Computer Vision News - February 2020

3 Summary EfficientNet with PyTo ch 1 2) In order to pursue better accuracy and efficiency, it is critical to balance all dimensions of the network during ConvNet scaling. Efficient-Net offers a schematic way to perform this dimension scaling. This is done simply by the following formula: where Here, is a user specified coefficient that controls the needed resources, which in our case are floating point operations - FLOPs. are constants that can be determined by a grid search and set how to assign the more resources between the depth, width and resolution. Intuitively, the FLOPs of a regular convolution operation is proportional to d,w 2 ,r 2 whichmeans that the constrain in the equation suggests that increasing the model size will increase the total FLOPs by . Model architecture The formula above suggest a method to create a family of architectures, from Efficient-Net-B0 up to Efficient-Net-B7. The idea is to start with and do a small grid search over d,w,r. These optimal parameters will give the initial architecture Efficient-Net-B0. Next under fixed architecture it is possible to change Φ and in this way to create the architecture Efficient-Net-B1 up to Efficient-Net-B7, where the number of FLOPs is controlled. The following figure demonstrates accuracy vs FLOPS and show that this family of architectures achieve its goals: better accuracy with a lower number of FLOPS. ℎ: = Φ , ℎ: = Φ , : = Φ ⋅ 2 ⋅ 2 ≈ 2 and , , ≥ 1 , , Φ 2 Φ Φ Φ = 1 Φ