-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathdqn.py
183 lines (151 loc) · 6.61 KB
/
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
import torch
import torch.nn as nn
import torch.optim as optim
#from drl_env import ReplayBuffer
import random
import numpy as np
from copy import deepcopy
device = torch.device('cpu')
if (torch.cuda.is_available()):
device = torch.device('cuda:0')
torch.cuda.empty_cache()
#print("Device set to : " + str(torch.cuda.get_device_name(device)))
else:
pass
#device = torch.device('cpu:0')
class RolloutBuffer(object):
def __init__(self, capacity):
self.capacity = capacity
self.buffer = []
def add(self, s0, a, r, s1, done):
if len(self.buffer) >= self.capacity:
self.buffer.pop(0)
self.buffer.append((s0, a, r, s1, done))
def sample(self, batch_size):
#print(random.sample(self.buffer, batch_size))
s0, a, r, s1, done = zip(*random.sample(self.buffer, batch_size))
s0 = torch.tensor(s0, dtype=torch.float)#.cuda()
s1 = torch.tensor(s1, dtype=torch.float)#.cuda()
a = torch.tensor(a, dtype=torch.long)#.cuda()
r = torch.tensor(r, dtype=torch.float)#.cuda()
done = torch.tensor(done, dtype=torch.float)#.cuda()
return s0, a, r, s1, done
def size(self):
return len(self.buffer)
class DQNAgent(nn.Module):
def __init__(self, state_size, action_size, device):
super(DQNAgent, self).__init__()
self.buffer = RolloutBuffer(2000)
self.state_size = state_size
self.action_size = action_size
self.gamma = 0.9 # discount rate
self.epsilon = 1.0 # exploration rate 1.0
self.epsilon_min = 0.001 #0.001
self.epsilon_decay = 0.99 #0.9
self.learning_rate = 0.02 #0.02
size = 256
#224:optimal:5/100,better than greedy:55/100, at least greedy:76/100,320.0211285299997s
#280:optimal:12/100,better than greedy:64/100, at least greedy:86/100,434.8715323809997s
self.nn = nn.Sequential(
nn.Linear(self.state_size, size),
nn.ReLU(),
nn.Linear(size, 2*size),
nn.ReLU(),
#nn.Linear(2*size, 2*size),
#nn.ReLU(),
nn.Linear(2*size, size),
nn.ReLU(),
nn.Linear(size, self.action_size)
)
#beta = (0.85,0.99)
#weight_decay=1e-4
self.optimizer = optim.Adam(self.parameters(), lr=self.learning_rate, betas=(0.85, 0.99), weight_decay=1e-4)#no need cuda
self.is_training = True
self.device = device
def model(self, x):
return self.nn(x)
def act(self, obs, idx):
state = torch.tensor(obs[idx], dtype=torch.float)#.cuda()
#print(random.random())
if random.random() > self.epsilon or not self.is_training:
q_value = self.model(state)
#print(q_value.size())
action = q_value.max(0)[1].item()
else:
action = random.randrange(self.action_size)
#print(action)
return action
def remember(self, state, action, reward, next_state, done):#, batch_size):
self.buffer.add(state, action, reward, next_state, done)
#if self.memory.size() < batch_size:
#self.memory.add(state, action, reward, next_state, done)
def replay(self, batch_size):
if self.epsilon > self.epsilon_min:
self.epsilon *= self.epsilon_decay
s0, a, r, s1, done = self.buffer.sample(batch_size)
q_values = self.model(s0)
next_q_values = self.model(s1)
next_q_value = next_q_values.max(1)[0]
#print(q_values.size(),a.unsqueeze(1).size())
q_value = q_values.gather(1, a.unsqueeze(1)).squeeze(1)
expected_q_value = r + self.gamma * next_q_value * (1 - done)
# Notice that detach the expected_q_value
loss = (q_value - expected_q_value.detach()).pow(2).mean()
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
def fgsm(self, batch_size, epsilon=0.2):
if self.epsilon > self.epsilon_min:
self.epsilon *= self.epsilon_decay
s0, a, r, s1, done = self.buffer.sample(batch_size)
fs0 = s0.clone().requires_grad_().to(device)
q_values = self.model(fs0)
next_q_values = self.model(s1)
next_q_value = next_q_values.max(1)[0]
#print(q_values.size(),a.unsqueeze(1).size())
q_value = q_values.gather(1, a.unsqueeze(1)).squeeze(1)
expected_q_value = r + self.gamma * next_q_value * (1 - done)
# Notice that detach the expected_q_value
loss = (q_value - expected_q_value.detach()).pow(2).mean()
#self.optimizer.zero_grad()
loss.backward()
sg_sign = torch.sign(fs0.grad.to(fs0.device))
fake_states = s0 + epsilon * sg_sign
fake_states.detach_()
q_values = self.model(fake_states)
q_value = q_values.gather(1, a.unsqueeze(1)).squeeze(1)
loss = (q_value - expected_q_value.detach()).pow(2).mean()
self.optimizer.zero_grad()
self.optimizer.step()
def pgd(self, batch_size, eps=0.3, alpha=2/255, iters=40):
if self.epsilon > self.epsilon_min:
self.epsilon *= self.epsilon_decay
s0, a, r, s1, done = self.buffer.sample(batch_size)
fs0 = s0.clone().requires_grad_().to(device)
q_values = self.model(fs0)
next_q_values = self.model(s1)
next_q_value = next_q_values.max(1)[0]
#print(q_values.size(),a.unsqueeze(1).size())
q_value = q_values.gather(1, a.unsqueeze(1)).squeeze(1)
expected_q_value = r + self.gamma * next_q_value * (1 - done)
# Notice that detach the expected_q_value
loss = (q_value - expected_q_value.detach()).pow(2).mean()
#self.optimizer.zero_grad()
loss.backward()
ori_states = deepcopy(s0)
for _ in range(iters):
adv_states = s0 + alpha*torch.sign(fs0.grad.to(fs0.device))
eta = torch.clamp(adv_states - ori_states, min=-eps, max=eps)
s0 = torch.clamp(ori_states + eta, min=0, max=1)
#print('pass')
fake_states = s0.detach_()
q_values = self.model(fake_states)
q_value = q_values.gather(1, a.unsqueeze(1)).squeeze(1)
loss = (q_value - expected_q_value.detach()).pow(2).mean()
self.optimizer.zero_grad()
self.optimizer.step()
def save(self, checkpoint_path):
torch.save([self.nn.state_dict(),self.optimizer.state_dict()], checkpoint_path)
def load(self, checkpoint_path):
self.nn.load_state_dict(torch.load(checkpoint_path)[0])
self.optimizer.load_state_dict(torch.load(checkpoint_path)[1])