-
Notifications
You must be signed in to change notification settings - Fork 2
/
ppo.py
257 lines (230 loc) · 9.37 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
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
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
# Proximal Policy Optimization
# https://arxiv.org/abs/1707.06347
# https://www.52coding.com.cn/2018/11/25/RL%20-%20PPO/
import gym
import numpy as np
import tensorflow as tf
from logging import getLogger
from collections import defaultdict
from concurrent.futures import ThreadPoolExecutor, as_completed
from src.base import BaseController
from src.util import discount_cumsum, mlp, cnn
from src.config import Config, ControllerType
logger = getLogger(__name__)
class PPOControl(BaseController):
def __init__(self, env, config: Config):
self.env = env
self.epsilon = config.controller.epsilon # clip ratio
self.gamma = config.controller.gamma
self.lam = config.controller.lambda_
self.pi_lr = config.trainer.lr # 1e-4
self.v_lr = 1e-3
self.max_workers = config.controller.max_workers
tfconfig = tf.ConfigProto(
gpu_options=tf.GPUOptions(
allow_growth=True,
visible_device_list='0'
)
)
self.sess = tf.Session(config=tfconfig)
self.raw_pixels = config.controller.raw_pixels
if self.raw_pixels:
state_space = [84, 84, 2]
else:
state_space = self.env.observation_space.shape
self.actor = PPOActor(self.sess, state_space, self.env.action_space.n,
self.pi_lr, self.epsilon, self.raw_pixels)
self.critic = PPOCritic(self.sess, state_space,
self.v_lr, self.raw_pixels)
self.build_model()
def build_model(self):
self.actor.build_model()
self.critic.build_model()
self.sess.run(tf.global_variables_initializer())
def action(self, observation, predict=False, return_q=False, epsilon=None, return_logp=True):
if return_q:
v = self.critic.value_of(observation)
return self.actor.action(observation), [v]
return self.actor.action(observation)[0]
def train(self, batch_buffers, i):
'''Update parameters
Args:
batch_buffers = [buf1, buf2, ...]
'''
batch_states = []
batch_actions = []
batch_rets = []
batch_advs = []
batch_logp_old = []
total_rewards = []
with ThreadPoolExecutor(max_workers=self.max_workers) as executor:
futures = [executor.submit(self.build_training_set, buf)
for buf in batch_buffers]
for future in futures:
states, actions, returns, advantages, logps, r = future.result()
batch_states.extend(states)
batch_actions.extend(actions)
batch_rets.extend(returns)
batch_advs.extend(advantages)
batch_logp_old.extend(logps)
total_rewards.append(r)
self.actor.train(batch_states, batch_actions,
batch_advs, batch_logp_old, np.mean(total_rewards), i)
self.critic.train(batch_states, batch_rets, i)
def build_training_set(self, buf):
rewards_to_go = discount_cumsum(buf.rewards, self.gamma)
buf.values = np.array(buf.values)
# compute GAE
deltas = buf.rewards[:-1] + self.gamma * \
buf.values[1:] - buf.values[:-1]
advs = discount_cumsum(deltas, self.gamma * self.lam)
# advantage normalization trick
adv_mean, adv_std = np.mean(advs), np.std(advs)
advs = (advs - adv_mean) / adv_std
# squeeze
actions = np.squeeze(buf.actions)
logps = np.squeeze(buf.logps)
if self.raw_pixels:
states = np.array(buf.states[:-1])
states = np.reshape(states, (-1, 84, 84, 2))
else:
states = buf.states[:-1]
return states, actions[:-1], rewards_to_go[:-1], advs, logps[:-1], sum(buf.rewards)
def save(self, path):
saver = tf.train.Saver()
save_path = saver.save(self.sess, path)
logger.info(f"Save weight to {save_path}")
def load(self, path):
try:
saver = tf.train.Saver()
saver.restore(self.sess, path)
logger.info(f"Load weight from {path}")
except Exception as e:
logger.error(e)
class PPOActor:
def __init__(self, sess, n_features, n_actions, lr, epsilon, raw_pixels):
self.sess = sess
if raw_pixels:
self.n_features = n_features
else:
self.n_features = n_features[0]
self.n_actions = n_actions
self.lr = lr
self.epsilon = epsilon
self.train_policy_iter = 80
self.target_kl = 0.01
self.raw_pixels = raw_pixels
def build_model(self):
clip_ratio = self.epsilon
# Input placeholder
if self.raw_pixels:
self.s_ph = tf.placeholder(tf.float32, [None] + self.n_features)
else:
self.s_ph = tf.placeholder(tf.float32, [None, self.n_features])
self.a_ph = tf.placeholder(tf.int32, [None])
self.logp_old_ph = tf.placeholder(tf.float32, [None])
self.adv_ph = tf.placeholder(tf.float32, [None])
# Construct model
with tf.variable_scope('pi'):
if self.raw_pixels:
logits = mlp(cnn(self.s_ph), [256, self.n_actions], tf.tanh)
else:
logits = mlp(self.s_ph, [128, 64, self.n_actions], tf.tanh)
self.logp_all = tf.nn.log_softmax(logits)
self.pi = tf.squeeze(tf.multinomial(logits, 1), axis=1)
self.logp_pi = tf.reduce_sum(tf.one_hot(
self.pi, depth=self.n_actions) * self.logp_all, axis=1)
logp = tf.reduce_sum(tf.one_hot(
self.a_ph, depth=self.n_actions) * self.logp_all, axis=1)
# PPO objectives
# pi(a|s) / pi_old(a|s)
ratio = tf.exp(logp - self.logp_old_ph)
min_adv = tf.where(self.adv_ph > 0, (1+clip_ratio)
* self.adv_ph, (1-clip_ratio)*self.adv_ph)
self.pi_loss = - \
tf.reduce_mean(tf.minimum(ratio * self.adv_ph, min_adv))
self.approx_kl = tf.reduce_mean(self.logp_old_ph - logp)
self.approx_ent = tf.reduce_mean(-logp)
self.pi_loss -= 0.01 * self.approx_ent
self.optimizer = tf.train.AdamOptimizer(self.lr).minimize(self.pi_loss)
def action(self, observation):
'''
Choose an action according to approximated softmax policy.
Args:
observation: An observation from the environment
Return:
The action choosed according to the policy
'''
my_action, logp = self.sess.run(
[self.pi, self.logp_pi], feed_dict={self.s_ph: observation})
return my_action, logp
def train(self, states, actions, advs, logp_old, avg_reward, i):
'''Update parameters
Args:
states = [s1, s2, ..., sn]
actions = [a1, a2, ..., an]
advs = [adv1, adv2, ..., advn]
logp_old = [logp1, logp2, ..., logpn]
i: episode number
'''
inputs = {
self.s_ph: states,
self.a_ph: actions,
self.adv_ph: advs,
self.logp_old_ph: logp_old
}
pi_loss_old, ent = self.sess.run(
[self.pi_loss, self.approx_ent], feed_dict=inputs)
for j in range(self.train_policy_iter):
_, kl = self.sess.run(
[self.optimizer, self.approx_kl], feed_dict=inputs)
kl = kl.mean()
if kl > 1.5 * self.target_kl:
logger.info(
'Early stopping at step %d due to reaching max kl.' % j)
break
pi_loss_new, kl = self.sess.run(
[self.pi_loss, self.approx_kl], feed_dict=inputs)
logger.info(
f"\n\tEpisode: {i}\n\tAvg Reward: {avg_reward:.2f}\n\t"
f"Loss_pi: {pi_loss_old:.3e}\n\tEntropy: {ent:.2f}\n\t"
f"KL: {kl:.2f}\n\tDelta_Pi_Loss: {(pi_loss_new - pi_loss_old):.2e}")
class PPOCritic:
def __init__(self, sess, n_features, lr, raw_pixels):
self.sess = sess
if raw_pixels:
self.n_features = n_features
else:
self.n_features = n_features[0]
self.lr = lr
self.model = None
self.train_value_iter = 80
self.raw_pixels = raw_pixels
def build_model(self):
with tf.variable_scope('v'):
if self.raw_pixels:
self.s_ph = tf.placeholder(
tf.float32, [None] + self.n_features)
x = cnn(self.s_ph)
hidden_sizes = [256, 1]
else:
self.s_ph = tf.placeholder(tf.float32, [None, self.n_features])
x = self.s_ph
hidden_sizes = [64, 64, 1]
self.ret_ph = tf.placeholder(tf.float32, [None])
self.value = tf.squeeze(mlp(x, hidden_sizes, tf.tanh), axis=1)
self.v_loss = tf.reduce_mean(
tf.losses.mean_squared_error(self.ret_ph, self.value))
self.optimizer = tf.train.AdamOptimizer(self.lr).minimize(self.v_loss)
def value_of(self, state):
v = self.sess.run(self.value, feed_dict={self.s_ph: state})
return v
def train(self, states, rets, i):
inputs = {
self.s_ph: states,
self.ret_ph: rets
}
for _ in range(self.train_value_iter):
_, loss = self.sess.run(
[self.optimizer, self.v_loss], feed_dict=inputs)
print(f"\tLoss_v = {loss:.2e}")