- 增加t5-copy模型,在t5-pegasus的基础上增加了pointer generator,用t5-pegasus的预训练任务继续训练
- 增加t5-copy-large模型,在t5-copy的基础上用公开的文本摘要数据集进行训练
- 增加examples,基于pytorch_lightning的多卡训练
数据集:AdvertiseGen
model | bleu | rouge-1 | rouge-2 | rouge-2 |
---|---|---|---|---|
t5-pegasus-base | 0.087 | 0.4299 | 0.1834 | 0.2675 |
t5-copy | 0.089 | 0.4257 | 0.1814 | 0.2626 |
使用t5-copy模型transformers的版本不能高于4.12.0 pytorch-lightning<=1.4.9
模型名 | MODEL_NAME |
---|---|
t5-pegasus-base | imxly/t5-pegasus |
t5-pegasus-small | imxly/t5-pegasus-small |
t5-copy | imxly/t5-copy |
t5-copy-summary | imxly/t5-copy-summary |
pytorch1.7.0 + transformers4.3.3
from tokenizer import T5PegasusTokenizer
from transformers.models.mt5.modeling_mt5 import MT5ForConditionalGeneration
model_path = './'
model = MT5ForConditionalGeneration.from_pretrained(model_path)
tokenizer = T5PegasusTokenizer.from_pretrained(model_path)
text = '蓝蓝的天上有一朵白白的云'
ids = tokenizer.encode(text, return_tensors='pt')
output = model.generate(ids,
decoder_start_token_id=tokenizer.cls_token_id,
eos_token_id=tokenizer.sep_token_id,
max_length=30).numpy()[0]
print(''.join(tokenizer.decode(output[1:])).replace(' ', ''))
感谢追一科技开源的t5-pegasus