-
Notifications
You must be signed in to change notification settings - Fork 0
/
actorCritic.py
175 lines (143 loc) · 7.08 KB
/
actorCritic.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
from env import Dino
from memory import Memory
import numpy as np
from collections import deque
import tensorflow as tf
import os
import shutil
import time
import random
class ActorCritic():
def __init__(self,stack_size=4,episodes=500,max_steps=20,mem_size=100000,batch_size=64):
self.env=Dino(skip_frame=0)
self.action_space=self.env.action_space
self.value_size=1
self.observation_space=self.env.observation_space
self.episodes=episodes
self.max_steps=max_steps
self.mem_size=mem_size
self.batch_size=batch_size
self.memory=deque(maxlen = mem_size)
# These are hyper parameters for the DQN
self.discount_factor = 0.99
self.critic_lr=0.005
self.actor_lr=0.001
self.skip_frame=self.env.skip_frame
self.stack_size = stack_size # We stack 4 frames
# Initialize deque with zero-images one array for each image
self.stacked_frames = deque([np.zeros(self.observation_space, dtype=np.int) for i in range(stack_size)], maxlen=stack_size)
self.actor,self.critic=self.build_models(self.stack_size)
def stack_frames(self, state, is_new_episode):
"""
"""
frame=np.reshape(state,self.observation_space)
if is_new_episode:
# Clear our stacked_frames
self.stacked_frames = deque([np.zeros(self.observation_space, dtype=np.int) for i in range(self.stack_size)], maxlen=self.stack_size)
# Because we're in a new episode, copy the same frame 4x
for _ in range(self.stack_size):
self.stacked_frames.append(frame)
# Stack the frames
stacked_state = np.stack(self.stacked_frames, axis=2)
# print(stacked_state.shape," <---")
else:
# Append frame to deque, automatically removes the oldest frame
self.stacked_frames.append(frame)
# Build the stacked state (first dimension specifies different frames)
stacked_state = np.stack(self.stacked_frames, axis=2)
return stacked_state
def append_sample(self,exp):
self.memory.append(exp)
def build_models(self,channels):
actor = tf.keras.models.Sequential()
# Add a Convolutional layer activation=tf.keras.layers.LeakyReLU(alpha=0.3)
actor.add(tf.keras.layers.Conv2D(32, (3, 3), activation="relu",kernel_initializer='he_uniform', input_shape=(*self.observation_space,channels)))
# Add a Max pooling layer
actor.add(tf.keras.layers.MaxPool2D())
# Add a Convolutional layer
actor.add(tf.keras.layers.Conv2D(64, (3, 3), activation="relu",kernel_initializer='he_uniform'))
# Add a Max pooling layer
actor.add(tf.keras.layers.MaxPool2D())
# Add a Convolutional layer
actor.add(tf.keras.layers.Conv2D(64, (3, 3), activation="relu",kernel_initializer='he_uniform'))
# Add a Max pooling layer
actor.add(tf.keras.layers.MaxPool2D())
# Add the flattened layer
actor.add(tf.keras.layers.Flatten())
# Add the hidden layer
actor.add(tf.keras.layers.Dense(512, activation="relu",kernel_initializer='he_uniform'))
# Adding a dropout layer
actor.add(tf.keras.layers.Dropout(0.3))
# Add the output layer
actor.add(tf.keras.layers.Dense(self.action_space,kernel_initializer='he_uniform', activation='softmax'))
# Compiling the model
actor.compile(optimizer=tf.keras.optimizers.Adam(lr=self.actor_lr), loss="categorical_crossentropy", metrics=[tf.keras.metrics.AUC()])
print (actor.summary())
critic = tf.keras.models.Sequential()
# Add a Convolutional layer activation=tf.keras.layers.LeakyReLU(alpha=0.3)
critic.add(tf.keras.layers.Conv2D(32, (3, 3), activation="relu",kernel_initializer='he_uniform', input_shape=(*self.observation_space,channels)))
# Add a Max pooling layer
critic.add(tf.keras.layers.MaxPool2D())
# Add a Convolutional layer
critic.add(tf.keras.layers.Conv2D(64, (3, 3), activation="relu",kernel_initializer='he_uniform'))
# Add a Max pooling layer
critic.add(tf.keras.layers.MaxPool2D())
# Add a Convolutional layer
critic.add(tf.keras.layers.Conv2D(64, (3, 3), activation="relu",kernel_initializer='he_uniform'))
# Add a Max pooling layer
critic.add(tf.keras.layers.MaxPool2D())
# Add the flattened layer
critic.add(tf.keras.layers.Flatten())
# Add the hidden layer
critic.add(tf.keras.layers.Dense(256,activation='relu',kernel_initializer='he_uniform'))
critic.add(tf.keras.layers.Dropout(0.3))
critic.add(tf.keras.layers.Dense(512,activation='relu',kernel_initializer='he_uniform'))
critic.add(tf.keras.layers.Dropout(0.3))
critic.add(tf.keras.layers.Dense(self.value_size,activation='linear',kernel_initializer='he_uniform'))
critic.compile(optimizer=tf.keras.optimizers.Adam(lr=self.critic_lr),loss='mse',metrics=['accuracy'])
print(critic.summary())
return actor,critic
# using the output of policy network, pick action stochastically
def get_action(self, state):
policy = self.actor.predict(np.expand_dims(state,axis=0), batch_size=1).flatten()
return np.random.choice(self.action_space, 1, p=policy)[0]
# update policy network every episode
def train_model(self):
# ##########################Learning start############################
if len(self.memory) > self.batch_size:
train_time_start=time.time()
### LEARNING PART
# Obtain random mini-batch from memory
batch = random.sample(self.memory,self.batch_size)
states_mb = np.array([each[0] for each in batch]).reshape(self.batch_size,*self.observation_space,self.stack_size)
actions_mb = np.array([each[1] for each in batch])
rewards_mb = np.array([each[2] for each in batch])
next_states_mb = np.array([each[3] for each in batch]).reshape(self.batch_size,*self.observation_space,self.stack_size)
dones_mb = np.array([each[4] for each in batch])
target = np.full((self.batch_size, self.value_size),0)
advantages = np.full((self.batch_size, self.action_space),0)
value = self.critic.predict(states_mb)
# print(value[0].shape,"<<---")
next_value = self.critic.predict(next_states_mb)
# update the target values
for i in range(self.batch_size):
if dones_mb[i]:
advantages[i][actions_mb[i]] = rewards_mb[i] - value[i]
target[i][0] = rewards_mb[i]
else: # non-terminal state
advantages[i][actions_mb[i]] = rewards_mb[i] + self.discount_factor * (next_value[i] - value[i])
target[i][0] = rewards_mb[i] + self.discount_factor * next_value[i]
log_dir = os.path.join("tensorboard","a2c")
file_writer = tf.summary.create_file_writer(log_dir + "/metrics")
file_writer.set_as_default()
tensorboard_callback = tf.keras.callbacks.TensorBoard(log_dir=log_dir)
# model fit
self.actor.fit(states_mb, advantages,batch_size=self.batch_size,callbacks=[tensorboard_callback], epochs=1, verbose=0,use_multiprocessing=True)
self.critic.fit(states_mb, target,batch_size=self.batch_size,callbacks=[tensorboard_callback], epochs=1, verbose=0,use_multiprocessing=True)
return time.time()-train_time_start
else:
return 0
# ##########################Learning end##############################
def save_model(self):
tf.keras.models.save_model(self.actor,'./actor.h5')
tf.keras.models.save_model(self.critic,'./critic.h5')