text
stringlengths 7
1.24M
| id
stringlengths 14
166
| metadata
dict | __index_level_0__
int64 0
519
|
|---|---|---|---|
#!/usr/bin/env python
# Copyright 2020 The HuggingFace Team. 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.
import fire
from transformers import AutoConfig, AutoModelForSeq2SeqLM, AutoTokenizer
def save_randomly_initialized_version(config_name: str, save_dir: str, **config_kwargs):
"""Save a randomly initialized version of a model using a pretrained config.
Args:
config_name: which config to use
save_dir: where to save the resulting model and tokenizer
config_kwargs: Passed to AutoConfig
Usage::
save_randomly_initialized_version("facebook/bart-large-cnn", "distilbart_random_cnn_6_3", encoder_layers=6, decoder_layers=3, num_beams=3)
"""
cfg = AutoConfig.from_pretrained(config_name, **config_kwargs)
model = AutoModelForSeq2SeqLM.from_config(cfg)
model.save_pretrained(save_dir)
AutoTokenizer.from_pretrained(config_name).save_pretrained(save_dir)
return model
if __name__ == "__main__":
fire.Fire(save_randomly_initialized_version)
|
transformers/examples/legacy/seq2seq/save_randomly_initialized_model.py/0
|
{
"file_path": "transformers/examples/legacy/seq2seq/save_randomly_initialized_model.py",
"repo_id": "transformers",
"token_count": 498
}
| 313
|
# Copyright 2020 The HuggingFace Team. 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.
export WANDB_PROJECT=distilbart-trainer
export BS=32
export m=sshleifer/student_cnn_12_6
export tok=facebook/bart-large
export MAX_TGT_LEN=142
python finetune_trainer.py \
--model_name_or_path $m --tokenizer_name $tok \
--data_dir cnn_dm \
--output_dir distilbart-cnn-12-6 --overwrite_output_dir \
--learning_rate=3e-5 \
--warmup_steps 500 --sortish_sampler \
--fp16 \
--n_val 500 \
--gradient_accumulation_steps=1 \
--per_device_train_batch_size=$BS --per_device_eval_batch_size=$BS \
--freeze_encoder --freeze_embeds \
--num_train_epochs=2 \
--save_steps 3000 --eval_steps 3000 \
--logging_first_step \
--max_target_length 56 --val_max_target_length $MAX_TGT_LEN --test_max_target_length $MAX_TGT_LEN\
--do_train --do_eval --do_predict \
--eval_strategy steps \
--predict_with_generate --sortish_sampler \
"$@"
|
transformers/examples/legacy/seq2seq/train_distilbart_cnn.sh/0
|
{
"file_path": "transformers/examples/legacy/seq2seq/train_distilbart_cnn.sh",
"repo_id": "transformers",
"token_count": 540
}
| 314
|
#!/usr/bin/env python
# coding=utf-8
# Copyright 2021 The HuggingFace Inc. team. 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.
import logging
import os
import sys
import warnings
from dataclasses import dataclass, field
from random import randint
from typing import Optional
import datasets
import evaluate
import numpy as np
from datasets import DatasetDict, load_dataset
import transformers
from transformers import (
AutoConfig,
AutoFeatureExtractor,
AutoModelForAudioClassification,
HfArgumentParser,
Trainer,
TrainingArguments,
set_seed,
)
from transformers.trainer_utils import get_last_checkpoint
from transformers.utils import check_min_version, send_example_telemetry
from transformers.utils.versions import require_version
logger = logging.getLogger(__name__)
# Will error if the minimal version of Transformers is not installed. Remove at your own risks.
check_min_version("4.45.0.dev0")
require_version("datasets>=1.14.0", "To fix: pip install -r examples/pytorch/audio-classification/requirements.txt")
def random_subsample(wav: np.ndarray, max_length: float, sample_rate: int = 16000):
"""Randomly sample chunks of `max_length` seconds from the input audio"""
sample_length = int(round(sample_rate * max_length))
if len(wav) <= sample_length:
return wav
random_offset = randint(0, len(wav) - sample_length - 1)
return wav[random_offset : random_offset + sample_length]
@dataclass
class DataTrainingArguments:
"""
Arguments pertaining to what data we are going to input our model for training and eval.
Using `HfArgumentParser` we can turn this class
into argparse arguments to be able to specify them on
the command line.
"""
dataset_name: Optional[str] = field(default=None, metadata={"help": "Name of a dataset from the datasets package"})
dataset_config_name: Optional[str] = field(
default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
)
train_file: Optional[str] = field(
default=None, metadata={"help": "A file containing the training audio paths and labels."}
)
eval_file: Optional[str] = field(
default=None, metadata={"help": "A file containing the validation audio paths and labels."}
)
train_split_name: str = field(
default="train",
metadata={
"help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
},
)
eval_split_name: str = field(
default="validation",
metadata={
"help": (
"The name of the training data set split to use (via the datasets library). Defaults to 'validation'"
)
},
)
audio_column_name: str = field(
default="audio",
metadata={"help": "The name of the dataset column containing the audio data. Defaults to 'audio'"},
)
label_column_name: str = field(
default="label", metadata={"help": "The name of the dataset column containing the labels. Defaults to 'label'"}
)
max_train_samples: Optional[int] = field(
default=None,
metadata={
"help": (
"For debugging purposes or quicker training, truncate the number of training examples to this "
"value if set."
)
},
)
max_eval_samples: Optional[int] = field(
default=None,
metadata={
"help": (
"For debugging purposes or quicker training, truncate the number of evaluation examples to this "
"value if set."
)
},
)
max_length_seconds: float = field(
default=20,
metadata={"help": "Audio clips will be randomly cut to this length during training if the value is set."},
)
@dataclass
class ModelArguments:
"""
Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
"""
model_name_or_path: str = field(
default="facebook/wav2vec2-base",
metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"},
)
config_name: Optional[str] = field(
default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"}
)
cache_dir: Optional[str] = field(
default=None, metadata={"help": "Where do you want to store the pretrained models downloaded from the Hub"}
)
model_revision: str = field(
default="main",
metadata={"help": "The specific model version to use (can be a branch name, tag name or commit id)."},
)
feature_extractor_name: Optional[str] = field(
default=None, metadata={"help": "Name or path of preprocessor config."}
)
freeze_feature_encoder: bool = field(
default=True, metadata={"help": "Whether to freeze the feature encoder layers of the model."}
)
attention_mask: bool = field(
default=True, metadata={"help": "Whether to generate an attention mask in the feature extractor."}
)
token: str = field(
default=None,
metadata={
"help": (
"The token to use as HTTP bearer authorization for remote files. If not specified, will use the token "
"generated when running `huggingface-cli login` (stored in `~/.huggingface`)."
)
},
)
trust_remote_code: bool = field(
default=False,
metadata={
"help": (
"Whether to trust the execution of code from datasets/models defined on the Hub."
" This option should only be set to `True` for repositories you trust and in which you have read the"
" code, as it will execute code present on the Hub on your local machine."
)
},
)
freeze_feature_extractor: Optional[bool] = field(
default=None, metadata={"help": "Whether to freeze the feature extractor layers of the model."}
)
ignore_mismatched_sizes: bool = field(
default=False,
metadata={"help": "Will enable to load a pretrained model whose head dimensions are different."},
)
def __post_init__(self):
if not self.freeze_feature_extractor and self.freeze_feature_encoder:
warnings.warn(
"The argument `--freeze_feature_extractor` is deprecated and "
"will be removed in a future version. Use `--freeze_feature_encoder` "
"instead. Setting `freeze_feature_encoder==True`.",
FutureWarning,
)
if self.freeze_feature_extractor and not self.freeze_feature_encoder:
raise ValueError(
"The argument `--freeze_feature_extractor` is deprecated and "
"should not be used in combination with `--freeze_feature_encoder`. "
"Only make use of `--freeze_feature_encoder`."
)
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# If we pass only one argument to the script and it's the path to a json file,
# let's parse it to get our arguments.
model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
else:
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
# Sending telemetry. Tracking the example usage helps us better allocate resources to maintain them. The
# information sent is the one passed as arguments along with your Python/PyTorch versions.
send_example_telemetry("run_audio_classification", model_args, data_args)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
handlers=[logging.StreamHandler(sys.stdout)],
)
if training_args.should_log:
# The default of training_args.log_level is passive, so we set log level at info here to have that default.
transformers.utils.logging.set_verbosity_info()
log_level = training_args.get_process_log_level()
logger.setLevel(log_level)
transformers.utils.logging.set_verbosity(log_level)
transformers.utils.logging.enable_default_handler()
transformers.utils.logging.enable_explicit_format()
# Log on each process the small summary:
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}, "
+ f"distributed training: {training_args.parallel_mode.value == 'distributed'}, 16-bits training: {training_args.fp16}"
)
logger.info(f"Training/evaluation parameters {training_args}")
# Set seed before initializing model.
set_seed(training_args.seed)
# Detecting last checkpoint.
last_checkpoint = None
if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
last_checkpoint = get_last_checkpoint(training_args.output_dir)
if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
"Use --overwrite_output_dir to train from scratch."
)
elif last_checkpoint is not None and training_args.resume_from_checkpoint is None:
logger.info(
f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
"the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
)
# Initialize our dataset and prepare it for the audio classification task.
raw_datasets = DatasetDict()
raw_datasets["train"] = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
split=data_args.train_split_name,
token=model_args.token,
trust_remote_code=model_args.trust_remote_code,
)
raw_datasets["eval"] = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
split=data_args.eval_split_name,
token=model_args.token,
trust_remote_code=model_args.trust_remote_code,
)
if data_args.audio_column_name not in raw_datasets["train"].column_names:
raise ValueError(
f"--audio_column_name {data_args.audio_column_name} not found in dataset '{data_args.dataset_name}'. "
"Make sure to set `--audio_column_name` to the correct audio column - one of "
f"{', '.join(raw_datasets['train'].column_names)}."
)
if data_args.label_column_name not in raw_datasets["train"].column_names:
raise ValueError(
f"--label_column_name {data_args.label_column_name} not found in dataset '{data_args.dataset_name}'. "
"Make sure to set `--label_column_name` to the correct text column - one of "
f"{', '.join(raw_datasets['train'].column_names)}."
)
# Setting `return_attention_mask=True` is the way to get a correctly masked mean-pooling over
# transformer outputs in the classifier, but it doesn't always lead to better accuracy
feature_extractor = AutoFeatureExtractor.from_pretrained(
model_args.feature_extractor_name or model_args.model_name_or_path,
return_attention_mask=model_args.attention_mask,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
token=model_args.token,
trust_remote_code=model_args.trust_remote_code,
)
# `datasets` takes care of automatically loading and resampling the audio,
# so we just need to set the correct target sampling rate.
raw_datasets = raw_datasets.cast_column(
data_args.audio_column_name, datasets.features.Audio(sampling_rate=feature_extractor.sampling_rate)
)
model_input_name = feature_extractor.model_input_names[0]
def train_transforms(batch):
"""Apply train_transforms across a batch."""
subsampled_wavs = []
for audio in batch[data_args.audio_column_name]:
wav = random_subsample(
audio["array"], max_length=data_args.max_length_seconds, sample_rate=feature_extractor.sampling_rate
)
subsampled_wavs.append(wav)
inputs = feature_extractor(subsampled_wavs, sampling_rate=feature_extractor.sampling_rate)
output_batch = {model_input_name: inputs.get(model_input_name)}
output_batch["labels"] = list(batch[data_args.label_column_name])
return output_batch
def val_transforms(batch):
"""Apply val_transforms across a batch."""
wavs = [audio["array"] for audio in batch[data_args.audio_column_name]]
inputs = feature_extractor(wavs, sampling_rate=feature_extractor.sampling_rate)
output_batch = {model_input_name: inputs.get(model_input_name)}
output_batch["labels"] = list(batch[data_args.label_column_name])
return output_batch
# Prepare label mappings.
# We'll include these in the model's config to get human readable labels in the Inference API.
labels = raw_datasets["train"].features[data_args.label_column_name].names
label2id, id2label = {}, {}
for i, label in enumerate(labels):
label2id[label] = str(i)
id2label[str(i)] = label
# Load the accuracy metric from the datasets package
metric = evaluate.load("accuracy", cache_dir=model_args.cache_dir)
# Define our compute_metrics function. It takes an `EvalPrediction` object (a namedtuple with
# `predictions` and `label_ids` fields) and has to return a dictionary string to float.
def compute_metrics(eval_pred):
"""Computes accuracy on a batch of predictions"""
predictions = np.argmax(eval_pred.predictions, axis=1)
return metric.compute(predictions=predictions, references=eval_pred.label_ids)
config = AutoConfig.from_pretrained(
model_args.config_name or model_args.model_name_or_path,
num_labels=len(labels),
label2id=label2id,
id2label=id2label,
finetuning_task="audio-classification",
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
token=model_args.token,
trust_remote_code=model_args.trust_remote_code,
)
model = AutoModelForAudioClassification.from_pretrained(
model_args.model_name_or_path,
from_tf=bool(".ckpt" in model_args.model_name_or_path),
config=config,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
token=model_args.token,
trust_remote_code=model_args.trust_remote_code,
ignore_mismatched_sizes=model_args.ignore_mismatched_sizes,
)
# freeze the convolutional waveform encoder
if model_args.freeze_feature_encoder:
model.freeze_feature_encoder()
if training_args.do_train:
if data_args.max_train_samples is not None:
raw_datasets["train"] = (
raw_datasets["train"].shuffle(seed=training_args.seed).select(range(data_args.max_train_samples))
)
# Set the training transforms
raw_datasets["train"].set_transform(train_transforms, output_all_columns=False)
if training_args.do_eval:
if data_args.max_eval_samples is not None:
raw_datasets["eval"] = (
raw_datasets["eval"].shuffle(seed=training_args.seed).select(range(data_args.max_eval_samples))
)
# Set the validation transforms
raw_datasets["eval"].set_transform(val_transforms, output_all_columns=False)
# Initialize our trainer
trainer = Trainer(
model=model,
args=training_args,
train_dataset=raw_datasets["train"] if training_args.do_train else None,
eval_dataset=raw_datasets["eval"] if training_args.do_eval else None,
compute_metrics=compute_metrics,
tokenizer=feature_extractor,
)
# Training
if training_args.do_train:
checkpoint = None
if training_args.resume_from_checkpoint is not None:
checkpoint = training_args.resume_from_checkpoint
elif last_checkpoint is not None:
checkpoint = last_checkpoint
train_result = trainer.train(resume_from_checkpoint=checkpoint)
trainer.save_model()
trainer.log_metrics("train", train_result.metrics)
trainer.save_metrics("train", train_result.metrics)
trainer.save_state()
# Evaluation
if training_args.do_eval:
metrics = trainer.evaluate()
trainer.log_metrics("eval", metrics)
trainer.save_metrics("eval", metrics)
# Write model card and (optionally) push to hub
kwargs = {
"finetuned_from": model_args.model_name_or_path,
"tasks": "audio-classification",
"dataset": data_args.dataset_name,
"tags": ["audio-classification"],
}
if training_args.push_to_hub:
trainer.push_to_hub(**kwargs)
else:
trainer.create_model_card(**kwargs)
if __name__ == "__main__":
main()
|
transformers/examples/pytorch/audio-classification/run_audio_classification.py/0
|
{
"file_path": "transformers/examples/pytorch/audio-classification/run_audio_classification.py",
"repo_id": "transformers",
"token_count": 6998
}
| 315
|
#!/usr/bin/env python
# coding=utf-8
# Copyright 2024 The HuggingFace Inc. team. 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
"""Finetuning 🤗 Transformers model for instance segmentation leveraging the Trainer API."""
import logging
import os
import sys
from dataclasses import dataclass, field
from functools import partial
from typing import Any, Dict, List, Mapping, Optional
import albumentations as A
import numpy as np
import torch
from datasets import load_dataset
from torchmetrics.detection.mean_ap import MeanAveragePrecision
import transformers
from transformers import (
AutoImageProcessor,
AutoModelForUniversalSegmentation,
HfArgumentParser,
Trainer,
TrainingArguments,
)
from transformers.image_processing_utils import BatchFeature
from transformers.trainer import EvalPrediction
from transformers.trainer_utils import get_last_checkpoint
from transformers.utils import check_min_version, send_example_telemetry
from transformers.utils.versions import require_version
logger = logging.getLogger(__name__)
# Will error if the minimal version of Transformers is not installed. Remove at your own risks.
check_min_version("4.45.0.dev0")
require_version("datasets>=2.0.0", "To fix: pip install -r examples/pytorch/instance-segmentation/requirements.txt")
@dataclass
class Arguments:
"""
Arguments pertaining to what data we are going to input our model for training and eval.
Using `HfArgumentParser` we can turn this class into argparse arguments to be able to specify
them on the command line.
"""
model_name_or_path: str = field(
default="facebook/mask2former-swin-tiny-coco-instance",
metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"},
)
dataset_name: str = field(
default="qubvel-hf/ade20k-mini",
metadata={
"help": "Name of a dataset from the hub (could be your own, possibly private dataset hosted on the hub)."
},
)
trust_remote_code: bool = field(
default=False,
metadata={
"help": (
"Whether to trust the execution of code from datasets/models defined on the Hub."
" This option should only be set to `True` for repositories you trust and in which you have read the"
" code, as it will execute code present on the Hub on your local machine."
)
},
)
image_height: Optional[int] = field(default=512, metadata={"help": "Image height after resizing."})
image_width: Optional[int] = field(default=512, metadata={"help": "Image width after resizing."})
token: str = field(
default=None,
metadata={
"help": (
"The token to use as HTTP bearer authorization for remote files. If not specified, will use the token "
"generated when running `huggingface-cli login` (stored in `~/.huggingface`)."
)
},
)
do_reduce_labels: bool = field(
default=False,
metadata={
"help": (
"If background class is labeled as 0 and you want to remove it from the labels, set this flag to True."
)
},
)
def augment_and_transform_batch(
examples: Mapping[str, Any], transform: A.Compose, image_processor: AutoImageProcessor
) -> BatchFeature:
batch = {
"pixel_values": [],
"mask_labels": [],
"class_labels": [],
}
for pil_image, pil_annotation in zip(examples["image"], examples["annotation"]):
image = np.array(pil_image)
semantic_and_instance_masks = np.array(pil_annotation)[..., :2]
# Apply augmentations
output = transform(image=image, mask=semantic_and_instance_masks)
aug_image = output["image"]
aug_semantic_and_instance_masks = output["mask"]
aug_instance_mask = aug_semantic_and_instance_masks[..., 1]
# Create mapping from instance id to semantic id
unique_semantic_id_instance_id_pairs = np.unique(aug_semantic_and_instance_masks.reshape(-1, 2), axis=0)
instance_id_to_semantic_id = {
instance_id: semantic_id for semantic_id, instance_id in unique_semantic_id_instance_id_pairs
}
# Apply the image processor transformations: resizing, rescaling, normalization
model_inputs = image_processor(
images=[aug_image],
segmentation_maps=[aug_instance_mask],
instance_id_to_semantic_id=instance_id_to_semantic_id,
return_tensors="pt",
)
batch["pixel_values"].append(model_inputs.pixel_values[0])
batch["mask_labels"].append(model_inputs.mask_labels[0])
batch["class_labels"].append(model_inputs.class_labels[0])
return batch
def collate_fn(examples):
batch = {}
batch["pixel_values"] = torch.stack([example["pixel_values"] for example in examples])
batch["class_labels"] = [example["class_labels"] for example in examples]
batch["mask_labels"] = [example["mask_labels"] for example in examples]
if "pixel_mask" in examples[0]:
batch["pixel_mask"] = torch.stack([example["pixel_mask"] for example in examples])
return batch
@dataclass
class ModelOutput:
class_queries_logits: torch.Tensor
masks_queries_logits: torch.Tensor
def nested_cpu(tensors):
if isinstance(tensors, (list, tuple)):
return type(tensors)(nested_cpu(t) for t in tensors)
elif isinstance(tensors, Mapping):
return type(tensors)({k: nested_cpu(t) for k, t in tensors.items()})
elif isinstance(tensors, torch.Tensor):
return tensors.cpu().detach()
else:
return tensors
class Evaluator:
"""
Compute metrics for the instance segmentation task.
"""
def __init__(
self,
image_processor: AutoImageProcessor,
id2label: Mapping[int, str],
threshold: float = 0.0,
):
"""
Initialize evaluator with image processor, id2label mapping and threshold for filtering predictions.
Args:
image_processor (AutoImageProcessor): Image processor for
`post_process_instance_segmentation` method.
id2label (Mapping[int, str]): Mapping from class id to class name.
threshold (float): Threshold to filter predicted boxes by confidence. Defaults to 0.0.
"""
self.image_processor = image_processor
self.id2label = id2label
self.threshold = threshold
self.metric = self.get_metric()
def get_metric(self):
metric = MeanAveragePrecision(iou_type="segm", class_metrics=True)
return metric
def reset_metric(self):
self.metric.reset()
def postprocess_target_batch(self, target_batch) -> List[Dict[str, torch.Tensor]]:
"""Collect targets in a form of list of dictionaries with keys "masks", "labels"."""
batch_masks = target_batch[0]
batch_labels = target_batch[1]
post_processed_targets = []
for masks, labels in zip(batch_masks, batch_labels):
post_processed_targets.append(
{
"masks": masks.to(dtype=torch.bool),
"labels": labels,
}
)
return post_processed_targets
def get_target_sizes(self, post_processed_targets) -> List[List[int]]:
target_sizes = []
for target in post_processed_targets:
target_sizes.append(target["masks"].shape[-2:])
return target_sizes
def postprocess_prediction_batch(self, prediction_batch, target_sizes) -> List[Dict[str, torch.Tensor]]:
"""Collect predictions in a form of list of dictionaries with keys "masks", "labels", "scores"."""
model_output = ModelOutput(class_queries_logits=prediction_batch[0], masks_queries_logits=prediction_batch[1])
post_processed_output = self.image_processor.post_process_instance_segmentation(
model_output,
threshold=self.threshold,
target_sizes=target_sizes,
return_binary_maps=True,
)
post_processed_predictions = []
for image_predictions, target_size in zip(post_processed_output, target_sizes):
if image_predictions["segments_info"]:
post_processed_image_prediction = {
"masks": image_predictions["segmentation"].to(dtype=torch.bool),
"labels": torch.tensor([x["label_id"] for x in image_predictions["segments_info"]]),
"scores": torch.tensor([x["score"] for x in image_predictions["segments_info"]]),
}
else:
# for void predictions, we need to provide empty tensors
post_processed_image_prediction = {
"masks": torch.zeros([0, *target_size], dtype=torch.bool),
"labels": torch.tensor([]),
"scores": torch.tensor([]),
}
post_processed_predictions.append(post_processed_image_prediction)
return post_processed_predictions
@torch.no_grad()
def __call__(self, evaluation_results: EvalPrediction, compute_result: bool = False) -> Mapping[str, float]:
"""
Update metrics with current evaluation results and return metrics if `compute_result` is True.
Args:
evaluation_results (EvalPrediction): Predictions and targets from evaluation.
compute_result (bool): Whether to compute and return metrics.
Returns:
Mapping[str, float]: Metrics in a form of dictionary {<metric_name>: <metric_value>}
"""
prediction_batch = nested_cpu(evaluation_results.predictions)
target_batch = nested_cpu(evaluation_results.label_ids)
# For metric computation we need to provide:
# - targets in a form of list of dictionaries with keys "masks", "labels"
# - predictions in a form of list of dictionaries with keys "masks", "labels", "scores"
post_processed_targets = self.postprocess_target_batch(target_batch)
target_sizes = self.get_target_sizes(post_processed_targets)
post_processed_predictions = self.postprocess_prediction_batch(prediction_batch, target_sizes)
# Compute metrics
self.metric.update(post_processed_predictions, post_processed_targets)
if not compute_result:
return
metrics = self.metric.compute()
# Replace list of per class metrics with separate metric for each class
classes = metrics.pop("classes")
map_per_class = metrics.pop("map_per_class")
mar_100_per_class = metrics.pop("mar_100_per_class")
for class_id, class_map, class_mar in zip(classes, map_per_class, mar_100_per_class):
class_name = self.id2label[class_id.item()] if self.id2label is not None else class_id.item()
metrics[f"map_{class_name}"] = class_map
metrics[f"mar_100_{class_name}"] = class_mar
metrics = {k: round(v.item(), 4) for k, v in metrics.items()}
# Reset metric for next evaluation
self.reset_metric()
return metrics
def setup_logging(training_args: TrainingArguments) -> None:
"""Setup logging according to `training_args`."""
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
handlers=[logging.StreamHandler(sys.stdout)],
)
if training_args.should_log:
# The default of training_args.log_level is passive, so we set log level at info here to have that default.
transformers.utils.logging.set_verbosity_info()
log_level = training_args.get_process_log_level()
logger.setLevel(log_level)
transformers.utils.logging.set_verbosity(log_level)
transformers.utils.logging.enable_default_handler()
transformers.utils.logging.enable_explicit_format()
def find_last_checkpoint(training_args: TrainingArguments) -> Optional[str]:
"""Find the last checkpoint in the output directory according to parameters specified in `training_args`."""
checkpoint = None
if training_args.resume_from_checkpoint is not None:
checkpoint = training_args.resume_from_checkpoint
elif os.path.isdir(training_args.output_dir) and not training_args.overwrite_output_dir:
checkpoint = get_last_checkpoint(training_args.output_dir)
if checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
"Use --overwrite_output_dir to overcome."
)
elif checkpoint is not None and training_args.resume_from_checkpoint is None:
logger.info(
f"Checkpoint detected, resuming training at {checkpoint}. To avoid this behavior, change "
"the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
)
return checkpoint
def main():
# See all possible arguments in https://huggingface.co/docs/transformers/main_classes/trainer#transformers.TrainingArguments
# or by passing the --help flag to this script.
parser = HfArgumentParser([Arguments, TrainingArguments])
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# If we pass only one argument to the script and it's the path to a json file,
# let's parse it to get our arguments.
args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
else:
args, training_args = parser.parse_args_into_dataclasses()
# Set default training arguments for instance segmentation
training_args.eval_do_concat_batches = False
training_args.batch_eval_metrics = True
training_args.remove_unused_columns = False
# # Sending telemetry. Tracking the example usage helps us better allocate resources to maintain them. The
# # information sent is the one passed as arguments along with your Python/PyTorch versions.
send_example_telemetry("run_instance_segmentation", args)
# Setup logging and log on each process the small summary:
setup_logging(training_args)
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}, "
+ f"distributed training: {training_args.parallel_mode.value == 'distributed'}, 16-bits training: {training_args.fp16}"
)
logger.info(f"Training/evaluation parameters {training_args}")
# Load last checkpoint from output_dir if it exists (and we are not overwriting it)
checkpoint = find_last_checkpoint(training_args)
# ------------------------------------------------------------------------------------------------
# Load dataset, prepare splits
# ------------------------------------------------------------------------------------------------
dataset = load_dataset(args.dataset_name, trust_remote_code=args.trust_remote_code)
# We need to specify the label2id mapping for the model
# it is a mapping from semantic class name to class index.
# In case your dataset does not provide it, you can create it manually:
# label2id = {"background": 0, "cat": 1, "dog": 2}
label2id = dataset["train"][0]["semantic_class_to_id"]
if args.do_reduce_labels:
label2id = {name: idx for name, idx in label2id.items() if idx != 0} # remove background class
label2id = {name: idx - 1 for name, idx in label2id.items()} # shift class indices by -1
id2label = {v: k for k, v in label2id.items()}
# ------------------------------------------------------------------------------------------------
# Load pretrained config, model and image processor
# ------------------------------------------------------------------------------------------------
model = AutoModelForUniversalSegmentation.from_pretrained(
args.model_name_or_path,
label2id=label2id,
id2label=id2label,
ignore_mismatched_sizes=True,
token=args.token,
)
image_processor = AutoImageProcessor.from_pretrained(
args.model_name_or_path,
do_resize=True,
size={"height": args.image_height, "width": args.image_width},
do_reduce_labels=args.do_reduce_labels,
reduce_labels=args.do_reduce_labels, # TODO: remove when mask2former support `do_reduce_labels`
token=args.token,
)
# ------------------------------------------------------------------------------------------------
# Define image augmentations and dataset transforms
# ------------------------------------------------------------------------------------------------
train_augment_and_transform = A.Compose(
[
A.HorizontalFlip(p=0.5),
A.RandomBrightnessContrast(p=0.5),
A.HueSaturationValue(p=0.1),
],
)
validation_transform = A.Compose(
[A.NoOp()],
)
# Make transform functions for batch and apply for dataset splits
train_transform_batch = partial(
augment_and_transform_batch, transform=train_augment_and_transform, image_processor=image_processor
)
validation_transform_batch = partial(
augment_and_transform_batch, transform=validation_transform, image_processor=image_processor
)
dataset["train"] = dataset["train"].with_transform(train_transform_batch)
dataset["validation"] = dataset["validation"].with_transform(validation_transform_batch)
# ------------------------------------------------------------------------------------------------
# Model training and evaluation with Trainer API
# ------------------------------------------------------------------------------------------------
compute_metrics = Evaluator(image_processor=image_processor, id2label=id2label, threshold=0.0)
trainer = Trainer(
model=model,
args=training_args,
train_dataset=dataset["train"] if training_args.do_train else None,
eval_dataset=dataset["validation"] if training_args.do_eval else None,
tokenizer=image_processor,
data_collator=collate_fn,
compute_metrics=compute_metrics,
)
# Training
if training_args.do_train:
train_result = trainer.train(resume_from_checkpoint=checkpoint)
trainer.save_model()
trainer.log_metrics("train", train_result.metrics)
trainer.save_metrics("train", train_result.metrics)
trainer.save_state()
# Final evaluation
if training_args.do_eval:
metrics = trainer.evaluate(eval_dataset=dataset["validation"], metric_key_prefix="test")
trainer.log_metrics("test", metrics)
trainer.save_metrics("test", metrics)
# Write model card and (optionally) push to hub
kwargs = {
"finetuned_from": args.model_name_or_path,
"dataset": args.dataset_name,
"tags": ["image-segmentation", "instance-segmentation", "vision"],
}
if training_args.push_to_hub:
trainer.push_to_hub(**kwargs)
else:
trainer.create_model_card(**kwargs)
if __name__ == "__main__":
main()
|
transformers/examples/pytorch/instance-segmentation/run_instance_segmentation.py/0
|
{
"file_path": "transformers/examples/pytorch/instance-segmentation/run_instance_segmentation.py",
"repo_id": "transformers",
"token_count": 7484
}
| 316
|
<!---
Copyright 2024 The HuggingFace Team. 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.
-->
# Object detection examples
This directory contains 2 scripts that showcase how to fine-tune any model supported by the [`AutoModelForObjectDetection` API](https://huggingface.co/docs/transformers/main/en/model_doc/auto#transformers.AutoModelForObjectDetection) (such as [DETR](https://huggingface.co/docs/transformers/main/en/model_doc/detr), [DETA](https://huggingface.co/docs/transformers/main/en/model_doc/deta), [Deformable DETR](https://huggingface.co/docs/transformers/main/en/model_doc/deformable_detr)) using PyTorch.
Content:
* [PyTorch version, Trainer](#pytorch-version-trainer)
* [PyTorch version, no Trainer](#pytorch-version-no-trainer)
* [Reload and perform inference](#reload-and-perform-inference)
* [Note on custom data](#note-on-custom-data)
## PyTorch version, Trainer
Based on the script [`run_object_detection.py`](https://github.com/huggingface/transformers/blob/main/examples/pytorch/object-detection/run_object_detection.py).
The script leverages the [🤗 Trainer API](https://huggingface.co/docs/transformers/main_classes/trainer) to automatically take care of the training for you, running on distributed environments right away.
Here we show how to fine-tune a [DETR](https://huggingface.co/facebook/detr-resnet-50) model on the [CPPE-5](https://huggingface.co/datasets/cppe-5) dataset:
```bash
python run_object_detection.py \
--model_name_or_path facebook/detr-resnet-50 \
--dataset_name cppe-5 \
--do_train true \
--do_eval true \
--output_dir detr-finetuned-cppe-5-10k-steps \
--num_train_epochs 100 \
--image_square_size 600 \
--fp16 true \
--learning_rate 5e-5 \
--weight_decay 1e-4 \
--dataloader_num_workers 4 \
--dataloader_prefetch_factor 2 \
--per_device_train_batch_size 8 \
--gradient_accumulation_steps 1 \
--remove_unused_columns false \
--eval_do_concat_batches false \
--ignore_mismatched_sizes true \
--metric_for_best_model eval_map \
--greater_is_better true \
--load_best_model_at_end true \
--logging_strategy epoch \
--evaluation_strategy epoch \
--save_strategy epoch \
--save_total_limit 2 \
--push_to_hub true \
--push_to_hub_model_id detr-finetuned-cppe-5-10k-steps \
--hub_strategy end \
--seed 1337
```
> Note:
`--eval_do_concat_batches false` is required for correct evaluation of detection models;
`--ignore_mismatched_sizes true` is required to load detection model for finetuning with different number of classes.
The resulting model can be seen here: https://huggingface.co/qubvel-hf/qubvel-hf/detr-resnet-50-finetuned-10k-cppe5. The corresponding Weights and Biases report [here](https://api.wandb.ai/links/qubvel-hf-co/bnm0r5ex). Note that it's always advised to check the original paper to know the details regarding training hyperparameters. Hyperparameters for current example were not tuned. To improve model quality you could try:
- changing image size parameters (`--shortest_edge`/`--longest_edge`)
- changing training parameters, such as learning rate, batch size, warmup, optimizer and many more (see [TrainingArguments](https://huggingface.co/docs/transformers/main_classes/trainer#transformers.TrainingArguments))
- adding more image augmentations (we created a helpful [HF Space](https://huggingface.co/spaces/qubvel-hf/albumentations-demo) to choose some)
Note that you can replace the model and dataset by simply setting the `model_name_or_path` and `dataset_name` arguments respectively, with model or dataset from the [hub](https://huggingface.co/).
For dataset, make sure it provides labels in the same format as [CPPE-5](https://huggingface.co/datasets/cppe-5) dataset and boxes are provided in [COCO format](https://albumentations.ai/docs/getting_started/bounding_boxes_augmentation/#coco).
## PyTorch version, no Trainer
Based on the script [`run_object_detection_no_trainer.py`](https://github.com/huggingface/transformers/blob/main/examples/pytorch/object-detection/run_object_detection.py).
The script leverages [🤗 `Accelerate`](https://github.com/huggingface/accelerate), which allows to write your own training loop in PyTorch, but have it run instantly on any (distributed) environment, including CPU, multi-CPU, GPU, multi-GPU and TPU. It also supports mixed precision.
First, run:
```bash
accelerate config
```
and reply to the questions asked regarding the environment on which you'd like to train. Then
```bash
accelerate test
```
that will check everything is ready for training. Finally, you can launch training with
```bash
accelerate launch run_object_detection_no_trainer.py \
--model_name_or_path "facebook/detr-resnet-50" \
--dataset_name cppe-5 \
--output_dir "detr-resnet-50-finetuned" \
--num_train_epochs 100 \
--image_square_size 600 \
--per_device_train_batch_size 8 \
--per_device_eval_batch_size 8 \
--checkpointing_steps epoch \
--learning_rate 5e-5 \
--ignore_mismatched_sizes \
--with_tracking \
--push_to_hub
```
and boom, you're training, possibly on multiple GPUs, logging everything to all trackers found in your environment (like Weights and Biases, Tensorboard) and regularly pushing your model to the hub (with the repo name being equal to `args.output_dir` at your HF username) 🤗
With the default settings, the script fine-tunes a [DETR](https://huggingface.co/facebook/detr-resnet-50) model on the [CPPE-5](https://huggingface.co/datasets/cppe-5) dataset. The resulting model can be seen here: https://huggingface.co/qubvel-hf/detr-resnet-50-finetuned-10k-cppe5-no-trainer.
## Reload and perform inference
This means that after training, you can easily load your trained model and perform inference as follows::
```python
import requests
import torch
from PIL import Image
from transformers import AutoImageProcessor, AutoModelForObjectDetection
# Name of repo on the hub or path to a local folder
model_name = "qubvel-hf/detr-resnet-50-finetuned-10k-cppe5"
image_processor = AutoImageProcessor.from_pretrained(model_name)
model = AutoModelForObjectDetection.from_pretrained(model_name)
# Load image for inference
url = "https://images.pexels.com/photos/8413299/pexels-photo-8413299.jpeg?auto=compress&cs=tinysrgb&w=630&h=375&dpr=2"
image = Image.open(requests.get(url, stream=True).raw)
# Prepare image for the model
inputs = image_processor(images=image, return_tensors="pt")
with torch.no_grad():
outputs = model(**inputs)
# Post process model predictions
# this include conversion to Pascal VOC format and filtering non confident boxes
width, height = image.size
target_sizes = torch.tensor([height, width]).unsqueeze(0) # add batch dim
results = image_processor.post_process_object_detection(outputs, threshold=0.5, target_sizes=target_sizes)[0]
for score, label, box in zip(results["scores"], results["labels"], results["boxes"]):
box = [round(i, 2) for i in box.tolist()]
print(
f"Detected {model.config.id2label[label.item()]} with confidence "
f"{round(score.item(), 3)} at location {box}"
)
```
And visualize with the following code:
```python
from PIL import ImageDraw
draw = ImageDraw.Draw(image)
for score, label, box in zip(results["scores"], results["labels"], results["boxes"]):
box = [round(i, 2) for i in box.tolist()]
x, y, x2, y2 = tuple(box)
draw.rectangle((x, y, x2, y2), outline="red", width=1)
draw.text((x, y), model.config.id2label[label.item()], fill="white")
image
```
## Note on custom data
In case you'd like to use the script with custom data, you could prepare your data with the following way:
```bash
custom_dataset/
└── train
├── 0001.jpg
├── 0002.jpg
├── ...
└── metadata.jsonl
└── validation
└── ...
└── test
└── ...
```
Where `metadata.jsonl` is a file with the following structure:
```json
{"file_name": "0001.jpg", "objects": {"bbox": [[302.0, 109.0, 73.0, 52.0]], "categories": [0], "id": [1], "area": [50.0]}}
{"file_name": "0002.jpg", "objects": {"bbox": [[810.0, 100.0, 57.0, 28.0]], "categories": [1], "id": [2], "area": [40.0]}}
...
```
Trining script support bounding boxes in COCO format (x_min, y_min, width, height).
Then, you cat load the dataset with just a few lines of code:
```python
from datasets import load_dataset
# Load dataset
dataset = load_dataset("imagefolder", data_dir="custom_dataset/")
# >>> DatasetDict({
# ... train: Dataset({
# ... features: ['image', 'objects'],
# ... num_rows: 2
# ... })
# ... })
# Push to hub (assumes you have ran the huggingface-cli login command in a terminal/notebook)
dataset.push_to_hub("name of repo on the hub")
# optionally, you can push to a private repo on the hub
# dataset.push_to_hub("name of repo on the hub", private=True)
```
And the final step, for training you should provide id2label mapping in the following way:
```python
id2label = {0: "Car", 1: "Bird", ...}
```
Just find it in code and replace for simplicity, or save `json` locally and with the dataset on the hub!
See also: [Dataset Creation Guide](https://huggingface.co/docs/datasets/image_dataset#create-an-image-dataset)
|
transformers/examples/pytorch/object-detection/README.md/0
|
{
"file_path": "transformers/examples/pytorch/object-detection/README.md",
"repo_id": "transformers",
"token_count": 3363
}
| 317
|
# coding=utf-8
# Copyright 2018 HuggingFace Inc..
#
# 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.
import json
import logging
import os
import sys
from unittest.mock import patch
from transformers import ViTMAEForPreTraining, Wav2Vec2ForPreTraining
from transformers.testing_utils import (
CaptureLogger,
TestCasePlus,
backend_device_count,
is_torch_fp16_available_on_device,
slow,
torch_device,
)
SRC_DIRS = [
os.path.join(os.path.dirname(__file__), dirname)
for dirname in [
"text-generation",
"text-classification",
"token-classification",
"language-modeling",
"multiple-choice",
"question-answering",
"summarization",
"translation",
"image-classification",
"speech-recognition",
"audio-classification",
"speech-pretraining",
"image-pretraining",
"semantic-segmentation",
"object-detection",
"instance-segmentation",
]
]
sys.path.extend(SRC_DIRS)
if SRC_DIRS is not None:
import run_audio_classification
import run_clm
import run_generation
import run_glue
import run_image_classification
import run_instance_segmentation
import run_mae
import run_mlm
import run_ner
import run_object_detection
import run_qa as run_squad
import run_semantic_segmentation
import run_seq2seq_qa as run_squad_seq2seq
import run_speech_recognition_ctc
import run_speech_recognition_ctc_adapter
import run_speech_recognition_seq2seq
import run_summarization
import run_swag
import run_translation
import run_wav2vec2_pretraining_no_trainer
logging.basicConfig(level=logging.DEBUG)
logger = logging.getLogger()
def get_results(output_dir):
results = {}
path = os.path.join(output_dir, "all_results.json")
if os.path.exists(path):
with open(path, "r") as f:
results = json.load(f)
else:
raise ValueError(f"can't find {path}")
return results
stream_handler = logging.StreamHandler(sys.stdout)
logger.addHandler(stream_handler)
class ExamplesTests(TestCasePlus):
def test_run_glue(self):
tmp_dir = self.get_auto_remove_tmp_dir()
testargs = f"""
run_glue.py
--model_name_or_path distilbert/distilbert-base-uncased
--output_dir {tmp_dir}
--overwrite_output_dir
--train_file ./tests/fixtures/tests_samples/MRPC/train.csv
--validation_file ./tests/fixtures/tests_samples/MRPC/dev.csv
--do_train
--do_eval
--per_device_train_batch_size=2
--per_device_eval_batch_size=1
--learning_rate=1e-4
--max_steps=10
--warmup_steps=2
--seed=42
--max_seq_length=128
""".split()
if is_torch_fp16_available_on_device(torch_device):
testargs.append("--fp16")
with patch.object(sys, "argv", testargs):
run_glue.main()
result = get_results(tmp_dir)
self.assertGreaterEqual(result["eval_accuracy"], 0.75)
def test_run_clm(self):
tmp_dir = self.get_auto_remove_tmp_dir()
testargs = f"""
run_clm.py
--model_name_or_path distilbert/distilgpt2
--train_file ./tests/fixtures/sample_text.txt
--validation_file ./tests/fixtures/sample_text.txt
--do_train
--do_eval
--block_size 128
--per_device_train_batch_size 5
--per_device_eval_batch_size 5
--num_train_epochs 2
--output_dir {tmp_dir}
--overwrite_output_dir
""".split()
if backend_device_count(torch_device) > 1:
# Skipping because there are not enough batches to train the model + would need a drop_last to work.
return
if torch_device == "cpu":
testargs.append("--use_cpu")
with patch.object(sys, "argv", testargs):
run_clm.main()
result = get_results(tmp_dir)
self.assertLess(result["perplexity"], 100)
def test_run_clm_config_overrides(self):
# test that config_overrides works, despite the misleading dumps of default un-updated
# config via tokenizer
tmp_dir = self.get_auto_remove_tmp_dir()
testargs = f"""
run_clm.py
--model_type gpt2
--tokenizer_name openai-community/gpt2
--train_file ./tests/fixtures/sample_text.txt
--output_dir {tmp_dir}
--config_overrides n_embd=10,n_head=2
""".split()
if torch_device == "cpu":
testargs.append("--use_cpu")
logger = run_clm.logger
with patch.object(sys, "argv", testargs):
with CaptureLogger(logger) as cl:
run_clm.main()
self.assertIn('"n_embd": 10', cl.out)
self.assertIn('"n_head": 2', cl.out)
def test_run_mlm(self):
tmp_dir = self.get_auto_remove_tmp_dir()
testargs = f"""
run_mlm.py
--model_name_or_path distilbert/distilroberta-base
--train_file ./tests/fixtures/sample_text.txt
--validation_file ./tests/fixtures/sample_text.txt
--output_dir {tmp_dir}
--overwrite_output_dir
--do_train
--do_eval
--prediction_loss_only
--num_train_epochs=1
""".split()
if torch_device == "cpu":
testargs.append("--use_cpu")
with patch.object(sys, "argv", testargs):
run_mlm.main()
result = get_results(tmp_dir)
self.assertLess(result["perplexity"], 42)
def test_run_ner(self):
# with so little data distributed training needs more epochs to get the score on par with 0/1 gpu
epochs = 7 if backend_device_count(torch_device) > 1 else 2
tmp_dir = self.get_auto_remove_tmp_dir()
testargs = f"""
run_ner.py
--model_name_or_path google-bert/bert-base-uncased
--train_file tests/fixtures/tests_samples/conll/sample.json
--validation_file tests/fixtures/tests_samples/conll/sample.json
--output_dir {tmp_dir}
--overwrite_output_dir
--do_train
--do_eval
--warmup_steps=2
--learning_rate=2e-4
--per_device_train_batch_size=2
--per_device_eval_batch_size=2
--num_train_epochs={epochs}
--seed 7
""".split()
if torch_device == "cpu":
testargs.append("--use_cpu")
with patch.object(sys, "argv", testargs):
run_ner.main()
result = get_results(tmp_dir)
self.assertGreaterEqual(result["eval_accuracy"], 0.75)
self.assertLess(result["eval_loss"], 0.5)
def test_run_squad(self):
tmp_dir = self.get_auto_remove_tmp_dir()
testargs = f"""
run_qa.py
--model_name_or_path google-bert/bert-base-uncased
--version_2_with_negative
--train_file tests/fixtures/tests_samples/SQUAD/sample.json
--validation_file tests/fixtures/tests_samples/SQUAD/sample.json
--output_dir {tmp_dir}
--overwrite_output_dir
--max_steps=10
--warmup_steps=2
--do_train
--do_eval
--learning_rate=2e-4
--per_device_train_batch_size=2
--per_device_eval_batch_size=1
""".split()
with patch.object(sys, "argv", testargs):
run_squad.main()
result = get_results(tmp_dir)
self.assertGreaterEqual(result["eval_f1"], 30)
self.assertGreaterEqual(result["eval_exact"], 30)
def test_run_squad_seq2seq(self):
tmp_dir = self.get_auto_remove_tmp_dir()
testargs = f"""
run_seq2seq_qa.py
--model_name_or_path google-t5/t5-small
--context_column context
--question_column question
--answer_column answers
--version_2_with_negative
--train_file tests/fixtures/tests_samples/SQUAD/sample.json
--validation_file tests/fixtures/tests_samples/SQUAD/sample.json
--output_dir {tmp_dir}
--overwrite_output_dir
--max_steps=10
--warmup_steps=2
--do_train
--do_eval
--learning_rate=2e-4
--per_device_train_batch_size=2
--per_device_eval_batch_size=1
--predict_with_generate
""".split()
with patch.object(sys, "argv", testargs):
run_squad_seq2seq.main()
result = get_results(tmp_dir)
self.assertGreaterEqual(result["eval_f1"], 30)
self.assertGreaterEqual(result["eval_exact"], 30)
def test_run_swag(self):
tmp_dir = self.get_auto_remove_tmp_dir()
testargs = f"""
run_swag.py
--model_name_or_path google-bert/bert-base-uncased
--train_file tests/fixtures/tests_samples/swag/sample.json
--validation_file tests/fixtures/tests_samples/swag/sample.json
--output_dir {tmp_dir}
--overwrite_output_dir
--max_steps=20
--warmup_steps=2
--do_train
--do_eval
--learning_rate=2e-4
--per_device_train_batch_size=2
--per_device_eval_batch_size=1
""".split()
with patch.object(sys, "argv", testargs):
run_swag.main()
result = get_results(tmp_dir)
self.assertGreaterEqual(result["eval_accuracy"], 0.8)
def test_generation(self):
testargs = ["run_generation.py", "--prompt=Hello", "--length=10", "--seed=42"]
if is_torch_fp16_available_on_device(torch_device):
testargs.append("--fp16")
model_type, model_name = (
"--model_type=gpt2",
"--model_name_or_path=sshleifer/tiny-gpt2",
)
with patch.object(sys, "argv", testargs + [model_type, model_name]):
result = run_generation.main()
self.assertGreaterEqual(len(result[0]), 10)
@slow
def test_run_summarization(self):
tmp_dir = self.get_auto_remove_tmp_dir()
testargs = f"""
run_summarization.py
--model_name_or_path google-t5/t5-small
--train_file tests/fixtures/tests_samples/xsum/sample.json
--validation_file tests/fixtures/tests_samples/xsum/sample.json
--output_dir {tmp_dir}
--overwrite_output_dir
--max_steps=50
--warmup_steps=8
--do_train
--do_eval
--learning_rate=2e-4
--per_device_train_batch_size=2
--per_device_eval_batch_size=1
--predict_with_generate
""".split()
with patch.object(sys, "argv", testargs):
run_summarization.main()
result = get_results(tmp_dir)
self.assertGreaterEqual(result["eval_rouge1"], 10)
self.assertGreaterEqual(result["eval_rouge2"], 2)
self.assertGreaterEqual(result["eval_rougeL"], 7)
self.assertGreaterEqual(result["eval_rougeLsum"], 7)
@slow
def test_run_translation(self):
tmp_dir = self.get_auto_remove_tmp_dir()
testargs = f"""
run_translation.py
--model_name_or_path sshleifer/student_marian_en_ro_6_1
--source_lang en
--target_lang ro
--train_file tests/fixtures/tests_samples/wmt16/sample.json
--validation_file tests/fixtures/tests_samples/wmt16/sample.json
--output_dir {tmp_dir}
--overwrite_output_dir
--max_steps=50
--warmup_steps=8
--do_train
--do_eval
--learning_rate=3e-3
--per_device_train_batch_size=2
--per_device_eval_batch_size=1
--predict_with_generate
--source_lang en_XX
--target_lang ro_RO
--max_source_length 512
""".split()
with patch.object(sys, "argv", testargs):
run_translation.main()
result = get_results(tmp_dir)
self.assertGreaterEqual(result["eval_bleu"], 30)
def test_run_image_classification(self):
tmp_dir = self.get_auto_remove_tmp_dir()
testargs = f"""
run_image_classification.py
--output_dir {tmp_dir}
--model_name_or_path google/vit-base-patch16-224-in21k
--dataset_name hf-internal-testing/cats_vs_dogs_sample
--trust_remote_code
--do_train
--do_eval
--learning_rate 1e-4
--per_device_train_batch_size 2
--per_device_eval_batch_size 1
--remove_unused_columns False
--overwrite_output_dir True
--dataloader_num_workers 16
--metric_for_best_model accuracy
--max_steps 10
--train_val_split 0.1
--seed 42
--label_column_name labels
""".split()
if is_torch_fp16_available_on_device(torch_device):
testargs.append("--fp16")
with patch.object(sys, "argv", testargs):
run_image_classification.main()
result = get_results(tmp_dir)
self.assertGreaterEqual(result["eval_accuracy"], 0.8)
def test_run_speech_recognition_ctc(self):
tmp_dir = self.get_auto_remove_tmp_dir()
testargs = f"""
run_speech_recognition_ctc.py
--output_dir {tmp_dir}
--model_name_or_path hf-internal-testing/tiny-random-wav2vec2
--dataset_name hf-internal-testing/librispeech_asr_dummy
--dataset_config_name clean
--train_split_name validation
--eval_split_name validation
--trust_remote_code
--do_train
--do_eval
--learning_rate 1e-4
--per_device_train_batch_size 2
--per_device_eval_batch_size 1
--remove_unused_columns False
--overwrite_output_dir True
--preprocessing_num_workers 16
--max_steps 10
--seed 42
""".split()
if is_torch_fp16_available_on_device(torch_device):
testargs.append("--fp16")
with patch.object(sys, "argv", testargs):
run_speech_recognition_ctc.main()
result = get_results(tmp_dir)
self.assertLess(result["eval_loss"], result["train_loss"])
def test_run_speech_recognition_ctc_adapter(self):
tmp_dir = self.get_auto_remove_tmp_dir()
testargs = f"""
run_speech_recognition_ctc_adapter.py
--output_dir {tmp_dir}
--model_name_or_path hf-internal-testing/tiny-random-wav2vec2
--dataset_name hf-internal-testing/librispeech_asr_dummy
--dataset_config_name clean
--train_split_name validation
--eval_split_name validation
--trust_remote_code
--do_train
--do_eval
--learning_rate 1e-4
--per_device_train_batch_size 2
--per_device_eval_batch_size 1
--remove_unused_columns False
--overwrite_output_dir True
--preprocessing_num_workers 16
--max_steps 10
--target_language tur
--seed 42
""".split()
if is_torch_fp16_available_on_device(torch_device):
testargs.append("--fp16")
with patch.object(sys, "argv", testargs):
run_speech_recognition_ctc_adapter.main()
result = get_results(tmp_dir)
self.assertTrue(os.path.isfile(os.path.join(tmp_dir, "./adapter.tur.safetensors")))
self.assertLess(result["eval_loss"], result["train_loss"])
def test_run_speech_recognition_seq2seq(self):
tmp_dir = self.get_auto_remove_tmp_dir()
testargs = f"""
run_speech_recognition_seq2seq.py
--output_dir {tmp_dir}
--model_name_or_path hf-internal-testing/tiny-random-speech-encoder-decoder
--dataset_name hf-internal-testing/librispeech_asr_dummy
--dataset_config_name clean
--train_split_name validation
--eval_split_name validation
--trust_remote_code
--do_train
--do_eval
--learning_rate 1e-4
--per_device_train_batch_size 2
--per_device_eval_batch_size 4
--remove_unused_columns False
--overwrite_output_dir True
--preprocessing_num_workers 16
--max_steps 10
--seed 42
""".split()
if is_torch_fp16_available_on_device(torch_device):
testargs.append("--fp16")
with patch.object(sys, "argv", testargs):
run_speech_recognition_seq2seq.main()
result = get_results(tmp_dir)
self.assertLess(result["eval_loss"], result["train_loss"])
def test_run_audio_classification(self):
tmp_dir = self.get_auto_remove_tmp_dir()
testargs = f"""
run_audio_classification.py
--output_dir {tmp_dir}
--model_name_or_path hf-internal-testing/tiny-random-wav2vec2
--dataset_name anton-l/superb_demo
--trust_remote_code
--dataset_config_name ks
--train_split_name test
--eval_split_name test
--audio_column_name audio
--label_column_name label
--do_train
--do_eval
--learning_rate 1e-4
--per_device_train_batch_size 2
--per_device_eval_batch_size 1
--remove_unused_columns False
--overwrite_output_dir True
--num_train_epochs 10
--max_steps 50
--seed 42
""".split()
if is_torch_fp16_available_on_device(torch_device):
testargs.append("--fp16")
with patch.object(sys, "argv", testargs):
run_audio_classification.main()
result = get_results(tmp_dir)
self.assertLess(result["eval_loss"], result["train_loss"])
def test_run_wav2vec2_pretraining(self):
tmp_dir = self.get_auto_remove_tmp_dir()
testargs = f"""
run_wav2vec2_pretraining_no_trainer.py
--output_dir {tmp_dir}
--model_name_or_path hf-internal-testing/tiny-random-wav2vec2
--dataset_name hf-internal-testing/librispeech_asr_dummy
--dataset_config_names clean
--dataset_split_names validation
--trust_remote_code
--learning_rate 1e-4
--per_device_train_batch_size 4
--per_device_eval_batch_size 4
--preprocessing_num_workers 16
--max_train_steps 2
--validation_split_percentage 5
--seed 42
""".split()
with patch.object(sys, "argv", testargs):
run_wav2vec2_pretraining_no_trainer.main()
model = Wav2Vec2ForPreTraining.from_pretrained(tmp_dir)
self.assertIsNotNone(model)
def test_run_vit_mae_pretraining(self):
tmp_dir = self.get_auto_remove_tmp_dir()
testargs = f"""
run_mae.py
--output_dir {tmp_dir}
--dataset_name hf-internal-testing/cats_vs_dogs_sample
--trust_remote_code
--do_train
--do_eval
--learning_rate 1e-4
--per_device_train_batch_size 2
--per_device_eval_batch_size 1
--remove_unused_columns False
--overwrite_output_dir True
--dataloader_num_workers 16
--metric_for_best_model accuracy
--max_steps 10
--train_val_split 0.1
--seed 42
""".split()
if is_torch_fp16_available_on_device(torch_device):
testargs.append("--fp16")
with patch.object(sys, "argv", testargs):
run_mae.main()
model = ViTMAEForPreTraining.from_pretrained(tmp_dir)
self.assertIsNotNone(model)
def test_run_semantic_segmentation(self):
tmp_dir = self.get_auto_remove_tmp_dir()
testargs = f"""
run_semantic_segmentation.py
--output_dir {tmp_dir}
--dataset_name huggingface/semantic-segmentation-test-sample
--do_train
--do_eval
--remove_unused_columns False
--overwrite_output_dir True
--max_steps 10
--learning_rate=2e-4
--per_device_train_batch_size=2
--per_device_eval_batch_size=1
--seed 32
""".split()
if is_torch_fp16_available_on_device(torch_device):
testargs.append("--fp16")
with patch.object(sys, "argv", testargs):
run_semantic_segmentation.main()
result = get_results(tmp_dir)
self.assertGreaterEqual(result["eval_overall_accuracy"], 0.1)
@patch.dict(os.environ, {"WANDB_DISABLED": "true"})
def test_run_object_detection(self):
tmp_dir = self.get_auto_remove_tmp_dir()
testargs = f"""
run_object_detection.py
--model_name_or_path qubvel-hf/detr-resnet-50-finetuned-10k-cppe5
--output_dir {tmp_dir}
--dataset_name qubvel-hf/cppe-5-sample
--do_train
--do_eval
--remove_unused_columns False
--overwrite_output_dir True
--eval_do_concat_batches False
--max_steps 10
--learning_rate=1e-6
--per_device_train_batch_size=2
--per_device_eval_batch_size=1
--seed 32
""".split()
if is_torch_fp16_available_on_device(torch_device):
testargs.append("--fp16")
with patch.object(sys, "argv", testargs):
run_object_detection.main()
result = get_results(tmp_dir)
self.assertGreaterEqual(result["test_map"], 0.1)
@patch.dict(os.environ, {"WANDB_DISABLED": "true"})
def test_run_instance_segmentation(self):
tmp_dir = self.get_auto_remove_tmp_dir()
testargs = f"""
run_instance_segmentation.py
--model_name_or_path qubvel-hf/finetune-instance-segmentation-ade20k-mini-mask2former
--output_dir {tmp_dir}
--dataset_name qubvel-hf/ade20k-nano
--do_reduce_labels
--image_height 256
--image_width 256
--do_train
--num_train_epochs 1
--learning_rate 1e-5
--lr_scheduler_type constant
--per_device_train_batch_size 2
--per_device_eval_batch_size 1
--do_eval
--evaluation_strategy epoch
--seed 32
""".split()
if is_torch_fp16_available_on_device(torch_device):
testargs.append("--fp16")
with patch.object(sys, "argv", testargs):
run_instance_segmentation.main()
result = get_results(tmp_dir)
self.assertGreaterEqual(result["test_map"], 0.1)
|
transformers/examples/pytorch/test_pytorch_examples.py/0
|
{
"file_path": "transformers/examples/pytorch/test_pytorch_examples.py",
"repo_id": "transformers",
"token_count": 12107
}
| 318
|
# coding=utf-8
# Copyright 2020 The Google AI Language Team Authors, The HuggingFace Inc. team and Microsoft Corporation.
# 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.
"""Training and inference using the library models for sequence classification on GLUE (Bert, Albert) with PABEE."""
import argparse
import glob
import json
import logging
import os
import random
import numpy as np
import torch
from pabee.modeling_pabee_albert import AlbertForSequenceClassificationWithPabee
from pabee.modeling_pabee_bert import BertForSequenceClassificationWithPabee
from torch import nn
from torch.utils.data import DataLoader, RandomSampler, SequentialSampler, TensorDataset
from torch.utils.data.distributed import DistributedSampler
from tqdm import tqdm, trange
import transformers
from transformers import (
WEIGHTS_NAME,
AdamW,
AlbertConfig,
AlbertTokenizer,
BertConfig,
BertTokenizer,
get_linear_schedule_with_warmup,
)
from transformers import glue_compute_metrics as compute_metrics
from transformers import glue_convert_examples_to_features as convert_examples_to_features
from transformers import glue_output_modes as output_modes
from transformers import glue_processors as processors
from transformers.trainer_utils import is_main_process
try:
from torch.utils.tensorboard import SummaryWriter
except ImportError:
from tensorboardX import SummaryWriter
logger = logging.getLogger(__name__)
MODEL_CLASSES = {
"bert": (BertConfig, BertForSequenceClassificationWithPabee, BertTokenizer),
"albert": (AlbertConfig, AlbertForSequenceClassificationWithPabee, AlbertTokenizer),
}
def set_seed(args):
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if args.n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
def train(args, train_dataset, model, tokenizer):
"""Train the model"""
if args.local_rank in [-1, 0]:
tb_writer = SummaryWriter()
args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
train_sampler = RandomSampler(train_dataset) if args.local_rank == -1 else DistributedSampler(train_dataset)
train_dataloader = DataLoader(train_dataset, sampler=train_sampler, batch_size=args.train_batch_size)
if args.max_steps > 0:
t_total = args.max_steps
args.num_train_epochs = args.max_steps // (len(train_dataloader) // args.gradient_accumulation_steps) + 1
else:
t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs
# Prepare optimizer and schedule (linear warmup and decay)
no_decay = ["bias", "LayerNorm.weight"]
optimizer_grouped_parameters = [
{
"params": [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)],
"weight_decay": args.weight_decay,
},
{"params": [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], "weight_decay": 0.0},
]
optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
scheduler = get_linear_schedule_with_warmup(
optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total
)
# Check if saved optimizer or scheduler states exist
if os.path.isfile(os.path.join(args.model_name_or_path, "optimizer.pt")) and os.path.isfile(
os.path.join(args.model_name_or_path, "scheduler.pt")
):
# Load in optimizer and scheduler states
optimizer.load_state_dict(torch.load(os.path.join(args.model_name_or_path, "optimizer.pt")))
scheduler.load_state_dict(torch.load(os.path.join(args.model_name_or_path, "scheduler.pt")))
if args.fp16:
try:
from apex import amp
except ImportError:
raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level)
# multi-gpu training (should be after apex fp16 initialization)
if args.n_gpu > 1:
model = nn.DataParallel(model)
# Distributed training (should be after apex fp16 initialization)
if args.local_rank != -1:
model = nn.parallel.DistributedDataParallel(
model,
device_ids=[args.local_rank],
output_device=args.local_rank,
find_unused_parameters=True,
)
# Train!
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_dataset))
logger.info(" Num Epochs = %d", args.num_train_epochs)
logger.info(" Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size)
logger.info(
" Total train batch size (w. parallel, distributed & accumulation) = %d",
args.train_batch_size
* args.gradient_accumulation_steps
* (torch.distributed.get_world_size() if args.local_rank != -1 else 1),
)
logger.info(" Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
logger.info(" Total optimization steps = %d", t_total)
global_step = 0
epochs_trained = 0
steps_trained_in_current_epoch = 0
# Check if continuing training from a checkpoint
if os.path.exists(args.model_name_or_path):
# set global_step to global_step of last saved checkpoint from model path
global_step = int(args.model_name_or_path.split("-")[-1].split("/")[0])
epochs_trained = global_step // (len(train_dataloader) // args.gradient_accumulation_steps)
steps_trained_in_current_epoch = global_step % (len(train_dataloader) // args.gradient_accumulation_steps)
logger.info(" Continuing training from checkpoint, will skip to saved global_step")
logger.info(" Continuing training from epoch %d", epochs_trained)
logger.info(" Continuing training from global step %d", global_step)
logger.info(
" Will skip the first %d steps in the first epoch",
steps_trained_in_current_epoch,
)
tr_loss, logging_loss = 0.0, 0.0
model.zero_grad()
train_iterator = trange(
epochs_trained,
int(args.num_train_epochs),
desc="Epoch",
disable=args.local_rank not in [-1, 0],
)
set_seed(args) # Added here for reproducibility
for _ in train_iterator:
epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
for step, batch in enumerate(epoch_iterator):
# Skip past any already trained steps if resuming training
if steps_trained_in_current_epoch > 0:
steps_trained_in_current_epoch -= 1
continue
model.train()
batch = tuple(t.to(args.device) for t in batch)
inputs = {
"input_ids": batch[0],
"attention_mask": batch[1],
"labels": batch[3],
}
inputs["token_type_ids"] = batch[2]
outputs = model(**inputs)
loss = outputs[0] # model outputs are always tuple in transformers (see doc)
if args.n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu parallel training
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
tr_loss += loss.item()
if (step + 1) % args.gradient_accumulation_steps == 0:
if args.fp16:
nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
else:
nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
optimizer.step()
scheduler.step() # Update learning rate schedule
model.zero_grad()
global_step += 1
if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
logs = {}
if (
args.local_rank == -1 and args.evaluate_during_training
): # Only evaluate when single GPU otherwise metrics may not average well
results = evaluate(args, model, tokenizer)
for key, value in results.items():
eval_key = "eval_{}".format(key)
logs[eval_key] = value
loss_scalar = (tr_loss - logging_loss) / args.logging_steps
learning_rate_scalar = scheduler.get_lr()[0]
logs["learning_rate"] = learning_rate_scalar
logs["loss"] = loss_scalar
logging_loss = tr_loss
for key, value in logs.items():
tb_writer.add_scalar(key, value, global_step)
print(json.dumps({**logs, **{"step": global_step}}))
if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0:
# Save model checkpoint
output_dir = os.path.join(args.output_dir, "checkpoint-{}".format(global_step))
model_to_save = (
model.module if hasattr(model, "module") else model
) # Take care of distributed/parallel training
model_to_save.save_pretrained(output_dir)
tokenizer.save_pretrained(output_dir)
torch.save(args, os.path.join(output_dir, "training_args.bin"))
logger.info("Saving model checkpoint to %s", output_dir)
torch.save(optimizer.state_dict(), os.path.join(output_dir, "optimizer.pt"))
torch.save(scheduler.state_dict(), os.path.join(output_dir, "scheduler.pt"))
logger.info("Saving optimizer and scheduler states to %s", output_dir)
if args.max_steps > 0 and global_step > args.max_steps:
epoch_iterator.close()
break
if args.max_steps > 0 and global_step > args.max_steps:
train_iterator.close()
break
if args.local_rank in [-1, 0]:
tb_writer.close()
return global_step, tr_loss / global_step
def evaluate(args, model, tokenizer, prefix="", patience=0):
if args.model_type == "albert":
model.albert.set_regression_threshold(args.regression_threshold)
model.albert.set_patience(patience)
model.albert.reset_stats()
elif args.model_type == "bert":
model.bert.set_regression_threshold(args.regression_threshold)
model.bert.set_patience(patience)
model.bert.reset_stats()
else:
raise NotImplementedError()
# Loop to handle MNLI double evaluation (matched, mis-matched)
eval_task_names = ("mnli", "mnli-mm") if args.task_name == "mnli" else (args.task_name,)
eval_outputs_dirs = (args.output_dir, args.output_dir + "-MM") if args.task_name == "mnli" else (args.output_dir,)
results = {}
for eval_task, eval_output_dir in zip(eval_task_names, eval_outputs_dirs):
eval_dataset = load_and_cache_examples(args, eval_task, tokenizer, evaluate=True)
if not os.path.exists(eval_output_dir) and args.local_rank in [-1, 0]:
os.makedirs(eval_output_dir)
args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
# Note that DistributedSampler samples randomly
eval_sampler = SequentialSampler(eval_dataset)
eval_dataloader = DataLoader(eval_dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
# multi-gpu eval
if args.n_gpu > 1 and not isinstance(model, nn.DataParallel):
model = nn.DataParallel(model)
# Eval!
logger.info("***** Running evaluation {} *****".format(prefix))
logger.info(" Num examples = %d", len(eval_dataset))
logger.info(" Batch size = %d", args.eval_batch_size)
eval_loss = 0.0
nb_eval_steps = 0
preds = None
out_label_ids = None
for batch in tqdm(eval_dataloader, desc="Evaluating"):
model.eval()
batch = tuple(t.to(args.device) for t in batch)
with torch.no_grad():
inputs = {
"input_ids": batch[0],
"attention_mask": batch[1],
"labels": batch[3],
}
inputs["token_type_ids"] = batch[2]
outputs = model(**inputs)
tmp_eval_loss, logits = outputs[:2]
eval_loss += tmp_eval_loss.mean().item()
nb_eval_steps += 1
if preds is None:
preds = logits.detach().cpu().numpy()
out_label_ids = inputs["labels"].detach().cpu().numpy()
else:
preds = np.append(preds, logits.detach().cpu().numpy(), axis=0)
out_label_ids = np.append(out_label_ids, inputs["labels"].detach().cpu().numpy(), axis=0)
eval_loss = eval_loss / nb_eval_steps
if args.output_mode == "classification":
preds = np.argmax(preds, axis=1)
elif args.output_mode == "regression":
preds = np.squeeze(preds)
result = compute_metrics(eval_task, preds, out_label_ids)
results.update(result)
output_eval_file = os.path.join(eval_output_dir, prefix, "eval_results.txt")
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results {} *****".format(prefix))
for key in sorted(result.keys()):
logger.info(" %s = %s", key, str(result[key]))
print(" %s = %s" % (key, str(result[key])))
writer.write("%s = %s\n" % (key, str(result[key])))
if args.eval_all_checkpoints and patience != 0:
if args.model_type == "albert":
model.albert.log_stats()
elif args.model_type == "bert":
model.bert.log_stats()
else:
raise NotImplementedError()
return results
def load_and_cache_examples(args, task, tokenizer, evaluate=False):
if args.local_rank not in [-1, 0] and not evaluate:
torch.distributed.barrier() # Make sure only the first process in distributed training process the dataset, and the others will use the cache
processor = processors[task]()
output_mode = output_modes[task]
# Load data features from cache or dataset file
cached_features_file = os.path.join(
args.data_dir,
"cached_{}_{}_{}_{}".format(
"dev" if evaluate else "train",
list(filter(None, args.model_name_or_path.split("/"))).pop(),
str(args.max_seq_length),
str(task),
),
)
if os.path.exists(cached_features_file) and not args.overwrite_cache:
logger.info("Loading features from cached file %s", cached_features_file)
features = torch.load(cached_features_file)
else:
logger.info("Creating features from dataset file at %s", args.data_dir)
label_list = processor.get_labels()
if task in ["mnli", "mnli-mm"] and args.model_type in ["roberta", "xlmroberta"]:
# HACK(label indices are swapped in RoBERTa pretrained model)
label_list[1], label_list[2] = label_list[2], label_list[1]
examples = (
processor.get_dev_examples(args.data_dir) if evaluate else processor.get_train_examples(args.data_dir)
)
features = convert_examples_to_features(
examples,
tokenizer,
label_list=label_list,
max_length=args.max_seq_length,
output_mode=output_mode,
)
if args.local_rank in [-1, 0]:
logger.info("Saving features into cached file %s", cached_features_file)
torch.save(features, cached_features_file)
if args.local_rank == 0 and not evaluate:
torch.distributed.barrier() # Make sure only the first process in distributed training process the dataset, and the others will use the cache
# Convert to Tensors and build dataset
all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long)
all_attention_mask = torch.tensor([f.attention_mask for f in features], dtype=torch.long)
all_token_type_ids = torch.tensor([f.token_type_ids for f in features], dtype=torch.long)
if output_mode == "classification":
all_labels = torch.tensor([f.label for f in features], dtype=torch.long)
elif output_mode == "regression":
all_labels = torch.tensor([f.label for f in features], dtype=torch.float)
dataset = TensorDataset(all_input_ids, all_attention_mask, all_token_type_ids, all_labels)
return dataset
def main():
parser = argparse.ArgumentParser()
# Required parameters
parser.add_argument(
"--data_dir",
default=None,
type=str,
required=True,
help="The input data dir. Should contain the .tsv files (or other data files) for the task.",
)
parser.add_argument(
"--model_type",
default=None,
type=str,
required=True,
help="Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys()),
)
parser.add_argument(
"--model_name_or_path",
default=None,
type=str,
required=True,
help="Path to pre-trained model or shortcut name.",
)
parser.add_argument(
"--task_name",
default=None,
type=str,
required=True,
help="The name of the task to train selected in the list: " + ", ".join(processors.keys()),
)
parser.add_argument(
"--output_dir",
default=None,
type=str,
required=True,
help="The output directory where the model predictions and checkpoints will be written.",
)
parser.add_argument(
"--patience",
default="0",
type=str,
required=False,
)
parser.add_argument(
"--regression_threshold",
default=0,
type=float,
required=False,
)
# Other parameters
parser.add_argument(
"--config_name",
default="",
type=str,
help="Pretrained config name or path if not the same as model_name",
)
parser.add_argument(
"--tokenizer_name",
default="",
type=str,
help="Pretrained tokenizer name or path if not the same as model_name",
)
parser.add_argument(
"--cache_dir",
default="",
type=str,
help="Where do you want to store the pre-trained models downloaded from huggingface.co",
)
parser.add_argument(
"--max_seq_length",
default=128,
type=int,
help=(
"The maximum total input sequence length after tokenization. Sequences longer "
"than this will be truncated, sequences shorter will be padded."
),
)
parser.add_argument("--do_train", action="store_true", help="Whether to run training.")
parser.add_argument("--do_eval", action="store_true", help="Whether to run eval on the dev set.")
parser.add_argument(
"--evaluate_during_training",
action="store_true",
help="Run evaluation during training at each logging step.",
)
parser.add_argument(
"--do_lower_case",
action="store_true",
help="Set this flag if you are using an uncased model.",
)
parser.add_argument(
"--per_gpu_train_batch_size",
default=8,
type=int,
help="Batch size per GPU/CPU for training.",
)
parser.add_argument(
"--per_gpu_eval_batch_size",
default=1,
type=int,
help="Batch size per GPU/CPU for evaluation.",
)
parser.add_argument(
"--gradient_accumulation_steps",
type=int,
default=1,
help="Number of updates steps to accumulate before performing a backward/update pass.",
)
parser.add_argument(
"--learning_rate",
default=5e-5,
type=float,
help="The initial learning rate for Adam.",
)
parser.add_argument("--weight_decay", default=0.0, type=float, help="Weight decay if we apply some.")
parser.add_argument("--adam_epsilon", default=1e-8, type=float, help="Epsilon for Adam optimizer.")
parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
parser.add_argument(
"--num_train_epochs",
default=3.0,
type=float,
help="Total number of training epochs to perform.",
)
parser.add_argument(
"--max_steps",
default=-1,
type=int,
help="If > 0: set total number of training steps to perform. Override num_train_epochs.",
)
parser.add_argument("--warmup_steps", default=0, type=int, help="Linear warmup over warmup_steps.")
parser.add_argument("--logging_steps", type=int, default=500, help="Log every X updates steps.")
parser.add_argument(
"--save_steps",
type=int,
default=500,
help="Save checkpoint every X updates steps.",
)
parser.add_argument(
"--eval_all_checkpoints",
action="store_true",
help="Evaluate all checkpoints starting with the same prefix as model_name ending and ending with step number",
)
parser.add_argument("--no_cuda", action="store_true", help="Avoid using CUDA when available")
parser.add_argument(
"--overwrite_output_dir",
action="store_true",
help="Overwrite the content of the output directory",
)
parser.add_argument(
"--overwrite_cache",
action="store_true",
help="Overwrite the cached training and evaluation sets",
)
parser.add_argument("--seed", type=int, default=42, help="random seed for initialization")
parser.add_argument(
"--fp16",
action="store_true",
help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit",
)
parser.add_argument(
"--fp16_opt_level",
type=str,
default="O1",
help=(
"For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']. "
"See details at https://nvidia.github.io/apex/amp.html"
),
)
parser.add_argument(
"--local_rank",
type=int,
default=-1,
help="For distributed training: local_rank",
)
parser.add_argument("--server_ip", type=str, default="", help="For distant debugging.")
parser.add_argument("--server_port", type=str, default="", help="For distant debugging.")
args = parser.parse_args()
if (
os.path.exists(args.output_dir)
and os.listdir(args.output_dir)
and args.do_train
and not args.overwrite_output_dir
):
raise ValueError(
"Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(
args.output_dir
)
)
# Setup distant debugging if needed
if args.server_ip and args.server_port:
# Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
import ptvsd
print("Waiting for debugger attach")
ptvsd.enable_attach(address=(args.server_ip, args.server_port), redirect_output=True)
ptvsd.wait_for_attach()
# Setup CUDA, GPU & distributed training
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
args.n_gpu = torch.cuda.device_count()
else: # Initializes the distributed backend which will take care of synchronizing nodes/GPUs
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
torch.distributed.init_process_group(backend="nccl")
args.n_gpu = 1
args.device = device
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN,
)
logger.warning(
"Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
args.local_rank,
device,
args.n_gpu,
bool(args.local_rank != -1),
args.fp16,
)
# Set the verbosity to info of the Transformers logger (on main process only):
if is_main_process(args.local_rank):
transformers.utils.logging.set_verbosity_info()
transformers.utils.logging.enable_default_handler()
transformers.utils.logging.enable_explicit_format()
# Set seed
set_seed(args)
# Prepare GLUE task
args.task_name = args.task_name.lower()
if args.task_name not in processors:
raise ValueError("Task not found: %s" % (args.task_name))
processor = processors[args.task_name]()
args.output_mode = output_modes[args.task_name]
label_list = processor.get_labels()
num_labels = len(label_list)
if args.patience != "0" and args.per_gpu_eval_batch_size != 1:
raise ValueError("The eval batch size must be 1 with PABEE inference on.")
# Load pretrained model and tokenizer
if args.local_rank not in [-1, 0]:
torch.distributed.barrier() # Make sure only the first process in distributed training will download model & vocab
args.model_type = args.model_type.lower()
config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
config = config_class.from_pretrained(
args.config_name if args.config_name else args.model_name_or_path,
num_labels=num_labels,
finetuning_task=args.task_name,
cache_dir=args.cache_dir if args.cache_dir else None,
)
tokenizer = tokenizer_class.from_pretrained(
args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
do_lower_case=args.do_lower_case,
cache_dir=args.cache_dir if args.cache_dir else None,
)
model = model_class.from_pretrained(
args.model_name_or_path,
from_tf=bool(".ckpt" in args.model_name_or_path),
config=config,
cache_dir=args.cache_dir if args.cache_dir else None,
)
if args.local_rank == 0:
torch.distributed.barrier() # Make sure only the first process in distributed training will download model & vocab
model.to(args.device)
print("Total Model Parameters:", sum(param.numel() for param in model.parameters()))
output_layers_param_num = sum(param.numel() for param in model.classifiers.parameters())
print("Output Layers Parameters:", output_layers_param_num)
single_output_layer_param_num = sum(param.numel() for param in model.classifiers[0].parameters())
print(
"Added Output Layers Parameters:",
output_layers_param_num - single_output_layer_param_num,
)
logger.info("Training/evaluation parameters %s", args)
# Training
if args.do_train:
train_dataset = load_and_cache_examples(args, args.task_name, tokenizer, evaluate=False)
global_step, tr_loss = train(args, train_dataset, model, tokenizer)
logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)
# Saving best-practices: if you use defaults names for the model, you can reload it using from_pretrained()
if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
logger.info("Saving model checkpoint to %s", args.output_dir)
# Save a trained model, configuration and tokenizer using `save_pretrained()`.
# They can then be reloaded using `from_pretrained()`
model_to_save = (
model.module if hasattr(model, "module") else model
) # Take care of distributed/parallel training
model_to_save.save_pretrained(args.output_dir)
tokenizer.save_pretrained(args.output_dir)
# Good practice: save your training arguments together with the trained model
torch.save(args, os.path.join(args.output_dir, "training_args.bin"))
# Load a trained model and vocabulary that you have fine-tuned
model = model_class.from_pretrained(args.output_dir)
tokenizer = tokenizer_class.from_pretrained(args.output_dir)
model.to(args.device)
# Evaluation
results = {}
if args.do_eval and args.local_rank in [-1, 0]:
patience_list = [int(x) for x in args.patience.split(",")]
tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
checkpoints = [args.output_dir]
if args.eval_all_checkpoints:
checkpoints = [
os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + "/**/" + WEIGHTS_NAME, recursive=True))
]
logger.info("Evaluate the following checkpoints: %s", checkpoints)
for checkpoint in checkpoints:
global_step = checkpoint.split("-")[-1] if len(checkpoints) > 1 else ""
prefix = checkpoint.split("/")[-1] if checkpoint.find("checkpoint") != -1 else ""
model = model_class.from_pretrained(checkpoint)
model.to(args.device)
print(f"Evaluation for checkpoint {prefix}")
for patience in patience_list:
result = evaluate(args, model, tokenizer, prefix=prefix, patience=patience)
result = {k + "_{}".format(global_step): v for k, v in result.items()}
results.update(result)
return results
if __name__ == "__main__":
main()
|
transformers/examples/research_projects/bert-loses-patience/run_glue_with_pabee.py/0
|
{
"file_path": "transformers/examples/research_projects/bert-loses-patience/run_glue_with_pabee.py",
"repo_id": "transformers",
"token_count": 12995
}
| 319
|
# DeeBERT: Early Exiting for *BERT
This is the code base for the paper [DeeBERT: Dynamic Early Exiting for Accelerating BERT Inference](https://www.aclweb.org/anthology/2020.acl-main.204/), modified from its [original code base](https://github.com/castorini/deebert).
The original code base also has information for downloading sample models that we have trained in advance.
## Usage
There are three scripts in the folder which can be run directly.
In each script, there are several things to modify before running:
* `PATH_TO_DATA`: path to the GLUE dataset.
* `--output_dir`: path for saving fine-tuned models. Default: `./saved_models`.
* `--plot_data_dir`: path for saving evaluation results. Default: `./results`. Results are printed to stdout and also saved to `npy` files in this directory to facilitate plotting figures and further analyses.
* `MODEL_TYPE`: bert or roberta
* `MODEL_SIZE`: base or large
* `DATASET`: SST-2, MRPC, RTE, QNLI, QQP, or MNLI
#### train_deebert.sh
This is for fine-tuning DeeBERT models.
#### eval_deebert.sh
This is for evaluating each exit layer for fine-tuned DeeBERT models.
#### entropy_eval.sh
This is for evaluating fine-tuned DeeBERT models, given a number of different early exit entropy thresholds.
## Citation
Please cite our paper if you find the resource useful:
```bibtex
@inproceedings{xin-etal-2020-deebert,
title = "{D}ee{BERT}: Dynamic Early Exiting for Accelerating {BERT} Inference",
author = "Xin, Ji and
Tang, Raphael and
Lee, Jaejun and
Yu, Yaoliang and
Lin, Jimmy",
booktitle = "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
month = jul,
year = "2020",
address = "Online",
publisher = "Association for Computational Linguistics",
url = "https://www.aclweb.org/anthology/2020.acl-main.204",
pages = "2246--2251",
}
```
|
transformers/examples/research_projects/deebert/README.md/0
|
{
"file_path": "transformers/examples/research_projects/deebert/README.md",
"repo_id": "transformers",
"token_count": 621
}
| 320
|
# coding=utf-8
# Copyright 2019-present, the HuggingFace Inc. team.
#
# 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.
"""
Preprocessing script before distillation.
"""
import argparse
import logging
import pickle
import random
import time
import numpy as np
from transformers import BertTokenizer, GPT2Tokenizer, RobertaTokenizer
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s", datefmt="%m/%d/%Y %H:%M:%S", level=logging.INFO
)
logger = logging.getLogger(__name__)
def main():
parser = argparse.ArgumentParser(
description="Preprocess the data to avoid re-doing it several times by (tokenization + token_to_ids)."
)
parser.add_argument("--file_path", type=str, default="data/dump.txt", help="The path to the data.")
parser.add_argument("--tokenizer_type", type=str, default="bert", choices=["bert", "roberta", "gpt2"])
parser.add_argument("--tokenizer_name", type=str, default="bert-base-uncased", help="The tokenizer to use.")
parser.add_argument("--dump_file", type=str, default="data/dump", help="The dump file prefix.")
args = parser.parse_args()
logger.info(f"Loading Tokenizer ({args.tokenizer_name})")
if args.tokenizer_type == "bert":
tokenizer = BertTokenizer.from_pretrained(args.tokenizer_name)
bos = tokenizer.special_tokens_map["cls_token"] # `[CLS]`
sep = tokenizer.special_tokens_map["sep_token"] # `[SEP]`
elif args.tokenizer_type == "roberta":
tokenizer = RobertaTokenizer.from_pretrained(args.tokenizer_name)
bos = tokenizer.special_tokens_map["cls_token"] # `<s>`
sep = tokenizer.special_tokens_map["sep_token"] # `</s>`
elif args.tokenizer_type == "gpt2":
tokenizer = GPT2Tokenizer.from_pretrained(args.tokenizer_name)
bos = tokenizer.special_tokens_map["bos_token"] # `<|endoftext|>`
sep = tokenizer.special_tokens_map["eos_token"] # `<|endoftext|>`
logger.info(f"Loading text from {args.file_path}")
with open(args.file_path, "r", encoding="utf8") as fp:
data = fp.readlines()
logger.info("Start encoding")
logger.info(f"{len(data)} examples to process.")
rslt = []
iter = 0
interval = 10000
start = time.time()
for text in data:
text = f"{bos} {text.strip()} {sep}"
token_ids = tokenizer.encode(text, add_special_tokens=False)
rslt.append(token_ids)
iter += 1
if iter % interval == 0:
end = time.time()
logger.info(f"{iter} examples processed. - {(end-start):.2f}s/{interval}expl")
start = time.time()
logger.info("Finished binarization")
logger.info(f"{len(data)} examples processed.")
dp_file = f"{args.dump_file}.{args.tokenizer_name}.pickle"
vocab_size = tokenizer.vocab_size
if vocab_size < (1 << 16):
rslt_ = [np.uint16(d) for d in rslt]
else:
rslt_ = [np.int32(d) for d in rslt]
random.shuffle(rslt_)
logger.info(f"Dump to {dp_file}")
with open(dp_file, "wb") as handle:
pickle.dump(rslt_, handle, protocol=pickle.HIGHEST_PROTOCOL)
if __name__ == "__main__":
main()
|
transformers/examples/research_projects/distillation/scripts/binarized_data.py/0
|
{
"file_path": "transformers/examples/research_projects/distillation/scripts/binarized_data.py",
"repo_id": "transformers",
"token_count": 1429
}
| 321
|
import torch
from transformers import AutoModel
class FSNERModel(torch.nn.Module):
"""
The FSNER model implements a few-shot named entity recognition method from the paper `Example-Based Named Entity Recognition <https://arxiv.org/abs/2008.10570>`__ by
Morteza Ziyadi, Yuting Sun, Abhishek Goswami, Jade Huang, Weizhu Chen. To identify entity spans in a new domain, it
uses a train-free few-shot learning approach inspired by question-answering.
"""
def __init__(self, pretrained_model_name_or_path="sayef/fsner-bert-base-uncased"):
super(FSNERModel, self).__init__()
self.bert = AutoModel.from_pretrained(pretrained_model_name_or_path, return_dict=True)
self.cos = torch.nn.CosineSimilarity(3, 1e-08)
self.softmax = torch.nn.Softmax(dim=1)
def BERT(self, **inputs):
return self.bert(**inputs).last_hidden_state
def VectorSum(self, token_embeddings):
return token_embeddings.sum(2, keepdim=True)
def Atten(self, q_rep, S_rep, T=1):
return self.softmax(T * self.cos(q_rep, S_rep))
def forward(self, W_query, W_supports):
"""
Find scores of each token being start and end token for an entity.
Args:
W_query (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
Indices of query sequence tokens in the vocabulary.
W_supports (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
Indices of support sequence tokens in the vocabulary.
Returns:
p_start (`torch.FloatTensor` of shape `(batch_size, sequence_length)`): Scores of each token as
being start token of an entity
p_end (`torch.FloatTensor` of shape `(batch_size, sequence_length)`): Scores of each token as
being end token of an entity
"""
support_sizes = W_supports["sizes"].tolist()
start_token_id = W_supports["start_token_id"].item()
end_token_id = W_supports["end_token_id"].item()
del W_supports["sizes"]
del W_supports["start_token_id"]
del W_supports["end_token_id"]
q = self.BERT(**W_query)
S = self.BERT(**W_supports)
p_starts = None
p_ends = None
start_token_masks = W_supports["input_ids"] == start_token_id
end_token_masks = W_supports["input_ids"] == end_token_id
for i, size in enumerate(support_sizes):
if i == 0:
s = 0
else:
s = support_sizes[i - 1]
s_start = S[s : s + size][start_token_masks[s : s + size]]
s_end = S[s : s + size][end_token_masks[s : s + size]]
p_start = torch.matmul(q[i], s_start.T).sum(1).softmax(0)
p_end = torch.matmul(q[i], s_end.T).sum(1).softmax(0)
if p_starts is not None:
p_starts = torch.vstack((p_starts, p_start))
p_ends = torch.vstack((p_ends, p_end))
else:
p_starts = p_start
p_ends = p_end
return p_starts, p_ends
|
transformers/examples/research_projects/fsner/src/fsner/model.py/0
|
{
"file_path": "transformers/examples/research_projects/fsner/src/fsner/model.py",
"repo_id": "transformers",
"token_count": 1436
}
| 322
|
import os
from dataclasses import replace
import jax
import wandb
from bigbird_flax import Args, DataCollator, FlaxBigBirdForNaturalQuestions, Trainer, build_tx, train_step, val_step
from datasets import load_dataset
from flax import jax_utils
from transformers import BigBirdTokenizerFast
if __name__ == "__main__":
print("#################### AVAILABLE DEVICES ####################")
print(jax.devices())
print("###########################################################")
# setup for wandb sweep
args = Args()
logger = wandb.init(project="bigbird-natural-questions", config=args.__dict__)
wandb_args = dict(logger.config)
del wandb_args["batch_size"]
args = replace(args, **wandb_args)
base_dir = args.base_dir + "-" + wandb.run.id
args = replace(args, base_dir=base_dir)
print(args)
tr_dataset = load_dataset("json", data_files=args.tr_data_path)["train"]
val_dataset = load_dataset("json", data_files=args.val_data_path)["train"]
# drop extra batch for now
indices = range(len(tr_dataset) - len(tr_dataset) % args.batch_size)
tr_dataset = tr_dataset.shuffle().select(indices)
indices = range(len(val_dataset) - len(val_dataset) % args.batch_size)
val_dataset = val_dataset.shuffle().select(indices)
if os.environ.get("TRAIN_ON_SMALL", "false") == "true":
tr_dataset = tr_dataset.shuffle().select(range(80000))
val_dataset = val_dataset.shuffle().select(range(8000))
print(tr_dataset)
print(val_dataset)
model = FlaxBigBirdForNaturalQuestions.from_pretrained(
args.model_id, block_size=args.block_size, num_random_blocks=args.num_random_blocks
)
tokenizer = BigBirdTokenizerFast.from_pretrained(args.model_id)
data_collator = DataCollator(pad_id=tokenizer.pad_token_id, max_length=4096)
tx_args = {
"lr": args.lr,
"init_lr": args.init_lr,
"warmup_steps": args.warmup_steps,
"num_train_steps": args.max_epochs * (len(tr_dataset) // args.batch_size),
"weight_decay": args.weight_decay,
}
tx, lr = build_tx(**tx_args)
trainer = Trainer(
args=args,
data_collator=data_collator,
model_save_fn=model.save_pretrained,
train_step_fn=train_step,
val_step_fn=val_step,
logger=logger,
scheduler_fn=lr,
)
ckpt_dir = None
state = trainer.create_state(model, tx, num_train_steps=tx_args["num_train_steps"], ckpt_dir=ckpt_dir)
try:
trainer.train(state, tr_dataset, val_dataset)
except KeyboardInterrupt:
print("Oooops; TRAINING STOPPED UNFORTUNATELY")
print("SAVING WEIGHTS IN `final-weights`")
params = jax_utils.unreplicate(state.params)
model.save_pretrained(os.path.join(args.base_dir, "final-weights"), params=params)
|
transformers/examples/research_projects/jax-projects/big_bird/train.py/0
|
{
"file_path": "transformers/examples/research_projects/jax-projects/big_bird/train.py",
"repo_id": "transformers",
"token_count": 1164
}
| 323
|
# Long Form Question Answering
Author: @yjernite
This folder contains the code for the Long Form Question answering [demo](http://35.226.96.115:8080/) as well as methods to train and use a fully end-to-end Long Form Question Answering system using the [🤗transformers](https://github.com/huggingface/transformers) and [🤗datasets](https://github.com/huggingface/datasets) libraries.
You can use these methods to train your own system by following along the associate [notebook](https://github.com/huggingface/notebooks/blob/master/longform-qa/Long_Form_Question_Answering_with_ELI5_and_Wikipedia.ipynb) or [blog post](https://yjernite.github.io/lfqa.html).
|
transformers/examples/research_projects/longform-qa/README.md/0
|
{
"file_path": "transformers/examples/research_projects/longform-qa/README.md",
"repo_id": "transformers",
"token_count": 208
}
| 324
|
# 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.
"""Fine-pruning Masked BERT on sequence classification on GLUE."""
import argparse
import glob
import json
import logging
import os
import random
import numpy as np
import torch
from emmental import MaskedBertConfig, MaskedBertForSequenceClassification
from torch import nn
from torch.utils.data import DataLoader, RandomSampler, SequentialSampler, TensorDataset
from torch.utils.data.distributed import DistributedSampler
from tqdm import tqdm, trange
from transformers import (
WEIGHTS_NAME,
AdamW,
BertConfig,
BertForSequenceClassification,
BertTokenizer,
get_linear_schedule_with_warmup,
)
from transformers import glue_compute_metrics as compute_metrics
from transformers import glue_convert_examples_to_features as convert_examples_to_features
from transformers import glue_output_modes as output_modes
from transformers import glue_processors as processors
try:
from torch.utils.tensorboard import SummaryWriter
except ImportError:
from tensorboardX import SummaryWriter
logger = logging.getLogger(__name__)
MODEL_CLASSES = {
"bert": (BertConfig, BertForSequenceClassification, BertTokenizer),
"masked_bert": (MaskedBertConfig, MaskedBertForSequenceClassification, BertTokenizer),
}
def set_seed(args):
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if args.n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
def schedule_threshold(
step: int,
total_step: int,
warmup_steps: int,
initial_threshold: float,
final_threshold: float,
initial_warmup: int,
final_warmup: int,
final_lambda: float,
):
if step <= initial_warmup * warmup_steps:
threshold = initial_threshold
elif step > (total_step - final_warmup * warmup_steps):
threshold = final_threshold
else:
spars_warmup_steps = initial_warmup * warmup_steps
spars_schedu_steps = (final_warmup + initial_warmup) * warmup_steps
mul_coeff = 1 - (step - spars_warmup_steps) / (total_step - spars_schedu_steps)
threshold = final_threshold + (initial_threshold - final_threshold) * (mul_coeff**3)
regu_lambda = final_lambda * threshold / final_threshold
return threshold, regu_lambda
def regularization(model: nn.Module, mode: str):
regu, counter = 0, 0
for name, param in model.named_parameters():
if "mask_scores" in name:
if mode == "l1":
regu += torch.norm(torch.sigmoid(param), p=1) / param.numel()
elif mode == "l0":
regu += torch.sigmoid(param - 2 / 3 * np.log(0.1 / 1.1)).sum() / param.numel()
else:
raise ValueError("Don't know this mode.")
counter += 1
return regu / counter
def train(args, train_dataset, model, tokenizer, teacher=None):
"""Train the model"""
if args.local_rank in [-1, 0]:
tb_writer = SummaryWriter(log_dir=args.output_dir)
args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
train_sampler = RandomSampler(train_dataset) if args.local_rank == -1 else DistributedSampler(train_dataset)
train_dataloader = DataLoader(train_dataset, sampler=train_sampler, batch_size=args.train_batch_size)
if args.max_steps > 0:
t_total = args.max_steps
args.num_train_epochs = args.max_steps // (len(train_dataloader) // args.gradient_accumulation_steps) + 1
else:
t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs
# Prepare optimizer and schedule (linear warmup and decay)
no_decay = ["bias", "LayerNorm.weight"]
optimizer_grouped_parameters = [
{
"params": [p for n, p in model.named_parameters() if "mask_score" in n and p.requires_grad],
"lr": args.mask_scores_learning_rate,
},
{
"params": [
p
for n, p in model.named_parameters()
if "mask_score" not in n and p.requires_grad and not any(nd in n for nd in no_decay)
],
"lr": args.learning_rate,
"weight_decay": args.weight_decay,
},
{
"params": [
p
for n, p in model.named_parameters()
if "mask_score" not in n and p.requires_grad and any(nd in n for nd in no_decay)
],
"lr": args.learning_rate,
"weight_decay": 0.0,
},
]
optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
scheduler = get_linear_schedule_with_warmup(
optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total
)
# Check if saved optimizer or scheduler states exist
if os.path.isfile(os.path.join(args.model_name_or_path, "optimizer.pt")) and os.path.isfile(
os.path.join(args.model_name_or_path, "scheduler.pt")
):
# Load in optimizer and scheduler states
optimizer.load_state_dict(torch.load(os.path.join(args.model_name_or_path, "optimizer.pt")))
scheduler.load_state_dict(torch.load(os.path.join(args.model_name_or_path, "scheduler.pt")))
if args.fp16:
try:
from apex import amp
except ImportError:
raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level)
# multi-gpu training (should be after apex fp16 initialization)
if args.n_gpu > 1:
model = nn.DataParallel(model)
# Distributed training (should be after apex fp16 initialization)
if args.local_rank != -1:
model = nn.parallel.DistributedDataParallel(
model,
device_ids=[args.local_rank],
output_device=args.local_rank,
find_unused_parameters=True,
)
# Train!
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_dataset))
logger.info(" Num Epochs = %d", args.num_train_epochs)
logger.info(" Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size)
logger.info(
" Total train batch size (w. parallel, distributed & accumulation) = %d",
args.train_batch_size
* args.gradient_accumulation_steps
* (torch.distributed.get_world_size() if args.local_rank != -1 else 1),
)
logger.info(" Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
logger.info(" Total optimization steps = %d", t_total)
# Distillation
if teacher is not None:
logger.info(" Training with distillation")
global_step = 0
# Global TopK
if args.global_topk:
threshold_mem = None
epochs_trained = 0
steps_trained_in_current_epoch = 0
# Check if continuing training from a checkpoint
if os.path.exists(args.model_name_or_path):
# set global_step to global_step of last saved checkpoint from model path
try:
global_step = int(args.model_name_or_path.split("-")[-1].split("/")[0])
except ValueError:
global_step = 0
epochs_trained = global_step // (len(train_dataloader) // args.gradient_accumulation_steps)
steps_trained_in_current_epoch = global_step % (len(train_dataloader) // args.gradient_accumulation_steps)
logger.info(" Continuing training from checkpoint, will skip to saved global_step")
logger.info(" Continuing training from epoch %d", epochs_trained)
logger.info(" Continuing training from global step %d", global_step)
logger.info(" Will skip the first %d steps in the first epoch", steps_trained_in_current_epoch)
tr_loss, logging_loss = 0.0, 0.0
model.zero_grad()
train_iterator = trange(
epochs_trained,
int(args.num_train_epochs),
desc="Epoch",
disable=args.local_rank not in [-1, 0],
)
set_seed(args) # Added here for reproducibility
for _ in train_iterator:
epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
for step, batch in enumerate(epoch_iterator):
# Skip past any already trained steps if resuming training
if steps_trained_in_current_epoch > 0:
steps_trained_in_current_epoch -= 1
continue
model.train()
batch = tuple(t.to(args.device) for t in batch)
threshold, regu_lambda = schedule_threshold(
step=global_step,
total_step=t_total,
warmup_steps=args.warmup_steps,
final_threshold=args.final_threshold,
initial_threshold=args.initial_threshold,
final_warmup=args.final_warmup,
initial_warmup=args.initial_warmup,
final_lambda=args.final_lambda,
)
# Global TopK
if args.global_topk:
if threshold == 1.0:
threshold = -1e2 # Or an indefinitely low quantity
else:
if (threshold_mem is None) or (global_step % args.global_topk_frequency_compute == 0):
# Sort all the values to get the global topK
concat = torch.cat(
[param.view(-1) for name, param in model.named_parameters() if "mask_scores" in name]
)
n = concat.numel()
kth = max(n - (int(n * threshold) + 1), 1)
threshold_mem = concat.kthvalue(kth).values.item()
threshold = threshold_mem
else:
threshold = threshold_mem
inputs = {"input_ids": batch[0], "attention_mask": batch[1], "labels": batch[3]}
if args.model_type != "distilbert":
inputs["token_type_ids"] = (
batch[2] if args.model_type in ["bert", "masked_bert", "xlnet", "albert"] else None
) # XLM, DistilBERT, RoBERTa, and XLM-RoBERTa don't use segment_ids
if "masked" in args.model_type:
inputs["threshold"] = threshold
outputs = model(**inputs)
loss, logits_stu = outputs # model outputs are always tuple in transformers (see doc)
# Distillation loss
if teacher is not None:
if "token_type_ids" not in inputs:
inputs["token_type_ids"] = None if args.teacher_type == "xlm" else batch[2]
with torch.no_grad():
(logits_tea,) = teacher(
input_ids=inputs["input_ids"],
token_type_ids=inputs["token_type_ids"],
attention_mask=inputs["attention_mask"],
)
loss_logits = nn.functional.kl_div(
input=nn.functional.log_softmax(logits_stu / args.temperature, dim=-1),
target=nn.functional.softmax(logits_tea / args.temperature, dim=-1),
reduction="batchmean",
) * (args.temperature**2)
loss = args.alpha_distil * loss_logits + args.alpha_ce * loss
# Regularization
if args.regularization is not None:
regu_ = regularization(model=model, mode=args.regularization)
loss = loss + regu_lambda * regu_
if args.n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu parallel training
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
tr_loss += loss.item()
if (step + 1) % args.gradient_accumulation_steps == 0 or (
# last step in epoch but step is always smaller than gradient_accumulation_steps
len(epoch_iterator) <= args.gradient_accumulation_steps and (step + 1) == len(epoch_iterator)
):
if args.fp16:
nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
else:
nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
tb_writer.add_scalar("threshold", threshold, global_step)
for name, param in model.named_parameters():
if not param.requires_grad:
continue
tb_writer.add_scalar("parameter_mean/" + name, param.data.mean(), global_step)
tb_writer.add_scalar("parameter_std/" + name, param.data.std(), global_step)
tb_writer.add_scalar("parameter_min/" + name, param.data.min(), global_step)
tb_writer.add_scalar("parameter_max/" + name, param.data.max(), global_step)
tb_writer.add_scalar("grad_mean/" + name, param.grad.data.mean(), global_step)
tb_writer.add_scalar("grad_std/" + name, param.grad.data.std(), global_step)
if args.regularization is not None and "mask_scores" in name:
if args.regularization == "l1":
perc = (torch.sigmoid(param) > threshold).sum().item() / param.numel()
elif args.regularization == "l0":
perc = (torch.sigmoid(param - 2 / 3 * np.log(0.1 / 1.1))).sum().item() / param.numel()
tb_writer.add_scalar("retained_weights_perc/" + name, perc, global_step)
optimizer.step()
scheduler.step() # Update learning rate schedule
model.zero_grad()
global_step += 1
if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
logs = {}
if (
args.local_rank == -1 and args.evaluate_during_training
): # Only evaluate when single GPU otherwise metrics may not average well
results = evaluate(args, model, tokenizer)
for key, value in results.items():
eval_key = "eval_{}".format(key)
logs[eval_key] = value
loss_scalar = (tr_loss - logging_loss) / args.logging_steps
learning_rate_scalar = scheduler.get_lr()
logs["learning_rate"] = learning_rate_scalar[0]
if len(learning_rate_scalar) > 1:
for idx, lr in enumerate(learning_rate_scalar[1:]):
logs[f"learning_rate/{idx+1}"] = lr
logs["loss"] = loss_scalar
if teacher is not None:
logs["loss/distil"] = loss_logits.item()
if args.regularization is not None:
logs["loss/regularization"] = regu_.item()
if (teacher is not None) or (args.regularization is not None):
if (teacher is not None) and (args.regularization is not None):
logs["loss/instant_ce"] = (
loss.item()
- regu_lambda * logs["loss/regularization"]
- args.alpha_distil * logs["loss/distil"]
) / args.alpha_ce
elif teacher is not None:
logs["loss/instant_ce"] = (
loss.item() - args.alpha_distil * logs["loss/distil"]
) / args.alpha_ce
else:
logs["loss/instant_ce"] = loss.item() - regu_lambda * logs["loss/regularization"]
logging_loss = tr_loss
for key, value in logs.items():
tb_writer.add_scalar(key, value, global_step)
print(json.dumps({**logs, **{"step": global_step}}))
if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0:
# Save model checkpoint
output_dir = os.path.join(args.output_dir, "checkpoint-{}".format(global_step))
if not os.path.exists(output_dir):
os.makedirs(output_dir)
model_to_save = (
model.module if hasattr(model, "module") else model
) # Take care of distributed/parallel training
model_to_save.save_pretrained(output_dir)
tokenizer.save_pretrained(output_dir)
torch.save(args, os.path.join(output_dir, "training_args.bin"))
logger.info("Saving model checkpoint to %s", output_dir)
torch.save(optimizer.state_dict(), os.path.join(output_dir, "optimizer.pt"))
torch.save(scheduler.state_dict(), os.path.join(output_dir, "scheduler.pt"))
logger.info("Saving optimizer and scheduler states to %s", output_dir)
if args.max_steps > 0 and global_step > args.max_steps:
epoch_iterator.close()
break
if args.max_steps > 0 and global_step > args.max_steps:
train_iterator.close()
break
if args.local_rank in [-1, 0]:
tb_writer.close()
return global_step, tr_loss / global_step
def evaluate(args, model, tokenizer, prefix=""):
# Loop to handle MNLI double evaluation (matched, mis-matched)
eval_task_names = ("mnli", "mnli-mm") if args.task_name == "mnli" else (args.task_name,)
eval_outputs_dirs = (args.output_dir, args.output_dir + "/MM") if args.task_name == "mnli" else (args.output_dir,)
results = {}
for eval_task, eval_output_dir in zip(eval_task_names, eval_outputs_dirs):
eval_dataset = load_and_cache_examples(args, eval_task, tokenizer, evaluate=True)
if not os.path.exists(eval_output_dir) and args.local_rank in [-1, 0]:
os.makedirs(eval_output_dir)
args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
# Note that DistributedSampler samples randomly
eval_sampler = SequentialSampler(eval_dataset)
eval_dataloader = DataLoader(eval_dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
# multi-gpu eval
if args.n_gpu > 1 and not isinstance(model, nn.DataParallel):
model = nn.DataParallel(model)
# Eval!
logger.info("***** Running evaluation {} *****".format(prefix))
logger.info(" Num examples = %d", len(eval_dataset))
logger.info(" Batch size = %d", args.eval_batch_size)
eval_loss = 0.0
nb_eval_steps = 0
preds = None
out_label_ids = None
# Global TopK
if args.global_topk:
threshold_mem = None
for batch in tqdm(eval_dataloader, desc="Evaluating"):
model.eval()
batch = tuple(t.to(args.device) for t in batch)
with torch.no_grad():
inputs = {"input_ids": batch[0], "attention_mask": batch[1], "labels": batch[3]}
if args.model_type != "distilbert":
inputs["token_type_ids"] = (
batch[2] if args.model_type in ["bert", "masked_bert", "xlnet", "albert"] else None
) # XLM, DistilBERT, RoBERTa, and XLM-RoBERTa don't use segment_ids
if "masked" in args.model_type:
inputs["threshold"] = args.final_threshold
if args.global_topk:
if threshold_mem is None:
concat = torch.cat(
[param.view(-1) for name, param in model.named_parameters() if "mask_scores" in name]
)
n = concat.numel()
kth = max(n - (int(n * args.final_threshold) + 1), 1)
threshold_mem = concat.kthvalue(kth).values.item()
inputs["threshold"] = threshold_mem
outputs = model(**inputs)
tmp_eval_loss, logits = outputs[:2]
eval_loss += tmp_eval_loss.mean().item()
nb_eval_steps += 1
if preds is None:
preds = logits.detach().cpu().numpy()
out_label_ids = inputs["labels"].detach().cpu().numpy()
else:
preds = np.append(preds, logits.detach().cpu().numpy(), axis=0)
out_label_ids = np.append(out_label_ids, inputs["labels"].detach().cpu().numpy(), axis=0)
eval_loss = eval_loss / nb_eval_steps
if args.output_mode == "classification":
from scipy.special import softmax
probs = softmax(preds, axis=-1)
entropy = np.exp((-probs * np.log(probs)).sum(axis=-1).mean())
preds = np.argmax(preds, axis=1)
elif args.output_mode == "regression":
preds = np.squeeze(preds)
result = compute_metrics(eval_task, preds, out_label_ids)
results.update(result)
if entropy is not None:
result["eval_avg_entropy"] = entropy
output_eval_file = os.path.join(eval_output_dir, prefix, "eval_results.txt")
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results {} *****".format(prefix))
for key in sorted(result.keys()):
logger.info(" %s = %s", key, str(result[key]))
writer.write("%s = %s\n" % (key, str(result[key])))
return results
def load_and_cache_examples(args, task, tokenizer, evaluate=False):
if args.local_rank not in [-1, 0] and not evaluate:
torch.distributed.barrier() # Make sure only the first process in distributed training process the dataset, and the others will use the cache
processor = processors[task]()
output_mode = output_modes[task]
# Load data features from cache or dataset file
cached_features_file = os.path.join(
args.data_dir,
"cached_{}_{}_{}_{}".format(
"dev" if evaluate else "train",
list(filter(None, args.model_name_or_path.split("/"))).pop(),
str(args.max_seq_length),
str(task),
),
)
if os.path.exists(cached_features_file) and not args.overwrite_cache:
logger.info("Loading features from cached file %s", cached_features_file)
features = torch.load(cached_features_file)
else:
logger.info("Creating features from dataset file at %s", args.data_dir)
label_list = processor.get_labels()
if task in ["mnli", "mnli-mm"] and args.model_type in ["roberta", "xlmroberta"]:
# HACK(label indices are swapped in RoBERTa pretrained model)
label_list[1], label_list[2] = label_list[2], label_list[1]
examples = (
processor.get_dev_examples(args.data_dir) if evaluate else processor.get_train_examples(args.data_dir)
)
features = convert_examples_to_features(
examples,
tokenizer,
max_length=args.max_seq_length,
label_list=label_list,
output_mode=output_mode,
)
if args.local_rank in [-1, 0]:
logger.info("Saving features into cached file %s", cached_features_file)
torch.save(features, cached_features_file)
if args.local_rank == 0 and not evaluate:
torch.distributed.barrier() # Make sure only the first process in distributed training process the dataset, and the others will use the cache
# Convert to Tensors and build dataset
all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long)
all_attention_mask = torch.tensor([f.attention_mask for f in features], dtype=torch.long)
all_token_type_ids = torch.tensor([f.token_type_ids for f in features], dtype=torch.long)
if output_mode == "classification":
all_labels = torch.tensor([f.label for f in features], dtype=torch.long)
elif output_mode == "regression":
all_labels = torch.tensor([f.label for f in features], dtype=torch.float)
dataset = TensorDataset(all_input_ids, all_attention_mask, all_token_type_ids, all_labels)
return dataset
def main():
parser = argparse.ArgumentParser()
# Required parameters
parser.add_argument(
"--data_dir",
default=None,
type=str,
required=True,
help="The input data dir. Should contain the .tsv files (or other data files) for the task.",
)
parser.add_argument(
"--model_type",
default=None,
type=str,
required=True,
help="Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys()),
)
parser.add_argument(
"--model_name_or_path",
default=None,
type=str,
required=True,
help="Path to pretrained model or model identifier from huggingface.co/models",
)
parser.add_argument(
"--task_name",
default=None,
type=str,
required=True,
help="The name of the task to train selected in the list: " + ", ".join(processors.keys()),
)
parser.add_argument(
"--output_dir",
default=None,
type=str,
required=True,
help="The output directory where the model predictions and checkpoints will be written.",
)
# Other parameters
parser.add_argument(
"--config_name",
default="",
type=str,
help="Pretrained config name or path if not the same as model_name",
)
parser.add_argument(
"--tokenizer_name",
default="",
type=str,
help="Pretrained tokenizer name or path if not the same as model_name",
)
parser.add_argument(
"--cache_dir",
default="",
type=str,
help="Where do you want to store the pre-trained models downloaded from huggingface.co",
)
parser.add_argument(
"--max_seq_length",
default=128,
type=int,
help=(
"The maximum total input sequence length after tokenization. Sequences longer "
"than this will be truncated, sequences shorter will be padded."
),
)
parser.add_argument("--do_train", action="store_true", help="Whether to run training.")
parser.add_argument("--do_eval", action="store_true", help="Whether to run eval on the dev set.")
parser.add_argument(
"--evaluate_during_training",
action="store_true",
help="Run evaluation during training at each logging step.",
)
parser.add_argument(
"--do_lower_case",
action="store_true",
help="Set this flag if you are using an uncased model.",
)
parser.add_argument(
"--per_gpu_train_batch_size",
default=8,
type=int,
help="Batch size per GPU/CPU for training.",
)
parser.add_argument(
"--per_gpu_eval_batch_size",
default=8,
type=int,
help="Batch size per GPU/CPU for evaluation.",
)
parser.add_argument("--learning_rate", default=5e-5, type=float, help="The initial learning rate for Adam.")
# Pruning parameters
parser.add_argument(
"--mask_scores_learning_rate",
default=1e-2,
type=float,
help="The Adam initial learning rate of the mask scores.",
)
parser.add_argument(
"--initial_threshold", default=1.0, type=float, help="Initial value of the threshold (for scheduling)."
)
parser.add_argument(
"--final_threshold", default=0.7, type=float, help="Final value of the threshold (for scheduling)."
)
parser.add_argument(
"--initial_warmup",
default=1,
type=int,
help=(
"Run `initial_warmup` * `warmup_steps` steps of threshold warmup during which threshold stays "
"at its `initial_threshold` value (sparsity schedule)."
),
)
parser.add_argument(
"--final_warmup",
default=2,
type=int,
help=(
"Run `final_warmup` * `warmup_steps` steps of threshold cool-down during which threshold stays "
"at its final_threshold value (sparsity schedule)."
),
)
parser.add_argument(
"--pruning_method",
default="topK",
type=str,
help=(
"Pruning Method (l0 = L0 regularization, magnitude = Magnitude pruning, topK = Movement pruning,"
" sigmoied_threshold = Soft movement pruning)."
),
)
parser.add_argument(
"--mask_init",
default="constant",
type=str,
help="Initialization method for the mask scores. Choices: constant, uniform, kaiming.",
)
parser.add_argument(
"--mask_scale", default=0.0, type=float, help="Initialization parameter for the chosen initialization method."
)
parser.add_argument("--regularization", default=None, help="Add L0 or L1 regularization to the mask scores.")
parser.add_argument(
"--final_lambda",
default=0.0,
type=float,
help="Regularization intensity (used in conjunction with `regularization`.",
)
parser.add_argument("--global_topk", action="store_true", help="Global TopK on the Scores.")
parser.add_argument(
"--global_topk_frequency_compute",
default=25,
type=int,
help="Frequency at which we compute the TopK global threshold.",
)
# Distillation parameters (optional)
parser.add_argument(
"--teacher_type",
default=None,
type=str,
help=(
"Teacher type. Teacher tokenizer and student (model) tokenizer must output the same tokenization. Only for"
" distillation."
),
)
parser.add_argument(
"--teacher_name_or_path",
default=None,
type=str,
help="Path to the already fine-tuned teacher model. Only for distillation.",
)
parser.add_argument(
"--alpha_ce", default=0.5, type=float, help="Cross entropy loss linear weight. Only for distillation."
)
parser.add_argument(
"--alpha_distil", default=0.5, type=float, help="Distillation loss linear weight. Only for distillation."
)
parser.add_argument(
"--temperature", default=2.0, type=float, help="Distillation temperature. Only for distillation."
)
parser.add_argument(
"--gradient_accumulation_steps",
type=int,
default=1,
help="Number of updates steps to accumulate before performing a backward/update pass.",
)
parser.add_argument("--weight_decay", default=0.0, type=float, help="Weight decay if we apply some.")
parser.add_argument("--adam_epsilon", default=1e-8, type=float, help="Epsilon for Adam optimizer.")
parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
parser.add_argument(
"--num_train_epochs",
default=3.0,
type=float,
help="Total number of training epochs to perform.",
)
parser.add_argument(
"--max_steps",
default=-1,
type=int,
help="If > 0: set total number of training steps to perform. Override num_train_epochs.",
)
parser.add_argument("--warmup_steps", default=0, type=int, help="Linear warmup over warmup_steps.")
parser.add_argument("--logging_steps", type=int, default=50, help="Log every X updates steps.")
parser.add_argument("--save_steps", type=int, default=50, help="Save checkpoint every X updates steps.")
parser.add_argument(
"--eval_all_checkpoints",
action="store_true",
help="Evaluate all checkpoints starting with the same prefix as model_name ending and ending with step number",
)
parser.add_argument("--no_cuda", action="store_true", help="Avoid using CUDA when available")
parser.add_argument(
"--overwrite_output_dir",
action="store_true",
help="Overwrite the content of the output directory",
)
parser.add_argument(
"--overwrite_cache",
action="store_true",
help="Overwrite the cached training and evaluation sets",
)
parser.add_argument("--seed", type=int, default=42, help="random seed for initialization")
parser.add_argument(
"--fp16",
action="store_true",
help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit",
)
parser.add_argument(
"--fp16_opt_level",
type=str,
default="O1",
help=(
"For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']. "
"See details at https://nvidia.github.io/apex/amp.html"
),
)
parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
args = parser.parse_args()
# Regularization
if args.regularization == "null":
args.regularization = None
if (
os.path.exists(args.output_dir)
and os.listdir(args.output_dir)
and args.do_train
and not args.overwrite_output_dir
):
raise ValueError(
f"Output directory ({args.output_dir}) already exists and is not empty. Use --overwrite_output_dir to"
" overcome."
)
# Setup CUDA, GPU & distributed training
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
args.n_gpu = 0 if args.no_cuda else torch.cuda.device_count()
else: # Initializes the distributed backend which will take care of synchronizing nodes/GPUs
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
torch.distributed.init_process_group(backend="nccl")
args.n_gpu = 1
args.device = device
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN,
)
logger.warning(
"Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
args.local_rank,
device,
args.n_gpu,
bool(args.local_rank != -1),
args.fp16,
)
# Set seed
set_seed(args)
# Prepare GLUE task
args.task_name = args.task_name.lower()
if args.task_name not in processors:
raise ValueError("Task not found: %s" % (args.task_name))
processor = processors[args.task_name]()
args.output_mode = output_modes[args.task_name]
label_list = processor.get_labels()
num_labels = len(label_list)
# Load pretrained model and tokenizer
if args.local_rank not in [-1, 0]:
torch.distributed.barrier() # Make sure only the first process in distributed training will download model & vocab
args.model_type = args.model_type.lower()
config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
config = config_class.from_pretrained(
args.config_name if args.config_name else args.model_name_or_path,
num_labels=num_labels,
finetuning_task=args.task_name,
cache_dir=args.cache_dir if args.cache_dir else None,
pruning_method=args.pruning_method,
mask_init=args.mask_init,
mask_scale=args.mask_scale,
)
tokenizer = tokenizer_class.from_pretrained(
args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
cache_dir=args.cache_dir if args.cache_dir else None,
do_lower_case=args.do_lower_case,
)
model = model_class.from_pretrained(
args.model_name_or_path,
from_tf=bool(".ckpt" in args.model_name_or_path),
config=config,
cache_dir=args.cache_dir if args.cache_dir else None,
)
if args.teacher_type is not None:
assert args.teacher_name_or_path is not None
assert args.alpha_distil > 0.0
assert args.alpha_distil + args.alpha_ce > 0.0
teacher_config_class, teacher_model_class, _ = MODEL_CLASSES[args.teacher_type]
teacher_config = teacher_config_class.from_pretrained(args.teacher_name_or_path)
teacher = teacher_model_class.from_pretrained(
args.teacher_name_or_path,
from_tf=False,
config=teacher_config,
cache_dir=args.cache_dir if args.cache_dir else None,
)
teacher.to(args.device)
else:
teacher = None
if args.local_rank == 0:
torch.distributed.barrier() # Make sure only the first process in distributed training will download model & vocab
model.to(args.device)
logger.info("Training/evaluation parameters %s", args)
# Training
if args.do_train:
train_dataset = load_and_cache_examples(args, args.task_name, tokenizer, evaluate=False)
global_step, tr_loss = train(args, train_dataset, model, tokenizer, teacher=teacher)
logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)
# Saving best-practices: if you use defaults names for the model, you can reload it using from_pretrained()
if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
logger.info("Saving model checkpoint to %s", args.output_dir)
# Save a trained model, configuration and tokenizer using `save_pretrained()`.
# They can then be reloaded using `from_pretrained()`
model_to_save = (
model.module if hasattr(model, "module") else model
) # Take care of distributed/parallel training
model_to_save.save_pretrained(args.output_dir)
tokenizer.save_pretrained(args.output_dir)
# Good practice: save your training arguments together with the trained model
torch.save(args, os.path.join(args.output_dir, "training_args.bin"))
# Load a trained model and vocabulary that you have fine-tuned
model = model_class.from_pretrained(args.output_dir)
tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
model.to(args.device)
# Evaluation
results = {}
if args.do_eval and args.local_rank in [-1, 0]:
tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
checkpoints = [args.output_dir]
if args.eval_all_checkpoints:
checkpoints = [
os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + "/**/" + WEIGHTS_NAME, recursive=True))
]
logger.info("Evaluate the following checkpoints: %s", checkpoints)
for checkpoint in checkpoints:
global_step = checkpoint.split("-")[-1] if len(checkpoints) > 1 else ""
prefix = checkpoint.split("/")[-1] if checkpoint.find("checkpoint") != -1 else ""
model = model_class.from_pretrained(checkpoint)
model.to(args.device)
result = evaluate(args, model, tokenizer, prefix=prefix)
result = {k + "_{}".format(global_step): v for k, v in result.items()}
results.update(result)
return results
if __name__ == "__main__":
main()
|
transformers/examples/research_projects/movement-pruning/masked_run_glue.py/0
|
{
"file_path": "transformers/examples/research_projects/movement-pruning/masked_run_glue.py",
"repo_id": "transformers",
"token_count": 18298
}
| 325
|
import itertools
import json
import linecache
import os
import pickle
import re
import socket
import string
from collections import Counter
from logging import getLogger
from pathlib import Path
from typing import Callable, Dict, Iterable, List
import git
import torch
from torch.utils.data import Dataset
from transformers import BartTokenizer, RagTokenizer, T5Tokenizer
def encode_line(tokenizer, line, max_length, padding_side, pad_to_max_length=True, return_tensors="pt"):
extra_kw = {"add_prefix_space": True} if isinstance(tokenizer, BartTokenizer) and not line.startswith(" ") else {}
tokenizer.padding_side = padding_side
return tokenizer(
[line],
max_length=max_length,
padding="max_length" if pad_to_max_length else None,
truncation=True,
return_tensors=return_tensors,
add_special_tokens=True,
**extra_kw,
)
def trim_batch(
input_ids,
pad_token_id,
attention_mask=None,
):
"""Remove columns that are populated exclusively by pad_token_id"""
keep_column_mask = input_ids.ne(pad_token_id).any(dim=0)
if attention_mask is None:
return input_ids[:, keep_column_mask]
else:
return (input_ids[:, keep_column_mask], attention_mask[:, keep_column_mask])
class Seq2SeqDataset(Dataset):
def __init__(
self,
tokenizer,
data_dir,
max_source_length,
max_target_length,
type_path="train",
n_obs=None,
src_lang=None,
tgt_lang=None,
prefix="",
):
super().__init__()
self.src_file = Path(data_dir).joinpath(type_path + ".source")
self.tgt_file = Path(data_dir).joinpath(type_path + ".target")
self.src_lens = self.get_char_lens(self.src_file)
self.max_source_length = max_source_length
self.max_target_length = max_target_length
assert min(self.src_lens) > 0, f"found empty line in {self.src_file}"
self.tokenizer = tokenizer
self.prefix = prefix
if n_obs is not None:
self.src_lens = self.src_lens[:n_obs]
self.src_lang = src_lang
self.tgt_lang = tgt_lang
def __len__(self):
return len(self.src_lens)
def __getitem__(self, index) -> Dict[str, torch.Tensor]:
index = index + 1 # linecache starts at 1
source_line = self.prefix + linecache.getline(str(self.src_file), index).rstrip("\n")
tgt_line = linecache.getline(str(self.tgt_file), index).rstrip("\n")
assert source_line, f"empty source line for index {index}"
assert tgt_line, f"empty tgt line for index {index}"
# Need to add eos token manually for T5
if isinstance(self.tokenizer, T5Tokenizer):
source_line += self.tokenizer.eos_token
tgt_line += self.tokenizer.eos_token
# Pad source and target to the right
source_tokenizer = (
self.tokenizer.question_encoder if isinstance(self.tokenizer, RagTokenizer) else self.tokenizer
)
target_tokenizer = self.tokenizer.generator if isinstance(self.tokenizer, RagTokenizer) else self.tokenizer
source_inputs = encode_line(source_tokenizer, source_line, self.max_source_length, "right")
target_inputs = encode_line(target_tokenizer, tgt_line, self.max_target_length, "right")
source_ids = source_inputs["input_ids"].squeeze()
target_ids = target_inputs["input_ids"].squeeze()
src_mask = source_inputs["attention_mask"].squeeze()
return {
"input_ids": source_ids,
"attention_mask": src_mask,
"decoder_input_ids": target_ids,
}
@staticmethod
def get_char_lens(data_file):
return [len(x) for x in Path(data_file).open().readlines()]
def collate_fn(self, batch) -> Dict[str, torch.Tensor]:
input_ids = torch.stack([x["input_ids"] for x in batch])
masks = torch.stack([x["attention_mask"] for x in batch])
target_ids = torch.stack([x["decoder_input_ids"] for x in batch])
tgt_pad_token_id = (
self.tokenizer.generator.pad_token_id
if isinstance(self.tokenizer, RagTokenizer)
else self.tokenizer.pad_token_id
)
src_pad_token_id = (
self.tokenizer.question_encoder.pad_token_id
if isinstance(self.tokenizer, RagTokenizer)
else self.tokenizer.pad_token_id
)
y = trim_batch(target_ids, tgt_pad_token_id)
source_ids, source_mask = trim_batch(input_ids, src_pad_token_id, attention_mask=masks)
batch = {
"input_ids": source_ids,
"attention_mask": source_mask,
"decoder_input_ids": y,
}
return batch
logger = getLogger(__name__)
def flatten_list(summary_ids: List[List]):
return list(itertools.chain.from_iterable(summary_ids))
def save_git_info(folder_path: str) -> None:
"""Save git information to output_dir/git_log.json"""
repo_infos = get_git_info()
save_json(repo_infos, os.path.join(folder_path, "git_log.json"))
def save_json(content, path, indent=4, **json_dump_kwargs):
with open(path, "w") as f:
json.dump(content, f, indent=indent, **json_dump_kwargs)
def load_json(path):
with open(path) as f:
return json.load(f)
def get_git_info():
repo = git.Repo(search_parent_directories=True)
repo_infos = {
"repo_id": str(repo),
"repo_sha": str(repo.head.object.hexsha),
"repo_branch": str(repo.active_branch),
"hostname": str(socket.gethostname()),
}
return repo_infos
def lmap(f: Callable, x: Iterable) -> List:
"""list(map(f, x))"""
return list(map(f, x))
def pickle_save(obj, path):
"""pickle.dump(obj, path)"""
with open(path, "wb") as f:
return pickle.dump(obj, f)
def normalize_answer(s):
"""Lower text and remove punctuation, articles and extra whitespace."""
def remove_articles(text):
return re.sub(r"\b(a|an|the)\b", " ", text)
def white_space_fix(text):
return " ".join(text.split())
def remove_punc(text):
exclude = set(string.punctuation)
return "".join(ch for ch in text if ch not in exclude)
def lower(text):
return text.lower()
return white_space_fix(remove_articles(remove_punc(lower(s))))
def f1_score(prediction, ground_truth):
prediction_tokens = normalize_answer(prediction).split()
ground_truth_tokens = normalize_answer(ground_truth).split()
common = Counter(prediction_tokens) & Counter(ground_truth_tokens)
num_same = sum(common.values())
if num_same == 0:
return 0
precision = 1.0 * num_same / len(prediction_tokens)
recall = 1.0 * num_same / len(ground_truth_tokens)
f1 = (2 * precision * recall) / (precision + recall)
return f1
def exact_match_score(prediction, ground_truth):
return normalize_answer(prediction) == normalize_answer(ground_truth)
def calculate_exact_match(output_lns: List[str], reference_lns: List[str]) -> Dict:
assert len(output_lns) == len(reference_lns)
em = 0
for hypo, pred in zip(output_lns, reference_lns):
em += exact_match_score(hypo, pred)
if len(output_lns) > 0:
em /= len(output_lns)
return {"em": em}
def is_rag_model(model_prefix):
return model_prefix.startswith("rag")
def set_extra_model_params(extra_params, hparams, config):
equivalent_param = {p: p for p in extra_params}
# T5 models don't have `dropout` param, they have `dropout_rate` instead
equivalent_param["dropout"] = "dropout_rate"
for p in extra_params:
if getattr(hparams, p, None):
if not hasattr(config, p) and not hasattr(config, equivalent_param[p]):
logger.info("config doesn't have a `{}` attribute".format(p))
delattr(hparams, p)
continue
set_p = p if hasattr(config, p) else equivalent_param[p]
setattr(config, set_p, getattr(hparams, p))
delattr(hparams, p)
return hparams, config
|
transformers/examples/research_projects/rag-end2end-retriever/utils_rag.py/0
|
{
"file_path": "transformers/examples/research_projects/rag-end2end-retriever/utils_rag.py",
"repo_id": "transformers",
"token_count": 3495
}
| 326
|
import json
import os
import shutil
import sys
import tempfile
import unittest
from unittest import TestCase
from unittest.mock import patch
import faiss
import numpy as np
from datasets import Dataset
from transformers import BartConfig, BartTokenizer, DPRConfig, DPRQuestionEncoderTokenizer, RagConfig
from transformers.file_utils import is_datasets_available, is_faiss_available, is_psutil_available, is_torch_available
from transformers.integrations import is_ray_available
from transformers.models.bert.tokenization_bert import VOCAB_FILES_NAMES as DPR_VOCAB_FILES_NAMES
from transformers.models.rag.retrieval_rag import CustomHFIndex, RagRetriever
from transformers.models.roberta.tokenization_roberta import VOCAB_FILES_NAMES as BART_VOCAB_FILES_NAMES
from transformers.testing_utils import require_ray
sys.path.append(os.path.join(os.getcwd())) # noqa: E402 # noqa: E402 # isort:skip
if is_torch_available():
from distributed_pytorch_retriever import RagPyTorchDistributedRetriever # noqa: E402 # isort:skip
else:
RagPyTorchDistributedRetriever = None
if is_ray_available():
import ray # noqa: E402 # isort:skip
from distributed_ray_retriever import RagRayDistributedRetriever, RayRetriever # noqa: E402 # isort:skip
else:
ray = None
RagRayDistributedRetriever = None
RayRetriever = None
def require_distributed_retrieval(test_case):
"""
Decorator marking a test that requires a set of dependencies necessary for pefrorm retrieval with
:class:`~transformers.RagRetriever`.
These tests are skipped when respective libraries are not installed.
"""
if not (is_datasets_available() and is_faiss_available() and is_psutil_available()):
test_case = unittest.skip("test requires Datasets, Faiss, psutil")(test_case)
return test_case
@require_distributed_retrieval
class RagRetrieverTest(TestCase):
def setUp(self):
self.tmpdirname = tempfile.mkdtemp()
self.retrieval_vector_size = 8
# DPR tok
vocab_tokens = [
"[UNK]",
"[CLS]",
"[SEP]",
"[PAD]",
"[MASK]",
"want",
"##want",
"##ed",
"wa",
"un",
"runn",
"##ing",
",",
"low",
"lowest",
]
dpr_tokenizer_path = os.path.join(self.tmpdirname, "dpr_tokenizer")
os.makedirs(dpr_tokenizer_path, exist_ok=True)
self.vocab_file = os.path.join(dpr_tokenizer_path, DPR_VOCAB_FILES_NAMES["vocab_file"])
with open(self.vocab_file, "w", encoding="utf-8") as vocab_writer:
vocab_writer.write("".join([x + "\n" for x in vocab_tokens]))
# BART tok
vocab = [
"l",
"o",
"w",
"e",
"r",
"s",
"t",
"i",
"d",
"n",
"\u0120",
"\u0120l",
"\u0120n",
"\u0120lo",
"\u0120low",
"er",
"\u0120lowest",
"\u0120newer",
"\u0120wider",
"<unk>",
]
vocab_tokens = dict(zip(vocab, range(len(vocab))))
merges = ["#version: 0.2", "\u0120 l", "\u0120l o", "\u0120lo w", "e r", ""]
self.special_tokens_map = {"unk_token": "<unk>"}
bart_tokenizer_path = os.path.join(self.tmpdirname, "bart_tokenizer")
os.makedirs(bart_tokenizer_path, exist_ok=True)
self.vocab_file = os.path.join(bart_tokenizer_path, BART_VOCAB_FILES_NAMES["vocab_file"])
self.merges_file = os.path.join(bart_tokenizer_path, BART_VOCAB_FILES_NAMES["merges_file"])
with open(self.vocab_file, "w", encoding="utf-8") as fp:
fp.write(json.dumps(vocab_tokens) + "\n")
with open(self.merges_file, "w", encoding="utf-8") as fp:
fp.write("\n".join(merges))
def get_dpr_tokenizer(self) -> DPRQuestionEncoderTokenizer:
return DPRQuestionEncoderTokenizer.from_pretrained(os.path.join(self.tmpdirname, "dpr_tokenizer"))
def get_bart_tokenizer(self) -> BartTokenizer:
return BartTokenizer.from_pretrained(os.path.join(self.tmpdirname, "bart_tokenizer"))
def tearDown(self):
shutil.rmtree(self.tmpdirname)
def get_dummy_dataset(self):
dataset = Dataset.from_dict(
{
"id": ["0", "1"],
"text": ["foo", "bar"],
"title": ["Foo", "Bar"],
"embeddings": [np.ones(self.retrieval_vector_size), 2 * np.ones(self.retrieval_vector_size)],
}
)
dataset.add_faiss_index("embeddings", string_factory="Flat", metric_type=faiss.METRIC_INNER_PRODUCT)
return dataset
def get_dummy_pytorch_distributed_retriever(
self, init_retrieval: bool, port=12345
) -> RagPyTorchDistributedRetriever:
dataset = self.get_dummy_dataset()
config = RagConfig(
retrieval_vector_size=self.retrieval_vector_size,
question_encoder=DPRConfig().to_dict(),
generator=BartConfig().to_dict(),
)
with patch("transformers.models.rag.retrieval_rag.load_dataset") as mock_load_dataset:
mock_load_dataset.return_value = dataset
retriever = RagPyTorchDistributedRetriever(
config,
question_encoder_tokenizer=self.get_dpr_tokenizer(),
generator_tokenizer=self.get_bart_tokenizer(),
)
if init_retrieval:
retriever.init_retrieval(port)
return retriever
def get_dummy_ray_distributed_retriever(self, init_retrieval: bool) -> RagRayDistributedRetriever:
# Have to run in local mode because sys.path modifications at top of
# file are not propogated to remote workers.
# https://stackoverflow.com/questions/54338013/parallel-import-a-python-file-from-sibling-folder
ray.init(local_mode=True)
config = RagConfig(
retrieval_vector_size=self.retrieval_vector_size,
question_encoder=DPRConfig().to_dict(),
generator=BartConfig().to_dict(),
)
remote_cls = ray.remote(RayRetriever)
workers = [remote_cls.remote() for _ in range(1)]
with patch("transformers.models.rag.retrieval_rag.load_dataset") as mock_load_dataset:
mock_load_dataset.return_value = self.get_dummy_dataset()
retriever = RagRayDistributedRetriever(
config,
question_encoder_tokenizer=self.get_dpr_tokenizer(),
generator_tokenizer=self.get_bart_tokenizer(),
retrieval_workers=workers,
)
if init_retrieval:
retriever.init_retrieval()
return retriever
def get_dummy_custom_hf_index_pytorch_retriever(self, init_retrieval: bool, from_disk: bool, port=12345):
dataset = self.get_dummy_dataset()
config = RagConfig(
retrieval_vector_size=self.retrieval_vector_size,
question_encoder=DPRConfig().to_dict(),
generator=BartConfig().to_dict(),
index_name="custom",
)
if from_disk:
config.passages_path = os.path.join(self.tmpdirname, "dataset")
config.index_path = os.path.join(self.tmpdirname, "index.faiss")
dataset.get_index("embeddings").save(os.path.join(self.tmpdirname, "index.faiss"))
dataset.drop_index("embeddings")
dataset.save_to_disk(os.path.join(self.tmpdirname, "dataset"))
del dataset
retriever = RagPyTorchDistributedRetriever(
config,
question_encoder_tokenizer=self.get_dpr_tokenizer(),
generator_tokenizer=self.get_bart_tokenizer(),
)
else:
retriever = RagPyTorchDistributedRetriever(
config,
question_encoder_tokenizer=self.get_dpr_tokenizer(),
generator_tokenizer=self.get_bart_tokenizer(),
index=CustomHFIndex(config.retrieval_vector_size, dataset),
)
if init_retrieval:
retriever.init_retrieval(port)
return retriever
def get_dummy_custom_hf_index_ray_retriever(self, init_retrieval: bool, from_disk: bool):
# Have to run in local mode because sys.path modifications at top of
# file are not propogated to remote workers.
# https://stackoverflow.com/questions/54338013/parallel-import-a-python-file-from-sibling-folder
ray.init(local_mode=True)
dataset = self.get_dummy_dataset()
config = RagConfig(
retrieval_vector_size=self.retrieval_vector_size,
question_encoder=DPRConfig().to_dict(),
generator=BartConfig().to_dict(),
index_name="custom",
)
remote_cls = ray.remote(RayRetriever)
workers = [remote_cls.remote() for _ in range(1)]
if from_disk:
config.passages_path = os.path.join(self.tmpdirname, "dataset")
config.index_path = os.path.join(self.tmpdirname, "index.faiss")
dataset.get_index("embeddings").save(os.path.join(self.tmpdirname, "index.faiss"))
dataset.drop_index("embeddings")
dataset.save_to_disk(os.path.join(self.tmpdirname, "dataset"))
del dataset
retriever = RagRayDistributedRetriever(
config,
question_encoder_tokenizer=self.get_dpr_tokenizer(),
generator_tokenizer=self.get_bart_tokenizer(),
retrieval_workers=workers,
index=CustomHFIndex.load_from_disk(
vector_size=config.retrieval_vector_size,
dataset_path=config.passages_path,
index_path=config.index_path,
),
)
else:
retriever = RagRayDistributedRetriever(
config,
question_encoder_tokenizer=self.get_dpr_tokenizer(),
generator_tokenizer=self.get_bart_tokenizer(),
retrieval_workers=workers,
index=CustomHFIndex(config.retrieval_vector_size, dataset),
)
if init_retrieval:
retriever.init_retrieval()
return retriever
def distributed_retriever_check(self, retriever: RagRetriever, hidden_states: np.array, n_docs: int) -> None:
retrieved_doc_embeds, doc_ids, doc_dicts = retriever.retrieve(hidden_states, n_docs=n_docs)
self.assertEqual(retrieved_doc_embeds.shape, (2, n_docs, self.retrieval_vector_size))
self.assertEqual(len(doc_dicts), 2)
self.assertEqual(sorted(doc_dicts[0]), ["embeddings", "id", "text", "title"])
self.assertEqual(len(doc_dicts[0]["id"]), n_docs)
self.assertEqual(doc_dicts[0]["id"][0], "1") # max inner product is reached with second doc
self.assertEqual(doc_dicts[1]["id"][0], "0") # max inner product is reached with first doc
self.assertListEqual(doc_ids.tolist(), [[1], [0]])
def test_pytorch_distributed_retriever_retrieve(self):
n_docs = 1
hidden_states = np.array(
[np.ones(self.retrieval_vector_size), -np.ones(self.retrieval_vector_size)], dtype=np.float32
)
self.distributed_retriever_check(
self.get_dummy_pytorch_distributed_retriever(init_retrieval=True), hidden_states, n_docs
)
def test_custom_hf_index_pytorch_retriever_retrieve(self):
n_docs = 1
hidden_states = np.array(
[np.ones(self.retrieval_vector_size), -np.ones(self.retrieval_vector_size)], dtype=np.float32
)
self.distributed_retriever_check(
self.get_dummy_custom_hf_index_pytorch_retriever(init_retrieval=True, from_disk=False),
hidden_states,
n_docs,
)
def test_custom_pytorch_distributed_retriever_retrieve_from_disk(self):
n_docs = 1
hidden_states = np.array(
[np.ones(self.retrieval_vector_size), -np.ones(self.retrieval_vector_size)], dtype=np.float32
)
self.distributed_retriever_check(
self.get_dummy_custom_hf_index_pytorch_retriever(init_retrieval=True, from_disk=True),
hidden_states,
n_docs,
)
@require_ray
def test_ray_distributed_retriever_retrieve(self):
n_docs = 1
hidden_states = np.array(
[np.ones(self.retrieval_vector_size), -np.ones(self.retrieval_vector_size)], dtype=np.float32
)
self.distributed_retriever_check(
self.get_dummy_ray_distributed_retriever(init_retrieval=True), hidden_states, n_docs
)
ray.shutdown()
@require_ray
def test_custom_hf_index_ray_retriever_retrieve(self):
n_docs = 1
hidden_states = np.array(
[np.ones(self.retrieval_vector_size), -np.ones(self.retrieval_vector_size)], dtype=np.float32
)
with self.assertRaises(ValueError):
self.distributed_retriever_check(
self.get_dummy_custom_hf_index_ray_retriever(init_retrieval=True, from_disk=False),
hidden_states,
n_docs,
)
ray.shutdown()
@require_ray
def test_custom_ray_distributed_retriever_retrieve_from_disk(self):
n_docs = 1
hidden_states = np.array(
[np.ones(self.retrieval_vector_size), -np.ones(self.retrieval_vector_size)], dtype=np.float32
)
self.distributed_retriever_check(
self.get_dummy_custom_hf_index_ray_retriever(init_retrieval=True, from_disk=True), hidden_states, n_docs
)
ray.shutdown()
|
transformers/examples/research_projects/rag/test_distributed_retriever.py/0
|
{
"file_path": "transformers/examples/research_projects/rag/test_distributed_retriever.py",
"repo_id": "transformers",
"token_count": 6637
}
| 327
|
# as due to their complexity multi-gpu tests could impact other tests, and to aid debug we have those in a separate module.
import os
import sys
from pathlib import Path
import torch
from transformers.testing_utils import TestCasePlus, execute_subprocess_async, require_torch_multi_gpu
from utils import load_json
CUDA_AVAILABLE = torch.cuda.is_available()
ARTICLES = [" Sam ate lunch today.", "Sams lunch ingredients."]
SUMMARIES = ["A very interesting story about what I ate for lunch.", "Avocado, celery, turkey, coffee"]
CHEAP_ARGS = {
"max_tokens_per_batch": None,
"supervise_forward": True,
"normalize_hidden": True,
"label_smoothing": 0.2,
"eval_max_gen_length": None,
"eval_beams": 1,
"val_metric": "loss",
"save_top_k": 1,
"adafactor": True,
"early_stopping_patience": 2,
"logger_name": "default",
"length_penalty": 0.5,
"cache_dir": "",
"task": "summarization",
"num_workers": 2,
"alpha_hid": 0,
"freeze_embeds": True,
"enc_only": False,
"tgt_suffix": "",
"resume_from_checkpoint": None,
"sortish_sampler": True,
"student_decoder_layers": 1,
"val_check_interval": 1.0,
"output_dir": "",
"fp16": False, # TODO(SS): set this to CUDA_AVAILABLE if ci installs apex or start using native amp
"no_teacher": False,
"fp16_opt_level": "O1",
"gpus": 1 if CUDA_AVAILABLE else 0,
"n_tpu_cores": 0,
"max_grad_norm": 1.0,
"do_train": True,
"do_predict": True,
"accumulate_grad_batches": 1,
"server_ip": "",
"server_port": "",
"seed": 42,
"model_name_or_path": "sshleifer/bart-tiny-random",
"config_name": "",
"tokenizer_name": "facebook/bart-large",
"do_lower_case": False,
"learning_rate": 0.3,
"lr_scheduler": "linear",
"weight_decay": 0.0,
"adam_epsilon": 1e-08,
"warmup_steps": 0,
"max_epochs": 1,
"train_batch_size": 2,
"eval_batch_size": 2,
"max_source_length": 12,
"max_target_length": 12,
"val_max_target_length": 12,
"test_max_target_length": 12,
"fast_dev_run": False,
"no_cache": False,
"n_train": -1,
"n_val": -1,
"n_test": -1,
"student_encoder_layers": 1,
"freeze_encoder": False,
"auto_scale_batch_size": False,
"overwrite_output_dir": False,
"student": None,
}
def _dump_articles(path: Path, articles: list):
content = "\n".join(articles)
Path(path).open("w").writelines(content)
def make_test_data_dir(tmp_dir):
for split in ["train", "val", "test"]:
_dump_articles(os.path.join(tmp_dir, f"{split}.source"), ARTICLES)
_dump_articles(os.path.join(tmp_dir, f"{split}.target"), SUMMARIES)
return tmp_dir
class TestSummarizationDistillerMultiGPU(TestCasePlus):
@classmethod
def setUpClass(cls):
return cls
@require_torch_multi_gpu
def test_multi_gpu(self):
updates = {
"no_teacher": True,
"freeze_encoder": True,
"gpus": 2,
"overwrite_output_dir": True,
"sortish_sampler": True,
}
self._test_distiller_cli_fork(updates, check_contents=False)
def _test_distiller_cli_fork(self, updates, check_contents=True):
default_updates = {
"label_smoothing": 0.0,
"early_stopping_patience": -1,
"train_batch_size": 1,
"eval_batch_size": 2,
"max_epochs": 2,
"alpha_mlm": 0.2,
"alpha_ce": 0.8,
"do_predict": True,
"model_name_or_path": "sshleifer/tinier_bart",
"teacher": CHEAP_ARGS["model_name_or_path"],
"val_check_interval": 0.5,
}
default_updates.update(updates)
args_d: dict = CHEAP_ARGS.copy()
tmp_dir = make_test_data_dir(tmp_dir=self.get_auto_remove_tmp_dir())
output_dir = self.get_auto_remove_tmp_dir()
args_d.update(data_dir=tmp_dir, output_dir=output_dir, **default_updates)
def convert(k, v):
if k in ["tgt_suffix", "server_ip", "server_port", "out", "n_tpu_cores"]:
return ""
if v is False or v is None:
return ""
if v is True: # or len(str(v))==0:
return f"--{k}"
return f"--{k}={v}"
cli_args = [x for x in (convert(k, v) for k, v in args_d.items()) if len(x)]
cmd = [sys.executable, f"{self.test_file_dir}/distillation.py"] + cli_args
execute_subprocess_async(cmd, env=self.get_env())
contents = os.listdir(output_dir)
contents = {os.path.basename(p) for p in contents}
ckpt_files = [p for p in contents if p.endswith("ckpt")]
assert len(ckpt_files) > 0
self.assertIn("test_generations.txt", contents)
self.assertIn("test_results.txt", contents)
# get the following from the module, (we don't have access to `model` here)
metrics_save_path = os.path.join(output_dir, "metrics.json")
val_metric = "rouge2"
metrics = load_json(metrics_save_path)
# {'test': [{'test_avg_loss': 10.63731575012207, 'test_avg_rouge1': 0.0, 'test_avg_rouge2': 0.0, 'test_avg_rougeL': 0.0, 'test_avg_gen_time': 0.1822289228439331, 'test_avg_gen_len': 142.0, 'step_count': 1}]}
print(metrics)
last_step_stats = metrics["val"][-1]
self.assertGreaterEqual(last_step_stats["val_avg_gen_time"], 0.01)
self.assertIsInstance(last_step_stats[f"val_avg_{val_metric}"], float)
self.assertEqual(len(metrics["test"]), 1)
desired_n_evals = int(args_d["max_epochs"] * (1 / args_d["val_check_interval"]) / 2 + 1)
self.assertEqual(len(metrics["val"]), desired_n_evals)
|
transformers/examples/research_projects/seq2seq-distillation/_test_seq2seq_examples_multi_gpu.py/0
|
{
"file_path": "transformers/examples/research_projects/seq2seq-distillation/_test_seq2seq_examples_multi_gpu.py",
"repo_id": "transformers",
"token_count": 2691
}
| 328
|
#!/usr/bin/env python
# coding=utf-8
# Copyright 2022 The HuggingFace Inc. team. 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
"""Fine-tuning a 🤗 Transformers pretrained speech model on the XTREME-S benchmark tasks"""
import json
import logging
import os
import re
import sys
from collections import OrderedDict, defaultdict
from dataclasses import dataclass, field
from typing import Dict, List, Optional, Union
import datasets
import numpy as np
import torch
from datasets import DatasetDict, load_dataset, load_metric
import transformers
from transformers import (
AutoConfig,
AutoFeatureExtractor,
AutoModelForAudioClassification,
AutoModelForCTC,
AutoModelForSpeechSeq2Seq,
AutoProcessor,
AutoTokenizer,
HfArgumentParser,
Seq2SeqTrainer,
Seq2SeqTrainingArguments,
Trainer,
set_seed,
)
from transformers.trainer_utils import get_last_checkpoint, is_main_process
from transformers.utils import check_min_version
from transformers.utils.versions import require_version
# Will error if the minimal version of Transformers is not installed. Remove at your own risks.
check_min_version("4.18.0.dev0")
require_version("datasets>=1.18.0", "To fix: pip install -r examples/pytorch/speech-recognition/requirements.txt")
logger = logging.getLogger(__name__)
def list_field(default=None, metadata=None):
return field(default_factory=lambda: default, metadata=metadata)
TASK_TO_TARGET_COLUMN_NAME = {
"fleurs-asr": "transcription",
"fleurs-lang_id": "lang_id",
"mls": "transcription",
"voxpopuli": "transcription",
"covost2": "translation",
"minds14": "intent_class",
"babel": "transcription",
}
@dataclass
class ModelArguments:
"""
Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
"""
model_name_or_path: str = field(
metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"}
)
tokenizer_name_or_path: Optional[str] = field(
default=None,
metadata={"help": "Path to pretrained tokenizer or tokenizer identifier from huggingface.co/models"},
)
cache_dir: Optional[str] = field(
default=None,
metadata={
"help": "Where do you want to store the pretrained models and datasets downloaded from huggingface.co"
},
)
freeze_feature_encoder: bool = field(
default=True, metadata={"help": "Whether to freeze the feature encoder layers of the model."}
)
attention_dropout: float = field(
default=0.0, metadata={"help": "The dropout ratio for the attention probabilities."}
)
activation_dropout: float = field(
default=0.0, metadata={"help": "The dropout ratio for activations inside the fully connected layer."}
)
feat_proj_dropout: float = field(default=0.0, metadata={"help": "The dropout ratio for the projected features."})
hidden_dropout: float = field(
default=0.0,
metadata={
"help": "The dropout probability for all fully connected layers in the embeddings, encoder, and pooler."
},
)
final_dropout: float = field(
default=0.0,
metadata={"help": "The dropout probability for the final projection layer."},
)
mask_time_prob: float = field(
default=0.05,
metadata={
"help": (
"Probability of each feature vector along the time axis to be chosen as the start of the vector "
"span to be masked. Approximately ``mask_time_prob * sequence_length // mask_time_length`` feature "
"vectors will be masked along the time axis."
)
},
)
mask_time_length: int = field(
default=10,
metadata={"help": "Length of vector span to mask along the time axis."},
)
mask_feature_prob: float = field(
default=0.0,
metadata={
"help": (
"Probability of each feature vector along the feature axis to be chosen as the start of the vectorspan"
" to be masked. Approximately ``mask_feature_prob * sequence_length // mask_feature_length`` feature"
" bins will be masked along the time axis."
)
},
)
mask_feature_length: int = field(
default=10,
metadata={"help": "Length of vector span to mask along the feature axis."},
)
layerdrop: float = field(default=0.0, metadata={"help": "The LayerDrop probability."})
ctc_zero_infinity: bool = field(
default=False,
metadata={"help": "Whether to zero infinite losses and the associated gradients of `torch.nn.CTCLoss`."},
)
ctc_loss_reduction: Optional[str] = field(
default="mean", metadata={"help": "The way the ctc loss should be reduced. Should be one of 'mean' or 'sum'."}
)
@dataclass
class DataTrainingArguments:
"""
Arguments pertaining to what data we are going to input our model for training and eval.
Using `HfArgumentParser` we can turn this class
into argparse arguments to be able to specify them on
the command line.
"""
dataset_name: str = field(
default="google/xtreme_s",
metadata={"help": "The name of the dataset to use (via the datasets library). Defaults to 'google/xtreme_s'"},
)
task: str = field(
default=None,
metadata={
"help": (
"The task name of the benchmark to use (via the datasets library). Should be on of: "
"'fleurs-asr', 'mls', 'voxpopuli', 'covost2', 'minds14', 'fleurs-lang_id', 'babel'."
)
},
)
language: str = field(
default="all",
metadata={"help": "The language id as defined in the datasets config name or `all` for all languages."},
)
language_group: str = field(
default=None,
metadata={
"help": (
"The language group to select a subset of languages to train on. "
"This option is only used the 'fleurs-asr' task. Should be one of: "
"'western_european_we', 'eastern_european_ee', 'central_asia_middle_north_african_cmn', "
"'sub_saharan_african_ssa', 'south_asian_sa', 'south_east_asian_sea', 'chinese_japanase_korean_cjk'."
)
},
)
train_split_name: str = field(
default="train",
metadata={
"help": "The name of the training dataset split to use (via the datasets library). Defaults to 'train'"
},
)
eval_split_name: str = field(
default="validation",
metadata={
"help": (
"The name of the evaluation dataset split to use (via the datasets library). Defaults to 'validation'"
)
},
)
predict_split_name: str = field(
default="test",
metadata={
"help": "The name of the prediction dataset split to use (via the datasets library). Defaults to 'test'"
},
)
audio_column_name: str = field(
default="audio",
metadata={"help": "The name of the dataset column containing the audio data. Defaults to 'audio'"},
)
target_column_name: str = field(
default=None,
metadata={
"help": (
"The name of the dataset column containing the target data (transcription/translation/label). If None,"
" the name will be inferred from the task. Defaults to None."
)
},
)
overwrite_cache: bool = field(
default=False, metadata={"help": "Overwrite the cached preprocessed datasets or not."}
)
preprocessing_num_workers: Optional[int] = field(
default=None,
metadata={"help": "The number of processes to use for the preprocessing."},
)
max_train_samples: Optional[int] = field(
default=None,
metadata={
"help": (
"For debugging purposes or quicker training, truncate the number of training examples to this "
"value if set."
)
},
)
max_eval_samples: Optional[int] = field(
default=None,
metadata={
"help": (
"For debugging purposes or quicker training, truncate the number of validation examples to this "
"value if set."
)
},
)
max_predict_samples: Optional[int] = field(
default=None,
metadata={
"help": (
"For debugging purposes or quicker training, truncate the number of prediction examples to this "
"value if set."
)
},
)
chars_to_ignore: Optional[List[str]] = list_field(
default=', ? . ! - ; : " “ % ‘ ” �'.split(" "),
metadata={"help": "A list of characters to remove from the transcripts."},
)
max_duration_in_seconds: float = field(
default=30.0,
metadata={
"help": (
"Filter audio files that are longer than `max_duration_in_seconds` seconds to"
" 'max_duration_in_seconds`"
)
},
)
min_duration_in_seconds: float = field(
default=0.0, metadata={"help": "Filter audio files that are shorter than `min_duration_in_seconds` seconds"}
)
preprocessing_only: bool = field(
default=False,
metadata={
"help": (
"Whether to only do data preprocessing and skip training. This is especially useful when data"
" preprocessing errors out in distributed training due to timeout. In this case, one should run the"
" preprocessing in a non-distributed setup with `preprocessing_only=True` so that the cached datasets"
" can consequently be loaded in distributed training"
)
},
)
use_auth_token: bool = field(
default=False,
metadata={
"help": (
"If :obj:`True`, will use the token generated when running"
":obj:`huggingface-cli login` as HTTP bearer authorization for remote files."
)
},
)
unk_token: str = field(
default="[UNK]",
metadata={"help": "The unk token for the tokenizer"},
)
pad_token: str = field(
default="[PAD]",
metadata={"help": "The padding token for the tokenizer"},
)
word_delimiter_token: str = field(
default="|",
metadata={"help": "The word delimiter token for the tokenizer"},
)
phoneme_language: Optional[str] = field(
default=None,
metadata={
"help": (
"The target language that should be used be"
" passed to the tokenizer for tokenization. Note that"
" this is only relevant if the model classifies the"
" input audio to a sequence of phoneme sequences."
)
},
)
per_lang_metrics: bool = field(
default=True,
metadata={
"help": (
"If `True`, compute the test metrics separately for each language, and average the results. "
"If `False` compute the average test metrics in a single pass for all languages at once."
)
},
)
@dataclass
class SpeechDataCollatorWithPadding:
processor: AutoProcessor
decoder_start_token_id: Optional[int] = None
padding: Union[bool, str] = "longest"
pad_labels: Optional[int] = True
pad_to_multiple_of: Optional[int] = None
pad_to_multiple_of_labels: Optional[int] = None
def __call__(self, features: List[Dict[str, Union[List[int], torch.Tensor]]]) -> Dict[str, torch.Tensor]:
# split inputs and labels since they have to be of different lengths and need
# different padding methods
input_features = [{"input_values": feature["input_values"]} for feature in features]
batch = self.processor.pad(
input_features,
padding=self.padding,
pad_to_multiple_of=self.pad_to_multiple_of,
return_tensors="pt",
)
if self.pad_labels:
label_features = [{"input_ids": feature["labels"]} for feature in features]
labels_batch = self.processor.pad(
labels=label_features,
padding=self.padding,
pad_to_multiple_of=self.pad_to_multiple_of_labels,
return_tensors="pt",
)
# replace padding with -100 to ignore loss correctly
labels = labels_batch["input_ids"].masked_fill(labels_batch.attention_mask.ne(1), -100)
# if bos token is appended in previous tokenization step,
# cut bos token here as it's append later anyways
if (
self.decoder_start_token_id is not None
and (labels[:, 0] == self.decoder_start_token_id).all().cpu().item()
):
labels = labels[:, 1:]
batch["labels"] = labels
else:
batch["labels"] = torch.tensor([feature["labels"] for feature in features])
return batch
def create_vocabulary_from_data(
datasets: DatasetDict,
word_delimiter_token: Optional[str] = None,
unk_token: Optional[str] = None,
pad_token: Optional[str] = None,
):
# Given training and test labels create vocabulary
def extract_all_chars(batch):
all_text = " ".join(batch["target_text"])
vocab = list(set(all_text))
return {"vocab": [vocab], "all_text": [all_text]}
vocabs = datasets.map(
extract_all_chars,
batched=True,
batch_size=-1,
keep_in_memory=True,
remove_columns=datasets["train"].column_names,
)
# take union of all unique characters in each dataset
vocab_set = (
(set(vocabs["train"]["vocab"][0]) if "train" in vocabs else set())
| (set(vocabs["eval"]["vocab"][0]) if "eval" in vocabs else set())
| (set(vocabs["predict"]["vocab"][0]) if "predict" in vocabs else set())
)
vocab_dict = {v: k for k, v in enumerate(sorted(vocab_set))}
# replace white space with delimiter token
if word_delimiter_token is not None:
vocab_dict[word_delimiter_token] = vocab_dict[" "]
del vocab_dict[" "]
# add unk and pad token
if unk_token is not None:
vocab_dict[unk_token] = len(vocab_dict)
if pad_token is not None:
vocab_dict[pad_token] = len(vocab_dict)
return vocab_dict
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser((ModelArguments, DataTrainingArguments, Seq2SeqTrainingArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# If we pass only one argument to the script and it's the path to a json file,
# let's parse it to get our arguments.
model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
else:
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
# Detecting last checkpoint.
last_checkpoint = None
if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
last_checkpoint = get_last_checkpoint(training_args.output_dir)
if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
"Use --overwrite_output_dir to overcome."
)
elif last_checkpoint is not None:
logger.info(
f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
"the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
handlers=[logging.StreamHandler(sys.stdout)],
)
logger.setLevel(logging.INFO if is_main_process(training_args.local_rank) else logging.WARN)
# Log on each process the small summary:
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}, "
f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
# Set the verbosity to info of the Transformers logger (on main process only):
if is_main_process(training_args.local_rank):
transformers.utils.logging.set_verbosity_info()
logger.info("Training/evaluation parameters %s", training_args)
# Set seed before initializing model.
set_seed(training_args.seed)
# 1. First, let's load the dataset
raw_datasets = DatasetDict()
task_name = data_args.task
lang_id = data_args.language
if task_name is None:
raise ValueError(
"Set --task should be set to '<xtreme_s_task>' (e.g. 'fleurs-asr', 'mls', 'covost2', 'minds14') "
)
if lang_id is None:
raise ValueError(
"Set --language should be set to the language id of the sub dataset "
"config to be used (e.g. 'pl', 'en.tr', 'fr-FR') or 'all'"
" for multi-lingual fine-tuning."
)
if data_args.language_group is not None:
if data_args.task != "fleurs-asr":
raise ValueError("--language_group should only be used with --task=fleurs-asr")
if data_args.language != "all":
raise ValueError("--language_group should only be used with --language=all")
if data_args.target_column_name is None:
target_column_name = TASK_TO_TARGET_COLUMN_NAME[task_name]
else:
target_column_name = data_args.target_column_name
# here we differentiate between tasks with text as the target and classification tasks
is_text_target = target_column_name in ("transcription", "translation")
config_name = ".".join([task_name.split("-")[0], lang_id])
if training_args.do_train:
raw_datasets["train"] = load_dataset(
data_args.dataset_name,
config_name,
split=data_args.train_split_name,
token=data_args.use_auth_token,
cache_dir=model_args.cache_dir,
)
if data_args.audio_column_name not in raw_datasets["train"].column_names:
raise ValueError(
f"--audio_column_name '{data_args.audio_column_name}' not found in dataset '{data_args.dataset_name}'."
" Make sure to set `--audio_column_name` to the correct audio column - one of"
f" {', '.join(raw_datasets['train'].column_names)}."
)
if target_column_name not in raw_datasets["train"].column_names:
raise ValueError(
f"--target_column_name {target_column_name} not found in dataset '{data_args.dataset_name}'. "
"Make sure to set `--target_column_name` to the correct text column - one of "
f"{', '.join(raw_datasets['train'].column_names)}."
)
if data_args.max_train_samples is not None:
raw_datasets["train"] = raw_datasets["train"].select(range(data_args.max_train_samples))
if training_args.do_eval:
raw_datasets["eval"] = load_dataset(
data_args.dataset_name,
config_name,
split=data_args.eval_split_name,
token=data_args.use_auth_token,
cache_dir=model_args.cache_dir,
)
if data_args.max_eval_samples is not None:
raw_datasets["eval"] = raw_datasets["eval"].select(range(data_args.max_eval_samples))
if training_args.do_predict:
raw_datasets["predict"] = load_dataset(
data_args.dataset_name,
config_name,
split=data_args.predict_split_name,
token=data_args.use_auth_token,
cache_dir=model_args.cache_dir,
)
if data_args.max_predict_samples is not None:
raw_datasets["predict"] = raw_datasets["predict"].select(range(data_args.max_predict_samples))
lang_list = next(iter(raw_datasets.values())).features["lang_id"].names
if not is_text_target:
label_list = next(iter(raw_datasets.values())).features[target_column_name].names
num_labels = len(label_list)
num_workers = data_args.preprocessing_num_workers
lang_group = data_args.language_group
if lang_group is not None:
with training_args.main_process_first(desc="language group filter"):
lang_group_id = next(iter(raw_datasets.values())).features["lang_group_id"].str2int(lang_group)
raw_datasets = raw_datasets.filter(
lambda lang_group: lang_group == lang_group_id,
num_proc=num_workers,
input_columns=["lang_group_id"],
)
# 2. We remove some special characters from the datasets
# that make training complicated and do not help in transcribing the speech
# E.g. characters, such as `,` and `.` do not really have an acoustic characteristic
# that could be easily picked up by the model
chars_to_ignore_regex = (
f'[{"".join(data_args.chars_to_ignore)}]' if data_args.chars_to_ignore is not None else None
)
def remove_special_characters(batch):
if chars_to_ignore_regex is not None:
batch["target_text"] = re.sub(chars_to_ignore_regex, "", batch[target_column_name]).lower() + " "
else:
batch["target_text"] = batch[target_column_name].lower() + " "
return batch
if is_text_target:
with training_args.main_process_first(desc="dataset map special characters removal"):
raw_datasets = raw_datasets.map(
remove_special_characters,
remove_columns=[target_column_name],
desc="remove special characters from datasets",
)
# save special tokens for tokenizer
word_delimiter_token = data_args.word_delimiter_token
unk_token = data_args.unk_token
pad_token = data_args.pad_token
# 3. Next, let's load the config as we might need it to create
# the tokenizer
config = AutoConfig.from_pretrained(
model_args.model_name_or_path, cache_dir=model_args.cache_dir, token=data_args.use_auth_token
)
if is_text_target:
# 4. (Optional, for ASR and translation) If no tokenizer file is defined,
# we create the vocabulary of the model by extracting all unique characters from
# the training and evaluation datasets
# We need to make sure that only first rank saves vocabulary
# make sure all processes wait until vocab is created
tokenizer_name_or_path = model_args.tokenizer_name_or_path
tokenizer_kwargs = {}
if tokenizer_name_or_path is None:
# save vocab in training output dir
tokenizer_name_or_path = training_args.output_dir
vocab_file = os.path.join(tokenizer_name_or_path, "vocab.json")
with training_args.main_process_first():
if training_args.overwrite_output_dir and os.path.isfile(vocab_file):
os.remove(vocab_file)
with training_args.main_process_first(desc="dataset map vocabulary creation"):
if not os.path.isfile(vocab_file):
os.makedirs(tokenizer_name_or_path, exist_ok=True)
vocab_dict = create_vocabulary_from_data(
raw_datasets,
word_delimiter_token=word_delimiter_token,
unk_token=unk_token,
pad_token=pad_token,
)
# save vocab dict to be loaded into tokenizer
with open(vocab_file, "w") as file:
json.dump(vocab_dict, file)
# if tokenizer has just been created
# it is defined by `tokenizer_class` if present in config else by `model_type`
if not config.is_encoder_decoder:
tokenizer_kwargs = {
"config": config if config.tokenizer_class is not None else None,
"tokenizer_type": config.model_type if config.tokenizer_class is None else None,
"unk_token": unk_token,
"pad_token": pad_token,
"word_delimiter_token": word_delimiter_token,
}
else:
tokenizer_kwargs = {}
# 5. Now we can instantiate the feature extractor, tokenizer and model
# Note for distributed training, the .from_pretrained methods guarantee that only
# one local process can concurrently download model & vocab.
# load feature_extractor and tokenizer
if is_text_target:
tokenizer = AutoTokenizer.from_pretrained(
tokenizer_name_or_path,
token=data_args.use_auth_token,
**tokenizer_kwargs,
)
feature_extractor = AutoFeatureExtractor.from_pretrained(
model_args.model_name_or_path, cache_dir=model_args.cache_dir, token=data_args.use_auth_token
)
# adapt config
# (speech translation requires pre-configured seq2seq models)
if task_name != "covost2":
config.update(
{
"feat_proj_dropout": model_args.feat_proj_dropout,
"attention_dropout": model_args.attention_dropout,
"hidden_dropout": model_args.hidden_dropout,
"final_dropout": model_args.final_dropout,
"mask_time_prob": model_args.mask_time_prob,
"mask_time_length": model_args.mask_time_length,
"mask_feature_prob": model_args.mask_feature_prob,
"mask_feature_length": model_args.mask_feature_length,
"gradient_checkpointing": training_args.gradient_checkpointing,
"layerdrop": model_args.layerdrop,
"ctc_zero_infinity": model_args.ctc_zero_infinity,
"ctc_loss_reduction": model_args.ctc_loss_reduction,
"activation_dropout": model_args.activation_dropout,
}
)
if training_args.do_train:
if is_text_target:
config.pad_token_id = tokenizer.pad_token_id
config.vocab_size = len(tokenizer)
else:
label_to_id = {v: i for i, v in enumerate(label_list)}
config.label2id = label_to_id
config.id2label = {id: label for label, id in label_to_id.items()}
config.num_labels = num_labels
# create model
if target_column_name == "transcription":
model = AutoModelForCTC.from_pretrained(
model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
config=config,
token=data_args.use_auth_token,
)
elif config.is_encoder_decoder:
model = AutoModelForSpeechSeq2Seq.from_pretrained(
model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
config=config,
token=data_args.use_auth_token,
)
if model.config.decoder_start_token_id is None:
raise ValueError("Make sure that `config.decoder_start_token_id` is correctly defined")
else:
model = AutoModelForAudioClassification.from_pretrained(
model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
config=config,
token=data_args.use_auth_token,
)
# freeze encoder
if model_args.freeze_feature_encoder:
model.freeze_feature_encoder()
# 6. Now we preprocess the datasets including loading the audio, resampling and normalization
# Thankfully, `datasets` takes care of automatically loading and resampling the audio,
# so that we just need to set the correct target sampling rate and normalize the input
# via the `feature_extractor`
# make sure that dataset decodes audio with correct sampling rate
dataset_sampling_rate = next(iter(raw_datasets.values())).features[data_args.audio_column_name].sampling_rate
if dataset_sampling_rate != feature_extractor.sampling_rate:
raw_datasets = raw_datasets.cast_column(
data_args.audio_column_name, datasets.features.Audio(sampling_rate=feature_extractor.sampling_rate)
)
# derive max & min input length for sample rate & max duration
max_input_length = data_args.max_duration_in_seconds * feature_extractor.sampling_rate
min_input_length = data_args.min_duration_in_seconds * feature_extractor.sampling_rate
audio_column_name = data_args.audio_column_name
# `phoneme_language` is only relevant if the model is fine-tuned on phoneme classification
phoneme_language = data_args.phoneme_language
# Preprocessing the datasets.
# We need to read the audio files as arrays and tokenize the targets.
def prepare_dataset(batch):
# load audio
sample = batch[audio_column_name]
inputs = feature_extractor(sample["array"], sampling_rate=sample["sampling_rate"])
batch["input_values"] = inputs.input_values[0]
batch["length"] = len(batch["input_values"])
# encode targets
additional_kwargs = {}
if phoneme_language is not None:
additional_kwargs["phonemizer_lang"] = phoneme_language
if is_text_target:
batch["labels"] = tokenizer(batch["target_text"], **additional_kwargs).input_ids
else:
batch["labels"] = batch[target_column_name]
batch["lang"] = batch["lang_id"]
return batch
with training_args.main_process_first(desc="dataset map preprocessing"):
vectorized_datasets = raw_datasets.map(
prepare_dataset,
remove_columns=next(iter(raw_datasets.values())).column_names,
num_proc=num_workers,
desc="preprocess datasets",
)
if training_args.do_train:
def is_audio_in_length_range(length):
return length > min_input_length and length < max_input_length
# filter data that is shorter than min_input_length
vectorized_datasets["train"] = vectorized_datasets["train"].filter(
is_audio_in_length_range,
num_proc=num_workers,
input_columns=["length"],
)
# 7. Next, we can prepare for the training step.
# Let's use the appropriate XTREME-S evaluation metric,
# instantiate a data collator and the trainer
# Define evaluation metrics during training, *i.e.* word error rate, character error rate
eval_metric = load_metric("xtreme_s", task_name)
# for large datasets it is advised to run the preprocessing on a
# single machine first with ``args.preprocessing_only`` since there will mostly likely
# be a timeout when running the script in distributed mode.
# In a second step ``args.preprocessing_only`` can then be set to `False` to load the
# cached dataset
if data_args.preprocessing_only:
logger.info(f"Data preprocessing finished. Files cached at {vectorized_datasets.cache_files}")
return
def asr_logits_argmax(logits, labels):
return logits.argmax(dim=-1)
def compute_asr_metric(pred):
pred.label_ids[pred.label_ids == -100] = tokenizer.pad_token_id
pred_str = tokenizer.batch_decode(pred.predictions)
# we do not want to group tokens when computing the metrics
label_str = tokenizer.batch_decode(pred.label_ids, group_tokens=False)
metric = eval_metric.compute(predictions=pred_str, references=label_str)
return metric
def compute_classification_metric(pred):
pred_ids = np.argmax(pred.predictions, axis=1)
metric = eval_metric.compute(predictions=pred_ids, references=pred.label_ids)
return metric
# Now save everything to be able to create a single processor later
if is_main_process(training_args.local_rank):
# save feature extractor, tokenizer and config
feature_extractor.save_pretrained(training_args.output_dir)
if is_text_target:
tokenizer.save_pretrained(training_args.output_dir)
config.save_pretrained(training_args.output_dir)
# wait until configs are saved in the main process before loading the processor
if training_args.local_rank != -1:
torch.distributed.barrier()
if is_text_target:
processor = AutoProcessor.from_pretrained(training_args.output_dir)
else:
processor = AutoFeatureExtractor.from_pretrained(training_args.output_dir)
# Instantiate custom data collator
data_collator = SpeechDataCollatorWithPadding(processor=processor, pad_labels=is_text_target)
# Initialize Trainer
if target_column_name == "translation":
trainer = Seq2SeqTrainer(
model=model,
data_collator=data_collator,
args=training_args,
preprocess_logits_for_metrics=asr_logits_argmax if training_args.predict_with_generate else None,
compute_metrics=compute_asr_metric if training_args.predict_with_generate else None,
train_dataset=vectorized_datasets["train"] if training_args.do_train else None,
eval_dataset=vectorized_datasets["eval"] if training_args.do_eval else None,
tokenizer=feature_extractor,
)
else:
trainer = Trainer(
model=model,
data_collator=data_collator,
args=training_args,
preprocess_logits_for_metrics=asr_logits_argmax if is_text_target else None,
compute_metrics=compute_asr_metric if is_text_target else compute_classification_metric,
train_dataset=vectorized_datasets["train"] if training_args.do_train else None,
eval_dataset=vectorized_datasets["eval"] if training_args.do_eval else None,
tokenizer=feature_extractor,
)
# 8. Finally, we can start training
# Training
if training_args.do_train:
# use last checkpoint if exist
if last_checkpoint is not None:
checkpoint = last_checkpoint
elif os.path.isdir(model_args.model_name_or_path):
checkpoint = model_args.model_name_or_path
else:
checkpoint = None
train_result = trainer.train(resume_from_checkpoint=checkpoint)
trainer.save_model()
metrics = train_result.metrics
max_train_samples = (
data_args.max_train_samples
if data_args.max_train_samples is not None
else len(vectorized_datasets["train"])
)
metrics["train_samples"] = min(max_train_samples, len(vectorized_datasets["train"]))
trainer.log_metrics("train", metrics)
trainer.save_metrics("train", metrics)
trainer.save_state()
# Evaluation on the test set
results = {}
if training_args.do_predict:
logger.info(f"*** Evaluating on the `{data_args.predict_split_name}` set ***")
if data_args.per_lang_metrics:
# separate the `test` dataset into language-specific subsets and compute metrics for each of them
metrics = {}
average_metrics = defaultdict(list)
for lang_id in range(len(lang_list)):
lang_name = lang_list[lang_id]
with training_args.main_process_first(desc="per-language dataset filter"):
lang_dataset = vectorized_datasets["predict"].filter(
lambda lang: lang == lang_id,
num_proc=num_workers,
input_columns=["lang"],
)
lang_metrics = trainer.evaluate(lang_dataset)
redundant_metrics = ["eval_runtime", "eval_samples_per_second", "eval_steps_per_second", "eval_epoch"]
for metric_name, value in lang_metrics.items():
average_metrics[metric_name].append(value)
if metric_name not in redundant_metrics:
metrics[f"{metric_name}_{lang_name}"] = value
for metric_name, value in average_metrics.items():
metrics[metric_name] = np.mean(value)
else:
metrics = trainer.evaluate(vectorized_datasets["predict"])
max_predict_samples = (
data_args.max_predict_samples
if data_args.max_predict_samples is not None
else len(vectorized_datasets["predict"])
)
metrics["predict_samples"] = min(max_predict_samples, len(vectorized_datasets["predict"]))
# make sure that the `predict` metrics end up in the log history for the model card
trainer.log(OrderedDict(sorted(metrics.items())))
trainer.log_metrics("predict", metrics)
trainer.save_metrics("predict", metrics)
# Write model card and (optionally) push to hub
kwargs = {
"finetuned_from": model_args.model_name_or_path,
"tasks": task_name,
"tags": [task_name, data_args.dataset_name],
"dataset_args": (
f"Config: {config_name}, Training split: {data_args.train_split_name}, Eval split:"
f" {data_args.eval_split_name}, Predict split: {data_args.predict_split_name}"
),
"dataset": f"{data_args.dataset_name.upper()} - {config_name.upper()}",
"language": data_args.language,
}
if training_args.push_to_hub:
trainer.push_to_hub(**kwargs)
else:
trainer.create_model_card(**kwargs)
return results
if __name__ == "__main__":
main()
|
transformers/examples/research_projects/xtreme-s/run_xtreme_s.py/0
|
{
"file_path": "transformers/examples/research_projects/xtreme-s/run_xtreme_s.py",
"repo_id": "transformers",
"token_count": 16446
}
| 329
|
# Training a masked language model end-to-end from scratch on TPUs
In this example, we're going to demonstrate how to train a TensorFlow model from 🤗 Transformers from scratch. If you're interested in some background theory on training Hugging Face models with TensorFlow on TPU, please check out our
[tutorial doc](https://huggingface.co/docs/transformers/main/perf_train_tpu_tf) on this topic!
If you're interested in smaller-scale TPU training from a pre-trained checkpoint, you can also check out the [TPU fine-tuning example](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/tpu_training-tf.ipynb).
This example will demonstrate pre-training language models at the 100M-1B parameter scale, similar to BERT or GPT-2. More concretely, we will show how to train a [RoBERTa](https://huggingface.co/docs/transformers/model_doc/roberta) (base model) from scratch on the [WikiText dataset (v1)](https://huggingface.co/datasets/wikitext).
We've tried to ensure that all the practices we show you here are scalable, though - with relatively few changes, the code could be scaled up to much larger models.
Google's gargantuan [PaLM model](https://arxiv.org/abs/2204.02311), with
over 500B parameters, is a good example of how far you can go with pure TPU training, though gathering the dataset and the budget to train at that scale is not an easy task!
### Table of contents
- [Setting up a TPU-VM](#setting-up-a-tpu-vm)
- [Training a tokenizer](#training-a-tokenizer)
- [Preparing the dataset](#preparing-the-dataset)
- [Training the model](#training-the-model)
- [Inference](#inference)
## Setting up a TPU-VM
Since this example focuses on using TPUs, the first step is to set up access to TPU hardware. For this example, we chose to use a TPU v3-8 VM. Follow [this guide](https://cloud.google.com/tpu/docs/run-calculation-tensorflow) to quickly create a TPU VM with TensorFlow pre-installed.
> 💡 **Note**: You don't need a TPU-enabled hardware for tokenizer training and TFRecord shard preparation.
## Training a tokenizer
To train a language model from scratch, the first step is to tokenize text. In most Hugging Face examples, we begin from a pre-trained model and use its tokenizer. However, in this example, we're going to train a tokenizer from scratch as well. The script for this is `train_unigram.py`. An example command is:
```bash
python train_unigram.py --batch_size 1000 --vocab_size 25000 --export_to_hub
```
The script will automatically load the `train` split of the WikiText dataset and train a [Unigram tokenizer](https://huggingface.co/course/chapter6/7?fw=pt) on it.
> 💡 **Note**: In order for `export_to_hub` to work, you must authenticate yourself with the `huggingface-cli`. Run `huggingface-cli login` and follow the on-screen instructions.
## Preparing the dataset
The next step is to prepare the dataset. This consists of loading a text dataset from the Hugging Face Hub, tokenizing it and grouping it into chunks of a fixed length ready for training. The script for this is `prepare_tfrecord_shards.py`.
The reason we create TFRecord output files from this step is that these files work well with [`tf.data` pipelines](https://www.tensorflow.org/guide/data_performance). This makes them very suitable for scalable TPU training - the dataset can easily be sharded and read in parallel just by tweaking a few parameters in the pipeline. An example command is:
```bash
python prepare_tfrecord_shards.py \
--tokenizer_name_or_path tf-tpu/unigram-tokenizer-wikitext \
--shard_size 5000 \
--split test
--max_length 128 \
--output_dir gs://tf-tpu-training-resources
```
**Notes**:
* While running the above script, you need to specify the `split` accordingly. The example command above will only filter the `test` split of the dataset.
* If you append `gs://` in your `output_dir` the TFRecord shards will be directly serialized to a Google Cloud Storage (GCS) bucket. Ensure that you have already [created the GCS bucket](https://cloud.google.com/storage/docs).
* If you're using a TPU node, you must stream data from a GCS bucket. Otherwise, if you're using a TPU VM,you can store the data locally. You may need to [attach](https://cloud.google.com/tpu/docs/setup-persistent-disk) a persistent storage to the VM.
* Additional CLI arguments are also supported. We encourage you to run `python prepare_tfrecord_shards.py -h` to know more about them.
## Training the model
Once that's done, the model is ready for training. By default, training takes place on TPU, but you can use the `--no_tpu` flag to train on CPU for testing purposes. An example command is:
```bash
python3 run_mlm.py \
--train_dataset gs://tf-tpu-training-resources/train/ \
--eval_dataset gs://tf-tpu-training-resources/validation/ \
--tokenizer tf-tpu/unigram-tokenizer-wikitext \
--output_dir trained_model
```
If you had specified a `hub_model_id` while launching training, then your model will be pushed to a model repository on the Hugging Face Hub. You can find such an example repository here:
[tf-tpu/roberta-base-epochs-500-no-wd](https://huggingface.co/tf-tpu/roberta-base-epochs-500-no-wd).
## Inference
Once the model is trained, you can use 🤗 Pipelines to perform inference:
```python
from transformers import pipeline
model_id = "tf-tpu/roberta-base-epochs-500-no-wd"
unmasker = pipeline("fill-mask", model=model_id, framework="tf")
unmasker("Goal of my life is to [MASK].")
[{'score': 0.1003185287117958,
'token': 52,
'token_str': 'be',
'sequence': 'Goal of my life is to be.'},
{'score': 0.032648514956235886,
'token': 5,
'token_str': '',
'sequence': 'Goal of my life is to .'},
{'score': 0.02152673341333866,
'token': 138,
'token_str': 'work',
'sequence': 'Goal of my life is to work.'},
{'score': 0.019547373056411743,
'token': 984,
'token_str': 'act',
'sequence': 'Goal of my life is to act.'},
{'score': 0.01939118467271328,
'token': 73,
'token_str': 'have',
'sequence': 'Goal of my life is to have.'}]
```
You can also try out inference using the [Inference Widget](https://huggingface.co/tf-tpu/roberta-base-epochs-500-no-wd?text=Goal+of+my+life+is+to+%5BMASK%5D.) from the model page.
|
transformers/examples/tensorflow/language-modeling-tpu/README.md/0
|
{
"file_path": "transformers/examples/tensorflow/language-modeling-tpu/README.md",
"repo_id": "transformers",
"token_count": 1947
}
| 330
|
<!---
Copyright 2021 The HuggingFace Team. 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.
-->
# Summarization example
This script shows an example of training a *summarization* model with the 🤗 Transformers library.
For straightforward use-cases you may be able to use these scripts without modification, although we have also
included comments in the code to indicate areas that you may need to adapt to your own projects.
### Multi-GPU and TPU usage
By default, these scripts use a `MirroredStrategy` and will use multiple GPUs effectively if they are available. TPUs
can also be used by passing the name of the TPU resource with the `--tpu` argument.
### Example command
```
python run_summarization.py \
--model_name_or_path facebook/bart-base \
--dataset_name cnn_dailymail \
--dataset_config "3.0.0" \
--output_dir /tmp/tst-summarization \
--per_device_train_batch_size 8 \
--per_device_eval_batch_size 16 \
--num_train_epochs 3 \
--do_train \
--do_eval
```
|
transformers/examples/tensorflow/summarization/README.md/0
|
{
"file_path": "transformers/examples/tensorflow/summarization/README.md",
"repo_id": "transformers",
"token_count": 415
}
| 331
|
from collections import Counter
import datasets
import transformers
from transformers.convert_slow_tokenizer import SLOW_TO_FAST_CONVERTERS
from transformers.utils import logging
logging.set_verbosity_info()
TOKENIZER_CLASSES = {
name: (getattr(transformers, name), getattr(transformers, name + "Fast")) for name in SLOW_TO_FAST_CONVERTERS
}
dataset = datasets.load_dataset("facebook/xnli", split="test+validation") # no-script
total = 0
perfect = 0
imperfect = 0
wrong = 0
def check_diff(spm_diff, tok_diff, slow, fast):
if spm_diff == list(reversed(tok_diff)):
# AAA -> AA+A vs A+AA case.
return True
elif len(spm_diff) == len(tok_diff) and fast.decode(spm_diff) == fast.decode(tok_diff):
# Second order OK
# Barrich -> Barr + ich vs Bar + rich
return True
spm_reencoded = slow.encode(slow.decode(spm_diff))
tok_reencoded = fast.encode(fast.decode(spm_diff))
if spm_reencoded != spm_diff and spm_reencoded == tok_reencoded:
# Type 3 error.
# Snehagatha ->
# Sne, h, aga, th, a
# Sne, ha, gat, ha
# Encoding the wrong with sp does not even recover what spm gave us
# It fits tokenizer however...
return True
return False
def check_LTR_mark(line, idx, fast):
enc = fast.encode_plus(line)[0]
offsets = enc.offsets
curr, prev = offsets[idx], offsets[idx - 1]
if curr is not None and line[curr[0] : curr[1]] == "\u200f":
return True
if prev is not None and line[prev[0] : prev[1]] == "\u200f":
return True
def check_details(line, spm_ids, tok_ids, slow, fast):
# Encoding can be the same with same result AAA -> A + AA vs AA + A
# We can check that we use at least exactly the same number of tokens.
for i, (spm_id, tok_id) in enumerate(zip(spm_ids, tok_ids)):
if spm_id != tok_id:
break
first = i
for i, (spm_id, tok_id) in enumerate(zip(reversed(spm_ids), reversed(tok_ids))):
if spm_id != tok_id:
break
last = len(spm_ids) - i
spm_diff = spm_ids[first:last]
tok_diff = tok_ids[first:last]
if check_diff(spm_diff, tok_diff, slow, fast):
return True
if check_LTR_mark(line, first, fast):
return True
if last - first > 5:
# We might have twice a single problem, attempt to subdivide the disjointed tokens into smaller problems
spms = Counter(spm_ids[first:last])
toks = Counter(tok_ids[first:last])
removable_tokens = {spm_ for (spm_, si) in spms.items() if toks.get(spm_, 0) == si}
min_width = 3
for i in range(last - first - min_width):
if all(spm_ids[first + i + j] in removable_tokens for j in range(min_width)):
possible_matches = [
k
for k in range(last - first - min_width)
if tok_ids[first + k : first + k + min_width] == spm_ids[first + i : first + i + min_width]
]
for j in possible_matches:
if check_diff(spm_ids[first : first + i], tok_ids[first : first + j], sp, tok) and check_details(
line,
spm_ids[first + i : last],
tok_ids[first + j : last],
slow,
fast,
):
return True
print(f"Spm: {[fast.decode([spm_ids[i]]) for i in range(first, last)]}")
try:
print(f"Tok: {[fast.decode([tok_ids[i]]) for i in range(first, last)]}")
except Exception:
pass
fast.decode(spm_ids[:first])
fast.decode(spm_ids[last:])
wrong = fast.decode(spm_ids[first:last])
print()
print(wrong)
return False
def test_string(slow, fast, text):
global perfect
global imperfect
global wrong
global total
slow_ids = slow.encode(text)
fast_ids = fast.encode(text)
skip_assert = False
total += 1
if slow_ids != fast_ids:
if check_details(text, slow_ids, fast_ids, slow, fast):
skip_assert = True
imperfect += 1
else:
wrong += 1
else:
perfect += 1
if total % 10000 == 0:
print(f"({perfect} / {imperfect} / {wrong} ----- {perfect + imperfect + wrong})")
if skip_assert:
return
assert (
slow_ids == fast_ids
), f"line {text} : \n\n{slow_ids}\n{fast_ids}\n\n{slow.tokenize(text)}\n{fast.tokenize(text)}"
def test_tokenizer(slow, fast):
global batch_total
for i in range(len(dataset)):
# premise, all languages
for text in dataset[i]["premise"].values():
test_string(slow, fast, text)
# hypothesis, all languages
for text in dataset[i]["hypothesis"]["translation"]:
test_string(slow, fast, text)
if __name__ == "__main__":
for name, (slow_class, fast_class) in TOKENIZER_CLASSES.items():
checkpoint_names = list(slow_class.max_model_input_sizes.keys())
for checkpoint in checkpoint_names:
imperfect = 0
perfect = 0
wrong = 0
total = 0
print(f"========================== Checking {name}: {checkpoint} ==========================")
slow = slow_class.from_pretrained(checkpoint, force_download=True)
fast = fast_class.from_pretrained(checkpoint, force_download=True)
test_tokenizer(slow, fast)
print(f"Accuracy {perfect * 100 / total:.2f}")
|
transformers/scripts/check_tokenizers.py/0
|
{
"file_path": "transformers/scripts/check_tokenizers.py",
"repo_id": "transformers",
"token_count": 2570
}
| 332
|
# Copyright 2021 The HuggingFace Team, the AllenNLP library authors. 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.
"""
Script to close stale issue. Taken in part from the AllenNLP repository.
https://github.com/allenai/allennlp.
"""
import os
from datetime import datetime as dt
import github.GithubException
from github import Github
LABELS_TO_EXEMPT = [
"good first issue",
"good second issue",
"good difficult issue",
"feature request",
"new model",
"wip",
]
def main():
g = Github(os.environ["GITHUB_TOKEN"])
repo = g.get_repo("huggingface/transformers")
open_issues = repo.get_issues(state="open")
for i, issue in enumerate(open_issues):
print(i, issue)
comments = sorted(list(issue.get_comments()), key=lambda i: i.created_at, reverse=True)
last_comment = comments[0] if len(comments) > 0 else None
if (
last_comment is not None and last_comment.user.login == "github-actions[bot]"
and (dt.utcnow() - issue.updated_at.replace(tzinfo=None)).days > 7
and (dt.utcnow() - issue.created_at.replace(tzinfo=None)).days >= 30
and not any(label.name.lower() in LABELS_TO_EXEMPT for label in issue.get_labels())
):
# print(f"Would close issue {issue.number} since it has been 7 days of inactivity since bot mention.")
try:
issue.edit(state="closed")
except github.GithubException as e:
print("Couldn't close the issue:", repr(e))
elif (
(dt.utcnow() - issue.updated_at.replace(tzinfo=None)).days > 23
and (dt.utcnow() - issue.created_at.replace(tzinfo=None)).days >= 30
and not any(label.name.lower() in LABELS_TO_EXEMPT for label in issue.get_labels())
):
# print(f"Would add stale comment to {issue.number}")
try:
issue.create_comment(
"This issue has been automatically marked as stale because it has not had "
"recent activity. If you think this still needs to be addressed "
"please comment on this thread.\n\nPlease note that issues that do not follow the "
"[contributing guidelines](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md) "
"are likely to be ignored."
)
except github.GithubException as e:
print("Couldn't create comment:", repr(e))
if __name__ == "__main__":
main()
|
transformers/scripts/stale.py/0
|
{
"file_path": "transformers/scripts/stale.py",
"repo_id": "transformers",
"token_count": 1217
}
| 333
|
#!/usr/bin/env python
# coding=utf-8
# Copyright 2024 The HuggingFace Inc. team. 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.
import re
from ..utils import cached_file
# docstyle-ignore
CHAT_MESSAGE_PROMPT = """
Human: <<task>>
Assistant: """
DEFAULT_PROMPTS_REPO = "huggingface-tools/default-prompts"
PROMPT_FILES = {"chat": "chat_prompt_template.txt", "run": "run_prompt_template.txt"}
def download_prompt(prompt_or_repo_id, agent_name, mode="run"):
"""
Downloads and caches the prompt from a repo and returns it contents (if necessary).
"""
if prompt_or_repo_id is None:
prompt_or_repo_id = DEFAULT_PROMPTS_REPO
# prompt is considered a repo ID when it does not contain any kind of space
if re.search("\\s", prompt_or_repo_id) is not None:
return prompt_or_repo_id
prompt_file = cached_file(
prompt_or_repo_id, PROMPT_FILES[mode], repo_type="dataset", user_agent={"agent": agent_name}
)
with open(prompt_file, "r", encoding="utf-8") as f:
return f.read()
DEFAULT_CODE_SYSTEM_PROMPT = """You will be given a task to solve, your job is to come up with a series of simple commands in Python that will perform the task.
To help you, I will give you access to a set of tools that you can use. Each tool is a Python function and has a description explaining the task it performs, the inputs it expects and the outputs it returns.
You should first explain which tool you will use to perform the task and for what reason, then write the code in Python.
Each instruction in Python should be a simple assignment. You can print intermediate results if it makes sense to do so.
In the end, use tool 'final_answer' to return your answer, its argument will be what gets returned.
You can use imports in your code, but only from the following list of modules: <<authorized_imports>>
Be sure to provide a 'Code:' token, else the run will fail.
Tools:
<<tool_descriptions>>
Examples:
---
Task: "Answer the question in the variable `question` about the image stored in the variable `image`. The question is in French."
Thought: I will use the following tools: `translator` to translate the question into English and then `image_qa` to answer the question on the input image.
Code:
```py
translated_question = translator(question=question, src_lang="French", tgt_lang="English")
print(f"The translated question is {translated_question}.")
answer = image_qa(image=image, question=translated_question)
final_answer(f"The answer is {answer}")
```<end_action>
---
Task: "Identify the oldest person in the `document` and create an image showcasing the result."
Thought: I will use the following tools: `document_qa` to find the oldest person in the document, then `image_generator` to generate an image according to the answer.
Code:
```py
answer = document_qa(document, question="What is the oldest person?")
print(f"The answer is {answer}.")
image = image_generator(answer)
final_answer(image)
```<end_action>
---
Task: "Generate an image using the text given in the variable `caption`."
Thought: I will use the following tool: `image_generator` to generate an image.
Code:
```py
image = image_generator(prompt=caption)
final_answer(image)
```<end_action>
---
Task: "Summarize the text given in the variable `text` and read it out loud."
Thought: I will use the following tools: `summarizer` to create a summary of the input text, then `text_reader` to read it out loud.
Code:
```py
summarized_text = summarizer(text)
print(f"Summary: {summarized_text}")
audio_summary = text_reader(summarized_text)
final_answer(audio_summary)
```<end_action>
---
Task: "Answer the question in the variable `question` about the text in the variable `text`. Use the answer to generate an image."
Thought: I will use the following tools: `text_qa` to create the answer, then `image_generator` to generate an image according to the answer.
Code:
```py
answer = text_qa(text=text, question=question)
print(f"The answer is {answer}.")
image = image_generator(answer)
final_answer(image)
```<end_action>
---
Task: "Caption the following `image`."
Thought: I will use the following tool: `image_captioner` to generate a caption for the image.
Code:
```py
caption = image_captioner(image)
final_answer(caption)
```<end_action>
---
Above example were using tools that might not exist for you. You only have acces to those Tools:
<<tool_names>>
Remember to make sure that variables you use are all defined.
Be sure to provide a 'Code:\n```' sequence before the code and '```<end_action>' after, else you will get an error.
DO NOT pass the arguments as a dict as in 'answer = ask_search_agent({'query': "What is the place where James Bond lives?"})', but use the arguments directly as in 'answer = ask_search_agent(query="What is the place where James Bond lives?")'.
Now Begin! If you solve the task correctly, you will receive a reward of $1,000,000.
"""
DEFAULT_REACT_JSON_SYSTEM_PROMPT = """You are an expert assistant who can solve any task using JSON tool calls. You will be given a task to solve as best you can.
To do so, you have been given access to the following tools: <<tool_names>>
The way you use the tools is by specifying a json blob, ending with '<end_action>'.
Specifically, this json should have an `action` key (name of the tool to use) and an `action_input` key (input to the tool).
The $ACTION_JSON_BLOB should only contain a SINGLE action, do NOT return a list of multiple actions. It should be formatted in json. Do not try to escape special characters. Here is the template of a valid $ACTION_JSON_BLOB:
{
"action": $TOOL_NAME,
"action_input": $INPUT
}<end_action>
Make sure to have the $INPUT as a dictionary in the right format for the tool you are using, and do not put variable names as input if you can find the right values.
You should ALWAYS use the following format:
Thought: you should always think about one action to take. Then use the action as follows:
Action:
$ACTION_JSON_BLOB
Observation: the result of the action
... (this Thought/Action/Observation can repeat N times, you should take several steps when needed. The $ACTION_JSON_BLOB must only use a SINGLE action at a time.)
You can use the result of the previous action as input for the next action.
The observation will always be a string: it can represent a file, like "image_1.jpg".
Then you can use it as input for the next action. You can do it for instance as follows:
Observation: "image_1.jpg"
Thought: I need to transform the image that I received in the previous observation to make it green.
Action:
{
"action": "image_transformer",
"action_input": {"image": "image_1.jpg"}
}<end_action>
To provide the final answer to the task, use an action blob with "action": "final_answer" tool. It is the only way to complete the task, else you will be stuck on a loop. So your final output should look like this:
Action:
{
"action": "final_answer",
"action_input": {"answer": "insert your final answer here"}
}<end_action>
Here are a few examples using notional tools:
---
Task: "Generate an image of the oldest person in this document."
Thought: I will proceed step by step and use the following tools: `document_qa` to find the oldest person in the document, then `image_generator` to generate an image according to the answer.
Action:
{
"action": "document_qa",
"action_input": {"document": "document.pdf", "question": "Who is the oldest person mentioned?"}
}<end_action>
Observation: "The oldest person in the document is John Doe, a 55 year old lumberjack living in Newfoundland."
Thought: I will now generate an image showcasing the oldest person.
Action:
{
"action": "image_generator",
"action_input": {"text": ""A portrait of John Doe, a 55-year-old man living in Canada.""}
}<end_action>
Observation: "image.png"
Thought: I will now return the generated image.
Action:
{
"action": "final_answer",
"action_input": "image.png"
}<end_action>
---
Task: "What is the result of the following operation: 5 + 3 + 1294.678?"
Thought: I will use python code evaluator to compute the result of the operation and then return the final answer using the `final_answer` tool
Action:
{
"action": "python_interpreter",
"action_input": {"code": "5 + 3 + 1294.678"}
}<end_action>
Observation: 1302.678
Thought: Now that I know the result, I will now return it.
Action:
{
"action": "final_answer",
"action_input": "1302.678"
}<end_action>
---
Task: "Which city has the highest population , Guangzhou or Shanghai?"
Thought: I need to get the populations for both cities and compare them: I will use the tool `search` to get the population of both cities.
Action:
{
"action": "search",
"action_input": "Population Guangzhou"
}<end_action>
Observation: ['Guangzhou has a population of 15 million inhabitants as of 2021.']
Thought: Now let's get the population of Shanghai using the tool 'search'.
Action:
{
"action": "search",
"action_input": "Population Shanghai"
}
Observation: '26 million (2019)'
Thought: Now I know that Shanghai has a larger population. Let's return the result.
Action:
{
"action": "final_answer",
"action_input": "Shanghai"
}<end_action>
Above example were using notional tools that might not exist for you. You only have acces to those tools:
<<tool_descriptions>>
Here are the rules you should always follow to solve your task:
1. ALWAYS provide a 'Thought:' sequence, and an 'Action:' sequence that ends with <end_action>, else you will fail.
2. Always use the right arguments for the tools. Never use variable names in the 'action_input' field, use the value instead.
3. Call a tool only when needed: do not call the search agent if you do not need information, try to solve the task yourself.
4. Never re-do a tool call that you previously did with the exact same parameters.
Now Begin! If you solve the task correctly, you will receive a reward of $1,000,000.
"""
DEFAULT_REACT_CODE_SYSTEM_PROMPT = """You are an expert assistant who can solve any task using code blobs. You will be given a task to solve as best you can.
To do so, you have been given access to a list of tools: these tools are basically Python functions which you can call with code.
To solve the task, you must plan forward to proceed in a series of steps, in a cycle of 'Thought:', 'Code:', and 'Observation:' sequences.
At each step, in the 'Thought:' sequence, you should first explain your reasoning towards solving the task and the tools that you want to use.
Then in the 'Code:' sequence, you should write the code in simple Python. The code sequence must end with '<end_action>' sequence.
During each intermediate step, you can use 'print()' to save whatever important information you will then need.
These print outputs will then appear in the 'Observation:' field, which will be available as input for the next step.
In the end you have to return a final answer using the `final_answer` tool.
Here are a few examples using notional tools:
---
Task: "Generate an image of the oldest person in this document."
Thought: I will proceed step by step and use the following tools: `document_qa` to find the oldest person in the document, then `image_generator` to generate an image according to the answer.
Code:
```py
answer = document_qa(document=document, question="Who is the oldest person mentioned?")
print(answer)
```<end_action>
Observation: "The oldest person in the document is John Doe, a 55 year old lumberjack living in Newfoundland."
Thought: I will now generate an image showcasing the oldest person.
Code:
```py
image = image_generator("A portrait of John Doe, a 55-year-old man living in Canada.")
final_answer(image)
```<end_action>
---
Task: "What is the result of the following operation: 5 + 3 + 1294.678?"
Thought: I will use python code to compute the result of the operation and then return the final answer using the `final_answer` tool
Code:
```py
result = 5 + 3 + 1294.678
final_answer(result)
```<end_action>
---
Task: "Which city has the highest population: Guangzhou or Shanghai?"
Thought: I need to get the populations for both cities and compare them: I will use the tool `search` to get the population of both cities.
Code:
```py
population_guangzhou = search("Guangzhou population")
print("Population Guangzhou:", population_guangzhou)
population_shanghai = search("Shanghai population")
print("Population Shanghai:", population_shanghai)
```<end_action>
Observation:
Population Guangzhou: ['Guangzhou has a population of 15 million inhabitants as of 2021.']
Population Shanghai: '26 million (2019)'
Thought: Now I know that Shanghai has the highest population.
Code:
```py
final_answer("Shanghai")
```<end_action>
---
Task: "What is the current age of the pope, raised to the power 0.36?"
Thought: I will use the tool `search` to get the age of the pope, then raise it to the power 0.36.
Code:
```py
pope_age = search(query="current pope age")
print("Pope age:", pope_age)
```<end_action>
Observation:
Pope age: "The pope Francis is currently 85 years old."
Thought: I know that the pope is 85 years old. Let's compute the result using python code.
Code:
```py
pope_current_age = 85 ** 0.36
final_answer(pope_current_age)
```<end_action>
Above example were using notional tools that might not exist for you. You only have acces to those tools:
<<tool_descriptions>>
You also can perform computations in the Python code that you generate.
Here are the rules you should always follow to solve your task:
1. Always provide a 'Thought:' sequence, and a 'Code:\n```py' sequence ending with '```<end_action>' sequence, else you will fail.
2. Use only variables that you have defined!
3. Always use the right arguments for the tools. DO NOT pass the arguments as a dict as in 'answer = ask_search_agent({'query': "What is the place where James Bond lives?"})', but use the arguments directly as in 'answer = ask_search_agent(query="What is the place where James Bond lives?")'.
4. Take care to not chain too many sequential tool calls in the same code block, especially when the output format is unpredictable. For instance, a call to search has an unpredictable return format, so do not have another tool call that depends on its output in the same block: rather output results with print() to use them in the next block.
5. Call a tool only when needed, and never re-do a tool call that you previously did with the exact same parameters.
6. Don't name any new variable with the same name as a tool: for instance don't name a variable 'final_answer'.
7. Never create any notional variables in our code, as having these in your logs might derail you from the true variables.
8. You can use imports in your code, but only from the following list of modules: <<authorized_imports>>
9. The state persists between code executions: so if in one step you've created variables or imported modules, these will all persist.
10. Don't give up! You're in charge of solving the task, not providing directions to solve it.
Now Begin! If you solve the task correctly, you will receive a reward of $1,000,000.
"""
SYSTEM_PROMPT_FACTS = """Below I will present you a task.
You will now build a comprehensive preparatory survey of which facts we have at our disposal and which ones we still need.
To do so, you will have to read the task and identify things that must be discovered in order to successfully complete it.
Don't make any assumptions. For each item, provide a thorough reasoning. Here is how you will structure this survey:
---
### 1. Facts given in the task
List here the specific facts given in the task that could help you (there might be nothing here).
### 2. Facts to look up
List here any facts that we may need to look up.
Also list where to find each of these, for instance a website, a file... - maybe the task contains some sources that you should re-use here.
### 3. Facts to derive
List here anything that we want to derive from the above by logical reasoning, for instance computation or simulation.
Keep in mind that "facts" will typically be specific names, dates, values, etc. Your answer should use the below headings:
### 1. Facts given in the task
### 2. Facts to look up
### 3. Facts to derive
Do not add anything else."""
SYSTEM_PROMPT_PLAN = """You are a world expert at making efficient plans to solve any task using a set of carefully crafted tools.
Now for the given task, develop a step-by-step high-level plan taking into account the above inputs and list of facts.
This plan should involve individual tasks based on the avilable tools, that if executed correctly will yield the correct answer.
Do not skip steps, do not add any superfluous steps. Only write the high-level plan, DO NOT DETAIL INDIVIDUAL TOOL CALLS.
After writing the final step of the plan, write the '\n<end_plan>' tag and stop there."""
USER_PROMPT_PLAN = """
Here is your task:
Task:
```
{task}
```
Your plan can leverage any of these tools:
{tool_descriptions}
List of facts that you know:
```
{answer_facts}
```
Now begin! Write your plan below."""
SYSTEM_PROMPT_FACTS_UPDATE = """
You are a world expert at gathering known and unknown facts based on a conversation.
Below you will find a task, and ahistory of attempts made to solve the task. You will have to produce a list of these:
### 1. Facts given in the task
### 2. Facts that we have learned
### 3. Facts still to look up
### 4. Facts still to derive
Find the task and history below."""
USER_PROMPT_FACTS_UPDATE = """Earlier we've built a list of facts.
But since in your previous steps you may have learned useful new facts or invalidated some false ones.
Please update your list of facts based on the previous history, and provide these headings:
### 1. Facts given in the task
### 2. Facts that we have learned
### 3. Facts still to look up
### 4. Facts still to derive
Now write your new list of facts below."""
SYSTEM_PROMPT_PLAN_UPDATE = """You are a world expert at making efficient plans to solve any task using a set of carefully crafted tools.
You have been given a task:
```
{task}
```
Find below the record of what has been tried so far to solve it. Then you will be asked to make an updated plan to solve the task.
If the previous tries so far have met some success, you can make an updated plan based on these actions.
If you are stalled, you can make a completely new plan starting from scratch.
"""
USER_PROMPT_PLAN_UPDATE = """You're still working towards solving this task:
```
{task}
```
You have access to these tools:
{tool_descriptions}
Here is the up to date list of facts that you know:
```
{facts_update}
```
Now for the given task, develop a step-by-step high-level plan taking into account the above inputs and list of facts.
This plan should involve individual tasks based on the avilable tools, that if executed correctly will yield the correct answer.
Beware that you have {remaining_steps} steps remaining.
Do not skip steps, do not add any superfluous steps. Only write the high-level plan, DO NOT DETAIL INDIVIDUAL TOOL CALLS.
After writing the final step of the plan, write the '\n<end_plan>' tag and stop there.
Now write your new plan below."""
SYSTEM_PROMPT_PLAN_STRUCTURED = """Output a step-by-step plan to solve the task using the given tools.
This plan should involve individual tasks based on the avilable tools, that if executed correctly will yield the correct answer. Each step should be structured as follows:
Step #n: {
"description": <description of what the step does and its output>
"tool": <tool to use>,
"params": {
<parameters to pass to the tool as a valid dict>
}
"output_var": <output variable name>
}
Each step must be necessary to reach the final answer. Steps should reuse outputs produced by earlier steps. The last step must be the final answer.
Below are some examples:
Example 1:
------
Inputs:
---
Task:
How many encoder blocks were in the first attention-only ML architecture published?
[FACTS LIST]:
### 1. Facts given in the task
- The paper first introduced an attention-only ML architecture.
- The specific information required is the page number where the number of encoder blocks is stated.
- No local files are provided for access.
### 2. Facts to look up
- The title and authors of the paper that first introduced an attention-only ML architecture.
- Source: Online search (e.g., Google Scholar, arXiv, or other academic databases)
- The full text of the identified paper.
- Source: Online academic repositories (e.g., arXiv, journal websites)
- The specific page number in the paper where the number of encoder blocks is mentioned.
- Source: The content of the identified paper
### 3. Facts to derive
- By identifying the correct paper and locating the specific page, we will derive the page number where the number of encoder blocks is stated.
- Logical steps: Identify the correct paper, access its content, search for the term "encoder blocks," and note the page number where this information is found.
```
[STEP 1 TOOL CALL]: {'tool_name': 'code interpreter', 'tool_arguments': '# Step 1: Identify the title and authors of the paper that first introduced an attention-only ML architecture.\nanswer = ask_search_agent(query="Can you find the title and authors of the paper that first introduced an attention-only machine learning architecture? Please provide the full citation.")\nprint(answer)'}
[OUTPUT OF STEP 1] Observation: **Title**: Attention Is All You Need
**Authors**: Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, Illia Polosukhin
[STEP 2 TOOL CALL]: {'tool_name': 'code interpreter', 'tool_arguments': '# Step 1: Find the full text of the identified paper on arXiv\\npaper_url = "https://arxiv.org/pdf/1706.03762.pdf"\\nprint(paper_url)'}
[OUTPUT OF STEP 2] Observation: https://arxiv.org/pdf/1706.03762.pdf
---
Output plan:
---
Step #1: {
"description": "Open the PDF of the paper from the provided URL and search within the text of the paper for the mention of "encoder blocks"",
"tool": "inspect_file_as_text",
"params": {
"file_path": "https://arxiv.org/pdf/1706.03762.pdf",
"question": "On which page is the number of encoder blocks mentioned?"
},
"output_var": "page_number"
}
Step #2: {
"description": "Provide the final answer",
"tool": "final_answer",
"params": {
"answer": "{page_number}"
},
"output_var": ""
}
------
Example 2:
------
Inputs:
---
Task:
How many golf balls fits into a Boeing-747?
[FACTS LIST]:
### 1. Facts given in the task
- The task requires calculating the number of golf balls that fir into a Boeing-747
### 2. Facts to look up
- The volume of a golf ball
- The volume of a Boeing-747
### 3. Facts to derive
- Once the volumes are known the final answer can be calculated
---
Output plan:
---
Step #1: {
"description": "Find the volume of a Boeing-747",
"tool": "web_search",
"params": {
"query": "What is the internal volume of a Boeing-747 in cubic meters?"
},
"output_var": "boeing_volume"
}
Step #2: {
"description": "Find the volume of a standard golf ball",
"tool": "ask_search_agent",
"params": {
"query": "What is the volume of a standard golf ball in cubic centimeters?"
},
"output_var": "golf_ball_volume"
}
Step #3: {
"description": "Convert the volume of a golf ball from cubic centimeters to cubic meters. Calculate the number of golf balls that fit into the Boeing-747 by dividing the internal volume of the Boeing-747 by the volume of a golf ball.",
"tool": "python_code",
"params": {
"code": "golf_ball_volume_m3 = golf_ball_volume / 1e6\nnumber_of_golf_balls = boeing_volume / golf_ball_volume_m3"
},
"output_var": "number_of_golf_balls"
}
Step #4: {
"description": "Provide the final answer",
"tool": "final_answer",
"params": {
"answer": "{number_of_golf_balls}"
},
"output_var": ""
}
------
Above example were using tools that might not exist for you.
Your goal is to create a plan to solve the task."""
USER_PROMPT_PLAN_STRUCTURED = """
Here are your inputs:
Task:
```
{task}
```
Your plan can leverage any of these tools:
{tool_descriptions}
These tools are Python functions which you can call with code. You also have access to a Python interpreter so you can run Python code.
List of facts that you know:
```
{answer_facts}
```
Now for the given task, create a plan taking into account the list of facts.
After writing the final step of the plan, write the '\n<end_plan>' tag and stop there. Output the plan only and nothing else."""
SYSTEM_PROMPT_PLAN_UPDATE_STRUCTURED = """Output a step-by-step plan to solve the task using the given tools.
This plan should involve individual tasks based on the avilable tools, that if executed correctly will yield the correct answer. Each step should be structured as follows:
Step #n: {{
"description": <description of what the step does and its output>
"tool": <tool to use>,
"params": {{
<parameters to pass to the tool as a valid dict>
}}
"output_var": <output variable name>
}}
Each step must be necessary to reach the final answer. Steps should reuse outputs produced by earlier steps. The last step must be the final answer.
Below are some examples:
Example 1:
------
Inputs:
---
Task:
How many encoder blocks were in the first attention-only ML architecture published?
[FACTS LIST]:
### 1. Facts given in the task
- The paper first introduced an attention-only ML architecture.
- The specific information required is the page number where the number of encoder blocks is stated.
- No local files are provided for access.
### 2. Facts to look up
- The title and authors of the paper that first introduced an attention-only ML architecture.
- Source: Online search (e.g., Google Scholar, arXiv, or other academic databases)
- The full text of the identified paper.
- Source: Online academic repositories (e.g., arXiv, journal websites)
- The specific page number in the paper where the number of encoder blocks is mentioned.
- Source: The content of the identified paper
### 3. Facts to derive
- By identifying the correct paper and locating the specific page, we will derive the page number where the number of encoder blocks is stated.
- Logical steps: Identify the correct paper, access its content, search for the term "encoder blocks," and note the page number where this information is found.
```
[STEP 1 TOOL CALL]: {{'tool_name': 'code interpreter', 'tool_arguments': '# Step 1: Identify the title and authors of the paper that first introduced an attention-only ML architecture.\nanswer = ask_search_agent(query="Can you find the title and authors of the paper that first introduced an attention-only machine learning architecture? Please provide the full citation.")\nprint(answer)'}}
[OUTPUT OF STEP 1] Observation: **Title**: Attention Is All You Need
**Authors**: Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, Illia Polosukhin
[STEP 2 TOOL CALL]: {{'tool_name': 'code interpreter', 'tool_arguments': '# Step 1: Find the full text of the identified paper on arXiv\\npaper_url = "https://arxiv.org/pdf/1706.03762.pdf"\\nprint(paper_url)'}}
[OUTPUT OF STEP 2] Observation: https://arxiv.org/pdf/1706.03762.pdf
---
Output plan:
---
Step #1: {{
"description": "Open the PDF of the paper from the provided URL and search within the text of the paper for the mention of "encoder blocks"",
"tool": "inspect_file_as_text",
"params": {{
"file_path": "https://arxiv.org/pdf/1706.03762.pdf",
"question": "On which page is the number of encoder blocks mentioned?"
}},
"output_var": "page_number"
}}
Step #2: {{
"description": "Provide the final answer",
"tool": "final_answer",
"params": {{
"answer": "{{page_number}}"
}},
"output_var": ""
}}
------
Example 2:
------
Inputs:
---
Task:
How many golf balls fits into a Boeing-747?
[FACTS LIST]:
### 1. Facts given in the task
- The task requires calculating the number of golf balls that fir into a Boeing-747
### 2. Facts to look up
- The volume of a golf ball
- The volume of a Boeing-747
### 3. Facts to derive
- Once the volumes are known the final answer can be calculated
---
Output plan:
---
Step #1: {{
"description": "Find the volume of a Boeing-747",
"tool": "web_search",
"params": {{
"query": "What is the internal volume of a Boeing-747 in cubic meters?"
}},
"output_var": "boeing_volume"
}}
Step #2: {{
"description": "Find the volume of a standard golf ball",
"tool": "ask_search_agent",
"params": {{
"query": "What is the volume of a standard golf ball in cubic centimeters?"
}},
"output_var": "golf_ball_volume"
}}
Step #3: {{
"description": "Convert the volume of a golf ball from cubic centimeters to cubic meters. Calculate the number of golf balls that fit into the Boeing-747 by dividing the internal volume of the Boeing-747 by the volume of a golf ball.",
"tool": "python_code",
"params": {{
"code": "golf_ball_volume_m3 = golf_ball_volume / 1e6\nnumber_of_golf_balls = boeing_volume / golf_ball_volume_m3"
}},
"output_var": "number_of_golf_balls"
}}
Step #4: {{
"description": "Provide the final answer",
"tool": "final_answer",
"params": {{
"answer": "{{number_of_golf_balls}}"
}},
"output_var": ""
}}
------
Above example were using tools that might not exist for you.
Find below the record of what has been tried so far to solve it. Your goal is to create an updated plan to solve the task."""
USER_PROMPT_PLAN_UPDATE_STRUCTURED = """
Here are your inputs:
Task:
```
{task}
```
Your plan can leverage any of these tools:
{tool_descriptions}
These tools are Python functions which you can call with code. You also have access to a Python interpreter so you can run Python code.
List of facts that you know:
```
{facts_update}
```
Now for the given task, create a plan taking into account the above inputs and list of facts.
Beware that you have {remaining_steps} steps remaining.
After writing the final step of the plan, write the '\n<end_plan>' tag and stop there. Output the plan only and nothing else."""
PLAN_UPDATE_FINAL_PLAN_REDACTION = """I still need to solve the task I was given:
```
{task}
```
Here is my new/updated plan of action to solve the task:
```
{plan_update}
```"""
SUPPORTED_PLAN_TYPES = ["default", "structured"]
PROMPTS_FOR_INITIAL_PLAN = {
"default": {"system": SYSTEM_PROMPT_PLAN, "user": USER_PROMPT_PLAN},
"structured": {"system": SYSTEM_PROMPT_PLAN_STRUCTURED, "user": USER_PROMPT_PLAN_STRUCTURED},
}
PROMPTS_FOR_PLAN_UPDATE = {
"default": {"system": SYSTEM_PROMPT_PLAN_UPDATE, "user": USER_PROMPT_PLAN_UPDATE},
"structured": {"system": SYSTEM_PROMPT_PLAN_UPDATE_STRUCTURED, "user": USER_PROMPT_PLAN_UPDATE_STRUCTURED},
}
|
transformers/src/transformers/agents/prompts.py/0
|
{
"file_path": "transformers/src/transformers/agents/prompts.py",
"repo_id": "transformers",
"token_count": 9084
}
| 334
|
# Copyright 2021 The HuggingFace Team. 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.
import difflib
import json
import os
import re
from argparse import ArgumentParser, Namespace
from dataclasses import dataclass
from datetime import date
from itertools import chain
from pathlib import Path
from typing import Any, Callable, Dict, List, Optional, Pattern, Tuple, Union
import yaml
from ..models import auto as auto_module
from ..models.auto.configuration_auto import model_type_to_module_name
from ..utils import is_flax_available, is_tf_available, is_torch_available, logging
from . import BaseTransformersCLICommand
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
CURRENT_YEAR = date.today().year
TRANSFORMERS_PATH = Path(__file__).parent.parent
REPO_PATH = TRANSFORMERS_PATH.parent.parent
@dataclass
class ModelPatterns:
"""
Holds the basic information about a new model for the add-new-model-like command.
Args:
model_name (`str`): The model name.
checkpoint (`str`): The checkpoint to use for doc examples.
model_type (`str`, *optional*):
The model type, the identifier used internally in the library like `bert` or `xlm-roberta`. Will default to
`model_name` lowercased with spaces replaced with minuses (-).
model_lower_cased (`str`, *optional*):
The lowercased version of the model name, to use for the module name or function names. Will default to
`model_name` lowercased with spaces and minuses replaced with underscores.
model_camel_cased (`str`, *optional*):
The camel-cased version of the model name, to use for the class names. Will default to `model_name`
camel-cased (with spaces and minuses both considered as word separators.
model_upper_cased (`str`, *optional*):
The uppercased version of the model name, to use for the constant names. Will default to `model_name`
uppercased with spaces and minuses replaced with underscores.
config_class (`str`, *optional*):
The tokenizer class associated with this model. Will default to `"{model_camel_cased}Config"`.
tokenizer_class (`str`, *optional*):
The tokenizer class associated with this model (leave to `None` for models that don't use a tokenizer).
image_processor_class (`str`, *optional*):
The image processor class associated with this model (leave to `None` for models that don't use an image
processor).
feature_extractor_class (`str`, *optional*):
The feature extractor class associated with this model (leave to `None` for models that don't use a feature
extractor).
processor_class (`str`, *optional*):
The processor class associated with this model (leave to `None` for models that don't use a processor).
"""
model_name: str
checkpoint: str
model_type: Optional[str] = None
model_lower_cased: Optional[str] = None
model_camel_cased: Optional[str] = None
model_upper_cased: Optional[str] = None
config_class: Optional[str] = None
tokenizer_class: Optional[str] = None
image_processor_class: Optional[str] = None
feature_extractor_class: Optional[str] = None
processor_class: Optional[str] = None
def __post_init__(self):
if self.model_type is None:
self.model_type = self.model_name.lower().replace(" ", "-")
if self.model_lower_cased is None:
self.model_lower_cased = self.model_name.lower().replace(" ", "_").replace("-", "_")
if self.model_camel_cased is None:
# Split the model name on - and space
words = self.model_name.split(" ")
words = list(chain(*[w.split("-") for w in words]))
# Make sure each word is capitalized
words = [w[0].upper() + w[1:] for w in words]
self.model_camel_cased = "".join(words)
if self.model_upper_cased is None:
self.model_upper_cased = self.model_name.upper().replace(" ", "_").replace("-", "_")
if self.config_class is None:
self.config_class = f"{self.model_camel_cased}Config"
ATTRIBUTE_TO_PLACEHOLDER = {
"config_class": "[CONFIG_CLASS]",
"tokenizer_class": "[TOKENIZER_CLASS]",
"image_processor_class": "[IMAGE_PROCESSOR_CLASS]",
"feature_extractor_class": "[FEATURE_EXTRACTOR_CLASS]",
"processor_class": "[PROCESSOR_CLASS]",
"checkpoint": "[CHECKPOINT]",
"model_type": "[MODEL_TYPE]",
"model_upper_cased": "[MODEL_UPPER_CASED]",
"model_camel_cased": "[MODEL_CAMELCASED]",
"model_lower_cased": "[MODEL_LOWER_CASED]",
"model_name": "[MODEL_NAME]",
}
def is_empty_line(line: str) -> bool:
"""
Determines whether a line is empty or not.
"""
return len(line) == 0 or line.isspace()
def find_indent(line: str) -> int:
"""
Returns the number of spaces that start a line indent.
"""
search = re.search(r"^(\s*)(?:\S|$)", line)
if search is None:
return 0
return len(search.groups()[0])
def parse_module_content(content: str) -> List[str]:
"""
Parse the content of a module in the list of objects it defines.
Args:
content (`str`): The content to parse
Returns:
`List[str]`: The list of objects defined in the module.
"""
objects = []
current_object = []
lines = content.split("\n")
# Doc-styler takes everything between two triple quotes in docstrings, so we need a fake """ here to go with this.
end_markers = [")", "]", "}", '"""']
for line in lines:
# End of an object
is_valid_object = len(current_object) > 0
if is_valid_object and len(current_object) == 1:
is_valid_object = not current_object[0].startswith("# Copied from")
if not is_empty_line(line) and find_indent(line) == 0 and is_valid_object:
# Closing parts should be included in current object
if line in end_markers:
current_object.append(line)
objects.append("\n".join(current_object))
current_object = []
else:
objects.append("\n".join(current_object))
current_object = [line]
else:
current_object.append(line)
# Add last object
if len(current_object) > 0:
objects.append("\n".join(current_object))
return objects
def extract_block(content: str, indent_level: int = 0) -> str:
"""Return the first block in `content` with the indent level `indent_level`.
The first line in `content` should be indented at `indent_level` level, otherwise an error will be thrown.
This method will immediately stop the search when a (non-empty) line with indent level less than `indent_level` is
encountered.
Args:
content (`str`): The content to parse
indent_level (`int`, *optional*, default to 0): The indent level of the blocks to search for
Returns:
`str`: The first block in `content` with the indent level `indent_level`.
"""
current_object = []
lines = content.split("\n")
# Doc-styler takes everything between two triple quotes in docstrings, so we need a fake """ here to go with this.
end_markers = [")", "]", "}", '"""']
for idx, line in enumerate(lines):
if idx == 0 and indent_level > 0 and not is_empty_line(line) and find_indent(line) != indent_level:
raise ValueError(
f"When `indent_level > 0`, the first line in `content` should have indent level {indent_level}. Got "
f"{find_indent(line)} instead."
)
if find_indent(line) < indent_level and not is_empty_line(line):
break
# End of an object
is_valid_object = len(current_object) > 0
if (
not is_empty_line(line)
and not line.endswith(":")
and find_indent(line) == indent_level
and is_valid_object
):
# Closing parts should be included in current object
if line.lstrip() in end_markers:
current_object.append(line)
return "\n".join(current_object)
else:
current_object.append(line)
# Add last object
if len(current_object) > 0:
return "\n".join(current_object)
def add_content_to_text(
text: str,
content: str,
add_after: Optional[Union[str, Pattern]] = None,
add_before: Optional[Union[str, Pattern]] = None,
exact_match: bool = False,
) -> str:
"""
A utility to add some content inside a given text.
Args:
text (`str`): The text in which we want to insert some content.
content (`str`): The content to add.
add_after (`str` or `Pattern`):
The pattern to test on a line of `text`, the new content is added after the first instance matching it.
add_before (`str` or `Pattern`):
The pattern to test on a line of `text`, the new content is added before the first instance matching it.
exact_match (`bool`, *optional*, defaults to `False`):
A line is considered a match with `add_after` or `add_before` if it matches exactly when `exact_match=True`,
otherwise, if `add_after`/`add_before` is present in the line.
<Tip warning={true}>
The arguments `add_after` and `add_before` are mutually exclusive, and one exactly needs to be provided.
</Tip>
Returns:
`str`: The text with the new content added if a match was found.
"""
if add_after is None and add_before is None:
raise ValueError("You need to pass either `add_after` or `add_before`")
if add_after is not None and add_before is not None:
raise ValueError("You can't pass both `add_after` or `add_before`")
pattern = add_after if add_before is None else add_before
def this_is_the_line(line):
if isinstance(pattern, Pattern):
return pattern.search(line) is not None
elif exact_match:
return pattern == line
else:
return pattern in line
new_lines = []
for line in text.split("\n"):
if this_is_the_line(line):
if add_before is not None:
new_lines.append(content)
new_lines.append(line)
if add_after is not None:
new_lines.append(content)
else:
new_lines.append(line)
return "\n".join(new_lines)
def add_content_to_file(
file_name: Union[str, os.PathLike],
content: str,
add_after: Optional[Union[str, Pattern]] = None,
add_before: Optional[Union[str, Pattern]] = None,
exact_match: bool = False,
):
"""
A utility to add some content inside a given file.
Args:
file_name (`str` or `os.PathLike`): The name of the file in which we want to insert some content.
content (`str`): The content to add.
add_after (`str` or `Pattern`):
The pattern to test on a line of `text`, the new content is added after the first instance matching it.
add_before (`str` or `Pattern`):
The pattern to test on a line of `text`, the new content is added before the first instance matching it.
exact_match (`bool`, *optional*, defaults to `False`):
A line is considered a match with `add_after` or `add_before` if it matches exactly when `exact_match=True`,
otherwise, if `add_after`/`add_before` is present in the line.
<Tip warning={true}>
The arguments `add_after` and `add_before` are mutually exclusive, and one exactly needs to be provided.
</Tip>
"""
with open(file_name, "r", encoding="utf-8") as f:
old_content = f.read()
new_content = add_content_to_text(
old_content, content, add_after=add_after, add_before=add_before, exact_match=exact_match
)
with open(file_name, "w", encoding="utf-8") as f:
f.write(new_content)
def replace_model_patterns(
text: str, old_model_patterns: ModelPatterns, new_model_patterns: ModelPatterns
) -> Tuple[str, str]:
"""
Replace all patterns present in a given text.
Args:
text (`str`): The text to treat.
old_model_patterns (`ModelPatterns`): The patterns for the old model.
new_model_patterns (`ModelPatterns`): The patterns for the new model.
Returns:
`Tuple(str, str)`: A tuple of with the treated text and the replacement actually done in it.
"""
# The order is crucially important as we will check and replace in that order. For instance the config probably
# contains the camel-cased named, but will be treated before.
attributes_to_check = ["config_class"]
# Add relevant preprocessing classes
for attr in ["tokenizer_class", "image_processor_class", "feature_extractor_class", "processor_class"]:
if getattr(old_model_patterns, attr) is not None and getattr(new_model_patterns, attr) is not None:
attributes_to_check.append(attr)
# Special cases for checkpoint and model_type
if old_model_patterns.checkpoint not in [old_model_patterns.model_type, old_model_patterns.model_lower_cased]:
attributes_to_check.append("checkpoint")
if old_model_patterns.model_type != old_model_patterns.model_lower_cased:
attributes_to_check.append("model_type")
else:
text = re.sub(
rf'(\s*)model_type = "{old_model_patterns.model_type}"',
r'\1model_type = "[MODEL_TYPE]"',
text,
)
# Special case when the model camel cased and upper cased names are the same for the old model (like for GPT2) but
# not the new one. We can't just do a replace in all the text and will need a special regex
if old_model_patterns.model_upper_cased == old_model_patterns.model_camel_cased:
old_model_value = old_model_patterns.model_upper_cased
if re.search(rf"{old_model_value}_[A-Z_]*[^A-Z_]", text) is not None:
text = re.sub(rf"{old_model_value}([A-Z_]*)([^a-zA-Z_])", r"[MODEL_UPPER_CASED]\1\2", text)
else:
attributes_to_check.append("model_upper_cased")
attributes_to_check.extend(["model_camel_cased", "model_lower_cased", "model_name"])
# Now let's replace every other attribute by their placeholder
for attr in attributes_to_check:
text = text.replace(getattr(old_model_patterns, attr), ATTRIBUTE_TO_PLACEHOLDER[attr])
# Finally we can replace the placeholder byt the new values.
replacements = []
for attr, placeholder in ATTRIBUTE_TO_PLACEHOLDER.items():
if placeholder in text:
replacements.append((getattr(old_model_patterns, attr), getattr(new_model_patterns, attr)))
text = text.replace(placeholder, getattr(new_model_patterns, attr))
# If we have two inconsistent replacements, we don't return anything (ex: GPT2->GPT_NEW and GPT2->GPTNew)
old_replacement_values = [old for old, new in replacements]
if len(set(old_replacement_values)) != len(old_replacement_values):
return text, ""
replacements = simplify_replacements(replacements)
replacements = [f"{old}->{new}" for old, new in replacements]
return text, ",".join(replacements)
def simplify_replacements(replacements):
"""
Simplify a list of replacement patterns to make sure there are no needless ones.
For instance in the sequence "Bert->BertNew, BertConfig->BertNewConfig, bert->bert_new", the replacement
"BertConfig->BertNewConfig" is implied by "Bert->BertNew" so not needed.
Args:
replacements (`List[Tuple[str, str]]`): List of patterns (old, new)
Returns:
`List[Tuple[str, str]]`: The list of patterns simplified.
"""
if len(replacements) <= 1:
# Nothing to simplify
return replacements
# Next let's sort replacements by length as a replacement can only "imply" another replacement if it's shorter.
replacements.sort(key=lambda x: len(x[0]))
idx = 0
while idx < len(replacements):
old, new = replacements[idx]
# Loop through all replacements after
j = idx + 1
while j < len(replacements):
old_2, new_2 = replacements[j]
# If the replacement is implied by the current one, we can drop it.
if old_2.replace(old, new) == new_2:
replacements.pop(j)
else:
j += 1
idx += 1
return replacements
def get_module_from_file(module_file: Union[str, os.PathLike]) -> str:
"""
Returns the module name corresponding to a module file.
"""
full_module_path = Path(module_file).absolute()
module_parts = full_module_path.with_suffix("").parts
# Find the first part named transformers, starting from the end.
idx = len(module_parts) - 1
while idx >= 0 and module_parts[idx] != "transformers":
idx -= 1
if idx < 0:
raise ValueError(f"{module_file} is not a transformers module.")
return ".".join(module_parts[idx:])
SPECIAL_PATTERNS = {
"_CHECKPOINT_FOR_DOC =": "checkpoint",
"_CONFIG_FOR_DOC =": "config_class",
"_TOKENIZER_FOR_DOC =": "tokenizer_class",
"_IMAGE_PROCESSOR_FOR_DOC =": "image_processor_class",
"_FEAT_EXTRACTOR_FOR_DOC =": "feature_extractor_class",
"_PROCESSOR_FOR_DOC =": "processor_class",
}
_re_class_func = re.compile(r"^(?:class|def)\s+([^\s:\(]+)\s*(?:\(|\:)", flags=re.MULTILINE)
def remove_attributes(obj, target_attr):
"""Remove `target_attr` in `obj`."""
lines = obj.split(os.linesep)
target_idx = None
for idx, line in enumerate(lines):
# search for assignment
if line.lstrip().startswith(f"{target_attr} = "):
target_idx = idx
break
# search for function/method definition
elif line.lstrip().startswith(f"def {target_attr}("):
target_idx = idx
break
# target not found
if target_idx is None:
return obj
line = lines[target_idx]
indent_level = find_indent(line)
# forward pass to find the ending of the block (including empty lines)
parsed = extract_block("\n".join(lines[target_idx:]), indent_level)
num_lines = len(parsed.split("\n"))
for idx in range(num_lines):
lines[target_idx + idx] = None
# backward pass to find comments or decorator
for idx in range(target_idx - 1, -1, -1):
line = lines[idx]
if (line.lstrip().startswith("#") or line.lstrip().startswith("@")) and find_indent(line) == indent_level:
lines[idx] = None
else:
break
new_obj = os.linesep.join([x for x in lines if x is not None])
return new_obj
def duplicate_module(
module_file: Union[str, os.PathLike],
old_model_patterns: ModelPatterns,
new_model_patterns: ModelPatterns,
dest_file: Optional[str] = None,
add_copied_from: bool = True,
attrs_to_remove: List[str] = None,
):
"""
Create a new module from an existing one and adapting all function and classes names from old patterns to new ones.
Args:
module_file (`str` or `os.PathLike`): Path to the module to duplicate.
old_model_patterns (`ModelPatterns`): The patterns for the old model.
new_model_patterns (`ModelPatterns`): The patterns for the new model.
dest_file (`str` or `os.PathLike`, *optional*): Path to the new module.
add_copied_from (`bool`, *optional*, defaults to `True`):
Whether or not to add `# Copied from` statements in the duplicated module.
"""
if dest_file is None:
dest_file = str(module_file).replace(
old_model_patterns.model_lower_cased, new_model_patterns.model_lower_cased
)
with open(module_file, "r", encoding="utf-8") as f:
content = f.read()
content = re.sub(r"# Copyright (\d+)\s", f"# Copyright {CURRENT_YEAR} ", content)
objects = parse_module_content(content)
# Loop and treat all objects
new_objects = []
for obj in objects:
special_pattern = False
for pattern, attr in SPECIAL_PATTERNS.items():
if pattern in obj:
obj = obj.replace(getattr(old_model_patterns, attr), getattr(new_model_patterns, attr))
new_objects.append(obj)
special_pattern = True
break
if special_pattern:
continue
# Regular classes functions
old_obj = obj
obj, replacement = replace_model_patterns(obj, old_model_patterns, new_model_patterns)
has_copied_from = re.search(r"^#\s+Copied from", obj, flags=re.MULTILINE) is not None
if add_copied_from and not has_copied_from and _re_class_func.search(obj) is not None and len(replacement) > 0:
# Copied from statement must be added just before the class/function definition, which may not be the
# first line because of decorators.
module_name = get_module_from_file(module_file)
old_object_name = _re_class_func.search(old_obj).groups()[0]
obj = add_content_to_text(
obj, f"# Copied from {module_name}.{old_object_name} with {replacement}", add_before=_re_class_func
)
# In all cases, we remove Copied from statement with indent on methods.
obj = re.sub("\n[ ]+# Copied from [^\n]*\n", "\n", obj)
new_objects.append(obj)
content = "\n".join(new_objects)
# Remove some attributes that we don't want to copy to the new file(s)
if attrs_to_remove is not None:
for attr in attrs_to_remove:
content = remove_attributes(content, target_attr=attr)
with open(dest_file, "w", encoding="utf-8") as f:
f.write(content)
def filter_framework_files(
files: List[Union[str, os.PathLike]], frameworks: Optional[List[str]] = None
) -> List[Union[str, os.PathLike]]:
"""
Filter a list of files to only keep the ones corresponding to a list of frameworks.
Args:
files (`List[Union[str, os.PathLike]]`): The list of files to filter.
frameworks (`List[str]`, *optional*): The list of allowed frameworks.
Returns:
`List[Union[str, os.PathLike]]`: The list of filtered files.
"""
if frameworks is None:
frameworks = get_default_frameworks()
framework_to_file = {}
others = []
for f in files:
parts = Path(f).name.split("_")
if "modeling" not in parts:
others.append(f)
continue
if "tf" in parts:
framework_to_file["tf"] = f
elif "flax" in parts:
framework_to_file["flax"] = f
else:
framework_to_file["pt"] = f
return [framework_to_file[f] for f in frameworks if f in framework_to_file] + others
def get_model_files(model_type: str, frameworks: Optional[List[str]] = None) -> Dict[str, Union[Path, List[Path]]]:
"""
Retrieves all the files associated to a model.
Args:
model_type (`str`): A valid model type (like "bert" or "gpt2")
frameworks (`List[str]`, *optional*):
If passed, will only keep the model files corresponding to the passed frameworks.
Returns:
`Dict[str, Union[Path, List[Path]]]`: A dictionary with the following keys:
- **doc_file** -- The documentation file for the model.
- **model_files** -- All the files in the model module.
- **test_files** -- The test files for the model.
"""
module_name = model_type_to_module_name(model_type)
model_module = TRANSFORMERS_PATH / "models" / module_name
model_files = list(model_module.glob("*.py"))
model_files = filter_framework_files(model_files, frameworks=frameworks)
doc_file = REPO_PATH / "docs" / "source" / "en" / "model_doc" / f"{model_type}.md"
# Basic pattern for test files
test_files = [
f"test_modeling_{module_name}.py",
f"test_modeling_tf_{module_name}.py",
f"test_modeling_flax_{module_name}.py",
f"test_tokenization_{module_name}.py",
f"test_image_processing_{module_name}.py",
f"test_feature_extraction_{module_name}.py",
f"test_processor_{module_name}.py",
]
test_files = filter_framework_files(test_files, frameworks=frameworks)
# Add the test directory
test_files = [REPO_PATH / "tests" / "models" / module_name / f for f in test_files]
# Filter by existing files
test_files = [f for f in test_files if f.exists()]
return {"doc_file": doc_file, "model_files": model_files, "module_name": module_name, "test_files": test_files}
_re_checkpoint_for_doc = re.compile(r"^_CHECKPOINT_FOR_DOC\s+=\s+(\S*)\s*$", flags=re.MULTILINE)
def find_base_model_checkpoint(
model_type: str, model_files: Optional[Dict[str, Union[Path, List[Path]]]] = None
) -> str:
"""
Finds the model checkpoint used in the docstrings for a given model.
Args:
model_type (`str`): A valid model type (like "bert" or "gpt2")
model_files (`Dict[str, Union[Path, List[Path]]`, *optional*):
The files associated to `model_type`. Can be passed to speed up the function, otherwise will be computed.
Returns:
`str`: The checkpoint used.
"""
if model_files is None:
model_files = get_model_files(model_type)
module_files = model_files["model_files"]
for fname in module_files:
if "modeling" not in str(fname):
continue
with open(fname, "r", encoding="utf-8") as f:
content = f.read()
if _re_checkpoint_for_doc.search(content) is not None:
checkpoint = _re_checkpoint_for_doc.search(content).groups()[0]
# Remove quotes
checkpoint = checkpoint.replace('"', "")
checkpoint = checkpoint.replace("'", "")
return checkpoint
# TODO: Find some kind of fallback if there is no _CHECKPOINT_FOR_DOC in any of the modeling file.
return ""
def get_default_frameworks():
"""
Returns the list of frameworks (PyTorch, TensorFlow, Flax) that are installed in the environment.
"""
frameworks = []
if is_torch_available():
frameworks.append("pt")
if is_tf_available():
frameworks.append("tf")
if is_flax_available():
frameworks.append("flax")
return frameworks
_re_model_mapping = re.compile("MODEL_([A-Z_]*)MAPPING_NAMES")
def retrieve_model_classes(model_type: str, frameworks: Optional[List[str]] = None) -> Dict[str, List[str]]:
"""
Retrieve the model classes associated to a given model.
Args:
model_type (`str`): A valid model type (like "bert" or "gpt2")
frameworks (`List[str]`, *optional*):
The frameworks to look for. Will default to `["pt", "tf", "flax"]`, passing a smaller list will restrict
the classes returned.
Returns:
`Dict[str, List[str]]`: A dictionary with one key per framework and the list of model classes associated to
that framework as values.
"""
if frameworks is None:
frameworks = get_default_frameworks()
modules = {
"pt": auto_module.modeling_auto if is_torch_available() else None,
"tf": auto_module.modeling_tf_auto if is_tf_available() else None,
"flax": auto_module.modeling_flax_auto if is_flax_available() else None,
}
model_classes = {}
for framework in frameworks:
new_model_classes = []
if modules[framework] is None:
raise ValueError(f"You selected {framework} in the frameworks, but it is not installed.")
model_mappings = [attr for attr in dir(modules[framework]) if _re_model_mapping.search(attr) is not None]
for model_mapping_name in model_mappings:
model_mapping = getattr(modules[framework], model_mapping_name)
if model_type in model_mapping:
new_model_classes.append(model_mapping[model_type])
if len(new_model_classes) > 0:
# Remove duplicates
model_classes[framework] = list(set(new_model_classes))
return model_classes
def retrieve_info_for_model(model_type, frameworks: Optional[List[str]] = None):
"""
Retrieves all the information from a given model_type.
Args:
model_type (`str`): A valid model type (like "bert" or "gpt2")
frameworks (`List[str]`, *optional*):
If passed, will only keep the info corresponding to the passed frameworks.
Returns:
`Dict`: A dictionary with the following keys:
- **frameworks** (`List[str]`): The list of frameworks that back this model type.
- **model_classes** (`Dict[str, List[str]]`): The model classes implemented for that model type.
- **model_files** (`Dict[str, Union[Path, List[Path]]]`): The files associated with that model type.
- **model_patterns** (`ModelPatterns`): The various patterns for the model.
"""
if model_type not in auto_module.MODEL_NAMES_MAPPING:
raise ValueError(f"{model_type} is not a valid model type.")
model_name = auto_module.MODEL_NAMES_MAPPING[model_type]
config_class = auto_module.configuration_auto.CONFIG_MAPPING_NAMES[model_type]
if model_type in auto_module.tokenization_auto.TOKENIZER_MAPPING_NAMES:
tokenizer_classes = auto_module.tokenization_auto.TOKENIZER_MAPPING_NAMES[model_type]
tokenizer_class = tokenizer_classes[0] if tokenizer_classes[0] is not None else tokenizer_classes[1]
else:
tokenizer_class = None
image_processor_classes = auto_module.image_processing_auto.IMAGE_PROCESSOR_MAPPING_NAMES.get(model_type, None)
if isinstance(image_processor_classes, tuple):
image_processor_class = image_processor_classes[0] # we take the slow image processor class.
else:
image_processor_class = image_processor_classes
feature_extractor_class = auto_module.feature_extraction_auto.FEATURE_EXTRACTOR_MAPPING_NAMES.get(model_type, None)
processor_class = auto_module.processing_auto.PROCESSOR_MAPPING_NAMES.get(model_type, None)
model_files = get_model_files(model_type, frameworks=frameworks)
model_camel_cased = config_class.replace("Config", "")
available_frameworks = []
for fname in model_files["model_files"]:
if "modeling_tf" in str(fname):
available_frameworks.append("tf")
elif "modeling_flax" in str(fname):
available_frameworks.append("flax")
elif "modeling" in str(fname):
available_frameworks.append("pt")
if frameworks is None:
frameworks = get_default_frameworks()
frameworks = [f for f in frameworks if f in available_frameworks]
model_classes = retrieve_model_classes(model_type, frameworks=frameworks)
model_upper_cased = model_camel_cased.upper()
model_patterns = ModelPatterns(
model_name,
checkpoint=find_base_model_checkpoint(model_type, model_files=model_files),
model_type=model_type,
model_camel_cased=model_camel_cased,
model_lower_cased=model_files["module_name"],
model_upper_cased=model_upper_cased,
config_class=config_class,
tokenizer_class=tokenizer_class,
image_processor_class=image_processor_class,
feature_extractor_class=feature_extractor_class,
processor_class=processor_class,
)
return {
"frameworks": frameworks,
"model_classes": model_classes,
"model_files": model_files,
"model_patterns": model_patterns,
}
def clean_frameworks_in_init(
init_file: Union[str, os.PathLike], frameworks: Optional[List[str]] = None, keep_processing: bool = True
):
"""
Removes all the import lines that don't belong to a given list of frameworks or concern tokenizers/feature
extractors/image processors/processors in an init.
Args:
init_file (`str` or `os.PathLike`): The path to the init to treat.
frameworks (`List[str]`, *optional*):
If passed, this will remove all imports that are subject to a framework not in frameworks
keep_processing (`bool`, *optional*, defaults to `True`):
Whether or not to keep the preprocessing (tokenizer, feature extractor, image processor, processor) imports
in the init.
"""
if frameworks is None:
frameworks = get_default_frameworks()
names = {"pt": "torch"}
to_remove = [names.get(f, f) for f in ["pt", "tf", "flax"] if f not in frameworks]
if not keep_processing:
to_remove.extend(["sentencepiece", "tokenizers", "vision"])
if len(to_remove) == 0:
# Nothing to do
return
remove_pattern = "|".join(to_remove)
re_conditional_imports = re.compile(rf"^\s*if not is_({remove_pattern})_available\(\):\s*$")
re_try = re.compile(r"\s*try:")
re_else = re.compile(r"\s*else:")
re_is_xxx_available = re.compile(rf"is_({remove_pattern})_available")
with open(init_file, "r", encoding="utf-8") as f:
content = f.read()
lines = content.split("\n")
new_lines = []
idx = 0
while idx < len(lines):
# Conditional imports in try-except-else blocks
if (re_conditional_imports.search(lines[idx]) is not None) and (re_try.search(lines[idx - 1]) is not None):
# Remove the preceding `try:`
new_lines.pop()
idx += 1
# Iterate until `else:`
while is_empty_line(lines[idx]) or re_else.search(lines[idx]) is None:
idx += 1
idx += 1
indent = find_indent(lines[idx])
while find_indent(lines[idx]) >= indent or is_empty_line(lines[idx]):
idx += 1
# Remove the import from utils
elif re_is_xxx_available.search(lines[idx]) is not None:
line = lines[idx]
for framework in to_remove:
line = line.replace(f", is_{framework}_available", "")
line = line.replace(f"is_{framework}_available, ", "")
line = line.replace(f"is_{framework}_available,", "")
line = line.replace(f"is_{framework}_available", "")
if len(line.strip()) > 0:
new_lines.append(line)
idx += 1
# Otherwise we keep the line, except if it's a tokenizer import and we don't want to keep it.
elif keep_processing or (
re.search(r'^\s*"(tokenization|processing|feature_extraction|image_processing)', lines[idx]) is None
and re.search(r"^\s*from .(tokenization|processing|feature_extraction|image_processing)", lines[idx])
is None
):
new_lines.append(lines[idx])
idx += 1
else:
idx += 1
with open(init_file, "w", encoding="utf-8") as f:
f.write("\n".join(new_lines))
def add_model_to_main_init(
old_model_patterns: ModelPatterns,
new_model_patterns: ModelPatterns,
frameworks: Optional[List[str]] = None,
with_processing: bool = True,
):
"""
Add a model to the main init of Transformers.
Args:
old_model_patterns (`ModelPatterns`): The patterns for the old model.
new_model_patterns (`ModelPatterns`): The patterns for the new model.
frameworks (`List[str]`, *optional*):
If specified, only the models implemented in those frameworks will be added.
with_processsing (`bool`, *optional*, defaults to `True`):
Whether the tokenizer/feature extractor/processor of the model should also be added to the init or not.
"""
with open(TRANSFORMERS_PATH / "__init__.py", "r", encoding="utf-8") as f:
content = f.read()
lines = content.split("\n")
idx = 0
new_lines = []
framework = None
while idx < len(lines):
new_framework = False
if not is_empty_line(lines[idx]) and find_indent(lines[idx]) == 0:
framework = None
elif lines[idx].lstrip().startswith("if not is_torch_available"):
framework = "pt"
new_framework = True
elif lines[idx].lstrip().startswith("if not is_tf_available"):
framework = "tf"
new_framework = True
elif lines[idx].lstrip().startswith("if not is_flax_available"):
framework = "flax"
new_framework = True
if new_framework:
# For a new framework, we need to skip until the else: block to get where the imports are.
while lines[idx].strip() != "else:":
new_lines.append(lines[idx])
idx += 1
# Skip if we are in a framework not wanted.
if framework is not None and frameworks is not None and framework not in frameworks:
new_lines.append(lines[idx])
idx += 1
elif re.search(rf'models.{old_model_patterns.model_lower_cased}( |")', lines[idx]) is not None:
block = [lines[idx]]
indent = find_indent(lines[idx])
idx += 1
while find_indent(lines[idx]) > indent:
block.append(lines[idx])
idx += 1
if lines[idx].strip() in [")", "]", "],"]:
block.append(lines[idx])
idx += 1
block = "\n".join(block)
new_lines.append(block)
add_block = True
if not with_processing:
processing_classes = [
old_model_patterns.tokenizer_class,
old_model_patterns.image_processor_class,
old_model_patterns.feature_extractor_class,
old_model_patterns.processor_class,
]
# Only keep the ones that are not None
processing_classes = [c for c in processing_classes if c is not None]
for processing_class in processing_classes:
block = block.replace(f' "{processing_class}",', "")
block = block.replace(f', "{processing_class}"', "")
block = block.replace(f" {processing_class},", "")
block = block.replace(f", {processing_class}", "")
if processing_class in block:
add_block = False
if add_block:
new_lines.append(replace_model_patterns(block, old_model_patterns, new_model_patterns)[0])
else:
new_lines.append(lines[idx])
idx += 1
with open(TRANSFORMERS_PATH / "__init__.py", "w", encoding="utf-8") as f:
f.write("\n".join(new_lines))
def insert_tokenizer_in_auto_module(old_model_patterns: ModelPatterns, new_model_patterns: ModelPatterns):
"""
Add a tokenizer to the relevant mappings in the auto module.
Args:
old_model_patterns (`ModelPatterns`): The patterns for the old model.
new_model_patterns (`ModelPatterns`): The patterns for the new model.
"""
if old_model_patterns.tokenizer_class is None or new_model_patterns.tokenizer_class is None:
return
with open(TRANSFORMERS_PATH / "models" / "auto" / "tokenization_auto.py", "r", encoding="utf-8") as f:
content = f.read()
lines = content.split("\n")
idx = 0
# First we get to the TOKENIZER_MAPPING_NAMES block.
while not lines[idx].startswith(" TOKENIZER_MAPPING_NAMES = OrderedDict("):
idx += 1
idx += 1
# That block will end at this prompt:
while not lines[idx].startswith("TOKENIZER_MAPPING = _LazyAutoMapping"):
# Either all the tokenizer block is defined on one line, in which case, it ends with "),"
if lines[idx].endswith(","):
block = lines[idx]
# Otherwise it takes several lines until we get to a "),"
else:
block = []
while not lines[idx].startswith(" ),"):
block.append(lines[idx])
idx += 1
block = "\n".join(block)
idx += 1
# If we find the model type and tokenizer class in that block, we have the old model tokenizer block
if f'"{old_model_patterns.model_type}"' in block and old_model_patterns.tokenizer_class in block:
break
new_block = block.replace(old_model_patterns.model_type, new_model_patterns.model_type)
new_block = new_block.replace(old_model_patterns.tokenizer_class, new_model_patterns.tokenizer_class)
new_lines = lines[:idx] + [new_block] + lines[idx:]
with open(TRANSFORMERS_PATH / "models" / "auto" / "tokenization_auto.py", "w", encoding="utf-8") as f:
f.write("\n".join(new_lines))
AUTO_CLASSES_PATTERNS = {
"configuration_auto.py": [
' ("{model_type}", "{model_name}"),',
' ("{model_type}", "{config_class}"),',
' ("{model_type}", "{pretrained_archive_map}"),',
],
"feature_extraction_auto.py": [' ("{model_type}", "{feature_extractor_class}"),'],
"image_processing_auto.py": [' ("{model_type}", "{image_processor_class}"),'],
"modeling_auto.py": [' ("{model_type}", "{any_pt_class}"),'],
"modeling_tf_auto.py": [' ("{model_type}", "{any_tf_class}"),'],
"modeling_flax_auto.py": [' ("{model_type}", "{any_flax_class}"),'],
"processing_auto.py": [' ("{model_type}", "{processor_class}"),'],
}
def add_model_to_auto_classes(
old_model_patterns: ModelPatterns, new_model_patterns: ModelPatterns, model_classes: Dict[str, List[str]]
):
"""
Add a model to the relevant mappings in the auto module.
Args:
old_model_patterns (`ModelPatterns`): The patterns for the old model.
new_model_patterns (`ModelPatterns`): The patterns for the new model.
model_classes (`Dict[str, List[str]]`): A dictionary framework to list of model classes implemented.
"""
for filename in AUTO_CLASSES_PATTERNS:
# Extend patterns with all model classes if necessary
new_patterns = []
for pattern in AUTO_CLASSES_PATTERNS[filename]:
if re.search("any_([a-z]*)_class", pattern) is not None:
framework = re.search("any_([a-z]*)_class", pattern).groups()[0]
if framework in model_classes:
new_patterns.extend(
[
pattern.replace("{" + f"any_{framework}_class" + "}", cls)
for cls in model_classes[framework]
]
)
elif "{config_class}" in pattern:
new_patterns.append(pattern.replace("{config_class}", old_model_patterns.config_class))
elif "{image_processor_class}" in pattern:
if (
old_model_patterns.image_processor_class is not None
and new_model_patterns.image_processor_class is not None
):
new_patterns.append(
pattern.replace("{image_processor_class}", old_model_patterns.image_processor_class)
)
elif "{feature_extractor_class}" in pattern:
if (
old_model_patterns.feature_extractor_class is not None
and new_model_patterns.feature_extractor_class is not None
):
new_patterns.append(
pattern.replace("{feature_extractor_class}", old_model_patterns.feature_extractor_class)
)
elif "{processor_class}" in pattern:
if old_model_patterns.processor_class is not None and new_model_patterns.processor_class is not None:
new_patterns.append(pattern.replace("{processor_class}", old_model_patterns.processor_class))
else:
new_patterns.append(pattern)
# Loop through all patterns.
for pattern in new_patterns:
full_name = TRANSFORMERS_PATH / "models" / "auto" / filename
old_model_line = pattern
new_model_line = pattern
for attr in ["model_type", "model_name"]:
old_model_line = old_model_line.replace("{" + attr + "}", getattr(old_model_patterns, attr))
new_model_line = new_model_line.replace("{" + attr + "}", getattr(new_model_patterns, attr))
new_model_line = new_model_line.replace(
old_model_patterns.model_camel_cased, new_model_patterns.model_camel_cased
)
add_content_to_file(full_name, new_model_line, add_after=old_model_line)
# Tokenizers require special handling
insert_tokenizer_in_auto_module(old_model_patterns, new_model_patterns)
DOC_OVERVIEW_TEMPLATE = """## Overview
The {model_name} model was proposed in [<INSERT PAPER NAME HERE>](<INSERT PAPER LINK HERE>) by <INSERT AUTHORS HERE>.
<INSERT SHORT SUMMARY HERE>
The abstract from the paper is the following:
*<INSERT PAPER ABSTRACT HERE>*
Tips:
<INSERT TIPS ABOUT MODEL HERE>
This model was contributed by [INSERT YOUR HF USERNAME HERE](https://huggingface.co/<INSERT YOUR HF USERNAME HERE>).
The original code can be found [here](<INSERT LINK TO GITHUB REPO HERE>).
"""
def duplicate_doc_file(
doc_file: Union[str, os.PathLike],
old_model_patterns: ModelPatterns,
new_model_patterns: ModelPatterns,
dest_file: Optional[Union[str, os.PathLike]] = None,
frameworks: Optional[List[str]] = None,
):
"""
Duplicate a documentation file and adapts it for a new model.
Args:
module_file (`str` or `os.PathLike`): Path to the doc file to duplicate.
old_model_patterns (`ModelPatterns`): The patterns for the old model.
new_model_patterns (`ModelPatterns`): The patterns for the new model.
dest_file (`str` or `os.PathLike`, *optional*): Path to the new doc file.
Will default to the a file named `{new_model_patterns.model_type}.md` in the same folder as `module_file`.
frameworks (`List[str]`, *optional*):
If passed, will only keep the model classes corresponding to this list of frameworks in the new doc file.
"""
with open(doc_file, "r", encoding="utf-8") as f:
content = f.read()
content = re.sub(r"<!--\s*Copyright (\d+)\s", f"<!--Copyright {CURRENT_YEAR} ", content)
if frameworks is None:
frameworks = get_default_frameworks()
if dest_file is None:
dest_file = Path(doc_file).parent / f"{new_model_patterns.model_type}.md"
# Parse the doc file in blocks. One block per section/header
lines = content.split("\n")
blocks = []
current_block = []
for line in lines:
if line.startswith("#"):
blocks.append("\n".join(current_block))
current_block = [line]
else:
current_block.append(line)
blocks.append("\n".join(current_block))
new_blocks = []
in_classes = False
for block in blocks:
# Copyright
if not block.startswith("#"):
new_blocks.append(block)
# Main title
elif re.search(r"^#\s+\S+", block) is not None:
new_blocks.append(f"# {new_model_patterns.model_name}\n")
# The config starts the part of the doc with the classes.
elif not in_classes and old_model_patterns.config_class in block.split("\n")[0]:
in_classes = True
new_blocks.append(DOC_OVERVIEW_TEMPLATE.format(model_name=new_model_patterns.model_name))
new_block, _ = replace_model_patterns(block, old_model_patterns, new_model_patterns)
new_blocks.append(new_block)
# In classes
elif in_classes:
in_classes = True
block_title = block.split("\n")[0]
block_class = re.search(r"^#+\s+(\S.*)$", block_title).groups()[0]
new_block, _ = replace_model_patterns(block, old_model_patterns, new_model_patterns)
if "Tokenizer" in block_class:
# We only add the tokenizer if necessary
if old_model_patterns.tokenizer_class != new_model_patterns.tokenizer_class:
new_blocks.append(new_block)
elif "ImageProcessor" in block_class:
# We only add the image processor if necessary
if old_model_patterns.image_processor_class != new_model_patterns.image_processor_class:
new_blocks.append(new_block)
elif "FeatureExtractor" in block_class:
# We only add the feature extractor if necessary
if old_model_patterns.feature_extractor_class != new_model_patterns.feature_extractor_class:
new_blocks.append(new_block)
elif "Processor" in block_class:
# We only add the processor if necessary
if old_model_patterns.processor_class != new_model_patterns.processor_class:
new_blocks.append(new_block)
elif block_class.startswith("Flax"):
# We only add Flax models if in the selected frameworks
if "flax" in frameworks:
new_blocks.append(new_block)
elif block_class.startswith("TF"):
# We only add TF models if in the selected frameworks
if "tf" in frameworks:
new_blocks.append(new_block)
elif len(block_class.split(" ")) == 1:
# We only add PyTorch models if in the selected frameworks
if "pt" in frameworks:
new_blocks.append(new_block)
else:
new_blocks.append(new_block)
with open(dest_file, "w", encoding="utf-8") as f:
f.write("\n".join(new_blocks))
def insert_model_in_doc_toc(old_model_patterns, new_model_patterns):
"""
Insert the new model in the doc TOC, in the same section as the old model.
Args:
old_model_patterns (`ModelPatterns`): The patterns for the old model.
new_model_patterns (`ModelPatterns`): The patterns for the new model.
"""
toc_file = REPO_PATH / "docs" / "source" / "en" / "_toctree.yml"
with open(toc_file, "r", encoding="utf8") as f:
content = yaml.safe_load(f)
# Get to the model API doc
api_idx = 0
while content[api_idx]["title"] != "API":
api_idx += 1
api_doc = content[api_idx]["sections"]
model_idx = 0
while api_doc[model_idx]["title"] != "Models":
model_idx += 1
model_doc = api_doc[model_idx]["sections"]
# Find the base model in the Toc
old_model_type = old_model_patterns.model_type
section_idx = 0
while section_idx < len(model_doc):
sections = [entry["local"] for entry in model_doc[section_idx]["sections"]]
if f"model_doc/{old_model_type}" in sections:
break
section_idx += 1
if section_idx == len(model_doc):
old_model = old_model_patterns.model_name
new_model = new_model_patterns.model_name
print(f"Did not find {old_model} in the table of content, so you will need to add {new_model} manually.")
return
# Add the new model in the same toc
toc_entry = {"local": f"model_doc/{new_model_patterns.model_type}", "title": new_model_patterns.model_name}
model_doc[section_idx]["sections"].append(toc_entry)
model_doc[section_idx]["sections"] = sorted(model_doc[section_idx]["sections"], key=lambda s: s["title"].lower())
api_doc[model_idx]["sections"] = model_doc
content[api_idx]["sections"] = api_doc
with open(toc_file, "w", encoding="utf-8") as f:
f.write(yaml.dump(content, allow_unicode=True))
def create_new_model_like(
model_type: str,
new_model_patterns: ModelPatterns,
add_copied_from: bool = True,
frameworks: Optional[List[str]] = None,
old_checkpoint: Optional[str] = None,
):
"""
Creates a new model module like a given model of the Transformers library.
Args:
model_type (`str`): The model type to duplicate (like "bert" or "gpt2")
new_model_patterns (`ModelPatterns`): The patterns for the new model.
add_copied_from (`bool`, *optional*, defaults to `True`):
Whether or not to add "Copied from" statements to all classes in the new model modeling files.
frameworks (`List[str]`, *optional*):
If passed, will limit the duplicate to the frameworks specified.
old_checkpoint (`str`, *optional*):
The name of the base checkpoint for the old model. Should be passed along when it can't be automatically
recovered from the `model_type`.
"""
# Retrieve all the old model info.
model_info = retrieve_info_for_model(model_type, frameworks=frameworks)
model_files = model_info["model_files"]
old_model_patterns = model_info["model_patterns"]
if old_checkpoint is not None:
old_model_patterns.checkpoint = old_checkpoint
if len(old_model_patterns.checkpoint) == 0:
raise ValueError(
"The old model checkpoint could not be recovered from the model type. Please pass it to the "
"`old_checkpoint` argument."
)
keep_old_processing = True
for processing_attr in ["image_processor_class", "feature_extractor_class", "processor_class", "tokenizer_class"]:
if getattr(old_model_patterns, processing_attr) != getattr(new_model_patterns, processing_attr):
keep_old_processing = False
model_classes = model_info["model_classes"]
# 1. We create the module for our new model.
old_module_name = model_files["module_name"]
module_folder = TRANSFORMERS_PATH / "models" / new_model_patterns.model_lower_cased
os.makedirs(module_folder, exist_ok=True)
files_to_adapt = model_files["model_files"]
if keep_old_processing:
files_to_adapt = [
f
for f in files_to_adapt
if "tokenization" not in str(f)
and "processing" not in str(f)
and "feature_extraction" not in str(f)
and "image_processing" not in str(f)
]
os.makedirs(module_folder, exist_ok=True)
for module_file in files_to_adapt:
new_module_name = module_file.name.replace(
old_model_patterns.model_lower_cased, new_model_patterns.model_lower_cased
)
dest_file = module_folder / new_module_name
duplicate_module(
module_file,
old_model_patterns,
new_model_patterns,
dest_file=dest_file,
add_copied_from=add_copied_from and "modeling" in new_module_name,
)
clean_frameworks_in_init(
module_folder / "__init__.py", frameworks=frameworks, keep_processing=not keep_old_processing
)
# 2. We add our new model to the models init and the main init
add_content_to_file(
TRANSFORMERS_PATH / "models" / "__init__.py",
f" {new_model_patterns.model_lower_cased},",
add_after=f" {old_module_name},",
exact_match=True,
)
add_model_to_main_init(
old_model_patterns, new_model_patterns, frameworks=frameworks, with_processing=not keep_old_processing
)
# 3. Add test files
files_to_adapt = model_files["test_files"]
if keep_old_processing:
files_to_adapt = [
f
for f in files_to_adapt
if "tokenization" not in str(f)
and "processor" not in str(f)
and "feature_extraction" not in str(f)
and "image_processing" not in str(f)
]
def disable_fx_test(filename: Path) -> bool:
with open(filename) as fp:
content = fp.read()
new_content = re.sub(r"fx_compatible\s*=\s*True", "fx_compatible = False", content)
with open(filename, "w") as fp:
fp.write(new_content)
return content != new_content
disabled_fx_test = False
tests_folder = REPO_PATH / "tests" / "models" / new_model_patterns.model_lower_cased
os.makedirs(tests_folder, exist_ok=True)
with open(tests_folder / "__init__.py", "w"):
pass
for test_file in files_to_adapt:
new_test_file_name = test_file.name.replace(
old_model_patterns.model_lower_cased, new_model_patterns.model_lower_cased
)
dest_file = test_file.parent.parent / new_model_patterns.model_lower_cased / new_test_file_name
duplicate_module(
test_file,
old_model_patterns,
new_model_patterns,
dest_file=dest_file,
add_copied_from=False,
attrs_to_remove=["pipeline_model_mapping", "is_pipeline_test_to_skip"],
)
disabled_fx_test = disabled_fx_test | disable_fx_test(dest_file)
if disabled_fx_test:
print(
"The tests for symbolic tracing with torch.fx were disabled, you can add those once symbolic tracing works"
" for your new model."
)
# 4. Add model to auto classes
add_model_to_auto_classes(old_model_patterns, new_model_patterns, model_classes)
# 5. Add doc file
doc_file = REPO_PATH / "docs" / "source" / "en" / "model_doc" / f"{old_model_patterns.model_type}.md"
duplicate_doc_file(doc_file, old_model_patterns, new_model_patterns, frameworks=frameworks)
insert_model_in_doc_toc(old_model_patterns, new_model_patterns)
# 6. Warn the user for duplicate patterns
if old_model_patterns.model_type == old_model_patterns.checkpoint:
print(
"The model you picked has the same name for the model type and the checkpoint name "
f"({old_model_patterns.model_type}). As a result, it's possible some places where the new checkpoint "
f"should be, you have {new_model_patterns.model_type} instead. You should search for all instances of "
f"{new_model_patterns.model_type} in the new files and check they're not badly used as checkpoints."
)
elif old_model_patterns.model_lower_cased == old_model_patterns.checkpoint:
print(
"The model you picked has the same name for the model type and the checkpoint name "
f"({old_model_patterns.model_lower_cased}). As a result, it's possible some places where the new "
f"checkpoint should be, you have {new_model_patterns.model_lower_cased} instead. You should search for "
f"all instances of {new_model_patterns.model_lower_cased} in the new files and check they're not badly "
"used as checkpoints."
)
if (
old_model_patterns.model_type == old_model_patterns.model_lower_cased
and new_model_patterns.model_type != new_model_patterns.model_lower_cased
):
print(
"The model you picked has the same name for the model type and the lowercased model name "
f"({old_model_patterns.model_lower_cased}). As a result, it's possible some places where the new "
f"model type should be, you have {new_model_patterns.model_lower_cased} instead. You should search for "
f"all instances of {new_model_patterns.model_lower_cased} in the new files and check they're not badly "
"used as the model type."
)
if not keep_old_processing and old_model_patterns.tokenizer_class is not None:
print(
"The constants at the start of the new tokenizer file created needs to be manually fixed. If your new "
"model has a tokenizer fast, you will also need to manually add the converter in the "
"`SLOW_TO_FAST_CONVERTERS` constant of `convert_slow_tokenizer.py`."
)
def add_new_model_like_command_factory(args: Namespace):
return AddNewModelLikeCommand(config_file=args.config_file, path_to_repo=args.path_to_repo)
class AddNewModelLikeCommand(BaseTransformersCLICommand):
@staticmethod
def register_subcommand(parser: ArgumentParser):
add_new_model_like_parser = parser.add_parser("add-new-model-like")
add_new_model_like_parser.add_argument(
"--config_file", type=str, help="A file with all the information for this model creation."
)
add_new_model_like_parser.add_argument(
"--path_to_repo", type=str, help="When not using an editable install, the path to the Transformers repo."
)
add_new_model_like_parser.set_defaults(func=add_new_model_like_command_factory)
def __init__(self, config_file=None, path_to_repo=None, *args):
if config_file is not None:
with open(config_file, "r", encoding="utf-8") as f:
config = json.load(f)
self.old_model_type = config["old_model_type"]
self.model_patterns = ModelPatterns(**config["new_model_patterns"])
self.add_copied_from = config.get("add_copied_from", True)
self.frameworks = config.get("frameworks", get_default_frameworks())
self.old_checkpoint = config.get("old_checkpoint", None)
else:
(
self.old_model_type,
self.model_patterns,
self.add_copied_from,
self.frameworks,
self.old_checkpoint,
) = get_user_input()
self.path_to_repo = path_to_repo
def run(self):
if self.path_to_repo is not None:
# Adapt constants
global TRANSFORMERS_PATH
global REPO_PATH
REPO_PATH = Path(self.path_to_repo)
TRANSFORMERS_PATH = REPO_PATH / "src" / "transformers"
create_new_model_like(
model_type=self.old_model_type,
new_model_patterns=self.model_patterns,
add_copied_from=self.add_copied_from,
frameworks=self.frameworks,
old_checkpoint=self.old_checkpoint,
)
def get_user_field(
question: str,
default_value: Optional[str] = None,
is_valid_answer: Optional[Callable] = None,
convert_to: Optional[Callable] = None,
fallback_message: Optional[str] = None,
) -> Any:
"""
A utility function that asks a question to the user to get an answer, potentially looping until it gets a valid
answer.
Args:
question (`str`): The question to ask the user.
default_value (`str`, *optional*): A potential default value that will be used when the answer is empty.
is_valid_answer (`Callable`, *optional*):
If set, the question will be asked until this function returns `True` on the provided answer.
convert_to (`Callable`, *optional*):
If set, the answer will be passed to this function. If this function raises an error on the procided
answer, the question will be asked again.
fallback_message (`str`, *optional*):
A message that will be displayed each time the question is asked again to the user.
Returns:
`Any`: The answer provided by the user (or the default), passed through the potential conversion function.
"""
if not question.endswith(" "):
question = question + " "
if default_value is not None:
question = f"{question} [{default_value}] "
valid_answer = False
while not valid_answer:
answer = input(question)
if default_value is not None and len(answer) == 0:
answer = default_value
if is_valid_answer is not None:
valid_answer = is_valid_answer(answer)
elif convert_to is not None:
try:
answer = convert_to(answer)
valid_answer = True
except Exception:
valid_answer = False
else:
valid_answer = True
if not valid_answer:
print(fallback_message)
return answer
def convert_to_bool(x: str) -> bool:
"""
Converts a string to a bool.
"""
if x.lower() in ["1", "y", "yes", "true"]:
return True
if x.lower() in ["0", "n", "no", "false"]:
return False
raise ValueError(f"{x} is not a value that can be converted to a bool.")
def get_user_input():
"""
Ask the user for the necessary inputs to add the new model.
"""
model_types = list(auto_module.configuration_auto.MODEL_NAMES_MAPPING.keys())
# Get old model type
valid_model_type = False
while not valid_model_type:
old_model_type = input(
"What is the model you would like to duplicate? Please provide the lowercase `model_type` (e.g. roberta): "
)
if old_model_type in model_types:
valid_model_type = True
else:
print(f"{old_model_type} is not a valid model type.")
near_choices = difflib.get_close_matches(old_model_type, model_types)
if len(near_choices) >= 1:
if len(near_choices) > 1:
near_choices = " or ".join(near_choices)
print(f"Did you mean {near_choices}?")
old_model_info = retrieve_info_for_model(old_model_type)
old_tokenizer_class = old_model_info["model_patterns"].tokenizer_class
old_image_processor_class = old_model_info["model_patterns"].image_processor_class
old_feature_extractor_class = old_model_info["model_patterns"].feature_extractor_class
old_processor_class = old_model_info["model_patterns"].processor_class
old_frameworks = old_model_info["frameworks"]
old_checkpoint = None
if len(old_model_info["model_patterns"].checkpoint) == 0:
old_checkpoint = get_user_field(
"We couldn't find the name of the base checkpoint for that model, please enter it here."
)
model_name = get_user_field(
"What is the name (with no special casing) for your new model in the paper (e.g. RoBERTa)? "
)
default_patterns = ModelPatterns(model_name, model_name)
model_type = get_user_field(
"What identifier would you like to use for the `model_type` of this model? ",
default_value=default_patterns.model_type,
)
model_lower_cased = get_user_field(
"What lowercase name would you like to use for the module (folder) of this model? ",
default_value=default_patterns.model_lower_cased,
)
model_camel_cased = get_user_field(
"What prefix (camel-cased) would you like to use for the model classes of this model (e.g. Roberta)? ",
default_value=default_patterns.model_camel_cased,
)
model_upper_cased = get_user_field(
"What prefix (upper-cased) would you like to use for the constants relative to this model? ",
default_value=default_patterns.model_upper_cased,
)
config_class = get_user_field(
"What will be the name of the config class for this model? ", default_value=f"{model_camel_cased}Config"
)
checkpoint = get_user_field(
"Please give a checkpoint identifier (on the model Hub) for this new model (e.g. facebook/FacebookAI/roberta-base): "
)
old_processing_classes = [
c if not isinstance(c, tuple) else c[0]
for c in [old_image_processor_class, old_feature_extractor_class, old_tokenizer_class, old_processor_class]
if c is not None
]
old_processing_classes = ", ".join(old_processing_classes)
keep_processing = get_user_field(
f"Will your new model use the same processing class as {old_model_type} ({old_processing_classes}) (yes/no)? ",
convert_to=convert_to_bool,
fallback_message="Please answer yes/no, y/n, true/false or 1/0. ",
)
if keep_processing:
image_processor_class = old_image_processor_class
feature_extractor_class = old_feature_extractor_class
processor_class = old_processor_class
tokenizer_class = old_tokenizer_class
else:
if old_tokenizer_class is not None:
tokenizer_class = get_user_field(
"What will be the name of the tokenizer class for this model? ",
default_value=f"{model_camel_cased}Tokenizer",
)
else:
tokenizer_class = None
if old_image_processor_class is not None:
image_processor_class = get_user_field(
"What will be the name of the image processor class for this model? ",
default_value=f"{model_camel_cased}ImageProcessor",
)
else:
image_processor_class = None
if old_feature_extractor_class is not None:
feature_extractor_class = get_user_field(
"What will be the name of the feature extractor class for this model? ",
default_value=f"{model_camel_cased}FeatureExtractor",
)
else:
feature_extractor_class = None
if old_processor_class is not None:
processor_class = get_user_field(
"What will be the name of the processor class for this model? ",
default_value=f"{model_camel_cased}Processor",
)
else:
processor_class = None
model_patterns = ModelPatterns(
model_name,
checkpoint,
model_type=model_type,
model_lower_cased=model_lower_cased,
model_camel_cased=model_camel_cased,
model_upper_cased=model_upper_cased,
config_class=config_class,
tokenizer_class=tokenizer_class,
image_processor_class=image_processor_class,
feature_extractor_class=feature_extractor_class,
processor_class=processor_class,
)
add_copied_from = get_user_field(
"Should we add # Copied from statements when creating the new modeling file (yes/no)? ",
convert_to=convert_to_bool,
default_value="yes",
fallback_message="Please answer yes/no, y/n, true/false or 1/0.",
)
all_frameworks = get_user_field(
"Should we add a version of your new model in all the frameworks implemented by"
f" {old_model_type} ({old_frameworks}) (yes/no)? ",
convert_to=convert_to_bool,
default_value="yes",
fallback_message="Please answer yes/no, y/n, true/false or 1/0.",
)
if all_frameworks:
frameworks = None
else:
frameworks = get_user_field(
"Please enter the list of framworks you want (pt, tf, flax) separated by spaces",
is_valid_answer=lambda x: all(p in ["pt", "tf", "flax"] for p in x.split(" ")),
)
frameworks = list(set(frameworks.split(" ")))
return (old_model_type, model_patterns, add_copied_from, frameworks, old_checkpoint)
|
transformers/src/transformers/commands/add_new_model_like.py/0
|
{
"file_path": "transformers/src/transformers/commands/add_new_model_like.py",
"repo_id": "transformers",
"token_count": 29705
}
| 335
|
# coding=utf-8
# Copyright 2020 The HuggingFace Inc. team.
#
# 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.
"""Convert Seq2Seq TF Hub checkpoint."""
import argparse
from . import (
BertConfig,
BertGenerationConfig,
BertGenerationDecoder,
BertGenerationEncoder,
load_tf_weights_in_bert_generation,
logging,
)
logging.set_verbosity_info()
def convert_tf_checkpoint_to_pytorch(tf_hub_path, pytorch_dump_path, is_encoder_named_decoder, vocab_size, is_encoder):
# Initialise PyTorch model
bert_config = BertConfig.from_pretrained(
"google-bert/bert-large-cased",
vocab_size=vocab_size,
max_position_embeddings=512,
is_decoder=True,
add_cross_attention=True,
)
bert_config_dict = bert_config.to_dict()
del bert_config_dict["type_vocab_size"]
config = BertGenerationConfig(**bert_config_dict)
if is_encoder:
model = BertGenerationEncoder(config)
else:
model = BertGenerationDecoder(config)
print(f"Building PyTorch model from configuration: {config}")
# Load weights from tf checkpoint
load_tf_weights_in_bert_generation(
model,
tf_hub_path,
model_class="bert",
is_encoder_named_decoder=is_encoder_named_decoder,
is_encoder=is_encoder,
)
# Save pytorch-model
print(f"Save PyTorch model and config to {pytorch_dump_path}")
model.save_pretrained(pytorch_dump_path)
if __name__ == "__main__":
parser = argparse.ArgumentParser()
# Required parameters
parser.add_argument(
"--tf_hub_path", default=None, type=str, required=True, help="Path to the TensorFlow checkpoint path."
)
parser.add_argument(
"--pytorch_dump_path", default=None, type=str, required=True, help="Path to the output PyTorch model."
)
parser.add_argument(
"--is_encoder_named_decoder",
action="store_true",
help="If decoder has to be renamed to encoder in PyTorch model.",
)
parser.add_argument("--is_encoder", action="store_true", help="If model is an encoder.")
parser.add_argument("--vocab_size", default=50358, type=int, help="Vocab size of model")
args = parser.parse_args()
convert_tf_checkpoint_to_pytorch(
args.tf_hub_path,
args.pytorch_dump_path,
args.is_encoder_named_decoder,
args.vocab_size,
is_encoder=args.is_encoder,
)
|
transformers/src/transformers/convert_tf_hub_seq_to_seq_bert_to_pytorch.py/0
|
{
"file_path": "transformers/src/transformers/convert_tf_hub_seq_to_seq_bert_to_pytorch.py",
"repo_id": "transformers",
"token_count": 1134
}
| 336
|
# Copyright 2020 The HuggingFace Team. 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.
from .dependency_versions_table import deps
from .utils.versions import require_version, require_version_core
# define which module versions we always want to check at run time
# (usually the ones defined in `install_requires` in setup.py)
#
# order specific notes:
# - tqdm must be checked before tokenizers
pkgs_to_check_at_runtime = [
"python",
"tqdm",
"regex",
"requests",
"packaging",
"filelock",
"numpy",
"tokenizers",
"huggingface-hub",
"safetensors",
"accelerate",
"pyyaml",
]
for pkg in pkgs_to_check_at_runtime:
if pkg in deps:
if pkg == "tokenizers":
# must be loaded here, or else tqdm check may fail
from .utils import is_tokenizers_available
if not is_tokenizers_available():
continue # not required, check version only if installed
elif pkg == "accelerate":
# must be loaded here, or else tqdm check may fail
from .utils import is_accelerate_available
# Maybe switch to is_torch_available in the future here so that Accelerate is hard dep of
# Transformers with PyTorch
if not is_accelerate_available():
continue # not required, check version only if installed
require_version_core(deps[pkg])
else:
raise ValueError(f"can't find {pkg} in {deps.keys()}, check dependency_versions_table.py")
def dep_version_check(pkg, hint=None):
require_version(deps[pkg], hint)
|
transformers/src/transformers/dependency_versions_check.py/0
|
{
"file_path": "transformers/src/transformers/dependency_versions_check.py",
"repo_id": "transformers",
"token_count": 759
}
| 337
|
# coding=utf-8
# Copyright 2022 The HuggingFace Inc. team
#
# 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.
import inspect
from typing import List, Tuple
import numpy as np
import tensorflow as tf
from ..tf_utils import stable_softmax
from ..utils import add_start_docstrings
from ..utils.logging import get_logger
logger = get_logger(__name__)
TF_LOGITS_PROCESSOR_INPUTS_DOCSTRING = r"""
Args:
input_ids (`tf.Tensor` of shape `(batch_size, sequence_length)`):
Indices of input sequence tokens in the vocabulary.
Indices can be obtained using [`PreTrainedTokenizer`]. See [`PreTrainedTokenizer.encode`] and
[`PreTrainedTokenizer.__call__`] for details.
[What are input IDs?](../glossary#input-ids)
scores (`tf.Tensor` of shape `(batch_size, config.vocab_size)`):
Prediction scores of a language modeling head. These can be logits for each vocabulary when not using beam
search or log softmax for each vocabulary token when using beam search.
cur_len (`int`):
The current length of valid input sequence tokens. In the TF implementation, the input_ids' sequence length
is the maximum length generate can produce, and we need to know which of its tokens are valid.
kwargs (`Dict[str, Any]`, *optional*):
Additional logits processor specific kwargs.
Return:
`tf.Tensor` of shape `(batch_size, config.vocab_size)`: The processed prediction scores.
"""
class TFLogitsProcessor:
"""Abstract base class for all logit processors that can be applied during generation."""
@add_start_docstrings(TF_LOGITS_PROCESSOR_INPUTS_DOCSTRING)
def __call__(self, input_ids: tf.Tensor, scores: tf.Tensor, cur_len: int) -> tf.Tensor:
"""TF method for processing logits."""
raise NotImplementedError(
f"{self.__class__} is an abstract class. Only classes inheriting this class can be called."
)
class TFLogitsWarper:
"""Abstract base class for all logit warpers that can be applied during generation with multinomial sampling."""
@add_start_docstrings(TF_LOGITS_PROCESSOR_INPUTS_DOCSTRING)
def __call__(self, input_ids: tf.Tensor, scores: tf.Tensor, cur_len: int) -> tf.Tensor:
"""TF method for warping logits."""
raise NotImplementedError(
f"{self.__class__} is an abstract class. Only classes inheriting this class can be called."
)
class TFLogitsProcessorList(list):
"""
This class can be used to create a list of [`TFLogitsProcessor`] to subsequently process a `scores` input tensor.
This class inherits from list and adds a specific *__call__* method to apply each [`TFLogitsProcessor`] to the
inputs.
"""
@add_start_docstrings(TF_LOGITS_PROCESSOR_INPUTS_DOCSTRING)
def __call__(self, input_ids: tf.Tensor, scores: tf.Tensor, cur_len: int, **kwargs) -> tf.Tensor:
for processor in self:
function_args = inspect.signature(processor.__call__).parameters
if len(function_args) > 3:
if not all(arg in kwargs for arg in list(function_args.keys())[2:]):
raise ValueError(
f"Make sure that all the required parameters: {list(function_args.keys())} for "
f"{processor.__class__} are passed to the logits processor."
)
scores = processor(input_ids, scores, cur_len, **kwargs)
else:
scores = processor(input_ids, scores, cur_len)
return scores
class TFTemperatureLogitsWarper(TFLogitsWarper):
r"""
[`TFLogitsWarper`] for temperature (exponential scaling output probability distribution).
Args:
temperature (`float`):
The value used to module the logits distribution.
"""
def __init__(self, temperature: float):
if not isinstance(temperature, float) or not (temperature > 0):
raise ValueError(f"`temperature` has to be a strictly positive float, but is {temperature}")
self.temperature = temperature
def __call__(self, input_ids: tf.Tensor, scores: tf.Tensor, cur_len: int) -> tf.Tensor:
scores = scores / self.temperature
return scores
class TFTopKLogitsWarper(TFLogitsWarper):
r"""
[`TFLogitsWarper`] that performs top-k, i.e. restricting to the k highest probability elements.
Args:
top_k (`int`):
The number of highest probability vocabulary tokens to keep for top-k-filtering.
filter_value (`float`, *optional*, defaults to -inf):
All filtered values will be set to this float value.
min_tokens_to_keep (`int`, *optional*, defaults to 1):
Minimum number of tokens that cannot be filtered.
"""
def __init__(self, top_k: int, filter_value: float = -float("Inf"), min_tokens_to_keep: int = 1):
if not isinstance(top_k, int) or top_k <= 0:
raise ValueError(f"`top_k` has to be a strictly positive integer, but is {top_k}")
self.top_k = max(top_k, min_tokens_to_keep)
self.filter_value = filter_value
def __call__(self, input_ids: tf.Tensor, scores: tf.Tensor, cur_len: int) -> tf.Tensor:
top_k = min(self.top_k, scores.shape[-1]) # Safety check
# Boolean mask containing all tokens with a probability less than the last token of the top-k
indices_to_remove = scores < tf.math.top_k(scores, k=top_k)[0][..., -1:]
next_scores = tf.where(indices_to_remove, self.filter_value, scores)
return next_scores
class TFTopPLogitsWarper(TFLogitsWarper):
"""
[`TFLogitsWarper`] that performs top-p, i.e. restricting to top tokens summing to <= prob_cut_off.
Args:
top_p (`float`):
If set to < 1, only the smallest set of most probable tokens with probabilities that add up to `top_p` or
higher are kept for generation.
filter_value (`float`, *optional*, defaults to -inf):
All filtered values will be set to this float value.
min_tokens_to_keep (`int`, *optional*, defaults to 1):
Minimum number of tokens that cannot be filtered.
"""
def __init__(self, top_p: float, filter_value: float = -float("Inf"), min_tokens_to_keep: int = 1):
if not isinstance(top_p, float) or (top_p < 0 or top_p > 1.0):
raise ValueError(f"`top_p` has to be a float > 0 and < 1, but is {top_p}")
if not isinstance(min_tokens_to_keep, int) or (min_tokens_to_keep < 1):
raise ValueError(f"`min_tokens_to_keep` has to be a positive integer, but is {min_tokens_to_keep}")
self.top_p = top_p
self.filter_value = filter_value
self.min_tokens_to_keep = min_tokens_to_keep
def __call__(self, input_ids: tf.Tensor, scores: tf.Tensor, cur_len: int) -> tf.Tensor:
topk_scores, topk_indices = tf.math.top_k(scores, scores.shape[-1])
mask_scores = tf.fill(scores.shape, self.filter_value)
cumulative_probs = tf.math.cumsum(stable_softmax(topk_scores, axis=-1), axis=-1)
score_mask = cumulative_probs < self.top_p
# Also include the token that is higher than top_p (the first false = shift and insert a True on the left)
score_mask = tf.concat((tf.ones([score_mask.shape[0], 1], dtype=tf.bool), score_mask[:, :-1]), axis=-1)
# Ensure min tokens to keep
score_mask = tf.concat(
(
tf.ones([score_mask.shape[0], self.min_tokens_to_keep], dtype=tf.bool),
score_mask[:, self.min_tokens_to_keep :],
),
axis=-1,
)
# Mask the values that do not fit the criteria
topk_next_scores = tf.where(score_mask, topk_scores, mask_scores)
# Undo the topk sorting: converts the 2D matrix of per-row original indices of shape (batch_size, vocab_size)
# to a 3D tensor of shape (batch_size, vocab_size, 2) containing the original score coordinate, from which we
# can scatter (i.e. `scatter_indices[row, col, :]` is a tensor containing `[row, topk_indices[row, col]]`)
scatter_rows = tf.tile(tf.expand_dims(tf.range(topk_indices.shape[0]), axis=-1), [1, topk_indices.shape[-1]])
scatter_indices = tf.stack((scatter_rows, topk_indices), axis=-1)
next_scores = tf.scatter_nd(scatter_indices, topk_next_scores, shape=topk_next_scores.shape)
return next_scores
class TFMinLengthLogitsProcessor(TFLogitsProcessor):
r"""
[`TFLogitsProcessor`] enforcing a min-length by setting EOS probability to 0.
Args:
min_length (`int`):
The minimum length below which the score of `eos_token_id` is set to `-float("Inf")`.
eos_token_id (`int`):
The id of the *end-of-sequence* token.
"""
def __init__(self, min_length: int, eos_token_id: int):
if not isinstance(min_length, int) or min_length < 0:
raise ValueError(f"`min_length` has to be a positive integer, but is {min_length}")
if not isinstance(eos_token_id, int) or eos_token_id < 0:
raise ValueError(f"`eos_token_id` has to be a positive integer, but is {eos_token_id}")
self.min_length = min_length
self.eos_token_id = eos_token_id
def _apply_eos_token_mask(self, scores: tf.Tensor) -> tf.Tensor:
eos_token_id_mask = tf.range(scores.shape[-1]) == self.eos_token_id
scores = tf.where(eos_token_id_mask, float("-inf"), scores)
return scores
def __call__(self, input_ids: tf.Tensor, scores: tf.Tensor, cur_len: int) -> tf.Tensor:
# applies eos token masking if the first argument is true
scores = tf.cond(
tf.less(cur_len, self.min_length),
lambda: self._apply_eos_token_mask(scores),
lambda: tf.identity(scores),
)
return scores
class TFRepetitionPenaltyLogitsProcessor(TFLogitsProcessor):
r"""
[`TFLogitsProcessor`] enforcing an exponential penalty on repeated sequences.
Args:
repetition_penalty (`float`):
The parameter for repetition penalty. 1.0 means no penalty. See [this
paper](https://arxiv.org/pdf/1909.05858.pdf) for more details.
"""
def __init__(self, penalty: float):
if not isinstance(penalty, float) or not (penalty > 0):
raise ValueError(f"`penalty` has to be a strictly positive float, but is {penalty}")
self.penalty = penalty
def _create_score_penalties(self, input_ids: tf.Tensor, logits: tf.Tensor) -> tf.Tensor:
# We want to populate the penalties in the positions of `input_ids`. Since XLA can't handle shapes unknown
# before runtime, `tf.unique` can't be used. Therefore, we may have redundant updates, when a given row has
# the same token multiple times.
# Gathers the penalties to apply
logit_penalties = tf.gather(logits, input_ids, axis=1, batch_dims=1)
logit_penalties = tf.where(logit_penalties > 0, 1 / self.penalty, logit_penalties)
logit_penalties = tf.where(logit_penalties < 0, self.penalty, logit_penalties)
# Scatters the penalties
token_penalties = tf.ones(logits.shape)
batch_size = input_ids.shape[0]
seq_len = tf.shape(input_ids)[1] # the sequence length has dynamic size, hence the dynamic shape
indexable_prev_input_ids = tf.concat(
(
tf.expand_dims(tf.repeat(tf.range(batch_size), seq_len), axis=-1),
tf.expand_dims(tf.reshape(input_ids, [-1]), axis=-1),
),
axis=1,
)
token_penalties = tf.tensor_scatter_nd_update(
token_penalties, indices=indexable_prev_input_ids, updates=tf.reshape(logit_penalties, [-1])
)
return token_penalties
def __call__(self, input_ids: tf.Tensor, scores: tf.Tensor, cur_len: int) -> tf.Tensor:
score_penalties = self._create_score_penalties(input_ids[:, :cur_len], scores)
scores = tf.math.multiply(scores, score_penalties)
return scores
class TFNoBadWordsLogitsProcessor(TFLogitsProcessor):
"""
[`TFLogitsProcessor`] that enforces that specified sequences will never be sampled.
Args:
bad_words_ids (`List[List[int]]`):
List of 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, make sure to set `add_prefix_space=True` when initializing
the tokenizer, and use `tokenizer(bad_words, add_special_tokens=False).input_ids`. The `add_prefix_space`
argument is only supported for some slow tokenizers, as fast tokenizers' prefixing behaviours come from
`pre tokenizers`. Read more [here](https://huggingface.co/docs/tokenizers/api/pre-tokenizers).
eos_token_id (`int`):
The id of the *end-of-sequence* token.
"""
def __init__(self, bad_words_ids: List[List[int]], eos_token_id: int):
if not isinstance(bad_words_ids, List) or len(bad_words_ids) == 0:
raise ValueError(f"`bad_words_ids` has to be a non-empty list, but is {bad_words_ids}.")
if any(not isinstance(bad_word_ids, list) for bad_word_ids in bad_words_ids):
raise ValueError(f"`bad_words_ids` has to be a list of lists, but is {bad_words_ids}.")
if any(
any((not isinstance(token_id, (int, np.integer)) or token_id < 0) for token_id in bad_word_ids)
for bad_word_ids in bad_words_ids
):
raise ValueError(
f"Each list in `bad_words_ids` has to be a list of positive integers, but is {bad_words_ids}."
)
# stores the information about bad words in three tensors:
# 1. a rectangular tensor with the forbidden sequences (padded with `-1`), for full data comparisons
self.bad_word_seqs_ids = tf.ragged.constant(bad_words_ids).to_tensor(default_value=-1)
# 2. a tensor with the unpadded length of each forbidden sequence, for quick length comparisons
bad_word_seqs_len = [len(bad_words) for bad_words in bad_words_ids]
if any(word_len == 0 for word_len in bad_word_seqs_len):
raise ValueError(f"Banned words token sequences {bad_words_ids} cannot have an empty list")
self.bad_word_seqs_len = tf.convert_to_tensor(bad_word_seqs_len, dtype=tf.int32)
# 3. a tensor containing the last token for each sequence, for easy access to the tokens that may be banned
self.seq_forbidden_tokens = tf.convert_to_tensor([bad_words[-1] for bad_words in bad_words_ids])
def _calc_row_banned_bad_tokens(self, row_input_ids: tf.Tensor) -> tf.Tensor:
def _tokens_match(bad_word_seq_number):
def _len_one():
# If the bad sequence only has one token, always mask it
return tf.cond(
tf.math.equal(self.bad_word_seqs_len[bad_word_seq_number], 1),
lambda: tf.ones((), dtype=tf.bool),
_len_greater_than_cur_len,
)
def _len_greater_than_cur_len():
# Otherwise, if the bad sequence is longer than the current length they can't ever match
return tf.cond(
tf.math.greater(self.bad_word_seqs_len[bad_word_seq_number], tf.shape(row_input_ids)[0]),
lambda: tf.zeros((), dtype=tf.bool),
_match_found,
)
def _match_found():
# Finaly, runs the actual comparison. Can only be called if the previous comparisons do not yield
# an answer (otherwise we get indexing exceptions)
compare_len = self.bad_word_seqs_len[bad_word_seq_number] - 1
return tf.cond(
tf.math.reduce_all(
tf.math.equal(
row_input_ids[-compare_len:], self.bad_word_seqs_ids[bad_word_seq_number, :compare_len]
)
),
lambda: tf.ones((), dtype=tf.bool),
lambda: tf.zeros((), dtype=tf.bool),
)
match = _len_one()
return match
# Compares the current row against all bad word sequences, obtaining a mask with the matches.
match_mask = tf.map_fn(_tokens_match, tf.range(self.bad_word_seqs_ids.shape[0]), fn_output_signature=tf.bool)
row_banned_tokens = self.seq_forbidden_tokens[match_mask]
return row_banned_tokens
def __call__(self, input_ids: tf.Tensor, scores: tf.Tensor, cur_len: int) -> tf.Tensor:
# We want to mask some banned tokens, at a score level. Since the banned tokens depend on the previous
# `input_ids`, they may have a different length for each row, and they may even be empty for some rows.
# To remain simple and XLA-compatible, we work on a per-row fashion.
# TODO (Joao): this function might trigger XLA retracing as `cur_len` increases. Fix it if it becomes
# a frequent choke point. (make `cur_len` a tensor?)
def _get_row_updated_score(row_inputs: Tuple[tf.Tensor]) -> tf.Tensor:
row_input_ids, row_score = row_inputs
banned_tokens = self._calc_row_banned_bad_tokens(row_input_ids[:cur_len])
banned_tokens_mask = tf.scatter_nd(
indices=tf.expand_dims(banned_tokens, axis=-1),
updates=tf.ones_like(banned_tokens, dtype=tf.bool),
shape=row_score.shape,
)
row_score = tf.where(banned_tokens_mask, -float("inf"), row_score)
return row_score
scores = tf.map_fn(_get_row_updated_score, (input_ids, scores), fn_output_signature=tf.float32)
return scores
class TFNoRepeatNGramLogitsProcessor(TFLogitsProcessor):
r"""
[`TFLogitsProcessor`] that enforces no repetition of n-grams. See
[Fairseq](https://github.com/pytorch/fairseq/blob/a07cb6f40480928c9e0548b737aadd36ee66ac76/fairseq/sequence_generator.py#L345).
Args:
ngram_size (`int`):
All ngrams of size `ngram_size` can only occur once.
"""
def __init__(self, ngram_size: int):
if not isinstance(ngram_size, int) or ngram_size <= 0:
raise ValueError(f"`ngram_size` has to be a strictly positive integer, but is {ngram_size}")
self.ngram_size = ngram_size
def calc_banned_ngram_tokens(self, input_ids, num_hypos, cur_len):
# Copied from fairseq for no_repeat_ngram in beam_search
if cur_len + 1 < self.ngram_size:
# return no banned tokens if we haven't generated ngram_size tokens yet
return [[] for _ in range(num_hypos)]
generated_ngrams = [{} for _ in range(num_hypos)]
prev_input_ids = input_ids[:, :cur_len]
for idx in range(num_hypos):
gen_tokens = prev_input_ids[idx].numpy().tolist()
generated_ngram = generated_ngrams[idx]
for ngram in zip(*[gen_tokens[i:] for i in range(self.ngram_size)]):
prev_ngram_tuple = tuple(ngram[:-1])
generated_ngram[prev_ngram_tuple] = generated_ngram.get(prev_ngram_tuple, []) + [ngram[-1]]
def _get_generated_ngrams(hypo_idx):
# Before decoding the next token, prevent decoding of ngrams that have already appeared
start_idx = cur_len + 1 - self.ngram_size
ngram_idx = tuple(prev_input_ids[hypo_idx, start_idx:cur_len].numpy().tolist())
return generated_ngrams[hypo_idx].get(ngram_idx, [])
banned_tokens = [_get_generated_ngrams(hypo_idx) for hypo_idx in range(num_hypos)]
return banned_tokens
def __call__(self, input_ids: tf.Tensor, scores: tf.Tensor, cur_len: int) -> tf.Tensor:
# TODO (joao): enable XLA on this logits processor. See discussion and attempts in
# https://github.com/huggingface/transformers/pull/16974
if not tf.executing_eagerly():
raise NotImplementedError("TFNoRepeatNGramLogitsProcessor is only implemented for eager execution.")
batch_size, vocab_size = scores.shape
banned_tokens = self.calc_banned_ngram_tokens(input_ids, batch_size, cur_len)
# create banned_tokens boolean mask
banned_tokens_indices_mask = []
for banned_tokens_slice in banned_tokens:
banned_tokens_indices_mask.append(
[True if token in banned_tokens_slice else False for token in range(vocab_size)]
)
scores = tf.where(tf.convert_to_tensor(banned_tokens_indices_mask, dtype=tf.bool), -float("inf"), scores)
return scores
class TFForcedBOSTokenLogitsProcessor(TFLogitsProcessor):
r"""
[`TFLogitsProcessor`] that enforces the specified token as the first generated token.
Args:
bos_token_id (`int`):
The id of the token to force as the first generated token.
"""
def __init__(self, bos_token_id: int):
if bos_token_id < 0:
raise ValueError(f"The forced bos token id must be a non-negative integer, got {bos_token_id}")
self.bos_token_id = bos_token_id
def __call__(self, input_ids: tf.Tensor, scores: tf.Tensor, cur_len: int) -> tf.Tensor:
if cur_len == 1:
batch_size, num_tokens = scores.shape
# sets the score to 0 in the bos_token_id column
scores = tf.zeros((batch_size, 1))
# sets the score to -inf everywhere else
if self.bos_token_id > 0:
scores = tf.concat((tf.broadcast_to(-float("inf"), (batch_size, self.bos_token_id)), scores), axis=-1)
if self.bos_token_id < (num_tokens - 1):
scores = tf.concat(
(scores, tf.broadcast_to(-float("inf"), (batch_size, (num_tokens - 1) - self.bos_token_id))),
axis=-1,
)
return scores
class TFForcedEOSTokenLogitsProcessor(TFLogitsProcessor):
r"""
[`TFLogitsProcessor`] that enforces the specified token as the last generated token when `max_length` is reached.
Args:
max_length (`int`):
The maximum length of the sequence to be generated.
eos_token_id (`int`):
The id of the token to force as the last generated token when `max_length` is reached.
"""
def __init__(self, max_length: int, eos_token_id: int):
self.max_length = max_length
if eos_token_id < 0:
raise ValueError(f"The forced eos token id must be a non-negative integer, got {eos_token_id}")
self.eos_token_id = eos_token_id
def __call__(self, input_ids: tf.Tensor, scores: tf.Tensor, cur_len: int) -> tf.Tensor:
if cur_len == self.max_length - 1:
batch_size, num_tokens = scores.shape
# sets the score to 0 in the eos_token_id column
scores = tf.zeros((batch_size, 1))
# sets the score to -inf everywhere else
if self.eos_token_id > 0:
scores = tf.concat((tf.broadcast_to(-float("inf"), (batch_size, self.eos_token_id)), scores), axis=-1)
if self.eos_token_id < (num_tokens - 1):
scores = tf.concat(
(scores, tf.broadcast_to(-float("inf"), (batch_size, (num_tokens - 1) - self.eos_token_id))),
axis=-1,
)
return scores
class TFSuppressTokensAtBeginLogitsProcessor(TFLogitsProcessor):
r"""
[`TFSuppressTokensAtBeginLogitsProcessor`] suppresses a list of tokens as soon as the `generate` function starts
generating using `begin_index` tokens. This should ensure that the tokens defined by `begin_suppress_tokens` at not
sampled at the begining of the generation.
"""
def __init__(self, begin_suppress_tokens, begin_index):
self.begin_suppress_tokens = list(begin_suppress_tokens)
self.begin_index = begin_index
def __call__(self, input_ids: tf.Tensor, scores: tf.Tensor, cur_len: int) -> tf.Tensor:
suppressed_indices = []
for token in self.begin_suppress_tokens:
if token < scores.shape[-1]: # to ensure we don't go beyond the vocab size
suppressed_indices.extend([[i, token] for i in range(scores.shape[0])])
if len(suppressed_indices) > 0:
scores = tf.cond(
tf.equal(cur_len, self.begin_index),
lambda: tf.tensor_scatter_nd_update(
scores,
indices=suppressed_indices,
updates=[-float("inf") for _ in range(scores.shape[0] * len(self.begin_suppress_tokens))],
),
lambda: scores,
)
return scores
class TFSuppressTokensLogitsProcessor(TFLogitsProcessor):
r"""This processor can be used to suppress a list of tokens. The processor will set their log probs to `-inf` so that they
are not sampled."""
def __init__(self, suppress_tokens):
self.suppress_tokens = list(suppress_tokens)
def __call__(self, input_ids: tf.Tensor, scores: tf.Tensor, cur_len: int) -> tf.Tensor:
suppressed_indices = []
for token in self.suppress_tokens:
if token < scores.shape[-1]: # to ensure we don't go beyond the vocab size
suppressed_indices.extend([[i, token] for i in range(scores.shape[0])])
if len(suppressed_indices) > 0:
scores = tf.tensor_scatter_nd_update(
scores,
indices=[[i, token] for i in range(scores.shape[0]) for token in self.suppress_tokens],
updates=[-float("inf") for _ in range(scores.shape[0] * len(self.suppress_tokens))],
)
return scores
class TFForceTokensLogitsProcessor(TFLogitsProcessor):
r"""This processor takes a list of pairs of integers which indicates a mapping from generation indices to token
indices that will be forced before sampling. The processor will set their log probs to `0` and all other tokens to
`-inf` so that they are sampled at their corresponding index."""
def __init__(self, force_token_map: List[List[int]]):
force_token_map = dict(force_token_map)
# Converts the dictionary of format {index: token} containing the tokens to be forced to an array, where the
# index of the array corresponds to the index of the token to be forced, for XLA compatibility.
# Indexes without forced tokens will have an negative value.
force_token_array = np.ones((max(force_token_map.keys()) + 1), dtype=np.int32) * -1
for index, token in force_token_map.items():
if token is not None:
force_token_array[index] = token
self.force_token_array = tf.convert_to_tensor(force_token_array, dtype=tf.int32)
def __call__(self, input_ids: tf.Tensor, scores: tf.Tensor, cur_len: int) -> tf.Tensor:
def _force_token(generation_idx):
batch_size = scores.shape[0]
current_token = self.force_token_array[generation_idx]
new_scores = tf.zeros_like(scores, dtype=scores.dtype) + tf.constant([scores.dtype.min])
indices = tf.stack((tf.range(batch_size), tf.tile([current_token], [batch_size])), axis=1)
updates = tf.zeros((batch_size,), dtype=scores.dtype)
new_scores = tf.tensor_scatter_nd_update(new_scores, indices, updates)
return new_scores
scores = tf.cond(
tf.greater_equal(cur_len, tf.shape(self.force_token_array)[0]),
# If the current length is geq than the length of force_token_array, the processor does nothing.
lambda: tf.identity(scores),
# Otherwise, it may force a certain token.
lambda: tf.cond(
tf.greater_equal(self.force_token_array[cur_len], 0),
# Only valid (positive) tokens are forced
lambda: _force_token(cur_len),
# Otherwise, the processor does nothing.
lambda: scores,
),
)
return scores
|
transformers/src/transformers/generation/tf_logits_process.py/0
|
{
"file_path": "transformers/src/transformers/generation/tf_logits_process.py",
"repo_id": "transformers",
"token_count": 12359
}
| 338
|
# coding=utf-8
# Copyright 2024 NetEase, Inc. and the HuggingFace Inc. team. 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.
from ..utils import is_accelerate_available, is_eetq_available, logging
if is_eetq_available():
import eetq
import torch.nn as nn
if is_accelerate_available():
from accelerate import init_empty_weights
logger = logging.get_logger(__name__)
def _replace_with_eetq_linear(
model,
modules_to_not_convert=None,
current_key_name=None,
quantization_config=None,
has_been_replaced=False,
pre_quantized=False,
):
"""
Private method that wraps the recursion for module replacement.
Returns the converted model and a boolean that indicates if the conversion has been successfull or not.
"""
if current_key_name is None:
current_key_name = []
for name, module in model.named_children():
current_key_name.append(name)
if (isinstance(module, nn.Linear)) and name not in modules_to_not_convert:
# Check if the current key is not in the `modules_to_not_convert`
current_key_name_str = ".".join(current_key_name)
if not any(
(key + "." in current_key_name_str) or (key == current_key_name_str) for key in modules_to_not_convert
):
with init_empty_weights():
in_features = module.in_features
out_features = module.out_features
model._modules[name] = eetq.EetqLinear(
in_features, out_features, module.bias is not None, module.weight.device
)
if pre_quantized:
model._modules[name].register_scale(module.weight.device)
has_been_replaced = True
# Force requires grad to False to avoid unexpected errors
model._modules[name].requires_grad_(False)
if len(list(module.children())) > 0:
_, has_been_replaced = _replace_with_eetq_linear(
module,
modules_to_not_convert,
current_key_name,
quantization_config,
has_been_replaced=has_been_replaced,
pre_quantized=pre_quantized,
)
# Remove the last key for recursion
current_key_name.pop(-1)
return model, has_been_replaced
def replace_with_eetq_linear(
model, modules_to_not_convert=None, current_key_name=None, quantization_config=None, pre_quantized=False
):
"""
A helper function to replace all `torch.nn.Linear` modules by `eetq.EetqLinear` modules from the `eetq`
library. This will enable running your models using high performance int8 weight-only gemm kerner from
FasterTransformer and TensorRT-LLM. Make sure `eetq` compiled with the correct CUDA
version of your hardware is installed before running this function. EETQ shall be installed via the source
'https://github.com/NetEase-FuXi/EETQ'
The function will be run recursively and replace all `torch.nn.Linear` modules except for the `lm_head` that should
be kept as a `torch.nn.Linear` module. The replacement is done under `init_empty_weights` context manager so no
CPU/GPU memory is required to run this function. Each weight will be quantized along the channel.
Parameters:
model (`torch.nn.Module`):
Input model or `torch.nn.Module` as the function is run recursively.
modules_to_not_convert (`List[`str`]`, *optional*, defaults to `["lm_head"]`):
Names of the modules to not convert in `EetqLinear`. In practice we keep the `lm_head` in full precision
for numerical stability reasons.
current_key_name (`List[`str`]`, *optional*):
An array to track the current key of the recursion. This is used to check whether the current key (part of
it) is not in the list of modules to not convert (for instances modules that are offloaded to `cpu` or
`disk`).
"""
modules_to_not_convert = ["lm_head"] if modules_to_not_convert is None else modules_to_not_convert
if quantization_config.modules_to_not_convert is not None:
modules_to_not_convert.extend(quantization_config.modules_to_not_convert)
modules_to_not_convert = list(set(modules_to_not_convert))
model, has_been_replaced = _replace_with_eetq_linear(
model, modules_to_not_convert, current_key_name, quantization_config, pre_quantized=pre_quantized
)
if not has_been_replaced:
logger.warning(
"You are loading your model using eetq but no linear modules were found in your model."
" Please double check your model architecture, or submit an issue on github if you think this is"
" a bug."
)
return model
|
transformers/src/transformers/integrations/eetq.py/0
|
{
"file_path": "transformers/src/transformers/integrations/eetq.py",
"repo_id": "transformers",
"token_count": 2106
}
| 339
|
#include <torch/extension.h>
#include "ATen/ATen.h"
typedef at::BFloat16 bf16;
void cuda_forward(int B, int T, int C, float *w, float *u, float *k, float *v, float *y);
void cuda_forward_bf16(int B, int T, int C, float *w, bf16 *u, bf16 *k, bf16 *v, bf16 *y);
void cuda_forward_with_state(int B, int T, int C, float *w, float *u, float *k, float *v, float *y, float *s);
void cuda_forward_with_state_bf16(int B, int T, int C, float *w, bf16 *u, bf16 *k, bf16 *v, bf16 *y, float *s);
void cuda_backward(int B, int T, int C, float *w, float *u, float *k, float *v, float *y, float *gy, float *gw, float *gu, float *gk, float *gv);
void cuda_backward_bf16(int B, int T, int C, float *w, bf16 *u, bf16 *k, bf16 *v, bf16 *y, bf16 *gy, bf16 *gw, bf16 *gu, bf16 *gk, bf16 *gv);
void forward(torch::Tensor &w, torch::Tensor &u, torch::Tensor &k, torch::Tensor &v, torch::Tensor &y) {
const int B = k.size(0);
const int T = k.size(1);
const int C = k.size(2);
cuda_forward(B, T, C, w.data_ptr<float>(), u.data_ptr<float>(), k.data_ptr<float>(), v.data_ptr<float>(), y.data_ptr<float>());
}
void forward_bf16(torch::Tensor &w, torch::Tensor &u, torch::Tensor &k, torch::Tensor &v, torch::Tensor &y) {
const int B = k.size(0);
const int T = k.size(1);
const int C = k.size(2);
cuda_forward_bf16(B, T, C, w.data_ptr<float>(), u.data_ptr<bf16>(), k.data_ptr<bf16>(), v.data_ptr<bf16>(), y.data_ptr<bf16>());
}
void forward_with_state(torch::Tensor &w, torch::Tensor &u, torch::Tensor &k, torch::Tensor &v, torch::Tensor &y, torch::Tensor &s) {
const int B = k.size(0);
const int T = k.size(1);
const int C = k.size(2);
cuda_forward_with_state(B, T, C, w.data_ptr<float>(), u.data_ptr<float>(), k.data_ptr<float>(), v.data_ptr<float>(), y.data_ptr<float>(), s.data_ptr<float>());
}
void forward_with_state_bf16(torch::Tensor &w, torch::Tensor &u, torch::Tensor &k, torch::Tensor &v, torch::Tensor &y, torch::Tensor &s) {
const int B = k.size(0);
const int T = k.size(1);
const int C = k.size(2);
cuda_forward_with_state_bf16(B, T, C, w.data_ptr<float>(), u.data_ptr<bf16>(), k.data_ptr<bf16>(), v.data_ptr<bf16>(), y.data_ptr<bf16>(), s.data_ptr<float>());
}
void backward(torch::Tensor &w, torch::Tensor &u, torch::Tensor &k, torch::Tensor &v, torch::Tensor &y, torch::Tensor &gy, torch::Tensor &gw, torch::Tensor &gu, torch::Tensor &gk, torch::Tensor &gv) {
const int B = k.size(0);
const int T = k.size(1);
const int C = k.size(2);
cuda_backward(B, T, C, w.data_ptr<float>(), u.data_ptr<float>(), k.data_ptr<float>(), v.data_ptr<float>(), y.data_ptr<float>(), gy.data_ptr<float>(), gw.data_ptr<float>(), gu.data_ptr<float>(), gk.data_ptr<float>(), gv.data_ptr<float>());
}
void backward_bf16(torch::Tensor &w, torch::Tensor &u, torch::Tensor &k, torch::Tensor &v, torch::Tensor &y, torch::Tensor &gy, torch::Tensor &gw, torch::Tensor &gu, torch::Tensor &gk, torch::Tensor &gv) {
const int B = k.size(0);
const int T = k.size(1);
const int C = k.size(2);
cuda_backward_bf16(B, T, C, w.data_ptr<float>(), u.data_ptr<bf16>(), k.data_ptr<bf16>(), v.data_ptr<bf16>(), y.data_ptr<bf16>(),
gy.data_ptr<bf16>(), gw.data_ptr<bf16>(), gu.data_ptr<bf16>(), gk.data_ptr<bf16>(), gv.data_ptr<bf16>());
}
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
m.def("forward", &forward, "wkv forward");
m.def("forward_bf16", &forward_bf16, "wkv forward bf16");
m.def("forward_with_state", &forward_with_state, "wkv forward with state");
m.def("forward_with_state_bf16", &forward_with_state_bf16, "wkv forward with state bf16");
m.def("backward", &backward, "wkv backward");
m.def("backward_bf16", &backward_bf16, "wkv backward bf16");
}
TORCH_LIBRARY(wkv, m) {
m.def("forward", forward);
m.def("forward_bf16", forward_bf16);
m.def("forward_with_state", forward_with_state);
m.def("forward_with_state_bf16", forward_with_state_bf16);
m.def("backward", backward);
m.def("backward_bf16", backward_bf16);
}
|
transformers/src/transformers/kernels/rwkv/wkv_op.cpp/0
|
{
"file_path": "transformers/src/transformers/kernels/rwkv/wkv_op.cpp",
"repo_id": "transformers",
"token_count": 1807
}
| 340
|
# Copyright 2020 The HuggingFace Team. 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.
import warnings
from dataclasses import dataclass
from typing import Optional, Tuple
import torch
from .utils import ModelOutput
@dataclass
class BaseModelOutput(ModelOutput):
"""
Base class for model's outputs, with potential hidden states and attentions.
Args:
last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
Sequence of hidden-states at the output of the last layer of the model.
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
"""
last_hidden_state: torch.FloatTensor = None
hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
@dataclass
class BaseModelOutputWithNoAttention(ModelOutput):
"""
Base class for model's outputs, with potential hidden states.
Args:
last_hidden_state (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
Sequence of hidden-states at the output of the last layer of the model.
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, num_channels, height, width)`.
Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
"""
last_hidden_state: torch.FloatTensor = None
hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
@dataclass
class BaseModelOutputWithPooling(ModelOutput):
"""
Base class for model's outputs that also contains a pooling of the last hidden states.
Args:
last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
Sequence of hidden-states at the output of the last layer of the model.
pooler_output (`torch.FloatTensor` of shape `(batch_size, hidden_size)`):
Last layer hidden-state of the first token of the sequence (classification token) after further processing
through the layers used for the auxiliary pretraining task. E.g. for BERT-family of models, this returns
the classification token after processing through a linear layer and a tanh activation function. The linear
layer weights are trained from the next sentence prediction (classification) objective during pretraining.
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
"""
last_hidden_state: torch.FloatTensor = None
pooler_output: torch.FloatTensor = None
hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
@dataclass
class BaseModelOutputWithPoolingAndNoAttention(ModelOutput):
"""
Base class for model's outputs that also contains a pooling of the last hidden states.
Args:
last_hidden_state (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
Sequence of hidden-states at the output of the last layer of the model.
pooler_output (`torch.FloatTensor` of shape `(batch_size, hidden_size)`):
Last layer hidden-state after a pooling operation on the spatial dimensions.
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, num_channels, height, width)`.
Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
"""
last_hidden_state: torch.FloatTensor = None
pooler_output: torch.FloatTensor = None
hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
@dataclass
class BaseModelOutputWithPast(ModelOutput):
"""
Base class for model's outputs that may also contain a past key/values (to speed up sequential decoding).
Args:
last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
Sequence of hidden-states at the output of the last layer of the model.
If `past_key_values` is used only the last hidden-state of the sequences of shape `(batch_size, 1,
hidden_size)` is output.
past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
`(batch_size, num_heads, sequence_length, embed_size_per_head)`) and optionally if
`config.is_encoder_decoder=True` 2 additional tensors of shape `(batch_size, num_heads,
encoder_sequence_length, embed_size_per_head)`.
Contains pre-computed hidden-states (key and values in the self-attention blocks and optionally if
`config.is_encoder_decoder=True` in the cross-attention blocks) that can be used (see `past_key_values`
input) to speed up sequential decoding.
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
"""
last_hidden_state: torch.FloatTensor = None
past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
@dataclass
class BaseModelOutputWithCrossAttentions(ModelOutput):
"""
Base class for model's outputs, with potential hidden states and attentions.
Args:
last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
Sequence of hidden-states at the output of the last layer of the model.
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
cross_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` and `config.add_cross_attention=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights of the decoder's cross-attention layer, after the attention softmax, used to compute the
weighted average in the cross-attention heads.
"""
last_hidden_state: torch.FloatTensor = None
hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
cross_attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
@dataclass
class BaseModelOutputWithPoolingAndCrossAttentions(ModelOutput):
"""
Base class for model's outputs that also contains a pooling of the last hidden states.
Args:
last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
Sequence of hidden-states at the output of the last layer of the model.
pooler_output (`torch.FloatTensor` of shape `(batch_size, hidden_size)`):
Last layer hidden-state of the first token of the sequence (classification token) after further processing
through the layers used for the auxiliary pretraining task. E.g. for BERT-family of models, this returns
the classification token after processing through a linear layer and a tanh activation function. The linear
layer weights are trained from the next sentence prediction (classification) objective during pretraining.
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
cross_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` and `config.add_cross_attention=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights of the decoder's cross-attention layer, after the attention softmax, used to compute the
weighted average in the cross-attention heads.
past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
`(batch_size, num_heads, sequence_length, embed_size_per_head)`) and optionally if
`config.is_encoder_decoder=True` 2 additional tensors of shape `(batch_size, num_heads,
encoder_sequence_length, embed_size_per_head)`.
Contains pre-computed hidden-states (key and values in the self-attention blocks and optionally if
`config.is_encoder_decoder=True` in the cross-attention blocks) that can be used (see `past_key_values`
input) to speed up sequential decoding.
"""
last_hidden_state: torch.FloatTensor = None
pooler_output: torch.FloatTensor = None
hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
cross_attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
@dataclass
class BaseModelOutputWithPastAndCrossAttentions(ModelOutput):
"""
Base class for model's outputs that may also contain a past key/values (to speed up sequential decoding).
Args:
last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
Sequence of hidden-states at the output of the last layer of the model.
If `past_key_values` is used only the last hidden-state of the sequences of shape `(batch_size, 1,
hidden_size)` is output.
past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
`(batch_size, num_heads, sequence_length, embed_size_per_head)`) and optionally if
`config.is_encoder_decoder=True` 2 additional tensors of shape `(batch_size, num_heads,
encoder_sequence_length, embed_size_per_head)`.
Contains pre-computed hidden-states (key and values in the self-attention blocks and optionally if
`config.is_encoder_decoder=True` in the cross-attention blocks) that can be used (see `past_key_values`
input) to speed up sequential decoding.
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
cross_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` and `config.add_cross_attention=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights of the decoder's cross-attention layer, after the attention softmax, used to compute the
weighted average in the cross-attention heads.
"""
last_hidden_state: torch.FloatTensor = None
past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
cross_attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
@dataclass
class MoECausalLMOutputWithPast(ModelOutput):
"""
Base class for causal language model (or autoregressive) outputs as well as Mixture of Expert's router hidden
states terms, to train a MoE model.
Args:
loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
Language modeling loss (for next-token prediction).
logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
`(batch_size, num_heads, sequence_length, embed_size_per_head)`)
Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
`past_key_values` input) to speed up sequential decoding.
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
z_loss (`torch.FloatTensor`, *optional*, returned when `labels` is provided):
z_loss for the sparse modules.
aux_loss (`torch.FloatTensor`, *optional*, returned when `labels` is provided):
aux_loss for the sparse modules.
router_logits (`tuple(torch.FloatTensor)`, *optional*, returned when `output_router_logits=True` is passed or when `config.add_router_probs=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, sequence_length, num_experts)`.
Router logits of the encoder model, useful to compute the auxiliary loss and the z_loss for the sparse
modules.
"""
loss: Optional[torch.FloatTensor] = None
logits: torch.FloatTensor = None
past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
z_loss: torch.FloatTensor = None
aux_loss: torch.FloatTensor = None
router_logits: Optional[Tuple[torch.FloatTensor]] = None
@dataclass
class MoEModelOutput(ModelOutput):
"""
Base class for model's outputs, with potential hidden states and attentions.
Args:
last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
Sequence of hidden-states at the output of the last layer of the model.
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
router_probs (`tuple(torch.FloatTensor)`, *optional*, returned when `output_router_probs=True` and `config.add_router_probs=True` is passed or when `config.output_router_probs=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, sequence_length, num_experts)`.
Raw router probabilities that are computed by MoE routers, these terms are used to compute the auxiliary
loss and the z_loss for Mixture of Experts models.
"""
last_hidden_state: torch.FloatTensor = None
hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
router_probs: Optional[Tuple[torch.FloatTensor]] = None
@dataclass
class MoeModelOutputWithPast(ModelOutput):
"""
Base class for model's outputs, with potential hidden states and attentions.
Args:
last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
Sequence of hidden-states at the output of the last layer of the model.
past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
`(batch_size, num_heads, sequence_length, embed_size_per_head)`) and optionally if
`config.is_encoder_decoder=True` 2 additional tensors of shape `(batch_size, num_heads,
encoder_sequence_length, embed_size_per_head)`.
Contains pre-computed hidden-states (key and values in the self-attention blocks and optionally if
`config.is_encoder_decoder=True` in the cross-attention blocks) that can be used (see `past_key_values`
input) to speed up sequential decoding.
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
router_logits (`tuple(torch.FloatTensor)`, *optional*, returned when `output_router_probs=True` and `config.add_router_probs=True` is passed or when `config.output_router_probs=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, sequence_length, num_experts)`.
Raw router logtis (post-softmax) that are computed by MoE routers, these terms are used to compute the auxiliary
loss for Mixture of Experts models.
"""
last_hidden_state: torch.FloatTensor = None
past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
router_logits: Optional[Tuple[torch.FloatTensor]] = None
@dataclass
class MoeCausalLMOutputWithPast(ModelOutput):
"""
Base class for causal language model (or autoregressive) with mixture of experts outputs.
Args:
loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
Language modeling loss (for next-token prediction).
logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
aux_loss (`torch.FloatTensor`, *optional*, returned when `labels` is provided):
aux_loss for the sparse modules.
router_logits (`tuple(torch.FloatTensor)`, *optional*, returned when `output_router_probs=True` and `config.add_router_probs=True` is passed or when `config.output_router_probs=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, sequence_length, num_experts)`.
Raw router logtis (post-softmax) that are computed by MoE routers, these terms are used to compute the auxiliary
loss for Mixture of Experts models.
past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
`(batch_size, num_heads, sequence_length, embed_size_per_head)`)
Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
`past_key_values` input) to speed up sequential decoding.
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
"""
loss: Optional[torch.FloatTensor] = None
aux_loss: Optional[torch.FloatTensor] = None
logits: torch.FloatTensor = None
past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
router_logits: Optional[Tuple[torch.FloatTensor]] = None
@dataclass
class MoEModelOutputWithPastAndCrossAttentions(ModelOutput):
"""
Base class for model's outputs that may also contain a past key/values (to speed up sequential decoding) as well as
Mixture of Expert's router hidden states terms, to train a MoE model.
Args:
last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
Sequence of hidden-states at the output of the last layer of the model.
If `past_key_values` is used only the last hidden-state of the sequences of shape `(batch_size, 1,
hidden_size)` is output.
past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
`(batch_size, num_heads, sequence_length, embed_size_per_head)`) and optionally if
`config.is_encoder_decoder=True` 2 additional tensors of shape `(batch_size, num_heads,
encoder_sequence_length, embed_size_per_head)`.
Contains pre-computed hidden-states (key and values in the self-attention blocks and optionally if
`config.is_encoder_decoder=True` in the cross-attention blocks) that can be used (see `past_key_values`
input) to speed up sequential decoding.
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
cross_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` and `config.add_cross_attention=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights of the decoder's cross-attention layer, after the attention softmax, used to compute the
weighted average in the cross-attention heads.
router_probs (`tuple(torch.FloatTensor)`, *optional*, returned when `output_router_probs=True` and `config.add_router_probs=True` is passed or when `config.output_router_probs=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, sequence_length, num_experts)`.
Raw router probabilities that are computed by MoE routers, these terms are used to compute the auxiliary
loss and the z_loss for Mixture of Experts models.
"""
last_hidden_state: torch.FloatTensor = None
past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
cross_attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
router_probs: Optional[Tuple[torch.FloatTensor]] = None
@dataclass
class Seq2SeqModelOutput(ModelOutput):
"""
Base class for model encoder's outputs that also contains : pre-computed hidden states that can speed up sequential
decoding.
Args:
last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
Sequence of hidden-states at the output of the last layer of the decoder of the model.
If `past_key_values` is used only the last hidden-state of the sequences of shape `(batch_size, 1,
hidden_size)` is output.
past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
`(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
`(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
decoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the decoder at the output of each layer plus the optional initial embedding outputs.
decoder_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights of the decoder, after the attention softmax, used to compute the weighted average in the
self-attention heads.
cross_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights of the decoder's cross-attention layer, after the attention softmax, used to compute the
weighted average in the cross-attention heads.
encoder_last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
Sequence of hidden-states at the output of the last layer of the encoder of the model.
encoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the encoder at the output of each layer plus the optional initial embedding outputs.
encoder_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights of the encoder, after the attention softmax, used to compute the weighted average in the
self-attention heads.
"""
last_hidden_state: torch.FloatTensor = None
past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
decoder_hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
decoder_attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
cross_attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
encoder_last_hidden_state: Optional[torch.FloatTensor] = None
encoder_hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
encoder_attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
@dataclass
class Seq2SeqMoEModelOutput(ModelOutput):
"""
Base class for model encoder's outputs that also contains : pre-computed hidden states that can speed up sequential
decoding.
Args:
last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
Sequence of hidden-states at the output of the last layer of the decoder of the model.
If `past_key_values` is used only the last hidden-state of the sequences of shape `(batch_size, 1,
hidden_size)` is output.
past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
`(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
`(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
decoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the decoder at the output of each layer plus the optional initial embedding outputs.
decoder_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights of the decoder, after the attention softmax, used to compute the weighted average in the
self-attention heads.
decoder_router_logits (`tuple(torch.FloatTensor)`, *optional*, returned when `output_router_logits=True` is passed or when `config.add_router_probs=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, sequence_length, num_experts)`.
Router logits of the decoder model, useful to compute the auxiliary loss for Mixture of Experts models.
cross_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights of the decoder's cross-attention layer, after the attention softmax, used to compute the
weighted average in the cross-attention heads.
encoder_last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
Sequence of hidden-states at the output of the last layer of the encoder of the model.
encoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the encoder at the output of each layer plus the optional initial embedding outputs.
encoder_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights of the encoder, after the attention softmax, used to compute the weighted average in the
self-attention heads.
encoder_router_logits (`tuple(torch.FloatTensor)`, *optional*, returned when `output_router_logits=True` is passed or when `config.add_router_probs=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, sequence_length, num_experts)`.
Router logits of the encoder model, useful to compute the auxiliary loss and the z_loss for the sparse
modules.
"""
last_hidden_state: torch.FloatTensor = None
past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
decoder_hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
decoder_attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
decoder_router_logits: Optional[Tuple[torch.FloatTensor]] = None
cross_attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
encoder_last_hidden_state: Optional[torch.FloatTensor] = None
encoder_hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
encoder_attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
encoder_router_logits: Optional[Tuple[torch.FloatTensor]] = None
@dataclass
class CausalLMOutput(ModelOutput):
"""
Base class for causal language model (or autoregressive) outputs.
Args:
loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
Language modeling loss (for next-token prediction).
logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
"""
loss: Optional[torch.FloatTensor] = None
logits: torch.FloatTensor = None
hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
@dataclass
class CausalLMOutputWithPast(ModelOutput):
"""
Base class for causal language model (or autoregressive) outputs.
Args:
loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
Language modeling loss (for next-token prediction).
logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
`(batch_size, num_heads, sequence_length, embed_size_per_head)`)
Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
`past_key_values` input) to speed up sequential decoding.
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
"""
loss: Optional[torch.FloatTensor] = None
logits: torch.FloatTensor = None
past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
@dataclass
class CausalLMOutputWithCrossAttentions(ModelOutput):
"""
Base class for causal language model (or autoregressive) outputs.
Args:
loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
Language modeling loss (for next-token prediction).
logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
cross_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Cross attentions weights after the attention softmax, used to compute the weighted average in the
cross-attention heads.
past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
Tuple of `torch.FloatTensor` tuples of length `config.n_layers`, with each tuple containing the cached key,
value states of the self-attention and the cross-attention layers if model is used in encoder-decoder
setting. Only relevant if `config.is_decoder = True`.
Contains pre-computed hidden-states (key and values in the attention blocks) that can be used (see
`past_key_values` input) to speed up sequential decoding.
"""
loss: Optional[torch.FloatTensor] = None
logits: torch.FloatTensor = None
past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
cross_attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
@dataclass
class SequenceClassifierOutputWithPast(ModelOutput):
"""
Base class for outputs of sentence classification models.
Args:
loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
Classification (or regression if config.num_labels==1) loss.
logits (`torch.FloatTensor` of shape `(batch_size, config.num_labels)`):
Classification (or regression if config.num_labels==1) scores (before SoftMax).
past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
`(batch_size, num_heads, sequence_length, embed_size_per_head)`)
Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
`past_key_values` input) to speed up sequential decoding.
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
"""
loss: Optional[torch.FloatTensor] = None
logits: torch.FloatTensor = None
past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
@dataclass
class MaskedLMOutput(ModelOutput):
"""
Base class for masked language models outputs.
Args:
loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
Masked language modeling (MLM) loss.
logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
"""
loss: Optional[torch.FloatTensor] = None
logits: torch.FloatTensor = None
hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
@dataclass
class Seq2SeqLMOutput(ModelOutput):
"""
Base class for sequence-to-sequence language models outputs.
Args:
loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
Language modeling loss.
logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
`(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
`(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
decoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the decoder at the output of each layer plus the initial embedding outputs.
decoder_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights of the decoder, after the attention softmax, used to compute the weighted average in the
self-attention heads.
cross_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights of the decoder's cross-attention layer, after the attention softmax, used to compute the
weighted average in the cross-attention heads.
encoder_last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
Sequence of hidden-states at the output of the last layer of the encoder of the model.
encoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the encoder at the output of each layer plus the initial embedding outputs.
encoder_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights of the encoder, after the attention softmax, used to compute the weighted average in the
self-attention heads.
"""
loss: Optional[torch.FloatTensor] = None
logits: torch.FloatTensor = None
past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
decoder_hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
decoder_attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
cross_attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
encoder_last_hidden_state: Optional[torch.FloatTensor] = None
encoder_hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
encoder_attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
@dataclass
class Seq2SeqMoEOutput(ModelOutput):
"""
Base class for sequence-to-sequence language models outputs.
Args:
loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
Language modeling loss.
logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
`(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
`(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
decoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the decoder at the output of each layer plus the initial embedding outputs.
decoder_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights of the decoder, after the attention softmax, used to compute the weighted average in the
self-attention heads.
decoder_router_logits (`tuple(torch.FloatTensor)`, *optional*, returned when `output_router_logits=True` is passed or when `config.add_router_probs=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, sequence_length, num_experts)`.
Router logits of the decoder model, useful to compute the auxiliary loss for Mixture of Experts models.
cross_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights of the decoder's cross-attention layer, after the attention softmax, used to compute the
weighted average in the cross-attention heads.
encoder_last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
Sequence of hidden-states at the output of the last layer of the encoder of the model.
encoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the encoder at the output of each layer plus the initial embedding outputs.
encoder_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights of the encoder, after the attention softmax, used to compute the weighted average in the
self-attention heads.
encoder_router_logits (`tuple(torch.FloatTensor)`, *optional*, returned when `output_router_logits=True` is passed or when `config.add_router_probs=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, sequence_length, num_experts)`.
Router logits of the encoder model, useful to compute the auxiliary loss and z_loss for Mixture of Experts
models.
"""
loss: Optional[torch.FloatTensor] = None
logits: torch.FloatTensor = None
encoder_z_loss: torch.FloatTensor = None
decoder_z_loss: torch.FloatTensor = None
encoder_aux_loss: torch.FloatTensor = None
decoder_aux_loss: torch.FloatTensor = None
past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
decoder_hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
decoder_attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
decoder_router_logits: Optional[Tuple[torch.FloatTensor]] = None
cross_attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
encoder_last_hidden_state: Optional[torch.FloatTensor] = None
encoder_hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
encoder_attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
encoder_router_logits: Optional[Tuple[torch.FloatTensor]] = None
@dataclass
class NextSentencePredictorOutput(ModelOutput):
"""
Base class for outputs of models predicting if two sentences are consecutive or not.
Args:
loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `next_sentence_label` is provided):
Next sequence prediction (classification) loss.
logits (`torch.FloatTensor` of shape `(batch_size, 2)`):
Prediction scores of the next sequence prediction (classification) head (scores of True/False continuation
before SoftMax).
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
"""
loss: Optional[torch.FloatTensor] = None
logits: torch.FloatTensor = None
hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
@dataclass
class SequenceClassifierOutput(ModelOutput):
"""
Base class for outputs of sentence classification models.
Args:
loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
Classification (or regression if config.num_labels==1) loss.
logits (`torch.FloatTensor` of shape `(batch_size, config.num_labels)`):
Classification (or regression if config.num_labels==1) scores (before SoftMax).
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
"""
loss: Optional[torch.FloatTensor] = None
logits: torch.FloatTensor = None
hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
@dataclass
class Seq2SeqSequenceClassifierOutput(ModelOutput):
"""
Base class for outputs of sequence-to-sequence sentence classification models.
Args:
loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `label` is provided):
Classification (or regression if config.num_labels==1) loss.
logits (`torch.FloatTensor` of shape `(batch_size, config.num_labels)`):
Classification (or regression if config.num_labels==1) scores (before SoftMax).
past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
`(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
`(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
decoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the decoder at the output of each layer plus the initial embedding outputs.
decoder_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights of the decoder, after the attention softmax, used to compute the weighted average in the
self-attention heads.
cross_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights of the decoder's cross-attention layer, after the attention softmax, used to compute the
weighted average in the cross-attention heads.
encoder_last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
Sequence of hidden-states at the output of the last layer of the encoder of the model.
encoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the encoder at the output of each layer plus the initial embedding outputs.
encoder_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights of the encoder, after the attention softmax, used to compute the weighted average in the
self-attention heads.
"""
loss: Optional[torch.FloatTensor] = None
logits: torch.FloatTensor = None
past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
decoder_hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
decoder_attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
cross_attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
encoder_last_hidden_state: Optional[torch.FloatTensor] = None
encoder_hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
encoder_attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
@dataclass
class MultipleChoiceModelOutput(ModelOutput):
"""
Base class for outputs of multiple choice models.
Args:
loss (`torch.FloatTensor` of shape *(1,)*, *optional*, returned when `labels` is provided):
Classification loss.
logits (`torch.FloatTensor` of shape `(batch_size, num_choices)`):
*num_choices* is the second dimension of the input tensors. (see *input_ids* above).
Classification scores (before SoftMax).
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
"""
loss: Optional[torch.FloatTensor] = None
logits: torch.FloatTensor = None
hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
@dataclass
class TokenClassifierOutput(ModelOutput):
"""
Base class for outputs of token classification models.
Args:
loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided) :
Classification loss.
logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.num_labels)`):
Classification scores (before SoftMax).
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
"""
loss: Optional[torch.FloatTensor] = None
logits: torch.FloatTensor = None
hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
@dataclass
class QuestionAnsweringModelOutput(ModelOutput):
"""
Base class for outputs of question answering models.
Args:
loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
Total span extraction loss is the sum of a Cross-Entropy for the start and end positions.
start_logits (`torch.FloatTensor` of shape `(batch_size, sequence_length)`):
Span-start scores (before SoftMax).
end_logits (`torch.FloatTensor` of shape `(batch_size, sequence_length)`):
Span-end scores (before SoftMax).
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
"""
loss: Optional[torch.FloatTensor] = None
start_logits: torch.FloatTensor = None
end_logits: torch.FloatTensor = None
hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
@dataclass
class Seq2SeqQuestionAnsweringModelOutput(ModelOutput):
"""
Base class for outputs of sequence-to-sequence question answering models.
Args:
loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
Total span extraction loss is the sum of a Cross-Entropy for the start and end positions.
start_logits (`torch.FloatTensor` of shape `(batch_size, sequence_length)`):
Span-start scores (before SoftMax).
end_logits (`torch.FloatTensor` of shape `(batch_size, sequence_length)`):
Span-end scores (before SoftMax).
past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
`(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
`(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
decoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the decoder at the output of each layer plus the initial embedding outputs.
decoder_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights of the decoder, after the attention softmax, used to compute the weighted average in the
self-attention heads.
cross_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights of the decoder's cross-attention layer, after the attention softmax, used to compute the
weighted average in the cross-attention heads.
encoder_last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
Sequence of hidden-states at the output of the last layer of the encoder of the model.
encoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the encoder at the output of each layer plus the initial embedding outputs.
encoder_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights of the encoder, after the attention softmax, used to compute the weighted average in the
self-attention heads.
"""
loss: Optional[torch.FloatTensor] = None
start_logits: torch.FloatTensor = None
end_logits: torch.FloatTensor = None
past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
decoder_hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
decoder_attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
cross_attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
encoder_last_hidden_state: Optional[torch.FloatTensor] = None
encoder_hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
encoder_attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
@dataclass
class SemanticSegmenterOutput(ModelOutput):
"""
Base class for outputs of semantic segmentation models.
Args:
loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
Classification (or regression if config.num_labels==1) loss.
logits (`torch.FloatTensor` of shape `(batch_size, config.num_labels, logits_height, logits_width)`):
Classification scores for each pixel.
<Tip warning={true}>
The logits returned do not necessarily have the same size as the `pixel_values` passed as inputs. This is
to avoid doing two interpolations and lose some quality when a user needs to resize the logits to the
original image size as post-processing. You should always check your logits shape and resize as needed.
</Tip>
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, patch_size, hidden_size)`.
Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, patch_size,
sequence_length)`.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
"""
loss: Optional[torch.FloatTensor] = None
logits: torch.FloatTensor = None
hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
@dataclass
class ImageClassifierOutput(ModelOutput):
"""
Base class for outputs of image classification models.
Args:
loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
Classification (or regression if config.num_labels==1) loss.
logits (`torch.FloatTensor` of shape `(batch_size, config.num_labels)`):
Classification (or regression if config.num_labels==1) scores (before SoftMax).
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each stage) of shape `(batch_size, sequence_length, hidden_size)`. Hidden-states
(also called feature maps) of the model at the output of each stage.
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, patch_size,
sequence_length)`.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
"""
loss: Optional[torch.FloatTensor] = None
logits: torch.FloatTensor = None
hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
@dataclass
class ImageClassifierOutputWithNoAttention(ModelOutput):
"""
Base class for outputs of image classification models.
Args:
loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
Classification (or regression if config.num_labels==1) loss.
logits (`torch.FloatTensor` of shape `(batch_size, config.num_labels)`):
Classification (or regression if config.num_labels==1) scores (before SoftMax).
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each stage) of shape `(batch_size, num_channels, height, width)`. Hidden-states (also
called feature maps) of the model at the output of each stage.
"""
loss: Optional[torch.FloatTensor] = None
logits: torch.FloatTensor = None
hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
@dataclass
class DepthEstimatorOutput(ModelOutput):
"""
Base class for outputs of depth estimation models.
Args:
loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
Classification (or regression if config.num_labels==1) loss.
predicted_depth (`torch.FloatTensor` of shape `(batch_size, height, width)`):
Predicted depth for each pixel.
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, num_channels, height, width)`.
Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, patch_size,
sequence_length)`.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
"""
loss: Optional[torch.FloatTensor] = None
predicted_depth: torch.FloatTensor = None
hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
@dataclass
class ImageSuperResolutionOutput(ModelOutput):
"""
Base class for outputs of image super resolution models.
Args:
loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
Reconstruction loss.
reconstruction (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
Reconstructed images, possibly upscaled.
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each stage) of shape `(batch_size, sequence_length, hidden_size)`. Hidden-states
(also called feature maps) of the model at the output of each stage.
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, patch_size,
sequence_length)`.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
"""
loss: Optional[torch.FloatTensor] = None
reconstruction: torch.FloatTensor = None
hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
@dataclass
class Wav2Vec2BaseModelOutput(ModelOutput):
"""
Base class for models that have been trained with the Wav2Vec2 loss objective.
Args:
last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
Sequence of hidden-states at the output of the last layer of the model.
extract_features (`torch.FloatTensor` of shape `(batch_size, sequence_length, conv_dim[-1])`):
Sequence of extracted feature vectors of the last convolutional layer of the model.
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer) of
shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the model at the output of each layer plus the initial embedding outputs.
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
"""
last_hidden_state: torch.FloatTensor = None
extract_features: torch.FloatTensor = None
hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
@dataclass
class XVectorOutput(ModelOutput):
"""
Output type of [`Wav2Vec2ForXVector`].
Args:
loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
Classification loss.
logits (`torch.FloatTensor` of shape `(batch_size, config.xvector_output_dim)`):
Classification hidden states before AMSoftmax.
embeddings (`torch.FloatTensor` of shape `(batch_size, config.xvector_output_dim)`):
Utterance embeddings used for vector similarity-based retrieval.
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer) of
shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the model at the output of each layer plus the initial embedding outputs.
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
"""
loss: Optional[torch.FloatTensor] = None
logits: torch.FloatTensor = None
embeddings: torch.FloatTensor = None
hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
@dataclass
class BackboneOutput(ModelOutput):
"""
Base class for outputs of backbones.
Args:
feature_maps (`tuple(torch.FloatTensor)` of shape `(batch_size, num_channels, height, width)`):
Feature maps of the stages.
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer) of
shape `(batch_size, sequence_length, hidden_size)` or `(batch_size, num_channels, height, width)`,
depending on the backbone.
Hidden-states of the model at the output of each stage plus the initial embedding outputs.
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`. Only applicable if the backbone uses attention.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
"""
feature_maps: Tuple[torch.FloatTensor] = None
hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
@dataclass
class BaseModelOutputWithPoolingAndProjection(ModelOutput):
"""
Base class for model's outputs that also contains a pooling of the last hidden states.
Args:
last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
Sequence of hidden-states at the output of the last layer of the model.
pooler_output (`torch.FloatTensor` of shape `(batch_size, hidden_size)`):
Last layer hidden-state of the first token of the sequence (classification token) after further processing
through the layers used for the auxiliary pretraining task. E.g. for BERT-family of models, this returns
the classification token after processing through a linear layer and a tanh activation function. The linear
layer weights are trained from the next sentence prediction (classification) objective during pretraining.
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
projection_state (`tuple(torch.FloatTensor)`, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` of shape `(batch_size,config.project_dim)`.
Text embeddings before the projection layer, used to mimic the last hidden state of the teacher encoder.
"""
last_hidden_state: torch.FloatTensor = None
pooler_output: torch.FloatTensor = None
hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
projection_state: Optional[Tuple[torch.FloatTensor]] = None
@dataclass
class Seq2SeqSpectrogramOutput(ModelOutput):
"""
Base class for sequence-to-sequence spectrogram outputs.
Args:
loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
Spectrogram generation loss.
spectrogram (`torch.FloatTensor` of shape `(batch_size, sequence_length, num_bins)`):
The predicted spectrogram.
past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
`(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
`(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
decoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the decoder at the output of each layer plus the initial embedding outputs.
decoder_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights of the decoder, after the attention softmax, used to compute the weighted average in the
self-attention heads.
cross_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights of the decoder's cross-attention layer, after the attention softmax, used to compute the
weighted average in the cross-attention heads.
encoder_last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
Sequence of hidden-states at the output of the last layer of the encoder of the model.
encoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the encoder at the output of each layer plus the initial embedding outputs.
encoder_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights of the encoder, after the attention softmax, used to compute the weighted average in the
self-attention heads.
"""
loss: Optional[torch.FloatTensor] = None
spectrogram: torch.FloatTensor = None
past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
decoder_hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
decoder_attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
cross_attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
encoder_last_hidden_state: Optional[torch.FloatTensor] = None
encoder_hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
encoder_attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
@dataclass
class Seq2SeqTSModelOutput(ModelOutput):
"""
Base class for time series model's encoder outputs that also contains pre-computed hidden states that can speed up
sequential decoding.
Args:
last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
Sequence of hidden-states at the output of the last layer of the decoder of the model.
If `past_key_values` is used only the last hidden-state of the sequences of shape `(batch_size, 1,
hidden_size)` is output.
past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
`(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
`(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
decoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the decoder at the output of each layer plus the optional initial embedding outputs.
decoder_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights of the decoder, after the attention softmax, used to compute the weighted average in the
self-attention heads.
cross_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights of the decoder's cross-attention layer, after the attention softmax, used to compute the
weighted average in the cross-attention heads.
encoder_last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
Sequence of hidden-states at the output of the last layer of the encoder of the model.
encoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the encoder at the output of each layer plus the optional initial embedding outputs.
encoder_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights of the encoder, after the attention softmax, used to compute the weighted average in the
self-attention heads.
loc (`torch.FloatTensor` of shape `(batch_size,)` or `(batch_size, input_size)`, *optional*):
Shift values of each time series' context window which is used to give the model inputs of the same
magnitude and then used to shift back to the original magnitude.
scale (`torch.FloatTensor` of shape `(batch_size,)` or `(batch_size, input_size)`, *optional*):
Scaling values of each time series' context window which is used to give the model inputs of the same
magnitude and then used to rescale back to the original magnitude.
static_features (`torch.FloatTensor` of shape `(batch_size, feature size)`, *optional*):
Static features of each time series' in a batch which are copied to the covariates at inference time.
"""
last_hidden_state: torch.FloatTensor = None
past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
decoder_hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
decoder_attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
cross_attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
encoder_last_hidden_state: Optional[torch.FloatTensor] = None
encoder_hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
encoder_attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
loc: Optional[torch.FloatTensor] = None
scale: Optional[torch.FloatTensor] = None
static_features: Optional[torch.FloatTensor] = None
@dataclass
class Seq2SeqTSPredictionOutput(ModelOutput):
"""
Base class for time series model's decoder outputs that also contain the loss as well as the parameters of the
chosen distribution.
Args:
loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when a `future_values` is provided):
Distributional loss.
params (`torch.FloatTensor` of shape `(batch_size, num_samples, num_params)`):
Parameters of the chosen distribution.
past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
`(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
`(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
decoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the decoder at the output of each layer plus the initial embedding outputs.
decoder_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights of the decoder, after the attention softmax, used to compute the weighted average in the
self-attention heads.
cross_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights of the decoder's cross-attention layer, after the attention softmax, used to compute the
weighted average in the cross-attention heads.
encoder_last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
Sequence of hidden-states at the output of the last layer of the encoder of the model.
encoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the encoder at the output of each layer plus the initial embedding outputs.
encoder_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights of the encoder, after the attention softmax, used to compute the weighted average in the
self-attention heads.
loc (`torch.FloatTensor` of shape `(batch_size,)` or `(batch_size, input_size)`, *optional*):
Shift values of each time series' context window which is used to give the model inputs of the same
magnitude and then used to shift back to the original magnitude.
scale (`torch.FloatTensor` of shape `(batch_size,)` or `(batch_size, input_size)`, *optional*):
Scaling values of each time series' context window which is used to give the model inputs of the same
magnitude and then used to rescale back to the original magnitude.
static_features (`torch.FloatTensor` of shape `(batch_size, feature size)`, *optional*):
Static features of each time series' in a batch which are copied to the covariates at inference time.
"""
loss: Optional[torch.FloatTensor] = None
params: Optional[Tuple[torch.FloatTensor]] = None
past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
decoder_hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
decoder_attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
cross_attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
encoder_last_hidden_state: Optional[torch.FloatTensor] = None
encoder_hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
encoder_attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
loc: Optional[torch.FloatTensor] = None
scale: Optional[torch.FloatTensor] = None
static_features: Optional[torch.FloatTensor] = None
@dataclass
class SampleTSPredictionOutput(ModelOutput):
"""
Base class for time series model's predictions outputs that contains the sampled values from the chosen
distribution.
Args:
sequences (`torch.FloatTensor` of shape `(batch_size, num_samples, prediction_length)` or `(batch_size, num_samples, prediction_length, input_size)`):
Sampled values from the chosen distribution.
"""
sequences: torch.FloatTensor = None
@dataclass
class MaskedImageModelingOutput(ModelOutput):
"""
Base class for outputs of masked image completion / in-painting models.
Args:
loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `bool_masked_pos` is provided):
Reconstruction loss.
reconstruction (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
Reconstructed / completed images.
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or
when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each stage) of shape `(batch_size, sequence_length, hidden_size)`. Hidden-states
(also called feature maps) of the model at the output of each stage.
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when
`config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, patch_size,
sequence_length)`. Attentions weights after the attention softmax, used to compute the weighted average in
the self-attention heads.
"""
loss: Optional[torch.FloatTensor] = None
reconstruction: torch.FloatTensor = None
hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
@property
def logits(self):
warnings.warn(
"logits attribute is deprecated and will be removed in version 5 of Transformers."
" Please use the reconstruction attribute to retrieve the final output instead.",
FutureWarning,
)
return self.reconstruction
|
transformers/src/transformers/modeling_outputs.py/0
|
{
"file_path": "transformers/src/transformers/modeling_outputs.py",
"repo_id": "transformers",
"token_count": 40604
}
| 341
|
# coding=utf-8
# Copyright 2023 The HuggingFace Inc. team. 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.
"""ALIGN model configuration"""
import os
from typing import TYPE_CHECKING, List, Union
if TYPE_CHECKING:
pass
from ...configuration_utils import PretrainedConfig
from ...utils import logging
logger = logging.get_logger(__name__)
class AlignTextConfig(PretrainedConfig):
r"""
This is the configuration class to store the configuration of a [`AlignTextModel`]. It is used to instantiate a
ALIGN text encoder according to the specified arguments, defining the model architecture. Instantiating a
configuration with the defaults will yield a similar configuration to that of the text encoder of the ALIGN
[kakaobrain/align-base](https://huggingface.co/kakaobrain/align-base) architecture. The default values here are
copied from BERT.
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PretrainedConfig`] for more information.
Args:
vocab_size (`int`, *optional*, defaults to 30522):
Vocabulary size of the Align Text model. Defines the number of different tokens that can be represented by
the `inputs_ids` passed when calling [`AlignTextModel`].
hidden_size (`int`, *optional*, defaults to 768):
Dimensionality of the encoder layers and the pooler layer.
num_hidden_layers (`int`, *optional*, defaults to 12):
Number of hidden layers in the Transformer encoder.
num_attention_heads (`int`, *optional*, defaults to 12):
Number of attention heads for each attention layer in the Transformer encoder.
intermediate_size (`int`, *optional*, defaults to 3072):
Dimensionality of the "intermediate" (often named feed-forward) layer in the Transformer encoder.
hidden_act (`str` or `Callable`, *optional*, defaults to `"gelu"`):
The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
`"relu"`, `"silu"` and `"gelu_new"` are supported.
hidden_dropout_prob (`float`, *optional*, defaults to 0.1):
The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
attention_probs_dropout_prob (`float`, *optional*, defaults to 0.1):
The dropout ratio for the attention probabilities.
max_position_embeddings (`int`, *optional*, defaults to 512):
The maximum sequence length that this model might ever be used with. Typically set this to something large
just in case (e.g., 512 or 1024 or 2048).
type_vocab_size (`int`, *optional*, defaults to 2):
The vocabulary size of the `token_type_ids` passed when calling [`AlignTextModel`].
initializer_range (`float`, *optional*, defaults to 0.02):
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
layer_norm_eps (`float`, *optional*, defaults to 1e-12):
The epsilon used by the layer normalization layers.
pad_token_id (`int`, *optional*, defaults to 0):
Padding token id.
position_embedding_type (`str`, *optional*, defaults to `"absolute"`):
Type of position embedding. Choose one of `"absolute"`, `"relative_key"`, `"relative_key_query"`. For
positional embeddings use `"absolute"`. For more information on `"relative_key"`, please refer to
[Self-Attention with Relative Position Representations (Shaw et al.)](https://arxiv.org/abs/1803.02155).
For more information on `"relative_key_query"`, please refer to *Method 4* in [Improve Transformer Models
with Better Relative Position Embeddings (Huang et al.)](https://arxiv.org/abs/2009.13658).
use_cache (`bool`, *optional*, defaults to `True`):
Whether or not the model should return the last key/values attentions (not used by all models). Only
relevant if `config.is_decoder=True`.
Example:
```python
>>> from transformers import AlignTextConfig, AlignTextModel
>>> # Initializing a AlignTextConfig with kakaobrain/align-base style configuration
>>> configuration = AlignTextConfig()
>>> # Initializing a AlignTextModel (with random weights) from the kakaobrain/align-base style configuration
>>> model = AlignTextModel(configuration)
>>> # Accessing the model configuration
>>> configuration = model.config
```"""
model_type = "align_text_model"
def __init__(
self,
vocab_size=30522,
hidden_size=768,
num_hidden_layers=12,
num_attention_heads=12,
intermediate_size=3072,
hidden_act="gelu",
hidden_dropout_prob=0.1,
attention_probs_dropout_prob=0.1,
max_position_embeddings=512,
type_vocab_size=2,
initializer_range=0.02,
layer_norm_eps=1e-12,
pad_token_id=0,
position_embedding_type="absolute",
use_cache=True,
**kwargs,
):
super().__init__(**kwargs)
self.vocab_size = vocab_size
self.hidden_size = hidden_size
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
self.hidden_act = hidden_act
self.intermediate_size = intermediate_size
self.hidden_dropout_prob = hidden_dropout_prob
self.attention_probs_dropout_prob = attention_probs_dropout_prob
self.max_position_embeddings = max_position_embeddings
self.type_vocab_size = type_vocab_size
self.initializer_range = initializer_range
self.layer_norm_eps = layer_norm_eps
self.position_embedding_type = position_embedding_type
self.use_cache = use_cache
self.pad_token_id = pad_token_id
@classmethod
def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
cls._set_token_in_kwargs(kwargs)
config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
# get the text config dict if we are loading from AlignConfig
if config_dict.get("model_type") == "align":
config_dict = config_dict["text_config"]
if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
logger.warning(
f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
)
return cls.from_dict(config_dict, **kwargs)
class AlignVisionConfig(PretrainedConfig):
r"""
This is the configuration class to store the configuration of a [`AlignVisionModel`]. It is used to instantiate a
ALIGN vision encoder according to the specified arguments, defining the model architecture. Instantiating a
configuration with the defaults will yield a similar configuration to that of the vision encoder of the ALIGN
[kakaobrain/align-base](https://huggingface.co/kakaobrain/align-base) architecture. The default values are copied
from EfficientNet (efficientnet-b7)
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PretrainedConfig`] for more information.
Args:
num_channels (`int`, *optional*, defaults to 3):
The number of input channels.
image_size (`int`, *optional*, defaults to 600):
The input image size.
width_coefficient (`float`, *optional*, defaults to 2.0):
Scaling coefficient for network width at each stage.
depth_coefficient (`float`, *optional*, defaults to 3.1):
Scaling coefficient for network depth at each stage.
depth_divisor `int`, *optional*, defaults to 8):
A unit of network width.
kernel_sizes (`List[int]`, *optional*, defaults to `[3, 3, 5, 3, 5, 5, 3]`):
List of kernel sizes to be used in each block.
in_channels (`List[int]`, *optional*, defaults to `[32, 16, 24, 40, 80, 112, 192]`):
List of input channel sizes to be used in each block for convolutional layers.
out_channels (`List[int]`, *optional*, defaults to `[16, 24, 40, 80, 112, 192, 320]`):
List of output channel sizes to be used in each block for convolutional layers.
depthwise_padding (`List[int]`, *optional*, defaults to `[]`):
List of block indices with square padding.
strides (`List[int]`, *optional*, defaults to `[1, 2, 2, 2, 1, 2, 1]`):
List of stride sizes to be used in each block for convolutional layers.
num_block_repeats (`List[int]`, *optional*, defaults to `[1, 2, 2, 3, 3, 4, 1]`):
List of the number of times each block is to repeated.
expand_ratios (`List[int]`, *optional*, defaults to `[1, 6, 6, 6, 6, 6, 6]`):
List of scaling coefficient of each block.
squeeze_expansion_ratio (`float`, *optional*, defaults to 0.25):
Squeeze expansion ratio.
hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
The non-linear activation function (function or string) in each block. If string, `"gelu"`, `"relu"`,
`"selu", `"gelu_new"`, `"silu"` and `"mish"` are supported.
hidden_dim (`int`, *optional*, defaults to 1280):
The hidden dimension of the layer before the classification head.
pooling_type (`str` or `function`, *optional*, defaults to `"mean"`):
Type of final pooling to be applied before the dense classification head. Available options are [`"mean"`,
`"max"`]
initializer_range (`float`, *optional*, defaults to 0.02):
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
batch_norm_eps (`float`, *optional*, defaults to 1e-3):
The epsilon used by the batch normalization layers.
batch_norm_momentum (`float`, *optional*, defaults to 0.99):
The momentum used by the batch normalization layers.
drop_connect_rate (`float`, *optional*, defaults to 0.2):
The drop rate for skip connections.
Example:
```python
>>> from transformers import AlignVisionConfig, AlignVisionModel
>>> # Initializing a AlignVisionConfig with kakaobrain/align-base style configuration
>>> configuration = AlignVisionConfig()
>>> # Initializing a AlignVisionModel (with random weights) from the kakaobrain/align-base style configuration
>>> model = AlignVisionModel(configuration)
>>> # Accessing the model configuration
>>> configuration = model.config
```"""
model_type = "align_vision_model"
def __init__(
self,
num_channels: int = 3,
image_size: int = 600,
width_coefficient: float = 2.0,
depth_coefficient: float = 3.1,
depth_divisor: int = 8,
kernel_sizes: List[int] = [3, 3, 5, 3, 5, 5, 3],
in_channels: List[int] = [32, 16, 24, 40, 80, 112, 192],
out_channels: List[int] = [16, 24, 40, 80, 112, 192, 320],
depthwise_padding: List[int] = [],
strides: List[int] = [1, 2, 2, 2, 1, 2, 1],
num_block_repeats: List[int] = [1, 2, 2, 3, 3, 4, 1],
expand_ratios: List[int] = [1, 6, 6, 6, 6, 6, 6],
squeeze_expansion_ratio: float = 0.25,
hidden_act: str = "swish",
hidden_dim: int = 2560,
pooling_type: str = "mean",
initializer_range: float = 0.02,
batch_norm_eps: float = 0.001,
batch_norm_momentum: float = 0.99,
drop_connect_rate: float = 0.2,
**kwargs,
):
super().__init__(**kwargs)
self.num_channels = num_channels
self.image_size = image_size
self.width_coefficient = width_coefficient
self.depth_coefficient = depth_coefficient
self.depth_divisor = depth_divisor
self.kernel_sizes = kernel_sizes
self.in_channels = in_channels
self.out_channels = out_channels
self.depthwise_padding = depthwise_padding
self.strides = strides
self.num_block_repeats = num_block_repeats
self.expand_ratios = expand_ratios
self.squeeze_expansion_ratio = squeeze_expansion_ratio
self.hidden_act = hidden_act
self.hidden_dim = hidden_dim
self.pooling_type = pooling_type
self.initializer_range = initializer_range
self.batch_norm_eps = batch_norm_eps
self.batch_norm_momentum = batch_norm_momentum
self.drop_connect_rate = drop_connect_rate
self.num_hidden_layers = sum(num_block_repeats) * 4
@classmethod
def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
cls._set_token_in_kwargs(kwargs)
config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
# get the vision config dict if we are loading from AlignConfig
if config_dict.get("model_type") == "align":
config_dict = config_dict["vision_config"]
if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
logger.warning(
f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
)
return cls.from_dict(config_dict, **kwargs)
class AlignConfig(PretrainedConfig):
r"""
[`AlignConfig`] is the configuration class to store the configuration of a [`AlignModel`]. It is used to
instantiate a ALIGN model according to the specified arguments, defining the text model and vision model configs.
Instantiating a configuration with the defaults will yield a similar configuration to that of the ALIGN
[kakaobrain/align-base](https://huggingface.co/kakaobrain/align-base) architecture.
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PretrainedConfig`] for more information.
Args:
text_config (`dict`, *optional*):
Dictionary of configuration options used to initialize [`AlignTextConfig`].
vision_config (`dict`, *optional*):
Dictionary of configuration options used to initialize [`AlignVisionConfig`].
projection_dim (`int`, *optional*, defaults to 640):
Dimensionality of text and vision projection layers.
temperature_init_value (`float`, *optional*, defaults to 1.0):
The initial value of the *temperature* parameter. Default is used as per the original ALIGN implementation.
initializer_range (`float`, *optional*, defaults to 0.02):
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
kwargs (*optional*):
Dictionary of keyword arguments.
Example:
```python
>>> from transformers import AlignConfig, AlignModel
>>> # Initializing a AlignConfig with kakaobrain/align-base style configuration
>>> configuration = AlignConfig()
>>> # Initializing a AlignModel (with random weights) from the kakaobrain/align-base style configuration
>>> model = AlignModel(configuration)
>>> # Accessing the model configuration
>>> configuration = model.config
>>> # We can also initialize a AlignConfig from a AlignTextConfig and a AlignVisionConfig
>>> from transformers import AlignTextConfig, AlignVisionConfig
>>> # Initializing ALIGN Text and Vision configurations
>>> config_text = AlignTextConfig()
>>> config_vision = AlignVisionConfig()
>>> config = AlignConfig.from_text_vision_configs(config_text, config_vision)
```"""
model_type = "align"
def __init__(
self,
text_config=None,
vision_config=None,
projection_dim=640,
temperature_init_value=1.0,
initializer_range=0.02,
**kwargs,
):
super().__init__(**kwargs)
if text_config is None:
text_config = {}
logger.info("text_config is None. Initializing the AlignTextConfig with default values.")
if vision_config is None:
vision_config = {}
logger.info("vision_config is None. Initializing the AlignVisionConfig with default values.")
self.text_config = AlignTextConfig(**text_config)
self.vision_config = AlignVisionConfig(**vision_config)
self.projection_dim = projection_dim
self.temperature_init_value = temperature_init_value
self.initializer_range = initializer_range
@classmethod
def from_text_vision_configs(cls, text_config: AlignTextConfig, vision_config: AlignVisionConfig, **kwargs):
r"""
Instantiate a [`AlignConfig`] (or a derived class) from align text model configuration and align vision model
configuration.
Returns:
[`AlignConfig`]: An instance of a configuration object
"""
return cls(text_config=text_config.to_dict(), vision_config=vision_config.to_dict(), **kwargs)
|
transformers/src/transformers/models/align/configuration_align.py/0
|
{
"file_path": "transformers/src/transformers/models/align/configuration_align.py",
"repo_id": "transformers",
"token_count": 6810
}
| 342
|
# coding=utf-8
# Copyright 2021 The HuggingFace Inc. team.
#
# 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.
"""AutoFeatureExtractor class."""
import importlib
import json
import os
import warnings
from collections import OrderedDict
from typing import Dict, Optional, Union
# Build the list of all feature extractors
from ...configuration_utils import PretrainedConfig
from ...dynamic_module_utils import get_class_from_dynamic_module, resolve_trust_remote_code
from ...feature_extraction_utils import FeatureExtractionMixin
from ...utils import CONFIG_NAME, FEATURE_EXTRACTOR_NAME, get_file_from_repo, logging
from .auto_factory import _LazyAutoMapping
from .configuration_auto import (
CONFIG_MAPPING_NAMES,
AutoConfig,
model_type_to_module_name,
replace_list_option_in_docstrings,
)
logger = logging.get_logger(__name__)
FEATURE_EXTRACTOR_MAPPING_NAMES = OrderedDict(
[
("audio-spectrogram-transformer", "ASTFeatureExtractor"),
("beit", "BeitFeatureExtractor"),
("chinese_clip", "ChineseCLIPFeatureExtractor"),
("clap", "ClapFeatureExtractor"),
("clip", "CLIPFeatureExtractor"),
("clipseg", "ViTFeatureExtractor"),
("clvp", "ClvpFeatureExtractor"),
("conditional_detr", "ConditionalDetrFeatureExtractor"),
("convnext", "ConvNextFeatureExtractor"),
("cvt", "ConvNextFeatureExtractor"),
("dac", "DacFeatureExtractor"),
("data2vec-audio", "Wav2Vec2FeatureExtractor"),
("data2vec-vision", "BeitFeatureExtractor"),
("deformable_detr", "DeformableDetrFeatureExtractor"),
("deit", "DeiTFeatureExtractor"),
("detr", "DetrFeatureExtractor"),
("dinat", "ViTFeatureExtractor"),
("donut-swin", "DonutFeatureExtractor"),
("dpt", "DPTFeatureExtractor"),
("encodec", "EncodecFeatureExtractor"),
("flava", "FlavaFeatureExtractor"),
("glpn", "GLPNFeatureExtractor"),
("groupvit", "CLIPFeatureExtractor"),
("hubert", "Wav2Vec2FeatureExtractor"),
("imagegpt", "ImageGPTFeatureExtractor"),
("layoutlmv2", "LayoutLMv2FeatureExtractor"),
("layoutlmv3", "LayoutLMv3FeatureExtractor"),
("levit", "LevitFeatureExtractor"),
("maskformer", "MaskFormerFeatureExtractor"),
("mctct", "MCTCTFeatureExtractor"),
("mobilenet_v1", "MobileNetV1FeatureExtractor"),
("mobilenet_v2", "MobileNetV2FeatureExtractor"),
("mobilevit", "MobileViTFeatureExtractor"),
("nat", "ViTFeatureExtractor"),
("owlvit", "OwlViTFeatureExtractor"),
("perceiver", "PerceiverFeatureExtractor"),
("poolformer", "PoolFormerFeatureExtractor"),
("pop2piano", "Pop2PianoFeatureExtractor"),
("regnet", "ConvNextFeatureExtractor"),
("resnet", "ConvNextFeatureExtractor"),
("seamless_m4t", "SeamlessM4TFeatureExtractor"),
("seamless_m4t_v2", "SeamlessM4TFeatureExtractor"),
("segformer", "SegformerFeatureExtractor"),
("sew", "Wav2Vec2FeatureExtractor"),
("sew-d", "Wav2Vec2FeatureExtractor"),
("speech_to_text", "Speech2TextFeatureExtractor"),
("speecht5", "SpeechT5FeatureExtractor"),
("swiftformer", "ViTFeatureExtractor"),
("swin", "ViTFeatureExtractor"),
("swinv2", "ViTFeatureExtractor"),
("table-transformer", "DetrFeatureExtractor"),
("timesformer", "VideoMAEFeatureExtractor"),
("tvlt", "TvltFeatureExtractor"),
("unispeech", "Wav2Vec2FeatureExtractor"),
("unispeech-sat", "Wav2Vec2FeatureExtractor"),
("univnet", "UnivNetFeatureExtractor"),
("van", "ConvNextFeatureExtractor"),
("videomae", "VideoMAEFeatureExtractor"),
("vilt", "ViltFeatureExtractor"),
("vit", "ViTFeatureExtractor"),
("vit_mae", "ViTFeatureExtractor"),
("vit_msn", "ViTFeatureExtractor"),
("wav2vec2", "Wav2Vec2FeatureExtractor"),
("wav2vec2-bert", "Wav2Vec2FeatureExtractor"),
("wav2vec2-conformer", "Wav2Vec2FeatureExtractor"),
("wavlm", "Wav2Vec2FeatureExtractor"),
("whisper", "WhisperFeatureExtractor"),
("xclip", "CLIPFeatureExtractor"),
("yolos", "YolosFeatureExtractor"),
]
)
FEATURE_EXTRACTOR_MAPPING = _LazyAutoMapping(CONFIG_MAPPING_NAMES, FEATURE_EXTRACTOR_MAPPING_NAMES)
def feature_extractor_class_from_name(class_name: str):
for module_name, extractors in FEATURE_EXTRACTOR_MAPPING_NAMES.items():
if class_name in extractors:
module_name = model_type_to_module_name(module_name)
module = importlib.import_module(f".{module_name}", "transformers.models")
try:
return getattr(module, class_name)
except AttributeError:
continue
for _, extractor in FEATURE_EXTRACTOR_MAPPING._extra_content.items():
if getattr(extractor, "__name__", None) == class_name:
return extractor
# We did not fine the class, but maybe it's because a dep is missing. In that case, the class will be in the main
# init and we return the proper dummy to get an appropriate error message.
main_module = importlib.import_module("transformers")
if hasattr(main_module, class_name):
return getattr(main_module, class_name)
return None
def get_feature_extractor_config(
pretrained_model_name_or_path: Union[str, os.PathLike],
cache_dir: Optional[Union[str, os.PathLike]] = None,
force_download: bool = False,
resume_download: Optional[bool] = None,
proxies: Optional[Dict[str, str]] = None,
token: Optional[Union[bool, str]] = None,
revision: Optional[str] = None,
local_files_only: bool = False,
**kwargs,
):
"""
Loads the tokenizer configuration from a pretrained model tokenizer configuration.
Args:
pretrained_model_name_or_path (`str` or `os.PathLike`):
This can be either:
- a string, the *model id* of a pretrained model configuration hosted inside a model repo on
huggingface.co.
- a path to a *directory* containing a configuration file saved using the
[`~PreTrainedTokenizer.save_pretrained`] method, e.g., `./my_model_directory/`.
cache_dir (`str` or `os.PathLike`, *optional*):
Path to a directory in which a downloaded pretrained model configuration should be cached if the standard
cache should not be used.
force_download (`bool`, *optional*, defaults to `False`):
Whether or not to force to (re-)download the configuration files and override the cached versions if they
exist.
resume_download:
Deprecated and ignored. All downloads are now resumed by default when possible.
Will be removed in v5 of Transformers.
proxies (`Dict[str, str]`, *optional*):
A dictionary of proxy servers to use by protocol or endpoint, e.g., `{'http': 'foo.bar:3128',
'http://hostname': 'foo.bar:4012'}.` The proxies are used on each request.
token (`str` or *bool*, *optional*):
The token to use as HTTP bearer authorization for remote files. If `True`, will use the token generated
when running `huggingface-cli login` (stored in `~/.huggingface`).
revision (`str`, *optional*, defaults to `"main"`):
The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a
git-based system for storing models and other artifacts on huggingface.co, so `revision` can be any
identifier allowed by git.
local_files_only (`bool`, *optional*, defaults to `False`):
If `True`, will only try to load the tokenizer configuration from local files.
<Tip>
Passing `token=True` is required when you want to use a private model.
</Tip>
Returns:
`Dict`: The configuration of the tokenizer.
Examples:
```python
# Download configuration from huggingface.co and cache.
tokenizer_config = get_tokenizer_config("google-bert/bert-base-uncased")
# This model does not have a tokenizer config so the result will be an empty dict.
tokenizer_config = get_tokenizer_config("FacebookAI/xlm-roberta-base")
# Save a pretrained tokenizer locally and you can reload its config
from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("google-bert/bert-base-cased")
tokenizer.save_pretrained("tokenizer-test")
tokenizer_config = get_tokenizer_config("tokenizer-test")
```"""
use_auth_token = kwargs.pop("use_auth_token", None)
if use_auth_token is not None:
warnings.warn(
"The `use_auth_token` argument is deprecated and will be removed in v5 of Transformers. Please use `token` instead.",
FutureWarning,
)
if token is not None:
raise ValueError("`token` and `use_auth_token` are both specified. Please set only the argument `token`.")
token = use_auth_token
resolved_config_file = get_file_from_repo(
pretrained_model_name_or_path,
FEATURE_EXTRACTOR_NAME,
cache_dir=cache_dir,
force_download=force_download,
resume_download=resume_download,
proxies=proxies,
token=token,
revision=revision,
local_files_only=local_files_only,
)
if resolved_config_file is None:
logger.info(
"Could not locate the feature extractor configuration file, will try to use the model config instead."
)
return {}
with open(resolved_config_file, encoding="utf-8") as reader:
return json.load(reader)
class AutoFeatureExtractor:
r"""
This is a generic feature extractor class that will be instantiated as one of the feature extractor classes of the
library when created with the [`AutoFeatureExtractor.from_pretrained`] class method.
This class cannot be instantiated directly using `__init__()` (throws an error).
"""
def __init__(self):
raise EnvironmentError(
"AutoFeatureExtractor is designed to be instantiated "
"using the `AutoFeatureExtractor.from_pretrained(pretrained_model_name_or_path)` method."
)
@classmethod
@replace_list_option_in_docstrings(FEATURE_EXTRACTOR_MAPPING_NAMES)
def from_pretrained(cls, pretrained_model_name_or_path, **kwargs):
r"""
Instantiate one of the feature extractor classes of the library from a pretrained model vocabulary.
The feature extractor class to instantiate is selected based on the `model_type` property of the config object
(either passed as an argument or loaded from `pretrained_model_name_or_path` if possible), or when it's
missing, by falling back to using pattern matching on `pretrained_model_name_or_path`:
List options
Params:
pretrained_model_name_or_path (`str` or `os.PathLike`):
This can be either:
- a string, the *model id* of a pretrained feature_extractor hosted inside a model repo on
huggingface.co.
- a path to a *directory* containing a feature extractor file saved using the
[`~feature_extraction_utils.FeatureExtractionMixin.save_pretrained`] method, e.g.,
`./my_model_directory/`.
- a path or url to a saved feature extractor JSON *file*, e.g.,
`./my_model_directory/preprocessor_config.json`.
cache_dir (`str` or `os.PathLike`, *optional*):
Path to a directory in which a downloaded pretrained model feature extractor should be cached if the
standard cache should not be used.
force_download (`bool`, *optional*, defaults to `False`):
Whether or not to force to (re-)download the feature extractor files and override the cached versions
if they exist.
resume_download:
Deprecated and ignored. All downloads are now resumed by default when possible.
Will be removed in v5 of Transformers.
proxies (`Dict[str, str]`, *optional*):
A dictionary of proxy servers to use by protocol or endpoint, e.g., `{'http': 'foo.bar:3128',
'http://hostname': 'foo.bar:4012'}.` The proxies are used on each request.
token (`str` or *bool*, *optional*):
The token to use as HTTP bearer authorization for remote files. If `True`, will use the token generated
when running `huggingface-cli login` (stored in `~/.huggingface`).
revision (`str`, *optional*, defaults to `"main"`):
The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a
git-based system for storing models and other artifacts on huggingface.co, so `revision` can be any
identifier allowed by git.
return_unused_kwargs (`bool`, *optional*, defaults to `False`):
If `False`, then this function returns just the final feature extractor object. If `True`, then this
functions returns a `Tuple(feature_extractor, unused_kwargs)` where *unused_kwargs* is a dictionary
consisting of the key/value pairs whose keys are not feature extractor attributes: i.e., the part of
`kwargs` which has not been used to update `feature_extractor` and is otherwise ignored.
trust_remote_code (`bool`, *optional*, defaults to `False`):
Whether or not to allow for custom models defined on the Hub in their own modeling files. This option
should only be set to `True` for repositories you trust and in which you have read the code, as it will
execute code present on the Hub on your local machine.
kwargs (`Dict[str, Any]`, *optional*):
The values in kwargs of any keys which are feature extractor attributes will be used to override the
loaded values. Behavior concerning key/value pairs whose keys are *not* feature extractor attributes is
controlled by the `return_unused_kwargs` keyword parameter.
<Tip>
Passing `token=True` is required when you want to use a private model.
</Tip>
Examples:
```python
>>> from transformers import AutoFeatureExtractor
>>> # Download feature extractor from huggingface.co and cache.
>>> feature_extractor = AutoFeatureExtractor.from_pretrained("facebook/wav2vec2-base-960h")
>>> # If feature extractor files are in a directory (e.g. feature extractor was saved using *save_pretrained('./test/saved_model/')*)
>>> # feature_extractor = AutoFeatureExtractor.from_pretrained("./test/saved_model/")
```"""
use_auth_token = kwargs.pop("use_auth_token", None)
if use_auth_token is not None:
warnings.warn(
"The `use_auth_token` argument is deprecated and will be removed in v5 of Transformers. Please use `token` instead.",
FutureWarning,
)
if kwargs.get("token", None) is not None:
raise ValueError(
"`token` and `use_auth_token` are both specified. Please set only the argument `token`."
)
kwargs["token"] = use_auth_token
config = kwargs.pop("config", None)
trust_remote_code = kwargs.pop("trust_remote_code", None)
kwargs["_from_auto"] = True
config_dict, _ = FeatureExtractionMixin.get_feature_extractor_dict(pretrained_model_name_or_path, **kwargs)
feature_extractor_class = config_dict.get("feature_extractor_type", None)
feature_extractor_auto_map = None
if "AutoFeatureExtractor" in config_dict.get("auto_map", {}):
feature_extractor_auto_map = config_dict["auto_map"]["AutoFeatureExtractor"]
# If we don't find the feature extractor class in the feature extractor config, let's try the model config.
if feature_extractor_class is None and feature_extractor_auto_map is None:
if not isinstance(config, PretrainedConfig):
config = AutoConfig.from_pretrained(pretrained_model_name_or_path, **kwargs)
# It could be in `config.feature_extractor_type``
feature_extractor_class = getattr(config, "feature_extractor_type", None)
if hasattr(config, "auto_map") and "AutoFeatureExtractor" in config.auto_map:
feature_extractor_auto_map = config.auto_map["AutoFeatureExtractor"]
if feature_extractor_class is not None:
feature_extractor_class = feature_extractor_class_from_name(feature_extractor_class)
has_remote_code = feature_extractor_auto_map is not None
has_local_code = feature_extractor_class is not None or type(config) in FEATURE_EXTRACTOR_MAPPING
trust_remote_code = resolve_trust_remote_code(
trust_remote_code, pretrained_model_name_or_path, has_local_code, has_remote_code
)
if has_remote_code and trust_remote_code:
feature_extractor_class = get_class_from_dynamic_module(
feature_extractor_auto_map, pretrained_model_name_or_path, **kwargs
)
_ = kwargs.pop("code_revision", None)
if os.path.isdir(pretrained_model_name_or_path):
feature_extractor_class.register_for_auto_class()
return feature_extractor_class.from_dict(config_dict, **kwargs)
elif feature_extractor_class is not None:
return feature_extractor_class.from_dict(config_dict, **kwargs)
# Last try: we use the FEATURE_EXTRACTOR_MAPPING.
elif type(config) in FEATURE_EXTRACTOR_MAPPING:
feature_extractor_class = FEATURE_EXTRACTOR_MAPPING[type(config)]
return feature_extractor_class.from_dict(config_dict, **kwargs)
raise ValueError(
f"Unrecognized feature extractor in {pretrained_model_name_or_path}. Should have a "
f"`feature_extractor_type` key in its {FEATURE_EXTRACTOR_NAME} of {CONFIG_NAME}, or one of the following "
f"`model_type` keys in its {CONFIG_NAME}: {', '.join(c for c in FEATURE_EXTRACTOR_MAPPING_NAMES.keys())}"
)
@staticmethod
def register(config_class, feature_extractor_class, exist_ok=False):
"""
Register a new feature extractor for this class.
Args:
config_class ([`PretrainedConfig`]):
The configuration corresponding to the model to register.
feature_extractor_class ([`FeatureExtractorMixin`]): The feature extractor to register.
"""
FEATURE_EXTRACTOR_MAPPING.register(config_class, feature_extractor_class, exist_ok=exist_ok)
|
transformers/src/transformers/models/auto/feature_extraction_auto.py/0
|
{
"file_path": "transformers/src/transformers/models/auto/feature_extraction_auto.py",
"repo_id": "transformers",
"token_count": 7718
}
| 343
|
# Copyright 2020 The HuggingFace Team. 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.
from typing import TYPE_CHECKING
from ...utils import (
OptionalDependencyNotAvailable,
_LazyModule,
is_flax_available,
is_tf_available,
is_tokenizers_available,
is_torch_available,
)
_import_structure = {
"configuration_bart": ["BartConfig", "BartOnnxConfig"],
"tokenization_bart": ["BartTokenizer"],
}
try:
if not is_tokenizers_available():
raise OptionalDependencyNotAvailable()
except OptionalDependencyNotAvailable:
pass
else:
_import_structure["tokenization_bart_fast"] = ["BartTokenizerFast"]
try:
if not is_torch_available():
raise OptionalDependencyNotAvailable()
except OptionalDependencyNotAvailable:
pass
else:
_import_structure["modeling_bart"] = [
"BartForCausalLM",
"BartForConditionalGeneration",
"BartForQuestionAnswering",
"BartForSequenceClassification",
"BartModel",
"BartPreTrainedModel",
"BartPretrainedModel",
"PretrainedBartModel",
]
try:
if not is_tf_available():
raise OptionalDependencyNotAvailable()
except OptionalDependencyNotAvailable:
pass
else:
_import_structure["modeling_tf_bart"] = [
"TFBartForConditionalGeneration",
"TFBartForSequenceClassification",
"TFBartModel",
"TFBartPretrainedModel",
]
try:
if not is_flax_available():
raise OptionalDependencyNotAvailable()
except OptionalDependencyNotAvailable:
pass
else:
_import_structure["modeling_flax_bart"] = [
"FlaxBartDecoderPreTrainedModel",
"FlaxBartForCausalLM",
"FlaxBartForConditionalGeneration",
"FlaxBartForQuestionAnswering",
"FlaxBartForSequenceClassification",
"FlaxBartModel",
"FlaxBartPreTrainedModel",
]
if TYPE_CHECKING:
from .configuration_bart import BartConfig, BartOnnxConfig
from .tokenization_bart import BartTokenizer
try:
if not is_tokenizers_available():
raise OptionalDependencyNotAvailable()
except OptionalDependencyNotAvailable:
pass
else:
from .tokenization_bart_fast import BartTokenizerFast
try:
if not is_torch_available():
raise OptionalDependencyNotAvailable()
except OptionalDependencyNotAvailable:
pass
else:
from .modeling_bart import (
BartForCausalLM,
BartForConditionalGeneration,
BartForQuestionAnswering,
BartForSequenceClassification,
BartModel,
BartPreTrainedModel,
BartPretrainedModel,
PretrainedBartModel,
)
try:
if not is_tf_available():
raise OptionalDependencyNotAvailable()
except OptionalDependencyNotAvailable:
pass
else:
from .modeling_tf_bart import (
TFBartForConditionalGeneration,
TFBartForSequenceClassification,
TFBartModel,
TFBartPretrainedModel,
)
try:
if not is_flax_available():
raise OptionalDependencyNotAvailable()
except OptionalDependencyNotAvailable:
pass
else:
from .modeling_flax_bart import (
FlaxBartDecoderPreTrainedModel,
FlaxBartForCausalLM,
FlaxBartForConditionalGeneration,
FlaxBartForQuestionAnswering,
FlaxBartForSequenceClassification,
FlaxBartModel,
FlaxBartPreTrainedModel,
)
else:
import sys
sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)
|
transformers/src/transformers/models/bart/__init__.py/0
|
{
"file_path": "transformers/src/transformers/models/bart/__init__.py",
"repo_id": "transformers",
"token_count": 1751
}
| 344
|
# coding=utf-8
# Copyright 2022 The Salesforce Team Authors and The HuggingFace Team. 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.
"""PyTorch BLIP model."""
import math
import warnings
from dataclasses import dataclass
from typing import Any, Optional, Tuple, Union
import torch
import torch.utils.checkpoint
from torch import nn
from torch.nn.functional import normalize
from ...activations import ACT2FN
from ...modeling_outputs import BaseModelOutput, BaseModelOutputWithPooling
from ...modeling_utils import PreTrainedModel
from ...utils import (
ModelOutput,
add_start_docstrings,
add_start_docstrings_to_model_forward,
logging,
replace_return_docstrings,
)
from .configuration_blip import BlipConfig, BlipTextConfig, BlipVisionConfig
from .modeling_blip_text import BlipTextLMHeadModel, BlipTextModel
logger = logging.get_logger(__name__)
_CHECKPOINT_FOR_DOC = "Salesforce/blip-vqa-base"
# Copied from transformers.models.clip.modeling_clip.contrastive_loss
def contrastive_loss(logits: torch.Tensor) -> torch.Tensor:
return nn.functional.cross_entropy(logits, torch.arange(len(logits), device=logits.device))
# Copied from transformers.models.clip.modeling_clip.clip_loss with clip->blip
def blip_loss(similarity: torch.Tensor) -> torch.Tensor:
caption_loss = contrastive_loss(similarity)
image_loss = contrastive_loss(similarity.t())
return (caption_loss + image_loss) / 2.0
@dataclass
class BlipForConditionalGenerationModelOutput(ModelOutput):
"""
Adapted from the base class for vision model's outputs that also contains image embeddings of the pooling of the
last hidden states. This class also adds the loss term from the text decoder.
Args:
loss (`torch.FloatTensor`, *optional*, returned when `labels` is provided, `torch.FloatTensor` of shape `(1,)`):
Languge modeling loss from the text decoder.
logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`, *optional*):
Prediction scores of the language modeling head of the text decoder model.
image_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim)`, *optional*):
The image embeddings obtained after applying the Vision Transformer model to the input image.
last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
Sequence of hidden-states at the output of the last layer of the model.
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
"""
loss: Optional[Tuple[torch.FloatTensor]] = None
logits: Optional[Tuple[torch.FloatTensor]] = None
image_embeds: Optional[torch.FloatTensor] = None
last_hidden_state: torch.FloatTensor = None
hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
@property
def decoder_logits(self):
warnings.warn(
"`decoder_logits` attribute is deprecated and will be removed in version 5 of Transformers."
" Please use the `logits` attribute to retrieve the final output instead.",
FutureWarning,
)
return self.logits
@dataclass
class BlipTextVisionModelOutput(ModelOutput):
"""
Adapted from the base class for vision model's outputs that also contains image embeddings of the pooling of the
last hidden states. This class also adds the loss term from the text decoder.
Args:
loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
Languge modeling loss from the text decoder.
image_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim)` *optional* returned when model is initialized with `with_projection=True`):
The image embeddings obtained by applying the projection layer to the pooler_output.
last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
Sequence of hidden-states at the output of the last layer of the model.
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
"""
loss: Optional[torch.FloatTensor] = None
image_embeds: Optional[torch.FloatTensor] = None
last_hidden_state: torch.FloatTensor = None
hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
@dataclass
class BlipImageTextMatchingModelOutput(ModelOutput):
"""
Adapted from the base class for vision model's outputs that also contains image embeddings of the pooling of the
last hidden states. This class also adds the loss term from the text decoder as well as the image-text similarity
scores.
Args:
itm_score (`torch.FloatTensor`):
The image-text similarity scores.
loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
Languge modeling loss from the text decoder.
image_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim)` *optional* returned when model is initialized with `with_projection=True`):
The image embeddings obtained by applying the projection layer to the pooler_output.
last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
Sequence of hidden-states at the output of the last layer of the model.
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
vision_pooler_output (`torch.FloatTensor` of shape `(batch_size, hidden_size)`, *optional*):
Last layer hidden-state of the vision of the vision-only branch of the model.
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
question_embeds (`torch.FloatTensor`):
The question embeddings obtained by the text projection layer.
"""
itm_score: Optional[torch.FloatTensor] = None
loss: Optional[torch.FloatTensor] = None
image_embeds: Optional[torch.FloatTensor] = None
last_hidden_state: torch.FloatTensor = None
hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
vision_pooler_output: Optional[torch.FloatTensor] = None
attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
question_embeds: Optional[Tuple[torch.FloatTensor]] = None
@dataclass
class BlipOutput(ModelOutput):
"""
Args:
loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `return_loss` is `True`):
Contrastive loss for image-text similarity.
logits_per_image:(`torch.FloatTensor` of shape `(image_batch_size, text_batch_size)`):
The scaled dot product scores between `image_embeds` and `text_embeds`. This represents the image-text
similarity scores.
logits_per_text:(`torch.FloatTensor` of shape `(text_batch_size, image_batch_size)`):
The scaled dot product scores between `text_embeds` and `image_embeds`. This represents the text-image
similarity scores.
text_embeds(`torch.FloatTensor` of shape `(batch_size, output_dim`):
The text embeddings obtained by applying the projection layer to the pooled output of [`BlipTextModel`].
image_embeds(`torch.FloatTensor` of shape `(batch_size, output_dim`):
The image embeddings obtained by applying the projection layer to the pooled output of [`BlipVisionModel`].
text_model_output(`BaseModelOutputWithPooling`):
The output of the [`BlipTextModel`].
vision_model_output(`BaseModelOutputWithPooling`):
The output of the [`BlipVisionModel`].
"""
loss: Optional[torch.FloatTensor] = None
logits_per_image: torch.FloatTensor = None
logits_per_text: torch.FloatTensor = None
text_embeds: torch.FloatTensor = None
image_embeds: torch.FloatTensor = None
text_model_output: BaseModelOutputWithPooling = None
vision_model_output: BaseModelOutputWithPooling = None
def to_tuple(self) -> Tuple[Any]:
return tuple(
self[k] if k not in ["text_model_output", "vision_model_output"] else getattr(self, k).to_tuple()
for k in self.keys()
)
class BlipVisionEmbeddings(nn.Module):
def __init__(self, config: BlipVisionConfig):
super().__init__()
self.config = config
self.embed_dim = config.hidden_size
self.image_size = config.image_size
self.patch_size = config.patch_size
self.class_embedding = nn.Parameter(torch.randn(1, 1, self.embed_dim))
self.patch_embedding = nn.Conv2d(
in_channels=3, out_channels=self.embed_dim, kernel_size=self.patch_size, stride=self.patch_size
)
self.num_patches = (self.image_size // self.patch_size) ** 2
self.num_positions = self.num_patches + 1
self.position_embedding = nn.Parameter(torch.randn(1, self.num_positions, self.embed_dim))
def interpolate_pos_encoding(self, embeddings: torch.Tensor, height: int, width: int) -> torch.Tensor:
"""
This method allows to interpolate the pre-trained position encodings, to be able to use the model on higher
resolution images.
Source:
https://github.com/facebookresearch/dino/blob/de9ee3df6cf39fac952ab558447af1fa1365362a/vision_transformer.py#L174
"""
num_patches = embeddings.shape[1] - 1
num_positions = self.position_embedding.shape[1] - 1
if num_patches == num_positions and height == width:
return self.position_embedding
class_pos_embed = self.position_embedding[:, 0, :]
patch_pos_embed = self.position_embedding[:, 1:, :]
dim = embeddings.shape[-1]
h0 = height // self.config.patch_size
w0 = width // self.config.patch_size
# we add a small number to avoid floating point error in the interpolation
# see discussion at https://github.com/facebookresearch/dino/issues/8
h0, w0 = h0 + 0.1, w0 + 0.1
patch_pos_embed = patch_pos_embed.reshape(1, int(math.sqrt(num_positions)), int(math.sqrt(num_positions)), dim)
patch_pos_embed = patch_pos_embed.permute(0, 3, 1, 2)
patch_pos_embed = nn.functional.interpolate(
patch_pos_embed,
scale_factor=(h0 / math.sqrt(num_positions), w0 / math.sqrt(num_positions)),
mode="bicubic",
align_corners=False,
)
patch_pos_embed = patch_pos_embed.permute(0, 2, 3, 1).view(1, -1, dim)
return torch.cat((class_pos_embed.unsqueeze(0), patch_pos_embed), dim=1)
def forward(self, pixel_values: torch.FloatTensor, interpolate_pos_encoding: bool = False) -> torch.Tensor:
batch_size, _, height, width = pixel_values.shape
target_dtype = self.patch_embedding.weight.dtype
patch_embeds = self.patch_embedding(pixel_values.to(dtype=target_dtype)) # shape = [*, width, grid, grid]
patch_embeds = patch_embeds.flatten(2).transpose(1, 2)
class_embeds = self.class_embedding.expand(batch_size, 1, -1).to(target_dtype)
embeddings = torch.cat([class_embeds, patch_embeds], dim=1)
if interpolate_pos_encoding:
position_embedding = self.interpolate_pos_encoding(embeddings, height, width)
else:
position_embedding = self.position_embedding
embeddings = embeddings + position_embedding[:, : embeddings.size(1), :].to(target_dtype)
return embeddings
# Copied from transformers.models.clip.modeling_clip.CLIPTextEmbeddings with CLIP->Blip
class BlipTextEmbeddings(nn.Module):
def __init__(self, config: BlipTextConfig):
super().__init__()
embed_dim = config.hidden_size
self.token_embedding = nn.Embedding(config.vocab_size, embed_dim)
self.position_embedding = nn.Embedding(config.max_position_embeddings, embed_dim)
# position_ids (1, len position emb) is contiguous in memory and exported when serialized
self.register_buffer(
"position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)), persistent=False
)
def forward(
self,
input_ids: Optional[torch.LongTensor] = None,
position_ids: Optional[torch.LongTensor] = None,
inputs_embeds: Optional[torch.FloatTensor] = None,
) -> torch.Tensor:
seq_length = input_ids.shape[-1] if input_ids is not None else inputs_embeds.shape[-2]
if position_ids is None:
position_ids = self.position_ids[:, :seq_length]
if inputs_embeds is None:
inputs_embeds = self.token_embedding(input_ids)
position_embeddings = self.position_embedding(position_ids)
embeddings = inputs_embeds + position_embeddings
return embeddings
class BlipAttention(nn.Module):
"""Multi-headed attention from 'Attention Is All You Need' paper"""
def __init__(self, config):
super().__init__()
self.config = config
self.embed_dim = config.hidden_size
self.num_heads = config.num_attention_heads
self.head_dim = self.embed_dim // self.num_heads
if self.head_dim * self.num_heads != self.embed_dim:
raise ValueError(
f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:"
f" {self.num_heads})."
)
self.scale = self.head_dim**-0.5
self.dropout = nn.Dropout(config.attention_dropout)
self.qkv = nn.Linear(self.embed_dim, 3 * self.embed_dim)
self.projection = nn.Linear(self.embed_dim, self.embed_dim)
def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
def forward(
self,
hidden_states: torch.Tensor,
head_mask: Optional[torch.Tensor] = None,
output_attentions: Optional[bool] = False,
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
"""Input shape: Batch x Time x Channel"""
bsz, tgt_len, embed_dim = hidden_states.size()
mixed_qkv = (
self.qkv(hidden_states)
.reshape(bsz, tgt_len, 3, self.num_heads, embed_dim // self.num_heads)
.permute(2, 0, 3, 1, 4)
)
query_states, key_states, value_states = mixed_qkv[0], mixed_qkv[1], mixed_qkv[2]
# Take the dot product between "query" and "key" to get the raw attention scores.
attention_scores = torch.matmul(query_states, key_states.transpose(-1, -2))
attention_scores = attention_scores * self.scale
# Normalize the attention scores to probabilities.
attention_probs = nn.functional.softmax(attention_scores, dim=-1)
# This is actually dropping out entire tokens to attend to, which might
# seem a bit unusual, but is taken from the original Transformer paper.
attention_probs = self.dropout(attention_probs)
# Mask heads if we want to
if head_mask is not None:
attention_probs = attention_probs * head_mask
context_layer = torch.matmul(attention_probs, value_states).permute(0, 2, 1, 3)
new_context_layer_shape = context_layer.size()[:-2] + (self.embed_dim,)
context_layer = context_layer.reshape(new_context_layer_shape)
output = self.projection(context_layer)
outputs = (output, attention_probs) if output_attentions else (output, None)
return outputs
# Copied from transformers.models.clip.modeling_clip.CLIPMLP with CLIP->Blip
class BlipMLP(nn.Module):
def __init__(self, config):
super().__init__()
self.config = config
self.activation_fn = ACT2FN[config.hidden_act]
self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
hidden_states = self.fc1(hidden_states)
hidden_states = self.activation_fn(hidden_states)
hidden_states = self.fc2(hidden_states)
return hidden_states
class BlipEncoderLayer(nn.Module):
def __init__(self, config: BlipConfig):
super().__init__()
self.embed_dim = config.hidden_size
self.self_attn = BlipAttention(config)
self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
self.mlp = BlipMLP(config)
self.layer_norm2 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
def forward(
self,
hidden_states: torch.Tensor,
attention_mask: torch.Tensor,
output_attentions: Optional[bool] = False,
) -> Tuple[torch.FloatTensor]:
"""
Args:
hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
attention_mask (`torch.FloatTensor`): attention mask of size
`(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
`(config.encoder_attention_heads,)`.
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
returned tensors for more detail.
"""
residual = hidden_states
hidden_states = self.layer_norm1(hidden_states)
hidden_states, attn_weights = self.self_attn(
hidden_states=hidden_states,
head_mask=attention_mask,
output_attentions=output_attentions,
)
hidden_states = hidden_states + residual
residual = hidden_states
hidden_states = self.layer_norm2(hidden_states)
hidden_states = self.mlp(hidden_states)
hidden_states = hidden_states + residual
outputs = (hidden_states,)
if output_attentions:
outputs += (attn_weights,)
return outputs
class BlipPreTrainedModel(PreTrainedModel):
"""
An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
models.
"""
config_class = BlipConfig
base_model_prefix = "blip"
supports_gradient_checkpointing = True
def _init_weights(self, module):
"""Initialize the weights"""
factor = self.config.initializer_range
if isinstance(module, nn.Conv2d) or isinstance(module, nn.Embedding) or isinstance(module, nn.Linear):
module.weight.data.normal_(mean=0.0, std=factor)
if hasattr(module, "bias") and module.bias is not None:
module.bias.data.zero_()
if isinstance(module, BlipVisionEmbeddings):
if hasattr(self.config, "vision_config"):
factor = self.config.vision_config.initializer_range
nn.init.trunc_normal_(
module.position_embedding,
mean=0.0,
std=factor,
)
nn.init.trunc_normal_(
module.class_embedding,
mean=0.0,
std=factor,
)
elif isinstance(module, nn.LayerNorm):
module.bias.data.zero_()
module.weight.data.fill_(1.0)
elif isinstance(module, nn.Linear) and module.bias is not None:
module.bias.data.zero_()
BLIP_START_DOCSTRING = r"""
This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
etc.)
This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
and behavior.
Parameters:
config ([`BlipConfig`]): Model configuration class with all the parameters of the model.
Initializing with a config file does not load the weights associated with the model, only the
configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
"""
BLIP_TEXT_INPUTS_DOCSTRING = r"""
Args:
input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
it.
Indices can be obtained using [`AutoProcessor`]. See [`BlipProcessor.__call__`] for details.
[What are input IDs?](../glossary#input-ids)
attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
- 1 for tokens that are **not masked**,
- 0 for tokens that are **masked**.
[What are attention masks?](../glossary#attention-mask)
position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
config.max_position_embeddings - 1]`.
[What are position IDs?](../glossary#position-ids)
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
tensors for more detail.
output_hidden_states (`bool`, *optional*):
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
more detail.
return_dict (`bool`, *optional*):
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
"""
BLIP_VISION_INPUTS_DOCSTRING = r"""
Args:
pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
Pixel values. Padding will be ignored by default should you provide it. Pixel values can be obtained using
[`BlipImageProcessor`]. See [`BlipImageProcessor.__call__`] for details.
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
tensors for more detail.
output_hidden_states (`bool`, *optional*):
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
more detail.
return_dict (`bool`, *optional*):
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
interpolate_pos_encoding (`bool`, *optional*, defaults to `False`):
Whether to interpolate the pre-trained position encodings.
"""
BLIP_INPUTS_DOCSTRING = r"""
Args:
input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
it.
Indices can be obtained using [`AutoProcessor`]. See [`BlipProcessor.__call__`] for details.
[What are input IDs?](../glossary#input-ids)
attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
- 1 for tokens that are **not masked**,
- 0 for tokens that are **masked**.
[What are attention masks?](../glossary#attention-mask)
position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
config.max_position_embeddings - 1]`.
[What are position IDs?](../glossary#position-ids)
pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
Pixel values. Padding will be ignored by default should you provide it. Pixel values can be obtained using
[`BlipImageProcessor`]. See [`BlipImageProcessor.__call__`] for details.
return_loss (`bool`, *optional*):
Whether or not to return the contrastive loss.
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
tensors for more detail.
output_hidden_states (`bool`, *optional*):
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
more detail.
return_dict (`bool`, *optional*):
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
interpolate_pos_encoding (`bool`, *optional*, defaults to `False`):
Whether to interpolate the pre-trained position encodings.
"""
class BlipEncoder(nn.Module):
"""
Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a
[`BlipEncoderLayer`].
Args:
config (`BlipConfig`):
The corresponding vision configuration for the `BlipEncoder`.
"""
def __init__(self, config: BlipConfig):
super().__init__()
self.config = config
self.layers = nn.ModuleList([BlipEncoderLayer(config) for _ in range(config.num_hidden_layers)])
self.gradient_checkpointing = False
def forward(
self,
inputs_embeds,
attention_mask: Optional[torch.Tensor] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
) -> Union[Tuple, BaseModelOutput]:
r"""
Args:
inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
Embedded representation of the inputs. Should be float, not int tokens.
attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
- 1 for tokens that are **not masked**,
- 0 for tokens that are **masked**.
[What are attention masks?](../glossary#attention-mask)
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
returned tensors for more detail.
output_hidden_states (`bool`, *optional*):
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
for more detail.
return_dict (`bool`, *optional*):
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
"""
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
encoder_states = () if output_hidden_states else None
all_attentions = () if output_attentions else None
hidden_states = inputs_embeds
for idx, encoder_layer in enumerate(self.layers):
if output_hidden_states:
encoder_states = encoder_states + (hidden_states,)
if self.gradient_checkpointing and self.training:
layer_outputs = self._gradient_checkpointing_func(
encoder_layer.__call__,
hidden_states,
attention_mask,
output_attentions,
)
else:
layer_outputs = encoder_layer(
hidden_states,
attention_mask,
output_attentions=output_attentions,
)
hidden_states = layer_outputs[0]
if output_attentions:
all_attentions = all_attentions + (layer_outputs[1],)
if output_hidden_states:
encoder_states = encoder_states + (hidden_states,)
if not return_dict:
return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None)
return BaseModelOutput(
last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions
)
class BlipVisionModel(BlipPreTrainedModel):
main_input_name = "pixel_values"
config_class = BlipVisionConfig
def __init__(self, config: BlipVisionConfig):
super().__init__(config)
self.config = config
embed_dim = config.hidden_size
self.embeddings = BlipVisionEmbeddings(config)
self.encoder = BlipEncoder(config)
self.post_layernorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
self.post_init()
@add_start_docstrings_to_model_forward(BLIP_VISION_INPUTS_DOCSTRING)
@replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=BlipVisionConfig)
def forward(
self,
pixel_values: Optional[torch.FloatTensor] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
interpolate_pos_encoding: bool = False,
) -> Union[Tuple, BaseModelOutputWithPooling]:
r"""
Returns:
"""
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 pixel_values is None:
raise ValueError("You have to specify pixel_values")
hidden_states = self.embeddings(pixel_values, interpolate_pos_encoding=interpolate_pos_encoding)
encoder_outputs = self.encoder(
inputs_embeds=hidden_states,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
last_hidden_state = encoder_outputs[0]
last_hidden_state = self.post_layernorm(last_hidden_state)
pooled_output = last_hidden_state[:, 0, :]
pooled_output = self.post_layernorm(pooled_output)
if not return_dict:
return (last_hidden_state, pooled_output) + encoder_outputs[1:]
return BaseModelOutputWithPooling(
last_hidden_state=last_hidden_state,
pooler_output=pooled_output,
hidden_states=encoder_outputs.hidden_states,
attentions=encoder_outputs.attentions,
)
def get_input_embeddings(self):
return self.embeddings
@add_start_docstrings(
"""
This model is going to be deprecated in future versions. Please use `BlipForConditionalGeneration`, `BlipForQuestionAnswering` or `BlipForImageTextRetrieval` depending on your usecase.
""",
BLIP_START_DOCSTRING,
)
class BlipModel(BlipPreTrainedModel):
config_class = BlipConfig
def __init__(self, config: BlipConfig):
super().__init__(config)
if not isinstance(config.text_config, BlipTextConfig):
raise TypeError(
"config.text_config is expected to be of type BlipTextConfig but is of type"
f" {type(config.text_config)}."
)
if not isinstance(config.vision_config, BlipVisionConfig):
raise TypeError(
"config.vision_config is expected to be of type BlipVisionConfig but is of type"
f" {type(config.vision_config)}."
)
text_config = config.text_config
vision_config = config.vision_config
self.projection_dim = config.projection_dim
self.text_embed_dim = text_config.hidden_size
self.vision_embed_dim = vision_config.hidden_size
self.text_model = BlipTextModel(text_config)
self.vision_model = BlipVisionModel(vision_config)
self.visual_projection = nn.Linear(self.vision_embed_dim, self.projection_dim, bias=False)
self.text_projection = nn.Linear(self.text_embed_dim, self.projection_dim, bias=False)
self.logit_scale = nn.Parameter(torch.tensor(self.config.logit_scale_init_value))
logger.warning(
"`BlipModel` is going to be deprecated in future release, please use `BlipForConditionalGeneration`, `BlipForQuestionAnswering` or `BlipForImageTextRetrieval` depending on your usecase."
)
# Initialize weights and apply final processing
self.post_init()
@add_start_docstrings_to_model_forward(BLIP_TEXT_INPUTS_DOCSTRING)
def get_text_features(
self,
input_ids: Optional[torch.Tensor] = None,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.Tensor] = None,
return_dict: Optional[bool] = None,
) -> torch.FloatTensor:
r"""
Returns:
text_features (`torch.FloatTensor` of shape `(batch_size, output_dim`): The text embeddings obtained by
applying the projection layer to the pooled output of [`BlipTextModel`].
Examples:
```python
>>> from transformers import AutoProcessor, BlipModel
>>> model = BlipModel.from_pretrained("Salesforce/blip-image-captioning-base")
>>> processor = AutoProcessor.from_pretrained("Salesforce/blip-image-captioning-base")
>>> inputs = processor(text=["a photo of a cat", "a photo of a dog"], padding=True, return_tensors="pt")
>>> text_features = model.get_text_features(**inputs)
```"""
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
text_outputs = self.text_model(
input_ids=input_ids,
attention_mask=attention_mask,
position_ids=position_ids,
return_dict=return_dict,
)
pooled_output = text_outputs[1]
text_features = self.text_projection(pooled_output)
return text_features
@add_start_docstrings_to_model_forward(BLIP_VISION_INPUTS_DOCSTRING)
def get_image_features(
self,
pixel_values: Optional[torch.FloatTensor] = None,
return_dict: Optional[bool] = None,
interpolate_pos_encoding: bool = False,
) -> torch.FloatTensor:
r"""
Returns:
image_features (`torch.FloatTensor` of shape `(batch_size, output_dim`): The image embeddings obtained by
applying the projection layer to the pooled output of [`BlipVisionModel`].
Examples:
```python
>>> from PIL import Image
>>> import requests
>>> from transformers import AutoProcessor, BlipModel
>>> model = BlipModel.from_pretrained("Salesforce/blip-image-captioning-base")
>>> processor = AutoProcessor.from_pretrained("Salesforce/blip-image-captioning-base")
>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
>>> image = Image.open(requests.get(url, stream=True).raw)
>>> inputs = processor(images=image, return_tensors="pt")
>>> image_features = model.get_image_features(**inputs)
```"""
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
vision_outputs = self.vision_model(
pixel_values=pixel_values,
return_dict=return_dict,
interpolate_pos_encoding=interpolate_pos_encoding,
)
pooled_output = vision_outputs[1] # pooled_output
image_features = self.visual_projection(pooled_output)
return image_features
@add_start_docstrings_to_model_forward(BLIP_INPUTS_DOCSTRING)
def get_multimodal_features(
self,
input_ids: Optional[torch.LongTensor] = None,
pixel_values: Optional[torch.FloatTensor] = None,
attention_mask: Optional[torch.Tensor] = None,
return_dict: Optional[bool] = None,
interpolate_pos_encoding: bool = False,
) -> torch.FloatTensor:
r"""
Returns:
multimodal_features (`torch.FloatTensor` of shape `(batch_size, output_dim`): The multimodal embeddings
obtained by applying the image embeddings to the text encoder using the cross-attention mechanism.
Examples:
```python
>>> from PIL import Image
>>> import requests
>>> from transformers import AutoProcessor, BlipModel
>>> model = BlipModel.from_pretrained("Salesforce/blip-image-captioning-base")
>>> processor = AutoProcessor.from_pretrained("Salesforce/blip-image-captioning-base")
>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
>>> image = Image.open(requests.get(url, stream=True).raw)
>>> texts = ["a photo of a cat", "a photo of a dog"]
>>> inputs = processor(images=image, text=texts, padding=True, return_tensors="pt")
>>> multimodal_features = model.get_multimodal_features(**inputs)
```"""
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
vision_outputs = self.vision_model(
pixel_values=pixel_values,
output_attentions=True,
output_hidden_states=True,
return_dict=return_dict,
interpolate_pos_encoding=interpolate_pos_encoding,
)
image_embeds = vision_outputs[0]
image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long)
text_outputs = self.text_model(
input_ids=input_ids,
attention_mask=attention_mask,
encoder_hidden_states=image_embeds,
encoder_attention_mask=image_atts,
return_dict=return_dict,
)
pooled_output = text_outputs[1] # pooled_output
multimodal_features = self.text_projection(pooled_output)
return multimodal_features
@add_start_docstrings_to_model_forward(BLIP_INPUTS_DOCSTRING)
@replace_return_docstrings(output_type=BlipOutput, config_class=BlipConfig)
def forward(
self,
input_ids: Optional[torch.LongTensor] = None,
pixel_values: Optional[torch.FloatTensor] = None,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.LongTensor] = None,
return_loss: Optional[bool] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
interpolate_pos_encoding: bool = False,
) -> Union[Tuple, BlipOutput]:
r"""
Returns:
Examples:
```python
>>> from PIL import Image
>>> import requests
>>> from transformers import AutoProcessor, BlipModel
>>> model = BlipModel.from_pretrained("Salesforce/blip-image-captioning-base")
>>> processor = AutoProcessor.from_pretrained("Salesforce/blip-image-captioning-base")
>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
>>> image = Image.open(requests.get(url, stream=True).raw)
>>> inputs = processor(
... text=["a photo of a cat", "a photo of a dog"], images=image, return_tensors="pt", padding=True
... )
>>> outputs = model(**inputs)
>>> logits_per_image = outputs.logits_per_image # this is the image-text similarity score
>>> probs = logits_per_image.softmax(dim=1) # we can take the softmax to get the label probabilities
```"""
# Use BLIP model's config for some fields (if specified) instead of those of vision & text components.
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
vision_outputs = self.vision_model(
pixel_values=pixel_values,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
interpolate_pos_encoding=interpolate_pos_encoding,
)
text_outputs = self.text_model(
input_ids=input_ids,
attention_mask=attention_mask,
position_ids=position_ids,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
image_embeds = vision_outputs[1]
image_embeds = self.visual_projection(image_embeds)
text_embeds = text_outputs[1]
text_embeds = self.text_projection(text_embeds)
# normalized features
image_embeds = image_embeds / image_embeds.norm(p=2, dim=-1, keepdim=True)
text_embeds = text_embeds / text_embeds.norm(p=2, dim=-1, keepdim=True)
# cosine similarity as logits
logit_scale = self.logit_scale.exp()
logits_per_text = torch.matmul(text_embeds, image_embeds.t()) * logit_scale
logits_per_image = logits_per_text.t()
loss = None
if return_loss:
loss = blip_loss(logits_per_text)
if not return_dict:
output = (logits_per_image, logits_per_text, text_embeds, image_embeds, text_outputs, vision_outputs)
return ((loss,) + output) if loss is not None else output
return BlipOutput(
loss=loss,
logits_per_image=logits_per_image,
logits_per_text=logits_per_text,
text_embeds=text_embeds,
image_embeds=image_embeds,
text_model_output=text_outputs,
vision_model_output=vision_outputs,
)
@add_start_docstrings(
"""
BLIP Model for image captioning. The model consists of a vision encoder and a text decoder. One can optionally pass
`input_ids` to the model, which serve as a text prompt, to make the text decoder continue the prompt. Otherwise,
the decoder starts generating text from the [BOS] (beginning-of-sequence) token. will start generating the caption
from the text input. If no text input is provided, the decoder will start with the [BOS] token only.
""",
BLIP_START_DOCSTRING,
)
class BlipForConditionalGeneration(BlipPreTrainedModel):
config_class = BlipConfig
_tied_weights_keys = ["text_decoder.cls.predictions.decoder.bias"]
main_input_name = "pixel_values"
def __init__(self, config: BlipConfig):
super().__init__(config)
self.vision_model = BlipVisionModel(config.vision_config)
self.text_decoder = BlipTextLMHeadModel(config.text_config)
self.decoder_input_ids = config.text_config.bos_token_id
self.decoder_pad_token_id = config.text_config.pad_token_id
# Initialize weights and apply final processing
self.post_init()
def get_input_embeddings(self) -> nn.Module:
return self.vision_model.embeddings.patch_embedding
@add_start_docstrings_to_model_forward(BLIP_VISION_INPUTS_DOCSTRING)
@replace_return_docstrings(output_type=BlipForConditionalGenerationModelOutput, config_class=BlipVisionConfig)
def forward(
self,
pixel_values: torch.FloatTensor,
input_ids: Optional[torch.LongTensor] = None,
attention_mask: Optional[torch.LongTensor] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
labels: Optional[torch.LongTensor] = None,
return_dict: Optional[bool] = None,
interpolate_pos_encoding: bool = False,
) -> Union[Tuple, BlipForConditionalGenerationModelOutput]:
r"""
Returns:
Examples:
```python
>>> from PIL import Image
>>> import requests
>>> from transformers import AutoProcessor, BlipForConditionalGeneration
>>> processor = AutoProcessor.from_pretrained("Salesforce/blip-image-captioning-base")
>>> model = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-base")
>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
>>> image = Image.open(requests.get(url, stream=True).raw)
>>> text = "A picture of"
>>> inputs = processor(images=image, text=text, return_tensors="pt")
>>> outputs = model(**inputs)
```"""
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
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
)
vision_outputs = self.vision_model(
pixel_values=pixel_values,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
interpolate_pos_encoding=interpolate_pos_encoding,
)
image_embeds = vision_outputs[0]
outputs = self.text_decoder(
input_ids=input_ids,
attention_mask=attention_mask,
encoder_hidden_states=image_embeds,
labels=labels,
return_dict=return_dict,
reduction="mean",
)
if not return_dict:
outputs = (outputs[0], outputs[1], image_embeds, vision_outputs[0]) + vision_outputs[2:]
return tuple(output for output in outputs if output is not None)
return BlipForConditionalGenerationModelOutput(
loss=outputs.loss,
logits=outputs.logits,
image_embeds=image_embeds,
last_hidden_state=vision_outputs.last_hidden_state,
hidden_states=vision_outputs.hidden_states,
attentions=vision_outputs.attentions,
)
@torch.no_grad()
def generate(
self,
pixel_values: torch.FloatTensor,
input_ids: Optional[torch.LongTensor] = None,
attention_mask: Optional[torch.LongTensor] = None,
interpolate_pos_encoding: bool = False,
**generate_kwargs,
) -> torch.LongTensor:
r"""
Overrides *generate* function to be able to use the model as a conditional generator
Parameters:
pixel_values (*torch.FloatTensor* of shape *(batch_size, num_channels, image_height, image_width)*:
Input image to be processed
input_ids (*torch.LongTensor* of shape *(batch_size, sequence_length)*, *optional*):
The sequence used as a prompt for the generation.
attention_mask (*torch.LongTensor* of shape *(batch_size, sequence_length)*, *optional*):
Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
Examples:
```python
>>> from PIL import Image
>>> import requests
>>> from transformers import AutoProcessor, BlipForConditionalGeneration
>>> model = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-base")
>>> processor = AutoProcessor.from_pretrained("Salesforce/blip-image-captioning-base")
>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
>>> image = Image.open(requests.get(url, stream=True).raw)
>>> inputs = processor(images=image, return_tensors="pt")
>>> outputs = model.generate(**inputs)
>>> print(processor.decode(outputs[0], skip_special_tokens=True))
two cats sleeping on a couch
```
"""
batch_size = pixel_values.shape[0]
vision_outputs = self.vision_model(
pixel_values=pixel_values,
interpolate_pos_encoding=interpolate_pos_encoding,
)
image_embeds = vision_outputs[0]
image_attention_mask = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(image_embeds.device)
if isinstance(input_ids, list):
input_ids = torch.LongTensor(input_ids)
elif input_ids is None:
input_ids = (
torch.LongTensor([[self.decoder_input_ids, self.config.text_config.eos_token_id]])
.repeat(batch_size, 1)
.to(image_embeds.device)
)
input_ids[:, 0] = self.config.text_config.bos_token_id
attention_mask = attention_mask[:, :-1] if attention_mask is not None else None
outputs = self.text_decoder.generate(
input_ids=input_ids[:, :-1],
eos_token_id=self.config.text_config.sep_token_id,
pad_token_id=self.config.text_config.pad_token_id,
attention_mask=attention_mask,
encoder_hidden_states=image_embeds,
encoder_attention_mask=image_attention_mask,
**generate_kwargs,
)
return outputs
@add_start_docstrings(
"""
BLIP Model for visual question answering. The model consists of a vision encoder, a text encoder as well as a text
decoder. The vision encoder will encode the input image, the text encoder will encode the input question together
with the encoding of the image, and the text decoder will output the answer to the question.
""",
BLIP_START_DOCSTRING,
)
class BlipForQuestionAnswering(BlipPreTrainedModel):
config_class = BlipConfig
_tied_weights_keys = ["text_decoder.cls.predictions.decoder.bias"]
def __init__(self, config: BlipConfig):
super().__init__(config)
self.vision_model = BlipVisionModel(config.vision_config)
self.text_encoder = BlipTextModel(config.text_config, add_pooling_layer=False)
self.text_decoder = BlipTextLMHeadModel(config.text_config)
self.decoder_pad_token_id = config.text_config.pad_token_id
self.decoder_start_token_id = config.text_config.bos_token_id
# Initialize weights and apply final processing
self.post_init()
def get_input_embeddings(self) -> nn.Module:
return self.vision_model.embeddings.patch_embedding
@add_start_docstrings_to_model_forward(BLIP_VISION_INPUTS_DOCSTRING)
@replace_return_docstrings(output_type=BlipTextVisionModelOutput, config_class=BlipVisionConfig)
def forward(
self,
input_ids: torch.LongTensor,
pixel_values: torch.FloatTensor,
decoder_input_ids: Optional[torch.LongTensor] = None,
decoder_attention_mask: Optional[torch.LongTensor] = None,
attention_mask: Optional[torch.LongTensor] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
labels: Optional[torch.LongTensor] = None,
return_dict: Optional[bool] = None,
interpolate_pos_encoding: bool = False,
) -> Union[Tuple, BlipTextVisionModelOutput]:
r"""
Returns:
Examples:
```python
>>> from PIL import Image
>>> import requests
>>> from transformers import AutoProcessor, BlipForQuestionAnswering
>>> model = BlipForQuestionAnswering.from_pretrained("Salesforce/blip-vqa-base")
>>> processor = AutoProcessor.from_pretrained("Salesforce/blip-vqa-base")
>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
>>> image = Image.open(requests.get(url, stream=True).raw)
>>> # training
>>> text = "How many cats are in the picture?"
>>> label = "2"
>>> inputs = processor(images=image, text=text, return_tensors="pt")
>>> labels = processor(text=label, return_tensors="pt").input_ids
>>> inputs["labels"] = labels
>>> outputs = model(**inputs)
>>> loss = outputs.loss
>>> loss.backward()
>>> # inference
>>> text = "How many cats are in the picture?"
>>> inputs = processor(images=image, text=text, return_tensors="pt")
>>> outputs = model.generate(**inputs)
>>> print(processor.decode(outputs[0], skip_special_tokens=True))
2
```"""
if labels is None and decoder_input_ids is None:
raise ValueError(
"Either `decoder_input_ids` or `labels` should be passed when calling `forward` with"
" `BlipForQuestionAnswering`. if you are training the model make sure that `labels` is passed, if you"
" are using the model for inference make sure that `decoder_input_ids` is passed or call `generate`"
)
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
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
)
vision_outputs = self.vision_model(
pixel_values=pixel_values,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
interpolate_pos_encoding=interpolate_pos_encoding,
)
image_embeds = vision_outputs[0]
image_attention_mask = torch.ones(image_embeds.size()[:-1], dtype=torch.long)
question_embeds = self.text_encoder(
input_ids=input_ids,
attention_mask=attention_mask,
encoder_hidden_states=image_embeds,
encoder_attention_mask=image_attention_mask,
return_dict=return_dict,
)
if labels is not None and decoder_input_ids is None:
# labels are already shifted right, see: https://github.com/huggingface/transformers/pull/23153
decoder_input_ids = labels
question_embeds = question_embeds[0] if not return_dict else question_embeds.last_hidden_state
answer_output = self.text_decoder(
input_ids=decoder_input_ids,
attention_mask=decoder_attention_mask,
encoder_hidden_states=question_embeds,
encoder_attention_mask=attention_mask,
labels=labels,
return_dict=return_dict,
reduction="mean",
)
if labels is not None:
decoder_loss = answer_output.loss.mean() if return_dict else answer_output[0].mean()
else:
decoder_loss = None
if not return_dict:
outputs = (decoder_loss, image_embeds, vision_outputs[0]) + vision_outputs[2:]
return tuple(output for output in outputs if output is not None)
return BlipTextVisionModelOutput(
loss=decoder_loss,
image_embeds=image_embeds,
last_hidden_state=vision_outputs.last_hidden_state,
hidden_states=vision_outputs.hidden_states,
attentions=vision_outputs.attentions,
)
@torch.no_grad()
def generate(
self,
input_ids: torch.LongTensor,
pixel_values: torch.FloatTensor,
attention_mask: Optional[torch.LongTensor] = None,
interpolate_pos_encoding: bool = False,
**generate_kwargs,
) -> torch.LongTensor:
r"""
Overrides *generate* function to be able to use the model as a conditional generator
Parameters:
input_ids (*torch.LongTensor* of shape *(batch_size, sequence_length)*):
The sequence used as a prompt for the generation.
pixel_values (*torch.FloatTensor* of shape *(batch_size, num_channels, image_height, image_width)*:
Input image to be processed
attention_mask (*torch.LongTensor* of shape *(batch_size, sequence_length)*, *optional*):
Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`. `1` for
tokens that are NOT MASKED, `0` for MASKED tokens.
**generate_kwargs:
Additional arguments passed to the *generate* function of the decoder
Examples:
```python
>>> from PIL import Image
>>> import requests
>>> from transformers import AutoProcessor, BlipForQuestionAnswering
>>> model = BlipForQuestionAnswering.from_pretrained("Salesforce/blip-vqa-base")
>>> processor = AutoProcessor.from_pretrained("Salesforce/blip-vqa-base")
>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
>>> image = Image.open(requests.get(url, stream=True).raw)
>>> text = "How many cats are in the picture?"
>>> inputs = processor(images=image, text=text, return_tensors="pt")
>>> outputs = model.generate(**inputs)
>>> print(processor.decode(outputs[0], skip_special_tokens=True))
2
```
"""
vision_outputs = self.vision_model(
pixel_values=pixel_values,
interpolate_pos_encoding=interpolate_pos_encoding,
)
image_embeds = vision_outputs[0]
image_attention_mask = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(image_embeds.device)
if isinstance(input_ids, list):
input_ids = torch.LongTensor(input_ids)
question_outputs = self.text_encoder(
input_ids=input_ids,
attention_mask=attention_mask,
encoder_hidden_states=image_embeds,
encoder_attention_mask=image_attention_mask,
return_dict=False,
)
question_embeds = question_outputs[0]
question_attention_mask = torch.ones(question_embeds.size()[:-1], dtype=torch.long).to(question_embeds.device)
bos_ids = torch.full(
(question_embeds.size(0), 1), fill_value=self.decoder_start_token_id, device=question_embeds.device
)
outputs = self.text_decoder.generate(
input_ids=bos_ids,
eos_token_id=self.config.text_config.sep_token_id,
pad_token_id=self.config.text_config.pad_token_id,
encoder_hidden_states=question_embeds,
encoder_attention_mask=question_attention_mask,
**generate_kwargs,
)
return outputs
@add_start_docstrings(
"""
BLIP Model with a vision and text projector, and a classification head on top. The model is used in the context of
image-text retrieval. Given an image and a text, the model returns the probability of the text being relevant to
the image.
""",
BLIP_START_DOCSTRING,
)
class BlipForImageTextRetrieval(BlipPreTrainedModel):
config_class = BlipConfig
def __init__(self, config: BlipConfig):
super().__init__(config)
self.vision_model = BlipVisionModel(config.vision_config)
self.text_encoder = BlipTextModel(config.text_config, add_pooling_layer=False)
# vision projection layer
self.vision_proj = nn.Linear(config.vision_config.hidden_size, config.image_text_hidden_size)
# text projection layer
self.text_proj = nn.Linear(config.text_config.hidden_size, config.image_text_hidden_size)
# image text matching head
self.itm_head = nn.Linear(config.text_config.hidden_size, 2)
self.decoder_pad_token_id = (
config.text_config.pad_token_id
if not hasattr(config, "decoder_pad_token_id")
else config.decoder_pad_token_id
)
self.decoder_start_token_id = (
config.text_config.bos_token_id
if not hasattr(config, "decoder_start_token_id")
else config.decoder_start_token_id
)
# Initialize weights and apply final processing
self.post_init()
def get_input_embeddings(self) -> nn.Module:
return self.vision_model.embeddings.patch_embedding
@add_start_docstrings_to_model_forward(BLIP_VISION_INPUTS_DOCSTRING)
@replace_return_docstrings(output_type=BlipTextVisionModelOutput, config_class=BlipVisionConfig)
def forward(
self,
input_ids: torch.LongTensor,
pixel_values: torch.FloatTensor,
use_itm_head: Optional[bool] = True,
attention_mask: Optional[torch.LongTensor] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
interpolate_pos_encoding: bool = False,
) -> Union[Tuple, BlipTextVisionModelOutput]:
r"""
Returns:
Examples:
```python
>>> from PIL import Image
>>> import requests
>>> from transformers import AutoProcessor, BlipForImageTextRetrieval
>>> model = BlipForImageTextRetrieval.from_pretrained("Salesforce/blip-itm-base-coco")
>>> processor = AutoProcessor.from_pretrained("Salesforce/blip-itm-base-coco")
>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
>>> image = Image.open(requests.get(url, stream=True).raw)
>>> text = "an image of a cat"
>>> inputs = processor(images=image, text=text, return_tensors="pt")
>>> outputs = model(**inputs)
```
"""
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
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
)
vision_outputs = self.vision_model(
pixel_values=pixel_values,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
interpolate_pos_encoding=interpolate_pos_encoding,
)
image_embeds = vision_outputs[0]
image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long)
if use_itm_head:
question_embeds = self.text_encoder(
input_ids=input_ids,
attention_mask=attention_mask,
encoder_hidden_states=image_embeds,
encoder_attention_mask=image_atts,
return_dict=return_dict,
)
question_embeds = question_embeds[0] if not return_dict else question_embeds.last_hidden_state
output = self.itm_head(question_embeds[:, 0, :])
else:
question_embeds = self.text_encoder(
input_ids=input_ids,
attention_mask=attention_mask,
return_dict=return_dict,
)
question_embeds = question_embeds[0] if not return_dict else question_embeds.last_hidden_state
image_feat = normalize(self.vision_proj(image_embeds[:, 0, :]), dim=-1)
text_feat = normalize(self.text_proj(question_embeds[:, 0, :]), dim=-1)
output = image_feat @ text_feat.t()
if not return_dict:
outputs = (output, vision_outputs[0]) + vision_outputs[2:] + (question_embeds,)
return tuple(output for output in outputs if output is not None)
return BlipImageTextMatchingModelOutput(
itm_score=output,
last_hidden_state=vision_outputs.last_hidden_state,
hidden_states=vision_outputs.hidden_states,
attentions=vision_outputs.attentions,
question_embeds=question_embeds,
)
|
transformers/src/transformers/models/blip/modeling_blip.py/0
|
{
"file_path": "transformers/src/transformers/models/blip/modeling_blip.py",
"repo_id": "transformers",
"token_count": 27796
}
| 345
|
# Copyright 2023 The Intel Labs Team Authors, The Microsoft Research Team Authors and HuggingFace Inc. team. 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.
from typing import TYPE_CHECKING
from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_torch_available, is_vision_available
_import_structure = {
"configuration_bridgetower": [
"BridgeTowerConfig",
"BridgeTowerTextConfig",
"BridgeTowerVisionConfig",
],
"processing_bridgetower": ["BridgeTowerProcessor"],
}
try:
if not is_vision_available():
raise OptionalDependencyNotAvailable()
except OptionalDependencyNotAvailable:
pass
else:
_import_structure["image_processing_bridgetower"] = ["BridgeTowerImageProcessor"]
try:
if not is_torch_available():
raise OptionalDependencyNotAvailable()
except OptionalDependencyNotAvailable:
pass
else:
_import_structure["modeling_bridgetower"] = [
"BridgeTowerForContrastiveLearning",
"BridgeTowerForImageAndTextRetrieval",
"BridgeTowerForMaskedLM",
"BridgeTowerModel",
"BridgeTowerPreTrainedModel",
]
if TYPE_CHECKING:
from .configuration_bridgetower import (
BridgeTowerConfig,
BridgeTowerTextConfig,
BridgeTowerVisionConfig,
)
from .processing_bridgetower import BridgeTowerProcessor
try:
if not is_vision_available():
raise OptionalDependencyNotAvailable()
except OptionalDependencyNotAvailable:
pass
else:
from .image_processing_bridgetower import BridgeTowerImageProcessor
try:
if not is_torch_available():
raise OptionalDependencyNotAvailable()
except OptionalDependencyNotAvailable:
pass
else:
from .modeling_bridgetower import (
BridgeTowerForContrastiveLearning,
BridgeTowerForImageAndTextRetrieval,
BridgeTowerForMaskedLM,
BridgeTowerModel,
BridgeTowerPreTrainedModel,
)
else:
import sys
sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure)
|
transformers/src/transformers/models/bridgetower/__init__.py/0
|
{
"file_path": "transformers/src/transformers/models/bridgetower/__init__.py",
"repo_id": "transformers",
"token_count": 957
}
| 346
|
# coding=utf-8
# Copyright 2022 The OFA-Sys Team Authors and The HuggingFace Team. 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.
import argparse
import torch
from transformers import ChineseCLIPConfig, ChineseCLIPModel
def copy_attn_layer(hf_attn_layer, pt_weights, prefix):
q_proj, k_proj, v_proj = pt_weights[f"{prefix}.in_proj_weight"].chunk(3, dim=0)
q_proj_bias, k_proj_bias, v_proj_bias = pt_weights[f"{prefix}.in_proj_bias"].chunk(3, dim=0)
out_proj_weights = pt_weights[f"{prefix}.out_proj.weight"]
out_proj_bias = pt_weights[f"{prefix}.out_proj.bias"]
hf_attn_layer.q_proj.weight.data = q_proj
hf_attn_layer.q_proj.bias.data = q_proj_bias
hf_attn_layer.k_proj.weight.data = k_proj
hf_attn_layer.k_proj.bias.data = k_proj_bias
hf_attn_layer.v_proj.weight.data = v_proj
hf_attn_layer.v_proj.bias.data = v_proj_bias
hf_attn_layer.out_proj.weight.data = out_proj_weights
hf_attn_layer.out_proj.bias.data = out_proj_bias
def copy_mlp(hf_mlp, pt_weights, prefix):
copy_linear(hf_mlp.fc1, pt_weights, f"{prefix}.c_fc")
copy_linear(hf_mlp.fc2, pt_weights, f"{prefix}.c_proj")
def copy_linear(hf_linear, pt_weights, prefix):
hf_linear.weight.data = pt_weights[f"{prefix}.weight"].data
hf_linear.bias.data = pt_weights[f"{prefix}.bias"].data
def copy_layer(hf_layer, pt_weights, prefix):
# copy layer norms
copy_linear(hf_layer.layer_norm1, pt_weights, f"{prefix}.ln_1")
copy_linear(hf_layer.layer_norm2, pt_weights, f"{prefix}.ln_2")
# copy MLP
copy_mlp(hf_layer.mlp, pt_weights, f"{prefix}.mlp")
# copy attn
copy_attn_layer(hf_layer.self_attn, pt_weights, f"{prefix}.attn")
def copy_layers(hf_layers, pt_weights, prefix):
for layer_id, hf_layer in enumerate(hf_layers):
copy_layer(hf_layer, pt_weights, f"{prefix}.{layer_id}")
def copy_text_model_and_projection(hf_model, pt_weights):
# copy projection
hf_model.text_projection.weight.data = pt_weights["text_projection"].data.T
# copy text encoder
for name, param in hf_model.text_model.named_parameters():
param.data = pt_weights[f"bert.{name}"].data
def copy_vision_model_and_projection(hf_model, pt_weights):
# copy projection
hf_model.visual_projection.weight.data = pt_weights["visual.proj"].data.T
# copy layer norms
copy_linear(hf_model.vision_model.pre_layrnorm, pt_weights, "visual.ln_pre")
copy_linear(hf_model.vision_model.post_layernorm, pt_weights, "visual.ln_post")
# copy embeddings
hf_model.vision_model.embeddings.patch_embedding.weight.data = pt_weights["visual.conv1.weight"].data
hf_model.vision_model.embeddings.class_embedding.data = pt_weights["visual.class_embedding"].data
hf_model.vision_model.embeddings.position_embedding.weight.data = pt_weights["visual.positional_embedding"].data
# copy encoder
copy_layers(hf_model.vision_model.encoder.layers, pt_weights, "visual.transformer.resblocks")
@torch.no_grad()
def convert_chinese_clip_checkpoint(checkpoint_path, pytorch_dump_folder_path, config_path=None):
"""
Copy/paste/tweak model's weights to transformers design.
"""
assert config_path is not None, "Please specify the ChineseCLIP model config of the corresponding model size."
config = ChineseCLIPConfig.from_pretrained(config_path)
hf_model = ChineseCLIPModel(config).eval()
pt_weights = torch.load(checkpoint_path, map_location="cpu")["state_dict"]
pt_weights = {(name[7:] if name.startswith("module.") else name): value for name, value in pt_weights.items()}
copy_text_model_and_projection(hf_model, pt_weights)
copy_vision_model_and_projection(hf_model, pt_weights)
hf_model.logit_scale.data = pt_weights["logit_scale"].data
hf_model.save_pretrained(pytorch_dump_folder_path)
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument(
"--pytorch_dump_folder_path",
default=None,
type=str,
help="Path to the output folder storing converted hf PyTorch model.",
)
parser.add_argument(
"--checkpoint_path", default=None, type=str, help="Path to original github format ChineseCLIP checkpoint."
)
parser.add_argument(
"--config_path", default=None, required=True, type=str, help="Path to hf config.json of model to convert."
)
args = parser.parse_args()
convert_chinese_clip_checkpoint(args.checkpoint_path, args.pytorch_dump_folder_path, args.config_path)
print("The conversion is finished!")
|
transformers/src/transformers/models/chinese_clip/convert_chinese_clip_original_pytorch_to_hf.py/0
|
{
"file_path": "transformers/src/transformers/models/chinese_clip/convert_chinese_clip_original_pytorch_to_hf.py",
"repo_id": "transformers",
"token_count": 1990
}
| 347
|
# coding=utf-8
# Copyright 2021 The OpenAI Team Authors and The HuggingFace Team. 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.
"""PyTorch CLIP model."""
from dataclasses import dataclass
from typing import Any, Optional, Tuple, Union
import torch
import torch.utils.checkpoint
from torch import nn
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
from ...activations import ACT2FN
from ...modeling_attn_mask_utils import _create_4d_causal_attention_mask, _prepare_4d_attention_mask
from ...modeling_outputs import BaseModelOutput, BaseModelOutputWithPooling, ImageClassifierOutput
from ...modeling_utils import PreTrainedModel
from ...pytorch_utils import is_torch_greater_or_equal_than_2_2
from ...utils import (
ModelOutput,
add_code_sample_docstrings,
add_start_docstrings,
add_start_docstrings_to_model_forward,
is_flash_attn_2_available,
is_flash_attn_greater_or_equal_2_10,
logging,
replace_return_docstrings,
)
from .configuration_clip import CLIPConfig, CLIPTextConfig, CLIPVisionConfig
if is_flash_attn_2_available():
from ...modeling_flash_attention_utils import _flash_attention_forward
logger = logging.get_logger(__name__)
# General docstring
_CONFIG_FOR_DOC = "CLIPConfig"
_CHECKPOINT_FOR_DOC = "openai/clip-vit-base-patch32"
# Image classification docstring
_IMAGE_CLASS_CHECKPOINT = "openai/clip-vit-base-patch32"
_IMAGE_CLASS_EXPECTED_OUTPUT = "LABEL_0"
# contrastive loss function, adapted from
# https://sachinruk.github.io/blog/2021-03-07-clip.html
def contrastive_loss(logits: torch.Tensor) -> torch.Tensor:
return nn.functional.cross_entropy(logits, torch.arange(len(logits), device=logits.device))
def clip_loss(similarity: torch.Tensor) -> torch.Tensor:
caption_loss = contrastive_loss(similarity)
image_loss = contrastive_loss(similarity.t())
return (caption_loss + image_loss) / 2.0
def _get_vector_norm(tensor: torch.Tensor) -> torch.Tensor:
"""
This method is equivalent to tensor.norm(p=2, dim=-1, keepdim=True) and used to make
model `executorch` exportable. See issue https://github.com/pytorch/executorch/issues/3566
"""
square_tensor = torch.pow(tensor, 2)
sum_tensor = torch.sum(square_tensor, dim=-1, keepdim=True)
normed_tensor = torch.pow(sum_tensor, 0.5)
return normed_tensor
@dataclass
class CLIPVisionModelOutput(ModelOutput):
"""
Base class for vision model's outputs that also contains image embeddings of the pooling of the last hidden states.
Args:
image_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim)` *optional* returned when model is initialized with `with_projection=True`):
The image embeddings obtained by applying the projection layer to the pooler_output.
last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
Sequence of hidden-states at the output of the last layer of the model.
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
"""
image_embeds: Optional[torch.FloatTensor] = None
last_hidden_state: torch.FloatTensor = None
hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
@dataclass
class CLIPTextModelOutput(ModelOutput):
"""
Base class for text model's outputs that also contains a pooling of the last hidden states.
Args:
text_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim)` *optional* returned when model is initialized with `with_projection=True`):
The text embeddings obtained by applying the projection layer to the pooler_output.
last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
Sequence of hidden-states at the output of the last layer of the model.
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
"""
text_embeds: Optional[torch.FloatTensor] = None
last_hidden_state: torch.FloatTensor = None
hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
@dataclass
class CLIPOutput(ModelOutput):
"""
Args:
loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `return_loss` is `True`):
Contrastive loss for image-text similarity.
logits_per_image (`torch.FloatTensor` of shape `(image_batch_size, text_batch_size)`):
The scaled dot product scores between `image_embeds` and `text_embeds`. This represents the image-text
similarity scores.
logits_per_text (`torch.FloatTensor` of shape `(text_batch_size, image_batch_size)`):
The scaled dot product scores between `text_embeds` and `image_embeds`. This represents the text-image
similarity scores.
text_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim`):
The text embeddings obtained by applying the projection layer to the pooled output of [`CLIPTextModel`].
image_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim`):
The image embeddings obtained by applying the projection layer to the pooled output of [`CLIPVisionModel`].
text_model_output (`BaseModelOutputWithPooling`):
The output of the [`CLIPTextModel`].
vision_model_output (`BaseModelOutputWithPooling`):
The output of the [`CLIPVisionModel`].
"""
loss: Optional[torch.FloatTensor] = None
logits_per_image: torch.FloatTensor = None
logits_per_text: torch.FloatTensor = None
text_embeds: torch.FloatTensor = None
image_embeds: torch.FloatTensor = None
text_model_output: BaseModelOutputWithPooling = None
vision_model_output: BaseModelOutputWithPooling = None
def to_tuple(self) -> Tuple[Any]:
return tuple(
self[k] if k not in ["text_model_output", "vision_model_output"] else getattr(self, k).to_tuple()
for k in self.keys()
)
class CLIPVisionEmbeddings(nn.Module):
def __init__(self, config: CLIPVisionConfig):
super().__init__()
self.config = config
self.embed_dim = config.hidden_size
self.image_size = config.image_size
self.patch_size = config.patch_size
self.class_embedding = nn.Parameter(torch.randn(self.embed_dim))
self.patch_embedding = nn.Conv2d(
in_channels=config.num_channels,
out_channels=self.embed_dim,
kernel_size=self.patch_size,
stride=self.patch_size,
bias=False,
)
self.num_patches = (self.image_size // self.patch_size) ** 2
self.num_positions = self.num_patches + 1
self.position_embedding = nn.Embedding(self.num_positions, self.embed_dim)
self.register_buffer("position_ids", torch.arange(self.num_positions).expand((1, -1)), persistent=False)
def forward(self, pixel_values: torch.FloatTensor) -> torch.Tensor:
batch_size = pixel_values.shape[0]
target_dtype = self.patch_embedding.weight.dtype
patch_embeds = self.patch_embedding(pixel_values.to(dtype=target_dtype)) # shape = [*, width, grid, grid]
patch_embeds = patch_embeds.flatten(2).transpose(1, 2)
class_embeds = self.class_embedding.expand(batch_size, 1, -1)
embeddings = torch.cat([class_embeds, patch_embeds], dim=1)
embeddings = embeddings + self.position_embedding(self.position_ids)
return embeddings
class CLIPTextEmbeddings(nn.Module):
def __init__(self, config: CLIPTextConfig):
super().__init__()
embed_dim = config.hidden_size
self.token_embedding = nn.Embedding(config.vocab_size, embed_dim)
self.position_embedding = nn.Embedding(config.max_position_embeddings, embed_dim)
# position_ids (1, len position emb) is contiguous in memory and exported when serialized
self.register_buffer(
"position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)), persistent=False
)
def forward(
self,
input_ids: Optional[torch.LongTensor] = None,
position_ids: Optional[torch.LongTensor] = None,
inputs_embeds: Optional[torch.FloatTensor] = None,
) -> torch.Tensor:
seq_length = input_ids.shape[-1] if input_ids is not None else inputs_embeds.shape[-2]
if position_ids is None:
position_ids = self.position_ids[:, :seq_length]
if inputs_embeds is None:
inputs_embeds = self.token_embedding(input_ids)
position_embeddings = self.position_embedding(position_ids)
embeddings = inputs_embeds + position_embeddings
return embeddings
class CLIPAttention(nn.Module):
"""Multi-headed attention from 'Attention Is All You Need' paper"""
def __init__(self, config):
super().__init__()
self.config = config
self.embed_dim = config.hidden_size
self.num_heads = config.num_attention_heads
self.head_dim = self.embed_dim // self.num_heads
if self.head_dim * self.num_heads != self.embed_dim:
raise ValueError(
f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:"
f" {self.num_heads})."
)
self.scale = self.head_dim**-0.5
self.dropout = config.attention_dropout
self.k_proj = nn.Linear(self.embed_dim, self.embed_dim)
self.v_proj = nn.Linear(self.embed_dim, self.embed_dim)
self.q_proj = nn.Linear(self.embed_dim, self.embed_dim)
self.out_proj = nn.Linear(self.embed_dim, self.embed_dim)
def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
def forward(
self,
hidden_states: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
causal_attention_mask: Optional[torch.Tensor] = None,
output_attentions: Optional[bool] = False,
) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
"""Input shape: Batch x Time x Channel"""
bsz, tgt_len, embed_dim = hidden_states.size()
# get query proj
query_states = self.q_proj(hidden_states) * self.scale
key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
proj_shape = (bsz * self.num_heads, -1, self.head_dim)
query_states = self._shape(query_states, tgt_len, bsz).view(*proj_shape)
key_states = key_states.view(*proj_shape)
value_states = value_states.view(*proj_shape)
src_len = key_states.size(1)
attn_weights = torch.bmm(query_states, key_states.transpose(1, 2))
if attn_weights.size() != (bsz * self.num_heads, tgt_len, src_len):
raise ValueError(
f"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is"
f" {attn_weights.size()}"
)
# apply the causal_attention_mask first
if causal_attention_mask is not None:
if causal_attention_mask.size() != (bsz, 1, tgt_len, src_len):
raise ValueError(
f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is"
f" {causal_attention_mask.size()}"
)
attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + causal_attention_mask
attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
if attention_mask is not None:
if attention_mask.size() != (bsz, 1, tgt_len, src_len):
raise ValueError(
f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.size()}"
)
attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + attention_mask
attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
attn_weights = nn.functional.softmax(attn_weights, dim=-1)
if output_attentions:
# this operation is a bit akward, but it's required to
# make sure that attn_weights keeps its gradient.
# In order to do so, attn_weights have to reshaped
# twice and have to be reused in the following
attn_weights_reshaped = attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
attn_weights = attn_weights_reshaped.view(bsz * self.num_heads, tgt_len, src_len)
else:
attn_weights_reshaped = None
attn_probs = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)
attn_output = torch.bmm(attn_probs, value_states)
if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
raise ValueError(
f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
f" {attn_output.size()}"
)
attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
attn_output = attn_output.transpose(1, 2)
attn_output = attn_output.reshape(bsz, tgt_len, embed_dim)
attn_output = self.out_proj(attn_output)
return attn_output, attn_weights_reshaped
class CLIPFlashAttention2(CLIPAttention):
"""
CLIPAttention flash attention module. This module inherits from `CLIPAttention` as the weights of the module stays
untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
flash attention and deal with padding tokens in case the input contains any of them.
"""
# Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2.__init__
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
# TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
# flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
# Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
# Adapted from transformers.models.llama.modeling_llama.LlamaFlashAttention2.forward
def forward(
self,
hidden_states: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
causal_attention_mask: Optional[torch.Tensor] = None,
output_attentions: Optional[bool] = False,
) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
output_attentions = False
batch_size, q_len, _ = hidden_states.size()
query_states = self.q_proj(hidden_states)
key_states = self.k_proj(hidden_states)
value_states = self.v_proj(hidden_states)
# Flash attention requires the input to have the shape
# batch_size x seq_length x head_dim x hidden_dim
# therefore we just need to keep the original shape
query_states = query_states.view(batch_size, q_len, self.num_heads, self.head_dim)
key_states = key_states.view(batch_size, q_len, self.num_heads, self.head_dim)
value_states = value_states.view(batch_size, q_len, self.num_heads, self.head_dim)
dropout_rate = self.dropout if self.training else 0.0
# In PEFT, usually we cast the layer norms in float32 for training stability reasons
# therefore the input hidden states gets silently casted in float32. Hence, we need
# cast them back in the correct dtype just to be sure everything works as expected.
# This might slowdown training & inference so it is recommended to not cast the LayerNorms
# in fp32.
input_dtype = query_states.dtype
if input_dtype == torch.float32:
if torch.is_autocast_enabled():
target_dtype = torch.get_autocast_gpu_dtype()
# Handle the case where the model is quantized
elif hasattr(self.config, "_pre_quantization_dtype"):
target_dtype = self.config._pre_quantization_dtype
else:
target_dtype = self.q_proj.weight.dtype
logger.warning_once(
f"The input hidden states seems to be silently casted in float32, this might be related to"
f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
f" {target_dtype}."
)
query_states = query_states.to(target_dtype)
key_states = key_states.to(target_dtype)
value_states = value_states.to(target_dtype)
attn_output = _flash_attention_forward(
query_states,
key_states,
value_states,
attention_mask,
q_len,
dropout=dropout_rate,
is_causal=causal_attention_mask is not None,
use_top_left_mask=self._flash_attn_uses_top_left_mask,
)
attn_output = attn_output.reshape(batch_size, q_len, self.embed_dim).contiguous()
attn_output = self.out_proj(attn_output)
if not output_attentions:
attn_weights = None
return attn_output, attn_weights
class CLIPSdpaAttention(CLIPAttention):
"""
SDPA attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
`CLIPAttention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
SDPA API.
"""
# Adapted from CLIPAttention.forward
def forward(
self,
hidden_states: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
causal_attention_mask: Optional[torch.Tensor] = None,
output_attentions: Optional[bool] = False,
) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
if output_attentions:
# TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
logger.warning_once(
"CLIPModel is using CLIPSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not "
"support `output_attentions=True`. Falling back to the manual attention implementation, but specifying "
"the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can "
'be removed using the argument `attn_implementation="eager"` when loading the model.'
)
return super().forward(
hidden_states=hidden_states,
attention_mask=attention_mask,
causal_attention_mask=causal_attention_mask,
output_attentions=output_attentions,
)
# CLIP text model uses both `causal_attention_mask` and `attention_mask`
if attention_mask is not None and causal_attention_mask is not None:
attn_mask = attention_mask + causal_attention_mask
elif causal_attention_mask is not None:
attn_mask = causal_attention_mask
else:
attn_mask = attention_mask
bsz, tgt_len, embed_dim = hidden_states.size()
query_states = self.q_proj(hidden_states)
key_states = self.k_proj(hidden_states)
value_states = self.v_proj(hidden_states)
query_states = query_states.view(bsz, -1, self.num_heads, self.head_dim).transpose(1, 2)
key_states = key_states.view(bsz, -1, self.num_heads, self.head_dim).transpose(1, 2)
value_states = value_states.view(bsz, -1, self.num_heads, self.head_dim).transpose(1, 2)
# SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
# Reference: https://github.com/pytorch/pytorch/issues/112577.
if not is_torch_greater_or_equal_than_2_2 and query_states.device.type == "cuda" and attn_mask is not None:
query_states = query_states.contiguous()
key_states = key_states.contiguous()
value_states = value_states.contiguous()
# CLIP text model uses both `causal_attention_mask` and `attention_mask` sequentially.
attn_output = torch.nn.functional.scaled_dot_product_attention(
query_states,
key_states,
value_states,
attn_mask=attn_mask,
dropout_p=self.dropout if self.training else 0.0,
scale=self.scale,
)
attn_output = attn_output.transpose(1, 2)
attn_output = attn_output.reshape(bsz, tgt_len, embed_dim)
attn_output = self.out_proj(attn_output)
return attn_output, None
CLIP_ATTENTION_CLASSES = {
"eager": CLIPAttention,
"sdpa": CLIPSdpaAttention,
"flash_attention_2": CLIPFlashAttention2,
}
class CLIPMLP(nn.Module):
def __init__(self, config):
super().__init__()
self.config = config
self.activation_fn = ACT2FN[config.hidden_act]
self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
hidden_states = self.fc1(hidden_states)
hidden_states = self.activation_fn(hidden_states)
hidden_states = self.fc2(hidden_states)
return hidden_states
class CLIPEncoderLayer(nn.Module):
def __init__(self, config: CLIPConfig):
super().__init__()
self.embed_dim = config.hidden_size
self.self_attn = CLIP_ATTENTION_CLASSES[config._attn_implementation](config)
self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
self.mlp = CLIPMLP(config)
self.layer_norm2 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
def forward(
self,
hidden_states: torch.Tensor,
attention_mask: torch.Tensor,
causal_attention_mask: torch.Tensor,
output_attentions: Optional[bool] = False,
) -> Tuple[torch.FloatTensor]:
"""
Args:
hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
attention_mask (`torch.FloatTensor`): attention mask of size
`(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
`(config.encoder_attention_heads,)`.
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
returned tensors for more detail.
"""
residual = hidden_states
hidden_states = self.layer_norm1(hidden_states)
hidden_states, attn_weights = self.self_attn(
hidden_states=hidden_states,
attention_mask=attention_mask,
causal_attention_mask=causal_attention_mask,
output_attentions=output_attentions,
)
hidden_states = residual + hidden_states
residual = hidden_states
hidden_states = self.layer_norm2(hidden_states)
hidden_states = self.mlp(hidden_states)
hidden_states = residual + hidden_states
outputs = (hidden_states,)
if output_attentions:
outputs += (attn_weights,)
return outputs
class CLIPPreTrainedModel(PreTrainedModel):
"""
An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
models.
"""
config_class = CLIPConfig
base_model_prefix = "clip"
supports_gradient_checkpointing = True
_supports_sdpa = True
_supports_flash_attn_2 = True
def _init_weights(self, module):
"""Initialize the weights"""
factor = self.config.initializer_factor
if isinstance(module, CLIPTextEmbeddings):
module.token_embedding.weight.data.normal_(mean=0.0, std=factor * 0.02)
module.position_embedding.weight.data.normal_(mean=0.0, std=factor * 0.02)
elif isinstance(module, CLIPVisionEmbeddings):
factor = self.config.initializer_factor
nn.init.normal_(module.class_embedding, mean=0.0, std=module.embed_dim**-0.5 * factor)
nn.init.normal_(module.patch_embedding.weight, std=module.config.initializer_range * factor)
nn.init.normal_(module.position_embedding.weight, std=module.config.initializer_range * factor)
elif isinstance(module, CLIPAttention):
factor = self.config.initializer_factor
in_proj_std = (module.embed_dim**-0.5) * ((2 * module.config.num_hidden_layers) ** -0.5) * factor
out_proj_std = (module.embed_dim**-0.5) * factor
nn.init.normal_(module.q_proj.weight, std=in_proj_std)
nn.init.normal_(module.k_proj.weight, std=in_proj_std)
nn.init.normal_(module.v_proj.weight, std=in_proj_std)
nn.init.normal_(module.out_proj.weight, std=out_proj_std)
elif isinstance(module, CLIPMLP):
factor = self.config.initializer_factor
in_proj_std = (module.config.hidden_size**-0.5) * ((2 * module.config.num_hidden_layers) ** -0.5) * factor
fc_std = (2 * module.config.hidden_size) ** -0.5 * factor
nn.init.normal_(module.fc1.weight, std=fc_std)
nn.init.normal_(module.fc2.weight, std=in_proj_std)
elif isinstance(module, CLIPModel):
nn.init.normal_(
module.text_projection.weight,
std=module.text_embed_dim**-0.5 * self.config.initializer_factor,
)
nn.init.normal_(
module.visual_projection.weight,
std=module.vision_embed_dim**-0.5 * self.config.initializer_factor,
)
elif isinstance(module, CLIPVisionModelWithProjection):
nn.init.normal_(
module.visual_projection.weight,
std=self.config.hidden_size**-0.5 * self.config.initializer_factor,
)
elif isinstance(module, CLIPTextModelWithProjection):
nn.init.normal_(
module.text_projection.weight,
std=self.config.hidden_size**-0.5 * self.config.initializer_factor,
)
elif isinstance(module, CLIPForImageClassification):
nn.init.normal_(
module.classifier.weight,
std=self.config.vision_config.hidden_size**-0.5 * self.config.initializer_factor,
)
if isinstance(module, nn.LayerNorm):
module.bias.data.zero_()
module.weight.data.fill_(1.0)
if isinstance(module, nn.Linear) and module.bias is not None:
module.bias.data.zero_()
CLIP_START_DOCSTRING = r"""
This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
etc.)
This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
and behavior.
Parameters:
config ([`CLIPConfig`]): Model configuration class with all the parameters of the model.
Initializing with a config file does not load the weights associated with the model, only the
configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
"""
CLIP_TEXT_INPUTS_DOCSTRING = r"""
Args:
input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
it.
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
[`PreTrainedTokenizer.__call__`] for details.
[What are input IDs?](../glossary#input-ids)
attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
- 1 for tokens that are **not masked**,
- 0 for tokens that are **masked**.
[What are attention masks?](../glossary#attention-mask)
position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
config.max_position_embeddings - 1]`.
[What are position IDs?](../glossary#position-ids)
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
tensors for more detail.
output_hidden_states (`bool`, *optional*):
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
more detail.
return_dict (`bool`, *optional*):
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
"""
CLIP_VISION_INPUTS_DOCSTRING = r"""
Args:
pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
Pixel values. Padding will be ignored by default should you provide it. Pixel values can be obtained using
[`AutoImageProcessor`]. See [`CLIPImageProcessor.__call__`] for details.
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
tensors for more detail.
output_hidden_states (`bool`, *optional*):
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
more detail.
return_dict (`bool`, *optional*):
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
"""
CLIP_INPUTS_DOCSTRING = r"""
Args:
input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
it.
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
[`PreTrainedTokenizer.__call__`] for details.
[What are input IDs?](../glossary#input-ids)
attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
- 1 for tokens that are **not masked**,
- 0 for tokens that are **masked**.
[What are attention masks?](../glossary#attention-mask)
position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
config.max_position_embeddings - 1]`.
[What are position IDs?](../glossary#position-ids)
pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
Pixel values. Padding will be ignored by default should you provide it. Pixel values can be obtained using
[`AutoImageProcessor`]. See [`CLIPImageProcessor.__call__`] for details.
return_loss (`bool`, *optional*):
Whether or not to return the contrastive loss.
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
tensors for more detail.
output_hidden_states (`bool`, *optional*):
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
more detail.
return_dict (`bool`, *optional*):
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
"""
class CLIPEncoder(nn.Module):
"""
Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a
[`CLIPEncoderLayer`].
Args:
config: CLIPConfig
"""
def __init__(self, config: CLIPConfig):
super().__init__()
self.config = config
self.layers = nn.ModuleList([CLIPEncoderLayer(config) for _ in range(config.num_hidden_layers)])
self.gradient_checkpointing = False
def forward(
self,
inputs_embeds,
attention_mask: Optional[torch.Tensor] = None,
causal_attention_mask: Optional[torch.Tensor] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
) -> Union[Tuple, BaseModelOutput]:
r"""
Args:
inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
This is useful if you want more control over how to convert `input_ids` indices into associated vectors
than the model's internal embedding lookup matrix.
attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
- 1 for tokens that are **not masked**,
- 0 for tokens that are **masked**.
[What are attention masks?](../glossary#attention-mask)
causal_attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
Causal mask for the text model. Mask values selected in `[0, 1]`:
- 1 for tokens that are **not masked**,
- 0 for tokens that are **masked**.
[What are attention masks?](../glossary#attention-mask)
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
returned tensors for more detail.
output_hidden_states (`bool`, *optional*):
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
for more detail.
return_dict (`bool`, *optional*):
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
"""
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
encoder_states = () if output_hidden_states else None
all_attentions = () if output_attentions else None
hidden_states = inputs_embeds
for idx, encoder_layer in enumerate(self.layers):
if output_hidden_states:
encoder_states = encoder_states + (hidden_states,)
if self.gradient_checkpointing and self.training:
layer_outputs = self._gradient_checkpointing_func(
encoder_layer.__call__,
hidden_states,
attention_mask,
causal_attention_mask,
output_attentions,
)
else:
layer_outputs = encoder_layer(
hidden_states,
attention_mask,
causal_attention_mask,
output_attentions=output_attentions,
)
hidden_states = layer_outputs[0]
if output_attentions:
all_attentions = all_attentions + (layer_outputs[1],)
if output_hidden_states:
encoder_states = encoder_states + (hidden_states,)
if not return_dict:
return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None)
return BaseModelOutput(
last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions
)
class CLIPTextTransformer(nn.Module):
def __init__(self, config: CLIPTextConfig):
super().__init__()
self.config = config
embed_dim = config.hidden_size
self.embeddings = CLIPTextEmbeddings(config)
self.encoder = CLIPEncoder(config)
self.final_layer_norm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
# For `pooled_output` computation
self.eos_token_id = config.eos_token_id
# For attention mask, it differs between `flash_attention_2` and other attention implementations
self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"
@add_start_docstrings_to_model_forward(CLIP_TEXT_INPUTS_DOCSTRING)
@replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=CLIPTextConfig)
def forward(
self,
input_ids: Optional[torch.Tensor] = None,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.Tensor] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
) -> Union[Tuple, BaseModelOutputWithPooling]:
r"""
Returns:
"""
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 None:
raise ValueError("You have to specify input_ids")
input_shape = input_ids.size()
input_ids = input_ids.view(-1, input_shape[-1])
hidden_states = self.embeddings(input_ids=input_ids, position_ids=position_ids)
# CLIP's text model uses causal mask, prepare it here.
# https://github.com/openai/CLIP/blob/cfcffb90e69f37bf2ff1e988237a0fbe41f33c04/clip/model.py#L324
causal_attention_mask = _create_4d_causal_attention_mask(
input_shape, hidden_states.dtype, device=hidden_states.device
)
# expand attention_mask
if attention_mask is not None and not self._use_flash_attention_2:
# [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
attention_mask = _prepare_4d_attention_mask(attention_mask, hidden_states.dtype)
encoder_outputs = self.encoder(
inputs_embeds=hidden_states,
attention_mask=attention_mask,
causal_attention_mask=causal_attention_mask,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
last_hidden_state = encoder_outputs[0]
last_hidden_state = self.final_layer_norm(last_hidden_state)
if self.eos_token_id == 2:
# The `eos_token_id` was incorrect before PR #24773: Let's keep what have been done here.
# A CLIP model with such `eos_token_id` in the config can't work correctly with extra new tokens added
# ------------------------------------------------------------
# text_embeds.shape = [batch_size, sequence_length, transformer.width]
# take features from the eot embedding (eot_token is the highest number in each sequence)
# casting to torch.int for onnx compatibility: argmax doesn't support int64 inputs with opset 14
pooled_output = last_hidden_state[
torch.arange(last_hidden_state.shape[0], device=last_hidden_state.device),
input_ids.to(dtype=torch.int, device=last_hidden_state.device).argmax(dim=-1),
]
else:
# The config gets updated `eos_token_id` from PR #24773 (so the use of exta new tokens is possible)
pooled_output = last_hidden_state[
torch.arange(last_hidden_state.shape[0], device=last_hidden_state.device),
# We need to get the first position of `eos_token_id` value (`pad_token_ids` might equal to `eos_token_id`)
# Note: we assume each sequence (along batch dim.) contains an `eos_token_id` (e.g. prepared by the tokenizer)
(input_ids.to(dtype=torch.int, device=last_hidden_state.device) == self.eos_token_id)
.int()
.argmax(dim=-1),
]
if not return_dict:
return (last_hidden_state, pooled_output) + encoder_outputs[1:]
return BaseModelOutputWithPooling(
last_hidden_state=last_hidden_state,
pooler_output=pooled_output,
hidden_states=encoder_outputs.hidden_states,
attentions=encoder_outputs.attentions,
)
@add_start_docstrings(
"""The text model from CLIP without any head or projection on top.""",
CLIP_START_DOCSTRING,
)
class CLIPTextModel(CLIPPreTrainedModel):
config_class = CLIPTextConfig
_no_split_modules = ["CLIPTextEmbeddings", "CLIPEncoderLayer"]
def __init__(self, config: CLIPTextConfig):
super().__init__(config)
self.text_model = CLIPTextTransformer(config)
# Initialize weights and apply final processing
self.post_init()
def get_input_embeddings(self) -> nn.Module:
return self.text_model.embeddings.token_embedding
def set_input_embeddings(self, value):
self.text_model.embeddings.token_embedding = value
@add_start_docstrings_to_model_forward(CLIP_TEXT_INPUTS_DOCSTRING)
@replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=CLIPTextConfig)
def forward(
self,
input_ids: Optional[torch.Tensor] = None,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.Tensor] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
) -> Union[Tuple, BaseModelOutputWithPooling]:
r"""
Returns:
Examples:
```python
>>> from transformers import AutoTokenizer, CLIPTextModel
>>> model = CLIPTextModel.from_pretrained("openai/clip-vit-base-patch32")
>>> tokenizer = AutoTokenizer.from_pretrained("openai/clip-vit-base-patch32")
>>> inputs = tokenizer(["a photo of a cat", "a photo of a dog"], padding=True, return_tensors="pt")
>>> outputs = model(**inputs)
>>> last_hidden_state = outputs.last_hidden_state
>>> pooled_output = outputs.pooler_output # pooled (EOS token) states
```"""
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
return self.text_model(
input_ids=input_ids,
attention_mask=attention_mask,
position_ids=position_ids,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
class CLIPVisionTransformer(nn.Module):
def __init__(self, config: CLIPVisionConfig):
super().__init__()
self.config = config
embed_dim = config.hidden_size
self.embeddings = CLIPVisionEmbeddings(config)
self.pre_layrnorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
self.encoder = CLIPEncoder(config)
self.post_layernorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
@add_start_docstrings_to_model_forward(CLIP_VISION_INPUTS_DOCSTRING)
@replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=CLIPVisionConfig)
def forward(
self,
pixel_values: Optional[torch.FloatTensor] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
) -> Union[Tuple, BaseModelOutputWithPooling]:
r"""
Returns:
"""
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 pixel_values is None:
raise ValueError("You have to specify pixel_values")
hidden_states = self.embeddings(pixel_values)
hidden_states = self.pre_layrnorm(hidden_states)
encoder_outputs = self.encoder(
inputs_embeds=hidden_states,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
last_hidden_state = encoder_outputs[0]
pooled_output = last_hidden_state[:, 0, :]
pooled_output = self.post_layernorm(pooled_output)
if not return_dict:
return (last_hidden_state, pooled_output) + encoder_outputs[1:]
return BaseModelOutputWithPooling(
last_hidden_state=last_hidden_state,
pooler_output=pooled_output,
hidden_states=encoder_outputs.hidden_states,
attentions=encoder_outputs.attentions,
)
@add_start_docstrings(
"""The vision model from CLIP without any head or projection on top.""",
CLIP_START_DOCSTRING,
)
class CLIPVisionModel(CLIPPreTrainedModel):
config_class = CLIPVisionConfig
main_input_name = "pixel_values"
_no_split_modules = ["CLIPEncoderLayer"]
def __init__(self, config: CLIPVisionConfig):
super().__init__(config)
self.vision_model = CLIPVisionTransformer(config)
# Initialize weights and apply final processing
self.post_init()
def get_input_embeddings(self) -> nn.Module:
return self.vision_model.embeddings.patch_embedding
@add_start_docstrings_to_model_forward(CLIP_VISION_INPUTS_DOCSTRING)
@replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=CLIPVisionConfig)
def forward(
self,
pixel_values: Optional[torch.FloatTensor] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
) -> Union[Tuple, BaseModelOutputWithPooling]:
r"""
Returns:
Examples:
```python
>>> from PIL import Image
>>> import requests
>>> from transformers import AutoProcessor, CLIPVisionModel
>>> model = CLIPVisionModel.from_pretrained("openai/clip-vit-base-patch32")
>>> processor = AutoProcessor.from_pretrained("openai/clip-vit-base-patch32")
>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
>>> image = Image.open(requests.get(url, stream=True).raw)
>>> inputs = processor(images=image, return_tensors="pt")
>>> outputs = model(**inputs)
>>> last_hidden_state = outputs.last_hidden_state
>>> pooled_output = outputs.pooler_output # pooled CLS states
```"""
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
return self.vision_model(
pixel_values=pixel_values,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
@add_start_docstrings(CLIP_START_DOCSTRING)
class CLIPModel(CLIPPreTrainedModel):
config_class = CLIPConfig
_no_split_modules = ["CLIPTextEmbeddings", "CLIPEncoderLayer", "CLIPVisionEmbeddings"]
def __init__(self, config: CLIPConfig):
super().__init__(config)
if not isinstance(config.text_config, CLIPTextConfig):
raise TypeError(
"config.text_config is expected to be of type CLIPTextConfig but is of type"
f" {type(config.text_config)}."
)
if not isinstance(config.vision_config, CLIPVisionConfig):
raise TypeError(
"config.vision_config is expected to be of type CLIPVisionConfig but is of type"
f" {type(config.vision_config)}."
)
text_config = config.text_config
vision_config = config.vision_config
self.projection_dim = config.projection_dim
self.text_embed_dim = text_config.hidden_size
self.vision_embed_dim = vision_config.hidden_size
text_model = CLIPTextModel._from_config(text_config, attn_implementation=config._attn_implementation)
self.text_model = text_model.text_model
vision_model = CLIPVisionModel._from_config(vision_config, attn_implementation=config._attn_implementation)
self.vision_model = vision_model.vision_model
self.visual_projection = nn.Linear(self.vision_embed_dim, self.projection_dim, bias=False)
self.text_projection = nn.Linear(self.text_embed_dim, self.projection_dim, bias=False)
self.logit_scale = nn.Parameter(torch.tensor(self.config.logit_scale_init_value))
# Initialize weights and apply final processing
self.post_init()
@add_start_docstrings_to_model_forward(CLIP_TEXT_INPUTS_DOCSTRING)
def get_text_features(
self,
input_ids: Optional[torch.Tensor] = None,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.Tensor] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
) -> torch.FloatTensor:
r"""
Returns:
text_features (`torch.FloatTensor` of shape `(batch_size, output_dim`): The text embeddings obtained by
applying the projection layer to the pooled output of [`CLIPTextModel`].
Examples:
```python
>>> from transformers import AutoTokenizer, CLIPModel
>>> model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32")
>>> tokenizer = AutoTokenizer.from_pretrained("openai/clip-vit-base-patch32")
>>> inputs = tokenizer(["a photo of a cat", "a photo of a dog"], padding=True, return_tensors="pt")
>>> text_features = model.get_text_features(**inputs)
```"""
# Use CLIP model's config for some fields (if specified) instead of those of vision & text components.
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
text_outputs = self.text_model(
input_ids=input_ids,
attention_mask=attention_mask,
position_ids=position_ids,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
pooled_output = text_outputs[1]
text_features = self.text_projection(pooled_output)
return text_features
@add_start_docstrings_to_model_forward(CLIP_VISION_INPUTS_DOCSTRING)
def get_image_features(
self,
pixel_values: Optional[torch.FloatTensor] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
) -> torch.FloatTensor:
r"""
Returns:
image_features (`torch.FloatTensor` of shape `(batch_size, output_dim`): The image embeddings obtained by
applying the projection layer to the pooled output of [`CLIPVisionModel`].
Examples:
```python
>>> from PIL import Image
>>> import requests
>>> from transformers import AutoProcessor, CLIPModel
>>> model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32")
>>> processor = AutoProcessor.from_pretrained("openai/clip-vit-base-patch32")
>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
>>> image = Image.open(requests.get(url, stream=True).raw)
>>> inputs = processor(images=image, return_tensors="pt")
>>> image_features = model.get_image_features(**inputs)
```"""
# Use CLIP model's config for some fields (if specified) instead of those of vision & text components.
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
vision_outputs = self.vision_model(
pixel_values=pixel_values,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
pooled_output = vision_outputs[1] # pooled_output
image_features = self.visual_projection(pooled_output)
return image_features
@add_start_docstrings_to_model_forward(CLIP_INPUTS_DOCSTRING)
@replace_return_docstrings(output_type=CLIPOutput, config_class=CLIPConfig)
def forward(
self,
input_ids: Optional[torch.LongTensor] = None,
pixel_values: Optional[torch.FloatTensor] = None,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.LongTensor] = None,
return_loss: Optional[bool] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
) -> Union[Tuple, CLIPOutput]:
r"""
Returns:
Examples:
```python
>>> from PIL import Image
>>> import requests
>>> from transformers import AutoProcessor, CLIPModel
>>> model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32")
>>> processor = AutoProcessor.from_pretrained("openai/clip-vit-base-patch32")
>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
>>> image = Image.open(requests.get(url, stream=True).raw)
>>> inputs = processor(
... text=["a photo of a cat", "a photo of a dog"], images=image, return_tensors="pt", padding=True
... )
>>> outputs = model(**inputs)
>>> logits_per_image = outputs.logits_per_image # this is the image-text similarity score
>>> probs = logits_per_image.softmax(dim=1) # we can take the softmax to get the label probabilities
```"""
# Use CLIP model's config for some fields (if specified) instead of those of vision & text components.
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
vision_outputs = self.vision_model(
pixel_values=pixel_values,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
text_outputs = self.text_model(
input_ids=input_ids,
attention_mask=attention_mask,
position_ids=position_ids,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
image_embeds = vision_outputs[1]
image_embeds = self.visual_projection(image_embeds)
text_embeds = text_outputs[1]
text_embeds = self.text_projection(text_embeds)
# normalized features
image_embeds = image_embeds / _get_vector_norm(image_embeds)
text_embeds = text_embeds / _get_vector_norm(text_embeds)
# cosine similarity as logits
logit_scale = self.logit_scale.exp()
logits_per_text = torch.matmul(text_embeds, image_embeds.t().to(text_embeds.device)) * logit_scale.to(
text_embeds.device
)
logits_per_image = logits_per_text.t()
loss = None
if return_loss:
loss = clip_loss(logits_per_text)
if not return_dict:
output = (logits_per_image, logits_per_text, text_embeds, image_embeds, text_outputs, vision_outputs)
return ((loss,) + output) if loss is not None else output
return CLIPOutput(
loss=loss,
logits_per_image=logits_per_image,
logits_per_text=logits_per_text,
text_embeds=text_embeds,
image_embeds=image_embeds,
text_model_output=text_outputs,
vision_model_output=vision_outputs,
)
@add_start_docstrings(
"""
CLIP Text Model with a projection layer on top (a linear layer on top of the pooled output).
""",
CLIP_START_DOCSTRING,
)
class CLIPTextModelWithProjection(CLIPPreTrainedModel):
config_class = CLIPTextConfig
_no_split_modules = ["CLIPTextEmbeddings", "CLIPEncoderLayer"]
def __init__(self, config: CLIPTextConfig):
super().__init__(config)
text_model = CLIPTextModel._from_config(config, attn_implementation=config._attn_implementation)
self.text_model = text_model.text_model
self.text_projection = nn.Linear(config.hidden_size, config.projection_dim, bias=False)
# Initialize weights and apply final processing
self.post_init()
def get_input_embeddings(self) -> nn.Module:
return self.text_model.embeddings.token_embedding
def set_input_embeddings(self, value):
self.text_model.embeddings.token_embedding = value
@add_start_docstrings_to_model_forward(CLIP_TEXT_INPUTS_DOCSTRING)
@replace_return_docstrings(output_type=CLIPTextModelOutput, config_class=CLIPTextConfig)
def forward(
self,
input_ids: Optional[torch.Tensor] = None,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.Tensor] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
) -> Union[Tuple, CLIPTextModelOutput]:
r"""
Returns:
Examples:
```python
>>> from transformers import AutoTokenizer, CLIPTextModelWithProjection
>>> model = CLIPTextModelWithProjection.from_pretrained("openai/clip-vit-base-patch32")
>>> tokenizer = AutoTokenizer.from_pretrained("openai/clip-vit-base-patch32")
>>> inputs = tokenizer(["a photo of a cat", "a photo of a dog"], padding=True, return_tensors="pt")
>>> outputs = model(**inputs)
>>> text_embeds = outputs.text_embeds
```"""
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
text_outputs = self.text_model(
input_ids=input_ids,
attention_mask=attention_mask,
position_ids=position_ids,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
pooled_output = text_outputs[1]
text_embeds = self.text_projection(pooled_output)
if not return_dict:
outputs = (text_embeds, text_outputs[0]) + text_outputs[2:]
return tuple(output for output in outputs if output is not None)
return CLIPTextModelOutput(
text_embeds=text_embeds,
last_hidden_state=text_outputs.last_hidden_state,
hidden_states=text_outputs.hidden_states,
attentions=text_outputs.attentions,
)
@add_start_docstrings(
"""
CLIP Vision Model with a projection layer on top (a linear layer on top of the pooled output).
""",
CLIP_START_DOCSTRING,
)
class CLIPVisionModelWithProjection(CLIPPreTrainedModel):
config_class = CLIPVisionConfig
main_input_name = "pixel_values"
def __init__(self, config: CLIPVisionConfig):
super().__init__(config)
vision_model = CLIPVisionModel._from_config(config, attn_implementation=config._attn_implementation)
self.vision_model = vision_model.vision_model
self.visual_projection = nn.Linear(config.hidden_size, config.projection_dim, bias=False)
# Initialize weights and apply final processing
self.post_init()
def get_input_embeddings(self) -> nn.Module:
return self.vision_model.embeddings.patch_embedding
@add_start_docstrings_to_model_forward(CLIP_VISION_INPUTS_DOCSTRING)
@replace_return_docstrings(output_type=CLIPVisionModelOutput, config_class=CLIPVisionConfig)
def forward(
self,
pixel_values: Optional[torch.FloatTensor] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
) -> Union[Tuple, CLIPVisionModelOutput]:
r"""
Returns:
Examples:
```python
>>> from PIL import Image
>>> import requests
>>> from transformers import AutoProcessor, CLIPVisionModelWithProjection
>>> model = CLIPVisionModelWithProjection.from_pretrained("openai/clip-vit-base-patch32")
>>> processor = AutoProcessor.from_pretrained("openai/clip-vit-base-patch32")
>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
>>> image = Image.open(requests.get(url, stream=True).raw)
>>> inputs = processor(images=image, return_tensors="pt")
>>> outputs = model(**inputs)
>>> image_embeds = outputs.image_embeds
```"""
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
vision_outputs = self.vision_model(
pixel_values=pixel_values,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
pooled_output = vision_outputs[1] # pooled_output
image_embeds = self.visual_projection(pooled_output)
if not return_dict:
outputs = (image_embeds, vision_outputs[0]) + vision_outputs[2:]
return tuple(output for output in outputs if output is not None)
return CLIPVisionModelOutput(
image_embeds=image_embeds,
last_hidden_state=vision_outputs.last_hidden_state,
hidden_states=vision_outputs.hidden_states,
attentions=vision_outputs.attentions,
)
@add_start_docstrings(
"""
CLIP vision encoder with an image classification head on top (a linear layer on top of the pooled final hidden states of
the patch tokens) e.g. for ImageNet.
""",
CLIP_START_DOCSTRING,
)
class CLIPForImageClassification(CLIPPreTrainedModel):
main_input_name = "pixel_values"
def __init__(self, config: CLIPConfig) -> None:
super().__init__(config)
self.num_labels = config.num_labels
vision_model = CLIPVisionModel._from_config(
config.vision_config, attn_implementation=config._attn_implementation
)
self.vision_model = vision_model.vision_model
# Classifier head
self.classifier = (
nn.Linear(config.vision_config.hidden_size, config.num_labels) if config.num_labels > 0 else nn.Identity()
)
# Initialize weights and apply final processing
self.post_init()
@add_start_docstrings_to_model_forward(CLIP_INPUTS_DOCSTRING)
@add_code_sample_docstrings(
checkpoint=_IMAGE_CLASS_CHECKPOINT,
output_type=ImageClassifierOutput,
config_class=_CONFIG_FOR_DOC,
expected_output=_IMAGE_CLASS_EXPECTED_OUTPUT,
)
def forward(
self,
pixel_values: Optional[torch.Tensor] = None,
labels: Optional[torch.Tensor] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
) -> Union[tuple, ImageClassifierOutput]:
r"""
labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
Labels for computing the image classification/regression loss. Indices should be in `[0, ...,
config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
"""
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
outputs = self.vision_model(
pixel_values,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
sequence_output = outputs[0]
# average pool the patch tokens
sequence_output = torch.mean(sequence_output[:, 1:, :], dim=1)
# apply classifier
logits = self.classifier(sequence_output)
loss = None
if labels is not None:
# move labels to correct device to enable model parallelism
labels = labels.to(logits.device)
if self.config.problem_type is None:
if self.num_labels == 1:
self.config.problem_type = "regression"
elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
self.config.problem_type = "single_label_classification"
else:
self.config.problem_type = "multi_label_classification"
if self.config.problem_type == "regression":
loss_fct = MSELoss()
if self.num_labels == 1:
loss = loss_fct(logits.squeeze(), labels.squeeze())
else:
loss = loss_fct(logits, labels)
elif self.config.problem_type == "single_label_classification":
loss_fct = CrossEntropyLoss()
loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
elif self.config.problem_type == "multi_label_classification":
loss_fct = BCEWithLogitsLoss()
loss = loss_fct(logits, labels)
if not return_dict:
output = (logits,) + outputs[2:]
return ((loss,) + output) if loss is not None else output
return ImageClassifierOutput(
loss=loss,
logits=logits,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
|
transformers/src/transformers/models/clip/modeling_clip.py/0
|
{
"file_path": "transformers/src/transformers/models/clip/modeling_clip.py",
"repo_id": "transformers",
"token_count": 29729
}
| 348
|
# coding=utf-8
# Copyright 2023 The HuggingFace Inc. team.
#
# 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.
"""English Normalizer class for CLVP."""
import re
class EnglishNormalizer:
def __init__(self):
# List of (regular expression, replacement) pairs for abbreviations:
self._abbreviations = [
(re.compile("\\b%s\\." % x[0], re.IGNORECASE), x[1])
for x in [
("mrs", "misess"),
("mr", "mister"),
("dr", "doctor"),
("st", "saint"),
("co", "company"),
("jr", "junior"),
("maj", "major"),
("gen", "general"),
("drs", "doctors"),
("rev", "reverend"),
("lt", "lieutenant"),
("hon", "honorable"),
("sgt", "sergeant"),
("capt", "captain"),
("esq", "esquire"),
("ltd", "limited"),
("col", "colonel"),
("ft", "fort"),
]
]
self.ones = ["", "one", "two", "three", "four", "five", "six", "seven", "eight", "nine"]
self.teens = [
"ten",
"eleven",
"twelve",
"thirteen",
"fourteen",
"fifteen",
"sixteen",
"seventeen",
"eighteen",
"nineteen",
]
self.tens = ["", "", "twenty", "thirty", "forty", "fifty", "sixty", "seventy", "eighty", "ninety"]
def number_to_words(self, num: int) -> str:
"""
Converts numbers(`int`) to words(`str`).
Please note that it only supports upto - "'nine hundred ninety-nine quadrillion, nine hundred ninety-nine
trillion, nine hundred ninety-nine billion, nine hundred ninety-nine million, nine hundred ninety-nine
thousand, nine hundred ninety-nine'" or `number_to_words(999_999_999_999_999_999)`.
"""
if num == 0:
return "zero"
elif num < 0:
return "minus " + self.number_to_words(abs(num))
elif num < 10:
return self.ones[num]
elif num < 20:
return self.teens[num - 10]
elif num < 100:
return self.tens[num // 10] + ("-" + self.number_to_words(num % 10) if num % 10 != 0 else "")
elif num < 1000:
return (
self.ones[num // 100] + " hundred" + (" " + self.number_to_words(num % 100) if num % 100 != 0 else "")
)
elif num < 1_000_000:
return (
self.number_to_words(num // 1000)
+ " thousand"
+ (", " + self.number_to_words(num % 1000) if num % 1000 != 0 else "")
)
elif num < 1_000_000_000:
return (
self.number_to_words(num // 1_000_000)
+ " million"
+ (", " + self.number_to_words(num % 1_000_000) if num % 1_000_000 != 0 else "")
)
elif num < 1_000_000_000_000:
return (
self.number_to_words(num // 1_000_000_000)
+ " billion"
+ (", " + self.number_to_words(num % 1_000_000_000) if num % 1_000_000_000 != 0 else "")
)
elif num < 1_000_000_000_000_000:
return (
self.number_to_words(num // 1_000_000_000_000)
+ " trillion"
+ (", " + self.number_to_words(num % 1_000_000_000_000) if num % 1_000_000_000_000 != 0 else "")
)
elif num < 1_000_000_000_000_000_000:
return (
self.number_to_words(num // 1_000_000_000_000_000)
+ " quadrillion"
+ (
", " + self.number_to_words(num % 1_000_000_000_000_000)
if num % 1_000_000_000_000_000 != 0
else ""
)
)
else:
return "number out of range"
def convert_to_ascii(self, text: str) -> str:
"""
Converts unicode to ascii
"""
return text.encode("ascii", "ignore").decode("utf-8")
def _expand_dollars(self, m: str) -> str:
"""
This method is used to expand numerical dollar values into spoken words.
"""
match = m.group(1)
parts = match.split(".")
if len(parts) > 2:
return match + " dollars" # Unexpected format
dollars = int(parts[0]) if parts[0] else 0
cents = int(parts[1]) if len(parts) > 1 and parts[1] else 0
if dollars and cents:
dollar_unit = "dollar" if dollars == 1 else "dollars"
cent_unit = "cent" if cents == 1 else "cents"
return "%s %s, %s %s" % (dollars, dollar_unit, cents, cent_unit)
elif dollars:
dollar_unit = "dollar" if dollars == 1 else "dollars"
return "%s %s" % (dollars, dollar_unit)
elif cents:
cent_unit = "cent" if cents == 1 else "cents"
return "%s %s" % (cents, cent_unit)
else:
return "zero dollars"
def _remove_commas(self, m: str) -> str:
"""
This method is used to remove commas from sentences.
"""
return m.group(1).replace(",", "")
def _expand_decimal_point(self, m: str) -> str:
"""
This method is used to expand '.' into spoken word ' point '.
"""
return m.group(1).replace(".", " point ")
def _expand_ordinal(self, num: str) -> str:
"""
This method is used to expand ordinals such as '1st', '2nd' into spoken words.
"""
ordinal_suffixes = {1: "st", 2: "nd", 3: "rd"}
num = int(num.group(0)[:-2])
if 10 <= num % 100 and num % 100 <= 20:
suffix = "th"
else:
suffix = ordinal_suffixes.get(num % 10, "th")
return self.number_to_words(num) + suffix
def _expand_number(self, m: str) -> str:
"""
This method acts as a preprocessing step for numbers between 1000 and 3000 (same as the original repository,
link :
https://github.com/neonbjb/tortoise-tts/blob/4003544b6ff4b68c09856e04d3eff9da26d023c2/tortoise/utils/tokenizer.py#L86)
"""
num = int(m.group(0))
if num > 1000 and num < 3000:
if num == 2000:
return "two thousand"
elif num > 2000 and num < 2010:
return "two thousand " + self.number_to_words(num % 100)
elif num % 100 == 0:
return self.number_to_words(num // 100) + " hundred"
else:
return self.number_to_words(num)
else:
return self.number_to_words(num)
def normalize_numbers(self, text: str) -> str:
"""
This method is used to normalize numbers within a text such as converting the numbers to words, removing
commas, etc.
"""
text = re.sub(re.compile(r"([0-9][0-9\,]+[0-9])"), self._remove_commas, text)
text = re.sub(re.compile(r"£([0-9\,]*[0-9]+)"), r"\1 pounds", text)
text = re.sub(re.compile(r"\$([0-9\.\,]*[0-9]+)"), self._expand_dollars, text)
text = re.sub(re.compile(r"([0-9]+\.[0-9]+)"), self._expand_decimal_point, text)
text = re.sub(re.compile(r"[0-9]+(st|nd|rd|th)"), self._expand_ordinal, text)
text = re.sub(re.compile(r"[0-9]+"), self._expand_number, text)
return text
def expand_abbreviations(self, text: str) -> str:
"""
Expands the abbreviate words.
"""
for regex, replacement in self._abbreviations:
text = re.sub(regex, replacement, text)
return text
def collapse_whitespace(self, text: str) -> str:
"""
Removes multiple whitespaces
"""
return re.sub(re.compile(r"\s+"), " ", text)
def __call__(self, text):
"""
Converts text to ascii, numbers / number-like quantities to their spelt-out counterparts and expands
abbreviations
"""
text = self.convert_to_ascii(text)
text = text.lower()
text = self.normalize_numbers(text)
text = self.expand_abbreviations(text)
text = self.collapse_whitespace(text)
text = text.replace('"', "")
return text
|
transformers/src/transformers/models/clvp/number_normalizer.py/0
|
{
"file_path": "transformers/src/transformers/models/clvp/number_normalizer.py",
"repo_id": "transformers",
"token_count": 4412
}
| 349
|
# coding=utf-8
# Copyright 2018 Salesforce and 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.
"""Salesforce CTRL configuration"""
from ...configuration_utils import PretrainedConfig
from ...utils import logging
logger = logging.get_logger(__name__)
class CTRLConfig(PretrainedConfig):
"""
This is the configuration class to store the configuration of a [`CTRLModel`] or a [`TFCTRLModel`]. It is used to
instantiate a CTRL model according to the specified arguments, defining the model architecture. Instantiating a
configuration with the defaults will yield a similar configuration to that of the
[Salesforce/ctrl](https://huggingface.co/Salesforce/ctrl) architecture from SalesForce.
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PretrainedConfig`] for more information.
Args:
vocab_size (`int`, *optional*, defaults to 246534):
Vocabulary size of the CTRL model. Defines the number of different tokens that can be represented by the
`inputs_ids` passed when calling [`CTRLModel`] or [`TFCTRLModel`].
n_positions (`int`, *optional*, defaults to 256):
The maximum sequence length that this model might ever be used with. Typically set this to something large
just in case (e.g., 512 or 1024 or 2048).
n_embd (`int`, *optional*, defaults to 1280):
Dimensionality of the embeddings and hidden states.
dff (`int`, *optional*, defaults to 8192):
Dimensionality of the inner dimension of the feed forward networks (FFN).
n_layer (`int`, *optional*, defaults to 48):
Number of hidden layers in the Transformer encoder.
n_head (`int`, *optional*, defaults to 16):
Number of attention heads for each attention layer in the Transformer encoder.
resid_pdrop (`float`, *optional*, defaults to 0.1):
The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
embd_pdrop (`int`, *optional*, defaults to 0.1):
The dropout ratio for the embeddings.
layer_norm_epsilon (`float`, *optional*, defaults to 1e-06):
The epsilon to use in the layer normalization layers
initializer_range (`float`, *optional*, defaults to 0.02):
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
use_cache (`bool`, *optional*, defaults to `True`):
Whether or not the model should return the last key/values attentions (not used by all models).
Examples:
```python
>>> from transformers import CTRLConfig, CTRLModel
>>> # Initializing a CTRL configuration
>>> configuration = CTRLConfig()
>>> # Initializing a model (with random weights) from the configuration
>>> model = CTRLModel(configuration)
>>> # Accessing the model configuration
>>> configuration = model.config
```"""
model_type = "ctrl"
keys_to_ignore_at_inference = ["past_key_values"]
attribute_map = {
"max_position_embeddings": "n_positions",
"hidden_size": "n_embd",
"num_attention_heads": "n_head",
"num_hidden_layers": "n_layer",
}
def __init__(
self,
vocab_size=246534,
n_positions=256,
n_embd=1280,
dff=8192,
n_layer=48,
n_head=16,
resid_pdrop=0.1,
embd_pdrop=0.1,
layer_norm_epsilon=1e-6,
initializer_range=0.02,
use_cache=True,
**kwargs,
):
self.vocab_size = vocab_size
self.n_positions = n_positions
self.n_embd = n_embd
self.n_layer = n_layer
self.n_head = n_head
self.dff = dff
self.resid_pdrop = resid_pdrop
self.embd_pdrop = embd_pdrop
self.layer_norm_epsilon = layer_norm_epsilon
self.initializer_range = initializer_range
self.use_cache = use_cache
super().__init__(**kwargs)
|
transformers/src/transformers/models/ctrl/configuration_ctrl.py/0
|
{
"file_path": "transformers/src/transformers/models/ctrl/configuration_ctrl.py",
"repo_id": "transformers",
"token_count": 1711
}
| 350
|
# coding=utf-8
# Copyright 2022 The HuggingFace Inc. team. 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.
"""Loading of Deformable DETR's CUDA kernels"""
import os
from pathlib import Path
def load_cuda_kernels():
from torch.utils.cpp_extension import load
root = Path(__file__).resolve().parent.parent.parent / "kernels" / "deformable_detr"
src_files = [
root / filename
for filename in [
"vision.cpp",
os.path.join("cpu", "ms_deform_attn_cpu.cpp"),
os.path.join("cuda", "ms_deform_attn_cuda.cu"),
]
]
load(
"MultiScaleDeformableAttention",
src_files,
with_cuda=True,
extra_include_paths=[str(root)],
extra_cflags=["-DWITH_CUDA=1"],
extra_cuda_cflags=[
"-DCUDA_HAS_FP16=1",
"-D__CUDA_NO_HALF_OPERATORS__",
"-D__CUDA_NO_HALF_CONVERSIONS__",
"-D__CUDA_NO_HALF2_OPERATORS__",
],
)
import MultiScaleDeformableAttention as MSDA
return MSDA
|
transformers/src/transformers/models/deformable_detr/load_custom.py/0
|
{
"file_path": "transformers/src/transformers/models/deformable_detr/load_custom.py",
"repo_id": "transformers",
"token_count": 638
}
| 351
|
# coding=utf-8
# Copyright 2023 HuggingFace Inc. team. 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.
"""Tokenization classes for GPTSANJapanese."""
import collections
import json
import os
import re
from typing import List, Optional, Tuple, Union
import numpy as np
from ....tokenization_utils import PreTrainedTokenizer
from ....tokenization_utils_base import (
BatchEncoding,
PreTokenizedInput,
PreTokenizedInputPair,
TextInput,
TextInputPair,
TruncationStrategy,
)
from ....utils import PaddingStrategy, logging
logger = logging.get_logger(__name__)
VOCAB_FILES_NAMES = {"vocab_file": "vocab.txt", "emoji_file": "emoji.json"}
def load_vocab_and_emoji(vocab_file, emoji_file):
"""Loads a vocabulary file and emoji file into a dictionary."""
with open(emoji_file, "r", encoding="utf-8") as f:
emoji = json.loads(f.read())
vocab = collections.OrderedDict()
raw_vocab = collections.OrderedDict()
ids_to_tokens = collections.OrderedDict()
with open(vocab_file, "r", encoding="utf-8") as f:
token = f.readlines()
token = [[t.rstrip("\n")] if (t == ",\n" or "," not in t) else t.rstrip("\n").split(",") for t in token]
for idx, b in enumerate(token):
ids_to_tokens[idx] = b
raw_vocab[",".join(b)] = idx
for wd in b:
vocab[wd] = idx
return vocab, raw_vocab, ids_to_tokens, emoji
class GPTSanJapaneseTokenizer(PreTrainedTokenizer):
"""
This tokenizer is based on GPTNeoXJapaneseTokenizer and has the following modifications
- Decoding byte0~byte255 tokens correctly
- Added bagofword token handling
- Return token_type_ids for Prefix-LM model
The bagofword token represents a repetition of the previous token and is converted to 3 consecutive tokens when
decoding In addition, the original Japanese special Sub-Word-Encoding has been released in this repository
(https://github.com/tanreinama/Japanese-BPEEncoder_V2). The token_type_ids is a mask indicating the prefix input
position of the Prefix-LM model. To specify a prefix position, specify a prefix input for prefix_text, or specify a
sentence of the prefix part and the part after it as a text pair of batch input.
Example:
```python
>>> from transformers import GPTSanJapaneseTokenizer
>>> tokenizer = GPTSanJapaneseTokenizer.from_pretrained("Tanrei/GPTSAN-japanese")
>>> # You can confirm both 慶応 and 慶應 are encoded to 17750
>>> tokenizer("吾輩は猫である🐯。実は慶応(慶應)大学出身")["input_ids"]
[35993, 35998, 34347, 31459, 30647, 31448, 25, 30659, 35729, 35676, 32417, 30647, 17750, 35589, 17750, 35590, 321, 1281]
>>> # Both 慶応 and 慶應 are decoded to 慶応
>>> tokenizer.decode(tokenizer("吾輩は猫である🐯。実は慶応(慶應)大学出身")["input_ids"])
'吾輩は猫である🐯。実は慶応(慶応)大学出身'
```
Example for Prefix-LM:
```python
>>> from transformers import GPTSanJapaneseTokenizer
>>> tokenizer = GPTSanJapaneseTokenizer.from_pretrained("Tanrei/GPTSAN-japanese")
>>> tokenizer("実は慶応(慶應)大学出身", prefix_text="吾輩は猫である🐯。")["input_ids"]
[35993, 34347, 31459, 30647, 31448, 25, 30659, 35729, 35676, 35998, 32417, 30647, 17750, 35589, 17750, 35590, 321, 1281]
>>> # Mask for Prefix-LM inputs
>>> tokenizer("実は慶応(慶應)大学出身", prefix_text="吾輩は猫である🐯。")["token_type_ids"]
[1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0]
```
Example for batch encode:
```python
>>> from transformers import GPTSanJapaneseTokenizer
>>> tokenizer = GPTSanJapaneseTokenizer.from_pretrained("Tanrei/GPTSAN-japanese")
>>> tokenizer([["武田信玄", "は、"], ["織田信長", "の配下の、"]], padding=True)["input_ids"]
[[35993, 35998, 8640, 25948, 35993, 35998, 30647, 35675, 35999, 35999], [35993, 35998, 10382, 9868, 35993, 35998, 30646, 9459, 30646, 35675]]
>>> # Mask for Prefix-LM inputs
>>> tokenizer([["武田信玄", "は、"], ["織田信長", "の配下の、"]], padding=True)["token_type_ids"]
[[1, 0, 0, 0, 0, 0, 0, 0, 0, 0], [1, 0, 0, 0, 0, 0, 0, 0, 0, 0]]
>>> # Mask for padding
>>> tokenizer([["武田信玄", "は、"], ["織田信長", "の配下の、"]], padding=True)["attention_mask"]
[[1, 1, 1, 1, 1, 1, 1, 1, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1]]
```
Args:
vocab_file (`str`):
File containing the vocabulary.
emoji_file (`str`):
File containing the emoji.
unk_token (`str`, *optional*, defaults to `"<|nottoken|>"`):
The token used for unknown charactor
pad_token (`str`, *optional*, defaults to `"<|separator|>"`):
The token used for padding
bos_token (`str`, *optional*, defaults to `"<|startoftext|>"`):
The beginning of sequence token.
eos_token (`str`, *optional*, defaults to `"<|endoftext|>"`):
The end of sequence token.
sep_token (`str`, *optional*, defaults to `"<|segmenter|>"`):
A special token to separate token to prefix part and general input part.
do_clean_text (`bool`, *optional*, defaults to `False`):
Whether or not to clean text for URL, EMAIL, TEL, Japanese DATE and Japanese PRICE.
"""
vocab_files_names = VOCAB_FILES_NAMES
model_input_names = ["input_ids", "attention_mask", "token_type_ids"]
def __init__(
self,
vocab_file,
emoji_file,
unk_token="<|nottoken|>",
pad_token="<|separator|>",
bos_token="<|startoftext|>",
eos_token="<|endoftext|>",
sep_token="<|segmenter|>",
do_clean_text=False,
**kwargs,
):
if not os.path.isfile(vocab_file):
raise ValueError(
f"Can't find a vocabulary file at path '{vocab_file}'. To load the vocabulary from a Google pretrained"
" model use `tokenizer = GPTSanJapaneseTokenizer.from_pretrained(PRETRAINED_MODEL_NAME)`"
)
if not os.path.isfile(emoji_file):
raise ValueError(
f"Can't find a emoji file at path '{emoji_file}'. To load the emoji information from a Google"
" pretrained model use `tokenizer = GPTSanJapaneseTokenizer.from_pretrained(PRETRAINED_MODEL_NAME)`"
)
self.do_clean_text = do_clean_text
self.vocab, self.raw_vocab, self.ids_to_tokens, self.emoji = load_vocab_and_emoji(vocab_file, emoji_file)
self.subword_tokenizer = SubWordJapaneseTokenizer(
vocab=self.vocab, ids_to_tokens=self.ids_to_tokens, emoji=self.emoji
)
super().__init__(
unk_token=unk_token,
pad_token=pad_token,
bos_token=bos_token,
eos_token=eos_token,
sep_token=sep_token,
do_clean_text=do_clean_text,
**kwargs,
)
@property
def vocab_size(self):
# self.vocab contains support for character fluctuation unique to Japanese, and has a large number of vocab
return len(self.raw_vocab)
def get_vocab(self):
return dict(self.raw_vocab, **self.added_tokens_encoder)
def _tokenize(self, text):
return self.subword_tokenizer.tokenize(text, clean=self.do_clean_text)
def _convert_token_to_id(self, token):
"""Converts a token (str) in an id using the vocab."""
return self.vocab.get(token, self.vocab.get(self.unk_token))
def _convert_id_to_token(self, index):
"""Converts an index (integer) in a token (str) using the vocab."""
return self.subword_tokenizer.convert_id_to_token(index)
def convert_tokens_to_string(self, tokens):
"""Converts a sequence of tokens (string) in a single string."""
words = []
byte_tokens = []
for word in tokens:
if word[:6] == "<|byte" and word[-2:] == "|>":
byte_tokens.append(int(word[6:-2]))
else:
if len(byte_tokens) > 0:
words.append(bytearray(byte_tokens).decode("utf-8", errors="replace"))
byte_tokens = []
if word[:7] == "<|emoji" and word[-2:] == "|>":
words.append(self.emoji["emoji_inv"][word])
elif word == "<SP>":
words.append(" ")
elif word == "<BR>":
words.append("\n")
elif word == "<TAB>":
words.append("\t")
elif word == "<BLOCK>":
words.append("▀")
elif word == "<KIGOU>":
words.append("ǀ")
elif word == "<U2000U2BFF>":
words.append("‖")
elif word == "<|bagoftoken|>":
if len(words) > 0:
words.append(words[-1])
words.append(words[-1])
words.append(words[-1])
elif word.startswith("<|") and word.endswith("|>"):
words.append("")
else:
words.append(word)
if len(byte_tokens) > 0:
words.append(bytearray(byte_tokens).decode("utf-8", errors="replace"))
text = "".join(words)
return text
def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
index = 0
if os.path.isdir(save_directory):
vocab_file = os.path.join(
save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]
)
emoji_file = os.path.join(
save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["emoji_file"]
)
else:
vocab_file = (
(filename_prefix + "-" if filename_prefix else "") + save_directory + VOCAB_FILES_NAMES["vocab_file"]
)
emoji_file = (
(filename_prefix + "-" if filename_prefix else "") + save_directory + VOCAB_FILES_NAMES["emoji_file"]
)
with open(vocab_file, "w", encoding="utf-8") as writer:
for token_index, token in self.ids_to_tokens.items():
if index != token_index:
logger.warning(
f"Saving vocabulary to {vocab_file}: vocabulary indices are not consecutive."
" Please check that the vocabulary is not corrupted!"
)
index = token_index
writer.write(",".join(token) + "\n")
index += 1
with open(emoji_file, "w", encoding="utf-8") as writer:
json.dump(self.emoji, writer)
return vocab_file, emoji_file
def create_token_type_ids_from_sequences(
self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
) -> List[int]:
# docstyle-ignore
"""
The tokenizer returns token_type_ids as separators between the Prefix part and the rest.
token_type_ids is 1 for the Prefix part and 0 for the rest of the token.
Example:
```python
>>> from transformers import GPTSanJapaneseTokenizer
>>> tokenizer = GPTSanJapaneseTokenizer.from_pretrained("Tanrei/GPTSAN-japanese")
>>> x_token = tokenizer("ｱｲｳｴ")
>>> # input_ids: | SOT | SEG | ｱ | ｲ | ｳ | ｴ |
>>> # token_type_ids: | 1 | 0 | 0 | 0 | 0 | 0 |
>>> x_token = tokenizer("", prefix_text="ｱｲｳｴ")
>>> # input_ids: | SOT | ｱ | ｲ | ｳ | ｴ | SEG |
>>> # token_type_ids: | 1 | 1 | 1 | 1 | 1 | 0 |
>>> x_token = tokenizer("ｳｴ", prefix_text="ｱｲ")
>>> # input_ids: | SOT | ｱ | ｲ | SEG | ｳ | ｴ |
>>> # token_type_ids: | 1 | 1 | 1 | 0 | 0 | 0 |
```"""
prefix_len = 0
if self.sep_token in self.vocab:
segid = self.vocab[self.sep_token]
if segid in token_ids_0:
prefix_len = token_ids_0.index(segid)
if token_ids_1 is None:
total_len = len(token_ids_0)
else:
total_len = len(token_ids_0 + token_ids_1)
return prefix_len * [1] + (total_len - prefix_len) * [0]
def prepare_for_tokenization(self, text, prefix_text=None, add_sep_token=None, **kwargs):
# GPTSAN inserts extra SEP tokens in Prefix-LM in addition to SOT for text generation.
# SOT at the beginning of the text, and SEP at the separator between the Prefix part and the rest.
if add_sep_token is None:
add_sep_token = self.sep_token not in text # If insert un-prefix position explicitly
prepared = self.bos_token if self.bos_token in self.vocab else ""
prepared += prefix_text if prefix_text is not None else ""
if add_sep_token:
prepared += self.sep_token if self.sep_token in self.vocab else ""
prepared += text
return (prepared, kwargs)
def _batch_encode_plus(
self,
batch_text_or_text_pairs: Union[
List[TextInput], List[TextInputPair], List[PreTokenizedInput], List[PreTokenizedInputPair]
],
add_special_tokens: bool = True,
padding_strategy: PaddingStrategy = PaddingStrategy.DO_NOT_PAD,
truncation_strategy: TruncationStrategy = TruncationStrategy.DO_NOT_TRUNCATE,
max_length: Optional[int] = None,
stride: int = 0,
is_split_into_words: bool = False,
pad_to_multiple_of: Optional[int] = None,
return_tensors: Optional[str] = None,
return_token_type_ids: Optional[bool] = None,
return_attention_mask: Optional[bool] = None,
return_overflowing_tokens: bool = False,
return_special_tokens_mask: bool = False,
return_offsets_mapping: bool = False,
return_length: bool = False,
verbose: bool = True,
**kwargs,
) -> BatchEncoding:
# This tokenizer converts input text pairs into Prefix input and subsequent input
if isinstance(batch_text_or_text_pairs[0], tuple) or isinstance(tuple(batch_text_or_text_pairs[0]), list):
# As a single text with an explicit un-prefix position
batch_prefix_texts = []
for pref, txt in batch_text_or_text_pairs:
batch_prefix_texts.append(pref + self.sep_token + txt)
batch_text_or_text_pairs = batch_prefix_texts
return super()._batch_encode_plus(
batch_text_or_text_pairs,
add_special_tokens,
padding_strategy,
truncation_strategy,
max_length,
stride,
is_split_into_words,
pad_to_multiple_of,
return_tensors,
return_token_type_ids,
return_attention_mask,
return_overflowing_tokens,
return_special_tokens_mask,
return_offsets_mapping,
return_length,
verbose,
**kwargs,
)
class SubWordJapaneseTokenizer:
"""
This tokenizer is based on GPTNeoXJapaneseTokenizer and has the following modifications
- Decoding byte0~byte255 tokens correctly
- Added bagofword token handling
https://github.com/tanreinama/Japanese-BPEEncoder_V2 This tokenizer class is under MIT Lisence according to the
original repository.
MIT License
Copyright (c) 2020 tanreinama
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.
"""
def __init__(self, vocab, ids_to_tokens, emoji):
self.vocab = vocab # same as swe
self.ids_to_tokens = ids_to_tokens # same as bpe
self.emoji = emoji
self.maxlen = np.max([len(w) for w in self.vocab.keys()])
self.content_repatter1 = re.compile(r"(https?|ftp)(:\/\/[-_\.!~*\'()a-zA-Z0-9;\/?:\@&=\+$,%#]+)")
self.content_repatter2 = re.compile(r"[A-Za-z0-9\._+]*@[\-_0-9A-Za-z]+(\.[A-Za-z]+)*")
self.content_repatter3 = re.compile(r"[\(]{0,1}[0-9]{2,4}[\)\-\(]{0,1}[0-9]{2,4}[\)\-]{0,1}[0-9]{3,4}")
self.content_repatter4 = re.compile(
r"([12]\d{3}[/\-年])*(0?[1-9]|1[0-2])[/\-月]((0?[1-9]|[12][0-9]|3[01])日?)*(\d{1,2}|:|\d{1,2}時|\d{1,2}分|\(日\)|\(月\)|\(火\)|\(水\)|\(木\)|\(金\)|\(土\)|㈰|㈪|㈫|㈬|㈭|㈮|㈯)*"
)
self.content_repatter5 = re.compile(
r"(明治|大正|昭和|平成|令和|㍾|㍽|㍼|㍻|\u32ff)\d{1,2}年(0?[1-9]|1[0-2])月(0?[1-9]|[12][0-9]|3[01])日(\d{1,2}|:|\d{1,2}時|\d{1,2}分|\(日\)|\(月\)|\(火\)|\(水\)|\(木\)|\(金\)|\(土\)|㈰|㈪|㈫|㈬|㈭|㈮|㈯)*"
)
self.content_repatter6 = re.compile(
r"((0|[1-9]\d*|[1-9]\d{0,2}(,\d{3})+)*億)*((0|[1-9]\d*|[1-9]\d{0,2}(,\d{3})+)*万)*((0|[1-9]\d*|[1-9]\d{0,2}(,\d{3})+)*千)*(0|[1-9]\d*|[1-9]\d{0,2}(,\d{3})+)*(千円|万円|千万円|円|千ドル|万ドル|千万ドル|ドル|千ユーロ|万ユーロ|千万ユーロ|ユーロ)+(\(税込\)|\(税抜\)|\+tax)*"
)
keisen = "─━│┃┄┅┆┇┈┉┊┋┌┍┎┏┐┑┒┓└┕┖┗┘┙┚┛├┝┞┟┠┡┢┣┤┥┦┧┨┩┪┫┬┭┮┯┰┱┲┳┴┵┶┷┸┹┺┻┼┽┾┿╀╁╂╃╄╅╆╇╈╉╊╋╌╍╎╏═║╒╓╔╕╖╗╘╙╚╛╜╝╞╟╠╡╢╣╤╥╦╧╨╩╪╫╬╭╮╯╰╱╲╳╴╵╶╷╸╹╺╻╼╽╾╿"
blocks = "▀▁▂▃▄▅▆▇█▉▊▋▌▍▎▏▐░▒▓▔▕▖▗▘▙▚▛▜▝▞▟"
self.content_trans1 = str.maketrans({k: "<BLOCK>" for k in keisen + blocks})
def __len__(self):
return len(self.ids_to_tokens)
def clean_text(self, content):
content = self.content_repatter1.sub("<URL>", content)
content = self.content_repatter2.sub("<EMAIL>", content)
content = self.content_repatter3.sub("<TEL>", content)
content = self.content_repatter4.sub("<DATE>", content)
content = self.content_repatter5.sub("<DATE>", content)
content = self.content_repatter6.sub("<PRICE>", content)
content = content.translate(self.content_trans1)
while "<BLOCK><BLOCK>" in content:
content = content.replace("<BLOCK><BLOCK>", "<BLOCK>")
return content
def tokenize(self, text, clean=False):
text = text.replace(" ", "<SP>")
text = text.replace(" ", "<SP>")
text = text.replace("\r\n", "<BR>")
text = text.replace("\n", "<BR>")
text = text.replace("\r", "<BR>")
text = text.replace("\t", "<TAB>")
text = text.replace("—", "ー")
text = text.replace("−", "ー")
for k, v in self.emoji["emoji"].items():
if k in text:
text = text.replace(k, v)
if clean:
text = self.clean_text(text)
def check_simbol(x):
e = x.encode()
if len(x) == 1 and len(e) == 2:
c = (int(e[0]) << 8) + int(e[1])
if (
(c >= 0xC2A1 and c <= 0xC2BF)
or (c >= 0xC780 and c <= 0xC783)
or (c >= 0xCAB9 and c <= 0xCBBF)
or (c >= 0xCC80 and c <= 0xCDA2)
):
return True
return False
def checku2e(x):
e = x.encode()
if len(x) == 1 and len(e) == 3:
c = (int(e[0]) << 16) + (int(e[1]) << 8) + int(e[2])
if c >= 0xE28080 and c <= 0xE2B07F:
return True
return False
pos = 0
result = []
while pos < len(text):
end = min(len(text), pos + self.maxlen + 1) if text[pos] == "<" else pos + 3
candidates = [] # (token_id, token, pos)
for e in range(end, pos, -1):
wd = text[pos:e]
if wd in self.vocab:
if wd[0] == "<" and len(wd) > 2:
candidates = [(self.vocab[wd], wd, e)]
break
else:
candidates.append((self.vocab[wd], wd, e))
if len(candidates) > 0:
# the smallest token_id is adopted
_, wd, e = sorted(candidates, key=lambda x: x[0])[0]
result.append(wd)
pos = e
else:
end = pos + 1
wd = text[pos:end]
if check_simbol(wd):
result.append("<KIGOU>")
elif checku2e(wd):
result.append("<U2000U2BFF>")
else:
for i in wd.encode("utf-8"):
result.append("<|byte%d|>" % i)
pos = end
return result
def convert_id_to_token(self, index):
return self.ids_to_tokens[index][0]
|
transformers/src/transformers/models/deprecated/gptsan_japanese/tokenization_gptsan_japanese.py/0
|
{
"file_path": "transformers/src/transformers/models/deprecated/gptsan_japanese/tokenization_gptsan_japanese.py",
"repo_id": "transformers",
"token_count": 10988
}
| 352
|
# coding=utf-8
# Copyright 2018 Google AI, Google Brain and Carnegie Mellon University 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.
"""
Tokenization classes for Transformer XL model. Adapted from https://github.com/kimiyoung/transformer-xl.
"""
import glob
import os
import pickle
import re
from collections import Counter, OrderedDict
from typing import List, Optional, Tuple
import numpy as np
from ....tokenization_utils import PreTrainedTokenizer
from ....utils import (
cached_file,
is_sacremoses_available,
is_torch_available,
logging,
requires_backends,
strtobool,
torch_only_method,
)
if is_sacremoses_available():
import sacremoses as sm
if is_torch_available():
import torch
logger = logging.get_logger(__name__)
VOCAB_FILES_NAMES = {
"pretrained_vocab_file": "vocab.pkl",
"pretrained_vocab_file_torch": "vocab.bin",
"vocab_file": "vocab.txt",
}
PRETRAINED_CORPUS_ARCHIVE_MAP = {
"transfo-xl/transfo-xl-wt103": "https://huggingface.co/transfo-xl/transfo-xl-wt103/resolve/main/corpus.bin",
}
CORPUS_NAME = "corpus.bin"
MATCH_NUMBERS = r"(?<=\d)[,.](?=\d)", r" @\g<0>@ "
DETOKENIZE_NUMBERS = [(r" @\,@ ", r","), (r" @\.@ ", r".")]
def tokenize_numbers(text_array: List[str]) -> List[str]:
"""
Splits large comma-separated numbers and floating point values. This is done by replacing commas with ' @,@ ' and
dots with ' @.@ '.
Args:
text_array: An already tokenized text as list.
Returns:
A list of strings with tokenized numbers.
Example:
```python
>>> tokenize_numbers(["$", "5,000", "1.73", "m"])
['$', '5', '@,@', '000', '1', '@.@', '73', 'm']
```"""
tokenized = []
for i in range(len(text_array)):
reg, sub = MATCH_NUMBERS
replaced = re.sub(reg, sub, text_array[i]).split()
tokenized.extend(replaced)
return tokenized
def detokenize_numbers(text: str) -> str:
"""
Inverts the operation of *tokenize_numbers*. This is replacing ' @,@ ' and ' @.@' by ',' and '.'.
Args:
text: A string where the number should be detokenized.
Returns:
A detokenized string.
Example:
```python
>>> detokenize_numbers("$ 5 @,@ 000 1 @.@ 73 m")
'$ 5,000 1.73 m'
```"""
for reg, sub in DETOKENIZE_NUMBERS:
text = re.sub(reg, sub, text)
return text
class TransfoXLTokenizer(PreTrainedTokenizer):
"""
Construct a Transformer-XL tokenizer adapted from Vocab class in [the original
code](https://github.com/kimiyoung/transformer-xl). The Transformer-XL tokenizer is a word-level tokenizer (no
sub-word tokenization).
This tokenizer inherits from [`PreTrainedTokenizer`] which contains most of the main methods. Users should refer to
this superclass for more information regarding those methods.
Args:
special (`List[str]`, *optional*):
A list of special tokens (to be treated by the original implementation of this tokenizer).
min_freq (`int`, *optional*, defaults to 0):
The minimum number of times a token has to be present in order to be kept in the vocabulary (otherwise it
will be mapped to `unk_token`).
max_size (`int`, *optional*):
The maximum size of the vocabulary. If left unset, it will default to the size of the vocabulary found
after excluding the tokens according to the `min_freq` rule.
lower_case (`bool`, *optional*, defaults to `False`):
Whether or not to lowercase the input when tokenizing.
delimiter (`str`, *optional*):
The delimiter used between tokens.
vocab_file (`str`, *optional*):
File containing the vocabulary (from the original implementation).
pretrained_vocab_file (`str`, *optional*):
File containing the vocabulary as saved with the `save_pretrained()` method.
never_split (`List[str]`, *optional*):
List of tokens that should never be split. If no list is specified, will simply use the existing special
tokens.
unk_token (`str`, *optional*, defaults to `"<unk>"`):
The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this
token instead.
eos_token (`str`, *optional*, defaults to `"<eos>"`):
The end of sequence token.
additional_special_tokens (`List[str]`, *optional*, defaults to `['<formula>']`):
A list of additional special tokens (for the HuggingFace functionality).
language (`str`, *optional*, defaults to `"en"`):
The language of this tokenizer (used for mose preprocessing).
"""
vocab_files_names = VOCAB_FILES_NAMES
model_input_names = ["input_ids"]
def __init__(
self,
special=None,
min_freq=0,
max_size=None,
lower_case=False,
delimiter=None,
vocab_file=None,
pretrained_vocab_file: str = None,
never_split=None,
unk_token="<unk>",
eos_token="<eos>",
additional_special_tokens=["<formula>"],
language="en",
**kwargs,
):
logger.error(
"`TransfoXL` was deprecated due to security issues linked to `pickle.load` in `TransfoXLTokenizer`. "
"See more details on this model's documentation page: "
"`https://github.com/huggingface/transformers/blob/main/docs/source/en/model_doc/transfo-xl.md`."
)
requires_backends(self, "sacremoses")
if special is None:
special = []
self.counter = Counter()
self.special = special
self.min_freq = min_freq
self.max_size = max_size
self.lower_case = lower_case
self.delimiter = delimiter
self.vocab_file = vocab_file
self.punctuation_symbols = '!"#$%&()*+,-./\\:;<=>?@[\\]^_`{|}~'
self.punction_without_space_before_pattern = re.compile(rf"[^\s][{self.punctuation_symbols}]")
self.punctuation_with_space_around_pattern = self._compile_space_around_punctuation_pattern()
self.language = language
self.moses_punct_normalizer = sm.MosesPunctNormalizer(language)
self.moses_tokenizer = sm.MosesTokenizer(language)
self.moses_detokenizer = sm.MosesDetokenizer(language)
self.idx2sym = []
self.sym2idx = OrderedDict()
# This try... catch... is not beautiful but honestly this tokenizer was not made to be used
# in a library like ours, at all.
try:
vocab_dict = None
if pretrained_vocab_file is not None:
# Priority on pickle files (support PyTorch and TF)
if not strtobool(os.environ.get("TRUST_REMOTE_CODE", "False")):
raise ValueError(
"This part uses `pickle.load` which is insecure and will execute arbitrary code that is "
"potentially malicious. It's recommended to never unpickle data that could have come from an "
"untrusted source, or that could have been tampered with. If you already verified the pickle "
"data and decided to use it, you can set the environment variable "
"`TRUST_REMOTE_CODE` to `True` to allow it."
)
with open(pretrained_vocab_file, "rb") as f:
vocab_dict = pickle.load(f)
# Loading a torch-saved transfo-xl vocab dict with pickle results in an integer
# Entering this if statement means that we tried to load a torch-saved file with pickle, and we failed.
# We therefore load it with torch, if it's available.
if isinstance(vocab_dict, int):
if not is_torch_available():
raise ImportError(
"Not trying to load dict with PyTorch as you need to install pytorch to load "
"from a PyTorch pretrained vocabulary, "
"or activate it with environment variables USE_TORCH=1 and USE_TF=0."
)
vocab_dict = torch.load(pretrained_vocab_file)
if vocab_dict is not None:
for key, value in vocab_dict.items():
if key not in self.__dict__ or key in ["sym2idx", "idx2sym"]:
self.__dict__[key] = value
elif vocab_file is not None:
self.build_vocab()
except Exception as e:
raise ValueError(
f"Unable to parse file {pretrained_vocab_file}. Unknown format. "
"If you tried to load a model saved through TransfoXLTokenizerFast, "
"please note they are not compatible."
) from e
if vocab_file is not None:
self.build_vocab()
super().__init__(
special=special,
min_freq=min_freq,
max_size=max_size,
lower_case=lower_case,
delimiter=delimiter,
vocab_file=vocab_file,
pretrained_vocab_file=pretrained_vocab_file,
never_split=never_split,
unk_token=unk_token,
eos_token=eos_token,
additional_special_tokens=additional_special_tokens,
language=language,
**kwargs,
)
# these are not required to initialize the parent class as only used when tokenizing.
if never_split is None:
never_split = self.all_special_tokens
self.never_split = never_split
@property
def do_lower_case(self):
return self.lower_case
def _compile_space_around_punctuation_pattern(self):
look_ahead_for_special_token = f"(?=[{self.punctuation_symbols}])"
look_ahead_to_match_all_except_space = r"(?=[^\s])"
return re.compile(r"" + look_ahead_for_special_token + look_ahead_to_match_all_except_space)
def count_file(self, path, verbose=False, add_eos=False):
if verbose:
logger.info(f"counting file {path} ...")
assert os.path.exists(path), f"Input file {path} not found"
sents = []
with open(path, "r", encoding="utf-8") as f:
for idx, line in enumerate(f):
if verbose and idx > 0 and idx % 500000 == 0:
logger.info(f" line {idx}")
symbols = self.tokenize(line, add_eos=add_eos)
self.counter.update(symbols)
sents.append(symbols)
return sents
def count_sents(self, sents, verbose=False):
"""
sents : a list of sentences, each a list of tokenized symbols
"""
if verbose:
logger.info(f"counting {len(sents)} sents ...")
for idx, symbols in enumerate(sents):
if verbose and idx > 0 and idx % 500000 == 0:
logger.info(f" line {idx}")
self.counter.update(symbols)
def _build_from_file(self, vocab_file):
self.idx2sym = []
self.sym2idx = OrderedDict()
with open(vocab_file, "r", encoding="utf-8") as f:
for line in f:
symb = line.strip().split()[0]
self.add_symbol(symb)
if "<UNK>" in self.sym2idx:
self.unk_idx = self.sym2idx["<UNK>"]
elif "<unk>" in self.sym2idx:
self.unk_idx = self.sym2idx["<unk>"]
else:
raise ValueError("Token not in vocabulary and no <unk> token in vocabulary for replacement.")
def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
if os.path.isdir(save_directory):
vocab_file = os.path.join(
save_directory,
(filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["pretrained_vocab_file"],
)
else:
vocab_file = (filename_prefix + "-" if filename_prefix else "") + save_directory
with open(vocab_file, "wb") as f:
pickle.dump(self.__dict__, f)
return (vocab_file,)
def build_vocab(self):
if self.vocab_file:
logger.info(f"building vocab from {self.vocab_file}")
self._build_from_file(self.vocab_file)
logger.info(f"Final vocab size {len(self.sym2idx)}")
else:
logger.info(f"building vocab with min_freq={self.min_freq}, max_size={self.max_size}")
self.idx2sym = []
self.sym2idx = OrderedDict()
for sym in self.special:
self.add_special(sym)
for sym, cnt in self.counter.most_common(self.max_size):
if cnt < self.min_freq:
break
self.add_symbol(sym)
logger.info(f"Final vocab size {len(self.sym2idx)} from {len(self.counter)} unique tokens")
@torch_only_method
def encode_file(self, path, ordered=False, verbose=False, add_eos=True, add_double_eos=False):
if verbose:
logger.info(f"encoding file {path} ...")
assert os.path.exists(path), f"Output file {path} not found"
encoded = []
with open(path, "r", encoding="utf-8") as f:
for idx, line in enumerate(f):
if verbose and idx > 0 and idx % 500000 == 0:
logger.info(f" line {idx}")
symbols = self.tokenize(line, add_eos=add_eos, add_double_eos=add_double_eos)
encoded.append(self.convert_to_tensor(symbols))
if ordered:
encoded = torch.cat(encoded)
return encoded
@torch_only_method
def encode_sents(self, sents, ordered=False, verbose=False):
if verbose:
logger.info(f"encoding {len(sents)} sents ...")
encoded = []
for idx, symbols in enumerate(sents):
if verbose and idx > 0 and idx % 500000 == 0:
logger.info(f" line {idx}")
encoded.append(self.convert_to_tensor(symbols))
if ordered:
encoded = torch.cat(encoded)
return encoded
def add_special(self, sym):
if sym not in self.sym2idx:
self.idx2sym.append(sym)
self.sym2idx[sym] = len(self.idx2sym) - 1
setattr(self, f"{sym.strip('<>')}_idx", self.sym2idx[sym])
def add_symbol(self, sym):
if sym not in self.sym2idx:
self.idx2sym.append(sym)
self.sym2idx[sym] = len(self.idx2sym) - 1
def move_added_token(self, token: str, target_idx: int):
"""
Moves an added token to a specific position in the vocab. This method should be used when resizing an embedding
layer other than the last one in the `AdaptiveEmbedding` in order to move the token in the tokenizer from the
default position (at the very end) to the desired one.
Args:
token: The token to move to a specific position in the vocab.
target_idx: The position where the token should be moved to.
"""
assert token in self.added_tokens_encoder, "Token which should be moved has to be an added token"
assert token not in self.idx2sym, "Token which should be moved is already in vocab"
# Insert sym into vocab
self.idx2sym.insert(target_idx, token)
self.sym2idx[token] = target_idx
# Shift following indices in sym2idx
for idx in range(target_idx + 1, len(self.idx2sym)):
current_sym = self.idx2sym[idx]
self.sym2idx[current_sym] = idx
# Delete token from added_tokens
old_index = self._added_tokens_encoder.pop(token)
self._added_tokens_decoder.pop(old_index)
def moses_punct_norm(self, text):
return self.moses_punct_normalizer.normalize(text)
def moses_tokenize(self, text):
return self.moses_tokenizer.tokenize(
text, aggressive_dash_splits=True, return_str=False, escape=False, protected_patterns=self.never_split
)
def moses_pipeline(self, text: str) -> List[str]:
"""
Does basic tokenization using [`sacremoses.MosesPunctNormalizer`] and [`sacremoses.MosesTokenizer`] with
*aggressive_dash_splits=True* (see [`sacremoses.tokenize.MosesTokenizer.tokenize`]). Additionally, large
comma-separated numbers and floating point values are split. E.g. "23,000 people are 1.80m tall" -> "23 @,@ 000
people are 1 @.@ 80m tall"
Args:
text: Text to be tokenize
Returns:
A list of tokenized string
Example:
```python
>>> tokenizer = TransfoXLTokenizer.from_pretrained("transfo-xl/transfo-xl-wt103")
>>> tokenizer.moses_pipeline("23,000 people are 1.80 m tall")
['23', '@,@', '000', 'people', 'are', '1', '@.@', '80', 'm', 'tall']
```"""
text = self.moses_punct_norm(text)
text = self.moses_tokenize(text)
text = tokenize_numbers(text)
return text
def _convert_id_to_token(self, idx):
"""Converts an id in a token (BPE) using the vocab."""
assert 0 <= idx < len(self), f"Index {idx} out of vocabulary range"
return self.idx2sym[idx]
def _convert_token_to_id(self, sym):
"""Converts a token (str) in an id using the vocab."""
if sym in self.sym2idx:
return self.sym2idx[sym]
else:
# logger.info(f'encounter unk {sym}')
# assert '<eos>' not in sym
if hasattr(self, "unk_idx"):
return self.sym2idx.get(sym, self.unk_idx)
# Backward compatibility with pre-trained models
elif "<unk>" in self.sym2idx:
return self.sym2idx["<unk>"]
elif "<UNK>" in self.sym2idx:
return self.sym2idx["<UNK>"]
else:
raise ValueError("Token not in vocabulary and no <unk> token in vocabulary for replacement.")
def convert_tokens_to_string(self, tokens):
"""
Converts a sequence of tokens (string) in a single string. Additionally, the split numbers are converted back
into it's original form.
"""
out_string = self.moses_detokenizer.detokenize(tokens)
return detokenize_numbers(out_string).strip()
@torch_only_method
def convert_to_tensor(self, symbols):
return torch.LongTensor(self.convert_tokens_to_ids(symbols))
@property
def vocab_size(self):
return len(self.idx2sym)
def get_vocab(self):
vocab = self.sym2idx.copy()
vocab.update(self.added_tokens_encoder)
return vocab
def _tokenize(self, line, add_eos=False, add_double_eos=False):
line = line.strip()
# convert to lower case
if self.lower_case:
line = line.lower()
# empty delimiter '' will evaluate False
if self.delimiter == "":
symbols = line
else:
symbols = self.moses_pipeline(line)
if add_double_eos: # lm1b
return ["<S>"] + symbols + ["<S>"]
elif add_eos:
return symbols + ["<eos>"]
else:
return symbols
class LMOrderedIterator:
def __init__(self, data, bsz, bptt, device="cpu", ext_len=None):
"""
data -- LongTensor -- the LongTensor is strictly ordered
"""
self.bsz = bsz
self.bptt = bptt
self.ext_len = ext_len if ext_len is not None else 0
self.device = device
# Work out how cleanly we can divide the dataset into bsz parts.
self.n_step = data.size(0) // bsz
# Trim off any extra elements that wouldn't cleanly fit (remainders).
data = data.narrow(0, 0, self.n_step * bsz)
# Evenly divide the data across the bsz batches.
self.data = data.view(bsz, -1).t().contiguous().to(device)
# Number of mini-batches
self.n_batch = (self.n_step + self.bptt - 1) // self.bptt
def get_batch(self, i, bptt=None):
if bptt is None:
bptt = self.bptt
seq_len = min(bptt, self.data.size(0) - 1 - i)
end_idx = i + seq_len
beg_idx = max(0, i - self.ext_len)
data = self.data[beg_idx:end_idx]
target = self.data[i + 1 : i + 1 + seq_len]
data_out = data.transpose(0, 1).contiguous().to(self.device)
target_out = target.transpose(0, 1).contiguous().to(self.device)
return data_out, target_out, seq_len
def get_fixlen_iter(self, start=0):
for i in range(start, self.data.size(0) - 1, self.bptt):
yield self.get_batch(i)
def get_varlen_iter(self, start=0, std=5, min_len=5, max_deviation=3):
max_len = self.bptt + max_deviation * std
i = start
while True:
bptt = self.bptt if np.random.random() < 0.95 else self.bptt / 2.0
bptt = min(max_len, max(min_len, int(np.random.normal(bptt, std))))
data, target, seq_len = self.get_batch(i, bptt)
i += seq_len
yield data, target, seq_len
if i >= self.data.size(0) - 2:
break
def __iter__(self):
return self.get_fixlen_iter()
class LMShuffledIterator:
def __init__(self, data, bsz, bptt, device="cpu", ext_len=None, shuffle=False):
"""
data -- list[LongTensor] -- there is no order among the LongTensors
"""
self.data = data
self.bsz = bsz
self.bptt = bptt
self.ext_len = ext_len if ext_len is not None else 0
self.device = device
self.shuffle = shuffle
def get_sent_stream(self):
# index iterator
epoch_indices = np.random.permutation(len(self.data)) if self.shuffle else np.array(range(len(self.data)))
# sentence iterator
for idx in epoch_indices:
yield self.data[idx]
@torch_only_method
def stream_iterator(self, sent_stream):
# streams for each data in the batch
streams = [None] * self.bsz
data = torch.LongTensor(self.bptt, self.bsz)
target = torch.LongTensor(self.bptt, self.bsz)
n_retain = 0
while True:
# data : [n_retain+bptt x bsz]
# target : [bptt x bsz]
data[n_retain:].fill_(-1)
target.fill_(-1)
valid_batch = True
for i in range(self.bsz):
n_filled = 0
try:
while n_filled < self.bptt:
if streams[i] is None or len(streams[i]) <= 1:
streams[i] = next(sent_stream)
# number of new tokens to fill in
n_new = min(len(streams[i]) - 1, self.bptt - n_filled)
# first n_retain tokens are retained from last batch
data[n_retain + n_filled : n_retain + n_filled + n_new, i] = streams[i][:n_new]
target[n_filled : n_filled + n_new, i] = streams[i][1 : n_new + 1]
streams[i] = streams[i][n_new:]
n_filled += n_new
except StopIteration:
valid_batch = False
break
if not valid_batch:
return
data_out = data.transpose(0, 1).contiguous().to(self.device)
target_out = target.transpose(0, 1).contiguous().to(self.device)
yield data_out, target_out, self.bptt
n_retain = min(data.size(0), self.ext_len)
if n_retain > 0:
data[:n_retain] = data[-n_retain:]
data.resize_(n_retain + self.bptt, data.size(1))
def __iter__(self):
# sent_stream is an iterator
sent_stream = self.get_sent_stream()
for batch in self.stream_iterator(sent_stream):
yield batch
class LMMultiFileIterator(LMShuffledIterator):
def __init__(self, paths, vocab, bsz, bptt, device="cpu", ext_len=None, shuffle=False):
self.paths = paths
self.vocab = vocab
self.bsz = bsz
self.bptt = bptt
self.ext_len = ext_len if ext_len is not None else 0
self.device = device
self.shuffle = shuffle
def get_sent_stream(self, path):
sents = self.vocab.encode_file(path, add_double_eos=True)
if self.shuffle:
np.random.shuffle(sents)
sent_stream = iter(sents)
return sent_stream
def __iter__(self):
if self.shuffle:
np.random.shuffle(self.paths)
for path in self.paths:
# sent_stream is an iterator
sent_stream = self.get_sent_stream(path)
for batch in self.stream_iterator(sent_stream):
yield batch
class TransfoXLCorpus:
@classmethod
@torch_only_method
def from_pretrained(cls, pretrained_model_name_or_path, cache_dir=None, *inputs, **kwargs):
"""
Instantiate a pre-processed corpus.
"""
vocab = TransfoXLTokenizer.from_pretrained(pretrained_model_name_or_path, *inputs, **kwargs)
is_local = os.path.isdir(pretrained_model_name_or_path)
# redirect to the cache, if necessary
try:
resolved_corpus_file = cached_file(pretrained_model_name_or_path, CORPUS_NAME, cache_dir=cache_dir)
except EnvironmentError:
logger.error(
f"Corpus '{pretrained_model_name_or_path}' was not found in corpus list"
f" ({', '.join(PRETRAINED_CORPUS_ARCHIVE_MAP.keys())}. We assumed '{pretrained_model_name_or_path}'"
f" was a path or url but couldn't find files {CORPUS_NAME} at this path or url."
)
return None
if is_local:
logger.info(f"loading corpus file {resolved_corpus_file}")
else:
logger.info(f"loading corpus file {CORPUS_NAME} from cache at {resolved_corpus_file}")
# Instantiate tokenizer.
corpus = cls(*inputs, **kwargs)
corpus_dict = torch.load(resolved_corpus_file)
for key, value in corpus_dict.items():
corpus.__dict__[key] = value
corpus.vocab = vocab
if corpus.train is not None:
corpus.train = torch.tensor(corpus.train, dtype=torch.long)
if corpus.valid is not None:
corpus.valid = torch.tensor(corpus.valid, dtype=torch.long)
if corpus.test is not None:
corpus.test = torch.tensor(corpus.test, dtype=torch.long)
return corpus
def __init__(self, *args, **kwargs):
self.vocab = TransfoXLTokenizer(*args, **kwargs)
self.dataset = None
self.train = None
self.valid = None
self.test = None
def build_corpus(self, path, dataset):
self.dataset = dataset
if self.dataset in ["ptb", "wt2", "enwik8", "text8"]:
self.vocab.count_file(os.path.join(path, "train.txt"))
self.vocab.count_file(os.path.join(path, "valid.txt"))
self.vocab.count_file(os.path.join(path, "test.txt"))
elif self.dataset == "wt103":
self.vocab.count_file(os.path.join(path, "train.txt"))
elif self.dataset == "lm1b":
train_path_pattern = os.path.join(
path,
"1-billion-word-language-modeling-benchmark-r13output",
"training-monolingual.tokenized.shuffled",
"news.en-*",
)
train_paths = glob.glob(train_path_pattern)
# the vocab will load from file when build_vocab() is called
self.vocab.build_vocab()
if self.dataset in ["ptb", "wt2", "wt103"]:
self.train = self.vocab.encode_file(os.path.join(path, "train.txt"), ordered=True)
self.valid = self.vocab.encode_file(os.path.join(path, "valid.txt"), ordered=True)
self.test = self.vocab.encode_file(os.path.join(path, "test.txt"), ordered=True)
elif self.dataset in ["enwik8", "text8"]:
self.train = self.vocab.encode_file(os.path.join(path, "train.txt"), ordered=True, add_eos=False)
self.valid = self.vocab.encode_file(os.path.join(path, "valid.txt"), ordered=True, add_eos=False)
self.test = self.vocab.encode_file(os.path.join(path, "test.txt"), ordered=True, add_eos=False)
elif self.dataset == "lm1b":
self.train = train_paths
self.valid = self.vocab.encode_file(os.path.join(path, "valid.txt"), ordered=False, add_double_eos=True)
self.test = self.vocab.encode_file(os.path.join(path, "test.txt"), ordered=False, add_double_eos=True)
def get_iterator(self, split, *args, **kwargs):
if split == "train":
if self.dataset in ["ptb", "wt2", "wt103", "enwik8", "text8"]:
data_iter = LMOrderedIterator(self.train, *args, **kwargs)
elif self.dataset == "lm1b":
kwargs["shuffle"] = True
data_iter = LMMultiFileIterator(self.train, self.vocab, *args, **kwargs)
elif split in ["valid", "test"]:
data = self.valid if split == "valid" else self.test
if self.dataset in ["ptb", "wt2", "wt103", "enwik8", "text8"]:
data_iter = LMOrderedIterator(data, *args, **kwargs)
elif self.dataset == "lm1b":
data_iter = LMShuffledIterator(data, *args, **kwargs)
else:
data_iter = None
raise ValueError(f"Split not recognized: {split}")
return data_iter
@torch_only_method
def get_lm_corpus(datadir, dataset):
fn = os.path.join(datadir, "cache.pt")
fn_pickle = os.path.join(datadir, "cache.pkl")
if os.path.exists(fn):
logger.info("Loading cached dataset...")
corpus = torch.load(fn_pickle)
elif os.path.exists(fn):
logger.info("Loading cached dataset from pickle...")
if not strtobool(os.environ.get("TRUST_REMOTE_CODE", "False")):
raise ValueError(
"This part uses `pickle.load` which is insecure and will execute arbitrary code that is potentially "
"malicious. It's recommended to never unpickle data that could have come from an untrusted source, or "
"that could have been tampered with. If you already verified the pickle data and decided to use it, "
"you can set the environment variable `TRUST_REMOTE_CODE` to `True` to allow it."
)
with open(fn, "rb") as fp:
corpus = pickle.load(fp)
else:
logger.info(f"Producing dataset {dataset}...")
kwargs = {}
if dataset in ["wt103", "wt2"]:
kwargs["special"] = ["<eos>"]
kwargs["lower_case"] = False
elif dataset == "ptb":
kwargs["special"] = ["<eos>"]
kwargs["lower_case"] = True
elif dataset == "lm1b":
kwargs["special"] = []
kwargs["lower_case"] = False
kwargs["vocab_file"] = os.path.join(datadir, "1b_word_vocab.txt")
elif dataset in ["enwik8", "text8"]:
pass
corpus = TransfoXLCorpus(datadir, dataset, **kwargs)
torch.save(corpus, fn)
return corpus
|
transformers/src/transformers/models/deprecated/transfo_xl/tokenization_transfo_xl.py/0
|
{
"file_path": "transformers/src/transformers/models/deprecated/transfo_xl/tokenization_transfo_xl.py",
"repo_id": "transformers",
"token_count": 14955
}
| 353
|
# Copyright 2020 The HuggingFace Team. 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.
import argparse
import os
import torch
from transformers.utils import WEIGHTS_NAME
DIALOGPT_MODELS = ["small", "medium", "large"]
OLD_KEY = "lm_head.decoder.weight"
NEW_KEY = "lm_head.weight"
def convert_dialogpt_checkpoint(checkpoint_path: str, pytorch_dump_folder_path: str):
d = torch.load(checkpoint_path)
d[NEW_KEY] = d.pop(OLD_KEY)
os.makedirs(pytorch_dump_folder_path, exist_ok=True)
torch.save(d, os.path.join(pytorch_dump_folder_path, WEIGHTS_NAME))
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--dialogpt_path", default=".", type=str)
args = parser.parse_args()
for MODEL in DIALOGPT_MODELS:
checkpoint_path = os.path.join(args.dialogpt_path, f"{MODEL}_ft.pkl")
pytorch_dump_folder_path = f"./DialoGPT-{MODEL}"
convert_dialogpt_checkpoint(
checkpoint_path,
pytorch_dump_folder_path,
)
|
transformers/src/transformers/models/dialogpt/convert_dialogpt_original_pytorch_checkpoint_to_pytorch.py/0
|
{
"file_path": "transformers/src/transformers/models/dialogpt/convert_dialogpt_original_pytorch_checkpoint_to_pytorch.py",
"repo_id": "transformers",
"token_count": 560
}
| 354
|
# Copyright 2021 AlQuraishi Laboratory
# Copyright 2021 DeepMind Technologies Limited
#
# 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.
"""Constants used in AlphaFold."""
import collections
import copy
import functools
from importlib import resources
from typing import Dict, List, Mapping, Sequence, Tuple
import numpy as np
# Internal import (35fd).
# Distance from one CA to next CA [trans configuration: omega = 180].
ca_ca = 3.80209737096
# Format: The list for each AA type contains chi1, chi2, chi3, chi4 in
# this order (or a relevant subset from chi1 onwards). ALA and GLY don't have
# chi angles so their chi angle lists are empty.
chi_angles_atoms: Dict[str, List[List[str]]] = {
"ALA": [],
# Chi5 in arginine is always 0 +- 5 degrees, so ignore it.
"ARG": [["N", "CA", "CB", "CG"], ["CA", "CB", "CG", "CD"], ["CB", "CG", "CD", "NE"], ["CG", "CD", "NE", "CZ"]],
"ASN": [["N", "CA", "CB", "CG"], ["CA", "CB", "CG", "OD1"]],
"ASP": [["N", "CA", "CB", "CG"], ["CA", "CB", "CG", "OD1"]],
"CYS": [["N", "CA", "CB", "SG"]],
"GLN": [["N", "CA", "CB", "CG"], ["CA", "CB", "CG", "CD"], ["CB", "CG", "CD", "OE1"]],
"GLU": [["N", "CA", "CB", "CG"], ["CA", "CB", "CG", "CD"], ["CB", "CG", "CD", "OE1"]],
"GLY": [],
"HIS": [["N", "CA", "CB", "CG"], ["CA", "CB", "CG", "ND1"]],
"ILE": [["N", "CA", "CB", "CG1"], ["CA", "CB", "CG1", "CD1"]],
"LEU": [["N", "CA", "CB", "CG"], ["CA", "CB", "CG", "CD1"]],
"LYS": [["N", "CA", "CB", "CG"], ["CA", "CB", "CG", "CD"], ["CB", "CG", "CD", "CE"], ["CG", "CD", "CE", "NZ"]],
"MET": [["N", "CA", "CB", "CG"], ["CA", "CB", "CG", "SD"], ["CB", "CG", "SD", "CE"]],
"PHE": [["N", "CA", "CB", "CG"], ["CA", "CB", "CG", "CD1"]],
"PRO": [["N", "CA", "CB", "CG"], ["CA", "CB", "CG", "CD"]],
"SER": [["N", "CA", "CB", "OG"]],
"THR": [["N", "CA", "CB", "OG1"]],
"TRP": [["N", "CA", "CB", "CG"], ["CA", "CB", "CG", "CD1"]],
"TYR": [["N", "CA", "CB", "CG"], ["CA", "CB", "CG", "CD1"]],
"VAL": [["N", "CA", "CB", "CG1"]],
}
# If chi angles given in fixed-length array, this matrix determines how to mask
# them for each AA type. The order is as per restype_order (see below).
chi_angles_mask: List[List[float]] = [
[0.0, 0.0, 0.0, 0.0], # ALA
[1.0, 1.0, 1.0, 1.0], # ARG
[1.0, 1.0, 0.0, 0.0], # ASN
[1.0, 1.0, 0.0, 0.0], # ASP
[1.0, 0.0, 0.0, 0.0], # CYS
[1.0, 1.0, 1.0, 0.0], # GLN
[1.0, 1.0, 1.0, 0.0], # GLU
[0.0, 0.0, 0.0, 0.0], # GLY
[1.0, 1.0, 0.0, 0.0], # HIS
[1.0, 1.0, 0.0, 0.0], # ILE
[1.0, 1.0, 0.0, 0.0], # LEU
[1.0, 1.0, 1.0, 1.0], # LYS
[1.0, 1.0, 1.0, 0.0], # MET
[1.0, 1.0, 0.0, 0.0], # PHE
[1.0, 1.0, 0.0, 0.0], # PRO
[1.0, 0.0, 0.0, 0.0], # SER
[1.0, 0.0, 0.0, 0.0], # THR
[1.0, 1.0, 0.0, 0.0], # TRP
[1.0, 1.0, 0.0, 0.0], # TYR
[1.0, 0.0, 0.0, 0.0], # VAL
]
# The following chi angles are pi periodic: they can be rotated by a multiple
# of pi without affecting the structure.
chi_pi_periodic: List[List[float]] = [
[0.0, 0.0, 0.0, 0.0], # ALA
[0.0, 0.0, 0.0, 0.0], # ARG
[0.0, 0.0, 0.0, 0.0], # ASN
[0.0, 1.0, 0.0, 0.0], # ASP
[0.0, 0.0, 0.0, 0.0], # CYS
[0.0, 0.0, 0.0, 0.0], # GLN
[0.0, 0.0, 1.0, 0.0], # GLU
[0.0, 0.0, 0.0, 0.0], # GLY
[0.0, 0.0, 0.0, 0.0], # HIS
[0.0, 0.0, 0.0, 0.0], # ILE
[0.0, 0.0, 0.0, 0.0], # LEU
[0.0, 0.0, 0.0, 0.0], # LYS
[0.0, 0.0, 0.0, 0.0], # MET
[0.0, 1.0, 0.0, 0.0], # PHE
[0.0, 0.0, 0.0, 0.0], # PRO
[0.0, 0.0, 0.0, 0.0], # SER
[0.0, 0.0, 0.0, 0.0], # THR
[0.0, 0.0, 0.0, 0.0], # TRP
[0.0, 1.0, 0.0, 0.0], # TYR
[0.0, 0.0, 0.0, 0.0], # VAL
[0.0, 0.0, 0.0, 0.0], # UNK
]
# Atoms positions relative to the 8 rigid groups, defined by the pre-omega, phi,
# psi and chi angles:
# 0: 'backbone group',
# 1: 'pre-omega-group', (empty)
# 2: 'phi-group', (currently empty, because it defines only hydrogens)
# 3: 'psi-group',
# 4,5,6,7: 'chi1,2,3,4-group'
# The atom positions are relative to the axis-end-atom of the corresponding
# rotation axis. The x-axis is in direction of the rotation axis, and the y-axis
# is defined such that the dihedral-angle-definiting atom (the last entry in
# chi_angles_atoms above) is in the xy-plane (with a positive y-coordinate).
# format: [atomname, group_idx, rel_position]
rigid_group_atom_positions: Dict[str, List[Tuple[str, int, Tuple[float, float, float]]]] = {
"ALA": [
("N", 0, (-0.525, 1.363, 0.000)),
("CA", 0, (0.000, 0.000, 0.000)),
("C", 0, (1.526, -0.000, -0.000)),
("CB", 0, (-0.529, -0.774, -1.205)),
("O", 3, (0.627, 1.062, 0.000)),
],
"ARG": [
("N", 0, (-0.524, 1.362, -0.000)),
("CA", 0, (0.000, 0.000, 0.000)),
("C", 0, (1.525, -0.000, -0.000)),
("CB", 0, (-0.524, -0.778, -1.209)),
("O", 3, (0.626, 1.062, 0.000)),
("CG", 4, (0.616, 1.390, -0.000)),
("CD", 5, (0.564, 1.414, 0.000)),
("NE", 6, (0.539, 1.357, -0.000)),
("NH1", 7, (0.206, 2.301, 0.000)),
("NH2", 7, (2.078, 0.978, -0.000)),
("CZ", 7, (0.758, 1.093, -0.000)),
],
"ASN": [
("N", 0, (-0.536, 1.357, 0.000)),
("CA", 0, (0.000, 0.000, 0.000)),
("C", 0, (1.526, -0.000, -0.000)),
("CB", 0, (-0.531, -0.787, -1.200)),
("O", 3, (0.625, 1.062, 0.000)),
("CG", 4, (0.584, 1.399, 0.000)),
("ND2", 5, (0.593, -1.188, 0.001)),
("OD1", 5, (0.633, 1.059, 0.000)),
],
"ASP": [
("N", 0, (-0.525, 1.362, -0.000)),
("CA", 0, (0.000, 0.000, 0.000)),
("C", 0, (1.527, 0.000, -0.000)),
("CB", 0, (-0.526, -0.778, -1.208)),
("O", 3, (0.626, 1.062, -0.000)),
("CG", 4, (0.593, 1.398, -0.000)),
("OD1", 5, (0.610, 1.091, 0.000)),
("OD2", 5, (0.592, -1.101, -0.003)),
],
"CYS": [
("N", 0, (-0.522, 1.362, -0.000)),
("CA", 0, (0.000, 0.000, 0.000)),
("C", 0, (1.524, 0.000, 0.000)),
("CB", 0, (-0.519, -0.773, -1.212)),
("O", 3, (0.625, 1.062, -0.000)),
("SG", 4, (0.728, 1.653, 0.000)),
],
"GLN": [
("N", 0, (-0.526, 1.361, -0.000)),
("CA", 0, (0.000, 0.000, 0.000)),
("C", 0, (1.526, 0.000, 0.000)),
("CB", 0, (-0.525, -0.779, -1.207)),
("O", 3, (0.626, 1.062, -0.000)),
("CG", 4, (0.615, 1.393, 0.000)),
("CD", 5, (0.587, 1.399, -0.000)),
("NE2", 6, (0.593, -1.189, -0.001)),
("OE1", 6, (0.634, 1.060, 0.000)),
],
"GLU": [
("N", 0, (-0.528, 1.361, 0.000)),
("CA", 0, (0.000, 0.000, 0.000)),
("C", 0, (1.526, -0.000, -0.000)),
("CB", 0, (-0.526, -0.781, -1.207)),
("O", 3, (0.626, 1.062, 0.000)),
("CG", 4, (0.615, 1.392, 0.000)),
("CD", 5, (0.600, 1.397, 0.000)),
("OE1", 6, (0.607, 1.095, -0.000)),
("OE2", 6, (0.589, -1.104, -0.001)),
],
"GLY": [
("N", 0, (-0.572, 1.337, 0.000)),
("CA", 0, (0.000, 0.000, 0.000)),
("C", 0, (1.517, -0.000, -0.000)),
("O", 3, (0.626, 1.062, -0.000)),
],
"HIS": [
("N", 0, (-0.527, 1.360, 0.000)),
("CA", 0, (0.000, 0.000, 0.000)),
("C", 0, (1.525, 0.000, 0.000)),
("CB", 0, (-0.525, -0.778, -1.208)),
("O", 3, (0.625, 1.063, 0.000)),
("CG", 4, (0.600, 1.370, -0.000)),
("CD2", 5, (0.889, -1.021, 0.003)),
("ND1", 5, (0.744, 1.160, -0.000)),
("CE1", 5, (2.030, 0.851, 0.002)),
("NE2", 5, (2.145, -0.466, 0.004)),
],
"ILE": [
("N", 0, (-0.493, 1.373, -0.000)),
("CA", 0, (0.000, 0.000, 0.000)),
("C", 0, (1.527, -0.000, -0.000)),
("CB", 0, (-0.536, -0.793, -1.213)),
("O", 3, (0.627, 1.062, -0.000)),
("CG1", 4, (0.534, 1.437, -0.000)),
("CG2", 4, (0.540, -0.785, -1.199)),
("CD1", 5, (0.619, 1.391, 0.000)),
],
"LEU": [
("N", 0, (-0.520, 1.363, 0.000)),
("CA", 0, (0.000, 0.000, 0.000)),
("C", 0, (1.525, -0.000, -0.000)),
("CB", 0, (-0.522, -0.773, -1.214)),
("O", 3, (0.625, 1.063, -0.000)),
("CG", 4, (0.678, 1.371, 0.000)),
("CD1", 5, (0.530, 1.430, -0.000)),
("CD2", 5, (0.535, -0.774, 1.200)),
],
"LYS": [
("N", 0, (-0.526, 1.362, -0.000)),
("CA", 0, (0.000, 0.000, 0.000)),
("C", 0, (1.526, 0.000, 0.000)),
("CB", 0, (-0.524, -0.778, -1.208)),
("O", 3, (0.626, 1.062, -0.000)),
("CG", 4, (0.619, 1.390, 0.000)),
("CD", 5, (0.559, 1.417, 0.000)),
("CE", 6, (0.560, 1.416, 0.000)),
("NZ", 7, (0.554, 1.387, 0.000)),
],
"MET": [
("N", 0, (-0.521, 1.364, -0.000)),
("CA", 0, (0.000, 0.000, 0.000)),
("C", 0, (1.525, 0.000, 0.000)),
("CB", 0, (-0.523, -0.776, -1.210)),
("O", 3, (0.625, 1.062, -0.000)),
("CG", 4, (0.613, 1.391, -0.000)),
("SD", 5, (0.703, 1.695, 0.000)),
("CE", 6, (0.320, 1.786, -0.000)),
],
"PHE": [
("N", 0, (-0.518, 1.363, 0.000)),
("CA", 0, (0.000, 0.000, 0.000)),
("C", 0, (1.524, 0.000, -0.000)),
("CB", 0, (-0.525, -0.776, -1.212)),
("O", 3, (0.626, 1.062, -0.000)),
("CG", 4, (0.607, 1.377, 0.000)),
("CD1", 5, (0.709, 1.195, -0.000)),
("CD2", 5, (0.706, -1.196, 0.000)),
("CE1", 5, (2.102, 1.198, -0.000)),
("CE2", 5, (2.098, -1.201, -0.000)),
("CZ", 5, (2.794, -0.003, -0.001)),
],
"PRO": [
("N", 0, (-0.566, 1.351, -0.000)),
("CA", 0, (0.000, 0.000, 0.000)),
("C", 0, (1.527, -0.000, 0.000)),
("CB", 0, (-0.546, -0.611, -1.293)),
("O", 3, (0.621, 1.066, 0.000)),
("CG", 4, (0.382, 1.445, 0.0)),
# ('CD', 5, (0.427, 1.440, 0.0)),
("CD", 5, (0.477, 1.424, 0.0)), # manually made angle 2 degrees larger
],
"SER": [
("N", 0, (-0.529, 1.360, -0.000)),
("CA", 0, (0.000, 0.000, 0.000)),
("C", 0, (1.525, -0.000, -0.000)),
("CB", 0, (-0.518, -0.777, -1.211)),
("O", 3, (0.626, 1.062, -0.000)),
("OG", 4, (0.503, 1.325, 0.000)),
],
"THR": [
("N", 0, (-0.517, 1.364, 0.000)),
("CA", 0, (0.000, 0.000, 0.000)),
("C", 0, (1.526, 0.000, -0.000)),
("CB", 0, (-0.516, -0.793, -1.215)),
("O", 3, (0.626, 1.062, 0.000)),
("CG2", 4, (0.550, -0.718, -1.228)),
("OG1", 4, (0.472, 1.353, 0.000)),
],
"TRP": [
("N", 0, (-0.521, 1.363, 0.000)),
("CA", 0, (0.000, 0.000, 0.000)),
("C", 0, (1.525, -0.000, 0.000)),
("CB", 0, (-0.523, -0.776, -1.212)),
("O", 3, (0.627, 1.062, 0.000)),
("CG", 4, (0.609, 1.370, -0.000)),
("CD1", 5, (0.824, 1.091, 0.000)),
("CD2", 5, (0.854, -1.148, -0.005)),
("CE2", 5, (2.186, -0.678, -0.007)),
("CE3", 5, (0.622, -2.530, -0.007)),
("NE1", 5, (2.140, 0.690, -0.004)),
("CH2", 5, (3.028, -2.890, -0.013)),
("CZ2", 5, (3.283, -1.543, -0.011)),
("CZ3", 5, (1.715, -3.389, -0.011)),
],
"TYR": [
("N", 0, (-0.522, 1.362, 0.000)),
("CA", 0, (0.000, 0.000, 0.000)),
("C", 0, (1.524, -0.000, -0.000)),
("CB", 0, (-0.522, -0.776, -1.213)),
("O", 3, (0.627, 1.062, -0.000)),
("CG", 4, (0.607, 1.382, -0.000)),
("CD1", 5, (0.716, 1.195, -0.000)),
("CD2", 5, (0.713, -1.194, -0.001)),
("CE1", 5, (2.107, 1.200, -0.002)),
("CE2", 5, (2.104, -1.201, -0.003)),
("OH", 5, (4.168, -0.002, -0.005)),
("CZ", 5, (2.791, -0.001, -0.003)),
],
"VAL": [
("N", 0, (-0.494, 1.373, -0.000)),
("CA", 0, (0.000, 0.000, 0.000)),
("C", 0, (1.527, -0.000, -0.000)),
("CB", 0, (-0.533, -0.795, -1.213)),
("O", 3, (0.627, 1.062, -0.000)),
("CG1", 4, (0.540, 1.429, -0.000)),
("CG2", 4, (0.533, -0.776, 1.203)),
],
}
# A list of atoms (excluding hydrogen) for each AA type. PDB naming convention.
residue_atoms: Dict[str, List[str]] = {
"ALA": ["C", "CA", "CB", "N", "O"],
"ARG": ["C", "CA", "CB", "CG", "CD", "CZ", "N", "NE", "O", "NH1", "NH2"],
"ASP": ["C", "CA", "CB", "CG", "N", "O", "OD1", "OD2"],
"ASN": ["C", "CA", "CB", "CG", "N", "ND2", "O", "OD1"],
"CYS": ["C", "CA", "CB", "N", "O", "SG"],
"GLU": ["C", "CA", "CB", "CG", "CD", "N", "O", "OE1", "OE2"],
"GLN": ["C", "CA", "CB", "CG", "CD", "N", "NE2", "O", "OE1"],
"GLY": ["C", "CA", "N", "O"],
"HIS": ["C", "CA", "CB", "CG", "CD2", "CE1", "N", "ND1", "NE2", "O"],
"ILE": ["C", "CA", "CB", "CG1", "CG2", "CD1", "N", "O"],
"LEU": ["C", "CA", "CB", "CG", "CD1", "CD2", "N", "O"],
"LYS": ["C", "CA", "CB", "CG", "CD", "CE", "N", "NZ", "O"],
"MET": ["C", "CA", "CB", "CG", "CE", "N", "O", "SD"],
"PHE": ["C", "CA", "CB", "CG", "CD1", "CD2", "CE1", "CE2", "CZ", "N", "O"],
"PRO": ["C", "CA", "CB", "CG", "CD", "N", "O"],
"SER": ["C", "CA", "CB", "N", "O", "OG"],
"THR": ["C", "CA", "CB", "CG2", "N", "O", "OG1"],
"TRP": ["C", "CA", "CB", "CG", "CD1", "CD2", "CE2", "CE3", "CZ2", "CZ3", "CH2", "N", "NE1", "O"],
"TYR": ["C", "CA", "CB", "CG", "CD1", "CD2", "CE1", "CE2", "CZ", "N", "O", "OH"],
"VAL": ["C", "CA", "CB", "CG1", "CG2", "N", "O"],
}
# Naming swaps for ambiguous atom names.
# Due to symmetries in the amino acids the naming of atoms is ambiguous in
# 4 of the 20 amino acids.
# (The LDDT paper lists 7 amino acids as ambiguous, but the naming ambiguities
# in LEU, VAL and ARG can be resolved by using the 3d constellations of
# the 'ambiguous' atoms and their neighbours)
# TODO: ^ interpret this
residue_atom_renaming_swaps: Dict[str, Dict[str, str]] = {
"ASP": {"OD1": "OD2"},
"GLU": {"OE1": "OE2"},
"PHE": {"CD1": "CD2", "CE1": "CE2"},
"TYR": {"CD1": "CD2", "CE1": "CE2"},
}
# Van der Waals radii [Angstroem] of the atoms (from Wikipedia)
van_der_waals_radius: Dict[str, float] = {
"C": 1.7,
"N": 1.55,
"O": 1.52,
"S": 1.8,
}
Bond = collections.namedtuple("Bond", ["atom1_name", "atom2_name", "length", "stddev"])
BondAngle = collections.namedtuple(
"BondAngle",
["atom1_name", "atom2_name", "atom3name", "angle_rad", "stddev"],
)
def map_structure_with_atom_order(in_list: list, first_call: bool = True) -> list:
# Maps strings in a nested list structure to their corresponding index in atom_order
if first_call:
in_list = copy.deepcopy(in_list)
for i in range(len(in_list)):
if isinstance(in_list[i], list):
in_list[i] = map_structure_with_atom_order(in_list[i], first_call=False)
elif isinstance(in_list[i], str):
in_list[i] = atom_order[in_list[i]]
else:
raise TypeError("Unexpected type when mapping nested lists!")
return in_list
@functools.lru_cache(maxsize=None)
def load_stereo_chemical_props() -> (
Tuple[
Mapping[str, List[Bond]],
Mapping[str, List[Bond]],
Mapping[str, List[BondAngle]],
]
):
"""Load stereo_chemical_props.txt into a nice structure.
Load literature values for bond lengths and bond angles and translate bond angles into the length of the opposite
edge of the triangle ("residue_virtual_bonds").
Returns:
residue_bonds: dict that maps resname --> list of Bond tuples residue_virtual_bonds: dict that maps resname -->
list of Bond tuples residue_bond_angles: dict that maps resname --> list of BondAngle tuples
"""
# TODO: this file should be downloaded in a setup script
stereo_chemical_props = resources.read_text("openfold.resources", "stereo_chemical_props.txt")
lines_iter = iter(stereo_chemical_props.splitlines())
# Load bond lengths.
residue_bonds: Dict[str, List[Bond]] = {}
next(lines_iter) # Skip header line.
for line in lines_iter:
if line.strip() == "-":
break
bond, resname, bond_length, stddev = line.split()
atom1, atom2 = bond.split("-")
if resname not in residue_bonds:
residue_bonds[resname] = []
residue_bonds[resname].append(Bond(atom1, atom2, float(bond_length), float(stddev)))
residue_bonds["UNK"] = []
# Load bond angles.
residue_bond_angles: Dict[str, List[BondAngle]] = {}
next(lines_iter) # Skip empty line.
next(lines_iter) # Skip header line.
for line in lines_iter:
if line.strip() == "-":
break
bond, resname, angle_degree, stddev_degree = line.split()
atom1, atom2, atom3 = bond.split("-")
if resname not in residue_bond_angles:
residue_bond_angles[resname] = []
residue_bond_angles[resname].append(
BondAngle(
atom1,
atom2,
atom3,
float(angle_degree) / 180.0 * np.pi,
float(stddev_degree) / 180.0 * np.pi,
)
)
residue_bond_angles["UNK"] = []
def make_bond_key(atom1_name: str, atom2_name: str) -> str:
"""Unique key to lookup bonds."""
return "-".join(sorted([atom1_name, atom2_name]))
# Translate bond angles into distances ("virtual bonds").
residue_virtual_bonds: Dict[str, List[Bond]] = {}
for resname, bond_angles in residue_bond_angles.items():
# Create a fast lookup dict for bond lengths.
bond_cache: Dict[str, Bond] = {}
for b in residue_bonds[resname]:
bond_cache[make_bond_key(b.atom1_name, b.atom2_name)] = b
residue_virtual_bonds[resname] = []
for ba in bond_angles:
bond1 = bond_cache[make_bond_key(ba.atom1_name, ba.atom2_name)]
bond2 = bond_cache[make_bond_key(ba.atom2_name, ba.atom3name)]
# Compute distance between atom1 and atom3 using the law of cosines
# c^2 = a^2 + b^2 - 2ab*cos(gamma).
gamma = ba.angle_rad
length = np.sqrt(bond1.length**2 + bond2.length**2 - 2 * bond1.length * bond2.length * np.cos(gamma))
# Propagation of uncertainty assuming uncorrelated errors.
dl_outer = 0.5 / length
dl_dgamma = (2 * bond1.length * bond2.length * np.sin(gamma)) * dl_outer
dl_db1 = (2 * bond1.length - 2 * bond2.length * np.cos(gamma)) * dl_outer
dl_db2 = (2 * bond2.length - 2 * bond1.length * np.cos(gamma)) * dl_outer
stddev = np.sqrt(
(dl_dgamma * ba.stddev) ** 2 + (dl_db1 * bond1.stddev) ** 2 + (dl_db2 * bond2.stddev) ** 2
)
residue_virtual_bonds[resname].append(Bond(ba.atom1_name, ba.atom3name, length, stddev))
return (residue_bonds, residue_virtual_bonds, residue_bond_angles)
# Between-residue bond lengths for general bonds (first element) and for Proline
# (second element).
between_res_bond_length_c_n: Tuple[float, float] = (1.329, 1.341)
between_res_bond_length_stddev_c_n: Tuple[float, float] = (0.014, 0.016)
# Between-residue cos_angles.
between_res_cos_angles_c_n_ca: Tuple[float, float] = (-0.5203, 0.0353) # degrees: 121.352 +- 2.315
between_res_cos_angles_ca_c_n: Tuple[float, float] = (-0.4473, 0.0311) # degrees: 116.568 +- 1.995
# This mapping is used when we need to store atom data in a format that requires
# fixed atom data size for every residue (e.g. a numpy array).
atom_types: List[str] = [
"N",
"CA",
"C",
"CB",
"O",
"CG",
"CG1",
"CG2",
"OG",
"OG1",
"SG",
"CD",
"CD1",
"CD2",
"ND1",
"ND2",
"OD1",
"OD2",
"SD",
"CE",
"CE1",
"CE2",
"CE3",
"NE",
"NE1",
"NE2",
"OE1",
"OE2",
"CH2",
"NH1",
"NH2",
"OH",
"CZ",
"CZ2",
"CZ3",
"NZ",
"OXT",
]
atom_order: Dict[str, int] = {atom_type: i for i, atom_type in enumerate(atom_types)}
atom_type_num = len(atom_types) # := 37.
# A compact atom encoding with 14 columns
# pylint: disable=line-too-long
# pylint: disable=bad-whitespace
restype_name_to_atom14_names: Dict[str, List[str]] = {
"ALA": ["N", "CA", "C", "O", "CB", "", "", "", "", "", "", "", "", ""],
"ARG": ["N", "CA", "C", "O", "CB", "CG", "CD", "NE", "CZ", "NH1", "NH2", "", "", ""],
"ASN": ["N", "CA", "C", "O", "CB", "CG", "OD1", "ND2", "", "", "", "", "", ""],
"ASP": ["N", "CA", "C", "O", "CB", "CG", "OD1", "OD2", "", "", "", "", "", ""],
"CYS": ["N", "CA", "C", "O", "CB", "SG", "", "", "", "", "", "", "", ""],
"GLN": ["N", "CA", "C", "O", "CB", "CG", "CD", "OE1", "NE2", "", "", "", "", ""],
"GLU": ["N", "CA", "C", "O", "CB", "CG", "CD", "OE1", "OE2", "", "", "", "", ""],
"GLY": ["N", "CA", "C", "O", "", "", "", "", "", "", "", "", "", ""],
"HIS": ["N", "CA", "C", "O", "CB", "CG", "ND1", "CD2", "CE1", "NE2", "", "", "", ""],
"ILE": ["N", "CA", "C", "O", "CB", "CG1", "CG2", "CD1", "", "", "", "", "", ""],
"LEU": ["N", "CA", "C", "O", "CB", "CG", "CD1", "CD2", "", "", "", "", "", ""],
"LYS": ["N", "CA", "C", "O", "CB", "CG", "CD", "CE", "NZ", "", "", "", "", ""],
"MET": ["N", "CA", "C", "O", "CB", "CG", "SD", "CE", "", "", "", "", "", ""],
"PHE": ["N", "CA", "C", "O", "CB", "CG", "CD1", "CD2", "CE1", "CE2", "CZ", "", "", ""],
"PRO": ["N", "CA", "C", "O", "CB", "CG", "CD", "", "", "", "", "", "", ""],
"SER": ["N", "CA", "C", "O", "CB", "OG", "", "", "", "", "", "", "", ""],
"THR": ["N", "CA", "C", "O", "CB", "OG1", "CG2", "", "", "", "", "", "", ""],
"TRP": ["N", "CA", "C", "O", "CB", "CG", "CD1", "CD2", "NE1", "CE2", "CE3", "CZ2", "CZ3", "CH2"],
"TYR": ["N", "CA", "C", "O", "CB", "CG", "CD1", "CD2", "CE1", "CE2", "CZ", "OH", "", ""],
"VAL": ["N", "CA", "C", "O", "CB", "CG1", "CG2", "", "", "", "", "", "", ""],
"UNK": ["", "", "", "", "", "", "", "", "", "", "", "", "", ""],
}
# pylint: enable=line-too-long
# pylint: enable=bad-whitespace
# This is the standard residue order when coding AA type as a number.
# Reproduce it by taking 3-letter AA codes and sorting them alphabetically.
restypes: List[str] = [
"A",
"R",
"N",
"D",
"C",
"Q",
"E",
"G",
"H",
"I",
"L",
"K",
"M",
"F",
"P",
"S",
"T",
"W",
"Y",
"V",
]
restype_order: Dict[str, int] = {restype: i for i, restype in enumerate(restypes)}
restype_num = len(restypes) # := 20.
unk_restype_index = restype_num # Catch-all index for unknown restypes.
restypes_with_x: List[str] = restypes + ["X"]
restype_order_with_x: Dict[str, int] = {restype: i for i, restype in enumerate(restypes_with_x)}
def sequence_to_onehot(sequence: str, mapping: Mapping[str, int], map_unknown_to_x: bool = False) -> np.ndarray:
"""Maps the given sequence into a one-hot encoded matrix.
Args:
sequence: An amino acid sequence.
mapping: A dictionary mapping amino acids to integers.
map_unknown_to_x: If True, any amino acid that is not in the mapping will be
mapped to the unknown amino acid 'X'. If the mapping doesn't contain amino acid 'X', an error will be thrown.
If False, any amino acid not in the mapping will throw an error.
Returns:
A numpy array of shape (seq_len, num_unique_aas) with one-hot encoding of the sequence.
Raises:
ValueError: If the mapping doesn't contain values from 0 to
num_unique_aas - 1 without any gaps.
"""
num_entries = max(mapping.values()) + 1
if sorted(set(mapping.values())) != list(range(num_entries)):
raise ValueError(
"The mapping must have values from 0 to num_unique_aas-1 without any gaps. Got: %s"
% sorted(mapping.values())
)
one_hot_arr = np.zeros((len(sequence), num_entries), dtype=np.int32)
for aa_index, aa_type in enumerate(sequence):
if map_unknown_to_x:
if aa_type.isalpha() and aa_type.isupper():
aa_id = mapping.get(aa_type, mapping["X"])
else:
raise ValueError(f"Invalid character in the sequence: {aa_type}")
else:
aa_id = mapping[aa_type]
one_hot_arr[aa_index, aa_id] = 1
return one_hot_arr
restype_1to3: Dict[str, str] = {
"A": "ALA",
"R": "ARG",
"N": "ASN",
"D": "ASP",
"C": "CYS",
"Q": "GLN",
"E": "GLU",
"G": "GLY",
"H": "HIS",
"I": "ILE",
"L": "LEU",
"K": "LYS",
"M": "MET",
"F": "PHE",
"P": "PRO",
"S": "SER",
"T": "THR",
"W": "TRP",
"Y": "TYR",
"V": "VAL",
}
# NB: restype_3to1 differs from Bio.PDB.protein_letters_3to1 by being a simple
# 1-to-1 mapping of 3 letter names to one letter names. The latter contains
# many more, and less common, three letter names as keys and maps many of these
# to the same one letter name (including 'X' and 'U' which we don't use here).
restype_3to1: Dict[str, str] = {v: k for k, v in restype_1to3.items()}
# Define a restype name for all unknown residues.
unk_restype = "UNK"
resnames: List[str] = [restype_1to3[r] for r in restypes] + [unk_restype]
resname_to_idx: Dict[str, int] = {resname: i for i, resname in enumerate(resnames)}
# The mapping here uses hhblits convention, so that B is mapped to D, J and O
# are mapped to X, U is mapped to C, and Z is mapped to E. Other than that the
# remaining 20 amino acids are kept in alphabetical order.
# There are 2 non-amino acid codes, X (representing any amino acid) and
# "-" representing a missing amino acid in an alignment. The id for these
# codes is put at the end (20 and 21) so that they can easily be ignored if
# desired.
HHBLITS_AA_TO_ID: Dict[str, int] = {
"A": 0,
"B": 2,
"C": 1,
"D": 2,
"E": 3,
"F": 4,
"G": 5,
"H": 6,
"I": 7,
"J": 20,
"K": 8,
"L": 9,
"M": 10,
"N": 11,
"O": 20,
"P": 12,
"Q": 13,
"R": 14,
"S": 15,
"T": 16,
"U": 1,
"V": 17,
"W": 18,
"X": 20,
"Y": 19,
"Z": 3,
"-": 21,
}
# Partial inversion of HHBLITS_AA_TO_ID.
ID_TO_HHBLITS_AA: Dict[int, str] = {
0: "A",
1: "C", # Also U.
2: "D", # Also B.
3: "E", # Also Z.
4: "F",
5: "G",
6: "H",
7: "I",
8: "K",
9: "L",
10: "M",
11: "N",
12: "P",
13: "Q",
14: "R",
15: "S",
16: "T",
17: "V",
18: "W",
19: "Y",
20: "X", # Includes J and O.
21: "-",
}
restypes_with_x_and_gap: List[str] = restypes + ["X", "-"]
MAP_HHBLITS_AATYPE_TO_OUR_AATYPE: Tuple[int, ...] = tuple(
restypes_with_x_and_gap.index(ID_TO_HHBLITS_AA[i]) for i in range(len(restypes_with_x_and_gap))
)
def _make_standard_atom_mask() -> np.ndarray:
"""Returns [num_res_types, num_atom_types] mask array."""
# +1 to account for unknown (all 0s).
mask = np.zeros([restype_num + 1, atom_type_num], dtype=np.int32)
for restype, restype_letter in enumerate(restypes):
restype_name = restype_1to3[restype_letter]
atom_names = residue_atoms[restype_name]
for atom_name in atom_names:
atom_type = atom_order[atom_name]
mask[restype, atom_type] = 1
return mask
STANDARD_ATOM_MASK = _make_standard_atom_mask()
# A one hot representation for the first and second atoms defining the axis
# of rotation for each chi-angle in each residue.
def chi_angle_atom(atom_index: int) -> np.ndarray:
"""Define chi-angle rigid groups via one-hot representations."""
chi_angles_index = {}
one_hots = []
for k, v in chi_angles_atoms.items():
indices = [atom_types.index(s[atom_index]) for s in v]
indices.extend([-1] * (4 - len(indices)))
chi_angles_index[k] = indices
for r in restypes:
res3 = restype_1to3[r]
one_hot = np.eye(atom_type_num)[chi_angles_index[res3]]
one_hots.append(one_hot)
one_hots.append(np.zeros([4, atom_type_num])) # Add zeros for residue `X`.
one_hot = np.stack(one_hots, axis=0)
one_hot = np.transpose(one_hot, [0, 2, 1])
return one_hot
chi_atom_1_one_hot = chi_angle_atom(1)
chi_atom_2_one_hot = chi_angle_atom(2)
# An array like chi_angles_atoms but using indices rather than names.
chi_angles_atom_indices_list: List[List[List[str]]] = [chi_angles_atoms[restype_1to3[r]] for r in restypes]
chi_angles_atom_indices_ours: list = map_structure_with_atom_order(chi_angles_atom_indices_list)
chi_angles_atom_indices = np.array(
[chi_atoms + ([[0, 0, 0, 0]] * (4 - len(chi_atoms))) for chi_atoms in chi_angles_atom_indices_list]
)
# Mapping from (res_name, atom_name) pairs to the atom's chi group index
# and atom index within that group.
chi_groups_for_atom: Dict[Tuple[str, str], List[Tuple[int, int]]] = collections.defaultdict(list)
for res_name, chi_angle_atoms_for_res in chi_angles_atoms.items():
for chi_group_i, chi_group in enumerate(chi_angle_atoms_for_res):
for atom_i, atom in enumerate(chi_group):
chi_groups_for_atom[(res_name, atom)].append((chi_group_i, atom_i))
chi_groups_for_atom = dict(chi_groups_for_atom)
def _make_rigid_transformation_4x4(ex: np.ndarray, ey: np.ndarray, translation: np.ndarray) -> np.ndarray:
"""Create a rigid 4x4 transformation matrix from two axes and transl."""
# Normalize ex.
ex_normalized = ex / np.linalg.norm(ex)
# make ey perpendicular to ex
ey_normalized = ey - np.dot(ey, ex_normalized) * ex_normalized
ey_normalized /= np.linalg.norm(ey_normalized)
# compute ez as cross product
eznorm = np.cross(ex_normalized, ey_normalized)
m = np.stack([ex_normalized, ey_normalized, eznorm, translation]).transpose()
m = np.concatenate([m, [[0.0, 0.0, 0.0, 1.0]]], axis=0)
return m
# create an array with (restype, atomtype) --> rigid_group_idx
# and an array with (restype, atomtype, coord) for the atom positions
# and compute affine transformation matrices (4,4) from one rigid group to the
# previous group
restype_atom37_to_rigid_group = np.zeros([21, 37], dtype=int)
restype_atom37_mask = np.zeros([21, 37], dtype=np.float32)
restype_atom37_rigid_group_positions = np.zeros([21, 37, 3], dtype=np.float32)
restype_atom14_to_rigid_group = np.zeros([21, 14], dtype=int)
restype_atom14_mask = np.zeros([21, 14], dtype=np.float32)
restype_atom14_rigid_group_positions = np.zeros([21, 14, 3], dtype=np.float32)
restype_rigid_group_default_frame = np.zeros([21, 8, 4, 4], dtype=np.float32)
def _make_rigid_group_constants() -> None:
"""Fill the arrays above."""
for restype, restype_letter in enumerate(restypes):
resname = restype_1to3[restype_letter]
for atomname, group_idx, atom_position in rigid_group_atom_positions[resname]:
atomtype = atom_order[atomname]
restype_atom37_to_rigid_group[restype, atomtype] = group_idx
restype_atom37_mask[restype, atomtype] = 1
restype_atom37_rigid_group_positions[restype, atomtype, :] = atom_position
atom14idx = restype_name_to_atom14_names[resname].index(atomname)
restype_atom14_to_rigid_group[restype, atom14idx] = group_idx
restype_atom14_mask[restype, atom14idx] = 1
restype_atom14_rigid_group_positions[restype, atom14idx, :] = atom_position
for restype, restype_letter in enumerate(restypes):
resname = restype_1to3[restype_letter]
atom_positions: Dict[str, np.ndarray] = {
name: np.array(pos) for name, _, pos in rigid_group_atom_positions[resname]
}
# backbone to backbone is the identity transform
restype_rigid_group_default_frame[restype, 0, :, :] = np.eye(4)
# pre-omega-frame to backbone (currently dummy identity matrix)
restype_rigid_group_default_frame[restype, 1, :, :] = np.eye(4)
# phi-frame to backbone
mat = _make_rigid_transformation_4x4(
ex=atom_positions["N"] - atom_positions["CA"],
ey=np.array([1.0, 0.0, 0.0]),
translation=atom_positions["N"],
)
restype_rigid_group_default_frame[restype, 2, :, :] = mat
# psi-frame to backbone
mat = _make_rigid_transformation_4x4(
ex=atom_positions["C"] - atom_positions["CA"],
ey=atom_positions["CA"] - atom_positions["N"],
translation=atom_positions["C"],
)
restype_rigid_group_default_frame[restype, 3, :, :] = mat
# chi1-frame to backbone
if chi_angles_mask[restype][0]:
base_atom_names = chi_angles_atoms[resname][0]
base_atom_positions = [atom_positions[name] for name in base_atom_names]
mat = _make_rigid_transformation_4x4(
ex=base_atom_positions[2] - base_atom_positions[1],
ey=base_atom_positions[0] - base_atom_positions[1],
translation=base_atom_positions[2],
)
restype_rigid_group_default_frame[restype, 4, :, :] = mat
# chi2-frame to chi1-frame
# chi3-frame to chi2-frame
# chi4-frame to chi3-frame
# luckily all rotation axes for the next frame start at (0,0,0) of the
# previous frame
for chi_idx in range(1, 4):
if chi_angles_mask[restype][chi_idx]:
axis_end_atom_name = chi_angles_atoms[resname][chi_idx][2]
axis_end_atom_position = atom_positions[axis_end_atom_name]
mat = _make_rigid_transformation_4x4(
ex=axis_end_atom_position,
ey=np.array([-1.0, 0.0, 0.0]),
translation=axis_end_atom_position,
)
restype_rigid_group_default_frame[restype, 4 + chi_idx, :, :] = mat
_make_rigid_group_constants()
def make_atom14_dists_bounds(
overlap_tolerance: float = 1.5,
bond_length_tolerance_factor: int = 15,
) -> Dict[str, np.ndarray]:
"""compute upper and lower bounds for bonds to assess violations."""
restype_atom14_bond_lower_bound = np.zeros([21, 14, 14], np.float32)
restype_atom14_bond_upper_bound = np.zeros([21, 14, 14], np.float32)
restype_atom14_bond_stddev = np.zeros([21, 14, 14], np.float32)
residue_bonds, residue_virtual_bonds, _ = load_stereo_chemical_props()
for restype, restype_letter in enumerate(restypes):
resname = restype_1to3[restype_letter]
atom_list = restype_name_to_atom14_names[resname]
# create lower and upper bounds for clashes
for atom1_idx, atom1_name in enumerate(atom_list):
if not atom1_name:
continue
atom1_radius = van_der_waals_radius[atom1_name[0]]
for atom2_idx, atom2_name in enumerate(atom_list):
if (not atom2_name) or atom1_idx == atom2_idx:
continue
atom2_radius = van_der_waals_radius[atom2_name[0]]
lower = atom1_radius + atom2_radius - overlap_tolerance
upper = 1e10
restype_atom14_bond_lower_bound[restype, atom1_idx, atom2_idx] = lower
restype_atom14_bond_lower_bound[restype, atom2_idx, atom1_idx] = lower
restype_atom14_bond_upper_bound[restype, atom1_idx, atom2_idx] = upper
restype_atom14_bond_upper_bound[restype, atom2_idx, atom1_idx] = upper
# overwrite lower and upper bounds for bonds and angles
for b in residue_bonds[resname] + residue_virtual_bonds[resname]:
atom1_idx = atom_list.index(b.atom1_name)
atom2_idx = atom_list.index(b.atom2_name)
lower = b.length - bond_length_tolerance_factor * b.stddev
upper = b.length + bond_length_tolerance_factor * b.stddev
restype_atom14_bond_lower_bound[restype, atom1_idx, atom2_idx] = lower
restype_atom14_bond_lower_bound[restype, atom2_idx, atom1_idx] = lower
restype_atom14_bond_upper_bound[restype, atom1_idx, atom2_idx] = upper
restype_atom14_bond_upper_bound[restype, atom2_idx, atom1_idx] = upper
restype_atom14_bond_stddev[restype, atom1_idx, atom2_idx] = b.stddev
restype_atom14_bond_stddev[restype, atom2_idx, atom1_idx] = b.stddev
return {
"lower_bound": restype_atom14_bond_lower_bound, # shape (21,14,14)
"upper_bound": restype_atom14_bond_upper_bound, # shape (21,14,14)
"stddev": restype_atom14_bond_stddev, # shape (21,14,14)
}
restype_atom14_ambiguous_atoms = np.zeros((21, 14), dtype=np.float32)
restype_atom14_ambiguous_atoms_swap_idx: np.ndarray = np.tile(np.arange(14, dtype=int), (21, 1))
def _make_atom14_ambiguity_feats() -> None:
for res, pairs in residue_atom_renaming_swaps.items():
res_idx = restype_order[restype_3to1[res]]
for atom1, atom2 in pairs.items():
atom1_idx = restype_name_to_atom14_names[res].index(atom1)
atom2_idx = restype_name_to_atom14_names[res].index(atom2)
restype_atom14_ambiguous_atoms[res_idx, atom1_idx] = 1
restype_atom14_ambiguous_atoms[res_idx, atom2_idx] = 1
restype_atom14_ambiguous_atoms_swap_idx[res_idx, atom1_idx] = atom2_idx
restype_atom14_ambiguous_atoms_swap_idx[res_idx, atom2_idx] = atom1_idx
_make_atom14_ambiguity_feats()
def aatype_to_str_sequence(aatype: Sequence[int]) -> str:
return "".join([restypes_with_x[aatype[i]] for i in range(len(aatype))])
|
transformers/src/transformers/models/esm/openfold_utils/residue_constants.py/0
|
{
"file_path": "transformers/src/transformers/models/esm/openfold_utils/residue_constants.py",
"repo_id": "transformers",
"token_count": 19142
}
| 355
|
# coding=utf-8
# Copyright 2023 The Espnet authors, IMS Toucan authors, and the HuggingFace Inc. team. 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.
"""PyTorch FastSpeech2Conformer model."""
import math
from dataclasses import dataclass
from typing import Optional, Tuple, Union
import torch
from torch import nn
from ...modeling_outputs import BaseModelOutput
from ...modeling_utils import PreTrainedModel
from ...utils import ModelOutput, add_start_docstrings, logging, replace_return_docstrings
from .configuration_fastspeech2_conformer import (
FastSpeech2ConformerConfig,
FastSpeech2ConformerHifiGanConfig,
FastSpeech2ConformerWithHifiGanConfig,
)
logger = logging.get_logger(__name__)
@dataclass
class FastSpeech2ConformerModelOutput(ModelOutput):
"""
Output type of [`FastSpeech2ConformerModel`].
Args:
loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
Spectrogram generation loss.
spectrogram (`torch.FloatTensor` of shape `(batch_size, sequence_length, num_bins)`):
The predicted spectrogram.
encoder_last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
Sequence of hidden-states at the output of the last layer of the encoder of the model.
encoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the encoder at the output of each layer plus the initial embedding outputs.
encoder_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights of the encoder, after the attention softmax, used to compute the weighted average in the
self-attention heads.
decoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the decoder at the output of each layer plus the initial embedding outputs.
decoder_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights of the decoder, after the attention softmax, used to compute the weighted average in the
self-attention heads.
duration_outputs (`torch.LongTensor` of shape `(batch_size, max_text_length + 1)`, *optional*):
Outputs of the duration predictor.
pitch_outputs (`torch.FloatTensor` of shape `(batch_size, max_text_length + 1, 1)`, *optional*):
Outputs of the pitch predictor.
energy_outputs (`torch.FloatTensor` of shape `(batch_size, max_text_length + 1, 1)`, *optional*):
Outputs of the energy predictor.
"""
loss: Optional[torch.FloatTensor] = None
spectrogram: torch.FloatTensor = None
encoder_last_hidden_state: Optional[torch.FloatTensor] = None
encoder_hidden_states: Optional[Tuple[torch.FloatTensor]] = None
encoder_attentions: Optional[Tuple[torch.FloatTensor]] = None
decoder_hidden_states: Optional[Tuple[torch.FloatTensor]] = None
decoder_attentions: Optional[Tuple[torch.FloatTensor]] = None
duration_outputs: torch.LongTensor = None
pitch_outputs: torch.FloatTensor = None
energy_outputs: torch.FloatTensor = None
@dataclass
class FastSpeech2ConformerWithHifiGanOutput(FastSpeech2ConformerModelOutput):
"""
Output type of [`FastSpeech2ConformerWithHifiGan`].
Args:
waveform (`torch.FloatTensor` of shape `(batch_size, audio_length)`):
Speech output as a result of passing the predicted mel spectrogram through the vocoder.
loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
Spectrogram generation loss.
spectrogram (`torch.FloatTensor` of shape `(batch_size, sequence_length, num_bins)`):
The predicted spectrogram.
encoder_last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
Sequence of hidden-states at the output of the last layer of the encoder of the model.
encoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the encoder at the output of each layer plus the initial embedding outputs.
encoder_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights of the encoder, after the attention softmax, used to compute the weighted average in the
self-attention heads.
decoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the decoder at the output of each layer plus the initial embedding outputs.
decoder_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights of the decoder, after the attention softmax, used to compute the weighted average in the
self-attention heads.
duration_outputs (`torch.LongTensor` of shape `(batch_size, max_text_length + 1)`, *optional*):
Outputs of the duration predictor.
pitch_outputs (`torch.FloatTensor` of shape `(batch_size, max_text_length + 1, 1)`, *optional*):
Outputs of the pitch predictor.
energy_outputs (`torch.FloatTensor` of shape `(batch_size, max_text_length + 1, 1)`, *optional*):
Outputs of the energy predictor.
"""
waveform: torch.FloatTensor = None
_CONFIG_FOR_DOC = "FastSpeech2ConformerConfig"
FASTSPEECH2_CONFORMER_START_DOCSTRING = r"""
This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
etc.)
This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
and behavior.
Parameters:
config ([`FastSpeech2ConformerConfig`]):
Model configuration class with all the parameters of the model. Initializing with a config file does not
load the weights associated with the model, only the configuration. Check out the
[`~PreTrainedModel.from_pretrained`] method to load the model weights.
"""
HIFIGAN_START_DOCSTRING = r"""
This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
etc.)
This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
and behavior.
Parameters:
config ([`FastSpeech2ConformerConfig`]):
Model configuration class with all the parameters of the model. Initializing with a config file does not
load the weights associated with the model, only the configuration. Check out the
[`~PreTrainedModel.from_pretrained`] method to load the model weights.
"""
FASTSPEECH2_CONFORMER_WITH_HIFIGAN_START_DOCSTRING = r"""
This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
etc.)
This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
and behavior.
Parameters:
config ([`FastSpeech2ConformerWithHifiGanConfig`]):
Model configuration class with all the parameters of the model. Initializing with a config file does not
load the weights associated with the model, only the configuration. Check out the
[`~PreTrainedModel.from_pretrained`] method to load the model weights.
"""
def length_regulator(encoded_embeddings, duration_labels, speaking_speed=1.0):
"""
Length regulator for feed-forward Transformer.
This is the length regulator module described in `FastSpeech: Fast, Robust and Controllable Text to Speech`
https://arxiv.org/pdf/1905.09263.pdf. The length regulator expands char or phoneme-level embedding features to
frame-level by repeating each feature based on the corresponding predicted durations.
Args:
encoded_embeddings (`torch.Tensor` of shape `(batch_size, max_text_length, embedding_dim)`):
Batch of sequences of char or phoneme embeddings.
duration_labels (`torch.LongTensor` of shape `(batch_size, time)`):
Batch of durations of each frame.
speaking_speed (`float`, *optional*, defaults to 1.0):
Value to control speed of speech.
Returns:
`torch.Tensor`:
Replicated input tensor based on durations (batch_size, time*, embedding_dim).
"""
if speaking_speed <= 0:
raise ValueError("`speaking_speed` must be greater than 0.")
elif speaking_speed != 1.0:
duration_labels = torch.round(duration_labels.float() * speaking_speed).long()
if duration_labels.sum() == 0:
duration_labels[duration_labels.sum(dim=1).eq(0)] = 1
# Calculate the maximum length needed
max_len = torch.sum(duration_labels, dim=1).max()
# Create a padded tensor to hold the results
hidden_states = torch.zeros(
(encoded_embeddings.size(0), max_len, encoded_embeddings.size(2)),
dtype=torch.float,
device=encoded_embeddings.device,
)
# Loop through the batch and fill in the data
for i, (encoded_embedding, target_duration) in enumerate(zip(encoded_embeddings, duration_labels)):
repeated = torch.repeat_interleave(encoded_embedding, target_duration, dim=0)
hidden_states[i, : repeated.size(0)] = repeated
return hidden_states
class FastSpeech2ConformerDurationPredictor(nn.Module):
"""
Duration predictor module.
This is a module of duration predictor described in the paper 'FastSpeech: Fast, Robust and Controllable Text to
Speech' https://arxiv.org/pdf/1905.09263.pdf The duration predictor predicts a duration of each frame in log domain
from the hidden embeddings of encoder.
Note:
The calculation domain of outputs is different between in `forward` and in `inference`. In `forward`, the
outputs are calculated in log domain but in `inference`, those are calculated in linear domain.
"""
def __init__(self, config: FastSpeech2ConformerConfig):
super().__init__()
self.conv_layers = nn.ModuleList()
self.log_domain_offset = 1.0
for layer_idx in range(config.duration_predictor_layers):
num_chans = config.duration_predictor_channels
input_channels = config.hidden_size if layer_idx == 0 else num_chans
layer = FastSpeech2ConformerPredictorLayer(
input_channels,
num_chans,
config.duration_predictor_kernel_size,
config.duration_predictor_dropout_rate,
)
self.conv_layers.append(layer)
self.linear = nn.Linear(config.duration_predictor_channels, 1)
def forward(self, encoder_hidden_states):
"""
Args:
hidden_states (`torch.Tensor` of shape `(batch_size, max_text_length, input_dim)`):
Batch of input sequences.
padding_masks (`torch.ByteTensor` of shape `(batch_size, max_text_length)`, *optional*):
Batch of masks indicating padded part.
Returns:
`torch.Tensor`: Batch of predicted durations in log domain `(batch_size, max_text_length)`.
"""
# (batch_size, input_dim, max_text_length)
hidden_states = encoder_hidden_states.transpose(1, -1)
for layer in self.conv_layers:
hidden_states = layer(hidden_states)
# NOTE: calculate in log domain, (batch_size, max_text_length)
hidden_states = self.linear(hidden_states.transpose(1, -1)).squeeze(-1)
if not self.training:
# NOTE: calculate in linear domain
hidden_states = torch.clamp(torch.round(hidden_states.exp() - self.log_domain_offset), min=0).long()
return hidden_states
# Copied from transformers.models.speecht5.modeling_speecht5.SpeechT5BatchNormConvLayer
class FastSpeech2ConformerBatchNormConvLayer(nn.Module):
def __init__(self, config, layer_id=0):
super().__init__()
if layer_id == 0:
in_conv_dim = config.num_mel_bins
else:
in_conv_dim = config.speech_decoder_postnet_units
if layer_id == config.speech_decoder_postnet_layers - 1:
out_conv_dim = config.num_mel_bins
else:
out_conv_dim = config.speech_decoder_postnet_units
self.conv = nn.Conv1d(
in_conv_dim,
out_conv_dim,
kernel_size=config.speech_decoder_postnet_kernel,
stride=1,
padding=(config.speech_decoder_postnet_kernel - 1) // 2,
bias=False,
)
self.batch_norm = nn.BatchNorm1d(out_conv_dim)
if layer_id < config.speech_decoder_postnet_layers - 1:
self.activation = nn.Tanh()
else:
self.activation = None
self.dropout = nn.Dropout(config.speech_decoder_postnet_dropout)
def forward(self, hidden_states):
hidden_states = self.conv(hidden_states)
hidden_states = self.batch_norm(hidden_states)
if self.activation is not None:
hidden_states = self.activation(hidden_states)
hidden_states = self.dropout(hidden_states)
return hidden_states
class FastSpeech2ConformerSpeechDecoderPostnet(nn.Module):
def __init__(self, config):
super().__init__()
self.config = config
self.feat_out = nn.Linear(config.hidden_size, config.num_mel_bins * config.reduction_factor)
self.layers = nn.ModuleList(
[FastSpeech2ConformerBatchNormConvLayer(config, i) for i in range(config.speech_decoder_postnet_layers)]
)
def forward(self, hidden_states: torch.Tensor):
outputs_before_postnet = self.feat_out(hidden_states).view(hidden_states.size(0), -1, self.config.num_mel_bins)
layer_output = outputs_before_postnet.transpose(1, 2)
for layer in self.layers:
layer_output = layer(layer_output)
outputs_after_postnet = outputs_before_postnet + layer_output.transpose(1, 2)
return outputs_before_postnet, outputs_after_postnet
class FastSpeech2ConformerPredictorLayer(nn.Module):
def __init__(self, input_channels, num_chans, kernel_size, dropout_rate):
super().__init__()
self.conv = nn.Conv1d(
input_channels,
num_chans,
kernel_size,
stride=1,
padding=(kernel_size - 1) // 2,
)
self.activation = nn.ReLU()
self.layer_norm = nn.LayerNorm(num_chans)
self.dropout = nn.Dropout(dropout_rate)
def forward(self, hidden_states):
hidden_states = self.conv(hidden_states)
hidden_states = self.activation(hidden_states)
# Perform layer norm on dimension 1
hidden_states = hidden_states.transpose(1, -1)
hidden_states = self.layer_norm(hidden_states)
hidden_states = hidden_states.transpose(1, -1)
hidden_states = self.dropout(hidden_states)
return hidden_states
class FastSpeech2ConformerVariancePredictor(nn.Module):
def __init__(
self,
config: FastSpeech2ConformerConfig,
num_layers=2,
num_chans=384,
kernel_size=3,
dropout_rate=0.5,
):
"""
Initilize variance predictor module.
Args:
input_dim (`int`): Input dimension.
num_layers (`int`, *optional*, defaults to 2): Number of convolutional layers.
num_chans (`int`, *optional*, defaults to 384): Number of channels of convolutional layers.
kernel_size (`int`, *optional*, defaults to 3): Kernel size of convolutional layers.
dropout_rate (`float`, *optional*, defaults to 0.5): Dropout rate.
"""
super().__init__()
self.conv_layers = nn.ModuleList()
for idx in range(num_layers):
input_channels = config.hidden_size if idx == 0 else num_chans
layer = FastSpeech2ConformerPredictorLayer(input_channels, num_chans, kernel_size, dropout_rate)
self.conv_layers.append(layer)
self.linear = nn.Linear(num_chans, 1)
def forward(self, encoder_hidden_states, padding_masks=None):
"""
Calculate forward propagation.
Args:
encoder_hidden_states (`torch.Tensor` of shape `(batch_size, max_text_length, input_dim)`):
Batch of input sequences.
padding_masks (`torch.ByteTensor` of shape `(batch_size, max_text_length)`, *optional*):
Batch of masks indicating padded part.
Returns:
Tensor: Batch of predicted sequences `(batch_size, max_text_length, 1)`.
"""
# (batch_size, input_dim, max_text_length)
hidden_states = encoder_hidden_states.transpose(1, -1)
for layer in self.conv_layers:
hidden_states = layer(hidden_states)
hidden_states = self.linear(hidden_states.transpose(1, 2))
if padding_masks is not None:
hidden_states = hidden_states.masked_fill(padding_masks, 0.0)
return hidden_states
class FastSpeech2ConformerVarianceEmbedding(nn.Module):
def __init__(
self,
in_channels=1,
out_channels=384,
kernel_size=1,
padding=0,
dropout_rate=0.0,
):
super().__init__()
self.conv = nn.Conv1d(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=kernel_size,
padding=padding,
)
self.dropout = nn.Dropout(dropout_rate)
def forward(self, hidden_states):
hidden_states = hidden_states.transpose(1, 2)
hidden_states = self.conv(hidden_states)
hidden_states = self.dropout(hidden_states)
hidden_states = hidden_states.transpose(1, 2)
return hidden_states
class FastSpeech2ConformerAttention(nn.Module):
"""
Multi-Head attention layer with relative position encoding. Details can be found in
https://github.com/espnet/espnet/pull/2816. Paper: https://arxiv.org/abs/1901.02860.
"""
def __init__(self, config: FastSpeech2ConformerConfig, module_config):
"""Construct an FastSpeech2ConformerAttention object."""
super().__init__()
# We assume d_v always equals dim_key
self.num_heads = module_config["num_attention_heads"]
self.hidden_size = config.hidden_size
self.dim_key = self.hidden_size // self.num_heads
self.head_dim = self.hidden_size // self.num_heads
self.linear_q = nn.Linear(self.hidden_size, self.hidden_size)
self.linear_k = nn.Linear(self.hidden_size, self.hidden_size)
self.linear_v = nn.Linear(self.hidden_size, self.hidden_size)
self.linear_out = nn.Linear(self.hidden_size, self.hidden_size)
self.dropout = nn.Dropout(p=module_config["attention_dropout_rate"])
# linear transformation for positional encoding
self.linear_pos = nn.Linear(self.hidden_size, self.hidden_size, bias=False)
# these two learnable bias are used in matrix c and matrix d
# as described in https://arxiv.org/abs/1901.02860 Section 3.3
self.pos_bias_u = nn.Parameter(torch.Tensor(self.num_heads, self.head_dim))
self.pos_bias_v = nn.Parameter(torch.Tensor(self.num_heads, self.head_dim))
def shift_relative_position_tensor(self, pos_tensor):
"""
Args:
pos_tensor (torch.Tensor of shape (batch_size, head, time1, 2*time1-1)): Input tensor.
"""
zero_pad = torch.zeros((*pos_tensor.size()[:3], 1), device=pos_tensor.device, dtype=pos_tensor.dtype)
pos_tensor_padded = torch.cat([zero_pad, pos_tensor], dim=-1)
pos_tensor_padded = pos_tensor_padded.view(*pos_tensor.size()[:2], pos_tensor.size(3) + 1, pos_tensor.size(2))
# only keep the positions from 0 to time2
pos_tensor = pos_tensor_padded[:, :, 1:].view_as(pos_tensor)[:, :, :, : pos_tensor.size(-1) // 2 + 1]
return pos_tensor
def forward(
self,
hidden_states: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
pos_emb: Optional[torch.Tensor] = None,
output_attentions: Optional[torch.Tensor] = False,
) -> Tuple[torch.Tensor, torch.Tensor]:
"""
Compute 'Scaled Dot Product Attention' with rel. positional encoding.
Args:
hidden_states (`torch.Tensor` of shape `(batch, time2, size)`): Values of the hidden states
attention_mask (`torch.Tensor` of shape `(batch, time1, time2)`): Mask tensor.
pos_emb (`torch.Tensor` of shape `(batch, 2*time1-1, size)`): Positional embedding tensor.
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
returned tensors for more detail.
Returns:
`torch.Tensor`: Output tensor of shape `(batch, time1, d_model)`.
"""
bsz, q_len, _ = hidden_states.size()
query_states = self.linear_q(hidden_states).view(bsz, -1, self.num_heads, self.head_dim)
key_states = self.linear_k(hidden_states).view(bsz, -1, self.num_heads, self.head_dim)
value_states = self.linear_v(hidden_states).view(bsz, -1, self.num_heads, self.head_dim)
bsz_pos = pos_emb.size(0)
pos_encoding = self.linear_pos(pos_emb).view(bsz_pos, -1, self.num_heads, self.head_dim)
# (batch_size, head, time1, dim_key)
query_with_bias_u = (query_states + self.pos_bias_u).transpose(1, 2)
# (batch_size, head, time1, dim_key)
query_with_bias_v = (query_states + self.pos_bias_v).transpose(1, 2)
# compute attention score
# first compute matrix a and matrix c
# as described in https://arxiv.org/abs/1901.02860 Section 3.3
# (batch_size, head, time1, time2)
matrix_ac = torch.matmul(query_with_bias_u, key_states.permute(0, 2, 3, 1))
# compute matrix b and matrix d
# (batch_size, head, time1, 2*time1-1)
matrix_bd = torch.matmul(query_with_bias_v, pos_encoding.permute(0, 2, 3, 1))
matrix_bd = self.shift_relative_position_tensor(matrix_bd)
# (batch_size, head, time1, time2)
scores = (matrix_ac + matrix_bd) / math.sqrt(self.dim_key)
# Forward attention
if attention_mask is not None:
expected_size = (bsz, 1, q_len)
if attention_mask.size() != expected_size:
raise ValueError(f"Attention mask should be of size {expected_size}, but is {attention_mask.size()}")
attention_mask = attention_mask.unsqueeze(1).eq(0)
min_value = float(torch.finfo(scores.dtype).min)
scores = scores.masked_fill(attention_mask, min_value)
attn_weights = torch.softmax(scores, dim=-1).masked_fill(attention_mask, 0.0)
else:
attn_weights = torch.softmax(scores, dim=-1)
attn_weights = self.dropout(attn_weights)
attn_output = torch.matmul(attn_weights, value_states.transpose(1, 2))
attn_output = attn_output.transpose(1, 2).contiguous().view(bsz, q_len, -1)
attn_output = self.linear_out(attn_output)
if not output_attentions:
attn_weights = None
return attn_output, attn_weights
class FastSpeech2ConformerConvolutionModule(nn.Module):
def __init__(self, config: FastSpeech2ConformerConfig, module_config):
super().__init__()
# kernel_size should be an odd number for 'SAME' padding
channels = config.hidden_size
kernel_size = module_config["kernel_size"]
self.pointwise_conv1 = nn.Conv1d(channels, 2 * channels, kernel_size=1, stride=1, padding=0, bias=True)
self.depthwise_conv = nn.Conv1d(
channels, channels, kernel_size, stride=1, padding=(kernel_size - 1) // 2, groups=channels, bias=True
)
self.norm = nn.BatchNorm1d(channels)
self.pointwise_conv2 = nn.Conv1d(channels, channels, kernel_size=1, stride=1, padding=0, bias=True)
def forward(self, hidden_states):
"""
Compute convolution module.
Args:
hidden_states (`torch.Tensor` of shape `(batch, time, channels)`): Input tensor.
Returns:
`torch.Tensor`: Output tensor of shape `(batch, time, channels)`.
"""
# exchange the temporal dimension and the feature dimension
hidden_states = hidden_states.transpose(1, 2)
# GLU mechanism, (batch_size, 2*channel, dim)
hidden_states = self.pointwise_conv1(hidden_states)
# (batch_size, channel, dim)
hidden_states = nn.functional.glu(hidden_states, dim=1)
# 1D Depthwise Conv
hidden_states = self.depthwise_conv(hidden_states)
hidden_states = self.norm(hidden_states)
hidden_states = hidden_states * torch.sigmoid(hidden_states)
hidden_states = self.pointwise_conv2(hidden_states)
return hidden_states.transpose(1, 2)
class FastSpeech2ConformerEncoderLayer(nn.Module):
def __init__(self, config: FastSpeech2ConformerConfig, module_config):
super().__init__()
# self-attention module definition
self.self_attn = FastSpeech2ConformerAttention(config, module_config)
# feed-forward module definition
self.feed_forward = FastSpeech2ConformerMultiLayeredConv1d(config, module_config)
self.macaron_style = config.use_macaron_style_in_conformer
if self.macaron_style:
self.feed_forward_macaron = FastSpeech2ConformerMultiLayeredConv1d(config, module_config)
self.ff_macaron_layer_norm = nn.LayerNorm(config.hidden_size)
self.ff_scale = 0.5
else:
self.ff_scale = 1.0
# convolution module definition
self.use_cnn_module = config.use_cnn_in_conformer
if self.use_cnn_module:
self.conv_module = FastSpeech2ConformerConvolutionModule(config, module_config)
self.conv_layer_norm = nn.LayerNorm(config.hidden_size)
self.final_layer_norm = nn.LayerNorm(config.hidden_size)
self.ff_layer_norm = nn.LayerNorm(config.hidden_size)
self.self_attn_layer_norm = nn.LayerNorm(config.hidden_size)
self.dropout = nn.Dropout(module_config["dropout_rate"])
self.size = config.hidden_size
self.normalize_before = module_config["normalize_before"]
self.concat_after = module_config["concat_after"]
if self.concat_after:
self.concat_linear = nn.Linear(config.hidden_size + config.hidden_size, config.hidden_size)
def forward(
self,
hidden_states: torch.Tensor,
pos_emb: Optional[torch.Tensor] = None,
attention_mask: Optional[torch.Tensor] = None,
output_attentions: Optional[torch.Tensor] = False,
):
"""
Compute encoded features.
Args:
hidden_states (`torch.Tensor` of shape `(batch, time, size)`): Input tensor.
pos_emb (`torch.Tensor` of shape `(1, time, size)`): Positional embeddings tensor.
attention_mask (`torch.Tensor` of shape `(batch, time)`): Attention mask tensor for the input.
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
returned tensors for more detail.
Returns:
`torch.Tensor`: Output tensor of shape `(batch, time, size)`.
"""
# whether to use macaron style
if self.macaron_style:
residual = hidden_states
if self.normalize_before:
hidden_states = self.ff_macaron_layer_norm(hidden_states)
hidden_states = residual + self.ff_scale * self.dropout(self.feed_forward_macaron(hidden_states))
if not self.normalize_before:
hidden_states = self.ff_macaron_layer_norm(hidden_states)
# multi-headed self-attention module
residual = hidden_states
if self.normalize_before:
hidden_states = self.self_attn_layer_norm(hidden_states)
attention_output, attention_scores = self.self_attn(
hidden_states, attention_mask=attention_mask, pos_emb=pos_emb, output_attentions=output_attentions
)
if self.concat_after:
x_concat = torch.cat((hidden_states, attention_output), dim=-1)
hidden_states = self.concat_linear(x_concat)
hidden_states = residual + hidden_states
else:
hidden_states = self.dropout(attention_output)
hidden_states = residual + hidden_states
if not self.normalize_before:
hidden_states = self.self_attn_layer_norm(hidden_states)
# convolution module
if self.use_cnn_module:
residual = hidden_states
if self.normalize_before:
hidden_states = self.conv_layer_norm(hidden_states)
hidden_states = self.conv_module(hidden_states)
hidden_states = self.dropout(hidden_states)
hidden_states = residual + hidden_states
if not self.normalize_before:
hidden_states = self.conv_layer_norm(hidden_states)
# feed forward module
residual = hidden_states
if self.normalize_before:
hidden_states = self.ff_layer_norm(hidden_states)
hidden_states = self.feed_forward(hidden_states)
hidden_states = self.dropout(hidden_states)
hidden_states = residual + self.ff_scale * hidden_states
if not self.normalize_before:
hidden_states = self.ff_layer_norm(hidden_states)
if self.conv_module is not None:
hidden_states = self.final_layer_norm(hidden_states)
outputs = (hidden_states,)
if output_attentions:
outputs += (attention_scores,)
return outputs
class FastSpeech2ConformerMultiLayeredConv1d(nn.Module):
"""
Multi-layered conv1d for Transformer block.
This is a module of multi-layered conv1d designed to replace positionwise feed-forward network in Transformer
block, which is introduced in 'FastSpeech: Fast, Robust and Controllable Text to Speech'
https://arxiv.org/pdf/1905.09263.pdf
"""
def __init__(self, config: FastSpeech2ConformerConfig, module_config):
"""
Initialize FastSpeech2ConformerMultiLayeredConv1d module.
Args:
input_channels (`int`): Number of input channels.
hidden_channels (`int`): Number of hidden channels.
kernel_size (`int`): Kernel size of conv1d.
dropout_rate (`float`): Dropout rate.
"""
super().__init__()
input_channels = config.hidden_size
hidden_channels = module_config["linear_units"]
kernel_size = config.positionwise_conv_kernel_size
self.conv1 = nn.Conv1d(input_channels, hidden_channels, kernel_size, stride=1, padding=(kernel_size - 1) // 2)
self.conv2 = nn.Conv1d(hidden_channels, input_channels, kernel_size, stride=1, padding=(kernel_size - 1) // 2)
self.dropout = nn.Dropout(module_config["dropout_rate"])
def forward(self, hidden_states):
"""
Calculate forward propagation.
Args:
hidden_states (torch.Tensor): Batch of input tensors (batch_size, time, input_channels).
Returns:
torch.Tensor: Batch of output tensors (batch_size, time, hidden_channels).
"""
hidden_states = hidden_states.transpose(-1, 1)
hidden_states = self.conv1(hidden_states)
hidden_states = torch.relu(hidden_states)
hidden_states = self.dropout(hidden_states)
hidden_states = self.conv2(hidden_states)
hidden_states = hidden_states.transpose(-1, 1)
return hidden_states
class FastSpeech2ConformerRelPositionalEncoding(nn.Module):
"""
Args:
Relative positional encoding module (new implementation). Details can be found in
https://github.com/espnet/espnet/pull/2816. See : Appendix Batch in https://arxiv.org/abs/1901.02860
config (`FastSpeech2ConformerConfig`):
FastSpeech2ConformerConfig instance.
module_config (`dict`):
Dictionary containing the encoder or decoder module configuration from the `FastSpeech2ConformerConfig`.
"""
def __init__(self, config: FastSpeech2ConformerConfig, module_config):
"""
Construct an PositionalEncoding object.
"""
super().__init__()
self.embed_dim = config.hidden_size
self.input_scale = math.sqrt(self.embed_dim)
self.dropout = nn.Dropout(p=module_config["positional_dropout_rate"])
self.pos_enc = None
self.max_len = 5000
self.extend_pos_enc(torch.tensor(0.0).expand(1, self.max_len))
def extend_pos_enc(self, x):
"""Reset the positional encodings."""
if self.pos_enc is not None:
# self.pos_enc contains both positive and negative parts
# the length of self.pos_enc is 2 * input_len - 1
if self.pos_enc.size(1) >= x.size(1) * 2 - 1:
if self.pos_enc.dtype != x.dtype or self.pos_enc.device != x.device:
self.pos_enc = self.pos_enc.to(dtype=x.dtype, device=x.device)
return
# Suppose `i` means to the position of query vector and `j` means the
# position of key vector. We use position relative positions when keys
# are to the left (i>j) and negative relative positions otherwise (i<j).
pos_enc_positive = torch.zeros(x.size(1), self.embed_dim)
pos_enc_negative = torch.zeros(x.size(1), self.embed_dim)
position = torch.arange(0, x.size(1), dtype=torch.int64).float().unsqueeze(1)
div_term = torch.exp(
torch.arange(0, self.embed_dim, 2, dtype=torch.int64).float() * -(math.log(10000.0) / self.embed_dim)
)
pos_enc_positive[:, 0::2] = torch.sin(position * div_term)
pos_enc_positive[:, 1::2] = torch.cos(position * div_term)
pos_enc_negative[:, 0::2] = torch.sin(-1 * position * div_term)
pos_enc_negative[:, 1::2] = torch.cos(-1 * position * div_term)
# Reserve the order of positive indices and concat both positive and
# negative indices. This is used to support the shifting trick
# as in https://arxiv.org/abs/1901.02860
pos_enc_positive = torch.flip(pos_enc_positive, [0]).unsqueeze(0)
pos_enc_negative = pos_enc_negative[1:].unsqueeze(0)
pos_enc = torch.cat([pos_enc_positive, pos_enc_negative], dim=1)
self.pos_enc = pos_enc.to(device=x.device, dtype=x.dtype)
def forward(self, feature_representation):
"""
Args:
feature_representation (`torch.Tensor` of shape (batch_size, time, `*`)):
Input tensor.
Returns:
`torch.Tensor`: Encoded tensor (batch_size, time, `*`).
"""
self.extend_pos_enc(feature_representation)
hidden_states = feature_representation * self.input_scale
center_idx = self.pos_enc.size(1) // 2
pos_emb = self.pos_enc[:, center_idx - hidden_states.size(1) + 1 : center_idx + hidden_states.size(1)]
return self.dropout(hidden_states), self.dropout(pos_emb)
class FastSpeech2ConformerEncoder(nn.Module):
"""
FastSpeech2ConformerEncoder encoder module.
Args:
config (`FastSpeech2ConformerConfig`):
FastSpeech2ConformerConfig instance.
module_config (`dict`):
Dictionary containing the encoder or decoder module configuration from the `FastSpeech2ConformerConfig`.
use_encoder_input_layer (`bool`, *optional*, defaults to `False`):
Input layer type.
"""
def __init__(
self,
config: FastSpeech2ConformerConfig,
module_config,
use_encoder_input_layer=False,
):
super().__init__()
self.embed = None
if use_encoder_input_layer:
self.embed = nn.Embedding(
num_embeddings=config.vocab_size, embedding_dim=config.hidden_size, padding_idx=0
)
self.pos_enc = FastSpeech2ConformerRelPositionalEncoding(config, module_config)
self.conformer_layers = nn.ModuleList(
[FastSpeech2ConformerEncoderLayer(config, module_config) for _ in range(module_config["layers"])]
)
def forward(
self,
input_tensor: torch.LongTensor,
attention_mask: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
output_attentions: Optional[bool] = False,
return_dict: Optional[bool] = None,
):
"""
Args:
input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you
provide it.
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
[`PreTrainedTokenizer.__call__`] for details.
[What are input IDs?](../glossary#input-ids)
attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
- 1 for tokens that are **not masked**,
- 0 for tokens that are **masked**.
[What are attention masks?](../glossary#attention-mask)
output_hidden_states (`bool`, *optional*):
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
for more detail.
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
returned tensors for more detail.
return_dict (`bool`, *optional*):
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
Returns:
`torch.Tensor`:
Output tensor of shape `(batch, time, attention_dim)`.
"""
feature_representation = input_tensor
if self.embed is not None:
feature_representation = self.embed(feature_representation)
hidden_states, pos_emb = self.pos_enc(feature_representation)
all_hidden_states = () if output_hidden_states else None
all_self_attentions = () if output_attentions else None
for conformer_layer in self.conformer_layers:
if output_hidden_states:
all_hidden_states = all_hidden_states + (hidden_states,)
layer_outputs = conformer_layer(hidden_states, pos_emb, attention_mask, output_attentions)
hidden_states = layer_outputs[0]
if output_attentions:
all_self_attentions = all_self_attentions + (layer_outputs[1],)
# Add last layer
if output_hidden_states:
all_hidden_states = all_hidden_states + (hidden_states,)
if not return_dict:
return tuple(v for v in [hidden_states, all_hidden_states, all_self_attentions] if v is not None)
return BaseModelOutput(
last_hidden_state=hidden_states, hidden_states=all_hidden_states, attentions=all_self_attentions
)
class FastSpeech2ConformerLoss(nn.Module):
def __init__(self, config: FastSpeech2ConformerConfig):
super().__init__()
use_masking = config.use_masking
use_weighted_masking = config.use_weighted_masking
if use_masking and use_weighted_masking:
raise ValueError("Either use_masking or use_weighted_masking can be True, but not both.")
self.use_masking = use_masking
self.use_weighted_masking = use_weighted_masking
# define criterions
reduction = "none" if self.use_weighted_masking else "mean"
self.l1_criterion = nn.L1Loss(reduction=reduction)
self.mse_criterion = nn.MSELoss(reduction=reduction)
self.duration_criterion = nn.MSELoss(reduction=reduction)
self.log_domain_offset = 1.0
def forward(
self,
outputs_after_postnet,
outputs_before_postnet,
duration_outputs,
pitch_outputs,
energy_outputs,
spectrogram_labels,
duration_labels,
pitch_labels,
energy_labels,
duration_mask,
spectrogram_mask,
):
"""
Args:
outputs_after_postnet (`torch.Tensor` of shape `(batch_size, max_spectrogram_length, num_mel_bins)`):
Batch of outputs after postnet.
outputs_before_postnet (`torch.Tensor` of shape `(batch_size, max_spectrogram_length, num_mel_bins)`):
Batch of outputs before postnet.
duration_outputs (`torch.LongTensor` of shape `(batch_size, max_text_length)`):
Batch of outputs of duration predictor.
pitch_outputs (`torch.Tensor` of shape `(batch_size, max_text_length, 1)`):
Batch of outputs of pitch predictor.
energy_outputs (`torch.Tensor` of shape `(batch_size, max_text_length, 1)`):
Batch of outputs of energy predictor.
spectrogram_labels (`torch.Tensor` of shape `(batch_size, max_spectrogram_length, num_mel_bins)`):
Batch of target features.
duration_labels (`torch.LongTensor` of shape `(batch_size, max_text_length)`): Batch of durations.
pitch_labels (`torch.Tensor` of shape `(batch_size, max_text_length, 1)`):
Batch of target token-averaged pitch.
energy_labels (`torch.Tensor` of shape `(batch_size, max_text_length, 1)`):
Batch of target token-averaged energy.
duration_mask (`torch.LongTensor`):
Mask used to discern which values the duration loss should be calculated for.
spectrogram_mask (`torch.LongTensor`):
Mask used to discern which values the spectrogam loss should be calculated for.
Returns:
`tuple(torch.FloatTensor)`: Tuple of tensors containing, in order, the L1 loss value, duration predictor
loss value, pitch predictor loss value, and energy predictor loss value.
"""
pitch_and_energy_masks = duration_mask.unsqueeze(-1)
# apply mask to remove padded part
if self.use_masking:
outputs_before_postnet = outputs_before_postnet.masked_select(spectrogram_mask)
if outputs_after_postnet is not None:
outputs_after_postnet = outputs_after_postnet.masked_select(spectrogram_mask)
spectrogram_labels = spectrogram_labels.masked_select(spectrogram_mask)
duration_outputs = duration_outputs.masked_select(duration_mask)
duration_labels = duration_labels.masked_select(duration_mask)
pitch_outputs = pitch_outputs.masked_select(pitch_and_energy_masks)
energy_outputs = energy_outputs.masked_select(pitch_and_energy_masks)
pitch_labels = pitch_labels.masked_select(pitch_and_energy_masks)
energy_labels = energy_labels.masked_select(pitch_and_energy_masks)
# calculate loss
l1_loss = self.l1_criterion(outputs_before_postnet, spectrogram_labels)
if outputs_after_postnet is not None:
l1_loss = l1_loss + self.l1_criterion(outputs_after_postnet, spectrogram_labels)
duration_labels = torch.log(duration_labels.float() + self.log_domain_offset)
duration_loss = self.duration_criterion(duration_outputs, duration_labels)
pitch_loss = self.mse_criterion(pitch_outputs, pitch_labels)
energy_loss = self.mse_criterion(energy_outputs, energy_labels)
# make weighted mask and apply it
if self.use_weighted_masking:
spectrogram_mask = nn.functional.pad(
spectrogram_mask.transpose(1, 2),
[0, spectrogram_labels.size(1) - spectrogram_mask.size(1), 0, 0, 0, 0],
value=False,
).transpose(1, 2)
out_weights = spectrogram_mask.float() / spectrogram_mask.sum(dim=1, keepdim=True).float()
out_weights /= spectrogram_labels.size(0) * spectrogram_labels.size(2)
duration_weights = duration_mask.float() / duration_mask.sum(dim=1, keepdim=True).float()
duration_weights /= duration_labels.size(0)
# apply weight
l1_loss = l1_loss.mul(out_weights).masked_select(spectrogram_mask).sum()
duration_loss = duration_loss.mul(duration_weights).masked_select(duration_mask).sum()
pitch_weights = duration_weights.unsqueeze(-1)
pitch_loss = pitch_loss.mul(pitch_weights).masked_select(pitch_and_energy_masks).sum()
energy_loss = energy_loss.mul(pitch_weights).masked_select(pitch_and_energy_masks).sum()
return l1_loss + duration_loss + pitch_loss + energy_loss
class FastSpeech2ConformerPreTrainedModel(PreTrainedModel):
"""
An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
models.
"""
config_class = FastSpeech2ConformerConfig
base_model_prefix = "fastspeech2_conformer"
main_input_name = "input_ids"
def _init_weights(self, module):
"""Initialize the weights"""
if isinstance(module, (nn.LayerNorm)):
module.bias.data.zero_()
module.weight.data.fill_(1.0)
elif isinstance(module, nn.Conv1d):
nn.init.kaiming_normal_(module.weight)
if module.bias is not None:
key = math.sqrt(module.groups / (module.in_channels * module.kernel_size[0]))
nn.init.uniform_(module.bias, a=-key, b=key)
elif isinstance(module, nn.Embedding):
module.weight.data.normal_()
if module.padding_idx is not None:
module.weight.data[module.padding_idx].zero_()
elif isinstance(module, FastSpeech2ConformerAttention):
nn.init.xavier_uniform_(module.pos_bias_u)
nn.init.xavier_uniform_(module.pos_bias_v)
def _set_gradient_checkpointing(self, module, value=False):
if isinstance(module, FastSpeech2ConformerEncoder):
module.gradient_checkpointing = value
@add_start_docstrings(
"""FastSpeech2Conformer Model.""",
FASTSPEECH2_CONFORMER_START_DOCSTRING,
)
class FastSpeech2ConformerModel(FastSpeech2ConformerPreTrainedModel):
"""
FastSpeech 2 module.
This is a module of FastSpeech 2 described in 'FastSpeech 2: Fast and High-Quality End-to-End Text to Speech'
https://arxiv.org/abs/2006.04558. Instead of quantized pitch and energy, we use token-averaged value introduced in
FastPitch: Parallel Text-to-speech with Pitch Prediction. The encoder and decoder are Conformers instead of regular
Transformers.
"""
def __init__(self, config: FastSpeech2ConformerConfig):
super().__init__(config)
self.config = config
# store hyperparameters
self.vocab_size = config.vocab_size
self.num_mel_bins = config.num_mel_bins
self.hidden_size = config.hidden_size
self.reduction_factor = config.reduction_factor
self.stop_gradient_from_pitch_predictor = config.stop_gradient_from_pitch_predictor
self.stop_gradient_from_energy_predictor = config.stop_gradient_from_energy_predictor
self.multilingual_model = config.num_languages is not None and config.num_languages > 1
if self.multilingual_model:
self.language_id_embedding = torch.nn.Embedding(config.num_languages, self.hidden_size)
self.multispeaker_model = config.num_speakers is not None and config.num_speakers > 1
if self.multispeaker_model:
self.speaker_id_embedding = torch.nn.Embedding(config.num_speakers, config.hidden_size)
self.speaker_embed_dim = config.speaker_embed_dim
if self.speaker_embed_dim:
self.projection = nn.Linear(config.hidden_size + self.speaker_embed_dim, config.hidden_size)
self.encoder = FastSpeech2ConformerEncoder(config, config.encoder_config, use_encoder_input_layer=True)
self.duration_predictor = FastSpeech2ConformerDurationPredictor(config)
self.pitch_predictor = FastSpeech2ConformerVariancePredictor(
config,
num_layers=config.pitch_predictor_layers,
num_chans=config.pitch_predictor_channels,
kernel_size=config.pitch_predictor_kernel_size,
dropout_rate=config.pitch_predictor_dropout,
)
# continuous pitch + FastPitch style avg
self.pitch_embed = FastSpeech2ConformerVarianceEmbedding(
out_channels=self.hidden_size,
kernel_size=config.pitch_embed_kernel_size,
padding=(config.pitch_embed_kernel_size - 1) // 2,
dropout_rate=config.pitch_embed_dropout,
)
self.energy_predictor = FastSpeech2ConformerVariancePredictor(
config,
num_layers=config.energy_predictor_layers,
num_chans=config.energy_predictor_channels,
kernel_size=config.energy_predictor_kernel_size,
dropout_rate=config.energy_predictor_dropout,
)
# continuous energy + FastPitch style avg
self.energy_embed = FastSpeech2ConformerVarianceEmbedding(
out_channels=self.hidden_size,
kernel_size=config.energy_embed_kernel_size,
padding=(config.energy_embed_kernel_size - 1) // 2,
dropout_rate=config.energy_embed_dropout,
)
# The decoder is an encoder
self.decoder = FastSpeech2ConformerEncoder(config, config.decoder_config, use_encoder_input_layer=False)
self.speech_decoder_postnet = FastSpeech2ConformerSpeechDecoderPostnet(config)
self.criterion = FastSpeech2ConformerLoss(config)
self.post_init()
@replace_return_docstrings(output_type=FastSpeech2ConformerModelOutput, config_class=_CONFIG_FOR_DOC)
def forward(
self,
input_ids: torch.LongTensor,
attention_mask: Optional[torch.LongTensor] = None,
spectrogram_labels: Optional[torch.FloatTensor] = None,
duration_labels: Optional[torch.LongTensor] = None,
pitch_labels: Optional[torch.FloatTensor] = None,
energy_labels: Optional[torch.FloatTensor] = None,
speaker_ids: Optional[torch.LongTensor] = None,
lang_ids: Optional[torch.LongTensor] = None,
speaker_embedding: Optional[torch.FloatTensor] = None,
return_dict: Optional[bool] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
) -> Union[Tuple, FastSpeech2ConformerModelOutput]:
"""
Args:
input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
Input sequence of text vectors.
attention_mask (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*, defaults to `None`):
Mask to avoid performing convolution and attention on padding token indices. Mask values selected in
`[0, 1]`: 0 for tokens that are **masked**, 1 for tokens that are **not masked**.
spectrogram_labels (`torch.FloatTensor` of shape `(batch_size, max_spectrogram_length, num_mel_bins)`, *optional*, defaults to `None`):
Batch of padded target features.
duration_labels (`torch.LongTensor` of shape `(batch_size, sequence_length + 1)`, *optional*, defaults to `None`):
Batch of padded durations.
pitch_labels (`torch.FloatTensor` of shape `(batch_size, sequence_length + 1, 1)`, *optional*, defaults to `None`):
Batch of padded token-averaged pitch.
energy_labels (`torch.FloatTensor` of shape `(batch_size, sequence_length + 1, 1)`, *optional*, defaults to `None`):
Batch of padded token-averaged energy.
speaker_ids (`torch.LongTensor` of shape `(batch_size, 1)`, *optional*, defaults to `None`):
Speaker ids used to condition features of speech output by the model.
lang_ids (`torch.LongTensor` of shape `(batch_size, 1)`, *optional*, defaults to `None`):
Language ids used to condition features of speech output by the model.
speaker_embedding (`torch.FloatTensor` of shape `(batch_size, embedding_dim)`, *optional*, defaults to `None`):
Embedding containing conditioning signals for the features of the speech.
return_dict (`bool`, *optional*, defaults to `None`):
Whether or not to return a [`FastSpeech2ConformerModelOutput`] instead of a plain tuple.
output_attentions (`bool`, *optional*, defaults to `None`):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
returned tensors for more detail.
output_hidden_states (`bool`, *optional*, defaults to `None`):
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
for more detail.
Returns:
Example:
```python
>>> from transformers import (
... FastSpeech2ConformerTokenizer,
... FastSpeech2ConformerModel,
... FastSpeech2ConformerHifiGan,
... )
>>> tokenizer = FastSpeech2ConformerTokenizer.from_pretrained("espnet/fastspeech2_conformer")
>>> inputs = tokenizer("some text to convert to speech", return_tensors="pt")
>>> input_ids = inputs["input_ids"]
>>> model = FastSpeech2ConformerModel.from_pretrained("espnet/fastspeech2_conformer")
>>> output_dict = model(input_ids, return_dict=True)
>>> spectrogram = output_dict["spectrogram"]
>>> vocoder = FastSpeech2ConformerHifiGan.from_pretrained("espnet/fastspeech2_conformer_hifigan")
>>> waveform = vocoder(spectrogram)
>>> print(waveform.shape)
torch.Size([1, 49664])
```
"""
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
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
)
if attention_mask is None:
attention_mask = torch.ones(input_ids.shape, device=input_ids.device)
has_missing_labels = (
spectrogram_labels is None or duration_labels is None or pitch_labels is None or energy_labels is None
)
if self.training and has_missing_labels:
raise ValueError("All labels must be provided to run in training mode.")
# forward encoder
text_masks = attention_mask.unsqueeze(-2)
encoder_outputs = self.encoder(
input_ids,
text_masks,
output_hidden_states=output_hidden_states,
output_attentions=output_attentions,
return_dict=return_dict,
)
hidden_states = encoder_outputs[0]
# Integrate with language id, speaker id, and speaker embedding
if self.multispeaker_model and speaker_ids is not None:
speaker_id_embeddings = self.speaker_id_embedding(speaker_ids.view(-1))
hidden_states = hidden_states + speaker_id_embeddings.unsqueeze(1)
if self.multilingual_model and lang_ids is not None:
language_id_embbedings = self.language_id_embedding(lang_ids.view(-1))
hidden_states = hidden_states + language_id_embbedings.unsqueeze(1)
if self.speaker_embed_dim is not None and speaker_embedding is not None:
embeddings_expanded = (
nn.functional.normalize(speaker_embedding).unsqueeze(1).expand(-1, hidden_states.size(1), -1)
)
hidden_states = self.projection(torch.cat([hidden_states, embeddings_expanded], dim=-1))
# forward duration predictor and variance predictors
duration_mask = ~attention_mask.bool()
if self.stop_gradient_from_pitch_predictor:
pitch_predictions = self.pitch_predictor(hidden_states.detach(), duration_mask.unsqueeze(-1))
else:
pitch_predictions = self.pitch_predictor(hidden_states, duration_mask.unsqueeze(-1))
if self.stop_gradient_from_energy_predictor:
energy_predictions = self.energy_predictor(hidden_states.detach(), duration_mask.unsqueeze(-1))
else:
energy_predictions = self.energy_predictor(hidden_states, duration_mask.unsqueeze(-1))
duration_predictions = self.duration_predictor(hidden_states)
duration_predictions = duration_predictions.masked_fill(duration_mask, 0.0)
if not self.training:
# use prediction in inference
embedded_pitch_curve = self.pitch_embed(pitch_predictions)
embedded_energy_curve = self.energy_embed(energy_predictions)
hidden_states = hidden_states + embedded_energy_curve + embedded_pitch_curve
hidden_states = length_regulator(hidden_states, duration_predictions, self.config.speaking_speed)
else:
# use groundtruth in training
embedded_pitch_curve = self.pitch_embed(pitch_labels)
embedded_energy_curve = self.energy_embed(energy_labels)
hidden_states = hidden_states + embedded_energy_curve + embedded_pitch_curve
hidden_states = length_regulator(hidden_states, duration_labels)
# forward decoder
if not self.training:
hidden_mask = None
else:
spectrogram_mask = (spectrogram_labels != -100).any(dim=-1)
spectrogram_mask = spectrogram_mask.int()
if self.reduction_factor > 1:
length_dim = spectrogram_mask.shape[1] - spectrogram_mask.shape[1] % self.reduction_factor
spectrogram_mask = spectrogram_mask[:, :, :length_dim]
hidden_mask = spectrogram_mask.unsqueeze(-2)
decoder_outputs = self.decoder(
hidden_states,
hidden_mask,
output_hidden_states=output_hidden_states,
output_attentions=output_attentions,
return_dict=return_dict,
)
outputs_before_postnet, outputs_after_postnet = self.speech_decoder_postnet(decoder_outputs[0])
loss = None
if self.training:
# calculate loss
loss_duration_mask = ~duration_mask
loss_spectrogram_mask = spectrogram_mask.unsqueeze(-1).bool()
loss = self.criterion(
outputs_after_postnet=outputs_after_postnet,
outputs_before_postnet=outputs_before_postnet,
duration_outputs=duration_predictions,
pitch_outputs=pitch_predictions,
energy_outputs=energy_predictions,
spectrogram_labels=spectrogram_labels,
duration_labels=duration_labels,
pitch_labels=pitch_labels,
energy_labels=energy_labels,
duration_mask=loss_duration_mask,
spectrogram_mask=loss_spectrogram_mask,
)
if not return_dict:
postnet_outputs = (outputs_after_postnet,)
audio_feature_predictions = (
duration_predictions,
pitch_predictions,
energy_predictions,
)
outputs = postnet_outputs + encoder_outputs + decoder_outputs[1:] + audio_feature_predictions
return ((loss,) + outputs) if loss is not None else outputs
return FastSpeech2ConformerModelOutput(
loss=loss,
spectrogram=outputs_after_postnet,
encoder_last_hidden_state=encoder_outputs.last_hidden_state,
encoder_hidden_states=encoder_outputs.hidden_states,
encoder_attentions=encoder_outputs.attentions,
decoder_hidden_states=decoder_outputs.hidden_states,
decoder_attentions=decoder_outputs.attentions,
duration_outputs=duration_predictions,
pitch_outputs=pitch_predictions,
energy_outputs=energy_predictions,
)
# Copied from transformers.models.speecht5.modeling_speecht5.HifiGanResidualBlock
class HifiGanResidualBlock(nn.Module):
def __init__(self, channels, kernel_size=3, dilation=(1, 3, 5), leaky_relu_slope=0.1):
super().__init__()
self.leaky_relu_slope = leaky_relu_slope
self.convs1 = nn.ModuleList(
[
nn.Conv1d(
channels,
channels,
kernel_size,
stride=1,
dilation=dilation[i],
padding=self.get_padding(kernel_size, dilation[i]),
)
for i in range(len(dilation))
]
)
self.convs2 = nn.ModuleList(
[
nn.Conv1d(
channels,
channels,
kernel_size,
stride=1,
dilation=1,
padding=self.get_padding(kernel_size, 1),
)
for _ in range(len(dilation))
]
)
def get_padding(self, kernel_size, dilation=1):
return (kernel_size * dilation - dilation) // 2
def apply_weight_norm(self):
for layer in self.convs1:
nn.utils.weight_norm(layer)
for layer in self.convs2:
nn.utils.weight_norm(layer)
def remove_weight_norm(self):
for layer in self.convs1:
nn.utils.remove_weight_norm(layer)
for layer in self.convs2:
nn.utils.remove_weight_norm(layer)
def forward(self, hidden_states):
for conv1, conv2 in zip(self.convs1, self.convs2):
residual = hidden_states
hidden_states = nn.functional.leaky_relu(hidden_states, self.leaky_relu_slope)
hidden_states = conv1(hidden_states)
hidden_states = nn.functional.leaky_relu(hidden_states, self.leaky_relu_slope)
hidden_states = conv2(hidden_states)
hidden_states = hidden_states + residual
return hidden_states
@add_start_docstrings(
"""HiFi-GAN vocoder.""",
HIFIGAN_START_DOCSTRING,
)
# Copied from transformers.models.speecht5.modeling_speecht5.SpeechT5HifiGan with SpeechT5->FastSpeech2Conformer
class FastSpeech2ConformerHifiGan(PreTrainedModel):
config_class = FastSpeech2ConformerHifiGanConfig
main_input_name = "spectrogram"
def __init__(self, config: FastSpeech2ConformerHifiGanConfig):
super().__init__(config)
self.num_kernels = len(config.resblock_kernel_sizes)
self.num_upsamples = len(config.upsample_rates)
self.conv_pre = nn.Conv1d(
config.model_in_dim,
config.upsample_initial_channel,
kernel_size=7,
stride=1,
padding=3,
)
self.upsampler = nn.ModuleList()
for i, (upsample_rate, kernel_size) in enumerate(zip(config.upsample_rates, config.upsample_kernel_sizes)):
self.upsampler.append(
nn.ConvTranspose1d(
config.upsample_initial_channel // (2**i),
config.upsample_initial_channel // (2 ** (i + 1)),
kernel_size=kernel_size,
stride=upsample_rate,
padding=(kernel_size - upsample_rate) // 2,
)
)
self.resblocks = nn.ModuleList()
for i in range(len(self.upsampler)):
channels = config.upsample_initial_channel // (2 ** (i + 1))
for kernel_size, dilation in zip(config.resblock_kernel_sizes, config.resblock_dilation_sizes):
self.resblocks.append(HifiGanResidualBlock(channels, kernel_size, dilation, config.leaky_relu_slope))
self.conv_post = nn.Conv1d(channels, 1, kernel_size=7, stride=1, padding=3)
self.register_buffer("mean", torch.zeros(config.model_in_dim))
self.register_buffer("scale", torch.ones(config.model_in_dim))
# Initialize weights and apply final processing
self.post_init()
def _init_weights(self, module):
"""Initialize the weights."""
if isinstance(module, (nn.Linear, nn.Conv1d)):
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
if module.bias is not None:
module.bias.data.zero_()
def apply_weight_norm(self):
nn.utils.weight_norm(self.conv_pre)
for layer in self.upsampler:
nn.utils.weight_norm(layer)
for layer in self.resblocks:
layer.apply_weight_norm()
nn.utils.weight_norm(self.conv_post)
def remove_weight_norm(self):
nn.utils.remove_weight_norm(self.conv_pre)
for layer in self.upsampler:
nn.utils.remove_weight_norm(layer)
for layer in self.resblocks:
layer.remove_weight_norm()
nn.utils.remove_weight_norm(self.conv_post)
def forward(self, spectrogram: torch.FloatTensor) -> torch.FloatTensor:
r"""
Converts a log-mel spectrogram into a speech waveform. Passing a batch of log-mel spectrograms returns a batch
of speech waveforms. Passing a single, un-batched log-mel spectrogram returns a single, un-batched speech
waveform.
Args:
spectrogram (`torch.FloatTensor`):
Tensor containing the log-mel spectrograms. Can be batched and of shape `(batch_size, sequence_length,
config.model_in_dim)`, or un-batched and of shape `(sequence_length, config.model_in_dim)`.
Returns:
`torch.FloatTensor`: Tensor containing the speech waveform. If the input spectrogram is batched, will be of
shape `(batch_size, num_frames,)`. If un-batched, will be of shape `(num_frames,)`.
"""
if self.config.normalize_before:
spectrogram = (spectrogram - self.mean) / self.scale
is_batched = spectrogram.dim() == 3
if not is_batched:
spectrogram = spectrogram.unsqueeze(0)
hidden_states = spectrogram.transpose(2, 1)
hidden_states = self.conv_pre(hidden_states)
for i in range(self.num_upsamples):
hidden_states = nn.functional.leaky_relu(hidden_states, self.config.leaky_relu_slope)
hidden_states = self.upsampler[i](hidden_states)
res_state = self.resblocks[i * self.num_kernels](hidden_states)
for j in range(1, self.num_kernels):
res_state += self.resblocks[i * self.num_kernels + j](hidden_states)
hidden_states = res_state / self.num_kernels
hidden_states = nn.functional.leaky_relu(hidden_states)
hidden_states = self.conv_post(hidden_states)
hidden_states = torch.tanh(hidden_states)
if not is_batched:
# remove batch dim and collapse tensor to 1-d audio waveform
waveform = hidden_states.squeeze(0).transpose(1, 0).view(-1)
else:
# remove seq-len dim since this collapses to 1
waveform = hidden_states.squeeze(1)
return waveform
@add_start_docstrings(
"The FastSpeech2ConformerModel with a FastSpeech2ConformerHifiGan vocoder head that performs text-to-speech (waveform).",
FASTSPEECH2_CONFORMER_WITH_HIFIGAN_START_DOCSTRING,
)
class FastSpeech2ConformerWithHifiGan(PreTrainedModel):
config_class = FastSpeech2ConformerWithHifiGanConfig
def __init__(self, config: FastSpeech2ConformerWithHifiGanConfig):
super().__init__(config)
self.model = FastSpeech2ConformerModel(config.model_config)
self.vocoder = FastSpeech2ConformerHifiGan(config.vocoder_config)
self.config = config
@replace_return_docstrings(
output_type=FastSpeech2ConformerWithHifiGanOutput, config_class=FastSpeech2ConformerWithHifiGanConfig
)
def forward(
self,
input_ids: torch.LongTensor,
attention_mask: Optional[torch.LongTensor] = None,
spectrogram_labels: Optional[torch.FloatTensor] = None,
duration_labels: Optional[torch.LongTensor] = None,
pitch_labels: Optional[torch.FloatTensor] = None,
energy_labels: Optional[torch.FloatTensor] = None,
speaker_ids: Optional[torch.LongTensor] = None,
lang_ids: Optional[torch.LongTensor] = None,
speaker_embedding: Optional[torch.FloatTensor] = None,
return_dict: Optional[bool] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
) -> Union[Tuple, FastSpeech2ConformerModelOutput]:
"""
Args:
input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
Input sequence of text vectors.
attention_mask (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*, defaults to `None`):
Mask to avoid performing convolution and attention on padding token indices. Mask values selected in
`[0, 1]`: 0 for tokens that are **masked**, 1 for tokens that are **not masked**.
spectrogram_labels (`torch.FloatTensor` of shape `(batch_size, max_spectrogram_length, num_mel_bins)`, *optional*, defaults to `None`):
Batch of padded target features.
duration_labels (`torch.LongTensor` of shape `(batch_size, sequence_length + 1)`, *optional*, defaults to `None`):
Batch of padded durations.
pitch_labels (`torch.FloatTensor` of shape `(batch_size, sequence_length + 1, 1)`, *optional*, defaults to `None`):
Batch of padded token-averaged pitch.
energy_labels (`torch.FloatTensor` of shape `(batch_size, sequence_length + 1, 1)`, *optional*, defaults to `None`):
Batch of padded token-averaged energy.
speaker_ids (`torch.LongTensor` of shape `(batch_size, 1)`, *optional*, defaults to `None`):
Speaker ids used to condition features of speech output by the model.
lang_ids (`torch.LongTensor` of shape `(batch_size, 1)`, *optional*, defaults to `None`):
Language ids used to condition features of speech output by the model.
speaker_embedding (`torch.FloatTensor` of shape `(batch_size, embedding_dim)`, *optional*, defaults to `None`):
Embedding containing conditioning signals for the features of the speech.
return_dict (`bool`, *optional*, defaults to `None`):
Whether or not to return a [`FastSpeech2ConformerModelOutput`] instead of a plain tuple.
output_attentions (`bool`, *optional*, defaults to `None`):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
returned tensors for more detail.
output_hidden_states (`bool`, *optional*, defaults to `None`):
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
for more detail.
Returns:
Example:
```python
>>> from transformers import (
... FastSpeech2ConformerTokenizer,
... FastSpeech2ConformerWithHifiGan,
... )
>>> tokenizer = FastSpeech2ConformerTokenizer.from_pretrained("espnet/fastspeech2_conformer")
>>> inputs = tokenizer("some text to convert to speech", return_tensors="pt")
>>> input_ids = inputs["input_ids"]
>>> model = FastSpeech2ConformerWithHifiGan.from_pretrained("espnet/fastspeech2_conformer_with_hifigan")
>>> output_dict = model(input_ids, return_dict=True)
>>> waveform = output_dict["waveform"]
>>> print(waveform.shape)
torch.Size([1, 49664])
```
"""
return_dict = return_dict if return_dict is not None else self.config.model_config.use_return_dict
output_attentions = (
output_attentions if output_attentions is not None else self.config.model_config.output_attentions
)
output_hidden_states = (
output_hidden_states if output_hidden_states is not None else self.config.model_config.output_hidden_states
)
model_outputs = self.model(
input_ids,
attention_mask,
spectrogram_labels=spectrogram_labels,
duration_labels=duration_labels,
pitch_labels=pitch_labels,
energy_labels=energy_labels,
speaker_ids=speaker_ids,
lang_ids=lang_ids,
speaker_embedding=speaker_embedding,
return_dict=return_dict,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
)
if not return_dict:
has_missing_labels = (
spectrogram_labels is None or duration_labels is None or pitch_labels is None or energy_labels is None
)
if has_missing_labels:
spectrogram = model_outputs[0]
else:
spectrogram = model_outputs[1]
else:
spectrogram = model_outputs["spectrogram"]
waveform = self.vocoder(spectrogram)
if not return_dict:
return model_outputs + (waveform,)
return FastSpeech2ConformerWithHifiGanOutput(waveform=waveform, **model_outputs)
|
transformers/src/transformers/models/fastspeech2_conformer/modeling_fastspeech2_conformer.py/0
|
{
"file_path": "transformers/src/transformers/models/fastspeech2_conformer/modeling_fastspeech2_conformer.py",
"repo_id": "transformers",
"token_count": 33002
}
| 356
|
# coding=utf-8
# Copyright 2021 Google AI and The HuggingFace Inc. team. 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.
"""FNet model configuration"""
from ...configuration_utils import PretrainedConfig
from ...utils import logging
logger = logging.get_logger(__name__)
class FNetConfig(PretrainedConfig):
r"""
This is the configuration class to store the configuration of a [`FNetModel`]. It is used to instantiate an FNet
model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
defaults will yield a similar configuration to that of the FNet
[google/fnet-base](https://huggingface.co/google/fnet-base) architecture.
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PretrainedConfig`] for more information.
Args:
vocab_size (`int`, *optional*, defaults to 32000):
Vocabulary size of the FNet model. Defines the number of different tokens that can be represented by the
`inputs_ids` passed when calling [`FNetModel`] or [`TFFNetModel`].
hidden_size (`int`, *optional*, defaults to 768):
Dimension of the encoder layers and the pooler layer.
num_hidden_layers (`int`, *optional*, defaults to 12):
Number of hidden layers in the Transformer encoder.
intermediate_size (`int`, *optional*, defaults to 3072):
Dimension of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
hidden_act (`str` or `function`, *optional*, defaults to `"gelu_new"`):
The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
`"relu"`, `"selu"` and `"gelu_new"` are supported.
hidden_dropout_prob (`float`, *optional*, defaults to 0.1):
The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
max_position_embeddings (`int`, *optional*, defaults to 512):
The maximum sequence length that this model might ever be used with. Typically set this to something large
just in case (e.g., 512 or 1024 or 2048).
type_vocab_size (`int`, *optional*, defaults to 4):
The vocabulary size of the `token_type_ids` passed when calling [`FNetModel`] or [`TFFNetModel`].
initializer_range (`float`, *optional*, defaults to 0.02):
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
layer_norm_eps (`float`, *optional*, defaults to 1e-12):
The epsilon used by the layer normalization layers.
use_tpu_fourier_optimizations (`bool`, *optional*, defaults to `False`):
Determines whether to use TPU optimized FFTs. If `True`, the model will favor axis-wise FFTs transforms.
Set to `False` for GPU/CPU hardware, in which case n-dimensional FFTs are used.
tpu_short_seq_length (`int`, *optional*, defaults to 512):
The sequence length that is expected by the model when using TPUs. This will be used to initialize the DFT
matrix only when *use_tpu_fourier_optimizations* is set to `True` and the input sequence is shorter than or
equal to 4096 tokens.
Example:
```python
>>> from transformers import FNetConfig, FNetModel
>>> # Initializing a FNet fnet-base style configuration
>>> configuration = FNetConfig()
>>> # Initializing a model (with random weights) from the fnet-base style configuration
>>> model = FNetModel(configuration)
>>> # Accessing the model configuration
>>> configuration = model.config
```"""
model_type = "fnet"
def __init__(
self,
vocab_size=32000,
hidden_size=768,
num_hidden_layers=12,
intermediate_size=3072,
hidden_act="gelu_new",
hidden_dropout_prob=0.1,
max_position_embeddings=512,
type_vocab_size=4,
initializer_range=0.02,
layer_norm_eps=1e-12,
use_tpu_fourier_optimizations=False,
tpu_short_seq_length=512,
pad_token_id=3,
bos_token_id=1,
eos_token_id=2,
**kwargs,
):
super().__init__(pad_token_id=pad_token_id, bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)
self.vocab_size = vocab_size
self.max_position_embeddings = max_position_embeddings
self.hidden_size = hidden_size
self.num_hidden_layers = num_hidden_layers
self.intermediate_size = intermediate_size
self.hidden_act = hidden_act
self.hidden_dropout_prob = hidden_dropout_prob
self.initializer_range = initializer_range
self.type_vocab_size = type_vocab_size
self.layer_norm_eps = layer_norm_eps
self.use_tpu_fourier_optimizations = use_tpu_fourier_optimizations
self.tpu_short_seq_length = tpu_short_seq_length
|
transformers/src/transformers/models/fnet/configuration_fnet.py/0
|
{
"file_path": "transformers/src/transformers/models/fnet/configuration_fnet.py",
"repo_id": "transformers",
"token_count": 2042
}
| 357
|
# Copyright 2021 The HuggingFace Team. 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.
from typing import TYPE_CHECKING
from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_flax_available, is_torch_available
_import_structure = {
"configuration_gpt_neo": ["GPTNeoConfig", "GPTNeoOnnxConfig"],
}
try:
if not is_torch_available():
raise OptionalDependencyNotAvailable()
except OptionalDependencyNotAvailable:
pass
else:
_import_structure["modeling_gpt_neo"] = [
"GPTNeoForCausalLM",
"GPTNeoForQuestionAnswering",
"GPTNeoForSequenceClassification",
"GPTNeoForTokenClassification",
"GPTNeoModel",
"GPTNeoPreTrainedModel",
"load_tf_weights_in_gpt_neo",
]
try:
if not is_flax_available():
raise OptionalDependencyNotAvailable()
except OptionalDependencyNotAvailable:
pass
else:
_import_structure["modeling_flax_gpt_neo"] = [
"FlaxGPTNeoForCausalLM",
"FlaxGPTNeoModel",
"FlaxGPTNeoPreTrainedModel",
]
if TYPE_CHECKING:
from .configuration_gpt_neo import GPTNeoConfig, GPTNeoOnnxConfig
try:
if not is_torch_available():
raise OptionalDependencyNotAvailable()
except OptionalDependencyNotAvailable:
pass
else:
from .modeling_gpt_neo import (
GPTNeoForCausalLM,
GPTNeoForQuestionAnswering,
GPTNeoForSequenceClassification,
GPTNeoForTokenClassification,
GPTNeoModel,
GPTNeoPreTrainedModel,
load_tf_weights_in_gpt_neo,
)
try:
if not is_flax_available():
raise OptionalDependencyNotAvailable()
except OptionalDependencyNotAvailable:
pass
else:
from .modeling_flax_gpt_neo import FlaxGPTNeoForCausalLM, FlaxGPTNeoModel, FlaxGPTNeoPreTrainedModel
else:
import sys
sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)
|
transformers/src/transformers/models/gpt_neo/__init__.py/0
|
{
"file_path": "transformers/src/transformers/models/gpt_neo/__init__.py",
"repo_id": "transformers",
"token_count": 1041
}
| 358
|
# Copyright 2021 The EleutherAI and HuggingFace Teams. 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.
from typing import TYPE_CHECKING
from ...utils import (
OptionalDependencyNotAvailable,
_LazyModule,
is_flax_available,
is_tf_available,
is_torch_available,
)
_import_structure = {"configuration_gptj": ["GPTJConfig", "GPTJOnnxConfig"]}
try:
if not is_torch_available():
raise OptionalDependencyNotAvailable()
except OptionalDependencyNotAvailable:
pass
else:
_import_structure["modeling_gptj"] = [
"GPTJForCausalLM",
"GPTJForQuestionAnswering",
"GPTJForSequenceClassification",
"GPTJModel",
"GPTJPreTrainedModel",
]
try:
if not is_tf_available():
raise OptionalDependencyNotAvailable()
except OptionalDependencyNotAvailable:
pass
else:
_import_structure["modeling_tf_gptj"] = [
"TFGPTJForCausalLM",
"TFGPTJForQuestionAnswering",
"TFGPTJForSequenceClassification",
"TFGPTJModel",
"TFGPTJPreTrainedModel",
]
try:
if not is_flax_available():
raise OptionalDependencyNotAvailable()
except OptionalDependencyNotAvailable:
pass
else:
_import_structure["modeling_flax_gptj"] = [
"FlaxGPTJForCausalLM",
"FlaxGPTJModel",
"FlaxGPTJPreTrainedModel",
]
if TYPE_CHECKING:
from .configuration_gptj import GPTJConfig, GPTJOnnxConfig
try:
if not is_torch_available():
raise OptionalDependencyNotAvailable()
except OptionalDependencyNotAvailable:
pass
else:
from .modeling_gptj import (
GPTJForCausalLM,
GPTJForQuestionAnswering,
GPTJForSequenceClassification,
GPTJModel,
GPTJPreTrainedModel,
)
try:
if not is_tf_available():
raise OptionalDependencyNotAvailable()
except OptionalDependencyNotAvailable:
pass
else:
from .modeling_tf_gptj import (
TFGPTJForCausalLM,
TFGPTJForQuestionAnswering,
TFGPTJForSequenceClassification,
TFGPTJModel,
TFGPTJPreTrainedModel,
)
try:
if not is_flax_available():
raise OptionalDependencyNotAvailable()
except OptionalDependencyNotAvailable:
pass
else:
from .modeling_flax_gptj import FlaxGPTJForCausalLM, FlaxGPTJModel, FlaxGPTJPreTrainedModel
else:
import sys
sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)
|
transformers/src/transformers/models/gptj/__init__.py/0
|
{
"file_path": "transformers/src/transformers/models/gptj/__init__.py",
"repo_id": "transformers",
"token_count": 1315
}
| 359
|
# coding=utf-8
# Copyright 2022 The HuggingFace Inc. team. 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.
"""
Convert GroupViT checkpoints from the original repository.
URL: https://github.com/NVlabs/GroupViT
"""
import argparse
import requests
import torch
from PIL import Image
from transformers import CLIPProcessor, GroupViTConfig, GroupViTModel
def rename_key(name):
# vision encoder
if "img_encoder.pos_embed" in name:
name = name.replace("img_encoder.pos_embed", "vision_model.embeddings.position_embeddings")
if "img_encoder.patch_embed.proj" in name:
name = name.replace("img_encoder.patch_embed.proj", "vision_model.embeddings.patch_embeddings.projection")
if "img_encoder.patch_embed.norm" in name:
name = name.replace("img_encoder.patch_embed.norm", "vision_model.embeddings.layernorm")
if "img_encoder.layers" in name:
name = name.replace("img_encoder.layers", "vision_model.encoder.stages")
if "blocks" in name and "res" not in name:
name = name.replace("blocks", "layers")
if "attn" in name and "pre_assign" not in name:
name = name.replace("attn", "self_attn")
if "proj" in name and "self_attn" in name and "text" not in name:
name = name.replace("proj", "out_proj")
if "pre_assign_attn.attn.proj" in name:
name = name.replace("pre_assign_attn.attn.proj", "pre_assign_attn.attn.out_proj")
if "norm1" in name:
name = name.replace("norm1", "layer_norm1")
if "norm2" in name and "pre_assign" not in name:
name = name.replace("norm2", "layer_norm2")
if "img_encoder.norm" in name:
name = name.replace("img_encoder.norm", "vision_model.layernorm")
# text encoder
if "text_encoder.token_embedding" in name:
name = name.replace("text_encoder.token_embedding", "text_model.embeddings.token_embedding")
if "text_encoder.positional_embedding" in name:
name = name.replace("text_encoder.positional_embedding", "text_model.embeddings.position_embedding.weight")
if "text_encoder.transformer.resblocks." in name:
name = name.replace("text_encoder.transformer.resblocks.", "text_model.encoder.layers.")
if "ln_1" in name:
name = name.replace("ln_1", "layer_norm1")
if "ln_2" in name:
name = name.replace("ln_2", "layer_norm2")
if "c_fc" in name:
name = name.replace("c_fc", "fc1")
if "c_proj" in name:
name = name.replace("c_proj", "fc2")
if "text_encoder" in name:
name = name.replace("text_encoder", "text_model")
if "ln_final" in name:
name = name.replace("ln_final", "final_layer_norm")
# projection layers
if "img_projector.linear_hidden." in name:
name = name.replace("img_projector.linear_hidden.", "visual_projection.")
if "img_projector.linear_out." in name:
name = name.replace("img_projector.linear_out.", "visual_projection.3.")
if "text_projector.linear_hidden" in name:
name = name.replace("text_projector.linear_hidden", "text_projection")
if "text_projector.linear_out" in name:
name = name.replace("text_projector.linear_out", "text_projection.3")
return name
def convert_state_dict(orig_state_dict, config):
for key in orig_state_dict.copy().keys():
val = orig_state_dict.pop(key)
if "qkv" in key:
# weights and biases of the key, value and query projections of vision encoder's attention layers require special treatment:
# we need to split them up into separate matrices/vectors
key_split = key.split(".")
stage_num, layer_num = int(key_split[2]), int(key_split[4])
dim = config.vision_config.hidden_size
if "weight" in key:
orig_state_dict[
f"vision_model.encoder.stages.{stage_num}.layers.{layer_num}.self_attn.q_proj.weight"
] = val[:dim, :]
orig_state_dict[
f"vision_model.encoder.stages.{stage_num}.layers.{layer_num}.self_attn.k_proj.weight"
] = val[dim : dim * 2, :]
orig_state_dict[
f"vision_model.encoder.stages.{stage_num}.layers.{layer_num}.self_attn.v_proj.weight"
] = val[-dim:, :]
else:
orig_state_dict[
f"vision_model.encoder.stages.{stage_num}.layers.{layer_num}.self_attn.q_proj.bias"
] = val[:dim]
orig_state_dict[
f"vision_model.encoder.stages.{stage_num}.layers.{layer_num}.self_attn.k_proj.bias"
] = val[dim : dim * 2]
orig_state_dict[
f"vision_model.encoder.stages.{stage_num}.layers.{layer_num}.self_attn.v_proj.bias"
] = val[-dim:]
elif "in_proj" in key:
# weights and biases of the key, value and query projections of text encoder's attention layers require special treatment:
# we need to split them up into separate matrices/vectors
key_split = key.split(".")
layer_num = int(key_split[3])
dim = config.text_config.hidden_size
if "weight" in key:
orig_state_dict[f"text_model.encoder.layers.{layer_num}.self_attn.q_proj.weight"] = val[:dim, :]
orig_state_dict[f"text_model.encoder.layers.{layer_num}.self_attn.k_proj.weight"] = val[
dim : dim * 2, :
]
orig_state_dict[f"text_model.encoder.layers.{layer_num}.self_attn.v_proj.weight"] = val[-dim:, :]
else:
orig_state_dict[f"text_model.encoder.layers.{layer_num}.self_attn.q_proj.bias"] = val[:dim]
orig_state_dict[f"text_model.encoder.layers.{layer_num}.self_attn.k_proj.bias"] = val[dim : dim * 2]
orig_state_dict[f"text_model.encoder.layers.{layer_num}.self_attn.v_proj.bias"] = val[-dim:]
else:
new_name = rename_key(key)
# squeeze if necessary
if (
"text_projection.0" in new_name
or "text_projection.3" in new_name
or "visual_projection.0" in new_name
or "visual_projection.3" in new_name
):
orig_state_dict[new_name] = val.squeeze_()
else:
orig_state_dict[new_name] = val
return orig_state_dict
# We will verify our results on an image of cute cats
def prepare_img():
url = "http://images.cocodataset.org/val2017/000000039769.jpg"
im = Image.open(requests.get(url, stream=True).raw)
return im
@torch.no_grad()
def convert_groupvit_checkpoint(
checkpoint_path, pytorch_dump_folder_path, model_name="groupvit-gcc-yfcc", push_to_hub=False
):
"""
Copy/paste/tweak model's weights to the Transformers design.
"""
config = GroupViTConfig()
model = GroupViTModel(config).eval()
state_dict = torch.load(checkpoint_path, map_location="cpu")["model"]
new_state_dict = convert_state_dict(state_dict, config)
missing_keys, unexpected_keys = model.load_state_dict(new_state_dict, strict=False)
assert missing_keys == ["text_model.embeddings.position_ids"]
assert (unexpected_keys == ["multi_label_logit_scale"]) or (len(unexpected_keys) == 0)
# verify result
processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32")
image = prepare_img()
inputs = processor(text=["a photo of a cat", "a photo of a dog"], images=image, padding=True, return_tensors="pt")
with torch.no_grad():
outputs = model(**inputs)
if model_name == "groupvit-gcc-yfcc":
expected_logits = torch.tensor([[13.3523, 6.3629]])
elif model_name == "groupvit-gcc-redcaps":
expected_logits = torch.tensor([[16.1873, 8.6230]])
else:
raise ValueError(f"Model name {model_name} not supported.")
assert torch.allclose(outputs.logits_per_image, expected_logits, atol=1e-3)
processor.save_pretrained(pytorch_dump_folder_path)
model.save_pretrained(pytorch_dump_folder_path)
print("Successfully saved processor and model to", pytorch_dump_folder_path)
if push_to_hub:
print("Pushing to the hub...")
processor.push_to_hub(model_name, organization="nielsr")
model.push_to_hub(model_name, organization="nielsr")
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument(
"--pytorch_dump_folder_path", default=None, type=str, help="Path to dump the processor and PyTorch model."
)
parser.add_argument("--checkpoint_path", default=None, type=str, help="Path to GroupViT checkpoint")
parser.add_argument(
"--model_name",
default="groupvit-gccy-fcc",
type=str,
help="Name of the model. Expecting either 'groupvit-gcc-yfcc' or 'groupvit-gcc-redcaps'",
)
parser.add_argument(
"--push_to_hub",
action="store_true",
help="Whether or not to push the converted model and processor to the 🤗 hub using the provided `model_name`.",
)
args = parser.parse_args()
convert_groupvit_checkpoint(args.checkpoint_path, args.pytorch_dump_folder_path, args.model_name, args.push_to_hub)
|
transformers/src/transformers/models/groupvit/convert_groupvit_nvlab_to_hf.py/0
|
{
"file_path": "transformers/src/transformers/models/groupvit/convert_groupvit_nvlab_to_hf.py",
"repo_id": "transformers",
"token_count": 4252
}
| 360
|
# coding=utf-8
# Copyright 2024 The HuggingFace Inc. team. 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.
"""Idefics2 model configuration"""
import os
from typing import Union
from ...configuration_utils import PretrainedConfig
from ...utils import logging
from ..auto import CONFIG_MAPPING
logger = logging.get_logger(__name__)
class Idefics2VisionConfig(PretrainedConfig):
r"""
This is the configuration class to store the configuration of a [`Idefics2VisionModel`]. It is used to instantiate a
Idefics2 vision encoder according to the specified arguments, defining the model architecture. Instantiating a
configuration with the defaults will yield a similar configuration to that of the SigLIP checkpoint
[google/siglip-base-patch16-224](https://huggingface.co/google/siglip-base-patch16-224) used in the Idefics2 model
[HuggingFaceM4/idefics2-8b](https://huggingface.co/HuggingFaceM4/idefics2-8b).
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PretrainedConfig`] for more information.
Args:
hidden_size (`int`, *optional*, defaults to 768):
Dimensionality of the encoder layers and the pooler layer.
intermediate_size (`int`, *optional*, defaults to 3072):
Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
num_hidden_layers (`int`, *optional*, defaults to 12):
Number of hidden layers in the Transformer encoder.
num_attention_heads (`int`, *optional*, defaults to 12):
Number of attention heads for each attention layer in the Transformer encoder.
num_channels (`int`, *optional*, defaults to 3):
Number of channels in the input images.
image_size (`int`, *optional*, defaults to 224):
The size (resolution) of each image.
patch_size (`int`, *optional*, defaults to 32):
The size (resolution) of each patch.
hidden_act (`str` or `function`, *optional*, defaults to `"gelu_pytorch_tanh"`):
The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
`"relu"`, `"selu"` and `"gelu_new"` `"quick_gelu"` are supported.
layer_norm_eps (`float`, *optional*, defaults to 1e-06):
The epsilon used by the layer normalization layers.
attention_dropout (`float`, *optional*, defaults to 0.0):
The dropout ratio for the attention probabilities.
intializer_range (`float`, *optional*, defaults to 0.02):
The standard deviation for initializing all weight matrices in the model.
Example:
```python
>>> from transformers.models.idefics2.modeling_idefics2 import Idefics2VisionTransformer
>>> from transformers.models.idefics2.configuration_idefics2 import Idefics2VisionConfig
>>> # Initializing a Idefics2VisionConfig with google/siglip-base-patch16-224 style configuration
>>> configuration = Idefics2VisionConfig()
>>> # Initializing a Idefics2VisionTransformer (with random weights) from the google/siglip-base-patch16-224 style configuration
>>> model = Idefics2VisionTransformer(configuration)
>>> # Accessing the model configuration
>>> configuration = model.config
```"""
model_type = "idefics2"
def __init__(
self,
hidden_size=768,
intermediate_size=3072,
num_hidden_layers=12,
num_attention_heads=12,
num_channels=3,
image_size=224,
patch_size=32,
hidden_act="gelu_pytorch_tanh",
layer_norm_eps=1e-6,
attention_dropout=0.0,
initializer_range=0.02,
**kwargs,
):
super().__init__(**kwargs)
self.hidden_size = hidden_size
self.intermediate_size = intermediate_size
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
self.num_channels = num_channels
self.patch_size = patch_size
self.image_size = image_size
self.attention_dropout = attention_dropout
self.layer_norm_eps = layer_norm_eps
self.hidden_act = hidden_act
self.initializer_range = initializer_range
@classmethod
def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
cls._set_token_in_kwargs(kwargs)
config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
# get the vision config dict if we are loading from Idefics2Config
if config_dict.get("model_type") == "idefics2":
config_dict = config_dict["vision_config"]
if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
logger.warning(
f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
)
return cls.from_dict(config_dict, **kwargs)
class Idefics2PerceiverConfig(PretrainedConfig):
r"""
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PretrainedConfig`] for more information.
Args:
hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
The non-linear activation function (function or string) in the perceiver block.
resampler_n_latents (`int`, *optional*, defaults to 64):
Number of latent embeddings to resample ("compress") the input sequence to (usually < 128).
resampler_depth (`int`, *optional*, defaults to 3):
Depth of the Perceiver Resampler (Transformer w/ cross attention). Should be shallow (<= 3).
resampler_n_heads (`int`, *optional*, defaults to 16):
Number of heads in each Transformer block (for multi-headed self-attention).
resampler_head_dim (`int`, *optional*, defaults to 96):
Dimensionality of each head projection in the Transformer block.
num_key_value_heads (`int`, *optional*, defaults to 4):
Number of key-value heads in the perceiver attention block.
attention_dropout (`float`, *optional*, defaults to 0.0):
The dropout ratio for the attention probabilities.
"""
model_type = "idefics2"
def __init__(
self,
hidden_act="silu",
resampler_n_latents=64,
resampler_depth=3,
resampler_n_heads=16,
resampler_head_dim=96,
num_key_value_heads=4,
attention_dropout=0.0,
**kwargs,
):
self.hidden_act = hidden_act
self.resampler_n_latents = resampler_n_latents
self.resampler_depth = resampler_depth
self.resampler_n_heads = resampler_n_heads
self.num_key_value_heads = num_key_value_heads
self.resampler_head_dim = resampler_head_dim
self.attention_dropout = attention_dropout
if self.num_key_value_heads > self.resampler_n_heads:
raise ValueError(
f"num_key_value_heads={self.num_key_value_heads} must be less than or equal to"
f" resampler_n_heads={self.resampler_n_heads}"
)
super().__init__(**kwargs)
class Idefics2Config(PretrainedConfig):
r"""
This is the configuration class to store the configuration of a [`Idefics2Model`]. It is used to instantiate a
Idefics2 model according to the specified arguments, defining the model architecture. Instantiating a
configuration with the defaults will yield a similar configuration to that of the model of the Idefics2
[HuggingFaceM4/idefics2-8b](https://huggingface.co/HuggingFaceM4/idefics2-8b) architecture.
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PretrainedConfig`] for more information.
Args:
use_cache (`bool`, *optional*, defaults to `True`):
Whether or not the model should cache the key/value pairs of the attention mechanism.
image_token_id (`int`, *optional*, defaults to 32001):
The id of the "image" token.
tie_word_embeddings (`bool`, *optional*, defaults to `False`):
Whether or not to tie the word embeddings with the token embeddings.
vision_config (`IdeficsVisionConfig` or `dict`, *optional*):
Custom vision config or dict
perceiver_config (`IdeficsPerceiverConfig` or `dict`, *optional*):
Custom perceiver config or dict
text_config (`MistralConfig` or `dict`, *optional*):
Custom text config or dict for the text model
Example:
```python
>>> from transformers import Idefics2Model, Idefics2Config
>>> # Initializing configuration
>>> configuration = Idefics2Config()
>>> # Initializing a model from the configuration
>>> model = Idefics2Model(configuration)
>>> # Accessing the model configuration
>>> configuration = model.config
```"""
model_type = "idefics2"
is_composition = True
def __init__(
self,
use_cache=True,
image_token_id=32_001,
tie_word_embeddings=False,
vision_config=None,
perceiver_config=None,
text_config=None,
**kwargs,
):
self.image_token_id = image_token_id
self.use_cache = use_cache
self.tie_word_embeddings = tie_word_embeddings
if perceiver_config is None:
self.perceiver_config = Idefics2PerceiverConfig()
logger.info("perciver_config is None, using default perceiver config")
elif isinstance(perceiver_config, dict):
self.perceiver_config = Idefics2PerceiverConfig(**perceiver_config)
elif isinstance(perceiver_config, Idefics2PerceiverConfig):
self.perceiver_config = perceiver_config
if vision_config is None:
self.vision_config = Idefics2VisionConfig()
logger.info("vision_config is None, using default vision config")
elif isinstance(vision_config, dict):
self.vision_config = Idefics2VisionConfig(**vision_config)
elif isinstance(vision_config, Idefics2VisionConfig):
self.vision_config = vision_config
if isinstance(text_config, dict):
text_config["model_type"] = text_config["model_type"] if "model_type" in text_config else "mistral"
text_config = CONFIG_MAPPING[text_config["model_type"]](**text_config)
elif text_config is None:
logger.info("text_config is None, using default text config")
text_config = CONFIG_MAPPING["mistral"](
max_position_embeddings=4096 * 8,
rms_norm_eps=1e-5,
# None in the original configuration_mistral, we set it to the unk_token_id
pad_token_id=0,
tie_word_embeddings=False,
)
self.text_config = text_config
super().__init__(**kwargs, tie_word_embeddings=tie_word_embeddings)
|
transformers/src/transformers/models/idefics2/configuration_idefics2.py/0
|
{
"file_path": "transformers/src/transformers/models/idefics2/configuration_idefics2.py",
"repo_id": "transformers",
"token_count": 4585
}
| 361
|
# coding=utf-8
# Copyright 2023 Microsoft Research and The HuggingFace Inc. team. 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.
from typing import TYPE_CHECKING
from ...utils import (
OptionalDependencyNotAvailable,
_LazyModule,
is_torch_available,
is_vision_available,
)
_import_structure = {
"configuration_kosmos2": ["Kosmos2Config"],
"processing_kosmos2": ["Kosmos2Processor"],
}
try:
if not is_torch_available():
raise OptionalDependencyNotAvailable()
except OptionalDependencyNotAvailable:
pass
else:
_import_structure["modeling_kosmos2"] = [
"Kosmos2ForConditionalGeneration",
"Kosmos2Model",
"Kosmos2PreTrainedModel",
]
if TYPE_CHECKING:
from .configuration_kosmos2 import Kosmos2Config
from .processing_kosmos2 import Kosmos2Processor
try:
if not is_torch_available():
raise OptionalDependencyNotAvailable()
except OptionalDependencyNotAvailable:
pass
else:
from .modeling_kosmos2 import (
Kosmos2ForConditionalGeneration,
Kosmos2Model,
Kosmos2PreTrainedModel,
)
else:
import sys
sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure)
|
transformers/src/transformers/models/kosmos2/__init__.py/0
|
{
"file_path": "transformers/src/transformers/models/kosmos2/__init__.py",
"repo_id": "transformers",
"token_count": 645
}
| 362
|
# Copyright 2021 The HuggingFace Team. 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.
from typing import TYPE_CHECKING
from ...utils import (
OptionalDependencyNotAvailable,
_LazyModule,
is_tf_available,
is_tokenizers_available,
is_torch_available,
)
_import_structure = {
"configuration_led": ["LEDConfig"],
"tokenization_led": ["LEDTokenizer"],
}
try:
if not is_tokenizers_available():
raise OptionalDependencyNotAvailable()
except OptionalDependencyNotAvailable:
pass
else:
_import_structure["tokenization_led_fast"] = ["LEDTokenizerFast"]
try:
if not is_torch_available():
raise OptionalDependencyNotAvailable()
except OptionalDependencyNotAvailable:
pass
else:
_import_structure["modeling_led"] = [
"LEDForConditionalGeneration",
"LEDForQuestionAnswering",
"LEDForSequenceClassification",
"LEDModel",
"LEDPreTrainedModel",
]
try:
if not is_tf_available():
raise OptionalDependencyNotAvailable()
except OptionalDependencyNotAvailable:
pass
else:
_import_structure["modeling_tf_led"] = ["TFLEDForConditionalGeneration", "TFLEDModel", "TFLEDPreTrainedModel"]
if TYPE_CHECKING:
from .configuration_led import LEDConfig
from .tokenization_led import LEDTokenizer
try:
if not is_tokenizers_available():
raise OptionalDependencyNotAvailable()
except OptionalDependencyNotAvailable:
pass
else:
from .tokenization_led_fast import LEDTokenizerFast
try:
if not is_torch_available():
raise OptionalDependencyNotAvailable()
except OptionalDependencyNotAvailable:
pass
else:
from .modeling_led import (
LEDForConditionalGeneration,
LEDForQuestionAnswering,
LEDForSequenceClassification,
LEDModel,
LEDPreTrainedModel,
)
try:
if not is_tf_available():
raise OptionalDependencyNotAvailable()
except OptionalDependencyNotAvailable:
pass
else:
from .modeling_tf_led import TFLEDForConditionalGeneration, TFLEDModel, TFLEDPreTrainedModel
else:
import sys
sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)
|
transformers/src/transformers/models/led/__init__.py/0
|
{
"file_path": "transformers/src/transformers/models/led/__init__.py",
"repo_id": "transformers",
"token_count": 1050
}
| 363
|
# coding=utf-8
# Copyright 2024 The HuggingFace Inc. team.
#
# 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.
"""
Processor class for LLaVa-NeXT.
"""
from typing import List, Optional, Union
from ...feature_extraction_utils import BatchFeature
from ...image_processing_utils import select_best_resolution
from ...image_utils import ImageInput, get_image_size, to_numpy_array
from ...processing_utils import ProcessorMixin
from ...tokenization_utils_base import PaddingStrategy, PreTokenizedInput, TextInput, TruncationStrategy
from ...utils import TensorType, logging
logger = logging.get_logger(__name__)
class LlavaNextProcessor(ProcessorMixin):
r"""
Constructs a LLaVa-NeXT processor which wraps a LLaVa-NeXT image processor and a LLaMa tokenizer into a single processor.
[`LlavaNextProcessor`] offers all the functionalities of [`LlavaNextImageProcessor`] and [`LlamaTokenizerFast`]. See the
[`~LlavaNextProcessor.__call__`] and [`~LlavaNextProcessor.decode`] for more information.
Args:
image_processor ([`LlavaNextImageProcessor`], *optional*):
The image processor is a required input.
tokenizer ([`LlamaTokenizerFast`], *optional*):
The tokenizer is a required input.
patch_size (`int`, *optional*):
Patch size from the vision tower.
vision_feature_select_strategy (`str`, *optional*):
The feature selection strategy used to select the vision feature from the vision backbone.
Shoudl be same as in model's config
chat_template (`str`, *optional*): A Jinja template which will be used to convert lists of messages
in a chat into a tokenizable string.
image_token (`str`, *optional*, defaults to `"<image>"`):
Special token used to denote image location.
"""
attributes = ["image_processor", "tokenizer"]
valid_kwargs = ["chat_template", "patch_size", "vision_feature_select_strategy", "image_token"]
image_processor_class = "AutoImageProcessor"
tokenizer_class = "AutoTokenizer"
def __init__(
self,
image_processor=None,
tokenizer=None,
patch_size=None,
vision_feature_select_strategy=None,
chat_template=None,
image_token="<image>", # set the default and let users change if they have peculiar special tokens in rare cases
**kwargs,
):
self.patch_size = patch_size
self.vision_feature_select_strategy = vision_feature_select_strategy
self.image_token = image_token
super().__init__(image_processor, tokenizer, chat_template=chat_template)
def __call__(
self,
text: Union[TextInput, PreTokenizedInput, List[TextInput], List[PreTokenizedInput]],
images: ImageInput = None,
padding: Union[bool, str, PaddingStrategy] = False,
truncation: Union[bool, str, TruncationStrategy] = None,
max_length: Optional[int] = None,
do_pad: Optional[bool] = True,
return_tensors: Optional[Union[str, TensorType]] = TensorType.PYTORCH,
) -> BatchFeature:
"""
Main method to prepare for the model one or several sequences(s) and image(s). This method forwards the `text`
and `kwargs` arguments to LlamaTokenizerFast's [`~LlamaTokenizerFast.__call__`] if `text` is not `None` to encode
the text. To prepare the image(s), this method forwards the `images` and `kwrags` arguments to
LlavaNextImageProcessor's [`~LlavaNextImageProcessor.__call__`] if `images` is not `None`. Please refer to the doctsring
of the above two methods for more information.
Args:
text (`str`, `List[str]`, `List[List[str]]`):
The sequence or batch of sequences to be encoded. Each sequence can be a string or a list of strings
(pretokenized string). If the sequences are provided as list of strings (pretokenized), you must set
`is_split_into_words=True` (to lift the ambiguity with a batch of sequences).
images (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, `List[PIL.Image.Image]`, `List[np.ndarray]`, `List[torch.Tensor]`):
The image or batch of images to be prepared. Each image can be a PIL image, NumPy array or PyTorch
tensor. Both channels-first and channels-last formats are supported.
padding (`bool`, `str` or [`~utils.PaddingStrategy`], *optional*, defaults to `False`):
Select a strategy to pad the returned sequences (according to the model's padding side and padding
index) among:
- `True` or `'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
sequence if provided).
- `'max_length'`: Pad to a maximum length specified with the argument `max_length` or to the maximum
acceptable input length for the model if that argument is not provided.
- `False` or `'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of different
lengths).
max_length (`int`, *optional*):
Maximum length of the returned list and optionally padding length (see above).
do_pad (`bool`, *optional*, defaults to self.do_pad):
Whether to pad the image. If `True` will pad the images in the batch to the largest image in the batch
and create a pixel mask. Padding will be applied to the bottom and right of the image with zeros.
truncation (`bool`, *optional*):
Activates truncation to cut input sequences longer than `max_length` to `max_length`.
return_tensors (`str` or [`~utils.TensorType`], *optional*):
If set, will return tensors of a particular framework. Acceptable values are:
- `'tf'`: Return TensorFlow `tf.constant` objects.
- `'pt'`: Return PyTorch `torch.Tensor` objects.
- `'np'`: Return NumPy `np.ndarray` objects.
- `'jax'`: Return JAX `jnp.ndarray` objects.
Returns:
[`BatchFeature`]: A [`BatchFeature`] with the following fields:
- **input_ids** -- List of token ids to be fed to a model. Returned when `text` is not `None`.
- **attention_mask** -- List of indices specifying which tokens should be attended to by the model (when
`return_attention_mask=True` or if *"attention_mask"* is in `self.model_input_names` and if `text` is not
`None`).
- **pixel_values** -- Pixel values to be fed to a model. Returned when `images` is not `None`.
"""
if images is not None:
image_inputs = self.image_processor(images, do_pad=do_pad, return_tensors=return_tensors)
else:
image_inputs = {}
if isinstance(text, str):
text = [text]
elif not isinstance(text, list) and not isinstance(text[0], str):
raise ValueError("Invalid input text. Please provide a string, or a list of strings")
if self.patch_size is None or self.vision_feature_select_strategy is None:
prompt_strings = text
logger.warning_once(
"Expanding inputs for image tokens in LLaVa-NeXT should be done in processing. "
"Please add `patch_size` and `vision_feature_select_strategy` to the model's processing config or set directly "
"with `processor.patch_size = {{patch_size}}` and processor.vision_feature_select_strategy = {{vision_feature_select_strategy}}`. "
"Using processors without these attributes in the config is deprecated and will throw an error in v4.47."
)
# cannot infer image expansion length if no images are found
elif not image_inputs:
prompt_strings = text
else:
image_sizes = image_inputs["image_sizes"]
height, width = get_image_size(to_numpy_array(image_inputs["pixel_values"][0][0]))
prompt_strings = []
for image_size, sample in zip(image_sizes, text):
# Replace the image token with the expanded image token sequence
orig_height, orig_width = image_size
num_image_tokens = self._get_number_of_features(orig_height, orig_width, height, width)
if self.vision_feature_select_strategy == "default":
num_image_tokens -= 1
sample = sample.replace(self.image_token, self.image_token * num_image_tokens)
prompt_strings.append(sample)
text_inputs = self.tokenizer(
prompt_strings,
return_tensors=return_tensors,
padding=padding,
truncation=truncation,
max_length=max_length,
)
return BatchFeature(data={**text_inputs, **image_inputs})
def _get_number_of_features(self, orig_height: int, orig_width: int, height: int, width: int) -> int:
image_grid_pinpoints = self.image_processor.image_grid_pinpoints
height_best_resolution, width_best_resolution = select_best_resolution(
[orig_height, orig_width], image_grid_pinpoints
)
scale_height, scale_width = height_best_resolution // height, width_best_resolution // width
patches_height = height // self.patch_size
patches_width = width // self.patch_size
unpadded_features, newline_features = self._get_unpadded_features(
orig_height, orig_width, patches_height, patches_width, scale_height, scale_width
)
# The base patch covers the entire image (+1 for the CLS)
base_features = patches_height * patches_width + 1
num_image_tokens = unpadded_features + newline_features + base_features
return num_image_tokens
def _get_unpadded_features(self, height, width, patches_height, patches_width, scale_height, scale_width):
"""
Get number of features for a given image with height/width. LLaVA-NeXT is different from LLaVA
because it divided each image into patches depending on its resolution. Therefore we need to calculate how many
patches an image is divided into and get the number of features from that.
"""
current_width = patches_height * scale_height
current_height = patches_width * scale_width
original_aspect_ratio = width / height
current_aspect_ratio = current_width / current_height
if original_aspect_ratio > current_aspect_ratio:
new_height = (height * current_width) // width
padding = (current_height - new_height) // 2
current_height -= padding * 2
else:
new_width = (width * current_height) // height
padding = (current_width - new_width) // 2
current_width -= padding * 2
unpadded_features = current_height * current_width
newline_features = current_height
return (unpadded_features, newline_features)
# Copied from transformers.models.clip.processing_clip.CLIPProcessor.batch_decode with CLIP->Llama
def batch_decode(self, *args, **kwargs):
"""
This method forwards all its arguments to LlamaTokenizerFast's [`~PreTrainedTokenizer.batch_decode`]. Please
refer to the docstring of this method for more information.
"""
return self.tokenizer.batch_decode(*args, **kwargs)
# Copied from transformers.models.clip.processing_clip.CLIPProcessor.decode with CLIP->Llama
def decode(self, *args, **kwargs):
"""
This method forwards all its arguments to LlamaTokenizerFast's [`~PreTrainedTokenizer.decode`]. Please refer to
the docstring of this method for more information.
"""
return self.tokenizer.decode(*args, **kwargs)
@property
# Copied from transformers.models.clip.processing_clip.CLIPProcessor.model_input_names
def model_input_names(self):
tokenizer_input_names = self.tokenizer.model_input_names
image_processor_input_names = self.image_processor.model_input_names
return list(dict.fromkeys(tokenizer_input_names + image_processor_input_names))
|
transformers/src/transformers/models/llava_next/processing_llava_next.py/0
|
{
"file_path": "transformers/src/transformers/models/llava_next/processing_llava_next.py",
"repo_id": "transformers",
"token_count": 4939
}
| 364
|
# coding=utf-8
# Copyright 2022, The LongT5 Authors and HuggingFace Inc.
#
# 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.
"""LongT5 model configuration"""
from typing import Mapping
from ...configuration_utils import PretrainedConfig
from ...onnx import OnnxSeq2SeqConfigWithPast
from ...utils import logging
logger = logging.get_logger(__name__)
class LongT5Config(PretrainedConfig):
r"""
This is the configuration class to store the configuration of a [`LongT5Model`] or a [`FlaxLongT5Model`]. It is
used to instantiate a LongT5 model according to the specified arguments, defining the model architecture.
Instantiating a configuration with the defaults will yield a similar configuration to that of the LongT5
[google/long-t5-local-base](https://huggingface.co/google/long-t5-local-base) architecture.
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PretrainedConfig`] for more information.
Arguments:
vocab_size (`int`, *optional*, defaults to 32128):
Vocabulary size of the LongT5 model. Defines the number of different tokens that can be represented by the
`inputs_ids` passed when calling [`LongT5Model`].
d_model (`int`, *optional*, defaults to 512):
Size of the encoder layers and the pooler layer.
d_kv (`int`, *optional*, defaults to 64):
Size of the key, query, value projections per attention head. `d_kv` has to be equal to `d_model //
num_heads`.
d_ff (`int`, *optional*, defaults to 2048):
Size of the intermediate feed forward layer in each `LongT5Block`.
num_layers (`int`, *optional*, defaults to 6):
Number of hidden layers in the Transformer encoder.
num_decoder_layers (`int`, *optional*):
Number of hidden layers in the Transformer decoder. Will use the same value as `num_layers` if not set.
num_heads (`int`, *optional*, defaults to 8):
Number of attention heads for each attention layer in the Transformer encoder.
local_radius (`int`, *optional*, defaults to 127)
Number of tokens to the left/right for each token to locally self-attend in a local attention mechanism.
global_block_size (`int`, *optional*, defaults to 16)
Lenght of blocks an input sequence is divided into for a global token representation. Used only for
`encoder_attention_type = "transient-global"`.
relative_attention_num_buckets (`int`, *optional*, defaults to 32):
The number of buckets to use for each attention layer.
relative_attention_max_distance (`int`, *optional*, defaults to 128):
The maximum distance of the longer sequences for the bucket separation.
dropout_rate (`float`, *optional*, defaults to 0.1):
The ratio for all dropout layers.
layer_norm_eps (`float`, *optional*, defaults to 1e-6):
The epsilon used by the layer normalization layers.
initializer_factor (`float`, *optional*, defaults to 1):
A factor for initializing all weight matrices (should be kept to 1, used internally for initialization
testing).
feed_forward_proj (`string`, *optional*, defaults to `"relu"`):
Type of feed forward layer to be used. Should be one of `"relu"` or `"gated-gelu"`. LongT5v1.1 uses the
`"gated-gelu"` feed forward projection. Original LongT5 implementation uses `"gated-gelu"`.
encoder_attention_type (`string`, *optional*, defaults to `"local"`):
Type of encoder attention to be used. Should be one of `"local"` or `"transient-global"`, which are
supported by LongT5 implementation.
use_cache (`bool`, *optional*, defaults to `True`):
Whether or not the model should return the last key/values attentions (not used by all models).
"""
model_type = "longt5"
keys_to_ignore_at_inference = ["past_key_values"]
attribute_map = {"hidden_size": "d_model", "num_attention_heads": "num_heads", "num_hidden_layers": "num_layers"}
def __init__(
self,
vocab_size=32128,
d_model=512,
d_kv=64,
d_ff=2048,
num_layers=6,
num_decoder_layers=None,
num_heads=8,
local_radius=127,
global_block_size=16,
relative_attention_num_buckets=32,
relative_attention_max_distance=128,
dropout_rate=0.1,
layer_norm_epsilon=1e-6,
initializer_factor=1.0,
feed_forward_proj="relu",
is_encoder_decoder=True,
encoder_attention_type="local",
use_cache=True,
pad_token_id=0,
eos_token_id=1,
**kwargs,
):
self.vocab_size = vocab_size
self.d_model = d_model
self.d_kv = d_kv
self.d_ff = d_ff
self.num_layers = num_layers
# default = symmetry
self.num_decoder_layers = num_decoder_layers if num_decoder_layers is not None else self.num_layers
self.num_heads = num_heads
self.local_radius = local_radius
self.global_block_size = global_block_size
self.relative_attention_num_buckets = relative_attention_num_buckets
self.relative_attention_max_distance = relative_attention_max_distance
self.dropout_rate = dropout_rate
self.layer_norm_epsilon = layer_norm_epsilon
self.initializer_factor = initializer_factor
self.feed_forward_proj = feed_forward_proj
self.encoder_attention_type = encoder_attention_type
self.use_cache = use_cache
act_info = self.feed_forward_proj.split("-")
self.dense_act_fn = act_info[-1]
self.is_gated_act = act_info[0] == "gated"
if len(act_info) > 1 and act_info[0] != "gated" or len(act_info) > 2:
raise ValueError(
f"`feed_forward_proj`: {feed_forward_proj} is not a valid activation function of the dense layer. "
"Please make sure `feed_forward_proj` is of the format `gated-{ACT_FN}` or `{ACT_FN}`, e.g. "
"'gated-gelu' or 'relu'"
)
# for backwards compatibility
if feed_forward_proj == "gated-gelu":
self.dense_act_fn = "gelu_new"
super().__init__(
pad_token_id=pad_token_id,
eos_token_id=eos_token_id,
is_encoder_decoder=is_encoder_decoder,
**kwargs,
)
class LongT5OnnxConfig(OnnxSeq2SeqConfigWithPast):
@property
def inputs(self) -> Mapping[str, Mapping[int, str]]:
common_inputs = {
"input_ids": {0: "batch", 1: "encoder_sequence"},
"attention_mask": {0: "batch", 1: "encoder_sequence"},
}
if self.use_past:
common_inputs["attention_mask"][1] = "past_encoder_sequence + sequence"
common_inputs["decoder_input_ids"] = {0: "batch"}
common_inputs["decoder_attention_mask"] = {0: "batch", 1: "past_decoder_sequence + sequence"}
else:
common_inputs["decoder_input_ids"] = {0: "batch", 1: "decoder_sequence"}
common_inputs["decoder_attention_mask"] = {0: "batch", 1: "decoder_sequence"}
if self.use_past:
self.fill_with_past_key_values_(common_inputs, direction="inputs")
return common_inputs
@property
def default_onnx_opset(self) -> int:
return 13
|
transformers/src/transformers/models/longt5/configuration_longt5.py/0
|
{
"file_path": "transformers/src/transformers/models/longt5/configuration_longt5.py",
"repo_id": "transformers",
"token_count": 3264
}
| 365
|
# Copyright 2020 The HuggingFace Team. 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.
import argparse
import json
import os
import socket
import time
import warnings
from pathlib import Path
from typing import Dict, List, Union
from zipfile import ZipFile
import numpy as np
import torch
from huggingface_hub.hf_api import list_models
from torch import nn
from tqdm import tqdm
from transformers import MarianConfig, MarianMTModel, MarianTokenizer
def remove_suffix(text: str, suffix: str):
if text.endswith(suffix):
return text[: -len(suffix)]
return text # or whatever
def remove_prefix(text: str, prefix: str):
if text.startswith(prefix):
return text[len(prefix) :]
return text # or whatever
def convert_encoder_layer(opus_dict, layer_prefix: str, converter: dict):
sd = {}
for k in opus_dict:
if not k.startswith(layer_prefix):
continue
stripped = remove_prefix(k, layer_prefix)
v = opus_dict[k].T # besides embeddings, everything must be transposed.
sd[converter[stripped]] = torch.tensor(v).squeeze()
return sd
def load_layers_(layer_lst: nn.ModuleList, opus_state: dict, converter, is_decoder=False):
for i, layer in enumerate(layer_lst):
layer_tag = f"decoder_l{i + 1}_" if is_decoder else f"encoder_l{i + 1}_"
sd = convert_encoder_layer(opus_state, layer_tag, converter)
layer.load_state_dict(sd, strict=False)
def find_pretrained_model(src_lang: str, tgt_lang: str) -> List[str]:
"""Find models that can accept src_lang as input and return tgt_lang as output."""
prefix = "Helsinki-NLP/opus-mt-"
model_list = list_models()
model_ids = [x.id for x in model_list if x.id.startswith("Helsinki-NLP")]
src_and_targ = [
remove_prefix(m, prefix).lower().split("-") for m in model_ids if "+" not in m
] # + cant be loaded.
matching = [f"{prefix}{a}-{b}" for (a, b) in src_and_targ if src_lang in a and tgt_lang in b]
return matching
def add_emb_entries(wemb, final_bias, n_special_tokens=1):
vsize, d_model = wemb.shape
embs_to_add = np.zeros((n_special_tokens, d_model))
new_embs = np.concatenate([wemb, embs_to_add])
bias_to_add = np.zeros((n_special_tokens, 1))
new_bias = np.concatenate((final_bias, bias_to_add), axis=1)
return new_embs, new_bias
def _cast_yaml_str(v):
bool_dct = {"true": True, "false": False}
if not isinstance(v, str):
return v
elif v in bool_dct:
return bool_dct[v]
try:
return int(v)
except (TypeError, ValueError):
return v
def cast_marian_config(raw_cfg: Dict[str, str]) -> Dict:
return {k: _cast_yaml_str(v) for k, v in raw_cfg.items()}
CONFIG_KEY = "special:model.yml"
def load_config_from_state_dict(opus_dict):
import yaml
cfg_str = "".join([chr(x) for x in opus_dict[CONFIG_KEY]])
yaml_cfg = yaml.load(cfg_str[:-1], Loader=yaml.BaseLoader)
return cast_marian_config(yaml_cfg)
def find_model_file(dest_dir): # this one better
model_files = list(Path(dest_dir).glob("*.npz"))
if len(model_files) != 1:
raise ValueError(f"Found more than one model file: {model_files}")
model_file = model_files[0]
return model_file
# Group Names Logic: change long opus model names to something shorter, like opus-mt-en-ROMANCE
ROM_GROUP = (
"fr+fr_BE+fr_CA+fr_FR+wa+frp+oc+ca+rm+lld+fur+lij+lmo+es+es_AR+es_CL+es_CO+es_CR+es_DO+es_EC+es_ES+es_GT"
"+es_HN+es_MX+es_NI+es_PA+es_PE+es_PR+es_SV+es_UY+es_VE+pt+pt_br+pt_BR+pt_PT+gl+lad+an+mwl+it+it_IT+co"
"+nap+scn+vec+sc+ro+la"
)
GROUPS = [
("cmn+cn+yue+ze_zh+zh_cn+zh_CN+zh_HK+zh_tw+zh_TW+zh_yue+zhs+zht+zh", "ZH"),
(ROM_GROUP, "ROMANCE"),
("de+nl+fy+af+da+fo+is+no+nb+nn+sv", "NORTH_EU"),
("da+fo+is+no+nb+nn+sv", "SCANDINAVIA"),
("se+sma+smj+smn+sms", "SAMI"),
("nb_NO+nb+nn_NO+nn+nog+no_nb+no", "NORWAY"),
("ga+cy+br+gd+kw+gv", "CELTIC"), # https://en.wikipedia.org/wiki/Insular_Celtic_languages
]
GROUP_TO_OPUS_NAME = {
"opus-mt-ZH-de": "cmn+cn+yue+ze_zh+zh_cn+zh_CN+zh_HK+zh_tw+zh_TW+zh_yue+zhs+zht+zh-de",
"opus-mt-ZH-fi": "cmn+cn+yue+ze_zh+zh_cn+zh_CN+zh_HK+zh_tw+zh_TW+zh_yue+zhs+zht+zh-fi",
"opus-mt-ZH-sv": "cmn+cn+yue+ze_zh+zh_cn+zh_CN+zh_HK+zh_tw+zh_TW+zh_yue+zhs+zht+zh-sv",
"opus-mt-SCANDINAVIA-SCANDINAVIA": "da+fo+is+no+nb+nn+sv-da+fo+is+no+nb+nn+sv",
"opus-mt-NORTH_EU-NORTH_EU": "de+nl+fy+af+da+fo+is+no+nb+nn+sv-de+nl+fy+af+da+fo+is+no+nb+nn+sv",
"opus-mt-de-ZH": "de-cmn+cn+yue+ze_zh+zh_cn+zh_CN+zh_HK+zh_tw+zh_TW+zh_yue+zhs+zht+zh",
"opus-mt-en_el_es_fi-en_el_es_fi": "en+el+es+fi-en+el+es+fi",
"opus-mt-en-ROMANCE": (
"en-fr+fr_BE+fr_CA+fr_FR+wa+frp+oc+ca+rm+lld+fur+lij+lmo+es+es_AR+es_CL+es_CO+es_CR+es_DO"
"+es_EC+es_ES+es_GT+es_HN+es_MX+es_NI+es_PA+es_PE+es_PR+es_SV+es_UY+es_VE+pt+pt_br+pt_BR"
"+pt_PT+gl+lad+an+mwl+it+it_IT+co+nap+scn+vec+sc+ro+la"
),
"opus-mt-en-CELTIC": "en-ga+cy+br+gd+kw+gv",
"opus-mt-es-NORWAY": "es-nb_NO+nb+nn_NO+nn+nog+no_nb+no",
"opus-mt-fi_nb_no_nn_ru_sv_en-SAMI": "fi+nb+no+nn+ru+sv+en-se+sma+smj+smn+sms",
"opus-mt-fi-ZH": "fi-cmn+cn+yue+ze_zh+zh_cn+zh_CN+zh_HK+zh_tw+zh_TW+zh_yue+zhs+zht+zh",
"opus-mt-fi-NORWAY": "fi-nb_NO+nb+nn_NO+nn+nog+no_nb+no",
"opus-mt-ROMANCE-en": (
"fr+fr_BE+fr_CA+fr_FR+wa+frp+oc+ca+rm+lld+fur+lij+lmo+es+es_AR+es_CL+es_CO+es_CR+es_DO"
"+es_EC+es_ES+es_GT+es_HN+es_MX+es_NI+es_PA+es_PE+es_PR+es_SV+es_UY+es_VE+pt+pt_br+pt_BR"
"+pt_PT+gl+lad+an+mwl+it+it_IT+co+nap+scn+vec+sc+ro+la-en"
),
"opus-mt-CELTIC-en": "ga+cy+br+gd+kw+gv-en",
"opus-mt-sv-ZH": "sv-cmn+cn+yue+ze_zh+zh_cn+zh_CN+zh_HK+zh_tw+zh_TW+zh_yue+zhs+zht+zh",
"opus-mt-sv-NORWAY": "sv-nb_NO+nb+nn_NO+nn+nog+no_nb+no",
}
OPUS_GITHUB_URL = "https://github.com/Helsinki-NLP/OPUS-MT-train/blob/master/models/"
ORG_NAME = "Helsinki-NLP/"
def convert_opus_name_to_hf_name(x):
"""For OPUS-MT-Train/ DEPRECATED"""
for substr, grp_name in GROUPS:
x = x.replace(substr, grp_name)
return x.replace("+", "_")
def convert_hf_name_to_opus_name(hf_model_name):
"""
Relies on the assumption that there are no language codes like pt_br in models that are not in GROUP_TO_OPUS_NAME.
"""
hf_model_name = remove_prefix(hf_model_name, ORG_NAME)
if hf_model_name in GROUP_TO_OPUS_NAME:
opus_w_prefix = GROUP_TO_OPUS_NAME[hf_model_name]
else:
opus_w_prefix = hf_model_name.replace("_", "+")
return remove_prefix(opus_w_prefix, "opus-mt-")
def get_system_metadata(repo_root):
import git
return {
"helsinki_git_sha": git.Repo(path=repo_root, search_parent_directories=True).head.object.hexsha,
"transformers_git_sha": git.Repo(path=".", search_parent_directories=True).head.object.hexsha,
"port_machine": socket.gethostname(),
"port_time": time.strftime("%Y-%m-%d-%H:%M"),
}
# docstyle-ignore
FRONT_MATTER_TEMPLATE = """---
language:
{}
tags:
- translation
license: apache-2.0
---
"""
DEFAULT_REPO = "Tatoeba-Challenge"
DEFAULT_MODEL_DIR = os.path.join(DEFAULT_REPO, "models")
def write_model_card(
hf_model_name: str,
repo_root=DEFAULT_REPO,
save_dir=Path("marian_converted"),
dry_run=False,
extra_metadata={},
) -> str:
"""
Copy the most recent model's readme section from opus, and add metadata. upload command: aws s3 sync model_card_dir
s3://models.huggingface.co/bert/Helsinki-NLP/ --dryrun
"""
import pandas as pd
hf_model_name = remove_prefix(hf_model_name, ORG_NAME)
opus_name: str = convert_hf_name_to_opus_name(hf_model_name)
if repo_root not in ("OPUS-MT-train", "Tatoeba-Challenge"):
raise ValueError(f"Repos root is {repo_root}. Expected either OPUS-MT-train or Tatoeba-Challenge")
opus_readme_path = Path(repo_root).joinpath("models", opus_name, "README.md")
if not (opus_readme_path.exists()):
raise ValueError(f"Readme file {opus_readme_path} not found")
opus_src, opus_tgt = [x.split("+") for x in opus_name.split("-")]
readme_url = f"https://github.com/Helsinki-NLP/{repo_root}/tree/master/models/{opus_name}/README.md"
s, t = ",".join(opus_src), ",".join(opus_tgt)
metadata = {
"hf_name": hf_model_name,
"source_languages": s,
"target_languages": t,
"opus_readme_url": readme_url,
"original_repo": repo_root,
"tags": ["translation"],
}
metadata.update(extra_metadata)
metadata.update(get_system_metadata(repo_root))
# combine with opus markdown
extra_markdown = (
f"### {hf_model_name}\n\n* source group: {metadata['src_name']} \n* target group: "
f"{metadata['tgt_name']} \n* OPUS readme: [{opus_name}]({readme_url})\n"
)
content = opus_readme_path.open().read()
content = content.split("\n# ")[-1] # Get the lowest level 1 header in the README -- the most recent model.
splat = content.split("*")[2:]
print(splat[3])
content = "*".join(splat)
content = (
FRONT_MATTER_TEMPLATE.format(metadata["src_alpha2"])
+ extra_markdown
+ "\n* "
+ content.replace("download", "download original weights")
)
items = "\n\n".join([f"- {k}: {v}" for k, v in metadata.items()])
sec3 = "\n### System Info: \n" + items
content += sec3
if dry_run:
return content, metadata
sub_dir = save_dir / f"opus-mt-{hf_model_name}"
sub_dir.mkdir(exist_ok=True)
dest = sub_dir / "README.md"
dest.open("w").write(content)
pd.Series(metadata).to_json(sub_dir / "metadata.json")
# if dry_run:
return content, metadata
def make_registry(repo_path="Opus-MT-train/models"):
if not (Path(repo_path) / "fr-en" / "README.md").exists():
raise ValueError(
f"repo_path:{repo_path} does not exist: "
"You must run: git clone git@github.com:Helsinki-NLP/Opus-MT-train.git before calling."
)
results = {}
for p in Path(repo_path).iterdir():
n_dash = p.name.count("-")
if n_dash == 0:
continue
else:
lns = list(open(p / "README.md").readlines())
results[p.name] = _parse_readme(lns)
return [(k, v["pre-processing"], v["download"], v["download"][:-4] + ".test.txt") for k, v in results.items()]
def convert_all_sentencepiece_models(model_list=None, repo_path=None, dest_dir=Path("marian_converted")):
"""Requires 300GB"""
save_dir = Path("marian_ckpt")
dest_dir = Path(dest_dir)
dest_dir.mkdir(exist_ok=True)
save_paths = []
if model_list is None:
model_list: list = make_registry(repo_path=repo_path)
for k, prepro, download, test_set_url in tqdm(model_list):
if "SentencePiece" not in prepro: # dont convert BPE models.
continue
if not os.path.exists(save_dir / k):
download_and_unzip(download, save_dir / k)
pair_name = convert_opus_name_to_hf_name(k)
convert(save_dir / k, dest_dir / f"opus-mt-{pair_name}")
save_paths.append(dest_dir / f"opus-mt-{pair_name}")
return save_paths
def lmap(f, x) -> List:
return list(map(f, x))
def fetch_test_set(test_set_url):
import wget
fname = wget.download(test_set_url, "opus_test.txt")
lns = Path(fname).open().readlines()
src = lmap(str.strip, lns[::4])
gold = lmap(str.strip, lns[1::4])
mar_model = lmap(str.strip, lns[2::4])
if not (len(gold) == len(mar_model) == len(src)):
raise ValueError(f"Gold, marian and source lengths {len(gold)}, {len(mar_model)}, {len(src)} mismatched")
os.remove(fname)
return src, mar_model, gold
def convert_whole_dir(path=Path("marian_ckpt/")):
for subdir in tqdm(list(path.ls())):
dest_dir = f"marian_converted/{subdir.name}"
if (dest_dir / "pytorch_model.bin").exists():
continue
convert(source_dir, dest_dir)
def _parse_readme(lns):
"""Get link and metadata from opus model card equivalent."""
subres = {}
for ln in [x.strip() for x in lns]:
if not ln.startswith("*"):
continue
ln = ln[1:].strip()
for k in ["download", "dataset", "models", "model", "pre-processing"]:
if ln.startswith(k):
break
else:
continue
if k in ["dataset", "model", "pre-processing"]:
splat = ln.split(":")
_, v = splat
subres[k] = v
elif k == "download":
v = ln.split("(")[-1][:-1]
subres[k] = v
return subres
def save_tokenizer_config(dest_dir: Path, separate_vocabs=False):
dname = dest_dir.name.split("-")
dct = {"target_lang": dname[-1], "source_lang": "-".join(dname[:-1]), "separate_vocabs": separate_vocabs}
save_json(dct, dest_dir / "tokenizer_config.json")
def add_to_vocab_(vocab: Dict[str, int], special_tokens: List[str]):
start = max(vocab.values()) + 1
added = 0
for tok in special_tokens:
if tok in vocab:
continue
vocab[tok] = start + added
added += 1
return added
def find_vocab_file(model_dir):
return list(model_dir.glob("*vocab.yml"))[0]
def find_src_vocab_file(model_dir):
return list(model_dir.glob("*src.vocab.yml"))[0]
def find_tgt_vocab_file(model_dir):
return list(model_dir.glob("*trg.vocab.yml"))[0]
def add_special_tokens_to_vocab(model_dir: Path, separate_vocab=False) -> None:
if separate_vocab:
vocab = load_yaml(find_src_vocab_file(model_dir))
vocab = {k: int(v) for k, v in vocab.items()}
num_added = add_to_vocab_(vocab, ["<pad>"])
save_json(vocab, model_dir / "vocab.json")
vocab = load_yaml(find_tgt_vocab_file(model_dir))
vocab = {k: int(v) for k, v in vocab.items()}
num_added = add_to_vocab_(vocab, ["<pad>"])
save_json(vocab, model_dir / "target_vocab.json")
save_tokenizer_config(model_dir, separate_vocabs=separate_vocab)
else:
vocab = load_yaml(find_vocab_file(model_dir))
vocab = {k: int(v) for k, v in vocab.items()}
num_added = add_to_vocab_(vocab, ["<pad>"])
print(f"added {num_added} tokens to vocab")
save_json(vocab, model_dir / "vocab.json")
save_tokenizer_config(model_dir)
def check_equal(marian_cfg, k1, k2):
v1, v2 = marian_cfg[k1], marian_cfg[k2]
if v1 != v2:
raise ValueError(f"hparams {k1},{k2} differ: {v1} != {v2}")
def check_marian_cfg_assumptions(marian_cfg):
assumed_settings = {
"layer-normalization": False,
"right-left": False,
"transformer-ffn-depth": 2,
"transformer-aan-depth": 2,
"transformer-no-projection": False,
"transformer-postprocess-emb": "d",
"transformer-postprocess": "dan", # Dropout, add, normalize
"transformer-preprocess": "",
"type": "transformer",
"ulr-dim-emb": 0,
"dec-cell-base-depth": 2,
"dec-cell-high-depth": 1,
"transformer-aan-nogate": False,
}
for k, v in assumed_settings.items():
actual = marian_cfg[k]
if actual != v:
raise ValueError(f"Unexpected config value for {k} expected {v} got {actual}")
BIAS_KEY = "decoder_ff_logit_out_b"
BART_CONVERTER = { # for each encoder and decoder layer
"self_Wq": "self_attn.q_proj.weight",
"self_Wk": "self_attn.k_proj.weight",
"self_Wv": "self_attn.v_proj.weight",
"self_Wo": "self_attn.out_proj.weight",
"self_bq": "self_attn.q_proj.bias",
"self_bk": "self_attn.k_proj.bias",
"self_bv": "self_attn.v_proj.bias",
"self_bo": "self_attn.out_proj.bias",
"self_Wo_ln_scale": "self_attn_layer_norm.weight",
"self_Wo_ln_bias": "self_attn_layer_norm.bias",
"ffn_W1": "fc1.weight",
"ffn_b1": "fc1.bias",
"ffn_W2": "fc2.weight",
"ffn_b2": "fc2.bias",
"ffn_ffn_ln_scale": "final_layer_norm.weight",
"ffn_ffn_ln_bias": "final_layer_norm.bias",
# Decoder Cross Attention
"context_Wk": "encoder_attn.k_proj.weight",
"context_Wo": "encoder_attn.out_proj.weight",
"context_Wq": "encoder_attn.q_proj.weight",
"context_Wv": "encoder_attn.v_proj.weight",
"context_bk": "encoder_attn.k_proj.bias",
"context_bo": "encoder_attn.out_proj.bias",
"context_bq": "encoder_attn.q_proj.bias",
"context_bv": "encoder_attn.v_proj.bias",
"context_Wo_ln_scale": "encoder_attn_layer_norm.weight",
"context_Wo_ln_bias": "encoder_attn_layer_norm.bias",
}
class OpusState:
def __init__(self, source_dir, eos_token_id=0):
npz_path = find_model_file(source_dir)
self.state_dict = np.load(npz_path)
cfg = load_config_from_state_dict(self.state_dict)
if cfg["dim-vocabs"][0] != cfg["dim-vocabs"][1]:
raise ValueError
if "Wpos" in self.state_dict:
raise ValueError("Wpos key in state dictionary")
self.state_dict = dict(self.state_dict)
if cfg["tied-embeddings-all"]:
cfg["tied-embeddings-src"] = True
cfg["tied-embeddings"] = True
self.share_encoder_decoder_embeddings = cfg["tied-embeddings-src"]
# create the tokenizer here because we need to know the eos_token_id
self.source_dir = source_dir
self.tokenizer = self.load_tokenizer()
# retrieve EOS token and set correctly
tokenizer_has_eos_token_id = (
hasattr(self.tokenizer, "eos_token_id") and self.tokenizer.eos_token_id is not None
)
eos_token_id = self.tokenizer.eos_token_id if tokenizer_has_eos_token_id else 0
if cfg["tied-embeddings-src"]:
self.wemb, self.final_bias = add_emb_entries(self.state_dict["Wemb"], self.state_dict[BIAS_KEY], 1)
self.pad_token_id = self.wemb.shape[0] - 1
cfg["vocab_size"] = self.pad_token_id + 1
else:
self.wemb, _ = add_emb_entries(self.state_dict["encoder_Wemb"], self.state_dict[BIAS_KEY], 1)
self.dec_wemb, self.final_bias = add_emb_entries(
self.state_dict["decoder_Wemb"], self.state_dict[BIAS_KEY], 1
)
# still assuming that vocab size is same for encoder and decoder
self.pad_token_id = self.wemb.shape[0] - 1
cfg["vocab_size"] = self.pad_token_id + 1
cfg["decoder_vocab_size"] = self.pad_token_id + 1
if cfg["vocab_size"] != self.tokenizer.vocab_size:
raise ValueError(
f"Original vocab size {cfg['vocab_size']} and new vocab size {len(self.tokenizer.encoder)} mismatched."
)
# self.state_dict['Wemb'].sha
self.state_keys = list(self.state_dict.keys())
if "Wtype" in self.state_dict:
raise ValueError("Wtype key in state dictionary")
self._check_layer_entries()
self.cfg = cfg
hidden_size, intermediate_shape = self.state_dict["encoder_l1_ffn_W1"].shape
if hidden_size != cfg["dim-emb"]:
raise ValueError(f"Hidden size {hidden_size} and configured size {cfg['dim_emb']} mismatched")
# Process decoder.yml
decoder_yml = cast_marian_config(load_yaml(source_dir / "decoder.yml"))
check_marian_cfg_assumptions(cfg)
self.hf_config = MarianConfig(
vocab_size=cfg["vocab_size"],
decoder_vocab_size=cfg.get("decoder_vocab_size", cfg["vocab_size"]),
share_encoder_decoder_embeddings=cfg["tied-embeddings-src"],
decoder_layers=cfg["dec-depth"],
encoder_layers=cfg["enc-depth"],
decoder_attention_heads=cfg["transformer-heads"],
encoder_attention_heads=cfg["transformer-heads"],
decoder_ffn_dim=cfg["transformer-dim-ffn"],
encoder_ffn_dim=cfg["transformer-dim-ffn"],
d_model=cfg["dim-emb"],
activation_function=cfg["transformer-ffn-activation"],
pad_token_id=self.pad_token_id,
eos_token_id=eos_token_id,
forced_eos_token_id=eos_token_id,
bos_token_id=0,
max_position_embeddings=cfg["dim-emb"],
scale_embedding=True,
normalize_embedding="n" in cfg["transformer-preprocess"],
static_position_embeddings=not cfg["transformer-train-position-embeddings"],
tie_word_embeddings=cfg["tied-embeddings"],
dropout=0.1, # see opus-mt-train repo/transformer-dropout param.
# default: add_final_layer_norm=False,
num_beams=decoder_yml["beam-size"],
decoder_start_token_id=self.pad_token_id,
bad_words_ids=[[self.pad_token_id]],
max_length=512,
)
def _check_layer_entries(self):
self.encoder_l1 = self.sub_keys("encoder_l1")
self.decoder_l1 = self.sub_keys("decoder_l1")
self.decoder_l2 = self.sub_keys("decoder_l2")
if len(self.encoder_l1) != 16:
warnings.warn(f"Expected 16 keys for each encoder layer, got {len(self.encoder_l1)}")
if len(self.decoder_l1) != 26:
warnings.warn(f"Expected 26 keys for each decoder layer, got {len(self.decoder_l1)}")
if len(self.decoder_l2) != 26:
warnings.warn(f"Expected 26 keys for each decoder layer, got {len(self.decoder_l1)}")
@property
def extra_keys(self):
extra = []
for k in self.state_keys:
if (
k.startswith("encoder_l")
or k.startswith("decoder_l")
or k in [CONFIG_KEY, "Wemb", "encoder_Wemb", "decoder_Wemb", "Wpos", "decoder_ff_logit_out_b"]
):
continue
else:
extra.append(k)
return extra
def sub_keys(self, layer_prefix):
return [remove_prefix(k, layer_prefix) for k in self.state_dict if k.startswith(layer_prefix)]
def load_tokenizer(self):
# save tokenizer
add_special_tokens_to_vocab(self.source_dir, not self.share_encoder_decoder_embeddings)
return MarianTokenizer.from_pretrained(str(self.source_dir))
def load_marian_model(self) -> MarianMTModel:
state_dict, cfg = self.state_dict, self.hf_config
if not cfg.static_position_embeddings:
raise ValueError("config.static_position_embeddings should be True")
model = MarianMTModel(cfg)
if "hidden_size" in cfg.to_dict():
raise ValueError("hidden_size is in config")
load_layers_(
model.model.encoder.layers,
state_dict,
BART_CONVERTER,
)
load_layers_(model.model.decoder.layers, state_dict, BART_CONVERTER, is_decoder=True)
# handle tensors not associated with layers
if self.cfg["tied-embeddings-src"]:
wemb_tensor = nn.Parameter(torch.FloatTensor(self.wemb))
bias_tensor = nn.Parameter(torch.FloatTensor(self.final_bias))
model.model.shared.weight = wemb_tensor
model.model.encoder.embed_tokens = model.model.decoder.embed_tokens = model.model.shared
else:
wemb_tensor = nn.Parameter(torch.FloatTensor(self.wemb))
model.model.encoder.embed_tokens.weight = wemb_tensor
decoder_wemb_tensor = nn.Parameter(torch.FloatTensor(self.dec_wemb))
bias_tensor = nn.Parameter(torch.FloatTensor(self.final_bias))
model.model.decoder.embed_tokens.weight = decoder_wemb_tensor
# handle tied embeddings, otherwise "from_pretrained" loads them incorrectly
if self.cfg["tied-embeddings"]:
model.lm_head.weight.data = model.model.decoder.embed_tokens.weight.data.clone()
model.final_logits_bias = bias_tensor
if "Wpos" in state_dict:
print("Unexpected: got Wpos")
wpos_tensor = torch.tensor(state_dict["Wpos"])
model.model.encoder.embed_positions.weight = wpos_tensor
model.model.decoder.embed_positions.weight = wpos_tensor
if cfg.normalize_embedding:
if "encoder_emb_ln_scale_pre" not in state_dict:
raise ValueError("encoder_emb_ln_scale_pre is not in state dictionary")
raise NotImplementedError("Need to convert layernorm_embedding")
if self.extra_keys:
raise ValueError(f"Failed to convert {self.extra_keys}")
if model.get_input_embeddings().padding_idx != self.pad_token_id:
raise ValueError(
f"Padding tokens {model.get_input_embeddings().padding_idx} and {self.pad_token_id} mismatched"
)
return model
def download_and_unzip(url, dest_dir):
try:
import wget
except ImportError:
raise ImportError("you must pip install wget")
filename = wget.download(url)
unzip(filename, dest_dir)
os.remove(filename)
def convert(source_dir: Path, dest_dir):
dest_dir = Path(dest_dir)
dest_dir.mkdir(exist_ok=True)
opus_state = OpusState(source_dir)
# save tokenizer
opus_state.tokenizer.save_pretrained(dest_dir)
# save_json(opus_state.cfg, dest_dir / "marian_original_config.json")
# ^^ Uncomment to save human readable marian config for debugging
model = opus_state.load_marian_model()
model = model.half()
model.save_pretrained(dest_dir)
model.from_pretrained(dest_dir) # sanity check
def load_yaml(path):
import yaml
with open(path, encoding="utf-8") as f:
return yaml.load(f, Loader=yaml.BaseLoader)
def save_json(content: Union[Dict, List], path: str) -> None:
with open(path, "w") as f:
json.dump(content, f)
def unzip(zip_path: str, dest_dir: str) -> None:
with ZipFile(zip_path, "r") as zipObj:
zipObj.extractall(dest_dir)
if __name__ == "__main__":
"""
Tatoeba conversion instructions in scripts/tatoeba/README.md
"""
parser = argparse.ArgumentParser()
# Required parameters
parser.add_argument("--src", type=str, help="path to marian model sub dir", default="en-de")
parser.add_argument("--dest", type=str, default=None, help="Path to the output PyTorch model.")
args = parser.parse_args()
source_dir = Path(args.src)
if not source_dir.exists():
raise ValueError(f"Source directory {source_dir} not found")
dest_dir = f"converted-{source_dir.name}" if args.dest is None else args.dest
convert(source_dir, dest_dir)
|
transformers/src/transformers/models/marian/convert_marian_to_pytorch.py/0
|
{
"file_path": "transformers/src/transformers/models/marian/convert_marian_to_pytorch.py",
"repo_id": "transformers",
"token_count": 12723
}
| 366
|
# coding=utf-8
# Copyright 2022 Meta Platforms, Inc. and The HuggingFace Inc. team. 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.
"""PyTorch Mask2Former model."""
import math
import warnings
from dataclasses import dataclass
from typing import Dict, List, Optional, Tuple
import numpy as np
import torch
from torch import Tensor, nn
from ...activations import ACT2FN
from ...file_utils import (
ModelOutput,
add_start_docstrings,
add_start_docstrings_to_model_forward,
is_scipy_available,
replace_return_docstrings,
requires_backends,
)
from ...modeling_outputs import BaseModelOutput, BaseModelOutputWithCrossAttentions
from ...modeling_utils import PreTrainedModel
from ...pytorch_utils import is_torch_greater_or_equal_than_2_1
from ...utils import is_accelerate_available, logging
from ...utils.backbone_utils import load_backbone
from ...utils.import_utils import is_torchdynamo_compiling
from .configuration_mask2former import Mask2FormerConfig
if is_scipy_available():
from scipy.optimize import linear_sum_assignment
if is_accelerate_available():
from accelerate import PartialState
from accelerate.utils import reduce
logger = logging.get_logger(__name__)
_CONFIG_FOR_DOC = "Mask2FormerConfig"
_CHECKPOINT_FOR_DOC = "facebook/mask2former-swin-small-coco-instance"
_IMAGE_PROCESSOR_FOR_DOC = "Mask2FormerImageProcessor"
@dataclass
class Mask2FormerPixelDecoderOutput(ModelOutput):
"""
Mask2Former's pixel decoder module output, practically a Multi-Scale Deformable Attention based decoder. It returns
the mask features and the multiscale features.
Args:
multi_scale_features (`tuple(torch.FloatTensor)`):
Tuple of multi-scale features of scales [1/8, 1/16, 1/32] and shape `(batch_size, num_channels, height,
width)`from the Multi-Scale Deformable Attenntion based Pixel Decoder.
mask_features (`torch.FloatTensor`):
Tensor of shape `(batch_size, num_channels, height, width)`, 1/4 scale features from the last Pixel Decoder
Layer.
attentions (`tuple(torch.FloatTensor)`, *optional*):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`. Attentions weights from pixel decoder. Returned when `output_attentions=True` is passed
or when `config.output_attentions=True`
"""
multi_scale_features: Tuple[torch.FloatTensor] = None
mask_features: torch.FloatTensor = None
attentions: Optional[Tuple[torch.FloatTensor]] = None
@dataclass
class Mask2FormerMaskedAttentionDecoderOutput(BaseModelOutputWithCrossAttentions):
"""
Base class for outputs of the Transformer decoder. This class adds two attributes to
BaseModelOutputWithCrossAttentions for mask predictions logits and a tuple of intermediate decoder activations,
i.e. the output of each decoder layer, each of them gone through a layernorm.
Args:
last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
Sequence of hidden-states at the output of the last layer of the model.
hidden_states (`tuple(torch.FloatTensor)`, *optional*):
Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer) of
shape `(batch_size, sequence_length, hidden_size)`. Hidden-states of the model at the output of each layer
plus the initial embedding outputs. Returned when `output_hidden_states=True`.
attentions (`tuple(torch.FloatTensor)`, *optional*):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`. Attentions weights after the attention softmax, used to compute the weighted average in
the self-attention heads. Returned when `output_attentions=True`.
masks_queries_logits (`tuple(torch.FloatTensor)` of shape `(batch_size, num_queries, height, width)`):
Tuple of mask predictions from all layers of the transformer decoder.
intermediate_hidden_states (`tuple(torch.FloatTensor)` of shape `(num_queries, 1, hidden_size)`):
Intermediate decoder activations, i.e. the output of each decoder layer, each of them gone through a
layernorm.
"""
last_hidden_state: torch.FloatTensor = None
hidden_states: Optional[Tuple[torch.FloatTensor]] = None
attentions: Optional[torch.FloatTensor] = None
masks_queries_logits: Tuple[torch.FloatTensor] = None
intermediate_hidden_states: Tuple[torch.FloatTensor] = None
@dataclass
class Mask2FormerPixelLevelModuleOutput(ModelOutput):
"""
Mask2Former's pixel level module output. It returns the output of the encoder (optional) and all hidden states
(multi-scale features) from the `decoder`. By default, the `encoder` is a Swin Backbone and the `decoder` is a
Multi-Scale Deformable Attention based decoder.
The `decoder_last_hidden_state` are the **per-pixel embeddings** while `decoder_hidden_states` refer to multi-scale
feature maps produced using **multi-scaling strategy** defined in the paper.
Args:
encoder_last_hidden_state (`torch.FloatTensor`):
Last hidden states (final feature map of shape `(batch_size, num_channels, height, width)`) of the last
stage of the encoder.
encoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*):
Tuple of `torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`. Hidden states (also
called feature maps) of the model at the output of each stage. Returned if output_hidden_states is set to
True.
decoder_last_hidden_state (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)):
1/4 scale features from the last Pixel Decoder Layer.
decoder_hidden_states (`tuple(torch.FloatTensor)`):
Tuple of `torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`. Hidden states (also
called feature maps) of the model at the output of each stage.
"""
encoder_last_hidden_state: torch.FloatTensor = None
encoder_hidden_states: Optional[Tuple[torch.FloatTensor]] = None
decoder_last_hidden_state: torch.FloatTensor = None
decoder_hidden_states: Tuple[torch.FloatTensor] = None
@dataclass
class Mask2FormerModelOutput(ModelOutput):
"""
Class for outputs of [`Mask2FormerModel`]. This class returns all the needed hidden states to compute the logits.
Args:
encoder_last_hidden_state (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`, *optional*):
Last hidden states (final feature map) of the last stage of the encoder model (backbone). Returned when
`output_hidden_states=True` is passed.
encoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*):
Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each stage) of
shape `(batch_size, num_channels, height, width)`. Hidden-states (also called feature maps) of the encoder
model at the output of each stage. Returned when `output_hidden_states=True` is passed.
pixel_decoder_last_hidden_state (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`, *optional*):
Last hidden states (final feature map) of the last stage of the pixel decoder model.
pixel_decoder_hidden_states (`tuple(torch.FloatTensor)`, , *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each stage) of
shape `(batch_size, num_channels, height, width)`. Hidden-states (also called feature maps) of the pixel
decoder model at the output of each stage. Returned when `output_hidden_states=True` is passed.
transformer_decoder_last_hidden_state (`tuple(torch.FloatTensor)`):
Final output of the transformer decoder `(batch_size, sequence_length, hidden_size)`.
transformer_decoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*):
Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each stage) of
shape `(batch_size, sequence_length, hidden_size)`. Hidden-states (also called feature maps) of the
transformer decoder at the output of each stage. Returned when `output_hidden_states=True` is passed.
transformer_decoder_intermediate_states (`tuple(torch.FloatTensor)` of shape `(num_queries, 1, hidden_size)`):
Intermediate decoder activations, i.e. the output of each decoder layer, each of them gone through a
layernorm.
masks_queries_logits (`tuple(torch.FloatTensor)` of shape `(batch_size, num_queries, height, width)`)
Mask Predictions from each layer in the transformer decoder.
attentions (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `output_attentions=True` is passed):
Tuple of `tuple(torch.FloatTensor)` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`. Self attentions weights from transformer decoder.
"""
encoder_last_hidden_state: torch.FloatTensor = None
pixel_decoder_last_hidden_state: torch.FloatTensor = None
transformer_decoder_last_hidden_state: torch.FloatTensor = None
encoder_hidden_states: Optional[Tuple[torch.FloatTensor]] = None
pixel_decoder_hidden_states: Optional[Tuple[torch.FloatTensor]] = None
transformer_decoder_hidden_states: Optional[Tuple[torch.FloatTensor]] = None
transformer_decoder_intermediate_states: Tuple[torch.FloatTensor] = None
masks_queries_logits: Tuple[torch.FloatTensor] = None
attentions: Optional[Tuple[torch.FloatTensor]] = None
@dataclass
class Mask2FormerForUniversalSegmentationOutput(ModelOutput):
"""
Class for outputs of [`Mask2FormerForUniversalSegmentationOutput`].
This output can be directly passed to [`~Mask2FormerImageProcessor.post_process_semantic_segmentation`] or
[`~Mask2FormerImageProcessor.post_process_instance_segmentation`] or
[`~Mask2FormerImageProcessor.post_process_panoptic_segmentation`] to compute final segmentation maps. Please, see
[`~Mask2FormerImageProcessor] for details regarding usage.
Args:
loss (`torch.Tensor`, *optional*):
The computed loss, returned when labels are present.
class_queries_logits (`torch.FloatTensor`):
A tensor of shape `(batch_size, num_queries, num_labels + 1)` representing the proposed classes for each
query. Note the `+ 1` is needed because we incorporate the null class.
masks_queries_logits (`torch.FloatTensor`):
A tensor of shape `(batch_size, num_queries, height, width)` representing the proposed masks for each
query.
auxiliary_logits (`List[Dict(str, torch.FloatTensor)]`, *optional*):
List of class and mask predictions from each layer of the transformer decoder.
encoder_last_hidden_state (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
Last hidden states (final feature map) of the last stage of the encoder model (backbone).
encoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each stage) of
shape `(batch_size, num_channels, height, width)`. Hidden-states (also called feature maps) of the encoder
model at the output of each stage.
pixel_decoder_last_hidden_state (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
Last hidden states (final feature map) of the last stage of the pixel decoder model.
pixel_decoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each stage) of
shape `(batch_size, num_channels, height, width)`. Hidden-states (also called feature maps) of the pixel
decoder model at the output of each stage.
transformer_decoder_last_hidden_state (`tuple(torch.FloatTensor)`):
Final output of the transformer decoder `(batch_size, sequence_length, hidden_size)`.
transformer_decoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each stage) of
shape `(batch_size, sequence_length, hidden_size)`. Hidden-states (also called feature maps) of the
transformer decoder at the output of each stage.
attentions (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `tuple(torch.FloatTensor)` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`. Self and Cross Attentions weights from transformer decoder.
"""
loss: Optional[torch.FloatTensor] = None
class_queries_logits: torch.FloatTensor = None
masks_queries_logits: torch.FloatTensor = None
auxiliary_logits: Optional[List[Dict[str, torch.FloatTensor]]] = None
encoder_last_hidden_state: torch.FloatTensor = None
pixel_decoder_last_hidden_state: torch.FloatTensor = None
transformer_decoder_last_hidden_state: torch.FloatTensor = None
encoder_hidden_states: Optional[Tuple[torch.FloatTensor]] = None
pixel_decoder_hidden_states: Optional[Tuple[torch.FloatTensor]] = None
transformer_decoder_hidden_states: Optional[torch.FloatTensor] = None
attentions: Optional[Tuple[torch.FloatTensor]] = None
# Adapted from https://github.com/facebookresearch/detectron2/blob/main/projects/PointRend/point_rend/point_features.py
def sample_point(
input_features: torch.Tensor, point_coordinates: torch.Tensor, add_dim=False, **kwargs
) -> torch.Tensor:
"""
A wrapper around `torch.nn.functional.grid_sample` to support 3D point_coordinates tensors.
Args:
input_features (`torch.Tensor` of shape (batch_size, channels, height, width)):
A tensor that contains features map on a height * width grid
point_coordinates (`torch.Tensor` of shape (batch_size, num_points, 2) or (batch_size, grid_height, grid_width,:
2)):
A tensor that contains [0, 1] * [0, 1] normalized point coordinates
add_dim (`bool`):
boolean value to keep track of added dimension
Returns:
point_features (`torch.Tensor` of shape (batch_size, channels, num_points) or (batch_size, channels,
height_grid, width_grid):
A tensor that contains features for points in `point_coordinates`.
"""
if point_coordinates.dim() == 3:
add_dim = True
point_coordinates = point_coordinates.unsqueeze(2)
# use nn.function.grid_sample to get features for points in `point_coordinates` via bilinear interpolation
point_features = torch.nn.functional.grid_sample(input_features, 2.0 * point_coordinates - 1.0, **kwargs)
if add_dim:
point_features = point_features.squeeze(3)
return point_features
# Copied from transformers.models.maskformer.modeling_maskformer.dice_loss
def dice_loss(inputs: Tensor, labels: Tensor, num_masks: int) -> Tensor:
r"""
Compute the DICE loss, similar to generalized IOU for masks as follows:
$$ \mathcal{L}_{\text{dice}(x, y) = 1 - \frac{2 * x \cap y }{x \cup y + 1}} $$
In practice, since `labels` is a binary mask, (only 0s and 1s), dice can be computed as follow
$$ \mathcal{L}_{\text{dice}(x, y) = 1 - \frac{2 * x * y }{x + y + 1}} $$
Args:
inputs (`torch.Tensor`):
A tensor representing a mask.
labels (`torch.Tensor`):
A tensor with the same shape as inputs. Stores the binary classification labels for each element in inputs
(0 for the negative class and 1 for the positive class).
num_masks (`int`):
The number of masks present in the current batch, used for normalization.
Returns:
`torch.Tensor`: The computed loss.
"""
probs = inputs.sigmoid().flatten(1)
numerator = 2 * (probs * labels).sum(-1)
denominator = probs.sum(-1) + labels.sum(-1)
loss = 1 - (numerator + 1) / (denominator + 1)
loss = loss.sum() / num_masks
return loss
def sigmoid_cross_entropy_loss(inputs: torch.Tensor, labels: torch.Tensor, num_masks: int) -> torch.Tensor:
r"""
Args:
inputs (`torch.Tensor`):
A float tensor of arbitrary shape.
labels (`torch.Tensor`):
A tensor with the same shape as inputs. Stores the binary classification labels for each element in inputs
(0 for the negative class and 1 for the positive class).
Returns:
loss (`torch.Tensor`): The computed loss.
"""
criterion = nn.BCEWithLogitsLoss(reduction="none")
cross_entropy_loss = criterion(inputs, labels)
loss = cross_entropy_loss.mean(1).sum() / num_masks
return loss
# Copied from transformers.models.maskformer.modeling_maskformer.pair_wise_dice_loss
def pair_wise_dice_loss(inputs: Tensor, labels: Tensor) -> Tensor:
"""
A pair wise version of the dice loss, see `dice_loss` for usage.
Args:
inputs (`torch.Tensor`):
A tensor representing a mask
labels (`torch.Tensor`):
A tensor with the same shape as inputs. Stores the binary classification labels for each element in inputs
(0 for the negative class and 1 for the positive class).
Returns:
`torch.Tensor`: The computed loss between each pairs.
"""
inputs = inputs.sigmoid().flatten(1)
numerator = 2 * torch.matmul(inputs, labels.T)
# using broadcasting to get a [num_queries, NUM_CLASSES] matrix
denominator = inputs.sum(-1)[:, None] + labels.sum(-1)[None, :]
loss = 1 - (numerator + 1) / (denominator + 1)
return loss
def pair_wise_sigmoid_cross_entropy_loss(inputs: torch.Tensor, labels: torch.Tensor) -> torch.Tensor:
r"""
A pair wise version of the cross entropy loss, see `sigmoid_cross_entropy_loss` for usage.
Args:
inputs (`torch.Tensor`):
A tensor representing a mask.
labels (`torch.Tensor`):
A tensor with the same shape as inputs. Stores the binary classification labels for each element in inputs
(0 for the negative class and 1 for the positive class).
Returns:
loss (`torch.Tensor`): The computed loss between each pairs.
"""
height_and_width = inputs.shape[1]
criterion = nn.BCEWithLogitsLoss(reduction="none")
cross_entropy_loss_pos = criterion(inputs, torch.ones_like(inputs))
cross_entropy_loss_neg = criterion(inputs, torch.zeros_like(inputs))
loss_pos = torch.matmul(cross_entropy_loss_pos / height_and_width, labels.T)
loss_neg = torch.matmul(cross_entropy_loss_neg / height_and_width, (1 - labels).T)
loss = loss_pos + loss_neg
return loss
# Adapted from https://github.com/facebookresearch/Mask2Former/blob/main/mask2former/modeling/matcher.py
class Mask2FormerHungarianMatcher(nn.Module):
"""This class computes an assignment between the labels and the predictions of the network.
For efficiency reasons, the labels don't include the no_object. Because of this, in general, there are more
predictions than labels. In this case, we do a 1-to-1 matching of the best predictions, while the others are
un-matched (and thus treated as non-objects).
"""
def __init__(
self, cost_class: float = 1.0, cost_mask: float = 1.0, cost_dice: float = 1.0, num_points: int = 12544
):
"""Creates the matcher
Params:
cost_class (`float`, *optional*, defaults to 1.0):
Relative weight of the classification error in the matching cost.
cost_mask (`float`, *optional*, defaults to 1.0):
This is the relative weight of the focal loss of the binary mask in the matching cost.
cost_dice (`float`, *optional*, defaults to 1.0):
This is the relative weight of the dice loss of the binary mask in the matching cost.
num_points (`int`, *optional*, defaults to 12544):
No. of points to sample on which the mask loss will be calculated. The same set of K points are
uniformly sampled for all prediction and ground truth masks to construct the cost matrix for bipartite
matching.
"""
super().__init__()
if cost_class == 0 and cost_mask == 0 and cost_dice == 0:
raise ValueError("All costs cant be 0")
self.num_points = num_points
self.cost_class = cost_class
self.cost_mask = cost_mask
self.cost_dice = cost_dice
@torch.no_grad()
def forward(
self,
masks_queries_logits: torch.Tensor,
class_queries_logits: torch.Tensor,
mask_labels: torch.Tensor,
class_labels: torch.Tensor,
) -> List[Tuple[Tensor]]:
"""
Params:
masks_queries_logits (`torch.Tensor`):
A tensor of dim `batch_size, num_queries, num_labels` with the classification logits.
class_queries_logits (`torch.Tensor`):
A tensor of dim `batch_size, num_queries, height, width` with the predicted masks.
class_labels (`torch.Tensor`):
A tensor of dim `num_target_boxes` (where num_target_boxes is the number of ground-truth objects in the
target) containing the class labels.
mask_labels (`torch.Tensor`):
A tensor of dim `num_target_boxes, height, width` containing the target masks.
Returns:
matched_indices (`List[Tuple[Tensor]]`): A list of size batch_size, containing tuples of (index_i, index_j)
where:
- index_i is the indices of the selected predictions (in order)
- index_j is the indices of the corresponding selected labels (in order)
For each batch element, it holds:
len(index_i) = len(index_j) = min(num_queries, num_target_boxes).
"""
indices: List[Tuple[np.array]] = []
# iterate through batch size
batch_size = masks_queries_logits.shape[0]
for i in range(batch_size):
pred_probs = class_queries_logits[i].softmax(-1)
pred_mask = masks_queries_logits[i]
# Compute the classification cost. Contrary to the loss, we don't use the NLL, but approximate it in 1 - proba[target class]. The 1 is a constant that doesn't change the matching, it can be ommitted.
cost_class = -pred_probs[:, class_labels[i]]
target_mask = mask_labels[i].to(pred_mask)
target_mask = target_mask[:, None]
pred_mask = pred_mask[:, None]
# Sample ground truth and predicted masks
point_coordinates = torch.rand(1, self.num_points, 2, device=pred_mask.device)
target_coordinates = point_coordinates.repeat(target_mask.shape[0], 1, 1)
target_mask = sample_point(target_mask, target_coordinates, align_corners=False).squeeze(1)
pred_coordinates = point_coordinates.repeat(pred_mask.shape[0], 1, 1)
pred_mask = sample_point(pred_mask, pred_coordinates, align_corners=False).squeeze(1)
# compute the cross entropy loss between each mask pairs -> shape (num_queries, num_labels)
cost_mask = pair_wise_sigmoid_cross_entropy_loss(pred_mask, target_mask)
# Compute the dice loss betwen each mask pairs -> shape (num_queries, num_labels)
cost_dice = pair_wise_dice_loss(pred_mask, target_mask)
# final cost matrix
cost_matrix = self.cost_mask * cost_mask + self.cost_class * cost_class + self.cost_dice * cost_dice
# eliminate infinite values in cost_matrix to avoid the error ``ValueError: cost matrix is infeasible``
cost_matrix = torch.minimum(cost_matrix, torch.tensor(1e10))
cost_matrix = torch.maximum(cost_matrix, torch.tensor(-1e10))
# do the assigmented using the hungarian algorithm in scipy
assigned_indices: Tuple[np.array] = linear_sum_assignment(cost_matrix.cpu())
indices.append(assigned_indices)
# It could be stacked in one tensor
matched_indices = [
(torch.as_tensor(i, dtype=torch.int64), torch.as_tensor(j, dtype=torch.int64)) for i, j in indices
]
return matched_indices
# Adapted from https://github.com/facebookresearch/Mask2Former/blob/main/mask2former/modeling/criterion.py
class Mask2FormerLoss(nn.Module):
def __init__(self, config: Mask2FormerConfig, weight_dict: Dict[str, float]):
"""
The Mask2Former Loss. The loss is computed very similar to DETR. The process happens in two steps: 1) we
compute hungarian assignment between ground truth masks and the outputs of the model 2) we supervise each pair
of matched ground-truth / prediction (supervise class and mask)
Args:
config (`Mask2FormerConfig`):
The configuration for Mask2Former model also containing loss calculation specific parameters.
weight_dict (`Dict[str, float]`):
A dictionary of weights to be applied to the different losses.
"""
super().__init__()
requires_backends(self, ["scipy"])
self.num_labels = config.num_labels
self.weight_dict = weight_dict
# Weight to apply to the null class
self.eos_coef = config.no_object_weight
empty_weight = torch.ones(self.num_labels + 1)
empty_weight[-1] = self.eos_coef
self.register_buffer("empty_weight", empty_weight)
# pointwise mask loss parameters
self.num_points = config.train_num_points
self.oversample_ratio = config.oversample_ratio
self.importance_sample_ratio = config.importance_sample_ratio
self.matcher = Mask2FormerHungarianMatcher(
cost_class=1.0,
cost_dice=config.dice_weight,
cost_mask=config.mask_weight,
num_points=self.num_points,
)
def _max_by_axis(self, sizes: List[List[int]]) -> List[int]:
maxes = sizes[0]
for sublist in sizes[1:]:
for index, item in enumerate(sublist):
maxes[index] = max(maxes[index], item)
return maxes
# Adapted from nested_tensor_from_tensor_list() in original implementation
def _pad_images_to_max_in_batch(self, tensors: List[Tensor]) -> Tuple[Tensor, Tensor]:
# get the maximum size in the batch
max_size = self._max_by_axis([list(tensor.shape) for tensor in tensors])
# compute final size
batch_shape = [len(tensors)] + max_size
batch_size, _, height, width = batch_shape
dtype = tensors[0].dtype
device = tensors[0].device
padded_tensors = torch.zeros(batch_shape, dtype=dtype, device=device)
padding_masks = torch.ones((batch_size, height, width), dtype=torch.bool, device=device)
# pad the tensors to the size of the biggest one
for tensor, padded_tensor, padding_mask in zip(tensors, padded_tensors, padding_masks):
padded_tensor[: tensor.shape[0], : tensor.shape[1], : tensor.shape[2]].copy_(tensor)
padding_mask[: tensor.shape[1], : tensor.shape[2]] = False
return padded_tensors, padding_masks
def loss_labels(
self, class_queries_logits: Tensor, class_labels: List[Tensor], indices: Tuple[np.array]
) -> Dict[str, Tensor]:
"""Compute the losses related to the labels using cross entropy.
Args:
class_queries_logits (`torch.Tensor`):
A tensor of shape `batch_size, num_queries, num_labels`
class_labels (`List[torch.Tensor]`):
List of class labels of shape `(labels)`.
indices (`Tuple[np.array])`:
The indices computed by the Hungarian matcher.
Returns:
`Dict[str, Tensor]`: A dict of `torch.Tensor` containing the following key:
- **loss_cross_entropy** -- The loss computed using cross entropy on the predicted and ground truth labels.
"""
pred_logits = class_queries_logits
batch_size, num_queries, _ = pred_logits.shape
criterion = nn.CrossEntropyLoss(weight=self.empty_weight)
idx = self._get_predictions_permutation_indices(indices) # shape of (batch_size, num_queries)
target_classes_o = torch.cat(
[target[j] for target, (_, j) in zip(class_labels, indices)]
) # shape of (batch_size, num_queries)
target_classes = torch.full(
(batch_size, num_queries), fill_value=self.num_labels, dtype=torch.int64, device=pred_logits.device
)
target_classes[idx] = target_classes_o
# Permute target_classes (batch_size, num_queries, num_labels) -> (batch_size, num_labels, num_queries)
pred_logits_transposed = pred_logits.transpose(1, 2)
loss_ce = criterion(pred_logits_transposed, target_classes)
losses = {"loss_cross_entropy": loss_ce}
return losses
def loss_masks(
self,
masks_queries_logits: torch.Tensor,
mask_labels: List[torch.Tensor],
indices: Tuple[np.array],
num_masks: int,
) -> Dict[str, torch.Tensor]:
"""Compute the losses related to the masks using sigmoid_cross_entropy_loss and dice loss.
Args:
masks_queries_logits (`torch.Tensor`):
A tensor of shape `(batch_size, num_queries, height, width)`.
mask_labels (`torch.Tensor`):
List of mask labels of shape `(labels, height, width)`.
indices (`Tuple[np.array])`:
The indices computed by the Hungarian matcher.
num_masks (`int)`:
The number of masks, used for normalization.
Returns:
losses (`Dict[str, Tensor]`): A dict of `torch.Tensor` containing two keys:
- **loss_mask** -- The loss computed using sigmoid cross entropy loss on the predicted and ground truth.
masks.
- **loss_dice** -- The loss computed using dice loss on the predicted on the predicted and ground truth,
masks.
"""
src_idx = self._get_predictions_permutation_indices(indices)
tgt_idx = self._get_targets_permutation_indices(indices)
# shape (batch_size * num_queries, height, width)
pred_masks = masks_queries_logits[src_idx]
# shape (batch_size, num_queries, height, width)
# pad all and stack the targets to the num_labels dimension
target_masks, _ = self._pad_images_to_max_in_batch(mask_labels)
target_masks = target_masks[tgt_idx]
# No need to upsample predictions as we are using normalized coordinates
pred_masks = pred_masks[:, None]
target_masks = target_masks[:, None]
# Sample point coordinates
with torch.no_grad():
point_coordinates = self.sample_points_using_uncertainty(
pred_masks,
lambda logits: self.calculate_uncertainty(logits),
self.num_points,
self.oversample_ratio,
self.importance_sample_ratio,
)
point_labels = sample_point(target_masks, point_coordinates, align_corners=False).squeeze(1)
point_logits = sample_point(pred_masks, point_coordinates, align_corners=False).squeeze(1)
losses = {
"loss_mask": sigmoid_cross_entropy_loss(point_logits, point_labels, num_masks),
"loss_dice": dice_loss(point_logits, point_labels, num_masks),
}
del pred_masks
del target_masks
return losses
def _get_predictions_permutation_indices(self, indices):
# Permute predictions following indices
batch_indices = torch.cat([torch.full_like(src, i) for i, (src, _) in enumerate(indices)])
predictions_indices = torch.cat([src for (src, _) in indices])
return batch_indices, predictions_indices
def _get_targets_permutation_indices(self, indices):
# Permute labels following indices
batch_indices = torch.cat([torch.full_like(tgt, i) for i, (_, tgt) in enumerate(indices)])
target_indices = torch.cat([tgt for (_, tgt) in indices])
return batch_indices, target_indices
def calculate_uncertainty(self, logits: torch.Tensor) -> torch.Tensor:
"""
In Mask2Former paper, uncertainty is estimated as L1 distance between 0.0 and the logit prediction in 'logits'
for the foreground class in `classes`.
Args:
logits (`torch.Tensor`):
A tensor of shape (R, 1, ...) for class-specific or class-agnostic, where R is the total number of predicted masks in all images and C is:
the number of foreground classes. The values are logits.
Returns:
scores (`torch.Tensor`): A tensor of shape (R, 1, ...) that contains uncertainty scores with the most
uncertain locations having the highest uncertainty score.
"""
uncertainty_scores = -(torch.abs(logits))
return uncertainty_scores
def sample_points_using_uncertainty(
self,
logits: torch.Tensor,
uncertainty_function,
num_points: int,
oversample_ratio: int,
importance_sample_ratio: float,
) -> torch.Tensor:
"""
This function is meant for sampling points in [0, 1] * [0, 1] coordinate space based on their uncertainty. The
uncertainty is calculated for each point using the passed `uncertainty function` that takes points logit
prediction as input.
Args:
logits (`float`):
Logit predictions for P points.
uncertainty_function:
A function that takes logit predictions for P points and returns their uncertainties.
num_points (`int`):
The number of points P to sample.
oversample_ratio (`int`):
Oversampling parameter.
importance_sample_ratio (`float`):
Ratio of points that are sampled via importance sampling.
Returns:
point_coordinates (`torch.Tensor`):
Coordinates for P sampled points.
"""
num_boxes = logits.shape[0]
num_points_sampled = int(num_points * oversample_ratio)
# Get random point coordinates
point_coordinates = torch.rand(num_boxes, num_points_sampled, 2, device=logits.device)
# Get sampled prediction value for the point coordinates
point_logits = sample_point(logits, point_coordinates, align_corners=False)
# Calculate the uncertainties based on the sampled prediction values of the points
point_uncertainties = uncertainty_function(point_logits)
num_uncertain_points = int(importance_sample_ratio * num_points)
num_random_points = num_points - num_uncertain_points
idx = torch.topk(point_uncertainties[:, 0, :], k=num_uncertain_points, dim=1)[1]
shift = num_points_sampled * torch.arange(num_boxes, dtype=torch.long, device=logits.device)
idx += shift[:, None]
point_coordinates = point_coordinates.view(-1, 2)[idx.view(-1), :].view(num_boxes, num_uncertain_points, 2)
if num_random_points > 0:
point_coordinates = torch.cat(
[point_coordinates, torch.rand(num_boxes, num_random_points, 2, device=logits.device)],
dim=1,
)
return point_coordinates
def forward(
self,
masks_queries_logits: torch.Tensor,
class_queries_logits: torch.Tensor,
mask_labels: List[torch.Tensor],
class_labels: List[torch.Tensor],
auxiliary_predictions: Optional[Dict[str, torch.Tensor]] = None,
) -> Dict[str, torch.Tensor]:
"""
This performs the loss computation.
Args:
masks_queries_logits (`torch.Tensor`):
A tensor of shape `(batch_size, num_queries, height, width)`.
class_queries_logits (`torch.Tensor`):
A tensor of shape `(batch_size, num_queries, num_labels)`.
mask_labels (`torch.Tensor`):
List of mask labels of shape `(labels, height, width)`.
class_labels (`List[torch.Tensor]`):
List of class labels of shape `(labels)`.
auxiliary_predictions (`Dict[str, torch.Tensor]`, *optional*):
if `use_auxiliary_loss` was set to `true` in [`Mask2FormerConfig`], then it contains the logits from
the inner layers of the Mask2FormerMaskedAttentionDecoder.
Returns:
losses (`Dict[str, Tensor]`): A dict of `torch.Tensor` containing three keys:
- **loss_cross_entropy** -- The loss computed using cross entropy on the predicted and ground truth labels.
- **loss_mask** -- The loss computed using sigmoid cross_entropy loss on the predicted and ground truth
masks.
- **loss_dice** -- The loss computed using dice loss on the predicted on the predicted and ground truth
masks.
if `use_auxiliary_loss` was set to `true` in [`Mask2FormerConfig`], the dictionary contains additional
losses for each auxiliary predictions.
"""
# retrieve the matching between the outputs of the last layer and the labels
indices = self.matcher(masks_queries_logits, class_queries_logits, mask_labels, class_labels)
# compute the average number of target masks for normalization purposes
num_masks = self.get_num_masks(class_labels, device=class_labels[0].device)
# get all the losses
losses: Dict[str, Tensor] = {
**self.loss_masks(masks_queries_logits, mask_labels, indices, num_masks),
**self.loss_labels(class_queries_logits, class_labels, indices),
}
# in case of auxiliary losses, we repeat this process with the output of each intermediate layer.
if auxiliary_predictions is not None:
for idx, aux_outputs in enumerate(auxiliary_predictions):
masks_queries_logits = aux_outputs["masks_queries_logits"]
class_queries_logits = aux_outputs["class_queries_logits"]
loss_dict = self.forward(masks_queries_logits, class_queries_logits, mask_labels, class_labels)
loss_dict = {f"{key}_{idx}": value for key, value in loss_dict.items()}
losses.update(loss_dict)
return losses
def get_num_masks(self, class_labels: torch.Tensor, device: torch.device) -> torch.Tensor:
"""
Computes the average number of target masks across the batch, for normalization purposes.
"""
num_masks = sum([len(classes) for classes in class_labels])
num_masks = torch.as_tensor(num_masks, dtype=torch.float, device=device)
world_size = 1
if is_accelerate_available():
if PartialState._shared_state != {}:
num_masks = reduce(num_masks)
world_size = PartialState().num_processes
num_masks = torch.clamp(num_masks / world_size, min=1)
return num_masks
# Copied from transformers.models.deformable_detr.modeling_deformable_detr.multi_scale_deformable_attention
def multi_scale_deformable_attention(
value: Tensor, value_spatial_shapes: Tensor, sampling_locations: Tensor, attention_weights: Tensor
) -> Tensor:
batch_size, _, num_heads, hidden_dim = value.shape
_, num_queries, num_heads, num_levels, num_points, _ = sampling_locations.shape
value_list = value.split([height.item() * width.item() for height, width in value_spatial_shapes], dim=1)
sampling_grids = 2 * sampling_locations - 1
sampling_value_list = []
for level_id, (height, width) in enumerate(value_spatial_shapes):
# batch_size, height*width, num_heads, hidden_dim
# -> batch_size, height*width, num_heads*hidden_dim
# -> batch_size, num_heads*hidden_dim, height*width
# -> batch_size*num_heads, hidden_dim, height, width
value_l_ = (
value_list[level_id].flatten(2).transpose(1, 2).reshape(batch_size * num_heads, hidden_dim, height, width)
)
# batch_size, num_queries, num_heads, num_points, 2
# -> batch_size, num_heads, num_queries, num_points, 2
# -> batch_size*num_heads, num_queries, num_points, 2
sampling_grid_l_ = sampling_grids[:, :, :, level_id].transpose(1, 2).flatten(0, 1)
# batch_size*num_heads, hidden_dim, num_queries, num_points
sampling_value_l_ = nn.functional.grid_sample(
value_l_, sampling_grid_l_, mode="bilinear", padding_mode="zeros", align_corners=False
)
sampling_value_list.append(sampling_value_l_)
# (batch_size, num_queries, num_heads, num_levels, num_points)
# -> (batch_size, num_heads, num_queries, num_levels, num_points)
# -> (batch_size, num_heads, 1, num_queries, num_levels*num_points)
attention_weights = attention_weights.transpose(1, 2).reshape(
batch_size * num_heads, 1, num_queries, num_levels * num_points
)
output = (
(torch.stack(sampling_value_list, dim=-2).flatten(-2) * attention_weights)
.sum(-1)
.view(batch_size, num_heads * hidden_dim, num_queries)
)
return output.transpose(1, 2).contiguous()
# Copied from transformers.models.maskformer.modeling_maskformer.MaskFormerSinePositionEmbedding with MaskFormer->Mask2Former
class Mask2FormerSinePositionEmbedding(nn.Module):
"""
This is a more standard version of the position embedding, very similar to the one used by the Attention is all you
need paper, generalized to work on images.
"""
def __init__(
self, num_pos_feats: int = 64, temperature: int = 10000, normalize: bool = False, scale: Optional[float] = None
):
super().__init__()
if scale is not None and normalize is False:
raise ValueError("normalize should be True if scale is passed")
self.num_pos_feats = num_pos_feats
self.temperature = temperature
self.normalize = normalize
self.scale = 2 * math.pi if scale is None else scale
def forward(self, x: Tensor, mask: Optional[Tensor] = None) -> Tensor:
if mask is None:
mask = torch.zeros((x.size(0), x.size(2), x.size(3)), device=x.device, dtype=torch.bool)
not_mask = (~mask).to(x.dtype)
y_embed = not_mask.cumsum(1)
x_embed = not_mask.cumsum(2)
if self.normalize:
eps = 1e-6
y_embed = y_embed / (y_embed[:, -1:, :] + eps) * self.scale
x_embed = x_embed / (x_embed[:, :, -1:] + eps) * self.scale
dim_t = torch.arange(self.num_pos_feats, dtype=torch.int64, device=x.device).type_as(x)
dim_t = self.temperature ** (2 * torch.div(dim_t, 2, rounding_mode="floor") / self.num_pos_feats)
pos_x = x_embed[:, :, :, None] / dim_t
pos_y = y_embed[:, :, :, None] / dim_t
pos_x = torch.stack((pos_x[:, :, :, 0::2].sin(), pos_x[:, :, :, 1::2].cos()), dim=4).flatten(3)
pos_y = torch.stack((pos_y[:, :, :, 0::2].sin(), pos_y[:, :, :, 1::2].cos()), dim=4).flatten(3)
pos = torch.cat((pos_y, pos_x), dim=3).permute(0, 3, 1, 2)
return pos
# Modified from transformers.models.detr.modeling_deformable_detr.DeformableDetrMultiscaleDeformableAttention
class Mask2FormerPixelDecoderEncoderMultiscaleDeformableAttention(nn.Module):
"""
Multiscale deformable attention as proposed in Deformable DETR.
"""
def __init__(self, embed_dim: int, num_heads: int, n_levels: int, n_points: int):
super().__init__()
if embed_dim % num_heads != 0:
raise ValueError(
f"embed_dim (d_model) must be divisible by num_heads, but got {embed_dim} and {num_heads}"
)
dim_per_head = embed_dim // num_heads
# check if dim_per_head is power of 2
if not ((dim_per_head & (dim_per_head - 1) == 0) and dim_per_head != 0):
warnings.warn(
"You'd better set embed_dim (d_model) in DeformableDetrMultiscaleDeformableAttention to make the"
" dimension of each attention head a power of 2 which is more efficient in the authors' CUDA"
" implementation."
)
self.im2col_step = 128
self.d_model = embed_dim
self.n_levels = n_levels
self.n_heads = num_heads
self.n_points = n_points
self.sampling_offsets = nn.Linear(embed_dim, num_heads * n_levels * n_points * 2)
self.attention_weights = nn.Linear(embed_dim, num_heads * n_levels * n_points)
self.value_proj = nn.Linear(embed_dim, embed_dim)
self.output_proj = nn.Linear(embed_dim, embed_dim)
def with_pos_embed(self, tensor: torch.Tensor, position_embeddings: Optional[Tensor]):
return tensor if position_embeddings is None else tensor + position_embeddings
def forward(
self,
hidden_states: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
encoder_hidden_states=None,
encoder_attention_mask=None,
position_embeddings: Optional[torch.Tensor] = None,
reference_points=None,
spatial_shapes=None,
level_start_index=None,
output_attentions: bool = False,
):
# add position embeddings to the hidden states before projecting to queries and keys
if position_embeddings is not None:
hidden_states = self.with_pos_embed(hidden_states, position_embeddings)
batch_size, num_queries, _ = hidden_states.shape
batch_size, sequence_length, _ = encoder_hidden_states.shape
if (spatial_shapes[:, 0] * spatial_shapes[:, 1]).sum() != sequence_length:
raise ValueError(
"Make sure to align the spatial shapes with the sequence length of the encoder hidden states"
)
value = self.value_proj(encoder_hidden_states)
if attention_mask is not None:
# we invert the attention_mask
value = value.masked_fill(attention_mask[..., None], float(0))
value = value.view(batch_size, sequence_length, self.n_heads, self.d_model // self.n_heads)
sampling_offsets = self.sampling_offsets(hidden_states).view(
batch_size, num_queries, self.n_heads, self.n_levels, self.n_points, 2
)
attention_weights = self.attention_weights(hidden_states).view(
batch_size, num_queries, self.n_heads, self.n_levels * self.n_points
)
attention_weights = nn.functional.softmax(attention_weights, -1).view(
batch_size, num_queries, self.n_heads, self.n_levels, self.n_points
)
# batch_size, num_queries, n_heads, n_levels, n_points, 2
if reference_points.shape[-1] == 2:
offset_normalizer = torch.stack([spatial_shapes[..., 1], spatial_shapes[..., 0]], -1)
sampling_locations = (
reference_points[:, :, None, :, None, :]
+ sampling_offsets / offset_normalizer[None, None, None, :, None, :]
)
elif reference_points.shape[-1] == 4:
sampling_locations = (
reference_points[:, :, None, :, None, :2]
+ sampling_offsets / self.n_points * reference_points[:, :, None, :, None, 2:] * 0.5
)
else:
raise ValueError(f"Last dim of reference_points must be 2 or 4, but got {reference_points.shape[-1]}")
output = multi_scale_deformable_attention(value, spatial_shapes, sampling_locations, attention_weights)
output = self.output_proj(output)
return output, attention_weights
class Mask2FormerPixelDecoderEncoderLayer(nn.Module):
def __init__(self, config: Mask2FormerConfig):
super().__init__()
self.embed_dim = config.feature_size
self.self_attn = Mask2FormerPixelDecoderEncoderMultiscaleDeformableAttention(
embed_dim=self.embed_dim,
num_heads=config.num_attention_heads,
n_levels=3,
n_points=4,
)
self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
self.dropout = config.dropout
self.activation_fn = nn.functional.relu
self.activation_dropout = config.dropout
self.fc1 = nn.Linear(self.embed_dim, config.encoder_feedforward_dim)
self.fc2 = nn.Linear(config.encoder_feedforward_dim, self.embed_dim)
self.final_layer_norm = nn.LayerNorm(self.embed_dim)
def forward(
self,
hidden_states: torch.Tensor,
attention_mask: torch.Tensor,
position_embeddings: torch.Tensor = None,
reference_points=None,
spatial_shapes=None,
level_start_index=None,
output_attentions: bool = False,
):
"""
Args:
hidden_states (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
Input to the layer.
attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`):
Attention mask.
position_embeddings (`torch.FloatTensor`, *optional*):
Position embeddings, to be added to `hidden_states`.
reference_points (`torch.FloatTensor`, *optional*):
Reference points.
spatial_shapes (`torch.LongTensor`, *optional*):
Spatial shapes of the backbone feature maps.
level_start_index (`torch.LongTensor`, *optional*):
Level start index.
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
returned tensors for more detail.
"""
residual = hidden_states
# Apply Multi-scale Deformable Attention Module on the multi-scale feature maps.
hidden_states, attn_weights = self.self_attn(
hidden_states=hidden_states,
attention_mask=attention_mask,
encoder_hidden_states=hidden_states,
encoder_attention_mask=attention_mask,
position_embeddings=position_embeddings,
reference_points=reference_points,
spatial_shapes=spatial_shapes,
level_start_index=level_start_index,
output_attentions=output_attentions,
)
hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
hidden_states = residual + hidden_states
hidden_states = self.self_attn_layer_norm(hidden_states)
residual = hidden_states
hidden_states = self.activation_fn(self.fc1(hidden_states))
hidden_states = nn.functional.dropout(hidden_states, p=self.activation_dropout, training=self.training)
hidden_states = self.fc2(hidden_states)
hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
hidden_states = residual + hidden_states
hidden_states = self.final_layer_norm(hidden_states)
if self.training:
if torch.isinf(hidden_states).any() or torch.isnan(hidden_states).any():
clamp_value = torch.finfo(hidden_states.dtype).max - 1000
hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)
outputs = (hidden_states,)
if output_attentions:
outputs += (attn_weights.transpose(1, 0),)
return outputs
# Modified from from transformers.models.detr.modeling_deformable_detr.DeformableDetrEncoder with DeformableDetrEncoder->Mask2FormerPixelDecoderEncoderOnly
class Mask2FormerPixelDecoderEncoderOnly(nn.Module):
"""
Transformer encoder consisting of *config.encoder_layers* deformable attention layers. Each layer is a
[`Mask2FormerPixelDecoderEncoderLayer`]. The encoder updates the flattened multi-scale feature maps through
multiple deformable attention layers.
Args:
config: Mask2FormerConfig
"""
def __init__(self, config: Mask2FormerConfig):
super().__init__()
self.config = config
self.dropout = config.dropout
self.layers = nn.ModuleList(
[Mask2FormerPixelDecoderEncoderLayer(config) for _ in range(config.encoder_layers)]
)
@staticmethod
def get_reference_points(spatial_shapes, valid_ratios, device):
"""
Get reference points for each feature map. Used in decoder.
Args:
spatial_shapes (`torch.LongTensor`):
Spatial shapes of each feature map, has shape of `(num_feature_levels, 2)`.
valid_ratios (`torch.FloatTensor`):
Valid ratios of each feature map, has shape of `(batch_size, num_feature_levels, 2)`.
device (`torch.device`):
Device on which to create the tensors.
Returns:
`torch.FloatTensor` of shape `(batch_size, num_queries, num_feature_levels, 2)`
"""
reference_points_list = []
for lvl, (height, width) in enumerate(spatial_shapes):
ref_y, ref_x = torch.meshgrid(
torch.linspace(0.5, height - 0.5, height, dtype=valid_ratios.dtype, device=device),
torch.linspace(0.5, width - 0.5, width, dtype=valid_ratios.dtype, device=device),
indexing="ij",
)
ref_y = ref_y.reshape(-1)[None] / (valid_ratios[:, None, lvl, 1] * height)
ref_x = ref_x.reshape(-1)[None] / (valid_ratios[:, None, lvl, 0] * width)
ref = torch.stack((ref_x, ref_y), -1)
reference_points_list.append(ref)
reference_points = torch.cat(reference_points_list, 1)
reference_points = reference_points[:, :, None] * valid_ratios[:, None]
return reference_points
def forward(
self,
inputs_embeds=None,
attention_mask=None,
position_embeddings=None,
spatial_shapes=None,
level_start_index=None,
valid_ratios=None,
output_attentions=None,
output_hidden_states=None,
return_dict=None,
):
r"""
Args:
inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
Flattened feature map (output of the backbone + projection layer) that is passed to the encoder.
attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
Mask to avoid performing attention on padding pixel features. Mask values selected in `[0, 1]`:
- 1 for pixel features that are real (i.e. **not masked**),
- 0 for pixel features that are padding (i.e. **masked**).
[What are attention masks?](../glossary#attention-mask)
position_embeddings (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
Position embeddings that are added to the queries and keys in each self-attention layer.
spatial_shapes (`torch.LongTensor` of shape `(num_feature_levels, 2)`):
Spatial shapes of each feature map.
level_start_index (`torch.LongTensor` of shape `(num_feature_levels)`):
Starting index of each feature map.
valid_ratios (`torch.FloatTensor` of shape `(batch_size, num_feature_levels, 2)`):
Ratio of valid area in each feature level.
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
returned tensors for more detail.
output_hidden_states (`bool`, *optional*):
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
for more detail.
return_dict (`bool`, *optional*):
Whether or not to return a [`~file_utils.ModelOutput`] instead of a plain tuple.
"""
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
hidden_states = inputs_embeds
reference_points = self.get_reference_points(spatial_shapes, valid_ratios, device=inputs_embeds.device)
all_hidden_states = () if output_hidden_states else None
all_attentions = () if output_attentions else None
for i, encoder_layer in enumerate(self.layers):
if output_hidden_states:
all_hidden_states += (hidden_states.transpose(1, 0),)
layer_outputs = encoder_layer(
hidden_states,
attention_mask,
position_embeddings=position_embeddings,
reference_points=reference_points,
spatial_shapes=spatial_shapes,
level_start_index=level_start_index,
output_attentions=output_attentions,
)
hidden_states = layer_outputs[0]
if output_attentions:
all_attentions = all_attentions + (layer_outputs[1],)
if output_hidden_states:
all_hidden_states += (hidden_states.transpose(1, 0),)
return BaseModelOutput(
last_hidden_state=hidden_states, hidden_states=all_hidden_states, attentions=all_attentions
)
# Modified from from transformers.models.detr.modeling_deformable_detr.DeformableDetrModel with DeformableDetrModel->Mask2FormerPixelDecoder
class Mask2FormerPixelDecoder(nn.Module):
def __init__(self, config: Mask2FormerConfig, feature_channels):
super().__init__()
self.config = config
feature_dim = config.feature_size
mask_dim = config.mask_feature_size
num_pos_features = feature_dim // 2
self.position_embedding = Mask2FormerSinePositionEmbedding(num_pos_feats=num_pos_features, normalize=True)
self.num_feature_levels = 3
transformer_in_channels = feature_channels[-self.num_feature_levels :]
self.transformer_feature_strides = config.feature_strides[-self.num_feature_levels :]
self.feature_channels = feature_channels
self.level_embed = nn.Parameter(torch.Tensor(self.num_feature_levels, feature_dim))
# Create input projection layers
if self.num_feature_levels > 1:
input_projections_list = []
for in_channels in transformer_in_channels[::-1]:
input_projections_list.append(
nn.Sequential(
nn.Conv2d(in_channels, feature_dim, kernel_size=1),
nn.GroupNorm(32, feature_dim),
)
)
self.input_projections = nn.ModuleList(input_projections_list)
else:
self.input_projections = nn.ModuleList(
[
nn.Sequential(
nn.Conv2d(transformer_in_channels[-1], feature_dim, kernel_size=1),
nn.GroupNorm(32, feature_dim),
)
]
)
self.encoder = Mask2FormerPixelDecoderEncoderOnly(config)
self.mask_projection = nn.Conv2d(feature_dim, mask_dim, kernel_size=1, stride=1, padding=0)
# Extra FPN levels
stride = min(self.transformer_feature_strides)
self.common_stride = config.common_stride
self.num_fpn_levels = int(np.log2(stride) - np.log2(self.common_stride))
lateral_convs = []
output_convs = []
for idx, in_channels in enumerate(self.feature_channels[: self.num_fpn_levels]):
lateral_conv = nn.Sequential(
nn.Conv2d(in_channels, feature_dim, kernel_size=1, bias=False),
nn.GroupNorm(32, feature_dim),
)
output_conv = nn.Sequential(
nn.Conv2d(feature_dim, feature_dim, kernel_size=3, stride=1, padding=1, bias=False),
nn.GroupNorm(32, feature_dim),
nn.ReLU(),
)
self.add_module("adapter_{}".format(idx + 1), lateral_conv)
self.add_module("layer_{}".format(idx + 1), output_conv)
lateral_convs.append(lateral_conv)
output_convs.append(output_conv)
# Order convolutional layers from low to high resolution
self.lateral_convolutions = lateral_convs[::-1]
self.output_convolutions = output_convs[::-1]
def get_valid_ratio(self, mask, dtype=torch.float32):
"""Get the valid ratio of all feature maps."""
_, height, width = mask.shape
valid_height = torch.sum(~mask[:, :, 0], 1)
valid_width = torch.sum(~mask[:, 0, :], 1)
valid_ratio_heigth = valid_height.to(dtype) / height
valid_ratio_width = valid_width.to(dtype) / width
valid_ratio = torch.stack([valid_ratio_width, valid_ratio_heigth], -1)
return valid_ratio
def forward(
self,
features,
encoder_outputs=None,
output_attentions=None,
output_hidden_states=None,
return_dict=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
)
# Apply 1x1 convolution to reduce the channel dimension to d_model (256 by default)
input_embeds = []
position_embeddings = []
for level, x in enumerate(features[::-1][: self.num_feature_levels]):
input_embeds.append(self.input_projections[level](x))
position_embeddings.append(self.position_embedding(x))
masks = [
torch.zeros((x.size(0), x.size(2), x.size(3)), device=x.device, dtype=torch.bool) for x in input_embeds
]
# Prepare encoder inputs (by flattening)
spatial_shapes = [(embed.shape[2], embed.shape[3]) for embed in input_embeds]
input_embeds_flat = torch.cat([embed.flatten(2).transpose(1, 2) for embed in input_embeds], 1)
spatial_shapes = torch.as_tensor(spatial_shapes, dtype=torch.long, device=input_embeds_flat.device)
masks_flat = torch.cat([mask.flatten(1) for mask in masks], 1)
position_embeddings = [embed.flatten(2).transpose(1, 2) for embed in position_embeddings]
level_pos_embed_flat = [x + self.level_embed[i].view(1, 1, -1) for i, x in enumerate(position_embeddings)]
level_pos_embed_flat = torch.cat(level_pos_embed_flat, 1)
level_start_index = torch.cat((spatial_shapes.new_zeros((1,)), spatial_shapes.prod(1).cumsum(0)[:-1]))
valid_ratios = torch.stack([self.get_valid_ratio(mask, dtype=input_embeds_flat.dtype) for mask in masks], 1)
# Send input_embeds_flat + masks_flat + level_pos_embed_flat (backbone + proj layer output) through encoder
if encoder_outputs is None:
encoder_outputs = self.encoder(
inputs_embeds=input_embeds_flat,
attention_mask=masks_flat,
position_embeddings=level_pos_embed_flat,
spatial_shapes=spatial_shapes,
level_start_index=level_start_index,
valid_ratios=valid_ratios,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
last_hidden_state = encoder_outputs.last_hidden_state
batch_size = last_hidden_state.shape[0]
split_sizes = [None] * self.num_feature_levels
for i in range(self.num_feature_levels):
if i < self.num_feature_levels - 1:
split_sizes[i] = level_start_index[i + 1] - level_start_index[i]
else:
split_sizes[i] = last_hidden_state.shape[1] - level_start_index[i]
encoder_output = torch.split(last_hidden_state, [size.item() for size in split_sizes], dim=1)
# Compute final features
outputs = [
x.transpose(1, 2).view(batch_size, -1, spatial_shapes[i][0], spatial_shapes[i][1])
for i, x in enumerate(encoder_output)
]
# Append extra FPN levels to outputs, ordered from low to high resolution
for idx, feature in enumerate(features[: self.num_fpn_levels][::-1]):
lateral_conv = self.lateral_convolutions[idx]
output_conv = self.output_convolutions[idx]
current_fpn = lateral_conv(feature)
# Following FPN implementation, we use nearest upsampling here
out = current_fpn + nn.functional.interpolate(
outputs[-1], size=current_fpn.shape[-2:], mode="bilinear", align_corners=False
)
out = output_conv(out)
outputs.append(out)
num_cur_levels = 0
multi_scale_features = []
for out in outputs:
if num_cur_levels < self.num_feature_levels:
multi_scale_features.append(out)
num_cur_levels += 1
return Mask2FormerPixelDecoderOutput(
mask_features=self.mask_projection(outputs[-1]),
multi_scale_features=tuple(multi_scale_features),
attentions=encoder_outputs.attentions,
)
class Mask2FormerPixelLevelModule(nn.Module):
def __init__(self, config: Mask2FormerConfig):
"""
Pixel Level Module proposed in [Masked-attention Mask Transformer for Universal Image
Segmentation](https://arxiv.org/abs/2112.01527). It runs the input image through a backbone and a pixel
decoder, generating multi-scale feature maps and pixel embeddings.
Args:
config ([`Mask2FormerConfig`]):
The configuration used to instantiate this model.
"""
super().__init__()
self.encoder = load_backbone(config)
self.decoder = Mask2FormerPixelDecoder(config, feature_channels=self.encoder.channels)
def forward(self, pixel_values: Tensor, output_hidden_states: bool = False) -> Mask2FormerPixelLevelModuleOutput:
backbone_features = self.encoder(pixel_values).feature_maps
decoder_output = self.decoder(backbone_features, output_hidden_states=output_hidden_states)
return Mask2FormerPixelLevelModuleOutput(
encoder_last_hidden_state=backbone_features[-1],
encoder_hidden_states=tuple(backbone_features) if output_hidden_states else None,
decoder_last_hidden_state=decoder_output.mask_features,
decoder_hidden_states=decoder_output.multi_scale_features,
)
# Modified from transformers.models.detr.modeling_detr.DetrAttention with Detr->Mask2Former
class Mask2FormerAttention(nn.Module):
"""
Multi-headed attention from 'Attention Is All You Need' paper. Here, we add position embeddings to the queries and
keys (as explained in the DETR paper).
"""
def __init__(
self,
embed_dim: int,
num_heads: int,
dropout: float = 0.0,
is_decoder: bool = False,
bias: bool = True,
):
super().__init__()
self.embed_dim = embed_dim
self.num_heads = num_heads
self.dropout = dropout
self.head_dim = embed_dim // num_heads
if self.head_dim * num_heads != self.embed_dim:
raise ValueError(
f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:"
f" {num_heads})."
)
self.scaling = self.head_dim**-0.5
self.k_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
self.v_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
def _shape(self, tensor: torch.Tensor, seq_len: int, batch_size: int):
return tensor.view(batch_size, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
def with_pos_embed(self, tensor: torch.Tensor, position_embeddings: Optional[Tensor]):
return tensor if position_embeddings is None else tensor + position_embeddings
def forward(
self,
hidden_states: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
position_embeddings: Optional[torch.Tensor] = None,
key_value_states: Optional[torch.Tensor] = None,
key_value_position_embeddings: Optional[torch.Tensor] = None,
output_attentions: bool = False,
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
"""Input shape: Batch x Time x Channel"""
hidden_states = hidden_states.permute(1, 0, 2) if hidden_states is not None else None
position_embeddings = position_embeddings.permute(1, 0, 2) if position_embeddings is not None else None
key_value_states = key_value_states.permute(1, 0, 2) if key_value_states is not None else None
key_value_position_embeddings = (
key_value_position_embeddings.permute(1, 0, 2) if key_value_position_embeddings is not None else None
)
# if key_value_states are provided this layer is used as a cross-attention layer
# for the decoder
is_cross_attention = key_value_states is not None
batch_size, target_len, embed_dim = hidden_states.size()
# add position embeddings to the hidden states before projecting to queries and keys
if position_embeddings is not None:
hidden_states_original = hidden_states
hidden_states = self.with_pos_embed(hidden_states, position_embeddings)
# add key-value position embeddings to the key value states
if key_value_position_embeddings is not None:
key_value_states_original = key_value_states
key_value_states = self.with_pos_embed(key_value_states, key_value_position_embeddings)
# get query proj
query_states = self.q_proj(hidden_states) * self.scaling
# get key, value proj
if is_cross_attention:
# cross_attentions
key_states = self._shape(self.k_proj(key_value_states), -1, batch_size)
value_states = self._shape(self.v_proj(key_value_states_original), -1, batch_size)
else:
# self_attention
key_states = self._shape(self.k_proj(hidden_states), -1, batch_size)
value_states = self._shape(self.v_proj(hidden_states_original), -1, batch_size)
proj_shape = (batch_size * self.num_heads, -1, self.head_dim)
query_states = self._shape(query_states, target_len, batch_size).view(*proj_shape)
key_states = key_states.view(*proj_shape)
value_states = value_states.view(*proj_shape)
source_len = key_states.size(1)
attn_weights = torch.bmm(query_states, key_states.transpose(1, 2))
if attn_weights.size() != (batch_size * self.num_heads, target_len, source_len):
raise ValueError(
f"Attention weights should be of size {(batch_size * self.num_heads, target_len, source_len)}, but is"
f" {attn_weights.size()}"
)
if attention_mask is not None:
if attention_mask.size() != (batch_size * self.num_heads, target_len, source_len):
raise ValueError(
f"Attention mask should be of size {(target_len, batch_size * self.num_heads, source_len)}, but is"
f" {attention_mask.size()}"
)
attn_weights += attention_mask
attn_weights = nn.functional.softmax(attn_weights, dim=-1)
if output_attentions:
# this operation is a bit awkward, but it's required to
# make sure that attn_weights keeps its gradient.
# In order to do so, attn_weights have to reshaped
# twice and have to be reused in the following
attn_weights_reshaped = attn_weights.view(batch_size, self.num_heads, target_len, source_len)
attn_weights = attn_weights_reshaped.view(batch_size * self.num_heads, target_len, source_len)
else:
attn_weights_reshaped = None
attn_probs = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)
attn_output = torch.bmm(attn_probs, value_states)
if attn_output.size() != (batch_size * self.num_heads, target_len, self.head_dim):
raise ValueError(
f"`attn_output` should be of size {(batch_size, self.num_heads, target_len, self.head_dim)}, but is"
f" {attn_output.size()}"
)
attn_output = attn_output.view(batch_size, self.num_heads, target_len, self.head_dim)
attn_output = attn_output.transpose(1, 2)
attn_output = attn_output.reshape(batch_size, target_len, embed_dim)
attn_output = self.out_proj(attn_output).permute(1, 0, 2)
return attn_output, attn_weights_reshaped
class Mask2FormerMaskedAttentionDecoderLayer(nn.Module):
"""
The Mask2FormerMaskedAttentionDecoderLayer is made up of self-attention, cross (masked) attention as well as FFN
blocks. The cross attention block used as part of `Mask2FormerMaskedAttentionDecoderLayer` is actually a `masked
attention` block that restricts the attention to localized features centered around predicted segments which leads
to faster convergence and improved performance. The order of self and cross (i.e. masked) attention blocks have
also been swapped in Mask2FormerMaskedAttentionDecoder compared to a standard DetrDecoder as an optimization
improvement.
Args:
config (`Mask2FormerConfig`):
The configuration used to initialize the Mask2FormerMaskedAttentionDecoder.
"""
def __init__(self, config: Mask2FormerConfig):
super().__init__()
self.config = config
self.embed_dim = self.config.hidden_dim
self.pre_norm = self.config.pre_norm
self.self_attn = Mask2FormerAttention(
embed_dim=self.embed_dim,
num_heads=config.num_attention_heads,
dropout=config.dropout,
is_decoder=True,
)
self.dropout = self.config.dropout
self.activation_fn = ACT2FN[self.config.activation_function]
self.activation_dropout = self.config.dropout
self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
self.cross_attn = nn.MultiheadAttention(self.embed_dim, self.config.num_attention_heads, self.config.dropout)
self.cross_attn_layer_norm = nn.LayerNorm(self.embed_dim)
self.fc1 = nn.Linear(self.embed_dim, self.config.dim_feedforward)
self.fc2 = nn.Linear(self.config.dim_feedforward, self.embed_dim)
self.final_layer_norm = nn.LayerNorm(self.embed_dim)
def with_pos_embed(self, tensor, pos: Optional[Tensor]):
return tensor if pos is None else tensor + pos
def forward_post(
self,
hidden_states: torch.Tensor,
level_index: int = None,
attention_mask: Optional[torch.Tensor] = None,
position_embeddings: Optional[torch.Tensor] = None,
query_position_embeddings: Optional[torch.Tensor] = None,
encoder_hidden_states: Optional[torch.Tensor] = None,
encoder_attention_mask: Optional[torch.Tensor] = None,
output_attentions: Optional[bool] = False,
):
# Masked(Cross)-Attention Block
cross_attn_weights = None
self_attn_weights = None
residual = hidden_states
hidden_states, cross_attn_weights = self.cross_attn(
query=self.with_pos_embed(hidden_states, query_position_embeddings),
key=self.with_pos_embed(encoder_hidden_states[level_index], position_embeddings[level_index]),
value=encoder_hidden_states[level_index],
attn_mask=encoder_attention_mask,
key_padding_mask=None,
)
hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
hidden_states = residual + hidden_states
hidden_states = self.cross_attn_layer_norm(hidden_states)
# Self Attention Block
residual = hidden_states
hidden_states, self_attn_weights = self.self_attn(
hidden_states=hidden_states,
position_embeddings=query_position_embeddings,
attention_mask=None,
output_attentions=True,
)
hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
hidden_states = residual + hidden_states
hidden_states = self.self_attn_layer_norm(hidden_states)
# Fully Connected
residual = hidden_states
hidden_states = self.activation_fn(self.fc1(hidden_states))
hidden_states = nn.functional.dropout(hidden_states, p=self.activation_dropout, training=self.training)
hidden_states = self.fc2(hidden_states)
hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
hidden_states = residual + hidden_states
hidden_states = self.final_layer_norm(hidden_states)
outputs = (hidden_states,)
if output_attentions:
outputs += (self_attn_weights, cross_attn_weights)
return outputs
def forward_pre(
self,
hidden_states: torch.Tensor,
level_index: int = None,
attention_mask: Optional[torch.Tensor] = None,
position_embeddings: Optional[torch.Tensor] = None,
query_position_embeddings: Optional[torch.Tensor] = None,
encoder_hidden_states: Optional[torch.Tensor] = None,
encoder_attention_mask: Optional[torch.Tensor] = None,
output_attentions: Optional[bool] = False,
):
# Masked(Cross)-Attention Block
cross_attn_weights = None
self_attn_weights = None
residual = hidden_states
hidden_states = self.cross_attn_layer_norm(hidden_states)
hidden_states, cross_attn_weights = self.cross_attn(
query=self.with_pos_embed(hidden_states, query_position_embeddings),
key=self.with_pos_embed(encoder_hidden_states[level_index], position_embeddings[level_index]),
value=encoder_hidden_states[level_index],
attn_mask=encoder_attention_mask,
key_padding_mask=None,
)
hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
hidden_states = residual + hidden_states
# Self Attention Block
residual = hidden_states
hidden_states = self.self_attn_layer_norm(hidden_states)
hidden_states, self_attn_weights = self.self_attn(
hidden_states=hidden_states,
position_embeddings=query_position_embeddings,
attention_mask=None,
output_attentions=True,
)
hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
hidden_states = residual + hidden_states
# Fully Connected
residual = hidden_states
hidden_states = self.final_layer_norm(hidden_states)
hidden_states = self.activation_fn(self.fc1(hidden_states))
hidden_states = nn.functional.dropout(hidden_states, p=self.activation_dropout, training=self.training)
hidden_states = self.fc2(hidden_states)
hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
hidden_states = residual + hidden_states
outputs = (hidden_states,)
if output_attentions:
outputs += (self_attn_weights, cross_attn_weights)
return outputs
def forward(
self,
hidden_states: torch.Tensor,
level_index: int = None,
attention_mask: Optional[torch.Tensor] = None,
position_embeddings: Optional[torch.Tensor] = None,
query_position_embeddings: Optional[torch.Tensor] = None,
encoder_hidden_states: Optional[torch.Tensor] = None,
encoder_attention_mask: Optional[torch.Tensor] = None,
output_attentions: Optional[bool] = False,
):
"""
Args:
hidden_states (`torch.FloatTensor`):
Input to the layer of shape `(seq_len, batch, embed_dim)`.
attention_mask (`torch.FloatTensor`):
Attention mask of shape `(1, seq_len, tgt_len, src_len)`.
position_embeddings (`torch.FloatTensor`, *optional*):
Position embeddings that are added to the keys in the masked-attention layer.
query_position_embeddings (`torch.FloatTensor`, *optional*):
Position embeddings that are added to the queries and keys in the self-attention layer.
encoder_hidden_states (`torch.FloatTensor`):
Cross attention input to the layer of shape `(seq_len, batch, embed_dim)`.
encoder_attention_mask (`torch.FloatTensor`):
Encoder attention mask of size`(1, seq_len, tgt_len, src_len)`.
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
returned tensors for more detail.
"""
if self.pre_norm:
outputs = self.forward_pre(
hidden_states=hidden_states,
level_index=level_index,
position_embeddings=position_embeddings,
query_position_embeddings=query_position_embeddings,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
output_attentions=output_attentions,
)
else:
outputs = self.forward_post(
hidden_states=hidden_states,
level_index=level_index,
position_embeddings=position_embeddings,
query_position_embeddings=query_position_embeddings,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
output_attentions=output_attentions,
)
return outputs
class Mask2FormerMaskedAttentionDecoder(nn.Module):
"""
Transformer decoder consisting of *config.decoder_layers* layers. Each layer is a
[`Mask2FormerMaskedAttentionDecoderLayer`]. The decoder updates the query embeddings through multiple cross
(masked) and self-attention layers. The decoder uses a new **masked attention** mechanism instead of the standard
cross-attention, which extracts localized features by constraining cross-attention to within the foreground region
of the predicted mask for each query, instead of attending to the full feature map.
Args:
config (`Mask2FormerConfig`):
Configuration used to instantiate Mask2FormerMaskedAttentionDecoder.
"""
def __init__(self, config: Mask2FormerConfig):
super().__init__()
self.config = config
self.mask_feature_size = config.mask_feature_size
self.dropout = config.dropout
self.layerdrop = config.dropout
self.num_feature_levels = 3 # level embedding (3 scales)
self.decoder_layers = config.decoder_layers - 1
self.layers = nn.ModuleList(
[Mask2FormerMaskedAttentionDecoderLayer(self.config) for _ in range(self.decoder_layers)]
)
self.layernorm = nn.LayerNorm(config.hidden_dim)
self.mask_predictor = Mask2FormerMaskPredictor(
hidden_size=config.hidden_dim,
num_heads=config.num_attention_heads,
mask_feature_size=self.mask_feature_size,
)
self.gradient_checkpointing = False
def forward(
self,
inputs_embeds: torch.Tensor = None,
multi_stage_positional_embeddings: torch.Tensor = None,
pixel_embeddings: torch.Tensor = None,
encoder_hidden_states: torch.Tensor = None,
query_position_embeddings: torch.Tensor = None,
feature_size_list: List = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
):
r"""
Args:
inputs_embeds (`torch.FloatTensor` of shape `(num_queries, batch_size, hidden_size)`):
The query embeddings that are passed into the decoder.
multi_stage_positional_embeddings (`torch.FloatTensor` of shape `(height*width, batch_size, num_channels)`):
Position embeddings that are added to the keys in each cross(masked)-attention layer.
pixel_embeddings (`torch.FloatTensor`):
Tensor of shape `(batch_size, num_channels, height, width)`, 1/4 scale features from the last Pixel
Decoder.
query_position_embeddings (`torch.FloatTensor` of shape `(num_queries, batch_size, hidden_size)`):
, *optional*): Position embeddings that are added to the queries and keys in each self-attention layer.
encoder_hidden_states (`torch.FloatTensor` of shape `(batch_size, encoder_sequence_length, hidden_size)`):
Sequence of hidden-states at the output of the last layer of the encoder. Used in the
cross(masked)-attention of the decoder.
feature_size_list (`List[torch.Size]`):
This is a list containing shapes (height & width) of multi-scale features from the Pixel Decoder.
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
returned tensors for more detail.
output_hidden_states (`bool`, *optional*):
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
for more detail.
return_dict (`bool`, *optional*):
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
"""
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 inputs_embeds is not None:
hidden_states = inputs_embeds
# intermediate hidden states with layernorm applied - required for predicting class logits
intermediate = ()
# decoder layers
all_hidden_states = () if output_hidden_states else None
attentions = () if output_attentions else None
# intermediate mask predictions from transformer decoder layers
intermediate_mask_predictions = ()
intermediate_hidden_states = self.layernorm(inputs_embeds)
intermediate += (intermediate_hidden_states,)
predicted_mask, attention_mask = self.mask_predictor(
intermediate_hidden_states, pixel_embeddings, feature_size_list[0]
)
intermediate_mask_predictions += (predicted_mask,)
for idx, decoder_layer in enumerate(self.layers):
if output_hidden_states:
all_hidden_states += (hidden_states,)
dropout_probability = torch.rand([])
if self.training and (dropout_probability < self.layerdrop):
continue
if self.gradient_checkpointing and self.training:
layer_outputs = self._gradient_checkpointing_func(
decoder_layer.__call__,
hidden_states,
attention_mask,
encoder_hidden_states,
None,
None,
output_attentions,
)
else:
level_index = idx % self.num_feature_levels
attention_mask[torch.where(attention_mask.sum(-1) == attention_mask.shape[-1])] = False
layer_outputs = decoder_layer(
hidden_states,
level_index=level_index,
position_embeddings=multi_stage_positional_embeddings,
query_position_embeddings=query_position_embeddings,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=attention_mask,
output_attentions=output_attentions,
)
intermediate_hidden_states = self.layernorm(layer_outputs[0])
predicted_mask, attention_mask = self.mask_predictor(
intermediate_hidden_states,
pixel_embeddings,
feature_size_list[(idx + 1) % self.num_feature_levels],
)
intermediate_mask_predictions += (predicted_mask,)
# add intermediate hidden states with layer norm applied which will be used for predicting class logits
intermediate += (intermediate_hidden_states,)
hidden_states = layer_outputs[0]
if output_attentions:
attentions += (layer_outputs[1],)
# add hidden states from the last decoder layer
if output_hidden_states:
all_hidden_states += (hidden_states,)
hidden_states = hidden_states.transpose(1, 0)
if not return_dict:
outputs = [hidden_states, all_hidden_states, attentions, intermediate, intermediate_mask_predictions]
return tuple(v for v in outputs if v is not None)
return Mask2FormerMaskedAttentionDecoderOutput(
last_hidden_state=hidden_states,
hidden_states=all_hidden_states,
attentions=attentions,
intermediate_hidden_states=intermediate,
masks_queries_logits=intermediate_mask_predictions,
)
# Copied from transformers.models.maskformer.modeling_maskformer.PredictionBlock with MaskFormer->Mask2Former
class Mask2FormerPredictionBlock(nn.Module):
def __init__(self, in_dim: int, out_dim: int, activation: nn.Module) -> None:
super().__init__()
self.layers = [nn.Linear(in_dim, out_dim), activation]
# Maintain submodule indexing as if part of a Sequential block
for i, layer in enumerate(self.layers):
self.add_module(str(i), layer)
def forward(self, input: Tensor) -> Tensor:
hidden_state = input
for layer in self.layers:
hidden_state = layer(hidden_state)
return hidden_state
class Mask2FormerMLPPredictionHead(nn.Module):
def __init__(self, input_dim: int, hidden_dim: int, output_dim: int, num_layers: int = 3):
"""
A classic Multi Layer Perceptron (MLP).
Args:
input_dim (`int`):
The input dimensions.
hidden_dim (`int`):
The hidden dimensions.
output_dim (`int`):
The output dimensions.
num_layers (int, *optional*, defaults to 3):
The number of layers.
"""
super().__init__()
in_dims = [input_dim] + [hidden_dim] * (num_layers - 1)
out_dims = [hidden_dim] * (num_layers - 1) + [output_dim]
self.layers = []
for i, (in_dim, out_dim) in enumerate(zip(in_dims, out_dims)):
activation = nn.ReLU() if i < num_layers - 1 else nn.Identity()
layer = Mask2FormerPredictionBlock(in_dim, out_dim, activation=activation)
self.layers.append(layer)
# Provide backwards compatibility from when the class inherited from nn.Sequential
# In nn.Sequential subclasses, the name given to the layer is its index in the sequence.
# In nn.Module subclasses they derived from the instance attribute they are assigned to e.g.
# self.my_layer_name = Layer()
# We can't give instance attributes integer names i.e. self.0 is not permitted and so need to register
# explicitly
self.add_module(str(i), layer)
def forward(self, input: Tensor) -> Tensor:
hidden_state = input
for layer in self.layers:
hidden_state = layer(hidden_state)
return hidden_state
class Mask2FormerMaskPredictor(nn.Module):
def __init__(self, hidden_size: int, num_heads: int, mask_feature_size: torch.Tensor):
"""
This class is used to get the predicted mask for a given Mask2FormerMaskedAttentionDecoder layer. It also
generates the binarized attention mask associated with the given predicted mask. The attention mask obtained
using predicted mask of the (l-1)th decoder layer is fed to the cross(masked)-attention block of the next
decoder layer as input.
Args:
hidden_size (`int`):
The feature dimension of the Mask2FormerMaskedAttentionDecoder
num_heads (`int`):
The number of heads used in the Mask2FormerMaskedAttentionDecoder
mask_feature_size (`torch.Tensor`):
one of the output dimensions of the predicted masks for each query
"""
super().__init__()
self.hidden_size = hidden_size
self.num_heads = num_heads
self.mask_embedder = Mask2FormerMLPPredictionHead(self.hidden_size, self.hidden_size, mask_feature_size)
def forward(self, outputs: torch.Tensor, pixel_embeddings: torch.Tensor, attention_mask_target_size: int = None):
mask_embeddings = self.mask_embedder(outputs.transpose(0, 1))
is_tracing = torch.jit.is_tracing() or isinstance(outputs, torch.fx.Proxy) or is_torchdynamo_compiling()
# Sum up over the channels
if is_tracing and not is_torch_greater_or_equal_than_2_1:
# Equivalent to einsum('bqc, bchw -> bqhw') but jit friendly
batch_size, num_queries, num_channels = mask_embeddings.shape
_, _, height, width = pixel_embeddings.shape
outputs_mask = torch.zeros((batch_size, num_queries, height, width), device=mask_embeddings.device)
for c in range(num_channels):
outputs_mask += mask_embeddings[..., c][..., None, None] * pixel_embeddings[:, None, c]
else:
outputs_mask = torch.einsum("bqc, bchw -> bqhw", mask_embeddings, pixel_embeddings)
attention_mask = nn.functional.interpolate(
outputs_mask, size=attention_mask_target_size, mode="bilinear", align_corners=False
)
attention_mask = attention_mask.sigmoid().flatten(2).unsqueeze(1).repeat(1, self.num_heads, 1, 1)
attention_mask = (attention_mask.flatten(0, 1) < 0.5).bool()
attention_mask = attention_mask.detach()
return outputs_mask, attention_mask
class Mask2FormerTransformerModule(nn.Module):
"""
The Mask2Former's transformer module.
"""
def __init__(self, in_features: int, config: Mask2FormerConfig):
super().__init__()
hidden_dim = config.hidden_dim
self.num_feature_levels = 3
self.position_embedder = Mask2FormerSinePositionEmbedding(num_pos_feats=hidden_dim // 2, normalize=True)
self.queries_embedder = nn.Embedding(config.num_queries, hidden_dim)
self.queries_features = nn.Embedding(config.num_queries, hidden_dim)
self.input_projections = []
for _ in range(self.num_feature_levels):
if in_features != hidden_dim or config.enforce_input_projection:
self.input_projections.append(nn.Conv2d(in_features, hidden_dim, kernel_size=1))
else:
self.input_projections.append(nn.Sequential())
self.decoder = Mask2FormerMaskedAttentionDecoder(config=config)
self.level_embed = nn.Embedding(self.num_feature_levels, hidden_dim)
def forward(
self,
multi_scale_features: List[Tensor],
mask_features: Tensor,
output_hidden_states: bool = False,
output_attentions: bool = False,
) -> Mask2FormerMaskedAttentionDecoderOutput:
multi_stage_features = []
multi_stage_positional_embeddings = []
size_list = []
for i in range(self.num_feature_levels):
size_list.append(multi_scale_features[i].shape[-2:])
multi_stage_positional_embeddings.append(self.position_embedder(multi_scale_features[i], None).flatten(2))
multi_stage_features.append(
self.input_projections[i](multi_scale_features[i]).flatten(2)
+ self.level_embed.weight[i][None, :, None]
)
# Flatten (batch_size, num_channels, height, width) -> (height*width, batch_size, num_channels)
multi_stage_positional_embeddings[-1] = multi_stage_positional_embeddings[-1].permute(2, 0, 1)
multi_stage_features[-1] = multi_stage_features[-1].permute(2, 0, 1)
_, batch_size, _ = multi_stage_features[0].shape
# [num_queries, batch_size, num_channels]
query_embeddings = self.queries_embedder.weight.unsqueeze(1).repeat(1, batch_size, 1)
query_features = self.queries_features.weight.unsqueeze(1).repeat(1, batch_size, 1)
decoder_output = self.decoder(
inputs_embeds=query_features,
multi_stage_positional_embeddings=multi_stage_positional_embeddings,
pixel_embeddings=mask_features,
encoder_hidden_states=multi_stage_features,
query_position_embeddings=query_embeddings,
feature_size_list=size_list,
output_hidden_states=output_hidden_states,
output_attentions=output_attentions,
return_dict=True,
)
return decoder_output
MASK2FORMER_START_DOCSTRING = r"""
This model is a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) sub-class. Use
it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and
behavior.
Parameters:
config ([`Mask2FormerConfig`]): Model configuration class with all the parameters of the model.
Initializing with a config file does not load the weights associated with the model, only the
configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
"""
MASK2FORMER_INPUTS_DOCSTRING = r"""
Args:
pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
Pixel values. Pixel values can be obtained using [`AutoImageProcessor`]. See
[`AutoImageProcessor.preprocess`] for details.
pixel_mask (`torch.LongTensor` of shape `(batch_size, height, width)`, *optional*):
Mask to avoid performing attention on padding pixel values. Mask values selected in `[0, 1]`:
- 1 for pixels that are real (i.e. **not masked**),
- 0 for pixels that are padding (i.e. **masked**).
[What are attention masks?](../glossary#attention-mask)
output_hidden_states (`bool`, *optional*):
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
more detail.
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of Detr's decoder attention layers.
return_dict (`bool`, *optional*):
Whether or not to return a [`~Mask2FormerModelOutput`] instead of a plain tuple.
"""
class Mask2FormerPreTrainedModel(PreTrainedModel):
config_class = Mask2FormerConfig
base_model_prefix = "model"
main_input_name = "pixel_values"
def _init_weights(self, module: nn.Module):
xavier_std = self.config.init_xavier_std
std = self.config.init_std
if isinstance(module, Mask2FormerTransformerModule):
if module.input_projections is not None:
for input_projection in module.input_projections:
if not isinstance(input_projection, nn.Sequential):
nn.init.xavier_uniform_(input_projection.weight, gain=xavier_std)
nn.init.constant_(input_projection.bias, 0)
elif isinstance(module, Mask2FormerPixelDecoderEncoderMultiscaleDeformableAttention):
nn.init.constant_(module.sampling_offsets.weight.data, 0.0)
thetas = torch.arange(module.n_heads, dtype=torch.int64).float() * (2.0 * math.pi / module.n_heads)
grid_init = torch.stack([thetas.cos(), thetas.sin()], -1)
grid_init = (
(grid_init / grid_init.abs().max(-1, keepdim=True)[0])
.view(module.n_heads, 1, 1, 2)
.repeat(1, module.n_levels, module.n_points, 1)
)
for i in range(module.n_points):
grid_init[:, :, i, :] *= i + 1
with torch.no_grad():
module.sampling_offsets.bias = nn.Parameter(grid_init.view(-1))
nn.init.constant_(module.attention_weights.weight.data, 0.0)
nn.init.constant_(module.attention_weights.bias.data, 0.0)
nn.init.xavier_uniform_(module.value_proj.weight.data)
nn.init.constant_(module.value_proj.bias.data, 0.0)
nn.init.xavier_uniform_(module.output_proj.weight.data)
nn.init.constant_(module.output_proj.bias.data, 0.0)
elif isinstance(module, Mask2FormerMaskedAttentionDecoderLayer):
for p in module.parameters():
if p.dim() > 1:
nn.init.xavier_uniform_(p, gain=xavier_std)
elif isinstance(module, Mask2FormerPixelLevelModule):
for submodule in module.modules():
if isinstance(submodule, (nn.Conv2d, nn.Linear)):
submodule.weight.data.normal_(mean=0.0, std=std)
if submodule.bias is not None:
submodule.bias.data.zero_()
elif isinstance(module, Mask2FormerPixelDecoder):
for p in module.parameters():
if p.dim() > 1:
nn.init.xavier_uniform_(p)
nn.init.normal_(module.level_embed, std=0)
elif isinstance(module, Mask2FormerPixelDecoderEncoderOnly):
for p in module.parameters():
if p.dim() > 1:
nn.init.xavier_uniform_(p)
elif isinstance(module, (nn.Linear, nn.Conv2d, nn.BatchNorm2d)):
module.weight.data.normal_(mean=0.0, std=std)
if module.bias is not None:
module.bias.data.zero_()
elif isinstance(module, nn.Embedding):
module.weight.data.normal_(mean=0.0, std=std)
if module.padding_idx is not None:
module.weight.data[module.padding_idx].zero_()
if hasattr(module, "reference_points"):
nn.init.xavier_uniform_(module.reference_points.weight.data, gain=1.0)
nn.init.constant_(module.reference_points.bias.data, 0.0)
@add_start_docstrings(
"The bare Mask2Former Model outputting raw hidden-states without any specific head on top.",
MASK2FORMER_START_DOCSTRING,
)
class Mask2FormerModel(Mask2FormerPreTrainedModel):
main_input_name = "pixel_values"
def __init__(self, config: Mask2FormerConfig):
super().__init__(config)
self.pixel_level_module = Mask2FormerPixelLevelModule(config)
self.transformer_module = Mask2FormerTransformerModule(in_features=config.feature_size, config=config)
self.post_init()
@add_start_docstrings_to_model_forward(MASK2FORMER_INPUTS_DOCSTRING)
@replace_return_docstrings(output_type=Mask2FormerModelOutput, config_class=_CONFIG_FOR_DOC)
def forward(
self,
pixel_values: Tensor,
pixel_mask: Optional[Tensor] = None,
output_hidden_states: Optional[bool] = None,
output_attentions: Optional[bool] = None,
return_dict: Optional[bool] = None,
) -> Mask2FormerModelOutput:
r"""
Returns:
`Mask2FormerModelOutput`
Examples:
```python
>>> import torch
>>> from PIL import Image
>>> import requests
>>> from transformers import AutoImageProcessor, Mask2FormerModel
>>> # load image
>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
>>> image = Image.open(requests.get(url, stream=True).raw)
>>> # load image preprocessor and Mask2FormerModel trained on COCO instance segmentation dataset
>>> image_processor = AutoImageProcessor.from_pretrained("facebook/mask2former-swin-small-coco-instance")
>>> model = Mask2FormerModel.from_pretrained("facebook/mask2former-swin-small-coco-instance")
>>> inputs = image_processor(image, return_tensors="pt")
>>> # forward pass
>>> with torch.no_grad():
... outputs = model(**inputs)
>>> # model outputs last hidden states of shape (batch_size, num_queries, hidden_size)
>>> print(outputs.transformer_decoder_last_hidden_state.shape)
torch.Size([1, 100, 256])
```
"""
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
batch_size, _, height, width = pixel_values.shape
if pixel_mask is None:
pixel_mask = torch.ones((batch_size, height, width), device=pixel_values.device)
pixel_level_module_output = self.pixel_level_module(
pixel_values=pixel_values, output_hidden_states=output_hidden_states
)
transformer_module_output = self.transformer_module(
multi_scale_features=pixel_level_module_output.decoder_hidden_states,
mask_features=pixel_level_module_output.decoder_last_hidden_state,
output_hidden_states=True,
output_attentions=output_attentions,
)
encoder_hidden_states = None
pixel_decoder_hidden_states = None
transformer_decoder_hidden_states = None
transformer_decoder_intermediate_states = None
if output_hidden_states:
encoder_hidden_states = pixel_level_module_output.encoder_hidden_states
pixel_decoder_hidden_states = pixel_level_module_output.decoder_hidden_states
transformer_decoder_hidden_states = transformer_module_output.hidden_states
transformer_decoder_intermediate_states = transformer_module_output.intermediate_hidden_states
output = Mask2FormerModelOutput(
encoder_last_hidden_state=pixel_level_module_output.encoder_last_hidden_state,
pixel_decoder_last_hidden_state=pixel_level_module_output.decoder_last_hidden_state,
transformer_decoder_last_hidden_state=transformer_module_output.last_hidden_state,
encoder_hidden_states=encoder_hidden_states,
pixel_decoder_hidden_states=pixel_decoder_hidden_states,
transformer_decoder_hidden_states=transformer_decoder_hidden_states,
transformer_decoder_intermediate_states=transformer_decoder_intermediate_states,
attentions=transformer_module_output.attentions,
masks_queries_logits=transformer_module_output.masks_queries_logits,
)
if not return_dict:
output = tuple(v for v in output.values() if v is not None)
return output
@add_start_docstrings(
"The Mask2Former Model with heads on top for instance/semantic/panoptic segmentation.",
MASK2FORMER_START_DOCSTRING,
)
class Mask2FormerForUniversalSegmentation(Mask2FormerPreTrainedModel):
main_input_name = "pixel_values"
def __init__(self, config: Mask2FormerConfig):
super().__init__(config)
self.model = Mask2FormerModel(config)
self.weight_dict: Dict[str, float] = {
"loss_cross_entropy": config.class_weight,
"loss_mask": config.mask_weight,
"loss_dice": config.dice_weight,
}
self.class_predictor = nn.Linear(config.hidden_dim, config.num_labels + 1)
self.criterion = Mask2FormerLoss(config=config, weight_dict=self.weight_dict)
self.post_init()
def get_loss_dict(
self,
masks_queries_logits: Tensor,
class_queries_logits: Tensor,
mask_labels: Tensor,
class_labels: Tensor,
auxiliary_predictions: Dict[str, Tensor],
) -> Dict[str, Tensor]:
loss_dict: Dict[str, Tensor] = self.criterion(
masks_queries_logits=masks_queries_logits,
class_queries_logits=class_queries_logits,
mask_labels=mask_labels,
class_labels=class_labels,
auxiliary_predictions=auxiliary_predictions,
)
# weight each loss by `self.weight_dict[<LOSS_NAME>]` including auxiliary losses
for key, weight in self.weight_dict.items():
for loss_key, loss in loss_dict.items():
if key in loss_key:
loss *= weight
return loss_dict
def get_loss(self, loss_dict: Dict[str, Tensor]) -> Tensor:
return sum(loss_dict.values())
def get_auxiliary_logits(self, classes: torch.Tensor, output_masks: torch.Tensor):
auxiliary_logits: List[Dict(str, Tensor)] = []
for aux_binary_masks, aux_classes in zip(output_masks[:-1], classes[:-1]):
auxiliary_logits.append({"masks_queries_logits": aux_binary_masks, "class_queries_logits": aux_classes})
return auxiliary_logits
@add_start_docstrings_to_model_forward(MASK2FORMER_INPUTS_DOCSTRING)
@replace_return_docstrings(output_type=Mask2FormerForUniversalSegmentationOutput, config_class=_CONFIG_FOR_DOC)
def forward(
self,
pixel_values: Tensor,
mask_labels: Optional[List[Tensor]] = None,
class_labels: Optional[List[Tensor]] = None,
pixel_mask: Optional[Tensor] = None,
output_hidden_states: Optional[bool] = None,
output_auxiliary_logits: Optional[bool] = None,
output_attentions: Optional[bool] = None,
return_dict: Optional[bool] = None,
) -> Mask2FormerForUniversalSegmentationOutput:
r"""
mask_labels (`List[torch.Tensor]`, *optional*):
List of mask labels of shape `(num_labels, height, width)` to be fed to a model
class_labels (`List[torch.LongTensor]`, *optional*):
list of target class labels of shape `(num_labels, height, width)` to be fed to a model. They identify the
labels of `mask_labels`, e.g. the label of `mask_labels[i][j]` if `class_labels[i][j]`.
Returns:
`Mask2FormerUniversalSegmentationOutput`
Examples:
Instance segmentation example:
```python
>>> from transformers import AutoImageProcessor, Mask2FormerForUniversalSegmentation
>>> from PIL import Image
>>> import requests
>>> import torch
>>> # Load Mask2Former trained on COCO instance segmentation dataset
>>> image_processor = AutoImageProcessor.from_pretrained("facebook/mask2former-swin-small-coco-instance")
>>> model = Mask2FormerForUniversalSegmentation.from_pretrained(
... "facebook/mask2former-swin-small-coco-instance"
... )
>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
>>> image = Image.open(requests.get(url, stream=True).raw)
>>> inputs = image_processor(image, return_tensors="pt")
>>> with torch.no_grad():
... outputs = model(**inputs)
>>> # Model predicts class_queries_logits of shape `(batch_size, num_queries)`
>>> # and masks_queries_logits of shape `(batch_size, num_queries, height, width)`
>>> class_queries_logits = outputs.class_queries_logits
>>> masks_queries_logits = outputs.masks_queries_logits
>>> # Perform post-processing to get instance segmentation map
>>> pred_instance_map = image_processor.post_process_semantic_segmentation(
... outputs, target_sizes=[image.size[::-1]]
... )[0]
>>> print(pred_instance_map.shape)
torch.Size([480, 640])
```
Semantic segmentation example:
```python
>>> from transformers import AutoImageProcessor, Mask2FormerForUniversalSegmentation
>>> from PIL import Image
>>> import requests
>>> import torch
>>> # Load Mask2Former trained on ADE20k semantic segmentation dataset
>>> image_processor = AutoImageProcessor.from_pretrained("facebook/mask2former-swin-small-ade-semantic")
>>> model = Mask2FormerForUniversalSegmentation.from_pretrained("facebook/mask2former-swin-small-ade-semantic")
>>> url = (
... "https://huggingface.co/datasets/hf-internal-testing/fixtures_ade20k/resolve/main/ADE_val_00000001.jpg"
... )
>>> image = Image.open(requests.get(url, stream=True).raw)
>>> inputs = image_processor(image, return_tensors="pt")
>>> with torch.no_grad():
... outputs = model(**inputs)
>>> # Model predicts class_queries_logits of shape `(batch_size, num_queries)`
>>> # and masks_queries_logits of shape `(batch_size, num_queries, height, width)`
>>> class_queries_logits = outputs.class_queries_logits
>>> masks_queries_logits = outputs.masks_queries_logits
>>> # Perform post-processing to get semantic segmentation map
>>> pred_semantic_map = image_processor.post_process_semantic_segmentation(
... outputs, target_sizes=[image.size[::-1]]
... )[0]
>>> print(pred_semantic_map.shape)
torch.Size([512, 683])
```
Panoptic segmentation example:
```python
>>> from transformers import AutoImageProcessor, Mask2FormerForUniversalSegmentation
>>> from PIL import Image
>>> import requests
>>> import torch
>>> # Load Mask2Former trained on CityScapes panoptic segmentation dataset
>>> image_processor = AutoImageProcessor.from_pretrained("facebook/mask2former-swin-small-cityscapes-panoptic")
>>> model = Mask2FormerForUniversalSegmentation.from_pretrained(
... "facebook/mask2former-swin-small-cityscapes-panoptic"
... )
>>> url = "https://cdn-media.huggingface.co/Inference-API/Sample-results-on-the-Cityscapes-dataset-The-above-images-show-how-our-method-can-handle.png"
>>> image = Image.open(requests.get(url, stream=True).raw)
>>> inputs = image_processor(image, return_tensors="pt")
>>> with torch.no_grad():
... outputs = model(**inputs)
>>> # Model predicts class_queries_logits of shape `(batch_size, num_queries)`
>>> # and masks_queries_logits of shape `(batch_size, num_queries, height, width)`
>>> class_queries_logits = outputs.class_queries_logits
>>> masks_queries_logits = outputs.masks_queries_logits
>>> # Perform post-processing to get panoptic segmentation map
>>> pred_panoptic_map = image_processor.post_process_panoptic_segmentation(
... outputs, target_sizes=[image.size[::-1]]
... )[0]["segmentation"]
>>> print(pred_panoptic_map.shape)
torch.Size([338, 676])
```
"""
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
outputs = self.model(
pixel_values=pixel_values,
pixel_mask=pixel_mask,
output_hidden_states=output_hidden_states or self.config.use_auxiliary_loss,
output_attentions=output_attentions,
return_dict=True,
)
loss, loss_dict, auxiliary_logits = None, None, None
class_queries_logits = ()
for decoder_output in outputs.transformer_decoder_intermediate_states:
class_prediction = self.class_predictor(decoder_output.transpose(0, 1))
class_queries_logits += (class_prediction,)
masks_queries_logits = outputs.masks_queries_logits
auxiliary_logits = self.get_auxiliary_logits(class_queries_logits, masks_queries_logits)
if mask_labels is not None and class_labels is not None:
loss_dict = self.get_loss_dict(
masks_queries_logits=masks_queries_logits[-1],
class_queries_logits=class_queries_logits[-1],
mask_labels=mask_labels,
class_labels=class_labels,
auxiliary_predictions=auxiliary_logits,
)
loss = self.get_loss(loss_dict)
encoder_hidden_states = None
pixel_decoder_hidden_states = None
transformer_decoder_hidden_states = None
if output_hidden_states:
encoder_hidden_states = outputs.encoder_hidden_states
pixel_decoder_hidden_states = outputs.pixel_decoder_hidden_states
transformer_decoder_hidden_states = outputs.transformer_decoder_hidden_states
output_auxiliary_logits = (
self.config.output_auxiliary_logits if output_auxiliary_logits is None else output_auxiliary_logits
)
if not output_auxiliary_logits:
auxiliary_logits = None
output = Mask2FormerForUniversalSegmentationOutput(
loss=loss,
class_queries_logits=class_queries_logits[-1],
masks_queries_logits=masks_queries_logits[-1],
auxiliary_logits=auxiliary_logits,
encoder_last_hidden_state=outputs.encoder_last_hidden_state,
pixel_decoder_last_hidden_state=outputs.pixel_decoder_last_hidden_state,
transformer_decoder_last_hidden_state=outputs.transformer_decoder_last_hidden_state,
encoder_hidden_states=encoder_hidden_states,
pixel_decoder_hidden_states=pixel_decoder_hidden_states,
transformer_decoder_hidden_states=transformer_decoder_hidden_states,
attentions=outputs.attentions,
)
if not return_dict:
output = tuple(v for v in output.values() if v is not None)
if loss is not None:
output = (loss) + output
return output
|
transformers/src/transformers/models/mask2former/modeling_mask2former.py/0
|
{
"file_path": "transformers/src/transformers/models/mask2former/modeling_mask2former.py",
"repo_id": "transformers",
"token_count": 50995
}
| 367
|
# coding=utf-8
# Copyright 2023 The HuggingFace Inc. team. 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.
"""Processor class for MGP-STR."""
import warnings
from transformers import AutoTokenizer
from transformers.utils import is_torch_available
from transformers.utils.generic import ExplicitEnum
from ...processing_utils import ProcessorMixin
if is_torch_available():
import torch
class DecodeType(ExplicitEnum):
CHARACTER = "char"
BPE = "bpe"
WORDPIECE = "wp"
SUPPORTED_ANNOTATION_FORMATS = (DecodeType.CHARACTER, DecodeType.BPE, DecodeType.WORDPIECE)
class MgpstrProcessor(ProcessorMixin):
r"""
Constructs a MGP-STR processor which wraps an image processor and MGP-STR tokenizers into a single
[`MgpstrProcessor`] offers all the functionalities of `ViTImageProcessor`] and [`MgpstrTokenizer`]. See the
[`~MgpstrProcessor.__call__`] and [`~MgpstrProcessor.batch_decode`] for more information.
Args:
image_processor (`ViTImageProcessor`, *optional*):
An instance of `ViTImageProcessor`. The image processor is a required input.
tokenizer ([`MgpstrTokenizer`], *optional*):
The tokenizer is a required input.
"""
attributes = ["image_processor", "char_tokenizer"]
image_processor_class = "ViTImageProcessor"
char_tokenizer_class = "MgpstrTokenizer"
def __init__(self, image_processor=None, tokenizer=None, **kwargs):
feature_extractor = None
if "feature_extractor" in kwargs:
warnings.warn(
"The `feature_extractor` argument is deprecated and will be removed in v5, use `image_processor`"
" instead.",
FutureWarning,
)
feature_extractor = kwargs.pop("feature_extractor")
image_processor = image_processor if image_processor is not None else feature_extractor
if image_processor is None:
raise ValueError("You need to specify an `image_processor`.")
if tokenizer is None:
raise ValueError("You need to specify a `tokenizer`.")
self.char_tokenizer = tokenizer
self.bpe_tokenizer = AutoTokenizer.from_pretrained("openai-community/gpt2")
self.wp_tokenizer = AutoTokenizer.from_pretrained("google-bert/bert-base-uncased")
super().__init__(image_processor, tokenizer)
def __call__(self, text=None, images=None, return_tensors=None, **kwargs):
"""
When used in normal mode, this method forwards all its arguments to ViTImageProcessor's
[`~ViTImageProcessor.__call__`] and returns its output. This method also forwards the `text` and `kwargs`
arguments to MgpstrTokenizer's [`~MgpstrTokenizer.__call__`] if `text` is not `None` to encode the text. Please
refer to the doctsring of the above methods for more information.
"""
if images is None and text is None:
raise ValueError("You need to specify either an `images` or `text` input to process.")
if images is not None:
inputs = self.image_processor(images, return_tensors=return_tensors, **kwargs)
if text is not None:
encodings = self.char_tokenizer(text, return_tensors=return_tensors, **kwargs)
if text is None:
return inputs
elif images is None:
return encodings
else:
inputs["labels"] = encodings["input_ids"]
return inputs
def batch_decode(self, sequences):
"""
Convert a list of lists of token ids into a list of strings by calling decode.
Args:
sequences (`torch.Tensor`):
List of tokenized input ids.
Returns:
`Dict[str, any]`: Dictionary of all the outputs of the decoded results.
generated_text (`List[str]`): The final results after fusion of char, bpe, and wp. scores
(`List[float]`): The final scores after fusion of char, bpe, and wp. char_preds (`List[str]`): The list
of character decoded sentences. bpe_preds (`List[str]`): The list of bpe decoded sentences. wp_preds
(`List[str]`): The list of wp decoded sentences.
This method forwards all its arguments to PreTrainedTokenizer's [`~PreTrainedTokenizer.batch_decode`]. Please
refer to the docstring of this method for more information.
"""
char_preds, bpe_preds, wp_preds = sequences
batch_size = char_preds.size(0)
char_strs, char_scores = self._decode_helper(char_preds, "char")
bpe_strs, bpe_scores = self._decode_helper(bpe_preds, "bpe")
wp_strs, wp_scores = self._decode_helper(wp_preds, "wp")
final_strs = []
final_scores = []
for i in range(batch_size):
scores = [char_scores[i], bpe_scores[i], wp_scores[i]]
strs = [char_strs[i], bpe_strs[i], wp_strs[i]]
max_score_index = scores.index(max(scores))
final_strs.append(strs[max_score_index])
final_scores.append(scores[max_score_index])
out = {}
out["generated_text"] = final_strs
out["scores"] = final_scores
out["char_preds"] = char_strs
out["bpe_preds"] = bpe_strs
out["wp_preds"] = wp_strs
return out
def _decode_helper(self, pred_logits, format):
"""
Convert a list of lists of bpe token ids into a list of strings by calling bpe tokenizer.
Args:
pred_logits (`torch.Tensor`):
List of model prediction logits.
format (`Union[DecoderType, str]`):
Type of model prediction. Must be one of ['char', 'bpe', 'wp'].
Returns:
`tuple`:
dec_strs(`str`): The decode strings of model prediction. conf_scores(`List[float]`): The confidence
score of model prediction.
"""
if format == DecodeType.CHARACTER:
decoder = self.char_decode
eos_token = 1
eos_str = "[s]"
elif format == DecodeType.BPE:
decoder = self.bpe_decode
eos_token = 2
eos_str = "#"
elif format == DecodeType.WORDPIECE:
decoder = self.wp_decode
eos_token = 102
eos_str = "[SEP]"
else:
raise ValueError(f"Format {format} is not supported.")
dec_strs, conf_scores = [], []
batch_size = pred_logits.size(0)
batch_max_length = pred_logits.size(1)
_, preds_index = pred_logits.topk(1, dim=-1, largest=True, sorted=True)
preds_index = preds_index.view(-1, batch_max_length)[:, 1:]
preds_str = decoder(preds_index)
preds_max_prob, _ = torch.nn.functional.softmax(pred_logits, dim=2).max(dim=2)
preds_max_prob = preds_max_prob[:, 1:]
for index in range(batch_size):
pred_eos = preds_str[index].find(eos_str)
pred = preds_str[index][:pred_eos]
pred_index = preds_index[index].cpu().tolist()
pred_eos_index = pred_index.index(eos_token) if eos_token in pred_index else -1
pred_max_prob = preds_max_prob[index][: pred_eos_index + 1]
confidence_score = pred_max_prob.cumprod(dim=0)[-1] if pred_max_prob.nelement() != 0 else 0.0
dec_strs.append(pred)
conf_scores.append(confidence_score)
return dec_strs, conf_scores
def char_decode(self, sequences):
"""
Convert a list of lists of char token ids into a list of strings by calling char tokenizer.
Args:
sequences (`torch.Tensor`):
List of tokenized input ids.
Returns:
`List[str]`: The list of char decoded sentences.
"""
decode_strs = [seq.replace(" ", "") for seq in self.char_tokenizer.batch_decode(sequences)]
return decode_strs
def bpe_decode(self, sequences):
"""
Convert a list of lists of bpe token ids into a list of strings by calling bpe tokenizer.
Args:
sequences (`torch.Tensor`):
List of tokenized input ids.
Returns:
`List[str]`: The list of bpe decoded sentences.
"""
return self.bpe_tokenizer.batch_decode(sequences)
def wp_decode(self, sequences):
"""
Convert a list of lists of word piece token ids into a list of strings by calling word piece tokenizer.
Args:
sequences (`torch.Tensor`):
List of tokenized input ids.
Returns:
`List[str]`: The list of wp decoded sentences.
"""
decode_strs = [seq.replace(" ", "") for seq in self.wp_tokenizer.batch_decode(sequences)]
return decode_strs
|
transformers/src/transformers/models/mgp_str/processing_mgp_str.py/0
|
{
"file_path": "transformers/src/transformers/models/mgp_str/processing_mgp_str.py",
"repo_id": "transformers",
"token_count": 4028
}
| 368
|
# coding=utf-8
# Copyright 2020 The HuggingFace Team. 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.
"""MobileBERT model configuration"""
from collections import OrderedDict
from typing import Mapping
from ...configuration_utils import PretrainedConfig
from ...onnx import OnnxConfig
from ...utils import logging
logger = logging.get_logger(__name__)
class MobileBertConfig(PretrainedConfig):
r"""
This is the configuration class to store the configuration of a [`MobileBertModel`] or a [`TFMobileBertModel`]. It
is used to instantiate a MobileBERT model according to the specified arguments, defining the model architecture.
Instantiating a configuration with the defaults will yield a similar configuration to that of the MobileBERT
[google/mobilebert-uncased](https://huggingface.co/google/mobilebert-uncased) architecture.
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PretrainedConfig`] for more information.
Args:
vocab_size (`int`, *optional*, defaults to 30522):
Vocabulary size of the MobileBERT model. Defines the number of different tokens that can be represented by
the `inputs_ids` passed when calling [`MobileBertModel`] or [`TFMobileBertModel`].
hidden_size (`int`, *optional*, defaults to 512):
Dimensionality of the encoder layers and the pooler layer.
num_hidden_layers (`int`, *optional*, defaults to 24):
Number of hidden layers in the Transformer encoder.
num_attention_heads (`int`, *optional*, defaults to 4):
Number of attention heads for each attention layer in the Transformer encoder.
intermediate_size (`int`, *optional*, defaults to 512):
Dimensionality of the "intermediate" (often named feed-forward) layer in the Transformer encoder.
hidden_act (`str` or `function`, *optional*, defaults to `"relu"`):
The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
`"relu"`, `"silu"` and `"gelu_new"` are supported.
hidden_dropout_prob (`float`, *optional*, defaults to 0.0):
The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
attention_probs_dropout_prob (`float`, *optional*, defaults to 0.1):
The dropout ratio for the attention probabilities.
max_position_embeddings (`int`, *optional*, defaults to 512):
The maximum sequence length that this model might ever be used with. Typically set this to something large
just in case (e.g., 512 or 1024 or 2048).
type_vocab_size (`int`, *optional*, defaults to 2):
The vocabulary size of the `token_type_ids` passed when calling [`MobileBertModel`] or
[`TFMobileBertModel`].
initializer_range (`float`, *optional*, defaults to 0.02):
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
layer_norm_eps (`float`, *optional*, defaults to 1e-12):
The epsilon used by the layer normalization layers.
pad_token_id (`int`, *optional*, defaults to 0):
The ID of the token in the word embedding to use as padding.
embedding_size (`int`, *optional*, defaults to 128):
The dimension of the word embedding vectors.
trigram_input (`bool`, *optional*, defaults to `True`):
Use a convolution of trigram as input.
use_bottleneck (`bool`, *optional*, defaults to `True`):
Whether to use bottleneck in BERT.
intra_bottleneck_size (`int`, *optional*, defaults to 128):
Size of bottleneck layer output.
use_bottleneck_attention (`bool`, *optional*, defaults to `False`):
Whether to use attention inputs from the bottleneck transformation.
key_query_shared_bottleneck (`bool`, *optional*, defaults to `True`):
Whether to use the same linear transformation for query&key in the bottleneck.
num_feedforward_networks (`int`, *optional*, defaults to 4):
Number of FFNs in a block.
normalization_type (`str`, *optional*, defaults to `"no_norm"`):
The normalization type in MobileBERT.
classifier_dropout (`float`, *optional*):
The dropout ratio for the classification head.
Examples:
```python
>>> from transformers import MobileBertConfig, MobileBertModel
>>> # Initializing a MobileBERT configuration
>>> configuration = MobileBertConfig()
>>> # Initializing a model (with random weights) from the configuration above
>>> model = MobileBertModel(configuration)
>>> # Accessing the model configuration
>>> configuration = model.config
```
"""
model_type = "mobilebert"
def __init__(
self,
vocab_size=30522,
hidden_size=512,
num_hidden_layers=24,
num_attention_heads=4,
intermediate_size=512,
hidden_act="relu",
hidden_dropout_prob=0.0,
attention_probs_dropout_prob=0.1,
max_position_embeddings=512,
type_vocab_size=2,
initializer_range=0.02,
layer_norm_eps=1e-12,
pad_token_id=0,
embedding_size=128,
trigram_input=True,
use_bottleneck=True,
intra_bottleneck_size=128,
use_bottleneck_attention=False,
key_query_shared_bottleneck=True,
num_feedforward_networks=4,
normalization_type="no_norm",
classifier_activation=True,
classifier_dropout=None,
**kwargs,
):
super().__init__(pad_token_id=pad_token_id, **kwargs)
self.vocab_size = vocab_size
self.hidden_size = hidden_size
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
self.hidden_act = hidden_act
self.intermediate_size = intermediate_size
self.hidden_dropout_prob = hidden_dropout_prob
self.attention_probs_dropout_prob = attention_probs_dropout_prob
self.max_position_embeddings = max_position_embeddings
self.type_vocab_size = type_vocab_size
self.initializer_range = initializer_range
self.layer_norm_eps = layer_norm_eps
self.embedding_size = embedding_size
self.trigram_input = trigram_input
self.use_bottleneck = use_bottleneck
self.intra_bottleneck_size = intra_bottleneck_size
self.use_bottleneck_attention = use_bottleneck_attention
self.key_query_shared_bottleneck = key_query_shared_bottleneck
self.num_feedforward_networks = num_feedforward_networks
self.normalization_type = normalization_type
self.classifier_activation = classifier_activation
if self.use_bottleneck:
self.true_hidden_size = intra_bottleneck_size
else:
self.true_hidden_size = hidden_size
self.classifier_dropout = classifier_dropout
# Copied from transformers.models.bert.configuration_bert.BertOnnxConfig with Bert->MobileBert
class MobileBertOnnxConfig(OnnxConfig):
@property
def inputs(self) -> Mapping[str, Mapping[int, str]]:
if self.task == "multiple-choice":
dynamic_axis = {0: "batch", 1: "choice", 2: "sequence"}
else:
dynamic_axis = {0: "batch", 1: "sequence"}
return OrderedDict(
[
("input_ids", dynamic_axis),
("attention_mask", dynamic_axis),
("token_type_ids", dynamic_axis),
]
)
|
transformers/src/transformers/models/mobilebert/configuration_mobilebert.py/0
|
{
"file_path": "transformers/src/transformers/models/mobilebert/configuration_mobilebert.py",
"repo_id": "transformers",
"token_count": 3121
}
| 369
|
# coding=utf-8
# Copyright 2018 The HuggingFace Inc. team, Microsoft Corporation.
# 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.
"""TF 2.0 MPNet model."""
from __future__ import annotations
import math
import warnings
from typing import Optional, Tuple, Union
import numpy as np
import tensorflow as tf
from ...activations_tf import get_tf_activation
from ...modeling_tf_outputs import (
TFBaseModelOutput,
TFBaseModelOutputWithPooling,
TFMaskedLMOutput,
TFMultipleChoiceModelOutput,
TFQuestionAnsweringModelOutput,
TFSequenceClassifierOutput,
TFTokenClassifierOutput,
)
from ...modeling_tf_utils import (
TFMaskedLanguageModelingLoss,
TFModelInputType,
TFMultipleChoiceLoss,
TFPreTrainedModel,
TFQuestionAnsweringLoss,
TFSequenceClassificationLoss,
TFTokenClassificationLoss,
get_initializer,
keras,
keras_serializable,
unpack_inputs,
)
from ...tf_utils import check_embeddings_within_bounds, shape_list, stable_softmax
from ...utils import (
add_code_sample_docstrings,
add_start_docstrings,
add_start_docstrings_to_model_forward,
logging,
)
from .configuration_mpnet import MPNetConfig
logger = logging.get_logger(__name__)
_CHECKPOINT_FOR_DOC = "microsoft/mpnet-base"
_CONFIG_FOR_DOC = "MPNetConfig"
class TFMPNetPreTrainedModel(TFPreTrainedModel):
"""
An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
models.
"""
config_class = MPNetConfig
base_model_prefix = "mpnet"
class TFMPNetEmbeddings(keras.layers.Layer):
"""Construct the embeddings from word, position embeddings."""
def __init__(self, config, **kwargs):
super().__init__(**kwargs)
self.padding_idx = 1
self.config = config
self.hidden_size = config.hidden_size
self.max_position_embeddings = config.max_position_embeddings
self.initializer_range = config.initializer_range
self.LayerNorm = keras.layers.LayerNormalization(epsilon=config.layer_norm_eps, name="LayerNorm")
self.dropout = keras.layers.Dropout(rate=config.hidden_dropout_prob)
def build(self, input_shape=None):
with tf.name_scope("word_embeddings"):
self.weight = self.add_weight(
name="weight",
shape=[self.config.vocab_size, self.hidden_size],
initializer=get_initializer(initializer_range=self.initializer_range),
)
with tf.name_scope("position_embeddings"):
self.position_embeddings = self.add_weight(
name="embeddings",
shape=[self.max_position_embeddings, self.hidden_size],
initializer=get_initializer(initializer_range=self.initializer_range),
)
if self.built:
return
self.built = True
if getattr(self, "LayerNorm", None) is not None:
with tf.name_scope(self.LayerNorm.name):
self.LayerNorm.build([None, None, self.config.hidden_size])
def create_position_ids_from_input_ids(self, input_ids):
"""
Replace non-padding symbols with their position numbers. Position numbers begin at padding_idx+1. Padding
symbols are ignored. This is modified from fairseq's `utils.make_positions`.
Args:
input_ids: tf.Tensor
Returns: tf.Tensor
"""
mask = tf.cast(tf.math.not_equal(input_ids, self.padding_idx), dtype=input_ids.dtype)
incremental_indices = tf.math.cumsum(mask, axis=1) * mask
return incremental_indices + self.padding_idx
def call(self, input_ids=None, position_ids=None, inputs_embeds=None, training=False):
"""
Applies embedding based on inputs tensor.
Returns:
final_embeddings (`tf.Tensor`): output embedding tensor.
"""
assert not (input_ids is None and inputs_embeds is None)
if input_ids is not None:
check_embeddings_within_bounds(input_ids, self.config.vocab_size)
inputs_embeds = tf.gather(params=self.weight, indices=input_ids)
input_shape = shape_list(inputs_embeds)[:-1]
if position_ids is None:
if input_ids is not None:
# Create the position ids from the input token ids. Any padded tokens remain padded.
position_ids = self.create_position_ids_from_input_ids(input_ids=input_ids)
else:
position_ids = tf.expand_dims(
tf.range(start=self.padding_idx + 1, limit=input_shape[-1] + self.padding_idx + 1), axis=0
)
position_embeds = tf.gather(params=self.position_embeddings, indices=position_ids)
final_embeddings = inputs_embeds + position_embeds
final_embeddings = self.LayerNorm(inputs=final_embeddings)
final_embeddings = self.dropout(inputs=final_embeddings, training=training)
return final_embeddings
# Copied from transformers.models.bert.modeling_tf_bert.TFBertPooler with Bert->MPNet
class TFMPNetPooler(keras.layers.Layer):
def __init__(self, config: MPNetConfig, **kwargs):
super().__init__(**kwargs)
self.dense = keras.layers.Dense(
units=config.hidden_size,
kernel_initializer=get_initializer(config.initializer_range),
activation="tanh",
name="dense",
)
self.config = config
def call(self, hidden_states: tf.Tensor) -> tf.Tensor:
# We "pool" the model by simply taking the hidden state corresponding
# to the first token.
first_token_tensor = hidden_states[:, 0]
pooled_output = self.dense(inputs=first_token_tensor)
return pooled_output
def build(self, input_shape=None):
if self.built:
return
self.built = True
if getattr(self, "dense", None) is not None:
with tf.name_scope(self.dense.name):
self.dense.build([None, None, self.config.hidden_size])
class TFMPNetSelfAttention(keras.layers.Layer):
def __init__(self, config, **kwargs):
super().__init__(**kwargs)
if config.hidden_size % config.num_attention_heads != 0:
raise ValueError(
f"The hidden size ({config.hidden_size}) is not a multiple of the number of attention "
f"heads ({config.num_attention_heads}"
)
self.num_attention_heads = config.num_attention_heads
assert config.hidden_size % config.num_attention_heads == 0
self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
self.all_head_size = self.num_attention_heads * self.attention_head_size
self.q = keras.layers.Dense(
self.all_head_size, kernel_initializer=get_initializer(config.initializer_range), name="q"
)
self.k = keras.layers.Dense(
self.all_head_size, kernel_initializer=get_initializer(config.initializer_range), name="k"
)
self.v = keras.layers.Dense(
self.all_head_size, kernel_initializer=get_initializer(config.initializer_range), name="v"
)
self.o = keras.layers.Dense(
config.hidden_size, kernel_initializer=get_initializer(config.initializer_range), name="o"
)
self.dropout = keras.layers.Dropout(config.attention_probs_dropout_prob)
self.config = config
def transpose_for_scores(self, x, batch_size):
# Reshape from [batch_size, seq_length, all_head_size] to [batch_size, seq_length, num_attention_heads, attention_head_size]
x = tf.reshape(x, (batch_size, -1, self.num_attention_heads, self.attention_head_size))
return tf.transpose(x, perm=[0, 2, 1, 3])
def call(self, hidden_states, attention_mask, head_mask, output_attentions, position_bias=None, training=False):
batch_size = shape_list(hidden_states)[0]
q = self.q(hidden_states)
k = self.k(hidden_states)
v = self.v(hidden_states)
q = self.transpose_for_scores(q, batch_size)
k = self.transpose_for_scores(k, batch_size)
v = self.transpose_for_scores(v, batch_size)
attention_scores = tf.matmul(q, k, transpose_b=True)
dk = tf.cast(shape_list(k)[-1], attention_scores.dtype)
attention_scores = attention_scores / tf.math.sqrt(dk)
# Apply relative position embedding (precomputed in MPNetEncoder) if provided.
if position_bias is not None:
attention_scores += position_bias
if attention_mask is not None:
attention_scores = attention_scores + attention_mask
attention_probs = stable_softmax(attention_scores, axis=-1)
attention_probs = self.dropout(attention_probs, training=training)
if head_mask is not None:
attention_probs = attention_probs * head_mask
c = tf.matmul(attention_probs, v)
c = tf.transpose(c, perm=[0, 2, 1, 3])
c = tf.reshape(c, (batch_size, -1, self.all_head_size))
o = self.o(c)
outputs = (o, attention_probs) if output_attentions else (o,)
return outputs
def build(self, input_shape=None):
if self.built:
return
self.built = True
if getattr(self, "q", None) is not None:
with tf.name_scope(self.q.name):
self.q.build([None, None, self.config.hidden_size])
if getattr(self, "k", None) is not None:
with tf.name_scope(self.k.name):
self.k.build([None, None, self.config.hidden_size])
if getattr(self, "v", None) is not None:
with tf.name_scope(self.v.name):
self.v.build([None, None, self.config.hidden_size])
if getattr(self, "o", None) is not None:
with tf.name_scope(self.o.name):
self.o.build([None, None, self.config.hidden_size])
class TFMPNetAttention(keras.layers.Layer):
def __init__(self, config, **kwargs):
super().__init__(**kwargs)
self.attn = TFMPNetSelfAttention(config, name="attn")
self.LayerNorm = keras.layers.LayerNormalization(epsilon=config.layer_norm_eps, name="LayerNorm")
self.dropout = keras.layers.Dropout(config.hidden_dropout_prob)
self.config = config
def prune_heads(self, heads):
raise NotImplementedError
def call(self, input_tensor, attention_mask, head_mask, output_attentions, position_bias=None, training=False):
self_outputs = self.attn(
input_tensor, attention_mask, head_mask, output_attentions, position_bias=position_bias, training=training
)
attention_output = self.LayerNorm(self.dropout(self_outputs[0]) + input_tensor)
outputs = (attention_output,) + self_outputs[1:] # add attentions if we output them
return outputs
def build(self, input_shape=None):
if self.built:
return
self.built = True
if getattr(self, "attn", None) is not None:
with tf.name_scope(self.attn.name):
self.attn.build(None)
if getattr(self, "LayerNorm", None) is not None:
with tf.name_scope(self.LayerNorm.name):
self.LayerNorm.build([None, None, self.config.hidden_size])
# Copied from transformers.models.bert.modeling_tf_bert.TFBertIntermediate with Bert->MPNet
class TFMPNetIntermediate(keras.layers.Layer):
def __init__(self, config: MPNetConfig, **kwargs):
super().__init__(**kwargs)
self.dense = keras.layers.Dense(
units=config.intermediate_size, kernel_initializer=get_initializer(config.initializer_range), name="dense"
)
if isinstance(config.hidden_act, str):
self.intermediate_act_fn = get_tf_activation(config.hidden_act)
else:
self.intermediate_act_fn = config.hidden_act
self.config = config
def call(self, hidden_states: tf.Tensor) -> tf.Tensor:
hidden_states = self.dense(inputs=hidden_states)
hidden_states = self.intermediate_act_fn(hidden_states)
return hidden_states
def build(self, input_shape=None):
if self.built:
return
self.built = True
if getattr(self, "dense", None) is not None:
with tf.name_scope(self.dense.name):
self.dense.build([None, None, self.config.hidden_size])
# Copied from transformers.models.bert.modeling_tf_bert.TFBertOutput with Bert->MPNet
class TFMPNetOutput(keras.layers.Layer):
def __init__(self, config: MPNetConfig, **kwargs):
super().__init__(**kwargs)
self.dense = keras.layers.Dense(
units=config.hidden_size, kernel_initializer=get_initializer(config.initializer_range), name="dense"
)
self.LayerNorm = keras.layers.LayerNormalization(epsilon=config.layer_norm_eps, name="LayerNorm")
self.dropout = keras.layers.Dropout(rate=config.hidden_dropout_prob)
self.config = config
def call(self, hidden_states: tf.Tensor, input_tensor: tf.Tensor, training: bool = False) -> tf.Tensor:
hidden_states = self.dense(inputs=hidden_states)
hidden_states = self.dropout(inputs=hidden_states, training=training)
hidden_states = self.LayerNorm(inputs=hidden_states + input_tensor)
return hidden_states
def build(self, input_shape=None):
if self.built:
return
self.built = True
if getattr(self, "dense", None) is not None:
with tf.name_scope(self.dense.name):
self.dense.build([None, None, self.config.intermediate_size])
if getattr(self, "LayerNorm", None) is not None:
with tf.name_scope(self.LayerNorm.name):
self.LayerNorm.build([None, None, self.config.hidden_size])
class TFMPNetLayer(keras.layers.Layer):
def __init__(self, config, **kwargs):
super().__init__(**kwargs)
self.attention = TFMPNetAttention(config, name="attention")
self.intermediate = TFMPNetIntermediate(config, name="intermediate")
self.out = TFMPNetOutput(config, name="output")
def call(self, hidden_states, attention_mask, head_mask, output_attentions, position_bias=None, training=False):
self_attention_outputs = self.attention(
hidden_states, attention_mask, head_mask, output_attentions, position_bias=position_bias, training=training
)
attention_output = self_attention_outputs[0]
outputs = self_attention_outputs[1:] # add self attentions if we output attention weights
intermediate_output = self.intermediate(attention_output)
layer_output = self.out(intermediate_output, attention_output, training=training)
outputs = (layer_output,) + outputs # add attentions if we output them
return outputs
def build(self, input_shape=None):
if self.built:
return
self.built = True
if getattr(self, "attention", None) is not None:
with tf.name_scope(self.attention.name):
self.attention.build(None)
if getattr(self, "intermediate", None) is not None:
with tf.name_scope(self.intermediate.name):
self.intermediate.build(None)
if getattr(self, "out", None) is not None:
with tf.name_scope(self.out.name):
self.out.build(None)
class TFMPNetEncoder(keras.layers.Layer):
def __init__(self, config, **kwargs):
super().__init__(**kwargs)
self.config = config
self.n_heads = config.num_attention_heads
self.output_attentions = config.output_attentions
self.output_hidden_states = config.output_hidden_states
self.relative_attention_num_buckets = config.relative_attention_num_buckets
self.initializer_range = config.initializer_range
self.layer = [TFMPNetLayer(config, name=f"layer_._{i}") for i in range(config.num_hidden_layers)]
self.relative_attention_num_buckets = config.relative_attention_num_buckets
def build(self, input_shape=None):
if self.built:
return
self.built = True
with tf.name_scope("relative_attention_bias"):
self.relative_attention_bias = self.add_weight(
name="embeddings",
shape=[self.relative_attention_num_buckets, self.n_heads],
initializer=get_initializer(self.initializer_range),
)
if getattr(self, "layer", None) is not None:
for layer in self.layer:
with tf.name_scope(layer.name):
layer.build(None)
def call(
self,
hidden_states,
attention_mask,
head_mask,
output_attentions,
output_hidden_states,
return_dict,
training=False,
):
position_bias = self.compute_position_bias(hidden_states)
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_outputs = layer_module(
hidden_states,
attention_mask,
head_mask[i],
output_attentions,
position_bias=position_bias,
training=training,
)
hidden_states = layer_outputs[0]
if output_attentions:
all_attentions = all_attentions + (layer_outputs[1],)
# Add last layer
if output_hidden_states:
all_hidden_states = all_hidden_states + (hidden_states,)
if not return_dict:
return tuple(v for v in [hidden_states, all_hidden_states, all_attentions] if v is not None)
return TFBaseModelOutput(
last_hidden_state=hidden_states, hidden_states=all_hidden_states, attentions=all_attentions
)
@staticmethod
def _relative_position_bucket(relative_position, num_buckets=32, max_distance=128):
ret = 0
n = -relative_position
num_buckets //= 2
ret += tf.cast(tf.math.less(n, 0), dtype=relative_position.dtype) * num_buckets
n = tf.math.abs(n)
# now n is in the range [0, inf)
max_exact = num_buckets // 2
is_small = tf.math.less(n, max_exact)
val_if_large = max_exact + tf.cast(
tf.math.log(n / max_exact) / math.log(max_distance / max_exact) * (num_buckets - max_exact),
dtype=relative_position.dtype,
)
val_if_large = tf.math.minimum(val_if_large, num_buckets - 1)
ret += tf.where(is_small, n, val_if_large)
return ret
def compute_position_bias(self, x, position_ids=None):
"""Compute binned relative position bias"""
input_shape = shape_list(x)
qlen, klen = input_shape[1], input_shape[1]
if position_ids is not None:
context_position = position_ids[:, :, None]
memory_position = position_ids[:, None, :]
else:
context_position = tf.range(qlen)[:, None]
memory_position = tf.range(klen)[None, :]
relative_position = memory_position - context_position # shape (qlen, klen)
rp_bucket = self._relative_position_bucket(
relative_position,
num_buckets=self.relative_attention_num_buckets,
)
values = tf.gather(self.relative_attention_bias, rp_bucket) # shape (qlen, klen, num_heads)
values = tf.expand_dims(tf.transpose(values, [2, 0, 1]), axis=0) # shape (1, num_heads, qlen, klen)
return values
@keras_serializable
class TFMPNetMainLayer(keras.layers.Layer):
config_class = MPNetConfig
def __init__(self, config, **kwargs):
super().__init__(**kwargs)
self.config = config
self.num_hidden_layers = config.num_hidden_layers
self.initializer_range = config.initializer_range
self.output_attentions = config.output_attentions
self.output_hidden_states = config.output_hidden_states
self.return_dict = config.use_return_dict
self.encoder = TFMPNetEncoder(config, name="encoder")
self.pooler = TFMPNetPooler(config, name="pooler")
# The embeddings must be the last declaration in order to follow the weights order
self.embeddings = TFMPNetEmbeddings(config, name="embeddings")
# Copied from transformers.models.bert.modeling_tf_bert.TFBertMainLayer.get_input_embeddings
def get_input_embeddings(self) -> keras.layers.Layer:
return self.embeddings
# Copied from transformers.models.bert.modeling_tf_bert.TFBertMainLayer.set_input_embeddings
def set_input_embeddings(self, value: tf.Variable):
self.embeddings.weight = value
self.embeddings.vocab_size = shape_list(value)[0]
# Copied from transformers.models.bert.modeling_tf_bert.TFBertMainLayer._prune_heads
def _prune_heads(self, heads_to_prune):
"""
Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
class PreTrainedModel
"""
raise NotImplementedError
@unpack_inputs
def call(
self,
input_ids=None,
attention_mask=None,
position_ids=None,
head_mask=None,
inputs_embeds=None,
output_attentions=None,
output_hidden_states=None,
return_dict=None,
training=False,
):
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:
input_shape = shape_list(input_ids)
elif inputs_embeds is not None:
input_shape = shape_list(inputs_embeds)[:-1]
else:
raise ValueError("You have to specify either input_ids or inputs_embeds")
if attention_mask is None:
attention_mask = tf.fill(input_shape, 1)
embedding_output = self.embeddings(
input_ids,
position_ids,
inputs_embeds,
training=training,
)
# We create a 3D attention mask from a 2D tensor mask.
# Sizes are [batch_size, 1, 1, to_seq_length]
# So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
# this attention mask is more simple than the triangular masking of causal attention
# used in OpenAI GPT, we just need to prepare the broadcast dimension here.
extended_attention_mask = tf.reshape(attention_mask, (input_shape[0], 1, 1, input_shape[1]))
# Since attention_mask is 1.0 for positions we want to attend and 0.0 for
# masked positions, this operation will create a tensor which is 0.0 for
# positions we want to attend and -10000.0 for masked positions.
# Since we are adding it to the raw scores before the softmax, this is
# effectively the same as removing these entirely.
extended_attention_mask = tf.cast(extended_attention_mask, embedding_output.dtype)
one_cst = tf.constant(1.0, dtype=embedding_output.dtype)
ten_thousand_cst = tf.constant(-10000.0, dtype=embedding_output.dtype)
extended_attention_mask = tf.multiply(tf.subtract(one_cst, extended_attention_mask), ten_thousand_cst)
# 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]
if head_mask is not None:
raise NotImplementedError
else:
head_mask = [None] * self.num_hidden_layers
encoder_outputs = self.encoder(
embedding_output,
extended_attention_mask,
head_mask,
output_attentions,
output_hidden_states,
return_dict,
training=training,
)
sequence_output = encoder_outputs[0]
pooled_output = self.pooler(sequence_output)
if not return_dict:
return (
sequence_output,
pooled_output,
) + encoder_outputs[1:]
return TFBaseModelOutputWithPooling(
last_hidden_state=sequence_output,
pooler_output=pooled_output,
hidden_states=encoder_outputs.hidden_states,
attentions=encoder_outputs.attentions,
)
def build(self, input_shape=None):
if self.built:
return
self.built = True
if getattr(self, "encoder", None) is not None:
with tf.name_scope(self.encoder.name):
self.encoder.build(None)
if getattr(self, "pooler", None) is not None:
with tf.name_scope(self.pooler.name):
self.pooler.build(None)
if getattr(self, "embeddings", None) is not None:
with tf.name_scope(self.embeddings.name):
self.embeddings.build(None)
MPNET_START_DOCSTRING = r"""
This model inherits from [`TFPreTrainedModel`]. Check the superclass documentation for the generic methods the
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
etc.)
This model is also a [keras.Model](https://www.tensorflow.org/api_docs/python/tf/keras/Model) subclass. Use it
as a regular TF 2.0 Keras Model and refer to the TF 2.0 documentation for all matter related to general usage and
behavior.
<Tip>
TensorFlow models and layers in `transformers` accept two formats as input:
- having all inputs as keyword arguments (like PyTorch models), or
- having all inputs as a list, tuple or dict in the first positional argument.
The reason the second format is supported is that Keras methods prefer this format when passing inputs to models
and layers. Because of this support, when using methods like `model.fit()` things should "just work" for you - just
pass your inputs and labels in any format that `model.fit()` supports! If, however, you want to use the second
format outside of Keras methods like `fit()` and `predict()`, such as when creating your own layers or models with
the Keras `Functional` API, there are three possibilities you can use to gather all the input Tensors in the first
positional argument:
- a single Tensor with `input_ids` only and nothing else: `model(input_ids)`
- a list of varying length with one or several input Tensors IN THE ORDER given in the docstring:
`model([input_ids, attention_mask])` or `model([input_ids, attention_mask, token_type_ids])`
- a dictionary with one or several input Tensors associated to the input names given in the docstring:
`model({"input_ids": input_ids, "token_type_ids": token_type_ids})`
Note that when creating models and layers with
[subclassing](https://keras.io/guides/making_new_layers_and_models_via_subclassing/) then you don't need to worry
about any of this, as you can just pass inputs like you would to any other Python function!
</Tip>
Args:
config ([`MPNetConfig`]): Model configuration class with all the parameters of the model.
Initializing with a config file does not load the weights associated with the model, only the
configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
"""
MPNET_INPUTS_DOCSTRING = r"""
Args:
input_ids (`Numpy array` or `tf.Tensor` of shape `({0})`):
Indices of input sequence tokens in the vocabulary.
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.__call__`] and
[`PreTrainedTokenizer.encode`] for details.
[What are input IDs?](../glossary#input-ids)
attention_mask (`Numpy array` or `tf.Tensor` of shape `({0})`, *optional*):
Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
- 1 for tokens that are **not masked**,
- 0 for tokens that are **masked**.
[What are attention masks?](../glossary#attention-mask)
position_ids (`Numpy array` or `tf.Tensor` of shape `({0})`, *optional*):
Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
config.max_position_embeddings - 1]`.
[What are position IDs?](../glossary#position-ids)
head_mask (`Numpy array` or `tf.Tensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:
- 1 indicates the head is **not masked**,
- 0 indicates the head is **masked**.
inputs_embeds (`tf.Tensor` of shape `({0}, hidden_size)`, *optional*):
Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
model's internal embedding lookup matrix.
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
tensors for more detail. This argument can be used only in eager mode, in graph mode the value in the
config will be used instead.
output_hidden_states (`bool`, *optional*):
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
more detail. This argument can be used only in eager mode, in graph mode the value in the config will be
used instead.
return_dict (`bool`, *optional*):
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. This argument can be used in
eager mode, in graph mode the value will always be set to True.
training (`bool`, *optional*, defaults to `False`):
Whether or not to use the model in training mode (some modules like dropout modules have different
behaviors between training and evaluation).
"""
@add_start_docstrings(
"The bare MPNet Model transformer outputting raw hidden-states without any specific head on top.",
MPNET_START_DOCSTRING,
)
class TFMPNetModel(TFMPNetPreTrainedModel):
def __init__(self, config, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs)
self.mpnet = TFMPNetMainLayer(config, name="mpnet")
@unpack_inputs
@add_start_docstrings_to_model_forward(MPNET_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
@add_code_sample_docstrings(
checkpoint=_CHECKPOINT_FOR_DOC,
output_type=TFBaseModelOutput,
config_class=_CONFIG_FOR_DOC,
)
def call(
self,
input_ids: TFModelInputType | None = None,
attention_mask: Optional[Union[np.array, tf.Tensor]] = None,
position_ids: Optional[Union[np.array, tf.Tensor]] = None,
head_mask: Optional[Union[np.array, tf.Tensor]] = None,
inputs_embeds: tf.Tensor | None = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
training: bool = False,
) -> Union[TFBaseModelOutput, Tuple[tf.Tensor]]:
outputs = self.mpnet(
input_ids=input_ids,
attention_mask=attention_mask,
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,
training=training,
)
return outputs
def build(self, input_shape=None):
if self.built:
return
self.built = True
if getattr(self, "mpnet", None) is not None:
with tf.name_scope(self.mpnet.name):
self.mpnet.build(None)
class TFMPNetLMHead(keras.layers.Layer):
"""MPNet head for masked and permuted language modeling"""
def __init__(self, config, input_embeddings, **kwargs):
super().__init__(**kwargs)
self.config = config
self.hidden_size = config.hidden_size
self.dense = keras.layers.Dense(
config.hidden_size, kernel_initializer=get_initializer(config.initializer_range), name="dense"
)
self.layer_norm = keras.layers.LayerNormalization(epsilon=config.layer_norm_eps, name="layer_norm")
self.act = get_tf_activation("gelu")
# The output weights are the same as the input embeddings, but there is
# an output-only bias for each token.
self.decoder = input_embeddings
def build(self, input_shape=None):
self.bias = self.add_weight(shape=(self.config.vocab_size,), initializer="zeros", trainable=True, name="bias")
if self.built:
return
self.built = True
if getattr(self, "dense", None) is not None:
with tf.name_scope(self.dense.name):
self.dense.build([None, None, self.config.hidden_size])
if getattr(self, "layer_norm", None) is not None:
with tf.name_scope(self.layer_norm.name):
self.layer_norm.build([None, None, self.config.hidden_size])
def get_output_embeddings(self):
return self.decoder
def set_output_embeddings(self, value):
self.decoder.weight = value
self.decoder.vocab_size = shape_list(value)[0]
def get_bias(self):
return {"bias": self.bias}
def set_bias(self, value):
self.bias = value["bias"]
self.config.vocab_size = shape_list(value["bias"])[0]
def call(self, hidden_states):
hidden_states = self.dense(hidden_states)
hidden_states = self.act(hidden_states)
hidden_states = self.layer_norm(hidden_states)
# project back to size of vocabulary with bias
seq_length = shape_list(tensor=hidden_states)[1]
hidden_states = tf.reshape(tensor=hidden_states, shape=[-1, self.hidden_size])
hidden_states = tf.matmul(a=hidden_states, b=self.decoder.weight, transpose_b=True)
hidden_states = tf.reshape(tensor=hidden_states, shape=[-1, seq_length, self.config.vocab_size])
hidden_states = tf.nn.bias_add(value=hidden_states, bias=self.bias)
return hidden_states
@add_start_docstrings("""MPNet Model with a `language modeling` head on top.""", MPNET_START_DOCSTRING)
class TFMPNetForMaskedLM(TFMPNetPreTrainedModel, TFMaskedLanguageModelingLoss):
_keys_to_ignore_on_load_missing = [r"pooler"]
def __init__(self, config, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs)
self.mpnet = TFMPNetMainLayer(config, name="mpnet")
self.lm_head = TFMPNetLMHead(config, self.mpnet.embeddings, name="lm_head")
def get_lm_head(self):
return self.lm_head
def get_prefix_bias_name(self):
warnings.warn("The method get_prefix_bias_name is deprecated. Please use `get_bias` instead.", FutureWarning)
return self.name + "/" + self.lm_head.name
@unpack_inputs
@add_start_docstrings_to_model_forward(MPNET_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
@add_code_sample_docstrings(
checkpoint=_CHECKPOINT_FOR_DOC,
output_type=TFMaskedLMOutput,
config_class=_CONFIG_FOR_DOC,
)
def call(
self,
input_ids: TFModelInputType | None = None,
attention_mask: np.ndarray | tf.Tensor | None = None,
position_ids: np.ndarray | tf.Tensor | None = None,
head_mask: np.ndarray | tf.Tensor | None = None,
inputs_embeds: tf.Tensor | None = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
labels: tf.Tensor | None = None,
training: bool = False,
) -> Union[TFMaskedLMOutput, Tuple[tf.Tensor]]:
r"""
labels (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
Labels for computing the masked language modeling loss. Indices should be in `[-100, 0, ...,
config.vocab_size]` (see `input_ids` docstring) Tokens with indices set to `-100` are ignored (masked), the
loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`
"""
outputs = self.mpnet(
input_ids,
attention_mask=attention_mask,
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,
training=training,
)
sequence_output = outputs[0]
prediction_scores = self.lm_head(sequence_output)
loss = None if labels is None else self.hf_compute_loss(labels, prediction_scores)
if not return_dict:
output = (prediction_scores,) + outputs[2:]
return ((loss,) + output) if loss is not None else output
return TFMaskedLMOutput(
loss=loss,
logits=prediction_scores,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
def build(self, input_shape=None):
if self.built:
return
self.built = True
if getattr(self, "mpnet", None) is not None:
with tf.name_scope(self.mpnet.name):
self.mpnet.build(None)
if getattr(self, "lm_head", None) is not None:
with tf.name_scope(self.lm_head.name):
self.lm_head.build(None)
class TFMPNetClassificationHead(keras.layers.Layer):
"""Head for sentence-level classification tasks."""
def __init__(self, config, **kwargs):
super().__init__(**kwargs)
self.dense = keras.layers.Dense(
config.hidden_size,
kernel_initializer=get_initializer(config.initializer_range),
activation="tanh",
name="dense",
)
self.dropout = keras.layers.Dropout(config.hidden_dropout_prob)
self.out_proj = keras.layers.Dense(
config.num_labels, kernel_initializer=get_initializer(config.initializer_range), name="out_proj"
)
self.config = config
def call(self, features, training=False):
x = features[:, 0, :] # take <s> token (equiv. to [CLS])
x = self.dropout(x, training=training)
x = self.dense(x)
x = self.dropout(x, training=training)
x = self.out_proj(x)
return x
def build(self, input_shape=None):
if self.built:
return
self.built = True
if getattr(self, "dense", None) is not None:
with tf.name_scope(self.dense.name):
self.dense.build([None, None, self.config.hidden_size])
if getattr(self, "out_proj", None) is not None:
with tf.name_scope(self.out_proj.name):
self.out_proj.build([None, None, self.config.hidden_size])
@add_start_docstrings(
"""
MPNet Model transformer with a sequence classification/regression head on top (a linear layer on top of the pooled
output) e.g. for GLUE tasks.
""",
MPNET_START_DOCSTRING,
)
class TFMPNetForSequenceClassification(TFMPNetPreTrainedModel, TFSequenceClassificationLoss):
_keys_to_ignore_on_load_missing = [r"pooler"]
def __init__(self, config, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs)
self.num_labels = config.num_labels
self.mpnet = TFMPNetMainLayer(config, name="mpnet")
self.classifier = TFMPNetClassificationHead(config, name="classifier")
@unpack_inputs
@add_start_docstrings_to_model_forward(MPNET_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
@add_code_sample_docstrings(
checkpoint=_CHECKPOINT_FOR_DOC,
output_type=TFSequenceClassifierOutput,
config_class=_CONFIG_FOR_DOC,
)
def call(
self,
input_ids: TFModelInputType | None = None,
attention_mask: Optional[Union[np.array, tf.Tensor]] = None,
position_ids: Optional[Union[np.array, tf.Tensor]] = None,
head_mask: Optional[Union[np.array, tf.Tensor]] = None,
inputs_embeds: tf.Tensor | None = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
labels: tf.Tensor | None = None,
training: bool = False,
) -> Union[TFSequenceClassifierOutput, Tuple[tf.Tensor]]:
r"""
labels (`tf.Tensor` of shape `(batch_size,)`, *optional*):
Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
"""
outputs = self.mpnet(
input_ids,
attention_mask=attention_mask,
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,
training=training,
)
sequence_output = outputs[0]
logits = self.classifier(sequence_output, training=training)
loss = None if labels is None else self.hf_compute_loss(labels, logits)
if not return_dict:
output = (logits,) + outputs[2:]
return ((loss,) + output) if loss is not None else output
return TFSequenceClassifierOutput(
loss=loss,
logits=logits,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
def build(self, input_shape=None):
if self.built:
return
self.built = True
if getattr(self, "mpnet", None) is not None:
with tf.name_scope(self.mpnet.name):
self.mpnet.build(None)
if getattr(self, "classifier", None) is not None:
with tf.name_scope(self.classifier.name):
self.classifier.build(None)
@add_start_docstrings(
"""
MPNet Model with a multiple choice classification head on top (a linear layer on top of the pooled output and a
softmax) e.g. for RocStories/SWAG tasks.
""",
MPNET_START_DOCSTRING,
)
class TFMPNetForMultipleChoice(TFMPNetPreTrainedModel, TFMultipleChoiceLoss):
def __init__(self, config, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs)
self.mpnet = TFMPNetMainLayer(config, name="mpnet")
self.dropout = keras.layers.Dropout(config.hidden_dropout_prob)
self.classifier = keras.layers.Dense(
1, kernel_initializer=get_initializer(config.initializer_range), name="classifier"
)
self.config = config
@unpack_inputs
@add_start_docstrings_to_model_forward(MPNET_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length"))
@add_code_sample_docstrings(
checkpoint=_CHECKPOINT_FOR_DOC,
output_type=TFMultipleChoiceModelOutput,
config_class=_CONFIG_FOR_DOC,
)
def call(
self,
input_ids: TFModelInputType | None = None,
attention_mask: np.ndarray | tf.Tensor | None = None,
position_ids: np.ndarray | tf.Tensor | None = None,
head_mask: np.ndarray | tf.Tensor | None = None,
inputs_embeds: tf.Tensor | None = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
labels: tf.Tensor | None = None,
training: bool = False,
) -> Union[TFMultipleChoiceModelOutput, Tuple[tf.Tensor]]:
r"""
labels (`tf.Tensor` of shape `(batch_size,)`, *optional*):
Labels for computing the multiple choice classification loss. Indices should be in `[0, ..., num_choices]`
where `num_choices` is the size of the second dimension of the input tensors. (See `input_ids` above)
"""
if input_ids is not None:
num_choices = shape_list(input_ids)[1]
seq_length = shape_list(input_ids)[2]
else:
num_choices = shape_list(inputs_embeds)[1]
seq_length = shape_list(inputs_embeds)[2]
flat_input_ids = tf.reshape(input_ids, (-1, seq_length)) if input_ids is not None else None
flat_attention_mask = tf.reshape(attention_mask, (-1, seq_length)) if attention_mask is not None else None
flat_position_ids = tf.reshape(position_ids, (-1, seq_length)) if position_ids is not None else None
flat_inputs_embeds = (
tf.reshape(inputs_embeds, (-1, seq_length, shape_list(inputs_embeds)[3]))
if inputs_embeds is not None
else None
)
outputs = self.mpnet(
flat_input_ids,
flat_attention_mask,
flat_position_ids,
head_mask,
flat_inputs_embeds,
output_attentions,
output_hidden_states,
return_dict=return_dict,
training=training,
)
pooled_output = outputs[1]
pooled_output = self.dropout(pooled_output, training=training)
logits = self.classifier(pooled_output)
reshaped_logits = tf.reshape(logits, (-1, num_choices))
loss = None if labels is None else self.hf_compute_loss(labels, reshaped_logits)
if not return_dict:
output = (reshaped_logits,) + outputs[2:]
return ((loss,) + output) if loss is not None else output
return TFMultipleChoiceModelOutput(
loss=loss,
logits=reshaped_logits,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
def build(self, input_shape=None):
if self.built:
return
self.built = True
if getattr(self, "mpnet", None) is not None:
with tf.name_scope(self.mpnet.name):
self.mpnet.build(None)
if getattr(self, "classifier", None) is not None:
with tf.name_scope(self.classifier.name):
self.classifier.build([None, None, self.config.hidden_size])
@add_start_docstrings(
"""
MPNet Model with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for
Named-Entity-Recognition (NER) tasks.
""",
MPNET_START_DOCSTRING,
)
class TFMPNetForTokenClassification(TFMPNetPreTrainedModel, TFTokenClassificationLoss):
_keys_to_ignore_on_load_missing = [r"pooler"]
def __init__(self, config, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs)
self.num_labels = config.num_labels
self.mpnet = TFMPNetMainLayer(config, name="mpnet")
self.dropout = keras.layers.Dropout(config.hidden_dropout_prob)
self.classifier = keras.layers.Dense(
config.num_labels, kernel_initializer=get_initializer(config.initializer_range), name="classifier"
)
self.config = config
@unpack_inputs
@add_start_docstrings_to_model_forward(MPNET_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
@add_code_sample_docstrings(
checkpoint=_CHECKPOINT_FOR_DOC,
output_type=TFTokenClassifierOutput,
config_class=_CONFIG_FOR_DOC,
)
def call(
self,
input_ids: TFModelInputType | None = None,
attention_mask: np.ndarray | tf.Tensor | None = None,
position_ids: np.ndarray | tf.Tensor | None = None,
head_mask: np.ndarray | tf.Tensor | None = None,
inputs_embeds: tf.Tensor | None = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
labels: tf.Tensor | None = None,
training: bool = False,
) -> Union[TFTokenClassifierOutput, Tuple[tf.Tensor]]:
r"""
labels (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
Labels for computing the token classification loss. Indices should be in `[0, ..., config.num_labels - 1]`.
"""
outputs = self.mpnet(
input_ids=input_ids,
attention_mask=attention_mask,
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,
training=training,
)
sequence_output = outputs[0]
sequence_output = self.dropout(sequence_output, training=training)
logits = self.classifier(sequence_output)
loss = None if labels is None else self.hf_compute_loss(labels, logits)
if not return_dict:
output = (logits,) + outputs[1:]
return ((loss,) + output) if loss is not None else output
return TFTokenClassifierOutput(
loss=loss,
logits=logits,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
def build(self, input_shape=None):
if self.built:
return
self.built = True
if getattr(self, "mpnet", None) is not None:
with tf.name_scope(self.mpnet.name):
self.mpnet.build(None)
if getattr(self, "classifier", None) is not None:
with tf.name_scope(self.classifier.name):
self.classifier.build([None, None, self.config.hidden_size])
@add_start_docstrings(
"""
MPNet Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear
layers on top of the hidden-states output to compute `span start logits` and `span end logits`).
""",
MPNET_START_DOCSTRING,
)
class TFMPNetForQuestionAnswering(TFMPNetPreTrainedModel, TFQuestionAnsweringLoss):
_keys_to_ignore_on_load_missing = [r"pooler"]
def __init__(self, config, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs)
self.num_labels = config.num_labels
self.mpnet = TFMPNetMainLayer(config, name="mpnet")
self.qa_outputs = keras.layers.Dense(
config.num_labels, kernel_initializer=get_initializer(config.initializer_range), name="qa_outputs"
)
self.config = config
@unpack_inputs
@add_start_docstrings_to_model_forward(MPNET_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
@add_code_sample_docstrings(
checkpoint=_CHECKPOINT_FOR_DOC,
output_type=TFQuestionAnsweringModelOutput,
config_class=_CONFIG_FOR_DOC,
)
def call(
self,
input_ids: TFModelInputType | None = None,
attention_mask: Optional[Union[np.array, tf.Tensor]] = None,
position_ids: Optional[Union[np.array, tf.Tensor]] = None,
head_mask: Optional[Union[np.array, tf.Tensor]] = None,
inputs_embeds: tf.Tensor | None = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
start_positions: tf.Tensor | None = None,
end_positions: tf.Tensor | None = None,
training: bool = False,
**kwargs,
) -> Union[TFQuestionAnsweringModelOutput, Tuple[tf.Tensor]]:
r"""
start_positions (`tf.Tensor` of shape `(batch_size,)`, *optional*):
Labels for position (index) of the start of the labelled span for computing the token classification loss.
Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
are not taken into account for computing the loss.
end_positions (`tf.Tensor` of shape `(batch_size,)`, *optional*):
Labels for position (index) of the end of the labelled span for computing the token classification loss.
Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
are not taken into account for computing the loss.
"""
outputs = self.mpnet(
input_ids,
attention_mask=attention_mask,
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,
training=training,
)
sequence_output = outputs[0]
logits = self.qa_outputs(sequence_output)
start_logits, end_logits = tf.split(logits, 2, axis=-1)
start_logits = tf.squeeze(start_logits, axis=-1)
end_logits = tf.squeeze(end_logits, axis=-1)
loss = None
if start_positions is not None and end_positions is not None:
labels = {"start_position": start_positions, "end_position": end_positions}
loss = self.hf_compute_loss(labels, (start_logits, end_logits))
if not return_dict:
output = (start_logits, end_logits) + outputs[2:]
return ((loss,) + output) if loss is not None else output
return TFQuestionAnsweringModelOutput(
loss=loss,
start_logits=start_logits,
end_logits=end_logits,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
def build(self, input_shape=None):
if self.built:
return
self.built = True
if getattr(self, "mpnet", None) is not None:
with tf.name_scope(self.mpnet.name):
self.mpnet.build(None)
if getattr(self, "qa_outputs", None) is not None:
with tf.name_scope(self.qa_outputs.name):
self.qa_outputs.build([None, None, self.config.hidden_size])
|
transformers/src/transformers/models/mpnet/modeling_tf_mpnet.py/0
|
{
"file_path": "transformers/src/transformers/models/mpnet/modeling_tf_mpnet.py",
"repo_id": "transformers",
"token_count": 23798
}
| 370
|
# coding=utf-8
# Copyright 2023 Meta AI and The HuggingFace Inc. team. 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.
"""MusicGen model configuration"""
from ...configuration_utils import PretrainedConfig
from ...utils import logging
from ..auto.configuration_auto import AutoConfig
logger = logging.get_logger(__name__)
class MusicgenDecoderConfig(PretrainedConfig):
r"""
This is the configuration class to store the configuration of an [`MusicgenDecoder`]. It is used to instantiate a
MusicGen decoder according to the specified arguments, defining the model architecture. Instantiating a
configuration with the defaults will yield a similar configuration to that of the MusicGen
[facebook/musicgen-small](https://huggingface.co/facebook/musicgen-small) architecture.
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PretrainedConfig`] for more information.
Args:
vocab_size (`int`, *optional*, defaults to 2048):
Vocabulary size of the MusicgenDecoder model. Defines the number of different tokens that can be
represented by the `inputs_ids` passed when calling [`MusicgenDecoder`].
hidden_size (`int`, *optional*, defaults to 1024):
Dimensionality of the layers and the pooler layer.
num_hidden_layers (`int`, *optional*, defaults to 24):
Number of decoder layers.
num_attention_heads (`int`, *optional*, defaults to 16):
Number of attention heads for each attention layer in the Transformer block.
ffn_dim (`int`, *optional*, defaults to 4096):
Dimensionality of the "intermediate" (often named feed-forward) layer in the Transformer block.
activation_function (`str` or `function`, *optional*, defaults to `"gelu"`):
The non-linear activation function (function or string) in the decoder and pooler. If string, `"gelu"`,
`"relu"`, `"silu"` and `"gelu_new"` are supported.
dropout (`float`, *optional*, defaults to 0.1):
The dropout probability for all fully connected layers in the embeddings, text_encoder, and pooler.
attention_dropout (`float`, *optional*, defaults to 0.0):
The dropout ratio for the attention probabilities.
activation_dropout (`float`, *optional*, defaults to 0.0):
The dropout ratio for activations inside the fully connected layer.
max_position_embeddings (`int`, *optional*, defaults to 2048):
The maximum sequence length that this model might ever be used with. Typically, set this to something large
just in case (e.g., 512 or 1024 or 2048).
initializer_factor (`float`, *optional*, defaults to 0.02):
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
layerdrop (`float`, *optional*, defaults to 0.0):
The LayerDrop probability for the decoder. See the [LayerDrop paper](see https://arxiv.org/abs/1909.11556)
for more details.
scale_embedding (`bool`, *optional*, defaults to `False`):
Scale embeddings by diving by sqrt(hidden_size).
use_cache (`bool`, *optional*, defaults to `True`):
Whether the model should return the last key/values attentions (not used by all models)
num_codebooks (`int`, *optional*, defaults to 4):
The number of parallel codebooks forwarded to the model.
tie_word_embeddings(`bool`, *optional*, defaults to `False`):
Whether input and output word embeddings should be tied.
audio_channels (`int`, *optional*, defaults to 1
Number of channels in the audio data. Either 1 for mono or 2 for stereo. Stereo models generate a separate
audio stream for the left/right output channels. Mono models generate a single audio stream output.
"""
model_type = "musicgen_decoder"
keys_to_ignore_at_inference = ["past_key_values"]
def __init__(
self,
vocab_size=2048,
max_position_embeddings=2048,
num_hidden_layers=24,
ffn_dim=4096,
num_attention_heads=16,
layerdrop=0.0,
use_cache=True,
activation_function="gelu",
hidden_size=1024,
dropout=0.1,
attention_dropout=0.0,
activation_dropout=0.0,
initializer_factor=0.02,
scale_embedding=False,
num_codebooks=4,
audio_channels=1,
pad_token_id=2048,
bos_token_id=2048,
eos_token_id=None,
tie_word_embeddings=False,
**kwargs,
):
self.vocab_size = vocab_size
self.max_position_embeddings = max_position_embeddings
self.hidden_size = hidden_size
self.ffn_dim = ffn_dim
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
self.dropout = dropout
self.attention_dropout = attention_dropout
self.activation_dropout = activation_dropout
self.activation_function = activation_function
self.initializer_factor = initializer_factor
self.layerdrop = layerdrop
self.use_cache = use_cache
self.scale_embedding = scale_embedding # scale factor will be sqrt(d_model) if True
self.num_codebooks = num_codebooks
if audio_channels not in [1, 2]:
raise ValueError(f"Expected 1 (mono) or 2 (stereo) audio channels, got {audio_channels} channels.")
self.audio_channels = audio_channels
super().__init__(
pad_token_id=pad_token_id,
bos_token_id=bos_token_id,
eos_token_id=eos_token_id,
tie_word_embeddings=tie_word_embeddings,
**kwargs,
)
class MusicgenConfig(PretrainedConfig):
r"""
This is the configuration class to store the configuration of a [`MusicgenModel`]. It is used to instantiate a
MusicGen model according to the specified arguments, defining the text encoder, audio encoder and MusicGen decoder
configs.
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PretrainedConfig`] for more information.
Args:
kwargs (*optional*):
Dictionary of keyword arguments. Notably:
- **text_encoder** ([`PretrainedConfig`], *optional*) -- An instance of a configuration object that
defines the text encoder config.
- **audio_encoder** ([`PretrainedConfig`], *optional*) -- An instance of a configuration object that
defines the audio encoder config.
- **decoder** ([`PretrainedConfig`], *optional*) -- An instance of a configuration object that defines
the decoder config.
Example:
```python
>>> from transformers import (
... MusicgenConfig,
... MusicgenDecoderConfig,
... T5Config,
... EncodecConfig,
... MusicgenForConditionalGeneration,
... )
>>> # Initializing text encoder, audio encoder, and decoder model configurations
>>> text_encoder_config = T5Config()
>>> audio_encoder_config = EncodecConfig()
>>> decoder_config = MusicgenDecoderConfig()
>>> configuration = MusicgenConfig.from_sub_models_config(
... text_encoder_config, audio_encoder_config, decoder_config
... )
>>> # Initializing a MusicgenForConditionalGeneration (with random weights) from the facebook/musicgen-small style configuration
>>> model = MusicgenForConditionalGeneration(configuration)
>>> # Accessing the model configuration
>>> configuration = model.config
>>> config_text_encoder = model.config.text_encoder
>>> config_audio_encoder = model.config.audio_encoder
>>> config_decoder = model.config.decoder
>>> # Saving the model, including its configuration
>>> model.save_pretrained("musicgen-model")
>>> # loading model and config from pretrained folder
>>> musicgen_config = MusicgenConfig.from_pretrained("musicgen-model")
>>> model = MusicgenForConditionalGeneration.from_pretrained("musicgen-model", config=musicgen_config)
```"""
model_type = "musicgen"
is_composition = True
def __init__(self, **kwargs):
super().__init__(**kwargs)
if "text_encoder" not in kwargs or "audio_encoder" not in kwargs or "decoder" not in kwargs:
raise ValueError("Config has to be initialized with text_encoder, audio_encoder and decoder config")
text_encoder_config = kwargs.pop("text_encoder")
text_encoder_model_type = text_encoder_config.pop("model_type")
audio_encoder_config = kwargs.pop("audio_encoder")
audio_encoder_model_type = audio_encoder_config.pop("model_type")
decoder_config = kwargs.pop("decoder")
self.text_encoder = AutoConfig.for_model(text_encoder_model_type, **text_encoder_config)
self.audio_encoder = AutoConfig.for_model(audio_encoder_model_type, **audio_encoder_config)
self.decoder = MusicgenDecoderConfig(**decoder_config)
self.is_encoder_decoder = True
@classmethod
def from_sub_models_config(
cls,
text_encoder_config: PretrainedConfig,
audio_encoder_config: PretrainedConfig,
decoder_config: MusicgenDecoderConfig,
**kwargs,
):
r"""
Instantiate a [`MusicgenConfig`] (or a derived class) from text encoder, audio encoder and decoder
configurations.
Returns:
[`MusicgenConfig`]: An instance of a configuration object
"""
return cls(
text_encoder=text_encoder_config.to_dict(),
audio_encoder=audio_encoder_config.to_dict(),
decoder=decoder_config.to_dict(),
**kwargs,
)
@property
# This is a property because you might want to change the codec model on the fly
def sampling_rate(self):
return self.audio_encoder.sampling_rate
@property
def _attn_implementation(self):
# This property is made private for now (as it cannot be changed and a PreTrainedModel.use_attn_implementation method needs to be implemented.)
if hasattr(self, "_attn_implementation_internal"):
if self._attn_implementation_internal is None:
# `config.attn_implementation` should never be None, for backward compatibility.
return "eager"
else:
return self._attn_implementation_internal
else:
return "eager"
@_attn_implementation.setter
def _attn_implementation(self, value):
self._attn_implementation_internal = value
self.decoder._attn_implementation = value
|
transformers/src/transformers/models/musicgen/configuration_musicgen.py/0
|
{
"file_path": "transformers/src/transformers/models/musicgen/configuration_musicgen.py",
"repo_id": "transformers",
"token_count": 4280
}
| 371
|
# coding=utf-8
# Copyright 2024 HuggingFace Inc. team. All rights reserved.
# Copyright (c) 2024, 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.
"""Nemotron model configuration"""
from ...configuration_utils import PretrainedConfig
from ...modeling_rope_utils import rope_config_validation
from ...utils import logging
logger = logging.get_logger(__name__)
class NemotronConfig(PretrainedConfig):
r"""
This is the configuration class to store the configuration of a [`NemotronModel`]. It is used to instantiate an Nemotron
model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
defaults will yield a similar configuration to that of the Nemotron-8B.
e.g. [nvidia/nemotron-3-8b-base-4k-hf](https://huggingface.co/nvidia/nemotron-3-8b-base-4k-hf).
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PretrainedConfig`] for more information.
Args:
vocab_size (`int`, *optional*, defaults to 256000):
Vocabulary size of the Nemotron model. Defines the number of different tokens that can be represented by the
`inputs_ids` passed when calling [`NemotronModel`]
hidden_size (`int`, *optional*, defaults to 6144):
Dimension of the hidden representations.
intermediate_size (`int`, *optional*, defaults to 24576):
Dimension of the MLP representations.
num_hidden_layers (`int`, *optional*, defaults to 32):
Number of hidden layers in the Transformer decoder.
num_attention_heads (`int`, *optional*, defaults to 48):
Number of attention heads for each attention layer in the Transformer decoder.
head_dim (`int`, *optional*):
Projection weights dimension in multi-head attention. Set to hidden_size // num_attention_heads if None
num_key_value_heads (`int`, *optional*):
This is the number of key_value heads that should be used to implement Grouped Query Attention. If
`num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
`num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When
converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
by meanpooling all the original heads within that group. For more details checkout [this
paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to
`num_attention_heads`.
hidden_act (`str` or `function`, *optional*, defaults to `"relu2"`):
The non-linear activation function (function or string) in the decoder.
max_position_embeddings (`int`, *optional*, defaults to 4096):
The maximum sequence length that this model might ever be used with.
initializer_range (`float`, *optional*, defaults to 0.0134):
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
norm_eps (`float`, *optional*, defaults to 1e-05):
The epsilon used by the normalization layers.
use_cache (`bool`, *optional*, defaults to `True`):
Whether or not the model should return the last key/values attentions (not used by all models). Only
relevant if `config.is_decoder=True`.
pad_token_id (`int`, *optional*):
Padding token id.
bos_token_id (`int`, *optional*, defaults to 2):
Beginning of stream token id.
eos_token_id (`int`, *optional*, defaults to 3):
End of stream token id.
tie_word_embeddings (`bool`, *optional*, defaults to `False`):
Whether to tie weight embeddings
rope_theta (`float`, *optional*, defaults to 10000.0):
The base period of the RoPE embeddings.
partial_rotary_factor (`float`, *optional*, defaults to 0.5): Percentage of the query and keys which will have rotary embedding.
attention_bias (`bool`, *optional*, defaults to `False`):
Whether to use a bias in the query, key, value and output projection layers during self-attention.
attention_dropout (`float`, *optional*, defaults to 0.0):
The dropout ratio for the attention probabilities.
mlp_bias (`bool`, *optional*, defaults to `False`):
Whether to use a bias in up_proj and down_proj layers in the MLP layers.
```python
>>> from transformers import NemotronModel, NemotronConfig
>>> # Initializing a Nemotron nemotron-15b style configuration
>>> configuration = NemotronConfig()
>>> # Initializing a model from the nemotron-15b style configuration
>>> model = NemotronModel(configuration)
>>> # Accessing the model configuration
>>> configuration = model.config
```"""
model_type = "nemotron"
keys_to_ignore_at_inference = ["past_key_values"]
def __init__(
self,
vocab_size=256000,
hidden_size=6144,
intermediate_size=24576,
num_hidden_layers=32,
num_attention_heads=48,
head_dim=None,
num_key_value_heads=None,
hidden_act="relu2",
max_position_embeddings=4096,
initializer_range=0.0134,
norm_eps=1e-5,
use_cache=True,
pad_token_id=None,
bos_token_id=2,
eos_token_id=3,
tie_word_embeddings=False,
rope_theta=10000.0,
partial_rotary_factor=0.5,
attention_bias=False,
attention_dropout=0.0,
mlp_bias=False,
**kwargs,
):
self.vocab_size = vocab_size
self.max_position_embeddings = max_position_embeddings
self.hidden_size = hidden_size
self.intermediate_size = intermediate_size
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
self.head_dim = head_dim if head_dim is not None else hidden_size // num_attention_heads
self.num_key_value_heads = num_key_value_heads
self.hidden_act = hidden_act
self.initializer_range = initializer_range
self.norm_eps = norm_eps
self.use_cache = use_cache
self.rope_theta = rope_theta
self.partial_rotary_factor = partial_rotary_factor
rope_config_validation(self)
self.attention_bias = attention_bias
self.attention_dropout = attention_dropout
self.mlp_bias = mlp_bias
super().__init__(
pad_token_id=pad_token_id,
bos_token_id=bos_token_id,
eos_token_id=eos_token_id,
tie_word_embeddings=tie_word_embeddings,
**kwargs,
)
|
transformers/src/transformers/models/nemotron/configuration_nemotron.py/0
|
{
"file_path": "transformers/src/transformers/models/nemotron/configuration_nemotron.py",
"repo_id": "transformers",
"token_count": 2820
}
| 372
|
# coding=utf-8
# Copyright 2022 UW-Madison and The HuggingFace Inc. team. 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.
"""Nystromformer model configuration"""
from ...configuration_utils import PretrainedConfig
from ...utils import logging
logger = logging.get_logger(__name__)
class NystromformerConfig(PretrainedConfig):
r"""
This is the configuration class to store the configuration of a [`NystromformerModel`]. It is used to instantiate
an Nystromformer model according to the specified arguments, defining the model architecture. Instantiating a
configuration with the defaults will yield a similar configuration to that of the Nystromformer
[uw-madison/nystromformer-512](https://huggingface.co/uw-madison/nystromformer-512) architecture.
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PretrainedConfig`] for more information.
Args:
vocab_size (`int`, *optional*, defaults to 30000):
Vocabulary size of the Nystromformer model. Defines the number of different tokens that can be represented
by the `inputs_ids` passed when calling [`NystromformerModel`].
hidden_size (`int`, *optional*, defaults to 768):
Dimension of the encoder layers and the pooler layer.
num_hidden_layers (`int`, *optional*, defaults to 12):
Number of hidden layers in the Transformer encoder.
num_attention_heads (`int`, *optional*, defaults to 12):
Number of attention heads for each attention layer in the Transformer encoder.
intermediate_size (`int`, *optional*, defaults to 3072):
Dimension of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
hidden_act (`str` or `function`, *optional*, defaults to `"gelu"`):
The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
`"relu"`, `"selu"` and `"gelu_new"` are supported.
hidden_dropout_prob (`float`, *optional*, defaults to 0.1):
The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
attention_probs_dropout_prob (`float`, *optional*, defaults to 0.1):
The dropout ratio for the attention probabilities.
max_position_embeddings (`int`, *optional*, defaults to 512):
The maximum sequence length that this model might ever be used with. Typically set this to something large
just in case (e.g., 512 or 1024 or 2048).
type_vocab_size (`int`, *optional*, defaults to 2):
The vocabulary size of the `token_type_ids` passed when calling [`NystromformerModel`].
segment_means_seq_len (`int`, *optional*, defaults to 64):
Sequence length used in segment-means.
num_landmarks (`int`, *optional*, defaults to 64):
The number of landmark (or Nystrom) points to use in Nystrom approximation of the softmax self-attention
matrix.
conv_kernel_size (`int`, *optional*, defaults to 65):
The kernel size of depthwise convolution used in Nystrom approximation.
inv_coeff_init_option (`bool`, *optional*, defaults to `False`):
Whether or not to use exact coefficient computation for the initial values for the iterative method of
calculating the Moore-Penrose inverse of a matrix.
initializer_range (`float`, *optional*, defaults to 0.02):
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
layer_norm_eps (`float`, *optional*, defaults to 1e-12):
The epsilon used by the layer normalization layers.
Example:
```python
>>> from transformers import NystromformerModel, NystromformerConfig
>>> # Initializing a Nystromformer uw-madison/nystromformer-512 style configuration
>>> configuration = NystromformerConfig()
>>> # Initializing a model from the uw-madison/nystromformer-512 style configuration
>>> model = NystromformerModel(configuration)
>>> # Accessing the model configuration
>>> configuration = model.config
```"""
model_type = "nystromformer"
def __init__(
self,
vocab_size=30000,
hidden_size=768,
num_hidden_layers=12,
num_attention_heads=12,
intermediate_size=3072,
hidden_act="gelu_new",
hidden_dropout_prob=0.1,
attention_probs_dropout_prob=0.1,
max_position_embeddings=510,
type_vocab_size=2,
segment_means_seq_len=64,
num_landmarks=64,
conv_kernel_size=65,
inv_coeff_init_option=False,
initializer_range=0.02,
layer_norm_eps=1e-5,
pad_token_id=1,
bos_token_id=0,
eos_token_id=2,
**kwargs,
):
self.vocab_size = vocab_size
self.max_position_embeddings = max_position_embeddings
self.hidden_size = hidden_size
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
self.intermediate_size = intermediate_size
self.hidden_act = hidden_act
self.hidden_dropout_prob = hidden_dropout_prob
self.attention_probs_dropout_prob = attention_probs_dropout_prob
self.initializer_range = initializer_range
self.type_vocab_size = type_vocab_size
self.segment_means_seq_len = segment_means_seq_len
self.num_landmarks = num_landmarks
self.conv_kernel_size = conv_kernel_size
self.inv_coeff_init_option = inv_coeff_init_option
self.layer_norm_eps = layer_norm_eps
super().__init__(pad_token_id=pad_token_id, bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)
|
transformers/src/transformers/models/nystromformer/configuration_nystromformer.py/0
|
{
"file_path": "transformers/src/transformers/models/nystromformer/configuration_nystromformer.py",
"repo_id": "transformers",
"token_count": 2344
}
| 373
|
# coding=utf-8
# Copyright 2018 The OpenAI Team Authors and 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.
"""PyTorch OpenAI GPT model."""
import json
import math
import os
from dataclasses import dataclass
from typing import Any, Dict, Optional, Tuple, Union
import torch
from torch import nn
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
from ...activations import gelu_new, silu
from ...modeling_outputs import BaseModelOutput, CausalLMOutput, SequenceClassifierOutput
from ...modeling_utils import PreTrainedModel, SequenceSummary
from ...pytorch_utils import Conv1D, find_pruneable_heads_and_indices, prune_conv1d_layer
from ...utils import (
ModelOutput,
add_code_sample_docstrings,
add_start_docstrings,
add_start_docstrings_to_model_forward,
logging,
replace_return_docstrings,
)
from .configuration_openai import OpenAIGPTConfig
logger = logging.get_logger(__name__)
_CHECKPOINT_FOR_DOC = "openai-community/openai-gpt"
_CONFIG_FOR_DOC = "OpenAIGPTConfig"
def load_tf_weights_in_openai_gpt(model, config, openai_checkpoint_folder_path):
"""Load tf pre-trained weights in a pytorch model (from NumPy arrays here)"""
import re
import numpy as np
if ".ckpt" in openai_checkpoint_folder_path:
openai_checkpoint_folder_path = os.path.dirname(openai_checkpoint_folder_path)
logger.info(f"Loading weights from {openai_checkpoint_folder_path}")
with open(openai_checkpoint_folder_path + "/parameters_names.json", "r", encoding="utf-8") as names_handle:
names = json.load(names_handle)
with open(openai_checkpoint_folder_path + "/params_shapes.json", "r", encoding="utf-8") as shapes_handle:
shapes = json.load(shapes_handle)
offsets = np.cumsum([np.prod(shape) for shape in shapes])
init_params = [np.load(openai_checkpoint_folder_path + f"/params_{n}.npy") for n in range(10)]
init_params = np.split(np.concatenate(init_params, 0), offsets)[:-1]
init_params = [param.reshape(shape) for param, shape in zip(init_params, shapes)]
# This was used when we had a single embedding matrix for positions and tokens
# init_params[0] = np.concatenate([init_params[1], init_params[0]], 0)
# del init_params[1]
init_params = [arr.squeeze() for arr in init_params]
# Check that the token and position embeddings weight dimensions map those of the init parameters.
if model.tokens_embed.weight.shape != init_params[1].shape:
raise ValueError(
f"tokens_embed.weight.shape: {model.tokens_embed.weight.shape} does not match init_param[1].shape:"
f" {init_params[1].shape}"
)
if model.positions_embed.weight.shape != init_params[0].shape:
raise ValueError(
f"positions_embed.weight.shape: {model.positions_embed.weight.shape} does not match init_param[0].shape:"
f" {init_params[0].shape}"
)
model.tokens_embed.weight.data = torch.from_numpy(init_params[1])
model.positions_embed.weight.data = torch.from_numpy(init_params[0])
names.pop(0)
# Pop position and token embedding arrays
init_params.pop(0)
init_params.pop(0)
for name, array in zip(names, init_params): # names[1:n_transfer], init_params[1:n_transfer]):
name = name[6:] # skip "model/"
if name[-2:] != ":0":
raise ValueError(f"Layer {name} does not end with :0")
name = name[:-2]
name = name.split("/")
pointer = model
for m_name in name:
if re.fullmatch(r"[A-Za-z]+\d+", m_name):
scope_names = re.split(r"(\d+)", m_name)
else:
scope_names = [m_name]
if scope_names[0] == "g":
pointer = getattr(pointer, "weight")
elif scope_names[0] == "b":
pointer = getattr(pointer, "bias")
elif scope_names[0] == "w":
pointer = getattr(pointer, "weight")
else:
pointer = getattr(pointer, scope_names[0])
if len(scope_names) >= 2:
num = int(scope_names[1])
pointer = pointer[num]
# Ensure that the pointer and array have compatible shapes.
if pointer.shape != array.shape:
raise ValueError(f"Pointer shape {pointer.shape} and array shape {array.shape} mismatched")
logger.info(f"Initialize PyTorch weight {name}")
pointer.data = torch.from_numpy(array)
return model
ACT_FNS = {"relu": nn.ReLU(), "silu": silu, "gelu": gelu_new, "swish": silu}
class Attention(nn.Module):
def __init__(self, nx, n_positions, config, scale=False):
super().__init__()
n_state = nx # in Attention: n_state=768 (nx=n_embd)
# [switch nx => n_state from Block to Attention to keep identical to TF implementation]
if n_state % config.n_head != 0:
raise ValueError(f"Attention n_state shape: {n_state} must be divisible by config.n_head {config.n_head}")
self.register_buffer(
"bias",
torch.tril(torch.ones(n_positions, n_positions)).view(1, 1, n_positions, n_positions),
persistent=False,
)
self.n_head = config.n_head
self.split_size = n_state
self.scale = scale
self.c_attn = Conv1D(n_state * 3, nx)
self.c_proj = Conv1D(n_state, nx)
self.attn_dropout = nn.Dropout(config.attn_pdrop)
self.resid_dropout = nn.Dropout(config.resid_pdrop)
self.pruned_heads = set()
def prune_heads(self, heads):
if len(heads) == 0:
return
heads, index = find_pruneable_heads_and_indices(
heads, self.n_head, self.split_size // self.n_head, self.pruned_heads
)
index_attn = torch.cat([index, index + self.split_size, index + (2 * self.split_size)])
# Prune conv1d layers
self.c_attn = prune_conv1d_layer(self.c_attn, index_attn, dim=1)
self.c_proj = prune_conv1d_layer(self.c_proj, index, dim=0)
# Update hyper params
self.split_size = (self.split_size // self.n_head) * (self.n_head - len(heads))
self.n_head = self.n_head - len(heads)
self.pruned_heads = self.pruned_heads.union(heads)
def _attn(self, q, k, v, attention_mask=None, head_mask=None, output_attentions=False):
w = torch.matmul(q, k)
if self.scale:
w = w / math.sqrt(v.size(-1))
# w = w * self.bias + -1e9 * (1 - self.bias) # TF implementation method: mask_attn_weights
# XD: self.b may be larger than w, so we need to crop it
b = self.bias[:, :, : w.size(-2), : w.size(-1)]
w = w * b + -1e4 * (1 - b)
if attention_mask is not None:
# Apply the attention mask
w = w + attention_mask
w = nn.functional.softmax(w, dim=-1)
w = self.attn_dropout(w)
# Mask heads if we want to
if head_mask is not None:
w = w * head_mask
outputs = [torch.matmul(w, v)]
if output_attentions:
outputs.append(w)
return outputs
def merge_heads(self, x):
x = x.permute(0, 2, 1, 3).contiguous()
new_x_shape = x.size()[:-2] + (x.size(-2) * x.size(-1),)
return x.view(*new_x_shape) # in Tensorflow implementation: fct merge_states
def split_heads(self, x, k=False):
new_x_shape = x.size()[:-1] + (self.n_head, x.size(-1) // self.n_head)
x = x.view(*new_x_shape) # in Tensorflow implementation: fct split_states
if k:
return x.permute(0, 2, 3, 1)
else:
return x.permute(0, 2, 1, 3)
def forward(self, x, attention_mask=None, head_mask=None, output_attentions=False):
x = self.c_attn(x)
query, key, value = x.split(self.split_size, dim=2)
query = self.split_heads(query)
key = self.split_heads(key, k=True)
value = self.split_heads(value)
attn_outputs = self._attn(query, key, value, attention_mask, head_mask, output_attentions)
a = attn_outputs[0]
a = self.merge_heads(a)
a = self.c_proj(a)
a = self.resid_dropout(a)
outputs = [a] + attn_outputs[1:]
return outputs # a, (attentions)
class MLP(nn.Module):
def __init__(self, n_state, config): # in MLP: n_state=3072 (4 * n_embd)
super().__init__()
nx = config.n_embd
self.c_fc = Conv1D(n_state, nx)
self.c_proj = Conv1D(nx, n_state)
self.act = ACT_FNS[config.afn]
self.dropout = nn.Dropout(config.resid_pdrop)
def forward(self, x):
h = self.act(self.c_fc(x))
h2 = self.c_proj(h)
return self.dropout(h2)
class Block(nn.Module):
def __init__(self, n_positions, config, scale=False):
super().__init__()
nx = config.n_embd
self.attn = Attention(nx, n_positions, config, scale)
self.ln_1 = nn.LayerNorm(nx, eps=config.layer_norm_epsilon)
self.mlp = MLP(4 * nx, config)
self.ln_2 = nn.LayerNorm(nx, eps=config.layer_norm_epsilon)
def forward(self, x, attention_mask=None, head_mask=None, output_attentions=False):
attn_outputs = self.attn(
x,
attention_mask=attention_mask,
head_mask=head_mask,
output_attentions=output_attentions,
)
a = attn_outputs[0]
n = self.ln_1(x + a)
m = self.mlp(n)
h = self.ln_2(n + m)
outputs = [h] + attn_outputs[1:]
return outputs
class OpenAIGPTPreTrainedModel(PreTrainedModel):
"""
An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
models.
"""
config_class = OpenAIGPTConfig
load_tf_weights = load_tf_weights_in_openai_gpt
base_model_prefix = "transformer"
def _init_weights(self, module):
"""Initialize the weights."""
if isinstance(module, (nn.Linear, Conv1D)):
# Slightly different from the TF version which uses truncated_normal for initialization
# cf https://github.com/pytorch/pytorch/pull/5617
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
if module.bias is not None:
module.bias.data.zero_()
elif isinstance(module, nn.Embedding):
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
if module.padding_idx is not None:
module.weight.data[module.padding_idx].zero_()
elif isinstance(module, nn.LayerNorm):
module.bias.data.zero_()
module.weight.data.fill_(1.0)
@dataclass
class OpenAIGPTDoubleHeadsModelOutput(ModelOutput):
"""
Base class for outputs of models predicting if two sentences are consecutive or not.
Args:
loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
Language modeling loss.
mc_loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `mc_labels` is provided):
Multiple choice classification loss.
logits (`torch.FloatTensor` of shape `(batch_size, num_choices, sequence_length, config.vocab_size)`):
Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
mc_logits (`torch.FloatTensor` of shape `(batch_size, num_choices)`):
Prediction scores of the multiple choice classification head (scores for each choice before SoftMax).
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer) of
shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the model at the output of each layer plus the initial embedding outputs.
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
"""
loss: Optional[torch.FloatTensor] = None
mc_loss: Optional[torch.FloatTensor] = None
logits: torch.FloatTensor = None
mc_logits: torch.FloatTensor = None
hidden_states: Optional[Tuple[torch.FloatTensor]] = None
attentions: Optional[Tuple[torch.FloatTensor]] = None
OPENAI_GPT_START_DOCSTRING = r"""
This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
etc.)
This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
and behavior.
Parameters:
config ([`OpenAIGPTConfig`]): Model configuration class with all the parameters of the model.
Initializing with a config file does not load the weights associated with the model, only the
configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
"""
OPENAI_GPT_INPUTS_DOCSTRING = r"""
Args:
input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
Indices of input sequence tokens in the vocabulary.
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
[`PreTrainedTokenizer.__call__`] for details.
[What are input IDs?](../glossary#input-ids)
attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, *optional*):
Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
- 1 for tokens that are **not masked**,
- 0 for tokens that are **masked**.
[What are attention masks?](../glossary#attention-mask)
token_type_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
Segment token indices to indicate first and second portions of the inputs. Indices are selected in `[0,
1]`:
- 0 corresponds to a *sentence A* token,
- 1 corresponds to a *sentence B* token.
[What are token type IDs?](../glossary#token-type-ids)
position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
config.max_position_embeddings - 1]`.
[What are position IDs?](../glossary#position-ids)
head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:
- 1 indicates the head is **not masked**,
- 0 indicates the head is **masked**.
inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
model's internal embedding lookup matrix.
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
tensors for more detail.
output_hidden_states (`bool`, *optional*):
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
more detail.
return_dict (`bool`, *optional*):
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
"""
@add_start_docstrings(
"The bare OpenAI GPT transformer model outputting raw hidden-states without any specific head on top.",
OPENAI_GPT_START_DOCSTRING,
)
class OpenAIGPTModel(OpenAIGPTPreTrainedModel):
def __init__(self, config):
super().__init__(config)
self.tokens_embed = nn.Embedding(config.vocab_size, config.n_embd)
self.positions_embed = nn.Embedding(config.n_positions, config.n_embd)
self.drop = nn.Dropout(config.embd_pdrop)
self.h = nn.ModuleList([Block(config.n_positions, config, scale=True) for _ in range(config.n_layer)])
self.register_buffer("position_ids", torch.arange(config.n_positions), persistent=False)
# Initialize weights and apply final processing
self.post_init()
def get_input_embeddings(self):
return self.tokens_embed
def set_input_embeddings(self, new_embeddings):
self.tokens_embed = new_embeddings
def _prune_heads(self, heads_to_prune):
"""
Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer}
"""
for layer, heads in heads_to_prune.items():
self.h[layer].attn.prune_heads(heads)
@add_start_docstrings_to_model_forward(OPENAI_GPT_INPUTS_DOCSTRING)
@add_code_sample_docstrings(
checkpoint=_CHECKPOINT_FOR_DOC,
output_type=BaseModelOutput,
config_class=_CONFIG_FOR_DOC,
)
def forward(
self,
input_ids: Optional[torch.LongTensor] = None,
attention_mask: Optional[torch.FloatTensor] = None,
token_type_ids: Optional[torch.LongTensor] = None,
position_ids: Optional[torch.LongTensor] = None,
head_mask: Optional[torch.FloatTensor] = None,
inputs_embeds: Optional[torch.FloatTensor] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
) -> Union[Tuple[torch.Tensor], BaseModelOutput]:
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:
self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask)
input_shape = input_ids.size()
input_ids = input_ids.view(-1, input_shape[-1])
elif inputs_embeds is not None:
input_shape = inputs_embeds.size()[:-1]
else:
raise ValueError("You have to specify either input_ids or inputs_embeds")
if position_ids is None:
# Code is different from when we had a single embedding matrix from position and token embeddings
position_ids = self.position_ids[None, : input_shape[-1]]
# Attention mask.
if attention_mask is not None:
# We create a 3D attention mask from a 2D tensor mask.
# Sizes are [batch_size, 1, 1, to_seq_length]
# So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
# this attention mask is more simple than the triangular masking of causal attention
# used in OpenAI GPT, we just need to prepare the broadcast dimension here.
attention_mask = attention_mask.unsqueeze(1).unsqueeze(2)
# Since attention_mask is 1.0 for positions we want to attend and 0.0 for
# masked positions, this operation will create a tensor which is 0.0 for
# positions we want to attend and the dtype's smallest value for masked positions.
# Since we are adding it to the raw scores before the softmax, this is
# effectively the same as removing these entirely.
attention_mask = attention_mask.to(dtype=next(self.parameters()).dtype) # fp16 compatibility
attention_mask = (1.0 - attention_mask) * torch.finfo(self.dtype).min
# Prepare head mask if needed
head_mask = self.get_head_mask(head_mask, self.config.n_layer)
if inputs_embeds is None:
inputs_embeds = self.tokens_embed(input_ids)
position_embeds = self.positions_embed(position_ids)
if token_type_ids is not None:
token_type_ids = token_type_ids.view(-1, token_type_ids.size(-1))
token_type_embeds = self.tokens_embed(token_type_ids)
else:
token_type_embeds = 0
hidden_states = inputs_embeds + position_embeds + token_type_embeds
hidden_states = self.drop(hidden_states)
output_shape = input_shape + (hidden_states.size(-1),)
all_attentions = () if output_attentions else None
all_hidden_states = () if output_hidden_states else None
for i, block in enumerate(self.h):
if output_hidden_states:
all_hidden_states = all_hidden_states + (hidden_states,)
outputs = block(hidden_states, attention_mask, head_mask[i], output_attentions=output_attentions)
hidden_states = outputs[0]
if output_attentions:
all_attentions = all_attentions + (outputs[1],)
hidden_states = hidden_states.view(*output_shape)
# Add last layer
if output_hidden_states:
all_hidden_states = all_hidden_states + (hidden_states,)
if not return_dict:
return tuple(v for v in [hidden_states, all_hidden_states, all_attentions] if v is not None)
return BaseModelOutput(
last_hidden_state=hidden_states,
hidden_states=all_hidden_states,
attentions=all_attentions,
)
@add_start_docstrings(
"""
OpenAI GPT Model transformer with a language modeling head on top (linear layer with weights tied to the input
embeddings).
""",
OPENAI_GPT_START_DOCSTRING,
)
class OpenAIGPTLMHeadModel(OpenAIGPTPreTrainedModel):
_tied_weights_keys = ["lm_head.weight"]
def __init__(self, config):
super().__init__(config)
self.transformer = OpenAIGPTModel(config)
self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
# Initialize weights and apply final processing
self.post_init()
def get_output_embeddings(self):
return self.lm_head
def set_output_embeddings(self, new_embeddings):
self.lm_head = new_embeddings
@add_start_docstrings_to_model_forward(OPENAI_GPT_INPUTS_DOCSTRING)
@add_code_sample_docstrings(
checkpoint=_CHECKPOINT_FOR_DOC,
output_type=CausalLMOutput,
config_class=_CONFIG_FOR_DOC,
)
def forward(
self,
input_ids: Optional[torch.LongTensor] = None,
attention_mask: Optional[torch.FloatTensor] = None,
token_type_ids: Optional[torch.LongTensor] = None,
position_ids: Optional[torch.LongTensor] = None,
head_mask: Optional[torch.FloatTensor] = None,
inputs_embeds: Optional[torch.FloatTensor] = None,
labels: Optional[torch.LongTensor] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
) -> Union[Tuple[torch.Tensor], CausalLMOutput]:
r"""
labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
`labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100`
are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]`
"""
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
transformer_outputs = self.transformer(
input_ids,
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,
)
hidden_states = transformer_outputs[0]
lm_logits = self.lm_head(hidden_states)
loss = None
if labels is not None:
# Shift so that tokens < n predict n
shift_logits = lm_logits[..., :-1, :].contiguous()
shift_labels = labels[..., 1:].contiguous()
# Flatten the tokens
loss_fct = CrossEntropyLoss()
loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
if not return_dict:
output = (lm_logits,) + transformer_outputs[1:]
return ((loss,) + output) if loss is not None else output
return CausalLMOutput(
loss=loss,
logits=lm_logits,
hidden_states=transformer_outputs.hidden_states,
attentions=transformer_outputs.attentions,
)
def prepare_inputs_for_generation(self, input_ids: torch.LongTensor, **kwargs) -> Dict[str, Any]:
return {"input_ids": input_ids}
@add_start_docstrings(
"""
OpenAI GPT Model transformer with a language modeling and a multiple-choice classification head on top e.g. for
RocStories/SWAG tasks. The two heads are two linear layers. The language modeling head has its weights tied to the
input embeddings, the classification head takes as input the input of a specified classification token index in the
input sequence).
""",
OPENAI_GPT_START_DOCSTRING,
)
class OpenAIGPTDoubleHeadsModel(OpenAIGPTPreTrainedModel):
_tied_weights_keys = ["lm_head.weight"]
def __init__(self, config):
super().__init__(config)
config.num_labels = 1
self.transformer = OpenAIGPTModel(config)
self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
self.multiple_choice_head = SequenceSummary(config)
# Initialize weights and apply final processing
self.post_init()
def get_output_embeddings(self):
return self.lm_head
def set_output_embeddings(self, new_embeddings):
self.lm_head = new_embeddings
@add_start_docstrings_to_model_forward(OPENAI_GPT_INPUTS_DOCSTRING)
@replace_return_docstrings(output_type=OpenAIGPTDoubleHeadsModelOutput, config_class=_CONFIG_FOR_DOC)
def forward(
self,
input_ids: Optional[torch.LongTensor] = None,
attention_mask: Optional[torch.FloatTensor] = None,
token_type_ids: Optional[torch.LongTensor] = None,
position_ids: Optional[torch.LongTensor] = None,
head_mask: Optional[torch.FloatTensor] = None,
inputs_embeds: Optional[torch.FloatTensor] = None,
mc_token_ids: Optional[torch.LongTensor] = None,
labels: Optional[torch.LongTensor] = None,
mc_labels: Optional[torch.LongTensor] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
) -> Union[Tuple[torch.Tensor], OpenAIGPTDoubleHeadsModelOutput]:
r"""
mc_token_ids (`torch.LongTensor` of shape `(batch_size, num_choices)`, *optional*, default to index of the last token of the input):
Index of the classification token in each input sequence. Selected in the range `[0, input_ids.size(-1) -
1]`.
labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
`labels = input_ids` Indices are selected in `[-1, 0, ..., config.vocab_size]` All labels set to `-100` are
ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]`
mc_labels (`torch.LongTensor` of shape `(batch_size)`, *optional*):
Labels for computing the multiple choice classification loss. Indices should be in `[0, ..., num_choices]`
where *num_choices* is the size of the second dimension of the input tensors. (see *input_ids* above)
Return:
Examples:
```python
>>> from transformers import AutoTokenizer, OpenAIGPTDoubleHeadsModel
>>> import torch
>>> tokenizer = AutoTokenizer.from_pretrained("openai-community/openai-gpt")
>>> model = OpenAIGPTDoubleHeadsModel.from_pretrained("openai-community/openai-gpt")
>>> tokenizer.add_special_tokens(
... {"cls_token": "[CLS]"}
... ) # Add a [CLS] to the vocabulary (we should train it also!)
>>> model.resize_token_embeddings(len(tokenizer))
>>> choices = ["Hello, my dog is cute [CLS]", "Hello, my cat is cute [CLS]"]
>>> input_ids = torch.tensor([tokenizer.encode(s) for s in choices]).unsqueeze(0) # Batch size 1, 2 choices
>>> mc_token_ids = torch.tensor([input_ids.size(-1) - 1, input_ids.size(-1) - 1]).unsqueeze(0) # Batch size 1
>>> outputs = model(input_ids, mc_token_ids=mc_token_ids)
>>> lm_logits = outputs.logits
>>> mc_logits = outputs.mc_logits
```"""
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
transformer_outputs = self.transformer(
input_ids,
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,
)
hidden_states = transformer_outputs[0]
lm_logits = self.lm_head(hidden_states)
mc_logits = self.multiple_choice_head(hidden_states, mc_token_ids).squeeze(-1)
lm_loss, mc_loss = None, None
if mc_labels is not None:
loss_fct = CrossEntropyLoss()
mc_loss = loss_fct(mc_logits.view(-1, mc_logits.size(-1)), mc_labels.view(-1))
if labels is not None:
shift_logits = lm_logits[..., :-1, :].contiguous()
shift_labels = labels[..., 1:].contiguous()
loss_fct = CrossEntropyLoss()
lm_loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
if not return_dict:
output = (lm_logits, mc_logits) + transformer_outputs[1:]
if mc_loss is not None:
output = (mc_loss,) + output
return ((lm_loss,) + output) if lm_loss is not None else output
return OpenAIGPTDoubleHeadsModelOutput(
loss=lm_loss,
mc_loss=mc_loss,
logits=lm_logits,
mc_logits=mc_logits,
hidden_states=transformer_outputs.hidden_states,
attentions=transformer_outputs.attentions,
)
@add_start_docstrings(
"""
The Original OpenAI GPT Model transformer with a sequence classification head on top (linear layer).
[`OpenAIGPTForSequenceClassification`] uses the last token in order to do the classification, as other causal
models (e.g. GPT-2) do. Since it does classification on the last token, it requires to know the position of the
last token. If a `pad_token_id` is defined in the configuration, it finds the last token that is not a padding
token in each row. If no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since
it cannot guess the padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take
the last value in each row of the batch).
""",
OPENAI_GPT_START_DOCSTRING,
)
class OpenAIGPTForSequenceClassification(OpenAIGPTPreTrainedModel):
def __init__(self, config):
super().__init__(config)
self.num_labels = config.num_labels
self.transformer = OpenAIGPTModel(config)
self.score = nn.Linear(config.n_embd, self.num_labels, bias=False)
# Initialize weights and apply final processing
self.post_init()
@add_start_docstrings_to_model_forward(OPENAI_GPT_INPUTS_DOCSTRING)
@add_code_sample_docstrings(
checkpoint=_CHECKPOINT_FOR_DOC,
output_type=SequenceClassifierOutput,
config_class=_CONFIG_FOR_DOC,
)
def forward(
self,
input_ids: Optional[torch.LongTensor] = None,
attention_mask: Optional[torch.FloatTensor] = None,
token_type_ids: Optional[torch.LongTensor] = None,
position_ids: Optional[torch.LongTensor] = None,
head_mask: Optional[torch.FloatTensor] = None,
inputs_embeds: Optional[torch.FloatTensor] = None,
labels: Optional[torch.LongTensor] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
) -> Union[Tuple[torch.Tensor], SequenceClassifierOutput]:
r"""
labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
"""
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
transformer_outputs = self.transformer(
input_ids,
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,
)
hidden_states = transformer_outputs[0]
logits = self.score(hidden_states)
if input_ids is not None:
batch_size, sequence_length = input_ids.shape[:2]
else:
batch_size, sequence_length = inputs_embeds.shape[:2]
# Ensure the batch size is > 1 if there is no padding.
if self.config.pad_token_id is None and batch_size != 1:
raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.")
if self.config.pad_token_id is None:
sequence_lengths = -1
else:
if input_ids is not None:
# if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1
sequence_lengths = sequence_lengths % input_ids.shape[-1]
sequence_lengths = sequence_lengths.to(logits.device)
else:
sequence_lengths = -1
logger.warning_once(
f"{self.__class__.__name__} will not detect padding tokens in `inputs_embeds`. Results may be "
"unexpected if using padding tokens in conjunction with `inputs_embeds.`"
)
pooled_logits = logits[range(batch_size), sequence_lengths]
loss = None
if labels is not None:
if self.config.problem_type is None:
if self.num_labels == 1:
self.config.problem_type = "regression"
elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
self.config.problem_type = "single_label_classification"
else:
self.config.problem_type = "multi_label_classification"
if self.config.problem_type == "regression":
loss_fct = MSELoss()
if self.num_labels == 1:
loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
else:
loss = loss_fct(pooled_logits, labels)
elif self.config.problem_type == "single_label_classification":
loss_fct = CrossEntropyLoss()
loss = loss_fct(pooled_logits.view(-1, self.num_labels), labels.view(-1))
elif self.config.problem_type == "multi_label_classification":
loss_fct = BCEWithLogitsLoss()
loss = loss_fct(pooled_logits, labels)
if not return_dict:
output = (pooled_logits,) + transformer_outputs[1:]
return ((loss,) + output) if loss is not None else output
return SequenceClassifierOutput(
loss=loss,
logits=pooled_logits,
hidden_states=transformer_outputs.hidden_states,
attentions=transformer_outputs.attentions,
)
|
transformers/src/transformers/models/openai/modeling_openai.py/0
|
{
"file_path": "transformers/src/transformers/models/openai/modeling_openai.py",
"repo_id": "transformers",
"token_count": 16258
}
| 374
|
# coding=utf-8
# Copyright 2023 The HuggingFace Inc. team. 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.
"""PatchTST model configuration"""
from typing import List, Optional, Union
from transformers.configuration_utils import PretrainedConfig
from transformers.utils import logging
logger = logging.get_logger(__name__)
class PatchTSTConfig(PretrainedConfig):
r"""
This is the configuration class to store the configuration of an [`PatchTSTModel`]. It is used to instantiate an
PatchTST model according to the specified arguments, defining the model architecture.
[ibm/patchtst](https://huggingface.co/ibm/patchtst) architecture.
Configuration objects inherit from [`PretrainedConfig`] can be used to control the model outputs. Read the
documentation from [`PretrainedConfig`] for more information.
Args:
num_input_channels (`int`, *optional*, defaults to 1):
The size of the target variable which by default is 1 for univariate targets. Would be > 1 in case of
multivariate targets.
context_length (`int`, *optional*, defaults to 32):
The context length of the input sequence.
distribution_output (`str`, *optional*, defaults to `"student_t"`):
The distribution emission head for the model when loss is "nll". Could be either "student_t", "normal" or
"negative_binomial".
loss (`str`, *optional*, defaults to `"mse"`):
The loss function for the model corresponding to the `distribution_output` head. For parametric
distributions it is the negative log likelihood ("nll") and for point estimates it is the mean squared
error "mse".
patch_length (`int`, *optional*, defaults to 1):
Define the patch length of the patchification process.
patch_stride (`int`, *optional*, defaults to 1):
Define the stride of the patchification process.
num_hidden_layers (`int`, *optional*, defaults to 3):
Number of hidden layers.
d_model (`int`, *optional*, defaults to 128):
Dimensionality of the transformer layers.
num_attention_heads (`int`, *optional*, defaults to 4):
Number of attention heads for each attention layer in the Transformer encoder.
share_embedding (`bool`, *optional*, defaults to `True`):
Sharing the input embedding across all channels.
channel_attention (`bool`, *optional*, defaults to `False`):
Activate channel attention block in the Transformer to allow channels to attend each other.
ffn_dim (`int`, *optional*, defaults to 512):
Dimension of the "intermediate" (often named feed-forward) layer in the Transformer encoder.
norm_type (`str` , *optional*, defaults to `"batchnorm"`):
Normalization at each Transformer layer. Can be `"batchnorm"` or `"layernorm"`.
norm_eps (`float`, *optional*, defaults to 1e-05):
A value added to the denominator for numerical stability of normalization.
attention_dropout (`float`, *optional*, defaults to 0.0):
The dropout probability for the attention probabilities.
positional_dropout (`float`, *optional*, defaults to 0.0):
The dropout probability in the positional embedding layer.
path_dropout (`float`, *optional*, defaults to 0.0):
The dropout path in the residual block.
ff_dropout (`float`, *optional*, defaults to 0.0):
The dropout probability used between the two layers of the feed-forward networks.
bias (`bool`, *optional*, defaults to `True`):
Whether to add bias in the feed-forward networks.
activation_function (`str`, *optional*, defaults to `"gelu"`):
The non-linear activation function (string) in the Transformer.`"gelu"` and `"relu"` are supported.
pre_norm (`bool`, *optional*, defaults to `True`):
Normalization is applied before self-attention if pre_norm is set to `True`. Otherwise, normalization is
applied after residual block.
positional_encoding_type (`str`, *optional*, defaults to `"sincos"`):
Positional encodings. Options `"random"` and `"sincos"` are supported.
use_cls_token (`bool`, *optional*, defaults to `False`):
Whether cls token is used.
init_std (`float`, *optional*, defaults to 0.02):
The standard deviation of the truncated normal weight initialization distribution.
share_projection (`bool`, *optional*, defaults to `True`):
Sharing the projection layer across different channels in the forecast head.
scaling (`Union`, *optional*, defaults to `"std"`):
Whether to scale the input targets via "mean" scaler, "std" scaler or no scaler if `None`. If `True`, the
scaler is set to "mean".
do_mask_input (`bool`, *optional*):
Apply masking during the pretraining.
mask_type (`str`, *optional*, defaults to `"random"`):
Masking type. Only `"random"` and `"forecast"` are currently supported.
random_mask_ratio (`float`, *optional*, defaults to 0.5):
Masking ratio applied to mask the input data during random pretraining.
num_forecast_mask_patches (`int` or `list`, *optional*, defaults to `[2]`):
Number of patches to be masked at the end of each batch sample. If it is an integer,
all the samples in the batch will have the same number of masked patches. If it is a list,
samples in the batch will be randomly masked by numbers defined in the list. This argument is only used
for forecast pretraining.
channel_consistent_masking (`bool`, *optional*, defaults to `False`):
If channel consistent masking is True, all the channels will have the same masking pattern.
unmasked_channel_indices (`list`, *optional*):
Indices of channels that are not masked during pretraining. Values in the list are number between 1 and
`num_input_channels`
mask_value (`int`, *optional*, defaults to 0):
Values in the masked patches will be filled by `mask_value`.
pooling_type (`str`, *optional*, defaults to `"mean"`):
Pooling of the embedding. `"mean"`, `"max"` and `None` are supported.
head_dropout (`float`, *optional*, defaults to 0.0):
The dropout probability for head.
prediction_length (`int`, *optional*, defaults to 24):
The prediction horizon that the model will output.
num_targets (`int`, *optional*, defaults to 1):
Number of targets for regression and classification tasks. For classification, it is the number of
classes.
output_range (`list`, *optional*):
Output range for regression task. The range of output values can be set to enforce the model to produce
values within a range.
num_parallel_samples (`int`, *optional*, defaults to 100):
The number of samples is generated in parallel for probabilistic prediction.
```python
>>> from transformers import PatchTSTConfig, PatchTSTModel
>>> # Initializing an PatchTST configuration with 12 time steps for prediction
>>> configuration = PatchTSTConfig(prediction_length=12)
>>> # Randomly initializing a model (with random weights) from the configuration
>>> model = PatchTSTModel(configuration)
>>> # Accessing the model configuration
>>> configuration = model.config
```"""
model_type = "patchtst"
attribute_map = {
"hidden_size": "d_model",
"num_attention_heads": "num_attention_heads",
"num_hidden_layers": "num_hidden_layers",
}
def __init__(
self,
# time series specific configuration
num_input_channels: int = 1,
context_length: int = 32,
distribution_output: str = "student_t",
loss: str = "mse",
# PatchTST arguments
patch_length: int = 1,
patch_stride: int = 1,
# Transformer architecture configuration
num_hidden_layers: int = 3,
d_model: int = 128,
num_attention_heads: int = 4,
share_embedding: bool = True,
channel_attention: bool = False,
ffn_dim: int = 512,
norm_type: str = "batchnorm",
norm_eps: float = 1e-05,
attention_dropout: float = 0.0,
positional_dropout: float = 0.0,
path_dropout: float = 0.0,
ff_dropout: float = 0.0,
bias: bool = True,
activation_function: str = "gelu",
pre_norm: bool = True,
positional_encoding_type: str = "sincos",
use_cls_token: bool = False,
init_std: float = 0.02,
share_projection: bool = True,
scaling: Optional[Union[str, bool]] = "std",
# mask pretraining
do_mask_input: Optional[bool] = None,
mask_type: str = "random",
random_mask_ratio: float = 0.5,
num_forecast_mask_patches: Optional[Union[List[int], int]] = [2],
channel_consistent_masking: Optional[bool] = False,
unmasked_channel_indices: Optional[List[int]] = None,
mask_value: int = 0,
# head
pooling_type: str = "mean",
head_dropout: float = 0.0,
prediction_length: int = 24,
num_targets: int = 1,
output_range: Optional[List] = None,
# distribution head
num_parallel_samples: int = 100,
**kwargs,
):
# time series specific configuration
self.context_length = context_length
self.num_input_channels = num_input_channels # n_vars
self.loss = loss
self.distribution_output = distribution_output
self.num_parallel_samples = num_parallel_samples
# Transformer architecture configuration
self.d_model = d_model
self.num_attention_heads = num_attention_heads
self.ffn_dim = ffn_dim
self.num_hidden_layers = num_hidden_layers
self.attention_dropout = attention_dropout
self.share_embedding = share_embedding
self.channel_attention = channel_attention
self.norm_type = norm_type
self.norm_eps = norm_eps
self.positional_dropout = positional_dropout
self.path_dropout = path_dropout
self.ff_dropout = ff_dropout
self.bias = bias
self.activation_function = activation_function
self.pre_norm = pre_norm
self.positional_encoding_type = positional_encoding_type
self.use_cls_token = use_cls_token
self.init_std = init_std
self.scaling = scaling
# PatchTST parameters
self.patch_length = patch_length
self.patch_stride = patch_stride
# Mask pretraining
self.do_mask_input = do_mask_input
self.mask_type = mask_type
self.random_mask_ratio = random_mask_ratio # for random masking
self.num_forecast_mask_patches = num_forecast_mask_patches # for forecast masking
self.channel_consistent_masking = channel_consistent_masking
self.unmasked_channel_indices = unmasked_channel_indices
self.mask_value = mask_value
# general head params
self.pooling_type = pooling_type
self.head_dropout = head_dropout
# For prediction head
self.share_projection = share_projection
self.prediction_length = prediction_length
# For prediction and regression head
self.num_parallel_samples = num_parallel_samples
# Regression
self.num_targets = num_targets
self.output_range = output_range
super().__init__(**kwargs)
|
transformers/src/transformers/models/patchtst/configuration_patchtst.py/0
|
{
"file_path": "transformers/src/transformers/models/patchtst/configuration_patchtst.py",
"repo_id": "transformers",
"token_count": 4664
}
| 375
|
# coding=utf-8
# Copyright (c) 2020, VinAI Research and the HuggingFace Inc. team.
# Copyright 2018 The Open AI Team Authors and The HuggingFace Inc. team.
#
# 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.
"""Tokenization classes for PhoBERT"""
import os
import re
from shutil import copyfile
from typing import List, Optional, Tuple
from ...tokenization_utils import PreTrainedTokenizer
from ...utils import logging
logger = logging.get_logger(__name__)
VOCAB_FILES_NAMES = {
"vocab_file": "vocab.txt",
"merges_file": "bpe.codes",
}
def get_pairs(word):
"""
Return set of symbol pairs in a word.
Word is represented as tuple of symbols (symbols being variable-length strings).
"""
pairs = set()
prev_char = word[0]
for char in word[1:]:
pairs.add((prev_char, char))
prev_char = char
pairs = set(pairs)
return pairs
class PhobertTokenizer(PreTrainedTokenizer):
"""
Construct a PhoBERT tokenizer. Based on Byte-Pair-Encoding.
This tokenizer inherits from [`PreTrainedTokenizer`] which contains most of the main methods. Users should refer to
this superclass for more information regarding those methods.
Args:
vocab_file (`str`):
Path to the vocabulary file.
merges_file (`str`):
Path to the merges file.
bos_token (`st`, *optional*, defaults to `"<s>"`):
The beginning of sequence token that was used during pretraining. Can be used a sequence classifier token.
<Tip>
When building a sequence using special tokens, this is not the token that is used for the beginning of
sequence. The token used is the `cls_token`.
</Tip>
eos_token (`str`, *optional*, defaults to `"</s>"`):
The end of sequence token.
<Tip>
When building a sequence using special tokens, this is not the token that is used for the end of sequence.
The token used is the `sep_token`.
</Tip>
sep_token (`str`, *optional*, defaults to `"</s>"`):
The separator token, which is used when building a sequence from multiple sequences, e.g. two sequences for
sequence classification or for a text and a question for question answering. It is also used as the last
token of a sequence built with special tokens.
cls_token (`str`, *optional*, defaults to `"<s>"`):
The classifier token which is used when doing sequence classification (classification of the whole sequence
instead of per-token classification). It is the first token of the sequence when built with special tokens.
unk_token (`str`, *optional*, defaults to `"<unk>"`):
The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this
token instead.
pad_token (`str`, *optional*, defaults to `"<pad>"`):
The token used for padding, for example when batching sequences of different lengths.
mask_token (`str`, *optional*, defaults to `"<mask>"`):
The token used for masking values. This is the token used when training this model with masked language
modeling. This is the token which the model will try to predict.
"""
vocab_files_names = VOCAB_FILES_NAMES
def __init__(
self,
vocab_file,
merges_file,
bos_token="<s>",
eos_token="</s>",
sep_token="</s>",
cls_token="<s>",
unk_token="<unk>",
pad_token="<pad>",
mask_token="<mask>",
**kwargs,
):
self.vocab_file = vocab_file
self.merges_file = merges_file
self.encoder = {}
self.encoder[str(bos_token)] = 0
self.encoder[str(pad_token)] = 1
self.encoder[str(eos_token)] = 2
self.encoder[str(unk_token)] = 3
self.add_from_file(vocab_file)
self.decoder = {v: k for k, v in self.encoder.items()}
with open(merges_file, encoding="utf-8") as merges_handle:
merges = merges_handle.read().split("\n")[:-1]
merges = [tuple(merge.split()[:-1]) for merge in merges]
self.bpe_ranks = dict(zip(merges, range(len(merges))))
self.cache = {}
super().__init__(
bos_token=bos_token,
eos_token=eos_token,
unk_token=unk_token,
sep_token=sep_token,
cls_token=cls_token,
pad_token=pad_token,
mask_token=mask_token,
**kwargs,
)
def build_inputs_with_special_tokens(
self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
) -> List[int]:
"""
Build model inputs from a sequence or a pair of sequence for sequence classification tasks by concatenating and
adding special tokens. A PhoBERT sequence has the following format:
- single sequence: `<s> X </s>`
- pair of sequences: `<s> A </s></s> B </s>`
Args:
token_ids_0 (`List[int]`):
List of IDs to which the special tokens will be added.
token_ids_1 (`List[int]`, *optional*):
Optional second list of IDs for sequence pairs.
Returns:
`List[int]`: List of [input IDs](../glossary#input-ids) with the appropriate special tokens.
"""
if token_ids_1 is None:
return [self.cls_token_id] + token_ids_0 + [self.sep_token_id]
cls = [self.cls_token_id]
sep = [self.sep_token_id]
return cls + token_ids_0 + sep + sep + token_ids_1 + sep
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]:
"""
Retrieve sequence ids from a token list that has no special tokens added. This method is called when adding
special tokens using the tokenizer `prepare_for_model` method.
Args:
token_ids_0 (`List[int]`):
List of IDs.
token_ids_1 (`List[int]`, *optional*):
Optional second list of IDs for sequence pairs.
already_has_special_tokens (`bool`, *optional*, defaults to `False`):
Whether or not the token list is already formatted with special tokens for the model.
Returns:
`List[int]`: A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
"""
if already_has_special_tokens:
return super().get_special_tokens_mask(
token_ids_0=token_ids_0, token_ids_1=token_ids_1, already_has_special_tokens=True
)
if token_ids_1 is None:
return [1] + ([0] * len(token_ids_0)) + [1]
return [1] + ([0] * len(token_ids_0)) + [1, 1] + ([0] * len(token_ids_1)) + [1]
def create_token_type_ids_from_sequences(
self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
) -> List[int]:
"""
Create a mask from the two sequences passed to be used in a sequence-pair classification task. PhoBERT does not
make use of token type ids, therefore a list of zeros is returned.
Args:
token_ids_0 (`List[int]`):
List of IDs.
token_ids_1 (`List[int]`, *optional*):
Optional second list of IDs for sequence pairs.
Returns:
`List[int]`: List of zeros.
"""
sep = [self.sep_token_id]
cls = [self.cls_token_id]
if token_ids_1 is None:
return len(cls + token_ids_0 + sep) * [0]
return len(cls + token_ids_0 + sep + sep + token_ids_1 + sep) * [0]
@property
def vocab_size(self):
return len(self.encoder)
def get_vocab(self):
return dict(self.encoder, **self.added_tokens_encoder)
def bpe(self, token):
if token in self.cache:
return self.cache[token]
word = tuple(token)
word = tuple(list(word[:-1]) + [word[-1] + "</w>"])
pairs = get_pairs(word)
if not pairs:
return token
while True:
bigram = min(pairs, key=lambda pair: self.bpe_ranks.get(pair, float("inf")))
if bigram not in self.bpe_ranks:
break
first, second = bigram
new_word = []
i = 0
while i < len(word):
try:
j = word.index(first, i)
except ValueError:
new_word.extend(word[i:])
break
else:
new_word.extend(word[i:j])
i = j
if word[i] == first and i < len(word) - 1 and word[i + 1] == second:
new_word.append(first + second)
i += 2
else:
new_word.append(word[i])
i += 1
new_word = tuple(new_word)
word = new_word
if len(word) == 1:
break
else:
pairs = get_pairs(word)
word = "@@ ".join(word)
word = word[:-4]
self.cache[token] = word
return word
def _tokenize(self, text):
"""Tokenize a string."""
split_tokens = []
words = re.findall(r"\S+\n?", text)
for token in words:
split_tokens.extend(list(self.bpe(token).split(" ")))
return split_tokens
def _convert_token_to_id(self, token):
"""Converts a token (str) in an id using the vocab."""
return self.encoder.get(token, self.encoder.get(self.unk_token))
def _convert_id_to_token(self, index):
"""Converts an index (integer) in a token (str) using the vocab."""
return self.decoder.get(index, self.unk_token)
def convert_tokens_to_string(self, tokens):
"""Converts a sequence of tokens (string) in a single string."""
out_string = " ".join(tokens).replace("@@ ", "").strip()
return out_string
def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
if not os.path.isdir(save_directory):
logger.error(f"Vocabulary path ({save_directory}) should be a directory")
return
out_vocab_file = os.path.join(
save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]
)
out_merge_file = os.path.join(
save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["merges_file"]
)
if os.path.abspath(self.vocab_file) != os.path.abspath(out_vocab_file) and os.path.isfile(self.vocab_file):
copyfile(self.vocab_file, out_vocab_file)
elif not os.path.isfile(self.vocab_file):
with open(out_vocab_file, "wb") as fi:
content_spiece_model = self.sp_model.serialized_model_proto()
fi.write(content_spiece_model)
if os.path.abspath(self.merges_file) != os.path.abspath(out_merge_file):
copyfile(self.merges_file, out_merge_file)
return out_vocab_file, out_merge_file
# def decode(self, token_ids, skip_special_tokens=False, clean_up_tokenization_spaces=True):
# filtered_tokens = ' '.join(self.convert_ids_to_tokens(token_ids, skip_special_tokens=skip_special_tokens))
# tokens_generated_so_far = re.sub('(@@ )', '', string=filtered_tokens)
# tokens_generated_so_far = re.sub('(@@ ?$)', '', string=tokens_generated_so_far)
# return ''.join(tokens_generated_so_far)
def add_from_file(self, f):
"""
Loads a pre-existing dictionary from a text file and adds its symbols to this instance.
"""
if isinstance(f, str):
try:
with open(f, "r", encoding="utf-8") as fd:
self.add_from_file(fd)
except FileNotFoundError as fnfe:
raise fnfe
except UnicodeError:
raise Exception(f"Incorrect encoding detected in {f}, please rebuild the dataset")
return
lines = f.readlines()
for lineTmp in lines:
line = lineTmp.strip()
idx = line.rfind(" ")
if idx == -1:
raise ValueError("Incorrect dictionary format, expected '<token> <cnt>'")
word = line[:idx]
self.encoder[word] = len(self.encoder)
|
transformers/src/transformers/models/phobert/tokenization_phobert.py/0
|
{
"file_path": "transformers/src/transformers/models/phobert/tokenization_phobert.py",
"repo_id": "transformers",
"token_count": 5906
}
| 376
|
# coding=utf-8
# Copyright 2023 Authors: Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan,
# Kaitao Song, Ding Liang, Tong Lu, Ping Luo, Ling Shao and The HuggingFace Inc. team.
# 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.
"""Convert Pvt checkpoints from the original library."""
import argparse
from pathlib import Path
import requests
import torch
from PIL import Image
from transformers import PvtConfig, PvtForImageClassification, PvtImageProcessor
from transformers.utils import logging
logging.set_verbosity_info()
logger = logging.get_logger(__name__)
# here we list all keys to be renamed (original name on the left, our name on the right)
def create_rename_keys(config):
rename_keys = []
for i in range(config.num_encoder_blocks):
# Remane embedings' paramters
rename_keys.append((f"pos_embed{i + 1}", f"pvt.encoder.patch_embeddings.{i}.position_embeddings"))
rename_keys.append((f"patch_embed{i + 1}.proj.weight", f"pvt.encoder.patch_embeddings.{i}.projection.weight"))
rename_keys.append((f"patch_embed{i + 1}.proj.bias", f"pvt.encoder.patch_embeddings.{i}.projection.bias"))
rename_keys.append((f"patch_embed{i + 1}.norm.weight", f"pvt.encoder.patch_embeddings.{i}.layer_norm.weight"))
rename_keys.append((f"patch_embed{i + 1}.norm.bias", f"pvt.encoder.patch_embeddings.{i}.layer_norm.bias"))
for j in range(config.depths[i]):
# Rename blocks' parameters
rename_keys.append(
(f"block{i + 1}.{j}.attn.q.weight", f"pvt.encoder.block.{i}.{j}.attention.self.query.weight")
)
rename_keys.append(
(f"block{i + 1}.{j}.attn.q.bias", f"pvt.encoder.block.{i}.{j}.attention.self.query.bias")
)
rename_keys.append(
(f"block{i + 1}.{j}.attn.kv.weight", f"pvt.encoder.block.{i}.{j}.attention.self.kv.weight")
)
rename_keys.append((f"block{i + 1}.{j}.attn.kv.bias", f"pvt.encoder.block.{i}.{j}.attention.self.kv.bias"))
if config.sequence_reduction_ratios[i] > 1:
rename_keys.append(
(
f"block{i + 1}.{j}.attn.norm.weight",
f"pvt.encoder.block.{i}.{j}.attention.self.layer_norm.weight",
)
)
rename_keys.append(
(f"block{i + 1}.{j}.attn.norm.bias", f"pvt.encoder.block.{i}.{j}.attention.self.layer_norm.bias")
)
rename_keys.append(
(
f"block{i + 1}.{j}.attn.sr.weight",
f"pvt.encoder.block.{i}.{j}.attention.self.sequence_reduction.weight",
)
)
rename_keys.append(
(
f"block{i + 1}.{j}.attn.sr.bias",
f"pvt.encoder.block.{i}.{j}.attention.self.sequence_reduction.bias",
)
)
rename_keys.append(
(f"block{i + 1}.{j}.attn.proj.weight", f"pvt.encoder.block.{i}.{j}.attention.output.dense.weight")
)
rename_keys.append(
(f"block{i + 1}.{j}.attn.proj.bias", f"pvt.encoder.block.{i}.{j}.attention.output.dense.bias")
)
rename_keys.append((f"block{i + 1}.{j}.norm1.weight", f"pvt.encoder.block.{i}.{j}.layer_norm_1.weight"))
rename_keys.append((f"block{i + 1}.{j}.norm1.bias", f"pvt.encoder.block.{i}.{j}.layer_norm_1.bias"))
rename_keys.append((f"block{i + 1}.{j}.norm2.weight", f"pvt.encoder.block.{i}.{j}.layer_norm_2.weight"))
rename_keys.append((f"block{i + 1}.{j}.norm2.bias", f"pvt.encoder.block.{i}.{j}.layer_norm_2.bias"))
rename_keys.append((f"block{i + 1}.{j}.mlp.fc1.weight", f"pvt.encoder.block.{i}.{j}.mlp.dense1.weight"))
rename_keys.append((f"block{i + 1}.{j}.mlp.fc1.bias", f"pvt.encoder.block.{i}.{j}.mlp.dense1.bias"))
rename_keys.append((f"block{i + 1}.{j}.mlp.fc2.weight", f"pvt.encoder.block.{i}.{j}.mlp.dense2.weight"))
rename_keys.append((f"block{i + 1}.{j}.mlp.fc2.bias", f"pvt.encoder.block.{i}.{j}.mlp.dense2.bias"))
# Rename cls token
rename_keys.extend(
[
("cls_token", "pvt.encoder.patch_embeddings.3.cls_token"),
]
)
# Rename norm layer and classifier layer
rename_keys.extend(
[
("norm.weight", "pvt.encoder.layer_norm.weight"),
("norm.bias", "pvt.encoder.layer_norm.bias"),
("head.weight", "classifier.weight"),
("head.bias", "classifier.bias"),
]
)
return rename_keys
# we split up the matrix of each encoder layer into queries, keys and values
def read_in_k_v(state_dict, config):
# for each of the encoder blocks:
for i in range(config.num_encoder_blocks):
for j in range(config.depths[i]):
# read in weights + bias of keys and values (which is a single matrix in the original implementation)
kv_weight = state_dict.pop(f"pvt.encoder.block.{i}.{j}.attention.self.kv.weight")
kv_bias = state_dict.pop(f"pvt.encoder.block.{i}.{j}.attention.self.kv.bias")
# next, add keys and values (in that order) to the state dict
state_dict[f"pvt.encoder.block.{i}.{j}.attention.self.key.weight"] = kv_weight[: config.hidden_sizes[i], :]
state_dict[f"pvt.encoder.block.{i}.{j}.attention.self.key.bias"] = kv_bias[: config.hidden_sizes[i]]
state_dict[f"pvt.encoder.block.{i}.{j}.attention.self.value.weight"] = kv_weight[
config.hidden_sizes[i] :, :
]
state_dict[f"pvt.encoder.block.{i}.{j}.attention.self.value.bias"] = kv_bias[config.hidden_sizes[i] :]
def rename_key(dct, old, new):
val = dct.pop(old)
dct[new] = val
# We will verify our results on an image of cute cats
def prepare_img():
url = "http://images.cocodataset.org/val2017/000000039769.jpg"
im = Image.open(requests.get(url, stream=True).raw)
return im
@torch.no_grad()
def convert_pvt_checkpoint(pvt_size, pvt_checkpoint, pytorch_dump_folder_path):
"""
Copy/paste/tweak model's weights to our PVT structure.
"""
# define default Pvt configuration
if pvt_size == "tiny":
config_path = "Zetatech/pvt-tiny-224"
elif pvt_size == "small":
config_path = "Zetatech/pvt-small-224"
elif pvt_size == "medium":
config_path = "Zetatech/pvt-medium-224"
elif pvt_size == "large":
config_path = "Zetatech/pvt-large-224"
else:
raise ValueError(f"Available model's size: 'tiny', 'small', 'medium', 'large', but " f"'{pvt_size}' was given")
config = PvtConfig(name_or_path=config_path)
# load original model from https://github.com/whai362/PVT
state_dict = torch.load(pvt_checkpoint, map_location="cpu")
rename_keys = create_rename_keys(config)
for src, dest in rename_keys:
rename_key(state_dict, src, dest)
read_in_k_v(state_dict, config)
# load HuggingFace model
model = PvtForImageClassification(config).eval()
model.load_state_dict(state_dict)
# Check outputs on an image, prepared by PVTFeatureExtractor
image_processor = PvtImageProcessor(size=config.image_size)
encoding = image_processor(images=prepare_img(), return_tensors="pt")
pixel_values = encoding["pixel_values"]
outputs = model(pixel_values)
logits = outputs.logits.detach().cpu()
if pvt_size == "tiny":
expected_slice_logits = torch.tensor([-1.4192, -1.9158, -0.9702])
elif pvt_size == "small":
expected_slice_logits = torch.tensor([0.4353, -0.1960, -0.2373])
elif pvt_size == "medium":
expected_slice_logits = torch.tensor([-0.2914, -0.2231, 0.0321])
elif pvt_size == "large":
expected_slice_logits = torch.tensor([0.3740, -0.7739, -0.4214])
else:
raise ValueError(f"Available model's size: 'tiny', 'small', 'medium', 'large', but " f"'{pvt_size}' was given")
assert torch.allclose(logits[0, :3], expected_slice_logits, atol=1e-4)
Path(pytorch_dump_folder_path).mkdir(exist_ok=True)
print(f"Saving model pytorch_model.bin to {pytorch_dump_folder_path}")
model.save_pretrained(pytorch_dump_folder_path)
print(f"Saving image processor to {pytorch_dump_folder_path}")
image_processor.save_pretrained(pytorch_dump_folder_path)
if __name__ == "__main__":
parser = argparse.ArgumentParser()
# Required parameters
parser.add_argument(
"--pvt_size",
default="tiny",
type=str,
help="Size of the PVT pretrained model you'd like to convert.",
)
parser.add_argument(
"--pvt_checkpoint",
default="pvt_tiny.pth",
type=str,
help="Checkpoint of the PVT pretrained model you'd like to convert.",
)
parser.add_argument(
"--pytorch_dump_folder_path", default=None, type=str, help="Path to the output PyTorch model directory."
)
args = parser.parse_args()
convert_pvt_checkpoint(args.pvt_size, args.pvt_checkpoint, args.pytorch_dump_folder_path)
|
transformers/src/transformers/models/pvt/convert_pvt_to_pytorch.py/0
|
{
"file_path": "transformers/src/transformers/models/pvt/convert_pvt_to_pytorch.py",
"repo_id": "transformers",
"token_count": 4511
}
| 377
|
# Copyright 2024 The Qwen Team and The HuggingFace Inc. team. 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.
from typing import TYPE_CHECKING
from ...utils import (
OptionalDependencyNotAvailable,
_LazyModule,
is_torch_available,
)
_import_structure = {
"configuration_qwen2_moe": ["Qwen2MoeConfig"],
}
try:
if not is_torch_available():
raise OptionalDependencyNotAvailable()
except OptionalDependencyNotAvailable:
pass
else:
_import_structure["modeling_qwen2_moe"] = [
"Qwen2MoeForCausalLM",
"Qwen2MoeModel",
"Qwen2MoePreTrainedModel",
"Qwen2MoeForSequenceClassification",
"Qwen2MoeForTokenClassification",
]
if TYPE_CHECKING:
from .configuration_qwen2_moe import Qwen2MoeConfig
try:
if not is_torch_available():
raise OptionalDependencyNotAvailable()
except OptionalDependencyNotAvailable:
pass
else:
from .modeling_qwen2_moe import (
Qwen2MoeForCausalLM,
Qwen2MoeForSequenceClassification,
Qwen2MoeForTokenClassification,
Qwen2MoeModel,
Qwen2MoePreTrainedModel,
)
else:
import sys
sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)
|
transformers/src/transformers/models/qwen2_moe/__init__.py/0
|
{
"file_path": "transformers/src/transformers/models/qwen2_moe/__init__.py",
"repo_id": "transformers",
"token_count": 707
}
| 378
|
# Copyright 2024 The HuggingFace Inc. team. 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.
import argparse
import os
import warnings
import torch
from accelerate import init_empty_weights
from transformers import GemmaTokenizer, RecurrentGemmaConfig, RecurrentGemmaForCausalLM
try:
from transformers import GemmaTokenizerFast
except ImportError as e:
warnings.warn(e)
warnings.warn(
"The converted tokenizer will be the `slow` tokenizer. To use the fast, update your `tokenizers` library and re-run the tokenizer conversion"
)
GemmaTokenizerFast = None
import regex as re
"""
Sample usage:
```
python src/transformers/models/gemma/convert_gemma_weights_to_hf.py \
--input_dir /path/to/downloaded/gemma/weights --model_size 7B --output_dir /output/path
```
Thereafter, models can be loaded via:
```py
from transformers import GemmaForCausalLM, GemmaTokenizerFast
model = GemmaForCausalLM.from_pretrained("/output/path")
tokenizer = GemmaTokenizerFast.from_pretrained("/output/path")
```
Important note: you need to be able to host the whole model in RAM to execute this script (even if the biggest versions
come in several checkpoints they each contain a part of each weight of the model, so we need to load them all in RAM).
"""
gemma_2b_config = RecurrentGemmaConfig(
num_attention_heads=10,
num_key_value_heads=1,
hidden_size=2560,
intermediate_size=15360,
vocab_size=256000,
num_hidden_layers=26,
)
gemma_7b_config = RecurrentGemmaConfig()
CONFIG_MAPPING = {"2B": gemma_2b_config, "7B": gemma_7b_config}
LAYER_NAME_MAPPING = {"embedder.weight": "model.embed_tokens.weight"}
def write_model(save_path, input_base_path, config, safe_serialization=True, push_to_hub=False, dtype=torch.float32):
print(f"Fetching all parameters from the checkpoint at '{input_base_path}'")
model_state_dict = torch.load(input_base_path, map_location="cpu")
REPLACEMENT = {
"blocks.": "layers.",
".ffw_down.b": ".down_proj.b",
".ffw_down.w": ".down_proj.w",
".ffw_up.b": ".up_proj.bias",
".ffw_up.w": ".up_proj.weight",
"recurrent_block": "temporal_block",
"attention_block": "temporal_block",
"temporal_block.proj_final": "temporal_block.out_proj",
"norm.scale": "norm.weight",
".proj_k": ".k_proj",
".proj_q": ".q_proj",
".proj_v": ".v_proj",
".proj_final": ".o_proj",
"embedder.input_embedding": "embed_tokens.weight",
"conv_1d.w": "conv_1d.weight",
"conv_1d.b": "conv_1d.bias",
"input_gate.w": "input_gate.weight",
"input_gate.b": "input_gate.bias",
"a_param": "recurrent_param",
"a_gate.b": "recurrent_gate.bias",
"a_gate.w": "recurrent_gate.weight",
}
state_dict = {}
for k, v in model_state_dict.items():
k = "model." + k
pattern = re.compile("|".join(map(re.escape, REPLACEMENT.keys())))
key = pattern.sub(lambda match: REPLACEMENT[match.group(0)], k)
if "conv_1d.weight" in key:
v = v[:, None, :].transpose(0, 2)
if "up_proj.weight" in key:
state_dict[key.replace("up_proj", "gate_proj")] = v[0].T.contiguous()
v = v[1].T.contiguous()
if "up_proj.bias" in key:
state_dict[key.replace("up_proj", "gate_proj")] = v[0, 0, 0].clone()
v = v[1, 0, 0].contiguous()
if "recurrent_gate.bias" in key:
state_dict[key.replace("gate.", "gate_")] = v.contiguous().clone()
elif "recurrent_gate.weight" in key:
state_dict[key.replace("gate.", "gate_")] = v.contiguous().clone()
elif "input_gate.b" in key:
state_dict[key.replace("gate.", "gate_")] = v.contiguous().clone()
elif "input_gate.w" in key:
state_dict[key.replace("gate.", "gate_")] = v.contiguous().clone()
elif "embed_tokens" in key:
state_dict[key] = v[: config.vocab_size, :].contiguous().clone()
state_dict["lm_head.weight"] = v[: config.vocab_size, :].contiguous().clone()
else:
state_dict[key] = v.contiguous()
torch.set_default_dtype(dtype)
print("Loading the checkpoint in a Gemma model.")
with init_empty_weights():
model = RecurrentGemmaForCausalLM(config)
model.load_state_dict(state_dict, assign=True, strict=True)
model.config.torch_dtype = torch.float32
del model.config._name_or_path
print("Saving in the Transformers format.")
if push_to_hub:
print(f"pushing the model to {save_path}")
else:
model.save_pretrained(save_path, safe_serialization=safe_serialization)
def write_tokenizer(input_tokenizer_path, save_path, push_to_hub=False):
# Initialize the tokenizer based on the `spm` model
tokenizer_class = GemmaTokenizer if GemmaTokenizerFast is None else GemmaTokenizerFast
print(f"Saving a {tokenizer_class.__name__} to {save_path}.")
tokenizer = tokenizer_class(input_tokenizer_path)
if push_to_hub:
tokenizer.push_to_hub(save_path)
else:
tokenizer.save_pretrained(save_path)
def main():
parser = argparse.ArgumentParser()
parser.add_argument(
"--input_checkpoint",
help="Absolute path to the target Gemma weights.",
default="/home/arthur/transformers_recurrentgemma/google/recurrent-gemma-2b-it/ToBeDeleted/2b-it.pt",
)
parser.add_argument(
"--tokenizer_checkpoint",
help="Location of Gemma tokenizer model",
)
parser.add_argument(
"--model_size",
default="2B",
choices=["2B", "7B", "tokenizer_only"],
help="'f' models correspond to the finetuned versions, and are specific to the Gemma2 official release. For more details on Gemma2, checkout the original repo: https://huggingface.co/google/gemma-7b",
)
parser.add_argument(
"--output_dir",
default="google/recurrent-gemma-2b-it-hf",
help="Location to write HF model and tokenizer",
)
parser.add_argument(
"--pickle_serialization",
help="Whether or not to save using `safetensors`.",
action="store_true",
default=False,
)
parser.add_argument(
"--convert_tokenizer",
help="Whether or not to convert the tokenizer as well.",
action="store_true",
default=False,
)
parser.add_argument(
"--push_to_hub",
help="Whether or not to push the model to the hub at `output_dir` instead of saving it locally.",
action="store_true",
default=False,
)
parser.add_argument(
"--dtype",
default="float32",
help="Target dtype of the converted model",
)
args = parser.parse_args()
if args.convert_tokenizer:
if args.tokenizer_checkpoint is None:
raise ValueError("Path to the tokenizer is required when passing --convert_tokenizer")
spm_path = os.path.join(args.tokenizer_checkpoint)
write_tokenizer(spm_path, args.output_dir, args.push_to_hub)
config = CONFIG_MAPPING[args.model_size]
dtype = getattr(torch, args.dtype)
write_model(
config=config,
input_base_path=args.input_checkpoint,
save_path=args.output_dir,
safe_serialization=not args.pickle_serialization,
push_to_hub=args.push_to_hub,
dtype=dtype,
)
if __name__ == "__main__":
main()
|
transformers/src/transformers/models/recurrent_gemma/convert_recurrent_gemma_to_hf.py/0
|
{
"file_path": "transformers/src/transformers/models/recurrent_gemma/convert_recurrent_gemma_to_hf.py",
"repo_id": "transformers",
"token_count": 3313
}
| 379
|
# coding=utf-8
# Copyright 2023 The HuggingFace Inc. team.
#
# 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.
"""Convert a RWKV checkpoint from BlinkDL to the Hugging Face format."""
import argparse
import gc
import json
import os
import re
import torch
from huggingface_hub import hf_hub_download
from transformers import AutoModelForCausalLM, AutoTokenizer, PreTrainedTokenizerFast, RwkvConfig
from transformers.modeling_utils import WEIGHTS_INDEX_NAME, shard_checkpoint
NUM_HIDDEN_LAYERS_MAPPING = {
"169M": 12,
"430M": 24,
"1B5": 24,
"3B": 32,
"7B": 32,
"14B": 40,
}
HIDEN_SIZE_MAPPING = {
"169M": 768,
"430M": 1024,
"1B5": 2048,
"3B": 2560,
"7B": 4096,
"14B": 5120,
}
def convert_state_dict(state_dict):
state_dict_keys = list(state_dict.keys())
for name in state_dict_keys:
weight = state_dict.pop(name)
# emb -> embedding
if name.startswith("emb."):
name = name.replace("emb.", "embeddings.")
# ln_0 -> pre_ln (only present at block 0)
if name.startswith("blocks.0.ln0"):
name = name.replace("blocks.0.ln0", "blocks.0.pre_ln")
# att -> attention
name = re.sub(r"blocks\.(\d+)\.att", r"blocks.\1.attention", name)
# ffn -> feed_forward
name = re.sub(r"blocks\.(\d+)\.ffn", r"blocks.\1.feed_forward", name)
# time_mix_k -> time_mix_key and reshape
if name.endswith(".time_mix_k"):
name = name.replace(".time_mix_k", ".time_mix_key")
# time_mix_v -> time_mix_value and reshape
if name.endswith(".time_mix_v"):
name = name.replace(".time_mix_v", ".time_mix_value")
# time_mix_r -> time_mix_key and reshape
if name.endswith(".time_mix_r"):
name = name.replace(".time_mix_r", ".time_mix_receptance")
if name != "head.weight":
name = "rwkv." + name
state_dict[name] = weight
return state_dict
def convert_rmkv_checkpoint_to_hf_format(
repo_id, checkpoint_file, output_dir, size=None, tokenizer_file=None, push_to_hub=False, model_name=None
):
# 1. If possible, build the tokenizer.
if tokenizer_file is None:
print("No `--tokenizer_file` provided, we will use the default tokenizer.")
vocab_size = 50277
tokenizer = AutoTokenizer.from_pretrained("EleutherAI/gpt-neox-20b")
else:
tokenizer = PreTrainedTokenizerFast(tokenizer_file=tokenizer_file)
vocab_size = len(tokenizer)
tokenizer.save_pretrained(output_dir)
# 2. Build the config
possible_sizes = list(NUM_HIDDEN_LAYERS_MAPPING.keys())
if size is None:
# Try to infer size from the checkpoint name
for candidate in possible_sizes:
if candidate in checkpoint_file:
size = candidate
break
if size is None:
raise ValueError("Could not infer the size, please provide it with the `--size` argument.")
if size not in possible_sizes:
raise ValueError(f"`size` should be one of {possible_sizes}, got {size}.")
config = RwkvConfig(
vocab_size=vocab_size,
num_hidden_layers=NUM_HIDDEN_LAYERS_MAPPING[size],
hidden_size=HIDEN_SIZE_MAPPING[size],
)
config.save_pretrained(output_dir)
# 3. Download model file then convert state_dict
model_file = hf_hub_download(repo_id, checkpoint_file)
state_dict = torch.load(model_file, map_location="cpu")
state_dict = convert_state_dict(state_dict)
# 4. Split in shards and save
shards, index = shard_checkpoint(state_dict)
for shard_file, shard in shards.items():
torch.save(shard, os.path.join(output_dir, shard_file))
if index is not None:
save_index_file = os.path.join(output_dir, WEIGHTS_INDEX_NAME)
# Save the index as well
with open(save_index_file, "w", encoding="utf-8") as f:
content = json.dumps(index, indent=2, sort_keys=True) + "\n"
f.write(content)
# 5. Clean up shards (for some reason the file PyTorch saves take the same space as the whole state_dict
print(
"Cleaning up shards. This may error with an OOM error, it this is the case don't worry you still have converted the model."
)
shard_files = list(shards.keys())
del state_dict
del shards
gc.collect()
for shard_file in shard_files:
state_dict = torch.load(os.path.join(output_dir, shard_file))
torch.save({k: v.cpu().clone() for k, v in state_dict.items()}, os.path.join(output_dir, shard_file))
del state_dict
gc.collect()
if push_to_hub:
if model_name is None:
raise ValueError("Please provide a `model_name` to push the model to the Hub.")
model = AutoModelForCausalLM.from_pretrained(output_dir)
model.push_to_hub(model_name, max_shard_size="2GB")
tokenizer.push_to_hub(model_name)
if __name__ == "__main__":
parser = argparse.ArgumentParser()
# Required parameters
parser.add_argument(
"--repo_id", default=None, type=str, required=True, help="Repo ID from which to pull the checkpoint."
)
parser.add_argument(
"--checkpoint_file", default=None, type=str, required=True, help="Name of the checkpoint file in the repo."
)
parser.add_argument(
"--output_dir", default=None, type=str, required=True, help="Where to save the converted model."
)
parser.add_argument(
"--tokenizer_file",
default=None,
type=str,
help="Path to the tokenizer file to use (if not provided, only the model is converted).",
)
parser.add_argument(
"--size",
default=None,
type=str,
help="Size of the model. Will be inferred from the `checkpoint_file` if not passed.",
)
parser.add_argument(
"--push_to_hub",
action="store_true",
help="Push to the Hub the converted model.",
)
parser.add_argument(
"--model_name",
default=None,
type=str,
help="Name of the pushed model on the Hub, including the username / organization.",
)
args = parser.parse_args()
convert_rmkv_checkpoint_to_hf_format(
args.repo_id,
args.checkpoint_file,
args.output_dir,
size=args.size,
tokenizer_file=args.tokenizer_file,
push_to_hub=args.push_to_hub,
model_name=args.model_name,
)
|
transformers/src/transformers/models/rwkv/convert_rwkv_checkpoint_to_hf.py/0
|
{
"file_path": "transformers/src/transformers/models/rwkv/convert_rwkv_checkpoint_to_hf.py",
"repo_id": "transformers",
"token_count": 2938
}
| 380
|
# coding=utf-8
# Copyright 2023 The HuggingFace Inc. team. 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.
"""Fast Tokenization class for SeamlessM4T."""
import os
from shutil import copyfile
from typing import List, Optional, Tuple, Union
from tokenizers import processors
from ...tokenization_utils import (
BatchEncoding,
PreTokenizedInput,
TextInput,
)
from ...tokenization_utils_fast import PreTrainedTokenizerFast
from ...utils import PaddingStrategy, is_sentencepiece_available, logging
if is_sentencepiece_available():
from .tokenization_seamless_m4t import SeamlessM4TTokenizer
else:
SeamlessM4TTokenizer = None
logger = logging.get_logger(__name__)
VOCAB_FILES_NAMES = {"vocab_file": "sentencepiece.bpe.model", "tokenizer_file": "tokenizer.json"}
class SeamlessM4TTokenizerFast(PreTrainedTokenizerFast):
"""
Construct a "fast" SeamlessM4T tokenizer (backed by HuggingFace's *tokenizers* library). Based on
[BPE](https://huggingface.co/docs/tokenizers/python/latest/components.html?highlight=BPE#models).
This tokenizer inherits from [`PreTrainedTokenizerFast`] which contains most of the main methods. Users should
refer to this superclass for more information regarding those methods.
The tokenization method is `<language code> <tokens> <eos>` for source language documents, and `<eos> <language
code> <tokens> <eos>` for target language documents.
Examples:
```python
>>> from transformers import SeamlessM4TTokenizerFast
>>> tokenizer = SeamlessM4TTokenizerFast.from_pretrained(
... "facebook/hf-seamless-m4t-medium", src_lang="eng", tgt_lang="fra"
... )
>>> example_english_phrase = " UN Chief Says There Is No Military Solution in Syria"
>>> expected_translation_french = "Le chef de l'ONU affirme qu'il n'y a pas de solution militaire en Syrie."
>>> inputs = tokenizer(example_english_phrase, text_target=expected_translation_french, return_tensors="pt")
```
Args:
vocab_file (`str`, *optional*):
Path to the vocabulary file.
tokenizer_file (`str`, *optional*):
The path to a tokenizer file to use instead of the vocab file.
bos_token (`str`, *optional*, defaults to `"<s>"`):
The beginning of sequence token that was used during pretraining. Can be used a sequence classifier token.
<Tip>
When building a sequence using special tokens, this is not the token that is used for the beginning of
sequence. The token used is the `cls_token`.
</Tip>
eos_token (`str`, *optional*, defaults to `"</s>"`):
The end of sequence token.
<Tip>
When building a sequence using special tokens, this is not the token that is used for the end of sequence.
The token used is the `sep_token`.
</Tip>
sep_token (`str`, *optional*, defaults to `"</s>"`):
The separator token, which is used when building a sequence from multiple sequences, e.g. two sequences for
sequence classification or for a text and a question for question answering. It is also used as the last
token of a sequence built with special tokens.
cls_token (`str`, *optional*, defaults to `"<s>"`):
The classifier token which is used when doing sequence classification (classification of the whole sequence
instead of per-token classification). It is the first token of the sequence when built with special tokens.
unk_token (`str`, *optional*, defaults to `"<unk>"`):
The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this
token instead.
pad_token (`str`, *optional*, defaults to `"<pad>"`):
The token used for padding, for example when batching sequences of different lengths.
src_lang (`str`, *optional*, defaults to `"eng"`):
The language to use as source language for translation.
tgt_lang (`str`, *optional*, defaults to `"fra"`):
The language to use as target language for translation.
additional_special_tokens (tuple or list of `str` or `tokenizers.AddedToken`, *optional*):
A tuple or a list of additional special tokens.
"""
vocab_files_names = VOCAB_FILES_NAMES
slow_tokenizer_class = SeamlessM4TTokenizer
model_input_names = ["input_ids", "attention_mask"]
prefix_tokens: List[int] = []
suffix_tokens: List[int] = []
def __init__(
self,
vocab_file=None,
tokenizer_file=None,
bos_token="<s>",
eos_token="</s>",
sep_token="</s>",
cls_token="<s>",
unk_token="<unk>",
pad_token="<pad>",
src_lang="eng",
tgt_lang="fra",
additional_special_tokens=None,
**kwargs,
):
super().__init__(
vocab_file=vocab_file,
tokenizer_file=tokenizer_file,
bos_token=bos_token,
eos_token=eos_token,
sep_token=sep_token,
cls_token=cls_token,
unk_token=unk_token,
pad_token=pad_token,
src_lang=src_lang,
tgt_lang=tgt_lang,
additional_special_tokens=additional_special_tokens,
**kwargs,
)
self.vocab_file = vocab_file
self._src_lang = f"__{src_lang}__" if "__" not in src_lang else src_lang
self._tgt_lang = f"__{tgt_lang}__" if "__" not in tgt_lang else tgt_lang
self.set_src_lang_special_tokens(self._src_lang)
self.set_tgt_lang_special_tokens(self._tgt_lang)
@property
def can_save_slow_tokenizer(self) -> bool:
return os.path.isfile(self.vocab_file) if self.vocab_file else False
@property
# Copied from transformers.models.nllb.tokenization_nllb.NllbTokenizer.src_lang
def src_lang(self) -> str:
return self._src_lang
@src_lang.setter
def src_lang(self, new_src_lang: str) -> None:
if "__" not in new_src_lang:
self._src_lang = f"__{new_src_lang}__"
else:
self._src_lang = new_src_lang
self.set_src_lang_special_tokens(self._src_lang)
@property
def tgt_lang(self) -> str:
return self._tgt_lang
@tgt_lang.setter
def tgt_lang(self, new_tgt_lang: str) -> None:
if "__" not in new_tgt_lang:
self._tgt_lang = f"__{new_tgt_lang}__"
else:
self._tgt_lang = new_tgt_lang
self.set_tgt_lang_special_tokens(self._tgt_lang)
def build_inputs_with_special_tokens(
self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
) -> List[int]:
"""
Build model inputs from a sequence or a pair of sequence for sequence classification tasks by concatenating and
adding special tokens. The special tokens depend on calling set_lang.
An SeamlessM4T sequence has the following format, where `X` represents the sequence:
- `input_ids` (for encoder) `[src_lang_code] X [eos]`
- `decoder_input_ids`: (for decoder) `[eos, tgt_lang_code] X [eos]`
BOS is never used. Pairs of sequences are not the expected use case, but they will be handled without a
separator.
Args:
token_ids_0 (`List[int]`):
List of IDs to which the special tokens will be added.
token_ids_1 (`List[int]`, *optional*):
Optional second list of IDs for sequence pairs.
Returns:
`List[int]`: list of [input IDs](../glossary#input-ids) with the appropriate special tokens.
"""
if token_ids_1 is None:
return self.prefix_tokens + token_ids_0 + self.suffix_tokens
# We don't expect to process pairs, but leave the pair logic for API consistency
return self.prefix_tokens + token_ids_0 + token_ids_1 + self.suffix_tokens
# Copied from transformers.models.nllb.tokenization_nllb_fast.NllbTokenizerFast.create_token_type_ids_from_sequences
def create_token_type_ids_from_sequences(
self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
) -> List[int]:
"""
Create a mask from the two sequences passed to be used in a sequence-pair classification task. nllb does not
make use of token type ids, therefore a list of zeros is returned.
Args:
token_ids_0 (`List[int]`):
List of IDs.
token_ids_1 (`List[int]`, *optional*):
Optional second list of IDs for sequence pairs.
Returns:
`List[int]`: List of zeros.
"""
sep = [self.sep_token_id]
cls = [self.cls_token_id]
if token_ids_1 is None:
return len(cls + token_ids_0 + sep) * [0]
return len(cls + token_ids_0 + sep + sep + token_ids_1 + sep) * [0]
def _build_translation_inputs(
self, raw_inputs, return_tensors: str, src_lang: Optional[str], tgt_lang: Optional[str], **extra_kwargs
):
"""Used by translation pipeline, to prepare inputs for the generate function"""
if src_lang is None or tgt_lang is None:
raise ValueError("Translation requires a `src_lang` and a `tgt_lang` for this model")
self.src_lang = src_lang
inputs = self(raw_inputs, add_special_tokens=True, return_tensors=return_tensors, **extra_kwargs)
if "__" not in tgt_lang:
tgt_lang = f"__{tgt_lang}__"
tgt_lang_id = self.convert_tokens_to_ids(tgt_lang)
inputs["forced_bos_token_id"] = tgt_lang_id
return inputs
# Copied from transformers.models.nllb.tokenization_nllb_fast.NllbTokenizerFast.prepare_seq2seq_batch with "fra_Latn"->"fra", "eng_Latn"->"eng"
def prepare_seq2seq_batch(
self,
src_texts: List[str],
src_lang: str = "eng",
tgt_texts: Optional[List[str]] = None,
tgt_lang: str = "fra",
**kwargs,
) -> BatchEncoding:
self.src_lang = src_lang
self.tgt_lang = tgt_lang
return super().prepare_seq2seq_batch(src_texts, tgt_texts, **kwargs)
# Copied from transformers.models.nllb.tokenization_nllb_fast.NllbTokenizerFast._switch_to_input_mode
def _switch_to_input_mode(self):
return self.set_src_lang_special_tokens(self.src_lang)
# Copied from transformers.models.nllb.tokenization_nllb_fast.NllbTokenizerFast._switch_to_target_mode
def _switch_to_target_mode(self):
return self.set_tgt_lang_special_tokens(self.tgt_lang)
def set_src_lang_special_tokens(self, src_lang) -> None:
"""Reset the special tokens to the source lang setting.
Prefix=[src_lang_code], suffix = [eos]
"""
self.cur_lang_code = self.convert_tokens_to_ids(src_lang)
if self.cur_lang_code == self.unk_token_id:
logger.warning_once(
f"`tgt_lang={src_lang}` has not be found in the `vocabulary`. Behaviour will probably be unexpected because the language token id will be replaced by the unknown token id."
)
self.init_kwargs["src_lang"] = src_lang
self.prefix_tokens = [self.cur_lang_code]
self.suffix_tokens = [self.eos_token_id]
prefix_tokens_str = self.convert_ids_to_tokens(self.prefix_tokens)
suffix_tokens_str = self.convert_ids_to_tokens(self.suffix_tokens)
self._tokenizer.post_processor = processors.TemplateProcessing(
single=prefix_tokens_str + ["$A"] + suffix_tokens_str,
pair=prefix_tokens_str + ["$A", "$B"] + suffix_tokens_str,
special_tokens=list(zip(prefix_tokens_str + suffix_tokens_str, self.prefix_tokens + self.suffix_tokens)),
)
def set_tgt_lang_special_tokens(self, lang: str) -> None:
"""Reset the special tokens to the target lang setting.
Prefix=[eos, tgt_lang_code] and suffix=[eos].
"""
self.cur_lang_code = self.convert_tokens_to_ids(lang)
if self.cur_lang_code == self.unk_token_id:
logger.warning_once(
f"`tgt_lang={lang}` has not be found in the `vocabulary`. Behaviour will probably be unexpected because the language token id will be replaced by the unknown token id."
)
self.init_kwargs["tgt_lang"] = lang
self.prefix_tokens = [self.eos_token_id, self.cur_lang_code]
self.suffix_tokens = [self.eos_token_id]
prefix_tokens_str = self.convert_ids_to_tokens(self.prefix_tokens)
suffix_tokens_str = self.convert_ids_to_tokens(self.suffix_tokens)
self._tokenizer.post_processor = processors.TemplateProcessing(
single=prefix_tokens_str + ["$A"] + suffix_tokens_str,
pair=prefix_tokens_str + ["$A", "$B"] + suffix_tokens_str,
special_tokens=list(zip(prefix_tokens_str + suffix_tokens_str, self.prefix_tokens + self.suffix_tokens)),
)
# Copied from transformers.models.nllb.tokenization_nllb_fast.NllbTokenizerFast.save_vocabulary
def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
if not self.can_save_slow_tokenizer:
raise ValueError(
"Your fast tokenizer does not have the necessary information to save the vocabulary for a slow "
"tokenizer."
)
if not os.path.isdir(save_directory):
logger.error(f"Vocabulary path ({save_directory}) should be a directory.")
return
out_vocab_file = os.path.join(
save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]
)
if os.path.abspath(self.vocab_file) != os.path.abspath(out_vocab_file):
copyfile(self.vocab_file, out_vocab_file)
return (out_vocab_file,)
@classmethod
def _from_pretrained(
cls,
resolved_vocab_files,
pretrained_model_name_or_path,
init_configuration,
*init_inputs,
token=None,
cache_dir=None,
local_files_only=False,
_commit_hash=None,
_is_local=False,
**kwargs,
):
tokenizer = super()._from_pretrained(
resolved_vocab_files,
pretrained_model_name_or_path,
init_configuration,
*init_inputs,
token=token,
cache_dir=cache_dir,
local_files_only=local_files_only,
_commit_hash=_commit_hash,
_is_local=_is_local,
**kwargs,
)
# ensure also set after from pretrained
tokenizer.set_src_lang_special_tokens(tokenizer._src_lang)
tokenizer.set_tgt_lang_special_tokens(tokenizer._tgt_lang)
return tokenizer
def __call__(
self,
text: Union[TextInput, PreTokenizedInput, List[TextInput], List[PreTokenizedInput]] = None,
text_pair: Optional[Union[TextInput, PreTokenizedInput, List[TextInput], List[PreTokenizedInput]]] = None,
text_target: Union[TextInput, PreTokenizedInput, List[TextInput], List[PreTokenizedInput]] = None,
text_pair_target: Optional[
Union[TextInput, PreTokenizedInput, List[TextInput], List[PreTokenizedInput]]
] = None,
padding: Union[bool, str, PaddingStrategy] = True,
pad_to_multiple_of: Optional[int] = 2,
src_lang: Optional[str] = None,
tgt_lang: Optional[str] = None,
**kwargs,
):
"""
Args:
text (`str`, `List[str]`, `List[List[str]]`, *optional*):
The sequence or batch of sequences to be encoded. Each sequence can be a string or a list of strings
(pretokenized string). If the sequences are provided as list of strings (pretokenized), you must set
`is_split_into_words=True` (to lift the ambiguity with a batch of sequences).
text_pair (`str`, `List[str]`, `List[List[str]]`, *optional*):
The sequence or batch of sequences to be encoded. Each sequence can be a string or a list of strings
(pretokenized string). If the sequences are provided as list of strings (pretokenized), you must set
`is_split_into_words=True` (to lift the ambiguity with a batch of sequences).
text_target (`str`, `List[str]`, `List[List[str]]`, *optional*):
The sequence or batch of sequences to be encoded as target texts. Each sequence can be a string or a
list of strings (pretokenized string). If the sequences are provided as list of strings (pretokenized),
you must set `is_split_into_words=True` (to lift the ambiguity with a batch of sequences).
text_pair_target (`str`, `List[str]`, `List[List[str]]`, *optional*):
The sequence or batch of sequences to be encoded as target texts. Each sequence can be a string or a
list of strings (pretokenized string). If the sequences are provided as list of strings (pretokenized),
you must set `is_split_into_words=True` (to lift the ambiguity with a batch of sequences).
padding (`bool`, `str` or [`~utils.PaddingStrategy`], *optional*, defaults to `True`):
Select a strategy to pad the returned sequences (according to the model's padding side and padding
index) among:
- `True` or `'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
sequence if provided).
- `'max_length'`: Pad to a maximum length specified with the argument `max_length` or to the maximum
acceptable input length for the model if that argument is not provided.
- `False` or `'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of different
lengths).
pad_to_multiple_of (`int`, *optional*):
If set will pad the sequence to a multiple of the provided value.
This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability
`>= 7.5` (Volta).
src_lang (`str`, *optional*):
A string representing the source language. If not specified, the last `src_lang` specified (either
during initialization or when calling this tokenizer) will be used.
tgt_lang (`str`, *optional*):
A string representing the target language. If not specified, the last `tgt_lang` specified (either
during initialization or when calling this tokenizer) will be used.
kwargs (*optional*):
Remaining dictionary of keyword arguments that will be passed to [`PreTrainedTokenizerFast.__call__`].
"""
if src_lang is not None:
self.src_lang = src_lang
if tgt_lang is not None:
self.tgt_lang = tgt_lang
output = super().__call__(
text=text,
text_pair=text_pair,
text_target=text_target,
text_pair_target=text_pair_target,
padding=padding,
pad_to_multiple_of=pad_to_multiple_of,
**kwargs,
)
return output
|
transformers/src/transformers/models/seamless_m4t/tokenization_seamless_m4t_fast.py/0
|
{
"file_path": "transformers/src/transformers/models/seamless_m4t/tokenization_seamless_m4t_fast.py",
"repo_id": "transformers",
"token_count": 8452
}
| 381
|
# coding=utf-8
# Copyright 2024 The HuggingFace Inc. team. 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.
"""PyTorch SegGpt model."""
import collections.abc
import math
from dataclasses import dataclass
from typing import Dict, List, Optional, Tuple, Union
import torch
import torch.utils.checkpoint
from torch import nn
from torch.nn import functional as F
from ...activations import ACT2FN
from ...modeling_utils import PreTrainedModel
from ...utils import (
ModelOutput,
add_start_docstrings,
add_start_docstrings_to_model_forward,
logging,
replace_return_docstrings,
)
from .configuration_seggpt import SegGptConfig
logger = logging.get_logger(__name__)
# General docstring
_CONFIG_FOR_DOC = "SegGptConfig"
# Base docstring
_CHECKPOINT_FOR_DOC = "BAAI/seggpt-vit-large"
_EXPECTED_OUTPUT_SHAPE = [3, 896, 448]
@dataclass
class SegGptEncoderOutput(ModelOutput):
"""
Output type of [`SegGptEncoderOutput`].
Args:
last_hidden_state (`torch.FloatTensor` of shape `(batch_size, patch_height, patch_width, hidden_size)`):
Sequence of hidden-states at the output of the last layer of the model.
hidden_states (`Tuple[torch.FloatTensor]`, `optional`, returned when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
of shape `(batch_size, patch_height, patch_width, hidden_size)`.
attentions (`Tuple[torch.FloatTensor]`, `optional`, returned when `config.output_attentions=True`):
Tuple of *torch.FloatTensor* (one for each layer) of shape
`(batch_size, num_heads, seq_len, seq_len)`.
intermediate_hidden_states (`Tuple[torch.FloatTensor]`, *optional*, returned when `config.intermediate_hidden_state_indices` is set):
Tuple of `torch.FloatTensor` of shape `(batch_size, patch_height, patch_width, hidden_size)`.
Each element in the Tuple corresponds to the output of the layer specified in `config.intermediate_hidden_state_indices`.
Additionaly, each feature passes through a LayerNorm.
"""
last_hidden_state: torch.FloatTensor
hidden_states: Optional[Tuple[torch.FloatTensor]] = None
attentions: Optional[Tuple[torch.FloatTensor]] = None
intermediate_hidden_states: Optional[Tuple[torch.FloatTensor]] = None
@dataclass
class SegGptImageSegmentationOutput(ModelOutput):
"""
Output type of [`SegGptImageSegmentationOutput`].
Args:
loss (`torch.FloatTensor`, *optional*, returned when `labels` is provided):
The loss value.
pred_masks (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
The predicted masks.
hidden_states (`Tuple[torch.FloatTensor]`, `optional`, returned when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
of shape `(batch_size, patch_height, patch_width, hidden_size)`.
attentions (`Tuple[torch.FloatTensor]`, `optional`, returned when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape
`(batch_size, num_heads, seq_len, seq_len)`.
"""
loss: Optional[torch.FloatTensor] = None
pred_masks: Optional[torch.FloatTensor] = None
hidden_states: Optional[Tuple[torch.FloatTensor]] = None
attentions: Optional[Tuple[torch.FloatTensor]] = None
# Copied from transformers.models.sam.modeling_sam.SamPatchEmbeddings with Sam->SegGpt
class SegGptPatchEmbeddings(nn.Module):
"""
This class turns `pixel_values` of shape `(batch_size, num_channels, height, width)` into the initial
`hidden_states` (patch embeddings) of shape `(batch_size, seq_length, hidden_size)` to be consumed by a
Transformer.
"""
def __init__(self, config):
super().__init__()
image_size, patch_size = config.image_size, config.patch_size
num_channels, hidden_size = config.num_channels, config.hidden_size
image_size = image_size if isinstance(image_size, collections.abc.Iterable) else (image_size, image_size)
patch_size = patch_size if isinstance(patch_size, collections.abc.Iterable) else (patch_size, patch_size)
num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
self.image_size = image_size
self.patch_size = patch_size
self.num_channels = num_channels
self.num_patches = num_patches
self.projection = nn.Conv2d(num_channels, hidden_size, kernel_size=patch_size, stride=patch_size)
def forward(self, pixel_values):
batch_size, num_channels, height, width = pixel_values.shape
if num_channels != self.num_channels:
raise ValueError(
"Make sure that the channel dimension of the pixel values match with the one set in the configuration."
)
if height != self.image_size[0] or width != self.image_size[1]:
raise ValueError(
f"Input image size ({height}*{width}) doesn't match model ({self.image_size[0]}*{self.image_size[1]})."
)
embeddings = self.projection(pixel_values).permute(0, 2, 3, 1)
return embeddings
class SegGptEmbeddings(nn.Module):
"""
Construct the embeddings from patch, position embeddings for input and prompt.
"""
def __init__(self, config: SegGptConfig) -> None:
super().__init__()
self.mask_token = nn.Parameter(torch.zeros(1, 1, 1, config.hidden_size))
self.segment_token_input = nn.Parameter(torch.zeros(1, 1, 1, config.hidden_size))
self.segment_token_prompt = nn.Parameter(torch.zeros(1, 1, 1, config.hidden_size))
# token for seg types
self.type_token_semantic = nn.Parameter(torch.zeros(1, 1, 1, config.hidden_size))
self.type_token_instance = nn.Parameter(torch.zeros(1, 1, 1, config.hidden_size))
self.patch_embeddings = SegGptPatchEmbeddings(config)
num_positions = (config.pretrain_image_size // config.patch_size) ** 2 + 1
self.position_embeddings = nn.Parameter(torch.randn(1, num_positions, config.hidden_size))
self.dropout = nn.Dropout(config.hidden_dropout_prob)
def interpolate_pos_encoding(self, height: int, width: int) -> torch.Tensor:
patch_pos_embed = self.position_embeddings[:, 1:]
num_patches = patch_pos_embed.shape[1]
pretrain_patch_size = int(math.sqrt(num_patches))
if pretrain_patch_size != height or pretrain_patch_size != width:
patch_pos_embed = F.interpolate(
patch_pos_embed.reshape(1, pretrain_patch_size, pretrain_patch_size, -1).permute(0, 3, 1, 2),
size=(height, width),
mode="bicubic",
align_corners=False,
)
return patch_pos_embed.permute(0, 2, 3, 1)
else:
return patch_pos_embed.reshape(1, height, width, -1)
def forward(
self,
pixel_values: torch.Tensor,
prompt_pixel_values: torch.Tensor,
bool_masked_pos: Optional[torch.BoolTensor] = None,
embedding_type: Optional[str] = None,
) -> torch.Tensor:
input_embeddings = self.patch_embeddings(pixel_values)
prompt_embeddings = self.patch_embeddings(prompt_pixel_values)
batch_size, patch_height, patch_width, _ = input_embeddings.shape
mask_token = self.mask_token.expand(batch_size, patch_height, patch_width, -1)
# replace the masked visual tokens by mask_token
w = bool_masked_pos.unsqueeze(-1).type_as(mask_token).reshape(-1, patch_height, patch_width, 1)
prompt_embeddings = prompt_embeddings * (1 - w) + mask_token * w
embedding_type = embedding_type if embedding_type is not None else "instance"
# add positional encoding to each token
pos_embed = self.interpolate_pos_encoding(patch_height, patch_width)
# add segment token
input_embeddings = input_embeddings + self.segment_token_input
prompt_embeddings = prompt_embeddings + self.segment_token_prompt
# add position embedding skipping CLS
input_embeddings = input_embeddings + pos_embed
prompt_embeddings = prompt_embeddings + pos_embed
# add type embedding to each token
if embedding_type == "semantic":
type_embedding = self.type_token_semantic
elif embedding_type == "instance":
type_embedding = self.type_token_instance
else:
raise ValueError(f"Embedding type should be either 'semantic' or 'instance', but got {embedding_type}")
input_embeddings = input_embeddings + type_embedding
prompt_embeddings = prompt_embeddings + type_embedding
embeddings = torch.cat((input_embeddings, prompt_embeddings), dim=0)
return embeddings
class SegGptAttention(nn.Module):
"""Multi-head Attention block with relative position embeddings."""
def __init__(self, config):
super().__init__()
image_size, patch_size = config.image_size, config.patch_size
image_size = image_size if isinstance(image_size, collections.abc.Iterable) else (image_size, image_size)
patch_size = patch_size if isinstance(patch_size, collections.abc.Iterable) else (patch_size, patch_size)
input_size = (image_size[0] // config.patch_size, image_size[1] // config.patch_size)
head_dim = config.hidden_size // config.num_attention_heads
self.num_attention_heads = config.num_attention_heads
self.scale = head_dim**-0.5
self.qkv = nn.Linear(config.hidden_size, config.hidden_size * 3, bias=config.qkv_bias)
self.proj = nn.Linear(config.hidden_size, config.hidden_size)
self.use_relative_position_embeddings = config.use_relative_position_embeddings
if self.use_relative_position_embeddings:
if input_size is None:
raise ValueError("Input size must be provided if using relative positional encoding.")
# initialize relative positional embeddings
self.rel_pos_h = nn.Parameter(torch.zeros(2 * input_size[0] - 1, head_dim))
self.rel_pos_w = nn.Parameter(torch.zeros(2 * input_size[1] - 1, head_dim))
def get_rel_pos(self, q_size: int, k_size: int, rel_pos: torch.Tensor) -> torch.Tensor:
"""
Get relative positional embeddings according to the relative positions of
query and key sizes.
Args:
q_size (int):
size of the query.
k_size (int):
size of key k.
rel_pos (`torch.Tensor`):
relative position embeddings (L, channel).
Returns:
Extracted positional embeddings according to relative positions.
"""
max_rel_dist = int(2 * max(q_size, k_size) - 1)
# Interpolate rel pos.
rel_pos_resized = F.interpolate(
rel_pos.reshape(1, rel_pos.shape[0], -1).permute(0, 2, 1),
size=max_rel_dist,
mode="linear",
)
rel_pos_resized = rel_pos_resized.reshape(-1, max_rel_dist).permute(1, 0)
# Scale the coords with short length if shapes for q and k are different.
q_coords = torch.arange(q_size)[:, None] * max(k_size / q_size, 1.0)
k_coords = torch.arange(k_size)[None, :] * max(q_size / k_size, 1.0)
relative_coords = (q_coords - k_coords) + (k_size - 1) * max(q_size / k_size, 1.0)
return rel_pos_resized[relative_coords.long()]
def add_decomposed_rel_pos(
self,
attn: torch.Tensor,
query: torch.Tensor,
rel_pos_h: torch.Tensor,
rel_pos_w: torch.Tensor,
q_size: Tuple[int, int],
k_size: Tuple[int, int],
) -> torch.Tensor:
"""
Calculate decomposed Relative Positional Embeddings from :paper:`mvitv2`.
https://github.com/facebookresearch/mvit/blob/19786631e330df9f3622e5402b4a419a263a2c80/mvit/models/attention.py
Args:
attn (`torch.Tensor`):
attention map.
query (`torch.Tensor`):
query q in the attention layer with shape (batch_size, query_height * query_width, channel).
rel_pos_h (`torch.Tensor`):
relative position embeddings (Lh, channel) for height axis.
rel_pos_w (`torch.Tensor`):
relative position embeddings (Lw, channel) for width axis.
q_size (tuple):
spatial sequence size of query q with (query_height, query_width).
k_size (tuple):
spatial sequence size of key k with (key_height, key_width).
Returns:
attn (`torch.Tensor`):
attention map with added relative positional embeddings.
"""
query_height, query_width = q_size
key_height, key_width = k_size
relative_position_height = self.get_rel_pos(query_height, key_height, rel_pos_h)
relative_position_width = self.get_rel_pos(query_width, key_width, rel_pos_w)
batch_size, _, dim = query.shape
reshaped_query = query.reshape(batch_size, query_height, query_width, dim)
rel_h = torch.einsum("bhwc,hkc->bhwk", reshaped_query, relative_position_height)
rel_w = torch.einsum("bhwc,wkc->bhwk", reshaped_query, relative_position_width)
attn = attn.reshape(batch_size, query_height, query_width, key_height, key_width)
attn = attn + rel_h[:, :, :, :, None] + rel_w[:, :, :, None, :]
attn = attn.reshape(batch_size, query_height * query_width, key_height * key_width)
return attn
def forward(self, hidden_states: torch.Tensor, output_attentions=False) -> torch.Tensor:
batch_size, height, width, _ = hidden_states.shape
# qkv with shape (3, batch_size, nHead, height * width, channel)
qkv = (
self.qkv(hidden_states)
.reshape(batch_size, height * width, 3, self.num_attention_heads, -1)
.permute(2, 0, 3, 1, 4)
)
# q, k, v with shape (batch_size * nHead, height * width, channel)
query, key, value = qkv.reshape(3, batch_size * self.num_attention_heads, height * width, -1).unbind(0)
attn_weights = (query * self.scale) @ key.transpose(-2, -1)
if self.use_relative_position_embeddings:
attn_weights = self.add_decomposed_rel_pos(
attn_weights, query, self.rel_pos_h, self.rel_pos_w, (height, width), (height, width)
)
attn_weights = torch.nn.functional.softmax(attn_weights, dtype=torch.float32, dim=-1).to(query.dtype)
if output_attentions:
# this operation is a bit awkward, but it's required to
# make sure that attn_weights keeps its gradient.
# In order to do so, attn_weights have to reshaped
# twice and have to be reused in the following
attn_weights_reshaped = attn_weights.view(batch_size, self.num_attention_heads, height * width, -1)
attn_weights = attn_weights_reshaped.view(batch_size * self.num_attention_heads, height * width, -1)
else:
attn_weights_reshaped = None
attn_output = (attn_weights @ value).reshape(batch_size, self.num_attention_heads, height, width, -1)
attn_output = attn_output.permute(0, 2, 3, 1, 4).reshape(batch_size, height, width, -1)
attn_output = self.proj(attn_output)
return (attn_output, attn_weights_reshaped)
# Copied from transformers.models.sam.modeling_sam.SamMLPBlock with SamMLPBlock->SegGptMlp
class SegGptMlp(nn.Module):
def __init__(self, config):
super().__init__()
self.lin1 = nn.Linear(config.hidden_size, config.mlp_dim)
self.lin2 = nn.Linear(config.mlp_dim, config.hidden_size)
self.act = ACT2FN[config.hidden_act]
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
hidden_states = self.lin1(hidden_states)
hidden_states = self.act(hidden_states)
hidden_states = self.lin2(hidden_states)
return hidden_states
# Copied from transformers.models.beit.modeling_beit.drop_path
def drop_path(input: torch.Tensor, drop_prob: float = 0.0, training: bool = False) -> torch.Tensor:
"""
Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
Comment by Ross Wightman: This is the same as the DropConnect impl I created for EfficientNet, etc networks,
however, the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...
See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for changing the
layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use 'survival rate' as the
argument.
"""
if drop_prob == 0.0 or not training:
return input
keep_prob = 1 - drop_prob
shape = (input.shape[0],) + (1,) * (input.ndim - 1) # work with diff dim tensors, not just 2D ConvNets
random_tensor = keep_prob + torch.rand(shape, dtype=input.dtype, device=input.device)
random_tensor.floor_() # binarize
output = input.div(keep_prob) * random_tensor
return output
# Copied from transformers.models.beit.modeling_beit.BeitDropPath with Beit->SegGpt
class SegGptDropPath(nn.Module):
"""Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks)."""
def __init__(self, drop_prob: Optional[float] = None) -> None:
super().__init__()
self.drop_prob = drop_prob
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
return drop_path(hidden_states, self.drop_prob, self.training)
def extra_repr(self) -> str:
return "p={}".format(self.drop_prob)
class SegGptLayer(nn.Module):
def __init__(self, config: SegGptConfig, drop_path_rate: float) -> None:
super().__init__()
self.attention = SegGptAttention(config)
self.mlp = SegGptMlp(config)
self.drop_path = SegGptDropPath(drop_path_rate) if drop_path_rate > 0.0 else nn.Identity()
self.layernorm_before = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
self.layernorm_after = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
def forward(
self,
hidden_states: torch.Tensor,
ensemble_cond: int,
feature_ensemble: bool = False,
output_attentions: bool = False,
) -> Union[Tuple[torch.Tensor, torch.Tensor], Tuple[torch.Tensor]]:
self_attention_outputs = self.attention(
self.layernorm_before(hidden_states), # in SegGpt, layernorm is applied before self-attention
output_attentions=output_attentions,
)
attention_output = self_attention_outputs[0]
outputs = self_attention_outputs[1:] # add self attentions if we output attention weights
if feature_ensemble and attention_output.shape[0] // 2 >= ensemble_cond:
prompt, inputs = attention_output.split(attention_output.shape[1] // 2, dim=1)
if ensemble_cond == 2:
num_prompts = attention_output.shape[0] // 2
inputs = inputs.reshape(2, num_prompts, -1)
inputs = inputs.mean(dim=1, keepdim=True).expand_as(inputs)
inputs = inputs.reshape(*prompt.shape)
else:
inputs = inputs.mean(dim=0, keepdim=True).expand_as(inputs)
attention_output = torch.cat([prompt, inputs], dim=1)
# first residual connection
hidden_states = self.drop_path(attention_output) + hidden_states
residual = hidden_states
hidden_states = self.layernorm_after(hidden_states)
hidden_states = self.mlp(hidden_states)
hidden_states = residual + self.drop_path(hidden_states)
outputs = (hidden_states,) + outputs
return outputs
class SegGptEncoder(nn.Module):
def __init__(self, config: SegGptConfig) -> None:
super().__init__()
self.config = config
dpr = [x.item() for x in torch.linspace(0, config.drop_path_rate, config.num_hidden_layers)]
self.layers = nn.ModuleList([SegGptLayer(config, dpr[i]) for i in range(config.num_hidden_layers)])
self.layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
self.gradient_checkpointing = False
def forward(
self,
hidden_states: torch.Tensor,
feature_ensemble: bool = False,
output_attentions: bool = False,
output_hidden_states: bool = False,
return_dict: bool = True,
) -> Union[tuple, SegGptEncoderOutput]:
all_hidden_states = () if output_hidden_states else None
all_self_attentions = () if output_attentions else None
intermediate_hidden_states = []
for i, layer_module in enumerate(self.layers):
if output_hidden_states:
all_hidden_states = all_hidden_states + (hidden_states,)
# Condition to check if we have the appropriate number of prompts to ensemble
ensemble_cond = 2 if self.config.merge_index > i else 1
if self.gradient_checkpointing and self.training:
layer_outputs = self._gradient_checkpointing_func(
layer_module.__call__,
hidden_states,
ensemble_cond,
feature_ensemble,
output_attentions,
)
else:
layer_outputs = layer_module(hidden_states, ensemble_cond, feature_ensemble, output_attentions)
hidden_states = layer_outputs[0]
if i == self.config.merge_index:
hidden_states = (
hidden_states[: hidden_states.shape[0] // 2] + hidden_states[hidden_states.shape[0] // 2 :]
) * 0.5
if i in self.config.intermediate_hidden_state_indices:
intermediate_hidden_states.append(self.layernorm(hidden_states))
if output_attentions:
all_self_attentions = all_self_attentions + (layer_outputs[1],)
if output_hidden_states:
all_hidden_states = all_hidden_states + (hidden_states,)
if not return_dict:
return tuple(
v
for v in [hidden_states, all_hidden_states, all_self_attentions, intermediate_hidden_states]
if v is not None
)
return SegGptEncoderOutput(
last_hidden_state=hidden_states,
hidden_states=all_hidden_states,
attentions=all_self_attentions,
intermediate_hidden_states=intermediate_hidden_states,
)
# Copied from transformers.models.convnext.modeling_convnext.ConvNextLayerNorm with ConvNext->SegGpt
class SegGptLayerNorm(nn.Module):
r"""LayerNorm that supports two data formats: channels_last (default) or channels_first.
The ordering of the dimensions in the inputs. channels_last corresponds to inputs with shape (batch_size, height,
width, channels) while channels_first corresponds to inputs with shape (batch_size, channels, height, width).
"""
def __init__(self, normalized_shape, eps=1e-6, data_format="channels_last"):
super().__init__()
self.weight = nn.Parameter(torch.ones(normalized_shape))
self.bias = nn.Parameter(torch.zeros(normalized_shape))
self.eps = eps
self.data_format = data_format
if self.data_format not in ["channels_last", "channels_first"]:
raise NotImplementedError(f"Unsupported data format: {self.data_format}")
self.normalized_shape = (normalized_shape,)
def forward(self, x: torch.Tensor) -> torch.Tensor:
if self.data_format == "channels_last":
x = torch.nn.functional.layer_norm(x, self.normalized_shape, self.weight, self.bias, self.eps)
elif self.data_format == "channels_first":
input_dtype = x.dtype
x = x.float()
u = x.mean(1, keepdim=True)
s = (x - u).pow(2).mean(1, keepdim=True)
x = (x - u) / torch.sqrt(s + self.eps)
x = x.to(dtype=input_dtype)
x = self.weight[:, None, None] * x + self.bias[:, None, None]
return x
class SegGptDecoderHead(nn.Module):
def __init__(self, config):
super().__init__()
self.conv = nn.Conv2d(
config.decoder_hidden_size,
config.decoder_hidden_size,
kernel_size=3,
padding=1,
)
self.layernorm = SegGptLayerNorm(
normalized_shape=config.decoder_hidden_size, eps=config.layer_norm_eps, data_format="channels_first"
)
self.act_fct = ACT2FN[config.hidden_act]
self.head = nn.Conv2d(config.decoder_hidden_size, 3, kernel_size=1, bias=True) # decoder to patch
def forward(self, hidden_states: torch.FloatTensor):
hidden_states = self.conv(hidden_states)
hidden_states = self.layernorm(hidden_states)
hidden_states = self.act_fct(hidden_states)
hidden_states = self.head(hidden_states)
return hidden_states
class SegGptDecoder(nn.Module):
def __init__(self, config):
super().__init__()
self.decoder_embed = nn.Linear(
config.hidden_size * len(config.intermediate_hidden_state_indices),
config.patch_size**2 * config.decoder_hidden_size,
bias=True,
)
self.decoder_pred = SegGptDecoderHead(config)
self.patch_size = config.patch_size
self.decoder_hidden_size = config.decoder_hidden_size
self.config = config
def _reshape_hidden_states(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:
batch_size, patch_height, patch_width, _ = hidden_states.shape
hidden_states = hidden_states.reshape(
batch_size, patch_height, patch_width, self.patch_size, self.patch_size, self.decoder_hidden_size
)
hidden_states = hidden_states.permute(0, 5, 1, 3, 2, 4)
hidden_states = hidden_states.reshape(
shape=(batch_size, -1, patch_height * self.patch_size, patch_width * self.patch_size)
)
return hidden_states
def forward(self, hidden_states: torch.FloatTensor):
hidden_states = self.decoder_embed(hidden_states)
hidden_states = self._reshape_hidden_states(hidden_states)
hidden_states = self.decoder_pred(hidden_states)
return hidden_states
class SegGptPreTrainedModel(PreTrainedModel):
"""
An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
models.
"""
config_class = SegGptConfig
base_model_prefix = "model"
main_input_name = "pixel_values"
supports_gradient_checkpointing = True
_no_split_modules = ["SegGptEmbeddings", "SegGptLayer"]
def _init_weights(self, module: Union[nn.Linear, nn.Conv2d, nn.LayerNorm]) -> None:
"""Initialize the weights"""
std = self.config.initializer_range
if isinstance(module, (nn.Linear, nn.Conv2d)):
# Upcast the input in `fp32` and cast it back to desired `dtype` to avoid
# `trunc_normal_cpu` not implemented in `half` issues
module.weight.data = nn.init.trunc_normal_(module.weight.data.to(torch.float32), mean=0.0, std=std).to(
module.weight.dtype
)
if module.bias is not None:
module.bias.data.zero_()
elif isinstance(module, nn.LayerNorm):
module.bias.data.zero_()
module.weight.data.fill_(1.0)
elif isinstance(module, SegGptAttention):
module.rel_pos_h.data = nn.init.trunc_normal_(
module.rel_pos_h.data.to(torch.float32),
mean=0.0,
std=std,
).to(module.rel_pos_h.dtype)
module.rel_pos_w.data = nn.init.trunc_normal_(
module.rel_pos_w.data.to(torch.float32),
mean=0.0,
std=std,
).to(module.rel_pos_w.dtype)
elif isinstance(module, SegGptEmbeddings):
module.position_embeddings.data = nn.init.trunc_normal_(
module.position_embeddings.data.to(torch.float32),
mean=0.0,
std=std,
).to(module.position_embeddings.dtype)
torch.nn.init.normal_(module.mask_token, std=std)
torch.nn.init.normal_(module.segment_token_input, std=std)
torch.nn.init.normal_(module.segment_token_prompt, std=std)
torch.nn.init.normal_(module.type_token_semantic, std=std)
torch.nn.init.normal_(module.type_token_instance, std=std)
SEGGPT_START_DOCSTRING = r"""
This model is a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass. Use it
as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and
behavior.
Parameters:
config ([`SegGptConfig`]): Model configuration class with all the parameters of the model.
Initializing with a config file does not load the weights associated with the model, only the
configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
"""
SEGGPT_INPUTS_DOCSTRING = r"""
Args:
pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
Pixel values. Pixel values can be obtained using [`AutoImageProcessor`]. See [`SegGptImageProcessor.__call__`]
for details.
prompt_pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
Prompt pixel values. Prompt pixel values can be obtained using [`AutoImageProcessor`]. See
[`SegGptImageProcessor.__call__`] for details.
prompt_masks (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
Prompt mask. Prompt mask can be obtained using [`AutoImageProcessor`]. See [`SegGptImageProcessor.__call__`] for
details.
bool_masked_pos (`torch.BoolTensor` of shape `(batch_size, num_patches)`, *optional*):
Boolean masked positions. Indicates which patches are masked (1) and which aren't (0).
feature_ensemble (`bool`, *optional*):
Boolean indicating whether to use feature ensemble or not. If `True`, the model will use feature ensemble
if we have at least two prompts. If `False`, the model will not use feature ensemble. This argument should
be considered when doing few-shot inference on an input image i.e. more than one prompt for the same image.
embedding_type (`str`, *optional*):
Embedding type. Indicates whether the prompt is a semantic or instance embedding. Can be either
instance or semantic.
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
tensors for more detail.
output_hidden_states (`bool`, *optional*):
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
more detail.
return_dict (`bool`, *optional*):
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
"""
@add_start_docstrings(
"The bare SegGpt Model transformer outputting raw hidden-states without any specific head on top.",
SEGGPT_START_DOCSTRING,
)
class SegGptModel(SegGptPreTrainedModel):
def __init__(self, config: SegGptConfig):
super().__init__(config)
self.config = config
self.embeddings = SegGptEmbeddings(config)
self.encoder = SegGptEncoder(config)
# Initialize weights and apply final processing
self.post_init()
def get_input_embeddings(self) -> SegGptPatchEmbeddings:
return self.embeddings.patch_embeddings
def _prune_heads(self, heads_to_prune: Dict[int, List[int]]) -> None:
"""
Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
class PreTrainedModel
"""
for layer, heads in heads_to_prune.items():
self.encoder.layer[layer].attention.prune_heads(heads)
@add_start_docstrings_to_model_forward(SEGGPT_INPUTS_DOCSTRING)
@replace_return_docstrings(output_type=SegGptEncoderOutput, config_class=_CONFIG_FOR_DOC)
def forward(
self,
pixel_values: torch.Tensor,
prompt_pixel_values: torch.Tensor,
prompt_masks: torch.Tensor,
bool_masked_pos: Optional[torch.BoolTensor] = None,
feature_ensemble: Optional[bool] = None,
embedding_type: Optional[str] = None,
labels: Optional[torch.FloatTensor] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
) -> Union[Tuple, SegGptEncoderOutput]:
r"""
labels (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`, `optional`):
Ground truth mask for input images.
Returns:
Examples:
```python
>>> from transformers import SegGptImageProcessor, SegGptModel
>>> from PIL import Image
>>> import requests
>>> image_input_url = "https://raw.githubusercontent.com/baaivision/Painter/main/SegGPT/SegGPT_inference/examples/hmbb_2.jpg"
>>> image_prompt_url = "https://raw.githubusercontent.com/baaivision/Painter/main/SegGPT/SegGPT_inference/examples/hmbb_1.jpg"
>>> mask_prompt_url = "https://raw.githubusercontent.com/baaivision/Painter/main/SegGPT/SegGPT_inference/examples/hmbb_1_target.png"
>>> image_input = Image.open(requests.get(image_input_url, stream=True).raw)
>>> image_prompt = Image.open(requests.get(image_prompt_url, stream=True).raw)
>>> mask_prompt = Image.open(requests.get(mask_prompt_url, stream=True).raw).convert("L")
>>> checkpoint = "BAAI/seggpt-vit-large"
>>> model = SegGptModel.from_pretrained(checkpoint)
>>> image_processor = SegGptImageProcessor.from_pretrained(checkpoint)
>>> inputs = image_processor(images=image_input, prompt_images=image_prompt, prompt_masks=mask_prompt, return_tensors="pt")
>>> outputs = model(**inputs)
>>> list(outputs.last_hidden_state.shape)
[1, 56, 28, 1024]
```
"""
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
feature_ensemble = feature_ensemble if feature_ensemble is not None else False
expected_dtype = self.embeddings.patch_embeddings.projection.weight.dtype
pixel_values = pixel_values.to(expected_dtype)
prompt_pixel_values = prompt_pixel_values.to(expected_dtype)
# Prepare inputs
pixel_values = torch.cat((prompt_pixel_values, pixel_values), dim=2)
prompt_pixel_values = (
torch.cat((prompt_masks, prompt_masks), dim=2)
if labels is None
else torch.cat((prompt_masks, labels), dim=2)
)
if bool_masked_pos is None and labels is not None:
logger.warning_once(
"Labels were provided, but bool_masked_pos were not. It will be set to default value. If you're training the model, make sure to provide a bool_masked_pos."
)
# We concat on height axis so SegGPT can handle as a single image, hence we need to mask the portion
# of the mask prompt pixels that will be destinated to the prediction as they don't add any information.
# This is only the case for inference. In training, the model concat of prompt mask and label is masked
# and reconstructed together (In-Context Painting).
if bool_masked_pos is None:
num_patches = self.embeddings.patch_embeddings.num_patches
bool_masked_pos = torch.zeros(num_patches, dtype=torch.bool).to(pixel_values.device)
bool_masked_pos[num_patches // 2 :] = 1
bool_masked_pos = bool_masked_pos.unsqueeze(0)
embedding_output = self.embeddings(
pixel_values, prompt_pixel_values, embedding_type=embedding_type, bool_masked_pos=bool_masked_pos
)
encoder_outputs = self.encoder(
embedding_output,
feature_ensemble=feature_ensemble,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
return encoder_outputs
def patchify(tensor: torch.Tensor, patch_size: int) -> torch.Tensor:
batch_size, num_channels, height, width = tensor.shape
patch_height = height // patch_size
patch_width = width // patch_size
tensor = tensor.reshape(shape=(batch_size, num_channels, patch_height, patch_size, patch_width, patch_size))
tensor = tensor.permute(0, 2, 4, 3, 5, 1)
tensor = tensor.reshape(shape=(batch_size, patch_height * patch_width, patch_size**2 * 3))
return tensor
def unpatchify(tensor: torch.Tensor, patch_height: int, patch_width: int) -> torch.Tensor:
batch_size = tensor.shape[0]
patch_size = int((tensor.shape[-1] / 3) ** 0.5)
if patch_height * patch_width != tensor.shape[1]:
raise ValueError(
f"Number of patches {tensor.shape[1]} does not match patch height ({patch_height}) and width ({patch_width})."
)
tensor = tensor.reshape(shape=(batch_size, patch_height, patch_width, patch_size, patch_size, 3))
tensor = tensor.permute(0, 5, 1, 3, 2, 4)
tensor = tensor.reshape(shape=(batch_size, 3, patch_height * patch_size, patch_width * patch_size))
return tensor
class SegGptLoss(nn.Module):
def __init__(self, config):
super().__init__()
self.beta = config.beta
self.patch_size = config.patch_size
def forward(
self,
prompt_masks: torch.FloatTensor,
pred_masks: torch.FloatTensor,
labels: torch.FloatTensor,
bool_masked_pos: torch.BoolTensor,
):
"""Computes the L1 loss between the predicted masks and the ground truth masks.
Args:
prompt_masks (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
Pixel values from mask prompt.
pred_masks (`torch.FloatTensor` of shape `(batch_size, num_channels, 2*height, width)`):
Predicted masks.
labels (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
Ground truth mask for input images.
bool_masked_pos (`torch.BoolTensor` of shape `(batch_size, num_patches)`):
Boolean masked positions. Indicates which patches are masked (1) and which aren't (0).
Returns:
`torch.FloatTensor`: The mean L1 loss between the predicted masks and the ground truth masks.
"""
ground_truth = torch.cat((prompt_masks, labels), dim=2)
mask = bool_masked_pos[:, :, None].repeat(1, 1, self.patch_size**2 * 3)
mask = unpatchify(mask, ground_truth.shape[2] // self.patch_size, ground_truth.shape[3] // self.patch_size)
loss = F.smooth_l1_loss(pred_masks, ground_truth, reduction="none", beta=self.beta)
loss = (loss * mask).sum() / mask.sum() # mean loss on removed patches
return loss
@add_start_docstrings(
"SegGpt model with a decoder on top for one-shot image segmentation.",
SEGGPT_START_DOCSTRING,
)
class SegGptForImageSegmentation(SegGptPreTrainedModel):
def __init__(self, config: SegGptConfig):
super().__init__(config)
self.config = config
self.model = SegGptModel(config)
self.decoder = SegGptDecoder(config)
# Initialize weights and apply final processing
self.post_init()
@add_start_docstrings_to_model_forward(SEGGPT_INPUTS_DOCSTRING)
@replace_return_docstrings(output_type=SegGptImageSegmentationOutput, config_class=_CONFIG_FOR_DOC)
def forward(
self,
pixel_values: torch.Tensor,
prompt_pixel_values: torch.Tensor,
prompt_masks: torch.Tensor,
bool_masked_pos: Optional[torch.BoolTensor] = None,
feature_ensemble: Optional[bool] = None,
embedding_type: Optional[str] = None,
labels: Optional[torch.FloatTensor] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
) -> Union[Tuple, SegGptImageSegmentationOutput]:
r"""
labels (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`, `optional`):
Ground truth mask for input images.
Returns:
Examples:
```python
>>> from transformers import SegGptImageProcessor, SegGptForImageSegmentation
>>> from PIL import Image
>>> import requests
>>> image_input_url = "https://raw.githubusercontent.com/baaivision/Painter/main/SegGPT/SegGPT_inference/examples/hmbb_2.jpg"
>>> image_prompt_url = "https://raw.githubusercontent.com/baaivision/Painter/main/SegGPT/SegGPT_inference/examples/hmbb_1.jpg"
>>> mask_prompt_url = "https://raw.githubusercontent.com/baaivision/Painter/main/SegGPT/SegGPT_inference/examples/hmbb_1_target.png"
>>> image_input = Image.open(requests.get(image_input_url, stream=True).raw)
>>> image_prompt = Image.open(requests.get(image_prompt_url, stream=True).raw)
>>> mask_prompt = Image.open(requests.get(mask_prompt_url, stream=True).raw).convert("L")
>>> checkpoint = "BAAI/seggpt-vit-large"
>>> model = SegGptForImageSegmentation.from_pretrained(checkpoint)
>>> image_processor = SegGptImageProcessor.from_pretrained(checkpoint)
>>> inputs = image_processor(images=image_input, prompt_images=image_prompt, prompt_masks=mask_prompt, return_tensors="pt")
>>> outputs = model(**inputs)
>>> result = image_processor.post_process_semantic_segmentation(outputs, target_sizes=[image_input.size[::-1]])[0]
>>> print(list(result.shape))
[170, 297]
```
"""
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 bool_masked_pos is None:
num_patches = self.model.embeddings.patch_embeddings.num_patches
bool_masked_pos = torch.zeros(num_patches, dtype=torch.bool).to(pixel_values.device)
bool_masked_pos[num_patches // 2 :] = 1
bool_masked_pos = bool_masked_pos.unsqueeze(0)
outputs = self.model(
pixel_values=pixel_values,
prompt_pixel_values=prompt_pixel_values,
prompt_masks=prompt_masks,
bool_masked_pos=bool_masked_pos,
feature_ensemble=feature_ensemble,
embedding_type=embedding_type,
labels=labels,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
intermediate_hidden_states = outputs.intermediate_hidden_states if return_dict else outputs[-1]
intermediate_hidden_states = torch.cat(intermediate_hidden_states, dim=-1)
pred_masks = self.decoder(intermediate_hidden_states)
loss = None
if labels is not None:
loss_fn = SegGptLoss(self.config)
loss = loss_fn(prompt_masks, pred_masks, labels, bool_masked_pos)
if not return_dict:
output = (pred_masks,)
if output_hidden_states:
output = output + (outputs[1],)
if output_attentions:
idx = 2 if output_hidden_states else 1
output = output + (outputs[idx],)
if loss is not None:
output = (loss,) + output
return output
return SegGptImageSegmentationOutput(
loss=loss,
pred_masks=pred_masks,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
|
transformers/src/transformers/models/seggpt/modeling_seggpt.py/0
|
{
"file_path": "transformers/src/transformers/models/seggpt/modeling_seggpt.py",
"repo_id": "transformers",
"token_count": 19315
}
| 382
|
# Copyright 2021 The HuggingFace Team. 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.
from typing import TYPE_CHECKING
from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_flax_available, is_torch_available
_import_structure = {"configuration_speech_encoder_decoder": ["SpeechEncoderDecoderConfig"]}
try:
if not is_torch_available():
raise OptionalDependencyNotAvailable()
except OptionalDependencyNotAvailable:
pass
else:
_import_structure["modeling_speech_encoder_decoder"] = ["SpeechEncoderDecoderModel"]
try:
if not is_flax_available():
raise OptionalDependencyNotAvailable()
except OptionalDependencyNotAvailable:
pass
else:
_import_structure["modeling_flax_speech_encoder_decoder"] = ["FlaxSpeechEncoderDecoderModel"]
if TYPE_CHECKING:
from .configuration_speech_encoder_decoder import SpeechEncoderDecoderConfig
try:
if not is_torch_available():
raise OptionalDependencyNotAvailable()
except OptionalDependencyNotAvailable:
pass
else:
from .modeling_speech_encoder_decoder import SpeechEncoderDecoderModel
try:
if not is_flax_available():
raise OptionalDependencyNotAvailable()
except OptionalDependencyNotAvailable:
pass
else:
from .modeling_flax_speech_encoder_decoder import FlaxSpeechEncoderDecoderModel
else:
import sys
sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)
|
transformers/src/transformers/models/speech_encoder_decoder/__init__.py/0
|
{
"file_path": "transformers/src/transformers/models/speech_encoder_decoder/__init__.py",
"repo_id": "transformers",
"token_count": 678
}
| 383
|
# coding=utf-8
# Copyright 2023 The HuggingFace Inc. team. 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.
"""Convert SpeechT5 HiFi-GAN checkpoint."""
import argparse
import numpy as np
import torch
from transformers import SpeechT5HifiGan, SpeechT5HifiGanConfig, logging
logging.set_verbosity_info()
logger = logging.get_logger("transformers.models.speecht5")
def load_weights(checkpoint, hf_model, config):
hf_model.apply_weight_norm()
hf_model.conv_pre.weight_g.data = checkpoint["input_conv.weight_g"]
hf_model.conv_pre.weight_v.data = checkpoint["input_conv.weight_v"]
hf_model.conv_pre.bias.data = checkpoint["input_conv.bias"]
for i in range(len(config.upsample_rates)):
hf_model.upsampler[i].weight_g.data = checkpoint[f"upsamples.{i}.1.weight_g"]
hf_model.upsampler[i].weight_v.data = checkpoint[f"upsamples.{i}.1.weight_v"]
hf_model.upsampler[i].bias.data = checkpoint[f"upsamples.{i}.1.bias"]
for i in range(len(config.upsample_rates) * len(config.resblock_kernel_sizes)):
for j in range(len(config.resblock_dilation_sizes)):
hf_model.resblocks[i].convs1[j].weight_g.data = checkpoint[f"blocks.{i}.convs1.{j}.1.weight_g"]
hf_model.resblocks[i].convs1[j].weight_v.data = checkpoint[f"blocks.{i}.convs1.{j}.1.weight_v"]
hf_model.resblocks[i].convs1[j].bias.data = checkpoint[f"blocks.{i}.convs1.{j}.1.bias"]
hf_model.resblocks[i].convs2[j].weight_g.data = checkpoint[f"blocks.{i}.convs2.{j}.1.weight_g"]
hf_model.resblocks[i].convs2[j].weight_v.data = checkpoint[f"blocks.{i}.convs2.{j}.1.weight_v"]
hf_model.resblocks[i].convs2[j].bias.data = checkpoint[f"blocks.{i}.convs2.{j}.1.bias"]
hf_model.conv_post.weight_g.data = checkpoint["output_conv.1.weight_g"]
hf_model.conv_post.weight_v.data = checkpoint["output_conv.1.weight_v"]
hf_model.conv_post.bias.data = checkpoint["output_conv.1.bias"]
hf_model.remove_weight_norm()
@torch.no_grad()
def convert_hifigan_checkpoint(
checkpoint_path,
stats_path,
pytorch_dump_folder_path,
config_path=None,
repo_id=None,
):
if config_path is not None:
config = SpeechT5HifiGanConfig.from_pretrained(config_path)
else:
config = SpeechT5HifiGanConfig()
model = SpeechT5HifiGan(config)
orig_checkpoint = torch.load(checkpoint_path)
load_weights(orig_checkpoint["model"]["generator"], model, config)
stats = np.load(stats_path)
mean = stats[0].reshape(-1)
scale = stats[1].reshape(-1)
model.mean = torch.from_numpy(mean).float()
model.scale = torch.from_numpy(scale).float()
model.save_pretrained(pytorch_dump_folder_path)
if repo_id:
print("Pushing to the hub...")
model.push_to_hub(repo_id)
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--checkpoint_path", required=True, default=None, type=str, help="Path to original checkpoint")
parser.add_argument("--stats_path", required=True, default=None, type=str, help="Path to stats.npy file")
parser.add_argument("--config_path", default=None, type=str, help="Path to hf config.json of model to convert")
parser.add_argument(
"--pytorch_dump_folder_path", required=True, default=None, type=str, help="Path to the output PyTorch model."
)
parser.add_argument(
"--push_to_hub", default=None, type=str, help="Where to upload the converted model on the 🤗 hub."
)
args = parser.parse_args()
convert_hifigan_checkpoint(
args.checkpoint_path,
args.stats_path,
args.pytorch_dump_folder_path,
args.config_path,
args.push_to_hub,
)
|
transformers/src/transformers/models/speecht5/convert_hifigan.py/0
|
{
"file_path": "transformers/src/transformers/models/speecht5/convert_hifigan.py",
"repo_id": "transformers",
"token_count": 1706
}
| 384
|
# coding=utf-8
# Copyright 2024 MBZUAI and The HuggingFace Inc. team. 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.
"""TensorFlow SwiftFormer model."""
import collections.abc
from typing import Optional, Tuple, Union
import tensorflow as tf
from ...activations_tf import get_tf_activation
from ...modeling_tf_outputs import (
TFBaseModelOutputWithNoAttention,
TFImageClassifierOutputWithNoAttention,
)
from ...modeling_tf_utils import TFPreTrainedModel, keras, keras_serializable, unpack_inputs
from ...utils import (
add_start_docstrings,
add_start_docstrings_to_model_forward,
logging,
)
from .configuration_swiftformer import SwiftFormerConfig
logger = logging.get_logger(__name__)
# General docstring
_CONFIG_FOR_DOC = "SwiftFormerConfig"
# Base docstring
_CHECKPOINT_FOR_DOC = "MBZUAI/swiftformer-xs"
_EXPECTED_OUTPUT_SHAPE = [1, 220, 7, 7]
# Image classification docstring
_IMAGE_CLASS_CHECKPOINT = "MBZUAI/swiftformer-xs"
_IMAGE_CLASS_EXPECTED_OUTPUT = "tabby, tabby cat"
class TFSwiftFormerPatchEmbeddingSequential(keras.layers.Layer):
"""
The sequential component of the patch embedding layer.
Input: tensor of shape `[batch_size, in_channels, height, width]`
Output: tensor of shape `[batch_size, out_channels, height/4, width/4]`
"""
def __init__(self, config: SwiftFormerConfig, **kwargs):
super().__init__(**kwargs)
self.out_chs = config.embed_dims[0]
self.zero_padding = keras.layers.ZeroPadding2D(padding=(1, 1))
self.conv1 = keras.layers.Conv2D(self.out_chs // 2, kernel_size=3, strides=2, name="0")
self.batch_norm1 = keras.layers.BatchNormalization(epsilon=config.batch_norm_eps, momentum=0.9, name="1")
self.conv2 = keras.layers.Conv2D(self.out_chs, kernel_size=3, strides=2, name="3")
self.batch_norm2 = keras.layers.BatchNormalization(epsilon=config.batch_norm_eps, momentum=0.9, name="4")
self.config = config
def call(self, x: tf.Tensor, training: bool = False) -> tf.Tensor:
x = self.zero_padding(x)
x = self.conv1(x)
x = self.batch_norm1(x, training=training)
x = get_tf_activation("relu")(x)
x = self.zero_padding(x)
x = self.conv2(x)
x = self.batch_norm2(x, training=training)
x = get_tf_activation("relu")(x)
return x
def build(self, input_shape=None):
if self.built:
return
if getattr(self, "conv1", None) is not None:
with tf.name_scope(self.conv1.name):
self.conv1.build(self.config.num_channels)
if getattr(self, "batch_norm1", None) is not None:
with tf.name_scope(self.batch_norm1.name):
self.batch_norm1.build((None, None, None, self.out_chs // 2))
if getattr(self, "conv2", None) is not None:
with tf.name_scope(self.conv2.name):
self.conv2.build((None, None, None, self.out_chs // 2))
if getattr(self, "batch_norm2", None) is not None:
with tf.name_scope(self.batch_norm2.name):
self.batch_norm2.build((None, None, None, self.out_chs))
self.built = True
class TFSwiftFormerPatchEmbedding(keras.layers.Layer):
"""
Patch Embedding Layer constructed of two 2D convolutional layers.
Input: tensor of shape `[batch_size, in_channels, height, width]`
Output: tensor of shape `[batch_size, out_channels, height/4, width/4]`
"""
def __init__(self, config: SwiftFormerConfig, **kwargs):
super().__init__(**kwargs)
self.patch_embedding = TFSwiftFormerPatchEmbeddingSequential(config, name="patch_embedding")
def call(self, x: tf.Tensor, training: bool = False) -> tf.Tensor:
return self.patch_embedding(x, training=training)
def build(self, input_shape=None):
if self.built:
return
if getattr(self, "patch_embedding", None) is not None:
with tf.name_scope(self.patch_embedding.name):
self.patch_embedding.build(None)
self.built = True
class TFSwiftFormerDropPath(keras.layers.Layer):
"""Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks)."""
def __init__(self, config: SwiftFormerConfig, **kwargs) -> None:
super().__init__(**kwargs)
raise NotImplementedError("Drop path is not implemented in TF port")
def call(self, hidden_states: tf.Tensor, training: bool = False) -> tf.Tensor:
raise NotImplementedError("Drop path is not implemented in TF port")
class TFSwiftFormerEmbeddings(keras.layers.Layer):
"""
Embeddings layer consisting of a single 2D convolutional and batch normalization layer.
Input: tensor of shape `[batch_size, channels, height, width]`
Output: tensor of shape `[batch_size, channels, height/stride, width/stride]`
"""
def __init__(self, config: SwiftFormerConfig, index: int, **kwargs):
super().__init__(**kwargs)
patch_size = config.down_patch_size
stride = config.down_stride
padding = config.down_pad
embed_dims = config.embed_dims
self.in_chans = embed_dims[index]
self.embed_dim = embed_dims[index + 1]
patch_size = patch_size if isinstance(patch_size, collections.abc.Iterable) else (patch_size, patch_size)
stride = stride if isinstance(stride, collections.abc.Iterable) else (stride, stride)
padding = padding if isinstance(padding, collections.abc.Iterable) else (padding, padding)
self.pad = keras.layers.ZeroPadding2D(padding=padding)
self.proj = keras.layers.Conv2D(self.embed_dim, kernel_size=patch_size, strides=stride, name="proj")
self.norm = keras.layers.BatchNormalization(epsilon=config.batch_norm_eps, momentum=0.9, name="norm")
def call(self, x: tf.Tensor, training: bool = False) -> tf.Tensor:
x = self.pad(x)
x = self.proj(x)
x = self.norm(x, training=training)
return x
def build(self, input_shape=None):
if self.built:
return
if getattr(self, "proj", None) is not None:
with tf.name_scope(self.proj.name):
self.proj.build(self.in_chans)
if getattr(self, "norm", None) is not None:
with tf.name_scope(self.norm.name):
self.norm.build((None, None, None, self.embed_dim))
self.built = True
class TFSwiftFormerConvEncoder(keras.layers.Layer):
"""
`SwiftFormerConvEncoder` with 3*3 and 1*1 convolutions.
Input: tensor of shape `[batch_size, channels, height, width]`
Output: tensor of shape `[batch_size, channels, height, width]`
"""
def __init__(self, config: SwiftFormerConfig, dim: int, **kwargs):
super().__init__(**kwargs)
hidden_dim = int(config.mlp_ratio * dim)
self.dim = dim
self.pad = keras.layers.ZeroPadding2D(padding=(1, 1))
self.depth_wise_conv = keras.layers.Conv2D(dim, kernel_size=3, groups=dim, name="depth_wise_conv")
self.norm = keras.layers.BatchNormalization(epsilon=config.batch_norm_eps, momentum=0.9, name="norm")
self.point_wise_conv1 = keras.layers.Conv2D(hidden_dim, kernel_size=1, name="point_wise_conv1")
self.act = get_tf_activation("gelu")
self.point_wise_conv2 = keras.layers.Conv2D(dim, kernel_size=1, name="point_wise_conv2")
self.drop_path = keras.layers.Dropout(name="drop_path", rate=config.drop_conv_encoder_rate)
self.hidden_dim = int(config.mlp_ratio * self.dim)
def build(self, input_shape=None):
if self.built:
return
self.layer_scale = self.add_weight(
name="layer_scale",
shape=self.dim,
initializer="ones",
trainable=True,
)
if getattr(self, "depth_wise_conv", None) is not None:
with tf.name_scope(self.depth_wise_conv.name):
self.depth_wise_conv.build(self.dim)
if getattr(self, "norm", None) is not None:
with tf.name_scope(self.norm.name):
self.norm.build((None, None, None, self.dim))
if getattr(self, "point_wise_conv1", None) is not None:
with tf.name_scope(self.point_wise_conv1.name):
self.point_wise_conv1.build(self.dim)
if getattr(self, "point_wise_conv2", None) is not None:
with tf.name_scope(self.point_wise_conv2.name):
self.point_wise_conv2.build(self.hidden_dim)
if getattr(self, "drop_path", None) is not None:
with tf.name_scope(self.drop_path.name):
self.drop_path.build(None)
self.built = True
def call(self, x: tf.Tensor, training: bool = False) -> tf.Tensor:
input = x
x = self.pad(x)
x = self.depth_wise_conv(x)
x = self.norm(x, training=training)
x = self.point_wise_conv1(x)
x = self.act(x)
x = self.point_wise_conv2(x)
x = input + self.drop_path(self.layer_scale * x)
return x
class TFSwiftFormerMlp(keras.layers.Layer):
"""
MLP layer with 1*1 convolutions.
Input: tensor of shape `[batch_size, channels, height, width]`
Output: tensor of shape `[batch_size, channels, height, width]`
"""
def __init__(self, config: SwiftFormerConfig, in_features: int, **kwargs):
super().__init__(**kwargs)
hidden_features = int(in_features * config.mlp_ratio)
self.norm1 = keras.layers.BatchNormalization(epsilon=config.batch_norm_eps, momentum=0.9, name="norm1")
self.fc1 = keras.layers.Conv2D(hidden_features, 1, name="fc1")
act_layer = get_tf_activation(config.hidden_act)
self.act = act_layer
self.fc2 = keras.layers.Conv2D(in_features, 1, name="fc2")
self.drop = keras.layers.Dropout(rate=config.drop_mlp_rate)
self.hidden_features = hidden_features
self.in_features = in_features
def call(self, x: tf.Tensor, training: bool = False) -> tf.Tensor:
x = self.norm1(x, training=training)
x = self.fc1(x)
x = self.act(x)
x = self.drop(x, training=training)
x = self.fc2(x)
x = self.drop(x, training=training)
return x
def build(self, input_shape=None):
if self.built:
return
if getattr(self, "norm1", None) is not None:
with tf.name_scope(self.norm1.name):
self.norm1.build((None, None, None, self.in_features))
if getattr(self, "fc1", None) is not None:
with tf.name_scope(self.fc1.name):
self.fc1.build((None, None, None, self.in_features))
if getattr(self, "fc2", None) is not None:
with tf.name_scope(self.fc2.name):
self.fc2.build((None, None, None, self.hidden_features))
self.built = True
class TFSwiftFormerEfficientAdditiveAttention(keras.layers.Layer):
"""
Efficient Additive Attention module for SwiftFormer.
Input: tensor of shape `[batch_size, channels, height, width]`
Output: tensor of shape `[batch_size, channels, height, width]`
"""
def __init__(self, config: SwiftFormerConfig, dim: int = 512, **kwargs):
super().__init__(**kwargs)
self.dim = dim
self.to_query = keras.layers.Dense(dim, name="to_query")
self.to_key = keras.layers.Dense(dim, name="to_key")
self.scale_factor = dim**-0.5
self.proj = keras.layers.Dense(dim, name="proj")
self.final = keras.layers.Dense(dim, name="final")
def build(self, input_shape=None):
if self.built:
return
self.w_g = self.add_weight(
name="w_g",
shape=(self.dim, 1),
initializer=keras.initializers.RandomNormal(mean=0, stddev=1),
trainable=True,
)
if getattr(self, "to_query", None) is not None:
with tf.name_scope(self.to_query.name):
self.to_query.build(self.dim)
if getattr(self, "to_key", None) is not None:
with tf.name_scope(self.to_key.name):
self.to_key.build(self.dim)
if getattr(self, "proj", None) is not None:
with tf.name_scope(self.proj.name):
self.proj.build(self.dim)
if getattr(self, "final", None) is not None:
with tf.name_scope(self.final.name):
self.final.build(self.dim)
self.built = True
def call(self, x: tf.Tensor) -> tf.Tensor:
query = self.to_query(x)
key = self.to_key(x)
query = tf.math.l2_normalize(query, dim=-1)
key = tf.math.l2_normalize(key, dim=-1)
query_weight = query @ self.w_g
scaled_query_weight = query_weight * self.scale_factor
scaled_query_weight = tf.nn.softmax(scaled_query_weight, axis=-1)
global_queries = tf.math.reduce_sum(scaled_query_weight * query, axis=1)
global_queries = tf.tile(tf.expand_dims(global_queries, 1), (1, key.shape[1], 1))
out = self.proj(global_queries * key) + query
out = self.final(out)
return out
class TFSwiftFormerLocalRepresentation(keras.layers.Layer):
"""
Local Representation module for SwiftFormer that is implemented by 3*3 depth-wise and point-wise convolutions.
Input: tensor of shape `[batch_size, channels, height, width]`
Output: tensor of shape `[batch_size, channels, height, width]`
"""
def __init__(self, config: SwiftFormerConfig, dim: int, **kwargs):
super().__init__(**kwargs)
self.dim = dim
self.pad = keras.layers.ZeroPadding2D(padding=(1, 1))
self.depth_wise_conv = keras.layers.Conv2D(dim, kernel_size=3, groups=dim, name="depth_wise_conv")
self.norm = keras.layers.BatchNormalization(epsilon=config.batch_norm_eps, momentum=0.9, name="norm")
self.point_wise_conv1 = keras.layers.Conv2D(dim, kernel_size=1, name="point_wise_conv1")
self.act = get_tf_activation("gelu")
self.point_wise_conv2 = keras.layers.Conv2D(dim, kernel_size=1, name="point_wise_conv2")
self.drop_path = keras.layers.Identity(name="drop_path")
def build(self, input_shape=None):
if self.built:
return
self.layer_scale = self.add_weight(
name="layer_scale",
shape=(self.dim),
initializer="ones",
trainable=True,
)
if getattr(self, "depth_wise_conv", None) is not None:
with tf.name_scope(self.depth_wise_conv.name):
self.depth_wise_conv.build((None, None, None, self.dim))
if getattr(self, "norm", None) is not None:
with tf.name_scope(self.norm.name):
self.norm.build((None, None, None, self.dim))
if getattr(self, "point_wise_conv1", None) is not None:
with tf.name_scope(self.point_wise_conv1.name):
self.point_wise_conv1.build(self.dim)
if getattr(self, "point_wise_conv2", None) is not None:
with tf.name_scope(self.point_wise_conv2.name):
self.point_wise_conv2.build(self.dim)
if getattr(self, "drop_path", None) is not None:
with tf.name_scope(self.drop_path.name):
self.drop_path.build(None)
self.built = True
def call(self, x: tf.Tensor, training: bool = False) -> tf.Tensor:
input = x
x = self.pad(x)
x = self.depth_wise_conv(x)
x = self.norm(x, training=training)
x = self.point_wise_conv1(x)
x = self.act(x)
x = self.point_wise_conv2(x)
x = input + self.drop_path(self.layer_scale * x, training=training)
return x
class TFSwiftFormerEncoderBlock(keras.layers.Layer):
"""
SwiftFormer Encoder Block for SwiftFormer. It consists of (1) Local representation module, (2)
SwiftFormerEfficientAdditiveAttention, and (3) MLP block.
Input: tensor of shape `[batch_size, channels, height, width]`
Output: tensor of shape `[batch_size, channels,height, width]`
"""
def __init__(self, config: SwiftFormerConfig, dim: int, drop_path: float = 0.0, **kwargs):
super().__init__(**kwargs)
layer_scale_init_value = config.layer_scale_init_value
use_layer_scale = config.use_layer_scale
self.local_representation = TFSwiftFormerLocalRepresentation(config, dim=dim, name="local_representation")
self.attn = TFSwiftFormerEfficientAdditiveAttention(config, dim=dim, name="attn")
self.linear = TFSwiftFormerMlp(config, in_features=dim, name="linear")
self.drop_path = TFSwiftFormerDropPath(config) if drop_path > 0.0 else keras.layers.Identity()
self.use_layer_scale = use_layer_scale
if use_layer_scale:
self.dim = dim
self.layer_scale_init_value = layer_scale_init_value
def build(self, input_shape=None):
if self.built:
return
self.layer_scale_1 = self.add_weight(
name="layer_scale_1",
shape=self.dim,
initializer=keras.initializers.constant(self.layer_scale_init_value),
trainable=True,
)
self.layer_scale_2 = self.add_weight(
name="layer_scale_2",
shape=self.dim,
initializer=keras.initializers.constant(self.layer_scale_init_value),
trainable=True,
)
if getattr(self, "local_representation", None) is not None:
with tf.name_scope(self.local_representation.name):
self.local_representation.build(None)
if getattr(self, "attn", None) is not None:
with tf.name_scope(self.attn.name):
self.attn.build(None)
if getattr(self, "linear", None) is not None:
with tf.name_scope(self.linear.name):
self.linear.build(None)
self.built = True
def call(self, x: tf.Tensor, training: bool = False):
x = self.local_representation(x, training=training)
batch_size, height, width, channels = x.shape
res = tf.reshape(x, [-1, height * width, channels])
res = self.attn(res)
res = tf.reshape(res, [-1, height, width, channels])
if self.use_layer_scale:
x = x + self.drop_path(self.layer_scale_1 * res, training=training)
x = x + self.drop_path(self.layer_scale_2 * self.linear(x), training=training)
else:
x = x + self.drop_path(res, training=training)
x = x + self.drop_path(self.linear(x), training=training)
return x
class TFSwiftFormerStage(keras.layers.Layer):
"""
A Swiftformer stage consisting of a series of `SwiftFormerConvEncoder` blocks and a final
`SwiftFormerEncoderBlock`.
Input: tensor in shape `[batch_size, channels, height, width]`
Output: tensor in shape `[batch_size, channels, height, width]`
"""
def __init__(self, config: SwiftFormerConfig, index: int, **kwargs) -> None:
super().__init__(**kwargs)
layer_depths = config.depths
dim = config.embed_dims[index]
depth = layer_depths[index]
self.blocks = []
for block_idx in range(depth):
block_dpr = config.drop_path_rate * (block_idx + sum(layer_depths[:index])) / (sum(layer_depths) - 1)
if depth - block_idx <= 1:
self.blocks.append(
TFSwiftFormerEncoderBlock(config, dim=dim, drop_path=block_dpr, name=f"blocks_._{block_idx}")
)
else:
self.blocks.append(TFSwiftFormerConvEncoder(config, dim=dim, name=f"blocks_._{block_idx}"))
def call(self, input: tf.Tensor, training: bool = False) -> tf.Tensor:
for i, block in enumerate(self.blocks):
input = block(input, training=training)
return input
def build(self, input_shape=None):
for layer in self.blocks:
with tf.name_scope(layer.name):
layer.build(None)
class TFSwiftFormerEncoder(keras.layers.Layer):
def __init__(self, config: SwiftFormerConfig, **kwargs) -> None:
super().__init__(**kwargs)
self.config = config
embed_dims = config.embed_dims
downsamples = config.downsamples
layer_depths = config.depths
# Transformer model
self.network = []
name_i = 0
for i in range(len(layer_depths)):
stage = TFSwiftFormerStage(config, index=i, name=f"network_._{name_i}")
self.network.append(stage)
name_i += 1
if i >= len(layer_depths) - 1:
break
if downsamples[i] or embed_dims[i] != embed_dims[i + 1]:
# downsampling between two stages
self.network.append(TFSwiftFormerEmbeddings(config, index=i, name=f"network_._{name_i}"))
name_i += 1
self.gradient_checkpointing = False
def call(
self,
hidden_states: tf.Tensor,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
training: bool = False,
) -> Union[tuple, TFBaseModelOutputWithNoAttention]:
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
all_hidden_states = (hidden_states,) if output_hidden_states else None
for i, block in enumerate(self.network):
hidden_states = block(hidden_states, training=training)
if output_hidden_states:
all_hidden_states = all_hidden_states + (hidden_states,)
hidden_states = tf.transpose(hidden_states, perm=[0, 3, 1, 2])
if all_hidden_states:
all_hidden_states = tuple(tf.transpose(s, perm=[0, 3, 1, 2]) for s in all_hidden_states)
if not return_dict:
return tuple(v for v in [hidden_states, all_hidden_states] if v is not None)
return TFBaseModelOutputWithNoAttention(
last_hidden_state=hidden_states,
hidden_states=all_hidden_states,
)
def build(self, input_shape=None):
for layer in self.network:
with tf.name_scope(layer.name):
layer.build(None)
class TFSwiftFormerPreTrainedModel(TFPreTrainedModel):
"""
An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
models.
"""
config_class = SwiftFormerConfig
base_model_prefix = "swiftformer"
main_input_name = "pixel_values"
TFSWIFTFORMER_START_DOCSTRING = r"""
This model inherits from [`TFPreTrainedModel`]. Check the superclass documentation for the generic methods the
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
etc.)
This model is also a [keras.Model](https://www.tensorflow.org/api_docs/python/tf/keras/Model) subclass. Use it
as a regular TF 2.0 Keras Model and refer to the TF 2.0 documentation for all matter related to general usage and
behavior.
<Tip>
TF 2.0 models accepts two formats as inputs:
- having all inputs as keyword arguments (like PyTorch models), or
- having all inputs as a list, tuple or dict in the first positional arguments.
This second option is useful when using [`keras.Model.fit`] method which currently requires having all the
tensors in the first argument of the model call function: `model(inputs)`.
If you choose this second option, there are three possibilities you can use to gather all the input Tensors in the
first positional argument :
- a single Tensor with `input_ids` only and nothing else: `model(input_ids)`
- a list of varying length with one or several input Tensors IN THE ORDER given in the docstring:
`model([input_ids, attention_mask])` or `model([input_ids, attention_mask, token_type_ids])`
- a dictionary with one or several input Tensors associated to the input names given in the docstring:
`model({"input_ids": input_ids, "token_type_ids": token_type_ids})`
</Tip>
Parameters:
config ([`SwiftFormerConfig`]): Model configuration class with all the parameters of the model.
Initializing with a config file does not load the weights associated with the model, only the
configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
"""
TFSWIFTFORMER_INPUTS_DOCSTRING = r"""
Args:
pixel_values (`tf.Tensor` of shape `(batch_size, num_channels, height, width)`):
Pixel values. Pixel values can be obtained using [`AutoImageProcessor`]. See [`ViTImageProcessor.__call__`]
for details.
output_hidden_states (`bool`, *optional*):
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
more detail.
return_dict (`bool`, *optional*):
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
training (`bool`, *optional*, defaults to `False`):
Whether or not to run the model in training mode.
"""
@keras_serializable
class TFSwiftFormerMainLayer(keras.layers.Layer):
config_class = SwiftFormerConfig
def __init__(self, config: SwiftFormerConfig, **kwargs):
super().__init__(**kwargs)
self.config = config
self.patch_embed = TFSwiftFormerPatchEmbedding(config, name="patch_embed")
self.encoder = TFSwiftFormerEncoder(config, name="encoder")
@unpack_inputs
def call(
self,
pixel_values: Optional[tf.Tensor] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
training: bool = False,
) -> Union[Tuple, TFBaseModelOutputWithNoAttention]:
r""" """
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
# TF 2.0 image layers can't use NCHW format when running on CPU.
# We transpose to NHWC format and then transpose back after the full forward pass.
# (batch_size, num_channels, height, width) -> (batch_size, height, width, num_channels)
pixel_values = tf.transpose(pixel_values, perm=[0, 2, 3, 1])
if pixel_values is None:
raise ValueError("You have to specify pixel_values")
embedding_output = self.patch_embed(pixel_values, training=training)
encoder_outputs = self.encoder(
embedding_output,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
training=training,
)
if not return_dict:
return tuple(v for v in encoder_outputs if v is not None)
return TFBaseModelOutputWithNoAttention(
last_hidden_state=encoder_outputs.last_hidden_state,
hidden_states=encoder_outputs.hidden_states,
)
def build(self, input_shape=None):
if self.built:
return
if getattr(self, "patch_embed", None) is not None:
with tf.name_scope(self.patch_embed.name):
self.patch_embed.build(None)
if getattr(self, "encoder", None) is not None:
with tf.name_scope(self.encoder.name):
self.encoder.build(None)
self.built = True
@add_start_docstrings(
"The bare TFSwiftFormer Model transformer outputting raw hidden-states without any specific head on top.",
TFSWIFTFORMER_START_DOCSTRING,
)
class TFSwiftFormerModel(TFSwiftFormerPreTrainedModel):
def __init__(self, config: SwiftFormerConfig, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs)
self.swiftformer = TFSwiftFormerMainLayer(config, name="swiftformer")
@unpack_inputs
@add_start_docstrings_to_model_forward(TFSWIFTFORMER_INPUTS_DOCSTRING)
def call(
self,
pixel_values: Optional[tf.Tensor] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
training: bool = False,
) -> Union[TFBaseModelOutputWithNoAttention, Tuple[tf.Tensor]]:
outputs = self.swiftformer(
pixel_values=pixel_values,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
training=training,
)
return outputs
def build(self, input_shape=None):
if self.built:
return
if getattr(self, "swiftformer", None) is not None:
with tf.name_scope(self.swiftformer.name):
self.swiftformer.build(None)
self.built = True
@add_start_docstrings(
"""
TFSwiftFormer Model transformer with an image classification head on top (e.g. for ImageNet).
""",
TFSWIFTFORMER_START_DOCSTRING,
)
class TFSwiftFormerForImageClassification(TFSwiftFormerPreTrainedModel):
def __init__(self, config: SwiftFormerConfig, **kwargs) -> None:
super().__init__(config, **kwargs)
self.num_labels = config.num_labels
self.swiftformer = TFSwiftFormerMainLayer(config, name="swiftformer")
# Classifier head
self.norm = keras.layers.BatchNormalization(epsilon=config.batch_norm_eps, momentum=0.9, name="norm")
self.head = (
keras.layers.Dense(self.num_labels, name="head")
if self.num_labels > 0
else keras.layers.Identity(name="head")
)
self.dist_head = (
keras.layers.Dense(self.num_labels, name="dist_head")
if self.num_labels > 0
else keras.layers.Identity(name="dist_head")
)
def hf_compute_loss(self, labels, logits):
if self.config.problem_type is None:
if self.num_labels == 1:
self.config.problem_type = "regression"
elif self.num_labels > 1 and (labels.dtype == tf.int64 or labels.dtype == tf.int32):
self.config.problem_type = "single_label_classification"
else:
self.config.problem_type = "multi_label_classification"
if self.config.problem_type == "regression":
loss_fct = keras.losses.MSE
if self.num_labels == 1:
loss = loss_fct(labels.squeeze(), logits.squeeze())
else:
loss = loss_fct(labels, logits)
elif self.config.problem_type == "single_label_classification":
loss_fct = keras.losses.SparseCategoricalCrossentropy(
from_logits=True, reduction=keras.losses.Reduction.NONE
)
loss = loss_fct(labels, logits)
elif self.config.problem_type == "multi_label_classification":
loss_fct = keras.losses.SparseCategoricalCrossentropy(
from_logits=True,
reduction=keras.losses.Reduction.NONE,
)
loss = loss_fct(labels, logits)
else:
loss = None
return loss
@unpack_inputs
@add_start_docstrings_to_model_forward(TFSWIFTFORMER_INPUTS_DOCSTRING)
def call(
self,
pixel_values: Optional[tf.Tensor] = None,
labels: Optional[tf.Tensor] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
training: bool = False,
) -> Union[tuple, TFImageClassifierOutputWithNoAttention]:
r"""
labels (`tf.Tensor` of shape `(batch_size,)`, *optional*):
Labels for computing the image classification/regression loss. Indices should be in `[0, ...,
config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
"""
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
# run base model
outputs = self.swiftformer(
pixel_values,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
training=training,
)
sequence_output = outputs.last_hidden_state if return_dict else outputs[0]
sequence_output = tf.transpose(sequence_output, perm=[0, 2, 3, 1])
# run classification head
sequence_output = self.norm(sequence_output, training=training)
sequence_output = tf.transpose(sequence_output, perm=[0, 3, 1, 2])
_, num_channels, height, width = sequence_output.shape
sequence_output = tf.reshape(sequence_output, [-1, num_channels, height * width])
sequence_output = tf.reduce_mean(sequence_output, axis=-1)
cls_out = self.head(sequence_output)
distillation_out = self.dist_head(sequence_output)
logits = (cls_out + distillation_out) / 2
# calculate loss
loss = None if labels is None else self.hf_compute_loss(labels=labels, logits=logits)
if not return_dict:
output = (logits,) + outputs[1:]
return ((loss,) + output) if loss is not None else output
return TFImageClassifierOutputWithNoAttention(
loss=loss,
logits=logits,
hidden_states=outputs.hidden_states,
)
def build(self, input_shape=None):
if self.built:
return
if getattr(self, "swiftformer", None) is not None:
with tf.name_scope(self.swiftformer.name):
self.swiftformer.build(None)
if getattr(self, "norm", None) is not None:
with tf.name_scope(self.norm.name):
self.norm.build((None, None, None, self.config.embed_dims[-1]))
if getattr(self, "head", None) is not None:
with tf.name_scope(self.head.name):
self.head.build(self.config.embed_dims[-1])
if getattr(self, "dist_head", None) is not None:
with tf.name_scope(self.dist_head.name):
self.dist_head.build(self.config.embed_dims[-1])
self.built = True
|
transformers/src/transformers/models/swiftformer/modeling_tf_swiftformer.py/0
|
{
"file_path": "transformers/src/transformers/models/swiftformer/modeling_tf_swiftformer.py",
"repo_id": "transformers",
"token_count": 15128
}
| 385
|
# Copyright 2022 The HuggingFace Team. 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.
from typing import TYPE_CHECKING
from ...utils import (
OptionalDependencyNotAvailable,
_LazyModule,
is_flax_available,
is_sentencepiece_available,
is_tf_available,
is_tokenizers_available,
is_torch_available,
)
_import_structure = {
"configuration_switch_transformers": [
"SwitchTransformersConfig",
"SwitchTransformersOnnxConfig",
]
}
try:
if not is_torch_available():
raise OptionalDependencyNotAvailable()
except OptionalDependencyNotAvailable:
pass
else:
_import_structure["modeling_switch_transformers"] = [
"SwitchTransformersEncoderModel",
"SwitchTransformersForConditionalGeneration",
"SwitchTransformersModel",
"SwitchTransformersPreTrainedModel",
"SwitchTransformersTop1Router",
"SwitchTransformersSparseMLP",
]
if TYPE_CHECKING:
from .configuration_switch_transformers import (
SwitchTransformersConfig,
SwitchTransformersOnnxConfig,
)
try:
if not is_torch_available():
raise OptionalDependencyNotAvailable()
except OptionalDependencyNotAvailable:
pass
else:
from .modeling_switch_transformers import (
SwitchTransformersEncoderModel,
SwitchTransformersForConditionalGeneration,
SwitchTransformersModel,
SwitchTransformersPreTrainedModel,
SwitchTransformersSparseMLP,
SwitchTransformersTop1Router,
)
else:
import sys
sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)
|
transformers/src/transformers/models/switch_transformers/__init__.py/0
|
{
"file_path": "transformers/src/transformers/models/switch_transformers/__init__.py",
"repo_id": "transformers",
"token_count": 810
}
| 386
|
# coding=utf-8
# Copyright 2022 The HuggingFace Inc. team.
#
# 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.
"""Convert TimeSformer checkpoints from the original repository: https://github.com/MCG-NJU/TimeSformer"""
import argparse
import json
import gdown
import numpy as np
import torch
from huggingface_hub import hf_hub_download
from transformers import TimesformerConfig, TimesformerForVideoClassification, VideoMAEImageProcessor
def get_timesformer_config(model_name):
config = TimesformerConfig()
if "large" in model_name:
config.num_frames = 96
if "hr" in model_name:
config.num_frames = 16
config.image_size = 448
repo_id = "huggingface/label-files"
if "k400" in model_name:
config.num_labels = 400
filename = "kinetics400-id2label.json"
elif "k600" in model_name:
config.num_labels = 600
filename = "kinetics600-id2label.json"
elif "ssv2" in model_name:
config.num_labels = 174
filename = "something-something-v2-id2label.json"
else:
raise ValueError("Model name should either contain 'k400', 'k600' or 'ssv2'.")
id2label = json.load(open(hf_hub_download(repo_id, filename, repo_type="dataset"), "r"))
id2label = {int(k): v for k, v in id2label.items()}
config.id2label = id2label
config.label2id = {v: k for k, v in id2label.items()}
return config
def rename_key(name):
if "encoder." in name:
name = name.replace("encoder.", "")
if "cls_token" in name:
name = name.replace("cls_token", "timesformer.embeddings.cls_token")
if "pos_embed" in name:
name = name.replace("pos_embed", "timesformer.embeddings.position_embeddings")
if "time_embed" in name:
name = name.replace("time_embed", "timesformer.embeddings.time_embeddings")
if "patch_embed.proj" in name:
name = name.replace("patch_embed.proj", "timesformer.embeddings.patch_embeddings.projection")
if "patch_embed.norm" in name:
name = name.replace("patch_embed.norm", "timesformer.embeddings.norm")
if "blocks" in name:
name = name.replace("blocks", "timesformer.encoder.layer")
if "attn.proj" in name:
name = name.replace("attn.proj", "attention.output.dense")
if "attn" in name and "bias" not in name and "temporal" not in name:
name = name.replace("attn", "attention.self")
if "attn" in name and "temporal" not in name:
name = name.replace("attn", "attention.attention")
if "temporal_norm1" in name:
name = name.replace("temporal_norm1", "temporal_layernorm")
if "temporal_attn.proj" in name:
name = name.replace("temporal_attn", "temporal_attention.output.dense")
if "temporal_fc" in name:
name = name.replace("temporal_fc", "temporal_dense")
if "norm1" in name and "temporal" not in name:
name = name.replace("norm1", "layernorm_before")
if "norm2" in name:
name = name.replace("norm2", "layernorm_after")
if "mlp.fc1" in name:
name = name.replace("mlp.fc1", "intermediate.dense")
if "mlp.fc2" in name:
name = name.replace("mlp.fc2", "output.dense")
if "norm.weight" in name and "fc" not in name and "temporal" not in name:
name = name.replace("norm.weight", "timesformer.layernorm.weight")
if "norm.bias" in name and "fc" not in name and "temporal" not in name:
name = name.replace("norm.bias", "timesformer.layernorm.bias")
if "head" in name:
name = name.replace("head", "classifier")
return name
def convert_state_dict(orig_state_dict, config):
for key in orig_state_dict.copy().keys():
val = orig_state_dict.pop(key)
if key.startswith("model."):
key = key.replace("model.", "")
if "qkv" in key:
key_split = key.split(".")
layer_num = int(key_split[1])
prefix = "timesformer.encoder.layer."
if "temporal" in key:
postfix = ".temporal_attention.attention.qkv."
else:
postfix = ".attention.attention.qkv."
if "weight" in key:
orig_state_dict[f"{prefix}{layer_num}{postfix}weight"] = val
else:
orig_state_dict[f"{prefix}{layer_num}{postfix}bias"] = val
else:
orig_state_dict[rename_key(key)] = val
return orig_state_dict
# We will verify our results on a video of eating spaghetti
# Frame indices used: [164 168 172 176 181 185 189 193 198 202 206 210 215 219 223 227]
def prepare_video():
file = hf_hub_download(
repo_id="hf-internal-testing/spaghetti-video", filename="eating_spaghetti.npy", repo_type="dataset"
)
video = np.load(file)
return list(video)
def convert_timesformer_checkpoint(checkpoint_url, pytorch_dump_folder_path, model_name, push_to_hub):
config = get_timesformer_config(model_name)
model = TimesformerForVideoClassification(config)
# download original checkpoint, hosted on Google Drive
output = "pytorch_model.bin"
gdown.cached_download(checkpoint_url, output, quiet=False)
files = torch.load(output, map_location="cpu")
if "model" in files:
state_dict = files["model"]
elif "module" in files:
state_dict = files["module"]
else:
state_dict = files["model_state"]
new_state_dict = convert_state_dict(state_dict, config)
model.load_state_dict(new_state_dict)
model.eval()
# verify model on basic input
image_processor = VideoMAEImageProcessor(image_mean=[0.5, 0.5, 0.5], image_std=[0.5, 0.5, 0.5])
video = prepare_video()
inputs = image_processor(video[:8], return_tensors="pt")
outputs = model(**inputs)
logits = outputs.logits
model_names = [
# Kinetics-400 checkpoints (hr = high resolution input of 448px instead of 224px)
"timesformer-base-finetuned-k400",
"timesformer-large-finetuned-k400",
"timesformer-hr-finetuned-k400",
# Kinetics-600 checkpoints (hr = high resolution input of 448px instead of 224px)
"timesformer-base-finetuned-k600",
"timesformer-large-finetuned-k600",
"timesformer-hr-finetuned-k600",
# Something-Something-v2 checkpoints (hr = high resolution input of 448px instead of 224px)
"timesformer-base-finetuned-ssv2",
"timesformer-large-finetuned-ssv2",
"timesformer-hr-finetuned-ssv2",
]
# NOTE: logits were tested with image_mean and image_std equal to [0.5, 0.5, 0.5] and [0.5, 0.5, 0.5]
if model_name == "timesformer-base-finetuned-k400":
expected_shape = torch.Size([1, 400])
expected_slice = torch.tensor([-0.3016, -0.7713, -0.4205])
elif model_name == "timesformer-base-finetuned-k600":
expected_shape = torch.Size([1, 600])
expected_slice = torch.tensor([-0.7267, -0.7466, 3.2404])
elif model_name == "timesformer-base-finetuned-ssv2":
expected_shape = torch.Size([1, 174])
expected_slice = torch.tensor([-0.9059, 0.6433, -3.1457])
elif model_name == "timesformer-large-finetuned-k400":
expected_shape = torch.Size([1, 400])
expected_slice = torch.tensor([0, 0, 0])
elif model_name == "timesformer-large-finetuned-k600":
expected_shape = torch.Size([1, 600])
expected_slice = torch.tensor([0, 0, 0])
elif model_name == "timesformer-large-finetuned-ssv2":
expected_shape = torch.Size([1, 174])
expected_slice = torch.tensor([0, 0, 0])
elif model_name == "timesformer-hr-finetuned-k400":
expected_shape = torch.Size([1, 400])
expected_slice = torch.tensor([-0.9617, -3.7311, -3.7708])
elif model_name == "timesformer-hr-finetuned-k600":
expected_shape = torch.Size([1, 600])
expected_slice = torch.tensor([2.5273, 0.7127, 1.8848])
elif model_name == "timesformer-hr-finetuned-ssv2":
expected_shape = torch.Size([1, 174])
expected_slice = torch.tensor([-3.6756, -0.7513, 0.7180])
else:
raise ValueError(f"Model name not supported. Should be one of {model_names}")
# verify logits
assert logits.shape == expected_shape
assert torch.allclose(logits[0, :3], expected_slice, atol=1e-4)
print("Logits ok!")
if pytorch_dump_folder_path is not None:
print(f"Saving model and image processor to {pytorch_dump_folder_path}")
image_processor.save_pretrained(pytorch_dump_folder_path)
model.save_pretrained(pytorch_dump_folder_path)
if push_to_hub:
print("Pushing to the hub...")
model.push_to_hub(f"fcakyon/{model_name}")
if __name__ == "__main__":
parser = argparse.ArgumentParser()
# Required parameters
parser.add_argument(
"--checkpoint_url",
default="https://drive.google.com/u/1/uc?id=17yvuYp9L4mn-HpIcK5Zo6K3UoOy1kA5l&export=download",
type=str,
help=(
"URL of the original PyTorch checkpoint (on Google Drive) you'd like to convert. Should be a direct"
" download link."
),
)
parser.add_argument(
"--pytorch_dump_folder_path",
default="",
type=str,
help="Path to the output PyTorch model directory.",
)
parser.add_argument("--model_name", default="timesformer-base-finetuned-k400", type=str, help="Name of the model.")
parser.add_argument(
"--push_to_hub", action="store_true", help="Whether or not to push the converted model to the 🤗 hub."
)
args = parser.parse_args()
convert_timesformer_checkpoint(
args.checkpoint_url, args.pytorch_dump_folder_path, args.model_name, args.push_to_hub
)
|
transformers/src/transformers/models/timesformer/convert_timesformer_to_pytorch.py/0
|
{
"file_path": "transformers/src/transformers/models/timesformer/convert_timesformer_to_pytorch.py",
"repo_id": "transformers",
"token_count": 4205
}
| 387
|
# coding=utf-8
# Copyright 2022 Facebook AI and The HuggingFace Inc. team. 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.
"""TF 2.0 ViT MAE (masked autoencoder) model."""
from __future__ import annotations
import collections.abc
import math
from copy import deepcopy
from dataclasses import dataclass
from typing import Optional, Tuple, Union
import numpy as np
import tensorflow as tf
from ...activations_tf import get_tf_activation
from ...file_utils import (
ModelOutput,
add_start_docstrings,
add_start_docstrings_to_model_forward,
replace_return_docstrings,
)
from ...modeling_tf_outputs import TFBaseModelOutput
from ...modeling_tf_utils import (
TFModelInputType,
TFPreTrainedModel,
get_initializer,
keras,
keras_serializable,
unpack_inputs,
)
from ...tf_utils import shape_list, stable_softmax
from ...utils import logging
from .configuration_vit_mae import ViTMAEConfig
logger = logging.get_logger(__name__)
_CONFIG_FOR_DOC = "ViTMAEConfig"
_CHECKPOINT_FOR_DOC = "facebook/vit-mae-base"
@dataclass
class TFViTMAEModelOutput(ModelOutput):
"""
Class for TFViTMAEModel's outputs, with potential hidden states and attentions.
Args:
last_hidden_state (`tf.Tensor` of shape `(batch_size, sequence_length, hidden_size)`):
Sequence of hidden-states at the output of the last layer of the model.
mask (`tf.Tensor` of shape `(batch_size, sequence_length)`):
Tensor indicating which patches are masked (1) and which are not (0).
ids_restore (`tf.Tensor` of shape `(batch_size, sequence_length)`):
Tensor containing the original index of the (shuffled) masked patches.
hidden_states (`tuple(tf.Tensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `tf.Tensor` (one for the output of the embeddings + one for the output of each layer) of shape
`(batch_size, sequence_length, hidden_size)`. Hidden-states of the model at the output of each layer plus
the initial embedding outputs.
attentions (`tuple(tf.Tensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `tf.Tensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`. Attentions weights after the attention softmax, used to compute the weighted average in
the self-attention heads.
"""
last_hidden_state: tf.Tensor = None
mask: tf.Tensor = None
ids_restore: tf.Tensor = None
hidden_states: Tuple[tf.Tensor] | None = None
attentions: Tuple[tf.Tensor] | None = None
@dataclass
class TFViTMAEDecoderOutput(ModelOutput):
"""
Class for TFViTMAEDecoder's outputs, with potential hidden states and attentions.
Args:
logits (`tf.Tensor` of shape `(batch_size, sequence_length, patch_size ** 2 * num_channels)`):
Pixel reconstruction logits.
hidden_states (`tuple(tf.Tensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `tf.Tensor` (one for the output of the embeddings + one for the output of each layer) of shape
`(batch_size, sequence_length, hidden_size)`. Hidden-states of the model at the output of each layer plus
the initial embedding outputs.
attentions (`tuple(tf.Tensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `tf.Tensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`. Attentions weights after the attention softmax, used to compute the weighted average in
the self-attention heads.
"""
logits: tf.Tensor = None
hidden_states: Tuple[tf.Tensor] | None = None
attentions: Tuple[tf.Tensor] | None = None
@dataclass
class TFViTMAEForPreTrainingOutput(ModelOutput):
"""
Class for TFViTMAEForPreTraining's outputs, with potential hidden states and attentions.
Args:
loss (`tf.Tensor` of shape `(1,)`):
Pixel reconstruction loss.
logits (`tf.Tensor` of shape `(batch_size, sequence_length, patch_size ** 2 * num_channels)`):
Pixel reconstruction logits.
mask (`tf.Tensor` of shape `(batch_size, sequence_length)`):
Tensor indicating which patches are masked (1) and which are not (0).
ids_restore (`tf.Tensor` of shape `(batch_size, sequence_length)`):
Tensor containing the original index of the (shuffled) masked patches.
hidden_states (`tuple(tf.Tensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `tf.Tensor` (one for the output of the embeddings + one for the output of each layer) of shape
`(batch_size, sequence_length, hidden_size)`. Hidden-states of the model at the output of each layer plus
the initial embedding outputs.
attentions (`tuple(tf.Tensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `tf.Tensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`. Attentions weights after the attention softmax, used to compute the weighted average in
the self-attention heads.
"""
loss: tf.Tensor | None = None
logits: tf.Tensor = None
mask: tf.Tensor = None
ids_restore: tf.Tensor = None
hidden_states: Tuple[tf.Tensor] | None = None
attentions: Tuple[tf.Tensor] | None = None
def get_2d_sincos_pos_embed(embed_dim, grid_size, add_cls_token=False):
"""
Create 2D sin/cos positional embeddings.
Args:
embed_dim (`int`):
Embedding dimension.
grid_size (`int`):
The grid height and width.
add_cls_token (`bool`, *optional*, defaults to `False`):
Whether or not to add a classification (CLS) token.
Returns:
(`tf.Tensor` of shape (grid_size*grid_size, embed_dim) or (1+grid_size*grid_size, embed_dim): the position
embeddings (with or without classification token)
"""
grid_h = tf.range(grid_size, dtype=tf.float32)
grid_w = tf.range(grid_size, dtype=tf.float32)
grid = tf.meshgrid(grid_w, grid_h) # here w goes first
grid = tf.stack(grid, axis=0)
grid = tf.reshape(grid, [2, 1, grid_size, grid_size])
pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
if add_cls_token:
pos_embed = tf.concat([tf.zeros((1, embed_dim)), pos_embed], axis=0)
return pos_embed
def get_2d_sincos_pos_embed_from_grid(embed_dim, grid):
if embed_dim % 2 != 0:
raise ValueError("embed_dim must be even")
# 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 = tf.concat([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)
"""
if embed_dim % 2 != 0:
raise ValueError("embed_dim must be even")
omega = tf.range(embed_dim // 2, dtype="float32")
omega /= embed_dim / 2.0
omega = 1.0 / 10000**omega # (D/2,)
pos = tf.reshape(pos, [-1]) # (M,)
out = tf.einsum("m,d->md", pos, omega) # (M, D/2), outer product
# half of the positions get sinusoidal pattern and the rest gets
# cosine pattern and then they are concatenated
emb_sin = tf.sin(out) # (M, D/2)
emb_cos = tf.cos(out) # (M, D/2)
emb = tf.concat([emb_sin, emb_cos], axis=1) # (M, D)
return emb
class TFViTMAEEmbeddings(keras.layers.Layer):
"""
Construct the CLS token, position and patch embeddings.
"""
def __init__(self, config: ViTMAEConfig, **kwargs):
super().__init__(**kwargs)
self.patch_embeddings = TFViTMAEPatchEmbeddings(config, name="patch_embeddings")
self.num_patches = self.patch_embeddings.num_patches
self.config = config
def build(self, input_shape=None):
self.cls_token = self.add_weight(
shape=(1, 1, self.config.hidden_size),
initializer=tf.random_normal_initializer(stddev=self.config.initializer_range),
trainable=True,
name="cls_token",
)
self.position_embeddings = self.add_weight(
shape=(1, self.num_patches + 1, self.config.hidden_size),
initializer="zeros",
trainable=False, # fixed sin-cos embedding
name="position_embeddings",
)
pos_embed = get_2d_sincos_pos_embed(
self.position_embeddings.shape[-1],
int(self.patch_embeddings.num_patches**0.5),
add_cls_token=True,
)[None, ...]
self.position_embeddings.assign(pos_embed)
if self.built:
return
self.built = True
if getattr(self, "patch_embeddings", None) is not None:
with tf.name_scope(self.patch_embeddings.name):
self.patch_embeddings.build(None)
def interpolate_pos_encoding(self, embeddings, height, width) -> tf.Tensor:
"""
This method allows to interpolate the pre-trained position encodings, to be able to use the model on higher
resolution images.
Source:
https://github.com/facebookresearch/dino/blob/de9ee3df6cf39fac952ab558447af1fa1365362a/vision_transformer.py#L174
"""
batch_size, seq_len, dim = shape_list(embeddings)
num_patches = seq_len - 1
_, num_positions, _ = shape_list(self.position_embeddings)
num_positions -= 1
if num_patches == num_positions and height == width:
return self.position_embeddings
class_pos_embed = self.position_embeddings[:, :1]
patch_pos_embed = self.position_embeddings[:, 1:]
h0 = height // self.config.patch_size
w0 = width // self.config.patch_size
patch_pos_embed = tf.image.resize(
images=tf.reshape(
patch_pos_embed, shape=(1, int(math.sqrt(num_positions)), int(math.sqrt(num_positions)), dim)
),
size=(h0, w0),
method="bicubic",
)
patch_pos_embed = tf.reshape(tensor=patch_pos_embed, shape=(1, -1, dim))
return tf.concat(values=(class_pos_embed, patch_pos_embed), axis=1)
def random_masking(self, sequence: tf.Tensor, noise: tf.Tensor | None = None):
"""
Perform per-sample random masking by per-sample shuffling. Per-sample shuffling is done by argsort random
noise.
Args:
sequence (`tf.Tensor` of shape `(batch_size, sequence_length, dim)`)
noise (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*) which is
mainly used for testing purposes to control randomness and maintain the reproducibility
"""
batch_size, seq_length, dim = shape_list(sequence)
len_keep = int(seq_length * (1 - self.config.mask_ratio))
if noise is None:
noise = tf.random.uniform(shape=(batch_size, seq_length), minval=0.0, maxval=1.0) # noise in [0, 1)
# sort noise for each sample
ids_shuffle = tf.argsort(noise, axis=1) # ascend: small is keep, large is remove
ids_restore = tf.argsort(ids_shuffle, axis=1)
# keep the first subset
ids_keep = ids_shuffle[:, :len_keep]
sequence_unmasked = tf.gather(
sequence,
axis=1,
batch_dims=1,
indices=ids_keep,
)
# generate the binary mask: 0 is keep, 1 is remove
# this hack is needed because TF's EagerTensors don't support
# assignment
mask_keep = tf.zeros((batch_size, len_keep))
mask_remove = tf.ones((batch_size, seq_length - len_keep))
mask = tf.concat([mask_keep, mask_remove], axis=-1)
# unshuffle to get the binary mask
mask = tf.gather(mask, axis=1, batch_dims=1, indices=ids_restore)
return sequence_unmasked, mask, ids_restore
def call(
self, pixel_values: tf.Tensor, noise: tf.Tensor = None, interpolate_pos_encoding: bool = False
) -> tf.Tensor:
batch_size, num_channels, height, width = shape_list(pixel_values)
embeddings = self.patch_embeddings(pixel_values, interpolate_pos_encoding=interpolate_pos_encoding)
if interpolate_pos_encoding:
position_embeddings = self.interpolate_pos_encoding(embeddings, height, width)
else:
position_embeddings = self.position_embeddings
# add position embeddings w/o cls token
embeddings = embeddings + position_embeddings[:, 1:, :]
# masking: length -> length * config.mask_ratio
embeddings, mask, ids_restore = self.random_masking(embeddings, noise)
# append cls token
cls_token = self.cls_token + position_embeddings[:, :1, :]
cls_tokens = tf.tile(cls_token, (shape_list(embeddings)[0], 1, 1))
embeddings = tf.concat([cls_tokens, embeddings], axis=1)
return embeddings, mask, ids_restore
class TFViTMAEPatchEmbeddings(keras.layers.Layer):
"""
This class turns `pixel_values` of shape `(batch_size, num_channels, height, width)` into the initial
`hidden_states` (patch embeddings) of shape `(batch_size, seq_length, hidden_size)` to be consumed by a
Transformer.
"""
def __init__(self, config: ViTMAEConfig, **kwargs):
super().__init__(**kwargs)
image_size, patch_size = config.image_size, config.patch_size
num_channels, hidden_size = config.num_channels, config.hidden_size
image_size = image_size if isinstance(image_size, collections.abc.Iterable) else (image_size, image_size)
patch_size = patch_size if isinstance(patch_size, collections.abc.Iterable) else (patch_size, patch_size)
num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
self.image_size = image_size
self.patch_size = patch_size
self.num_patches = num_patches
self.num_channels = num_channels
self.config = config
self.projection = keras.layers.Conv2D(
filters=hidden_size,
kernel_size=patch_size,
strides=patch_size,
padding="valid",
data_format="channels_last",
kernel_initializer="glorot_uniform", # following torch.nn.Linear
bias_initializer="zeros",
name="projection",
)
def call(
self, pixel_values: tf.Tensor, training: bool = False, interpolate_pos_encoding: bool = False
) -> tf.Tensor:
batch_size, num_channels, height, width = shape_list(pixel_values)
if tf.executing_eagerly():
if num_channels != self.num_channels:
raise ValueError(
"Make sure that the channel dimension of the pixel values match with the one set in the"
" configuration."
)
if not interpolate_pos_encoding and (height != self.image_size[0] or width != self.image_size[1]):
raise ValueError(
f"Input image size ({height}*{width}) doesn't match model"
f" ({self.image_size[0]}*{self.image_size[1]})."
)
# When running on CPU, `keras.layers.Conv2D` doesn't support `NCHW` format.
# So change the input format from `NCHW` to `NHWC`.
# shape = (batch_size, in_height, in_width, in_channels=num_channels)
pixel_values = tf.transpose(pixel_values, perm=(0, 2, 3, 1))
projection = self.projection(pixel_values)
# Change the 2D spatial dimensions to a single temporal dimension.
# shape = (batch_size, num_patches, out_channels=embed_dim)
num_patches = (width // self.patch_size[1]) * (height // self.patch_size[0])
x = tf.reshape(tensor=projection, shape=(batch_size, num_patches, -1))
return x
def build(self, input_shape=None):
if self.built:
return
self.built = True
if getattr(self, "projection", None) is not None:
with tf.name_scope(self.projection.name):
self.projection.build([None, None, None, self.num_channels])
# Copied from transformers.models.vit.modeling_tf_vit.TFViTSelfAttention with ViT->ViTMAE
class TFViTMAESelfAttention(keras.layers.Layer):
def __init__(self, config: ViTMAEConfig, **kwargs):
super().__init__(**kwargs)
if config.hidden_size % config.num_attention_heads != 0:
raise ValueError(
f"The hidden size ({config.hidden_size}) is not a multiple of the number "
f"of attention heads ({config.num_attention_heads})"
)
self.num_attention_heads = config.num_attention_heads
self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
self.all_head_size = self.num_attention_heads * self.attention_head_size
self.sqrt_att_head_size = math.sqrt(self.attention_head_size)
self.query = keras.layers.Dense(
units=self.all_head_size, kernel_initializer=get_initializer(config.initializer_range), name="query"
)
self.key = keras.layers.Dense(
units=self.all_head_size, kernel_initializer=get_initializer(config.initializer_range), name="key"
)
self.value = keras.layers.Dense(
units=self.all_head_size, kernel_initializer=get_initializer(config.initializer_range), name="value"
)
self.dropout = keras.layers.Dropout(rate=config.attention_probs_dropout_prob)
self.config = config
def transpose_for_scores(self, tensor: tf.Tensor, batch_size: int) -> tf.Tensor:
# Reshape from [batch_size, seq_length, all_head_size] to [batch_size, seq_length, num_attention_heads, attention_head_size]
tensor = tf.reshape(tensor=tensor, shape=(batch_size, -1, self.num_attention_heads, self.attention_head_size))
# Transpose the tensor from [batch_size, seq_length, num_attention_heads, attention_head_size] to [batch_size, num_attention_heads, seq_length, attention_head_size]
return tf.transpose(tensor, perm=[0, 2, 1, 3])
def call(
self,
hidden_states: tf.Tensor,
head_mask: tf.Tensor,
output_attentions: bool,
training: bool = False,
) -> Tuple[tf.Tensor]:
batch_size = shape_list(hidden_states)[0]
mixed_query_layer = self.query(inputs=hidden_states)
mixed_key_layer = self.key(inputs=hidden_states)
mixed_value_layer = self.value(inputs=hidden_states)
query_layer = self.transpose_for_scores(mixed_query_layer, batch_size)
key_layer = self.transpose_for_scores(mixed_key_layer, batch_size)
value_layer = self.transpose_for_scores(mixed_value_layer, batch_size)
# Take the dot product between "query" and "key" to get the raw attention scores.
# (batch size, num_heads, seq_len_q, seq_len_k)
attention_scores = tf.matmul(query_layer, key_layer, transpose_b=True)
dk = tf.cast(self.sqrt_att_head_size, dtype=attention_scores.dtype)
attention_scores = tf.divide(attention_scores, dk)
# Normalize the attention scores to probabilities.
attention_probs = stable_softmax(logits=attention_scores, axis=-1)
# This is actually dropping out entire tokens to attend to, which might
# seem a bit unusual, but is taken from the original Transformer paper.
attention_probs = self.dropout(inputs=attention_probs, training=training)
# Mask heads if we want to
if head_mask is not None:
attention_probs = tf.multiply(attention_probs, head_mask)
attention_output = tf.matmul(attention_probs, value_layer)
attention_output = tf.transpose(attention_output, perm=[0, 2, 1, 3])
# (batch_size, seq_len_q, all_head_size)
attention_output = tf.reshape(tensor=attention_output, shape=(batch_size, -1, self.all_head_size))
outputs = (attention_output, attention_probs) if output_attentions else (attention_output,)
return outputs
def build(self, input_shape=None):
if self.built:
return
self.built = True
if getattr(self, "query", None) is not None:
with tf.name_scope(self.query.name):
self.query.build([None, None, self.config.hidden_size])
if getattr(self, "key", None) is not None:
with tf.name_scope(self.key.name):
self.key.build([None, None, self.config.hidden_size])
if getattr(self, "value", None) is not None:
with tf.name_scope(self.value.name):
self.value.build([None, None, self.config.hidden_size])
# Copied from transformers.models.vit.modeling_tf_vit.TFViTSelfOutput with ViT->ViTMAE
class TFViTMAESelfOutput(keras.layers.Layer):
"""
The residual connection is defined in TFViTMAELayer instead of here (as is the case with other models), due to the
layernorm applied before each block.
"""
def __init__(self, config: ViTMAEConfig, **kwargs):
super().__init__(**kwargs)
self.dense = keras.layers.Dense(
units=config.hidden_size, kernel_initializer=get_initializer(config.initializer_range), name="dense"
)
self.dropout = keras.layers.Dropout(rate=config.hidden_dropout_prob)
self.config = config
def call(self, hidden_states: tf.Tensor, input_tensor: tf.Tensor, training: bool = False) -> tf.Tensor:
hidden_states = self.dense(inputs=hidden_states)
hidden_states = self.dropout(inputs=hidden_states, training=training)
return hidden_states
def build(self, input_shape=None):
if self.built:
return
self.built = True
if getattr(self, "dense", None) is not None:
with tf.name_scope(self.dense.name):
self.dense.build([None, None, self.config.hidden_size])
# Copied from transformers.models.vit.modeling_tf_vit.TFViTAttention with ViT->ViTMAE
class TFViTMAEAttention(keras.layers.Layer):
def __init__(self, config: ViTMAEConfig, **kwargs):
super().__init__(**kwargs)
self.self_attention = TFViTMAESelfAttention(config, name="attention")
self.dense_output = TFViTMAESelfOutput(config, name="output")
def prune_heads(self, heads):
raise NotImplementedError
def call(
self,
input_tensor: tf.Tensor,
head_mask: tf.Tensor,
output_attentions: bool,
training: bool = False,
) -> Tuple[tf.Tensor]:
self_outputs = self.self_attention(
hidden_states=input_tensor, head_mask=head_mask, output_attentions=output_attentions, training=training
)
attention_output = self.dense_output(
hidden_states=self_outputs[0], input_tensor=input_tensor, training=training
)
outputs = (attention_output,) + self_outputs[1:] # add attentions if we output them
return outputs
def build(self, input_shape=None):
if self.built:
return
self.built = True
if getattr(self, "self_attention", None) is not None:
with tf.name_scope(self.self_attention.name):
self.self_attention.build(None)
if getattr(self, "dense_output", None) is not None:
with tf.name_scope(self.dense_output.name):
self.dense_output.build(None)
# Copied from transformers.models.vit.modeling_tf_vit.TFViTIntermediate with ViT->ViTMAE
class TFViTMAEIntermediate(keras.layers.Layer):
def __init__(self, config: ViTMAEConfig, **kwargs):
super().__init__(**kwargs)
self.dense = keras.layers.Dense(
units=config.intermediate_size, kernel_initializer=get_initializer(config.initializer_range), name="dense"
)
if isinstance(config.hidden_act, str):
self.intermediate_act_fn = get_tf_activation(config.hidden_act)
else:
self.intermediate_act_fn = config.hidden_act
self.config = config
def call(self, hidden_states: tf.Tensor) -> tf.Tensor:
hidden_states = self.dense(inputs=hidden_states)
hidden_states = self.intermediate_act_fn(hidden_states)
return hidden_states
def build(self, input_shape=None):
if self.built:
return
self.built = True
if getattr(self, "dense", None) is not None:
with tf.name_scope(self.dense.name):
self.dense.build([None, None, self.config.hidden_size])
# Copied from transformers.models.vit.modeling_tf_vit.TFViTOutput with ViT->ViTMAE
class TFViTMAEOutput(keras.layers.Layer):
def __init__(self, config: ViTMAEConfig, **kwargs):
super().__init__(**kwargs)
self.dense = keras.layers.Dense(
units=config.hidden_size, kernel_initializer=get_initializer(config.initializer_range), name="dense"
)
self.dropout = keras.layers.Dropout(rate=config.hidden_dropout_prob)
self.config = config
def call(self, hidden_states: tf.Tensor, input_tensor: tf.Tensor, training: bool = False) -> tf.Tensor:
hidden_states = self.dense(inputs=hidden_states)
hidden_states = self.dropout(inputs=hidden_states, training=training)
hidden_states = hidden_states + input_tensor
return hidden_states
def build(self, input_shape=None):
if self.built:
return
self.built = True
if getattr(self, "dense", None) is not None:
with tf.name_scope(self.dense.name):
self.dense.build([None, None, self.config.intermediate_size])
# Copied from transformers.models.vit.modeling_tf_vit.TFViTLayer with ViT->ViTMAE
class TFViTMAELayer(keras.layers.Layer):
"""This corresponds to the Block class in the timm implementation."""
def __init__(self, config: ViTMAEConfig, **kwargs):
super().__init__(**kwargs)
self.attention = TFViTMAEAttention(config, name="attention")
self.intermediate = TFViTMAEIntermediate(config, name="intermediate")
self.vit_output = TFViTMAEOutput(config, name="output")
self.layernorm_before = keras.layers.LayerNormalization(epsilon=config.layer_norm_eps, name="layernorm_before")
self.layernorm_after = keras.layers.LayerNormalization(epsilon=config.layer_norm_eps, name="layernorm_after")
self.config = config
def call(
self,
hidden_states: tf.Tensor,
head_mask: tf.Tensor,
output_attentions: bool,
training: bool = False,
) -> Tuple[tf.Tensor]:
attention_outputs = self.attention(
# in ViTMAE, layernorm is applied before self-attention
input_tensor=self.layernorm_before(inputs=hidden_states),
head_mask=head_mask,
output_attentions=output_attentions,
training=training,
)
attention_output = attention_outputs[0]
# first residual connection
hidden_states = attention_output + hidden_states
# in ViTMAE, layernorm is also applied after self-attention
layer_output = self.layernorm_after(inputs=hidden_states)
intermediate_output = self.intermediate(hidden_states=layer_output)
# second residual connection is done here
layer_output = self.vit_output(
hidden_states=intermediate_output, input_tensor=hidden_states, training=training
)
outputs = (layer_output,) + attention_outputs[1:] # add attentions if we output them
return outputs
def build(self, input_shape=None):
if self.built:
return
self.built = True
if getattr(self, "attention", None) is not None:
with tf.name_scope(self.attention.name):
self.attention.build(None)
if getattr(self, "intermediate", None) is not None:
with tf.name_scope(self.intermediate.name):
self.intermediate.build(None)
if getattr(self, "vit_output", None) is not None:
with tf.name_scope(self.vit_output.name):
self.vit_output.build(None)
if getattr(self, "layernorm_before", None) is not None:
with tf.name_scope(self.layernorm_before.name):
self.layernorm_before.build([None, None, self.config.hidden_size])
if getattr(self, "layernorm_after", None) is not None:
with tf.name_scope(self.layernorm_after.name):
self.layernorm_after.build([None, None, self.config.hidden_size])
# Copied from transformers.models.vit.modeling_tf_vit.TFViTEncoder with ViT->ViTMAE
class TFViTMAEEncoder(keras.layers.Layer):
def __init__(self, config: ViTMAEConfig, **kwargs):
super().__init__(**kwargs)
self.layer = [TFViTMAELayer(config, name=f"layer_._{i}") for i in range(config.num_hidden_layers)]
def call(
self,
hidden_states: tf.Tensor,
head_mask: tf.Tensor,
output_attentions: bool,
output_hidden_states: bool,
return_dict: bool,
training: bool = False,
) -> Union[TFBaseModelOutput, Tuple[tf.Tensor]]:
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_outputs = layer_module(
hidden_states=hidden_states,
head_mask=head_mask[i],
output_attentions=output_attentions,
training=training,
)
hidden_states = layer_outputs[0]
if output_attentions:
all_attentions = all_attentions + (layer_outputs[1],)
# Add last layer
if output_hidden_states:
all_hidden_states = all_hidden_states + (hidden_states,)
if not return_dict:
return tuple(v for v in [hidden_states, all_hidden_states, all_attentions] if v is not None)
return TFBaseModelOutput(
last_hidden_state=hidden_states, hidden_states=all_hidden_states, attentions=all_attentions
)
def build(self, input_shape=None):
if self.built:
return
self.built = True
if getattr(self, "layer", None) is not None:
for layer in self.layer:
with tf.name_scope(layer.name):
layer.build(None)
@keras_serializable
class TFViTMAEMainLayer(keras.layers.Layer):
config_class = ViTMAEConfig
def __init__(self, config: ViTMAEConfig, **kwargs):
super().__init__(**kwargs)
self.config = config
self.embeddings = TFViTMAEEmbeddings(config, name="embeddings")
self.encoder = TFViTMAEEncoder(config, name="encoder")
self.layernorm = keras.layers.LayerNormalization(epsilon=config.layer_norm_eps, name="layernorm")
def get_input_embeddings(self) -> keras.layers.Layer:
return self.embeddings.patch_embeddings
def _prune_heads(self, heads_to_prune):
"""
Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
class PreTrainedModel
"""
raise NotImplementedError
@unpack_inputs
def call(
self,
pixel_values: TFModelInputType | None = None,
noise: tf.Tensor = None,
head_mask: np.ndarray | tf.Tensor | None = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
training: bool = False,
interpolate_pos_encoding: bool = False,
) -> Union[TFViTMAEModelOutput, Tuple[tf.Tensor]]:
embedding_output, mask, ids_restore = self.embeddings(
pixel_values=pixel_values,
training=training,
noise=noise,
interpolate_pos_encoding=interpolate_pos_encoding,
)
# 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]
if head_mask is not None:
raise NotImplementedError
else:
head_mask = [None] * self.config.num_hidden_layers
encoder_outputs = self.encoder(
embedding_output,
head_mask=head_mask,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
training=training,
)
sequence_output = encoder_outputs[0]
sequence_output = self.layernorm(inputs=sequence_output)
if not return_dict:
return (sequence_output, mask, ids_restore) + encoder_outputs[1:]
return TFViTMAEModelOutput(
last_hidden_state=sequence_output,
mask=mask,
ids_restore=ids_restore,
hidden_states=encoder_outputs.hidden_states,
attentions=encoder_outputs.attentions,
)
def build(self, input_shape=None):
if self.built:
return
self.built = True
if getattr(self, "embeddings", None) is not None:
with tf.name_scope(self.embeddings.name):
self.embeddings.build(None)
if getattr(self, "encoder", None) is not None:
with tf.name_scope(self.encoder.name):
self.encoder.build(None)
if getattr(self, "layernorm", None) is not None:
with tf.name_scope(self.layernorm.name):
self.layernorm.build([None, None, self.config.hidden_size])
class TFViTMAEPreTrainedModel(TFPreTrainedModel):
"""
An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
models.
"""
config_class = ViTMAEConfig
base_model_prefix = "vit"
main_input_name = "pixel_values"
VIT_MAE_START_DOCSTRING = r"""
This model inherits from [`TFPreTrainedModel`]. Check the superclass documentation for the generic methods the
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
etc.)
This model is also a [keras.Model](https://www.tensorflow.org/api_docs/python/tf/keras/Model) subclass. Use it
as a regular TF 2.0 Keras Model and refer to the TF 2.0 documentation for all matter related to general usage and
behavior.
<Tip>
TensorFlow models and layers in `transformers` accept two formats as input:
- having all inputs as keyword arguments (like PyTorch models), or
- having all inputs as a list, tuple or dict in the first positional argument.
The reason the second format is supported is that Keras methods prefer this format when passing inputs to models
and layers. Because of this support, when using methods like `model.fit()` things should "just work" for you - just
pass your inputs and labels in any format that `model.fit()` supports! If, however, you want to use the second
format outside of Keras methods like `fit()` and `predict()`, such as when creating your own layers or models with
the Keras `Functional` API, there are three possibilities you can use to gather all the input Tensors in the first
positional argument:
- a single Tensor with `pixel_values` only and nothing else: `model(pixel_values)`
- a list of varying length with one or several input Tensors IN THE ORDER given in the docstring:
`model([pixel_values, attention_mask])` or `model([pixel_values, attention_mask, token_type_ids])`
- a dictionary with one or several input Tensors associated to the input names given in the docstring:
`model({"pixel_values": pixel_values, "token_type_ids": token_type_ids})`
Note that when creating models and layers with
[subclassing](https://keras.io/guides/making_new_layers_and_models_via_subclassing/) then you don't need to worry
about any of this, as you can just pass inputs like you would to any other Python function!
</Tip>
Args:
config ([`ViTMAEConfig`]): Model configuration class with all the parameters of the model.
Initializing with a config file does not load the weights associated with the model, only the
configuration. Check out the [`~TFPreTrainedModel.from_pretrained`] method to load the model weights.
"""
VIT_MAE_INPUTS_DOCSTRING = r"""
Args:
pixel_values (`np.ndarray`, `tf.Tensor`, `List[tf.Tensor]` ``Dict[str, tf.Tensor]` or `Dict[str, np.ndarray]` and each example must have the shape `(batch_size, num_channels, height, width)`):
Pixel values. Pixel values can be obtained using [`AutoImageProcessor`]. See [`ViTImageProcessor.__call__`]
for details.
head_mask (`np.ndarray` or `tf.Tensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:
- 1 indicates the head is **not masked**,
- 0 indicates the head is **masked**.
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
tensors for more detail. This argument can be used only in eager mode, in graph mode the value in the
config will be used instead.
output_hidden_states (`bool`, *optional*):
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
more detail. This argument can be used only in eager mode, in graph mode the value in the config will be
used instead.
return_dict (`bool`, *optional*):
Whether or not to return a [`~file_utils.ModelOutput`] instead of a plain tuple. This argument can be used
in eager mode, in graph mode the value will always be set to True.
training (`bool`, *optional*, defaults to `False``):
Whether or not to use the model in training mode (some modules like dropout modules have different
behaviors between training and evaluation).
interpolate_pos_encoding (`bool`, *optional*, defaults to `False`):
Whether to interpolate the position encodings at the encoder and decoder.
"""
@add_start_docstrings(
"The bare ViTMAE Model transformer outputting raw hidden-states without any specific head on top.",
VIT_MAE_START_DOCSTRING,
)
class TFViTMAEModel(TFViTMAEPreTrainedModel):
def __init__(self, config: ViTMAEConfig, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs)
self.vit = TFViTMAEMainLayer(config, name="vit")
def get_input_embeddings(self):
return self.vit.get_input_embeddings()
@unpack_inputs
@add_start_docstrings_to_model_forward(VIT_MAE_INPUTS_DOCSTRING)
@replace_return_docstrings(output_type=TFViTMAEModelOutput, config_class=_CONFIG_FOR_DOC)
def call(
self,
pixel_values: TFModelInputType | None = None,
noise: tf.Tensor = None,
head_mask: np.ndarray | tf.Tensor | None = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
training: bool = False,
interpolate_pos_encoding: bool = False,
) -> Union[TFViTMAEModelOutput, Tuple[tf.Tensor]]:
r"""
Returns:
Examples:
```python
>>> from transformers import AutoImageProcessor, TFViTMAEModel
>>> from PIL import Image
>>> import requests
>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
>>> image = Image.open(requests.get(url, stream=True).raw)
>>> image_processor = AutoImageProcessor.from_pretrained("facebook/vit-mae-base")
>>> model = TFViTMAEModel.from_pretrained("facebook/vit-mae-base")
>>> inputs = image_processor(images=image, return_tensors="tf")
>>> outputs = model(**inputs)
>>> last_hidden_states = outputs.last_hidden_state
```"""
outputs = self.vit(
pixel_values=pixel_values,
noise=noise,
head_mask=head_mask,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
training=training,
interpolate_pos_encoding=interpolate_pos_encoding,
)
return outputs
def build(self, input_shape=None):
if self.built:
return
self.built = True
if getattr(self, "vit", None) is not None:
with tf.name_scope(self.vit.name):
self.vit.build(None)
class TFViTMAEDecoder(keras.layers.Layer):
def __init__(self, config, num_patches, **kwargs):
super().__init__(**kwargs)
self.decoder_embed = keras.layers.Dense(config.decoder_hidden_size, name="decoder_embed")
decoder_config = deepcopy(config)
decoder_config.hidden_size = config.decoder_hidden_size
decoder_config.num_hidden_layers = config.decoder_num_hidden_layers
decoder_config.num_attention_heads = config.decoder_num_attention_heads
decoder_config.intermediate_size = config.decoder_intermediate_size
self.decoder_layers = [
TFViTMAELayer(decoder_config, name=f"decoder_layers.{j}") for j in range(config.decoder_num_hidden_layers)
]
self.decoder_norm = keras.layers.LayerNormalization(epsilon=config.layer_norm_eps, name="decoder_norm")
self.decoder_pred = keras.layers.Dense(
config.patch_size**2 * config.num_channels,
kernel_initializer=get_initializer(config.initializer_range),
name="decoder_pred",
) # encoder to decoder
self.config = config
self.num_patches = num_patches
def build(self, input_shape=None):
self.mask_token = self.add_weight(
shape=(1, 1, self.config.decoder_hidden_size),
initializer=tf.random_normal_initializer(stddev=self.config.initializer_range),
trainable=True,
name="mask_token",
)
self.decoder_pos_embed = self.add_weight(
shape=(1, self.num_patches + 1, self.config.decoder_hidden_size),
initializer="zeros",
trainable=False,
name="decoder_pos_embed",
)
decoder_pos_embed = get_2d_sincos_pos_embed(
self.decoder_pos_embed.shape[-1],
int(self.num_patches**0.5),
add_cls_token=True,
)[None, ...]
self.decoder_pos_embed.assign(decoder_pos_embed)
if self.built:
return
self.built = True
if getattr(self, "decoder_embed", None) is not None:
with tf.name_scope(self.decoder_embed.name):
self.decoder_embed.build([None, None, self.config.hidden_size])
if getattr(self, "decoder_norm", None) is not None:
with tf.name_scope(self.decoder_norm.name):
self.decoder_norm.build([None, None, self.config.decoder_hidden_size])
if getattr(self, "decoder_pred", None) is not None:
with tf.name_scope(self.decoder_pred.name):
self.decoder_pred.build([None, None, self.config.decoder_hidden_size])
if getattr(self, "decoder_layers", None) is not None:
for layer in self.decoder_layers:
with tf.name_scope(layer.name):
layer.build(None)
def interpolate_pos_encoding(self, embeddings) -> tf.Tensor:
"""
This method is a modified version of the interpolation function for ViT-mae model at the deocder, that
allows to interpolate the pre-trained decoder position encodings, to be able to use the model on higher
resolution images.
Source:
https://github.com/facebookresearch/dino/blob/de9ee3df6cf39fac952ab558447af1fa1365362a/vision_transformer.py#L174
"""
# [batch_size, num_patches + 1, hidden_size]
_, num_positions, dim = shape_list(self.decoder_pos_embed)
# -1 removes the class dimension since we later append it without interpolation
seq_len = shape_list(embeddings)[1] - 1
num_positions = num_positions - 1
# Separation of class token and patch tokens
class_pos_embed = self.decoder_pos_embed[:, :1, :]
patch_pos_embed = self.decoder_pos_embed[:, 1:, :]
# interpolate the position embeddings
patch_pos_embed = tf.image.resize(
images=tf.reshape(patch_pos_embed, shape=(1, 1, -1, dim)),
size=(1, seq_len),
method="bicubic",
)
# [1, seq_len, hidden_size]
patch_pos_embed = tf.reshape(tensor=patch_pos_embed, shape=(1, -1, dim))
# Adding the class token back
return tf.concat(values=(class_pos_embed, patch_pos_embed), axis=1)
def call(
self,
hidden_states,
ids_restore,
output_attentions=False,
output_hidden_states=False,
return_dict=True,
interpolate_pos_encoding=False,
):
# embed tokens
x = self.decoder_embed(hidden_states)
# append mask tokens to sequence
mask_tokens = tf.tile(
self.mask_token,
(shape_list(x)[0], shape_list(ids_restore)[1] + 1 - shape_list(x)[1], 1),
)
x_ = tf.concat([x[:, 1:, :], mask_tokens], axis=1) # no cls token
x_ = tf.gather(x_, axis=1, batch_dims=1, indices=ids_restore) # unshuffle
x = tf.concat([x[:, :1, :], x_], axis=1) # append cls token
if interpolate_pos_encoding:
decoder_pos_embed = self.interpolate_pos_encoding(x)
else:
decoder_pos_embed = self.decoder_pos_embed
# add pos embed
hidden_states = x + decoder_pos_embed
# apply Transformer layers (blocks)
all_hidden_states = () if output_hidden_states else None
all_self_attentions = () if output_attentions else None
for i, layer_module in enumerate(self.decoder_layers):
if output_hidden_states:
all_hidden_states = all_hidden_states + (hidden_states,)
layer_outputs = layer_module(
hidden_states,
head_mask=None,
output_attentions=output_attentions,
)
hidden_states = layer_outputs[0]
if output_attentions:
all_self_attentions = all_self_attentions + (layer_outputs[1],)
if output_hidden_states:
all_hidden_states = all_hidden_states + (hidden_states,)
hidden_states = self.decoder_norm(hidden_states)
# predictor projection
logits = self.decoder_pred(hidden_states)
# remove cls token
logits = logits[:, 1:, :]
if not return_dict:
return tuple(v for v in [logits, all_hidden_states, all_self_attentions] if v is not None)
return TFViTMAEDecoderOutput(logits=logits, hidden_states=all_hidden_states, attentions=all_self_attentions)
@add_start_docstrings(
"The ViTMAE Model transformer with the decoder on top for self-supervised pre-training.",
VIT_MAE_START_DOCSTRING,
)
class TFViTMAEForPreTraining(TFViTMAEPreTrainedModel):
def __init__(self, config):
super().__init__(config)
self.config = config
self.vit = TFViTMAEMainLayer(config, name="vit")
self.decoder = TFViTMAEDecoder(
config,
num_patches=self.vit.embeddings.num_patches,
name="decoder",
)
def get_input_embeddings(self):
return self.vit.get_input_embeddings()
def _prune_heads(self, heads_to_prune):
raise NotImplementedError
def patchify(self, pixel_values, interpolate_pos_encoding: bool = False):
"""
Args:
pixel_values (`tf.Tensor` of shape `(batch_size, height, width, num_channels)` or `(batch_size, num_channels, height, width)`):
Pixel values.
interpolate_pos_encoding (`bool`, default `False`):
interpolation flag passed during the forward pass.
Returns:
`tf.Tensor` of shape `(batch_size, num_patches, patch_size**2 * num_channels)`:
Patchified pixel values.
"""
patch_size, num_channels = self.config.patch_size, self.config.num_channels
# make sure channels are last
if shape_list(pixel_values)[1] == num_channels:
pixel_values = tf.transpose(pixel_values, perm=(0, 2, 3, 1))
# sanity checks
if not interpolate_pos_encoding:
tf.debugging.assert_equal(
shape_list(pixel_values)[1],
shape_list(pixel_values)[2],
message="Make sure the pixel values have a squared size",
)
tf.debugging.assert_equal(
shape_list(pixel_values)[1] % patch_size,
0,
message="Make sure the pixel values have a size that is divisible by the patch size",
)
tf.debugging.assert_equal(
shape_list(pixel_values)[3],
num_channels,
message=(
"Make sure the number of channels of the pixel values is equal to the one set in the configuration"
),
)
# patchify
batch_size = shape_list(pixel_values)[0]
num_patches_h = shape_list(pixel_values)[1] // patch_size
num_patches_w = shape_list(pixel_values)[2] // patch_size
patchified_pixel_values = tf.reshape(
pixel_values,
(batch_size, num_patches_h, patch_size, num_patches_w, patch_size, num_channels),
)
patchified_pixel_values = tf.einsum("nhpwqc->nhwpqc", patchified_pixel_values)
patchified_pixel_values = tf.reshape(
patchified_pixel_values,
(batch_size, num_patches_h * num_patches_w, patch_size**2 * num_channels),
)
return patchified_pixel_values
def unpatchify(self, patchified_pixel_values, original_image_size: Optional[Tuple[int, int]] = None):
"""
Args:
patchified_pixel_values (`tf.Tensor` of shape `(batch_size, num_patches, patch_size**2 * num_channels)`:
Patchified pixel values.
original_image_size (`Tuple[int, int]`, *optional*):
Original image size.
Returns:
`tf.Tensor` of shape `(batch_size, height, width, num_channels)`:
Pixel values.
"""
patch_size, num_channels = self.config.patch_size, self.config.num_channels
original_image_size = (
original_image_size
if original_image_size is not None
else (self.config.image_size, self.config.image_size)
)
original_height, original_width = original_image_size
num_patches_h = original_height // patch_size
num_patches_w = original_width // patch_size
# sanity check
tf.debugging.assert_equal(
num_patches_h * num_patches_w,
shape_list(patchified_pixel_values)[1],
message=f"The number of patches in the patchified pixel values is {shape_list(patchified_pixel_values)[1]} does not match the patches of original image {num_patches_w}*{num_patches_h}",
)
# unpatchify
batch_size = shape_list(patchified_pixel_values)[0]
patchified_pixel_values = tf.reshape(
patchified_pixel_values,
(batch_size, num_patches_h, num_patches_w, patch_size, patch_size, num_channels),
)
patchified_pixel_values = tf.einsum("nhwpqc->nhpwqc", patchified_pixel_values)
pixel_values = tf.reshape(
patchified_pixel_values,
(batch_size, num_patches_h * patch_size, num_patches_w * patch_size, num_channels),
)
return pixel_values
def forward_loss(self, pixel_values, pred, mask, interpolate_pos_encoding: bool = False):
"""
Args:
pixel_values (`tf.Tensor` of shape `(batch_size, height, width, num_channels)`):
Pixel values.
pred (`tf.Tensor` of shape `(batch_size, num_patches, patch_size**2 * num_channels)`:
Predicted pixel values.
mask (`tf.Tensor` of shape `(batch_size, sequence_length)`):
Tensor indicating which patches are masked (1) and which are not (0).
interpolate_pos_encoding (`bool`, *optional*, default `False`):
interpolation flag passed during the forward pass.
Returns:
`tf.Tensor`: Pixel reconstruction loss.
"""
target = self.patchify(pixel_values, interpolate_pos_encoding=interpolate_pos_encoding)
if self.config.norm_pix_loss:
mean = tf.reduce_mean(target, axis=-1, keepdims=True)
var = tf.math.reduce_variance(target, axis=-1, keepdims=True)
target = (target - mean) / (var + 1.0e-6) ** 0.5
loss = (pred - target) ** 2
loss = tf.reduce_mean(loss, axis=-1) # [batch_size, num_patches], mean loss per patch
loss = tf.reduce_sum(loss * mask) / tf.reduce_sum(mask) # mean loss on removed patches
loss = tf.reshape(loss, (1,))
return loss
@unpack_inputs
@add_start_docstrings_to_model_forward(VIT_MAE_INPUTS_DOCSTRING)
@replace_return_docstrings(output_type=TFViTMAEForPreTrainingOutput, config_class=_CONFIG_FOR_DOC)
def call(
self,
pixel_values: TFModelInputType | None = None,
noise: tf.Tensor = None,
head_mask: np.ndarray | tf.Tensor | None = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
training: bool = False,
interpolate_pos_encoding: bool = False,
) -> Union[TFViTMAEForPreTrainingOutput, Tuple[tf.Tensor]]:
r"""
Returns:
Examples:
```python
>>> from transformers import AutoImageProcessor, TFViTMAEForPreTraining
>>> from PIL import Image
>>> import requests
>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
>>> image = Image.open(requests.get(url, stream=True).raw)
>>> image_processor = AutoImageProcessor.from_pretrained("facebook/vit-mae-base")
>>> model = TFViTMAEForPreTraining.from_pretrained("facebook/vit-mae-base")
>>> inputs = image_processor(images=image, return_tensors="pt")
>>> outputs = model(**inputs)
>>> loss = outputs.loss
>>> mask = outputs.mask
>>> ids_restore = outputs.ids_restore
```"""
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
outputs = self.vit(
pixel_values=pixel_values,
noise=noise,
head_mask=head_mask,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
training=training,
interpolate_pos_encoding=interpolate_pos_encoding,
)
latent = outputs.last_hidden_state
ids_restore = outputs.ids_restore
mask = outputs.mask
# [batch_size, num_patches, patch_size**2*3]
decoder_outputs = self.decoder(latent, ids_restore, interpolate_pos_encoding=interpolate_pos_encoding)
logits = decoder_outputs.logits
loss = self.forward_loss(pixel_values, logits, mask, interpolate_pos_encoding=interpolate_pos_encoding)
if not return_dict:
output = (logits, mask, ids_restore) + outputs[2:]
return ((loss,) + output) if loss is not None else output
return TFViTMAEForPreTrainingOutput(
loss=loss,
logits=logits,
mask=mask,
ids_restore=ids_restore,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
def build(self, input_shape=None):
if self.built:
return
self.built = True
if getattr(self, "vit", None) is not None:
with tf.name_scope(self.vit.name):
self.vit.build(None)
if getattr(self, "decoder", None) is not None:
with tf.name_scope(self.decoder.name):
self.decoder.build(None)
|
transformers/src/transformers/models/vit_mae/modeling_tf_vit_mae.py/0
|
{
"file_path": "transformers/src/transformers/models/vit_mae/modeling_tf_vit_mae.py",
"repo_id": "transformers",
"token_count": 24969
}
| 388
|
# coding=utf-8
# Copyright 2023 The HuggingFace Inc. team. 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.
"""Convert VITS checkpoint."""
import argparse
import json
import tempfile
import torch
from huggingface_hub import hf_hub_download
from transformers import VitsConfig, VitsModel, VitsTokenizer, logging
logging.set_verbosity_info()
logger = logging.get_logger("transformers.models.vits")
MAPPING_TEXT_ENCODER = {
"enc_p.emb": "text_encoder.embed_tokens",
"enc_p.encoder.attn_layers.*.conv_k": "text_encoder.encoder.layers.*.attention.k_proj",
"enc_p.encoder.attn_layers.*.conv_v": "text_encoder.encoder.layers.*.attention.v_proj",
"enc_p.encoder.attn_layers.*.conv_q": "text_encoder.encoder.layers.*.attention.q_proj",
"enc_p.encoder.attn_layers.*.conv_o": "text_encoder.encoder.layers.*.attention.out_proj",
"enc_p.encoder.attn_layers.*.emb_rel_k": "text_encoder.encoder.layers.*.attention.emb_rel_k",
"enc_p.encoder.attn_layers.*.emb_rel_v": "text_encoder.encoder.layers.*.attention.emb_rel_v",
"enc_p.encoder.norm_layers_1.*.gamma": "text_encoder.encoder.layers.*.layer_norm.weight",
"enc_p.encoder.norm_layers_1.*.beta": "text_encoder.encoder.layers.*.layer_norm.bias",
"enc_p.encoder.ffn_layers.*.conv_1": "text_encoder.encoder.layers.*.feed_forward.conv_1",
"enc_p.encoder.ffn_layers.*.conv_2": "text_encoder.encoder.layers.*.feed_forward.conv_2",
"enc_p.encoder.norm_layers_2.*.gamma": "text_encoder.encoder.layers.*.final_layer_norm.weight",
"enc_p.encoder.norm_layers_2.*.beta": "text_encoder.encoder.layers.*.final_layer_norm.bias",
"enc_p.proj": "text_encoder.project",
}
MAPPING_STOCHASTIC_DURATION_PREDICTOR = {
"dp.pre": "duration_predictor.conv_pre",
"dp.proj": "duration_predictor.conv_proj",
"dp.convs.convs_sep.*": "duration_predictor.conv_dds.convs_dilated.*",
"dp.convs.convs_1x1.*": "duration_predictor.conv_dds.convs_pointwise.*",
"dp.convs.norms_1.*.gamma": "duration_predictor.conv_dds.norms_1.*.weight",
"dp.convs.norms_1.*.beta": "duration_predictor.conv_dds.norms_1.*.bias",
"dp.convs.norms_2.*.gamma": "duration_predictor.conv_dds.norms_2.*.weight",
"dp.convs.norms_2.*.beta": "duration_predictor.conv_dds.norms_2.*.bias",
"dp.flows.0.logs": "duration_predictor.flows.0.log_scale",
"dp.flows.0.m": "duration_predictor.flows.0.translate",
"dp.flows.*.pre": "duration_predictor.flows.*.conv_pre",
"dp.flows.*.proj": "duration_predictor.flows.*.conv_proj",
"dp.flows.*.convs.convs_1x1.0": "duration_predictor.flows.*.conv_dds.convs_pointwise.0",
"dp.flows.*.convs.convs_1x1.1": "duration_predictor.flows.*.conv_dds.convs_pointwise.1",
"dp.flows.*.convs.convs_1x1.2": "duration_predictor.flows.*.conv_dds.convs_pointwise.2",
"dp.flows.*.convs.convs_sep.0": "duration_predictor.flows.*.conv_dds.convs_dilated.0",
"dp.flows.*.convs.convs_sep.1": "duration_predictor.flows.*.conv_dds.convs_dilated.1",
"dp.flows.*.convs.convs_sep.2": "duration_predictor.flows.*.conv_dds.convs_dilated.2",
"dp.flows.*.convs.norms_1.0.gamma": "duration_predictor.flows.*.conv_dds.norms_1.0.weight",
"dp.flows.*.convs.norms_1.0.beta": "duration_predictor.flows.*.conv_dds.norms_1.0.bias",
"dp.flows.*.convs.norms_1.1.gamma": "duration_predictor.flows.*.conv_dds.norms_1.1.weight",
"dp.flows.*.convs.norms_1.1.beta": "duration_predictor.flows.*.conv_dds.norms_1.1.bias",
"dp.flows.*.convs.norms_1.2.gamma": "duration_predictor.flows.*.conv_dds.norms_1.2.weight",
"dp.flows.*.convs.norms_1.2.beta": "duration_predictor.flows.*.conv_dds.norms_1.2.bias",
"dp.flows.*.convs.norms_2.0.gamma": "duration_predictor.flows.*.conv_dds.norms_2.0.weight",
"dp.flows.*.convs.norms_2.0.beta": "duration_predictor.flows.*.conv_dds.norms_2.0.bias",
"dp.flows.*.convs.norms_2.1.gamma": "duration_predictor.flows.*.conv_dds.norms_2.1.weight",
"dp.flows.*.convs.norms_2.1.beta": "duration_predictor.flows.*.conv_dds.norms_2.1.bias",
"dp.flows.*.convs.norms_2.2.gamma": "duration_predictor.flows.*.conv_dds.norms_2.2.weight",
"dp.flows.*.convs.norms_2.2.beta": "duration_predictor.flows.*.conv_dds.norms_2.2.bias",
"dp.post_pre": "duration_predictor.post_conv_pre",
"dp.post_proj": "duration_predictor.post_conv_proj",
"dp.post_convs.convs_sep.*": "duration_predictor.post_conv_dds.convs_dilated.*",
"dp.post_convs.convs_1x1.*": "duration_predictor.post_conv_dds.convs_pointwise.*",
"dp.post_convs.norms_1.*.gamma": "duration_predictor.post_conv_dds.norms_1.*.weight",
"dp.post_convs.norms_1.*.beta": "duration_predictor.post_conv_dds.norms_1.*.bias",
"dp.post_convs.norms_2.*.gamma": "duration_predictor.post_conv_dds.norms_2.*.weight",
"dp.post_convs.norms_2.*.beta": "duration_predictor.post_conv_dds.norms_2.*.bias",
"dp.post_flows.0.logs": "duration_predictor.post_flows.0.log_scale",
"dp.post_flows.0.m": "duration_predictor.post_flows.0.translate",
"dp.post_flows.*.pre": "duration_predictor.post_flows.*.conv_pre",
"dp.post_flows.*.proj": "duration_predictor.post_flows.*.conv_proj",
"dp.post_flows.*.convs.convs_1x1.0": "duration_predictor.post_flows.*.conv_dds.convs_pointwise.0",
"dp.post_flows.*.convs.convs_1x1.1": "duration_predictor.post_flows.*.conv_dds.convs_pointwise.1",
"dp.post_flows.*.convs.convs_1x1.2": "duration_predictor.post_flows.*.conv_dds.convs_pointwise.2",
"dp.post_flows.*.convs.convs_sep.0": "duration_predictor.post_flows.*.conv_dds.convs_dilated.0",
"dp.post_flows.*.convs.convs_sep.1": "duration_predictor.post_flows.*.conv_dds.convs_dilated.1",
"dp.post_flows.*.convs.convs_sep.2": "duration_predictor.post_flows.*.conv_dds.convs_dilated.2",
"dp.post_flows.*.convs.norms_1.0.gamma": "duration_predictor.post_flows.*.conv_dds.norms_1.0.weight",
"dp.post_flows.*.convs.norms_1.0.beta": "duration_predictor.post_flows.*.conv_dds.norms_1.0.bias",
"dp.post_flows.*.convs.norms_1.1.gamma": "duration_predictor.post_flows.*.conv_dds.norms_1.1.weight",
"dp.post_flows.*.convs.norms_1.1.beta": "duration_predictor.post_flows.*.conv_dds.norms_1.1.bias",
"dp.post_flows.*.convs.norms_1.2.gamma": "duration_predictor.post_flows.*.conv_dds.norms_1.2.weight",
"dp.post_flows.*.convs.norms_1.2.beta": "duration_predictor.post_flows.*.conv_dds.norms_1.2.bias",
"dp.post_flows.*.convs.norms_2.0.gamma": "duration_predictor.post_flows.*.conv_dds.norms_2.0.weight",
"dp.post_flows.*.convs.norms_2.0.beta": "duration_predictor.post_flows.*.conv_dds.norms_2.0.bias",
"dp.post_flows.*.convs.norms_2.1.gamma": "duration_predictor.post_flows.*.conv_dds.norms_2.1.weight",
"dp.post_flows.*.convs.norms_2.1.beta": "duration_predictor.post_flows.*.conv_dds.norms_2.1.bias",
"dp.post_flows.*.convs.norms_2.2.gamma": "duration_predictor.post_flows.*.conv_dds.norms_2.2.weight",
"dp.post_flows.*.convs.norms_2.2.beta": "duration_predictor.post_flows.*.conv_dds.norms_2.2.bias",
"dp.cond": "duration_predictor.cond", # num_speakers > 1
}
MAPPING_FLOW = {
"flow.flows.*.pre": "flow.flows.*.conv_pre",
"flow.flows.*.enc.in_layers.0": "flow.flows.*.wavenet.in_layers.0",
"flow.flows.*.enc.in_layers.1": "flow.flows.*.wavenet.in_layers.1",
"flow.flows.*.enc.in_layers.2": "flow.flows.*.wavenet.in_layers.2",
"flow.flows.*.enc.in_layers.3": "flow.flows.*.wavenet.in_layers.3",
"flow.flows.*.enc.res_skip_layers.0": "flow.flows.*.wavenet.res_skip_layers.0",
"flow.flows.*.enc.res_skip_layers.1": "flow.flows.*.wavenet.res_skip_layers.1",
"flow.flows.*.enc.res_skip_layers.2": "flow.flows.*.wavenet.res_skip_layers.2",
"flow.flows.*.enc.res_skip_layers.3": "flow.flows.*.wavenet.res_skip_layers.3",
"flow.flows.*.enc.cond_layer": "flow.flows.*.wavenet.cond_layer", # num_speakers > 1
"flow.flows.*.post": "flow.flows.*.conv_post",
}
MAPPING_GENERATOR = {
"dec.conv_pre": "decoder.conv_pre",
"dec.ups.0": "decoder.upsampler.0",
"dec.ups.1": "decoder.upsampler.1",
"dec.ups.2": "decoder.upsampler.2",
"dec.ups.3": "decoder.upsampler.3",
"dec.resblocks.*.convs1.0": "decoder.resblocks.*.convs1.0",
"dec.resblocks.*.convs1.1": "decoder.resblocks.*.convs1.1",
"dec.resblocks.*.convs1.2": "decoder.resblocks.*.convs1.2",
"dec.resblocks.*.convs2.0": "decoder.resblocks.*.convs2.0",
"dec.resblocks.*.convs2.1": "decoder.resblocks.*.convs2.1",
"dec.resblocks.*.convs2.2": "decoder.resblocks.*.convs2.2",
"dec.conv_post": "decoder.conv_post",
"dec.cond": "decoder.cond", # num_speakers > 1
}
MAPPING_POSTERIOR_ENCODER = {
"enc_q.pre": "posterior_encoder.conv_pre",
"enc_q.enc.in_layers.*": "posterior_encoder.wavenet.in_layers.*",
"enc_q.enc.res_skip_layers.*": "posterior_encoder.wavenet.res_skip_layers.*",
"enc_q.enc.cond_layer": "posterior_encoder.wavenet.cond_layer", # num_speakers > 1
"enc_q.proj": "posterior_encoder.conv_proj",
}
MAPPING = {
**MAPPING_TEXT_ENCODER,
**MAPPING_STOCHASTIC_DURATION_PREDICTOR,
**MAPPING_FLOW,
**MAPPING_GENERATOR,
**MAPPING_POSTERIOR_ENCODER,
"emb_g": "embed_speaker", # num_speakers > 1
}
TOP_LEVEL_KEYS = []
IGNORE_KEYS = []
def set_recursively(hf_pointer, key, value, full_name, weight_type):
for attribute in key.split("."):
hf_pointer = getattr(hf_pointer, attribute)
if weight_type is not None:
hf_shape = getattr(hf_pointer, weight_type).shape
else:
hf_shape = hf_pointer.shape
# strip off the kernel dimension at the end (original weights are Conv1d)
if key.endswith(".k_proj") or key.endswith(".v_proj") or key.endswith(".q_proj") or key.endswith(".out_proj"):
value = value.squeeze(-1)
if hf_shape != value.shape:
raise ValueError(
f"Shape of hf {key + '.' + weight_type if weight_type is not None else ''} is {hf_shape}, but should be"
f" {value.shape} for {full_name}"
)
if weight_type == "weight":
hf_pointer.weight.data = value
elif weight_type == "weight_g":
hf_pointer.weight_g.data = value
elif weight_type == "weight_v":
hf_pointer.weight_v.data = value
elif weight_type == "bias":
hf_pointer.bias.data = value
elif weight_type == "running_mean":
hf_pointer.running_mean.data = value
elif weight_type == "running_var":
hf_pointer.running_var.data = value
elif weight_type == "num_batches_tracked":
hf_pointer.num_batches_tracked.data = value
else:
hf_pointer.data = value
logger.info(f"{key + ('.' + weight_type if weight_type is not None else '')} was initialized from {full_name}.")
def should_ignore(name, ignore_keys):
for key in ignore_keys:
if key.endswith(".*"):
if name.startswith(key[:-1]):
return True
elif ".*." in key:
prefix, suffix = key.split(".*.")
if prefix in name and suffix in name:
return True
elif key in name:
return True
return False
def recursively_load_weights(fairseq_dict, hf_model):
unused_weights = []
for name, value in fairseq_dict.items():
if should_ignore(name, IGNORE_KEYS):
logger.info(f"{name} was ignored")
continue
is_used = False
for key, mapped_key in MAPPING.items():
if key.endswith(".*"):
key = key[:-1]
elif "*" in key:
prefix, suffix = key.split(".*.")
if prefix in name and suffix in name:
key = suffix
if key in name:
is_used = True
if mapped_key.endswith(".*"):
layer_index = name.split(key)[-1].split(".")[0]
mapped_key = mapped_key.replace("*", layer_index)
elif "*" in mapped_key:
layer_index = name.split(key)[0].split(".")[-2]
# remap the layer index since we removed the Flip layers
if "flow.flows" in mapped_key:
layer_index = str(int(layer_index) // 2)
if "duration_predictor.flows" in mapped_key or "duration_predictor.post_flows" in mapped_key:
layer_index = str(int(layer_index) // 2 + 1)
mapped_key = mapped_key.replace("*", layer_index)
if "weight_g" in name:
weight_type = "weight_g"
elif "weight_v" in name:
weight_type = "weight_v"
elif "bias" in name:
weight_type = "bias"
elif "weight" in name:
weight_type = "weight"
elif "running_mean" in name:
weight_type = "running_mean"
elif "running_var" in name:
weight_type = "running_var"
elif "num_batches_tracked" in name:
weight_type = "num_batches_tracked"
else:
weight_type = None
set_recursively(hf_model, mapped_key, value, name, weight_type)
continue
if not is_used:
unused_weights.append(name)
logger.warning(f"Unused weights: {unused_weights}")
@torch.no_grad()
def convert_checkpoint(
pytorch_dump_folder_path,
checkpoint_path=None,
config_path=None,
vocab_path=None,
language=None,
num_speakers=None,
sampling_rate=None,
repo_id=None,
):
"""
Copy/paste/tweak model's weights to transformers design.
"""
if config_path is not None:
config = VitsConfig.from_pretrained(config_path)
else:
config = VitsConfig()
if num_speakers:
config.num_speakers = num_speakers
config.speaker_embedding_size = 256
if sampling_rate:
config.sampling_rate = sampling_rate
if checkpoint_path is None:
logger.info(f"***Converting model: facebook/mms-tts {language}***")
vocab_path = hf_hub_download(
repo_id="facebook/mms-tts",
filename="vocab.txt",
subfolder=f"models/{language}",
)
config_file = hf_hub_download(
repo_id="facebook/mms-tts",
filename="config.json",
subfolder=f"models/{language}",
)
checkpoint_path = hf_hub_download(
repo_id="facebook/mms-tts",
filename="G_100000.pth",
subfolder=f"models/{language}",
)
with open(config_file, "r") as f:
data = f.read()
hps = json.loads(data)
is_uroman = hps["data"]["training_files"].split(".")[-1] == "uroman"
if is_uroman:
logger.warning("For this checkpoint, you should use `uroman` to convert input text before tokenizing it!")
else:
logger.info(f"***Converting model: {checkpoint_path}***")
is_uroman = False
# original VITS checkpoint
if vocab_path is None:
_pad = "_"
_punctuation = ';:,.!?¡¿—…"«»“” '
_letters = "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz"
_letters_ipa = "ɑɐɒæɓʙβɔɕçɗɖðʤəɘɚɛɜɝɞɟʄɡɠɢʛɦɧħɥʜɨɪʝɭɬɫɮʟɱɯɰŋɳɲɴøɵɸθœɶʘɹɺɾɻʀʁɽʂʃʈʧʉʊʋⱱʌɣɤʍχʎʏʑʐʒʔʡʕʢǀǁǂǃˈˌːˑʼʴʰʱʲʷˠˤ˞↓↑→↗↘'̩'ᵻ"
symbols = _pad + _punctuation + _letters + _letters_ipa
symbol_to_id = {s: i for i, s in enumerate(symbols)}
phonemize = True
else:
# Save vocab as temporary json file
symbols = [line.replace("\n", "") for line in open(vocab_path, encoding="utf-8").readlines()]
symbol_to_id = {s: i for i, s in enumerate(symbols)}
# MMS-TTS does not use a <pad> token, so we set to the token used to space characters
_pad = symbols[0]
phonemize = False
with tempfile.NamedTemporaryFile() as tf:
with open(tf.name, "w", encoding="utf-8") as f:
f.write(json.dumps(symbol_to_id, indent=2, sort_keys=True, ensure_ascii=False) + "\n")
tokenizer = VitsTokenizer(tf.name, language=language, phonemize=phonemize, is_uroman=is_uroman, pad_token=_pad)
config.vocab_size = len(symbols)
model = VitsModel(config)
model.decoder.apply_weight_norm()
orig_checkpoint = torch.load(checkpoint_path, map_location=torch.device("cpu"))
recursively_load_weights(orig_checkpoint["model"], model)
model.decoder.remove_weight_norm()
model.save_pretrained(pytorch_dump_folder_path)
tokenizer.save_pretrained(pytorch_dump_folder_path)
if repo_id:
print("Pushing to the hub...")
tokenizer.push_to_hub(repo_id)
model.push_to_hub(repo_id)
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--checkpoint_path", default=None, type=str, help="Local path to original checkpoint")
parser.add_argument("--vocab_path", default=None, type=str, help="Path to vocab.txt")
parser.add_argument("--config_path", default=None, type=str, help="Path to hf config.json of model to convert")
parser.add_argument("--language", default=None, type=str, help="Tokenizer language (three-letter code)")
parser.add_argument("--num_speakers", default=None, type=int, help="Number of speakers")
parser.add_argument(
"--sampling_rate", default=None, type=int, help="Sampling rate on which the model was trained."
)
parser.add_argument(
"--pytorch_dump_folder_path", required=True, default=None, type=str, help="Path to the output PyTorch model."
)
parser.add_argument(
"--push_to_hub", default=None, type=str, help="Where to upload the converted model on the 🤗 hub."
)
args = parser.parse_args()
convert_checkpoint(
args.pytorch_dump_folder_path,
args.checkpoint_path,
args.config_path,
args.vocab_path,
args.language,
args.num_speakers,
args.sampling_rate,
args.push_to_hub,
)
|
transformers/src/transformers/models/vits/convert_original_checkpoint.py/0
|
{
"file_path": "transformers/src/transformers/models/vits/convert_original_checkpoint.py",
"repo_id": "transformers",
"token_count": 8722
}
| 389
|
# coding=utf-8
# Copyright 2021 The HuggingFace Inc. team.
#
# 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.
"""
Speech processor class for Wav2Vec2
"""
import warnings
from contextlib import contextmanager
from ...processing_utils import ProcessorMixin
from .feature_extraction_wav2vec2 import Wav2Vec2FeatureExtractor
from .tokenization_wav2vec2 import Wav2Vec2CTCTokenizer
class Wav2Vec2Processor(ProcessorMixin):
r"""
Constructs a Wav2Vec2 processor which wraps a Wav2Vec2 feature extractor and a Wav2Vec2 CTC tokenizer into a single
processor.
[`Wav2Vec2Processor`] offers all the functionalities of [`Wav2Vec2FeatureExtractor`] and [`PreTrainedTokenizer`].
See the docstring of [`~Wav2Vec2Processor.__call__`] and [`~Wav2Vec2Processor.decode`] for more information.
Args:
feature_extractor (`Wav2Vec2FeatureExtractor`):
An instance of [`Wav2Vec2FeatureExtractor`]. The feature extractor is a required input.
tokenizer ([`PreTrainedTokenizer`]):
An instance of [`PreTrainedTokenizer`]. The tokenizer is a required input.
"""
feature_extractor_class = "Wav2Vec2FeatureExtractor"
tokenizer_class = "AutoTokenizer"
def __init__(self, feature_extractor, tokenizer):
super().__init__(feature_extractor, tokenizer)
self.current_processor = self.feature_extractor
self._in_target_context_manager = False
@classmethod
def from_pretrained(cls, pretrained_model_name_or_path, **kwargs):
try:
return super().from_pretrained(pretrained_model_name_or_path, **kwargs)
except OSError:
warnings.warn(
f"Loading a tokenizer inside {cls.__name__} from a config that does not"
" include a `tokenizer_class` attribute is deprecated and will be "
"removed in v5. Please add `'tokenizer_class': 'Wav2Vec2CTCTokenizer'`"
" attribute to either your `config.json` or `tokenizer_config.json` "
"file to suppress this warning: ",
FutureWarning,
)
feature_extractor = Wav2Vec2FeatureExtractor.from_pretrained(pretrained_model_name_or_path, **kwargs)
tokenizer = Wav2Vec2CTCTokenizer.from_pretrained(pretrained_model_name_or_path, **kwargs)
return cls(feature_extractor=feature_extractor, tokenizer=tokenizer)
def __call__(self, *args, **kwargs):
"""
When used in normal mode, this method forwards all its arguments to Wav2Vec2FeatureExtractor's
[`~Wav2Vec2FeatureExtractor.__call__`] and returns its output. If used in the context
[`~Wav2Vec2Processor.as_target_processor`] this method forwards all its arguments to PreTrainedTokenizer's
[`~PreTrainedTokenizer.__call__`]. Please refer to the docstring of the above two methods for more information.
"""
# For backward compatibility
if self._in_target_context_manager:
return self.current_processor(*args, **kwargs)
if "raw_speech" in kwargs:
warnings.warn("Using `raw_speech` as a keyword argument is deprecated. Use `audio` instead.")
audio = kwargs.pop("raw_speech")
else:
audio = kwargs.pop("audio", None)
sampling_rate = kwargs.pop("sampling_rate", None)
text = kwargs.pop("text", None)
if len(args) > 0:
audio = args[0]
args = args[1:]
if audio is None and text is None:
raise ValueError("You need to specify either an `audio` or `text` input to process.")
if audio is not None:
inputs = self.feature_extractor(audio, *args, sampling_rate=sampling_rate, **kwargs)
if text is not None:
encodings = self.tokenizer(text, **kwargs)
if text is None:
return inputs
elif audio is None:
return encodings
else:
inputs["labels"] = encodings["input_ids"]
return inputs
def pad(self, *args, **kwargs):
"""
When used in normal mode, this method forwards all its arguments to Wav2Vec2FeatureExtractor's
[`~Wav2Vec2FeatureExtractor.pad`] and returns its output. If used in the context
[`~Wav2Vec2Processor.as_target_processor`] this method forwards all its arguments to PreTrainedTokenizer's
[`~PreTrainedTokenizer.pad`]. Please refer to the docstring of the above two methods for more information.
"""
# For backward compatibility
if self._in_target_context_manager:
return self.current_processor.pad(*args, **kwargs)
input_features = kwargs.pop("input_features", None)
labels = kwargs.pop("labels", None)
if len(args) > 0:
input_features = args[0]
args = args[1:]
if input_features is not None:
input_features = self.feature_extractor.pad(input_features, *args, **kwargs)
if labels is not None:
labels = self.tokenizer.pad(labels, **kwargs)
if labels is None:
return input_features
elif input_features is None:
return labels
else:
input_features["labels"] = labels["input_ids"]
return input_features
def batch_decode(self, *args, **kwargs):
"""
This method forwards all its arguments to PreTrainedTokenizer's [`~PreTrainedTokenizer.batch_decode`]. Please
refer to the docstring of this method for more information.
"""
return self.tokenizer.batch_decode(*args, **kwargs)
def decode(self, *args, **kwargs):
"""
This method forwards all its arguments to PreTrainedTokenizer's [`~PreTrainedTokenizer.decode`]. Please refer
to the docstring of this method for more information.
"""
return self.tokenizer.decode(*args, **kwargs)
@contextmanager
def as_target_processor(self):
"""
Temporarily sets the tokenizer for processing the input. Useful for encoding the labels when fine-tuning
Wav2Vec2.
"""
warnings.warn(
"`as_target_processor` is deprecated and will be removed in v5 of Transformers. You can process your "
"labels by using the argument `text` of the regular `__call__` method (either in the same call as "
"your audio inputs, or in a separate call."
)
self._in_target_context_manager = True
self.current_processor = self.tokenizer
yield
self.current_processor = self.feature_extractor
self._in_target_context_manager = False
|
transformers/src/transformers/models/wav2vec2/processing_wav2vec2.py/0
|
{
"file_path": "transformers/src/transformers/models/wav2vec2/processing_wav2vec2.py",
"repo_id": "transformers",
"token_count": 2840
}
| 390
|
# coding=utf-8
# Copyright 2018 The HuggingFace Inc. team.
#
# 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.
"""Convert BERT checkpoint."""
import argparse
import os
import torch
from transformers import (
XLNetConfig,
XLNetForQuestionAnswering,
XLNetForSequenceClassification,
XLNetLMHeadModel,
load_tf_weights_in_xlnet,
)
from transformers.utils import CONFIG_NAME, WEIGHTS_NAME, logging
GLUE_TASKS_NUM_LABELS = {
"cola": 2,
"mnli": 3,
"mrpc": 2,
"sst-2": 2,
"sts-b": 1,
"qqp": 2,
"qnli": 2,
"rte": 2,
"wnli": 2,
}
logging.set_verbosity_info()
def convert_xlnet_checkpoint_to_pytorch(
tf_checkpoint_path, bert_config_file, pytorch_dump_folder_path, finetuning_task=None
):
# Initialise PyTorch model
config = XLNetConfig.from_json_file(bert_config_file)
finetuning_task = finetuning_task.lower() if finetuning_task is not None else ""
if finetuning_task in GLUE_TASKS_NUM_LABELS:
print(f"Building PyTorch XLNetForSequenceClassification model from configuration: {config}")
config.finetuning_task = finetuning_task
config.num_labels = GLUE_TASKS_NUM_LABELS[finetuning_task]
model = XLNetForSequenceClassification(config)
elif "squad" in finetuning_task:
config.finetuning_task = finetuning_task
model = XLNetForQuestionAnswering(config)
else:
model = XLNetLMHeadModel(config)
# Load weights from tf checkpoint
load_tf_weights_in_xlnet(model, config, tf_checkpoint_path)
# Save pytorch-model
pytorch_weights_dump_path = os.path.join(pytorch_dump_folder_path, WEIGHTS_NAME)
pytorch_config_dump_path = os.path.join(pytorch_dump_folder_path, CONFIG_NAME)
print(f"Save PyTorch model to {os.path.abspath(pytorch_weights_dump_path)}")
torch.save(model.state_dict(), pytorch_weights_dump_path)
print(f"Save configuration file to {os.path.abspath(pytorch_config_dump_path)}")
with open(pytorch_config_dump_path, "w", encoding="utf-8") as f:
f.write(config.to_json_string())
if __name__ == "__main__":
parser = argparse.ArgumentParser()
# Required parameters
parser.add_argument(
"--tf_checkpoint_path", default=None, type=str, required=True, help="Path to the TensorFlow checkpoint path."
)
parser.add_argument(
"--xlnet_config_file",
default=None,
type=str,
required=True,
help=(
"The config json file corresponding to the pre-trained XLNet model. \n"
"This specifies the model architecture."
),
)
parser.add_argument(
"--pytorch_dump_folder_path",
default=None,
type=str,
required=True,
help="Path to the folder to store the PyTorch model or dataset/vocab.",
)
parser.add_argument(
"--finetuning_task",
default=None,
type=str,
help="Name of a task on which the XLNet TensorFlow model was fine-tuned",
)
args = parser.parse_args()
print(args)
convert_xlnet_checkpoint_to_pytorch(
args.tf_checkpoint_path, args.xlnet_config_file, args.pytorch_dump_folder_path, args.finetuning_task
)
|
transformers/src/transformers/models/xlnet/convert_xlnet_original_tf_checkpoint_to_pytorch.py/0
|
{
"file_path": "transformers/src/transformers/models/xlnet/convert_xlnet_original_tf_checkpoint_to_pytorch.py",
"repo_id": "transformers",
"token_count": 1468
}
| 391
|
# coding=utf-8
# Copyright 2022 The HuggingFace Inc. team. 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.
"""YOSO model configuration"""
from ...configuration_utils import PretrainedConfig
from ...utils import logging
logger = logging.get_logger(__name__)
class YosoConfig(PretrainedConfig):
r"""
This is the configuration class to store the configuration of a [`YosoModel`]. It is used to instantiate an YOSO
model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
defaults will yield a similar configuration to that of the YOSO
[uw-madison/yoso-4096](https://huggingface.co/uw-madison/yoso-4096) architecture.
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PretrainedConfig`] for more information.
Args:
vocab_size (`int`, *optional*, defaults to 50265):
Vocabulary size of the YOSO model. Defines the number of different tokens that can be represented by the
`inputs_ids` passed when calling [`YosoModel`].
hidden_size (`int`, *optional*, defaults to 768):
Dimension of the encoder layers and the pooler layer.
num_hidden_layers (`int`, *optional*, defaults to 12):
Number of hidden layers in the Transformer encoder.
num_attention_heads (`int`, *optional*, defaults to 12):
Number of attention heads for each attention layer in the Transformer encoder.
intermediate_size (`int`, *optional*, defaults to 3072):
Dimension of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
hidden_act (`str` or `function`, *optional*, defaults to `"gelu"`):
The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
`"relu"`, `"selu"` and `"gelu_new"` are supported.
hidden_dropout_prob (`float`, *optional*, defaults to 0.1):
The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
attention_probs_dropout_prob (`float`, *optional*, defaults to 0.1):
The dropout ratio for the attention probabilities.
max_position_embeddings (`int`, *optional*, defaults to 512):
The maximum sequence length that this model might ever be used with. Typically set this to something large
just in case (e.g., 512 or 1024 or 2048).
type_vocab_size (`int`, *optional*, defaults to 2):
The vocabulary size of the `token_type_ids` passed when calling [`YosoModel`].
initializer_range (`float`, *optional*, defaults to 0.02):
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
layer_norm_eps (`float`, *optional*, defaults to 1e-12):
The epsilon used by the layer normalization layers.
position_embedding_type (`str`, *optional*, defaults to `"absolute"`):
Type of position embedding. Choose one of `"absolute"`, `"relative_key"`, `"relative_key_query"`.
use_expectation (`bool`, *optional*, defaults to `True`):
Whether or not to use YOSO Expectation. Overrides any effect of num_hash.
hash_code_len (`int`, *optional*, defaults to 9):
The length of hashes generated by the hash functions.
num_hash (`int`, *optional*, defaults to 64):
Number of hash functions used in [`YosoSelfAttention`].
conv_window (`int`, *optional*):
Kernel size of depth-wise convolution.
use_fast_hash (`bool`, *optional*, defaults to `False`):
Whether or not to use custom cuda kernels which perform fast random projection via hadamard transform.
lsh_backward (`bool`, *optional*, defaults to `True`):
Whether or not to perform backpropagation using Locality Sensitive Hashing.
Example:
```python
>>> from transformers import YosoConfig, YosoModel
>>> # Initializing a YOSO uw-madison/yoso-4096 style configuration
>>> configuration = YosoConfig()
>>> # Initializing a model (with random weights) from the uw-madison/yoso-4096 style configuration
>>> model = YosoModel(configuration)
>>> # Accessing the model configuration
>>> configuration = model.config
```"""
model_type = "yoso"
def __init__(
self,
vocab_size=50265,
hidden_size=768,
num_hidden_layers=12,
num_attention_heads=12,
intermediate_size=3072,
hidden_act="gelu",
hidden_dropout_prob=0.1,
attention_probs_dropout_prob=0.1,
max_position_embeddings=4096,
type_vocab_size=1,
initializer_range=0.02,
layer_norm_eps=1e-12,
position_embedding_type="absolute",
use_expectation=True,
hash_code_len=9,
num_hash=64,
conv_window=None,
use_fast_hash=True,
lsh_backward=True,
pad_token_id=1,
bos_token_id=0,
eos_token_id=2,
**kwargs,
):
super().__init__(pad_token_id=pad_token_id, bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)
self.vocab_size = vocab_size
self.max_position_embeddings = max_position_embeddings
self.hidden_size = hidden_size
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
self.intermediate_size = intermediate_size
self.hidden_act = hidden_act
self.hidden_dropout_prob = hidden_dropout_prob
self.attention_probs_dropout_prob = attention_probs_dropout_prob
self.initializer_range = initializer_range
self.type_vocab_size = type_vocab_size
self.layer_norm_eps = layer_norm_eps
self.position_embedding_type = position_embedding_type
self.use_expectation = use_expectation
self.hash_code_len = hash_code_len
self.num_hash = num_hash
self.conv_window = conv_window
self.use_fast_hash = use_fast_hash
self.lsh_backward = lsh_backward
|
transformers/src/transformers/models/yoso/configuration_yoso.py/0
|
{
"file_path": "transformers/src/transformers/models/yoso/configuration_yoso.py",
"repo_id": "transformers",
"token_count": 2543
}
| 392
|
# coding=utf-8
# Copyright 2018 The HuggingFace Inc. team.
#
# 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.
import json
import os
import warnings
from pathlib import Path
from typing import TYPE_CHECKING, Any, Dict, List, Optional, Tuple, Union
from huggingface_hub import model_info
from ..configuration_utils import PretrainedConfig
from ..dynamic_module_utils import get_class_from_dynamic_module
from ..feature_extraction_utils import PreTrainedFeatureExtractor
from ..image_processing_utils import BaseImageProcessor
from ..models.auto.configuration_auto import AutoConfig
from ..models.auto.feature_extraction_auto import FEATURE_EXTRACTOR_MAPPING, AutoFeatureExtractor
from ..models.auto.image_processing_auto import IMAGE_PROCESSOR_MAPPING, AutoImageProcessor
from ..models.auto.modeling_auto import AutoModelForDepthEstimation, AutoModelForImageToImage
from ..models.auto.tokenization_auto import TOKENIZER_MAPPING, AutoTokenizer
from ..tokenization_utils import PreTrainedTokenizer
from ..utils import (
CONFIG_NAME,
HUGGINGFACE_CO_RESOLVE_ENDPOINT,
cached_file,
extract_commit_hash,
find_adapter_config_file,
is_kenlm_available,
is_offline_mode,
is_peft_available,
is_pyctcdecode_available,
is_tf_available,
is_torch_available,
logging,
)
from .audio_classification import AudioClassificationPipeline
from .automatic_speech_recognition import AutomaticSpeechRecognitionPipeline
from .base import (
ArgumentHandler,
CsvPipelineDataFormat,
JsonPipelineDataFormat,
PipedPipelineDataFormat,
Pipeline,
PipelineDataFormat,
PipelineException,
PipelineRegistry,
get_default_model_and_revision,
infer_framework_load_model,
)
from .depth_estimation import DepthEstimationPipeline
from .document_question_answering import DocumentQuestionAnsweringPipeline
from .feature_extraction import FeatureExtractionPipeline
from .fill_mask import FillMaskPipeline
from .image_classification import ImageClassificationPipeline
from .image_feature_extraction import ImageFeatureExtractionPipeline
from .image_segmentation import ImageSegmentationPipeline
from .image_to_image import ImageToImagePipeline
from .image_to_text import ImageToTextPipeline
from .mask_generation import MaskGenerationPipeline
from .object_detection import ObjectDetectionPipeline
from .question_answering import QuestionAnsweringArgumentHandler, QuestionAnsweringPipeline
from .table_question_answering import TableQuestionAnsweringArgumentHandler, TableQuestionAnsweringPipeline
from .text2text_generation import SummarizationPipeline, Text2TextGenerationPipeline, TranslationPipeline
from .text_classification import TextClassificationPipeline
from .text_generation import TextGenerationPipeline
from .text_to_audio import TextToAudioPipeline
from .token_classification import (
AggregationStrategy,
NerPipeline,
TokenClassificationArgumentHandler,
TokenClassificationPipeline,
)
from .video_classification import VideoClassificationPipeline
from .visual_question_answering import VisualQuestionAnsweringPipeline
from .zero_shot_audio_classification import ZeroShotAudioClassificationPipeline
from .zero_shot_classification import ZeroShotClassificationArgumentHandler, ZeroShotClassificationPipeline
from .zero_shot_image_classification import ZeroShotImageClassificationPipeline
from .zero_shot_object_detection import ZeroShotObjectDetectionPipeline
if is_tf_available():
import tensorflow as tf
from ..models.auto.modeling_tf_auto import (
TFAutoModel,
TFAutoModelForCausalLM,
TFAutoModelForImageClassification,
TFAutoModelForMaskedLM,
TFAutoModelForQuestionAnswering,
TFAutoModelForSeq2SeqLM,
TFAutoModelForSequenceClassification,
TFAutoModelForTableQuestionAnswering,
TFAutoModelForTokenClassification,
TFAutoModelForVision2Seq,
TFAutoModelForZeroShotImageClassification,
)
if is_torch_available():
import torch
from ..models.auto.modeling_auto import (
AutoModel,
AutoModelForAudioClassification,
AutoModelForCausalLM,
AutoModelForCTC,
AutoModelForDocumentQuestionAnswering,
AutoModelForImageClassification,
AutoModelForImageSegmentation,
AutoModelForMaskedLM,
AutoModelForMaskGeneration,
AutoModelForObjectDetection,
AutoModelForQuestionAnswering,
AutoModelForSemanticSegmentation,
AutoModelForSeq2SeqLM,
AutoModelForSequenceClassification,
AutoModelForSpeechSeq2Seq,
AutoModelForTableQuestionAnswering,
AutoModelForTextToSpectrogram,
AutoModelForTextToWaveform,
AutoModelForTokenClassification,
AutoModelForVideoClassification,
AutoModelForVision2Seq,
AutoModelForVisualQuestionAnswering,
AutoModelForZeroShotImageClassification,
AutoModelForZeroShotObjectDetection,
)
if TYPE_CHECKING:
from ..modeling_tf_utils import TFPreTrainedModel
from ..modeling_utils import PreTrainedModel
from ..tokenization_utils_fast import PreTrainedTokenizerFast
logger = logging.get_logger(__name__)
# Register all the supported tasks here
TASK_ALIASES = {
"sentiment-analysis": "text-classification",
"ner": "token-classification",
"vqa": "visual-question-answering",
"text-to-speech": "text-to-audio",
}
SUPPORTED_TASKS = {
"audio-classification": {
"impl": AudioClassificationPipeline,
"tf": (),
"pt": (AutoModelForAudioClassification,) if is_torch_available() else (),
"default": {"model": {"pt": ("superb/wav2vec2-base-superb-ks", "372e048")}},
"type": "audio",
},
"automatic-speech-recognition": {
"impl": AutomaticSpeechRecognitionPipeline,
"tf": (),
"pt": (AutoModelForCTC, AutoModelForSpeechSeq2Seq) if is_torch_available() else (),
"default": {"model": {"pt": ("facebook/wav2vec2-base-960h", "55bb623")}},
"type": "multimodal",
},
"text-to-audio": {
"impl": TextToAudioPipeline,
"tf": (),
"pt": (AutoModelForTextToWaveform, AutoModelForTextToSpectrogram) if is_torch_available() else (),
"default": {"model": {"pt": ("suno/bark-small", "645cfba")}},
"type": "text",
},
"feature-extraction": {
"impl": FeatureExtractionPipeline,
"tf": (TFAutoModel,) if is_tf_available() else (),
"pt": (AutoModel,) if is_torch_available() else (),
"default": {
"model": {
"pt": ("distilbert/distilbert-base-cased", "935ac13"),
"tf": ("distilbert/distilbert-base-cased", "935ac13"),
}
},
"type": "multimodal",
},
"text-classification": {
"impl": TextClassificationPipeline,
"tf": (TFAutoModelForSequenceClassification,) if is_tf_available() else (),
"pt": (AutoModelForSequenceClassification,) if is_torch_available() else (),
"default": {
"model": {
"pt": ("distilbert/distilbert-base-uncased-finetuned-sst-2-english", "af0f99b"),
"tf": ("distilbert/distilbert-base-uncased-finetuned-sst-2-english", "af0f99b"),
},
},
"type": "text",
},
"token-classification": {
"impl": TokenClassificationPipeline,
"tf": (TFAutoModelForTokenClassification,) if is_tf_available() else (),
"pt": (AutoModelForTokenClassification,) if is_torch_available() else (),
"default": {
"model": {
"pt": ("dbmdz/bert-large-cased-finetuned-conll03-english", "f2482bf"),
"tf": ("dbmdz/bert-large-cased-finetuned-conll03-english", "f2482bf"),
},
},
"type": "text",
},
"question-answering": {
"impl": QuestionAnsweringPipeline,
"tf": (TFAutoModelForQuestionAnswering,) if is_tf_available() else (),
"pt": (AutoModelForQuestionAnswering,) if is_torch_available() else (),
"default": {
"model": {
"pt": ("distilbert/distilbert-base-cased-distilled-squad", "626af31"),
"tf": ("distilbert/distilbert-base-cased-distilled-squad", "626af31"),
},
},
"type": "text",
},
"table-question-answering": {
"impl": TableQuestionAnsweringPipeline,
"pt": (AutoModelForTableQuestionAnswering,) if is_torch_available() else (),
"tf": (TFAutoModelForTableQuestionAnswering,) if is_tf_available() else (),
"default": {
"model": {
"pt": ("google/tapas-base-finetuned-wtq", "69ceee2"),
"tf": ("google/tapas-base-finetuned-wtq", "69ceee2"),
},
},
"type": "text",
},
"visual-question-answering": {
"impl": VisualQuestionAnsweringPipeline,
"pt": (AutoModelForVisualQuestionAnswering,) if is_torch_available() else (),
"tf": (),
"default": {
"model": {"pt": ("dandelin/vilt-b32-finetuned-vqa", "4355f59")},
},
"type": "multimodal",
},
"document-question-answering": {
"impl": DocumentQuestionAnsweringPipeline,
"pt": (AutoModelForDocumentQuestionAnswering,) if is_torch_available() else (),
"tf": (),
"default": {
"model": {"pt": ("impira/layoutlm-document-qa", "52e01b3")},
},
"type": "multimodal",
},
"fill-mask": {
"impl": FillMaskPipeline,
"tf": (TFAutoModelForMaskedLM,) if is_tf_available() else (),
"pt": (AutoModelForMaskedLM,) if is_torch_available() else (),
"default": {
"model": {
"pt": ("distilbert/distilroberta-base", "ec58a5b"),
"tf": ("distilbert/distilroberta-base", "ec58a5b"),
}
},
"type": "text",
},
"summarization": {
"impl": SummarizationPipeline,
"tf": (TFAutoModelForSeq2SeqLM,) if is_tf_available() else (),
"pt": (AutoModelForSeq2SeqLM,) if is_torch_available() else (),
"default": {
"model": {"pt": ("sshleifer/distilbart-cnn-12-6", "a4f8f3e"), "tf": ("google-t5/t5-small", "d769bba")}
},
"type": "text",
},
# This task is a special case as it's parametrized by SRC, TGT languages.
"translation": {
"impl": TranslationPipeline,
"tf": (TFAutoModelForSeq2SeqLM,) if is_tf_available() else (),
"pt": (AutoModelForSeq2SeqLM,) if is_torch_available() else (),
"default": {
("en", "fr"): {"model": {"pt": ("google-t5/t5-base", "686f1db"), "tf": ("google-t5/t5-base", "686f1db")}},
("en", "de"): {"model": {"pt": ("google-t5/t5-base", "686f1db"), "tf": ("google-t5/t5-base", "686f1db")}},
("en", "ro"): {"model": {"pt": ("google-t5/t5-base", "686f1db"), "tf": ("google-t5/t5-base", "686f1db")}},
},
"type": "text",
},
"text2text-generation": {
"impl": Text2TextGenerationPipeline,
"tf": (TFAutoModelForSeq2SeqLM,) if is_tf_available() else (),
"pt": (AutoModelForSeq2SeqLM,) if is_torch_available() else (),
"default": {"model": {"pt": ("google-t5/t5-base", "686f1db"), "tf": ("google-t5/t5-base", "686f1db")}},
"type": "text",
},
"text-generation": {
"impl": TextGenerationPipeline,
"tf": (TFAutoModelForCausalLM,) if is_tf_available() else (),
"pt": (AutoModelForCausalLM,) if is_torch_available() else (),
"default": {"model": {"pt": ("openai-community/gpt2", "6c0e608"), "tf": ("openai-community/gpt2", "6c0e608")}},
"type": "text",
},
"zero-shot-classification": {
"impl": ZeroShotClassificationPipeline,
"tf": (TFAutoModelForSequenceClassification,) if is_tf_available() else (),
"pt": (AutoModelForSequenceClassification,) if is_torch_available() else (),
"default": {
"model": {
"pt": ("facebook/bart-large-mnli", "c626438"),
"tf": ("FacebookAI/roberta-large-mnli", "130fb28"),
},
"config": {
"pt": ("facebook/bart-large-mnli", "c626438"),
"tf": ("FacebookAI/roberta-large-mnli", "130fb28"),
},
},
"type": "text",
},
"zero-shot-image-classification": {
"impl": ZeroShotImageClassificationPipeline,
"tf": (TFAutoModelForZeroShotImageClassification,) if is_tf_available() else (),
"pt": (AutoModelForZeroShotImageClassification,) if is_torch_available() else (),
"default": {
"model": {
"pt": ("openai/clip-vit-base-patch32", "f4881ba"),
"tf": ("openai/clip-vit-base-patch32", "f4881ba"),
}
},
"type": "multimodal",
},
"zero-shot-audio-classification": {
"impl": ZeroShotAudioClassificationPipeline,
"tf": (),
"pt": (AutoModel,) if is_torch_available() else (),
"default": {
"model": {
"pt": ("laion/clap-htsat-fused", "973b6e5"),
}
},
"type": "multimodal",
},
"image-classification": {
"impl": ImageClassificationPipeline,
"tf": (TFAutoModelForImageClassification,) if is_tf_available() else (),
"pt": (AutoModelForImageClassification,) if is_torch_available() else (),
"default": {
"model": {
"pt": ("google/vit-base-patch16-224", "5dca96d"),
"tf": ("google/vit-base-patch16-224", "5dca96d"),
}
},
"type": "image",
},
"image-feature-extraction": {
"impl": ImageFeatureExtractionPipeline,
"tf": (TFAutoModel,) if is_tf_available() else (),
"pt": (AutoModel,) if is_torch_available() else (),
"default": {
"model": {
"pt": ("google/vit-base-patch16-224", "3f49326"),
"tf": ("google/vit-base-patch16-224", "3f49326"),
}
},
"type": "image",
},
"image-segmentation": {
"impl": ImageSegmentationPipeline,
"tf": (),
"pt": (AutoModelForImageSegmentation, AutoModelForSemanticSegmentation) if is_torch_available() else (),
"default": {"model": {"pt": ("facebook/detr-resnet-50-panoptic", "fc15262")}},
"type": "multimodal",
},
"image-to-text": {
"impl": ImageToTextPipeline,
"tf": (TFAutoModelForVision2Seq,) if is_tf_available() else (),
"pt": (AutoModelForVision2Seq,) if is_torch_available() else (),
"default": {
"model": {
"pt": ("ydshieh/vit-gpt2-coco-en", "65636df"),
"tf": ("ydshieh/vit-gpt2-coco-en", "65636df"),
}
},
"type": "multimodal",
},
"object-detection": {
"impl": ObjectDetectionPipeline,
"tf": (),
"pt": (AutoModelForObjectDetection,) if is_torch_available() else (),
"default": {"model": {"pt": ("facebook/detr-resnet-50", "2729413")}},
"type": "multimodal",
},
"zero-shot-object-detection": {
"impl": ZeroShotObjectDetectionPipeline,
"tf": (),
"pt": (AutoModelForZeroShotObjectDetection,) if is_torch_available() else (),
"default": {"model": {"pt": ("google/owlvit-base-patch32", "17740e1")}},
"type": "multimodal",
},
"depth-estimation": {
"impl": DepthEstimationPipeline,
"tf": (),
"pt": (AutoModelForDepthEstimation,) if is_torch_available() else (),
"default": {"model": {"pt": ("Intel/dpt-large", "e93beec")}},
"type": "image",
},
"video-classification": {
"impl": VideoClassificationPipeline,
"tf": (),
"pt": (AutoModelForVideoClassification,) if is_torch_available() else (),
"default": {"model": {"pt": ("MCG-NJU/videomae-base-finetuned-kinetics", "4800870")}},
"type": "video",
},
"mask-generation": {
"impl": MaskGenerationPipeline,
"tf": (),
"pt": (AutoModelForMaskGeneration,) if is_torch_available() else (),
"default": {"model": {"pt": ("facebook/sam-vit-huge", "997b15")}},
"type": "multimodal",
},
"image-to-image": {
"impl": ImageToImagePipeline,
"tf": (),
"pt": (AutoModelForImageToImage,) if is_torch_available() else (),
"default": {"model": {"pt": ("caidas/swin2SR-classical-sr-x2-64", "4aaedcb")}},
"type": "image",
},
}
NO_FEATURE_EXTRACTOR_TASKS = set()
NO_IMAGE_PROCESSOR_TASKS = set()
NO_TOKENIZER_TASKS = set()
# Those model configs are special, they are generic over their task, meaning
# any tokenizer/feature_extractor might be use for a given model so we cannot
# use the statically defined TOKENIZER_MAPPING and FEATURE_EXTRACTOR_MAPPING to
# see if the model defines such objects or not.
MULTI_MODEL_AUDIO_CONFIGS = {"SpeechEncoderDecoderConfig"}
MULTI_MODEL_VISION_CONFIGS = {"VisionEncoderDecoderConfig", "VisionTextDualEncoderConfig"}
for task, values in SUPPORTED_TASKS.items():
if values["type"] == "text":
NO_FEATURE_EXTRACTOR_TASKS.add(task)
NO_IMAGE_PROCESSOR_TASKS.add(task)
elif values["type"] in {"image", "video"}:
NO_TOKENIZER_TASKS.add(task)
elif values["type"] in {"audio"}:
NO_TOKENIZER_TASKS.add(task)
NO_IMAGE_PROCESSOR_TASKS.add(task)
elif values["type"] != "multimodal":
raise ValueError(f"SUPPORTED_TASK {task} contains invalid type {values['type']}")
PIPELINE_REGISTRY = PipelineRegistry(supported_tasks=SUPPORTED_TASKS, task_aliases=TASK_ALIASES)
def get_supported_tasks() -> List[str]:
"""
Returns a list of supported task strings.
"""
return PIPELINE_REGISTRY.get_supported_tasks()
def get_task(model: str, token: Optional[str] = None, **deprecated_kwargs) -> str:
use_auth_token = deprecated_kwargs.pop("use_auth_token", None)
if use_auth_token is not None:
warnings.warn(
"The `use_auth_token` argument is deprecated and will be removed in v5 of Transformers. Please use `token` instead.",
FutureWarning,
)
if token is not None:
raise ValueError("`token` and `use_auth_token` are both specified. Please set only the argument `token`.")
token = use_auth_token
if is_offline_mode():
raise RuntimeError("You cannot infer task automatically within `pipeline` when using offline mode")
try:
info = model_info(model, token=token)
except Exception as e:
raise RuntimeError(f"Instantiating a pipeline without a task set raised an error: {e}")
if not info.pipeline_tag:
raise RuntimeError(
f"The model {model} does not seem to have a correct `pipeline_tag` set to infer the task automatically"
)
if getattr(info, "library_name", "transformers") != "transformers":
raise RuntimeError(f"This model is meant to be used with {info.library_name} not with transformers")
task = info.pipeline_tag
return task
def check_task(task: str) -> Tuple[str, Dict, Any]:
"""
Checks an incoming task string, to validate it's correct and return the default Pipeline and Model classes, and
default models if they exist.
Args:
task (`str`):
The task defining which pipeline will be returned. Currently accepted tasks are:
- `"audio-classification"`
- `"automatic-speech-recognition"`
- `"conversational"`
- `"depth-estimation"`
- `"document-question-answering"`
- `"feature-extraction"`
- `"fill-mask"`
- `"image-classification"`
- `"image-feature-extraction"`
- `"image-segmentation"`
- `"image-to-text"`
- `"image-to-image"`
- `"object-detection"`
- `"question-answering"`
- `"summarization"`
- `"table-question-answering"`
- `"text2text-generation"`
- `"text-classification"` (alias `"sentiment-analysis"` available)
- `"text-generation"`
- `"text-to-audio"` (alias `"text-to-speech"` available)
- `"token-classification"` (alias `"ner"` available)
- `"translation"`
- `"translation_xx_to_yy"`
- `"video-classification"`
- `"visual-question-answering"` (alias `"vqa"` available)
- `"zero-shot-classification"`
- `"zero-shot-image-classification"`
- `"zero-shot-object-detection"`
Returns:
(normalized_task: `str`, task_defaults: `dict`, task_options: (`tuple`, None)) The normalized task name
(removed alias and options). The actual dictionary required to initialize the pipeline and some extra task
options for parametrized tasks like "translation_XX_to_YY"
"""
return PIPELINE_REGISTRY.check_task(task)
def clean_custom_task(task_info):
import transformers
if "impl" not in task_info:
raise RuntimeError("This model introduces a custom pipeline without specifying its implementation.")
pt_class_names = task_info.get("pt", ())
if isinstance(pt_class_names, str):
pt_class_names = [pt_class_names]
task_info["pt"] = tuple(getattr(transformers, c) for c in pt_class_names)
tf_class_names = task_info.get("tf", ())
if isinstance(tf_class_names, str):
tf_class_names = [tf_class_names]
task_info["tf"] = tuple(getattr(transformers, c) for c in tf_class_names)
return task_info, None
def pipeline(
task: str = None,
model: Optional[Union[str, "PreTrainedModel", "TFPreTrainedModel"]] = None,
config: Optional[Union[str, PretrainedConfig]] = None,
tokenizer: Optional[Union[str, PreTrainedTokenizer, "PreTrainedTokenizerFast"]] = None,
feature_extractor: Optional[Union[str, PreTrainedFeatureExtractor]] = None,
image_processor: Optional[Union[str, BaseImageProcessor]] = None,
framework: Optional[str] = None,
revision: Optional[str] = None,
use_fast: bool = True,
token: Optional[Union[str, bool]] = None,
device: Optional[Union[int, str, "torch.device"]] = None,
device_map=None,
torch_dtype=None,
trust_remote_code: Optional[bool] = None,
model_kwargs: Dict[str, Any] = None,
pipeline_class: Optional[Any] = None,
**kwargs,
) -> Pipeline:
"""
Utility factory method to build a [`Pipeline`].
Pipelines are made of:
- A [tokenizer](tokenizer) in charge of mapping raw textual input to token.
- A [model](model) to make predictions from the inputs.
- Some (optional) post processing for enhancing model's output.
Args:
task (`str`):
The task defining which pipeline will be returned. Currently accepted tasks are:
- `"audio-classification"`: will return a [`AudioClassificationPipeline`].
- `"automatic-speech-recognition"`: will return a [`AutomaticSpeechRecognitionPipeline`].
- `"depth-estimation"`: will return a [`DepthEstimationPipeline`].
- `"document-question-answering"`: will return a [`DocumentQuestionAnsweringPipeline`].
- `"feature-extraction"`: will return a [`FeatureExtractionPipeline`].
- `"fill-mask"`: will return a [`FillMaskPipeline`]:.
- `"image-classification"`: will return a [`ImageClassificationPipeline`].
- `"image-feature-extraction"`: will return an [`ImageFeatureExtractionPipeline`].
- `"image-segmentation"`: will return a [`ImageSegmentationPipeline`].
- `"image-to-image"`: will return a [`ImageToImagePipeline`].
- `"image-to-text"`: will return a [`ImageToTextPipeline`].
- `"mask-generation"`: will return a [`MaskGenerationPipeline`].
- `"object-detection"`: will return a [`ObjectDetectionPipeline`].
- `"question-answering"`: will return a [`QuestionAnsweringPipeline`].
- `"summarization"`: will return a [`SummarizationPipeline`].
- `"table-question-answering"`: will return a [`TableQuestionAnsweringPipeline`].
- `"text2text-generation"`: will return a [`Text2TextGenerationPipeline`].
- `"text-classification"` (alias `"sentiment-analysis"` available): will return a
[`TextClassificationPipeline`].
- `"text-generation"`: will return a [`TextGenerationPipeline`]:.
- `"text-to-audio"` (alias `"text-to-speech"` available): will return a [`TextToAudioPipeline`]:.
- `"token-classification"` (alias `"ner"` available): will return a [`TokenClassificationPipeline`].
- `"translation"`: will return a [`TranslationPipeline`].
- `"translation_xx_to_yy"`: will return a [`TranslationPipeline`].
- `"video-classification"`: will return a [`VideoClassificationPipeline`].
- `"visual-question-answering"`: will return a [`VisualQuestionAnsweringPipeline`].
- `"zero-shot-classification"`: will return a [`ZeroShotClassificationPipeline`].
- `"zero-shot-image-classification"`: will return a [`ZeroShotImageClassificationPipeline`].
- `"zero-shot-audio-classification"`: will return a [`ZeroShotAudioClassificationPipeline`].
- `"zero-shot-object-detection"`: will return a [`ZeroShotObjectDetectionPipeline`].
model (`str` or [`PreTrainedModel`] or [`TFPreTrainedModel`], *optional*):
The model that will be used by the pipeline to make predictions. This can be a model identifier or an
actual instance of a pretrained model inheriting from [`PreTrainedModel`] (for PyTorch) or
[`TFPreTrainedModel`] (for TensorFlow).
If not provided, the default for the `task` will be loaded.
config (`str` or [`PretrainedConfig`], *optional*):
The configuration that will be used by the pipeline to instantiate the model. This can be a model
identifier or an actual pretrained model configuration inheriting from [`PretrainedConfig`].
If not provided, the default configuration file for the requested model will be used. That means that if
`model` is given, its default configuration will be used. However, if `model` is not supplied, this
`task`'s default model's config is used instead.
tokenizer (`str` or [`PreTrainedTokenizer`], *optional*):
The tokenizer that will be used by the pipeline to encode data for the model. This can be a model
identifier or an actual pretrained tokenizer inheriting from [`PreTrainedTokenizer`].
If not provided, the default tokenizer for the given `model` will be loaded (if it is a string). If `model`
is not specified or not a string, then the default tokenizer for `config` is loaded (if it is a string).
However, if `config` is also not given or not a string, then the default tokenizer for the given `task`
will be loaded.
feature_extractor (`str` or [`PreTrainedFeatureExtractor`], *optional*):
The feature extractor that will be used by the pipeline to encode data for the model. This can be a model
identifier or an actual pretrained feature extractor inheriting from [`PreTrainedFeatureExtractor`].
Feature extractors are used for non-NLP models, such as Speech or Vision models as well as multi-modal
models. Multi-modal models will also require a tokenizer to be passed.
If not provided, the default feature extractor for the given `model` will be loaded (if it is a string). If
`model` is not specified or not a string, then the default feature extractor for `config` is loaded (if it
is a string). However, if `config` is also not given or not a string, then the default feature extractor
for the given `task` will be loaded.
framework (`str`, *optional*):
The framework to use, either `"pt"` for PyTorch or `"tf"` for TensorFlow. The specified framework must be
installed.
If no framework is specified, will default to the one currently installed. If no framework is specified and
both frameworks are installed, will default to the framework of the `model`, or to PyTorch if no model is
provided.
revision (`str`, *optional*, defaults to `"main"`):
When passing a task name or a string model identifier: The specific model version to use. It can be a
branch name, a tag name, or a commit id, since we use a git-based system for storing models and other
artifacts on huggingface.co, so `revision` can be any identifier allowed by git.
use_fast (`bool`, *optional*, defaults to `True`):
Whether or not to use a Fast tokenizer if possible (a [`PreTrainedTokenizerFast`]).
use_auth_token (`str` or *bool*, *optional*):
The token to use as HTTP bearer authorization for remote files. If `True`, will use the token generated
when running `huggingface-cli login` (stored in `~/.huggingface`).
device (`int` or `str` or `torch.device`):
Defines the device (*e.g.*, `"cpu"`, `"cuda:1"`, `"mps"`, or a GPU ordinal rank like `1`) on which this
pipeline will be allocated.
device_map (`str` or `Dict[str, Union[int, str, torch.device]`, *optional*):
Sent directly as `model_kwargs` (just a simpler shortcut). When `accelerate` library is present, set
`device_map="auto"` to compute the most optimized `device_map` automatically (see
[here](https://huggingface.co/docs/accelerate/main/en/package_reference/big_modeling#accelerate.cpu_offload)
for more information).
<Tip warning={true}>
Do not use `device_map` AND `device` at the same time as they will conflict
</Tip>
torch_dtype (`str` or `torch.dtype`, *optional*):
Sent directly as `model_kwargs` (just a simpler shortcut) to use the available precision for this model
(`torch.float16`, `torch.bfloat16`, ... or `"auto"`).
trust_remote_code (`bool`, *optional*, defaults to `False`):
Whether or not to allow for custom code defined on the Hub in their own modeling, configuration,
tokenization or even pipeline files. This option should only be set to `True` for repositories you trust
and in which you have read the code, as it will execute code present on the Hub on your local machine.
model_kwargs (`Dict[str, Any]`, *optional*):
Additional dictionary of keyword arguments passed along to the model's `from_pretrained(...,
**model_kwargs)` function.
kwargs (`Dict[str, Any]`, *optional*):
Additional keyword arguments passed along to the specific pipeline init (see the documentation for the
corresponding pipeline class for possible values).
Returns:
[`Pipeline`]: A suitable pipeline for the task.
Examples:
```python
>>> from transformers import pipeline, AutoModelForTokenClassification, AutoTokenizer
>>> # Sentiment analysis pipeline
>>> analyzer = pipeline("sentiment-analysis")
>>> # Question answering pipeline, specifying the checkpoint identifier
>>> oracle = pipeline(
... "question-answering", model="distilbert/distilbert-base-cased-distilled-squad", tokenizer="google-bert/bert-base-cased"
... )
>>> # Named entity recognition pipeline, passing in a specific model and tokenizer
>>> model = AutoModelForTokenClassification.from_pretrained("dbmdz/bert-large-cased-finetuned-conll03-english")
>>> tokenizer = AutoTokenizer.from_pretrained("google-bert/bert-base-cased")
>>> recognizer = pipeline("ner", model=model, tokenizer=tokenizer)
```"""
if model_kwargs is None:
model_kwargs = {}
# Make sure we only pass use_auth_token once as a kwarg (it used to be possible to pass it in model_kwargs,
# this is to keep BC).
use_auth_token = model_kwargs.pop("use_auth_token", None)
if use_auth_token is not None:
warnings.warn(
"The `use_auth_token` argument is deprecated and will be removed in v5 of Transformers. Please use `token` instead.",
FutureWarning,
)
if token is not None:
raise ValueError("`token` and `use_auth_token` are both specified. Please set only the argument `token`.")
token = use_auth_token
code_revision = kwargs.pop("code_revision", None)
commit_hash = kwargs.pop("_commit_hash", None)
hub_kwargs = {
"revision": revision,
"token": token,
"trust_remote_code": trust_remote_code,
"_commit_hash": commit_hash,
}
if task is None and model is None:
raise RuntimeError(
"Impossible to instantiate a pipeline without either a task or a model "
"being specified. "
"Please provide a task class or a model"
)
if model is None and tokenizer is not None:
raise RuntimeError(
"Impossible to instantiate a pipeline with tokenizer specified but not the model as the provided tokenizer"
" may not be compatible with the default model. Please provide a PreTrainedModel class or a"
" path/identifier to a pretrained model when providing tokenizer."
)
if model is None and feature_extractor is not None:
raise RuntimeError(
"Impossible to instantiate a pipeline with feature_extractor specified but not the model as the provided"
" feature_extractor may not be compatible with the default model. Please provide a PreTrainedModel class"
" or a path/identifier to a pretrained model when providing feature_extractor."
)
if isinstance(model, Path):
model = str(model)
if commit_hash is None:
pretrained_model_name_or_path = None
if isinstance(config, str):
pretrained_model_name_or_path = config
elif config is None and isinstance(model, str):
pretrained_model_name_or_path = model
if not isinstance(config, PretrainedConfig) and pretrained_model_name_or_path is not None:
# We make a call to the config file first (which may be absent) to get the commit hash as soon as possible
resolved_config_file = cached_file(
pretrained_model_name_or_path,
CONFIG_NAME,
_raise_exceptions_for_gated_repo=False,
_raise_exceptions_for_missing_entries=False,
_raise_exceptions_for_connection_errors=False,
cache_dir=model_kwargs.get("cache_dir"),
**hub_kwargs,
)
hub_kwargs["_commit_hash"] = extract_commit_hash(resolved_config_file, commit_hash)
else:
hub_kwargs["_commit_hash"] = getattr(config, "_commit_hash", None)
# Config is the primordial information item.
# Instantiate config if needed
if isinstance(config, str):
config = AutoConfig.from_pretrained(
config, _from_pipeline=task, code_revision=code_revision, **hub_kwargs, **model_kwargs
)
hub_kwargs["_commit_hash"] = config._commit_hash
elif config is None and isinstance(model, str):
# Check for an adapter file in the model path if PEFT is available
if is_peft_available():
# `find_adapter_config_file` doesn't accept `trust_remote_code`
_hub_kwargs = {k: v for k, v in hub_kwargs.items() if k != "trust_remote_code"}
maybe_adapter_path = find_adapter_config_file(
model,
token=hub_kwargs["token"],
revision=hub_kwargs["revision"],
_commit_hash=hub_kwargs["_commit_hash"],
)
if maybe_adapter_path is not None:
with open(maybe_adapter_path, "r", encoding="utf-8") as f:
adapter_config = json.load(f)
model = adapter_config["base_model_name_or_path"]
config = AutoConfig.from_pretrained(
model, _from_pipeline=task, code_revision=code_revision, **hub_kwargs, **model_kwargs
)
hub_kwargs["_commit_hash"] = config._commit_hash
custom_tasks = {}
if config is not None and len(getattr(config, "custom_pipelines", {})) > 0:
custom_tasks = config.custom_pipelines
if task is None and trust_remote_code is not False:
if len(custom_tasks) == 1:
task = list(custom_tasks.keys())[0]
else:
raise RuntimeError(
"We can't infer the task automatically for this model as there are multiple tasks available. Pick "
f"one in {', '.join(custom_tasks.keys())}"
)
if task is None and model is not None:
if not isinstance(model, str):
raise RuntimeError(
"Inferring the task automatically requires to check the hub with a model_id defined as a `str`. "
f"{model} is not a valid model_id."
)
task = get_task(model, token)
# Retrieve the task
if task in custom_tasks:
normalized_task = task
targeted_task, task_options = clean_custom_task(custom_tasks[task])
if pipeline_class is None:
if not trust_remote_code:
raise ValueError(
"Loading this pipeline requires you to execute the code in the pipeline file in that"
" repo on your local machine. Make sure you have read the code there to avoid malicious use, then"
" set the option `trust_remote_code=True` to remove this error."
)
class_ref = targeted_task["impl"]
pipeline_class = get_class_from_dynamic_module(
class_ref,
model,
code_revision=code_revision,
**hub_kwargs,
)
else:
normalized_task, targeted_task, task_options = check_task(task)
if pipeline_class is None:
pipeline_class = targeted_task["impl"]
# Use default model/config/tokenizer for the task if no model is provided
if model is None:
# At that point framework might still be undetermined
model, default_revision = get_default_model_and_revision(targeted_task, framework, task_options)
revision = revision if revision is not None else default_revision
logger.warning(
f"No model was supplied, defaulted to {model} and revision"
f" {revision} ({HUGGINGFACE_CO_RESOLVE_ENDPOINT}/{model}).\n"
"Using a pipeline without specifying a model name and revision in production is not recommended."
)
if config is None and isinstance(model, str):
config = AutoConfig.from_pretrained(model, _from_pipeline=task, **hub_kwargs, **model_kwargs)
hub_kwargs["_commit_hash"] = config._commit_hash
if device_map is not None:
if "device_map" in model_kwargs:
raise ValueError(
'You cannot use both `pipeline(... device_map=..., model_kwargs={"device_map":...})` as those'
" arguments might conflict, use only one.)"
)
if device is not None:
logger.warning(
"Both `device` and `device_map` are specified. `device` will override `device_map`. You"
" will most likely encounter unexpected behavior. Please remove `device` and keep `device_map`."
)
model_kwargs["device_map"] = device_map
if torch_dtype is not None:
if "torch_dtype" in model_kwargs:
raise ValueError(
'You cannot use both `pipeline(... torch_dtype=..., model_kwargs={"torch_dtype":...})` as those'
" arguments might conflict, use only one.)"
)
if isinstance(torch_dtype, str) and hasattr(torch, torch_dtype):
torch_dtype = getattr(torch, torch_dtype)
model_kwargs["torch_dtype"] = torch_dtype
model_name = model if isinstance(model, str) else None
# Load the correct model if possible
# Infer the framework from the model if not already defined
if isinstance(model, str) or framework is None:
model_classes = {"tf": targeted_task["tf"], "pt": targeted_task["pt"]}
framework, model = infer_framework_load_model(
model,
model_classes=model_classes,
config=config,
framework=framework,
task=task,
**hub_kwargs,
**model_kwargs,
)
model_config = model.config
hub_kwargs["_commit_hash"] = model.config._commit_hash
load_tokenizer = (
type(model_config) in TOKENIZER_MAPPING
or model_config.tokenizer_class is not None
or isinstance(tokenizer, str)
)
load_feature_extractor = type(model_config) in FEATURE_EXTRACTOR_MAPPING or feature_extractor is not None
load_image_processor = type(model_config) in IMAGE_PROCESSOR_MAPPING or image_processor is not None
# If `model` (instance of `PretrainedModel` instead of `str`) is passed (and/or same for config), while
# `image_processor` or `feature_extractor` is `None`, the loading will fail. This happens particularly for some
# vision tasks when calling `pipeline()` with `model` and only one of the `image_processor` and `feature_extractor`.
# TODO: we need to make `NO_IMAGE_PROCESSOR_TASKS` and `NO_FEATURE_EXTRACTOR_TASKS` more robust to avoid such issue.
# This block is only temporarily to make CI green.
if load_image_processor and load_feature_extractor:
load_feature_extractor = False
if (
tokenizer is None
and not load_tokenizer
and normalized_task not in NO_TOKENIZER_TASKS
# Using class name to avoid importing the real class.
and (
model_config.__class__.__name__ in MULTI_MODEL_AUDIO_CONFIGS
or model_config.__class__.__name__ in MULTI_MODEL_VISION_CONFIGS
)
):
# This is a special category of models, that are fusions of multiple models
# so the model_config might not define a tokenizer, but it seems to be
# necessary for the task, so we're force-trying to load it.
load_tokenizer = True
if (
image_processor is None
and not load_image_processor
and normalized_task not in NO_IMAGE_PROCESSOR_TASKS
# Using class name to avoid importing the real class.
and model_config.__class__.__name__ in MULTI_MODEL_VISION_CONFIGS
):
# This is a special category of models, that are fusions of multiple models
# so the model_config might not define a tokenizer, but it seems to be
# necessary for the task, so we're force-trying to load it.
load_image_processor = True
if (
feature_extractor is None
and not load_feature_extractor
and normalized_task not in NO_FEATURE_EXTRACTOR_TASKS
# Using class name to avoid importing the real class.
and model_config.__class__.__name__ in MULTI_MODEL_AUDIO_CONFIGS
):
# This is a special category of models, that are fusions of multiple models
# so the model_config might not define a tokenizer, but it seems to be
# necessary for the task, so we're force-trying to load it.
load_feature_extractor = True
if task in NO_TOKENIZER_TASKS:
# These will never require a tokenizer.
# the model on the other hand might have a tokenizer, but
# the files could be missing from the hub, instead of failing
# on such repos, we just force to not load it.
load_tokenizer = False
if task in NO_FEATURE_EXTRACTOR_TASKS:
load_feature_extractor = False
if task in NO_IMAGE_PROCESSOR_TASKS:
load_image_processor = False
if load_tokenizer:
# Try to infer tokenizer from model or config name (if provided as str)
if tokenizer is None:
if isinstance(model_name, str):
tokenizer = model_name
elif isinstance(config, str):
tokenizer = config
else:
# Impossible to guess what is the right tokenizer here
raise Exception(
"Impossible to guess which tokenizer to use. "
"Please provide a PreTrainedTokenizer class or a path/identifier to a pretrained tokenizer."
)
# Instantiate tokenizer if needed
if isinstance(tokenizer, (str, tuple)):
if isinstance(tokenizer, tuple):
# For tuple we have (tokenizer name, {kwargs})
use_fast = tokenizer[1].pop("use_fast", use_fast)
tokenizer_identifier = tokenizer[0]
tokenizer_kwargs = tokenizer[1]
else:
tokenizer_identifier = tokenizer
tokenizer_kwargs = model_kwargs.copy()
tokenizer_kwargs.pop("torch_dtype", None)
tokenizer = AutoTokenizer.from_pretrained(
tokenizer_identifier, use_fast=use_fast, _from_pipeline=task, **hub_kwargs, **tokenizer_kwargs
)
if load_image_processor:
# Try to infer image processor from model or config name (if provided as str)
if image_processor is None:
if isinstance(model_name, str):
image_processor = model_name
elif isinstance(config, str):
image_processor = config
# Backward compatibility, as `feature_extractor` used to be the name
# for `ImageProcessor`.
elif feature_extractor is not None and isinstance(feature_extractor, BaseImageProcessor):
image_processor = feature_extractor
else:
# Impossible to guess what is the right image_processor here
raise Exception(
"Impossible to guess which image processor to use. "
"Please provide a PreTrainedImageProcessor class or a path/identifier "
"to a pretrained image processor."
)
# Instantiate image_processor if needed
if isinstance(image_processor, (str, tuple)):
image_processor = AutoImageProcessor.from_pretrained(
image_processor, _from_pipeline=task, **hub_kwargs, **model_kwargs
)
if load_feature_extractor:
# Try to infer feature extractor from model or config name (if provided as str)
if feature_extractor is None:
if isinstance(model_name, str):
feature_extractor = model_name
elif isinstance(config, str):
feature_extractor = config
else:
# Impossible to guess what is the right feature_extractor here
raise Exception(
"Impossible to guess which feature extractor to use. "
"Please provide a PreTrainedFeatureExtractor class or a path/identifier "
"to a pretrained feature extractor."
)
# Instantiate feature_extractor if needed
if isinstance(feature_extractor, (str, tuple)):
feature_extractor = AutoFeatureExtractor.from_pretrained(
feature_extractor, _from_pipeline=task, **hub_kwargs, **model_kwargs
)
if (
feature_extractor._processor_class
and feature_extractor._processor_class.endswith("WithLM")
and isinstance(model_name, str)
):
try:
import kenlm # to trigger `ImportError` if not installed
from pyctcdecode import BeamSearchDecoderCTC
if os.path.isdir(model_name) or os.path.isfile(model_name):
decoder = BeamSearchDecoderCTC.load_from_dir(model_name)
else:
language_model_glob = os.path.join(
BeamSearchDecoderCTC._LANGUAGE_MODEL_SERIALIZED_DIRECTORY, "*"
)
alphabet_filename = BeamSearchDecoderCTC._ALPHABET_SERIALIZED_FILENAME
allow_patterns = [language_model_glob, alphabet_filename]
decoder = BeamSearchDecoderCTC.load_from_hf_hub(model_name, allow_patterns=allow_patterns)
kwargs["decoder"] = decoder
except ImportError as e:
logger.warning(f"Could not load the `decoder` for {model_name}. Defaulting to raw CTC. Error: {e}")
if not is_kenlm_available():
logger.warning("Try to install `kenlm`: `pip install kenlm")
if not is_pyctcdecode_available():
logger.warning("Try to install `pyctcdecode`: `pip install pyctcdecode")
if task == "translation" and model.config.task_specific_params:
for key in model.config.task_specific_params:
if key.startswith("translation"):
task = key
warnings.warn(
f'"translation" task was used, instead of "translation_XX_to_YY", defaulting to "{task}"',
UserWarning,
)
break
if tokenizer is not None:
kwargs["tokenizer"] = tokenizer
if feature_extractor is not None:
kwargs["feature_extractor"] = feature_extractor
if torch_dtype is not None:
kwargs["torch_dtype"] = torch_dtype
if image_processor is not None:
kwargs["image_processor"] = image_processor
if device is not None:
kwargs["device"] = device
return pipeline_class(model=model, framework=framework, task=task, **kwargs)
|
transformers/src/transformers/pipelines/__init__.py/0
|
{
"file_path": "transformers/src/transformers/pipelines/__init__.py",
"repo_id": "transformers",
"token_count": 21375
}
| 393
|
import numpy as np
import torch
from torch.utils.data import Dataset, IterableDataset
from ..utils.generic import ModelOutput
class PipelineDataset(Dataset):
def __init__(self, dataset, process, params):
self.dataset = dataset
self.process = process
self.params = params
def __len__(self):
return len(self.dataset)
def __getitem__(self, i):
item = self.dataset[i]
processed = self.process(item, **self.params)
return processed
class PipelineIterator(IterableDataset):
def __init__(self, loader, infer, params, loader_batch_size=None):
"""
Roughly equivalent to
```
for item in loader:
yield infer(item, **params)
```
Arguments:
loader (`torch.utils.data.DataLoader` or `Iterable`):
The iterator that will be used to apply `infer` on.
infer (any function):
The function to apply of each element of `loader`.
params (`dict`):
The parameters passed to `infer` along with every item
loader_batch_size (`int`, *optional*):
If specified, the items of `loader` are supposed to come as batch, and are loader_batched here
making it roughly behave as
```
for items in loader:
for i in loader_batch_size:
item = items[i]
yield infer(item, **params)
```"""
self.loader = loader
self.infer = infer
self.params = params
if loader_batch_size == 1:
# Let's spare some time by deactivating altogether
loader_batch_size = None
self.loader_batch_size = loader_batch_size
# Internal bookkeeping
self._loader_batch_index = None
self._loader_batch_data = None
def __len__(self):
return len(self.loader)
def __iter__(self):
self.iterator = iter(self.loader)
return self
def loader_batch_item(self):
"""
Return item located at `loader_batch_index` within the current `loader_batch_data`.
"""
if isinstance(self._loader_batch_data, torch.Tensor):
# Batch data is simple tensor, just fetch the slice
result = self._loader_batch_data[self._loader_batch_index].unsqueeze(0)
else:
# Batch data is assumed to be BaseModelOutput (or dict)
loader_batched = {}
for k, element in self._loader_batch_data.items():
if isinstance(element, ModelOutput):
# Convert ModelOutput to tuple first
element = element.to_tuple()
if isinstance(element[0], torch.Tensor):
loader_batched[k] = tuple(el[self._loader_batch_index].unsqueeze(0) for el in element)
elif isinstance(element[0], np.ndarray):
loader_batched[k] = tuple(np.expand_dims(el[self._loader_batch_index], 0) for el in element)
continue
if k in {"hidden_states", "past_key_values", "attentions"} and isinstance(element, tuple):
# Those are stored as lists of tensors so need specific unbatching.
if isinstance(element[0], torch.Tensor):
loader_batched[k] = tuple(el[self._loader_batch_index].unsqueeze(0) for el in element)
elif isinstance(element[0], np.ndarray):
loader_batched[k] = tuple(np.expand_dims(el[self._loader_batch_index], 0) for el in element)
continue
if element is None:
# This can happen for optional data that get passed around
loader_batched[k] = None
elif isinstance(element[self._loader_batch_index], torch.Tensor):
# Take correct batch data, but make it looked like batch_size=1
# For compatibility with other methods within transformers
loader_batched[k] = element[self._loader_batch_index].unsqueeze(0)
elif isinstance(element[self._loader_batch_index], np.ndarray):
# Take correct batch data, but make it looked like batch_size=1
# For compatibility with other methods within transformers
loader_batched[k] = np.expand_dims(element[self._loader_batch_index], 0)
else:
# This is typically a list, so no need to `unsqueeze`.
loader_batched[k] = element[self._loader_batch_index]
# Recreate the element by reusing the original class to make it look
# batch_size=1
result = self._loader_batch_data.__class__(loader_batched)
self._loader_batch_index += 1
return result
def __next__(self):
if self._loader_batch_index is not None and self._loader_batch_index < self.loader_batch_size:
# We are currently unrolling a batch so we just need to return
# the current item within a batch
return self.loader_batch_item()
# We're out of items within a batch
item = next(self.iterator)
processed = self.infer(item, **self.params)
# We now have a batch of "inferred things".
if self.loader_batch_size is not None:
# Try to infer the size of the batch
if isinstance(processed, torch.Tensor):
first_tensor = processed
elif isinstance(processed, tuple):
first_tensor = processed[0]
else:
key = list(processed.keys())[0]
first_tensor = processed[key]
if isinstance(first_tensor, list):
observed_batch_size = len(first_tensor)
else:
observed_batch_size = first_tensor.shape[0]
if 0 < observed_batch_size < self.loader_batch_size:
# could be last batch so we can't unroll as many
# elements.
self.loader_batch_size = observed_batch_size
# Setting internal index to unwrap the batch
self._loader_batch_data = processed[0] if isinstance(processed, tuple) else processed
self._loader_batch_index = 0
return self.loader_batch_item()
else:
# We're not unrolling batches
return processed
class PipelineChunkIterator(PipelineIterator):
def __init__(self, loader, infer, params, loader_batch_size=None):
"""
Roughly equivalent to
```
for iterator in loader:
for item in iterator:
yield infer(item, **params)
```
Arguments:
loader (`torch.utils.data.DataLoader` or `Iterable`):
The iterator that will be used to apply `infer` on.
infer (any function):
The function to apply of each element of `loader`.
params (`dict`):
The parameters passed to `infer` along with every item
"""
super().__init__(loader, infer, params)
def __iter__(self):
self.iterator = iter(self.loader)
self.subiterator = None
return self
def __next__(self):
if self.subiterator is None:
"Subiterator None means we haven't started a `preprocess` iterator. so start it"
self.subiterator = self.infer(next(self.iterator), **self.params)
try:
# Try to return next item
processed = next(self.subiterator)
except StopIteration:
# When a preprocess iterator ends, we can start lookig at the next item
# ChunkIterator will keep feeding until ALL elements of iterator
# all have created their subiterator and have been iterating against.
#
# Another way to look at it, is we're basically flattening lists of lists
# into a single list, but with generators
self.subiterator = self.infer(next(self.iterator), **self.params)
processed = next(self.subiterator)
return processed
class PipelinePackIterator(PipelineIterator):
"""
Roughly equivalent to
```
packed = []
for item in loader:
packed.append(item)
if item["is_last"]:
yield packed
packed = []
```
but it also handles cases where `item` are batched (meaning it's a dict of Tensor with first dimension > 1. In
that case it does
```
packed = []
for batch in loader:
# item is batched
for item in batch:
packed.append(item)
if item["is_last"]:
yield packed
packed = []
```
Arguments:
loader (`torch.utils.data.DataLoader` or `Iterable`):
The iterator that will be used to apply `infer` on.
infer (any function):
The function to apply of each element of `loader`.
params (`dict`):
The parameters passed to `infer` along with every item
loader_batch_size (`int`, *optional*):
If specified, the items of `loader` are supposed to come as batch, and are loader_batched here making
it roughly behave as
```
for items in loader:
for i in loader_batch_size:
item = items[i]
yield infer(item, **params)
```"""
def __iter__(self):
self.iterator = iter(self.loader)
return self
def __next__(self):
# Extremely similar to PipelineIterator in its unpacking mechanism
# BUT, we have an extra required item which is the presence of `is_last`
# That is because everything is flattened by `PipelineChunkIterator` we
# need to keep track of how to regroup here in the original `process`
# boundaries so that `process` and `postprocess` see the same data.
# This iterator accumulates items (possibly while unbatching) until it
# its a `is_last` and then just passes it on to the caller.
is_last = False
accumulator = []
if self._loader_batch_index is not None and self._loader_batch_index < self.loader_batch_size:
while self._loader_batch_index < self.loader_batch_size:
item = self.loader_batch_item()
is_last = item.pop("is_last")
accumulator.append(item)
if is_last:
return accumulator
while not is_last:
processed = self.infer(next(self.iterator), **self.params)
if self.loader_batch_size is not None:
if isinstance(processed, torch.Tensor):
first_tensor = processed
else:
key = list(processed.keys())[0]
first_tensor = processed[key]
if isinstance(first_tensor, list):
observed_batch_size = len(first_tensor)
else:
observed_batch_size = first_tensor.shape[0]
if 0 < observed_batch_size < self.loader_batch_size:
# could be last batch so we can't unroll as many
# elements.
self.loader_batch_size = observed_batch_size
self._loader_batch_data = processed
self._loader_batch_index = 0
while self._loader_batch_index < self.loader_batch_size:
item = self.loader_batch_item()
is_last = item.pop("is_last")
accumulator.append(item)
if is_last:
return accumulator
else:
item = processed
is_last = item.pop("is_last")
accumulator.append(item)
return accumulator
class KeyDataset(Dataset):
def __init__(self, dataset: Dataset, key: str):
self.dataset = dataset
self.key = key
def __len__(self):
return len(self.dataset)
def __getitem__(self, i):
return self.dataset[i][self.key]
class KeyPairDataset(Dataset):
def __init__(self, dataset: Dataset, key1: str, key2: str):
self.dataset = dataset
self.key1 = key1
self.key2 = key2
def __len__(self):
return len(self.dataset)
def __getitem__(self, i):
return {"text": self.dataset[i][self.key1], "text_pair": self.dataset[i][self.key2]}
|
transformers/src/transformers/pipelines/pt_utils.py/0
|
{
"file_path": "transformers/src/transformers/pipelines/pt_utils.py",
"repo_id": "transformers",
"token_count": 5958
}
| 394
|
#!/usr/bin/env python
# coding=utf-8
# Copyright 2021 The HuggingFace Inc. team. 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.
from huggingface_hub import get_full_repo_name # for backward compatibility
from huggingface_hub.constants import HF_HUB_DISABLE_TELEMETRY as DISABLE_TELEMETRY # for backward compatibility
from packaging import version
from .. import __version__
from .backbone_utils import BackboneConfigMixin, BackboneMixin
from .chat_template_utils import DocstringParsingException, TypeHintParsingException, get_json_schema
from .constants import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD, IMAGENET_STANDARD_MEAN, IMAGENET_STANDARD_STD
from .doc import (
add_code_sample_docstrings,
add_end_docstrings,
add_start_docstrings,
add_start_docstrings_to_model_forward,
copy_func,
replace_return_docstrings,
)
from .generic import (
ContextManagers,
ExplicitEnum,
ModelOutput,
PaddingStrategy,
TensorType,
add_model_info_to_auto_map,
add_model_info_to_custom_pipelines,
cached_property,
can_return_loss,
expand_dims,
filter_out_non_signature_kwargs,
find_labels,
flatten_dict,
infer_framework,
is_jax_tensor,
is_numpy_array,
is_tensor,
is_tf_symbolic_tensor,
is_tf_tensor,
is_torch_device,
is_torch_dtype,
is_torch_tensor,
reshape,
squeeze,
strtobool,
tensor_size,
to_numpy,
to_py_obj,
torch_float,
torch_int,
transpose,
working_or_temp_dir,
)
from .hub import (
CLOUDFRONT_DISTRIB_PREFIX,
HF_MODULES_CACHE,
HUGGINGFACE_CO_PREFIX,
HUGGINGFACE_CO_RESOLVE_ENDPOINT,
PYTORCH_PRETRAINED_BERT_CACHE,
PYTORCH_TRANSFORMERS_CACHE,
S3_BUCKET_PREFIX,
TRANSFORMERS_CACHE,
TRANSFORMERS_DYNAMIC_MODULE_NAME,
EntryNotFoundError,
PushInProgress,
PushToHubMixin,
RepositoryNotFoundError,
RevisionNotFoundError,
cached_file,
default_cache_path,
define_sagemaker_information,
download_url,
extract_commit_hash,
get_cached_models,
get_file_from_repo,
has_file,
http_user_agent,
is_offline_mode,
is_remote_url,
move_cache,
send_example_telemetry,
try_to_load_from_cache,
)
from .import_utils import (
ACCELERATE_MIN_VERSION,
ENV_VARS_TRUE_AND_AUTO_VALUES,
ENV_VARS_TRUE_VALUES,
TORCH_FX_REQUIRED_VERSION,
USE_JAX,
USE_TF,
USE_TORCH,
XLA_FSDPV2_MIN_VERSION,
DummyObject,
OptionalDependencyNotAvailable,
_LazyModule,
ccl_version,
direct_transformers_import,
get_torch_version,
is_accelerate_available,
is_apex_available,
is_aqlm_available,
is_auto_awq_available,
is_auto_gptq_available,
is_av_available,
is_bitsandbytes_available,
is_bs4_available,
is_coloredlogs_available,
is_cv2_available,
is_cython_available,
is_datasets_available,
is_decord_available,
is_detectron2_available,
is_eetq_available,
is_essentia_available,
is_faiss_available,
is_fbgemm_gpu_available,
is_flash_attn_2_available,
is_flash_attn_greater_or_equal,
is_flash_attn_greater_or_equal_2_10,
is_flax_available,
is_fsdp_available,
is_ftfy_available,
is_g2p_en_available,
is_galore_torch_available,
is_gguf_available,
is_grokadamw_available,
is_hqq_available,
is_in_notebook,
is_ipex_available,
is_jieba_available,
is_jinja_available,
is_jumanpp_available,
is_kenlm_available,
is_keras_nlp_available,
is_levenshtein_available,
is_librosa_available,
is_liger_kernel_available,
is_lomo_available,
is_mlx_available,
is_natten_available,
is_ninja_available,
is_nltk_available,
is_onnx_available,
is_openai_available,
is_optimum_available,
is_pandas_available,
is_peft_available,
is_phonemizer_available,
is_pretty_midi_available,
is_protobuf_available,
is_psutil_available,
is_py3nvml_available,
is_pyctcdecode_available,
is_pytesseract_available,
is_pytest_available,
is_pytorch_quantization_available,
is_quanto_available,
is_rjieba_available,
is_sacremoses_available,
is_safetensors_available,
is_sagemaker_dp_enabled,
is_sagemaker_mp_enabled,
is_scipy_available,
is_sentencepiece_available,
is_seqio_available,
is_sklearn_available,
is_soundfile_availble,
is_spacy_available,
is_speech_available,
is_sudachi_available,
is_sudachi_projection_available,
is_tensorflow_probability_available,
is_tensorflow_text_available,
is_tf2onnx_available,
is_tf_available,
is_timm_available,
is_tokenizers_available,
is_torch_available,
is_torch_bf16_available,
is_torch_bf16_available_on_device,
is_torch_bf16_cpu_available,
is_torch_bf16_gpu_available,
is_torch_compile_available,
is_torch_cuda_available,
is_torch_deterministic,
is_torch_fp16_available_on_device,
is_torch_fx_available,
is_torch_fx_proxy,
is_torch_mlu_available,
is_torch_mps_available,
is_torch_musa_available,
is_torch_neuroncore_available,
is_torch_npu_available,
is_torch_sdpa_available,
is_torch_tensorrt_fx_available,
is_torch_tf32_available,
is_torch_tpu_available,
is_torch_xla_available,
is_torch_xpu_available,
is_torchao_available,
is_torchaudio_available,
is_torchdistx_available,
is_torchdynamo_available,
is_torchdynamo_compiling,
is_torchvision_available,
is_training_run_on_sagemaker,
is_uroman_available,
is_vision_available,
requires_backends,
torch_only_method,
)
from .peft_utils import (
ADAPTER_CONFIG_NAME,
ADAPTER_SAFE_WEIGHTS_NAME,
ADAPTER_WEIGHTS_NAME,
check_peft_version,
find_adapter_config_file,
)
WEIGHTS_NAME = "pytorch_model.bin"
WEIGHTS_INDEX_NAME = "pytorch_model.bin.index.json"
TF2_WEIGHTS_NAME = "tf_model.h5"
TF2_WEIGHTS_INDEX_NAME = "tf_model.h5.index.json"
TF_WEIGHTS_NAME = "model.ckpt"
FLAX_WEIGHTS_NAME = "flax_model.msgpack"
FLAX_WEIGHTS_INDEX_NAME = "flax_model.msgpack.index.json"
SAFE_WEIGHTS_NAME = "model.safetensors"
SAFE_WEIGHTS_INDEX_NAME = "model.safetensors.index.json"
CONFIG_NAME = "config.json"
FEATURE_EXTRACTOR_NAME = "preprocessor_config.json"
IMAGE_PROCESSOR_NAME = FEATURE_EXTRACTOR_NAME
PROCESSOR_NAME = "processor_config.json"
CHAT_TEMPLATE_NAME = "chat_template.json"
GENERATION_CONFIG_NAME = "generation_config.json"
MODEL_CARD_NAME = "modelcard.json"
SENTENCEPIECE_UNDERLINE = "▁"
SPIECE_UNDERLINE = SENTENCEPIECE_UNDERLINE # Kept for backward compatibility
MULTIPLE_CHOICE_DUMMY_INPUTS = [
[[0, 1, 0, 1], [1, 0, 0, 1]]
] * 2 # Needs to have 0s and 1s only since XLM uses it for langs too.
DUMMY_INPUTS = [[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]]
DUMMY_MASK = [[1, 1, 1, 1, 1], [1, 1, 1, 0, 0], [0, 0, 0, 1, 1]]
def check_min_version(min_version):
if version.parse(__version__) < version.parse(min_version):
if "dev" in min_version:
error_message = (
"This example requires a source install from HuggingFace Transformers (see "
"`https://huggingface.co/docs/transformers/installation#install-from-source`),"
)
else:
error_message = f"This example requires a minimum version of {min_version},"
error_message += f" but the version found is {__version__}.\n"
raise ImportError(
error_message
+ "Check out https://github.com/huggingface/transformers/tree/main/examples#important-note for the examples corresponding to other "
"versions of HuggingFace Transformers."
)
|
transformers/src/transformers/utils/__init__.py/0
|
{
"file_path": "transformers/src/transformers/utils/__init__.py",
"repo_id": "transformers",
"token_count": 3589
}
| 395
|
# This file is autogenerated by the command `make fix-copies`, do not edit.
from ..utils import DummyObject, requires_backends
class TFBertTokenizer(metaclass=DummyObject):
_backends = ["tensorflow_text"]
def __init__(self, *args, **kwargs):
requires_backends(self, ["tensorflow_text"])
|
transformers/src/transformers/utils/dummy_tensorflow_text_objects.py/0
|
{
"file_path": "transformers/src/transformers/utils/dummy_tensorflow_text_objects.py",
"repo_id": "transformers",
"token_count": 109
}
| 396
|
# Generated by the protocol buffer compiler. DO NOT EDIT!
# source: sentencepiece_model.proto
# Copyright 2022 The HuggingFace Team. 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.
from google.protobuf import descriptor as _descriptor
from google.protobuf import message as _message
from google.protobuf import reflection as _reflection
from google.protobuf import symbol_database as _symbol_database
# @@protoc_insertion_point(imports)
_sym_db = _symbol_database.Default()
DESCRIPTOR = _descriptor.FileDescriptor(
name="sentencepiece_model.proto",
package="sentencepiece",
syntax="proto2",
serialized_options=b"H\003",
create_key=_descriptor._internal_create_key,
serialized_pb=(
b'\n\x19sentencepiece_model.proto\x12\rsentencepiece"\xa1\n\n\x0bTrainerSpec\x12\r\n\x05input\x18\x01'
b" \x03(\t\x12\x14\n\x0cinput_format\x18\x07 \x01(\t\x12\x14\n\x0cmodel_prefix\x18\x02"
b" \x01(\t\x12\x41\n\nmodel_type\x18\x03"
b" \x01(\x0e\x32$.sentencepiece.TrainerSpec.ModelType:\x07UNIGRAM\x12\x18\n\nvocab_size\x18\x04"
b" \x01(\x05:\x04\x38\x30\x30\x30\x12\x17\n\x0f\x61\x63\x63\x65pt_language\x18\x05 \x03(\t\x12"
b' \n\x15self_test_sample_size\x18\x06 \x01(\x05:\x01\x30\x12"\n\x12\x63haracter_coverage\x18\n'
b" \x01(\x02:\x06\x30.9995\x12\x1e\n\x13input_sentence_size\x18\x0b"
b" \x01(\x04:\x01\x30\x12$\n\x16shuffle_input_sentence\x18\x13 \x01(\x08:\x04true\x12"
b' \n\x14mining_sentence_size\x18\x0c \x01(\x05\x42\x02\x18\x01\x12"\n\x16training_sentence_size\x18\r'
b" \x01(\x05\x42\x02\x18\x01\x12(\n\x17seed_sentencepiece_size\x18\x0e"
b" \x01(\x05:\x07\x31\x30\x30\x30\x30\x30\x30\x12\x1e\n\x10shrinking_factor\x18\x0f"
b" \x01(\x02:\x04\x30.75\x12!\n\x13max_sentence_length\x18\x12"
b" \x01(\x05:\x04\x34\x31\x39\x32\x12\x17\n\x0bnum_threads\x18\x10"
b" \x01(\x05:\x02\x31\x36\x12\x1d\n\x12num_sub_iterations\x18\x11"
b" \x01(\x05:\x01\x32\x12$\n\x18max_sentencepiece_length\x18\x14"
b" \x01(\x05:\x02\x31\x36\x12%\n\x17split_by_unicode_script\x18\x15"
b" \x01(\x08:\x04true\x12\x1d\n\x0fsplit_by_number\x18\x17"
b" \x01(\x08:\x04true\x12!\n\x13split_by_whitespace\x18\x16"
b" \x01(\x08:\x04true\x12)\n\x1atreat_whitespace_as_suffix\x18\x18"
b" \x01(\x08:\x05\x66\x61lse\x12\x1b\n\x0csplit_digits\x18\x19"
b" \x01(\x08:\x05\x66\x61lse\x12\x17\n\x0f\x63ontrol_symbols\x18\x1e"
b" \x03(\t\x12\x1c\n\x14user_defined_symbols\x18\x1f \x03(\t\x12\x16\n\x0erequired_chars\x18$"
b" \x01(\t\x12\x1c\n\rbyte_fallback\x18# \x01(\x08:\x05\x66\x61lse\x12+\n\x1dvocabulary_output_piece_score\x18"
b' \x01(\x08:\x04true\x12\x1e\n\x10hard_vocab_limit\x18! \x01(\x08:\x04true\x12\x1c\n\ruse_all_vocab\x18"'
b" \x01(\x08:\x05\x66\x61lse\x12\x11\n\x06unk_id\x18( \x01(\x05:\x01\x30\x12\x11\n\x06\x62os_id\x18)"
b" \x01(\x05:\x01\x31\x12\x11\n\x06\x65os_id\x18* \x01(\x05:\x01\x32\x12\x12\n\x06pad_id\x18+"
b" \x01(\x05:\x02-1\x12\x18\n\tunk_piece\x18- \x01(\t:\x05<unk>\x12\x16\n\tbos_piece\x18."
b" \x01(\t:\x03<s>\x12\x17\n\teos_piece\x18/ \x01(\t:\x04</s>\x12\x18\n\tpad_piece\x18\x30"
b" \x01(\t:\x05<pad>\x12\x1a\n\x0bunk_surface\x18, \x01(\t:\x05 \xe2\x81\x87"
b" \x12+\n\x1ctrain_extremely_large_corpus\x18\x31"
b' \x01(\x08:\x05\x66\x61lse"5\n\tModelType\x12\x0b\n\x07UNIGRAM\x10\x01\x12\x07\n\x03\x42PE\x10\x02\x12\x08\n\x04WORD\x10\x03\x12\x08\n\x04\x43HAR\x10\x04*\t\x08\xc8\x01\x10\x80\x80\x80\x80\x02"\xd1\x01\n\x0eNormalizerSpec\x12\x0c\n\x04name\x18\x01'
b" \x01(\t\x12\x1c\n\x14precompiled_charsmap\x18\x02 \x01(\x0c\x12\x1e\n\x10\x61\x64\x64_dummy_prefix\x18\x03"
b" \x01(\x08:\x04true\x12&\n\x18remove_extra_whitespaces\x18\x04 \x01(\x08:\x04true\x12"
b" \n\x12\x65scape_whitespaces\x18\x05 \x01(\x08:\x04true\x12\x1e\n\x16normalization_rule_tsv\x18\x06"
b' \x01(\t*\t\x08\xc8\x01\x10\x80\x80\x80\x80\x02"y\n\x0cSelfTestData\x12\x33\n\x07samples\x18\x01'
b' \x03(\x0b\x32".sentencepiece.SelfTestData.Sample\x1a)\n\x06Sample\x12\r\n\x05input\x18\x01'
b" \x01(\t\x12\x10\n\x08\x65xpected\x18\x02"
b' \x01(\t*\t\x08\xc8\x01\x10\x80\x80\x80\x80\x02"\xfe\x03\n\nModelProto\x12\x37\n\x06pieces\x18\x01'
b" \x03(\x0b\x32'.sentencepiece.ModelProto.SentencePiece\x12\x30\n\x0ctrainer_spec\x18\x02"
b" \x01(\x0b\x32\x1a.sentencepiece.TrainerSpec\x12\x36\n\x0fnormalizer_spec\x18\x03"
b" \x01(\x0b\x32\x1d.sentencepiece.NormalizerSpec\x12\x33\n\x0eself_test_data\x18\x04"
b" \x01(\x0b\x32\x1b.sentencepiece.SelfTestData\x12\x38\n\x11\x64\x65normalizer_spec\x18\x05"
b" \x01(\x0b\x32\x1d.sentencepiece.NormalizerSpec\x1a\xd2\x01\n\rSentencePiece\x12\r\n\x05piece\x18\x01"
b" \x01(\t\x12\r\n\x05score\x18\x02 \x01(\x02\x12\x42\n\x04type\x18\x03"
b' \x01(\x0e\x32,.sentencepiece.ModelProto.SentencePiece.Type:\x06NORMAL"T\n\x04Type\x12\n\n\x06NORMAL\x10\x01\x12\x0b\n\x07UNKNOWN\x10\x02\x12\x0b\n\x07\x43ONTROL\x10\x03\x12\x10\n\x0cUSER_DEFINED\x10\x04\x12\x08\n\x04\x42YTE\x10\x06\x12\n\n\x06UNUSED\x10\x05*\t\x08\xc8\x01\x10\x80\x80\x80\x80\x02*\t\x08\xc8\x01\x10\x80\x80\x80\x80\x02\x42\x02H\x03'
),
)
_TRAINERSPEC_MODELTYPE = _descriptor.EnumDescriptor(
name="ModelType",
full_name="sentencepiece.TrainerSpec.ModelType",
filename=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
values=[
_descriptor.EnumValueDescriptor(
name="UNIGRAM",
index=0,
number=1,
serialized_options=None,
type=None,
create_key=_descriptor._internal_create_key,
),
_descriptor.EnumValueDescriptor(
name="BPE",
index=1,
number=2,
serialized_options=None,
type=None,
create_key=_descriptor._internal_create_key,
),
_descriptor.EnumValueDescriptor(
name="WORD",
index=2,
number=3,
serialized_options=None,
type=None,
create_key=_descriptor._internal_create_key,
),
_descriptor.EnumValueDescriptor(
name="CHAR",
index=3,
number=4,
serialized_options=None,
type=None,
create_key=_descriptor._internal_create_key,
),
],
containing_type=None,
serialized_options=None,
serialized_start=1294,
serialized_end=1347,
)
_sym_db.RegisterEnumDescriptor(_TRAINERSPEC_MODELTYPE)
_MODELPROTO_SENTENCEPIECE_TYPE = _descriptor.EnumDescriptor(
name="Type",
full_name="sentencepiece.ModelProto.SentencePiece.Type",
filename=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
values=[
_descriptor.EnumValueDescriptor(
name="NORMAL",
index=0,
number=1,
serialized_options=None,
type=None,
create_key=_descriptor._internal_create_key,
),
_descriptor.EnumValueDescriptor(
name="UNKNOWN",
index=1,
number=2,
serialized_options=None,
type=None,
create_key=_descriptor._internal_create_key,
),
_descriptor.EnumValueDescriptor(
name="CONTROL",
index=2,
number=3,
serialized_options=None,
type=None,
create_key=_descriptor._internal_create_key,
),
_descriptor.EnumValueDescriptor(
name="USER_DEFINED",
index=3,
number=4,
serialized_options=None,
type=None,
create_key=_descriptor._internal_create_key,
),
_descriptor.EnumValueDescriptor(
name="BYTE",
index=4,
number=6,
serialized_options=None,
type=None,
create_key=_descriptor._internal_create_key,
),
_descriptor.EnumValueDescriptor(
name="UNUSED",
index=5,
number=5,
serialized_options=None,
type=None,
create_key=_descriptor._internal_create_key,
),
],
containing_type=None,
serialized_options=None,
serialized_start=2100,
serialized_end=2184,
)
_sym_db.RegisterEnumDescriptor(_MODELPROTO_SENTENCEPIECE_TYPE)
_TRAINERSPEC = _descriptor.Descriptor(
name="TrainerSpec",
full_name="sentencepiece.TrainerSpec",
filename=None,
file=DESCRIPTOR,
containing_type=None,
create_key=_descriptor._internal_create_key,
fields=[
_descriptor.FieldDescriptor(
name="input",
full_name="sentencepiece.TrainerSpec.input",
index=0,
number=1,
type=9,
cpp_type=9,
label=3,
has_default_value=False,
default_value=[],
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="input_format",
full_name="sentencepiece.TrainerSpec.input_format",
index=1,
number=7,
type=9,
cpp_type=9,
label=1,
has_default_value=False,
default_value=b"".decode("utf-8"),
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="model_prefix",
full_name="sentencepiece.TrainerSpec.model_prefix",
index=2,
number=2,
type=9,
cpp_type=9,
label=1,
has_default_value=False,
default_value=b"".decode("utf-8"),
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="model_type",
full_name="sentencepiece.TrainerSpec.model_type",
index=3,
number=3,
type=14,
cpp_type=8,
label=1,
has_default_value=True,
default_value=1,
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="vocab_size",
full_name="sentencepiece.TrainerSpec.vocab_size",
index=4,
number=4,
type=5,
cpp_type=1,
label=1,
has_default_value=True,
default_value=8000,
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="accept_language",
full_name="sentencepiece.TrainerSpec.accept_language",
index=5,
number=5,
type=9,
cpp_type=9,
label=3,
has_default_value=False,
default_value=[],
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="self_test_sample_size",
full_name="sentencepiece.TrainerSpec.self_test_sample_size",
index=6,
number=6,
type=5,
cpp_type=1,
label=1,
has_default_value=True,
default_value=0,
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="character_coverage",
full_name="sentencepiece.TrainerSpec.character_coverage",
index=7,
number=10,
type=2,
cpp_type=6,
label=1,
has_default_value=True,
default_value=float(0.9995),
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="input_sentence_size",
full_name="sentencepiece.TrainerSpec.input_sentence_size",
index=8,
number=11,
type=4,
cpp_type=4,
label=1,
has_default_value=True,
default_value=0,
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="shuffle_input_sentence",
full_name="sentencepiece.TrainerSpec.shuffle_input_sentence",
index=9,
number=19,
type=8,
cpp_type=7,
label=1,
has_default_value=True,
default_value=True,
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="mining_sentence_size",
full_name="sentencepiece.TrainerSpec.mining_sentence_size",
index=10,
number=12,
type=5,
cpp_type=1,
label=1,
has_default_value=False,
default_value=0,
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=b"\030\001",
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="training_sentence_size",
full_name="sentencepiece.TrainerSpec.training_sentence_size",
index=11,
number=13,
type=5,
cpp_type=1,
label=1,
has_default_value=False,
default_value=0,
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=b"\030\001",
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="seed_sentencepiece_size",
full_name="sentencepiece.TrainerSpec.seed_sentencepiece_size",
index=12,
number=14,
type=5,
cpp_type=1,
label=1,
has_default_value=True,
default_value=1000000,
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="shrinking_factor",
full_name="sentencepiece.TrainerSpec.shrinking_factor",
index=13,
number=15,
type=2,
cpp_type=6,
label=1,
has_default_value=True,
default_value=float(0.75),
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="max_sentence_length",
full_name="sentencepiece.TrainerSpec.max_sentence_length",
index=14,
number=18,
type=5,
cpp_type=1,
label=1,
has_default_value=True,
default_value=4192,
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="num_threads",
full_name="sentencepiece.TrainerSpec.num_threads",
index=15,
number=16,
type=5,
cpp_type=1,
label=1,
has_default_value=True,
default_value=16,
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="num_sub_iterations",
full_name="sentencepiece.TrainerSpec.num_sub_iterations",
index=16,
number=17,
type=5,
cpp_type=1,
label=1,
has_default_value=True,
default_value=2,
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="max_sentencepiece_length",
full_name="sentencepiece.TrainerSpec.max_sentencepiece_length",
index=17,
number=20,
type=5,
cpp_type=1,
label=1,
has_default_value=True,
default_value=16,
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="split_by_unicode_script",
full_name="sentencepiece.TrainerSpec.split_by_unicode_script",
index=18,
number=21,
type=8,
cpp_type=7,
label=1,
has_default_value=True,
default_value=True,
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="split_by_number",
full_name="sentencepiece.TrainerSpec.split_by_number",
index=19,
number=23,
type=8,
cpp_type=7,
label=1,
has_default_value=True,
default_value=True,
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="split_by_whitespace",
full_name="sentencepiece.TrainerSpec.split_by_whitespace",
index=20,
number=22,
type=8,
cpp_type=7,
label=1,
has_default_value=True,
default_value=True,
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="treat_whitespace_as_suffix",
full_name="sentencepiece.TrainerSpec.treat_whitespace_as_suffix",
index=21,
number=24,
type=8,
cpp_type=7,
label=1,
has_default_value=True,
default_value=False,
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="split_digits",
full_name="sentencepiece.TrainerSpec.split_digits",
index=22,
number=25,
type=8,
cpp_type=7,
label=1,
has_default_value=True,
default_value=False,
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="control_symbols",
full_name="sentencepiece.TrainerSpec.control_symbols",
index=23,
number=30,
type=9,
cpp_type=9,
label=3,
has_default_value=False,
default_value=[],
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="user_defined_symbols",
full_name="sentencepiece.TrainerSpec.user_defined_symbols",
index=24,
number=31,
type=9,
cpp_type=9,
label=3,
has_default_value=False,
default_value=[],
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="required_chars",
full_name="sentencepiece.TrainerSpec.required_chars",
index=25,
number=36,
type=9,
cpp_type=9,
label=1,
has_default_value=False,
default_value=b"".decode("utf-8"),
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="byte_fallback",
full_name="sentencepiece.TrainerSpec.byte_fallback",
index=26,
number=35,
type=8,
cpp_type=7,
label=1,
has_default_value=True,
default_value=False,
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="vocabulary_output_piece_score",
full_name="sentencepiece.TrainerSpec.vocabulary_output_piece_score",
index=27,
number=32,
type=8,
cpp_type=7,
label=1,
has_default_value=True,
default_value=True,
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="hard_vocab_limit",
full_name="sentencepiece.TrainerSpec.hard_vocab_limit",
index=28,
number=33,
type=8,
cpp_type=7,
label=1,
has_default_value=True,
default_value=True,
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="use_all_vocab",
full_name="sentencepiece.TrainerSpec.use_all_vocab",
index=29,
number=34,
type=8,
cpp_type=7,
label=1,
has_default_value=True,
default_value=False,
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="unk_id",
full_name="sentencepiece.TrainerSpec.unk_id",
index=30,
number=40,
type=5,
cpp_type=1,
label=1,
has_default_value=True,
default_value=0,
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="bos_id",
full_name="sentencepiece.TrainerSpec.bos_id",
index=31,
number=41,
type=5,
cpp_type=1,
label=1,
has_default_value=True,
default_value=1,
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="eos_id",
full_name="sentencepiece.TrainerSpec.eos_id",
index=32,
number=42,
type=5,
cpp_type=1,
label=1,
has_default_value=True,
default_value=2,
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="pad_id",
full_name="sentencepiece.TrainerSpec.pad_id",
index=33,
number=43,
type=5,
cpp_type=1,
label=1,
has_default_value=True,
default_value=-1,
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="unk_piece",
full_name="sentencepiece.TrainerSpec.unk_piece",
index=34,
number=45,
type=9,
cpp_type=9,
label=1,
has_default_value=True,
default_value=b"<unk>".decode("utf-8"),
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="bos_piece",
full_name="sentencepiece.TrainerSpec.bos_piece",
index=35,
number=46,
type=9,
cpp_type=9,
label=1,
has_default_value=True,
default_value=b"<s>".decode("utf-8"),
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="eos_piece",
full_name="sentencepiece.TrainerSpec.eos_piece",
index=36,
number=47,
type=9,
cpp_type=9,
label=1,
has_default_value=True,
default_value=b"</s>".decode("utf-8"),
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="pad_piece",
full_name="sentencepiece.TrainerSpec.pad_piece",
index=37,
number=48,
type=9,
cpp_type=9,
label=1,
has_default_value=True,
default_value=b"<pad>".decode("utf-8"),
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="unk_surface",
full_name="sentencepiece.TrainerSpec.unk_surface",
index=38,
number=44,
type=9,
cpp_type=9,
label=1,
has_default_value=True,
default_value=b" \342\201\207 ".decode("utf-8"),
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="train_extremely_large_corpus",
full_name="sentencepiece.TrainerSpec.train_extremely_large_corpus",
index=39,
number=49,
type=8,
cpp_type=7,
label=1,
has_default_value=True,
default_value=False,
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
],
extensions=[],
nested_types=[],
enum_types=[
_TRAINERSPEC_MODELTYPE,
],
serialized_options=None,
is_extendable=True,
syntax="proto2",
extension_ranges=[
(200, 536870912),
],
oneofs=[],
serialized_start=45,
serialized_end=1358,
)
_NORMALIZERSPEC = _descriptor.Descriptor(
name="NormalizerSpec",
full_name="sentencepiece.NormalizerSpec",
filename=None,
file=DESCRIPTOR,
containing_type=None,
create_key=_descriptor._internal_create_key,
fields=[
_descriptor.FieldDescriptor(
name="name",
full_name="sentencepiece.NormalizerSpec.name",
index=0,
number=1,
type=9,
cpp_type=9,
label=1,
has_default_value=False,
default_value=b"".decode("utf-8"),
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="precompiled_charsmap",
full_name="sentencepiece.NormalizerSpec.precompiled_charsmap",
index=1,
number=2,
type=12,
cpp_type=9,
label=1,
has_default_value=False,
default_value=b"",
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="add_dummy_prefix",
full_name="sentencepiece.NormalizerSpec.add_dummy_prefix",
index=2,
number=3,
type=8,
cpp_type=7,
label=1,
has_default_value=True,
default_value=True,
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="remove_extra_whitespaces",
full_name="sentencepiece.NormalizerSpec.remove_extra_whitespaces",
index=3,
number=4,
type=8,
cpp_type=7,
label=1,
has_default_value=True,
default_value=True,
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="escape_whitespaces",
full_name="sentencepiece.NormalizerSpec.escape_whitespaces",
index=4,
number=5,
type=8,
cpp_type=7,
label=1,
has_default_value=True,
default_value=True,
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="normalization_rule_tsv",
full_name="sentencepiece.NormalizerSpec.normalization_rule_tsv",
index=5,
number=6,
type=9,
cpp_type=9,
label=1,
has_default_value=False,
default_value=b"".decode("utf-8"),
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
],
extensions=[],
nested_types=[],
enum_types=[],
serialized_options=None,
is_extendable=True,
syntax="proto2",
extension_ranges=[
(200, 536870912),
],
oneofs=[],
serialized_start=1361,
serialized_end=1570,
)
_SELFTESTDATA_SAMPLE = _descriptor.Descriptor(
name="Sample",
full_name="sentencepiece.SelfTestData.Sample",
filename=None,
file=DESCRIPTOR,
containing_type=None,
create_key=_descriptor._internal_create_key,
fields=[
_descriptor.FieldDescriptor(
name="input",
full_name="sentencepiece.SelfTestData.Sample.input",
index=0,
number=1,
type=9,
cpp_type=9,
label=1,
has_default_value=False,
default_value=b"".decode("utf-8"),
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="expected",
full_name="sentencepiece.SelfTestData.Sample.expected",
index=1,
number=2,
type=9,
cpp_type=9,
label=1,
has_default_value=False,
default_value=b"".decode("utf-8"),
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
],
extensions=[],
nested_types=[],
enum_types=[],
serialized_options=None,
is_extendable=False,
syntax="proto2",
extension_ranges=[],
oneofs=[],
serialized_start=1641,
serialized_end=1682,
)
_SELFTESTDATA = _descriptor.Descriptor(
name="SelfTestData",
full_name="sentencepiece.SelfTestData",
filename=None,
file=DESCRIPTOR,
containing_type=None,
create_key=_descriptor._internal_create_key,
fields=[
_descriptor.FieldDescriptor(
name="samples",
full_name="sentencepiece.SelfTestData.samples",
index=0,
number=1,
type=11,
cpp_type=10,
label=3,
has_default_value=False,
default_value=[],
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
],
extensions=[],
nested_types=[
_SELFTESTDATA_SAMPLE,
],
enum_types=[],
serialized_options=None,
is_extendable=True,
syntax="proto2",
extension_ranges=[
(200, 536870912),
],
oneofs=[],
serialized_start=1572,
serialized_end=1693,
)
_MODELPROTO_SENTENCEPIECE = _descriptor.Descriptor(
name="SentencePiece",
full_name="sentencepiece.ModelProto.SentencePiece",
filename=None,
file=DESCRIPTOR,
containing_type=None,
create_key=_descriptor._internal_create_key,
fields=[
_descriptor.FieldDescriptor(
name="piece",
full_name="sentencepiece.ModelProto.SentencePiece.piece",
index=0,
number=1,
type=9,
cpp_type=9,
label=1,
has_default_value=False,
default_value=b"".decode("utf-8"),
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="score",
full_name="sentencepiece.ModelProto.SentencePiece.score",
index=1,
number=2,
type=2,
cpp_type=6,
label=1,
has_default_value=False,
default_value=float(0),
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="type",
full_name="sentencepiece.ModelProto.SentencePiece.type",
index=2,
number=3,
type=14,
cpp_type=8,
label=1,
has_default_value=True,
default_value=1,
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
],
extensions=[],
nested_types=[],
enum_types=[
_MODELPROTO_SENTENCEPIECE_TYPE,
],
serialized_options=None,
is_extendable=True,
syntax="proto2",
extension_ranges=[
(200, 536870912),
],
oneofs=[],
serialized_start=1985,
serialized_end=2195,
)
_MODELPROTO = _descriptor.Descriptor(
name="ModelProto",
full_name="sentencepiece.ModelProto",
filename=None,
file=DESCRIPTOR,
containing_type=None,
create_key=_descriptor._internal_create_key,
fields=[
_descriptor.FieldDescriptor(
name="pieces",
full_name="sentencepiece.ModelProto.pieces",
index=0,
number=1,
type=11,
cpp_type=10,
label=3,
has_default_value=False,
default_value=[],
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="trainer_spec",
full_name="sentencepiece.ModelProto.trainer_spec",
index=1,
number=2,
type=11,
cpp_type=10,
label=1,
has_default_value=False,
default_value=None,
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="normalizer_spec",
full_name="sentencepiece.ModelProto.normalizer_spec",
index=2,
number=3,
type=11,
cpp_type=10,
label=1,
has_default_value=False,
default_value=None,
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="self_test_data",
full_name="sentencepiece.ModelProto.self_test_data",
index=3,
number=4,
type=11,
cpp_type=10,
label=1,
has_default_value=False,
default_value=None,
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
_descriptor.FieldDescriptor(
name="denormalizer_spec",
full_name="sentencepiece.ModelProto.denormalizer_spec",
index=4,
number=5,
type=11,
cpp_type=10,
label=1,
has_default_value=False,
default_value=None,
message_type=None,
enum_type=None,
containing_type=None,
is_extension=False,
extension_scope=None,
serialized_options=None,
file=DESCRIPTOR,
create_key=_descriptor._internal_create_key,
),
],
extensions=[],
nested_types=[
_MODELPROTO_SENTENCEPIECE,
],
enum_types=[],
serialized_options=None,
is_extendable=True,
syntax="proto2",
extension_ranges=[
(200, 536870912),
],
oneofs=[],
serialized_start=1696,
serialized_end=2206,
)
_TRAINERSPEC.fields_by_name["model_type"].enum_type = _TRAINERSPEC_MODELTYPE
_TRAINERSPEC_MODELTYPE.containing_type = _TRAINERSPEC
_SELFTESTDATA_SAMPLE.containing_type = _SELFTESTDATA
_SELFTESTDATA.fields_by_name["samples"].message_type = _SELFTESTDATA_SAMPLE
_MODELPROTO_SENTENCEPIECE.fields_by_name["type"].enum_type = _MODELPROTO_SENTENCEPIECE_TYPE
_MODELPROTO_SENTENCEPIECE.containing_type = _MODELPROTO
_MODELPROTO_SENTENCEPIECE_TYPE.containing_type = _MODELPROTO_SENTENCEPIECE
_MODELPROTO.fields_by_name["pieces"].message_type = _MODELPROTO_SENTENCEPIECE
_MODELPROTO.fields_by_name["trainer_spec"].message_type = _TRAINERSPEC
_MODELPROTO.fields_by_name["normalizer_spec"].message_type = _NORMALIZERSPEC
_MODELPROTO.fields_by_name["self_test_data"].message_type = _SELFTESTDATA
_MODELPROTO.fields_by_name["denormalizer_spec"].message_type = _NORMALIZERSPEC
DESCRIPTOR.message_types_by_name["TrainerSpec"] = _TRAINERSPEC
DESCRIPTOR.message_types_by_name["NormalizerSpec"] = _NORMALIZERSPEC
DESCRIPTOR.message_types_by_name["SelfTestData"] = _SELFTESTDATA
DESCRIPTOR.message_types_by_name["ModelProto"] = _MODELPROTO
_sym_db.RegisterFileDescriptor(DESCRIPTOR)
TrainerSpec = _reflection.GeneratedProtocolMessageType(
"TrainerSpec",
(_message.Message,),
{
"DESCRIPTOR": _TRAINERSPEC,
"__module__": "sentencepiece_model_pb2",
# @@protoc_insertion_point(class_scope:sentencepiece.TrainerSpec)
},
)
_sym_db.RegisterMessage(TrainerSpec)
NormalizerSpec = _reflection.GeneratedProtocolMessageType(
"NormalizerSpec",
(_message.Message,),
{
"DESCRIPTOR": _NORMALIZERSPEC,
"__module__": "sentencepiece_model_pb2",
# @@protoc_insertion_point(class_scope:sentencepiece.NormalizerSpec)
},
)
_sym_db.RegisterMessage(NormalizerSpec)
SelfTestData = _reflection.GeneratedProtocolMessageType(
"SelfTestData",
(_message.Message,),
{
"Sample": _reflection.GeneratedProtocolMessageType(
"Sample",
(_message.Message,),
{
"DESCRIPTOR": _SELFTESTDATA_SAMPLE,
"__module__": "sentencepiece_model_pb2",
# @@protoc_insertion_point(class_scope:sentencepiece.SelfTestData.Sample)
},
),
"DESCRIPTOR": _SELFTESTDATA,
"__module__": "sentencepiece_model_pb2",
# @@protoc_insertion_point(class_scope:sentencepiece.SelfTestData)
},
)
_sym_db.RegisterMessage(SelfTestData)
_sym_db.RegisterMessage(SelfTestData.Sample)
ModelProto = _reflection.GeneratedProtocolMessageType(
"ModelProto",
(_message.Message,),
{
"SentencePiece": _reflection.GeneratedProtocolMessageType(
"SentencePiece",
(_message.Message,),
{
"DESCRIPTOR": _MODELPROTO_SENTENCEPIECE,
"__module__": "sentencepiece_model_pb2",
# @@protoc_insertion_point(class_scope:sentencepiece.ModelProto.SentencePiece)
},
),
"DESCRIPTOR": _MODELPROTO,
"__module__": "sentencepiece_model_pb2",
# @@protoc_insertion_point(class_scope:sentencepiece.ModelProto)
},
)
_sym_db.RegisterMessage(ModelProto)
_sym_db.RegisterMessage(ModelProto.SentencePiece)
DESCRIPTOR._options = None
_TRAINERSPEC.fields_by_name["mining_sentence_size"]._options = None
_TRAINERSPEC.fields_by_name["training_sentence_size"]._options = None
# @@protoc_insertion_point(module_scope)
|
transformers/src/transformers/utils/sentencepiece_model_pb2.py/0
|
{
"file_path": "transformers/src/transformers/utils/sentencepiece_model_pb2.py",
"repo_id": "transformers",
"token_count": 28261
}
| 397
|
# coding=utf-8
# Copyright 2024 HuggingFace Inc.
#
# 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.
import os
import tempfile
import unittest
import uuid
import pytest
from transformers.agents.agent_types import AgentText
from transformers.agents.agents import AgentMaxIterationsError, CodeAgent, ReactCodeAgent, ReactJsonAgent, Toolbox
from transformers.agents.default_tools import PythonInterpreterTool
from transformers.testing_utils import require_torch
def get_new_path(suffix="") -> str:
directory = tempfile.mkdtemp()
return os.path.join(directory, str(uuid.uuid4()) + suffix)
def fake_react_json_llm(messages, stop_sequences=None, grammar=None) -> str:
prompt = str(messages)
if "special_marker" not in prompt:
return """
Thought: I should multiply 2 by 3.6452. special_marker
Action:
{
"action": "python_interpreter",
"action_input": {"code": "2*3.6452"}
}
"""
else: # We're at step 2
return """
Thought: I can now answer the initial question
Action:
{
"action": "final_answer",
"action_input": {"answer": "7.2904"}
}
"""
def fake_react_code_llm(messages, stop_sequences=None, grammar=None) -> str:
prompt = str(messages)
if "special_marker" not in prompt:
return """
Thought: I should multiply 2 by 3.6452. special_marker
Code:
```py
result = 2**3.6452
print(result)
```<end_code>
"""
else: # We're at step 2
return """
Thought: I can now answer the initial question
Code:
```py
final_answer(7.2904)
```<end_code>
"""
def fake_react_code_llm_error(messages, stop_sequences=None) -> str:
prompt = str(messages)
if "special_marker" not in prompt:
return """
Thought: I should multiply 2 by 3.6452. special_marker
Code:
```py
print = 2
```<end_code>
"""
else: # We're at step 2
return """
Thought: I can now answer the initial question
Code:
```py
final_answer("got an error")
```<end_code>
"""
def fake_react_code_functiondef(messages, stop_sequences=None) -> str:
prompt = str(messages)
if "special_marker" not in prompt:
return """
Thought: Let's define the function. special_marker
Code:
```py
import numpy as np
def moving_average(x, w):
return np.convolve(x, np.ones(w), 'valid') / w
```<end_code>
"""
else: # We're at step 2
return """
Thought: I can now answer the initial question
Code:
```py
x, w = [0, 1, 2, 3, 4, 5], 2
res = moving_average(x, w)
final_answer(res)
```<end_code>
"""
def fake_code_llm_oneshot(messages, stop_sequences=None, grammar=None) -> str:
return """
Thought: I should multiply 2 by 3.6452. special_marker
Code:
```py
result = python_interpreter(code="2*3.6452")
final_answer(result)
```
"""
def fake_code_llm_no_return(messages, stop_sequences=None, grammar=None) -> str:
return """
Thought: I should multiply 2 by 3.6452. special_marker
Code:
```py
result = python_interpreter(code="2*3.6452")
print(result)
```
"""
class AgentTests(unittest.TestCase):
def test_fake_code_agent(self):
agent = CodeAgent(tools=[PythonInterpreterTool()], llm_engine=fake_code_llm_oneshot)
output = agent.run("What is 2 multiplied by 3.6452?")
assert isinstance(output, str)
assert output == "7.2904"
def test_fake_react_json_agent(self):
agent = ReactJsonAgent(tools=[PythonInterpreterTool()], llm_engine=fake_react_json_llm)
output = agent.run("What is 2 multiplied by 3.6452?")
assert isinstance(output, str)
assert output == "7.2904"
assert agent.logs[0]["task"] == "What is 2 multiplied by 3.6452?"
assert agent.logs[1]["observation"] == "7.2904"
assert agent.logs[1]["rationale"].strip() == "Thought: I should multiply 2 by 3.6452. special_marker"
assert (
agent.logs[2]["llm_output"]
== """
Thought: I can now answer the initial question
Action:
{
"action": "final_answer",
"action_input": {"answer": "7.2904"}
}
"""
)
def test_fake_react_code_agent(self):
agent = ReactCodeAgent(tools=[PythonInterpreterTool()], llm_engine=fake_react_code_llm)
output = agent.run("What is 2 multiplied by 3.6452?")
assert isinstance(output, float)
assert output == 7.2904
assert agent.logs[0]["task"] == "What is 2 multiplied by 3.6452?"
assert float(agent.logs[1]["observation"].strip()) - 12.511648 < 1e-6
assert agent.logs[2]["tool_call"] == {
"tool_arguments": "final_answer(7.2904)",
"tool_name": "code interpreter",
}
def test_react_code_agent_code_errors_show_offending_lines(self):
agent = ReactCodeAgent(tools=[PythonInterpreterTool()], llm_engine=fake_react_code_llm_error)
output = agent.run("What is 2 multiplied by 3.6452?")
assert isinstance(output, AgentText)
assert output == "got an error"
assert "Evaluation stopped at line 'print = 2' because of" in str(agent.logs)
def test_setup_agent_with_empty_toolbox(self):
ReactJsonAgent(llm_engine=fake_react_json_llm, tools=[])
def test_react_fails_max_iterations(self):
agent = ReactCodeAgent(
tools=[PythonInterpreterTool()],
llm_engine=fake_code_llm_no_return, # use this callable because it never ends
max_iterations=5,
)
agent.run("What is 2 multiplied by 3.6452?")
assert len(agent.logs) == 7
assert type(agent.logs[-1]["error"]) is AgentMaxIterationsError
@require_torch
def test_init_agent_with_different_toolsets(self):
toolset_1 = []
agent = ReactCodeAgent(tools=toolset_1, llm_engine=fake_react_code_llm)
assert (
len(agent.toolbox.tools) == 1
) # when no tools are provided, only the final_answer tool is added by default
toolset_2 = [PythonInterpreterTool(), PythonInterpreterTool()]
agent = ReactCodeAgent(tools=toolset_2, llm_engine=fake_react_code_llm)
assert (
len(agent.toolbox.tools) == 2
) # deduplication of tools, so only one python_interpreter tool is added in addition to final_answer
toolset_3 = Toolbox(toolset_2)
agent = ReactCodeAgent(tools=toolset_3, llm_engine=fake_react_code_llm)
assert (
len(agent.toolbox.tools) == 2
) # same as previous one, where toolset_3 is an instantiation of previous one
# check that add_base_tools will not interfere with existing tools
with pytest.raises(KeyError) as e:
agent = ReactJsonAgent(tools=toolset_3, llm_engine=fake_react_json_llm, add_base_tools=True)
assert "already exists in the toolbox" in str(e)
# check that python_interpreter base tool does not get added to code agents
agent = ReactCodeAgent(tools=[], llm_engine=fake_react_code_llm, add_base_tools=True)
assert len(agent.toolbox.tools) == 6 # added final_answer tool + 5 base tools (excluding interpreter)
def test_function_persistence_across_steps(self):
agent = ReactCodeAgent(
tools=[], llm_engine=fake_react_code_functiondef, max_iterations=2, additional_authorized_imports=["numpy"]
)
res = agent.run("ok")
assert res[0] == 0.5
|
transformers/tests/agents/test_agents.py/0
|
{
"file_path": "transformers/tests/agents/test_agents.py",
"repo_id": "transformers",
"token_count": 3062
}
| 398
|
# Copyright 2020 The HuggingFace Team. 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.
import dataclasses
import io
import itertools
import json
import os
import unittest
from copy import deepcopy
from functools import partial
import datasets
from parameterized import parameterized
import tests.trainer.test_trainer
import transformers
from tests.trainer.test_trainer import TrainerIntegrationCommon # noqa
from transformers import AutoModel, TrainingArguments, is_torch_available, logging
from transformers.integrations.deepspeed import (
HfDeepSpeedConfig,
is_deepspeed_available,
unset_hf_deepspeed_config,
)
from transformers.testing_utils import (
CaptureLogger,
CaptureStd,
CaptureStderr,
LoggingLevel,
TestCasePlus,
backend_device_count,
execute_subprocess_async,
mockenv_context,
require_deepspeed,
require_optuna,
require_torch_accelerator,
require_torch_multi_accelerator,
slow,
torch_device,
)
from transformers.trainer_utils import get_last_checkpoint, set_seed
from transformers.utils import SAFE_WEIGHTS_NAME, is_torch_bf16_available_on_device
if is_torch_available():
import torch
from tests.trainer.test_trainer import ( # noqa
RegressionModelConfig,
RegressionPreTrainedModel,
)
# hack to restore original logging level pre #21700
get_regression_trainer = partial(tests.trainer.test_trainer.get_regression_trainer, log_level="info")
set_seed(42)
# default torch.distributed port
DEFAULT_MASTER_PORT = "10999"
T5_SMALL = "google-t5/t5-small"
T5_TINY = "patrickvonplaten/t5-tiny-random"
GPT2_TINY = "sshleifer/tiny-gpt2"
GPTJ_TINY = "hf-internal-testing/tiny-random-gptj"
def load_json(path):
with open(path) as f:
return json.load(f)
def get_master_port(real_launcher=False):
"""
When using a single gpu launcher emulation (i.e. not deepspeed or python -m torch.distributed)
the issue is that once the port is tied it can't be used anywhere else outside of this process,
since torch.dist doesn't free the port until the process exits. Therefore for the sake of being
able to run both emulated launcher and normal launcher tests we need 2 distinct ports.
This function will give the right port in the right context. For real launcher it'll give the
base port, for emulated launcher it'll give the base port + 1. In both cases a string is
returned.
Args:
`real_launcher`: whether a real launcher is going to be used, or the emulated one
"""
master_port_base = os.environ.get("DS_TEST_PORT", DEFAULT_MASTER_PORT)
if not real_launcher:
master_port_base = str(int(master_port_base) + 1)
return master_port_base
def require_deepspeed_aio(test_case):
"""
Decorator marking a test that requires deepspeed aio (nvme)
"""
if not is_deepspeed_available():
return unittest.skip(reason="test requires deepspeed")(test_case)
import deepspeed
from deepspeed.ops.aio import AsyncIOBuilder
if not deepspeed.ops.__compatible_ops__[AsyncIOBuilder.NAME]:
return unittest.skip(reason="test requires deepspeed async-io")(test_case)
else:
return test_case
if is_deepspeed_available():
from deepspeed.utils import logger as deepspeed_logger # noqa
from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
from transformers.integrations.deepspeed import deepspeed_config, is_deepspeed_zero3_enabled # noqa
def get_launcher(distributed=False):
# 1. explicitly set --num_nodes=1 just in case these tests end up run on a multi-node setup
# - it won't be able to handle that
# 2. for now testing with just 2 gpus max (since some quality tests may give different
# results with mode gpus because we use very little data)
num_gpus = min(2, backend_device_count(torch_device)) if distributed else 1
master_port = get_master_port(real_launcher=True)
return f"deepspeed --num_nodes 1 --num_gpus {num_gpus} --master_port {master_port}".split()
ZERO2 = "zero2"
ZERO3 = "zero3"
FP16 = "fp16"
BF16 = "bf16"
HF_OPTIM = "hf_optim"
HF_SCHEDULER = "hf_scheduler"
DS_OPTIM = "ds_optim"
DS_SCHEDULER = "ds_scheduler"
optims = [HF_OPTIM, DS_OPTIM]
schedulers = [HF_SCHEDULER, DS_SCHEDULER]
stages = [ZERO2, ZERO3]
if is_torch_bf16_available_on_device(torch_device):
dtypes = [FP16, BF16]
else:
dtypes = [FP16]
def parameterized_custom_name_func(func, param_num, param):
# customize the test name generator function as we want both params to appear in the sub-test
# name, as by default it shows only the first param
param_based_name = parameterized.to_safe_name("_".join(str(x) for x in param.args))
return f"{func.__name__}_{param_based_name}"
# Cartesian-product of zero stages with models to test
params = list(itertools.product(stages, dtypes))
params_with_optims_and_schedulers = list(itertools.product(stages, dtypes, optims, schedulers))
@require_deepspeed
@require_torch_accelerator
class CoreIntegrationDeepSpeed(TestCasePlus, TrainerIntegrationCommon):
"""
Testing non-Trainer DeepSpeed integration
"""
def setUp(self):
super().setUp()
master_port = get_master_port(real_launcher=False)
self.dist_env_1_gpu = {
"MASTER_ADDR": "localhost",
"MASTER_PORT": master_port,
"RANK": "0",
"LOCAL_RANK": "0",
"WORLD_SIZE": "1",
}
def tearDown(self):
super().tearDown()
# reset the ds config global so that tests state doesn't leak
unset_hf_deepspeed_config()
def test_init_zero3_fp16(self):
# test that zero.Init() works correctly under zero3/fp16
ds_config = {
"train_batch_size": 1,
"zero_optimization": {
"stage": 3,
},
}
dschf = HfDeepSpeedConfig(ds_config)
self.assertTrue(dschf.is_zero3())
self.assertTrue(is_deepspeed_zero3_enabled())
with LoggingLevel(logging.INFO):
with mockenv_context(**self.dist_env_1_gpu):
logger = logging.get_logger("transformers.modeling_utils")
with CaptureLogger(logger) as cl:
AutoModel.from_pretrained(T5_TINY)
self.assertIn("Detected DeepSpeed ZeRO-3", cl.out)
# now remove zero optimization
del ds_config["zero_optimization"]
dschf = HfDeepSpeedConfig(ds_config)
self.assertFalse(dschf.is_zero3())
self.assertFalse(is_deepspeed_zero3_enabled())
with LoggingLevel(logging.INFO):
with mockenv_context(**self.dist_env_1_gpu):
logger = logging.get_logger("transformers.modeling_utils")
with CaptureLogger(logger) as cl:
AutoModel.from_pretrained(T5_TINY)
self.assertNotIn("Detected DeepSpeed ZeRO-3", cl.out)
def test_init_zero3_missing_params(self):
# test that zero.Init() for missing parameters works correctly under zero3
import deepspeed
import torch
from transformers.models.gpt2.modeling_gpt2 import GPT2PreTrainedModel
class TinyGPT2WithUninitializedWeights(GPT2PreTrainedModel):
def __init__(self, config):
super().__init__(config)
self.transformer = AutoModel.from_pretrained(GPT2_TINY, config=config)
self.new_head = torch.nn.Linear(config.hidden_size, config.vocab_size, bias=True)
def forward(self, *args, **kwargs):
transformer_outputs = self.transformer(*args, **kwargs)
hidden_states = transformer_outputs[0]
return self.new_head(hidden_states).float()
def _init_weights(self, module):
super()._init_weights(module)
if module is self.new_head:
self.new_head.weight.data.fill_(-100.0)
self.new_head.bias.data.fill_(+100.0)
ds_config = {
"train_batch_size": 1,
"zero_optimization": {
"stage": 3,
},
}
dschf = HfDeepSpeedConfig(ds_config)
self.assertTrue(dschf.is_zero3())
self.assertTrue(is_deepspeed_zero3_enabled())
with LoggingLevel(logging.INFO):
with mockenv_context(**self.dist_env_1_gpu):
logger = logging.get_logger("transformers.modeling_utils")
with CaptureLogger(logger) as cl:
model = TinyGPT2WithUninitializedWeights.from_pretrained(GPT2_TINY)
self.assertIn("Detected DeepSpeed ZeRO-3", cl.out)
self.assertRegex(cl.out, r"newly initialized.*new_head\.bias.*new_head\.weight")
with deepspeed.zero.GatheredParameters([model.new_head.weight, model.new_head.bias]):
self.assertTrue(
torch.allclose(model.new_head.weight, torch.tensor(-100.0, device=model.new_head.weight.device)),
)
self.assertTrue(
torch.allclose(model.new_head.bias, torch.tensor(+100.0, device=model.new_head.bias.device)),
)
# now remove zero optimization
del ds_config["zero_optimization"]
dschf = HfDeepSpeedConfig(ds_config)
self.assertFalse(dschf.is_zero3())
self.assertFalse(is_deepspeed_zero3_enabled())
with LoggingLevel(logging.INFO):
with mockenv_context(**self.dist_env_1_gpu):
logger = logging.get_logger("transformers.modeling_utils")
with CaptureLogger(logger) as cl:
model = TinyGPT2WithUninitializedWeights.from_pretrained(GPT2_TINY)
self.assertNotIn("Detected DeepSpeed ZeRO-3", cl.out)
self.assertRegex(cl.out, r"newly initialized.*new_head\.bias.*new_head\.weight")
self.assertTrue(
torch.allclose(model.new_head.weight, torch.tensor(-100.0, device=model.new_head.weight.device)),
)
self.assertTrue(
torch.allclose(model.new_head.bias, torch.tensor(+100.0, device=model.new_head.bias.device)),
)
def test_arange_bf16(self):
# Tests that configuring DeepSpeed with 16 bits does not cause float `torch.arange()` tensors to be cast down.
# NOTE -- this assumes that the function calls have the following downcast-preventing pattern, i.e.
# `torch.arange(...,dtype=torch.int64)` followed by a cast like `.to(torch.float32)`. 🚨 If this pattern is
# NOT applied (e.g. `torch.arange(...,dtype=torch.float32)` is used), DeepSpeed can automatically cast it down
# at init time. See https://github.com/huggingface/transformers/issues/28685 for more info.
ds_config = {
"train_batch_size": 1,
"zero_optimization": {
"stage": 3,
},
"bf16": {"enabled": True},
}
dschf = HfDeepSpeedConfig(ds_config)
self.assertTrue(dschf.is_zero3())
self.assertTrue(is_deepspeed_zero3_enabled())
with LoggingLevel(logging.INFO):
with mockenv_context(**self.dist_env_1_gpu):
logger = logging.get_logger("transformers.modeling_utils")
with CaptureLogger(logger) as cl:
model = AutoModel.from_pretrained(GPTJ_TINY)
self.assertIn("Detected DeepSpeed ZeRO-3", cl.out)
# The model weights are in BF16 as per deepspeed config
self.assertTrue(str(model.h[0].attn.q_proj.weight.dtype) == "torch.bfloat16")
good_deepspeed_sin_cos = model.h[0].attn.embed_positions
# Monkeypatches the function that creates RoPE embeddings using the INCORRECT torch.arange() pattern, and
# then recreates the model
def bad_deepspeed_create_sinusoidal_positions(num_pos: int, dim: int) -> torch.Tensor:
inv_freq = 1.0 / (10000 ** (torch.arange(0, dim, 2, dtype=torch.int64) / dim))
# Incorrect pattern here: torch.arange has dtype=torch.float32 as its argument, and it will automatically
# converted to BF16 by DeepSpeed
sinusoid_inp = torch.einsum("i , j -> i j", torch.arange(num_pos, dtype=inv_freq.dtype), inv_freq)
return torch.cat((torch.sin(sinusoid_inp), torch.cos(sinusoid_inp)), dim=1)
good_deepspeed_create_sinusoidal_positions = transformers.models.gptj.modeling_gptj.create_sinusoidal_positions
transformers.models.gptj.modeling_gptj.create_sinusoidal_positions = bad_deepspeed_create_sinusoidal_positions
with LoggingLevel(logging.INFO):
with mockenv_context(**self.dist_env_1_gpu):
logger = logging.get_logger("transformers.modeling_utils")
with CaptureLogger(logger) as cl:
model = AutoModel.from_pretrained(GPTJ_TINY)
self.assertIn("Detected DeepSpeed ZeRO-3", cl.out)
self.assertTrue(str(model.h[0].attn.q_proj.weight.dtype) == "torch.bfloat16")
bad_deepspeed_sin_cos = model.h[0].attn.embed_positions
# Compares the two values: the two sets of values are different, and the correct one matches the torch
# (i.e. outside DeepSpeed) version.
good_torch_sin_cos = good_deepspeed_create_sinusoidal_positions(
model.config.max_position_embeddings, model.config.rotary_dim
)
self.assertFalse(torch.allclose(good_deepspeed_sin_cos, bad_deepspeed_sin_cos))
self.assertTrue(torch.allclose(good_torch_sin_cos, good_deepspeed_sin_cos.cpu()))
# Finally, we can see that the incorrect pattern is okay on vanilla torch, demostrating that this issue is
# exclusive to DeepSpeed
bad_torch_sin_cos = bad_deepspeed_create_sinusoidal_positions(
model.config.max_position_embeddings, model.config.rotary_dim
)
self.assertTrue(torch.allclose(bad_torch_sin_cos, good_torch_sin_cos))
class TrainerIntegrationDeepSpeedWithCustomConfig(TestCasePlus):
def setUp(self):
super().setUp()
args = TrainingArguments(".")
self.n_epochs = args.num_train_epochs
self.batch_size = args.train_batch_size
master_port = get_master_port(real_launcher=False)
self.dist_env_1_gpu = {
"MASTER_ADDR": "localhost",
"MASTER_PORT": master_port,
"RANK": "0",
"LOCAL_RANK": "0",
"WORLD_SIZE": "1",
}
self.ds_config_file = {
"zero2": f"{self.test_file_dir_str}/ds_config_zero2.json",
"zero3": f"{self.test_file_dir_str}/ds_config_zero3.json",
}
# use self.get_config_dict(stage) to use these to ensure the original is not modified
with io.open(self.ds_config_file[ZERO2], "r", encoding="utf-8") as f:
config_zero2 = json.load(f)
with io.open(self.ds_config_file[ZERO3], "r", encoding="utf-8") as f:
config_zero3 = json.load(f)
# The following setting slows things down, so don't enable it by default unless needed by a test.
# It's in the file as a demo for users since we want everything to work out of the box even if slower.
config_zero3["zero_optimization"]["stage3_gather_16bit_weights_on_model_save"] = False
self.ds_config_dict = {
"zero2": config_zero2,
"zero3": config_zero3,
}
def tearDown(self):
super().tearDown()
# reset the ds config global so that tests state doesn't leak
unset_hf_deepspeed_config()
def get_config_dict(self, stage):
# As some tests modify the dict, always make a copy
return deepcopy(self.ds_config_dict[stage])
@require_deepspeed
@require_torch_accelerator
class TrainerIntegrationDeepSpeed(TrainerIntegrationDeepSpeedWithCustomConfig, TrainerIntegrationCommon):
"""
This class is for testing directly via get_regression_trainer
It mixes in `TrainerIntegrationCommon` which already has a lot of helper validation methods
which we can re-use here.
Important: this class' setup can only work with a single gpu because it runs within the current
pytest worker. For multi-gpu tests use TestDeepSpeedWithLauncher.
Note: if any of the tests of this class get run there will be at least one gpu occupied by them
until this pytest worker exits. This is because the gpu memory allocated by the cuda-kernels
won't be released until this pytest worker exits.
This may appear as some run-away tests if you watch `nvidia-smi` while other tests that fork new
processes are run. So there will be one or two "stale" processes reported in `nvidia-smi`. This
is not a bug.
"""
# --- These tests are enough to run on one of zero stages --- #
def test_hf_ds_config_mismatch(self):
ds_config = self.get_config_dict(ZERO2)
# Purposefully configure these values to mismatch TrainingArguments values.
# This currently doesn't cover all keys (but it could)
per_device_train_batch_size = 2
ds_config["train_micro_batch_size_per_gpu"] = per_device_train_batch_size + 2
ds_config["train_batch_size"] = 1000
gradient_accumulation_steps = 2
ds_config["gradient_accumulation_steps"] = gradient_accumulation_steps + 2
max_grad_norm = 1.0
ds_config["gradient_clipping"] = max_grad_norm + 0.1
adam_beta1, adam_beta2 = 0.9, 0.99
ds_config["optimizer"]["params"]["betas"] = [adam_beta1 - 0.1, adam_beta2 - 0.1]
fp16 = True
ds_config["fp16"]["enabled"] = not fp16
keys = [
"per_device_train_batch_size",
"train_batch_size",
"gradient_accumulation_steps",
"max_grad_norm",
"betas",
"fp16",
]
with mockenv_context(**self.dist_env_1_gpu):
trainer = get_regression_trainer(
local_rank=0,
fp16=fp16,
deepspeed=ds_config,
per_device_train_batch_size=per_device_train_batch_size,
gradient_accumulation_steps=gradient_accumulation_steps,
max_grad_norm=max_grad_norm,
adam_beta1=adam_beta1,
adam_beta2=adam_beta2,
)
with self.assertRaises(Exception) as context:
trainer.train()
for key in keys:
self.assertTrue(
key in str(context.exception),
f"{key} is not in the exception message:\n{context.exception}",
)
# Test various combos
# 1. DS scheduler + DS optimizer: this is already tested by most other tests
# 2. HF scheduler + HF optimizer:
# 3. DS scheduler + HF optimizer:
# 4. HF scheduler + DS optimizer:
def test_hf_scheduler_hf_optimizer(self):
a = 0
with mockenv_context(**self.dist_env_1_gpu):
ds_config_zero2_dict = self.get_config_dict(ZERO2)
del ds_config_zero2_dict["optimizer"] # force default HF Trainer optimizer
del ds_config_zero2_dict["scheduler"] # force default HF Trainer scheduler
ds_config_zero2_dict["zero_optimization"]["offload_optimizer"]["device"] = "none"
ds_config_zero2_dict["fp16"]["initial_scale_power"] = 1 # force optimizer on the first step
trainer = get_regression_trainer(a=a, local_rank=0, fp16=True, deepspeed=ds_config_zero2_dict)
trainer.train()
new_a = trainer.model.a.item()
self.assertNotEqual(new_a, a)
def test_ds_scheduler_hf_optimizer(self):
a = 0
with mockenv_context(**self.dist_env_1_gpu):
ds_config_zero2_dict = self.get_config_dict(ZERO2)
del ds_config_zero2_dict["optimizer"] # force default HF Trainer optimizer
ds_config_zero2_dict["zero_optimization"]["offload_optimizer"]["device"] = "none"
ds_config_zero2_dict["fp16"]["initial_scale_power"] = 1 # force optimizer on the first step
trainer = get_regression_trainer(a=a, local_rank=0, fp16=True, deepspeed=ds_config_zero2_dict)
trainer.train()
new_a = trainer.model.a.item()
self.assertNotEqual(new_a, a)
def test_hf_scheduler_ds_optimizer(self):
a = 0
with mockenv_context(**self.dist_env_1_gpu):
ds_config_zero2_dict = self.get_config_dict(ZERO2)
del ds_config_zero2_dict["scheduler"] # force default HF Trainer scheduler
ds_config_zero2_dict["zero_optimization"]["offload_optimizer"]["device"] = "none"
ds_config_zero2_dict["fp16"]["initial_scale_power"] = 1 # force optimizer on the first step
trainer = get_regression_trainer(a=a, local_rank=0, fp16=True, deepspeed=ds_config_zero2_dict)
trainer.train()
new_a = trainer.model.a.item()
self.assertNotEqual(new_a, a)
@require_deepspeed_aio
def test_stage3_nvme_offload(self):
with mockenv_context(**self.dist_env_1_gpu):
# this actually doesn't have to be on NVMe, any storage will do since this test only
# runs a simple check that we can use some directory as if it were NVMe
nvme_path = self.get_auto_remove_tmp_dir()
nvme_config = {"device": "nvme", "nvme_path": nvme_path}
ds_config_zero3_dict = self.get_config_dict(ZERO3)
ds_config_zero3_dict["zero_optimization"]["offload_optimizer"] = nvme_config
ds_config_zero3_dict["zero_optimization"]["offload_param"] = nvme_config
ds_config_zero3_dict["zero_optimization"]["stage3_gather_16bit_weights_on_model_save"] = True
trainer = get_regression_trainer(local_rank=0, fp16=True, deepspeed=ds_config_zero3_dict)
with CaptureLogger(deepspeed_logger) as cl:
trainer.train()
self.assertIn("DeepSpeed info", cl.out, "expected DeepSpeed logger output but got none")
@require_optuna
def test_hyperparameter_search(self):
with mockenv_context(**self.dist_env_1_gpu):
ds_config_zero3_dict = self.get_config_dict(ZERO3)
# hyperparameter_search requires model_init() to recreate the model for each trial
def model_init():
config = RegressionModelConfig(a=0, b=0, double_output=False)
model = RegressionPreTrainedModel(config)
return model
trainer = get_regression_trainer(
local_rank=0,
fp16=True,
model_init=model_init,
deepspeed=ds_config_zero3_dict,
)
n_trials = 3
with CaptureLogger(deepspeed_logger) as cl:
with CaptureStd() as cs:
trainer.hyperparameter_search(direction="maximize", n_trials=n_trials)
self.assertIn("DeepSpeed info", cl.out, "expected DeepSpeed logger output but got none")
self.assertIn(f"Trial {n_trials-1} finished with value", cs.err, "expected hyperparameter_search output")
self.assertIn("Best is trial", cs.err, "expected hyperparameter_search output")
# --- These tests need to run on both zero stages --- #
@parameterized.expand(params, name_func=parameterized_custom_name_func)
def test_hf_optimizer_with_offload(self, stage, dtype):
# non-DS optimizers can be used with ZERO-offload (as long as they have both CPU and GPU implementation (except LAMB))
ds_config_dict = self.get_config_dict(stage)
del ds_config_dict["optimizer"] # force default HF Trainer optimizer
# force cpu offload
ds_config_dict["zero_optimization"]["offload_optimizer"]["device"] = "cpu"
ds_config_dict["zero_force_ds_cpu_optimizer"] = False # offload is not efficient w/o CPUAdam
with mockenv_context(**self.dist_env_1_gpu):
kwargs = {"local_rank": 0, "deepspeed": ds_config_dict}
kwargs[dtype] = True
trainer = get_regression_trainer(**kwargs)
with CaptureLogger(deepspeed_logger) as cl:
trainer.train()
self.assertIn("DeepSpeed info", cl.out, "expected DeepSpeed logger output but got none")
@parameterized.expand(params, name_func=parameterized_custom_name_func)
def test_fake_notebook_no_launcher(self, stage, dtype):
# this setup emulates a notebook where a launcher needs to be emulated by hand
# note that unittest resets sys.stdout each test, so `CaptureStd` will work here to capture
# DeepSpeed log if this test happens to run first in this pytest worker. But it will fail if
# it's run not as a first test as `sys.stdout` will no longer be the same. So we either have
# to reset `deepspeed_logger.handlers[0].setStream(sys.stdout)` or directly capture from the deepspeed_logger.
with mockenv_context(**self.dist_env_1_gpu):
kwargs = {"local_rank": 0, "deepspeed": self.get_config_dict(stage)}
kwargs[dtype] = True
trainer = get_regression_trainer(**kwargs)
with CaptureLogger(deepspeed_logger) as cl:
trainer.train()
self.assertIn("DeepSpeed info", cl.out, "expected DeepSpeed logger output but got none")
@parameterized.expand(params, name_func=parameterized_custom_name_func)
def test_early_get_last_lr(self, stage, dtype):
# with deepspeed's fp16 and dynamic loss scale enabled the optimizer/scheduler steps may
# not run for the first few dozen steps while loss scale is too large, and thus during
# that time `get_last_lr` will fail if called during that warm up stage,
#
# setting `logging_steps=1` forces an early `trainer._maybe_log_save_evaluate()` which calls
# `self.lr_scheduler.get_last_lr()` and originally it'd fail on the very first step.
with mockenv_context(**self.dist_env_1_gpu):
a = b = 0.0
kwargs = {
"a": a,
"b": b,
"local_rank": 0,
"train_len": 8,
"deepspeed": self.get_config_dict(stage),
"per_device_train_batch_size": 8,
"logging_steps": 1,
}
kwargs[dtype] = True
trainer = get_regression_trainer(**kwargs)
trainer.train()
post_train_a = trainer.model.a.item()
# XXX: for some reason the following check fails with zero3/fp16 and any/bf16 - not a
# broken but a different qualitative outcome - as if optimizer did run
# oddly getting 1.0 for both a and b from 0.0 - there is a bug somewhere
# print(trainer.model.a.item())
# print(trainer.model.b.item())
# need to investigate at some point
if (stage == ZERO3 and dtype == FP16) or (dtype == BF16):
self.skipTest(reason="When using zero3/fp16 or any/bf16 the optimizer seems run oddly")
# it's enough that train didn't fail for this test, but we must check that
# optimizer/scheduler didn't run (since if it did this test isn't testing the right thing)
self.assertEqual(post_train_a, a)
@parameterized.expand(params, name_func=parameterized_custom_name_func)
def test_gradient_accumulation(self, stage, dtype):
# this test measures that we get identical weights and similar loss with:
# 1. per_device_train_batch_size=8, gradient_accumulation_steps=1
# 2. per_device_train_batch_size=4, gradient_accumulation_steps=2
# since the 2nd should produce the effective batch of 1st, with the same results
#
# I can get an identical loss for a small train_len=32, plus the power of the initial
# dynamic loss scale value set to:
# "fp16.initial_scale_power": 1
# plus having the same WarmupLR's warmup_min_lr == warmup_max_lr in the config file
# but for some reason going to train_len=64 the weights, weights start to mismatch with this setup.
# the culprit seems to be `initial_scale_power` - putting it back to its default 32 keeps the weights identical
train_len = 64
a = b = 0.0
kwargs = {
"a": a,
"b": b,
"local_rank": 0,
"train_len": train_len,
"deepspeed": self.get_config_dict(stage),
}
kwargs[dtype] = True
with mockenv_context(**self.dist_env_1_gpu):
no_grad_accum_trainer = get_regression_trainer(
**kwargs,
per_device_train_batch_size=16,
gradient_accumulation_steps=1,
)
no_grad_accum_result = no_grad_accum_trainer.train()
no_grad_accum_loss = no_grad_accum_result.training_loss
no_grad_accum_a = no_grad_accum_trainer.model.a.item()
no_grad_accum_b = no_grad_accum_trainer.model.b.item()
# make sure the optimizer kicked in - if it hasn't changed from the original value of a then make train_len bigger
self.assertNotEqual(no_grad_accum_a, a)
with mockenv_context(**self.dist_env_1_gpu):
yes_grad_accum_trainer = get_regression_trainer(
**kwargs,
per_device_train_batch_size=4,
gradient_accumulation_steps=4,
)
yes_grad_accum_result = yes_grad_accum_trainer.train()
yes_grad_accum_loss = yes_grad_accum_result.training_loss
yes_grad_accum_a = yes_grad_accum_trainer.model.a.item()
yes_grad_accum_b = yes_grad_accum_trainer.model.b.item()
self.assertNotEqual(yes_grad_accum_a, a)
# training with half the batch size but accumulation steps as 2 should give the same
# weights, but sometimes get a slight difference still of 1e-6
self.assertAlmostEqual(no_grad_accum_a, yes_grad_accum_a, places=5)
self.assertAlmostEqual(no_grad_accum_b, yes_grad_accum_b, places=5)
# Relative difference. See the note above how to get identical loss on a small bs
self.assertTrue((no_grad_accum_loss - yes_grad_accum_loss) / (no_grad_accum_loss + 1e-15) <= 1e-3)
# NOTE: Currently a disabled test. In the future we should re-enable it.
# Issue resolves around Zero-3 w/ DPO/TRL + DeepSpeed
# As well as Zero-3 inference
# Related PR: https://github.com/huggingface/transformers/pull/32299
# def test_missed_zero3_init(self):
# from transformers import Trainer # noqa
# with mockenv_context(**self.dist_env_1_gpu):
# model = AutoModel.from_pretrained(T5_TINY)
# training_args = TrainingArguments(
# output_dir="./test_missed_zero3_init",
# deepspeed=self.get_config_dict(ZERO3),
# )
# with self.assertRaises(
# ValueError, msg="Model was not initialized with `Zero-3` despite being configured."
# ):
# _ = Trainer(
# model=model,
# args=training_args,
# )
# # Now do it properly, triggered from our `TrainingArguments` earlier
# model = AutoModel.from_pretrained(T5_TINY)
# trainer = Trainer(
# model=model,
# args=training_args,
# )
# assert trainer.is_deepspeed_enabled
# assert model._transformers_zero3_init_used
def check_saved_checkpoints_deepspeed(self, output_dir, freq, total, stage, dtype):
# adapted from TrainerIntegrationCommon.check_saved_checkpoints
file_list = [SAFE_WEIGHTS_NAME, "training_args.bin", "trainer_state.json", "config.json"]
if stage == ZERO2:
ds_file_list = ["mp_rank_00_model_states.pt"]
elif stage == ZERO3:
ds_file_list = ["zero_pp_rank_0_mp_rank_00_model_states.pt"]
else:
raise ValueError(f"unknown stage {stage}")
if dtype == "bf16":
ds_file_list.append("bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt")
for step in range(freq, total, freq):
checkpoint = os.path.join(output_dir, f"checkpoint-{step}")
self.assertTrue(os.path.isdir(checkpoint), f"[{stage}] {checkpoint} dir is not found")
# common files
for filename in file_list:
path = os.path.join(checkpoint, filename)
self.assertTrue(os.path.isfile(path), f"[{stage}] {path} is not found")
# ds files
ds_path = os.path.join(checkpoint, f"global_step{step}")
for filename in ds_file_list:
# filename = os.path.join(path, filename)
# print(filename)
path = os.path.join(ds_path, filename)
self.assertTrue(os.path.isfile(path), f"[{stage}] {path} is not found")
@parameterized.expand(params, name_func=parameterized_custom_name_func)
def test_save_checkpoints(self, stage, dtype):
# adapted from TrainerIntegrationTest.test_save_checkpoints
freq = 5
output_dir = self.get_auto_remove_tmp_dir()
ds_config_dict = self.get_config_dict(stage)
if dtype == FP16:
ds_config_dict["fp16"]["initial_scale_power"] = 1 # force optimizer on the first step
# XXX:
if stage == ZERO3:
ds_config_dict["zero_optimization"]["stage3_gather_16bit_weights_on_model_save"] = True
# save checkpoints
with mockenv_context(**self.dist_env_1_gpu):
kwargs = {
"output_dir": output_dir,
"save_steps": freq,
"deepspeed": ds_config_dict,
}
kwargs[dtype] = True
trainer = get_regression_trainer(**kwargs)
trainer.train()
total = int(self.n_epochs * 64 / self.batch_size)
self.check_saved_checkpoints_deepspeed(output_dir, freq, total, stage, dtype)
@parameterized.expand(params, name_func=parameterized_custom_name_func)
def test_can_resume_training_errors(self, stage, dtype):
with mockenv_context(**self.dist_env_1_gpu):
ds_config_dict = self.get_config_dict(stage)
output_dir = self.get_auto_remove_tmp_dir()
kwargs = {"output_dir": output_dir, "deepspeed": ds_config_dict}
kwargs[dtype] = True
trainer = get_regression_trainer(**kwargs)
# 1. fail to find any checkpoint - due a fresh output_dir
with self.assertRaises(Exception) as context:
trainer.train(resume_from_checkpoint=True)
self.assertTrue(
"No valid checkpoint found in output directory" in str(context.exception),
f"got exception: {context.exception}",
)
# 2. fail to find a bogus checkpoint
with self.assertRaises(Exception) as context:
checkpoint = os.path.join(output_dir, "checkpoint-5")
trainer.train(resume_from_checkpoint=f"{checkpoint}-bogus")
@parameterized.expand(params_with_optims_and_schedulers, name_func=parameterized_custom_name_func)
def test_can_resume_training_normal(self, stage, dtype, optim, scheduler):
# adapted from TrainerIntegrationTest.test_can_resume_training
# test normal resume for each stage separately, error-handling is tested in a different test
# ToDo: Currently, hf_optim + hf_scheduler resumes with the correct states and
# also has same losses for few steps but then slowly diverges. Need to figure it out.
if optim == HF_OPTIM and scheduler == HF_SCHEDULER:
self.skipTest(reason="hf_optim + hf_scheduler resumes with the correct states but slowly diverges")
output_dir = self.get_auto_remove_tmp_dir("./xxx", after=False)
ds_config_dict = self.get_config_dict(stage)
if dtype == FP16:
ds_config_dict["fp16"]["initial_scale_power"] = 1 # force optimizer on the first step
# XXX:
if stage == ZERO3:
ds_config_dict["zero_optimization"]["stage3_gather_16bit_weights_on_model_save"] = True
if optim == HF_OPTIM:
del ds_config_dict["optimizer"]
if scheduler == HF_SCHEDULER:
del ds_config_dict["scheduler"]
kwargs = {
"output_dir": output_dir,
"train_len": 128,
"save_steps": 5,
"learning_rate": 0.1,
"deepspeed": ds_config_dict,
}
kwargs[dtype] = True
with mockenv_context(**self.dist_env_1_gpu):
trainer = get_regression_trainer(**kwargs)
trainer.train()
(a, b) = trainer.model.a.item(), trainer.model.b.item()
state = dataclasses.asdict(trainer.state)
checkpoint = os.path.join(output_dir, "checkpoint-5")
# Reinitialize trainer
trainer = get_regression_trainer(**kwargs)
trainer.train(resume_from_checkpoint=checkpoint)
(a1, b1) = trainer.model.a.item(), trainer.model.b.item()
state1 = dataclasses.asdict(trainer.state)
self.assertEqual(a, a1)
self.assertEqual(b, b1)
self.check_trainer_state_are_the_same(state, state1)
# Now check with a later checkpoint that it also works when we span over one epoch
checkpoint = os.path.join(output_dir, "checkpoint-15")
# Reinitialize trainer and load model
trainer = get_regression_trainer(**kwargs)
trainer.train(resume_from_checkpoint=checkpoint)
(a1, b1) = trainer.model.a.item(), trainer.model.b.item()
state1 = dataclasses.asdict(trainer.state)
self.assertEqual(a, a1)
self.assertEqual(b, b1)
self.check_trainer_state_are_the_same(state, state1)
# Finally, should be able to resume with the same trainer/same deepspeed engine instance
# XXX: but currently this not possible due DS bug: https://github.com/microsoft/DeepSpeed/issues/1612
# trainer.train(resume_from_checkpoint=checkpoint)
# a workaround needs to be used that re-creates the deepspeed engine
@parameterized.expand(params, name_func=parameterized_custom_name_func)
def test_load_state_dict_from_zero_checkpoint(self, stage, dtype):
# test that we can load fp32 weights directly from the zero checkpoint into the current model
output_dir = self.get_auto_remove_tmp_dir() # "./xxx", after=False, before=False)
ds_config_dict = self.get_config_dict(stage)
kwargs = {
"output_dir": output_dir,
"train_len": 4,
"per_device_train_batch_size": 4,
"num_train_epochs": 1,
"save_strategy": "steps",
"save_steps": 1,
"learning_rate": 0.1,
"deepspeed": ds_config_dict,
}
kwargs[dtype] = True
with mockenv_context(**self.dist_env_1_gpu):
trainer = get_regression_trainer(**kwargs)
trainer.train()
(a, b) = trainer.model.a.item(), trainer.model.b.item()
state = dataclasses.asdict(trainer.state)
checkpoint_dir = get_last_checkpoint(output_dir)
model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
(a1, b1) = model.a.item(), model.b.item()
state1 = dataclasses.asdict(trainer.state)
self.assertEqual(a, a1)
self.assertEqual(b, b1)
self.check_trainer_state_are_the_same(state, state1)
def test_ds_config_object(self):
# test that we can switch from zero2 to zero3 in the same process for example
# test is_zero, etc.
output_dir = self.get_auto_remove_tmp_dir()
kwargs = {"output_dir": output_dir, "train_len": 8, "fp16": True}
ds_config_zero3_dict = self.get_config_dict(ZERO3)
ds_config_zero2_dict = self.get_config_dict(ZERO2)
with mockenv_context(**self.dist_env_1_gpu):
trainer = get_regression_trainer(deepspeed=ds_config_zero3_dict, **kwargs)
self.assertTrue(is_deepspeed_zero3_enabled())
# test we can repeat that and with train this time
trainer = get_regression_trainer(deepspeed=ds_config_zero3_dict, **kwargs)
trainer.train()
self.assertTrue(is_deepspeed_zero3_enabled())
# test zero3 is disabled
trainer = get_regression_trainer(deepspeed=ds_config_zero2_dict, **kwargs)
self.assertFalse(is_deepspeed_zero3_enabled())
# check config obj
config = deepspeed_config()
self.assertTrue(bool(config), "Deepspeed config should be accessible")
# with accelerate integration below line is additionally required for this test to pass
trainer.accelerator.state._reset_state()
del trainer
# now weakref should gc the global and we shouldn't get anything here
config = deepspeed_config()
self.assertFalse(is_deepspeed_zero3_enabled())
self.assertFalse(bool(config), "Deepspeed config should not be accessible")
@parameterized.expand(params, name_func=parameterized_custom_name_func)
def test_load_best_model(self, stage, dtype):
# Test that forced deepspeed reinit doesn't break the model. the forced re-init after
# loading the best model in Trainer is there to workaround this bug in Deepspeed
# https://github.com/microsoft/DeepSpeed/issues/1612
#
# The test is derived from a repro script submitted in this Issue:
# https://github.com/huggingface/transformers/issues/17114
#
# One additional feature of this test is that we use a non-AdamW optimizer to test that
# deepspeed doesn't fallback to AdamW, which would prevent the optimizer states from loading
# correctly
from transformers import T5ForConditionalGeneration, T5Tokenizer, Trainer # noqa
output_dir = self.get_auto_remove_tmp_dir() # "./xxx", after=False, before=False)
ds_config_dict = self.get_config_dict(stage)
del ds_config_dict["optimizer"] # will use HF Trainer optimizer
del ds_config_dict["scheduler"] # will use HF Trainer scheduler
ds_config_dict["zero_force_ds_cpu_optimizer"] = False # offload is not efficient w/o CPUAdam
# must use this setting to get the reload path exercised
ds_config_dict["zero_optimization"]["stage3_gather_16bit_weights_on_model_save"] = True
with mockenv_context(**self.dist_env_1_gpu):
args_dict = {
"per_device_train_batch_size": 1,
"per_device_eval_batch_size": 1,
"gradient_accumulation_steps": 1,
"learning_rate": 1e-4,
"num_train_epochs": 1,
"do_train": True,
"do_eval": True,
"optim": "adafactor",
"eval_strategy": "steps",
"eval_steps": 1,
"save_strategy": "steps",
"save_steps": 1,
"load_best_model_at_end": True,
"max_steps": 1,
"deepspeed": ds_config_dict,
"report_to": "none",
}
training_args = TrainingArguments(output_dir, **args_dict)
tokenizer = T5Tokenizer.from_pretrained(T5_TINY)
model = T5ForConditionalGeneration.from_pretrained(T5_TINY)
def _add_eos_to_examples(example):
example["input_text"] = f"question: {example['question']} context: {example['context']}"
example["target_text"] = example["answers"]["text"][0] if len(example["answers"]["text"]) > 0 else ""
return example
def _convert_to_features(example_batch):
input_encodings = tokenizer.batch_encode_plus(
example_batch["input_text"], pad_to_max_length=True, max_length=512, truncation=True
)
target_encodings = tokenizer.batch_encode_plus(
example_batch["target_text"], pad_to_max_length=True, max_length=16, truncation=True
)
encodings = {
"input_ids": input_encodings["input_ids"],
"attention_mask": input_encodings["attention_mask"],
"labels": target_encodings["input_ids"],
}
return encodings
def get_dataset():
data_file = str(self.tests_dir / "fixtures/tests_samples/SQUAD/sample.json")
data_files = {"train": data_file, "validation": data_file}
raw_datasets = datasets.load_dataset("json", data_files=data_files, field="data")
train_dataset = raw_datasets["train"].map(_add_eos_to_examples).map(_convert_to_features, batched=True)
valid_dataset = deepcopy(train_dataset)
return train_dataset, valid_dataset
train_dataset, eval_dataset = get_dataset()
trainer = Trainer(
model=model,
tokenizer=tokenizer,
args=training_args,
train_dataset=train_dataset,
eval_dataset=eval_dataset,
)
trainer.train() # crash 1 was here
trainer.evaluate() # crash 2 was here
@slow
@require_deepspeed
@require_torch_accelerator
class TestDeepSpeedWithLauncher(TestCasePlus):
"""This class is for testing via an external script - can do multiple gpus"""
# Tests to devise #
#
# 1. predict_with_generate on multigpu - need to figure out how to give input sequences so that
# the 2 gpus will generate prediction sequences that aren't of the same length - this is because
# we had to code a special feature to sync the gpus when the predicted sequences aren't of the
# same length. In general this will tested as a side-effect through a variety of other tests -
# it'll simply hang trying to synchronize with other gpus if this problem is encountered. So as
# long as we have a few full tests running on zero3 + predict_with_generate this should be
# mostly covered.
#
# but there are 5 variations on beam search in `generate`- with identical code branched with `if
# synced_gpus`
#
# 2. most tests should probably be run on both: zero2 and zero3 configs
#
@parameterized.expand(params, name_func=parameterized_custom_name_func)
@require_torch_multi_accelerator
def test_basic_distributed(self, stage, dtype):
self.run_and_check(stage=stage, dtype=dtype, distributed=True)
def test_do_eval_no_train(self):
# testing only zero3 since zero2 makes no sense with inference
self.run_and_check(
stage=ZERO3,
dtype=FP16,
eval_steps=1,
distributed=False,
do_train=False,
do_eval=True,
)
@parameterized.expand(params, name_func=parameterized_custom_name_func)
def test_fp32_non_distributed(self, stage, dtype):
# real model needs too much GPU memory under stage2+fp32, so using tiny random model here -
# therefore no quality checks, just basic completion checks are done
self.run_and_check(
stage=stage,
dtype=dtype,
model_name=T5_TINY,
distributed=False,
do_train=True,
do_eval=True,
quality_checks=False,
fp32=True,
)
@parameterized.expand(params, name_func=parameterized_custom_name_func)
@require_torch_multi_accelerator
def test_fp32_distributed(self, stage, dtype):
# real model needs too much GPU memory under stage2+fp32, so using tiny random model here -
# therefore no quality checks, just basic completion checks are done
self.run_and_check(
stage=stage,
dtype=dtype,
model_name=T5_TINY,
distributed=True,
do_train=True,
do_eval=True,
quality_checks=False,
fp32=True,
)
@parameterized.expand(params, name_func=parameterized_custom_name_func)
def test_resume_train_not_from_ds_checkpoint(self, stage, dtype):
# do normal training and then resume not from the deepspeed checkpoint but explicitly from
# the saved model dir
do_train = True
do_eval = False
kwargs = {
"stage": stage,
"dtype": dtype,
"eval_steps": 1,
"distributed": True,
"do_train": do_train,
"do_eval": do_eval,
}
# 1. normal training
output_dir = self.run_and_check(**kwargs)
# 2. now resume explicitly from the saved weights, by passing --model_name_or_path output_dir
# - i.e. the same path the model was saved to in step 1
output_dir = self.run_trainer(**kwargs, model_name=output_dir)
self.do_checks(output_dir, do_train=do_train, do_eval=do_eval)
@parameterized.expand(["bf16", "fp16", "fp32"])
@require_torch_multi_accelerator
def test_inference(self, dtype):
if dtype == "bf16" and not is_torch_bf16_available_on_device(torch_device):
self.skipTest(reason="test requires bfloat16 hardware support")
# this is just inference, so no optimizer should be loaded
# it only works for z3 (makes no sense with z1-z2)
fp32 = True if dtype == "fp32" else False
self.run_and_check(
stage=ZERO3,
dtype=FP16,
model_name=T5_TINY,
distributed=True,
do_train=False,
do_eval=True,
quality_checks=False,
fp32=fp32,
)
def do_checks(self, output_dir, do_train=True, do_eval=True, quality_checks=True):
if do_train:
train_metrics = load_json(os.path.join(output_dir, "train_results.json"))
self.assertIn("train_samples_per_second", train_metrics)
if quality_checks:
self.assertGreater(train_metrics["train_samples_per_second"], 0.5)
if do_eval:
eval_metrics = load_json(os.path.join(output_dir, "eval_results.json"))
self.assertIn("eval_bleu", eval_metrics)
if quality_checks:
self.assertGreater(eval_metrics["eval_bleu"], 1)
# XXX: need to do better validation beyond just that the run was successful
def run_and_check(
self,
stage,
dtype,
model_name: str = T5_SMALL,
eval_steps: int = 10,
distributed: bool = True,
do_train: bool = True,
do_eval: bool = True,
quality_checks: bool = True,
fp32: bool = False,
extra_args_str: str = None,
remove_args_str: str = None,
):
# we are doing quality testing so using a small real model
output_dir = self.run_trainer(
stage=stage,
dtype=dtype,
model_name=model_name,
eval_steps=eval_steps,
num_train_epochs=1,
do_train=do_train,
do_eval=do_eval,
distributed=distributed,
fp32=fp32,
extra_args_str=extra_args_str,
remove_args_str=remove_args_str,
)
self.do_checks(output_dir, do_train=do_train, do_eval=do_eval, quality_checks=quality_checks)
return output_dir
def run_trainer(
self,
stage: str,
dtype: str,
model_name: str,
eval_steps: int = 10,
num_train_epochs: int = 1,
do_train: bool = False,
do_eval: bool = True,
distributed: bool = True,
fp32: bool = False,
extra_args_str: str = None,
remove_args_str: str = None,
):
max_len = 32
data_dir = self.test_file_dir / "../fixtures/tests_samples/wmt_en_ro"
output_dir = self.get_auto_remove_tmp_dir()
args = f"""
--model_name_or_path {model_name}
--train_file {data_dir}/train.json
--validation_file {data_dir}/val.json
--output_dir {output_dir}
--overwrite_output_dir
--max_source_length {max_len}
--max_target_length {max_len}
--val_max_target_length {max_len}
--warmup_steps 8
--predict_with_generate
--save_steps 0
--eval_steps {eval_steps}
--group_by_length
--label_smoothing_factor 0.1
--source_lang en
--target_lang ro
--report_to none
""".split()
args.extend(["--source_prefix", '"translate English to Romanian: "'])
if not fp32:
args.extend([f"--{dtype}"])
actions = 0
if do_train:
actions += 1
args.extend(
f"""
--do_train
--num_train_epochs {str(num_train_epochs)}
--max_train_samples 16
--per_device_train_batch_size 2
--learning_rate 3e-3
""".split()
)
if do_eval:
actions += 1
args.extend(
"""
--do_eval
--max_eval_samples 16
--per_device_eval_batch_size 2
""".split()
)
assert actions > 0, "need at least do_train or do_eval for the test to run"
if extra_args_str is not None:
args.extend(extra_args_str.split())
# currently only works for bool args
if remove_args_str is not None:
remove_args = remove_args_str.split()
args = [x for x in args if x not in remove_args]
ds_args = f"--deepspeed {self.test_file_dir_str}/ds_config_{stage}.json".split()
script = [f"{self.examples_dir_str}/pytorch/translation/run_translation.py"]
launcher = get_launcher(distributed)
cmd = launcher + script + args + ds_args
# keep for quick debug
# print(" ".join([f"\nPYTHONPATH={self.src_dir_str}"] +cmd)); die
execute_subprocess_async(cmd, env=self.get_env())
return output_dir
@parameterized.expand(params, name_func=parameterized_custom_name_func)
def test_clm(self, stage, dtype):
# this test exercises model.resize_token_embeddings() which requires param gathering outside
# of forward - it's not used by `run_translation.py`, but it is in `run_clm.py`
data_dir = self.tests_dir / "fixtures"
output_dir = self.get_auto_remove_tmp_dir()
args = f"""
--model_name_or_path {GPT2_TINY}
--train_file {data_dir}/sample_text.txt
--validation_file {data_dir}/sample_text.txt
--output_dir {output_dir}
--overwrite_output_dir
--do_train
--do_eval
--max_train_samples 16
--max_eval_samples 16
--per_device_train_batch_size 2
--per_device_eval_batch_size 2
--num_train_epochs 1
--warmup_steps 8
--block_size 64
--report_to none
""".split()
args.extend([f"--{dtype}"])
ds_args = f"--deepspeed {self.test_file_dir_str}/ds_config_{stage}.json".split()
script = [f"{self.examples_dir_str}/pytorch/language-modeling/run_clm.py"]
launcher = get_launcher(distributed=True)
cmd = launcher + script + args + ds_args
# keep for quick debug
# print(" ".join([f"\nPYTHONPATH={self.src_dir_str}"] +cmd)); die
execute_subprocess_async(cmd, env=self.get_env())
def test_clm_from_config_zero3_fp16(self):
# this test exercises AutoModel.from_config(config) - to ensure zero.Init is called
data_dir = self.tests_dir / "fixtures"
output_dir = self.get_auto_remove_tmp_dir()
args = f"""
--model_type gpt2
--tokenizer_name {GPT2_TINY}
--train_file {data_dir}/sample_text.txt
--validation_file {data_dir}/sample_text.txt
--output_dir {output_dir}
--overwrite_output_dir
--do_train
--max_train_samples 4
--per_device_train_batch_size 2
--num_train_epochs 1
--warmup_steps 8
--block_size 8
--fp16
--report_to none
""".split()
ds_args = f"--deepspeed {self.test_file_dir_str}/ds_config_zero3.json".split()
script = [f"{self.examples_dir_str}/pytorch/language-modeling/run_clm.py"]
launcher = get_launcher(distributed=True)
cmd = launcher + script + args + ds_args
# keep for quick debug
# print(" ".join([f"\nPYTHONPATH={self.src_dir_str}"] +cmd)); die
with CaptureStderr() as cs:
execute_subprocess_async(cmd, env=self.get_env())
self.assertIn("Detected DeepSpeed ZeRO-3", cs.err)
|
transformers/tests/deepspeed/test_deepspeed.py/0
|
{
"file_path": "transformers/tests/deepspeed/test_deepspeed.py",
"repo_id": "transformers",
"token_count": 26044
}
| 399
|
# coding=utf-8
# Copyright 2020 The HuggingFace Team Inc.
#
# 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 clone 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.
import unittest
from transformers import is_torch_available
from transformers.testing_utils import require_torch
if is_torch_available():
import torch
from transformers.generation import DisjunctiveConstraint
@require_torch
class ConstraintTest(unittest.TestCase):
def test_input_types(self):
# For consistency across different places the DisjunctiveConstraint is called,
# dc.token_ids is a list of integers. It is also initialized only by integers.
cset = [[1, 2, 4], [1, 2, 3, 4]]
dc = DisjunctiveConstraint(cset)
self.assertTrue(isinstance(dc.token_ids, list))
with self.assertRaises(ValueError):
DisjunctiveConstraint(torch.LongTensor([[1, 2, 4], [1, 2, 3]]))
with self.assertRaises(ValueError):
DisjunctiveConstraint([torch.LongTensor([1, 2, 4]), torch.LongTensor([1, 2, 3, 4, 5])])
def test_check_illegal_input(self):
# We can't have constraints that are complete subsets of another. This leads to a preverse
# interpretation of "constraint fulfillment": does generating [1,2,3] fulfill the constraint?
# It would mean that it generated [1,2] which fulfills it, but it's in the middle of potentially
# fulfilling [1,2,3,4]. If we believe that [1,2,3] does fulfill the constraint, then the algorithm
# will necessarily never reach [1,2,3,4], giving users a false sense of control (better to just not allow it).
cset = [[1, 2], [1, 2, 3, 4]]
with self.assertRaises(ValueError):
DisjunctiveConstraint(cset) # fails here
def test_example_progression(self):
cset = [[1, 2, 3], [1, 2, 4]]
dc = DisjunctiveConstraint(cset)
stepped, completed, reset = dc.update(1)
desired = stepped is True and completed is False and reset is False
self.assertTrue(desired)
self.assertTrue(not dc.completed)
self.assertTrue(dc.current_seq == [1])
stepped, completed, reset = dc.update(2)
desired = stepped is True and completed is False and reset is False
self.assertTrue(desired)
self.assertTrue(not dc.completed)
self.assertTrue(dc.current_seq == [1, 2])
stepped, completed, reset = dc.update(3)
desired = stepped is True and completed is True and reset is False
self.assertTrue(desired)
self.assertTrue(dc.completed) # Completed!
self.assertTrue(dc.current_seq == [1, 2, 3])
def test_example_progression_unequal_three_mid_and_reset(self):
cset = [[1, 2, 3], [1, 2, 4, 5], [1, 2, 5]]
dc = DisjunctiveConstraint(cset)
stepped, completed, reset = dc.update(1)
self.assertTrue(not dc.completed)
self.assertTrue(dc.current_seq == [1])
stepped, completed, reset = dc.update(2)
self.assertTrue(not dc.completed)
self.assertTrue(dc.current_seq == [1, 2])
stepped, completed, reset = dc.update(4)
self.assertTrue(not dc.completed)
self.assertTrue(dc.current_seq == [1, 2, 4])
stepped, completed, reset = dc.update(5)
self.assertTrue(dc.completed) # Completed!
self.assertTrue(dc.current_seq == [1, 2, 4, 5])
dc.reset()
stepped, completed, reset = dc.update(1)
self.assertTrue(not dc.completed)
self.assertTrue(dc.remaining() == 3)
self.assertTrue(dc.current_seq == [1])
stepped, completed, reset = dc.update(2)
self.assertTrue(not dc.completed)
self.assertTrue(dc.remaining() == 2)
self.assertTrue(dc.current_seq == [1, 2])
stepped, completed, reset = dc.update(5)
self.assertTrue(dc.completed) # Completed!
self.assertTrue(dc.remaining() == 0)
self.assertTrue(dc.current_seq == [1, 2, 5])
|
transformers/tests/generation/test_beam_constraints.py/0
|
{
"file_path": "transformers/tests/generation/test_beam_constraints.py",
"repo_id": "transformers",
"token_count": 1723
}
| 400
|
# coding=utf-8
# Copyright 2020 The HuggingFace Team. 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.
from __future__ import annotations
import copy
import tempfile
import unittest
from transformers import CONFIG_MAPPING, AutoConfig, BertConfig, GPT2Config, T5Config, TapasConfig, is_tf_available
from transformers.testing_utils import (
DUMMY_UNKNOWN_IDENTIFIER,
SMALL_MODEL_IDENTIFIER,
RequestCounter,
require_tensorflow_probability,
require_tf,
slow,
)
from ..bert.test_modeling_bert import BertModelTester
if is_tf_available():
from transformers import (
TFAutoModel,
TFAutoModelForCausalLM,
TFAutoModelForMaskedLM,
TFAutoModelForPreTraining,
TFAutoModelForQuestionAnswering,
TFAutoModelForSeq2SeqLM,
TFAutoModelForSequenceClassification,
TFAutoModelForTableQuestionAnswering,
TFAutoModelForTokenClassification,
TFAutoModelWithLMHead,
TFBertForMaskedLM,
TFBertForPreTraining,
TFBertForQuestionAnswering,
TFBertForSequenceClassification,
TFBertModel,
TFFunnelBaseModel,
TFFunnelModel,
TFGPT2LMHeadModel,
TFRobertaForMaskedLM,
TFT5ForConditionalGeneration,
TFTapasForQuestionAnswering,
)
from transformers.models.auto.modeling_tf_auto import (
TF_MODEL_FOR_CAUSAL_LM_MAPPING,
TF_MODEL_FOR_MASKED_LM_MAPPING,
TF_MODEL_FOR_PRETRAINING_MAPPING,
TF_MODEL_FOR_QUESTION_ANSWERING_MAPPING,
TF_MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING,
TF_MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING,
TF_MODEL_MAPPING,
)
class NewModelConfig(BertConfig):
model_type = "new-model"
if is_tf_available():
class TFNewModel(TFBertModel):
config_class = NewModelConfig
@require_tf
class TFAutoModelTest(unittest.TestCase):
@slow
def test_model_from_pretrained(self):
model_name = "google-bert/bert-base-cased"
config = AutoConfig.from_pretrained(model_name)
self.assertIsNotNone(config)
self.assertIsInstance(config, BertConfig)
model = TFAutoModel.from_pretrained(model_name)
self.assertIsNotNone(model)
self.assertIsInstance(model, TFBertModel)
@slow
def test_model_for_pretraining_from_pretrained(self):
model_name = "google-bert/bert-base-cased"
config = AutoConfig.from_pretrained(model_name)
self.assertIsNotNone(config)
self.assertIsInstance(config, BertConfig)
model = TFAutoModelForPreTraining.from_pretrained(model_name)
self.assertIsNotNone(model)
self.assertIsInstance(model, TFBertForPreTraining)
@slow
def test_model_for_causal_lm(self):
model_name = "openai-community/gpt2"
config = AutoConfig.from_pretrained(model_name)
self.assertIsNotNone(config)
self.assertIsInstance(config, GPT2Config)
model = TFAutoModelForCausalLM.from_pretrained(model_name)
model, loading_info = TFAutoModelForCausalLM.from_pretrained(model_name, output_loading_info=True)
self.assertIsNotNone(model)
self.assertIsInstance(model, TFGPT2LMHeadModel)
@slow
def test_lmhead_model_from_pretrained(self):
model_name = "openai-community/gpt2"
config = AutoConfig.from_pretrained(model_name)
self.assertIsNotNone(config)
self.assertIsInstance(config, GPT2Config)
model = TFAutoModelWithLMHead.from_pretrained(model_name)
self.assertIsNotNone(model)
self.assertIsInstance(model, TFGPT2LMHeadModel)
@slow
def test_model_for_masked_lm(self):
model_name = "google-bert/bert-base-uncased"
config = AutoConfig.from_pretrained(model_name)
self.assertIsNotNone(config)
self.assertIsInstance(config, BertConfig)
model = TFAutoModelForMaskedLM.from_pretrained(model_name)
model, loading_info = TFAutoModelForMaskedLM.from_pretrained(model_name, output_loading_info=True)
self.assertIsNotNone(model)
self.assertIsInstance(model, TFBertForMaskedLM)
@slow
def test_model_for_encoder_decoder_lm(self):
model_name = "google-t5/t5-base"
config = AutoConfig.from_pretrained(model_name)
self.assertIsNotNone(config)
self.assertIsInstance(config, T5Config)
model = TFAutoModelForSeq2SeqLM.from_pretrained(model_name)
model, loading_info = TFAutoModelForSeq2SeqLM.from_pretrained(model_name, output_loading_info=True)
self.assertIsNotNone(model)
self.assertIsInstance(model, TFT5ForConditionalGeneration)
@slow
def test_sequence_classification_model_from_pretrained(self):
# model_name = 'openai-community/gpt2'
for model_name in ["google-bert/bert-base-uncased"]:
config = AutoConfig.from_pretrained(model_name)
self.assertIsNotNone(config)
self.assertIsInstance(config, BertConfig)
model = TFAutoModelForSequenceClassification.from_pretrained(model_name)
self.assertIsNotNone(model)
self.assertIsInstance(model, TFBertForSequenceClassification)
@slow
def test_question_answering_model_from_pretrained(self):
# model_name = 'openai-community/gpt2'
for model_name in ["google-bert/bert-base-uncased"]:
config = AutoConfig.from_pretrained(model_name)
self.assertIsNotNone(config)
self.assertIsInstance(config, BertConfig)
model = TFAutoModelForQuestionAnswering.from_pretrained(model_name)
self.assertIsNotNone(model)
self.assertIsInstance(model, TFBertForQuestionAnswering)
@slow
@require_tensorflow_probability
def test_table_question_answering_model_from_pretrained(self):
model_name = "google/tapas-base"
config = AutoConfig.from_pretrained(model_name)
self.assertIsNotNone(config)
self.assertIsInstance(config, TapasConfig)
model = TFAutoModelForTableQuestionAnswering.from_pretrained(model_name)
model, loading_info = TFAutoModelForTableQuestionAnswering.from_pretrained(
model_name, output_loading_info=True
)
self.assertIsNotNone(model)
self.assertIsInstance(model, TFTapasForQuestionAnswering)
def test_from_pretrained_identifier(self):
model = TFAutoModelWithLMHead.from_pretrained(SMALL_MODEL_IDENTIFIER)
self.assertIsInstance(model, TFBertForMaskedLM)
self.assertEqual(model.num_parameters(), 14410)
self.assertEqual(model.num_parameters(only_trainable=True), 14410)
def test_from_identifier_from_model_type(self):
model = TFAutoModelWithLMHead.from_pretrained(DUMMY_UNKNOWN_IDENTIFIER)
self.assertIsInstance(model, TFRobertaForMaskedLM)
self.assertEqual(model.num_parameters(), 14410)
self.assertEqual(model.num_parameters(only_trainable=True), 14410)
def test_from_pretrained_with_tuple_values(self):
# For the auto model mapping, FunnelConfig has two models: FunnelModel and FunnelBaseModel
model = TFAutoModel.from_pretrained("sgugger/funnel-random-tiny")
self.assertIsInstance(model, TFFunnelModel)
config = copy.deepcopy(model.config)
config.architectures = ["FunnelBaseModel"]
model = TFAutoModel.from_config(config)
model.build_in_name_scope()
self.assertIsInstance(model, TFFunnelBaseModel)
with tempfile.TemporaryDirectory() as tmp_dir:
model.save_pretrained(tmp_dir)
model = TFAutoModel.from_pretrained(tmp_dir)
self.assertIsInstance(model, TFFunnelBaseModel)
def test_new_model_registration(self):
try:
AutoConfig.register("new-model", NewModelConfig)
auto_classes = [
TFAutoModel,
TFAutoModelForCausalLM,
TFAutoModelForMaskedLM,
TFAutoModelForPreTraining,
TFAutoModelForQuestionAnswering,
TFAutoModelForSequenceClassification,
TFAutoModelForTokenClassification,
]
for auto_class in auto_classes:
with self.subTest(auto_class.__name__):
# Wrong config class will raise an error
with self.assertRaises(ValueError):
auto_class.register(BertConfig, TFNewModel)
auto_class.register(NewModelConfig, TFNewModel)
# Trying to register something existing in the Transformers library will raise an error
with self.assertRaises(ValueError):
auto_class.register(BertConfig, TFBertModel)
# Now that the config is registered, it can be used as any other config with the auto-API
tiny_config = BertModelTester(self).get_config()
config = NewModelConfig(**tiny_config.to_dict())
model = auto_class.from_config(config)
model.build_in_name_scope()
self.assertIsInstance(model, TFNewModel)
with tempfile.TemporaryDirectory() as tmp_dir:
model.save_pretrained(tmp_dir)
new_model = auto_class.from_pretrained(tmp_dir)
self.assertIsInstance(new_model, TFNewModel)
finally:
if "new-model" in CONFIG_MAPPING._extra_content:
del CONFIG_MAPPING._extra_content["new-model"]
for mapping in (
TF_MODEL_MAPPING,
TF_MODEL_FOR_PRETRAINING_MAPPING,
TF_MODEL_FOR_QUESTION_ANSWERING_MAPPING,
TF_MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING,
TF_MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING,
TF_MODEL_FOR_CAUSAL_LM_MAPPING,
TF_MODEL_FOR_MASKED_LM_MAPPING,
):
if NewModelConfig in mapping._extra_content:
del mapping._extra_content[NewModelConfig]
def test_repo_not_found(self):
with self.assertRaisesRegex(
EnvironmentError, "bert-base is not a local folder and is not a valid model identifier"
):
_ = TFAutoModel.from_pretrained("bert-base")
def test_revision_not_found(self):
with self.assertRaisesRegex(
EnvironmentError, r"aaaaaa is not a valid git identifier \(branch name, tag name or commit id\)"
):
_ = TFAutoModel.from_pretrained(DUMMY_UNKNOWN_IDENTIFIER, revision="aaaaaa")
def test_model_file_not_found(self):
with self.assertRaisesRegex(
EnvironmentError,
"hf-internal-testing/config-no-model does not appear to have a file named pytorch_model.bin",
):
_ = TFAutoModel.from_pretrained("hf-internal-testing/config-no-model")
def test_model_from_pt_suggestion(self):
with self.assertRaisesRegex(EnvironmentError, "Use `from_pt=True` to load this model"):
_ = TFAutoModel.from_pretrained("hf-internal-testing/tiny-bert-pt-only")
def test_cached_model_has_minimum_calls_to_head(self):
# Make sure we have cached the model.
_ = TFAutoModel.from_pretrained("hf-internal-testing/tiny-random-bert")
with RequestCounter() as counter:
_ = TFAutoModel.from_pretrained("hf-internal-testing/tiny-random-bert")
self.assertEqual(counter["GET"], 0)
self.assertEqual(counter["HEAD"], 1)
self.assertEqual(counter.total_calls, 1)
# With a sharded checkpoint
_ = TFAutoModel.from_pretrained("ArthurZ/tiny-random-bert-sharded")
with RequestCounter() as counter:
_ = TFAutoModel.from_pretrained("ArthurZ/tiny-random-bert-sharded")
self.assertEqual(counter["GET"], 0)
self.assertEqual(counter["HEAD"], 1)
self.assertEqual(counter.total_calls, 1)
|
transformers/tests/models/auto/test_modeling_tf_auto.py/0
|
{
"file_path": "transformers/tests/models/auto/test_modeling_tf_auto.py",
"repo_id": "transformers",
"token_count": 5501
}
| 401
|
# coding=utf-8
# Copyright 2020 The HuggingFace Team. 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.
import os
import pickle
import unittest
from transformers import AutoTokenizer
from transformers.models.bert.tokenization_bert import BertTokenizer
from transformers.models.bert_japanese.tokenization_bert_japanese import (
VOCAB_FILES_NAMES,
BertJapaneseTokenizer,
CharacterTokenizer,
JumanppTokenizer,
MecabTokenizer,
SudachiTokenizer,
WordpieceTokenizer,
)
from transformers.testing_utils import custom_tokenizers, require_jumanpp, require_sudachi_projection
from ...test_tokenization_common import TokenizerTesterMixin
@custom_tokenizers
class BertJapaneseTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
from_pretrained_id = "cl-tohoku/bert-base-japanese"
tokenizer_class = BertJapaneseTokenizer
test_rust_tokenizer = False
space_between_special_tokens = True
def setUp(self):
super().setUp()
vocab_tokens = [
"[UNK]",
"[CLS]",
"[SEP]",
"こんにちは",
"こん",
"にちは",
"ばんは",
"##こん",
"##にちは",
"##ばんは",
"世界",
"##世界",
"、",
"##、",
"。",
"##。",
"アップルストア",
"外国",
"##人",
"参政",
"##権",
"此れ",
"は",
"猫",
"です",
]
self.vocab_file = os.path.join(self.tmpdirname, VOCAB_FILES_NAMES["vocab_file"])
with open(self.vocab_file, "w", encoding="utf-8") as vocab_writer:
vocab_writer.write("".join([x + "\n" for x in vocab_tokens]))
def get_input_output_texts(self, tokenizer):
input_text = "こんにちは、世界。 \nこんばんは、世界。"
output_text = "こんにちは 、 世界 。 こんばんは 、 世界 。"
return input_text, output_text
def get_clean_sequence(self, tokenizer):
input_text, output_text = self.get_input_output_texts(tokenizer)
ids = tokenizer.encode(output_text, add_special_tokens=False)
text = tokenizer.decode(ids, clean_up_tokenization_spaces=False)
return text, ids
def test_pretokenized_inputs(self):
pass # TODO add if relevant
def test_maximum_encoding_length_pair_input(self):
pass # TODO add if relevant
def test_maximum_encoding_length_single_input(self):
pass # TODO add if relevant
def test_full_tokenizer(self):
tokenizer = self.tokenizer_class(self.vocab_file)
tokens = tokenizer.tokenize("こんにちは、世界。\nこんばんは、世界。")
self.assertListEqual(tokens, ["こんにちは", "、", "世界", "。", "こん", "##ばんは", "、", "世界", "。"])
self.assertListEqual(tokenizer.convert_tokens_to_ids(tokens), [3, 12, 10, 14, 4, 9, 12, 10, 14])
def test_pickle_mecab_tokenizer(self):
tokenizer = self.tokenizer_class(self.vocab_file, word_tokenizer_type="mecab")
self.assertIsNotNone(tokenizer)
text = "こんにちは、世界。\nこんばんは、世界。"
tokens = tokenizer.tokenize(text)
self.assertListEqual(tokens, ["こんにちは", "、", "世界", "。", "こん", "##ばんは", "、", "世界", "。"])
self.assertListEqual(tokenizer.convert_tokens_to_ids(tokens), [3, 12, 10, 14, 4, 9, 12, 10, 14])
filename = os.path.join(self.tmpdirname, "tokenizer.bin")
with open(filename, "wb") as handle:
pickle.dump(tokenizer, handle)
with open(filename, "rb") as handle:
tokenizer_new = pickle.load(handle)
tokens_loaded = tokenizer_new.tokenize(text)
self.assertListEqual(tokens, tokens_loaded)
def test_mecab_full_tokenizer_with_mecab_kwargs(self):
tokenizer = self.tokenizer_class(
self.vocab_file, word_tokenizer_type="mecab", mecab_kwargs={"mecab_dic": "ipadic"}
)
text = "ｱｯﾌﾟﾙストア"
tokens = tokenizer.tokenize(text)
self.assertListEqual(tokens, ["アップルストア"])
def test_mecab_tokenizer_ipadic(self):
tokenizer = MecabTokenizer(mecab_dic="ipadic")
self.assertListEqual(
tokenizer.tokenize(" \tｱｯﾌﾟﾙストアでiPhone８ が \n 発売された 。 "),
["アップルストア", "で", "iPhone", "8", "が", "発売", "さ", "れ", "た", "。"],
)
def test_mecab_tokenizer_unidic_lite(self):
try:
tokenizer = MecabTokenizer(mecab_dic="unidic_lite")
except ModuleNotFoundError:
return
self.assertListEqual(
tokenizer.tokenize(" \tｱｯﾌﾟﾙストアでiPhone８ が \n 発売された 。 "),
["アップル", "ストア", "で", "iPhone", "8", "が", "発売", "さ", "れ", "た", "。"],
)
def test_mecab_tokenizer_unidic(self):
try:
import unidic
self.assertTrue(
os.path.isdir(unidic.DICDIR),
"The content of unidic was not downloaded. Run `python -m unidic download` before running this test case. Note that this requires 2.1GB on disk.",
)
tokenizer = MecabTokenizer(mecab_dic="unidic")
except ModuleNotFoundError:
return
self.assertListEqual(
tokenizer.tokenize(" \tｱｯﾌﾟﾙストアでiPhone８ が \n 発売された 。 "),
["アップル", "ストア", "で", "iPhone", "8", "が", "発売", "さ", "れ", "た", "。"],
)
def test_mecab_tokenizer_lower(self):
tokenizer = MecabTokenizer(do_lower_case=True, mecab_dic="ipadic")
self.assertListEqual(
tokenizer.tokenize(" \tｱｯﾌﾟﾙストアでiPhone８ が \n 発売された 。 "),
["アップルストア", "で", "iphone", "8", "が", "発売", "さ", "れ", "た", "。"],
)
def test_mecab_tokenizer_with_option(self):
try:
tokenizer = MecabTokenizer(
do_lower_case=True, normalize_text=False, mecab_option="-d /usr/local/lib/mecab/dic/jumandic"
)
except RuntimeError:
# if dict doesn't exist in the system, previous code raises this error.
return
self.assertListEqual(
tokenizer.tokenize(" \tｱｯﾌﾟﾙストアでiPhone８ が \n 発売された 。 "),
["ｱｯﾌﾟﾙストア", "で", "iPhone", "８", "が", "発売", "さ", "れた", "\u3000", "。"],
)
def test_mecab_tokenizer_no_normalize(self):
tokenizer = MecabTokenizer(normalize_text=False, mecab_dic="ipadic")
self.assertListEqual(
tokenizer.tokenize(" \tｱｯﾌﾟﾙストアでiPhone８ が \n 発売された 。 "),
["ｱｯﾌﾟﾙストア", "で", "iPhone", "８", "が", "発売", "さ", "れ", "た", " ", "。"],
)
@require_sudachi_projection
def test_pickle_sudachi_tokenizer(self):
tokenizer = self.tokenizer_class(self.vocab_file, word_tokenizer_type="sudachi")
self.assertIsNotNone(tokenizer)
text = "こんにちは、世界。\nこんばんは、世界。"
tokens = tokenizer.tokenize(text)
self.assertListEqual(tokens, ["こんにちは", "、", "世界", "。", "こん", "##ばんは", "、", "世界", "。"])
self.assertListEqual(tokenizer.convert_tokens_to_ids(tokens), [3, 12, 10, 14, 4, 9, 12, 10, 14])
filename = os.path.join(self.tmpdirname, "tokenizer.bin")
with open(filename, "wb") as handle:
pickle.dump(tokenizer, handle)
with open(filename, "rb") as handle:
tokenizer_new = pickle.load(handle)
tokens_loaded = tokenizer_new.tokenize(text)
self.assertListEqual(tokens, tokens_loaded)
@require_sudachi_projection
def test_sudachi_tokenizer_core(self):
tokenizer = SudachiTokenizer(sudachi_dict_type="core")
# fmt: off
self.assertListEqual(
tokenizer.tokenize(" \tｱｯﾌﾟﾙストアでiPhone８ が \n 発売された 。 "),
[" ", "\t", "アップル", "ストア", "で", "iPhone", "8", " ", "が", " ", " ", "\n ", "発売", "さ", "れ", "た", " ", "。", " ", " "],
)
# fmt: on
@require_sudachi_projection
def test_sudachi_tokenizer_split_mode_A(self):
tokenizer = SudachiTokenizer(sudachi_dict_type="core", sudachi_split_mode="A")
self.assertListEqual(tokenizer.tokenize("外国人参政権"), ["外国", "人", "参政", "権"])
@require_sudachi_projection
def test_sudachi_tokenizer_split_mode_B(self):
tokenizer = SudachiTokenizer(sudachi_dict_type="core", sudachi_split_mode="B")
self.assertListEqual(tokenizer.tokenize("外国人参政権"), ["外国人", "参政権"])
@require_sudachi_projection
def test_sudachi_tokenizer_split_mode_C(self):
tokenizer = SudachiTokenizer(sudachi_dict_type="core", sudachi_split_mode="C")
self.assertListEqual(tokenizer.tokenize("外国人参政権"), ["外国人参政権"])
@require_sudachi_projection
def test_sudachi_full_tokenizer_with_sudachi_kwargs_split_mode_B(self):
tokenizer = self.tokenizer_class(
self.vocab_file, word_tokenizer_type="sudachi", sudachi_kwargs={"sudachi_split_mode": "B"}
)
self.assertListEqual(tokenizer.tokenize("外国人参政権"), ["外国", "##人", "参政", "##権"])
@require_sudachi_projection
def test_sudachi_tokenizer_projection(self):
tokenizer = SudachiTokenizer(
sudachi_dict_type="core", sudachi_split_mode="A", sudachi_projection="normalized_nouns"
)
self.assertListEqual(tokenizer.tokenize("これはねこです。"), ["此れ", "は", "猫", "です", "。"])
@require_sudachi_projection
def test_sudachi_full_tokenizer_with_sudachi_kwargs_sudachi_projection(self):
tokenizer = self.tokenizer_class(
self.vocab_file, word_tokenizer_type="sudachi", sudachi_kwargs={"sudachi_projection": "normalized_nouns"}
)
self.assertListEqual(tokenizer.tokenize("これはねこです。"), ["此れ", "は", "猫", "です", "。"])
@require_sudachi_projection
def test_sudachi_tokenizer_lower(self):
tokenizer = SudachiTokenizer(do_lower_case=True, sudachi_dict_type="core")
self.assertListEqual(tokenizer.tokenize(" \tｱｯﾌﾟﾙストアでiPhone８ が \n 発売された 。 "),[" ", "\t", "アップル", "ストア", "で", "iphone", "8", " ", "が", " ", " ", "\n ", "発売", "さ", "れ", "た", " ", "。", " ", " "]) # fmt: skip
@require_sudachi_projection
def test_sudachi_tokenizer_no_normalize(self):
tokenizer = SudachiTokenizer(normalize_text=False, sudachi_dict_type="core")
self.assertListEqual(tokenizer.tokenize(" \tｱｯﾌﾟﾙストアでiPhone８ が \n 発売された 。 "),[" ", "\t", "ｱｯﾌﾟﾙ", "ストア", "で", "iPhone", "８", " ", "が", " ", " ", "\n ", "発売", "さ", "れ", "た", "\u3000", "。", " ", " "]) # fmt: skip
@require_sudachi_projection
def test_sudachi_tokenizer_trim_whitespace(self):
tokenizer = SudachiTokenizer(trim_whitespace=True, sudachi_dict_type="core")
self.assertListEqual(
tokenizer.tokenize(" \tｱｯﾌﾟﾙストアでiPhone８ が \n 発売された 。 "),
["アップル", "ストア", "で", "iPhone", "8", "が", "発売", "さ", "れ", "た", "。"],
)
@require_jumanpp
def test_pickle_jumanpp_tokenizer(self):
tokenizer = self.tokenizer_class(self.vocab_file, word_tokenizer_type="jumanpp")
self.assertIsNotNone(tokenizer)
text = "こんにちは、世界。\nこんばんは、世界。"
tokens = tokenizer.tokenize(text)
self.assertListEqual(tokens, ["こんにちは", "、", "世界", "。", "こん", "##ばんは", "、", "世界", "。"])
self.assertListEqual(tokenizer.convert_tokens_to_ids(tokens), [3, 12, 10, 14, 4, 9, 12, 10, 14])
filename = os.path.join(self.tmpdirname, "tokenizer.bin")
with open(filename, "wb") as handle:
pickle.dump(tokenizer, handle)
with open(filename, "rb") as handle:
tokenizer_new = pickle.load(handle)
tokens_loaded = tokenizer_new.tokenize(text)
self.assertListEqual(tokens, tokens_loaded)
@require_jumanpp
def test_jumanpp_tokenizer(self):
tokenizer = JumanppTokenizer()
self.assertListEqual(
tokenizer.tokenize(" \tｱｯﾌﾟﾙストアでiPhone８ が \n 発売された 。 "),["アップル", "ストア", "で", "iPhone", "8", "\u3000", "が", "\u3000", "\u3000", "\u3000", "発売", "さ", "れた", "\u3000", "。"]) # fmt: skip
@require_jumanpp
def test_jumanpp_tokenizer_lower(self):
tokenizer = JumanppTokenizer(do_lower_case=True)
self.assertListEqual(tokenizer.tokenize(" \tｱｯﾌﾟﾙストアでiPhone８ が \n 発売された 。 "),["アップル", "ストア", "で", "iphone", "8", "\u3000", "が", "\u3000", "\u3000", "\u3000", "発売", "さ", "れた", "\u3000", "。"],) # fmt: skip
@require_jumanpp
def test_jumanpp_tokenizer_no_normalize(self):
tokenizer = JumanppTokenizer(normalize_text=False)
self.assertListEqual(tokenizer.tokenize(" \tｱｯﾌﾟﾙストアでiPhone８ が \n 発売された 。 "),["ｱ", "ｯ", "ﾌ", "ﾟ", "ﾙ", "ストア", "で", "iPhone", "８", "\u3000", "が", "\u3000", "\u3000", "\u3000", "発売", "さ", "れた", "\u3000", "。"],) # fmt: skip
@require_jumanpp
def test_jumanpp_tokenizer_trim_whitespace(self):
tokenizer = JumanppTokenizer(trim_whitespace=True)
self.assertListEqual(
tokenizer.tokenize(" \tｱｯﾌﾟﾙストアでiPhone８ が \n 発売された 。 "),
["アップル", "ストア", "で", "iPhone", "8", "が", "発売", "さ", "れた", "。"],
)
@require_jumanpp
def test_jumanpp_full_tokenizer_with_jumanpp_kwargs_trim_whitespace(self):
tokenizer = self.tokenizer_class(
self.vocab_file, word_tokenizer_type="jumanpp", jumanpp_kwargs={"trim_whitespace": True}
)
text = "こんにちは、世界。\nこんばんは、世界。"
tokens = tokenizer.tokenize(text)
self.assertListEqual(tokens, ["こんにちは", "、", "世界", "。", "こん", "##ばんは", "、", "世界", "。"])
self.assertListEqual(tokenizer.convert_tokens_to_ids(tokens), [3, 12, 10, 14, 4, 9, 12, 10, 14])
@require_jumanpp
def test_jumanpp_tokenizer_ext(self):
tokenizer = JumanppTokenizer()
self.assertListEqual(
tokenizer.tokenize("ありがとうございますm(_ _)ｍ見つけるのが大変です。"),
["ありがとう", "ございます", "m(_ _)m", "見つける", "の", "が", "大変です", "。"],
)
def test_wordpiece_tokenizer(self):
vocab_tokens = ["[UNK]", "[CLS]", "[SEP]", "こんにちは", "こん", "にちは", "ばんは", "##こん", "##にちは", "##ばんは"] # fmt: skip
vocab = {}
for i, token in enumerate(vocab_tokens):
vocab[token] = i
tokenizer = WordpieceTokenizer(vocab=vocab, unk_token="[UNK]")
self.assertListEqual(tokenizer.tokenize(""), [])
self.assertListEqual(tokenizer.tokenize("こんにちは"), ["こんにちは"])
self.assertListEqual(tokenizer.tokenize("こんばんは"), ["こん", "##ばんは"])
self.assertListEqual(tokenizer.tokenize("こんばんは こんばんにちは こんにちは"), ["こん", "##ばんは", "[UNK]", "こんにちは"]) # fmt: skip
def test_sentencepiece_tokenizer(self):
tokenizer = BertJapaneseTokenizer.from_pretrained("nlp-waseda/roberta-base-japanese-with-auto-jumanpp")
subword_tokenizer = tokenizer.subword_tokenizer
tokens = subword_tokenizer.tokenize("国境 の 長い トンネル を 抜ける と 雪国 であった 。")
self.assertListEqual(tokens, ["▁国境", "▁の", "▁長い", "▁トンネル", "▁を", "▁抜ける", "▁と", "▁雪", "国", "▁であった", "▁。"]) # fmt: skip
tokens = subword_tokenizer.tokenize("こんばんは こんばん にち は こんにちは")
self.assertListEqual(tokens, ["▁こん", "ばん", "は", "▁こん", "ばん", "▁に", "ち", "▁は", "▁こんにちは"])
def test_sequence_builders(self):
tokenizer = self.tokenizer_class.from_pretrained("cl-tohoku/bert-base-japanese")
text = tokenizer.encode("ありがとう。", add_special_tokens=False)
text_2 = tokenizer.encode("どういたしまして。", add_special_tokens=False)
encoded_sentence = tokenizer.build_inputs_with_special_tokens(text)
encoded_pair = tokenizer.build_inputs_with_special_tokens(text, text_2)
# 2 is for "[CLS]", 3 is for "[SEP]"
assert encoded_sentence == [2] + text + [3]
assert encoded_pair == [2] + text + [3] + text_2 + [3]
@custom_tokenizers
class BertJapaneseCharacterTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
from_pretrained_id = "cl-tohoku/bert-base-japanese"
tokenizer_class = BertJapaneseTokenizer
test_rust_tokenizer = False
def setUp(self):
super().setUp()
vocab_tokens = ["[UNK]", "[CLS]", "[SEP]", "こ", "ん", "に", "ち", "は", "ば", "世", "界", "、", "。"]
self.vocab_file = os.path.join(self.tmpdirname, VOCAB_FILES_NAMES["vocab_file"])
with open(self.vocab_file, "w", encoding="utf-8") as vocab_writer:
vocab_writer.write("".join([x + "\n" for x in vocab_tokens]))
def get_tokenizer(self, **kwargs):
return BertJapaneseTokenizer.from_pretrained(self.tmpdirname, subword_tokenizer_type="character", **kwargs)
def get_input_output_texts(self, tokenizer):
input_text = "こんにちは、世界。 \nこんばんは、世界。"
output_text = "こ ん に ち は 、 世 界 。 こ ん ば ん は 、 世 界 。"
return input_text, output_text
def test_pretokenized_inputs(self):
pass # TODO add if relevant
def test_maximum_encoding_length_pair_input(self):
pass # TODO add if relevant
def test_maximum_encoding_length_single_input(self):
pass # TODO add if relevant
def test_full_tokenizer(self):
tokenizer = self.tokenizer_class(self.vocab_file, subword_tokenizer_type="character")
tokens = tokenizer.tokenize("こんにちは、世界。 \nこんばんは、世界。")
self.assertListEqual(tokens, ["こ", "ん", "に", "ち", "は", "、", "世", "界", "。", "こ", "ん", "ば", "ん", "は", "、", "世", "界", "。"]) # fmt: skip
self.assertListEqual(
tokenizer.convert_tokens_to_ids(tokens), [3, 4, 5, 6, 7, 11, 9, 10, 12, 3, 4, 8, 4, 7, 11, 9, 10, 12]
)
def test_character_tokenizer(self):
vocab_tokens = ["[UNK]", "[CLS]", "[SEP]", "こ", "ん", "に", "ち", "は", "ば", "世", "界", "、", "。"]
vocab = {}
for i, token in enumerate(vocab_tokens):
vocab[token] = i
tokenizer = CharacterTokenizer(vocab=vocab, unk_token="[UNK]")
self.assertListEqual(tokenizer.tokenize(""), [])
self.assertListEqual(tokenizer.tokenize("こんにちは"), ["こ", "ん", "に", "ち", "は"])
self.assertListEqual(tokenizer.tokenize("こんにちほ"), ["こ", "ん", "に", "ち", "[UNK]"])
def test_sequence_builders(self):
tokenizer = self.tokenizer_class.from_pretrained("cl-tohoku/bert-base-japanese-char")
text = tokenizer.encode("ありがとう。", add_special_tokens=False)
text_2 = tokenizer.encode("どういたしまして。", add_special_tokens=False)
encoded_sentence = tokenizer.build_inputs_with_special_tokens(text)
encoded_pair = tokenizer.build_inputs_with_special_tokens(text, text_2)
# 2 is for "[CLS]", 3 is for "[SEP]"
assert encoded_sentence == [2] + text + [3]
assert encoded_pair == [2] + text + [3] + text_2 + [3]
@custom_tokenizers
class AutoTokenizerCustomTest(unittest.TestCase):
def test_tokenizer_bert_japanese(self):
EXAMPLE_BERT_JAPANESE_ID = "cl-tohoku/bert-base-japanese"
tokenizer = AutoTokenizer.from_pretrained(EXAMPLE_BERT_JAPANESE_ID)
self.assertIsInstance(tokenizer, BertJapaneseTokenizer)
class BertTokenizerMismatchTest(unittest.TestCase):
def test_tokenizer_mismatch_warning(self):
EXAMPLE_BERT_JAPANESE_ID = "cl-tohoku/bert-base-japanese"
with self.assertLogs("transformers", level="WARNING") as cm:
BertTokenizer.from_pretrained(EXAMPLE_BERT_JAPANESE_ID)
self.assertTrue(
cm.records[0].message.startswith(
"The tokenizer class you load from this checkpoint is not the same type as the class this function"
" is called from."
)
)
EXAMPLE_BERT_ID = "google-bert/bert-base-cased"
with self.assertLogs("transformers", level="WARNING") as cm:
BertJapaneseTokenizer.from_pretrained(EXAMPLE_BERT_ID)
self.assertTrue(
cm.records[0].message.startswith(
"The tokenizer class you load from this checkpoint is not the same type as the class this function"
" is called from."
)
)
|
transformers/tests/models/bert_japanese/test_tokenization_bert_japanese.py/0
|
{
"file_path": "transformers/tests/models/bert_japanese/test_tokenization_bert_japanese.py",
"repo_id": "transformers",
"token_count": 10470
}
| 402
|
# coding=utf-8
# Copyright 2023 The HuggingFace Inc. team. 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.
"""Testing suite for the PyTorch BLIP-2 model."""
import inspect
import tempfile
import unittest
import numpy as np
import requests
from transformers import CONFIG_MAPPING, Blip2Config, Blip2QFormerConfig, Blip2VisionConfig
from transformers.testing_utils import (
require_torch,
require_torch_fp16,
require_torch_gpu,
require_torch_multi_accelerator,
require_vision,
slow,
torch_device,
)
from transformers.utils import is_torch_available, is_vision_available
from ...generation.test_utils import GenerationTesterMixin
from ...test_configuration_common import ConfigTester
from ...test_modeling_common import (
ModelTesterMixin,
_config_zero_init,
floats_tensor,
ids_tensor,
random_attention_mask,
)
from ...test_pipeline_mixin import PipelineTesterMixin
if is_torch_available():
import torch
from torch import nn
from transformers import (
Blip2ForConditionalGeneration,
Blip2ForImageTextRetrieval,
Blip2Model,
Blip2TextModelWithProjection,
Blip2VisionModel,
Blip2VisionModelWithProjection,
)
if is_vision_available():
from PIL import Image
from transformers import Blip2Processor
class Blip2VisionModelTester:
def __init__(
self,
parent,
batch_size=12,
image_size=30,
patch_size=2,
num_channels=3,
is_training=True,
hidden_size=32,
projection_dim=32,
num_hidden_layers=2,
num_attention_heads=4,
intermediate_size=37,
dropout=0.1,
attention_dropout=0.1,
initializer_range=1e-10,
scope=None,
):
self.parent = parent
self.batch_size = batch_size
self.image_size = image_size
self.patch_size = patch_size
self.num_channels = num_channels
self.is_training = is_training
self.hidden_size = hidden_size
self.projection_dim = projection_dim
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
self.intermediate_size = intermediate_size
self.dropout = dropout
self.attention_dropout = attention_dropout
self.initializer_range = initializer_range
self.scope = scope
# in ViT, the seq length equals the number of patches + 1 (we add 1 for the [CLS] token)
num_patches = (image_size // patch_size) ** 2
self.seq_length = num_patches + 1
def prepare_config_and_inputs(self):
pixel_values = floats_tensor([self.batch_size, self.num_channels, self.image_size, self.image_size])
config = self.get_config()
return config, pixel_values
def get_config(self):
return Blip2VisionConfig(
image_size=self.image_size,
patch_size=self.patch_size,
num_channels=self.num_channels,
hidden_size=self.hidden_size,
projection_dim=self.projection_dim,
num_hidden_layers=self.num_hidden_layers,
num_attention_heads=self.num_attention_heads,
intermediate_size=self.intermediate_size,
dropout=self.dropout,
attention_dropout=self.attention_dropout,
initializer_range=self.initializer_range,
)
def create_and_check_model(self, config, pixel_values):
model = Blip2VisionModel(config=config)
model.to(torch_device)
model.eval()
with torch.no_grad():
result = model(pixel_values)
# expected sequence length = num_patches + 1 (we add 1 for the [CLS] token)
image_size = (self.image_size, self.image_size)
patch_size = (self.patch_size, self.patch_size)
num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, num_patches + 1, self.hidden_size))
self.parent.assertEqual(result.pooler_output.shape, (self.batch_size, self.hidden_size))
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
config, pixel_values = config_and_inputs
inputs_dict = {"pixel_values": pixel_values}
return config, inputs_dict
@require_torch
class Blip2VisionModelTest(ModelTesterMixin, unittest.TestCase):
"""
Here we also overwrite some of the tests of test_modeling_common.py, as BLIP-2's vision encoder does not use input_ids, inputs_embeds,
attention_mask and seq_length.
"""
all_model_classes = (Blip2VisionModel,) if is_torch_available() else ()
fx_compatible = False
test_pruning = False
test_resize_embeddings = False
test_head_masking = False
def setUp(self):
self.model_tester = Blip2VisionModelTester(self)
self.config_tester = ConfigTester(
self, config_class=Blip2VisionConfig, has_text_modality=False, hidden_size=37
)
def test_config(self):
self.config_tester.run_common_tests()
@unittest.skip(reason="BLIP-2's vision encoder does not use inputs_embeds")
def test_inputs_embeds(self):
pass
def test_model_get_set_embeddings(self):
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
model = model_class(config)
self.assertIsInstance(model.get_input_embeddings(), (nn.Module))
x = model.get_output_embeddings()
self.assertTrue(x is None or isinstance(x, nn.Linear))
def test_forward_signature(self):
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
model = model_class(config)
signature = inspect.signature(model.forward)
# signature.parameters is an OrderedDict => so arg_names order is deterministic
arg_names = [*signature.parameters.keys()]
expected_arg_names = ["pixel_values"]
self.assertListEqual(arg_names[:1], expected_arg_names)
def test_model(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_model(*config_and_inputs)
@unittest.skip
def test_training(self):
pass
@unittest.skip
def test_training_gradient_checkpointing(self):
pass
@unittest.skip(
reason="This architecure seem to not compute gradients properly when using GC, check: https://github.com/huggingface/transformers/pull/27124"
)
def test_training_gradient_checkpointing_use_reentrant(self):
pass
@unittest.skip(
reason="This architecure seem to not compute gradients properly when using GC, check: https://github.com/huggingface/transformers/pull/27124"
)
def test_training_gradient_checkpointing_use_reentrant_false(self):
pass
@unittest.skip(reason="Blip2VisionModel has no base class and is not available in MODEL_MAPPING")
def test_save_load_fast_init_from_base(self):
pass
@unittest.skip(reason="Blip2VisionModel has no base class and is not available in MODEL_MAPPING")
def test_save_load_fast_init_to_base(self):
pass
@slow
def test_model_from_pretrained(self):
model_name = "Salesforce/blip2-opt-2.7b"
model = Blip2VisionModel.from_pretrained(model_name)
self.assertIsNotNone(model)
class Blip2QFormerModelTester:
def __init__(
self,
parent,
batch_size=12,
seq_length=7,
is_training=True,
use_input_mask=True,
use_labels=True,
vocab_size=99,
hidden_size=32,
projection_dim=32,
num_hidden_layers=2,
num_attention_heads=4,
intermediate_size=37,
dropout=0.1,
attention_dropout=0.1,
max_position_embeddings=512,
initializer_range=0.02,
bos_token_id=0,
scope=None,
use_qformer_text_input=False,
):
self.parent = parent
self.batch_size = batch_size
self.seq_length = seq_length
self.is_training = is_training
self.use_input_mask = use_input_mask
self.use_labels = use_labels
self.vocab_size = vocab_size
self.hidden_size = hidden_size
self.projection_dim = projection_dim
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
self.intermediate_size = intermediate_size
self.dropout = dropout
self.attention_dropout = attention_dropout
self.max_position_embeddings = max_position_embeddings
self.initializer_range = initializer_range
self.scope = scope
self.bos_token_id = bos_token_id
self.use_qformer_text_input = use_qformer_text_input
def prepare_config_and_inputs(self):
input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
input_mask = None
if self.use_input_mask:
input_mask = random_attention_mask([self.batch_size, self.seq_length])
if input_mask is not None:
batch_size, seq_length = input_mask.shape
rnd_start_indices = np.random.randint(1, seq_length - 1, size=(batch_size,))
for batch_idx, start_index in enumerate(rnd_start_indices):
input_mask[batch_idx, :start_index] = 1
input_mask[batch_idx, start_index:] = 0
config = self.get_config()
return config, input_ids, input_mask
def get_config(self):
return Blip2QFormerConfig(
vocab_size=self.vocab_size,
hidden_size=self.hidden_size,
projection_dim=self.projection_dim,
num_hidden_layers=self.num_hidden_layers,
num_attention_heads=self.num_attention_heads,
intermediate_size=self.intermediate_size,
dropout=self.dropout,
attention_dropout=self.attention_dropout,
max_position_embeddings=self.max_position_embeddings,
initializer_range=self.initializer_range,
bos_token_id=self.bos_token_id,
use_qformer_text_input=self.use_qformer_text_input,
)
# this class is based on `OPTModelTester` found in tests/models/opt/test_modeling_opt.py
class Blip2TextModelDecoderOnlyTester:
def __init__(
self,
parent,
batch_size=12,
seq_length=7,
is_training=True,
use_labels=False,
vocab_size=99,
hidden_size=16,
num_hidden_layers=2,
num_attention_heads=4,
intermediate_size=4,
hidden_act="gelu",
hidden_dropout_prob=0.1,
attention_probs_dropout_prob=0.1,
max_position_embeddings=20,
eos_token_id=2,
pad_token_id=1,
bos_token_id=0,
embed_dim=16,
num_labels=3,
word_embed_proj_dim=16,
type_sequence_label_size=2,
):
self.parent = parent
self.batch_size = batch_size
self.seq_length = seq_length
self.is_training = is_training
self.use_labels = use_labels
self.vocab_size = vocab_size
self.hidden_size = hidden_size
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
self.intermediate_size = intermediate_size
self.hidden_act = hidden_act
self.hidden_dropout_prob = hidden_dropout_prob
self.attention_probs_dropout_prob = attention_probs_dropout_prob
self.max_position_embeddings = max_position_embeddings
self.eos_token_id = eos_token_id
self.pad_token_id = pad_token_id
self.bos_token_id = bos_token_id
self.embed_dim = embed_dim
self.num_labels = num_labels
self.type_sequence_label_size = type_sequence_label_size
self.word_embed_proj_dim = word_embed_proj_dim
self.is_encoder_decoder = False
def prepare_config_and_inputs(self):
config = self.get_config()
input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size).clamp(3)
input_ids[:, -1] = self.eos_token_id # Eos Token
attention_mask = input_ids.ne(self.pad_token_id)
return config, input_ids, attention_mask
def get_config(self):
return CONFIG_MAPPING["opt"](
vocab_size=self.vocab_size,
hidden_size=self.hidden_size,
num_hidden_layers=self.num_hidden_layers,
num_attention_heads=self.num_attention_heads,
ffn_dim=self.intermediate_size,
dropout=self.hidden_dropout_prob,
attention_dropout=self.attention_probs_dropout_prob,
max_position_embeddings=self.max_position_embeddings,
eos_token_id=self.eos_token_id,
bos_token_id=self.bos_token_id,
pad_token_id=self.pad_token_id,
embed_dim=self.embed_dim,
is_encoder_decoder=False,
word_embed_proj_dim=self.word_embed_proj_dim,
)
# this model tester uses a decoder-only language model (OPT)
class Blip2ForConditionalGenerationDecoderOnlyModelTester:
def __init__(
self, parent, vision_kwargs=None, qformer_kwargs=None, text_kwargs=None, is_training=True, num_query_tokens=10
):
if vision_kwargs is None:
vision_kwargs = {}
if qformer_kwargs is None:
qformer_kwargs = {}
if text_kwargs is None:
text_kwargs = {}
self.parent = parent
self.vision_model_tester = Blip2VisionModelTester(parent, **vision_kwargs)
self.qformer_model_tester = Blip2QFormerModelTester(parent, **qformer_kwargs)
self.text_model_tester = Blip2TextModelDecoderOnlyTester(parent, **text_kwargs)
self.batch_size = self.text_model_tester.batch_size # need bs for batching_equivalence test
self.seq_length = self.text_model_tester.seq_length # need seq_length for common tests
self.is_training = is_training
self.num_query_tokens = num_query_tokens
def prepare_config_and_inputs(self):
_, pixel_values = self.vision_model_tester.prepare_config_and_inputs()
_, input_ids, attention_mask = self.text_model_tester.prepare_config_and_inputs()
config = self.get_config()
return config, input_ids, attention_mask, pixel_values
def get_config(self):
return Blip2Config.from_vision_qformer_text_configs(
vision_config=self.vision_model_tester.get_config(),
qformer_config=self.qformer_model_tester.get_config(),
text_config=self.text_model_tester.get_config(),
num_query_tokens=self.num_query_tokens,
)
def create_and_check_for_conditional_generation(self, config, input_ids, attention_mask, pixel_values):
model = Blip2ForConditionalGeneration(config).to(torch_device).eval()
with torch.no_grad():
result = model(pixel_values, input_ids, attention_mask)
expected_seq_length = self.num_query_tokens + self.text_model_tester.seq_length
self.parent.assertEqual(
result.logits.shape,
(self.vision_model_tester.batch_size, expected_seq_length, self.text_model_tester.vocab_size),
)
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
config, input_ids, attention_mask, pixel_values = config_and_inputs
inputs_dict = {
"pixel_values": pixel_values,
"input_ids": input_ids,
"attention_mask": attention_mask,
"labels": input_ids,
}
return config, inputs_dict
@require_torch
class Blip2ForConditionalGenerationDecoderOnlyTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):
all_model_classes = (Blip2ForConditionalGeneration,) if is_torch_available() else ()
fx_compatible = False
test_head_masking = False
test_pruning = False
test_resize_embeddings = False
test_attention_outputs = False
test_torchscript = False
def setUp(self):
self.model_tester = Blip2ForConditionalGenerationDecoderOnlyModelTester(self)
def test_for_conditional_generation(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_for_conditional_generation(*config_and_inputs)
@unittest.skip(reason="Hidden_states is tested in individual model tests")
def test_hidden_states_output(self):
pass
@unittest.skip(reason="Inputs_embeds is tested in individual model tests")
def test_inputs_embeds(self):
pass
@unittest.skip(reason="Retain_grad is tested in individual model tests")
def test_retain_grad_hidden_states_attentions(self):
pass
@unittest.skip(reason="Blip2Model does not have input/output embeddings")
def test_model_get_set_embeddings(self):
pass
@unittest.skip(reason="There's no base Blip2Model")
def test_save_load_fast_init_from_base(self):
pass
@unittest.skip(reason="There's no base Blip2Model")
def test_save_load_fast_init_to_base(self):
pass
def test_forward_signature(self):
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
model = model_class(config)
signature = inspect.signature(model.forward)
# signature.parameters is an OrderedDict => so arg_names order is deterministic
arg_names = [*signature.parameters.keys()]
expected_arg_names = ["pixel_values"]
self.assertListEqual(arg_names[:1], expected_arg_names)
def test_load_vision_qformer_text_config(self):
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
# Save Blip2Config and check if we can load Blip2VisionConfig from it
with tempfile.TemporaryDirectory() as tmp_dir_name:
config.save_pretrained(tmp_dir_name)
vision_config = Blip2VisionConfig.from_pretrained(tmp_dir_name)
self.assertDictEqual(config.vision_config.to_dict(), vision_config.to_dict())
# Save Blip2Config and check if we can load Blip2QFormerConfig from it
with tempfile.TemporaryDirectory() as tmp_dir_name:
config.save_pretrained(tmp_dir_name)
qformer_config = Blip2QFormerConfig.from_pretrained(tmp_dir_name)
self.assertDictEqual(config.qformer_config.to_dict(), qformer_config.to_dict())
@slow
def test_model_from_pretrained(self):
model_name = "Salesforce/blip2-opt-2.7b"
model = Blip2ForConditionalGeneration.from_pretrained(model_name)
self.assertIsNotNone(model)
# this class is based on `T5ModelTester` found in tests/models/t5/test_modeling_t5.py
class Blip2TextModelTester:
def __init__(
self,
parent,
vocab_size=99,
batch_size=12,
encoder_seq_length=7,
decoder_seq_length=9,
# For common tests
is_training=True,
use_attention_mask=True,
use_labels=True,
hidden_size=32,
num_hidden_layers=2,
num_attention_heads=4,
d_ff=37,
relative_attention_num_buckets=8,
dropout_rate=0.1,
initializer_factor=0.002,
eos_token_id=1,
pad_token_id=0,
decoder_start_token_id=0,
scope=None,
decoder_layers=None,
):
self.parent = parent
self.batch_size = batch_size
self.encoder_seq_length = encoder_seq_length
self.decoder_seq_length = decoder_seq_length
# For common tests
self.seq_length = self.decoder_seq_length
self.is_training = is_training
self.use_attention_mask = use_attention_mask
self.use_labels = use_labels
self.vocab_size = vocab_size
self.hidden_size = hidden_size
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
self.d_ff = d_ff
self.relative_attention_num_buckets = relative_attention_num_buckets
self.dropout_rate = dropout_rate
self.initializer_factor = initializer_factor
self.eos_token_id = eos_token_id
self.pad_token_id = pad_token_id
self.decoder_start_token_id = decoder_start_token_id
self.scope = None
self.decoder_layers = decoder_layers
def prepare_config_and_inputs(self):
input_ids = ids_tensor([self.batch_size, self.encoder_seq_length], self.vocab_size)
decoder_input_ids = ids_tensor([self.batch_size, self.decoder_seq_length], self.vocab_size)
attention_mask = None
decoder_attention_mask = None
if self.use_attention_mask:
attention_mask = ids_tensor([self.batch_size, self.encoder_seq_length], vocab_size=2)
decoder_attention_mask = ids_tensor([self.batch_size, self.decoder_seq_length], vocab_size=2)
lm_labels = None
if self.use_labels:
lm_labels = ids_tensor([self.batch_size, self.decoder_seq_length], self.vocab_size)
config = self.get_config()
return (
config,
input_ids,
decoder_input_ids,
attention_mask,
decoder_attention_mask,
lm_labels,
)
def get_config(self):
return CONFIG_MAPPING["t5"](
vocab_size=self.vocab_size,
d_model=self.hidden_size,
d_ff=self.d_ff,
d_kv=self.hidden_size // self.num_attention_heads,
num_layers=self.num_hidden_layers,
num_decoder_layers=self.decoder_layers,
num_heads=self.num_attention_heads,
relative_attention_num_buckets=self.relative_attention_num_buckets,
dropout_rate=self.dropout_rate,
initializer_factor=self.initializer_factor,
eos_token_id=self.eos_token_id,
bos_token_id=self.pad_token_id,
pad_token_id=self.pad_token_id,
decoder_start_token_id=self.decoder_start_token_id,
)
# this model tester uses an encoder-decoder language model (T5)
class Blip2ModelTester:
def __init__(
self, parent, vision_kwargs=None, qformer_kwargs=None, text_kwargs=None, is_training=True, num_query_tokens=10
):
if vision_kwargs is None:
vision_kwargs = {}
if qformer_kwargs is None:
qformer_kwargs = {}
if text_kwargs is None:
text_kwargs = {}
self.parent = parent
self.vision_model_tester = Blip2VisionModelTester(parent, **vision_kwargs)
self.qformer_model_tester = Blip2QFormerModelTester(parent, **qformer_kwargs)
self.text_model_tester = Blip2TextModelTester(parent, **text_kwargs)
self.batch_size = self.text_model_tester.batch_size # need bs for batching_equivalence test
self.seq_length = self.text_model_tester.seq_length # need seq_length for common tests
self.is_training = is_training
self.num_query_tokens = num_query_tokens
def prepare_config_and_inputs(self):
_, pixel_values = self.vision_model_tester.prepare_config_and_inputs()
(
_,
input_ids,
decoder_input_ids,
attention_mask,
decoder_attention_mask,
lm_labels,
) = self.text_model_tester.prepare_config_and_inputs()
config = self.get_config()
return config, input_ids, attention_mask, pixel_values, decoder_input_ids, decoder_attention_mask, lm_labels
def get_config(self):
return Blip2Config.from_vision_qformer_text_configs(
vision_config=self.vision_model_tester.get_config(),
qformer_config=self.qformer_model_tester.get_config(),
text_config=self.text_model_tester.get_config(),
num_query_tokens=self.num_query_tokens,
)
def create_and_check_for_conditional_generation(
self, config, input_ids, attention_mask, pixel_values, decoder_input_ids, decoder_attention_mask, labels
):
model = Blip2ForConditionalGeneration(config).to(torch_device).eval()
with torch.no_grad():
result = model(pixel_values, input_ids, attention_mask, decoder_input_ids, decoder_attention_mask)
self.parent.assertEqual(
result.logits.shape,
(
self.vision_model_tester.batch_size,
self.text_model_tester.seq_length,
self.text_model_tester.vocab_size,
),
)
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
(
config,
input_ids,
attention_mask,
pixel_values,
decoder_input_ids,
decoder_attention_mask,
labels,
) = config_and_inputs
inputs_dict = {
"pixel_values": pixel_values,
"input_ids": input_ids,
"attention_mask": attention_mask,
"decoder_input_ids": decoder_input_ids,
"decoder_attention_mask": decoder_attention_mask,
"labels": labels,
}
return config, inputs_dict
@require_torch
class Blip2ModelTest(ModelTesterMixin, PipelineTesterMixin, GenerationTesterMixin, unittest.TestCase):
all_model_classes = (Blip2ForConditionalGeneration, Blip2Model) if is_torch_available() else ()
pipeline_model_mapping = (
{
"feature-extraction": Blip2Model,
"image-to-text": Blip2ForConditionalGeneration,
"visual-question-answering": Blip2ForConditionalGeneration,
}
if is_torch_available()
else {}
)
fx_compatible = False
test_head_masking = False
test_pruning = False
test_resize_embeddings = False
test_attention_outputs = False
test_torchscript = False
# TODO: Fix the failed tests
def is_pipeline_test_to_skip(
self, pipeline_test_casse_name, config_class, model_architecture, tokenizer_name, processor_name
):
if pipeline_test_casse_name == "VisualQuestionAnsweringPipelineTests":
# Get `RuntimeError: "LayerNormKernelImpl" not implemented for 'Half'`.
return True
return False
def setUp(self):
self.model_tester = Blip2ModelTester(self)
def test_for_conditional_generation(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_for_conditional_generation(*config_and_inputs)
@unittest.skip(reason="Hidden_states is tested in individual model tests")
def test_hidden_states_output(self):
pass
@unittest.skip(reason="Inputs_embeds is tested in individual model tests")
def test_inputs_embeds(self):
pass
@unittest.skip(reason="Retain_grad is tested in individual model tests")
def test_retain_grad_hidden_states_attentions(self):
pass
@unittest.skip(reason="Blip2Model does not have input/output embeddings")
def test_model_get_set_embeddings(self):
pass
@unittest.skip(reason="There's no base Blip2Model")
def test_save_load_fast_init_from_base(self):
pass
@unittest.skip(reason="There's no base Blip2Model")
def test_save_load_fast_init_to_base(self):
pass
@unittest.skip(reason="Does not work on the tiny model as we keep hitting edge cases.")
def test_cpu_offload(self):
pass
def test_forward_signature(self):
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
model = model_class(config)
signature = inspect.signature(model.forward)
# signature.parameters is an OrderedDict => so arg_names order is deterministic
arg_names = [*signature.parameters.keys()]
expected_arg_names = ["pixel_values"]
self.assertListEqual(arg_names[:1], expected_arg_names)
def test_load_vision_qformer_text_config(self):
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
# Save Blip2Config and check if we can load Blip2VisionConfig from it
with tempfile.TemporaryDirectory() as tmp_dir_name:
config.save_pretrained(tmp_dir_name)
vision_config = Blip2VisionConfig.from_pretrained(tmp_dir_name)
self.assertDictEqual(config.vision_config.to_dict(), vision_config.to_dict())
# Save Blip2Config and check if we can load Blip2QFormerConfig from it
with tempfile.TemporaryDirectory() as tmp_dir_name:
config.save_pretrained(tmp_dir_name)
qformer_config = Blip2QFormerConfig.from_pretrained(tmp_dir_name)
self.assertDictEqual(config.qformer_config.to_dict(), qformer_config.to_dict())
@slow
def test_model_from_pretrained(self):
model_name = "Salesforce/blip2-opt-2.7b"
model = Blip2ForConditionalGeneration.from_pretrained(model_name)
self.assertIsNotNone(model)
def test_get_text_features(self):
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
inputs_dict = {
"input_ids": torch.LongTensor([[1, 2, 3, 4, 5, 6, 7, 8, 9, 10]]).to(torch_device),
"attention_mask": torch.LongTensor([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1]]).to(torch_device),
"decoder_input_ids": torch.LongTensor([[1, 2, 3, 4, 5, 6, 7, 8, 9, 10]]).to(torch_device),
}
model = Blip2Model(config).to(torch_device)
model.eval()
text_features = model.get_text_features(**inputs_dict)
self.assertEqual(text_features[0].shape, (1, 10, config.text_config.vocab_size))
def test_get_image_features(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
keys_to_pop = ["input_ids", "attention_mask", "decoder_input_ids", "decoder_attention_mask", "labels"]
for key in keys_to_pop:
inputs_dict.pop(key)
model = Blip2Model(config).to(torch_device)
model.eval()
image_features = model.get_image_features(**inputs_dict)
self.assertEqual(
image_features[0].shape,
(
self.model_tester.vision_model_tester.batch_size,
self.model_tester.vision_model_tester.seq_length,
config.vision_config.hidden_size,
),
)
def test_get_qformer_features(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
keys_to_pop = ["input_ids", "attention_mask", "decoder_input_ids", "decoder_attention_mask", "labels"]
for key in keys_to_pop:
inputs_dict.pop(key)
model = Blip2Model(config).to(torch_device)
model.eval()
qformer_features = model.get_qformer_features(**inputs_dict)
self.assertEqual(
qformer_features[0].shape,
(self.model_tester.vision_model_tester.batch_size, 10, config.vision_config.hidden_size),
)
# override from common to deal with nested configurations (`vision_config`, `text_config` and `qformer_config`)
def test_initialization(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
configs_no_init = _config_zero_init(config)
for key in ["vision_config", "qformer_config", "text_config"]:
setattr(configs_no_init, key, _config_zero_init(getattr(configs_no_init, key)))
for model_class in self.all_model_classes:
model = model_class(config=configs_no_init)
for name, param in model.named_parameters():
if param.requires_grad:
self.assertIn(
((param.data.mean() * 1e9).round() / 1e9).item(),
[0.0, 1.0],
msg=f"Parameter {name} of model {model_class} seems not properly initialized",
)
class Blip2TextModelWithProjectionTester:
def __init__(self, parent, vision_kwargs=None, qformer_kwargs=None, is_training=True):
if vision_kwargs is None:
vision_kwargs = {}
if qformer_kwargs is None:
qformer_kwargs = {"use_qformer_text_input": True}
self.parent = parent
self.vision_model_tester = Blip2VisionModelTester(parent, **vision_kwargs)
self.qformer_model_tester = Blip2QFormerModelTester(parent, **qformer_kwargs)
self.is_training = is_training
self.batch_size = self.vision_model_tester.batch_size # need bs for batching_equivalence test
def get_config(self):
return Blip2Config.from_vision_qformer_text_configs(
vision_config=self.vision_model_tester.get_config(),
qformer_config=self.qformer_model_tester.get_config(),
)
def prepare_config_and_inputs(self):
_, input_ids, attention_mask = self.qformer_model_tester.prepare_config_and_inputs()
config = self.get_config()
return config, input_ids, attention_mask
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
config, input_ids, attention_mask = config_and_inputs
inputs_dict = {
"input_ids": input_ids,
"attention_mask": attention_mask,
}
return config, inputs_dict
def create_and_check_model(self, config, input_ids, attention_mask):
model = Blip2TextModelWithProjection(config=config)
model.to(torch_device)
model.eval()
with torch.no_grad():
result = model(input_ids, attention_mask=attention_mask, output_attentions=True, output_hidden_states=True)
self.parent.assertEqual(
result.last_hidden_state.shape,
(self.vision_model_tester.batch_size, input_ids.shape[1], self.qformer_model_tester.hidden_size),
)
self.parent.assertEqual(
result.text_embeds.shape,
(
self.vision_model_tester.batch_size,
input_ids.shape[1],
config.image_text_hidden_size,
),
)
with torch.no_grad():
result2 = model(
input_ids,
attention_mask=attention_mask,
return_dict=not config.use_return_dict,
output_attentions=True,
output_hidden_states=True,
)
self.parent.assertTrue(torch.allclose(result.text_embeds, result2[0]))
self.parent.assertTrue(torch.allclose(result.last_hidden_state, result2[1]))
self.parent.assertTrue(torch.allclose(result.hidden_states[0], result2[2][0]))
self.parent.assertTrue(torch.allclose(result.hidden_states[1], result2[2][1]))
self.parent.assertTrue(torch.allclose(result.attentions[0], result2[3][0]))
self.parent.assertTrue(torch.allclose(result.attentions[1], result2[3][1]))
@require_torch
class Blip2TextModelWithProjectionTest(ModelTesterMixin, unittest.TestCase):
all_model_classes = (Blip2TextModelWithProjection,) if is_torch_available() else ()
fx_compatible = False
test_pruning = False
test_head_masking = False
test_resize_embeddings = False
test_attention_outputs = False
test_torchscript = False
def setUp(self):
self.model_tester = Blip2TextModelWithProjectionTester(self)
def test_model(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_model(*config_and_inputs)
@unittest.skip(reason="Training is not yet supported")
def test_training(self):
pass
@unittest.skip(reason="Training is not yet supported")
def test_training_gradient_checkpointing(self):
pass
@unittest.skip(reason="Hidden_states is tested in individual model tests")
def test_hidden_states_output(self):
pass
@unittest.skip(reason="Blip2TextModelWithProjection does not use inputs_embeds")
def test_inputs_embeds(self):
pass
@unittest.skip(reason="Blip2TextModelWithProjection does not support input and output embeddings")
def test_model_get_set_embeddings(self):
pass
@unittest.skip(reason="Retain_grad is tested in individual model tests")
def test_retain_grad_hidden_states_attentions(self):
pass
@unittest.skip(reason="Blip2TextModelWithProjection does not have input/output embeddings")
def test_model_common_attributes(self):
pass
@unittest.skip(reason="Blip2TextModelWithProjection has no base class and is not available in MODEL_MAPPING")
def test_save_load_fast_init_from_base(self):
pass
@unittest.skip(reason="Blip2TextModelWithProjection has no base class and is not available in MODEL_MAPPING")
def test_save_load_fast_init_to_base(self):
pass
def test_forward_signature(self):
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
model = model_class(config)
signature = inspect.signature(model.forward)
# signature.parameters is an OrderedDict => so arg_names order is deterministic
arg_names = [*signature.parameters.keys()]
expected_arg_names = ["input_ids", "attention_mask", "position_ids"]
self.assertListEqual(arg_names[: len(expected_arg_names)], expected_arg_names)
@slow
@require_torch_gpu
def test_model_from_pretrained(self):
model_name = "Salesforce/blip2-itm-vit-g"
model = Blip2TextModelWithProjection.from_pretrained(model_name)
self.assertIsNotNone(model)
self.assertTrue(hasattr(model, "text_projection"))
_, input_ids, attention_mask = self.model_tester.prepare_config_and_inputs()
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(input_ids=input_ids, attention_mask=attention_mask)
self.assertEqual(
outputs.text_embeds.shape,
(
self.model_tester.qformer_model_tester.batch_size,
input_ids.shape[1],
model.config.image_text_hidden_size,
),
)
class Blip2VisionModelWithProjectionTester:
def __init__(self, parent, vision_kwargs=None, qformer_kwargs=None, is_training=True):
if vision_kwargs is None:
vision_kwargs = {}
if qformer_kwargs is None:
qformer_kwargs = {"use_qformer_text_input": True}
self.parent = parent
self.vision_model_tester = Blip2VisionModelTester(parent, **vision_kwargs)
self.qformer_model_tester = Blip2QFormerModelTester(parent, **qformer_kwargs)
self.is_training = is_training
self.num_hidden_layers = self.vision_model_tester.num_hidden_layers
self.num_attention_heads = self.vision_model_tester.num_attention_heads
self.seq_length = self.vision_model_tester.seq_length
self.hidden_size = self.vision_model_tester.hidden_size
self.batch_size = self.vision_model_tester.batch_size # need bs for batching_equivalence test
def get_config(self):
return Blip2Config.from_vision_qformer_text_configs(
vision_config=self.vision_model_tester.get_config(),
qformer_config=self.qformer_model_tester.get_config(),
)
def prepare_config_and_inputs(self):
_, pixel_values = self.vision_model_tester.prepare_config_and_inputs()
config = self.get_config()
return config, pixel_values
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
config, pixel_values = config_and_inputs
inputs_dict = {"pixel_values": pixel_values}
return config, inputs_dict
def create_and_check_model(self, config, pixel_values):
model = Blip2VisionModelWithProjection(config=config)
model.to(torch_device)
model.eval()
with torch.no_grad():
result = model(pixel_values, output_attentions=True, output_hidden_states=True)
self.parent.assertEqual(
result.last_hidden_state.shape,
(
self.vision_model_tester.batch_size,
self.vision_model_tester.seq_length,
self.qformer_model_tester.hidden_size,
),
)
self.parent.assertEqual(
result.image_embeds.shape,
(
self.vision_model_tester.batch_size,
config.vision_config.hidden_size,
config.image_text_hidden_size,
),
)
with torch.no_grad():
result2 = model(
pixel_values,
return_dict=not config.use_return_dict,
output_attentions=True,
output_hidden_states=True,
)
self.parent.assertTrue(torch.allclose(result.image_embeds, result2[0]))
self.parent.assertTrue(torch.allclose(result.last_hidden_state, result2[1]))
self.parent.assertTrue(torch.allclose(result.hidden_states[0], result2[2][0]))
self.parent.assertTrue(torch.allclose(result.hidden_states[1], result2[2][1]))
self.parent.assertTrue(torch.allclose(result.attentions[0], result2[3][0]))
self.parent.assertTrue(torch.allclose(result.attentions[1], result2[3][1]))
@require_torch
class Blip2VisionModelWithProjectionTest(ModelTesterMixin, unittest.TestCase):
all_model_classes = (Blip2VisionModelWithProjection,) if is_torch_available() else ()
fx_compatible = False
test_pruning = False
test_head_masking = False
test_resize_embeddings = False
test_torchscript = False
def setUp(self):
self.model_tester = Blip2VisionModelWithProjectionTester(self)
def test_model(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_model(*config_and_inputs)
@unittest.skip(reason="Training is not yet supported")
def test_training(self):
pass
@unittest.skip(reason="Training is not yet supported")
def test_training_gradient_checkpointing(self):
pass
@unittest.skip(reason="Training is not yet supported")
def test_training_gradient_checkpointing_use_reentrant(self):
pass
@unittest.skip(reason="Training is not yet supported")
def test_training_gradient_checkpointing_use_reentrant_false(self):
pass
@unittest.skip(reason="Blip2VisionModelWithProjection does not use inputs_embeds")
def test_inputs_embeds(self):
pass
@unittest.skip(reason="Blip2VisionModelWithProjection does not support input and output embeddings")
def test_model_get_set_embeddings(self):
pass
@unittest.skip(reason="Retain_grad is tested in individual model tests")
def test_retain_grad_hidden_states_attentions(self):
pass
def test_model_common_attributes(self):
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
model = model_class(config)
self.assertIsInstance(model.get_input_embeddings(), (nn.Module))
x = model.get_output_embeddings()
self.assertTrue(x is None or isinstance(x, nn.Linear))
@unittest.skip(reason="Blip2VisionModelWithProjection has no base class and is not available in MODEL_MAPPING")
def test_save_load_fast_init_from_base(self):
pass
@unittest.skip(reason="Blip2VisionModelWithProjection has no base class and is not available in MODEL_MAPPING")
def test_save_load_fast_init_to_base(self):
pass
def test_forward_signature(self):
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
model = model_class(config)
signature = inspect.signature(model.forward)
# signature.parameters is an OrderedDict => so arg_names order is deterministic
arg_names = [*signature.parameters.keys()]
expected_arg_names = ["pixel_values"]
self.assertListEqual(arg_names[: len(expected_arg_names)], expected_arg_names)
@slow
@require_torch_gpu
def test_model_from_pretrained(self):
model_name = "Salesforce/blip2-itm-vit-g"
model = Blip2VisionModelWithProjection.from_pretrained(model_name)
self.assertIsNotNone(model)
self.assertTrue(hasattr(model, "vision_projection"))
_, pixel_values = self.model_tester.prepare_config_and_inputs()
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(pixel_values=pixel_values)
self.assertEqual(
outputs.image_embeds.shape,
(
self.model_tester.vision_model_tester.batch_size,
model.config.num_query_tokens,
model.config.image_text_hidden_size,
),
)
class Blip2TextRetrievalModelTester:
def __init__(self, parent, vision_kwargs=None, qformer_kwargs=None, is_training=True):
if vision_kwargs is None:
vision_kwargs = {}
if qformer_kwargs is None:
qformer_kwargs = {"use_qformer_text_input": True}
self.parent = parent
self.vision_model_tester = Blip2VisionModelTester(parent, **vision_kwargs)
self.qformer_model_tester = Blip2QFormerModelTester(parent, **qformer_kwargs)
self.is_training = is_training
self.batch_size = self.vision_model_tester.batch_size # need bs for batching_equivalence test
def get_config(self):
return Blip2Config.from_vision_qformer_text_configs(
vision_config=self.vision_model_tester.get_config(),
qformer_config=self.qformer_model_tester.get_config(),
)
def prepare_config_and_inputs(self):
_, input_ids, attention_mask = self.qformer_model_tester.prepare_config_and_inputs()
_, pixel_values = self.vision_model_tester.prepare_config_and_inputs()
config = self.get_config()
return config, input_ids, attention_mask, pixel_values
def create_and_check_model(self, config, input_ids, attention_mask, pixel_values):
model = Blip2ForImageTextRetrieval(config).to(torch_device).eval()
with torch.no_grad():
result = model(pixel_values, input_ids, attention_mask, use_image_text_matching_head=True)
self.parent.assertEqual(
result.logits_per_image.shape,
(self.vision_model_tester.batch_size, 2),
)
with torch.no_grad():
result = model(pixel_values, input_ids, attention_mask)
self.parent.assertEqual(
result.logits_per_image.shape,
(self.vision_model_tester.batch_size, self.qformer_model_tester.batch_size),
)
self.parent.assertEqual(
result.logits_per_text.shape, (self.qformer_model_tester.batch_size, self.vision_model_tester.batch_size)
)
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
config, input_ids, attention_mask, pixel_values = config_and_inputs
inputs_dict = {
"input_ids": input_ids,
"attention_mask": attention_mask,
"pixel_values": pixel_values,
}
return config, inputs_dict
@require_torch
class Blip2TextRetrievalModelTest(ModelTesterMixin, unittest.TestCase):
all_model_classes = (Blip2ForImageTextRetrieval,) if is_torch_available() else ()
fx_compatible = False
test_head_masking = False
test_pruning = False
test_resize_embeddings = False
test_attention_outputs = False
test_torchscript = False
def setUp(self):
self.model_tester = Blip2TextRetrievalModelTester(self)
def test_model(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_model(*config_and_inputs)
@unittest.skip(reason="Hidden_states is tested in individual model tests")
def test_hidden_states_output(self):
pass
@unittest.skip(reason="Inputs_embeds is tested in individual model tests")
def test_inputs_embeds(self):
pass
@unittest.skip(reason="Blip2ForImageTextRetrieval does not support input and output embeddings")
def test_model_get_set_embeddings(self):
pass
@unittest.skip(reason="Retain_grad is tested in individual model tests")
def test_retain_grad_hidden_states_attentions(self):
pass
@unittest.skip(reason="Blip2Model does not have input/output embeddings")
def test_model_common_attributes(self):
pass
def test_forward_signature(self):
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
model = model_class(config)
signature = inspect.signature(model.forward)
# signature.parameters is an OrderedDict => so arg_names order is deterministic
arg_names = [*signature.parameters.keys()]
expected_arg_names = ["pixel_values", "input_ids", "attention_mask"]
expected_arg_names.extend(
["use_image_text_matching_head"] if "use_image_text_matching_head" in arg_names else []
)
self.assertListEqual(arg_names[: len(expected_arg_names)], expected_arg_names)
def test_load_vision_qformer_text_config(self):
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
# Save Blip2Config and check if we can load Blip2VisionConfig from it
with tempfile.TemporaryDirectory() as tmp_dir_name:
config.save_pretrained(tmp_dir_name)
vision_config = Blip2VisionConfig.from_pretrained(tmp_dir_name)
self.assertDictEqual(config.vision_config.to_dict(), vision_config.to_dict())
# Save Blip2Config and check if we can load Blip2QFormerConfig from it
with tempfile.TemporaryDirectory() as tmp_dir_name:
config.save_pretrained(tmp_dir_name)
qformer_config = Blip2QFormerConfig.from_pretrained(tmp_dir_name)
self.assertDictEqual(config.qformer_config.to_dict(), qformer_config.to_dict())
@slow
@require_torch_gpu
def test_model_from_pretrained(self):
model_name = "Salesforce/blip2-itm-vit-g"
model = Blip2ForImageTextRetrieval.from_pretrained(model_name)
self.assertIsNotNone(model)
_, input_ids, attention_mask, pixel_values = self.model_tester.prepare_config_and_inputs()
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(
pixel_values=pixel_values,
input_ids=input_ids,
attention_mask=attention_mask,
use_image_text_matching_head=True,
)
self.assertEqual(outputs.logits_per_image.shape, (self.model_tester.qformer_model_tester.batch_size, 2))
with torch.no_grad():
outputs = model(
pixel_values=pixel_values,
input_ids=input_ids,
attention_mask=attention_mask,
)
self.assertEqual(
outputs.logits_per_image.shape,
(self.model_tester.vision_model_tester.batch_size, self.model_tester.qformer_model_tester.batch_size),
)
@unittest.skip(reason="Training is not yet supported")
def test_training(self):
pass
@unittest.skip(reason="Training is not yet supported")
def test_training_gradient_checkpointing(self):
pass
@unittest.skip(reason="Training is not yet supported")
def test_training_gradient_checkpointing_use_reentrant(self):
pass
@unittest.skip(reason="Training is not yet supported")
def test_training_gradient_checkpointing_use_reentrant_false(self):
pass
def test_initialization(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
configs_no_init = _config_zero_init(config)
for model_class in self.all_model_classes:
model = model_class(config=configs_no_init)
for name, param in model.named_parameters():
if param.requires_grad:
# check if `logit_scale` is initilized as per the original implementation
if name == "logit_scale":
self.assertAlmostEqual(
param.data.item(),
np.log(1 / 0.07),
delta=1e-3,
msg=f"Parameter {name} of model {model_class} seems not properly initialized",
)
elif name == "temp":
self.assertAlmostEqual(
param.data.item(),
0.07,
delta=1e-3,
msg=f"Parameter {name} of model {model_class} seems not properly initialized",
)
else:
self.assertIn(
((param.data.mean() * 1e9).round() / 1e9).item(),
[0.0, 1.0],
msg=f"Parameter {name} of model {model_class} seems not properly initialized",
)
# We will verify our results on an image of cute cats
def prepare_img():
url = "https://huggingface.co/hf-internal-testing/blip-test-image/resolve/main/demo.jpg"
image = Image.open(requests.get(url, stream=True).raw)
return image
@require_vision
@require_torch
@slow
class Blip2ModelIntegrationTest(unittest.TestCase):
def test_inference_opt(self):
processor = Blip2Processor.from_pretrained("Salesforce/blip2-opt-2.7b")
model = Blip2ForConditionalGeneration.from_pretrained(
"Salesforce/blip2-opt-2.7b", torch_dtype=torch.float16
).to(torch_device)
# prepare image
image = prepare_img()
inputs = processor(images=image, return_tensors="pt").to(torch_device, dtype=torch.float16)
predictions = model.generate(**inputs)
generated_text = processor.batch_decode(predictions, skip_special_tokens=True)[0].strip()
# Test output
self.assertEqual(predictions[0].tolist(), [2, 102, 693, 2828, 15, 5, 4105, 19, 10, 2335, 50118])
self.assertEqual("a woman sitting on the beach with a dog", generated_text)
# image and context
prompt = "Question: which city is this? Answer:"
inputs = processor(images=image, text=prompt, return_tensors="pt").to(torch_device, dtype=torch.float16)
# max_length for BLIP includes prompt length from now on, use max_new_tokens
predictions = model.generate(**inputs, max_new_tokens=11)
generated_text = processor.batch_decode(predictions, skip_special_tokens=True)[0].strip()
# Test output
self.assertEqual(
predictions[0].tolist(),
[2, 24, 18, 45, 10, 343, 6, 24, 18, 10, 4105, 50118],
)
self.assertEqual(generated_text, "it's not a city, it's a beach")
def test_inference_interpolate_pos_encoding(self):
processor = Blip2Processor.from_pretrained("Salesforce/blip2-opt-2.7b")
model = Blip2ForConditionalGeneration.from_pretrained(
"Salesforce/blip2-opt-2.7b", torch_dtype=torch.float16
).to(torch_device)
processor.image_processor.size = {"height": 500, "width": 500}
image = prepare_img()
inputs = processor(images=image, return_tensors="pt").to(torch_device)
predictions = model.generate(**inputs, interpolate_pos_encoding=True)
generated_text = processor.batch_decode(predictions, skip_special_tokens=True)[0].strip()
self.assertEqual(predictions[0].tolist(), [2, 102, 693, 8, 2335, 15, 5, 4105, 50118])
self.assertEqual(generated_text, "a woman and dog on the beach")
def test_inference_opt_batched_beam_search(self):
processor = Blip2Processor.from_pretrained("Salesforce/blip2-opt-2.7b")
model = Blip2ForConditionalGeneration.from_pretrained(
"Salesforce/blip2-opt-2.7b", torch_dtype=torch.float16
).to(torch_device)
# prepare image
image = prepare_img()
inputs = processor(images=[image, image], return_tensors="pt").to(torch_device, dtype=torch.float16)
predictions = model.generate(**inputs, num_beams=2)
# Test output (in this case, slightly different from greedy search)
self.assertEqual(predictions[0].tolist(), [2, 102, 693, 2828, 15, 5, 4105, 19, 69, 2335, 50118])
self.assertEqual(predictions[1].tolist(), [2, 102, 693, 2828, 15, 5, 4105, 19, 69, 2335, 50118])
def test_inference_t5(self):
processor = Blip2Processor.from_pretrained("Salesforce/blip2-flan-t5-xl")
model = Blip2ForConditionalGeneration.from_pretrained(
"Salesforce/blip2-flan-t5-xl", torch_dtype=torch.float16
).to(torch_device)
# prepare image
image = prepare_img()
inputs = processor(images=image, return_tensors="pt").to(torch_device, dtype=torch.float16)
predictions = model.generate(**inputs)
generated_text = processor.batch_decode(predictions, skip_special_tokens=True)[0].strip()
# Test output
self.assertEqual(predictions[0].tolist(), [0, 2335, 1556, 28, 1782, 30, 8, 2608, 1])
self.assertEqual("woman playing with dog on the beach", generated_text)
# image and context
prompt = "Question: which city is this? Answer:"
inputs = processor(images=image, text=prompt, return_tensors="pt").to(torch_device, dtype=torch.float16)
predictions = model.generate(**inputs)
generated_text = processor.batch_decode(predictions, skip_special_tokens=True)[0].strip()
# Test output
self.assertEqual(
predictions[0].tolist(),
[0, 3, 7, 152, 67, 839, 1],
)
self.assertEqual(generated_text, "san diego")
def test_inference_t5_batched_beam_search(self):
processor = Blip2Processor.from_pretrained("Salesforce/blip2-flan-t5-xl")
model = Blip2ForConditionalGeneration.from_pretrained(
"Salesforce/blip2-flan-t5-xl", torch_dtype=torch.float16
).to(torch_device)
# prepare image
image = prepare_img()
inputs = processor(images=[image, image], return_tensors="pt").to(torch_device, dtype=torch.float16)
predictions = model.generate(**inputs, num_beams=2)
# Test output (in this case, slightly different from greedy search)
self.assertEqual(predictions[0].tolist(), [0, 2335, 1556, 28, 1782, 30, 8, 2608, 1])
self.assertEqual(predictions[1].tolist(), [0, 2335, 1556, 28, 1782, 30, 8, 2608, 1])
@require_torch_multi_accelerator
def test_inference_opt_multi_accelerator(self):
processor = Blip2Processor.from_pretrained("Salesforce/blip2-opt-2.7b")
model = Blip2ForConditionalGeneration.from_pretrained(
"Salesforce/blip2-opt-2.7b", torch_dtype=torch.float16, device_map="balanced"
)
# prepare image
image = prepare_img()
inputs = processor(images=image, return_tensors="pt").to(0, dtype=torch.float16)
predictions = model.generate(**inputs)
generated_text = processor.batch_decode(predictions, skip_special_tokens=True)[0].strip()
# Test output
self.assertEqual(predictions[0].tolist(), [2, 102, 693, 2828, 15, 5, 4105, 19, 10, 2335, 50118])
self.assertEqual("a woman sitting on the beach with a dog", generated_text)
# image and context
prompt = "Question: which city is this? Answer:"
inputs = processor(images=image, text=prompt, return_tensors="pt").to(0, dtype=torch.float16)
predictions = model.generate(**inputs, max_new_tokens=11)
generated_text = processor.batch_decode(predictions, skip_special_tokens=True)[0].strip()
# Test output
self.assertEqual(
predictions[0].tolist(),
[2, 24, 18, 45, 10, 343, 6, 24, 18, 10, 4105, 50118],
)
self.assertEqual(generated_text, "it's not a city, it's a beach")
@require_torch_multi_accelerator
def test_inference_t5_multi_accelerator(self):
processor = Blip2Processor.from_pretrained("Salesforce/blip2-flan-t5-xl")
device_map = device_map = {
"query_tokens": 0,
"vision_model": 0,
"language_model": 1,
"language_projection": 0,
"qformer": 0,
}
model = Blip2ForConditionalGeneration.from_pretrained(
"Salesforce/blip2-flan-t5-xl", torch_dtype=torch.float16, device_map=device_map
)
# prepare image
image = prepare_img()
inputs = processor(images=image, return_tensors="pt").to(f"{torch_device}:0", dtype=torch.float16)
predictions = model.generate(**inputs)
generated_text = processor.batch_decode(predictions, skip_special_tokens=True)[0].strip()
# Test output
self.assertEqual(predictions[0].tolist(), [0, 2335, 1556, 28, 1782, 30, 8, 2608, 1])
self.assertEqual("woman playing with dog on the beach", generated_text)
# image and context
prompt = "Question: which city is this? Answer:"
inputs = processor(images=image, text=prompt, return_tensors="pt").to(f"{torch_device}:0", dtype=torch.float16)
predictions = model.generate(**inputs)
generated_text = processor.batch_decode(predictions, skip_special_tokens=True)[0].strip()
# Test output
self.assertEqual(
predictions[0].tolist(),
[0, 3, 7, 152, 67, 839, 1],
)
self.assertEqual(generated_text, "san diego")
def test_expansion_in_processing(self):
processor = Blip2Processor.from_pretrained("Salesforce/blip2-opt-2.7b")
model = Blip2ForConditionalGeneration.from_pretrained(
"Salesforce/blip2-opt-2.7b", torch_dtype=torch.float16
).to(torch_device)
image = prepare_img()
prompt = "Question: which city is this? Answer:"
# Make sure we will go the legacy path by setting these args to None
processor.num_query_tokens = None
model.config.image_token_index = None
inputs = processor(images=image, text=prompt, return_tensors="pt").to(torch_device, dtype=torch.float16)
predictions = model.generate(**inputs, do_sample=False, max_new_tokens=15)
generated_text = processor.batch_decode(predictions, skip_special_tokens=True)[0].strip()
# Add args to the config to trigger new logic when inputs are expanded in processing file
processor.num_query_tokens = model.config.num_query_tokens
processor.tokenizer.add_special_tokens({"additional_special_tokens": ["<image>"]})
model.config.image_token_index = len(processor.tokenizer) - 1
model.resize_token_embeddings(processor.tokenizer.vocab_size, pad_to_multiple_of=64)
# Generate again with new inputs
inputs = processor(images=image, text=prompt, return_tensors="pt").to(torch_device, dtype=torch.float16)
predictions_expanded = model.generate(**inputs, do_sample=False, max_new_tokens=15)
generated_text_expanded = processor.batch_decode(predictions_expanded, skip_special_tokens=True)[0].strip()
self.assertTrue(generated_text_expanded == generated_text)
@require_torch_gpu
def test_inference_itm(self):
model_name = "Salesforce/blip2-itm-vit-g"
processor = Blip2Processor.from_pretrained(model_name)
model = Blip2ForImageTextRetrieval.from_pretrained(model_name).to(torch_device)
image = prepare_img()
text = "A woman and her dog sitting in a beach"
inputs = processor(images=image, text=text, return_tensors="pt").to(torch_device)
# forward pass
out_itm = model(**inputs, use_image_text_matching_head=True)
out = model(**inputs)
# verify
expected_scores = torch.Tensor([[0.0238, 0.9762]])
self.assertTrue(torch.allclose(torch.nn.Softmax()(out_itm[0].cpu()), expected_scores, rtol=1e-3, atol=1e-3))
self.assertTrue(torch.allclose(out[0].cpu(), torch.Tensor([[0.4406]]), rtol=1e-3, atol=1e-3))
@require_torch_gpu
@require_torch_fp16
def test_inference_itm_fp16(self):
model_name = "Salesforce/blip2-itm-vit-g"
processor = Blip2Processor.from_pretrained(model_name)
model = Blip2ForImageTextRetrieval.from_pretrained(model_name, torch_dtype=torch.float16).to(torch_device)
image = prepare_img()
text = "A woman and her dog sitting in a beach"
inputs = processor(images=image, text=text, return_tensors="pt").to(torch_device, dtype=torch.float16)
# forward pass
out_itm = model(**inputs, use_image_text_matching_head=True)
out = model(**inputs)
# verify
expected_scores = torch.Tensor([[0.0239, 0.9761]])
self.assertTrue(
torch.allclose(torch.nn.Softmax()(out_itm[0].cpu().float()), expected_scores, rtol=1e-3, atol=1e-3)
)
self.assertTrue(torch.allclose(out[0].cpu().float(), torch.Tensor([[0.4406]]), rtol=1e-3, atol=1e-3))
@require_torch_gpu
@require_torch_fp16
def test_inference_vision_with_projection_fp16(self):
model_name = "Salesforce/blip2-itm-vit-g"
processor = Blip2Processor.from_pretrained(model_name)
model = Blip2VisionModelWithProjection.from_pretrained(model_name, torch_dtype=torch.float16).to(torch_device)
image = prepare_img()
inputs = processor(images=image, return_tensors="pt").to(torch_device, dtype=torch.float16)
# forward pass
out = model(**inputs)
# verify
expected_image_embeds = [
-0.093994140625,
-0.075927734375,
0.031890869140625,
0.053009033203125,
0.0352783203125,
-0.01190185546875,
]
self.assertTrue(np.allclose(out.image_embeds[0][0][:6].tolist(), expected_image_embeds, atol=1e-3))
@require_torch_gpu
@require_torch_fp16
def test_inference_text_with_projection_fp16(self):
model_name = "Salesforce/blip2-itm-vit-g"
processor = Blip2Processor.from_pretrained(model_name)
model = Blip2TextModelWithProjection.from_pretrained(model_name, torch_dtype=torch.float16).to(torch_device)
inputs = processor(text="a woman sitting on the beach with a dog", padding=True, return_tensors="pt").to(
torch_device
)
# forward pass
out = model(**inputs)
# verify
expected_text_embeds = [
-0.1082763671875,
0.053192138671875,
-0.02825927734375,
0.0169830322265625,
0.08648681640625,
-0.04656982421875,
]
self.assertTrue(np.allclose(out.text_embeds[0][0][:6].tolist(), expected_text_embeds, atol=1e-3))
|
transformers/tests/models/blip_2/test_modeling_blip_2.py/0
|
{
"file_path": "transformers/tests/models/blip_2/test_modeling_blip_2.py",
"repo_id": "transformers",
"token_count": 30439
}
| 403
|
# coding=utf-8
# Copyright 2018 HuggingFace Inc. team.
#
# 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.
import tempfile
import unittest
from transformers import AddedToken, CamembertTokenizer, CamembertTokenizerFast
from transformers.testing_utils import get_tests_dir, require_sentencepiece, require_tokenizers, slow
from transformers.utils import is_torch_available
from ...test_tokenization_common import TokenizerTesterMixin
SAMPLE_VOCAB = get_tests_dir("fixtures/test_sentencepiece.model")
SAMPLE_BPE_VOCAB = get_tests_dir("fixtures/test_sentencepiece_bpe.model")
FRAMEWORK = "pt" if is_torch_available() else "tf"
@require_sentencepiece
@require_tokenizers
class CamembertTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
from_pretrained_id = "almanach/camembert-base"
tokenizer_class = CamembertTokenizer
rust_tokenizer_class = CamembertTokenizerFast
test_rust_tokenizer = True
test_sentencepiece = True
def setUp(self):
super().setUp()
# We have a SentencePiece fixture for testing
tokenizer = CamembertTokenizer(SAMPLE_VOCAB)
tokenizer.save_pretrained(self.tmpdirname)
@unittest.skip(
"Token maps are not equal because someone set the probability of ('<unk>NOTUSED', -100), so it's never encoded for fast"
)
def test_special_tokens_map_equal(self):
return
def test_convert_token_and_id(self):
"""Test ``_convert_token_to_id`` and ``_convert_id_to_token``."""
token = "<pad>"
token_id = 1 # 1 is the offset id, but in the spm vocab it's 3
self.assertEqual(self.get_tokenizer().convert_tokens_to_ids(token), token_id)
self.assertEqual(self.get_tokenizer().convert_ids_to_tokens(token_id), token)
def test_get_vocab(self):
vocab_keys = list(self.get_tokenizer().get_vocab().keys())
self.assertEqual(vocab_keys[0], "<s>NOTUSED")
self.assertEqual(vocab_keys[1], "<pad>")
self.assertEqual(vocab_keys[-1], "<mask>")
self.assertEqual(len(vocab_keys), 1_005)
def test_vocab_size(self):
self.assertEqual(self.get_tokenizer().vocab_size, 1_000)
def test_rust_and_python_bpe_tokenizers(self):
tokenizer = CamembertTokenizer(SAMPLE_BPE_VOCAB)
tokenizer.save_pretrained(self.tmpdirname)
rust_tokenizer = CamembertTokenizerFast.from_pretrained(self.tmpdirname)
sequence = "I was born in 92000, and this is falsé."
ids = tokenizer.encode(sequence)
rust_ids = rust_tokenizer.encode(sequence)
self.assertListEqual(ids, rust_ids)
ids = tokenizer.encode(sequence, add_special_tokens=False)
rust_ids = rust_tokenizer.encode(sequence, add_special_tokens=False)
self.assertListEqual(ids, rust_ids)
# <unk> tokens are not the same for `rust` than for `slow`.
# Because spm gives back raw token instead of `unk` in EncodeAsPieces
# tokens = tokenizer.tokenize(sequence)
tokens = tokenizer.convert_ids_to_tokens(ids)
rust_tokens = rust_tokenizer.tokenize(sequence)
self.assertListEqual(tokens, rust_tokens)
def test_rust_and_python_full_tokenizers(self):
if not self.test_rust_tokenizer:
self.skipTest(reason="test_rust_tokenizer is set to False")
tokenizer = self.get_tokenizer()
rust_tokenizer = self.get_rust_tokenizer()
sequence = "I was born in 92000, and this is falsé."
tokens = tokenizer.tokenize(sequence)
rust_tokens = rust_tokenizer.tokenize(sequence)
self.assertListEqual(tokens, rust_tokens)
ids = tokenizer.encode(sequence, add_special_tokens=False)
rust_ids = rust_tokenizer.encode(sequence, add_special_tokens=False)
self.assertListEqual(ids, rust_ids)
rust_tokenizer = self.get_rust_tokenizer()
ids = tokenizer.encode(sequence)
rust_ids = rust_tokenizer.encode(sequence)
self.assertListEqual(ids, rust_ids)
@slow
def test_tokenizer_integration(self):
expected_encoding = {'input_ids': [[5, 54, 7196, 297, 30, 23, 776, 18, 11, 3215, 3705, 8252, 22, 3164, 1181, 2116, 29, 16, 813, 25, 791, 3314, 20, 3446, 38, 27575, 120, 6, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [5, 468, 17, 11, 9088, 20, 1517, 8, 22804, 18818, 10, 38, 629, 607, 607, 142, 19, 7196, 867, 56, 10326, 24, 2267, 20, 416, 5072, 15612, 233, 734, 7, 2399, 27, 16, 3015, 1649, 7, 24, 20, 4338, 2399, 27, 13, 3400, 14, 13, 6189, 8, 930, 9, 6]], 'attention_mask': [[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]]} # fmt: skip
# camembert is a french model. So we also use french texts.
sequences = [
"Le transformeur est un modèle d'apprentissage profond introduit en 2017, "
"utilisé principalement dans le domaine du traitement automatique des langues (TAL).",
"À l'instar des réseaux de neurones récurrents (RNN), les transformeurs sont conçus "
"pour gérer des données séquentielles, telles que le langage naturel, pour des tâches "
"telles que la traduction et la synthèse de texte.",
]
self.tokenizer_integration_test_util(
expected_encoding=expected_encoding,
model_name="almanach/camembert-base",
revision="3a0641d9a1aeb7e848a74299e7e4c4bca216b4cf",
sequences=sequences,
)
# Overwritten because we have to use from slow (online pretrained is wrong, the tokenizer.json has a whole)
def test_added_tokens_serialization(self):
self.maxDiff = None
# Utility to test the added vocab
def _test_added_vocab_and_eos(expected, tokenizer_class, expected_eos, temp_dir):
tokenizer = tokenizer_class.from_pretrained(temp_dir)
self.assertTrue(str(expected_eos) not in tokenizer.additional_special_tokens)
self.assertIn(new_eos, tokenizer.added_tokens_decoder.values())
self.assertEqual(tokenizer.added_tokens_decoder[tokenizer.eos_token_id], new_eos)
self.assertTrue(all(item in tokenizer.added_tokens_decoder.items() for item in expected.items()))
return tokenizer
new_eos = AddedToken("[NEW_EOS]", rstrip=False, lstrip=True, normalized=False)
for tokenizer, pretrained_name, kwargs in self.tokenizers_list:
with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name})"):
# Load a slow tokenizer from the hub, init with the new token for fast to also include it
tokenizer = self.tokenizer_class.from_pretrained(pretrained_name, eos_token=new_eos)
EXPECTED_ADDED_TOKENS_DECODER = tokenizer.added_tokens_decoder
with self.subTest("Hub -> Slow: Test loading a slow tokenizer from the hub)"):
self.assertEqual(tokenizer._eos_token, new_eos)
self.assertIn(new_eos, list(tokenizer.added_tokens_decoder.values()))
with tempfile.TemporaryDirectory() as tmp_dir_2:
tokenizer.save_pretrained(tmp_dir_2)
with self.subTest(
"Hub -> Slow -> Slow: Test saving this slow tokenizer and reloading it in the fast class"
):
_test_added_vocab_and_eos(
EXPECTED_ADDED_TOKENS_DECODER, self.tokenizer_class, new_eos, tmp_dir_2
)
if self.rust_tokenizer_class is not None:
with self.subTest(
"Hub -> Slow -> Fast: Test saving this slow tokenizer and reloading it in the fast class"
):
tokenizer_fast = _test_added_vocab_and_eos(
EXPECTED_ADDED_TOKENS_DECODER, self.rust_tokenizer_class, new_eos, tmp_dir_2
)
with tempfile.TemporaryDirectory() as tmp_dir_3:
tokenizer_fast.save_pretrained(tmp_dir_3)
with self.subTest(
"Hub -> Slow -> Fast -> Fast: Test saving this fast tokenizer and reloading it in the fast class"
):
_test_added_vocab_and_eos(
EXPECTED_ADDED_TOKENS_DECODER, self.rust_tokenizer_class, new_eos, tmp_dir_3
)
with self.subTest(
"Hub -> Slow -> Fast -> Slow: Test saving this slow tokenizer and reloading it in the slow class"
):
_test_added_vocab_and_eos(
EXPECTED_ADDED_TOKENS_DECODER, self.rust_tokenizer_class, new_eos, tmp_dir_3
)
with self.subTest("Hub -> Fast: Test loading a fast tokenizer from the hub)"):
if self.rust_tokenizer_class is not None:
tokenizer_fast = self.rust_tokenizer_class.from_pretrained(
pretrained_name, eos_token=new_eos, from_slow=True
)
self.assertEqual(tokenizer_fast._eos_token, new_eos)
self.assertIn(new_eos, list(tokenizer_fast.added_tokens_decoder.values()))
# We can't test the following because for BC we kept the default rstrip lstrip in slow not fast. Will comment once normalization is alright
with self.subTest("Hub -> Fast == Hub -> Slow: make sure slow and fast tokenizer match"):
with self.subTest("Hub -> Fast == Hub -> Slow: make sure slow and fast tokenizer match"):
self.assertTrue(
all(
item in tokenizer.added_tokens_decoder.items()
for item in EXPECTED_ADDED_TOKENS_DECODER.items()
)
)
EXPECTED_ADDED_TOKENS_DECODER = tokenizer_fast.added_tokens_decoder
with tempfile.TemporaryDirectory() as tmp_dir_4:
tokenizer_fast.save_pretrained(tmp_dir_4)
with self.subTest("Hub -> Fast -> Fast: saving Fast1 locally and loading"):
_test_added_vocab_and_eos(
EXPECTED_ADDED_TOKENS_DECODER, self.rust_tokenizer_class, new_eos, tmp_dir_4
)
with self.subTest("Hub -> Fast -> Slow: saving Fast1 locally and loading"):
_test_added_vocab_and_eos(
EXPECTED_ADDED_TOKENS_DECODER, self.tokenizer_class, new_eos, tmp_dir_4
)
|
transformers/tests/models/camembert/test_tokenization_camembert.py/0
|
{
"file_path": "transformers/tests/models/camembert/test_tokenization_camembert.py",
"repo_id": "transformers",
"token_count": 5751
}
| 404
|
# coding=utf-8
# Copyright 2024 HuggingFace Inc.
#
# 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.
"""Tests for the dac feature extractor."""
import itertools
import random
import unittest
import numpy as np
from transformers import DacFeatureExtractor
from transformers.testing_utils import require_torch
from transformers.utils.import_utils import is_torch_available
from ...test_sequence_feature_extraction_common import SequenceFeatureExtractionTestMixin
if is_torch_available():
import torch
global_rng = random.Random()
# Copied from tests.models.whisper.test_feature_extraction_whisper.floats_list
def floats_list(shape, scale=1.0, rng=None, name=None):
"""Creates a random float32 tensor"""
if rng is None:
rng = global_rng
values = []
for batch_idx in range(shape[0]):
values.append([])
for _ in range(shape[1]):
values[-1].append(rng.random() * scale)
return values
@require_torch
# Copied from transformers.tests.encodec.test_feature_extraction_dac.EncodecFeatureExtractionTester with Encodec->Dac
class DacFeatureExtractionTester(unittest.TestCase):
# Ignore copy
def __init__(
self,
parent,
batch_size=7,
min_seq_length=400,
max_seq_length=2000,
feature_size=1,
padding_value=0.0,
sampling_rate=16000,
hop_length=512,
):
self.parent = parent
self.batch_size = batch_size
self.min_seq_length = min_seq_length
self.max_seq_length = max_seq_length
self.hop_length = hop_length
self.seq_length_diff = (self.max_seq_length - self.min_seq_length) // (self.batch_size - 1)
self.feature_size = feature_size
self.padding_value = padding_value
self.sampling_rate = sampling_rate
# Ignore copy
def prepare_feat_extract_dict(self):
return {
"feature_size": self.feature_size,
"padding_value": self.padding_value,
"sampling_rate": self.sampling_rate,
"hop_length": self.hop_length,
}
def prepare_inputs_for_common(self, equal_length=False, numpify=False):
def _flatten(list_of_lists):
return list(itertools.chain(*list_of_lists))
if equal_length:
audio_inputs = floats_list((self.batch_size, self.max_seq_length))
else:
# make sure that inputs increase in size
audio_inputs = [
_flatten(floats_list((x, self.feature_size)))
for x in range(self.min_seq_length, self.max_seq_length, self.seq_length_diff)
]
if numpify:
audio_inputs = [np.asarray(x) for x in audio_inputs]
return audio_inputs
@require_torch
# Copied from transformers.tests.encodec.test_feature_extraction_dac.EnCodecFeatureExtractionTest with Encodec->Dac
class DacFeatureExtractionTest(SequenceFeatureExtractionTestMixin, unittest.TestCase):
feature_extraction_class = DacFeatureExtractor
def setUp(self):
self.feat_extract_tester = DacFeatureExtractionTester(self)
def test_call(self):
# Tests that all call wrap to encode_plus and batch_encode_plus
feat_extract = self.feature_extraction_class(**self.feat_extract_tester.prepare_feat_extract_dict())
# create three inputs of length 800, 1000, and 1200
audio_inputs = [floats_list((1, x))[0] for x in range(800, 1400, 200)]
np_audio_inputs = [np.asarray(audio_input) for audio_input in audio_inputs]
# Test not batched input
encoded_sequences_1 = feat_extract(audio_inputs[0], return_tensors="np").input_values
encoded_sequences_2 = feat_extract(np_audio_inputs[0], return_tensors="np").input_values
self.assertTrue(np.allclose(encoded_sequences_1, encoded_sequences_2, atol=1e-3))
# Test batched
encoded_sequences_1 = feat_extract(audio_inputs, padding=True, return_tensors="np").input_values
encoded_sequences_2 = feat_extract(np_audio_inputs, padding=True, return_tensors="np").input_values
for enc_seq_1, enc_seq_2 in zip(encoded_sequences_1, encoded_sequences_2):
self.assertTrue(np.allclose(enc_seq_1, enc_seq_2, atol=1e-3))
def test_double_precision_pad(self):
feature_extractor = self.feature_extraction_class(**self.feat_extract_tester.prepare_feat_extract_dict())
np_audio_inputs = np.random.rand(100).astype(np.float64)
py_audio_inputs = np_audio_inputs.tolist()
for inputs in [py_audio_inputs, np_audio_inputs]:
np_processed = feature_extractor.pad([{"input_values": inputs}], return_tensors="np")
self.assertTrue(np_processed.input_values.dtype == np.float32)
pt_processed = feature_extractor.pad([{"input_values": inputs}], return_tensors="pt")
self.assertTrue(pt_processed.input_values.dtype == torch.float32)
def _load_datasamples(self, num_samples):
from datasets import load_dataset
ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
# automatic decoding with librispeech
audio_samples = ds.sort("id").select(range(num_samples))[:num_samples]["audio"]
return [x["array"] for x in audio_samples]
def test_integration(self):
# fmt: off
EXPECTED_INPUT_VALUES = torch.tensor(
[ 2.3803711e-03, 2.0751953e-03, 1.9836426e-03, 2.1057129e-03,
1.6174316e-03, 3.0517578e-04, 9.1552734e-05, 3.3569336e-04,
9.7656250e-04, 1.8310547e-03, 2.0141602e-03, 2.1057129e-03,
1.7395020e-03, 4.5776367e-04, -3.9672852e-04, 4.5776367e-04,
1.0070801e-03, 9.1552734e-05, 4.8828125e-04, 1.1596680e-03,
7.3242188e-04, 9.4604492e-04, 1.8005371e-03, 1.8310547e-03,
8.8500977e-04, 4.2724609e-04, 4.8828125e-04, 7.3242188e-04,
1.0986328e-03, 2.1057129e-03]
)
# fmt: on
input_audio = self._load_datasamples(1)
feature_extractor = DacFeatureExtractor()
input_values = feature_extractor(input_audio, return_tensors="pt")["input_values"]
self.assertEqual(input_values.shape, (1, 1, 93696))
self.assertTrue(torch.allclose(input_values[0, 0, :30], EXPECTED_INPUT_VALUES, atol=1e-4))
audio_input_end = torch.tensor(input_audio[0][-30:], dtype=torch.float32)
self.assertTrue(torch.allclose(input_values[0, 0, -46:-16], audio_input_end, atol=1e-4))
# Ignore copy
@unittest.skip("The DAC model doesn't support stereo logic")
def test_integration_stereo(self):
pass
# Ignore copy
def test_truncation_and_padding(self):
input_audio = self._load_datasamples(2)
# would be easier if the stride was like
feature_extractor = DacFeatureExtractor()
# pad and trunc raise an error ?
with self.assertRaisesRegex(
ValueError,
"^Both padding and truncation were set. Make sure you only set one.$",
):
truncated_outputs = feature_extractor(
input_audio, padding="max_length", truncation=True, return_tensors="pt"
).input_values
# force truncate to max_length
truncated_outputs = feature_extractor(
input_audio, truncation=True, max_length=48000, return_tensors="pt"
).input_values
self.assertEqual(truncated_outputs.shape, (2, 1, 48128))
# pad:
padded_outputs = feature_extractor(input_audio, padding=True, return_tensors="pt").input_values
self.assertEqual(padded_outputs.shape, (2, 1, 93696))
# force pad to max length
truncated_outputs = feature_extractor(
input_audio, padding="max_length", max_length=100000, return_tensors="pt"
).input_values
self.assertEqual(truncated_outputs.shape, (2, 1, 100352))
# force no pad
with self.assertRaisesRegex(
ValueError,
"^Unable to create tensor, you should probably activate padding with 'padding=True' to have batched tensors with the same length.$",
):
truncated_outputs = feature_extractor(input_audio, padding=False, return_tensors="pt").input_values
truncated_outputs = feature_extractor(input_audio[0], padding=False, return_tensors="pt").input_values
self.assertEqual(truncated_outputs.shape, (1, 1, 93680))
|
transformers/tests/models/dac/test_feature_extraction_dac.py/0
|
{
"file_path": "transformers/tests/models/dac/test_feature_extraction_dac.py",
"repo_id": "transformers",
"token_count": 3868
}
| 405
|
# coding=utf-8
# Copyright 2022 The HuggingFace Inc. team. 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.
"""Testing suite for the PyTorch Dinat model."""
import collections
import unittest
from transformers import DinatConfig
from transformers.testing_utils import require_natten, require_torch, require_vision, slow, torch_device
from transformers.utils import cached_property, is_torch_available, is_vision_available
from ...test_backbone_common import BackboneTesterMixin
from ...test_configuration_common import ConfigTester
from ...test_modeling_common import ModelTesterMixin, _config_zero_init, floats_tensor, ids_tensor
from ...test_pipeline_mixin import PipelineTesterMixin
if is_torch_available():
import torch
from torch import nn
from transformers import DinatBackbone, DinatForImageClassification, DinatModel
if is_vision_available():
from PIL import Image
from transformers import AutoImageProcessor
class DinatModelTester:
def __init__(
self,
parent,
batch_size=13,
image_size=64,
patch_size=4,
num_channels=3,
embed_dim=16,
depths=[1, 2, 1],
num_heads=[2, 4, 8],
kernel_size=3,
dilations=[[3], [1, 2], [1]],
mlp_ratio=2.0,
qkv_bias=True,
hidden_dropout_prob=0.0,
attention_probs_dropout_prob=0.0,
drop_path_rate=0.1,
hidden_act="gelu",
patch_norm=True,
initializer_range=0.02,
layer_norm_eps=1e-5,
is_training=True,
scope=None,
use_labels=True,
num_labels=10,
out_features=["stage1", "stage2"],
out_indices=[1, 2],
):
self.parent = parent
self.batch_size = batch_size
self.image_size = image_size
self.patch_size = patch_size
self.num_channels = num_channels
self.embed_dim = embed_dim
self.depths = depths
self.num_heads = num_heads
self.kernel_size = kernel_size
self.dilations = dilations
self.mlp_ratio = mlp_ratio
self.qkv_bias = qkv_bias
self.hidden_dropout_prob = hidden_dropout_prob
self.attention_probs_dropout_prob = attention_probs_dropout_prob
self.drop_path_rate = drop_path_rate
self.hidden_act = hidden_act
self.patch_norm = patch_norm
self.layer_norm_eps = layer_norm_eps
self.initializer_range = initializer_range
self.is_training = is_training
self.scope = scope
self.use_labels = use_labels
self.num_labels = num_labels
self.out_features = out_features
self.out_indices = out_indices
def prepare_config_and_inputs(self):
pixel_values = floats_tensor([self.batch_size, self.num_channels, self.image_size, self.image_size])
labels = None
if self.use_labels:
labels = ids_tensor([self.batch_size], self.num_labels)
config = self.get_config()
return config, pixel_values, labels
def get_config(self):
return DinatConfig(
num_labels=self.num_labels,
image_size=self.image_size,
patch_size=self.patch_size,
num_channels=self.num_channels,
embed_dim=self.embed_dim,
depths=self.depths,
num_heads=self.num_heads,
kernel_size=self.kernel_size,
dilations=self.dilations,
mlp_ratio=self.mlp_ratio,
qkv_bias=self.qkv_bias,
hidden_dropout_prob=self.hidden_dropout_prob,
attention_probs_dropout_prob=self.attention_probs_dropout_prob,
drop_path_rate=self.drop_path_rate,
hidden_act=self.hidden_act,
patch_norm=self.patch_norm,
layer_norm_eps=self.layer_norm_eps,
initializer_range=self.initializer_range,
out_features=self.out_features,
out_indices=self.out_indices,
)
def create_and_check_model(self, config, pixel_values, labels):
model = DinatModel(config=config)
model.to(torch_device)
model.eval()
result = model(pixel_values)
expected_height = expected_width = (config.image_size // config.patch_size) // (2 ** (len(config.depths) - 1))
expected_dim = int(config.embed_dim * 2 ** (len(config.depths) - 1))
self.parent.assertEqual(
result.last_hidden_state.shape, (self.batch_size, expected_height, expected_width, expected_dim)
)
def create_and_check_for_image_classification(self, config, pixel_values, labels):
model = DinatForImageClassification(config)
model.to(torch_device)
model.eval()
result = model(pixel_values, labels=labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_labels))
# test greyscale images
config.num_channels = 1
model = DinatForImageClassification(config)
model.to(torch_device)
model.eval()
pixel_values = floats_tensor([self.batch_size, 1, self.image_size, self.image_size])
result = model(pixel_values)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_labels))
def create_and_check_backbone(self, config, pixel_values, labels):
model = DinatBackbone(config=config)
model.to(torch_device)
model.eval()
result = model(pixel_values)
# verify hidden states
self.parent.assertEqual(len(result.feature_maps), len(config.out_features))
self.parent.assertListEqual(list(result.feature_maps[0].shape), [self.batch_size, model.channels[0], 16, 16])
# verify channels
self.parent.assertEqual(len(model.channels), len(config.out_features))
# verify backbone works with out_features=None
config.out_features = None
model = DinatBackbone(config=config)
model.to(torch_device)
model.eval()
result = model(pixel_values)
# verify feature maps
self.parent.assertEqual(len(result.feature_maps), 1)
self.parent.assertListEqual(list(result.feature_maps[0].shape), [self.batch_size, model.channels[-1], 4, 4])
# verify channels
self.parent.assertEqual(len(model.channels), 1)
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
config, pixel_values, labels = config_and_inputs
inputs_dict = {"pixel_values": pixel_values}
return config, inputs_dict
@require_natten
@require_torch
class DinatModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
all_model_classes = (
(
DinatModel,
DinatForImageClassification,
DinatBackbone,
)
if is_torch_available()
else ()
)
pipeline_model_mapping = (
{"image-feature-extraction": DinatModel, "image-classification": DinatForImageClassification}
if is_torch_available()
else {}
)
fx_compatible = False
test_torchscript = False
test_pruning = False
test_resize_embeddings = False
test_head_masking = False
def setUp(self):
self.model_tester = DinatModelTester(self)
self.config_tester = ConfigTester(
self, config_class=DinatConfig, embed_dim=37, common_properties=["patch_size", "num_channels"]
)
def test_config(self):
self.config_tester.run_common_tests()
def test_model(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_model(*config_and_inputs)
def test_for_image_classification(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_for_image_classification(*config_and_inputs)
def test_backbone(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_backbone(*config_and_inputs)
@unittest.skip(reason="Dinat does not use inputs_embeds")
def test_inputs_embeds(self):
pass
@unittest.skip(reason="Dinat does not use feedforward chunking")
def test_feed_forward_chunking(self):
pass
def test_model_get_set_embeddings(self):
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
model = model_class(config)
self.assertIsInstance(model.get_input_embeddings(), (nn.Module))
x = model.get_output_embeddings()
self.assertTrue(x is None or isinstance(x, nn.Linear))
def test_attention_outputs(self):
self.skipTest(reason="Dinat's attention operation is handled entirely by NATTEN.")
def check_hidden_states_output(self, inputs_dict, config, model_class, image_size):
model = model_class(config)
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(**self._prepare_for_class(inputs_dict, model_class))
hidden_states = outputs.hidden_states
expected_num_layers = getattr(
self.model_tester, "expected_num_hidden_layers", len(self.model_tester.depths) + 1
)
self.assertEqual(len(hidden_states), expected_num_layers)
# Dinat has a different seq_length
patch_size = (
config.patch_size
if isinstance(config.patch_size, collections.abc.Iterable)
else (config.patch_size, config.patch_size)
)
height = image_size[0] // patch_size[0]
width = image_size[1] // patch_size[1]
self.assertListEqual(
list(hidden_states[0].shape[-3:]),
[height, width, self.model_tester.embed_dim],
)
if model_class.__name__ != "DinatBackbone":
reshaped_hidden_states = outputs.reshaped_hidden_states
self.assertEqual(len(reshaped_hidden_states), expected_num_layers)
batch_size, num_channels, height, width = reshaped_hidden_states[0].shape
reshaped_hidden_states = (
reshaped_hidden_states[0].view(batch_size, num_channels, height, width).permute(0, 2, 3, 1)
)
self.assertListEqual(
list(reshaped_hidden_states.shape[-3:]),
[height, width, self.model_tester.embed_dim],
)
def test_hidden_states_output(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
image_size = (
self.model_tester.image_size
if isinstance(self.model_tester.image_size, collections.abc.Iterable)
else (self.model_tester.image_size, self.model_tester.image_size)
)
for model_class in self.all_model_classes:
inputs_dict["output_hidden_states"] = True
self.check_hidden_states_output(inputs_dict, config, model_class, image_size)
# check that output_hidden_states also work using config
del inputs_dict["output_hidden_states"]
config.output_hidden_states = True
self.check_hidden_states_output(inputs_dict, config, model_class, image_size)
@slow
def test_model_from_pretrained(self):
model_name = "shi-labs/dinat-mini-in1k-224"
model = DinatModel.from_pretrained(model_name)
self.assertIsNotNone(model)
def test_initialization(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
configs_no_init = _config_zero_init(config)
for model_class in self.all_model_classes:
model = model_class(config=configs_no_init)
for name, param in model.named_parameters():
if "embeddings" not in name and param.requires_grad:
self.assertIn(
((param.data.mean() * 1e9).round() / 1e9).item(),
[0.0, 1.0],
msg=f"Parameter {name} of model {model_class} seems not properly initialized",
)
@require_natten
@require_vision
@require_torch
class DinatModelIntegrationTest(unittest.TestCase):
@cached_property
def default_image_processor(self):
return AutoImageProcessor.from_pretrained("shi-labs/dinat-mini-in1k-224") if is_vision_available() else None
@slow
def test_inference_image_classification_head(self):
model = DinatForImageClassification.from_pretrained("shi-labs/dinat-mini-in1k-224").to(torch_device)
image_processor = self.default_image_processor
image = Image.open("./tests/fixtures/tests_samples/COCO/000000039769.png")
inputs = image_processor(images=image, return_tensors="pt").to(torch_device)
# forward pass
with torch.no_grad():
outputs = model(**inputs)
# verify the logits
expected_shape = torch.Size((1, 1000))
self.assertEqual(outputs.logits.shape, expected_shape)
expected_slice = torch.tensor([-0.1545, -0.7667, 0.4642]).to(torch_device)
self.assertTrue(torch.allclose(outputs.logits[0, :3], expected_slice, atol=1e-4))
@require_torch
@require_natten
class DinatBackboneTest(unittest.TestCase, BackboneTesterMixin):
all_model_classes = (DinatBackbone,) if is_torch_available() else ()
config_class = DinatConfig
def setUp(self):
self.model_tester = DinatModelTester(self)
|
transformers/tests/models/dinat/test_modeling_dinat.py/0
|
{
"file_path": "transformers/tests/models/dinat/test_modeling_dinat.py",
"repo_id": "transformers",
"token_count": 6137
}
| 406
|
# coding=utf-8
# Copyright 2020 Huggingface
#
# 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.
import tempfile
import unittest
from transformers import DPRConfig, is_torch_available
from transformers.testing_utils import require_torch, slow, torch_device
from ...test_configuration_common import ConfigTester
from ...test_modeling_common import ModelTesterMixin, ids_tensor, random_attention_mask
from ...test_pipeline_mixin import PipelineTesterMixin
if is_torch_available():
import torch
from transformers import DPRContextEncoder, DPRQuestionEncoder, DPRReader, DPRReaderTokenizer
class DPRModelTester:
def __init__(
self,
parent,
batch_size=13,
seq_length=7,
is_training=False,
use_input_mask=True,
use_token_type_ids=True,
use_labels=True,
vocab_size=99,
hidden_size=32,
num_hidden_layers=2,
num_attention_heads=4,
intermediate_size=37,
hidden_act="gelu",
hidden_dropout_prob=0.1,
attention_probs_dropout_prob=0.1,
max_position_embeddings=512,
type_vocab_size=16,
type_sequence_label_size=2,
initializer_range=0.02,
num_labels=3,
num_choices=4,
scope=None,
projection_dim=0,
):
self.parent = parent
self.batch_size = batch_size
self.seq_length = seq_length
self.is_training = is_training
self.use_input_mask = use_input_mask
self.use_token_type_ids = use_token_type_ids
self.use_labels = use_labels
self.vocab_size = vocab_size
self.hidden_size = hidden_size
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
self.intermediate_size = intermediate_size
self.hidden_act = hidden_act
self.hidden_dropout_prob = hidden_dropout_prob
self.attention_probs_dropout_prob = attention_probs_dropout_prob
self.max_position_embeddings = max_position_embeddings
self.type_vocab_size = type_vocab_size
self.type_sequence_label_size = type_sequence_label_size
self.initializer_range = initializer_range
self.num_labels = num_labels
self.num_choices = num_choices
self.scope = scope
self.projection_dim = projection_dim
def prepare_config_and_inputs(self):
input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
input_mask = None
if self.use_input_mask:
input_mask = random_attention_mask([self.batch_size, self.seq_length])
token_type_ids = None
if self.use_token_type_ids:
token_type_ids = ids_tensor([self.batch_size, self.seq_length], self.type_vocab_size)
sequence_labels = None
token_labels = None
choice_labels = None
if self.use_labels:
sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels)
choice_labels = ids_tensor([self.batch_size], self.num_choices)
config = self.get_config()
return config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
def get_config(self):
return DPRConfig(
projection_dim=self.projection_dim,
vocab_size=self.vocab_size,
hidden_size=self.hidden_size,
num_hidden_layers=self.num_hidden_layers,
num_attention_heads=self.num_attention_heads,
intermediate_size=self.intermediate_size,
hidden_act=self.hidden_act,
hidden_dropout_prob=self.hidden_dropout_prob,
attention_probs_dropout_prob=self.attention_probs_dropout_prob,
max_position_embeddings=self.max_position_embeddings,
type_vocab_size=self.type_vocab_size,
initializer_range=self.initializer_range,
)
def create_and_check_context_encoder(
self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
):
model = DPRContextEncoder(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids)
result = model(input_ids, token_type_ids=token_type_ids)
result = model(input_ids)
self.parent.assertEqual(result.pooler_output.shape, (self.batch_size, self.projection_dim or self.hidden_size))
def create_and_check_question_encoder(
self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
):
model = DPRQuestionEncoder(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids)
result = model(input_ids, token_type_ids=token_type_ids)
result = model(input_ids)
self.parent.assertEqual(result.pooler_output.shape, (self.batch_size, self.projection_dim or self.hidden_size))
def create_and_check_reader(
self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
):
model = DPRReader(config=config)
model.to(torch_device)
model.eval()
result = model(
input_ids,
attention_mask=input_mask,
)
self.parent.assertEqual(result.start_logits.shape, (self.batch_size, self.seq_length))
self.parent.assertEqual(result.end_logits.shape, (self.batch_size, self.seq_length))
self.parent.assertEqual(result.relevance_logits.shape, (self.batch_size,))
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
(
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
) = config_and_inputs
inputs_dict = {"input_ids": input_ids}
return config, inputs_dict
@require_torch
class DPRModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
all_model_classes = (
(
DPRContextEncoder,
DPRQuestionEncoder,
DPRReader,
)
if is_torch_available()
else ()
)
pipeline_model_mapping = {"feature-extraction": DPRQuestionEncoder} if is_torch_available() else {}
test_resize_embeddings = False
test_missing_keys = False # why?
test_pruning = False
test_head_masking = False
def setUp(self):
self.model_tester = DPRModelTester(self)
self.config_tester = ConfigTester(self, config_class=DPRConfig, hidden_size=37)
def test_config(self):
self.config_tester.run_common_tests()
def test_context_encoder_model(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_context_encoder(*config_and_inputs)
def test_question_encoder_model(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_question_encoder(*config_and_inputs)
def test_reader_model(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_reader(*config_and_inputs)
def test_init_changed_config(self):
config = self.model_tester.prepare_config_and_inputs()[0]
model = DPRQuestionEncoder(config=config)
model.to(torch_device)
model.eval()
with tempfile.TemporaryDirectory() as tmp_dirname:
model.save_pretrained(tmp_dirname)
model = DPRQuestionEncoder.from_pretrained(tmp_dirname, projection_dim=512)
self.assertIsNotNone(model)
@slow
def test_model_from_pretrained(self):
model_name = "facebook/dpr-ctx_encoder-single-nq-base"
model = DPRContextEncoder.from_pretrained(model_name)
self.assertIsNotNone(model)
model_name = "facebook/dpr-ctx_encoder-single-nq-base"
model = DPRContextEncoder.from_pretrained(model_name)
self.assertIsNotNone(model)
model_name = "facebook/dpr-ctx_encoder-single-nq-base"
model = DPRQuestionEncoder.from_pretrained(model_name)
self.assertIsNotNone(model)
model_name = "facebook/dpr-ctx_encoder-single-nq-base"
model = DPRReader.from_pretrained(model_name)
self.assertIsNotNone(model)
@require_torch
class DPRModelIntegrationTest(unittest.TestCase):
@slow
def test_inference_no_head(self):
model = DPRQuestionEncoder.from_pretrained("facebook/dpr-question_encoder-single-nq-base", return_dict=False)
model.to(torch_device)
input_ids = torch.tensor(
[[101, 7592, 1010, 2003, 2026, 3899, 10140, 1029, 102]], dtype=torch.long, device=torch_device
) # [CLS] hello, is my dog cute? [SEP]
output = model(input_ids)[0] # embedding shape = (1, 768)
# compare the actual values for a slice.
expected_slice = torch.tensor(
[
[
0.03236253,
0.12753335,
0.16818509,
0.00279786,
0.3896933,
0.24264945,
0.2178971,
-0.02335227,
-0.08481959,
-0.14324117,
]
],
dtype=torch.float,
device=torch_device,
)
self.assertTrue(torch.allclose(output[:, :10], expected_slice, atol=1e-4))
@slow
def test_reader_inference(self):
tokenizer = DPRReaderTokenizer.from_pretrained("facebook/dpr-reader-single-nq-base")
model = DPRReader.from_pretrained("facebook/dpr-reader-single-nq-base")
model.to(torch_device)
encoded_inputs = tokenizer(
questions="What is love ?",
titles="Haddaway",
texts="What Is Love is a song recorded by the artist Haddaway",
padding=True,
return_tensors="pt",
)
encoded_inputs.to(torch_device)
outputs = model(**encoded_inputs)
# compare the actual values for a slice.
expected_start_logits = torch.tensor(
[[-10.3005, -10.7765, -11.4872, -11.6841, -11.9312, -10.3002, -9.8544, -11.7378, -12.0821, -10.2975]],
dtype=torch.float,
device=torch_device,
)
expected_end_logits = torch.tensor(
[[-11.0684, -11.7041, -11.5397, -10.3465, -10.8791, -6.8443, -11.9959, -11.0364, -10.0096, -6.8405]],
dtype=torch.float,
device=torch_device,
)
self.assertTrue(torch.allclose(outputs.start_logits[:, :10], expected_start_logits, atol=1e-4))
self.assertTrue(torch.allclose(outputs.end_logits[:, :10], expected_end_logits, atol=1e-4))
|
transformers/tests/models/dpr/test_modeling_dpr.py/0
|
{
"file_path": "transformers/tests/models/dpr/test_modeling_dpr.py",
"repo_id": "transformers",
"token_count": 5288
}
| 407
|
# coding=utf-8
# Copyright 2023 The HuggingFace Team. 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.
import math
import unittest
from parameterized import parameterized
from transformers import GPTBigCodeConfig, is_torch_available
from transformers.testing_utils import require_torch, slow, torch_device
from ...generation.test_utils import GenerationTesterMixin
from ...test_configuration_common import ConfigTester
from ...test_modeling_common import ModelTesterMixin, floats_tensor, ids_tensor, random_attention_mask
from ...test_pipeline_mixin import PipelineTesterMixin
if is_torch_available():
import torch
from transformers import (
GPT2TokenizerFast,
GPTBigCodeForCausalLM,
GPTBigCodeForSequenceClassification,
GPTBigCodeForTokenClassification,
GPTBigCodeModel,
)
from transformers.models.gpt_bigcode.modeling_gpt_bigcode import GPTBigCodeAttention
from transformers.pytorch_utils import is_torch_greater_or_equal_than_1_12
else:
is_torch_greater_or_equal_than_1_12 = False
class GPTBigCodeModelTester:
def __init__(
self,
parent,
batch_size=14,
seq_length=7,
is_training=True,
use_token_type_ids=True,
use_input_mask=True,
use_labels=True,
use_mc_token_ids=True,
vocab_size=99,
hidden_size=32,
num_hidden_layers=2,
num_attention_heads=4,
intermediate_size=37,
hidden_act="relu",
hidden_dropout_prob=0.1,
attention_probs_dropout_prob=0.1,
max_position_embeddings=512,
type_vocab_size=16,
type_sequence_label_size=2,
initializer_range=0.02,
num_labels=3,
num_choices=4,
multi_query=True,
scope=None,
):
self.parent = parent
self.batch_size = batch_size
self.seq_length = seq_length
self.is_training = is_training
self.use_token_type_ids = use_token_type_ids
self.use_input_mask = use_input_mask
self.use_labels = use_labels
self.use_mc_token_ids = use_mc_token_ids
self.vocab_size = vocab_size
self.hidden_size = hidden_size
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
self.intermediate_size = intermediate_size
self.hidden_act = hidden_act
self.hidden_dropout_prob = hidden_dropout_prob
self.attention_probs_dropout_prob = attention_probs_dropout_prob
self.max_position_embeddings = max_position_embeddings
self.type_vocab_size = type_vocab_size
self.type_sequence_label_size = type_sequence_label_size
self.initializer_range = initializer_range
self.num_labels = num_labels
self.num_choices = num_choices
self.scope = None
self.bos_token_id = vocab_size - 1
self.eos_token_id = vocab_size - 2
self.pad_token_id = vocab_size - 3
self.multi_query = multi_query
def get_large_model_config(self):
return GPTBigCodeConfig.from_pretrained("bigcode/gpt_bigcode-santacoder")
def prepare_config_and_inputs(
self, gradient_checkpointing=False, scale_attn_by_inverse_layer_idx=False, reorder_and_upcast_attn=False
):
input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
input_mask = None
if self.use_input_mask:
input_mask = random_attention_mask([self.batch_size, self.seq_length])
token_type_ids = None
if self.use_token_type_ids:
token_type_ids = ids_tensor([self.batch_size, self.seq_length], self.type_vocab_size)
mc_token_ids = None
if self.use_mc_token_ids:
mc_token_ids = ids_tensor([self.batch_size, self.num_choices], self.seq_length)
sequence_labels = None
token_labels = None
choice_labels = None
if self.use_labels:
sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels)
choice_labels = ids_tensor([self.batch_size], self.num_choices)
config = self.get_config(
gradient_checkpointing=gradient_checkpointing,
scale_attn_by_inverse_layer_idx=scale_attn_by_inverse_layer_idx,
reorder_and_upcast_attn=reorder_and_upcast_attn,
)
head_mask = ids_tensor([self.num_hidden_layers, self.num_attention_heads], 2)
return (
config,
input_ids,
input_mask,
head_mask,
token_type_ids,
mc_token_ids,
sequence_labels,
token_labels,
choice_labels,
)
def get_config(
self, gradient_checkpointing=False, scale_attn_by_inverse_layer_idx=False, reorder_and_upcast_attn=False
):
return GPTBigCodeConfig(
vocab_size=self.vocab_size,
n_embd=self.hidden_size,
n_layer=self.num_hidden_layers,
n_head=self.num_attention_heads,
n_inner=self.intermediate_size,
activation_function=self.hidden_act,
resid_pdrop=self.hidden_dropout_prob,
attn_pdrop=self.attention_probs_dropout_prob,
n_positions=self.max_position_embeddings,
type_vocab_size=self.type_vocab_size,
initializer_range=self.initializer_range,
use_cache=True,
bos_token_id=self.bos_token_id,
eos_token_id=self.eos_token_id,
pad_token_id=self.pad_token_id,
gradient_checkpointing=gradient_checkpointing,
scale_attn_by_inverse_layer_idx=scale_attn_by_inverse_layer_idx,
reorder_and_upcast_attn=reorder_and_upcast_attn,
attention_softmax_in_fp32=False,
scale_attention_softmax_in_fp32=False,
multi_query=self.multi_query,
)
def get_pipeline_config(self):
config = self.get_config()
config.vocab_size = 300
return config
def prepare_config_and_inputs_for_decoder(self):
(
config,
input_ids,
input_mask,
head_mask,
token_type_ids,
mc_token_ids,
sequence_labels,
token_labels,
choice_labels,
) = self.prepare_config_and_inputs()
encoder_hidden_states = floats_tensor([self.batch_size, self.seq_length, self.hidden_size])
encoder_attention_mask = ids_tensor([self.batch_size, self.seq_length], vocab_size=2)
return (
config,
input_ids,
input_mask,
head_mask,
token_type_ids,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
)
def create_and_check_gpt_bigcode_model(self, config, input_ids, input_mask, head_mask, token_type_ids, *args):
model = GPTBigCodeModel(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, token_type_ids=token_type_ids, head_mask=head_mask)
result = model(input_ids, token_type_ids=token_type_ids)
result = model(input_ids)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
self.parent.assertEqual(len(result.past_key_values), config.n_layer)
def create_and_check_gpt_bigcode_model_past(self, config, input_ids, input_mask, head_mask, token_type_ids, *args):
model = GPTBigCodeModel(config=config)
model.to(torch_device)
model.eval()
# first forward pass
outputs = model(input_ids, token_type_ids=token_type_ids, use_cache=True)
outputs_use_cache_conf = model(input_ids, token_type_ids=token_type_ids)
outputs_no_past = model(input_ids, token_type_ids=token_type_ids, use_cache=False)
self.parent.assertTrue(len(outputs) == len(outputs_use_cache_conf))
self.parent.assertTrue(len(outputs) == len(outputs_no_past) + 1)
output, past = outputs.to_tuple()
# create hypothetical next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size)
next_token_types = ids_tensor([self.batch_size, 1], self.type_vocab_size)
# append to next input_ids and token_type_ids
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
next_token_type_ids = torch.cat([token_type_ids, next_token_types], dim=-1)
output_from_no_past = model(next_input_ids, token_type_ids=next_token_type_ids)["last_hidden_state"]
output_from_past = model(next_tokens, token_type_ids=next_token_types, past_key_values=past)[
"last_hidden_state"
]
# select random slice
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
output_from_no_past_slice = output_from_no_past[:, -1, random_slice_idx].detach()
output_from_past_slice = output_from_past[:, 0, random_slice_idx].detach()
# test that outputs are equal for slice
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
def create_and_check_gpt_bigcode_model_attention_mask_past(
self, config, input_ids, input_mask, head_mask, token_type_ids, *args
):
model = GPTBigCodeModel(config=config)
model.to(torch_device)
model.eval()
# create attention mask
attn_mask = torch.ones(input_ids.shape, dtype=torch.long, device=torch_device)
half_seq_length = self.seq_length // 2
attn_mask[:, half_seq_length:] = 0
# first forward pass
output, past = model(input_ids, attention_mask=attn_mask).to_tuple()
# create hypothetical next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size)
# change a random masked slice from input_ids
random_seq_idx_to_change = ids_tensor((1,), half_seq_length).item() + 1
random_other_next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size).squeeze(-1)
input_ids[:, -random_seq_idx_to_change] = random_other_next_tokens
# append to next input_ids and attn_mask
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
attn_mask = torch.cat(
[attn_mask, torch.ones((attn_mask.shape[0], 1), dtype=torch.long, device=torch_device)],
dim=1,
)
# get two different outputs
output_from_no_past = model(next_input_ids, attention_mask=attn_mask)["last_hidden_state"]
output_from_past = model(next_tokens, past_key_values=past, attention_mask=attn_mask)["last_hidden_state"]
# select random slice
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
output_from_no_past_slice = output_from_no_past[:, -1, random_slice_idx].detach()
output_from_past_slice = output_from_past[:, 0, random_slice_idx].detach()
# test that outputs are equal for slice
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
def create_and_check_gpt_bigcode_model_past_large_inputs(
self, config, input_ids, input_mask, head_mask, token_type_ids, *args
):
model = GPTBigCodeModel(config=config)
model.to(torch_device)
model.eval()
# first forward pass
outputs = model(input_ids, token_type_ids=token_type_ids, attention_mask=input_mask, use_cache=True)
output, past = outputs.to_tuple()
# create hypothetical next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
next_token_types = ids_tensor([self.batch_size, 3], self.type_vocab_size)
next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
# append to next input_ids and token_type_ids
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
next_token_type_ids = torch.cat([token_type_ids, next_token_types], dim=-1)
next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)
output_from_no_past = model(
next_input_ids, token_type_ids=next_token_type_ids, attention_mask=next_attention_mask
)["last_hidden_state"]
output_from_past = model(
next_tokens, token_type_ids=next_token_types, attention_mask=next_attention_mask, past_key_values=past
)["last_hidden_state"]
self.parent.assertTrue(output_from_past.shape[1] == next_tokens.shape[1])
# select random slice
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
# test that outputs are equal for slice
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
def create_and_check_lm_head_model(self, config, input_ids, input_mask, head_mask, token_type_ids, *args):
model = GPTBigCodeForCausalLM(config)
model.to(torch_device)
model.eval()
result = model(input_ids, token_type_ids=token_type_ids, labels=input_ids)
self.parent.assertEqual(result.loss.shape, ())
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
def create_and_check_forward_and_backwards(
self, config, input_ids, input_mask, head_mask, token_type_ids, *args, gradient_checkpointing=False
):
model = GPTBigCodeForCausalLM(config)
model.to(torch_device)
if gradient_checkpointing:
model.gradient_checkpointing_enable()
result = model(input_ids, token_type_ids=token_type_ids, labels=input_ids)
self.parent.assertEqual(result.loss.shape, ())
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
result.loss.backward()
def create_and_check_gpt_bigcode_for_sequence_classification(
self, config, input_ids, input_mask, head_mask, token_type_ids, mc_token_ids, sequence_labels, *args
):
config.num_labels = self.num_labels
model = GPTBigCodeForSequenceClassification(config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=sequence_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_labels))
def create_and_check_gpt_bigcode_for_token_classification(
self, config, input_ids, input_mask, head_mask, token_type_ids, mc_token_ids, sequence_labels, *args
):
config.num_labels = self.num_labels
model = GPTBigCodeForTokenClassification(config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.num_labels))
def create_and_check_gpt_bigcode_weight_initialization(self, config, *args):
model = GPTBigCodeModel(config)
model_std = model.config.initializer_range / math.sqrt(2 * model.config.n_layer)
for key in model.state_dict().keys():
if "c_proj" in key and "weight" in key:
self.parent.assertLessEqual(abs(torch.std(model.state_dict()[key]) - model_std), 0.001)
self.parent.assertLessEqual(abs(torch.mean(model.state_dict()[key]) - 0.0), 0.01)
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
(
config,
input_ids,
input_mask,
head_mask,
token_type_ids,
mc_token_ids,
sequence_labels,
token_labels,
choice_labels,
) = config_and_inputs
inputs_dict = {
"input_ids": input_ids,
"token_type_ids": token_type_ids,
"head_mask": head_mask,
}
return config, inputs_dict
@require_torch
class GPTBigCodeModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase):
# TODO: Update the tests to use valid pretrained models.
all_model_classes = (
(
GPTBigCodeModel,
GPTBigCodeForCausalLM,
GPTBigCodeForSequenceClassification,
GPTBigCodeForTokenClassification,
)
if is_torch_available()
else ()
)
all_generative_model_classes = (GPTBigCodeForCausalLM,) if is_torch_available() else ()
pipeline_model_mapping = (
{
"feature-extraction": GPTBigCodeModel,
"text-classification": GPTBigCodeForSequenceClassification,
"text-generation": GPTBigCodeForCausalLM,
"token-classification": GPTBigCodeForTokenClassification,
"zero-shot": GPTBigCodeForSequenceClassification,
}
if is_torch_available()
else {}
)
fx_compatible = False
test_missing_keys = False
test_pruning = False
test_torchscript = False
multi_query = True
# special case for DoubleHeads model
def _prepare_for_class(self, inputs_dict, model_class, return_labels=False):
inputs_dict = super()._prepare_for_class(inputs_dict, model_class, return_labels=return_labels)
return inputs_dict
def setUp(self):
self.model_tester = GPTBigCodeModelTester(self, multi_query=self.multi_query)
self.config_tester = ConfigTester(self, config_class=GPTBigCodeConfig, n_embd=37)
def tearDown(self):
import gc
gc.collect()
def test_config(self):
self.config_tester.run_common_tests()
@unittest.skip(reason="MQA models does not support retain_grad")
def test_retain_grad_hidden_states_attentions(self):
pass
@unittest.skip(reason="Contrastive search not supported due to non-standard caching mechanism")
def test_contrastive_generate(self):
pass
@unittest.skip(reason="Contrastive search not supported due to non-standard caching mechanism")
def test_contrastive_generate_dict_outputs_use_cache(self):
pass
@unittest.skip(reason="CPU offload seems to be broken for some reason - tiny models keep hitting corner cases")
def test_cpu_offload(self):
pass
@unittest.skip(reason="Disk offload seems to be broken for some reason - tiny models keep hitting corner cases")
def test_disk_offload(self):
pass
@unittest.skip(reason="BigCodeGPT has a non-standard KV cache format.")
def test_past_key_values_format(self):
pass
def test_gpt_bigcode_model(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_gpt_bigcode_model(*config_and_inputs)
def test_gpt_bigcode_model_past(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_gpt_bigcode_model_past(*config_and_inputs)
def test_gpt_bigcode_model_att_mask_past(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_gpt_bigcode_model_attention_mask_past(*config_and_inputs)
def test_gpt_bigcode_model_past_large_inputs(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_gpt_bigcode_model_past_large_inputs(*config_and_inputs)
def test_gpt_bigcode_lm_head_model(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_lm_head_model(*config_and_inputs)
def test_gpt_bigcode_sequence_classification_model(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_gpt_bigcode_for_sequence_classification(*config_and_inputs)
def test_gpt_bigcode_token_classification_model(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_gpt_bigcode_for_token_classification(*config_and_inputs)
def test_gpt_bigcode_gradient_checkpointing(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_forward_and_backwards(*config_and_inputs, gradient_checkpointing=True)
def test_gpt_bigcode_scale_attn_by_inverse_layer_idx(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs(scale_attn_by_inverse_layer_idx=True)
self.model_tester.create_and_check_forward_and_backwards(*config_and_inputs)
def test_gpt_bigcode_reorder_and_upcast_attn(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs(reorder_and_upcast_attn=True)
self.model_tester.create_and_check_forward_and_backwards(*config_and_inputs)
def test_gpt_bigcode_weight_initialization(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_gpt_bigcode_weight_initialization(*config_and_inputs)
@require_torch
class GPTBigCodeMHAModelTest(GPTBigCodeModelTest):
# `parameterized_class` breaks with mixins, so we use inheritance instead
multi_query = False
@unittest.skipIf(
not is_torch_greater_or_equal_than_1_12,
reason="`GPTBigCode` checkpoints use `PytorchGELUTanh` which requires `torch>=1.12.0`.",
)
@slow
@require_torch
class GPTBigCodeModelLanguageGenerationTest(unittest.TestCase):
def test_generate_simple(self):
model = GPTBigCodeForCausalLM.from_pretrained("bigcode/gpt_bigcode-santacoder").to(torch_device)
tokenizer = GPT2TokenizerFast.from_pretrained("bigcode/gpt_bigcode-santacoder")
input_ids = tokenizer("def print_hello_world():", return_tensors="pt").input_ids.to(torch_device)
output_sequence = model.generate(input_ids)
output_sentence = tokenizer.decode(output_sequence[0], skip_special_tokens=True)
expected_output = """def print_hello_world():\n print("Hello World!")\n\n\ndef print_hello_"""
self.assertEqual(output_sentence, expected_output)
def test_generate_batched(self):
tokenizer = GPT2TokenizerFast.from_pretrained("bigcode/gpt_bigcode-santacoder")
tokenizer.pad_token = tokenizer.eos_token
tokenizer.padding_side = "left"
model = GPTBigCodeForCausalLM.from_pretrained("bigcode/gpt_bigcode-santacoder").to(torch_device)
inputs = tokenizer(["def print_hello_world():", "def say_hello():"], return_tensors="pt", padding=True).to(
torch_device
)
outputs = model.generate(**inputs)
outputs = tokenizer.batch_decode(outputs, skip_special_tokens=True)
expected_output = [
'def print_hello_world():\n print("Hello World!")\n\n\ndef print_hello_',
'def say_hello():\n print("Hello, World!")\n\n\nsay_hello()',
]
self.assertListEqual(outputs, expected_output)
@require_torch
class GPTBigCodeMQATest(unittest.TestCase):
def get_attention(self, multi_query):
config = GPTBigCodeConfig.from_pretrained(
"bigcode/gpt_bigcode-santacoder",
multi_query=multi_query,
attn_pdrop=0,
resid_pdrop=0,
)
return GPTBigCodeAttention(config)
@parameterized.expand([(seed, is_train_mode) for seed in range(5) for is_train_mode in [True, False]])
def test_mqa_reduces_to_mha(self, seed, is_train_mode=True):
torch.manual_seed(seed)
# CREATE MQA AND MHA ATTENTIONS
attention_mqa = self.get_attention(True)
attention_mha = self.get_attention(False)
# ENFORCE MATCHING WEIGHTS
num_heads = attention_mqa.num_heads
embed_dim = attention_mqa.embed_dim
head_dim = attention_mqa.head_dim
with torch.no_grad():
mqa_q_weight = attention_mqa.c_attn.weight[:embed_dim, :].view(num_heads, 1, head_dim, embed_dim)
mqa_kv_weight = attention_mqa.c_attn.weight[embed_dim:, :].view(1, 2, head_dim, embed_dim)
mha_c_weight = torch.cat(
[mqa_q_weight, mqa_kv_weight.expand(num_heads, 2, head_dim, embed_dim)], dim=1
).view(3 * num_heads * head_dim, embed_dim)
mqa_q_bias = attention_mqa.c_attn.bias[:embed_dim].view(num_heads, 1, head_dim)
mqa_kv_bias = attention_mqa.c_attn.bias[embed_dim:].view(1, 2, head_dim)
mha_c_bias = torch.cat([mqa_q_bias, mqa_kv_bias.expand(num_heads, 2, head_dim)], dim=1).view(
3 * num_heads * head_dim
)
attention_mha.c_attn.weight.copy_(mha_c_weight)
attention_mha.c_attn.bias.copy_(mha_c_bias)
attention_mha.c_proj.weight.copy_(attention_mqa.c_proj.weight)
attention_mha.c_proj.bias.copy_(attention_mqa.c_proj.bias)
# PUT THE MODEL INTO THE CORRECT MODE
attention_mha.train(is_train_mode)
attention_mqa.train(is_train_mode)
# RUN AN INPUT THROUGH THE MODELS
num_tokens = 5
hidden_states = torch.randn(1, num_tokens, embed_dim)
attention_mha_result = attention_mha(hidden_states)[0]
attention_mqa_result = attention_mqa(hidden_states)[0]
# CHECK THAT ALL OUTPUTS ARE THE SAME
self.assertTrue(torch.allclose(attention_mha_result, attention_mqa_result, atol=1e-5))
|
transformers/tests/models/gpt_bigcode/test_modeling_gpt_bigcode.py/0
|
{
"file_path": "transformers/tests/models/gpt_bigcode/test_modeling_gpt_bigcode.py",
"repo_id": "transformers",
"token_count": 11779
}
| 408
|
# coding=utf-8
# Copyright 2024 The HuggingFace Inc. team. 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.
"""Testing suite for the PyTorch Granite model."""
import tempfile
import unittest
import pytest
from parameterized import parameterized
from transformers import AutoTokenizer, GraniteConfig, is_torch_available, set_seed
from transformers.testing_utils import (
require_bitsandbytes,
require_flash_attn,
require_read_token,
require_torch,
require_torch_gpu,
require_torch_sdpa,
slow,
torch_device,
)
from ...generation.test_utils import GenerationTesterMixin
from ...test_configuration_common import ConfigTester
from ...test_modeling_common import ModelTesterMixin, ids_tensor
from ...test_pipeline_mixin import PipelineTesterMixin
if is_torch_available():
import torch
from transformers import (
GraniteForCausalLM,
GraniteModel,
)
from transformers.models.granite.modeling_granite import (
GraniteRotaryEmbedding,
)
class GraniteModelTester:
def __init__(
self,
parent,
batch_size=13,
seq_length=7,
is_training=True,
use_input_mask=True,
use_token_type_ids=False,
use_labels=True,
vocab_size=99,
hidden_size=32,
num_hidden_layers=2,
num_attention_heads=4,
intermediate_size=37,
hidden_act="gelu",
hidden_dropout_prob=0.1,
attention_probs_dropout_prob=0.1,
max_position_embeddings=512,
type_vocab_size=16,
type_sequence_label_size=2,
initializer_range=0.02,
num_labels=3,
num_choices=4,
pad_token_id=0,
scope=None,
):
self.parent = parent
self.batch_size = batch_size
self.seq_length = seq_length
self.is_training = is_training
self.use_input_mask = use_input_mask
self.use_token_type_ids = use_token_type_ids
self.use_labels = use_labels
self.vocab_size = vocab_size
self.hidden_size = hidden_size
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
self.intermediate_size = intermediate_size
self.hidden_act = hidden_act
self.hidden_dropout_prob = hidden_dropout_prob
self.attention_probs_dropout_prob = attention_probs_dropout_prob
self.max_position_embeddings = max_position_embeddings
self.type_vocab_size = type_vocab_size
self.type_sequence_label_size = type_sequence_label_size
self.initializer_range = initializer_range
self.num_labels = num_labels
self.num_choices = num_choices
self.pad_token_id = pad_token_id
self.scope = scope
def prepare_config_and_inputs(self):
input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
input_mask = None
if self.use_input_mask:
input_mask = torch.tril(torch.ones(self.batch_size, self.seq_length)).to(torch_device)
token_type_ids = None
if self.use_token_type_ids:
token_type_ids = ids_tensor([self.batch_size, self.seq_length], self.type_vocab_size)
sequence_labels = None
token_labels = None
choice_labels = None
if self.use_labels:
sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels)
choice_labels = ids_tensor([self.batch_size], self.num_choices)
config = self.get_config()
return config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
def get_config(self):
return GraniteConfig(
vocab_size=self.vocab_size,
hidden_size=self.hidden_size,
num_hidden_layers=self.num_hidden_layers,
num_attention_heads=self.num_attention_heads,
intermediate_size=self.intermediate_size,
hidden_act=self.hidden_act,
hidden_dropout_prob=self.hidden_dropout_prob,
attention_probs_dropout_prob=self.attention_probs_dropout_prob,
max_position_embeddings=self.max_position_embeddings,
type_vocab_size=self.type_vocab_size,
is_decoder=False,
initializer_range=self.initializer_range,
pad_token_id=self.pad_token_id,
)
def create_and_check_model(
self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
):
model = GraniteModel(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask)
result = model(input_ids)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_model_as_decoder(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.add_cross_attention = True
model = GraniteModel(config)
model.to(torch_device)
model.eval()
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
)
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
)
result = model(input_ids, attention_mask=input_mask)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_for_causal_lm(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
model = GraniteForCausalLM(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, labels=token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
def create_and_check_decoder_model_past_large_inputs(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.is_decoder = True
config.add_cross_attention = True
model = GraniteForCausalLM(config=config)
model.to(torch_device)
model.eval()
# first forward pass
outputs = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
use_cache=True,
)
past_key_values = outputs.past_key_values
# create hypothetical multiple next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
# append to next input_ids and
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)
output_from_no_past = model(
next_input_ids,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
output_hidden_states=True,
)["hidden_states"][0]
output_from_past = model(
next_tokens,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
past_key_values=past_key_values,
output_hidden_states=True,
)["hidden_states"][0]
# select random slice
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
# test that outputs are equal for slice
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
(
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
) = config_and_inputs
inputs_dict = {"input_ids": input_ids, "attention_mask": input_mask}
return config, inputs_dict
@require_torch
class GraniteModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase):
all_model_classes = (
(
GraniteModel,
GraniteForCausalLM,
)
if is_torch_available()
else ()
)
all_generative_model_classes = (GraniteForCausalLM,) if is_torch_available() else ()
pipeline_model_mapping = (
{
"feature-extraction": GraniteModel,
"text-generation": GraniteForCausalLM,
}
if is_torch_available()
else {}
)
test_headmasking = False
test_pruning = False
fx_compatible = False
# Need to use `0.8` instead of `0.9` for `test_cpu_offload`
# This is because we are hitting edge cases with the causal_mask buffer
model_split_percents = [0.5, 0.7, 0.8]
# used in `test_torch_compile`
_torch_compile_test_ckpt = "ibm/PowerLM-3b"
def setUp(self):
self.model_tester = GraniteModelTester(self)
self.config_tester = ConfigTester(self, config_class=GraniteConfig, hidden_size=37)
def test_config(self):
self.config_tester.run_common_tests()
def test_model(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_model(*config_and_inputs)
def test_model_various_embeddings(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
for type in ["absolute", "relative_key", "relative_key_query"]:
config_and_inputs[0].position_embedding_type = type
self.model_tester.create_and_check_model(*config_and_inputs)
# def test_granite_sequence_classification_model(self):
# config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
# config.num_labels = 3
# input_ids = input_dict["input_ids"]
# attention_mask = input_ids.ne(1).to(torch_device)
# sequence_labels = ids_tensor([self.model_tester.batch_size], self.model_tester.type_sequence_label_size)
# model = GraniteForSequenceClassification(config)
# model.to(torch_device)
# model.eval()
# result = model(input_ids, attention_mask=attention_mask, labels=sequence_labels)
# self.assertEqual(result.logits.shape, (self.model_tester.batch_size, self.model_tester.num_labels))
# def test_granite_sequence_classification_model_for_single_label(self):
# config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
# config.num_labels = 3
# config.problem_type = "single_label_classification"
# input_ids = input_dict["input_ids"]
# attention_mask = input_ids.ne(1).to(torch_device)
# sequence_labels = ids_tensor([self.model_tester.batch_size], self.model_tester.type_sequence_label_size)
# model = GraniteForSequenceClassification(config)
# model.to(torch_device)
# model.eval()
# result = model(input_ids, attention_mask=attention_mask, labels=sequence_labels)
# self.assertEqual(result.logits.shape, (self.model_tester.batch_size, self.model_tester.num_labels))
# def test_granite_sequence_classification_model_for_multi_label(self):
# config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
# config.num_labels = 3
# config.problem_type = "multi_label_classification"
# input_ids = input_dict["input_ids"]
# attention_mask = input_ids.ne(1).to(torch_device)
# sequence_labels = ids_tensor(
# [self.model_tester.batch_size, config.num_labels], self.model_tester.type_sequence_label_size
# ).to(torch.float)
# model = GraniteForSequenceClassification(config)
# model.to(torch_device)
# model.eval()
# result = model(input_ids, attention_mask=attention_mask, labels=sequence_labels)
# self.assertEqual(result.logits.shape, (self.model_tester.batch_size, self.model_tester.num_labels))
# def test_granite_token_classification_model(self):
# config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
# config.num_labels = 3
# input_ids = input_dict["input_ids"]
# attention_mask = input_ids.ne(1).to(torch_device)
# token_labels = ids_tensor([self.model_tester.batch_size, self.model_tester.seq_length], config.num_labels)
# model = GraniteForTokenClassification(config=config)
# model.to(torch_device)
# model.eval()
# result = model(input_ids, attention_mask=attention_mask, labels=token_labels)
# self.assertEqual(
# result.logits.shape,
# (self.model_tester.batch_size, self.model_tester.seq_length, self.model_tester.num_labels),
# )
@unittest.skip("Granite buffers include complex numbers, which breaks this test")
def test_save_load_fast_init_from_base(self):
pass
@parameterized.expand([("linear",), ("dynamic",)])
def test_model_rope_scaling_from_config(self, scaling_type):
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
short_input = ids_tensor([1, 10], config.vocab_size)
long_input = ids_tensor([1, int(config.max_position_embeddings * 1.5)], config.vocab_size)
set_seed(42) # Fixed seed at init time so the two models get the same random weights
original_model = GraniteModel(config)
original_model.to(torch_device)
original_model.eval()
original_short_output = original_model(short_input).last_hidden_state
original_long_output = original_model(long_input).last_hidden_state
set_seed(42) # Fixed seed at init time so the two models get the same random weights
config.rope_scaling = {"type": scaling_type, "factor": 10.0}
scaled_model = GraniteModel(config)
scaled_model.to(torch_device)
scaled_model.eval()
scaled_short_output = scaled_model(short_input).last_hidden_state
scaled_long_output = scaled_model(long_input).last_hidden_state
# Dynamic scaling does not change the RoPE embeddings until it receives an input longer than the original
# maximum sequence length, so the outputs for the short input should match.
if scaling_type == "dynamic":
self.assertTrue(torch.allclose(original_short_output, scaled_short_output, atol=1e-5))
else:
self.assertFalse(torch.allclose(original_short_output, scaled_short_output, atol=1e-5))
# The output should be different for long inputs
self.assertFalse(torch.allclose(original_long_output, scaled_long_output, atol=1e-5))
def test_model_rope_scaling(self):
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
scaling_factor = 10
short_input_length = 10
long_input_length = int(config.max_position_embeddings * 1.5)
# Inputs
x = torch.randn(1, dtype=torch.float32, device=torch_device) # used exlusively to get the dtype and the device
position_ids_short = torch.arange(short_input_length, dtype=torch.long, device=torch_device)
position_ids_short = position_ids_short.unsqueeze(0)
position_ids_long = torch.arange(long_input_length, dtype=torch.long, device=torch_device)
position_ids_long = position_ids_long.unsqueeze(0)
# Sanity check original RoPE
original_rope = GraniteRotaryEmbedding(config=config).to(torch_device)
original_cos_short, original_sin_short = original_rope(x, position_ids_short)
original_cos_long, original_sin_long = original_rope(x, position_ids_long)
torch.testing.assert_close(original_cos_short, original_cos_long[:, :short_input_length, :])
torch.testing.assert_close(original_sin_short, original_sin_long[:, :short_input_length, :])
# Sanity check linear RoPE scaling
# New position "x" should match original position with index "x/scaling_factor"
config.rope_scaling = {"type": "linear", "factor": scaling_factor}
linear_scaling_rope = GraniteRotaryEmbedding(config=config).to(torch_device)
linear_cos_short, linear_sin_short = linear_scaling_rope(x, position_ids_short)
linear_cos_long, linear_sin_long = linear_scaling_rope(x, position_ids_long)
torch.testing.assert_close(linear_cos_short, linear_cos_long[:, :short_input_length, :])
torch.testing.assert_close(linear_sin_short, linear_sin_long[:, :short_input_length, :])
for new_position in range(0, long_input_length, scaling_factor):
original_position = int(new_position // scaling_factor)
torch.testing.assert_close(linear_cos_long[:, new_position, :], original_cos_long[:, original_position, :])
torch.testing.assert_close(linear_sin_long[:, new_position, :], original_sin_long[:, original_position, :])
# Sanity check Dynamic NTK RoPE scaling
# Scaling should only be observed after a long input is fed. We can observe that the frequencies increase
# with scaling_factor (or that `inv_freq` decreases)
config.rope_scaling = {"type": "dynamic", "factor": scaling_factor}
ntk_scaling_rope = GraniteRotaryEmbedding(config=config).to(torch_device)
ntk_cos_short, ntk_sin_short = ntk_scaling_rope(x, position_ids_short)
ntk_cos_long, ntk_sin_long = ntk_scaling_rope(x, position_ids_long)
torch.testing.assert_close(ntk_cos_short, original_cos_short)
torch.testing.assert_close(ntk_sin_short, original_sin_short)
with self.assertRaises(AssertionError):
torch.testing.assert_close(ntk_cos_long, original_cos_long)
with self.assertRaises(AssertionError):
torch.testing.assert_close(ntk_sin_long, original_sin_long)
self.assertTrue((ntk_scaling_rope.inv_freq <= original_rope.inv_freq).all())
# Sanity check Yarn RoPE scaling
# Scaling should be over the entire input
config.rope_scaling = {"type": "yarn", "factor": scaling_factor}
yarn_scaling_rope = GraniteRotaryEmbedding(config=config).to(torch_device)
yarn_cos_short, yarn_sin_short = yarn_scaling_rope(x, position_ids_short)
yarn_cos_long, yarn_sin_long = yarn_scaling_rope(x, position_ids_long)
torch.testing.assert_close(yarn_cos_short, yarn_cos_long[:, :short_input_length, :])
torch.testing.assert_close(yarn_sin_short, yarn_sin_long[:, :short_input_length, :])
with self.assertRaises(AssertionError):
torch.testing.assert_close(yarn_cos_short, original_cos_short)
with self.assertRaises(AssertionError):
torch.testing.assert_close(yarn_sin_short, original_sin_short)
with self.assertRaises(AssertionError):
torch.testing.assert_close(yarn_cos_long, original_cos_long)
with self.assertRaises(AssertionError):
torch.testing.assert_close(yarn_sin_long, original_sin_long)
@require_flash_attn
@require_torch_gpu
@require_bitsandbytes
@pytest.mark.flash_attn_test
@require_read_token
@slow
def test_flash_attn_2_generate_padding_right(self):
"""
Overwritting the common test as the test is flaky on tiny models
"""
model = GraniteForCausalLM.from_pretrained(
"ibm/PowerLM-3b",
load_in_4bit=True,
device_map={"": 0},
)
tokenizer = AutoTokenizer.from_pretrained("ibm/PowerLM-3b")
texts = ["hi", "Hello this is a very long sentence"]
tokenizer.padding_side = "right"
tokenizer.pad_token = tokenizer.eos_token
inputs = tokenizer(texts, return_tensors="pt", padding=True).to(0)
output_native = model.generate(**inputs, max_new_tokens=20, do_sample=False)
output_native = tokenizer.batch_decode(output_native)
model = GraniteForCausalLM.from_pretrained(
"ibm/PowerLM-3b",
load_in_4bit=True,
device_map={"": 0},
attn_implementation="flash_attention_2",
)
output_fa_2 = model.generate(**inputs, max_new_tokens=20, do_sample=False)
output_fa_2 = tokenizer.batch_decode(output_fa_2)
self.assertListEqual(output_native, output_fa_2)
@require_flash_attn
@require_torch_gpu
@slow
def test_use_flash_attention_2_true(self):
"""
NOTE: this is the only test testing that the legacy `use_flash_attention=2` argument still works as intended.
"""
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
with tempfile.TemporaryDirectory() as tmp_dir:
model = model_class(config)
model.save_pretrained(tmp_dir)
new_model = GraniteForCausalLM.from_pretrained(
tmp_dir, use_flash_attention_2=True, torch_dtype=torch.float16
).to("cuda")
self.assertTrue(new_model.config._attn_implementation == "flash_attention_2")
has_flash = False
for name, submodule in new_model.named_modules():
if "FlashAttention" in submodule.__class__.__name__:
has_flash = True
break
if not has_flash:
raise ValueError("The flash model should have flash attention layers")
@parameterized.expand([("float16",), ("bfloat16",), ("float32",)])
@require_torch_sdpa
@slow
def test_eager_matches_sdpa_inference(self, torch_dtype: str):
"""
skipping the test since mup is very flaky and gets consistently different outputs
"""
self.skipTest("skipping the test since mup is very flaky and gets consistently different outputs")
@require_torch_gpu
class GraniteIntegrationTest(unittest.TestCase):
# This variable is used to determine which CUDA device are we using for our runners (A10 or T4)
# Depending on the hardware we get different logits / generations
cuda_compute_capability_major_version = None
@classmethod
def setUpClass(cls):
if is_torch_available() and torch.cuda.is_available():
# 8 is for A100 / A10 and 7 for T4
cls.cuda_compute_capability_major_version = torch.cuda.get_device_capability()[0]
@slow
@require_read_token
def test_model_3b_logits_bf16(self):
input_ids = [1, 306, 4658, 278, 6593, 310, 2834, 338]
model = GraniteForCausalLM.from_pretrained(
"ibm/PowerLM-3b", device_map="auto", torch_dtype=torch.bfloat16, attn_implementation="eager"
)
with torch.no_grad():
out = model(torch.tensor([input_ids]).to(torch_device))
# Expected mean on dim = -1
# fmt: off
EXPECTED_MEAN = torch.tensor([[-1.8799, -3.1269, -2.8297, -2.3755, -2.7364, -2.2389, -2.5914, -2.4154]])
self.assertTrue(torch.allclose(EXPECTED_MEAN.to(torch_device), out.logits.mean(-1), atol=1e-2, rtol=1e-2))
# slicing logits[0, 0, 0:15]
EXPECTED_SLICE = torch.tensor([[4.8125, -2.0156, -2.0156, -2.0000, -2.0000, -2.8438, -2.0156, -2.0000, -2.0000, -2.0000, -2.0000, -2.0000, -2.0000, -2.0000, -2.0000]])
# fmt: on
self.assertTrue(
torch.allclose(
EXPECTED_SLICE.to(torch_device),
out.logits[0, 0, :15],
atol=1e-3,
rtol=1e-3,
)
)
@slow
@require_read_token
def test_model_3b_logits(self):
input_ids = [1, 306, 4658, 278, 6593, 310, 2834, 338]
model = GraniteForCausalLM.from_pretrained("ibm/PowerLM-3b", device_map="auto", torch_dtype=torch.float16)
with torch.no_grad():
out = model(torch.tensor([input_ids]).to(torch_device))
# fmt: off
# Expected mean on dim = -1
EXPECTED_MEAN = torch.tensor([[0.0000, 0.0000, -3.4374, -2.1636, -2.6245, -3.0029, -3.8229, -3.1158]])
self.assertTrue(torch.allclose(EXPECTED_MEAN.to(torch_device), out.logits.mean(-1), atol=1e-2, rtol=1e-2))
|
transformers/tests/models/granite/test_modeling_granite.py/0
|
{
"file_path": "transformers/tests/models/granite/test_modeling_granite.py",
"repo_id": "transformers",
"token_count": 11514
}
| 409
|
# coding=utf-8
# Copyright 2020 The HuggingFace Team. 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.
import copy
import unittest
from transformers import IBertConfig, is_torch_available
from transformers.testing_utils import require_torch, slow, torch_device
from ...test_configuration_common import ConfigTester
from ...test_modeling_common import ModelTesterMixin, ids_tensor, random_attention_mask
from ...test_pipeline_mixin import PipelineTesterMixin
if is_torch_available():
import torch
from torch import nn
from transformers import (
IBertForMaskedLM,
IBertForMultipleChoice,
IBertForQuestionAnswering,
IBertForSequenceClassification,
IBertForTokenClassification,
IBertModel,
)
from transformers.models.ibert.modeling_ibert import (
IBertEmbeddings,
IntGELU,
IntLayerNorm,
IntSoftmax,
QuantAct,
QuantEmbedding,
QuantLinear,
create_position_ids_from_input_ids,
)
class IBertModelTester:
def __init__(
self,
parent,
batch_size=13,
seq_length=7,
is_training=True,
use_input_mask=True,
use_token_type_ids=True,
use_labels=True,
vocab_size=99,
hidden_size=32,
num_hidden_layers=2,
num_attention_heads=4,
intermediate_size=37,
hidden_act="gelu",
hidden_dropout_prob=0.1,
attention_probs_dropout_prob=0.1,
max_position_embeddings=512,
type_vocab_size=16,
type_sequence_label_size=2,
initializer_range=0.02,
num_labels=3,
num_choices=4,
scope=None,
):
self.parent = parent
self.batch_size = batch_size
self.seq_length = seq_length
self.is_training = is_training
self.use_input_mask = use_input_mask
self.use_token_type_ids = use_token_type_ids
self.use_labels = use_labels
self.vocab_size = vocab_size
self.hidden_size = hidden_size
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
self.intermediate_size = intermediate_size
self.hidden_act = hidden_act
self.hidden_dropout_prob = hidden_dropout_prob
self.attention_probs_dropout_prob = attention_probs_dropout_prob
self.max_position_embeddings = max_position_embeddings
self.type_vocab_size = type_vocab_size
self.type_sequence_label_size = type_sequence_label_size
self.initializer_range = initializer_range
self.num_labels = num_labels
self.num_choices = num_choices
self.scope = scope
def prepare_config_and_inputs(self):
input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
input_mask = None
if self.use_input_mask:
input_mask = random_attention_mask([self.batch_size, self.seq_length])
token_type_ids = None
if self.use_token_type_ids:
token_type_ids = ids_tensor([self.batch_size, self.seq_length], self.type_vocab_size)
sequence_labels = None
token_labels = None
choice_labels = None
if self.use_labels:
sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels)
choice_labels = ids_tensor([self.batch_size], self.num_choices)
config = self.get_config()
return config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
def get_config(self):
return IBertConfig(
vocab_size=self.vocab_size,
hidden_size=self.hidden_size,
num_hidden_layers=self.num_hidden_layers,
num_attention_heads=self.num_attention_heads,
intermediate_size=self.intermediate_size,
hidden_act=self.hidden_act,
hidden_dropout_prob=self.hidden_dropout_prob,
attention_probs_dropout_prob=self.attention_probs_dropout_prob,
max_position_embeddings=self.max_position_embeddings,
type_vocab_size=self.type_vocab_size,
initializer_range=self.initializer_range,
quant_mode=True,
)
def get_pipeline_config(self):
config = self.get_config()
config.vocab_size = 300
return config
def create_and_check_model(
self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
):
model = IBertModel(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids)
result = model(input_ids, token_type_ids=token_type_ids)
result = model(input_ids)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
self.parent.assertEqual(result.pooler_output.shape, (self.batch_size, self.hidden_size))
def create_and_check_for_masked_lm(
self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
):
model = IBertForMaskedLM(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
def create_and_check_for_token_classification(
self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
):
config.num_labels = self.num_labels
model = IBertForTokenClassification(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.num_labels))
def create_and_check_for_multiple_choice(
self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
):
config.num_choices = self.num_choices
model = IBertForMultipleChoice(config=config)
model.to(torch_device)
model.eval()
multiple_choice_inputs_ids = input_ids.unsqueeze(1).expand(-1, self.num_choices, -1).contiguous()
multiple_choice_token_type_ids = token_type_ids.unsqueeze(1).expand(-1, self.num_choices, -1).contiguous()
multiple_choice_input_mask = input_mask.unsqueeze(1).expand(-1, self.num_choices, -1).contiguous()
result = model(
multiple_choice_inputs_ids,
attention_mask=multiple_choice_input_mask,
token_type_ids=multiple_choice_token_type_ids,
labels=choice_labels,
)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_choices))
def create_and_check_for_question_answering(
self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
):
model = IBertForQuestionAnswering(config=config)
model.to(torch_device)
model.eval()
result = model(
input_ids,
attention_mask=input_mask,
token_type_ids=token_type_ids,
start_positions=sequence_labels,
end_positions=sequence_labels,
)
self.parent.assertEqual(result.start_logits.shape, (self.batch_size, self.seq_length))
self.parent.assertEqual(result.end_logits.shape, (self.batch_size, self.seq_length))
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
(
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
) = config_and_inputs
inputs_dict = {"input_ids": input_ids, "token_type_ids": token_type_ids, "attention_mask": input_mask}
return config, inputs_dict
@require_torch
class IBertModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
test_pruning = False
test_torchscript = False
test_head_masking = False
test_resize_embeddings = False
all_model_classes = (
(
IBertForMaskedLM,
IBertModel,
IBertForSequenceClassification,
IBertForTokenClassification,
IBertForMultipleChoice,
IBertForQuestionAnswering,
)
if is_torch_available()
else ()
)
pipeline_model_mapping = (
{
"feature-extraction": IBertModel,
"fill-mask": IBertForMaskedLM,
"question-answering": IBertForQuestionAnswering,
"text-classification": IBertForSequenceClassification,
"token-classification": IBertForTokenClassification,
"zero-shot": IBertForSequenceClassification,
}
if is_torch_available()
else {}
)
def setUp(self):
self.model_tester = IBertModelTester(self)
self.config_tester = ConfigTester(self, config_class=IBertConfig, hidden_size=37)
def test_config(self):
self.config_tester.run_common_tests()
def test_model(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_model(*config_and_inputs)
def test_model_various_embeddings(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
# I-BERT only supports absolute embedding
for type in ["absolute"]:
config_and_inputs[0].position_embedding_type = type
self.model_tester.create_and_check_model(*config_and_inputs)
def test_for_masked_lm(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_for_masked_lm(*config_and_inputs)
def test_for_token_classification(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_for_token_classification(*config_and_inputs)
def test_for_multiple_choice(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_for_multiple_choice(*config_and_inputs)
def test_for_question_answering(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_for_question_answering(*config_and_inputs)
@slow
def test_model_from_pretrained(self):
model_name = "kssteven/ibert-roberta-base"
model = IBertModel.from_pretrained(model_name)
self.assertIsNotNone(model)
def test_create_position_ids_respects_padding_index(self):
"""This is a regression test for https://github.com/huggingface/transformers/issues/1761
The position ids should be masked with the embedding object's padding index. Therefore, the
first available non-padding position index is IBertEmbeddings.padding_idx + 1
"""
config = self.model_tester.prepare_config_and_inputs()[0]
model = IBertEmbeddings(config=config)
input_ids = torch.as_tensor([[12, 31, 13, model.padding_idx]])
expected_positions = torch.as_tensor(
[[0 + model.padding_idx + 1, 1 + model.padding_idx + 1, 2 + model.padding_idx + 1, model.padding_idx]]
)
position_ids = create_position_ids_from_input_ids(input_ids, model.padding_idx)
self.assertEqual(position_ids.shape, expected_positions.shape)
self.assertTrue(torch.all(torch.eq(position_ids, expected_positions)))
def test_create_position_ids_from_inputs_embeds(self):
"""This is a regression test for https://github.com/huggingface/transformers/issues/1761
The position ids should be masked with the embedding object's padding index. Therefore, the
first available non-padding position index is IBertEmbeddings.padding_idx + 1
"""
config = self.model_tester.prepare_config_and_inputs()[0]
embeddings = IBertEmbeddings(config=config)
inputs_embeds = torch.empty(2, 4, 30)
expected_single_positions = [
0 + embeddings.padding_idx + 1,
1 + embeddings.padding_idx + 1,
2 + embeddings.padding_idx + 1,
3 + embeddings.padding_idx + 1,
]
expected_positions = torch.as_tensor([expected_single_positions, expected_single_positions])
position_ids = embeddings.create_position_ids_from_inputs_embeds(inputs_embeds)
self.assertEqual(position_ids.shape, expected_positions.shape)
self.assertTrue(torch.all(torch.eq(position_ids, expected_positions)))
# Override
def test_model_get_set_embeddings(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
model = model_class(config)
self.assertIsInstance(model.get_input_embeddings(), QuantEmbedding)
model.set_input_embeddings(nn.Embedding(10, 10))
x = model.get_output_embeddings()
self.assertTrue(x is None or isinstance(x, nn.Linear))
# Override
def test_feed_forward_chunking(self):
pass # I-BERT does not support chunking
# Override
def test_inputs_embeds(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
model = model_class(config)
model.to(torch_device)
model.eval()
inputs = copy.deepcopy(self._prepare_for_class(inputs_dict, model_class))
if not self.is_encoder_decoder:
input_ids = inputs["input_ids"]
del inputs["input_ids"]
else:
encoder_input_ids = inputs["input_ids"]
decoder_input_ids = inputs.get("decoder_input_ids", encoder_input_ids)
del inputs["input_ids"]
inputs.pop("decoder_input_ids", None)
wte = model.get_input_embeddings()
if not self.is_encoder_decoder:
embed, embed_scaling_factor = wte(input_ids)
inputs["inputs_embeds"] = embed
else:
inputs["inputs_embeds"] = wte(encoder_input_ids)
inputs["decoder_inputs_embeds"] = wte(decoder_input_ids)
with torch.no_grad():
model(**inputs)[0]
@unittest.skip(reason="ibert overrides scaling to None if inputs_embeds")
def test_inputs_embeds_matches_input_ids(self):
pass
@require_torch
class IBertModelIntegrationTest(unittest.TestCase):
def test_quant_embedding(self):
weight_bit = 8
embedding = QuantEmbedding(2, 4, quant_mode=True, weight_bit=weight_bit)
embedding_weight = torch.tensor([[-1.0, -2.0, -3.0, -4.0], [5.0, 6.0, 7.0, 8.0]])
embedding.weight = nn.Parameter(embedding_weight)
expected_scaling_factor = embedding_weight.abs().max() / (2 ** (weight_bit - 1) - 1)
x, x_scaling_factor = embedding(torch.tensor(0))
y, y_scaling_factor = embedding(torch.tensor(1))
# scaling factor should follow the symmetric quantization rule
self.assertTrue(torch.allclose(x_scaling_factor, expected_scaling_factor, atol=1e-4))
self.assertTrue(torch.allclose(x_scaling_factor, expected_scaling_factor, atol=1e-4))
self.assertTrue(torch.allclose(y_scaling_factor, expected_scaling_factor, atol=1e-4))
# quantization error should not exceed the scaling factor
self.assertTrue(torch.allclose(x, embedding_weight[0], atol=expected_scaling_factor))
self.assertTrue(torch.allclose(y, embedding_weight[1], atol=expected_scaling_factor))
def test_quant_act(self):
def _test_range():
act = QuantAct(activation_bit, act_range_momentum, quant_mode=True)
# First pass
x = torch.tensor([[-1.0, -2.0, -3.0, -4.0], [5.0, 6.0, 7.0, 8.0]])
x_scaling_factor = torch.tensor(1.0)
y, y_scaling_factor = act(x, x_scaling_factor)
y_int = y / y_scaling_factor
# After the first pass, x_min and x_max should be initialized with x.min() and x.max()
expected_x_min, expected_x_max = x.min(), x.max()
self.assertTrue(torch.allclose(act.x_min, expected_x_min, atol=1e-4))
self.assertTrue(torch.allclose(act.x_max, expected_x_max, atol=1e-4))
# scaling factor should follow the symmetric quantization rule
expected_range = torch.max(expected_x_min.abs(), expected_x_max.abs())
expected_scaling_factor = expected_range / (2 ** (activation_bit - 1) - 1)
self.assertTrue(torch.allclose(y_scaling_factor, expected_scaling_factor, atol=1e-4))
# quantization error should not exceed the scaling factor
self.assertTrue(torch.allclose(x, y, atol=expected_scaling_factor))
# output should be integer
self.assertTrue(torch.allclose(y_int, y_int.round(), atol=1e-4))
# Second Pass
x = torch.tensor([[-1.0, -2.0, -3.0, -4.0], [5.0, 6.0, 7.0, 8.0]]) * 2
x_scaling_factor = torch.tensor(1.0)
y, y_scaling_factor = act(x, x_scaling_factor)
y_int = y / y_scaling_factor
# From the second pass, x_min and x_max should be updated with moving average
expected_x_min = expected_x_min * act_range_momentum + x.min() * (1 - act_range_momentum)
expected_x_max = expected_x_max * act_range_momentum + x.max() * (1 - act_range_momentum)
self.assertTrue(torch.allclose(act.x_min, expected_x_min, atol=1e-4))
self.assertTrue(torch.allclose(act.x_max, expected_x_max, atol=1e-4))
# scaling factor should follow the symmetric quantization rule
expected_range = torch.max(expected_x_min.abs(), expected_x_max.abs())
expected_scaling_factor = expected_range / (2 ** (activation_bit - 1) - 1)
self.assertTrue(torch.allclose(y_scaling_factor, expected_scaling_factor, atol=1e-4))
# quantization error should not exceed the scaling factor
x = x.clamp(min=-expected_range, max=expected_range)
self.assertTrue(torch.allclose(x, y, atol=expected_scaling_factor))
# output should be integer
self.assertTrue(torch.allclose(y_int, y_int.round(), atol=1e-4))
# Third pass, with eval()
act.eval()
x = torch.tensor([[-1.0, -2.0, -3.0, -4.0], [5.0, 6.0, 7.0, 8.0]]) * 3
# In eval mode, min/max and scaling factor must be fixed
self.assertTrue(torch.allclose(act.x_min, expected_x_min, atol=1e-4))
self.assertTrue(torch.allclose(act.x_max, expected_x_max, atol=1e-4))
self.assertTrue(torch.allclose(y_scaling_factor, expected_scaling_factor, atol=1e-4))
def _test_identity():
# test if identity and identity_scaling_factor are given
# should add the input values
act = QuantAct(activation_bit, act_range_momentum, quant_mode=True)
x = torch.tensor([[-1.0, -2.0, -3.0, -4.0], [5.0, 6.0, 7.0, 8.0]])
y = torch.tensor([[6.0, -7.0, 1.0, -2.0], [3.0, -4.0, -8.0, 5.0]])
x_scaling_factor = torch.tensor(1.0)
y_scaling_factor = torch.tensor(0.5)
z, z_scaling_factor = act(x, x_scaling_factor, y, y_scaling_factor)
z_int = z / z_scaling_factor
self.assertTrue(torch.allclose(x + y, z, atol=0.1))
self.assertTrue(torch.allclose(z_int, z_int.round(), atol=1e-4))
activation_bit = 8
act_range_momentum = 0.95
_test_range()
_test_identity()
def test_quant_linear(self):
def _test(per_channel):
linear_q = QuantLinear(2, 4, quant_mode=True, per_channel=per_channel, weight_bit=weight_bit)
linear_dq = QuantLinear(2, 4, quant_mode=False, per_channel=per_channel, weight_bit=weight_bit)
linear_weight = torch.tensor([[-1.0, 2.0, 3.0, -4.0], [5.0, -6.0, -7.0, 8.0]]).T
linear_q.weight = nn.Parameter(linear_weight)
linear_dq.weight = nn.Parameter(linear_weight)
q, q_scaling_factor = linear_q(x, x_scaling_factor)
q_int = q / q_scaling_factor
dq, dq_scaling_factor = linear_dq(x, x_scaling_factor)
if per_channel:
q_max = linear_weight.abs().max(dim=1).values
else:
q_max = linear_weight.abs().max()
expected_scaling_factor = q_max / (2 ** (weight_bit - 1) - 1)
# scaling factor should follow the symmetric quantization rule
self.assertTrue(torch.allclose(linear_q.fc_scaling_factor, expected_scaling_factor, atol=1e-4))
# output of the normal linear layer and the quantized linear layer should be similar
self.assertTrue(torch.allclose(q, dq, atol=0.5))
# output of the quantized linear layer should be integer
self.assertTrue(torch.allclose(q_int, q_int.round(), atol=1e-4))
weight_bit = 8
x = torch.tensor([[2.0, -5.0], [-3.0, 4.0]])
x_scaling_factor = torch.tensor([1.0])
_test(True)
_test(False)
def test_int_gelu(self):
gelu_q = IntGELU(quant_mode=True)
gelu_dq = nn.GELU()
x_int = torch.arange(-10000, 10001, 1)
x_scaling_factor = torch.tensor(0.001)
x = x_int * x_scaling_factor
q, q_scaling_factor = gelu_q(x, x_scaling_factor)
q_int = q / q_scaling_factor
dq = gelu_dq(x)
# output of the normal GELU and the quantized GELU should be similar
self.assertTrue(torch.allclose(q, dq, atol=0.5))
# output of the quantized GELU layer should be integer
self.assertTrue(torch.allclose(q_int, q_int.round(), atol=1e-4))
def test_force_dequant_gelu(self):
x_int = torch.arange(-10000, 10001, 1)
x_scaling_factor = torch.tensor(0.001)
x = x_int * x_scaling_factor
gelu_dq = IntGELU(quant_mode=False)
gelu_fdqs_dict = {
True: [
IntGELU(quant_mode=True, force_dequant="nonlinear"),
IntGELU(quant_mode=True, force_dequant="gelu"),
],
False: [
IntGELU(quant_mode=True, force_dequant="none"),
IntGELU(quant_mode=True, force_dequant="softmax"),
IntGELU(quant_mode=True, force_dequant="layernorm"),
],
}
dq, dq_scaling_factor = gelu_dq(x, x_scaling_factor)
for label, gelu_fdqs in gelu_fdqs_dict.items():
for gelu_fdq in gelu_fdqs:
q, q_scaling_factor = gelu_fdq(x, x_scaling_factor)
if label:
self.assertTrue(torch.allclose(q, dq, atol=1e-4))
else:
self.assertFalse(torch.allclose(q, dq, atol=1e-4))
def test_int_softmax(self):
output_bit = 8
softmax_q = IntSoftmax(output_bit, quant_mode=True)
softmax_dq = nn.Softmax()
def _test(array):
x_int = torch.tensor(array)
x_scaling_factor = torch.tensor(0.1)
x = x_int * x_scaling_factor
q, q_scaling_factor = softmax_q(x, x_scaling_factor)
q_int = q / q_scaling_factor
dq = softmax_dq(x)
# output of the normal Softmax and the quantized Softmax should be similar
self.assertTrue(torch.allclose(q, dq, atol=0.5))
# output of the quantized GELU layer should be integer
self.assertTrue(torch.allclose(q_int, q_int.round(), atol=1e-4))
# Output of the quantize Softmax should not exceed the output_bit
self.assertTrue(q.abs().max() < 2**output_bit)
array = [[i + j for j in range(10)] for i in range(-10, 10)]
_test(array)
array = [[i + j for j in range(50)] for i in range(-10, 10)]
_test(array)
array = [[i + 100 * j for j in range(2)] for i in range(-10, 10)]
_test(array)
def test_force_dequant_softmax(self):
output_bit = 8
array = [[i + j for j in range(10)] for i in range(-10, 10)]
x_int = torch.tensor(array)
x_scaling_factor = torch.tensor(0.1)
x = x_int * x_scaling_factor
softmax_dq = IntSoftmax(output_bit, quant_mode=False)
softmax_fdqs_dict = {
True: [
IntSoftmax(output_bit, quant_mode=True, force_dequant="nonlinear"),
IntSoftmax(output_bit, quant_mode=True, force_dequant="softmax"),
],
False: [
IntSoftmax(output_bit, quant_mode=True, force_dequant="none"),
IntSoftmax(output_bit, quant_mode=True, force_dequant="gelu"),
IntSoftmax(output_bit, quant_mode=True, force_dequant="layernorm"),
],
}
dq, dq_scaling_factor = softmax_dq(x, x_scaling_factor)
for label, softmax_fdqs in softmax_fdqs_dict.items():
for softmax_fdq in softmax_fdqs:
q, q_scaling_factor = softmax_fdq(x, x_scaling_factor)
if label:
self.assertTrue(torch.allclose(q, dq, atol=1e-4))
else:
self.assertFalse(torch.allclose(q, dq, atol=1e-4))
def test_int_layernorm(self):
output_bit = 8
# some random matrix
array = [[[i * j * j + j for j in range(5, 15)]] for i in range(-10, 10)]
x_int = torch.tensor(array)
x_scaling_factor = torch.tensor(0.1)
x = x_int * x_scaling_factor
ln_q = IntLayerNorm(x.shape[1:], 1e-5, quant_mode=True, output_bit=output_bit)
ln_dq = nn.LayerNorm(x.shape[1:], 1e-5)
ln_q.weight = nn.Parameter(torch.ones(x.shape[1:]))
ln_q.bias = nn.Parameter(torch.ones(x.shape[1:]))
ln_dq.weight = nn.Parameter(torch.ones(x.shape[1:]))
ln_dq.bias = nn.Parameter(torch.ones(x.shape[1:]))
q, q_scaling_factor = ln_q(x, x_scaling_factor)
q_int = q / q_scaling_factor
dq = ln_dq(x)
# output of the normal LN and the quantized LN should be similar
self.assertTrue(torch.allclose(q, dq, atol=0.5))
# output of the quantized GELU layer should be integer
self.assertTrue(torch.allclose(q_int, q_int.round(), atol=1e-4))
def test_force_dequant_layernorm(self):
output_bit = 8
array = [[[i * j * j + j for j in range(5, 15)]] for i in range(-10, 10)]
x_int = torch.tensor(array)
x_scaling_factor = torch.tensor(0.1)
x = x_int * x_scaling_factor
ln_dq = IntLayerNorm(x.shape[1:], 1e-5, quant_mode=False, output_bit=output_bit)
ln_fdqs_dict = {
True: [
IntLayerNorm(x.shape[1:], 1e-5, quant_mode=True, output_bit=output_bit, force_dequant="nonlinear"),
IntLayerNorm(x.shape[1:], 1e-5, quant_mode=True, output_bit=output_bit, force_dequant="layernorm"),
],
False: [
IntLayerNorm(x.shape[1:], 1e-5, quant_mode=True, output_bit=output_bit, force_dequant="none"),
IntLayerNorm(x.shape[1:], 1e-5, quant_mode=True, output_bit=output_bit, force_dequant="gelu"),
IntLayerNorm(x.shape[1:], 1e-5, quant_mode=True, output_bit=output_bit, force_dequant="softmax"),
],
}
ln_dq.weight = nn.Parameter(torch.ones(x.shape[1:]))
ln_dq.bias = nn.Parameter(torch.ones(x.shape[1:]))
dq, dq_scaling_factor = ln_dq(x, x_scaling_factor)
for label, ln_fdqs in ln_fdqs_dict.items():
for ln_fdq in ln_fdqs:
ln_fdq.weight = nn.Parameter(torch.ones(x.shape[1:]))
ln_fdq.bias = nn.Parameter(torch.ones(x.shape[1:]))
q, q_scaling_factor = ln_fdq(x, x_scaling_factor)
if label:
self.assertTrue(torch.allclose(q, dq, atol=1e-4))
else:
self.assertFalse(torch.allclose(q, dq, atol=1e-4))
def quantize(self, model):
# Helper function that quantizes the given model
# Recursively convert all the `quant_mode` attributes as `True`
if hasattr(model, "quant_mode"):
model.quant_mode = True
elif isinstance(model, nn.Sequential):
for n, m in model.named_children():
self.quantize(m)
elif isinstance(model, nn.ModuleList):
for n in model:
self.quantize(n)
else:
for attr in dir(model):
mod = getattr(model, attr)
if isinstance(mod, nn.Module) and mod != model:
self.quantize(mod)
@slow
def test_inference_masked_lm(self):
# I-BERT should be "equivalent" to RoBERTa if not quantized
# Test coped from `test_modeling_roberta.py`
model = IBertForMaskedLM.from_pretrained("kssteven/ibert-roberta-base")
input_ids = torch.tensor([[0, 31414, 232, 328, 740, 1140, 12695, 69, 46078, 1588, 2]])
output = model(input_ids)[0]
expected_shape = torch.Size((1, 11, 50265))
self.assertEqual(output.shape, expected_shape)
expected_slice = torch.tensor(
[[[33.8802, -4.3103, 22.7761], [4.6539, -2.8098, 13.6253], [1.8228, -3.6898, 8.8600]]]
)
self.assertTrue(torch.allclose(output[:, :3, :3], expected_slice, atol=1e-4))
# I-BERT should be "similar" to RoBERTa if quantized
self.quantize(model)
output = model(input_ids)[0]
self.assertEqual(output.shape, expected_shape)
self.assertTrue(torch.allclose(output[:, :3, :3], expected_slice, atol=0.1))
@slow
def test_inference_classification_head(self):
# I-BERT should be "equivalent" to RoBERTa if not quantized
# Test coped from `test_modeling_roberta.py`
model = IBertForSequenceClassification.from_pretrained("kssteven/ibert-roberta-large-mnli")
input_ids = torch.tensor([[0, 31414, 232, 328, 740, 1140, 12695, 69, 46078, 1588, 2]])
output = model(input_ids)[0]
expected_shape = torch.Size((1, 3))
self.assertEqual(output.shape, expected_shape)
expected_tensor = torch.tensor([[-0.9469, 0.3913, 0.5118]])
self.assertTrue(torch.allclose(output, expected_tensor, atol=1e-4))
# I-BERT should be "similar" to RoBERTa if quantized
self.quantize(model)
output = model(input_ids)[0]
self.assertEqual(output.shape, expected_shape)
self.assertTrue(torch.allclose(output, expected_tensor, atol=0.1))
|
transformers/tests/models/ibert/test_modeling_ibert.py/0
|
{
"file_path": "transformers/tests/models/ibert/test_modeling_ibert.py",
"repo_id": "transformers",
"token_count": 14908
}
| 410
|
# coding=utf-8
# Copyright 2023 The HuggingFace Inc. team. 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.
"""Testing suite for the PyTorch InstructBLIP model."""
import inspect
import tempfile
import unittest
import numpy as np
import requests
from transformers import (
CONFIG_MAPPING,
InstructBlipConfig,
InstructBlipProcessor,
InstructBlipQFormerConfig,
InstructBlipVisionConfig,
)
from transformers.testing_utils import (
require_accelerate,
require_bitsandbytes,
require_torch,
require_vision,
slow,
torch_device,
)
from transformers.utils import is_torch_available, is_vision_available
from ...generation.test_utils import GenerationTesterMixin
from ...test_configuration_common import ConfigTester
from ...test_modeling_common import (
ModelTesterMixin,
floats_tensor,
ids_tensor,
random_attention_mask,
)
if is_torch_available():
import torch
from torch import nn
from transformers import InstructBlipForConditionalGeneration, InstructBlipVisionModel
if is_vision_available():
from PIL import Image
class InstructBlipVisionModelTester:
def __init__(
self,
parent,
batch_size=12,
image_size=30,
patch_size=2,
num_channels=3,
is_training=True,
hidden_size=32,
projection_dim=32,
num_hidden_layers=2,
num_attention_heads=4,
intermediate_size=37,
dropout=0.1,
attention_dropout=0.1,
initializer_range=1e-10,
scope=None,
):
self.parent = parent
self.batch_size = batch_size
self.image_size = image_size
self.patch_size = patch_size
self.num_channels = num_channels
self.is_training = is_training
self.hidden_size = hidden_size
self.projection_dim = projection_dim
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
self.intermediate_size = intermediate_size
self.dropout = dropout
self.attention_dropout = attention_dropout
self.initializer_range = initializer_range
self.scope = scope
# in case of a vision transformer, the seq length equals the number of patches + 1 (we add 1 for the [CLS] token)
num_patches = (image_size // patch_size) ** 2
self.seq_length = num_patches + 1
def prepare_config_and_inputs(self):
pixel_values = floats_tensor([self.batch_size, self.num_channels, self.image_size, self.image_size])
config = self.get_config()
return config, pixel_values
def get_config(self):
return InstructBlipVisionConfig(
image_size=self.image_size,
patch_size=self.patch_size,
num_channels=self.num_channels,
hidden_size=self.hidden_size,
projection_dim=self.projection_dim,
num_hidden_layers=self.num_hidden_layers,
num_attention_heads=self.num_attention_heads,
intermediate_size=self.intermediate_size,
dropout=self.dropout,
attention_dropout=self.attention_dropout,
initializer_range=self.initializer_range,
)
def create_and_check_model(self, config, pixel_values):
model = InstructBlipVisionModel(config=config)
model.to(torch_device)
model.eval()
with torch.no_grad():
result = model(pixel_values)
# expected sequence length = num_patches + 1 (we add 1 for the [CLS] token)
image_size = (self.image_size, self.image_size)
patch_size = (self.patch_size, self.patch_size)
num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, num_patches + 1, self.hidden_size))
self.parent.assertEqual(result.pooler_output.shape, (self.batch_size, self.hidden_size))
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
config, pixel_values = config_and_inputs
inputs_dict = {"pixel_values": pixel_values}
return config, inputs_dict
@require_torch
class InstructBlipVisionModelTest(ModelTesterMixin, unittest.TestCase):
"""
Here we also overwrite some of the tests of test_modeling_common.py, as InstructBLIP's vision encoder does not use input_ids, inputs_embeds,
attention_mask and seq_length.
"""
all_model_classes = (InstructBlipVisionModel,) if is_torch_available() else ()
fx_compatible = False
test_pruning = False
test_resize_embeddings = False
test_head_masking = False
def setUp(self):
self.model_tester = InstructBlipVisionModelTester(self)
self.config_tester = ConfigTester(
self, config_class=InstructBlipVisionConfig, has_text_modality=False, hidden_size=37
)
def test_config(self):
self.config_tester.run_common_tests()
@unittest.skip(reason="InstructBLIP's vision encoder does not use inputs_embeds")
def test_inputs_embeds(self):
pass
def test_model_get_set_embeddings(self):
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
model = model_class(config)
self.assertIsInstance(model.get_input_embeddings(), (nn.Module))
x = model.get_output_embeddings()
self.assertTrue(x is None or isinstance(x, nn.Linear))
def test_forward_signature(self):
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
model = model_class(config)
signature = inspect.signature(model.forward)
# signature.parameters is an OrderedDict => so arg_names order is deterministic
arg_names = [*signature.parameters.keys()]
expected_arg_names = ["pixel_values"]
self.assertListEqual(arg_names[:1], expected_arg_names)
def test_model(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_model(*config_and_inputs)
@unittest.skip(reason="InstructBlipVisionModel is an internal building block, doesn't support standalone training")
def test_training(self):
pass
@unittest.skip(reason="InstructBlipVisionModel is an internal building block, doesn't support standalone training")
def test_training_gradient_checkpointing(self):
pass
@unittest.skip(
reason="This architecure seem to not compute gradients properly when using GC, check: https://github.com/huggingface/transformers/pull/27124"
)
def test_training_gradient_checkpointing_use_reentrant(self):
pass
@unittest.skip(
reason="This architecure seem to not compute gradients properly when using GC, check: https://github.com/huggingface/transformers/pull/27124"
)
def test_training_gradient_checkpointing_use_reentrant_false(self):
pass
@unittest.skip(reason="InstructBlipVisionModel has no base class and is not available in MODEL_MAPPING")
def test_save_load_fast_init_from_base(self):
pass
@unittest.skip(reason="InstructBlipVisionModel has no base class and is not available in MODEL_MAPPING")
def test_save_load_fast_init_to_base(self):
pass
@slow
def test_model_from_pretrained(self):
model_name = "Salesforce/instructblip-flan-t5-xl"
model = InstructBlipVisionModel.from_pretrained(model_name)
self.assertIsNotNone(model)
class InstructBlipQFormerModelTester:
def __init__(
self,
parent,
batch_size=12,
seq_length=7,
is_training=True,
use_input_mask=True,
use_labels=True,
vocab_size=99,
hidden_size=32,
projection_dim=32,
num_hidden_layers=2,
num_attention_heads=4,
intermediate_size=37,
dropout=0.1,
attention_dropout=0.1,
max_position_embeddings=512,
initializer_range=0.02,
bos_token_id=0,
scope=None,
):
self.parent = parent
self.batch_size = batch_size
self.seq_length = seq_length
self.is_training = is_training
self.use_input_mask = use_input_mask
self.use_labels = use_labels
self.vocab_size = vocab_size
self.hidden_size = hidden_size
self.projection_dim = projection_dim
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
self.intermediate_size = intermediate_size
self.dropout = dropout
self.attention_dropout = attention_dropout
self.max_position_embeddings = max_position_embeddings
self.initializer_range = initializer_range
self.scope = scope
self.bos_token_id = bos_token_id
def prepare_config_and_inputs(self):
input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
qformer_input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
input_mask = None
if self.use_input_mask:
input_mask = random_attention_mask([self.batch_size, self.seq_length])
qformer_attention_mask = ids_tensor([self.batch_size, self.seq_length], vocab_size=2)
if input_mask is not None:
batch_size, seq_length = input_mask.shape
rnd_start_indices = np.random.randint(1, seq_length - 1, size=(batch_size,))
for batch_idx, start_index in enumerate(rnd_start_indices):
input_mask[batch_idx, :start_index] = 1
input_mask[batch_idx, start_index:] = 0
config = self.get_config()
return config, input_ids, input_mask, qformer_input_ids, qformer_attention_mask
def get_config(self):
return InstructBlipQFormerConfig(
vocab_size=self.vocab_size,
hidden_size=self.hidden_size,
projection_dim=self.projection_dim,
num_hidden_layers=self.num_hidden_layers,
num_attention_heads=self.num_attention_heads,
intermediate_size=self.intermediate_size,
dropout=self.dropout,
attention_dropout=self.attention_dropout,
max_position_embeddings=self.max_position_embeddings,
initializer_range=self.initializer_range,
bos_token_id=self.bos_token_id,
)
# this class is based on `OPTModelTester` found in tests/models/opt/test_modeling_opt.py
class InstructBlipTextModelDecoderOnlyTester:
def __init__(
self,
parent,
batch_size=12,
seq_length=7,
is_training=True,
use_labels=False,
vocab_size=99,
hidden_size=16,
num_hidden_layers=2,
num_attention_heads=4,
intermediate_size=4,
hidden_act="gelu",
hidden_dropout_prob=0.1,
attention_probs_dropout_prob=0.1,
max_position_embeddings=20,
eos_token_id=2,
pad_token_id=1,
bos_token_id=0,
embed_dim=16,
num_labels=3,
word_embed_proj_dim=16,
type_sequence_label_size=2,
):
self.parent = parent
self.batch_size = batch_size
self.seq_length = seq_length
self.is_training = is_training
self.use_labels = use_labels
self.vocab_size = vocab_size
self.hidden_size = hidden_size
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
self.intermediate_size = intermediate_size
self.hidden_act = hidden_act
self.hidden_dropout_prob = hidden_dropout_prob
self.attention_probs_dropout_prob = attention_probs_dropout_prob
self.max_position_embeddings = max_position_embeddings
self.eos_token_id = eos_token_id
self.pad_token_id = pad_token_id
self.bos_token_id = bos_token_id
self.embed_dim = embed_dim
self.num_labels = num_labels
self.type_sequence_label_size = type_sequence_label_size
self.word_embed_proj_dim = word_embed_proj_dim
self.is_encoder_decoder = False
def prepare_config_and_inputs(self):
config = self.get_config()
input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size).clamp(3)
input_ids[:, -1] = self.eos_token_id # Eos Token
attention_mask = input_ids.ne(self.pad_token_id)
return config, input_ids, attention_mask
def get_config(self):
return CONFIG_MAPPING["opt"](
vocab_size=self.vocab_size,
hidden_size=self.hidden_size,
num_hidden_layers=self.num_hidden_layers,
num_attention_heads=self.num_attention_heads,
ffn_dim=self.intermediate_size,
dropout=self.hidden_dropout_prob,
attention_dropout=self.attention_probs_dropout_prob,
max_position_embeddings=self.max_position_embeddings,
eos_token_id=self.eos_token_id,
bos_token_id=self.bos_token_id,
pad_token_id=self.pad_token_id,
embed_dim=self.embed_dim,
is_encoder_decoder=False,
word_embed_proj_dim=self.word_embed_proj_dim,
)
# this model tester uses a decoder-only language model (OPT)
class InstructBlipForConditionalGenerationDecoderOnlyModelTester:
def __init__(
self, parent, vision_kwargs=None, qformer_kwargs=None, text_kwargs=None, is_training=True, num_query_tokens=10
):
if vision_kwargs is None:
vision_kwargs = {}
if qformer_kwargs is None:
qformer_kwargs = {}
if text_kwargs is None:
text_kwargs = {}
self.parent = parent
self.vision_model_tester = InstructBlipVisionModelTester(parent, **vision_kwargs)
self.qformer_model_tester = InstructBlipQFormerModelTester(parent, **qformer_kwargs)
self.text_model_tester = InstructBlipTextModelDecoderOnlyTester(parent, **text_kwargs)
self.batch_size = self.text_model_tester.batch_size # need bs for batching_equivalence test
self.seq_length = self.text_model_tester.seq_length # need seq_length for common tests
self.is_training = is_training
self.num_query_tokens = num_query_tokens
def prepare_config_and_inputs(self):
_, pixel_values = self.vision_model_tester.prepare_config_and_inputs()
_, _, _, qformer_input_ids, qformer_attention_mask = self.qformer_model_tester.prepare_config_and_inputs()
_, input_ids, attention_mask = self.text_model_tester.prepare_config_and_inputs()
config = self.get_config()
return config, input_ids, attention_mask, qformer_input_ids, qformer_attention_mask, pixel_values
def get_config(self):
return InstructBlipConfig.from_vision_qformer_text_configs(
vision_config=self.vision_model_tester.get_config(),
qformer_config=self.qformer_model_tester.get_config(),
text_config=self.text_model_tester.get_config(),
num_query_tokens=self.num_query_tokens,
)
def create_and_check_for_conditional_generation(
self, config, input_ids, attention_mask, qformer_input_ids, qformer_attention_mask, pixel_values
):
model = InstructBlipForConditionalGeneration(config).to(torch_device).eval()
with torch.no_grad():
result = model(
pixel_values,
input_ids=input_ids,
attention_mask=attention_mask,
qformer_input_ids=qformer_input_ids,
qformer_attention_mask=qformer_attention_mask,
)
expected_seq_length = self.num_query_tokens + self.text_model_tester.seq_length
self.parent.assertEqual(
result.logits.shape,
(self.vision_model_tester.batch_size, expected_seq_length, self.text_model_tester.vocab_size),
)
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
config, input_ids, attention_mask, qformer_input_ids, qformer_attention_mask, pixel_values = config_and_inputs
inputs_dict = {
"pixel_values": pixel_values,
"input_ids": input_ids,
"attention_mask": attention_mask,
"qformer_input_ids": qformer_input_ids,
"qformer_attention_mask": qformer_attention_mask,
"labels": input_ids,
}
return config, inputs_dict
@require_torch
class InstructBlipForConditionalGenerationDecoderOnlyTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):
all_model_classes = (InstructBlipForConditionalGeneration,) if is_torch_available() else ()
fx_compatible = False
test_head_masking = False
test_pruning = False
test_resize_embeddings = False
test_attention_outputs = False
test_torchscript = False
def setUp(self):
self.model_tester = InstructBlipForConditionalGenerationDecoderOnlyModelTester(self)
def test_for_conditional_generation(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_for_conditional_generation(*config_and_inputs)
@unittest.skip(reason="Hidden_states is tested in individual model tests")
def test_hidden_states_output(self):
pass
@unittest.skip(reason="InstructBlipForConditionalGeneration doesn't support inputs_embeds")
def test_inputs_embeds(self):
pass
@unittest.skip(reason="Tied weights are tested in individual model tests")
def test_tied_weights_keys(self):
pass
@unittest.skip(reason="Retain_grad is tested in individual model tests")
def test_retain_grad_hidden_states_attentions(self):
pass
@unittest.skip(reason="InstructBlipModel does not have input/output embeddings")
def test_model_get_set_embeddings(self):
pass
@unittest.skip(reason="There's no base InstructBlipModel")
def test_save_load_fast_init_from_base(self):
pass
@unittest.skip(reason="There's no base InstructBlipModel")
def test_save_load_fast_init_to_base(self):
pass
def test_forward_signature(self):
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
model = model_class(config)
signature = inspect.signature(model.forward)
# signature.parameters is an OrderedDict => so arg_names order is deterministic
arg_names = [*signature.parameters.keys()]
expected_arg_names = ["pixel_values"]
self.assertListEqual(arg_names[:1], expected_arg_names)
def test_load_vision_qformer_text_config(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
# Save InstructBlipConfig and check if we can load InstructBlipVisionConfig from it
with tempfile.TemporaryDirectory() as tmp_dir_name:
config.save_pretrained(tmp_dir_name)
vision_config = InstructBlipVisionConfig.from_pretrained(tmp_dir_name)
self.assertDictEqual(config.vision_config.to_dict(), vision_config.to_dict())
# Save InstructBlipConfig and check if we can load InstructBlipQFormerConfig from it
with tempfile.TemporaryDirectory() as tmp_dir_name:
config.save_pretrained(tmp_dir_name)
qformer_config = InstructBlipQFormerConfig.from_pretrained(tmp_dir_name)
self.assertDictEqual(config.qformer_config.to_dict(), qformer_config.to_dict())
@slow
def test_model_from_pretrained(self):
model_name = "Salesforce/instructblip-flan-t5-xl"
model = InstructBlipForConditionalGeneration.from_pretrained(model_name)
self.assertIsNotNone(model)
# We will verify our results on an image of cute cats
def prepare_img():
url = "https://huggingface.co/hf-internal-testing/blip-test-image/resolve/main/demo.jpg"
image = Image.open(requests.get(url, stream=True).raw)
return image
@require_vision
@require_torch
@slow
class InstructBlipModelIntegrationTest(unittest.TestCase):
@require_bitsandbytes
@require_accelerate
def test_inference_vicuna_7b(self):
processor = InstructBlipProcessor.from_pretrained("Salesforce/instructblip-vicuna-7b")
model = InstructBlipForConditionalGeneration.from_pretrained(
"Salesforce/instructblip-vicuna-7b", load_in_8bit=True, low_cpu_mem_usage=True
)
url = "https://raw.githubusercontent.com/salesforce/LAVIS/main/docs/_static/Confusing-Pictures.jpg"
image = Image.open(requests.get(url, stream=True).raw).convert("RGB")
prompt = "What is unusual about this image?"
inputs = processor(images=image, text=prompt, return_tensors="pt").to(torch_device, torch.float16)
# verify logits
with torch.no_grad():
logits = model(**inputs).logits
expected_slice = torch.tensor(
[[-3.3926, -12.2969, 8.4922], [-5.0195, -11.9531, 8.1406], [-4.0039, -13.3594, 9.2578]],
device=torch_device,
)
self.assertTrue(torch.allclose(logits[0, :3, :3].float(), expected_slice, atol=1e-3))
# verify generation
outputs = model.generate(**inputs, max_new_tokens=30)
generated_text = processor.batch_decode(outputs, skip_special_tokens=True)[0].strip()
expected_outputs = [2, 450, 22910, 9565, 310, 445, 1967, 338, 393, 263, 767, 338, 13977, 292, 22095, 373, 278, 1250, 310, 263, 13328, 20134, 29963, 1550, 372, 338, 19500, 1623, 263, 19587, 4272] # fmt: off
self.assertEqual(outputs[0].tolist(), expected_outputs)
self.assertEqual(
generated_text,
"The unusual aspect of this image is that a man is ironing clothes on the back of a yellow SUV while it is driving down a busy city",
)
def test_inference_flant5_xl(self):
processor = InstructBlipProcessor.from_pretrained("Salesforce/instructblip-flan-t5-xl")
model = InstructBlipForConditionalGeneration.from_pretrained(
"Salesforce/instructblip-flan-t5-xl",
torch_dtype=torch.bfloat16,
low_cpu_mem_usage=True,
).to(torch_device)
url = "https://raw.githubusercontent.com/salesforce/LAVIS/main/docs/_static/Confusing-Pictures.jpg"
image = Image.open(requests.get(url, stream=True).raw).convert("RGB")
prompt = "What is unusual about this image?"
inputs = processor(images=image, text=prompt, return_tensors="pt").to(torch_device)
for k, v in inputs.items():
if torch.is_floating_point(v):
inputs[k] = v.to(torch.bfloat16)
outputs = model.generate(
**inputs,
do_sample=False,
num_beams=5,
max_length=256,
min_length=1,
top_p=0.9,
repetition_penalty=1.5,
length_penalty=1.0,
temperature=1,
)
generated_text = processor.batch_decode(outputs, skip_special_tokens=True)[0]
expected_outputs = [0, 37, 1023, 9850, 7, 3, 9, 388, 3575, 53, 4954, 30, 8, 223, 13, 3, 9, 4459, 4049, 16, 8, 2214, 13, 3, 9, 3164, 690, 2815, 5, 37, 388, 19, 5119, 3, 9, 4459, 8677, 28, 3, 9, 2756, 4459, 6177, 6, 11, 3, 88, 19, 338, 46, 3575, 53, 1476, 12, 743, 112, 2491, 5, 37, 1023, 19, 7225, 788, 12, 8, 685, 24, 34, 1267, 3, 9, 388, 3575, 53, 4954, 30, 8, 223, 13, 3, 9, 4049, 16, 8, 2214, 13, 3, 9, 3164, 690, 2815, 5, 94, 19, 487, 24, 8, 388, 19, 1119, 12, 1097, 540, 57, 692, 112, 10428, 30, 8, 223, 13, 8, 4049, 6, 68, 34, 19, 92, 487, 24, 3, 88, 19, 1119, 12, 1097, 97, 57, 692, 112, 10428, 30, 8, 223, 13, 8, 4049, 16, 8, 2214, 13, 3, 9, 3164, 690, 2815, 5, 3, 13865, 13, 8, 1053, 21, 8, 388, 31, 7, 2874, 6, 34, 19, 964, 24, 3, 88, 19, 1119, 12, 1097, 97, 57, 692, 112, 10428, 30, 8, 223, 13, 8, 4049, 16, 8, 2214, 13, 3, 9, 3164, 690, 2815, 5, 1] # fmt: skip
expected_outputs = [0, 37, 7225, 1023, 9850, 7, 3, 9, 388, 3575, 53, 4954, 30, 8, 223, 13, 3, 9, 4459, 4049, 16, 8, 2214, 13, 3, 9, 3164, 690, 2815, 5, 37, 388, 19, 5119, 3, 9, 4459, 8677, 28, 46, 3575, 53, 1476, 5223, 12, 34, 6, 15495, 24, 3, 88, 19, 692, 112, 293, 10428, 44, 234, 1066, 145, 338, 3, 9, 50, 1106, 3522, 144, 42, 2192, 7919, 31, 7, 5, 37, 1023, 92, 1267, 3, 9, 381, 13, 119, 3203, 16, 8, 2458, 6, 379, 14264, 6, 9256, 7, 6, 11, 11718, 7, 5, 1] # fmt: skip
self.assertEqual(outputs[0].tolist(), expected_outputs)
self.assertEqual(
generated_text,
"The unusual image depicts a man ironing clothes on the back of a yellow van in the middle of a busy city street. The man is wearing a yellow shirt with an ironing board attached to it, suggesting that he is doing his own laundry at home rather than using a laundromat or dry cleaner's. The image also shows a number of other vehicles in the background, including buses, taxis, and motorcycles.",
)
def test_inference_interpolate_pos_encoding(self):
processor = InstructBlipProcessor.from_pretrained("Salesforce/instructblip-flan-t5-xl")
model = InstructBlipForConditionalGeneration.from_pretrained(
"Salesforce/instructblip-flan-t5-xl",
torch_dtype=torch.bfloat16,
low_cpu_mem_usage=True,
).to(torch_device)
processor.image_processor.size = {"height": 500, "width": 500}
image = prepare_img()
prompt = "What's in the image?"
inputs = processor(images=image, text=prompt, return_tensors="pt").to(torch_device)
predictions = model.generate(**inputs, interpolate_pos_encoding=True)
generated_text = processor.batch_decode(predictions, skip_special_tokens=True)[0].strip()
self.assertEqual(
predictions[0].tolist(), [0, 37, 1023, 753, 3, 9, 2335, 3823, 30, 8, 2608, 28, 3, 9, 1782, 5, 1]
)
self.assertEqual(generated_text, "The image features a woman sitting on the beach with a dog.")
def test_expansion_in_processing(self):
processor = InstructBlipProcessor.from_pretrained("Salesforce/instructblip-flan-t5-xl")
model = InstructBlipForConditionalGeneration.from_pretrained(
"Salesforce/instructblip-flan-t5-xl",
torch_dtype=torch.bfloat16,
low_cpu_mem_usage=True,
).to(torch_device)
image = prepare_img()
prompt = "What's in the image?"
# Make sure we will go the legacy path by setting these args to None
processor.num_query_tokens = None
model.config.image_token_index = None
inputs = processor(images=image, text=prompt, return_tensors="pt").to(torch_device, dtype=torch.float16)
predictions = model.generate(**inputs, do_sample=False, max_new_tokens=15)
generated_text = processor.batch_decode(predictions, skip_special_tokens=True)[0].strip()
# Add args to the config to trigger new logic when inputs are expanded in processing file
processor.num_query_tokens = model.config.num_query_tokens
processor.tokenizer.add_special_tokens({"additional_special_tokens": ["<image>"]})
model.config.image_token_index = len(processor.tokenizer) - 1
model.resize_token_embeddings(processor.tokenizer.vocab_size, pad_to_multiple_of=64)
# Generate again with new inputs
inputs = processor(images=image, text=prompt, return_tensors="pt").to(torch_device, dtype=torch.float16)
predictions_expanded = model.generate(**inputs, do_sample=False, max_new_tokens=15)
generated_text_expanded = processor.batch_decode(predictions_expanded, skip_special_tokens=True)[0].strip()
self.assertTrue(generated_text_expanded == generated_text)
|
transformers/tests/models/instructblip/test_modeling_instructblip.py/0
|
{
"file_path": "transformers/tests/models/instructblip/test_modeling_instructblip.py",
"repo_id": "transformers",
"token_count": 12293
}
| 411
|
# coding=utf-8
# Copyright Iz Beltagy, Matthew E. Peters, Arman Cohan and The HuggingFace Inc. team. 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.
from __future__ import annotations
import unittest
from transformers import LEDConfig, is_tf_available
from transformers.testing_utils import require_tf, slow
from ...test_configuration_common import ConfigTester
from ...test_modeling_tf_common import TFModelTesterMixin, ids_tensor
from ...test_pipeline_mixin import PipelineTesterMixin
if is_tf_available():
import tensorflow as tf
from transformers import TFLEDForConditionalGeneration, TFLEDModel
@require_tf
class TFLEDModelTester:
config_cls = LEDConfig
config_updates = {}
hidden_act = "gelu"
def __init__(
self,
parent,
batch_size=13,
seq_length=7,
is_training=True,
use_labels=False,
vocab_size=99,
hidden_size=32,
num_hidden_layers=2,
num_attention_heads=4,
intermediate_size=37,
hidden_dropout_prob=0.1,
attention_probs_dropout_prob=0.1,
max_position_embeddings=20,
eos_token_id=2,
pad_token_id=1,
bos_token_id=0,
attention_window=4,
):
self.parent = parent
self.batch_size = batch_size
self.seq_length = seq_length
self.is_training = is_training
self.use_labels = use_labels
self.vocab_size = vocab_size
self.hidden_size = hidden_size
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
self.intermediate_size = intermediate_size
self.hidden_dropout_prob = hidden_dropout_prob
self.attention_probs_dropout_prob = attention_probs_dropout_prob
self.max_position_embeddings = max_position_embeddings
self.eos_token_id = eos_token_id
self.pad_token_id = pad_token_id
self.bos_token_id = bos_token_id
self.attention_window = attention_window
# `ModelTesterMixin.test_attention_outputs` is expecting attention tensors to be of size
# [num_attention_heads, encoder_seq_length, encoder_key_length], but TFLongformerSelfAttention
# returns attention of shape [num_attention_heads, encoder_seq_length, self.attention_window + 1]
# because its local attention only attends to `self.attention_window` and one before and one after
self.key_length = self.attention_window + 2
# because of padding `encoder_seq_length`, is different from `seq_length`. Relevant for
# the `test_attention_outputs` and `test_hidden_states_output` tests
self.encoder_seq_length = (
self.seq_length + (self.attention_window - self.seq_length % self.attention_window) % self.attention_window
)
def prepare_config_and_inputs_for_common(self):
input_ids = ids_tensor([self.batch_size, self.seq_length - 1], self.vocab_size)
eos_tensor = tf.expand_dims(tf.constant([self.eos_token_id] * self.batch_size), 1)
input_ids = tf.concat([input_ids, eos_tensor], axis=1)
decoder_input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
config = self.config_cls(
vocab_size=self.vocab_size,
d_model=self.hidden_size,
encoder_layers=self.num_hidden_layers,
decoder_layers=self.num_hidden_layers,
encoder_attention_heads=self.num_attention_heads,
decoder_attention_heads=self.num_attention_heads,
encoder_ffn_dim=self.intermediate_size,
decoder_ffn_dim=self.intermediate_size,
dropout=self.hidden_dropout_prob,
attention_dropout=self.attention_probs_dropout_prob,
max_position_embeddings=self.max_position_embeddings,
eos_token_ids=[2],
bos_token_id=self.bos_token_id,
pad_token_id=self.pad_token_id,
decoder_start_token_id=self.pad_token_id,
attention_window=self.attention_window,
**self.config_updates,
)
inputs_dict = prepare_led_inputs_dict(config, input_ids, decoder_input_ids)
global_attention_mask = tf.concat(
[tf.zeros_like(input_ids)[:, :-1], tf.ones_like(input_ids)[:, -1:]],
axis=-1,
)
inputs_dict["global_attention_mask"] = global_attention_mask
return config, inputs_dict
def check_decoder_model_past_large_inputs(self, config, inputs_dict):
model = TFLEDModel(config=config).get_decoder()
input_ids = inputs_dict["input_ids"]
input_ids = input_ids[:1, :]
attention_mask = inputs_dict["attention_mask"][:1, :]
self.batch_size = 1
# first forward pass
outputs = model(input_ids, attention_mask=attention_mask, use_cache=True)
output, past_key_values = outputs.to_tuple()
# create hypothetical next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
next_attn_mask = tf.cast(ids_tensor((self.batch_size, 3), 2), tf.int8)
# append to next input_ids and
next_input_ids = tf.concat([input_ids, next_tokens], axis=-1)
next_attention_mask = tf.concat([attention_mask, next_attn_mask], axis=-1)
output_from_no_past = model(next_input_ids, attention_mask=next_attention_mask)[0]
output_from_past = model(next_tokens, attention_mask=next_attention_mask, past_key_values=past_key_values)[0]
self.parent.assertEqual(next_tokens.shape[1], output_from_past.shape[1])
# select random slice
random_slice_idx = int(ids_tensor((1,), output_from_past.shape[-1]))
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx]
output_from_past_slice = output_from_past[:, :, random_slice_idx]
# test that outputs are equal for slice
tf.debugging.assert_near(output_from_past_slice, output_from_no_past_slice, rtol=1e-3)
def prepare_led_inputs_dict(
config,
input_ids,
decoder_input_ids,
attention_mask=None,
decoder_attention_mask=None,
head_mask=None,
decoder_head_mask=None,
):
if attention_mask is None:
attention_mask = tf.cast(tf.math.not_equal(input_ids, config.pad_token_id), tf.int8)
if decoder_attention_mask is None:
decoder_attention_mask = tf.concat(
[
tf.ones(decoder_input_ids[:, :1].shape, dtype=tf.int8),
tf.cast(tf.math.not_equal(decoder_input_ids[:, 1:], config.pad_token_id), tf.int8),
],
axis=-1,
)
if head_mask is None:
head_mask = tf.ones((config.encoder_layers, config.encoder_attention_heads))
if decoder_head_mask is None:
decoder_head_mask = tf.ones((config.decoder_layers, config.decoder_attention_heads))
return {
"input_ids": input_ids,
"attention_mask": attention_mask,
"decoder_input_ids": decoder_input_ids,
"decoder_attention_mask": decoder_attention_mask,
"head_mask": head_mask,
"decoder_head_mask": decoder_head_mask,
}
@require_tf
class TFLEDModelTest(TFModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
all_model_classes = (TFLEDForConditionalGeneration, TFLEDModel) if is_tf_available() else ()
all_generative_model_classes = (TFLEDForConditionalGeneration,) if is_tf_available() else ()
pipeline_model_mapping = (
{
"feature-extraction": TFLEDModel,
"summarization": TFLEDForConditionalGeneration,
"text2text-generation": TFLEDForConditionalGeneration,
"translation": TFLEDForConditionalGeneration,
}
if is_tf_available()
else {}
)
is_encoder_decoder = True
test_pruning = False
test_head_masking = False
test_onnx = False
def setUp(self):
self.model_tester = TFLEDModelTester(self)
self.config_tester = ConfigTester(self, config_class=LEDConfig)
def test_config(self):
self.config_tester.run_common_tests()
def test_decoder_model_past_large_inputs(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs_for_common()
self.model_tester.check_decoder_model_past_large_inputs(*config_and_inputs)
def test_attention_outputs(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
inputs_dict["global_attention_mask"] = tf.zeros_like(inputs_dict["attention_mask"])
num_global_attn_indices = 2
inputs_dict["global_attention_mask"] = tf.where(
tf.range(self.model_tester.seq_length)[None, :] < num_global_attn_indices,
1,
inputs_dict["global_attention_mask"],
)
config.return_dict = True
seq_length = self.model_tester.seq_length
encoder_seq_length = self.model_tester.encoder_seq_length
def check_decoder_attentions_output(outputs):
decoder_attentions = outputs.decoder_attentions
self.assertEqual(len(decoder_attentions), self.model_tester.num_hidden_layers)
self.assertListEqual(
list(decoder_attentions[0].shape[-3:]),
[self.model_tester.num_attention_heads, seq_length, seq_length],
)
def check_encoder_attentions_output(outputs):
attentions = [t.numpy() for t in outputs.encoder_attentions]
global_attentions = [t.numpy() for t in outputs.encoder_global_attentions]
self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)
self.assertEqual(len(global_attentions), self.model_tester.num_hidden_layers)
self.assertListEqual(
list(attentions[0].shape[-3:]),
[self.model_tester.num_attention_heads, seq_length, seq_length],
)
self.assertListEqual(
list(global_attentions[0].shape[-3:]),
[self.model_tester.num_attention_heads, encoder_seq_length, num_global_attn_indices],
)
for model_class in self.all_model_classes:
inputs_dict["output_attentions"] = True
inputs_dict["use_cache"] = False
config.output_hidden_states = False
model = model_class(config)
outputs = model(self._prepare_for_class(inputs_dict, model_class))
out_len = len(outputs)
self.assertEqual(config.output_hidden_states, False)
check_encoder_attentions_output(outputs)
if self.is_encoder_decoder:
model = model_class(config)
outputs = model(self._prepare_for_class(inputs_dict, model_class))
self.assertEqual(config.output_hidden_states, False)
check_decoder_attentions_output(outputs)
# Check that output attentions can also be changed via the config
del inputs_dict["output_attentions"]
config.output_attentions = True
model = model_class(config)
outputs = model(self._prepare_for_class(inputs_dict, model_class))
self.assertEqual(config.output_hidden_states, False)
check_encoder_attentions_output(outputs)
# Check attention is always last and order is fine
inputs_dict["output_attentions"] = True
config.output_hidden_states = True
model = model_class(config)
outputs = model(self._prepare_for_class(inputs_dict, model_class))
self.assertEqual(out_len + (2 if self.is_encoder_decoder else 1), len(outputs))
self.assertEqual(model.config.output_hidden_states, True)
check_encoder_attentions_output(outputs)
@unittest.skip("LED keeps using potentially symbolic tensors in conditionals and breaks tracing.")
def test_saved_model_creation(self):
pass
def test_generate_with_headmasking(self):
# TODO: Head-masking not yet implement
pass
def _long_tensor(tok_lst):
return tf.constant(tok_lst, dtype=tf.int32)
TOLERANCE = 1e-4
@slow
@require_tf
class TFLEDModelIntegrationTest(unittest.TestCase):
def test_inference_no_head(self):
model = TFLEDForConditionalGeneration.from_pretrained("allenai/led-base-16384").led
# change to intended input here
input_ids = _long_tensor([512 * [0, 31414, 232, 328, 740, 1140, 12695, 69]])
decoder_input_ids = _long_tensor([128 * [0, 31414, 232, 328, 740, 1140, 12695, 69]])
inputs_dict = prepare_led_inputs_dict(model.config, input_ids, decoder_input_ids)
output = model(**inputs_dict)[0]
expected_shape = (1, 1024, 768)
self.assertEqual(output.shape, expected_shape)
# change to expected output here
expected_slice = tf.convert_to_tensor(
[[2.3050, 2.8279, 0.6531], [-1.8457, -0.1455, -3.5661], [-1.0186, 0.4586, -2.2043]],
)
tf.debugging.assert_near(output[:, :3, :3], expected_slice, atol=1e-3)
def test_inference_with_head(self):
model = TFLEDForConditionalGeneration.from_pretrained("allenai/led-base-16384")
# change to intended input here
input_ids = _long_tensor([512 * [0, 31414, 232, 328, 740, 1140, 12695, 69]])
decoder_input_ids = _long_tensor([128 * [0, 31414, 232, 328, 740, 1140, 12695, 69]])
inputs_dict = prepare_led_inputs_dict(model.config, input_ids, decoder_input_ids)
output = model(**inputs_dict)[0]
expected_shape = (1, 1024, model.config.vocab_size)
self.assertEqual(output.shape, expected_shape)
# change to expected output here
expected_slice = tf.convert_to_tensor(
[[33.6507, 6.4572, 16.8089], [5.8739, -2.4238, 11.2902], [-3.2139, -4.3149, 4.2783]],
)
tf.debugging.assert_near(output[:, :3, :3], expected_slice, atol=1e-3, rtol=1e-3)
|
transformers/tests/models/led/test_modeling_tf_led.py/0
|
{
"file_path": "transformers/tests/models/led/test_modeling_tf_led.py",
"repo_id": "transformers",
"token_count": 6429
}
| 412
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.