EfficientNet
EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks · Also known as: EfficientNet, compound scaling CNN, EfficientNet-B0 through B7, EfficientNetV2
EfficientNet is a family of convolutional neural network architectures introduced by Mingxing Tan and Quoc V. Le (Google Brain) at ICML 2019 that systematically co-scales network depth, width, and input resolution using a single compound coefficient, achieving state-of-the-art image classification accuracy with substantially fewer parameters and FLOPs than prior networks such as ResNet and Inception.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
EfficientNet is appropriate for image classification, feature extraction, and transfer learning tasks where accuracy per parameter or per FLOP is important — including deployments on resource-constrained devices. Choose a smaller variant (B0–B2) for mobile or embedded targets and a larger one (B5–B7) when GPU memory is abundant. It assumes the task is visual and benefits from pre-training on large image corpora. It is less suitable for non-image modalities, 1-D time series, or very small datasets where a simpler CNN may suffice.
Strengths & limitations
- Consistent accuracy–efficiency trade-off: EfficientNet-B7 matched the then-state-of-the-art ImageNet top-1 accuracy (84.4%) with 8.4x fewer parameters than the next-best model.
- Principled scaling: the compound coefficient provides a transparent, reproducible way to trade compute for accuracy, avoiding ad hoc architecture engineering.
- Strong transfer learning backbone: pre-trained EfficientNet features generalise well to diverse downstream vision tasks with modest fine-tuning data.
- Scalable family: a single design philosophy covers the full range from mobile (B0, 5.3 M parameters) to high-accuracy server (B7, 66 M parameters).
- Wide ecosystem support: pre-trained weights are available in TensorFlow/Keras, PyTorch (torchvision, timm), and JAX/Flax.
- Training from scratch is slow: the compound scaling and MBConv blocks make full ImageNet training substantially longer than simpler architectures like ResNet.
- Memory-intensive at high resolution: B6 and B7 require large GPU memory and smaller batch sizes, which can destabilise batch normalisation.
- Sensitivity to regularisation: EfficientNet requires careful dropout and data augmentation tuning (RandAugment, Mixup, CutMix); without it, larger variants overfit on small datasets.
- Not designed for non-image data: the MBConv backbone is specialised for 2-D spatial inputs and does not transfer naturally to text, tabular, or graph data.
- Superseded in some benchmarks: vision transformers (ViT, Swin) and EfficientNetV2 outperform the original family on very large datasets and modern training recipes.
Frequently asked
Which EfficientNet variant should I start with?
Begin with B0 or B1 for an accuracy-efficiency baseline. Move to B3–B5 if accuracy is insufficient and GPU memory permits. B6 and B7 are mainly relevant for leaderboard-level performance on large, diverse datasets and require high-resolution inputs (528px and 600px respectively).
What is the compound coefficient phi in practice?
Phi is an integer that determines how aggressively depth, width, and resolution are scaled relative to the B0 baseline. B0 corresponds to phi=0; B1 to phi=0.5; through to B7 at phi=6. You do not set phi directly — you simply choose the variant label (B0–B7).
Can EfficientNet be used without ImageNet pre-training?
Yes, but training from scratch requires a large dataset and careful regularisation. In practice, fine-tuning from ImageNet weights converges faster and generalises better for most downstream tasks. For very small datasets (fewer than a few thousand images), use pre-trained weights and freeze most of the trunk.
How does EfficientNet compare to Vision Transformers?
On standard ImageNet benchmarks with moderate data, EfficientNet variants are competitive with similarly sized vision transformers. Transformers tend to outperform CNNs at very large scales and with large pre-training corpora (e.g., JFT-300M), while EfficientNet remains more parameter-efficient in the mid-range. EfficientNetV2 narrows the gap with improved training recipes.
Sources
- Tan, M. & Le, Q. V. (2019). EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks. Proceedings of the 36th International Conference on Machine Learning (ICML 2019), PMLR 97, 6105–6114. link ↗
- Goodfellow, I., Bengio, Y. & Courville, A. (2016). Deep Learning. MIT Press. ISBN: 978-0-262-03561-3
How to cite this page
ScholarGate. (2026, June 3). EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks. ScholarGate. https://scholargate.app/en/deep-learning/efficientnet
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- MobileNetDeep learning↔ compare
- Neural Architecture SearchDeep learning↔ compare
- ResNetDeep learning↔ compare
- Transfer LearningMachine learning↔ compare