-
Notifications
You must be signed in to change notification settings - Fork 2
/
policy_gradients.py
131 lines (108 loc) · 4.29 KB
/
policy_gradients.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
import tensorflow as tf
import numpy as np
import gym
import os
from gym.wrappers import Monitor
class Net(object):
"""docstring for Net"""
def __init__(self, layers, batch_size, learning_rate = 5e-3):
self.layers = layers
self.batch_size = batch_size
self.learning_rate = learning_rate
def build_model(self, size_of_state, num_of_actions, session):
self.states = tf.placeholder(shape = [None, size_of_state], name = "states", dtype = tf.float32)
self.actions = tf.placeholder(shape = [None], name = "actions_for_each_state", dtype = tf.int32)
# self.rewards = tf.placeholder(shape = [None], name = "rewards_for_each_state", dtype = tf.float32)
self.advantages = tf.placeholder(shape = [None], name = "Advantages_for_each_state", dtype = tf.float32)
self.architecture = {}
temp_input = self.states
for index, num_units in enumerate(self.layers[:-1]):
self.architecture[index] = tf.layers.dense(temp_input, num_units, activation = tf.nn.relu, kernel_initializer = tf.contrib.layers.xavier_initializer())
temp_input = self.architecture[index]
# self.architecture[len(self.layers) - 1] is the output layer
self.architecture[len(self.layers) - 1] = tf.layers.dense(temp_input, num_of_actions)
logits = self.architecture[len(self.layers) - 1]
self.probs = tf.nn.softmax(logits)
self.log_probs = tf.nn.sparse_softmax_cross_entropy_with_logits(logits = logits, labels = self.actions)
self.loss = tf.reduce_mean(self.log_probs * self.advantages)
self.optimizer = tf.train.AdamOptimizer(self.learning_rate).minimize(self.loss)
session.run(tf.global_variables_initializer())
def train(self, session, actual_states, actual_actions, actual_advantages):
feed_in = {self.states:actual_states, self.actions:actual_actions, self.advantages:actual_advantages}
loss, _ = session.run([self.loss, self.optimizer], feed_in)
return loss
def compute_advantage(rewards, gamma):
adv = [0] * len(rewards)
for index, reward in enumerate(rewards):
if index == 0:
adv[-1] = rewards[-1]
adv[(-1 * index) - 1] += rewards[(-1 * index) - 1] + gamma*adv[-1 * index]
adv = np.array(adv)
mean = np.mean(adv)
std = np.std(adv)
adv = (adv - mean)/std
for advan in adv:
print(advan)
return adv
def main(option):
env = gym.make("CartPole-v0")
size_of_state = env.observation_space.shape[0]
# print(size_of_state)
num_of_actions = env.action_space.n
# print(num_of_actions)
batch_size = 25
gamma = 0.99
session = tf.Session()
my_net = Net([64, num_of_actions], batch_size)
my_net.build_model(size_of_state, num_of_actions,session)
saver = tf.train.Saver()
ckpt = tf.train.get_checkpoint_state(os.path.dirname('PG_checkpoints/'))
if ckpt and ckpt.model_checkpoint_path:
# print("hi")
saver.restore(session, ckpt.model_checkpoint_path)
if option == "play":
state = env.reset()
episode_reward = 0
while True:
env.render()
feed_in = {my_net.states:np.array([state])}
probs = session.run([my_net.probs], feed_in)[0][0]
possible_actions = np.arange(num_of_actions)
action_taken = np.random.choice(possible_actions, p = probs)
state, reward, done, _ = env.step(action_taken)
episode_reward += reward
if done:
break
else:
iterations = 1
states, actions, rewards, advantages = [], [], [], []
advantages = np.array(advantages)
while iterations != 2000:
state = env.reset()
temp_rewards = []
episode_reward = 0
while True:
feed_in = {my_net.states:np.array([state])}
probs = session.run([my_net.probs], feed_in)[0][0]
possible_actions = np.arange(num_of_actions)
action_taken = np.random.choice(possible_actions, p = probs)
actions.append(action_taken)
states.append(state)
state, reward, done, _ = env.step(action_taken)
temp_rewards.append(reward)
episode_reward += reward
if done:
rewards.append(temp_rewards)
# advantages = advantages + compute_advantage(temp_rewards, gamma)
advantages = np.concatenate(advantages, compute_advantage(temp_rewards, gamma))
# print(episode_reward)
saver.save(session, 'PG_checkpoints/PG')
break
iterations += 1
if iterations % batch_size == 0:
loss = my_net.train(session, states, actions, advantages)
# print(" - ", loss)
states, actions, rewards, advantages = [], [], [], []
if __name__ == '__main__':
option = input("Play or Train?")
main(option)