How do I fix this error? '_cuda_setDevice'

[ArchitectTaeyoon]

Hello,

I have been following the instruction up to the training stage, where I got stuck.

#!./scripts/train_cyclegan.sh
python train.py --dataroot ./datasets/maps --name maps_cyclegan --model cycle_gan

And I get the following error;

Traceback (most recent call last):
  File "train.py", line 8, in <module>
    opt = TrainOptions().parse()
  File "/Users/taeyoonkim/Documents/ML/pix2pix/pytorch-CycleGAN-and-pix2pix/options/base_options.py", line 116, in parse
    torch.cuda.set_device(opt.gpu_ids[0])
  File "/Users/taeyoonkim/anaconda2/lib/python2.7/site-packages/torch/cuda/__init__.py", line 262, in set_device
    torch._C._cuda_setDevice(device)
AttributeError: 'module' object has no attribute '_cuda_setDevice'

Seems like this is a popular question, and after looking at the previous answers, I have been trying to install different versions of Pytorch, but didn't succeed. I couldn't figure out how to run it on the CPU like the other answer suggests. (--gpu_ids 0)

Running on Mac OS High Sierra 10.13.3
Python 2.7 version.
Installed : pip, anaconda2, pytorch (python2.7, CUDA 9.2), visdom, dominate, brew, bash, wget, CUDA developer toolkit, cuDNN

I also installed python3 using brew, but switching the path became problematic, so I stuck to 2.7 versions. I have tried installing pytorch 0.4.0, although the terminal seems to be using 0.4.1
I tried downgrading to pytorch 0.3.1 and the same error appears. (#267)

[junyanz]

The current code only supports PyTorch 0.4+. See Q & A for more details.

[ArchitectTaeyoon]

yup I tried the latest versions (0.4+) first. Same error.

[junyanz]

Hmm, do you have an NVIDIA GPU on your Mac? If not, you can set --gpu_ids -1 to use CPU mode.

[ArchitectTaeyoon]

yes I have an NVIDIA GPU.
I would like to use the GPU, since it is much faster..

I tried the CPU mode command anyway, but it won't work.
Am I meant to insert --gpu_ids -1 line into .profile ?

Thank you for answering my questions! :)

[junyanz]

add --gpu_ids -1 to your python command.
I am not sure if your issue is caused by this repo. You may want to run some PyTorch official examples and see if your PyTorch is installed correctly.

[ArchitectTaeyoon]

okay thanks! I will try the examples and come back :)

[ssnl]

The PyTorch binary for osx does not support CUDA. So if you run with gpu_ids not -1, you will see such error. You can build from pytorch github source to support CUDA.

[ArchitectTaeyoon]

Oh I see! Thank you for the feedback.
I’m not so sure how I can build the support for CUDA - I will also try the same setup with python 3.7 on a linux environment using docker on a Windows laptop.

[ssnl]

Both linux and windows binaries work with CUDA. Have fun!

[ArchitectTaeyoon]

Thank you ! :)

[dsy199610]

--gpu_ids -1

where should I type this command?

[junyanz]

You add it when you call python.py...

[QasimWani]

i'm having the same issue as @ArchitectTaeyoon with python3.8.
@ArchitectTaeyoon how were u able to solve/fix your error?

[RoyceDavison]

i'm having the same issue as @ArchitectTaeyoon with python3.8.
@ArchitectTaeyoon how were u able to solve/fix your error?

An example will be like (suppose I want to apply a pre-trained model):
python ./test.py --dataroot ./datasets/monet2photo/testA/ --name monet2photo_pretrained --model test --no_dropout --gpu_ids -1
Just add --gpu_ids -1 to your python command.

[aktaseren]

Hmm, do you have an NVIDIA GPU on your Mac? If not, you can set --gpu_ids -1 to use CPU mode.

That worked for me as well. Thank you.