In my machine learning laboratory course, I built a basic neural network class using only NumPy, rather than deep learning libraries like Tensorflow or Keras. The class can generate multi-layered fully connected neural networks (FCNs) that can be trained and used easily. In this project, a dataset was produced by our class to train and test the networks. The dataset can't be shared for privacy reasons.
The dataset was pre-processed for the training and testing purpose using pandas and scikit-learn. Then a network is created with two layers having three and four neurons respectively. The network can't be too big because a bigger network results in exploding gradients, which was not addressed, as it was meant to be a simple implementation. The network was then trained and tested with the private dataset.