Get inf ctc loss during training about icefall HOT 4 OPEN

By the way, you can replace ctc_loss != float('inf') with

torch.isinf(ctc_loss)

from icefall.

The minimum of number frames (before subsampling) is 11.

If you are using conformer_ctc, which has a subsampling factor of 4, you will get ( (11 - 1) / 2 - 1) / 2 = 2 frames.
One of the assumptions of CTC is that the output label length is not greater than that of the input length.
If that assumption does not hold, you will get an empty lattice, and hence an inf value for the total score.

And what is the recommended minimum threshold for this value?

You can do one of the following:

• (1) Exclude inf values from ctc loss
  

  icefall/egs/librispeech/ASR/conformer_ctc/train.py

  Lines 367 to 373 in 243fb97

  	ctc_loss = k2.ctc_loss(
  	decoding_graph=decoding_graph,
  	dense_fsa_vec=dense_fsa_vec,
  	output_beam=params.beam_size,
  	reduction=params.reduction,
  	use_double_scores=params.use_double_scores,
  	)

(Note you have to first set params.reduction to none, filter inf values from ctc_loss, and then compute the sum.

• (2) Filter utterances that are too short from the manifest.

from icefall.

Thanks, I will try to simply filter short utterances and adjust some threshold for experiments.

from icefall.

My solution:

ctc_loss = k2.ctc_loss(decoding_graph=decoding_graph, 
dense_fsa_vec=dense_fsa_vec,
output_beam=params.beam_size,
reduction='none',
use_double_scores=params.use_double_scores,
    )

#filter inf when performing sum reduction with ctc_loss
ctc_loss = torch.sum(torch.where(ctc_loss != float('inf'),ctc_loss , torch.tensor(0, dtype=torch.float32).to(device)))

from icefall.