train loss

[ZackGoing]

Total time: 18:52:17.857143, iter: 0:00:07.628324, epoch: 1:59:21.331554
[Iteration 8904] [learning rate 0.001] [Total loss 150.54] [img size 672]
level_21 total 6 objects: xy/gt 1.189, wh/gt 0.078, angle/gt 0.475, conf 28.056
level_42 total 14 objects: xy/gt 1.211, wh/gt 0.041, angle/gt 0.049, conf 42.867
level_84 total 7 objects: xy/gt 1.332, wh/gt 0.062, angle/gt 0.135, conf 40.987
Max GPU memory usage: 3.0111474990844727 GigaBytes

the conf is still high at iteration 8904,is it normal?

[duanzhiihao]

Yes it's normal since the confidence loss (i.e., self.loss4obj in the code) is the summation of conf. losses of all anchor boxes. There are thousands of anchor boxes in each image so the confidence loss is usually very large.

RAPiD/models/rapid.py

Line 114 in df9a529

self.loss4obj = nn.BCELoss(reduction='sum')

[ZackGoing]

would you like to share your train.log?Thx

[duanzhiihao]

Sorry, I don't have access to the training logs.

Did you try to visualize the predicted bounding boxes using the trained weights?

[ZackGoing]

not yet