ML Autoencoder MNIST Digits

Description

This project features an autoencoder model trained to encode, compress, and decode hand-written digits. There are two files, model_functions.py which contains the functions and structure of the model.
Secondly, there is interactive.py which provides a small GUI screen to draw or load a random digit between 0 and 9. The user's drawing can be inputted into the trained model by pressing a key in which the neural network outputs its auto-encoded version.

Installation

• Pip install h5py (built with 3.1.0)
• Pip install tensorflow (built with 2.5.0)
• Pip install numpy (built with 1.19.5)
• Pip install pygame (built with 2.0.1 - for interactive.py)

Usage

Before running any python files please make sure you have completed ALL above installation steps.

Read model_functions.py function build_model, to see the structure of the autoencoder.
If you run model_functions.py, it will build and train a new model which will overwrite the existing saved model and weights. The newly trained model will be evaluated, and it's accuracy outputted.

Upon running interactive.py, a window will pop up. Pressing D passes the input to the Auto Encoder (AE), which populates the right group of pixels with output from the AE. Pressing T loads a random image into the input section. Pressing C clears both the input and output sections.

Autoencoder Details

(InputLayer) Input Shape: (28, 28, 1)
(Flatten) 784 Units
(Dense) 256 Units
(LeakyReLU) Activation Layer
(Dense) 128 Units
(LeakyReLU) Activation Layer
(Dense) 256 Units
(LeakyReLU) Activation Layer
(Dense) 784 Units
(LeakyReLU) Activation Layer
(Output Layer) Output Shape: (28, 28, 1)

Testing accuracy: ~81%

Credits

• Author: Lee Taylor

Note

This is my first autoencoder machine learning project, inspired by the book, Generative Deep Learning by David Foster. Project also inspired by YouTube tutorial by sentdex.