Mostafa Badawy
Mostafa Badawy
Which Pytorch version do you use?
Because my implementation of group convolution is based on straightforward tensorflow graph operations. However, I think the authors used a fast cuDNN implementation based on caffe.
You're welcome. Also, note that the group convolution operator hasn't been implemented yet officially in Tensorflow.
@Pelups I have an update to this. I found out that the authors have counted mult-add as 'one' unit not as 'two' units. So, the number I have achieved if...
Sorry for the late reply. I've trained on TinyImageNet-200 and the loss was very small after converging. On ImageNet, I stopped the training because I was very busy doing my...