tf_unet
tf_unet copied to clipboard
jakeret
Padding Options
Hi jakeret,
I have been using your U-Net implementation for a few months now. Thank you for sharing it out! Do you see an issue if I try to implement a 'SAME' padding for convolutions for this specific U-Net network?
Zhao
Great to hear that the package is useful to you. Any kind of contribution is highly welcomed.
In principle I don't see a problem. Ideally, the padding becomes a parameter where the default is the current behaviour.
Looking forward to your Pull Request
Isn't adding padding during segmentation might decrease the segmentation accuracy? I could not find any reference for this, but presumably, when we do zero-padding, we add some information that is not present in the original image. What do you think?
For the maxpooling the padding shouldn't have an impact. For the conv/deconv this could indeed introduce a bias but i'm not sure how big the impact would be