Kernel size can't be greater than actual input size

[edwardyoon]

I've downloaded pre-trained model and deep speech.pytorch-1.1. When I try to transcribe audio, it throws RuntimeError: Calculated padded input size per channel: (61 x 6). Kernel size: (21 x 11). Kernel size can't be greater than actual input size. Any advice?

I'm using CUDA-10.0, installations were all finished successfully.

edward@GPU-machine:~/deepspeech.pytorch-1.1$ python transcribe.py --model_path models/librispeech_pretrained.pth --audio_path fox_question.wav 
Traceback (most recent call last):
  File "transcribe.py", line 88, in <module>
    out = model(Variable(spect, volatile=True))
  File "/home/edward/.local/lib/python3.6/site-packages/torch/nn/modules/module.py", line 477, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/edward/deepspeech.pytorch-1.1/model.py", line 172, in forward
    x = self.conv(x)
  File "/home/edward/.local/lib/python3.6/site-packages/torch/nn/modules/module.py", line 477, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/edward/.local/lib/python3.6/site-packages/torch/nn/modules/container.py", line 91, in forward
    input = module(input)
  File "/home/edward/.local/lib/python3.6/site-packages/torch/nn/modules/module.py", line 477, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/edward/.local/lib/python3.6/site-packages/torch/nn/modules/conv.py", line 301, in forward
    self.padding, self.dilation, self.groups)
RuntimeError: Calculated padded input size per channel: (61 x 6). Kernel size: (21 x 11). Kernel size can't be greater than actual input size at /pytorch/aten/src/THNN/generic/SpatialConvolutionMM.c:48

'''

[edwardyoon]

I used this sample http://www.voiptroubleshooter.com/open_speech/american.html?fbclid=IwAR38V9crNqoPL2x0Nt-KpQysWGJMPRB3m-QxNU-NpvUYEekyoHdxqMR5iOA

[edwardyoon]

NOTE: spect.shape: torch.Size([1, 1, 161, 1])
16kHz audio file also the same. Shape is always torch.Size([1, 1, 161, 1])

[edwardyoon]

solved. I've installed PyTorch 1.1 and CUDA 10.0, there were few changes.

[boyuangong]

Hi, I met the same problem using torch 0.4. But when change to torch 1.1. There's another problem same as #373 . Do you meet this problem? I met this problem in the test phase. Thanks.

[maggieezzat]

solved. I've installed PyTorch 1.1 and CUDA 10.0, there were few changes.

Hello @edwardyoon , I have the same problem, PyTorch 1.1 and CUDA 10.0, What are the few changes you made ?

[dimitriStoidis]

Hello @maggieezzat , I have the same issue, have you found any solution?

[turlvo]

@michacal I guess you already have got a answer.
There is a answer in #373

You should add padding to the second convolution layer.
In model.py change line 137 to:

nn.Conv2d(32, 32, kernel_size=(21, 11), stride=(2, 1), padding=(0, 5))

[vpapaioannou]

Although in the documentation is said that "It is up to the user to add the proper padding", it is quite tedious to calculate for each layer the proper padding. At least in PyTorch 1.4.0, the problem actually comes from torch/nn/modules/conv.py#L341. However, learning by PyTorch 1.5, https://github.com/pytorch/pytorch/blob/0475d7b08d7cba11fc0b3ea7ccbf0046fa0b1cf7/torch/nn/functional.py there is a function _pad used at https://github.com/pytorch/pytorch/blob/master/torch/nn/modules/conv.py#L342 that can be used to solve the problem in PyTorch 1.4 as well, In PyTorch 1.4.0 this function is called pad at line 2802 of the corresponding functional.py file.

Note that PyTorch 1.4.0 is the latest stable release, at least for Anaconda( April 6 2020). Additionally, the documentation is neither clear how to pad nor the padding_mode. Code torch/nn/modules/conv.py#L341, /home/bill/anaconda3/envs/py36/lib/python3.6/site-packages/torch/nn/functional.py#L2802 sheds some light on this. Nonetheless, further clarification is needed about the available options of padding. I have opened #36089 about this.