AF_Net/model.py at main · RenMin1991/AF_Net · GitHub

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
# -*- coding: utf-8 -*-
'''
    maxout + NetVLAD
    Ren Min
    20181122
'''

import math
import torch
import torch.nn as nn
import torch.nn.functional as F
import pdb

from netvlad import NetVLAD
from layers import ConvOffset2D


class mfm(nn.Module):
    def __init__(self, in_channels, out_channels, kernel_size=3, stride=1, padding=1, type=1):
        super(mfm, self).__init__()
        self.out_channels = out_channels
        if type == 1:
            self.filter = nn.Conv2d(in_channels, 2*out_channels, kernel_size=kernel_size, stride=stride, padding=padding)
        else:
            self.filter = nn.Linear(in_channels, 2*out_channels)

    def forward(self, x):
        x = self.filter(x)
        out = torch.split(x, self.out_channels, 1)
        return torch.max(out[0], out[1])

class Maxout_4(nn.Module):
    def __init__(self, num_classes_th, num_classes_la, num_classes_in, num_classes_cs):
        super(Maxout_4, self).__init__()
        self.features = nn.Sequential(
            mfm(1, 48, 9, 1, 0),
            nn.MaxPool2d(kernel_size=2, stride=2, ceil_mode=True),
            mfm(48, 96, 5, 1, 0),
            nn.MaxPool2d(kernel_size=2, stride=2, ceil_mode=True),
            mfm(96, 128, 5, 1, 0),
            nn.MaxPool2d(kernel_size=2, stride=2, ceil_mode=True),
            mfm(128, 192, 4, 1, 0),
            nn.MaxPool2d(kernel_size=2, stride=2, ceil_mode=True),
            )
        self.fc1 = mfm(5*5*192, 256, type=0)
        self.dropout = nn.Dropout(0.7)
        self.fc2_th = nn.Linear(256, num_classes_th)
        self.fc2_la = nn.Linear(256, num_classes_la)
        #self.fc2_in = nn.Linear(256, num_classes_in)
        self.fc2_cs = nn.Linear(256, num_classes_cs)

    def forward(self, x):
        x = self.features(x)
        x = x.view(x.size(0), -1)
        x = self.fc1(x)
        x = self.dropout(x)
        out_th = self.fc2_th(x)
        out_la = self.fc2_la(x)
        #out_in = self.fc2_in(x)
        out_cs = self.fc2_cs(x)
        return out_th, out_la, out_cs, x


fc1_pca_weights = torch.load('fc1_pca_weights.pth')
fc1_pca_weights = fc1_pca_weights.float()

class Maxout_VLAD(nn.Module):
    def __init__(self, num_classes_th, num_classes_la, num_classes_in, num_classes_cs):
        super(Maxout_VLAD, self).__init__()

        self.mfm1 = mfm(1, 48, 9, 1, 0)
        self.pool1 = nn.MaxPool2d(kernel_size=2, stride=2, ceil_mode=True)
        self.offset1 = ConvOffset2D(filters=48)

        self.mfm2 = mfm(48, 96, 5, 1, 0)
        self.pool2 = nn.MaxPool2d(kernel_size=2, stride=2, ceil_mode=True)
        self.offset2 = ConvOffset2D(filters=96)

        self.mfm3 = mfm(96, 128, 5, 1, 0)
        self.pool3 = nn.MaxPool2d(kernel_size=2, stride=2, ceil_mode=True)
        self.offset3 = ConvOffset2D(filters=128)

        self.mfm4 = mfm(128, 192, 4, 1, 0)
        self.pool4 = nn.MaxPool2d(kernel_size=2, stride=2, ceil_mode=True)

        self.offset4 = ConvOffset2D(filters=192)

        self.netvlad = NetVLAD(num_clusters=25, num_ghost=0, dim=192, alpha=50., normalize_input=True, L2_normalize=False)

        self.fc1_pca = mfm(25*192, 256, type=0)
        #self.bn = nn.BatchNorm1d(256)
        self.dropout = nn.Dropout(0.7)
        self.fc2_th_pca = nn.Linear(256, num_classes_th)
        self.fc2_la_pca = nn.Linear(256, num_classes_la)
        self.fc2_in = nn.Linear(256, num_classes_in)
        self.fc2_cs_pca = nn.Linear(256, num_classes_cs)

        #self._init_params()

    #def _init_params(self):
        #self.fc1_pca.filter.weight = nn.Parameter(fc1_pca_weights)

    def forward(self, x):
        x = self.mfm1(x)
        x = self.pool1(x)
        x = self.offset1(x)

        x = self.mfm2(x)
        x = self.pool2(x)
        x = self.offset2(x)

        x = self.mfm3(x)
        x = self.pool3(x)
        x = self.offset3(x)

        x = self.mfm4(x)
        x = self.pool4(x)
        x = self.offset4(x)

        x = self.netvlad(x)

        #x = x.view(x.size(0), -1)
        x = self.fc1_pca(x)

        #x = self.bn(x)
        #x = F.normalize(x, p=2, dim=1)

        x = self.dropout(x)
        out_th = self.fc2_th_pca(x)
        out_la = self.fc2_la_pca(x)
        out_in = self.fc2_in(x)
        out_cs = self.fc2_cs_pca(x)
        return out_th, out_la, out_in, out_cs, x