File size: 9,800 Bytes
3070d13 12a6671 3070d13 3137fe7 3070d13 7a44ad9 3070d13 3137fe7 12a6671 3070d13 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 | from __future__ import annotations
import argparse
import tempfile
from collections import Counter
from pathlib import Path
import unittest
from unittest.mock import patch
MISSING_DEPENDENCY: str | None = None
try:
import numpy as np
import pandas as pd
import torch
from PIL import Image
from milk10k_effb2_metadata.data import (
HybridEpochSampler,
PairedMilk10kMetadataDataset,
fit_metadata_spec,
hybrid_target_counts,
)
from milk10k_effb2_metadata.training import validate_balance_args
from milk10k_effb2_metadata.cli import parse_args
from milk10k_effb2_metadata.runner import append_augmented_train_rows
except ModuleNotFoundError as exc: # pragma: no cover - local minimal env may omit ML deps.
MISSING_DEPENDENCY = exc.name
def balance_args(**overrides):
values = {
"balance_mode": "hybrid",
"weighted_sampler": False,
"balance_head_ratio": 2.0,
"balance_tail_floor": 100,
"balance_min_source_count": 20,
}
values.update(overrides)
return argparse.Namespace(**values)
@unittest.skipIf(MISSING_DEPENDENCY is not None, f"Missing ML test dependency: {MISSING_DEPENDENCY}")
class HybridBalanceTest(unittest.TestCase):
def setUp(self) -> None:
# BCC=250, NV=100, INF=40, MAL_OTH=9.
self.labels = [0] * 250 + [1] * 100 + [2] * 40 + [3] * 9
self.targets, self.strong_labels = hybrid_target_counts(self.labels, balance_args())
def test_targets_cap_head_oversample_tail_and_leave_ultra_rare_alone(self) -> None:
self.assertEqual(self.targets.tolist(), [200, 100, 100, 9])
self.assertEqual(self.strong_labels, {2})
def test_sampler_has_expected_counts_and_only_tail_duplicates(self) -> None:
sampler = HybridEpochSampler(self.labels, self.targets, seed=42)
indices = list(sampler)
sampled_labels = Counter(self.labels[index] for index in indices)
self.assertEqual(sampled_labels, Counter({0: 200, 1: 100, 2: 100, 3: 9}))
bcc_indices = [index for index in indices if self.labels[index] == 0]
inf_indices = [index for index in indices if self.labels[index] == 2]
self.assertEqual(len(bcc_indices), len(set(bcc_indices)))
self.assertLess(len(set(inf_indices)), len(inf_indices))
def test_sampler_is_reproducible_and_changes_head_subset_by_epoch(self) -> None:
first = HybridEpochSampler(self.labels, self.targets, seed=7)
second = HybridEpochSampler(self.labels, self.targets, seed=7)
first.set_epoch(3)
second.set_epoch(3)
self.assertEqual(list(first), list(second))
epoch_three = set(index for index in first if self.labels[index] == 0)
first.set_epoch(4)
epoch_four = set(index for index in first if self.labels[index] == 0)
self.assertNotEqual(epoch_three, epoch_four)
def test_dataset_routes_only_tail_to_strong_transform(self) -> None:
with tempfile.TemporaryDirectory() as tmp:
image_path = Path(tmp) / "image.png"
Image.new("RGB", (4, 4), color=(10, 20, 30)).save(image_path)
rows = []
for label in ("BCC", "INF"):
rows.append(
{
"lesion_id": label,
"label": label,
"clinical_path": str(image_path),
"dermoscopic_path": str(image_path),
"clinical_age_approx": 50,
"dermoscopic_age_approx": 50,
"clinical_skin_tone_class": 2,
"dermoscopic_skin_tone_class": 2,
"clinical_sex": "unknown",
"dermoscopic_sex": "unknown",
"clinical_site": "unknown",
"dermoscopic_site": "unknown",
}
)
spec = {"sex_values": ["unknown"], "site_values": ["unknown"], "monet_columns": []}
regular = lambda image: torch.zeros(3, image.height, image.width)
strong = lambda image: torch.ones(3, image.height, image.width)
dataset = PairedMilk10kMetadataDataset(
pd.DataFrame(rows),
{"BCC": 0, "INF": 1},
spec,
regular,
strong_transform=strong,
strong_augment_labels={1},
)
self.assertTrue(torch.equal(dataset[0]["clinical"], torch.zeros(3, 4, 4)))
self.assertTrue(torch.equal(dataset[1]["clinical"], torch.ones(3, 4, 4)))
self.assertTrue(torch.equal(dataset[1]["dermoscopic"], torch.ones(3, 4, 4)))
def test_balance_argument_validation(self) -> None:
validate_balance_args(balance_args())
with self.assertRaisesRegex(ValueError, "weighted-sampler"):
validate_balance_args(balance_args(weighted_sampler=True))
with self.assertRaisesRegex(ValueError, "head-ratio"):
validate_balance_args(balance_args(balance_head_ratio=0))
with self.assertRaisesRegex(ValueError, "tail-floor"):
validate_balance_args(balance_args(balance_tail_floor=-1))
with self.assertRaisesRegex(ValueError, "min-source-count"):
validate_balance_args(balance_args(balance_min_source_count=0))
validate_balance_args(balance_args(branch_dropout_prob=0.0))
validate_balance_args(balance_args(branch_dropout_prob=0.5))
with self.assertRaisesRegex(ValueError, "branch-dropout-prob"):
validate_balance_args(balance_args(branch_dropout_prob=-0.1))
with self.assertRaisesRegex(ValueError, "branch-dropout-prob"):
validate_balance_args(balance_args(branch_dropout_prob=1.0))
with self.assertRaisesRegex(ValueError, "Auxiliary CE"):
validate_balance_args(balance_args(aux_fusion_ce_weight=-0.1))
with self.assertRaisesRegex(ValueError, "Auxiliary CE"):
validate_balance_args(balance_args(aux_clinical_ce_weight=-0.1))
with self.assertRaisesRegex(ValueError, "Auxiliary CE"):
validate_balance_args(balance_args(aux_dermoscopic_ce_weight=-0.1))
with self.assertRaisesRegex(ValueError, "consistency-weight"):
validate_balance_args(balance_args(consistency_weight=-0.1))
with self.assertRaisesRegex(ValueError, "consistency-temperature"):
validate_balance_args(balance_args(consistency_temperature=0.0))
def test_train_all_rejects_validation_only_options(self) -> None:
validate_balance_args(balance_args(train_all=True, k_folds=1, calibrate_bias=False, fit_temperature=False, lws_epochs=0))
with self.assertRaisesRegex(ValueError, "k-folds"):
validate_balance_args(balance_args(train_all=True, k_folds=2, calibrate_bias=False, fit_temperature=False, lws_epochs=0))
with self.assertRaisesRegex(ValueError, "calibrate-bias"):
validate_balance_args(balance_args(train_all=True, k_folds=1, calibrate_bias=True, fit_temperature=False, lws_epochs=0))
with self.assertRaisesRegex(ValueError, "fit-temperature"):
validate_balance_args(balance_args(train_all=True, k_folds=1, calibrate_bias=False, fit_temperature=True, lws_epochs=0))
with self.assertRaisesRegex(ValueError, "lws-epochs"):
validate_balance_args(balance_args(train_all=True, k_folds=1, calibrate_bias=False, fit_temperature=False, lws_epochs=1))
def test_train_all_cli_uses_23_finetune_default_only_for_train_all(self) -> None:
with patch("sys.argv", ["train"]):
self.assertEqual(parse_args().finetune_epochs, 20)
with patch("sys.argv", ["train", "--train-all"]):
args = parse_args()
self.assertTrue(args.train_all)
self.assertEqual(args.freeze_epochs, 8)
self.assertEqual(args.finetune_epochs, 23)
with patch("sys.argv", ["train", "--train-all", "--finetune-epochs", "5"]):
self.assertEqual(parse_args().finetune_epochs, 5)
def test_disable_monet_metadata_excludes_concept_columns(self) -> None:
df = pd.DataFrame(
{
"clinical_sex": ["unknown"],
"dermoscopic_sex": ["unknown"],
"clinical_site": ["unknown"],
"dermoscopic_site": ["unknown"],
"clinical_MONET_hair": [0.1],
"dermoscopic_MONET_hair": [0.2],
}
)
self.assertEqual(fit_metadata_spec(df)["monet_columns"], ["MONET_hair"])
spec = fit_metadata_spec(df, disable_monet_metadata=True)
self.assertEqual(spec["monet_columns"], [])
self.assertTrue(spec["monet_disabled"])
def test_augmented_rows_are_filtered_by_original_train_source(self) -> None:
base = pd.DataFrame({"lesion_id": ["TRAIN", "VAL"], "label": ["A", "A"]})
train = base.iloc[[0]].copy()
val = base.iloc[[1]].copy()
augmented = pd.DataFrame(
{
"lesion_id": ["TRAIN__sdpair_000", "VAL__sdpair_000"],
"label": ["A", "A"],
"is_augmented": [True, True],
"ignore_metadata": [False, False],
}
)
args = argparse.Namespace(
augmented_data_dir=Path("augmented"),
augmented_max_per_class=0,
zero_augmented_metadata=False,
seed=42,
)
with patch("milk10k_effb2_metadata.runner.load_augmented_subset", return_value=augmented):
result = append_augmented_train_rows(base, train, val, ["A"], args)
self.assertEqual(result["lesion_id"].tolist(), ["TRAIN", "TRAIN__sdpair_000"])
if __name__ == "__main__":
unittest.main()
|