PH#4: EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks
The paper that gave us the EfficientNet family of CNN models, and taught us how to efficiently scale up CNNs.
Why you should read this paper?
Efficient Nets are the most commonly used CNN models, applied across several tasks and consistently give very good trade offs between #train-able parameters vs accuracy.
Haiku:
Want to scale up your model?
Try balancing growth in:
Depth, width & resolution.
Bigger image (γ),
needs more receptive field,
= more model depth(α).
Bigger image,
Need to capture fine-grained patterns,
= more channels (width)(β)
Use Multi Objective NAS,
Optimize for Acc & FLOPS,
Produce EfficientNet-B0.
Use grid search
to find optimal α,β,γ
For Efficient-B0 : α=1.2, β=1.1, γ=1.15
Compound Scale EffNet-B0,
Create EfficientNet-B1 to B7
Achieve SOTA!