Skip to main content

Data Augmentation

Summary​

  • what is data augmentation
  • how to do data augmentation
  • applying data augmentation
    • applying to pre-processing layers to the data set
    • applying to the model layers itself

Content​

What is Data Augmentation​

Data augmentation is a technique in machine learning used to reduce overfitting when training a machine learning model, by training models on several slightly-modified copies of existing data

How to Do Data Augmentation​

Data augmentation can be done as a part of the model itself.

NOTE: That data augmentation is only active when Model.fit for obvious reasons. When calling Model.evaluate or Model.predict, augmentation layers will be inactive

NOTE: Even though, the augmentation layers are inactive, layers such as Rescaling & Resizing layers will be active.

import tensorflow as tf
import matplotlib.pyplot as plt

data_augmentation = tf.keras.Sequential(
    [
        tf.keras.layers.Rescaling(1 / 255),
        tf.keras.layers.RandomRotation(0.2),
        tf.keras.layers.RandomFlip("horizontal_and_vertical"),
    ]
)

plt.figure(figsize=(10, 10))

for i in range(9):
    augmented_image = data_augmentation(tf.expand_dims(image, 0))
    ax = plt.subplot(3, 3, i + 1)
    plt.imshow(augmented_image[0])
    plt.axis("off")

Data augmentation output image

Applying Data Augmentation​

Data augmentation can be defined as a sequence of layers and applied to the model

Define pre-processing & data augmentation layers​

import tensorflow as tf
import matplotlib.pyplot as plt
import pandas as pd

IMG_SIZE = 180

img_preprocessing = tf.keras.Sequential(
    [
        # image resizing can be done in the model so when predicting the image
        # will be auto  resized
        tf.keras.layers.Resizing(IMG_SIZE, IMG_SIZE),
        tf.keras.layers.Rescaling(1.0 / 255),
    ]
)

data_augmentation = tf.keras.Sequential(
    [
        tf.keras.layers.RandomFlip("horizontal"),
        tf.keras.layers.RandomRotation(0.2),
        tf.keras.layers.RandomZoom(0.2),
    ]
)

Building the model​

model = tf.keras.Sequential(
    [
        img_preprocessing,
        data_augmentation,
        tf.keras.layers.Conv2D(10, 3, activation="relu", input_shape=(180, 180, 3)),
        tf.keras.layers.MaxPooling2D(),
        tf.keras.layers.Conv2D(10, 3, activation="relu"),
        tf.keras.layers.MaxPooling2D(),
        tf.keras.layers.Conv2D(20, 3, activation="relu"),
        tf.keras.layers.MaxPooling2D(),
        tf.keras.layers.Flatten(),
        tf.keras.layers.Dense(1, activation="sigmoid"),
    ]
)

model.compile(
    loss=tf.keras.losses.BinaryCrossentropy(),
    optimizer=tf.keras.optimizers.Adam(),
    metrics=["accuracy"],
)

history = model.fit(train_data, epochs=4, validation_data=test_data)