Feature extraction approach in TensorFlow - ML Experiment: Train & Evaluate

Experiment - Feature extraction approach

Problem:We want to classify images of cats and dogs using a neural network. Currently, the model is trained from scratch on a small dataset.

Current Metrics:Training accuracy: 95%, Validation accuracy: 70%, Training loss: 0.15, Validation loss: 0.65

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

Your Task

Reduce overfitting by using a feature extraction approach with a pre-trained model, aiming for validation accuracy above 85% while keeping training accuracy below 90%.

Use TensorFlow and Keras only.

Do not increase the dataset size.

Do not train the entire pre-trained model, only the top layers.

Hint 1

Hint 2

Hint 3

Hint 4

Solution

TensorFlow

import tensorflow as tf
from tensorflow.keras import layers, models
from tensorflow.keras.applications import MobileNetV2
from tensorflow.keras.preprocessing.image import ImageDataGenerator

# Load pre-trained MobileNetV2 without top layers
base_model = MobileNetV2(input_shape=(160, 160, 3), include_top=False, weights='imagenet')
base_model.trainable = False  # Freeze base model

# Add classification head
model = models.Sequential([
    base_model,
    layers.GlobalAveragePooling2D(),
    layers.Dense(128, activation='relu'),
    layers.Dropout(0.3),
    layers.Dense(1, activation='sigmoid')
])

model.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy'])

# Data generators with augmentation
train_datagen = ImageDataGenerator(rescale=1./255, rotation_range=20, width_shift_range=0.2, height_shift_range=0.2, horizontal_flip=True, validation_split=0.2)

train_generator = train_datagen.flow_from_directory(
    'cats_and_dogs/train', target_size=(160, 160), batch_size=32, class_mode='binary', subset='training')

validation_generator = train_datagen.flow_from_directory(
    'cats_and_dogs/train', target_size=(160, 160), batch_size=32, class_mode='binary', subset='validation')

# Train only the top layers
history = model.fit(train_generator, epochs=10, validation_data=validation_generator)

Replaced training from scratch with MobileNetV2 pre-trained model as feature extractor.

Froze the base model layers to prevent overfitting.

Added a small classifier on top with dropout for regularization.

Used data augmentation to improve generalization.

Results Interpretation

Before: Training accuracy 95%, Validation accuracy 70%, showing overfitting.

After: Training accuracy 88%, Validation accuracy 87%, showing better generalization.

Using a pre-trained model as a fixed feature extractor helps reduce overfitting and improves validation accuracy by leveraging learned features from a large dataset.

Bonus Experiment

Try fine-tuning the last few layers of the pre-trained model after initial training to see if validation accuracy improves further.

💡 Hint

Unfreeze the last 10 layers of MobileNetV2 and train with a very low learning rate.

Practice

1. What is the main purpose of using feature extraction with a pre-trained TensorFlow model?

easy

A. To use learned features from a large dataset to help a new task

B. To train a model from scratch with random weights

C. To increase the size of the dataset automatically

D. To reduce the number of layers in the model

Solution

1. Step 1: Understand feature extraction concept

   Feature extraction uses a pre-trained model to get useful features learned from a large dataset.

2. Step 2: Identify the benefit for new tasks

   This helps when you have little data or want faster training by reusing learned features.

3. Final Answer:

   To use learned features from a large dataset to help a new task -> Option A

4. Quick Check:

   Feature extraction = reuse learned features [OK]

Hint: Feature extraction reuses knowledge from big datasets [OK]

Common Mistakes:

• Thinking feature extraction trains from scratch
• Confusing feature extraction with data augmentation
• Believing it reduces model layers automatically

2. Which TensorFlow code snippet correctly freezes the base model layers for feature extraction?

easy

A. base_model.trainable = True

B. base_model.trainable = False

C. base_model.freeze()

D. base_model.layers.freeze = True

Solution

1. Step 1: Recall how to freeze layers in TensorFlow

   Setting trainable = False on the base model freezes its weights.

2. Step 2: Check each option's correctness

   base_model.trainable = False correctly sets base_model.trainable = False. Others are invalid syntax or do not freeze layers.

3. Final Answer:

   base_model.trainable = False -> Option B

4. Quick Check:

   Freeze layers = trainable False [OK]

Hint: Freeze layers by setting trainable to False [OK]

Common Mistakes:

• Using trainable = True to freeze layers
• Calling non-existent freeze() method
• Trying to set freeze attribute on layers directly

3. Given this code, what is the output shape of features?

base_model = tf.keras.applications.MobileNetV2(input_shape=(128,128,3), include_top=False)
base_model.trainable = False
inputs = tf.keras.Input(shape=(128,128,3))
features = base_model(inputs)
print(features.shape)

medium

A. (None, 4, 4, 1280)

B. (None, 7, 7, 512)

C. (None, 1, 1, 1000)

D. (None, 128, 128, 3)

Solution

1. Step 1: Understand MobileNetV2 output with include_top=False

   MobileNetV2 outputs feature maps of size (4,4,1280) for 128x128 input when top layers are excluded.

2. Step 2: Confirm batch dimension is None

   The batch size is None (unspecified), so output shape is (None, 4, 4, 1280).

3. Final Answer:

   (None, 4, 4, 1280) -> Option A

4. Quick Check:

   MobileNetV2 feature shape = (None,4,4,1280) [OK]

Hint: MobileNetV2 outputs 4x4x1280 features for 128x128 input [OK]

Common Mistakes:

• Confusing output with classification layer shape
• Assuming output shape equals input shape
• Ignoring batch dimension None

4. Identify the error in this feature extraction code snippet:

base_model = tf.keras.applications.ResNet50(weights='imagenet', include_top=False)
base_model.trainable = False
inputs = tf.keras.Input(shape=(224,224,3))
x = base_model(inputs)
x = tf.keras.layers.Dense(10)(x)
model = tf.keras.Model(inputs, x)

medium

A. Base model weights not loaded correctly

B. Input shape is incorrect for ResNet50

C. Dense layer applied to 4D tensor without flattening

D. trainable should be True to freeze layers

Solution

1. Step 1: Check tensor shapes before Dense layer

   Base model outputs 4D tensor (batch, height, width, channels). Dense expects 2D input.

2. Step 2: Identify missing flatten or global pooling

   Applying Dense directly to 4D tensor causes error; flatten or global pooling needed first.

3. Final Answer:

   Dense layer applied to 4D tensor without flattening -> Option C

4. Quick Check:

   Dense needs 2D input, flatten first [OK]

Hint: Flatten or pool before Dense layer [OK]

Common Mistakes:

• Applying Dense directly to 4D output
• Confusing input shape requirements
• Misunderstanding trainable flag usage

5. You want to use feature extraction with MobileNetV2 on 64x64 images but get an error when adding a Dense layer after the base model. What is the best way to fix this?

hard

A. Set base_model.trainable = True to allow training

B. Change input shape to 224x224 to match MobileNetV2 default

C. Remove the Dense layer and use base model output directly

D. Add a GlobalAveragePooling2D layer before the Dense layer

Solution

1. Step 1: Understand why Dense layer causes error

   MobileNetV2 outputs 4D feature maps; Dense layers need 2D input.

2. Step 2: Use GlobalAveragePooling2D to convert 4D to 2D

   Adding GlobalAveragePooling2D flattens spatial dimensions, making output suitable for Dense.

3. Step 3: Evaluate other options

   Changing input size is not required; training base model is optional; removing Dense loses classification ability.

4. Final Answer:

   Add a GlobalAveragePooling2D layer before the Dense layer -> Option D

5. Quick Check:

   Pool features before Dense layer [OK]

Hint: Use pooling to flatten features before Dense [OK]

Common Mistakes:

• Changing input size unnecessarily
• Trying to train frozen base model
• Skipping pooling and causing shape errors