/
agent.py
268 lines (196 loc) · 8.8 KB
/
agent.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
256
257
258
259
260
261
262
263
264
265
266
267
268
import numpy as np
import random
import time
import os
import logging
import models
from utils.config import load_model_config
import torch.nn.functional as F
import torch
import pickle as pkl
import os
# Set up logger
logging.basicConfig(
format='%(asctime)s:%(levelname)s:%(message)s',
level=logging.INFO
)
"""
Contains the definition of the agent that will run in an
environment.
"""
USE_GPU = True
class DQAgent:
def __init__(self,graph,args):
self.ckpt = args.ckpt
if not args.test_only and not os.path.exists(f'{self.ckpt}/model_ckpts'):
os.mkdir(f'{self.ckpt}/model_ckpts')
self.train = graph
self.test = pkl.load(open(args.test_file, 'rb'))
self.graphs = graph
self.embed_dim = 64
self.model_name = args.model
self.k = 20
self.alpha = 0.1
self.gamma = 0.99
self.lambd = 0.
self.n_step=3
self.epsilon_=1.0
self.epsilon_min=0.02
self.discount_factor =0.999990
self.eps_end=0.02
self.eps_start=1
self.eps_step=20000
self.t=1
self.memory = []
self.memory_n=[]
if self.model_name == 'S2V_QN_1':
args_init = load_model_config()[self.model_name]
self.model = models.S2V_QN_1(**args_init)
elif self.model_name == 'S2V_QN_2':
args_init = load_model_config()[self.model_name]
self.model = models.S2V_QN_2(**args_init)
elif self.model_name== 'GCN_QN_1':
args_init = load_model_config()[self.model_name]
self.model = models.GCN_QN_1(**args_init)
elif self.model_name == 'LINE_QN':
args_init = load_model_config()[self.model_name]
self.model = models.LINE_QN(**args_init)
elif self.model_name == 'W2V_QN':
args_init = load_model_config()[self.model_name]
self.model = models.W2V_QN(G=self.graphs[self.games], **args_init)
self.criterion = torch.nn.MSELoss(reduction='sum')
self.optimizer = torch.optim.Adam(self.model.parameters(), lr=1.e-3)
#self.optimizer_target = torch.optim.Adam(self.model_target.parameters(), lr=1.e-4)
self.T = 5
self.t = 1
if args.load_model != False:
self.model.load_state_dict(torch.load(args.load_model))
if USE_GPU:
self.criterion = self.criterion.cuda()
self.model = self.model.cuda()
"""
p : embedding dimension
"""
def reset(self, g, test=False):
if test:
self.graphs = self.test
else:
self.graphs = self.train
#self.memory=[]
#self.memory_n=[]
self.games = g
#self.epsilon=1
#if (len(self.memory) != 0) and (len(self.memory) % 300000 == 0):
# self.memory = random.sample(self.memory,120000)
if (len(self.memory_n) != 0) and (len(self.memory_n) % 300000 == 0):
self.memory_n =random.sample(self.memory_n,120000)
self.minibatch_length = 32
self.nodes = self.graphs[self.games].nodes()
self.adj = self.graphs[self.games].adj()
self.adj = self.adj.todense()
self.adj = torch.from_numpy(np.expand_dims(self.adj.astype(int), axis=0))
self.adj = self.adj.type(torch.FloatTensor)
self.last_action = 0
self.last_observation = torch.zeros(1, self.nodes, 1, dtype=torch.float)
self.last_reward = -0.01
self.last_done=0
self.iter=1
if test:
self.k = 20
self.alpha = 0.1
self.gamma = 0.99
self.lambd = 0.
self.n_step=3
self.epsilon_=1.0
self.epsilon_min=0.02
self.discount_factor =0.999990
self.eps_end=0.02
self.eps_start=1
self.eps_step=20000
self.t=1
self.memory = []
self.memory_n=[]
def act(self, observation):
if USE_GPU:
observation = observation.cuda()
if self.epsilon_ > np.random.rand():
return np.random.choice(np.where(observation.cpu().numpy()[0,:,0] == 0)[0])
else:
q_a = self.model(observation, self.adj)
q_a=q_a.detach().cpu().numpy()
return np.where((q_a[0, :, 0] == np.max(q_a[0, :, 0][observation.cpu().numpy()[0, :, 0] == 0])))[0][0]
def reward(self, observation, action, reward,done, train=True):
if len(self.memory_n) > self.minibatch_length + self.n_step: #or self.games > 2:
(last_observation_tens, action_tens, reward_tens, observation_tens, done_tens,adj_tens) = self.get_sample()
target = reward_tens + self.gamma *(1-done_tens)*torch.max(self.model(observation_tens, adj_tens).cpu() + observation_tens * (-1e5), dim=1)[0]
target_f = self.model(last_observation_tens, adj_tens)
target_p = target_f.clone()
target_f[range(self.minibatch_length),action_tens,:] = target.cuda()
loss=self.criterion(target_p, target_f)
if train:
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
#print(self.model.mu_2.weight[0])
#print(self.t, loss)
#self.epsilon = self.eps_end + max(0., (self.eps_start- self.eps_end) * (self.eps_step - self.t) / self.eps_step)
if self.epsilon_ > self.epsilon_min:
self.epsilon_ *= self.discount_factor
if self.iter>1:
self.remember(self.last_observation, self.last_action, self.last_reward, observation.clone(),self.last_done*1)
if done & self.iter> self.n_step:
self.remember_n(False)
new_observation = observation.clone()
new_observation[:,action,:]=1
self.remember(observation,action,reward,new_observation,done*1)
if self.iter > self.n_step:
self.remember_n(done)
self.iter+=1
self.t += 1
self.last_action = action
self.last_observation = observation.clone()
self.last_reward = reward
self.last_done = done
def get_sample(self):
minibatch = random.sample(self.memory_n, self.minibatch_length - 1)
minibatch.append(self.memory_n[-1])
last_observation_tens = minibatch[0][0]
action_tens = torch.Tensor([minibatch[0][1]]).type(torch.LongTensor)
reward_tens = torch.Tensor([[minibatch[0][2]]])
observation_tens = minibatch[0][3]
done_tens =torch.Tensor([[minibatch[0][4]]])
adj_tens = self.graphs[minibatch[0][5]].adj().todense()
adj_tens = torch.from_numpy(np.expand_dims(adj_tens.astype(int), axis=0)).type(torch.FloatTensor)
for last_observation_, action_, reward_, observation_, done_, games_ in minibatch[-self.minibatch_length + 1:]:
last_observation_tens=torch.cat((last_observation_tens,last_observation_))
action_tens = torch.cat((action_tens, torch.Tensor([action_]).type(torch.LongTensor)))
reward_tens = torch.cat((reward_tens, torch.Tensor([[reward_]])))
observation_tens = torch.cat((observation_tens, observation_))
done_tens = torch.cat((done_tens,torch.Tensor([[done_]])))
adj_ = self.graphs[games_].adj().todense()
adj = torch.from_numpy(np.expand_dims(adj_.astype(int), axis=0)).type(torch.FloatTensor)
adj_tens = torch.cat((adj_tens, adj))
return (last_observation_tens, action_tens, reward_tens, observation_tens,done_tens, adj_tens)
def remember(self, last_observation, last_action, last_reward, observation,done):
self.memory.append((last_observation, last_action, last_reward, observation,done, self.games))
def remember_n(self,done):
if not done:
step_init = self.memory[-self.n_step]
cum_reward=step_init[2]
for step in range(1,self.n_step):
cum_reward+=self.memory[-step][2]
self.memory_n.append((step_init[0], step_init[1], cum_reward, self.memory[-1][-3],self.memory[-1][-2], self.memory[-1][-1]))
else:
for i in range(1,self.n_step):
step_init = self.memory[-self.n_step+i]
cum_reward=step_init[2]
for step in range(1,self.n_step-i):
cum_reward+=self.memory[-step][2]
if i==self.n_step-1:
self.memory_n.append(
(step_init[0], step_init[1], cum_reward, self.memory[-1][-3], False, self.memory[-1][-1]))
else:
self.memory_n.append((step_init[0], step_init[1], cum_reward,self.memory[-1][-3], False, self.memory[-1][-1]))
def save_model(self, it):
torch.save(self.model.state_dict(), f'{self.ckpt}/model_ckpts/{it}.pt')
Agent = DQAgent