SignX/training_phoenix/README.md

Walk-Through for PHOENIX-2014T

This document describes the rough procedure to train a SLTUnet model.

Step 1. Download and Preprocess Dataset

1. Get the phoenix2014T dataset from here or using

   wget https://www-i6.informatik.rwth-aachen.de/ftp/pub/rwth-phoenix/2016/phoenix-2014-T.v3.tar.gz
   

2. Get MuST-C En-De dataset from FBK; note we used the data in v1.0

   We applied tokenization and subword modeling to these dataset. See preprocess_phoenix.sh for reference.

Step 2. Pretrain SMKD Embeddings

We adopt the SMKD method to pretrain sign embeddings and further adapt it for sign language translation. smkd shows the adapted source code.

To pretrain SMKD embeddings,

1. preprocess the dataset

   python preprocess/dataset_preprocess.py --dataset phoenix2014 --dataset-root PHOENIX-2014
   -T-release-v3/PHOENIX-2014-T/
   

2. launch training

   python main.py --work-dir exp/resnet34 --config baseline.yaml --device 0,1
   

3. checkpoint averaging (optional)

   Among all saved checkpoints, select top-K (e.g. 5) checkpoint and put their (abs)path into a file named checkpoint under exp/resnet34

   python ckpt_avg.py  --path exp/resnet34 --checkpoints 5 --output avg
   

4. extract sign features

   python main.py --load-weights avg/average.pt --phase features --device 0 --num-feature-aug 10 --work-dir exp/resnet34 --config baseline.yaml
   

   Then combine different training features

   python sign_feature_cmb.py train\*h5 
   

At the end, you will have train/dev/test.h5 files as the sign feature inputs

Step 3. Train SLTUnet Model

See the given running scripts train.sh for reference.

Step 4. Decoding and Evaluation

1. we saved top-10 checkpoints based on dev set performance. we averaged them before final evaluation.

   python models/checkpoint_averaging.py  --path path-to-best-ckpt-dir --checkpoints 10 --output avg --gpu 0
   

2. See the given running scripts test.sh for decoding.

3. Regarding evaluation, please checkout eval/metrics.py for details.

   For future evaluation and dataset construction, we suggest retaining the punctuations and

adopt detokenized BLEU. E.g.

```bash
python eval/metrics.py -t slt -hyp model-output-file -ref gold-reference-file
```