MapDia/ours.py at main · FrontierLabs/MapDia · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
#!/usr/bin/env python3
# -*- coding:utf-8 -*-
"""
@file: ours.py
@time: 2024/5/21 9:28
@desc:
"""
import json

import torch
from tqdm import tqdm

from ranker_utils import rank_model_retrieval
from utils import load_rank_model, load_qwen, load_config
from test_data import load_episodes


config = load_config()


def _add_or_replace_eos_token(tokenizer: "PreTrainedTokenizer", eos_token: str) -> None:
    is_added = tokenizer.eos_token_id is None
    is_oov = eos_token not in tokenizer.get_vocab()
    tokenizer.add_special_tokens({"eos_token": eos_token})

    if is_added:
        print("Add eos token: {}".format(tokenizer.eos_token))
    else:
        print("Replace eos token: {}".format(tokenizer.eos_token))

    if is_oov:
        print("New tokens have been added, make sure `resize_vocab` is True.")


rank_device = torch.device('cuda:3')
rank_path = config['model_pathes']['ranker']['model']
vhead_file = config['model_pathes']['ranker']['vhead']

rank_model, rank_tokenizer = load_rank_model(rank_path, vhead_file, rank_device)

xiaoming_device = torch.device('cuda:2')

xiaoming_path = config['model_pathes']['user_model']
xiaoming_model, xiaoming_tokenizer = load_qwen(xiaoming_path, xiaoming_device)

# xiaoming_tokenizer没有设置<|im_end|>为eos_token
_add_or_replace_eos_token(xiaoming_tokenizer, '<|im_end|>')

callback_device = torch.device('cuda:1')
callback_path = config['model_pathes']['callback']
callback_model, callback_tokenizer = load_qwen(callback_path, callback_device)

xiaoming_template = """<|im_start|>system
你是小明，正在与对话机器人Lucy进行聊天。<|im_end|>
"""

# callback的history里面是有topic和sub_topic的，要注意还原回去
callback_template = """<|im_start|>system
你是Lucy，与小明进行聊天，依据历史对话和当前对话，进行回复。聊天的过程中，需要结合给定历史对话，考虑是否可以将当前对话主题转移到历史对话主题，时机的判断以Thoughts形式展示。<|im_end|>
<|im_start|>user
对话历史如下：
###
{history}
###
当前对话如下：
###
{dialog}
###
<|im_end|>
<|im_start|>assistant"""

template = """<|im_start|>system
你是一个对话标题打分助手，针对给定的对话与标题，依据其相关程度进行打分。<|im_end|>
<|im_start|>user
{dialog}<|im_start|>assistant
{sub_topic}<|im_end|>"""


def check_callback_output(callback_output):
    try:
        thoughts, dialog = callback_output.split('\n')
    except:
        return False
    if not thoughts.startswith('Thoughts'):
        return False
    if not dialog.startswith('Lucy'):
        return False
    return True


def chat(history, dialog, dialog_with_thoughts, greedy=False):
    xiaoming_prompt = xiaoming_template
    for idx, d in enumerate(dialog.strip().replace('\n\n', '\n').split('\n')):
        if idx % 2 == 0:
            xiaoming_prompt += '<|im_start|>user\n' + d.replace('小明：', '') + '<|im_end|>\n'
        else:
            xiaoming_prompt += '<|im_start|>assistant\n' + d.replace('Lucy：', '') + '<|im_end|>\n'
    xiaoming_input = xiaoming_tokenizer(xiaoming_prompt, return_tensors='pt').to(xiaoming_model.device)
    with torch.no_grad():
        if greedy:
            xiaoming_output = xiaoming_model.generate(**xiaoming_input, do_sample=False,
                                                      max_new_tokens=100,
                                                      eos_token_id=xiaoming_tokenizer.eos_token_id,
                                                      pad_token_id=xiaoming_tokenizer.eos_token_id)
        else:
            xiaoming_output = xiaoming_model.generate(**xiaoming_input, do_sample=True,
                                                      max_new_tokens=100,
                                                      temperature=1.,
                                                      eos_token_id=xiaoming_tokenizer.eos_token_id,
                                                      pad_token_id=xiaoming_tokenizer.eos_token_id)
    xiaoming_output = '小明：' + xiaoming_tokenizer.decode(xiaoming_output[0]).split('<|im_start|>user')[-1].replace(
        '<|im_end|>', '').strip()
    dialog += '\n\n' + xiaoming_output
    dialog_with_thoughts += '\n\n' + xiaoming_output

    callback_prompt = callback_template.replace('{history}', history).replace('{dialog}', dialog)
    callback_input = callback_tokenizer(callback_prompt, return_tensors='pt').to(callback_model.device)

    # 如果失败，最多生成三次
    for i in range(3):
        with torch.no_grad():
            callback_output = callback_model.generate(**callback_input, do_sample=True,
                                                      max_new_tokens=250,
                                                      temperature=1.,
                                                      eos_token_id=xiaoming_tokenizer.eos_token_id,
                                                      pad_token_id=xiaoming_tokenizer.eos_token_id)
            callback_output = callback_tokenizer.decode(callback_output[0]).split('<|im_start|>assistant')[1].replace(
                '<|im_end|>', '').strip()
        if check_callback_output(callback_output):
            break
        else:
            print(callback_output)

    dialog_with_thoughts += '\n' + callback_output
    # remove thoughts
    try:
        dialog += '\n' + callback_output.split('\n')[1]
    except:
        dialog += '\n' + callback_output
    return dialog, dialog_with_thoughts


def format_history(history, topic, sub_topic):
    history = '主题：{}\n细化主题：{}\n\n'.format(topic, sub_topic) + history
    return history


episodes = load_episodes(config)

## 检索一次
round = 10
for episode in tqdm(episodes):

    dialog = dialog_with_thoughts = episode['start']
    topic = episode['topic']
    sub_topic = episode['sub_topic']
    rank = rank_model_retrieval(dialog, sub_topic, rank_model, rank_tokenizer)

    top_index = rank[0]
    history = episode['dialog'][top_index]
    history = format_history(history, topic[top_index], sub_topic[top_index])
    episode['retrieve_once'] = [top_index]
    for r in range(round):
        if r == 0:
            dialog, dialog_with_thoughts = chat(history, dialog, dialog_with_thoughts, greedy=True)
        else:
            dialog, dialog_with_thoughts = chat(history, dialog, dialog_with_thoughts)
    episode['search_once_dialog_with_thoughts'] = dialog_with_thoughts
    torch.cuda.empty_cache()

## 检索每轮
round = 10
for episode in tqdm(episodes):

    dialog = dialog_with_thoughts = episode['start']
    topic = episode['topic']
    sub_topic = episode['sub_topic']

    episode['retrieve_each_round'] = []
    for r in range(round):
        rank = rank_model_retrieval(dialog, sub_topic, rank_model, rank_tokenizer)
        top_index = rank[0]
        episode['retrieve_each_round'].append(top_index)
        history = episode['dialog'][top_index]
        history = format_history(history, topic[top_index], sub_topic[top_index])
        if r == 0:
            dialog, dialog_with_thoughts = chat(history, dialog, dialog_with_thoughts, greedy=True)
        else:
            dialog, dialog_with_thoughts = chat(history, dialog, dialog_with_thoughts)
    episode['search_each_round_dialog_with_thoughts'] = dialog_with_thoughts
    torch.cuda.empty_cache()

with open('data/ours_result.json', 'w', encoding='utf-8') as f:
    f.write(json.dumps(episodes, ensure_ascii=False))