Fix for training in batches.

[akash13singh]

Hi

This pull requests has changes for:

1. Allowing new task training in batches
2. Global placeholder for bottleneck to allow for training in batches.
3. Global placeholder for true lables Y_true) to speed up processing and avoid parameter passing
4. Simple boolean flag (DO_SERIALIZATION) to choose between pre-computing bottlenecks and training for new task
5. Updated README file with instructions
6. Removed timestamp from .npy filenames to avoid manual setting of filenameas

Please review :)