CPU and GPU memory consumption during inference

[pkhigh]

DOUBT

While running the demo notebook for MaskRCNN model. I want to know why the script uses :

1. Only 1.4 GB GPU memory ?
2. 2.95 GB CPU memory ?
3. The notebook loads more than 20 CPU threads ?
4. Is there a way to control the total CPU threads and memory?

STEPS TO REPRODUCE

config_file = '../configs/mask_rcnn_r50_fpn_1x.py'
checkpoint_file = '../checkpoints/mask_rcnn_r50_fpn_2x_20181010-41d35c05.pth'
model = init_detector(config_file, checkpoint_file, device='cuda:0')
result = inference_detector(model, img)

[hellock]

The demo just process a single image, thus the memory should be small. Could you verify that the CPU threads are related to the jupyter notebook itself or mmdetection?

[pkhigh]

@hellock I created a python script and ran the model on 1000 images in a loop. I want to know why does the CPU memory consumption is 2.95 GB ? Yes all the threads are of mmdetection because I verfied by running it in .py script. Is there a way to control threads ?

[haikuoyao]

@hellock @pkhigh

I think I met a very similar issue.
#1571

Did you sort it out?

Thanks.

[haikuoyao]

Try torch.set_num_threads(1)

[hezhu1996]

@haikuoyao Hi, I have the same problem when training FasterRcnn. It uses a large amount of CPU and keep increasing. where do you add this "torch.set_num_threads(1)" in the code? thanks

[haikuoyao]

@TWDH I met in reference not in training.
So set in reference.py for torch

[solauky]

@hellock Hi, I have the same problem when training MaskRCNN. It uses a large amount of CPU and keep increasing. How should deal with？Thank you~