UNet version

[mavaylon1]

I'm testing to see the benefits of having the crf layer with a unet. Currently the model returns a loss that oscillates between 3&4 with the crf, while the unet (from the original paper) converges to a loss of .021. Any advice?

[Brandy0k]

I met the same problem with u

[Brandy0k]

Under this scenario, the model tends to predict the whole image as foreground.