Results on VGG16/ImageNet

[capitaso]

Hello, thanks for your work.

I modified your code to prune 3x3 conv layers using im2col in order to prune VGG16. The result is, about 40% accuracy loss from baseline before re-training, and 3.7% loss after re-training. This is much worse than your result. I used the experimental settings written in the paper as far as possible, and for some settings that could not be found in the paper, I used the default settings. I also tried some other settings (--use_new_input, --n_calibration_batches=XX, and so on), all failed.

I realized that I pruned lots of channels in Conv5-1 and -2 layers (they are not very redundant and should not be pruned, according to your previous paper about channel pruning.) Therefore, one possibility for improvement is manually excluding those layers from the pruning targets. However, I think AMC should be able to do it automatically.

Do you have any idea how to reproduce the paper's result? For the codes of pruning 3x3 conv, please see the pull request. Is the implementation correct?