-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathddpg.py
executable file
·162 lines (135 loc) · 7.4 KB
/
ddpg.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
import tensorflow as tf
import tensorflow.contrib as tc
from actor import Actor
from critic import Critic
from config import CONFIG as C
class DDPG(object):
def __init__(self, state_dim=(84, 84, 4), action_dim=3):
self.s_ph = tf.placeholder(tf.uint8, shape=(None,) + (state_dim), name='s_ph')
self.a_ph = tf.placeholder(tf.float32, shape=(None, action_dim), name='a_ph')
self.r_ph = tf.placeholder(tf.float32, shape=(None, 1), name='r_ph')
self.d_ph = tf.placeholder(tf.float32, shape=(None, 1), name='d_ph')
self.ns_ph = tf.placeholder(tf.uint8, shape=(None,) + (state_dim), name='ns_ph')
self.online_actor = Actor(action_dim, 'online')
self.online_critic = Critic(1, 'online')
self.target_actor = Actor(action_dim, 'target')
self.target_critic = Critic(1, 'target')
self.pred_a = self.online_actor(self.s_ph)
self.pred_Q = self.online_critic(self.s_ph, self.a_ph)
self.pred_Q_with_actor = self.online_critic(self.s_ph, self.pred_a)
self.pred_next_a = self.target_actor(self.ns_ph)
self.pred_next_Q = self.target_critic(self.ns_ph, self.pred_next_a)
# Set up update target networks
self.init_ops, self.update_ops = self._target_updates(C.tau)
# Set up actor optimization
actor_optimizer = tf.train.AdamOptimizer(0.0001)
with tf.name_scope('actor_loss'):
actor_loss = -tf.reduce_mean(self.pred_Q_with_actor)
actor_grads = tf.gradients(actor_loss, self.online_actor.trainable_vars())
if C.clip_norm is not None:
with tf.name_scope('grads_clip'):
with tf.name_scope('actor'):
actor_grads = [tf.clip_by_norm(grad, C.clip_norm) for grad in actor_grads]
assert len(actor_grads) == len(self.online_actor.trainable_vars())
self.actor_optimize_op = actor_optimizer.apply_gradients(zip(actor_grads, self.online_actor.trainable_vars()))
# Set up critic optimization
critic_optimizer = tf.train.AdamOptimizer(0.001)
with tf.name_scope('critic_loss'):
target_Q = self.r_ph + (1.0 - self.d_ph)*C.discount_factor*self.pred_next_Q
critic_loss = tf.reduce_mean(tf.square(self.pred_Q - tf.stop_gradient(target_Q)))
if C.critic_l2_reg > 0.0:
critic_reg_vars = [v for v in self.online_critic.trainable_vars() if 'pred_Q' not in v.name]
critic_reg = tc.layers.apply_regularization(tc.layers.l2_regularizer(C.critic_l2_reg), weights_list=critic_reg_vars)
critic_loss += critic_reg
critic_grads = tf.gradients(critic_loss, self.online_critic.trainable_vars())
if C.clip_norm is not None:
with tf.name_scope('grads_clip/'):
with tf.name_scope('critic'):
critic_grads = [tf.clip_by_norm(grad, C.clip_norm) for grad in critic_grads]
assert len(critic_grads) == len(self.online_critic.trainable_vars())
self.critic_optimize_op = critic_optimizer.apply_gradients(zip(critic_grads, self.online_critic.trainable_vars()))
def _target_updates(self, tau):
online_actor_vars = self.online_actor.global_vars()
online_critic_vars = self.online_critic.global_vars()
target_actor_vars = self.target_actor.global_vars()
target_critic_vars = self.target_critic.global_vars()
assert len(online_actor_vars) == len(target_actor_vars)
assert len(online_critic_vars) == len(target_critic_vars)
soft_ops = []
init_ops = []
with tf.name_scope('updates'):
with tf.name_scope('actor'):
for o, t in zip(online_actor_vars, target_actor_vars):
init_ops.append(t.assign(o))
soft_ops.append(t.assign((1.0 - tau)*t + tau*o))
with tf.name_scope('critic'):
for o, t in zip(online_critic_vars, target_critic_vars):
init_ops.append(t.assign(o))
soft_ops.append(t.assign((1.0 - tau)*t + tau*o))
with tf.name_scope('group_ops'):
io = tf.group(*init_ops)
so = tf.group(*soft_ops)
assert len(init_ops) == len(soft_ops)
return io, so
def initialize_target(self):
tf.get_default_session().run(self.init_ops)
def update_target(self):
tf.get_default_session().run(self.update_ops)
def action(self, s):
if s.ndim == 3:
s = [s]
return tf.get_default_session().run(self.pred_a, feed_dict={self.s_ph: s})
def train_critic(self, sb, ab, rb, db, nsb):
tf.get_default_session().run(self.critic_optimize_op, feed_dict={self.s_ph: sb, self.a_ph: ab, self.r_ph: rb, self.d_ph: db, self.ns_ph: nsb})
def train_actor(self, sb):
tf.get_default_session().run(self.actor_optimize_op, feed_dict={self.s_ph: sb})
def train(self, sb, ab, rb, db, nsb):
tf.get_default_session().run([self.critic_optimize_op, self.actor_optimize_op], feed_dict={self.s_ph: sb, self.a_ph: ab, self.r_ph: rb, self.d_ph: db, self.ns_ph: nsb})
def save(self, directory, i):
saver = tf.train.Saver()
d = saver.save(tf.get_default_session(), f'{directory}/model_{i}.ckpt')
return d
def load(self, directory, i):
saver = tf.train.Saver()
saver.restore(tf.get_default_session(), f'{directory}/model_{i}.ckpt')
if __name__ == '__main__':
import os
import numpy as np
os.system('rm -rf test/') # Clean up
sess = tf.InteractiveSession()
ddpg = DDPG()
sess.run(tf.global_variables_initializer())
writer = tf.summary.FileWriter('./test', sess.graph)
sb = (np.random.rand(32, 84, 84, 4)*255).astype('uint8')
ab = np.random.rand(32, 3)
rb = np.random.randn(32, 1)
db = np.random.randint(2, size=(32, 1))
nsb = (np.random.rand(32, 84, 84, 4)*255).astype('uint8')
ddpg.initialize_target()
for o, t in zip(ddpg.online_actor.global_vars(), ddpg.target_actor.global_vars()):
assert np.allclose(o.eval(), t.eval())
for o, t in zip(ddpg.online_critic.global_vars(), ddpg.target_critic.global_vars()):
assert np.allclose(o.eval(), t.eval())
print(ddpg.action(sb[0]))
print(ddpg.action(sb[:3]))
ddpg.train_critic(sb, ab, rb, db, nsb)
t0a = []
t0c = []
for o, t in zip(ddpg.online_actor.trainable_vars(), ddpg.target_actor.trainable_vars()):
t0a.append(t.eval())
assert np.allclose(o.eval(), t.eval())
for o, t in zip(ddpg.online_critic.trainable_vars(), ddpg.target_critic.trainable_vars()):
t0c.append(t.eval())
assert not np.allclose(o.eval(), t.eval())
ddpg.train_actor(sb)
for o, t in zip(ddpg.online_actor.trainable_vars(), ddpg.target_actor.trainable_vars()):
assert not np.allclose(o.eval(), t.eval())
# for o, t in zip(ddpg.online_critic.trainable_vars(), ddpg.target_critic.trainable_vars()):
# assert not np.allclose(o.eval(), t.eval())
ddpg.update_target()
for i, (o, t) in enumerate(zip(ddpg.online_actor.trainable_vars(), ddpg.target_actor.trainable_vars())):
assert np.allclose(t0a[i], (t.eval() - C.tau*o.eval())/(1 - C.tau))
assert not np.allclose(o.eval(), t.eval())
for i, (o, t) in enumerate(zip(ddpg.online_critic.trainable_vars(), ddpg.target_critic.trainable_vars())):
assert np.allclose(t0c[i], (t.eval() - C.tau*o.eval())/(1 - C.tau))
assert not np.allclose(o.eval(), t.eval())