PSMNet debug(二)

最新推荐文章于 2024-09-13 22:08:10 发布

原创最新推荐文章于 2024-09-13 22:08:10 发布 · 737 阅读

1 ·

CC 4.0 BY-SA版权

本文详细探讨了PSMNet中特征提取模块的实现，包括convbn层、特征提取层、不同尺度的特征融合及左右特征图的合并过程，展示了各层输出的尺寸变化，深入理解其在立体匹配中的作用。

摘要生成于 C知道，由 DeepSeek-R1 满血版支持，前往体验 >

submodule.py

1
3
384
1248

def convbn(in_planes, out_planes, kernel_size, stride, pad, dilation):

return nn.Sequential(nn.Conv2d(in_planes, out_planes, kernel_size=kernel_size, stride=stride, padding=dilation if dilation > 1 else pad, dilation = dilation, bias=False),
nn.BatchNorm2d(out_planes))

class feature_extraction(nn.Module):

def __init__(self):
print('feature_extraction 123')
super(feature_extraction, self).__init__()
self.inplanes = 32
self.firstconv = nn.Sequential(convbn(3, 32, 3, 2, 1, 1), #下采样，缩小一倍。w,h各除以2.filter num:32。
nn.ReLU(inplace=True),
convbn(32, 32, 3, 1, 1, 1),
nn.ReLU(inplace=True),
convbn(32, 32, 3, 1, 1, 1),
nn.ReLU(inplace=True))

对应下面。

def forward(self, x):

output = self.firstconv(x)
print('submodule firstconv:')
print(output.size()[0])
print(output.size()[1])
print(output.size()[2])
print(output.size()[3])

submodule firstconv:
1
32
192
624

self.lastconv = nn.Sequential(convbn(320, 128, 3, 1, 1, 1),
nn.ReLU(inplace=True),
nn.Conv2d(128, 32, kernel_size=1, padding=0, stride = 1, bias=False))

output_feature = torch.cat((output_raw, output_skip, output_branch4, output_branch3, output_branch2, output_branch1),

#output_branch4, output_branch3, output_branch2, output_branch1 第2个通道分别是32，output_raw是64，output_skip是128。其它通道都是一样的。1,96,312。

和起来第2个通道是32. (32*4=128) + 128 + 64 = 320。

submodule torch.cat:
1
320
96
312

1)
print('submodule torch.cat:')
print(output_feature.size()[0])
print(output_feature.size()[1])
print(output_feature.size()[2])
print(output_feature.size()[3])

output_feature = self.lastconv(output_feature)
print('submodule lastconv:')
print(output_feature.size()[0])
print(output_feature.size()[1])
print(output_feature.size()[2])
print(output_feature.size()[3])

submodule torch.cat:
1
320
96
312
submodule lastconv:
1
32
96
312

self.branch2 = nn.Sequential(nn.AvgPool2d((32, 32), stride=(32,32)),#32倍缩小。第3个通道96变为3，第4个通道312变为9
convbn(128, 32, 1, 1, 0, 1), # 第2个通道128变为了32。
nn.ReLU(inplace=True))

output_skip = self.layer4(output)
print('submodule layer4 output_skip:')
print(output_skip.size()[0])
print(output_skip.size()[1])
print(output_skip.size()[2])
print(output_skip.size()[3])

output_branch1 = self.branch1(output_skip)
output_branch1 = F.upsample(output_branch1, (output_skip.size()[2],output_skip.size()[3]),mode='bilinear')

output_branch2 = self.branch2(output_skip)
print('submodule branch2 output_branch2:')
print(output_branch2.size()[0])
print(output_branch2.size()[1])
print(output_branch2.size()[2])
print(output_branch2.size()[3])
output_branch2 = F.upsample(output_branch2, (output_skip.size()[2],output_skip.size()[3]),mode='bilinear')
print('submodule upsample output_branch2:')
print(output_branch2.size()[0])
print(output_branch2.size()[1])
print(output_branch2.size()[2])
print(output_branch2.size()[3])

打印：

submodule layer4 output_skip:
1
128
96
312
submodule branch2 output_branch2:
1
32
3
9
submodule upsample output_branch2:
1
32
96
312

self.branch3 = nn.Sequential(nn.AvgPool2d((16, 16), stride=(16,16)),#16倍缩小。96变为6.312变为19。
convbn(128, 32, 1, 1, 0, 1),# 第2维的通道128变为32
nn.ReLU(inplace=True))

output_branch3 = self.branch3(output_skip)
print('submodule branch3 output_branch3:')
print(output_branch3.size()[0])
print(output_branch3.size()[1])
print(output_branch3.size()[2])
print(output_branch3.size()[3])
output_branch3 = F.upsample(output_branch3, (output_skip.size()[2],output_skip.size()[3]),mode='bilinear')
print('submodule upsample output_branch3:')
print(output_branch3.size()[0])
print(output_branch3.size()[1])
print(output_branch3.size()[2])
print(output_branch3.size()[3])

submodule layer4 output_skip:
1
128
96
312

submodule branch3 output_branch3:
1
32
6
19
submodule upsample output_branch3:
1
32
96
312

self.layer2 = self._make_layer(BasicBlock, 64, 16, 2,1,1)

output_raw = self.layer2(output)
print('submodule layer2 output_raw:')
print(output_raw.size()[0])
print(output_raw.size()[1])
print(output_raw.size()[2])
print(output_raw.size()[3])

打印结果：

submodule layer2 output_raw:
1
64
96
312

class feature_extraction(nn.Module):

def _make_layer(self, block, planes, blocks, stride, pad, dilation):
downsample = None
if stride != 1 or self.inplanes != planes * block.expansion:
downsample = nn.Sequential(
nn.Conv2d(self.inplanes, planes * block.expansion,
kernel_size=1, stride=stride, bias=False),
nn.BatchNorm2d(planes * block.expansion),)

layers = []
layers.append(block(self.inplanes, planes, stride, downsample, pad, dilation))
self.inplanes = planes * block.expansion
for i in range(1, blocks):
layers.append(block(self.inplanes, planes,1,None,pad,dilation))

return nn.Sequential(*layers)

def __init__(self):

self.layer1 = self._make_layer(BasicBlock, 32, 3, 1,1,1)
self.layer2 = self._make_layer(BasicBlock, 64, 16, 2,1,1)
self.layer3 = self._make_layer(BasicBlock, 128, 3, 1,1,1)
self.layer4 = self._make_layer(BasicBlock, 128, 3, 1,1,2)

output = self.layer1(output)
output_raw = self.layer2(output)
print('submodule layer2 output_raw:')
print(output_raw.size()[0])
print(output_raw.size()[1])
print(output_raw.size()[2])
print(output_raw.size()[3])

output = self.layer3(output_raw)
output_skip = self.layer4(output)
print('submodule layer4 output_skip:')
print(output_skip.size()[0])
print(output_skip.size()[1])
print(output_skip.size()[2])
print(output_skip.size()[3])
打印结果，看第2个通道的。64和128。

submodule layer2 output_raw:
1
64
96
312
submodule layer4 output_skip:
1
128
96
312

输入彩色图像。输出特征图像是H/4=96，W/4=312。32个通道。

384
1248

1
32
96
312

左右特征图合并。视差也以最大的除以4。

cost = Variable(torch.FloatTensor(refimg_fea.size()[0], refimg_fea.size()[1]*2, self.maxdisp/4, refimg_fea.size()[2], refimg_fea.size()[3]).zero_())

self.maxdisp/4是D。

refimg_fea.size()[1]*2是左右图合并的。特征通道个数。

stackhourglass refim

refimg_fea = self.feature_extraction(left)
print('PSMNet forward 0000')
targetimg_fea = self.feature_extraction(right)
print('PSMNet forward 1111')

print('stackhourglass refimg_fea disp')
print(refimg_fea.size()[0])
print(refimg_fea.size()[1])
print(refimg_fea.size()[2])
print(refimg_fea.size()[3])

print('stackhourglass targetimg_fea disp')
print(targetimg_fea.size()[0])
print(targetimg_fea.size()[1])
print(targetimg_fea.size()[2])
print(targetimg_fea.size()[3])

g_fea disp
1
32
96
312
stackhourglass targetimg_fea disp
1
32
96
312

cost = Variable(torch.FloatTensor(refimg_fea.size()[0], refimg_fea.size()[1]*2, self.maxdisp/4, refimg_fea.size()[2], refimg_fea.size()[3]).zero_())
print('stackhourglass cost disp')
print(cost.size()[0])
print(cost.size()[1])#2个特征图合并。32*2 = 64
print(cost.size()[2])#视差D
print(cost.size()[3])#高度
print(cost.size()[4])#高度

打印

stackhourglass cost disp
1
64
48
96
312

10.

stackhourglass.py

for i in range(self.maxdisp/4):#48*4 = 192
#print(i)
if i > 0 :
cost[:, :refimg_fea.size()[1], i, :,i:] = refimg_fea[:,:,:,i:] # 1:32 存储左特征图
cost[:, refimg_fea.size()[1]:, i, :,i:] = targetimg_fea[:,:,:,:-i] #32:64 存储右特征图
else:
cost[:, :refimg_fea.size()[1], i, :,:] = refimg_fea # 1:32 存储左特征图
cost[:, refimg_fea.size()[1]:, i, :,:] = targetimg_fea #32:64 存储右特征图