for the base model trained on the Librispeech dataset in the data2vec paper
+
+Note that the input is expected to be single channel, sampled at 16 kHz
+
+```shell script
+$ python fairseq_cli/hydra_train.py -m --config-dir examples/data2vec/config/audio/pretraining \
+--config-name base_librispeech task.data=/path/to/manifests common.user_dir=examples/data2vec
+```
+
+Note: you can simulate 16 GPUs by using k GPUs and adding command line parameters
+`distributed_training.distributed_world_size=k` `+optimization.update_freq='[x]'` where x = 16/k
+
+### Fine-tune a pre-trained model with CTC:
+
+Fine-tuning a model requires parallel audio and labels file, as well as a vocabulary file in fairseq format.
+A letter vocabulary can be downloaded [here](https://dl.fbaipublicfiles.com/fairseq/wav2vec/dict.ltr.txt).
+An example [script](../wav2vec/libri_labels.py) that generates labels for the Librispeech dataset from the tsv file produced by wav2vec_manifest.py can be used as follows:
+
+```shell script
+split=train
+$ python libri_labels.py /path/to/tsv --output-dir /output/dir --output-name $split
+```
+
+Fine-tuning on 100h of Librispeech with letter targets:
+```shell script
+$ fairseq-hydra-train \
+ distributed_training.distributed_port=$PORT \
+ task.data=/path/to/data \
+ model.w2v_path=/path/to/model.pt \
+ --config-dir /path/to/fairseq-py/examples/wav2vec/config/finetuning \
+ --config-name base_100h common.user_dir=examples/data2vec
+```
+
+There are other config files in the config/finetuning directory that can be used to fine-tune on other splits.
+You can specify the right config via the `--config-name` parameter.
+
+Decoding with a language model during training requires flashlight [python bindings](https://github.com/facebookresearch/flashlight/tree/master/bindings/python) (previously called [wav2letter](https://github.com/facebookresearch/wav2letter).
+If you want to use a language model, add `+criterion.wer_args='[/path/to/kenlm, /path/to/lexicon, 2, -1]'` to the command line.
+
+### Evaluating a CTC model:
+
+Evaluating a CTC model with a language model requires [flashlight python bindings](https://github.com/facebookresearch/flashlight/tree/master/bindings/python) (previously called [wav2letter](https://github.com/facebookresearch/wav2letter) to be installed.
+
+Fairseq transformer language model used in the wav2vec 2.0 paper can be obtained from the [wav2letter model repository](https://github.com/facebookresearch/wav2letter/tree/master/recipes/sota/2019).
+Be sure to upper-case the language model vocab after downloading it.
+
+Letter dictionary for pre-trained models can be found [here](https://dl.fbaipublicfiles.com/fairseq/wav2vec/dict.ltr.txt).
+
+Next, run the evaluation command:
+
+```shell script
+python examples/speech_recognition/new/infer.py --config-dir examples/speech_recognition/new/conf \
+--config-name infer task=audio_finetuning task.data=/path/to/manifests common.user_dir=examples/data2vec \
+task.labels=ltr decoding.type=kenlm \
+decoding.lmweight=${lmweight} decoding.wordscore=${wordscore} decoding.silweight=${silscore} \
+decoding.lexicon=/path/to/lexicon \
+decoding.lmpath=/path/to/lm decoding.unique_wer_file=True \
+dataset.gen_subset=dev_clean,dev_other,test_clean,test_other \
+common_eval.path=/path/to/checkpoint.pt decoding.beam=1500 distributed_training.distributed_world_size=${num_gpus}
+```
+
+To get raw numbers, use decoding.type=viterbi and omit the lexicon. To use the transformer language model, use decoding.type=fairseqlm.
+
+## Training a new NLP model with the CLI tools
+
+Please follow the [RoBERTa](../roberta/README.md) instructions to preprocess your data. To train a data2vec model on run:
+
+```shell script
+$ python fairseq_cli/hydra_train.py -m --config-dir examples/data2vec/config/text/pretraining \
+--config-name base task.data=/path/to/data common.user_dir=examples/data2vec
+```
+
+As for speech models, you can simulate 16 gpus by using the update_freq parameter.
+
+### Finetuning data2vec-text on GLUE
+
+Please use a command similar to this:
+
+```shell
+$ python fairseq_cli/hydra_train.py -m --config-dir examples/roberta/config/finetuning \
+ --config-name $task task.data=$data_path checkpoint.restore_file="${/path/to/pretrained/model.pt}"
+```
diff --git a/fairseq/examples/data2vec/__init__.py b/fairseq/examples/data2vec/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/fairseq/examples/data2vec/config/audio/classification/base_classification.yaml b/fairseq/examples/data2vec/config/audio/classification/base_classification.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fdb9c8d3d7c11b15c50d9a661603438f3ad0ffec
--- /dev/null
+++ b/fairseq/examples/data2vec/config/audio/classification/base_classification.yaml
@@ -0,0 +1,70 @@
+# @package _group_
+
+common:
+ fp16: true
+ log_format: json
+ log_interval: 200
+ all_gather_list_size: 70000
+ tensorboard_logdir: tb
+ min_loss_scale: 1e-6
+
+checkpoint:
+ save_interval: 1
+ no_epoch_checkpoints: true
+ best_checkpoint_metric: mAP
+ maximize_best_checkpoint_metric: true
+
+task:
+ _name: audio_classification
+ data: ???
+ normalize: true
+ labels: lbl
+
+dataset:
+ num_workers: 6
+ max_tokens: 2560000
+ skip_invalid_size_inputs_valid_test: true
+ valid_subset: eval
+ validate_interval: 5
+
+distributed_training:
+ ddp_backend: legacy_ddp
+ distributed_world_size: 8
+
+criterion:
+ _name: model
+ can_sum: false
+ log_keys:
+ - _predictions
+ - _targets
+
+optimization:
+ max_update: 30000
+ lr: [0.00006] # scratch 53-5
+
+optimizer:
+ _name: adam
+ adam_betas: (0.9,0.98)
+ adam_eps: 1e-08
+
+lr_scheduler:
+ _name: cosine
+ warmup_updates: 5000
+
+model:
+ _name: audio_classification
+ model_path: ???
+ apply_mask: true
+ mask_prob: 0.6
+ mask_length: 5 # scratch 1
+ mask_channel_prob: 0
+ mask_channel_length: 64
+ layerdrop: 0.1
+ dropout: 0.1
+ activation_dropout: 0.1
+ attention_dropout: 0.2
+ feature_grad_mult: 0 # scratch 1
+ label_mixup: true
+ source_mixup: 0.5
+ prediction_mode: lin_softmax # scratch average_sigmoid
+
diff --git a/fairseq/examples/data2vec/config/audio/classification/run_config/slurm_1.yaml b/fairseq/examples/data2vec/config/audio/classification/run_config/slurm_1.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..881a1583f871d272b43c6a2d13ef6768f0923bb2
--- /dev/null
+++ b/fairseq/examples/data2vec/config/audio/classification/run_config/slurm_1.yaml
@@ -0,0 +1,35 @@
+# @package _global_
+
+hydra:
+ job:
+ config:
+ override_dirname:
+ kv_sep: ':'
+ item_sep: '/'
+ exclude_keys:
+ - run_config
+ - distributed_training.distributed_port
+ - distributed_training.distributed_world_size
+ - model.pretrained_model_path
+ - model.target_network_path
+ - next_script
+ - task.cache_in_scratch
+ - task.data
+ - checkpoint.save_interval_updates
+ - checkpoint.keep_interval_updates
+ - checkpoint.save_on_overflow
+ sweep:
+ dir: /checkpoint/${env:USER}/${env:PREFIX}/${hydra.job.config_name}_${hydra.launcher.gpus_per_node}/${hydra.job.override_dirname}
+ subdir: ''
+ launcher:
+ submitit_folder: ${hydra.sweep.dir}
+ timeout_min: 4320
+ cpus_per_task: 10
+ gpus_per_node: 8
+ tasks_per_node: 8
+ mem_gb: 450
+ nodes: 1
+ name: ${env:PREFIX}_${hydra.job.config_name}
+ partition: devlab,learnlab,learnfair,scavenge
+ constraint: volta32gb,ib4
+ max_num_timeout: 30
diff --git a/fairseq/examples/data2vec/config/audio/classification/run_config/slurm_1g.yaml b/fairseq/examples/data2vec/config/audio/classification/run_config/slurm_1g.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..de7894d9cf0ac85b8d5a4282c1c9612f14f2d2b3
--- /dev/null
+++ b/fairseq/examples/data2vec/config/audio/classification/run_config/slurm_1g.yaml
@@ -0,0 +1,35 @@
+# @package _global_
+
+hydra:
+ job:
+ config:
+ override_dirname:
+ kv_sep: ':'
+ item_sep: '/'
+ exclude_keys:
+ - run_config
+ - distributed_training.distributed_port
+ - distributed_training.distributed_world_size
+ - model.pretrained_model_path
+ - model.target_network_path
+ - next_script
+ - task.cache_in_scratch
+ - task.data
+ - checkpoint.save_interval_updates
+ - checkpoint.keep_interval_updates
+ - checkpoint.save_on_overflow
+ sweep:
+ dir: /checkpoint/${env:USER}/${env:PREFIX}/${hydra.job.config_name}_${hydra.launcher.gpus_per_node}/${hydra.job.override_dirname}
+ subdir: ''
+ launcher:
+ submitit_folder: ${hydra.sweep.dir}
+ timeout_min: 4320
+ cpus_per_task: 10
+ gpus_per_node: 1
+ tasks_per_node: 1
+ mem_gb: 100
+ nodes: 1
+ name: ${env:PREFIX}_${hydra.job.config_name}
+ partition: devlab,learnlab,learnfair,scavenge
+ constraint: volta32gb
+ max_num_timeout: 30
diff --git a/fairseq/examples/data2vec/config/audio/classification/run_config/slurm_2.yaml b/fairseq/examples/data2vec/config/audio/classification/run_config/slurm_2.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..b016cac9b5e4a5f978d83124697a9c01ec885164
--- /dev/null
+++ b/fairseq/examples/data2vec/config/audio/classification/run_config/slurm_2.yaml
@@ -0,0 +1,35 @@
+# @package _global_
+
+hydra:
+ job:
+ config:
+ override_dirname:
+ kv_sep: ':'
+ item_sep: '/'
+ exclude_keys:
+ - run_config
+ - distributed_training.distributed_port
+ - distributed_training.distributed_world_size
+ - model.pretrained_model_path
+ - model.target_network_path
+ - next_script
+ - task.cache_in_scratch
+ - task.data
+ - checkpoint.save_interval_updates
+ - checkpoint.keep_interval_updates
+ - checkpoint.save_on_overflow
+ sweep:
+ dir: /checkpoint/${env:USER}/${env:PREFIX}/${hydra.job.config_name}_${hydra.launcher.gpus_per_node}/${hydra.job.override_dirname}
+ subdir: ''
+ launcher:
+ submitit_folder: ${hydra.sweep.dir}
+ timeout_min: 4320
+ cpus_per_task: 10
+ gpus_per_node: 8
+ tasks_per_node: 8
+ mem_gb: 450
+ nodes: 2
+ name: ${env:PREFIX}_${hydra.job.config_name}
+ partition: devlab,learnlab,learnfair,scavenge
+ constraint: volta32gb,ib4
+ max_num_timeout: 30
diff --git a/fairseq/examples/data2vec/config/audio/pretraining/audioset.yaml b/fairseq/examples/data2vec/config/audio/pretraining/audioset.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..dd30fbedd5818debf49ef1c87c84af4b6c58fb2d
--- /dev/null
+++ b/fairseq/examples/data2vec/config/audio/pretraining/audioset.yaml
@@ -0,0 +1,91 @@
+# @package _group_
+
+common:
+ fp16: true
+ log_format: json
+ log_interval: 200
+ tensorboard_logdir: tb
+ min_loss_scale: 1e-6
+ user_dir: /private/home/abaevski/fairseq-py/examples/data2vec
+
+checkpoint:
+ save_interval: 1
+ save_interval_updates: 25000
+ keep_interval_updates: 1
+ no_epoch_checkpoints: true
+
+task:
+ _name: audio_pretraining
+ data: /private/home/abaevski/data/audioset
+ max_sample_size: 320000
+ min_sample_size: 32000
+ normalize: true
+
+dataset:
+ num_workers: 6
+ max_tokens: 3400000
+ skip_invalid_size_inputs_valid_test: true
+ validate_interval: 5
+ required_batch_size_multiple: 1
+ disable_validation: true
+
+distributed_training:
+ distributed_world_size: 24
+ ddp_backend: legacy_ddp
+
+criterion:
+ _name: model
+ log_keys:
+ - ema_decay
+ - target_var
+ - pred_var
+# - avg_self_attn
+# - weights
+
+optimization:
+ max_update: 200000
+ lr: [0.0005]
+
+optimizer:
+ _name: adam
+ adam_betas: (0.9,0.98)
+ adam_eps: 1e-06
+ weight_decay: 0.01
+
+lr_scheduler:
+ _name: cosine
+ warmup_updates: 10000
+
+model:
+ _name: data2vec_audio
+ extractor_mode: layer_norm
+ encoder_layerdrop: 0.05
+ dropout_input: 0.0
+ dropout_features: 0.0
+ feature_grad_mult: 1.0
+ encoder_embed_dim: 768
+
+ mask_prob: 0.65
+ mask_length: 10
+
+ loss_beta: 0
+ loss_scale: null
+
+ instance_norm_target_layer: true
+ layer_norm_targets: true
+ average_top_k_layers: 12
+
+ self_attn_norm_type: deepnorm
+ final_norm_type: deepnorm
+
+ pos_conv_depth: 5
+ conv_pos: 95
+
+ ema_decay: 0.999
+ ema_end_decay: 0.9999
+ ema_anneal_end_step: 30000
+ ema_transformer_only: true
+ ema_layers_only: false
+
+ require_same_masks: true
+ mask_dropout: 0
diff --git a/fairseq/examples/data2vec/config/audio/pretraining/base_librispeech.yaml b/fairseq/examples/data2vec/config/audio/pretraining/base_librispeech.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..c332c5a3f8e5e48150a3c0c3649d81234521bc16
--- /dev/null
+++ b/fairseq/examples/data2vec/config/audio/pretraining/base_librispeech.yaml
@@ -0,0 +1,83 @@
+# @package _group_
+
+common:
+ fp16: true
+ log_format: json
+ log_interval: 200
+ tensorboard_logdir: tb
+
+checkpoint:
+ save_interval: 5
+ save_interval_updates: 25000
+ keep_interval_updates: 1
+ no_epoch_checkpoints: true
+
+task:
+ _name: audio_pretraining
+ data: ???
+ max_sample_size: 320000
+ min_sample_size: 32000
+ normalize: true
+
+dataset:
+ num_workers: 6
+ max_tokens: 3800000
+ skip_invalid_size_inputs_valid_test: true
+ validate_interval: 5
+ required_batch_size_multiple: 1
+ disable_validation: true
+
+distributed_training:
+ distributed_world_size: 16
+ ddp_backend: legacy_ddp
+
+criterion:
+ _name: model
+ log_keys:
+ - ema_decay
+ - target_var
+ - pred_var
+
+optimization:
+ max_update: 400000
+ lr: [0.0005]
+
+optimizer:
+ _name: adam
+ adam_betas: (0.9,0.98)
+ adam_eps: 1e-06
+ weight_decay: 0.01
+
+lr_scheduler:
+ _name: tri_stage
+ phase_ratio: [0.03,0.9,0.07]
+
+model:
+ _name: data2vec_audio
+ extractor_mode: layer_norm
+ encoder_layerdrop: 0.05
+ dropout_input: 0.0
+ dropout_features: 0.0
+ feature_grad_mult: 1.0
+ encoder_embed_dim: 768
+
+ mask_prob: 0.65
+ mask_length: 10
+
+ loss_beta: 0
+ loss_scale: null
+
+ instance_norm_target_layer: true
+ average_top_k_layers: 8
+
+ pos_conv_depth: 5
+ conv_pos: 95
+
+ ema_decay: 0.999
+ ema_end_decay: 0.9999
+ ema_anneal_end_step: 30000
+ ema_transformer_only: true
+ ema_layers_only: true
+
+ require_same_masks: true
+ mask_dropout: 0
diff --git a/fairseq/examples/data2vec/config/audio/pretraining/run_config/local.yaml b/fairseq/examples/data2vec/config/audio/pretraining/run_config/local.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..45595f9eea7c369a1200d802bfa1883c1bdfe573
--- /dev/null
+++ b/fairseq/examples/data2vec/config/audio/pretraining/run_config/local.yaml
@@ -0,0 +1,15 @@
+# @package _global_
+hydra:
+ sweep:
+ dir: ${env:PWD}/tmp_dbg/${now:%H-%M-%S}
+
+distributed_training:
+ distributed_world_size: 1
+ nprocs_per_node: 1
+ distributed_port: -1
+
+common:
+ log_interval: 1
+
+dataset:
+ num_workers: 0
diff --git a/fairseq/examples/data2vec/config/audio/pretraining/run_config/slurm_1.yaml b/fairseq/examples/data2vec/config/audio/pretraining/run_config/slurm_1.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..732f0188993a14c847ab5d5e8addb24b0b536311
--- /dev/null
+++ b/fairseq/examples/data2vec/config/audio/pretraining/run_config/slurm_1.yaml
@@ -0,0 +1,37 @@
+# @package _global_
+
+hydra:
+ job:
+ config:
+ override_dirname:
+ kv_sep: ':'
+ item_sep: '/'
+ exclude_keys:
+ - run_config
+ - distributed_training.distributed_port
+ - distributed_training.distributed_world_size
+ - model.pretrained_model_path
+ - model.target_network_path
+ - next_script
+ - task.cache_in_scratch
+ - task.data
+ - checkpoint.save_interval_updates
+ - checkpoint.keep_interval_updates
+ - checkpoint.save_on_overflow
+ - common.log_interval
+ - common.user_dir
+ sweep:
+ dir: /checkpoint/${env:USER}/${env:PREFIX}/${hydra.job.config_name}_${hydra.launcher.gpus_per_node}/${hydra.job.override_dirname}
+ subdir: ''
+ launcher:
+ submitit_folder: ${hydra.sweep.dir}
+ timeout_min: 4320
+ cpus_per_task: 80
+ gpus_per_node: 8
+ tasks_per_node: 1
+ mem_gb: 450
+ nodes: 1
+ name: ${env:PREFIX}_${hydra.job.config_name}
+ partition: devlab,learnlab,learnfair,scavenge
+ constraint: volta32gb,ib4
+ max_num_timeout: 30
diff --git a/fairseq/examples/data2vec/config/audio/pretraining/run_config/slurm_1_aws.yaml b/fairseq/examples/data2vec/config/audio/pretraining/run_config/slurm_1_aws.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..e2bab5675a561a6a01b5db06530b119a282bbf4a
--- /dev/null
+++ b/fairseq/examples/data2vec/config/audio/pretraining/run_config/slurm_1_aws.yaml
@@ -0,0 +1,36 @@
+# @package _global_
+
+hydra:
+ job:
+ config:
+ override_dirname:
+ kv_sep: ':'
+ item_sep: '/'
+ exclude_keys:
+ - run_config
+ - distributed_training.distributed_port
+ - distributed_training.distributed_world_size
+ - model.pretrained_model_path
+ - model.target_network_path
+ - next_script
+ - task.cache_in_scratch
+ - task.data
+ - checkpoint.save_interval_updates
+ - checkpoint.keep_interval_updates
+ - checkpoint.save_on_overflow
+ - common.log_interval
+ - common.user_dir
+ sweep:
+ dir: /checkpoint/${env:USER}/${env:PREFIX}/${hydra.job.config_name}_${hydra.launcher.gpus_per_node}/${hydra.job.override_dirname}
+ subdir: ''
+ launcher:
+ submitit_folder: ${hydra.sweep.dir}
+ timeout_min: 4320
+ cpus_per_task: 80
+ gpus_per_node: 8
+ tasks_per_node: 1
+ mem_gb: 0
+ nodes: 1
+ name: ${env:PREFIX}_${hydra.job.config_name}
+ partition: wav2vec,learnlab,learnfair
+ max_num_timeout: 30
diff --git a/fairseq/examples/data2vec/config/audio/pretraining/run_config/slurm_2.yaml b/fairseq/examples/data2vec/config/audio/pretraining/run_config/slurm_2.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..ec53dc2a984babdf7e61930350868e38f0d769c3
--- /dev/null
+++ b/fairseq/examples/data2vec/config/audio/pretraining/run_config/slurm_2.yaml
@@ -0,0 +1,37 @@
+# @package _global_
+
+hydra:
+ job:
+ config:
+ override_dirname:
+ kv_sep: ':'
+ item_sep: '/'
+ exclude_keys:
+ - run_config
+ - distributed_training.distributed_port
+ - distributed_training.distributed_world_size
+ - model.pretrained_model_path
+ - model.target_network_path
+ - next_script
+ - task.cache_in_scratch
+ - task.data
+ - checkpoint.save_interval_updates
+ - checkpoint.keep_interval_updates
+ - checkpoint.save_on_overflow
+ - common.log_interval
+ - common.user_dir
+ sweep:
+ dir: /checkpoint/${env:USER}/${env:PREFIX}/${hydra.job.config_name}_${hydra.launcher.gpus_per_node}/${hydra.job.override_dirname}
+ subdir: ''
+ launcher:
+ submitit_folder: ${hydra.sweep.dir}
+ timeout_min: 4320
+ cpus_per_task: 10
+ gpus_per_node: 8
+ tasks_per_node: 8
+ mem_gb: 450
+ nodes: 2
+ name: ${env:PREFIX}_${hydra.job.config_name}
+ partition: devlab,learnlab,learnfair,scavenge
+ constraint: volta32gb,ib4
+ max_num_timeout: 30
diff --git a/fairseq/examples/data2vec/config/audio/pretraining/run_config/slurm_2_aws.yaml b/fairseq/examples/data2vec/config/audio/pretraining/run_config/slurm_2_aws.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..70cc8cbb5b8078fb4dbebd66964892efca6607ac
--- /dev/null
+++ b/fairseq/examples/data2vec/config/audio/pretraining/run_config/slurm_2_aws.yaml
@@ -0,0 +1,37 @@
+# @package _global_
+
+hydra:
+ job:
+ config:
+ override_dirname:
+ kv_sep: ':'
+ item_sep: '/'
+ exclude_keys:
+ - run_config
+ - distributed_training.distributed_port
+ - distributed_training.distributed_world_size
+ - model.pretrained_model_path
+ - model.target_network_path
+ - next_script
+ - task.cache_in_scratch
+ - task.data
+ - task.post_save_script
+ - checkpoint.save_interval_updates
+ - checkpoint.keep_interval_updates
+ - checkpoint.save_on_overflow
+ - common.log_interval
+ - common.user_dir
+ sweep:
+ dir: /fsx-wav2vec/${env:USER}/${env:PREFIX}/${hydra.job.config_name}_${hydra.launcher.gpus_per_node}/${hydra.job.override_dirname}
+ subdir: ''
+ launcher:
+ submitit_folder: ${hydra.sweep.dir}
+ timeout_min: 4320
+ cpus_per_task: 10
+ gpus_per_node: 8
+ tasks_per_node: 8
+ mem_gb: 0
+ nodes: 2
+ name: ${env:PREFIX}_${hydra.job.config_name}
+ partition: wav2vec,learnlab,learnfair
+ max_num_timeout: 30
diff --git a/fairseq/examples/data2vec/config/audio/pretraining/run_config/slurm_3.yaml b/fairseq/examples/data2vec/config/audio/pretraining/run_config/slurm_3.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..14b47d14e6282d5a846cc7c41c6b2080c5ed8c62
--- /dev/null
+++ b/fairseq/examples/data2vec/config/audio/pretraining/run_config/slurm_3.yaml
@@ -0,0 +1,36 @@
+# @package _global_
+
+hydra:
+ job:
+ config:
+ override_dirname:
+ kv_sep: ':'
+ item_sep: '/'
+ exclude_keys:
+ - run_config
+ - distributed_training.distributed_port
+ - distributed_training.distributed_world_size
+ - model.pretrained_model_path
+ - model.target_network_path
+ - next_script
+ - task.cache_in_scratch
+ - task.data
+ - checkpoint.save_interval_updates
+ - checkpoint.keep_interval_updates
+ - checkpoint.save_on_overflow
+ - common.log_interval
+ sweep:
+ dir: /checkpoint/${env:USER}/${env:PREFIX}/${hydra.job.config_name}_${hydra.launcher.gpus_per_node}/${hydra.job.override_dirname}
+ subdir: ''
+ launcher:
+ submitit_folder: ${hydra.sweep.dir}
+ timeout_min: 4320
+ cpus_per_task: 80
+ gpus_per_node: 8
+ tasks_per_node: 1
+ mem_gb: 450
+ nodes: 3
+ name: ${env:PREFIX}_${hydra.job.config_name}
+ partition: devlab,learnlab,learnfair,scavenge
+ constraint: volta32gb,ib4
+ max_num_timeout: 30
diff --git a/fairseq/examples/data2vec/config/audio/pretraining/run_config/slurm_4.yaml b/fairseq/examples/data2vec/config/audio/pretraining/run_config/slurm_4.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..c54d735fb2dc0b0af2d467caa7f64405af18ea10
--- /dev/null
+++ b/fairseq/examples/data2vec/config/audio/pretraining/run_config/slurm_4.yaml
@@ -0,0 +1,36 @@
+# @package _global_
+
+hydra:
+ job:
+ config:
+ override_dirname:
+ kv_sep: ':'
+ item_sep: '/'
+ exclude_keys:
+ - run_config
+ - distributed_training.distributed_port
+ - distributed_training.distributed_world_size
+ - model.pretrained_model_path
+ - model.target_network_path
+ - next_script
+ - task.cache_in_scratch
+ - task.data
+ - checkpoint.save_interval_updates
+ - checkpoint.keep_interval_updates
+ - checkpoint.save_on_overflow
+ - common.log_interval
+ sweep:
+ dir: /checkpoint/${env:USER}/${env:PREFIX}/${hydra.job.config_name}_${hydra.launcher.gpus_per_node}/${hydra.job.override_dirname}
+ subdir: ''
+ launcher:
+ submitit_folder: ${hydra.sweep.dir}
+ timeout_min: 4320
+ cpus_per_task: 10
+ gpus_per_node: 8
+ tasks_per_node: 8
+ mem_gb: 450
+ nodes: 4
+ name: ${env:PREFIX}_${hydra.job.config_name}
+ partition: devlab,learnlab,learnfair,scavenge
+ constraint: volta32gb,ib4
+ max_num_timeout: 30
diff --git a/fairseq/examples/data2vec/config/audio/pretraining/run_config/slurm_4_aws.yaml b/fairseq/examples/data2vec/config/audio/pretraining/run_config/slurm_4_aws.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..0231b2690d63f758d13eff9c3f5c5826cefba050
--- /dev/null
+++ b/fairseq/examples/data2vec/config/audio/pretraining/run_config/slurm_4_aws.yaml
@@ -0,0 +1,37 @@
+# @package _global_
+
+hydra:
+ job:
+ config:
+ override_dirname:
+ kv_sep: ':'
+ item_sep: '/'
+ exclude_keys:
+ - run_config
+ - distributed_training.distributed_port
+ - distributed_training.distributed_world_size
+ - model.pretrained_model_path
+ - model.target_network_path
+ - next_script
+ - task.cache_in_scratch
+ - task.data
+ - task.post_save_script
+ - checkpoint.save_interval_updates
+ - checkpoint.keep_interval_updates
+ - checkpoint.save_on_overflow
+ - common.log_interval
+ - common.user_dir
+ sweep:
+ dir: /fsx-wav2vec/${env:USER}/${env:PREFIX}/${hydra.job.config_name}_${hydra.launcher.gpus_per_node}/${hydra.job.override_dirname}
+ subdir: ''
+ launcher:
+ submitit_folder: ${hydra.sweep.dir}
+ timeout_min: 4320
+ cpus_per_task: 10
+ gpus_per_node: 8
+ tasks_per_node: 8
+ mem_gb: 0
+ nodes: 4
+ name: ${env:PREFIX}_${hydra.job.config_name}
+ partition: wav2vec,learnlab,learnfair
+ max_num_timeout: 30
diff --git a/fairseq/examples/data2vec/config/audio/pretraining/run_config/slurm_6_aws.yaml b/fairseq/examples/data2vec/config/audio/pretraining/run_config/slurm_6_aws.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..9a4e43a987e8810b42a4601af5c51072c19fe6b7
--- /dev/null
+++ b/fairseq/examples/data2vec/config/audio/pretraining/run_config/slurm_6_aws.yaml
@@ -0,0 +1,36 @@
+# @package _global_
+
+hydra:
+ job:
+ config:
+ override_dirname:
+ kv_sep: ':'
+ item_sep: '/'
+ exclude_keys:
+ - run_config
+ - distributed_training.distributed_port
+ - distributed_training.distributed_world_size
+ - model.pretrained_model_path
+ - model.target_network_path
+ - next_script
+ - task.cache_in_scratch
+ - task.data
+ - checkpoint.save_interval_updates
+ - checkpoint.keep_interval_updates
+ - checkpoint.save_on_overflow
+ - common.log_interval
+ - common.user_dir
+ sweep:
+ dir: /fsx-wav2vec/${env:USER}/${env:PREFIX}/${hydra.job.config_name}_${hydra.launcher.gpus_per_node}/${hydra.job.override_dirname}
+ subdir: ''
+ launcher:
+ submitit_folder: ${hydra.sweep.dir}
+ timeout_min: 4320
+ cpus_per_task: 10
+ gpus_per_node: 8
+ tasks_per_node: 8
+ mem_gb: 0
+ nodes: 6
+ name: ${env:PREFIX}_${hydra.job.config_name}
+ partition: wav2vec,learnlab,learnfair
+ max_num_timeout: 30
diff --git a/fairseq/examples/data2vec/config/audio/pretraining/run_config/slurm_8_aws.yaml b/fairseq/examples/data2vec/config/audio/pretraining/run_config/slurm_8_aws.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..78c9f57aeb2111f6a3e7d99005f25fbd359c7a72
--- /dev/null
+++ b/fairseq/examples/data2vec/config/audio/pretraining/run_config/slurm_8_aws.yaml
@@ -0,0 +1,36 @@
+# @package _global_
+
+hydra:
+ job:
+ config:
+ override_dirname:
+ kv_sep: ':'
+ item_sep: '/'
+ exclude_keys:
+ - run_config
+ - distributed_training.distributed_port
+ - distributed_training.distributed_world_size
+ - model.pretrained_model_path
+ - model.target_network_path
+ - next_script
+ - task.cache_in_scratch
+ - task.data
+ - checkpoint.save_interval_updates
+ - checkpoint.keep_interval_updates
+ - checkpoint.save_on_overflow
+ - common.log_interval
+ - common.user_dir
+ sweep:
+ dir: /fsx-wav2vec/${env:USER}/${env:PREFIX}/${hydra.job.config_name}_${hydra.launcher.gpus_per_node}/${hydra.job.override_dirname}
+ subdir: ''
+ launcher:
+ submitit_folder: ${hydra.sweep.dir}
+ timeout_min: 4320
+ cpus_per_task: 10
+ gpus_per_node: 8
+ tasks_per_node: 8
+ mem_gb: 0
+ nodes: 8
+ name: ${env:PREFIX}_${hydra.job.config_name}
+ partition: wav2vec,learnlab,learnfair
+ max_num_timeout: 30
diff --git a/fairseq/examples/data2vec/config/text/pretraining/base.yaml b/fairseq/examples/data2vec/config/text/pretraining/base.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..c6b07c4052aaf64d5b73bd983e395fa983b54cd5
--- /dev/null
+++ b/fairseq/examples/data2vec/config/text/pretraining/base.yaml
@@ -0,0 +1,77 @@
+# @package _group_
+common:
+ fp16: true
+ log_format: json
+ log_interval: 200
+ tensorboard_logdir: tb
+
+checkpoint:
+ no_epoch_checkpoints: true
+ save_interval_updates: 50000
+ keep_interval_updates: 1
+
+distributed_training:
+ distributed_world_size: 16
+ ddp_backend: legacy_ddp
+
+task:
+ _name: masked_lm
+ data: ???
+ sample_break_mode: complete_doc
+ tokens_per_sample: 512
+ include_target_tokens: true
+ random_token_prob: 0
+ leave_unmasked_prob: 0
+ mask_prob: 0.35
+ mask_multiple_length: 4
+
+criterion: model
+
+dataset:
+ max_tokens: 8192
+ ignore_unused_valid_subsets: true
+ skip_invalid_size_inputs_valid_test: true
+
+optimizer:
+ _name: adam
+ weight_decay: 0.01
+ adam_betas: (0.9,0.98)
+ adam_eps: 1e-06
+
+lr_scheduler:
+ _name: cosine
+ warmup_updates: 10000
+
+optimization:
+ clip_norm: 5
+ lr: [0.0002]
+ max_update: 1000000
+ update_freq: [1]
+
+model:
+ _name: data2vec_text
+ head_layers: 2
+ average_top_k_layers: 10
+ layer_norm_target_layer: true
+ loss_scale: 1
+ ema_decay: 0.999
+ ema_end_decay: 0.9999
+ ema_anneal_end_step: 300000
+ loss_beta: 4
+ ema_transformer_layers_only: true
+
+ transformer:
+ dropout: 0.1
+ attention_dropout: 0.1
+ layernorm_embedding: true
+ activation_fn: gelu
+ no_scale_embedding: true
+ max_source_positions: 512
+ encoder:
+ embed_dim: 768
+ ffn_embed_dim: 3072
+ layers: 12
+ attention_heads: 12
+ normalize_before: false
+ learned_pos: true
+ layerdrop: 0
diff --git a/fairseq/examples/data2vec/config/text/pretraining/run_config/local.yaml b/fairseq/examples/data2vec/config/text/pretraining/run_config/local.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..45595f9eea7c369a1200d802bfa1883c1bdfe573
--- /dev/null
+++ b/fairseq/examples/data2vec/config/text/pretraining/run_config/local.yaml
@@ -0,0 +1,15 @@
+# @package _global_
+hydra:
+ sweep:
+ dir: ${env:PWD}/tmp_dbg/${now:%H-%M-%S}
+
+distributed_training:
+ distributed_world_size: 1
+ nprocs_per_node: 1
+ distributed_port: -1
+
+common:
+ log_interval: 1
+
+dataset:
+ num_workers: 0
diff --git a/fairseq/examples/data2vec/config/text/pretraining/run_config/slurm_1_aws.yaml b/fairseq/examples/data2vec/config/text/pretraining/run_config/slurm_1_aws.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..4bac45a58de960ea347167f6da6d8d73910da4e1
--- /dev/null
+++ b/fairseq/examples/data2vec/config/text/pretraining/run_config/slurm_1_aws.yaml
@@ -0,0 +1,37 @@
+# @package _global_
+
+hydra:
+ job:
+ config:
+ override_dirname:
+ kv_sep: '_'
+ item_sep: '/'
+ exclude_keys:
+ - run_config
+ - distributed_training.distributed_port
+ - distributed_training.distributed_world_size
+ - model.pretrained_model_path
+ - model.target_network_path
+ - next_script
+ - task.cache_in_scratch
+ - task.data
+ - checkpoint.save_interval_updates
+ - checkpoint.keep_interval_updates
+ - checkpoint.save_on_overflow
+ - common.log_interval
+ - common.user_dir
+ sweep:
+ dir: /fsx-wav2vec/${env:USER}/${env:PREFIX}/${hydra.job.config_name}_${hydra.launcher.gpus_per_node}/${hydra.job.override_dirname}
+ subdir: ''
+ launcher:
+ submitit_folder: ${hydra.sweep.dir}/submitit
+ timeout_min: 4320
+ cpus_per_task: 80
+ gpus_per_node: 8
+ tasks_per_node: 1
+ mem_gb: 0
+ nodes: 1
+ name: ${env:PREFIX}_${hydra.job.config_name}
+ partition: wav2vec
+ max_num_timeout: 30
+ exclude: a100-st-p4d24xlarge-471
diff --git a/fairseq/examples/data2vec/config/text/pretraining/run_config/slurm_2.yaml b/fairseq/examples/data2vec/config/text/pretraining/run_config/slurm_2.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..006a0f2116db5a85da558cada65873f1150eb717
--- /dev/null
+++ b/fairseq/examples/data2vec/config/text/pretraining/run_config/slurm_2.yaml
@@ -0,0 +1,37 @@
+# @package _global_
+
+hydra:
+ job:
+ config:
+ override_dirname:
+ kv_sep: ':'
+ item_sep: '/'
+ exclude_keys:
+ - run_config
+ - distributed_training.distributed_port
+ - distributed_training.distributed_world_size
+ - model.pretrained_model_path
+ - model.target_network_path
+ - next_script
+ - task.cache_in_scratch
+ - task.data
+ - checkpoint.save_interval_updates
+ - checkpoint.keep_interval_updates
+ - checkpoint.save_on_overflow
+ - common.log_interval
+ - common.user_dir
+ sweep:
+ dir: /checkpoint/${env:USER}/${env:PREFIX}/${hydra.job.config_name}_${hydra.launcher.gpus_per_node}/${hydra.job.override_dirname}
+ subdir: ''
+ launcher:
+ submitit_folder: ${hydra.sweep.dir}
+ timeout_min: 4320
+ cpus_per_task: 80
+ gpus_per_node: 8
+ tasks_per_node: 1
+ mem_gb: 450
+ nodes: 2
+ name: ${env:PREFIX}_${hydra.job.config_name}
+ partition: devlab,learnlab,learnfair,scavenge
+ constraint: volta32gb,ib4
+ max_num_timeout: 30
diff --git a/fairseq/examples/data2vec/config/text/pretraining/run_config/slurm_2_aws.yaml b/fairseq/examples/data2vec/config/text/pretraining/run_config/slurm_2_aws.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..4292198b4ef9608b6550652fd72d55a9acf85cc0
--- /dev/null
+++ b/fairseq/examples/data2vec/config/text/pretraining/run_config/slurm_2_aws.yaml
@@ -0,0 +1,37 @@
+# @package _global_
+
+hydra:
+ job:
+ config:
+ override_dirname:
+ kv_sep: '_'
+ item_sep: '/'
+ exclude_keys:
+ - run_config
+ - distributed_training.distributed_port
+ - distributed_training.distributed_world_size
+ - model.pretrained_model_path
+ - model.target_network_path
+ - next_script
+ - task.cache_in_scratch
+ - task.data
+ - checkpoint.save_interval_updates
+ - checkpoint.keep_interval_updates
+ - checkpoint.save_on_overflow
+ - common.log_interval
+ - common.user_dir
+ sweep:
+ dir: /fsx-wav2vec/${env:USER}/${env:PREFIX}/${hydra.job.config_name}_${hydra.launcher.gpus_per_node}/${hydra.job.override_dirname}
+ subdir: ''
+ launcher:
+ submitit_folder: ${hydra.sweep.dir}/submitit
+ timeout_min: 4320
+ cpus_per_task: 10
+ gpus_per_node: 8
+ tasks_per_node: 8
+ mem_gb: 0
+ nodes: 2
+ name: ${env:PREFIX}_${hydra.job.config_name}
+ partition: wav2vec
+ max_num_timeout: 30
+ exclude: a100-st-p4d24xlarge-471
diff --git a/fairseq/examples/data2vec/config/text/pretraining/run_config/slurm_3.yaml b/fairseq/examples/data2vec/config/text/pretraining/run_config/slurm_3.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..0e1555d20f9380b9d82d5852be6799d35fa3d078
--- /dev/null
+++ b/fairseq/examples/data2vec/config/text/pretraining/run_config/slurm_3.yaml
@@ -0,0 +1,36 @@
+# @package _global_
+
+hydra:
+ job:
+ config:
+ override_dirname:
+ kv_sep: ':'
+ item_sep: '/'
+ exclude_keys:
+ - run_config
+ - distributed_training.distributed_port
+ - distributed_training.distributed_world_size
+ - model.pretrained_model_path
+ - model.target_network_path
+ - next_script
+ - task.cache_in_scratch
+ - task.data
+ - checkpoint.save_interval_updates
+ - checkpoint.keep_interval_updates
+ - checkpoint.save_on_overflow
+ - common.log_interval
+ sweep:
+ dir: /checkpoint/${env:USER}/${env:PREFIX}/${hydra.job.config_name}_${hydra.launcher.gpus_per_node}/${hydra.job.override_dirname}
+ subdir: ''
+ launcher:
+ submitit_folder: ${hydra.sweep.dir}
+ timeout_min: 4320
+ cpus_per_task: 10
+ gpus_per_node: 8
+ tasks_per_node: 8
+ mem_gb: 450
+ nodes: 3
+ name: ${env:PREFIX}_${hydra.job.config_name}
+ partition: devlab,learnlab,learnfair,scavenge
+ constraint: volta32gb,ib4
+ max_num_timeout: 30
diff --git a/fairseq/examples/data2vec/config/text/pretraining/run_config/slurm_4.yaml b/fairseq/examples/data2vec/config/text/pretraining/run_config/slurm_4.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..c54d735fb2dc0b0af2d467caa7f64405af18ea10
--- /dev/null
+++ b/fairseq/examples/data2vec/config/text/pretraining/run_config/slurm_4.yaml
@@ -0,0 +1,36 @@
+# @package _global_
+
+hydra:
+ job:
+ config:
+ override_dirname:
+ kv_sep: ':'
+ item_sep: '/'
+ exclude_keys:
+ - run_config
+ - distributed_training.distributed_port
+ - distributed_training.distributed_world_size
+ - model.pretrained_model_path
+ - model.target_network_path
+ - next_script
+ - task.cache_in_scratch
+ - task.data
+ - checkpoint.save_interval_updates
+ - checkpoint.keep_interval_updates
+ - checkpoint.save_on_overflow
+ - common.log_interval
+ sweep:
+ dir: /checkpoint/${env:USER}/${env:PREFIX}/${hydra.job.config_name}_${hydra.launcher.gpus_per_node}/${hydra.job.override_dirname}
+ subdir: ''
+ launcher:
+ submitit_folder: ${hydra.sweep.dir}
+ timeout_min: 4320
+ cpus_per_task: 10
+ gpus_per_node: 8
+ tasks_per_node: 8
+ mem_gb: 450
+ nodes: 4
+ name: ${env:PREFIX}_${hydra.job.config_name}
+ partition: devlab,learnlab,learnfair,scavenge
+ constraint: volta32gb,ib4
+ max_num_timeout: 30
diff --git a/fairseq/examples/data2vec/config/text/pretraining/run_config/slurm_4_aws.yaml b/fairseq/examples/data2vec/config/text/pretraining/run_config/slurm_4_aws.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..5df84cd6da99d06fc32f3596cc1d0f8d5e62ee1f
--- /dev/null
+++ b/fairseq/examples/data2vec/config/text/pretraining/run_config/slurm_4_aws.yaml
@@ -0,0 +1,41 @@
+# @package _global_
+
+hydra:
+ job:
+ config:
+ override_dirname:
+ kv_sep: '_'
+ item_sep: '/'
+ exclude_keys:
+ - run_config
+ - distributed_training.distributed_port
+ - distributed_training.distributed_world_size
+ - model.pretrained_model_path
+ - model.target_network_path
+ - next_script
+ - task.cache_in_scratch
+ - task.data
+ - checkpoint.save_interval_updates
+ - checkpoint.keep_interval_updates
+ - checkpoint.save_on_overflow
+ - common.log_interval
+ - common.user_dir
+ sweep:
+ dir: /fsx-wav2vec/${env:USER}/${env:PREFIX}/${hydra.job.config_name}_${hydra.launcher.gpus_per_node}/${hydra.job.override_dirname}
+ subdir: ''
+ launcher:
+ submitit_folder: ${hydra.sweep.dir}/submitit
+ timeout_min: 4320
+ cpus_per_task: 10
+ gpus_per_node: 8
+ tasks_per_node: 8
+ mem_gb: 0
+ nodes: 4
+ name: ${env:PREFIX}_${hydra.job.config_name}
+ partition: wav2vec
+ max_num_timeout: 30
+ exclude: a100-st-p4d24xlarge-471
+
+distributed_training:
+ distributed_world_size: 32
+ ddp_backend: legacy_ddp
diff --git a/fairseq/examples/data2vec/config/text/pretraining/run_config/slurm_8_aws.yaml b/fairseq/examples/data2vec/config/text/pretraining/run_config/slurm_8_aws.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..5b32c23a665f9374d1c919c52dcbf19665c47f4f
--- /dev/null
+++ b/fairseq/examples/data2vec/config/text/pretraining/run_config/slurm_8_aws.yaml
@@ -0,0 +1,41 @@
+# @package _global_
+
+hydra:
+ job:
+ config:
+ override_dirname:
+ kv_sep: '_'
+ item_sep: '/'
+ exclude_keys:
+ - run_config
+ - distributed_training.distributed_port
+ - distributed_training.distributed_world_size
+ - model.pretrained_model_path
+ - model.target_network_path
+ - next_script
+ - task.cache_in_scratch
+ - task.data
+ - checkpoint.save_interval_updates
+ - checkpoint.keep_interval_updates
+ - checkpoint.save_on_overflow
+ - common.log_interval
+ - common.user_dir
+ sweep:
+ dir: /fsx-wav2vec/${env:USER}/${env:PREFIX}/${hydra.job.config_name}_${hydra.launcher.gpus_per_node}/${hydra.job.override_dirname}
+ subdir: ''
+ launcher:
+ submitit_folder: ${hydra.sweep.dir}/submitit
+ timeout_min: 4320
+ cpus_per_task: 10
+ gpus_per_node: 8
+ tasks_per_node: 8
+ mem_gb: 0
+ nodes: 8
+ name: pt
+ partition: wav2vec
+ max_num_timeout: 30
+ exclude: a100-st-p4d24xlarge-471
+
+distributed_training:
+ distributed_world_size: 64
+ ddp_backend: legacy_ddp
diff --git a/fairseq/examples/data2vec/config/v2/base_audio_only_task.yaml b/fairseq/examples/data2vec/config/v2/base_audio_only_task.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..65a9ab3e738a8361b229b6c42fc11c7c0480f264
--- /dev/null
+++ b/fairseq/examples/data2vec/config/v2/base_audio_only_task.yaml
@@ -0,0 +1,113 @@
+# @package _group_
+
+common:
+ fp16: true
+ log_format: json
+ log_interval: 200
+ tensorboard_logdir: tb
+ min_loss_scale: 1e-6
+ fp16_no_flatten_grads: false
+ user_dir: ${env:PWD}/examples/data2vec
+
+checkpoint:
+ save_interval: 1
+ save_interval_updates: 25000
+ keep_interval_updates: 1
+ no_epoch_checkpoints: true
+
+task:
+ _name: audio_pretraining
+ data: /private/home/abaevski/data/librispeech/full
+ max_sample_size: 320000
+ min_sample_size: 32000
+ normalize: true
+ precompute_mask_config: {}
+
+dataset:
+ num_workers: 6
+ max_tokens: 1000000
+ skip_invalid_size_inputs_valid_test: true
+ validate_interval: 5
+ required_batch_size_multiple: 1
+ disable_validation: true
+
+distributed_training:
+ distributed_world_size: 8
+ ddp_backend: legacy_ddp
+
+criterion:
+ _name: model
+ log_keys:
+ - ema_decay
+ - target_var
+ - pred_var
+ - model_norm
+ - ema_norm
+ - masked_pct
+
+optimization:
+ max_update: 400000
+ lr: [0.00075]
+ debug_param_names: true
+
+optimizer:
+ _name: adam
+ adam_betas: [ 0.9,0.98 ]
+ adam_eps: 1e-06
+ weight_decay: 0.01
+
+lr_scheduler:
+ _name: cosine
+ warmup_updates: 8000
+
+model:
+ _name: data2vec_multi
+
+ loss_beta: 0
+ loss_scale: null
+
+ depth: 12
+ embed_dim: 768
+ clone_batch: 8
+
+ ema_decay: 0.999
+ ema_end_decay: 0.99999
+ ema_anneal_end_step: 75000
+ ema_encoder_only: false
+
+ average_top_k_layers: 8
+ instance_norm_target_layer: true
+ layer_norm_target_layer: false
+ layer_norm_targets: false
+
+ layerdrop: 0.05
+ norm_eps: 1e-5
+
+ supported_modality: AUDIO
+
+ modalities:
+ audio:
+ feature_encoder_spec: '[(512, 10, 5)] + [(512, 3, 2)] * 4 + [(512,2,2)] + [(512,2,2)]'
+ conv_pos_depth: 5
+ conv_pos_width: 95
+ conv_pos_groups: 16
+ prenet_depth: 0
+ mask_prob: 0.5
+ mask_prob_adjust: 0.05
+ inverse_mask: false
+ mask_length: 5
+ mask_noise_std: 0.01
+ mask_dropout: 0
+ add_masks: false
+ ema_local_encoder: false
+ use_alibi_encoder: true
+ prenet_layerdrop: 0.05
+ prenet_dropout: 0.1
+ learned_alibi_scale: true
+ learned_alibi_scale_per_head: true
+ decoder:
+ input_dropout: 0.1
+ decoder_dim: 384
+ decoder_groups: 16
+ decoder_kernel: 7
+ decoder_layers: 4
diff --git a/fairseq/examples/data2vec/config/v2/base_images_only_task.yaml b/fairseq/examples/data2vec/config/v2/base_images_only_task.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..ff0c247b1387f10945145788ed465177ac5d554e
--- /dev/null
+++ b/fairseq/examples/data2vec/config/v2/base_images_only_task.yaml
@@ -0,0 +1,116 @@
+# @package _group_
+
+common:
+ fp16: true
+ log_format: json
+ log_interval: 200
+ tensorboard_logdir: tb
+ min_loss_scale: 1e-6
+ fp16_no_flatten_grads: true
+ user_dir: ${env:PWD}/examples/data2vec
+
+checkpoint:
+ save_interval: 5
+ save_interval_updates: 25000
+ keep_interval_updates: 1
+ no_epoch_checkpoints: true
+
+task:
+ _name: mae_image_pretraining
+ data: /datasets01/imagenet_full_size/061417/
+ rebuild_batches: true
+ local_cache_path: /scratch/cache_abaevski/imagenet
+ key: source
+ precompute_mask_config: {}
+
+dataset:
+ num_workers: 10
+ batch_size: 16
+ skip_invalid_size_inputs_valid_test: true
+ required_batch_size_multiple: 1
+ disable_validation: true
+
+distributed_training:
+ distributed_world_size: 16
+ ddp_backend: c10d
+
+criterion:
+ _name: model
+ log_keys:
+ - ema_decay
+ - target_var
+ - pred_var
+ - model_norm
+ - ema_norm
+ - masked_pct
+
+optimization:
+ max_update: 375300
+ lr: [ 0.001 ]
+ debug_param_names: true
+ clip_norm: 4
+
+optimizer:
+ _name: composite
+ dynamic_groups: true
+ groups:
+ default:
+ lr_float: 1e-3
+ optimizer:
+ _name: adam
+ adam_betas: [0.9,0.95]
+ weight_decay: 0.05
+ lr_scheduler:
+ _name: cosine
+ warmup_updates: 50040
+
+lr_scheduler: pass_through
+
+model:
+ _name: data2vec_multi
+
+ ema_decay: 0.9998
+ ema_end_decay: 0.99999
+ ema_anneal_end_step: 100000
+ instance_norm_target_layer: true
+ layer_norm_target_layer: false
+ layer_norm_targets: true
+ end_of_block_targets: false
+
+ depth: 10
+ average_top_k_layers: 10
+ clone_batch: 16
+
+ norm_eps: 1e-6
+
+ min_target_var: 0
+ min_pred_var: 0
+
+ encoder_dropout: 0
+ post_mlp_drop: 0
+ attention_dropout: 0
+ activation_dropout: 0
+
+ supported_modality: IMAGE
+ cls_loss: 0.01
+
+ ema_encoder_only: false
+
+ modalities:
+ image:
+ inverse_mask: true
+ mask_prob: 0.8
+ mask_prob_adjust: 0.07
+ mask_length: 3
+ mask_noise_std: 0.01
+ prenet_depth: 2
+ ema_local_encoder: true
+ num_extra_tokens: 1
+ init_extra_token_zero: false
+ use_alibi_encoder: false
+ decoder:
+ decoder_dim: 768
+ decoder_groups: 16
+ decoder_kernel: 3
+ decoder_layers: 6
+ input_dropout: 0
\ No newline at end of file
diff --git a/fairseq/examples/data2vec/config/v2/base_text_only_task.yaml b/fairseq/examples/data2vec/config/v2/base_text_only_task.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..62f22eb0fe50c1b8d8febfb01529b32c14dece93
--- /dev/null
+++ b/fairseq/examples/data2vec/config/v2/base_text_only_task.yaml
@@ -0,0 +1,112 @@
+# @package _group_
+
+common:
+ fp16: true
+ log_format: json
+ log_interval: 200
+ tensorboard_logdir: tb
+ fp16_no_flatten_grads: true
+ user_dir: ${env:PWD}/examples/data2vec
+
+checkpoint:
+ no_epoch_checkpoints: true
+ save_interval_updates: 50000
+ keep_interval_updates: 1
+
+distributed_training:
+ distributed_world_size: 16
+ ddp_backend: legacy_ddp
+
+task:
+ _name: masked_lm
+ data: /fsx-wav2vec/abaevski/data/nlp/bookwiki_aml-full-mmap2-bin
+ sample_break_mode: none
+ tokens_per_sample: 512
+ include_target_tokens: true
+ random_token_prob: 0
+ leave_unmasked_prob: 0
+ include_index: True
+ skip_masking: True
+ d2v2_multi: True
+
+criterion:
+ _name: model
+ log_keys:
+ - ema_decay
+ - target_var
+ - pred_var
+ - model_norm
+ - ema_norm
+ - masked_pct
+
+dataset:
+ batch_size: 4
+ ignore_unused_valid_subsets: true
+ skip_invalid_size_inputs_valid_test: true
+ disable_validation: true
+
+optimization:
+ clip_norm: 1
+ lr: [0.0002]
+ max_update: 1000000
+ update_freq: [1]
+
+optimizer:
+ _name: composite
+ dynamic_groups: true
+ groups:
+ default:
+ lr_float: 0.0002
+ optimizer:
+ _name: adam
+ adam_betas: [0.9,0.98]
+ adam_eps: 1e-06
+ weight_decay: 0.01
+ lr_scheduler:
+ _name: cosine
+ warmup_updates: 4000
+
+lr_scheduler: pass_through
+
+model:
+ _name: data2vec_multi
+
+ loss_beta: 0
+ loss_scale: 1
+
+ depth: 12
+ embed_dim: 768
+ clone_batch: 8
+
+ ema_decay: 0.9999
+ ema_end_decay: 0.99999
+ ema_anneal_end_step: 100000
+ ema_encoder_only: true
+
+ average_top_k_layers: 12
+ layer_norm_target_layer: false
+ instance_norm_target_layer: true
+ batch_norm_target_layer: false
+ instance_norm_targets: false
+ layer_norm_targets: false
+
+ layerdrop: 0
+ norm_eps: 1e-5
+
+ supported_modality: TEXT
+
+ modalities:
+ text:
+ mask_prob: 0.48
+ mask_length: 1
+ mask_noise_std: 0.01
+ prenet_depth: 0
+ decoder:
+ input_dropout: 0.1
+ decoder_dim: 768
+ decoder_groups: 1
+ decoder_kernel: 9
+ decoder_layers: 5
+ decoder_residual: false
+ projection_layers: 2
+ projection_ratio: 2.0
diff --git a/fairseq/examples/data2vec/config/v2/huge_images14_only_task.yaml b/fairseq/examples/data2vec/config/v2/huge_images14_only_task.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..a8a15253f2df07198bf541a76259f9784d43f276
--- /dev/null
+++ b/fairseq/examples/data2vec/config/v2/huge_images14_only_task.yaml
@@ -0,0 +1,122 @@
+# @package _group_
+
+common:
+ fp16: true
+ log_format: json
+ log_interval: 200
+ tensorboard_logdir: tb
+ min_loss_scale: 1e-6
+ fp16_no_flatten_grads: true
+ user_dir: ${env:PWD}/examples/data2vec
+
+checkpoint:
+ save_interval: 5
+ save_interval_updates: 25000
+ keep_interval_updates: 1
+ no_epoch_checkpoints: true
+
+task:
+ _name: mae_image_pretraining
+ data: /datasets01/imagenet_full_size/061417/
+ rebuild_batches: true
+ local_cache_path: /scratch/cache_abaevski/imagenet
+ key: source
+ precompute_mask_config: {}
+
+dataset:
+ num_workers: 10
+ batch_size: 8
+ skip_invalid_size_inputs_valid_test: true
+ required_batch_size_multiple: 1
+ disable_validation: true
+
+distributed_training:
+ distributed_world_size: 32
+ ddp_backend: c10d
+
+criterion:
+ _name: model
+ log_keys:
+ - ema_decay
+ - target_var
+ - pred_var
+ - model_norm
+ - ema_norm
+ - masked_pct
+
+optimization:
+ max_update: 500000
+ lr: [ 0.0004 ]
+ debug_param_names: true
+ clip_norm: 4
+
+optimizer:
+ _name: composite
+ dynamic_groups: true
+ groups:
+ default:
+ lr_float: 4e-4
+ optimizer:
+ _name: adam
+ adam_betas: [0.9,0.95]
+ weight_decay: 0.05
+ lr_scheduler:
+ _name: cosine
+ warmup_updates: 50040
+
+lr_scheduler: pass_through
+
+model:
+ _name: data2vec_multi
+
+ ema_decay: 0.9998
+ ema_end_decay: 1
+ ema_anneal_end_step: 300000
+ instance_norm_target_layer: true
+ layer_norm_target_layer: false
+ layer_norm_targets: true
+ end_of_block_targets: false
+
+ depth: 32
+ embed_dim: 1280
+ num_heads: 16
+
+ average_top_k_layers: 24
+ clone_batch: 16
+
+ norm_eps: 1e-6
+
+ min_target_var: 0
+ min_pred_var: 0
+
+ encoder_dropout: 0
+ post_mlp_drop: 0
+ attention_dropout: 0
+ activation_dropout: 0
+
+ supported_modality: IMAGE
+ cls_loss: 0.01
+
+ ema_encoder_only: false
+
+ modalities:
+ image:
+ patch_size: 14
+ inverse_mask: true
+ mask_prob: 0.75
+ mask_prob_adjust: 0.1
+ mask_length: 3
+ mask_noise_std: 0.01
+ prenet_depth: 0
+ ema_local_encoder: true
+ num_extra_tokens: 1
+ init_extra_token_zero: false
+ use_alibi_encoder: false
+ embed_dim: 1280
+ decoder:
+ decoder_dim: 1024
+ decoder_groups: 16
+ decoder_kernel: 5
+ decoder_layers: 3
+ final_layer_norm: false
+ input_dropout: 0
\ No newline at end of file
diff --git a/fairseq/examples/data2vec/config/v2/huge_images_only_task.yaml b/fairseq/examples/data2vec/config/v2/huge_images_only_task.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..7a352ac3c741a70d958074cabd5d57021d7e9753
--- /dev/null
+++ b/fairseq/examples/data2vec/config/v2/huge_images_only_task.yaml
@@ -0,0 +1,120 @@
+# @package _group_
+
+common:
+ fp16: true
+ log_format: json
+ log_interval: 200
+ tensorboard_logdir: tb
+ min_loss_scale: 1e-6
+ fp16_no_flatten_grads: true
+ user_dir: ${env:PWD}/examples/data2vec
+
+checkpoint:
+ save_interval: 5
+ save_interval_updates: 25000
+ keep_interval_updates: 1
+ no_epoch_checkpoints: true
+
+task:
+ _name: mae_image_pretraining
+ data: /datasets01/imagenet_full_size/061417/
+ rebuild_batches: true
+ local_cache_path: /scratch/cache_abaevski/imagenet
+ key: source
+ precompute_mask_config: {}
+
+dataset:
+ num_workers: 10
+ batch_size: 8
+ skip_invalid_size_inputs_valid_test: true
+ required_batch_size_multiple: 1
+ disable_validation: true
+
+distributed_training:
+ distributed_world_size: 16
+ ddp_backend: c10d
+
+criterion:
+ _name: model
+ log_keys:
+ - ema_decay
+ - target_var
+ - pred_var
+ - model_norm
+ - ema_norm
+ - masked_pct
+
+optimization:
+ max_update: 375300
+ lr: [ 0.0004 ]
+ debug_param_names: true
+ clip_norm: 4
+
+optimizer:
+ _name: composite
+ dynamic_groups: true
+ groups:
+ default:
+ lr_float: 4e-4
+ optimizer:
+ _name: adam
+ adam_betas: [0.9,0.95]
+ weight_decay: 0.05
+ lr_scheduler:
+ _name: cosine
+ warmup_updates: 50040
+
+lr_scheduler: pass_through
+
+model:
+ _name: data2vec_multi
+
+ ema_decay: 0.9998
+ ema_end_decay: 0.99995
+ ema_anneal_end_step: 150000
+ instance_norm_target_layer: true
+ layer_norm_target_layer: false
+ layer_norm_targets: true
+ end_of_block_targets: false
+
+ depth: 32
+ embed_dim: 1280
+ num_heads: 16
+
+ average_top_k_layers: 24
+ clone_batch: 16
+
+ norm_eps: 1e-6
+
+ min_target_var: 0
+ min_pred_var: 0
+
+ encoder_dropout: 0
+ post_mlp_drop: 0
+ attention_dropout: 0
+ activation_dropout: 0
+
+ supported_modality: IMAGE
+ cls_loss: 0.01
+
+ ema_encoder_only: false
+
+ modalities:
+ image:
+ inverse_mask: true
+ mask_prob: 0.75
+ mask_prob_adjust: 0.1
+ mask_length: 3
+ mask_noise_std: 0.01
+ prenet_depth: 0
+ ema_local_encoder: true
+ num_extra_tokens: 1
+ init_extra_token_zero: false
+ use_alibi_encoder: false
+ embed_dim: 1280
+ decoder:
+ decoder_dim: 1024
+ decoder_groups: 16
+ decoder_kernel: 5
+ decoder_layers: 3
+ input_dropout: 0
\ No newline at end of file
diff --git a/fairseq/examples/data2vec/config/v2/large_audio_only_task.yaml b/fairseq/examples/data2vec/config/v2/large_audio_only_task.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..3f615897215b8c7cee45caab00d69b3342c7ba20
--- /dev/null
+++ b/fairseq/examples/data2vec/config/v2/large_audio_only_task.yaml
@@ -0,0 +1,122 @@
+# @package _group_
+
+common:
+ fp16: true
+ log_format: json
+ log_interval: 200
+ tensorboard_logdir: tb
+ min_loss_scale: 1e-6
+ fp16_no_flatten_grads: true
+ user_dir: ${env:PWD}/examples/data2vec
+
+checkpoint:
+ save_interval: 1
+ save_interval_updates: 25000
+ keep_interval_updates: 1
+ no_epoch_checkpoints: true
+
+task:
+ _name: audio_pretraining
+ data: /fsx-wav2vec/abaevski/data/librivox/no_silence
+ max_sample_size: 320000
+ min_sample_size: 32000
+ normalize: true
+ precompute_mask_config: {}
+
+dataset:
+ num_workers: 8
+ max_tokens: 320000
+ skip_invalid_size_inputs_valid_test: true
+ validate_interval: 5
+ required_batch_size_multiple: 1
+ disable_validation: true
+
+distributed_training:
+ distributed_world_size: 48
+ ddp_backend: c10d
+
+criterion:
+ _name: model
+ log_keys:
+ - ema_decay
+ - target_var
+ - pred_var
+ - model_norm
+ - ema_norm
+ - masked_pct
+
+optimization:
+ max_update: 600000
+ debug_param_names: true
+ clip_norm: 1
+
+optimizer:
+ _name: composite
+ dynamic_groups: true
+ groups:
+ default:
+ lr_float: 0.0004
+ optimizer:
+ _name: adam
+ adam_betas: [0.9,0.98]
+ adam_eps: 1e-06
+ weight_decay: 0.01
+ lr_scheduler:
+ _name: cosine
+ warmup_updates: 10000
+
+lr_scheduler: pass_through
+
+model:
+ _name: data2vec_multi
+
+ loss_beta: 0
+ loss_scale: null
+
+ depth: 16
+ embed_dim: 1024
+ num_heads: 16
+
+ clone_batch: 12
+
+ ema_decay: 0.9997
+ ema_end_decay: 1
+ ema_anneal_end_step: 300000
+ ema_encoder_only: false
+
+ average_top_k_layers: 16
+ instance_norm_target_layer: true
+ layer_norm_target_layer: false
+ layer_norm_targets: false
+
+ layerdrop: 0
+ norm_eps: 1e-5
+
+ supported_modality: AUDIO
+
+ modalities:
+ audio:
+ feature_encoder_spec: '[(512, 10, 5)] + [(512, 3, 2)] * 4 + [(512,2,2)] + [(512,2,2)]'
+ conv_pos_depth: 5
+ conv_pos_width: 95
+ conv_pos_groups: 16
+ prenet_depth: 8
+ mask_prob: 0.55
+ mask_prob_adjust: 0.1
+ inverse_mask: false
+ mask_length: 5
+ mask_noise_std: 0.01
+ mask_dropout: 0
+ add_masks: false
+ ema_local_encoder: false
+ use_alibi_encoder: true
+ prenet_layerdrop: 0
+ prenet_dropout: 0.1
+ learned_alibi_scale: true
+ learned_alibi_scale_per_head: true
+ decoder:
+ input_dropout: 0.1
+ decoder_dim: 768
+ decoder_groups: 16
+ decoder_kernel: 7
+ decoder_layers: 4
diff --git a/fairseq/examples/data2vec/config/v2/large_images_only_task.yaml b/fairseq/examples/data2vec/config/v2/large_images_only_task.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..6b957fc129deee8cae03a20a4f336716123a45ef
--- /dev/null
+++ b/fairseq/examples/data2vec/config/v2/large_images_only_task.yaml
@@ -0,0 +1,120 @@
+# @package _group_
+
+common:
+ fp16: true
+ log_format: json
+ log_interval: 200
+ tensorboard_logdir: tb
+ min_loss_scale: 1e-6
+ fp16_no_flatten_grads: true
+ user_dir: ${env:PWD}/examples/data2vec
+
+checkpoint:
+ save_interval: 5
+ save_interval_updates: 25000
+ keep_interval_updates: 1
+ no_epoch_checkpoints: true
+
+task:
+ _name: mae_image_pretraining
+ data: /datasets01/imagenet_full_size/061417/
+ rebuild_batches: true
+ local_cache_path: /scratch/cache_abaevski/imagenet
+ key: source
+ precompute_mask_config: {}
+
+dataset:
+ num_workers: 10
+ batch_size: 8
+ skip_invalid_size_inputs_valid_test: true
+ required_batch_size_multiple: 1
+ disable_validation: true
+
+distributed_training:
+ distributed_world_size: 16
+ ddp_backend: c10d
+
+criterion:
+ _name: model
+ log_keys:
+ - ema_decay
+ - target_var
+ - pred_var
+ - model_norm
+ - ema_norm
+ - masked_pct
+
+optimization:
+ max_update: 375300
+ lr: [ 0.0004 ]
+ debug_param_names: true
+ clip_norm: 4
+
+optimizer:
+ _name: composite
+ dynamic_groups: true
+ groups:
+ default:
+ lr_float: 4e-4
+ optimizer:
+ _name: adam
+ adam_betas: [0.9,0.95]
+ weight_decay: 0.05
+ lr_scheduler:
+ _name: cosine
+ warmup_updates: 50040
+
+lr_scheduler: pass_through
+
+model:
+ _name: data2vec_multi
+
+ ema_decay: 0.9998
+ ema_end_decay: 0.99999
+ ema_anneal_end_step: 150000
+ instance_norm_target_layer: true
+ layer_norm_target_layer: false
+ layer_norm_targets: true
+ end_of_block_targets: false
+
+ depth: 24
+ embed_dim: 1024
+ num_heads: 16
+
+ average_top_k_layers: 18
+ clone_batch: 16
+
+ norm_eps: 1e-6
+
+ min_target_var: 0
+ min_pred_var: 0
+
+ encoder_dropout: 0
+ post_mlp_drop: 0
+ attention_dropout: 0
+ activation_dropout: 0
+
+ supported_modality: IMAGE
+ cls_loss: 0.01
+
+ ema_encoder_only: false
+
+ modalities:
+ image:
+ inverse_mask: true
+ mask_prob: 0.75
+ mask_prob_adjust: 0.1
+ mask_length: 3
+ mask_noise_std: 0.01
+ prenet_depth: 0
+ ema_local_encoder: true
+ num_extra_tokens: 1
+ init_extra_token_zero: false
+ use_alibi_encoder: false
+ embed_dim: 1024
+ decoder:
+ decoder_dim: 1024
+ decoder_groups: 16
+ decoder_kernel: 5
+ decoder_layers: 3
+ input_dropout: 0
\ No newline at end of file
diff --git a/fairseq/examples/data2vec/config/v2/large_text_only_task.yaml b/fairseq/examples/data2vec/config/v2/large_text_only_task.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fd69048e77d6febb0eee6e7fbe459069900f202c
--- /dev/null
+++ b/fairseq/examples/data2vec/config/v2/large_text_only_task.yaml
@@ -0,0 +1,112 @@
+# @package _group_
+
+common:
+ fp16: true
+ log_format: json
+ log_interval: 200
+ tensorboard_logdir: tb
+ min_loss_scale: 1e-6
+ fp16_no_flatten_grads: true
+ user_dir: ${env:PWD}/examples/data2vec
+
+checkpoint:
+ save_interval_updates: 50000
+ keep_interval_updates: 1
+ no_epoch_checkpoints: true
+
+task:
+ _name: masked_lm
+ data: /fsx-wav2vec/abaevski/data/nlp/bookwiki_aml-full-mmap2-bin
+ sample_break_mode: none
+ tokens_per_sample: 512
+ include_target_tokens: true
+ random_token_prob: 0
+ leave_unmasked_prob: 0
+ include_index: True
+ skip_masking: True
+ d2v2_multi: True
+
+dataset:
+ batch_size: 2
+ ignore_unused_valid_subsets: true
+ skip_invalid_size_inputs_valid_test: true
+ disable_validation: true
+
+distributed_training:
+ distributed_world_size: 32
+ ddp_backend: c10d
+
+criterion:
+ _name: model
+ log_keys:
+ - ema_decay
+ - target_var
+ - pred_var
+ - model_norm
+ - ema_norm
+ - masked_pct
+
+optimization:
+ max_update: 600000
+ clip_norm: 1
+
+optimizer:
+ _name: composite
+ dynamic_groups: true
+ groups:
+ default:
+ lr_float: 0.0001
+ optimizer:
+ _name: adam
+ adam_betas: [0.9,0.98]
+ adam_eps: 1e-06
+ weight_decay: 0.01
+ lr_scheduler:
+ _name: cosine
+ warmup_updates: 4000
+
+lr_scheduler: pass_through
+
+model:
+ _name: data2vec_multi
+
+ loss_beta: 0
+ loss_scale: 1
+
+ depth: 24
+ num_heads: 16
+ embed_dim: 1024
+ clone_batch: 8
+
+ ema_decay: 0.9999
+ ema_end_decay: 0.99999
+ ema_anneal_end_step: 100000
+ ema_encoder_only: true
+
+ average_top_k_layers: 24
+ layer_norm_target_layer: true
+ instance_norm_target_layer: false
+ batch_norm_target_layer: false
+ instance_norm_targets: true
+ layer_norm_targets: false
+
+ layerdrop: 0
+ norm_eps: 1e-5
+
+ supported_modality: TEXT
+
+ modalities:
+ text:
+ mask_prob: 0.5
+ mask_length: 1
+ mask_noise_std: 0.01
+ prenet_depth: 0
+ decoder:
+ input_dropout: 0.1
+ decoder_dim: 768
+ decoder_groups: 1
+ decoder_kernel: 9
+ decoder_layers: 5
+ decoder_residual: false
+ projection_layers: 2
+ projection_ratio: 2.0
diff --git a/fairseq/examples/data2vec/config/v2/large_text_only_task_pgrp_1M.yaml b/fairseq/examples/data2vec/config/v2/large_text_only_task_pgrp_1M.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..739e6f6724eefef493d6b46ad1b815d88431def9
--- /dev/null
+++ b/fairseq/examples/data2vec/config/v2/large_text_only_task_pgrp_1M.yaml
@@ -0,0 +1,123 @@
+# @package _group_
+
+common:
+ fp16: true
+ log_format: json
+ log_interval: 200
+ tensorboard_logdir: tb
+ fp16_no_flatten_grads: true
+ user_dir: ${env:PWD}/examples/data2vec
+
+checkpoint:
+ no_epoch_checkpoints: true
+ save_interval_updates: 50000
+ keep_interval_updates: 1
+
+distributed_training:
+ distributed_world_size: 32
+ ddp_backend: legacy_ddp
+
+task:
+ _name: masked_lm
+ data: /fsx-wav2vec/abaevski/data/nlp/bookwiki_aml-full-mmap2-bin
+ sample_break_mode: none
+ tokens_per_sample: 512
+ include_target_tokens: true
+ random_token_prob: 0
+ leave_unmasked_prob: 0
+ include_index: True
+ skip_masking: True
+ d2v2_multi: True
+
+criterion:
+ _name: model
+ log_keys:
+ - ema_decay
+ - target_var
+ - pred_var
+ - model_norm
+ - ema_norm
+ - masked_pct
+
+dataset:
+ batch_size: 2
+ ignore_unused_valid_subsets: true
+ skip_invalid_size_inputs_valid_test: true
+ disable_validation: true
+
+optimization:
+ clip_norm: 1
+ lr: [3e-4]
+ max_update: 1000000
+ update_freq: [1]
+
+optimizer:
+ _name: composite
+ groups:
+ default:
+ lr_float: 1e-4
+ optimizer:
+ _name: adam
+ adam_betas: [0.9,0.98]
+ adam_eps: 1e-06
+ weight_decay: 0.01
+ lr_scheduler:
+ _name: cosine
+ warmup_updates: 4000
+ decoder:
+ lr_float: 1e-4
+ optimizer:
+ _name: adam
+ adam_betas: [0.9,0.98]
+ adam_eps: 1e-06
+ weight_decay: 0.01
+ lr_scheduler:
+ _name: cosine
+ warmup_updates: 4000
+
+lr_scheduler: pass_through
+
+model:
+ _name: data2vec_multi
+
+ loss_beta: 4
+ loss_scale: 1
+
+ depth: 24
+ num_heads: 16
+ embed_dim: 1024
+ clone_batch: 8
+
+ ema_decay: 0.9999
+ ema_end_decay: 0.99999
+ ema_anneal_end_step: 100000
+ ema_encoder_only: true
+
+ average_top_k_layers: 24
+ layer_norm_target_layer: true
+ instance_norm_target_layer: false
+ batch_norm_target_layer: false
+ instance_norm_targets: true
+ layer_norm_targets: false
+
+ layerdrop: 0
+ norm_eps: 1e-5
+
+ supported_modality: TEXT
+ decoder_group: true
+
+ modalities:
+ text:
+ mask_prob: 0.5
+ mask_length: 1
+ mask_noise_std: 0.01
+ prenet_depth: 0
+ decoder:
+ input_dropout: 0.1
+ decoder_dim: 768
+ decoder_groups: 1
+ decoder_kernel: 9
+ decoder_layers: 5
+ decoder_residual: false
+ projection_layers: 2
+ projection_ratio: 2.0
diff --git a/fairseq/examples/data2vec/config/v2/run_config/local.yaml b/fairseq/examples/data2vec/config/v2/run_config/local.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..45595f9eea7c369a1200d802bfa1883c1bdfe573
--- /dev/null
+++ b/fairseq/examples/data2vec/config/v2/run_config/local.yaml
@@ -0,0 +1,15 @@
+# @package _global_
+hydra:
+ sweep:
+ dir: ${env:PWD}/tmp_dbg/${now:%H-%M-%S}
+
+distributed_training:
+ distributed_world_size: 1
+ nprocs_per_node: 1
+ distributed_port: -1
+
+common:
+ log_interval: 1
+
+dataset:
+ num_workers: 0
diff --git a/fairseq/examples/data2vec/config/v2/run_config/slurm_1.yaml b/fairseq/examples/data2vec/config/v2/run_config/slurm_1.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..732f0188993a14c847ab5d5e8addb24b0b536311
--- /dev/null
+++ b/fairseq/examples/data2vec/config/v2/run_config/slurm_1.yaml
@@ -0,0 +1,37 @@
+# @package _global_
+
+hydra:
+ job:
+ config:
+ override_dirname:
+ kv_sep: ':'
+ item_sep: '/'
+ exclude_keys:
+ - run_config
+ - distributed_training.distributed_port
+ - distributed_training.distributed_world_size
+ - model.pretrained_model_path
+ - model.target_network_path
+ - next_script
+ - task.cache_in_scratch
+ - task.data
+ - checkpoint.save_interval_updates
+ - checkpoint.keep_interval_updates
+ - checkpoint.save_on_overflow
+ - common.log_interval
+ - common.user_dir
+ sweep:
+ dir: /checkpoint/${env:USER}/${env:PREFIX}/${hydra.job.config_name}_${hydra.launcher.gpus_per_node}/${hydra.job.override_dirname}
+ subdir: ''
+ launcher:
+ submitit_folder: ${hydra.sweep.dir}
+ timeout_min: 4320
+ cpus_per_task: 80
+ gpus_per_node: 8
+ tasks_per_node: 1
+ mem_gb: 450
+ nodes: 1
+ name: ${env:PREFIX}_${hydra.job.config_name}
+ partition: devlab,learnlab,learnfair,scavenge
+ constraint: volta32gb,ib4
+ max_num_timeout: 30
diff --git a/fairseq/examples/data2vec/config/v2/run_config/slurm_1_aws.yaml b/fairseq/examples/data2vec/config/v2/run_config/slurm_1_aws.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..b2184f8cfa252f34ec71b965728d286ccd930eeb
--- /dev/null
+++ b/fairseq/examples/data2vec/config/v2/run_config/slurm_1_aws.yaml
@@ -0,0 +1,37 @@
+# @package _global_
+
+hydra:
+ job:
+ config:
+ override_dirname:
+ kv_sep: ':'
+ item_sep: '/'
+ exclude_keys:
+ - run_config
+ - distributed_training.distributed_port
+ - distributed_training.distributed_world_size
+ - model.pretrained_model_path
+ - model.target_network_path
+ - next_script
+ - task.cache_in_scratch
+ - task.local_cache_path
+ - task.data
+ - checkpoint.save_interval_updates
+ - checkpoint.keep_interval_updates
+ - checkpoint.save_on_overflow
+ - common.log_interval
+ - common.user_dir
+ sweep:
+ dir: /checkpoint/${env:USER}/${env:PREFIX}/${hydra.job.config_name}_${hydra.launcher.gpus_per_node}/${hydra.job.override_dirname}
+ subdir: ''
+ launcher:
+ submitit_folder: ${hydra.sweep.dir}
+ timeout_min: 4320
+ cpus_per_task: 80
+ gpus_per_node: 8
+ tasks_per_node: 1
+ mem_gb: 0
+ nodes: 1
+ name: ${env:PREFIX}_${hydra.job.config_name}
+ partition: wav2vec,learnlab,learnfair
+ max_num_timeout: 30
diff --git a/fairseq/examples/data2vec/config/v2/run_config/slurm_2.yaml b/fairseq/examples/data2vec/config/v2/run_config/slurm_2.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..ec53dc2a984babdf7e61930350868e38f0d769c3
--- /dev/null
+++ b/fairseq/examples/data2vec/config/v2/run_config/slurm_2.yaml
@@ -0,0 +1,37 @@
+# @package _global_
+
+hydra:
+ job:
+ config:
+ override_dirname:
+ kv_sep: ':'
+ item_sep: '/'
+ exclude_keys:
+ - run_config
+ - distributed_training.distributed_port
+ - distributed_training.distributed_world_size
+ - model.pretrained_model_path
+ - model.target_network_path
+ - next_script
+ - task.cache_in_scratch
+ - task.data
+ - checkpoint.save_interval_updates
+ - checkpoint.keep_interval_updates
+ - checkpoint.save_on_overflow
+ - common.log_interval
+ - common.user_dir
+ sweep:
+ dir: /checkpoint/${env:USER}/${env:PREFIX}/${hydra.job.config_name}_${hydra.launcher.gpus_per_node}/${hydra.job.override_dirname}
+ subdir: ''
+ launcher:
+ submitit_folder: ${hydra.sweep.dir}
+ timeout_min: 4320
+ cpus_per_task: 10
+ gpus_per_node: 8
+ tasks_per_node: 8
+ mem_gb: 450
+ nodes: 2
+ name: ${env:PREFIX}_${hydra.job.config_name}
+ partition: devlab,learnlab,learnfair,scavenge
+ constraint: volta32gb,ib4
+ max_num_timeout: 30
diff --git a/fairseq/examples/data2vec/config/v2/run_config/slurm_2_aws.yaml b/fairseq/examples/data2vec/config/v2/run_config/slurm_2_aws.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..553765597ff5437e9044d1bc0a5f1087438594be
--- /dev/null
+++ b/fairseq/examples/data2vec/config/v2/run_config/slurm_2_aws.yaml
@@ -0,0 +1,39 @@
+# @package _global_
+
+hydra:
+ job:
+ config:
+ override_dirname:
+ kv_sep: ':'
+ item_sep: '/'
+ exclude_keys:
+ - run_config
+ - distributed_training.distributed_port
+ - distributed_training.distributed_world_size
+ - model.pretrained_model_path
+ - model.target_network_path
+ - next_script
+ - task.cache_in_scratch
+ - task.local_cache_path
+ - task.data
+ - task.post_save_script
+ - checkpoint.save_interval_updates
+ - checkpoint.keep_interval_updates
+ - checkpoint.save_on_overflow
+ - common.log_interval
+ - common.user_dir
+ - model.model_path
+ sweep:
+ dir: /fsx-wav2vec/${env:USER}/${env:PREFIX}/${hydra.job.config_name}_${hydra.launcher.gpus_per_node}/${hydra.job.override_dirname}
+ subdir: ''
+ launcher:
+ submitit_folder: ${hydra.sweep.dir}
+ timeout_min: 4320
+ cpus_per_task: 12
+ gpus_per_node: 8
+ tasks_per_node: 8
+ mem_gb: 0
+ nodes: 2
+ name: ${env:PREFIX}_${hydra.job.config_name}
+ partition: wav2vec
+ max_num_timeout: 30
diff --git a/fairseq/examples/data2vec/config/v2/run_config/slurm_3.yaml b/fairseq/examples/data2vec/config/v2/run_config/slurm_3.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..14b47d14e6282d5a846cc7c41c6b2080c5ed8c62
--- /dev/null
+++ b/fairseq/examples/data2vec/config/v2/run_config/slurm_3.yaml
@@ -0,0 +1,36 @@
+# @package _global_
+
+hydra:
+ job:
+ config:
+ override_dirname:
+ kv_sep: ':'
+ item_sep: '/'
+ exclude_keys:
+ - run_config
+ - distributed_training.distributed_port
+ - distributed_training.distributed_world_size
+ - model.pretrained_model_path
+ - model.target_network_path
+ - next_script
+ - task.cache_in_scratch
+ - task.data
+ - checkpoint.save_interval_updates
+ - checkpoint.keep_interval_updates
+ - checkpoint.save_on_overflow
+ - common.log_interval
+ sweep:
+ dir: /checkpoint/${env:USER}/${env:PREFIX}/${hydra.job.config_name}_${hydra.launcher.gpus_per_node}/${hydra.job.override_dirname}
+ subdir: ''
+ launcher:
+ submitit_folder: ${hydra.sweep.dir}
+ timeout_min: 4320
+ cpus_per_task: 80
+ gpus_per_node: 8
+ tasks_per_node: 1
+ mem_gb: 450
+ nodes: 3
+ name: ${env:PREFIX}_${hydra.job.config_name}
+ partition: devlab,learnlab,learnfair,scavenge
+ constraint: volta32gb,ib4
+ max_num_timeout: 30
diff --git a/fairseq/examples/data2vec/config/v2/run_config/slurm_4.yaml b/fairseq/examples/data2vec/config/v2/run_config/slurm_4.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..c54d735fb2dc0b0af2d467caa7f64405af18ea10
--- /dev/null
+++ b/fairseq/examples/data2vec/config/v2/run_config/slurm_4.yaml
@@ -0,0 +1,36 @@
+# @package _global_
+
+hydra:
+ job:
+ config:
+ override_dirname:
+ kv_sep: ':'
+ item_sep: '/'
+ exclude_keys:
+ - run_config
+ - distributed_training.distributed_port
+ - distributed_training.distributed_world_size
+ - model.pretrained_model_path
+ - model.target_network_path
+ - next_script
+ - task.cache_in_scratch
+ - task.data
+ - checkpoint.save_interval_updates
+ - checkpoint.keep_interval_updates
+ - checkpoint.save_on_overflow
+ - common.log_interval
+ sweep:
+ dir: /checkpoint/${env:USER}/${env:PREFIX}/${hydra.job.config_name}_${hydra.launcher.gpus_per_node}/${hydra.job.override_dirname}
+ subdir: ''
+ launcher:
+ submitit_folder: ${hydra.sweep.dir}
+ timeout_min: 4320
+ cpus_per_task: 10
+ gpus_per_node: 8
+ tasks_per_node: 8
+ mem_gb: 450
+ nodes: 4
+ name: ${env:PREFIX}_${hydra.job.config_name}
+ partition: devlab,learnlab,learnfair,scavenge
+ constraint: volta32gb,ib4
+ max_num_timeout: 30
diff --git a/fairseq/examples/data2vec/config/v2/run_config/slurm_4_aws.yaml b/fairseq/examples/data2vec/config/v2/run_config/slurm_4_aws.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..a77f62aece140d09373c225504e2025c799e2ea1
--- /dev/null
+++ b/fairseq/examples/data2vec/config/v2/run_config/slurm_4_aws.yaml
@@ -0,0 +1,37 @@
+# @package _global_
+
+hydra:
+ job:
+ config:
+ override_dirname:
+ kv_sep: ':'
+ item_sep: '/'
+ exclude_keys:
+ - run_config
+ - distributed_training.distributed_port
+ - distributed_training.distributed_world_size
+ - model.pretrained_model_path
+ - model.target_network_path
+ - next_script
+ - task.cache_in_scratch
+ - task.data
+ - task.post_save_script
+ - checkpoint.save_interval_updates
+ - checkpoint.keep_interval_updates
+ - checkpoint.save_on_overflow
+ - common.log_interval
+ - common.user_dir
+ sweep:
+ dir: /fsx-wav2vec/${env:USER}/${env:PREFIX}/${hydra.job.config_name}_${hydra.launcher.gpus_per_node}/${hydra.job.override_dirname}
+ subdir: ''
+ launcher:
+ submitit_folder: ${hydra.sweep.dir}
+ timeout_min: 4320
+ cpus_per_task: 12
+ gpus_per_node: 8
+ tasks_per_node: 8
+ mem_gb: 0
+ nodes: 4
+ name: ${env:PREFIX}_${hydra.job.config_name}
+ partition: wav2vec
+ max_num_timeout: 30
diff --git a/fairseq/examples/data2vec/config/v2/run_config/slurm_6_aws.yaml b/fairseq/examples/data2vec/config/v2/run_config/slurm_6_aws.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..20e06582be5673de5b5c19fb996172c9ddac15b2
--- /dev/null
+++ b/fairseq/examples/data2vec/config/v2/run_config/slurm_6_aws.yaml
@@ -0,0 +1,36 @@
+# @package _global_
+
+hydra:
+ job:
+ config:
+ override_dirname:
+ kv_sep: ':'
+ item_sep: '/'
+ exclude_keys:
+ - run_config
+ - distributed_training.distributed_port
+ - distributed_training.distributed_world_size
+ - model.pretrained_model_path
+ - model.target_network_path
+ - next_script
+ - task.cache_in_scratch
+ - task.data
+ - checkpoint.save_interval_updates
+ - checkpoint.keep_interval_updates
+ - checkpoint.save_on_overflow
+ - common.log_interval
+ - common.user_dir
+ sweep:
+ dir: /fsx-wav2vec/${env:USER}/${env:PREFIX}/${hydra.job.config_name}_${hydra.launcher.gpus_per_node}/${hydra.job.override_dirname}
+ subdir: ''
+ launcher:
+ submitit_folder: ${hydra.sweep.dir}
+ timeout_min: 4320
+ cpus_per_task: 12
+ gpus_per_node: 8
+ tasks_per_node: 8
+ mem_gb: 0
+ nodes: 6
+ name: ${env:PREFIX}_${hydra.job.config_name}
+ partition: wav2vec,learnlab,learnfair
+ max_num_timeout: 30
diff --git a/fairseq/examples/data2vec/config/v2/run_config/slurm_8.yaml b/fairseq/examples/data2vec/config/v2/run_config/slurm_8.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..e3ec2c28475076f0c9a1e77c7affa00ff6ed4008
--- /dev/null
+++ b/fairseq/examples/data2vec/config/v2/run_config/slurm_8.yaml
@@ -0,0 +1,37 @@
+# @package _global_
+
+hydra:
+ job:
+ config:
+ override_dirname:
+ kv_sep: ':'
+ item_sep: '/'
+ exclude_keys:
+ - run_config
+ - distributed_training.distributed_port
+ - distributed_training.distributed_world_size
+ - model.pretrained_model_path
+ - model.target_network_path
+ - next_script
+ - task.cache_in_scratch
+ - task.data
+ - checkpoint.save_interval_updates
+ - checkpoint.keep_interval_updates
+ - checkpoint.save_on_overflow
+ - common.log_interval
+ - common.user_dir
+ sweep:
+ dir: /checkpoint/${env:USER}/${env:PREFIX}/${hydra.job.config_name}_${hydra.launcher.gpus_per_node}/${hydra.job.override_dirname}
+ subdir: ''
+ launcher:
+ submitit_folder: ${hydra.sweep.dir}
+ timeout_min: 4320
+ cpus_per_task: 10
+ gpus_per_node: 8
+ tasks_per_node: 8
+ mem_gb: 450
+ nodes: 8
+ name: ${env:PREFIX}_${hydra.job.config_name}
+ partition: devlab,learnlab,learnfair,scavenge
+ constraint: volta32gb,ib4
+ max_num_timeout: 30
diff --git a/fairseq/examples/data2vec/config/v2/run_config/slurm_8_aws.yaml b/fairseq/examples/data2vec/config/v2/run_config/slurm_8_aws.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..a9dce876cc9f5c26da77d8c8f35508cf3652c8d3
--- /dev/null
+++ b/fairseq/examples/data2vec/config/v2/run_config/slurm_8_aws.yaml
@@ -0,0 +1,36 @@
+# @package _global_
+
+hydra:
+ job:
+ config:
+ override_dirname:
+ kv_sep: ':'
+ item_sep: '/'
+ exclude_keys:
+ - run_config
+ - distributed_training.distributed_port
+ - distributed_training.distributed_world_size
+ - model.pretrained_model_path
+ - model.target_network_path
+ - next_script
+ - task.cache_in_scratch
+ - task.data
+ - checkpoint.save_interval_updates
+ - checkpoint.keep_interval_updates
+ - checkpoint.save_on_overflow
+ - common.log_interval
+ - common.user_dir
+ sweep:
+ dir: /fsx-wav2vec/${env:USER}/${env:PREFIX}/${hydra.job.config_name}_${hydra.launcher.gpus_per_node}/${hydra.job.override_dirname}
+ subdir: ''
+ launcher:
+ submitit_folder: ${hydra.sweep.dir}
+ timeout_min: 4320
+ cpus_per_task: 12
+ gpus_per_node: 8
+ tasks_per_node: 8
+ mem_gb: 0
+ nodes: 8
+ name: ${env:PREFIX}_${hydra.job.config_name}
+ partition: wav2vec,learnlab,learnfair
+ max_num_timeout: 30
diff --git a/fairseq/examples/data2vec/config/v2/text_finetuning/cola.yaml b/fairseq/examples/data2vec/config/v2/text_finetuning/cola.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d4ac4ec8b83107d280b2eb7d6b7ad718dc19803a
--- /dev/null
+++ b/fairseq/examples/data2vec/config/v2/text_finetuning/cola.yaml
@@ -0,0 +1,60 @@
+# @package _group_
+
+common:
+ fp16: true
+ fp16_init_scale: 4
+ threshold_loss_scale: 1
+ fp16_scale_window: 128
+ log_format: json
+ log_interval: 200
+ user_dir: ${env:PWD}/examples/data2vec
+
+task:
+ _name: sentence_prediction
+ data: ???
+ init_token: 0
+ separator_token: 2
+ num_classes: 2
+ max_positions: 512
+ d2v2_multi: True
+
+checkpoint:
+ best_checkpoint_metric: mcc
+ maximize_best_checkpoint_metric: true
+ no_epoch_checkpoints: true
+
+distributed_training:
+ find_unused_parameters: true
+ distributed_world_size: 1
+ nprocs_per_node: 1
+ distributed_port: -1
+
+criterion:
+ _name: sentence_prediction
+ report_mcc: True
+
+dataset:
+ batch_size: 16
+ required_batch_size_multiple: 1
+ max_tokens: 4400
+ num_workers: 1
+
+optimizer:
+ _name: adam
+ weight_decay: 0.1
+ adam_betas: (0.9,0.98)
+ adam_eps: 1e-06
+
+lr_scheduler:
+ _name: polynomial_decay
+ warmup_updates: 320
+
+optimization:
+ clip_norm: 0.0
+ lr: [2e-05]
+ max_update: 5336
+ max_epoch: 10
+
+model:
+ _name: data2vec_text_classification
+ model_path: ???
diff --git a/fairseq/examples/data2vec/config/v2/text_finetuning/mnli.yaml b/fairseq/examples/data2vec/config/v2/text_finetuning/mnli.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..1a9d6e52f072890273bb50b7b7ba9b6286c27f14
--- /dev/null
+++ b/fairseq/examples/data2vec/config/v2/text_finetuning/mnli.yaml
@@ -0,0 +1,60 @@
+# @package _group_
+
+common:
+ fp16: true
+ fp16_init_scale: 4
+ threshold_loss_scale: 1
+ fp16_scale_window: 128
+ log_format: json
+ log_interval: 200
+ user_dir: ${env:PWD}/examples/data2vec
+
+task:
+ _name: sentence_prediction
+ data: ???
+ init_token: 0
+ separator_token: 2
+ num_classes: 3
+ max_positions: 512
+ d2v2_multi: True
+
+checkpoint:
+ best_checkpoint_metric: accuracy
+ maximize_best_checkpoint_metric: true
+ no_epoch_checkpoints: true
+
+distributed_training:
+ find_unused_parameters: true
+ distributed_world_size: 1
+ nprocs_per_node: 1
+ distributed_port: -1
+
+criterion:
+ _name: sentence_prediction
+
+dataset:
+ batch_size: 32
+ required_batch_size_multiple: 1
+ max_tokens: 4400
+ valid_subset: valid,valid1
+ num_workers: 1
+
+optimizer:
+ _name: adam
+ weight_decay: 0.1
+ adam_betas: (0.9,0.98)
+ adam_eps: 1e-06
+
+lr_scheduler:
+ _name: polynomial_decay
+ warmup_updates: 7432
+
+optimization:
+ clip_norm: 0.0
+ lr: [2e-05]
+ max_update: 123873
+ max_epoch: 10
+
+model:
+ _name: data2vec_text_classification
+ model_path: ???
diff --git a/fairseq/examples/data2vec/config/v2/text_finetuning/mrpc.yaml b/fairseq/examples/data2vec/config/v2/text_finetuning/mrpc.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..8f93d9d9ea66fe6e341d3cd25749286bcc91e828
--- /dev/null
+++ b/fairseq/examples/data2vec/config/v2/text_finetuning/mrpc.yaml
@@ -0,0 +1,60 @@
+# @package _group_
+
+common:
+ fp16: true
+ fp16_init_scale: 4
+ threshold_loss_scale: 1
+ fp16_scale_window: 128
+ log_format: json
+ log_interval: 200
+ user_dir: ${env:PWD}/examples/data2vec
+
+task:
+ _name: sentence_prediction
+ data: ???
+ init_token: 0
+ separator_token: 2
+ num_classes: 2
+ max_positions: 512
+ d2v2_multi: True
+
+checkpoint:
+ best_checkpoint_metric: acc_and_f1
+ maximize_best_checkpoint_metric: true
+ no_epoch_checkpoints: true
+
+distributed_training:
+ find_unused_parameters: true
+ distributed_world_size: 1
+ nprocs_per_node: 1
+ distributed_port: -1
+
+criterion:
+ _name: sentence_prediction
+ report_acc_and_f1: True
+
+dataset:
+ batch_size: 16
+ required_batch_size_multiple: 1
+ max_tokens: 4400
+ num_workers: 1
+
+optimizer:
+ _name: adam
+ weight_decay: 0.1
+ adam_betas: (0.9,0.98)
+ adam_eps: 1e-06
+
+lr_scheduler:
+ _name: polynomial_decay
+ warmup_updates: 137
+
+optimization:
+ clip_norm: 0.0
+ lr: [2e-05]
+ max_update: 2296
+ max_epoch: 10
+
+model:
+ _name: data2vec_text_classification
+ model_path: ???
diff --git a/fairseq/examples/data2vec/config/v2/text_finetuning/qnli.yaml b/fairseq/examples/data2vec/config/v2/text_finetuning/qnli.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..739fb53b694287e02f6521e2fe330e5e0c4f455e
--- /dev/null
+++ b/fairseq/examples/data2vec/config/v2/text_finetuning/qnli.yaml
@@ -0,0 +1,59 @@
+# @package _group_
+
+common:
+ fp16: true
+ fp16_init_scale: 4
+ threshold_loss_scale: 1
+ fp16_scale_window: 128
+ log_format: json
+ log_interval: 200
+ user_dir: ${env:PWD}/examples/data2vec
+
+task:
+ _name: sentence_prediction
+ data: ???
+ init_token: 0
+ separator_token: 2
+ num_classes: 2
+ max_positions: 512
+ d2v2_multi: True
+
+checkpoint:
+ best_checkpoint_metric: accuracy
+ maximize_best_checkpoint_metric: true
+ no_epoch_checkpoints: true
+
+distributed_training:
+ find_unused_parameters: true
+ distributed_world_size: 1
+ nprocs_per_node: 1
+ distributed_port: -1
+
+criterion:
+ _name: sentence_prediction
+
+dataset:
+ batch_size: 32
+ required_batch_size_multiple: 1
+ max_tokens: 4400
+ num_workers: 1
+
+optimizer:
+ _name: adam
+ weight_decay: 0.1
+ adam_betas: (0.9,0.98)
+ adam_eps: 1e-06
+
+lr_scheduler:
+ _name: polynomial_decay
+ warmup_updates: 1986
+
+optimization:
+ clip_norm: 0.0
+ lr: [2e-05]
+ max_update: 33112
+ max_epoch: 10
+
+model:
+ _name: data2vec_text_classification
+ model_path: ???
diff --git a/fairseq/examples/data2vec/config/v2/text_finetuning/qqp.yaml b/fairseq/examples/data2vec/config/v2/text_finetuning/qqp.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..9accbaa5210bc7105d7eb58f5982ae6cb5826e1c
--- /dev/null
+++ b/fairseq/examples/data2vec/config/v2/text_finetuning/qqp.yaml
@@ -0,0 +1,60 @@
+# @package _group_
+
+common:
+ fp16: true
+ fp16_init_scale: 4
+ threshold_loss_scale: 1
+ fp16_scale_window: 128
+ log_format: json
+ log_interval: 200
+ user_dir: ${env:PWD}/examples/data2vec
+
+task:
+ _name: sentence_prediction
+ data: ???
+ init_token: 0
+ separator_token: 2
+ num_classes: 2
+ max_positions: 512
+ d2v2_multi: True
+
+checkpoint:
+ best_checkpoint_metric: acc_and_f1
+ maximize_best_checkpoint_metric: true
+ no_epoch_checkpoints: true
+
+distributed_training:
+ find_unused_parameters: true
+ distributed_world_size: 1
+ nprocs_per_node: 1
+ distributed_port: -1
+
+criterion:
+ _name: sentence_prediction
+ report_acc_and_f1: True
+
+dataset:
+ batch_size: 32
+ required_batch_size_multiple: 1
+ max_tokens: 4400
+ num_workers: 1
+
+optimizer:
+ _name: adam
+ weight_decay: 0.1
+ adam_betas: (0.9,0.98)
+ adam_eps: 1e-06
+
+lr_scheduler:
+ _name: polynomial_decay
+ warmup_updates: 28318
+
+optimization:
+ clip_norm: 0.0
+ lr: [2e-05]
+ max_update: 113272
+ max_epoch: 10
+
+model:
+ _name: data2vec_text_classification
+ model_path: ???
diff --git a/fairseq/examples/data2vec/config/v2/text_finetuning/rte.yaml b/fairseq/examples/data2vec/config/v2/text_finetuning/rte.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..ea07764d9835c9dd22bde49ee7152a4a78fa87ba
--- /dev/null
+++ b/fairseq/examples/data2vec/config/v2/text_finetuning/rte.yaml
@@ -0,0 +1,59 @@
+# @package _group_
+
+common:
+ fp16: true
+ fp16_init_scale: 4
+ threshold_loss_scale: 1
+ fp16_scale_window: 128
+ log_format: json
+ log_interval: 200
+ user_dir: ${env:PWD}/examples/data2vec
+
+task:
+ _name: sentence_prediction
+ data: ???
+ init_token: 0
+ separator_token: 2
+ num_classes: 2
+ max_positions: 512
+ d2v2_multi: True
+
+checkpoint:
+ best_checkpoint_metric: accuracy
+ maximize_best_checkpoint_metric: true
+ no_epoch_checkpoints: true
+
+distributed_training:
+ find_unused_parameters: true
+ distributed_world_size: 1
+ nprocs_per_node: 1
+ distributed_port: -1
+
+criterion:
+ _name: sentence_prediction
+
+dataset:
+ batch_size: 16
+ required_batch_size_multiple: 1
+ max_tokens: 4400
+ num_workers: 1
+
+optimizer:
+ _name: adam
+ weight_decay: 0.1
+ adam_betas: (0.9,0.98)
+ adam_eps: 1e-06
+
+lr_scheduler:
+ _name: polynomial_decay
+ warmup_updates: 122
+
+optimization:
+ clip_norm: 0.0
+ lr: [2e-05]
+ max_update: 2036
+ max_epoch: 10
+
+model:
+ _name: data2vec_text_classification
+ model_path: ???
diff --git a/fairseq/examples/data2vec/config/v2/text_finetuning/run_config/local.yaml b/fairseq/examples/data2vec/config/v2/text_finetuning/run_config/local.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..45595f9eea7c369a1200d802bfa1883c1bdfe573
--- /dev/null
+++ b/fairseq/examples/data2vec/config/v2/text_finetuning/run_config/local.yaml
@@ -0,0 +1,15 @@
+# @package _global_
+hydra:
+ sweep:
+ dir: ${env:PWD}/tmp_dbg/${now:%H-%M-%S}
+
+distributed_training:
+ distributed_world_size: 1
+ nprocs_per_node: 1
+ distributed_port: -1
+
+common:
+ log_interval: 1
+
+dataset:
+ num_workers: 0
diff --git a/fairseq/examples/data2vec/config/v2/text_finetuning/sst_2.yaml b/fairseq/examples/data2vec/config/v2/text_finetuning/sst_2.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..a273e5b9434cabd732d8a37b9fc8063ac8399d8f
--- /dev/null
+++ b/fairseq/examples/data2vec/config/v2/text_finetuning/sst_2.yaml
@@ -0,0 +1,59 @@
+# @package _group_
+
+common:
+ fp16: true
+ fp16_init_scale: 4
+ threshold_loss_scale: 1
+ fp16_scale_window: 128
+ log_format: json
+ log_interval: 200
+ user_dir: ${env:PWD}/examples/data2vec
+
+task:
+ _name: sentence_prediction
+ data: ???
+ init_token: 0
+ separator_token: 2
+ num_classes: 2
+ max_positions: 512
+ d2v2_multi: True
+
+checkpoint:
+ best_checkpoint_metric: accuracy
+ maximize_best_checkpoint_metric: true
+ no_epoch_checkpoints: true
+
+distributed_training:
+ find_unused_parameters: true
+ distributed_world_size: 1
+ nprocs_per_node: 1
+ distributed_port: -1
+
+criterion:
+ _name: sentence_prediction
+
+dataset:
+ batch_size: 32
+ required_batch_size_multiple: 1
+ max_tokens: 4400
+ num_workers: 1
+
+optimizer:
+ _name: adam
+ weight_decay: 0.1
+ adam_betas: (0.9,0.98)
+ adam_eps: 1e-06
+
+lr_scheduler:
+ _name: polynomial_decay
+ warmup_updates: 1256
+
+optimization:
+ clip_norm: 0.0
+ lr: [2e-05]
+ max_update: 20935
+ max_epoch: 10
+
+model:
+ _name: data2vec_text_classification
+ model_path: ???
diff --git a/fairseq/examples/data2vec/config/v2/text_finetuning/sts_b.yaml b/fairseq/examples/data2vec/config/v2/text_finetuning/sts_b.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fb009ab95b53fd88342ea75b3e5e1eab344b0679
--- /dev/null
+++ b/fairseq/examples/data2vec/config/v2/text_finetuning/sts_b.yaml
@@ -0,0 +1,61 @@
+# @package _group_
+
+common:
+ fp16: true
+ fp16_init_scale: 4
+ threshold_loss_scale: 1
+ fp16_scale_window: 128
+ log_format: json
+ log_interval: 200
+ user_dir: ${env:PWD}/examples/data2vec
+
+task:
+ _name: sentence_prediction
+ data: ???
+ init_token: 0
+ separator_token: 2
+ num_classes: 1
+ max_positions: 512
+ d2v2_multi: True
+
+checkpoint:
+ best_checkpoint_metric: pearson_and_spearman
+ maximize_best_checkpoint_metric: true
+ no_epoch_checkpoints: true
+
+distributed_training:
+ find_unused_parameters: true
+ distributed_world_size: 1
+ nprocs_per_node: 1
+ distributed_port: -1
+
+criterion:
+ _name: sentence_prediction
+ regression_target: true
+ report_pearson_and_spearman: True
+
+dataset:
+ batch_size: 16
+ required_batch_size_multiple: 1
+ max_tokens: 4400
+ num_workers: 1
+
+optimizer:
+ _name: adam
+ weight_decay: 0.1
+ adam_betas: (0.9,0.98)
+ adam_eps: 1e-06
+
+lr_scheduler:
+ _name: polynomial_decay
+ warmup_updates: 214
+
+optimization:
+ clip_norm: 0.0
+ lr: [4e-05]
+ max_update: 3598
+ max_epoch: 10
+
+model:
+ _name: data2vec_text_classification
+ model_path: ???
diff --git a/fairseq/examples/data2vec/config/vision/finetuning/imagenet.yaml b/fairseq/examples/data2vec/config/vision/finetuning/imagenet.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d6d4864cca01098e331c6eeb8c174851bdd8bc6c
--- /dev/null
+++ b/fairseq/examples/data2vec/config/vision/finetuning/imagenet.yaml
@@ -0,0 +1,52 @@
+# @package _group_
+
+common:
+ fp16: true
+ log_format: json
+ log_interval: 200
+ tensorboard_logdir: tb
+
+checkpoint:
+ save_interval: 1
+ save_interval_updates: 25000
+ keep_interval_updates: 1
+ no_epoch_checkpoints: true
+ best_checkpoint_metric: accuracy
+
+task:
+ _name: image_classification
+ data: /datasets01/imagenet_full_size/061417
+
+dataset:
+ num_workers: 6
+ batch_size: 64
+ skip_invalid_size_inputs_valid_test: true
+ required_batch_size_multiple: 1
+ valid_subset: val
+
+distributed_training:
+ distributed_world_size: 8
+ ddp_backend: c10d
+
+criterion:
+ _name: model
+ log_keys:
+ - correct
+
+optimization:
+ max_update: 100000
+ lr: [0.0005]
+
+optimizer:
+ _name: adam
+ adam_betas: (0.9,0.98)
+ adam_eps: 1e-06
+ weight_decay: 0.01
+
+lr_scheduler:
+ _name: cosine
+ warmup_updates: 10000
+
+model:
+ _name: data2vec_image_classification
+ model_path: ???
diff --git a/fairseq/examples/data2vec/config/vision/finetuning/mae_imagenet_clean.yaml b/fairseq/examples/data2vec/config/vision/finetuning/mae_imagenet_clean.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..17d4c0a8f57140ee2e24555c46aff875039b5d39
--- /dev/null
+++ b/fairseq/examples/data2vec/config/vision/finetuning/mae_imagenet_clean.yaml
@@ -0,0 +1,65 @@
+# @package _group_
+
+common:
+ fp16: true
+ log_format: json
+ log_interval: 200
+ tensorboard_logdir: tb
+ fp16_no_flatten_grads: true
+
+checkpoint:
+ save_interval: 1
+ save_interval_updates: 25000
+ keep_interval_updates: 1
+ no_epoch_checkpoints: true
+ best_checkpoint_metric: accuracy
+ maximize_best_checkpoint_metric: true
+
+task:
+ _name: mae_image_classification
+ data: /datasets01/imagenet_full_size/061417
+
+dataset:
+ num_workers: 6
+ batch_size: 32
+ skip_invalid_size_inputs_valid_test: true
+ required_batch_size_multiple: 2
+ valid_subset: val
+
+distributed_training:
+ distributed_world_size: 16
+ ddp_backend: c10d
+
+criterion:
+ _name: model
+ log_keys:
+ - correct
+
+optimization:
+ max_update: 250200
+ lr: [0.001]
+
+optimizer:
+ _name: composite
+ dynamic_groups: true
+ groups:
+ default:
+ lr_float: 0.001
+ optimizer:
+ _name: adam
+ adam_betas: [0.9,0.95]
+ weight_decay: 0.05
+ lr_scheduler:
+ _name: cosine
+ warmup_updates: 16000
+ min_lr: 1e-6
+
+
+lr_scheduler: pass_through
+
+model:
+ _name: mae_image_classification
+ mixup: 0.7
+ mixup_prob: 0.9
+
+ model_path: ???
diff --git a/fairseq/examples/data2vec/config/vision/finetuning/mae_imagenet_huge_clean.yaml b/fairseq/examples/data2vec/config/vision/finetuning/mae_imagenet_huge_clean.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..2d2eb57bac1a89bc705f21dfa9768cfbfb14e26c
--- /dev/null
+++ b/fairseq/examples/data2vec/config/vision/finetuning/mae_imagenet_huge_clean.yaml
@@ -0,0 +1,68 @@
+# @package _group_
+
+common:
+ fp16: true
+ log_format: json
+ log_interval: 200
+ tensorboard_logdir: tb
+ fp16_no_flatten_grads: true
+
+checkpoint:
+ save_interval: 1
+ save_interval_updates: 25000
+ keep_interval_updates: 1
+ no_epoch_checkpoints: true
+ best_checkpoint_metric: accuracy
+ maximize_best_checkpoint_metric: true
+
+task:
+ _name: mae_image_classification
+ data: /datasets01/imagenet_full_size/061417
+
+dataset:
+ num_workers: 6
+ batch_size: 32
+ skip_invalid_size_inputs_valid_test: true
+ required_batch_size_multiple: 2
+ valid_subset: val
+
+distributed_training:
+ distributed_world_size: 16
+ ddp_backend: c10d
+
+criterion:
+ _name: model
+ log_keys:
+ - correct
+
+optimization:
+ max_update: 125200
+ lr: [0.0005]
+ clip_norm: 4
+
+optimizer:
+ _name: composite
+ dynamic_groups: true
+ groups:
+ default:
+ lr_float: 0.0005
+ optimizer:
+ _name: adam
+ adam_betas: [0.9,0.95]
+ weight_decay: 0.05
+ lr_scheduler:
+ _name: cosine
+ warmup_updates: 16000
+ min_lr: 1e-20
+
+
+lr_scheduler: pass_through
+
+model:
+ _name: mae_image_classification
+ mixup: 0.7
+ mixup_prob: 0.9
+ layer_decay: 0.75
+ drop_path_rate: 0.2
+
+ model_path: ???
diff --git a/fairseq/examples/data2vec/config/vision/finetuning/mae_imagenet_large_clean.yaml b/fairseq/examples/data2vec/config/vision/finetuning/mae_imagenet_large_clean.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..3a9413cef6a41e09b541c7f88b5bbd785bba2289
--- /dev/null
+++ b/fairseq/examples/data2vec/config/vision/finetuning/mae_imagenet_large_clean.yaml
@@ -0,0 +1,68 @@
+# @package _group_
+
+common:
+ fp16: true
+ log_format: json
+ log_interval: 200
+ tensorboard_logdir: tb
+ fp16_no_flatten_grads: true
+
+checkpoint:
+ save_interval: 1
+ save_interval_updates: 25000
+ keep_interval_updates: 1
+ no_epoch_checkpoints: true
+ best_checkpoint_metric: accuracy
+ maximize_best_checkpoint_metric: true
+
+task:
+ _name: mae_image_classification
+ data: /datasets01/imagenet_full_size/061417
+
+dataset:
+ num_workers: 6
+ batch_size: 32
+ skip_invalid_size_inputs_valid_test: true
+ required_batch_size_multiple: 2
+ valid_subset: val
+
+distributed_training:
+ distributed_world_size: 16
+ ddp_backend: c10d
+
+criterion:
+ _name: model
+ log_keys:
+ - correct
+
+optimization:
+ max_update: 125200
+ lr: [0.0005]
+ clip_norm: 4
+
+optimizer:
+ _name: composite
+ dynamic_groups: true
+ groups:
+ default:
+ lr_float: 0.0005
+ optimizer:
+ _name: adam
+ adam_betas: [0.9,0.95]
+ weight_decay: 0.05
+ lr_scheduler:
+ _name: cosine
+ warmup_updates: 16000
+ min_lr: 1e-7
+
+
+lr_scheduler: pass_through
+
+model:
+ _name: mae_image_classification
+ mixup: 0.7
+ mixup_prob: 0.9
+ layer_decay: 0.75
+ drop_path_rate: 0.2
+
+ model_path: ???
diff --git a/fairseq/examples/data2vec/config/vision/finetuning/run_config/local.yaml b/fairseq/examples/data2vec/config/vision/finetuning/run_config/local.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..45595f9eea7c369a1200d802bfa1883c1bdfe573
--- /dev/null
+++ b/fairseq/examples/data2vec/config/vision/finetuning/run_config/local.yaml
@@ -0,0 +1,15 @@
+# @package _global_
+hydra:
+ sweep:
+ dir: ${env:PWD}/tmp_dbg/${now:%H-%M-%S}
+
+distributed_training:
+ distributed_world_size: 1
+ nprocs_per_node: 1
+ distributed_port: -1
+
+common:
+ log_interval: 1
+
+dataset:
+ num_workers: 0
diff --git a/fairseq/examples/data2vec/config/vision/finetuning/run_config/slurm_1.yaml b/fairseq/examples/data2vec/config/vision/finetuning/run_config/slurm_1.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..732f0188993a14c847ab5d5e8addb24b0b536311
--- /dev/null
+++ b/fairseq/examples/data2vec/config/vision/finetuning/run_config/slurm_1.yaml
@@ -0,0 +1,37 @@
+# @package _global_
+
+hydra:
+ job:
+ config:
+ override_dirname:
+ kv_sep: ':'
+ item_sep: '/'
+ exclude_keys:
+ - run_config
+ - distributed_training.distributed_port
+ - distributed_training.distributed_world_size
+ - model.pretrained_model_path
+ - model.target_network_path
+ - next_script
+ - task.cache_in_scratch
+ - task.data
+ - checkpoint.save_interval_updates
+ - checkpoint.keep_interval_updates
+ - checkpoint.save_on_overflow
+ - common.log_interval
+ - common.user_dir
+ sweep:
+ dir: /checkpoint/${env:USER}/${env:PREFIX}/${hydra.job.config_name}_${hydra.launcher.gpus_per_node}/${hydra.job.override_dirname}
+ subdir: ''
+ launcher:
+ submitit_folder: ${hydra.sweep.dir}
+ timeout_min: 4320
+ cpus_per_task: 80
+ gpus_per_node: 8
+ tasks_per_node: 1
+ mem_gb: 450
+ nodes: 1
+ name: ${env:PREFIX}_${hydra.job.config_name}
+ partition: devlab,learnlab,learnfair,scavenge
+ constraint: volta32gb,ib4
+ max_num_timeout: 30
diff --git a/fairseq/examples/data2vec/config/vision/finetuning/run_config/slurm_1_aws.yaml b/fairseq/examples/data2vec/config/vision/finetuning/run_config/slurm_1_aws.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..e2bab5675a561a6a01b5db06530b119a282bbf4a
--- /dev/null
+++ b/fairseq/examples/data2vec/config/vision/finetuning/run_config/slurm_1_aws.yaml
@@ -0,0 +1,36 @@
+# @package _global_
+
+hydra:
+ job:
+ config:
+ override_dirname:
+ kv_sep: ':'
+ item_sep: '/'
+ exclude_keys:
+ - run_config
+ - distributed_training.distributed_port
+ - distributed_training.distributed_world_size
+ - model.pretrained_model_path
+ - model.target_network_path
+ - next_script
+ - task.cache_in_scratch
+ - task.data
+ - checkpoint.save_interval_updates
+ - checkpoint.keep_interval_updates
+ - checkpoint.save_on_overflow
+ - common.log_interval
+ - common.user_dir
+ sweep:
+ dir: /checkpoint/${env:USER}/${env:PREFIX}/${hydra.job.config_name}_${hydra.launcher.gpus_per_node}/${hydra.job.override_dirname}
+ subdir: ''
+ launcher:
+ submitit_folder: ${hydra.sweep.dir}
+ timeout_min: 4320
+ cpus_per_task: 80
+ gpus_per_node: 8
+ tasks_per_node: 1
+ mem_gb: 0
+ nodes: 1
+ name: ${env:PREFIX}_${hydra.job.config_name}
+ partition: wav2vec,learnlab,learnfair
+ max_num_timeout: 30
diff --git a/fairseq/examples/data2vec/config/vision/finetuning/run_config/slurm_2.yaml b/fairseq/examples/data2vec/config/vision/finetuning/run_config/slurm_2.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..c8b0f02a9b7795d900804c557cae6576153bbaf0
--- /dev/null
+++ b/fairseq/examples/data2vec/config/vision/finetuning/run_config/slurm_2.yaml
@@ -0,0 +1,38 @@
+# @package _global_
+
+hydra:
+ job:
+ config:
+ override_dirname:
+ kv_sep: ':'
+ item_sep: '/'
+ exclude_keys:
+ - run_config
+ - distributed_training.distributed_port
+ - distributed_training.distributed_world_size
+ - model.pretrained_model_path
+ - model.target_network_path
+ - next_script
+ - task.cache_in_scratch
+ - task.data
+ - checkpoint.save_interval_updates
+ - checkpoint.keep_interval_updates
+ - checkpoint.save_on_overflow
+ - common.log_interval
+ - common.user_dir
+ - task.local_cache_path
+ sweep:
+ dir: /checkpoint/${env:USER}/${env:PREFIX}/${hydra.job.config_name}_${hydra.launcher.gpus_per_node}/${hydra.job.override_dirname}
+ subdir: ''
+ launcher:
+ submitit_folder: ${hydra.sweep.dir}
+ timeout_min: 4320
+ cpus_per_task: 10
+ gpus_per_node: 8
+ tasks_per_node: 8
+ mem_gb: 450
+ nodes: 2
+ name: ${env:PREFIX}_${hydra.job.config_name}
+ partition: devlab,learnlab,learnfair,scavenge
+ constraint: volta32gb,ib4
+ max_num_timeout: 30
diff --git a/fairseq/examples/data2vec/config/vision/finetuning/run_config/slurm_2_aws.yaml b/fairseq/examples/data2vec/config/vision/finetuning/run_config/slurm_2_aws.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..93d0d9c20ab007eefb2afec2f3c950dcecb44487
--- /dev/null
+++ b/fairseq/examples/data2vec/config/vision/finetuning/run_config/slurm_2_aws.yaml
@@ -0,0 +1,38 @@
+# @package _global_
+
+hydra:
+ job:
+ config:
+ override_dirname:
+ kv_sep: ':'
+ item_sep: '/'
+ exclude_keys:
+ - run_config
+ - distributed_training.distributed_port
+ - distributed_training.distributed_world_size
+ - model.pretrained_model_path
+ - model.target_network_path
+ - next_script
+ - task.cache_in_scratch
+ - task.data
+ - checkpoint.save_interval_updates
+ - checkpoint.keep_interval_updates
+ - checkpoint.save_on_overflow
+ - common.log_interval
+ - common.user_dir
+ - task.local_cache_path
+ - model.model_path
+ sweep:
+ dir: /fsx-wav2vec/${env:USER}/${env:PREFIX}/${hydra.job.config_name}_${hydra.launcher.gpus_per_node}/${hydra.job.override_dirname}
+ subdir: ''
+ launcher:
+ submitit_folder: ${hydra.sweep.dir}
+ timeout_min: 4320
+ cpus_per_task: 10
+ gpus_per_node: 8
+ tasks_per_node: 8
+ mem_gb: 0
+ nodes: 2
+ name: ${env:PREFIX}_${hydra.job.config_name}
+ partition: wav2vec,learnlab,learnfair
+ max_num_timeout: 30
diff --git a/fairseq/examples/data2vec/config/vision/finetuning/run_config/slurm_3.yaml b/fairseq/examples/data2vec/config/vision/finetuning/run_config/slurm_3.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..14b47d14e6282d5a846cc7c41c6b2080c5ed8c62
--- /dev/null
+++ b/fairseq/examples/data2vec/config/vision/finetuning/run_config/slurm_3.yaml
@@ -0,0 +1,36 @@
+# @package _global_
+
+hydra:
+ job:
+ config:
+ override_dirname:
+ kv_sep: ':'
+ item_sep: '/'
+ exclude_keys:
+ - run_config
+ - distributed_training.distributed_port
+ - distributed_training.distributed_world_size
+ - model.pretrained_model_path
+ - model.target_network_path
+ - next_script
+ - task.cache_in_scratch
+ - task.data
+ - checkpoint.save_interval_updates
+ - checkpoint.keep_interval_updates
+ - checkpoint.save_on_overflow
+ - common.log_interval
+ sweep:
+ dir: /checkpoint/${env:USER}/${env:PREFIX}/${hydra.job.config_name}_${hydra.launcher.gpus_per_node}/${hydra.job.override_dirname}
+ subdir: ''
+ launcher:
+ submitit_folder: ${hydra.sweep.dir}
+ timeout_min: 4320
+ cpus_per_task: 80
+ gpus_per_node: 8
+ tasks_per_node: 1
+ mem_gb: 450
+ nodes: 3
+ name: ${env:PREFIX}_${hydra.job.config_name}
+ partition: devlab,learnlab,learnfair,scavenge
+ constraint: volta32gb,ib4
+ max_num_timeout: 30
diff --git a/fairseq/examples/data2vec/config/vision/finetuning/run_config/slurm_4.yaml b/fairseq/examples/data2vec/config/vision/finetuning/run_config/slurm_4.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..c54d735fb2dc0b0af2d467caa7f64405af18ea10
--- /dev/null
+++ b/fairseq/examples/data2vec/config/vision/finetuning/run_config/slurm_4.yaml
@@ -0,0 +1,36 @@
+# @package _global_
+
+hydra:
+ job:
+ config:
+ override_dirname:
+ kv_sep: ':'
+ item_sep: '/'
+ exclude_keys:
+ - run_config
+ - distributed_training.distributed_port
+ - distributed_training.distributed_world_size
+ - model.pretrained_model_path
+ - model.target_network_path
+ - next_script
+ - task.cache_in_scratch
+ - task.data
+ - checkpoint.save_interval_updates
+ - checkpoint.keep_interval_updates
+ - checkpoint.save_on_overflow
+ - common.log_interval
+ sweep:
+ dir: /checkpoint/${env:USER}/${env:PREFIX}/${hydra.job.config_name}_${hydra.launcher.gpus_per_node}/${hydra.job.override_dirname}
+ subdir: ''
+ launcher:
+ submitit_folder: ${hydra.sweep.dir}
+ timeout_min: 4320
+ cpus_per_task: 10
+ gpus_per_node: 8
+ tasks_per_node: 8
+ mem_gb: 450
+ nodes: 4
+ name: ${env:PREFIX}_${hydra.job.config_name}
+ partition: devlab,learnlab,learnfair,scavenge
+ constraint: volta32gb,ib4
+ max_num_timeout: 30
diff --git a/fairseq/examples/data2vec/config/vision/finetuning/run_config/slurm_4_aws.yaml b/fairseq/examples/data2vec/config/vision/finetuning/run_config/slurm_4_aws.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d5d11cb755dcdafdc23ba8c0c44961323c274635
--- /dev/null
+++ b/fairseq/examples/data2vec/config/vision/finetuning/run_config/slurm_4_aws.yaml
@@ -0,0 +1,36 @@
+# @package _global_
+
+hydra:
+ job:
+ config:
+ override_dirname:
+ kv_sep: ':'
+ item_sep: '/'
+ exclude_keys:
+ - run_config
+ - distributed_training.distributed_port
+ - distributed_training.distributed_world_size
+ - model.pretrained_model_path
+ - model.target_network_path
+ - next_script
+ - task.cache_in_scratch
+ - task.data
+ - checkpoint.save_interval_updates
+ - checkpoint.keep_interval_updates
+ - checkpoint.save_on_overflow
+ - common.log_interval
+ - common.user_dir
+ sweep:
+ dir: /checkpoint/${env:USER}/${env:PREFIX}/${hydra.job.config_name}_${hydra.launcher.gpus_per_node}/${hydra.job.override_dirname}
+ subdir: ''
+ launcher:
+ submitit_folder: ${hydra.sweep.dir}
+ timeout_min: 4320
+ cpus_per_task: 10
+ gpus_per_node: 8
+ tasks_per_node: 8
+ mem_gb: 0
+ nodes: 4
+ name: ${env:PREFIX}_${hydra.job.config_name}
+ partition: wav2vec,learnlab,learnfair
+ max_num_timeout: 30
diff --git a/fairseq/examples/data2vec/config/vision/finetuning/run_config/slurm_6_aws.yaml b/fairseq/examples/data2vec/config/vision/finetuning/run_config/slurm_6_aws.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..906f08a602f498a4ce5ee24979861fab88ff8e96
--- /dev/null
+++ b/fairseq/examples/data2vec/config/vision/finetuning/run_config/slurm_6_aws.yaml
@@ -0,0 +1,36 @@
+# @package _global_
+
+hydra:
+ job:
+ config:
+ override_dirname:
+ kv_sep: ':'
+ item_sep: '/'
+ exclude_keys:
+ - run_config
+ - distributed_training.distributed_port
+ - distributed_training.distributed_world_size
+ - model.pretrained_model_path
+ - model.target_network_path
+ - next_script
+ - task.cache_in_scratch
+ - task.data
+ - checkpoint.save_interval_updates
+ - checkpoint.keep_interval_updates
+ - checkpoint.save_on_overflow
+ - common.log_interval
+ - common.user_dir
+ sweep:
+ dir: /checkpoint/${env:USER}/${env:PREFIX}/${hydra.job.config_name}_${hydra.launcher.gpus_per_node}/${hydra.job.override_dirname}
+ subdir: ''
+ launcher:
+ submitit_folder: ${hydra.sweep.dir}
+ timeout_min: 4320
+ cpus_per_task: 10
+ gpus_per_node: 8
+ tasks_per_node: 8
+ mem_gb: 0
+ nodes: 6
+ name: ${env:PREFIX}_${hydra.job.config_name}
+ partition: wav2vec,learnlab,learnfair
+ max_num_timeout: 30
diff --git a/fairseq/examples/data2vec/config/vision/finetuning/run_config/slurm_8_aws.yaml b/fairseq/examples/data2vec/config/vision/finetuning/run_config/slurm_8_aws.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d60e13f8ba6541267e7b09eb9d6bb61fb9c618d1
--- /dev/null
+++ b/fairseq/examples/data2vec/config/vision/finetuning/run_config/slurm_8_aws.yaml
@@ -0,0 +1,36 @@
+# @package _global_
+
+hydra:
+ job:
+ config:
+ override_dirname:
+ kv_sep: ':'
+ item_sep: '/'
+ exclude_keys:
+ - run_config
+ - distributed_training.distributed_port
+ - distributed_training.distributed_world_size
+ - model.pretrained_model_path
+ - model.target_network_path
+ - next_script
+ - task.cache_in_scratch
+ - task.data
+ - checkpoint.save_interval_updates
+ - checkpoint.keep_interval_updates
+ - checkpoint.save_on_overflow
+ - common.log_interval
+ - common.user_dir
+ sweep:
+ dir: /checkpoint/${env:USER}/${env:PREFIX}/${hydra.job.config_name}_${hydra.launcher.gpus_per_node}/${hydra.job.override_dirname}
+ subdir: ''
+ launcher:
+ submitit_folder: ${hydra.sweep.dir}
+ timeout_min: 4320
+ cpus_per_task: 10
+ gpus_per_node: 8
+ tasks_per_node: 8
+ mem_gb: 0
+ nodes: 8
+ name: ${env:PREFIX}_${hydra.job.config_name}
+ partition: wav2vec,learnlab,learnfair
+ max_num_timeout: 30
diff --git a/fairseq/examples/data2vec/config/vision/pretraining/base_imagenet.yaml b/fairseq/examples/data2vec/config/vision/pretraining/base_imagenet.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..9bfc0f32b68a740909bcfd7515a55bdbae69edaa
--- /dev/null
+++ b/fairseq/examples/data2vec/config/vision/pretraining/base_imagenet.yaml
@@ -0,0 +1,52 @@
+# @package _group_
+
+common:
+ fp16: true
+ log_format: json
+ log_interval: 200
+ tensorboard_logdir: tb
+
+checkpoint:
+ save_interval: 5
+ save_interval_updates: 25000
+ keep_interval_updates: 1
+ no_epoch_checkpoints: true
+
+task:
+ _name: image_pretraining
+ data: /datasets01/imagenet_full_size/061417/
+
+dataset:
+ num_workers: 6
+ batch_size: 64
+ skip_invalid_size_inputs_valid_test: true
+ required_batch_size_multiple: 1
+ disable_validation: true
+
+distributed_training:
+ distributed_world_size: 16
+ ddp_backend: c10d
+
+criterion:
+ _name: model
+ log_keys:
+ - ema_decay
+ - target_var
+ - pred_var
+
+optimization:
+ max_update: 400000
+ lr: [0.0005]
+
+optimizer:
+ _name: adam
+ adam_betas: (0.9,0.98)
+ adam_eps: 1e-06
+ weight_decay: 0.01
+
+lr_scheduler:
+ _name: cosine
+ warmup_updates: 10000
+
+model:
+ _name: data2vec_vision
diff --git a/fairseq/examples/data2vec/config/vision/pretraining/base_imagenet_d2v1.yaml b/fairseq/examples/data2vec/config/vision/pretraining/base_imagenet_d2v1.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..5fd399b117f74413b3ddcfd60d97022db78071a7
--- /dev/null
+++ b/fairseq/examples/data2vec/config/vision/pretraining/base_imagenet_d2v1.yaml
@@ -0,0 +1,64 @@
+# @package _group_
+
+common:
+ fp16: true
+ log_format: json
+ log_interval: 200
+ tensorboard_logdir: tb
+
+checkpoint:
+ save_interval: 5
+ save_interval_updates: 25000
+ keep_interval_updates: 1
+ no_epoch_checkpoints: true
+
+task:
+ _name: image_pretraining
+ data: /datasets01/imagenet_full_size/061417
+
+dataset:
+ num_workers: 6
+ batch_size: 128
+ skip_invalid_size_inputs_valid_test: true
+ required_batch_size_multiple: 2
+ disable_validation: true
+
+distributed_training:
+ distributed_world_size: 16
+ ddp_backend: legacy_ddp
+
+criterion:
+ _name: model
+ log_keys:
+ - ema_decay
+ - target_var
+ - pred_var
+
+optimization:
+ max_update: 375300 #300*1251
+ lr: [0.0005]
+ clip_norm: 3.0
+
+optimizer:
+ _name: adam
+ adam_betas: (0.9,0.999)
+ adam_eps: 1e-08
+ weight_decay: 0.05
+
+lr_scheduler:
+ _name: cosine
+ warmup_updates: 12510 # it should be 10 epochs
+
+model:
+ _name: data2vec_vision
+
+ attention_dropout: 0.05
+
+ ema_decay: 0.999
+ ema_end_decay: 0.9998
+ layer_norm_targets: True
+ average_top_k_layers: 6
+
+ loss_beta: 2.0
+
+ drop_path: 0.25
diff --git a/fairseq/examples/data2vec/config/vision/pretraining/base_mae_imagenet.yaml b/fairseq/examples/data2vec/config/vision/pretraining/base_mae_imagenet.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d7872b5e0445cd909e71a3f7ab36732aa40c5b69
--- /dev/null
+++ b/fairseq/examples/data2vec/config/vision/pretraining/base_mae_imagenet.yaml
@@ -0,0 +1,64 @@
+# @package _group_
+
+common:
+ fp16: true
+ log_format: json
+ log_interval: 200
+ tensorboard_logdir: tb
+ fp16_no_flatten_grads: true
+
+checkpoint:
+ save_interval: 5
+ save_interval_updates: 25000
+ keep_interval_updates: 1
+ no_epoch_checkpoints: true
+
+task:
+ _name: mae_image_pretraining
+ data: /datasets01/imagenet_full_size/061417/
+ rebuild_batches: true
+
+dataset:
+ num_workers: 6
+ batch_size: 64
+ skip_invalid_size_inputs_valid_test: true
+ required_batch_size_multiple: 1
+ disable_validation: true
+
+distributed_training:
+ distributed_world_size: 16
+ ddp_backend: c10d
+
+criterion:
+ _name: model
+
+optimization:
+ max_update: 375300
+ lr: [0.0006]
+
+optimizer:
+ _name: composite
+ groups:
+ with_decay:
+ lr_float: 6e-4
+ optimizer:
+ _name: adam
+ adam_betas: [0.9,0.95]
+ weight_decay: 0.05
+ lr_scheduler:
+ _name: cosine
+ warmup_updates: 50040
+ no_decay:
+ lr_float: 6e-4
+ optimizer:
+ _name: adam
+ adam_betas: [0.9,0.95]
+ weight_decay: 0
+ lr_scheduler:
+ _name: cosine
+ warmup_updates: 50040
+
+lr_scheduler: pass_through
+
+model:
+ _name: mae
diff --git a/fairseq/examples/data2vec/config/vision/pretraining/run_config/local.yaml b/fairseq/examples/data2vec/config/vision/pretraining/run_config/local.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..45595f9eea7c369a1200d802bfa1883c1bdfe573
--- /dev/null
+++ b/fairseq/examples/data2vec/config/vision/pretraining/run_config/local.yaml
@@ -0,0 +1,15 @@
+# @package _global_
+hydra:
+ sweep:
+ dir: ${env:PWD}/tmp_dbg/${now:%H-%M-%S}
+
+distributed_training:
+ distributed_world_size: 1
+ nprocs_per_node: 1
+ distributed_port: -1
+
+common:
+ log_interval: 1
+
+dataset:
+ num_workers: 0
diff --git a/fairseq/examples/data2vec/config/vision/pretraining/run_config/slurm_1.yaml b/fairseq/examples/data2vec/config/vision/pretraining/run_config/slurm_1.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..732f0188993a14c847ab5d5e8addb24b0b536311
--- /dev/null
+++ b/fairseq/examples/data2vec/config/vision/pretraining/run_config/slurm_1.yaml
@@ -0,0 +1,37 @@
+# @package _global_
+
+hydra:
+ job:
+ config:
+ override_dirname:
+ kv_sep: ':'
+ item_sep: '/'
+ exclude_keys:
+ - run_config
+ - distributed_training.distributed_port
+ - distributed_training.distributed_world_size
+ - model.pretrained_model_path
+ - model.target_network_path
+ - next_script
+ - task.cache_in_scratch
+ - task.data
+ - checkpoint.save_interval_updates
+ - checkpoint.keep_interval_updates
+ - checkpoint.save_on_overflow
+ - common.log_interval
+ - common.user_dir
+ sweep:
+ dir: /checkpoint/${env:USER}/${env:PREFIX}/${hydra.job.config_name}_${hydra.launcher.gpus_per_node}/${hydra.job.override_dirname}
+ subdir: ''
+ launcher:
+ submitit_folder: ${hydra.sweep.dir}
+ timeout_min: 4320
+ cpus_per_task: 80
+ gpus_per_node: 8
+ tasks_per_node: 1
+ mem_gb: 450
+ nodes: 1
+ name: ${env:PREFIX}_${hydra.job.config_name}
+ partition: devlab,learnlab,learnfair,scavenge
+ constraint: volta32gb,ib4
+ max_num_timeout: 30
diff --git a/fairseq/examples/data2vec/config/vision/pretraining/run_config/slurm_1_aws.yaml b/fairseq/examples/data2vec/config/vision/pretraining/run_config/slurm_1_aws.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..e2bab5675a561a6a01b5db06530b119a282bbf4a
--- /dev/null
+++ b/fairseq/examples/data2vec/config/vision/pretraining/run_config/slurm_1_aws.yaml
@@ -0,0 +1,36 @@
+# @package _global_
+
+hydra:
+ job:
+ config:
+ override_dirname:
+ kv_sep: ':'
+ item_sep: '/'
+ exclude_keys:
+ - run_config
+ - distributed_training.distributed_port
+ - distributed_training.distributed_world_size
+ - model.pretrained_model_path
+ - model.target_network_path
+ - next_script
+ - task.cache_in_scratch
+ - task.data
+ - checkpoint.save_interval_updates
+ - checkpoint.keep_interval_updates
+ - checkpoint.save_on_overflow
+ - common.log_interval
+ - common.user_dir
+ sweep:
+ dir: /checkpoint/${env:USER}/${env:PREFIX}/${hydra.job.config_name}_${hydra.launcher.gpus_per_node}/${hydra.job.override_dirname}
+ subdir: ''
+ launcher:
+ submitit_folder: ${hydra.sweep.dir}
+ timeout_min: 4320
+ cpus_per_task: 80
+ gpus_per_node: 8
+ tasks_per_node: 1
+ mem_gb: 0
+ nodes: 1
+ name: ${env:PREFIX}_${hydra.job.config_name}
+ partition: wav2vec,learnlab,learnfair
+ max_num_timeout: 30
diff --git a/fairseq/examples/data2vec/config/vision/pretraining/run_config/slurm_2.yaml b/fairseq/examples/data2vec/config/vision/pretraining/run_config/slurm_2.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..c8b0f02a9b7795d900804c557cae6576153bbaf0
--- /dev/null
+++ b/fairseq/examples/data2vec/config/vision/pretraining/run_config/slurm_2.yaml
@@ -0,0 +1,38 @@
+# @package _global_
+
+hydra:
+ job:
+ config:
+ override_dirname:
+ kv_sep: ':'
+ item_sep: '/'
+ exclude_keys:
+ - run_config
+ - distributed_training.distributed_port
+ - distributed_training.distributed_world_size
+ - model.pretrained_model_path
+ - model.target_network_path
+ - next_script
+ - task.cache_in_scratch
+ - task.data
+ - checkpoint.save_interval_updates
+ - checkpoint.keep_interval_updates
+ - checkpoint.save_on_overflow
+ - common.log_interval
+ - common.user_dir
+ - task.local_cache_path
+ sweep:
+ dir: /checkpoint/${env:USER}/${env:PREFIX}/${hydra.job.config_name}_${hydra.launcher.gpus_per_node}/${hydra.job.override_dirname}
+ subdir: ''
+ launcher:
+ submitit_folder: ${hydra.sweep.dir}
+ timeout_min: 4320
+ cpus_per_task: 10
+ gpus_per_node: 8
+ tasks_per_node: 8
+ mem_gb: 450
+ nodes: 2
+ name: ${env:PREFIX}_${hydra.job.config_name}
+ partition: devlab,learnlab,learnfair,scavenge
+ constraint: volta32gb,ib4
+ max_num_timeout: 30
diff --git a/fairseq/examples/data2vec/config/vision/pretraining/run_config/slurm_2_aws.yaml b/fairseq/examples/data2vec/config/vision/pretraining/run_config/slurm_2_aws.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..032e53a3043c5205b06533eb2019996f32a5de9a
--- /dev/null
+++ b/fairseq/examples/data2vec/config/vision/pretraining/run_config/slurm_2_aws.yaml
@@ -0,0 +1,37 @@
+# @package _global_
+
+hydra:
+ job:
+ config:
+ override_dirname:
+ kv_sep: ':'
+ item_sep: '/'
+ exclude_keys:
+ - run_config
+ - distributed_training.distributed_port
+ - distributed_training.distributed_world_size
+ - model.pretrained_model_path
+ - model.target_network_path
+ - next_script
+ - task.cache_in_scratch
+ - task.data
+ - checkpoint.save_interval_updates
+ - checkpoint.keep_interval_updates
+ - checkpoint.save_on_overflow
+ - common.log_interval
+ - common.user_dir
+ - task.local_cache_path
+ sweep:
+ dir: /fsx-wav2vec/${env:USER}/${env:PREFIX}/${hydra.job.config_name}_${hydra.launcher.gpus_per_node}/${hydra.job.override_dirname}
+ subdir: ''
+ launcher:
+ submitit_folder: ${hydra.sweep.dir}
+ timeout_min: 4320
+ cpus_per_task: 10
+ gpus_per_node: 8
+ tasks_per_node: 8
+ mem_gb: 0
+ nodes: 2
+ name: ${env:PREFIX}_${hydra.job.config_name}
+ partition: wav2vec,learnlab,learnfair
+ max_num_timeout: 30
diff --git a/fairseq/examples/data2vec/config/vision/pretraining/run_config/slurm_3.yaml b/fairseq/examples/data2vec/config/vision/pretraining/run_config/slurm_3.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..14b47d14e6282d5a846cc7c41c6b2080c5ed8c62
--- /dev/null
+++ b/fairseq/examples/data2vec/config/vision/pretraining/run_config/slurm_3.yaml
@@ -0,0 +1,36 @@
+# @package _global_
+
+hydra:
+ job:
+ config:
+ override_dirname:
+ kv_sep: ':'
+ item_sep: '/'
+ exclude_keys:
+ - run_config
+ - distributed_training.distributed_port
+ - distributed_training.distributed_world_size
+ - model.pretrained_model_path
+ - model.target_network_path
+ - next_script
+ - task.cache_in_scratch
+ - task.data
+ - checkpoint.save_interval_updates
+ - checkpoint.keep_interval_updates
+ - checkpoint.save_on_overflow
+ - common.log_interval
+ sweep:
+ dir: /checkpoint/${env:USER}/${env:PREFIX}/${hydra.job.config_name}_${hydra.launcher.gpus_per_node}/${hydra.job.override_dirname}
+ subdir: ''
+ launcher:
+ submitit_folder: ${hydra.sweep.dir}
+ timeout_min: 4320
+ cpus_per_task: 80
+ gpus_per_node: 8
+ tasks_per_node: 1
+ mem_gb: 450
+ nodes: 3
+ name: ${env:PREFIX}_${hydra.job.config_name}
+ partition: devlab,learnlab,learnfair,scavenge
+ constraint: volta32gb,ib4
+ max_num_timeout: 30
diff --git a/fairseq/examples/data2vec/config/vision/pretraining/run_config/slurm_4.yaml b/fairseq/examples/data2vec/config/vision/pretraining/run_config/slurm_4.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..c54d735fb2dc0b0af2d467caa7f64405af18ea10
--- /dev/null
+++ b/fairseq/examples/data2vec/config/vision/pretraining/run_config/slurm_4.yaml
@@ -0,0 +1,36 @@
+# @package _global_
+
+hydra:
+ job:
+ config:
+ override_dirname:
+ kv_sep: ':'
+ item_sep: '/'
+ exclude_keys:
+ - run_config
+ - distributed_training.distributed_port
+ - distributed_training.distributed_world_size
+ - model.pretrained_model_path
+ - model.target_network_path
+ - next_script
+ - task.cache_in_scratch
+ - task.data
+ - checkpoint.save_interval_updates
+ - checkpoint.keep_interval_updates
+ - checkpoint.save_on_overflow
+ - common.log_interval
+ sweep:
+ dir: /checkpoint/${env:USER}/${env:PREFIX}/${hydra.job.config_name}_${hydra.launcher.gpus_per_node}/${hydra.job.override_dirname}
+ subdir: ''
+ launcher:
+ submitit_folder: ${hydra.sweep.dir}
+ timeout_min: 4320
+ cpus_per_task: 10
+ gpus_per_node: 8
+ tasks_per_node: 8
+ mem_gb: 450
+ nodes: 4
+ name: ${env:PREFIX}_${hydra.job.config_name}
+ partition: devlab,learnlab,learnfair,scavenge
+ constraint: volta32gb,ib4
+ max_num_timeout: 30
diff --git a/fairseq/examples/data2vec/config/vision/pretraining/run_config/slurm_4_aws.yaml b/fairseq/examples/data2vec/config/vision/pretraining/run_config/slurm_4_aws.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d5d11cb755dcdafdc23ba8c0c44961323c274635
--- /dev/null
+++ b/fairseq/examples/data2vec/config/vision/pretraining/run_config/slurm_4_aws.yaml
@@ -0,0 +1,36 @@
+# @package _global_
+
+hydra:
+ job:
+ config:
+ override_dirname:
+ kv_sep: ':'
+ item_sep: '/'
+ exclude_keys:
+ - run_config
+ - distributed_training.distributed_port
+ - distributed_training.distributed_world_size
+ - model.pretrained_model_path
+ - model.target_network_path
+ - next_script
+ - task.cache_in_scratch
+ - task.data
+ - checkpoint.save_interval_updates
+ - checkpoint.keep_interval_updates
+ - checkpoint.save_on_overflow
+ - common.log_interval
+ - common.user_dir
+ sweep:
+ dir: /checkpoint/${env:USER}/${env:PREFIX}/${hydra.job.config_name}_${hydra.launcher.gpus_per_node}/${hydra.job.override_dirname}
+ subdir: ''
+ launcher:
+ submitit_folder: ${hydra.sweep.dir}
+ timeout_min: 4320
+ cpus_per_task: 10
+ gpus_per_node: 8
+ tasks_per_node: 8
+ mem_gb: 0
+ nodes: 4
+ name: ${env:PREFIX}_${hydra.job.config_name}
+ partition: wav2vec,learnlab,learnfair
+ max_num_timeout: 30
diff --git a/fairseq/examples/data2vec/config/vision/pretraining/run_config/slurm_6_aws.yaml b/fairseq/examples/data2vec/config/vision/pretraining/run_config/slurm_6_aws.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..906f08a602f498a4ce5ee24979861fab88ff8e96
--- /dev/null
+++ b/fairseq/examples/data2vec/config/vision/pretraining/run_config/slurm_6_aws.yaml
@@ -0,0 +1,36 @@
+# @package _global_
+
+hydra:
+ job:
+ config:
+ override_dirname:
+ kv_sep: ':'
+ item_sep: '/'
+ exclude_keys:
+ - run_config
+ - distributed_training.distributed_port
+ - distributed_training.distributed_world_size
+ - model.pretrained_model_path
+ - model.target_network_path
+ - next_script
+ - task.cache_in_scratch
+ - task.data
+ - checkpoint.save_interval_updates
+ - checkpoint.keep_interval_updates
+ - checkpoint.save_on_overflow
+ - common.log_interval
+ - common.user_dir
+ sweep:
+ dir: /checkpoint/${env:USER}/${env:PREFIX}/${hydra.job.config_name}_${hydra.launcher.gpus_per_node}/${hydra.job.override_dirname}
+ subdir: ''
+ launcher:
+ submitit_folder: ${hydra.sweep.dir}
+ timeout_min: 4320
+ cpus_per_task: 10
+ gpus_per_node: 8
+ tasks_per_node: 8
+ mem_gb: 0
+ nodes: 6
+ name: ${env:PREFIX}_${hydra.job.config_name}
+ partition: wav2vec,learnlab,learnfair
+ max_num_timeout: 30
diff --git a/fairseq/examples/data2vec/config/vision/pretraining/run_config/slurm_8_aws.yaml b/fairseq/examples/data2vec/config/vision/pretraining/run_config/slurm_8_aws.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d60e13f8ba6541267e7b09eb9d6bb61fb9c618d1
--- /dev/null
+++ b/fairseq/examples/data2vec/config/vision/pretraining/run_config/slurm_8_aws.yaml
@@ -0,0 +1,36 @@
+# @package _global_
+
+hydra:
+ job:
+ config:
+ override_dirname:
+ kv_sep: ':'
+ item_sep: '/'
+ exclude_keys:
+ - run_config
+ - distributed_training.distributed_port
+ - distributed_training.distributed_world_size
+ - model.pretrained_model_path
+ - model.target_network_path
+ - next_script
+ - task.cache_in_scratch
+ - task.data
+ - checkpoint.save_interval_updates
+ - checkpoint.keep_interval_updates
+ - checkpoint.save_on_overflow
+ - common.log_interval
+ - common.user_dir
+ sweep:
+ dir: /checkpoint/${env:USER}/${env:PREFIX}/${hydra.job.config_name}_${hydra.launcher.gpus_per_node}/${hydra.job.override_dirname}
+ subdir: ''
+ launcher:
+ submitit_folder: ${hydra.sweep.dir}
+ timeout_min: 4320
+ cpus_per_task: 10
+ gpus_per_node: 8
+ tasks_per_node: 8
+ mem_gb: 0
+ nodes: 8
+ name: ${env:PREFIX}_${hydra.job.config_name}
+ partition: wav2vec,learnlab,learnfair
+ max_num_timeout: 30
diff --git a/fairseq/examples/data2vec/data/__init__.py b/fairseq/examples/data2vec/data/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..d76112bfc25af1bafe8da4209a7bfbb05cdde043
--- /dev/null
+++ b/fairseq/examples/data2vec/data/__init__.py
@@ -0,0 +1,17 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+from .image_dataset import ImageDataset
+from .path_dataset import PathDataset
+from .mae_image_dataset import MaeImageDataset
+from .mae_finetuning_image_dataset import MaeFinetuningImageDataset
+
+
+__all__ = [
+ "ImageDataset",
+ "MaeImageDataset",
+ "MaeFinetuningImageDataset",
+ "PathDataset",
+]
\ No newline at end of file
diff --git a/fairseq/examples/data2vec/data/add_class_target_dataset.py b/fairseq/examples/data2vec/data/add_class_target_dataset.py
new file mode 100644
index 0000000000000000000000000000000000000000..c346c83e586f4cffcfbcab3e616ae38457c194fb
--- /dev/null
+++ b/fairseq/examples/data2vec/data/add_class_target_dataset.py
@@ -0,0 +1,63 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+
+from fairseq.data import BaseWrapperDataset, data_utils
+
+
+class AddClassTargetDataset(BaseWrapperDataset):
+ def __init__(
+ self,
+ dataset,
+ labels,
+ multi_class,
+ num_classes=None,
+ label_indices=None,
+ add_to_input=True,
+ ):
+ super().__init__(dataset)
+
+ self.label_indices = label_indices
+ self.labels = labels
+ self.multi_class = multi_class
+ self.add_to_input = add_to_input
+ if num_classes is None and multi_class:
+ assert self.label_indices is not None
+ num_classes = len(self.label_indices)
+
+ self.num_classes = num_classes
+
+ def __getitem__(self, index):
+ item = self.dataset[index]
+ item_labels = self.labels[index]
+ if self.multi_class:
+ item["label"] = torch.zeros(self.num_classes)
+ for il in item_labels:
+ if self.label_indices is not None:
+ il = self.label_indices[il]
+ item["label"][il] = 1.0
+ else:
+ item["label"] = torch.tensor(
+ self.labels[index]
+ if self.label_indices is None
+ else self.label_indices[self.labels[index]]
+ )
+
+ return item
+
+ def collater(self, samples):
+ collated = self.dataset.collater(samples)
+ if len(collated) == 0:
+ return collated
+
+ indices = set(collated["id"].tolist())
+ target = [s["label"] for s in samples if s["id"] in indices]
+ collated["label"] = torch.stack(target, dim=0)
+
+ if self.add_to_input:
+ collated["net_input"]["label"] = collated["label"]
+
+ return collated
diff --git a/fairseq/examples/data2vec/data/image_dataset.py b/fairseq/examples/data2vec/data/image_dataset.py
new file mode 100644
index 0000000000000000000000000000000000000000..7f551057e8e52a3c32a7a49b2f28f9ddd0a9ab40
--- /dev/null
+++ b/fairseq/examples/data2vec/data/image_dataset.py
@@ -0,0 +1,127 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+
+import logging
+
+import numpy as np
+import os
+from typing import Optional, Callable, Set
+
+import torch
+
+from torchvision.datasets.vision import VisionDataset
+from torchvision.transforms import ToTensor
+
+from fairseq.data import FairseqDataset
+
+
+logger = logging.getLogger(__name__)
+
+
+class ImageDataset(FairseqDataset, VisionDataset):
+ def __init__(
+ self,
+ root: str,
+ extensions: Set[str],
+ load_classes: bool,
+ transform: Optional[Callable] = None,
+ shuffle=True,
+ ):
+ FairseqDataset.__init__(self)
+ VisionDataset.__init__(self, root=root, transform=transform)
+
+ self.shuffle = shuffle
+ self.tensor_transform = ToTensor()
+
+ self.classes = None
+ self.labels = None
+ if load_classes:
+ classes = [d.name for d in os.scandir(root) if d.is_dir()]
+ classes.sort()
+ self.classes = {cls_name: i for i, cls_name in enumerate(classes)}
+ logger.info(f"loaded {len(self.classes)} classes")
+ self.labels = []
+
+ def walk_path(root_path):
+ for root, _, fnames in sorted(os.walk(root_path, followlinks=True)):
+ for fname in sorted(fnames):
+ fname_ext = os.path.splitext(fname)
+ if fname_ext[-1].lower() not in extensions:
+ continue
+
+ path = os.path.join(root, fname)
+ yield path
+
+ logger.info(f"finding images in {root}")
+ if self.classes is not None:
+ self.files = []
+ self.labels = []
+ for c, i in self.classes.items():
+ for f in walk_path(os.path.join(root, c)):
+ self.files.append(f)
+ self.labels.append(i)
+ else:
+ self.files = [f for f in walk_path(root)]
+
+ logger.info(f"loaded {len(self.files)} examples")
+
+ def __getitem__(self, index):
+ from PIL import Image
+
+ fpath = self.files[index]
+
+ with open(fpath, "rb") as f:
+ img = Image.open(f).convert("RGB")
+
+ if self.transform is None:
+ img = self.tensor_transform(img)
+ else:
+ img = self.transform(img)
+ assert torch.is_tensor(img)
+
+ res = {"id": index, "img": img}
+
+ if self.labels is not None:
+ res["label"] = self.labels[index]
+
+ return res
+
+ def __len__(self):
+ return len(self.files)
+
+ def collater(self, samples):
+ if len(samples) == 0:
+ return {}
+
+ collated_img = torch.stack([s["img"] for s in samples], dim=0)
+
+ res = {
+ "id": torch.LongTensor([s["id"] for s in samples]),
+ "net_input": {
+ "img": collated_img,
+ },
+ }
+
+ if "label" in samples[0]:
+ res["net_input"]["label"] = torch.LongTensor([s["label"] for s in samples])
+
+ return res
+
+ def num_tokens(self, index):
+ return 1
+
+ def size(self, index):
+ return 1
+
+ def ordered_indices(self):
+ """Return an ordered list of indices. Batches will be constructed based
+ on this order."""
+ if self.shuffle:
+ order = [np.random.permutation(len(self))]
+ else:
+ order = [np.arange(len(self))]
+
+ return order[0]
diff --git a/fairseq/examples/data2vec/data/mae_finetuning_image_dataset.py b/fairseq/examples/data2vec/data/mae_finetuning_image_dataset.py
new file mode 100644
index 0000000000000000000000000000000000000000..28cbcb38ac14b36fc8f871f9a59da7ef7906f3e1
--- /dev/null
+++ b/fairseq/examples/data2vec/data/mae_finetuning_image_dataset.py
@@ -0,0 +1,135 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+
+import logging
+
+import numpy as np
+import os
+
+import torch
+
+from torchvision import datasets, transforms
+
+from timm.data import create_transform
+from timm.data.constants import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD
+import PIL
+
+from fairseq.data import FairseqDataset
+from .mae_image_dataset import caching_loader
+
+
+logger = logging.getLogger(__name__)
+
+
+def build_transform(is_train, input_size, color_jitter, aa, reprob, remode, recount):
+ mean = IMAGENET_DEFAULT_MEAN
+ std = IMAGENET_DEFAULT_STD
+ # train transform
+ if is_train:
+ # this should always dispatch to transforms_imagenet_train
+ transform = create_transform(
+ input_size=input_size,
+ is_training=True,
+ color_jitter=color_jitter,
+ auto_augment=aa,
+ interpolation="bicubic",
+ re_prob=reprob,
+ re_mode=remode,
+ re_count=recount,
+ mean=mean,
+ std=std,
+ )
+ return transform
+
+ # eval transform
+ t = []
+ if input_size <= 224:
+ crop_pct = 224 / 256
+ else:
+ crop_pct = 1.0
+ size = int(input_size / crop_pct)
+ t.append(
+ transforms.Resize(
+ size, interpolation=PIL.Image.BICUBIC
+ ), # to maintain same ratio w.r.t. 224 images
+ )
+ t.append(transforms.CenterCrop(input_size))
+
+ t.append(transforms.ToTensor())
+ t.append(transforms.Normalize(mean, std))
+ return transforms.Compose(t)
+
+
+class MaeFinetuningImageDataset(FairseqDataset):
+ def __init__(
+ self,
+ root: str,
+ split: str,
+ is_train: bool,
+ input_size,
+ color_jitter=None,
+ aa="rand-m9-mstd0.5-inc1",
+ reprob=0.25,
+ remode="pixel",
+ recount=1,
+ local_cache_path=None,
+ shuffle=True,
+ ):
+ FairseqDataset.__init__(self)
+
+ self.shuffle = shuffle
+
+ transform = build_transform(
+ is_train, input_size, color_jitter, aa, reprob, remode, recount
+ )
+
+ path = os.path.join(root, split)
+ loader = caching_loader(local_cache_path, datasets.folder.default_loader)
+
+ self.dataset = datasets.ImageFolder(path, loader=loader, transform=transform)
+
+ logger.info(f"loaded {len(self.dataset)} examples")
+
+ def __getitem__(self, index):
+ img, label = self.dataset[index]
+ return {"id": index, "img": img, "label": label}
+
+ def __len__(self):
+ return len(self.dataset)
+
+ def collater(self, samples):
+ if len(samples) == 0:
+ return {}
+
+ collated_img = torch.stack([s["img"] for s in samples], dim=0)
+
+ res = {
+ "id": torch.LongTensor([s["id"] for s in samples]),
+ "net_input": {
+ "imgs": collated_img,
+ },
+ }
+
+ if "label" in samples[0]:
+ res["net_input"]["labels"] = torch.LongTensor([s["label"] for s in samples])
+
+ return res
+
+ def num_tokens(self, index):
+ return 1
+
+ def size(self, index):
+ return 1
+
+ def ordered_indices(self):
+ """Return an ordered list of indices. Batches will be constructed based
+ on this order."""
+ if self.shuffle:
+ order = [np.random.permutation(len(self))]
+ else:
+ order = [np.arange(len(self))]
+
+ return order[0]
diff --git a/fairseq/examples/data2vec/data/mae_image_dataset.py b/fairseq/examples/data2vec/data/mae_image_dataset.py
new file mode 100644
index 0000000000000000000000000000000000000000..4aacb94895d0e98ff0039a14c5f5ff3f940d78c4
--- /dev/null
+++ b/fairseq/examples/data2vec/data/mae_image_dataset.py
@@ -0,0 +1,418 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+
+from functools import partial
+import logging
+import math
+import random
+import time
+
+import numpy as np
+import os
+
+import torch
+
+from torchvision import datasets, transforms
+from .path_dataset import PathDataset
+
+from fairseq.data import FairseqDataset
+from fairseq.data.data_utils import compute_block_mask_1d, compute_block_mask_2d
+
+from shutil import copyfile
+
+logger = logging.getLogger(__name__)
+
+
+def load(path, loader, cache):
+ if hasattr(caching_loader, "cache_root"):
+ cache = caching_loader.cache_root
+
+ cached_path = cache + path
+
+ num_tries = 3
+ for curr_try in range(num_tries):
+ try:
+ if curr_try == 2:
+ return loader(path)
+ if not os.path.exists(cached_path) or curr_try > 0:
+ os.makedirs(os.path.dirname(cached_path), exist_ok=True)
+ copyfile(path, cached_path)
+ os.chmod(cached_path, 0o777)
+ return loader(cached_path)
+ except Exception as e:
+ logger.warning(str(e))
+ if "Errno 13" in str(e):
+ caching_loader.cache_root = f"/scratch/{random.randint(0, 69420)}"
+ logger.warning(f"setting cache root to {caching_loader.cache_root}")
+ cached_path = caching_loader.cache_root + path
+ if curr_try == (num_tries - 1):
+ raise
+ time.sleep(2)
+
+
+def caching_loader(cache_root: str, loader):
+ if cache_root is None:
+ return loader
+
+ if cache_root == "slurm_tmpdir":
+ cache_root = os.environ["SLURM_TMPDIR"]
+ assert len(cache_root) > 0
+
+ if not cache_root.endswith("/"):
+ cache_root += "/"
+
+ return partial(load, loader=loader, cache=cache_root)
+
+
+class RandomResizedCropAndInterpolationWithTwoPic:
+ """Crop the given PIL Image to random size and aspect ratio with random interpolation.
+
+ A crop of random size (default: of 0.08 to 1.0) of the original size and a random
+ aspect ratio (default: of 3/4 to 4/3) of the original aspect ratio is made. This crop
+ is finally resized to given size.
+ This is popularly used to train the Inception networks.
+
+ Args:
+ size: expected output size of each edge
+ scale: range of size of the origin size cropped
+ ratio: range of aspect ratio of the origin aspect ratio cropped
+ interpolation: Default: PIL.Image.BILINEAR
+ """
+
+ def __init__(
+ self,
+ size,
+ second_size=None,
+ scale=(0.08, 1.0),
+ ratio=(3.0 / 4.0, 4.0 / 3.0),
+ interpolation="bilinear",
+ second_interpolation="lanczos",
+ ):
+ if isinstance(size, tuple):
+ self.size = size
+ else:
+ self.size = (size, size)
+ if second_size is not None:
+ if isinstance(second_size, tuple):
+ self.second_size = second_size
+ else:
+ self.second_size = (second_size, second_size)
+ else:
+ self.second_size = None
+ if (scale[0] > scale[1]) or (ratio[0] > ratio[1]):
+ logger.warning("range should be of kind (min, max)")
+
+ if interpolation == "random":
+ from PIL import Image
+
+ self.interpolation = (Image.BILINEAR, Image.BICUBIC)
+ else:
+ self.interpolation = self._pil_interp(interpolation)
+
+ self.second_interpolation = (
+ self._pil_interp(second_interpolation)
+ if second_interpolation is not None
+ else None
+ )
+ self.scale = scale
+ self.ratio = ratio
+
+ def _pil_interp(self, method):
+ from PIL import Image
+
+ if method == "bicubic":
+ return Image.BICUBIC
+ elif method == "lanczos":
+ return Image.LANCZOS
+ elif method == "hamming":
+ return Image.HAMMING
+ else:
+ # default bilinear, do we want to allow nearest?
+ return Image.BILINEAR
+
+ @staticmethod
+ def get_params(img, scale, ratio):
+ """Get parameters for ``crop`` for a random sized crop.
+
+ Args:
+ img (PIL Image): Image to be cropped.
+ scale (tuple): range of size of the origin size cropped
+ ratio (tuple): range of aspect ratio of the origin aspect ratio cropped
+
+ Returns:
+ tuple: params (i, j, h, w) to be passed to ``crop`` for a random
+ sized crop.
+ """
+ area = img.size[0] * img.size[1]
+
+ for attempt in range(10):
+ target_area = random.uniform(*scale) * area
+ log_ratio = (math.log(ratio[0]), math.log(ratio[1]))
+ aspect_ratio = math.exp(random.uniform(*log_ratio))
+
+ w = int(round(math.sqrt(target_area * aspect_ratio)))
+ h = int(round(math.sqrt(target_area / aspect_ratio)))
+
+ if w <= img.size[0] and h <= img.size[1]:
+ i = random.randint(0, img.size[1] - h)
+ j = random.randint(0, img.size[0] - w)
+ return i, j, h, w
+
+ # Fallback to central crop
+ in_ratio = img.size[0] / img.size[1]
+ if in_ratio < min(ratio):
+ w = img.size[0]
+ h = int(round(w / min(ratio)))
+ elif in_ratio > max(ratio):
+ h = img.size[1]
+ w = int(round(h * max(ratio)))
+ else: # whole image
+ w = img.size[0]
+ h = img.size[1]
+ i = (img.size[1] - h) // 2
+ j = (img.size[0] - w) // 2
+ return i, j, h, w
+
+ def __call__(self, img):
+ import torchvision.transforms.functional as F
+
+ """
+ Args:
+ img (PIL Image): Image to be cropped and resized.
+
+ Returns:
+ PIL Image: Randomly cropped and resized image.
+ """
+ i, j, h, w = self.get_params(img, self.scale, self.ratio)
+ if isinstance(self.interpolation, (tuple, list)):
+ interpolation = random.choice(self.interpolation)
+ else:
+ interpolation = self.interpolation
+ if self.second_size is None:
+ return F.resized_crop(img, i, j, h, w, self.size, interpolation)
+ else:
+ return F.resized_crop(
+ img, i, j, h, w, self.size, interpolation
+ ), F.resized_crop(
+ img, i, j, h, w, self.second_size, self.second_interpolation
+ )
+
+
+class MaeImageDataset(FairseqDataset):
+ def __init__(
+ self,
+ root: str,
+ split: str,
+ input_size,
+ local_cache_path=None,
+ shuffle=True,
+ key="imgs",
+ beit_transforms=False,
+ target_transform=False,
+ no_transform=False,
+ compute_mask=False,
+ patch_size: int = 16,
+ mask_prob: float = 0.75,
+ mask_prob_adjust: float = 0,
+ mask_length: int = 1,
+ inverse_mask: bool = False,
+ expand_adjacent: bool = False,
+ mask_dropout: float = 0,
+ non_overlapping: bool = False,
+ require_same_masks: bool = True,
+ clone_batch: int = 1,
+ dataset_type: str = "imagefolder",
+ ):
+ FairseqDataset.__init__(self)
+
+ self.shuffle = shuffle
+ self.key = key
+
+ loader = caching_loader(local_cache_path, datasets.folder.default_loader)
+
+ self.transform_source = None
+ self.transform_target = None
+
+ if target_transform:
+ self.transform_source = transforms.ColorJitter(0.4, 0.4, 0.4)
+ self.transform_target = transforms.ColorJitter(0.4, 0.4, 0.4)
+
+ if no_transform:
+ if input_size <= 224:
+ crop_pct = 224 / 256
+ else:
+ crop_pct = 1.0
+ size = int(input_size / crop_pct)
+
+ self.transform_train = transforms.Compose(
+ [
+ transforms.Resize(size, interpolation=3),
+ transforms.CenterCrop(input_size),
+ ]
+ )
+
+ self.transform_train = transforms.Resize((input_size, input_size))
+ elif beit_transforms:
+ beit_transform_list = []
+ if not target_transform:
+ beit_transform_list.append(transforms.ColorJitter(0.4, 0.4, 0.4))
+ beit_transform_list.extend(
+ [
+ transforms.RandomHorizontalFlip(p=0.5),
+ RandomResizedCropAndInterpolationWithTwoPic(
+ size=input_size,
+ second_size=None,
+ interpolation="bicubic",
+ second_interpolation=None,
+ ),
+ ]
+ )
+ self.transform_train = transforms.Compose(beit_transform_list)
+ else:
+ self.transform_train = transforms.Compose(
+ [
+ transforms.RandomResizedCrop(
+ input_size, scale=(0.2, 1.0), interpolation=3
+ ), # 3 is bicubic
+ transforms.RandomHorizontalFlip(),
+ ]
+ )
+ self.final_transform = transforms.Compose(
+ [
+ transforms.ToTensor(),
+ transforms.Normalize(
+ mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]
+ ),
+ ]
+ )
+
+ if dataset_type == "imagefolder":
+ self.dataset = datasets.ImageFolder(
+ os.path.join(root, split), loader=loader
+ )
+ elif dataset_type == "path":
+ self.dataset = PathDataset(
+ root,
+ loader,
+ None,
+ None,
+ mean=[0.485, 0.456, 0.406],
+ std=[0.229, 0.224, 0.225],
+ )
+ else:
+ raise Exception(f"invalid dataset type {dataset_type}")
+
+ logger.info(
+ f"initial transform: {self.transform_train}, "
+ f"source transform: {self.transform_source}, "
+ f"target transform: {self.transform_target}, "
+ f"final transform: {self.final_transform}"
+ )
+ logger.info(f"loaded {len(self.dataset)} examples")
+
+ self.is_compute_mask = compute_mask
+ self.patches = (input_size // patch_size) ** 2
+ self.mask_prob = mask_prob
+ self.mask_prob_adjust = mask_prob_adjust
+ self.mask_length = mask_length
+ self.inverse_mask = inverse_mask
+ self.expand_adjacent = expand_adjacent
+ self.mask_dropout = mask_dropout
+ self.non_overlapping = non_overlapping
+ self.require_same_masks = require_same_masks
+ self.clone_batch = clone_batch
+
+ def __getitem__(self, index):
+ img, _ = self.dataset[index]
+
+ img = self.transform_train(img)
+
+ source = None
+ target = None
+ if self.transform_source is not None:
+ source = self.final_transform(self.transform_source(img))
+ if self.transform_target is not None:
+ target = self.final_transform(self.transform_target(img))
+
+ if source is None:
+ img = self.final_transform(img)
+
+ v = {"id": index, self.key: source if source is not None else img}
+ if target is not None:
+ v["target"] = target
+
+ if self.is_compute_mask:
+ if self.mask_length == 1:
+ mask = compute_block_mask_1d(
+ shape=(self.clone_batch, self.patches),
+ mask_prob=self.mask_prob,
+ mask_length=self.mask_length,
+ mask_prob_adjust=self.mask_prob_adjust,
+ inverse_mask=self.inverse_mask,
+ require_same_masks=True,
+ )
+ else:
+ mask = compute_block_mask_2d(
+ shape=(self.clone_batch, self.patches),
+ mask_prob=self.mask_prob,
+ mask_length=self.mask_length,
+ mask_prob_adjust=self.mask_prob_adjust,
+ inverse_mask=self.inverse_mask,
+ require_same_masks=True,
+ expand_adjcent=self.expand_adjacent,
+ mask_dropout=self.mask_dropout,
+ non_overlapping=self.non_overlapping,
+ )
+
+ v["precomputed_mask"] = mask
+
+ return v
+
+ def __len__(self):
+ return len(self.dataset)
+
+ def collater(self, samples):
+ if len(samples) == 0:
+ return {}
+
+ collated_img = torch.stack([s[self.key] for s in samples], dim=0)
+
+ res = {
+ "id": torch.LongTensor([s["id"] for s in samples]),
+ "net_input": {
+ self.key: collated_img,
+ },
+ }
+
+ if "target" in samples[0]:
+ collated_target = torch.stack([s["target"] for s in samples], dim=0)
+ res["net_input"]["target"] = collated_target
+
+ if "precomputed_mask" in samples[0]:
+ collated_mask = torch.cat([s["precomputed_mask"] for s in samples], dim=0)
+ res["net_input"]["precomputed_mask"] = collated_mask
+
+ return res
+
+ def num_tokens(self, index):
+ return 1
+
+ def size(self, index):
+ return 1
+
+ @property
+ def sizes(self):
+ return np.full((len(self),), 1)
+
+ def ordered_indices(self):
+ """Return an ordered list of indices. Batches will be constructed based
+ on this order."""
+ if self.shuffle:
+ order = [np.random.permutation(len(self))]
+ else:
+ order = [np.arange(len(self))]
+
+ return order[0]
diff --git a/fairseq/examples/data2vec/data/modality.py b/fairseq/examples/data2vec/data/modality.py
new file mode 100644
index 0000000000000000000000000000000000000000..aa23ac94f7ea59410947eb57a52f9df7291d6ba3
--- /dev/null
+++ b/fairseq/examples/data2vec/data/modality.py
@@ -0,0 +1,14 @@
+# Copyright (c) 2017-present, Facebook, Inc.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the LICENSE file in
+# the root directory of this source tree. An additional grant of patent rights
+# can be found in the PATENTS file in the same directory.
+
+from enum import Enum, auto
+
+
+class Modality(Enum):
+ AUDIO = auto()
+ IMAGE = auto()
+ TEXT = auto()
diff --git a/fairseq/examples/data2vec/data/path_dataset.py b/fairseq/examples/data2vec/data/path_dataset.py
new file mode 100644
index 0000000000000000000000000000000000000000..02010058e60ea9103643d69bc43e58dbd59c9143
--- /dev/null
+++ b/fairseq/examples/data2vec/data/path_dataset.py
@@ -0,0 +1,64 @@
+import glob
+import os
+from typing import List, Optional, Tuple
+
+import logging
+import numpy as np
+import torchvision.transforms.functional as TF
+import PIL
+from PIL import Image
+from torchvision.datasets import VisionDataset
+
+logger = logging.getLogger(__name__)
+
+
+class PathDataset(VisionDataset):
+ def __init__(
+ self,
+ root: List[str],
+ loader: None = None,
+ transform: Optional[str] = None,
+ extra_transform: Optional[str] = None,
+ mean: Optional[List[float]] = None,
+ std: Optional[List[float]] = None,
+ ):
+ super().__init__(root=root)
+
+ PIL.Image.MAX_IMAGE_PIXELS = 256000001
+
+ self.files = []
+ for folder in self.root:
+ self.files.extend(
+ sorted(glob.glob(os.path.join(folder, "**", "*.jpg"), recursive=True))
+ )
+ self.files.extend(
+ sorted(glob.glob(os.path.join(folder, "**", "*.png"), recursive=True))
+ )
+
+ self.transform = transform
+ self.extra_transform = extra_transform
+ self.mean = mean
+ self.std = std
+
+ self.loader = loader
+
+ logger.info(f"loaded {len(self.files)} samples from {root}")
+
+ assert (mean is None) == (std is None)
+
+ def __len__(self) -> int:
+ return len(self.files)
+
+ def __getitem__(self, idx) -> Tuple[np.ndarray, np.ndarray]:
+ path = self.files[idx]
+
+ if self.loader is not None:
+ return self.loader(path), None
+
+ img = Image.open(path).convert("RGB")
+ if self.transform is not None:
+ img = self.transform(img)
+ img = TF.to_tensor(img)
+ if self.mean is not None and self.std is not None:
+ img = TF.normalize(img, self.mean, self.std)
+ return img, None
diff --git a/fairseq/examples/data2vec/fb_convert_beit_cp.py b/fairseq/examples/data2vec/fb_convert_beit_cp.py
new file mode 100644
index 0000000000000000000000000000000000000000..cf42ace762d12353d98ebdcd77f649d9b7025dca
--- /dev/null
+++ b/fairseq/examples/data2vec/fb_convert_beit_cp.py
@@ -0,0 +1,165 @@
+#!/usr/bin/env python3
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import argparse
+import torch
+
+from omegaconf import OmegaConf
+
+from fairseq.criterions.model_criterion import ModelCriterionConfig
+from fairseq.dataclass.configs import FairseqConfig
+
+from tasks import ImageClassificationConfig, ImagePretrainingConfig
+from models.data2vec_image_classification import (
+ Data2VecImageClassificationConfig,
+ Data2VecImageClassificationModel,
+)
+from models.data2vec_vision import Data2VecVisionConfig, Data2VecVisionModel
+
+
+def get_parser():
+ parser = argparse.ArgumentParser(
+ description="convert beit checkpoint into data2vec - vision checkpoint"
+ )
+ # fmt: off
+ parser.add_argument('checkpoint', help='checkpoint to convert')
+ parser.add_argument('--output', required=True, metavar='PATH', help='where to output converted checkpoint')
+ parser.add_argument('--type', type=str, choices=['vision', 'image_classification'], default='image_classification', help='type of model to upgrade')
+ parser.add_argument('--inception_norms', action='store_true', default=False)
+ # fmt: on
+
+ return parser
+
+
+def update_checkpoint(model_dict, prefix, is_nested):
+
+ replace_paths = {
+ "cls_token": "model.cls_emb" if is_nested else "cls_emb",
+ "patch_embed": "model.patch_embed" if is_nested else "patch_embed",
+ "mask_token": "mask_emb",
+ }
+
+ starts_with = {
+ "patch_embed.proj": "model.patch_embed.conv"
+ if is_nested
+ else "patch_embed.conv",
+ "lm_head": "final_proj",
+ "fc_norm": "fc_norm",
+ "head": "head",
+ }
+
+ partial = {
+ "mlp.fc1": "mlp.0",
+ "mlp.fc2": "mlp.2",
+ }
+
+ for k in list(model_dict.keys()):
+ for sw, r in starts_with.items():
+ if k.startswith(sw):
+ replace_paths[k] = k.replace(sw, r)
+ for p, r in partial.items():
+ if p in k:
+ replace_paths[k] = prefix + k.replace(p, r)
+
+ if prefix != "":
+ for k in list(model_dict.keys()):
+ if k not in replace_paths:
+ replace_paths[k] = prefix + k
+
+ for k in list(model_dict.keys()):
+ if k in replace_paths:
+ model_dict[replace_paths[k]] = model_dict[k]
+ if k != replace_paths[k]:
+ del model_dict[k]
+
+ return model_dict
+
+
+def main():
+ parser = get_parser()
+ args = parser.parse_args()
+
+ cp = torch.load(args.checkpoint, map_location="cpu")
+
+ cfg = FairseqConfig(
+ criterion=ModelCriterionConfig(_name="model", log_keys=["correct"]),
+ )
+
+ if args.type == "image_classification":
+
+ cfg.task = ImageClassificationConfig(
+ _name="image_classification",
+ data=".",
+ )
+
+ if args.inception_norms:
+ cfg.task.normalization_mean = [0.5, 0.5, 0.5]
+ cfg.task.normalization_std = [0.5, 0.5, 0.5]
+
+ cfg.model = Data2VecImageClassificationConfig(
+ _name="data2vec_image_classification",
+ )
+ cfg.model.pretrained_model_args = FairseqConfig(
+ model=Data2VecVisionConfig(
+ _name="data2vec_vision", shared_rel_pos_bias=False
+ ),
+ task=ImagePretrainingConfig(
+ _name="image_pretraining",
+ ),
+ )
+
+ cfg = OmegaConf.create(cfg)
+
+ state = {
+ "cfg": OmegaConf.to_container(cfg, resolve=True, enum_to_str=True),
+ "model": cp["module"],
+ "best_loss": None,
+ "optimizer": None,
+ "extra_state": {},
+ }
+
+ model = Data2VecImageClassificationModel(cfg.model)
+ model.load_state_dict(
+ update_checkpoint(state["model"], prefix="model.encoder.", is_nested=True),
+ strict=True,
+ )
+ elif args.type == "vision":
+ cfg.task = ImagePretrainingConfig(
+ _name="image_pretraining",
+ data=".",
+ )
+
+ if args.inception_norms:
+ cfg.task.normalization_mean = [0.5, 0.5, 0.5]
+ cfg.task.normalization_std = [0.5, 0.5, 0.5]
+
+ cfg.model = Data2VecVisionConfig(
+ _name="data2vec_vision",
+ )
+ cfg = OmegaConf.create(cfg)
+
+ state = {
+ "cfg": OmegaConf.to_container(cfg, resolve=True, enum_to_str=True),
+ "model": cp["model"],
+ "best_loss": None,
+ "optimizer": None,
+ "extra_state": {},
+ }
+
+ model = Data2VecVisionModel(cfg.model)
+ model.load_state_dict(
+ update_checkpoint(state["model"], prefix="encoder.", is_nested=False),
+ strict=True,
+ )
+ else:
+ raise Exception("unsupported type " + args.type)
+
+ print(state["cfg"], state.keys())
+ torch.save(state, args.output)
+
+
+if __name__ == "__main__":
+ main()
diff --git a/fairseq/examples/data2vec/models/__init__.py b/fairseq/examples/data2vec/models/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/fairseq/examples/data2vec/models/audio_classification.py b/fairseq/examples/data2vec/models/audio_classification.py
new file mode 100644
index 0000000000000000000000000000000000000000..06d21582671082806249076f355aea06dff8d66a
--- /dev/null
+++ b/fairseq/examples/data2vec/models/audio_classification.py
@@ -0,0 +1,614 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import contextlib
+import logging
+import re
+from dataclasses import dataclass, field
+from typing import Any, Optional
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import numpy as np
+from omegaconf import II, MISSING, open_dict
+
+from fairseq import checkpoint_utils, tasks
+from fairseq.dataclass import FairseqDataclass
+from fairseq.dataclass.utils import convert_namespace_to_omegaconf
+from fairseq.models import (
+ BaseFairseqModel,
+ register_model,
+)
+from fairseq.models.wav2vec.wav2vec2 import MASKING_DISTRIBUTION_CHOICES
+from fairseq.modules import TransposeLast
+from fairseq.tasks import FairseqTask
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class AudioClassificationConfig(FairseqDataclass):
+ model_path: str = field(
+ default=MISSING, metadata={"help": "path to wav2vec 2.0 model"}
+ )
+ no_pretrained_weights: bool = field(
+ default=False, metadata={"help": "if true, does not load pretrained weights"}
+ )
+ dropout_input: float = field(
+ default=0.0,
+ metadata={"help": "dropout to apply to the input (after feat extr)"},
+ )
+ final_dropout: float = field(
+ default=0.0,
+ metadata={"help": "dropout after transformer and before final projection"},
+ )
+ dropout: float = field(
+ default=0.0, metadata={"help": "dropout probability inside wav2vec 2.0 model"}
+ )
+ attention_dropout: float = field(
+ default=0.0,
+ metadata={
+ "help": "dropout probability for attention weights inside wav2vec 2.0 model"
+ },
+ )
+ activation_dropout: float = field(
+ default=0.0,
+ metadata={
+ "help": "dropout probability after activation in FFN inside wav2vec 2.0 model"
+ },
+ )
+
+ # masking
+ apply_mask: bool = field(
+ default=False, metadata={"help": "apply masking during fine-tuning"}
+ )
+ mask_length: int = field(
+ default=10, metadata={"help": "repeat the mask indices multiple times"}
+ )
+ mask_prob: float = field(
+ default=0.5,
+ metadata={
+ "help": "probability of replacing a token with mask (normalized by length)"
+ },
+ )
+ mask_selection: MASKING_DISTRIBUTION_CHOICES = field(
+ default="static", metadata={"help": "how to choose masks"}
+ )
+ mask_other: float = field(
+ default=0,
+ metadata={
+ "help": "secondary mask argument (used for more complex distributions), "
+ "see help in compute_mask_indices"
+ },
+ )
+ no_mask_overlap: bool = field(
+ default=False, metadata={"help": "whether to allow masks to overlap"}
+ )
+ mask_min_space: Optional[int] = field(
+ default=1,
+ metadata={"help": "min space between spans (if no overlap is enabled)"},
+ )
+ require_same_masks: bool = field(
+ default=True,
+ metadata={
+ "help": "whether to number of masked timesteps must be the same across all "
+ "examples in a batch"
+ },
+ )
+ mask_dropout: float = field(
+ default=0.0,
+ metadata={"help": "percent of masks to unmask for each sample"},
+ )
+
+ # channel masking
+ mask_channel_length: int = field(
+ default=10, metadata={"help": "length of the mask for features (channels)"}
+ )
+ mask_channel_prob: float = field(
+ default=0.0, metadata={"help": "probability of replacing a feature with 0"}
+ )
+ mask_channel_selection: MASKING_DISTRIBUTION_CHOICES = field(
+ default="static",
+ metadata={"help": "how to choose mask length for channel masking"},
+ )
+ mask_channel_other: float = field(
+ default=0,
+ metadata={
+ "help": "secondary mask argument (used for more complex distributions), "
+ "see help in compute_mask_indicesh"
+ },
+ )
+ no_mask_channel_overlap: bool = field(
+ default=False, metadata={"help": "whether to allow channel masks to overlap"}
+ )
+ freeze_finetune_updates: int = field(
+ default=0, metadata={"help": "dont finetune wav2vec for this many updates"}
+ )
+ feature_grad_mult: float = field(
+ default=0.0, metadata={"help": "reset feature grad mult in wav2vec 2.0 to this"}
+ )
+ layerdrop: float = field(
+ default=0.0, metadata={"help": "probability of dropping a layer in wav2vec 2.0"}
+ )
+ mask_channel_min_space: Optional[int] = field(
+ default=1,
+ metadata={"help": "min space between spans (if no overlap is enabled)"},
+ )
+ mask_channel_before: bool = False
+ normalize: bool = II("task.normalize")
+ data: str = II("task.data")
+ # this holds the loaded wav2vec args
+ d2v_args: Any = None
+ offload_activations: bool = field(
+ default=False, metadata={"help": "offload_activations"}
+ )
+ min_params_to_wrap: int = field(
+ default=int(1e8),
+ metadata={
+ "help": "minimum number of params for a layer to be wrapped with FSDP() when "
+ "training with --ddp-backend=fully_sharded. Smaller values will "
+ "improve memory efficiency, but may make torch.distributed "
+ "communication less efficient due to smaller input sizes. This option "
+ "is set to 0 (i.e., always wrap) when --checkpoint-activations or "
+ "--offload-activations are passed."
+ },
+ )
+
+ checkpoint_activations: bool = field(
+ default=False,
+ metadata={"help": "recompute activations and save memory for extra compute"},
+ )
+ ddp_backend: str = II("distributed_training.ddp_backend")
+
+ prediction_mode: str = "lin_softmax"
+ eval_prediction_mode: Optional[str] = None
+ conv_kernel: int = -1
+ conv_stride: int = 1
+ two_convs: bool = False
+ extreme_factor: float = 1.0
+
+ conv_feature_layers: Optional[str] = field(
+ default=None,
+ metadata={
+ "help": "string describing convolutional feature extraction layers in form of a python list that contains "
+ "[(dim, kernel_size, stride), ...]"
+ },
+ )
+
+ mixup_prob: float = 1.0
+ source_mixup: float = -1
+ same_mixup: bool = True
+ label_mixup: bool = False
+
+ gain_mode: str = "none"
+
+
+@register_model("audio_classification", dataclass=AudioClassificationConfig)
+class AudioClassificationModel(BaseFairseqModel):
+ def __init__(self, cfg: AudioClassificationConfig, num_classes):
+ super().__init__()
+
+ self.apply_mask = cfg.apply_mask
+ self.cfg = cfg
+
+ arg_overrides = {
+ "dropout": cfg.dropout,
+ "activation_dropout": cfg.activation_dropout,
+ "dropout_input": cfg.dropout_input,
+ "attention_dropout": cfg.attention_dropout,
+ "mask_length": cfg.mask_length,
+ "mask_prob": cfg.mask_prob,
+ "require_same_masks": getattr(cfg, "require_same_masks", True),
+ "mask_dropout": getattr(cfg, "mask_dropout", 0),
+ "mask_selection": cfg.mask_selection,
+ "mask_other": cfg.mask_other,
+ "no_mask_overlap": cfg.no_mask_overlap,
+ "mask_channel_length": cfg.mask_channel_length,
+ "mask_channel_prob": cfg.mask_channel_prob,
+ "mask_channel_before": cfg.mask_channel_before,
+ "mask_channel_selection": cfg.mask_channel_selection,
+ "mask_channel_other": cfg.mask_channel_other,
+ "no_mask_channel_overlap": cfg.no_mask_channel_overlap,
+ "encoder_layerdrop": cfg.layerdrop,
+ "feature_grad_mult": cfg.feature_grad_mult,
+ "checkpoint_activations": cfg.checkpoint_activations,
+ "offload_activations": cfg.offload_activations,
+ "min_params_to_wrap": cfg.min_params_to_wrap,
+ "mixup": -1,
+ }
+
+ if cfg.conv_feature_layers is not None:
+ arg_overrides["conv_feature_layers"] = cfg.conv_feature_layers
+
+ if cfg.d2v_args is None:
+ state = checkpoint_utils.load_checkpoint_to_cpu(
+ cfg.model_path, arg_overrides
+ )
+ d2v_args = state.get("cfg", None)
+ if d2v_args is None:
+ d2v_args = convert_namespace_to_omegaconf(state["args"])
+ d2v_args.criterion = None
+ d2v_args.lr_scheduler = None
+ cfg.d2v_args = d2v_args
+
+ logger.info(d2v_args)
+
+ else:
+ state = None
+ d2v_args = cfg.d2v_args
+
+ model_normalized = d2v_args.task.get(
+ "normalize", d2v_args.model.get("normalize", False)
+ )
+ assert cfg.normalize == model_normalized, (
+ "Fine-tuning works best when data normalization is the same. "
+ "Please check that --normalize is set or unset for both pre-training and here"
+ )
+
+ if hasattr(cfg, "checkpoint_activations") and cfg.checkpoint_activations:
+ with open_dict(d2v_args):
+ d2v_args.model.checkpoint_activations = cfg.checkpoint_activations
+
+ d2v_args.task.data = cfg.data
+ task = tasks.setup_task(d2v_args.task)
+ model = task.build_model(d2v_args.model, from_checkpoint=True)
+
+ model.remove_pretraining_modules()
+
+ if state is not None and not cfg.no_pretrained_weights:
+ self.load_model_weights(state, model, cfg)
+
+ d = d2v_args.model.encoder_embed_dim
+
+ self.d2v_model = model
+
+ self.final_dropout = nn.Dropout(cfg.final_dropout)
+ self.freeze_finetune_updates = cfg.freeze_finetune_updates
+ self.num_updates = 0
+
+ for p in self.parameters():
+ p.param_group = "pretrained"
+
+ if cfg.prediction_mode == "proj_avg_proj":
+ self.proj = nn.Linear(d, d * 2)
+ self.proj2 = nn.Linear(d * 2, num_classes)
+
+ for p in self.proj.parameters():
+ p.param_group = "projection"
+ for p in self.proj2.parameters():
+ p.param_group = "projection"
+ elif self.cfg.prediction_mode == "summary_proj":
+ self.proj = nn.Linear(d // 3, num_classes)
+ for p in self.proj.parameters():
+ p.param_group = "projection"
+ elif self.cfg.conv_kernel > 1 and not self.cfg.two_convs:
+ self.proj = nn.Sequential(
+ TransposeLast(),
+ nn.Conv1d(d, num_classes, kernel_size=self.cfg.conv_kernel, stride=self.cfg.conv_stride),
+ TransposeLast(),
+ )
+ for p in self.proj.parameters():
+ p.param_group = "projection"
+ elif self.cfg.conv_kernel > 0 and self.cfg.two_convs:
+ self.proj = nn.Sequential(
+ TransposeLast(),
+ nn.Conv1d(d, d, kernel_size=self.cfg.conv_kernel, stride=self.cfg.conv_stride),
+ TransposeLast(),
+ nn.GELU(),
+ nn.Linear(d, num_classes),
+ )
+ for p in self.proj.parameters():
+ p.param_group = "projection"
+ else:
+ self.proj = nn.Linear(d, num_classes)
+ for p in self.proj.parameters():
+ p.param_group = "projection"
+
+ def upgrade_state_dict_named(self, state_dict, name):
+ super().upgrade_state_dict_named(state_dict, name)
+ return state_dict
+
+ @classmethod
+ def build_model(cls, cfg: AudioClassificationConfig, task: FairseqTask):
+ """Build a new model instance."""
+
+ assert hasattr(task, "labels"), f"Task {task} must have an attribute 'labels'"
+
+ return cls(cfg, len(task.labels))
+
+ def load_model_weights(self, state, model, cfg):
+ if cfg.ddp_backend == "fully_sharded":
+ from fairseq.distributed import FullyShardedDataParallel
+
+ for name, module in model.named_modules():
+ if "encoder.layers" in name and len(name.split(".")) == 3:
+ # Only for layers, we do a special handling and load the weights one by one
+ # We dont load all weights together as that wont be memory efficient and may
+ # cause oom
+ new_dict = {
+ k.replace(name + ".", ""): v
+ for (k, v) in state["model"].items()
+ if name + "." in k
+ }
+ assert isinstance(module, FullyShardedDataParallel)
+ with module.summon_full_params():
+ module.load_state_dict(new_dict, strict=True)
+ module._reset_lazy_init()
+
+ # Once layers are loaded, filter them out and load everything else.
+ r = re.compile("encoder.layers.\d.")
+ filtered_list = list(filter(r.match, state["model"].keys()))
+
+ new_big_dict = {
+ k: v for (k, v) in state["model"].items() if k not in filtered_list
+ }
+
+ model.load_state_dict(new_big_dict, strict=False)
+ else:
+ if "_ema" in state["model"]:
+ del state["model"]["_ema"]
+ model.load_state_dict(state["model"], strict=False)
+
+ def set_num_updates(self, num_updates):
+ """Set the number of parameters updates."""
+ super().set_num_updates(num_updates)
+ self.num_updates = num_updates
+
+ def compute_gain(self, sound, fs=16_000, min_db=-80.0, mode="A_weighting"):
+ if fs == 16000:
+ n_fft = 2048
+ elif fs == 44100:
+ n_fft = 4096
+ else:
+ raise Exception("Invalid fs {}".format(fs))
+ stride = n_fft // 2
+
+ def a_weight(fs, n_fft, min_db=-80.0):
+ freq = np.linspace(0, fs // 2, n_fft // 2 + 1)
+ freq_sq = np.power(freq, 2)
+ freq_sq[0] = 1.0
+ weight = 2.0 + 20.0 * (
+ 2 * np.log10(12194)
+ + 2 * np.log10(freq_sq)
+ - np.log10(freq_sq + 12194 ** 2)
+ - np.log10(freq_sq + 20.6 ** 2)
+ - 0.5 * np.log10(freq_sq + 107.7 ** 2)
+ - 0.5 * np.log10(freq_sq + 737.9 ** 2)
+ )
+ weight = np.maximum(weight, min_db)
+
+ return weight
+
+ gain = []
+ for i in range(0, len(sound) - n_fft + 1, stride):
+ if mode == "RMSE":
+ g = np.mean(sound[i : i + n_fft] ** 2)
+ elif mode == "A_weighting":
+ spec = np.fft.rfft(np.hanning(n_fft + 1)[:-1] * sound[i : i + n_fft])
+ power_spec = np.abs(spec) ** 2
+ a_weighted_spec = power_spec * np.power(10, a_weight(fs, n_fft) / 10)
+ g = np.sum(a_weighted_spec)
+ else:
+ raise Exception("Invalid mode {}".format(mode))
+ gain.append(g)
+
+ gain = np.array(gain)
+ gain = np.maximum(gain, np.power(10, min_db / 10))
+ gain_db = 10 * np.log10(gain)
+
+ return gain_db
+
+ # adapted from https://github.com/mil-tokyo/bc_learning_sound/blob/master/utils.py
+ def compute_gain_torch(self, sound, fs=16_000, min_db=-80.0, mode="A_weighting"):
+ if fs == 16000:
+ n_fft = 2048
+ elif fs == 44100:
+ n_fft = 4096
+ else:
+ raise Exception("Invalid fs {}".format(fs))
+
+ if mode == "A_weighting":
+ if not hasattr(self, f"a_weight"):
+ self.a_weight = {}
+
+ if fs not in self.a_weight:
+
+ def a_weight(fs, n_fft, min_db=-80.0):
+ freq = np.linspace(0, fs // 2, n_fft // 2 + 1)
+ freq_sq = freq ** 2
+ freq_sq[0] = 1.0
+ weight = 2.0 + 20.0 * (
+ 2 * np.log10(12194)
+ + 2 * np.log10(freq_sq)
+ - np.log10(freq_sq + 12194 ** 2)
+ - np.log10(freq_sq + 20.6 ** 2)
+ - 0.5 * np.log10(freq_sq + 107.7 ** 2)
+ - 0.5 * np.log10(freq_sq + 737.9 ** 2)
+ )
+ weight = np.maximum(weight, min_db)
+
+ return weight
+
+ self.a_weight[fs] = torch.from_numpy(
+ np.power(10, a_weight(fs, n_fft, min_db) / 10)
+ ).to(device=sound.device)
+
+ sound = sound.unfold(-1, n_fft, n_fft // 2)
+
+ if mode == "RMSE":
+ sound = sound ** 2
+ g = sound.mean(-1)
+ elif mode == "A_weighting":
+ w = torch.hann_window(n_fft, device=sound.device) * sound
+ spec = torch.fft.rfft(w)
+ power_spec = spec.abs() ** 2
+ a_weighted_spec = power_spec * self.a_weight[fs]
+ g = a_weighted_spec.sum(-1)
+ else:
+ raise Exception("Invalid mode {}".format(mode))
+
+ gain = torch.maximum(g, torch.tensor(10 ** (min_db / 10), device=g.device))
+ gain_db = 10 * torch.log10(gain)
+
+ return gain_db
+
+ def forward(self, source, padding_mask, label=None, **kwargs):
+
+ if self.cfg.source_mixup >= 0 and self.training and self.cfg.mixup_prob > 0:
+ with torch.no_grad():
+ mixed_source = source
+ mix_mask = None
+ if self.cfg.mixup_prob < 1:
+ mix_mask = (
+ torch.empty((source.size(0),), device=source.device)
+ .bernoulli_(self.cfg.mixup_prob)
+ .bool()
+ )
+ mixed_source = source[mix_mask]
+
+ r = (
+ torch.FloatTensor(
+ 1 if self.cfg.same_mixup else mixed_source.size(0)
+ )
+ .uniform_(max(1e-6, self.cfg.source_mixup), 1)
+ .to(dtype=source.dtype, device=source.device)
+ )
+
+ mixup_perm = torch.randperm(source.size(0))
+ s2 = source[mixup_perm]
+
+ if self.cfg.gain_mode == "none":
+ p = r.unsqueeze(-1)
+ if mix_mask is not None:
+ s2 = s2[mix_mask]
+ else:
+ if self.cfg.gain_mode == "naive_rms":
+ G1 = source.pow(2).mean(dim=-1).sqrt()
+ else:
+ G1, _ = self.compute_gain_torch(
+ source, mode=self.cfg.gain_mode
+ ).max(-1)
+ G1 = G1.to(dtype=source.dtype)
+
+ G2 = G1[mixup_perm]
+
+ if mix_mask is not None:
+ G1 = G1[mix_mask]
+ G2 = G2[mix_mask]
+ s2 = s2[mix_mask]
+
+ p = 1 / (1 + 10 ** ((G1 - G2) / 20) * (1 - r) / r)
+ p = p.unsqueeze(-1)
+
+ mixed = (p * mixed_source) + (1 - p) * s2
+
+ if mix_mask is None:
+ source = mixed / torch.sqrt(p ** 2 + (1 - p) ** 2)
+ else:
+ source[mix_mask] = mixed / torch.sqrt(p ** 2 + (1 - p) ** 2)
+
+ if label is not None and self.cfg.label_mixup:
+ r = r.unsqueeze(-1)
+ if mix_mask is None:
+ label = label * r + (1 - r) * label[mixup_perm]
+ else:
+ label[mix_mask] = (
+ label[mix_mask] * r + (1 - r) * label[mixup_perm][mix_mask]
+ )
+
+ d2v_args = {
+ "source": source,
+ "padding_mask": padding_mask,
+ "mask": self.apply_mask and self.training,
+ }
+
+ ft = self.freeze_finetune_updates <= self.num_updates
+
+ with torch.no_grad() if not ft else contextlib.ExitStack():
+ res = self.d2v_model.extract_features(**d2v_args)
+
+ x = res["x"]
+ padding_mask = res["padding_mask"]
+ if padding_mask is not None:
+ x[padding_mask] = 0
+
+ x = self.final_dropout(x)
+
+ if self.training or (
+ self.cfg.eval_prediction_mode is None or self.cfg.eval_prediction_mode == ""
+ ):
+ prediction_mode = self.cfg.prediction_mode
+ else:
+ prediction_mode = self.cfg.eval_prediction_mode
+
+ if prediction_mode == "average_before":
+ x = x.mean(dim=1)
+
+ if prediction_mode != "summary_mha" and prediction_mode != "summary_proj" and prediction_mode != "cls":
+ x = self.proj(x)
+
+ logits = True
+ if prediction_mode == "lin_softmax":
+ x = F.logsigmoid(x.float())
+ x = torch.logsumexp(x + x, dim=1) - torch.logsumexp(x, dim=1)
+ x = x.clamp(max=0)
+ x = x - torch.log(-(torch.expm1(x)))
+ elif prediction_mode == "extremized_odds":
+ x = x.float().sum(dim=1)
+ x = x * self.cfg.extreme_factor
+ elif prediction_mode == "average_before":
+ x = x.float()
+ elif prediction_mode == "average":
+ x = x.float().mean(dim=1)
+ elif prediction_mode == "average_sigmoid":
+ x = torch.sigmoid(x.float())
+ x = x.mean(dim=1)
+ logits = False
+ elif prediction_mode == "max":
+ x, _ = x.float().max(dim=1)
+ elif prediction_mode == "max_sigmoid":
+ x = torch.sigmoid(x.float())
+ x, _ = x.float().max(dim=1)
+ logits = False
+ elif prediction_mode == "proj_avg_proj":
+ x = x.mean(dim=1)
+ x = self.proj2(x)
+ elif prediction_mode == "summary_mha" or prediction_mode == "summary_proj":
+ x = self.d2v_model.summary(
+ x, padding_mask, proj=prediction_mode == "summary_proj"
+ )
+ x = x.type_as(source)
+ x = self.proj(x)
+ elif prediction_mode == "cls":
+ x = x[:,0]
+ x = self.proj(x)
+ else:
+ raise Exception(f"unknown prediction mode {prediction_mode}")
+
+ if label is None:
+ return torch.sigmoid(x) if logits else x
+
+ x = torch.nan_to_num(x)
+
+ if logits:
+ loss = F.binary_cross_entropy_with_logits(
+ x, label.float(), reduction="none"
+ )
+ else:
+ loss = F.binary_cross_entropy(x, label.float(), reduction="none")
+
+ result = {
+ "losses": {
+ "main": loss,
+ },
+ "sample_size": label.sum(),
+ }
+
+ if not self.training:
+ result["_predictions"] = torch.sigmoid(x) if logits else x
+ result["_targets"] = label
+
+ return result
diff --git a/fairseq/examples/data2vec/models/data2vec2.py b/fairseq/examples/data2vec/models/data2vec2.py
new file mode 100644
index 0000000000000000000000000000000000000000..0c61b3708182437b3ef88683640792f095a36644
--- /dev/null
+++ b/fairseq/examples/data2vec/models/data2vec2.py
@@ -0,0 +1,813 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import logging
+import math
+from dataclasses import dataclass, field
+from typing import Optional, Callable
+from functools import partial
+import numpy as np
+
+from omegaconf import II
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.distributed as dist
+
+from fairseq.modules import EMAModule, EMAModuleConfig
+
+from fairseq.dataclass import FairseqDataclass
+from fairseq.models import BaseFairseqModel, register_model
+
+from examples.data2vec.data.modality import Modality
+
+from examples.data2vec.models.modalities.base import (
+ MaskSeed,
+ D2vModalityConfig,
+ ModalitySpecificEncoder,
+ get_annealed_rate,
+)
+from examples.data2vec.models.modalities.modules import (
+ D2vDecoderConfig,
+ AltBlock,
+ Decoder1d,
+)
+
+from examples.data2vec.models.modalities.audio import (
+ D2vAudioConfig,
+ AudioEncoder,
+)
+from examples.data2vec.models.modalities.images import (
+ D2vImageConfig,
+ ImageEncoder,
+)
+from examples.data2vec.models.modalities.text import (
+ D2vTextConfig,
+ TextEncoder,
+)
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class D2vModalitiesConfig(FairseqDataclass):
+ audio: D2vAudioConfig = D2vAudioConfig()
+ image: D2vImageConfig = D2vImageConfig()
+ text: D2vTextConfig = D2vTextConfig()
+
+
+@dataclass
+class Data2VecMultiConfig(FairseqDataclass):
+
+ loss_beta: float = field(
+ default=0, metadata={"help": "beta for smooth l1 loss. 0 means use l2 loss"}
+ )
+ loss_scale: Optional[float] = field(
+ default=None,
+ metadata={
+ "help": "scale the reconstruction loss by this constant. if None then scales by 1/sqrt(dim)"
+ },
+ )
+
+ depth: int = 8
+ start_drop_path_rate: float = 0
+ end_drop_path_rate: float = 0
+ num_heads: int = 12
+ norm_eps: float = 1e-6
+ norm_affine: bool = True
+ encoder_dropout: float = 0.1
+ post_mlp_drop: float = 0.1
+ attention_dropout: float = 0.1
+ activation_dropout: float = 0.0
+ dropout_input: float = 0.0
+ layerdrop: float = 0.0
+ embed_dim: int = 768
+ mlp_ratio: float = 4
+ layer_norm_first: bool = False
+
+ average_top_k_layers: int = field(
+ default=8, metadata={"help": "how many layers to average"}
+ )
+
+ end_of_block_targets: bool = False
+
+ clone_batch: int = 1
+
+ layer_norm_target_layer: bool = False
+ batch_norm_target_layer: bool = False
+ instance_norm_target_layer: bool = False
+ instance_norm_targets: bool = False
+ layer_norm_targets: bool = False
+
+ ema_decay: float = field(default=0.999, metadata={"help": "initial ema decay rate"})
+ ema_same_dtype: bool = True
+ log_norms: bool = True
+ ema_end_decay: float = field(
+ default=0.9999, metadata={"help": "final ema decay rate"}
+ )
+
+ # when to finish annealing ema decay rate
+ ema_anneal_end_step: int = II("optimization.max_update")
+
+ ema_encoder_only: bool = field(
+ default=True,
+ metadata={
+ "help": "whether to momentum update only the shared transformer encoder"
+ },
+ )
+
+ max_update: int = II("optimization.max_update")
+
+ modalities: D2vModalitiesConfig = D2vModalitiesConfig()
+
+ shared_decoder: Optional[D2vDecoderConfig] = None
+
+ min_target_var: float = field(
+ default=0.1, metadata={"help": "stop training if target var falls below this"}
+ )
+ min_pred_var: float = field(
+ default=0.01,
+ metadata={"help": "stop training if prediction var falls below this"},
+ )
+
+ supported_modality: Optional[Modality] = None
+ mae_init: bool = False
+
+ seed: int = II("common.seed")
+
+ skip_ema: bool = False
+
+ cls_loss: float = 0
+ recon_loss: float = 0
+ d2v_loss: float = 1
+
+ decoder_group: bool = False
+
+
+@register_model("data2vec_multi", dataclass=Data2VecMultiConfig)
+class Data2VecMultiModel(BaseFairseqModel):
+ def make_modality_encoder(
+ self,
+ cfg: D2vModalityConfig,
+ embed_dim: int,
+ make_block: Callable[[float], nn.ModuleList],
+ norm_layer: Callable[[int], nn.LayerNorm],
+ layer_norm_first: bool,
+ alibi_biases,
+ task,
+ ) -> ModalitySpecificEncoder:
+ if cfg.type == Modality.AUDIO:
+ enc_cls = AudioEncoder
+ elif cfg.type == Modality.IMAGE:
+ enc_cls = ImageEncoder
+ elif cfg.type == Modality.TEXT:
+ enc_cls = TextEncoder
+ if hasattr(task, "text_task"):
+ task = task.text_task
+ else:
+ raise Exception(f"unsupported modality {cfg.type}")
+
+ return enc_cls(
+ cfg,
+ embed_dim,
+ make_block,
+ norm_layer,
+ layer_norm_first,
+ alibi_biases,
+ task,
+ )
+
+ def __init__(self, cfg: Data2VecMultiConfig, modalities, skip_ema=False, task=None):
+ super().__init__()
+ self.cfg = cfg
+ self.modalities = modalities
+ self.task = task
+
+ make_layer_norm = partial(
+ nn.LayerNorm, eps=cfg.norm_eps, elementwise_affine=cfg.norm_affine
+ )
+
+ def make_block(drop_path, dim=None, heads=None):
+ return AltBlock(
+ cfg.embed_dim if dim is None else dim,
+ cfg.num_heads if heads is None else heads,
+ cfg.mlp_ratio,
+ qkv_bias=True,
+ drop=cfg.encoder_dropout,
+ attn_drop=cfg.attention_dropout,
+ mlp_drop=cfg.activation_dropout,
+ post_mlp_drop=cfg.post_mlp_drop,
+ drop_path=drop_path,
+ norm_layer=make_layer_norm,
+ layer_norm_first=cfg.layer_norm_first,
+ ffn_targets=not cfg.end_of_block_targets,
+ )
+
+ self.alibi_biases = {}
+ self.modality_encoders = nn.ModuleDict()
+ for mod in self.modalities:
+ mod_cfg = getattr(cfg.modalities, mod.name.lower())
+ enc = self.make_modality_encoder(
+ mod_cfg,
+ cfg.embed_dim,
+ make_block,
+ make_layer_norm,
+ cfg.layer_norm_first,
+ self.alibi_biases,
+ task,
+ )
+ self.modality_encoders[mod.name] = enc
+
+ self.ema = None
+
+ self.average_top_k_layers = cfg.average_top_k_layers
+ self.loss_beta = cfg.loss_beta
+ self.loss_scale = cfg.loss_scale
+
+ self.dropout_input = nn.Dropout(cfg.dropout_input)
+
+ dpr = np.linspace(cfg.start_drop_path_rate, cfg.end_drop_path_rate, cfg.depth)
+
+ self.blocks = nn.ModuleList([make_block(dpr[i]) for i in range(cfg.depth)])
+
+ self.norm = None
+ if cfg.layer_norm_first:
+ self.norm = make_layer_norm(cfg.embed_dim)
+
+ if self.cfg.mae_init:
+ self.apply(self._init_weights)
+ else:
+ from fairseq.modules.transformer_sentence_encoder import init_bert_params
+
+ self.apply(init_bert_params)
+
+ for mod_enc in self.modality_encoders.values():
+ mod_enc.reset_parameters()
+
+ if not skip_ema:
+ self.ema = self.make_ema_teacher(cfg.ema_decay)
+ self.shared_decoder = (
+ Decoder1d(cfg.shared_decoder, cfg.embed_dim)
+ if self.cfg.shared_decoder is not None
+ else None
+ )
+ if self.shared_decoder is not None:
+ self.shared_decoder.apply(self._init_weights)
+
+ self.recon_proj = None
+ if cfg.recon_loss > 0:
+ self.recon_proj = nn.Linear(cfg.embed_dim, cfg.embed_dim)
+
+ for pn, p in self.named_parameters():
+ if len(p.shape) == 1 or pn.endswith(".bias") or "alibi_scale" in pn:
+ p.optim_overrides = {"optimizer": {"weight_decay_scale": 0}}
+ if cfg.decoder_group and "decoder" in pn:
+ p.param_group = "decoder"
+
+ self.num_updates = 0
+
+ def _init_weights(self, m):
+
+ try:
+ from apex.normalization import FusedLayerNorm
+
+ fn = FusedLayerNorm
+ except:
+ fn = nn.LayerNorm
+
+ if isinstance(m, nn.Linear):
+ torch.nn.init.xavier_uniform_(m.weight)
+ if isinstance(m, nn.Linear) and m.bias is not None:
+ nn.init.constant_(m.bias, 0)
+ elif isinstance(m, nn.LayerNorm) or isinstance(m, fn):
+ if m.bias is not None:
+ nn.init.constant_(m.bias, 0)
+ if m.weight is not None:
+ nn.init.constant_(m.weight, 1.0)
+
+ @torch.no_grad()
+ def make_ema_teacher(self, ema_decay):
+ ema_config = EMAModuleConfig(
+ ema_decay=ema_decay,
+ ema_fp32=True,
+ log_norms=self.cfg.log_norms,
+ add_missing_params=False,
+ )
+
+ model_copy = self.make_target_model()
+
+ return EMAModule(
+ model_copy,
+ ema_config,
+ copy_model=False,
+ )
+
+ def make_target_model(self):
+ logger.info("making target model")
+
+ model_copy = Data2VecMultiModel(
+ self.cfg, self.modalities, skip_ema=True, task=self.task
+ )
+
+ if self.cfg.ema_encoder_only:
+ model_copy = model_copy.blocks
+ for p_s, p_t in zip(self.blocks.parameters(), model_copy.parameters()):
+ p_t.data.copy_(p_s.data)
+ else:
+ for p_s, p_t in zip(self.parameters(), model_copy.parameters()):
+ p_t.data.copy_(p_s.data)
+
+ for mod_enc in model_copy.modality_encoders.values():
+ mod_enc.decoder = None
+ if not mod_enc.modality_cfg.ema_local_encoder:
+ mod_enc.local_encoder = None
+ mod_enc.project_features = None
+
+ model_copy.requires_grad_(False)
+ return model_copy
+
+ def set_num_updates(self, num_updates):
+ super().set_num_updates(num_updates)
+
+ if self.ema is not None and (
+ (self.num_updates == 0 and num_updates > 1)
+ or self.num_updates >= num_updates
+ ):
+ pass
+ elif self.training and self.ema is not None:
+ ema_weight_decay = None
+ if self.cfg.ema_decay != self.cfg.ema_end_decay:
+ if num_updates >= self.cfg.ema_anneal_end_step:
+ decay = self.cfg.ema_end_decay
+ else:
+ decay = get_annealed_rate(
+ self.cfg.ema_decay,
+ self.cfg.ema_end_decay,
+ num_updates,
+ self.cfg.ema_anneal_end_step,
+ )
+ self.ema.set_decay(decay, weight_decay=ema_weight_decay)
+ if self.ema.get_decay() < 1:
+ self.ema.step(self.blocks if self.cfg.ema_encoder_only else self)
+
+ self.num_updates = num_updates
+
+ def state_dict(self, destination=None, prefix="", keep_vars=False):
+ state = super().state_dict(destination, prefix, keep_vars)
+
+ if self.ema is not None:
+ state[prefix + "_ema"] = self.ema.fp32_params
+
+ return state
+
+ def _load_from_state_dict(self, state_dict, prefix, *args, **kwargs):
+ k = prefix + "_ema"
+ if self.ema is not None:
+ assert k in state_dict
+ self.ema.restore(state_dict[k], True)
+ del state_dict[k]
+ elif k in state_dict:
+ del state_dict[k]
+
+ return super()._load_from_state_dict(state_dict, prefix, *args, **kwargs)
+
+ @classmethod
+ def build_model(cls, cfg: Data2VecMultiConfig, task=None):
+ """Build a new model instance."""
+ if task is None or not hasattr(task, "supported_modalities"):
+ modalities = (
+ [cfg.supported_modality]
+ if cfg.supported_modality is not None
+ else [
+ Modality.AUDIO,
+ Modality.IMAGE,
+ Modality.TEXT,
+ ]
+ )
+ else:
+ modalities = task.supported_modalities
+ return cls(cfg, modalities, task=task, skip_ema=cfg.skip_ema)
+
+ def forward(
+ self,
+ source,
+ target=None,
+ id=None,
+ mode=None,
+ padding_mask=None,
+ mask=True,
+ features_only=False,
+ force_remove_masked=False,
+ remove_extra_tokens=True,
+ precomputed_mask=None,
+ ):
+ if mode is None:
+ assert self.cfg.supported_modality is not None
+ mode = self.cfg.supported_modality
+
+ if isinstance(mode, Modality):
+ mode = mode.name
+
+ feature_extractor = self.modality_encoders[mode]
+
+ mask_seeds = None
+ if id is not None:
+ mask_seeds = MaskSeed(seed=self.cfg.seed, update=self.num_updates, ids=id)
+
+ extractor_out = feature_extractor(
+ source,
+ padding_mask,
+ mask,
+ remove_masked=not features_only or force_remove_masked,
+ clone_batch=self.cfg.clone_batch if not features_only else 1,
+ mask_seeds=mask_seeds,
+ precomputed_mask=precomputed_mask,
+ )
+
+ x = extractor_out["x"]
+ encoder_mask = extractor_out["encoder_mask"]
+ masked_padding_mask = extractor_out["padding_mask"]
+ masked_alibi_bias = extractor_out.get("alibi_bias", None)
+ alibi_scale = extractor_out.get("alibi_scale", None)
+
+ if self.dropout_input is not None:
+ x = self.dropout_input(x)
+
+ layer_results = []
+ for i, blk in enumerate(self.blocks):
+ if (
+ not self.training
+ or self.cfg.layerdrop == 0
+ or (np.random.random() > self.cfg.layerdrop)
+ ):
+ ab = masked_alibi_bias
+ if ab is not None and alibi_scale is not None:
+ scale = (
+ alibi_scale[i]
+ if alibi_scale.size(0) > 1
+ else alibi_scale.squeeze(0)
+ )
+ ab = ab * scale.type_as(ab)
+
+ x, lr = blk(
+ x,
+ padding_mask=masked_padding_mask,
+ alibi_bias=ab,
+ )
+ if features_only:
+ layer_results.append(lr)
+
+ if self.norm is not None:
+ x = self.norm(x)
+
+ if features_only:
+ if remove_extra_tokens:
+ x = x[:, feature_extractor.modality_cfg.num_extra_tokens :]
+ if masked_padding_mask is not None:
+ masked_padding_mask = masked_padding_mask[
+ :, feature_extractor.modality_cfg.num_extra_tokens :
+ ]
+
+ return {
+ "x": x,
+ "padding_mask": masked_padding_mask,
+ "layer_results": layer_results,
+ "mask": encoder_mask,
+ }
+
+ xs = []
+
+ if self.shared_decoder is not None:
+ dx = self.forward_decoder(
+ x,
+ feature_extractor,
+ self.shared_decoder,
+ encoder_mask,
+ )
+ xs.append(dx)
+ if feature_extractor.decoder is not None:
+ dx = self.forward_decoder(
+ x,
+ feature_extractor,
+ feature_extractor.decoder,
+ encoder_mask,
+ )
+ xs.append(dx)
+ orig_x = x
+
+ assert len(xs) > 0
+
+ p = next(self.ema.model.parameters())
+ device = x.device
+ dtype = x.dtype
+ ema_device = p.device
+ ema_dtype = p.dtype
+
+ if not self.cfg.ema_same_dtype:
+ dtype = ema_dtype
+
+ if ema_device != device or ema_dtype != dtype:
+ logger.info(f"adjusting ema dtype to {dtype} and device to {device}")
+ self.ema.model = self.ema.model.to(dtype=dtype, device=device)
+ ema_dtype = dtype
+
+ def to_device(d):
+ for k, p in d.items():
+ if isinstance(d[k], dict):
+ to_device(d[k])
+ else:
+ d[k] = p.to(device=device)
+
+ to_device(self.ema.fp32_params)
+ tm = self.ema.model
+
+ with torch.no_grad():
+ tm.eval()
+
+ if self.cfg.ema_encoder_only:
+ assert target is None
+ ema_input = extractor_out["local_features"]
+ ema_input = feature_extractor.contextualized_features(
+ ema_input.to(dtype=ema_dtype),
+ padding_mask,
+ mask=False,
+ remove_masked=False,
+ )
+ ema_blocks = tm
+ else:
+ ema_blocks = tm.blocks
+ if feature_extractor.modality_cfg.ema_local_encoder:
+ inp = (
+ target.to(dtype=ema_dtype)
+ if target is not None
+ else source.to(dtype=ema_dtype)
+ )
+ ema_input = tm.modality_encoders[mode](
+ inp,
+ padding_mask,
+ mask=False,
+ remove_masked=False,
+ )
+ else:
+ assert target is None
+ ema_input = extractor_out["local_features"]
+ ema_feature_enc = tm.modality_encoders[mode]
+ ema_input = ema_feature_enc.contextualized_features(
+ ema_input.to(dtype=ema_dtype),
+ padding_mask,
+ mask=False,
+ remove_masked=False,
+ )
+
+ ema_padding_mask = ema_input["padding_mask"]
+ ema_alibi_bias = ema_input.get("alibi_bias", None)
+ ema_alibi_scale = ema_input.get("alibi_scale", None)
+ ema_input = ema_input["x"]
+
+ y = []
+ ema_x = []
+ extra_tokens = feature_extractor.modality_cfg.num_extra_tokens
+ for i, blk in enumerate(ema_blocks):
+ ab = ema_alibi_bias
+ if ab is not None and alibi_scale is not None:
+ scale = (
+ ema_alibi_scale[i]
+ if ema_alibi_scale.size(0) > 1
+ else ema_alibi_scale.squeeze(0)
+ )
+ ab = ab * scale.type_as(ab)
+
+ ema_input, lr = blk(
+ ema_input,
+ padding_mask=ema_padding_mask,
+ alibi_bias=ab,
+ )
+ y.append(lr[:, extra_tokens:])
+ ema_x.append(ema_input[:, extra_tokens:])
+
+ y = self.make_targets(y, self.average_top_k_layers)
+ orig_targets = y
+
+ if self.cfg.clone_batch > 1:
+ y = y.repeat_interleave(self.cfg.clone_batch, 0)
+
+ masked = encoder_mask.mask.unsqueeze(-1)
+ masked_b = encoder_mask.mask.bool()
+ y = y[masked_b]
+
+ if xs[0].size(1) == masked_b.size(1):
+ xs = [x[masked_b] for x in xs]
+ else:
+ xs = [x.reshape(-1, x.size(-1)) for x in xs]
+
+ sample_size = masked.sum().long()
+
+ result = {
+ "losses": {},
+ "sample_size": sample_size,
+ }
+
+ sample_size = result["sample_size"]
+
+ if self.cfg.cls_loss > 0:
+ assert extra_tokens > 0
+ cls_target = orig_targets.mean(dim=1)
+ if self.cfg.clone_batch > 1:
+ cls_target = cls_target.repeat_interleave(self.cfg.clone_batch, 0)
+ cls_pred = x[:, extra_tokens - 1]
+ result["losses"]["cls"] = self.d2v_loss(cls_pred, cls_target) * (
+ self.cfg.cls_loss * sample_size
+ )
+
+ if self.cfg.recon_loss > 0:
+
+ with torch.no_grad():
+ target = feature_extractor.patchify(source)
+ mean = target.mean(dim=-1, keepdim=True)
+ var = target.var(dim=-1, keepdim=True)
+ target = (target - mean) / (var + 1.0e-6) ** 0.5
+
+ if self.cfg.clone_batch > 1:
+ target = target.repeat_interleave(self.cfg.clone_batch, 0)
+
+ if masked_b is not None:
+ target = target[masked_b]
+
+ recon = xs[0]
+ if self.recon_proj is not None:
+ recon = self.recon_proj(recon)
+
+ result["losses"]["recon"] = (
+ self.d2v_loss(recon, target.float()) * self.cfg.recon_loss
+ )
+
+ if self.cfg.d2v_loss > 0:
+ for i, x in enumerate(xs):
+ reg_loss = self.d2v_loss(x, y)
+ n = f"{mode}_regression_{i}" if len(xs) > 1 else f"{mode}_regression"
+ result["losses"][n] = reg_loss * self.cfg.d2v_loss
+
+ suffix = "" if len(self.modalities) == 1 else f"_{mode}"
+ with torch.no_grad():
+ if encoder_mask is not None:
+ result["masked_pct"] = 1 - (
+ encoder_mask.ids_keep.size(1) / encoder_mask.ids_restore.size(1)
+ )
+ for i, x in enumerate(xs):
+ n = f"pred_var{suffix}_{i}" if len(xs) > 1 else f"pred_var{suffix}"
+ result[n] = self.compute_var(x.float())
+ if self.ema is not None:
+ for k, v in self.ema.logs.items():
+ result[k] = v
+
+ y = y.float()
+ result[f"target_var{suffix}"] = self.compute_var(y)
+
+ if self.num_updates > 5000:
+ if result[f"target_var{suffix}"] < self.cfg.min_target_var:
+ logger.error(
+ f"target var is {result[f'target_var{suffix}'].item()} < {self.cfg.min_target_var}, exiting ({mode})"
+ )
+ raise Exception(
+ f"target var is {result[f'target_var{suffix}'].item()} < {self.cfg.min_target_var}, exiting ({mode})"
+ )
+
+ for k in result.keys():
+ if k.startswith("pred_var") and result[k] < self.cfg.min_pred_var:
+ logger.error(
+ f"{k} is {result[k].item()} < {self.cfg.min_pred_var}, exiting ({mode})"
+ )
+ raise Exception(
+ f"{k} is {result[k].item()} < {self.cfg.min_pred_var}, exiting ({mode})"
+ )
+
+ result["ema_decay"] = self.ema.get_decay() * 1000
+
+ return result
+
+ def forward_decoder(
+ self,
+ x,
+ feature_extractor,
+ decoder,
+ mask_info,
+ ):
+ x = feature_extractor.decoder_input(x, mask_info)
+ x = decoder(*x)
+
+ return x
+
+ def d2v_loss(self, x, y):
+ x = x.view(-1, x.size(-1)).float()
+ y = y.view(-1, x.size(-1))
+
+ if self.loss_beta == 0:
+ loss = F.mse_loss(x, y, reduction="none")
+ else:
+ loss = F.smooth_l1_loss(x, y, reduction="none", beta=self.loss_beta)
+
+ if self.loss_scale is not None:
+ scale = self.loss_scale
+ else:
+ scale = 1 / math.sqrt(x.size(-1))
+
+ reg_loss = loss * scale
+
+ return reg_loss
+
+ def make_targets(self, y, num_layers):
+
+ with torch.no_grad():
+ target_layer_results = y[-num_layers:]
+
+ permuted = False
+ if self.cfg.instance_norm_target_layer or self.cfg.batch_norm_target_layer:
+ target_layer_results = [
+ tl.transpose(1, 2) for tl in target_layer_results # BTC -> BCT
+ ]
+ permuted = True
+ if self.cfg.batch_norm_target_layer:
+ target_layer_results = [
+ F.batch_norm(
+ tl.float(), running_mean=None, running_var=None, training=True
+ )
+ for tl in target_layer_results
+ ]
+ if self.cfg.instance_norm_target_layer:
+ target_layer_results = [
+ F.instance_norm(tl.float()) for tl in target_layer_results
+ ]
+ if permuted:
+ target_layer_results = [
+ tl.transpose(1, 2) for tl in target_layer_results # BCT -> BTC
+ ]
+ if self.cfg.layer_norm_target_layer:
+ target_layer_results = [
+ F.layer_norm(tl.float(), tl.shape[-1:])
+ for tl in target_layer_results
+ ]
+
+ y = target_layer_results[0].float()
+ for tl in target_layer_results[1:]:
+ y.add_(tl.float())
+ y = y.div_(len(target_layer_results))
+
+ if self.cfg.layer_norm_targets:
+ y = F.layer_norm(y, y.shape[-1:])
+
+ if self.cfg.instance_norm_targets:
+ y = F.instance_norm(y.transpose(1, 2)).transpose(1, 2)
+
+ return y
+
+ @staticmethod
+ def compute_var(y):
+ y = y.view(-1, y.size(-1))
+ if dist.is_initialized():
+ zc = torch.tensor(y.size(0)).cuda()
+ zs = y.sum(dim=0)
+ zss = (y**2).sum(dim=0)
+
+ dist.all_reduce(zc)
+ dist.all_reduce(zs)
+ dist.all_reduce(zss)
+
+ var = zss / (zc - 1) - (zs**2) / (zc * (zc - 1))
+ return torch.sqrt(var + 1e-6).mean()
+ else:
+ return torch.sqrt(y.var(dim=0) + 1e-6).mean()
+
+ def extract_features(
+ self, source, mode=None, padding_mask=None, mask=False, remove_extra_tokens=True
+ ):
+ res = self.forward(
+ source,
+ mode=mode,
+ padding_mask=padding_mask,
+ mask=mask,
+ features_only=True,
+ remove_extra_tokens=remove_extra_tokens,
+ )
+ return res
+
+ def remove_pretraining_modules(self, modality=None, keep_decoder=False):
+ self.ema = None
+ self.cfg.clone_batch = 1
+ self.recon_proj = None
+
+ if not keep_decoder:
+ self.shared_decoder = None
+
+ modality = modality.lower() if modality is not None else None
+ for k in list(self.modality_encoders.keys()):
+ if modality is not None and k.lower() != modality:
+ del self.modality_encoders[k]
+ else:
+ self.modality_encoders[k].remove_pretraining_modules(
+ keep_decoder=keep_decoder
+ )
+ if not keep_decoder:
+ self.modality_encoders[k].decoder = None
diff --git a/fairseq/examples/data2vec/models/data2vec_audio.py b/fairseq/examples/data2vec/models/data2vec_audio.py
new file mode 100644
index 0000000000000000000000000000000000000000..261c2f104c4686700553b306189915e65cd29563
--- /dev/null
+++ b/fairseq/examples/data2vec/models/data2vec_audio.py
@@ -0,0 +1,537 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import logging
+import math
+from dataclasses import dataclass, field
+from typing import Optional
+
+from omegaconf import II
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.distributed as dist
+
+from fairseq.modules import EMAModule, EMAModuleConfig
+from fairseq.data.data_utils import compute_mask_indices
+from fairseq.models import BaseFairseqModel, register_model
+from fairseq.models.wav2vec import (
+ ConvFeatureExtractionModel,
+ Wav2Vec2Config,
+ TransformerEncoder,
+)
+from fairseq.modules import (
+ GradMultiply,
+ LayerNorm,
+)
+from fairseq.utils import index_put
+
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class Data2VecAudioConfig(Wav2Vec2Config):
+
+ loss_beta: float = field(
+ default=0, metadata={"help": "beta for smooth l1 loss. 0 means use l2 loss"}
+ )
+ loss_scale: Optional[float] = field(
+ default=None,
+ metadata={
+ "help": "scale the reconstruction loss by this constant. if None then scales by 1/sqrt(dim)"
+ },
+ )
+ average_top_k_layers: int = field(
+ default=8, metadata={"help": "how many layers to average"}
+ )
+
+ layer_norm_target_layer: bool = False
+ instance_norm_target_layer: bool = False
+ instance_norm_targets: bool = False
+ layer_norm_targets: bool = False
+ batch_norm_target_layer: bool = False
+ group_norm_target_layer: bool = False
+
+ ema_decay: float = field(default=0.999, metadata={"help": "initial ema decay rate"})
+ ema_end_decay: float = field(
+ default=0.9999, metadata={"help": "final ema decay rate"}
+ )
+
+ # when to finish annealing ema decay rate
+ ema_anneal_end_step: int = II("optimization.max_update")
+
+ ema_transformer_only: bool = field(
+ default=True,
+ metadata={"help": "whether to momentum update only the transformer"},
+ )
+ ema_layers_only: bool = field(
+ default=True,
+ metadata={"help": "whether to momentum update only the transformer layers"},
+ )
+
+ max_update: int = II("optimization.max_update")
+
+ min_target_var: float = field(
+ default=0.1, metadata={"help": "stop training if target var falls below this"}
+ )
+ min_pred_var: float = field(
+ default=0.01,
+ metadata={"help": "stop training if prediction var falls below this"},
+ )
+
+
+def get_annealed_rate(start, end, curr_step, total_steps):
+ r = end - start
+ pct_remaining = 1 - curr_step / total_steps
+ return end - r * pct_remaining
+
+
+@register_model("data2vec_audio", dataclass=Data2VecAudioConfig)
+class Data2VecAudioModel(BaseFairseqModel):
+ def __init__(self, cfg: Data2VecAudioConfig):
+ super().__init__()
+ self.cfg = cfg
+
+ feature_enc_layers = eval(cfg.conv_feature_layers)
+ self.extractor_embed = feature_enc_layers[-1][0]
+
+ self.ema = None
+ self.embed = cfg.encoder_embed_dim
+
+ self.average_top_k_layers = cfg.average_top_k_layers
+ self.loss_beta = cfg.loss_beta
+ self.loss_scale = cfg.loss_scale
+
+ self.feature_extractor = ConvFeatureExtractionModel(
+ conv_layers=feature_enc_layers,
+ dropout=0.0,
+ mode=cfg.extractor_mode,
+ conv_bias=cfg.conv_bias,
+ )
+
+ self.post_extract_proj = nn.Linear(self.extractor_embed, cfg.encoder_embed_dim)
+
+ self.mask_prob = cfg.mask_prob
+ self.mask_selection = cfg.mask_selection
+ self.mask_other = cfg.mask_other
+ self.mask_length = cfg.mask_length
+ self.no_mask_overlap = cfg.no_mask_overlap
+ self.mask_min_space = cfg.mask_min_space
+
+ self.mask_channel_prob = cfg.mask_channel_prob
+ self.mask_channel_before = cfg.mask_channel_before
+ self.mask_channel_selection = cfg.mask_channel_selection
+ self.mask_channel_other = cfg.mask_channel_other
+ self.mask_channel_length = cfg.mask_channel_length
+ self.no_mask_channel_overlap = cfg.no_mask_channel_overlap
+ self.mask_channel_min_space = cfg.mask_channel_min_space
+
+ self.dropout_input = nn.Dropout(cfg.dropout_input)
+ self.dropout_features = nn.Dropout(cfg.dropout_features)
+
+ self.feature_grad_mult = cfg.feature_grad_mult
+
+ self.mask_emb = nn.Parameter(
+ torch.FloatTensor(cfg.encoder_embed_dim).uniform_()
+ )
+
+ self.encoder = TransformerEncoder(cfg)
+ self.layer_norm = LayerNorm(self.extractor_embed)
+
+ self.final_proj = nn.Linear(self.embed, self.embed)
+
+ self.num_updates = 0
+
+ def make_ema_teacher(self):
+ ema_config = EMAModuleConfig(
+ ema_decay=self.cfg.ema_decay,
+ ema_fp32=True,
+ )
+ skip_keys = set()
+ if self.cfg.ema_layers_only:
+ self.cfg.ema_transformer_only = True
+ for k, _ in self.encoder.pos_conv.named_parameters():
+ skip_keys.add(f"pos_conv.{k}")
+
+ self.ema = EMAModule(
+ self.encoder if self.cfg.ema_transformer_only else self,
+ ema_config,
+ skip_keys=skip_keys,
+ )
+
+ def set_num_updates(self, num_updates):
+ super().set_num_updates(num_updates)
+
+ if self.ema is None and self.final_proj is not None:
+ logger.info(f"making ema teacher")
+ self.make_ema_teacher()
+ elif self.training and self.ema is not None:
+ if self.cfg.ema_decay != self.cfg.ema_end_decay:
+ if num_updates >= self.cfg.ema_anneal_end_step:
+ decay = self.cfg.ema_end_decay
+ else:
+ decay = get_annealed_rate(
+ self.cfg.ema_decay,
+ self.cfg.ema_end_decay,
+ num_updates,
+ self.cfg.ema_anneal_end_step,
+ )
+ self.ema.set_decay(decay)
+ if self.ema.get_decay() < 1:
+ self.ema.step(self.encoder if self.cfg.ema_transformer_only else self)
+
+ self.num_updates = num_updates
+
+ def state_dict(self, destination=None, prefix="", keep_vars=False):
+ state = super().state_dict(destination, prefix, keep_vars)
+
+ if self.ema is not None:
+ state[prefix + "_ema"] = self.ema.fp32_params
+
+ return state
+
+ def _load_from_state_dict(self, state_dict, prefix, *args, **kwargs):
+ if self.ema is not None:
+ k = prefix + "_ema"
+ assert k in state_dict
+ self.ema.restore(state_dict[k], True)
+ del state_dict[k]
+ return super()._load_from_state_dict(state_dict, prefix, *args, **kwargs)
+
+ @classmethod
+ def build_model(cls, cfg: Data2VecAudioConfig, task=None):
+ """Build a new model instance."""
+
+ return cls(cfg)
+
+ def apply_mask(
+ self,
+ x,
+ padding_mask,
+ mask_indices=None,
+ mask_channel_indices=None,
+ ):
+ B, T, C = x.shape
+
+ if self.mask_channel_prob > 0 and self.mask_channel_before:
+ mask_channel_indices = compute_mask_indices(
+ (B, C),
+ None,
+ self.mask_channel_prob,
+ self.mask_channel_length,
+ self.mask_channel_selection,
+ self.mask_channel_other,
+ no_overlap=self.no_mask_channel_overlap,
+ min_space=self.mask_channel_min_space,
+ )
+ mask_channel_indices = (
+ torch.from_numpy(mask_channel_indices)
+ .to(x.device)
+ .unsqueeze(1)
+ .expand(-1, T, -1)
+ )
+ x[mask_channel_indices] = 0
+
+ if self.mask_prob > 0:
+ if mask_indices is None:
+ mask_indices = compute_mask_indices(
+ (B, T),
+ padding_mask,
+ self.mask_prob,
+ self.mask_length,
+ self.mask_selection,
+ self.mask_other,
+ min_masks=1,
+ no_overlap=self.no_mask_overlap,
+ min_space=self.mask_min_space,
+ require_same_masks=self.cfg.require_same_masks,
+ mask_dropout=self.cfg.mask_dropout,
+ )
+ mask_indices = torch.from_numpy(mask_indices).to(x.device)
+ x = index_put(x, mask_indices, self.mask_emb)
+ else:
+ mask_indices = None
+
+ if self.mask_channel_prob > 0 and not self.mask_channel_before:
+ if mask_channel_indices is None:
+ mask_channel_indices = compute_mask_indices(
+ (B, C),
+ None,
+ self.mask_channel_prob,
+ self.mask_channel_length,
+ self.mask_channel_selection,
+ self.mask_channel_other,
+ no_overlap=self.no_mask_channel_overlap,
+ min_space=self.mask_channel_min_space,
+ )
+ mask_channel_indices = (
+ torch.from_numpy(mask_channel_indices)
+ .to(x.device)
+ .unsqueeze(1)
+ .expand(-1, T, -1)
+ )
+ x = index_put(x, mask_channel_indices, 0)
+
+ return x, mask_indices
+
+ def _get_feat_extract_output_lengths(self, input_lengths: torch.LongTensor):
+ """
+ Computes the output length of the convolutional layers
+ """
+
+ def _conv_out_length(input_length, kernel_size, stride):
+ return torch.floor((input_length - kernel_size) / stride + 1)
+
+ conv_cfg_list = eval(self.cfg.conv_feature_layers)
+
+ for i in range(len(conv_cfg_list)):
+ input_lengths = _conv_out_length(
+ input_lengths, conv_cfg_list[i][1], conv_cfg_list[i][2]
+ )
+
+ return input_lengths.to(torch.long)
+
+ def forward(
+ self,
+ source,
+ padding_mask=None,
+ mask=True,
+ features_only=False,
+ layer=None,
+ mask_indices=None,
+ mask_channel_indices=None,
+ padding_count=None,
+ ):
+ features = source
+
+ if self.feature_grad_mult > 0:
+ features = self.feature_extractor(features)
+ if self.feature_grad_mult != 1.0:
+ features = GradMultiply.apply(features, self.feature_grad_mult)
+ else:
+ with torch.no_grad():
+ features = self.feature_extractor(features)
+
+ features = features.transpose(1, 2)
+
+ features = self.layer_norm(features)
+
+ orig_padding_mask = padding_mask
+
+ if padding_mask is not None and padding_mask.any():
+ input_lengths = (1 - padding_mask.long()).sum(-1)
+ # apply conv formula to get real output_lengths
+ output_lengths = self._get_feat_extract_output_lengths(input_lengths)
+
+ padding_mask = torch.zeros(
+ features.shape[:2], dtype=features.dtype, device=features.device
+ )
+
+ # these two operations makes sure that all values
+ # before the output lengths indices are attended to
+ padding_mask[
+ (
+ torch.arange(padding_mask.shape[0], device=padding_mask.device),
+ output_lengths - 1,
+ )
+ ] = 1
+ padding_mask = (1 - padding_mask.flip([-1]).cumsum(-1).flip([-1])).bool()
+ else:
+ padding_mask = None
+
+ if self.post_extract_proj is not None:
+ features = self.post_extract_proj(features)
+
+ pre_encoder_features = None
+ if self.cfg.ema_transformer_only:
+ pre_encoder_features = features.clone()
+
+ features = self.dropout_input(features)
+
+ if mask:
+ x, mask_indices = self.apply_mask(
+ features,
+ padding_mask,
+ mask_indices=mask_indices,
+ mask_channel_indices=mask_channel_indices,
+ )
+ else:
+ x = features
+ mask_indices = None
+
+ x, layer_results = self.encoder(
+ x,
+ padding_mask=padding_mask,
+ layer=layer,
+ )
+
+ if features_only:
+ return {
+ "x": x,
+ "padding_mask": padding_mask,
+ "layer_results": layer_results,
+ }
+
+ result = {
+ "losses": {},
+ }
+
+ with torch.no_grad():
+ self.ema.model.eval()
+
+ if self.cfg.ema_transformer_only:
+ y, layer_results = self.ema.model.extract_features(
+ pre_encoder_features,
+ padding_mask=padding_mask,
+ min_layer=self.cfg.encoder_layers - self.average_top_k_layers,
+ )
+ y = {
+ "x": y,
+ "padding_mask": padding_mask,
+ "layer_results": layer_results,
+ }
+ else:
+ y = self.ema.model.extract_features(
+ source=source,
+ padding_mask=orig_padding_mask,
+ mask=False,
+ )
+
+ target_layer_results = [l[2] for l in y["layer_results"]]
+
+ permuted = False
+ if self.cfg.instance_norm_target_layer or self.cfg.batch_norm_target_layer:
+ target_layer_results = [
+ tl.permute(1, 2, 0) for tl in target_layer_results # TBC -> BCT
+ ]
+ permuted = True
+
+ if self.cfg.batch_norm_target_layer:
+ target_layer_results = [
+ F.batch_norm(
+ tl.float(), running_mean=None, running_var=None, training=True
+ )
+ for tl in target_layer_results
+ ]
+
+ if self.cfg.instance_norm_target_layer:
+ target_layer_results = [
+ F.instance_norm(tl.float()) for tl in target_layer_results
+ ]
+
+ if permuted:
+ target_layer_results = [
+ tl.transpose(1, 2) for tl in target_layer_results # BCT -> BTC
+ ]
+
+ if self.cfg.group_norm_target_layer:
+ target_layer_results = [
+ F.layer_norm(tl.float(), tl.shape[-2:])
+ for tl in target_layer_results
+ ]
+
+ if self.cfg.layer_norm_target_layer:
+ target_layer_results = [
+ F.layer_norm(tl.float(), tl.shape[-1:])
+ for tl in target_layer_results
+ ]
+
+ y = sum(target_layer_results) / len(target_layer_results)
+
+ if self.cfg.layer_norm_targets:
+ y = F.layer_norm(y.float(), y.shape[-1:])
+
+ if self.cfg.instance_norm_targets:
+ y = F.instance_norm(y.float().transpose(1, 2)).transpose(1, 2)
+
+ if not permuted:
+ y = y.transpose(0, 1)
+
+ y = y[mask_indices]
+
+ x = x[mask_indices]
+ x = self.final_proj(x)
+
+ sz = x.size(-1)
+
+ if self.loss_beta == 0:
+ loss = F.mse_loss(x.float(), y.float(), reduction="none").sum(dim=-1)
+ else:
+ loss = F.smooth_l1_loss(
+ x.float(), y.float(), reduction="none", beta=self.loss_beta
+ ).sum(dim=-1)
+
+ if self.loss_scale is not None:
+ scale = self.loss_scale
+ else:
+ scale = 1 / math.sqrt(sz)
+
+ result["losses"]["regression"] = loss.sum() * scale
+
+ if "sample_size" not in result:
+ result["sample_size"] = loss.numel()
+
+ with torch.no_grad():
+ result["target_var"] = self.compute_var(y)
+ result["pred_var"] = self.compute_var(x.float())
+
+ if self.num_updates > 5000 and result["target_var"] < self.cfg.min_target_var:
+ logger.error(
+ f"target var is {result['target_var'].item()} < {self.cfg.min_target_var}, exiting"
+ )
+ raise Exception(
+ f"target var is {result['target_var'].item()} < {self.cfg.min_target_var}, exiting"
+ )
+ if self.num_updates > 5000 and result["pred_var"] < self.cfg.min_pred_var:
+ logger.error(
+ f"pred var is {result['pred_var'].item()} < {self.cfg.min_pred_var}, exiting"
+ )
+ raise Exception(
+ f"pred var is {result['pred_var'].item()} < {self.cfg.min_pred_var}, exiting"
+ )
+
+ if self.ema is not None:
+ result["ema_decay"] = self.ema.get_decay() * 1000
+
+ return result
+
+ @staticmethod
+ def compute_var(y):
+ y = y.view(-1, y.size(-1))
+ if dist.is_initialized():
+ zc = torch.tensor(y.size(0)).cuda()
+ zs = y.sum(dim=0)
+ zss = (y ** 2).sum(dim=0)
+
+ dist.all_reduce(zc)
+ dist.all_reduce(zs)
+ dist.all_reduce(zss)
+
+ var = zss / (zc - 1) - (zs ** 2) / (zc * (zc - 1))
+ return torch.sqrt(var + 1e-6).mean()
+ else:
+ return torch.sqrt(y.var(dim=0) + 1e-6).mean()
+
+ def extract_features(
+ self, source, padding_mask, mask=False, layer=None
+ ):
+ res = self.forward(
+ source,
+ padding_mask,
+ mask=mask,
+ features_only=True,
+ layer=layer,
+ )
+ return res
+
+ def remove_pretraining_modules(self, last_layer=None):
+ self.final_proj = None
+ self.ema = None
+ if last_layer is not None:
+ self.encoder.layers = nn.ModuleList(
+ l for i, l in enumerate(self.encoder.layers) if i <= last_layer
+ )
diff --git a/fairseq/examples/data2vec/models/data2vec_image_classification.py b/fairseq/examples/data2vec/models/data2vec_image_classification.py
new file mode 100644
index 0000000000000000000000000000000000000000..851c9ce45531d38a28f688ff4d9fa8a45e705544
--- /dev/null
+++ b/fairseq/examples/data2vec/models/data2vec_image_classification.py
@@ -0,0 +1,143 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+# The code in this file is adapted from the BeiT implementation which can be found here:
+# https://github.com/microsoft/unilm/tree/master/beit
+
+import logging
+
+from dataclasses import dataclass
+from typing import Any
+
+from omegaconf import II, MISSING
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from fairseq import checkpoint_utils, tasks
+
+from fairseq.dataclass import FairseqDataclass
+from fairseq.models import BaseFairseqModel, register_model
+
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class Data2VecImageClassificationConfig(FairseqDataclass):
+ model_path: str = MISSING
+ no_pretrained_weights: bool = False
+ num_classes: int = 1000
+ mixup: float = 0.8
+ cutmix: float = 1.0
+ label_smoothing: float = 0.1
+
+ pretrained_model_args: Any = None
+ data: str = II("task.data")
+
+
+@register_model(
+ "data2vec_image_classification", dataclass=Data2VecImageClassificationConfig
+)
+class Data2VecImageClassificationModel(BaseFairseqModel):
+ def __init__(self, cfg: Data2VecImageClassificationConfig):
+ super().__init__()
+ self.cfg = cfg
+
+ if cfg.pretrained_model_args is None:
+ state = checkpoint_utils.load_checkpoint_to_cpu(cfg.model_path, {})
+ pretrained_args = state.get("cfg", None)
+ pretrained_args.criterion = None
+ pretrained_args.lr_scheduler = None
+ cfg.pretrained_model_args = pretrained_args
+
+ logger.info(pretrained_args)
+ else:
+ state = None
+ pretrained_args = cfg.pretrained_model_args
+
+ pretrained_args.task.data = cfg.data
+ task = tasks.setup_task(pretrained_args.task)
+ model = task.build_model(pretrained_args.model, from_checkpoint=True)
+
+ model.remove_pretraining_modules()
+
+ self.model = model
+
+ if state is not None and not cfg.no_pretrained_weights:
+ self.load_model_weights(state, model, cfg)
+
+ self.fc_norm = nn.LayerNorm(pretrained_args.model.embed_dim)
+ self.head = nn.Linear(pretrained_args.model.embed_dim, cfg.num_classes)
+
+ self.head.weight.data.mul_(1e-3)
+ self.head.bias.data.mul_(1e-3)
+
+ self.mixup_fn = None
+
+ if cfg.mixup > 0 or cfg.cutmix > 0:
+ from timm.data import Mixup
+
+ self.mixup_fn = Mixup(
+ mixup_alpha=cfg.mixup,
+ cutmix_alpha=cfg.cutmix,
+ cutmix_minmax=None,
+ prob=1.0,
+ switch_prob=0.5,
+ mode="batch",
+ label_smoothing=cfg.label_smoothing,
+ num_classes=cfg.num_classes,
+ )
+
+ def load_model_weights(self, state, model, cfg):
+ if "_ema" in state["model"]:
+ del state["model"]["_ema"]
+ model.load_state_dict(state["model"], strict=True)
+
+ @classmethod
+ def build_model(cls, cfg: Data2VecImageClassificationConfig, task=None):
+ """Build a new model instance."""
+
+ return cls(cfg)
+
+ def forward(
+ self,
+ img,
+ label=None,
+ ):
+ if self.training and self.mixup_fn is not None and label is not None:
+ img, label = self.mixup_fn(img, label)
+
+ x = self.model(img, mask=False)
+ x = x[:, 1:]
+ x = self.fc_norm(x.mean(1))
+ x = self.head(x)
+
+ if label is None:
+ return x
+
+ if self.training and self.mixup_fn is not None:
+ loss = -label * F.log_softmax(x.float(), dim=-1)
+ else:
+ loss = F.cross_entropy(
+ x.float(),
+ label,
+ label_smoothing=self.cfg.label_smoothing if self.training else 0,
+ reduction="none",
+ )
+
+ result = {
+ "losses": {"regression": loss},
+ "sample_size": img.size(0),
+ }
+
+ if not self.training:
+ with torch.no_grad():
+ pred = x.argmax(-1)
+ correct = (pred == label).sum()
+ result["correct"] = correct
+
+ return result
diff --git a/fairseq/examples/data2vec/models/data2vec_text.py b/fairseq/examples/data2vec/models/data2vec_text.py
new file mode 100644
index 0000000000000000000000000000000000000000..cb3c8b383a45ec67d0f8ec2939fa71d5d076a107
--- /dev/null
+++ b/fairseq/examples/data2vec/models/data2vec_text.py
@@ -0,0 +1,517 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+from dataclasses import dataclass, field
+from typing import Optional
+import logging
+import math
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from omegaconf import II
+
+from fairseq.dataclass import FairseqDataclass
+from fairseq.modules import EMAModule, EMAModuleConfig
+from fairseq.models import (
+ FairseqEncoder,
+ FairseqEncoderModel,
+ register_model,
+)
+from fairseq.models.roberta.model import RobertaLMHead, RobertaClassificationHead
+from fairseq.models.transformer import TransformerEncoder, TransformerConfig
+from fairseq.modules.transformer_sentence_encoder import init_bert_params
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class Data2VecTextConfig(FairseqDataclass):
+ max_positions: int = II("task.tokens_per_sample")
+
+ head_layers: int = 1
+
+ transformer: TransformerConfig = TransformerConfig()
+
+ load_checkpoint_heads: bool = field(
+ default=False,
+ metadata={"help": "(re-)register and load heads when loading checkpoints"},
+ )
+
+ loss_beta: float = field(
+ default=0, metadata={"help": "beta for smooth l1 loss. 0 means use l2 loss"}
+ )
+ loss_scale: Optional[float] = field(
+ default=None,
+ metadata={
+ "help": "scale the reconstruction loss by this constant. if None then scales by 1/sqrt(dim)"
+ },
+ )
+ average_top_k_layers: int = field(
+ default=8, metadata={"help": "how many layers to average"}
+ )
+
+ layer_norm_target_layer: bool = False
+ instance_norm_target_layer: bool = False
+ batch_norm_target_layer: bool = False
+ instance_norm_targets: bool = False
+ layer_norm_targets: bool = False
+
+ ema_decay: float = field(default=0.999, metadata={"help": "initial ema decay rate"})
+ ema_end_decay: float = field(
+ default=0.9999, metadata={"help": "final ema decay rate"}
+ )
+
+ # when to finish annealing ema decay rate
+ ema_anneal_end_step: int = II("optimization.max_update")
+
+ ema_transformer_layers_only: bool = field(
+ default=True,
+ metadata={"help": "whether to momentum update only the transformer layers"},
+ )
+
+
+def get_annealed_rate(start, end, curr_step, total_steps):
+ r = end - start
+ pct_remaining = 1 - curr_step / total_steps
+ return end - r * pct_remaining
+
+
+@register_model("data2vec_text", dataclass=Data2VecTextConfig)
+class Data2VecTextModel(FairseqEncoderModel):
+ def __init__(self, cfg: Data2VecTextConfig, encoder):
+ super().__init__(encoder)
+ self.cfg = cfg
+
+ # We follow BERT's random weight initialization
+ self.apply(init_bert_params)
+
+ self.classification_heads = nn.ModuleDict()
+
+ @classmethod
+ def build_model(cls, cfg, task):
+ """Build a new model instance."""
+
+ encoder = Data2VecTextEncoder(cfg, task.source_dictionary, task.cfg.data)
+
+ return cls(cfg, encoder)
+
+ def forward(
+ self,
+ src_tokens,
+ target_tokens=None,
+ features_only=False,
+ return_all_hiddens=False,
+ classification_head_name=None,
+ **kwargs,
+ ):
+ if classification_head_name is not None:
+ features_only = True
+
+ res = self.encoder(
+ src_tokens, target_tokens, features_only, return_all_hiddens, **kwargs
+ )
+
+ if isinstance(res, tuple):
+ x, extra = res
+ else:
+ return res
+
+ if classification_head_name is not None:
+ x = self.classification_heads[classification_head_name](x)
+ return x, extra
+
+ def get_normalized_probs(self, net_output, log_probs, sample=None):
+ """Get normalized probabilities (or log probs) from a net's output."""
+ logits = net_output[0].float()
+ if log_probs:
+ return F.log_softmax(logits, dim=-1)
+ else:
+ return F.softmax(logits, dim=-1)
+
+ def register_classification_head(
+ self, name, num_classes=None, inner_dim=None, **kwargs
+ ):
+ """Register a classification head."""
+ if name in self.classification_heads:
+ prev_num_classes = self.classification_heads[name].out_proj.out_features
+ prev_inner_dim = self.classification_heads[name].dense.out_features
+ if num_classes != prev_num_classes or inner_dim != prev_inner_dim:
+ logger.warning(
+ 're-registering head "{}" with num_classes {} (prev: {}) '
+ "and inner_dim {} (prev: {})".format(
+ name, num_classes, prev_num_classes, inner_dim, prev_inner_dim
+ )
+ )
+ self.classification_heads[name] = RobertaClassificationHead(
+ input_dim=self.cfg.transformer.encoder.embed_dim,
+ inner_dim=inner_dim or self.cfg.transformer.encoder.embed_dim,
+ num_classes=num_classes,
+ activation_fn="tanh",
+ pooler_dropout=0,
+ )
+
+ @property
+ def supported_targets(self):
+ return {"self"}
+
+ def upgrade_state_dict_named(self, state_dict, name):
+ prefix = name + "." if name != "" else ""
+
+ # rename decoder -> encoder before upgrading children modules
+ for k in list(state_dict.keys()):
+ if k.startswith(prefix + "decoder"):
+ new_k = prefix + "encoder" + k[len(prefix + "decoder") :]
+ state_dict[new_k] = state_dict[k]
+ del state_dict[k]
+
+ # rename emb_layer_norm -> layernorm_embedding
+ for k in list(state_dict.keys()):
+ if ".emb_layer_norm." in k:
+ new_k = k.replace(".emb_layer_norm.", ".layernorm_embedding.")
+ state_dict[new_k] = state_dict[k]
+ del state_dict[k]
+
+ if self.encoder.regression_head is not None:
+ if ".lm_head." in k:
+ new_k = k.replace(".lm_head.", ".regression_head.")
+ state_dict[new_k] = state_dict[k]
+ del state_dict[k]
+ else:
+ if ".regression_head." in k:
+ del state_dict[k]
+
+ # upgrade children modules
+ super().upgrade_state_dict_named(state_dict, name)
+
+ # Handle new classification heads present in the state dict.
+ current_head_names = (
+ []
+ if not hasattr(self, "classification_heads")
+ or self.classification_heads is None
+ else self.classification_heads.keys()
+ )
+ keys_to_delete = []
+ for k in state_dict.keys():
+ if not k.startswith(prefix + "classification_heads."):
+ continue
+
+ head_name = k[len(prefix + "classification_heads.") :].split(".")[0]
+ num_classes = state_dict[
+ prefix + "classification_heads." + head_name + ".out_proj.weight"
+ ].size(0)
+ inner_dim = state_dict[
+ prefix + "classification_heads." + head_name + ".dense.weight"
+ ].size(0)
+
+ if self.cfg.load_checkpoint_heads:
+ if head_name not in current_head_names:
+ self.register_classification_head(head_name, num_classes, inner_dim)
+ else:
+ if head_name not in current_head_names:
+ logger.warning(
+ "deleting classification head ({}) from checkpoint "
+ "not present in current model: {}".format(head_name, k)
+ )
+ keys_to_delete.append(k)
+ elif (
+ num_classes
+ != self.classification_heads[head_name].out_proj.out_features
+ or inner_dim
+ != self.classification_heads[head_name].dense.out_features
+ ):
+ logger.warning(
+ "deleting classification head ({}) from checkpoint "
+ "with different dimensions than current model: {}".format(
+ head_name, k
+ )
+ )
+ keys_to_delete.append(k)
+ for k in keys_to_delete:
+ del state_dict[k]
+
+ # Copy any newly-added classification heads into the state dict
+ # with their current weights.
+ if (
+ hasattr(self, "classification_heads")
+ and self.classification_heads is not None
+ and len(self.classification_heads) > 0
+ ):
+ cur_state = self.classification_heads.state_dict()
+ for k, v in cur_state.items():
+ if prefix + "classification_heads." + k not in state_dict:
+ logger.info("Overwriting " + prefix + "classification_heads." + k)
+ state_dict[prefix + "classification_heads." + k] = v
+
+ for k in list(state_dict.keys()):
+ if k.startswith(prefix + "encoder.lm_head.") or k.startswith(
+ prefix + "encoder.emb_head."
+ ):
+ del state_dict[k]
+
+ self.encoder.lm_head = None
+
+ if self.encoder.target_model is None:
+ for k in list(state_dict.keys()):
+ if k.startswith(prefix + "encoder.target_model."):
+ del state_dict[k]
+
+ if (self.encoder.ema is None) and (prefix + "encoder._ema" in state_dict):
+ del state_dict[prefix + "encoder._ema"]
+
+ def remove_pretraining_modules(self, last_layer=None):
+ self.encoder.lm_head = None
+ self.encoder.regression_head = None
+ self.encoder.ema = None
+ self.classification_heads = None
+
+ if last_layer is not None:
+ self.encoder.sentence_encoder.layers = nn.ModuleList(
+ l
+ for i, l in enumerate(self.encoder.sentence_encoder.layers)
+ if i <= last_layer
+ )
+ self.encoder.sentence_encoder.layer_norm = None
+
+
+class Data2VecTextEncoder(FairseqEncoder):
+ def __init__(self, cfg: Data2VecTextConfig, dictionary, task_data):
+ super().__init__(dictionary)
+
+ self.cfg = cfg
+
+ embed_tokens = self.build_embedding(
+ len(dictionary), cfg.transformer.encoder.embed_dim, dictionary.pad()
+ )
+
+ self.sentence_encoder = self.build_encoder(cfg, dictionary, embed_tokens)
+ self.mask_idx = dictionary.index("")
+ assert self.mask_idx != dictionary.unk(), dictionary.symbols
+
+ self.ema = None
+ self.average_top_k_layers = cfg.average_top_k_layers
+ self.loss_scale = cfg.loss_scale
+
+ assert self.cfg.head_layers >= 1
+
+ embed_dim = cfg.transformer.encoder.embed_dim
+ curr_dim = embed_dim
+ projs = []
+ for i in range(self.cfg.head_layers - 1):
+ next_dim = embed_dim * 2 if i == 0 else curr_dim
+ projs.append(nn.Linear(curr_dim, next_dim))
+ projs.append(nn.GELU())
+ curr_dim = next_dim
+
+ projs.append(nn.Linear(curr_dim, embed_dim))
+ self.regression_head = nn.Sequential(*projs)
+
+ self.num_updates = 0
+
+ def build_embedding(self, vocab_size, embedding_dim, padding_idx):
+ return nn.Embedding(vocab_size, embedding_dim, padding_idx)
+
+ def build_encoder(self, cfg, dictionary, embed_tokens):
+ encoder = TransformerEncoder(cfg.transformer, dictionary, embed_tokens, return_fc=True)
+ encoder.apply(init_bert_params)
+ return encoder
+
+ def build_lm_head(self, embed_dim, output_dim, activation_fn, weight):
+ return RobertaLMHead(embed_dim, output_dim, activation_fn, weight)
+
+ def make_ema_teacher(self):
+ ema_config = EMAModuleConfig(
+ ema_decay=self.cfg.ema_decay,
+ ema_fp32=True,
+ )
+ skip_keys = set()
+ if self.cfg.ema_transformer_layers_only:
+ for k, _ in self.sentence_encoder.embed_positions.named_parameters():
+ skip_keys.add(f"embed_tokens.{k}")
+ for k, _ in self.sentence_encoder.embed_positions.named_parameters():
+ skip_keys.add(f"embed_positions.{k}")
+ if self.sentence_encoder.layernorm_embedding is not None:
+ for (
+ k,
+ _,
+ ) in self.sentence_encoder.layernorm_embedding.named_parameters():
+ skip_keys.add(f"layernorm_embedding.{k}")
+ if self.sentence_encoder.layer_norm is not None:
+ for k, _ in self.sentence_encoder.layer_norm.named_parameters():
+ skip_keys.add(f"layernorm_embedding.{k}")
+
+ self.ema = EMAModule(
+ self.sentence_encoder,
+ ema_config,
+ skip_keys=skip_keys,
+ )
+
+ def set_num_updates(self, num_updates):
+ super().set_num_updates(num_updates)
+
+ if self.ema is None and self.regression_head is not None:
+ logger.info(f"making ema teacher")
+ self.make_ema_teacher()
+ elif self.training and self.ema is not None:
+ if self.cfg.ema_decay != self.cfg.ema_end_decay:
+ if num_updates >= self.cfg.ema_anneal_end_step:
+ decay = self.cfg.ema_end_decay
+ else:
+ decay = get_annealed_rate(
+ self.cfg.ema_decay,
+ self.cfg.ema_end_decay,
+ num_updates,
+ self.cfg.ema_anneal_end_step,
+ )
+ self.ema.set_decay(decay)
+ if self.ema.get_decay() < 1:
+ self.ema.step(self.sentence_encoder)
+
+ def state_dict(self, destination=None, prefix="", keep_vars=False):
+ state = super().state_dict(destination, prefix, keep_vars)
+ if self.ema is not None:
+ state[prefix + "_ema"] = self.ema.fp32_params
+ return state
+
+ def _load_from_state_dict(self, state_dict, prefix, *args, **kwargs):
+ if self.ema is not None:
+ k = prefix + "_ema"
+ assert k in state_dict
+ self.ema.restore(state_dict[k], True)
+ del state_dict[k]
+ return super()._load_from_state_dict(state_dict, prefix, *args, **kwargs)
+
+ def forward(
+ self,
+ src_tokens,
+ target_tokens=None,
+ features_only=False,
+ return_all_hiddens=False,
+ masked_tokens=None,
+ **unused,
+ ):
+ """
+ Args:
+ src_tokens (LongTensor): input tokens of shape `(batch, src_len)`
+ features_only (bool, optional): skip LM head and just return
+ features. If True, the output will be of shape
+ `(batch, src_len, embed_dim)`.
+ return_all_hiddens (bool, optional): also return all of the
+ intermediate hidden states (default: False).
+
+ Returns:
+ tuple:
+ - the LM output of shape `(batch, src_len, vocab)`
+ - a dictionary of additional data, where 'inner_states'
+ is a list of hidden states. Note that the hidden
+ states have shape `(src_len, batch, vocab)`.
+ """
+
+ x, extra = self.extract_features(
+ src_tokens, return_all_hiddens=return_all_hiddens
+ )
+
+ if features_only:
+ return x, extra
+
+ assert target_tokens is not None
+
+ with torch.no_grad():
+ # use EMA parameter as the teacher
+ self.ema.model.eval()
+
+ encoder_out = self.ema.model(
+ target_tokens,
+ return_all_hiddens=True,
+ )
+ y = encoder_out["fc_results"]
+
+ y = y[-self.average_top_k_layers :]
+
+ permuted = False
+ if self.cfg.instance_norm_target_layer or self.cfg.batch_norm_target_layer:
+ y = [tl.permute(1, 2, 0) for tl in y] # TBC -> BCT
+ permuted = True
+
+ if self.cfg.batch_norm_target_layer:
+ y = [
+ F.batch_norm(
+ tl.float(), running_mean=None, running_var=None, training=True
+ )
+ for tl in y
+ ]
+
+ if self.cfg.instance_norm_target_layer:
+ y = [F.instance_norm(tl.float()) for tl in y]
+
+ if permuted:
+ y = [tl.transpose(1, 2) for tl in y] # BCT -> BTC
+
+ if self.cfg.layer_norm_target_layer:
+ y = [F.layer_norm(tl.float(), tl.shape[-1:]) for tl in y]
+
+ y = sum(y) / len(y)
+
+ if not permuted:
+ y = y.transpose(0, 1)
+
+ if self.cfg.layer_norm_targets:
+ y = F.layer_norm(y.float(), y.shape[-1:])
+
+ if self.cfg.instance_norm_targets:
+ y = F.instance_norm(y.transpose(1, 2)).transpose(1, 2)
+
+ masked_indices = src_tokens.eq(self.mask_idx)
+
+ x = x[masked_indices]
+ y = y[masked_indices]
+
+ x = self.regression_head(x)
+
+ sz = x.size(-1)
+ if self.cfg.loss_beta == 0:
+ loss = F.mse_loss(x.float(), y.float(), reduction="none").sum(dim=-1)
+ else:
+ loss = F.smooth_l1_loss(
+ x.float(), y.float(), reduction="none", beta=self.cfg.loss_beta
+ ).sum(dim=-1)
+
+ result = {
+ "losses": {
+ "main": loss.sum() / math.sqrt(sz)
+ if self.loss_scale <= 0
+ else loss.sum() * self.loss_scale,
+ },
+ "sample_size": loss.numel(),
+ }
+
+ # logging other values
+ other_logs = {
+ "ema_decay": self.ema.get_decay() * 1000
+ }
+ result["logs"] = other_logs
+ return result
+
+ def extract_features(self, src_tokens, return_all_hiddens=False, **kwargs):
+ encoder_out = self.sentence_encoder(
+ src_tokens,
+ return_all_hiddens=return_all_hiddens,
+ token_embeddings=kwargs.get("token_embeddings", None),
+ )
+ # T x B x C -> B x T x C
+ features = encoder_out["encoder_out"][0].transpose(0, 1)
+ inner_states = encoder_out["encoder_states"] if return_all_hiddens else None
+ return features, {
+ "inner_states": inner_states,
+ "encoder_embedding": encoder_out["encoder_embedding"][0],
+ }
+
+ def output_layer(self, features, masked_tokens=None, **unused):
+ return self.lm_head(features, masked_tokens)
+
+ def max_positions(self):
+ """Maximum output length supported by the encoder."""
+ return self.cfg.max_positions
diff --git a/fairseq/examples/data2vec/models/data2vec_text_classification.py b/fairseq/examples/data2vec/models/data2vec_text_classification.py
new file mode 100644
index 0000000000000000000000000000000000000000..e787b916dcff5b2dae669dc2cdc5cd8ff1ff8564
--- /dev/null
+++ b/fairseq/examples/data2vec/models/data2vec_text_classification.py
@@ -0,0 +1,141 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+# The code in this file is adapted from the BeiT implementation which can be found here:
+# https://github.com/microsoft/unilm/tree/master/beit
+
+import logging
+
+from dataclasses import dataclass
+from typing import Any
+
+from omegaconf import II, MISSING
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from fairseq import checkpoint_utils, tasks
+
+from fairseq.dataclass import FairseqDataclass
+from fairseq.models import BaseFairseqModel, register_model
+from fairseq.models.roberta.model import RobertaClassificationHead
+
+from examples.data2vec.data.modality import Modality
+
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class Data2VecTextClassificationConfig(FairseqDataclass):
+ pooler_dropout: float = 0.0
+ pooler_activation_fn: str = "tanh"
+ quant_noise_pq: int = 0
+ quant_noise_pq_block_size: int = 8
+ spectral_norm_classification_head: bool = False
+
+ model_path: str = MISSING
+ no_pretrained_weights: bool = False
+
+ pretrained_model_args: Any = None
+
+
+@register_model(
+ "data2vec_text_classification", dataclass=Data2VecTextClassificationConfig
+)
+class Data2VecTextClassificationModel(BaseFairseqModel):
+ def __init__(self, cfg: Data2VecTextClassificationConfig):
+ super().__init__()
+ self.cfg = cfg
+
+ if cfg.pretrained_model_args is None:
+ state = checkpoint_utils.load_checkpoint_to_cpu(cfg.model_path, {})
+ pretrained_args = state.get("cfg", None)
+ pretrained_args.criterion = None
+ pretrained_args.lr_scheduler = None
+ cfg.pretrained_model_args = pretrained_args
+
+ logger.info(pretrained_args)
+ else:
+ state = None
+ pretrained_args = cfg.pretrained_model_args
+
+ task = tasks.setup_task(pretrained_args.task)
+ model = task.build_model(pretrained_args.model, from_checkpoint=True)
+
+ model.remove_pretraining_modules()
+
+ self.model = model
+
+ if state is not None and not cfg.no_pretrained_weights:
+ self.load_model_weights(state, model, cfg)
+
+ self.classification_heads = nn.ModuleDict()
+
+
+ def load_model_weights(self, state, model, cfg):
+ for k in list(state["model"].keys()):
+ if (
+ k.startswith("shared_decoder") or
+ k.startswith("_ema") or
+ "decoder" in k
+ ):
+ logger.info(f"Deleting {k} from checkpoint")
+ del state["model"][k]
+ model.load_state_dict(state["model"], strict=True)
+
+ @classmethod
+ def build_model(cls, cfg: Data2VecTextClassificationConfig, task=None):
+ """Build a new model instance."""
+
+ return cls(cfg)
+
+ def register_classification_head(
+ self, name, num_classes=None, inner_dim=None, **kwargs
+ ):
+ """Register a classification head."""
+ if name in self.classification_heads:
+ prev_num_classes = self.classification_heads[name].out_proj.out_features
+ prev_inner_dim = self.classification_heads[name].dense.out_features
+ if num_classes != prev_num_classes or inner_dim != prev_inner_dim:
+ logger.warning(
+ 're-registering head "{}" with num_classes {} (prev: {}) '
+ "and inner_dim {} (prev: {})".format(
+ name, num_classes, prev_num_classes, inner_dim, prev_inner_dim
+ )
+ )
+ embed_dim = self.cfg.pretrained_model_args.model.embed_dim
+ self.classification_heads[name] = RobertaClassificationHead(
+ input_dim=embed_dim,
+ inner_dim=inner_dim or embed_dim,
+ num_classes=num_classes,
+ activation_fn=self.cfg.pooler_activation_fn,
+ pooler_dropout=self.cfg.pooler_dropout,
+ q_noise=self.cfg.quant_noise_pq,
+ qn_block_size=self.cfg.quant_noise_pq_block_size,
+ do_spectral_norm=self.cfg.spectral_norm_classification_head,
+ )
+
+ def forward(
+ self,
+ source,
+ id,
+ padding_mask,
+ features_only=True,
+ remove_extra_tokens=True,
+ classification_head_name=None,
+ ):
+ encoder_out = self.model(
+ source,
+ id=id,
+ mode=Modality.TEXT,
+ padding_mask=padding_mask,
+ mask=False,
+ features_only=features_only,
+ remove_extra_tokens=remove_extra_tokens
+ )
+ logits = self.classification_heads[classification_head_name](encoder_out["x"])
+ return logits, encoder_out
diff --git a/fairseq/examples/data2vec/models/data2vec_vision.py b/fairseq/examples/data2vec/models/data2vec_vision.py
new file mode 100644
index 0000000000000000000000000000000000000000..2f8989442941b4ac7b7f2ee050ac821948687857
--- /dev/null
+++ b/fairseq/examples/data2vec/models/data2vec_vision.py
@@ -0,0 +1,727 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+# The code in this file is adapted from the BeiT implementation which can be found here:
+# https://github.com/microsoft/unilm/tree/master/beit
+
+import logging
+import math
+import numpy as np
+import random
+
+from dataclasses import dataclass, field
+from typing import Optional
+
+from omegaconf import II
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.distributed as dist
+
+from fairseq.modules import EMAModule, EMAModuleConfig
+from fairseq.dataclass import FairseqDataclass
+from fairseq.models import BaseFairseqModel, register_model
+
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class Data2VecVisionConfig(FairseqDataclass):
+ layer_scale_init_value: float = field(
+ default=1e-4, metadata={"help": "rescale layer outputs, 0 to disable"}
+ )
+ num_mask_patches: int = field(
+ default=75,
+ metadata={"help": "number of the visual tokens/patches need be masked"},
+ )
+ min_mask_patches_per_block: int = 16
+ max_mask_patches_per_block: int = 196
+ image_size: int = 224
+ patch_size: int = 16
+ in_channels: int = 3
+
+ shared_rel_pos_bias: bool = True
+
+ drop_path: float = 0.1
+ attention_dropout: float = 0.0
+
+ depth: int = 12
+ embed_dim: int = 768
+ num_heads: int = 12
+ mlp_ratio: int = 4
+
+ loss_beta: float = field(
+ default=0, metadata={"help": "beta for smooth l1 loss. 0 means use l2 loss"}
+ )
+ loss_scale: Optional[float] = field(
+ default=None,
+ metadata={
+ "help": "scale the reconstruction loss by this constant. if None then scales by 1/sqrt(dim)"
+ },
+ )
+ average_top_k_layers: int = field(
+ default=8, metadata={"help": "how many layers to average"}
+ )
+
+ end_of_block_targets: bool = True
+ layer_norm_target_layer: bool = False
+ instance_norm_target_layer: bool = False
+ batch_norm_target_layer: bool = False
+ instance_norm_targets: bool = False
+ layer_norm_targets: bool = False
+
+ ema_decay: float = field(default=0.999, metadata={"help": "initial ema decay rate"})
+ ema_end_decay: float = field(
+ default=0.9999, metadata={"help": "final ema decay rate"}
+ )
+
+ # when to finish annealing ema decay rate
+ ema_anneal_end_step: int = II("optimization.max_update")
+
+ ema_transformer_only: bool = field(
+ default=True,
+ metadata={"help": "whether to momentum update only the transformer layers"},
+ )
+
+
+def get_annealed_rate(start, end, curr_step, total_steps):
+ r = end - start
+ pct_remaining = 1 - curr_step / total_steps
+ return end - r * pct_remaining
+
+
+@register_model("data2vec_vision", dataclass=Data2VecVisionConfig)
+class Data2VecVisionModel(BaseFairseqModel):
+ def __init__(self, cfg: Data2VecVisionConfig):
+ super().__init__()
+ self.cfg = cfg
+
+ self.ema = None
+
+ self.average_top_k_layers = cfg.average_top_k_layers
+ self.loss_beta = cfg.loss_beta
+ self.loss_scale = (
+ cfg.loss_scale
+ if cfg.loss_scale is not None
+ else 1 / math.sqrt(cfg.embed_dim)
+ )
+
+ self.patch_embed = PatchEmbed(
+ img_size=cfg.image_size,
+ patch_size=cfg.patch_size,
+ in_chans=cfg.in_channels,
+ embed_dim=cfg.embed_dim,
+ )
+
+ patch_size = self.patch_embed.patch_size
+ self.window_size = (
+ cfg.image_size // patch_size[0],
+ cfg.image_size // patch_size[1],
+ )
+
+ self.cls_emb = nn.Parameter(torch.FloatTensor(1, 1, cfg.embed_dim))
+ self.mask_emb = nn.Parameter(torch.FloatTensor(1, 1, cfg.embed_dim))
+
+ nn.init.trunc_normal_(self.cls_emb, 0.02)
+ nn.init.trunc_normal_(self.mask_emb, 0.02)
+
+ self.encoder = TransformerEncoder(cfg, self.patch_embed.patch_shape)
+
+ self.final_proj = nn.Linear(cfg.embed_dim, cfg.embed_dim)
+ self.num_updates = 0
+
+ def make_ema_teacher(self):
+ ema_config = EMAModuleConfig(
+ ema_decay=self.cfg.ema_decay,
+ ema_fp32=True,
+ )
+ self.ema = EMAModule(
+ self.encoder if self.cfg.ema_transformer_only else self,
+ ema_config,
+ )
+
+ def set_num_updates(self, num_updates):
+ super().set_num_updates(num_updates)
+
+ if self.ema is None and self.final_proj is not None:
+ logger.info(f"making ema teacher")
+ self.make_ema_teacher()
+ elif self.training and self.ema is not None:
+ if self.cfg.ema_decay != self.cfg.ema_end_decay:
+ if num_updates >= self.cfg.ema_anneal_end_step:
+ decay = self.cfg.ema_end_decay
+ else:
+ decay = get_annealed_rate(
+ self.cfg.ema_decay,
+ self.cfg.ema_end_decay,
+ num_updates,
+ self.cfg.ema_anneal_end_step,
+ )
+ self.ema.set_decay(decay)
+ if self.ema.get_decay() < 1:
+ self.ema.step(self.encoder if self.cfg.ema_transformer_only else self)
+
+ self.num_updates = num_updates
+
+ def state_dict(self, destination=None, prefix="", keep_vars=False):
+ state = super().state_dict(destination, prefix, keep_vars)
+
+ if self.ema is not None:
+ state[prefix + "_ema"] = self.ema.fp32_params
+
+ return state
+
+ def _load_from_state_dict(self, state_dict, prefix, *args, **kwargs):
+ if self.ema is not None:
+ k = prefix + "_ema"
+ assert k in state_dict
+ self.ema.restore(state_dict[k], True)
+ del state_dict[k]
+ return super()._load_from_state_dict(state_dict, prefix, *args, **kwargs)
+
+ @classmethod
+ def build_model(cls, cfg: Data2VecVisionConfig, task=None):
+ """Build a new model instance."""
+
+ return cls(cfg)
+
+ def make_mask(self, bsz, num_masks, min_masks, max_masks):
+ height, width = self.window_size
+
+ masks = np.zeros(shape=(bsz, height, width), dtype=np.int)
+
+ for i in range(bsz):
+ mask = masks[i]
+ mask_count = 0
+
+ min_aspect = 0.3
+ max_aspect = 1 / min_aspect
+ log_aspect_ratio = (math.log(min_aspect), math.log(max_aspect))
+
+ def _mask(mask, max_mask_patches):
+ delta = 0
+ for attempt in range(10):
+ target_area = random.uniform(min_masks, max_mask_patches)
+ aspect_ratio = math.exp(random.uniform(*log_aspect_ratio))
+ h = int(round(math.sqrt(target_area * aspect_ratio)))
+ w = int(round(math.sqrt(target_area / aspect_ratio)))
+ if w < width and h < height:
+ top = random.randint(0, height - h)
+ left = random.randint(0, width - w)
+
+ num_masked = mask[top : top + h, left : left + w].sum()
+ # Overlap
+ if 0 < h * w - num_masked <= max_mask_patches:
+ for i in range(top, top + h):
+ for j in range(left, left + w):
+ if mask[i, j] == 0:
+ mask[i, j] = 1
+ delta += 1
+
+ if delta > 0:
+ break
+ return delta
+
+ while mask_count < num_masks:
+ max_mask_patches = min(num_masks - mask_count, max_masks)
+
+ delta = _mask(mask, max_mask_patches)
+ if delta == 0:
+ break
+ else:
+ mask_count += delta
+
+ return torch.from_numpy(masks)
+
+ def forward(
+ self,
+ img,
+ mask: bool = True,
+ layer_results: bool = False,
+ ):
+ x = self.patch_embed(img)
+ batch_size, seq_len, _ = x.size()
+
+ if mask:
+ mask_indices = self.make_mask(
+ img.size(0),
+ self.cfg.num_mask_patches,
+ self.cfg.min_mask_patches_per_block,
+ self.cfg.max_mask_patches_per_block,
+ )
+ bool_mask = mask_indices.view(mask_indices.size(0), -1).bool()
+ else:
+ mask_indices = bool_mask = None
+
+ cls_tokens = self.cls_emb.expand(batch_size, -1, -1)
+ x = torch.cat((cls_tokens, x), dim=1)
+
+ if self.ema is not None:
+ with torch.no_grad():
+ self.ema.model.eval()
+
+ if self.cfg.ema_transformer_only:
+ y = self.ema.model(
+ x,
+ layer_results="end" if self.cfg.end_of_block_targets else "fc",
+ )
+ else:
+ y = self.ema.model(
+ img,
+ mask=False,
+ layer_results=True,
+ )
+
+ y = y[-self.cfg.average_top_k_layers :]
+
+ permuted = False
+ if self.cfg.instance_norm_target_layer or self.cfg.batch_norm_target_layer:
+ y = [tl.transpose(1, 2) for tl in y] # BTC -> BCT
+ permuted = True
+
+ if self.cfg.batch_norm_target_layer:
+ y = [
+ F.batch_norm(
+ tl.float(), running_mean=None, running_var=None, training=True
+ )
+ for tl in y
+ ]
+
+ if self.cfg.instance_norm_target_layer:
+ y = [F.instance_norm(tl.float()) for tl in y]
+
+ if permuted:
+ y = [tl.transpose(1, 2) for tl in y] # BCT -> BTC
+
+ if self.cfg.layer_norm_target_layer:
+ y = [F.layer_norm(tl.float(), tl.shape[-1:]) for tl in y]
+
+ y = sum(y) / len(y)
+
+ if self.cfg.layer_norm_targets:
+ y = F.layer_norm(y.float(), y.shape[-1:])
+
+ if self.cfg.instance_norm_targets:
+ y = F.instance_norm(y.float().transpose(1, 2)).transpose(1, 2)
+
+ y = y[bool_mask].float()
+
+ if mask_indices is not None:
+ mask_token = self.mask_emb.expand(batch_size, seq_len, -1)
+ w = mask_indices.view(mask_indices.size(0), -1, 1).type_as(mask_token)
+ x[:, 1:] = x[:, 1:] * (1 - w) + mask_token * w
+
+ if layer_results:
+ enc_layer_results = "end" if self.cfg.end_of_block_targets else "fc"
+ else:
+ enc_layer_results = None
+
+ x = self.encoder(x, layer_results=enc_layer_results)
+ if layer_results or mask_indices is None:
+ return x
+
+ x = x[bool_mask].float()
+
+ if self.loss_beta == 0:
+ loss = F.mse_loss(x, y, reduction="none").sum(dim=-1)
+ else:
+ loss = F.smooth_l1_loss(x, y, reduction="none", beta=self.loss_beta).sum(
+ dim=-1
+ )
+
+ if self.loss_scale > 0:
+ loss = loss * self.loss_scale
+
+ result = {
+ "losses": {"regression": loss.sum()},
+ "sample_size": loss.numel(),
+ "target_var": self.compute_var(y),
+ "pred_var": self.compute_var(x),
+ "ema_decay": self.ema.get_decay() * 1000,
+ }
+ return result
+
+ @staticmethod
+ def compute_var(y):
+ y = y.view(-1, y.size(-1))
+ if dist.is_initialized():
+ zc = torch.tensor(y.size(0)).cuda()
+ zs = y.sum(dim=0)
+ zss = (y ** 2).sum(dim=0)
+
+ dist.all_reduce(zc)
+ dist.all_reduce(zs)
+ dist.all_reduce(zss)
+
+ var = zss / (zc - 1) - (zs ** 2) / (zc * (zc - 1))
+ return torch.sqrt(var + 1e-6).mean()
+ else:
+ return torch.sqrt(y.var(dim=0) + 1e-6).mean()
+
+ def remove_pretraining_modules(self, last_layer=None):
+ self.final_proj = None
+ self.ema = None
+ self.encoder.norm = nn.Identity()
+ self.mask_emb = None
+ if last_layer is not None:
+ self.encoder.layers = nn.ModuleList(
+ l for i, l in enumerate(self.encoder.layers) if i <= last_layer
+ )
+
+
+class PatchEmbed(nn.Module):
+ """Image to Patch Embedding"""
+
+ def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768):
+ super().__init__()
+ if isinstance(img_size, int):
+ img_size = img_size, img_size
+ if isinstance(patch_size, int):
+ patch_size = patch_size, patch_size
+ num_patches = (img_size[1] // patch_size[1]) * (img_size[0] // patch_size[0])
+ self.patch_shape = (img_size[0] // patch_size[0], img_size[1] // patch_size[1])
+ self.img_size = img_size
+ self.patch_size = patch_size
+ self.num_patches = num_patches
+
+ self.conv = nn.Conv2d(
+ in_chans, embed_dim, kernel_size=patch_size, stride=patch_size
+ )
+
+ def forward(self, x):
+ # BCHW -> BTC
+ x = self.conv(x).flatten(2).transpose(1, 2)
+ return x
+
+
+class Attention(nn.Module):
+ def __init__(
+ self,
+ dim,
+ num_heads=8,
+ qkv_bias=True,
+ attn_drop=0.0,
+ proj_drop=0.0,
+ window_size=None,
+ attn_head_dim=None,
+ ):
+ super().__init__()
+ self.num_heads = num_heads
+ head_dim = dim // num_heads
+ if attn_head_dim is not None:
+ head_dim = attn_head_dim
+ all_head_dim = head_dim * self.num_heads
+ self.scale = head_dim ** -0.5
+
+ self.qkv = nn.Linear(dim, all_head_dim * 3, bias=False)
+ if qkv_bias:
+ self.q_bias = nn.Parameter(torch.zeros(all_head_dim))
+ self.v_bias = nn.Parameter(torch.zeros(all_head_dim))
+ else:
+ self.q_bias = None
+ self.v_bias = None
+
+ if window_size:
+ self.window_size = window_size
+ self.num_relative_distance = (2 * window_size[0] - 1) * (
+ 2 * window_size[1] - 1
+ ) + 3
+ self.relative_position_bias_table = nn.Parameter(
+ torch.zeros(self.num_relative_distance, num_heads)
+ ) # 2*Wh-1 * 2*Ww-1, nH
+ # cls to token & token 2 cls & cls to cls
+
+ # get pair-wise relative position index for each token inside the window
+ coords_h = torch.arange(window_size[0])
+ coords_w = torch.arange(window_size[1])
+ coords = torch.stack(torch.meshgrid([coords_h, coords_w])) # 2, Wh, Ww
+ coords_flatten = torch.flatten(coords, 1) # 2, Wh*Ww
+ relative_coords = (
+ coords_flatten[:, :, None] - coords_flatten[:, None, :]
+ ) # 2, Wh*Ww, Wh*Ww
+ relative_coords = relative_coords.permute(
+ 1, 2, 0
+ ).contiguous() # Wh*Ww, Wh*Ww, 2
+ relative_coords[:, :, 0] += window_size[0] - 1 # shift to start from 0
+ relative_coords[:, :, 1] += window_size[1] - 1
+ relative_coords[:, :, 0] *= 2 * window_size[1] - 1
+ relative_position_index = torch.zeros(
+ size=(window_size[0] * window_size[1] + 1,) * 2,
+ dtype=relative_coords.dtype,
+ )
+ relative_position_index[1:, 1:] = relative_coords.sum(-1) # Wh*Ww, Wh*Ww
+ relative_position_index[0, 0:] = self.num_relative_distance - 3
+ relative_position_index[0:, 0] = self.num_relative_distance - 2
+ relative_position_index[0, 0] = self.num_relative_distance - 1
+
+ self.register_buffer("relative_position_index", relative_position_index)
+ else:
+ self.window_size = None
+ self.relative_position_bias_table = None
+ self.relative_position_index = None
+
+ self.attn_drop = nn.Dropout(attn_drop)
+ self.proj = nn.Linear(all_head_dim, dim)
+ self.proj_drop = nn.Dropout(proj_drop)
+
+ def forward(self, x, rel_pos_bias=None):
+ B, N, C = x.shape
+ qkv_bias = None
+ if self.q_bias is not None:
+ qkv_bias = torch.cat(
+ (
+ self.q_bias,
+ torch.zeros_like(self.v_bias, requires_grad=False),
+ self.v_bias,
+ )
+ )
+ # qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
+ qkv = F.linear(input=x, weight=self.qkv.weight, bias=qkv_bias)
+ qkv = qkv.reshape(B, N, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
+ q, k, v = (
+ qkv[0],
+ qkv[1],
+ qkv[2],
+ ) # make torchscript happy (cannot use tensor as tuple)
+
+ q = q * self.scale
+ attn = q @ k.transpose(-2, -1)
+
+ if self.relative_position_bias_table is not None:
+ assert 1==2
+ relative_position_bias = self.relative_position_bias_table[
+ self.relative_position_index.view(-1)
+ ].view(
+ self.window_size[0] * self.window_size[1] + 1,
+ self.window_size[0] * self.window_size[1] + 1,
+ -1,
+ ) # Wh*Ww,Wh*Ww,nH
+ relative_position_bias = relative_position_bias.permute(
+ 2, 0, 1
+ ).contiguous() # nH, Wh*Ww, Wh*Ww
+ attn = attn + relative_position_bias.unsqueeze(0)
+ print("attn.size() :", attn.size())
+ print("rel_pos_bias.size() :", rel_pos_bias.size())
+ if rel_pos_bias is not None:
+ attn = attn + rel_pos_bias
+ attn = attn.softmax(dim=-1)
+ attn = self.attn_drop(attn)
+
+ x = (attn @ v).transpose(1, 2).reshape(B, N, -1)
+ x = self.proj(x)
+ x = self.proj_drop(x)
+ return x
+
+
+class RelativePositionBias(nn.Module):
+ def __init__(self, window_size, num_heads):
+ super().__init__()
+ self.window_size = window_size
+ self.num_relative_distance = (2 * window_size[0] - 1) * (
+ 2 * window_size[1] - 1
+ ) + 3
+ self.relative_position_bias_table = nn.Parameter(
+ torch.zeros(self.num_relative_distance, num_heads)
+ )
+
+ # get pair-wise relative position index for each token inside the window
+ coords_h = torch.arange(window_size[0])
+ coords_w = torch.arange(window_size[1])
+ coords = torch.stack(torch.meshgrid([coords_h, coords_w])) # 2, Wh, Ww
+ coords_flatten = torch.flatten(coords, 1) # 2, Wh*Ww
+ relative_coords = (
+ coords_flatten[:, :, None] - coords_flatten[:, None, :]
+ ) # 2, Wh*Ww, Wh*Ww
+ relative_coords = relative_coords.permute(
+ 1, 2, 0
+ ).contiguous() # Wh*Ww, Wh*Ww, 2
+ relative_coords[:, :, 0] += window_size[0] - 1 # shift to start from 0
+ relative_coords[:, :, 1] += window_size[1] - 1
+ relative_coords[:, :, 0] *= 2 * window_size[1] - 1
+ relative_position_index = torch.zeros(
+ size=(window_size[0] * window_size[1] + 1,) * 2, dtype=relative_coords.dtype
+ )
+ relative_position_index[1:, 1:] = relative_coords.sum(-1) # Wh*Ww, Wh*Ww
+ relative_position_index[0, 0:] = self.num_relative_distance - 3
+ relative_position_index[0:, 0] = self.num_relative_distance - 2
+ relative_position_index[0, 0] = self.num_relative_distance - 1
+
+ self.register_buffer("relative_position_index", relative_position_index)
+
+ def forward(self):
+ relative_position_bias = self.relative_position_bias_table[
+ self.relative_position_index.view(-1)
+ ].view(
+ self.window_size[0] * self.window_size[1] + 1,
+ self.window_size[0] * self.window_size[1] + 1,
+ -1,
+ ) # Wh*Ww,Wh*Ww,nH
+ print("self.window_size :", self.window_size)
+ print("self.num_relative_distance :", self.num_relative_distance)
+ print("self.relative_position_index :", self.relative_position_index.size(), self.relative_position_index)
+ print("relative_position_bias.size(), relative_position_bias :",relative_position_bias.size(), relative_position_bias)
+ print("self.relative_position_bias_table.size(), self.relative_position_bias_table :",self.relative_position_bias_table.size(), self.relative_position_bias_table)
+ return relative_position_bias.permute(2, 0, 1).contiguous() # nH, Wh*Ww, Wh*Ww
+
+
+class DropPath(nn.Module):
+ """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks)."""
+
+ def __init__(self, drop_prob=None):
+ super(DropPath, self).__init__()
+ self.drop_prob = drop_prob
+
+ def forward(self, x):
+ if self.drop_prob == 0.0 or not self.training:
+ return x
+ keep_prob = 1 - self.drop_prob
+ shape = (x.shape[0],) + (1,) * (
+ x.ndim - 1
+ ) # work with diff dim tensors, not just 2D ConvNets
+ random_tensor = keep_prob + torch.rand(shape, dtype=x.dtype, device=x.device)
+ random_tensor.floor_()
+ output = x.div(keep_prob) * random_tensor
+ return output
+
+ def extra_repr(self) -> str:
+ return "p={}".format(self.drop_prob)
+
+
+class Block(nn.Module):
+ def __init__(
+ self,
+ dim,
+ num_heads,
+ mlp_ratio=4.0,
+ drop=0.0,
+ attn_drop=0.0,
+ drop_path=0.0,
+ init_values=None,
+ window_size=None,
+ ):
+ super().__init__()
+
+ self.norm1 = nn.LayerNorm(dim)
+ self.attn = Attention(
+ dim,
+ num_heads=num_heads,
+ attn_drop=attn_drop,
+ proj_drop=drop,
+ window_size=window_size,
+ )
+
+ self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
+ self.norm2 = nn.LayerNorm(dim)
+ mlp_hidden_dim = int(dim * mlp_ratio)
+
+ self.mlp = nn.Sequential(
+ nn.Linear(dim, mlp_hidden_dim),
+ nn.GELU(),
+ nn.Linear(mlp_hidden_dim, dim),
+ nn.Dropout(drop),
+ )
+
+ if init_values > 0:
+ self.gamma_1 = nn.Parameter(
+ init_values * torch.ones((dim)), requires_grad=True
+ )
+ self.gamma_2 = nn.Parameter(
+ init_values * torch.ones((dim)), requires_grad=True
+ )
+ else:
+ self.gamma_1, self.gamma_2 = None, None
+
+ def forward(self, x, rel_pos_bias=None):
+ print("inside block :", x.size())
+ if self.gamma_1 is None:
+ x = x + self.drop_path(self.attn(self.norm1(x), rel_pos_bias=rel_pos_bias))
+ fc_feature = self.drop_path(self.mlp(self.norm2(x)))
+ x = x + fc_feature
+ else:
+ x = x + self.drop_path(
+ self.gamma_1 * self.attn(self.norm1(x), rel_pos_bias=rel_pos_bias)
+ )
+ fc_feature = self.drop_path(self.gamma_2 * self.mlp(self.norm2(x)))
+ x = x + fc_feature
+ return x, fc_feature
+
+
+class TransformerEncoder(nn.Module):
+ def __init__(self, cfg: Data2VecVisionConfig, patch_shape):
+ super().__init__()
+
+ self.rel_pos_bias = None
+ if cfg.shared_rel_pos_bias:
+ self.rel_pos_bias = RelativePositionBias(
+ window_size=patch_shape, num_heads=cfg.num_heads
+ )
+
+ dpr = [
+ x.item() for x in torch.linspace(0, cfg.drop_path, cfg.depth)
+ ] # stochastic depth decay rule
+
+ print("TransformerEncoder > patch_shape :", patch_shape)
+ self.blocks = nn.ModuleList(
+ Block(
+ dim=cfg.embed_dim,
+ num_heads=cfg.num_heads,
+ attn_drop=cfg.attention_dropout,
+ drop_path=dpr[i],
+ init_values=cfg.layer_scale_init_value,
+ window_size=patch_shape if not cfg.shared_rel_pos_bias else None,
+ )
+ for i in range(cfg.depth)
+ )
+
+ self.norm = nn.LayerNorm(cfg.embed_dim)
+
+ self.apply(self.init_weights)
+ self.fix_init_weight()
+
+ def init_weights(self, m):
+ std = 0.02
+ if isinstance(m, nn.Linear):
+ nn.init.trunc_normal_(m.weight, std=std)
+ if isinstance(m, nn.Linear) and m.bias is not None:
+ nn.init.constant_(m.bias, 0)
+ elif isinstance(m, nn.LayerNorm):
+ nn.init.constant_(m.bias, 0)
+ nn.init.constant_(m.weight, 1.0)
+ elif isinstance(m, nn.Conv2d):
+ nn.init.trunc_normal_(m.weight, std=std)
+ if m.bias is not None:
+ nn.init.constant_(m.bias, 0)
+
+ def fix_init_weight(self):
+ def rescale(param, layer_id):
+ param.div_(math.sqrt(2.0 * layer_id))
+
+ for layer_id, layer in enumerate(self.blocks):
+ rescale(layer.attn.proj.weight.data, layer_id + 1)
+ rescale(layer.mlp[2].weight.data, layer_id + 1)
+
+ def extract_features(self, x, layer_results):
+
+ rel_pos_bias = self.rel_pos_bias() if self.rel_pos_bias is not None else None
+
+ z = []
+ for i, blk in enumerate(self.blocks):
+ x, fc_feature = blk(x, rel_pos_bias=rel_pos_bias)
+ if layer_results == "end":
+ z.append(x)
+ elif layer_results == "fc":
+ z.append(fc_feature)
+
+ return z if layer_results else self.norm(x)
+
+ def forward(self, x, layer_results=None):
+ x = self.extract_features(x, layer_results=layer_results)
+ if layer_results:
+ return [z[:, 1:] for z in x]
+
+ x = x[:, 1:]
+ return x
diff --git a/fairseq/examples/data2vec/models/mae.py b/fairseq/examples/data2vec/models/mae.py
new file mode 100644
index 0000000000000000000000000000000000000000..a3b5f72a4a4886bcbdfc62f3670f32e96dcc696b
--- /dev/null
+++ b/fairseq/examples/data2vec/models/mae.py
@@ -0,0 +1,829 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+# The code in this file is adapted from the BeiT implementation which can be found here:
+# https://github.com/microsoft/unilm/tree/master/beit
+
+import logging
+from dataclasses import dataclass
+from functools import partial
+
+from timm.models.vision_transformer import PatchEmbed, Block
+
+import torch
+import torch.nn as nn
+
+import numpy as np
+
+from fairseq.dataclass import FairseqDataclass
+from fairseq.models import BaseFairseqModel, register_model
+from fairseq.models.wav2vec.wav2vec2 import TransformerSentenceEncoderLayer
+
+try:
+ from apex.normalization import FusedLayerNorm
+except:
+ FusedLayerNorm = nn.LayerNorm
+
+import torch.nn.functional as F
+
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class MaeConfig(FairseqDataclass):
+ input_size: int = 224
+ in_chans: int = 3
+ patch_size: int = 16
+ embed_dim: int = 768
+ depth: int = 12
+ num_heads: int = 12
+ decoder_embed_dim: int = 512
+ decoder_depth: int = 8
+ decoder_num_heads: int = 16
+ mlp_ratio: int = 4
+ norm_eps: float = 1e-6
+
+ drop_path_rate: float = 0.0
+
+ mask_ratio: float = 0.75
+ norm_pix_loss: bool = True
+
+ w2v_block: bool = False
+ alt_block: bool = False
+ alt_block2: bool = False
+ alt_attention: bool = False
+ block_dropout: float = 0
+ attention_dropout: float = 0
+ activation_dropout: float = 0
+ layer_norm_first: bool = False
+
+ fused_ln: bool = True
+ end_of_block_targets: bool = True
+
+ no_decoder_embed: bool = False
+ no_decoder_pos_embed: bool = False
+ mask_noise_std: float = 0
+
+ single_qkv: bool = False
+ use_rel_pos_bias: bool = False
+ no_cls: bool = False
+
+
+def modify_relative_position_bias(orig_bias, bsz, mask):
+ if mask is None:
+ return orig_bias.unsqueeze(0).repeat(
+ bsz, 1, 1, 1
+ ) # heads x seq_len x seq_len => bsz x heads x seq_len x seq_len
+ heads, max_seq_len, max_seq_len = orig_bias.shape # includes CLS token
+ mask_for_rel_pos_bias = torch.cat(
+ (torch.zeros(bsz, 1, dtype=mask.dtype, device=mask.device), mask), dim=1
+ ).bool() # bsz x seqlen (add CLS token)
+ unmasked_for_rel_pos_bias = ~mask_for_rel_pos_bias
+ unmasked_for_rel_pos_bias = unmasked_for_rel_pos_bias.unsqueeze(1).repeat(
+ 1, heads, 1
+ ) # bsz x seq_len => bsz x heads x seq_len
+ b_t_t_rel_pos_bias = orig_bias.unsqueeze(0).repeat(
+ bsz, 1, 1, 1
+ ) # heads x seq_len x seq_len => bsz x heads x seq_len x seq_len
+ b_t_t_rel_pos_bias = b_t_t_rel_pos_bias.masked_select(
+ unmasked_for_rel_pos_bias.unsqueeze(-1)
+ )
+ b_t_t_rel_pos_bias = b_t_t_rel_pos_bias.view(bsz, heads, -1, max_seq_len)
+ new_len = b_t_t_rel_pos_bias.size(-2)
+ b_t_t_rel_pos_bias = b_t_t_rel_pos_bias.masked_select(
+ unmasked_for_rel_pos_bias.unsqueeze(-2)
+ )
+ b_t_t_rel_pos_bias = b_t_t_rel_pos_bias.view(bsz, heads, new_len, new_len)
+ return b_t_t_rel_pos_bias
+
+
+class AltBlock(nn.Module):
+ def __init__(
+ self,
+ dim,
+ num_heads,
+ mlp_ratio=4.0,
+ qkv_bias=False,
+ qk_scale=None,
+ drop=0.0,
+ attn_drop=0.0,
+ drop_path=0.0,
+ act_layer=nn.GELU,
+ norm_layer=nn.LayerNorm,
+ layer_norm_first=True,
+ ffn_targets=False,
+ use_rel_pos_bias=False,
+ window_size=None,
+ alt_attention=False,
+ ):
+ super().__init__()
+
+ self.layer_norm_first = layer_norm_first
+ self.ffn_targets = ffn_targets
+
+ from timm.models.vision_transformer import Attention, DropPath, Mlp
+
+ self.norm1 = norm_layer(dim)
+ self.use_rel_pos_bias = use_rel_pos_bias
+ if use_rel_pos_bias:
+ self.attn = AltAttention(
+ dim,
+ num_heads=num_heads,
+ qkv_bias=qkv_bias,
+ qk_scale=qk_scale,
+ attn_drop=attn_drop,
+ proj_drop=drop,
+ window_size=window_size,
+ )
+ else:
+ if alt_attention:
+ from .multi.modules import AltAttention as AltAttention2
+ self.attn = AltAttention2(
+ dim,
+ num_heads=num_heads,
+ qkv_bias=qkv_bias,
+ qk_scale=qk_scale,
+ attn_drop=attn_drop,
+ proj_drop=drop,
+ )
+ else:
+ self.attn = Attention(
+ dim,
+ num_heads=num_heads,
+ qkv_bias=qkv_bias,
+ qk_scale=qk_scale,
+ attn_drop=attn_drop,
+ proj_drop=drop,
+ )
+ # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
+ self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
+ self.norm2 = norm_layer(dim)
+ mlp_hidden_dim = int(dim * mlp_ratio)
+ self.mlp = Mlp(
+ in_features=dim,
+ hidden_features=mlp_hidden_dim,
+ act_layer=act_layer,
+ drop=drop,
+ )
+
+ def forward(self, x, rel_pos_bias=None, pos_mask=None):
+ if self.layer_norm_first:
+ if self.use_rel_pos_bias:
+ x = x + self.drop_path(
+ self.attn(
+ self.norm1(x), rel_pos_bias=rel_pos_bias, pos_mask=pos_mask
+ )
+ )
+ else:
+ x = x + self.drop_path(self.attn(self.norm1(x)))
+ t = self.mlp(self.norm2(x))
+ x = x + self.drop_path(t)
+ if not self.ffn_targets:
+ t = x
+ return x, t
+ else:
+ if self.use_rel_pos_bias:
+ x = x + self.drop_path(
+ self.attn(x, rel_pos_bias=rel_pos_bias, pos_mask=pos_mask)
+ )
+ else:
+ x = x + self.drop_path(self.attn(x))
+ r = x = self.norm1(x)
+ x = self.mlp(x)
+ t = x
+ x = self.norm2(r + self.drop_path(x))
+ if not self.ffn_targets:
+ t = x
+ return x, t
+
+
+class AltAttention(nn.Module):
+ def __init__(
+ self,
+ dim,
+ num_heads=8,
+ qkv_bias=True,
+ qk_scale=None,
+ attn_drop=0.0,
+ proj_drop=0.0,
+ window_size=None,
+ attn_head_dim=None,
+ ):
+ super().__init__()
+ self.num_heads = num_heads
+ head_dim = dim // num_heads
+ if attn_head_dim is not None:
+ head_dim = attn_head_dim
+ all_head_dim = head_dim * self.num_heads
+ self.scale = qk_scale or head_dim ** -0.5
+
+ self.qkv = nn.Linear(dim, all_head_dim * 3, bias=False)
+ if qkv_bias:
+ self.q_bias = nn.Parameter(torch.zeros(all_head_dim))
+ self.v_bias = nn.Parameter(torch.zeros(all_head_dim))
+ else:
+ self.q_bias = None
+ self.v_bias = None
+
+ if window_size:
+ self.window_size = window_size
+ self.num_relative_distance = (2 * window_size[0] - 1) * (
+ 2 * window_size[1] - 1
+ ) + 3
+ self.relative_position_bias_table = nn.Parameter(
+ torch.zeros(self.num_relative_distance, num_heads)
+ ) # 2*Wh-1 * 2*Ww-1, nH
+ # cls to token & token 2 cls & cls to cls
+
+ # get pair-wise relative position index for each token inside the window
+ coords_h = torch.arange(window_size[0])
+ coords_w = torch.arange(window_size[1])
+ coords = torch.stack(torch.meshgrid([coords_h, coords_w])) # 2, Wh, Ww
+ coords_flatten = torch.flatten(coords, 1) # 2, Wh*Ww
+ relative_coords = (
+ coords_flatten[:, :, None] - coords_flatten[:, None, :]
+ ) # 2, Wh*Ww, Wh*Ww
+ relative_coords = relative_coords.permute(
+ 1, 2, 0
+ ).contiguous() # Wh*Ww, Wh*Ww, 2
+ relative_coords[:, :, 0] += window_size[0] - 1 # shift to start from 0
+ relative_coords[:, :, 1] += window_size[1] - 1
+ relative_coords[:, :, 0] *= 2 * window_size[1] - 1
+ relative_position_index = torch.zeros(
+ size=(window_size[0] * window_size[1] + 1,) * 2,
+ dtype=relative_coords.dtype,
+ )
+ relative_position_index[1:, 1:] = relative_coords.sum(-1) # Wh*Ww, Wh*Ww
+ relative_position_index[0, 0:] = self.num_relative_distance - 3
+ relative_position_index[0:, 0] = self.num_relative_distance - 2
+ relative_position_index[0, 0] = self.num_relative_distance - 1
+
+ self.register_buffer("relative_position_index", relative_position_index)
+ else:
+ self.window_size = None
+ self.relative_position_bias_table = None
+ self.relative_position_index = None
+
+ self.attn_drop = nn.Dropout(attn_drop)
+ self.proj = nn.Linear(all_head_dim, dim)
+ self.proj_drop = nn.Dropout(proj_drop)
+
+ def forward(self, x, rel_pos_bias=None, pos_mask=None):
+ B, N, C = x.shape
+ qkv_bias = None
+ if self.q_bias is not None:
+ qkv_bias = torch.cat(
+ (
+ self.q_bias,
+ torch.zeros_like(self.v_bias, requires_grad=False),
+ self.v_bias,
+ )
+ )
+ # qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
+ qkv = F.linear(input=x, weight=self.qkv.weight, bias=qkv_bias)
+ qkv = qkv.reshape(B, N, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
+ q, k, v = (
+ qkv[0],
+ qkv[1],
+ qkv[2],
+ ) # make torchscript happy (cannot use tensor as tuple)
+
+ q = q * self.scale
+ attn = q @ k.transpose(-2, -1)
+
+ if self.relative_position_bias_table is not None:
+ relative_position_bias = self.relative_position_bias_table[
+ self.relative_position_index.view(-1)
+ ].view(
+ self.window_size[0] * self.window_size[1] + 1,
+ self.window_size[0] * self.window_size[1] + 1,
+ -1,
+ ) # Wh*Ww,Wh*Ww,nH
+ relative_position_bias = relative_position_bias.permute(
+ 2, 0, 1
+ ).contiguous() # nH, Wh*Ww, Wh*Ww
+ attn = attn + modify_relative_position_bias(
+ relative_position_bias, x.size(0), pos_mask
+ )
+
+ if rel_pos_bias is not None:
+ attn = attn + rel_pos_bias
+
+ attn = attn.softmax(dim=-1)
+ attn = self.attn_drop(attn)
+
+ x = (attn @ v).transpose(1, 2).reshape(B, N, -1)
+ x = self.proj(x)
+ x = self.proj_drop(x)
+ return x
+
+
+class RelativePositionBias(nn.Module):
+ def __init__(self, window_size, num_heads):
+ super().__init__()
+ self.window_size = window_size
+ self.num_relative_distance = (2 * window_size[0] - 1) * (
+ 2 * window_size[1] - 1
+ ) + 3
+ self.relative_position_bias_table = nn.Parameter(
+ torch.zeros(self.num_relative_distance, num_heads)
+ )
+
+ # get pair-wise relative position index for each token inside the window
+ coords_h = torch.arange(window_size[0])
+ coords_w = torch.arange(window_size[1])
+ coords = torch.stack(torch.meshgrid([coords_h, coords_w])) # 2, Wh, Ww
+ coords_flatten = torch.flatten(coords, 1) # 2, Wh*Ww
+ relative_coords = (
+ coords_flatten[:, :, None] - coords_flatten[:, None, :]
+ ) # 2, Wh*Ww, Wh*Ww
+ relative_coords = relative_coords.permute(
+ 1, 2, 0
+ ).contiguous() # Wh*Ww, Wh*Ww, 2
+ relative_coords[:, :, 0] += window_size[0] - 1 # shift to start from 0
+ relative_coords[:, :, 1] += window_size[1] - 1
+ relative_coords[:, :, 0] *= 2 * window_size[1] - 1
+ relative_position_index = torch.zeros(
+ size=(window_size[0] * window_size[1] + 1,) * 2, dtype=relative_coords.dtype
+ )
+ relative_position_index[1:, 1:] = relative_coords.sum(-1) # Wh*Ww, Wh*Ww
+ relative_position_index[0, 0:] = self.num_relative_distance - 3
+ relative_position_index[0:, 0] = self.num_relative_distance - 2
+ relative_position_index[0, 0] = self.num_relative_distance - 1
+
+ self.register_buffer("relative_position_index", relative_position_index)
+
+ def forward(self):
+ relative_position_bias = self.relative_position_bias_table[
+ self.relative_position_index.view(-1)
+ ].view(
+ self.window_size[0] * self.window_size[1] + 1,
+ self.window_size[0] * self.window_size[1] + 1,
+ -1,
+ ) # Wh*Ww,Wh*Ww,nH
+ return relative_position_bias.permute(2, 0, 1).contiguous() # nH, Wh*Ww, Wh*Ww
+
+
+def get_2d_sincos_pos_embed(embed_dim, grid_size, cls_token=False):
+ """
+ grid_size: int of the grid height and width
+ return:
+ pos_embed: [grid_size*grid_size, embed_dim] or [1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token)
+ """
+ grid_h = np.arange(grid_size, dtype=np.float32)
+ grid_w = np.arange(grid_size, dtype=np.float32)
+ grid = np.meshgrid(grid_w, grid_h) # here w goes first
+ grid = np.stack(grid, axis=0)
+
+ grid = grid.reshape([2, 1, grid_size, grid_size])
+ pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
+ if cls_token:
+ pos_embed = np.concatenate([np.zeros([1, embed_dim]), pos_embed], axis=0)
+ return pos_embed
+
+
+def get_2d_sincos_pos_embed_from_grid(embed_dim, grid):
+ assert embed_dim % 2 == 0
+
+ # use half of dimensions to encode grid_h
+ emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0]) # (H*W, D/2)
+ emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1]) # (H*W, D/2)
+
+ emb = np.concatenate([emb_h, emb_w], axis=1) # (H*W, D)
+ return emb
+
+
+def get_1d_sincos_pos_embed_from_grid(embed_dim, pos):
+ """
+ embed_dim: output dimension for each position
+ pos: a list of positions to be encoded: size (M,)
+ out: (M, D)
+ """
+ assert embed_dim % 2 == 0
+ omega = np.arange(embed_dim // 2, dtype=np.float)
+ omega /= embed_dim / 2.0
+ omega = 1.0 / 10000 ** omega # (D/2,)
+
+ pos = pos.reshape(-1) # (M,)
+ out = np.einsum("m,d->md", pos, omega) # (M, D/2), outer product
+
+ emb_sin = np.sin(out) # (M, D/2)
+ emb_cos = np.cos(out) # (M, D/2)
+
+ emb = np.concatenate([emb_sin, emb_cos], axis=1) # (M, D)
+ return emb
+
+
+def interpolate_pos_embed(model, checkpoint_model):
+ if "pos_embed" in checkpoint_model:
+ pos_embed_checkpoint = checkpoint_model["pos_embed"]
+ embedding_size = pos_embed_checkpoint.shape[-1]
+ num_patches = model.patch_embed.num_patches
+ num_extra_tokens = model.pos_embed.shape[-2] - num_patches
+ # height (== width) for the checkpoint position embedding
+ orig_size = int((pos_embed_checkpoint.shape[-2] - num_extra_tokens) ** 0.5)
+ # height (== width) for the new position embedding
+ new_size = int(num_patches ** 0.5)
+ # class_token and dist_token are kept unchanged
+ if orig_size != new_size:
+ print(
+ "Position interpolate from %dx%d to %dx%d"
+ % (orig_size, orig_size, new_size, new_size)
+ )
+ extra_tokens = pos_embed_checkpoint[:, :num_extra_tokens]
+ # only the position tokens are interpolated
+ pos_tokens = pos_embed_checkpoint[:, num_extra_tokens:]
+ pos_tokens = pos_tokens.reshape(
+ -1, orig_size, orig_size, embedding_size
+ ).permute(0, 3, 1, 2)
+ pos_tokens = torch.nn.functional.interpolate(
+ pos_tokens,
+ size=(new_size, new_size),
+ mode="bicubic",
+ align_corners=False,
+ )
+ pos_tokens = pos_tokens.permute(0, 2, 3, 1).flatten(1, 2)
+ new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=1)
+ checkpoint_model["pos_embed"] = new_pos_embed
+
+
+@register_model("mae", dataclass=MaeConfig)
+class MaeModel(BaseFairseqModel):
+ def __init__(self, cfg: MaeConfig):
+ super().__init__()
+ self.cfg = cfg
+
+ self.mask_ratio = cfg.mask_ratio
+
+ # --------------------------------------------------------------------------
+ # MAE encoder specifics
+ self.patch_embed = PatchEmbed(
+ cfg.input_size, cfg.patch_size, cfg.in_chans, cfg.embed_dim
+ )
+ num_patches = self.patch_embed.num_patches
+
+ self.cls_token = nn.Parameter(torch.zeros(1, 1, cfg.embed_dim)) if not cfg.no_cls else None
+ self.pos_embed = nn.Parameter(
+ torch.zeros(1, num_patches + int(not cfg.no_cls), cfg.embed_dim), requires_grad=False
+ ) # fixed sin-cos embedding
+
+ norm_layer = partial(nn.LayerNorm, eps=cfg.norm_eps)
+
+ dpr = [
+ x.item() for x in torch.linspace(0, cfg.drop_path_rate, cfg.depth)
+ ] # stochastic depth decay rule
+
+ def make_block(drop_path):
+ if cfg.w2v_block:
+ return TransformerSentenceEncoderLayer(
+ embedding_dim=cfg.embed_dim,
+ ffn_embedding_dim=cfg.embed_dim * cfg.mlp_ratio,
+ num_attention_heads=cfg.num_heads,
+ dropout=cfg.block_dropout,
+ attention_dropout=cfg.attention_dropout,
+ activation_dropout=cfg.activation_dropout,
+ activation_fn="gelu",
+ layer_norm_first=cfg.layer_norm_first,
+ drop_path=drop_path,
+ norm_eps=1e-6,
+ single_qkv=cfg.single_qkv,
+ fused_ln=cfg.fused_ln,
+ )
+ elif cfg.alt_block:
+ window_size = (
+ cfg.input_size // self.patch_embed.patch_size[0],
+ cfg.input_size // self.patch_embed.patch_size[1],
+ )
+ return AltBlock(
+ cfg.embed_dim,
+ cfg.num_heads,
+ cfg.mlp_ratio,
+ qkv_bias=True,
+ qk_scale=None,
+ norm_layer=norm_layer,
+ drop_path=drop_path,
+ layer_norm_first=cfg.layer_norm_first,
+ ffn_targets=not cfg.end_of_block_targets,
+ use_rel_pos_bias=cfg.use_rel_pos_bias,
+ window_size=window_size
+ if (self.cfg.use_rel_pos_bias and not self.cfg.shared_rel_pos_bias)
+ else None,
+ alt_attention=cfg.alt_attention,
+ )
+ elif cfg.alt_block2:
+ from .multi.modules import AltBlock as AltBlock2
+ return AltBlock2(
+ cfg.embed_dim,
+ cfg.num_heads,
+ cfg.mlp_ratio,
+ qkv_bias=True,
+ qk_scale=None,
+ norm_layer=norm_layer,
+ drop_path=drop_path,
+ layer_norm_first=cfg.layer_norm_first,
+ ffn_targets=not cfg.end_of_block_targets,
+ )
+ else:
+ return Block(
+ cfg.embed_dim,
+ cfg.num_heads,
+ cfg.mlp_ratio,
+ qkv_bias=True,
+ qk_scale=None,
+ norm_layer=norm_layer,
+ drop_path=drop_path,
+ )
+
+ self.blocks = nn.ModuleList([make_block(dpr[i]) for i in range(cfg.depth)])
+ self.norm = norm_layer(cfg.embed_dim)
+ # --------------------------------------------------------------------------
+
+ # --------------------------------------------------------------------------
+ # MAE decoder specifics
+ self.decoder_embed = (
+ nn.Linear(cfg.embed_dim, cfg.decoder_embed_dim, bias=True)
+ if not cfg.no_decoder_embed
+ else None
+ )
+
+ self.mask_token = (
+ nn.Parameter(
+ torch.zeros(
+ 1,
+ 1,
+ cfg.decoder_embed_dim
+ if not cfg.no_decoder_embed
+ else cfg.embed_dim,
+ )
+ )
+ if cfg.mask_noise_std <= 0
+ else None
+ )
+
+ self.decoder_pos_embed = (
+ nn.Parameter(
+ torch.zeros(
+ 1,
+ num_patches + 1,
+ cfg.decoder_embed_dim
+ if not cfg.no_decoder_embed
+ else cfg.embed_dim,
+ ),
+ requires_grad=False,
+ )
+ if not cfg.no_decoder_pos_embed
+ else None
+ )
+
+ self.decoder_blocks = nn.ModuleList(
+ [
+ Block(
+ cfg.decoder_embed_dim,
+ cfg.decoder_num_heads,
+ cfg.mlp_ratio,
+ qkv_bias=True,
+ qk_scale=None,
+ norm_layer=norm_layer,
+ )
+ for _ in range(cfg.decoder_depth)
+ ]
+ )
+
+ self.decoder_norm = norm_layer(cfg.decoder_embed_dim)
+ self.decoder_pred = nn.Linear(
+ cfg.decoder_embed_dim, cfg.patch_size ** 2 * cfg.in_chans, bias=True
+ ) # decoder to patch
+ # --------------------------------------------------------------------------
+
+ self.norm_pix_loss = cfg.norm_pix_loss
+
+ self.initialize_weights()
+
+ for pn, p in self.named_parameters():
+ if len(p.shape) == 1 or pn.endswith(".bias"):
+ p.param_group = "no_decay"
+ else:
+ p.param_group = "with_decay"
+
+ def initialize_weights(self):
+ # initialization
+ # initialize (and freeze) pos_embed by sin-cos embedding
+ pos_embed = get_2d_sincos_pos_embed(
+ self.pos_embed.shape[-1],
+ int(self.patch_embed.num_patches ** 0.5),
+ cls_token=not self.cfg.no_cls,
+ )
+ self.pos_embed.data.copy_(torch.from_numpy(pos_embed).float().unsqueeze(0))
+
+ if self.decoder_pos_embed is not None:
+ decoder_pos_embed = get_2d_sincos_pos_embed(
+ self.decoder_pos_embed.shape[-1],
+ int(self.patch_embed.num_patches ** 0.5),
+ cls_token=not self.cfg.no_cls,
+ )
+ self.decoder_pos_embed.data.copy_(
+ torch.from_numpy(decoder_pos_embed).float().unsqueeze(0)
+ )
+
+ # initialize patch_embed like nn.Linear (instead of nn.Conv2d)
+ w = self.patch_embed.proj.weight.data
+ torch.nn.init.xavier_uniform_(w.view([w.shape[0], -1]))
+
+ # timm's trunc_normal_(std=.02) is effectively normal_(std=0.02) as cutoff is too big (2.)
+ if self.cls_token is not None:
+ torch.nn.init.normal_(self.cls_token, std=0.02)
+
+ if self.mask_token is not None:
+ torch.nn.init.normal_(self.mask_token, std=0.02)
+
+ # initialize nn.Linear and nn.LayerNorm
+ self.apply(self._init_weights)
+
+ def _init_weights(self, m):
+ if isinstance(m, nn.Linear):
+ # we use xavier_uniform following official JAX ViT:
+ torch.nn.init.xavier_uniform_(m.weight)
+ if isinstance(m, nn.Linear) and m.bias is not None:
+ nn.init.constant_(m.bias, 0)
+ elif isinstance(m, nn.LayerNorm) or isinstance(m, FusedLayerNorm):
+ nn.init.constant_(m.bias, 0)
+ nn.init.constant_(m.weight, 1.0)
+
+ def patchify(self, imgs):
+ """
+ imgs: (N, 3, H, W)
+ x: (N, L, patch_size**2 *3)
+ """
+ p = self.patch_embed.patch_size[0]
+ assert imgs.shape[2] == imgs.shape[3] and imgs.shape[2] % p == 0
+
+ h = w = imgs.shape[2] // p
+ x = imgs.reshape(shape=(imgs.shape[0], 3, h, p, w, p))
+ x = torch.einsum("nchpwq->nhwpqc", x)
+ x = x.reshape(shape=(imgs.shape[0], h * w, p ** 2 * 3))
+ return x
+
+ def unpatchify(self, x):
+ """
+ x: (N, L, patch_size**2 *3)
+ imgs: (N, 3, H, W)
+ """
+ p = self.patch_embed.patch_size[0]
+ h = w = int(x.shape[1] ** 0.5)
+ assert h * w == x.shape[1]
+
+ x = x.reshape(shape=(x.shape[0], h, w, p, p, 3))
+ x = torch.einsum("nhwpqc->nchpwq", x)
+ imgs = x.reshape(shape=(x.shape[0], 3, h * p, h * p))
+ return imgs
+
+ def random_masking(self, x, mask_ratio):
+ """
+ Perform per-sample random masking by per-sample shuffling.
+ Per-sample shuffling is done by argsort random noise.
+ x: [N, L, D], sequence
+ """
+ N, L, D = x.shape # batch, length, dim
+ len_keep = int(L * (1 - mask_ratio))
+
+ noise = torch.rand(N, L, device=x.device) # noise in [0, 1]
+
+ # sort noise for each sample
+ ids_shuffle = torch.argsort(
+ noise, dim=1
+ ) # ascend: small is keep, large is remove
+ ids_restore = torch.argsort(ids_shuffle, dim=1)
+
+ # keep the first subset
+ ids_keep = ids_shuffle[:, :len_keep]
+ x_masked = torch.gather(x, dim=1, index=ids_keep.unsqueeze(-1).repeat(1, 1, D))
+
+ # generate the binary mask: 0 is keep, 1 is remove
+ mask = torch.ones([N, L], device=x.device)
+ mask[:, :len_keep] = 0
+ # unshuffle to get the binary mask
+ mask = torch.gather(mask, dim=1, index=ids_restore)
+
+ return x_masked, mask, ids_restore # x_masked is actually unmasked x
+
+ @classmethod
+ def build_model(cls, cfg: MaeConfig, task=None):
+ """Build a new model instance."""
+
+ return cls(cfg)
+
+ def forward_encoder(self, x, mask_ratio):
+ # embed patches
+ x = self.patch_embed(x)
+
+ # add pos embed w/o cls token
+ # if self.cls_token is not None:
+ # x = x + self.pos_embed
+ # else:
+ x = x + self.pos_embed[:, 1:, :]
+
+ # masking: length -> length * mask_ratio
+ if mask_ratio > 0:
+ x, mask, ids_restore = self.random_masking(x, mask_ratio)
+ else:
+ mask = ids_restore = None
+
+ # append cls token
+ if self.cls_token is not None:
+ cls_token = self.cls_token + self.pos_embed[:, :1, :]
+ cls_tokens = cls_token.expand(x.shape[0], -1, -1)
+ x = torch.cat((cls_tokens, x), dim=1)
+
+ # apply Transformer blocks
+ for blk in self.blocks:
+ x = blk(x)
+
+ if self.norm is not None:
+ x = self.norm(x)
+
+ return x, mask, ids_restore
+
+ def forward_decoder(self, x, ids_restore):
+ # embed tokens
+ x = self.decoder_embed(x)
+
+ # append mask tokens to sequence
+ mask_tokens = self.mask_token.repeat(
+ x.shape[0], ids_restore.shape[1] + 1 - x.shape[1], 1
+ )
+ if self.cls_token is not None:
+ x_ = torch.cat([x[:, 1:, :], mask_tokens], dim=1) # no cls token
+ else:
+ x_ = torch.cat([x, mask_tokens], dim=1) # no cls token
+
+ x_ = torch.gather(
+ x_, dim=1, index=ids_restore.unsqueeze(-1).repeat(1, 1, x.shape[2])
+ ) # unshuffle
+
+ if self.cls_token is not None:
+ x = torch.cat([x[:, :1, :], x_], dim=1) # append cls token
+
+ # add pos embed
+ x = x + self.decoder_pos_embed
+
+ # apply Transformer blocks
+ for blk in self.decoder_blocks:
+ x = blk(x)
+ x = self.decoder_norm(x)
+
+ # predictor projection
+ x = self.decoder_pred(x)
+
+ if self.cls_token is not None:
+ # remove cls token
+ x = x[:, 1:, :]
+
+ return x
+
+ def forward_loss(self, imgs, pred, mask):
+ """
+ imgs: [N, 3, H, W]
+ pred: [N, L, p*p*3]
+ mask: [N, L], 0 is keep, 1 is remove,
+ """
+ target = self.patchify(imgs)
+ if self.norm_pix_loss:
+ mean = target.mean(dim=-1, keepdim=True)
+ var = target.var(dim=-1, keepdim=True)
+ target = (target - mean) / (var + 1.0e-6) ** 0.5
+
+ loss = (pred - target) ** 2
+ loss = loss.mean(dim=-1) # [N, L], mean loss per patch
+
+ loss = (loss * mask).sum()
+ return loss, mask.sum()
+
+ def forward(self, imgs, predictions_only=False):
+ latent, mask, ids_restore = self.forward_encoder(
+ imgs, self.mask_ratio if not predictions_only else 0
+ )
+
+ if predictions_only:
+ return latent
+
+ pred = self.forward_decoder(latent, ids_restore) # [N, L, p*p*3]
+ loss, sample_size = self.forward_loss(imgs, pred, mask)
+
+ result = {
+ "losses": {"regression": loss},
+ "sample_size": sample_size,
+ }
+ return result
+
+ def remove_pretraining_modules(self):
+ self.decoder_embed = None
+ self.decoder_blocks = None
+ self.decoder_norm = None
+ self.decoder_pos_embed = None
+ self.decoder_pred = None
+ self.mask_token = None
+ if self.cfg.layer_norm_first:
+ self.norm = None
diff --git a/fairseq/examples/data2vec/models/mae_image_classification.py b/fairseq/examples/data2vec/models/mae_image_classification.py
new file mode 100644
index 0000000000000000000000000000000000000000..e304618dc51668fb644f868c0e33c341e72bd9a7
--- /dev/null
+++ b/fairseq/examples/data2vec/models/mae_image_classification.py
@@ -0,0 +1,386 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+# The code in this file is adapted from the BeiT implementation which can be found here:
+# https://github.com/microsoft/unilm/tree/master/beit
+
+import logging
+
+from dataclasses import dataclass
+from enum import Enum, auto
+from typing import Any, Optional
+
+import numpy as np
+from omegaconf import II, MISSING
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from fairseq import checkpoint_utils, tasks
+from omegaconf import open_dict
+
+from fairseq.dataclass import FairseqDataclass
+from fairseq.models import BaseFairseqModel, register_model
+from .mae import interpolate_pos_embed
+
+
+logger = logging.getLogger(__name__)
+
+
+class PredictionMode(Enum):
+ MEAN_POOLING = auto()
+ CLS_TOKEN = auto()
+ LIN_SOFTMAX = auto()
+
+
+@dataclass
+class MaeImageClassificationConfig(FairseqDataclass):
+ model_path: str = MISSING
+ no_pretrained_weights: bool = False
+ linear_classifier: bool = False
+ num_classes: int = 1000
+ mixup: float = 0.8
+ cutmix: float = 1.0
+ label_smoothing: float = 0.1
+
+ drop_path_rate: float = 0.1
+ layer_decay: float = 0.65
+
+ mixup_prob: float = 1.0
+ mixup_switch_prob: float = 0.5
+ mixup_mode: str = "batch"
+
+ pretrained_model_args: Any = None
+ data: str = II("task.data")
+
+ norm_eps: Optional[float] = None
+
+ remove_alibi: bool = False
+
+ # regularization overwrites
+ encoder_dropout: float = 0
+ post_mlp_drop: float = 0
+ attention_dropout: float = 0
+ activation_dropout: float = 0.0
+ dropout_input: float = 0.0
+ layerdrop: float = 0.0
+
+ prenet_layerdrop: float = 0
+ prenet_dropout: float = 0
+
+ use_fc_norm: bool = True
+ prediction_mode: PredictionMode = PredictionMode.MEAN_POOLING
+
+ no_decay_blocks: bool = True
+
+
+def get_layer_id_for_vit(name, num_layers):
+ """
+ Assign a parameter with its layer id
+ Following BEiT: https://github.com/microsoft/unilm/blob/master/beit/optim_factory.py#L33
+ """
+ if name in ["cls_token", "pos_embed"]:
+ return 0
+ elif name.startswith("patch_embed"):
+ return 0
+ elif name.startswith("rel_pos_bias"):
+ return num_layers - 1
+ elif name.startswith("blocks"):
+ return int(name.split(".")[1]) + 1
+ else:
+ return num_layers
+
+
+@register_model("mae_image_classification", dataclass=MaeImageClassificationConfig)
+class MaeImageClassificationModel(BaseFairseqModel):
+ def __init__(self, cfg: MaeImageClassificationConfig):
+ super().__init__()
+ self.cfg = cfg
+
+ if cfg.pretrained_model_args is None:
+ state = checkpoint_utils.load_checkpoint_to_cpu(cfg.model_path, {})
+ pretrained_args = state.get("cfg", None)
+
+ pretrained_args.criterion = None
+ pretrained_args.lr_scheduler = None
+
+ logger.info(pretrained_args.model)
+
+ with open_dict(pretrained_args.model):
+ pretrained_args.model.drop_path_rate = cfg.drop_path_rate
+ if cfg.norm_eps is not None:
+ pretrained_args.model.norm_eps = cfg.norm_eps
+
+ cfg.pretrained_model_args = pretrained_args
+
+ logger.info(pretrained_args)
+ else:
+ state = None
+ pretrained_args = cfg.pretrained_model_args
+
+ if "data" in pretrained_args.task:
+ pretrained_args.task.data = cfg.data
+ elif "image" in pretrained_args.task:
+ pretrained_args.task.image.data = cfg.data
+
+ if "modalities" in pretrained_args.model:
+ prenet_blocks = pretrained_args.model["modalities"]["image"]["prenet_depth"]
+ model_blocks = pretrained_args.model["depth"]
+ with open_dict(pretrained_args):
+ dpr = np.linspace(0, cfg.drop_path_rate, model_blocks).tolist()
+ pretrained_args.model["modalities"]["image"][
+ "start_drop_path_rate"
+ ] = dpr[0]
+ pretrained_args.model["modalities"]["image"][
+ "end_drop_path_rate"
+ ] = max(0, dpr[prenet_blocks - 1])
+ pretrained_args.model["start_drop_path_rate"] = dpr[prenet_blocks]
+ pretrained_args.model["end_drop_path_rate"] = dpr[-1]
+
+ if "mae_masking" in pretrained_args.model["modalities"]["image"]:
+ del pretrained_args.model["modalities"]["image"]["mae_masking"]
+
+ if cfg.remove_alibi:
+ pretrained_args.model["modalities"]["image"][
+ "use_alibi_encoder"
+ ] = False
+ if (
+ state is not None
+ and "modality_encoders.IMAGE.alibi_bias" in state["model"]
+ ):
+ del state["model"]["modality_encoders.IMAGE.alibi_bias"]
+
+ pretrained_args.model["encoder_dropout"] = cfg.encoder_dropout
+ pretrained_args.model["post_mlp_drop"] = cfg.post_mlp_drop
+ pretrained_args.model["attention_dropout"] = cfg.attention_dropout
+ pretrained_args.model["activation_dropout"] = cfg.activation_dropout
+ pretrained_args.model["dropout_input"] = cfg.dropout_input
+ pretrained_args.model["layerdrop"] = cfg.layerdrop
+
+ pretrained_args.model["modalities"]["image"][
+ "prenet_layerdrop"
+ ] = cfg.prenet_layerdrop
+ pretrained_args.model["modalities"]["image"][
+ "prenet_dropout"
+ ] = cfg.prenet_dropout
+ else:
+ # not d2v multi
+ with open_dict(pretrained_args):
+ pretrained_args.model["drop_path_rate"] = cfg.drop_path_rate
+ pretrained_args.model["block_dropout"] = cfg.encoder_dropout
+ pretrained_args.model["attention_dropout"] = cfg.attention_dropout
+ pretrained_args.model["activation_dropout"] = cfg.activation_dropout
+
+ task = tasks.setup_task(pretrained_args.task)
+ model = task.build_model(pretrained_args.model, from_checkpoint=True)
+
+ self.d2v_multi = "data2vec_multi" in pretrained_args.model._name
+ self.linear_classifier = cfg.linear_classifier
+
+ self.model = model
+
+ if state is not None and not cfg.no_pretrained_weights:
+ interpolate_pos_embed(model, state)
+
+ if "modality_encoders.IMAGE.positional_encoder.pos_embed" in state["model"]:
+ state["model"][
+ "modality_encoders.IMAGE.positional_encoder.positions"
+ ] = state["model"][
+ "modality_encoders.IMAGE.positional_encoder.pos_embed"
+ ]
+ del state["model"][
+ "modality_encoders.IMAGE.positional_encoder.pos_embed"
+ ]
+ if "modality_encoders.IMAGE.encoder_mask" in state["model"]:
+ del state["model"]["modality_encoders.IMAGE.encoder_mask"]
+
+ model.load_state_dict(state["model"], strict=True)
+
+ if self.d2v_multi:
+ model.remove_pretraining_modules(modality="image")
+ else:
+ model.remove_pretraining_modules()
+
+ if self.linear_classifier:
+ model.requires_grad_(False)
+
+ self.fc_norm = None
+ if self.cfg.use_fc_norm:
+ self.fc_norm = nn.LayerNorm(pretrained_args.model.embed_dim, eps=1e-6)
+ nn.init.constant_(self.fc_norm.bias, 0)
+ nn.init.constant_(self.fc_norm.weight, 1.0)
+
+ self.head = nn.Linear(pretrained_args.model.embed_dim, cfg.num_classes)
+
+ nn.init.trunc_normal_(self.head.weight, std=0.02)
+ nn.init.constant_(self.head.bias, 0)
+
+ self.mixup_fn = None
+
+ if cfg.mixup > 0 or cfg.cutmix > 0:
+ from timm.data import Mixup
+
+ self.mixup_fn = Mixup(
+ mixup_alpha=cfg.mixup,
+ cutmix_alpha=cfg.cutmix,
+ cutmix_minmax=None,
+ prob=cfg.mixup_prob,
+ switch_prob=cfg.mixup_switch_prob,
+ mode=cfg.mixup_mode,
+ label_smoothing=cfg.label_smoothing,
+ num_classes=cfg.num_classes,
+ )
+
+ if self.model.norm is not None:
+ for pn, p in self.model.norm.named_parameters():
+ if len(p.shape) == 1 or pn.endswith(".bias"):
+ p.optim_overrides = {"optimizer": {"weight_decay_scale": 0}}
+
+ if self.fc_norm is not None:
+ for pn, p in self.fc_norm.named_parameters():
+ if len(p.shape) == 1 or pn.endswith(".bias"):
+ p.optim_overrides = {"optimizer": {"weight_decay_scale": 0}}
+
+ for pn, p in self.head.named_parameters():
+ if len(p.shape) == 1 or pn.endswith(".bias"):
+ p.optim_overrides = {"optimizer": {"weight_decay_scale": 0}}
+
+ if self.d2v_multi:
+ mod_encs = list(model.modality_encoders.values())
+ assert len(mod_encs) == 1, len(mod_encs)
+ blocks = list(mod_encs[0].context_encoder.blocks) + list(model.blocks)
+ else:
+ blocks = model.blocks
+
+ num_layers = len(blocks) + 1
+ layer_scales = list(
+ cfg.layer_decay ** (num_layers - i) for i in range(num_layers + 1)
+ )
+
+ if self.d2v_multi:
+ for n, p in self.model.named_parameters():
+ optimizer_override_dict = {}
+
+ if len(p.shape) == 1 or n.endswith(".bias"):
+ optimizer_override_dict["weight_decay_scale"] = 0
+
+ p.optim_overrides = {"optimizer": optimizer_override_dict}
+
+ if cfg.layer_decay > 0:
+ for i, b in enumerate(blocks):
+ lid = i + 1
+ if layer_scales[lid] == 1.0:
+ continue
+
+ for n, p in b.named_parameters():
+ optim_override = getattr(p, "optim_overrides", {})
+ if "optimizer" not in optim_override:
+ optim_override["optimizer"] = {}
+
+ if cfg.no_decay_blocks:
+ optim_override["optimizer"]["lr_scale"] = layer_scales[lid]
+ p.optim_overrides = optim_override
+ else:
+ optim_override["optimizer"] = {
+ "lr_scale": layer_scales[lid]
+ }
+ p.optim_overrides = optim_override
+
+ else:
+ for n, p in self.model.named_parameters():
+ optimizer_override_dict = {}
+ layer_id = get_layer_id_for_vit(n, num_layers)
+
+ if len(p.shape) == 1 or n.endswith(".bias"):
+ optimizer_override_dict["weight_decay_scale"] = 0
+
+ if cfg.layer_decay > 0:
+ optimizer_override_dict["lr_scale"] = layer_scales[layer_id]
+ p.optim_overrides = {"optimizer": optimizer_override_dict}
+
+ @classmethod
+ def build_model(cls, cfg: MaeImageClassificationConfig, task=None):
+ """Build a new model instance."""
+
+ return cls(cfg)
+
+ def forward(
+ self,
+ imgs,
+ labels=None,
+ ):
+ if self.training and self.mixup_fn is not None and labels is not None:
+ imgs, labels = self.mixup_fn(imgs, labels)
+
+ if self.linear_classifier:
+ with torch.no_grad():
+ x = self.model_forward(imgs)
+ else:
+ x = self.model_forward(imgs)
+
+ if self.cfg.prediction_mode == PredictionMode.MEAN_POOLING:
+ x = x.mean(dim=1)
+ elif self.cfg.prediction_mode == PredictionMode.CLS_TOKEN:
+ x = x[:, 0]
+ elif self.cfg.prediction_mode == PredictionMode.LIN_SOFTMAX:
+ dtype = x.dtype
+ x = F.logsigmoid(x.float())
+ x = torch.logsumexp(x + x, dim=1) - torch.logsumexp(x + 1e-6, dim=1)
+ x = x.clamp(max=0)
+ x = x - torch.log(-(torch.expm1(x)))
+ x = torch.nan_to_num(x, nan=0, posinf=0, neginf=0)
+ x = x.to(dtype=dtype)
+ else:
+ raise Exception(f"unknown prediction mode {self.cfg.prediction_mode.name}")
+
+ if self.fc_norm is not None:
+ x = self.fc_norm(x)
+
+ x = self.head(x)
+
+ if labels is None:
+ return x
+
+ if self.training and self.mixup_fn is not None:
+ loss = -labels * F.log_softmax(x.float(), dim=-1)
+ else:
+ loss = F.cross_entropy(
+ x.float(),
+ labels,
+ label_smoothing=self.cfg.label_smoothing if self.training else 0,
+ reduction="none",
+ )
+
+ result = {
+ "losses": {"regression": loss},
+ "sample_size": imgs.size(0),
+ }
+
+ if not self.training:
+ with torch.no_grad():
+ pred = x.argmax(-1)
+ correct = (pred == labels).sum()
+ result["correct"] = correct
+
+ return result
+
+ def model_forward(self, imgs):
+ if self.d2v_multi:
+ x = self.model.extract_features(
+ imgs,
+ mode="IMAGE",
+ mask=False,
+ remove_extra_tokens=(
+ self.cfg.prediction_mode != PredictionMode.CLS_TOKEN
+ ),
+ )["x"]
+ else:
+ x = self.model(imgs, predictions_only=True)
+ if (
+ "no_cls" not in self.model.cfg or not self.model.cfg.no_cls
+ ) and not self.cfg.prediction_mode == PredictionMode.CLS_TOKEN:
+ x = x[:, 1:]
+ return x
diff --git a/fairseq/examples/data2vec/models/modalities/__init__.py b/fairseq/examples/data2vec/models/modalities/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/fairseq/examples/data2vec/models/modalities/audio.py b/fairseq/examples/data2vec/models/modalities/audio.py
new file mode 100644
index 0000000000000000000000000000000000000000..80d2857b244436a680808c3aab26638da42be6d5
--- /dev/null
+++ b/fairseq/examples/data2vec/models/modalities/audio.py
@@ -0,0 +1,192 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+from functools import partial
+import torch
+import torch.nn as nn
+import numpy as np
+from dataclasses import dataclass, field
+from typing import Callable, Dict, Optional
+from fairseq.models.wav2vec import ConvFeatureExtractionModel
+from fairseq.modules import (
+ LayerNorm,
+ SamePad,
+ TransposeLast,
+)
+from fairseq.tasks import FairseqTask
+from .base import D2vModalityConfig, ModalitySpecificEncoder, get_alibi_bias
+from .modules import BlockEncoder, Decoder1d
+from examples.data2vec.data.modality import Modality
+
+
+@dataclass
+class D2vAudioConfig(D2vModalityConfig):
+ type: Modality = Modality.AUDIO
+ extractor_mode: str = "layer_norm"
+ feature_encoder_spec: str = field(
+ default="[(512, 10, 5)] + [(512, 3, 2)] * 4 + [(512,2,2)] + [(512,2,2)]",
+ metadata={
+ "help": "string describing convolutional feature extraction layers in form of a python list that contains "
+ "[(dim, kernel_size, stride), ...]"
+ },
+ )
+ conv_pos_width: int = field(
+ default=95,
+ metadata={"help": "number of filters for convolutional positional embeddings"},
+ )
+ conv_pos_groups: int = field(
+ default=16,
+ metadata={"help": "number of groups for convolutional positional embedding"},
+ )
+ conv_pos_depth: int = field(
+ default=5,
+ metadata={"help": "depth of positional encoder network"},
+ )
+ conv_pos_pre_ln: bool = False
+
+
+class AudioEncoder(ModalitySpecificEncoder):
+
+ modality_cfg: D2vAudioConfig
+
+ def __init__(
+ self,
+ modality_cfg: D2vAudioConfig,
+ embed_dim: int,
+ make_block: Callable[[float], nn.ModuleList],
+ norm_layer: Callable[[int], nn.LayerNorm],
+ layer_norm_first: bool,
+ alibi_biases: Dict,
+ task: Optional[FairseqTask],
+ ):
+
+ self.feature_enc_layers = eval(modality_cfg.feature_encoder_spec)
+ feature_embed_dim = self.feature_enc_layers[-1][0]
+
+ local_encoder = ConvFeatureExtractionModel(
+ conv_layers=self.feature_enc_layers,
+ dropout=0.0,
+ mode=modality_cfg.extractor_mode,
+ conv_bias=False,
+ )
+
+ project_features = nn.Sequential(
+ TransposeLast(),
+ nn.LayerNorm(feature_embed_dim),
+ nn.Linear(feature_embed_dim, embed_dim),
+ )
+
+ num_pos_layers = modality_cfg.conv_pos_depth
+ k = max(3, modality_cfg.conv_pos_width // num_pos_layers)
+
+ positional_encoder = nn.Sequential(
+ TransposeLast(),
+ *[
+ nn.Sequential(
+ nn.Conv1d(
+ embed_dim,
+ embed_dim,
+ kernel_size=k,
+ padding=k // 2,
+ groups=modality_cfg.conv_pos_groups,
+ ),
+ SamePad(k),
+ TransposeLast(),
+ LayerNorm(embed_dim, elementwise_affine=False),
+ TransposeLast(),
+ nn.GELU(),
+ )
+ for _ in range(num_pos_layers)
+ ],
+ TransposeLast(),
+ )
+
+ if modality_cfg.conv_pos_pre_ln:
+ positional_encoder = nn.Sequential(LayerNorm(embed_dim), positional_encoder)
+
+ dpr = np.linspace(
+ modality_cfg.start_drop_path_rate,
+ modality_cfg.end_drop_path_rate,
+ modality_cfg.prenet_depth,
+ )
+ context_encoder = BlockEncoder(
+ nn.ModuleList(make_block(dpr[i]) for i in range(modality_cfg.prenet_depth)),
+ norm_layer(embed_dim) if not layer_norm_first else None,
+ layer_norm_first,
+ modality_cfg.prenet_layerdrop,
+ modality_cfg.prenet_dropout,
+ )
+
+ decoder = (
+ Decoder1d(modality_cfg.decoder, embed_dim)
+ if modality_cfg.decoder is not None
+ else None
+ )
+
+ alibi_bias_fn = partial(get_alibi_bias, alibi_biases=alibi_biases)
+
+ super().__init__(
+ modality_cfg=modality_cfg,
+ embed_dim=embed_dim,
+ local_encoder=local_encoder,
+ project_features=project_features,
+ fixed_positional_encoder=None,
+ relative_positional_encoder=positional_encoder,
+ context_encoder=context_encoder,
+ decoder=decoder,
+ get_alibi_bias=alibi_bias_fn,
+ )
+
+ def convert_padding_mask(self, x, padding_mask):
+ def get_feat_extract_output_lengths(input_lengths: torch.LongTensor):
+ """
+ Computes the output length of the convolutional layers
+ """
+
+ def _conv_out_length(input_length, kernel_size, stride):
+ return torch.floor((input_length - kernel_size) / stride + 1)
+
+ for i in range(len(self.feature_enc_layers)):
+ input_lengths = _conv_out_length(
+ input_lengths,
+ self.feature_enc_layers[i][1],
+ self.feature_enc_layers[i][2],
+ )
+
+ return input_lengths.to(torch.long)
+
+ if padding_mask is not None:
+ input_lengths = (1 - padding_mask.long()).sum(-1)
+ # apply conv formula to get real output_lengths
+ output_lengths = get_feat_extract_output_lengths(input_lengths)
+
+ if padding_mask.any():
+ padding_mask = torch.zeros(x.shape[:2], dtype=x.dtype, device=x.device)
+
+ # these two operations makes sure that all values
+ # before the output lengths indices are attended to
+ padding_mask[
+ (
+ torch.arange(padding_mask.shape[0], device=padding_mask.device),
+ output_lengths - 1,
+ )
+ ] = 1
+ padding_mask = (
+ 1 - padding_mask.flip([-1]).cumsum(-1).flip([-1])
+ ).bool()
+ else:
+ padding_mask = torch.zeros(
+ x.shape[:2], dtype=torch.bool, device=x.device
+ )
+
+ return padding_mask
+
+ def reset_parameters(self):
+ super().reset_parameters()
+ for mod in self.project_features.children():
+ if isinstance(mod, nn.Linear):
+ mod.reset_parameters()
+ if self.decoder is not None:
+ self.decoder.reset_parameters()
diff --git a/fairseq/examples/data2vec/models/modalities/base.py b/fairseq/examples/data2vec/models/modalities/base.py
new file mode 100644
index 0000000000000000000000000000000000000000..642cc8466152276c968312ba27a881467bfbd50e
--- /dev/null
+++ b/fairseq/examples/data2vec/models/modalities/base.py
@@ -0,0 +1,684 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import logging
+import math
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from collections import namedtuple
+from dataclasses import dataclass
+from functools import partial
+from omegaconf import MISSING, II
+from typing import Optional, Callable
+from fairseq.data.data_utils import compute_mask_indices
+from fairseq.modules import GradMultiply
+from fairseq.utils import index_put
+from examples.data2vec.data.modality import Modality
+from .modules import D2vDecoderConfig
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class D2vModalityConfig:
+ type: Modality = MISSING
+ prenet_depth: int = 4
+ prenet_layerdrop: float = 0
+ prenet_dropout: float = 0
+ start_drop_path_rate: float = 0
+ end_drop_path_rate: float = 0
+
+ num_extra_tokens: int = 0
+ init_extra_token_zero: bool = True
+
+ mask_noise_std: float = 0.01
+ mask_prob_min: Optional[float] = None
+ mask_prob: float = 0.7
+ inverse_mask: bool = False
+ mask_prob_adjust: float = 0
+ keep_masked_pct: float = 0
+
+ mask_length: int = 5
+ add_masks: bool = False
+ remove_masks: bool = False
+ mask_dropout: float = 0.0
+ encoder_zero_mask: bool = True
+
+ mask_channel_prob: float = 0.0
+ mask_channel_length: int = 64
+
+ ema_local_encoder: bool = False # used in data2vec_multi
+ local_grad_mult: float = 1.0
+
+ use_alibi_encoder: bool = False
+ alibi_scale: float = 1.0
+ learned_alibi: bool = False
+ alibi_max_pos: Optional[int] = None
+ learned_alibi_scale: bool = False
+ learned_alibi_scale_per_head: bool = False
+ learned_alibi_scale_per_layer: bool = False
+
+ num_alibi_heads: int = II("model.num_heads")
+ model_depth: int = II("model.depth")
+
+ decoder: Optional[D2vDecoderConfig] = D2vDecoderConfig()
+
+
+MaskSeed = namedtuple("MaskSeed", ["seed", "update", "ids"])
+MaskInfo = namedtuple("MaskInfo", ["x_unmasked", "mask", "ids_restore", "ids_keep"])
+
+
+class ModalitySpecificEncoder(nn.Module):
+ def __init__(
+ self,
+ modality_cfg: D2vModalityConfig,
+ embed_dim: int,
+ local_encoder: nn.Module,
+ project_features: nn.Module,
+ fixed_positional_encoder: Optional[nn.Module],
+ relative_positional_encoder: Optional[nn.Module],
+ context_encoder: nn.Module,
+ decoder: nn.Module,
+ get_alibi_bias: Optional[Callable[[int, int, str, str], torch.Tensor]],
+ ):
+ super().__init__()
+
+ self.modality_cfg = modality_cfg
+ self.local_encoder = local_encoder
+ self.project_features = project_features
+ self.fixed_positional_encoder = fixed_positional_encoder
+ self.relative_positional_encoder = relative_positional_encoder
+ self.context_encoder = context_encoder
+
+ self.decoder = decoder
+ self.get_alibi_bias = get_alibi_bias if modality_cfg.use_alibi_encoder else None
+
+ self.local_grad_mult = self.modality_cfg.local_grad_mult
+
+ self.extra_tokens = None
+ if modality_cfg.num_extra_tokens > 0:
+ self.extra_tokens = nn.Parameter(
+ torch.zeros(1, modality_cfg.num_extra_tokens, embed_dim)
+ )
+ if not modality_cfg.init_extra_token_zero:
+ nn.init.normal_(self.extra_tokens)
+ elif self.extra_tokens.size(1) > 1:
+ nn.init.normal_(self.extra_tokens[:, 1:])
+
+ self.alibi_scale = None
+ if self.get_alibi_bias is not None:
+ self.alibi_scale = nn.Parameter(
+ torch.full(
+ (
+ (modality_cfg.prenet_depth + modality_cfg.model_depth)
+ if modality_cfg.learned_alibi_scale_per_layer
+ else 1,
+ 1,
+ self.modality_cfg.num_alibi_heads
+ if modality_cfg.learned_alibi_scale_per_head
+ else 1,
+ 1,
+ 1,
+ ),
+ modality_cfg.alibi_scale,
+ dtype=torch.float,
+ ),
+ requires_grad=modality_cfg.learned_alibi_scale,
+ )
+
+ if modality_cfg.learned_alibi and self.get_alibi_bias is not None:
+ assert modality_cfg.alibi_max_pos is not None
+ alibi_bias = self.get_alibi_bias(
+ batch_size=1,
+ time_steps=modality_cfg.alibi_max_pos,
+ heads=modality_cfg.num_alibi_heads,
+ scale=1.0,
+ dtype=torch.float,
+ device="cpu",
+ )
+ self.alibi_bias = nn.Parameter(alibi_bias)
+ self.get_alibi_bias = partial(
+ _learned_alibi_bias, alibi_bias=self.alibi_bias
+ )
+
+ def upgrade_state_dict_named(self, state_dict, name):
+ k = f"{name}.alibi_scale"
+ if k in state_dict and state_dict[k].dim() == 4:
+ state_dict[k] = state_dict[k].unsqueeze(0)
+
+ return state_dict
+
+ def convert_padding_mask(self, x, padding_mask):
+ return padding_mask
+
+ def decoder_input(self, x, mask_info: MaskInfo):
+ inp_drop = self.modality_cfg.decoder.input_dropout
+ if inp_drop > 0:
+ x = F.dropout(x, inp_drop, training=self.training, inplace=True)
+
+ num_extra = self.modality_cfg.num_extra_tokens
+
+ if mask_info is not None:
+ num_masked = mask_info.ids_restore.shape[1] - x.shape[1] + num_extra
+
+ mask_tokens = x.new_empty(
+ x.size(0),
+ num_masked,
+ x.size(-1),
+ ).normal_(0, self.modality_cfg.mask_noise_std)
+
+ x_ = torch.cat([x[:, num_extra:], mask_tokens], dim=1)
+ x = torch.gather(x_, dim=1, index=mask_info.ids_restore)
+
+ if self.modality_cfg.decoder.add_positions_masked:
+ assert self.fixed_positional_encoder is not None
+ pos = self.fixed_positional_encoder(x, None)
+ x = x + (pos * mask_info.mask.unsqueeze(-1))
+ else:
+ x = x[:, num_extra:]
+
+ if self.modality_cfg.decoder.add_positions_all:
+ assert self.fixed_positional_encoder is not None
+ x = x + self.fixed_positional_encoder(x, None)
+
+ return x, mask_info
+
+ def local_features(self, features):
+ if self.local_grad_mult > 0:
+ if self.local_grad_mult == 1.0:
+ x = self.local_encoder(features)
+ else:
+ x = GradMultiply.apply(
+ self.local_encoder(features), self.local_grad_mult
+ )
+ else:
+ with torch.no_grad():
+ x = self.local_encoder(features)
+
+ x = self.project_features(x)
+ return x
+
+ def contextualized_features(
+ self,
+ x,
+ padding_mask,
+ mask,
+ remove_masked,
+ clone_batch: int = 1,
+ mask_seeds: Optional[torch.Tensor] = None,
+ precomputed_mask=None,
+ ):
+
+ if padding_mask is not None:
+ padding_mask = self.convert_padding_mask(x, padding_mask)
+
+ local_features = x
+ if mask and clone_batch == 1:
+ local_features = local_features.clone()
+
+ orig_B, orig_T, _ = x.shape
+ pre_mask_B = orig_B
+ mask_info = None
+
+ x_pos = None
+ if self.fixed_positional_encoder is not None:
+ x = x + self.fixed_positional_encoder(x, padding_mask)
+
+ if mask:
+ if clone_batch > 1:
+ x = x.repeat_interleave(clone_batch, 0)
+ if mask_seeds is not None:
+ clone_hash = [
+ int(hash((mask_seeds.seed, ind)) % 1e10)
+ for ind in range(clone_batch - 1)
+ ]
+ clone_hash = torch.tensor([0] + clone_hash).long().view(1, -1)
+
+ id = mask_seeds.ids
+ id = id.repeat_interleave(clone_batch, 0)
+ id = id.view(-1, clone_batch) + clone_hash.to(id)
+ id = id.view(-1)
+ mask_seeds = MaskSeed(
+ seed=mask_seeds.seed, update=mask_seeds.update, ids=id
+ )
+ if padding_mask is not None:
+ padding_mask = padding_mask.repeat_interleave(clone_batch, 0)
+
+ x, mask_info = self.compute_mask(
+ x,
+ padding_mask,
+ mask_seed=mask_seeds,
+ apply=self.relative_positional_encoder is not None or not remove_masked,
+ precomputed_mask=precomputed_mask,
+ )
+
+ if self.relative_positional_encoder is not None:
+ x_pos = self.relative_positional_encoder(x)
+
+ masked_padding_mask = padding_mask
+ if mask and remove_masked:
+ x = mask_info.x_unmasked
+ if x_pos is not None:
+ x = x + gather_unmasked(x_pos, mask_info)
+
+ if padding_mask is not None and padding_mask.any():
+ masked_padding_mask = gather_unmasked_mask(padding_mask, mask_info)
+ if not masked_padding_mask.any():
+ masked_padding_mask = None
+ else:
+ masked_padding_mask = None
+
+ elif x_pos is not None:
+ x = x + x_pos
+
+ alibi_bias = None
+ alibi_scale = self.alibi_scale
+
+ if self.get_alibi_bias is not None:
+ alibi_bias = self.get_alibi_bias(
+ batch_size=pre_mask_B,
+ time_steps=orig_T,
+ heads=self.modality_cfg.num_alibi_heads,
+ dtype=torch.float32,
+ device=x.device,
+ )
+
+ if alibi_scale is not None:
+ alibi_scale = alibi_scale.clamp_min(0)
+ if alibi_scale.size(0) == 1:
+ alibi_bias = alibi_bias * alibi_scale.squeeze(0).type_as(alibi_bias)
+ alibi_scale = None
+
+ if clone_batch > 1:
+ alibi_bias = alibi_bias.repeat_interleave(clone_batch, 0)
+
+ if mask_info is not None and remove_masked:
+ alibi_bias = masked_alibi(alibi_bias, mask_info)
+
+ if self.extra_tokens is not None:
+ num = self.extra_tokens.size(1)
+ x = torch.cat([self.extra_tokens.expand(x.size(0), -1, -1), x], dim=1)
+ if masked_padding_mask is not None:
+ # B x T
+ masked_padding_mask = F.pad(masked_padding_mask, (num, 0))
+ if alibi_bias is not None:
+ # B x H x T x T
+ alibi_bias = F.pad(alibi_bias, (num, 0, num, 0))
+
+ x = self.context_encoder(
+ x,
+ masked_padding_mask,
+ alibi_bias,
+ alibi_scale[: self.modality_cfg.prenet_depth]
+ if alibi_scale is not None
+ else None,
+ )
+
+ return {
+ "x": x,
+ "local_features": local_features,
+ "padding_mask": masked_padding_mask,
+ "alibi_bias": alibi_bias,
+ "alibi_scale": alibi_scale[self.modality_cfg.prenet_depth :]
+ if alibi_scale is not None and alibi_scale.size(0) > 1
+ else alibi_scale,
+ "encoder_mask": mask_info,
+ }
+
+ def forward(
+ self,
+ features,
+ padding_mask,
+ mask: bool,
+ remove_masked: bool,
+ clone_batch: int = 1,
+ mask_seeds: Optional[torch.Tensor] = None,
+ precomputed_mask=None,
+ ):
+ x = self.local_features(features)
+ return self.contextualized_features(
+ x,
+ padding_mask,
+ mask,
+ remove_masked,
+ clone_batch,
+ mask_seeds,
+ precomputed_mask,
+ )
+
+ def reset_parameters(self):
+ pass
+
+ def compute_mask(
+ self,
+ x,
+ padding_mask,
+ mask_seed: Optional[MaskSeed],
+ apply,
+ precomputed_mask,
+ ):
+ if precomputed_mask is not None:
+ mask = precomputed_mask
+ mask_info = self.make_maskinfo(x, mask)
+ else:
+ B, T, C = x.shape
+ cfg = self.modality_cfg
+
+ mask_prob = cfg.mask_prob
+
+ if (
+ cfg.mask_prob_min is not None
+ and cfg.mask_prob_min >= 0
+ and cfg.mask_prob_min < mask_prob
+ ):
+ mask_prob = np.random.uniform(cfg.mask_prob_min, mask_prob)
+
+ if mask_prob > 0:
+ if cfg.mask_length == 1:
+ mask_info = random_masking(x, mask_prob, mask_seed)
+ else:
+ if self.modality_cfg.inverse_mask:
+ mask_prob = 1 - mask_prob
+
+ mask = compute_mask_indices(
+ (B, T),
+ padding_mask,
+ mask_prob,
+ cfg.mask_length,
+ min_masks=1,
+ require_same_masks=True,
+ mask_dropout=cfg.mask_dropout,
+ add_masks=cfg.add_masks,
+ seed=mask_seed.seed if mask_seed is not None else None,
+ epoch=mask_seed.update if mask_seed is not None else None,
+ indices=mask_seed.ids if mask_seed is not None else None,
+ )
+
+ mask = torch.from_numpy(mask).to(device=x.device)
+ if self.modality_cfg.inverse_mask:
+ mask = 1 - mask
+ mask_info = self.make_maskinfo(x, mask)
+ else:
+ mask_info = None
+
+ if apply:
+ x = self.apply_mask(x, mask_info)
+
+ return x, mask_info
+
+ def make_maskinfo(self, x, mask, shape=None):
+ if shape is None:
+ B, T, D = x.shape
+ else:
+ B, T, D = shape
+
+ mask = mask.to(torch.uint8)
+ ids_shuffle = mask.argsort(dim=1)
+ ids_restore = ids_shuffle.argsort(dim=1).unsqueeze(-1).expand(-1, -1, D)
+
+ len_keep = T - mask[0].sum()
+ if self.modality_cfg.keep_masked_pct > 0:
+ len_keep += round((T - int(len_keep)) * self.modality_cfg.keep_masked_pct)
+
+ ids_keep = ids_shuffle[:, :len_keep]
+
+ if shape is not None:
+ x_unmasked = None
+ else:
+ ids_keep = ids_keep.unsqueeze(-1).expand(-1, -1, D)
+ x_unmasked = torch.gather(x, dim=1, index=ids_keep)
+
+ mask_info = MaskInfo(
+ x_unmasked=x_unmasked,
+ mask=mask,
+ ids_restore=ids_restore,
+ ids_keep=ids_keep,
+ )
+ return mask_info
+
+ def apply_mask(self, x, mask_info):
+ cfg = self.modality_cfg
+ B, T, C = x.shape
+
+ if mask_info is not None:
+ mask = mask_info.mask
+ if cfg.encoder_zero_mask:
+ x = x * (1 - mask.type_as(x).unsqueeze(-1))
+ else:
+ num_masks = mask.sum().item()
+ masks = x.new_empty(num_masks, x.size(-1)).normal_(
+ 0, cfg.mask_noise_std
+ )
+ x = index_put(x, mask, masks)
+ if cfg.mask_channel_prob > 0:
+ mask_channel = compute_mask_indices(
+ (B, C),
+ None,
+ cfg.mask_channel_prob,
+ cfg.mask_channel_length,
+ )
+ mask_channel = (
+ torch.from_numpy(mask_channel)
+ .to(x.device)
+ .unsqueeze(1)
+ .expand(-1, T, -1)
+ )
+ x = index_put(x, mask_channel, 0)
+ return x
+
+ def remove_pretraining_modules(self, keep_decoder=False):
+ if not keep_decoder:
+ self.decoder = None
+
+
+def get_annealed_rate(start, end, curr_step, total_steps):
+ if curr_step >= total_steps:
+ return end
+ r = end - start
+ pct_remaining = 1 - curr_step / total_steps
+ return end - r * pct_remaining
+
+
+# adapted from MAE
+def random_masking(x, mask_ratio, mask_seed: Optional[MaskSeed]):
+ N, L, D = x.shape # batch, length, dim
+ len_keep = int(L * (1 - mask_ratio))
+
+ generator = None
+ if mask_seed is not None:
+ seed = int(
+ hash((mask_seed.seed, mask_seed.update, mask_seed.ids.sum().item())) % 1e6
+ )
+ generator = torch.Generator(device=x.device)
+ generator.manual_seed(seed)
+
+ noise = torch.rand(N, L, generator=generator, device=x.device) # noise in [0, 1]
+
+ # sort noise for each sample
+ ids_shuffle = noise.argsort(dim=1) # ascend: small is keep, large is remove
+ ids_restore = ids_shuffle.argsort(dim=1)
+
+ # keep the first subset
+ ids_keep = ids_shuffle[:, :len_keep]
+ ids_keep = ids_keep.unsqueeze(-1).expand(-1, -1, D)
+ x_unmasked = torch.gather(x, dim=1, index=ids_keep)
+
+ # generate the binary mask: 0 is keep, 1 is remove
+ mask = torch.ones([N, L], dtype=x.dtype, device=x.device)
+ mask[:, :len_keep] = 0
+ # unshuffle to get the binary mask
+ mask = torch.gather(mask, dim=1, index=ids_restore)
+
+ ids_restore = ids_restore.unsqueeze(-1).expand(-1, -1, D)
+
+ return MaskInfo(
+ x_unmasked=x_unmasked, mask=mask, ids_restore=ids_restore, ids_keep=ids_keep
+ )
+
+
+def gather_unmasked(x: torch.Tensor, mask_info: MaskInfo) -> torch.Tensor:
+ return torch.gather(
+ x,
+ dim=1,
+ index=mask_info.ids_keep,
+ )
+
+
+def gather_unmasked_mask(x: torch.Tensor, mask_info: MaskInfo) -> torch.Tensor:
+ return torch.gather(
+ x,
+ dim=1,
+ index=mask_info.ids_keep[..., 0], # ignore the feature dimension
+ )
+
+
+def get_alibi(
+ max_positions: int,
+ attention_heads: int,
+ dims: int = 1,
+ distance: str = "manhattan",
+):
+ def get_slopes(n):
+ def get_slopes_power_of_2(n):
+ start = 2 ** (-(2 ** -(math.log2(n) - 3)))
+ ratio = start
+ return [start * ratio**i for i in range(n)]
+
+ # In the paper, we only train models that have 2^a heads for some
+ # a. This function has some good properties that only occur when
+ # the input is a power of 2. To maintain that even when the number
+ # of heads is not a power of 2, we use this workaround.
+ if math.log2(n).is_integer():
+ return get_slopes_power_of_2(n)
+ else:
+ closest_power_of_2 = 2 ** math.floor(math.log2(n))
+ return (
+ get_slopes_power_of_2(closest_power_of_2)
+ + get_slopes(2 * closest_power_of_2)[0::2][: n - closest_power_of_2]
+ )
+
+ maxpos = max_positions
+ attn_heads = attention_heads
+ slopes = torch.Tensor(get_slopes(attn_heads))
+
+ if dims == 1:
+ # prepare alibi position linear bias. Note that wav2vec2 is non
+ # autoregressive model so we want a symmetric mask with 0 on the
+ # diagonal and other wise linear decreasing valuees
+ pos_bias = (
+ torch.abs(
+ torch.arange(maxpos).unsqueeze(0) - torch.arange(maxpos).unsqueeze(1)
+ )
+ * -1
+ )
+ elif dims == 2:
+ if distance == "manhattan":
+ df = lambda x1, y1, x2, y2: abs(x1 - x2) + abs(y1 - y2)
+ elif distance == "euclidean":
+ df = lambda x1, y1, x2, y2: math.sqrt((x1 - x2) ** 2 + (y1 - y2) ** 2)
+
+ n = math.sqrt(max_positions)
+ assert n.is_integer(), n
+ n = int(n)
+
+ pos_bias = torch.zeros((max_positions, max_positions))
+
+ for i in range(n):
+ for j in range(n):
+ for k in range(n):
+ for l in range(n):
+ new_x = i * n + j
+ new_y = k * n + l
+ pos_bias[new_x, new_y] = -df(i, j, k, l)
+
+ else:
+ raise Exception(f"unsupported number of alibi dims: {dims}")
+
+ alibi_bias = slopes.unsqueeze(1).unsqueeze(1) * pos_bias.unsqueeze(0).expand(
+ attn_heads, -1, -1
+ )
+
+ return alibi_bias
+
+
+def get_alibi_bias(
+ alibi_biases,
+ batch_size,
+ time_steps,
+ heads,
+ dtype,
+ device,
+ dims=1,
+ distance="manhattan",
+):
+ cache_key = f"{dims}_{heads}_{distance}"
+
+ buffered = alibi_biases.get(cache_key, None)
+
+ target_size = heads * batch_size
+ if (
+ buffered is None
+ or buffered.size(0) < target_size
+ or buffered.size(1) < time_steps
+ or buffered.dtype != dtype
+ or buffered.device != device
+ ):
+ bt = max(time_steps, buffered.size(1) if buffered is not None else 0)
+ bn = max(target_size, buffered.size(0) if buffered is not None else 0) // heads
+
+ buffered = (
+ get_alibi(bt, heads, dims=dims, distance=distance)
+ .to(dtype=dtype, device=device)
+ .repeat(bn, 1, 1)
+ )
+
+ alibi_biases[cache_key] = buffered
+
+ b = buffered[:target_size, :time_steps, :time_steps]
+ b = b.view(batch_size, heads, time_steps, time_steps)
+ return b
+
+
+def _learned_alibi_bias(
+ alibi_bias,
+ batch_size,
+ time_steps,
+ heads,
+ scale,
+ dtype,
+ device,
+):
+ assert alibi_bias.size(1) == heads, alibi_bias.shape
+ assert alibi_bias.dtype == dtype, alibi_bias.dtype
+ assert alibi_bias.device == device, alibi_bias.device
+
+ if alibi_bias.size(-1) < time_steps:
+ psz = math.ceil((time_steps - alibi_bias.size(-1)) / 2)
+ alibi_bias = F.pad(alibi_bias, (psz, psz, psz, psz), mode="replicate")
+
+ alibi_bias = alibi_bias.expand(batch_size, -1, -1, -1) * scale
+ return alibi_bias[..., :time_steps, :time_steps]
+
+
+def masked_alibi(alibi_bias, mask_info):
+ H = alibi_bias.size(1)
+
+ orig_bias = alibi_bias
+
+ index = mask_info.ids_keep.unsqueeze(1)[..., 0].unsqueeze(-1)
+ alibi_bias = torch.gather(
+ orig_bias,
+ dim=-2,
+ index=index.expand(-1, H, -1, mask_info.ids_restore.size(1)),
+ )
+ alibi_bias = torch.gather(
+ alibi_bias,
+ dim=-1,
+ index=index.transpose(-1, -2).expand(-1, H, alibi_bias.size(-2), -1),
+ )
+
+ return alibi_bias
diff --git a/fairseq/examples/data2vec/models/modalities/images.py b/fairseq/examples/data2vec/models/modalities/images.py
new file mode 100644
index 0000000000000000000000000000000000000000..a6b738cb075312ce90b72594e6f1b8368223a6f1
--- /dev/null
+++ b/fairseq/examples/data2vec/models/modalities/images.py
@@ -0,0 +1,256 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import numpy as np
+from functools import partial
+from dataclasses import dataclass
+from typing import Callable, Dict, Optional
+from timm.models.layers import to_2tuple
+from fairseq.tasks import FairseqTask
+from examples.data2vec.models.mae import get_2d_sincos_pos_embed, PatchEmbed
+from .base import (
+ D2vModalityConfig,
+ ModalitySpecificEncoder,
+ get_alibi_bias,
+ MaskSeed,
+)
+from .modules import (
+ BlockEncoder,
+ Decoder2d,
+ FixedPositionalEncoder,
+ TransformerDecoder,
+ EncDecTransformerDecoder,
+)
+from examples.data2vec.data.modality import Modality
+
+
+@dataclass
+class D2vImageConfig(D2vModalityConfig):
+ type: Modality = Modality.IMAGE
+
+ input_size: int = 224
+ in_chans: int = 3
+ patch_size: int = 16
+ embed_dim: int = 768
+
+ alibi_dims: int = 2
+ alibi_distance: str = "manhattan"
+
+ fixed_positions: bool = True
+
+ transformer_decoder: bool = False
+ enc_dec_transformer: bool = False
+
+
+class ImageEncoder(ModalitySpecificEncoder):
+
+ modality_cfg: D2vImageConfig
+
+ def __init__(
+ self,
+ modality_cfg: D2vImageConfig,
+ embed_dim: int,
+ make_block: Callable[[float, Optional[int], Optional[int]], nn.ModuleList],
+ norm_layer: Callable[[int], nn.LayerNorm],
+ layer_norm_first: bool,
+ alibi_biases: Dict,
+ task: Optional[FairseqTask],
+ ):
+
+ img_size = to_2tuple(modality_cfg.input_size)
+ patch_size = to_2tuple(modality_cfg.patch_size)
+ num_patches = (img_size[1] // patch_size[1]) * (img_size[0] // patch_size[0])
+
+ local_encoder = PatchEmbed(
+ modality_cfg.input_size,
+ modality_cfg.patch_size,
+ modality_cfg.in_chans,
+ modality_cfg.embed_dim,
+ )
+
+ w = local_encoder.proj.weight.data
+ torch.nn.init.xavier_uniform_(w.view([w.shape[0], -1]))
+
+ if modality_cfg.embed_dim != embed_dim:
+ local_encoder = nn.Sequential(
+ local_encoder,
+ nn.Linear(modality_cfg.embed_dim, embed_dim),
+ )
+
+ project_features = nn.Identity()
+
+ pos_embed = nn.Parameter(
+ torch.zeros(1, num_patches, embed_dim), requires_grad=False
+ )
+
+ side_n = int(num_patches ** 0.5)
+
+ emb = get_2d_sincos_pos_embed(
+ pos_embed.shape[-1],
+ side_n,
+ cls_token=False,
+ )
+ pos_embed.data.copy_(torch.from_numpy(emb).float().unsqueeze(0))
+ fixed_positional_encoder = (
+ FixedPositionalEncoder(pos_embed) if modality_cfg.fixed_positions else None
+ )
+
+ dpr = np.linspace(
+ modality_cfg.start_drop_path_rate,
+ modality_cfg.end_drop_path_rate,
+ modality_cfg.prenet_depth,
+ )
+
+ context_encoder = BlockEncoder(
+ nn.ModuleList(make_block(dpr[i]) for i in range(modality_cfg.prenet_depth)),
+ norm_layer(embed_dim) if not layer_norm_first else None,
+ layer_norm_first,
+ modality_cfg.prenet_layerdrop,
+ modality_cfg.prenet_dropout,
+ )
+
+ if modality_cfg.transformer_decoder:
+ if modality_cfg.enc_dec_transformer:
+ decoder = EncDecTransformerDecoder(modality_cfg.decoder, embed_dim)
+ else:
+ dec_enc = BlockEncoder(
+ nn.ModuleList(
+ make_block(0, modality_cfg.decoder.decoder_dim, 8)
+ for _ in range(modality_cfg.decoder.decoder_layers)
+ ),
+ None,
+ layer_norm_first,
+ 0,
+ 0,
+ )
+ decoder = TransformerDecoder(modality_cfg.decoder, embed_dim, dec_enc)
+ else:
+ decoder = (
+ Decoder2d(modality_cfg.decoder, embed_dim, side_n, side_n)
+ if modality_cfg.decoder is not None
+ else None
+ )
+
+ alibi_bias_fn = partial(
+ get_alibi_bias,
+ alibi_biases=alibi_biases,
+ heads=modality_cfg.num_alibi_heads,
+ dims=modality_cfg.alibi_dims,
+ distance=modality_cfg.alibi_distance,
+ )
+
+ super().__init__(
+ modality_cfg=modality_cfg,
+ embed_dim=embed_dim,
+ local_encoder=local_encoder,
+ project_features=project_features,
+ fixed_positional_encoder=fixed_positional_encoder,
+ relative_positional_encoder=None,
+ context_encoder=context_encoder,
+ decoder=decoder,
+ get_alibi_bias=alibi_bias_fn,
+ )
+
+ def reset_parameters(self):
+ super().reset_parameters()
+ if self.decoder is not None:
+ self.decoder.reset_parameters()
+
+ @torch.no_grad()
+ def patchify(self, imgs):
+ """
+ imgs: (N, 3, H, W)
+ x: (N, L, patch_size**2 *3)
+ """
+ p = self.modality_cfg.patch_size
+ h = w = imgs.shape[2] // p
+ x = imgs.reshape(shape=(imgs.shape[0], 3, h, p, w, p))
+ x = torch.einsum("nchpwq->nhwpqc", x)
+ x = x.reshape(shape=(imgs.shape[0], h * w, p ** 2 * 3))
+
+ return x
+
+ @torch.no_grad()
+ def unpatchify(self, x):
+ """
+ x: (N, L, patch_size**2 *3)
+ imgs: (N, 3, H, W)
+ """
+ p = self.modality_cfg.patch_size
+ h = w = int(x.shape[1] ** 0.5)
+ assert h * w == x.shape[1]
+
+ x = x.reshape(shape=(x.shape[0], h, w, p, p, 3))
+ x = torch.einsum("nhwpqc->nchpwq", x)
+ imgs = x.reshape(shape=(x.shape[0], 3, h * p, h * p))
+ return imgs
+
+ def compute_mask(
+ self,
+ x,
+ padding_mask,
+ mask_seed: Optional[MaskSeed],
+ apply,
+ shape=None,
+ precomputed_mask=None,
+ ):
+ mlen = self.modality_cfg.mask_length
+ if mlen <= 1:
+ return super().compute_mask(
+ x, padding_mask, mask_seed, apply, precomputed_mask
+ )
+
+ if precomputed_mask is not None:
+ mask = precomputed_mask
+ else:
+ from fairseq.data.data_utils import compute_block_mask_2d
+
+ if shape is not None:
+ B, L, D = shape
+ else:
+ B, L, D = x.shape
+
+ mask = compute_block_mask_2d(
+ shape=(B, L),
+ mask_prob=self.modality_cfg.mask_prob,
+ mask_length=self.modality_cfg.mask_length,
+ mask_prob_adjust=self.modality_cfg.mask_prob_adjust,
+ inverse_mask=self.modality_cfg.inverse_mask,
+ require_same_masks=True,
+ mask_dropout=self.modality_cfg.mask_dropout,
+ )
+
+ mask_info = self.make_maskinfo(x, mask, shape)
+ if apply:
+ x = self.apply_mask(x, mask_info)
+
+ return x, mask_info
+
+ def decoder_input(self, x, mask_info):
+ if (
+ not self.modality_cfg.transformer_decoder
+ or not self.modality_cfg.enc_dec_transformer
+ ):
+ return super().decoder_input(x, mask_info)
+
+ inp_drop = self.modality_cfg.decoder.input_dropout
+ if inp_drop > 0:
+ x = F.dropout(x, inp_drop, training=self.training, inplace=True)
+
+ kv = x[:, self.modality_cfg.num_extra_tokens :]
+
+ assert self.fixed_positional_encoder is not None
+ pos = self.fixed_positional_encoder(x, None).expand(x.size(0), -1, -1)
+
+ mask = mask_info.mask.bool()
+ if self.modality_cfg.decoder.add_positions_all:
+ kv = kv + pos[~mask].view(kv.shape)
+
+ q = pos[mask].view(x.size(0), -1, x.size(-1))
+
+ return q, kv
diff --git a/fairseq/examples/data2vec/models/modalities/modules.py b/fairseq/examples/data2vec/models/modalities/modules.py
new file mode 100644
index 0000000000000000000000000000000000000000..a4e1a4ea07bf3bbf92ec06ea260b0d9a733ff35d
--- /dev/null
+++ b/fairseq/examples/data2vec/models/modalities/modules.py
@@ -0,0 +1,589 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import numpy as np
+from dataclasses import dataclass
+from fairseq.modules import (
+ LayerNorm,
+ SamePad,
+ SamePad2d,
+ TransposeLast,
+)
+
+
+@dataclass
+class D2vDecoderConfig:
+ decoder_dim: int = 384
+ decoder_groups: int = 16
+ decoder_kernel: int = 5
+ decoder_layers: int = 5
+ input_dropout: float = 0.1
+
+ add_positions_masked: bool = False
+ add_positions_all: bool = False
+
+ decoder_residual: bool = True
+ projection_layers: int = 1
+ projection_ratio: float = 2.0
+
+
+class FixedPositionalEncoder(nn.Module):
+ def __init__(self, pos_embed):
+ super().__init__()
+ self.positions = pos_embed
+
+ def forward(self, x, padding_mask):
+ return self.positions
+
+
+class TextFeatPositionalEncoder(nn.Module):
+ """
+ Original encoder expects (B, T) long input. This module wraps it to take
+ local_encoder output which are (B, T, D) float tensors
+ """
+
+ def __init__(self, pos_encoder):
+ super().__init__()
+ self.pos_encoder = pos_encoder
+
+ def forward(self, x, padding_mask):
+ # assume padded token embeddings are 0s
+ # TODO: consider using padding_mask as input
+ return self.pos_encoder(x[..., 0])
+
+
+class BlockEncoder(nn.Module):
+ def __init__(self, blocks, norm_layer, layer_norm_first, layerdrop, dropout):
+ super().__init__()
+ self.blocks = blocks
+ self.norm = norm_layer
+ self.layer_norm_first = layer_norm_first
+ self.layerdrop = layerdrop
+ self.dropout = nn.Dropout(dropout, inplace=True)
+
+ def forward(self, x, padding_mask, alibi_bias, alibi_scale):
+ if self.norm is not None and not self.layer_norm_first:
+ x = self.norm(x)
+
+ x = self.dropout(x)
+
+ for i, blk in enumerate(self.blocks):
+ if (
+ not self.training
+ or self.layerdrop == 0
+ or (np.random.random() > self.layerdrop)
+ ):
+ ab = alibi_bias
+ if ab is not None and alibi_scale is not None:
+ scale = (
+ alibi_scale[i]
+ if alibi_scale.size(0) > 1
+ else alibi_scale.squeeze(0)
+ )
+ ab = ab * scale.type_as(ab)
+ x, _ = blk(x, padding_mask, ab)
+
+ if self.norm is not None and self.layer_norm_first:
+ x = self.norm(x)
+
+ return x
+
+
+class DecoderBase(nn.Module):
+ decoder_cfg: D2vDecoderConfig
+
+ def __init__(self, cfg: D2vDecoderConfig):
+ super().__init__()
+
+ self.decoder_cfg = cfg
+
+ def reset_parameters(self):
+ for mod in self.proj.modules():
+ if isinstance(mod, nn.Linear):
+ mod.reset_parameters()
+
+ def add_residual(self, x, residual, i, mask_info):
+ if (
+ residual is None
+ or not self.decoder_cfg.decoder_residual
+ or residual.size(1) != x.size(1)
+ ):
+ return x
+
+ ret = x + residual
+
+ return ret
+
+
+class Decoder1d(DecoderBase):
+ def __init__(self, cfg: D2vDecoderConfig, input_dim):
+ super().__init__(cfg)
+
+ def make_block(in_dim):
+ block = [
+ nn.Conv1d(
+ in_dim,
+ cfg.decoder_dim,
+ kernel_size=cfg.decoder_kernel,
+ padding=cfg.decoder_kernel // 2,
+ groups=cfg.decoder_groups,
+ ),
+ SamePad(cfg.decoder_kernel),
+ TransposeLast(),
+ LayerNorm(cfg.decoder_dim, elementwise_affine=False),
+ TransposeLast(),
+ nn.GELU(),
+ ]
+
+ return nn.Sequential(*block)
+
+ self.blocks = nn.Sequential(
+ *[
+ make_block(input_dim if i == 0 else cfg.decoder_dim)
+ for i in range(cfg.decoder_layers)
+ ]
+ )
+
+ projs = []
+ curr_dim = cfg.decoder_dim
+ for i in range(cfg.projection_layers - 1):
+ next_dim = int(curr_dim * cfg.projection_ratio) if i == 0 else curr_dim
+ projs.append(nn.Linear(curr_dim, next_dim))
+ projs.append(nn.GELU())
+ curr_dim = next_dim
+ projs.append(nn.Linear(curr_dim, input_dim))
+ if len(projs) == 1:
+ self.proj = projs[0]
+ else:
+ self.proj = nn.Sequential(*projs)
+
+ def forward(self, x, mask_info):
+
+ x = x.transpose(1, 2)
+
+ residual = x
+
+ for i, layer in enumerate(self.blocks):
+ x = layer(x)
+ x = self.add_residual(x, residual, i, mask_info)
+ residual = x
+
+ x = x.transpose(1, 2)
+ x = self.proj(x)
+ return x
+
+
+class Decoder2d(DecoderBase):
+ def __init__(self, cfg: D2vDecoderConfig, input_dim, h_size, w_size):
+ super().__init__(cfg)
+
+ self.h_size = h_size
+ self.w_size = w_size
+
+ def make_block(in_dim):
+ block = [
+ nn.Conv2d(
+ in_dim,
+ cfg.decoder_dim,
+ kernel_size=cfg.decoder_kernel,
+ padding=cfg.decoder_kernel // 2,
+ groups=cfg.decoder_groups,
+ ),
+ SamePad2d(cfg.decoder_kernel),
+ TransposeLast(tranpose_dim=-3),
+ LayerNorm(cfg.decoder_dim, elementwise_affine=False),
+ TransposeLast(tranpose_dim=-3),
+ nn.GELU(),
+ ]
+
+ return nn.Sequential(*block)
+
+ self.blocks = nn.Sequential(
+ *[
+ make_block(input_dim if i == 0 else cfg.decoder_dim)
+ for i in range(cfg.decoder_layers)
+ ]
+ )
+
+ self.proj = nn.Linear(cfg.decoder_dim, input_dim)
+
+ def forward(self, x, mask_info):
+ B, T, C = x.shape
+
+ x = x.transpose(1, 2).reshape(B, C, self.h_size, self.w_size)
+
+ residual = x
+
+ for i, layer in enumerate(self.blocks):
+ x = layer(x)
+ x = self.add_residual(x, residual, i, mask_info)
+ residual = x
+
+ x = x.reshape(B, -1, T).transpose(1, 2)
+ x = self.proj(x)
+ return x
+
+
+class TransformerDecoder(nn.Module):
+ decoder_cfg: D2vDecoderConfig
+
+ def __init__(self, cfg: D2vDecoderConfig, input_dim, encoder):
+ super().__init__()
+
+ self.decoder_cfg = cfg
+
+ self.input_proj = nn.Linear(input_dim, cfg.decoder_dim)
+
+ self.encoder = encoder
+
+ self.proj = nn.Linear(cfg.decoder_dim, input_dim)
+
+ def reset_parameters(self):
+ from fairseq.modules.transformer_sentence_encoder import init_bert_params
+
+ self.apply(init_bert_params)
+
+ def forward(self, x, mask_info):
+ x = self.input_proj(x)
+ x = self.encoder(x, None, None, 1)
+ x = self.proj(x)
+ return x
+
+
+class AltBlock(nn.Module):
+ def __init__(
+ self,
+ dim,
+ num_heads,
+ mlp_ratio=4.0,
+ qkv_bias=False,
+ qk_scale=None,
+ drop=0.0,
+ attn_drop=0.0,
+ mlp_drop=0.0,
+ post_mlp_drop=0.0,
+ drop_path=0.0,
+ act_layer=nn.GELU,
+ norm_layer=nn.LayerNorm,
+ layer_norm_first=True,
+ ffn_targets=False,
+ cosine_attention=False,
+ ):
+ super().__init__()
+
+ self.layer_norm_first = layer_norm_first
+ self.ffn_targets = ffn_targets
+
+ from timm.models.vision_transformer import DropPath, Mlp
+
+ self.norm1 = norm_layer(dim)
+ self.attn = AltAttention(
+ dim,
+ num_heads=num_heads,
+ qkv_bias=qkv_bias,
+ qk_scale=qk_scale,
+ attn_drop=attn_drop,
+ proj_drop=drop,
+ cosine_attention=cosine_attention,
+ )
+
+ self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
+ self.norm2 = norm_layer(dim)
+ mlp_hidden_dim = int(dim * mlp_ratio)
+ self.mlp = Mlp(
+ in_features=dim,
+ hidden_features=mlp_hidden_dim,
+ act_layer=act_layer,
+ drop=mlp_drop,
+ )
+ self.post_mlp_dropout = nn.Dropout(post_mlp_drop, inplace=False)
+
+ def forward(self, x, padding_mask=None, alibi_bias=None):
+ if self.layer_norm_first:
+ x = x + self.drop_path(self.attn(self.norm1(x), padding_mask, alibi_bias))
+ r = x = self.mlp(self.norm2(x))
+ t = x
+ x = r + self.drop_path(self.post_mlp_dropout(x))
+ if not self.ffn_targets:
+ t = x
+ else:
+ x = x + self.drop_path(self.attn(x, padding_mask, alibi_bias))
+ r = x = self.norm1(x)
+ x = self.mlp(x)
+ t = x
+ x = self.norm2(r + self.drop_path(self.post_mlp_dropout(x)))
+ if not self.ffn_targets:
+ t = x
+
+ return x, t
+
+
+class AltAttention(nn.Module):
+ def __init__(
+ self,
+ dim,
+ num_heads=8,
+ qkv_bias=False,
+ qk_scale=None,
+ attn_drop=0.0,
+ proj_drop=0.0,
+ cosine_attention=False,
+ ):
+ super().__init__()
+ self.num_heads = num_heads
+ head_dim = dim // num_heads
+ self.scale = qk_scale or head_dim ** -0.5
+
+ self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+ self.attn_drop = nn.Dropout(attn_drop)
+ self.proj = nn.Linear(dim, dim)
+ self.proj_drop = nn.Dropout(proj_drop)
+
+ self.cosine_attention = cosine_attention
+
+ if cosine_attention:
+ self.logit_scale = nn.Parameter(
+ torch.log(10 * torch.ones((num_heads, 1, 1))), requires_grad=True
+ )
+
+ def forward(self, x, padding_mask=None, alibi_bias=None):
+ B, N, C = x.shape
+ qkv = (
+ self.qkv(x)
+ .reshape(B, N, 3, self.num_heads, C // self.num_heads)
+ .permute(2, 0, 3, 1, 4) # qkv x B x H x L x D
+ )
+ q, k, v = (
+ qkv[0],
+ qkv[1],
+ qkv[2],
+ ) # make torchscript happy (cannot use tensor as tuple)
+
+ dtype = q.dtype
+
+ if self.cosine_attention:
+ # cosine attention
+ attn = F.normalize(q, dim=-1) @ F.normalize(k, dim=-1).transpose(-2, -1)
+ logit_scale = torch.clamp(
+ self.logit_scale, max=torch.log(torch.tensor(1.0 / 0.01))
+ ).exp()
+ attn = attn * logit_scale
+ else:
+ q = q * self.scale
+ attn = q @ k.transpose(-2, -1)
+
+ if alibi_bias is not None:
+ attn = attn.type_as(alibi_bias)
+ attn[:, : alibi_bias.size(1)] += alibi_bias
+
+ if padding_mask is not None and padding_mask.any():
+ attn = attn.masked_fill(
+ padding_mask.unsqueeze(1).unsqueeze(2).to(torch.bool),
+ float("-inf"),
+ )
+
+ attn = attn.softmax(dim=-1, dtype=torch.float32).to(dtype=dtype)
+ attn = self.attn_drop(attn)
+ x = (attn @ v).transpose(1, 2) #
+ x = x.reshape(B, N, C)
+ x = self.proj(x)
+ x = self.proj_drop(x)
+ return x
+
+
+class EncDecAttention(nn.Module):
+ def __init__(
+ self,
+ q_dim,
+ kv_dim,
+ num_heads=8,
+ qkv_bias=False,
+ qk_scale=None,
+ attn_drop=0.0,
+ proj_drop=0.0,
+ cosine_attention=False,
+ ):
+ super().__init__()
+ self.num_heads = num_heads
+ head_dim = q_dim // num_heads
+ self.scale = qk_scale or head_dim ** -0.5
+
+ self.q_proj = nn.Linear(q_dim, q_dim, bias=qkv_bias)
+ self.kv_proj = nn.Linear(kv_dim, 2 * q_dim, bias=qkv_bias)
+ self.attn_drop = nn.Dropout(attn_drop)
+ self.proj = nn.Linear(q_dim, q_dim)
+ self.proj_drop = nn.Dropout(proj_drop)
+
+ self.cosine_attention = cosine_attention
+
+ if cosine_attention:
+ self.logit_scale = nn.Parameter(
+ torch.log(10 * torch.ones((num_heads, 1, 1))), requires_grad=True
+ )
+
+ def forward(self, q, kv, padding_mask=None, alibi_bias=None):
+ B, N, C = q.shape
+
+ q = (
+ self.q_proj(q)
+ .reshape(B, N, self.num_heads, C // self.num_heads)
+ .permute(0, 2, 1, 3)
+ ) # B x H x L x D
+ kv = (
+ self.kv_proj(kv)
+ .reshape(B, -1, 2, self.num_heads, C // self.num_heads)
+ .permute(2, 0, 3, 1, 4)
+ ) # kv x B x H x L x D
+ k, v = (
+ kv[0],
+ kv[1],
+ ) # make torchscript happy (cannot use tensor as tuple)
+
+ dtype = q.dtype
+
+ if self.cosine_attention:
+ # cosine attention
+ attn = F.normalize(q, dim=-1) @ F.normalize(k, dim=-1).transpose(-2, -1)
+ logit_scale = torch.clamp(
+ self.logit_scale, max=torch.log(torch.tensor(1.0 / 0.01))
+ ).exp()
+ attn = attn * logit_scale
+ else:
+ q = q * self.scale
+ attn = q @ k.transpose(-2, -1)
+
+ if alibi_bias is not None:
+ attn = attn.type_as(alibi_bias)
+ attn[:, : alibi_bias.size(1)] += alibi_bias
+
+ if padding_mask is not None and padding_mask.any():
+ attn = attn.masked_fill(
+ padding_mask.unsqueeze(1).unsqueeze(2).to(torch.bool),
+ float("-inf"),
+ )
+
+ attn = attn.softmax(dim=-1, dtype=torch.float32).to(dtype=dtype)
+ attn = self.attn_drop(attn)
+ x = (attn @ v).transpose(1, 2) #
+ x = x.reshape(B, N, C)
+ x = self.proj(x)
+ x = self.proj_drop(x)
+ return x
+
+
+class EncDecBlock(nn.Module):
+ def __init__(
+ self,
+ q_dim,
+ kv_dim,
+ num_heads,
+ mlp_ratio=4.0,
+ qkv_bias=False,
+ qk_scale=None,
+ drop=0.0,
+ attn_drop=0.0,
+ mlp_drop=0.0,
+ post_mlp_drop=0.0,
+ drop_path=0.0,
+ act_layer=nn.GELU,
+ norm_layer=nn.LayerNorm,
+ layer_norm_first=True,
+ cosine_attention=False,
+ first_residual=True,
+ ):
+ super().__init__()
+
+ self.layer_norm_first = layer_norm_first
+
+ from timm.models.vision_transformer import DropPath, Mlp
+
+ self.norm1 = norm_layer(q_dim)
+ self.attn = EncDecAttention(
+ q_dim,
+ kv_dim,
+ num_heads=num_heads,
+ qkv_bias=qkv_bias,
+ qk_scale=qk_scale,
+ attn_drop=attn_drop,
+ proj_drop=drop,
+ cosine_attention=cosine_attention,
+ )
+
+ self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
+ self.norm2 = norm_layer(q_dim)
+ mlp_hidden_dim = int(q_dim * mlp_ratio)
+ self.mlp = Mlp(
+ in_features=q_dim,
+ hidden_features=mlp_hidden_dim,
+ act_layer=act_layer,
+ drop=mlp_drop,
+ )
+ self.post_mlp_dropout = nn.Dropout(post_mlp_drop, inplace=False)
+ self.first_residual = first_residual
+
+ def forward(self, q, kv, padding_mask=None, alibi_bias=None):
+ r = q if self.first_residual else 0
+ if self.layer_norm_first:
+ x = r + self.drop_path(
+ self.attn(self.norm1(q), kv, padding_mask, alibi_bias)
+ )
+ r = x = self.mlp(self.norm2(x))
+ x = r + self.drop_path(self.post_mlp_dropout(x))
+ else:
+ x = r + self.drop_path(self.attn(q, kv, padding_mask, alibi_bias))
+ r = x = self.norm1(x)
+ x = self.mlp(x)
+ x = self.norm2(r + self.drop_path(self.post_mlp_dropout(x)))
+
+ return x
+
+
+class EncDecTransformerDecoder(nn.Module):
+ def __init__(self, cfg: D2vDecoderConfig, input_dim):
+ super().__init__()
+
+ self.input_proj = nn.Linear(input_dim, cfg.decoder_dim)
+
+ self.blocks = nn.Sequential(
+ *[
+ EncDecBlock(
+ q_dim=cfg.decoder_dim,
+ kv_dim=input_dim,
+ num_heads=8,
+ mlp_ratio=4.0,
+ qkv_bias=True,
+ qk_scale=None,
+ drop=0.0,
+ attn_drop=0.0,
+ mlp_drop=0.0,
+ post_mlp_drop=0.0,
+ drop_path=0.0,
+ act_layer=nn.GELU,
+ norm_layer=nn.LayerNorm,
+ layer_norm_first=False,
+ cosine_attention=False,
+ first_residual=i > 0,
+ )
+ for i in range(cfg.decoder_layers)
+ ]
+ )
+
+ self.proj = nn.Linear(cfg.decoder_dim, input_dim)
+
+ def reset_parameters(self):
+ from fairseq.modules.transformer_sentence_encoder import init_bert_params
+
+ self.apply(init_bert_params)
+
+ def forward(self, x, kv):
+ x = self.input_proj(x)
+ for i, layer in enumerate(self.blocks):
+ x = layer(x, kv)
+
+ x = self.proj(x)
+ return x
diff --git a/fairseq/examples/data2vec/models/modalities/text.py b/fairseq/examples/data2vec/models/modalities/text.py
new file mode 100644
index 0000000000000000000000000000000000000000..adfac1ca481bd5b79862137a277549e801268ea1
--- /dev/null
+++ b/fairseq/examples/data2vec/models/modalities/text.py
@@ -0,0 +1,161 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import math
+from dataclasses import dataclass
+from functools import partial
+from typing import Callable, Dict, Optional
+
+import torch.nn as nn
+import torch.nn.functional as F
+import numpy as np
+from fairseq.modules import PositionalEmbedding, FairseqDropout, LayerNorm
+from fairseq.tasks import FairseqTask
+from .base import D2vModalityConfig, ModalitySpecificEncoder, get_alibi_bias
+from .modules import BlockEncoder, Decoder1d
+from examples.data2vec.data.modality import Modality
+
+
+@dataclass
+class D2vTextConfig(D2vModalityConfig):
+ type: Modality = Modality.TEXT
+ max_source_positions: int = 512
+ learned_pos: bool = True
+ dropout: float = 0.1 # used for both local_encoder and contextualized encoder. tied with global transformer in data2vec_text
+
+ no_scale_embedding: bool = True
+ layernorm_embedding: bool = True
+ no_token_positional_embeddings: bool = False
+
+
+class TextEncoder(ModalitySpecificEncoder):
+
+ modality_cfg: D2vTextConfig
+
+ def __init__(
+ self,
+ modality_cfg: D2vTextConfig,
+ embed_dim: int,
+ make_block: Callable[[float], nn.ModuleList],
+ norm_layer: Callable[[int], nn.LayerNorm],
+ layer_norm_first: bool,
+ alibi_biases: Dict,
+ task: Optional[FairseqTask],
+ ):
+ self.pad_idx = task.source_dictionary.pad()
+ self.vocab_size = len(task.source_dictionary)
+
+ local_encoder = TextLocalEncoder(
+ vocab_size=self.vocab_size,
+ embed_dim=embed_dim,
+ max_source_positions=modality_cfg.max_source_positions,
+ pad_idx=self.pad_idx,
+ no_scale_embedding=modality_cfg.no_scale_embedding,
+ layernorm_embedding=modality_cfg.layernorm_embedding,
+ dropout=modality_cfg.dropout,
+ no_token_positional_embeddings=modality_cfg.no_token_positional_embeddings,
+ learned_pos=modality_cfg.learned_pos,
+ )
+ dpr = np.linspace(
+ modality_cfg.start_drop_path_rate,
+ modality_cfg.end_drop_path_rate,
+ modality_cfg.prenet_depth,
+ )
+ context_encoder = BlockEncoder(
+ nn.ModuleList(make_block(dpr[i]) for i in range(modality_cfg.prenet_depth)),
+ norm_layer(embed_dim)
+ if not layer_norm_first and modality_cfg.prenet_depth > 0
+ else None,
+ layer_norm_first,
+ modality_cfg.prenet_layerdrop,
+ modality_cfg.prenet_dropout if modality_cfg.prenet_depth > 0 else 0.0,
+ )
+ decoder = (
+ Decoder1d(modality_cfg.decoder, embed_dim)
+ if modality_cfg.decoder is not None
+ else None
+ )
+
+ alibi_bias_fn = partial(get_alibi_bias, alibi_biases=alibi_biases)
+
+ super().__init__(
+ modality_cfg=modality_cfg,
+ embed_dim=embed_dim,
+ local_encoder=local_encoder,
+ project_features=nn.Identity(),
+ fixed_positional_encoder=None,
+ relative_positional_encoder=None,
+ context_encoder=context_encoder,
+ decoder=decoder,
+ get_alibi_bias=alibi_bias_fn,
+ )
+
+ def reset_parameters(self):
+ super().reset_parameters()
+
+ def convert_padding_mask(self, x, padding_mask):
+ if padding_mask is None or padding_mask.size(1) == x.size(1):
+ return padding_mask
+
+ diff = self.downsample - padding_mask.size(1) % self.downsample
+ if 0 < diff < self.downsample:
+ padding_mask = F.pad(padding_mask, (0, diff), value=True)
+
+ padding_mask = padding_mask.view(padding_mask.size(0), -1, self.downsample)
+ padding_mask = padding_mask.all(-1)
+ if padding_mask.size(1) > x.size(1):
+ padding_mask = padding_mask[:, : x.size(1)]
+
+ assert x.size(1) == padding_mask.size(
+ 1
+ ), f"{x.size(1), padding_mask.size(1), diff, self.downsample}"
+
+ return padding_mask
+
+
+class TextLocalEncoder(nn.Module):
+ def __init__(
+ self,
+ vocab_size,
+ embed_dim,
+ max_source_positions,
+ pad_idx,
+ no_scale_embedding,
+ layernorm_embedding,
+ dropout,
+ no_token_positional_embeddings,
+ learned_pos,
+ ):
+ super().__init__()
+ self.pad_idx = pad_idx
+ self.dropout_module = FairseqDropout(dropout)
+
+ self.embed_tokens = nn.Embedding(vocab_size, embed_dim, pad_idx)
+ self.embed_scale = 1.0 if no_scale_embedding else math.sqrt(embed_dim)
+ self.embed_positions = (
+ PositionalEmbedding(
+ max_source_positions,
+ embed_dim,
+ pad_idx,
+ learned=learned_pos,
+ )
+ if not no_token_positional_embeddings
+ else None
+ )
+ self.embed_scale = 1.0 if no_scale_embedding else math.sqrt(embed_dim)
+
+ self.layernorm_embedding = None
+ if layernorm_embedding:
+ self.layernorm_embedding = LayerNorm(embed_dim)
+
+ def forward(self, src_tokens):
+ x = self.embed_scale * self.embed_tokens(src_tokens)
+ if self.embed_positions is not None:
+ x = x + self.embed_positions(src_tokens)
+
+ if self.layernorm_embedding is not None:
+ x = self.layernorm_embedding(x)
+ x = self.dropout_module(x)
+ return x
diff --git a/fairseq/examples/data2vec/models/utils.py b/fairseq/examples/data2vec/models/utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..0e2f240d4f27ac1b13403b38097ba58aaea57eb0
--- /dev/null
+++ b/fairseq/examples/data2vec/models/utils.py
@@ -0,0 +1,55 @@
+import math
+import torch
+
+def get_alibi(
+ max_positions: int,
+ attention_heads: int,
+):
+ def get_slopes(n):
+ def get_slopes_power_of_2(n):
+ start = 2 ** (-(2 ** -(math.log2(n) - 3)))
+ ratio = start
+ return [start * ratio ** i for i in range(n)]
+
+ # In the paper, we only train models that have 2^a heads for some
+ # a. This function has some good properties that only occur when
+ # the input is a power of 2. To maintain that even when the number
+ # of heads is not a power of 2, we use this workaround.
+ if math.log2(n).is_integer():
+ return get_slopes_power_of_2(n)
+ else:
+ closest_power_of_2 = 2 ** math.floor(math.log2(n))
+ return (
+ get_slopes_power_of_2(closest_power_of_2)
+ + get_slopes(2 * closest_power_of_2)[0::2][: n - closest_power_of_2]
+ )
+
+ maxpos = max_positions
+ attn_heads = attention_heads
+ slopes = torch.Tensor(get_slopes(attn_heads))
+ # prepare alibi position linear bias. Note that wav2vec2 is non
+ # autoregressive model so we want a symmetric mask with 0 on the
+ # diagonal and other wise linear decreasing valuees
+ pos_bias = (
+ torch.abs(
+ torch.arange(maxpos).unsqueeze(0) - torch.arange(maxpos).unsqueeze(1)
+ )
+ * -1
+ )
+ alibi_bias = slopes.unsqueeze(1).unsqueeze(1) * pos_bias.unsqueeze(0).expand(
+ attn_heads, -1, -1
+ )
+ return alibi_bias
+
+def masked_alibi(alibi_bias, mask_indices, orig_B, orig_T):
+ alibi_bias = alibi_bias.view(orig_B, -1, orig_T, orig_T)
+ H = alibi_bias.size(1)
+ alibi_mask = mask_indices.unsqueeze(1)
+ alibi_bias = alibi_bias.masked_select(alibi_mask.unsqueeze(-1))
+ alibi_bias = alibi_bias.view(orig_B, H, -1, orig_T)
+ M = alibi_bias.size(-2)
+ alibi_bias = alibi_bias.masked_select(alibi_mask.unsqueeze(-2))
+ alibi_bias = alibi_bias.view(-1, M, M)
+ return alibi_bias
+
+
diff --git a/fairseq/examples/data2vec/scripts/convert_audioset_labels.py b/fairseq/examples/data2vec/scripts/convert_audioset_labels.py
new file mode 100644
index 0000000000000000000000000000000000000000..7d720e606a4e2f12e15608bcee1ca2f551070fb3
--- /dev/null
+++ b/fairseq/examples/data2vec/scripts/convert_audioset_labels.py
@@ -0,0 +1,63 @@
+#!/usr/bin/env python3
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import argparse
+import os
+
+
+def get_parser():
+ parser = argparse.ArgumentParser(description="convert audioset labels")
+ # fmt: off
+ parser.add_argument('in_file', help='audioset csv file to convert')
+ parser.add_argument('--manifest', required=True, metavar='PATH', help='wav2vec-like manifest')
+ parser.add_argument('--descriptors', required=True, metavar='PATH', help='path to label descriptor file')
+ parser.add_argument('--output', required=True, metavar='PATH', help='where to output converted labels')
+ # fmt: on
+
+ return parser
+
+
+def main():
+ parser = get_parser()
+ args = parser.parse_args()
+
+ label_descriptors = {}
+ with open(args.descriptors, "r") as ldf:
+ next(ldf)
+ for line in ldf:
+ if line.strip() == "":
+ continue
+
+ items = line.split(",")
+ assert len(items) > 2, line
+ idx = items[0]
+ lbl = items[1]
+ assert lbl not in label_descriptors, lbl
+ label_descriptors[lbl] = idx
+
+ labels = {}
+ with open(args.in_file, "r") as ifd:
+ for line in ifd:
+ if line.lstrip().startswith("#"):
+ continue
+ items = line.rstrip().split(",")
+ id = items[0].strip()
+ start = items[1].strip()
+ end = items[2].strip()
+ lbls = [label_descriptors[it.strip(' "')] for it in items[3:]]
+ labels[id] = [start, end, ",".join(lbls)]
+
+ with open(args.manifest, "r") as mf, open(args.output, "w") as of:
+ next(mf)
+ for line in mf:
+ path, _ = line.split("\t")
+ id = os.path.splitext(os.path.basename(path))[0]
+ lbl = labels[id]
+ print("\t".join(lbl), file=of)
+
+
+if __name__ == "__main__":
+ main()
diff --git a/fairseq/examples/data2vec/scripts/multi/finetune_all_fair_aws_local_lr.sh b/fairseq/examples/data2vec/scripts/multi/finetune_all_fair_aws_local_lr.sh
new file mode 100644
index 0000000000000000000000000000000000000000..41bcd31fc5cfdf4c7a41b10f51e9b055e1e8ece4
--- /dev/null
+++ b/fairseq/examples/data2vec/scripts/multi/finetune_all_fair_aws_local_lr.sh
@@ -0,0 +1,18 @@
+#!/bin/bash
+
+set -eu
+
+job_id="$1"
+task_id="$2"
+dir="$3"
+
+echo "job_id: $job_id, task_id: $task_id, dir: $dir"
+
+mkdir -p "$dir/log"
+sbatch_args="-p wav2vec --nodes=1 --ntasks-per-node=1"
+sbatch_args="$sbatch_args --gpus-per-node=1 --cpus-per-task=8 --mem=0 --time=24:00:00"
+sbatch_args="$sbatch_args -d afterok:$job_id -o $dir/log/decode_sweep_%A.out"
+sbatch_args="$sbatch_args -e $dir/log/decode_sweep_%A.err"
+
+sbatch $sbatch_args examples/data2vec/scripts/multi/finetune_all_fair_local_lr.sh $dir
+
diff --git a/fairseq/examples/data2vec/scripts/multi/finetune_all_fair_aws_local_lr_nodep.sh b/fairseq/examples/data2vec/scripts/multi/finetune_all_fair_aws_local_lr_nodep.sh
new file mode 100644
index 0000000000000000000000000000000000000000..fc85908b7265612b03afe0c1d3423b2fee4a7ebc
--- /dev/null
+++ b/fairseq/examples/data2vec/scripts/multi/finetune_all_fair_aws_local_lr_nodep.sh
@@ -0,0 +1,16 @@
+#!/bin/bash
+
+set -eu
+
+dir="$1"
+
+echo "dir: $dir"
+
+mkdir -p "$dir/log"
+sbatch_args="-p wav2vec --nodes=1 --ntasks-per-node=1"
+sbatch_args="$sbatch_args --gpus-per-node=1 --cpus-per-task=8 --mem=0 --time=24:00:00"
+sbatch_args="$sbatch_args -o $dir/log/decode_sweep_%A.out"
+sbatch_args="$sbatch_args -e $dir/log/decode_sweep_%A.err"
+
+sbatch $sbatch_args examples/data2vec/scripts/multi/finetune_all_fair_local_lr.sh $dir
+
diff --git a/fairseq/examples/data2vec/scripts/multi/finetune_all_fair_local_lr.sh b/fairseq/examples/data2vec/scripts/multi/finetune_all_fair_local_lr.sh
new file mode 100644
index 0000000000000000000000000000000000000000..121226972b73c50875ac2575740ad20f2856c66d
--- /dev/null
+++ b/fairseq/examples/data2vec/scripts/multi/finetune_all_fair_local_lr.sh
@@ -0,0 +1,28 @@
+#!/usr/bin/env zsh
+
+dir="$1"
+cp="$dir/checkpoints/checkpoint_last.pt"
+
+echo "dir: $dir"
+
+declare -A tasks
+tasks[cola]="/fsx-wav2vec/abaevski/data/nlp/GLUE/CoLA-bin"
+tasks[qnli]="/fsx-wav2vec/abaevski/data/nlp/GLUE/QNLI-bin"
+tasks[mrpc]="/fsx-wav2vec/abaevski/data/nlp/GLUE/MRPC-bin"
+tasks[rte]="/fsx-wav2vec/abaevski/data/nlp/GLUE/RTE-bin"
+tasks[sst_2]="/fsx-wav2vec/abaevski/data/nlp/GLUE/SST-2-bin"
+tasks[mnli]="/fsx-wav2vec/abaevski/data/nlp/GLUE/MNLI-bin"
+tasks[qqp]="/fsx-wav2vec/abaevski/data/nlp/GLUE/QQP-bin"
+tasks[sts_b]="/fsx-wav2vec/abaevski/data/nlp/GLUE/STS-B-bin"
+
+lrs=(5e-6 8e-6 1e-5 2e-5)
+
+for task data_path in ${(kv)tasks}; do
+ for lr in $lrs; do
+ echo $lr $task
+ PYTHONPATH=. PREFIX="${PREFIX}" SUFFIX="" \
+ python fairseq_cli/hydra_train.py -m --config-dir examples/data2vec/config/multi/text_finetuning \
+ --config-name $task +run_config=local task.data="$data_path" common.log_interval=200 dataset.num_workers=1 \
+ model.model_path="$cp" hydra.sweep.dir="$dir/finetune_lr/$task/$lr" "optimization.lr=[${lr}]" +model=text_wrap
+ done
+done
diff --git a/fairseq/examples/data2vec/scripts/text/finetune_all_char_fair_aws_local_lr.sh b/fairseq/examples/data2vec/scripts/text/finetune_all_char_fair_aws_local_lr.sh
new file mode 100644
index 0000000000000000000000000000000000000000..18b862c2408ae95c7c0c53401fe2637d52fb9a52
--- /dev/null
+++ b/fairseq/examples/data2vec/scripts/text/finetune_all_char_fair_aws_local_lr.sh
@@ -0,0 +1,17 @@
+#!/bin/bash
+
+set -eu
+
+job_id="$1"
+task_id="$2"
+dir="$3"
+
+echo "job_id: $job_id, task_id: $task_id, dir: $dir"
+
+mkdir -p "$dir/log"
+sbatch_args="-p wav2vec --nodes=1 --ntasks-per-node=1"
+sbatch_args="$sbatch_args --gpus-per-node=1 --cpus-per-task=8 --mem=0 --time=24:00:00"
+sbatch_args="$sbatch_args -d afterok:$job_id -o $dir/log/ft_%A.out"
+sbatch_args="$sbatch_args -e $dir/log/ft_%A.err"
+
+sbatch $sbatch_args examples/data2vec/scripts/text/finetune_all_char_fair_local_lr.sh $dir
diff --git a/fairseq/examples/data2vec/scripts/text/finetune_all_fair.sh b/fairseq/examples/data2vec/scripts/text/finetune_all_fair.sh
new file mode 100644
index 0000000000000000000000000000000000000000..34a2df399025d63fb1593a668d74ab9a8508d2cb
--- /dev/null
+++ b/fairseq/examples/data2vec/scripts/text/finetune_all_fair.sh
@@ -0,0 +1,21 @@
+#!/usr/bin/env zsh
+
+job_id=$1
+task_id=$2
+dir="$3"
+cp="$dir/$task_id/checkpoints/checkpoint_last.pt"
+
+echo "job_id: $job_id, task_id: $task_id, dir: $dir"
+
+declare -A tasks
+tasks[cola]="/private/home/jgu/data/GLUE/CoLA-bin"
+tasks[qnli]="/private/home/jgu/data/GLUE/QNLI-bin"
+tasks[mrpc]="/private/home/jgu/data/GLUE/MRPC-bin"
+tasks[rte]="/private/home/jgu/data/GLUE/RTE-bin"
+tasks[sst_2]="/private/home/jgu/data/GLUE/SST-2-bin"
+
+for task data_path in ${(kv)tasks}; do
+ PYTHONPATH=. PREFIX="${PREFIX}" SUFFIX="" nohup python fairseq_cli/hydra_train.py -m --config-dir examples/roberta/config/finetuning \
+ --config-name $task hydra/launcher=submitit_slurm +run_config=slurm_1g task.data="$data_path" hydra.launcher.name=finetune_${task}_${PREFIX} \
+ checkpoint.restore_file="$cp" +hydra.launcher.additional_parameters.dependency="afterok:$job_id" hydra.sweep.dir="$dir/finetune/$task" &
+done
diff --git a/fairseq/examples/data2vec/scripts/text/finetune_all_fair_aws.sh b/fairseq/examples/data2vec/scripts/text/finetune_all_fair_aws.sh
new file mode 100644
index 0000000000000000000000000000000000000000..b417c20024d296bcc5828a2213f25ab19261d771
--- /dev/null
+++ b/fairseq/examples/data2vec/scripts/text/finetune_all_fair_aws.sh
@@ -0,0 +1,21 @@
+#!/usr/bin/env zsh
+
+job_id=$1
+task_id=$2
+dir="$3"
+cp="$dir/checkpoints/checkpoint_last.pt"
+
+echo "job_id: $job_id, task_id: $task_id, dir: $dir"
+
+declare -A tasks
+tasks[cola]="/fsx-wav2vec/abaevski/data/nlp/GLUE/CoLA-bin"
+tasks[qnli]="/fsx-wav2vec/abaevski/data/nlp/GLUE/QNLI-bin"
+tasks[mrpc]="/fsx-wav2vec/abaevski/data/nlp/GLUE/MRPC-bin"
+tasks[rte]="/fsx-wav2vec/abaevski/data/nlp/GLUE/RTE-bin"
+tasks[sst_2]="/fsx-wav2vec/abaevski/data/nlp/GLUE/SST-2-bin"
+
+for task data_path in ${(kv)tasks}; do
+ PYTHONPATH=. PREFIX="${PREFIX}" SUFFIX="" nohup python fairseq_cli/hydra_train.py -m --config-dir examples/roberta/config/finetuning \
+ --config-name $task hydra/launcher=submitit_slurm +run_config=slurm_1g_aws task.data="$data_path" hydra.launcher.name=finetune_${task}_${PREFIX} \
+ checkpoint.restore_file="$cp" +hydra.launcher.additional_parameters.dependency="afterok:$job_id" hydra.sweep.dir="$dir/finetune/$task" &
+done
diff --git a/fairseq/examples/data2vec/scripts/text/finetune_all_fair_aws_local_lr.sh b/fairseq/examples/data2vec/scripts/text/finetune_all_fair_aws_local_lr.sh
new file mode 100644
index 0000000000000000000000000000000000000000..64dbcb111e9709956798acbe9d006640e656849b
--- /dev/null
+++ b/fairseq/examples/data2vec/scripts/text/finetune_all_fair_aws_local_lr.sh
@@ -0,0 +1,17 @@
+#!/bin/bash
+
+set -eu
+
+job_id="$1"
+task_id="$2"
+dir="$3"
+
+echo "job_id: $job_id, task_id: $task_id, dir: $dir"
+
+mkdir -p "$dir/log"
+sbatch_args="-p wav2vec --nodes=1 --ntasks-per-node=1"
+sbatch_args="$sbatch_args --gpus-per-node=1 --cpus-per-task=8 --mem=0 --time=24:00:00"
+sbatch_args="$sbatch_args -d afterok:$job_id -o $dir/log/decode_sweep_%A.out"
+sbatch_args="$sbatch_args -e $dir/log/decode_sweep_%A.err"
+
+sbatch $sbatch_args examples/data2vec/scripts/text/finetune_all_fair_local_lr.sh $dir
diff --git a/fairseq/examples/data2vec/scripts/text/finetune_all_fair_aws_lr.sh b/fairseq/examples/data2vec/scripts/text/finetune_all_fair_aws_lr.sh
new file mode 100644
index 0000000000000000000000000000000000000000..d75c5495731cc583d481c2521777cd3e5e74e3ad
--- /dev/null
+++ b/fairseq/examples/data2vec/scripts/text/finetune_all_fair_aws_lr.sh
@@ -0,0 +1,23 @@
+#!/usr/bin/env zsh
+
+job_id=$1
+task_id=$2
+dir="$3"
+cp="$dir/checkpoints/checkpoint_last.pt"
+
+echo "job_id: $job_id, task_id: $task_id, dir: $dir"
+
+declare -A tasks
+tasks[cola]="/fsx-wav2vec/abaevski/data/nlp/GLUE/CoLA-bin"
+tasks[qnli]="/fsx-wav2vec/abaevski/data/nlp/GLUE/QNLI-bin"
+tasks[mrpc]="/fsx-wav2vec/abaevski/data/nlp/GLUE/MRPC-bin"
+tasks[rte]="/fsx-wav2vec/abaevski/data/nlp/GLUE/RTE-bin"
+tasks[sst_2]="/fsx-wav2vec/abaevski/data/nlp/GLUE/SST-2-bin"
+
+for task data_path in ${(kv)tasks}; do
+ for lr in 5e-6 8e-6 1e-5 2e-5 5e-5 8e-5 1e-4 2e-4; do
+ PYTHONPATH=. PREFIX="${PREFIX}" SUFFIX="" nohup python fairseq_cli/hydra_train.py -m --config-dir examples/roberta/config/finetuning \
+ --config-name $task hydra/launcher=submitit_slurm +run_config=slurm_1g_aws task.data="$data_path" hydra.launcher.name=finetune_${task}_${PREFIX} \
+ checkpoint.restore_file="$cp" +hydra.launcher.additional_parameters.dependency="afterok:$job_id" hydra.sweep.dir="$dir/finetune_lr/$task/$lr" "optimization.lr=[${lr}]" &
+ done
+done
diff --git a/fairseq/examples/data2vec/scripts/text/finetune_all_fair_local_lr.sh b/fairseq/examples/data2vec/scripts/text/finetune_all_fair_local_lr.sh
new file mode 100644
index 0000000000000000000000000000000000000000..8be98c08475bd570f01065c892b94dd1c2cb1636
--- /dev/null
+++ b/fairseq/examples/data2vec/scripts/text/finetune_all_fair_local_lr.sh
@@ -0,0 +1,25 @@
+#!/usr/bin/env zsh
+
+dir="$1"
+cp="$dir/checkpoints/checkpoint_last.pt"
+
+echo "dir: $dir"
+
+declare -A tasks
+tasks[cola]="/fsx-wav2vec/abaevski/data/nlp/GLUE/CoLA-bin"
+tasks[qnli]="/fsx-wav2vec/abaevski/data/nlp/GLUE/QNLI-bin"
+tasks[mrpc]="/fsx-wav2vec/abaevski/data/nlp/GLUE/MRPC-bin"
+tasks[rte]="/fsx-wav2vec/abaevski/data/nlp/GLUE/RTE-bin"
+tasks[sst_2]="/fsx-wav2vec/abaevski/data/nlp/GLUE/SST-2-bin"
+
+lrs=(5e-6 8e-6 1e-5 2e-5)
+
+for task data_path in ${(kv)tasks}; do
+ for lr in $lrs; do
+ echo $lr $task
+ PYTHONPATH=. PREFIX="${PREFIX}" SUFFIX="" \
+ python fairseq_cli/hydra_train.py -m --config-dir examples/roberta/config/finetuning \
+ --config-name $task +run_config=local task.data="$data_path" common.log_interval=200 dataset.num_workers=1 \
+ checkpoint.restore_file="$cp" hydra.sweep.dir="$dir/finetune_lr/$task/$lr" "optimization.lr=[${lr}]"
+ done
+done
diff --git a/fairseq/examples/data2vec/scripts/text/finetune_all_fair_nodep.sh b/fairseq/examples/data2vec/scripts/text/finetune_all_fair_nodep.sh
new file mode 100644
index 0000000000000000000000000000000000000000..d02bcc0f7588a35613055b1b13f10e53dc30a188
--- /dev/null
+++ b/fairseq/examples/data2vec/scripts/text/finetune_all_fair_nodep.sh
@@ -0,0 +1,19 @@
+#!/usr/bin/env zsh
+
+dir="$1"
+cp="$dir/checkpoints/checkpoint_last.pt"
+
+echo "dir: $dir"
+
+declare -A tasks
+tasks[cola]="/private/home/jgu/data/GLUE/CoLA-bin"
+tasks[qnli]="/private/home/jgu/data/GLUE/QNLI-bin"
+tasks[mrpc]="/private/home/jgu/data/GLUE/MRPC-bin"
+tasks[rte]="/private/home/jgu/data/GLUE/RTE-bin"
+tasks[sst_2]="/private/home/jgu/data/GLUE/SST-2-bin"
+
+for task data_path in ${(kv)tasks}; do
+ PYTHONPATH=. PREFIX="${PREFIX}" SUFFIX="" nohup python fairseq_cli/hydra_train.py -m --config-dir examples/roberta/config/finetuning \
+ --config-name $task hydra/launcher=submitit_slurm +run_config=slurm_1g task.data="$data_path" hydra.launcher.name=finetune_${task}_${PREFIX} \
+ checkpoint.restore_file="$cp" hydra.sweep.dir="$dir/finetune/$task" &
+done
diff --git a/fairseq/examples/data2vec/scripts/text/finetune_all_fair_nodep_aws.sh b/fairseq/examples/data2vec/scripts/text/finetune_all_fair_nodep_aws.sh
new file mode 100644
index 0000000000000000000000000000000000000000..75538354e11b6ccb8bb1922e05fad790a9e4049a
--- /dev/null
+++ b/fairseq/examples/data2vec/scripts/text/finetune_all_fair_nodep_aws.sh
@@ -0,0 +1,19 @@
+#!/usr/bin/env zsh
+
+dir="$1"
+cp="$dir/checkpoints/checkpoint_last.pt"
+
+echo "dir: $dir"
+
+declare -A tasks
+tasks[cola]="/fsx-wav2vec/abaevski/data/nlp/GLUE/CoLA-bin"
+tasks[qnli]="/fsx-wav2vec/abaevski/data/nlp/GLUE/QNLI-bin"
+tasks[mrpc]="/fsx-wav2vec/abaevski/data/nlp/GLUE/MRPC-bin"
+tasks[rte]="/fsx-wav2vec/abaevski/data/nlp/GLUE/RTE-bin"
+tasks[sst_2]="/fsx-wav2vec/abaevski/data/nlp/GLUE/SST-2-bin"
+
+for task data_path in ${(kv)tasks}; do
+ PYTHONPATH=. PREFIX="${PREFIX}" SUFFIX="" nohup python fairseq_cli/hydra_train.py -m --config-dir examples/roberta/config/finetuning \
+ --config-name $task hydra/launcher=submitit_slurm +run_config=slurm_1g_aws task.data="$data_path" hydra.launcher.name=finetune_${task}_${PREFIX} \
+ checkpoint.restore_file="$cp" hydra.sweep.dir="$dir/finetune/$task" &
+done
diff --git a/fairseq/examples/data2vec/scripts/text/finetune_all_fair_nodep_aws_local_lr.sh b/fairseq/examples/data2vec/scripts/text/finetune_all_fair_nodep_aws_local_lr.sh
new file mode 100644
index 0000000000000000000000000000000000000000..16c1358b2ffa0194c5a8884ec7236badba2d99f5
--- /dev/null
+++ b/fairseq/examples/data2vec/scripts/text/finetune_all_fair_nodep_aws_local_lr.sh
@@ -0,0 +1,15 @@
+#!/bin/bash
+
+set -eu
+
+dir="$1"
+
+echo "dir: $dir"
+
+mkdir -p "$dir/log"
+sbatch_args="-p wav2vec --nodes=1 --ntasks-per-node=1"
+sbatch_args="$sbatch_args --gpus-per-node=1 --cpus-per-task=8 --mem=0 --time=24:00:00"
+sbatch_args="$sbatch_args -o $dir/log/decode_sweep_%A.out"
+sbatch_args="$sbatch_args -e $dir/log/decode_sweep_%A.err"
+
+sbatch $sbatch_args examples/data2vec/scripts/text/finetune_all_fair_local_lr.sh $dir
diff --git a/fairseq/examples/data2vec/scripts/text/finetune_all_fair_nodep_aws_lr.sh b/fairseq/examples/data2vec/scripts/text/finetune_all_fair_nodep_aws_lr.sh
new file mode 100644
index 0000000000000000000000000000000000000000..fb5ddbe22c2034011c11432bd0daa1b9c31df512
--- /dev/null
+++ b/fairseq/examples/data2vec/scripts/text/finetune_all_fair_nodep_aws_lr.sh
@@ -0,0 +1,21 @@
+#!/usr/bin/env zsh
+
+dir="$1"
+cp="$dir/checkpoints/checkpoint_last.pt"
+
+echo "dir: $dir"
+
+declare -A tasks
+tasks[cola]="/fsx-wav2vec/abaevski/data/nlp/GLUE/CoLA-bin"
+tasks[qnli]="/fsx-wav2vec/abaevski/data/nlp/GLUE/QNLI-bin"
+tasks[mrpc]="/fsx-wav2vec/abaevski/data/nlp/GLUE/MRPC-bin"
+tasks[rte]="/fsx-wav2vec/abaevski/data/nlp/GLUE/RTE-bin"
+tasks[sst_2]="/fsx-wav2vec/abaevski/data/nlp/GLUE/SST-2-bin"
+
+for task data_path in ${(kv)tasks}; do
+ for lr in 5e-6 8e-6 1e-5 2e-5 5e-5 8e-5 1e-4 2e-4; do
+ PYTHONPATH=. PREFIX="${PREFIX}" SUFFIX="" nohup python fairseq_cli/hydra_train.py -m --config-dir examples/roberta/config/finetuning \
+ --config-name $task hydra/launcher=submitit_slurm +run_config=slurm_1g_aws task.data="$data_path" hydra.launcher.name=finetune_${task}_${PREFIX} \
+ checkpoint.restore_file="$cp" hydra.sweep.dir="$dir/finetune_lr/$task/$lr" "optimization.lr=[${lr}]" &
+ done
+done
diff --git a/fairseq/examples/data2vec/scripts/text/finetune_all_fair_nodep_aws_lr_nopos.sh b/fairseq/examples/data2vec/scripts/text/finetune_all_fair_nodep_aws_lr_nopos.sh
new file mode 100644
index 0000000000000000000000000000000000000000..1ffab1c85064f8719854efd2cc91cbd532995c12
--- /dev/null
+++ b/fairseq/examples/data2vec/scripts/text/finetune_all_fair_nodep_aws_lr_nopos.sh
@@ -0,0 +1,21 @@
+#!/usr/bin/env zsh
+
+dir="$1"
+cp="$dir/checkpoints/checkpoint_last.pt"
+
+echo "dir: $dir"
+
+declare -A tasks
+tasks[cola]="/fsx-wav2vec/abaevski/data/nlp/GLUE/CoLA-bin"
+tasks[qnli]="/fsx-wav2vec/abaevski/data/nlp/GLUE/QNLI-bin"
+tasks[mrpc]="/fsx-wav2vec/abaevski/data/nlp/GLUE/MRPC-bin"
+tasks[rte]="/fsx-wav2vec/abaevski/data/nlp/GLUE/RTE-bin"
+tasks[sst_2]="/fsx-wav2vec/abaevski/data/nlp/GLUE/SST-2-bin"
+
+for task data_path in ${(kv)tasks}; do
+ for lr in 5e-6 8e-6 1e-5 2e-5 5e-5 8e-5 1e-4 2e-4; do
+ PYTHONPATH=. PREFIX="${PREFIX}" SUFFIX="" nohup python fairseq_cli/hydra_train.py -m --config-dir examples/roberta/config/finetuning \
+ --config-name $task hydra/launcher=submitit_slurm +run_config=slurm_1g_aws task.data="$data_path" hydra.launcher.name=finetune_${task}_${PREFIX} \
+ checkpoint.restore_file="$cp" hydra.sweep.dir="$dir/finetune_lr/$task/$lr" "optimization.lr=[${lr}]" +model.encoder_learned_pos=False &
+ done
+done
diff --git a/fairseq/examples/data2vec/scripts/text/finetune_all_large_fair_aws_local_lr.sh b/fairseq/examples/data2vec/scripts/text/finetune_all_large_fair_aws_local_lr.sh
new file mode 100644
index 0000000000000000000000000000000000000000..c3c58adcb81b53fc30765d7506b92935fc1978bc
--- /dev/null
+++ b/fairseq/examples/data2vec/scripts/text/finetune_all_large_fair_aws_local_lr.sh
@@ -0,0 +1,17 @@
+#!/bin/bash
+
+set -eu
+
+job_id="$1"
+task_id="$2"
+dir="$3"
+
+echo "job_id: $job_id, task_id: $task_id, dir: $dir"
+
+mkdir -p "$dir/log"
+sbatch_args="-p wav2vec --nodes=1 --ntasks-per-node=1"
+sbatch_args="$sbatch_args --gpus-per-node=1 --cpus-per-task=8 --mem=0 --time=24:00:00"
+sbatch_args="$sbatch_args -d afterok:$job_id -o $dir/log/decode_sweep_%A.out"
+sbatch_args="$sbatch_args -e $dir/log/decode_sweep_%A.err"
+
+sbatch $sbatch_args examples/data2vec/scripts/text/finetune_all_large_fair_local_lr.sh $dir
diff --git a/fairseq/examples/data2vec/scripts/text/finetune_all_large_fair_local_lr.sh b/fairseq/examples/data2vec/scripts/text/finetune_all_large_fair_local_lr.sh
new file mode 100644
index 0000000000000000000000000000000000000000..5efb00e0df355404c760c8029b0639ae1476d7db
--- /dev/null
+++ b/fairseq/examples/data2vec/scripts/text/finetune_all_large_fair_local_lr.sh
@@ -0,0 +1,26 @@
+#!/usr/bin/env zsh
+
+dir="$1"
+cp="$dir/checkpoints/checkpoint_last.pt"
+
+echo "dir: $dir"
+
+declare -A tasks
+tasks[cola]="/fsx-wav2vec/abaevski/data/nlp/GLUE/CoLA-bin"
+tasks[qnli]="/fsx-wav2vec/abaevski/data/nlp/GLUE/QNLI-bin"
+tasks[mrpc]="/fsx-wav2vec/abaevski/data/nlp/GLUE/MRPC-bin"
+tasks[rte]="/fsx-wav2vec/abaevski/data/nlp/GLUE/RTE-bin"
+tasks[sst_2]="/fsx-wav2vec/abaevski/data/nlp/GLUE/SST-2-bin"
+
+lrs=(5e-6 8e-6 1e-5 2e-5)
+
+for task data_path in ${(kv)tasks}; do
+ for lr in $lrs; do
+ echo $lr $task
+ PYTHONPATH=. PREFIX="${PREFIX}" SUFFIX="" \
+ python fairseq_cli/hydra_train.py -m --config-dir examples/roberta/config/finetuning \
+ --config-name $task +run_config=local task.data="$data_path" common.log_interval=200 dataset.num_workers=1 \
+ checkpoint.restore_file="$cp" hydra.sweep.dir="$dir/finetune_lr/$task/$lr" "optimization.lr=[${lr}]" \
+ model._name=roberta_large
+ done
+done
diff --git a/fairseq/examples/data2vec/scripts/text/finetune_all_large_fair_nodep_aws_local_lr.sh b/fairseq/examples/data2vec/scripts/text/finetune_all_large_fair_nodep_aws_local_lr.sh
new file mode 100644
index 0000000000000000000000000000000000000000..4fb21bce79cab2f2f238178beb38c85daa901296
--- /dev/null
+++ b/fairseq/examples/data2vec/scripts/text/finetune_all_large_fair_nodep_aws_local_lr.sh
@@ -0,0 +1,15 @@
+#!/bin/bash
+
+set -eu
+
+dir="$1"
+
+echo "dir: $dir"
+
+mkdir -p "$dir/log"
+sbatch_args="-p wav2vec --nodes=1 --ntasks-per-node=1"
+sbatch_args="$sbatch_args --gpus-per-node=1 --cpus-per-task=8 --mem=0 --time=24:00:00"
+sbatch_args="$sbatch_args -o $dir/log/decode_sweep_%A.out"
+sbatch_args="$sbatch_args -e $dir/log/decode_sweep_%A.err"
+
+sbatch $sbatch_args examples/data2vec/scripts/text/finetune_all_large_fair_local_lr.sh $dir
diff --git a/fairseq/examples/data2vec/scripts/text/finetune_sst2_qnli_sweep_fair_nodep.sh b/fairseq/examples/data2vec/scripts/text/finetune_sst2_qnli_sweep_fair_nodep.sh
new file mode 100644
index 0000000000000000000000000000000000000000..d7b43bee80b906619446c773b6473fb2e4673626
--- /dev/null
+++ b/fairseq/examples/data2vec/scripts/text/finetune_sst2_qnli_sweep_fair_nodep.sh
@@ -0,0 +1,20 @@
+#!/usr/bin/env zsh
+
+dir="$1"
+cp="$dir/checkpoints/checkpoint_last.pt"
+
+echo "dir: $dir"
+
+declare -A tasks
+tasks[qnli]="/private/home/jgu/data/GLUE/QNLI-bin"
+tasks[sst_2]="/private/home/jgu/data/GLUE/SST-2-bin"
+
+lrs="5e-6 1e-5 2e-5 5e-5 1e-4 2e-4 5e-4 1e-3"
+
+for task data_path in ${(kv)tasks}; do
+ for lr in $(echo "$lrs"); do
+ PYTHONPATH=. PREFIX="${PREFIX}" SUFFIX="" nohup python fairseq_cli/hydra_train.py -m --config-dir examples/roberta/config/finetuning \
+ --config-name $task hydra/launcher=submitit_slurm +run_config=slurm_1g task.data="$data_path" hydra.launcher.name=finetune_${task}_${PREFIX} \
+ checkpoint.restore_file="$cp" hydra.sweep.dir="$dir/finetune_sweep/$task/lr_$lr" "optimization.lr=[${lr}]" &
+ done
+done
diff --git a/fairseq/examples/data2vec/scripts/text/glue.py b/fairseq/examples/data2vec/scripts/text/glue.py
new file mode 100644
index 0000000000000000000000000000000000000000..5382d31834e0b7db5ad5ecee7b09072adff747e6
--- /dev/null
+++ b/fairseq/examples/data2vec/scripts/text/glue.py
@@ -0,0 +1,34 @@
+from valids import parser, main as valids_main
+import os.path as osp
+
+
+args = parser.parse_args()
+args.target = "valid_accuracy"
+args.best_biggest = True
+args.best = True
+args.last = 0
+args.path_contains = None
+
+res = valids_main(args, print_output=False)
+
+grouped = {}
+for k, v in res.items():
+ k = osp.dirname(k)
+ run = osp.dirname(k)
+ task = osp.basename(k)
+ val = v["valid_accuracy"]
+
+ if run not in grouped:
+ grouped[run] = {}
+
+ grouped[run][task] = val
+
+for run, tasks in grouped.items():
+ print(run)
+ avg = sum(float(v) for v in tasks.values()) / len(tasks)
+ avg_norte = sum(float(v) for k,v in tasks.items() if k != 'rte') / (len(tasks) -1)
+ try:
+ print(f"{tasks['cola']}\t{tasks['qnli']}\t{tasks['mrpc']}\t{tasks['rte']}\t{tasks['sst_2']}\t{avg:.2f}\t{avg_norte:.2f}")
+ except:
+ print(tasks)
+ print()
diff --git a/fairseq/examples/data2vec/scripts/text/glue_lr.py b/fairseq/examples/data2vec/scripts/text/glue_lr.py
new file mode 100644
index 0000000000000000000000000000000000000000..75bdfe0368032065fe66008f59db32ff5de26ff9
--- /dev/null
+++ b/fairseq/examples/data2vec/scripts/text/glue_lr.py
@@ -0,0 +1,143 @@
+import os.path as osp
+import re
+from collections import defaultdict
+
+from valids import parser, main as valids_main
+
+
+TASK_TO_METRIC = {
+ "cola": "mcc",
+ "qnli": "accuracy",
+ "mrpc": "acc_and_f1",
+ "rte": "accuracy",
+ "sst_2": "accuracy",
+ "mnli": "accuracy",
+ "qqp": "acc_and_f1",
+ "sts_b": "pearson_and_spearman",
+}
+TASKS = ["cola", "qnli", "mrpc", "rte", "sst_2", "mnli", "qqp", "sts_b"]
+
+
+def get_best_stat_str(task_vals, show_subdir):
+ task_to_best_val = {}
+ task_to_best_dir = {}
+ for task, subdir_to_val in task_vals.items():
+ task_to_best_val[task] = max(subdir_to_val.values())
+ task_to_best_dir[task] = max(subdir_to_val.keys(), key=lambda x: subdir_to_val[x])
+
+ # import pdb; pdb.set_trace()
+ N1 = len(task_to_best_val)
+ N2 = len([k for k in task_to_best_val if k != "rte"])
+ avg1 = sum(task_to_best_val.values()) / N1
+ avg2 = sum(v for task, v in task_to_best_val.items() if task != "rte") / N2
+
+ try:
+ msg = ""
+ for task in TASKS:
+ dir = task_to_best_dir.get(task, 'null')
+ val = task_to_best_val.get(task, -100)
+ msg += f"({dir}, {val})\t" if show_subdir else f"{val}\t"
+ msg += f"{avg1:.2f}\t{avg2:.2f}"
+ except Exception as e:
+ msg = str(e)
+ msg += str(sorted(task_vals.items()))
+ return msg
+
+def get_all_stat_str(task_vals):
+ msg = ""
+ for task in [task for task in TASKS if task in task_vals]:
+ msg += f"=== {task}\n"
+ for subdir in sorted(task_vals[task].keys()):
+ msg += f"\t{subdir}\t{task_vals[task][subdir]}\n"
+ return msg
+
+def get_tabular_stat_str(task_vals):
+ """assume subdir is /run_*/0"""
+ msg = ""
+ for task in [task for task in TASKS if task in task_vals]:
+ msg += f"=== {task}\n"
+ param_to_runs = defaultdict(dict)
+ for subdir in task_vals[task]:
+ match = re.match("(.*)/(run_.*)/0", subdir)
+ assert match, "subdir"
+ param, run = match.groups()
+ param_to_runs[param][run] = task_vals[task][subdir]
+ params = sorted(param_to_runs, key=lambda x: float(x))
+ runs = sorted(set(run for runs in param_to_runs.values() for run in runs))
+ msg += ("runs:" + "\t".join(runs) + "\n")
+ msg += ("params:" + "\t".join(params) + "\n")
+ for param in params:
+ msg += "\t".join([str(param_to_runs[param].get(run, None)) for run in runs])
+ msg += "\n"
+ # for subdir in sorted(task_vals[task].keys()):
+ # msg += f"\t{subdir}\t{task_vals[task][subdir]}\n"
+ return msg
+
+
+
+def main():
+ parser.add_argument("--show_glue", action="store_true", help="show glue metric for each task instead of accuracy")
+ parser.add_argument("--print_mode", default="best", help="best|all|tabular")
+ parser.add_argument("--show_subdir", action="store_true", help="print the subdir that has the best results for each run")
+ parser.add_argument("--override_target", default="valid_accuracy", help="override target")
+
+ args = parser.parse_args()
+ args.target = args.override_target
+ args.best_biggest = True
+ args.best = True
+ args.last = 0
+ args.path_contains = None
+
+ res = valids_main(args, print_output=False)
+ grouped_acc = {}
+ grouped_met = {} # use official metric for each task
+ for path, v in res.items():
+ path = "/".join([args.base, path])
+ path = re.sub("//*", "/", path)
+ match = re.match("(.*)finetune[^/]*/([^/]*)/(.*)", path)
+ if not match:
+ continue
+ run, task, subdir = match.groups()
+
+ if run not in grouped_acc:
+ grouped_acc[run] = {}
+ grouped_met[run] = {}
+ if task not in grouped_acc[run]:
+ grouped_acc[run][task] = {}
+ grouped_met[run][task] = {}
+
+ if v is not None:
+ grouped_acc[run][task][subdir] = float(v.get("valid_accuracy", -100))
+ grouped_met[run][task][subdir] = float(v.get(f"valid_{TASK_TO_METRIC[task]}", -100))
+ else:
+ print(f"{path} has None return")
+
+ header = "\t".join(TASKS)
+ for run in sorted(grouped_acc):
+ print(run)
+ if args.print_mode == "all":
+ if args.show_glue:
+ print("===== GLUE =====")
+ print(get_all_stat_str(grouped_met[run]))
+ else:
+ print("===== ACC =====")
+ print(get_all_stat_str(grouped_acc[run]))
+ elif args.print_mode == "best":
+ print(f" {header}")
+ if args.show_glue:
+ print(f"GLEU: {get_best_stat_str(grouped_met[run], args.show_subdir)}")
+ else:
+ print(f"ACC: {get_best_stat_str(grouped_acc[run], args.show_subdir)}")
+ elif args.print_mode == "tabular":
+ if args.show_glue:
+ print("===== GLUE =====")
+ print(get_tabular_stat_str(grouped_met[run]))
+ else:
+ print("===== ACC =====")
+ print(get_tabular_stat_str(grouped_acc[run]))
+ else:
+ raise ValueError(args.print_mode)
+ print()
+
+if __name__ == "__main__":
+ main()
diff --git a/fairseq/examples/data2vec/scripts/text/unprocess_data.py b/fairseq/examples/data2vec/scripts/text/unprocess_data.py
new file mode 100644
index 0000000000000000000000000000000000000000..f1acb624b87de207f7eee820f3b3cba1940599b2
--- /dev/null
+++ b/fairseq/examples/data2vec/scripts/text/unprocess_data.py
@@ -0,0 +1,188 @@
+import json
+import os
+import tqdm
+from fairseq.data import Dictionary, data_utils
+
+
+def load_dictionary(dict_path):
+ return Dictionary.load(dict_path)
+
+def load_dataset(split_path, src_dict):
+ dataset = data_utils.load_indexed_dataset(
+ split_path,
+ src_dict,
+ combine=False, # set to true for loading `train*`
+ )
+ if dataset is None:
+ raise FileNotFoundError(f"Dataset not found: {split_path}")
+ return dataset
+
+def load_bpe(enc_path):
+ with open(enc_path) as f:
+ bpe2idx = json.load(f)
+ idx2bpe = {v: k for k, v in bpe2idx.items()}
+ return bpe2idx, idx2bpe
+
+def detokenize(tokens, src_dict, idx2bpe):
+ raw_inds = map(int, src_dict.string(tokens).split())
+ raw_chrs = "".join([idx2bpe[raw_ind] for raw_ind in raw_inds])
+ raw_chrs = raw_chrs.replace("\u0120", " ")
+ return raw_chrs
+
+def _main(src_root, src_dict_path, src_bpe_path, src_splits, tgt_root, tgt_splits):
+ src_dict = load_dictionary(src_dict_path)
+ bpe2idx, idx2bpe = load_bpe(src_bpe_path)
+
+ assert len(src_splits) == len(tgt_splits)
+ for src_split, tgt_split in zip(src_splits, tgt_splits):
+ src_dataset = load_dataset(f"{src_root}/{src_split}", src_dict)
+ tgt_path = f"{tgt_root}/{tgt_split}.txt"
+ print(f"processing {src_split} (dump to {tgt_path})...")
+ os.makedirs(os.path.dirname(tgt_path), exist_ok=True)
+ with open(tgt_path, "w") as f:
+ for tokens in tqdm.tqdm(src_dataset):
+ raw_str = detokenize(tokens, src_dict, idx2bpe)
+ f.write(raw_str + "\n")
+
+def main_pt():
+ src_root = "/datasets01/bookwiki_CC-NEWS_openwebtext_stories-mmap2-bin/121219/bookwiki_CC-NEWS_openwebtext_stories-mmap2-bin"
+ src_dict_path = f"{src_root}/dict.txt"
+ src_bpe_path = f"{src_root}/encoder.json"
+ src_splits = [
+ "bookwiki_aml-mmap2-bin/shard0/train",
+ "bookwiki_aml-mmap2-bin/shard1/train",
+ "bookwiki_aml-mmap2-bin/shard2/train",
+ "bookwiki_aml-mmap2-bin/shard3/train",
+ "bookwiki_aml-mmap2-bin/shard4/train",
+ "bookwiki_aml-mmap2-bin/valid/valid",
+ ]
+
+ tgt_root = "/checkpoint/wnhsu/data/data2vec2/data/text/bookwiki_aml-full-mmap2-txt"
+ tgt_splits = [
+ "train0",
+ "train1",
+ "train2",
+ "train3",
+ "train4",
+ "valid",
+ ]
+ _main(src_root, src_dict_path, src_bpe_path, src_splits, tgt_root, tgt_splits)
+
+def main_ft():
+ src_root = "/fsx-wav2vec/wnhsu/data/data2vec2/data/text/GLUE"
+ src_dict_path = f"{src_root}/dict.txt"
+ src_bpe_path = f"{src_root}/encoder.json"
+ src_splits = [
+ "CoLA-bin/input0/train",
+ "CoLA-bin/input0/valid",
+ "CoLA-bin/input0/test",
+
+ "MNLI-bin/input0/train",
+ "MNLI-bin/input0/valid",
+ "MNLI-bin/input0/test",
+ "MNLI-bin/input0/test1",
+ "MNLI-bin/input1/train",
+ "MNLI-bin/input1/valid",
+ "MNLI-bin/input1/test",
+ "MNLI-bin/input1/test1",
+
+ "MRPC-bin/input0/train",
+ "MRPC-bin/input0/valid",
+ "MRPC-bin/input0/test",
+ "MRPC-bin/input1/train",
+ "MRPC-bin/input1/valid",
+ "MRPC-bin/input1/test",
+
+ "QNLI-bin/input0/train",
+ "QNLI-bin/input0/valid",
+ "QNLI-bin/input0/test",
+ "QNLI-bin/input1/train",
+ "QNLI-bin/input1/valid",
+ "QNLI-bin/input1/test",
+
+ "QQP-bin/input0/train",
+ "QQP-bin/input0/valid",
+ "QQP-bin/input0/test",
+ "QQP-bin/input1/train",
+ "QQP-bin/input1/valid",
+ "QQP-bin/input1/test",
+
+ "RTE-bin/input0/train",
+ "RTE-bin/input0/valid",
+ "RTE-bin/input0/test",
+ "RTE-bin/input1/train",
+ "RTE-bin/input1/valid",
+ "RTE-bin/input1/test",
+
+ "SST-2-bin/input0/train",
+ "SST-2-bin/input0/valid",
+ "SST-2-bin/input0/test",
+
+ "STS-B-bin/input0/train",
+ "STS-B-bin/input0/valid",
+ "STS-B-bin/input0/test",
+ "STS-B-bin/input1/train",
+ "STS-B-bin/input1/valid",
+ "STS-B-bin/input1/test",
+ ]
+
+ tgt_root = "/fsx-wav2vec/wnhsu/data/data2vec2/data/text/GLUE_chr"
+ tgt_splits = [
+ "CoLA-bin/input0/train",
+ "CoLA-bin/input0/valid",
+ "CoLA-bin/input0/test",
+
+ "MNLI-bin/input0/train",
+ "MNLI-bin/input0/valid",
+ "MNLI-bin/input0/test",
+ "MNLI-bin/input0/test1",
+ "MNLI-bin/input1/train",
+ "MNLI-bin/input1/valid",
+ "MNLI-bin/input1/test",
+ "MNLI-bin/input1/test1",
+
+ "MRPC-bin/input0/train",
+ "MRPC-bin/input0/valid",
+ "MRPC-bin/input0/test",
+ "MRPC-bin/input1/train",
+ "MRPC-bin/input1/valid",
+ "MRPC-bin/input1/test",
+
+ "QNLI-bin/input0/train",
+ "QNLI-bin/input0/valid",
+ "QNLI-bin/input0/test",
+ "QNLI-bin/input1/train",
+ "QNLI-bin/input1/valid",
+ "QNLI-bin/input1/test",
+
+ "QQP-bin/input0/train",
+ "QQP-bin/input0/valid",
+ "QQP-bin/input0/test",
+ "QQP-bin/input1/train",
+ "QQP-bin/input1/valid",
+ "QQP-bin/input1/test",
+
+ "RTE-bin/input0/train",
+ "RTE-bin/input0/valid",
+ "RTE-bin/input0/test",
+ "RTE-bin/input1/train",
+ "RTE-bin/input1/valid",
+ "RTE-bin/input1/test",
+
+ "SST-2-bin/input0/train",
+ "SST-2-bin/input0/valid",
+ "SST-2-bin/input0/test",
+
+ "STS-B-bin/input0/train",
+ "STS-B-bin/input0/valid",
+ "STS-B-bin/input0/test",
+ "STS-B-bin/input1/train",
+ "STS-B-bin/input1/valid",
+ "STS-B-bin/input1/test",
+ ]
+ _main(src_root, src_dict_path, src_bpe_path, src_splits, tgt_root, tgt_splits)
+
+
+if __name__ == "__main__":
+ main_pt()
+ main_ft()
diff --git a/fairseq/examples/data2vec/scripts/text/valids.py b/fairseq/examples/data2vec/scripts/text/valids.py
new file mode 100644
index 0000000000000000000000000000000000000000..b2e5cfb25d0e21cffeee21f719c008c82e1fa329
--- /dev/null
+++ b/fairseq/examples/data2vec/scripts/text/valids.py
@@ -0,0 +1,301 @@
+import os, argparse, re, json, copy, math
+from collections import OrderedDict
+import numpy as np
+
+parser = argparse.ArgumentParser(description='Process some integers.')
+parser.add_argument('base', help='base log path')
+parser.add_argument('--file_name', default='train.log', help='the log file name')
+parser.add_argument('--target', default='valid_loss', help='target metric')
+parser.add_argument('--last', type=int, default=999999999, help='print last n matches')
+parser.add_argument('--last_files', type=int, default=None, help='print last x files')
+parser.add_argument('--everything', action='store_true', help='print everything instead of only last match')
+parser.add_argument('--path_contains', help='only consider matching file pattern')
+parser.add_argument('--group_on', help='if set, groups by this metric and shows table of differences')
+parser.add_argument('--epoch', help='epoch for comparison', type=int)
+parser.add_argument('--skip_empty', action='store_true', help='skip empty results')
+parser.add_argument('--skip_containing', help='skips entries containing this attribute')
+parser.add_argument('--unique_epochs', action='store_true', help='only consider the last line fore each epoch')
+parser.add_argument('--best', action='store_true', help='print the last best result')
+parser.add_argument('--avg_params', help='average these params through entire log')
+parser.add_argument('--extract_prev', help='extracts this metric from previous line')
+
+parser.add_argument('--remove_metric', help='extracts this metric from previous line')
+
+parser.add_argument('--compact', action='store_true', help='if true, just prints checkpoint best val')
+parser.add_argument('--hydra', action='store_true', help='if true, uses hydra param conventions')
+
+parser.add_argument('--best_biggest', action='store_true', help='if true, best is the biggest number, not smallest')
+parser.add_argument('--key_len', type=int, default=10, help='max length of key')
+
+parser.add_argument('--best_only', action='store_true', help='if set, only prints the best value')
+parser.add_argument('--flat', action='store_true', help='just print the best results')
+
+
+def main(args, print_output):
+ ret = {}
+
+ entries = []
+
+ def extract_metric(s, metric):
+ try:
+ j = json.loads(s)
+ except:
+ return None
+ if args.epoch is not None and ('epoch' not in j or j['epoch'] != args.epoch):
+ return None
+ return j[metric] if metric in j else None
+
+
+ def extract_params(s):
+ s = s.replace(args.base, '', 1)
+ if args.path_contains is not None:
+ s = s.replace(args.path_contains, '', 1)
+
+ if args.hydra:
+ num_matches = re.findall(r'(?:/|__)([^/:]+):(\d+\.?\d*)', s)
+ # str_matches = re.findall(r'(?:/|__)([^/:]+):([^\.]*[^\d\.]+)(?:/|__)', s)
+ str_matches = re.findall(r'(?:/|__)?((?:(?!(?:\:|__)).)+):([^\.]*[^\d\.]+\d*)(?:/|__)', s)
+ lr_matches = re.findall(r'optimization.(lr):\[([\d\.,]+)\]', s)
+ task_matches = re.findall(r'.*/(\d+)$', s)
+ else:
+ num_matches = re.findall(r'\.?([^\.]+?)(\d+(e\-\d+)?(?:\.\d+)?)(\.|$)', s)
+ str_matches = re.findall(r'[/\.]([^\.]*[^\d\.]+\d*)(?=\.)', s)
+ lr_matches = []
+ task_matches = []
+
+ cp_matches = re.findall(r'checkpoint(?:_\d+)?_(\d+).pt', s)
+
+ items = OrderedDict()
+ for m in str_matches:
+ if isinstance(m, tuple):
+ if 'checkpoint' not in m[0]:
+ items[m[0]] = m[1]
+ else:
+ items[m] = ''
+
+ for m in num_matches:
+ items[m[0]] = m[1]
+
+ for m in lr_matches:
+ items[m[0]] = m[1]
+
+ for m in task_matches:
+ items["hydra_task"] = m
+
+ for m in cp_matches:
+ items['checkpoint'] = m
+
+ return items
+
+ abs_best = None
+
+ sources = []
+ for root, _, files in os.walk(args.base):
+ if args.path_contains is not None and not args.path_contains in root:
+ continue
+ for f in files:
+ if f.endswith(args.file_name):
+ sources.append((root, f))
+
+ if args.last_files is not None:
+ sources = sources[-args.last_files:]
+
+ for root, file in sources:
+ with open(os.path.join(root, file), 'r') as fin:
+ found = []
+ avg = {}
+ prev = None
+ for line in fin:
+ line = line.rstrip()
+ if line.find(args.target) != -1 and (
+ args.skip_containing is None or line.find(args.skip_containing) == -1):
+ try:
+ idx = line.index("{")
+ line = line[idx:]
+ line_json = json.loads(line)
+ except:
+ continue
+ if prev is not None:
+ try:
+ prev.update(line_json)
+ line_json = prev
+ except:
+ pass
+ if args.target in line_json:
+ found.append(line_json)
+ if args.avg_params:
+ avg_params = args.avg_params.split(',')
+ for p in avg_params:
+ m = extract_metric(line, p)
+ if m is not None:
+ prev_v, prev_c = avg.get(p, (0, 0))
+ avg[p] = prev_v + float(m), prev_c + 1
+ if args.extract_prev:
+ try:
+ prev = json.loads(line)
+ except:
+ pass
+ best = None
+ if args.best:
+ curr_best = None
+ for i in range(len(found)):
+ cand_best = found[i][args.target] if args.target in found[i] else None
+
+ def cmp(a, b):
+ a = float(a)
+ b = float(b)
+ if args.best_biggest:
+ return a > b
+ return a < b
+
+ if cand_best is not None and not math.isnan(float(cand_best)) and (
+ curr_best is None or cmp(cand_best, curr_best)):
+ curr_best = cand_best
+ if abs_best is None or cmp(curr_best, abs_best):
+ abs_best = curr_best
+ best = found[i]
+ if args.unique_epochs or args.epoch:
+ last_found = []
+ last_epoch = None
+ for i in reversed(range(len(found))):
+ epoch = found[i]['epoch']
+ if args.epoch and args.epoch != epoch:
+ continue
+ if epoch != last_epoch:
+ last_epoch = epoch
+ last_found.append(found[i])
+ found = list(reversed(last_found))
+
+ if len(found) == 0:
+ if print_output and (args.last_files is not None or not args.skip_empty):
+ # print(root.split('/')[-1])
+ print(root[len(args.base):])
+ print('Nothing')
+ else:
+ if not print_output:
+ ret[root[len(args.base):]] = best
+ continue
+
+ if args.compact:
+ # print('{}\t{}'.format(root.split('/')[-1], curr_best))
+ print('{}\t{}'.format(root[len(args.base)+1:], curr_best))
+ continue
+
+ if args.group_on is None and not args.best_only:
+ # print(root.split('/')[-1])
+ print(root[len(args.base):])
+ if not args.everything:
+ if best is not None and args.group_on is None and not args.best_only and not args.flat:
+ print(best, '(best)')
+ if args.group_on is None and args.last and not args.best_only and not args.flat:
+ for f in found[-args.last:]:
+ if args.extract_prev is not None:
+ try:
+ print('{}\t{}'.format(f[args.extract_prev], f[args.target]))
+ except Exception as e:
+ print('Exception!', e)
+ else:
+ print(f)
+ try:
+ metric = found[-1][args.target] if not args.best or best is None else best[args.target]
+ except:
+ print(found[-1])
+ raise
+ if metric is not None:
+ entries.append((extract_params(root), metric))
+ else:
+ for f in found:
+ print(f)
+ if not args.group_on and print_output:
+ print()
+
+ if len(avg) > 0:
+ for k, (v, c) in avg.items():
+ print(f'{k}: {v/c}')
+
+ if args.best_only:
+ print(abs_best)
+
+ if args.flat:
+ print("\t".join(m for _, m in entries))
+
+ if args.group_on is not None:
+ by_val = OrderedDict()
+ for e, m in entries:
+ k = args.group_on
+ if k not in e:
+ m_keys = [x for x in e.keys() if x.startswith(k)]
+ if len(m_keys) == 0:
+ val = "False"
+ else:
+ assert len(m_keys) == 1
+ k = m_keys[0]
+ val = m_keys[0]
+ else:
+ val = e[args.group_on]
+ if val == "":
+ val = "True"
+ scrubbed_entry = copy.deepcopy(e)
+ if k in scrubbed_entry:
+ del scrubbed_entry[k]
+ if args.remove_metric and args.remove_metric in scrubbed_entry:
+ val += '_' + scrubbed_entry[args.remove_metric]
+ del scrubbed_entry[args.remove_metric]
+ by_val.setdefault(tuple(scrubbed_entry.items()), dict())[val] = m
+ distinct_vals = set()
+ for v in by_val.values():
+ distinct_vals.update(v.keys())
+ try:
+ distinct_vals = {int(d) for d in distinct_vals}
+ except:
+ print(distinct_vals)
+ print()
+ print("by_val", len(by_val))
+ for k,v in by_val.items():
+ print(k, '=>', v)
+ print()
+
+ # , by_val, entries)
+ raise
+ from natsort import natsorted
+ svals = list(map(str, natsorted(distinct_vals)))
+ print('{}\t{}'.format(args.group_on, '\t'.join(svals)))
+ sums = OrderedDict({n:[] for n in svals})
+ for k, v in by_val.items():
+ kstr = '.'.join(':'.join(x) for x in k)
+ vstr = ''
+ for mv in svals:
+ x = v[mv] if mv in v else ''
+ vstr += '\t{}'.format(round(x, 5) if isinstance(x, float) else x)
+ try:
+ sums[mv].append(float(x))
+ except:
+ pass
+ print('{}{}'.format(kstr[:args.key_len], vstr))
+ if any(len(x) > 0 for x in sums.values()):
+ print('min:', end='')
+ for v in sums.values():
+ min = np.min(v)
+ print(f'\t{round(min, 5)}', end='')
+ print()
+ print('max:', end='')
+ for v in sums.values():
+ max = np.max(v)
+ print(f'\t{round(max, 5)}', end='')
+ print()
+ print('avg:', end='')
+ for v in sums.values():
+ mean = np.mean(v)
+ print(f'\t{round(mean, 5)}', end='')
+ print()
+ print('median:', end='')
+ for v in sums.values():
+ median = np.median(v)
+ print(f'\t{round(median, 5)}', end='')
+ print()
+
+ return ret
+
+if __name__ == "__main__":
+ args = parser.parse_args()
+ main(args, print_output=True)
\ No newline at end of file
diff --git a/fairseq/examples/data2vec/tasks/__init__.py b/fairseq/examples/data2vec/tasks/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..a7422e4b30d3aeb1069d2cf11bb3622cf3e5686f
--- /dev/null
+++ b/fairseq/examples/data2vec/tasks/__init__.py
@@ -0,0 +1,18 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+from .image_pretraining import ImagePretrainingTask, ImagePretrainingConfig
+from .image_classification import ImageClassificationTask, ImageClassificationConfig
+from .mae_image_pretraining import MaeImagePretrainingTask, MaeImagePretrainingConfig
+
+
+__all__ = [
+ "ImageClassificationTask",
+ "ImageClassificationConfig",
+ "ImagePretrainingTask",
+ "ImagePretrainingConfig",
+ "MaeImagePretrainingTask",
+ "MaeImagePretrainingConfig",
+]
\ No newline at end of file
diff --git a/fairseq/examples/data2vec/tasks/audio_classification.py b/fairseq/examples/data2vec/tasks/audio_classification.py
new file mode 100644
index 0000000000000000000000000000000000000000..2925a04cf9c6220db04a567ac023ba074ee222c5
--- /dev/null
+++ b/fairseq/examples/data2vec/tasks/audio_classification.py
@@ -0,0 +1,167 @@
+# Copyright (c) 2017-present, Facebook, Inc.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the LICENSE file in
+# the root directory of this source tree. An additional grant of patent rights
+# can be found in the PATENTS file in the same directory.
+
+import logging
+import os
+import numpy as np
+import math
+import torch
+
+from sklearn import metrics as sklearn_metrics
+from dataclasses import dataclass
+
+from fairseq.tasks.audio_pretraining import AudioPretrainingTask, AudioPretrainingConfig
+from fairseq.tasks import register_task
+from fairseq.logging import metrics
+
+from ..data.add_class_target_dataset import AddClassTargetDataset
+
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class AudioClassificationConfig(AudioPretrainingConfig):
+ label_descriptors: str = "label_descriptors.csv"
+ labels: str = "lbl"
+
+
+@register_task("audio_classification", dataclass=AudioClassificationConfig)
+class AudioClassificationTask(AudioPretrainingTask):
+ """ """
+
+ cfg: AudioClassificationConfig
+
+ def __init__(
+ self,
+ cfg: AudioClassificationConfig,
+ ):
+ super().__init__(cfg)
+
+ self.state.add_factory("labels", self.load_labels)
+
+ def load_labels(self):
+ labels = {}
+ path = os.path.join(self.cfg.data, self.cfg.label_descriptors)
+ with open(path, "r") as ldf:
+ for line in ldf:
+ if line.strip() == "":
+ continue
+ items = line.split(",")
+ idx = items[0]
+ lbl = items[1]
+ assert lbl not in labels, lbl
+ labels[lbl] = idx
+ return labels
+
+ @property
+ def labels(self):
+ return self.state.labels
+
+ def load_dataset(
+ self, split: str, task_cfg: AudioClassificationConfig = None, **kwargs
+ ):
+ super().load_dataset(split, task_cfg, **kwargs)
+
+ task_cfg = task_cfg or self.cfg
+
+ data_path = self.cfg.data
+ label_path = os.path.join(data_path, f"{split}.{task_cfg.labels}")
+ skipped_indices = getattr(self.datasets[split], "skipped_indices", set())
+ labels = []
+ with open(label_path, "r") as f:
+ for i, line in enumerate(f):
+ if i not in skipped_indices:
+ lbl_items = line.rstrip().split("\t")
+ labels.append([int(x) for x in lbl_items[2].split(",")])
+
+ assert len(labels) == len(self.datasets[split]), (
+ f"labels length ({len(labels)}) and dataset length "
+ f"({len(self.datasets[split])}) do not match"
+ )
+
+ self.datasets[split] = AddClassTargetDataset(
+ self.datasets[split],
+ labels,
+ multi_class=True,
+ add_to_input=True,
+ num_classes=len(self.labels),
+ )
+
+ def calculate_stats(self, output, target):
+
+ classes_num = target.shape[-1]
+ stats = []
+
+ # Accuracy, only used for single-label classification such as esc-50, not for multiple label one such as AudioSet
+ # acc = sklearn_metrics.accuracy_score(np.argmax(target, 1), np.argmax(output, 1))
+
+ # Class-wise statistics
+ for k in range(classes_num):
+ # Average precision
+ avg_precision = sklearn_metrics.average_precision_score(
+ target[:, k], output[:, k], average=None
+ )
+
+ dict = {
+ "AP": avg_precision,
+ }
+
+ # # AUC
+ # try:
+ # auc = sklearn_metrics.roc_auc_score(target[:, k], output[:, k], average=None)
+ # except:
+ # auc = 0
+ #
+ # # Precisions, recalls
+ # (precisions, recalls, thresholds) = sklearn_metrics.precision_recall_curve(
+ # target[:, k], output[:, k]
+ # )
+ #
+ # # FPR, TPR
+ # (fpr, tpr, thresholds) = sklearn_metrics.roc_curve(target[:, k], output[:, k])
+ #
+ # save_every_steps = 1000 # Sample statistics to reduce size
+ # dict = {
+ # "precisions": precisions[0::save_every_steps],
+ # "recalls": recalls[0::save_every_steps],
+ # "AP": avg_precision,
+ # "fpr": fpr[0::save_every_steps],
+ # "fnr": 1.0 - tpr[0::save_every_steps],
+ # "auc": auc,
+ # # note acc is not class-wise, this is just to keep consistent with other metrics
+ # "acc": acc,
+ # }
+ stats.append(dict)
+
+ return stats
+
+ def valid_step(self, sample, model, criterion):
+ loss, sample_size, logging_output = super().valid_step(sample, model, criterion)
+ return loss, sample_size, logging_output
+
+ def reduce_metrics(self, logging_outputs, criterion):
+ super().reduce_metrics(logging_outputs, criterion)
+ if "_predictions" in logging_outputs[0]:
+ metrics.log_concat_tensor(
+ "_predictions",
+ torch.cat([l["_predictions"].cpu() for l in logging_outputs], dim=0),
+ )
+ metrics.log_concat_tensor(
+ "_targets",
+ torch.cat([l["_targets"].cpu() for l in logging_outputs], dim=0),
+ )
+
+ def compute_stats(meters):
+ if meters["_predictions"].tensor.shape[0] < 100:
+ return 0
+ stats = self.calculate_stats(
+ meters["_predictions"].tensor, meters["_targets"].tensor
+ )
+ return np.nanmean([stat["AP"] for stat in stats])
+
+ metrics.log_derived("mAP", compute_stats)
diff --git a/fairseq/examples/data2vec/tasks/image_classification.py b/fairseq/examples/data2vec/tasks/image_classification.py
new file mode 100644
index 0000000000000000000000000000000000000000..1ea4c2afeeb59345d91418a22fda5fc576c5c961
--- /dev/null
+++ b/fairseq/examples/data2vec/tasks/image_classification.py
@@ -0,0 +1,129 @@
+# Copyright (c) 2017-present, Facebook, Inc.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the LICENSE file in
+# the root directory of this source tree. An additional grant of patent rights
+# can be found in the PATENTS file in the same directory.
+
+import os.path as osp
+import logging
+
+from dataclasses import dataclass
+import torch
+from torchvision import transforms
+
+from fairseq.dataclass import FairseqDataclass
+from fairseq.tasks import register_task
+from fairseq.logging import metrics
+
+try:
+ from ..data import ImageDataset
+except:
+ import sys
+
+ sys.path.append("..")
+ from data import ImageDataset
+
+from .image_pretraining import (
+ ImagePretrainingConfig,
+ ImagePretrainingTask,
+ IMG_EXTENSIONS,
+)
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class ImageClassificationConfig(ImagePretrainingConfig):
+ pass
+
+
+@register_task("image_classification", dataclass=ImageClassificationConfig)
+class ImageClassificationTask(ImagePretrainingTask):
+
+ cfg: ImageClassificationConfig
+
+ @classmethod
+ def setup_task(cls, cfg: ImageClassificationConfig, **kwargs):
+ return cls(cfg)
+
+ def load_dataset(self, split: str, task_cfg: FairseqDataclass = None, **kwargs):
+ data_path = self.cfg.data
+ cfg = task_cfg or self.cfg
+
+ path_with_split = osp.join(data_path, split)
+ if osp.exists(path_with_split):
+ data_path = path_with_split
+
+ from timm.data import create_transform
+
+ if split == "train":
+ # this should always dispatch to transforms_imagenet_train
+ transform = create_transform(
+ input_size=cfg.input_size,
+ is_training=True,
+ auto_augment="rand-m9-mstd0.5-inc1",
+ interpolation="bicubic",
+ re_prob=0.25,
+ re_mode="pixel",
+ re_count=1,
+ mean=cfg.normalization_mean,
+ std=cfg.normalization_std,
+ )
+ if not cfg.input_size > 32:
+ transform.transforms[0] = transforms.RandomCrop(
+ cfg.input_size, padding=4
+ )
+ else:
+ t = []
+ if cfg.input_size > 32:
+ crop_pct = 1
+ if cfg.input_size < 384:
+ crop_pct = 224 / 256
+ size = int(cfg.input_size / crop_pct)
+ t.append(
+ transforms.Resize(
+ size, interpolation=3
+ ), # to maintain same ratio w.r.t. 224 images
+ )
+ t.append(transforms.CenterCrop(cfg.input_size))
+
+ t.append(transforms.ToTensor())
+ t.append(
+ transforms.Normalize(cfg.normalization_mean, cfg.normalization_std)
+ )
+ transform = transforms.Compose(t)
+ logger.info(transform)
+
+ self.datasets[split] = ImageDataset(
+ root=data_path,
+ extensions=IMG_EXTENSIONS,
+ load_classes=True,
+ transform=transform,
+ )
+ for k in self.datasets.keys():
+ if k != split:
+ assert self.datasets[k].classes == self.datasets[split].classes
+
+ def build_model(self, model_cfg: FairseqDataclass, from_checkpoint=False):
+ model = super().build_model(model_cfg, from_checkpoint)
+
+ actualized_cfg = getattr(model, "cfg", None)
+ if actualized_cfg is not None:
+ if hasattr(actualized_cfg, "pretrained_model_args"):
+ model_cfg.pretrained_model_args = actualized_cfg.pretrained_model_args
+
+ return model
+
+ def reduce_metrics(self, logging_outputs, criterion):
+ super().reduce_metrics(logging_outputs, criterion)
+
+ if "correct" in logging_outputs[0]:
+ zero = torch.scalar_tensor(0.0)
+ correct = sum(log.get("correct", zero) for log in logging_outputs)
+ metrics.log_scalar_sum("_correct", correct)
+
+ metrics.log_derived(
+ "accuracy",
+ lambda meters: 100 * meters["_correct"].sum / meters["sample_size"].sum,
+ )
diff --git a/fairseq/examples/data2vec/tasks/image_pretraining.py b/fairseq/examples/data2vec/tasks/image_pretraining.py
new file mode 100644
index 0000000000000000000000000000000000000000..cd688fd1366f3ee4145726307d41be3174a44151
--- /dev/null
+++ b/fairseq/examples/data2vec/tasks/image_pretraining.py
@@ -0,0 +1,110 @@
+# Copyright (c) 2017-present, Facebook, Inc.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the LICENSE file in
+# the root directory of this source tree. An additional grant of patent rights
+# can be found in the PATENTS file in the same directory.
+
+import logging
+import sys
+import os.path as osp
+
+from dataclasses import dataclass, field
+from typing import List
+from omegaconf import MISSING
+
+import torch
+from torchvision import transforms
+
+from fairseq.dataclass import FairseqDataclass
+from fairseq.tasks import FairseqTask, register_task
+
+try:
+ from ..data import ImageDataset
+except:
+ sys.path.append("..")
+ from data import ImageDataset
+
+logger = logging.getLogger(__name__)
+
+IMG_EXTENSIONS = {
+ ".jpg",
+ ".jpeg",
+ ".png",
+ ".ppm",
+ ".bmp",
+ ".pgm",
+ ".tif",
+ ".tiff",
+ ".webp",
+}
+
+
+@dataclass
+class ImagePretrainingConfig(FairseqDataclass):
+ data: str = field(default=MISSING, metadata={"help": "path to data directory"})
+ input_size: int = 224
+ normalization_mean: List[float] = (0.485, 0.456, 0.406)
+ normalization_std: List[float] = (0.229, 0.224, 0.225)
+
+
+@register_task("image_pretraining", dataclass=ImagePretrainingConfig)
+class ImagePretrainingTask(FairseqTask):
+ """ """
+
+ cfg: ImagePretrainingConfig
+
+ @classmethod
+ def setup_task(cls, cfg: ImagePretrainingConfig, **kwargs):
+ """Setup the task (e.g., load dictionaries).
+
+ Args:
+ cfg (AudioPretrainingConfig): configuration of this task
+ """
+
+ return cls(cfg)
+
+ def load_dataset(self, split: str, task_cfg: FairseqDataclass = None, **kwargs):
+ data_path = self.cfg.data
+ cfg = task_cfg or self.cfg
+
+ path_with_split = osp.join(data_path, split)
+ if osp.exists(path_with_split):
+ data_path = path_with_split
+
+ transform = transforms.Compose(
+ [
+ transforms.ColorJitter(0.4, 0.4, 0.4),
+ transforms.RandomHorizontalFlip(p=0.5),
+ transforms.RandomResizedCrop(
+ size=cfg.input_size,
+ interpolation=transforms.InterpolationMode.BICUBIC,
+ ),
+ transforms.ToTensor(),
+ transforms.Normalize(
+ mean=torch.tensor(cfg.normalization_mean),
+ std=torch.tensor(cfg.normalization_std),
+ ),
+ ]
+ )
+
+ logger.info(transform)
+
+ self.datasets[split] = ImageDataset(
+ root=data_path,
+ extensions=IMG_EXTENSIONS,
+ load_classes=False,
+ transform=transform,
+ )
+
+ @property
+ def source_dictionary(self):
+ return None
+
+ @property
+ def target_dictionary(self):
+ return None
+
+ def max_positions(self):
+ """Maximum input length supported by the encoder."""
+ return sys.maxsize, sys.maxsize
diff --git a/fairseq/examples/data2vec/tasks/mae_image_classification.py b/fairseq/examples/data2vec/tasks/mae_image_classification.py
new file mode 100644
index 0000000000000000000000000000000000000000..1bf935879fd0de1c50533b1e42a147089fd123f4
--- /dev/null
+++ b/fairseq/examples/data2vec/tasks/mae_image_classification.py
@@ -0,0 +1,100 @@
+# Copyright (c) 2017-present, Facebook, Inc.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the LICENSE file in
+# the root directory of this source tree. An additional grant of patent rights
+# can be found in the PATENTS file in the same directory.
+
+import logging
+import sys
+import torch
+
+from typing import Optional
+from dataclasses import dataclass, field
+from omegaconf import MISSING
+
+from fairseq.dataclass import FairseqDataclass
+from fairseq.tasks import FairseqTask, register_task
+from fairseq.logging import metrics
+
+try:
+ from ..data import MaeFinetuningImageDataset
+except:
+ sys.path.append("..")
+ from data import MaeFinetuningImageDataset
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class MaeImageClassificationConfig(FairseqDataclass):
+ data: str = field(default=MISSING, metadata={"help": "path to data directory"})
+ input_size: int = 224
+ local_cache_path: Optional[str] = None
+
+ rebuild_batches: bool = True
+
+
+@register_task("mae_image_classification", dataclass=MaeImageClassificationConfig)
+class MaeImageClassificationTask(FairseqTask):
+ """ """
+
+ cfg: MaeImageClassificationConfig
+
+ @classmethod
+ def setup_task(cls, cfg: MaeImageClassificationConfig, **kwargs):
+ """Setup the task (e.g., load dictionaries).
+
+ Args:
+ cfg (AudioPretrainingConfig): configuration of this task
+ """
+
+ return cls(cfg)
+
+ def load_dataset(self, split: str, task_cfg: FairseqDataclass = None, **kwargs):
+ data_path = self.cfg.data
+ cfg = task_cfg or self.cfg
+
+ self.datasets[split] = MaeFinetuningImageDataset(
+ root=data_path,
+ split=split,
+ is_train=split == "train",
+ input_size=cfg.input_size,
+ local_cache_path=cfg.local_cache_path,
+ shuffle=split == "train",
+ )
+
+ def build_model(self, model_cfg: FairseqDataclass, from_checkpoint=False):
+ model = super().build_model(model_cfg, from_checkpoint)
+
+ actualized_cfg = getattr(model, "cfg", None)
+ if actualized_cfg is not None:
+ if hasattr(actualized_cfg, "pretrained_model_args"):
+ model_cfg.pretrained_model_args = actualized_cfg.pretrained_model_args
+
+ return model
+
+ def reduce_metrics(self, logging_outputs, criterion):
+ super().reduce_metrics(logging_outputs, criterion)
+
+ if "correct" in logging_outputs[0]:
+ zero = torch.scalar_tensor(0.0)
+ correct = sum(log.get("correct", zero) for log in logging_outputs)
+ metrics.log_scalar_sum("_correct", correct)
+
+ metrics.log_derived(
+ "accuracy",
+ lambda meters: 100 * meters["_correct"].sum / meters["sample_size"].sum,
+ )
+
+ @property
+ def source_dictionary(self):
+ return None
+
+ @property
+ def target_dictionary(self):
+ return None
+
+ def max_positions(self):
+ """Maximum input length supported by the encoder."""
+ return sys.maxsize, sys.maxsize
diff --git a/fairseq/examples/data2vec/tasks/mae_image_pretraining.py b/fairseq/examples/data2vec/tasks/mae_image_pretraining.py
new file mode 100644
index 0000000000000000000000000000000000000000..35a14891ca6039943a4d7add0ed92ba45db8d1e0
--- /dev/null
+++ b/fairseq/examples/data2vec/tasks/mae_image_pretraining.py
@@ -0,0 +1,119 @@
+# Copyright (c) 2017-present, Facebook, Inc.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the LICENSE file in
+# the root directory of this source tree. An additional grant of patent rights
+# can be found in the PATENTS file in the same directory.
+
+import logging
+import sys
+
+from typing import Optional, List
+from dataclasses import dataclass, field
+from omegaconf import MISSING, II
+
+from fairseq.data import SubsampleDataset
+from fairseq.dataclass import FairseqDataclass
+from fairseq.tasks import FairseqTask, register_task
+
+try:
+ from ..data import MaeImageDataset
+except:
+ sys.path.append("..")
+ from data import MaeImageDataset
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class ImageMaskingConfig:
+ patch_size: int = II("model.modalities.image.patch_size")
+ mask_prob: float = II("model.modalities.image.mask_prob")
+ mask_prob_adjust: float = II("model.modalities.image.mask_prob_adjust")
+ mask_length: int = II("model.modalities.image.mask_length")
+ inverse_mask: bool = II("model.modalities.image.inverse_mask")
+ mask_dropout: float = II("model.modalities.image.mask_dropout")
+ clone_batch: int = II("model.clone_batch")
+ expand_adjacent: bool = False
+ non_overlapping: bool = False
+
+
+@dataclass
+class MaeImagePretrainingConfig(FairseqDataclass):
+ data: str = field(default=MISSING, metadata={"help": "path to data directory"})
+ multi_data: Optional[List[str]] = None
+ input_size: int = 224
+ local_cache_path: Optional[str] = None
+ key: str = "imgs"
+
+ beit_transforms: bool = False
+ target_transform: bool = False
+ no_transform: bool = False
+
+ rebuild_batches: bool = True
+
+ precompute_mask_config: Optional[ImageMaskingConfig] = None
+
+ subsample: float = 1
+ seed: int = II("common.seed")
+ dataset_type: str = "imagefolder"
+
+
+@register_task("mae_image_pretraining", dataclass=MaeImagePretrainingConfig)
+class MaeImagePretrainingTask(FairseqTask):
+ """ """
+
+ cfg: MaeImagePretrainingConfig
+
+ @classmethod
+ def setup_task(cls, cfg: MaeImagePretrainingConfig, **kwargs):
+ """Setup the task (e.g., load dictionaries).
+
+ Args:
+ cfg (AudioPretrainingConfig): configuration of this task
+ """
+
+ return cls(cfg)
+
+ def load_dataset(self, split: str, task_cfg: FairseqDataclass = None, **kwargs):
+ data_path = self.cfg.data
+ cfg = task_cfg or self.cfg
+
+ compute_mask = cfg.precompute_mask_config is not None
+ mask_args = {}
+ if compute_mask:
+ mask_args = cfg.precompute_mask_config
+
+ self.datasets[split] = MaeImageDataset(
+ root=data_path if cfg.multi_data is None else cfg.multi_data,
+ split=split,
+ input_size=cfg.input_size,
+ local_cache_path=cfg.local_cache_path,
+ key=cfg.key,
+ beit_transforms=cfg.beit_transforms,
+ target_transform=cfg.target_transform,
+ no_transform=cfg.no_transform,
+ compute_mask=compute_mask,
+ dataset_type=cfg.dataset_type,
+ **mask_args,
+ )
+
+ if cfg.subsample < 1:
+ self.datasets[split] = SubsampleDataset(
+ self.datasets[split],
+ cfg.subsample,
+ shuffle=True,
+ seed=cfg.seed,
+ )
+
+ @property
+ def source_dictionary(self):
+ return None
+
+ @property
+ def target_dictionary(self):
+ return None
+
+ def max_positions(self):
+ """Maximum input length supported by the encoder."""
+ return sys.maxsize, sys.maxsize
diff --git a/fairseq/examples/data2vec/tasks/multimodal.py b/fairseq/examples/data2vec/tasks/multimodal.py
new file mode 100644
index 0000000000000000000000000000000000000000..74648e918fcc54b89ee3ebe9d14987a2a74c34b8
--- /dev/null
+++ b/fairseq/examples/data2vec/tasks/multimodal.py
@@ -0,0 +1,165 @@
+# Copyright (c) 2017-present, Facebook, Inc.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the LICENSE file in
+# the root directory of this source tree. An additional grant of patent rights
+# can be found in the PATENTS file in the same directory.
+
+import sys
+
+from dataclasses import dataclass
+from typing import Optional, List
+from omegaconf import II
+
+from fairseq.data.iterators import GroupedEpochBatchIterator
+
+from fairseq.dataclass import FairseqDataclass
+from fairseq.tasks import FairseqTask, register_task
+from fairseq.tasks.audio_pretraining import AudioPretrainingConfig, AudioPretrainingTask
+from fairseq.tasks.masked_lm import MaskedLMConfig, MaskedLMTask
+from .mae_image_pretraining import MaeImagePretrainingConfig, MaeImagePretrainingTask
+from examples.data2vec.data.modality import Modality
+
+from fairseq.data.audio.multi_modality_dataset import (
+ MultiModalityDataset,
+ ModalityDatasetItem,
+)
+
+
+@dataclass
+class MultimodalPretrainingConfig(FairseqDataclass):
+ audio: Optional[AudioPretrainingConfig] = None
+ image: Optional[MaeImagePretrainingConfig] = None
+ text: Optional[MaskedLMConfig] = None
+
+ audio_ratio: float = 1
+ image_ratio: float = 1
+ text_ratio: float = 1
+
+ max_tokens: Optional[int] = II("dataset.max_tokens")
+ batch_size: Optional[int] = II("dataset.batch_size")
+ update_freq: List[int] = II("optimization.update_freq")
+
+ rebuild_batches: bool = True
+
+
+@register_task("multimodal_pretraining", dataclass=MultimodalPretrainingConfig)
+class MultimodalPretrainingTask(FairseqTask):
+ """ """
+
+ cfg: MultimodalPretrainingConfig
+
+ def __init__(self, cfg: MultimodalPretrainingConfig):
+ super().__init__(cfg)
+ self.audio_task = (
+ AudioPretrainingTask(cfg.audio) if cfg.audio is not None else None
+ )
+ self.image_task = (
+ MaeImagePretrainingTask(cfg.image) if cfg.image is not None else None
+ )
+ self.text_task = MaskedLMTask(cfg.text) if cfg.text is not None else None
+
+ self.mult_ratios = []
+
+ @classmethod
+ def setup_task(cls, cfg: MultimodalPretrainingConfig, **kwargs):
+ """Setup the task (e.g., load dictionaries).
+
+ Args:
+ cfg (AudioPretrainingConfig): configuration of this task
+ """
+
+ return cls(cfg)
+
+ def load_dataset(self, split: str, task_cfg: FairseqDataclass = None, **kwargs):
+ datasets = []
+ self.mult_ratios = []
+
+ def load_ds(task, name, ratio):
+ if task is not None:
+ task.load_dataset(split)
+ ds = ModalityDatasetItem(
+ datasetname=name,
+ dataset=task.dataset(split),
+ max_positions=task.max_positions(),
+ max_tokens=self.cfg.max_tokens,
+ max_sentences=self.cfg.batch_size,
+ )
+ datasets.append(ds)
+ self.mult_ratios.append(ratio)
+
+ load_ds(self.audio_task, Modality.AUDIO, self.cfg.audio_ratio)
+ load_ds(self.image_task, Modality.IMAGE, self.cfg.image_ratio)
+ load_ds(self.text_task, Modality.TEXT, self.cfg.text_ratio)
+
+ assert len(datasets) > 0
+
+ self.datasets[split] = MultiModalityDataset(datasets)
+
+ @property
+ def supported_modalities(self):
+ modalities = []
+ if self.cfg.text is not None:
+ modalities.append(Modality.TEXT)
+ if self.cfg.audio is not None:
+ modalities.append(Modality.AUDIO)
+ if self.cfg.image is not None:
+ modalities.append(Modality.IMAGE)
+
+ return modalities
+
+ def get_batch_iterator(
+ self,
+ dataset,
+ max_tokens=None,
+ max_sentences=None,
+ max_positions=None,
+ ignore_invalid_inputs=False,
+ required_batch_size_multiple=1,
+ seed=1,
+ num_shards=1,
+ shard_id=0,
+ num_workers=0,
+ epoch=0,
+ data_buffer_size=0,
+ disable_iterator_cache=False,
+ skip_remainder_batch=False,
+ grouped_shuffling=False,
+ update_epoch_batch_itr=False,
+ ):
+
+ # initialize the dataset with the correct starting epoch
+ dataset.set_epoch(epoch)
+
+ batch_samplers = dataset.get_batch_samplers(
+ self.mult_ratios, required_batch_size_multiple, seed
+ )
+
+ # return a reusable, sharded iterator
+ epoch_iter = GroupedEpochBatchIterator(
+ dataset=dataset,
+ collate_fn=dataset.collater,
+ batch_samplers=batch_samplers,
+ seed=seed,
+ num_shards=num_shards,
+ shard_id=shard_id,
+ num_workers=num_workers,
+ epoch=epoch,
+ mult_rate=max(self.cfg.update_freq),
+ buffer_size=data_buffer_size,
+ skip_remainder_batch=skip_remainder_batch,
+ )
+ self.dataset_to_epoch_iter[dataset] = {} # refresh it every epoch
+ return epoch_iter
+
+ @property
+ def source_dictionary(self):
+ return None
+
+ @property
+ def target_dictionary(self):
+ return None
+
+ def max_positions(self):
+ """Maximum input length supported by the encoder."""
+ return sys.maxsize, sys.maxsize
diff --git a/fairseq/examples/discriminative_reranking_nmt/README.md b/fairseq/examples/discriminative_reranking_nmt/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..b155e855f2f94e30ad22262f260008fda8ac1804
--- /dev/null
+++ b/fairseq/examples/discriminative_reranking_nmt/README.md
@@ -0,0 +1,202 @@
+# Discriminative Reranking for Neural Machine Translation
+https://aclanthology.org/2021.acl-long.563/
+
+This folder contains source code for training DrNMT, a discriminatively trained reranker for neural machine translation.
+
+## Data preparation
+1. Follow the instructions under `examples/translation` to build a base MT model. Prepare three files, one with source sentences, one with ground truth target sentences, and one with hypotheses generated from the base MT model. Each line in the file contains one sentence in raw text (i.e. no sentencepiece, etc.). Below is an example of the files with _N_ hypotheses for each source sentence.
+
+```
+# Example of the source sentence file: (The file should contain L lines.)
+
+source_sentence_1
+source_sentence_2
+source_sentence_3
+...
+source_sentence_L
+
+# Example of the target sentence file: (The file should contain L lines.)
+
+target_sentence_1
+target_sentence_2
+target_sentence_3
+...
+target_sentence_L
+
+# Example of the hypotheses file: (The file should contain L*N lines.)
+
+source_sentence_1_hypo_1
+source_sentence_1_hypo_2
+...
+source_sentence_1_hypo_N
+source_sentence_2_hypo_1
+...
+source_sentence_2_hypo_N
+...
+source_sentence_L_hypo_1
+...
+source_sentence_L_hypo_N
+```
+
+2. Download the [XLMR model](https://github.com/fairinternal/fairseq-py/tree/main/examples/xlmr#pre-trained-models).
+```
+wget https://dl.fbaipublicfiles.com/fairseq/models/xlmr.base.tar.gz
+tar zxvf xlmr.base.tar.gz
+
+# The folder should contain dict.txt, model.pt and sentencepiece.bpe.model.
+```
+
+3. Prepare scores and BPE data.
+* `N`: Number of hypotheses per each source sentence. We use 50 in the paper.
+* `SPLIT`: Name of the data split, i.e. train, valid, test. Use split_name, split_name1, split_name2, ..., if there are multiple datasets for a split, e.g. train, train1, valid, valid1.
+* `NUM_SHARDS`: Number of shards. Set this to 1 for non-train splits.
+* `METRIC`: The metric for DrNMT to optimize for. We support either `bleu` or `ter`.
+```
+# For each data split, e.g. train, valid, test, etc., run the following:
+
+SOURCE_FILE=/path/to/source_sentence_file
+TARGET_FILE=/path/to/target_sentence_file
+HYPO_FILE=/path/to/hypo_file
+XLMR_DIR=/path/to/xlmr
+OUTPUT_DIR=/path/to/output
+
+python scripts/prep_data.py \
+ --input-source ${SOURCE_FILE} \
+ --input-target ${TARGET_FILE} \
+ --input-hypo ${HYPO_FILE} \
+ --output-dir ${OUTPUT_DIR} \
+ --split $SPLIT
+ --beam $N \
+ --sentencepiece-model ${XLMR_DIR}/sentencepiece.bpe.model \
+ --metric $METRIC \
+ --num-shards ${NUM_SHARDS}
+
+# The script will create ${OUTPUT_DIR}/$METRIC with ${NUM_SHARDS} splits.
+# Under split*/input_src, split*/input_tgt and split*/$METRIC, there will be $SPLIT.bpe and $SPLIT.$METRIC files, respectively.
+
+```
+
+4. Pre-process the data into fairseq format.
+```
+# use comma to separate if there are more than one train or valid set
+for suffix in src tgt ; do
+ fairseq-preprocess --only-source \
+ --trainpref ${OUTPUT_DIR}/$METRIC/split1/input_${suffix}/train.bpe \
+ --validpref ${OUTPUT_DIR}/$METRIC/split1/input_${suffix}/valid.bpe \
+ --destdir ${OUTPUT_DIR}/$METRIC/split1/input_${suffix} \
+ --workers 60 \
+ --srcdict ${XLMR_DIR}/dict.txt
+done
+
+for i in `seq 2 ${NUM_SHARDS}`; do
+ for suffix in src tgt ; do
+ fairseq-preprocess --only-source \
+ --trainpref ${OUTPUT_DIR}/$METRIC/split${i}/input_${suffix}/train.bpe \
+ --destdir ${OUTPUT_DIR}/$METRIC/split${i}/input_${suffix} \
+ --workers 60 \
+ --srcdict ${XLMR_DIR}/dict.txt
+
+ ln -s ${OUTPUT_DIR}/$METRIC/split1/input_${suffix}/valid* ${OUTPUT_DIR}/$METRIC/split${i}/input_${suffix}/.
+ done
+
+ ln -s ${OUTPUT_DIR}/$METRIC/split1/$METRIC/valid* ${OUTPUT_DIR}/$METRIC/split${i}/$METRIC/.
+done
+```
+
+## Training
+
+```
+EXP_DIR=/path/to/exp
+
+# An example of training the model with the config for De-En experiment in the paper.
+# The config uses 16 GPUs and 50 hypotheses.
+# For training with fewer number of GPUs, set
+# distributed_training.distributed_world_size=k +optimization.update_freq='[x]' where x = 16/k
+# For training with fewer number of hypotheses, set
+# task.mt_beam=N dataset.batch_size=N dataset.required_batch_size_multiple=N
+
+fairseq-hydra-train -m \
+ --config-dir config/ --config-name deen \
+ task.data=${OUTPUT_DIR}/$METRIC/split1/ \
+ task.num_data_splits=${NUM_SHARDS} \
+ model.pretrained_model=${XLMR_DIR}/model.pt \
+ common.user_dir=${FAIRSEQ_ROOT}/examples/discriminative_reranking_nmt \
+ checkpoint.save_dir=${EXP_DIR}
+
+```
+
+## Inference & scoring
+Perform DrNMT reranking (fw + reranker score)
+1. Tune weights on valid sets.
+```
+# genrate N hypotheses with the base MT model (fw score)
+VALID_SOURCE_FILE=/path/to/source_sentences # one sentence per line, converted to the sentencepiece used by the base MT model
+VALID_TARGET_FILE=/path/to/target_sentences # one sentence per line in raw text, i.e. no sentencepiece and tokenization
+MT_MODEL=/path/to/mt_model
+MT_DATA_PATH=/path/to/mt_data
+
+cat ${VALID_SOURCE_FILE} | \
+ fairseq-interactive ${MT_DATA_PATH} \
+ --max-tokens 4000 --buffer-size 16 \
+ --num-workers 32 --path ${MT_MODEL} \
+ --beam $N --nbest $N \
+ --post-process sentencepiece &> valid-hypo.out
+
+# replace "bleu" with "ter" to optimize for TER
+python drnmt_rerank.py \
+ ${OUTPUT_DIR}/$METRIC/split1/ \
+ --path ${EXP_DIR}/checkpoint_best.pt \
+ --in-text valid-hypo.out \
+ --results-path ${EXP_DIR} \
+ --gen-subset valid \
+ --target-text ${VALID_TARGET_FILE} \
+ --user-dir ${FAIRSEQ_ROOT}/examples/discriminative_reranking_nmt \
+ --bpe sentencepiece \
+ --sentencepiece-model ${XLMR_DIR}/sentencepiece.bpe.model \
+ --beam $N \
+ --batch-size $N \
+ --metric bleu \
+ --tune
+
+```
+
+2. Apply best weights on test sets
+```
+# genrate N hypotheses with the base MT model (fw score)
+TEST_SOURCE_FILE=/path/to/source_sentences # one sentence per line, converted to the sentencepiece used by the base MT model
+
+cat ${TEST_SOURCE_FILE} | \
+ fairseq-interactive ${MT_DATA_PATH} \
+ --max-tokens 4000 --buffer-size 16 \
+ --num-workers 32 --path ${MT_MODEL} \
+ --beam $N --nbest $N \
+ --post-process sentencepiece &> test-hypo.out
+
+# replace "bleu" with "ter" to evaluate TER
+# Add --target-text for evaluating BLEU/TER,
+# otherwise the script will only generate the hypotheses with the highest scores only.
+python drnmt_rerank.py \
+ ${OUTPUT_DIR}/$METRIC/split1/ \
+ --path ${EXP_DIR}/checkpoint_best.pt \
+ --in-text test-hypo.out \
+ --results-path ${EXP_DIR} \
+ --gen-subset test \
+ --user-dir ${FAIRSEQ_ROOT}/examples/discriminative_reranking_nmt \
+ --bpe sentencepiece \
+ --sentencepiece-model ${XLMR_DIR}/sentencepiece.bpe.model \
+ --beam $N \
+ --batch-size $N \
+ --metric bleu \
+ --fw-weight ${BEST_FW_WEIGHT} \
+ --lenpen ${BEST_LENPEN}
+```
+
+## Citation
+```bibtex
+@inproceedings{lee2021discriminative,
+ title={Discriminative Reranking for Neural Machine Translation},
+ author={Lee, Ann and Auli, Michael and Ranzato, Marc'Aurelio},
+ booktitle={ACL},
+ year={2021}
+}
+```
diff --git a/fairseq/examples/discriminative_reranking_nmt/__init__.py b/fairseq/examples/discriminative_reranking_nmt/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..0278f6a27340c7ff7e207d09348483d1b0d3a100
--- /dev/null
+++ b/fairseq/examples/discriminative_reranking_nmt/__init__.py
@@ -0,0 +1 @@
+from . import criterions, models, tasks # noqa
diff --git a/fairseq/examples/discriminative_reranking_nmt/config/deen.yaml b/fairseq/examples/discriminative_reranking_nmt/config/deen.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..3fc2d5fcf5bacbb842d181fcfcde80e55331fed7
--- /dev/null
+++ b/fairseq/examples/discriminative_reranking_nmt/config/deen.yaml
@@ -0,0 +1,56 @@
+# @package _group_
+
+common:
+ fp16: true
+ log_format: json
+ log_interval: 50
+ seed: 2
+
+checkpoint:
+ no_epoch_checkpoints: true
+ best_checkpoint_metric: bleu
+ maximize_best_checkpoint_metric: true
+
+task:
+ _name: discriminative_reranking_nmt
+ data: ???
+ num_data_splits: ???
+ include_src: true
+ mt_beam: 50
+ eval_target_metric: true
+ target_metric: bleu
+
+dataset:
+ batch_size: 50
+ num_workers: 6
+ required_batch_size_multiple: 50
+ valid_subset: ???
+
+criterion:
+ _name: kl_divergence_rereanking
+ target_dist_norm: minmax
+ temperature: 0.5
+
+optimization:
+ max_epoch: 200
+ lr: [0.00005]
+ update_freq: [32]
+
+optimizer:
+ _name: adam
+ adam_betas: (0.9,0.98)
+ adam_eps: 1e-06
+
+lr_scheduler:
+ _name: polynomial_decay
+ warmup_updates: 8000
+ total_num_update: 320000
+
+model:
+ _name: discriminative_nmt_reranker
+ pretrained_model: ???
+ classifier_dropout: 0.2
+
+distributed_training:
+ ddp_backend: no_c10d
+ distributed_world_size: 16
diff --git a/fairseq/examples/discriminative_reranking_nmt/criterions/__init__.py b/fairseq/examples/discriminative_reranking_nmt/criterions/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..7c257c2700f015cb123a976584aef72f0429eb0c
--- /dev/null
+++ b/fairseq/examples/discriminative_reranking_nmt/criterions/__init__.py
@@ -0,0 +1,6 @@
+from .discriminative_reranking_criterion import KLDivergenceRerankingCriterion
+
+
+__all__ = [
+ "KLDivergenceRerankingCriterion",
+]
diff --git a/fairseq/examples/discriminative_reranking_nmt/criterions/discriminative_reranking_criterion.py b/fairseq/examples/discriminative_reranking_nmt/criterions/discriminative_reranking_criterion.py
new file mode 100644
index 0000000000000000000000000000000000000000..c8f19e3858356aff5d1c47f9d6565f788578f252
--- /dev/null
+++ b/fairseq/examples/discriminative_reranking_nmt/criterions/discriminative_reranking_criterion.py
@@ -0,0 +1,139 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import math
+from dataclasses import dataclass, field
+
+import torch
+import torch.nn.functional as F
+
+from fairseq import utils
+from fairseq.logging import metrics
+from fairseq.criterions import FairseqCriterion, register_criterion
+from fairseq.dataclass import ChoiceEnum, FairseqDataclass
+
+
+_EPSILON = torch.finfo(torch.float32).eps
+TARGET_DIST_NORM_CHOICES = ChoiceEnum(["none", "minmax"])
+
+
+@dataclass
+class KLDivergenceRerankingCriterionConfig(FairseqDataclass):
+ target_dist_norm: TARGET_DIST_NORM_CHOICES = field(
+ default="none",
+ metadata={"help": "method to normalize the range of target scores"},
+ )
+ temperature: float = field(
+ default=1.0,
+ metadata={"help": "temperature in softmax for target distributions"},
+ )
+ forward_batch_size: int = field(
+ default=32,
+ metadata={
+ "help": "number of hypotheses per batch for model forward (set a value smaller than --mt-beam to avoid OOM when training with a large beam size)"
+ },
+ )
+
+
+@register_criterion(
+ "kl_divergence_rereanking", dataclass=KLDivergenceRerankingCriterionConfig
+)
+class KLDivergenceRerankingCriterion(FairseqCriterion):
+ def __init__(
+ self, task, target_dist_norm, temperature, forward_batch_size,
+ ):
+ super().__init__(task)
+ self.target_dist_norm = target_dist_norm
+ self.temperature = temperature
+ self.forward_batch_size = forward_batch_size
+
+ def forward(self, model, sample, reduce=True):
+ """Compute the loss for the given sample.
+
+ Returns a tuple with three elements:
+ 1) the loss
+ 2) the sample size, which is used as the denominator for the gradient
+ 3) logging outputs to display while training
+ """
+
+ sample_size = sample["id"].numel()
+ assert sample_size % self.task.cfg.mt_beam == 0, (
+ f"sample_size ({sample_size}) cannot be divided by beam size ({self.task.cfg.mt_beam})."
+ f"Please set --required-batch-size-multiple={self.task.cfg.mt_beam}."
+ )
+
+ # split into smaller batches for model forward
+ batch_out = []
+ for i in range(0, sample_size, self.forward_batch_size):
+ j = min(i + self.forward_batch_size, sample_size)
+
+ out = model(
+ src_tokens=sample["net_input"]["src_tokens"][i:j, :],
+ src_lengths=sample["net_input"]["src_lengths"][i:j],
+ )
+
+ batch_out.append(
+ model.sentence_forward(out, sample["net_input"]["src_tokens"][i:j, :])
+ )
+
+ batch_out = torch.cat(batch_out, dim=0).view(
+ self.task.cfg.mt_beam, sample_size // self.task.cfg.mt_beam, -1
+ ) # T x B x C
+ if model.joint_classification == "sent":
+ batch_out = model.joint_forward(batch_out)
+ scores = model.classification_forward(batch_out.view(sample_size, 1, -1)).view(
+ -1, self.task.cfg.mt_beam
+ ) # input: B x T x C
+
+ loss = self.compute_kl_loss(
+ scores, sample["target"][:, 0].view(-1, self.task.cfg.mt_beam)
+ )
+
+ sample_size = sample_size // self.task.cfg.mt_beam
+
+ logging_output = {
+ "loss": loss.detach(),
+ "ntokens": sample["ntokens"],
+ "nsentences": sample_size * self.task.cfg.mt_beam,
+ "sample_size": sample_size,
+ "scores": scores.detach(),
+ }
+
+ return loss, sample_size, logging_output
+
+ def compute_kl_loss(self, logits, target):
+ norm_target = target
+ if self.target_dist_norm == "minmax":
+ min_v = torch.min(target, 1, keepdim=True).values
+ max_v = torch.max(target, 1, keepdim=True).values
+ norm_target = (target - min_v) / (max_v - min_v + _EPSILON)
+
+ target_dist = F.softmax(
+ norm_target / self.temperature, dim=-1, dtype=torch.float32
+ )
+ model_dist = F.log_softmax(logits, dim=-1, dtype=torch.float32)
+ loss = -(target_dist * model_dist - target_dist * target_dist.log()).sum()
+ return loss
+
+ @staticmethod
+ def reduce_metrics(logging_outputs) -> None:
+ """Aggregate logging outputs from data parallel training."""
+ loss_sum = utils.item(sum(log.get("loss", 0) for log in logging_outputs))
+
+ sample_size = utils.item(
+ sum(log.get("sample_size", 0) for log in logging_outputs)
+ )
+
+ loss = loss_sum / sample_size / math.log(2)
+ metrics.log_scalar("loss", loss, sample_size, round=3)
+
+ @staticmethod
+ def logging_outputs_can_be_summed() -> bool:
+ """
+ Whether the logging outputs returned by `forward` can be summed
+ across workers prior to calling `reduce_metrics`. Setting this
+ to True will improves distributed training speed.
+ """
+ return True
diff --git a/fairseq/examples/discriminative_reranking_nmt/drnmt_rerank.py b/fairseq/examples/discriminative_reranking_nmt/drnmt_rerank.py
new file mode 100644
index 0000000000000000000000000000000000000000..2e0fc2bd29aedb0b477b7cc8e2c3b606acdd454a
--- /dev/null
+++ b/fairseq/examples/discriminative_reranking_nmt/drnmt_rerank.py
@@ -0,0 +1,364 @@
+#!/usr/bin/env python3 -u
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+"""
+Score raw text with a trained model.
+"""
+
+from collections import namedtuple
+import logging
+from multiprocessing import Pool
+import sys
+import os
+import random
+
+import numpy as np
+import sacrebleu
+import torch
+
+from fairseq import checkpoint_utils, options, utils
+
+
+logger = logging.getLogger("fairseq_cli.drnmt_rerank")
+logger.setLevel(logging.INFO)
+
+Batch = namedtuple("Batch", "ids src_tokens src_lengths")
+
+
+pool_init_variables = {}
+
+
+def init_loaded_scores(mt_scores, model_scores, hyp, ref):
+ global pool_init_variables
+ pool_init_variables["mt_scores"] = mt_scores
+ pool_init_variables["model_scores"] = model_scores
+ pool_init_variables["hyp"] = hyp
+ pool_init_variables["ref"] = ref
+
+
+def parse_fairseq_gen(filename, task):
+ source = {}
+ hypos = {}
+ scores = {}
+ with open(filename, "r", encoding="utf-8") as f:
+ for line in f:
+ line = line.strip()
+ if line.startswith("S-"): # source
+ uid, text = line.split("\t", 1)
+ uid = int(uid[2:])
+ source[uid] = text
+ elif line.startswith("D-"): # hypo
+ uid, score, text = line.split("\t", 2)
+ uid = int(uid[2:])
+ if uid not in hypos:
+ hypos[uid] = []
+ scores[uid] = []
+ hypos[uid].append(text)
+ scores[uid].append(float(score))
+ else:
+ continue
+
+ source_out = [source[i] for i in range(len(hypos))]
+ hypos_out = [h for i in range(len(hypos)) for h in hypos[i]]
+ scores_out = [s for i in range(len(scores)) for s in scores[i]]
+
+ return source_out, hypos_out, scores_out
+
+
+def read_target(filename):
+ with open(filename, "r", encoding="utf-8") as f:
+ output = [line.strip() for line in f]
+ return output
+
+
+def make_batches(args, src, hyp, task, max_positions, encode_fn):
+ assert len(src) * args.beam == len(
+ hyp
+ ), f"Expect {len(src) * args.beam} hypotheses for {len(src)} source sentences with beam size {args.beam}. Got {len(hyp)} hypotheses intead."
+ hyp_encode = [
+ task.source_dictionary.encode_line(encode_fn(h), add_if_not_exist=False).long()
+ for h in hyp
+ ]
+ if task.cfg.include_src:
+ src_encode = [
+ task.source_dictionary.encode_line(
+ encode_fn(s), add_if_not_exist=False
+ ).long()
+ for s in src
+ ]
+ tokens = [(src_encode[i // args.beam], h) for i, h in enumerate(hyp_encode)]
+ lengths = [(t1.numel(), t2.numel()) for t1, t2 in tokens]
+ else:
+ tokens = [(h,) for h in hyp_encode]
+ lengths = [(h.numel(),) for h in hyp_encode]
+
+ itr = task.get_batch_iterator(
+ dataset=task.build_dataset_for_inference(tokens, lengths),
+ max_tokens=args.max_tokens,
+ max_sentences=args.batch_size,
+ max_positions=max_positions,
+ ignore_invalid_inputs=args.skip_invalid_size_inputs_valid_test,
+ ).next_epoch_itr(shuffle=False)
+
+ for batch in itr:
+ yield Batch(
+ ids=batch["id"],
+ src_tokens=batch["net_input"]["src_tokens"],
+ src_lengths=batch["net_input"]["src_lengths"],
+ )
+
+
+def decode_rerank_scores(args):
+ if args.max_tokens is None and args.batch_size is None:
+ args.batch_size = 1
+
+ logger.info(args)
+
+ use_cuda = torch.cuda.is_available() and not args.cpu
+
+ # Load ensemble
+ logger.info("loading model(s) from {}".format(args.path))
+ models, _model_args, task = checkpoint_utils.load_model_ensemble_and_task(
+ [args.path], arg_overrides=eval(args.model_overrides),
+ )
+
+ for model in models:
+ if args.fp16:
+ model.half()
+ if use_cuda:
+ model.cuda()
+
+ # Initialize generator
+ generator = task.build_generator(args)
+
+ # Handle tokenization and BPE
+ tokenizer = task.build_tokenizer(args)
+ bpe = task.build_bpe(args)
+
+ def encode_fn(x):
+ if tokenizer is not None:
+ x = tokenizer.encode(x)
+ if bpe is not None:
+ x = bpe.encode(x)
+ return x
+
+ max_positions = utils.resolve_max_positions(
+ task.max_positions(), *[model.max_positions() for model in models]
+ )
+
+ src, hyp, mt_scores = parse_fairseq_gen(args.in_text, task)
+ model_scores = {}
+ logger.info("decode reranker score")
+ for batch in make_batches(args, src, hyp, task, max_positions, encode_fn):
+ src_tokens = batch.src_tokens
+ src_lengths = batch.src_lengths
+ if use_cuda:
+ src_tokens = src_tokens.cuda()
+ src_lengths = src_lengths.cuda()
+
+ sample = {
+ "net_input": {"src_tokens": src_tokens, "src_lengths": src_lengths},
+ }
+ scores = task.inference_step(generator, models, sample)
+
+ for id, sc in zip(batch.ids.tolist(), scores.tolist()):
+ model_scores[id] = sc[0]
+
+ model_scores = [model_scores[i] for i in range(len(model_scores))]
+
+ return src, hyp, mt_scores, model_scores
+
+
+def get_score(mt_s, md_s, w1, lp, tgt_len):
+ return mt_s / (tgt_len ** lp) * w1 + md_s
+
+
+def get_best_hyps(mt_scores, md_scores, hypos, fw_weight, lenpen, beam):
+ assert len(mt_scores) == len(md_scores) and len(mt_scores) == len(hypos)
+ hypo_scores = []
+ best_hypos = []
+ best_scores = []
+ offset = 0
+ for i in range(len(hypos)):
+ tgt_len = len(hypos[i].split())
+ hypo_scores.append(
+ get_score(mt_scores[i], md_scores[i], fw_weight, lenpen, tgt_len)
+ )
+
+ if (i + 1) % beam == 0:
+ max_i = np.argmax(hypo_scores)
+ best_hypos.append(hypos[offset + max_i])
+ best_scores.append(hypo_scores[max_i])
+ hypo_scores = []
+ offset += beam
+ return best_hypos, best_scores
+
+
+def eval_metric(args, hypos, ref):
+ if args.metric == "bleu":
+ score = sacrebleu.corpus_bleu(hypos, [ref]).score
+ else:
+ score = sacrebleu.corpus_ter(hypos, [ref]).score
+
+ return score
+
+
+def score_target_hypo(args, fw_weight, lp):
+ mt_scores = pool_init_variables["mt_scores"]
+ model_scores = pool_init_variables["model_scores"]
+ hyp = pool_init_variables["hyp"]
+ ref = pool_init_variables["ref"]
+ best_hypos, _ = get_best_hyps(
+ mt_scores, model_scores, hyp, fw_weight, lp, args.beam
+ )
+ rerank_eval = None
+ if ref:
+ rerank_eval = eval_metric(args, best_hypos, ref)
+ print(f"fw_weight {fw_weight}, lenpen {lp}, eval {rerank_eval}")
+
+ return rerank_eval
+
+
+def print_result(best_scores, best_hypos, output_file):
+ for i, (s, h) in enumerate(zip(best_scores, best_hypos)):
+ print(f"{i}\t{s}\t{h}", file=output_file)
+
+
+def main(args):
+ utils.import_user_module(args)
+
+ src, hyp, mt_scores, model_scores = decode_rerank_scores(args)
+
+ assert (
+ not args.tune or args.target_text is not None
+ ), "--target-text has to be set when tuning weights"
+ if args.target_text:
+ ref = read_target(args.target_text)
+ assert len(src) == len(
+ ref
+ ), f"different numbers of source and target sentences ({len(src)} vs. {len(ref)})"
+
+ orig_best_hypos = [hyp[i] for i in range(0, len(hyp), args.beam)]
+ orig_eval = eval_metric(args, orig_best_hypos, ref)
+
+ if args.tune:
+ logger.info("tune weights for reranking")
+
+ random_params = np.array(
+ [
+ [
+ random.uniform(
+ args.lower_bound_fw_weight, args.upper_bound_fw_weight
+ ),
+ random.uniform(args.lower_bound_lenpen, args.upper_bound_lenpen),
+ ]
+ for k in range(args.num_trials)
+ ]
+ )
+
+ logger.info("launching pool")
+ with Pool(
+ 32,
+ initializer=init_loaded_scores,
+ initargs=(mt_scores, model_scores, hyp, ref),
+ ) as p:
+ rerank_scores = p.starmap(
+ score_target_hypo,
+ [
+ (args, random_params[i][0], random_params[i][1],)
+ for i in range(args.num_trials)
+ ],
+ )
+ if args.metric == "bleu":
+ best_index = np.argmax(rerank_scores)
+ else:
+ best_index = np.argmin(rerank_scores)
+ best_fw_weight = random_params[best_index][0]
+ best_lenpen = random_params[best_index][1]
+ else:
+ assert (
+ args.lenpen is not None and args.fw_weight is not None
+ ), "--lenpen and --fw-weight should be set"
+ best_fw_weight, best_lenpen = args.fw_weight, args.lenpen
+
+ best_hypos, best_scores = get_best_hyps(
+ mt_scores, model_scores, hyp, best_fw_weight, best_lenpen, args.beam
+ )
+
+ if args.results_path is not None:
+ os.makedirs(args.results_path, exist_ok=True)
+ output_path = os.path.join(
+ args.results_path, "generate-{}.txt".format(args.gen_subset),
+ )
+ with open(output_path, "w", buffering=1, encoding="utf-8") as o:
+ print_result(best_scores, best_hypos, o)
+ else:
+ print_result(best_scores, best_hypos, sys.stdout)
+
+ if args.target_text:
+ rerank_eval = eval_metric(args, best_hypos, ref)
+ print(f"before reranking, {args.metric.upper()}:", orig_eval)
+ print(
+ f"after reranking with fw_weight={best_fw_weight}, lenpen={best_lenpen}, {args.metric.upper()}:",
+ rerank_eval,
+ )
+
+
+def cli_main():
+ parser = options.get_generation_parser(interactive=True)
+
+ parser.add_argument(
+ "--in-text",
+ default=None,
+ required=True,
+ help="text from fairseq-interactive output, containing source sentences and hypotheses",
+ )
+ parser.add_argument("--target-text", default=None, help="reference text")
+ parser.add_argument("--metric", type=str, choices=["bleu", "ter"], default="bleu")
+ parser.add_argument(
+ "--tune",
+ action="store_true",
+ help="if set, tune weights on fw scores and lenpen instead of applying fixed weights for reranking",
+ )
+ parser.add_argument(
+ "--lower-bound-fw-weight",
+ default=0.0,
+ type=float,
+ help="lower bound of search space",
+ )
+ parser.add_argument(
+ "--upper-bound-fw-weight",
+ default=3,
+ type=float,
+ help="upper bound of search space",
+ )
+ parser.add_argument(
+ "--lower-bound-lenpen",
+ default=0.0,
+ type=float,
+ help="lower bound of search space",
+ )
+ parser.add_argument(
+ "--upper-bound-lenpen",
+ default=3,
+ type=float,
+ help="upper bound of search space",
+ )
+ parser.add_argument(
+ "--fw-weight", type=float, default=None, help="weight on the fw model score"
+ )
+ parser.add_argument(
+ "--num-trials",
+ default=1000,
+ type=int,
+ help="number of trials to do for random search",
+ )
+
+ args = options.parse_args_and_arch(parser)
+ main(args)
+
+
+if __name__ == "__main__":
+ cli_main()
diff --git a/fairseq/examples/discriminative_reranking_nmt/models/__init__.py b/fairseq/examples/discriminative_reranking_nmt/models/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..c593ea5f1842794bfcc952fc93c679a5f16aeb98
--- /dev/null
+++ b/fairseq/examples/discriminative_reranking_nmt/models/__init__.py
@@ -0,0 +1,6 @@
+from .discriminative_reranking_model import DiscriminativeNMTReranker
+
+
+__all__ = [
+ "DiscriminativeNMTReranker",
+]
diff --git a/fairseq/examples/discriminative_reranking_nmt/models/discriminative_reranking_model.py b/fairseq/examples/discriminative_reranking_nmt/models/discriminative_reranking_model.py
new file mode 100644
index 0000000000000000000000000000000000000000..e4b5887f825df36f4e1e0384f38fefe790e485e6
--- /dev/null
+++ b/fairseq/examples/discriminative_reranking_nmt/models/discriminative_reranking_model.py
@@ -0,0 +1,365 @@
+from dataclasses import dataclass, field
+import os
+
+import torch
+import torch.nn as nn
+
+from fairseq import utils
+from fairseq.dataclass import ChoiceEnum, FairseqDataclass
+from fairseq.models import (
+ BaseFairseqModel,
+ register_model,
+)
+
+from fairseq.models.roberta.model import RobertaClassificationHead
+
+from fairseq.modules import (
+ LayerNorm,
+ TransformerSentenceEncoder,
+ TransformerSentenceEncoderLayer,
+)
+
+
+ACTIVATION_FN_CHOICES = ChoiceEnum(utils.get_available_activation_fns())
+JOINT_CLASSIFICATION_CHOICES = ChoiceEnum(["none", "sent"])
+SENTENCE_REP_CHOICES = ChoiceEnum(["head", "meanpool", "maxpool"])
+
+
+def update_init_roberta_model_state(state):
+ """
+ update the state_dict of a Roberta model for initializing
+ weights of the BertRanker
+ """
+ for k in list(state.keys()):
+ if ".lm_head." in k or "version" in k:
+ del state[k]
+ continue
+ # remove 'encoder/decoder.sentence_encoder.' from the key
+ assert k.startswith("encoder.sentence_encoder.") or k.startswith(
+ "decoder.sentence_encoder."
+ ), f"Cannot recognize parameter name {k}"
+ if "layernorm_embedding" in k:
+ new_k = k.replace(".layernorm_embedding.", ".emb_layer_norm.")
+ state[new_k[25:]] = state[k]
+ else:
+ state[k[25:]] = state[k]
+ del state[k]
+
+
+class BaseRanker(nn.Module):
+ def __init__(self, args, task):
+ super().__init__()
+
+ self.separator_token = task.dictionary.eos()
+ self.padding_idx = task.dictionary.pad()
+
+ def forward(self, src_tokens):
+ raise NotImplementedError
+
+ def get_segment_labels(self, src_tokens):
+ segment_boundary = (src_tokens == self.separator_token).long()
+ segment_labels = (
+ segment_boundary.cumsum(dim=1)
+ - segment_boundary
+ - (src_tokens == self.padding_idx).long()
+ )
+
+ return segment_labels
+
+ def get_positions(self, src_tokens, segment_labels):
+ segment_positions = (
+ torch.arange(src_tokens.shape[1])
+ .to(src_tokens.device)
+ .repeat(src_tokens.shape[0], 1)
+ )
+ segment_boundary = (src_tokens == self.separator_token).long()
+ _, col_idx = (segment_positions * segment_boundary).nonzero(as_tuple=True)
+ col_idx = torch.cat([torch.zeros(1).type_as(col_idx), col_idx])
+ offset = torch.cat(
+ [
+ torch.zeros(1).type_as(segment_boundary),
+ segment_boundary.sum(dim=1).cumsum(dim=0)[:-1],
+ ]
+ )
+ segment_positions -= col_idx[segment_labels + offset.unsqueeze(1)] * (
+ segment_labels != 0
+ )
+
+ padding_mask = src_tokens.ne(self.padding_idx)
+ segment_positions = (segment_positions + 1) * padding_mask.type_as(
+ segment_positions
+ ) + self.padding_idx
+
+ return segment_positions
+
+
+class BertRanker(BaseRanker):
+ def __init__(self, args, task):
+ super(BertRanker, self).__init__(args, task)
+
+ init_model = getattr(args, "pretrained_model", "")
+ self.joint_layers = nn.ModuleList()
+ if os.path.isfile(init_model):
+ print(f"initialize weight from {init_model}")
+
+ from fairseq import hub_utils
+
+ x = hub_utils.from_pretrained(
+ os.path.dirname(init_model),
+ checkpoint_file=os.path.basename(init_model),
+ )
+
+ in_state_dict = x["models"][0].state_dict()
+ init_args = x["args"].model
+
+ num_positional_emb = init_args.max_positions + task.dictionary.pad() + 1
+
+ # follow the setup in roberta
+ self.model = TransformerSentenceEncoder(
+ padding_idx=task.dictionary.pad(),
+ vocab_size=len(task.dictionary),
+ num_encoder_layers=getattr(
+ args, "encoder_layers", init_args.encoder_layers
+ ),
+ embedding_dim=init_args.encoder_embed_dim,
+ ffn_embedding_dim=init_args.encoder_ffn_embed_dim,
+ num_attention_heads=init_args.encoder_attention_heads,
+ dropout=init_args.dropout,
+ attention_dropout=init_args.attention_dropout,
+ activation_dropout=init_args.activation_dropout,
+ num_segments=2, # add language embeddings
+ max_seq_len=num_positional_emb,
+ offset_positions_by_padding=False,
+ encoder_normalize_before=True,
+ apply_bert_init=True,
+ activation_fn=init_args.activation_fn,
+ freeze_embeddings=args.freeze_embeddings,
+ n_trans_layers_to_freeze=args.n_trans_layers_to_freeze,
+ )
+
+ # still need to learn segment embeddings as we added a second language embedding
+ if args.freeze_embeddings:
+ for p in self.model.segment_embeddings.parameters():
+ p.requires_grad = False
+
+ update_init_roberta_model_state(in_state_dict)
+ print("loading weights from the pretrained model")
+ self.model.load_state_dict(
+ in_state_dict, strict=False
+ ) # ignore mismatch in language embeddings
+
+ ffn_embedding_dim = init_args.encoder_ffn_embed_dim
+ num_attention_heads = init_args.encoder_attention_heads
+ dropout = init_args.dropout
+ attention_dropout = init_args.attention_dropout
+ activation_dropout = init_args.activation_dropout
+ activation_fn = init_args.activation_fn
+
+ classifier_embed_dim = getattr(
+ args, "embed_dim", init_args.encoder_embed_dim
+ )
+ if classifier_embed_dim != init_args.encoder_embed_dim:
+ self.transform_layer = nn.Linear(
+ init_args.encoder_embed_dim, classifier_embed_dim
+ )
+ else:
+ self.model = TransformerSentenceEncoder(
+ padding_idx=task.dictionary.pad(),
+ vocab_size=len(task.dictionary),
+ num_encoder_layers=args.encoder_layers,
+ embedding_dim=args.embed_dim,
+ ffn_embedding_dim=args.ffn_embed_dim,
+ num_attention_heads=args.attention_heads,
+ dropout=args.dropout,
+ attention_dropout=args.attention_dropout,
+ activation_dropout=args.activation_dropout,
+ max_seq_len=task.max_positions()
+ if task.max_positions()
+ else args.tokens_per_sample,
+ num_segments=2,
+ offset_positions_by_padding=False,
+ encoder_normalize_before=args.encoder_normalize_before,
+ apply_bert_init=args.apply_bert_init,
+ activation_fn=args.activation_fn,
+ )
+
+ classifier_embed_dim = args.embed_dim
+ ffn_embedding_dim = args.ffn_embed_dim
+ num_attention_heads = args.attention_heads
+ dropout = args.dropout
+ attention_dropout = args.attention_dropout
+ activation_dropout = args.activation_dropout
+ activation_fn = args.activation_fn
+
+ self.joint_classification = args.joint_classification
+ if args.joint_classification == "sent":
+ if args.joint_normalize_before:
+ self.joint_layer_norm = LayerNorm(classifier_embed_dim)
+ else:
+ self.joint_layer_norm = None
+
+ self.joint_layers = nn.ModuleList(
+ [
+ TransformerSentenceEncoderLayer(
+ embedding_dim=classifier_embed_dim,
+ ffn_embedding_dim=ffn_embedding_dim,
+ num_attention_heads=num_attention_heads,
+ dropout=dropout,
+ attention_dropout=attention_dropout,
+ activation_dropout=activation_dropout,
+ activation_fn=activation_fn,
+ )
+ for _ in range(args.num_joint_layers)
+ ]
+ )
+
+ self.classifier = RobertaClassificationHead(
+ classifier_embed_dim,
+ classifier_embed_dim,
+ 1, # num_classes
+ "tanh",
+ args.classifier_dropout,
+ )
+
+ def forward(self, src_tokens, src_lengths):
+ segment_labels = self.get_segment_labels(src_tokens)
+ positions = self.get_positions(src_tokens, segment_labels)
+
+ inner_states, _ = self.model(
+ tokens=src_tokens,
+ segment_labels=segment_labels,
+ last_state_only=True,
+ positions=positions,
+ )
+
+ return inner_states[-1].transpose(0, 1) # T x B x C -> B x T x C
+
+ def sentence_forward(self, encoder_out, src_tokens=None, sentence_rep="head"):
+ # encoder_out: B x T x C
+ if sentence_rep == "head":
+ x = encoder_out[:, :1, :]
+ else: # 'meanpool', 'maxpool'
+ assert src_tokens is not None, "meanpool requires src_tokens input"
+ segment_labels = self.get_segment_labels(src_tokens)
+ padding_mask = src_tokens.ne(self.padding_idx)
+ encoder_mask = segment_labels * padding_mask.type_as(segment_labels)
+
+ if sentence_rep == "meanpool":
+ ntokens = torch.sum(encoder_mask, dim=1, keepdim=True)
+ x = torch.sum(
+ encoder_out * encoder_mask.unsqueeze(2), dim=1, keepdim=True
+ ) / ntokens.unsqueeze(2).type_as(encoder_out)
+ else: # 'maxpool'
+ encoder_out[
+ (encoder_mask == 0).unsqueeze(2).repeat(1, 1, encoder_out.shape[-1])
+ ] = -float("inf")
+ x, _ = torch.max(encoder_out, dim=1, keepdim=True)
+
+ if hasattr(self, "transform_layer"):
+ x = self.transform_layer(x)
+
+ return x # B x 1 x C
+
+ def joint_forward(self, x):
+ # x: T x B x C
+ if self.joint_layer_norm:
+ x = self.joint_layer_norm(x.transpose(0, 1))
+ x = x.transpose(0, 1)
+
+ for layer in self.joint_layers:
+ x, _ = layer(x, self_attn_padding_mask=None)
+ return x
+
+ def classification_forward(self, x):
+ # x: B x T x C
+ return self.classifier(x)
+
+
+@dataclass
+class DiscriminativeNMTRerankerConfig(FairseqDataclass):
+ pretrained_model: str = field(
+ default="", metadata={"help": "pretrained model to load"}
+ )
+ sentence_rep: SENTENCE_REP_CHOICES = field(
+ default="head",
+ metadata={
+ "help": "method to transform the output of the transformer stack to a sentence-level representation"
+ },
+ )
+
+ dropout: float = field(default=0.1, metadata={"help": "dropout probability"})
+ attention_dropout: float = field(
+ default=0.0, metadata={"help": "dropout probability for attention weights"}
+ )
+ activation_dropout: float = field(
+ default=0.0, metadata={"help": "dropout probability after activation in FFN"}
+ )
+ classifier_dropout: float = field(
+ default=0.0, metadata={"help": "classifier dropout probability"}
+ )
+ embed_dim: int = field(default=768, metadata={"help": "embedding dimension"})
+ ffn_embed_dim: int = field(
+ default=2048, metadata={"help": "embedding dimension for FFN"}
+ )
+ encoder_layers: int = field(default=12, metadata={"help": "num encoder layers"})
+ attention_heads: int = field(default=8, metadata={"help": "num attention heads"})
+ encoder_normalize_before: bool = field(
+ default=False, metadata={"help": "apply layernorm before each encoder block"}
+ )
+ apply_bert_init: bool = field(
+ default=False, metadata={"help": "use custom param initialization for BERT"}
+ )
+ activation_fn: ACTIVATION_FN_CHOICES = field(
+ default="relu", metadata={"help": "activation function to use"}
+ )
+ freeze_embeddings: bool = field(
+ default=False, metadata={"help": "freeze embeddings in the pretrained model"}
+ )
+ n_trans_layers_to_freeze: int = field(
+ default=0,
+ metadata={
+ "help": "number of layers to freeze in the pretrained transformer model"
+ },
+ )
+
+ # joint classfication
+ joint_classification: JOINT_CLASSIFICATION_CHOICES = field(
+ default="none",
+ metadata={"help": "method to compute joint features for classification"},
+ )
+ num_joint_layers: int = field(
+ default=1, metadata={"help": "number of joint layers"}
+ )
+ joint_normalize_before: bool = field(
+ default=False,
+ metadata={"help": "apply layer norm on the input to the joint layer"},
+ )
+
+
+@register_model(
+ "discriminative_nmt_reranker", dataclass=DiscriminativeNMTRerankerConfig
+)
+class DiscriminativeNMTReranker(BaseFairseqModel):
+ @classmethod
+ def build_model(cls, args, task):
+ model = BertRanker(args, task)
+ return DiscriminativeNMTReranker(args, model)
+
+ def __init__(self, args, model):
+ super().__init__()
+
+ self.model = model
+ self.sentence_rep = args.sentence_rep
+ self.joint_classification = args.joint_classification
+
+ def forward(self, src_tokens, src_lengths, **kwargs):
+ return self.model(src_tokens, src_lengths)
+
+ def sentence_forward(self, encoder_out, src_tokens):
+ return self.model.sentence_forward(encoder_out, src_tokens, self.sentence_rep)
+
+ def joint_forward(self, x):
+ return self.model.joint_forward(x)
+
+ def classification_forward(self, x):
+ return self.model.classification_forward(x)
diff --git a/fairseq/examples/discriminative_reranking_nmt/scripts/prep_data.py b/fairseq/examples/discriminative_reranking_nmt/scripts/prep_data.py
new file mode 100644
index 0000000000000000000000000000000000000000..7aa7d37edc2c3e4c1d293911b753abf2ef597a7e
--- /dev/null
+++ b/fairseq/examples/discriminative_reranking_nmt/scripts/prep_data.py
@@ -0,0 +1,136 @@
+#!/usr/bin/env python
+
+import argparse
+from multiprocessing import Pool
+from pathlib import Path
+
+import sacrebleu
+import sentencepiece as spm
+
+
+def read_text_file(filename):
+ with open(filename, "r") as f:
+ output = [line.strip() for line in f]
+
+ return output
+
+
+def get_bleu(in_sent, target_sent):
+ bleu = sacrebleu.corpus_bleu([in_sent], [[target_sent]])
+ out = " ".join(
+ map(str, [bleu.score, bleu.sys_len, bleu.ref_len] + bleu.counts + bleu.totals)
+ )
+ return out
+
+
+def get_ter(in_sent, target_sent):
+ ter = sacrebleu.corpus_ter([in_sent], [[target_sent]])
+ out = " ".join(map(str, [ter.score, ter.num_edits, ter.ref_length]))
+ return out
+
+
+def init(sp_model):
+ global sp
+ sp = spm.SentencePieceProcessor()
+ sp.Load(sp_model)
+
+
+def process(source_sent, target_sent, hypo_sent, metric):
+ source_bpe = " ".join(sp.EncodeAsPieces(source_sent))
+ hypo_bpe = [" ".join(sp.EncodeAsPieces(h)) for h in hypo_sent]
+
+ if metric == "bleu":
+ score_str = [get_bleu(h, target_sent) for h in hypo_sent]
+ else: # ter
+ score_str = [get_ter(h, target_sent) for h in hypo_sent]
+
+ return source_bpe, hypo_bpe, score_str
+
+
+def main(args):
+ assert (
+ args.split.startswith("train") or args.num_shards == 1
+ ), "--num-shards should be set to 1 for valid and test sets"
+ assert (
+ args.split.startswith("train")
+ or args.split.startswith("valid")
+ or args.split.startswith("test")
+ ), "--split should be set to train[n]/valid[n]/test[n]"
+
+ source_sents = read_text_file(args.input_source)
+ target_sents = read_text_file(args.input_target)
+
+ num_sents = len(source_sents)
+ assert num_sents == len(
+ target_sents
+ ), f"{args.input_source} and {args.input_target} should have the same number of sentences."
+
+ hypo_sents = read_text_file(args.input_hypo)
+ assert (
+ len(hypo_sents) % args.beam == 0
+ ), f"Number of hypotheses ({len(hypo_sents)}) cannot be divided by beam size ({args.beam})."
+
+ hypo_sents = [
+ hypo_sents[i : i + args.beam] for i in range(0, len(hypo_sents), args.beam)
+ ]
+ assert num_sents == len(
+ hypo_sents
+ ), f"{args.input_hypo} should contain {num_sents * args.beam} hypotheses but only has {len(hypo_sents) * args.beam}. (--beam={args.beam})"
+
+ output_dir = args.output_dir / args.metric
+ for ns in range(args.num_shards):
+ print(f"processing shard {ns+1}/{args.num_shards}")
+ shard_output_dir = output_dir / f"split{ns+1}"
+ source_output_dir = shard_output_dir / "input_src"
+ hypo_output_dir = shard_output_dir / "input_tgt"
+ metric_output_dir = shard_output_dir / args.metric
+
+ source_output_dir.mkdir(parents=True, exist_ok=True)
+ hypo_output_dir.mkdir(parents=True, exist_ok=True)
+ metric_output_dir.mkdir(parents=True, exist_ok=True)
+
+ if args.n_proc > 1:
+ with Pool(
+ args.n_proc, initializer=init, initargs=(args.sentencepiece_model,)
+ ) as p:
+ output = p.starmap(
+ process,
+ [
+ (source_sents[i], target_sents[i], hypo_sents[i], args.metric)
+ for i in range(ns, num_sents, args.num_shards)
+ ],
+ )
+ else:
+ init(args.sentencepiece_model)
+ output = [
+ process(source_sents[i], target_sents[i], hypo_sents[i], args.metric)
+ for i in range(ns, num_sents, args.num_shards)
+ ]
+
+ with open(source_output_dir / f"{args.split}.bpe", "w") as s_o, open(
+ hypo_output_dir / f"{args.split}.bpe", "w"
+ ) as h_o, open(metric_output_dir / f"{args.split}.{args.metric}", "w") as m_o:
+ for source_bpe, hypo_bpe, score_str in output:
+ assert len(hypo_bpe) == len(score_str)
+ for h, m in zip(hypo_bpe, score_str):
+ s_o.write(f"{source_bpe}\n")
+ h_o.write(f"{h}\n")
+ m_o.write(f"{m}\n")
+
+
+if __name__ == "__main__":
+ parser = argparse.ArgumentParser()
+ parser.add_argument("--input-source", type=Path, required=True)
+ parser.add_argument("--input-target", type=Path, required=True)
+ parser.add_argument("--input-hypo", type=Path, required=True)
+ parser.add_argument("--output-dir", type=Path, required=True)
+ parser.add_argument("--split", type=str, required=True)
+ parser.add_argument("--beam", type=int, required=True)
+ parser.add_argument("--sentencepiece-model", type=str, required=True)
+ parser.add_argument("--metric", type=str, choices=["bleu", "ter"], default="bleu")
+ parser.add_argument("--num-shards", type=int, default=1)
+ parser.add_argument("--n-proc", type=int, default=8)
+
+ args = parser.parse_args()
+
+ main(args)
diff --git a/fairseq/examples/discriminative_reranking_nmt/tasks/__init__.py b/fairseq/examples/discriminative_reranking_nmt/tasks/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..2d78ca98708121261aa365738a65c051b5b40626
--- /dev/null
+++ b/fairseq/examples/discriminative_reranking_nmt/tasks/__init__.py
@@ -0,0 +1,6 @@
+from .discriminative_reranking_task import DiscriminativeRerankingNMTTask
+
+
+__all__ = [
+ "DiscriminativeRerankingNMTTask",
+]
diff --git a/fairseq/examples/discriminative_reranking_nmt/tasks/discriminative_reranking_task.py b/fairseq/examples/discriminative_reranking_nmt/tasks/discriminative_reranking_task.py
new file mode 100644
index 0000000000000000000000000000000000000000..b4ed2a69aacf64543c2d42d310cc113a98b254b2
--- /dev/null
+++ b/fairseq/examples/discriminative_reranking_nmt/tasks/discriminative_reranking_task.py
@@ -0,0 +1,490 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+from dataclasses import dataclass, field
+
+import itertools
+import logging
+import os
+
+import numpy as np
+import torch
+
+from fairseq.logging import metrics
+from fairseq.data import (
+ ConcatDataset,
+ ConcatSentencesDataset,
+ data_utils,
+ Dictionary,
+ IdDataset,
+ indexed_dataset,
+ NestedDictionaryDataset,
+ NumSamplesDataset,
+ NumelDataset,
+ PrependTokenDataset,
+ RawLabelDataset,
+ RightPadDataset,
+ SortDataset,
+ TruncateDataset,
+ TokenBlockDataset,
+)
+from fairseq.dataclass import ChoiceEnum, FairseqDataclass
+from fairseq.tasks import FairseqTask, register_task
+from omegaconf import II, MISSING
+
+
+EVAL_BLEU_ORDER = 4
+TARGET_METRIC_CHOICES = ChoiceEnum(["bleu", "ter"])
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class DiscriminativeRerankingNMTConfig(FairseqDataclass):
+ data: str = field(default=MISSING, metadata={"help": "path to data directory"})
+ num_data_splits: int = field(
+ default=1, metadata={"help": "total number of data splits"}
+ )
+ no_shuffle: bool = field(
+ default=False, metadata={"help": "do not shuffle training data"}
+ )
+ max_positions: int = field(
+ default=512, metadata={"help": "number of positional embeddings to learn"}
+ )
+ include_src: bool = field(
+ default=False, metadata={"help": "include source sentence"}
+ )
+ mt_beam: int = field(default=50, metadata={"help": "beam size of input hypotheses"})
+ eval_target_metric: bool = field(
+ default=False,
+ metadata={"help": "evaluation with the target metric during validation"},
+ )
+ target_metric: TARGET_METRIC_CHOICES = field(
+ default="bleu", metadata={"help": "name of the target metric to optimize for"}
+ )
+ train_subset: str = field(
+ default=II("dataset.train_subset"),
+ metadata={"help": "data subset to use for training (e.g. train, valid, test)"},
+ )
+ seed: int = field(
+ default=II("common.seed"),
+ metadata={"help": "pseudo random number generator seed"},
+ )
+
+
+class RerankerScorer(object):
+ """Scores the target for a given (source (optional), target) input."""
+
+ def __init__(self, args, mt_beam):
+ self.mt_beam = mt_beam
+
+ @torch.no_grad()
+ def generate(self, models, sample, **kwargs):
+ """Score a batch of translations."""
+ net_input = sample["net_input"]
+
+ assert len(models) == 1, "does not support model ensemble"
+ model = models[0]
+
+ bs = net_input["src_tokens"].shape[0]
+ assert (
+ model.joint_classification == "none" or bs % self.mt_beam == 0
+ ), f"invalid batch size ({bs}) for joint classification with beam size ({self.mt_beam})"
+
+ model.eval()
+ logits = model(**net_input)
+
+ batch_out = model.sentence_forward(logits, net_input["src_tokens"])
+ if model.joint_classification == "sent":
+ batch_out = model.joint_forward(
+ batch_out.view(self.mt_beam, bs // self.mt_beam, -1)
+ )
+ scores = model.classification_forward(
+ batch_out.view(bs, 1, -1)
+ ) # input: B x T x C
+
+ return scores
+
+
+@register_task(
+ "discriminative_reranking_nmt", dataclass=DiscriminativeRerankingNMTConfig
+)
+class DiscriminativeRerankingNMTTask(FairseqTask):
+ """
+ Translation rerank task.
+ The input can be either (src, tgt) sentence pairs or tgt sentence only.
+ """
+
+ cfg: DiscriminativeRerankingNMTConfig
+
+ def __init__(self, cfg: DiscriminativeRerankingNMTConfig, data_dictionary=None):
+ super().__init__(cfg)
+ self.dictionary = data_dictionary
+ self._max_positions = cfg.max_positions
+ # args.tokens_per_sample = self._max_positions
+ # self.num_classes = 1 # for model
+
+ @classmethod
+ def load_dictionary(cls, cfg, filename):
+ """Load the dictionary from the filename"""
+ dictionary = Dictionary.load(filename)
+ dictionary.add_symbol("") # for loading pretrained XLMR model
+
+ return dictionary
+
+ @classmethod
+ def setup_task(cls, cfg: DiscriminativeRerankingNMTConfig, **kwargs):
+ # load data dictionary (assume joint dictionary)
+ data_path = cfg.data
+ data_dict = cls.load_dictionary(
+ cfg, os.path.join(data_path, "input_src/dict.txt")
+ )
+
+ logger.info("[input] src dictionary: {} types".format(len(data_dict)))
+
+ return DiscriminativeRerankingNMTTask(cfg, data_dict)
+
+ def load_dataset(self, split, epoch=0, combine=False, **kwargs):
+ """Load a given dataset split (e.g., train, valid, test)."""
+ if self.cfg.data.endswith("1"):
+ data_shard = (epoch - 1) % self.cfg.num_data_splits + 1
+ data_path = self.cfg.data[:-1] + str(data_shard)
+ else:
+ data_path = self.cfg.data
+
+ def get_path(type, data_split):
+ return os.path.join(data_path, str(type), data_split)
+
+ def make_dataset(type, dictionary, data_split, combine):
+ split_path = get_path(type, data_split)
+
+ dataset = data_utils.load_indexed_dataset(
+ split_path,
+ dictionary,
+ combine=combine,
+ )
+ return dataset
+
+ def load_split(data_split, metric):
+ input_src = None
+ if self.cfg.include_src:
+ input_src = make_dataset(
+ "input_src", self.dictionary, data_split, combine=False
+ )
+ assert input_src is not None, "could not find dataset: {}".format(
+ get_path("input_src", data_split)
+ )
+
+ input_tgt = make_dataset(
+ "input_tgt", self.dictionary, data_split, combine=False
+ )
+ assert input_tgt is not None, "could not find dataset: {}".format(
+ get_path("input_tgt", data_split)
+ )
+
+ label_path = f"{get_path(metric, data_split)}.{metric}"
+ assert os.path.exists(label_path), f"could not find dataset: {label_path}"
+
+ np_labels = np.loadtxt(label_path)
+ if self.cfg.target_metric == "ter":
+ np_labels = -np_labels
+ label = RawLabelDataset(np_labels)
+
+ return input_src, input_tgt, label
+
+ src_datasets = []
+ tgt_datasets = []
+ label_datasets = []
+
+ if split == self.cfg.train_subset:
+ for k in itertools.count():
+ split_k = "train" + (str(k) if k > 0 else "")
+ prefix = os.path.join(data_path, "input_tgt", split_k)
+ if not indexed_dataset.dataset_exists(prefix, impl=None):
+ if k > 0:
+ break
+ else:
+ raise FileNotFoundError(f"Dataset not found: {prefix}")
+ input_src, input_tgt, label = load_split(
+ split_k, self.cfg.target_metric
+ )
+ src_datasets.append(input_src)
+ tgt_datasets.append(input_tgt)
+ label_datasets.append(label)
+ else:
+ input_src, input_tgt, label = load_split(split, self.cfg.target_metric)
+ src_datasets.append(input_src)
+ tgt_datasets.append(input_tgt)
+ label_datasets.append(label)
+
+ if len(tgt_datasets) == 1:
+ input_tgt, label = tgt_datasets[0], label_datasets[0]
+ if self.cfg.include_src:
+ input_src = src_datasets[0]
+ else:
+ input_tgt = ConcatDataset(tgt_datasets)
+ label = ConcatDataset(label_datasets)
+ if self.cfg.include_src:
+ input_src = ConcatDataset(src_datasets)
+
+ input_tgt = TruncateDataset(input_tgt, self.cfg.max_positions)
+ if self.cfg.include_src:
+ input_src = PrependTokenDataset(input_src, self.dictionary.bos())
+ input_src = TruncateDataset(input_src, self.cfg.max_positions)
+ src_lengths = NumelDataset(input_src, reduce=False)
+ src_tokens = ConcatSentencesDataset(input_src, input_tgt)
+ else:
+ src_tokens = PrependTokenDataset(input_tgt, self.dictionary.bos())
+ src_lengths = NumelDataset(src_tokens, reduce=False)
+
+ dataset = {
+ "id": IdDataset(),
+ "net_input": {
+ "src_tokens": RightPadDataset(
+ src_tokens,
+ pad_idx=self.source_dictionary.pad(),
+ ),
+ "src_lengths": src_lengths,
+ },
+ "nsentences": NumSamplesDataset(),
+ "ntokens": NumelDataset(src_tokens, reduce=True),
+ "target": label,
+ }
+
+ dataset = NestedDictionaryDataset(
+ dataset,
+ sizes=[src_tokens.sizes],
+ )
+
+ assert (
+ len(dataset) % self.cfg.mt_beam == 0
+ ), "dataset size (%d) is not a multiple of beam size (%d)" % (
+ len(dataset),
+ self.cfg.mt_beam,
+ )
+
+ # no need to shuffle valid/test sets
+ if not self.cfg.no_shuffle and split == self.cfg.train_subset:
+
+ # need to keep all hypothese together
+ start_idx = np.arange(0, len(dataset), self.cfg.mt_beam)
+ with data_utils.numpy_seed(self.cfg.seed + epoch):
+ np.random.shuffle(start_idx)
+
+ idx = np.arange(0, self.cfg.mt_beam)
+ shuffle = np.tile(idx, (len(start_idx), 1)).reshape(-1) + np.tile(
+ start_idx, (self.cfg.mt_beam, 1)
+ ).transpose().reshape(-1)
+
+ dataset = SortDataset(
+ dataset,
+ sort_order=[shuffle],
+ )
+
+ logger.info(f"Loaded {split} with #samples: {len(dataset)}")
+
+ self.datasets[split] = dataset
+ return self.datasets[split]
+
+ def build_dataset_for_inference(self, src_tokens, src_lengths, **kwargs):
+ assert not self.cfg.include_src or len(src_tokens[0]) == 2
+ input_src = None
+ if self.cfg.include_src:
+ input_src = TokenBlockDataset(
+ [t[0] for t in src_tokens],
+ [l[0] for l in src_lengths],
+ block_size=None, # ignored for "eos" break mode
+ pad=self.source_dictionary.pad(),
+ eos=self.source_dictionary.eos(),
+ break_mode="eos",
+ )
+ input_src = PrependTokenDataset(input_src, self.dictionary.bos())
+ input_src = TruncateDataset(input_src, self.cfg.max_positions)
+
+ input_tgt = TokenBlockDataset(
+ [t[-1] for t in src_tokens],
+ [l[-1] for l in src_lengths],
+ block_size=None, # ignored for "eos" break mode
+ pad=self.source_dictionary.pad(),
+ eos=self.source_dictionary.eos(),
+ break_mode="eos",
+ )
+ input_tgt = TruncateDataset(input_tgt, self.cfg.max_positions)
+ if self.cfg.include_src:
+ src_tokens = ConcatSentencesDataset(input_src, input_tgt)
+ src_lengths = NumelDataset(input_src, reduce=False)
+ else:
+ input_tgt = PrependTokenDataset(input_tgt, self.dictionary.bos())
+ src_tokens = input_tgt
+ src_lengths = NumelDataset(src_tokens, reduce=False)
+
+ dataset = {
+ "id": IdDataset(),
+ "net_input": {
+ "src_tokens": RightPadDataset(
+ src_tokens,
+ pad_idx=self.source_dictionary.pad(),
+ ),
+ "src_lengths": src_lengths,
+ },
+ "nsentences": NumSamplesDataset(),
+ "ntokens": NumelDataset(src_tokens, reduce=True),
+ }
+
+ return NestedDictionaryDataset(
+ dataset,
+ sizes=[src_tokens.sizes],
+ )
+
+ def build_model(self, cfg: FairseqDataclass, from_checkpoint: bool = False):
+ return super().build_model(cfg)
+
+ def build_generator(self, args):
+ return RerankerScorer(args, mt_beam=self.cfg.mt_beam)
+
+ def max_positions(self):
+ return self._max_positions
+
+ @property
+ def source_dictionary(self):
+ return self.dictionary
+
+ @property
+ def target_dictionary(self):
+ return self.dictionary
+
+ def create_dummy_batch(self, device):
+ dummy_target = (
+ torch.zeros(self.cfg.mt_beam, EVAL_BLEU_ORDER * 2 + 3).long().to(device)
+ if not self.cfg.eval_ter
+ else torch.zeros(self.cfg.mt_beam, 3).long().to(device)
+ )
+
+ return {
+ "id": torch.zeros(self.cfg.mt_beam, 1).long().to(device),
+ "net_input": {
+ "src_tokens": torch.zeros(self.cfg.mt_beam, 4).long().to(device),
+ "src_lengths": torch.ones(self.cfg.mt_beam, 1).long().to(device),
+ },
+ "nsentences": 0,
+ "ntokens": 0,
+ "target": dummy_target,
+ }
+
+ def train_step(
+ self, sample, model, criterion, optimizer, update_num, ignore_grad=False
+ ):
+ if ignore_grad and sample is None:
+ sample = self.create_dummy_batch(model.device)
+
+ return super().train_step(
+ sample, model, criterion, optimizer, update_num, ignore_grad
+ )
+
+ def valid_step(self, sample, model, criterion):
+ if sample is None:
+ sample = self.create_dummy_batch(model.device)
+
+ loss, sample_size, logging_output = super().valid_step(sample, model, criterion)
+
+ if not self.cfg.eval_target_metric:
+ return loss, sample_size, logging_output
+
+ scores = logging_output["scores"]
+
+ if self.cfg.target_metric == "bleu":
+ assert sample["target"].shape[1] == EVAL_BLEU_ORDER * 2 + 3, (
+ "target does not contain enough information ("
+ + str(sample["target"].shape[1])
+ + "for evaluating BLEU"
+ )
+
+ max_id = torch.argmax(scores, dim=1)
+ select_id = max_id + torch.arange(
+ 0, sample_size * self.cfg.mt_beam, self.cfg.mt_beam
+ ).to(max_id.device)
+ bleu_data = sample["target"][select_id, 1:].sum(0).data
+
+ logging_output["_bleu_sys_len"] = bleu_data[0]
+ logging_output["_bleu_ref_len"] = bleu_data[1]
+
+ for i in range(EVAL_BLEU_ORDER):
+ logging_output["_bleu_counts_" + str(i)] = bleu_data[2 + i]
+ logging_output["_bleu_totals_" + str(i)] = bleu_data[
+ 2 + EVAL_BLEU_ORDER + i
+ ]
+
+ elif self.cfg.target_metric == "ter":
+ assert sample["target"].shape[1] == 3, (
+ "target does not contain enough information ("
+ + str(sample["target"].shape[1])
+ + "for evaluating TER"
+ )
+
+ max_id = torch.argmax(scores, dim=1)
+ select_id = max_id + torch.arange(
+ 0, sample_size * self.cfg.mt_beam, self.cfg.mt_beam
+ ).to(max_id.device)
+ ter_data = sample["target"][select_id, 1:].sum(0).data
+
+ logging_output["_ter_num_edits"] = -ter_data[0]
+ logging_output["_ter_ref_len"] = -ter_data[1]
+
+ return loss, sample_size, logging_output
+
+ def reduce_metrics(self, logging_outputs, criterion):
+ super().reduce_metrics(logging_outputs, criterion)
+
+ if not self.cfg.eval_target_metric:
+ return
+
+ def sum_logs(key):
+ return sum(log.get(key, 0) for log in logging_outputs)
+
+ if self.cfg.target_metric == "bleu":
+ counts, totals = [], []
+ for i in range(EVAL_BLEU_ORDER):
+ counts.append(sum_logs("_bleu_counts_" + str(i)))
+ totals.append(sum_logs("_bleu_totals_" + str(i)))
+
+ if max(totals) > 0:
+ # log counts as numpy arrays -- log_scalar will sum them correctly
+ metrics.log_scalar("_bleu_counts", np.array(counts))
+ metrics.log_scalar("_bleu_totals", np.array(totals))
+ metrics.log_scalar("_bleu_sys_len", sum_logs("_bleu_sys_len"))
+ metrics.log_scalar("_bleu_ref_len", sum_logs("_bleu_ref_len"))
+
+ def compute_bleu(meters):
+ import inspect
+ import sacrebleu
+
+ fn_sig = inspect.getfullargspec(sacrebleu.compute_bleu)[0]
+ if "smooth_method" in fn_sig:
+ smooth = {"smooth_method": "exp"}
+ else:
+ smooth = {"smooth": "exp"}
+ bleu = sacrebleu.compute_bleu(
+ correct=meters["_bleu_counts"].sum,
+ total=meters["_bleu_totals"].sum,
+ sys_len=meters["_bleu_sys_len"].sum,
+ ref_len=meters["_bleu_ref_len"].sum,
+ **smooth,
+ )
+ return round(bleu.score, 2)
+
+ metrics.log_derived("bleu", compute_bleu)
+ elif self.cfg.target_metric == "ter":
+ num_edits = sum_logs("_ter_num_edits")
+ ref_len = sum_logs("_ter_ref_len")
+
+ if ref_len > 0:
+ metrics.log_scalar("_ter_num_edits", num_edits)
+ metrics.log_scalar("_ter_ref_len", ref_len)
+
+ def compute_ter(meters):
+ score = meters["_ter_num_edits"].sum / meters["_ter_ref_len"].sum
+ return round(score.item(), 2)
+
+ metrics.log_derived("ter", compute_ter)
diff --git a/fairseq/examples/emotion_conversion/README.md b/fairseq/examples/emotion_conversion/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..caf22befe4cfb6fa07d0f7e9c1268eb6ec9fc392
--- /dev/null
+++ b/fairseq/examples/emotion_conversion/README.md
@@ -0,0 +1,214 @@
+# Textless speech emotion conversion using decomposed and discrete representations
+[Felix Kreuk](https://felixkreuk.github.io), Adam Polyak, Jade Copet, Eugene Kharitonov, Tu-Anh Nguyen, Morgane Rivière, Wei-Ning Hsu, Abdelrahman Mohamed, Emmanuel Dupoux, [Yossi Adi](https://adiyoss.github.io)
+
+_abstract_: Speech emotion conversion is the task of modifying the perceived emotion of a speech utterance while preserving the lexical content and speaker identity. In this study, we cast the problem of emotion conversion as a spoken language translation task. We decompose speech into discrete and disentangled learned representations, consisting of content units, F0, speaker, and emotion. First, we modify the speech content by translating the content units to a target emotion, and then predict the prosodic features based on these units. Finally, the speech waveform is generated by feeding the predicted representations into a neural vocoder. Such a paradigm allows us to go beyond spectral and parametric changes of the signal, and model non-verbal vocalizations, such as laughter insertion, yawning removal, etc. We demonstrate objectively and subjectively that the proposed method is superior to the baselines in terms of perceived emotion and audio quality. We rigorously evaluate all components of such a complex system and conclude with an extensive model analysis and ablation study to better emphasize the architectural choices, strengths and weaknesses of the proposed method. Samples and code will be publicly available under the following link: https://speechbot.github.io/emotion.
+
+## Installation
+First, create a conda virtual environment and activate it:
+```
+conda create -n emotion python=3.8 -y
+conda activate emotion
+```
+
+Then, clone this repository:
+```
+git clone https://github.com/facebookresearch/fairseq.git
+cd fairseq/examples/emotion_conversion
+git clone https://github.com/felixkreuk/speech-resynthesis
+```
+
+Next, download the EmoV discrete tokens:
+```
+wget https://dl.fbaipublicfiles.com/textless_nlp/emotion_conversion/data.tar.gz # (still in fairseq/examples/emotion_conversion)
+tar -xzvf data.tar.gz
+```
+
+Your `fairseq/examples/emotion_conversion` directory should like this:
+```
+drwxrwxr-x 3 felixkreuk felixkreuk 0 Feb 6 2022 data
+drwxrwxr-x 3 felixkreuk felixkreuk 0 Sep 28 10:41 emotion_models
+drwxr-xr-x 3 felixkreuk felixkreuk 0 Jun 29 05:43 fairseq_models
+drwxr-xr-x 3 felixkreuk felixkreuk 0 Sep 28 10:41 preprocess
+-rw-rw-r-- 1 felixkreuk felixkreuk 11K Dec 5 09:00 README.md
+-rw-rw-r-- 1 felixkreuk felixkreuk 88 Mar 6 2022 requirements.txt
+-rw-rw-r-- 1 felixkreuk felixkreuk 13K Jun 29 06:26 synthesize.py
+```
+
+Lastly, install fairseq and the other packages:
+```
+pip install --editable ./
+pip install -r examples/emotion_conversion/requirements.txt
+```
+
+## Data preprocessing
+
+### Convert your audio to discrete representations
+Please follow the steps described [here](https://github.com/pytorch/fairseq/tree/main/examples/hubert/simple_kmeans).
+To generate the same discrete representations please use the following:
+1. [HuBERT checkpoint](https://dl.fbaipublicfiles.com/hubert/hubert_base_ls960.pt)
+2. k-means model at `data/hubert_base_ls960_layer9_clusters200/data_hubert_base_ls960_layer9_clusters200.bin`
+
+### Construct data splits
+This step will use the discrete representations from the previous step and split them to train/valid/test sets for 3 tasks:
+1. Translation model pre-training (BART language denoising)
+2. Translation model training (content units emotion translation mechanism)
+3. HiFiGAN model training (for synthesizing audio from discrete representations)
+
+Your processed data should be at `data/`:
+1. `hubert_base_ls960_layer9_clusters200` - discrete representations extracted using HuBERT layer 9, clustered into 200 clusters.
+2. `data.tsv` - a tsv file pointing to the EmoV dataset in your environment (Please edit the first line of this file according to your path).
+
+The following command will create the above splits:
+```
+python examples/emotion_conversion/preprocess/create_core_manifest.py \
+ --tsv data/data.tsv \
+ --emov-km data/hubert_base_ls960_layer9_clusters200/data.km \
+ --km data/hubert_base_ls960_layer9_clusters200/vctk.km \
+ --dict data/hubert_base_ls960_layer9_clusters200/dict.txt \
+ --manifests-dir $DATA
+```
+* Set `$DATA` as the directory that will contain the processed data.
+
+### Extract F0
+To train the HiFiGAN vocoder we need to first extract the F0 curves:
+```
+python examples/emotion_conversion/preprocess/extract_f0.py \
+ --tsv data/data.tsv \
+ --extractor pyaapt \
+```
+
+## HiFiGAN training
+Now we are all set to train the HiFiGAN vocoder:
+```
+python examples/emotion_conversion/speech-resynthesis/train.py
+ --checkpoint_path \
+ --config examples/emotion_conversion/speech-resynthesis/configs/EmoV/emov_hubert-layer9-cluster200_fixed-spkr-embedder_f0-raw_gst.json
+```
+
+## Translation Pre-training
+Before translating emotions, we first need to pre-train the translation model as a denoising autoencoder (similarly to BART).
+```
+python train.py \
+ $DATA/fairseq-data/emov_multilingual_denoising_cross-speaker_dedup_nonzeroshot/tokenized \
+ --save-dir \
+ --tensorboard-logdir \
+ --langs neutral,amused,angry,sleepy,disgusted,vctk.km \
+ --dataset-impl mmap \
+ --task multilingual_denoising \
+ --arch transformer_small --criterion cross_entropy \
+ --multilang-sampling-alpha 1.0 --sample-break-mode eos --max-tokens 16384 \
+ --update-freq 1 --max-update 3000000 \
+ --dropout 0.1 --attention-dropout 0.1 --relu-dropout 0.0 \
+ --optimizer adam --weight-decay 0.01 --adam-eps 1e-06 \
+ --clip-norm 0.1 --lr-scheduler polynomial_decay --lr 0.0003 \
+ --total-num-update 3000000 --warmup-updates 10000 --fp16 \
+ --poisson-lambda 3.5 --mask 0.3 --mask-length span-poisson --replace-length 1 --rotate 0 --mask-random 0.1 --insert 0 --permute-sentences 1.0 \
+ --skip-invalid-size-inputs-valid-test \
+ --user-dir examples/emotion_conversion/fairseq_models
+```
+
+## Translation Training
+Now we are ready to train our emotion translation model:
+```
+python train.py \
+ --distributed-world-size 1 \
+ $DATA/fairseq-data/emov_multilingual_translation_cross-speaker_dedup/tokenized/ \
+ --save-dir \
+ --tensorboard-logdir \
+ --arch multilingual_small --task multilingual_translation \
+ --criterion label_smoothed_cross_entropy --label-smoothing 0.2 \
+ --lang-pairs neutral-amused,neutral-sleepy,neutral-disgusted,neutral-angry,amused-sleepy,amused-disgusted,amused-neutral,amused-angry,angry-amused,angry-sleepy,angry-disgusted,angry-neutral,disgusted-amused,disgusted-sleepy,disgusted-neutral,disgusted-angry,sleepy-amused,sleepy-neutral,sleepy-disgusted,sleepy-angry \
+ --optimizer adam --adam-betas "(0.9, 0.98)" --adam-eps 1e-06 \
+ --lr 1e-05 --clip-norm 0 --dropout 0.1 --attention-dropout 0.1 \
+ --weight-decay 0.01 --warmup-updates 2000 --lr-scheduler inverse_sqrt \
+ --max-tokens 4096 --update-freq 1 --max-update 100000 \
+ --required-batch-size-multiple 8 --fp16 --num-workers 4 \
+ --seed 2 --log-format json --log-interval 25 --save-interval-updates 1000 \
+ --no-epoch-checkpoints --keep-best-checkpoints 1 --keep-interval-updates 1 \
+ --finetune-from-model \
+ --user-dir examples/emotion_conversion/fairseq_models
+```
+* To share encoders/decoders use the `--share-encoders` and `--share-decoders` flags.
+* To add source/target emotion tokens use the `--encoder-langtok {'src'|'tgt'}` and `--decoder-langtok` flags.
+
+## F0-predictor Training
+The following command trains the F0 prediction module:
+```
+cd examples/emotion_conversion
+python -m emotion_models.pitch_predictor n_tokens=200 \
+ train_tsv="$DATA/denoising/emov/train.tsv" \
+ train_km="$DATA/denoising/emov/train.km" \
+ valid_tsv="$DATA/denoising/emov/valid.tsv" \
+ valid_km="$DATA/denoising/emov/valid.km"
+```
+* See `hyra.run.dir` to configure directory for saving models.
+
+## Duration-predictor Training
+The following command trains the duration prediction modules:
+```
+cd examples/emotion_conversion
+for emotion in "neutral" "amused" "angry" "disgusted" "sleepy"; do
+ python -m emotion_models.duration_predictor n_tokens=200 substring=$emotion \
+ train_tsv="$DATA/denoising/emov/train.tsv" \
+ train_km="$DATA/denoising/emov/train.km" \
+ valid_tsv="$DATA/denoising/emov/valid.tsv" \
+ valid_km="$DATA/denoising/emov/valid.km"
+done
+```
+* See `hyra.run.dir` to configure directory for saving models.
+* After the above command you should have 5 duration models in your checkpoint directory:
+```
+❯ ll duration_predictor/
+total 21M
+-rw-rw-r-- 1 felixkreuk felixkreuk 4.1M Nov 15 2021 amused.ckpt
+-rw-rw-r-- 1 felixkreuk felixkreuk 4.1M Nov 15 2021 angry.ckpt
+-rw-rw-r-- 1 felixkreuk felixkreuk 4.1M Nov 15 2021 disgusted.ckpt
+-rw-rw-r-- 1 felixkreuk felixkreuk 4.1M Nov 15 2021 neutral.ckpt
+-rw-rw-r-- 1 felixkreuk felixkreuk 4.1M Nov 15 2021 sleepy.ckpt
+```
+
+## Token Generation
+The following command uses `fairseq-generate` to generate the token sequences based on the source and target emotions.
+```
+fairseq-generate \
+ $DATA/fairseq-data/emov_multilingual_translation_cross-speaker_dedup/tokenized/ \
+ --task multilingual_translation \
+ --gen-subset test \
+ --path \
+ --beam 5 \
+ --batch-size 4 --max-len-a 1.8 --max-len-b 10 --lenpen 1 --min-len 1 \
+ --skip-invalid-size-inputs-valid-test --distributed-world-size 1 \
+ --source-lang neutral --target-lang amused \
+ --lang-pairs neutral-amused,neutral-sleepy,neutral-disgusted,neutral-angry,amused-sleepy,amused-disgusted,amused-neutral,amused-angry,angry-amused,angry-sleepy,angry-disgusted,angry-neutral,disgusted-amused,disgusted-sleepy,disgusted-neutral,disgusted-angry,sleepy-amused,sleepy-neutral,sleepy-disgusted,sleepy-angry \
+ --results-path \
+ --user-dir examples/emotion_conversion/fairseq_models
+```
+* Modify `--source-lang` and `--target-lang` to control for the source and target emotions.
+* See [fairseq documentation](https://fairseq.readthedocs.io/en/latest/command_line_tools.html#fairseq-generate) for a full overview of generation parameters (e.g., top-k/top-p sampling).
+
+## Waveform Synthesis
+Using the output of the above command, the HiFiGAN vocoder, and the prosody prediction modules (F0 and duration) we can now generate the output waveforms:
+```
+python examples/emotion_conversion/synthesize.py \
+ --result-path /generate-test.txt \
+ --data $DATA/fairseq-data/emov_multilingual_translation_cross-speaker_dedup/neutral-amused \
+ --orig-tsv examples/emotion_conversion/data/data.tsv \
+ --orig-km examples/emotion_conversion/data/hubert_base_ls960_layer9_clusters200/data.km \
+ --checkpoint-file /g_00400000 \
+ --dur-model duration_predictor/ \
+ --f0-model pitch_predictor/pitch_predictor.ckpt \
+ -s neutral -t amused \
+ --outdir ~/tmp/emotion_results/wavs/neutral-amused
+```
+* Please make sure the source and target emotions here match those of the previous command.
+
+# Citation
+If you find this useful in your research, please use the following BibTeX entry for citation.
+```
+@article{kreuk2021textless,
+ title={Textless speech emotion conversion using decomposed and discrete representations},
+ author={Kreuk, Felix and Polyak, Adam and Copet, Jade and Kharitonov, Eugene and Nguyen, Tu-Anh and Rivi{\`e}re, Morgane and Hsu, Wei-Ning and Mohamed, Abdelrahman and Dupoux, Emmanuel and Adi, Yossi},
+ journal={Conference on Empirical Methods in Natural Language Processing (EMNLP)},
+ year={2022}
+}
+```
diff --git a/fairseq/examples/emotion_conversion/emotion_models/__init__.py b/fairseq/examples/emotion_conversion/emotion_models/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/fairseq/examples/emotion_conversion/emotion_models/duration_predictor.py b/fairseq/examples/emotion_conversion/emotion_models/duration_predictor.py
new file mode 100644
index 0000000000000000000000000000000000000000..eb47df0a214b593987611e2d62c61c48a0ea0c67
--- /dev/null
+++ b/fairseq/examples/emotion_conversion/emotion_models/duration_predictor.py
@@ -0,0 +1,243 @@
+import logging
+import os
+
+import hydra
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops.layers.torch import Rearrange
+from torch.utils.data import DataLoader, Dataset
+
+from .utils import Accuracy
+
+logger = logging.getLogger(__name__)
+
+
+def save_ckpt(model, path, model_class):
+ ckpt = {
+ "state_dict": model.state_dict(),
+ "padding_token": model.padding_token,
+ "model_class": model_class,
+ }
+ torch.save(ckpt, path)
+
+
+def load_ckpt(path):
+ ckpt = torch.load(path)
+ ckpt["model_class"]["_target_"] = "emotion_models.duration_predictor.CnnPredictor"
+ model = hydra.utils.instantiate(ckpt["model_class"])
+ model.load_state_dict(ckpt["state_dict"])
+ model.padding_token = ckpt["padding_token"]
+ model = model.cpu()
+ model.eval()
+ return model
+
+
+class Collator:
+ def __init__(self, padding_idx):
+ self.padding_idx = padding_idx
+
+ def __call__(self, batch):
+ x = [item[0] for item in batch]
+ lengths = [len(item) for item in x]
+ x = torch.nn.utils.rnn.pad_sequence(x, batch_first=True, padding_value=self.padding_idx)
+ y = [item[1] for item in batch]
+ y = torch.nn.utils.rnn.pad_sequence(y, batch_first=True, padding_value=self.padding_idx)
+ mask = (x != self.padding_idx)
+ return x, y, mask, lengths
+
+
+class Predictor(nn.Module):
+ def __init__(self, n_tokens, emb_dim):
+ super(Predictor, self).__init__()
+ self.n_tokens = n_tokens
+ self.emb_dim = emb_dim
+ self.padding_token = n_tokens
+ # add 1 extra embedding for padding token, set the padding index to be the last token
+ # (tokens from the clustering start at index 0)
+ self.emb = nn.Embedding(n_tokens + 1, emb_dim, padding_idx=self.padding_token)
+
+ def inflate_input(self, batch):
+ """ get a sequence of tokens, predict their durations
+ and inflate them accordingly """
+ batch_durs = self.forward(batch)
+ batch_durs = torch.exp(batch_durs) - 1
+ batch_durs = batch_durs.round()
+ output = []
+ for seq, durs in zip(batch, batch_durs):
+ inflated_seq = []
+ for token, n in zip(seq, durs):
+ if token == self.padding_token:
+ break
+ n = int(n.item())
+ token = int(token.item())
+ inflated_seq.extend([token for _ in range(n)])
+ output.append(inflated_seq)
+ output = torch.LongTensor(output)
+ return output
+
+
+class CnnPredictor(Predictor):
+ def __init__(self, n_tokens, emb_dim, channels, kernel, output_dim, dropout, n_layers):
+ super(CnnPredictor, self).__init__(n_tokens=n_tokens, emb_dim=emb_dim)
+ layers = [
+ Rearrange("b t c -> b c t"),
+ nn.Conv1d(emb_dim, channels, kernel_size=kernel, padding=(kernel - 1) // 2),
+ Rearrange("b c t -> b t c"),
+ nn.ReLU(),
+ nn.LayerNorm(channels),
+ nn.Dropout(dropout),
+ ]
+ for _ in range(n_layers-1):
+ layers += [
+ Rearrange("b t c -> b c t"),
+ nn.Conv1d(channels, channels, kernel_size=kernel, padding=(kernel - 1) // 2),
+ Rearrange("b c t -> b t c"),
+ nn.ReLU(),
+ nn.LayerNorm(channels),
+ nn.Dropout(dropout),
+ ]
+ self.conv_layer = nn.Sequential(*layers)
+ self.proj = nn.Linear(channels, output_dim)
+
+ def forward(self, x):
+ x = self.emb(x)
+ x = self.conv_layer(x)
+ x = self.proj(x)
+ x = x.squeeze(-1)
+ return x
+
+
+def l2_log_loss(input, target):
+ return F.mse_loss(
+ input=input.float(),
+ target=torch.log(target.float() + 1),
+ reduce=False
+ )
+
+
+class DurationDataset(Dataset):
+ def __init__(self, tsv_path, km_path, substring=""):
+ lines = open(tsv_path, "r").readlines()
+ self.root, self.tsv = lines[0], lines[1:]
+ self.km = open(km_path, "r").readlines()
+ logger.info(f"loaded {len(self.km)} files")
+
+ if substring != "":
+ tsv, km = [], []
+ for tsv_line, km_line in zip(self.tsv, self.km):
+ if substring.lower() in tsv_line.lower():
+ tsv.append(tsv_line)
+ km.append(km_line)
+ self.tsv, self.km = tsv, km
+ logger.info(f"after filtering: {len(self.km)} files")
+
+ def __len__(self):
+ return len(self.km)
+
+ def __getitem__(self, i):
+ x = self.km[i]
+ x = x.split(" ")
+ x = list(map(int, x))
+
+ y = []
+ xd = []
+ count = 1
+ for x1, x2 in zip(x[:-1], x[1:]):
+ if x1 == x2:
+ count += 1
+ continue
+ else:
+ y.append(count)
+ xd.append(x1)
+ count = 1
+
+ xd = torch.LongTensor(xd)
+ y = torch.LongTensor(y)
+ return xd, y
+
+
+def train(cfg):
+ device = "cuda:0"
+ model = hydra.utils.instantiate(cfg[cfg.model]).to(device)
+ optimizer = hydra.utils.instantiate(cfg.optimizer, model.parameters())
+ # add 1 extra embedding for padding token, set the padding index to be the last token
+ # (tokens from the clustering start at index 0)
+ collate_fn = Collator(padding_idx=model.padding_token)
+ logger.info(f"data: {cfg.train_tsv}")
+ train_ds = DurationDataset(cfg.train_tsv, cfg.train_km, substring=cfg.substring)
+ valid_ds = DurationDataset(cfg.valid_tsv, cfg.valid_km, substring=cfg.substring)
+ train_dl = DataLoader(train_ds, batch_size=32, shuffle=True, collate_fn=collate_fn)
+ valid_dl = DataLoader(valid_ds, batch_size=32, shuffle=False, collate_fn=collate_fn)
+
+ best_loss = float("inf")
+ for epoch in range(cfg.epochs):
+ train_loss, train_loss_scaled = train_epoch(model, train_dl, l2_log_loss, optimizer, device)
+ valid_loss, valid_loss_scaled, *acc = valid_epoch(model, valid_dl, l2_log_loss, device)
+ acc0, acc1, acc2, acc3 = acc
+ if valid_loss_scaled < best_loss:
+ path = f"{os.getcwd()}/{cfg.substring}.ckpt"
+ save_ckpt(model, path, cfg[cfg.model])
+ best_loss = valid_loss_scaled
+ logger.info(f"saved checkpoint: {path}")
+ logger.info(f"[epoch {epoch}] train loss: {train_loss:.3f}, train scaled: {train_loss_scaled:.3f}")
+ logger.info(f"[epoch {epoch}] valid loss: {valid_loss:.3f}, valid scaled: {valid_loss_scaled:.3f}")
+ logger.info(f"acc: {acc0,acc1,acc2,acc3}")
+
+
+def train_epoch(model, loader, criterion, optimizer, device):
+ model.train()
+ epoch_loss = 0
+ epoch_loss_scaled = 0
+ for x, y, mask, _ in loader:
+ x, y, mask = x.to(device), y.to(device), mask.to(device)
+ yhat = model(x)
+ loss = criterion(yhat, y) * mask
+ loss = torch.mean(loss)
+ loss.backward()
+ nn.utils.clip_grad_norm_(model.parameters(), 1.0)
+ optimizer.step()
+ epoch_loss += loss.item()
+ # get normal scale loss
+ yhat_scaled = torch.exp(yhat) - 1
+ yhat_scaled = torch.round(yhat_scaled)
+ scaled_loss = torch.mean(torch.abs(yhat_scaled - y) * mask)
+ epoch_loss_scaled += scaled_loss.item()
+ return epoch_loss / len(loader), epoch_loss_scaled / len(loader)
+
+
+def valid_epoch(model, loader, criterion, device):
+ model.eval()
+ epoch_loss = 0
+ epoch_loss_scaled = 0
+ acc = Accuracy()
+ for x, y, mask, _ in loader:
+ x, y, mask = x.to(device), y.to(device), mask.to(device)
+ yhat = model(x)
+ loss = criterion(yhat, y) * mask
+ loss = torch.mean(loss)
+ epoch_loss += loss.item()
+ # get normal scale loss
+ yhat_scaled = torch.exp(yhat) - 1
+ yhat_scaled = torch.round(yhat_scaled)
+ scaled_loss = torch.sum(torch.abs(yhat_scaled - y) * mask) / mask.sum()
+ acc.update(yhat_scaled[mask].view(-1).float(), y[mask].view(-1).float())
+ epoch_loss_scaled += scaled_loss.item()
+ logger.info(f"example y: {y[0, :10].tolist()}")
+ logger.info(f"example yhat: {yhat_scaled[0, :10].tolist()}")
+ acc0 = acc.acc(tol=0)
+ acc1 = acc.acc(tol=1)
+ acc2 = acc.acc(tol=2)
+ acc3 = acc.acc(tol=3)
+ logger.info(f"accs: {acc0,acc1,acc2,acc3}")
+ return epoch_loss / len(loader), epoch_loss_scaled / len(loader), acc0, acc1, acc2, acc3
+
+
+@hydra.main(config_path=".", config_name="duration_predictor.yaml")
+def main(cfg):
+ logger.info(f"{cfg}")
+ train(cfg)
+
+
+if __name__ == "__main__":
+ main()
diff --git a/fairseq/examples/emotion_conversion/emotion_models/duration_predictor.yaml b/fairseq/examples/emotion_conversion/emotion_models/duration_predictor.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..0e976f484351a7d64eaedb70eb7866ce8555b419
--- /dev/null
+++ b/fairseq/examples/emotion_conversion/emotion_models/duration_predictor.yaml
@@ -0,0 +1,48 @@
+train_tsv: "/denoising/emov/train.tsv"
+train_km: "/denoising/emov/train.km"
+valid_tsv: "/denoising/emov/valid.tsv"
+valid_km: "/denoising/emov/valid.km"
+
+n_tokens: 200
+batch_size: 32
+lr: 0.0001
+epochs: 300
+model: "cnn"
+substring: ""
+
+rnn:
+ _target_: emotion_models.duration_predictor.RnnPredictor
+ n_tokens: ${n_tokens}
+ emb_dim: 128
+ rnn_hidden: 128
+ output_dim: 1
+ dropout: 0
+ n_layers: 1
+
+optimizer:
+ _target_: torch.optim.Adam
+ lr: ${lr}
+ betas: [0.9, 0.98]
+ eps: 0.000000001
+ weight_decay: 0
+
+cnn:
+ _target_: emotion_models.duration_predictor.CnnPredictor
+ n_tokens: ${n_tokens}
+ emb_dim: 128
+ channels: 256
+ kernel: 3
+ output_dim: 1
+ dropout: 0.5
+ n_layers: 1
+
+hydra:
+ run:
+ dir: /checkpoint/felixkreuk/experiments/duration_predictor/${hydra.job.override_dirname}
+ job:
+ config:
+ # configuration for the ${hydra.job.override_dirname} runtime variable
+ override_dirname:
+ kv_sep: '='
+ item_sep: ','
+ exclude_keys: ['train_tsv', 'train_km', 'valid_tsv', 'valid_km']
diff --git a/fairseq/examples/emotion_conversion/emotion_models/pitch_predictor.py b/fairseq/examples/emotion_conversion/emotion_models/pitch_predictor.py
new file mode 100644
index 0000000000000000000000000000000000000000..431446996c30e1dadb61a1834e5ca56b647fe8f1
--- /dev/null
+++ b/fairseq/examples/emotion_conversion/emotion_models/pitch_predictor.py
@@ -0,0 +1,559 @@
+import logging
+import os
+import random
+import sys
+from collections import defaultdict
+
+import hydra
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange
+from einops.layers.torch import Rearrange
+from scipy.io.wavfile import read
+from scipy.ndimage import gaussian_filter1d
+from torch.utils.data import DataLoader, Dataset
+from tqdm import tqdm
+
+dir_path = os.path.dirname(__file__)
+resynth_path = os.path.dirname(dir_path) + "/speech-resynthesis"
+sys.path.append(resynth_path)
+from dataset import parse_speaker, parse_style
+from .utils import F0Stat
+
+MAX_WAV_VALUE = 32768.0
+logger = logging.getLogger(__name__)
+
+
+def quantize_f0(speaker_to_f0, nbins, normalize, log):
+ f0_all = []
+ for speaker, f0 in speaker_to_f0.items():
+ f0 = f0.raw_data
+ if log:
+ f0 = f0.log()
+ mean = speaker_to_f0[speaker].mean_log if log else speaker_to_f0[speaker].mean
+ std = speaker_to_f0[speaker].std_log if log else speaker_to_f0[speaker].std
+ if normalize == "mean":
+ f0 = f0 - mean
+ elif normalize == "meanstd":
+ f0 = (f0 - mean) / std
+ f0_all.extend(f0.tolist())
+
+ hist, bin_x = np.histogram(f0_all, 100000)
+ cum_hist = np.cumsum(hist) / len(f0_all) * 100
+
+ bin_offset = []
+ bin_size = 100 / nbins
+ threshold = bin_size
+ for i in range(nbins - 1):
+ index = (np.abs(cum_hist - threshold)).argmin()
+ bin_offset.append(bin_x[index])
+ threshold += bin_size
+ bins = np.array(bin_offset)
+ bins = torch.FloatTensor(bins)
+
+ return bins
+
+
+def save_ckpt(model, path, model_class, f0_min, f0_max, f0_bins, speaker_stats):
+ ckpt = {
+ "state_dict": model.state_dict(),
+ "padding_token": model.padding_token,
+ "model_class": model_class,
+ "speaker_stats": speaker_stats,
+ "f0_min": f0_min,
+ "f0_max": f0_max,
+ "f0_bins": f0_bins,
+ }
+ torch.save(ckpt, path)
+
+
+def load_ckpt(path):
+ ckpt = torch.load(path)
+ ckpt["model_class"]["_target_"] = "emotion_models.pitch_predictor.CnnPredictor"
+ model = hydra.utils.instantiate(ckpt["model_class"])
+ model.load_state_dict(ckpt["state_dict"])
+ model.setup_f0_stats(
+ ckpt["f0_min"],
+ ckpt["f0_max"],
+ ckpt["f0_bins"],
+ ckpt["speaker_stats"],
+ )
+ return model
+
+
+def freq2bin(f0, f0_min, f0_max, bins):
+ f0 = f0.clone()
+ f0[f0 < f0_min] = f0_min
+ f0[f0 > f0_max] = f0_max
+ f0 = torch.bucketize(f0, bins)
+ return f0
+
+
+def bin2freq(x, f0_min, f0_max, bins, mode):
+ n_bins = len(bins) + 1
+ assert x.shape[-1] == n_bins
+ bins = torch.cat([torch.tensor([f0_min]), bins]).to(x.device)
+ if mode == "mean":
+ f0 = (x * bins).sum(-1, keepdims=True) / x.sum(-1, keepdims=True)
+ elif mode == "argmax":
+ idx = F.one_hot(x.argmax(-1), num_classes=n_bins)
+ f0 = (idx * bins).sum(-1, keepdims=True)
+ else:
+ raise NotImplementedError()
+ return f0[..., 0]
+
+
+def load_wav(full_path):
+ sampling_rate, data = read(full_path)
+ return data, sampling_rate
+
+
+def l1_loss(input, target):
+ return F.l1_loss(input=input.float(), target=target.float(), reduce=False)
+
+
+def l2_loss(input, target):
+ return F.mse_loss(input=input.float(), target=target.float(), reduce=False)
+
+
+class Collator:
+ def __init__(self, padding_idx):
+ self.padding_idx = padding_idx
+
+ def __call__(self, batch):
+ tokens = [item[0] for item in batch]
+ lengths = [len(item) for item in tokens]
+ tokens = torch.nn.utils.rnn.pad_sequence(
+ tokens, batch_first=True, padding_value=self.padding_idx
+ )
+ f0 = [item[1] for item in batch]
+ f0 = torch.nn.utils.rnn.pad_sequence(
+ f0, batch_first=True, padding_value=self.padding_idx
+ )
+ f0_raw = [item[2] for item in batch]
+ f0_raw = torch.nn.utils.rnn.pad_sequence(
+ f0_raw, batch_first=True, padding_value=self.padding_idx
+ )
+ spk = [item[3] for item in batch]
+ spk = torch.LongTensor(spk)
+ gst = [item[4] for item in batch]
+ gst = torch.LongTensor(gst)
+ mask = tokens != self.padding_idx
+ return tokens, f0, f0_raw, spk, gst, mask, lengths
+
+
+class CnnPredictor(nn.Module):
+ def __init__(
+ self,
+ n_tokens,
+ emb_dim,
+ channels,
+ kernel,
+ dropout,
+ n_layers,
+ spk_emb,
+ gst_emb,
+ n_bins,
+ f0_pred,
+ f0_log,
+ f0_norm,
+ ):
+ super(CnnPredictor, self).__init__()
+ self.n_tokens = n_tokens
+ self.emb_dim = emb_dim
+ self.f0_log = f0_log
+ self.f0_pred = f0_pred
+ self.padding_token = n_tokens
+ self.f0_norm = f0_norm
+ # add 1 extra embedding for padding token, set the padding index to be the last token
+ # (tokens from the clustering start at index 0)
+ self.token_emb = nn.Embedding(
+ n_tokens + 1, emb_dim, padding_idx=self.padding_token
+ )
+
+ self.spk_emb = spk_emb
+ self.gst_emb = nn.Embedding(20, gst_emb)
+ self.setup = False
+
+ feats = emb_dim + gst_emb
+ # feats = emb_dim + gst_emb + (256 if spk_emb else 0)
+ layers = [
+ nn.Sequential(
+ Rearrange("b t c -> b c t"),
+ nn.Conv1d(
+ feats, channels, kernel_size=kernel, padding=(kernel - 1) // 2
+ ),
+ Rearrange("b c t -> b t c"),
+ nn.ReLU(),
+ nn.LayerNorm(channels),
+ nn.Dropout(dropout),
+ )
+ ]
+ for _ in range(n_layers - 1):
+ layers += [
+ nn.Sequential(
+ Rearrange("b t c -> b c t"),
+ nn.Conv1d(
+ channels,
+ channels,
+ kernel_size=kernel,
+ padding=(kernel - 1) // 2,
+ ),
+ Rearrange("b c t -> b t c"),
+ nn.ReLU(),
+ nn.LayerNorm(channels),
+ nn.Dropout(dropout),
+ )
+ ]
+ self.conv_layer = nn.ModuleList(layers)
+ self.proj = nn.Linear(channels, n_bins)
+
+ def forward(self, x, gst=None):
+ x = self.token_emb(x)
+ feats = [x]
+
+ if gst is not None:
+ gst = self.gst_emb(gst)
+ gst = rearrange(gst, "b c -> b c 1")
+ gst = F.interpolate(gst, x.shape[1])
+ gst = rearrange(gst, "b c t -> b t c")
+ feats.append(gst)
+
+ x = torch.cat(feats, dim=-1)
+
+ for i, conv in enumerate(self.conv_layer):
+ if i != 0:
+ x = conv(x) + x
+ else:
+ x = conv(x)
+
+ x = self.proj(x)
+ x = x.squeeze(-1)
+
+ if self.f0_pred == "mean":
+ x = torch.sigmoid(x)
+ elif self.f0_pred == "argmax":
+ x = torch.softmax(x, dim=-1)
+ else:
+ raise NotImplementedError
+ return x
+
+ def setup_f0_stats(self, f0_min, f0_max, f0_bins, speaker_stats):
+ self.f0_min = f0_min
+ self.f0_max = f0_max
+ self.f0_bins = f0_bins
+ self.speaker_stats = speaker_stats
+ self.setup = True
+
+ def inference(self, x, spk_id=None, gst=None):
+ assert (
+ self.setup == True
+ ), "make sure that `setup_f0_stats` was called before inference!"
+ probs = self(x, gst)
+ f0 = bin2freq(probs, self.f0_min, self.f0_max, self.f0_bins, self.f0_pred)
+ for i in range(f0.shape[0]):
+ mean = (
+ self.speaker_stats[spk_id[i].item()].mean_log
+ if self.f0_log
+ else self.speaker_stats[spk_id[i].item()].mean
+ )
+ std = (
+ self.speaker_stats[spk_id[i].item()].std_log
+ if self.f0_log
+ else self.speaker_stats[spk_id[i].item()].std
+ )
+ if self.f0_norm == "mean":
+ f0[i] = f0[i] + mean
+ if self.f0_norm == "meanstd":
+ f0[i] = (f0[i] * std) + mean
+ if self.f0_log:
+ f0 = f0.exp()
+ return f0
+
+
+class PitchDataset(Dataset):
+ def __init__(
+ self,
+ tsv_path,
+ km_path,
+ substring,
+ spk,
+ spk2id,
+ gst,
+ gst2id,
+ f0_bins,
+ f0_bin_type,
+ f0_smoothing,
+ f0_norm,
+ f0_log,
+ ):
+ lines = open(tsv_path, "r").readlines()
+ self.root, self.tsv = lines[0], lines[1:]
+ self.root = self.root.strip()
+ self.km = open(km_path, "r").readlines()
+ print(f"loaded {len(self.km)} files")
+
+ self.spk = spk
+ self.spk2id = spk2id
+ self.gst = gst
+ self.gst2id = gst2id
+
+ self.f0_bins = f0_bins
+ self.f0_smoothing = f0_smoothing
+ self.f0_norm = f0_norm
+ self.f0_log = f0_log
+
+ if substring != "":
+ tsv, km = [], []
+ for tsv_line, km_line in zip(self.tsv, self.km):
+ if substring.lower() in tsv_line.lower():
+ tsv.append(tsv_line)
+ km.append(km_line)
+ self.tsv, self.km = tsv, km
+ print(f"after filtering: {len(self.km)} files")
+
+ self.speaker_stats = self._compute_f0_stats()
+ self.f0_min, self.f0_max = self._compute_f0_minmax()
+ if f0_bin_type == "adaptive":
+ self.f0_bins = quantize_f0(
+ self.speaker_stats, self.f0_bins, self.f0_norm, self.f0_log
+ )
+ elif f0_bin_type == "uniform":
+ self.f0_bins = torch.linspace(self.f0_min, self.f0_max, self.f0_bins + 1)[
+ 1:-1
+ ]
+ else:
+ raise NotImplementedError
+ print(f"f0 min: {self.f0_min}, f0 max: {self.f0_max}")
+ print(f"bins: {self.f0_bins} (shape: {self.f0_bins.shape})")
+
+ def __len__(self):
+ return len(self.km)
+
+ def _load_f0(self, tsv_line):
+ tsv_line = tsv_line.split("\t")[0]
+ f0 = self.root + "/" + tsv_line.replace(".wav", ".yaapt.f0.npy")
+ f0 = np.load(f0)
+ f0 = torch.FloatTensor(f0)
+ return f0
+
+ def _preprocess_f0(self, f0, spk):
+ mask = f0 != -999999 # process all frames
+ # mask = (f0 != 0) # only process voiced frames
+ mean = (
+ self.speaker_stats[spk].mean_log
+ if self.f0_log
+ else self.speaker_stats[spk].mean
+ )
+ std = (
+ self.speaker_stats[spk].std_log
+ if self.f0_log
+ else self.speaker_stats[spk].std
+ )
+ if self.f0_log:
+ f0[f0 == 0] = 1e-5
+ f0[mask] = f0[mask].log()
+ if self.f0_norm == "mean":
+ f0[mask] = f0[mask] - mean
+ if self.f0_norm == "meanstd":
+ f0[mask] = (f0[mask] - mean) / std
+ return f0
+
+ def _compute_f0_minmax(self):
+ f0_min, f0_max = float("inf"), -float("inf")
+ for tsv_line in tqdm(self.tsv, desc="computing f0 minmax"):
+ spk = self.spk2id[parse_speaker(tsv_line, self.spk)]
+ f0 = self._load_f0(tsv_line)
+ f0 = self._preprocess_f0(f0, spk)
+ f0_min = min(f0_min, f0.min().item())
+ f0_max = max(f0_max, f0.max().item())
+ return f0_min, f0_max
+
+ def _compute_f0_stats(self):
+ from functools import partial
+
+ speaker_stats = defaultdict(partial(F0Stat, True))
+ for tsv_line in tqdm(self.tsv, desc="computing speaker stats"):
+ spk = self.spk2id[parse_speaker(tsv_line, self.spk)]
+ f0 = self._load_f0(tsv_line)
+ mask = f0 != 0
+ f0 = f0[mask] # compute stats only on voiced parts
+ speaker_stats[spk].update(f0)
+ return speaker_stats
+
+ def __getitem__(self, i):
+ x = self.km[i]
+ x = x.split(" ")
+ x = list(map(int, x))
+ x = torch.LongTensor(x)
+
+ gst = parse_style(self.tsv[i], self.gst)
+ gst = self.gst2id[gst]
+ spk = parse_speaker(self.tsv[i], self.spk)
+ spk = self.spk2id[spk]
+
+ f0_raw = self._load_f0(self.tsv[i])
+ f0 = self._preprocess_f0(f0_raw.clone(), spk)
+
+ f0 = F.interpolate(f0.unsqueeze(0).unsqueeze(0), x.shape[0])[0, 0]
+ f0_raw = F.interpolate(f0_raw.unsqueeze(0).unsqueeze(0), x.shape[0])[0, 0]
+
+ f0 = freq2bin(f0, f0_min=self.f0_min, f0_max=self.f0_max, bins=self.f0_bins)
+ f0 = F.one_hot(f0.long(), num_classes=len(self.f0_bins) + 1).float()
+ if self.f0_smoothing > 0:
+ f0 = torch.tensor(
+ gaussian_filter1d(f0.float().numpy(), sigma=self.f0_smoothing)
+ )
+ return x, f0, f0_raw, spk, gst
+
+
+def train(cfg):
+ device = "cuda:0"
+ # add 1 extra embedding for padding token, set the padding index to be the last token
+ # (tokens from the clustering start at index 0)
+ padding_token = cfg.n_tokens
+ collate_fn = Collator(padding_idx=padding_token)
+ train_ds = PitchDataset(
+ cfg.train_tsv,
+ cfg.train_km,
+ substring=cfg.substring,
+ spk=cfg.spk,
+ spk2id=cfg.spk2id,
+ gst=cfg.gst,
+ gst2id=cfg.gst2id,
+ f0_bins=cfg.f0_bins,
+ f0_bin_type=cfg.f0_bin_type,
+ f0_smoothing=cfg.f0_smoothing,
+ f0_norm=cfg.f0_norm,
+ f0_log=cfg.f0_log,
+ )
+ valid_ds = PitchDataset(
+ cfg.valid_tsv,
+ cfg.valid_km,
+ substring=cfg.substring,
+ spk=cfg.spk,
+ spk2id=cfg.spk2id,
+ gst=cfg.gst,
+ gst2id=cfg.gst2id,
+ f0_bins=cfg.f0_bins,
+ f0_bin_type=cfg.f0_bin_type,
+ f0_smoothing=cfg.f0_smoothing,
+ f0_norm=cfg.f0_norm,
+ f0_log=cfg.f0_log,
+ )
+ train_dl = DataLoader(
+ train_ds,
+ num_workers=0,
+ batch_size=cfg.batch_size,
+ shuffle=True,
+ collate_fn=collate_fn,
+ )
+ valid_dl = DataLoader(
+ valid_ds, num_workers=0, batch_size=16, shuffle=False, collate_fn=collate_fn
+ )
+
+ f0_min = train_ds.f0_min
+ f0_max = train_ds.f0_max
+ f0_bins = train_ds.f0_bins
+ speaker_stats = train_ds.speaker_stats
+
+ model = hydra.utils.instantiate(cfg["model"]).to(device)
+ model.setup_f0_stats(f0_min, f0_max, f0_bins, speaker_stats)
+
+ optimizer = hydra.utils.instantiate(cfg.optimizer, model.parameters())
+
+ best_loss = float("inf")
+ for epoch in range(cfg.epochs):
+ train_loss, train_l2_loss, train_l2_voiced_loss = run_epoch(
+ model, train_dl, optimizer, device, cfg, mode="train"
+ )
+ valid_loss, valid_l2_loss, valid_l2_voiced_loss = run_epoch(
+ model, valid_dl, None, device, cfg, mode="valid"
+ )
+ print(
+ f"[epoch {epoch}] train loss: {train_loss:.3f}, l2 loss: {train_l2_loss:.3f}, l2 voiced loss: {train_l2_voiced_loss:.3f}"
+ )
+ print(
+ f"[epoch {epoch}] valid loss: {valid_loss:.3f}, l2 loss: {valid_l2_loss:.3f}, l2 voiced loss: {valid_l2_voiced_loss:.3f}"
+ )
+ if valid_l2_voiced_loss < best_loss:
+ path = f"{os.getcwd()}/pitch_predictor.ckpt"
+ save_ckpt(model, path, cfg["model"], f0_min, f0_max, f0_bins, speaker_stats)
+ best_loss = valid_l2_voiced_loss
+ print(f"saved checkpoint: {path}")
+ print(f"[epoch {epoch}] best loss: {best_loss:.3f}")
+
+
+def run_epoch(model, loader, optimizer, device, cfg, mode):
+ if mode == "train":
+ model.train()
+ else:
+ model.eval()
+
+ epoch_loss = 0
+ l1 = 0
+ l1_voiced = 0
+ for x, f0_bin, f0_raw, spk_id, gst, mask, _ in tqdm(loader):
+ x, f0_bin, f0_raw, spk_id, gst, mask = (
+ x.to(device),
+ f0_bin.to(device),
+ f0_raw.to(device),
+ spk_id.to(device),
+ gst.to(device),
+ mask.to(device),
+ )
+ b, t, n_bins = f0_bin.shape
+ yhat = model(x, gst)
+ nonzero_mask = (f0_raw != 0).logical_and(mask)
+ yhat_raw = model.inference(x, spk_id, gst)
+ expanded_mask = mask.unsqueeze(-1).expand(-1, -1, n_bins)
+ if cfg.f0_pred == "mean":
+ loss = F.binary_cross_entropy(
+ yhat[expanded_mask], f0_bin[expanded_mask]
+ ).mean()
+ elif cfg.f0_pred == "argmax":
+ loss = F.cross_entropy(
+ rearrange(yhat, "b t d -> (b t) d"),
+ rearrange(f0_bin.argmax(-1), "b t -> (b t)"),
+ reduce=False,
+ )
+ loss = rearrange(loss, "(b t) -> b t", b=b, t=t)
+ loss = (loss * mask).sum() / mask.float().sum()
+ else:
+ raise NotImplementedError
+ l1 += F.l1_loss(yhat_raw[mask], f0_raw[mask]).item()
+ l1_voiced += F.l1_loss(yhat_raw[nonzero_mask], f0_raw[nonzero_mask]).item()
+ epoch_loss += loss.item()
+
+ if mode == "train":
+ loss.backward()
+ nn.utils.clip_grad_norm_(model.parameters(), 1.0)
+ optimizer.step()
+
+ print(f"{mode} example y: {f0_bin.argmax(-1)[0, 50:60].tolist()}")
+ print(f"{mode} example yhat: {yhat.argmax(-1)[0, 50:60].tolist()}")
+ print(f"{mode} example y: {f0_raw[0, 50:60].round().tolist()}")
+ print(f"{mode} example yhat: {yhat_raw[0, 50:60].round().tolist()}")
+ return epoch_loss / len(loader), l1 / len(loader), l1_voiced / len(loader)
+
+
+@hydra.main(config_path=dir_path, config_name="pitch_predictor.yaml")
+def main(cfg):
+ np.random.seed(1)
+ random.seed(1)
+ torch.manual_seed(1)
+ from hydra.core.hydra_config import HydraConfig
+
+ overrides = {
+ x.split("=")[0]: x.split("=")[1]
+ for x in HydraConfig.get().overrides.task
+ if "/" not in x
+ }
+ print(f"{cfg}")
+ train(cfg)
+
+
+if __name__ == "__main__":
+ main()
diff --git a/fairseq/examples/emotion_conversion/emotion_models/pitch_predictor.yaml b/fairseq/examples/emotion_conversion/emotion_models/pitch_predictor.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d2dbb862c358a0af58c0c4abf37fbdb537da6b44
--- /dev/null
+++ b/fairseq/examples/emotion_conversion/emotion_models/pitch_predictor.yaml
@@ -0,0 +1,64 @@
+train_tsv: "/denoising/emov/train.tsv"
+train_km: "/denoising/emov/train.km"
+valid_tsv: "/denoising/emov/valid.tsv"
+valid_km: "/denoising/emov/valid.km"
+
+n_tokens: 200
+batch_size: 64
+lr: 0.0001
+epochs: 1000
+
+substring: ""
+loss: "l2"
+spk: "parent_parent_name"
+gst: "emotion"
+
+f0_bins: 50
+f0_pred: "mean" # [argmax, mean]
+f0_smoothing: 0.1
+f0_norm: "mean"
+f0_log: false
+f0_bin_type: "adaptive" # [uniform, adaptive]
+
+spk2id:
+ bea: 0
+ jenie: 1
+ josh: 2
+ sam: 3
+
+gst2id:
+ amused: 0
+ angry: 1
+ disgusted: 2
+ neutral: 3
+ sleepy: 4
+
+optimizer:
+ _target_: torch.optim.Adam
+ lr: ${lr}
+
+model:
+ _target_: emotion_models.pitch_predictor.CnnPredictor
+ n_tokens: ${n_tokens}
+ emb_dim: 256
+ channels: 256
+ kernel: 5
+ dropout: 0.1
+ n_layers: 6
+ spk_emb: true
+ gst_emb: 8
+ n_bins: ${f0_bins}
+ f0_pred: ${f0_pred}
+ f0_log: ${f0_log}
+ f0_norm: ${f0_norm}
+
+hydra:
+ run:
+ dir: /checkpoint/felixkreuk/experiments/pitch_predictor/${hydra.job.override_dirname}
+ job:
+ config:
+ # configuration for the ${hydra.job.override_dirname} runtime variable
+ override_dirname:
+ kv_sep: '='
+ item_sep: ','
+ exclude_keys: ['train_tsv', 'train_km', 'valid_tsv', 'valid_km']
diff --git a/fairseq/examples/emotion_conversion/emotion_models/utils.py b/fairseq/examples/emotion_conversion/emotion_models/utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..4199c310f8ee6e50c81ca0645ea30cc672fd0b5c
--- /dev/null
+++ b/fairseq/examples/emotion_conversion/emotion_models/utils.py
@@ -0,0 +1,78 @@
+import torch
+
+
+class Stat:
+ def __init__(self, keep_raw=False):
+ self.x = 0.0
+ self.x2 = 0.0
+ self.z = 0.0 # z = logx
+ self.z2 = 0.0
+ self.n = 0.0
+ self.u = 0.0
+ self.keep_raw = keep_raw
+ self.raw = []
+
+ def update(self, new_x):
+ new_z = new_x.log()
+
+ self.x += new_x.sum()
+ self.x2 += (new_x**2).sum()
+ self.z += new_z.sum()
+ self.z2 += (new_z**2).sum()
+ self.n += len(new_x)
+ self.u += 1
+
+ if self.keep_raw:
+ self.raw.append(new_x)
+
+ @property
+ def mean(self):
+ return self.x / self.n
+
+ @property
+ def std(self):
+ return (self.x2 / self.n - self.mean**2) ** 0.5
+
+ @property
+ def mean_log(self):
+ return self.z / self.n
+
+ @property
+ def std_log(self):
+ return (self.z2 / self.n - self.mean_log**2) ** 0.5
+
+ @property
+ def n_frms(self):
+ return self.n
+
+ @property
+ def n_utts(self):
+ return self.u
+
+ @property
+ def raw_data(self):
+ assert self.keep_raw, "does not support storing raw data!"
+ return torch.cat(self.raw)
+
+
+class F0Stat(Stat):
+ def update(self, new_x):
+ # assume unvoiced frames are 0 and consider only voiced frames
+ if new_x is not None:
+ super().update(new_x[new_x != 0])
+
+
+class Accuracy:
+ def __init__(self):
+ self.y, self.yhat = [], []
+
+ def update(self, yhat, y):
+ self.yhat.append(yhat)
+ self.y.append(y)
+
+ def acc(self, tol):
+ yhat = torch.cat(self.yhat)
+ y = torch.cat(self.y)
+ acc = torch.abs(yhat - y) <= tol
+ acc = acc.float().mean().item()
+ return acc
diff --git a/fairseq/examples/emotion_conversion/fairseq_models/__init__.py b/fairseq/examples/emotion_conversion/fairseq_models/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..441bc03db49c78556796db5d0b680b0cc7a99ed7
--- /dev/null
+++ b/fairseq/examples/emotion_conversion/fairseq_models/__init__.py
@@ -0,0 +1,226 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+from fairseq import utils
+from fairseq.models import (
+ FairseqMultiModel,
+ register_model,
+ register_model_architecture,
+)
+from fairseq.models.transformer import (
+ Embedding,
+ base_architecture,
+)
+from fairseq.models.multilingual_transformer import (
+ MultilingualTransformerModel,
+ base_multilingual_architecture,
+)
+from fairseq.utils import safe_hasattr
+from collections import OrderedDict
+
+
+@register_model("multilingual_transformer_from_mbart")
+class MultilingualTransformerModelFromMbart(MultilingualTransformerModel):
+ @classmethod
+ def build_model(cls, args, task):
+ """Build a new model instance."""
+ from fairseq.tasks.multilingual_translation import MultilingualTranslationTask
+
+ assert isinstance(task, MultilingualTranslationTask)
+
+ # make sure all arguments are present in older models
+ base_multilingual_architecture(args)
+
+ if not safe_hasattr(args, "max_source_positions"):
+ args.max_source_positions = 1024
+ if not safe_hasattr(args, "max_target_positions"):
+ args.max_target_positions = 1024
+
+ src_langs = [lang_pair.split("-")[0] for lang_pair in task.model_lang_pairs]
+ tgt_langs = [lang_pair.split("-")[1] for lang_pair in task.model_lang_pairs]
+
+ if args.share_encoders:
+ args.share_encoder_embeddings = True
+ if args.share_decoders:
+ args.share_decoder_embeddings = True
+
+ def build_embedding(dictionary, embed_dim, path=None):
+ num_embeddings = len(dictionary)
+ padding_idx = dictionary.pad()
+ emb = Embedding(num_embeddings, embed_dim, padding_idx)
+ # if provided, load from preloaded dictionaries
+ if path:
+ embed_dict = utils.parse_embedding(path)
+ utils.load_embedding(embed_dict, dictionary, emb)
+ return emb
+
+ # build shared embeddings (if applicable)
+ shared_encoder_embed_tokens, shared_decoder_embed_tokens = None, None
+ if args.share_all_embeddings:
+ if args.encoder_embed_dim != args.decoder_embed_dim:
+ raise ValueError(
+ "--share-all-embeddings requires --encoder-embed-dim to match --decoder-embed-dim"
+ )
+ if args.decoder_embed_path and (
+ args.decoder_embed_path != args.encoder_embed_path
+ ):
+ raise ValueError(
+ "--share-all-embeddings not compatible with --decoder-embed-path"
+ )
+ shared_encoder_embed_tokens = FairseqMultiModel.build_shared_embeddings(
+ dicts=task.dicts,
+ langs=task.langs,
+ embed_dim=args.encoder_embed_dim,
+ build_embedding=build_embedding,
+ pretrained_embed_path=args.encoder_embed_path,
+ )
+ shared_decoder_embed_tokens = shared_encoder_embed_tokens
+ args.share_decoder_input_output_embed = True
+ else:
+ if args.share_encoder_embeddings:
+ shared_encoder_embed_tokens = FairseqMultiModel.build_shared_embeddings(
+ dicts=task.dicts,
+ langs=src_langs,
+ embed_dim=args.encoder_embed_dim,
+ build_embedding=build_embedding,
+ pretrained_embed_path=args.encoder_embed_path,
+ )
+ if args.share_decoder_embeddings:
+ shared_decoder_embed_tokens = FairseqMultiModel.build_shared_embeddings(
+ dicts=task.dicts,
+ langs=tgt_langs,
+ embed_dim=args.decoder_embed_dim,
+ build_embedding=build_embedding,
+ pretrained_embed_path=args.decoder_embed_path,
+ )
+
+ # encoders/decoders for each language
+ lang_encoders, lang_decoders = {}, {}
+
+ def get_encoder(lang):
+ if lang not in lang_encoders:
+ if shared_encoder_embed_tokens is not None:
+ encoder_embed_tokens = shared_encoder_embed_tokens
+ else:
+ encoder_embed_tokens = build_embedding(
+ task.dicts[lang],
+ args.encoder_embed_dim,
+ args.encoder_embed_path,
+ )
+ lang_encoders[lang] = MultilingualTransformerModel._get_module_class(
+ True, args, task.dicts[lang], encoder_embed_tokens, src_langs
+ )
+ return lang_encoders[lang]
+
+ def get_decoder(lang):
+ if lang not in lang_decoders:
+ if shared_decoder_embed_tokens is not None:
+ decoder_embed_tokens = shared_decoder_embed_tokens
+ else:
+ decoder_embed_tokens = build_embedding(
+ task.dicts[lang],
+ args.decoder_embed_dim,
+ args.decoder_embed_path,
+ )
+ lang_decoders[lang] = MultilingualTransformerModel._get_module_class(
+ False, args, task.dicts[lang], decoder_embed_tokens, tgt_langs
+ )
+ return lang_decoders[lang]
+
+ # shared encoders/decoders (if applicable)
+ shared_encoder, shared_decoder = None, None
+ if args.share_encoders:
+ shared_encoder = get_encoder(src_langs[0])
+ if args.share_decoders:
+ shared_decoder = get_decoder(tgt_langs[0])
+
+ encoders, decoders = OrderedDict(), OrderedDict()
+ for lang_pair, src, tgt in zip(task.model_lang_pairs, src_langs, tgt_langs):
+ encoders[lang_pair] = (
+ shared_encoder if shared_encoder is not None else get_encoder(src)
+ )
+ decoders[lang_pair] = (
+ shared_decoder if shared_decoder is not None else get_decoder(tgt)
+ )
+
+ return MultilingualTransformerModelFromMbart(encoders, decoders)
+
+ def load_state_dict(self, state_dict, strict=True, model_cfg=None):
+ state_dict_subset = state_dict.copy()
+ lang_pairs = set([x.split(".")[1] for x in state_dict.keys()])
+ finetune_mode = not any("neutral" in lp for lp in lang_pairs)
+
+ if finetune_mode:
+ # load a pre-trained mBART/BART model
+ # we need this code because mBART/BART are not of type FairseqMultiModel but FairseqModel
+ # so we hackishly load the weights by replicating them for all lang pairs
+ print("loading pre-trained BART")
+ self_state_dict = self.state_dict()
+ for k, v in state_dict.items():
+ for lang_pair in self.models:
+ new_key = k if "models." in k else f"models.{lang_pair}.{k}"
+ # print(new_key)
+ if self_state_dict[new_key].shape == v.shape:
+ state_dict_subset[new_key] = v
+ elif any(
+ w in k
+ for w in [
+ "encoder.embed_tokens.weight",
+ "decoder.embed_tokens.weight",
+ "decoder.output_projection.weight",
+ ]
+ ):
+ # why vocab_size - 5? because there are `vocab_size` tokens from the language
+ # and 5 additional tokens in the denoising task: eos,bos,pad,unk,mask.
+ # but in the translation task there are only `vocab_size` + 4 (no mask).
+ print(
+ f"{k}: {self_state_dict[new_key].shape} != {v.shape}",
+ end="",
+ flush=True,
+ )
+ vocab_size = v.shape[0] - 5
+ state_dict_subset[new_key] = self_state_dict[new_key]
+ state_dict_subset[new_key] = v[: vocab_size + 4]
+ print(f" => fixed by using first {vocab_size + 4} dims")
+ else:
+ raise ValueError("unable to load model due to mimatched dims!")
+ del state_dict_subset[k]
+ else:
+ print("loading pre-trained emotion translation model")
+ for k, _ in state_dict.items():
+ assert k.startswith("models.")
+ lang_pair = k.split(".")[1]
+ if lang_pair not in self.models:
+ del state_dict_subset[k]
+
+ super().load_state_dict(state_dict_subset, strict=strict, model_cfg=model_cfg)
+
+
+@register_model_architecture("transformer", "transformer_small")
+def transformer_small(args):
+ args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512)
+ args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 512)
+ args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 4)
+ args.encoder_layers = getattr(args, "encoder_layers", 3)
+ args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 512)
+ args.decoder_ffn_embed_dim = getattr(args, "decoder_ffn_embed_dim", 512)
+ args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 4)
+ args.decoder_layers = getattr(args, "decoder_layers", 3)
+ base_architecture(args)
+
+
+@register_model_architecture(
+ "multilingual_transformer_from_mbart", "multilingual_small"
+)
+def multilingual_small(args):
+ args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512)
+ args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 512)
+ args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 4)
+ args.encoder_layers = getattr(args, "encoder_layers", 3)
+ args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 512)
+ args.decoder_ffn_embed_dim = getattr(args, "decoder_ffn_embed_dim", 512)
+ args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 4)
+ args.decoder_layers = getattr(args, "decoder_layers", 3)
+ base_multilingual_architecture(args)
diff --git a/fairseq/examples/emotion_conversion/preprocess/__init__.py b/fairseq/examples/emotion_conversion/preprocess/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/fairseq/examples/emotion_conversion/preprocess/build_hifigan_manifest.py b/fairseq/examples/emotion_conversion/preprocess/build_hifigan_manifest.py
new file mode 100644
index 0000000000000000000000000000000000000000..29c0d79ceeca7f03087781394b907b089197b4e9
--- /dev/null
+++ b/fairseq/examples/emotion_conversion/preprocess/build_hifigan_manifest.py
@@ -0,0 +1,38 @@
+import torchaudio
+import argparse
+import json
+
+def main():
+ parser = argparse.ArgumentParser(description="example: python create_hifigan_manifest.py --tsv /checkpoint/felixkreuk/datasets/vctk/splits/vctk_16khz/train.tsv --km /checkpoint/felixkreuk/experiments/hubert/hubert_feats/vctk_16khz_km_100/train.km --km_type hubert_100km > ~/tmp/tmp_mani.txt")
+ parser.add_argument("--tsv", required=True, help="path to fairseq tsv file")
+ parser.add_argument("--km", required=True, help="path to a km file generated by HuBERT clustering")
+ parser.add_argument("--km_type", required=True, help="name of the codes in the output json (for example: 'cpc_100km')")
+ args = parser.parse_args()
+
+ km_lines = open(args.km, "r").readlines()
+ tsv_lines = open(args.tsv, "r").readlines()
+ assert len(km_lines) == len(tsv_lines) - 1, "tsv and km files are not of the same length!"
+
+ wav_root = tsv_lines[0].strip()
+ tsv_lines = tsv_lines[1:]
+
+ for tsv_line, km_line in zip(tsv_lines, km_lines):
+ tsv_line, km_line = tsv_line.strip(), km_line.strip()
+ wav_basename, wav_num_frames = tsv_line.split("\t")
+ wav_path = wav_root + "/" + wav_basename
+ wav_info = torchaudio.info(wav_path)
+ assert int(wav_num_frames) == wav_info.num_frames, "tsv duration and actual duration don't match!"
+ wav_duration = wav_info.num_frames / wav_info.sample_rate
+ manifest_line = {"audio": wav_path, "duration": wav_duration, args.km_type: km_line}
+ print(json.dumps(manifest_line))
+
+if __name__ == "__main__":
+ """
+ usage:
+ python create_hifigan_manifest.py \
+ --tsv /checkpoint/felixkreuk/datasets/vctk/manifests/vctk_16khz/valid.tsv \
+ --km /checkpoint/felixkreuk/datasets/vctk/manifests/vctk_16khz/hubert_km_100/valid.km \
+ --km_type hubert \
+ > /checkpoint/felixkreuk/datasets/vctk/manifests/vctk_16khz/hubert_km_100/hifigan_valid_manifest.txt
+ """
+ main()
diff --git a/fairseq/examples/emotion_conversion/preprocess/build_translation_manifests.py b/fairseq/examples/emotion_conversion/preprocess/build_translation_manifests.py
new file mode 100644
index 0000000000000000000000000000000000000000..d38454a71368146c44dc73f6f37037dca0dd026d
--- /dev/null
+++ b/fairseq/examples/emotion_conversion/preprocess/build_translation_manifests.py
@@ -0,0 +1,258 @@
+from glob import glob
+import argparse
+from collections import defaultdict, Counter
+from itertools import combinations, product, groupby
+from pathlib import Path
+import os
+from sklearn.utils import shuffle
+import numpy as np
+import random
+from shutil import copy
+from subprocess import check_call
+
+np.random.seed(42)
+random.seed(42)
+
+
+def get_fname(s):
+ return s.split("\t")[0]
+
+def get_emotion(s):
+ return get_fname(s).split("_")[0].split("/")[1].lower()
+
+def get_utt_id(s):
+ return get_fname(s).split(".")[0].split("_")[-1]
+
+def dedup(seq):
+ """ >> remove_repetitions("1 2 2 3 100 2 2 1")
+ '1 2 3 100 2 1' """
+ seq = seq.strip().split(" ")
+ result = seq[:1]
+ reps = []
+ rep_counter = 1
+ for k in seq[1:]:
+ if k != result[-1]:
+ result += [k]
+ reps += [rep_counter]
+ rep_counter = 1
+ else:
+ rep_counter += 1
+ reps += [rep_counter]
+ assert len(reps) == len(result) and sum(reps) == len(seq)
+ return " ".join(result) + "\n" #, reps
+
+def remove_under_k(seq, k):
+ """ remove tokens that repeat less then k times in a row
+ >> remove_under_k("a a a a b c c c", 1) ==> a a a a c c c """
+ seq = seq.strip().split(" ")
+ result = []
+
+ freqs = [(k,len(list(g))) for k, g in groupby(seq)]
+ for c, f in freqs:
+ if f > k:
+ result += [c for _ in range(f)]
+ return " ".join(result) + "\n" #, reps
+
+
+def call(cmd):
+ print(cmd)
+ check_call(cmd, shell=True)
+
+
+def denoising_preprocess(path, lang, dict):
+ bin = 'fairseq-preprocess'
+ cmd = [
+ bin,
+ f'--trainpref {path}/train.{lang} --validpref {path}/valid.{lang} --testpref {path}/test.{lang}',
+ f'--destdir {path}/tokenized/{lang}',
+ '--only-source',
+ '--task multilingual_denoising',
+ '--workers 40',
+ ]
+ if dict != "":
+ cmd += [f'--srcdict {dict}']
+ cmd = " ".join(cmd)
+ call(cmd)
+
+
+def translation_preprocess(path, src_lang, trg_lang, dict, only_train=False):
+ bin = 'fairseq-preprocess'
+ cmd = [
+ bin,
+ f'--source-lang {src_lang} --target-lang {trg_lang}',
+ f'--trainpref {path}/train',
+ f'--destdir {path}/tokenized',
+ '--workers 40',
+ ]
+ if not only_train:
+ cmd += [f'--validpref {path}/valid --testpref {path}/test']
+ if dict != "":
+ cmd += [
+ f'--srcdict {dict}',
+ f'--tgtdict {dict}',
+ ]
+ cmd = " ".join(cmd)
+ call(cmd)
+
+
+def load_tsv_km(tsv_path, km_path):
+ assert tsv_path.exists() and km_path.exists()
+ tsv_lines = open(tsv_path, "r").readlines()
+ root, tsv_lines = tsv_lines[0], tsv_lines[1:]
+ km_lines = open(km_path, "r").readlines()
+ assert len(tsv_lines) == len(km_lines), ".tsv and .km should be the same length!"
+ return root, tsv_lines, km_lines
+
+
+def main():
+ desc = """
+ this script takes as input .tsv and .km files for EMOV dataset, and a pairs of emotions.
+ it generates parallel .tsv and .km files for these emotions. for exmaple:
+ ❯ python build_emov_translation_manifests.py \
+ /checkpoint/felixkreuk/datasets/emov/manifests/emov_16khz/train.tsv \
+ /checkpoint/felixkreuk/datasets/emov/manifests/emov_16khz/emov_16khz_km_100/train.km \
+ ~/tmp/emov_pairs \
+ --src-emotion amused --trg-emotion neutral \
+ --dedup --shuffle --cross-speaker --dry-run
+ """
+ parser = argparse.ArgumentParser(description=desc)
+ parser.add_argument("data", type=Path, help="path to a dir containing .tsv and .km files containing emov dataset")
+ parser.add_argument("output_path", type=Path, help="output directory with the manifests will be created")
+ parser.add_argument("-cs", "--cross-speaker", action='store_true', help="if set then translation will occur also between speakers, meaning the same sentence can be translated between different speakers (default: false)")
+ parser.add_argument("-dd", "--dedup", action='store_true', help="remove repeated tokens (example: 'aaabc=>abc')")
+ parser.add_argument("-sh", "--shuffle", action='store_true', help="shuffle the data")
+ parser.add_argument("-ae", "--autoencode", action='store_true', help="include training pairs from the same emotion (this includes examples of the same sentence uttered by different people and examples where the src and trg are the exact same seq)")
+ parser.add_argument("-dr", "--dry-run", action='store_true', help="don't write anything to disk")
+ parser.add_argument("-zs", "--zero-shot", action='store_true', help="if true, the denoising task will train on the same splits as the translation task (split by utterance id). if false, the denoising task will train on randomly sampled splits (not split by utterance id)")
+ parser.add_argument("--km-ext", default="km", help="")
+ parser.add_argument("--dict", default="/checkpoint/felixkreuk/datasets/emov/manifests/emov_16khz/fairseq.dict.txt", help="")
+ args = parser.parse_args()
+ SPEAKERS = ["bea", "jenie", "josh", "sam", "SAME"]
+ EMOTIONS = ['neutral', 'amused', 'angry', 'disgusted', 'sleepy']
+
+ suffix = ""
+ if args.cross_speaker: suffix += "_cross-speaker"
+ if args.dedup: suffix += "_dedup"
+ translation_suffix = ""
+ if args.autoencode: translation_suffix += "_autoencode"
+ denoising_suffix = ""
+ denoising_suffix += "_zeroshot" if args.zero_shot else "_nonzeroshot"
+
+ translation_dir = Path(args.output_path) / ("emov_multilingual_translation" + suffix + translation_suffix)
+ os.makedirs(translation_dir, exist_ok=True)
+ denoising_dir = Path(args.output_path) / ("emov_multilingual_denoising" + suffix + denoising_suffix)
+ os.makedirs(denoising_dir, exist_ok=True)
+
+ denoising_data = [p.name for p in (args.data / "denoising").glob("*") if "emov" not in p.name]
+
+ for split in ["train", "valid", "test"]:
+ root, tsv_lines, km_lines = load_tsv_km(
+ tsv_path = args.data / "denoising" / "emov" / f"{split}.tsv",
+ km_path = args.data / "denoising" / "emov" / f"{split}.{args.km_ext}"
+ )
+
+ # generate data for the multilingual denoising task
+ for EMOTION in EMOTIONS:
+ print("---")
+ print(split)
+ print(f"denoising: {EMOTION}")
+ emotion_tsv, emotion_km = [], []
+ for tsv_line, km_line in zip(tsv_lines, km_lines):
+ if EMOTION.lower() in tsv_line.lower():
+ km_line = km_line if not args.dedup else dedup(km_line)
+ emotion_tsv.append(tsv_line)
+ emotion_km.append(km_line)
+ print(f"{len(emotion_km)} samples")
+ open(denoising_dir / f"files.{split}.{EMOTION}", "w").writelines([root] + emotion_tsv)
+ open(denoising_dir / f"{split}.{EMOTION}", "w").writelines(emotion_km)
+
+ for data in denoising_data:
+ with open(args.data / "denoising" / data / f"{split}.{args.km_ext}", "r") as f1:
+ with open(denoising_dir / f"{split}.{data}", "w") as f2:
+ f2.writelines([l if not args.dedup else dedup(l) for l in f1.readlines()])
+
+ # start of translation preprocessing
+ root, tsv_lines, km_lines = load_tsv_km(
+ tsv_path = args.data / "translation" / f"{split}.tsv",
+ km_path = args.data / "translation" / f"{split}.{args.km_ext}"
+ )
+
+ # generate data for the multilingual translation task
+ for SRC_EMOTION in EMOTIONS:
+ TRG_EMOTIONS = EMOTIONS if args.autoencode else set(EMOTIONS) - set([SRC_EMOTION])
+ for TRG_EMOTION in TRG_EMOTIONS:
+ # when translating back to the same emotion - we dont want these emotion
+ # pairs to be part of the validation/test sets (because its not really emotion conversino)
+ # if SRC_EMOTION == TRG_EMOTION and split in ["valid", "test"]: continue
+ print("---")
+ print(split)
+ print(f"src emotions: {SRC_EMOTION}\ntrg emotions: {TRG_EMOTION}")
+
+ # create a dictionary with the following structure:
+ # output[SPEAKER][UTT_ID] = list with indexes of line from the tsv file
+ # that match the speaker and utterance id. for exmaple:
+ # output = {'sam': {'0493': [875, 1608, 1822], ...}, ...}
+ # meaning, for speaker 'sam', utterance id '0493', the indexes in tsv_lines
+ # are 875, 1608, 1822
+ spkr2utts = defaultdict(lambda: defaultdict(list))
+ for i, tsv_line in enumerate(tsv_lines):
+ speaker = tsv_line.split("/")[0]
+ if args.cross_speaker: speaker = "SAME"
+ assert speaker in SPEAKERS, "unknown speaker! make sure the .tsv contains EMOV data"
+ utt_id = get_utt_id(tsv_line)
+ spkr2utts[speaker][utt_id].append(i)
+
+ # create a tsv and km files with all the combinations for translation
+ src_tsv, trg_tsv, src_km, trg_km = [], [], [], []
+ for speaker, utt_ids in spkr2utts.items():
+ for utt_id, indices in utt_ids.items():
+ # generate all pairs
+ pairs = [(x,y) for x in indices for y in indices]
+ # self-translation
+ if SRC_EMOTION == TRG_EMOTION:
+ pairs = [(x,y) for (x,y) in pairs if x == y]
+ # filter according to src and trg emotions
+ pairs = [(x,y) for (x,y) in pairs
+ if get_emotion(tsv_lines[x]) == SRC_EMOTION and get_emotion(tsv_lines[y]) == TRG_EMOTION]
+
+ for idx1, idx2 in pairs:
+ assert get_utt_id(tsv_lines[idx1]) == get_utt_id(tsv_lines[idx2])
+ src_tsv.append(tsv_lines[idx1])
+ trg_tsv.append(tsv_lines[idx2])
+ km_line_idx1 = km_lines[idx1]
+ km_line_idx2 = km_lines[idx2]
+ km_line_idx1 = km_line_idx1 if not args.dedup else dedup(km_line_idx1)
+ km_line_idx2 = km_line_idx2 if not args.dedup else dedup(km_line_idx2)
+ src_km.append(km_line_idx1)
+ trg_km.append(km_line_idx2)
+ assert len(src_tsv) == len(trg_tsv) == len(src_km) == len(trg_km)
+ print(f"{len(src_tsv)} pairs")
+
+ if len(src_tsv) == 0:
+ raise Exception("ERROR: generated 0 pairs!")
+
+ if args.dry_run: continue
+
+ # create files
+ os.makedirs(translation_dir / f"{SRC_EMOTION}-{TRG_EMOTION}", exist_ok=True)
+ open(translation_dir / f"{SRC_EMOTION}-{TRG_EMOTION}" / f"files.{split}.{SRC_EMOTION}", "w").writelines([root] + src_tsv)
+ open(translation_dir / f"{SRC_EMOTION}-{TRG_EMOTION}" / f"files.{split}.{TRG_EMOTION}", "w").writelines([root] + trg_tsv)
+ open(translation_dir / f"{SRC_EMOTION}-{TRG_EMOTION}" / f"{split}.{SRC_EMOTION}", "w").writelines(src_km)
+ open(translation_dir / f"{SRC_EMOTION}-{TRG_EMOTION}" / f"{split}.{TRG_EMOTION}", "w").writelines(trg_km)
+
+
+ # fairseq-preprocess the denoising data
+ for EMOTION in EMOTIONS + denoising_data:
+ denoising_preprocess(denoising_dir, EMOTION, args.dict)
+ os.system(f"cp {args.dict} {denoising_dir}/tokenized/dict.txt")
+
+ # fairseq-preprocess the translation data
+ os.makedirs(translation_dir / "tokenized", exist_ok=True)
+ for SRC_EMOTION in EMOTIONS:
+ TRG_EMOTIONS = EMOTIONS if args.autoencode else set(EMOTIONS) - set([SRC_EMOTION])
+ for TRG_EMOTION in TRG_EMOTIONS:
+ translation_preprocess(translation_dir / f"{SRC_EMOTION}-{TRG_EMOTION}", SRC_EMOTION, TRG_EMOTION, args.dict)#, only_train=SRC_EMOTION==TRG_EMOTION)
+ os.system(f"cp -rf {translation_dir}/**/tokenized/* {translation_dir}/tokenized")
+
+if __name__ == "__main__":
+ main()
diff --git a/fairseq/examples/emotion_conversion/preprocess/create_core_manifest.py b/fairseq/examples/emotion_conversion/preprocess/create_core_manifest.py
new file mode 100644
index 0000000000000000000000000000000000000000..b55740e00becb6fc45d2d3c5396dc6ca85790e2a
--- /dev/null
+++ b/fairseq/examples/emotion_conversion/preprocess/create_core_manifest.py
@@ -0,0 +1,91 @@
+from pathlib import Path
+import os
+import sys
+import subprocess
+import argparse
+from datetime import datetime
+import logging
+
+logging.basicConfig(
+ level=logging.INFO,
+ format='%(asctime)s [%(levelname)s] %(message)s',
+ handlers=[logging.FileHandler('debug.log'), logging.StreamHandler()]
+)
+logger = logging.getLogger(__name__)
+
+
+def verify_dict_size(km, dict):
+ logger.info(f"verifying: {km}")
+ dict_size = len(open(dict, "r").readlines())
+ km_vocab = set(open(km, "r").read().replace("\n", " ").split(" "))
+ if "" in km_vocab: km_vocab.remove("")
+ km_vocab_size = len(km_vocab)
+ return dict_size == km_vocab_size
+
+
+def verify_files_exist(l):
+ for f in l:
+ if not f.exists():
+ logging.error(f"{f} doesn't exist!")
+ return False
+ return True
+
+
+def run_cmd(cmd, print_output=True):
+ try:
+ out = subprocess.check_output(cmd, stderr=subprocess.STDOUT, universal_newlines=True, shell=True)
+ if print_output:
+ logger.info(f"command output:\n{out}")
+ return out
+ except subprocess.CalledProcessError as grepexc:
+ logger.info(f"error executing command!:\n{cmd}")
+ logger.info(grepexc.output)
+
+def main():
+ parser = argparse.ArgumentParser()
+ parser.add_argument("--tsv", default="/checkpoint/felixkreuk/datasets/emov/manifests/emov_16khz/data.tsv", type=Path)
+ parser.add_argument("--emov-km", required=True, type=Path)
+ parser.add_argument("--km", nargs='+', required=True, type=Path)
+ parser.add_argument("--seed", type=int, default=1)
+ parser.add_argument("--dict", default="/checkpoint/felixkreuk/datasets/emov/manifests/emov_16khz/fairseq.dict.txt")
+ parser.add_argument("--manifests-dir", type=Path, default="/checkpoint/felixkreuk/datasets/emov/manifests/emov_16khz")
+ args = parser.parse_args()
+
+ manifests_dir = args.manifests_dir
+ date = datetime.now().strftime('%d%m%y')
+ outdir = manifests_dir / f"{date}"
+
+ # verify input and create folders
+ all_kms = args.km + [args.emov_km]
+ assert verify_files_exist(all_kms), "make sure the km dir contains: train-clean-all.km, blizzard2013.km, data.km"
+ for codes in all_kms:
+ assert verify_dict_size(codes, args.dict), "dict argument doesn't match the vocabulary of the km file!"
+ assert not outdir.exists(), "data dir already exists!"
+ outdir.mkdir(parents=True, exist_ok=True)
+
+ logger.info("generating denoising split (emov)")
+ run_cmd(f"python preprocess/split_km_tsv.py {args.tsv} {args.emov_km} --destdir {outdir}/denoising/emov -sh --seed {args.seed}")
+ for codes in args.km:
+ codes_name = os.path.basename(codes)
+ run_cmd(f"python preprocess/split_km.py {codes} --destdir {outdir}/denoising/{codes_name} -sh --seed {args.seed}")
+
+ logger.info("generating translation split")
+ run_cmd(f"python preprocess/split_emov_km_tsv_by_uttid.py {args.tsv} {args.emov_km} --destdir {outdir}/translation --seed {args.seed}")
+
+ emov_code_name = os.path.basename(args.emov_km)
+ logger.info("generating hifigan split")
+ run_cmd(
+ f"mkdir -p {outdir}/hifigan &&"
+ f"python preprocess/build_hifigan_manifest.py --km_type hubert --tsv {outdir}/denoising/emov/train.tsv --km {outdir}/denoising/emov/train.km > {outdir}/hifigan/train.txt &&"
+ f"python preprocess/build_hifigan_manifest.py --km_type hubert --tsv {outdir}/denoising/emov/valid.tsv --km {outdir}/denoising/emov/valid.km > {outdir}/hifigan/valid.txt &&"
+ f"python preprocess/build_hifigan_manifest.py --km_type hubert --tsv {outdir}/denoising/emov/test.tsv --km {outdir}/denoising/emov/test.km > {outdir}/hifigan/test.txt"
+ )
+
+ logger.info("generating fairseq manifests")
+ run_cmd(f"python preprocess/build_translation_manifests.py {outdir} {outdir}/fairseq-data -dd -cs --dict {args.dict}")
+
+ logger.info(f"finished processing data at:\n{outdir}")
+
+
+if __name__ == "__main__":
+ main()
diff --git a/fairseq/examples/emotion_conversion/preprocess/extract_f0.py b/fairseq/examples/emotion_conversion/preprocess/extract_f0.py
new file mode 100644
index 0000000000000000000000000000000000000000..4204aa4db1c39fac6a12a865741b5acca702230f
--- /dev/null
+++ b/fairseq/examples/emotion_conversion/preprocess/extract_f0.py
@@ -0,0 +1,57 @@
+import argparse
+from tqdm import tqdm
+from multiprocessing import Manager, Pool
+
+from scipy.io.wavfile import read
+from librosa.util import normalize
+import numpy as np
+import amfm_decompy.pYAAPT as pYAAPT
+import amfm_decompy.basic_tools as basic
+
+MAX_WAV_VALUE = 32768.0
+
+parser = argparse.ArgumentParser(description="")
+parser.add_argument("tsv", help="")
+parser.add_argument("--extractor", choices=["crepe", "pyaapt"], default="pyaapt", help="")
+parser.add_argument("--interp", action="store_true", help="")
+parser.add_argument("--n_workers", type=int, default=40, help="")
+args = parser.parse_args()
+
+tsv_lines = open(args.tsv, "r").readlines()
+root, tsv_lines = tsv_lines[0].strip(), tsv_lines[1:]
+
+
+def extract_f0(tsv_line):
+ wav_path, _ = tsv_line.split("\t")
+ wav_path = root.strip() + "/" + wav_path
+ sr, wav = read(wav_path)
+ wav = wav / MAX_WAV_VALUE
+ wav = normalize(wav) * 0.95
+
+ if args.extractor == "pyaapt":
+ frame_length = 20.0
+ pad = int(frame_length / 1000 * sr) // 2
+ wav = np.pad(wav.squeeze(), (pad, pad), "constant", constant_values=0)
+ signal = basic.SignalObj(wav, sr)
+ pitch = pYAAPT.yaapt(
+ signal,
+ **{
+ 'frame_length': frame_length,
+ 'frame_space': 5.0,
+ 'nccf_thresh1': 0.25,
+ 'tda_frame_length': 25.0
+ })
+ pitch = pitch.samp_interp[None, None, :] if args.interp else pitch.samp_values[None, None, :]
+ pitch = pitch[0, 0]
+ f0_path = wav_path.replace(".wav", ".yaapt")
+ f0_path += ".interp.f0" if args.interp else ".f0"
+ np.save(f0_path, pitch)
+
+
+def main():
+ with Pool(args.n_workers) as p:
+ r = list(tqdm(p.imap(extract_f0, tsv_lines), total=len(tsv_lines)))
+
+
+if __name__ == "__main__":
+ main()
diff --git a/fairseq/examples/emotion_conversion/preprocess/process_km.py b/fairseq/examples/emotion_conversion/preprocess/process_km.py
new file mode 100644
index 0000000000000000000000000000000000000000..864a022105f3c6f82b53a4553a438e76e211f40d
--- /dev/null
+++ b/fairseq/examples/emotion_conversion/preprocess/process_km.py
@@ -0,0 +1,40 @@
+import sys
+import argparse
+from tqdm import tqdm
+from build_emov_translation_manifests import dedup, remove_under_k
+
+
+if __name__ == "__main__":
+ """
+ this is a standalone script to process a km file
+ specifically, to dedup or remove tokens that repeat less
+ than k times in a row
+ """
+ parser = argparse.ArgumentParser(description="")
+ parser.add_argument("km", type=str, help="path to km file")
+ parser.add_argument("--dedup", action='store_true')
+ parser.add_argument("--remove-under-k", type=int, default=0)
+ parser.add_argument("--output", default=None)
+ args = parser.parse_args()
+
+ if not args.dedup and args.remove_under_k == 0:
+ print("nothing to do! quitting...")
+ sys.exit(0)
+
+ km = open(args.km, "r").readlines()
+ out = []
+ for line in tqdm(km):
+ if args.remove_under_k > 0:
+ line = remove_under_k(line, args.remove_under_k)
+ if args.dedup:
+ line = dedup(line)
+ out.append(line)
+
+ path = args.km if args.output is None else args.output
+ if args.remove_under_k > 0:
+ path = path.replace(".km", f"-k{args.remove_under_k}.km")
+ if args.dedup:
+ path = path.replace(".km", f"-deduped.km")
+
+ open(path, "w").writelines(out)
+ print(f"written to {path}")
diff --git a/fairseq/examples/emotion_conversion/preprocess/split_emov_km_tsv_by_uttid.py b/fairseq/examples/emotion_conversion/preprocess/split_emov_km_tsv_by_uttid.py
new file mode 100644
index 0000000000000000000000000000000000000000..94221afba711eb106ecd2275e5862a9a8b6148fd
--- /dev/null
+++ b/fairseq/examples/emotion_conversion/preprocess/split_emov_km_tsv_by_uttid.py
@@ -0,0 +1,70 @@
+from pathlib import Path
+import os
+import sys
+import argparse
+import random
+import numpy as np
+from tqdm import tqdm
+from sklearn.model_selection import train_test_split
+from build_translation_manifests import get_utt_id
+
+
+def train_val_test_split(tsv_lines, km_lines, valid_percent, test_percent, seed=42):
+ utt_ids = list(sorted(set([get_utt_id(x) for x in tsv_lines])))
+ utt_ids, valid_utt_ids, _, _ = train_test_split(utt_ids, utt_ids, test_size=valid_percent, shuffle=True, random_state=seed)
+ train_utt_ids, test_utt_ids, _, _ = train_test_split(utt_ids, utt_ids, test_size=test_percent, shuffle=True, random_state=seed)
+
+ train_idx = [i for i, line in enumerate(tsv_lines) if get_utt_id(line) in train_utt_ids]
+ valid_idx = [i for i, line in enumerate(tsv_lines) if get_utt_id(line) in valid_utt_ids]
+ test_idx = [i for i, line in enumerate(tsv_lines) if get_utt_id(line) in test_utt_ids]
+
+ train_tsv, train_km = [tsv_lines[i] for i in train_idx], [km_lines[i] for i in train_idx]
+ valid_tsv, valid_km = [tsv_lines[i] for i in valid_idx], [km_lines[i] for i in valid_idx]
+ test_tsv, test_km = [tsv_lines[i] for i in test_idx], [km_lines[i] for i in test_idx]
+
+ print(f"train {len(train_km)}")
+ print(f"valid {len(valid_km)}")
+ print(f"test {len(test_km)}")
+
+ return train_tsv, train_km, valid_tsv, valid_km, test_tsv, test_km
+
+
+if __name__ == "__main__":
+ """
+ this is a standalone script to process a km file
+ specifically, to dedup or remove tokens that repeat less
+ than k times in a row
+ """
+ parser = argparse.ArgumentParser(description="")
+ parser.add_argument("tsv", type=str, help="path to tsv file")
+ parser.add_argument("km", type=str, help="path to km file")
+ parser.add_argument("--destdir", required=True, type=str)
+ parser.add_argument("--valid-percent", type=float, default=0.05, help="percent to allocate to validation set")
+ parser.add_argument("--test-percent", type=float, default=0.05, help="percent to allocate to test set")
+ parser.add_argument("--seed", type=int, default=42, help="")
+ args = parser.parse_args()
+
+ np.random.seed(args.seed)
+ random.seed(args.seed)
+
+ os.makedirs(args.destdir, exist_ok=True)
+ km = open(args.km, "r").readlines()
+ tsv = open(args.tsv, "r").readlines()
+ root, tsv = tsv[0], tsv[1:]
+
+ assert args.tsv.endswith(".tsv") and args.km.endswith(".km")
+ assert len(tsv) == len(km)
+
+ train_tsv, train_km, valid_tsv, valid_km, test_tsv, test_km = train_val_test_split(tsv, km, args.valid_percent, args.test_percent, args.seed)
+
+ assert len(train_tsv) + len(valid_tsv) + len(test_tsv) == len(tsv)
+ assert len(train_tsv) == len(train_km) and len(valid_tsv) == len(valid_km) and len(test_tsv) == len(test_km)
+
+ dir = Path(args.destdir)
+ open(dir / f"train.tsv", "w").writelines([root] + train_tsv)
+ open(dir / f"valid.tsv", "w").writelines([root] + valid_tsv)
+ open(dir / f"test.tsv", "w").writelines([root] + test_tsv)
+ open(dir / f"train.km", "w").writelines(train_km)
+ open(dir / f"valid.km", "w").writelines(valid_km)
+ open(dir / f"test.km", "w").writelines(test_km)
+ print("done")
diff --git a/fairseq/examples/emotion_conversion/preprocess/split_km.py b/fairseq/examples/emotion_conversion/preprocess/split_km.py
new file mode 100644
index 0000000000000000000000000000000000000000..d145fc2bdefe72d5ce90b56f13d4e9f19a001d31
--- /dev/null
+++ b/fairseq/examples/emotion_conversion/preprocess/split_km.py
@@ -0,0 +1,50 @@
+from pathlib import Path
+import os
+import argparse
+import random
+import numpy as np
+from sklearn.utils import shuffle
+
+
+if __name__ == "__main__":
+ """
+ this is a standalone script to process a km file
+ specifically, to dedup or remove tokens that repeat less
+ than k times in a row
+ """
+ parser = argparse.ArgumentParser(description="")
+ parser.add_argument("km", type=str, help="path to km file")
+ parser.add_argument("--destdir", required=True, type=str)
+ parser.add_argument("--valid-percent", type=float, default=0.05, help="percent to allocate to validation set")
+ parser.add_argument("--test-percent", type=float, default=0.05, help="percent to allocate to test set")
+ parser.add_argument("-sh", "--shuffle", action="store_true", help="path to km file")
+ parser.add_argument("--seed", type=int, default=42, help="")
+ args = parser.parse_args()
+
+ np.random.seed(args.seed)
+ random.seed(args.seed)
+
+ os.makedirs(args.destdir, exist_ok=True)
+ km = open(args.km, "r").readlines()
+
+ if args.shuffle:
+ km = shuffle(km)
+ print(f"shuffled")
+
+ N = len(km)
+ N_tt = int(N * args.test_percent)
+ N_cv = int(N * args.valid_percent)
+ N_tr = N - N_tt - N_cv
+
+ train_km = km[:N_tr]
+ valid_km = km[N_tr:N_tr + N_cv]
+ test_km = km[N_tr + N_cv:]
+
+ dir = Path(args.destdir)
+ open(dir / f"train.km", "w").writelines(train_km)
+ open(dir / f"valid.km", "w").writelines(valid_km)
+ open(dir / f"test.km", "w").writelines(test_km)
+ print(f"train: {len(train_km)}")
+ print(f"valid: {len(valid_km)}")
+ print(f"test: {len(test_km)}")
+ print("done")
diff --git a/fairseq/examples/emotion_conversion/preprocess/split_km_tsv.py b/fairseq/examples/emotion_conversion/preprocess/split_km_tsv.py
new file mode 100644
index 0000000000000000000000000000000000000000..2113aa718d3bb75c08e5f9d01a79c2ff7ff2d4f4
--- /dev/null
+++ b/fairseq/examples/emotion_conversion/preprocess/split_km_tsv.py
@@ -0,0 +1,65 @@
+from pathlib import Path
+import os
+import argparse
+import random
+import numpy as np
+from sklearn.utils import shuffle
+
+
+if __name__ == "__main__":
+ """
+ this is a standalone script to process a km file
+ specifically, to dedup or remove tokens that repeat less
+ than k times in a row
+ """
+ parser = argparse.ArgumentParser(description="")
+ parser.add_argument("tsv", type=str, help="path to tsv file")
+ parser.add_argument("km", type=str, help="path to km file")
+ parser.add_argument("--destdir", required=True, type=str)
+ parser.add_argument("--valid-percent", type=float, default=0.05, help="percent to allocate to validation set")
+ parser.add_argument("--test-percent", type=float, default=0.05, help="percent to allocate to test set")
+ parser.add_argument("-sh", "--shuffle", action="store_true", help="path to km file")
+ parser.add_argument("--seed", type=int, default=42, help="")
+ args = parser.parse_args()
+
+ np.random.seed(args.seed)
+ random.seed(args.seed)
+
+ os.makedirs(args.destdir, exist_ok=True)
+ km = open(args.km, "r").readlines()
+ tsv = open(args.tsv, "r").readlines()
+ root, tsv = tsv[0], tsv[1:]
+
+ assert args.tsv.endswith(".tsv") and args.km.endswith(".km")
+ assert len(tsv) == len(km)
+
+ if args.shuffle:
+ tsv, km = shuffle(tsv, km)
+ print(f"shuffled")
+
+ N = len(tsv)
+ N_tt = int(N * args.test_percent)
+ N_cv = int(N * args.valid_percent)
+ N_tr = N - N_tt - N_cv
+
+ train_tsv = tsv[:N_tr]
+ valid_tsv = tsv[N_tr:N_tr + N_cv]
+ test_tsv = tsv[N_tr + N_cv:]
+ train_km = km[:N_tr]
+ valid_km = km[N_tr:N_tr + N_cv]
+ test_km = km[N_tr + N_cv:]
+
+ assert len(train_tsv) + len(valid_tsv) + len(test_tsv) == len(tsv)
+ assert len(train_tsv) == len(train_km) and len(valid_tsv) == len(valid_km) and len(test_tsv) == len(test_km)
+
+ dir = Path(args.destdir)
+ open(dir / f"train.tsv", "w").writelines([root] + train_tsv)
+ open(dir / f"valid.tsv", "w").writelines([root] + valid_tsv)
+ open(dir / f"test.tsv", "w").writelines([root] + test_tsv)
+ open(dir / f"train.km", "w").writelines(train_km)
+ open(dir / f"valid.km", "w").writelines(valid_km)
+ open(dir / f"test.km", "w").writelines(test_km)
+ print(f"train: {len(train_km)}")
+ print(f"valid: {len(valid_km)}")
+ print(f"test: {len(test_km)}")
+ print("done")
diff --git a/fairseq/examples/emotion_conversion/requirements.txt b/fairseq/examples/emotion_conversion/requirements.txt
new file mode 100644
index 0000000000000000000000000000000000000000..fc94c5a5470833d45468745a78fea1844e746dbb
--- /dev/null
+++ b/fairseq/examples/emotion_conversion/requirements.txt
@@ -0,0 +1,11 @@
+scipy
+einops
+amfm_decompy
+joblib
+numba
+decorator
+requests
+appdirs
+packaging
+six
+sklearn
diff --git a/fairseq/examples/emotion_conversion/synthesize.py b/fairseq/examples/emotion_conversion/synthesize.py
new file mode 100644
index 0000000000000000000000000000000000000000..327fdaf4ea38a08d63cfb2d9b9a18d27e8e326c7
--- /dev/null
+++ b/fairseq/examples/emotion_conversion/synthesize.py
@@ -0,0 +1,322 @@
+import logging
+import argparse
+import random
+import sys
+import os
+import numpy as np
+import torch
+import soundfile as sf
+import shutil
+import librosa
+import json
+from pathlib import Path
+from tqdm import tqdm
+import amfm_decompy.basic_tools as basic
+import amfm_decompy.pYAAPT as pYAAPT
+
+dir_path = os.path.dirname(__file__)
+resynth_path = os.path.dirname(os.path.abspath(__file__)) + "/speech-resynthesis"
+sys.path.append(resynth_path)
+
+from models import CodeGenerator
+from inference import scan_checkpoint, load_checkpoint, generate
+from emotion_models.pitch_predictor import load_ckpt as load_pitch_predictor
+from emotion_models.duration_predictor import load_ckpt as load_duration_predictor
+from dataset import load_audio, MAX_WAV_VALUE, parse_style, parse_speaker, EMOV_SPK2ID, EMOV_STYLE2ID
+
+
+logging.basicConfig(
+ level=logging.INFO,
+ format='%(asctime)s [%(levelname)s] %(message)s',
+ handlers=[logging.FileHandler('debug.log'), logging.StreamHandler()]
+)
+logger = logging.getLogger(__name__)
+
+
+class AttrDict(dict):
+ def __init__(self, *args, **kwargs):
+ super(AttrDict, self).__init__(*args, **kwargs)
+ self.__dict__ = self
+
+
+def parse_generation_file(fname):
+ lines = open(fname).read()
+ lines = lines.split('\n')
+
+ results = {}
+ for l in lines:
+ if len(l) == 0:
+ continue
+
+ if l[0] == 'H':
+ parts = l[2:].split('\t')
+ if len(parts) == 2:
+ sid, utt = parts
+ else:
+ sid, _, utt = parts
+ sid = int(sid)
+ utt = [int(x) for x in utt.split()]
+ if sid in results:
+ results[sid]['H'] = utt
+ else:
+ results[sid] = {'H': utt}
+ elif l[0] == 'S':
+ sid, utt = l[2:].split('\t')
+ sid = int(sid)
+ utt = [x for x in utt.split()]
+ if sid in results:
+ results[sid]['S'] = utt
+ else:
+ results[sid] = {'S': utt}
+ elif l[0] == 'T':
+ sid, utt = l[2:].split('\t')
+ sid = int(sid)
+ utt = [int(x) for x in utt.split()]
+ if sid in results:
+ results[sid]['T'] = utt
+ else:
+ results[sid] = {'T': utt}
+
+ for d, result in results.items():
+ if 'H' not in result:
+ result['H'] = result['S']
+
+ return results
+
+
+def get_code_to_fname(manifest, tokens):
+ if tokens is None:
+ code_to_fname = {}
+ with open(manifest) as f:
+ for line in f:
+ line = line.strip()
+ fname, code = line.split()
+ code = code.replace(',', ' ')
+ code_to_fname[code] = fname
+
+ return code_to_fname
+
+ with open(manifest) as f:
+ fnames = [l.strip() for l in f.readlines()]
+ root = Path(fnames[0])
+ fnames = fnames[1:]
+ if '\t' in fnames[0]:
+ fnames = [x.split()[0] for x in fnames]
+
+ with open(tokens) as f:
+ codes = [l.strip() for l in f.readlines()]
+
+ code_to_fname = {}
+ for fname, code in zip(fnames, codes):
+ code = code.replace(',', ' ')
+ code_to_fname[code] = str(root / fname)
+
+ return root, code_to_fname
+
+
+def code_to_str(s):
+ k = ' '.join([str(x) for x in s])
+ return k
+
+
+def get_praat_f0(audio, rate=16000, interp=False):
+ frame_length = 20.0
+ to_pad = int(frame_length / 1000 * rate) // 2
+
+ f0s = []
+ for y in audio.astype(np.float64):
+ y_pad = np.pad(y.squeeze(), (to_pad, to_pad), "constant", constant_values=0)
+ signal = basic.SignalObj(y_pad, rate)
+ pitch = pYAAPT.yaapt(signal, **{'frame_length': frame_length, 'frame_space': 5.0, 'nccf_thresh1': 0.25,
+ 'tda_frame_length': 25.0})
+ if interp:
+ f0s += [pitch.samp_interp[None, None, :]]
+ else:
+ f0s += [pitch.samp_values[None, None, :]]
+
+ f0 = np.vstack(f0s)
+ return f0
+
+
+def generate_from_code(generator, h, code, spkr=None, f0=None, gst=None, device="cpu"):
+ batch = {
+ 'code': torch.LongTensor(code).to(device).view(1, -1),
+ }
+ if spkr is not None:
+ batch['spkr'] = spkr.to(device).unsqueeze(0)
+ if f0 is not None:
+ batch['f0'] = f0.to(device)
+ if gst is not None:
+ batch['style'] = gst.to(device)
+
+ with torch.no_grad():
+ audio, rtf = generate(h, generator, batch)
+ audio = librosa.util.normalize(audio / 2 ** 15)
+
+ return audio
+
+
+@torch.no_grad()
+def synth(argv, interactive=False):
+ parser = argparse.ArgumentParser()
+ parser.add_argument('--result-path', type=Path, help='Translation Model Output', required=True)
+ parser.add_argument('--data', type=Path, help='a directory with the files: src.tsv, src.km, trg.tsv, trg.km, orig.tsv, orig.km')
+ parser.add_argument("--orig-tsv", default="/checkpoint/felixkreuk/datasets/emov/manifests/emov_16khz/data.tsv")
+ parser.add_argument("--orig-km", default="/checkpoint/felixkreuk/datasets/emov/manifests/emov_16khz/core_manifests/emov_16khz_km_100/data.km")
+
+ parser.add_argument('--checkpoint-file', type=Path, help='Generator Checkpoint', required=True)
+ parser.add_argument('--dur-model', type=Path, help='a token duration prediction model (if tokens were deduped)')
+ parser.add_argument('--f0-model', type=Path, help='a f0 prediction model')
+
+ parser.add_argument('-s', '--src-emotion', default=None)
+ parser.add_argument('-t', '--trg-emotion', default=None)
+ parser.add_argument('-N', type=int, default=10)
+ parser.add_argument('--split', default="test")
+
+ parser.add_argument('--outdir', type=Path, default=Path('results'))
+ parser.add_argument('--orig-filename', action='store_true')
+
+ parser.add_argument('--device', type=int, default=0)
+ a = parser.parse_args(argv)
+
+ seed = 52
+ random.seed(seed)
+ np.random.seed(seed)
+ torch.manual_seed(seed)
+
+ if os.path.isdir(a.checkpoint_file):
+ config_file = os.path.join(a.checkpoint_file, 'config.json')
+ else:
+ config_file = os.path.join(os.path.split(a.checkpoint_file)[0], 'config.json')
+ with open(config_file) as f:
+ data = f.read()
+ json_config = json.loads(data)
+ h = AttrDict(json_config)
+
+ generator = CodeGenerator(h).to(a.device)
+ if os.path.isdir(a.checkpoint_file):
+ cp_g = scan_checkpoint(a.checkpoint_file, 'g_')
+ else:
+ cp_g = a.checkpoint_file
+ state_dict_g = load_checkpoint(cp_g)
+ generator.load_state_dict(state_dict_g['generator'])
+
+ generator.eval()
+ generator.remove_weight_norm()
+
+ dur_models = {
+ "neutral": load_duration_predictor(f"{a.dur_model}/neutral.ckpt"),
+ "amused": load_duration_predictor(f"{a.dur_model}/amused.ckpt"),
+ "disgusted": load_duration_predictor(f"{a.dur_model}/disgusted.ckpt"),
+ "angry": load_duration_predictor(f"{a.dur_model}/angry.ckpt"),
+ "sleepy": load_duration_predictor(f"{a.dur_model}/sleepy.ckpt"),
+ }
+ logger.info(f"loaded duration prediction model from {a.dur_model}")
+
+ f0_model = load_pitch_predictor(a.f0_model).to(a.device)
+ logger.info(f"loaded f0 prediction model from {a.f0_model}")
+
+ # we need to know how to map code back to the filename
+ # (if we want the original files names as output)
+ results = parse_generation_file(a.result_path)
+ _, src_code_to_fname = get_code_to_fname(f'{a.data}/files.{a.split}.{a.src_emotion}', f'{a.data}/{a.split}.{a.src_emotion}')
+ _, tgt_code_to_fname = get_code_to_fname(f'{a.data}/files.{a.split}.{a.trg_emotion}', f'{a.data}/{a.split}.{a.trg_emotion}')
+
+ # we need the originals (before dedup) to get the ground-truth durations
+ orig_tsv = open(a.orig_tsv, 'r').readlines()
+ orig_tsv_root, orig_tsv = orig_tsv[0].strip(), orig_tsv[1:]
+ orig_km = open(a.orig_km, 'r').readlines()
+ fname_to_idx = {orig_tsv_root + "/" + line.split("\t")[0]: i for i, line in enumerate(orig_tsv)}
+
+ outdir = a.outdir
+ outdir.mkdir(parents=True, exist_ok=True)
+ (outdir / '0-source').mkdir(exist_ok=True)
+ (outdir / '1-src-tokens-src-style-src-f0').mkdir(exist_ok=True)
+ (outdir / '2-src-tokens-trg-style-src-f0').mkdir(exist_ok=True)
+ (outdir / '2.5-src-tokens-trg-style-src-f0').mkdir(exist_ok=True)
+ (outdir / '3-src-tokens-trg-style-pred-f0').mkdir(exist_ok=True)
+ (outdir / '4-gen-tokens-trg-style-pred-f0').mkdir(exist_ok=True)
+ (outdir / '5-target').mkdir(exist_ok=True)
+
+ N = 0
+ results = list(results.items())
+ random.shuffle(results)
+ for i, (sid, result) in tqdm(enumerate(results)):
+ N += 1
+ if N > a.N and a.N != -1:
+ break
+
+ if '[' in result['S'][0]:
+ result['S'] = result['S'][1:]
+ if '_' in result['S'][-1]:
+ result['S'] = result['S'][:-1]
+ src_ref = src_code_to_fname[code_to_str(result['S'])]
+ trg_ref = tgt_code_to_fname[code_to_str(result['T'])]
+
+ src_style, trg_style = None, None
+ src_spkr, trg_spkr = None, None
+ src_f0 = None
+ src_audio = (load_audio(src_ref)[0] / MAX_WAV_VALUE) * 0.95
+ trg_audio = (load_audio(trg_ref)[0] / MAX_WAV_VALUE) * 0.95
+ src_audio = torch.FloatTensor(src_audio).unsqueeze(0).cuda()
+ trg_audio = torch.FloatTensor(trg_audio).unsqueeze(0).cuda()
+
+ src_spkr = parse_speaker(src_ref, h.multispkr)
+ src_spkr = src_spkr if src_spkr in EMOV_SPK2ID else random.choice(list(EMOV_SPK2ID.keys()))
+ src_spkr = EMOV_SPK2ID[src_spkr]
+ src_spkr = torch.LongTensor([src_spkr])
+ trg_spkr = parse_speaker(trg_ref, h.multispkr)
+ trg_spkr = trg_spkr if trg_spkr in EMOV_SPK2ID else random.choice(list(EMOV_SPK2ID.keys()))
+ trg_spkr = EMOV_SPK2ID[trg_spkr]
+ trg_spkr = torch.LongTensor([trg_spkr])
+
+ src_style = EMOV_STYLE2ID[a.src_emotion]
+ src_style = torch.LongTensor([src_style]).cuda()
+ trg_style_str = a.trg_emotion
+ trg_style = EMOV_STYLE2ID[a.trg_emotion]
+ trg_style = torch.LongTensor([trg_style]).cuda()
+
+ src_tokens = list(map(int, orig_km[fname_to_idx[src_ref]].strip().split(" ")))
+ src_tokens = torch.LongTensor(src_tokens).unsqueeze(0)
+ src_tokens_dur_pred = torch.LongTensor(list(map(int, result['S']))).unsqueeze(0)
+ src_tokens_dur_pred = dur_models[trg_style_str].inflate_input(src_tokens_dur_pred)
+ gen_tokens = torch.LongTensor(result['H']).unsqueeze(0)
+ gen_tokens = dur_models[trg_style_str].inflate_input(gen_tokens)
+ trg_tokens = torch.LongTensor(result['T']).unsqueeze(0)
+ trg_tokens = dur_models[trg_style_str].inflate_input(trg_tokens)
+
+ src_f0 = get_praat_f0(src_audio.unsqueeze(0).cpu().numpy())
+ src_f0 = torch.FloatTensor(src_f0).cuda()
+
+ pred_src_f0 = f0_model.inference(torch.LongTensor(src_tokens).to(a.device), src_spkr, trg_style).unsqueeze(0)
+ pred_src_dur_pred_f0 = f0_model.inference(torch.LongTensor(src_tokens_dur_pred).to(a.device), src_spkr, trg_style).unsqueeze(0)
+ pred_gen_f0 = f0_model.inference(torch.LongTensor(gen_tokens).to(a.device), src_spkr, trg_style).unsqueeze(0)
+ pred_trg_f0 = f0_model.inference(torch.LongTensor(trg_tokens).to(a.device), src_spkr, trg_style).unsqueeze(0)
+
+ if a.orig_filename:
+ path = src_code_to_fname[code_to_str(result['S'])]
+ sid = str(sid) + "__" + Path(path).stem
+ shutil.copy(src_code_to_fname[code_to_str(result['S'])], outdir / '0-source' / f'{sid}.wav')
+
+ audio = generate_from_code(generator, h, src_tokens, spkr=src_spkr, f0=src_f0, gst=src_style, device=a.device)
+ sf.write(outdir / '1-src-tokens-src-style-src-f0' / f'{sid}.wav', audio, samplerate=h.sampling_rate)
+
+ audio = generate_from_code(generator, h, src_tokens, spkr=src_spkr, f0=src_f0, gst=trg_style, device=a.device)
+ sf.write(outdir / '2-src-tokens-trg-style-src-f0' / f'{sid}.wav', audio, samplerate=h.sampling_rate)
+
+ audio = generate_from_code(generator, h, src_tokens_dur_pred, spkr=src_spkr, f0=src_f0, gst=trg_style, device=a.device)
+ sf.write(outdir / '2.5-src-tokens-trg-style-src-f0' / f'{sid}.wav', audio, samplerate=h.sampling_rate)
+
+ audio = generate_from_code(generator, h, src_tokens_dur_pred, spkr=src_spkr, f0=pred_src_dur_pred_f0, gst=trg_style, device=a.device)
+ sf.write(outdir / '3-src-tokens-trg-style-pred-f0' / f'{sid}.wav', audio, samplerate=h.sampling_rate)
+
+ audio = generate_from_code(generator, h, gen_tokens, spkr=src_spkr, f0=pred_gen_f0, gst=trg_style, device=a.device)
+ sf.write(outdir / '4-gen-tokens-trg-style-pred-f0' / f'{sid}.wav', audio, samplerate=h.sampling_rate)
+
+ shutil.copy(tgt_code_to_fname[code_to_str(result['T'])], outdir / '5-target' / f'{sid}.wav')
+
+ logger.info("Done.")
+
+
+if __name__ == '__main__':
+ synth(sys.argv[1:])
diff --git a/fairseq/examples/fast_noisy_channel/README.md b/fairseq/examples/fast_noisy_channel/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..f2631a8c34d11bdf7d351c6807b6fe415f5715e1
--- /dev/null
+++ b/fairseq/examples/fast_noisy_channel/README.md
@@ -0,0 +1,345 @@
+# Language Models not just for Pre-training: Fast Online Neural Noisy Channel Modeling
+
+## Introduction
+- [Yee et al. (2019)](https://www.aclweb.org/anthology/D19-1571.pdf) introduce a simple and effective noisy channel modeling approach for neural machine translation. However, the noisy channel online decoding approach introduced in this paper is too slow to be practical.
+- To address this, [Bhosale et al. (2020)](http://www.statmt.org/wmt20/pdf/2020.wmt-1.68.pdf) introduces 3 simple approximations to make this approach very fast and practical without much loss in accuracy.
+- This README provides intructions on how to run online decoding or generation with the noisy channel modeling approach, including ways to make it very fast without much loss in accuracy.
+
+## Noisy Channel Modeling
+
+[Yee et al. (2019)](https://www.aclweb.org/anthology/D19-1571.pdf) applies the Bayes Rule to predict `P(y|x)`, the probability of the target `y` given the source `x`.
+```P(y|x) = P(x|y) * P(y) / P(x)```
+- `P(x|y)` predicts the source `x` given the target `y` and is referred to as the **channel model**
+- `P(y)` is a **language model** over the target `y`
+- `P(x)` is generally not modeled since it is constant for all `y`.
+
+We use Transformer models to parameterize the direct model `P(y|x)`, the channel model `P(x|y)` and the language model `P(y)`.
+
+During online decoding with beam search, we generate the top `K2` candidates per beam and score them with the following linear combination of the channel model, the language model as well as the direct model scores.
+
+```(1 / t) * log(P(y|x) + (1 / s) * ( λ1 * log(P(x|y)) + λ2 * log(P(y) ) )```
+- `t` - Target Prefix Length
+- `s` - Source Length
+- `λ1` - Channel Model Weight
+- `λ2` - Language Model Weight
+
+The top `beam_size` candidates based on the above combined scores are chosen to continue the beams in beam search. In beam search with a direct model alone, the scores from the direct model `P(y|x)` are used to choose the top candidates in beam search.
+
+This framework provides a great way to utlize strong target language models trained on large amounts of unlabeled data. Language models can prefer targets unrelated to the source, so we also need a channel model whose role is to ensure that the target preferred by the language model also translates back to the source.
+
+### Training Translation Models and Language Models
+
+For training Transformer models in fairseq for machine translation, refer to instructions [here](https://github.com/pytorch/fairseq/tree/main/examples/translation)
+
+For training Transformer models in fairseq for language modeling, refer to instructions [here](https://github.com/pytorch/fairseq/tree/main/examples/language_model)
+
+### Generation with Language Model for German-English translation with fairseq
+
+Here are instructions to generate using a direct model and a target-side language model.
+
+Note:
+- Download and install fairseq as per instructions [here](https://github.com/pytorch/fairseq)
+- Preprocess and binarize the dataset as per instructions in section [Test Data Preprocessing](#test-data-preprocessing)
+
+```sh
+binarized_data=data_dir/binarized
+direct_model=de_en_seed4.pt
+lm_model=en_lm.pt
+lm_data=lm_data
+wget https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/direct_models/seed4.pt -O ${direct_model}
+wget https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/lm_model/transformer_lm.pt -O ${lm_model}
+mkdir -p ${lm_data}
+wget https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/lm_model/lm_dict/dict.txt -O ${lm_data}/dict.txt
+
+k2=10
+lenpen=0.16
+lm_wt=0.14
+fairseq-generate ${binarized_data} \
+ --user-dir examples/fast_noisy_channel \
+ --beam 5 \
+ --path ${direct_model} \
+ --lm-model ${lm_model} \
+ --lm-data ${lm_data} \
+ --k2 ${k2} \
+ --combine-method lm_only \
+ --task noisy_channel_translation \
+ --lenpen ${lenpen} \
+ --lm-wt ${lm_wt} \
+ --gen-subset valid \
+ --remove-bpe \
+ --fp16 \
+ --batch-size 10
+```
+### Noisy Channel Generation for German-English translation with fairseq
+
+Here are instructions for noisy channel generation with a direct model, channel model and language model as explained in section [Noisy Channel Modeling](#noisy-channel-modeling).
+
+Note:
+- Download and install fairseq as per instructions [here](https://github.com/pytorch/fairseq)
+- Preprocess and binarize the dataset as per instructions in section [Test Data Preprocessing](#test-data-preprocessing)
+
+```sh
+binarized_data=data_dir/binarized
+direct_model=de_en_seed4.pt
+lm_model=en_lm.pt
+lm_data=lm_data
+ch_model=en_de.big.seed4.pt
+wget https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/direct_models/seed4.pt -O ${direct_model}
+wget https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/lm_model/transformer_lm.pt -O ${lm_model}
+mkdir -p ${lm_data}
+wget https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/lm_model/lm_dict/dict.txt -O ${lm_data}/dict.txt
+wget https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/big.seed4.pt -O ${ch_model}
+
+k2=10
+lenpen=0.21
+lm_wt=0.50
+bw_wt=0.30
+fairseq-generate ${binarized_data} \
+ --user-dir examples/fast_noisy_channel \
+ --beam 5 \
+ --path ${direct_model} \
+ --lm-model ${lm_model} \
+ --lm-data ${lm_data} \
+ --channel-model ${ch_model} \
+ --k2 ${k2} \
+ --combine-method noisy_channel \
+ --task noisy_channel_translation \
+ --lenpen ${lenpen} \
+ --lm-wt ${lm_wt} \
+ --ch-wt ${bw_wt} \
+ --gen-subset test \
+ --remove-bpe \
+ --fp16 \
+ --batch-size 1
+```
+## Fast Noisy Channel Modeling
+
+[Bhosale et al. (2020)](http://www.statmt.org/wmt20/pdf/2020.wmt-1.68.pdf) introduces 3 approximations that speed up online noisy channel decoding -
+- Smaller channel models (`Tranformer Base` with 1 encoder and decoder layer each vs. `Transformer Big`)
+ - This involves training a channel model that is possibly smaller and less accurate in terms of BLEU than a channel model of the same size as the direct model.
+ - Since the role of the channel model is mainly to assign low scores to generations from the language model if they don't translate back to the source, we may not need the most accurate channel model for this purpose.
+- Smaller output vocabulary size for the channel model (~30,000 -> ~1000)
+ - The channel model doesn't need to score the full output vocabulary, it just needs to score the source tokens, which are completely known.
+ - This is specified using the arguments `--channel-scoring-type src_vocab --top-k-vocab 500`
+ - This means that the output vocabulary for the channel model will be the source tokens for all examples in the batch and the top-K most frequent tokens in the vocabulary
+ - This reduces the memory consumption needed to store channel model scores significantly
+- Smaller number of candidates (`k2`) scored per beam
+ - This is specified by reducing the argument `--k2`
+
+
+### Fast Noisy Channel Generation for German-English translation with fairseq
+
+Here are instructions for **fast** noisy channel generation with a direct model, channel model and language model as explained in section [Fast Noisy Channel Modeling](#fast-noisy-channel-modeling). The main differences are that we use a smaller channel model, reduce `--k2`, set `--channel-scoring-type src_vocab --top-k-vocab 500` and increase the `--batch-size`.
+
+Note:
+- Download and install fairseq as per instructions [here](https://github.com/pytorch/fairseq)
+- Preprocess and binarize the dataset as per instructions in section [Test Data Preprocessing](#test-data-preprocessing)
+
+```sh
+binarized_data=data_dir/binarized
+direct_model=de_en_seed4.pt
+lm_model=en_lm.pt
+lm_data=lm_data
+small_ch_model=en_de.base_1_1.seed4.pt
+wget https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/direct_models/seed4.pt -O ${direct_model}
+wget https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/lm_model/transformer_lm.pt -O ${lm_model}
+mkdir -p ${lm_data}
+wget https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/lm_model/lm_dict/dict.txt -O ${lm_data}/dict.txt
+wget https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/base_1_1.seed4.pt -O ${small_ch_model}
+
+k2=3
+lenpen=0.23
+lm_wt=0.58
+bw_wt=0.26
+fairseq-generate ${binarized_data} \
+ --user-dir examples/fast_noisy_channel \
+ --beam 5 \
+ --path ${direct_model} \
+ --lm-model ${lm_model} \
+ --lm-data ${lm_data} \
+ --channel-model ${small_ch_model} \
+ --k2 ${k2} \
+ --combine-method noisy_channel \
+ --task noisy_channel_translation \
+ --lenpen ${lenpen} \
+ --lm-wt ${lm_wt} \
+ --ch-wt ${bw_wt} \
+ --gen-subset test \
+ --remove-bpe \
+ --fp16 \
+ --batch-size 50 \
+ --channel-scoring-type src_vocab --top-k-vocab 500
+```
+
+## Test Data Preprocessing
+
+For preprocessing and binarizing the test sets for Romanian-English and German-English translation, we use the following script -
+
+```sh
+FAIRSEQ=/path/to/fairseq
+cd $FAIRSEQ
+SCRIPTS=$FAIRSEQ/mosesdecoder/scripts
+if [ ! -d "${SCRIPTS}" ]; then
+ echo 'Cloning Moses github repository (for tokenization scripts)...'
+ git clone https://github.com/moses-smt/mosesdecoder.git
+fi
+TOKENIZER=$SCRIPTS/tokenizer/tokenizer.perl
+NORMALIZE=$SCRIPTS/tokenizer/normalize-punctuation.perl
+
+s=de
+t=en
+test=wmt18
+
+mkdir -p data_dir
+
+# Tokenization
+if [ $s == "ro" ] ; then
+ # Note: Get normalise-romanian.py and remove-diacritics.py from
+ # https://github.com/rsennrich/wmt16-scripts/tree/master/preprocess
+ sacrebleu -t $test -l $s-$t --echo src | \
+ $NORMALIZE -l $s | \
+ python normalise-romanian.py | \
+ python remove-diacritics.py | \
+ $TOKENIZER -l $s -a -q > data_dir/$test.$s-$t.$s
+else
+ sacrebleu -t $test -l $s-$t --echo src | perl $NORMALIZE -l $s | perl $TOKENIZER -threads 8 -a -l $s > data_dir/$test.$s-$t.$s
+fi
+
+sacrebleu -t $test -l $s-$t --echo ref | perl $NORMALIZE -l $t | perl $TOKENIZER -threads 8 -a -l $t > data_dir/$test.$s-$t.$t
+
+
+# Applying BPE
+src_bpe_code=/path/to/source/language/bpe/code
+tgt_bpe_code=/path/to/target/language/bpe/code
+src_dict=/path/to/source/language/dict
+tgt_dict=/path/to/target/language/dict
+
+FASTBPE=$FAIRSEQ/fastBPE
+if [ ! -d "${FASTBPE}" ] ; then
+ git clone https://github.com/glample/fastBPE.git
+ # Follow compilation instructions at https://github.com/glample/fastBPE
+ g++ -std=c++11 -pthread -O3 fastBPE/main.cc -IfastBPE -o fast
+fi
+
+${FASTBPE}/fast applybpe data_dir/bpe.$test.$s-$t.$s data_dir/$test.$s-$t.$s ${src_bpe_code}
+${FASTBPE}/fast applybpe data_dir/bpe.$test.$s-$t.$s data_dir/$test.$s-$t.$s ${tgt_bpe_code}
+
+fairseq-preprocess -s $s -t $t \
+ --testpref data_dir/bpe.$test.$s-$t \
+ --destdir data_dir/binarized \
+ --srcdict ${src_dict} \
+ --tgtdict ${tgt_dict}
+```
+
+## Calculating BLEU
+
+```sh
+DETOKENIZER=$SCRIPTS/tokenizer/detokenizer.perl
+cat ${generation_output} | grep -P "^H" | sort -V | cut -f 3- | $DETOKENIZER -l $t -q -a | sacrebleu -t $test -l $s-$t
+```
+
+
+## Romanian-English Translation
+
+The direct and channel models are trained using bitext data (WMT16) combined with backtranslated data (The monolingual data used for backtranslation comes from http://data.statmt.org/rsennrich/wmt16_backtranslations/ (Sennrich et al., 2016c))
+
+The backtranslated data is generated using an ensemble of 3 English-Romanian models trained on bitext training data (WMT16) with unrestricted sampling.
+
+### BPE Codes and Dictionary
+
+We learn a joint BPE vocabulary of 18K types on the bitext training data which is used for both the source and target.
+||Path|
+|----------|------|
+| BPE Code | [joint_bpe_18k](https://dl.fbaipublicfiles.com/fast_noisy_channel/ro_en/bpe_18k) |
+| Dictionary | [dict](https://dl.fbaipublicfiles.com/fast_noisy_channel/ro_en/dict) |
+
+### Direct Models
+For Ro-En with backtranslation, the direct and channel models use a Transformer-Big architecture.
+
+| Seed | Model |
+|----|----|
+| 2 | [ro_en_seed2.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/ro_en/direct_models/seed2.pt)
+| 4 | [ro_en_seed4.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/ro_en/direct_models/seed4.pt)
+| 6 | [ro_en_seed6.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/ro_en/direct_models/seed6.pt)
+
+### Channel Models
+For channel models, we follow the same steps as for the direct models. But backtranslated data is generated in the opposite direction using [this Romanian monolingual data](http://data.statmt.org/rsennrich/wmt16_backtranslations/).
+The best lenpen, LM weight and CH weight are obtained by sweeping over the validation set (wmt16/dev) using beam 5.
+| Model Size | Lenpen | LM Weight | CH Weight | Seed 2 | Seed 4 | Seed 6 |
+|----|----|----|----|----|----|----|
+| `big` | 0.84 | 0.64 | 0.56 | [big.seed2.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/ro_en/channel_models/big.seed2.pt) | [big.seed2.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/ro_en/channel_models/big.seed2.pt) | [big.seed2.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/ro_en/channel_models/big.seed2.pt) |
+| `base_1_1` | 0.63 | 0.40 | 0.37 | [base_1_1.seed2.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/ro_en/channel_models/base_1_1.seed2.pt) | [base_1_1.seed4.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/ro_en/channel_models/base_1_1.seed4.pt) | [base_1_1.seed6.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/ro_en/channel_models/base_1_1.seed6.pt) |
+
+### Language Model
+The model is trained on de-duplicated English Newscrawl data from 2007-2018 comprising 186 million sentences or 4.5B words after normalization and tokenization.
+| | Path |
+|----|----|
+| `--lm-model` | [transformer_en_lm](https://dl.fbaipublicfiles.com/fast_noisy_channel/ro_en/lm_model/transformer_lm.pt) |
+| `--lm-data` | [lm_data](https://dl.fbaipublicfiles.com/fast_noisy_channel/ro_en/lm_model/lm_dict)
+
+## German-English Translation
+
+### BPE Codes and Dictionaries
+
+| | Path|
+|----------|------|
+| Source BPE Code | [de_bpe_code_24K](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/de_bpe_code_24K) |
+| Target BPE Code | [en_bpe_code_24K](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/en_bpe_code_24K)
+| Source Dictionary | [de_dict](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/de_dict) |
+| Target Dictionary | [en_dict](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/en_dict) |
+
+### Direct Models
+We train on WMT’19 training data. Following [Ng et al., 2019](http://statmt.org/wmt19/pdf/53/WMT33.pdf), we apply language identification filtering and remove sentences longer than 250 tokens as well as sentence pairs with a source/target length ratio exceeding 1.5. This results in 26.8M sentence pairs.
+We use the Transformer-Big architecture for the direct model.
+
+| Seed | Model |
+|:----:|----|
+| 4 | [de_en_seed4.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/direct_models/seed4.pt)
+| 5 | [de_en_seed5.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/direct_models/seed5.pt)
+| 6 | [de_en_seed6.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/direct_models/seed6.pt)
+
+### Channel Models
+
+We train on WMT’19 training data. Following [Ng et al., 2019](http://statmt.org/wmt19/pdf/53/WMT33.pdf), we apply language identification filtering and remove sentences longer than 250 tokens as well as sentence pairs with a source/target length ratio exceeding 1.5. This results in 26.8M sentence pairs.
+
+| Model Size | Seed 4 | Seed 5 | Seed 6 |
+|----|----|----|----|
+| `big` | [big.seed4.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/big.seed4.pt) | [big.seed5.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/big.seed5.pt) | [big.seed6.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/big.seed6.pt) |
+| `big_1_1` | [big_1_1.seed4.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/big_1_1.seed4.pt) | [big_1_1.seed5.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/big_1_1.seed5.pt) | [big_1_1.seed6.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/big_1_1.seed6.pt) |
+| `base` | [base.seed4.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/base.seed4.pt) | [base.seed5.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/base.seed5.pt) | [base.seed6.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/base.seed6.pt) |
+| `base_1_1` | [base_1_1.seed4.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/base_1_1.seed4.pt) | [base_1_1.seed5.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/base_1_1.seed5.pt) | [base_1_1.seed6.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/base_1_1.seed6.pt) |
+| `half` | [half.seed4.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/half.seed4.pt) | [half.seed5.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/half.seed5.pt) | [half.seed6.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/half.seed6.pt) |
+| `half_1_1` | [half_1_1.seed4.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/half_1_1.seed4.pt) | [half_1_1.seed5.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/half_1_1.seed5.pt) | [half_1_1.seed6.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/half_1_1.seed6.pt) |
+| `quarter` | [quarter.seed4.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/quarter.seed4.pt) | [quarter.seed5.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/quarter.seed5.pt) | [quarter.seed6.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/quarter.seed6.pt) |
+| `quarter_1_1` | [quarter_1_1.seed4.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/quarter_1_1.seed4.pt) | [quarter_1_1.seed5.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/quarter_1_1.seed5.pt) | [quarter_1_1.seed6.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/quarter_1_1.seed6.pt) |
+| `8th` | [8th.seed4.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/8th.seed4.pt) | [8th.seed5.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/8th.seed5.pt) | [8th.seed6.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/8th.seed6.pt) |
+| `8th_1_1` | [8th_1_1.seed4.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/8th_1_1.seed4.pt) | [8th_1_1.seed5.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/8th_1_1.seed5.pt) | [8th_1_1.seed6.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/8th_1_1.seed6.pt) |
+| `16th` | [16th.seed4.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/16th.seed4.pt) | [16th.seed5.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/16th.seed5.pt) | [16th.seed6.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/16th.seed6.pt) |
+| `16th_1_1` | [16th_1_1.seed4.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/16th_1_1.seed4.pt) | [16th_1_1.seed5.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/16th_1_1.seed5.pt) | [16th_1_1.seed6.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/16th_1_1.seed6.pt) |
+
+### Language Model
+The model is trained on de-duplicated English Newscrawl data from 2007-2018 comprising 186 million sentences or 4.5B words after normalization and tokenization.
+| | Path |
+|----|----|
+| `--lm-model` | [transformer_en_lm](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/lm_model/transformer_lm.pt) |
+| `--lm-data` | [lm_data](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/lm_model/lm_dict/)
+
+
+## Citation
+
+```bibtex
+@inproceedings{bhosale2020language,
+ title={Language Models not just for Pre-training: Fast Online Neural Noisy Channel Modeling},
+ author={Shruti Bhosale and Kyra Yee and Sergey Edunov and Michael Auli},
+ booktitle={Proceedings of the Fifth Conference on Machine Translation (WMT)},
+ year={2020},
+}
+
+@inproceedings{yee2019simple,
+ title={Simple and Effective Noisy Channel Modeling for Neural Machine Translation},
+ author={Yee, Kyra and Dauphin, Yann and Auli, Michael},
+ booktitle={Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)},
+ pages={5700--5705},
+ year={2019}
+}
+```
diff --git a/fairseq/examples/fast_noisy_channel/__init__.py b/fairseq/examples/fast_noisy_channel/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..9b248c3a24e12ad3da885a7f328c714942de2e6b
--- /dev/null
+++ b/fairseq/examples/fast_noisy_channel/__init__.py
@@ -0,0 +1,8 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+from . import noisy_channel_translation # noqa
+from . import noisy_channel_sequence_generator # noqa
+from . import noisy_channel_beam_search # noqa
diff --git a/fairseq/examples/fast_noisy_channel/noisy_channel_beam_search.py b/fairseq/examples/fast_noisy_channel/noisy_channel_beam_search.py
new file mode 100644
index 0000000000000000000000000000000000000000..23869ebcd0c438f36e310c8ccddd3b5c07a71182
--- /dev/null
+++ b/fairseq/examples/fast_noisy_channel/noisy_channel_beam_search.py
@@ -0,0 +1,71 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+from fairseq.search import Search
+
+
+class NoisyChannelBeamSearch(Search):
+
+ def __init__(self, tgt_dict):
+ super().__init__(tgt_dict)
+ self.fw_scores_buf = None
+ self.lm_scores_buf = None
+
+ def _init_buffers(self, t):
+ # super()._init_buffers(t)
+ if self.fw_scores_buf is None:
+ self.scores_buf = t.new()
+ self.indices_buf = torch.LongTensor().to(device=t.device)
+ self.beams_buf = torch.LongTensor().to(device=t.device)
+ self.fw_scores_buf = t.new()
+ self.lm_scores_buf = t.new()
+
+ def combine_fw_bw(self, combine_method, fw_cum, bw, step):
+ if combine_method == "noisy_channel":
+ fw_norm = fw_cum.div(step + 1)
+ lprobs = bw + fw_norm
+ elif combine_method == "lm_only":
+ lprobs = bw + fw_cum
+
+ return lprobs
+
+ def step(self, step, fw_lprobs, scores, bw_lprobs, lm_lprobs, combine_method):
+ self._init_buffers(fw_lprobs)
+ bsz, beam_size, vocab_size = fw_lprobs.size()
+
+ if step == 0:
+ # at the first step all hypotheses are equally likely, so use
+ # only the first beam
+ fw_lprobs = fw_lprobs[:, ::beam_size, :].contiguous()
+ bw_lprobs = bw_lprobs[:, ::beam_size, :].contiguous()
+ # nothing to add since we are at the first step
+ fw_lprobs_cum = fw_lprobs
+
+ else:
+ # make probs contain cumulative scores for each hypothesis
+ raw_scores = (scores[:, :, step - 1].unsqueeze(-1))
+ fw_lprobs_cum = (fw_lprobs.add(raw_scores))
+
+ combined_lprobs = self.combine_fw_bw(combine_method, fw_lprobs_cum, bw_lprobs, step)
+
+ # choose the top k according to the combined noisy channel model score
+ torch.topk(
+ combined_lprobs.view(bsz, -1),
+ k=min(
+ # Take the best 2 x beam_size predictions. We'll choose the first
+ # beam_size of these which don't predict eos to continue with.
+ beam_size * 2,
+ combined_lprobs.view(bsz, -1).size(1) - 1, # -1 so we never select pad
+ ),
+ out=(self.scores_buf, self.indices_buf),
+ )
+ # save corresponding fw and lm scores
+ self.fw_scores_buf = torch.gather(fw_lprobs_cum.view(bsz, -1), 1, self.indices_buf)
+ self.lm_scores_buf = torch.gather(lm_lprobs.view(bsz, -1), 1, self.indices_buf)
+ # Project back into relative indices and beams
+ self.beams_buf = self.indices_buf // vocab_size
+ self.indices_buf.fmod_(vocab_size)
+ return self.scores_buf, self.fw_scores_buf, self.lm_scores_buf, self.indices_buf, self.beams_buf
diff --git a/fairseq/examples/fast_noisy_channel/noisy_channel_sequence_generator.py b/fairseq/examples/fast_noisy_channel/noisy_channel_sequence_generator.py
new file mode 100644
index 0000000000000000000000000000000000000000..ea8fae98e87e9f3e69bc51987703a6429eb0c92a
--- /dev/null
+++ b/fairseq/examples/fast_noisy_channel/noisy_channel_sequence_generator.py
@@ -0,0 +1,842 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+from typing import Dict, List, Optional
+
+import math
+import numpy as np
+
+import torch
+import torch.nn.functional as F
+from torch import Tensor
+
+from .noisy_channel_beam_search import NoisyChannelBeamSearch
+from fairseq.sequence_generator import EnsembleModel
+
+
+class NoisyChannelSequenceGenerator(object):
+ def __init__(
+ self,
+ combine_method,
+ tgt_dict,
+ src_dict=None,
+ beam_size=1,
+ max_len_a=0,
+ max_len_b=200,
+ min_len=1,
+ len_penalty=1.0,
+ unk_penalty=0.0,
+ retain_dropout=False,
+ temperature=1.0,
+ match_source_len=False,
+ no_repeat_ngram_size=0,
+ normalize_scores=True,
+ channel_models=None,
+ k2=10,
+ ch_weight=1.0,
+ channel_scoring_type='log_norm',
+ top_k_vocab=0,
+ lm_models=None,
+ lm_dict=None,
+ lm_weight=1.0,
+ normalize_lm_scores_by_tgt_len=False,
+ ):
+ """Generates translations of a given source sentence,
+ using beam search with noisy channel decoding.
+
+ Args:
+ combine_method (string, optional): Method to combine direct, LM and
+ channel model scores (default: None)
+ tgt_dict (~fairseq.data.Dictionary): target dictionary
+ src_dict (~fairseq.data.Dictionary): source dictionary
+ beam_size (int, optional): beam width (default: 1)
+ max_len_a/b (int, optional): generate sequences of maximum length
+ ax + b, where x is the source length
+ min_len (int, optional): the minimum length of the generated output
+ (not including end-of-sentence)
+ len_penalty (float, optional): length penalty, where <1.0 favors
+ shorter, >1.0 favors longer sentences (default: 1.0)
+ unk_penalty (float, optional): unknown word penalty, where <0
+ produces more unks, >0 produces fewer (default: 0.0)
+ retain_dropout (bool, optional): use dropout when generating
+ (default: False)
+ temperature (float, optional): temperature, where values
+ >1.0 produce more uniform samples and values <1.0 produce
+ sharper samples (default: 1.0)
+ match_source_len (bool, optional): outputs should match the source
+ length (default: False)
+ no_repeat_ngram_size (int, optional): Size of n-grams that we avoid
+ repeating in the generation (default: 0)
+ normalize_scores (bool, optional): normalize scores by the length
+ of the output (default: True)
+ channel_models (List[~fairseq.models.FairseqModel]): ensemble of models
+ translating from the target to the source
+ k2 (int, optional): Top K2 candidates to score per beam at each step (default:10)
+ ch_weight (int, optional): Weight associated with the channel model score
+ assuming that the direct model score has weight 1.0 (default: 1.0)
+ channel_scoring_type (str, optional): String specifying how to score
+ the channel model (default: 'log_norm')
+ top_k_vocab (int, optional): If `channel_scoring_type` is `'src_vocab'` or
+ `'src_vocab_batched'`, then this parameter specifies the number of
+ most frequent tokens to include in the channel model output vocabulary,
+ in addition to the source tokens in the input batch (default: 0)
+ lm_models (List[~fairseq.models.FairseqModel]): ensemble of models
+ generating text in the target language
+ lm_dict (~fairseq.data.Dictionary): LM Model dictionary
+ lm_weight (int, optional): Weight associated with the LM model score
+ assuming that the direct model score has weight 1.0 (default: 1.0)
+ normalize_lm_scores_by_tgt_len (bool, optional): Should we normalize LM scores
+ by the target length? By default, we normalize the combination of
+ LM and channel model scores by the source length
+ """
+ self.pad = tgt_dict.pad()
+ self.unk = tgt_dict.unk()
+ self.eos = tgt_dict.eos()
+ self.vocab_size = len(tgt_dict)
+ self.beam_size = beam_size
+ # the max beam size is the dictionary size - 1, since we never select pad
+ self.beam_size = min(beam_size, self.vocab_size - 1)
+ self.max_len_a = max_len_a
+ self.max_len_b = max_len_b
+ self.min_len = min_len
+ self.normalize_scores = normalize_scores
+ self.len_penalty = len_penalty
+ self.unk_penalty = unk_penalty
+ self.retain_dropout = retain_dropout
+ self.temperature = temperature
+ self.match_source_len = match_source_len
+ self.no_repeat_ngram_size = no_repeat_ngram_size
+ self.channel_models = channel_models
+ self.src_dict = src_dict
+ self.tgt_dict = tgt_dict
+ self.combine_method = combine_method
+ self.k2 = k2
+ self.ch_weight = ch_weight
+ self.channel_scoring_type = channel_scoring_type
+ self.top_k_vocab = top_k_vocab
+ self.lm_models = lm_models
+ self.lm_dict = lm_dict
+ self.lm_weight = lm_weight
+ self.log_softmax_fn = torch.nn.LogSoftmax(dim=1)
+ self.normalize_lm_scores_by_tgt_len = normalize_lm_scores_by_tgt_len
+
+ self.share_tgt_dict = (self.lm_dict == self.tgt_dict)
+ self.tgt_to_lm = make_dict2dict(tgt_dict, lm_dict)
+
+ self.ch_scoring_bsz = 3072
+
+ assert temperature > 0, '--temperature must be greater than 0'
+
+ self.search = NoisyChannelBeamSearch(tgt_dict)
+
+ @torch.no_grad()
+ def generate(
+ self,
+ models,
+ sample,
+ prefix_tokens=None,
+ bos_token=None,
+ **kwargs
+ ):
+ """Generate a batch of translations.
+ Args:
+ models (List[~fairseq.models.FairseqModel]): ensemble of models
+ sample (dict): batch
+ prefix_tokens (torch.LongTensor, optional): force decoder to begin
+ with these tokens
+ """
+ model = EnsembleModel(models)
+ incremental_states = torch.jit.annotate(
+ List[Dict[str, Dict[str, Optional[Tensor]]]],
+ [
+ torch.jit.annotate(Dict[str, Dict[str, Optional[Tensor]]], {})
+ for i in range(model.models_size)
+ ],
+ )
+ if not self.retain_dropout:
+ model.eval()
+
+ # model.forward normally channels prev_output_tokens into the decoder
+ # separately, but SequenceGenerator directly calls model.encoder
+ encoder_input = {
+ k: v for k, v in sample['net_input'].items()
+ if k != 'prev_output_tokens'
+ }
+ src_tokens = encoder_input['src_tokens']
+ src_lengths_no_eos = (src_tokens.ne(self.eos) & src_tokens.ne(self.pad)).long().sum(dim=1)
+ input_size = src_tokens.size()
+ # batch dimension goes first followed by source lengths
+ bsz = input_size[0]
+ src_len = input_size[1]
+ beam_size = self.beam_size
+
+ if self.match_source_len:
+ max_len = src_lengths_no_eos.max().item()
+ else:
+ max_len = min(
+ int(self.max_len_a * src_len + self.max_len_b),
+ # exclude the EOS marker
+ model.max_decoder_positions() - 1,
+ )
+
+ # compute the encoder output for each beam
+ encoder_outs = model.forward_encoder(encoder_input)
+ new_order = torch.arange(bsz).view(-1, 1).repeat(1, beam_size).view(-1)
+ new_order = new_order.to(src_tokens.device).long()
+ encoder_outs = model.reorder_encoder_out(encoder_outs, new_order)
+
+ src_lengths = encoder_input['src_lengths']
+ # initialize buffers
+ scores = src_tokens.new(bsz * beam_size, max_len + 1).float().fill_(0)
+ lm_prefix_scores = src_tokens.new(bsz * beam_size).float().fill_(0)
+
+ scores_buf = scores.clone()
+ tokens = src_tokens.new(bsz * beam_size, max_len + 2).long().fill_(self.pad)
+ tokens_buf = tokens.clone()
+ tokens[:, 0] = self.eos if bos_token is None else bos_token
+
+ # reorder source tokens so they may be used as a reference in generating P(S|T)
+ src_tokens = reorder_all_tokens(src_tokens, src_lengths, self.src_dict.eos_index)
+
+ src_tokens = src_tokens.repeat(1, beam_size).view(-1, src_len)
+ src_lengths = src_lengths.view(bsz, -1).repeat(1, beam_size).view(bsz*beam_size, -1)
+
+ attn, attn_buf = None, None
+ nonpad_idxs = None
+
+ # The cands_to_ignore indicates candidates that should be ignored.
+ # For example, suppose we're sampling and have already finalized 2/5
+ # samples. Then the cands_to_ignore would mark 2 positions as being ignored,
+ # so that we only finalize the remaining 3 samples.
+ cands_to_ignore = src_tokens.new_zeros(bsz, beam_size).eq(-1) # forward and backward-compatible False mask
+
+ # list of completed sentences
+ finalized = [[] for i in range(bsz)]
+ finished = [False for i in range(bsz)]
+ num_remaining_sent = bsz
+
+ # number of candidate hypos per step
+ cand_size = 2 * beam_size # 2 x beam size in case half are EOS
+
+ # offset arrays for converting between different indexing schemes
+ bbsz_offsets = (torch.arange(0, bsz) * beam_size).unsqueeze(1).type_as(tokens)
+ cand_offsets = torch.arange(0, cand_size).type_as(tokens)
+
+ # helper function for allocating buffers on the fly
+ buffers = {}
+
+ def buffer(name, type_of=tokens): # noqa
+ if name not in buffers:
+ buffers[name] = type_of.new()
+ return buffers[name]
+
+ def is_finished(sent, step, unfin_idx):
+ """
+ Check whether we've finished generation for a given sentence, by
+ comparing the worst score among finalized hypotheses to the best
+ possible score among unfinalized hypotheses.
+ """
+ assert len(finalized[sent]) <= beam_size
+ if len(finalized[sent]) == beam_size:
+ return True
+ return False
+
+ def finalize_hypos(step, bbsz_idx, eos_scores, combined_noisy_channel_eos_scores):
+ """
+ Finalize the given hypotheses at this step, while keeping the total
+ number of finalized hypotheses per sentence <= beam_size.
+
+ Note: the input must be in the desired finalization order, so that
+ hypotheses that appear earlier in the input are preferred to those
+ that appear later.
+
+ Args:
+ step: current time step
+ bbsz_idx: A vector of indices in the range [0, bsz*beam_size),
+ indicating which hypotheses to finalize
+ eos_scores: A vector of the same size as bbsz_idx containing
+ fw scores for each hypothesis
+ combined_noisy_channel_eos_scores: A vector of the same size as bbsz_idx containing
+ combined noisy channel scores for each hypothesis
+ """
+ assert bbsz_idx.numel() == eos_scores.numel()
+
+ # clone relevant token and attention tensors
+ tokens_clone = tokens.index_select(0, bbsz_idx)
+ tokens_clone = tokens_clone[:, 1:step + 2] # skip the first index, which is EOS
+ assert not tokens_clone.eq(self.eos).any()
+ tokens_clone[:, step] = self.eos
+ attn_clone = attn.index_select(0, bbsz_idx)[:, :, 1:step+2] if attn is not None else None
+
+ # compute scores per token position
+ pos_scores = scores.index_select(0, bbsz_idx)[:, :step+1]
+ pos_scores[:, step] = eos_scores
+ # convert from cumulative to per-position scores
+ pos_scores[:, 1:] = pos_scores[:, 1:] - pos_scores[:, :-1]
+
+ # normalize sentence-level scores
+ if self.normalize_scores:
+ combined_noisy_channel_eos_scores /= (step + 1) ** self.len_penalty
+
+ cum_unfin = []
+ prev = 0
+ for f in finished:
+ if f:
+ prev += 1
+ else:
+ cum_unfin.append(prev)
+
+ sents_seen = set()
+ for i, (idx, score) in enumerate(zip(bbsz_idx.tolist(), combined_noisy_channel_eos_scores.tolist())):
+ unfin_idx = idx // beam_size
+ sent = unfin_idx + cum_unfin[unfin_idx]
+
+ sents_seen.add((sent, unfin_idx))
+
+ if self.match_source_len and step > src_lengths_no_eos[unfin_idx]:
+ score = -math.inf
+
+ def get_hypo():
+
+ if attn_clone is not None:
+ # remove padding tokens from attn scores
+ hypo_attn = attn_clone[i][nonpad_idxs[sent]]
+ _, alignment = hypo_attn.max(dim=0)
+ else:
+ hypo_attn = None
+ alignment = None
+
+ return {
+ 'tokens': tokens_clone[i],
+ 'score': score,
+ 'attention': hypo_attn, # src_len x tgt_len
+ 'alignment': alignment,
+ 'positional_scores': pos_scores[i],
+ }
+
+ if len(finalized[sent]) < beam_size:
+ finalized[sent].append(get_hypo())
+
+ newly_finished = []
+ for sent, unfin_idx in sents_seen:
+ # check termination conditions for this sentence
+ if not finished[sent] and is_finished(sent, step, unfin_idx):
+ finished[sent] = True
+ newly_finished.append(unfin_idx)
+ return newly_finished
+
+ def noisy_channel_rescoring(lprobs, beam_size, bsz, src_tokens, tokens, k):
+ """Rescore the top k hypothesis from each beam using noisy channel modeling
+ Returns:
+ new_fw_lprobs: the direct model probabilities after pruning the top k
+ new_ch_lm_lprobs: the combined channel and language model probabilities
+ new_lm_lprobs: the language model probabilities after pruning the top k
+ """
+ with torch.no_grad():
+ lprobs_size = lprobs.size()
+ if prefix_tokens is not None and step < prefix_tokens.size(1):
+ probs_slice = lprobs.view(bsz, -1, lprobs.size(-1))[:, 0, :]
+ cand_scores = torch.gather(
+ probs_slice, dim=1,
+ index=prefix_tokens[:, step].view(-1, 1).data
+ ).expand(-1, beam_size).contiguous().view(bsz*beam_size, 1)
+ cand_indices = prefix_tokens[:, step].view(-1, 1).expand(bsz, beam_size).data.contiguous().view(bsz*beam_size, 1)
+
+ # need to calculate and save fw and lm probs for prefix tokens
+ fw_top_k = cand_scores
+ fw_top_k_idx = cand_indices
+ k = 1
+ else:
+ # take the top k best words for every sentence in batch*beam
+ fw_top_k, fw_top_k_idx = torch.topk(lprobs.view(beam_size*bsz, -1), k=k)
+ eos_idx = torch.nonzero(fw_top_k_idx.view(bsz*beam_size*k, -1) == self.eos)[:, 0]
+ ch_scores = fw_top_k.new_full((beam_size*bsz*k, ), 0)
+ src_size = torch.sum(src_tokens[:, :] != self.src_dict.pad_index, dim=1, keepdim=True, dtype=fw_top_k.dtype)
+
+ if self.combine_method != "lm_only":
+ temp_src_tokens_full = src_tokens[:, :].repeat(1, k).view(bsz*beam_size*k, -1)
+ not_padding = temp_src_tokens_full[:, 1:] != self.src_dict.pad_index
+ cur_tgt_size = step+2
+
+ # add eos to all candidate sentences except those that already end in eos
+ eos_tokens = tokens[:, 0].repeat(1, k).view(-1, 1)
+ eos_tokens[eos_idx] = self.tgt_dict.pad_index
+
+ if step == 0:
+ channel_input = torch.cat((fw_top_k_idx.view(-1, 1), eos_tokens), 1)
+ else:
+ # move eos from beginning to end of target sentence
+ channel_input = torch.cat((tokens[:, 1:step + 1].repeat(1, k).view(-1, step), fw_top_k_idx.view(-1, 1), eos_tokens), 1)
+
+ ch_input_lengths = torch.tensor(np.full(channel_input.size(0), cur_tgt_size))
+ ch_input_lengths[eos_idx] = cur_tgt_size-1
+ if self.channel_scoring_type == "unnormalized":
+ ch_encoder_output = channel_model.encoder(channel_input, src_lengths=ch_input_lengths)
+ ch_decoder_output, _ = channel_model.decoder(temp_src_tokens_full, encoder_out=ch_encoder_output, features_only=True)
+ del ch_encoder_output
+ ch_intermed_scores = channel_model.decoder.unnormalized_scores_given_target(ch_decoder_output, target_ids=temp_src_tokens_full[:, 1:])
+ ch_intermed_scores = ch_intermed_scores.float()
+ ch_intermed_scores *= not_padding.float()
+ ch_scores = torch.sum(ch_intermed_scores, dim=1)
+ elif self.channel_scoring_type == "k2_separate":
+ for k_idx in range(k):
+ k_eos_tokens = eos_tokens[k_idx::k, :]
+ if step == 0:
+ k_ch_input = torch.cat((fw_top_k_idx[:, k_idx:k_idx+1], k_eos_tokens), 1)
+ else:
+ # move eos from beginning to end of target sentence
+ k_ch_input = torch.cat((tokens[:, 1:step + 1], fw_top_k_idx[:, k_idx:k_idx+1], k_eos_tokens), 1)
+ k_ch_input_lengths = ch_input_lengths[k_idx::k]
+ k_ch_output = channel_model(k_ch_input, k_ch_input_lengths, src_tokens)
+ k_ch_lprobs = channel_model.get_normalized_probs(k_ch_output, log_probs=True)
+ k_ch_intermed_scores = torch.gather(k_ch_lprobs[:, :-1, :], 2, src_tokens[:, 1:].unsqueeze(2)).squeeze(2)
+ k_ch_intermed_scores *= not_padding.float()
+ ch_scores[k_idx::k] = torch.sum(k_ch_intermed_scores, dim=1)
+ elif self.channel_scoring_type == "src_vocab":
+ ch_encoder_output = channel_model.encoder(channel_input, src_lengths=ch_input_lengths)
+ ch_decoder_output, _ = channel_model.decoder(temp_src_tokens_full, encoder_out=ch_encoder_output, features_only=True)
+
+ del ch_encoder_output
+ ch_lprobs = normalized_scores_with_batch_vocab(
+ channel_model.decoder,
+ ch_decoder_output, src_tokens, k, bsz, beam_size,
+ self.src_dict.pad_index, top_k=self.top_k_vocab)
+ ch_scores = torch.sum(ch_lprobs, dim=1)
+ elif self.channel_scoring_type == "src_vocab_batched":
+ ch_bsz_size = temp_src_tokens_full.shape[0]
+ ch_lprobs_list = [None] * len(range(0, ch_bsz_size, self.ch_scoring_bsz))
+ for i, start_idx in enumerate(range(0, ch_bsz_size, self.ch_scoring_bsz)):
+ end_idx = min(start_idx + self.ch_scoring_bsz, ch_bsz_size)
+ temp_src_tokens_full_batch = temp_src_tokens_full[start_idx:end_idx, :]
+ channel_input_batch = channel_input[start_idx:end_idx, :]
+ ch_input_lengths_batch = ch_input_lengths[start_idx:end_idx]
+ ch_encoder_output_batch = channel_model.encoder(channel_input_batch, src_lengths=ch_input_lengths_batch)
+ ch_decoder_output_batch, _ = channel_model.decoder(temp_src_tokens_full_batch, encoder_out=ch_encoder_output_batch, features_only=True)
+ ch_lprobs_list[i] = normalized_scores_with_batch_vocab(
+ channel_model.decoder,
+ ch_decoder_output_batch, src_tokens, k, bsz, beam_size,
+ self.src_dict.pad_index, top_k=self.top_k_vocab,
+ start_idx=start_idx, end_idx=end_idx)
+ ch_lprobs = torch.cat(ch_lprobs_list, dim=0)
+ ch_scores = torch.sum(ch_lprobs, dim=1)
+ else:
+ ch_output = channel_model(channel_input, ch_input_lengths, temp_src_tokens_full)
+ ch_lprobs = channel_model.get_normalized_probs(ch_output, log_probs=True)
+ ch_intermed_scores = torch.gather(ch_lprobs[:, :-1, :], 2, temp_src_tokens_full[:, 1:].unsqueeze(2)).squeeze().view(bsz*beam_size*k, -1)
+ ch_intermed_scores *= not_padding.float()
+ ch_scores = torch.sum(ch_intermed_scores, dim=1)
+
+ else:
+ cur_tgt_size = 0
+ ch_scores = ch_scores.view(bsz*beam_size, k)
+ expanded_lm_prefix_scores = lm_prefix_scores.unsqueeze(1).expand(-1, k).flatten()
+
+ if self.share_tgt_dict:
+ lm_scores = get_lm_scores(lm, tokens[:, :step + 1].view(-1, step+1), lm_incremental_states, fw_top_k_idx.view(-1, 1), torch.tensor(np.full(tokens.size(0), step+1)), k)
+ else:
+ new_lm_input = dict2dict(tokens[:, :step + 1].view(-1, step+1), self.tgt_to_lm)
+ new_cands = dict2dict(fw_top_k_idx.view(-1, 1), self.tgt_to_lm)
+ lm_scores = get_lm_scores(lm, new_lm_input, lm_incremental_states, new_cands, torch.tensor(np.full(tokens.size(0), step+1)), k)
+
+ lm_scores.add_(expanded_lm_prefix_scores)
+ ch_lm_scores = combine_ch_lm(self.combine_method, ch_scores, lm_scores, src_size, cur_tgt_size)
+ # initialize all as min value
+ new_fw_lprobs = ch_scores.new(lprobs_size).fill_(-1e17).view(bsz*beam_size, -1)
+ new_ch_lm_lprobs = ch_scores.new(lprobs_size).fill_(-1e17).view(bsz*beam_size, -1)
+ new_lm_lprobs = ch_scores.new(lprobs_size).fill_(-1e17).view(bsz*beam_size, -1)
+ new_fw_lprobs[:, self.pad] = -math.inf
+ new_ch_lm_lprobs[:, self.pad] = -math.inf
+ new_lm_lprobs[:, self.pad] = -math.inf
+
+ new_fw_lprobs.scatter_(1, fw_top_k_idx, fw_top_k)
+ new_ch_lm_lprobs.scatter_(1, fw_top_k_idx, ch_lm_scores)
+ new_lm_lprobs.scatter_(1, fw_top_k_idx, lm_scores.view(-1, k))
+ return new_fw_lprobs, new_ch_lm_lprobs, new_lm_lprobs
+
+ def combine_ch_lm(combine_type, ch_scores, lm_scores1, src_size, tgt_size):
+ if self.channel_scoring_type == "unnormalized":
+ ch_scores = self.log_softmax_fn(
+ ch_scores.view(-1, self.beam_size * self.k2)
+ ).view(ch_scores.shape)
+ ch_scores = ch_scores * self.ch_weight
+ lm_scores1 = lm_scores1 * self.lm_weight
+
+ if combine_type == "lm_only":
+ # log P(T|S) + log P(T)
+ ch_scores = lm_scores1.view(ch_scores.size())
+ elif combine_type == "noisy_channel":
+ # 1/t log P(T|S) + 1/s log P(S|T) + 1/t log P(T)
+ if self.normalize_lm_scores_by_tgt_len:
+ ch_scores.div_(src_size)
+ lm_scores_norm = lm_scores1.view(ch_scores.size()).div(tgt_size)
+ ch_scores.add_(lm_scores_norm)
+ # 1/t log P(T|S) + 1/s log P(S|T) + 1/s log P(T)
+ else:
+ ch_scores.add_(lm_scores1.view(ch_scores.size()))
+ ch_scores.div_(src_size)
+
+ return ch_scores
+
+ if self.channel_models is not None:
+ channel_model = self.channel_models[0] # assume only one channel_model model
+ else:
+ channel_model = None
+
+ lm = EnsembleModel(self.lm_models)
+ lm_incremental_states = torch.jit.annotate(
+ List[Dict[str, Dict[str, Optional[Tensor]]]],
+ [
+ torch.jit.annotate(Dict[str, Dict[str, Optional[Tensor]]], {})
+ for i in range(lm.models_size)
+ ],
+ )
+
+ reorder_state = None
+ batch_idxs = None
+ for step in range(max_len + 1): # one extra step for EOS marker
+ # reorder decoder internal states based on the prev choice of beams
+ if reorder_state is not None:
+ if batch_idxs is not None:
+ # update beam indices to take into account removed sentences
+ corr = batch_idxs - torch.arange(batch_idxs.numel()).type_as(batch_idxs)
+ reorder_state.view(-1, beam_size).add_(corr.unsqueeze(-1) * beam_size)
+ model.reorder_incremental_state(incremental_states, reorder_state)
+ encoder_outs = model.reorder_encoder_out(encoder_outs, reorder_state)
+
+ lm.reorder_incremental_state(lm_incremental_states, reorder_state)
+
+ fw_lprobs, avg_attn_scores = model.forward_decoder(
+ tokens[:, :step + 1], encoder_outs, incremental_states, temperature=self.temperature,
+ )
+
+ fw_lprobs[:, self.pad] = -math.inf # never select pad
+ fw_lprobs[:, self.unk] -= self.unk_penalty # apply unk penalty
+ fw_lprobs, ch_lm_lprobs, lm_lprobs = noisy_channel_rescoring(fw_lprobs, beam_size, bsz, src_tokens, tokens, self.k2)
+
+ # handle min and max length constraints
+ if step >= max_len:
+ fw_lprobs[:, :self.eos] = -math.inf
+ fw_lprobs[:, self.eos + 1:] = -math.inf
+ elif step < self.min_len:
+ fw_lprobs[:, self.eos] = -math.inf
+
+ # handle prefix tokens (possibly with different lengths)
+ if prefix_tokens is not None and step < prefix_tokens.size(1):
+ prefix_toks = prefix_tokens[:, step].unsqueeze(-1).repeat(1, beam_size).view(-1)
+ prefix_mask = prefix_toks.ne(self.pad)
+
+ prefix_fw_lprobs = fw_lprobs.gather(-1, prefix_toks.unsqueeze(-1))
+ fw_lprobs[prefix_mask] = -math.inf
+ fw_lprobs[prefix_mask] = fw_lprobs[prefix_mask].scatter_(
+ -1, prefix_toks[prefix_mask].unsqueeze(-1), prefix_fw_lprobs
+ )
+
+ prefix_ch_lm_lprobs = ch_lm_lprobs.gather(-1, prefix_toks.unsqueeze(-1))
+ ch_lm_lprobs[prefix_mask] = -math.inf
+ ch_lm_lprobs[prefix_mask] = ch_lm_lprobs[prefix_mask].scatter_(
+ -1, prefix_toks[prefix_mask].unsqueeze(-1), prefix_ch_lm_lprobs
+ )
+
+ prefix_lm_lprobs = lm_lprobs.gather(-1, prefix_toks.unsqueeze(-1))
+ lm_lprobs[prefix_mask] = -math.inf
+ lm_lprobs[prefix_mask] = lm_lprobs[prefix_mask].scatter_(
+ -1, prefix_toks[prefix_mask].unsqueeze(-1), prefix_lm_lprobs
+ )
+
+ # if prefix includes eos, then we should make sure tokens and
+ # scores are the same across all beams
+ eos_mask = prefix_toks.eq(self.eos)
+ if eos_mask.any():
+ # validate that the first beam matches the prefix
+ first_beam = tokens[eos_mask].view(-1, beam_size, tokens.size(-1))[:, 0, 1:step + 1]
+ eos_mask_batch_dim = eos_mask.view(-1, beam_size)[:, 0]
+ target_prefix = prefix_tokens[eos_mask_batch_dim][:, :step]
+ assert (first_beam == target_prefix).all()
+
+ def replicate_first_beam(tensor, mask):
+ tensor = tensor.view(-1, beam_size, tensor.size(-1))
+ tensor[mask] = tensor[mask][:, :1, :]
+ return tensor.view(-1, tensor.size(-1))
+
+ # copy tokens, scores and lprobs from the first beam to all beams
+ tokens = replicate_first_beam(tokens, eos_mask_batch_dim)
+ scores = replicate_first_beam(scores, eos_mask_batch_dim)
+
+ fw_lprobs = replicate_first_beam(fw_lprobs, eos_mask_batch_dim)
+ ch_lm_lprobs = replicate_first_beam(ch_lm_lprobs, eos_mask_batch_dim)
+ lm_lprobs = replicate_first_beam(lm_lprobs, eos_mask_batch_dim)
+
+ if self.no_repeat_ngram_size > 0:
+ # for each beam and batch sentence, generate a list of previous ngrams
+ gen_ngrams = [{} for bbsz_idx in range(bsz * beam_size)]
+ for bbsz_idx in range(bsz * beam_size):
+ gen_tokens = tokens[bbsz_idx].tolist()
+ for ngram in zip(*[gen_tokens[i:] for i in range(self.no_repeat_ngram_size)]):
+ gen_ngrams[bbsz_idx][tuple(ngram[:-1])] = \
+ gen_ngrams[bbsz_idx].get(tuple(ngram[:-1]), []) + [ngram[-1]]
+
+ # Record attention scores
+ if avg_attn_scores is not None:
+ if attn is None:
+ attn = scores.new(bsz * beam_size, src_tokens.size(1), max_len + 2)
+ attn_buf = attn.clone()
+ nonpad_idxs = src_tokens.ne(self.pad)
+ attn[:, :, step + 1].copy_(avg_attn_scores)
+
+ scores = scores.type_as(fw_lprobs)
+ scores_buf = scores_buf.type_as(fw_lprobs)
+
+ self.search.set_src_lengths(src_lengths_no_eos)
+
+ if self.no_repeat_ngram_size > 0:
+ def calculate_banned_tokens(bbsz_idx):
+ # before decoding the next token, prevent decoding of ngrams that have already appeared
+ ngram_index = tuple(tokens[bbsz_idx, step + 2 - self.no_repeat_ngram_size:step + 1].tolist())
+ return gen_ngrams[bbsz_idx].get(ngram_index, [])
+
+ if step + 2 - self.no_repeat_ngram_size >= 0:
+ # no banned tokens if we haven't generated no_repeat_ngram_size tokens yet
+ banned_tokens = [calculate_banned_tokens(bbsz_idx) for bbsz_idx in range(bsz * beam_size)]
+ else:
+ banned_tokens = [[] for bbsz_idx in range(bsz * beam_size)]
+
+ for bbsz_idx in range(bsz * beam_size):
+ fw_lprobs[bbsz_idx, banned_tokens[bbsz_idx]] = -math.inf
+
+ combined_noisy_channel_scores, fw_lprobs_top_k, lm_lprobs_top_k, cand_indices, cand_beams = self.search.step(
+ step,
+ fw_lprobs.view(bsz, -1, self.vocab_size),
+ scores.view(bsz, beam_size, -1)[:, :, :step], ch_lm_lprobs.view(bsz, -1, self.vocab_size),
+ lm_lprobs.view(bsz, -1, self.vocab_size), self.combine_method
+ )
+
+ # cand_bbsz_idx contains beam indices for the top candidate
+ # hypotheses, with a range of values: [0, bsz*beam_size),
+ # and dimensions: [bsz, cand_size]
+ cand_bbsz_idx = cand_beams.add(bbsz_offsets)
+
+ # finalize hypotheses that end in eos (except for candidates to be ignored)
+ eos_mask = cand_indices.eq(self.eos)
+ eos_mask[:, :beam_size] &= ~cands_to_ignore
+
+ # only consider eos when it's among the top beam_size indices
+ eos_bbsz_idx = torch.masked_select(
+ cand_bbsz_idx[:, :beam_size], mask=eos_mask[:, :beam_size]
+ )
+
+ finalized_sents = set()
+ if eos_bbsz_idx.numel() > 0:
+ eos_scores = torch.masked_select(
+ fw_lprobs_top_k[:, :beam_size], mask=eos_mask[:, :beam_size]
+ )
+ combined_noisy_channel_eos_scores = torch.masked_select(
+ combined_noisy_channel_scores[:, :beam_size],
+ mask=eos_mask[:, :beam_size],
+ )
+
+ # finalize hypo using channel model score
+ finalized_sents = finalize_hypos(
+ step, eos_bbsz_idx, eos_scores, combined_noisy_channel_eos_scores)
+
+ num_remaining_sent -= len(finalized_sents)
+
+ assert num_remaining_sent >= 0
+ if num_remaining_sent == 0:
+ break
+
+ if len(finalized_sents) > 0:
+ new_bsz = bsz - len(finalized_sents)
+
+ # construct batch_idxs which holds indices of batches to keep for the next pass
+ batch_mask = cand_indices.new_ones(bsz)
+ batch_mask[cand_indices.new(finalized_sents)] = 0
+ batch_idxs = torch.nonzero(batch_mask).squeeze(-1)
+
+ eos_mask = eos_mask[batch_idxs]
+ cand_beams = cand_beams[batch_idxs]
+ bbsz_offsets.resize_(new_bsz, 1)
+ cand_bbsz_idx = cand_beams.add(bbsz_offsets)
+
+ lm_lprobs_top_k = lm_lprobs_top_k[batch_idxs]
+
+ fw_lprobs_top_k = fw_lprobs_top_k[batch_idxs]
+ cand_indices = cand_indices[batch_idxs]
+ if prefix_tokens is not None:
+ prefix_tokens = prefix_tokens[batch_idxs]
+ src_lengths_no_eos = src_lengths_no_eos[batch_idxs]
+ cands_to_ignore = cands_to_ignore[batch_idxs]
+
+ scores = scores.view(bsz, -1)[batch_idxs].view(new_bsz * beam_size, -1)
+ scores_buf.resize_as_(scores)
+ tokens = tokens.view(bsz, -1)[batch_idxs].view(new_bsz * beam_size, -1)
+ tokens_buf.resize_as_(tokens)
+ src_tokens = src_tokens.view(bsz, -1)[batch_idxs].view(new_bsz * beam_size, -1)
+ src_lengths = src_lengths.view(bsz, -1)[batch_idxs].view(new_bsz * beam_size, -1)
+ lm_prefix_scores = lm_prefix_scores.view(bsz, -1)[batch_idxs].view(new_bsz * beam_size, -1).squeeze()
+
+ if attn is not None:
+ attn = attn.view(bsz, -1)[batch_idxs].view(new_bsz * beam_size, attn.size(1), -1)
+ attn_buf.resize_as_(attn)
+ bsz = new_bsz
+ else:
+ batch_idxs = None
+
+ # Set active_mask so that values > cand_size indicate eos or
+ # ignored hypos and values < cand_size indicate candidate
+ # active hypos. After this, the min values per row are the top
+ # candidate active hypos.
+ eos_mask[:, :beam_size] |= cands_to_ignore
+ active_mask = torch.add(
+ eos_mask.type_as(cand_offsets) * cand_size,
+ cand_offsets[: eos_mask.size(1)],
+ )
+
+ # get the top beam_size active hypotheses, which are just the hypos
+ # with the smallest values in active_mask
+ active_hypos, new_cands_to_ignore = buffer('active_hypos'), buffer('new_cands_to_ignore')
+ torch.topk(
+ active_mask, k=beam_size, dim=1, largest=False,
+ out=(new_cands_to_ignore, active_hypos)
+ )
+
+ # update cands_to_ignore to ignore any finalized hypos
+ cands_to_ignore = new_cands_to_ignore.ge(cand_size)[:, :beam_size]
+ assert (~cands_to_ignore).any(dim=1).all()
+
+ active_bbsz_idx = buffer('active_bbsz_idx')
+ torch.gather(
+ cand_bbsz_idx, dim=1, index=active_hypos,
+ out=active_bbsz_idx,
+ )
+ active_scores = torch.gather(
+ fw_lprobs_top_k, dim=1, index=active_hypos,
+ out=scores[:, step].view(bsz, beam_size),
+ )
+
+ active_bbsz_idx = active_bbsz_idx.view(-1)
+ active_scores = active_scores.view(-1)
+
+ # copy tokens and scores for active hypotheses
+ torch.index_select(
+ tokens[:, :step + 1], dim=0, index=active_bbsz_idx,
+ out=tokens_buf[:, :step + 1],
+ )
+ torch.gather(
+ cand_indices, dim=1, index=active_hypos,
+ out=tokens_buf.view(bsz, beam_size, -1)[:, :, step + 1],
+ )
+ if step > 0:
+ torch.index_select(
+ scores[:, :step], dim=0, index=active_bbsz_idx,
+ out=scores_buf[:, :step],
+ )
+ torch.gather(
+ fw_lprobs_top_k, dim=1, index=active_hypos,
+ out=scores_buf.view(bsz, beam_size, -1)[:, :, step],
+ )
+ torch.gather(
+ lm_lprobs_top_k, dim=1, index=active_hypos,
+ out=lm_prefix_scores.view(bsz, beam_size)
+ )
+
+ # copy attention for active hypotheses
+ if attn is not None:
+ torch.index_select(
+ attn[:, :, :step + 2], dim=0, index=active_bbsz_idx,
+ out=attn_buf[:, :, :step + 2],
+ )
+
+ # swap buffers
+ tokens, tokens_buf = tokens_buf, tokens
+ scores, scores_buf = scores_buf, scores
+ if attn is not None:
+ attn, attn_buf = attn_buf, attn
+
+ # reorder incremental state in decoder
+ reorder_state = active_bbsz_idx
+
+ # sort by score descending
+ for sent in range(len(finalized)):
+ finalized[sent] = sorted(finalized[sent], key=lambda r: r['score'], reverse=True)
+
+ return finalized
+
+
+def get_lm_scores(model, input_tokens, incremental_states, cand_tokens, input_len, k):
+ with torch.no_grad():
+ lm_lprobs, avg_attn_scores = model.forward_decoder(
+ input_tokens, encoder_outs=None, incremental_states=incremental_states,
+ )
+
+ lm_lprobs_size = lm_lprobs.size(0)
+ probs_next_wrd = torch.gather(lm_lprobs.repeat(1, k).view(lm_lprobs_size*k, -1), 1, cand_tokens).squeeze().view(-1)
+
+ return probs_next_wrd
+
+
+def make_dict2dict(old_dict, new_dict):
+ dict2dict_map = {}
+ for sym in old_dict.symbols:
+ dict2dict_map[old_dict.index(sym)] = new_dict.index(sym)
+ return dict2dict_map
+
+
+def dict2dict(tokens, dict2dict_map):
+ if tokens.device == torch.device('cpu'):
+ tokens_tmp = tokens
+ else:
+ tokens_tmp = tokens.cpu()
+ return tokens_tmp.map_(
+ tokens_tmp,
+ lambda _, val, dict2dict_map=dict2dict_map : dict2dict_map[float(val)]
+ ).to(tokens.device)
+
+
+def reorder_tokens(tokens, lengths, eos):
+ # reorder source tokens so they may be used as reference for P(S|T)
+ return torch.cat((tokens.new([eos]), tokens[-lengths:-1], tokens[:-lengths]), 0)
+
+
+def reorder_all_tokens(tokens, lengths, eos):
+ # used to reorder src tokens from [ .. ] to [...]
+ # so source tokens can be used to predict P(S|T)
+ return torch.stack([reorder_tokens(token, length, eos) for token, length in zip(tokens, lengths)])
+
+
+def normalized_scores_with_batch_vocab(
+ model_decoder, features, target_ids, k, bsz, beam_size,
+ pad_idx, top_k=0, vocab_size_meter=None, start_idx=None,
+ end_idx=None, **kwargs):
+ """
+ Get normalized probabilities (or log probs) from a net's output
+ w.r.t. vocab consisting of target IDs in the batch
+ """
+ if model_decoder.adaptive_softmax is None:
+ weight = model_decoder.output_projection.weight
+ vocab_ids = torch.unique(
+ torch.cat(
+ (torch.unique(target_ids), torch.arange(top_k, device=target_ids.device))
+ )
+ )
+ id_map = dict(zip(vocab_ids.tolist(), range(len(vocab_ids))))
+ mapped_target_ids = target_ids.cpu().apply_(
+ lambda x, id_map=id_map: id_map[x]
+ ).to(target_ids.device)
+ expanded_target_ids = mapped_target_ids[:, :].repeat(1, k).view(bsz*beam_size*k, -1)
+ if start_idx is not None and end_idx is not None:
+ expanded_target_ids = expanded_target_ids[start_idx:end_idx, :]
+ logits = F.linear(features, weight[vocab_ids, :])
+ log_softmax = F.log_softmax(logits, dim=-1, dtype=torch.float32)
+ intermed_scores = torch.gather(
+ log_softmax[:, :-1, :],
+ 2,
+ expanded_target_ids[:, 1:].unsqueeze(2),
+ ).squeeze()
+ not_padding = expanded_target_ids[:, 1:] != pad_idx
+ intermed_scores *= not_padding.float()
+ return intermed_scores
+ else:
+ raise ValueError("adaptive softmax doesn't work with " +
+ "`normalized_scores_with_batch_vocab()`")
diff --git a/fairseq/examples/fast_noisy_channel/noisy_channel_translation.py b/fairseq/examples/fast_noisy_channel/noisy_channel_translation.py
new file mode 100644
index 0000000000000000000000000000000000000000..b74bdfd456f9b7c546ce528173c77431b4f57ac1
--- /dev/null
+++ b/fairseq/examples/fast_noisy_channel/noisy_channel_translation.py
@@ -0,0 +1,127 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+from fairseq.tasks.translation import TranslationTask
+from fairseq.tasks.language_modeling import LanguageModelingTask
+from fairseq import checkpoint_utils
+import argparse
+from fairseq.tasks import register_task
+import torch
+
+
+@register_task("noisy_channel_translation")
+class NoisyChannelTranslation(TranslationTask):
+ """
+ Rescore the top k candidates from each beam using noisy channel modeling
+ """
+
+ @staticmethod
+ def add_args(parser):
+ """Add task-specific arguments to the parser."""
+ TranslationTask.add_args(parser)
+ # fmt: off
+ parser.add_argument('--channel-model', metavar='FILE',
+ help='path to P(S|T) model. P(S|T) and P(T|S) must share source and target dictionaries.')
+ parser.add_argument('--combine-method', default='lm_only',
+ choices=['lm_only', 'noisy_channel'],
+ help="""method for combining direct and channel model scores.
+ lm_only: decode with P(T|S)P(T)
+ noisy_channel: decode with 1/t P(T|S) + 1/s(P(S|T)P(T))""")
+ parser.add_argument('--normalize-lm-scores-by-tgt-len', action='store_true', default=False,
+ help='normalize lm score by target length instead of source length')
+ parser.add_argument('--channel-scoring-type', default='log_norm', choices=['unnormalized', 'log_norm', 'k2_separate', 'src_vocab', 'src_vocab_batched'],
+ help="Normalize bw scores with log softmax or return bw scores without log softmax")
+ parser.add_argument('--top-k-vocab', default=0, type=int,
+ help='top k vocab IDs to use with `src_vocab` in channel model scoring')
+ parser.add_argument('--k2', default=50, type=int,
+ help='the top k2 candidates to rescore with the noisy channel model for each beam')
+ parser.add_argument('--ch-wt', default=1, type=float,
+ help='weight for the channel model')
+ parser.add_argument('--lm-model', metavar='FILE',
+ help='path to lm model file, to model P(T). P(T) must share the same vocab as the direct model on the target side')
+ parser.add_argument('--lm-data', metavar='FILE',
+ help='path to lm model training data for target language, used to properly load LM with correct dictionary')
+ parser.add_argument('--lm-wt', default=1, type=float,
+ help='the weight of the lm in joint decoding')
+ # fmt: on
+
+ def build_generator(
+ self, models, args, seq_gen_cls=None, extra_gen_cls_kwargs=None
+ ):
+ if getattr(args, "score_reference", False):
+ raise NotImplementedError()
+ else:
+ from .noisy_channel_sequence_generator import NoisyChannelSequenceGenerator
+ use_cuda = torch.cuda.is_available() and not self.args.cpu
+ assert self.args.lm_model is not None, '--lm-model required for noisy channel generation!'
+ assert self.args.lm_data is not None, '--lm-data required for noisy channel generation to map between LM and bitext vocabs'
+ if self.args.channel_model is not None:
+ import copy
+ ch_args_task = copy.deepcopy(self.args)
+ tmp = ch_args_task.source_lang
+ ch_args_task.source_lang = ch_args_task.target_lang
+ ch_args_task.target_lang = tmp
+ ch_args_task._name = 'translation'
+ channel_task = TranslationTask.setup_task(ch_args_task)
+
+ arg_dict = {}
+ arg_dict['task'] = 'language_modeling'
+ arg_dict['sample_break_mode'] = 'eos'
+ arg_dict['data'] = self.args.lm_data
+ arg_dict['output_dictionary_size'] = -1
+ lm_args = argparse.Namespace(**arg_dict)
+ lm_task = LanguageModelingTask.setup_task(lm_args)
+ lm_dict = lm_task.output_dictionary
+
+ if self.args.channel_model is not None:
+ channel_models, _ = checkpoint_utils.load_model_ensemble(self.args.channel_model.split(':'), task=channel_task)
+
+ for model in channel_models:
+ model.make_generation_fast_(
+ beamable_mm_beam_size=None if args.no_beamable_mm else args.beam,
+ need_attn=args.print_alignment,
+ )
+ if self.args.fp16:
+ model.half()
+ if use_cuda:
+ model.cuda()
+ else:
+ channel_models = None
+
+ lm_models, _ = checkpoint_utils.load_model_ensemble(self.args.lm_model.split(':'), task=lm_task)
+
+ for model in lm_models:
+ model.make_generation_fast_(
+ beamable_mm_beam_size=None if args.no_beamable_mm else args.beam,
+ need_attn=args.print_alignment,
+ )
+ if self.args.fp16:
+ model.half()
+ if use_cuda:
+ model.cuda()
+ return NoisyChannelSequenceGenerator(
+ combine_method=self.args.combine_method,
+ tgt_dict=self.target_dictionary,
+ src_dict=self.source_dictionary,
+ beam_size=getattr(args, 'beam', 5),
+ max_len_a=getattr(args, 'max_len_a', 0),
+ max_len_b=getattr(args, 'max_len_b', 200),
+ min_len=getattr(args, 'min_len', 1),
+ len_penalty=getattr(args, 'lenpen', 1),
+ unk_penalty=getattr(args, 'unkpen', 0),
+ temperature=getattr(args, 'temperature', 1.),
+ match_source_len=getattr(args, 'match_source_len', False),
+ no_repeat_ngram_size=getattr(args, 'no_repeat_ngram_size', 0),
+ normalize_scores=(not getattr(args, 'unnormalized', False)),
+ channel_models=channel_models,
+ k2=getattr(self.args, 'k2', 50),
+ ch_weight=getattr(self.args, 'ch_wt', 1),
+ channel_scoring_type=self.args.channel_scoring_type,
+ top_k_vocab=self.args.top_k_vocab,
+ lm_models=lm_models,
+ lm_dict=lm_dict,
+ lm_weight=getattr(self.args, 'lm_wt', 1),
+ normalize_lm_scores_by_tgt_len=getattr(self.args, 'normalize_lm_scores_by_tgt_len', False),
+ )
diff --git a/fairseq/examples/flores101/README.md b/fairseq/examples/flores101/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..635c13f40bd0ccab704735bc5c26ea0192ea98cd
--- /dev/null
+++ b/fairseq/examples/flores101/README.md
@@ -0,0 +1,223 @@
+
+
+
+
+# Flores101: Large-Scale Multilingual Machine Translation
+
+## Introduction
+
+Baseline pretrained models for small and large tracks of WMT 21 Large-Scale Multilingual Machine Translation competition.
+
+Flores Task at WMT 21: http://www.statmt.org/wmt21/large-scale-multilingual-translation-task.html
+
+Flores announement blog post: https://ai.facebook.com/blog/flores-researchers-kick-off-multilingual-translation-challenge-at-wmt-and-call-for-compute-grants/
+
+
+
+## Pretrained models
+
+Model | Num layers | Embed dimension | FFN dimension| Vocab Size | #params | Download
+---|---|---|---|---|---|---
+`flores101_mm100_615M` | 12 | 1024 | 4096 | 256,000 | 615M | https://dl.fbaipublicfiles.com/flores101/pretrained_models/flores101_mm100_615M.tar.gz
+`flores101_mm100_175M` | 6 | 512 | 2048 | 256,000 | 175M | https://dl.fbaipublicfiles.com/flores101/pretrained_models/flores101_mm100_175M.tar.gz
+
+
+These models are trained similar to [M2M-100](https://arxiv.org/abs/2010.11125) with additional support for the languages that are part of the WMT Large-Scale Multilingual Machine Translation track. Full list of languages can be found at the bottom.
+
+
+## Example Generation code
+
+### Download model, sentencepiece vocab
+
+```bash
+fairseq=/path/to/fairseq
+cd $fairseq
+
+# Download 615M param model.
+wget https://dl.fbaipublicfiles.com/flores101/pretrained_models/flores101_mm100_615M.tar.gz
+
+# Extract
+tar -xvzf flores101_mm100_615M.tar.gz
+```
+
+### Encode using our SentencePiece Model
+Note: Install SentencePiece from [here](https://github.com/google/sentencepiece)
+
+
+```bash
+fairseq=/path/to/fairseq
+cd $fairseq
+
+# Download example dataset From German to French
+sacrebleu --echo src -l de-fr -t wmt19 | head -n 20 > raw_input.de-fr.de
+sacrebleu --echo ref -l de-fr -t wmt19 | head -n 20 > raw_input.de-fr.fr
+
+for lang in de fr ; do
+ python scripts/spm_encode.py \
+ --model flores101_mm100_615M/sentencepiece.bpe.model \
+ --output_format=piece \
+ --inputs=raw_input.de-fr.${lang} \
+ --outputs=spm.de-fr.${lang}
+done
+```
+
+### Binarization
+
+```bash
+fairseq-preprocess \
+ --source-lang de --target-lang fr \
+ --testpref spm.de-fr \
+ --thresholdsrc 0 --thresholdtgt 0 \
+ --destdir data_bin \
+ --srcdict flores101_mm100_615M/dict.txt --tgtdict flores101_mm100_615M/dict.txt
+```
+
+### Generation
+
+
+```bash
+fairseq-generate \
+ data_bin \
+ --batch-size 1 \
+ --path flores101_mm100_615M/model.pt \
+ --fixed-dictionary flores101_mm100_615M/dict.txt \
+ -s de -t fr \
+ --remove-bpe 'sentencepiece' \
+ --beam 5 \
+ --task translation_multi_simple_epoch \
+ --lang-pairs flores101_mm100_615M/language_pairs.txt \
+ --decoder-langtok --encoder-langtok src \
+ --gen-subset test \
+ --fp16 \
+ --dataset-impl mmap \
+ --distributed-world-size 1 --distributed-no-spawn
+```
+
+### Supported Languages and lang code
+
+Language | lang code
+---|---
+Akrikaans | af
+Amharic | am
+Arabic | ar
+Assamese | as
+Asturian | ast
+Aymara | ay
+Azerbaijani | az
+Bashkir | ba
+Belarusian | be
+Bulgarian | bg
+Bengali | bn
+Breton | br
+Bosnian | bs
+Catalan | ca
+Cebuano | ceb
+Chokwe | cjk
+Czech | cs
+Welsh | cy
+Danish | da
+German | de
+Dyula| dyu
+Greek | el
+English | en
+Spanish | es
+Estonian | et
+Persian | fa
+Fulah | ff
+Finnish | fi
+French | fr
+Western Frisian | fy
+Irish | ga
+Scottish Gaelic | gd
+Galician | gl
+Gujarati | gu
+Hausa | ha
+Hebrew | he
+Hindi | hi
+Croatian | hr
+Haitian Creole | ht
+Hungarian | hu
+Armenian | hy
+Indonesian | id
+Igbo | ig
+Iloko | ilo
+Icelandic | is
+Italian | it
+Japanese | ja
+Javanese | jv
+Georgian | ka
+Kachin | kac
+Kamba | kam
+Kabuverdianu | kea
+Kongo | kg
+Kazakh | kk
+Central Khmer | km
+Kimbundu | kmb
+Northern Kurdish | kmr
+Kannada | kn
+Korean | ko
+Kurdish | ku
+Kyrgyz | ky
+Luxembourgish | lb
+Ganda | lg
+Lingala | ln
+Lao | lo
+Lithuanian | lt
+Luo | luo
+Latvian | lv
+Malagasy | mg
+Maori | mi
+Macedonian | mk
+Malayalam | ml
+Mongolian | mn
+Marathi | mr
+Malay | ms
+Maltese | mt
+Burmese | my
+Nepali | ne
+Dutch | nl
+Norwegian | no
+Northern Sotho | ns
+Nyanja | ny
+Occitan | oc
+Oromo | om
+Oriya | or
+Punjabi | pa
+Polish | pl
+Pashto | ps
+Portuguese | pt
+Quechua | qu
+Romanian | ro
+Russian | ru
+Sindhi | sd
+Shan | shn
+Sinhala | si
+Slovak | sk
+Slovenian | sl
+Shona | sn
+Somali | so
+Albanian | sq
+Serbian | sr
+Swati | ss
+Sundanese | su
+Swedish | sv
+Swahili | sw
+Tamil | ta
+Telugu | te
+Tajik | tg
+Thai | th
+Tigrinya | ti
+Tagalog | tl
+Tswana | tn
+Turkish | tr
+Ukrainian | uk
+Umbundu | umb
+Urdu | ur
+Uzbek | uz
+Vietnamese | vi
+Wolof | wo
+Xhosa | xh
+Yiddish | yi
+Yoruba | yo
+Chinese| zh
+Zulu | zu
diff --git a/fairseq/examples/flores101/flores_logo.png b/fairseq/examples/flores101/flores_logo.png
new file mode 100644
index 0000000000000000000000000000000000000000..6dfba86310ca0754f1212f48ff363ca7f8eb4286
--- /dev/null
+++ b/fairseq/examples/flores101/flores_logo.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:430c4eeefe89b1527d74ab88c04247cd7e21c1c62aaa05947625ef9ce8aa67ed
+size 33184
diff --git a/fairseq/examples/fully_sharded_data_parallel/README.md b/fairseq/examples/fully_sharded_data_parallel/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..b9e44fef48bee5faeee27b3d1d1b1eb96b6a477f
--- /dev/null
+++ b/fairseq/examples/fully_sharded_data_parallel/README.md
@@ -0,0 +1,177 @@
+# Fully Sharded Data Parallel (FSDP)
+
+## Overview
+Recent work by [Microsoft](https://arxiv.org/abs/1910.02054) and
+[Google](https://arxiv.org/abs/2004.13336) has shown that data parallel
+training can be made significantly more efficient by sharding the model
+parameters and optimizer state across data parallel workers. These ideas are
+encapsulated in the new **`FullyShardedDataParallel` (FSDP)** wrapper provided
+by [fairscale](https://github.com/facebookresearch/fairscale/).
+
+Compared to PyTorch DDP:
+* FSDP produces identical results as PyTorch DDP (it's still synchronous data parallel training)
+* FSDP shards parameters (FP16 + FP32) and optimizer state across data parallel GPUs
+* FSDP is faster than PyTorch DDP because the optimizer step is sharded, and the communication can be overlapped with the forward pass
+* FSDP enables training 13B parameter models on 8 GPUs and 175B parameter models on 128 GPUs
+
+FSDP is fully supported in fairseq via the following new arguments:
+* `--ddp-backend=fully_sharded`: enables full sharding via FSDP
+* `--cpu-offload`: offloads the optimizer state and FP32 model copy to CPU (combine with `--optimizer=cpu_adam`)
+* `--no-reshard-after-forward`: increases training speed for large models (1B+ params) and is similar to ZeRO stage 2
+* other popular options (`--fp16`, `--update-freq`, `--checkpoint-activations`, `--offload-activations`, etc.) continue to work as normal
+
+Limitations
+
+FSDP currently has several limitations compared to fairseq's default DDP backend (PyTorch DDP):
+* while FSDP is full compatible with pointwise Optimizers (e.g., Adam, AdamW, Adadelta, Adamax, SGD, etc.), it is not currently compatible with non-pointwise Optimizers (e.g., Adagrad, Adafactor, LAMB, etc.)
+* FSDP depends on flattening the parameters, so models that currently require `--fp16-no-flatten-grads` may not be supported
+
+See the [fairscale docs](https://fairscale.readthedocs.io/en/latest/api/nn/fsdp_tips.html) for a more detailed
+explanation of these and other limitations.
+
+
+
+How it works
+
+
+
+See the [fairscale docs](https://fairscale.readthedocs.io/en/latest/api/nn/fsdp_tips.html) for a more detailed
+explanation of how FSDP works.
+
+
+
+## Example usage
+
+The following examples illustrate how to train a very large language model with
+13 billion parameters on 1 GPU by offloading parameters and optimizer states to
+CPU, or on 8 GPUs by fully sharding the params and optimizer states across GPUs.
+
+These examples use the WikiText-103 dataset for demonstration purposes, but
+in practice a much larger dataset will be needed to achieve good results.
+Follow the [instructions here](https://github.com/pytorch/fairseq/blob/main/examples/roberta/README.pretraining.md#1-preprocess-the-data)
+to preprocess the WikiText-103 dataset using the GPT-2/RoBERTa vocabulary.
+
+### 13B params on 1 V100 GPU (with CPU offloading)
+
+The following command trains a 13B parameter GPT-3 model on a single V100 GPU
+using the `--cpu-offload` feature to offload parameters and optimizer states to
+CPU. In this setting, the optimizer step (Adam) happens on CPU. We also use the
+`--checkpoint-activations` feature (sometimes called [gradient checkpointing](https://pytorch.org/docs/stable/checkpoint.html)),
+which further saves memory in exchange for a small increase in computation.
+
+**Requirements:**
+- Install the latest master version of fairscale: `pip install git+https://github.com/facebookresearch/fairscale.git@master`
+- You'll need 32GB of GPU memory and ~256GB of system memory to train the 13B param model.
+- If you have less system memory, the 6.7B param model can be trained with ~128GB of system memory, just set `--arch transformer_lm_gpt3_6_7`
+- We use the CPU Adam optimizer from [DeepSpeed](https://github.com/microsoft/DeepSpeed), so you'll need to `pip install deepspeed` before running the command.
+
+**Notes:**
+- The command will take ~5 minutes to start training, during which time it will appear to be hung, since randomly initializing 13B weights can be slow.
+- The `--cpu-offload` feature requires training in mixed precision (`--fp16`).
+- Tune the `OMP_NUM_THREADS` env variable for best performance with CPU offloading.
+- The example command below stops training after 10 steps (`--max-update 10`) and does not save checkpoints (`--no-save`).
+
+```bash
+OMP_NUM_THREADS=20 CUDA_VISIBLE_DEVICES=0 \
+ fairseq-train data-bin/wikitext-103-roberta-bpe-bin \
+ --ddp-backend fully_sharded --fp16 --fp16-init-scale 4 \
+ --cpu-offload --checkpoint-activations \
+ --task language_modeling --tokens-per-sample 2048 --batch-size 8 \
+ --arch transformer_lm_gpt3_13 \
+ --optimizer cpu_adam --adam-betas "(0.9,0.98)" \
+ --lr 0.0001 --lr-scheduler polynomial_decay --warmup-updates 5 --total-num-update 10 \
+ --max-update 10 --no-save --log-format json --log-interval 1
+```
+
+Example output
+
+### 13B params on 8 V100 GPUs (with full parameter + optimizer state sharding)
+
+FSDP can also shard the parameters and optimizer states across multiple GPUs,
+reducing memory requirements significantly. On 8 x 32GB GPUs, sharding enables
+training the same 13B parameter model *without offloading the parameters to
+CPU*. However, without CPU offloading we'd only be able to fit a batch size of
+1 per GPU, which would cause training speed to suffer.
+
+We obtain the best performance on 8 GPUs by combining full sharding and CPU
+offloading. The following command trains the same 13B parameter GPT-3 model as
+before on 8 x 32GB V100 GPUs; training speed increases superlinearly from ~310
+words per second to ~3200 words per second.
+
+```bash
+OMP_NUM_THREADS=20 CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
+ fairseq-train data-bin/wikitext-103-roberta-bpe-bin \
+ --ddp-backend fully_sharded --fp16 --fp16-init-scale 4 \
+ --cpu-offload --checkpoint-activations \
+ --task language_modeling --tokens-per-sample 2048 --batch-size 8 \
+ --arch transformer_lm_gpt3_13 \
+ --optimizer cpu_adam --adam-betas "(0.9,0.98)" \
+ --lr 0.0001 --lr-scheduler polynomial_decay --warmup-updates 5 --total-num-update 10 \
+ --max-update 10 --no-save --log-format json --log-interval 1
+```
+
+Example output
+
+
+
+
+
+
+ 
+ 
+ 
+ 
+ 
+ 
+ 
+ 
+
+
+VLM is a masked language model style pre-training using only one encoder with masked modality model (MMM) for retrieval/generation/sequence labeling style tasks.
+
+
+
+### News
+[Oct. 2021] Initial release of implementation for the following papers:
+[VideoCLIP: Contrastive Pre-training for Zero-shot Video-Text Understanding](https://arxiv.org/pdf/2109.14084.pdf) (Xu et. al., EMNLP 2021)
+[VLM: Task-agnostic Video-Language Model Pre-training for Video Understanding](https://aclanthology.org/2021.findings-acl.370.pdf) (Xu et. al., ACL Findings 2021)
+
+
+### Installation
+We aim to minimize the dependency of this repo on other packages.
+We use fairseq as the main trainer (no models/datasets dependency on fairseq. We will support other trainer in future):
+```
+git clone https://github.com/pytorch/fairseq
+cd fairseq
+pip install -e . # also optionally follow fairseq README for apex installation for fp16 training.
+export MKL_THREADING_LAYER=GNU # fairseq may need this for numpy.
+```
+
+Then install this toolkit:
+```
+cd examples/MMPT # MMPT can be in any folder, not necessarily under fairseq/examples.
+pip install -e .
+```
+
+The code is developed under Python=3.8.8, Pytorch=1.8, cuda=11.0 with fairseq=1.0.0a0+af0389f and tested under Python=3.8.8 pytorch=1.9 cuda=11.0 fairseq=1.0.0a0+8e7bc73 during code release.
+Most models require `transformers==3.4` for API compatibility `pip install transformers==3.4`.
+In addition, some downstream tasks may need `conda install pandas`.
+
+
+### Usage
+#### Download Checkpoints
+We use pre-trained [S3D](https://github.com/antoine77340/S3D_HowTo100M) for video feature extraction. Please place the models as `pretrained_models/s3d_dict.npy` and `pretrained_models/s3d_howto100m.pth`.
+
+Download VideoCLIP checkpoint `https://dl.fbaipublicfiles.com/MMPT/retri/videoclip/checkpoint_best.pt` to `runs/retri/videoclip` or VLM checkpoint `https://dl.fbaipublicfiles.com/MMPT/mtm/vlm/checkpoint_best.pt` to `runs/mtm/vlm`.
+
+#### Demo of Inference
+run `python locallaunch.py projects/retri/videoclip.yaml --dryrun` to get all `.yaml`s for VideoCLIP.
+
+```python
+import torch
+
+from mmpt.models import MMPTModel
+
+
+model, tokenizer, aligner = MMPTModel.from_pretrained(
+ "projects/retri/videoclip/how2.yaml")
+
+model.eval()
+
+
+# B, T, FPS, H, W, C (VideoCLIP is trained on 30 fps of s3d)
+video_frames = torch.randn(1, 2, 30, 224, 224, 3)
+caps, cmasks = aligner._build_text_seq(
+ tokenizer("some text", add_special_tokens=False)["input_ids"]
+)
+
+caps, cmasks = caps[None, :], cmasks[None, :] # bsz=1
+
+with torch.no_grad():
+ output = model(video_frames, caps, cmasks, return_score=True)
+print(output["score"]) # dot-product
+```
+
+#### Data Preparation
+See [dataset](DATASET.md) for each dataset.
+
+#### Global Config for Training Pipeline
+We organize a global config file for a training/testing pipeline under projects (see a detailed [explanation](CONFIG.md)). For example, VideoCLIP in `projects/retri/videoclip.yaml` and VLM is in `projects/mtm/vlm.yaml`.
+
+We wrap all cmds into `locallaunch.py` and `mmpt_cli/localjob.py`. You can check concrete cmds by `--dryrun` and then drop it for actual run.
+
+First, run `python locallaunch.py projects/retri/videoclip.yaml --dryrun` will generate configs for all configs of pre-training, zero-shot evaluation, fine-tuning and testing, for VideoCLIP under `projects/retri/videoclip`.
+
+Then each (either training or evaluation) process will be configed by a concrete config file (we save all complex arguments into the concrete config file for reproducibility, including fairseq args). For example, run zero-shot evaluation on youcook,
+```
+python locallaunch.py projects/retri/videoclip/test_youcook_zs.yaml --jobtype local_predict # zero-shot evaluation.
+python locallaunch.py projects/retri/videoclip/youcook_videoclip.yaml --jobtype local_single --dryrun # fine-tuning: use --dryrun to check cmds and drop it to make an actual run; local_small will run on two gpus (as in paper).
+python locallaunch.py projects/retri/videoclip/test_youcook_videoclip.yaml --jobtype local_predict # testing on fine-tuned model.
+```
+
+Pretraining can be run as:
+```
+python locallaunch.py projects/retri/videoclip/how2.yaml --jobtype local_single --dryrun # check then drop dryrun; paper is ran on local_big as 8 gpus.
+```
+You may need to change `--jobtype`, check/extend `LocalJob` in `mmpt_cli/localjob.py` for multi-gpu/multi-node pre-training.
+
+The detailed instructions of pretraining and fine-tuning can be found at [pretraining instruction](pretraining.md) and [finetuning instruction](endtask.md).
+
+
+### Development
+Several components of this toolkit can be re-used for future research (and also our ongoing research).
+
+#### Framework Wrapper
+We currently only support fairseq, but most components can be easily fit into other frameworks like huggingface. This repo is a `--user-dir` of fairseq with fairseq wrapper. For example, `mmpt/tasks` includes a `FairseqMMTTask`, which manages `mmpt/datasets` with `FairseqDataset`, `mmpt/models` with `FairseqModel`, `mmpt/losses` with `FairseqCriterion`.
+
+#### Processors
+**Multi**modal research introduces the complexity on modality alignment from different input sources to losses. Inspired by [MMF](https://github.com/facebookresearch/mmf), this toolkit leverages `mmpt/processors` to handle various needs of data preprocessing and loading, **alleviating** the needs of multiple `torch.data.utils.Dataset` (that can be tricky for ablation study).
+Processors can also be decoupled from `torch.data.utils.Dataset` for offline preprocessing instead of on-the-fly data preprocessing.
+
+We decouple a `mmpt.MMDataset` as 3 types of processors: `MetaProcessor`, `VideoProcessor`, `TextProcessor` and `Aligner`. They can be configed in `dataset` field of a config file (e.g., see `projects/task/how2.yaml`).
+`MetaProcessor` is used to load the meta data about a dataset, aka, all video_ids of how2 dataset.
+`VideoProcessor` is used to load the video features about a dataset. For example, S3D features for each second of a video.
+`TextProcessor` is used to load the text (feature). For example, BERT pre-tokenized text clips for how2 dataset (with `start`s, `end`s of timestamps and `cap` for `token_ids`).
+`Aligner` is the core class for different baselines that prepares the training data. For example, sampling a clip, masking tokens for MLM, etc.
+
+#### Performance-tuned Components
+To speed up pre-training, this toolkit uses sharded features stored in mmaped numpy, backed by `ShardedTensor` in `mmpt/utils/shardedtensor.py` (adopted from MARGE paper). This reduces the loads of IO for multi-GPU training without loading all features for a video into the memory each time and `ShardedTensor` ensure features are stored in continuous disk space for near random access. This is used for both How2 video features and texts in `mmpt/processors/how2processor.py`.
+
+
+### Citation
+If this codebase is useful for your work, please cite the following papers:
+
+```BibTeX
+@inproceedings{xu-etal-2021-videoclip,
+ title = "{VideoCLIP}: Contrastive Pre-training for\\Zero-shot Video-Text Understanding",
+ author = "Xu, Hu and
+ Ghosh, Gargi and
+ Huang, Po-Yao and
+ Okhonko, Dmytro and
+ Aghajanyan, Armen and
+ Metze, Florian and
+ Zettlemoyer, Luke and
+ Feichtenhofer, Christoph",
+ booktitle = "Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
+ month = nov,
+ year = "2021",
+ address = "Online",
+ publisher = "Association for Computational Linguistics",
+}
+
+@inproceedings{xu-etal-2021-vlm,
+ title = "{VLM}: Task-agnostic Video-Language Model Pre-training for Video Understanding",
+ author = "Xu, Hu and
+ Ghosh, Gargi and
+ Huang, Po-Yao and
+ Arora, Prahal and
+ Aminzadeh, Masoumeh and
+ Feichtenhofer, Christoph and
+ Metze, Florian and
+ Zettlemoyer, Luke",
+ booktitle = "Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021",
+ month = aug,
+ year = "2021",
+ address = "Online",
+ publisher = "Association for Computational Linguistics",
+ url = "https://aclanthology.org/2021.findings-acl.370",
+ doi = "10.18653/v1/2021.findings-acl.370",
+ pages = "4227--4239",
+}
+```
+
+### Bug Reports
+This repo is in its initial stage, welcome bug reports to huxu@fb.com
+
+### Copyright
+The majority of Multimodal Pre-training (MMPT) is licensed under CC-BY-NC, however portions of the project are available under separate license terms: Evaluation Codes/Models: Howto100M and HuggingFace Transformers are licensed under the Apache2.0 license; COIN and NLG-eval are licensed under the MIT license; CrossTask is licensed under the BSD-3; DiDeMo is licensed under the BSD-2 license.
diff --git a/fairseq/examples/MMPT/endtask.md b/fairseq/examples/MMPT/endtask.md
new file mode 100644
index 0000000000000000000000000000000000000000..7690955327283ded4b37857a4a7b78463e0eb0f8
--- /dev/null
+++ b/fairseq/examples/MMPT/endtask.md
@@ -0,0 +1,41 @@
+# Zero-shot Transfer and Finetuning
+
+(If you are new to the ideas of `mmpt.processors`, see [README](README.md) first.)
+All finetuning datasets (specifically `processors`) are defined in `mmpt.processors.dsprocessor`.
+Given the complexity of different types of finetuning tasks, each task may have their own meta/video/text/aligner processors and `mmpt/evaluators/{Predictor,Metric}`.
+
+### Tasks
+
+Currently, we support 5 end datasets: `MSRVTT`, `Youcook`, `COIN`, `Crosstask` and `DiDeMo` with the following tasks:
+text-video retrieval: `MSRVTT`, `Youcook`, `DiDeMo`;
+video captioning: `Youcook`;
+Video Question and Answering: `MSRVTT-QA`.
+
+To add your own dataset, you can specify the corresponding processors and config them in the `dataset` field of a config file, such as `projects/task/vtt.yaml`.
+
+### Zero-shot Transfer (no Training)
+Zero-shot transfer will run the pre-trained model (e.g., VideoCLIP) directly on testing data. Configs with pattern: `projects/task/*_zs_*.yaml` are dedicated for zero-shot transfer.
+
+### Fine-tuning
+
+The training of a downstream task is similar to pretraining, execept you may need to specify the `restore_file` in `fairseq.checkpoint` and reset optimizers, see `projects/task/ft.yaml` that is included by `projects/task/vtt.yaml`.
+
+We typically do finetuning on 2 gpus (`local_small`).
+
+### Testing
+For each finetuning dataset, you may need to specify a testing config, similar to `projects/task/test_vtt.yaml`.
+
+We define `mmpt.evaluators.Predictor` for different types of prediction. For example, `MSRVTT` and `Youcook` are video-retrieval tasks and expecting to use `RetrievalPredictor`. You may need to define your new type of predictors and specify that in `predictor` field of a testing config.
+
+Each task may also have their own metric for evaluation. This can be created in `mmpt.evaluators.Metric` and specified in the `metric` field of a testing config.
+
+Launching a testing is as simple as training by specifying the path of a testing config:
+```python locallaunch.py projects/mfmmlm/test_vtt.yaml```
+Testing will be launched locally by default since prediction is computationally less expensive.
+
+### Third-party Libraries
+We list the following finetuning tasks that require third-party libraries.
+
+Youcook captioning: `https://github.com/Maluuba/nlg-eval`
+
+CrossTask: `https://github.com/DmZhukov/CrossTask`'s `dp` under `third-party/CrossTask` (`python setup.py build_ext --inplace`)
diff --git a/fairseq/examples/MMPT/locallaunch.py b/fairseq/examples/MMPT/locallaunch.py
new file mode 100644
index 0000000000000000000000000000000000000000..e20fd816fa3bed8b1af8f6a4d1a07ccb69a1fffa
--- /dev/null
+++ b/fairseq/examples/MMPT/locallaunch.py
@@ -0,0 +1,148 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+import argparse
+import os
+
+from omegaconf import OmegaConf
+
+from mmpt.utils import recursive_config, overwrite_dir
+from mmpt_cli.localjob import LocalJob
+
+
+class JobLauncher(object):
+ JOB_CONFIG = {
+ "local": LocalJob,
+ }
+
+ def __init__(self, yaml_file):
+ self.yaml_file = yaml_file
+ job_key = "local"
+
+ if yaml_file.endswith(".yaml"):
+ config = recursive_config(yaml_file)
+ if config.task_type is not None:
+ job_key = config.task_type.split("_")[0]
+ else:
+ raise ValueError("unknown extension of job file:", yaml_file)
+ self.job_key = job_key
+
+ def __call__(self, job_type=None, dryrun=False):
+ if job_type is not None:
+ self.job_key = job_type.split("_")[0]
+ print("[JobLauncher] job_key", self.job_key)
+ job = JobLauncher.JOB_CONFIG[self.job_key](
+ self.yaml_file, job_type=job_type, dryrun=dryrun)
+ return job.submit()
+
+
+class Pipeline(object):
+ """a job that loads yaml config."""
+
+ def __init__(self, fn):
+ """
+ load a yaml config of a job and save generated configs as yaml for each task.
+ return: a list of files to run as specified by `run_task`.
+ """
+ if fn.endswith(".py"):
+ # a python command.
+ self.backend = "python"
+ self.run_yamls = [fn]
+ return
+
+ job_config = recursive_config(fn)
+ if job_config.base_dir is None: # single file job config.
+ self.run_yamls = [fn]
+ return
+
+ self.project_dir = os.path.join("projects", job_config.project_dir)
+ self.run_dir = os.path.join("runs", job_config.project_dir)
+
+ if job_config.run_task is not None:
+ run_yamls = []
+ for stage in job_config.run_task:
+ # each stage can have multiple tasks running in parallel.
+ if OmegaConf.is_list(stage):
+ stage_yamls = []
+ for task_file in stage:
+ stage_yamls.append(
+ os.path.join(self.project_dir, task_file))
+ run_yamls.append(stage_yamls)
+ else:
+ run_yamls.append(os.path.join(self.project_dir, stage))
+ self.run_yamls = run_yamls
+ configs_to_save = self._overwrite_task(job_config)
+ self._save_configs(configs_to_save)
+
+ def __getitem__(self, idx):
+ yaml_files = self.run_yamls[idx]
+ if isinstance(yaml_files, list):
+ return [JobLauncher(yaml_file) for yaml_file in yaml_files]
+ return [JobLauncher(yaml_files)]
+
+ def __len__(self):
+ return len(self.run_yamls)
+
+ def _save_configs(self, configs_to_save: dict):
+ # save
+ os.makedirs(self.project_dir, exist_ok=True)
+ for config_file in configs_to_save:
+ config = configs_to_save[config_file]
+ print("saving", config_file)
+ OmegaConf.save(config=config, f=config_file)
+
+ def _overwrite_task(self, job_config):
+ configs_to_save = {}
+ self.base_project_dir = os.path.join("projects", job_config.base_dir)
+ self.base_run_dir = os.path.join("runs", job_config.base_dir)
+
+ for config_sets in job_config.task_group:
+ overwrite_config = job_config.task_group[config_sets]
+ if (
+ overwrite_config.task_list is None
+ or len(overwrite_config.task_list) == 0
+ ):
+ print(
+ "[warning]",
+ job_config.task_group,
+ "has no task_list specified.")
+ # we don't want this added to a final config.
+ task_list = overwrite_config.pop("task_list", None)
+ for config_file in task_list:
+ config_file_path = os.path.join(
+ self.base_project_dir, config_file)
+ config = recursive_config(config_file_path)
+ # overwrite it.
+ if overwrite_config:
+ config = OmegaConf.merge(config, overwrite_config)
+ overwrite_dir(config, self.run_dir, basedir=self.base_run_dir)
+ save_file_path = os.path.join(self.project_dir, config_file)
+ configs_to_save[save_file_path] = config
+ return configs_to_save
+
+
+def main(args):
+ job_type = args.jobtype if args.jobtype else None
+ # parse multiple pipelines.
+ pipelines = [Pipeline(fn) for fn in args.yamls.split(",")]
+
+ for pipe_id, pipeline in enumerate(pipelines):
+ if not hasattr(pipeline, "project_dir"):
+ for job in pipeline[0]:
+ job(job_type=job_type, dryrun=args.dryrun)
+
+
+if __name__ == "__main__":
+ parser = argparse.ArgumentParser()
+ parser.add_argument("yamls", type=str)
+ parser.add_argument(
+ "--dryrun",
+ action="store_true",
+ help="run config and prepare to submit without launch the job.",
+ )
+ parser.add_argument(
+ "--jobtype", type=str, default="",
+ help="force to run jobs as specified.")
+ args = parser.parse_args()
+ main(args)
diff --git a/fairseq/examples/MMPT/mmpt/__init__.py b/fairseq/examples/MMPT/mmpt/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..6ff86ddd5ce0c454281e6568c628bd3c49ea5024
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/__init__.py
@@ -0,0 +1,12 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+try:
+ # fairseq user dir
+ from .datasets import FairseqMMDataset
+ from .losses import FairseqCriterion
+ from .models import FairseqMMModel
+ from .tasks import FairseqMMTask
+except ImportError:
+ pass
diff --git a/fairseq/examples/MMPT/mmpt/datasets/__init__.py b/fairseq/examples/MMPT/mmpt/datasets/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..2578235e1771fdc7e6fcfb66a519cbe891d7e254
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/datasets/__init__.py
@@ -0,0 +1,10 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+from .mmdataset import *
+
+try:
+ from .fairseqmmdataset import *
+except ImportError:
+ pass
diff --git a/fairseq/examples/MMPT/mmpt/datasets/fairseqmmdataset.py b/fairseq/examples/MMPT/mmpt/datasets/fairseqmmdataset.py
new file mode 100644
index 0000000000000000000000000000000000000000..02c49141db69c44663bd438b947c268d06f8aa2b
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/datasets/fairseqmmdataset.py
@@ -0,0 +1,57 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+"""
+TODO (huxu): fairseq wrapper class for all dataset you defined: mostly MMDataset.
+"""
+
+from collections import OrderedDict
+
+from torch.utils.data import Dataset
+from torch.utils.data.dataloader import default_collate
+from fairseq.data import FairseqDataset, data_utils
+
+
+class FairseqMMDataset(FairseqDataset):
+ """
+ A wrapper class for MMDataset for fairseq.
+ """
+
+ def __init__(self, mmdataset):
+ if not isinstance(mmdataset, Dataset):
+ raise TypeError("mmdataset must be of type `torch.utils.data.dataset`.")
+ self.mmdataset = mmdataset
+
+ def set_epoch(self, epoch, **unused):
+ super().set_epoch(epoch)
+ self.epoch = epoch
+
+ def __getitem__(self, idx):
+ with data_utils.numpy_seed(43211, self.epoch, idx):
+ return self.mmdataset[idx]
+
+ def __len__(self):
+ return len(self.mmdataset)
+
+ def collater(self, samples):
+ if hasattr(self.mmdataset, "collator"):
+ return self.mmdataset.collator(samples)
+ if len(samples) == 0:
+ return {}
+ if isinstance(samples[0], dict):
+ batch = OrderedDict()
+ for key in samples[0]:
+ if samples[0][key] is not None:
+ batch[key] = default_collate([sample[key] for sample in samples])
+ return batch
+ else:
+ return default_collate(samples)
+
+ def size(self, index):
+ """dummy implementation: we don't use --max-tokens"""
+ return 1
+
+ def num_tokens(self, index):
+ """dummy implementation: we don't use --max-tokens"""
+ return 1
diff --git a/fairseq/examples/MMPT/mmpt/datasets/mmdataset.py b/fairseq/examples/MMPT/mmpt/datasets/mmdataset.py
new file mode 100644
index 0000000000000000000000000000000000000000..3d07283f917a430a8d9b1226c8fa6ab71450e8a9
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/datasets/mmdataset.py
@@ -0,0 +1,111 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+
+from collections import OrderedDict
+
+from torch.utils.data import Dataset
+from torch.utils.data.dataloader import default_collate
+
+from ..utils import set_seed
+
+
+class MMDataset(Dataset):
+ """
+ A generic multi-modal dataset.
+ Args:
+ `meta_processor`: a meta processor,
+ handling loading meta data and return video_id and text_id.
+ `video_processor`: a video processor,
+ handling e.g., decoding, loading .np files.
+ `text_processor`: a text processor,
+ handling e.g., tokenization.
+ `aligner`: combine the video and text feature
+ as one training example.
+ """
+
+ def __init__(
+ self,
+ meta_processor,
+ video_processor,
+ text_processor,
+ align_processor,
+ ):
+ self.split = meta_processor.split
+ self.meta_processor = meta_processor
+ self.video_processor = video_processor
+ self.text_processor = text_processor
+ self.align_processor = align_processor
+
+ def __len__(self):
+ return len(self.meta_processor)
+
+ def __getitem__(self, idx):
+ if self.split == "test":
+ set_seed(idx)
+ video_id, text_id = self.meta_processor[idx]
+ video_feature = self.video_processor(video_id)
+ text_feature = self.text_processor(text_id)
+ output = self.align_processor(video_id, video_feature, text_feature)
+ # TODO (huxu): the following is for debug purpose.
+ output.update({"idx": idx})
+ return output
+
+ def collater(self, samples):
+ """This collator is deprecated.
+ set self.collator = MMDataset.collater.
+ see collator in FairseqMMDataset.
+ """
+
+ if len(samples) == 0:
+ return {}
+ if isinstance(samples[0], dict):
+ batch = OrderedDict()
+ for key in samples[0]:
+ if samples[0][key] is not None:
+ batch[key] = default_collate(
+ [sample[key] for sample in samples])
+ # if torch.is_tensor(batch[key]):
+ # print(key, batch[key].size())
+ # else:
+ # print(key, len(batch[key]))
+ return batch
+ else:
+ return default_collate(samples)
+
+ def print_example(self, output):
+ print("[one example]", output["video_id"])
+ if (
+ hasattr(self.align_processor, "subsampling")
+ and self.align_processor.subsampling is not None
+ and self.align_processor.subsampling > 1
+ ):
+ for key in output:
+ if torch.is_tensor(output[key]):
+ output[key] = output[key][0]
+
+ # search tokenizer to translate ids back.
+ tokenizer = None
+ if hasattr(self.text_processor, "tokenizer"):
+ tokenizer = self.text_processor.tokenizer
+ elif hasattr(self.align_processor, "tokenizer"):
+ tokenizer = self.align_processor.tokenizer
+ if tokenizer is not None:
+ caps = output["caps"].tolist()
+ if isinstance(caps[0], list):
+ caps = caps[0]
+ print("caps", tokenizer.decode(caps))
+ print("caps", tokenizer.convert_ids_to_tokens(caps))
+
+ for key, value in output.items():
+ if torch.is_tensor(value):
+ if len(value.size()) >= 3: # attention_mask.
+ print(key, value.size())
+ print(key, "first", value[0, :, :])
+ print(key, "last", value[-1, :, :])
+ else:
+ print(key, value)
+ print("[end of one example]")
diff --git a/fairseq/examples/MMPT/mmpt/evaluators/__init__.py b/fairseq/examples/MMPT/mmpt/evaluators/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..2d06b9d7974db251143025124468f48cd230e89a
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/evaluators/__init__.py
@@ -0,0 +1,13 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+from .metric import *
+from .evaluator import *
+
+
+# experimental.
+try:
+ from .expmetric import *
+except ImportError:
+ pass
diff --git a/fairseq/examples/MMPT/mmpt/evaluators/evaluator.py b/fairseq/examples/MMPT/mmpt/evaluators/evaluator.py
new file mode 100644
index 0000000000000000000000000000000000000000..94d9c5ec9a6e84434dbced8d5647754c5a571570
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/evaluators/evaluator.py
@@ -0,0 +1,54 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+import os
+import glob
+import numpy as np
+
+from . import metric as metric_path
+from . import predictor as predictor_path
+
+
+class Evaluator(object):
+ """
+ perform evaluation on a single (downstream) task.
+ make this both offline and online.
+ TODO(huxu) saving evaluation results.
+ """
+
+ def __init__(self, config, eval_dataloader=None):
+ if config.metric is None:
+ raise ValueError("config.metric is", config.metric)
+ metric_cls = getattr(metric_path, config.metric)
+ self.metric = metric_cls(config)
+ if config.predictor is None:
+ raise ValueError("config.predictor is", config.predictor)
+ predictor_cls = getattr(predictor_path, config.predictor)
+ self.predictor = predictor_cls(config)
+ self.eval_dataloader = eval_dataloader
+
+ def __call__(self):
+ try:
+ print(self.predictor.pred_dir)
+ for pred_file in glob.glob(
+ self.predictor.pred_dir + "/*_merged.npy"):
+ outputs = np.load(pred_file)
+ results = self.metric.compute_metrics(outputs)
+ self.metric.print_computed_metrics(results)
+
+ outputs = np.load(os.path.join(
+ self.predictor.pred_dir, "merged.npy"))
+ results = self.metric.compute_metrics(outputs)
+ return {"results": results, "metric": self.metric}
+ except FileNotFoundError:
+ print("\n[missing]", self.predictor.pred_dir)
+ return {}
+
+ def evaluate(self, model, eval_dataloader=None, output_file="merged"):
+ if eval_dataloader is None:
+ eval_dataloader = self.eval_dataloader
+ outputs = self.predictor.predict_loop(
+ model, eval_dataloader, output_file)
+ results = self.metric.compute_metrics(**outputs)
+ return results
diff --git a/fairseq/examples/MMPT/mmpt/evaluators/metric.py b/fairseq/examples/MMPT/mmpt/evaluators/metric.py
new file mode 100644
index 0000000000000000000000000000000000000000..163724bb250cb1b7057b3fa4d75a9fafa9c181f5
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/evaluators/metric.py
@@ -0,0 +1,313 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import numpy as np
+import json
+
+
+class Metric(object):
+ def __init__(self, config, metric_names):
+ self.metric_names = metric_names
+
+ def best_metric(self, metric):
+ return metric[self.metric_names[0]]
+
+ def save_metrics(self, fn, metrics):
+ with open(fn, "w") as fw:
+ json.dump(fw, metrics)
+
+ def print_computed_metrics(self, metrics):
+ raise NotImplementedError
+
+
+class RetrievalMetric(Metric):
+ """
+ this is modified from `howto100m/metrics.py`.
+ History of changes:
+ refactor as a class.
+ add metric_key in __init__
+ """
+
+ def __init__(self, config, metric_names=["R1", "R5", "R10", "MR"]):
+ super().__init__(config, metric_names)
+ self.error = False # TODO(huxu): add to config to print error.
+
+ def compute_metrics(self, outputs, texts, **kwargs):
+ x = outputs
+ sx = np.sort(-x, axis=1)
+ d = np.diag(-x)
+ d = d[:, np.newaxis]
+ ind = sx - d
+ ind = np.where(ind == 0)
+ ind = ind[1]
+ metrics = {}
+ metrics["R1"] = float(np.sum(ind == 0)) / len(ind)
+ metrics["R5"] = float(np.sum(ind < 5)) / len(ind)
+ metrics["R10"] = float(np.sum(ind < 10)) / len(ind)
+ metrics["MR"] = np.median(ind) + 1
+
+ max_idx = np.argmax(outputs, axis=1)
+ if self.error:
+ # print top-20 errors.
+ error = []
+ for ex_idx in range(20):
+ error.append((texts[ex_idx], texts[max_idx[ex_idx]]))
+ metrics["error"] = error
+ return metrics
+
+ def print_computed_metrics(self, metrics):
+ r1 = metrics["R1"]
+ r5 = metrics["R5"]
+ r10 = metrics["R10"]
+ mr = metrics["MR"]
+ print(
+ "R@1: {:.4f} - R@5: {:.4f} - R@10: {:.4f} - Median R: {}".format(
+ r1, r5, r10, mr
+ )
+ )
+ if "error" in metrics:
+ print(metrics["error"])
+
+
+class DiDeMoMetric(Metric):
+ """
+ History of changes:
+ python 2.x to python 3.x.
+ merge utils.py into eval to save one file.
+ reference: https://github.com/LisaAnne/LocalizingMoments/blob/master/utils/eval.py
+ Code to evaluate your results on the DiDeMo dataset.
+ """
+ def __init__(self, config, metric_names=["rank1", "rank5", "miou"]):
+ super().__init__(config, metric_names)
+
+ def compute_metrics(self, outputs, targets, **kwargs):
+ assert len(outputs) == len(targets)
+ rank1, rank5, miou = self._eval_predictions(outputs, targets)
+ metrics = {
+ "rank1": rank1,
+ "rank5": rank5,
+ "miou": miou
+ }
+ return metrics
+
+ def print_computed_metrics(self, metrics):
+ rank1 = metrics["rank1"]
+ rank5 = metrics["rank5"]
+ miou = metrics["miou"]
+ # print("Average rank@1: %f" % rank1)
+ # print("Average rank@5: %f" % rank5)
+ # print("Average iou: %f" % miou)
+
+ print(
+ "Average rank@1: {:.4f} Average rank@5: {:.4f} Average iou: {:.4f}".format(
+ rank1, rank5, miou
+ )
+ )
+
+ def _iou(self, pred, gt):
+ intersection = max(0, min(pred[1], gt[1]) + 1 - max(pred[0], gt[0]))
+ union = max(pred[1], gt[1]) + 1 - min(pred[0], gt[0])
+ return float(intersection)/union
+
+ def _rank(self, pred, gt):
+ return pred.index(tuple(gt)) + 1
+
+ def _eval_predictions(self, segments, data):
+ '''
+ Inputs:
+ segments: For each item in the ground truth data, rank possible video segments given the description and video.
+ In DiDeMo, there are 21 posible moments extracted for each video so the list of video segments will be of length 21.
+ The first video segment should be the video segment that best corresponds to the text query.
+ There are 4180 sentence in the validation data, so when evaluating a model on the val dataset,
+ segments should be a list of lenght 4180, and each item in segments should be a list of length 21.
+ data: ground truth data
+ '''
+ average_ranks = []
+ average_iou = []
+ for s, d in zip(segments, data):
+ pred = s[0]
+ ious = [self._iou(pred, t) for t in d['times']]
+ average_iou.append(np.mean(np.sort(ious)[-3:]))
+ ranks = [self._rank(s, t) for t in d['times'] if tuple(t) in s] # if t in s] is added for s, e not in prediction.
+ average_ranks.append(np.mean(np.sort(ranks)[:3]))
+ rank1 = np.sum(np.array(average_ranks) <= 1)/float(len(average_ranks))
+ rank5 = np.sum(np.array(average_ranks) <= 5)/float(len(average_ranks))
+ miou = np.mean(average_iou)
+
+ # print("Average rank@1: %f" % rank1)
+ # print("Average rank@5: %f" % rank5)
+ # print("Average iou: %f" % miou)
+ return rank1, rank5, miou
+
+
+class NLGMetric(Metric):
+ def __init__(
+ self,
+ config,
+ metric_names=[
+ "Bleu_1", "Bleu_2", "Bleu_3", "Bleu_4",
+ "METEOR", "ROUGE_L", "CIDEr"
+ ]
+ ):
+ super().__init__(config, metric_names)
+ # please install NLGEval from `https://github.com/Maluuba/nlg-eval`
+ from nlgeval import NLGEval
+ self.nlg = NLGEval()
+
+ def compute_metrics(self, outputs, targets, **kwargs):
+ return self.nlg.compute_metrics(
+ hyp_list=outputs, ref_list=targets)
+
+ def print_computed_metrics(self, metrics):
+ Bleu_1 = metrics["Bleu_1"]
+ Bleu_2 = metrics["Bleu_2"]
+ Bleu_3 = metrics["Bleu_3"]
+ Bleu_4 = metrics["Bleu_4"]
+ METEOR = metrics["METEOR"]
+ ROUGE_L = metrics["ROUGE_L"]
+ CIDEr = metrics["CIDEr"]
+
+ print(
+ "Bleu_1: {:.4f} - Bleu_2: {:.4f} - Bleu_3: {:.4f} - Bleu_4: {:.4f} - METEOR: {:.4f} - ROUGE_L: {:.4f} - CIDEr: {:.4f}".format(
+ Bleu_1, Bleu_2, Bleu_3, Bleu_4, METEOR, ROUGE_L, CIDEr
+ )
+ )
+
+
+class QAMetric(Metric):
+ def __init__(
+ self,
+ config,
+ metric_names=["acc"]
+ ):
+ super().__init__(config, metric_names)
+
+ def compute_metrics(self, outputs, targets, **kwargs):
+ from sklearn.metrics import accuracy_score
+ return {"acc": accuracy_score(targets, outputs)}
+
+ def print_computed_metrics(self, metrics):
+ print("acc: {:.4f}".format(metrics["acc"]))
+
+
+class COINActionSegmentationMetric(Metric):
+ """
+ COIN dataset listed 3 repos for Action Segmentation.
+ Action Sets, NeuralNetwork-Viterbi, TCFPN-ISBA.
+ The first and second are the same.
+ https://github.com/alexanderrichard/action-sets/blob/master/eval.py
+
+ Future reference for the third:
+ `https://github.com/Zephyr-D/TCFPN-ISBA/blob/master/utils/metrics.py`
+ """
+ def __init__(self, config, metric_name=["frame_acc"]):
+ super().__init__(config, metric_name)
+
+ def compute_metrics(self, outputs, targets):
+ n_frames = 0
+ n_errors = 0
+ n_errors = sum(outputs != targets)
+ n_frames = len(targets)
+ return {"frame_acc": 1.0 - float(n_errors) / n_frames}
+
+ def print_computed_metrics(self, metrics):
+ fa = metrics["frame_acc"]
+ print("frame accuracy:", fa)
+
+
+class CrossTaskMetric(Metric):
+ def __init__(self, config, metric_names=["recall"]):
+ super().__init__(config, metric_names)
+
+ def compute_metrics(self, outputs, targets, **kwargs):
+ """refactored from line 166:
+ https://github.com/DmZhukov/CrossTask/blob/master/train.py"""
+
+ recalls = self._get_recalls(Y_true=targets, Y_pred=outputs)
+ results = {}
+ for task, rec in recalls.items():
+ results[str(task)] = rec
+
+ avg_recall = np.mean(list(recalls.values()))
+ results["recall"] = avg_recall
+ return results
+
+ def print_computed_metrics(self, metrics):
+ print('Recall: {0:0.3f}'.format(metrics["recall"]))
+ for task in metrics:
+ if task != "recall":
+ print('Task {0}. Recall = {1:0.3f}'.format(
+ task, metrics[task]))
+
+ def _get_recalls(self, Y_true, Y_pred):
+ """refactored from
+ https://github.com/DmZhukov/CrossTask/blob/master/train.py"""
+
+ step_match = {task: 0 for task in Y_true.keys()}
+ step_total = {task: 0 for task in Y_true.keys()}
+ for task, ys_true in Y_true.items():
+ ys_pred = Y_pred[task]
+ for vid in set(ys_pred.keys()).intersection(set(ys_true.keys())):
+ y_true = ys_true[vid]
+ y_pred = ys_pred[vid]
+ step_total[task] += (y_true.sum(axis=0) > 0).sum()
+ step_match[task] += (y_true*y_pred).sum()
+ recalls = {
+ task: step_match[task] / n for task, n in step_total.items()}
+ return recalls
+
+
+class ActionRecognitionMetric(Metric):
+ def __init__(
+ self,
+ config,
+ metric_names=["acc", "acc_splits", "r1_splits", "r5_splits", "r10_splits"]
+ ):
+ super().__init__(config, metric_names)
+
+ def compute_metrics(self, outputs, targets, splits, **kwargs):
+ all_video_embd = outputs
+ labels = targets
+ split1, split2, split3 = splits
+ accs = []
+ r1s = []
+ r5s = []
+ r10s = []
+ for split in range(3):
+ if split == 0:
+ s = split1
+ elif split == 1:
+ s = split2
+ else:
+ s = split3
+
+ X_pred = all_video_embd[np.where(s == 2)[0]]
+ label_test = labels[np.where(s == 2)[0]]
+ logits = X_pred
+ X_pred = np.argmax(X_pred, axis=1)
+ acc = np.sum(X_pred == label_test) / float(len(X_pred))
+ accs.append(acc)
+ # compute recall.
+ sorted_pred = (-logits).argsort(axis=-1)
+ label_test_sp = label_test.reshape(-1, 1)
+
+ r1 = np.mean((sorted_pred[:, :1] == label_test_sp).sum(axis=1), axis=0)
+ r5 = np.mean((sorted_pred[:, :5] == label_test_sp).sum(axis=1), axis=0)
+ r10 = np.mean((sorted_pred[:, :10] == label_test_sp).sum(axis=1), axis=0)
+ r1s.append(r1)
+ r5s.append(r5)
+ r10s.append(r10)
+
+ return {"acc": accs[0], "acc_splits": accs, "r1_splits": r1s, "r5_splits": r5s, "r10_splits": r10s}
+
+ def print_computed_metrics(self, metrics):
+ for split, acc in enumerate(metrics["acc_splits"]):
+ print("Top 1 accuracy on split {}: {}; r1 {}; r5 {}; r10 {}".format(
+ split + 1, acc,
+ metrics["r1_splits"][split],
+ metrics["r5_splits"][split],
+ metrics["r10_splits"][split],
+ )
+ )
diff --git a/fairseq/examples/MMPT/mmpt/evaluators/predictor.py b/fairseq/examples/MMPT/mmpt/evaluators/predictor.py
new file mode 100644
index 0000000000000000000000000000000000000000..2ffef6ab474a7a275d90784250d04222bd6dc70f
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/evaluators/predictor.py
@@ -0,0 +1,595 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+import os
+import random
+import json
+import numpy as np
+import torch
+import pickle
+import math
+
+from tqdm import tqdm
+
+
+class Predictor(object):
+ """this base class is used to save predictions to disk
+ (and being called by a evaluator later).
+ Predictor has minimum support of single gpu prediction.
+ """
+ def __init__(self, config):
+ self.pred_dir = None # on-the-fly eval does not save the results.
+ if hasattr(config, "eval") and config.eval is not None:
+ self.pred_dir = config.eval.save_path
+ os.makedirs(self.pred_dir, exist_ok=True)
+
+ def __call__(self, outputs):
+ """extract the prediction and save it."""
+ raise NotImplementedError
+
+ def predict_loop(self, model, eval_dataloader, output_file=None):
+ """on-the-fly prediction on a single gpu."""
+ self.full_scores = []
+ model.eval()
+ model = model.to(0)
+ with torch.no_grad():
+ for data in eval_dataloader:
+ data = self.to_ctx(data)
+ outputs = model(**data)
+ outputs.update(data)
+ self(outputs)
+ return self.finalize(output_file)
+
+ def finalize(self, output_file):
+ pass
+
+ def to_ctx(self, data, ctx=0, dtype=None):
+ if isinstance(data, dict):
+ for key in data:
+ if torch.is_tensor(data[key]):
+ if dtype is not None and data[key].dtype == torch.float32:
+ data[key] = data[key].to(dtype)
+ data[key] = data[key].to(ctx)
+ return data
+ else:
+ raise ValueError("non-dict type of batch is not supported yet.")
+
+
+class NLGPredictor(Predictor):
+ """Predicting Text from MMFusion models."""
+ """TODO: make a context."""
+ def __init__(self, config):
+ super().__init__(config)
+ from transformers import AutoTokenizer
+
+ self.tokenizer = AutoTokenizer.from_pretrained(
+ config.dataset.bert_name,
+ bos_token="[CLS]", eos_token="[SEP]")
+ self.bos_token_id = self.tokenizer.bos_token_id
+ self.eos_token_id = self.tokenizer.eos_token_id
+
+ def predict_loop(self, model, eval_dataloader, output_file=None):
+ """TODO: refactor base classes."""
+ ctx = 0
+ outputs = {"outputs": [], "targets": [[]]}
+ model.eval()
+ model = model.to(ctx)
+ with torch.no_grad():
+ for data in tqdm(eval_dataloader):
+ data = self.to_ctx(data, ctx)
+ self(data, model, outputs)
+ return self.finalize(outputs, output_file)
+
+ def __call__(self, data, model, outputs):
+ data.update({
+ "bos_token_id": self.bos_token_id,
+ "eos_token_id": self.eos_token_id
+ })
+
+ output = model.generate(**data)
+ assert len(output) == len(data["ref"])
+ for idx, _output in enumerate(output):
+ generated_text = self.tokenizer.decode(
+ _output, skip_special_tokens=True)
+ if generated_text == "":
+ generated_text = "none"
+ outputs["outputs"].append(generated_text)
+ outputs["targets"][0].append(data["ref"][idx])
+ if random.random() < 0.001:
+ print("_output", _output)
+ print("generated_text", generated_text)
+ print("ref", data["ref"][idx])
+
+ def finalize(self, outputs, output_file=None):
+ if output_file is not None:
+ with open(os.path.join(
+ self.pred_dir, output_file + ".json"), "w") as fw:
+ json.dump(outputs, fw, indent=4)
+ return outputs
+
+
+class RetrievalPredictor(Predictor):
+ """generated `pooled_video` and `pooled_text`."""
+ def __init__(self, config):
+ super().__init__(config)
+ from transformers import AutoTokenizer
+ self.tokenizer = AutoTokenizer.from_pretrained(
+ config.dataset.bert_name)
+
+ def predict_loop(
+ self,
+ model,
+ eval_dataloader,
+ output_file="retrieval.npy"
+ ):
+ """on-the-fly prediction on a single gpu."""
+ full_scores = []
+ texts = []
+ model.eval()
+ model = model.cuda()
+ with torch.no_grad():
+ for data in eval_dataloader:
+ # convert to dict.
+ if not isinstance(data, dict):
+ data = {
+ "caps": data[0],
+ "cmasks": data[1],
+ "vfeats": data[2],
+ "vmasks": data[3],
+ "video_id": data[4]
+ }
+ data = self.to_ctx(data)
+ outputs = model(**data)
+ outputs.update(data)
+ self(outputs, full_scores)
+ for _cap in data["caps"]:
+ texts.append(
+ self.tokenizer.decode(_cap, skip_special_tokens=True)
+ )
+
+ return self.finalize(full_scores, texts, output_file)
+
+ def __call__(self, sample, full_scores):
+ scores = self._get_pooled_outputs(sample)
+ self._append_scores(scores, full_scores)
+
+ def finalize(self, full_scores, texts, output_file=None):
+ outputs = self._aggregate_scores(full_scores)
+ if output_file is not None:
+ np.save(os.path.join(self.pred_dir, output_file + ".npy"), outputs)
+ return {"outputs": outputs, "texts": texts}
+
+ def _get_pooled_outputs(self, outputs):
+ if "pooled_video" in outputs:
+ return outputs["pooled_video"], outputs["pooled_text"]
+ else:
+ raise ValueError("unknown format of outputs.")
+
+ def _append_scores(self, scores, full_scores):
+ assert len(scores) == 2
+ if len(full_scores) == 0:
+ full_scores.append([])
+ full_scores.append([])
+ full_scores[0].append(scores[0].cpu().detach().numpy())
+ full_scores[1].append(scores[1].cpu().detach().numpy())
+
+ def _aggregate_scores(self, scores):
+ assert len(scores) == 2
+ video_hidden = np.concatenate(scores[0], axis=0)
+ text_hidden = np.concatenate(scores[1], axis=0)
+ # clear up.
+ self.full_scores = []
+ return np.matmul(text_hidden, video_hidden.T)
+
+
+class QAPredictor(Predictor):
+ """generated `pooled_video` and `pooled_text`."""
+ def __init__(self, config):
+ super().__init__(config)
+ """predictor maintains scores and aggregate them."""
+
+ def predict_loop(self, model, eval_dataloader, output_file="qa.npy"):
+ """on-the-fly prediction on a single gpu."""
+ self.full_scores = []
+ model.eval()
+ model = model.cuda()
+ with torch.no_grad():
+ for data in eval_dataloader:
+ # reshape ans and dup video 5 times.
+ v_len = data["vfeats"].size(1)
+ hidden_size = data["vfeats"].size(2)
+ data["vfeats"] = data["vfeats"].unsqueeze(1).repeat(1, 5, 1, 1).view(-1, v_len, hidden_size)
+ data["vmasks"] = data["vmasks"].unsqueeze(1).repeat(1, 5, 1).view(-1, v_len)
+
+ t_len = data["caps"].size(-1)
+ data["caps"] = data["caps"].view(-1, t_len)
+ data["cmasks"] = data["cmasks"].view(-1, t_len)
+
+ data = self.to_ctx(data)
+ outputs = model(**data)
+ outputs.update(data)
+ self(outputs)
+ return self.finalize(output_file)
+
+ def __call__(self, sample):
+ hidden_size = sample["pooled_video"].size(-1)
+ pooled_video = sample["pooled_video"].view(-1, 5, hidden_size)
+ pooled_text = sample["pooled_text"].view(-1, 5, hidden_size)
+ scores = torch.bmm(pooled_video, pooled_text.transpose(2, 1))
+ scores = scores.argmax(-1)
+ self._append_scores(scores[:, 0], sample["answers"], self.full_scores)
+
+ def finalize(self, output_file=None):
+ outputs, targets = self._aggregate_scores(self.full_scores)
+ if output_file is not None:
+ np.save(os.path.join(self.pred_dir, output_file + ".npy"), outputs)
+ return {"outputs": outputs, "targets": targets}
+
+ def _append_scores(self, scores, answers, full_scores):
+ if len(full_scores) == 0:
+ full_scores.append([])
+ full_scores.append([])
+ full_scores[0].append(scores.cpu().detach().numpy())
+ full_scores[1].append(answers.cpu().detach().numpy())
+
+ def _aggregate_scores(self, scores):
+ assert len(scores) == 2
+ outputs = np.concatenate(scores[0], axis=0)
+ targets = np.concatenate(scores[1], axis=0)
+ # clear up.
+ self.full_scores = []
+ return outputs, targets
+
+
+class CrossTaskPredictor(Predictor):
+ """
+ CrossTaskPredictor needs to compute the average of logits
+ for overlapped sliding-window.
+ """
+ def __init__(self, config):
+ super().__init__(config)
+ self.lsm = torch.nn.LogSoftmax(dim=1)
+ self.max_video_len = config.dataset.max_video_len
+ self.sliding_window = config.dataset.sliding_window
+ self.sliding_window_size = config.dataset.sliding_window_size
+ self.annotation_path = config.dataset.annotation_path
+
+ def predict_loop(self, model, eval_dataloader, output_file="result.pkl"):
+ """refactored from line 144:
+ https://github.com/DmZhukov/CrossTask/blob/master/train.py
+ """
+ ctx = 0
+ model.eval()
+ model = model.to(ctx)
+ # this is not a loss but just compute neg_log_prob.
+ Y_pred = {}
+ Y_true = {}
+ with torch.no_grad():
+ for batch in eval_dataloader:
+ self(batch, model, Y_pred, Y_true)
+ return self.finalize(Y_pred, Y_true, output_file)
+
+ def __call__(self, sample, model, Y_pred, Y_true):
+ # please install dp from `https://github.com/DmZhukov/CrossTask`
+ from dp import dp
+ vid, task = sample['video_id'][0], sample['task'][0]
+ sample = self.to_ctx(sample)
+ # compute the average logits over sliding windows.
+ output = model(**sample)
+ batch_logits = output["logits"].cpu()
+
+ video_len = sample["video_len"][0]
+
+ # the following version is slow.
+ logits = torch.zeros((video_len, batch_logits.size(1)))
+ logits_counts = torch.zeros((video_len, 1), dtype=torch.long)
+ # use the same loop as aligner to recover.
+ batch_logit_idx = 0
+ for window_start in range(0, video_len, self.sliding_window):
+ video_end = min(video_len - window_start, self.sliding_window_size)
+ logits[window_start: window_start + video_end] += batch_logits[
+ batch_logit_idx: batch_logit_idx + video_end]
+ batch_logit_idx += video_end
+ logits_counts[window_start: window_start + video_end] += torch.ones((video_end, 1), dtype=torch.long)
+
+ if (video_len - window_start) <= self.sliding_window_size:
+ break
+
+ logits /= logits_counts
+ assert logits.size() == (video_len, batch_logits.size(1)), "{}, {}".format(logits.size(), video_len)
+
+ O = self.lsm(logits)
+ y = np.zeros(O.size(), dtype=np.float32)
+ dp(y, -O.detach().cpu().numpy())
+ if task not in Y_pred:
+ Y_pred[task] = {}
+ Y_pred[task][vid] = y
+ annot_path = os.path.join(
+ self.annotation_path, task+'_'+vid+'.csv')
+ if os.path.exists(annot_path):
+ if task not in Y_true:
+ Y_true[task] = {}
+ Y_true[task][vid] = self._read_assignment(
+ *y.shape, annot_path)
+
+ def finalize(self, Y_pred, Y_true, output_file=None):
+ if output_file is not None:
+ with open(
+ os.path.join(self.pred_dir, output_file + ".pkl"),
+ "wb") as fw:
+ pickle.dump(
+ {"Y_pred": Y_pred, "Y_true": Y_true}, fw,
+ protocol=pickle.HIGHEST_PROTOCOL)
+ return {"outputs": Y_pred, "targets": Y_true}
+
+ def _read_assignment(self, T, K, path):
+ """
+ refactored from https://github.com/DmZhukov/CrossTask/blob/master/data.py
+ Howto interpret contraints on loss that is going to be minimized:
+ lambd is a big number;
+ self.lambd * C is a big number for all valid position (csv stores invalids)
+
+ def forward(self, O, Y, C):
+ return (Y*(self.lambd * C - self.lsm(O))).mean(dim=0).sum()
+
+ This will load the csv file and fill-in the step col from start to end rows.
+ """
+
+ Y = np.zeros([T, K], dtype=np.uint8)
+ with open(path, 'r') as f:
+ for line in f:
+ step, start, end = line.strip().split(',')
+ start = int(math.floor(float(start)))
+ end = int(math.ceil(float(end)))
+ step = int(step) - 1
+ Y[start:end, step] = 1
+ return Y
+
+
+class COINPredictor(Predictor):
+ """
+ COINPredictor is similar to CrossTask on sliding windows.
+ """
+ def __init__(self, config):
+ super().__init__(config)
+ self.max_video_len = config.dataset.max_video_len
+ self.sliding_window = config.dataset.sliding_window
+ self.sliding_window_size = config.dataset.sliding_window_size
+
+ def predict_loop(self, model, eval_dataloader, output_file="result.pkl"):
+ """refactored from line 144:
+ https://github.com/DmZhukov/CrossTask/blob/master/train.py
+ """
+ ctx = 0
+ model.eval()
+ model = model.to(ctx)
+ # this is not a loss but just compute neg_log_prob.
+ Y_pred = []
+ Y_true = []
+ with torch.no_grad():
+ for batch in eval_dataloader:
+ self(batch, model, Y_pred, Y_true)
+ return self.finalize(Y_pred, Y_true, output_file)
+
+ def __call__(self, sample, model, Y_pred, Y_true):
+ sample = self.to_ctx(sample)
+ # compute the average logits over sliding windows.
+ output = model(**sample)
+ logits = self._merge_windows(sample, output)
+ Y_pred.append(logits.argmax(dim=1))
+ Y_true.append(sample["video_targets"].squeeze(0).cpu())
+
+ def _merge_windows(self, sample, output):
+ targets = sample["targets"].reshape(-1).cpu()
+ valid_mask = targets != -100
+ targets = targets[valid_mask]
+ batch_logits = output["logits"].cpu()
+ batch_logits = batch_logits.reshape(-1, batch_logits.size(-1))
+ batch_logits = batch_logits[valid_mask]
+
+ video_len = sample["video_len"][0]
+
+ # the following version is slow.
+ logits = torch.zeros((video_len, batch_logits.size(1)))
+ logits_counts = torch.zeros((video_len, 1), dtype=torch.long)
+ # use the same loop as aligner to recover.
+ batch_logit_idx = 0
+ for window_start in range(0, video_len, self.sliding_window):
+ video_end = min(video_len - window_start, self.sliding_window_size)
+ logits[window_start: window_start + video_end] += batch_logits[
+ batch_logit_idx: batch_logit_idx + video_end]
+ batch_logit_idx += video_end
+ logits_counts[window_start: window_start + video_end] += torch.ones((video_end, 1), dtype=torch.long)
+ if (video_len - window_start) <= self.sliding_window_size:
+ break
+ logits /= logits_counts
+ assert logits.size() == (video_len, batch_logits.size(1)), "{}, {}".format(logits.size(), video_len)
+ return logits
+
+ def finalize(self, Y_pred, Y_true, output_file=None):
+ Y_pred = torch.cat(Y_pred, dim=0).numpy()
+ Y_true = torch.cat(Y_true, dim=0).numpy()
+ assert len(Y_pred) == len(Y_true)
+
+ error_mask = Y_pred != Y_true
+ print("sample error", Y_pred[error_mask][:10], Y_true[error_mask][:10])
+ print("sample error", Y_pred[error_mask][10:20], Y_true[error_mask][10:20])
+
+ if output_file is not None:
+ with open(
+ os.path.join(self.pred_dir, output_file + ".pkl"),
+ "wb") as fw:
+ pickle.dump(
+ {"Y_pred": Y_pred, "Y_true": Y_true}, fw,
+ protocol=pickle.HIGHEST_PROTOCOL)
+ return {"outputs": Y_pred, "targets": Y_true}
+
+
+class COINZSPredictor(COINPredictor):
+ """
+ COINZSPredictor for COIN zero-shot prediction.
+ """
+
+ def __init__(self, config):
+ super().__init__(config)
+ self.dataset_config = config.dataset
+
+ def predict_loop(self, model, eval_dataloader, output_file="result.pkl"):
+ """refactored from line 144:
+ https://github.com/DmZhukov/CrossTask/blob/master/train.py
+ """
+ ctx = 0
+ model.eval()
+ model = model.to(ctx)
+
+ with torch.no_grad():
+ outputs = eval_dataloader.dataset.meta_processor.meta_text_labels(
+ self.dataset_config)
+ outputs = self.to_ctx(outputs, ctx)
+ label_hidden_states = model.forward_text(**outputs).cpu()
+ label_sim = label_hidden_states @ label_hidden_states.t()
+ num_labels = label_sim.size(0)
+ eye_mask = ~torch.eye(num_labels, dtype=torch.bool)
+ label_sim = label_sim.masked_select(eye_mask).view(num_labels, num_labels - 1)
+ lbd = label_sim.max()
+
+ # this is not a loss but just compute neg_log_prob.
+ Y_pred = []
+ Y_true = []
+ with torch.no_grad():
+ for batch in eval_dataloader:
+ self(batch, label_hidden_states, model, lbd, Y_pred, Y_true)
+ return self.finalize(Y_pred, Y_true, output_file)
+
+ def reshape_subsample(self, sample):
+ for key in sample:
+ if torch.is_tensor(sample[key]):
+ sample[key] = self.flat_subsample(sample[key])
+ return sample
+
+ def flat_subsample(self, tensor):
+ if len(tensor.size()) > 1 and tensor.size(0) == 1:
+ tensor = tensor.squeeze(0)
+ return tensor
+
+ def __call__(self, sample, label_hidden_states, model, lbd, Y_pred, Y_true):
+ sample = self.reshape_subsample(sample)
+ sample = self.to_ctx(sample)
+ # compute the average logits over sliding windows.
+ sample["output_hidden_states"] = True
+ video_outputs = model.forward_video(**sample).cpu()
+ output = {"logits": video_outputs[:, 1:sample["vmasks"].size(1)+1] @ label_hidden_states.t()}
+ logits = self._merge_windows(sample, output)
+ # logic of zero-shot for sequence labeling.
+ logits_argmax = logits.argmax(dim=1) + 1 # 0 is "O" label.
+ logits_max = logits.max(dim=1)[0]
+
+ pred = torch.zeros_like(logits_argmax)
+ label_select = logits_max > lbd # 73 or 74
+ pred[label_select] = logits_argmax[label_select]
+
+ Y_pred.append(pred)
+ Y_true.append(sample["video_targets"].squeeze(0).cpu())
+
+ def finalize(self, Y_pred, Y_true, output_file=None):
+ Y_pred = torch.cat(Y_pred, dim=0).numpy()
+ Y_true = torch.cat(Y_true, dim=0).numpy()
+ assert len(Y_pred) == len(Y_true)
+
+ error_mask = Y_pred != Y_true
+ print("sample error", Y_pred[error_mask][:10], Y_true[error_mask][:10])
+ print("sample error", Y_pred[error_mask][10:20], Y_true[error_mask][10:20])
+
+ if output_file is not None:
+ with open(
+ os.path.join(self.pred_dir, output_file + ".pkl"),
+ "wb") as fw:
+ pickle.dump(
+ {"Y_pred": Y_pred, "Y_true": Y_true}, fw,
+ protocol=pickle.HIGHEST_PROTOCOL)
+ return {"outputs": Y_pred, "targets": Y_true}
+
+
+class DiDeMoPredictor(Predictor):
+ """reference: https://github.com/LisaAnne/LocalizingMoments/blob/master/utils/eval.py
+ https://github.com/LisaAnne/LocalizingMoments/blob/master/utils/data_processing.py
+ """
+ def __init__(self, config):
+ super().__init__(config)
+ # load targets.
+ with open(config.dataset.test_path) as data_file:
+ self.test_data = json.load(data_file)
+
+ def predict_loop(self, model, eval_dataloader, output_file="didemo.npy"):
+ """
+ TODO: two solutions here.
+ """
+ import itertools
+ # 21 chunks.
+ self.possible_segments = [(0,0), (1,1), (2,2), (3,3), (4,4), (5,5)]
+ for i in itertools.combinations(range(6), 2):
+ self.possible_segments.append(i)
+ # pick segments from a video.
+
+ """on-the-fly prediction on a single gpu."""
+ self.full_scores = []
+ model.eval()
+ model = model.cuda()
+ with torch.no_grad():
+ for data in eval_dataloader:
+ # TODO special forwarding logic here.
+ data = self.to_ctx(data)
+ data["output_hidden_states"] = True
+ hidden_video = model.forward_video(**data)
+ data["output_hidden_states"] = False
+ pooled_text = model.forward_text(**data)
+ outputs = {
+ "hidden_video": hidden_video,
+ "pooled_text": pooled_text
+ }
+ outputs.update(data)
+ self(outputs)
+ return self.finalize(output_file)
+
+ def __call__(self, sample):
+ # TODO: make an index select from self.possible_segments.
+ hidden_video = sample["hidden_video"]
+ pooled_text = sample["pooled_text"]
+ vmasks = sample["vmasks"]
+ # probably maintain valid results here.
+
+ hidden_video = hidden_video[:, 1:-1, :]
+ # probably maintain valid results here.
+ pooled_video = []
+ for s, e in self.possible_segments:
+ pooled_video.append(
+ torch.mean(
+ hidden_video[:, int(s*5):int((e+1)*5), :],
+ dim=1, keepdim=True)
+ )
+ pooled_video = torch.cat(pooled_video, dim=1)
+ scores = torch.bmm(
+ pooled_video, pooled_text.unsqueeze(-1)).squeeze(-1).cpu()
+
+ ranks = scores.argsort(dim=-1, descending=True)
+
+ for batch_idx, rank in enumerate(ranks):
+ rank_of_moment = []
+ for m_idx, moment in enumerate(rank):
+ s, e = self.possible_segments[moment.item()]
+ if torch.any(
+ vmasks[batch_idx, int(s*5):int((e+1)*5)]
+ ):
+ rank_of_moment.append((s, e))
+ self.full_scores.append(rank_of_moment)
+
+ def finalize(self, output_file=None):
+ outputs = self._aggregate_scores(self.full_scores)
+ if output_file is not None:
+ np.save(os.path.join(self.pred_dir, output_file + ".npy"), outputs)
+ return {"outputs": outputs, "targets": self.test_data}
+
+ def _aggregate_scores(self, scores):
+ self.full_scores = []
+ return scores
diff --git a/fairseq/examples/MMPT/mmpt/losses/__init__.py b/fairseq/examples/MMPT/mmpt/losses/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..8dc32c96d2d8aed25c59e69e9d9a2ff24a9a2a47
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/losses/__init__.py
@@ -0,0 +1,16 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+from .loss import *
+from .nce import *
+
+try:
+ from .fairseqmmloss import *
+except ImportError:
+ pass
+
+try:
+ from .expnce import *
+except ImportError:
+ pass
diff --git a/fairseq/examples/MMPT/mmpt/losses/fairseqmmloss.py b/fairseq/examples/MMPT/mmpt/losses/fairseqmmloss.py
new file mode 100644
index 0000000000000000000000000000000000000000..a95e5ecf45d90098c1719487bb9a11c36be7c507
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/losses/fairseqmmloss.py
@@ -0,0 +1,63 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""
+TODO (huxu): a general fairseq criterion for all your pre-defined losses.
+"""
+
+from fairseq.criterions import FairseqCriterion, register_criterion
+from fairseq.logging import metrics
+
+
+@register_criterion("mmloss")
+class MMCriterion(FairseqCriterion):
+ def __init__(self, task):
+ super().__init__(task)
+ # TODO (huxu): wrap forward call of loss_fn and eval_fn into task.
+ self.mmtask = task.mmtask
+
+ def forward(self, model, sample):
+ """Compute the loss for the given sample.
+ Returns a tuple with three elements:
+ 1) the loss
+ 2) the sample size, which is used as the denominator for the gradient
+ 3) logging outputs to display while training
+ """
+ outputs = self.mmtask(model, sample)
+
+ loss, loss_scalar, max_len, batch_size, sample_size = (
+ outputs["loss"],
+ outputs["loss_scalar"],
+ outputs["max_len"],
+ outputs["batch_size"],
+ outputs["sample_size"],
+ )
+
+ logging_output = {
+ "loss": loss_scalar,
+ "ntokens": max_len * batch_size, # dummy report.
+ "nsentences": batch_size, # dummy report.
+ "sample_size": sample_size,
+ }
+
+ return loss, 1, logging_output
+
+ @staticmethod
+ def reduce_metrics(logging_outputs) -> None:
+ """Aggregate logging outputs from data parallel training."""
+ """since we use NCE, our actual batch_size is 1 per GPU.
+ Then we take the mean of each worker."""
+ loss_sum = sum(log.get("loss", 0.0) for log in logging_outputs)
+ sample_size = sum(log.get("sample_size", 0) for log in logging_outputs)
+ metrics.log_scalar("loss", loss_sum / sample_size, round=3)
+
+ @staticmethod
+ def logging_outputs_can_be_summed() -> bool:
+ """
+ Whether the logging outputs returned by `forward` can be summed
+ across workers prior to calling `reduce_metrics`. Setting this
+ to True will improves distributed training speed.
+ """
+ return True
diff --git a/fairseq/examples/MMPT/mmpt/losses/loss.py b/fairseq/examples/MMPT/mmpt/losses/loss.py
new file mode 100644
index 0000000000000000000000000000000000000000..99c05d067edac220f9e53080f09f0b40d7dc1e8d
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/losses/loss.py
@@ -0,0 +1,87 @@
+# Copyright (c) Facebook, Inc. All Rights Reserved
+
+import torch
+
+from torch import nn
+
+
+class Loss(object):
+ def __call__(self, *args, **kwargs):
+ raise NotImplementedError
+
+
+# Dummy Loss for testing.
+class DummyLoss(Loss):
+ def __init__(self):
+ self.loss = nn.CrossEntropyLoss()
+
+ def __call__(self, logits, targets, **kwargs):
+ return self.loss(logits, targets)
+
+
+class DummyK400Loss(Loss):
+ """dummy k400 loss for MViT."""
+ def __init__(self):
+ self.loss = nn.CrossEntropyLoss()
+
+ def __call__(self, logits, targets, **kwargs):
+ return self.loss(
+ logits, torch.randint(0, 400, (logits.size(0),), device=logits.device))
+
+
+class CrossEntropy(Loss):
+ def __init__(self):
+ self.loss = nn.CrossEntropyLoss()
+
+ def __call__(self, logits, targets, **kwargs):
+ return self.loss(logits.reshape(-1, logits.size(-1)), targets.reshape(-1))
+
+
+class ArgmaxCrossEntropy(Loss):
+ def __init__(self):
+ self.loss = nn.CrossEntropyLoss()
+
+ def __call__(self, logits, targets, **kwargs):
+ return self.loss(logits, targets.argmax(dim=1))
+
+
+class BCE(Loss):
+ def __init__(self):
+ self.loss = nn.BCEWithLogitsLoss()
+
+ def __call__(self, logits, targets, **kwargs):
+ targets = targets.squeeze(0)
+ return self.loss(logits, targets)
+
+
+class NLGLoss(Loss):
+ def __init__(self):
+ self.loss = nn.CrossEntropyLoss()
+
+ def __call__(self, logits, text_label, **kwargs):
+ targets = text_label[text_label != -100]
+ return self.loss(logits, targets)
+
+
+class MSE(Loss):
+ def __init__(self):
+ self.loss = nn.MSELoss()
+
+ def __call__(self, logits, targets, **kwargs):
+ return self.loss(logits, targets)
+
+
+class L1(Loss):
+ def __init__(self):
+ self.loss = nn.L1Loss()
+
+ def __call__(self, logits, targets, **kwargs):
+ return self.loss(logits, targets)
+
+
+class SmoothL1(Loss):
+ def __init__(self):
+ self.loss = nn.SmoothL1Loss()
+
+ def __call__(self, logits, targets, **kwargs):
+ return self.loss(logits, targets)
diff --git a/fairseq/examples/MMPT/mmpt/losses/nce.py b/fairseq/examples/MMPT/mmpt/losses/nce.py
new file mode 100644
index 0000000000000000000000000000000000000000..ed7be8d372e371bb0e0d6166f76e01d3466d2306
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/losses/nce.py
@@ -0,0 +1,156 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""
+softmax-based NCE loss, used by this project.
+"""
+
+import torch
+
+from torch import nn
+
+from .loss import Loss
+
+
+class NCE(Loss):
+ def __init__(self):
+ # TODO (huxu): define temperature.
+ self.loss = nn.CrossEntropyLoss()
+
+ def __call__(self, align_scores, **kargs):
+ # note: we reuse the same shape as cls head in BERT (batch_size, 2)
+ # but NCE only needs one logits.
+ # (so we drop all weights in the second neg logits.)
+ align_scores = align_scores[:, :1]
+ # duplicate negative examples
+ batch_size = align_scores.size(0) // 2
+ pos_scores = align_scores[:batch_size]
+ neg_scores = align_scores[batch_size:].view(1, batch_size).repeat(
+ batch_size, 1)
+ scores = torch.cat([pos_scores, neg_scores], dim=1)
+ return self.loss(
+ scores,
+ torch.zeros(
+ (batch_size,),
+ dtype=torch.long,
+ device=align_scores.device),
+ )
+
+
+class T2VContraLoss(Loss):
+ """NCE for MM joint space, on softmax text2video matrix.
+ """
+ def __init__(self):
+ # TODO (huxu): define temperature.
+ self.loss = nn.CrossEntropyLoss()
+
+ def __call__(self, pooled_video, pooled_text, **kargs):
+ batch_size = pooled_video.size(0)
+ logits = torch.mm(pooled_text, pooled_video.transpose(1, 0))
+ targets = torch.arange(
+ batch_size,
+ dtype=torch.long,
+ device=pooled_video.device)
+ return self.loss(logits, targets)
+
+
+class V2TContraLoss(Loss):
+ """NCE for MM joint space, with softmax on video2text matrix."""
+
+ def __init__(self):
+ # TODO (huxu): define temperature.
+ self.loss = nn.CrossEntropyLoss()
+
+ def __call__(self, pooled_video, pooled_text, **kargs):
+ batch_size = pooled_video.size(0)
+ logits = torch.mm(pooled_video, pooled_text.transpose(1, 0))
+ targets = torch.arange(
+ batch_size,
+ dtype=torch.long,
+ device=pooled_video.device)
+ return self.loss(logits, targets)
+
+
+class MMContraLoss(Loss):
+ def __init__(self):
+ self.loss = nn.CrossEntropyLoss()
+
+ def __call__(self, pooled_video, pooled_text, **kwargs):
+ logits_per_video = pooled_video @ pooled_text.t()
+ logits_per_text = pooled_text @ pooled_video.t()
+
+ targets = torch.arange(
+ pooled_video.size(0),
+ dtype=torch.long,
+ device=pooled_video.device)
+ loss_video = self.loss(logits_per_video, targets)
+ loss_text = self.loss(logits_per_text, targets)
+ return loss_video + loss_text
+
+
+class MTM(Loss):
+ """Combination of MFM and MLM."""
+
+ def __init__(self):
+ self.loss = nn.CrossEntropyLoss()
+
+ def __call__(
+ self,
+ video_logits,
+ text_logits,
+ video_label,
+ text_label,
+ **kwargs
+ ):
+ text_logits = torch.cat([
+ text_logits,
+ torch.zeros(
+ (text_logits.size(0), 1), device=text_logits.device)
+ ], dim=1)
+ vt_logits = torch.cat([video_logits, text_logits], dim=0)
+ # loss for video.
+ video_label = torch.zeros(
+ (video_logits.size(0),),
+ dtype=torch.long,
+ device=video_logits.device
+ )
+
+ # loss for text.
+ text_label = text_label.reshape(-1)
+ labels_mask = text_label != -100
+ selected_text_label = text_label[labels_mask]
+
+ vt_label = torch.cat([video_label, selected_text_label], dim=0)
+ return self.loss(vt_logits, vt_label)
+
+
+class MFMMLM(Loss):
+ """Combination of MFM and MLM."""
+
+ def __init__(self):
+ self.loss = nn.CrossEntropyLoss()
+
+ def __call__(
+ self,
+ video_logits,
+ text_logits,
+ video_label,
+ text_label,
+ **kwargs
+ ):
+ # loss for video.
+ video_label = torch.zeros(
+ (video_logits.size(0),),
+ dtype=torch.long,
+ device=video_logits.device
+ )
+ masked_frame_loss = self.loss(video_logits, video_label)
+
+ # loss for text.
+ text_label = text_label.reshape(-1)
+ labels_mask = text_label != -100
+ selected_text_label = text_label[labels_mask]
+ masked_lm_loss = self.loss(text_logits, selected_text_label)
+ return masked_frame_loss + masked_lm_loss
diff --git a/fairseq/examples/MMPT/mmpt/models/__init__.py b/fairseq/examples/MMPT/mmpt/models/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..825250cd007f5e072b6c9d2376445b955a7aa71e
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/models/__init__.py
@@ -0,0 +1,17 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+from .mmfusion import *
+from .transformermodel import *
+from .mmfusionnlg import *
+
+try:
+ from .fairseqmmmodel import *
+except ImportError:
+ pass
+
+try:
+ from .expmmfusion import *
+except ImportError:
+ pass
diff --git a/fairseq/examples/MMPT/mmpt/models/fairseqmmmodel.py b/fairseq/examples/MMPT/mmpt/models/fairseqmmmodel.py
new file mode 100644
index 0000000000000000000000000000000000000000..b7dd643693dee8cfc20ca77d6cea798d07eaf15a
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/models/fairseqmmmodel.py
@@ -0,0 +1,51 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+from fairseq.models import (
+ BaseFairseqModel,
+ register_model,
+ register_model_architecture
+)
+
+
+@register_model("mmmodel")
+class FairseqMMModel(BaseFairseqModel):
+ """a fairseq wrapper of model built by `task`."""
+
+ @classmethod
+ def build_model(cls, args, task):
+ return FairseqMMModel(task.mmtask.model)
+
+ def __init__(self, mmmodel):
+ super().__init__()
+ self.mmmodel = mmmodel
+
+ def forward(self, *args, **kwargs):
+ return self.mmmodel(*args, **kwargs)
+
+ def upgrade_state_dict_named(self, state_dict, name):
+
+ super().upgrade_state_dict_named(state_dict, name)
+
+ keys_to_delete = []
+
+ for key in state_dict:
+ if key not in self.state_dict():
+ keys_to_delete.append(key)
+ for key in keys_to_delete:
+ print("[INFO]", key, "not used anymore.")
+ del state_dict[key]
+
+ # copy any newly defined parameters.
+ for key in self.state_dict():
+ if key not in state_dict:
+ print("[INFO] adding", key)
+ state_dict[key] = self.state_dict()[key]
+
+
+# a dummy arch, we config the model.
+@register_model_architecture("mmmodel", "mmarch")
+def mmarch(args):
+ pass
diff --git a/fairseq/examples/MMPT/mmpt/models/mmfusion.py b/fairseq/examples/MMPT/mmpt/models/mmfusion.py
new file mode 100644
index 0000000000000000000000000000000000000000..2509e26b67b467c3b18c76630881e66cf334a350
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/models/mmfusion.py
@@ -0,0 +1,926 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# Copyright (c) Facebook, Inc. All Rights Reserved
+
+
+import torch
+
+from torch import nn
+
+try:
+ from transformers import AutoConfig, AutoTokenizer
+except ImportError:
+ pass
+
+from . import transformermodel
+
+
+class MMPTModel(nn.Module):
+ """An e2e wrapper of inference model.
+ """
+ @classmethod
+ def from_pretrained(cls, config, checkpoint="checkpoint_best.pt"):
+ import os
+ from ..utils import recursive_config
+ from ..tasks import Task
+ config = recursive_config(config)
+ mmtask = Task.config_task(config)
+ checkpoint_path = os.path.join(config.eval.save_path, checkpoint)
+ mmtask.build_model(checkpoint=checkpoint_path)
+ # TODO(huxu): make the video encoder configurable.
+ from ..processors.models.s3dg import S3D
+ video_encoder = S3D('pretrained_models/s3d_dict.npy', 512)
+ video_encoder.load_state_dict(
+ torch.load('pretrained_models/s3d_howto100m.pth'))
+ from transformers import AutoTokenizer
+ tokenizer = AutoTokenizer.from_pretrained(
+ config.dataset.bert_name, use_fast=config.dataset.use_fast
+ )
+ from ..processors import Aligner
+ aligner = Aligner(config.dataset)
+ return (
+ MMPTModel(config, mmtask.model, video_encoder),
+ tokenizer,
+ aligner
+ )
+
+ def __init__(self, config, model, video_encoder, **kwargs):
+ super().__init__()
+ self.max_video_len = config.dataset.max_video_len
+ self.video_encoder = video_encoder
+ self.model = model
+
+ def forward(self, video_frames, caps, cmasks, return_score=False):
+ bsz = video_frames.size(0)
+ assert bsz == 1, "only bsz=1 is supported now."
+ seq_len = video_frames.size(1)
+ video_frames = video_frames.view(-1, *video_frames.size()[2:])
+ vfeats = self.video_encoder(video_frames.permute(0, 4, 1, 2, 3))
+ vfeats = vfeats['video_embedding']
+ vfeats = vfeats.view(bsz, seq_len, vfeats.size(-1))
+ padding = torch.zeros(
+ bsz, self.max_video_len - seq_len, vfeats.size(-1))
+ vfeats = torch.cat([vfeats, padding], dim=1)
+ vmasks = torch.cat([
+ torch.ones((bsz, seq_len), dtype=torch.bool),
+ torch.zeros((bsz, self.max_video_len - seq_len), dtype=torch.bool)
+ ],
+ dim=1
+ )
+ output = self.model(caps, cmasks, vfeats, vmasks)
+ if return_score:
+ output = {"score": torch.bmm(
+ output["pooled_video"][:, None, :],
+ output["pooled_text"][:, :, None]
+ ).squeeze(-1).squeeze(-1)}
+ return output
+
+
+class MMFusion(nn.Module):
+ """a MMPT wrapper class for MMBert style models.
+ TODO: move isolated mask to a subclass.
+ """
+ def __init__(self, config, **kwargs):
+ super().__init__()
+ transformer_config = AutoConfig.from_pretrained(
+ config.dataset.bert_name)
+ self.hidden_size = transformer_config.hidden_size
+ self.is_train = False
+ if config.dataset.train_path is not None:
+ self.is_train = True
+ # 0 means no iso; 1-12 means iso up to that layer.
+ self.num_hidden_layers = transformer_config.num_hidden_layers
+ self.last_iso_layer = 0
+ if config.dataset.num_iso_layer is not None:
+ self.last_iso_layer = config.dataset.num_iso_layer - 1 + 1
+
+ if config.model.mm_encoder_cls is not None:
+ mm_encoder_cls = getattr(transformermodel, config.model.mm_encoder_cls)
+ model_config = AutoConfig.from_pretrained(config.dataset.bert_name)
+ model_config.max_video_len = config.dataset.max_video_len
+ # TODO: a general way to add parameter for a model.
+ model_config.use_seg_emb = config.model.use_seg_emb
+ self.mm_encoder = mm_encoder_cls.from_pretrained(
+ config.dataset.bert_name, config=model_config)
+ elif config.model.video_encoder_cls is not None\
+ and config.model.text_encoder_cls is not None:
+ video_encoder_cls = getattr(transformermodel, config.model.video_encoder_cls)
+ model_config = AutoConfig.from_pretrained(config.dataset.bert_name)
+ model_config.max_video_len = config.dataset.max_video_len
+ # TODO: make each model a set of config class.
+ if hasattr(model_config, "num_layers"):
+ model_config.num_layers = config.model.num_hidden_video_layers
+ else:
+ model_config.num_hidden_layers = config.model.num_hidden_video_layers
+ self.video_encoder = video_encoder_cls.from_pretrained(
+ config.dataset.bert_name, config=model_config)
+ # exact same NLP model from Huggingface.
+ text_encoder_cls = getattr(transformermodel, config.model.text_encoder_cls)
+ self.text_encoder = text_encoder_cls.from_pretrained(
+ config.dataset.bert_name)
+ else:
+ raise ValueError("the encoder must be either MM or two backbones.")
+
+ def forward(
+ self,
+ caps,
+ cmasks,
+ vfeats,
+ vmasks,
+ **kwargs
+ ):
+ raise NotImplementedError(
+ "Please derive MMFusion module."
+ )
+
+ def _mm_on_the_fly(
+ self,
+ cmasks,
+ vmasks,
+ attention_mask
+ ):
+ """helper function for mask, seg_ids and token_type_ids."""
+ if attention_mask is None:
+ attention_mask = self._mm_attention_mask(cmasks, vmasks)
+
+ """
+ 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1
+ | first sequence | second sequence |
+ """
+ token_type_ids = torch.cat(
+ [
+ torch.zeros(
+ (vmasks.size(0), vmasks.size(1) + 2),
+ dtype=torch.long,
+ device=vmasks.device,
+ ),
+ torch.ones(
+ (cmasks.size(0), cmasks.size(1) - 2),
+ dtype=torch.long,
+ device=cmasks.device,
+ ),
+ ],
+ dim=1,
+ )
+ return attention_mask, token_type_ids
+
+ def _mm_attention_mask(self, cmasks, vmasks):
+ assert cmasks.size(0) == vmasks.size(0), "{}, {}, {}, {}".format(
+ str(cmasks.size()),
+ str(vmasks.size()),
+ str(cmasks.size(0)),
+ str(vmasks.size(0)),
+ )
+
+ mm_mask = torch.cat([cmasks[:, :1], vmasks, cmasks[:, 1:]], dim=1)
+ if self.last_iso_layer == 0:
+ # hard attention mask.
+ return mm_mask
+ else:
+ # a gpu iso mask; 0 : num_iso_layer is isolated;
+ # num_iso_layer: are MM-fused.
+ # make an iso layer
+ batch_size = cmasks.size(0)
+ iso_mask = self._make_iso_mask(batch_size, cmasks, vmasks)
+ mm_mask = mm_mask[:, None, :].repeat(1, mm_mask.size(-1), 1)
+ iso_mm_masks = []
+ # hard attention mask.
+ iso_mask = iso_mask[:, None, :, :].repeat(
+ 1, self.last_iso_layer, 1, 1)
+ iso_mm_masks.append(iso_mask)
+ if self.last_iso_layer < self.num_hidden_layers:
+ mm_mask = mm_mask[:, None, :, :].repeat(
+ 1, self.num_hidden_layers - self.last_iso_layer, 1, 1
+ )
+ iso_mm_masks.append(mm_mask)
+ iso_mm_masks = torch.cat(iso_mm_masks, dim=1)
+ return iso_mm_masks
+
+ def _make_iso_mask(self, batch_size, cmasks, vmasks):
+ cls_self_mask = torch.cat(
+ [
+ torch.ones(
+ (batch_size, 1), dtype=torch.bool, device=cmasks.device),
+ torch.zeros(
+ (batch_size, cmasks.size(1) + vmasks.size(1) - 1),
+ dtype=torch.bool, device=cmasks.device)
+ ], dim=1)
+
+ iso_video_mask = torch.cat(
+ [
+ # [CLS] is not used.
+ torch.zeros(
+ (batch_size, 1), dtype=torch.bool, device=cmasks.device
+ ),
+ vmasks,
+ # assume to be 1.
+ cmasks[:, 1:2],
+ # 2 means [CLS] + [SEP]
+ torch.zeros(
+ (batch_size, cmasks.size(1) - 2),
+ dtype=torch.bool,
+ device=cmasks.device,
+ ),
+ ],
+ dim=1,
+ )
+ iso_text_mask = torch.cat(
+ [
+ torch.zeros(
+ (batch_size, 2 + vmasks.size(1)),
+ dtype=torch.bool,
+ device=cmasks.device,
+ ), # [CLS] is not used.
+ cmasks[:, 2:], # assume to be 1.
+ ],
+ dim=1,
+ )
+ cls_self_mask = cls_self_mask[:, None, :]
+ iso_video_mask = iso_video_mask[:, None, :].repeat(
+ 1, vmasks.size(1) + 1, 1)
+ iso_text_mask = iso_text_mask[:, None, :].repeat(
+ 1, cmasks.size(1) - 2, 1)
+ return torch.cat([cls_self_mask, iso_video_mask, iso_text_mask], dim=1)
+
+ def _pooling_vt_layer(
+ self,
+ layered_sequence_output,
+ cmasks,
+ vmasks
+ ):
+ layer_idx = self.last_iso_layer \
+ if self.last_iso_layer > 0 else self.num_hidden_layers
+ hidden_state = layered_sequence_output[layer_idx]
+ # also output pooled_video and pooled_text.
+ batch_size = cmasks.size(0)
+ # pool the modality.
+ text_offset = vmasks.size(1) + 2 # [CLS] + [SEP]
+ # video tokens + [SEP]
+ video_outputs = hidden_state[:, 1:text_offset]
+ video_attention_mask = torch.cat(
+ [
+ vmasks,
+ torch.ones(
+ (batch_size, 1), dtype=torch.bool, device=vmasks.device),
+ ],
+ dim=1,
+ )
+ assert video_outputs.size(1) == video_attention_mask.size(1)
+ pooled_video = torch.sum(
+ video_outputs * video_attention_mask.unsqueeze(-1), dim=1
+ ) / video_attention_mask.sum(1, keepdim=True)
+ # pooled_video = torch.mean(video_outputs[0], dim=1)
+
+ # text tokens + [SEP]
+ text_attention_mask = cmasks[:, 2:]
+ text_outputs = hidden_state[:, text_offset:]
+ assert text_outputs.size(1) == text_attention_mask.size(1)
+ pooled_text = torch.sum(
+ text_outputs * text_attention_mask.unsqueeze(-1), dim=1
+ ) / text_attention_mask.sum(1, keepdim=True)
+ return pooled_video, pooled_text
+
+
+class MMFusionMFMMLM(MMFusion):
+ """forward function for MFM and MLM."""
+ def forward(
+ self,
+ caps,
+ cmasks,
+ vfeats,
+ vmasks,
+ attention_mask=None,
+ video_label=None,
+ text_label=None,
+ **kwargs
+ ):
+ output_hidden_states = False if self.is_train else True
+
+ target_vfeats, non_masked_frame_mask = None, None
+ if video_label is not None:
+ target_vfeats = vfeats.masked_select(
+ video_label.unsqueeze(-1)).view(
+ -1, vfeats.size(-1)
+ )
+ # mask video token.
+ vfeats[video_label] = 0.0
+ non_masked_frame_mask = vmasks.clone()
+ non_masked_frame_mask[video_label] = False
+
+ attention_mask, token_type_ids = self._mm_on_the_fly(
+ cmasks, vmasks, attention_mask)
+
+ outputs = self.mm_encoder(
+ input_ids=caps,
+ input_video_embeds=vfeats,
+ attention_mask=attention_mask,
+ token_type_ids=token_type_ids,
+ masked_frame_labels=video_label,
+ target_video_hidden_states=target_vfeats,
+ non_masked_frame_mask=non_masked_frame_mask,
+ masked_lm_labels=text_label,
+ output_hidden_states=output_hidden_states,
+ )
+
+ video_logits, text_logits = outputs[0], outputs[1]
+
+ if self.is_train: # return earlier for training.
+ return {
+ "video_logits": video_logits,
+ "text_logits": text_logits,
+ }
+
+ pooled_video, pooled_text = self._pooling_vt_layer(
+ outputs[2], cmasks, vmasks)
+ return {"pooled_video": pooled_video, "pooled_text": pooled_text}
+
+
+class MMFusionMTM(MMFusionMFMMLM):
+ def __init__(self, config, **kwargs):
+ super().__init__(config)
+ """
+ For reproducibility:
+ self.mm_encoder will be initialized then discarded.
+ """
+ from .transformermodel import MMBertForMTM
+ model_config = AutoConfig.from_pretrained(config.dataset.bert_name)
+ model_config.max_video_len = config.dataset.max_video_len
+ model_config.use_seg_emb = config.model.use_seg_emb
+ self.mm_encoder = MMBertForMTM.from_pretrained(
+ config.dataset.bert_name, config=model_config)
+
+
+class MMFusionShare(MMFusion):
+ """A retrival wrapper using mm_encoder as both video/text backbone.
+ TODO: move formally.
+ """
+ def forward(
+ self,
+ caps,
+ cmasks,
+ vfeats,
+ vmasks,
+ attention_mask=None,
+ video_label=None,
+ text_label=None,
+ output_hidden_states=False,
+ **kwargs
+ ):
+ pooled_video = self.forward_video(
+ vfeats,
+ vmasks,
+ caps,
+ cmasks,
+ output_hidden_states
+ )
+
+ pooled_text = self.forward_text(
+ caps,
+ cmasks,
+ output_hidden_states
+ )
+
+ return {"pooled_video": pooled_video, "pooled_text": pooled_text}
+
+ def forward_video(
+ self,
+ vfeats,
+ vmasks,
+ caps,
+ cmasks,
+ output_hidden_states=False,
+ **kwargs
+ ):
+ input_ids = caps[:, :2]
+
+ attention_mask = torch.cat([
+ cmasks[:, :1],
+ vmasks,
+ cmasks[:, 1:2]
+ ], dim=1)
+
+ token_type_ids = torch.zeros(
+ (vmasks.size(0), vmasks.size(1) + 2),
+ dtype=torch.long,
+ device=vmasks.device)
+
+ outputs = self.mm_encoder(
+ input_ids=input_ids,
+ input_video_embeds=vfeats,
+ attention_mask=attention_mask,
+ token_type_ids=token_type_ids,
+ output_hidden_states=True
+ )
+ video_outputs = outputs[0]
+
+ if output_hidden_states:
+ return video_outputs
+
+ batch_size = cmasks.size(0)
+
+ video_attention_mask = torch.cat(
+ [
+ torch.zeros(
+ (batch_size, 1), dtype=torch.bool, device=vmasks.device),
+ vmasks,
+ torch.ones(
+ (batch_size, 1), dtype=torch.bool, device=vmasks.device),
+ ],
+ dim=1,
+ )
+ assert video_outputs.size(1) == video_attention_mask.size(1)
+
+ video_attention_mask = video_attention_mask.type(video_outputs.dtype) \
+ / video_attention_mask.sum(1, keepdim=True)
+
+ pooled_video = torch.bmm(
+ video_outputs.transpose(2, 1),
+ video_attention_mask.unsqueeze(2)
+ ).squeeze(-1)
+ return pooled_video # video_outputs
+
+ def forward_text(
+ self,
+ caps,
+ cmasks,
+ output_hidden_states=False,
+ **kwargs
+ ):
+ input_ids = torch.cat([
+ caps[:, :1], caps[:, 2:],
+ ], dim=1)
+
+ attention_mask = torch.cat([
+ cmasks[:, :1],
+ cmasks[:, 2:]
+ ], dim=1)
+
+ token_type_ids = torch.cat([
+ torch.zeros(
+ (cmasks.size(0), 1),
+ dtype=torch.long,
+ device=cmasks.device),
+ torch.ones(
+ (cmasks.size(0), cmasks.size(1) - 2),
+ dtype=torch.long,
+ device=cmasks.device)
+ ], dim=1)
+
+ outputs = self.mm_encoder(
+ input_ids=input_ids,
+ input_video_embeds=None,
+ attention_mask=attention_mask,
+ token_type_ids=token_type_ids,
+ output_hidden_states=True
+ )
+ text_outputs = outputs[0]
+
+ if output_hidden_states:
+ return text_outputs
+
+ batch_size = caps.size(0)
+ # text tokens + [SEP]
+ text_attention_mask = torch.cat([
+ torch.zeros(
+ (batch_size, 1), dtype=torch.bool, device=cmasks.device),
+ cmasks[:, 2:]
+ ], dim=1)
+
+ assert text_outputs.size(1) == text_attention_mask.size(1)
+
+ text_attention_mask = text_attention_mask.type(text_outputs.dtype) \
+ / text_attention_mask.sum(1, keepdim=True)
+
+ pooled_text = torch.bmm(
+ text_outputs.transpose(2, 1),
+ text_attention_mask.unsqueeze(2)
+ ).squeeze(-1)
+ return pooled_text # text_outputs
+
+
+class MMFusionSeparate(MMFusionShare):
+ def forward_video(
+ self,
+ vfeats,
+ vmasks,
+ caps,
+ cmasks,
+ output_hidden_states=False,
+ **kwargs
+ ):
+ input_ids = caps[:, :2]
+
+ attention_mask = torch.cat([
+ cmasks[:, :1],
+ vmasks,
+ cmasks[:, 1:2]
+ ], dim=1)
+
+ token_type_ids = torch.zeros(
+ (vmasks.size(0), vmasks.size(1) + 2),
+ dtype=torch.long,
+ device=vmasks.device)
+
+ outputs = self.video_encoder(
+ input_ids=input_ids,
+ input_video_embeds=vfeats,
+ attention_mask=attention_mask,
+ token_type_ids=token_type_ids,
+ output_hidden_states=True
+ )
+ video_outputs = outputs[0]
+
+ if output_hidden_states:
+ return video_outputs
+
+ batch_size = cmasks.size(0)
+
+ video_attention_mask = torch.cat(
+ [
+ torch.zeros(
+ (batch_size, 1), dtype=torch.bool, device=vmasks.device),
+ vmasks,
+ torch.ones(
+ (batch_size, 1), dtype=torch.bool, device=vmasks.device),
+ ],
+ dim=1,
+ )
+ assert video_outputs.size(1) == video_attention_mask.size(1)
+
+ video_attention_mask = video_attention_mask.type(video_outputs.dtype) \
+ / video_attention_mask.sum(1, keepdim=True)
+
+ pooled_video = torch.bmm(
+ video_outputs.transpose(2, 1),
+ video_attention_mask.unsqueeze(2)
+ ).squeeze(-1)
+ return pooled_video # video_outputs
+
+ def forward_text(
+ self,
+ caps,
+ cmasks,
+ output_hidden_states=False,
+ **kwargs
+ ):
+ input_ids = torch.cat([
+ caps[:, :1], caps[:, 2:],
+ ], dim=1)
+
+ attention_mask = torch.cat([
+ cmasks[:, :1],
+ cmasks[:, 2:]
+ ], dim=1)
+ # different from sharing, we use all-0 type.
+ token_type_ids = torch.zeros(
+ (cmasks.size(0), cmasks.size(1) - 1),
+ dtype=torch.long,
+ device=cmasks.device)
+
+ outputs = self.text_encoder(
+ input_ids=input_ids,
+ attention_mask=attention_mask,
+ token_type_ids=token_type_ids,
+ output_hidden_states=True
+ )
+ text_outputs = outputs[0]
+
+ if output_hidden_states:
+ return text_outputs
+
+ batch_size = caps.size(0)
+ # text tokens + [SEP]
+ text_attention_mask = torch.cat([
+ torch.zeros(
+ (batch_size, 1), dtype=torch.bool, device=cmasks.device),
+ cmasks[:, 2:]
+ ], dim=1)
+
+ assert text_outputs.size(1) == text_attention_mask.size(1)
+
+ text_attention_mask = text_attention_mask.type(text_outputs.dtype) \
+ / text_attention_mask.sum(1, keepdim=True)
+
+ pooled_text = torch.bmm(
+ text_outputs.transpose(2, 1),
+ text_attention_mask.unsqueeze(2)
+ ).squeeze(-1)
+ return pooled_text # text_outputs
+
+
+class MMFusionJoint(MMFusion):
+ """fine-tuning wrapper for retrival task."""
+
+ def forward(
+ self,
+ caps,
+ cmasks,
+ vfeats,
+ vmasks,
+ attention_mask=None,
+ video_label=None,
+ text_label=None,
+ **kwargs
+ ):
+ # TODO (huxu): other ways to do negative examples; move the following
+ # into your criterion forward.
+ output_hidden_states = True
+
+ attention_mask, token_type_ids = self._mm_on_the_fly(
+ cmasks, vmasks, attention_mask)
+
+ separate_forward_split = (
+ None if self.is_train else vmasks.size(1) + 2
+ ) # [CLS] + [SEP]
+
+ outputs = self.mm_encoder(
+ input_ids=caps,
+ input_video_embeds=vfeats,
+ attention_mask=attention_mask,
+ token_type_ids=token_type_ids,
+ output_hidden_states=output_hidden_states,
+ separate_forward_split=separate_forward_split,
+ )
+
+ pooled_video, pooled_text = self._pooling_vt_layer(
+ outputs[2], cmasks, vmasks)
+ return {"pooled_video": pooled_video, "pooled_text": pooled_text}
+
+
+class MMFusionActionSegmentation(MMFusion):
+ """Fine-tuning wrapper for action segmentation.
+ TODO: rename this for VLM.
+ """
+ def forward(
+ self,
+ caps,
+ cmasks,
+ vfeats,
+ vmasks,
+ attention_mask=None,
+ **kwargs
+ ):
+ # ActionLocalization assume of batch_size=1, squeeze it.
+ caps = caps.view(-1, caps.size(-1))
+ cmasks = cmasks.view(-1, cmasks.size(-1))
+ vfeats = vfeats.view(-1, vfeats.size(2), vfeats.size(3))
+ vmasks = vmasks.view(-1, vmasks.size(-1))
+
+ # this may not cover all shapes of attention_mask.
+ attention_mask = attention_mask.view(
+ -1, attention_mask.size(2), attention_mask.size(3)) \
+ if attention_mask is not None else None
+
+ # TODO (huxu): other ways to do negative examples; move the following
+ # into your criterion forward.
+ output_hidden_states = True
+
+ # video forwarding, text is dummy; never use attention_mask.
+ attention_mask, token_type_ids = self._mm_on_the_fly(
+ cmasks, vmasks, attention_mask)
+
+ logits = self.mm_encoder(
+ input_ids=caps,
+ input_video_embeds=vfeats,
+ attention_mask=attention_mask,
+ token_type_ids=token_type_ids,
+ output_hidden_states=output_hidden_states,
+ )
+ return {"logits": logits[0][:, 1:vmasks.size(1)+1]}
+
+
+class MMFusionActionLocalization(MMFusion):
+ """fine-tuning model for retrival task."""
+
+ def __init__(self, config, **kwargs):
+ super().__init__(config)
+ tokenizer = AutoTokenizer.from_pretrained(
+ config.dataset.bert_name)
+ self.cls_token_id = tokenizer.cls_token_id
+ self.sep_token_id = tokenizer.sep_token_id
+ self.pad_token_id = tokenizer.pad_token_id
+
+ def forward(
+ self,
+ caps,
+ cmasks,
+ vfeats,
+ vmasks,
+ attention_mask=None,
+ **kwargs
+ ):
+ # ActionLocalization assume of batch_size=1, squeeze it.
+ caps = caps.squeeze(0)
+ cmasks = cmasks.squeeze(0)
+ vfeats = vfeats.squeeze(0)
+ vmasks = vmasks.squeeze(0)
+ attention_mask = attention_mask.squeeze(0) if attention_mask is not None else None
+
+ # TODO (huxu): other ways to do negative examples; move the following
+ # into your criterion forward.
+ output_hidden_states = True
+
+ # a len1 dummy video token.
+ dummy_vfeats = torch.zeros(
+ (caps.size(0), 1, vfeats.size(-1)), device=vfeats.device, dtype=vfeats.dtype)
+ dummy_vmasks = torch.ones(
+ (caps.size(0), 1), dtype=torch.bool,
+ device=vfeats.device)
+
+ dummy_caps = torch.LongTensor(
+ [[self.cls_token_id, self.sep_token_id,
+ self.pad_token_id, self.sep_token_id]],
+ ).to(caps.device).repeat(vfeats.size(0), 1)
+ dummy_cmasks = torch.BoolTensor(
+ [[0, 1, 0, 1]] # pad are valid for attention.
+ ).to(caps.device).repeat(vfeats.size(0), 1)
+
+ # video forwarding, text is dummy; never use attention_mask.
+ attention_mask, token_type_ids = self._mm_on_the_fly(
+ dummy_cmasks, vmasks, None)
+
+ outputs = self.mm_encoder(
+ input_ids=dummy_caps,
+ input_video_embeds=vfeats,
+ attention_mask=attention_mask,
+ token_type_ids=token_type_ids,
+ output_hidden_states=output_hidden_states,
+ )
+
+ layer_idx = self.last_iso_layer \
+ if self.last_iso_layer > 0 else self.num_hidden_layers
+
+ video_seq = outputs[2][layer_idx][:, 1:vmasks.size(1)+1].masked_select(
+ vmasks.unsqueeze(-1)
+ ).view(-1, self.hidden_size)
+
+ # text forwarding, video is dummy
+ attention_mask, token_type_ids = self._mm_on_the_fly(
+ cmasks, dummy_vmasks, None)
+
+ outputs = self.mm_encoder(
+ input_ids=caps,
+ input_video_embeds=dummy_vfeats,
+ attention_mask=attention_mask,
+ token_type_ids=token_type_ids,
+ output_hidden_states=output_hidden_states,
+ )
+
+ _, pooled_text = self._pooling_vt_layer(
+ outputs[2], cmasks, dummy_vmasks)
+ # this line is not right.
+ logits = torch.mm(video_seq, pooled_text.transpose(1, 0))
+ return {"logits": logits}
+
+
+# --------------- MMFusionSeparate for end tasks ---------------
+
+class MMFusionSeparateActionSegmentation(MMFusionSeparate):
+ """Fine-tuning wrapper for action segmentation."""
+ def forward(
+ self,
+ caps,
+ cmasks,
+ vfeats,
+ vmasks,
+ attention_mask=None,
+ **kwargs
+ ):
+ # ActionLocalization assume of batch_size=1, squeeze it.
+ caps = caps.view(-1, caps.size(-1))
+ cmasks = cmasks.view(-1, cmasks.size(-1))
+ vfeats = vfeats.view(-1, vfeats.size(2), vfeats.size(3))
+ vmasks = vmasks.view(-1, vmasks.size(-1))
+ logits = self.forward_video(
+ vfeats,
+ vmasks,
+ caps,
+ cmasks,
+ output_hidden_states=True
+ )
+ return {"logits": logits[:, 1:vmasks.size(1)+1]}
+
+
+class MMFusionSeparateActionLocalization(MMFusionSeparate):
+ def __init__(self, config, **kwargs):
+ super().__init__(config)
+ tokenizer = AutoTokenizer.from_pretrained(
+ config.dataset.bert_name)
+ self.cls_token_id = tokenizer.cls_token_id
+ self.sep_token_id = tokenizer.sep_token_id
+ self.pad_token_id = tokenizer.pad_token_id
+
+ def forward(
+ self,
+ caps,
+ cmasks,
+ vfeats,
+ vmasks,
+ **kwargs
+ ):
+ # ActionLocalization assume of batch_size=1, squeeze it.
+ caps = caps.squeeze(0)
+ cmasks = cmasks.squeeze(0)
+ vfeats = vfeats.squeeze(0)
+ vmasks = vmasks.squeeze(0)
+
+ # TODO (huxu): other ways to do negative examples; move the following
+ # into your criterion forward.
+ dummy_caps = torch.LongTensor(
+ [[self.cls_token_id, self.sep_token_id,
+ self.pad_token_id, self.sep_token_id]],
+ ).to(caps.device).repeat(vfeats.size(0), 1)
+ dummy_cmasks = torch.BoolTensor(
+ [[0, 1, 0, 1]] # pad are valid for attention.
+ ).to(caps.device).repeat(vfeats.size(0), 1)
+
+ outputs = self.forward_video(
+ vfeats,
+ vmasks,
+ dummy_caps,
+ dummy_cmasks,
+ output_hidden_states=True
+ )
+
+ video_seq = outputs[:, 1:vmasks.size(1)+1].masked_select(
+ vmasks.unsqueeze(-1)
+ ).view(-1, self.hidden_size)
+
+ pooled_text = self.forward_text(
+ caps,
+ cmasks,
+ output_hidden_states=False
+ )
+
+ # this line is not right.
+ logits = torch.mm(video_seq, pooled_text.transpose(1, 0))
+ return {"logits": logits}
+
+
+class MMFusionShareActionLocalization(MMFusionShare):
+ def __init__(self, config, **kwargs):
+ super().__init__(config)
+ tokenizer = AutoTokenizer.from_pretrained(
+ config.dataset.bert_name)
+ self.cls_token_id = tokenizer.cls_token_id
+ self.sep_token_id = tokenizer.sep_token_id
+ self.pad_token_id = tokenizer.pad_token_id
+
+ def forward(
+ self,
+ caps,
+ cmasks,
+ vfeats,
+ vmasks,
+ **kwargs
+ ):
+ # ActionLocalization assume of batch_size=1, squeeze it.
+ caps = caps.squeeze(0)
+ cmasks = cmasks.squeeze(0)
+ vfeats = vfeats.squeeze(0)
+ vmasks = vmasks.squeeze(0)
+
+ # TODO (huxu): other ways to do negative examples; move the following
+ # into your criterion forward.
+ dummy_caps = torch.LongTensor(
+ [[self.cls_token_id, self.sep_token_id,
+ self.pad_token_id, self.sep_token_id]],
+ ).to(caps.device).repeat(vfeats.size(0), 1)
+ dummy_cmasks = torch.BoolTensor(
+ [[0, 1, 0, 1]] # pad are valid for attention.
+ ).to(caps.device).repeat(vfeats.size(0), 1)
+
+ outputs = self.forward_video(
+ vfeats,
+ vmasks,
+ dummy_caps,
+ dummy_cmasks,
+ output_hidden_states=True
+ )
+
+ video_seq = outputs[:, 1:vmasks.size(1)+1].masked_select(
+ vmasks.unsqueeze(-1)
+ ).view(-1, self.hidden_size)
+
+ pooled_text = self.forward_text(
+ caps,
+ cmasks,
+ output_hidden_states=False
+ )
+
+ # this line is not right.
+ logits = torch.mm(video_seq, pooled_text.transpose(1, 0))
+ return {"logits": logits}
diff --git a/fairseq/examples/MMPT/mmpt/models/mmfusionnlg.py b/fairseq/examples/MMPT/mmpt/models/mmfusionnlg.py
new file mode 100644
index 0000000000000000000000000000000000000000..9207e77dab3025d7a26efcce0795183de1d34fc7
--- /dev/null
+++ b/fairseq/examples/MMPT/mmpt/models/mmfusionnlg.py
@@ -0,0 +1,999 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors, Facebook AI Research authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# Copyright (c) Facebook, Inc. All Rights Reserved
+
+
+import torch
+
+from torch.nn import functional as F
+
+from typing import Optional, Iterable
+
+try:
+ from transformers import BertPreTrainedModel
+ from transformers.modeling_bert import BertOnlyMLMHead
+
+ from transformers.file_utils import ModelOutput
+ from transformers.modeling_outputs import CausalLMOutput
+ from transformers.generation_utils import (
+ BeamHypotheses,
+ top_k_top_p_filtering
+ )
+except ImportError:
+ pass
+
+from .mmfusion import MMFusion
+from .transformermodel import MMBertModel
+from ..modules import VideoTokenMLP
+
+
+class MMFusionNLG(MMFusion):
+ def __init__(self, config, **kwargs):
+ super().__init__(config)
+ if config.model.max_decode_length is not None:
+ self.max_length = min(
+ config.model.max_decode_length,
+ config.dataset.max_len - config.dataset.max_video_len - 3
+ )
+ else:
+ self.max_length = \
+ config.dataset.max_len - config.dataset.max_video_len - 3
+ self.gen_param = config.gen_param if config.gen_param is not None \
+ else {}
+
+ def forward(
+ self,
+ caps,
+ cmasks,
+ vfeats,
+ vmasks,
+ attention_mask,
+ video_label=None,
+ text_label=None,
+ **kwargs
+ ):
+ """use pre-trained LM header for generation."""
+ attention_mask, token_type_ids = self._mm_on_the_fly(
+ cmasks, vmasks, attention_mask)
+
+ outputs = self.mm_encoder(
+ input_ids=caps,
+ input_video_embeds=vfeats,
+ attention_mask=attention_mask,
+ token_type_ids=token_type_ids,
+ masked_lm_labels=text_label,
+ )
+ return {"logits": outputs[0]}
+
+ @torch.no_grad()
+ def generate(
+ self,
+ caps, cmasks, vfeats, vmasks,
+ attention_mask=None,
+ bos_token_id=None,
+ eos_token_id=None,
+ **kwargs
+ ):
+ # a simplified interface from
+ # https://huggingface.co/transformers/v3.4.0/_modules/transformers/generation_utils.html#GenerationMixin.generate
+
+ # caps now only have
+ # [CLS], [SEP] (for video) and [CLS] (as bos_token)
+ assert caps.size(1) == 3
+
+ attention_mask, token_type_ids = self._mm_on_the_fly(
+ cmasks, vmasks, attention_mask)
+
+ output = self.mm_encoder.generate(
+ input_ids=caps,
+ input_video_embeds=vfeats,
+ attention_mask=attention_mask,
+ token_type_ids=token_type_ids,
+ bos_token_id=bos_token_id,
+ eos_token_id=eos_token_id,
+ max_length=self.max_length,
+ **self.gen_param
+ )
+ return output
+
+
+class MMBertForNLG(BertPreTrainedModel):
+ def __init__(self, config):
+ super().__init__(config)
+ self.bert = MMBertModel(config)
+ self.videomlp = VideoTokenMLP(config)
+ # we do not use `BertGenerationOnlyLMHead`
+ # because we can reuse pretraining.
+ self.cls = BertOnlyMLMHead(config)
+ self.hidden_size = config.hidden_size
+ self.init_weights()
+
+ def get_output_embeddings(self):
+ return self.cls.predictions.decoder
+
+ def forward(
+ self,
+ input_ids=None,
+ input_video_embeds=None,
+ attention_mask=None,
+ token_type_ids=None,
+ position_ids=None,
+ head_mask=None,
+ inputs_embeds=None,
+ masked_lm_labels=None,
+ output_attentions=None,
+ output_hidden_states=None,
+ return_dict=None,
+ ):
+ # similar to MMBertForMFMMLM without MFM.
+ video_tokens = self.videomlp(input_video_embeds)
+ outputs = self.bert(
+ input_ids,
+ video_tokens,
+ attention_mask=attention_mask,
+ token_type_ids=token_type_ids,
+ position_ids=position_ids,
+ head_mask=head_mask,
+ inputs_embeds=inputs_embeds,
+ output_attentions=output_attentions,
+ output_hidden_states=output_hidden_states,
+ return_dict=return_dict,
+ )
+
+ sequence_output = outputs[0]
+
+ prediction_scores = None
+ if masked_lm_labels is not None:
+ text_offset = input_video_embeds.size(1) + 1 # [CLS]
+ # recover caps format: [CLS] [SEP] text [SEP]
+ text_sequence_output = torch.cat(
+ [sequence_output[:, :1], sequence_output[:, text_offset:]],
+ dim=1
+ )
+
+ # only compute select tokens to training to speed up.
+ hidden_size = text_sequence_output.size(-1)
+ # masked_lm_labels = masked_lm_labels.reshape(-1)
+ labels_mask = masked_lm_labels != -100
+
+ selected_text_output = text_sequence_output.masked_select(
+ labels_mask.unsqueeze(-1)
+ ).view(-1, hidden_size)
+ prediction_scores = self.cls(selected_text_output)
+
+ if not return_dict:
+ output = (
+ prediction_scores,
+ ) + outputs[2:]
+ return output
+
+ # for generation.
+ text_offset = input_video_embeds.size(1) + 2 # [CLS]
+ text_sequence_output = sequence_output[:, text_offset:]
+ prediction_scores = self.cls(text_sequence_output)
+ return CausalLMOutput(
+ loss=None,
+ logits=prediction_scores,
+ )
+
+ def prepare_inputs_for_generation(
+ self,
+ input_ids,
+ input_video_embeds,
+ attention_mask=None,
+ token_type_ids=None,
+ **model_kwargs
+ ):
+ # must return a dictionary.
+ seq_len = input_ids.size(1) + input_video_embeds.size(1)
+ if attention_mask is not None:
+ if len(attention_mask.size()) == 4:
+ attention_mask = attention_mask[:, :, :seq_len, :seq_len]
+ elif len(attention_mask.size()) == 3:
+ attention_mask = attention_mask[:, :seq_len, :seq_len]
+ else:
+ attention_mask = attention_mask[:, :seq_len]
+ if token_type_ids is not None:
+ token_type_ids = token_type_ids[:, :seq_len]
+
+ return {
+ "input_ids": input_ids,
+ "input_video_embeds": input_video_embeds,
+ "attention_mask": attention_mask,
+ "token_type_ids": token_type_ids,
+ }
+
+ @torch.no_grad()
+ def generate(
+ self,
+ input_ids: Optional[torch.LongTensor] = None,
+ decoder_input_ids: Optional[torch.LongTensor] = None,
+ max_length: Optional[int] = None,
+ min_length: Optional[int] = None,
+ do_sample: Optional[bool] = None,
+ early_stopping: Optional[bool] = None,
+ num_beams: Optional[int] = None,
+ temperature: Optional[float] = None,
+ top_k: Optional[int] = None,
+ top_p: Optional[float] = None,
+ repetition_penalty: Optional[float] = None,
+ bad_words_ids: Optional[Iterable[int]] = None,
+ bos_token_id: Optional[int] = None,
+ pad_token_id: Optional[int] = None,
+ eos_token_id: Optional[int] = None,
+ length_penalty: Optional[float] = None,
+ no_repeat_ngram_size: Optional[int] = None,
+ num_return_sequences: Optional[int] = None,
+ attention_mask: Optional[torch.LongTensor] = None,
+ decoder_start_token_id: Optional[int] = None,
+ use_cache: Optional[bool] = None,
+ **model_kwargs
+ ) -> torch.LongTensor:
+ r"""
+ Generates sequences for models with a language modeling head. The method currently supports greedy decoding,
+ beam-search decoding, sampling with temperature, sampling with top-k or nucleus sampling.
+ Adapted in part from `Facebook's XLM beam search code
+ 
+
+
+See the [fairscale docs](https://fairscale.readthedocs.io/en/latest/api/nn/fsdp_tips.html) for a more detailed
+explanation of how FSDP works.
+
+