Add Orthogonal Initializer

[therealansh]

This PR adds Orthogonal Initializer (#51).

TODO:

• Add Initializer implementation
• Add support for export and import of layer in JSON format
• Add Documentation
• Add Tests

[therealansh]

Hey @zaleslaw , How are we writing tests for initialisers that are using RandomGenerator in it. Like using a statelessRandom will generate different Operand everytime. How can we "expect" something out of randomness?

[zaleslaw]

It's a good question, it's very difficult to write good tests in ML/DL area, especially unit tests for calculations in TensorFlow due to the known "non-determinism" problem.

So, I don't know why, but "stateless"Random operand in TF has seed and it produces stable results from run to run, please have a look at the RandomNormalTest and RandomNormal internals.

So the recipe: to use seed if it's available and hope that it will be stable:(

[zaleslaw]

Hi, @therealansh did you get test data from the Keras framework?
Could you please merge with the master branch, I'll run the TC builds for it

[therealansh]

Hi, @therealansh did you get test data from the Keras framework?
Could you please merge with the master branch, I'll run the TC builds for it

Yes i took the data from tf.keras.initializers.Orthogonal