Summary

TFLite is a feature of TF2.0 that takes a TF model and compresses and optimizes it to run on an embedded Linux device, or a low-power and low-binary device. Converting a TF model into a TFLite model can be done in three ways: from a saved model, a tf.keras model, or a concrete function. Once the model has been converted, a .tflite file will be created, which can then be transferred to the desired device and run using the TFLite interpreter. This model is optimized to use hardware acceleration and is stored in FlatBuffer format for quick read speeds. Other optimization techniques can be applied to the model, such as quantization, which converts the 32-bit floating point numbers into 8-bit fixed-point numbers, with a tradeoff of a minimal amount of accuracy. Some devices that TFLite can be run on are the Edge TPU, the NVIDIA Jetson Nano, and the Raspberry Pi. Google also...