Custom Deep Learning

• Create a customized Feedforward Neural Network
  • Available options:
    • Weight initialization: Random, Xavier, He
    • Activation functions: Identity, Sigmoid, Softmax, Tanh, ReLU
    • Loss functions: MSE, Cross Entropy
    • Optimizers: GD, Momentum based GD, Nesterov accerelated GD
    • Learning mode: online, mini-batch, batch
• Refer to the documentation of any class/method by using help(class/method) Eg: help(FNN), help(FNN.compile)
• For a high-level overview of the underlying theory refer:
  • Feedforward Neural Network
  • Optimizers

Installation

$ [sudo] pip3 install customdl

Development Installation

$ git clone https://github.com/Taarak9/Custom-DL.git

Usage

>>> from customdl import FNN

Handwritten Digit Recognition example

import numpy as np
from matplotlib import pyplot as plt
from mnist_loader import load_data_wrapper 
from customdl import FNN

# MNIST data split
training_data, validation_data, test_data = load_data_wrapper()

# Loss function: Cross Entropy
hdr = FNN(784, "ce")
hdr.add_layer(80, "sigmoid")
hdr.add_layer(10, "sigmoid")
hdr.compile()
hdr.fit(training_data, validation_data)
hdr.accuracy(test_data)

The mnist_loader used could be found here.

Features to be added

• Plots for monitoring loss and accuracy over epochs
• Regularization techniques: L1, L2, dropout
• Optimizers: Adam, RMSProp
• RBF NN