evgenykashin/srmnet

PyTorch implementation of "SRM : A Style-based Recalibration Module for Convolutional Neural Networks"

srmnet's Issues

The performance of SRM-Net on the cifar10.

Hi, thanks for your work. When I performed an experiment on the cifar10 using srmnet-32, I only got 91.2 accuracy, which is far away from the result you post (92.9). Could you obtain this accuracy stably during your experiments?

how to use class SRMLayer?

I want to use the Class SRMLAyer with channel=128 but it gives this error:
RuntimeError: Given groups=128, weight of size [128, 1, 2], expected input[1, 64, 2] to have 128 channels, but got 64 channels instead

How can I fix it?

Do not apply weight decay on SRM parameters

From authors: "Since SRM module learns to how to utilize mean and std statistics, it is better not to impose weight decay regularization." About SRM parameters only.

could you share srm_resnet50.pth,thanks

Train ImageNet models

Add comparison of models and fill results table

ImageNet training

Add an ImageNet version of ResNets. Train them to get benchmarks results.