Pretrained weights used in the paper.

Naming convention for trained models:
numkey<number of keypoints>_<affine or tps>_<loss_fn>.<epochs>.h5
dice is supervised, mse is unsupervised, msekpconsistency is unsupervised with keypoint consistency loss.

We also provide the self-supervised pretrained models.
Naming convention for pretrained models:
numkey<number of keypoints>_pretrain.<epochs>.h5