如何使用停用词列表提前停止自回归模型？

Question

我正在使用 GPT-Neo 模型来transformers生成文本。因为我使用的提示以 开头'{'，所以我想在'}'生成配对后停止该句子。我发现源代码中有一个StoppingCriteria方法，但没有进一步说明如何使用它。有人找到了提前停止模型生成的方法吗？谢谢！

这是我尝试过的：

from transformers import StoppingCriteria, AutoModelForCausalLM, AutoTokenizer
model_name = 'gpt2'
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, pad_token_id=tokenizer.eos_token_id, torch_dtype=dtype).eval()

class KeywordsStoppingCriteria(StoppingCriteria):
    def __init__(self, keywords_ids:list):
        self.keywords = keywords_ids

    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
        if input_ids in self.keywords:
            return True
        return False

stop_words = ['}', ' }', '\n']
stop_ids = [tokenizer.encode(w) for w in stop_words]
stop_ids.append(tokenizer.eos_token_id)
stop_criteria = KeywordsStoppingCriteria(stop_ids)

model.generate(
    text_inputs='some text:{', 
    StoppingCriteria=stop_criteria
)

Answer 1

我已经能够调整您的代码以使其工作。此外，请确保您使用的是最新版本的变压器，您可能需要升级。

import torch
from transformers import StoppingCriteria, AutoModelForCausalLM, AutoTokenizer, StoppingCriteriaList
model_name = 'gpt2'
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, pad_token_id=tokenizer.eos_token_id).eval()

class KeywordsStoppingCriteria(StoppingCriteria):
    def __init__(self, keywords_ids:list):
        self.keywords = keywords_ids

    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
        if input_ids[0][-1] in self.keywords:
            return True
        return False


stop_words = ['}', ' }', '\n']
stop_ids = [tokenizer.encode(w)[0] for w in stop_words]
stop_criteria = KeywordsStoppingCriteria(stop_ids)


inputs = tokenizer.encode('some text: {', add_special_tokens=False, return_tensors='pt')

output = model.generate(
    inputs,
    do_sample=True,
    stopping_criteria=StoppingCriteriaList([stop_criteria]),

)
print(tokenizer.decode(*output))