nas_fpn.py
6.58 KB
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
# Copyright (c) OpenMMLab. All rights reserved.
import torch.nn as nn
from mmcv.cnn import ConvModule
from mmcv.ops.merge_cells import GlobalPoolingCell, SumCell
from mmcv.runner import BaseModule, ModuleList
from ..builder import NECKS
@NECKS.register_module()
class NASFPN(BaseModule):
"""NAS-FPN.
Implementation of `NAS-FPN: Learning Scalable Feature Pyramid Architecture
for Object Detection <https://arxiv.org/abs/1904.07392>`_
Args:
in_channels (List[int]): Number of input channels per scale.
out_channels (int): Number of output channels (used at each scale)
num_outs (int): Number of output scales.
stack_times (int): The number of times the pyramid architecture will
be stacked.
start_level (int): Index of the start input backbone level used to
build the feature pyramid. Default: 0.
end_level (int): Index of the end input backbone level (exclusive) to
build the feature pyramid. Default: -1, which means the last level.
add_extra_convs (bool): It decides whether to add conv
layers on top of the original feature maps. Default to False.
If True, its actual mode is specified by `extra_convs_on_inputs`.
init_cfg (dict or list[dict], optional): Initialization config dict.
"""
def __init__(self,
in_channels,
out_channels,
num_outs,
stack_times,
start_level=0,
end_level=-1,
add_extra_convs=False,
norm_cfg=None,
init_cfg=dict(type='Caffe2Xavier', layer='Conv2d')):
super(NASFPN, self).__init__(init_cfg)
assert isinstance(in_channels, list)
self.in_channels = in_channels
self.out_channels = out_channels
self.num_ins = len(in_channels) # num of input feature levels
self.num_outs = num_outs # num of output feature levels
self.stack_times = stack_times
self.norm_cfg = norm_cfg
if end_level == -1:
self.backbone_end_level = self.num_ins
assert num_outs >= self.num_ins - start_level
else:
# if end_level < inputs, no extra level is allowed
self.backbone_end_level = end_level
assert end_level <= len(in_channels)
assert num_outs == end_level - start_level
self.start_level = start_level
self.end_level = end_level
self.add_extra_convs = add_extra_convs
# add lateral connections
self.lateral_convs = nn.ModuleList()
for i in range(self.start_level, self.backbone_end_level):
l_conv = ConvModule(
in_channels[i],
out_channels,
1,
norm_cfg=norm_cfg,
act_cfg=None)
self.lateral_convs.append(l_conv)
# add extra downsample layers (stride-2 pooling or conv)
extra_levels = num_outs - self.backbone_end_level + self.start_level
self.extra_downsamples = nn.ModuleList()
for i in range(extra_levels):
extra_conv = ConvModule(
out_channels, out_channels, 1, norm_cfg=norm_cfg, act_cfg=None)
self.extra_downsamples.append(
nn.Sequential(extra_conv, nn.MaxPool2d(2, 2)))
# add NAS FPN connections
self.fpn_stages = ModuleList()
for _ in range(self.stack_times):
stage = nn.ModuleDict()
# gp(p6, p4) -> p4_1
stage['gp_64_4'] = GlobalPoolingCell(
in_channels=out_channels,
out_channels=out_channels,
out_norm_cfg=norm_cfg)
# sum(p4_1, p4) -> p4_2
stage['sum_44_4'] = SumCell(
in_channels=out_channels,
out_channels=out_channels,
out_norm_cfg=norm_cfg)
# sum(p4_2, p3) -> p3_out
stage['sum_43_3'] = SumCell(
in_channels=out_channels,
out_channels=out_channels,
out_norm_cfg=norm_cfg)
# sum(p3_out, p4_2) -> p4_out
stage['sum_34_4'] = SumCell(
in_channels=out_channels,
out_channels=out_channels,
out_norm_cfg=norm_cfg)
# sum(p5, gp(p4_out, p3_out)) -> p5_out
stage['gp_43_5'] = GlobalPoolingCell(with_out_conv=False)
stage['sum_55_5'] = SumCell(
in_channels=out_channels,
out_channels=out_channels,
out_norm_cfg=norm_cfg)
# sum(p7, gp(p5_out, p4_2)) -> p7_out
stage['gp_54_7'] = GlobalPoolingCell(with_out_conv=False)
stage['sum_77_7'] = SumCell(
in_channels=out_channels,
out_channels=out_channels,
out_norm_cfg=norm_cfg)
# gp(p7_out, p5_out) -> p6_out
stage['gp_75_6'] = GlobalPoolingCell(
in_channels=out_channels,
out_channels=out_channels,
out_norm_cfg=norm_cfg)
self.fpn_stages.append(stage)
def forward(self, inputs):
"""Forward function."""
# build P3-P5
feats = [
lateral_conv(inputs[i + self.start_level])
for i, lateral_conv in enumerate(self.lateral_convs)
]
# build P6-P7 on top of P5
for downsample in self.extra_downsamples:
feats.append(downsample(feats[-1]))
p3, p4, p5, p6, p7 = feats
for stage in self.fpn_stages:
# gp(p6, p4) -> p4_1
p4_1 = stage['gp_64_4'](p6, p4, out_size=p4.shape[-2:])
# sum(p4_1, p4) -> p4_2
p4_2 = stage['sum_44_4'](p4_1, p4, out_size=p4.shape[-2:])
# sum(p4_2, p3) -> p3_out
p3 = stage['sum_43_3'](p4_2, p3, out_size=p3.shape[-2:])
# sum(p3_out, p4_2) -> p4_out
p4 = stage['sum_34_4'](p3, p4_2, out_size=p4.shape[-2:])
# sum(p5, gp(p4_out, p3_out)) -> p5_out
p5_tmp = stage['gp_43_5'](p4, p3, out_size=p5.shape[-2:])
p5 = stage['sum_55_5'](p5, p5_tmp, out_size=p5.shape[-2:])
# sum(p7, gp(p5_out, p4_2)) -> p7_out
p7_tmp = stage['gp_54_7'](p5, p4_2, out_size=p7.shape[-2:])
p7 = stage['sum_77_7'](p7, p7_tmp, out_size=p7.shape[-2:])
# gp(p7_out, p5_out) -> p6_out
p6 = stage['gp_75_6'](p7, p5, out_size=p6.shape[-2:])
return p3, p4, p5, p6, p7