-
Notifications
You must be signed in to change notification settings - Fork 2
/
model_ac_torch.py
132 lines (107 loc) · 5.45 KB
/
model_ac_torch.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
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
class Actor(torch.nn.Module):
def __init__(self, action_space):
super(Actor, self).__init__()
self.input_channel = 1
self.action_space = action_space
channel_cnn = 128
channel_fc = 128
# self.bn = nn.BatchNorm1d(self.input_channel)
self.actor_conv1 = nn.Conv1d(self.input_channel, channel_cnn, 4) # L_out = 8 - (4-1) -1 + 1 = 5
self.actor_conv2 = nn.Conv1d(self.input_channel, channel_cnn, 4)
self.actor_conv3 = nn.Conv1d(self.input_channel, channel_cnn, 4) # for available chunk sizes 6 version L_out = 6 - (4-1) -1 + 1 = 3
self.actor_fc_1 = nn.Linear(self.input_channel, channel_fc)
self.actor_fc_2 = nn.Linear(self.input_channel, channel_fc)
self.actor_fc_3 = nn.Linear(self.input_channel, channel_fc)
#===================Hide layer=========================
incoming_size = 2*channel_cnn*5 + 1 * channel_cnn*3 + 3 * channel_fc #
self.fc1 = nn.Linear(in_features=incoming_size, out_features= channel_fc)
# self.fc2 = nn.Linear(in_features=channel_fc, out_features=channel_fc)
self.fc3 = nn.Linear(in_features=channel_fc, out_features=self.action_space)
# self.fc4 = nn.Linear(in_features=channel_fc, out_features=1)
def forward(self, inputs):
throughputs_batch = inputs[:, 2:3, :] ## refer to env_train.py
# throughputs_batch = self.bn(throughputs_batch)
download_time_batch = inputs[:, 3:4, :]
# download_time_batch = self.bn(download_time_batch)
sizes_batch = inputs[:, 4:5, :self.action_space]
# sizes_batch = self.bn(sizes_batch)
x_1 = F.relu(self.actor_conv1(throughputs_batch))
x_2 = F.relu(self.actor_conv2(download_time_batch))
x_3 = F.relu(self.actor_conv3(sizes_batch))
x_4 = F.relu(self.actor_fc_1(inputs[:, 0:1, -1]))
x_5 = F.relu(self.actor_fc_2(inputs[:, 1:2, -1]))
x_6 = F.relu(self.actor_fc_3(inputs[:, 5:6, -1]))
x_1 = x_1.view(-1, self.num_flat_features(x_1))
x_2 = x_2.view(-1, self.num_flat_features(x_2))
x_3 = x_3.view(-1, self.num_flat_features(x_3))
x_4 = x_4.view(-1, self.num_flat_features(x_4))
x_5 = x_5.view(-1, self.num_flat_features(x_5))
x_6 = x_6.view(-1, self.num_flat_features(x_6))
x = torch.cat([x_1, x_2, x_3, x_4, x_5, x_6], 1)
x = F.relu(self.fc1(x))
# actor
# actor = F.relu(self.fc1(x))
# actor = F.relu(self.fc2(actor))
actor = F.softmax(self.fc3(x), dim=1)
return actor
def num_flat_features(self,x):
size=x.size()[1:] # all dimensions except the batch dimension
num_features=1
for s in size:
num_features*=s
return num_features
class Critic(torch.nn.Module):
def __init__(self, action_space):
super(Critic, self).__init__()
self.input_channel = 1
self.action_space = action_space
channel_cnn = 128
channel_fc = 128
# self.bn = nn.BatchNorm1d(self.input_channel)
self.critic_conv1 = nn.Conv1d(self.input_channel, channel_cnn, 4) # L_out = 8 - (4-1) -1 + 1 = 5
self.critic_conv2 = nn.Conv1d(self.input_channel, channel_cnn, 4)
self.critic_conv3 = nn.Conv1d(self.input_channel, channel_cnn, 4) # for available chunk sizes 6 version L_out = 6 - (4-1) -1 + 1 = 3
self.critic_fc_1 = nn.Linear(self.input_channel, channel_fc)
self.critic_fc_2 = nn.Linear(self.input_channel, channel_fc)
self.critic_fc_3 = nn.Linear(self.input_channel, channel_fc)
#===================Hide layer=========================
incoming_size = 2*channel_cnn*5 + 1 * channel_cnn*3 + 3 * channel_fc #
self.fc1 = nn.Linear(in_features=incoming_size, out_features= channel_fc)
# self.fc2 = nn.Linear(in_features=channel_fc, out_features=channel_fc)
self.fc3 = nn.Linear(in_features=channel_fc, out_features=1)
def forward(self, inputs):
throughputs_batch = inputs[:, 2:3, :] ## refer to env_train.py
# throughputs_batch = self.bn(throughputs_batch)
download_time_batch = inputs[:, 3:4, :]
# download_time_batch = self.bn(download_time_batch)
sizes_batch = inputs[:, 4:5, :self.action_space]
# sizes_batch = self.bn(sizes_batch)
x_1 = F.relu(self.critic_conv1(throughputs_batch))
x_2 = F.relu(self.critic_conv2(download_time_batch))
x_3 = F.relu(self.critic_conv3(sizes_batch))
x_4 = F.relu(self.critic_fc_1(inputs[:, 0:1, -1]))
x_5 = F.relu(self.critic_fc_2(inputs[:, 1:2, -1]))
x_6 = F.relu(self.critic_fc_3(inputs[:, 5:6, -1]))
x_1 = x_1.view(-1, self.num_flat_features(x_1))
x_2 = x_2.view(-1, self.num_flat_features(x_2))
x_3 = x_3.view(-1, self.num_flat_features(x_3))
x_4 = x_4.view(-1, self.num_flat_features(x_4))
x_5 = x_5.view(-1, self.num_flat_features(x_5))
x_6 = x_6.view(-1, self.num_flat_features(x_6))
x = torch.cat([x_1, x_2, x_3, x_4, x_5, x_6], 1)
x = F.relu(self.fc1(x))
# critic
# critic = F.relu(self.fc1(x))
# critic = F.relu(self.fc2(critic))
critic = self.fc3(x)
return critic
def num_flat_features(self,x):
size=x.size()[1:] # all dimensions except the batch dimension
num_features=1
for s in size:
num_features*=s
return num_features