Preprocessing images with ResNet

Question

ili3p opened this issue 9 years ago · 3 comments

$ Torch: not enough memory: you tried to allocate 123GB.
Happens here prepro_img_residule.lua:98: in main chunk, i.e.

local feat_train=torch.FloatTensor(sz, 14, 14, 2048) --ndims)

Do you really have 123GB+ RAM? 😃

Anyway, why 14x14x2048? Shouldn't be 7x7x2048?

Answer 1 · 2016-09-07T11:02:13.000Z

Oh, is it because you scale the images to 448 instead 224?

Answer 2 · 2016-09-30T01:34:30.000Z

Actually, yes, and that's why I only trained the model once. ;)

I'm training the version 2 code now, which include CNN part. In that case you don't need that much RAM

Answer 3 · 2016-12-02T15:43:40.000Z

Can we get a link to version 2 code. I don't have 128 GB RAM. I am getting a memory error when i try to extract feature for whole MSCOCO dataset