# Dataset Augmentaion

This article will cover how you can increase the size of your original dataset with the help of data augmentation.
Data augmentaion is a practice of altering samples in your dataset, making them distinct enough from the original sample to be considered a new sample, and keeping alterations small enough to keep them recognizable as a part of the dataset's original data domain.
<br><b>Examples:</b> Adding slight noise to audio samples and mirroring images.

# Image augmentaion
The simplest way to add data augmentaion to your training pipeline is to use <a href="https://albumentations.ai/docs/">Albumentations</a> library.

Starting from the most basic ones, here are some augmentaion tricks you can use:
<ul>
  <li>Original image:
  <br><img width=480 src="https://github.com/locsor/imagesForWiki/blob/main/augmentations/testImg.jpg?raw=true">
  
  ```python 
  image = Image.open("testImg.jpg")
  image_np = np.array(image)
  ```
  <li>Image flipping or mirroring:
  <br><img width=480 src="https://github.com/locsor/imagesForWiki/blob/main/augmentations/aug_1.png?raw=true">
  
  ```python 
  transform = A.Compose([A.HorizontalFlip(p=1.0)])
  transformed_image = transform(image=image_np)["image"]
  ```
  <br><img width=480 src="https://github.com/locsor/imagesForWiki/blob/main/augmentations/aug_2.png?raw=true">
  ```python 
  transform = A.Compose([A.VerticalFlip(p=1.0)])
  transformed_image = transform(image=image_np)["image"]
  ```
  <li>Image rotation:
  <br><img width=480 src="https://github.com/locsor/imagesForWiki/blob/main/augmentations/aug_3.png?raw=true">
  
  ```python 
  transform = A.Compose([A.Rotate(p=1.0, limit=45, border_mode=0)])
  transformed_image = transform(image=image_np)["image"]
  ```
  <li>HSV Jitter:
  <br><img width=480 src="https://github.com/locsor/imagesForWiki/blob/main/augmentations/aug_5.png?raw=true">
  
  ```python 
  transform = A.Compose([A.ColorJitter(p=1.0)])
  transformed_image = transform(image=image_np)["image"]
  ```
  <li>Gaussian Noise:
  <br><img width=480 src="https://github.com/locsor/imagesForWiki/blob/main/augmentations/aug_6.png?raw=true">

  ```python 
  transform = A.Compose([A.GaussNoise(p=1.0, var_limit=(1000.0, 5000.0))])
  transformed_image = transform(image=image_np)["image"]
  ```
</ul>
Augmentaions almost always combined with each other:
<br><img width=480 src="https://github.com/locsor/imagesForWiki/blob/main/augmentations/aug_7.png?raw=true">

```python 
  transform = A.Compose([
      A.HorizontalFlip(p=1.0),
      A.VerticalFlip(p=1.0),
      A.Rotate(p=1.0, limit=45, border_mode=0),
      A.RandomBrightnessContrast(p=1.0, brightness_limit=(0.15,0.25)),
      A.ColorJitter(p=1.0),
      A.GaussNoise(p=1.0, var_limit=(1000.0, 2000.0),),
  ])
  transformed_image = transform(image=image_np)["image"]
```

Above is an extremecase of image augmentaion, we still want to keep the resulting images as close to the original data distribution as possible:
<br><img width=480 src="https://github.com/locsor/imagesForWiki/blob/main/augmentations/aug_final.png?raw=true">

```python 
  transform = A.Compose([
      A.HorizontalFlip(p=0.5),
      A.Rotate(p=0.5, limit=15, border_mode=0),
      A.RandomBrightnessContrast(p=0.5, brightness_limit=(-0.1,0.1)),
      A.ColorJitter(p=0.5),
      A.GaussNoise(p=0.5, var_limit=(50.0, 250.0),),
  ])
  transformed_image = transform(image=image_np)["image"]
```