一、使用Chinese GPT2 Model

from transformers import BertTokenizer, GPT2LMHeadModel, TextGenerationPipeline

tokenizer = BertTokenizer.from_pretrained("uer/gpt2-chinese-cluecorpussmall")
model = GPT2LMHeadModel.from_pretrained("uer/gpt2-chinese-cluecorpussmall")
text = TextGenerationPipeline(model, tokenizer)
generate = text("2022年8月7日星期天，今天我收到了秋天的第一杯奶茶", max_length=100, do_sample=True)
print(generate)

输出结果：

二、使用transformers通用方法

from transformers import AutoTokenizer, AutoModelWithLMHead
import torch, os
import pandas as pd

os.environ["CUDA_VISIBLE_DEVICES"] = "0"

tokenizer = AutoTokenizer.from_pretrained("uer/gpt2-chinese-cluecorpussmall")
model = AutoModelWithLMHead.from_pretrained("uer/gpt2-chinese-cluecorpussmall")

config=model.config

print(config)

device = cuda if torch.cuda.is_available() else cpu
model = model.to(device)

# data = pd.read_excel(D:/pthon.1/NLP/bigbird/data/殖装_处理后.xlsx)
# texts = data[content].tolist()
texts = ["用我三生烟火，换你一世迷离 只缘感君一回顾，使我思君朝与暮",
         "一座城一个人，一盏花灯一场烟火，一棵古树一地雪玉残香，一人回眸凝望，一世繁华无殇",
         "纵然万劫不复，纵然相思入骨，我也待你眉眼如初，岁月如故",
         "于我虽一眼惊鸿，于你却似指尖清风"]


encoding = tokenizer(texts, return_tensors=pt, padding=True).to(device)

with torch.no_grad():
    generated_ids = model.generate(**encoding,
    max_length=200, # 生成序列的最大长度
    do_sample=True, #是否开启采样，默认是 False，即贪婪找最大条件概率的词
    top_k=20, # top-k-filtering 算法保留多少个 最高概率的词 作为候选，默认50
    repetition_penalty=1.0, #重复词惩罚
    temperature=1.0 # 增加了高概率词的可能性，降低了低概率词的可能性
    )
generated_texts = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)

for txt in generated_texts:
    print(txt)

输出结果：

三、simpletransformers