Image classification models trained on ImageNet (ILSVRC2012)

This collection of image classification models is trained on the ImageNet Large Scale Visual Recognition Challenge dataset (ILSVRC2012). These models use different neural network architectures and different input sizes to trade off accuracy and speed.

The plot below summarizes the accuracy and speed of each model. The model’s accuracy is measured in terms of its top 1 error rate - how often is the model’s top prediction correct. The model’s speed is measured in milliseconds per input on a Raspberry Pi 3B, running at 700MHz, using only the quad-core ARM CPU (and not the VideoCore GPU). See config.txt for the boot config file used to configure the Raspberry Pi test machines.

The following table contains links to information pages that provide more detail about each model. Each information page includes a link to a .ell model file, which can be compiled and deployed using the ELL compiler.