-
Notifications
You must be signed in to change notification settings - Fork 56
/
Copy pathpaddle_net.py
176 lines (155 loc) · 7 KB
/
paddle_net.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
"""策略价值网络"""
import paddle
import paddle.nn as nn
import numpy as np
import paddle.nn.functional as F
# 搭建残差块
class ResBlock(nn.Layer):
def __init__(self, num_filters=256):
super().__init__()
self.conv1 = nn.Conv2D(in_channels=num_filters, out_channels=num_filters, kernel_size=3, stride=1, padding=1)
self.conv1_bn = nn.BatchNorm2D(num_features=num_filters)
self.conv1_act = nn.ReLU()
self.conv2 = nn.Conv2D(in_channels=num_filters, out_channels=num_filters, kernel_size=3, stride=1, padding=1)
self.conv2_bn = nn.BatchNorm2D(num_features=num_filters)
self.conv2_act = nn.ReLU()
def forward(self, x):
y = self.conv1(x)
y = self.conv1_bn(y)
y = self.conv1_act(y)
y = self.conv2(y)
y = self.conv2_bn(y)
y = x + y
return self.conv2_act(y)
# 搭建骨干网络，输入：N, 9, 10, 9 --> N, C, H, W
class Net(nn.Layer):
def __init__(self, num_channels=256, num_res_blocks=13):
super().__init__()
# 初始化特征
self.conv_block = nn.Conv2D(in_channels=9, out_channels=num_channels, kernel_size=3, stride=1, padding=1)
self.conv_block_bn = nn.BatchNorm2D(num_features=256)
self.conv_block_act = nn.ReLU()
# 全局特征
self.global_conv = nn.Conv2D(in_channels=9, out_channels=512, kernel_size=(10, 9))
self.global_bn = nn.BatchNorm1D(512)
# 残差块抽取特征
self.res_blocks = nn.LayerList([ResBlock(num_filters=num_channels) for _ in range(num_res_blocks)])
# 策略头
self.global_policy_fc = nn.Linear(512, 2086)
self.policy_conv = nn.Conv2D(in_channels=num_channels, out_channels=16, kernel_size=1, stride=1)
self.policy_bn = nn.BatchNorm2D(16)
self.policy_act = nn.ReLU()
self.policy_fc = nn.Linear(16 * 9 * 10, 2086)
# 价值头
self.global_value_fc = nn.Linear(512, 256)
self.value_conv = nn.Conv2D(in_channels=num_channels, out_channels=8, kernel_size=1, stride=1)
self.value_bn = nn.BatchNorm2D(8)
self.value_act1 = nn.ReLU()
self.value_fc1 = nn.Linear(8 * 9 * 10, 256)
self.value_act2 = nn.ReLU()
self.value_fc2 = nn.Linear(256, 1)
# 定义前向传播
def forward(self, x):
# 公共头
global_x = self.global_conv(x)
global_x = paddle.reshape(global_x, [-1, 512])
global_x = self.global_bn(global_x)
x = self.conv_block(x)
x = self.conv_block_bn(x)
x = self.conv_block_act(x)
for layer in self.res_blocks:
x = layer(x)
# 策略头
policy = self.policy_conv(x)
policy = self.policy_bn(policy)
policy = self.policy_act(policy)
policy = paddle.reshape(policy, [-1, 16 * 10 * 9])
policy = self.policy_fc(policy)
global_policy = self.policy_act(self.global_policy_fc(global_x))
policy = F.log_softmax(policy + global_policy)
# 价值头
value = self.value_conv(x)
value = self.value_bn(value)
value = self.value_act1(value)
value = paddle.reshape(value, [-1, 8 * 10 * 9])
global_value = self.value_act1(self.global_value_fc(global_x))
value = self.value_fc1(value)
value = self.value_act1(value)
value = self.value_fc2(value + global_value)
value = F.tanh(value)
return policy, value
# 策略值网络，用来进行模型的训练
class PolicyValueNet:
def __init__(self, model_file=None, use_gpu=True):
self.use_gpu = use_gpu
self.l2_const = 2e-3 # l2 正则化
self.policy_value_net = Net()
self.optimizer = paddle.optimizer.Adam(learning_rate=0.001,
parameters=self.policy_value_net.parameters(),
weight_decay=self.l2_const)
if model_file:
net_params = paddle.load(model_file)
self.policy_value_net.set_state_dict(net_params)
# 输入一个批次的状态，输出一个批次的动作概率和状态价值
def policy_value(self, state_batch):
self.policy_value_net.eval()
state_batch = paddle.to_tensor(state_batch)
log_act_probs, value = self.policy_value_net(state_batch)
act_probs = np.exp(log_act_probs.numpy())
return act_probs, value.numpy()
# 输入棋盘，返回每个合法动作的（动作，概率）元组列表，以及棋盘状态的分数
def policy_value_fn(self, board):
self.policy_value_net.eval()
# 获取合法动作列表
legal_positions = board.availables
current_state = np.ascontiguousarray(board.current_state().reshape(-1, 9, 10, 9)).astype('float32')
current_state = paddle.to_tensor(current_state)
# 使用神经网络进行预测
log_act_probs, value = self.policy_value_net(current_state)
act_probs = np.exp(log_act_probs.numpy().flatten())
# 只取出合法动作
act_probs = zip(legal_positions, act_probs[legal_positions])
# 返回动作概率，以及状态价值
return act_probs, value.numpy()
# 得到模型参数
def get_policy_param(self):
net_params = self.policy_value_net.state_dict()
return net_params
# 保存模型
def save_model(self, model_file):
net_params = self.get_policy_param() # 取得模型参数
paddle.save(net_params, model_file)
# 执行一步训练
def train_step(self, state_batch, mcts_probs, winner_batch, lr=0.002):
self.policy_value_net.train()
# 包装变量
state_batch = paddle.to_tensor(state_batch)
mcts_probs = paddle.to_tensor(mcts_probs)
winner_batch = paddle.to_tensor(winner_batch)
# 清零梯度
self.optimizer.clear_gradients()
# 设置学习率
self.optimizer.set_lr(lr)
# 前向运算
log_act_probs, value = self.policy_value_net(state_batch)
value = paddle.reshape(x=value, shape=[-1])
# 价值损失
value_loss = F.mse_loss(input=value, label=winner_batch)
# 策略损失
policy_loss = -paddle.mean(paddle.sum(mcts_probs * log_act_probs, axis=1)) # 希望两个向量方向越一致越好
# 总的损失，注意l2惩罚已经包含在优化器内部
loss = value_loss + policy_loss
# 反向传播及优化
loss.backward()
self.optimizer.minimize(loss)
# 计算策略的熵，仅用于评估模型
entropy = -paddle.mean(
paddle.sum(paddle.exp(log_act_probs) * log_act_probs, axis=1)
)
return loss.numpy(), entropy.numpy()[0]
if __name__ == '__main__':
net = Net()
test_data = paddle.ones([8, 9, 10, 9])
x_act, x_val = net(test_data)
print(x_act.shape) # 8, 2086
print(x_val.shape) # 8, 1