Difference on epochs in network-slimming

[erichhhhho]

Hi, I found the default number of epochs in network-slimming(scratch training VGG-11) for Imagenet (which is 90 in the code) is different from the original paper, which is 60.

[Eric-mingjie]

Hi, Thanks for your interest in our code!

Yes, in the original network slimming paper, the epochs of ImageNet is 60. In this repo, we use the official Pytorch ImageNet training schedule which is 90 epochs.

[liuzhuang13]

Hi @erichhhhho I'm an author of both this paper and the Network Slimming paper.

Using 60 epochs in the original Network Slimming paper was due to the resource limit at that time, and there was a significant bug (a bug about activation functions in fc layers that was later found) in the original paper for the result on VGG-11 on ImageNet. So in this project, we fixed the bug and used 90 epochs (standard in many papers).

[erichhhhho]

@Eric-mingjie @liuzhuang13 I see. Thank you for your clarification.

[erichhhhho]

Btw, there is a bug in network-slimming cifar10 main_B.py (line 102)

if args.refine:
AttributeError: 'Namespace' object has no attribute 'refine'

The args.refine should already become args.scratch in your code, and will be redundant.

[Eric-mingjie]

Hi, @erichhhhho ! Thanks for pointing it out! I just pushed a fix.

[huangbiubiu]

Hi @erichhhhho I'm an author of both this paper and the Network Slimming paper.

Using 60 epochs in the original Network Slimming paper was due to the resource limit at that time, and there was a significant bug (a bug about activation functions in fc layers that was later found) in the original paper for the result on VGG-11 on ImageNet. So in this project, we fixed the bug and used 90 epochs (standard in many papers).

@liuzhuang13 I download the PyTorch model of scratch-E from the link of the trained model (https://github.com/Eric-mingjie/rethinking-network-pruning/tree/master/imagenet/network-slimming#models). I found that the value of epoch is 60 instead of 90 in the model dict. Is the scratch-E model trained 60 epoch? And what're the epoch settings of the other experiments, e.g., Unpruned in the Table. 4 of the paper RETHINKING THE VALUE OF NETWORK PRUNING? Thanks!

[Eric-mingjie]

The epoch is actually 90. Don't be bothered by the value of epoch. Standard ImageNet models are trained with 90 epochs.