Skip to content

Latest commit

 

History

History
39 lines (22 loc) · 2.21 KB

File metadata and controls

39 lines (22 loc) · 2.21 KB

AutoEncoder-and-Classifier-of-MNIST-images

We aim to build different Neural Networks in order to achieve 2 goals:

  1. AutoEncoder's implementation.
  2. Classifier's implementation.

Auto-Encoder is made up of one encoder and one decoder. User will be asked to give hyperparameters in order to build different Auto-Encoders and by observing their results he can save model which performed better.

Classifier will be made up of the half part of previous Auto-Encoder, more specifically only 'Encoder' part, and then a fully connected layer. With same mindset user has the freedom to build several models and plot his experiment's results or save whichever model he prefers.

Compilation & Execution

  • For AutoEncoder: python src/autoencoder.py -d data/train/train-images-idx3-ubyte

  • For Classifier: python src/classification.py -d data/train/train-images-idx3-ubyte -dl data/train/train-labels-idx1-ubyte -t data/test/t10k-images-idx3-ubyte -tl data/test/t10k-labels-idx1-ubyte -model dropoutautoencoder.h5

Results

From directory 'results_of_experiments' we are going display plots, loss vs epochs with specific hyperparameters after a series of experiments (tuning), from our best models. Note that inside this directory we saved some pretty good models for AutoEncoder, but also for Classifier.

  1. Firstly for Auto-Encoder:

alt text

  1. Secondly for Classifier:

alt text

  • Correct Labels

alt text

  • InCorrect Labels

alt text

Contributors

  1. Alexandra Apostolopoulou
  2. Spyros Briakos