Prediction Using Pre-trained EfficientNet

This lesson will provide a step-by-step guide to building inference scripts using EfficientNet. Moreover, this lesson has multiple interactive playgrounds for you to practice with.

Overview of EfficientNet

EfficientNet is another convolutional neural network architecture that performs well on Imagenet and other image classification tasks. During its initial release in 2019, it was considered a state-of-the-art model.

It’s an efficiency-oriented model that utilizes a scaling method that uniformly scales all dimensions (depth, width, and resolution) using a compound coefficient. Hence, it tends to outperform other architectures without extensive grid search of hyperparameters.

Variant (B0 - B7)

EfficientNet comes with eight different variants, and each one contains different parameters:

• resolution: This has different resolutions for each variant.
• depth and width: The channel size is a multiple of 8.
• resource limit: This is the increasing depth or width. Retaining the resolution might improve performance.

The table below highlights the desired resolution of the input image for all the variants.

Base Model	Resolution
EfficientNetB0	224
EfficientNetB1	240
EfficientNetB2	260
EfficientNetB3	300
EfficientNetB4	380
EfficientNetB5	456
EfficientNetB6	528
EfficientNetB7	600

PyTorch Image Model

Like ResNet, the PyTorch Image Model framework has a pre-trained model for EfficientNet. Currently, it provides the pre-trained weights for B0, B1, B2, B3, and B4.

Import

We can use the create_model function provided by timm and easily load the pre-trained model by using the following code snippet: