Accuracy of pruned vgg without pruning

[robinmiali]

Hi He,

I am quite interested in your paper and thanks a lot for sharing the code. In terms of Table 3: VGG, CIFAR10. I am trying to re-implement the experiment but it seems I could not get 80.38 top-1 accuracy for your algorithm.
The model you offered on Googledrive("VGG/vgg_pretrain/prune_precfg_epoch160_variance1") seems to be vgg16 with 'cfg': [64, 64, 'M', 128, 128, 'M', 256, 256, 256, 'M', 512, 512, 512, 'M', 512, 512, 512]. However using "sh VGG_cifar/scripts/pruning_vgg_my_method.sh", we are seemingly using the model with cfg= [32, 64, 128, 128, 256, 256, 256, 256, 256, 256, 256, 256, 256].

In terms of the "accuracy without finetuning", I am not sure I am getting it right. Is it the accuracy after one-time masking the original pretrained model (line 197 in file "VGG_cifar/pruning_cifar_vgg.py ")? If so, for the model given on GoogleDrive as mentioned above, I could only get about 51%.

Could you please tell me which type of vgg16 are we using? It would be really nice if you could inform me how to re-implement the 80.38 accuracy for pruned model without training:-)

Sincerely thank you!!!

Mia