Build Neural Network From Scratch Without Using ML Framework(tensorflow, pytorch...)

NOTE: This project is inspired by the Keras framework and is just for study purposes.

How to run project

Quick Test

To quickly test the project, you can copy the code from the neural-network-standalone.js file and paste it into the console tab of your browser.

Full Project Setup:

1. Download mnist dataset on kaggle https://www.kaggle.com/datasets/hojjatk/mnist-dataset
2. Extract and copy these files to mnist-data folder(create new one if needed):

• t10k-images.idx3-ubyte
• t10k-labels.idx1-ubyte
• train-images.idx3-ubyte
• train-labels.idx1-ubyte

3. Install dependencies first:

npm install

4. Run project

npm start

Compatibility:

• node v16.14.0
• npm 8.3.1

Example

Here is an example of how to create and summarize a model and save it after training:

let model = new Model('categorical_crossentropy', earlyStoppingCallback, epochEndCallback, ['loss', 'progress']);

model.add(new Dense(784, 10, 'leaky_relu'));

model.add(new Dense(10, 10, 'softmax'));

model.summary();

model.train(xTrain, yTrain, 10, 0.001);

model.save('./mnist_model.any_extension');