TFLite Model Inference Code

Created and coded by:
Behiç KILINÇKAYA (https://github.com/BehicKlncky) & Sencer YÜCEL (https://github.com/senceryucel)

This script aims to test the performance of a TFLite Classification Model.

MicroPython has been used in this script on OpenMV IDE.

Algorithm

It takes the .jpg image and crop it into desired number of equal parts (name of the variable to set crop number is CROP_COUNT). Then, for every mini-frame (cropped photo), TFLite Model runs and the result of classification is compared with the grand truth (.json annotation). The performance of the model is calculating based on this comparison.

Pre-Requirities

- H7 Plus or or higher version of an OpenMV board.

- OpenMV IDE.

- A model to test.

- A testing dataset in .jpg and their annotations in .json; example format of the .json file is in the repository [Recommended: 100+ photos with different scenarios (easy-medium-hard)].

Usage

1-) Open INFERENCE_CODE.py on OpenMV IDE, then plug your OpenMV board in to your computer.

2-) Construct the first lines of the script with your own paths to your own directories:

PATH_TO_JSON = "PATH_TO_ANNOTATION_JSON_FILE"
PATH_TO_DATASET = "PATH_TO_YOUR_DATASET"
PATH_TO_CROPPED_PHOTOS_TO_SAVE = "PATH_TO_CROPPED_PHOTOS_TO_SAVE"
PATH_TO_TFLITE_MODEL = "PATH_TO_MODEL.tflite"
PATH_TO_SAVE_INFERENCE_RESULTS = "INFERENCE_RESULTS.txt"
CROP_COUNT = 16

What CROP_COUNT is has been described in the Algorithm part.

3-) Connect your OpenMV Board to the IDE with Ctrl+E and execute the script with Ctrl+R.

4-) You have your output in txt format to the directory you set with the information of [accuracy, precision, recall, F1_score] of your TFLite model.