Transformation (Albumentations)

Albumentations

I used it because it’s better than PyTorch’s built-in transformations in both speed and variety.

Transformation setup

Suppose you used the following transformation for training:

1
transformation = A.Compose(
2
    [
3
        A.Resize(224, 224),
4
        A.CenterCrop(100, 100),
5
        A.HorizontalFlip(p=0.5),
6
        A.OneOf(
7
            [
8
                A.MotionBlur(p=0.2),
9
                A.MedianBlur(blur_limit=3, p=0.2),
10
                A.Blur(blur_limit=3, p=0.2),
11
            ],
12
            p=1,
13
        ),
14
        A.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225], ),
15
        albumentations.pytorch.transforms.ToTensorV2(),
16
    ]
17
)

Then during inference, you need the same resizing, cropping, and normalization:

1
transformation = A.Compose(
2
    [
3
        A.Resize(224, 224),
4
        A.CenterCrop(100, 100),
5
        A.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225], ),
6
        albumentations.pytorch.transforms.ToTensorV2(),
7
    ]
8
)

When to apply transformations

I passed the transformation as a parameter when creating the dataset and applied it in __getitem__.

Transformations I used

Training

1
transformation = A.Compose(
2
    [
3
        A.Resize(224, 224),
4
        A.HorizontalFlip(p=0.5),
5
        A.OneOf([A.GaussNoise()], p=0.4),
6
        A.OneOf(
7
            [
8
                A.MotionBlur(p=0.2),
9
                A.MedianBlur(blur_limit=3, p=0.2),
10
                A.Blur(blur_limit=3, p=0.2),
11
            ],
12
            p=1,
13
        ),
14
        A.OneOf(
15
            [
16
                A.HueSaturationValue(p=0.5),
17
                A.RGBShift(p=0.5),
18
                A.ChannelShuffle(p=0.5),
19
            ],
20
            p=1,
21
        ),
22
        A.ShiftScaleRotate(
23
            shift_limit=0.2,
24
            scale_limit=0.2,
25
            rotate_limit=10,
26
            border_mode=0,
27
            p=0.4,
28
        ),
29
        A.CoarseDropout(p=0.5),
30
        A.ColorJitter(p=0.3),
31
        A.RandomBrightnessContrast(p=0.7),
32
        # A.Rotate(limit=(-10, 10), p=0.4),
33
        A.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225], ),
34
        albumentations.pytorch.transforms.ToTensorV2(),
35
    ]
36
)

TTA

1
   A.GaussNoise(var_limit=(20.0, 60.0),p=1),
2
    A.MedianBlur(blur_limit=9, p=1),
3
    A.Blur(blur_limit=9, p=1),
4
    A.HueSaturationValue(hue_shift_limit=40, sat_shift_limit=40, val_shift_limit=40,p=1),
5
    A.RGBShift(r_shift_limit=50, g_shift_limit=50, b_shift_limit=50,p=1),
6
    A.ChannelDropout(p=1),
7
    A.ChannelShuffle(p=1),
8
    A.CoarseDropout(p=1),
9
    A.ColorJitter(brightness=0.5, contrast=0.5, saturation=0.5, hue=0.5,p=1),
10
    A.RandomBrightnessContrast(brightness_limit=0.5, contrast_limit=0.5,p=1),
11
    A.ShiftScaleRotate(
12
            shift_limit=0.2,
13
            scale_limit=0.2,
14
            rotate_limit=10,
15
            border_mode=0,
16
            p=1,
17
    ),