def __init__(self,
encoder: E.EncoderModule,
num_classes: int,
fpn_features=128,
dropout=0.25):
super().__init__()
decoder = D.FPNDecoder(
features=encoder.output_filters,
prediction_block=DoubleConvBNRelu,
bottleneck=FPNBottleneckBlockBN,
fpn_features=fpn_features,
prediction_features=fpn_features,
)
self.encoder = encoder
self.decoder = decoder
self.fpn_fuse = FPNFuse()
self.dropout = nn.Dropout2d(dropout, inplace=True)
output_features = sum(self.decoder.output_filters)
# Final Classifier
self.final_decoder = DoubleConvBNRelu(output_features,
fpn_features // 2)
self.logits = nn.Conv2d(fpn_features // 2, num_classes, kernel_size=1)
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 __init__(self, encoder: E.EncoderModule, num_classes: int,
fpn_features: int):
super().__init__()
self.encoder = encoder
# hard-coded assumption that encoder has first layer with stride of 4
self.decoder = D.FPNDecoder(features=encoder.output_filters[1:],
prediction_block=DoubleConvRelu,
bottleneck=FPNBottleneckBlockBN,
fpn_features=fpn_features)
self.fpn_fuse = FPNFuse()
output_features = sum(self.decoder.output_filters)
self.reduce = ConvBNRelu(output_features, fpn_features * 2)
self.smooth1 = DoubleConvRelu(
fpn_features * 2 + encoder.output_filters[0], fpn_features)
self.smooth2 = DoubleConvRelu(fpn_features, fpn_features // 2)
self.dropout = nn.Dropout2d(0.5, inplace=False)
self.logits = nn.Conv2d(fpn_features // 2, num_classes, kernel_size=1)