Official Implementations of Deep Conditional Generative Learning: Model and Error Analysis

Introduction

We introduce Conditional Föllmer Flow, an ODE based generative method for sampling from a conditional distribution. We establish that the ODE is defined over the unit time interval, and under mild conditions, the velocity field and flow map of this ODE exhibit Lipschitz properties. Furthermore, we prove that the distribution of the generated samples converges to the underlying target conditional distribution with a certain convergence rate, providing a robust theoretical foundation for our approach. Our numerical experiments showcase its effectiveness across a range of scenarios, from standard nonparametric conditional density estimation problems to more intricate challenges such as image data.

We also explore an one-step scheme to accelerate sampling process.

Preparations

• Third-party softwares for comparison on toy problems

  • NNKCDE
  • FlexCoDE

• pip install -r requirements for dependencies

Details are listed in README of sub-folders.