Code for the paper : architectural bias in transport-based generative models - precise asymptotics (arXiv)
- Simulations.ipynb provides the code necessary to train the generative model, and produce the samples.
- Theory.ipynb implements the theoretical characterization of Corollary 2.3 on low-dimensional projections of the generated density, for the example of a trimodal Gaussian mixture target (Fig. 1).
Versions: These notebooks employ Python 3.12 , and Pytorch 2.5.