GameStatusClassifier fails to load VideoMAE checkpoint due to classifier size mismatch

[homayoonh79]

Hello,

When I run the demo, I get the following error while initializing the GameStatusClassifier:

ERROR: Failed to load game state classification model:
size mismatch for bias: copying a param with shape torch.Size([3]) from checkpoint,
the shape in current model is torch.Size([2])

It looks like the checkpoint is trained for 3 classes (play, no-play, service), but the model is being initialized with 2 classes by default.

I suggest updating the _load_model method to explicitly set num_labels to match the checkpoint, for example:

from transformers import VideoMAEConfig, VideoMAEForVideoClassification

config = VideoMAEConfig.from_pretrained(model_path)
config.num_labels = 3 # Match checkpoint

model = VideoMAEForVideoClassification.from_pretrained(
model_path,
config=config,
dtype=torch.float16
)

This should resolve the size mismatch error.

Could you clarify if the repository intends the game state classifier to always be 3-class, or if there is a separate 2-class checkpoint available?

Thanks!

[masouduut94]

Thanks for the report.
It is intended for 3 classes and it can be found in the game state classifier configuration json. (play, no-play, serve)

I only can update my projects at weekends, so I have little time to update it. I hope I could keep updating this project much better than before.