issue while fine tuning the wav2vec model

🐛 Bug

When i run the fine tuning script on the wav2vec trained model. I’m getting below error everytime.

Traceback (most recent call last):
  File "fairseq_cli/hydra_train.py", line 45, in hydra_main
    distributed_utils.call_main(cfg, pre_main)
  File "/home/robot/airobotics/speech/fairseq/fairseq/distributed/utils.py", line 366, in call_main
    main(cfg, **kwargs)
  File "/home/robot/airobotics/speech/fairseq/fairseq_cli/train.py", line 85, in main
    task.load_dataset(valid_sub_split, combine=False, epoch=1)
  File "/home/robot/airobotics/speech/fairseq/fairseq/tasks/audio_pretraining.py", line 206, in load_dataset
    **self._get_mask_precompute_kwargs(task_cfg),
  File "/home/robot/airobotics/speech/fairseq/fairseq/data/audio/raw_audio_dataset.py", line 256, in __init__
    with open(manifest_path, "r") as f:
FileNotFoundError: [Errno 2] No such file or directory: '/home/robot/dataset/fairseq/libri_speech/audio/dev_other.tsv'

I’m following this documentation to train and fine tune the wav2vec model on librespeech dataset.

To Reproduce

Steps to reproduce the behavior (always include the command you ran): I have generated the manifest and label data according to the instructions providd in that page. But when i run this command: fairseq-hydra-train task.data=/home/robot/dataset/fairseq/libri_speech/audio model.w2v_path=/home/robot/dataset/fairseq/libri_speech/models/wav2vec_small.pt --config-dir config/finetuning/ --config-name base_100h I’m getting the above error.

Environment

fairseq branch: master os: ubuntu 20.04 cuda: nvcc: NVIDIA ® Cuda compiler driver Copyright © 2005-2021 NVIDIA Corporation Built on Sun_Feb_14_21:12:58_PST_2021 Cuda compilation tools, release 11.2, V11.2.152 Build cuda_11.2.r11.2/compiler.29618528_0

Additional context

Since the documentation fore the speech recognition with wav2vec is little complex in the readme gude provided. Lots of developers are facing issue and taking longer time to use this framework. So i think it is better if the documentation guide is improved for better understanding and less error.

[dzubke]

If you look at the base_100h config file at examples/wav2vec/configs/fintuning/base_100h.yaml you’ll see that the dataset.valid_subset field is set to dev_other which means the script is looking in your data directory (/home/robot/dataset/fairseq/libri_speech/audio/) for the files: dev_other.tsv, dev_other.wrd, and dev_other.ltr (if you’re doing character recognition) in the same way as it is looking for your training dataset files.

I don’t know how to turn validation off, so you’ll have to create a dev/validation set in the same way as you created your training set and put those dev_other files in the data directory. If your validation set is named something else like valid.tsv you can just change the value in the base_100h.yaml file for the dataset.valid_subset field.

[shuvohishab]

If you look at the base_100h config file at examples/wav2vec/configs/fintuning/base_100h.yaml you’ll see that the dataset.valid_subset field is set to dev_other which means the script is looking in your data directory (/home/robot/dataset/fairseq/libri_speech/audio/) for the files: dev_other.tsv, dev_other.wrd, and dev_other.ltr (if you’re doing character recognition) in the same way as it is looking for your training dataset files.

I don’t know how to turn validation off, so you’ll have to create a dev/validation set in the same way as you created your training set and put those dev_other files in the data directory. If your validation set is named something else like valid.tsv you can just change the value in the base_100h.yaml file for the dataset.valid_subset field.

Thank you!! @dzubke