-
Notifications
You must be signed in to change notification settings - Fork 0
/
td3.py
237 lines (190 loc) · 8.64 KB
/
td3.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
import chainer
import chainer.functions as F
from chainer import optimizers
from chainer import iterators
from chainer.datasets import tuple_dataset
from chainer.dataset import concat_examples
from chainer.distributions import Normal
from collections import deque
from models.td3_actor import TD3Actor
from models.td3_critic import TD3Critic
import gym
import numpy as np
# import cupy as cp
class TD3(object):
def __init__(self, state_dim, action_num, lr=1.0*1e-3, batch_size=100, device=-1):
super(TD3, self).__init__()
self._q1_optimizer = optimizers.Adam(alpha=lr)
self._q2_optimizer = optimizers.Adam(alpha=lr)
self._pi_optimizer = optimizers.Adam(alpha=lr)
self._batch_size = batch_size
self._q1 = TD3Critic(state_dim=state_dim, action_num=action_num)
self._q2 = TD3Critic(state_dim=state_dim, action_num=action_num)
self._pi = TD3Actor(state_dim=state_dim, action_num=action_num)
self._target_q1 = TD3Critic(state_dim=state_dim, action_num=action_num)
self._target_q2 = TD3Critic(state_dim=state_dim, action_num=action_num)
self._target_pi = TD3Actor(state_dim=state_dim, action_num=action_num)
if not device < 0:
self._q1.to_gpu()
self._q2.to_gpu()
self._pi.to_gpu()
self._target_q1.to_gpu()
self._target_q2.to_gpu()
self._target_pi.to_gpu()
self._q1_optimizer.setup(self._q1)
self._q2_optimizer.setup(self._q2)
self._pi_optimizer.setup(self._pi)
xp = np if device < 0 else cp
mean = xp.zeros(shape=(action_num), dtype=xp.float32)
sigma = xp.ones(shape=(action_num), dtype=np.float32)
self._exploration_noise = Normal(loc=mean, scale=sigma * 0.1)
self._action_noise = Normal(loc=mean, scale=sigma * 0.2)
self._device = device
self._initialized = False
self._action_num = action_num
def act_with_policy(self, env, s):
s = np.float32(s)
state = chainer.Variable(np.reshape(s, newshape=(1, ) + s.shape))
if not self._device < 0:
state.to_gpu()
a = self._pi(state)
if not self._device < 0:
a.to_cpu()
noise = self._sample_exploration_noise(shape=(1))
# print('a shape:', a.shape, ' noise shape: ', noise.shape)
assert a.shape == noise.shape
a = np.squeeze((a + noise).data, axis=0)
s_next, r, done, _ = env.step(a)
s_next = np.float32(s_next)
a = np.float32(a)
r = np.float32(r)
return s, a, r, s_next, done
def act_randomly(self, env, s):
s = np.float32(s)
a = env.action_space.sample()
s_next, r, done, _ = env.step(a)
s_next = np.float32(s_next)
a = np.float32(a)
r = np.float32(r)
return s, a, r, s_next, done
def evaluate_policy(self, env, render=False, save_video=False):
if save_video:
from OpenGL import GL
env = gym.wrappers.Monitor(env, directory='video',
write_upon_reset=True, force=True, resume=True, mode='evaluation')
render=False
rewards = []
for _ in range(10):
s = env.reset()
episode_reward = 0
while True:
if render:
env.render()
s = np.float32(s)
s = chainer.Variable(np.reshape(s, newshape=(1, ) + s.shape))
if not self._device < 0:
s.to_gpu()
a = self._pi(s)
if not self._device < 0:
a.to_cpu()
a = np.squeeze(a.data, axis=0)
s, r, done, _ = env.step(a)
episode_reward += r
if done:
rewards.append(episode_reward)
break
return rewards
def train(self, replay_buffer, iterations, d, clip_value, gamma, tau):
if not self._initialized:
self._initialize_target_networks()
iterator = self._prepare_iterator(replay_buffer)
for i in range(iterations):
batch = iterator.next()
s_current, action, r, s_next, non_terminal = concat_examples(
batch, device=self._device)
epsilon = F.clip(self._sample_action_noise(shape=(self._batch_size)),
-clip_value, clip_value)
target_pi = self._target_pi(s_current)
assert target_pi.shape == epsilon.shape
a_tilde = target_pi + epsilon
target_q1 = self._target_q1(s_next, a_tilde)
target_q2 = self._target_q2(s_next, a_tilde)
r = F.reshape(r, shape=(*r.shape, 1))
non_terminal = F.reshape(
non_terminal, shape=(*non_terminal.shape, 1))
min_q = F.minimum(target_q1, target_q2)
# print('r shape: ', r.shape)
# print('done shape: ', non_terminal.shape)
# print('min q shape: ', min_q.shape)
y = r + gamma * non_terminal * min_q
# print('y shape: ', y.shape)
# Remove reference to avoid unexpected gradient update
y.unchain()
q1 = self._q1(s_current, action)
q1_loss = F.mean_squared_error(y, q1)
q2 = self._q2(s_current, action)
q2_loss = F.mean_squared_error(y, q2)
critic_loss = q1_loss + q2_loss
self._q1_optimizer.target.cleargrads()
self._q2_optimizer.target.cleargrads()
critic_loss.backward()
critic_loss.unchain_backward()
self._q1_optimizer.update()
self._q2_optimizer.update()
if i % d == 0:
a = self._pi(s_current)
q1 = self._q1(s_current, a)
pi_loss = -F.mean(q1)
self._pi_optimizer.target.cleargrads()
pi_loss.backward()
pi_loss.unchain_backward()
self._pi_optimizer.update()
self._update_target_network(self._target_q1, self._q1, tau)
self._update_target_network(self._target_q2, self._q2, tau)
self._update_target_network(self._target_pi, self._pi, tau)
def _initialize_target_networks(self):
self._update_target_network(self._target_q1, self._q1, 1.0)
self._update_target_network(self._target_q2, self._q2, 1.0)
self._update_target_network(self._target_pi, self._pi, 1.0)
self._initialized = True
def _update_target_network(self, target, origin, tau):
for target_param, origin_param in zip(target.params(), origin.params()):
target_param.data = tau * origin_param.data + \
(1.0 - tau) * target_param.data
def _sample_action_noise(self, shape):
return self._action_noise.sample(shape)
def _sample_exploration_noise(self, shape):
return self._exploration_noise.sample(shape)
def _prepare_iterator(self, buffer):
return iterators.SerialIterator(buffer, self._batch_size)
if __name__ == "__main__":
state_dim = 5
action_num = 5
batch_size = 100
td3 = TD3(state_dim=state_dim, action_num=action_num,
batch_size=batch_size)
a_noise = td3._sample_action_noise(shape=(batch_size))
print('action noise shape: ', a_noise.shape, ' noise: ', a_noise)
mean = np.mean(a_noise.array)
var = np.var(a_noise.array)
print('mean: ', mean, ' sigma: ', np.sqrt(var))
assert a_noise.shape == (td3._batch_size, action_num)
e_noise = td3._sample_action_noise(shape=(1))
print('exploration noise shape: ', e_noise.shape, ' noise: ', e_noise)
mean = np.mean(e_noise.array)
var = np.var(e_noise.array)
print('mean: ', mean, ' sigma: ', np.sqrt(var))
assert a_noise.shape == (1, action_num)
for target_param, origin_param in zip(td3._target_pi.params(), td3._pi.params()):
print('before target param shape: ', target_param.shape,
' origin param shape: ', origin_param.shape)
print('target: ', target_param.data, ' origin: ', origin_param.data)
is_equal = np.array_equal(target_param.data, origin_param.data)
print('is target and origin equal?: ', is_equal)
td3._update_target_network(td3._target_pi, td3._pi, tau=0.1)
for target_param, origin_param in zip(td3._target_pi.params(), td3._pi.params()):
print('after target param shape: ', target_param.shape,
' origin param shape: ', origin_param.shape)
print('target: ', target_param.data, ' origin: ', origin_param.data)
is_equal = np.array_equal(target_param.data, origin_param.data)
print('is target and origin equal?: ', is_equal)