ValueError: not enough values to unpack (expected 2, got 0)

[efrathason]

Hi.
i try to run the trainig of dostillation with hubert and in the part of the hubert decode (before training) i got this error:

`Traceback (most recent call last):
  File "./pruned_transducer_stateless6/hubert_decode.py", line 204, in <module>
    main()
  File "/export/c06/efrat/miniconda3/envs/secondPytorch/lib/python3.8/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context
    return func(*args, **kwargs)
  File "./pruned_transducer_stateless6/hubert_decode.py", line 187, in main
    results_dict = decode_dataset(
  File "./pruned_transducer_stateless6/hubert_decode.py", line 83, in decode_dataset
    for batch_idx, batch in enumerate(dl):
  File "/export/c06/efrat/miniconda3/envs/secondPytorch/lib/python3.8/site-packages/torch/utils/data/dataloader.py", line 530, in __next__
    data = self._next_data()
  File "/export/c06/efrat/miniconda3/envs/secondPytorch/lib/python3.8/site-packages/torch/utils/data/dataloader.py", line 1224, in _next_data
    return self._process_data(data)
  File "/export/c06/efrat/miniconda3/envs/secondPytorch/lib/python3.8/site-packages/torch/utils/data/dataloader.py", line 1250, in _process_data
    data.reraise()
  File "/export/c06/efrat/miniconda3/envs/secondPytorch/lib/python3.8/site-packages/torch/_utils.py", line 457, in reraise
    raise exception
ValueError: Caught ValueError in DataLoader worker process 0.
Original Traceback (most recent call last):
  File "/export/c06/efrat/miniconda3/envs/secondPytorch/lib/python3.8/site-packages/torch/utils/data/_utils/worker.py", line 287, in _worker_loop
    data = fetcher.fetch(index)
  File "/export/c06/efrat/miniconda3/envs/secondPytorch/lib/python3.8/site-packages/torch/utils/data/_utils/fetch.py", line 51, in fetch
    data = self.dataset[possibly_batched_index]
  File "/export/c06/efrat/miniconda3/envs/secondPytorch/lib/python3.8/site-packages/lhotse/dataset/speech_recognition.py", line 125, in __getitem__
    supervision_intervals = self.input_strategy.supervision_intervals(cuts)
  File "/export/c06/efrat/miniconda3/envs/secondPytorch/lib/python3.8/site-packages/lhotse/dataset/input_strategies.py", line 249, in supervision_intervals
    start_samples, nums_samples = zip(
ValueError: not enough values to unpack (expected 2, got 0)

i run on 1 GPU if its relevant

can someone know why it happen? thanks!!

[marcoyang1998]

Hi,

I cannot reproduce your error. Are you using the latest version of Icefall?

Also, HuBERT model requires waveform as input (not fbank). Could you please check if the path to the .flac files exists on your local machine? You can the path in data/fbank/librispeech_cuts_test-clean.jsonl.gz.

[efrathason]

MonoCut(id='2277-149896-0000-0-0', start=0, duration=6.59, channel=0, supervisions=[], features=Features(type='kaldifeat-fbank', num_frames=659, num_features=80, frame_shift=0.01, sampling_rate=16000, start=0, duration=6.59, storage_type='lilcom_chunky', storage_path='data/fbank/librispeech_feats_test-clean.lca', storage_key='0,45628,14608', recording_id='2277-149896-0000', channels=0), recording=Recording(id='2277-149896-0000', sources=[AudioSource(type='file', channels=[0], source='/export/c06/efrat/icefall/egs/librispeech/ASR/download/LibriSpeech/dev-clean/2277/149896/2277-149896-0000.flac')], sampling_rate=16000, num_samples=105440, duration=6.59, transforms=None), custom=None)

[marcoyang1998]

Does /export/c06/efrat/icefall/egs/librispeech/ASR/download/LibriSpeech/dev-clean/2277/149896/2277-149896-0000.flac exist?

[efrathason]

yes

[csukuangfj]

Should the command

./pruned_transducer_stateless6/hubert_decode.py --exp-dir $exp_dir

be changed to

./pruned_transducer_stateless6/hubert_decode.py --exp-dir $exp_dir --input-strategy AudioSamples

?

[marcoyang1998]

I don't think it's necessary. Audiosamples is set as default in hubert_xlarge.py (see

icefall/egs/librispeech/ASR/pruned_transducer_stateless6/hubert_xlarge.py

Line 95 in bd7fa22

"input_strategy": "AudioSamples",

)

@efrathason
As a workaround, you may skip this stage (you just need to download the HuBERT model and dict file).

[efrathason]

i try to change the line, it steel throw the same error..

and if i skip this stage i got another error (whe i try to wxtracted the codebook):

Traceback (most recent call last):
  File "./pruned_transducer_stateless6/extract_codebook_index.py", line 92, in <module>
    main()
  File "./pruned_transducer_stateless6/extract_codebook_index.py", line 84, in main
    extractor.train_quantizer()
  File "/export/c06/efrat/icefall/egs/librispeech/ASR/pruned_transducer_stateless6/vq_utils.py", line 198, in train_quantizer
    train, valid = quantization.read_hdf5_data(self.embedding_file_path)
  File "/export/c06/efrat/miniconda3/envs/secondPytorch/lib/python3.8/site-packages/multi_quantization/quantization.py", line 798, in read_hdf5_data
    ans = np.empty((tot_frames, dim), dtype=np.float16)
ValueError: negative dimensions are not allowed

[danpovey]

when I pip install multi_quantization, it has no file quantization.py in it, only multi_quantization.py.

[efrathason]

i update my code and update the multi_quatization package and i still has this problem.
although, i saw in the master (not in my fork) that they call to the quantization.py too,
https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/pruned_transducer_stateless6/vq_utils.py#L37

what do i miss?

[danpovey]

Most likely the archive that it is trying to read is empty. Look at the file size. Possibly something went wrong in an earlier stage, when dumping the VQ indexes.

[efrathason]

ok. i try to prepare the manifest from scratch. it look that somthing in the coprus file not good or that i need to add sonthing to delete all the cuts without any supervision
i got this error:

Traceback (most recent call last):
  File "./local/validate_manifest.py", line 94, in <module>
    main()
  File "./local/validate_manifest.py", line 85, in main
    validate_one_supervision_per_cut(c)
  File "./local/validate_manifest.py", line 54, in validate_one_supervision_per_cut
    raise ValueError(f"{c.id} has {len(c.supervisions)} supervisions")
ValueError: 2277-149896-0000-0-0 has 0 supervisions

[marcoyang1998]

Yes, you need to fix the manifest first before proceeding. In your case, it seems that the cut 2277-149896-0000-0-0 has no supervision

[csukuangfj]

I am wondering if there are any tests for this recipe.