Noisy student training for wenet #1600
Conversation
|
@robin1001 hi Binbin, would you mind reviewing this PR? Thanks!! |
wenet/bin/train_nst.py
Outdated
| print("unsupervised data list = ", args.train_data_unsupervised) | ||
|
|
||
| cv_conf = copy.deepcopy(train_supervised_conf) | ||
| # cv_conf['speed_perturb'] = False |
There was a problem hiding this comment.
why use speed_perturb and spec_aug in cv?
There was a problem hiding this comment.
Hi robin, thanks for pointing out this issue, the speed_perturb, spech_aug as well as shuffle should be set to false for CV. I traced back my history and found it was a mis-comment during my recent review. In our experiment cluster, it was set to be False.
|
The flake8-bugbear is updated to version 22.12.6 from 22.10.27 and they modified the test for zip function. |
fixed, #1604, plz rebase your commits |
|
It works now, thanks! |
wenet/bin/train_nst.py
Outdated
| pin_memory=args.pin_memory, | ||
| num_workers=args.num_workers, | ||
| prefetch_factor=args.prefetch) | ||
| executor = Executor_nst() |
There was a problem hiding this comment.
better to Name it ExecutorNst
wenet/bin/train_nst.py
Outdated
| from tensorboardX import SummaryWriter | ||
| from torch.utils.data import DataLoader | ||
| from wenet.dataset.dataset import Dataset | ||
| from wenet.transformer.asr_model import init_asr_model |
There was a problem hiding this comment.
What's your base commits? we have refactor our model initialization in this PR #1280, call wenet.utils.init_model.init_model instead of wenet.transformer.asr_model.init_asr_model
TODO: modify run_nst.sh , remove train_nst, exector_nst
modifyed README.md
|
Could you please review this PR? @robin1001 and @xingchensong Thanks!!^^ |
wenet/utils/executor_nst.py
Outdated
| def __init__(self): | ||
| self.step = 0 | ||
|
|
||
| def train(self, model, optimizer, scheduler, data_loader_aishell, |
There was a problem hiding this comment.
better to name it
data_loader_supervised and data_loader_unsupervised
| --val_best | ||
| fi | ||
|
|
||
| # export model |
There was a problem hiding this comment.
export_jit is not required here.
| @@ -0,0 +1,11 @@ | |||
| data/train/shards/shards_000000000.tar | |||
| @@ -0,0 +1,143 @@ | |||
| # network architecture | |||
There was a problem hiding this comment.
This file is not required, right?
| @@ -0,0 +1,11 @@ | |||
| data/train/shards/shards_000000000.tar | |||
There was a problem hiding this comment.
please remove data_example/train dir
| @@ -0,0 +1 @@ | |||
| 你好你好 | |||
| @@ -0,0 +1,2 @@ | |||
| data/train/wenet_1khr_tar//dir0_000000.tar | |||
| data/train/wenet_1khr_tar//dir0_000001.tar | |||
examples/aishell/NST/run_nst.sh
Outdated
| if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ] && [ ${enable_nst} -eq 0 ]; then | ||
| echo "********step 3 start time : $now ********" | ||
| python split_data_list.py \ |
There was a problem hiding this comment.
python local/split_data_list.py
|
Great Job, LGTM. |
The data example directory, train_nst.py and executor.py has been removed, the problems you mentioned should be sovled. |
Here, we provide a recipe to run Noisy Student Training (NST) with LM filter strategy using AISHELL-1 as supervised data and WenetSpeech as unsupervised data from our paper.
The example codes are stored under examples/aishell/NST with a detailed guideline and results in readme.md. we mainly modified the script for "run.sh" and added "executor_nst.py" ,"train_nst.py" as well as some auxiliary codes and examples under local directory.