is openspeech.modules.conv2d_subsampling forward method made a mistake while calculating the output_lengths?

[wl-junlin]

❓ Questions & Help

def forward(self, inputs: torch.Tensor, input_lengths: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
  outputs, input_lengths = super().forward(inputs, input_lengths)
  output_lengths = input_lengths >> 2
  output_lengths -= 1
  return outputs, output_lengths

I believe the line <3, 4> of above forward method is redundant and wrong
In line 2, input_lengths has already been updated to the proper output_lengths.

Details

eg:
inputs.shape ==> [batch_size, time, features] == [2, 1000, 80]
input_lengths ==> [1000, 587]

then after line 2 executed
outputs.shape ==> [2, 249, num_out_features]
input_lengths ==> [249, 146]

the intput_lengths has been updated to the proper output_lengths.
but later in line <3, 4>, output_lenghts is assigned to input_lenghts divided by 4 which I believe leads to an error.

[sooftware]

@wl-junlin Oh Thank you!
I'll check. It seems to me that there is a problem too.

[sooftware]

@hasangchun I think our code is wrong. How do you think?

[upskyy]

@sooftware I think so. I think we can delete the two lines below, but I'll test it when I have time. [link]

def forward(self, inputs: torch.Tensor, input_lengths: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
    outputs, output_lengths = super().forward(inputs, input_lengths)
    # output_lengths = input_lengths >> 2
    # output_lengths -= 1
    return outputs, output_lengths

[sooftware]

@hasangchun Great. Add test file too.
@wl-junlin thanks for letting us know.