/
Pong_A2C_episodic.py
253 lines (194 loc) · 9.16 KB
/
Pong_A2C_episodic.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
# -*- coding: utf-8 -*-
import numpy as np
import tensorflow as tf
import gym
from collections import deque
from skimage.transform import resize
from skimage.color import rgb2gray
env = gym.make('PongDeterministic-v4')
# 하이퍼 파라미터
LEARNING_RATE = 0.00025
INPUT = env.observation_space.shape[0]
OUTPUT = 2 # 액션수를 두개로 줄임. UP or DOWN
DISCOUNT = 0.99
HEIGHT = 84
WIDTH = 84
HISTORY_SIZE = 4
model_path = 'save/pong-pg.ckpt'
def pre_proc(X):
'''입력데이터 전처리.
Args:
X(np.array): 받아온 이미지를 그레이 스케일링 후 84X84로 크기변경
그리고 정수값으로 저장하기위해(메모리 효율 높이기 위해) 255를 곱함
Returns:
np.array: 변경된 이미지
'''
# 바로 전 frame과 비교하여 max를 취함으로써 flickering을 제거
# x = np.maximum(X, X1)
# 그레이 스케일링과 리사이징을 하여 데이터 크기 수정
x = np.uint8(resize(rgb2gray(X), (HEIGHT, WIDTH), mode='reflect') * 255)
return x
def discount_rewards(r):
'''Discounted reward를 구하기 위한 함수
Args:
r(np.array): reward 값이 저장된 array
Returns:
discounted_r(np.array): Discounted 된 reward가 저장된 array
'''
discounted_r = np.zeros_like(r, dtype=np.float32)
running_add = 0
for t in reversed(range(len(r))):
running_add = running_add * DISCOUNT + r[t]
discounted_r[t] = running_add
return discounted_r
def get_init_state(history, s):
'''에피소드 시작 State를 초기화.
Args:
history(np.array): 5개의 프레임이 저장될 array
s(list): 초기화된 이미지
Note:
history[:,:,:3]에 모두 초기화된 이미지(s)를 넣어줌
'''
for i in range(HISTORY_SIZE):
history[:, :, i] = pre_proc(s)
def train_episodic(A2Cagent, x, y, r):
'''에피소드당 학습을 하기위한 함수
Args:
A2Cagent(ActorCritic): 학습될 네트워크
x(np.array): State가 저장되어있는 array
y(np.array): Action(one_hot)이 저장되어있는 array
r(np.array) : Discounted reward가 저장되어있는 array
Returns:
l(float): 네트워크에 의한 loss
'''
l, _ = A2Cagent.sess.run([A2Cagent.loss, A2Cagent.train], feed_dict={A2Cagent.X: x, A2Cagent.Y: y, A2Cagent.r: r})
return l
def play_atari(A2Cagent):
'''학습된 네트워크로 Play하기 위한 함수
Args:
PGagent(PolicyGradient): 학습된 네트워크
'''
print("Play Atari!")
episode = 0
while True:
s = env.reset()
history = np.zeros([84, 84, 5], dtype=np.uint8)
done = False
rall = 0
episode += 1
get_init_state(history, s)
while not done:
env.render()
action_p = A2Cagent.sess.run(
A2Cagent.a_prob,feed_dict={A2Cagent.X: np.reshape(np.float32(history[:,:,:4] / 255.), [-1, 84, 84, 4])})
s1, reward, done, _ = env.step(np.argmax(action_p)+2)
history[:, :, 4] = pre_proc(s1)
history[:, :, :4] = history[:, :, 1:]
rall += reward
print("[Episode {0:6f}] Reward: {1:4f} ".format(episode, rall))
class ActorCritic:
def __init__(self, sess, input_size, output_size):
self.sess = sess
self.input_size = input_size
self.output_size = output_size
self.height = HEIGHT
self.width = WIDTH
self.history_size = HISTORY_SIZE
self.build_network()
def build_network(self):
self.X = tf.placeholder('float', [None, self.height, self.width, self.history_size])
self.Y = tf.placeholder('float', [None, self.output_size])
self.r = tf.placeholder('float')
# Actor network
f1_a = tf.get_variable("f1_a", shape=[1, 1, 4, 1], initializer=tf.contrib.layers.xavier_initializer_conv2d())
f2_a = tf.get_variable("f2_a", shape=[4, 4, 1, 16], initializer=tf.contrib.layers.xavier_initializer_conv2d())
f3_a = tf.get_variable("f3_a", shape=[4, 4, 16, 32], initializer=tf.contrib.layers.xavier_initializer_conv2d())
w1_a = tf.get_variable("w1_a", shape=[6 * 6 * 32, 256], initializer=tf.contrib.layers.xavier_initializer())
w2_a = tf.get_variable("w2_a", shape=[256, OUTPUT], initializer=tf.contrib.layers.xavier_initializer())
c1_a = tf.nn.relu(tf.nn.conv2d(self.X, f1_a, strides=[1, 1, 1, 1], padding="VALID"))
c2_a = tf.nn.relu(tf.nn.conv2d(c1_a, f2_a, strides=[1, 4, 4, 1], padding="VALID"))
c3_a = tf.nn.relu(tf.nn.conv2d(c2_a, f3_a, strides=[1, 3, 3, 1], padding="VALID"))
l1_a = tf.reshape(c3_a, [-1, w1_a.get_shape().as_list()[0]])
l2_a = tf.nn.relu(tf.matmul(l1_a, w1_a))
self.a_prob = tf.nn.softmax(tf.matmul(l2_a, w2_a))
# Critic network(like DQN network)
f1_c = tf.get_variable("f1_c", shape=[8, 8, 4, 16], initializer=tf.contrib.layers.xavier_initializer_conv2d())
f2_c = tf.get_variable("f2_c", shape=[4, 4, 16, 32], initializer=tf.contrib.layers.xavier_initializer_conv2d())
w1_c = tf.get_variable("w1_c", shape=[9 * 9 * 32, 256], initializer=tf.contrib.layers.xavier_initializer())
w2_c = tf.get_variable("w2_c", shape=[256, 1], initializer=tf.contrib.layers.xavier_initializer())
c1_c = tf.nn.relu(tf.nn.conv2d(self.X, f1_c, strides=[1, 4, 4, 1], padding="VALID"))
c2_c = tf.nn.relu(tf.nn.conv2d(c1_c, f2_c, strides=[1, 2, 2, 1], padding="VALID"))
l1_c = tf.reshape(c2_c, [-1, w1_c.get_shape().as_list()[0]])
l2_c = tf.nn.relu(tf.matmul(l1_c, w1_c))
self.v = tf.matmul(l2_c, w2_c)
# A_t = R_t - V(S_t)
self.adv = self.r - self.v
# Policy loss
self.log_p = self.Y * tf.log(tf.clip_by_value(self.a_prob,1e-10,1.))
self.log_lik = self.log_p * tf.stop_gradient(self.adv)
self.p_loss = -tf.reduce_mean(tf.reduce_sum(self.log_lik, axis=1))
# Value loss
self.v_loss = tf.reduce_mean(tf.square(self.v - self.r), axis=1)
# Total loss
self.loss = self.p_loss + self.v_loss
self.train = tf.train.AdamOptimizer(LEARNING_RATE).minimize(self.loss)
self.saver = tf.train.Saver()
def get_action(self, state, max_prob):
action_p = self.sess.run(self.a_prob, feed_dict={self.X: np.reshape(np.float32(state / 255.),
[-1, HEIGHT, WIDTH, HISTORY_SIZE])})
# 각 액션의 확률로 액션을 결정
max_prob.append(np.max(action_p))
action = np.random.choice(np.arange(self.output_size), p=action_p[0])
return action
def main():
with tf.Session() as sess:
# VRAM이 부족하면 CPU로 학습
# with tf.Session(config = tf.ConfigProto(device_count ={'GPU' : 0})) as sess:
A2Cagent = ActorCritic(sess, INPUT, OUTPUT)
A2Cagent.sess.run(tf.global_variables_initializer())
episode = 0
recent_rlist = deque(maxlen=100)
recent_rlist.append(0)
# 최근 100개의 점수가 19점 넘을 때까지 학습
while np.mean(recent_rlist) <= 19:
episode += 1
state_memory = deque()
action_memory = deque()
reward_memory = deque()
# 공의 움직임을 알아보기 위한 History
history = np.zeros([84, 84, HISTORY_SIZE+1], dtype=np.uint8)
rall, count = 0, 0
done = False
s = env.reset()
max_prob = deque()
get_init_state(history, s)
while not done:
# env.render()
count += 1
# 액션 선택
action = A2Cagent.get_action(history[:,:,:HISTORY_SIZE], max_prob)
# action을 one_hot으로 표현
y = np.zeros(OUTPUT)
y[action] = 1
# 학습속도 개선을 위해 액션의 개수를 2개로 줄임 (UP or DOWN)
s1, reward, done, l = env.step(action + 2)
rall += reward
reward = np.clip(reward, -1, 1)
# 한 에피소드를 저장
state_memory.append(np.copy(np.float32(history[:,:,:HISTORY_SIZE]/255.)))
action_memory.append(np.copy(y))
reward_memory.append(np.copy(reward))
# 새로운 프레임을 히스토리 마지막에 넣어줌
history[:, :, HISTORY_SIZE] = pre_proc(s1)
history[:, :, :HISTORY_SIZE] = history[:, :, 1:]
# 에피소드가 끝났을때 학습
if done:
# Discounted return을 구함
rewards = discount_rewards(np.vstack(reward_memory))
train_episodic(A2Cagent, np.stack(state_memory, axis=0), np.stack(action_memory, axis =0), rewards)
recent_rlist.append(rall)
print("[Episode {0:6d}] Reward: {1:4f} Recent Reward: {2:4f} Max Prob: {3:5.10f}".format(episode, rall, np.mean(recent_rlist), np.mean(max_prob)))
if episode % 10 == 0:
A2Cagent.saver.save(A2Cagent.sess, model_path, global_step= episode)
if __name__ == "__main__":
main()