Fast Generation for Convolutional Autoregressive Models

TL;DR

This paper introduces a caching method to accelerate generation in convolutional autoregressive models, significantly reducing computation time and enabling practical deployment.

Contribution

The paper presents a novel caching technique that speeds up generation in convolutional autoregressive models like Wavenet and PixelCNN++, achieving up to 183x faster inference.

Findings

Up to 21x speedup for Wavenet

Up to 183x speedup for PixelCNN++

Enables real-time generation in practice

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 na\"{i}ve 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.