File size: 11,012 Bytes
f7ac35e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
# Copyright (c) OpenMMLab. All rights reserved.
from abc import ABCMeta

import torch
import torch.nn as nn

from ..utils import constant_init, normal_init
from .conv_module import ConvModule
from .registry import PLUGIN_LAYERS


class _NonLocalNd(nn.Module, metaclass=ABCMeta):
    """Basic Non-local module.

    This module is proposed in
    "Non-local Neural Networks"
    Paper reference: https://arxiv.org/abs/1711.07971
    Code reference: https://github.com/AlexHex7/Non-local_pytorch

    Args:
        in_channels (int): Channels of the input feature map.
        reduction (int): Channel reduction ratio. Default: 2.
        use_scale (bool): Whether to scale pairwise_weight by
            `1/sqrt(inter_channels)` when the mode is `embedded_gaussian`.
            Default: True.
        conv_cfg (None | dict): The config dict for convolution layers.
            If not specified, it will use `nn.Conv2d` for convolution layers.
            Default: None.
        norm_cfg (None | dict): The config dict for normalization layers.
            Default: None. (This parameter is only applicable to conv_out.)
        mode (str): Options are `gaussian`, `concatenation`,
            `embedded_gaussian` and `dot_product`. Default: embedded_gaussian.
    """

    def __init__(self,
                 in_channels,
                 reduction=2,
                 use_scale=True,
                 conv_cfg=None,
                 norm_cfg=None,
                 mode='embedded_gaussian',
                 **kwargs):
        super(_NonLocalNd, self).__init__()
        self.in_channels = in_channels
        self.reduction = reduction
        self.use_scale = use_scale
        self.inter_channels = max(in_channels // reduction, 1)
        self.mode = mode

        if mode not in [
                'gaussian', 'embedded_gaussian', 'dot_product', 'concatenation'
        ]:
            raise ValueError("Mode should be in 'gaussian', 'concatenation', "
                             f"'embedded_gaussian' or 'dot_product', but got "
                             f'{mode} instead.')

        # g, theta, phi are defaulted as `nn.ConvNd`.
        # Here we use ConvModule for potential usage.
        self.g = ConvModule(
            self.in_channels,
            self.inter_channels,
            kernel_size=1,
            conv_cfg=conv_cfg,
            act_cfg=None)
        self.conv_out = ConvModule(
            self.inter_channels,
            self.in_channels,
            kernel_size=1,
            conv_cfg=conv_cfg,
            norm_cfg=norm_cfg,
            act_cfg=None)

        if self.mode != 'gaussian':
            self.theta = ConvModule(
                self.in_channels,
                self.inter_channels,
                kernel_size=1,
                conv_cfg=conv_cfg,
                act_cfg=None)
            self.phi = ConvModule(
                self.in_channels,
                self.inter_channels,
                kernel_size=1,
                conv_cfg=conv_cfg,
                act_cfg=None)

        if self.mode == 'concatenation':
            self.concat_project = ConvModule(
                self.inter_channels * 2,
                1,
                kernel_size=1,
                stride=1,
                padding=0,
                bias=False,
                act_cfg=dict(type='ReLU'))

        self.init_weights(**kwargs)

    def init_weights(self, std=0.01, zeros_init=True):
        if self.mode != 'gaussian':
            for m in [self.g, self.theta, self.phi]:
                normal_init(m.conv, std=std)
        else:
            normal_init(self.g.conv, std=std)
        if zeros_init:
            if self.conv_out.norm_cfg is None:
                constant_init(self.conv_out.conv, 0)
            else:
                constant_init(self.conv_out.norm, 0)
        else:
            if self.conv_out.norm_cfg is None:
                normal_init(self.conv_out.conv, std=std)
            else:
                normal_init(self.conv_out.norm, std=std)

    def gaussian(self, theta_x, phi_x):
        # NonLocal1d pairwise_weight: [N, H, H]
        # NonLocal2d pairwise_weight: [N, HxW, HxW]
        # NonLocal3d pairwise_weight: [N, TxHxW, TxHxW]
        pairwise_weight = torch.matmul(theta_x, phi_x)
        pairwise_weight = pairwise_weight.softmax(dim=-1)
        return pairwise_weight

    def embedded_gaussian(self, theta_x, phi_x):
        # NonLocal1d pairwise_weight: [N, H, H]
        # NonLocal2d pairwise_weight: [N, HxW, HxW]
        # NonLocal3d pairwise_weight: [N, TxHxW, TxHxW]
        pairwise_weight = torch.matmul(theta_x, phi_x)
        if self.use_scale:
            # theta_x.shape[-1] is `self.inter_channels`
            pairwise_weight /= theta_x.shape[-1]**0.5
        pairwise_weight = pairwise_weight.softmax(dim=-1)
        return pairwise_weight

    def dot_product(self, theta_x, phi_x):
        # NonLocal1d pairwise_weight: [N, H, H]
        # NonLocal2d pairwise_weight: [N, HxW, HxW]
        # NonLocal3d pairwise_weight: [N, TxHxW, TxHxW]
        pairwise_weight = torch.matmul(theta_x, phi_x)
        pairwise_weight /= pairwise_weight.shape[-1]
        return pairwise_weight

    def concatenation(self, theta_x, phi_x):
        # NonLocal1d pairwise_weight: [N, H, H]
        # NonLocal2d pairwise_weight: [N, HxW, HxW]
        # NonLocal3d pairwise_weight: [N, TxHxW, TxHxW]
        h = theta_x.size(2)
        w = phi_x.size(3)
        theta_x = theta_x.repeat(1, 1, 1, w)
        phi_x = phi_x.repeat(1, 1, h, 1)

        concat_feature = torch.cat([theta_x, phi_x], dim=1)
        pairwise_weight = self.concat_project(concat_feature)
        n, _, h, w = pairwise_weight.size()
        pairwise_weight = pairwise_weight.view(n, h, w)
        pairwise_weight /= pairwise_weight.shape[-1]

        return pairwise_weight

    def forward(self, x):
        # Assume `reduction = 1`, then `inter_channels = C`
        # or `inter_channels = C` when `mode="gaussian"`

        # NonLocal1d x: [N, C, H]
        # NonLocal2d x: [N, C, H, W]
        # NonLocal3d x: [N, C, T, H, W]
        n = x.size(0)

        # NonLocal1d g_x: [N, H, C]
        # NonLocal2d g_x: [N, HxW, C]
        # NonLocal3d g_x: [N, TxHxW, C]
        g_x = self.g(x).view(n, self.inter_channels, -1)
        g_x = g_x.permute(0, 2, 1)

        # NonLocal1d theta_x: [N, H, C], phi_x: [N, C, H]
        # NonLocal2d theta_x: [N, HxW, C], phi_x: [N, C, HxW]
        # NonLocal3d theta_x: [N, TxHxW, C], phi_x: [N, C, TxHxW]
        if self.mode == 'gaussian':
            theta_x = x.view(n, self.in_channels, -1)
            theta_x = theta_x.permute(0, 2, 1)
            if self.sub_sample:
                phi_x = self.phi(x).view(n, self.in_channels, -1)
            else:
                phi_x = x.view(n, self.in_channels, -1)
        elif self.mode == 'concatenation':
            theta_x = self.theta(x).view(n, self.inter_channels, -1, 1)
            phi_x = self.phi(x).view(n, self.inter_channels, 1, -1)
        else:
            theta_x = self.theta(x).view(n, self.inter_channels, -1)
            theta_x = theta_x.permute(0, 2, 1)
            phi_x = self.phi(x).view(n, self.inter_channels, -1)

        pairwise_func = getattr(self, self.mode)
        # NonLocal1d pairwise_weight: [N, H, H]
        # NonLocal2d pairwise_weight: [N, HxW, HxW]
        # NonLocal3d pairwise_weight: [N, TxHxW, TxHxW]
        pairwise_weight = pairwise_func(theta_x, phi_x)

        # NonLocal1d y: [N, H, C]
        # NonLocal2d y: [N, HxW, C]
        # NonLocal3d y: [N, TxHxW, C]
        y = torch.matmul(pairwise_weight, g_x)
        # NonLocal1d y: [N, C, H]
        # NonLocal2d y: [N, C, H, W]
        # NonLocal3d y: [N, C, T, H, W]
        y = y.permute(0, 2, 1).contiguous().reshape(n, self.inter_channels,
                                                    *x.size()[2:])

        output = x + self.conv_out(y)

        return output


class NonLocal1d(_NonLocalNd):
    """1D Non-local module.

    Args:
        in_channels (int): Same as `NonLocalND`.
        sub_sample (bool): Whether to apply max pooling after pairwise
            function (Note that the `sub_sample` is applied on spatial only).
            Default: False.
        conv_cfg (None | dict): Same as `NonLocalND`.
            Default: dict(type='Conv1d').
    """

    def __init__(self,
                 in_channels,
                 sub_sample=False,
                 conv_cfg=dict(type='Conv1d'),
                 **kwargs):
        super(NonLocal1d, self).__init__(
            in_channels, conv_cfg=conv_cfg, **kwargs)

        self.sub_sample = sub_sample

        if sub_sample:
            max_pool_layer = nn.MaxPool1d(kernel_size=2)
            self.g = nn.Sequential(self.g, max_pool_layer)
            if self.mode != 'gaussian':
                self.phi = nn.Sequential(self.phi, max_pool_layer)
            else:
                self.phi = max_pool_layer


@PLUGIN_LAYERS.register_module()
class NonLocal2d(_NonLocalNd):
    """2D Non-local module.

    Args:
        in_channels (int): Same as `NonLocalND`.
        sub_sample (bool): Whether to apply max pooling after pairwise
            function (Note that the `sub_sample` is applied on spatial only).
            Default: False.
        conv_cfg (None | dict): Same as `NonLocalND`.
            Default: dict(type='Conv2d').
    """

    _abbr_ = 'nonlocal_block'

    def __init__(self,
                 in_channels,
                 sub_sample=False,
                 conv_cfg=dict(type='Conv2d'),
                 **kwargs):
        super(NonLocal2d, self).__init__(
            in_channels, conv_cfg=conv_cfg, **kwargs)

        self.sub_sample = sub_sample

        if sub_sample:
            max_pool_layer = nn.MaxPool2d(kernel_size=(2, 2))
            self.g = nn.Sequential(self.g, max_pool_layer)
            if self.mode != 'gaussian':
                self.phi = nn.Sequential(self.phi, max_pool_layer)
            else:
                self.phi = max_pool_layer


class NonLocal3d(_NonLocalNd):
    """3D Non-local module.

    Args:
        in_channels (int): Same as `NonLocalND`.
        sub_sample (bool): Whether to apply max pooling after pairwise
            function (Note that the `sub_sample` is applied on spatial only).
            Default: False.
        conv_cfg (None | dict): Same as `NonLocalND`.
            Default: dict(type='Conv3d').
    """

    def __init__(self,
                 in_channels,
                 sub_sample=False,
                 conv_cfg=dict(type='Conv3d'),
                 **kwargs):
        super(NonLocal3d, self).__init__(
            in_channels, conv_cfg=conv_cfg, **kwargs)
        self.sub_sample = sub_sample

        if sub_sample:
            max_pool_layer = nn.MaxPool3d(kernel_size=(1, 2, 2))
            self.g = nn.Sequential(self.g, max_pool_layer)
            if self.mode != 'gaussian':
                self.phi = nn.Sequential(self.phi, max_pool_layer)
            else:
                self.phi = max_pool_layer