-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathDDQN.py
169 lines (149 loc) · 5.68 KB
/
DDQN.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
import gym
import numpy as np
import keras
import random
from collections import deque
from keras.models import Sequential,Model
from keras.layers import Dense,Conv2D,Flatten,Dropout,MaxPooling2D
from keras.optimizers import Adam
import matplotlib.pyplot as plt
import cv2
import time
max_ep=200
train = True
class dqnagent():
def __init__(self, lr = 0.00025, ob_size = (84,84,4), action_size = 2, env = 'BreakoutNoFrameskip-v0'):
self.state_size = (84,84,4)
self.action_size = 4
# build model to extimate q value
self.model = self._build_model(lr)
# build target model
self.model_t = self._build_model(lr)
self.replay_memory = deque(maxlen = 100000) # experience replay_memory to store value
self.reward_memory = deque(maxlen = 100)
self.env = gym.make(env)
#self.env = Monitor(env, "/tmp",force = True)
self.ep_start = 1
self.ep_stop = .1
self.ep = 1
self.ep_decay = (1-.1)/1000000
self.batch_size = 32
self.gamma = 0.99
self.t = 0
# make target and main model same first then after end of every episode we will update it
self.update_target_model()
def add_memory(self,s,a,r,d,s2):
# adding experience replay memory
self.replay_memory.append((s, a, r, d, s2))
def choose_action(self,s):
ran = np.random.random()
# you can use non linear decay rate but we will use linear decay for to get good result ####---
self.t += 1
#self.ep = self.ep_stop + (self.ep_start - self.ep_stop)*np.exp(-self.ep_decay*self.t_)
self.ep = self.ep_start - self.t*self.ep_decay
if self.ep >= ran :
#self.ep -= self.ep_decay
return self.env.action_space.sample()
else:
a = self.model.predict(s)
return np.argmax(a[0])
def learn(self):
st_ = np.zeros((self.batch_size,84,84,4))
st_2 = np.zeros((self.batch_size,84,84,4))
out = np.zeros((self.batch_size,4))
batch = random.sample(self.replay_memory, self.batch_size)
i=0
for s, a , r, d, s2 in batch:
st_[i:i+1] = s
st_2[i:i+1] = s2
target = r
if d == False:
aa=np.argmax(self.model.predict(s2)[0])
target = r + self.gamma * ( self.model_t.predict(s2)[0][aa] )
out[i] = self.model.predict(s)
out[i][a] = target
i = i +1
self.model.fit(st_,out,epochs=1,verbose=0)
def _build_model(self,lr):
init = keras.initializers.RandomNormal(mean=0.0, stddev=0.05, seed=None)
shape_image=(84,84,4)
model = Sequential()
model.add(keras.layers.Lambda(lambda x: x / 255.0,input_shape = shape_image))
model.add(Conv2D(32,(8,8), strides=4,use_bias =True,bias_initializer='zeros',kernel_initializer = init,activation = 'relu'))
model.add(MaxPooling2D(pool_size=2))
model.add(Conv2D(64,(4,4), strides = 2, use_bias = True, bias_initializer = 'zeros',kernel_initializer = init, activation='relu'))
model.add(Conv2D(64,(3,3),use_bias= True, bias_initializer = 'zeros', kernel_initializer = init, activation = 'relu'))
model.add(Flatten())
model.add(Dense(512, activation='relu', kernel_initializer='he_uniform' ))
model.add(Dense(24, activation = 'relu',kernel_initializer='he_uniform' ))
model.add(Dense(self.action_size, activation='linear', kernel_initializer = 'he_uniform'))
model.compile(optimizer=keras.optimizers.RMSprop(lr,rho=0.95), loss = 'mse')
return model
def model_save(self):
self.model.save_weights("model_breakout_dqn.h5")
def env_re(self):
return self.env.reset()
def step(self,a):
#self.env.render()
return self.env.step(a)
def update_target_model(self):
self.model_t.set_weights(self.model.get_weights())
def model_load(self):
self.model.load_weights("model_breakout_dqn.h5")
self.model_t.load_weights("model_breakout_dqn.h5")
record = []
env_name = 'BreakoutNoFrameskip-v0'
batch_size = 32
count = 0
brain = dqnagent()
learning_start = 10000
st = time.time()
model_saved = False
if model_saved == True:
brain.model_load()
#env = wrappers.Monitor(brain.env,force=True, '/tmp/cartpole-experiment-1')
#env = Monitor(env, directory='/tmp/pp',video_callable=False,force=True, write_upon_reset=True)
update_ = 0
if train == True:
for i in range(100000):
s = brain.env_re()
s =cv2.resize(cv2.cvtColor(s, cv2.COLOR_BGR2GRAY),(84,84))
s = np.reshape(s,(1,84,84))
s = [s for _ in range(4)]
s = np.stack(s,axis=3)
s = np.array(s)
#print(s.shape)
d = False
R = 0
while not d:
update_ += 1
a = brain.choose_action(s)
#print(a)
s2, r, d, _ = brain.step(a)
s2 =cv2.resize(cv2.cvtColor(s2, cv2.COLOR_BGR2GRAY),(84,84))
s2= np.reshape(s2, (1,84,84,1) )
s2=np.concatenate((s2,s[:,:,:,0:3]),axis=3)
R +=r
if d == True :
r = -1
brain.add_memory(s,a,r,d,s2)
s = s2
count += 1
if count > learning_start and count %4 == 0:
brain.learn()
if d == True:
record.append(R)
print(i, R)
break
if update_ == 10000:
update_ = 0
brain.update_target_model()
if (i+1) % 100 == 0:
brain.model_save()
else:
brain.model.load_weights("model_cart.h5")
record = np.array(record)
plt.plot(record)
plt.xlabel('no of episode')
plt.ylabel('score')
plt.show()