Getting CUDA error when using GPU

[jliendo]

When running with CPU there is no issue...when I enable GPU then i get:

`(deep) deep/MtcnnMaskDetection git:(master) [D]% python main.py ../automl/efficientdet/video_calle.mp4
Downloading: "https://github.com/khrlimam/mtcnn-pytorch/releases/download/0.0.1/pnet-6b6ef92b.pth" to /home/javier/.cache/torch/checkpoints/pnet-6b6ef92b.pth
100%|##########################################################################################################| 28.1k/28.1k [00:00<00:00, 253kB/s]
Downloading: "https://github.com/khrlimam/mtcnn-pytorch/releases/download/0.0.1/rnet-b13c48bc.pth" to /home/javier/.cache/torch/checkpoints/rnet-b13c48bc.pth
100%|############################################################################################################| 394k/394k [00:00<00:00, 616kB/s]
Downloading: "https://github.com/khrlimam/mtcnn-pytorch/releases/download/0.0.1/onet-60cc8dd5.pth" to /home/javier/.cache/torch/checkpoints/onet-60cc8dd5.pth
100%|##########################################################################################################| 1.49M/1.49M [00:02<00:00, 559kB/s]

Traceback (most recent call last):
File "main.py", line 60, in
main(**parse_args())
File "main.py", line 49, in main
track_video(vcap)
File "main.py", line 33, in track_video
bboxes, scores = detections.fetch_faces(frame, return_landmarks=False)
File "/home/javier/repos/deep/MtcnnMaskDetection/utils/detections.py", line 24, in fetch_faces
bboxes, landmarks = detect_faces(image)
File "/home/javier/repos/deep/lib/python3.6/site-packages/torch_mtcnn/detector.py", line 58, in detect_faces
boxes = run_first_stage(image, pnet, scale=s, threshold=thresholds[0])
File "/home/javier/repos/deep/lib/python3.6/site-packages/torch_mtcnn/first_stage.py", line 34, in run_first_stage
probs = output[1].data.numpy()[0, 1, :, :]

TypeError: can't convert CUDA tensor to numpy. Use Tensor.cpu() to copy the tensor to host memory first.`

[tempdata73]

The problem is due to torch_mtcnn pip package, which does not currently support GPU. To fix this, you can go to the source file (in your case it is /home/javier/repos/deep/lib/python3.6/site-packages/torch_mtcnn/detector.py) and
add after line 26

pnet.cuda()
rnet.cuda()
onet.cuda()

replace lines 82-83 with

offsets = output[0].cpu().data.numpy()
probs = output[1].cpu().data.numpy()

Link to solution from original repo.