Fix problem of poor reconstruction accuracy

[vineetjohn]

Clip gradients

[vineetjohn]

Dropout embedded encoder and decoder inputs

[vineetjohn]

Add beam search decoder for inference

[vineetjohn]

Throttle learning rate EDIT: Done

[vineetjohn]

Pass encoder cell states to the decoder initial state This might bypass the disentanglement EDIT: Done

[vineetjohn]

Use temperature for variability Not applicable right now

[vineetjohn]

Switch to SGD from Adam Adam converges quicker but is prone to over-fitting EDIT: Done

[vineetjohn]

Add cost for repeated tokens? Not applicable right now

[vineetjohn]

Append hidden state to every time step Not applicable right now

[vineetjohn]

Increase RNN neurons EDIT: Done

[vineetjohn]

Reduce batch size to avoid over-generalization EDIT: Done

[vineetjohn]

Fixed. Solution was to reduce batch size.