Optimization process to replicate the results & strange interpolation mode

[tcourat]

Hello,

I'm trying to replicate the results in your article, but I'm having trouble getting similar performance with my own R2U-Net trained from scratch. The default parameters in the code do not lead to the desired performance. What optimization process and hyper-parameters did you use to get the pre-trained weights in the Google Drive ?

Also, why is the interpolation mode in the alignment loss set to "nearest" since it is unlikely to lead to a differentiable loss?

Thank you very much.