panzer/tf-mnist

TensorFlow MNIST

A homemade digit recognition system

Includes data generation, training, teaching, and prediction

Read about MNIST here

Inspiration

Why do this in the first place? Doesn't a better MNIST dataset already exist?

Yes. Much larger, much higher accuracy MNIST datasets exist. This was created for two reasons:

• I wanted experience generating a dataset myself
• I wanted to see what conditions (image resolution, dataset size, training time) would contribute to better results

Neural Network Structure

Conv2D (15x15, relu) -> Flatten -> Dense (64, relu) -> Dense (10, softmax)

Using a convolution layer is very common for image based classifiers. It allows for a better understanding of spacially related information.