Description

Code for the blog post Models trained with unnormalized density functions: A need for a course correction

This is a fork of the official repository for the paper Iterated Denoising Energy Matching for Sampling from Boltzmann Densities. It is used primarily to setup MCMC experiments on the same systems IDEM was trained on. The environment is the same as the original repository.

Running MCMC

For GMM:

python mcmc.py --energy gmm --n_chains 256000 --n_steps 100000 --step_size 1.25 --run_name gmm_final --save_endpoints True --endpoint_name gmm_endpoint1.npy

For DW4:

python mcmc.py --energy dw4 --n_chains 512000 --n_steps 100000 --step_size 0.5 --run_name dw4_old_baseline --box_size 2 --max_step 1 --save_endpoints True --endpoint_name dw4_endpoint1.npy

For LJ13:

python mcmc.py --energy lj13 --n_chains 512000 --n_steps 100000 --step_size 0.025 --run_name lj13_old_baseline --box_size 2.0 --max_step 0.1 --save_endpoints True --endpoint_name lj13_endpoint5.npy

For LJ55:

python mcmc.py --energy lj55 --n_chains 12800 --n_steps 200000 --step_size 0.0075 --run_name lj55_oldbaseline --box_size 2.0 --max_step 0.01 --num_equilibriation_steps 2000 --save_endpoints True --endpoint_name lj55_endpoint1.npy

Benchmarks

The benchmark evaluation codes are present in notebooks/<system>\_benchmark.ipynb files.