RLHS/main.py at main · SafeRoboticsLab/RLHS · 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
import argparse
import os
import random

from marketplace.env import Marketplace
from marketplace.chat import MarketplaceChatBot

from restaurant.env import Restaurant
from restaurant.chat import RestaurantChatBot

from course.env import OnlineCourse
from course.chat import OnlineCourseChatBot

from llm import load_model
import json
import pdb
import time
import torch


def create_directory(path):
    if not os.path.exists(path):
        os.makedirs(path)
        print(f"Directory {path} created.")
    else:
        print(f"Directory {path} already exists.")


def get_env(env_name):
    if args.env == "marketplace":
        ChatBot = MarketplaceChatBot
        Env = Marketplace
    elif args.env == "restaurant":
        ChatBot = RestaurantChatBot
        Env = Restaurant
    elif args.env == "online_course":
        ChatBot = OnlineCourseChatBot
        Env = OnlineCourse
    return ChatBot, Env


def gen_rlhf_train_data(env, ai_model, human_model, use_world_model, output, index):
    """This is the main one for training."""

    ChatBot, Env = get_env(args.env)

    chatbot = ChatBot(ai_model, human_model, use_world_model)
    output_path = f"{output}/results_{ai_model}_{human_model}_{index}.json"
    final_output_path = f"{output}/final_results_{ai_model}_{human_model}_{index}.json"
    env = Env(chatbot)

    all_results = []
    for i in range(10):
        results = env.generate_scenarios()
        all_results.extend(results)
        break
    all_results = random.sample(all_results, k=len(all_results))

    rlhf_data = env.generate_rlhf_data(all_results, output_path)
    with open(final_output_path, 'w') as file:
        json.dump(rlhf_data, file)


def gen_test_data_opt(env, ai_model_names, human_model_name, test_data, output, index, load_in_8bit, rlhf_type):
    ChatBot, Env = get_env(args.env)

    output_path = f"{output}/test_data_{index}.json"
    final_output_path = f"{output}/final_test_data_{index}.json"

    human_model, human_tokenizer = load_model(human_model_name, load_in_8bit=load_in_8bit)
    all_results = json.load(open(test_data, 'r'))
    create_directory(output)

    for ai_model_name in ai_model_names:
        ai_model, ai_tokenizer = load_model(ai_model_name)
        chatbot = ChatBot(ai_model_name, human_model_name, ai_model, ai_tokenizer, human_model, human_tokenizer)
        env = Env([chatbot])
        all_results = env.generate_test_data(all_results, output_path=output_path, rlhf_type=rlhf_type)
        output_path2 = f"{output}/test_data_{ai_model_name}_{index}.json"
        with open(output_path2, 'w') as file:
            json.dump(all_results, file)

    with open(final_output_path, 'w') as file:
        json.dump(all_results, file)


def test_inference_opt(env, ai_model_name, human_model_name, ai_model_directory, ckpts, test_data, output, index, load_in_8bit, rlhf_type):
    ChatBot, Env = get_env(args.env)

    output_path = f"{output}/test_inf_{ai_model_name}_{index}.json"
    final_output_path = f"{output}/final_test_inf_{ai_model_name}_{index}.json"

    human_model, human_tokenizer = load_model(human_model_name, load_in_8bit=load_in_8bit)
    all_results = json.load(open(test_data, 'r'))
    create_directory(output)

    for ckpt in ckpts:
        ckpts_path = f"{ai_model_directory}/{ckpt}"
        ai_model, ai_tokenizer = load_model(ai_model_name, model_ckpts=ckpts_path)
        chatbot = ChatBot(ai_model_name, human_model_name, ai_model, ai_tokenizer, human_model, human_tokenizer)
        env = Env([chatbot])
        all_results = env.generate_test_data(all_results, ckpts_list=[ckpt], output_path=output_path, rlhf_type=rlhf_type)
        output_path2 = f"{output}/test_inf_{ai_model_name}_{ckpt}.json"
        with open(output_path2, 'w') as file:
            json.dump(all_results, file)

        del ai_model
        del ai_tokenizer
        torch.cuda.empty_cache()

    with open(final_output_path, 'w') as file:
        json.dump(all_results, file)


def gen_new(env, ai_model_name, ai_model_directory, ckpts, test_data, output, index, rlhf_type):
    """This is just for human study"""
    ChatBot, Env = get_env(args.env)

    output_path = f"{output}/test_inf_{ai_model_name}_{index}.json"
    final_output_path = f"{output}/final_test_inf_{ai_model_name}_{index}.json"

    all_results = json.load(open(test_data, 'r'))
    create_directory(output)

    for ckpt in ckpts:
        ckpts_path = f"{ai_model_directory}/{ckpt}"
        ai_model, ai_tokenizer = load_model(ai_model_name, model_ckpts=ckpts_path)
        chatbot = ChatBot(ai_model_name, "None", ai_model, ai_tokenizer, "None", "None")
        env = Env([chatbot])
        all_results = env.generate_new(all_results, ckpts_list=[ckpt], output_path=output_path, rlhf_type=rlhf_type)
        output_path2 = f"{output}/test_inf_{ai_model_name}_{ckpt}.json"
        with open(output_path2, 'w') as file:
            json.dump(all_results, file)

        del ai_model
        del ai_tokenizer
        torch.cuda.empty_cache()

    with open(final_output_path, 'w') as file:
        json.dump(all_results, file)


def main(args):
    if args.task == "gen_rlhf_train_data":
        gen_rlhf_train_data(args.env, args.ai_model[0], args.human_model, args.use_world_model, args.output, args.index)
    elif args.task == "gen_test_data":
        gen_test_data_opt(args.env, args.ai_model, args.human_model, args.test_data, args.output, args.index, args.load_in_8bit, args.rlhf_type)
    elif args.task == "test_inference":
        test_inference_opt(args.env, args.ai_model[0], args.human_model, args.ai_model_directory, args.ai_model_ckpts, args.test_data, args.output, args.index, args.load_in_8bit, args.rlhf_type)
    else:
        gen_new(args.env, args.ai_model[0], args.ai_model_directory, args.ai_model_ckpts, args.test_data, args.output, args.index, args.rlhf_type)


if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    # Support model: llama-2-7b, llama-2-13b, llama-3-8b, llama-3-70b
    parser.add_argument('--ai_model', nargs='+', default=['llama-3-8b'], help="List of AI models or a single AI model")
    parser.add_argument('--ai_model_directory', default=None)
    parser.add_argument('--ai_model_ckpts', nargs='+', default=None, help="List of checkpoints")
    parser.add_argument('--human_model', default='llama-3.1-70b')
    parser.add_argument('--use_world_model', action='store_true', default=False, help="You can either use AI's world model to simulate the outcome or used deterministic outcome from the actual simulator.")
    parser.add_argument('--load_in_8bit', action='store_true', default=False, help="Whether the human model load in 8bit")
    parser.add_argument('--output', default='./data')
    parser.add_argument('--index', default=1, type=int)
    parser.add_argument('--task', choices=['gen_rlhf_train_data', "gen_test_data", "test_inference", "gen_new"], default='gen_rlhf_train_data', help="Task to be performed")
    parser.add_argument('--env', choices=['marketplace', "restaurant", "online_course"], default='marketplace', help="Environment to use")
    # Specific only to test_inference
    parser.add_argument('--rlhf_type', nargs='+', default=['base', 'phs', 'fhs', 'phs_v0', 'fhs_v0'], help="List of different type of RLHF is used")
    # Generate test data. If the test file is specified, then use this test file instead of generating all new scenarios.
    parser.add_argument('--test_data', default='test_data/test_marketplace.json')

    args = parser.parse_args()
    main(args)