PPT

Code and datasets for our paper "PPT: Pre-trained Prompt Tuning for Few-shot Learning"

1 Environment

The code requires the CUDA10.2 toolkit.

Install basic dependencies

pip install -r requirements.txt

Install apex

git clone https://github.com/NVIDIA/apex
cd apex
pip install -v --disable-pip-version-check --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./

Install DeepSpeed

The version we used is v0.3.9, It can be installed from its repo or

pip install deepspeed==0.3.9

Since there exist some bugs in DeepSpeed, you need to make some little modifications to this package. You can refer to this issue for more information. Specifically, you need to modify two lines of code in ${PATH_TO_PYTHON_SITE_PACKAGE}/deepspeed/runtime/zero/stage1.py and ${PATH_TO_PYTHON_SITE_PACKAGE}/deepspeed/runtime/engine.py. We provide the modified tools/ds_fix/stage1.py and tools/ds_fix/engine.py in our repo. You can simply replace ${PATH_TO_PYTHON_SITE_PACKAGE}/deepspeed/runtime/zero/stage1.py with stage1.py and ${PATH_TO_PYTHON_SITE_PACKAGE}/deepspeed/runtime/engine.py with engine.py that we provided.

2 Datasets

2.1 Downstream Datasets

The original datasets is obtained from huggingface.

The preprocessed datasets can be obtained from this link. If you do tuning (FT, PT, or PPT), you need to put the preprocessed data in downstream_data/.

2.2 Pre-training Data

Our pre-training data is sampled from openwebtext. If you would like to preprocess the data from scratch, please put the openwebtext.txt in pretrain_data/raw/. Run the following preprocessing scripts to construct the pre-training data:

bash scripts/tools/preprocess_pretrain_nsp.sh # Next Sentence Prediction
bash scripts/tools/preprocess_pretrain_nss.sh # Next Sentence Selection
bash scripts/tools/preprocess_pretrain_cls.sh # Single Sentence Classification
bash scripts/tools/preprocess_pretrain_nss_uni.sh # Unified Next Sentence Selection (for Unified PPT)

For reproductivity, we also provided the preprocessed pre-training data in this link. You can directly move the preprocessed pre-training data to pretrain_data/preprocessed/.

3 Pre-trained Checkpoints

3.1 Base Model

The original base model is obtained from huggingface. Before runing the code, please use the transforming scripts to transfer the original pytorch_model.bin model checkpoints to fit in our deepspeed + megatron framework:

mkdir -p checkpoints/t5-xxl/t5-MP4

python3 tools/transform.py \
--hf_path ${PATH_TO_PYTORCH_MODLE_BIN}
--save_path "./checkpoints/t5-xxl/t5-MP4"
--half

Note that our base model is T5.1.1

The transfered checkpoint should look like this:

t5-MP4/
├── 1
│   ├── mp_rank_00_model_states.pt
│   ├── mp_rank_01_model_states.pt
│   ├── mp_rank_02_model_states.pt
│   └── mp_rank_03_model_states.pt
└── latest_checkpointed_iteration.txt

3.2 Prompts

The pretrained prompts can be obtained from this link. You need to move the pre-tained prompts to pretrained_prompts/.

4 Run the code

All scripts are in the directory scripts.

Before running the code, please first change the WORKING_DIR to the current directory of this repo. If you are runing multiple scripts on a single node, you need to make sure that the MASTER_PORT of each script is different.

If the checkpoint is successfully loaded, the log printed to the stdout should contain messages like successfully loaded /path-to-checkpoint/t5-MP4/mp_rank_01_model_states.pt. Otherwise, WARNING: could not find the metadata file /***/latest_checkpointed_iteration.txt will not load any checkpoints and will start from random will display. Note that when you successfully load the model, you will see messages like The following zero checkpoints paths are missing: ['/path-to-checkpoint/200000/zero_pp_rank_0_mp_rank_00_optim_states.pt',... which mean optimizer states are not loaded. This DOES NOT affect the use of model inference and you can just ignore it.

4.1 Tuning

We use the cb dataset as an example. For t5-xxl model, PT and PPT can run on at least 4 * 32G V100 GPU. FT can run on at least 16 * 32G V100 GPU.

# few-shot 32 samples
bash scripts/cb/few-shot/ft.sh # Fine-tuning (FT)
bash scripts/cb/few-shot/pt.sh # Prompt Tuning (PT)
bash scripts/cb/few-shot/pt_pretrain.sh # Pre-trained Prompt Tuning (PPT)
bash scripts/cb/few-shot/pt_uni_pretrain.sh # Unified Pre-trained Prompt Tuning (Unified PPT)

# full data
bash scripts/cb/full/ft.sh # Fine-tuning (FT)
bash scripts/cb/full/pt.sh # Prompt Tuning (PT)
bash scripts/cb/full/pt_pretrain.sh # Pre-trained Prompt Tuning (PPT)
bash scripts/cb/full/pt_uni_pretrain.sh # Unified Pre-trained Prompt Tuning (Unified PPT)

4.2 Pre-training

bash scripts/pretrain/pretrain_nsp.sh # Next Sentence Prediction
bash scripts/pretrain/pretrain_nss.sh # Next Sentence Selelction
bash scripts/pretrain/pretrain_cls.sh # Single Sentence Classificatin
bash scripts/pretrain/pretrain_nss_uni.sh # Unified Next Sentence Selelction (for Unified PPT)

5 Cite

If you use the code, please cite the following paper:

@inproceedings{gu2022ppt,
  title={PPT: Pre-trained Prompt Tuning for Few-shot Learning},
  author={Gu, Yuxian and Han, Xu and Liu, Zhiyuan and Huang, Minlie},
  booktitle={Proceedings of ACL},
  year={2022}
}