Implement and train a neural network from scratch in Python without using Tensorflow or PyTorch. The network is designed for the MNIST dataset. It optimizes the weights and offsets of the network using SGD (Stochastic Gradient Descent). Depending on the number of iterations and other network parameters, it can achieve 97-98% accuracy on the test dataset.

An object called 'NN' represent the neural network model and its parameters, including the weights and offsets of the first and second layers, the input size, the hidden layer size, the output size, and the activation function. The weights and offsets are initially random, and then stochastic gradient descent is used to train the network and optimize its parameters. The step size is set to 0.01.

To train a neural network specified by the provided configuration file, execute the following command in terminal:

python3 ./Neural_Network.py <"configfilename.cfg"> <num_iterations>

The first layer's weights and offsets are saved in "first_layer.txt".

The second layer's weights and offsets are saved in "second_layer.txt".

The package is uploaded on Test PyPi. The url is https://test.pypi.org/project/neural-net-challenge/.