Allow pure generation

[vineetjohn]

• Use a single empirical style vector, and sample from the unit Gaussian to set the content vector
• Use the decoder weights to generate a sentence

[vineetjohn]

The current implementation is pretty rigid in it's requirement that there be a sentence fed to the model to transform, as opposed to simply sampling and generating from latent space.

This change might need a whole lot of re-structuring in the model class. https://github.com/vineetjohn/linguistic-style-transfer/blob/master/linguistic_style_transfer_model/models/adversarial_autoencoder.py

[vineetjohn]

During the initial stab at the implementation, I've noticed that using both negative or positive style embeddings and using a sampled content vector yields mostly positive sentence in generation mode.

[vineetjohn]

During the initial stab at the implementation, I've noticed that using both negative or positive style embeddings and using a sampled content vector yields mostly positive sentence in generation mode.

Turns out this was due to a coding error and I was overriding the conditioning style embedding during generation.

[vineetjohn]

For posterity, verified that the normal inference mode of transforming sentence labels are working fine.

[vineetjohn]

Tested style accuracy of generated sentences. It's 0.987