File size: 8,105 Bytes
2b21abc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
203
204
205
206
207
208
209
210
211
212
import torch
import torch.nn as nn


class MIMBlock(nn.Module):

    def __init__(self, in_channel, num_hidden, height, width, filter_size, stride, layer_norm):
        super(MIMBlock, self).__init__()

        self.convlstm_c = None
        self.num_hidden = num_hidden
        self.padding = filter_size // 2
        self._forget_bias = 1.0

        self.ct_weight = nn.Parameter(torch.zeros(num_hidden*2, height, width))
        self.oc_weight = nn.Parameter(torch.zeros(num_hidden, height, width))

        if layer_norm:
            self.conv_t_cc = nn.Sequential(
                nn.Conv2d(in_channel, num_hidden * 3, kernel_size=filter_size,
                          stride=stride, padding=self.padding, bias=False),
                nn.LayerNorm([num_hidden * 3, height, width])
            )
            self.conv_s_cc = nn.Sequential(
                nn.Conv2d(num_hidden, num_hidden * 4, kernel_size=filter_size,
                          stride=stride, padding=self.padding, bias=False),
                nn.LayerNorm([num_hidden * 4, height, width])
            )
            self.conv_x_cc = nn.Sequential(
                nn.Conv2d(num_hidden, num_hidden * 4, kernel_size=filter_size,
                          stride=stride, padding=self.padding, bias=False),
                nn.LayerNorm([num_hidden * 4, height, width])
            )
            self.conv_h_concat = nn.Sequential(
                nn.Conv2d(num_hidden, num_hidden * 4, kernel_size=filter_size,
                          stride=stride, padding=self.padding, bias=False),
                nn.LayerNorm([num_hidden * 4, height, width])
            )
            self.conv_x_concat = nn.Sequential(
                nn.Conv2d(num_hidden, num_hidden * 4, kernel_size=filter_size,
                          stride=stride, padding=self.padding, bias=False),
                nn.LayerNorm([num_hidden * 4, height, width])
            )
        else:
            self.conv_t_cc = nn.Sequential(
                nn.Conv2d(in_channel, num_hidden * 3, kernel_size=filter_size,
                          stride=stride, padding=self.padding, bias=False),
            )
            self.conv_s_cc = nn.Sequential(
                nn.Conv2d(num_hidden, num_hidden * 4, kernel_size=filter_size,
                          stride=stride, padding=self.padding, bias=False),
            )
            self.conv_x_cc = nn.Sequential(
                nn.Conv2d(num_hidden, num_hidden * 4, kernel_size=filter_size,
                          stride=stride, padding=self.padding, bias=False),
            )
            self.conv_h_concat = nn.Sequential(
                nn.Conv2d(num_hidden, num_hidden * 4, kernel_size=filter_size,
                          stride=stride, padding=self.padding, bias=False),
            )
            self.conv_x_concat = nn.Sequential(
                nn.Conv2d(num_hidden, num_hidden * 4, kernel_size=filter_size,
                          stride=stride, padding=self.padding, bias=False),
            )
        self.conv_last = nn.Conv2d(num_hidden * 2, num_hidden, kernel_size=1,
                                   stride=1, padding=0, bias=False)

    def _init_state(self, inputs):
        return torch.zeros_like(inputs)

    def MIMS(self, x, h_t, c_t):
        if h_t is None:
            h_t = self._init_state(x)
        if c_t is None:
            c_t = self._init_state(x)

        h_concat = self.conv_h_concat(h_t)
        i_h, g_h, f_h, o_h = torch.split(h_concat, self.num_hidden, dim=1)

        ct_activation = torch.mul(c_t.repeat(1,2,1,1), self.ct_weight)
        i_c, f_c = torch.split(ct_activation, self.num_hidden, dim=1)

        i_ = i_h + i_c
        f_ = f_h + f_c
        g_ = g_h
        o_ = o_h

        if x != None:
            x_concat = self.conv_x_concat(x)
            i_x, g_x, f_x, o_x = torch.split(x_concat, self.num_hidden, dim=1)

            i_ = i_ +  i_x
            f_ = f_ + f_x
            g_ = g_ +  g_x
            o_ = o_ + o_x

        i_ = torch.sigmoid(i_)
        f_ = torch.sigmoid(f_ + self._forget_bias)
        c_new = f_ * c_t + i_ * torch.tanh(g_)

        o_c = torch.mul(c_new, self.oc_weight)

        h_new = torch.sigmoid(o_ + o_c) * torch.tanh(c_new)

        return h_new, c_new

    def forward(self, x, diff_h, h, c, m):
        h = self._init_state(x) if h is None else h
        c = self._init_state(x) if c is None else c
        m = self._init_state(x) if m is None else m
        diff_h = self._init_state(x) if diff_h is None else diff_h

        t_cc = self.conv_t_cc(h)
        s_cc = self.conv_s_cc(m)
        x_cc = self.conv_x_cc(x)

        i_s, g_s, f_s, o_s = torch.split(s_cc, self.num_hidden, dim=1)
        i_t, g_t, o_t = torch.split(t_cc, self.num_hidden, dim=1)
        i_x, g_x, f_x, o_x = torch.split(x_cc, self.num_hidden, dim=1)

        i = torch.sigmoid(i_x + i_t)
        i_ = torch.sigmoid(i_x + i_s)
        g = torch.tanh(g_x + g_t)
        g_ = torch.tanh(g_x + g_s)
        f_ = torch.sigmoid(f_x + f_s + self._forget_bias)
        o = torch.sigmoid(o_x + o_t + o_s)
        new_m = f_ * m + i_ * g_
        
        c, self.convlstm_c = self.MIMS(diff_h, c, self.convlstm_c \
            if self.convlstm_c is None else self.convlstm_c.detach())

        new_c = c + i * g
        cell = torch.cat((new_c, new_m), 1)
        new_h = o * torch.tanh(self.conv_last(cell))

        return new_h, new_c, new_m


class MIMN(nn.Module):

    def __init__(self, in_channel, num_hidden, height, width, filter_size, stride, layer_norm):
        super(MIMN, self).__init__()

        self.num_hidden = num_hidden
        self.padding = filter_size // 2
        self._forget_bias = 1.0

        self.ct_weight = nn.Parameter(torch.zeros(num_hidden*2, height, width))
        self.oc_weight = nn.Parameter(torch.zeros(num_hidden, height, width))

        if layer_norm:
            self.conv_h_concat = nn.Sequential(
                nn.Conv2d(in_channel, num_hidden * 4, kernel_size=filter_size,
                          stride=stride, padding=self.padding, bias=False),
                nn.LayerNorm([num_hidden * 4, height, width])
            )
            self.conv_x_concat = nn.Sequential(
                nn.Conv2d(in_channel, num_hidden * 4, kernel_size=filter_size,
                          stride=stride, padding=self.padding, bias=False),
                nn.LayerNorm([num_hidden * 4, height, width])
            )
        else:
            self.conv_h_concat = nn.Sequential(
                nn.Conv2d(in_channel, num_hidden * 4, kernel_size=filter_size,
                          stride=stride, padding=self.padding, bias=False),
            )
            self.conv_x_concat = nn.Sequential(
                nn.Conv2d(in_channel, num_hidden * 4, kernel_size=filter_size,
                          stride=stride, padding=self.padding, bias=False),
            )
        self.conv_last = nn.Conv2d(num_hidden * 2, num_hidden, kernel_size=1,
                                   stride=1, padding=0, bias=False)

    def _init_state(self, inputs):
        return torch.zeros_like(inputs)

    def forward(self, x, h_t, c_t):
        if h_t is None:
            h_t = self._init_state(x)
        if c_t is None:
            c_t = self._init_state(x)

        h_concat = self.conv_h_concat(h_t)
        i_h, g_h, f_h, o_h = torch.split(h_concat, self.num_hidden, dim=1)

        ct_activation = torch.mul(c_t.repeat(1,2,1,1), self.ct_weight)
        i_c, f_c = torch.split(ct_activation, self.num_hidden, dim=1)

        i_ = i_h + i_c
        f_ = f_h + f_c
        g_ = g_h
        o_ = o_h

        if x != None:
            x_concat = self.conv_x_concat(x)
            i_x, g_x, f_x, o_x = torch.split(x_concat, self.num_hidden, dim=1)

            i_ = i_ + i_x
            f_ = f_ + f_x
            g_ = g_ + g_x
            o_ = o_ + o_x

        i_ = torch.sigmoid(i_)
        f_ = torch.sigmoid(f_ + self._forget_bias)
        c_new = f_ * c_t + i_ * torch.tanh(g_)

        o_c = torch.mul(c_new, self.oc_weight)

        h_new = torch.sigmoid(o_ + o_c) * torch.tanh(c_new)

        return h_new, c_new