I'm trying to use exactly the same picture for training and verification, but their reported losses are quite different, like:
This is quite weird....anybody has the same problem?
I found the problem......
"util.combine_img_prediction" seems to change the original img matrix (making the outside part totally black)
I'd suggest np.copy, rather than the matrix itself
I'm trying to use exactly the same picture for training and verification, but their reported losses are quite different, like: This is quite weird....anybody has the same problem?