-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathqnets.py
164 lines (132 loc) · 5.58 KB
/
qnets.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
import torch
import torch.nn as nn
import torch.optim as optim
from torch.nn import functional as F
from utils import *
class QNetContinuous(nn.Module):
def __init__(self, state_dim, action_dim, learning_rate):
super().__init__()
self.fc1 = nn.Linear(state_dim, 64)
self.fc2 = nn.Linear(action_dim, 32)
self.fcn = Sequential([64 + 32, 256, 1])
self.optimizer = optim.Adam(self.parameters(), learning_rate)
def forward(self, states, actions):
x1 = torch.relu(self.fc1(states))
x2 = torch.relu(self.fc2(actions))
x = torch.cat((x1, x2), 1)
return self.fcn(x)
def train(self, states, actions, expected_values):
q_values = self.forward(states, actions)
loss = F.smooth_l1_loss(expected_values, q_values)
optimize(self.parameters(), self.optimizer, loss)
class DQNContinuous(nn.Module):
''' Double Q-Network Critic '''
def __init__(self, state_dim, action_dim, learning_rate):
super().__init__()
self.q1 = QNetContinuous(state_dim, action_dim, learning_rate)
self.q2 = QNetContinuous(state_dim, action_dim, learning_rate)
def forward(self, states, actions):
q1 = self.q1(states, actions)
q2 = self.q2(states, actions)
return torch.min(q1, q2)
def train(self, states, actions, expected_values):
self.q1.train(states, actions, expected_values)
self.q2.train(states, actions, expected_values)
def requires_grad(self, boolean = True):
for param in self.parameters():
param.requires_grad = boolean
# import jax
# import jax.numpy as jnp
# import flax.linen as nn
# class FlaxQNetContinuous(nn.Module):
# '''Single Q-Network implemented in JAX/Flax'''
# action_dim: int
# @nn.compact
# def __call__(self, states, actions):
# x1 = nn.Dense(64)(states)
# x1 = nn.relu(x1)
# x2 = nn.Dense(32)(actions)
# x2 = nn.relu(x2)
# x = jnp.concatenate([x1, x2], axis=-1)
# x = nn.Dense(256)(x)
# x = nn.relu(x)
# x = nn.Dense(1)(x)
# return x
# class FlaxDQNContinuous(nn.Module):
# '''Double Q-Network Critic implemented in JAX/Flax'''
# action_dim: int
# def setup(self):
# self.q1 = FlaxQNetContinuous(self.action_dim)
# self.q2 = FlaxQNetContinuous(self.action_dim)
# def __call__(self, states, actions):
# q1 = self.q1(states, actions)
# q2 = self.q2(states, actions)
# return jnp.minimum(q1, q2)
# def train_step(self, params, states, actions, expected_values, optimizer):
# def loss_fn(params):
# q1 = self.q1.apply({'params': params['q1']}, states, actions)
# q2 = self.q2.apply({'params': params['q2']}, states, actions)
# loss1 = jnp.mean((q1 - expected_values) ** 2)
# loss2 = jnp.mean((q2 - expected_values) ** 2)
# return loss1 + loss2
# grad_fn = jax.value_and_grad(loss_fn)
# loss, grads = grad_fn(params)
# updates, optimizer = optimizer.update(grads, optimizer)
# params = optax.apply_updates(params, updates)
# return params, optimizer, loss
class QNetDiscrete(nn.Module):
def __init__(self, state_dim, action_dim, learning_rate, extractor = None):
super().__init__()
self.action_dim = action_dim
# linear feature extractor
if extractor is None:
self.extract = nn.Linear(state_dim, 128)
self.hidden_dim = 128
# if specified, like cnn
else:
self.extract = extractor(state_dim)
self.hidden_dim = self.extract.out_dim
self.fcn = Sequential([self.hidden_dim, 256, action_dim])
self.optimizer = optim.Adam(self.parameters(), learning_rate)
def forward(self, states):
x = self.extract(states)
x = self.fcn(x)
return x
def train(self, states, actions, expected_values):
Q_out = self.forward(states)
Q_values = Q_out.gather(1, actions)
loss = F.smooth_l1_loss(expected_values, Q_values)
optimize(self.parameters(), self.optimizer, loss)
def choose_action(self, state, epsilon = 0.2):
if np.random.uniform() < epsilon:
return np.random.randint(self.action_dim)
else:
with torch.no_grad():
dist = self.forward(T(state))
action = dist.probs.argmax(1)
return np.round(action.item(), 5)
class DQNDiscrete(nn.Module):
''' Double Q-Network Critic '''
def __init__(self, state_dim, action_dim, learning_rate, extractor = None):
super().__init__()
self.action_dim = action_dim
self.q1 = QNetDiscrete(state_dim, action_dim, learning_rate, extractor)
self.q2 = QNetDiscrete(state_dim, action_dim, learning_rate, extractor)
def forward(self, states):
q1 = self.q1(states)
q2 = self.q2(states)
return torch.min(q1, q2, axis = 1)
def train(self, states, actions, expected_values):
self.q1.train(states, actions, expected_values)
self.q2.train(states, actions, expected_values)
def requires_grad(self, boolean = True):
for param in self.parameters():
param.requires_grad = boolean
def choose_action(self, state, epsilon = 0.2):
if np.random.uniform() < epsilon:
return np.random.randint(self.action_dim)
else:
with torch.no_grad():
dist = self.forward(T(state))
action = dist.probs.argmax(1)
return np.round(action.item(), 5)