d-li14 / dgconv.pytorch

PyTorch implementation of Dynamic Grouping Convolution and Groupable ConvNet with pre-trained G-ResNeXt models

dgconv.pytorch

PyTorch implementation of Dynamic Grouping Convolution and Groupable ConvNet in Differentiable Learning-to-Group Channels via Groupable Convolutional Neural Networks.

• Kronecker Product is utilized to construct the sparse matrix efficiently and regularly.
• Discrete optimization is solved with the Straight-Through Estimator trick.
• Automatically learn the number of groups in an end-to-end differentiable fashion.

ResNeXt-50 on ImageNet

DGConv is used as a drop-in replacement of depthwise separable convolution in the original ResNeXt to build G-ResNeXt-50/101 network architectures. Here are some results of their performance comparison.

Architecture	LR decay strategy	Top-1 / Top-5 Accuracy
ResNeXt-50 (32x4d)	cosine (120 epochs)	78.198 / 93.916
G-ResNeXt	cosine (120 epochs)	78.592 / 94.106

Citation

@InProceedings{Zhang_2019_ICCV,
author = {Zhang, Zhaoyang and Li, Jingyu and Shao, Wenqi and Peng, Zhanglin and Zhang, Ruimao and Wang, Xiaogang and Luo, Ping},
title = {Differentiable Learning-to-Group Channels via Groupable Convolutional Neural Networks},
booktitle = {The IEEE International Conference on Computer Vision (ICCV)},
month = {Oct},
year = {2019}
}

dgconv.pytorch's Issues

can you provide the training code?

how to train the DGConv model?

The regularizer term U is the number of the used parameters in the convolution filter.
U_regularizer = 2**(self.K + torch.sum(self.gate))
And every DGConv block has recorded these regularizer terms. And use these regularizers as complexity constraints. I was confused, how to add these terms in the final loss function. I have read some discussions about this paper. It said that the author used NAS method to search the architecture. How do you think about this?

Can you provide the training file?

Can you provide the training file? Or some details of training setting?

Can you present the Baidu Pan download links for the trained weights?

Hello, @d-li14 , the google drive you gave us is unfriendly for mainland users, we are struggling to download g_resnext50_cosine.pth , can you present the baidu yunpan download links?
Thanks!