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× and 183× speedups respectively.
@article{arxiv.1704.06001,
title = {Fast Generation for Convolutional Autoregressive Models},
author = {Prajit Ramachandran and Tom Le Paine and Pooya Khorrami and Mohammad Babaeizadeh and Shiyu Chang and Yang Zhang and Mark A. Hasegawa-Johnson and Roy H. Campbell and Thomas S. Huang},
journal= {arXiv preprint arXiv:1704.06001},
year = {2017}
}