Pruning Problem of MobileNetV2 for CIFAR100 dataset

[tahsirmunna]

Hello @RudyChin ,

Thank you for your brilliant work.

I am facing some problems when I used the CIFAR100 dataset. For CIFAR10 it works well. I got the same result that you showed in the paper for CIFAR10.

But for CIFAR100 dataset, When I select 0.9 pruning away for 60 epochs even 1000 epochs, it just prunes till 80% flops where targeted flops should be 26.58M for 90% prunes. I tried in a multiple-way to change the hyper-parameters but it does not work.

Would you please, give me some suggestions that how I can solve this problem?

Here are my used hyper-parameters for your reference,

Namespace(datapath='./data', dataset='torchvision.datasets.CIFAR100', epoch=1000, name='prune_90', model='./mbnetv2c100-best.pth', batch_size=128, lr=0.01, lbda=3e-09, prune_away=0.9, constraint='flops', large_input=False, no_grow=False, pruner='FilterPrunerMBNetV2')

Thank you and eagerly waiting for your valuable reply.