TFLite Model Benchmarking

This tool is meant to analyze and Benchmark a given TensorFlow Lite (TFLite) model with respect to its timing and space requirements.

Tools Used

To analyze a TFLite model, we will benchmark it by the following attributes:

• Aggregate latency statistics
• Number of Weights
• Weights memory size
• Tensors Used

To measure the above attributes I am using the following open-source tools:

• TF Benchmark Tool: This tool is found in the tensorflow repo and is used to estimate the model's latency by measuring the initialization time, 1st inference time, average warmup time, average inference time.
• tflite_analyzer: This tool estimates the memory requirements of the model, i.e. the Number of weights, their memory size, the name of the tensors used, and their dynamic memory allocation.