Upgrade to keras 2.0. Also, build a better infrastructure for testing models and saving states - e.g. each time a model is tested, the code to create that model, the code to create the features, should all be saved. And features, therefore, should be more modular.

[x] upgrade to keras 2.0
[x] attempt to remove all global vars that can be easily removed
[x] break out features into more sub-features with better arguments used
[x] add in features and model code snapshot to history callback
- [x] actually holding off on this for now since just the 1st level feature function definition is used --> kind of actually want this for now. However, even without improvements to the logging system, if there were a set of pretty stable feature sub-functions that were used, that that would still be good to get going for code stability, clarity.
[x] figure out if the normalization isn't working as well as it used to due to upgrading to keras 2.0 (ask josh about this)
- [x] test old keras vs new keras AUC accuracy
  looks like it's comparable - mayyybe a little bit worse / variable but looks pretty solid.
[ ] hook up tensorboard to the convent output results, customize a bit as needed
- [x] 1) read tensorboard README: https://github.com/tensorflow/tensorflow/blob/master/tensorflow/tensorboard/README.md
- [ ] 2) read tensorboard tutorials https://www.tensorflow.org/get_started/summaries_and_tensorboard
  - [x] tutorial 1
  - [x] tutorial 2
  - [x] tutorial 3
[ ] start using the big URL dataset with convnet; compare against splits_5_5
[ ] look into html URI encoding: Punycode!

hillarysanders / deep_learning