def fpn256_resnext50(num_classes=1, num_channels=3):
assert num_channels == 3
encoder = E.SEResNeXt50Encoder()
decoder = D.FPNDecoder(features=encoder.output_filters,
prediction_block=DoubleConvRelu,
bottleneck=FPNBottleneckBlockBN,
fpn_features=256)
return SegmentationModel(encoder, decoder, num_classes)
def seresnext50_unet64(input_channels=3, num_classes=1, dropout=0.0, pretrained=True):
encoder = E.SEResNeXt50Encoder(pretrained=pretrained, layers=[0, 1, 2, 3, 4])
if input_channels != 3:
encoder.change_input_channels(input_channels)
return UnetV3SegmentationModel(
encoder,
num_classes=num_classes,
unet_channels=[64, 128, 256, 256],
dropout=dropout,
abn_block=partial(ABN, activation=ACT_RELU),
)
def seresnext50_unet_v2(input_channels=6,
num_classes=5,
dropout=0.0,
pretrained=True,
classifiers=True):
encoder = E.SEResNeXt50Encoder(pretrained=pretrained,
layers=[0, 1, 2, 3, 4])
return UnetV2SegmentationModel(
encoder,
num_classes=num_classes,
disaster_type_classes=len(DISASTER_TYPES) if classifiers else None,
damage_type_classes=len(DAMAGE_TYPES) if classifiers else None,
unet_channels=[64, 128, 256, 256],
dropout=dropout,
abn_block=partial(ABN, activation=ACT_RELU),
)
def resnext50_fpn(num_classes=1, fpn_features=128):
encoder = E.SEResNeXt50Encoder()
return FPNSegmentationModel(encoder, num_classes, fpn_features, dropout=0.5)
def seresnext50_fpncat128(num_classes=5, dropout=0.0, pretrained=True):
encoder = E.SEResNeXt50Encoder(pretrained=pretrained)
return FPNCatSegmentationModel(encoder, num_classes=num_classes, fpn_channels=128, dropout=dropout)
def hd_fpn_resnext50(num_classes=1, num_channels=3, fpn_features=128):
assert num_channels == 3
encoder = E.SEResNeXt50Encoder(layers=[0, 1, 2, 3, 4])
return HiResSegmentationModel(encoder, num_classes, fpn_features)
