This directory contains files and directories used for the baselines presented
in our paper. We developed the code with Tensorflow version 0.10.

The following scripts are used to generate and display the images and sets in
our data.
  * image_gen.js -- contains functions to render images on an HTML5 canvas, and
    randomly generate new environments.
  * display_images.py -- loads a JSON file of structured representation (e.g.,
    train.json) to generate an HTML page with canvas elements showing each
    image. 

The following scripts are used to train and evaluate baseline models. All are
run without command-line arguments.
  * convnet_model.py -- CNN, RNN, and CNN+RNN models. Two variables within this
    script (use_image and use_text) are used to control whether the text and/or
    image is provided to the model.
  * img_ft_rnn.py -- structured image features alongside text encoded using an
    RNN.
  * mlp.py -- MLP model using structured representation + text features.

The following scripts are helper scripts to compute features:
  * mem_features.py -- used to compute the features used in the maximum entropy
    and MLP models. Combines structured representation and ngram features of
    the text. The first argument is the original JSON file of the split, and
    the second argument is the filename to write to.
  * ngrams.py -- computes 2-6grams present in the training data (filename
    provided as the first argument in the command line).
  * number_grams.py -- computes 2-6 number grams (n-grams which contain a
    number, with the number removed) present in the training data (filename
    provided as the first argument in the command line).

The following scripts are used to preprocess data:
  * rnn_features.py: computes structured image features used in the structured
    image + text RNN model.
  * simple_cnn.py: helper functions for CNN, adapted from a Tensorflow tutorial
    and originally written by Aymeric Damien.
  * util.py: contains functions such as for loading data and finding the image
    filenames.

The following directories contain precomputed features:
  * ngram_data: lists of ngrams and number-grams (2-6) computed by ngrams.py
    and number_grams.py and used by mem_features.py.