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

Difference between your paper and source code. #43

Closed msson closed 5 years ago

msson commented 5 years ago

Hi,

During I check your codes, I found some different part between the paper and codes.

your paper mentioned, for the decoder part, output0_cat is applied by 1x1 conv (19,C) and then concat but I can see output0_cat does not pass 1x1 conv (19,C). Look at this part in your code which is Model.py (the last part of decoder)

self.conv = CBR(19 + classes, classes, 3, 1) concat_features = self.conv(torch.cat([comb_l2_l3, output0_cat], 1))

This shows that output0_cat + comb_l2_l3 passes CBR (19 + classes, classes, 3, 1) and it means output0_cat does not pass 1x1 conv (19,C).

I would like to know why there is the difference.

Thanks!

sacmehta commented 5 years ago

See #14 and #15

msson commented 5 years ago

@sacmehta Thank you.