[tune] Option --no-cuda is misleading in mnist_pytorch.py example

[vfdev-5]

System information

• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Ubuntu 16.04
• Ray installed from (source or binary): binary
• Ray version: 0.6.2
• Python version: 3.5
• Exact command to reproduce:

python3 mnist_pytorch.py

Describe the problem

Argument parser's option --no-cuda which is False by default is misleading as resources_per_trial does not allocate a GPU.

ray/python/ray/tune/examples/mnist_pytorch.py

Lines 45 to 49 in eddd60e

	parser.add_argument(
	'--no-cuda',
	action='store_true',
	default=False,
	help='disables CUDA training')

ray/python/ray/tune/examples/mnist_pytorch.py

Lines 178 to 180 in eddd60e

	"resources_per_trial": {
	"cpu": 3
	},

Thanks

[richardliaw]

Nice catch! Would you be interested in pushing a fix?

[vfdev-5]

@richardliaw yes, I can send a PR with a fix of both PyTorch examples :

• mnist_pytorch_trainable.py
• mnist_pytorch.py.

Idea is to add one GPU:

"resources_per_trial": { 
     "cpu": 3,
     "gpu": 1
}

what do you think ?

PS: IMO examples are a little bit outdated as Variable is used which was deprecated since 0.4.0.
Maybe we should update this too.

[richardliaw]

I think maybe do something like "gpu": int(not args.no_cuda), since you don't want to set it like that when CUDA is disabled.

Updating deprecated things would be great; thanks!

[vfdev-5]

I think maybe do something like "gpu": int(not args.no_cuda), since you don't want to set it like that when CUDA is disabled.

@richardliaw yes, but this is handled inside the trial code:

ray/python/ray/tune/examples/mnist_pytorch.py

Line 62 in eddd60e

args.cuda = not args.no_cuda and torch.cuda.is_available()

so we can leave savely "gpu": 1.

[richardliaw]

If you leave “gpu”: 1, the example will not run if Ray does not detect a GPU. Thanks, Richard

…

On Tue, Jan 29, 2019 at 2:37 AM vfdev ***@***.***> wrote: I think maybe do something like "gpu": int(not args.no_cuda), since you don't want to set it like that when CUDA is disabled. @richardliaw <https://github.com/richardliaw> yes, but this is handled inside the trial code: https://github.com/ray-project/ray/blob/eddd60e14e95a2aadb06192bd141d06c68d5f082/python/ray/tune/examples/mnist_pytorch.py#L62 so we can leave savely "gpu": 1. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#3873 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AEUc5bFOAUQgKygCBnihLCLHdRu14Yvoks5vICRugaJpZM4aU24_> .