training loss not decrease

[ccl-private]

trafficstars

I train the landcover dataset in the image segment model you provide, but the loss does not decrease.

[ccl-private]

loss always around 0.8, even though after 115 epoch.

[ccl-private]

I tried segmenter before, trained one night on two RTX3090 cards, and the loss still did not decrease. why does this happen?

[xperzy]

@wutianyiRosun , Could you please help with this issue?

[GuoQuanhao]

I tried segmenter before, trained one night on two RTX3090 cards, and the loss still did not decrease. why does this happen?

Hello, we currently make the performance of the model reach the SOTA, But the current segmentation training process is not complete, the loss func and augmentation and other solvers are in progress. You can follow my aistudio. I add some loss function and other tricks, for Trans2seg, my loss is from 5.4 to 0.02 for total training progress(16000steps), for DPT, from 5.4 to 2.0 for 18850/1600000(steps)