t5-pegasus pytorch
最新更新
- 重构代码，支持更多模型
- 支持transformers最新版本 老版代码点这里
模型效果对比
数据集：LCSTS_new 训练集取前一万条，验证集取前一千条
| model | bleu | rouge-1 | rouge-2 | rouge-l |
|---|---|---|---|---|
| t5-pegasus-base | 0.1276 | 0.3490 | 0.2123 | 0.3155 |
| t5-copy | 0.0938 | 0.3369 | 0.1955 | 0.3086 |
| Pegasus-238M-Chinese | 0.1200 | 0.3252 | 0.1957 | 0.2924 |
| Pegasus-523M-Chinese | 0.1233 | 0.3313 | 0.2032 | 0.2996 |
| cpt-large | 0.1366 | 0.3550 | 0.2242 | 0.3220 |
| prophet-zh | 0.1240 | 0.3419 | 0.2109 | 0.3107 |
huggingface模型
| model_type | model_type |
|---|---|
| t5-pegasus | imxly/t5-pegasus |
| t5copy | imxly/t5-copy |
| Pegasus | IDEA-CCNL/Randeng-Pegasus-238M-Chinese |
| Pegasus | IDEA-CCNL/Randeng-Pegasus-523M-Chinese |
| cpt | fnlp/cpt-large |
| prophet | imxly/prophetnet-zh |
训练命令
requirements
torch >=1.10.0
transformers
pytorch_lightning==1.4.9
torchmetrics==0.5.0
model_type见上方表格
python train.py \
--train_file train.json \
--dev_file dev.json \
--batch_size 6 \
--max_epochs 10 \
--max_source_length 512 \
--max_target_length 300 \
--model_path imxly/t5-pegasus \
--gpus 4 \
--lr 5e-5 \
--model_type t5-pegasus参考
https://github.com/ZhuiyiTechnology/t5-pegasus
https://github.com/fastnlp/CPT
https://github.com/IDEA-CCNL/Fengshenbang-LM
https://github.com/microsoft/ProphetNet