A handwritten digit recognizer built using: Python, Tensorflow and Keras.
The dataset used in that project is the MNIST dataset which is a very famous and clean dataset and that's why I skipped the cleaning step in that project.
The MNIST dataset is loaded first from the Keras dataset, you can load it using any other way.
I have divided the data into three groups:
training data (xtr,ytr) : used to train the model, it has 60k digits images represented by a 28X28 matrix and the test set has 10k images.
validation data (xval,yval) : used to tone the model, it has 8k digits images.
test data (xtest,ytest) : used for the final evaluation of the model, , it has 62k digits images
Due to the fact that the data is already very clean the pre-processing step was a short and nice one, I have just normalized the data by dividing them by 255 and used the one-hot-encoding method to represent the labels.
For the modelling step:
I've just started with a very basic neural network model - only 2 layers - and got a validation accuracy of 92.1% and test accuracy of 94.25% which was a good start due to the simplicity of that model.
And then I started adding more complexity by adding more layers - also tried different optimizers: Adam, and different learning rates: .01,.1,.5 - until I reached a validation accuracy of %97.48 and test accuracy of 98.65%, any changes after that was getting less accuracy
After that I decided to try a convolutional neural network, once again I started with a very basic CNN model, and I was quite surprised - Note: that's my first deep learning project :) - as I got a validation accuracy of 98.39% and test accuracy of 98.85% which is even better than the second complex model.
For my final model, I also used CNN but added more layers until I had an overfitting problem:), With that model, I managed to get a validation accuracy of 99.18% and test accuracy of 99.25%.
Aml-Hassan-Abd-El-hamid/Handwritten-Digit-Recognition
A handwritten digit recognizer built using: Python, Tensorflow and Keras
Jupyter Notebook