Project: Sign Language Digit Recognition

Project for Mathematics of Neural Networks 2MMA80, CNN for the classification task Dataset_sign_language.

Grading

DUE DATE 18th of January. The project is graded /10 with points given for:

Item	Points
Correctly set up the training process with split dataset, batched SGD, etc.	2.0
Define and train a neural network that can classify the given images	2.0
Document every step you take, others should understand what is happening from the text/comments	3.0
Describe and perform an experiment of your choice, report your results and draw conclusions	3.0

Extras:

Choose one of these:

Vary the width/depth
Compare different activation functions
Try data augmentation
Compare different SGD variants
Try different loss functions
Compare different initialization schemes (such as found in torch.nn.init)
Use the full sized color images and compare the difference with the grayscale images
Can the network recognized rotated images?
Analyse the evolution of the parameters during training.

Dependencies

To run the code we need to install the python Virtual Environment with all the dependencies.
Install venv on your machine:

sudo apt update && sudo apt upgrade
sudo install python3-venv

Then clone the repo and run the installation script:

git clone https://github.com/kativenOG/mnn_project
cd  mnn_project/deps
bash installer.sh

In a new terminal just run:

mnn
python3 main.py

Or run the code through the notebook using VSCode (or any other iPython environment):

mnn
code project.ipynb

To run the project with a specific set of params, just add the file (ex: params.json) to the directory and pas it as a inline argument to the CLI version:

mnn
python3 main.py params.json

Todo:

Add initialization scheme;
Choose one or more Extras;
Add more metrics to the Test;
Write the project Notebook explaining every step.