TFLITE ACAP

Overview

This examples shows how to run TLITE models in an Axis camera using the ACAP platform. The TFLITE models must have one output with lable scores provided as an int8 array. Models exported from Googels Teachable Machine are aligned with this output.

The final output are ACAPs that can run camera based on:

ARTPEC-8
ARMv7hf TPU (P3255, AXIS Q1615 Mk III, Q1715, M4308 )
ARMv7hf (All models without TPU. CPU will be used)

The package can be compiled as-is. The model included is mobilenet V2 224 but the idea is to replace this with your own model.

Prerequisite

A Linux PC (or Linux VM) with Docker and Git installed

Building

Clone this repository
Use Googels Teachable Machine to train your model.
Export the model in both TFLITE Edge TPU and TFLITE Quntization.
Unzip both files and place the files in source/model. The labels.txt should be placed in the same diorectory. It is recommeded to have file names that easily seperates the EdgeTPU from the Quantized model file. The labels.txt must be called labels.txt and must have an empty last line.
Edit the Dockerfile line 70 and 72 with the filename you saved in source/model/ e.g. /opt/app/model/model_quant.tflite. Make sure that the EdgeTPU and Quant file are set on the correct lines based on platform. Note that Dockerfile will copy the correct file to model/model.tflite to be included in the ACAP depending on the platform selected.
Compile the ACAP from tflite_1/ directory. Type:
. artpec8.sh
. edgetpu.sh
. armv7hf.sh
Install the eap-files in appropriate camera model

Usage

Clients may request inference using the the URL http://camera-ip/local/tflite/inference. The ACAP web page uses the same CGI to update the result every 500ms.

Response

{
  "device":"B8A44FXXXXXX",
  "timestamp":1681550011529,
  "duration":39,  //Amount of milliseconds used for inference
  "list":[
    { "label": string, "score": number 0-100},
    ...
  ]
}

You can set confidence to elimiate unwanted detections in the list.

Customization

You can customize the package in name, HTML, CGI, behavior and output.
If you are using a model with size different to 224x224, edit source/html/config/model.json.

The file main.c shows two examples to make inference and process the output

HTTP Request - for the web page an clients that integrate using HTTP
Timer - If the ACAP needs support other integration methods. Look at hte example code that iterates through the detection list and extracts the lable and its score.

Name

The ACAP has a package name (tflite) and a Nice Name (TFLITE xxxx). These names can be changed. Edit the following files:

source/manifest.json.artpec-8
source/manifest.json.cpu
source/manifest.json.edgetpu
source/Makefile (row 1)
source/main.c (row 21)

User interface page

source/html/index.html

Troubleshooting

If log file says malloc(): invalid size (unsorted) the labels file does not contain one empty line at the bottom.

pandosme/TFLITE_1