Skip to content

transforms

anomaly_match.image_processing.transforms

Image transform pipelines for weak, strong, and prediction augmentation.

NumpyToTensor

Convert numpy array to torch tensor for N-channel images.

Handles HWC to CHW conversion for arbitrary channel counts.

__call__(img)

Convert numpy array (H, W, C) to tensor (C, H, W).

Returns:

Type Description
Tensor

torch.Tensor: Image tensor in CHW format.
Source code in anomaly_match/image_processing/transforms.py 
28
29
30
31
32
33
34
35
36
37
38
def __call__(self, img) -> torch.Tensor:
    """Convert numpy array (H, W, C) to tensor (C, H, W).

    Returns:
        torch.Tensor: Image tensor in CHW format.
    """
    if isinstance(img, np.ndarray):
        # HWC to CHW conversion
        img = np.transpose(img, (2, 0, 1))
        img = torch.from_numpy(img.copy()).float() / 255.0
    return img

NumpyRandomHorizontalFlip(p=0.5)

Random horizontal flip for numpy arrays.

Source code in anomaly_match/image_processing/transforms.py 
44
45
def __init__(self, p=0.5):
    self.p = p

__call__(img)

Apply random horizontal flip to the image.

Returns:

Type Description
ndarray | Tensor

Image with random horizontal flip applied.
Source code in anomaly_match/image_processing/transforms.py 
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
def __call__(self, img) -> np.ndarray | torch.Tensor:
    """Apply random horizontal flip to the image.

    Returns:
        Image with random horizontal flip applied.
    """
    if isinstance(img, np.ndarray):
        if np.random.random() < self.p:
            return np.ascontiguousarray(img[:, ::-1, :])
        return img
    elif isinstance(img, torch.Tensor):
        if np.random.random() < self.p:
            return torch.flip(img, dims=[2])  # Flip width dimension (CHW format)
        return img
    return img

NumpyRandomTranslate(translate=(0, 0.125))

Random translation for numpy arrays.

Source code in anomaly_match/image_processing/transforms.py 
67
68
def __init__(self, translate=(0, 0.125)):
    self.translate = translate

__call__(img)

Apply random translation to the image.

Returns:

Type Description
ndarray | Tensor

Image with random translation applied.
Source code in anomaly_match/image_processing/transforms.py 
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
def __call__(self, img) -> np.ndarray | torch.Tensor:
    """Apply random translation to the image.

    Returns:
        Image with random translation applied.
    """
    if isinstance(img, np.ndarray):
        h, w = img.shape[:2]
        max_dx = self.translate[1] * w
        max_dy = self.translate[0] * h
        dx = np.random.uniform(-max_dx, max_dx)
        dy = np.random.uniform(-max_dy, max_dy)

        # Use numpy roll for translation
        img = np.roll(img, int(dx), axis=1)
        img = np.roll(img, int(dy), axis=0)
        return img
    elif isinstance(img, torch.Tensor):
        # For tensor (CHW format)
        _, h, w = img.shape
        max_dx = int(self.translate[1] * w)
        max_dy = int(self.translate[0] * h)
        dx = np.random.randint(-max_dx, max_dx + 1) if max_dx > 0 else 0
        dy = np.random.randint(-max_dy, max_dy + 1) if max_dy > 0 else 0
        return torch.roll(torch.roll(img, dx, dims=2), dy, dims=1)
    return img

get_weak_transforms(num_channels=3)

Get weak augmentation transforms.

Parameters:

Name Type Description Default
num_channels int

Number of image channels (3 for RGB, other for multispectral)

 3

Returns:

Type Description
Compose

torchvision.transforms.Compose: transforms.
Source code in anomaly_match/image_processing/transforms.py 
 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
def get_weak_transforms(num_channels: int = 3) -> transforms.Compose:
    """Get weak augmentation transforms.

    Args:
        num_channels: Number of image channels (3 for RGB, other for multispectral)

    Returns:
        torchvision.transforms.Compose: transforms.
    """
    if num_channels == 3:
        # Use PIL-based transforms for RGB
        return transforms.Compose(
            [
                transforms.ToTensor(),
                transforms.RandomHorizontalFlip(),
                transforms.RandomAffine(0, translate=(0, 0.125)),
            ]
        )
    else:
        # Use numpy-based transforms for multispectral
        return transforms.Compose(
            [
                NumpyRandomHorizontalFlip(),
                NumpyRandomTranslate(translate=(0, 0.125)),
                NumpyToTensor(),
            ]
        )

get_prediction_transforms(num_channels=3)

Get the standard image transform.

Parameters:

Name Type Description Default
num_channels int

Number of image channels (3 for RGB, other for multispectral)

 3

Returns:

Type Description
Compose

torchvision.transforms.Compose: transforms.
Source code in anomaly_match/image_processing/transforms.py 
127
128
129
130
131
132
133
134
135
136
137
138
139
def get_prediction_transforms(num_channels: int = 3) -> transforms.Compose:
    """Get the standard image transform.

    Args:
        num_channels: Number of image channels (3 for RGB, other for multispectral)

    Returns:
        torchvision.transforms.Compose: transforms.
    """
    if num_channels == 3:
        return transforms.Compose([transforms.ToTensor()])
    else:
        return transforms.Compose([NumpyToTensor()])

get_strong_transforms(num_channels=3)

Get strong augmentations for FixMatch.

Includes RandAugment followed by the same transforms as weak (ToTensor, RandomHorizontalFlip, RandomAffine).

Parameters:

Name Type Description Default
num_channels int

Number of image channels (3 for RGB, other for multispectral)

 3

Returns:

Type Description
Compose

torchvision.transforms.Compose: Strong augmentation pipeline.
Source code in anomaly_match/image_processing/transforms.py 
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
def get_strong_transforms(num_channels: int = 3) -> transforms.Compose:
    """Get strong augmentations for FixMatch.

    Includes RandAugment followed by the same transforms as weak (ToTensor,
    RandomHorizontalFlip, RandomAffine).

    Args:
        num_channels: Number of image channels (3 for RGB, other for multispectral)

    Returns:
        torchvision.transforms.Compose: Strong augmentation pipeline.
    """
    if num_channels == 3:
        # Use PIL-based transforms for RGB
        return transforms.Compose(
            [
                RandAugment(3, 5),
                transforms.ToTensor(),
                transforms.RandomHorizontalFlip(),
                transforms.RandomAffine(0, translate=(0, 0.125)),
            ]
        )
    else:
        # Use numpy-based transforms for multispectral
        return transforms.Compose(
            [
                MultispectralRandAugment(3, 5, num_channels=num_channels),
                NumpyRandomHorizontalFlip(),
                NumpyRandomTranslate(translate=(0, 0.125)),
                NumpyToTensor(),
            ]
        )