search

NVlabs / DG-Net

:couple: Joint Discriminative and Generative Learning for Person Re-identification. CVPR'19 (Oral) :couple:
https://www.zdzheng.xyz/publication/Joint-di2019
Other
1.27k stars 230 forks source link

Change the data set to train #62

Open zbx-hue opened 3 years ago

zbx-hue commented 3 years ago

I deleted some training data of market-1501. Trained a new teacher network. Changed ID Classes. It's still going to report an error. RuntimeError: CUDA error: device-side assert triggered What else do I need to change? Thank you

layumi commented 3 years ago

Hi @zbx-hue Could you show me the whole log? Do you use fp16? If so, you may refer to https://github.com/NVlabs/DG-Net/issues/61. Do you use multiple GPU? If so, could you try to run the code on one GPU first?

Besides, have you successfully tested and trained the model on Market-1501?

zbx-hue commented 3 years ago

Hi @layumi I try to train my data set. But something went wrong. So I did the following work to see if changing the data set would work. I deleted ten ID folders of train, train_all and val. Trained a new teacher network. The same error occurred with my own data set.

I'm not use fp16. Both multi-Gpus and single-Gpus have been tried. Make the same mistake. The complete Market-1501 can be successfully trained and tested.

zbx-hue commented 3 years ago

An error screenshot An error screenshot（1）