""" Comprehensive tests for data preprocessing pipeline. Tests preprocessing, transforms, dataset class, splits, and caching. """ import sys from pathlib import Path from typing import Tuple import numpy as np import pandas as pd import pytest import torch import yaml # Add src to path project_root = Path(__file__).parent.parent sys.path.insert(0, str(project_root)) from src.data import ( BrainTumorDataset, get_train_transforms, get_val_transforms, create_split, preprocess_image, CLASS_TO_IDX, IDX_TO_CLASS, ) from src.data.preprocess import find_brain_contour # ============================================================================ # Fixtures # ============================================================================ @pytest.fixture def tmp_dataset(tmp_path: Path) -> Tuple[Path, Path]: """ Create temporary dataset structure with dummy images. Returns: Tuple of (raw_dir, processed_dir) """ raw_dir = tmp_path / "raw" processed_dir = tmp_path / "processed" # Create directory structure for split in ["Training", "Testing"]: for class_name in ["glioma", "meningioma", "notumor", "pituitary"]: (raw_dir / split / class_name).mkdir(parents=True, exist_ok=True) (processed_dir / split / class_name).mkdir(parents=True, exist_ok=True) return raw_dir, processed_dir @pytest.fixture def sample_config() -> dict: """Load base.yaml config.""" config_path = project_root / "configs" / "base.yaml" if config_path.exists(): with open(config_path, 'r') as f: config = yaml.safe_load(f) return config else: # Return minimal config if file doesn't exist return { 'data': { 'img_size': 224, 'train_val_split': 0.85, }, 'augmentation': { 'rotation_degrees': 15, 'horizontal_flip_prob': 0.5, 'brightness': 0.2, 'contrast': 0.2, 'mean': [0.485, 0.456, 0.406], 'std': [0.229, 0.224, 0.225], } } @pytest.fixture def dummy_image_with_borders() -> np.ndarray: """Create a test image with black borders.""" # 300x300 image with 50px black border img = np.zeros((300, 300, 3), dtype=np.uint8) # White center region img[50:250, 50:250, :] = 255 return img @pytest.fixture def dummy_image_no_borders() -> np.ndarray: """Create a test image without black borders.""" # 224x224 image, all white img = np.ones((224, 224, 3), dtype=np.uint8) * 255 return img @pytest.fixture def dummy_manifest(tmp_path: Path) -> pd.DataFrame: """Create a dummy manifest for testing splits.""" records = [] # Create balanced dataset: 100 samples per class for class_name in ["glioma", "meningioma", "notumor", "pituitary"]: for i in range(100): records.append({ 'filepath': f"Training/{class_name}/sample_{i:03d}.npy", 'class': class_name, 'split': 'training' }) return pd.DataFrame(records) # ============================================================================ # Tests: Preprocessing # ============================================================================ def test_preprocess_crops_margins(tmp_path: Path, dummy_image_with_borders: np.ndarray): """Test that preprocessing crops black margins correctly.""" # Save dummy image img_path = tmp_path / "test_image.jpg" import cv2 cv2.imwrite(str(img_path), dummy_image_with_borders) # Preprocess result = preprocess_image(img_path, target_size=(224, 224)) # Assert output is smaller than input (margins cropped) assert result.shape == (224, 224, 3), f"Expected (224, 224, 3), got {result.shape}" # Assert output dtype is uint8 assert result.dtype == np.uint8, f"Expected uint8, got {result.dtype}" # The image should be resized to 224x224 assert result.shape[0] == 224 and result.shape[1] == 224 def test_preprocess_handles_already_cropped(tmp_path: Path, dummy_image_no_borders: np.ndarray): """Test that preprocessing handles images without borders.""" # Save dummy image img_path = tmp_path / "test_image.jpg" import cv2 cv2.imwrite(str(img_path), dummy_image_no_borders) # Preprocess - should not raise errors result = preprocess_image(img_path, target_size=(224, 224)) # Assert no errors and output is correct size assert result.shape == (224, 224, 3) assert result.dtype == np.uint8 def test_find_brain_contour(): """Test the contour finding function.""" # Create test image with clear foreground img = np.zeros((200, 200, 3), dtype=np.uint8) img[50:150, 50:150, :] = 255 # White square in center x, y, w, h = find_brain_contour(img, threshold=10) # Should find the white region assert x >= 45 and x <= 55, f"Expected x around 50, got {x}" assert y >= 45 and y <= 55, f"Expected y around 50, got {y}" assert w >= 95 and w <= 105, f"Expected w around 100, got {w}" assert h >= 95 and h <= 105, f"Expected h around 100, got {h}" def test_find_brain_contour_no_foreground(): """Test contour finding on all-black image.""" img = np.zeros((200, 200, 3), dtype=np.uint8) x, y, w, h = find_brain_contour(img, threshold=10) # Should return full image bounds when no contour found assert x == 0 and y == 0 assert w == 200 and h == 200 # ============================================================================ # Tests: Transforms # ============================================================================ def test_transforms_train_augmentation(dummy_image_no_borders: np.ndarray): """Test that training transforms include randomness.""" transform = get_train_transforms() # Apply transforms multiple times outputs = [] for _ in range(5): output = transform(dummy_image_no_borders) outputs.append(output) # Check output properties for output in outputs: assert output.shape == (3, 224, 224), f"Expected (3, 224, 224), got {output.shape}" assert output.dtype == torch.float32, f"Expected float32, got {output.dtype}" # Check that outputs differ due to randomness # Compare first two outputs - they should not be identical diff = torch.abs(outputs[0] - outputs[1]).sum().item() assert diff > 0, "Training transforms should produce different outputs (randomness)" def test_transforms_val_deterministic(dummy_image_no_borders: np.ndarray): """Test that validation transforms are deterministic.""" transform = get_val_transforms() # Apply to same image multiple times output1 = transform(dummy_image_no_borders) output2 = transform(dummy_image_no_borders) # Check shape and dtype assert output1.shape == (3, 224, 224) assert output1.dtype == torch.float32 # Outputs should be identical (deterministic) assert torch.allclose(output1, output2, atol=1e-6), \ "Validation transforms should be deterministic" def test_normalize_values(dummy_image_no_borders: np.ndarray): """Test that normalization produces expected value ranges.""" transform = get_val_transforms() # Apply normalization output = transform(dummy_image_no_borders) # Check that values are normalized (not in [0, 1] range anymore) # For white image (255), after dividing by 255 and normalizing with ImageNet stats, # we expect values around (1 - mean) / std # For ImageNet: mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225] # White pixel (1.0) normalized should be around (1 - 0.485) / 0.229 ≈ 2.25 for R channel # Just check that normalization was applied (values outside [0, 1]) assert output.max() > 1.5, "Expected normalized values > 1.5 for white image" assert output.min() > 1.5, "Expected normalized values > 1.5 for white image" # ============================================================================ # Tests: Dataset Class # ============================================================================ def test_dataset_class_mapping(tmp_path: Path): """Test BrainTumorDataset class mapping and basic functionality.""" # Create dummy processed data processed_dir = tmp_path / "processed" records = [] for i, class_name in enumerate(["glioma", "meningioma"]): class_dir = processed_dir / "Training" / class_name class_dir.mkdir(parents=True, exist_ok=True) # Create dummy .npy file dummy_img = np.ones((224, 224, 3), dtype=np.uint8) * (i * 100) img_path = class_dir / f"sample_{i}.npy" np.save(img_path, dummy_img) records.append({ 'filepath': f"Training/{class_name}/sample_{i}.npy", 'class': class_name, 'split': 'train' }) manifest_df = pd.DataFrame(records) # Create dataset dataset = BrainTumorDataset( data_dir=processed_dir, manifest_df=manifest_df, transform=None # No transform for easier testing ) # Test __len__ assert len(dataset) == 2, f"Expected 2 samples, got {len(dataset)}" # Test __getitem__ img, label = dataset[0] assert isinstance(img, torch.Tensor), "Image should be a tensor" assert isinstance(label, int), "Label should be an int" assert label in [0, 1], f"Label should be 0 or 1, got {label}" # Test class indices match expected mapping for i, (img, label) in enumerate(dataset): expected_class = ["glioma", "meningioma"][i] expected_label = CLASS_TO_IDX[expected_class] assert label == expected_label, \ f"Expected label {expected_label} for {expected_class}, got {label}" def test_dataset_with_transforms(tmp_path: Path): """Test that dataset correctly applies transforms.""" # Create dummy processed data processed_dir = tmp_path / "processed" class_dir = processed_dir / "Training" / "glioma" class_dir.mkdir(parents=True, exist_ok=True) # Create dummy .npy file (use white image to ensure normalized values > 1) dummy_img = np.ones((224, 224, 3), dtype=np.uint8) * 255 img_path = class_dir / "sample.npy" np.save(img_path, dummy_img) manifest_df = pd.DataFrame([{ 'filepath': 'Training/glioma/sample.npy', 'class': 'glioma', 'split': 'train' }]) # Create dataset with transforms transform = get_val_transforms() dataset = BrainTumorDataset( data_dir=processed_dir, manifest_df=manifest_df, transform=transform ) img, label = dataset[0] # Check that transform was applied assert img.shape == (3, 224, 224), "Transform should produce (3, 224, 224)" assert img.dtype == torch.float32, "Transform should produce float32" # Normalized white image should have values > 1 (after ImageNet normalization) assert img.max() > 1.0, "White image should be normalized to values > 1" def test_dataset_get_class_counts(tmp_path: Path): """Test class count calculation.""" # Create dummy manifest manifest_df = pd.DataFrame([ {'filepath': 'Training/glioma/s1.npy', 'class': 'glioma', 'split': 'train'}, {'filepath': 'Training/glioma/s2.npy', 'class': 'glioma', 'split': 'train'}, {'filepath': 'Training/meningioma/s1.npy', 'class': 'meningioma', 'split': 'train'}, ]) dataset = BrainTumorDataset( data_dir=tmp_path, manifest_df=manifest_df, transform=None ) counts = dataset.get_class_counts() assert counts['glioma'] == 2, f"Expected 2 glioma samples, got {counts['glioma']}" assert counts['meningioma'] == 1, f"Expected 1 meningioma sample, got {counts['meningioma']}" def test_dataset_get_class_weights(tmp_path: Path): """Test class weight calculation for imbalanced datasets.""" # Create imbalanced manifest manifest_df = pd.DataFrame([ {'filepath': 'Training/glioma/s1.npy', 'class': 'glioma', 'split': 'train'}, {'filepath': 'Training/glioma/s2.npy', 'class': 'glioma', 'split': 'train'}, {'filepath': 'Training/glioma/s3.npy', 'class': 'glioma', 'split': 'train'}, {'filepath': 'Training/meningioma/s1.npy', 'class': 'meningioma', 'split': 'train'}, ]) dataset = BrainTumorDataset( data_dir=tmp_path, manifest_df=manifest_df, transform=None ) weights = dataset.get_class_weights() # glioma (3 samples) should have lower weight than meningioma (1 sample) glioma_idx = CLASS_TO_IDX['glioma'] meningioma_idx = CLASS_TO_IDX['meningioma'] assert weights[meningioma_idx] > weights[glioma_idx], \ "Minority class should have higher weight" # ============================================================================ # Tests: Stratified Split # ============================================================================ def test_stratified_split(tmp_path: Path, dummy_manifest: pd.DataFrame): """Test stratified train/validation split.""" from sklearn.model_selection import train_test_split # Use sklearn directly for testing (same as create_split uses internally) train_df, val_df = train_test_split( dummy_manifest, train_size=0.85, stratify=dummy_manifest['class'], random_state=42 ) # Check train/val ratio is approximately 85/15 total = len(dummy_manifest) train_ratio = len(train_df) / total val_ratio = len(val_df) / total assert 0.84 <= train_ratio <= 0.86, \ f"Expected train ratio ~0.85, got {train_ratio:.3f}" assert 0.14 <= val_ratio <= 0.16, \ f"Expected val ratio ~0.15, got {val_ratio:.3f}" # Check stratification: class proportions should be preserved for class_name in ["glioma", "meningioma", "notumor", "pituitary"]: # Original proportion orig_prop = len(dummy_manifest[dummy_manifest['class'] == class_name]) / len(dummy_manifest) # Train proportion train_prop = len(train_df[train_df['class'] == class_name]) / len(train_df) # Val proportion val_prop = len(val_df[val_df['class'] == class_name]) / len(val_df) # Proportions should be close (within 1%) assert abs(train_prop - orig_prop) < 0.01, \ f"Train proportion for {class_name} differs: {train_prop:.3f} vs {orig_prop:.3f}" assert abs(val_prop - orig_prop) < 0.01, \ f"Val proportion for {class_name} differs: {val_prop:.3f} vs {orig_prop:.3f}" # Check no overlap between train and val train_files = set(train_df['filepath']) val_files = set(val_df['filepath']) overlap = train_files & val_files assert len(overlap) == 0, f"Found {len(overlap)} overlapping files between train and val" # ============================================================================ # Tests: Cached Data Format # ============================================================================ def test_cached_data_format(tmp_path: Path, dummy_image_no_borders: np.ndarray): """Test that cached .npy files have correct format.""" # Save as .npy (simulating preprocessing output) npy_path = tmp_path / "cached_sample.npy" np.save(npy_path, dummy_image_no_borders) # Load cached data loaded = np.load(npy_path) # Assert dtype is uint8 assert loaded.dtype == np.uint8, f"Expected uint8, got {loaded.dtype}" # Assert shape is (224, 224, 3) assert loaded.shape == (224, 224, 3), \ f"Expected shape (224, 224, 3), got {loaded.shape}" # Assert data integrity assert np.array_equal(loaded, dummy_image_no_borders), \ "Loaded data should match original" def test_cached_data_loading_in_dataset(tmp_path: Path): """Test that dataset correctly loads cached .npy files.""" # Create dummy cached data processed_dir = tmp_path / "processed" class_dir = processed_dir / "Training" / "glioma" class_dir.mkdir(parents=True, exist_ok=True) # Create and save dummy image dummy_img = np.random.randint(0, 256, (224, 224, 3), dtype=np.uint8) img_path = class_dir / "sample.npy" np.save(img_path, dummy_img) manifest_df = pd.DataFrame([{ 'filepath': 'Training/glioma/sample.npy', 'class': 'glioma', 'split': 'train' }]) # Load via dataset dataset = BrainTumorDataset( data_dir=processed_dir, manifest_df=manifest_df, transform=None ) img_tensor, label = dataset[0] # Convert back to numpy for comparison img_numpy = (img_tensor.permute(1, 2, 0) * 255).byte().numpy() # Should match original assert np.array_equal(img_numpy, dummy_img), \ "Dataset should load cached data correctly" # ============================================================================ # Tests: Error Handling # ============================================================================ def test_dataset_missing_file(tmp_path: Path): """Test that dataset raises error for missing files.""" manifest_df = pd.DataFrame([{ 'filepath': 'Training/glioma/nonexistent.npy', 'class': 'glioma', 'split': 'train' }]) dataset = BrainTumorDataset( data_dir=tmp_path, manifest_df=manifest_df, transform=None ) with pytest.raises(FileNotFoundError): _ = dataset[0] def test_dataset_invalid_manifest(tmp_path: Path): """Test that dataset raises error for invalid manifest.""" # Missing required columns manifest_df = pd.DataFrame([{ 'filepath': 'Training/glioma/sample.npy', # Missing 'class' and 'split' }]) with pytest.raises(ValueError, match="Manifest must contain columns"): BrainTumorDataset( data_dir=tmp_path, manifest_df=manifest_df, transform=None ) def test_dataset_unknown_class(tmp_path: Path): """Test that dataset raises error for unknown class names.""" manifest_df = pd.DataFrame([{ 'filepath': 'Training/unknown/sample.npy', 'class': 'unknown_class', 'split': 'train' }]) with pytest.raises(ValueError, match="Unknown classes in manifest"): BrainTumorDataset( data_dir=tmp_path, manifest_df=manifest_df, transform=None ) def test_preprocess_missing_file(tmp_path: Path): """Test that preprocess raises error for missing image file.""" nonexistent = tmp_path / "nonexistent.jpg" with pytest.raises(ValueError, match="Failed to load image"): preprocess_image(nonexistent) # ============================================================================ # Tests: Class Mapping Constants # ============================================================================ def test_class_mapping_consistency(): """Test that CLASS_TO_IDX and IDX_TO_CLASS are consistent.""" # Check all expected classes are present expected_classes = {'glioma', 'meningioma', 'notumor', 'pituitary'} assert set(CLASS_TO_IDX.keys()) == expected_classes, \ "CLASS_TO_IDX should contain all expected classes" # Check indices are in correct range assert set(CLASS_TO_IDX.values()) == {0, 1, 2, 3}, \ "CLASS_TO_IDX should map to indices 0-3" # Check IDX_TO_CLASS is inverse of CLASS_TO_IDX for class_name, idx in CLASS_TO_IDX.items(): assert IDX_TO_CLASS[idx] == class_name, \ f"IDX_TO_CLASS[{idx}] should be '{class_name}', got '{IDX_TO_CLASS[idx]}'" # ============================================================================ # Performance Tests # ============================================================================ def test_transforms_performance(dummy_image_no_borders: np.ndarray): """Test that transforms execute quickly (< 100ms for 10 images).""" import time transform = get_train_transforms() start = time.time() for _ in range(10): _ = transform(dummy_image_no_borders) elapsed = time.time() - start assert elapsed < 1.0, \ f"Transform performance too slow: {elapsed:.3f}s for 10 images" def test_dataset_loading_performance(tmp_path: Path): """Test that dataset can load samples quickly.""" import time # Create 10 dummy samples processed_dir = tmp_path / "processed" class_dir = processed_dir / "Training" / "glioma" class_dir.mkdir(parents=True, exist_ok=True) records = [] for i in range(10): dummy_img = np.random.randint(0, 256, (224, 224, 3), dtype=np.uint8) img_path = class_dir / f"sample_{i}.npy" np.save(img_path, dummy_img) records.append({ 'filepath': f'Training/glioma/sample_{i}.npy', 'class': 'glioma', 'split': 'train' }) manifest_df = pd.DataFrame(records) dataset = BrainTumorDataset( data_dir=processed_dir, manifest_df=manifest_df, transform=get_val_transforms() ) # Load all samples start = time.time() for i in range(len(dataset)): _ = dataset[i] elapsed = time.time() - start assert elapsed < 2.0, \ f"Dataset loading too slow: {elapsed:.3f}s for 10 samples" if __name__ == "__main__": # Run tests with pytest pytest.main([__file__, "-v"])