Multiple GPU training not working on AWS instances

CompVis / adaptive-style-transfer

source code for the ECCV18 paper A Style-Aware Content Loss for Real-time HD Style Transfer

https://compvis.github.io/adaptive-style-transfer/

GNU General Public License v3.0

723 stars 139 forks source link

Multiple GPU training not working on AWS instances #20

Closed miolamio closed 5 years ago

miolamio commented 5 years ago

Hi guys, thank you for sharing this fantastic project! I'm using AWS g3 instance for training with 2 x Tesla M60 cards, looks like the training is not utilizing both cards at the same time.

Arguments: CUDA_VISIBLE_DEVICES=0,1 python main.py --model_name=model_van-gogh --phase=inference --image_size=1280 --ii_dir=./samples/ --save_dir=./output/

One of the cards is 100% utilized while the other one remains unused.

asanakoy commented 5 years ago

Hi,

It makes total sense, since the code is written for a single GPU. Setting more than one visible device with CUDA_VISIBLE_DEVICES will not affect anything.