HoverNet uses resnet50 based imagenet pretrained weights for the encoder (see: https://github.com/vqdang/hover_net#training). I think we should also enable this feature.