EOS, SOS characters in dataloader and decoder

[paanguin]

I have a question about decoder inputs. I think the following pre-processing adds SOS and EOS token to label y.

end2end-asr-pytorch/models/asr/transformer.py

Line 261 in a22efdd

seq_in = [torch.cat([sos, y], dim=0) for y in seq]

end2end-asr-pytorch/models/asr/transformer.py

Line 262 in a22efdd

seq_out = [torch.cat([y, eos], dim=0) for y in seq]

It seems SpectrogramDataset also contain a process for adding SOS and EOS to label y.

end2end-asr-pytorch/utils/data_loader.py

Line 137 in a22efdd

transcript = constant.SOS_CHAR + transcript_file.read().replace('\n', '').lower() + constant.EOS_CHAR

But I think SpectrogramDataset should not do this. I think the decoder currently process the label like this:
y= HELLO

seq_in: SOS, SOS, H, E, L, L, O, EOS
seq_out: SOS, H, E, L, L, O, EOS, EOS

I'll be very grateful if you confirm whether this is correct or not.

[gentaiscool]

Thank you for reporting this issue @paanguin. Actually, we are working on revamping the project with a newer version soon.

Yes, I agree we need to fix this part. There should be a single SOS and EOS tokens. We should not include SOS and EOS tokens in SpectrogramDataset. We will fix this issue in the next update.