Deep-Computer-Go/model.py at main · vesran/Deep-Computer-Go · GitHub

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
import numpy as np
import tensorflow as tf
import tensorflow.keras as keras
from tensorflow.keras import layers
from tensorflow.keras import regularizers

import golois


##################################################################
# Parameters
##################################################################

PLANES = 21
MOVES = 361
N = 100_000
DIM = 19


##################################################################
# Blocks
##################################################################

def shaping_data():
    input_data = np.random.randint(2, size=(N, DIM, DIM, PLANES))
    input_data = input_data.astype('float32')

    policy = np.random.randint(MOVES, size=(N,))
    policy = keras.utils.to_categorical(policy)

    value = np.random.randint(2, size=(N,))
    value = value.astype('float32')

    end = np.random.randint(2, size=(N, DIM, DIM, 2))
    end = end.astype('float32')

    groups = np.zeros((N, DIM, DIM, 1))
    groups = groups.astype('float32')

    return input_data, policy, value, end, groups


def hswish(x):
    return x * tf.nn.relu6(x+3) * 0.166666666667


def se_block(in_block, ch, ratio=16):
    x = layers.Dropout(0.2)(in_block)
    x = layers.GlobalAveragePooling2D()(x)
    x = layers.Dense(ch//ratio, activation='relu')(x)
    x = layers.Dense(ch, activation='sigmoid')(x)
    return layers.Multiply()([in_block, x])


def channel_attention_module(in_block, filters, ratio):
    maxp = layers.GlobalMaxPooling2D()(in_block)
    avgp = layers.GlobalAveragePooling2D()(in_block)

    hidden_ff = layers.Dense(filters // ratio, activation='relu')
    out_ff = layers.Dense(filters)

    maxp = hidden_ff(maxp)
    maxp = out_ff(maxp)

    avgp = hidden_ff(avgp)
    avgp = out_ff(avgp)

    add_x = layers.add([maxp, avgp])
    activ_x = layers.Activation('sigmoid')(add_x)
    return layers.Multiply()([in_block, activ_x])


def spatial_attention_module(in_block):
    maxp = layers.Lambda(lambda x: tf.keras.backend.mean(x, axis=3, keepdims=True))(in_block)
    avgp = layers.Lambda(lambda x: tf.keras.backend.max(x, axis=3, keepdims=True))(in_block)

    max_avg = layers.Concatenate()([maxp, avgp])

    conv_x = layers.Conv2D(1, (7, 7), padding='same')(max_avg)

    activ_x = layers.Activation('sigmoid')(conv_x)
    return layers.Multiply()([in_block, activ_x])


def residual_block(x, filters):
    x1 = layers.Conv2D(filters, (3, 3), padding='same')(x)
    x1 = layers.BatchNormalization()(x1)
    x1 = hswish(x1)

    x1 = layers.Conv2D(filters, (3, 3), padding='same')(x1)
    x1 = layers.BatchNormalization()(x1)

    x1 = se_block(x1, filters, ratio=4)

    x = layers.add([x1, x])
    x = hswish(x)
    x = layers.BatchNormalization()(x)

    return x


def input_block(filters, inp):
    x = layers.Conv2D(filters, (3, 3), padding='same')(inp)
    x = layers.BatchNormalization()(x)
    x = hswish(x)
    x1 = layers.Conv2D(filters, (5, 5), padding='same')(inp)
    x1 = layers.BatchNormalization()(x1)
    x1 = hswish(x1)
    x = layers.add([x, x1])
    return x


def output_policy_block(policy_head):
    policy_head = layers.Conv2D(1, 1, padding='same', use_bias=False, kernel_regularizer=regularizers.l2(0.0001))(
        policy_head)
    policy_head = hswish(policy_head)
    policy_head = layers.BatchNormalization()(policy_head)
    policy_head = layers.Flatten()(policy_head)
    policy_head = layers.Activation('softmax', name='policy')(policy_head)
    return policy_head


def output_value_block(value_head):
    value_head = layers.GlobalAveragePooling2D()(value_head)
    value_head = layers.Dense(64, kernel_regularizer=regularizers.l2(0.0001))(value_head)
    value_head = hswish(value_head)
    value_head = layers.BatchNormalization()(value_head)
    value_head = layers.Dropout(0.2)(value_head)
    value_head = layers.Dense(1, activation='sigmoid', name='value', kernel_regularizer=regularizers.l2(0.0001))(
        value_head)
    return value_head


def create_model(filters, num_blocks=6):
    inp = tf.keras.Input(shape=(19, 19, PLANES), name='board')
    x = input_block(filters, inp)

    for i in range(num_blocks):
        x = residual_block(x, filters)

    policy_head = output_policy_block(x)
    value_head = output_value_block(x)

    model = keras.Model(inputs=inp, outputs=[policy_head, value_head])
    return model