RuntimeError at the inference time

[Charlottecuc]

Hi. I trained the DurIAN model using a Mandarin dataset and modified text_frontend.py a little bit.
However, at the inference time, I encountered the following problem:

Traceback (most recent call last):
  File "inference.py", line 56, in <module>
    test()
  File "inference.py", line 48, in test
    outputs = model.inference(inputs)
  File "/models/DurIAN-ivanvok/model/model.py", line 91, in inference
    alignments, _ = self.duration_model.inference(inputs)
  File "/models/DurIAN-ivanvok/model/duration.py", line 88, in inference
    outputs, durations = self._compute_weighted_forced_alignment(outputs[0])
  File "/models/DurIAN-ivanvok/model/duration.py", line 47, in _compute_weighted_forced_alignment
    durations = torch.bincount(alignment.argmax(dim=0))
RuntimeError: cannot perform reduction function argmax on a tensor with no elements because the operation does not have an identity

I printed the value of outputs[0].sum(dim=0):

tensor([1.7232e-08, 1.7707e-08, 2.1013e-08,  ..., 1.3151e+01, 1.5982e+01,
        1.9016e+01])

And I also printed the value of eos_idx, which is zero.

Could you tell me what is eos_idx = list((outputs[0].sum(dim=0) > 0.1).cpu().numpy()).index(False) used for, and how to solve the above problem? Thank you very much!

[ivanvovk]

@Charlottecuc how did you solved the alignment issue? As BCE or MSE? The line you pointed finds the end of sequence in alignment map.

[Charlottecuc]

Hi @ivanvovk Thank you very much for your reply. I used MFA to extract alignment information of my data. For the 'alignment issue', I'm not quite sure what are you referring to..
Besides, could you tell me what's the exact meaning of "outputs[0].sum(dim=0) > 0.1" ? Thank you very much~