google-deepmind / mc_gradients

Monte Carlo Gradient Estimation in Machine Learning

This is the example code for the following paper. If you use the code here please cite this paper.

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

Running the code

The code contains:

• the implementation the score function, pathwise and measure valued estimators gradient_estimators.py and their tests to ensure unbiasedness gradient_estimators_test.py.
• the implementation of control variates control_variates.py and their tests control_variates_tests.py.
• a main.py file to reproduce the Bayesian Logistic regression experiments in the paper.
• a config.py file used to configure experiments.

To run the code and install the required dependencies:

  source monte_carlo_gradients/run.sh

To run a test:

  python3 -m monte_carlo_gradients.gradient_estimators_test

Colab

You can run the code in the browser using Colab. The experiments from Section 3 can be reproduced using the following link: Intuitive Analysis of Gradient Estimators

Disclaimer

This is not an official Google product.

mc_gradients's Issues

DoubleSidedMaxwell missing from TFP compatible with TF 1.5

Currently the measure valued derivatives code cannot be run since the DoubleSidedMaxwell is not present in the TFP versions which allow TF 1.5.

Solution: upgrade to TF 2.1 and use in graph mode.

Requires both a TF dependency change and a code.

Update requirements.txt file

Hi,
I am unable to run the experiments by following the instructions in the README.

Here's what I get when I run python -m monte_carlo_gradients.gradients_estimators_test

ImportError: This version of TensorFlow Probability requires TensorFlow version >= 2.1; Detected an installation of version 1.15.0-rc2. Please upgrade TensorFlow to proceed

However, if I update to TF >= 2.1 some other parts of the code break with this version.

Could you update the requirements.txt file with dependencies that should work?

More info: Ubuntu 18.04, Python 3.7.6, no GPU