search

tencent-ailab / pika

a lightweight speech processing toolkit based on Pytorch and (Py)Kaldi
Apache License 2.0
338 stars 57 forks source link

eval_transducer.sh decoder error #3

Closed bliunlpr2020 closed 3 years ago

bliunlpr2020 commented 3 years ago

When I run eval_transducer.sh, got the error:

Traceback (most recent call last): File "/work0/pika//decoder/decode_transducer.py", line 264, in main() File "/work0/pika//decoder/decode_transducer.py", line 133, in main len_batch + 100) File "/work0/pika/decoder/transducer_decoder.py", line 182, in decode_batch b.advance(out[:, j], self.t_idx[:, j], x_len[j]) File "/work0/pika/decoder/beam_transducer.py", line 140, in advance self.scores -= self.lm_scorer_scale * self.lm_scores[prev_k] IndexError: tensors used as indices must be long, byte or bool tensors

And I print some outputs: beam_scores tensor([ -0.6430, -12.0813, -6.1094, ..., -8.2931, -11.2605, -11.8402], device='cuda:0', grad_fn=) torch.Size([4232]) flat_beam_scores tensor([ -0.6430, -12.0813, -6.1094, ..., -8.2931, -11.2605, -11.8402], device='cuda:0', grad_fn=) torch.Size([4232]) best_scores_id tensor([ 0, 156, 2995, 3703, 30, 718, 850, 1150], device='cuda:0') torch.Size([8]) <class 'torch.Tensor'> num_words 4232 <class 'int'> prev_k tensor([0.0000, 0.0369, 0.7077, 0.8750, 0.0071, 0.1697, 0.2009, 0.2717], device='cuda:0') <class 'torch.Tensor'>

I use the pytorch 1.7.1.

Any suggestions? Thanks!

cweng6 commented 3 years ago

It seems / or // is deprecated for integer division in pytorch 1.7.0, could you try replacing this line (https://github.com/tencent-ailab/pika/blob/23c4cddef4392bc035207187d3b5653e9a3f083e/decoder/beam_transducer.py#L125)

prev_k = best_scores_id / num_words

to,

prev_k = torch.floor_divide(best_scores_id, num_words)

should solve your problem. I am not sure if this change will affect back compatibility with the older PyTorch version. Will submit a PR once confirmed.

cweng6 commented 3 years ago

@bliunlpr2020 please let us know if your issue has been fixed or not.

bliunlpr2020 commented 3 years ago

My issue has been fixed. Thank you very much! Maybe eval_transducer.sh miss "--cmn" .

cweng6 commented 3 years ago

It seems so. As long as '--cmn' option complies with the one used in the compute_cmvn_stat.py script, you should be good. If the issue has been fixed, could you please close it. Thanks.