problem in finetune stage

[dianxin556]

No description provided.

[dianxin556]

hi, i use your code to train my own model, but in funtune stage i set train_bn=False, the training loss began to increase rapidly and then decrese very slowly. and for a period end ,the predict image is black where there is the mask.Do you have the similar situation when finetuning. Thanks a lot.

[MathiasGruber]

This happened in my training as well. I'm not sure why, but it's on my todo list to dig a bit more into the batch normalization layer of keras and figure out why this happens. If anyone already knows, please let me know :)

[EricZgw]

Have you solved this problem yet? @dianxin556 @MathiasGruber

[Squibit]

Did anyone figure this out yet? @dianxin556 @MathiasGruber @EricZgw

Edit: And is it a problem? Will it eventually go away with enough finetuning?

[burhr2]

I fine-tuned based on phase1 only. My ROI is mostly at the centre of the image having a clean white background. train-800+ Val- 100+ and test-200+ with 0.95ssim for my test set