comparision with proxylessnas

Question

liwei46 opened this issue 4 years ago · comments

Hello,

Thanks for sharing. I did not have time to understand all your code yet. But I am wondering have you do a comparison with proxylessnas? (link above)

Regards,
liwei

Thanks.

SherwoodZheng · Answer 1 · Thu Mar 12 2020 10:33:58 GMT+0800 (China Standard Time)

We do not compare our method with proxylessNAS, since

The architecture space is different, in https://github.com/mit-han-lab/proxylessnas they only release a chain structure search space and method on imagenet. However, the search space used in our method is cell-based.
the BPE we found is on cifar10 dataset, if you want directly search on Imagenet, you may need some finetuning of the BPE-1 and BPE-2
Even with a small dataset (cifar10), founding optimal BPE had spent about 200+ GPU/days, which means that it is now infeasible on ImageNet. However, we also have an experiment about transfer learning (search on CIFAR-10 transferring to Imagenet), which shows comparable performance with much less searching cost.

liwei46 · Answer 2 · Thu Mar 12 2020 11:15:22 GMT+0800 (China Standard Time)

Great! Thanks for the reply.

regarding search space comparison,
does the search space of proxylessNAS (chain based) is much smaller compared with yours (cell based)?

Regards,

SherwoodZheng · Answer 3 · Thu Mar 12 2020 13:11:44 GMT+0800 (China Standard Time)

hi, @liwei46 architectures in cell-based search space are 2 * 8^{14}, in proxyless NAS is about 6^{20}