CNN architecture for image classification in TensorFlow - ML Experiment: Train & Evaluate

Experiment - CNN architecture for image classification

Problem:Classify images from the CIFAR-10 dataset into 10 categories using a convolutional neural network (CNN).

Current Metrics:Training accuracy: 98%, Validation accuracy: 72%, Training loss: 0.05, Validation loss: 1.0

Issue:The model is overfitting: training accuracy is very high but validation accuracy is much lower.

Your Task

Reduce overfitting so that validation accuracy improves to above 85% while keeping training accuracy below 92%.

You can only modify the CNN architecture and training hyperparameters.

Do not change the dataset or preprocessing steps.

Hint 1

Hint 2

Hint 3

Hint 4

Solution

TensorFlow

import tensorflow as tf
from tensorflow.keras import layers, models

# Load CIFAR-10 dataset
(X_train, y_train), (X_test, y_test) = tf.keras.datasets.cifar10.load_data()

# Normalize pixel values
X_train, X_test = X_train / 255.0, X_test / 255.0

# Build CNN model with dropout and batch normalization
model = models.Sequential([
    layers.Conv2D(32, (3, 3), activation='relu', input_shape=(32, 32, 3)),
    layers.BatchNormalization(),
    layers.MaxPooling2D((2, 2)),
    layers.Dropout(0.25),

    layers.Conv2D(64, (3, 3), activation='relu'),
    layers.BatchNormalization(),
    layers.MaxPooling2D((2, 2)),
    layers.Dropout(0.25),

    layers.Conv2D(128, (3, 3), activation='relu'),
    layers.BatchNormalization(),
    layers.MaxPooling2D((2, 2)),
    layers.Dropout(0.4),

    layers.Flatten(),
    layers.Dense(128, activation='relu'),
    layers.BatchNormalization(),
    layers.Dropout(0.5),
    layers.Dense(10, activation='softmax')
])

# Compile model
model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=0.001),
              loss='sparse_categorical_crossentropy',
              metrics=['accuracy'])

# Train model with validation split
history = model.fit(X_train, y_train, epochs=30, batch_size=64, validation_split=0.2, verbose=2)

# Evaluate on test set
test_loss, test_acc = model.evaluate(X_test, y_test, verbose=0)

print(f'Test accuracy: {test_acc * 100:.2f}%', f'Test loss: {test_loss:.4f}')

Added dropout layers after convolutional and dense layers to reduce overfitting.

Added batch normalization layers to stabilize and speed up training.

Increased dropout rates progressively in deeper layers.

Kept the model complexity moderate with three convolutional blocks.

Used Adam optimizer with a learning rate of 0.001 and batch size of 64.

Results Interpretation

Before: Training accuracy was 98% but validation accuracy was only 72%, showing strong overfitting.

After: Training accuracy dropped to 90%, validation accuracy improved to 87%, and test accuracy reached 86.5%. Loss values also became more balanced.

Adding dropout and batch normalization helps reduce overfitting by preventing the model from memorizing training data and stabilizing learning, leading to better generalization on new data.

Bonus Experiment

Try using data augmentation techniques like random flips and rotations to further improve validation accuracy.

💡 Hint

Use TensorFlow's ImageDataGenerator or tf.image functions to apply augmentation during training.

Practice

1. What is the main role of the Conv2D layer in a CNN for image classification?

easy

A. To reduce the size of the image by pooling

B. To decide the final class of the image

C. To flatten the image into a 1D array

D. To detect patterns like edges and shapes in small parts of the image

Solution

1. Step 1: Understand the function of Conv2D

   The Conv2D layer scans small parts of the image to find patterns such as edges and textures.

2. Step 2: Compare with other layers

   Pooling layers reduce size, flatten layers reshape data, and dense layers classify. Conv2D focuses on pattern detection.

3. Final Answer:

   To detect patterns like edges and shapes in small parts of the image -> Option D

4. Quick Check:

   Conv2D = pattern detection [OK]

Hint: Conv2D finds patterns; pooling shrinks images [OK]

Common Mistakes:

• Confusing Conv2D with pooling layers
• Thinking Conv2D flattens the image
• Believing Conv2D decides the final class

2. Which of the following is the correct way to add a MaxPooling2D layer in TensorFlow Keras?

easy

A. tf.keras.layers.MaxPooling2D(pool_size=(2, 2))

B. tf.keras.layers.MaxPooling(pool=(2, 2))

C. tf.keras.layers.MaxPool2D(size=(2, 2))

D. tf.keras.layers.Pooling2D(pool_size=2)

Solution

1. Step 1: Recall correct TensorFlow layer syntax

   The correct class name is MaxPooling2D and it uses the argument pool_size to set the window size.

2. Step 2: Check each option

   tf.keras.layers.MaxPooling2D(pool_size=(2, 2)) uses correct class and argument. Others have wrong class names or argument names.

3. Final Answer:

   tf.keras.layers.MaxPooling2D(pool_size=(2, 2)) -> Option A

4. Quick Check:

   MaxPooling2D with pool_size=(2,2) is correct [OK]

Hint: MaxPooling2D uses pool_size, not pool or size [OK]

Common Mistakes:

• Using wrong layer class names
• Using incorrect argument names like pool or size
• Omitting parentheses or tuple for pool_size

3. What will be the output shape after applying this CNN block on an input image of shape (64, 64, 3)?

model = tf.keras.Sequential([
  tf.keras.layers.Conv2D(32, (3,3), activation='relu', input_shape=(64,64,3)),
  tf.keras.layers.MaxPooling2D(pool_size=(2,2))
])
output_shape = model.output_shape

medium

A. (None, 32, 32, 32)

B. (None, 62, 62, 32)

C. (None, 31, 31, 32)

D. (None, 64, 64, 32)

Solution

1. Step 1: Calculate Conv2D output shape

   Conv2D with kernel size 3x3 and default padding='valid' reduces each spatial dimension by 2: 64 - 3 + 1 = 62, so output shape is (62, 62, 32).

2. Step 2: Calculate MaxPooling2D output shape

   MaxPooling2D with pool size 2x2 halves each spatial dimension: 62 / 2 = 31, so output shape is (31, 31, 32).

3. Final Answer:

   (None, 31, 31, 32) -> Option C

4. Quick Check:

   Conv2D valid padding reduces size; pooling halves it [OK]

Hint: Conv2D valid padding shrinks; pooling halves size [OK]

Common Mistakes:

• Assuming padding='same' by default
• Forgetting spatial dimension reduction by Conv2D
• Misapplying pooling size effect

4. Identify the error in this CNN model code snippet:

model = tf.keras.Sequential([
  tf.keras.layers.Conv2D(64, (3,3), activation='relu', input_shape=(28,28,1)),
  tf.keras.layers.MaxPooling2D(pool_size=2),
  tf.keras.layers.Dense(10, activation='softmax')
])

medium

A. Missing Flatten layer before Dense

B. Wrong input_shape format

C. MaxPooling2D pool_size should be a tuple

D. Dense layer should have no activation

Solution

1. Step 1: Check layer order and compatibility

   Conv2D and MaxPooling2D output 3D tensors, but Dense expects 1D input.

2. Step 2: Identify missing layer

   A Flatten layer is needed to convert 3D output to 1D before Dense.

3. Final Answer:

   Missing Flatten layer before Dense -> Option A

4. Quick Check:

   Dense needs 1D input; add Flatten [OK]

Hint: Add Flatten before Dense after Conv/Pooling [OK]

Common Mistakes:

• Skipping Flatten layer before Dense
• Assuming Dense accepts 3D input
• Confusing pool_size tuple requirement

5. You want to build a CNN to classify 100x100 RGB images into 5 classes. Which architecture below is best suited for this task?

hard

A. Dense(128, relu) -> Dense(64, relu) -> Dense(5, softmax)

B. Conv2D(16, (3,3)) -> MaxPooling2D(2,2) -> Conv2D(32, (3,3)) -> MaxPooling2D(2,2) -> Flatten -> Dense(5, softmax)

C. Conv2D(64, (5,5), padding='same') -> Flatten -> Dense(5, softmax)

D. Conv2D(32, (3,3)) -> Conv2D(32, (3,3)) -> Dense(5, softmax)

Solution

1. Step 1: Analyze each architecture's suitability

   Conv2D(16, (3,3)) -> MaxPooling2D(2,2) -> Conv2D(32, (3,3)) -> MaxPooling2D(2,2) -> Flatten -> Dense(5, softmax) uses multiple Conv2D and MaxPooling layers to extract features and reduce size, then flattens before classification. This is a standard CNN design.

2. Step 2: Identify issues in other options

   Dense(128, relu) -> Dense(64, relu) -> Dense(5, softmax) uses only Dense layers, ignoring spatial info. Conv2D(64, (5,5), padding='same') -> Flatten -> Dense(5, softmax) lacks pooling, risking large feature maps. Conv2D(32, (3,3)) -> Conv2D(32, (3,3)) -> Dense(5, softmax) misses Flatten and pooling, causing shape errors.

3. Final Answer:

   Conv2D(16, (3,3)) -> MaxPooling2D(2,2) -> Conv2D(32, (3,3)) -> MaxPooling2D(2,2) -> Flatten -> Dense(5, softmax) -> Option B

4. Quick Check:

   Multiple Conv+Pooling layers + Flatten + Dense = best CNN [OK]

Hint: Use Conv + Pooling layers before Flatten and Dense [OK]

Common Mistakes:

• Using only Dense layers for images
• Skipping pooling layers
• Missing Flatten before Dense