__________________
< Keep it a secret >
------------------
\
\
\
___ _____ ___
/ \ / /| / \
| | / / | | |
| | /____/ | | |
| | | | | | |
| | | {} | / | |
^__^ | | |____|/ | |
(oo)\_______ | | |==| | |
(__)\ )\/\ | \___________/ |
||----w | | |
|| ||
1. 陆上模块提取特征层前10层
1.1 configs/nas_trinity/2stage_hitdet_out10.py
复制2stage_hitdet.py, 重命名为2stage_hitdet_out10.py,修改backbone:
1
2
3
4
5
6
7
8
model = dict(
type='FasterRCNN',
pretrained='./ImageNet-pretrained/fbhit_7747.pth',
backbone=dict(
type='FBNet_out10',
out_indices=(4, 8, 16, 22),
frozen_stages=-1,
arch='fbnet_hit'),
1.2 mmdet/models/backbones/fbnet_out10.py
复制fbnet.py, 重命名为fbnet_out10.py,修改forward函数:
在陆上模块前向传播时，输出前10层特征层。
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
def forward(self, x, alphas=None):
outs = []
cnt = 0
out10 = []
for i, layer in enumerate(self.layers):
# fuhao7i------------------
# 提取陆上模块1——10层
# =========================
x = layer(x)
if i in self.out_indices:
outs.append(x)
if i > 0 and i <= 10:
out10.append(x)
return outs, out10
1.3 mmdet/models/backbones/init.py
加入fbnet_out10
1
__all__ = ['ResNet', 'make_res_layer', 'ResNeXt', 'SSDVGG', 'HRNet', 'MobileNetV2', 'DetNas', 'FBNet', 'MnasNet', 'FBNet_out10']
1.4 mmdet/models/detectors/two_stage.py
新增一个extract_feat_out10函数
1
2
3
4
5
6
7
8
9
10
11
12
13
# fuhao7i------------------
# 提取陆上模块1——10层
# =========================
def extract_feat_out10(self, img):
x, out10 = self.backbone(img)
if self.with_neck:
x = self.neck(x)
if len(x) >= 2:
if x[1] is not None:
x = x
else:
x = x[0]
return x, out10
2. 和水下模块前10层特征层进行堆叠
2.1 mmdet/models/backbones/fbnet.py
首先修改build_backbone函数，因为前10层特征层进行堆叠时，通道数要加倍。
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
def build_backbone(self, arch, input_size):
genotypes = predefine_archs[arch]['genotypes']
strides = predefine_archs[arch]['strides']
out_channels = predefine_archs[arch]['out_channels']
self.layers = nn.ModuleList()
self.layers.append(ConvBNReLU(input_size, in_channels=3, out_channels=out_channels[0], kernel_size=3, stride=strides[0], padding=1,
bias=True, relu_type='relu', bn_type='bn'))
input_size = input_size // strides[0]
_in_channels = out_channels[0]
# fuhao7i------------------
# 前10层特征层堆叠，输入的特征层会翻倍(x1 + x2)
# =========================
index_out = 0
for genotype, stride, _out_channels in zip(genotypes[1:], strides[1:], out_channels[1:]):
if genotype.endswith('sb'):
self.layers.append(SUPER_PRIMITIVES[genotype](input_size, _in_channels, _out_channels, stride))
else:
self.layers.append(PRIMITIVES[genotype](input_size, _in_channels, _out_channels, stride))
input_size = input_size // stride
if index_out < 10:
_in_channels = _out_channels * 2
else:
_in_channels = _out_channels
print("backbone.layers = ", len(self.layers))
for m in self.modules():
if isinstance(m, nn.SyncBatchNorm):
m._specify_ddp_gpu_num(1)
然后修改forward函数:
1
2
3
4
5
6
7
8
9
10
11
def forward(self, x, out10, alphas=None):
outs = []
cnt = 0
for i, layer in enumerate(self.layers):
x = layer(x)
if i > 0 and i <= 10:
x = x + out10[i-1]
if i in self.out_indices:
outs.append(x)
return outs
3. 对重新构建的网络进行训练
tools/train.py –> mmdet/apis/train.py –> mmcv/runner/runner.py
