db_fpn.py 9.2 KB
Newer Older
W
WenmuZhou 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
# copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#    http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import paddle
from paddle import nn
import paddle.nn.functional as F
from paddle import ParamAttr
L
LDOUBLEV 已提交
23
import os
L
fix det  
LDOUBLEV 已提交
24 25
import sys

L
fix  
LDOUBLEV 已提交
26 27 28 29
__dir__ = os.path.dirname(os.path.abspath(__file__))
sys.path.append(__dir__)
sys.path.insert(0, os.path.abspath(os.path.join(__dir__, '../../..')))

L
LDOUBLEV 已提交
30
from ppocr.backbones.det_mobilenet_v3 import SEModule, ConvBNLayer
W
WenmuZhou 已提交
31 32


D
dyning 已提交
33
class DBFPN(nn.Layer):
W
WenmuZhou 已提交
34
    def __init__(self, in_channels, out_channels, **kwargs):
D
dyning 已提交
35
        super(DBFPN, self).__init__()
W
WenmuZhou 已提交
36
        self.out_channels = out_channels
W
WenmuZhou 已提交
37
        weight_attr = paddle.nn.initializer.KaimingUniform()
W
WenmuZhou 已提交
38

D
dyning 已提交
39
        self.in2_conv = nn.Conv2D(
W
WenmuZhou 已提交
40 41 42
            in_channels=in_channels[0],
            out_channels=self.out_channels,
            kernel_size=1,
littletomatodonkey's avatar
littletomatodonkey 已提交
43
            weight_attr=ParamAttr(initializer=weight_attr),
W
WenmuZhou 已提交
44
            bias_attr=False)
D
dyning 已提交
45
        self.in3_conv = nn.Conv2D(
W
WenmuZhou 已提交
46 47 48
            in_channels=in_channels[1],
            out_channels=self.out_channels,
            kernel_size=1,
littletomatodonkey's avatar
littletomatodonkey 已提交
49
            weight_attr=ParamAttr(initializer=weight_attr),
W
WenmuZhou 已提交
50
            bias_attr=False)
D
dyning 已提交
51
        self.in4_conv = nn.Conv2D(
W
WenmuZhou 已提交
52 53 54
            in_channels=in_channels[2],
            out_channels=self.out_channels,
            kernel_size=1,
littletomatodonkey's avatar
littletomatodonkey 已提交
55
            weight_attr=ParamAttr(initializer=weight_attr),
W
WenmuZhou 已提交
56
            bias_attr=False)
D
dyning 已提交
57
        self.in5_conv = nn.Conv2D(
W
WenmuZhou 已提交
58 59 60
            in_channels=in_channels[3],
            out_channels=self.out_channels,
            kernel_size=1,
littletomatodonkey's avatar
littletomatodonkey 已提交
61
            weight_attr=ParamAttr(initializer=weight_attr),
W
WenmuZhou 已提交
62
            bias_attr=False)
D
dyning 已提交
63
        self.p5_conv = nn.Conv2D(
W
WenmuZhou 已提交
64 65 66 67
            in_channels=self.out_channels,
            out_channels=self.out_channels // 4,
            kernel_size=3,
            padding=1,
littletomatodonkey's avatar
littletomatodonkey 已提交
68
            weight_attr=ParamAttr(initializer=weight_attr),
W
WenmuZhou 已提交
69
            bias_attr=False)
D
dyning 已提交
70
        self.p4_conv = nn.Conv2D(
W
WenmuZhou 已提交
71 72 73 74
            in_channels=self.out_channels,
            out_channels=self.out_channels // 4,
            kernel_size=3,
            padding=1,
littletomatodonkey's avatar
littletomatodonkey 已提交
75
            weight_attr=ParamAttr(initializer=weight_attr),
W
WenmuZhou 已提交
76
            bias_attr=False)
D
dyning 已提交
77
        self.p3_conv = nn.Conv2D(
W
WenmuZhou 已提交
78 79 80 81
            in_channels=self.out_channels,
            out_channels=self.out_channels // 4,
            kernel_size=3,
            padding=1,
littletomatodonkey's avatar
littletomatodonkey 已提交
82
            weight_attr=ParamAttr(initializer=weight_attr),
W
WenmuZhou 已提交
83
            bias_attr=False)
D
dyning 已提交
84
        self.p2_conv = nn.Conv2D(
W
WenmuZhou 已提交
85 86 87 88
            in_channels=self.out_channels,
            out_channels=self.out_channels // 4,
            kernel_size=3,
            padding=1,
littletomatodonkey's avatar
littletomatodonkey 已提交
89
            weight_attr=ParamAttr(initializer=weight_attr),
W
WenmuZhou 已提交
90 91 92 93 94 95 96 97 98 99
            bias_attr=False)

    def forward(self, x):
        c2, c3, c4, c5 = x

        in5 = self.in5_conv(c5)
        in4 = self.in4_conv(c4)
        in3 = self.in3_conv(c3)
        in2 = self.in2_conv(c2)

W
WenmuZhou 已提交
100 101 102 103 104 105
        out4 = in4 + F.upsample(
            in5, scale_factor=2, mode="nearest", align_mode=1)  # 1/16
        out3 = in3 + F.upsample(
            out4, scale_factor=2, mode="nearest", align_mode=1)  # 1/8
        out2 = in2 + F.upsample(
            out3, scale_factor=2, mode="nearest", align_mode=1)  # 1/4
W
WenmuZhou 已提交
106 107 108 109 110

        p5 = self.p5_conv(in5)
        p4 = self.p4_conv(out4)
        p3 = self.p3_conv(out3)
        p2 = self.p2_conv(out2)
W
WenmuZhou 已提交
111 112 113
        p5 = F.upsample(p5, scale_factor=8, mode="nearest", align_mode=1)
        p4 = F.upsample(p4, scale_factor=4, mode="nearest", align_mode=1)
        p3 = F.upsample(p3, scale_factor=2, mode="nearest", align_mode=1)
W
WenmuZhou 已提交
114 115 116

        fuse = paddle.concat([p5, p4, p3, p2], axis=1)
        return fuse
L
LDOUBLEV 已提交
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


class CALayer(nn.Layer):
    def __init__(self, in_channels, out_channels, kernel_size, shortcut=True):
        super(CALayer, self).__init__()
        weight_attr = paddle.nn.initializer.KaimingUniform()
        self.in_conv = nn.Conv2D(
            in_channels=in_channels,
            out_channels=self.out_channels,
            kernel_size=kernel_size,
            padding=int(kernel_size // 2),
            weight_attr=ParamAttr(initializer=weight_attr),
            bias_attr=False)
        self.se_block = SEModule(self.out_channels)
        self.shortcut = shortcut

    def forward(self, ins):
        x = self.in_conv(ins)
        if self.shortcut:
            out = x + self.se_block(x)
        else:
            out = self.se_block(x)
        return out


class CAFPN(nn.Layer):
L
fix det  
LDOUBLEV 已提交
143
    def __init__(self, in_channels, out_channels, shortcut=True, **kwargs):
L
LDOUBLEV 已提交
144
        super(CAFPN, self).__init__()
L
fix det  
LDOUBLEV 已提交
145 146 147
        self.out_channels = out_channels
        self.ins_conv = []
        self.inp_conv = []
L
LDOUBLEV 已提交
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

        for i in range(len(in_channels)):
            self.ins_conv.append(
                CALayer(
                    in_channels[i],
                    out_channels,
                    kernel_size=1,
                    shortcut=shortcut))
            self.inp_conv.append(
                CALayer(
                    out_channels,
                    out_channels // 4,
                    kernel_size=3,
                    shortcut=shortcut))

    def forward(self, x):
        c2, c3, c4, c5 = x

        in5 = self.ins_conv[3](c5)
        in4 = self.ins_conv[2](c4)
        in3 = self.ins_conv[1](c3)
        in2 = self.ins_conv[0](c2)

        out4 = in4 + F.upsample(
            in5, scale_factor=2, mode="nearest", align_mode=1)  # 1/16
        out3 = in3 + F.upsample(
            out4, scale_factor=2, mode="nearest", align_mode=1)  # 1/8
        out2 = in2 + F.upsample(
            out3, scale_factor=2, mode="nearest", align_mode=1)  # 1/4

        p5 = self.inp_conv[3](in5)
        p4 = self.inp_conv[2](out4)
        p3 = self.inp_conv[1](out3)
        p2 = self.inp_conv[0](out2)

        p5 = F.upsample(p5, scale_factor=8, mode="nearest", align_mode=1)
        p4 = F.upsample(p4, scale_factor=4, mode="nearest", align_mode=1)
        p3 = F.upsample(p3, scale_factor=2, mode="nearest", align_mode=1)

        fuse = paddle.concat([p5, p4, p3, p2], axis=1)
        return fuse
L
LDOUBLEV 已提交
189 190 191 192 193 194 195 196


class FEPAN(nn.Layer):
    def __init__(self, in_channels, out_channels, **kwargs):
        super(FEPAN, self).__init__()
        self.out_channels = out_channels
        weight_attr = paddle.nn.initializer.KaimingUniform()

L
fix det  
LDOUBLEV 已提交
197 198
        self.ins_conv = []
        self.inp_conv = []
L
LDOUBLEV 已提交
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
        # pan head
        self.pan_head_conv = []
        self.pan_lat_conv = []

        for i in range(len(in_channels)):
            self.ins_conv.append(
                nn.Conv2D(
                    in_channels=in_channels[0],
                    out_channels=self.out_channels,
                    kernel_size=1,
                    weight_attr=ParamAttr(initializer=weight_attr),
                    bias_attr=False))

            self.inp_conv.append(
                ConvBNLayer(
                    in_channels=self.out_channels,
                    out_channels=self.out_channels // 4,
                    kernel_size=9,
                    padding=4))

            if i > 0:
                self.pan_head_conv.append(
                    nn.Conv2D(
                        in_channels=self.out_channels // 4,
                        out_channels=self.out_channels // 4,
                        kernel_size=3,
                        padding=1,
                        stride=2,
                        weight_attr=ParamAttr(initializer=weight_attr),
                        bias_attr=False))
            self.pan_lat_conv.append(
                ConvBNLayer(
                    in_channels=self.out_channels // 4,
                    out_channels=self.out_channels // 4,
                    kernel_size=9,
                    padding=4))

    def forward(self, x):
        c2, c3, c4, c5 = x

        in5 = self.ins_conv[3](c5)
        in4 = self.ins_conv[2](c4)
        in3 = self.ins_conv[1](c3)
        in2 = self.ins_conv[0](c2)

        out4 = in4 + F.upsample(
            in5, scale_factor=2, mode="nearest", align_mode=1)  # 1/16
        out3 = in3 + F.upsample(
            out4, scale_factor=2, mode="nearest", align_mode=1)  # 1/8
        out2 = in2 + F.upsample(
            out3, scale_factor=2, mode="nearest", align_mode=1)  # 1/4

        f5 = self.inp_conv[3](in5)
        f4 = self.inp_conv[2](out4)
        f3 = self.inp_conv[1](out3)
        f2 = self.inp_conv[0](out2)

        pan3 = f3 + self.pan_head[0](f2)
        pan4 = f4 + self.pan_head[1](pan3)
        pan5 = f5 + self.pan_head[2](pan4)

        p2 = self.pan_lat[0](f2)
        p3 = self.pan_lat[1](pan3)
        p4 = self.pan_lat[2](pan4)
        p5 = self.pan_lat[3](pan5)

        p5 = F.upsample(p5, scale_factor=8, mode="nearest", align_mode=1)
        p4 = F.upsample(p4, scale_factor=4, mode="nearest", align_mode=1)
        p3 = F.upsample(p3, scale_factor=2, mode="nearest", align_mode=1)

        fuse = paddle.concat([p5, p4, p3, p2], axis=1)
        return fuse