MNIST Exploration

This repo contains an exploration of the MNIST data set using Tensorflow.
Several factors are being explored:

• Accuracy impacts of various types of model architectures/layers including -
  • Number of layers
  • Convolutions vs dense layers
  • Dropout
  • Batch normalization
  • Different optimization functions
• The effects of gaussian noise of classification performance
• The effects of mislabeled data on the classification performance

Repo Structure

• Exploration - Contains the explanations of the project
  • answers.md - Contains the questions and answers
  • additional_exploration.md - Additional comparison and exploration of architectures/hyperparameters
  • runlog - File containing brief descriptions of all the runs for quick reference
  • Images - Contains saved images for analysis
• Models - The saved jsons/numpy arrays of run performances
• src - The scripts used in the analysis
  • pymodels - Folder containing python files, one for each model
  • datagen.py - Data generator/augmenter
  • eval.py - The core training/testing module
  • graph.py - Used to visualize the results from the training/testing
  • summarize.py - Small utility used to quickly see the hyperparameters used in each run
  • utils.py - Utilities for saving models