-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathTrain.py
149 lines (125 loc) · 5.3 KB
/
Train.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
from keras.backend import set_session
from keras.regularizers import L2
from keras.optimizers import SGD
from keras.layers import Input, Dense, Conv2D, Flatten, BatchNormalization, ReLU, add, Reshape, Softmax
from keras.models import load_model, Model
import random
import numpy as np
import tensorflow as tf
import os
import struct
modelPathName = 'model'
trainPathName = 'dataset'
def ConvolutionalLayer(x, filters=64):
x = Conv2D(filters=filters, kernel_size=3, data_format='channels_first',
padding='same', use_bias=False, activation='linear', kernel_regularizer=L2(0.0001))(x)
x = BatchNormalization(axis=1)(x)
x = ReLU()(x)
return x
def ResidualLayer(in_x, filters=64):
x = ConvolutionalLayer(in_x, filters)
x = Conv2D(filters=filters, kernel_size=3, data_format='channels_first',
padding='same', use_bias=False, activation='linear', kernel_regularizer=L2(0.0001))(x)
x = BatchNormalization(axis=1)(x)
x = add([in_x, x])
x = ReLU()(x)
return x
def ValueHead(x, units=64):
x = Conv2D(filters=1, kernel_size=1, data_format='channels_first', padding='same',
use_bias=False, activation='linear', kernel_regularizer=L2(0.0001))(x)
x = BatchNormalization(axis=1)(x)
x = ReLU()(x)
x = Flatten()(x)
x = Dense(units, use_bias=False, activation='linear',
kernel_regularizer=L2(0.0001))(x)
x = ReLU()(x)
x = Dense(1, name='valueHead', use_bias=False, activation='tanh',
kernel_regularizer=L2(0.0001))(x)
return x
def PolicyHead(x, w, h):
x = Conv2D(
filters=1, kernel_size=1, data_format='channels_first', padding='same', use_bias=False, activation='linear', kernel_regularizer=L2(0.0001))(x)
x = Flatten()(x)
x = Softmax(name='policyHead')(x)
return x
def BuildModel(channel, size) -> Model:
input = Input(shape=(channel, size, size))
x = ConvolutionalLayer(input)
for i in range(5):
x = ResidualLayer(x)
x = ConvolutionalLayer(x)
valueHead = ValueHead(x)
policyHead = PolicyHead(x, size, size)
model = Model(inputs=[input], outputs=[valueHead, policyHead])
model.compile(
loss={'valueHead': 'mean_squared_error',
'policyHead': 'categorical_crossentropy'},
loss_weights={'valueHead': 0.5, 'policyHead': 0.5},
optimizer=SGD(learning_rate=0.01, momentum=0.9))
return model
def writeConv(outfile, layer):
for x in np.array(layer.weights).transpose(0, 4, 3, 1, 2).flatten():
outfile.write(struct.pack('f', x))
def writeNormalize(outfile, layer):
weights = np.array(layer.weights)
for i in [1, 0, 2, 3]:
for x in weights[i].flatten():
outfile.write(struct.pack('f', x))
def writeDense(outfile, layer):
for x in np.array(layer.weights).transpose(0, 2, 1).flatten():
outfile.write(struct.pack('f', x))
def saveModel(filePath):
with open(filePath, 'wb') as outfile:
outfile.write(struct.pack('iiiii', 0, 0, 0, 0, 0))
writeNormalize(outfile, model.layers[2])
writeConv(outfile, model.layers[1])
for i in range(5):
writeNormalize(outfile, model.layers[i * 7 + 5])
writeConv(outfile, model.layers[i * 7 + 4])
writeNormalize(outfile, model.layers[i * 7 + 8])
writeConv(outfile, model.layers[i * 7 + 7])
writeNormalize(outfile, model.layers[40])
writeConv(outfile, model.layers[39])
writeNormalize(outfile, model.layers[43])
writeConv(outfile, model.layers[42])
for i in range(np.array(model.layers[46].weights).shape[2]):
outfile.write(struct.pack('f', 0))
writeDense(outfile, model.layers[46])
for i in range(np.array(model.layers[50].weights).shape[2]):
outfile.write(struct.pack('f', 0))
writeDense(outfile, model.layers[50])
outfile.write(struct.pack('f', 0))
writeConv(outfile, model.layers[47])
print('Loading model...')
if os.path.exists(modelPathName):
model = load_model(modelPathName)
else:
model = BuildModel(6, 9)
model.save(modelPathName)
saveModel('network.weights')
print('Loading data...')
trainDatas = []
for file in os.listdir(trainPathName):
if os.path.exists(os.path.join(trainPathName, file)):
trainDatas.append(np.loadtxt(os.path.join(trainPathName, file)))
trainDatas = np.concatenate(trainDatas, axis=0)
print('Start training...')
for i in range(4):
trainData = trainDatas[np.random.choice(
trainDatas.shape[0], 256 * 1000, replace=False)]
trainInput = trainData[:, :6*9*9].reshape(-1, 6, 9, 9)
tarinValueOutput = trainData[:, 7*9*9:9*7*9+1].reshape(-1, 1)
tarinPolicyOutput = trainData[:, 6*9*9:7*9*9].reshape(-1, 9, 9)
for i in range(len(trainData)):
times = random.randint(1, 4)
trainInput[i] = np.rot90(trainInput[i], times, [1, 2])
tarinPolicyOutput[i] = np.rot90(tarinPolicyOutput[i], times, [0, 1])
if random.choice([True, False]):
trainInput[i] = np.flip(trainInput[i], [1])
tarinPolicyOutput[i] = np.flip(tarinPolicyOutput[i], [0])
tarinPolicyOutput = tarinPolicyOutput.reshape(-1, 81)
model.fit(trainInput, [tarinValueOutput,
tarinPolicyOutput], epochs=1, batch_size=256)
print('Save model...')
model.save(modelPathName)
saveModel('network.weights')