This is my try in implementing a deep neural network library from scratch.
- I used MNIST datatset of handwritten digits to test the implementation. in the main file, I tested with a network of 2 hidden layers of 8 neurons and an output layer of 10 neurons to give the result of the prediction.
- As an optimizer, I used Stochastic Gradient Descent with dividing the training data to mini batches.
- For the Activation Functions, I implemented the possibility to choose between having Sigmoid function for the whole network layers, or ReLu for hidden layers with Softmax for the output layer. But this version only works with the Sigmoid option for now.
- in main.py file, specify the size of each layer depending on your needs in a list like this :
layers_sizes = [ 28*28, 8, 8, 10]# 28*28 is the input layer, 10 is the output layer
- Load the data from any source you want ; i loaded the MNIST data :
from keras.datasets import mnist(train_X, train_y), (test_X, test_y) = mnist.load_data()
- transform your data into a list of tuples
(x,y)(datapoints) - instantiate The class Network with the layers sizes specified above
net = Network(layers_sizes)
- use the train method to train the model :
net.train(epoch, train_data, learning_rate, test_data=None, mini_batch_size=32)
To be Done.