[ICML'21] I-BERT: Integer-only BERT Quantization

I-BERT is also available in the master branch of HuggingFace!

Visit the following links for the HuggingFace implementation.

Github Link: https://github.com/huggingface/transformers/tree/master/src/transformers/models/ibert

Model Links:

Installation & Requirements

You can find more detailed installation guides from the Fairseq repo: https://github.com/pytorch/fairseq

1. Fairseq Installation

Reference: Fairseq

  • PyTorch version >= 1.4.0
  • Python version >= 3.6
  • Currently, I-BERT only supports training on GPU
git clone https://github.com/kssteven418/I-BERT.git
pip install --editable ./

2. Download pre-trained RoBERTa models

Reference: Fairseq RoBERTa

Download pretrained RoBERTa models from the links and unzip them.

# In I-BERT (root) directory
mkdir models && cd models
wget {link}
tar -xvf roberta.{base|large}.tar.gz

3. Download GLUE datasets

Reference: Fairseq Finetuning on GLUE

First, download the data from the GLUE website. Make sure to download the dataset in I-BERT (root) directory.

# In I-BERT (root) directory
wget https://gist.githubusercontent.com/W4ngatang/60c2bdb54d156a41194446737ce03e2e/raw/17b8dd0d724281ed7c3b2aeeda662b92809aadd5/download_glue_data.py
python download_glue_data.py --data_dir glue_data --tasks all

Then, preprocess the data.

# In I-BERT (root) directory
./examples/roberta/preprocess_GLUE_tasks.sh glue_data {task_name}

task_name can be one of the following: {ALL, QQP, MNLI, QNLI, MRPC, RTE, STS-B, SST-2, CoLA} .
ALL will preprocess all the tasks.
If the command is run propely, preprocessed datasets will be stored in I-BERT/{task_name}-bin

Now, you have the models and the datasets ready, so you are ready to run I-BERT!

Task-specific Model Finetuning

Before quantizing the model, you first have to finetune the pre-trained models to a specific downstream task.
Although you can finetune the model from the original Fairseq repo, we provide ibert-base branch where you can train non-quantized models without having to install the original Fairseq.
This branch is identical to the master branch of the original Fairseq repo, except for some loggings and run scripts that are irrelevant to the functionality.
If you already have finetuned models, you can skip this part.

Run the following commands to fetch and move to the ibert-base branch:

# In I-BERT (root) directory
git fetch
git checkout -t origin/ibert-base

Then, run the script:

# In I-BERT (root) directory
# CUDA_VISIBLE_DEVICES={device} python run.py --arch {roberta_base|roberta_large} --task {task_name}
CUDA_VISIBLE_DEVICES=0 python run.py --arch roberta_base --task MRPC

Checkpoints and validation logs will be stored at ./outputs directory. You can change this output location by adding the option --output-dir OUTPUT_DIR. The exact output location will look something like: ./outputs/none/MRPC-base/wd0.1_ad0.1_d0.1_lr2e-5/1219-101427_ckpt/checkpoint_best.pt.
By default, models are trained according to the task-specific hyperparameters specified in Fairseq Finetuning on GLUE. However, you can also specify the hyperparameters with the options (use the option -h for more details).

Quantiation & Quantization-Aware-Finetuning

Now, we come back to ibert branch for quantization.

git checkout ibert

And then run the script. This will first quantize the model and do quantization-aware-finetuning with the learning rate that you specify with the option --lr {lr}.

# In I-BERT (root) directory
# CUDA_VISIBLE_DEVICES={device} python run.py --arch {roberta_base|roberta_large} --task {task_name} \
# --restore-file {ckpt_path} --lr {lr}
CUDA_VISIBLE_DEVICES=0 python run.py --arch roberta_base --task MRPC --restore-file ckpt-best.pt --lr 1e-6

NOTE: Our work is still on progress. Currently, all integer operations are executed with floating point.


I-BERT has been developed as part of the following paper. We appreciate it if you would please cite the following paper if you found the library useful for your work:

  title={I-BERT: Integer-only BERT Quantization},
  author={Kim, Sehoon and Gholami, Amir and Yao, Zhewei and Mahoney, Michael W and Keutzer, Kurt},
  journal={International Conference on Machine Learning (Accepted)},