Training Log

[megvii-basedetection]

Training Log

[yinchimaoliang]

The training log is saved while training in the tensorboard format.

[Divadi]

Ah, I understand. I had wanted to check some training details (if a learning rate scheduler is used, convergence rate, etc). Also, I was wondering why you decided to use EMA; is there a significant difference?

[yinchimaoliang]

When EMA is applied, lr stays the same in the shole training process. In the early stage of our development, EMA is able to promote NDS about 1%, when BEVDepth is stable, no ablation study is conducted.

[zhaozp15]

@yinchimaoliang I also wonder if a learning rate scheduler is used.

[Divadi]

From what I can tell during training, the learning rate is constant without warmup either.

[zhaozp15]

Thank you so much for your reply!