Can I run you trained networks on a computer without GPU?

[loicland]

Can I run you trained networks on a computer without GPU?

[bermanmaxim]

There are ways to load GPU weights to CPU only (changing device) with torch.load, have a look at:
https://discuss.pytorch.org/t/loading-weights-for-cpu-model-while-trained-on-gpu/1032

[loicland]

Hi,

first you need to set the option --cuda to 0. But yes, the trained weights are only compatible with cuda, and you will need to retrain the model from scratch. It takes about 3 hours on my 1080. I haven't tried without CPUs, but it might take a while.

[amandaloh16]

If i already have a model.pth.tar file, I don't need a GPU computer to semantically segment new test data correct? What is the correct command line arg to do that? Is this correct? Say I have a folder of new data to label in NewClouds

CUDA_VISIBLE_DEVICES=0 python learning/main.py --cuda 0 --dataset sema3d --SEMA3D_PATH $SEMA3D_DIR --db_test_name NewClouds --db_train_name train
--epochs -1 --lr_steps '[350, 400, 450]' --test_nth_epoch 100 --model_config 'gru_10,f_8' --ptn_nfeat_stn 11
--nworkers 2 --odir "results/sema3d/trainval_best" --resume

[loicland]

Hi,

sorry a bit late to the party, but the link sent by @bermanmaxim works.

the simplest way would be to add:

if not args.cuda:
        model = model.cpu()

in the function resume

[mys007]

Hi, this should also work and allows you to load the model even if you have no cuda at all:

checkpoint = torch.load(args.resume, map_location=None if args.cuda else 'cpu')