This guide explains how to run a trained Ecto-Trigger model on real devices in the field. We provide instructions for two supported platforms:

Raspberry Pi

To execute the models on Raspberry Pi systems, you can choose to use the Tensorflow or TFLite runtime (reccomended). To use the TFLite runtime, check out the guidance here. Basic steps:

(On RPi)

python3 -m pip install tflite-runtime

Using Python, you can execute a quantised inference:

import numpy as np
from tflite_model_runner import TFLiteModelRunner
 
q_model = TFLiteModelRunner.load_tflite_model("model_weights/8/quant/8_int8.tflite")
 
input_image_array = np.random.uniform(0, 255, size=(q_model.get_input_details()[0]["shape"][1:])).astype(np.uint8)
input_image_array = np.expand_dims(input_image_array, axis=0)
 
q_model.set_tensor(q_model.get_input_details()[0]["index"], input_image_array)
q_model.invoke()
 
output = q_model.get_tensor(q_model.get_output_details()[0]["index"])
print(output[0]) # remember that the output will be in confidence range 0-255

Using this example, you can load images into the input_image_array, by replacing the part using numpy, e.g. you could use PiCamera to take images, and then process them with the model. output[0] will always contain the prediction from a given image, which is just given as an integer number.

ESP32-S3

Deploying models onto microcontroller platforms is a little more complicated, as these don't usually support python, so we have to compile code from scratch to execute on each device. To make things as easy as possible, we have provided an example project which uses our models on ESP32s3 chipset with the Platformio extension for VSCode.

We have made a separate repository for this, which includes full guidance and further details.

InsectCT_lilygo

If you have not yet trained or quantised a model first, use one of ours developed for insect detection, or follow the our usage guidance to train your own.