MACHINE LEARNING STARTER KIT (ml.reversal.org)
==============================================

START IN 3 STEPS (Windows)
  1. Install Python 3.10+ from https://www.python.org/downloads/ (tick "Add python.exe to PATH").
  2. Unzip this folder anywhere.
  3. Double-click run.bat. The first run sets up packages (a few minutes), then trains on the sample images.

macOS / Linux: open a terminal in this folder and run  ./run.sh   (same options as run.bat).

USE YOUR OWN IMAGES
  Put them in images\<label>\ folders, one folder per answer, e.g. images\crack\ and images
o_crack\.
  Delete the sample digit folders first. Aim for 30+ images per folder. PNG, JPG, BMP, GIF and WEBP work.

SEE WHAT YOUR MODEL THINKS
  run.bat panel      opens a page on this computer: pick a model, upload a picture, press Predict.
                     Shows each label's probability, "Uncertain" below 70%, a history, and Mark wrong.
  run.bat catsdogs   downloads the Oxford pets photos (about 800 MB, CC BY-SA 4.0) and trains a cats vs dogs
                     model with a pretrained network (MobileNetV3). Installs PyTorch the first time.
  run.bat images --engine mobilenet --images my_photos --name my-model   the same engine on your own photos

THE EXAMPLES
  run.bat images  [--tune 20 --size 32]   SUPERVISED: learns your folders, confusion matrix, tuning, fit check
  run.bat groups  [--groups 5]            UNSUPERVISED: groups images without looking at folder names
  run.bat numbers [--tune 30]             SUPERVISED REGRESSION: predicts a number, judged by MAE
                  [--csv your.csv --target column]

WHAT YOU GET
  results.json, confusion_matrix.csv, model.json (the trained model) and images.csv.
  Upload images.csv to a container on ml.reversal.org (target column: label) and use Auto train to compare
  every model type on our workers, or download the container package to tune on your own hardware.

GUIDE: https://ml.reversal.org/blueprint/
Models make mistakes. Scores describe your images and this split; results on new images will differ.
