How to Use EfficientNet in PyTorch: Syntax and Example

To use EfficientNet in PyTorch, import it from torchvision.models, load a pretrained model with efficientnet_b0(pretrained=True), and pass input tensors to get predictions. You can fine-tune or use it for feature extraction by modifying the final layer.

Syntax

Here is the basic syntax to load and use EfficientNet in PyTorch:

• from torchvision.models import efficientnet_b0: Import the EfficientNet B0 model.
• model = efficientnet_b0(pretrained=True): Load the pretrained model weights.
• model.eval(): Set the model to evaluation mode for inference.
• outputs = model(inputs): Pass input tensors to get predictions.

You can replace efficientnet_b0 with other variants like efficientnet_b1 to efficientnet_b7.

python

from torchvision.models import efficientnet_b0
import torch

# Load pretrained EfficientNet B0
model = efficientnet_b0(pretrained=True)
model.eval()

# Example input: batch of 1 image, 3 color channels, 224x224 pixels
inputs = torch.randn(1, 3, 224, 224)

# Get model predictions
outputs = model(inputs)

Example

This example shows how to load EfficientNet B0 pretrained on ImageNet, run a dummy input through it, and print the output shape.

python

from torchvision.models import efficientnet_b0
import torch

# Load pretrained EfficientNet B0
model = efficientnet_b0(pretrained=True)
model.eval()

# Create a dummy input tensor (batch size 1, 3 channels, 224x224 image)
inputs = torch.randn(1, 3, 224, 224)

# Forward pass to get predictions
outputs = model(inputs)

# Print output shape (should be [1, 1000] for ImageNet classes)
print('Output shape:', outputs.shape)

Output

Output shape: torch.Size([1, 1000])

Common Pitfalls

• Wrong input size: EfficientNet expects 224x224 images by default; using other sizes without resizing causes errors or poor results.
• Forgetting model.eval(): Not setting evaluation mode disables dropout and batch norm behavior needed for inference.
• Not normalizing inputs: Inputs should be normalized with ImageNet mean and std for best results.
• Using pretrained weights incorrectly: Ensure pretrained=True to load weights; otherwise, the model is random.

python

from torchvision.models import efficientnet_b0
import torch

# Wrong way: no eval mode
model = efficientnet_b0(pretrained=True)
inputs = torch.randn(1, 3, 224, 224)
outputs_wrong = model(inputs)  # Dropout active, results vary

# Right way: set eval mode
model.eval()
outputs_right = model(inputs)

Quick Reference

Tips for using EfficientNet in PyTorch:

• Use torchvision.models.efficientnet_b0 to efficientnet_b7 for different sizes.
• Input images must be resized to 224x224 and normalized with ImageNet stats.
• Call model.eval() before inference to disable training layers.
• For fine-tuning, replace the final classifier layer with your own.

Key Takeaways

Import EfficientNet from torchvision.models and load pretrained weights with pretrained=True.

Resize and normalize input images to 224x224 with ImageNet mean and std before passing to the model.

Always call model.eval() before inference to get consistent predictions.

You can fine-tune EfficientNet by replacing the final classification layer.

EfficientNet variants range from b0 (smallest) to b7 (largest) for different accuracy and speed trade-offs.