G-ResNeXt_GroupNet_DGConv

This repository reproduces the results of the following paper: Differentiable Learning-to-Group Channels via Groupable Convolutional Neural Networks re-implement of Group ConvNet, also be called as G-ResNext. It's from the paper, reproduction of the paper "Differentiable Learning-to-Group Channels via Groupable Convolutional Neural Networks". The architecture is the same as G-ResNeXt in table 1 of the paper. I just re-implemented the GroupNet bu using dynamic grouping convolution (DGConv) operation.

Now, i have construct the loss function which considering the complexity of group conv DGConv.

The total complexity of DGConv layers has been added to the loss function.

Guideline for train the G-ResNeXt-50, 101 on ImageNet.

• Just change the ImageNet data path, change the GPU ID for fast reproducing of the GroupConvNet.

• PyTorch 0.4.0+, 1.0 is ok.

• Pretrained weights can be downloaded from BaiduYunPan (Baidu drive).

Pre-trained weights

Baidu Wangpan: G-ResNeXt password:c0o6 .

For Chinese friends. 链接：https://pan.baidu.com/s/1X6qlvArdNsjpjnCV6zrnPA 提取码：c0o6

My reproduced results for the Group Constrained ResNeXt and unconstrained one.