Data Augmentation

Create more training data.

What is Data Augmentation?

Creating new training examples from existing ones.

Goal: More data → Better model

Image Augmentation

from tensorflow.keras.preprocessing.image import ImageDataGenerator

# Define augmentations
datagen = ImageDataGenerator(
    rotation_range=20,      # Rotate up to 20 degrees
    width_shift_range=0.2,  # Shift left/right
    height_shift_range=0.2, # Shift up/down
    horizontal_flip=True,   # Mirror image
    zoom_range=0.2,         # Zoom in/out
    fill_mode='nearest'
)

# Apply to images
augmented_images = datagen.flow(images, batch_size=32)

# Use in training
model.fit(augmented_images, epochs=50)

Albumentations (Better for Images)

import albumentations as A
import cv2

# Define pipeline
transform = A.Compose([
    A.RandomRotate90(),
    A.Flip(),
    A.Transpose(),
    A.GaussNoise(),
    A.OneOf([
        A.MotionBlur(p=0.2),
        A.MedianBlur(blur_limit=3, p=0.1),
        A.Blur(blur_limit=3, p=0.1),
    ], p=0.2),
    A.RandomBrightnessContrast(p=0.2),
])

# Apply
image = cv2.imread('photo.jpg')
augmented = transform(image=image)['image']

Text Augmentation

import nlpaug.augmenter.word as naw

# Synonym replacement
aug = naw.SynonymAug(aug_src='wordnet')
text = "The food in Miami is amazing"
augmented = aug.augment(text)
print(augmented)
# "The food in Miami is fantastic"

# Back translation (more realistic)
aug = naw.BackTranslationAug(
    from_model_name='facebook/wmt19-en-de',
    to_model_name='facebook/wmt19-de-en'
)
augmented = aug.augment(text)

Audio Augmentation

import librosa
import numpy as np

def augment_audio(audio, sr):
    # Time stretch
    audio_stretched = librosa.effects.time_stretch(audio, rate=1.1)
    
    # Pitch shift
    audio_pitched = librosa.effects.pitch_shift(audio, sr=sr, n_steps=2)
    
    # Add noise
    noise = np.random.randn(len(audio))
    audio_noisy = audio + 0.005 * noise
    
    return [audio_stretched, audio_pitched, audio_noisy]

Tabular Data Augmentation

from sklearn.utils import resample
import pandas as pd

# Oversample minority class
df_minority = df[df['target'] == 1]
df_majority = df[df['target'] == 0]

# Resample with replacement
df_minority_upsampled = resample(
    df_minority,
    replace=True,
    n_samples=len(df_majority),
    random_state=42
)

# Combine
df_balanced = pd.concat([df_majority, df_minority_upsampled])

SMOTE for Tabular Data

from imblearn.over_sampling import SMOTE

# Create synthetic samples
smote = SMOTE()
X_resampled, y_resampled = smote.fit_resample(X, y)

print(f"Original: {len(X)}")
print(f"After SMOTE: {len(X_resampled)}")

MixUp (Advanced)

Mix two images:

def mixup(x1, y1, x2, y2, alpha=0.2):
    # Random mixing ratio
    lam = np.random.beta(alpha, alpha)
    
    # Mix images
    x = lam * x1 + (1 - lam) * x2
    
    # Mix labels
    y = lam * y1 + (1 - lam) * y2
    
    return x, y

# Use in training
x_mix, y_mix = mixup(image1, label1, image2, label2)

CutMix

Cut part of image and paste another:

def cutmix(x1, y1, x2, y2, alpha=1.0):
    lam = np.random.beta(alpha, alpha)
    
    # Random box
    h, w = x1.shape[:2]
    cut_h = int(h * np.sqrt(1 - lam))
    cut_w = int(w * np.sqrt(1 - lam))
    
    cx = np.random.randint(w)
    cy = np.random.randint(h)
    
    x1y1 = np.clip([cx - cut_w // 2, cy - cut_h // 2], 0, [w, h])
    x2y2 = np.clip([cx + cut_w // 2, cy + cut_h // 2], 0, [w, h])
    
    # Replace region
    x1[x1y1[1]:x2y2[1], x1y1[0]:x2y2[0]] = x2[x1y1[1]:x2y2[1], x1y1[0]:x2y2[0]]
    
    return x1, y1

Best Practices

1. Images: Always augment (flip, rotate, crop)
2. Text: Be careful (meaning shouldn't change)
3. Validation: Never augment validation data!
4. Realistic: Augmentations should be realistic

Remember

• More data = better model
• Augment during training
• Don't overdo it
• Keep validation data clean