Generative Flows as a General Purpose Solution for Inverse Problems

This repository contains the official Pytorch implementation of the paper:

Generative Flows as a General Purpose Solution for Inverse Problems
José A. Chávez
https://arxiv.org/abs/2110.13285)

Abstract: Due to the success of generative flows to model data distributions, they have been explored in inverse problems. Given a pre-trained generative flow, previous work pro- posed to minimize the 2-norm of the latent variables as a regularization term. The intuition behind it was to en- sure high likelihood latent variables that produce the clos- est restoration. However, high-likelihood latent variables may generate unrealistic samples as we show in our exper- iments. We therefore propose a solver to directly produce high-likelihood reconstructions. We hypothesize that our approach could make generative flows a general purpose solver for inverse problems. Furthermore, we propose 1 × 1 coupling functions to introduce permutations in a genera- tive flow. It has the advantage that its inverse does not re- quire to be calculated in the generation process. Finally, we evaluate our method for denoising, deblurring, inpainting, and colorization. We observe a compelling improvement of our method over prior works.

System requirements:

• We recommend Linux, but Windows is supported.
• 64-bit Python 3.8 installation.
• Pytorch 1.5.0 or newer with GPU support.
• One NVIDIA GPU with at least 2GB of DRAM.

Training:

With the default configuration:

$ python train.py

Solver:

With the default configuration:

$ python solver.py

For denoising:

$ python solver.py --invprob=denoising