infer with multi gpus

[uefall]

Hi:

I'm trying to use infer_simple.py to generate segmentation mask for a large number of imgs
but I don't know how to set multi gpu mode

if I change

cfg.NUM_GPUS = 1 --- > cfg.NUM_GPUS = 8

the demo still use gpu0 only

if I should change

with c2_utils.NamedCudaScope(0):

or some other configs?

[ir413]

Hi @uefall, predictions for each image are computed using a single GPU. Detectron supports multi-gpu inference at the dataset level by splitting the datasets into cfg.NUM_GPUS chunks of images, each of which is handled by a single GPU (e.g. see this function). You can enable multi-gpu inference on datasets by specifying the multi-gpu-testing option in test_net and train_net tools. This option is not supported by infer and infer_simple tools at this time.

[MiZhangWhuer]

Hi @ir413, I have successfully complied caffe2 & Detectron under win7 platform, everything including python interface testing is OK except for multi-gpu inference. I tested the test_net.py on single GPU and it works, but when employing multi-gpu, it finally failed. I debug the scripts through PyCharm, and it indicates that some errors occurs on function p = subprocess.Popen in utils/subprocess.py.
After checking this function, I found that there isn't CUDA_VISIABLE_DEVICES envrionment variable in windows platform. So I comment this line " subprocess_env['CUDA_VISIBLE_DEVICES'] = str(gpu_ind)" in process_in_parallel function, but new issues occurs "AssertionError: Range subprocess failed (exit code: 1)
" (See snapshot)

How can I fix this error on windows platform? It seems that this error related to mutli-thread programming on python.

Our windows version link: https://github.com/MiZhangWhuer/Detectron-windows

Hope your reply.

[ir413]

Hi @MiZhangWhuer, it seems that the issue with multi-gpu testing is specific to Windows. Unfortunately I don't have any experience with Windows so I don't have any concrete advice. Are you sure that it is not possible to set CUDA_VISIBLE_DEVICES on Windows? (I can't find anything about it here) Have you checked the subprocess.Popen docs to see if any of its Windows specific behavior is causing this issue?

[MiZhangWhuer]

@ir413 Thanks for your response. It may be the specific behavior on windows platform. I have tried several times, but still failed.
By the way, is there any example to do only semantic segmentation using Detectron framework? I notice that someone say they can training Mask-RCNN without RPN by employing StopGradient method. Is it possible for only semantic segmentation work?

[CasonTsai]

@ir413 does Detectron supports batch>1 inference with single gpu?