search

Epiphqny / VisTR

[CVPR2021 Oral] End-to-End Video Instance Segmentation with Transformers
https://arxiv.org/abs/2011.14503
Apache License 2.0
738 stars 96 forks source link

About the results and the loss convergence #46

Closed cqg1997 closed 3 years ago

cqg1997 commented 3 years ago

I run your code in 8 V100 GPUs using resnet50 as backbone. But it seems that the loss convergence is unstable in training after the first few epochs. The AP of the last epoch is 0.33, but the AP of the 12th epoch is 0.35. Could you provide the loss curve and the APs of every epoch? It would be very helpful for me as a reference. Thanks.

Epiphqny commented 3 years ago

@cqg1997 The training of transformer is unstable, it is natural. Sorry I have not saved the log and have no card to train it now.