def __init__(self, in_channel=3, n_class=1):
super(Efficient_SIIMD, self).__init__()
efficient = EfficientNet.from_pretrained('efficientnet-b0')
self.block0 = nn.Sequential(efficient._conv_stem, efficient._bn0)
print(len(efficient._blocks))
self.block1 = nn.Sequential(*efficient._blocks[:5])
self.block2 = nn.Sequential(*efficient._blocks[5:10])
self.block3 = nn.Sequential(*efficient._blocks[10:16])
self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
self.classifier = Classifier(
efficient._blocks[-1]._project_conv.out_channels, 128, n_class)
# for para in self.block0.parameters():
# para.requires_grad = False
#
# for para in self.block1.parameters():
# para.requires_grad = False
#
# for para in self.block2.parameters():
# para.requires_grad = False
# for para in self.block3.parameters():
# para.requires_grad = False
for m in self.classifier.modules():
if isinstance(m, nn.Linear):
init_weights(m, init_type='kaiming')
elif isinstance(m, nn.BatchNorm1d):
init_weights(m, init_type='kaiming')
def __init__(self, in_channel, mid_channel, n_class, dropout=0.5):
super(Classifier, self).__init__()
# solution 1
self.layer1 = nn.Dropout(0.5)
self.layer2 = nn.Sequential(nn.Linear(in_channel, mid_channel),
nn.BatchNorm1d(mid_channel), nn.ReLU())
self.layer3 = nn.Linear(mid_channel, n_class)
for m in self.modules():
if isinstance(m, nn.Conv2d) or isinstance(m, nn.Linear):
init_weights(m, init_type='kaiming')
elif isinstance(m, nn.BatchNorm1d):
init_weights(m, init_type='kaiming')
def __init__(self, in_channel=3, n_class=1):
super(Regnet_SIIM, self).__init__()
backbone = regnet_800M(pretrained=True)
self.stem = list(list(backbone.children())[0].children())[0]
self.feature = nn.Sequential(*list(list(backbone.children())[0].children())[1:-1])
self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
self.classifier = Classifier(672, 320, n_class)
for m in self.classifier.modules():
if isinstance(m, nn.Linear):
init_weights(m, init_type='kaiming')
elif isinstance(m, nn.BatchNorm1d):
init_weights(m, init_type='kaiming')
def __init__(self, in_channel=3, n_class=1):
super(ResNet_SIIM, self).__init__()
resnet = models.resnet34(True)
self.layer0 = nn.Sequential(resnet.conv1, resnet.bn1, resnet.maxpool)
self.layer1 = resnet.layer1
self.layer2 = resnet.layer2
self.layer3 = resnet.layer3
self.layer4 = resnet.layer4
self.avgpool = resnet.avgpool
self.fc_reduce = nn.Linear(resnet.fc.in_features, 32)
# self.features = nn.Sequential(*list(resnet.children())[:-1])
self.classifier = Classifier(32, 32, n_class)
for m in self.classifier.modules():
if isinstance(m, nn.Linear):
init_weights(m, init_type='kaiming')
elif isinstance(m, nn.BatchNorm1d):
init_weights(m, init_type='kaiming')