search

zylo117 / Yet-Another-EfficientDet-Pytorch

The pytorch re-implement of the official efficientdet with SOTA performance in real time and pretrained weights.
GNU Lesser General Public License v3.0
5.2k stars 1.27k forks source link

Cls loss not decreasing #741

Open oliviervluijk opened 1 year ago

oliviervluijk commented 1 year ago

Hi everyone, I have already checked #285 & #714, but in both issues the tutorial works with their code. The shape tutorial indeed works with lr = 1e-4, but the logo tutorial doesn't work lr = 1e-4. With the logo tutorial my cls loss is all over the place (something like the picture below).

results_rip

EDIT: After 37 epochs, the loss is around ~1.1, but when running the eval.py script and look at the average precision and average recall, almost all numbers are 0.00.

Anyone knows what the problem could be?

hoxjun commented 1 year ago

I have same problem. #361 I changed learning rate and optimizer, but AP is still not good. Please help me TT