## Clean up code base * Merge all branches into one (then into main) * Extract functions and make modular ## Training debugging * Sanity check for loss functions -> overfit on one training image ## Architecture changes * (Optional) Try a smaller model without downscaling -> compression loses too much information * Change regularized loss * log scale * remove/ make weaker: mask decay * "completely" rethink extreme cases and tune loss regularizations. * SSIM loss??? (for occlusion?) * Change Unet branches final activations (see last meeting notes)
