Question about using or-resnet for original --> rot experiments

[dingjiansw101]

I use the network defined in the demo.py, it is ok to reproduce the similar results for original --> rot experiments on mnist. But when I changed to or-resnet, it got very low results, the accuracy is no more than 50%(even lower than the simple OR-CNN defined in demo.py ). I do not know what's wrong with it. The following is my code.
https://paste.ubuntu.com/p/NssKJbxDqZ/
What's more, when I do original --> rot experiments on cifar-10. I use the network defined in the demo.py. And got results of 35.26 and 44.97 for without/with orn respectively.
Look forward to your reply.

[ZhouYanzhao]

Hi @dingjiansw101
The OR-CNN defined in demo.py is used in our MNIST experiments. For CIFAR10/100 experiments, please refer to this model definition.

[dingjiansw101]

On the construct of my OR-Resnet, I can get a normal accuracy if I do not do random rotation operation on the test set. But when I random rotate the test images, it got low accuracy.

Can you give an implementation for pytorch? I am not sure what's wrong with my code.

Have you tried to do original --> rot experiments on mnist for resnet or vgg?

[ZhouYanzhao]

Hi @dingjiansw101,

1. I can port the CIFAR experiment after I upgrade the pytorch implementation. But it won't be available in the near future.
2. No. The models we used in the MNIST experiment are specified in the Fig.5 of the paper.