EfficientNet scaling in Computer Vision - Practice Problems & Coding Challenges

Challenge - 5 Problems

🧠 Conceptual

intermediate

What does compound scaling in EfficientNet do?

EfficientNet uses a compound scaling method to scale up the model. What does this method do?

A It randomly changes the network architecture during training to find the best model.

B It only increases the depth of the network while keeping width and resolution constant.

C It scales the number of output classes to improve accuracy.

D It scales depth, width, and resolution of the network uniformly using fixed coefficients.

💡 Hint

Think about how EfficientNet balances different model dimensions together.

Compound scaling uniformly scales depth, width, and input resolution with fixed coefficients to balance model size and accuracy.

❓ Model Choice

intermediate

Choosing EfficientNet variant for limited GPU memory

You want to train an EfficientNet model on a GPU with limited memory. Which variant should you choose to balance accuracy and memory use?

A EfficientNet-B0

B EfficientNet-B7

C EfficientNet-B5

D EfficientNet-B3

💡 Hint

Smaller variants use less memory but have lower accuracy.

EfficientNet-B0 is the smallest variant, requiring less memory and suitable for limited GPU resources.

❓ Predict Output

advanced

Output shape after EfficientNet scaling

Given the following code snippet using PyTorch, what is the output shape of the tensor after scaling?

Computer Vision

import torch
from torchvision.models import efficientnet_b0

model = efficientnet_b0()
input_tensor = torch.randn(1, 3, 224, 224)
output = model.features(input_tensor)
print(output.shape)

A torch.Size([1, 1280, 14, 14])

B torch.Size([1, 1280, 7, 7])

C torch.Size([1, 320, 7, 7])

D torch.Size([1, 1000])

💡 Hint

Look at the output channels (1280) and spatial size after the feature extractor in EfficientNet-B0.

The features part of EfficientNet-B0 outputs a tensor with shape [batch, 1280, 7, 7] before the avgpool and classifier.

❓ Hyperparameter

advanced

Effect of increasing input resolution in EfficientNet scaling

What is the main effect of increasing the input image resolution in EfficientNet's compound scaling?

A It increases the spatial size of feature maps, improving fine detail capture but increasing computation.

B It reduces the number of layers in the network to speed up training.

C It decreases the number of channels in each layer to reduce overfitting.

D It changes the activation functions to improve non-linearity.

💡 Hint

Think about what happens when you feed larger images into a convolutional network.

Increasing input resolution increases feature map sizes, allowing the model to capture more detail but requiring more computation.

❓ Metrics

expert

Comparing accuracy and FLOPS of EfficientNet variants

Which EfficientNet variant has approximately 19 billion FLOPS and achieves around 84.0% top-1 accuracy on ImageNet?

A EfficientNet-B7

B EfficientNet-B5

C EfficientNet-B6

D EfficientNet-B4

💡 Hint

Recall the FLOPS and accuracy increase with variant number.

EfficientNet-B6 has about 19B FLOPS and achieves ~84.0% top-1 accuracy on ImageNet.