This is a collection of Tensflow scripts for implementing neural networks to recognize hand-written digits using the MNIST dataset for training and testing.
This is a TensorFlow 2.0 compatible script using Keras. It creates a simple 3-layer neural network using stochastic gradient descent as the optimizer for the mean-squared-error loss function. Activation functions for each layer is the sigmoid function.
It performs poorly, reaching an accuracy of ~60% at best.
This is a more robust version of the previous script that uses 'adam', a much more efficient optimizer function, along with cross-entropy for the loss function. An additional dropout layer is also incorporated. This version also does not require the image label values to be one-hot encoded, allowing for runs to be quicker. It achieves an accuracy of 97% using half as many training epochs as the simpler model.
This is the best performing neural network for the MNIST dataset. It deploys a 2D convolutional neural network for the first section, image feature extraction. It then uses a final single 128-node hidden layer to process the image data, before outputting to the 10-node output layer. Best accuracy so far is 98%.
A script to load saved, trained models from the previous scripts and test the neural networks against our own set of created hand-written digits.
The images were created with an image program using a digital pen/tablet and
saved as 28x28 PNG files. They are in the images directory.
This is an original code written for Tensorflow 1 and is only included here for demonstration purposes.
It is adapted from the excellent tutorial presented at the Katakoda site