EfficientNet

What EfficientNet Does

EfficientNet is a family of image-classification backbones that offers several accuracy–compute sizes. Rather than making a model only deeper, wider, or higher-resolution, it scales all three dimensions together.

The family was introduced in EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks. EfficientNet-B0 is a compact architecture found with neural architecture search; B1 through B7 are systematically enlarged versions.

Core Idea

EfficientNet uses one compound coefficient ϕ\phi and fixed constants α\alpha, β\beta, and γ\gamma:

depth:      d = α^φ
width:      w = β^φ
resolution: r = γ^φ

subject to: α · β² · γ² ≈ 2
            α ≥ 1, β ≥ 1, γ ≥ 1

Increasing ϕ\phi by one targets roughly twice the FLOPs before practical rounding. The original search selected α=1.2\alpha=1.2, β=1.1\beta=1.1, and γ=1.15\gamma=1.15 for the B0 baseline.

EfficientNet Compound Scaling

Compound Scaling vs. Single-Dimension Scaling

Source: Tan & Le, 2019 — EfficientNet: Rethinking Model Scaling (ICML)

EfficientNet-B0 Architecture

EfficientNet-B0 Architecture with MBConv blocks

Source: Tan & Le, 2019 — EfficientNet: Rethinking Model Scaling (ICML)

How It Works

1. Start from EfficientNet-B0

The B0 search optimized accuracy and FLOPs in an MnasNet-style space. Its stem is followed by seven stages of mobile inverted bottleneck convolutions (MBConv), then a 1×11\times1 head, global average pooling, dropout, and a classifier.

Model Resolution Depth factor Width factor Parameters Reported top-1
B0 224 1.0 1.0 5.3M 77.1%
B3 300 1.4 1.2 12M 81.6%
B7 600 3.1 2.0 66M 84.3%

These are original-paper ImageNet results; comparisons require matching datasets and training recipes.

2. Process Features with MBConv

Most blocks:

  1. expand channels with a 1×11\times1 convolution;
  2. apply a depthwise 3×33\times3 or 5×55\times5 convolution;
  3. recalibrate channels with squeeze-and-excitation;
  4. project back with a linear 1×11\times1 convolution; and
  5. add a residual path when input and output shapes match.

The original architecture uses the swish activation f(x)=xσ(x)f(x)=x\sigma(x) and batch normalization. Its training recipe also uses dropout and stochastic depth.

3. Scale the Three Dimensions Together

More resolution exposes finer detail, more depth expands the receptive computation, and more width supplies channels to represent the additional patterns. Compound scaling aims to keep these capacities balanced instead of over-investing in one dimension.

4. Adapt the Design in EfficientNetV2

EfficientNetV2 uses fused-MBConv blocks in early stages, limits expensive high-resolution expansion, and progressively changes image size and regularization during training. The goal is faster training as well as efficient inference.

Why It Works

  • The B0 baseline is already optimized for an accuracy–cost objective.
  • MBConv and depthwise convolution reduce arithmetic cost, while squeeze-and-excitation selects informative channels.
  • Balanced scaling gives higher-resolution inputs enough network depth and width to use their additional detail.
  • A family of sizes lets practitioners choose a model after measuring their actual latency and memory constraints.

Limitations

  • FLOPs are only a proxy for latency; depthwise operations can be inefficient on some accelerators.
  • Large variants require high-resolution activations and substantial training memory despite their parameter efficiency.
  • The original training recipe is complex, using strong augmentation, dropout, and stochastic depth.
  • Fixed compound ratios do not guarantee the best model for every dataset, task, or device.
  • EfficientNetV2 and later CNN or transformer families may offer better trade-offs in a given deployment.

Influence and Variants

Related

Sources