训练中得到0map about hop HOT 5 CLOSED

@iuhiyuh Hi, we noticed that the loss may diverge at the beginning, as the training of BEVDet is unstable. If the loss does not decrease in the first epoch, you should try restarting the training.

from hop.

@iuhiyuh Hi, we noticed that the loss may diverge at the beginning, as the training of BEVDet is unstable. If the loss does not decrease in the first epoch, you should try restarting the training.

Thank you for your reply. I would like to understand why the learning rate (lr) isn't changing as expected. According to the configuration file with the 'warmup_iters=200' setting, there should be a warm-up.

from hop.

I notice that the training starts at "epoch": 10 instead of the first epoch from your log

from hop.

I notice that the training starts at "epoch": 10 instead of the first epoch from your log

Sorry, my mistake. The learning rate is indeed as expected.

I have restarted the training, and it appears to be running normally now.

By the way, do you have plans to release the configuration and corresponding code for the 'nuscenes' version (with ViT backbone)?"

from hop.

We have no plans to release this model now.

from hop.