Deep Neural network (DNN) implementation with MNIST db from scratch using NumPy. The layout of the network is set to be modular, so you can explore variations of layers and other hyperparameters with it. Regardless of the number of layers, the model will be fully connected so we can call it MLP (Multilayer perceptron). Feel free to tune the model way you like.
Gif demonstrating forward prop in the model with 2 hidden layers [source]
If you don't wish to explore the training and the modularity of the model, you can test it with the GUI provided with the pre-trained model. There is a simple Tkinter GUI implemented where you can draw your own digits. Those will be inputted to the model and you will be provided with the guess of the model. The implementation is not nearly as good as it could be, but with properly centered drawings, it seems to perform really nicely. The model included will be always the best current version possible. Applying some sort of centering algorithm to the picture would increase the accuracy regardless of the accuracy of the pre-trained model #TODO.
If you want you can train the model with your own parameters. Follow the instructions, please.
In the training you can choose custom values for these parameter:
- Number of hidden layers [positive number] ~Exmpl = 3
- Size of those layers [postive number] ~Exmpl = 256, 128 ,64
- Learning rate [positive number] ~Exmpl = 0.03
- Epochs [positive number] ~Exmpl = 10
- Batch size [positive number] ~Exmpl = 32
- Activation function [RelU, Sigmoid]
You need to have Python installed with Poetry to run this application. Clone the repository to your desired path.
# You should use
$ poetry shell
# Install dependencies
$ poetry install
# Run the GUI to test the pre-trained model
$ poetry run invoke startgui
# Download the data to csv format
$ poetry run invoke dowloadmnist
# Run the training cli
$ poetry run invoke train- Refactor more towards OOP to add more flexibility
- Add more datasets
- Add CNN , RNN implementation
- Evaluate docs such as time and space complexity
- Refactor loss function / add cross-entropy
- Add more precision to to gui / deploy the gui as website
- Add math docs to backprop