Fix shape mismatching in SuDORMRF's masknn

[z-wony]

As the length of input audio tensor, sometimes U-ConvBlock's tensor shape becomes odd number.
In this case, shape miss-matching is occurred in process of downsampling and upsampling.

This commit trims additional last data in upsampling sequence.

[z-wony]

Test Environment.

Test Model

• Model: SuDORMRFImprovedNet
• hparams: default settings
  • egs/librimix/SuDORMRFImprovedNet/run.sh
  • egs/librimix/SuDORMRFImprovedNet/local/conf.yml
  • masknet.upsampling_depth: 4

Test code

import sys 
sys.path.append('/home/jwkim/github/asteroid')
from asteroid.models import SuDORMRFImprovedNet
import torch

model = SuDORMRFImprovedNet.from_pretrained('best_model.pth')

mix = torch.rand(1, 32000 + 200)
ret = model.separate(mix)

Exception print

Traceback (most recent call last):
  File "/usr/lib/python3.8/runpy.py", line 192, in _run_module_as_main
    return _run_code(code, main_globals, None,
  File "/usr/lib/python3.8/runpy.py", line 85, in _run_code
    exec(code, run_globals)
  File "/home/jwkim/github/asteroid/separate.py", line 9, in <module>
    ret = model.separate(mix)
  File "/home/jwkim/github/asteroid/asteroid/models/base_models.py", line 64, in separate
    return separate.separate(self, *args, **kwargs)
  File "/home/jwkim/github/asteroid/asteroid/separate.py", line 84, in separate
    return torch_separate(model, wav, **kwargs)
  File "/home/jwkim/pTest/envs/ast/lib/python3.8/site-packages/torch/autograd/grad_mode.py", line 28, in decorate_context
    return func(*args, **kwargs)
  File "/home/jwkim/github/asteroid/asteroid/separate.py", line 105, in torch_separate
    out_wavs = separate_func(wav, **kwargs)
  File "/home/jwkim/github/asteroid/asteroid/models/base_models.py", line 89, in forward_wav
    return self(wav, *args, **kwargs)
  File "/home/jwkim/pTest/envs/ast/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1051, in _call_impl
    return forward_call(*input, **kwargs)
  File "/home/jwkim/github/asteroid/asteroid/models/base_models.py", line 219, in forward
    est_masks = self.forward_masker(tf_rep)
  File "/home/jwkim/github/asteroid/asteroid/models/base_models.py", line 248, in forward_masker
    return self.masker(tf_rep)
  File "/home/jwkim/pTest/envs/ast/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1051, in _call_impl
    return forward_call(*input, **kwargs)
  File "/home/jwkim/github/asteroid/asteroid/masknn/convolutional.py", line 745, in forward
    x = self.sm(x)
  File "/home/jwkim/pTest/envs/ast/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1051, in _call_impl
    return forward_call(*input, **kwargs)
  File "/home/jwkim/pTest/envs/ast/lib/python3.8/site-packages/torch/nn/modules/container.py", line 139, in forward
    input = module(input)
  File "/home/jwkim/pTest/envs/ast/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1051, in _call_impl
    return forward_call(*input, **kwargs)
  File "/home/jwkim/github/asteroid/asteroid/masknn/convolutional.py", line 879, in forward
    output[-1] = output[-1] + resampled_out_k
RuntimeError: The size of tensor a (403) must match the size of tensor b (404) at non-singleton dimension 2

Debugging print

• I printed shape of UConvBlock between downsampling and upsampling.
• asteroid/masknn/convolutional.py:875

        # Do the downsampling process from the previous level
        for k in range(1, self.depth):
            out_k = self.spp_dw[k](output[-1])
            output.append(out_k)

        #  DEBUG PRINT
        for d in range(self.depth):
            print(f'depth: {d}, shape: {output[d].shape}')

        # Gather them now in reverse order
        for _ in range(self.depth - 1): 
            resampled_out_k = self.upsampler(output.pop(-1))
            output[-1] = output[-1] + resampled_out_k

• print result

depth: 0, shape: torch.Size([1, 512, 1612])
depth: 1, shape: torch.Size([1, 512, 806])
depth: 2, shape: torch.Size([1, 512, 403])
depth: 3, shape: torch.Size([1, 512, 202])

Cause of exception

In below code, 'depth 3' ([1, 512, 202]) is upsampled to [1, 512, 404] (assigned to resampled_out_k).
But output[-1]'s shape is [1, 512, 403].
So, size miss matching exception is occurred.

        # Gather them now in reverse order
        for _ in range(self.depth - 1): 
            resampled_out_k = self.upsampler(output.pop(-1))
            output[-1] = output[-1] + resampled_out_k

[mpariente]

Thanks for the PR and the great detail in the explanation of the problem !

The CI failures are not due to your code, so it's merged