Megvii-BaseDetection / BEVDepth

Official code for BEVDepth.
MIT License
730 stars 101 forks source link

training speed very slow #123

Open xiaocongcsu opened 1 year ago

xiaocongcsu commented 1 year ago

The training speed is very slow, did anybody encounter this problem? image

And the GPU is in use: image

yinchimaoliang commented 1 year ago

You may try to raise workers_per_gpu.

clw5180 commented 1 year ago

You may try to raise workers_per_gpu.

I tried it but it's much slower...

zhangzaibin commented 1 year ago

Have you solved this problem? I encounter the same issue with the A100 GPU.

qfwysw commented 1 year ago

The processing of radar point cloud is too time-consuming. You can try the method in Bevstereo. Store the point cloud as an npy file, and then read it.

Zhangwenyao1 commented 1 year ago

The processing of radar point cloud is too time-consuming. You can try the method in Bevstereo. Store the point cloud as an npy file, and then read it.

Thanks for your advice! Can you give me a demo code?

Zhangwenyao1 commented 1 year ago

The training speed is very slow, did anybody encounter this problem? image

And the GPU is in use: image

Have you solved this problem?

xiaocongcsu commented 1 year ago

Not solved yet, I used the implementation in BEVDet.