Finetuning FCN or deeplab with more classes

I want to do semantic segmentation with 23 classes. But in the FCN or Deeplab is trained with 20 classes.

I implemented via VGG16 then upsample the output but I wanted to directly use the pre-trained semantic segmentation models like FCN or DeepLab. I could not find any reference in the PyTorch documentation. Could you please help me?