Chicken V-JEPA2 Action Classifier

Fine-tuned chicken behavior classifier used by Khoroos.

• Architecture: VJEPA2ForVideoClassification.
• Input: 64-frame RGB clips of individual birds.
• Processor: resize shortest edge to 292 pixels, center crop to 256 x 256, normalize using the values saved in video_preprocessor_config.json.
• Checkpoint dtype: float32.

Classes in output-column order

0. body_shaking
1. drinking
2. dust_bathing
3. feeding
4. head_scratching
5. litter_pecking
6. litter_scratching
7. pooping
8. preening
9. resting
10. running
11. standing
12. stretching
13. walking
14. wing_flapping

Load

Authenticate with a Hugging Face account authorized to access this private repository.

from transformers import AutoVideoProcessor, AutoModelForVideoClassification

repo_id = "amirivojdan/chicken_vjepa2_action"
processor = AutoVideoProcessor.from_pretrained(repo_id, token=True)
model = AutoModelForVideoClassification.from_pretrained(repo_id, token=True).eval()

Match the training sampler: uniformly sample 64 frames across longer clips; for shorter clips, repeat the final frame until 64 frames are available. Khoroos's ActionRecognizer handles this sampling and returns softmax probabilities in the class order above.

Predictions are model-assigned behavior labels for descriptive statistics; no welfare interpretation is provided. Evaluation metrics are not included in this export.

License

This fine-tuned model is licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license (CC BY-NC-SA 4.0). See LICENSE for the full terms.

Attribution: Ahmad Amirivojdan, amirivojdan/chicken_vjepa2_action. Indicate any modifications when sharing the model. Third-party components retain their applicable licenses and notices.