Generative Model with Dynamic Linear Flow

Generative Model with Dynamic Linear Flow |
Flow-based generative models are a family of exact log-likelihood models withtractable sampling and latent-variable inference, hence conceptually attractive formodeling complex distributions. However, flow-based models are limited by den-sity estimation performance issues as compared to state-of-the-art autoregressivemodels. Autoregressive models, which also belong to the family of likelihood-based methods, however suffer from limited parallelizability. In this paper, weproposeDynamic Linear Flow (DLF), a new family of invertible transformationswith partially autoregressive structure. Our method benefits from the efficientcomputation of flow-based methods and high density estimation performance ofautoregressive methods. We demonstrate that the proposed DLF yields state-of-the-art performance on ImageNet 32×32 and 64×64 out of all flow-based methods,and is competitive with the best autoregressive model. Additionally, our modelconverges 10 times faster than Glow (Kingma and Dhariwal, 2018). The code isavailab