change on FCN.py

[yeLer]

In view of the inability to converge in the loss of training, I made the following changes to the FCN.py file:
1 The initial learning rate is 1e-4, and the training iteration is 1e5 times, which becomes the original learning rate of 5e4 per iteration.
2 The loss function changes from sparse_softmax_cross_entropy_with_logits() to fcn's custom loss function.

Experiments show that the modified code can finally converge and have a good effect on the verification set.