gpt2.py

import os
import numpy as np
from transformers import BertTokenizerFast
import migraphx

# 加载词汇表
print("INFO: Complete loading the vocabulary")
vocab_file = os.path.join('../Resource/', 'vocab_shici.txt')
tokenizer = BertTokenizerFast(vocab_file, sep_token="[SEP]", pad_token="[PAD]", cls_token="[CLS]")

# 设置最大输入shape
maxInput={"input":[1,1000]}

# 加载模型
print("INFO: Parsing and compiling the model")
model = migraphx.parse_onnx("../Resource/GPT2_shici.onnx", map_input_dims=maxInput)

# 获取模型输入/输出节点信息
inputs = model.get_inputs()
outputs = model.get_outputs()
inputName = model.get_parameter_names()[0]
inputShape = inputs[inputName].lens()

# 编译
model.compile(t=migraphx.get_target("gpu"), device_id=0)

print('开始和GPT2对诗，输入CTRL + Z以退出')
while True:
    try:
        history = []
        text = input("question:")
        text_ids = tokenizer.encode(text, add_special_tokens=False)
        history.extend(text_ids)
        input_ids = [tokenizer.cls_token_id] 
        input_ids.extend(text_ids)
        input_ids = np.array(input_ids, dtype=np.int64)
        input_ids = np.expand_dims(input_ids, axis=0) 
        
        max_len = 50
        for _ in range(max_len):
                
            # 推理
            result = model.run({inputName: input_ids})
            logits = [float(x) for x in result[0].tolist()]
            
            # 对于[UNK]的概率设为无穷小，模型的预测结果不可能是[UNK]
            logits[tokenizer.convert_tokens_to_ids('[UNK]')] = -float('Inf')
            
            # 排序
            score = []
            for index in range((input_ids.shape[1]-1)*22557, input_ids.shape[1]*22557):
                score.append(logits[index])
            index_and_score = sorted(enumerate(score), key=lambda x: x[1], reverse=True)
            
            # 取概率值最大的作为预测结果
            next_token = index_and_score[0][0]
            if next_token == tokenizer.convert_tokens_to_ids('[SEP]'):    # 遇到[SEP]结束标志符，结束循环
                break
            history.append(next_token)                                    # 结果存放在response列表中
            
            next_token = np.array(next_token, dtype=np.int64)
            input_ids = np.append(input_ids, next_token)
            input_ids = np.expand_dims(input_ids, axis=0)
        
        text = tokenizer.convert_ids_to_tokens(history)                   
        print("chatbot:" + "".join(text))

    except KeyboardInterrupt:
        break