search

Vegeta2020 / SE-SSD

SE-SSD: Self-Ensembling Single-Stage Object Detector From Point Cloud, CVPR 2021.
Apache License 2.0
811 stars 128 forks source link

Loss does not converge #72

Closed Eaphan closed 2 years ago

Eaphan commented 2 years ago

I use the default config and take the CIA-SSD as pretrained model. But I still can't reproduce the performance. To be specific, the loss doesn't converge. The total loss starts from 0.5, increases to 0.6x and finnaly ends at about 0.55. Do the change meet our expectations? @Vegeta2020 Thanks for your attention.

Besides, there are some warnings in my training logs:

Warning: using Python fallback for SyncBatchNorm, possibly because apex was installed without --cuda_ext. The exception raised when attempting to import the cuda backend was: No module named 'syncbn'

I'm not sure whether the syncbn of Python fallback works.

Vegeta2020 commented 2 years ago

"syncbn" seems the issue of apex, you may have to install it well. The loss also doesn't reach the expected trend.