-
Notifications
You must be signed in to change notification settings - Fork 0
/
single_dqn.py
255 lines (211 loc) · 10.1 KB
/
single_dqn.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
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
import random
import numpy as np
import pandas as pd
import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
import os
from environment.graph import KnowledgeGraph
from environment.chatenv import StoryBotRetellEnv
import json
import logging
from datetime import datetime
# class of DQN
class DQN(nn.Module):
def __init__(self, env, gamma=0.99, epsilon=1.0, epsilon_decay=0.99, epsilon_min=0.01, lr=0.001) -> None:
super(DQN, self).__init__()
self.env = env
self.gamma = gamma
self.epsilon = epsilon
self.epsilon_decay = epsilon_decay
self.epsilon_min = epsilon_min
self.lr = lr
self.memory = []
self.model = nn.Sequential(
torch.nn.Linear(self.env.observation_space, 128),
torch.nn.ReLU(),
torch.nn.Linear(128, 128),
torch.nn.ReLU(),
torch.nn.Linear(128, len(self.env.action_space))
)
self.model.to(self.env.device)
self.optimizer = optim.Adam(self.model.parameters(), lr=self.lr)
def forward(self, x):
return self.model(x)
def act(self, state):
if np.random.rand() <= self.epsilon:
return random.randint(0, len(self.env.action_space)-1)
else:
return torch.argmax(self.forward(state)).item()
def remember(self, state, action, reward, next_state, done):
self.memory.append((state, action, reward, next_state, done))
def replay(self, batch_size=32):
if len(self.memory) < batch_size:
return
batch = random.sample(self.memory, batch_size)
for state, action, reward, next_state, done in batch:
target = reward
if not done:
target = reward + self.gamma * torch.max(self.forward(next_state))
target_f = self.forward(state)
target_f[0][action] = target
self.optimizer.zero_grad()
loss = F.mse_loss(self.forward(state), target_f)
loss.backward()
self.optimizer.step()
# if self.epsilon > self.epsilon_min:
# self.epsilon *= self.epsilon_decay
def update_epsilon(self):
if self.epsilon > self.epsilon_min:
self.epsilon *= self.epsilon_decay
def save(self, path):
torch.save(self.model.state_dict(), path)
# class of DQN agent
class DQNAgent:
def __init__(self, env, gamma=0.99, epsilon=1.0, epsilon_decay=0.7, epsilon_min=0.01, lr=0.001) -> None:
self.env = env
self.gamma = gamma
self.epsilon = epsilon
self.epsilon_decay = epsilon_decay
self.epsilon_min = epsilon_min
self.lr = lr
self.memory = []
self.model = DQN(self.env, self.gamma, self.epsilon, self.epsilon_decay, self.epsilon_min, self.lr)
def act(self, state):
return self.model.act(state)
def remember(self, state, action, reward, next_state, done):
self.model.remember(state, action, reward, next_state, done)
def replay(self, batch_size=32):
self.model.replay(batch_size)
def save(self, path):
self.model.save(path)
def load(self, path):
self.model.model.load_state_dict(torch.load(path))
self.model.model.eval()
# trainer of DQN agent
class DQNTrainer:
def __init__(self, env1: StoryBotRetellEnv, env2: StoryBotRetellEnv, agent1: DQNAgent, agent2: DQNAgent, \
story_summary_path, story_kg_path, epoch=100, episodes=100, batch_size=8) -> None:
# environment
self.env1 = env1
# self.env2 = env2
# agent
self.agent1 = agent1
self.agent2 = agent2
self.agent1_model_name = 'model/dqn1.pth'
self.agent2_model_name = 'model/dqn2.pth'
# training parameters
self.episodes = episodes
self.batch_size = batch_size
self.epoch = epoch
# all story names and knowledge graphs
self.story_summary_path = story_summary_path
self.story_kg_path = story_kg_path
with open(self.story_summary_path, 'r') as f:
self.story_summary = json.load(f)
self.story_name_list = list(self.story_summary.keys())
self.kg_dict = {}
for story_name in self.story_name_list:
kg = KnowledgeGraph(device=self.env1.embedding_model_device,
model=self.env1.embedding_model,
tokenizer=self.env1.embedding_tokenizer,
story_kg_file=os.path.join(self.story_kg_path, story_name+'.json'))
self.kg_dict[story_name] = kg
# current story name and knowledge graph
self.current_story_name = None
self.current_kg = None
def set_current_story_kg(self, story_name):
self.current_story_name = story_name
self.current_kg = self.kg_dict[story_name]
def merge_dicts(self, dict1, dict2):
merged_dict = dict1.copy()
for key, value in dict2.items():
if key in merged_dict:
merged_dict[key] += value
else:
merged_dict[key] = value
return merged_dict
def train(self):
dt_start_str = datetime.now().strftime("%Y%m%d_%H%M%S")
df = pd.DataFrame(columns=['epoch', 'episode', 'story_name', 'score1', 'score2', 'epsilon1', 'epsilon2'])
best_score_1 = -1
# best_score_2 = -1
dialogue_history_df1 = pd.DataFrame()
for e in range(self.epoch):
score1_list = []
score2_list = []
similarity1_list = []
similarity2_list = []
action_counter = {}
# for episode in range(self.episodes):
for episode, story in enumerate(self.story_name_list):
self.set_current_story_kg(story)
self.env1.reset(story_name=self.current_story_name, story_kg=self.current_kg)
# self.env2.reset(story_name=self.current_story_name, story_kg=self.current_kg)
output_dialogue2 = ''
output_kg2 = None
self.env1.render(input_sentence=output_dialogue2, input_kg=output_kg2)
while True:
# two agent talk with each other
# self.env1.render(input_sentence=output_dialogue2, input_kg=output_kg2)
done1, done1_msg = self.env1.done()
if not done1:
state1 = self.env1.observation()
state1 = torch.tensor(state1, dtype=torch.float32).unsqueeze(0)
action1 = self.agent1.act(state1)
output_dialogue1, output_kg1 = self.env1.step(action1)
next_state1 = self.env1.observation()
reward1, score1 = self.env1.reward()
done1, done1_msg = self.env1.done()
self.agent1.remember(state1, action1, reward1, next_state1, done1)
if done1:
break
# self.env1.render(input_sentence=output_dialogue1, input_kg=output_kg1)
done2, done2_msg = self.env1.done()
if not done2:
state2 = self.env1.observation()
state2 = torch.tensor(state2, dtype=torch.float32).unsqueeze(0)
action2 = self.agent2.act(state2)
output_dialogue2, output_kg2 = self.env1.step(action2)
next_state2 = self.env1.observation()
reward2, score2 = self.env1.reward()
done2, done2_msg = self.env1.done()
self.agent2.remember(state2, action2, reward2, next_state2, done2)
if done2:
break
self.agent1.replay(self.batch_size)
self.agent2.replay(self.batch_size)
final_score1 = self.env1.current_score
# final_score2 = self.env2.current_score
similarity1 = self.env1.dialogue_summary_similarity()
# similarity2 = self.env2.dialogue_summary_similarity()
action_counter = self.merge_dicts(action_counter, self.env1.count_actions())
# action_counter = self.merge_dicts(action_counter, self.env2.count_actions())
# append the result to df
df = df.append({'epoch': e, 'episode': episode, 'story_name': self.current_story_name, \
'score1': final_score1, \
'epsilon1': self.agent1.model.epsilon, 'epsilon2': self.agent2.model.epsilon, \
'similarity1': similarity1}, ignore_index=True)
dialogue_history_df1 = dialogue_history_df1.append(self.env1.dialogue_log_list, ignore_index=True)
dialogue_history_df1.to_csv(f'output/dialogue_history1_{dt_start_str}.csv', index=False)
score1_list.append(final_score1)
# score2_list.append(final_score2)
similarity1_list.append(similarity1)
# similarity2_list.append(similarity2)
self.agent1.model.update_epsilon()
self.agent2.model.update_epsilon()
score1_mean = np.mean(score1_list)
# score2_mean = np.mean(score2_list)
similarity1_mean = np.mean(similarity1_list)
# similarity2_mean = np.mean(similarity2_list)
logging.info('epoch: {:2}, score: {:.4f}, similarity: {:.4f}, action:{}'.format(e, score1_mean, similarity1_mean, action_counter))
if score1_mean > best_score_1:
best_score_1 = score1_mean
self.agent1.save(self.agent1_model_name)
# if score1_mean > best_score_1:
# best_score_2 = score2_mean
# self.agent2.save(self.agent2_model_name)
# self.agent1.save(self.agent1_model_name)
# self.agent2.save(self.agent2_model_name)
df.to_csv(f'output/result_{dt_start_str}.csv', index=False)