-
Notifications
You must be signed in to change notification settings - Fork 13
/
Copy pathpolicy_value_net.py
202 lines (175 loc) · 7.78 KB
/
policy_value_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
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
import torch.nn as nn
import torch
import math
import numpy as np
import torch.nn.functional as F
import torch.optim as optim
from torch.autograd import Variable
def set_learning_rate(optimizer, lr):
"""设置学习率"""
for param_group in optimizer.param_groups:
param_group['lr'] = lr
def conv3x3(in_planes, out_planes, stride=1):
"""3x3 convolution with padding"""
return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride,
padding=1, bias=False)
class BasicBlock(nn.Module):
def __init__(self, inplanes, planes, stride=1, downsample=None):
super(BasicBlock, self).__init__()
self.conv1 = conv3x3(inplanes, planes, stride)
self.bn1 = nn.BatchNorm2d(planes)
self.relu = nn.ReLU(inplace=True)
self.conv2 = conv3x3(planes, planes)
self.bn2 = nn.BatchNorm2d(planes)
self.downsample = downsample
self.stride = stride
def forward(self, x):
residual = x
# print(residual.size())
out = self.conv1(x)
out = self.bn1(out)
out = self.relu(out)
out = self.conv2(out)
out = self.bn2(out)
# print(out.size())
if self.downsample is not None:
residual = self.downsample(x)
out += residual
out = self.relu(out)
return out
class policy_value_net(nn.Module):
def __init__(self, block, inplanes, planes, stride=1):
super(policy_value_net, self).__init__()
self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=3, stride=stride,
padding=1, bias=False)
self.bn1 = nn.BatchNorm2d(planes)
self.relu = nn.ReLU(inplace=True)
self.res1 = block(planes, planes)
self.res2 = block(planes, planes)
self.res3 = block(planes, planes)
self.res4 = block(planes, planes)
self.res5 = block(planes, planes)
# 价值头
self.conv2 = nn.Conv2d(64, 4, kernel_size=3, stride=stride,
padding=1, bias=False)
self.bn2 = nn.BatchNorm2d(4)
self.fc1 = nn.Linear(324, 128)
self.fc2 = nn.Linear(128, 1)
# 策略头
self.conv3 = nn.Conv2d(64, 2, kernel_size=3, stride=stride,
padding=1, bias=False)
self.bn3 = nn.BatchNorm2d(2)
self.fc3 = nn.Linear(162, 140)
def forward(self,x):
out = self.conv1(x)
out = self.bn1(out)
out = self.relu(out)
out = self.res1(out)
out = self.res2(out)
out = self.res3(out)
out = self.res4(out)
out = self.res5(out)
# 价值头
value_out = self.conv2(out)
value_out = self.bn2(value_out)
value_out = self.relu(value_out)
value_out = value_out.view(-1, 324)
value_out = self.fc1(value_out)
value_out = F.tanh(self.fc2(value_out))
# 策略头
policy_out = self.conv3(out)
policy_out = self.bn3(policy_out)
policy_out = self.relu(policy_out)
policy_out = policy_out.view(-1, 162)
# policy_out = F.log_softmax(self.fc3(policy_out), dim=1) # softmax+log pytorch 0.4 支持dim
policy_out = F.log_softmax(self.fc3(policy_out))
return policy_out,value_out
# device = torch.device("cpu")
# x = torch.Tensor(4,26,9,9).to(device)
# net = policy_value_net(BasicBlock,26,64).to(device)
# print(net)
# value_out,policy_out = net(x)
# print("value:",value_out.size())
# print("policy:",policy_out.size())
class PolicyValueNet(object):
"""策略价值网络 """
def __init__(self,model_file=None, use_gpu=True):
self.use_gpu = use_gpu
self.l2_const = 1e-4 # 正则化系数
if self.use_gpu:
# device = torch.device("cuda:0")
self.policy_value_net = policy_value_net(BasicBlock,26,64).cuda()
else:
# device = torch.device("cpu")
self.policy_value_net = policy_value_net(BasicBlock,26,64)
self.optimizer = optim.Adam(self.policy_value_net.parameters(), weight_decay=self.l2_const)
if model_file:
self.policy_value_net.load_state_dict(torch.load('ckpt/%s.pth'% model_file))
def policy_value(self, state_batch):
"""
输入：一批次的状态
输出：一批次的落子概率和状态价值
"""
if self.use_gpu:
# device = torch.device("cuda:0")
state_batch = Variable(torch.FloatTensor(state_batch).cuda())
log_act_probs, value = self.policy_value_net(state_batch)
act_probs = np.exp(log_act_probs.data.cpu().numpy())
return act_probs, value.data.cpu().numpy()
else:
device = torch.device("cpu")
state_batch = Variable(torch.FloatTensor(state_batch))
log_act_probs, value = self.policy_value_net(state_batch)
act_probs = np.exp(log_act_probs.data.numpy())
return act_probs, value.data.numpy()
def policy_value_fn(self, game):
"""
输入：棋盘
输出：一个列表，由每一个可用落子的(action, probability)和棋盘状态价值组成
"""
legal_positions = game.actions() # 策略价值网络输出的是所有的落子概率，所以你需要剔除已落子的位置
current_state = np.ascontiguousarray(game.state()).reshape([1,26,9,9])
if self.use_gpu:
# device = torch.device("cuda:0")
log_act_probs, value = self.policy_value_net(Variable(torch.from_numpy(current_state)).cuda().float())
act_probs = np.exp(log_act_probs.data.cpu().numpy().flatten())
else:
# device = torch.device("cpu")
log_act_probs, value = self.policy_value_net(Variable(torch.from_numpy(current_state)).float())
act_probs = np.exp(log_act_probs.data.numpy().flatten())
# print("legal:",legal_positions)
# print("probs:", act_probs)
act_probs = zip(legal_positions, act_probs[legal_positions])
value = value.data[0][0]
return act_probs, value
def train_step(self, state_batch, mcts_probs, winner_batch, lr):
if self.use_gpu:
# device = torch.device("cuda:0")
state_batch = Variable(torch.FloatTensor(state_batch).cuda())
mcts_probs = Variable(torch.FloatTensor(mcts_probs).cuda())
winner_batch = Variable(torch.FloatTensor(winner_batch).cuda())
else:
# device = torch.device("cpu")
state_batch = Variable(torch.FloatTensor(state_batch))
mcts_probs = Variable(torch.FloatTensor(mcts_probs))
winner_batch = Variable(torch.FloatTensor(winner_batch))
self.optimizer.zero_grad()
set_learning_rate(self.optimizer, lr)
# 向前传播
log_act_probs, value = self.policy_value_net(state_batch)
# 损失公式： loss = (z - v)^2 - pi^T * log(p) + c||theta||^2，也就是value_loss+policy_loss
value_loss = F.mse_loss(value.view(-1), winner_batch)
policy_loss = -torch.mean(torch.sum(mcts_probs * log_act_probs, 1))
loss = value_loss + policy_loss
# 反向传播，优化损失
loss.backward()
self.optimizer.step()
# 计算熵，只是用于监控
entropy = -torch.mean(torch.sum(torch.exp(log_act_probs) * log_act_probs, 1))
return loss.data[0], entropy.data[0]
def get_policy_param(self):
net_params = self.policy_value_net.state_dict()
return net_params
def save_model(self, model_file):
""" 保存模型"""
torch.save(self.policy_value_net.state_dict(), 'ckpt/%s.pth'%(model_file))