Questions about training on custom dataset

lhoyer / DAFormer

[CVPR22] Official Implementation of DAFormer: Improving Network Architectures and Training Strategies for Domain-Adaptive Semantic Segmentation

Other

470 stars 93 forks source link

Questions about training on custom dataset #57

Closed qubewx closed 1 year ago

qubewx commented 1 year ago

When I was training, I found that the accuracy of verification set was very high at the beginning and then it began to decline. Besides, the segmentation graph generated by demo-image.py was not effective. Besides, I did not see the Lt loss function marked in the graph on the code. Thank you

lhoyer commented 1 year ago

Could you please provide further details to your issue:

Can you please upload the concerning log file?
What do you mean with "segmentation graph"? Do you mean the predicted semantic segmentation?
What do you mean with "not effective"?

You can find the backward of the target loss here: https://github.com/lhoyer/DAFormer/blob/1dc68355f6ce8c5abb58d3900b094097b8b13d1e/mmseg/models/uda/dacs.py#L312

qubewx commented 1 year ago

[ 20230301_211002.log Thanks for your reply. The above is my log file,the 4347th row shows the 8000 epoch' result,89% .But after this,the aACC always decline.Besize, Can you send me your train log?Thank you very much!

lhoyer commented 1 year ago

It seems that you only train with two classes on custom datasets. Without knowing your setup, I cannot provide insights what causes the issues, which you observe. The DAFormer checkpoints from https://github.com/lhoyer/DAFormer#checkpoints are accompanied with logs.