Emerging Convolutions for Generative Normalizing Flows

Abstract

Generative flows are attractive because they admit exact likelihood optimization and efficient image synthesis. Recently, Kingma & Dhariwal (2018) demonstrated with Glow that generative flows are capable of generating high quality images. We generalize the 1 x 1 convolutions proposed in Glow to invertible d x d convolutions, which are more flexible since they operate on both channel and spatial axes. We propose two methods to produce invertible convolutions that have receptive fields identical to standard convolutions: Emerging convolutions are obtained by chaining specific autoregressive convolutions, and periodic convolutions are decoupled in the frequency domain. Our experiments show that the flexibility of d x d convolutions significantly improves the performance of generative flow models on galaxy images, CIFAR10 and ImageNet.

Accepted at International Conference on Machine Learning (ICML) 2019

Cited by in corpus (31)

• Normalizing Flows: An Introduction and Review of Current Methods
• Normalizing Flows for Probabilistic Modeling and Inference
• Glow-TTS: A Generative Flow for Text-to-Speech via Monotonic Alignment Search
• Deep Generative Modelling: A Comparative Review of VAEs, GANs, Normalizing Flows, Energy-Based and Autoregressive Models
• Learning Likelihoods with Conditional Normalizing Flows
• Blow: a single-scale hyperconditioned flow for non-parallel raw-audio voice conversion
• SurVAE Flows: Surjections to Bridge the Gap between VAEs and Flows
• IDF++: Analyzing and Improving Integer Discrete Flows for Lossless Compression
• Augmented Normalizing Flows: Bridging the Gap Between Generative Flows and Latent Variable Models
• VFlow: More Expressive Generative Flows with Variational Data Augmentation
• Learning Discrete Distributions by Dequantization
• RG-Flow: A hierarchical and explainable flow model based on renormalization group and sparse prior
• Orthogonalizing Convolutional Layers with the Cayley Transform
• MintNet: Building Invertible Neural Networks with Masked Convolutions
• Integer Discrete Flows and Lossless Compression
• PolyGAN: High-Order Polynomial Generators
• MaCow: Masked Convolutional Generative Flow
• Implicit Normalizing Flows
• What if Neural Networks had SVDs?
• NanoFlow: Scalable Normalizing Flows with Sublinear Parameter Complexity
• The Convolution Exponential and Generalized Sylvester Flows
• Relational State-Space Model for Stochastic Multi-Object Systems
• Likelihood Contribution based Multi-scale Architecture for Generative Flows
• Training Invertible Linear Layers through Rank-One Perturbations
• Ordering Dimensions with Nested Dropout Normalizing Flows
• Normalizing Flows with Multi-Scale Autoregressive Priors
• One Reflection Suffice
• Self Normalizing Flows
• Invertible Attention
• PixelPyramids: Exact Inference Models from Lossless Image Pyramids
• A Linear Systems Theory of Normalizing Flows