DART

Implementation for ICLR2022 paper Differentiable Prompt Makes Pre-trained Language Models Better Few-shot Learners.

Environment

• [email protected]
• Use pip install -r requirements.txt to install dependencies.
• wandb account is required if the user wants to search for best hyper-parameter combinations.

Data source

• 16-shot GLUE dataset from LM-BFF.
• Generated data consists of 5 random splits (13/21/42/87/100) for a task, each has 16 samples.

How to run

• To run across each 5 splits in a task, use run.py:
  • In the arguments, encoder="inner" is the method proposed in the paper where verbalizers are other trainable tokens; encoder="manual" means verbalizers are selected fixed tokens; encoder="lstm" refers to the P-Tuning method.

$ python run.py -h
usage: run.py [-h] [--encoder {manual,lstm,inner,inner2}] [--task TASK]
              [--num_splits NUM_SPLITS] [--repeat REPEAT] [--load_manual]
              [--extra_mask_rate EXTRA_MASK_RATE]
              [--output_dir_suffix OUTPUT_DIR_SUFFIX]

optional arguments:
  -h, --help            show this help message and exit
  --encoder {manual,lstm,inner,inner2}
  --task TASK
  --num_splits NUM_SPLITS
  --repeat REPEAT
  --load_manual
  --extra_mask_rate EXTRA_MASK_RATE
  --output_dir_suffix OUTPUT_DIR_SUFFIX, -o OUTPUT_DIR_SUFFIX

• To train and evaluate on a single split with details recorded, use inference.py.
  • Before running, [task_name, label_list, prompt_type] should be configured in the code.
  • prompt_type="none" refers to fixed verbalizer training, while "inner" refers to the method proposed in the paper. ("inner2" is deprecated 2-stage training)
• To find optimal hyper-parameters for each task-split and reproduce our result, please use sweep.py:
  • Please refer to documentation for WandB for more details.
  • ❗NOTE: we follow LM-BFF to use the corresponding automatic search results with different data split seeds.

$ python sweep.py -h
usage: sweep.py [-h]
                [--task {SST-2,sst-5,mr,cr,mpqa,subj,trec,CoLA,MNLI,MNLI-mm,SNLI,QNLI,RTE-glue,MRPC,QQP}]
                [--encoder {none,mlp,lstm,inner,inner2}]
                [--seed_split {13,21,42,87,100} [{13,21,42,87,100} ...]]
                [--batch_size {4,8,16,24,32} [{4,8,16,24,32} ...]]
                [--sweep_id SWEEP_ID]

optional arguments:
  -h, --help            show this help message and exit
  --task {SST-2,sst-5,mr,cr,mpqa,subj,trec,CoLA,MNLI,MNLI-mm,SNLI,QNLI,RTE-glue,MRPC,QQP}
  --encoder {none,mlp,lstm,inner,inner2}
  --seed_split {13,21,42,87,100} [{13,21,42,87,100} ...]
  --batch_size {4,8,16,24,32} [{4,8,16,24,32} ...]
  --sweep_id SWEEP_ID

• To train and evaluate with more customized configurations, use cli.py.
• To analyze and visualize the results come from inference.py, use visualize.py and visualize_word_emb.py.

How to Cite

@inproceedings{
zhang2022differentiable,
title={Differentiable Prompt Makes Pre-trained Language Models Better Few-shot Learners},
author={Ningyu Zhang and Luoqiu Li and Xiang Chen and Shumin Deng and Zhen Bi and Chuanqi Tan and Fei Huang and Huajun Chen},
booktitle={International Conference on Learning Representations},
year={2022},
url={https://openreview.net/forum?id=ek9a0qIafW}
}