Mingjie Sun
Mingjie Sun
抱歉，不太熟悉mnn转换过程，但是code 本身应该没有问题
The current code is designed to support one-shot pruning. Please take a look at the pruning code. I think for VGG you can reuse the pruning code for iterative pruning....
We don't have it yet.
channel_selection 虽然为nn.Parameter,但是训练过程中不会update,因为这个parameter只是作为feature map tensor 的index来使用的，所以backpropogation的时候不会有gradient。
which command are you running?
请问是哪个文件里的?
https://github.com/Eric-mingjie/network-slimming/blob/master/resprune.py#L153
i think you are using a large prune ratio, try smaller prune ratio e.g. 0.1, 0.2 to see if it still exists?
把幅度小的bn对应的channel剪掉，这样对accuracy影响最小。
应该是的