Day 13: NPZ Files

Goal

Understand why pose data may be saved as NPZ.

Learn

• NPZ is a compressed NumPy file format. It can store multiple arrays in one file, which makes it useful for training pipelines.
• A SignLLM NPZ may store pose features, frame masks, confidence scores, frame count, labels, signer ID references, or normalized coordinates.
• JSON is easier for humans to read. NPZ is often faster and smaller for training, but it should be paired with metadata that humans can inspect.

Example

• Example structure: keypoints shape [frames, points, channels], confidence shape [frames, points], mask shape [frames], label = THANK-YOU, frame_count = 64.
• A related metadata row might say source_video: signer03_thankyou_0042.mp4, pose_file: signer03_thankyou_0042.npz, split: train, qa_status: approved.

Practice

1. Draw a pretend NPZ container with four arrays: keypoints, confidence, mask, and labels.
2. Write what each array means in plain English.

Checkpoint

Pipeline note