[AI from Scratch] Episode 231: Practical Data Augmentation — Expanding Image Datasets

Recap and Today’s Theme

Hello! In the previous episode, we covered saving and loading models using Keras, learning how to save trained models for future use. This helps streamline both development and deployment processes.

Today, we will dive into data augmentation, a technique used to generate new data from limited datasets and apply it in model training. Data augmentation is especially effective with image datasets, enabling models to learn a broader range of patterns and improving accuracy. In this article, we will explain how to augment image data using Keras and discuss its impact.

What Is Data Augmentation?

Data Augmentation is a technique that increases the size of a dataset by applying various transformations to the existing data. Deep learning models require large amounts of data, but it’s often challenging to gather sufficient data in all scenarios. Therefore, data augmentation becomes crucial.

Common Data Augmentation Techniques

1. Rotation: Rotating images by a set angle.
2. Horizontal/Vertical Flip: Flipping images horizontally or vertically.
3. Zoom: Enlarging or shrinking the image.
4. Shift: Shifting the image horizontally or vertically.
5. Noise Addition: Adding random noise to the image.
6. Color Transformation: Adjusting brightness, contrast, or hue.

Combining these transformations generates new data, helping the model learn diverse patterns and preventing overfitting.

Implementing Data Augmentation with Keras

Keras offers the ImageDataGenerator class, which simplifies image data augmentation. This class allows real-time augmentation during training, improving efficiency. Let’s implement data augmentation using ImageDataGenerator.

1. Importing Necessary Libraries

First, import TensorFlow and Keras libraries.

import tensorflow as tf
from tensorflow.keras.preprocessing.image import ImageDataGenerator
import matplotlib.pyplot as plt

2. Preparing the Dataset

We will use the CIFAR-10 dataset, which contains small color images from 10 different classes, to demonstrate data augmentation.

# Loading the CIFAR-10 dataset
(x_train, y_train), (x_test, y_test) = tf.keras.datasets.cifar10.load_data()

# Scaling image data to 0-1
x_train = x_train.astype('float32') / 255.0
x_test = x_test.astype('float32') / 255.0

3. Configuring Data Augmentation

Set up data augmentation using ImageDataGenerator:

# Configuring data augmentation
datagen = ImageDataGenerator(
    rotation_range=20,        # Random rotation up to 20 degrees
    width_shift_range=0.1,    # Shift up to 10% horizontally
    height_shift_range=0.1,   # Shift up to 10% vertically
    zoom_range=0.2,           # Random zoom up to 20%
    horizontal_flip=True,     # Random horizontal flip
    brightness_range=[0.8, 1.2]  # Adjust brightness
)

• rotation_range: Specifies the range for random rotations (up to 20 degrees here).
• width_shift_range and height_shift_range: Defines the maximum shift (10% of image width/height).
• zoom_range: Specifies the zoom range for scaling the image.
• horizontal_flip: Randomly flips the image horizontally.
• brightness_range: Adjusts image brightness within the specified range.

4. Visualizing Augmented Data

Visualize how the data is being augmented:

# Displaying augmented data
sample_image = x_train[0]  # Sample image
sample_image = sample_image.reshape((1, *sample_image.shape))

# Generating augmented images
gen = datagen.flow(sample_image, batch_size=1)

# Displaying generated images
plt.figure(figsize=(12, 6))
for i in range(5):
    plt.subplot(1, 5, i+1)
    plt.imshow(gen.next()[0])
    plt.axis('off')
plt.show()

• flow(): Generates batches of augmented data.
• plt.imshow(): Displays the augmented images.

This visualization shows variations based on the configured augmentation settings.

5. Training a Model with Augmented Data

Apply data augmentation during model training using a simple CNN model:

# Defining a simple CNN model
model = tf.keras.Sequential([
    tf.keras.layers.Conv2D(32, (3, 3), activation='relu', input_shape=(32, 32, 3)),
    tf.keras.layers.MaxPooling2D((2, 2)),
    tf.keras.layers.Conv2D(64, (3, 3), activation='relu'),
    tf.keras.layers.MaxPooling2D((2, 2)),
    tf.keras.layers.Flatten(),
    tf.keras.layers.Dense(64, activation='relu'),
    tf.keras.layers.Dense(10, activation='softmax')
])

# Compiling the model
model.compile(optimizer='adam',
              loss='sparse_categorical_crossentropy',
              metrics=['accuracy'])

# Training the model with augmented data
history = model.fit(datagen.flow(x_train, y_train, batch_size=64),
                    epochs=20,
                    validation_data=(x_test, y_test))

• datagen.flow(): Supplies batches of augmented data to the model.
• epochs=20: Trains the model over 20 epochs.

Benefits of Data Augmentation

1. Preventing Overfitting: Data augmentation exposes the model to various data variations, reducing the risk of overfitting to specific data.
2. Data Diversity: Real-world data contains noise and different conditions; data augmentation prepares the model for such variability.
3. Dataset Expansion: Without needing additional resources, data augmentation expands the dataset by generating new samples from existing data.

Summary

In this episode, we explored data augmentation using Keras, demonstrating how to increase image data and its impact on training models. Data augmentation is particularly effective for image recognition tasks, significantly enhancing model accuracy. By combining various transformations, you can train more robust models. In the next episode, we will explore transfer learning using pre-trained models. Stay tuned!

---

Notes

• ImageDataGenerator: A Keras class that performs real-time data augmentation and feeds it into the model.
• Data Augmentation: A technique to improve model generalization by preventing overfitting to specific training data patterns.