-
Notifications
You must be signed in to change notification settings - Fork 7
/
Copy pathrlmodule.py
228 lines (177 loc) · 6.48 KB
/
rlmodule.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
# -*- coding: utf-8 -*-
"""
@author: truthless
"""
import torch
import torch.nn as nn
import numpy as np
from collections import namedtuple
import random
class DiscretePolicy(nn.Module):
def __init__(self, cfg):
super(DiscretePolicy, self).__init__()
self.net = nn.Sequential(nn.Linear(cfg.s_dim, cfg.h_dim),
nn.ReLU(),
nn.Linear(cfg.h_dim, cfg.h_dim),
nn.ReLU(),
nn.Linear(cfg.h_dim, cfg.a_dim))
def forward(self, s):
# [b, s_dim] => [b, a_dim]
a_weights = self.net(s)
return a_weights
def select_action(self, s, sample=True):
"""
:param s: [s_dim]
:return: [1]
"""
# forward to get action probs
# [s_dim] => [a_dim]
a_weights = self.forward(s)
a_probs = torch.softmax(a_weights, 0)
# randomly sample from normal distribution, whose mean and variance come from policy network.
# [a_dim] => [1]
a = a_probs.multinomial(1) if sample else a_probs.argmax(0, True)
return a
def get_log_prob(self, s, a):
"""
:param s: [b, s_dim]
:param a: [b, 1]
:return: [b, 1]
"""
# forward to get action probs
# [b, s_dim] => [b, a_dim]
a_weights = self.forward(s)
a_probs = torch.softmax(a_weights, -1)
# [b, a_dim] => [b, 1]
trg_a_probs = a_probs.gather(-1, a)
log_prob = torch.log(trg_a_probs)
return log_prob
class MultiDiscretePolicy(nn.Module):
def __init__(self, cfg):
super(MultiDiscretePolicy, self).__init__()
self.net = nn.Sequential(nn.Linear(cfg.s_dim, cfg.h_dim),
nn.ReLU(),
nn.Linear(cfg.h_dim, cfg.h_dim),
nn.ReLU(),
nn.Linear(cfg.h_dim, cfg.a_dim))
def forward(self, s):
# [b, s_dim] => [b, a_dim]
a_weights = self.net(s)
return a_weights
def select_action(self, s, sample=True):
"""
:param s: [s_dim]
:return: [a_dim]
"""
# forward to get action probs
# [s_dim] => [a_dim]
a_weights = self.forward(s)
a_probs = torch.sigmoid(a_weights)
# [a_dim] => [a_dim, 2]
a_probs = a_probs.unsqueeze(1)
a_probs = torch.cat([1-a_probs, a_probs], 1)
# [a_dim, 2] => [a_dim]
a = a_probs.multinomial(1).squeeze(1) if sample else a_probs.argmax(1)
return a
def get_log_prob(self, s, a):
"""
:param s: [b, s_dim]
:param a: [b, a_dim]
:return: [b, 1]
"""
# forward to get action probs
# [b, s_dim] => [b, a_dim]
a_weights = self.forward(s)
a_probs = torch.sigmoid(a_weights)
# [b, a_dim] => [b, a_dim, 2]
a_probs = a_probs.unsqueeze(-1)
a_probs = torch.cat([1-a_probs, a_probs], -1)
# [b, a_dim, 2] => [b, a_dim]
trg_a_probs = a_probs.gather(-1, a.unsqueeze(-1)).squeeze(-1)
log_prob = torch.log(trg_a_probs)
return log_prob.sum(-1, keepdim=True)
class ContinuousPolicy(nn.Module):
def __init__(self, cfg):
super(ContinuousPolicy, self).__init__()
self.net = nn.Sequential(nn.Linear(cfg.s_dim, cfg.h_dim),
nn.ReLU(),
nn.Linear(cfg.h_dim, cfg.h_dim),
nn.ReLU())
self.net_mean = nn.Linear(cfg.h_dim, cfg.a_dim)
self.net_std = nn.Linear(cfg.h_dim, cfg.a_dim)
def forward(self, s):
# [b, s_dim] => [b, h_dim]
h = self.net(s)
# [b, h_dim] => [b, a_dim]
a_mean = self.net_mean(h)
a_log_std = self.net_std(h)
return a_mean, a_log_std
def select_action(self, s, sample=True):
"""
:param s: [s_dim]
:return: [a_dim]
"""
# forward to get action mean and log_std
# [s_dim] => [a_dim]
a_mean, a_log_std = self.forward(s)
# randomly sample from normal distribution, whose mean and variance come from policy network.
# [a_dim]
a = torch.normal(a_mean, a_log_std.exp()) if sample else a_mean
return a
def get_log_prob(self, s, a):
"""
:param s: [b, s_dim]
:param a: [b, a_dim]
:return: [b, 1]
"""
def normal_log_density(x, mean, log_std):
"""
x ~ N(mean, std)
this function will return log(prob(x)) while x belongs to guassian distrition(mean, std)
:param x: [b, a_dim]
:param mean: [b, a_dim]
:param log_std: [b, a_dim]
:return: [b, 1]
"""
std = log_std.exp()
var = std.pow(2)
log_density = - (x - mean).pow(2) / (2 * var) - 0.5 * np.log(2 * np.pi) - log_std
return log_density.sum(-1, keepdim=True)
# forward to get action mean and log_std
# [b, s_dim] => [b, a_dim]
a_mean, a_log_std = self.forward(s)
# [b, a_dim] => [b, 1]
log_prob = normal_log_density(a, a_mean, a_log_std)
return log_prob
class Value(nn.Module):
def __init__(self, cfg):
super(Value, self).__init__()
self.net = nn.Sequential(nn.Linear(cfg.s_dim, cfg.hv_dim),
nn.ReLU(),
nn.Linear(cfg.hv_dim, cfg.hv_dim),
nn.ReLU(),
nn.Linear(cfg.hv_dim, 1))
def forward(self, s):
"""
:param s: [b, s_dim]
:return: [b, 1]
"""
value = self.net(s)
return value
Transition = namedtuple('Transition', ('state', 'action', 'mask', 'next_state'))
class Memory(object):
def __init__(self):
self.memory = []
def push(self, *args):
"""Saves a transition."""
self.memory.append(Transition(*args))
def get_batch(self, batch_size=None):
if batch_size is None:
return Transition(*zip(*self.memory))
else:
random_batch = random.sample(self.memory, batch_size)
return Transition(*zip(*random_batch))
def append(self, new_memory):
self.memory += new_memory.memory
def __len__(self):
return len(self.memory)