Add R(2+1)D, S3D, and MViT V2 as feature extractors

[Copilot]

The codebase only supported older video feature extractors (C3D, MFNET, 3D-ResNet). This adds three modern, torchvision-backed alternatives with Kinetics-400 pretrained weights — no separate weight downloads required.

New extractors

Key	Model	Output dim	Input size
r2plus1d	R(2+1)D-18 — decomposed spatiotemporal convolutions	512	112×112
s3d	S3D — separable 3D Inception-style	1024	224×224
mvit	MViT V2 Small — multiscale video transformer	768	224×224

Changes

• network/r2plus1d.py, network/s3d.py, network/mvit.py — thin wrapper classes that strip the classification head and expose the backbone as a feature extractor
• utils/load_model.py — adds loading logic for the three new models; feature_extractor_path is now str | None — omitting it triggers torchvision's auto-download of pretrained weights (only c3d, mfnet, r3d101, r3d152 still require an explicit path)
• utils/utils.py — adds correct normalization and resize/crop transforms per model
• utils/types.py — extends FeatureExtractor union
• feature_extractor.py, AD_live_prediction.py — new choices exposed in CLI args

Usage

# Uses Kinetics-400 pretrained weights automatically — no path needed
python feature_extractor.py --model_type mvit --dataset_path /data/videos --save_dir features/

# Or supply custom weights
python feature_extractor.py --model_type r2plus1d --pretrained_3d /weights/my_r2plus1d.pth ...

---

🔒 GitHub Advanced Security automatically protects Copilot coding agent pull requests. You can protect all pull requests by enabling Advanced Security for your repositories. Learn more about Advanced Security.