Pre-trained models (ResNet, VGG, EfficientNet) in Computer Vision - Practice Problems & Coding Challenges

Challenge - 5 Problems

🧠 Conceptual

intermediate

Understanding the main advantage of using pre-trained models

Why do we often use pre-trained models like ResNet, VGG, or EfficientNet instead of training a model from scratch?

A They always require more data and time to train than new models.

B They allow faster training and better performance by using knowledge from large datasets.

C They are simpler models with fewer layers, so easier to understand.

D They only work for black and white images, not color images.

💡 Hint

Think about how learning from previous experience helps you solve new problems faster.

Pre-trained models have already learned useful features from large datasets like ImageNet. This helps new tasks train faster and often with better accuracy.

❓ Predict Output

intermediate

Output shape after passing an image through a pre-trained ResNet

Given a batch of 8 RGB images of size 224x224 passed through a pre-trained ResNet50 model (without the final classification layer), what is the shape of the output tensor?

Computer Vision

import torch
import torchvision.models as models

model = models.resnet50(pretrained=True)
model = torch.nn.Sequential(*list(model.children())[:-1])  # Remove final layer

input_tensor = torch.randn(8, 3, 224, 224)
output = model(input_tensor)
output_shape = output.shape

A (8, 512, 7, 7)

B (8, 1000)

C (8, 3, 224, 224)

D (8, 2048, 1, 1)

💡 Hint

Check the last convolutional layer output channels and spatial dimensions after pooling.

ResNet50 outputs a feature map with 2048 channels and spatial size 1x1 after global average pooling when the final classification layer is removed.

❓ Model Choice

advanced

Choosing the best pre-trained model for mobile deployment

You want to deploy an image classification model on a mobile device with limited memory and processing power. Which pre-trained model is the best choice?

A ResNet50 - medium size but heavier than EfficientNet

B VGG19 - very deep and large model

C EfficientNet - balances accuracy and efficiency

D ResNet152 - very deep with many layers

💡 Hint

Think about models designed for efficiency and smaller size.

EfficientNet models are designed to provide good accuracy with fewer parameters and less computation, making them suitable for mobile devices.

❓ Hyperparameter

advanced

Adjusting learning rate when fine-tuning a pre-trained model

When fine-tuning a pre-trained VGG16 model on a new dataset, which learning rate strategy is most appropriate?

A Use a very low learning rate (e.g., 0.0001) to avoid destroying learned features

B Use a very high learning rate (e.g., 0.1) to speed up training

C Use the same learning rate as training from scratch (e.g., 0.01)

D Use no learning rate because the model is pre-trained

💡 Hint

Think about how big changes affect already learned knowledge.

A low learning rate helps fine-tune the model gently without losing the useful features learned from the original dataset.

❓ Metrics

expert

Interpreting accuracy differences between pre-trained models

You fine-tune ResNet50 and EfficientNet-B0 on the same dataset. ResNet50 achieves 85% accuracy, EfficientNet-B0 achieves 88%. What is the most likely reason for this difference?

A EfficientNet-B0 has a better architecture that balances depth, width, and resolution.

B ResNet50 was trained with a higher learning rate causing overfitting.

C EfficientNet-B0 uses fewer parameters, so it always performs better.

D ResNet50 is older and cannot learn new datasets.

💡 Hint

Consider model design principles and efficiency.

EfficientNet uses a compound scaling method to optimize model size and accuracy, often outperforming older architectures like ResNet50.