radarFudan / transformers-as-statisticians

Transformers as Statisticians

This repository provides the official implementation for the paper:

Transformers as Statisticians: Provable In-Context Learning with In-Context Algorithm Selection.
Yu Bai, Fan Chen, Huan Wang, Caiming Xiong, Song Mei.
Paper: https://arxiv.org/abs/2306.04637.

Getting started

Create conda environment

conda env create -f environment.yml

Training

Train the encoder model on all tasks (enter src/ first):

chmod +x ./train_encoder_all.sh
./train_encoder_all.sh

Run train_lasso.sh for the specific task of sparse linear regression (which uses a different optimization method from the tasks in the above script). Run train_gpt_all.sh for training the decoder (gpt-like) model.

Check out the corresponding config files in conf/ for the detailed training configurations. To set up a training run with a customized config file, run python train.py --config conf/$myconf.yaml$, where $myconf.yaml$ is the customized config file.

Tracking training with wandb

populate conf/wandb.yaml with you wandb info.

Evaluation and plotting

See encoder_plotting.ipynb and gpt_plotting.ipynb for details.

Test runs

Add test_run: True into the conf to start a test run, which trains for a small number of steps with saving & wandb disabled.

Miscellanous

The code is built upon https://github.com/dtsip/in-context-learning.
More information about the code can also be found in the above repo.

If you use this code in your research, please cite our paper

@article{bai2023transformers,
  title={Transformers as Statisticians: Provable In-Context Learning with In-Context Algorithm Selection},
  author={Bai, Yu and Chen, Fan and Wang, Huan and Xiong, Caiming and Mei, Song},
  journal={arXiv preprint arXiv:2306.04637},
  year={2023}
}