- Evaluate multiple datasets in a single run (CIFAR10 + CIFAR100)
- Report Recall@K for a list of K values from one underlying search
- Save results to disk: summary.md, metrics.csv, per-dataset predictions.csv, confusion_matrix.csv
- Richer evaluation output: query_labels, topk_ids, topk_labels for downstream analysis
- Add --dataset and --top-k CLI overrides for benchmark command
- Update config schema: dataset→datasets, top_k→top_k_list
- 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.
- Extract model loading logic from benchmark CLI into task-owned prepare_benchmark
- Add RetrievalEncoder class wrapping DINO with optional hash compression
- Add accelerate dependency for device management
- Switch dataset from CIFAR-10 to CIFAR-100 with fine_label column
- 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
