Indicate GPU_id when training in Multi-gpu mode about fb.resnet.torch

[colesbury]

The first option is the right way to do it. There's a typo in the issue (CUDDA vs CUDA); not sure if you had a typo when you ran it. How did it fail?

[hli2020]

Problem solved! I made the typo CUDDA. so silly. thanks! now turn to try imagenet! finger crossed.

[ajeetksingh]

Hi,
I did the exact step as mentioned in the first step of @hli2020 sans the typo, but somehow, it's still assigning the process to other GPUs. Below is the command which I am running:

CUDA_VISIBLE_DEVICES=2,3 th main.lua -dataset cifar10 -nGPU 2 -batchSize 128 -depth 20 -netType resnet -LR 0.005 -nEpochs 150

Assuming I have to run it on the GPUs 2 and 3. GPUs are 0-indexed.

Thanks :)

[colesbury]

@ajeetksingh, what does this print out for you:

CUDA_VISIBLE_DEVICES=2,3 th -lcutorch -e "print(cutorch.getDeviceCount())"

[ajeetksingh]

The command prints 2.

Anyways, I resolved the issue. Apparently, the problem was with mapping of numbers to GPU ids.
e.g. CUDA_VISIBLE_DEVICES=2,3 would correspond to 0th and 1st GPU not the 2nd and 3rd.

So I just changed the command to CUDA_VISIBLE_DEVICES=0,1 and it started running on the 2nd and 3rd. I still don't know, why does it happen? :(