ML Made Simple

What is Data Augmentation?

Data augmentation is a set of transformations techniques that create new variations of the original data, helping to improve the generalization and robustness of machine learning models.

What are the benefits of using data augmentation?

1. Increased Dataset Size: Data augmentation effectively increases the effective size of the training dataset, allowing the model to see more diverse examples during training.
2. Improved Generalization: By exposing the model to various data variations, data augmentation helps the model generalize better to unseen examples.
3. Robustness: Augmenting the data with different transformations makes the model more robust to variations in the input data.

What are commonly used data augmentation techniques in PyTorch?

1. RandomResizedCrop:
   • Example: transforms.RandomResizedCrop(224, scale=(0.08, 1.0), ratio=(0.75, 1.3333333333333333), interpolation=2)
   • Explanation:
     • size: 224 specifies the output size of the crop.
     • scale: (0.08, 1.0) means the cropped area’s size will range between 8% and 100% of the original image.
     • ratio: (0.75, 1.33) means the aspect ratio of the cropped area will range between 0.75 and 1.33.
     • interpolation: 2 specifies bilinear interpolation.
     • bilinear interpolation: Bilinear interpolation is like blending or averaging the colors of the four nearest pixels to estimate the color at an in-between point. If you have a point between the upper-left pixel (red), upper-right pixel (blue), lower-left pixel (green), and lower-right pixel (yellow), bilinear interpolation would combine these colors to estimate the color at the in-between point.
2. RandomHorizontalFlip:
   • Example: transforms.RandomHorizontalFlip(p=0.5)
   • Explanation:
     • p: 0.5 means there is a 50% chance of the image being horizontally flipped.
3. RandomVerticalFlip:
   • Example: transforms.RandomVerticalFlip(p=0.5)
   • Explanation:
     • p: 0.5 means there is a 50% chance of the image being vertically flipped.
4. RandomRotation:
   • Example: transforms.RandomRotation(degrees=30, resample=False, expand=False, center=None)
   • Explanation:
     • degrees: 30 means the image will be rotated by a random angle between -30 and 30 degrees.
     • Other parameters are optional.
5. ColorJitter:
   • Example: transforms.ColorJitter(brightness=0.2, contrast=0.2, saturation=0.2, hue=0.2)
   • Explanation:
     • brightness, contrast, saturation, hue: 0.2 means the corresponding factor will be randomly adjusted within the range [-0.2, 0.2].
     • brightness: Increasing brightness makes an image look lighter, while decreasing it makes the image look darker.
     • contrast: High contrast makes the bright areas brighter and the dark areas darker, while low contrast reduces the difference between bright and dark.
     • saturation: Highly saturated colors appear vivid and intense, while desaturated colors appear more muted or grayscale.
     • hue: Changing the hue involves shifting the entire color spectrum, so, for instance, red can become blue.
6. RandomGrayscale:
   • Example: transforms.RandomGrayscale(p=0.1)
   • Explanation:
     • p: 0.1 means there is a 10% chance of the image being converted to grayscale.
7. RandomAffine:
   • Example: transforms.RandomAffine(degrees=45, translate=(0.1, 0.1), scale=(0.8, 1.2), shear=30, resample=False, fillcolor=0)
   • Explanation:
     • degrees: 45 means the image will be rotated by a random angle between -45 and 45 degrees.
     • translate: (0.1, 0.1) means the image will be translated horizontally and vertically by a fraction of 0.1. Imagine you have a picture, and you want to slide it horizontally or vertically. Translating an image means moving it to a different position without changing its shape or size.
     • scale: (0.8, 1.2) means the image will be scaled by a factor between 0.8 and 1.2.
     • shear: 30 means the image will be sheared by a random angle between -30 and 30 degrees. Shearing is a transformation that shifts one part of an object in a direction perpendicular to the other part. Think about tilting a rectangle so that the top part moves horizontally to the right or left while the bottom part stays in place.
     • fillcolor: 0 for black and 255 for white.
8. RandomPerspective:
   • Example: transforms.RandomPerspective(distortion_scale=0.5, p=0.5, interpolation=3)
   • Explanation:
     • distortion_scale: 0.5 controls the amount of perspective distortion.
     • perspective distortion: Imagine you’re looking at an object, like a rectangular sign, from a certain angle or distance. Perspective distortion happens when the object appears skewed or distorted because of the way you’re viewing it.
     • p: 0.5 means there is a 50% chance of applying the transformation.
     • interpolation: 3 specifies trilinear interpolation.
     • trilinear interpolation: Trilinear interpolation is an extension of bilinear interpolation to the third dimension. Trilinear interpolation extends this idea to 3D space, blending colors across adjacent layers in addition to the horizontal and vertical directions.
9. RandomErasing:
   • Example: transforms.RandomErasing(p=0.5, scale=(0.02, 0.33), ratio=(0.3, 3.3), value=0, inplace=False)
   • Explanation:
     • p: 0.5 means there is a 50% chance of applying the transformation.
     • scale: (0.02, 0.33) means the proportion of the erased area will be between 2% and 33% of the image.
     • ratio: (0.3, 3.3) means the aspect ratio of the erased area will range between 0.3 and 3.3.
     • value: 0 specifies the value used for erasing. 0 for black and 255 for white.
     • inplace: False means the operation is not performed in-place.
10. Normalize:
    • Example: transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225], inplace=False)
    • Explanation:
      • mean: [0.485, 0.456, 0.406] are the mean values for each channel.
      • std: [0.229, 0.224, 0.225] are the standard deviations for each channel.
      • inplace: False means the operation is not performed in-place.
11. ToTensor:
    • Example: transforms.ToTensor()
    • Explanation:
      • No specific parameters. Converts the image to a PyTorch tensor.

When to use which augmentation technique?

1. Image Classification:
   • Transformation: RandomResizedCrop, RandomHorizontalFlip, ColorJitter, Normalize
   • Application: In image classification tasks, augmentations help the model generalize better by exposing it to variations in scale, orientation, and color. This is crucial for recognizing objects in different conditions.
2. Object Detection:
   • Transformation: RandomResizedCrop, RandomHorizontalFlip, RandomVerticalFlip, RandomRotation
   • Application: Augmentations are used to generate diverse training samples for object detection models. These transformations help the model handle objects at different orientations and positions in images.
3. Semantic Segmentation:
   • Transformation: RandomResizedCrop, RandomHorizontalFlip, RandomVerticalFlip
   • Application: Augmentations are applied to both the input image and its corresponding segmentation mask. This ensures that the model learns to segment objects correctly under various conditions.
4. Facial Recognition:
   • Transformation: RandomRotation, ColorJitter, RandomGrayscale
   • Application: For facial recognition tasks, augmentations are applied to simulate variations in lighting conditions, facial expressions, and head orientations, making the model more robust.
5. Document Analysis:
   • Transformation: RandomPerspective, RandomRotation, RandomAffine
   • Application: Augmentations are useful for handling perspective distortions, rotations, and shearing in scanned documents. This helps models perform well on documents captured from different angles.
6. Satellite Image Analysis:
   • Transformation: RandomRotation, RandomResizedCrop, ColorJitter
   • Application: Augmentations are applied to satellite images to handle variations in weather conditions, sun angles, and different scales. This improves the model’s ability to recognize features in satellite imagery.
7. Medical Image Analysis:
   • Transformation: RandomResizedCrop, RandomRotation, RandomAffine
   • Application: In medical imaging, augmentations help models generalize across different patients and imaging conditions, improving their robustness to variations in data.