/
ann.py
310 lines (234 loc) · 12.2 KB
/
ann.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
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
import tensorflow as tf
from tensorflow.contrib.layers import apply_regularization, l1_l2_regularizer
import numpy as np
RMSPROP_DECAY = 0.9 # Decay term for RMSProp.
RMSPROP_MOMENTUM = 0.9 # Momentum in RMSProp.
RMSPROP_EPSILON = 1.0 # Epsilon term for RMSProp.
def get_param_name(s):
return s.split('/')[1].split(':')[0]
def get_scope_name(s):
return s.split('/')[0].split(':')[0]
class ann(object):
def __init__(self, scope, input_size, output_size, summary_writer):
print "going to initialize scope %s" % scope
self.summary_writer = summary_writer
self.scope = scope
with tf.variable_scope(scope) as vscope:
self.vscope = vscope
self.do_init(input_size, output_size)
print "scope %s has been initialized" % scope
def init_neurons(self, input_layer, wname, wnum, bias_name=None):
ishape = input_layer.get_shape()[1].value
dims = [ishape, wnum]
w = tf.get_variable(wname,
initializer=tf.random_normal(dims),
regularizer=l1_l2_regularizer(scale_l1=self.reg_beta, scale_l2=self.reg_beta),
dtype=tf.float32)
sw = tf.summary.histogram(wname, w)
self.summary_weights.append(sw)
self.transform_params[wname] = w
wext = tf.placeholder(tf.float32, dims, name=wname+'_ext')
w_transform_ops = w.assign(w * (1 - self.transform_lr) + wext * self.transform_lr)
self.transform_ops.append(w_transform_ops)
h = tf.matmul(input_layer, w)
if bias_name:
b = tf.get_variable(bias_name, initializer=tf.random_normal([wnum]), dtype=tf.float32)
sb = tf.summary.histogram(bias_name, b)
self.summary_weights.append(sb)
self.transform_params[bias_name] = b
bext = tf.placeholder(tf.float32, [wnum], name=bias_name+'_ext')
b_transform_ops = b.assign(b * (1 - self.transform_lr) + bext * self.transform_lr)
self.transform_ops.append(b_transform_ops)
h = tf.add(h, b)
return h
def init_layer(self, input_layer, wname, wnum, nonlinear, bias_name=None):
h = self.init_neurons(input_layer, wname, wnum, bias_name)
if nonlinear:
return nonlinear(h)
return h
def init_model(self, input_size, output_size):
layers = [('w0', 256), ('w1', 64)]
print "init_model scope: %s" % (tf.get_variable_scope().name)
x = tf.placeholder(tf.float32, [None, input_size], name='x')
action = tf.placeholder(tf.int32, [None, 1], name='action')
reward = tf.placeholder(tf.float32, [None, 1], name='reward')
self.add_summary(tf.summary.histogram('action', action))
self.add_summary(tf.summary.histogram('reward', reward))
input_dimension = input_size
input_layer = x
idx = 0
for wname, wnum in layers:
input_layer = self.init_layer(input_layer, wname, wnum, tf.nn.elu, 'b%d'%(idx))
#input_layer = self.init_layer(input_layer, l, tf.nn.tanh)
self.add_summary(tf.summary.histogram('h%d' % idx, input_layer))
idx += 1
self.policy = self.init_layer(input_layer, 'policy', output_size, tf.nn.softmax)
self.value = self.init_layer(input_layer, 'value', 1, None)
log_policy = tf.log(1e-6 + self.policy)
actions = tf.one_hot(action, output_size)
actions = tf.squeeze(actions, 1)
log_probability_per_action = tf.reduce_sum(tf.multiply(log_policy, actions), axis=1)
advantage = (reward - self.value)
self.add_summary(tf.summary.scalar("advantage_mean", tf.reduce_mean(advantage)))
self.add_summary(tf.summary.scalar("advantage_rms", tf.sqrt(tf.reduce_mean(tf.square(advantage)))))
self.cost_policy = -log_probability_per_action * advantage
self.add_summary(tf.summary.scalar("cost_policy_mean", tf.reduce_mean(self.cost_policy)))
self.add_summary(tf.summary.scalar("cost_policy_rms", tf.sqrt(tf.reduce_mean(tf.square(self.cost_policy)))))
self.cost_value = tf.square(self.value - reward)
self.add_summary(tf.summary.scalar("cost_value_mean", tf.reduce_mean(self.cost_value)))
self.add_summary(tf.summary.scalar("input_reward_mean", tf.reduce_mean(reward)))
self.add_summary(tf.summary.scalar("value_mean", tf.reduce_mean(self.value)))
#reg_loss = tf.add_n(tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES))
#self.add_summary(tf.summary.scalar("reg_loss", reg_loss))
#self.cost = tf.add_n([self.cost_value, reg_loss])
#self.add_summary(tf.summary.scalar("cost", self.cost))
#self.cost = self.cost_policy
def add_summary(self, s):
self.summary_all.append(s)
def setup_gradients(self, prefix, opt, cost):
grads = opt.compute_gradients(cost)
ret_grads = []
ret_names = []
ret_apply = []
for e in grads:
grad, var = e
if grad is None or var is None:
continue
#print "var: %s, gradient: %s" % (var, grad)
if self.scope != get_scope_name(var.name):
continue
pname = get_param_name(var.name)
gname = '%s/gradient_%s' % (prefix, pname)
print "gradient %s -> %s" % (var, gname)
# get all gradients
ret_grads.append(grad)
ret_names.append(gname)
pl = tf.placeholder(tf.float32, shape=var.get_shape(), name=gname)
clip = tf.clip_by_average_norm(pl, 1)
ret_apply.append((clip, var))
ag = tf.summary.histogram('%s/%s/apply_%s'% (self.scope, prefix, gname), clip)
self.summary_apply_gradients.append(ag)
return ret_grads, ret_names, ret_apply
def do_init(self, input_size, output_size):
self.learning_rate_start = 0.0025
self.reg_beta_start = 0.001
self.transform_lr_start = 1.0
self.train_num = 0
self.transform_ops = []
self.transform_params = {}
self.summary_all = []
self.summary_weights = []
self.episode_stats_update = []
self.summary_apply_gradients = []
global_step = tf.get_variable('global_step', [], initializer=tf.constant_initializer(0), trainable=False)
self.transform_lr = 0.00001 + tf.train.exponential_decay(self.transform_lr_start, global_step, 30000, 0.6, staircase=True)
self.learning_rate = 0.00001 + tf.train.exponential_decay(self.learning_rate_start, global_step, 30000, 0.6, staircase=True)
self.reg_beta = tf.train.exponential_decay(self.reg_beta_start, global_step, 30000, 0.6, staircase=True)
self.add_summary(tf.summary.scalar('reg_beta', self.reg_beta))
self.add_summary(tf.summary.scalar('transform_lr', self.transform_lr))
self.add_summary(tf.summary.scalar('learning_rate', self.learning_rate))
self.add_summary(tf.summary.scalar('global_step', global_step))
episodes_passed_p = tf.placeholder(tf.int32, [], name='episodes_passed')
episode_reward_p = tf.placeholder(tf.float32, [], name='episode_reward')
self.episodes_passed = tf.get_variable('episodes_passed', [], initializer=tf.constant_initializer(0),
trainable=False, dtype=tf.int32)
self.episode_stats_update.append(self.episodes_passed.assign(episodes_passed_p))
self.episode_reward = tf.get_variable('episode_reward', [], initializer=tf.constant_initializer(0), trainable=False)
self.episode_stats_update.append(self.episode_reward.assign(episode_reward_p))
self.add_summary(tf.summary.scalar('episodes_passed', self.episodes_passed))
self.add_summary(tf.summary.scalar('episode_reward', self.episode_reward))
self.init_model(input_size, output_size)
opt = tf.train.RMSPropOptimizer(self.learning_rate,
RMSPROP_DECAY,
momentum=RMSPROP_MOMENTUM,
epsilon=RMSPROP_EPSILON, name='optimizer')
self.gradient_names_policy = []
self.apply_grads_policy = []
self.gradient_names_value = []
self.apply_grads_value = []
self.compute_gradients_step_policy, self.gradient_names_policy, self.apply_grads_policy = self.setup_gradients("policy", opt, self.cost_policy)
self.compute_gradients_step_value, self.gradient_names_value, self.apply_grads_value = self.setup_gradients("value", opt, self.cost_value)
apply_gradients = self.apply_grads_policy + self.apply_grads_value
self.apply_gradients_step = opt.apply_gradients(apply_gradients, global_step=global_step)
config=tf.ConfigProto(
intra_op_parallelism_threads = 8,
inter_op_parallelism_threads = 8,
)
self.sess = tf.Session(config=config)
self.summary_weights_merged = tf.summary.merge(self.summary_weights)
self.summary_merged = tf.summary.merge(self.summary_all)
self.summary_apply_gradients_merged = tf.summary.merge(self.summary_apply_gradients)
init = [tf.global_variables_initializer(), tf.local_variables_initializer()]
self.sess.run(init)
def update_gradients(self, states, dret, names, grads):
for gname, grad in zip(names, grads):
#value = np.sum(grad) / float(len(states))
value = grad / float(len(states)) * 2.0
g = dret.get(gname)
if not g:
dret[gname] = value
else:
g.update(value)
#print "computed gradients %s, shape: %s" % (gname, grad.shape)
#print grad
def compute_gradients(self, states, action, reward):
self.train_num += 1
ops = [self.summary_merged, self.compute_gradients_step_policy, self.compute_gradients_step_value]
summary, grads_policy, grads_value = self.sess.run(ops, feed_dict={
self.scope + '/x:0': states,
self.scope + '/action:0': action,
self.scope + '/reward:0': reward,
})
self.summary_writer.add_summary(summary, self.train_num)
dret = {}
self.update_gradients(states, dret, self.gradient_names_policy, grads_policy)
self.update_gradients(states, dret, self.gradient_names_value, grads_value)
return dret
def apply_gradients(self, grads):
if len(grads) == 0:
print "empty gradients to apply"
return
feed_dict = {}
#print "apply: %s" % grads
for n, g in grads.iteritems():
gname = self.scope + '/' + n + ':0'
#print "apply gradients to %s" % (gname)
#print g
feed_dict[gname] = g
ops = [self.summary_weights_merged, self.summary_apply_gradients_merged, self.apply_gradients_step]
summary_weights, summary_apply, grads = self.sess.run(ops, feed_dict=feed_dict)
self.summary_writer.add_summary(summary_weights, self.train_num)
self.summary_writer.add_summary(summary_apply, self.train_num)
def predict_policy(self, states):
p = self.sess.run([self.policy], feed_dict={
self.scope + '/x:0': states,
})
return p[0]
def predict_value(self, states):
p = self.sess.run([self.value], feed_dict={
self.scope + '/x:0': states,
})
return p[0]
def predict_both(self, states):
p = self.sess.run([self.policy, self.value], feed_dict={
self.scope + '/x:0': states,
})
return p
def export_params(self):
return self.sess.run(self.transform_params)
def transform(self, x1, x2):
lr = 0.9
return x1 * lr + x2 * (1 - lr)
def import_params(self, d):
self.train_num += 1
d1 = {}
for k, v in d.iteritems():
d1[self.scope + '/' + k + '_ext:0'] = v
self.sess.run(self.transform_ops, feed_dict=d1)
summary = self.sess.run([self.summary_weights_merged])
self.summary_writer.add_summary(summary[0], self.train_num)
def update_episode_stats(self, episodes, reward):
summary = self.sess.run(self.episode_stats_update, feed_dict={
self.scope + '/episodes_passed:0': episodes,
self.scope + '/episode_reward:0': reward,
})