Ramón

Added

• ability to use only a subset of MNIST training data and get a validation set from it
• ability to shuffle dataset on each epoch
• ability to train replicates (experiment repeated n times with same training data,
  only random initialization / shuffling different)
• ability to specify other modules to use to load other datasets
  • an example is provided to load the kanji MNIST dataset
• logging of runs/experiments, with option to dump to a text file
• tests for MNIST module in datasets
• a CHANGELOG (this one)

Changed

• change argparser to use positional arguments command and config
  • before all arguments were "optional" (although the program would crash without them)
• many changes to training, in attempt to reproduce original paper + reconcile different versions
  • currently: use pdf of Gaussian for policy gradient of location network, and
    normalize both baseline, target of baseline, and advantage to decrease variance and
    to keep gradient from exploding

Fixed

• fix action network and glimpse network, did not have correct number of layers