DeepNeuralNetwork

Simple L-layer neural network implementation for binary classification.

Repository contents

• DeepNeuralNetwork.py — core implementation: initialization, forward/backward passes, training (L_layer_model) and predict.
• train_and_test.py — script to load HDF5 datasets from datasets/, train the model, save parameters to parameters.npz, and evaluate train/test accuracy.
• predict_from_test.py — loads saved parameters.npz and runs prediction on the test set.
• datasets/ — contains train_catvnoncat.h5 and test_catvnoncat.h5 used by the example scripts.

Requirements

• Python 3.8+
• numpy
• matplotlib
• h5py

Install dependencies:

python3 -m pip install --user numpy matplotlib h5py

Quickstart — Train and evaluate

1. Ensure the HDF5 datasets are in datasets/ and named train_catvnoncat.h5 and test_catvnoncat.h5.
2. Run the training script:

python3 train_and_test.py

This will:

• Load and preprocess the datasets (flatten images and scale by 255).
• Train an L-layer network with default architecture [n_x, 20, 7, 5, 1].
• Save trained parameters to parameters.npz.
• Print train and test accuracies.

Quickstart — Predict using saved parameters

After training (or if you already have parameters.npz), run:

python3 predict_from_test.py

This loads parameters.npz, preprocesses the test set, and prints accuracy.

Files to edit

• Edit train_and_test.py to change layers_dims, learning_rate, or number of iterations.
• Edit DeepNeuralNetwork.py to modify the architecture or add utilities.

Notes

• The scripts assume binary labels shaped (1, m).
• If your HDF5 keys differ, update the loader functions in train_and_test.py and predict_from_test.py.
• For larger datasets, consider minibatch gradient descent and checkpointing.

If you want, I can make the scripts accept command-line arguments for dataset paths and hyperparameters.

Here is training plot of the neural network for 2400 iteration for model size of [12288, 20, 7, 5, 1]