search

AndreKelm / RefineContourNet

Results of the paper "Object Contour and Edge Detection with RefineContourNet"
Other
21 stars 11 forks source link

How can I see the details of the net structure？ #5

Closed jjkkqq closed 3 years ago

jjkkqq commented 3 years ago

hi！ I load the net like : image but it seems that specific information about the network cannot be obtained: image （when I use dagnn.DagNN.loadobj to load ,there will be some wrongs） So, could you tell me where can get the details of your net? thank you.

AndreKelm commented 3 years ago

Hi, we have described the used structure of the network in the paper. I remember a MatconvNet script that can visualize more details. I will check it and let you know

jjkkqq commented 3 years ago

thank you for your reply Your paper describes the structure of the network, and I want to know more about the parameters in the network, such as convolution core size, stride size and so on thank you

AndreKelm commented 3 years ago

You can visualize MatConvNet architectures using visualize_matconvnet from @layumi. We used ResNet101 as a backbone, so a Residual Unit contains three conv layers with a bottleneck design. In total, we used about 81.4 million parameters with 148 conv layers. The core size is mainly 3, and the stride is 1. If the input and output have different dimensions, then the bottleneck connection path has a 1 x 1 conv layer. The whole network is strongly based on RefineNet and differs only in the loss function. Hope this helps.

CoachingJane commented 3 years ago

I encountered some problems during compilation. Could you tell me what the network detection speed (FPS) is? Thank you!