-
Notifications
You must be signed in to change notification settings - Fork 70
/
Copy pathDQNModel.py
204 lines (173 loc) · 8.55 KB
/
DQNModel.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
#!/usr/bin/env python
# -*- coding: utf-8 -*-
# File: DQNModel.py
# Author: Yuxin Wu <[email protected]>
# Modified: Amir Alansary <[email protected]>
import abc
import tensorflow as tf
from tensorpack import ModelDesc, InputDesc
from tensorpack.utils import logger
from tensorpack.tfutils import (
collection, summary, get_current_tower_context, optimizer, gradproc)
from tensorpack.tfutils.scope_utils import auto_reuse_variable_scope
class Model2D(ModelDesc):
def __init__(self, image_shape, channel, method, num_actions, gamma):
"""
:param image_shape: the shape of input 2d image
:param channel: history length and goes to channel dimension in kernel
:param method: dqn or double (default is double)
:param num_actions: number of actions
:param gamma: discount factor
"""
self.gamma = gamma
self.method = method
self.channel = channel
self.image_shape = image_shape
self.num_actions = num_actions
def inputs(self):
# Use a combined state for efficiency.
# The first h channels are the current state, and the last h channels are the next state.
return [InputDesc(tf.uint8,
(None,) + self.image_shape + (self.channel + 1,),
'comb_state'),
InputDesc(tf.int64, (None,), 'action'),
InputDesc(tf.float32, (None,), 'reward'),
InputDesc(tf.bool, (None,), 'isOver')]
@abc.abstractmethod
def _get_DQN_prediction(self, image):
pass
# decorate the function
@auto_reuse_variable_scope
def get_DQN_prediction(self, image):
return self._get_DQN_prediction(image)
def build_graph(self, *inputs):
comb_state, action, reward, isOver = inputs
comb_state = tf.cast(comb_state, tf.float32)
state = tf.slice(comb_state, [0, 0, 0, 0], [-1, -1, -1, self.channel], name='state')
self.predict_value = self.get_DQN_prediction(state)
if not get_current_tower_context().is_training:
return
reward = tf.clip_by_value(reward, -1, 1)
next_state = tf.slice(comb_state, [0, 0, 0, 1], [-1, -1, -1, self.channel], name='next_state')
action_onehot = tf.one_hot(action, self.num_actions, 1.0, 0.0)
pred_action_value = tf.reduce_sum(self.predict_value * action_onehot, 1) # N,
max_pred_reward = tf.reduce_mean(tf.reduce_max(
self.predict_value, 1), name='predict_reward')
summary.add_moving_summary(max_pred_reward)
with tf.variable_scope('target'):
targetQ_predict_value = self.get_DQN_prediction(next_state) # NxA
if 'Double' not in self.method:
# DQN or Dueling
best_v = tf.reduce_max(targetQ_predict_value, 1) # N,
else:
# Double-DQN or DuelingDouble
next_predict_value = self.get_DQN_prediction(next_state)
self.greedy_choice = tf.argmax(next_predict_value, 1) # N,
predict_onehot = tf.one_hot(self.greedy_choice, self.num_actions, 1.0, 0.0)
best_v = tf.reduce_sum(targetQ_predict_value * predict_onehot, 1)
target = reward + (1.0 - tf.cast(isOver, tf.float32)) * self.gamma * tf.stop_gradient(best_v)
cost = tf.losses.huber_loss(target, pred_action_value,
reduction=tf.losses.Reduction.MEAN)
summary.add_param_summary(('conv.*/W', ['histogram', 'rms']),
('fc.*/W', ['histogram', 'rms'])) # monitor all W
summary.add_moving_summary(cost)
return cost
def optimizer(self):
lr = tf.get_variable('learning_rate',initializer=1e-3, trainable=False)
opt = tf.train.AdamOptimizer(lr, epsilon=1e-3)
return optimizer.apply_grad_processors(
opt, [gradproc.GlobalNormClip(10), gradproc.SummaryGradient()])
@staticmethod
def update_target_param():
vars = tf.global_variables()
ops = []
G = tf.get_default_graph()
for v in vars:
target_name = v.op.name
if target_name.startswith('target'):
new_name = target_name.replace('target/', '')
logger.info("{} <- {}".format(target_name, new_name))
ops.append(v.assign(G.get_tensor_by_name(new_name + ':0')))
return tf.group(*ops, name='update_target_network')
class Model3D(ModelDesc):
def __init__(self, image_shape, channel, method, num_actions, gamma):
"""
:param image_shape: the shape of input 3d image
:param channel: history length and goes to channel dimension in kernel
:param method: dqn or double (default is double)
:param num_actions: number of actions
:param gamma: discount factor
See http://tensorpack.readthedocs.io/tutorial/training-interface.html for Mode lDesc documentation.
"""
self.gamma = gamma
self.method = method
self.channel = channel
self.image_shape = image_shape
self.num_actions = num_actions
def inputs(self):
# Use a combined state for efficiency.
# The first h channels are the current state, and the last h channels are the next state.
return [InputDesc(tf.uint8,
(None,) + self.image_shape + (self.channel + 1,),
'comb_state'),
InputDesc(tf.int64, (None,), 'action'),
InputDesc(tf.float32, (None,), 'reward'),
InputDesc(tf.bool, (None,), 'isOver')]
@abc.abstractmethod
def _get_DQN_prediction(self, image):
"""this method is overridden in DQN.py, where it will return a list of predicted Q-values"""
pass
# decorate the function
@auto_reuse_variable_scope
def get_DQN_prediction(self, image):
return self._get_DQN_prediction(image)
def build_graph(self, *inputs):
comb_state, action, reward, isOver = inputs
comb_state = tf.cast(comb_state, tf.float32)
state = tf.slice(comb_state, [0, 0, 0, 0, 0], [-1, -1, -1, -1, self.channel], name='state')
self.predict_value = self.get_DQN_prediction(state)
if not get_current_tower_context().is_training:
return
reward = tf.clip_by_value(reward, -1, 1)
next_state = tf.slice(comb_state, [0, 0, 0, 0, 1], [-1, -1, -1, -1, self.channel], name='next_state')
action_onehot = tf.one_hot(action, self.num_actions, 1.0, 0.0)
pred_action_value = tf.reduce_sum(self.predict_value * action_onehot, 1) # N,
max_pred_reward = tf.reduce_mean(tf.reduce_max(
self.predict_value, 1), name='predict_reward')
summary.add_moving_summary(max_pred_reward)
with tf.variable_scope('target'):
targetQ_predict_value = self.get_DQN_prediction(next_state) # NxA
if 'Double' not in self.method:
# DQN or Dueling
best_v = tf.reduce_max(targetQ_predict_value, 1) # N,
else:
# Double-DQN or DuelingDouble
next_predict_value = self.get_DQN_prediction(next_state)
self.greedy_choice = tf.argmax(next_predict_value, 1) # N,
predict_onehot = tf.one_hot(self.greedy_choice, self.num_actions, 1.0, 0.0)
best_v = tf.reduce_sum(targetQ_predict_value * predict_onehot, 1)
target = reward + (1.0 - tf.cast(isOver, tf.float32)) * self.gamma * tf.stop_gradient(best_v)
cost = tf.losses.huber_loss(target, pred_action_value,
reduction=tf.losses.Reduction.MEAN)
summary.add_param_summary(('conv.*/W', ['histogram', 'rms']),
('fc.*/W', ['histogram', 'rms'])) # monitor all W
summary.add_moving_summary(cost)
return cost
def optimizer(self):
lr = tf.get_variable('learning_rate', initializer=1e-3, trainable=False)
opt = tf.train.AdamOptimizer(lr, epsilon=1e-3)
return optimizer.apply_grad_processors(
opt, [gradproc.GlobalNormClip(10), gradproc.SummaryGradient()])
@staticmethod
def update_target_param():
"""periodically triggered by trainer"""
vars = tf.global_variables()
ops = []
G = tf.get_default_graph()
for v in vars:
target_name = v.op.name
if target_name.startswith('target'):
new_name = target_name.replace('target/', '')
logger.info("{} <- {}".format(target_name, new_name))
ops.append(v.assign(G.get_tensor_by_name(new_name + ':0')))
return tf.group(*ops, name='update_target_network')