File size: 10,620 Bytes
f549064
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright (c) OpenMMLab. All rights reserved.
import math

import torch
import torch.nn as nn
from mmcv.cnn import ConvModule, DepthwiseSeparableConvModule
from mmengine.model import BaseModule
from torch.nn.modules.batchnorm import _BatchNorm

from mmdet.registry import MODELS
from ..layers import CSPLayer


class Focus(nn.Module):
    """Focus width and height information into channel space.

    Args:
        in_channels (int): The input channels of this Module.
        out_channels (int): The output channels of this Module.
        kernel_size (int): The kernel size of the convolution. Default: 1
        stride (int): The stride of the convolution. Default: 1
        conv_cfg (dict): Config dict for convolution layer. Default: None,
            which means using conv2d.
        norm_cfg (dict): Config dict for normalization layer.
            Default: dict(type='BN', momentum=0.03, eps=0.001).
        act_cfg (dict): Config dict for activation layer.
            Default: dict(type='Swish').
    """

    def __init__(self,
                 in_channels,
                 out_channels,
                 kernel_size=1,
                 stride=1,
                 conv_cfg=None,
                 norm_cfg=dict(type='BN', momentum=0.03, eps=0.001),
                 act_cfg=dict(type='Swish')):
        super().__init__()
        self.conv = ConvModule(
            in_channels * 4,
            out_channels,
            kernel_size,
            stride,
            padding=(kernel_size - 1) // 2,
            conv_cfg=conv_cfg,
            norm_cfg=norm_cfg,
            act_cfg=act_cfg)

    def forward(self, x):
        # shape of x (b,c,w,h) -> y(b,4c,w/2,h/2)
        patch_top_left = x[..., ::2, ::2]
        patch_top_right = x[..., ::2, 1::2]
        patch_bot_left = x[..., 1::2, ::2]
        patch_bot_right = x[..., 1::2, 1::2]
        x = torch.cat(
            (
                patch_top_left,
                patch_bot_left,
                patch_top_right,
                patch_bot_right,
            ),
            dim=1,
        )
        return self.conv(x)


class SPPBottleneck(BaseModule):
    """Spatial pyramid pooling layer used in YOLOv3-SPP.

    Args:
        in_channels (int): The input channels of this Module.
        out_channels (int): The output channels of this Module.
        kernel_sizes (tuple[int]): Sequential of kernel sizes of pooling
            layers. Default: (5, 9, 13).
        conv_cfg (dict): Config dict for convolution layer. Default: None,
            which means using conv2d.
        norm_cfg (dict): Config dict for normalization layer.
            Default: dict(type='BN').
        act_cfg (dict): Config dict for activation layer.
            Default: dict(type='Swish').
        init_cfg (dict or list[dict], optional): Initialization config dict.
            Default: None.
    """

    def __init__(self,
                 in_channels,
                 out_channels,
                 kernel_sizes=(5, 9, 13),
                 conv_cfg=None,
                 norm_cfg=dict(type='BN', momentum=0.03, eps=0.001),
                 act_cfg=dict(type='Swish'),
                 init_cfg=None):
        super().__init__(init_cfg)
        mid_channels = in_channels // 2
        self.conv1 = ConvModule(
            in_channels,
            mid_channels,
            1,
            stride=1,
            conv_cfg=conv_cfg,
            norm_cfg=norm_cfg,
            act_cfg=act_cfg)
        self.poolings = nn.ModuleList([
            nn.MaxPool2d(kernel_size=ks, stride=1, padding=ks // 2)
            for ks in kernel_sizes
        ])
        conv2_channels = mid_channels * (len(kernel_sizes) + 1)
        self.conv2 = ConvModule(
            conv2_channels,
            out_channels,
            1,
            conv_cfg=conv_cfg,
            norm_cfg=norm_cfg,
            act_cfg=act_cfg)

    def forward(self, x):
        x = self.conv1(x)
        with torch.cuda.amp.autocast(enabled=False):
            x = torch.cat(
                [x] + [pooling(x) for pooling in self.poolings], dim=1)
        x = self.conv2(x)
        return x


@MODELS.register_module()
class CSPDarknet(BaseModule):
    """CSP-Darknet backbone used in YOLOv5 and YOLOX.

    Args:
        arch (str): Architecture of CSP-Darknet, from {P5, P6}.
            Default: P5.
        deepen_factor (float): Depth multiplier, multiply number of
            blocks in CSP layer by this amount. Default: 1.0.
        widen_factor (float): Width multiplier, multiply number of
            channels in each layer by this amount. Default: 1.0.
        out_indices (Sequence[int]): Output from which stages.
            Default: (2, 3, 4).
        frozen_stages (int): Stages to be frozen (stop grad and set eval
            mode). -1 means not freezing any parameters. Default: -1.
        use_depthwise (bool): Whether to use depthwise separable convolution.
            Default: False.
        arch_ovewrite(list): Overwrite default arch settings. Default: None.
        spp_kernal_sizes: (tuple[int]): Sequential of kernel sizes of SPP
            layers. Default: (5, 9, 13).
        conv_cfg (dict): Config dict for convolution layer. Default: None.
        norm_cfg (dict): Dictionary to construct and config norm layer.
            Default: dict(type='BN', requires_grad=True).
        act_cfg (dict): Config dict for activation layer.
            Default: dict(type='LeakyReLU', negative_slope=0.1).
        norm_eval (bool): Whether to set norm layers to eval mode, namely,
            freeze running stats (mean and var). Note: Effect on Batch Norm
            and its variants only.
        init_cfg (dict or list[dict], optional): Initialization config dict.
            Default: None.
    Example:
        >>> from mmdet.models import CSPDarknet
        >>> import torch
        >>> self = CSPDarknet(depth=53)
        >>> self.eval()
        >>> inputs = torch.rand(1, 3, 416, 416)
        >>> level_outputs = self.forward(inputs)
        >>> for level_out in level_outputs:
        ...     print(tuple(level_out.shape))
        ...
        (1, 256, 52, 52)
        (1, 512, 26, 26)
        (1, 1024, 13, 13)
    """
    # From left to right:
    # in_channels, out_channels, num_blocks, add_identity, use_spp
    arch_settings = {
        'P5': [[64, 128, 3, True, False], [128, 256, 9, True, False],
               [256, 512, 9, True, False], [512, 1024, 3, False, True]],
        'P6': [[64, 128, 3, True, False], [128, 256, 9, True, False],
               [256, 512, 9, True, False], [512, 768, 3, True, False],
               [768, 1024, 3, False, True]]
    }

    def __init__(self,
                 arch='P5',
                 deepen_factor=1.0,
                 widen_factor=1.0,
                 out_indices=(2, 3, 4),
                 frozen_stages=-1,
                 use_depthwise=False,
                 arch_ovewrite=None,
                 spp_kernal_sizes=(5, 9, 13),
                 conv_cfg=None,
                 norm_cfg=dict(type='BN', momentum=0.03, eps=0.001),
                 act_cfg=dict(type='Swish'),
                 norm_eval=False,
                 init_cfg=dict(
                     type='Kaiming',
                     layer='Conv2d',
                     a=math.sqrt(5),
                     distribution='uniform',
                     mode='fan_in',
                     nonlinearity='leaky_relu')):
        super().__init__(init_cfg)
        arch_setting = self.arch_settings[arch]
        if arch_ovewrite:
            arch_setting = arch_ovewrite
        assert set(out_indices).issubset(
            i for i in range(len(arch_setting) + 1))
        if frozen_stages not in range(-1, len(arch_setting) + 1):
            raise ValueError('frozen_stages must be in range(-1, '
                             'len(arch_setting) + 1). But received '
                             f'{frozen_stages}')

        self.out_indices = out_indices
        self.frozen_stages = frozen_stages
        self.use_depthwise = use_depthwise
        self.norm_eval = norm_eval
        conv = DepthwiseSeparableConvModule if use_depthwise else ConvModule

        self.stem = Focus(
            3,
            int(arch_setting[0][0] * widen_factor),
            kernel_size=3,
            conv_cfg=conv_cfg,
            norm_cfg=norm_cfg,
            act_cfg=act_cfg)
        self.layers = ['stem']

        for i, (in_channels, out_channels, num_blocks, add_identity,
                use_spp) in enumerate(arch_setting):
            in_channels = int(in_channels * widen_factor)
            out_channels = int(out_channels * widen_factor)
            num_blocks = max(round(num_blocks * deepen_factor), 1)
            stage = []
            conv_layer = conv(
                in_channels,
                out_channels,
                3,
                stride=2,
                padding=1,
                conv_cfg=conv_cfg,
                norm_cfg=norm_cfg,
                act_cfg=act_cfg)
            stage.append(conv_layer)
            if use_spp:
                spp = SPPBottleneck(
                    out_channels,
                    out_channels,
                    kernel_sizes=spp_kernal_sizes,
                    conv_cfg=conv_cfg,
                    norm_cfg=norm_cfg,
                    act_cfg=act_cfg)
                stage.append(spp)
            csp_layer = CSPLayer(
                out_channels,
                out_channels,
                num_blocks=num_blocks,
                add_identity=add_identity,
                use_depthwise=use_depthwise,
                conv_cfg=conv_cfg,
                norm_cfg=norm_cfg,
                act_cfg=act_cfg)
            stage.append(csp_layer)
            self.add_module(f'stage{i + 1}', nn.Sequential(*stage))
            self.layers.append(f'stage{i + 1}')

    def _freeze_stages(self):
        if self.frozen_stages >= 0:
            for i in range(self.frozen_stages + 1):
                m = getattr(self, self.layers[i])
                m.eval()
                for param in m.parameters():
                    param.requires_grad = False

    def train(self, mode=True):
        super(CSPDarknet, self).train(mode)
        self._freeze_stages()
        if mode and self.norm_eval:
            for m in self.modules():
                if isinstance(m, _BatchNorm):
                    m.eval()

    def forward(self, x):
        outs = []
        for i, layer_name in enumerate(self.layers):
            layer = getattr(self, layer_name)
            x = layer(x)
            if i in self.out_indices:
                outs.append(x)
        return tuple(outs)