Elastic-forcing-code / docs /readability-changes.md
Elastic-forcing's picture
Add training and inference source code
2ea2835 verified
|
Raw History Blame
2.17 kB

Source cleanup and validation

The public 1.3B and 14B releases now expose the training path used by the reported three-encoder experiments: DINOv3, V-JEPA 2 and VideoMAEv2 features are combined by the hybrid Nyström–Monte Carlo MMD objective, and selected feature gradients are replayed to the generator.

Earlier source snapshots also contained optional frame-level and plain Fréchet-distance objectives that were disabled in both released recipes. Their models, controllers, preprocessing tools, configuration fields and replay helpers have been removed so that the public tree follows the evaluated path.

Names that intentionally remain

The code/video_fd directory remains part of both releases. Despite its historical directory name, it contains frozen encoder adapters, temporal sampling and decoder/encoder VJP utilities used by the active three-encoder MMD pipeline.

Checkpoint methods and payload keys containing fd also remain where changing them would break loading existing three-encoder checkpoints. They describe the saved distribution-matching state; they do not enable a second loss branch.

Validation

Run the root checks from the repository root in an environment with the release dependencies installed. They are self-contained and do not require an old Git commit:

python3 scripts/check_readability.py
python3 scripts/check_loss_parity.py --release 1.3B
python3 scripts/check_loss_parity.py --release 14B
python3 1.3B/scripts/test_cpu.py
python3 14B/scripts/test_cpu.py

check_readability.py verifies the cleaned file layout, rejects stale auxiliary loss references and configuration keys, parses every Python source, checks shell syntax, and confirms the main scientific settings of both public recipes. check_loss_parity.py independently composes the three encoder branches and checks the active MMD values, metadata, detached normalization and feature gradients on CPU.

The first optimizer-step report in first-step-comparison.md records a historical comparison made before this cleanup. It remains useful evidence for the active training calculation, but it is not a source-identity claim for the cleaned tree.