Skip to content

No Parameters Left Behind: Sensitivity Guided Adaptive Learning Rate for Training Large Transformer Models (ICLR 2022)

License

Notifications You must be signed in to change notification settings

cliang1453/SAGE

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

4 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

SAGE: Sensitivity-guided Adaptive Learning Rate for Transformers

This repo contains our codes for the paper "No Parameters Left Behind: Sensitivity Guided Adaptive Learning Rate for Training Large Transformer Models" (ICLR 2022).


Getting Start

  1. Pull and run docker
    pytorch/pytorch:1.5.1-cuda10.1-cudnn7-devel
  2. Install requirements
    pip install -r requirements.txt

Data and Model

  1. Download data and pre-trained models
    ./download.sh
    Please refer to this link for details on the GLUE benchmark.
  2. Preprocess data
    ./experiments/glue/prepro.sh
    For the most updated data processing details, please refer to the mt-dnn repo.

Fine-tuning Pre-trained Models using SAGE

We provide an example script for fine-tuning a pre-trained BERT-base model on MNLI using Adamax-SAGE:

./scripts/train_mnli_usadamax.sh GPUID

A few notices:

  • learning_rate and beta3 are two of the most important hyper-parameters. learning_rate that works well for Adamax/AdamW-SAGE is usually 2 to 5 times larger than that works well for Adamax/AdamW, depending on the tasks. beta3 that works well for Adamax/AdamW-SAGE is usually in the range of 0.6 and 0.9, depending on the tasks.

  • To use AdamW-SAGE, set argument --optim=usadamw. The current codebase only contains the implementation of Adamax-SAGE and AdamW-SAGE. Please refer to module/bert_optim.py for details. Please refer to our paper for integrating SAGE on other optimizers.

  • To fine-tune a pre-trained RoBERTa-base model, set arguments --init_checkpoint to the model path and set --encoder_type to 2. Other supported models are listed in pretrained_models.py.

  • To fine-tune on other tasks, set arguments --train_datasets and --test_datasets to the corresponding task names.


Citation

@inproceedings{
liang2022no,
title={No Parameters Left Behind: Sensitivity Guided Adaptive Learning Rate for Training Large Transformer Models},
author={Chen Liang and Haoming Jiang and Simiao Zuo and Pengcheng He and Xiaodong Liu and Jianfeng Gao and Weizhu Chen and Tuo Zhao},
booktitle={International Conference on Learning Representations},
year={2022},
url={https://openreview.net/forum?id=cuvga_CiVND}
}

Contact Information

For help or issues related to this package, please submit a GitHub issue. For personal questions related to this paper, please contact Chen Liang ([email protected]).

About

No Parameters Left Behind: Sensitivity Guided Adaptive Learning Rate for Training Large Transformer Models (ICLR 2022)

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published