def resnet18_domain(classes=4, domains=30, url=None, **kwargs):
"""Constructs a ResNet-18 model.
Args:
num_classes (int): the number of classes of the classification model
domains (int): the number of domains included in the model (#source + #auxuliary)
"""
# Instantiate original ResNet
model_origin = ResNet(BasicBlock, [2, 2, 2, 2], **kwargs)
model_origin.load_state_dict(model_zoo.load_url(url))
# Instatiate Domain ResNet
model = DomainResNet(DomainBlock, [2, 2, 2, 2], domains=domains, **kwargs)
# Copy BN stats and params from the original to the domain-based
c = 0
for m_orig in model_origin.named_modules():
if 'bn' in m_orig[0] or 'downsample.1' in m_orig[0]:
for m_doms in model.named_modules():
if m_doms[0] == m_orig[0]:
c += 1
for i in range(domains):
m_doms[1].bns[i].running_var.data[:] = m_orig[
1].running_var.data[:]
m_doms[1].bns[i].running_mean.data[:] = m_orig[
1].running_mean.data[:]
m_doms[1].scale.data[i, :] = m_orig[1].weight.data[:]
m_doms[1].bias.data[i, :] = m_orig[1].bias.data[:]
elif 'conv' in m_orig[0] or 'downsample.0' in m_orig[0]:
for m_doms in model.named_modules():
if m_doms[0] == m_orig[0]:
m_doms[1].weight.data[:] = m_orig[1].weight.data[:]
# Init classifier
num_ftrs = model.fc.in_features
model.fc = nn.Linear(num_ftrs, classes)
for m in model.modules():
if isinstance(m, nn.Linear):
m.weight.data.normal_(0, 0.0001)
return model
def se_resnet152(num_classes=1_000):
"""Constructs a ResNet-152 model.
Args:
pretrained (bool): If True, returns a model pre-trained on ImageNet
"""
model = ResNet(SEBottleneck, [3, 8, 36, 3], num_classes=num_classes)
model.avgpool = nn.AdaptiveAvgPool2d(1)
return model
if __name__ == "__main__":
import torch
model = se_resnet50()
# 替换网络层
for name, module in model.named_modules():
print("layer name:{}, layer instance:{}".format(name, module))
in_feat_num = model.fc.in_features
model.fc = nn.Linear(in_feat_num, 102)
# forward
fake_img = torch.randn((1, 3, 224, 224)) # batchsize * channel * height * width
output = model(fake_img)
print(output.shape)