Feed Forward Neural Networks using NumPy

This library is a modification of my previous one. Click Here to check my previous library.

Installation

$ [sudo] pip3 install nicenet

Development Installation

$ git clone https://github.com/Subhash3/Neural_Net_Using_NumPy.git

Usage

>>> from nicenet import NeuralNetwork

Creating a Neural Network

inputs = 2
outputs = 1
network = NeuralNetwork(inputs, outputs, cost="mse")

# Add 2 hidden layers with 16 neurons each and activation function 'tanh'
network.add_layer(16, activation_function="tanh") 
network.add_layer(16, activation_function="tanh")

# Finish the neural network by adding the output layer with sigmoid activation function.
network.compile(activation_function="sigmoid")

Building a dataset

The package contains a Dataset class to create a dataset.

>>> from nicenet import Dataset

Make sure you have inputs and target values in seperate files in csv format.

input_file = "inputs.csv"
target_file = "targets.csv"

# Create a dataset object with the same inputs and outputs defined for the network.
dataset_handler = Dataset(inputs, outputs)
dataset_handler.make_dataset(input_file, target_file)
data, size = dataset_handler.get_raw_data()

If you want to manually make a dataset, follow these rules:

• Dataset must be a list of data samples.
• A data sample is a tuple containing inputs and target values.
• Input and target values are column vector of size (inputs x 1) and (outputs x 1) respectively.

For eg, a typical XOR data set looks something like :

>>> XOR_data = [
    (
        np.array([[0], [0]]),
        np.array([[0]])
    ),
    (
        np.array([[0], [1]]),
        np.array([[1]])
    ),
    (
        np.array([[1], [0]]),
        np.array([[1]])
    ),
    (
        np.array([[1], [1]]),
        np.array([[0]])
    )
]
>>> size = 4

Training The network

The library provides a Train function which accepts the dataset, dataset size, and two optional parameters epochs, and logging.

def Train(self, dataset: T_Dataset, size, epochs=100, logging=False, epoch_logging=True, prediction_evaulator=None):
	....
	....

For Eg: If you want to train your network for 1000 epochs.

>>> network.Train(data, size, epochs=1000)

Notice that I didn't change the value of logging as I want the output to be printed for each epoch.

Debugging

Plot a nice epoch vs error graph

>>> network.epoch_vs_error()

Know how well the model performed.

>>> network.evaluate()

To take a look at all the layers' info

>>> network.display()

Sometimes, learning rate might have to be altered for better convergence.

>>> network.set_learning_rate(0.1)

Exporting Model

You can export a trained model to a json file which can be loaded and used for predictions in the future.

filename = "model.json"
network.export_model(filename)

Load Model

To load a model from an exported model (json) file. load_model is a static function, so you must not call this on a NeuralNetwork object!.

filename = "model.json"
network = NeuralNetwork.load_model(filename)

Todo

- [x] Generalize the gradient descent algorithm
    - [x] Generalise the loss function => Write a separate class for it!
- [x] Implement Cross Entropy Loss
- [ ] Data scaling
    - [x] Min Max scaler
    - [ ] Data Standardization
- [x] Change the datasample type to a tuple instead of a list.
- [x] Show Progress bar if epoch_logging is False
- [x] Use a function as a parameter to Train method to compare predictions and actual targets.
- [x] convert all camel-cased vars to snake-case.

- [ ] API docs
    - [x] Add doc strings to all functions.
    - [x] Make the class/function declarations' docs collapsable.
    - [ ] Merge API md files and embed them in Readme.
    - [ ] Create a section, API, in README to provide documentation for all prototypes.

- [ ] Implement Batch Training
- [ ] Write a separate class for Scalers as the scaling methods increase.
- [ ] Linear and Relu activation functions
- [ ] Ability to perform regression
- [ ] Separate out outputlayer from other layers. => Create a separate class for output layer which inherits Layer.


- [ ] Convolution Nets
- [ ] Recurrent Nets