The following pretrained models are available for download:
Dataset | Model details | Link |
---|---|---|
MovingClevr | M2F with SACNN | here |
MovingClevrTex | M2F with SACNN | here |
MOVi A | M2F with SACNN | here |
MOVi C | M2F with SACNN | here |
MOVi D | M2F with SACNN | here |
MOVi E | M2F with SACNN | here |
MOVi A | M2F with Swin+WL | here |
MOVi C | M2F with Swin+WL | here |
MOVi D | M2F with Swin+WL | here |
MOVi E | M2F with Swin+WL | here |
KITTI | M2F with R18 | here |
KITTI | M2F with R18+WL | here |
KITTI | M2F with Swin+WL | here |
(SACNN - 6-layer CNN used in Slot Attention, Swin - Swin Transformer (weights for kitti), WL - Weighted Loss, R18 - ResNet18)
To load them with an appropriate architecture, consider the following snippet:
from types import SimpleNamespace
from mask2former_trainer_video import setup, Trainer
def load_model_cfg(config_path, weights_path):
args = SimpleNamespace(config_file=config_path,
opts=[
# Any extra arguments, such as number of slots.
'MODEL.WEIGHTS', str(weights_path),
], eval_only=True)
cfg = setup(args)
model = Trainer.build_model(cfg)
model.eval()
return model