LGFeb 2, 2021

AutoFreeze: Automatically Freezing Model Blocks to Accelerate Fine-tuning

arXiv:2102.01386v277 citations
Originality Highly original
AI Analysis

This work provides a significant speedup for fine-tuning large pre-trained models, which is a common practice in many machine learning domains, benefiting researchers and practitioners by reducing training time and cost.

The authors tackled the problem of slow fine-tuning of large pre-trained models like BERT, which can take many hours. They developed AutoFreeze, an adaptive system that accelerates fine-tuning while preserving accuracy. AutoFreeze achieved up to 2.55x speedup on a single GPU and up to 4.38x speedup on a 64 GPU cluster for end-to-end training time, or 5.03x cost reduction, without affecting model accuracy.

With the rapid adoption of machine learning (ML), a number of domains now use the approach of fine tuning models which were pre-trained on a large corpus of data. However, our experiments show that even fine-tuning on models like BERT can take many hours even when using modern accelerators like GPUs. While prior work proposes limiting the number of layers that are fine-tuned, e.g., freezing all layers but the last layer, we find that such static approaches lead to reduced accuracy. We propose, AutoFreeze, a system that uses an adaptive approach to choose which layers are trained and show how this can accelerate model fine-tuning while preserving accuracy. We also develop mechanisms to enable efficient caching of intermediate activations which can reduce the forward computation time when performing fine-tuning. We extend AutoFreeze to perform distributed fine-tuning and design two execution modes that minimize cost and running time respectively. Our evaluation on ten NLP tasks shows that AutoFreeze, with caching enabled, can improve fine-tuning on a single GPU by up to 2.55x. On a 64 GPU cluster, for fine-tuning on the AG's news dataset, AutoFreeze is able to achieve up to 4.38x speedup when optimizing for end-to-end training time and 5.03x reduction in total cost when optimizing for efficiency, without affecting model accuracy.

Code Implementations1 repo
Foundations

The foundational work for this paper's niche, ranked by how specifically the neighbourhood builds on it — not by global fame.

Your Notes