Datasets

[mariru]

Summary:

• issues: makes progress for predefined toy, simulated and benchmark data #125, relevant to Simple tests of full models #174

This PR adds functionality to load and simulate data in edward.
Functions for loading data are in /edward/datasets/load_data.py.
Functions for simulating data are in /edward/datasets/simulate_data.py

Now, data for a regression can be simulated as follows:

from edward.datasets import simulate_regression_data

data = simulate_regression_data()

For additional examples see example scripts in How to verify section.
This is work in progress, I would appreciate feedback. Most importantly, documentation needs to be added.

• add docstring documentation
• modify API standalone pages (if necessary)
• modify website pages (if necessary)

Intended Effect:

Easy loading of data. No need to rewrite functions for data preprocessing, data loading and data generation.

How to Verify:

Run following examples:

bayesian_linear_regression_plot.py
bayesian_linear_regression.py
bayesian_linear_regression_test.py
bayesian_nn_analytic_kl.py
bayesian_nn.py
gp_classification.py
hierarchical_logistic_regression.py
latent_space_model.py
mixture_density_network.py
mixture_gaussian_laplace.py
mixture_gaussian_map.py
mixture_gaussian.py

Side Effects:

N/A

Documentation:

needs to be added.

Reviewer Suggestions:

@diengadji @dustinvtran @akucukelbir

[dustinvtran]

This is really cool! No specific comments yet. Can you describe how you distinguish between load and simulate, and between simulated and toy? How are bigger data sets categorized?

[akucukelbir]

nice job @mariru !

i added some "checkboxes" so that we can make sure we have fully documented this before merging.

[dustinvtran]

Closing due to https://github.com/edwardlib/observations.