runtime error with pytorch 0.4.0

[xhzhao]

i came across this error with pytorch 0.4.0a0+0f198fa:

Epoch: [1][19/29] Time 9.658 (6.595) Data 0.022 (0.040) Loss 65.1772 (79.3325)
Epoch: [1][20/29] Time 10.012 (6.766) Data 0.024 (0.040) Loss 76.3577 (79.1838)
Epoch: [1][21/29] Time 10.893 (6.962) Data 0.019 (0.039) Loss 76.3393 (79.0483)
Epoch: [1][22/29] Time 11.721 (7.179) Data 0.020 (0.038) Loss 68.5335 (78.5704)
Epoch: [1][23/29] Time 13.135 (7.438) Data 0.020 (0.037) Loss 79.9024 (78.6283)
Epoch: [1][24/29] Time 13.349 (7.684) Data 0.039 (0.037) Loss 69.2027 (78.2355)
Epoch: [1][25/29] Time 13.621 (7.921) Data 0.023 (0.037) Loss 73.2235 (78.0351)
Epoch: [1][26/29] Time 14.609 (8.179) Data 0.038 (0.037) Loss 63.1697 (77.4633)
Epoch: [1][27/29] Time 17.431 (8.521) Data 0.025 (0.036) Loss 77.6363 (77.4697)
Epoch: [1][28/29] Time 26.266 (9.155) Data 0.024 (0.036) Loss 75.8170 (77.4107)
Epoch: [1][29/29] Time 5.676 (9.035) Data 0.027 (0.035) Loss 118.7269 (77.4568)
Training Summary Epoch: [1] Time taken (s): 262 Average Loss 78.835
0%| | 0/5 [00:00<?, ?it/s]train.py:344: UserWarning: volatile was removed and now has no effect. Use with torch.no_grad(): instead.
inputs = Variable(inputs, volatile=True)
check, i = 0
check, sequence[i] =
0
[torch.LongTensor of size ()]

Traceback (most recent call last):
File "train.py", line 360, in
decoded_output, _ = decoder.decode(out.data, sizes)
File "/home/zhaoxiao/test/deepspeech.pytorch-latest/decoder.py", line 199, in decode
remove_repetitions=True, return_offsets=True)
File "/home/zhaoxiao/test/deepspeech.pytorch-latest/decoder.py", line 156, in convert_to_strings
string, string_offsets = self.process_string(sequences[x], seq_len, remove_repetitions)
File "/home/zhaoxiao/test/deepspeech.pytorch-latest/decoder.py", line 172, in process_string
char = self.int_to_char[sequence[i]]
KeyError:
0
[torch.LongTensor of size ()]

it seems that the sequence[i] is not a int but a tensor.

[ryanleary]

PyTorch 0.4 is still unstable and therefore unsupported at this time. The error you're seeing is that 0.4 introduces 0-dimensional tensors (ie support for scalar values) when in previous versions the above indexing would return a native Python type. You can probably replace the sequence[i] with sequence[i].item() and it'll probably work.

No guarantees you don't run into other issues with 0.4. Lots of breaking changes.

[xhzhao]

thanks very much