fix bug - update 'deeplab.py' #180 #181

[samscientist]

With minor tweaks, the semantic segmentation is compatible with recent TensorFlow and Keras.
The PixelLib project seems to be on a halt, but for anyone who steps here, I hope this helps.

The errors I faced were... (every job was held in semantic/deeplab.py)

• ImportError: cannot import name '...' from 'tensorflow.python.keras.layers'
  → change tensorflow.python.keras.layers to tensorflow.keras.layers
• ValueError: A KerasTensor cannot be used as input to a TensorFlow function.
  (on code shape_before = tf.shape(x))
  → add Keras Layer subclass and use it to return the shape (x is KerasTensor and shape() is TensorFlow function)
• AttributeError: 'KerasHistory' object has no attribute 'layer'
  → change tensorflow.python.keras.models to tensorflow.keras.models

Additionally, I'm on Apple Silicon with the versions below.

• Python 3.12.4
• tensorflow==2.17.0
• pixellib==0.7.1