-
Notifications
You must be signed in to change notification settings - Fork 2
/
ppo.py
208 lines (161 loc) · 9.67 KB
/
ppo.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
import os
import glob
from datetime import datetime
import numpy as np
import tensorflow as tf
from tensorflow.python.keras.layers import Dense, Dropout
import gym
from distributions import Categorical, Normal
from utils import save_target_graph, restore_target_graph
class Agent:
def __init__(self, env, actor_net, critic_net, ent_coef=0.001, cliprange=0.1, max_grad_norm=None,
saver_max_to_keep=10):
def get_space_shape(space):
try:
return None, space.shape[0]
except IndexError:
return None,
self.max_to_keep = saver_max_to_keep
# Environment parameters
self.act_space = env.action_space
self.act_shape = get_space_shape(self.act_space)
self.obs_shape = get_space_shape(env.observation_space)
self.actor_net = actor_net
self.critic_net = critic_net
# Reset the graph
tf.reset_default_graph()
# Init
self.states = tf.placeholder(tf.float32, shape=self.obs_shape, name='states')
self.actions_old = tf.placeholder(tf.float32, shape=self.act_shape, name='actions_old')
self.values_old = tf.placeholder(tf.float32, shape=(None,), name='values_old')
self.neglogps_old = tf.placeholder(tf.float32, shape=(None,), name='neglogps_old')
self.gaes = tf.placeholder(tf.float32, shape=(None,), name='advantage')
self.q_values = tf.placeholder(tf.float32, shape=(None,), name='estimation')
self.drop_rate = tf.Variable(0.0, dtype=tf.float32, trainable=False)
self.actor_lr = tf.placeholder(tf.float32, name='actor_lr')
self.critic_lr = tf.placeholder(tf.float32, name='critic_lr')
# Build the agent
with tf.variable_scope('agent') as scope:
self.action_distrs = self.actor(self.states)
self.values = self.critic(self.states)
policy_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope.name)
# Sample actions from the given distribution
self.actions = self.action_distrs.sample()
self.neglogp = self.action_distrs.neglogp(self.actions)
self.neglogp_new = self.action_distrs.neglogp(self.actions_old)
try:
self.actions = tf.clip_by_value(self.actions, self.act_space.low, self.act_space.high)
except AttributeError:
pass
"""Losses"""
with tf.variable_scope('critic_loss'):
self.values_cliped = self.values_old + \
tf.clip_by_value(self.values - self.values_old, -cliprange, cliprange)
critic_loss = tf.square(self.q_values - self.values)
critic_loss_clipped = tf.square(self.q_values - self.values_cliped)
self.critic_loss = tf.reduce_mean(tf.maximum(critic_loss, critic_loss_clipped))
with tf.variable_scope('actor_loss'):
ratio = tf.exp(self.neglogps_old - self.neglogp_new)
actor_loss = self.gaes * ratio
actor_loss_clipped = self.gaes * tf.clip_by_value(ratio, 1.0 - cliprange, 1.0 + cliprange)
self.actor_loss = -tf.reduce_mean(tf.minimum(actor_loss, actor_loss_clipped))
with tf.variable_scope('entropy_loss'):
entropy = self.action_distrs.entropy()
try:
self.entropy_loss = tf.reduce_mean(entropy, axis=0)
except:
self.entropy_loss = entropy
with tf.variable_scope('total_loss'):
self.loss = self.critic_loss + self.actor_loss - self.entropy_loss*ent_coef
# Define all trainable variables
self.train_vars = policy_vars
# Calculate the gradients
def get_grads(loss):
grads = tf.gradients(loss, self.train_vars)
if max_grad_norm is not None:
grads, grad_norm = tf.clip_by_global_norm(grads, max_grad_norm)
return list(zip(grads, self.train_vars))
# Backpropagate
self.train_op_critic = tf.train.AdamOptimizer(self.critic_lr).apply_gradients(get_grads(self.critic_loss))
self.train_op_actor = tf.train.AdamOptimizer(self.actor_lr).apply_gradients(
get_grads(self.actor_loss - self.entropy_loss*ent_coef))
@staticmethod
def restore_weights(sess, path=None, lastpath=None):
"""Restore the weights if they exist"""
try:
if lastpath is None:
paths = glob.glob(path + '/*.npy')
lastpath = paths[-1]
restore_target_graph(sess, lastpath)
print('******** The model was restored! *********')
print(lastpath)
except:
print("******** The model not found! *********")
def save_weights(self, sess, path, global_step=0):
index = datetime.now().strftime('%Y.%m.%d-%H.%M.%S')
# Save weights
save_target_graph(sess, path, '/weights_{}_{:04d}'.format(index, global_step))
print('******** The model was saved! *********')
# Remove the earlier files to keep their total number equal to the `max_to_keep`
remove_paths = glob.glob(path + '/*.npy')[:-self.max_to_keep]
# Lazy way to loop over files and remove them
_ = [*map(os.remove, remove_paths)]
def get_actions_distribution(self, sess, samples, keep_prob):
return sess.run(self.action_distrs, {self.states: samples, self.drop_rate: keep_prob})
def get_value(self, sess, states, drop_rate):
return sess.run(self.values, {self.states: states, self.drop_rate: drop_rate})
# TODO: Add deterministic policy
def get_action(self, sess, states, drop_rate=0.0, stochastic=False):
# # Sample actions from the given distribution
# self.action = self.actions_distrs.sample(stochastic)
#
# # Clip range of the action if allowed
# try:
# self.action = tf.clip_by_value(self.action, self.act_space.low, self.act_space.high)
# except AttributeError:
# pass
return sess.run(self.actions, {self.states: states, self.drop_rate: drop_rate})
def evaluate_model(self, sess, samples):
return sess.run([self.actions, self.values, self.neglogp], feed_dict={self.states: samples})
def train_agent(self, sess, states, actions, values, neglogps, gaes, q_values):
return sess.run([self.loss, self.train_op], feed_dict={self.states: states,
self.actions_old: actions,
self.values_old: values,
self.neglogps_old: neglogps,
self.gaes: gaes,
self.q_values: q_values})
def train_actor(self, sess, states, actions, values, neglogps, gaes, q_values, lr, drop_rate):
return sess.run([self.actor_loss, self.train_op_actor], feed_dict={self.states: states,
self.actions_old: actions,
self.values_old: values,
self.neglogps_old: neglogps,
self.gaes: gaes,
self.q_values: q_values,
self.actor_lr: lr,
self.drop_rate: drop_rate})
def train_critic(self, sess, states, actions, values, neglogps, gaes, q_values, lr, drop_rate):
return sess.run([self.critic_loss, self.train_op_critic], feed_dict={self.states: states,
self.actions_old: actions,
self.values_old: values,
self.neglogps_old: neglogps,
self.gaes: gaes,
self.q_values: q_values,
self.critic_lr: lr,
self.drop_rate: drop_rate})
def actor(self, states, name='actor', reuse=False, trainable=True):
with tf.variable_scope(name, reuse=reuse):
features = self.actor_net(states, self.drop_rate, trainable=trainable)
if isinstance(self.act_space, gym.spaces.Discrete):
logits = Dense(self.act_space.n, None, trainable=trainable, name="layer_logits")(features)
distribution = Categorical(logits)
else:
mean = Dense(self.act_space.shape[0], None, trainable=trainable, name='mean')(features)
logstd = tf.get_variable('logstd', initializer=-0.5*np.ones(self.act_space.shape[0], dtype=np.float32))
# logstd = Dense(self.act_space.shape[0], None, trainable=trainable, name='logstd')(features)
distribution = Normal(mean=mean, logstd=logstd)
return distribution
def critic(self, states, name='critic', reuse=False, trainable=True):
with tf.variable_scope(name, reuse=reuse):
features = self.critic_net(states, self.drop_rate, trainable=trainable)
value = Dense(1, None, trainable=trainable, name="layer_logits")(features)
return tf.squeeze(value, axis=1)