- Add write_training_metrics() in new compressors/training_metrics.py
for appending epoch/step/lr/component rows as JSON Lines
- Wire --metrics-path and --log-every CLI options into train.py, passing
them to the training loop so metrics rows are written every N steps
- Accept absolute metrics paths or paths relative to output directory
- Add quantization component to loss log alongside existing distill/contrastive
- Replace inline torch.device() with get_device() utility
- Add test_hash_training_metrics.py covering multi-row JSONL append
Infrastructure:
- Pin torch 2.7.1 + CUDA 12.8 index for Linux/Windows in pyproject.toml
- Add .justfile rsync upload recipe with .stignore exclusion
- Exclude **/__marimo__ from rsync in .stignore
Dependencies updated: numpy 2.4.5, pandas 3.0.3, black 26.5.0,
click 8.4.0, contourpy, etc.
- Add FramePacket dataclass to encapsulate per-image pipeline state
- Rename internal methods with underscore prefix convention
- Replace separate filter_batch/crop_batch with unified process_batch method
- Update notebook to use new HashPipeline API
- Add Owlv2ForObjectDetection and Owlv2Processor imports to model_loader
- Refactor load_dino_model to return tuple of processor and model
- Rewrite generate_proposals_batch to group images by bbox count for efficient batching
- Add _normalize_single_bbox_list helper for bbox normalization
- Update verification.py to use new pipeline architecture with detect/segment/filter/crop steps
- Add crop_batch method to HashPipeline for cropping images using OWLv2 detection boxes
- Integrate crop_batch into pipeline forward pass (extract_hash and extract_features)
- Move extract_masked_region from compressors/proposal/utils.py to utils/image.py
- Add crop_image_by_bbox utility function in utils/image.py
- Update type annotations to use dict[str, Any] instead of bare dict
- Update .justfile to add memory server command
- Update marimo dependency to >=0.22.0
- Update nvidia CUDA package markers for platform compatibility
- Replace bbox-prompted segmentation with OWLv2 text-guided object detection
- Refactor HashPipeline from nn.Module to plain class with modular stage methods
- Add detect_batch, segment_batch, filter_batch for explicit pipeline stages
- Rename forward to forward_batch with text_labels API instead of bboxes
- Add mask_scoring_config, score_threshold, postprocess_threshold configuration
- Update model_loader to expose Dinov2Model type annotation
- Refactor model_loader.py: improve return type annotations from tuple[Any, Any] to tuple[AutoImageProcessor, AutoModel]
- Refactor proposal/core.py: add input validation for mask array dimensionality, handle 2D masks and batch dimensions gracefully
- Refactor proposal_segament.ipynb: replace inline model loading with centralized load_owlv2_model() and load_sam_model() functions, use batched detect_objects_batch() and generate_proposals_batch() APIs
- switch OWLv2 loader return type to Owlv2ForObjectDetection
- add detect_objects and detect_objects_batch with two-stage score filtering
- add DetectionResult typed dict and conversion helper for post-processed outputs
- export new detection APIs from proposal module
- Replace SAM2AutomaticMaskGenerator pipeline with Sam2Processor+Sam2Model
- Freeze SAM model parameters at load time, removing torch.no_grad() at call sites
- Rewrite proposal/core.py to use bbox prompts instead of automatic point sampling
- Add bboxes parameter to all HashPipeline public methods (forward, forward_dataset, extract_features, extract_features_dataset)
- Extract mask filtering logic (_filter_masks) from proposal into pipeline
- Rename object_score/ to filter/
- Add load_owlv2_model to model_loader
- Rename notebooks/test.py to habitat_sim_setup.py
- Remove dino_compressor.py and segament_compressor.py
- Rewrite pipeline.py to inline DINO into HashPipeline
- Maintain backward compatibility: SAMHashPipeline alias
- Update tests and benchmark.py
