Open DanielZ98 opened 8 months ago
I was just about to create an issue for the training log before I see this lol. My reason is somehow more abstract than the OP, as I'm just pretty unsure about how to determine if the classification and bbox losses are of an appropriate value, as well as dropping in a "proper" way when doing my own training with replaced backbones.
I am trying to reproduce this work, but I encountered some problems during the training process. When the training schedule reaches 4/40, approximately when the lr is near its peak, the loss increases dramatically. Could the author or anyone who has successfully reproduced it share the log file of the training process to help me identify the problem?