-
Notifications
You must be signed in to change notification settings - Fork 104
/
ppo.py
158 lines (145 loc) · 6.95 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
import numpy as np
import tensorflow as tf
from tf2rl.algos.vpg import VPG
class PPO(VPG):
"""
Proximal Policy Optimization (PPO) Agent: https://arxiv.org/abs/1707.06347
Command Line Args:
* ``--batch-size`` (int): Batch size of training. The default is ``32``.
* ``--gpu`` (int): GPU id. ``-1`` disables GPU. The default is ``0``.
* ``--horizon`` (int): The default is ``2048``.
* ``--normalize_adv``: Normalize Advantage.
* ``--enable-gae``: Enable GAE.
"""
def __init__(
self,
clip=True,
clip_ratio=0.2,
name="PPO",
**kwargs):
"""
Initialize PPO
Args:
clip (bool): Whether clip or not. The default is ``True``.
clip_ratio (float): Probability ratio is clipped between ``1-clip_ratio`` and ``1+clip_ratio``.
name (str): Name of agent. The default is ``"PPO"``.
state_shape (iterable of int):
action_dim (int):
is_discrete (bool):
actor:
critic:
actor_critic:
max_action (float): maximum action size.
actor_units (iterable of int): Numbers of units at hidden layers of actor. The default is ``(256, 256)``.
critic_units (iterable of int): Numbers of units at hidden layers of critic. The default is ``(256, 256)``.
lr_actor (float): Learning rate of actor. The default is ``1e-3``.
lr_critic (float): Learning rate of critic. The default is ``3e-3``.
hidden_activation_actor (str): Activation for actor. The default is ``"relu"``.
hidden_activation_critic (str): Activation for critic. The default is ``"relu"``.
horizon (int): Number of steps of online episode horizon. The horizon must be multiple of ``batch_size``. The default is ``2048``.
enable_gae (bool): Enable GAE. The default is ``True``.
normalize_adv (bool): Normalize Advantage. The default is ``True``.
entropy_coef (float): Entropy coefficient. The default is ``0.01``.
vfunc_coef (float): Mixing ratio factor for actor and critic. ``actor_loss + vfunc_coef*critic_loss``
batch_size (int): Batch size. The default is ``256``.
"""
super().__init__(name=name, **kwargs)
self.clip = clip
self.clip_ratio = clip_ratio
def train(self, states, actions, advantages, logp_olds, returns):
"""
Train PPO
Args:
states
actions
advantages
logp_olds
returns
"""
# Train actor and critic
if self.actor_critic is not None:
actor_loss, critic_loss, logp_news, ratio, ent = self._train_actor_critic_body(states, actions, advantages, logp_olds, returns)
else:
actor_loss, logp_news, ratio, ent = self._train_actor_body(
states, actions, advantages, logp_olds)
critic_loss = self._train_critic_body(states, returns)
# Visualize results in TensorBoard
tf.summary.scalar(name=self.policy_name+"/actor_loss",
data=actor_loss)
tf.summary.scalar(name=self.policy_name+"/logp_max",
data=np.max(logp_news))
tf.summary.scalar(name=self.policy_name+"/logp_min",
data=np.min(logp_news))
tf.summary.scalar(name=self.policy_name+"/logp_mean",
data=np.mean(logp_news))
tf.summary.scalar(name=self.policy_name+"/adv_max",
data=np.max(advantages))
tf.summary.scalar(name=self.policy_name+"/adv_min",
data=np.min(advantages))
tf.summary.scalar(name=self.policy_name+"/kl",
data=tf.reduce_mean(logp_olds - logp_news))
tf.summary.scalar(name=self.policy_name+"/ent",
data=ent)
tf.summary.scalar(name=self.policy_name+"/ratio",
data=tf.reduce_mean(ratio))
tf.summary.scalar(name=self.policy_name+"/critic_loss",
data=critic_loss)
return actor_loss, critic_loss
@tf.function
def _train_actor_critic_body(self, states, actions, advantages, logp_olds, returns):
with tf.device(self.device):
with tf.GradientTape() as tape:
_, _, current_V = self.actor_critic(states)
ent = tf.reduce_mean(
self.actor_critic.compute_entropy(states))
# Train actor
if self.clip:
logp_news = self.actor_critic.compute_log_probs(
states, actions)
ratio = tf.math.exp(
logp_news - tf.squeeze(logp_olds))
min_adv = tf.clip_by_value(
ratio,
1.0 - self.clip_ratio,
1.0 + self.clip_ratio) * tf.squeeze(advantages)
actor_loss = -tf.reduce_mean(tf.minimum(
ratio * tf.squeeze(advantages),
min_adv))
actor_loss -= self.entropy_coef * ent
else:
raise NotImplementedError
# Train critic
td_errors = tf.squeeze(returns) - current_V
critic_loss = tf.reduce_mean(tf.square(td_errors))
total_loss = actor_loss + self.vfunc_coef * critic_loss
grads = tape.gradient(
total_loss, self.actor_critic.trainable_variables)
self.actor_critic_optimizer.apply_gradients(
zip(grads, self.actor_critic.trainable_variables))
return actor_loss, critic_loss, logp_news, ratio, ent
@tf.function
def _train_actor_body(self, states, actions, advantages, logp_olds):
with tf.device(self.device):
# Update actor
with tf.GradientTape() as tape:
ent = tf.reduce_mean(
self.actor.compute_entropy(states))
if self.clip:
logp_news = self.actor.compute_log_probs(
states, actions)
ratio = tf.math.exp(logp_news - tf.squeeze(logp_olds))
min_adv = tf.clip_by_value(
ratio,
1.0 - self.clip_ratio,
1.0 + self.clip_ratio) * tf.squeeze(advantages)
actor_loss = -tf.reduce_mean(tf.minimum(
ratio * tf.squeeze(advantages),
min_adv))
actor_loss -= self.entropy_coef * ent
else:
raise NotImplementedError
actor_grad = tape.gradient(
actor_loss, self.actor.trainable_variables)
self.actor_optimizer.apply_gradients(
zip(actor_grad, self.actor.trainable_variables))
return actor_loss, logp_news, ratio, ent