Pre-trained models (VGG, ResNet, MobileNet) in TensorFlow - Practice Problems & Coding Challenges

Challenge - 5 Problems

❓ Model Choice

intermediate

Choosing the right pre-trained model for image classification

You want to classify images of everyday objects on a mobile device with limited memory and processing power. Which pre-trained model is the best choice?

A VGG16 - It has a very deep architecture but is large and slow.

B MobileNetV2 - It is designed for mobile and embedded devices with low latency.

C ResNet50 - It uses residual connections and is moderately sized.

D InceptionV3 - It has complex modules and is computationally heavy.

💡 Hint

Think about model size and speed for mobile devices.

MobileNetV2 is optimized for mobile devices with fewer parameters and faster inference, making it ideal for limited resources.

❓ Predict Output

intermediate

Output shape of a pre-trained VGG16 model

What is the output shape of the last layer of the VGG16 model loaded with include_top=True for ImageNet classification?

TensorFlow

import tensorflow as tf
model = tf.keras.applications.VGG16(include_top=True)
output_shape = model.output_shape
print(output_shape)

A (None, 7, 7, 512)

B (None, 224, 224, 3)

C (None, 1000)

D (None, 4096)

💡 Hint

Consider the final classification layer output for ImageNet classes.

With include_top=True, VGG16 ends with a Dense layer of 1000 units for ImageNet classes, so output shape is (None, 1000).

❓ Hyperparameter

advanced

Choosing the correct input size for ResNet50

You want to use the pre-trained ResNet50 model from TensorFlow Keras applications. What is the correct input image size to feed into this model?

A 128x128 pixels

B 299x299 pixels

C 64x64 pixels

D 224x224 pixels

💡 Hint

Check the standard input size for ResNet50 trained on ImageNet.

ResNet50 expects input images of size 224x224 pixels as it was trained on ImageNet with this size.

❓ Metrics

advanced

Evaluating MobileNetV2 performance on a custom dataset

You fine-tune MobileNetV2 on a small dataset and want to evaluate its performance. Which metric is most appropriate to measure classification accuracy?

A Accuracy

B Mean Squared Error (MSE)

C Mean Absolute Error (MAE)

D Root Mean Squared Error (RMSE)

💡 Hint

Think about the type of problem: classification or regression.

Accuracy measures the proportion of correct predictions in classification tasks, which fits fine-tuning MobileNetV2 for classification.

🔧 Debug

expert

Debugging a shape mismatch error with ResNet50

You load ResNet50 with include_top=False and try to add a Dense layer with 1000 units directly after the model output. You get a shape mismatch error. Why?

TensorFlow

import tensorflow as tf
base_model = tf.keras.applications.ResNet50(include_top=False, input_shape=(224,224,3))
x = base_model.output
output = tf.keras.layers.Dense(1000, activation='softmax')(x)
model = tf.keras.Model(inputs=base_model.input, outputs=output)

A The output of base_model is 4D (batch, height, width, channels), but Dense expects 2D input.

B The input shape is incorrect; ResNet50 requires 299x299 images.

C Dense layer cannot have 1000 units after a convolutional base.

D You must set include_top=True to add Dense layers.

💡 Hint

Check the shape of the output tensor from the convolutional base.

When include_top=False, the model output is a 4D tensor. Dense layers require 2D input, so you must flatten or pool before Dense.