/Mini-Deep-Learning-Framework

Primary LanguageJupyter Notebook

Mini-Deep-Learning-Framework

This is a mini deep learning framework that uses only pytorch’s tensor operations and python’s standard math library.

Prerequisties

  • pytorch
  • math
  • matplotlib

File descriptions

module.py - this is the main .py file encapsulating the key components for implementing a neural network. It contains the following classes:

  • Module - a backbone abstract class which defines the necessary methods for training a neural network.
  • Sequential - a container class of modules which processes input data sequentially. Inherits Module.
  • DenseLayer - a class implementing fully-connected layer. Inherits Module.
  • ReLU, Tanh, Sigmoid, Softmax activation functions. All of them inherit Module.

loss.py - this is a .py file implementing neural network loss functions. It contains the following classes:

  • Loss - an abstract class defining all the necessary methods that a neural network loss function should have.
  • MSE - a class implementing mean squared error loss. Inherits from Loss
  • CrossEntropy - a class implementing categorical cross-entropy loss. Inherits from Loss.

utilities.py - this is a .py file containig several utility methods such as generating the data for the given binary classification problem, computing the accuracy of the neural net, etc.

test.py - this is the main executable python file which trains and evaluates the two deep neural network models described in the report.

How to run

To run the (default) model with the MSE loss: python3 test.py
To run the model with the softmax loss: python3 test.py --loss softmax_loss