training

Burn-based training and evaluation for TinyDet and BigDet.

Contents

• models: TinyDet (single-logit) + BigDet (multibox) configs/constructors.
• dataset: DatasetConfig, RunSample loader; collate pads boxes to max_boxes, emits gt_boxes, gt_mask, and global features (mean/std RGB, aspect, box count). collate_from_burn_batch does the same for warehouse batches.
• util: TrainArgs (model/backend/max-boxes/loss weights/input source), run_train, eval helpers, checkpoint load helpers for TinyDet/BigDet, greedy IoU matcher, backend validation.
• bin/train: CLI for training with --model {tiny,big} (default tiny), --max-boxes, --lambda-box, --lambda-obj, --backend {ndarray,wgpu}, --input-source {warehouse,capture-logs}.
• bin/eval: CLI to load a checkpoint (TinyDet/BigDet) and compute precision/recall at an IoU threshold.

Models

• TinyDet: single-logit detector, best for single-box targets.
• BigDet: multibox detector; config includes max_boxes (default 64) and optional input_dim (defaults to box-only; training sets 4+8 for box+global features).
  • forward_multibox returns (boxes [B, max_boxes, 4], scores [B, max_boxes]), normalized/clamped to [0,1].
  • TinyDet remains backward compatible for existing single-box flows.

Loss/matching

• Collate pads/truncates GT to max_boxes and provides a mask.
• Greedy IoU matching (GT -> best pred) builds objectness + box targets; unassigned preds are negative.
• Loss: masked L1 box regression on matched preds + BCE objectness for all preds; weighted by --lambda-box/--lambda-obj (optional IoU loss hook).

Backends/features

• Backends: NdArray by default; WGPU with --features backend-wgpu.
• Input source: warehouse manifests by default; capture-log loading is a legacy dev path (--input-source capture-logs).
• CLI flags: --backend, --model, --max-boxes, --lambda-box, --lambda-obj, --input-source, --warehouse-manifest, dataset roots.

Tests

• Collate test (padding/mask/features).
• BigDet smoke train/test (one step, save/load).
• BigDet forward-shape test (boxes/scores in expected shapes and [0,1] range).

License

Apache-2.0 (see LICENSE in the repo root).