h-jia/tflite-find-arena-size

FInd arena size for TensorFlow Lite models

Find TensorFlow Lite Arena Size

TensorFlow Lite for microcontrollers requires an arena size to be specified, but there's no quick way of calculating the minimum size of the arena. This repository contains an application which loads a tflite model and uses a divide and conquer algorithm to find the optimal size. The size of the arena is dependent on the target architecture, so it's advisable to run this application on the same platform as where you want to deploy the model. What we've found so far is that the arena size on a 64-bit computer (e.g. under macOS) is always bigger than on 32-bits.

Example outputs

Using float32 models (non-quantized) with TensorFlow Lite 2.1.

Multi-layer perceptron with two hidden layers

Environment	Arena size
macOS Catalina	2272 bytes
Mbed OS (Cortex-M4F)	1392 bytes

1D Convolutional neural net with two hidden layers

Environment	Arena size
macOS Catalina	26400 bytes
Mbed OS (Cortex-M4F)	24896 bytes

Cloning this repository

This clones the repository and initializes all submodules:

$ git clone --recurse-submodules https://github.com/edgeimpulse/tflite-find-arena-size.git

Building locally (Docker)

1. Install Docker.

2. Build the application:

   $ docker build -t tflite-find-arena-size .
   

3. Place a tflite file in this folder and name it trained.tflite, and run:

   $ docker run --rm -v $PWD:/data tflite-find-arena-size /data/trained.tflite
   arena_size is 1420
   

Building locally (native)

1. On Linux: Install LLVM 9 (install instructions). The default cmake version on macOS Catalina already works.

2. Build the application:

   $ make
   

3. Run the application and pass in a tflite file:

   $ ./build/find-arena-size some-tflite-file.tflite
   {"arena_size": 291872}