wide_residual_nets_caffe icon indicating copy to clipboard operation
wide_residual_nets_caffe copied to clipboard

Published 20 hours ago • verified publisher icon revilokeb

.prototxt file

Open hakimkhalafi opened this issue 8 years ago • 4 comments

Hey, could you post the generated .prototxt file?

Thanks, Hakim

hakimkhalafi avatar Jul 18 '16 18:07 hakimkhalafi

@hakimkhalafi which .prototxt are you looking for? train_val.prototxt and solver.prototxt together with snapshots are available via links on the Readme.md

revilokeb avatar Jul 25 '16 21:07 revilokeb

@revilokeb
I would be interested into the train_val / solver you are using on imagenet! Plus, and advice on how big should I set the convolutions if testing on imagenet 256x256 instead of 128x128. I have done few trials starting from the paper network definition and building the train_val myself but my network does diverge.

engharat avatar Jul 28 '16 15:07 engharat

@engharat by diverging you mean that you are running out GPU RAM or that your error increases (instead of decreasing) when doing SGD?

I could upload my train_vals but unfortunately so far I could only spend a few days investigating imagenet as I got interrupted and needed to devote my attention to something else. So I could only run 4 simulations so far which did not give any stellar results. I am still planning to investigate this further, in particular impact of number of stages, number of residual blocks in each stage, number of features maps etc

I havent done any simulations on imagenet 256x256 using wrn's so far.

revilokeb avatar Jul 29 '16 21:07 revilokeb

The error does not decrease, but I have built it choosing myself the convolutions sizes. I would like to explore the same. Anyway, didn't you find suspicious the fact that in the original paper the authors didn't presented the WRN on imagenet? Maybe they didn't got stellar results too.

engharat avatar Jul 29 '16 22:07 engharat