-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathDDQN_with_Pong.py
185 lines (156 loc) · 6.87 KB
/
DDQN_with_Pong.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
import torch
import torch.nn as nn
import numpy as np
import Pong as game
import random
import os
import cv2
class MyNet(nn.Module):
def __init__(self,learning_rate):
super(MyNet,self).__init__()
self.ConvNet = nn.Sequential(
nn.Conv2d(4,8,3,1,padding=1, bias=False),
nn.ReLU(),
nn.Conv2d(8,8,1,2, bias=False),
nn.ReLU(),
nn.Conv2d(8,8,3,1,padding=1, bias=False),
nn.ReLU(),
nn.Conv2d(8,16,3,1,padding=1, bias=False),
nn.ReLU(),
nn.Conv2d(16,16,1,2, bias=False),
nn.ReLU(),
nn.Conv2d(16,16,3,1,padding=1, bias=False),
nn.ReLU(),
nn.Conv2d(16,16,1,2, bias=False),
nn.ReLU(),
nn.Conv2d(16,5,3,1,padding=1, bias=False),
nn.ReLU()
)
self.fc = nn.Sequential(
nn.Linear(48*5,48*5),
nn.ReLU(),
nn.Linear(48*5,3)
)
self.mls = nn.MSELoss()
self.opt = torch.optim.Adam(self.parameters(), lr = learning_rate,betas=(0.9, 0.999), eps=1e-08, weight_decay=0)
def forward(self, inputs):
inputs = self.ConvNet(inputs)
inputs = inputs.view(-1,48*5)
inputs = self.fc(inputs)
return inputs
width_net = 64
height_net = 48
store_count = 0
store_size = 10000
decline = 0.6
learning_rate = 0.0001
learn_time = 0
update_time = 50 # 1000
gama = 0.99
b_size = 32 #每次训练网络，从store_size个状态中取32组来训练
store_picture = np.zeros((store_size,2,4,height_net,width_net)) #存储store_size个state(即game截图)
store_other = np.zeros((store_size,3)) #存储store_size个用到的值，包括：采取的下一步行动a, 报酬值reward, 表示游戏是否结束的done_show
start_study = False
log_dir = './model/RL-pong.pth'
log_dir2 = './best/best-RL-pong.pth'
log_dir3 = './best/best-RL-pong-all.pth'
train = False
net1 = MyNet(learning_rate).cuda()
net2 = MyNet(learning_rate).cuda()
optimizer = torch.optim.Adam(net1.parameters(), lr = learning_rate,betas=(0.9, 0.999), eps=1e-08, weight_decay=0)
if os.path.exists(log_dir):
checkpoint = torch.load(log_dir)
net1.load_state_dict(checkpoint['model'])
net2.load_state_dict(checkpoint['model'])
net1.opt.load_state_dict(checkpoint['optimizer'])
net2.opt.load_state_dict(checkpoint['optimizer'])
learn_time = checkpoint['epoch']
print("--------------成功加载模型--------------")
else:
print("--------------训练全新模型--------------")
s = game.game_init()
s = np.reshape(cv2.resize(s,(width_net,height_net)),(1,1,height_net,width_net))
s = np.concatenate((s,s,s,s),axis = 1)
s = s / 255.
for i in range(5000000):
while True:
if random.randint(0,100) < 100*(decline ** (learn_time/10000)) and train: #if后面的式子表达的意思：训练轮数越多，每次越倾向于选择网络预测值中最大的a作为抉择，而不倾向于随机选择一个抉择
a = random.choice([0, 1, 2])
print("random")
else:
inputs = torch.Tensor(s).cuda()
out = net1(inputs).detach()
a = torch.argmax(out).data.item()
s_part,r,score_one_round,score,done_show = game.game_step(a)
s_part = np.reshape(cv2.resize(s_part,(width_net,height_net)),(1,1,height_net,width_net))
s_part = s_part / 255.
s_ = np.concatenate((s[:,1:4,:,:],s_part),axis = 1)
# if (np.all(s_[:,0,:,:] == s_[:,1,:,:]) and np.all(s_[:,0,:,:] == s_[:,2,:,:])):
# print("3个s")
done = True
store_picture[store_count % store_size][0] = s
store_picture[store_count % store_size][1] = s_
store_other[store_count % store_size][0] = a
store_other[store_count % store_size][1] = r
store_other[store_count % store_size][2] = done_show
store_count += 1
s = s_
rank_file_r = open("rank.txt","r")
best = int(rank_file_r.readline())
rank_file_r.close()
if score_one_round > best:
state = {'model':net1.state_dict(), 'epoch':learn_time}
torch.save(state, log_dir2)
rank_file_w = open("rank.txt","w")
rank_file_w.write("%d" % score_one_round)
print("********** best score_one_round updated!! *********")
rank_file_w.close()
rank_file_r2 = open("rank_all_round.txt","r")
best2 = int(rank_file_r2.readline())
rank_file_r2.close()
if score > best2:
state = {'model':net1.state_dict(), 'epoch':learn_time}
torch.save(state, log_dir3)
rank_file_w2 = open("rank_all_round.txt","w")
rank_file_w2.write("%d" % score)
print("********** best score updated!! *********")
rank_file_w2.close()
if store_count > store_size and train:
if learn_time % update_time == 0 :
net2.load_state_dict(net1.state_dict())
index = random.randint(0, store_size - b_size - 1)
b_s = torch.Tensor(store_picture[index:index + b_size, 0])
b_s = b_s.cuda()
b_s_ = torch.Tensor(store_picture[index:index + b_size, 1])
b_s_ = b_s_.cuda()
b_a = torch.Tensor(store_other[index:index + b_size, 0:1]).long() #转为LongTensor
b_a = b_a.cuda()
b_r = torch.Tensor(store_other[index:index + b_size, 1:2])
b_r = b_r.cuda()
b_done = torch.Tensor(store_other[index:index + b_size, 2:3])
b_done = b_done.cuda()
q = net1(b_s).gather(1,b_a) #dim = 1，使用b_a来选取net中最终采取的抉择带来的Q值，防止选最大一方为最后Q值后，忽略了随机选择的存在。
q_next = net2(b_s_).detach().max(1)[0].reshape(b_size, 1)
q_truth = b_r + gama*q_next*(1-b_done)
q = q.cuda()
q_truth = q_truth.cuda()
loss = nn.MSELoss()(q, q_truth)
optimizer.zero_grad()
loss.backward()
optimizer.step()
learn_time += 1
if not start_study:
print("===========start study===========")
start_study = True
print("epcho = %d" % learn_time)
break
if done:
print("epcho = %d" % learn_time)
print("loss = %f" % loss.item())
if learn_time % 1000 == 0 :
state = {'model':net1.state_dict(), 'optimizer':net1.opt.state_dict(), 'epoch':learn_time}
torch.save(state, log_dir)
f = open("scores.txt","a")
f.write("========= %d ========== \n" % learn_time)
f.close()
break