Fast generation for convolutional autoregressive models

Prajit Ramachandran, Tom Le Paine, Pooya Khorrami, Mohammad Babaeizadeh, Shiyu Chang, Yang Zhang, Mark Allan Hasegawa-Johnson, R H Campbell, Thomas S Huang

Research output: Contribution to conferencePaper

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× and 183× speedups respectively.

Original languageEnglish (US)
StatePublished - Jan 1 2019
Event5th International Conference on Learning Representations, ICLR 2017 - Toulon, France
Duration: Apr 24 2017Apr 26 2017

Conference

Conference5th International Conference on Learning Representations, ICLR 2017
CountryFrance
CityToulon
Period4/24/174/26/17

Fingerprint

training method
performance
Nave
Model-making
Cache
Performance Art

ASJC Scopus subject areas

  • Education
  • Computer Science Applications
  • Linguistics and Language
  • Language and Linguistics

Cite this

Ramachandran, P., Le Paine, T., Khorrami, P., Babaeizadeh, M., Chang, S., Zhang, Y., ... Huang, T. S. (2019). Fast generation for convolutional autoregressive models. Paper presented at 5th International Conference on Learning Representations, ICLR 2017, Toulon, France.

Fast generation for convolutional autoregressive models. / Ramachandran, Prajit; Le Paine, Tom; Khorrami, Pooya; Babaeizadeh, Mohammad; Chang, Shiyu; Zhang, Yang; Hasegawa-Johnson, Mark Allan; Campbell, R H; Huang, Thomas S.

2019. Paper presented at 5th International Conference on Learning Representations, ICLR 2017, Toulon, France.

Research output: Contribution to conferencePaper

Ramachandran, P, Le Paine, T, Khorrami, P, Babaeizadeh, M, Chang, S, Zhang, Y, Hasegawa-Johnson, MA, Campbell, RH & Huang, TS 2019, 'Fast generation for convolutional autoregressive models' Paper presented at 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, 4/24/17 - 4/26/17, .
Ramachandran P, Le Paine T, Khorrami P, Babaeizadeh M, Chang S, Zhang Y et al. Fast generation for convolutional autoregressive models. 2019. Paper presented at 5th International Conference on Learning Representations, ICLR 2017, Toulon, France.
Ramachandran, Prajit ; Le Paine, Tom ; Khorrami, Pooya ; Babaeizadeh, Mohammad ; Chang, Shiyu ; Zhang, Yang ; Hasegawa-Johnson, Mark Allan ; Campbell, R H ; Huang, Thomas S. / Fast generation for convolutional autoregressive models. Paper presented at 5th International Conference on Learning Representations, ICLR 2017, Toulon, France.
@conference{010c8a61786c43388d34263e985c116a,
title = "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 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.",
author = "Prajit Ramachandran and {Le Paine}, Tom and Pooya Khorrami and Mohammad Babaeizadeh and Shiyu Chang and Yang Zhang and Hasegawa-Johnson, {Mark Allan} and Campbell, {R H} and Huang, {Thomas S}",
year = "2019",
month = "1",
day = "1",
language = "English (US)",
note = "5th International Conference on Learning Representations, ICLR 2017 ; Conference date: 24-04-2017 Through 26-04-2017",

}

TY - CONF

T1 - Fast generation for convolutional autoregressive models

AU - Ramachandran, Prajit

AU - Le Paine, Tom

AU - Khorrami, Pooya

AU - Babaeizadeh, Mohammad

AU - Chang, Shiyu

AU - Zhang, Yang

AU - Hasegawa-Johnson, Mark Allan

AU - Campbell, R H

AU - Huang, Thomas S

PY - 2019/1/1

Y1 - 2019/1/1

N2 - 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× and 183× speedups respectively.

AB - 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× and 183× speedups respectively.

UR - http://www.scopus.com/inward/record.url?scp=85071124810&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=85071124810&partnerID=8YFLogxK

M3 - Paper

AN - SCOPUS:85071124810

ER -