README.md

Pretrained Model (ready to use)

This folder contains a pretrained, ready-to-use marble-sorting model so you can run the sorter without training your own.

If you don't have an NVIDIA GPU / training PC, or you just want to try the sorter quickly, use the files in this folder instead of completing Step 4 — Train the Model.

Contents

File	Description
marbel_coral.tflite	Edge TPU model — runs on the Coral USB Accelerator
labels.txt	Class id-to-name map (must match label_map.pbtxt)

The model is an SSD MobileNet V2 detector, quantized to INT8 and compiled for the Coral Edge TPU.

Classes (labels.txt)

0 black
1 green
2 orange
3 red

How to use it

The sorter loads its model from my-models/ (which is not tracked by git, see .gitignore). Copy this folder's contents into my-models/:

# From the project root, on the Raspberry Pi
mkdir -p my-models
cp my-models-example/marbel_coral.tflite my-models-example/labels.txt my-models/

Then run the sorter as usual:

sudo .venv-3.10/bin/python main.py

Want better accuracy for your own marbles/lighting? Train your own model by following Step 4 — Train the Model and place the output in my-models/ instead.