Matej Balog, Nilesh Tripuraneni, Zoubin Ghahramani, Adrian Weller

34th International Conference on Machine Learning (ICML 2017)

[PDF] [arXiv]

This repository contains scripts to reproduce experiments appearing in this academic paper.

Requirements:

• Standard Python packages: argparse, json, matplotlib, numpy, scipy, sys
• Only for generating samples yourself for the A* sampling experiment in Figure 3a: A* sampling
• Only for generating samples yourself for the low-rank perturbation experiments in Figure 4: libDAI and the subprocess Python package

Analytically computed MSE and variance of Gumbel and Exponential trick estimators of Z (left) and ln(Z) (right).

python fig1.py

MSE of estimators of Z (left) and ln(Z) (right) stemming from Fréchet (-1/2 < α < 0), Gumbel (α = 0) and Weibull tricks (α > 0).

python fig2.py

A faster but less accurate result can be obtained by setting the repetition parameter K to a value smaller than the default 100000. For example:

python fig2.py --K 1000

Sample size M required to reach a given MSE using Gumbel and Exponential trick estimators of ln(Z), using samples from A* sampling on a Robust Bayesian Regression task.

python fig3a_plot.py

The plot is produced using 100000 samples stored in data/astar_rbr_MK100000.json. To generate samples yourself, please follow these steps:

1. Obtain astar.py, osstar.py, heaps.py and robustbayesregr.py from Chris Maddison's A* sampling implementation.
2. Put all these scripts into the same directory where you store this repository.
3. Execute python fig3a_sample.py.

MSE of ln(Z) estimators for different values of α, using M=100 samples from the approximate MAP algorithm discussed in Section 5.2, with different error bounds 𝛿.

python fig3b_plot.py

The plot is produced using sample points stored in data/bandits_normal_delta0.1_M100000.json, data/bandits_normal_delta0.01_M100000.json, and data/bandits_normal_delta0.001_M100000.json.

MSEs of U(α) as estimators of ln(Z) on 10x10 attractive (left, middle) and mixed (right) spin glass model with different coupling strengths C.

python fig4_plot.py

The plot is produced using sample points stored in the data/ subdirectory. To produce samples yourself, you can follow these steps:

1. Download and compile libDAI.
2. Put the file spin_glass.cpp into the examples/ subdirectory of your libDAI installation.
3. Compile examples/spin_glass.cpp. For example, execute the following from the libDAI installation directory on Ubuntu:

   g++ -Iinclude -Wno-deprecated -Wall -W -Wextra -fpic -O3 -g -DDAI_DEBUG  -Llib -oexamples/spin_glass examples/spin_glass.cpp -ldai -lgmpxx -lgmp
   

   On Mac the following might work:

   g++ -Iinclude -I/opt/local/include -Wno-deprecated -Wall -W -Wextra -fPIC -DMACOSX -arch x86_64 -O3 -g -DDAI_DEBUG  -Llib -L/opt/local/lib -o examples/spin_glass examples/spin_glass.cpp -ldai -lgmpxx -lgmp -arch x86_64
   

4. Update the PATH_CPP variable in libdai.py with your libDAI installation location.
5. Execute python fig4_sample.py.

@inproceedings{balog2017relatives,
  author = {Matej Balog and Nilesh Tripuraneni and Zoubin Ghahramani and Adrian Weller},
  title={Lost Relatives of the {G}umbel Trick},
  booktitle = {34th International Conference on Machine Learning (ICML)},
  year = {2017},
  month = {August}
}