Pytorch FCN segmentation model architecture question

vision
1

Hi,

I am trying to do semantic segmentation and thought of trying out torchvision.models.segmentation.fcn_resnet50. I just see down convolution in this model and there is no upsampling happening through conv transpose layers. Then, how does the segmentation output be of same size as of the input?

Thanks

2

_SimpleSegmenationModel uses F.interpolate(x, size=input_shape) here and will thus interpolate the output to match the input shape.