"workers_per_gpu" settings do not work about mmaction2 HOT 5 CLOSED

Could you provide more info?

• Which config file are used?
• Does workers_per_gpu works for codebases other than mmaction2?
• There is a possibility that the multi-threading lib does not support AMD, but lets rule out other options first.

from mmaction2.

Could you provide more info?

• Which config file are used?
• Does workers_per_gpu works for codebases other than mmaction2?
• There is a possibility that the multi-threading lib does not support AMD, but lets rule out other options first.

• Which config file are used?
  tsn_r50_1x1x3_80e_ucf101_rgb.py

• Does workers_per_gpu works for codebases other than mmaction2?
  I am running the training sample code of tsn on mmaction2.

python tools/train.py configs/recognition/tsn/tsn_r50_1x1x3_80e_ucf101_rgb.py
--work-dir work_dirs/tsn_r50_1x1x3_80e_ucf101_rgb
--validate --seed 0 --deterministic

Thank you.

from mmaction2.

Here is our result with

python tools/train.py configs/recognition/tsn/tsn_r50_1x1x3_80e_ucf101_rgb.py --work-dir work_dirs/tsn_r50_1x1x3_80e_ucf101_rgb --validate --seed 0 --deterministic

The environment is: Intel(R) Xeon(R) Silver 4114 CPU @ 2.20GHz, GTX 1080Ti GPU.
The PID 168761-168764 are corresponding to the workers_per_gpu=4.
There is a possibility that the multi-threading lib does not support AMD. I think you can check whether workers_per_gpu works for other codebases.

from mmaction2.

Here is our result with

python tools/train.py configs/recognition/tsn/tsn_r50_1x1x3_80e_ucf101_rgb.py --work-dir work_dirs/tsn_r50_1x1x3_80e_ucf101_rgb --validate --seed 0 --deterministic

The environment is: Intel(R) Xeon(R) Silver 4114 CPU @ 2.20GHz, GTX 1080Ti GPU.
The PID 168761-168764 are corresponding to the workers_per_gpu=4.
There is a possibility that the multi-threading lib does not support AMD. I think you can check whether workers_per_gpu works for other codebases.

Thank you, I tested that pytorch's multithreading is invalid for 2990wx.

from mmaction2.

It is grad to see that we find the root of the problem and i will close this issue.

from mmaction2.