GlenCrawford / tensorflow_number_recognition

Digit recognition with a Tensorflow (1.14) neural network.

Number recognition with a neural network with Tensorflow

Build and train a Neural Network with Tensorflow to recognize handwritten digits with ~92% accuracy.

Adapted from this tutorial by Ellie Birbeck, updated for later Tensorflow versions and with documentation inlined as comments.

Note that this uses Tensorflow 1.14, not 2.

Setup

Setting up Python on a Mac

The MacOS built-in Python is 2.7.X, which is old and soon to be deprecated. Like Ruby, you can use a version manager tool to manage and use multiple versions instead of the default one.

• Install pyenv, a Python version manager (like RVM): brew install pyenv
• Install the latest stable of Python into pyenv: pyenv install 3.7.4
• Set the global default Python version: pyenv global 3.7.4
• To load pyenv into your bash shell sessions: echo -e 'if command -v pyenv 1>/dev/null 2>&1; then\n eval "$(pyenv init -)"\nfi' >> ~/.bash_profile
• And load it into your current session: source ~/.bash_profile
• Check that pyenv hijacks Python in the shell: which python
• Check the active version is what you installed: python -V
• And check that you have pip (the Python package manager): pip -V

Set up the project

cd to where you keep your projects, then:

mkdir tensorflow_number_recognition
cd tensorflow_number_recognition
git init

Initialize a virtual environment for this project

Using venv, the built-in module for virtual environments in Python.

python -m venv tensorflow_number_recognition
touch .gitignore
echo 'tensorflow_number_recognition' > .gitignore
source tensorflow_number_recognition/bin/activate

Install dependencies using pip

touch dependencies.txt

Add to the file:

image==1.5.20
numpy==1.14.5
tensorflow==1.14.0

And install the dependencies:

pip install -r dependencies.txt

Write the Python script

touch main.py
echo 'MNIST_data' > .gitignore

See main.py file in this repo.

Run the Python script to train and test the network

python main.py

Test the network manually with real input

curl -O https://raw.githubusercontent.com/do-community/tensorflow-digit-recognition/master/test_img.png

Then uncomment the last section of main.py and run it again.