dingxiaoh / aofp Goto Github PK

In the Algorithm 1 of the original paper, why do BP in line 11 ? These filters are not be pruned at that time.

I am very interesting in this nice work. I am eager to the PyTorch. thank your.

Hi, ShawnDing. I have an issue about the demo and could you please answer it for me?
The acc-top1 of VGG16 is 86.57% by using 'python aofp/aofp_standalone.py train' to train from scratch while the baseline in paper is 93.38% on CIFAR-10.
What's the performance did you get by using 'python aofp/aofp_standalone.py train'?

Hi ShawnDing, I've been watching this paper for a long time and it's so nice. But I still have a question about the experiment result on ImageNet in this paper. As far as I know, the FLOPs of ResNet-50 is about 4.09G. However, we can figure out the FLOPs by 1.66G/(1-56.73%), which is only 3.84G. So I'm confused on it. Could you please explain it for me?

I am very interesting in this nice work. I am eager to the PyTorch implementation and the ResNet model.