def test_get_composed_model(self):
for backbone_name in test_backbone_name_list:
for classifier_name in test_classifier_name_list:
print(backbone_name + '+' + classifier_name + ":")
model = get_composed_model(backbone_name, classifier_name)
testModel(model)
for backbone_name in test_UNet_backbone_list:
for classifier_name in test_UNet_classifier_list:
print(backbone_name + '+' + classifier_name + ":")
model = get_composed_model(backbone_name, classifier_name)
testModel(model)
class rSoftMax(nn.Module):
def __init__(self, radix, cardinality):
super().__init__()
self.radix = radix
self.cardinality = cardinality
def forward(self, x):
batch = x.size(0)
if self.radix > 1:
x = x.view(batch, self.cardinality, self.radix, -1).transpose(1, 2)
x = F.softmax(x, dim=1)
x = x.reshape(batch, -1)
else:
x = torch.sigmoid(x)
return x
if __name__ == '__main__':
model = ResNestBase([3, 4, 6, 3],
in_channels=3,
num_classes=1,
radix=2,
groups=1,
bottleneck_width=64,
deep_stem=True,
stem_width=32,
avg_down=True,
avd=True,
avd_first=False)
testModel(model)
out = torch.cat((x, y), dim=1)
out = self.fusion(out)
elif self.fusion_function == 'max':
out = torch.max(x, y)
elif self.fusion_function == 'sum':
out = torch.add(x, y)
return out
if __name__ == '__main__':
backbone = ResNet_BB(3, clf='PointRend')
classifier = DeepLabV3Head(backbone.out_channels, 2)
modelA = SimpleSegmentationModel(backbone,
classifier,
if_point_rend_upsample=True)
testModel(modelA, eval=True)
# backbone = TwoInput_NDDRLSC_BB(3, 3, clf='PointRend')
# classifier = DeepLabV3Head(backbone.out_channels, 2)
# modelA = TwoInputSegmentationModel(backbone, classifier, if_point_rend_upsample=True)
# test2IModel(modelA, eval=True)
# 原始torchvision给出 backbone-classfier模型：backbone 将input(c, h, w) 变为 （2048， h/8, w/8)， 二分类为(1, h/8, w/8), 最后再插值回去。
# 这样使得每个像素要对应之前的64个像素，显然是表达困难的... 需要重新实现！
# input output shapetorch.Size([8, 3, 224, 224])
# backbone output shapetorch.Size([8, 2048, 28, 28])
# classifier output shapetorch.Size([8, 1, 28, 28])
# result output shapetorch.Size([8, 1, 224, 224])
# input output shapetorch.Size([8, 3, 448, 448])
# backbone output shapetorch.Size([8, 2048, 56, 56])
# classifier output shapetorch.Size([8, 1, 56, 56])
def test_get_model(self):
for model_name in model_name_list:
print(model_name + ':')
model = get_model(model_name)
testModel(model)
x2 = self.down1(x1)
x3 = self.down2(x2)
x4 = self.down3(x3)
x5 = self.down4(x4)
x6 = self.up1(x5, x4)
x7 = self.up2(x6, x3)
x8 = self.up3(x7, x2)
x9 = self.up4(x8, x1)
x6 = F.interpolate(x6, size=size, mode='bilinear', align_corners=True)
x6 = self.after_upsample6(x6)
x6 = torch.mul(x6, self.layers_weights[0])
x7 = F.interpolate(x7, size=size, mode='bilinear', align_corners=True)
x7 = self.after_upsample7(x7)
x7 = torch.mul(x7, self.layers_weights[1])
x8 = F.interpolate(x8, size=size, mode='bilinear', align_corners=True)
x8 = self.after_upsample8(x8)
x8 = torch.mul(x8, self.layers_weights[2])
x9 = F.interpolate(x9, size=size, mode='bilinear', align_corners=True)
x9 = self.after_upsample9(x9)
x9 = torch.mul(x9, self.layers_weights[3])
x = torch.cat([x6, x7, x8, x9], dim=1)
x = self.outc(x)
return x
if __name__ == '__main__':
testModel(DeepMadSeg(5, 1))
filters_dict['contractive_' + str(i)][1],
filt_pair[0]))
filters_dict['expansive_' + str(i)] = (filt_pair[1], filt_pair[0])
filt_pair[1] = filt_pair[0]
filt_pair[0] = filt_pair[0] // 2
self.output = nn.Conv2d(filt_pair[1], n_labels, kernel_size=1)
filters_dict['output'] = (filt_pair[1], n_labels)
self.filters_dict = filters_dict
# final_forward
def forward(self, x):
c00, c0 = self.contractive_0(x)
c11, c1 = self.contractive_1(c0)
c22, c2 = self.contractive_2(c1)
c33, c3 = self.contractive_3(c2)
bottle = self.bottleneck(c3)
u3 = F.relu(self.expansive_3(bottle, c33))
u2 = F.relu(self.expansive_2(u3, c22))
u1 = F.relu(self.expansive_1(u2, c11))
u0 = F.relu(self.expansive_0(u1, c00))
return F.softmax(self.output(u0), dim=1)
if __name__ == '__main__':
testModel(U_Net(5, 1))
testModel(R2U_Net(5, 1))
testModel(AttU_Net(5, 1))
testModel(R2AttU_Net(5, 1))
testModel(NestedUNet(5, 1))