MegaDetector-Overhead smoke tests

Minimal end-to-end smoke tests that verify a fresh install can:

1. Construct and forward-pass all 6 OWL models (OWL-C, OWL-T, OWL-D-S/B/L/H) on a random (1, 3, 512, 512) input.
2. Build a tiny synthetic point-annotated dataset.
3. Run one training epoch of OWL-C (CPU, batch_size=1) on that dataset using tools/train.py.
4. Run evaluation on the resulting checkpoint using tools/test.py.

These are NOT a substitute for running on real data — they verify the plumbing only (configs parse, datasets load, models forward, gradients flow, checkpoints save/load, metrics compute).

Prerequisites

• uv sync has succeeded (see INSTALL.md).
• For OWL-D smoke tests only: DINOv3 backbone weights are present under weights/ at the repo root. See INSTALL.md for download instructions. Without weights, OWL-D tests are skipped.

Run

Activate the venv first (source .venv/bin/activate after uv sync), then run with plain python:

# 1. Forward-pass smoke test (all 6 OWL models)
python tests/smoke_forward.py

# 2. Synthetic dataset (writes /tmp/owl-smoketest/)
python tests/make_synthetic_dataset.py

# 3. OWL-C training smoke (one epoch, batch_size=1, CPU)
WANDB_MODE=disabled python tools/train.py train=owlc_smoketest

# 4. OWL-C evaluation smoke on the checkpoint produced above
CKPT=$(ls -t outputs/*/*/best_model.pth | head -1 | xargs realpath)
WANDB_MODE=disabled python tools/test.py test=owlc_smoketest \
    "++test.model.pth_file=$CKPT"

# 5. (Optional) OWL-D-S training+eval smoke (requires DINOv3 weights)
WANDB_MODE=disabled python tools/train.py train=owld_s_smoketest
CKPT=$(ls -t outputs/*/*/best_model.pth | head -1 | xargs realpath)
WANDB_MODE=disabled python tools/test.py test=owld_s_smoketest \
    "++test.model.pth_file=$CKPT"

Expected runtime on CPU: ~1 min for forward-pass + dataset, ~30 s for OWL-C train, ~5 s for OWL-C eval, ~25 s for OWL-D-S train (frozen backbone), ~5 s for OWL-D-S eval.

Metrics on synthetic data are meaningless (4 train + 2 val images, 1 epoch, batch_size=1, random init). What matters is that every step completes without error.

Expected output

The forward-pass test ends with:

=== Forward-pass smoke summary ===
  OWLC      PASS            out shape=(1, 1, 256, 256)
  OWLT      PASS            out shape=(1, 1, 256, 256)
  OWLD_S    PASS            out shapes=[(1, 1, 256, 256)]
  OWLD_B    PASS            out shapes=[(1, 1, 256, 256)]
  OWLD_L    PASS            out shapes=[(1, 1, 256, 256)]
  OWLD_H    PASS            out shapes=[(1, 1, 256, 256)]

6/6 models passed; exit=0

OWL-D-* will be skipped (CONSTRUCT_FAIL with FileNotFoundError on the weight path) if DINOv3 weights are not present under weights/.

The training smoke run completes with output like:

[TRAINING] - Epoch: [1] [4/4] eta: ... loss: ...
[VALIDATION] - Epoch: [1] ...
Best model saved - Epoch 1 - Validation value: ...
Training complete | Best f1_score: ... at epoch 1

The evaluation smoke run writes metrics_results.csv, confusion_matrix.csv, detections.csv, and plots/precision_recall_curve.png under outputs/<date>/<time>/.