Monte Carlo Gradient Estimation in Machine Learning

This is the example code for the following paper, adapted for presentation for the course Machine Learning Seminar (COMP0168) at UCL:

Shakir Mohamed, Mihaela Rosca, Michael Figurnov, Andriy Mnih Monte Carlo Gradient Estimation in Machine Learning. [JMLR].

Running the code

The code contains:

• the implementation the score function, pathwise and measure valued estimators gradient_estimators.py.
• the implementation of control variates control_variates.py.
• a main.py file, edited to compute the computational cost per iteration of the Bayesian Logistic regression experiments in the paper.
• a config.py file used to configure experiments.

To install the required dependencies:

source monte_carlo_gradients/setup.sh

A runtime analysis of the experiments from Section 3 can be found in runtime-exp_dummy-data.ipynb. To analyse the runtime for the Bayesian LR experimtents, run

python -m monte_carlo_gradients.main ${estimator} ${control-variate} ${N}

where estimator is one of ("score_function", "pathwise", "measure_valued"), control-variate is one of ("none", "moving_avg", "delta") and N is an positive integer. This saves the results in the directory ./results/N-${N}/estimator-${estimator}/cv-${control-variate}.

Note that no control-variate can be used with measure-valued gradient because the weak derivative decomposition will have to be analytically computed (which is a pain).

To save results for all methods with a particular sequence of values of N, edit the iterator in runtime-exp_bayesian-lr.sh appropriately and run

source runtime-exp_bayesian-lr.sh

The current implementation runs from N = 30 to N = 400 with a step size of 5. Finally, the results can be compared in analysis.ipynb.

Disclaimer

This code has largely been borrowed from the original work, and a few changes have been made to focus solely on the runtime computation. For any other experiments, we suggest referring to the original code base.