Bert Tiny Finetuned Qnli

🚀 BERT-tiny model finetuned with M-FAC

This project presents a BERT-tiny model finetuned using the state-of-the-art second-order optimizer M-FAC on the QNLI dataset, offering improved performance in text classification tasks.

🚀 Quick Start

This model is finetuned on the QNLI dataset with the state-of-the-art second-order optimizer M-FAC. For more details on M-FAC, check the NeurIPS 2021 paper: https://arxiv.org/pdf/2107.03356.pdf.

✨ Features

• Finetuned BERT-tiny model using M-FAC optimizer on QNLI dataset.
• Provides a fair comparison with the default Adam optimizer.
• Results can be reproduced with provided code and hyperparameters.

📦 Installation

To reproduce the results, you need to add the M-FAC optimizer code in https://github.com/huggingface/transformers/blob/master/examples/pytorch/text-classification/run_glue.py.

💻 Usage Examples

Basic Usage

The hyperparameters used by the M-FAC optimizer are as follows:

learning rate = 1e-4
number of gradients = 1024
dampening = 1e-6

To run the finetuning process, use the following bash script:

CUDA_VISIBLE_DEVICES=0 python run_glue.py \
  --seed 8276 \
  --model_name_or_path prajjwal1/bert-tiny \
  --task_name qnli \
  --do_train \
  --do_eval \
  --max_seq_length 128 \
  --per_device_train_batch_size 32 \
  --learning_rate 1e-4 \
  --num_train_epochs 5 \
  --output_dir out_dir/ \
  --optim MFAC \
  --optim_args '{"lr": 1e-4, "num_grads": 1024, "damp": 1e-6}'

📚 Documentation

Finetuning setup

For a fair comparison against the default Adam baseline, we finetune the model in the same framework as described here https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification and just swap the Adam optimizer with M-FAC.

Results

We share the best model out of 5 runs with the following score on the QNLI validation set:

accuracy = 81.54

The mean and standard deviation for 5 runs on the QNLI validation set are as follows:

We believe these results could be improved with modest tuning of hyperparameters: per_device_train_batch_size, learning_rate, num_train_epochs, num_grads, and damp. For the sake of fair comparison and a robust default setup, we use the same hyperparameters across all models (bert-tiny, bert-mini) and all datasets (SQuAD version 2 and GLUE).

Our code for M-FAC can be found here: https://github.com/IST-DASLab/M-FAC. A step-by-step tutorial on how to integrate and use M-FAC with any repository can be found here: https://github.com/IST-DASLab/M-FAC/tree/master/tutorials.

📄 License

BibTeX entry and citation info

@article{frantar2021m,
  title={M-FAC: Efficient Matrix-Free Approximations of Second-Order Information},
  author={Frantar, Elias and Kurtic, Eldar and Alistarh, Dan},
  journal={Advances in Neural Information Processing Systems},
  volume={35},
  year={2021}
}