Mahmoud Abas, Full-Stack Software Engineer & AI

Neural Networks - Building a CNN to classify images in the CIFAR-10 Dataset

In the print dialog, choose "Save as PDF" as the destination.

Neural Networks - Lesson 8

Building a CNN to Classify Images in the CIFAR-10 Dataset

Convolutional layers, a first CNN, and data augmentation

The CIFAR-10 Dataset

60,000 32x32 color images across 10 classes, 6,000 images per class
50,000 training images, 10,000 test images
Classes: airplane, automobile, bird, cat, deer, dog, frog, horse, ship, truck

python

from keras.datasets import cifar10
(x_train, y_train), (x_test, y_test) = cifar10.load_data()
# x_train shape: (50000, 32, 32, 3)

Preparing the Data

One-hot encode labels for categorical_crossentropy; scale pixels to [0, 1]

python

num_classes = 10
y_train = keras.utils.to_categorical(y_train, num_classes)
y_test = keras.utils.to_categorical(y_test, num_classes)
# y_train[444] -> [0,0,0,0,0,0,0,0,0,1]  (class 9)

x_train = x_train.astype('float32') / 255
x_test = x_test.astype('float32') / 255

Keras CNN Layers: `Conv2D`

Conv2D: Slides learned filters over the image to extract spatial features.

python

keras.layers.Conv2D(filters, kernel_size, strides=(1, 1), padding='valid', ...)

filters: number of filters used per location, the depth of the output
kernel_size: (height, width) of the sliding kernel
strides: step size in each dimension, default (1,1)
input_shape: required only on the first layer

Keras CNN Layers: `MaxPooling2D` and `Flatten`

MaxPooling2D: Downsamples spatial dimensions, keeping the strongest activation in each window.

python

keras.layers.MaxPooling2D(pool_size=(2, 2), strides=None, padding='valid')

pool_size: size of the pooling window; strides defaults to pool_size
Flatten turns a multi-dimensional feature map into a 1D vector, used when transitioning from conv layers to Dense layers

Building the First CNN

Conv

MaxPool

Dropout

Flatten

Dense

softmax

Strides of `2` in the first two conv layers quickly shrink the spatial dimensions

python

model_1 = Sequential()
model_1.add(Conv2D(32, kernel_size=(5,5), strides=(2,2), padding='same',
                    input_shape=x_train.shape[1:]))
model_1.add(Activation('relu'))
model_1.add(Conv2D(32, (5,5), strides=(2,2), activation="relu"))
model_1.add(MaxPooling2D(pool_size=(2, 2)))
model_1.add(Dropout(0.25))
model_1.add(Flatten())
model_1.add(Dense(512))
model_1.add(Activation('relu'))
model_1.add(Dropout(0.5))
model_1.add(Dense(num_classes))
model_1.add(Activation('softmax'))

Training the First CNN

Compile with categorical_crossentropy (one-hot labels) + RMSprop + accuracy

python

model_1.compile(loss='categorical_crossentropy',
                 optimizer='rmsprop',
                 metrics=['accuracy'])

model_1.fit(x_train, y_train, batch_size=32, epochs=15,
            validation_data=(x_test, y_test), shuffle=True)

text

Epoch 1/15  accuracy: 0.3136  val_accuracy: 0.4506
Epoch 8/15  accuracy: 0.5748  val_accuracy: 0.6114
Epoch 15/15 accuracy: 0.5731  val_accuracy: 0.5863

Exercise

A Deeper CNN

Conv

MaxPool

Conv

MaxPool

Flatten

Dense

Classification

Use stride 1 for all convolutional layers this time

Count the deeper model's parameters and compare them to the first CNN
Train for 5 epochs; compare training time, loss, and accuracy against the first CNN (train and validation)
Try different structures to see how much accuracy you can achieve

Data Augmentation: Why and How

Data Augmentation: Artificially expands the training set by creating modified copies of existing images, helping the model generalize better and reducing overfitting.

python

from tensorflow.keras.preprocessing.image import ImageDataGenerator

datagen = ImageDataGenerator(
    rotation_range=20,
    width_shift_range=0.1,
    height_shift_range=0.1,
    horizontal_flip=False,
    shear_range=0.1,
    zoom_range=0.1,
    fill_mode='nearest'
)

Training with Augmented Data

Feed the model batches from datagen.flow() instead of raw arrays

python

model_1.fit(
    datagen.flow(x_train, y_train, batch_size=batch_size),
    epochs=15,
    validation_data=(x_test, y_test)
)

text

Epoch 1/15   accuracy: 0.3195  val_accuracy: 0.5205
Epoch 9/15   accuracy: 0.6074  val_accuracy: 0.6604
Epoch 15/15  accuracy: 0.6221  val_accuracy: 0.6285

Validation accuracy improves and is more stable than without augmentation, peak ~66% vs ~61%
Exercise: tune augmentation parameters, can you make performance better, or worse?

Key Takeaways

Conv2D extracts spatial features via learned filters; MaxPooling2D downsamples; Flatten bridges to Dense layers
Even a small CNN can have 100K+ parameters, architecture choices (strides, filters) directly control model size
categorical_crossentropy pairs with one-hot encoded labels, versus sparse_categorical_crossentropy for integer labels
A basic CNN on CIFAR-10 plateaus around 55-61% validation accuracy, a starting point, not a ceiling
Data augmentation (rotation, shift, shear, zoom) expands effective training data and improves generalization
Deeper architectures (more Conv/MaxPool stages) and augmentation are the two levers explored to push accuracy higher

Open this lesson

Mahmoud AbasBuilding a CNN to classify images in the CIFAR-10 Dataset

num_classes = 10 y_train = keras.utils.to_categorical(y_train, num_classes) y_test = keras.utils.to_categorical(y_test, num_classes) # y_train[444] -> [0,0,0,0,0,0,0,0,0,1] (class 9) x_train = x_train.astype('float32') / 255 x_test = x_test.astype('float32') / 255

model_1 = Sequential() model_1.add(Conv2D(32, kernel_size=(5,5), strides=(2,2), padding='same', input_shape=x_train.shape[1:])) model_1.add(Activation('relu')) model_1.add(Conv2D(32, (5,5), strides=(2,2), activation="relu")) model_1.add(MaxPooling2D(pool_size=(2, 2))) model_1.add(Dropout(0.25)) model_1.add(Flatten()) model_1.add(Dense(512)) model_1.add(Activation('relu')) model_1.add(Dropout(0.5)) model_1.add(Dense(num_classes)) model_1.add(Activation('softmax'))

model_1.compile(loss='categorical_crossentropy', optimizer='rmsprop', metrics=['accuracy']) model_1.fit(x_train, y_train, batch_size=32, epochs=15, validation_data=(x_test, y_test), shuffle=True)

from tensorflow.keras.preprocessing.image import ImageDataGenerator datagen = ImageDataGenerator( rotation_range=20, width_shift_range=0.1, height_shift_range=0.1, horizontal_flip=False, shear_range=0.1, zoom_range=0.1, fill_mode='nearest' )

Building a CNN to Classify Images in the CIFAR-10 Dataset

The CIFAR-10 Dataset

Preparing the Data

Keras CNN Layers: Conv2D

Keras CNN Layers: MaxPooling2D and Flatten

Building the First CNN

Training the First CNN

A Deeper CNN

Data Augmentation: Why and How

Training with Augmented Data

Key Takeaways

Open this lesson

Building a CNN to Classify Images in the CIFAR-10 Dataset

The CIFAR-10 Dataset

Preparing the Data

Keras CNN Layers: Conv2D

Keras CNN Layers: MaxPooling2D and Flatten

Building the First CNN

Training the First CNN

A Deeper CNN

Data Augmentation: Why and How

Training with Augmented Data

Key Takeaways

Open this lesson

Keras CNN Layers: `Conv2D`

Keras CNN Layers: `MaxPooling2D` and `Flatten`

Keras CNN Layers: `Conv2D`

Keras CNN Layers: `MaxPooling2D` and `Flatten`