370 lines
16 KiB
Python
370 lines
16 KiB
Python
from __future__ import annotations
|
|
|
|
import importlib.util
|
|
from pathlib import Path
|
|
|
|
import pytest
|
|
|
|
|
|
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_rejects_protected_test_and_background_feedback() -> 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},
|
|
}
|
|
with pytest.raises(ValueError, match="protected test/background evidence"):
|
|
MODULE.build_sampling(
|
|
summary=summary, manifest=manifest, assessment=assessment, max_region_share=1.0
|
|
)
|
|
|
|
|
|
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, max_region_share=1.0
|
|
)
|
|
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,
|
|
max_region_share=1.0,
|
|
)
|
|
|
|
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, max_region_share=1.0
|
|
)
|
|
|
|
assert paths.count(str(Path("/tmp/industry-pos.png").resolve())) == 5
|
|
assert paths.count(str(Path("/tmp/industry-neg.png").resolve())) == 1
|
|
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"]
|
|
assert metadata["recall_dominant_regions"] == ["flanders"]
|
|
|
|
|
|
def test_sampling_rejects_background_feedback_after_a_protected_background_opening() -> None:
|
|
manifest = {"samples": [
|
|
{"sample_slug": "positive", "split": "train", "region": "flanders", "context": "industrial"},
|
|
{"sample_slug": "negative", "split": "train", "region": "flanders", "context": "industrial-hard-negative"},
|
|
]}
|
|
summary = {"tiles": [
|
|
{"sample_slug": "positive", "split": "train", "label_count": 1, "image_path": "/tmp/positive.png"},
|
|
{"sample_slug": "negative", "split": "train", "label_count": 0, "image_path": "/tmp/negative.png"},
|
|
]}
|
|
assessment = {
|
|
"status": "continue_training_loop",
|
|
"gates": {"min_region_f1": .45, "min_region_precision": .5, "min_region_recall": .4,
|
|
"max_pure_empty_false_positives": 0},
|
|
"calibration": {"regions": {
|
|
"flanders": {"f1": .25, "precision": .4, "recall": .2},
|
|
}},
|
|
"background": {"pure_empty_false_positives": 1},
|
|
}
|
|
|
|
with pytest.raises(ValueError, match="protected background evidence"):
|
|
MODULE.build_sampling(
|
|
summary=summary, manifest=manifest, assessment=assessment, max_region_share=1.0,
|
|
)
|
|
|
|
|
|
def test_region_cap_drops_only_repeats_and_preserves_every_unique_tile() -> None:
|
|
manifest = {"samples": [
|
|
{"sample_slug": "fl", "split": "train", "region": "flanders", "context": "industrial"},
|
|
{"sample_slug": "wa", "split": "train", "region": "wallonia", "context": "rural-town"},
|
|
{"sample_slug": "br", "split": "train", "region": "brussels", "context": "dense-urban"},
|
|
]}
|
|
summary = {"tiles": [
|
|
{"sample_slug": "fl", "split": "train", "label_count": 2, "image_path": f"/tmp/fl-{index}.png"}
|
|
for index in range(4)
|
|
] + [
|
|
{"sample_slug": "wa", "split": "train", "label_count": 2, "image_path": f"/tmp/wa-{index}.png"}
|
|
for index in range(2)
|
|
] + [
|
|
{"sample_slug": "br", "split": "train", "label_count": 2, "image_path": f"/tmp/br-{index}.png"}
|
|
for index in range(2)
|
|
]}
|
|
assessment = {
|
|
"status": "continue_training_loop",
|
|
"gates": {"min_region_f1": .45, "min_region_precision": .5, "min_region_recall": .4,
|
|
"max_pure_empty_false_positives": 0},
|
|
"calibration": {"regions": {
|
|
"flanders": {"f1": .2, "precision": .3, "recall": .2},
|
|
"wallonia": {"f1": .6, "precision": .6, "recall": .6},
|
|
"brussels": {"f1": .6, "precision": .6, "recall": .6},
|
|
}},
|
|
}
|
|
|
|
paths, metadata = MODULE.build_sampling(
|
|
summary=summary, manifest=manifest, assessment=assessment,
|
|
positive_repeat=5, max_region_share=.65,
|
|
)
|
|
|
|
assert all(str(Path(f"/tmp/fl-{index}.png").resolve()) in paths for index in range(4))
|
|
assert metadata["pre_cap_entries_by_region"]["flanders"] == 20
|
|
assert metadata["sampled_entries_by_region"]["flanders"] == 7
|
|
assert metadata["dropped_region_repeat_count"] == 13
|
|
assert metadata["sampled_entries_by_region"]["flanders"] / len(paths) <= .65
|
|
|
|
|
|
def test_region_cap_rotates_repeats_between_sampling_rounds() -> None:
|
|
manifest = {"samples": [
|
|
{"sample_slug": "fl", "split": "train", "region": "flanders", "context": "industrial"},
|
|
{"sample_slug": "wa", "split": "train", "region": "wallonia", "context": "rural-town"},
|
|
{"sample_slug": "br", "split": "train", "region": "brussels", "context": "dense-urban"},
|
|
]}
|
|
summary = {"tiles": [
|
|
{"sample_slug": "fl", "split": "train", "label_count": 2, "image_path": f"/tmp/fl-{index}.png"}
|
|
for index in range(4)
|
|
] + [
|
|
{"sample_slug": "wa", "split": "train", "label_count": 2, "image_path": f"/tmp/wa-{index}.png"}
|
|
for index in range(2)
|
|
] + [
|
|
{"sample_slug": "br", "split": "train", "label_count": 2, "image_path": f"/tmp/br-{index}.png"}
|
|
for index in range(2)
|
|
]}
|
|
assessment = {
|
|
"status": "continue_training_loop",
|
|
"gates": {"min_region_f1": .45, "min_region_precision": .5, "min_region_recall": .4,
|
|
"max_pure_empty_false_positives": 0},
|
|
"calibration": {"regions": {
|
|
"flanders": {"f1": .2, "precision": .3, "recall": .2},
|
|
"wallonia": {"f1": .6, "precision": .6, "recall": .6},
|
|
"brussels": {"f1": .6, "precision": .6, "recall": .6},
|
|
}},
|
|
}
|
|
|
|
first, first_metadata = MODULE.build_sampling(
|
|
summary=summary, manifest=manifest, assessment=assessment,
|
|
positive_repeat=5, max_region_share=.65, sampling_round=1,
|
|
)
|
|
second, second_metadata = MODULE.build_sampling(
|
|
summary=summary, manifest=manifest, assessment=assessment,
|
|
positive_repeat=5, max_region_share=.65, sampling_round=2,
|
|
)
|
|
|
|
assert first != second
|
|
assert set(first) == set(second)
|
|
assert first_metadata["sampling_round"] == 1
|
|
assert second_metadata["sampling_round"] == 2
|
|
|
|
|
|
def test_region_cap_preserves_failed_context_positive_before_hard_negative() -> None:
|
|
manifest = {"samples": [
|
|
{"sample_slug": "target", "split": "train", "region": "flanders", "context": "industrial"},
|
|
{"sample_slug": "negative", "split": "train", "region": "flanders", "context": "industrial-hard-negative"},
|
|
{"sample_slug": "wa", "split": "train", "region": "wallonia", "context": "rural-town"},
|
|
{"sample_slug": "br", "split": "train", "region": "brussels", "context": "dense-urban"},
|
|
]}
|
|
summary = {"tiles": [
|
|
{"sample_slug": "target", "split": "train", "label_count": 2, "image_path": "/tmp/target.png"},
|
|
{"sample_slug": "negative", "split": "train", "label_count": 0, "image_path": "/tmp/negative.png"},
|
|
{"sample_slug": "wa", "split": "train", "label_count": 1, "image_path": "/tmp/wa.png"},
|
|
{"sample_slug": "br", "split": "train", "label_count": 1, "image_path": "/tmp/br.png"},
|
|
]}
|
|
assessment = {
|
|
"status": "continue_training_loop",
|
|
"gates": {"min_region_f1": .45, "min_region_precision": .5, "min_region_recall": .4,
|
|
"max_pure_empty_false_positives": 0},
|
|
"calibration": {
|
|
"regions": {
|
|
"flanders": {"f1": .2, "precision": .2, "recall": .3},
|
|
"wallonia": {"f1": .6, "precision": .6, "recall": .6},
|
|
"brussels": {"f1": .6, "precision": .6, "recall": .6},
|
|
},
|
|
"samples": {"target": {"f1": .2, "precision": .2, "recall": .3}},
|
|
},
|
|
}
|
|
|
|
paths, metadata = MODULE.build_sampling(
|
|
summary=summary, manifest=manifest, assessment=assessment, max_region_share=.65,
|
|
)
|
|
|
|
assert paths.count(str(Path("/tmp/target.png").resolve())) == 2
|
|
assert paths.count(str(Path("/tmp/negative.png").resolve())) == 1
|
|
assert metadata["priority_positive_repeat_count"] == 4
|
|
|
|
|
|
def test_precision_guard_band_keeps_near_gate_region_stabilized() -> None:
|
|
manifest = {"samples": [
|
|
{"sample_slug": "wa-positive", "split": "train", "region": "wallonia", "context": "rural-town"},
|
|
{"sample_slug": "wa-negative", "split": "train", "region": "wallonia", "context": "farmland-hard-negative"},
|
|
]}
|
|
summary = {"tiles": [
|
|
{"sample_slug": "wa-positive", "split": "train", "label_count": 1, "image_path": "/tmp/wa-positive.png"},
|
|
{"sample_slug": "wa-negative", "split": "train", "label_count": 0, "image_path": "/tmp/wa-negative.png"},
|
|
]}
|
|
assessment = {
|
|
"status": "continue_training_loop",
|
|
"gates": {"min_region_f1": .45, "min_region_precision": .5, "min_region_recall": .4,
|
|
"max_pure_empty_false_positives": 0},
|
|
"calibration": {"regions": {
|
|
"wallonia": {"f1": .6, "precision": .52, "recall": .7},
|
|
}},
|
|
}
|
|
|
|
paths, metadata = MODULE.build_sampling(
|
|
summary=summary, manifest=manifest, assessment=assessment, max_region_share=1.0,
|
|
)
|
|
|
|
assert "wallonia" in metadata["weak_precision_regions"]
|
|
assert paths.count(str(Path("/tmp/wa-negative.png").resolve())) == 4
|
|
assert metadata["precision_guard_band"] == .03
|
|
|
|
|
|
def test_coastal_precision_failure_targets_port_and_dunes_negatives() -> None:
|
|
manifest = {"samples": [
|
|
{"sample_slug": "coastal-train", "split": "train", "region": "flanders", "context": "coastal-urban"},
|
|
{"sample_slug": "port-negative", "split": "train", "region": "flanders", "context": "port-hard-negative"},
|
|
{"sample_slug": "dunes-negative", "split": "train", "region": "flanders", "context": "dunes-negative"},
|
|
{"sample_slug": "coastal-cal", "split": "calibration", "region": "flanders", "context": "coastal-urban"},
|
|
]}
|
|
summary = {"tiles": [
|
|
{"sample_slug": "coastal-train", "split": "train", "label_count": 2, "image_path": "/tmp/coastal.png"},
|
|
{"sample_slug": "port-negative", "split": "train", "label_count": 0, "image_path": "/tmp/port.png"},
|
|
{"sample_slug": "dunes-negative", "split": "train", "label_count": 0, "image_path": "/tmp/dunes.png"},
|
|
]}
|
|
assessment = {
|
|
"status": "continue_training_loop",
|
|
"gates": {"min_region_f1": .45, "min_region_precision": .5, "min_region_recall": .4,
|
|
"max_pure_empty_false_positives": 0},
|
|
"calibration": {
|
|
"regions": {"flanders": {"f1": .3, "precision": .2, "recall": .4}},
|
|
"samples": {"coastal-cal": {"f1": .1, "precision": .05, "recall": .2}},
|
|
},
|
|
}
|
|
|
|
paths, metadata = MODULE.build_sampling(
|
|
summary=summary, manifest=manifest, assessment=assessment, max_region_share=1.0
|
|
)
|
|
|
|
assert paths.count(str(Path("/tmp/port.png").resolve())) == 6
|
|
assert paths.count(str(Path("/tmp/dunes.png").resolve())) == 6
|
|
assert "flanders:port-hard-negative" in metadata["targeted_negative_contexts"]
|
|
assert "flanders:dunes-negative" in metadata["targeted_negative_contexts"]
|