The any_parameter_encoder from tufts-ml

any_parameter_encoder's Introduction

Any-Parameter Encoder (APE)

Public code release accompanying the AABI 2019 paper "Rapid Model Comparison by Amortizing Across Models" by Lily H. Zhang and Michael C. Hughes.

To view the code used to generate the results and figures, visit the workshop branch.

The ape branch is a work-in-progress refactor of the codebase for future maintainability. We urge viewers to use this branch to understand our work through code.

any_parameter_encoder's People

Contributors

Stargazers

Watchers

any_parameter_encoder's Issues

Standardize an accessible interface for computing per-document representations

Would like to see TimedSVI, TimedMCMC, etc (all approaches) offer a standard interface to getting samples from the learned "q" distribution

def draw_samples_from_normal_approx_posterior(docs, topics, n_samples):
	'''
	Args
	-----
	docs : 2d array (n_docs, n_vocab)
	topics : 2d array (n_topics, n_vocab)
	n_samples : int

	Returns
	--------
	h_SK : 2d array (n_samples, n_topics)
		Each row is an unconstrained real vector
		Represents draw from q(h_n | x_n, topics)
	'''


def draw_samples_from_softmax_normal_approx_posterior(docs, topics, n_samples):
	'''
	Args
	-----
	docs : 2d array (n_docs, n_vocab)
	topics : 2d array (n_topics, n_vocab)
	n_samples : int

	Returns
	-------
	pi_SK : 2d array (n_samples, n_topics)
		Each row is a vector on the simplex (non-negative, sums to one)
		Represents draw from q(h_n | x_n, topics) that is transformed via softmax
	'''

Recommend Projects