Deepex

* Add get_wide_model() function
* Rename deep() to get_deep_model()

**Fix tf.nn.softmax() type error: softmax get an unexpected keyword argument 'axis' when multi-classification**: It's all because keras and tensorflow version are not match.

**More flexible category_index parameter**: Now, this parameter can be a mix type, like [A,B,c], A and B is list but c is a integer. It can achive more flexible way of split field, A and B will be split as col.1 and the rest of element will be split as col.2 with parameter c. Notice: c must be the last element, and the count of remaining element can be evenly divisible by c

Because of custom metrixs are used when model compile, however, when load exists model, compile will assert a error

Fix a activation bug

1. **Activation**: When declare a `DeepEX` object, custom activation can give to deep network
2. **FC7**: FC7 layer output can get from `get_fc7_output()` functionq