In this repository I implemented a neural network from scratch. I am very excited about this and it was verrrry eager to see it works and now it's done!

a simple implemention of a neural network.
also simple implemention of a single neuron classification and regression models.

Your data must be in this template:

X_train = [[x11, x12], [x21, x22]]
Y_train = [[y1], [y2]]

X_train is a list of data samples features. A datasample can be 1D or nD, anyway it should be in a list like I've definde above.
Y_train is also the same as X_train. It should be a list containing lists of datasample outputs each one in a list like above.

Import class Layer from neural_network.py and then create a list and define layers for the network.
To create a layer you have these options to be passed:

Layer(input_neurons, neuron_num, activation, activation_differ, w=0.1, lr=1)

Read about details on Layer docs line 53.

Then it's time to create the network or model. To create a network you should just pass the layers as a list to it:

net = NeuralNetwork(layers)

To train the model you should pass the loop of train. See a completed sample here:
By the way I used stochastic gradient descent here but I'll add mini batches later ...

for epoch in range(epochs):
    for x, y_true in zip(X_train, Y_train):
        y_pred = net(x)
        net.backward(y_true)

You can also get the loss by net.loss. This property is the loss just for one datapoint that goes into network. You can see loss for each time passing whole data batch into network:

for epoch in range(150):
    total_loss = 0  # for one batch or one time passing datapoints into network
    for x, y_true in zip(X_train, Y_train):
        y_pred = net(x)
        net.backward(y_true)
        total_loss += net.loss  # net.loss is the loss for one datapoint: (x, y_true)
        
    if epoch % 10 == 0:
        print(f"Epoch {epoch}, Training Loss: {total_loss/len(X_train):.4f}")

It will absolutely be easier for you to read and run this code if you are familier with mathmatics which is used in a neural network

• Ehsan-004

This project is open-source and does not have a specific license. Feel free to use, modify, and distribute it as you see fit.