"Draw a digit" is a simple Flask web app for mouse-or-finger-written digit recognition.
Demo:https://draw-a-digit.nw.r.appspot.com/ (shutdown)
Make sure you have Python3.7+ on your machine or virtual environment.
It is recommended to upgrade pip:
pip install --upgrade pip
then just run:
pip install -r requirements.txt
Run the app:
python3 app.py
The app will run on: http://localhost:8080
The model is already trained but if you want to modify something and retrain the network you first need to:
-
add these two lines to
requirements.txt:matplotlib scikit-learn -
execute again:
pip install -r requirements.txt -
launch the training with:
python3 model/CNN.py
If you want to run cross-validation, save it and make a plot of the average performance before the training, you have to uncomment these lines:
# histories = evaluate_model(model, trainX, trainY)
# save_histories(histories)
# histories_mean = mean_folds_history()
# model_performance(histories_mean)
take care of updating and pass to the functions the new default parameters if you changed them (e.g. n_folds, epochs)
The model is a 2xCNN layers with the following architecture:
For more details see model\CNN.py.
The 5-folds cross-validation accuracy of the model reached 99.5% after 10 epochs.