TASKS

In order of priority:

1) [EASY] Write the code such that modules which do not load instantaneously
are only loaded once needed. The idea is to be able to type python mymprogram.py --help
without having to wait ~20s for Theano and the various GPU libraries to load.

2) [EASY]] Rewrite the dataset code to use less memory. It is highly inefficient
right now. We should only load the images and captions and that's it. Then,
we create only the new datasets we need with a transformation based on the model
that was chosen.

3) [MEDIUM/HARD]] Write automatic checkpointing, with the ability to automatically resume a job.


4) [EASY] Write a convolutional model from scratch to finish.

5) [EASY] Write a convolutional-transpose convolution model from scratch to finish.

6) [MEDIUM] Download VGG-16 or VGG-19 and define the loss function as the difference of
the activations at some intermediate layer. Hopefully, this loss function will
be much better and reduce the L2 loss blurriness effect.