forked from peter1591/hearthstone-ai
-
Notifications
You must be signed in to change notification settings - Fork 0
/
model.py
251 lines (191 loc) · 6.96 KB
/
model.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
#!/usr/bin/python3
import os
import json
import tensorflow as tf
import data_reader
kOutputNodeName = 'final_argmax'
class NextInputGetter:
def __init__(self, data):
self._idx = 0
self._data = data
def get_next_slice(self, size):
ret = tf.slice(self._data, [0, self._idx], [-1, size])
self._idx = self._idx + size
return ret
def get_rest(self):
return tf.slice(self._data, [0, self._idx], [-1, -1])
class Model:
def __init__(self):
self._mode = None
self.kEnableCardIdEmbed = False
self.kEnableHandCardIdEmbed = False
self.kCardIdDimension = 3
self.kMinionConvolutionHidden1 = 10
self.kHeroConvolutionHidden1 = 2
self.kHandConvolutionHidden1 = 2
self.kResidualBlockFeatures = 50
self.kResidualBlocks = 3
def _model_hero(self, input_getter):
inputs = input_getter.get_next_slice(data_reader.kHeroFeatures)
hero1 = tf.layers.dense(
name='hero1',
inputs=inputs,
units=self.kHeroConvolutionHidden1,
reuse=tf.AUTO_REUSE,
activation=tf.nn.relu)
return [hero1]
def _get_embedded_onboard_card_id(self, card_id):
if not self.kEnableCardIdEmbed:
return None
card_id = tf.to_int32(card_id, name='card_id_to_int32')
with tf.variable_scope("onboard", reuse=tf.AUTO_REUSE):
card_id_matrix = tf.get_variable(
'card_embed_matrix',
[data_reader.kMaxCardId, self.kCardIdDimension],
initializer=tf.zeros_initializer())
card_id_embed = tf.nn.embedding_lookup(
card_id_matrix, card_id, name='card_id_embed')
card_id_embed = tf.reshape(card_id_embed, [-1, self.kCardIdDimension])
return card_id_embed
def _model_minion(self, input_getter):
inputs = []
card_id = input_getter.get_next_slice(1)
card_id_embed = self._get_embedded_onboard_card_id(card_id)
if card_id_embed is not None:
inputs.append(card_id_embed)
features = input_getter.get_next_slice(data_reader.kMinionFeatures)
inputs.append(features)
minion1 = tf.layers.dense(
name='minion1',
inputs=tf.concat(inputs, 1),
units=self.kMinionConvolutionHidden1,
reuse=tf.AUTO_REUSE,
activation=tf.nn.relu)
return minion1
def _model_minions(self, input_getter):
features = []
for _ in range(data_reader.kMinions):
features.append(self._model_minion(input_getter))
return features
def _get_embedded_hand_card_id(self, card_id):
if not self.kEnableHandCardIdEmbed:
return None
card_id = tf.to_int32(card_id, name='card_id_to_int32')
with tf.variable_scope("current_hand", reuse=tf.AUTO_REUSE):
card_id_matrix = tf.get_variable(
'card_embed_matrix',
[data_reader.kMaxCardId, self.kCardIdDimension],
initializer=tf.zeros_initializer())
card_id_embed = tf.nn.embedding_lookup(
card_id_matrix, card_id, name='card_id_embed')
card_id_embed = tf.reshape(card_id_embed, [-1, self.kCardIdDimension])
return card_id_embed
def _model_current_hand_card(self, input_getter):
outputs = []
card_id = input_getter.get_next_slice(1)
card_id_embed = self._get_embedded_hand_card_id(card_id)
if card_id_embed is not None:
outputs.append(card_id_embed)
card_features = input_getter.get_next_slice(
data_reader.kCurrentHandCardFeatures)
card1 = tf.layers.dense(
name='hand_card_1',
inputs=card_features,
units=self.kHandConvolutionHidden1,
reuse=tf.AUTO_REUSE,
activation=tf.nn.relu)
outputs.append(card1)
return outputs
def _model_current_hand(self, input_getter):
outputs = []
outputs.append(
input_getter.get_next_slice(data_reader.kCurrentHandFeatures))
for _ in range(data_reader.kCurrentHandCards):
outputs.extend(self._model_current_hand_card(input_getter))
return outputs
def _model_board_features(self, input_getter):
return [input_getter.get_rest()]
def _residual_block_unit(self, scope, inputs, hidden, add_input=None):
with tf.variable_scope(scope):
output = tf.contrib.layers.fully_connected(
inputs,
hidden,
activation_fn=None,
scope='dense')
if add_input is not None:
output = tf.add(output, add_input)
output = tf.nn.relu(output, 'relu')
return output
def _residual_block(self, idx, inputs, hidden):
name = 'residual_' + str(idx)
dense1 = self._residual_block_unit(
name + '_1',
inputs,
hidden)
dense2 = self._residual_block_unit(
name + '_2',
dense1,
hidden,
add_input=inputs)
return dense2
def set_mode(self, mode):
self._mode = mode
def get_model(self, features, labels):
feature = features["x"]
feature = tf.to_float(feature)
input_getter = NextInputGetter(feature)
inputs = []
inputs.extend(self._model_hero(input_getter)) # current hero
inputs.extend(self._model_hero(input_getter)) # opponent hero
inputs.extend(self._model_minions(input_getter)) # current minions
inputs.extend(self._model_minions(input_getter)) # opponent minions
inputs.extend(self._model_current_hand(input_getter)) # current hand
inputs.extend(self._model_board_features(input_getter)) # board features
inputs = tf.concat(inputs, 1)
dense1 = tf.layers.dense(
name='dense1',
inputs=inputs,
units=self.kResidualBlockFeatures,
activation=tf.nn.relu)
prev = dense1
for i in range(0, self.kResidualBlocks):
prev = self._residual_block(i+1, prev, self.kResidualBlockFeatures)
final = tf.layers.dense(
name='final',
inputs=prev,
units=1,
activation=None)
final_argmax = tf.argmax(
name=kOutputNodeName,
input=final,
axis=1)
predictions = {
"classes": final_argmax,
"probabilities": tf.nn.softmax(final, name="softmax_tensor")}
if self._mode == tf.estimator.ModeKeys.PREDICT:
return tf.estimator.EstimatorSpec(
mode=self._mode, predictions=final,
export_outputs={"y": tf.estimator.export.PredictOutput({"y": final})})
labels = tf.reshape(labels, [-1, 1])
loss = tf.losses.mean_squared_error(
labels=labels,
predictions=final)
if self._mode == tf.estimator.ModeKeys.TRAIN:
train_step = tf.train.AdamOptimizer(1e-4).minimize(
loss,
global_step=tf.train.get_global_step())
return tf.estimator.EstimatorSpec(
mode=self._mode,
loss=loss,
train_op=train_step)
else: # EVAL mode
binary_labels = tf.equal(labels, data_reader.kLabelIfFirstPlayerWin)
predictions = tf.greater(final,
data_reader.kLabelFirstPlayerWinIfGreaterThan)
eval_metric_ops = {
"accuracy": tf.metrics.accuracy(
labels=binary_labels,
predictions=predictions)}
return tf.estimator.EstimatorSpec(
mode=self._mode, loss=loss,
eval_metric_ops=eval_metric_ops)