geointel/backend/tests/test_failure_driven_yolo_sampling.py

from __future__ import annotations

import importlib.util
from pathlib import Path


SCRIPT = Path(__file__).parents[2] / "scripts" / "build_failure_driven_yolo_sampling.py"
SPEC = importlib.util.spec_from_file_location("failure_sampling", SCRIPT)
assert SPEC and SPEC.loader
MODULE = importlib.util.module_from_spec(SPEC)
SPEC.loader.exec_module(MODULE)


def test_dataset_validation_source_preserves_manifest_path(tmp_path: Path):
    source = tmp_path / "dataset.yaml"
    source.write_text(
        "path: /data/source\ntrain: /data/source/train.txt\n"
        "val: /data/source/internal-val.txt\nnames:\n  0: building\n",
        encoding="utf-8",
    )

    assert MODULE.dataset_validation_source(source) == "/data/source/internal-val.txt"


def test_sampling_repeats_only_failed_region_train_tiles() -> None:
    manifest = {
        "samples": [
            {"sample_slug": "train-fl", "split": "train", "region": "flanders"},
            {"sample_slug": "train-wa", "split": "train", "region": "wallonia"},
            {"sample_slug": "test-fl", "split": "test", "region": "flanders"},
        ]
    }
    summary = {
        "tiles": [
            {"sample_slug": "train-fl", "split": "train", "label_count": 2, "image_path": "/tmp/fl-pos.png"},
            {"sample_slug": "train-fl", "split": "train", "label_count": 0, "image_path": "/tmp/fl-neg.png"},
            {"sample_slug": "train-wa", "split": "train", "label_count": 1, "image_path": "/tmp/wa-pos.png"},
            {"sample_slug": "test-fl", "split": "val", "label_count": 1, "image_path": "/tmp/protected.png"},
        ]
    }
    assessment = {
        "status": "continue_training_loop",
        "gates": {
            "min_region_f1": 0.45,
            "min_region_precision": 0.5,
            "min_region_recall": 0.4,
            "max_pure_empty_false_positives": 0,
        },
        "test": {
            "regions": {
                "flanders": {"f1": 0.2, "precision": 0.3, "recall": 0.2},
                "wallonia": {"f1": 0.6, "precision": 0.6, "recall": 0.6},
            }
        },
        "background": {"pure_empty_false_positives": 2},
    }
    paths, metadata = MODULE.build_sampling(
        summary=summary, manifest=manifest, assessment=assessment
    )
    assert paths.count(str(Path("/tmp/fl-pos.png").resolve())) == 3
    assert paths.count(str(Path("/tmp/fl-neg.png").resolve())) == 4
    assert paths.count(str(Path("/tmp/wa-pos.png").resolve())) == 1
    assert not any("protected" in path for path in paths)
    assert metadata["protected_samples_in_training"] == []
    assert metadata["weak_recall_regions"] == ["flanders"]


def test_sampling_can_use_calibration_before_test_is_opened() -> None:
    manifest = {"samples": [{"sample_slug": "train-fl", "split": "train", "region": "flanders"}]}
    summary = {
        "tiles": [
            {"sample_slug": "train-fl", "split": "train", "label_count": 1, "image_path": "/tmp/fl.png"}
        ]
    }
    assessment = {
        "status": "continue_training_loop",
        "gates": {
            "min_region_f1": 0.45,
            "min_region_precision": 0.5,
            "min_region_recall": 0.4,
            "max_pure_empty_false_positives": 0,
        },
        "calibration": {
            "regions": {"flanders": {"f1": 0.4, "precision": 0.6, "recall": 0.35}}
        },
        "test": None,
        "background": None,
    }
    paths, metadata = MODULE.build_sampling(
        summary=summary, manifest=manifest, assessment=assessment
    )
    assert len(paths) == 3
    assert metadata["failure_evidence_source"] == "calibration"


def test_precision_correction_can_balance_positive_and_negative_tiles() -> None:
    manifest = {"samples": [{"sample_slug": "train-fl", "split": "train", "region": "flanders"}]}
    summary = {
        "tiles": [
            {"sample_slug": "train-fl", "split": "train", "label_count": 2, "image_path": "/tmp/fl-pos.png"},
            {"sample_slug": "train-fl", "split": "train", "label_count": 0, "image_path": "/tmp/fl-neg.png"},
        ]
    }
    assessment = {
        "status": "continue_training_loop",
        "gates": {
            "min_region_f1": 0.45,
            "min_region_precision": 0.5,
            "min_region_recall": 0.4,
            "max_pure_empty_false_positives": 0,
        },
        "calibration": {
            "regions": {"flanders": {"f1": 0.46, "precision": 0.45, "recall": 0.46}}
        },
    }

    paths, metadata = MODULE.build_sampling(
        summary=summary,
        manifest=manifest,
        assessment=assessment,
        precision_positive_repeat=2,
        negative_repeat=3,
    )

    assert paths.count(str(Path("/tmp/fl-pos.png").resolve())) == 2
    assert paths.count(str(Path("/tmp/fl-neg.png").resolve())) == 3
    assert metadata["precision_positive_repeat"] == 2


def test_sampling_targets_failed_calibration_contexts_without_using_protected_tiles() -> None:
    manifest = {
        "samples": [
            {"sample_slug": "train-industry", "split": "train", "region": "flanders", "context": "industrial"},
            {"sample_slug": "train-suburban", "split": "train", "region": "flanders", "context": "suburban"},
            {"sample_slug": "cal-industry", "split": "calibration", "region": "flanders", "context": "industrial"},
        ]
    }
    summary = {
        "tiles": [
            {"sample_slug": "train-industry", "split": "train", "label_count": 2, "image_path": "/tmp/industry-pos.png"},
            {"sample_slug": "train-industry", "split": "train", "label_count": 0, "image_path": "/tmp/industry-neg.png"},
            {"sample_slug": "train-suburban", "split": "train", "label_count": 2, "image_path": "/tmp/suburban-pos.png"},
            {"sample_slug": "cal-industry", "split": "val", "label_count": 2, "image_path": "/tmp/protected.png"},
        ]
    }
    assessment = {
        "status": "continue_training_loop",
        "gates": {
            "min_region_f1": 0.45,
            "min_region_precision": 0.5,
            "min_region_recall": 0.4,
            "max_pure_empty_false_positives": 0,
        },
        "calibration": {
            "regions": {"flanders": {"f1": 0.3, "precision": 0.35, "recall": 0.27}},
            "samples": {"cal-industry": {"f1": 0.2, "precision": 0.3, "recall": 0.15}},
        },
    }

    paths, metadata = MODULE.build_sampling(summary=summary, manifest=manifest, assessment=assessment)

    assert paths.count(str(Path("/tmp/industry-pos.png").resolve())) == 5
    assert paths.count(str(Path("/tmp/industry-neg.png").resolve())) == 6
    assert paths.count(str(Path("/tmp/suburban-pos.png").resolve())) == 3
    assert not any("protected" in path for path in paths)
    assert metadata["weak_recall_contexts"] == ["flanders:industrial"]
    assert metadata["weak_precision_contexts"] == ["flanders:industrial"]