Training fails with 100+ classes

[Raj-08]

Hi ,

i am using your code for semantic segmentation. when i train it for only 2 classes it works but when i increase it to a bigger number 14, 100 , 150 classes the training fails. there are some classes that are just dont get predicted. i am doing a unit test where i am trying to overfit on one image but it just doesnt happen. I have trained the same dataset with a different archtecture and it worked perfectly fine.