search

Andrew-booler / W-Net

An implementation of the W-Net: A Deep Model for Fully Unsupervised Image Segmentation
50 stars 18 forks source link

About Minibatch Loss Computation #2

Open redzhepdx opened 6 years ago

redzhepdx commented 6 years ago

I am working on Tensorflow implementation of W-Net and I have a problem with the Mini-Batch Loss computation of Soft Normalized Cut Loss. Could you explain that part?. I think you are computing NCutLoss for all images and then subtracting value from K value.

In general, you need to compute value and subtract it from K for every image. Also, you need to sum all loss values.

Andrew-booler commented 6 years ago

Sorry for missing your comment. I'm not sure how I computed the loss at that time, but I'm now computing the loss individually. You can find that my forward function will return a list in the same size as the batch.