Paper review: Score-Based Generative Modeling through Stochastic Differential Equations
Abstract: This memoir focuses on generative modeling, and more precisely score based generative modeling. While other downstream tasks such as classification or even detection have been successfully tackled with deep learning, generative modeling has proven to be more challenging. First promising results were obtained with Generative Adversarial Networks (GANs), but they are known to be hard to train (as it is a highly unstable process) and to suffer from mode collapse. More recently, diffusion models have been proposed, and have shown great results on image generation, take for instance OpenAI DALLE-2 a stunning text-to-image model.
However, diffusion models are computationally expensive, and still cannot attain the same generation speed as GANs. Similarly to diffusion models, score based generative models attempt to generate images by iteratively applying a sequence transformations to a noise sample, to go from the noise distribution (which is known) to the data distribution (which is learnt). This denoising process is done by estimating a particular gradient field (know as the score) on the perturbed data distribution, to then learn how to revert the noise corruption step. High-level wise, this process iteratively transports a sample from a particular distribution, to another one, which is an interesting theory that is already applied to other tasks.
In this study, inspired by the work and code provided by Song et al, we conducted experiments using a score based generative model trained on the FashionMNIST dataset, that uses Stochastic Differential Equations (SDE) to revert the noise corruption step. Our code is made available at https://github.com/greg2451/ score-based-generative-modeling.git and allows to easily train a score based generative model on FashionMNIST, and to play with image generation using various SDE solvers.
Loading