Oppiskele Regularization and Optimization Strategies | Transfer Learning in Computer Vision

Regularization and Optimization in Transfer Learning

When adapting pre-trained models, regularization and optimization strategies help prevent overfitting and improve generalization:

Dropout: randomly drops units during training to prevent co-adaptation;
Batch normalization: normalizes activations to stabilize and speed up training;
Data augmentation: expands the training set by applying random transformations (flip, rotate, zoom, etc.);
Differential learning rates: use lower learning rates for pre-trained layers and higher for new layers.

Note

Data augmentation is especially important when your target dataset is small.


              12345678910111213141516171819202122232425
            
# Example: Adding regularization and data augmentation
from tensorflow.keras import layers, models
from tensorflow.keras.preprocessing.image import ImageDataGenerator

# Data augmentation pipeline
train_datagen = ImageDataGenerator(
    rotation_range=20,
    width_shift_range=0.2,
    height_shift_range=0.2,
    horizontal_flip=True,
    zoom_range=0.2
)

# Build model with dropout and batch normalization
model = models.Sequential([
    VGG16(weights='imagenet', include_top=False, input_shape=(224, 224, 3)),
    layers.Flatten(),
    layers.Dense(256, activation='relu'),
    layers.BatchNormalization(),
    layers.Dropout(0.5),
    layers.Dense(10, activation='softmax')
])

# Compile model
model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])

from tensorflow.keras.preprocessing.image import ImageDataGenerator

# Data augmentation pipeline
train_datagen = ImageDataGenerator(
    rotation_range=20,
    width_shift_range=0.2,
    height_shift_range=0.2,
    horizontal_flip=True,
    zoom_range=0.2
)

This block defines a data generator that applies random transformations to training images. Instead of manually creating more images, it produces slightly altered versions in real time during training, which helps the model become more robust.

rotation_range=20 rotates each image by up to 20 degrees in any direction.
width_shift_range=0.2 and height_shift_range=0.2 move images horizontally or vertically by up to 20% of their width or height.
horizontal_flip=True mirrors images horizontally, useful when objects can appear facing either direction.
zoom_range=0.2 randomly zooms in or out by up to 20%.

The augmented images are generated on the fly every epoch, so the model sees slightly different inputs each time, improving generalization and reducing overfitting.

# Build model with dropout and batch normalization
model = models.Sequential([
    VGG16(weights='imagenet', include_top=False, input_shape=(224, 224, 3)),
    layers.Flatten(),
    layers.Dense(256, activation='relu'),
    layers.BatchNormalization(),
    layers.Dropout(0.5),
    layers.Dense(10, activation='softmax')
])

The model combines a pre-trained VGG16 feature extractor with new dense layers for classification. BatchNormalization() stabilizes activations, while Dropout() randomly disables neurons during training to prevent overfitting.

# Compile model
model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])

The model is compiled using the Adam optimizer and categorical cross-entropy loss, suitable for multi-class classification tasks.

Oliko kaikki selvää?

Kiitos palautteestasi!

Osio 2. Luku 3

Kysy tekoälyä

Kysy mitä tahansa tai kokeile jotakin ehdotetuista kysymyksistä aloittaaksesi keskustelumme

Awesome!

Completion rate improved to 9.09

Pyyhkäise näyttääksesi valikon