Prefix Tuning
Note:
作者在论文中提到使用真实的word去初始化prefix的操作（Initializing the prefix with activations of real words，significantly improves generation）。我在使用作者提供的代码时遇到了一些问题，因此按照代码的思路添加了利用真实词汇进行初始化的内容。
可以采用以下的方式运行：
Train
cd seq2seq;
python train_bart.py --mode xsum --preseqlen 200 --do_train yes --fp16 yes --bsz 16 --epoch 30 --gradient_accumulation_step 3 --learning_rate 0.00005 --mid_dim 800 --use_lowdata_token 'yes' --lowdata_token 'summarize'
其中use_lowdata_token表示是否采用real word初始化的方式；lowdata_token表示传入的real word.
Decode
cd seq2seq;
python train_bart.py --mode xsum --do_train no --prefix_model_path {checkpoint_path} --preseqlen {same as training} --mid_dim {same as training} --use_lowdata_token 'yes' --lowdata_token 'summarize'
Files:
.
├── gpt2 # Code for GPT2 style autoregressive LM
│ ├── train_e2e.py # high-level scripts to train.
│ ├── train_control.py # code that implements prefix-tuning.
│ ├── trainer_prefix.py # trainer code for the training loop.
│ ├── run_language_modeling.py # training code (contains data loading, model loading, and calls trainer)
│ ├── gen.py # high-level scripts to decode.
│ └── run_generation.py # decoding code.
│
├── seq2seq # Code for encoder-decoder architecture
│ ├── train_bart.py # high-level scripts to train.
│ ├── prefixTuning.py # code that implements prefix-tuning.
│ ├── finetune.py # training code (contains data loading, model loading, and calls trainer)
│ ├── lightning_base.py # helper code
│ ├── utils.py # helper code
│ └── callbacks.py # helper code
└── ...
To run the code for GPT2 style autoregressive LM, the code is in gpt2/. This corresponds to the table-to-text experiments in the paper.
To run the code for encoder-decoder architecture like BART, the code is in seq2seq. This corresponds to the summarization experiments in the paper.
The two primary scripts I used to run my codes are gpt2/train_e2e.py (for table-to-text) and seq2seq/train_bart.py(for summarization). they are set to default of good hyperparameters, and can be used to tune hyperparameter :)
Setup:
cd transformer; pip install -e .
Train via prefix-tuning:
cd gpt2;
python train_e2e.py --optim_prefix yes --preseqlen 5 --epoch 5 --learning_rate 0.00005 --mode webnlg --bsz 5 --seed 101
cd seq2seq;
python train_bart.py --mode xsum --preseqlen 200 --do_train yes --fp16 yes --bsz 16 --epoch 30 --gradient_accumulation_step 3 --learning_rate 0.00005 --mid_dim 800
Other baseline approaches
cd gpt2;
python train_e2e.py --tuning_mode {finetune/adaptertune} --epoch 5 --learning_rate 0.00005 --mode webnlg --bsz 5 --seed 101
cd seq2seq;
python train_e2e.py --tuning_mode finetune --epoch 5 --learning_rate 0.00005 --mode webnlg --bsz 5 --seed 101
Decode:
cd gpt2;
python gen.py {data2text/webnlg/...} yes test {checkpoint_path} no
cd seq2seq;
python train_bart.py --mode xsum --do_train no --prefix_model_path {checkpoint_path} --preseqlen {same as training} --mid_dim {same as training}
For details of the methods and results, please refer to our paper.
@misc{li2021prefixtuning,
title={Prefix-Tuning: Optimizing Continuous Prompts for Generation},
author={Xiang Lisa Li and Percy Liang},
year={2021},
eprint={2101.00190},
archivePrefix={arXiv},
primaryClass={cs.CL}
}
11 Mar 20, 2022
41 Nov 09, 2022
631 Jan 04, 2023
2.6k Dec 30, 2022
104 Sep 28, 2022
13 Dec 14, 2022
62 Dec 14, 2022
98 Nov 26, 2022
17 Dec 12, 2022
17 Dec 27, 2022
171 Dec 14, 2022
1 Jan 13, 2022
8 Oct 01, 2022
579 Dec 22, 2022
135 Dec 30, 2022
2 Apr 28, 2022
761 Dec 05, 2022
401 Jan 02, 2023
17 Dec 10, 2022
21 Jul 27, 2022