/P-tuning-v2

An optimized prompt tuning strategy comparable to fine-tuning across model scales and tasks.

Primary LanguagePythonApache License 2.0Apache-2.0

P-tuning v2

P-Tuning v2: Prompt Tuning Can Be Comparable to Finetuning Universally Across Scales and Tasks

An optimized prompt tuning strategy achieving comparable performance to fine-tuning on small/medium-sized models and sequence tagging challenges.

Find our previous version P-tuning v1 for knowledge probing and few-shot SuperGLUE. Your kindly starring our repo can greatly encourage us to work harder :)

P-tuning v2 leverages deep prompt tuning, which is to apply continuous prompts for every layer input of the pretrained transformer. Deep prompt tuning increases the capacity of continuous prompts and closes the gap to fine-tuning across various settings, especially for small models and hard tasks.

Thanks @rainatam's joint effort in re-organizing codes for publishing!

Reproduce Tips

Since experiments reported in our paper are all conducted on NVIDIA DGX-A100 servers (which might be difficult to acquire), we reimplement P-tuning v2's results on BERT-large/RoBERTa-large with:

  • Ubuntu servers with NVIDIA GeForce RTX 3090 (24G) GPUs
  • cuda 11.1
  • packages with certain versions (provided below)

We notice that the best hyper-parameters can be sensitive to your server environment and package version. If you do not have the exact same environment, we highly recommend you to run hyper-parameter search in your environment based on our example hyper-parameter search script in search_script and result collection scripts search.py.

Setup

We conduct our experiment with Anaconda3. If you have installed Anaconda3, then create the environment for P-tuning v2:

conda create -n pt2 python=3.8.5
conda activate pt2

After we setup basic conda environment, install pytorch related packages via:

conda install -n pt2 pytorch==1.7.1 torchvision==0.8.2 torchaudio==0.7.2 cudatoolkit=11.0 -c pytorch

Finally, install other python packages we need:

pip install -r requirements.txt

Data

For SuperGLUE and SQuAD datasets, we download them from the Huggingface Datasets APIs (embedded in our codes).

For sequence tagging (NER, SRL) datasets, we prepare a non-official packup here. After downloading, unzip the packup to the project root. Please use at your own risk.

Training

Run training scripts in run_script (e.g., RoBERTa for RTE):

bash run_script/run_rte_roberta.sh

Implemented Results

Currently we have released our reimplementation on following tasks and datasets. More implementation will be released soon.

Released results on BERT-large

BoolQ COPA RTE WiC WSC CoNLL04 OntoNotes 5.0 CoNLL12
Result 74.3 77.0 80.1 75.1 68.3 84.5 86.4 85.3
Total Epochs 100 80 60 80 80 40 30 45
Best Epoch 58 12 30 56 17 33 24 43

Released results on RoBERTa-large

BoolQ COPA RTE WiC WSC CoNLL03 CoNLL04 OntoNotes 5.0 CoNLL12 CoNLL05 WSJ CoNLL05 Brown SQuAD 1.1 SQuAD 2.0
Results 84.0 92.0 86.6 73.7 64.4 91.8 88.4 90.1 84.7 89.4 83.9 88.1/94.2 81.3/84.7
Total Epochs 100 120 100 50 10 30 80 60 45 15 - 30 10
Best Epoch 86 78 65 31 3 28 45 59 37 13 - 24 9

For other hyper-parameters, please refer to the training scripts. If you can not achieve the reported results at the best epoch, there is probably an environmental mismatch and hyper-parameter search is needed.

Citation

If you find our work useful, please kindly cite our paper:

@article{DBLP:journals/corr/abs-2110-07602,
  author    = {Xiao Liu and
               Kaixuan Ji and
               Yicheng Fu and
               Zhengxiao Du and
               Zhilin Yang and
               Jie Tang},
  title     = {P-Tuning v2: Prompt Tuning Can Be Comparable to Fine-tuning Universally
               Across Scales and Tasks},
  journal   = {CoRR},
  volume    = {abs/2110.07602},
  year      = {2021},
  url       = {https://arxiv.org/abs/2110.07602},
  eprinttype = {arXiv},
  eprint    = {2110.07602},
  timestamp = {Fri, 22 Oct 2021 13:33:09 +0200},
  biburl    = {https://dblp.org/rec/journals/corr/abs-2110-07602.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}