Skip to content

CHANGELOG.md

Latest commit

 

History

History
67 lines (62 loc) · 5.46 KB

CHANGELOG.md

File metadata and controls

67 lines (62 loc) · 5.46 KB

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[Unreleased]

Changed

  • jepa-web: Browser-session creation now validates config invariants, inference rejects shape mismatches explicitly, and the step API refuses calls past the configured total step count
  • README, CONTRIBUTING, and CLAUDE.md now describe the real 7-crate workspace, current browser-demo status, and the added operator runbooks
  • jepa-core: Representation::gather now uses indexed selection and preserves representation masks instead of dropping them silently
  • jepa-train: JepaComponents::forward_step now validates generated masks and passes real target indices to predictors
  • jepa-train: Generic JepaComponents::forward_step docs now explicitly call out its approximate masking semantics and point callers to strict vision helpers
  • jepa-vision: TransformerPredictor now conditions prediction tokens on target positions instead of ignoring them
  • jepa-compat: ONNX support now parses real ModelProto files and loads initializers into the checkpoint abstraction
  • scripts/run_parity_suite.sh now runs every checked-in strict image fixture by default and can target a single fixture or fixture directory explicitly
  • Documentation and agent context files now describe the project as alpha and explicitly call out the current trainer and ONNX limitations
  • Added in-repo planning docs for production gaps, milestone roadmap, and implementation work packages

Added

  • jepa-core: ObjectMasking strategy for masking whole object slots instead of spatial patches (C-JEPA support)
  • jepa-vision: SlotAttention<B> module with GRU-based iterative slot refinement and SlotEncoder<B> composing a frozen ViT backbone with slot attention
  • jepa-train: CausalJepaComponents training loop for C-JEPA — frozen encoder, object-level masking with identity anchoring, joint history + future MSE loss
  • jepa-world: ObjectDynamicsPredictor<B> transformer-based dynamics predictor for CEM planning in object-representation space
  • jepa-compat: CJepa variant in ModelFamily registry enum
  • jepa: --masking object CLI flag and --family cjepa model filter for C-JEPA workflows
  • jepa: CLI binary with 6 subcommands (models, inspect, checkpoint, train, encode, tui)
  • jepa: Interactive TUI dashboard with 6 tabs (Dashboard, Models, Training, Inference, Checkpoint, About) using ratatui and Catppuccin Mocha theme
  • jepa-vision: Strict masked image and video forward paths with no-leakage regression coverage
  • jepa-vision: Criterion coverage for strict IJepa::forward_step_strict in the maintained vision benchmark surface
  • jepa-world: try_new, try_push, try_total_cost, and try_plan runtime-validation helpers for caller-triggerable failure modes
  • jepa-compat: Parser-backed ONNX metadata inspection and initializer loading
  • Two additional canonical Python-exported strict I-JEPA image fixtures covering non-square grids and RGB asymmetric patches
  • External-facing operations, release-candidate, and performance-budget runbooks under docs/
  • Fuzz targets for masking, gather, energy, and checkpoint parsing
  • Coverage, fuzz, and benchmark-smoke CI jobs
  • ADR-0001 for strict masked encoder semantics
  • Quality-gate and release-process runbooks under docs/
  • Architecture, roadmap, and production-gap runbooks under docs/
  • jepa-core: Core traits (Encoder, Predictor, EnergyFn, MaskingStrategy, CollapseRegularizer) — RFC-001 through RFC-007
  • jepa-core: Energy functions: L2Energy, CosineEnergy, SmoothL1Energy — RFC-004
  • jepa-core: Masking strategies: BlockMasking, SpatiotemporalMasking, MultiBlockMasking — RFC-005
  • jepa-core: Collapse prevention: VICReg, BarlowTwins — RFC-006
  • jepa-core: EMA with CosineMomentumSchedule for target encoder updates — RFC-007
  • jepa-core: JepaConfig builder with validation and presets
  • jepa-vision: VitEncoder with TransformerBlock, MHSA, MLP — RFC-002
  • jepa-vision: PatchEmbedding for image patchification — RFC-002
  • jepa-vision: RotaryPositionEncoding2D for spatial awareness — RFC-002
  • jepa-vision: TransformerPredictor and IJepa model — RFC-003
  • jepa-vision: VitVideoEncoder, TubeletEmbedding, 3D RoPE, VJepa — RFC-003
  • jepa-world: Action, ActionConditionedPredictor trait — RFC-009
  • jepa-world: WorldModel, RandomShootingPlanner (CEM), L2Cost — RFC-009
  • jepa-world: HierarchicalJepa (H-JEPA) — RFC-010
  • jepa-world: ShortTermMemory ring buffer — RFC-010
  • jepa-train: JepaComponents forward step orchestration — RFC-008
  • jepa-train: WarmupCosineSchedule learning rate scheduler — RFC-008
  • jepa-train: CheckpointMeta serialization — RFC-008
  • jepa-train: TrainConfig, TrainMetrics, TrainStepOutput — RFC-008
  • jepa-compat: Safetensors checkpoint loading and conversion — RFC-011
  • jepa-compat: I-JEPA/V-JEPA key mapping patterns — RFC-011
  • CI workflow with check, test, clippy, fmt, and doc jobs
  • 365 unit/integration tests + doc tests across all crates
  • Property-based tests with proptest for numerical invariants
  • Criterion benchmarks for core and vision crates
  • 3 runnable examples (I-JEPA demo, training loop, world model planning)