Framework

This repository contains the implementation of a framework for a neural net developed during a project at the University of Würzburg. It has much less features than for example PyTorch or TensorFlow, but it was still able to achieve good results as a slideshow controller for a presentation using hand gestures.
The framework can be be used for creating neural networks and training them.

Installation using pip

To install the latest version of this package and all its dependencies simply run:

pip install git+https://github.com/M-Lampert/ML-Framework.git

Structure

• example_data\: Example training and testing data for demonstration purposes
• framework\: The program logic
  • init.py: Necessary for the installation via pip (see Installation using pip)
  • activation_layers.py: Activations layers
    • Sigmoid
    • ReLU
    • Tanh
    • Softmax
  • hyperparameter_search.py: Method to execute a random hyperparameter search
  • input_layer.py: Contains special types of input layers
    • DefaultInputLayer: Converts a column vector of shape (x) saved as numpy array to a row vector of shape (1, x) saved using the internal Tensor class
    • PictureInputLayer: Converts a two-dimensional (gray-scale) image saved as numpy array to a row vector of shape (1, x) saved using the internal Tensor class
  • layers.py: Typical layers that are no activation layers
    • FullyConnectedLayer
    • FlattenLayer
  • losses.py: Typical loss functions
    • MeanSquaredError
    • CrossEntropy
  • network.py: Network class with method for forward-, backward-pass, saving, loading and a convenient prediction method
  • optimizer.py: Class for the training algorithm stochastic gradient descent
  • pca.py: Class for Principal Component Analysis. Contains methods to analyze the given data and save, load and transform functionalities to use inside a preprocessing pipeline
  • tensor.py: Tensor class used inside the neural network for computations. Conveniently stores values and if needed their gradients
  • utils.py: Methods to get different performance scores and transformation methods between vectors and class labels
• Demo.ipynb: A demo on how to use the essential features of the framework
• makefile: Convenient commands for development. Type make help for more information. (Does not work natively on Windows)
• Demo_PCA.ipynb: Demonstration of the PCA functionality
• requirements.txt: All packages that are required to run the framework. Can be installed with pip install -r requirements.txt
• setup.py: Necessary for the installation via pip (see Installation using pip)
• test_utils.py: Precomputed values used in test.py
• test.py: Unit-tests testing the correctness of essential features