Diffusion_GMM

Code for the paper : Analysis of learning a flow-based generative model from limited sample complexity (link to paper)

Theoretical characterization

(Figs. 1, 2)

• Theory.ipynb provides a Jupyter notebook implementation of the theoretical characterization of Result III.1 for the summary statistics tracking the generative flow. Any choice of schedule functions $\alpha(t), \beta(t)$ can be specified as an input.

Simulations

(Fig. 2)

• Simulations.py implements the associated numerical experiments for the flow-based model. The DAE is trained using Pytorch. For instance, to run the training on $n=10$ samples from a target binary Gaussian mixture with centroid norm $\lVert \mu\lVert=1.5$ and cluster variance $\sigma=0.5$, run

python3 Simulations.py --n 10 --m 1.5 --s 0.5

Versions: These notebooks employ Python 3.12 .