Generative Model with Dynamic Linear Flow
This addresses the problem of improving density estimation and efficiency in generative modeling for researchers and practitioners, representing an incremental advancement by combining flow-based and autoregressive methods.
The paper tackles the performance limitations of flow-based generative models compared to autoregressive models by proposing Dynamic Linear Flow (DLF), a new invertible transformation with partially autoregressive structure, achieving state-of-the-art performance on ImageNet 32x32 and 64x64 among flow-based methods and competitive results with autoregressive models, with a 10x faster convergence than Glow.
Flow-based generative models are a family of exact log-likelihood models with tractable sampling and latent-variable inference, hence conceptually attractive for modeling complex distributions. However, flow-based models are limited by density estimation performance issues as compared to state-of-the-art autoregressive models. Autoregressive models, which also belong to the family of likelihood-based methods, however suffer from limited parallelizability. In this paper, we propose Dynamic Linear Flow (DLF), a new family of invertible transformations with partially autoregressive structure. Our method benefits from the efficient computation of flow-based methods and high density estimation performance of autoregressive methods. We demonstrate that the proposed DLF yields state-of-theart performance on ImageNet 32x32 and 64x64 out of all flow-based methods, and is competitive with the best autoregressive model. Additionally, our model converges 10 times faster than Glow (Kingma and Dhariwal, 2018). The code is available at https://github.com/naturomics/DLF.