IDPG: An Instance-Dependent Prompt Generation Method

This repository contains the code for our paper IDPG: An Instance-Dependent Prompt Generation Method

**************************** Updates ****************************

• 4/23: We released our training code.
• 4/9: We released our paper
• 4/7: Our paper has been accepted to NAACL 2022 main conference!

• PyTorch version >= 1.4.0
• Python version >= 3.6
• For training new models, you'll also need an NVIDIA GPU and NCCL
• To install IDPG and develop locally:

conda create -n IDPG python=3.6
conda activate IDPG
git clone https://github.com/CSerxy/IDPG.git
cd IDPG 
pip install --editable ./
pip install requests
pip install pytorch-metric-learning==0.9.90.dev0

# on MacOS:
# CFLAGS="-stdlib=libc++" pip install --editable ./

# on A100:
# conda install pytorch torchvision torchaudio cudatoolkit=11.1 -c pytorch -c nvidia
# pip3 install torch==1.8.1+cu111 torchvision==0.9.1+cu111 torchaudio==0.8.1 -f https://download.pytorch.org/whl/torch_stable.html

We mainly trained our model starting from below two checkpoints. We used roberta.large for main results, roberta.large.mnli for few-shot results.

We also provided a GloVe-based prompt generation method, with much fewer FLOPS but maintaining similar performance. Before running this model, you need to download GloVe embeddings first.

We provide example training scripts for both settings.

IDPG-PHM.sh, IDPG-PHM-GloVe.sh, IDPG-DNN.sh are the training scripts for full data setting.

few_IDPG-PHM.sh, few_IDPG-PHM-GloVe.sh are the training scripts for few-shot setting.

Take IDPG-PHM-GloVe.sh as an example. To run it, you need to download checkpoints and GloVe files first. Next, modify the SAVE and NEWSAVE variable, where SAVE is the path to your checkpoint folder, NEWSAVE is the path where you hope to store the intermediate checkpoints (i.e., fine-tuned checkpoints). Also, change glove_path to your local path as well.

We explain the main arguments in following:

• suffixlen: the generated prompt length in each Transformer layer.
• LRs: the tuned learning rate.
• seeds: the multiple random seeds.
• pdim: the hidden layer size in phm bottleneck

The result will be stored at the path OUT_FILE.

Below is an example of showing you how to train and evaluate IDPG-PHM-GloVE:

ARCH=roberta_large

SAVE=/PATH/TO/YOUR/CHECKPOINTS/FOLDER/
NEWSAVE=/PATH/TO/WHERE/YOU/STORE/THE/FINETUNE/CHECKPOINT/FOLDER/
glove_path=/PATH/TO/GloVe/FILES/

SAVE=/scratch/vgvinodv_root/vgvinodv1/zhuofeng/checkpoints/
NEWSAVE=/scratch/vgvinodv_root/vgvinodv1/zhuofeng/checkpoints/
glove_path=/home/zhuofeng/glove.6B.300d.txt
ROBERTA_PATH=$SAVE'roberta_large_checkpoint.pt'
#ROBERTA_PATH=$SAVE'nli_large_checkpoint.pt'

insertpositions="0"
suffixlen="5"
LRs="5e-3 1e-3 5e-4 1e-4"
seeds="1 2 3 4 5"

pdim="16"
mode="1"
mkdir -p "main_results"
OUT_FILE='main_results/IDPG-PHM-glove-layerb.txt'$pdim'-'$suffixlen


for LR in $LRs; do
    for insertposition in $insertpositions; do
        SUFFIX='-multi-phm-glove-layerb-'$mode'-'$pdim'-'$suffixlen'-'$insertposition'-f'$LR'_5-'
        TASKs='mpqa subj cr mr sst-2 qnli rte mrpc sts-b qqp'
        for TASK in $TASKs; do
            for seed in $seeds; do
                node=0
                SAVE_FILE=$NEWSAVE$TASK$SUFFIX$seed
                bash 'scripts/glove-multi-suffix-'$TASK'_finetune-phm-layerb.sh' $ROBERTA_PATH $SAVE_FILE $seed $pdim $node $suffixlen $insertposition $LR $mode $glove_path
            done
            wait
            for seed in $seeds; do
                 SAVE_FILE=$NEWSAVE$TASK$SUFFIX$seed'/'
                 CUDA_VISIBLE_DEVICES=0 python 'scripts/'$TASK'_get_result.py' -i $SAVE_FILE -o $OUT_FILE -n $seed -t $insertposition -l $LR
            done
            wait
            #SAVE_FILE=$NEWSAVE$TASK$SUFFIX
            #python 'scripts/bagging_'$TASK'.py' -i $SAVE_FILE -o $OUT_FILE
            #echo $TASK 'done'
        done
    done 
done

The codebase is based on fairseq. We tested our codes in an Nvidia A100 environment. We notice that the model's performance is sensitive to one's server environment and package version. You may find a slight performance difference if you do not have the exact same environment. We highly recommend you run hyper-parameter tuning in your own environment based on our sample scripts.

Please cite our paper if you use IDPG in your work:

@article{wu2022idpg,
  title={IDPG: An Instance-Dependent Prompt Generation Method},
  author={Wu, Zhuofeng and Wang, Sinong and Gu, Jiatao and Hou, Rui and Dong, Yuxiao and Vydiswaran, VG and Ma, Hao},
  journal={arXiv preprint arXiv:2204.04497},
  year={2022}
}