xiaobin-rong / deepvqe Goto Github PK

Thanks for your Repo.
I just want to know if you have reproduced the DeepVQE-S model ？In the original paper, it says only have 0.59M parameters.
“. DeepVQE-S is a downscaled version of our best model. The
DeepVQE-S microphone branch has 4 blocks with 16, 40, 56,
and 24 filters, the far end branch has 8 and 24 filters, and the decoding branch has 4 blocks with 40, 32, 32, and 27 filters. Additionally, the residual block is omitted in all the encoder blocks
and in the first and last decoder blocks to save more computing”
According to this part, I implemented the DeepVQE-S model, but it still have "933.32k parameters, 328 MMacs" from get_model_complexity_info(). It is far different from the number of parameters in the article. If you try it, can you synchronize the results?
I think the difference mainly comes from Bottleneck(). I set the Bottleneck() input_size=24x11, hidden_size=32x11. FrameSize is 20ms, hop length is 10ms.

1.请问你的clip_grad_norm_value设置为多少？
2.Adamw的weight_decay设置为多少？

 DeConv = torch.nn.ConvTranspose2d(4, 4, (3,3), (1,2), padding=(0,1), dilation=(1,4), groups=4)
 SDC = StreamConvTranspose2d(4, 4, (3,3), (1,2), padding=(0,1), dilation=(1,4), groups=4)

流式的结果没问题

torch.Size([1, 4, 100, 17])
torch.Size([1, 4, 100, 17])
tensor(2.3842e-07)

但是我换成如下形式后

DeConv = torch.nn.ConvTranspose2d(4, 4, (3,3), (1,2), padding=(0,1))
SDC = StreamConvTranspose2d(4, 4, (3,3), (1,2), padding=(0,1))

结果就出现了巨大偏差，这是什么原因呢？谢谢！

torch.Size([1, 4, 100, 11])
torch.Size([1, 4, 100, 11])
tensor(2.1431)

I have a question I'd like to discuss with the author and the community.
In the AlignBlock, the paper states that the input and output should be the same, for example, 128 input and 128 output.
However, when followed by a concatenation operation, the dimension becomes 256. In the subsequent layer, if it is designed for 128, how should this be handled?
I'm not sure if my understanding is correct.

Hi, Xiaobin, thanks for implementation for 'DeepVQE'.I'v read the paper and your codes.But I have some questions:
Why the input tensor shape is (B,F,T,2)?
What does 'B','F','T' mean?
Thanks in advance.

The author's gratitude for sharing. I would like to know what the differences are between deepvqe.py and deepvqe_v1.py. Looking forward to your response.

as you compared with TF-Grid net method.. did you simulate it? i want train and test it with my dataset