prepare function ignoring device, always using GPU

[nnmm]

Describe the bug

onnx_tf allocates a lot of memory on the GPU (two GPUs, actually) in the prepare function despite the device='CPU' parameter being present. This leads to out of memory crashes. It is also very slow.

To Reproduce
Download resnet50 from the model zoo and unpack into /tmp.

import onnx
from onnx_tf.backend import prepare
onnx_model = onnx.load('/tmp/resnet50/model.onnx')
prepare(onnx_model, device='CPU')

The primary GPU now has memory usage of 7705MiB and the secondary of 4573MiB.

Python, ONNX, ONNX-TF, Tensorflow version

This section can be obtained by running get_version.py from util folder.

• Python version: 3.5.2
• ONNX version: 1.3.0
• ONNX-TF version: 1.2.0
• Tensorflow version: 1.8.0

[tjingrant]

Hi, the easiest fix is to make sure onnx-tf doesn't see your GPU's. For instance, when you run the script, do this:

CUDA_VISIBLE_DEVICES= python script.py

[nnmm]

I tried that, but curiously, it didn't change anything.

…

On Wed, Nov 21, 2018, 18:23 Tian Jin ***@***.*** wrote: Hi, the easiest fix is to make sure onnx-tf doesn't see your GPU's. For instance, when you run the script, do this: CUDA_VISIBLE_DEVICES= python script.py — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#330 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AB1_xQNBvgqgUIHDYZs5i2LN-OnLeUy5ks5uxYwngaJpZM4Ys33m> .

[tjingrant]

@nnmm

I tried
CUDA_VISIBLE_DEVICES= python inference_test.py

and it works for me.

Another thing to try from this post comment tensorflow/tensorflow#2175 (comment)

Run this before running your python command (in the same shell prompt).
export CUDA_VISIBLE_DEVICES=

Are you on Windows?
I see this related issue (tensorflow/tensorflow#16284).

[nnmm]

I'm sorry, the issue was shell-related. I use fish and ran set -x CUDA_VISIBLE_DEVICES instead of set -x CUDA_VISIBLE_DEVICES ''. When I do it correctly, it does not use the GPU. The original issue of not respecting the device parameter still stands, however.

[tjingrant]

@nnmm I agree this is a nuisance.

We'll keep this issue as a low priority enhancement request since it does not hurt functionality. Sorry for your inconvenience.

[IceTDrinker]

Hello, classic random bump on an old issue, I just happen to run into this problem today.

Any news or maybe pointers to contribute a fix for this ?

[IceTDrinker]

btw workaround from : tensorflow/tensorflow#16284 (comment)

does work :

import os
os.environ["CUDA_VISIBLE_DEVICES"] = "-1"