Fast Generation for Convolutional Autoregressive Models

Abstract

Convolutional autoregressive models have recently demonstrated state-of-the-art performance on a number of generation tasks. While fast, parallel training methods have been crucial for their success, generation is typically implemented in a naïve fashion where redundant computations are unnecessarily repeated. This results in slow generation, making such models infeasible for production environments. In this work, we describe a method to speed up generation in convolutional autoregressive models. The key idea is to cache hidden states to avoid redundant computation. We apply our fast generation method to the Wavenet and PixelCNN++ models and achieve up to $21\times$ and $183\times$ speedups respectively.

arXiv commentAccepted at ICLR 2017 Workshop

References in corpus (6)

• WaveNet: A Generative Model for Raw Audio
• Conditional Image Generation with PixelCNN Decoders
• Pixel Recurrent Neural Networks
• A guide to convolution arithmetic for deep learning
• Action-Conditional Video Prediction using Deep Networks in Atari Games
• Neural Machine Translation in Linear Time

Cited by in corpus (7)

• Conditional Time Series Forecasting with Convolutional Neural Networks
• PixelDefend: Leveraging Generative Models to Understand and Defend against Adversarial Examples
• Deep autoregressive models for the efficient variational simulation of many-body quantum systems
• PointGrow: Autoregressively Learned Point Cloud Generation with Self-Attention
• HiLLoC: Lossless Image Compression with Hierarchical Latent Variable Models
• Planning in Dynamic Environments with Conditional Autoregressive Models
• Neural Approximation of an Auto-Regressive Process through Confidence Guided Sampling