-
Notifications
You must be signed in to change notification settings - Fork 0
/
drqn.py
245 lines (207 loc) · 8.95 KB
/
drqn.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
import tensorflow as tf
import numpy as np
from tensorflow.contrib import rnn
from tensorflow.contrib import slim
def huber_loss(y, t, delta=1.0):
with tf.name_scope('huber_loss', [y, t]):
err = tf.abs(y - t)
q = tf.minimum(err, delta)
return 0.5 * tf.square(q) + delta * (err - q)
class DRQN(object):
def __init__(self,
state_shape,
n_action,
n_steps,
hs=[32,64],
scope='drqn', data_format='NCHW',
is_training=True, reuse=None
):
self._state_shape = state_shape
self._n_action = n_action
self._n_steps = n_steps
self._hs = hs # currently not configurable
self._scope = scope
self._data_format = data_format
self._is_training = is_training
self._reuse = reuse
self._build()
def _arg_scope(self):
batch_norm_params = {
'is_training' : self._is_training,
'decay' : 0.995,
'fused' : True,
'scale' : True,
'reuse' : self._reuse,
'data_format' : self._data_format,
'scope' : 'batch_norm',
}
with slim.arg_scope([slim.fully_connected],
activation_fn = tf.nn.elu,
#weights_regularizer=slim.l2_regularizer(1e-4)
#normalizer_fn = slim.batch_norm,
#normalizer_params = batch_norm_params,
# don't use batch norm, for complicated reasons
) as sc:
return sc
def _build_fcn(self, x):
# [batch_size, 2] -> [batch_size, 64]
with tf.name_scope('fcn', [x]):
with slim.arg_scope(self._arg_scope()):
return slim.stack(x, slim.fully_connected, self._hs, scope='fc')
def _build_cnn(self, x):
with tf.name_scope('cnn', [x]):
with slim.arg_scope(self._arg_scope()):
return NotImplementedError("CNN Not Supported Yet!")
def _build_rnn(self, x, b, s):
# TODO : treat conv vs. fc differently somehow?
with tf.name_scope('rnn', [x]):
n_h = self._hs[-1]
x = slim.flatten(x) # in case input is conv.
x = tf.reshape(x, [b, s, n_h])
with slim.arg_scope(self._arg_scope()):
cell = rnn.BasicLSTMCell(n_h, state_is_tuple=True)
s0 = cell.zero_state(b, tf.float32)
#c_in = s0.c
#h_in = s0.h
c_in = tf.placeholder(
shape = s0.c.shape,
dtype = tf.float32,
name = 'c_in')
h_in = tf.placeholder(
shape = s0.h.shape,
dtype = tf.float32,
name = 'h_in')
s_in = rnn.LSTMStateTuple(c_in, h_in)
y, s_out = tf.nn.dynamic_rnn(
inputs=x,
cell=cell,
dtype=tf.float32,
initial_state=s_in,
scope='rnn')
# y = (batch_size, n_steps, n_actions)
#y = y[:,-1,:]
y = tf.reshape(y, [-1, n_h])
c_out = s_out.c
h_out = s_out.h
#cell = rnn.Conv2DLSTMCell(
# input_shape = (?)
# output_channels = self._n_action,
# kernel_shape = (3,3),
# use_bias=True,
# skip_connection=False,
# initializer=slim.initializers.xavier_initializer()
# )
return c_in, h_in, y, c_out, h_out
def _build_qn(self, x, n_b, n_t):
""" Build Q-Network """
with tf.name_scope('qn', [x]):
with slim.arg_scope(self._arg_scope()):
#xf = slim.fully_connected(x, 128,
# scope='xf', activation_fn=tf.nn.elu)
sa, sv = tf.split(x, 2, axis=1) # split into action-value streams
#adv = slim.fully_connected(sa, 64,
# scope='adv_0', activation_fn=tf.nn.elu)
adv = slim.fully_connected(sa, self._n_action,
scope='adv', activation_fn=None)
#val = slim.fully_connected(sv, 64,
# scope='val_0', activation_fn=tf.nn.elu)
val = slim.fully_connected(sv, 1,
scope='val', activation_fn=None)
q_y = val + (adv - tf.reduce_mean(adv, axis=1, keepdims=True))
a_y = tf.argmax(q_y, axis=1)
return q_y, a_y
def _build_loss(self, q_y, a_y, n_b, n_t):
with tf.name_scope('err', [q_y, a_y]):
# setup targets
# TODO : add eval flag to enable creating loss/evaluation targets
q_t = tf.placeholder(shape=[None], dtype=tf.float32, name='q_t')
a_t = tf.placeholder(shape=[None], dtype=tf.int32, name='a_t')
a_t_o = tf.one_hot(a_t, self._n_action, dtype=tf.float32)
q = tf.reduce_sum(q_y * a_t_o, axis=1)
decay = 0.99
q_m = tf.Variable(initial_value=1.0, trainable=False)
q_s = tf.Variable(initial_value=1.0, trainable=False)
#q_t_m, q_t_v = tf.nn.moments(q_t, axes=[0])
#q_t_s = tf.sqrt(q_t_v)
#u_m = q_m.assign(q_m*decay + q_t_m*(1.0-decay)) # offset
#u_s = q_s.assign(q_s*decay + q_t_s*(1.0-decay)) # scale
#with tf.control_dependencies([u_m, u_s]):
# q_n = (q - q_m) / (2.0 * q_s)
# q_t_n = (q_t - q_m) / (2.0 * q_s)
# q_err = huber_loss(q_n, q_t_n)
# OPT1 . relative error
q_s = tf.Variable(initial_value=1.0, trainable=False)
q_t_s = tf.reduce_mean(q_t)
#with tf.control_dependencies([q_s.assign(q_s*decay + q_t_s*(1.0-decay))]):
# q_n = q / q_s
# q_t_n = q_t / q_s
# q_err = huber_loss(q_n, q_t_n)
# OPT2.0 absolute error huber
q_err = huber_loss(q, q_t)
# OPT2.1 absolute error square
#q_err = tf.square(q_t-q)
# only the latter steps will be counted for loss ...
n_mask = tf.maximum(n_t//2, 1)
m_a = tf.zeros([n_b, n_t - n_mask], dtype=tf.float32)
m_b = tf.ones([n_b, n_mask], dtype=tf.float32)
mask = tf.concat([m_a, m_b], 1)
mask = tf.reshape(mask, [-1])
loss = tf.reduce_mean(q_err * mask)
return q, q_t, a_t, loss, q_s
#def _popart(self, y, t):
# with tf.name_scope('popart', [y,t]):
# t_n = tf.matmul(tf.inv(sigma), y - mu)
# # y = [n_b]
# #W = tf.eye(...)
def _build(self):
with tf.variable_scope(self._scope, reuse=self._reuse):
x_in = tf.placeholder(tf.float32, shape=[None] + list(self._state_shape), name='x_in')
batch_size = tf.placeholder(tf.int32, shape=[], name='n_b')
step_size = tf.placeholder(tf.int32, shape=[], name='n_t')
#cnn = self._build_cnn(self._inputs)
fcn = self._build_fcn(x_in)
c_in, h_in, y, c_out, h_out = self._build_rnn(fcn, batch_size, step_size)
q_y, a_y = self._build_qn(y, batch_size, step_size)
q, q_t, a_t, loss, q_s = self._build_loss(q_y, a_y, batch_size, step_size)
# save ; inputs
self._inputs = {
'n_b' : batch_size,
'n_t' : step_size,
'x_in' : x_in, # actual env input
'c_in' : c_in, # states bookkeeping
'h_in' : h_in, #
'q_t' : q_t, # target q
'a_t' : a_t # target action
}
self._outputs = {
'y' : y, # rnn output
'c_out' : c_out, # states bookkeeping
'h_out' : h_out, #
'q' : q,
'q_y' : q_y, # network q
'a_y' : a_y, # network action
'loss' : loss,
'q_s' : q_s
}
# create tensors lookup dictionary
self._tensors = self._inputs.copy()
self._tensors.update(self._outputs)
def __getitem__(self, name):
return self._tensors[name]
def predict(self):
return NotImplementedError("DRQN.predict() does not exist yet.")
def train(self):
return NotImplementedError("DRQN.train() does not exist yet.")
def get_trainable_variables(self):
return slim.get_trainable_variables(self._scope)
def main():
# test ...
drqn_a = DRQN([2], 2, 8, scope='actor')
drqn_c = DRQN([2], 2, 8, scope='critic')
va = drqn_a.get_trainable_variables()
vc = drqn_c.get_trainable_variables()
tau = 0.001
copy_ops = [c.assign(a.value()*tau + c.value() * (1.0-tau)) for (a,c) in zip(va,vc)]
copy_ops = tf.group(copy_ops)
if __name__ == "__main__":
main()