search

sacmehta / ESPNet

ESPNet: Efficient Spatial Pyramid of Dilated Convolutions for Semantic Segmentation
https://sacmehta.github.io/ESPNet/
MIT License
538 stars 111 forks source link

Question about CDilated(nn.Module) #40

Closed userb2020 closed 5 years ago

userb2020 commented 5 years ago

Hello. Your paper is really interesting and impressing. And I have a question about your network. I changed CDilated class in Model.py to speed up the network. I apply factorization to CDilated class. To be specific I changed the code as follows

cdilated

As far as I know, I learned that factorized convolution reduce the number of parameters and reduce the inference time. But, this changed network has 87 fps and your network has 146 fps.

Can you tell me your opinion about this results?

Thank you.

sacmehta commented 5 years ago

I am confused. Could you please clarify a bit more?

Did your changed increased the speed or decreased?

userb2020 commented 5 years ago

Oh, my changed network's speed decreased.

sacmehta commented 5 years ago

This is obvious. Ideally, the change you made is inline with Gaussian kernel factorization trick and should improve the speed. However, you are creating two kernels. The cost of creating a kernel is more and hence reducing the inference time. Hope this helps.

userb2020 commented 5 years ago

Ah, now I know the reason why I got low performance on inference speed. Thank you very much!