det_resnet_vd.py 7.73 KB
Newer Older
WenmuZhou's avatar
WenmuZhou committed
1
# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
LDOUBLEV's avatar
LDOUBLEV committed
2
#
WenmuZhou's avatar
WenmuZhou committed
3
4
5
# 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
LDOUBLEV's avatar
LDOUBLEV committed
6
7
8
#
#    http://www.apache.org/licenses/LICENSE-2.0
#
WenmuZhou's avatar
WenmuZhou committed
9
10
11
12
13
# 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.
LDOUBLEV's avatar
LDOUBLEV committed
14
15
16
17
18

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

19
import paddle
WenmuZhou's avatar
WenmuZhou committed
20
from paddle import ParamAttr
21
import paddle.nn as nn
WenmuZhou's avatar
WenmuZhou committed
22
import paddle.nn.functional as F
LDOUBLEV's avatar
LDOUBLEV committed
23
24
25
26

__all__ = ["ResNet"]


WenmuZhou's avatar
WenmuZhou committed
27
class ConvBNLayer(nn.Layer):
littletomatodonkey's avatar
littletomatodonkey committed
28
29
30
31
32
33
34
35
    def __init__(self,
                 in_channels,
                 out_channels,
                 kernel_size,
                 stride=1,
                 groups=1,
                 is_vd_mode=False,
                 act=None):
WenmuZhou's avatar
WenmuZhou committed
36
        super(ConvBNLayer, self).__init__()
37
38

        self.is_vd_mode = is_vd_mode
WenmuZhou's avatar
WenmuZhou committed
39
        self._pool2d_avg = nn.AvgPool2D(
40
            kernel_size=2, stride=2, padding=0, ceil_mode=True)
WenmuZhou's avatar
WenmuZhou committed
41
        self._conv = nn.Conv2D(
WenmuZhou's avatar
WenmuZhou committed
42
43
44
            in_channels=in_channels,
            out_channels=out_channels,
            kernel_size=kernel_size,
LDOUBLEV's avatar
LDOUBLEV committed
45
            stride=stride,
WenmuZhou's avatar
WenmuZhou committed
46
            padding=(kernel_size - 1) // 2,
LDOUBLEV's avatar
LDOUBLEV committed
47
48
            groups=groups,
            bias_attr=False)
littletomatodonkey's avatar
littletomatodonkey committed
49
        self._batch_norm = nn.BatchNorm(out_channels, act=act)
WenmuZhou's avatar
WenmuZhou committed
50

51
52
53
54
55
56
    def forward(self, inputs):
        if self.is_vd_mode:
            inputs = self._pool2d_avg(inputs)
        y = self._conv(inputs)
        y = self._batch_norm(y)
        return y
WenmuZhou's avatar
WenmuZhou committed
57
58


59
class BottleneckBlock(nn.Layer):
WenmuZhou's avatar
WenmuZhou committed
60
61
62
    def __init__(self,
                 in_channels,
                 out_channels,
63
64
                 stride,
                 shortcut=True,
littletomatodonkey's avatar
littletomatodonkey committed
65
                 if_first=False):
WenmuZhou's avatar
WenmuZhou committed
66
        super(BottleneckBlock, self).__init__()
67

WenmuZhou's avatar
WenmuZhou committed
68
69
70
71
        self.conv0 = ConvBNLayer(
            in_channels=in_channels,
            out_channels=out_channels,
            kernel_size=1,
littletomatodonkey's avatar
littletomatodonkey committed
72
            act='relu')
WenmuZhou's avatar
WenmuZhou committed
73
74
75
76
        self.conv1 = ConvBNLayer(
            in_channels=out_channels,
            out_channels=out_channels,
            kernel_size=3,
LDOUBLEV's avatar
LDOUBLEV committed
77
            stride=stride,
littletomatodonkey's avatar
littletomatodonkey committed
78
            act='relu')
WenmuZhou's avatar
WenmuZhou committed
79
80
81
82
        self.conv2 = ConvBNLayer(
            in_channels=out_channels,
            out_channels=out_channels * 4,
            kernel_size=1,
littletomatodonkey's avatar
littletomatodonkey committed
83
            act=None)
LDOUBLEV's avatar
LDOUBLEV committed
84

85
86
87
88
89
90
        if not shortcut:
            self.short = ConvBNLayer(
                in_channels=in_channels,
                out_channels=out_channels * 4,
                kernel_size=1,
                stride=1,
littletomatodonkey's avatar
littletomatodonkey committed
91
                is_vd_mode=False if if_first else True)
92
93
94
95
96
97
98

        self.shortcut = shortcut

    def forward(self, inputs):
        y = self.conv0(inputs)
        conv1 = self.conv1(y)
        conv2 = self.conv2(conv1)
WenmuZhou's avatar
WenmuZhou committed
99

100
101
102
103
        if self.shortcut:
            short = inputs
        else:
            short = self.short(inputs)
WenmuZhou's avatar
WenmuZhou committed
104
105
        y = paddle.add(x=short, y=conv2)
        y = F.relu(y)
WenmuZhou's avatar
WenmuZhou committed
106
        return y
LDOUBLEV's avatar
LDOUBLEV committed
107
108


WenmuZhou's avatar
WenmuZhou committed
109
class BasicBlock(nn.Layer):
littletomatodonkey's avatar
littletomatodonkey committed
110
111
112
113
114
115
116
    def __init__(
            self,
            in_channels,
            out_channels,
            stride,
            shortcut=True,
            if_first=False, ):
WenmuZhou's avatar
WenmuZhou committed
117
        super(BasicBlock, self).__init__()
118
        self.stride = stride
WenmuZhou's avatar
WenmuZhou committed
119
120
121
122
        self.conv0 = ConvBNLayer(
            in_channels=in_channels,
            out_channels=out_channels,
            kernel_size=3,
LDOUBLEV's avatar
LDOUBLEV committed
123
            stride=stride,
littletomatodonkey's avatar
littletomatodonkey committed
124
            act='relu')
WenmuZhou's avatar
WenmuZhou committed
125
126
127
128
        self.conv1 = ConvBNLayer(
            in_channels=out_channels,
            out_channels=out_channels,
            kernel_size=3,
littletomatodonkey's avatar
littletomatodonkey committed
129
            act=None)
WenmuZhou's avatar
WenmuZhou committed
130

131
132
133
134
135
136
        if not shortcut:
            self.short = ConvBNLayer(
                in_channels=in_channels,
                out_channels=out_channels,
                kernel_size=1,
                stride=1,
littletomatodonkey's avatar
littletomatodonkey committed
137
                is_vd_mode=False if if_first else True)
138
139
140
141
142
143

        self.shortcut = shortcut

    def forward(self, inputs):
        y = self.conv0(inputs)
        conv1 = self.conv1(y)
WenmuZhou's avatar
WenmuZhou committed
144

145
146
147
148
        if self.shortcut:
            short = inputs
        else:
            short = self.short(inputs)
WenmuZhou's avatar
WenmuZhou committed
149
150
        y = paddle.add(x=short, y=conv1)
        y = F.relu(y)
151
        return y
WenmuZhou's avatar
WenmuZhou committed
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
class ResNet(nn.Layer):
    def __init__(self, in_channels=3, layers=50, **kwargs):
        super(ResNet, self).__init__()

        self.layers = layers
        supported_layers = [18, 34, 50, 101, 152, 200]
        assert layers in supported_layers, \
            "supported layers are {} but input layer is {}".format(
                supported_layers, layers)

        if layers == 18:
            depth = [2, 2, 2, 2]
        elif layers == 34 or layers == 50:
            depth = [3, 4, 6, 3]
        elif layers == 101:
            depth = [3, 4, 23, 3]
        elif layers == 152:
            depth = [3, 8, 36, 3]
        elif layers == 200:
            depth = [3, 12, 48, 3]
        num_channels = [64, 256, 512,
                        1024] if layers >= 50 else [64, 64, 128, 256]
        num_filters = [64, 128, 256, 512]

        self.conv1_1 = ConvBNLayer(
            in_channels=in_channels,
            out_channels=32,
            kernel_size=3,
            stride=2,
littletomatodonkey's avatar
littletomatodonkey committed
183
            act='relu')
184
185
186
187
188
        self.conv1_2 = ConvBNLayer(
            in_channels=32,
            out_channels=32,
            kernel_size=3,
            stride=1,
littletomatodonkey's avatar
littletomatodonkey committed
189
            act='relu')
190
191
192
193
194
        self.conv1_3 = ConvBNLayer(
            in_channels=32,
            out_channels=64,
            kernel_size=3,
            stride=1,
littletomatodonkey's avatar
littletomatodonkey committed
195
            act='relu')
WenmuZhou's avatar
WenmuZhou committed
196
        self.pool2d_max = nn.MaxPool2D(kernel_size=3, stride=2, padding=1)
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212

        self.stages = []
        self.out_channels = []
        if layers >= 50:
            for block in range(len(depth)):
                block_list = []
                shortcut = False
                for i in range(depth[block]):
                    bottleneck_block = self.add_sublayer(
                        'bb_%d_%d' % (block, i),
                        BottleneckBlock(
                            in_channels=num_channels[block]
                            if i == 0 else num_filters[block] * 4,
                            out_channels=num_filters[block],
                            stride=2 if i == 0 and block != 0 else 1,
                            shortcut=shortcut,
littletomatodonkey's avatar
littletomatodonkey committed
213
                            if_first=block == i == 0))
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
                    shortcut = True
                    block_list.append(bottleneck_block)
                self.out_channels.append(num_filters[block] * 4)
                self.stages.append(nn.Sequential(*block_list))
        else:
            for block in range(len(depth)):
                block_list = []
                shortcut = False
                for i in range(depth[block]):
                    basic_block = self.add_sublayer(
                        'bb_%d_%d' % (block, i),
                        BasicBlock(
                            in_channels=num_channels[block]
                            if i == 0 else num_filters[block],
                            out_channels=num_filters[block],
                            stride=2 if i == 0 and block != 0 else 1,
                            shortcut=shortcut,
littletomatodonkey's avatar
littletomatodonkey committed
231
                            if_first=block == i == 0))
232
233
234
235
                    shortcut = True
                    block_list.append(basic_block)
                self.out_channels.append(num_filters[block])
                self.stages.append(nn.Sequential(*block_list))
WenmuZhou's avatar
WenmuZhou committed
236

237
238
239
240
241
242
243
244
245
246
    def forward(self, inputs):
        y = self.conv1_1(inputs)
        y = self.conv1_2(y)
        y = self.conv1_3(y)
        y = self.pool2d_max(y)
        out = []
        for block in self.stages:
            y = block(y)
            out.append(y)
        return out