In this chapter, you will run image classification by Keras Resnet-50 model using TensorFlow framework on CPU and Neuron core. Experience the steps to compile a pre-trained model for an Inferentia chip and how to optimize the performance.
ResNet model has more than about 20 million parameters, and each parameter is 4 bytes in FP32 format. The Neuron compiler automatically converts them to BFloat16 format (2 bytes per parameter). This is a more efficient data format for the Neuron core to support in hardware.
Please refer to here for running Resnet-50 model using PyTorch framework.