Blockwise Parallel Decoding for Deep Autoregressive Models

[kweonwooj]

Abstract

• propose novel block-wise parallel decoding scheme to make predictions for multiple time steps in parallel and then back off to the longest prefix validated by a scoring model
• apply to exsting SoTA in machine translation and image translation
• achieves 2x speed up w/o loss in quality (MT)
• achieves 3.3x speed up with slight loss in quality (MT)

Details

Introduction

• Non-Autoregressive Decoding
  • Problem : although encoding source sentence can be parallelized via self-attention, decoding target sentence is still autoregressive and hence slow and inefficient
  • Fully Non-Autoregressive models (by Gu et al 2017) is difficult to train, and leads to a large loss of quality
  • Discrete latent variable models (by Kaiser et al 2018) does not show SoTA quality
  • Iterative refinements (by Lee et al 2018) shows impressive results, but speed up is not significant

Blockwise Parallel Decoding

• restricted to Greedy Decoding
• Algorithm
  • Predict : predict k block tokens
  • Verify : find largest k~ that is quality-equivalent to greedy decoding
  • predict k blocks in parallel with each predicted token as oracle (this step can be re-used as next predict step)
  • verify the validity of k block tokens, and accept the best k~
  • Accept : extend result upto k~

Approximate Inference

• Top-k selection : relax accept condition by allowing exact match upto top k items
• Distance-based selection : in case of image, one can use distance metric d as a criteria
• Minimum Block Size : to ensure minimum speedup, we can constrain at least l words to be accepted. Ablation study says that this leads to drop in performance. (min_block_size=1 is best)

Training

• pre-train Transformer base model with WMT14 EnDe data for 100k steps
• modify decoder part, by extending to k output layers and fine-tune for 100k steps
  • due to memory constraint, unable to use mean of k cross-entropy loss, so select one of k sub-losses uniformly as a unbiased estimate of the full loss
• Knowledge distillation for smoother training

Machine Translation (Experiments)

• Methods
  • Regular : fix the pre-trained model and train modified k output layers only
  • Distillation : fix the pre-trained model and train modified k outputs with distillation
  • Fine-Tuning : fine-tune the pre-trained model with modified k outputs
  • Both : fine-tune the pre-trained model with modified k outputs with distillation
• Result
  • Combining Distillation and Fine-Tuning leads to significant improvement in speed while maintaining quality

Wall-Clock SpeedUp

• mean accepted block size is a proxy for speed up, actual wall-clock speed up of 3x is obtained with MT

Example

• Generation Process : see step 1 outputs 10 tokens simultaneously

Overall Performance

• 3x speed up with only 1 BLEU point loss (k=4,6 seems practical)

Personal Thoughts

• wow, great paper with simple yet an effective idea!
• must implement

Link : https://arxiv.org/pdf/1811.03115.pdf Authors : Stern et al 2018