RT-DETR on roadsign: 600+epoch, no convergence. Confusing


I am a postgraduate student, who want to write a paper by improving RT-DETR’s performance. I used RT-DETR and roadsign datasets, aiming to train a traffic-detection model and get the baseline. The problem is I use 600+ epoch but there is no sign of convergence. The result is given by picture.

My questions are:

  1. what happened? Is it a overfitting(but val_loss seems to be okay) or my dataset is too simple or other reason?
  2. what can i do? should i use regularization or change dataset or … ?

more information(if more info is needed, please tell me):
model: RT-DETR
dataset: 800+ pics, train/val/test: 8/1/1
epoch: 600+, optimizer: AdamW, bs=156, gpu=’0,1,2’, workers=24, cache=true,backbone=resnet18。
training on 3 GPUs(A6000)

1 Like

I looked into it for now

Thank you so much! That’s really helpful, I learned a lot :laughing:

1 Like