jAM (JAX implementation of Action Matching)

Action Matching is a method for learning the time evolution of distributions from samples. That is, suppose we observe the time evolution of some random variable $X_t$ with the density $q_t$, from $t_0$ to $t_1$. Having access to uncorrelated samples of $X_t$ (which are not a part of a single trajectory) at different points in time $t\in [t_0,t_1]$, Action Matching learns a model of the dynamics by learning how to move samples in time such that they are distributed according to $q_t$ for any $t$.

The key idea is to learn such function $s_t$ (which we parameterize with a neural network) that yields the continuity equation $$\partial_t q_t = -\nabla \cdot(q_t \nabla s_t).$$ Then the propagation of samples in time can be performed simply by simulating the corresponding ODE $$\frac{dx}{dt} = \nabla s_t(x).$$ For the method description see the Tutorials section below or the paper.

Tutorials

Method	Link
Action Matching
Entropic (Stochastic) Action Matching
Unbalanced Action Matching (with reweighting)

Running the code

Run all the code using main.py with different config and mode.

• config flag takes the path to the config file.
• mode flag takes one of the following values: "train", "eval", "fid_stats". All the modes require a config file. Mind the data.uniform_dequantization flag when evaluating statistics on the dataset for FID evaluation.
• workdir is the path to the working directory for storing states.

For instance, on the clusters with slurm, you would run the code like this inside your sbatch scripts

python main.py --config configs/am/cifar/generation.py \
               --workdir $PWD/checkpoint/${SLURM_JOB_ID} \
               --mode 'train'