Open TimbusCalin opened 4 years ago
I also mention that the accuracy is very high during the training and the masks are prepared accordingly(tested with other neural network).
how many classes are you using, are you applying softmax or sigmoid to those logits?
30 classes and applying softmax not sigmoid; to the best of my knowledge sigmoid can be applied in case of overlapping classes, just like in multilabel classification.
Will test again with new testing configuration.
When applying argmax on logits, the same result is for every pixel in the image.
Does anyone know how to solve this?