TODO.txt

- Complete and fix some issues with current dataset
  1. Needs to add dividends, lctq, loq, and capital expenditures
  2. Momentum calcs are wrong size (i.e., 6mo is actually 5mo)

- Support multi-class output

- Allow for hidden layers to have different sizes in both MLP and RNN models.
  Right now all hidden layers of a NN have a fixed size given be --num_hidden.

- Change keep_prob for dropout to a tf placeholder so that we dont need two
  seperate model objects which share weights for training and validation
  
- Allow RNN model step function to exclude the first num_unrollings
  from the loss function calculation

- Redo configs.py so that it is not such a kludge and give ability for
  user to make a command line arg mandatory

- Make training global step number pick up the last number from last save
instead of starting over at epoch #1

- (DONE: These features were added to classify_examples.py) 
  Create python script that takes predications and targets and calcs
  performance metrics on out of sample data by date (this is done by
  time_plot.R but it would be nice to have something that doesn't
  depend on installing R)
  
- (DONE) Test different size batch configurations, how to handle large 
  batch size x unrollings that produce zero steps? User "raise RuntimeError(msg) 
  to handle bad config" 

- (DONE) Test for bad configs such as training files doesn't exist

- (DONE) Report network geometry in verbose mode before training begins

- (DONE) Force secret key to be input at stdin in download_data.py

- (DONE) Make check point directory if doesn't exist.

- (DONE) Confirm that code works under tensorflow version r0.10

- (DONE) Create test suite that doesn't depend on prorprietary data.

- (DONE) Create a sample experiment that doesn't depend on proprietary data

- (DONE)  Implement non-rnn (simple multi-layer perceptron) model
    Maybe new class defined in deep_mlp_model.py