Write out graph definition when storing weights for restoration from graph definition and Tensorboard visualization

[damro]

For debugging as well as for book keeping purposes, it would be useful to write out the complete graph definition every time weights are saved.

We can store out the graph definition using tf.train.write_graph() as a ".pbtxt" file, (pickle it) and store it in gridfs:

https://www.tensorflow.org/versions/master/api_docs/python/tf/train/write_graph

This description can be then loaded into Tensorboard or used to rebuild the model without having to keep the actual script around / to make sure the correct script is being used to restart training.

[anayebi]

I like this, it might also be good to save a .png of the structure too just for visualization purposes

[chengxuz]

See also #28