-
Notifications
You must be signed in to change notification settings - Fork 0
/
dqn.py
153 lines (120 loc) · 5.11 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
import torch
from torch.optim.lr_scheduler import LambdaLR
import numpy as np
from memory import ReplayBuffer
def lr_lambda(epoch):
if epoch < 20000:
return 3e-4
else:
return 1e-4
class ActionValue(torch.nn.Module):
def __init__(self, input_shape, n_actions, alpha=3e-4, chkpt_file="weights/dqn.pt"):
super(ActionValue, self).__init__()
self.chkpt_file = chkpt_file
self.conv1 = torch.nn.Conv2d(input_shape[0], 32, kernel_size=8, stride=4)
self.conv2 = torch.nn.Conv2d(32, 64, kernel_size=4, stride=2)
self.conv3 = torch.nn.Conv2d(64, 64, kernel_size=3, stride=1)
self.fc1_input_dim = self._calculate_fc1_input_dim(input_shape)
self.fc1 = torch.nn.Linear(self.fc1_input_dim, 512)
self.out = torch.nn.Linear(512, n_actions)
self.optimizer = torch.optim.AdamW(self.parameters(), lr=alpha)
self.scheduler = LambdaLR(self.optimizer, lr_lambda)
self.loss = torch.nn.MSELoss() # use squared l1 instead of mse?
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
self.to(self.device)
self._initialize_weights()
def forward(self, x):
x = torch.nn.functional.relu(self.conv1(x))
x = torch.nn.functional.relu(self.conv2(x))
x = torch.nn.functional.relu(self.conv3(x))
x = x.view(x.size()[0], -1)
x = torch.nn.functional.relu(self.fc1(x))
return self.out(x)
def save_checkpoint(self):
torch.save(self.state_dict(), self.chkpt_file)
def load_checkpoint(self):
self.load_state_dict(torch.load(self.chkpt_file))
def _calculate_fc1_input_dim(self, input_shape):
dummy_input = torch.zeros(1, *input_shape)
x = torch.nn.functional.relu(self.conv1(dummy_input))
x = torch.nn.functional.relu(self.conv2(x))
x = torch.nn.functional.relu(self.conv3(x))
return x.numel()
def _initialize_weights(self):
for m in self.modules():
if isinstance(m, torch.nn.Conv2d) or isinstance(m, torch.nn.Linear):
torch.nn.init.orthogonal_(m.weight)
if isinstance(m, torch.nn.Linear):
m.weight.data.mul_(1 / 100)
if m.bias is not None:
torch.nn.init.constant_(m.bias, 0)
class DQNAgent:
def __init__(
self,
env_name,
input_shape,
n_actions,
alpha=3e-4,
gamma=0.99,
eps_min=0.1,
eps_dec=5e-7,
batch_size=64,
mem_size=100000,
replace_target_count=1000,
):
self.gamma = gamma
self.epsilon = 1.0
self.eps_min = eps_min
self.eps_dec = eps_dec
self.n_actions = n_actions
self.batch_size = batch_size
self.replace_target_count = replace_target_count
self.counter = 0
self.memory = ReplayBuffer(input_shape, int(mem_size), batch_size)
self.q = ActionValue(
input_shape, n_actions, alpha, f"weights/{env_name}_dqn.pt"
)
self.target_q = ActionValue(
input_shape, n_actions, alpha, f"weights/{env_name}_target_dqn.pt"
)
def choose_action(self, state):
if np.random.random() > self.epsilon:
state = torch.FloatTensor(state).unsqueeze(0).to(self.q.device)
actions = self.q(state)
return torch.argmax(actions).item()
return np.random.randint(0, self.n_actions)
def store_transition(self, state, action, reward, next_state, done):
self.memory.store_transition(state, action, reward, next_state, done)
def learn(self):
if self.memory.mem_counter < self.batch_size:
return
if self.counter % self.replace_target_count == 0:
self.update_target_parameters()
states, actions, rewards, next_states, dones = self.memory.sample()
states = torch.FloatTensor(states).to(self.q.device)
actions = torch.IntTensor(actions).to(self.q.device)
next_states = torch.FloatTensor(next_states).to(self.q.device)
rewards = torch.FloatTensor(rewards).to(self.q.device)
dones = torch.BoolTensor(dones).to(self.q.device)
self.q.optimizer.zero_grad()
# get Q value for chosen actions, need np.arange for proper indexing
q_pred = self.q(states)[np.arange(self.batch_size), actions]
# max returns tuple of max_val, index
target_vals = self.target_q(next_states).max(dim=1)[0]
target_vals[dones] = 0.0
q_target = rewards + self.gamma * target_vals
loss = self.q.loss(q_target, q_pred).to(self.q.device)
loss.backward()
self.q.optimizer.step()
self.counter += 1
self.decrement_epsilon()
def decrement_epsilon(self):
self.epsilon = max(self.eps_min, self.epsilon - self.eps_dec)
def update_target_parameters(self):
self.target_q.load_state_dict(dict(self.q.named_parameters()))
def save_checkpoint(self):
self.q.save_checkpoint()
self.target_q.save_checkpoint()
def load_checkpoint(self):
self.q.load_checkpoint()
self.target_q.load_checkpoint()