File size: 7,784 Bytes
3e06e1c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright (c) OpenMMLab. All rights reserved.
import math
from typing import Sequence, Tuple

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

from mmdet.registry import MODELS
from mmdet.utils import ConfigType, OptConfigType, OptMultiConfig
from ..layers import CSPLayer
from .csp_darknet import SPPBottleneck


@MODELS.register_module()
class CSPNeXt(BaseModule):
    """CSPNeXt backbone used in RTMDet.

    Args:
        arch (str): Architecture of CSPNeXt, from {P5, P6}.
            Defaults to P5.
        expand_ratio (float): Ratio to adjust the number of channels of the
            hidden layer. Defaults to 0.5.
        deepen_factor (float): Depth multiplier, multiply number of
            blocks in CSP layer by this amount. Defaults to 1.0.
        widen_factor (float): Width multiplier, multiply number of
            channels in each layer by this amount. Defaults to 1.0.
        out_indices (Sequence[int]): Output from which stages.
            Defaults to (2, 3, 4).
        frozen_stages (int): Stages to be frozen (stop grad and set eval
            mode). -1 means not freezing any parameters. Defaults to -1.
        use_depthwise (bool): Whether to use depthwise separable convolution.
            Defaults to False.
        arch_ovewrite (list): Overwrite default arch settings.
            Defaults to None.
        spp_kernel_sizes: (tuple[int]): Sequential of kernel sizes of SPP
            layers. Defaults to (5, 9, 13).
        channel_attention (bool): Whether to add channel attention in each
            stage. Defaults to True.
        conv_cfg (:obj:`ConfigDict` or dict, optional): Config dict for
            convolution layer. Defaults to None.
        norm_cfg (:obj:`ConfigDict` or dict): Dictionary to construct and
            config norm layer. Defaults to dict(type='BN', requires_grad=True).
        act_cfg (:obj:`ConfigDict` or dict): Config dict for activation layer.
            Defaults to dict(type='SiLU').
        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 (:obj:`ConfigDict` or dict or list[dict] or
            list[:obj:`ConfigDict`]): Initialization config dict.
    """
    # From left to right:
    # in_channels, out_channels, num_blocks, add_identity, use_spp
    arch_settings = {
        'P5': [[64, 128, 3, True, False], [128, 256, 6, True, False],
               [256, 512, 6, True, False], [512, 1024, 3, False, True]],
        'P6': [[64, 128, 3, True, False], [128, 256, 6, True, False],
               [256, 512, 6, True, False], [512, 768, 3, True, False],
               [768, 1024, 3, False, True]]
    }

    def __init__(
        self,
        arch: str = 'P5',
        deepen_factor: float = 1.0,
        widen_factor: float = 1.0,
        out_indices: Sequence[int] = (2, 3, 4),
        frozen_stages: int = -1,
        use_depthwise: bool = False,
        expand_ratio: float = 0.5,
        arch_ovewrite: dict = None,
        spp_kernel_sizes: Sequence[int] = (5, 9, 13),
        channel_attention: bool = True,
        conv_cfg: OptConfigType = None,
        norm_cfg: ConfigType = dict(type='BN', momentum=0.03, eps=0.001),
        act_cfg: ConfigType = dict(type='SiLU'),
        norm_eval: bool = False,
        init_cfg: OptMultiConfig = dict(
            type='Kaiming',
            layer='Conv2d',
            a=math.sqrt(5),
            distribution='uniform',
            mode='fan_in',
            nonlinearity='leaky_relu')
    ) -> None:
        super().__init__(init_cfg=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 = nn.Sequential(
            ConvModule(
                3,
                int(arch_setting[0][0] * widen_factor // 2),
                3,
                padding=1,
                stride=2,
                norm_cfg=norm_cfg,
                act_cfg=act_cfg),
            ConvModule(
                int(arch_setting[0][0] * widen_factor // 2),
                int(arch_setting[0][0] * widen_factor // 2),
                3,
                padding=1,
                stride=1,
                norm_cfg=norm_cfg,
                act_cfg=act_cfg),
            ConvModule(
                int(arch_setting[0][0] * widen_factor // 2),
                int(arch_setting[0][0] * widen_factor),
                3,
                padding=1,
                stride=1,
                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_kernel_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,
                use_cspnext_block=True,
                expand_ratio=expand_ratio,
                channel_attention=channel_attention,
                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) -> None:
        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) -> None:
        super().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: Tuple[Tensor, ...]) -> Tuple[Tensor, ...]:
        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)