-
Notifications
You must be signed in to change notification settings - Fork 2
/
interactive.py
216 lines (184 loc) · 6.76 KB
/
interactive.py
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
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
import numpy as np
import logging
import argparse
from tqdm.auto import tqdm
from core.gen_models import (
LocalModel, OpenAIModel, OpenAIChatModel, AzureOpenAIChatModel
)
from core.players import (
PersuadeeModel, PersuaderModel, P4GSystemPlanner,
PersuaderChatModel, PersuadeeChatModel, P4GChatSystemPlanner
)
from core.game import PersuasionGame
from core.mcts import MCTS, OpenLoopMCTS, OpenLoopMCTSParallel
from core.helpers import DialogSession
from utils.utils import dotdict
from utils.prompt_examples import EXP_DIALOG
logger = logging.getLogger(__name__)
def play_gdpzero(backbone_model, args):
args = dotdict({
"cpuct": 1.0,
"num_MCTS_sims": args.num_mcts_sims,
"max_realizations": args.max_realizations,
"Q_0": args.Q_0,
})
game_ontology = PersuasionGame.get_game_ontology()
sys_da = game_ontology['system']['dialog_acts']
user_da = game_ontology['user']['dialog_acts']
system_name = PersuasionGame.SYS
user_name = PersuasionGame.USR
exp_1 = DialogSession(system_name, user_name).from_history(EXP_DIALOG)
system = PersuaderChatModel(
sys_da,
backbone_model,
conv_examples=[exp_1],
inference_args={
"temperature": 0.7,
"do_sample": True, # for MCTS open loop
"return_full_text": False,
}
)
user = PersuadeeChatModel(
user_da,
inference_args={
"max_new_tokens": 128,
"temperature": 1.1,
"repetition_penalty": 1.0,
"do_sample": True, # for MCTS open loop
"return_full_text": False,
},
backbone_model=backbone_model,
conv_examples=[exp_1]
)
planner = P4GChatSystemPlanner(
dialog_acts=system.dialog_acts,
max_hist_num_turns=system.max_hist_num_turns,
user_dialog_acts=user.dialog_acts,
user_max_hist_num_turns=user.max_hist_num_turns,
generation_model=backbone_model,
conv_examples=[exp_1]
)
game = PersuasionGame(system, user)
state = game.init_dialog()
# init
state.add_single(game.SYS, 'greeting', "Hello. How are you?")
print("You are now the Persuadee. Type 'q' to quit, and 'r' to restart.")
print("Persuader: Hello. How are you?")
your_utt = input("You: ")
while your_utt.strip() != "q":
if your_utt.strip() == "r":
state = game.init_dialog()
state.add_single(game.SYS, 'greeting', "Hello. How are you?")
game.display(state)
your_utt = input("You: ")
continue
# used for da prediction
tmp_state = state.copy()
tmp_state.add_single(game.USR, 'neutral', your_utt.strip())
user_da = user.predict_da(tmp_state)
logging.info(f"user_da: {user_da}")
state.add_single(game.USR, user_da, your_utt.strip())
# planning
if isinstance(backbone_model, OpenAIModel):
backbone_model._cached_generate.cache_clear()
dialog_planner = OpenLoopMCTS(game, planner, args)
for i in tqdm(range(args.num_MCTS_sims)):
dialog_planner.search(state)
mcts_policy = dialog_planner.get_action_prob(state)
mcts_policy_next_da = system.dialog_acts[np.argmax(mcts_policy)]
logger.info(f"mcts_policy: {mcts_policy}")
logger.info(f"mcts_policy_next_da: {mcts_policy_next_da}")
logger.info(dialog_planner.Q)
sys_utt = dialog_planner.get_best_realization(state, np.argmax(mcts_policy))
logging.info(f"sys_da: [{mcts_policy_next_da}]")
print(f"Persuader: {sys_utt}")
state.add_single(game.SYS, mcts_policy_next_da, sys_utt)
your_utt = input("You: ")
return
def play_raw_prompt(backbone_model):
system_name = PersuasionGame.SYS
user_name = PersuasionGame.USR
exp_1 = DialogSession(system_name, user_name).from_history(EXP_DIALOG)
game_ontology = PersuasionGame.get_game_ontology()
sys_da = game_ontology['system']['dialog_acts']
user_da = game_ontology['user']['dialog_acts']
system = PersuaderChatModel(
sys_da,
backbone_model,
conv_examples=[exp_1]
)
user = PersuadeeChatModel(
user_da,
inference_args={
"max_new_tokens": 128,
"temperature": 1.1,
"repetition_penalty": 1.0,
"do_sample": True,
"return_full_text": False,
},
backbone_model=backbone_model,
conv_examples=[exp_1]
)
planner = P4GChatSystemPlanner(
dialog_acts=system.dialog_acts,
max_hist_num_turns=system.max_hist_num_turns,
user_dialog_acts=user.dialog_acts,
user_max_hist_num_turns=user.max_hist_num_turns,
generation_model=backbone_model,
conv_examples=[exp_1]
)
game = PersuasionGame(system, user)
state = game.init_dialog()
# init
state.add_single(game.SYS, 'greeting', "Hello. How are you?")
print("You are now the Persuadee. Type 'q' to quit, and 'r' to restart.")
print("Persuader: Hello. How are you?")
your_utt = input("You: ")
while your_utt.strip() != "q":
if your_utt.strip() == "r":
state = game.init_dialog()
state.add_single(game.SYS, 'greeting', "Hello. How are you?")
game.display(state)
your_utt = input("You: ")
continue
# used for da prediction
state.add_single(game.USR, 'neutral', your_utt.strip())
# planning
prior, v = planner.predict(state)
greedy_policy = system.dialog_acts[np.argmax(prior)]
next_best_state = game.get_next_state(state, np.argmax(prior))
greedy_pred_resp = next_best_state.history[-2][2]
logging.info(f"sys_da: [{greedy_policy}]")
print(f"Persuader: {greedy_pred_resp}")
state.add_single(game.SYS, greedy_policy, greedy_pred_resp)
your_utt = input("You: ")
return
def main(args):
if args.llm in ['code-davinci-002', 'text-davinci-003']:
backbone_model = OpenAIModel(args.llm)
elif args.llm in ['gpt-3.5-turbo']:
backbone_model = OpenAIChatModel(args.llm, args.gen_sentences)
elif args.llm == 'chatgpt':
backbone_model = AzureOpenAIChatModel(args.llm, args.gen_sentences)
if args.algo == 'gdpzero':
print("using GDPZero as planning algorithm")
play_gdpzero(backbone_model, args)
elif args.algo == 'raw-prompt':
print("using raw prompting as planning")
play_raw_prompt(backbone_model)
return
if __name__ == "__main__":
# logging mode
parser = argparse.ArgumentParser()
parser.add_argument("--log", type=int, default=logging.WARNING, help="logging mode", choices=[logging.INFO, logging.DEBUG, logging.WARNING])
parser.add_argument("--algo", type=str, default='gdpzero', choices=['gdpzero', 'raw-prompt'], help="planning algorithm")
# used by PDP-Zero
parser.add_argument('--llm', type=str, default="gpt-3.5-turbo", choices=["code-davinci-002", "gpt-3.5-turbo", "text-davinci-002", "chatgpt"], help='OpenAI model name')
parser.add_argument('--gen_sentences', type=int, default=3, help='number of sentences to generate from the llm. Longer ones will be truncated by nltk.')
parser.add_argument('--num_mcts_sims', type=int, default=10, help='number of mcts simulations')
parser.add_argument('--max_realizations', type=int, default=3, help='number of realizations per mcts state')
parser.add_argument('--Q_0', type=float, default=0.25, help='initial Q value for unitialized states. to control exploration')
args = parser.parse_args()
logging.basicConfig(level=args.log)
logger.setLevel(args.log)
main(args)