Skielex / 02456-deep-learning-project

Repository for project in course 02456 Deep Learning at DTU, fall 2017, by Niels Jeppesen (niejep).

Un- and Semi-supervised Learning for Ultrasonic Images with Convolutional Variational Autoencoders

Repository for project in course 02456 Deep Learning at DTU, fall 2017, by Niels Jeppesen (niejep).

Prerequisites

• python 3.5+
• keras 2.0+
• tensorflow 1.3+
• scikit-learn 0.19+

Files

• autoencoder.py implements a Variational Autoencoder and a Semi-superviser Variational Autoencoder/Classifier in Keras/Tensorflow.
• Unsupervivised VAE.ipynb is a Jupyter Notebook for recreating unsupervised MNIST result from the paper.
• Semi-supervised VAE.ipynb is a Jupyter Notebook for recreating semi-supervised MNIST result from the paper.

Training

It is possible to train the implementations in the notebooks by setting TRAIN = True in the beginning of the notebook. By default, training plot are saved to the output folder using Tensorboard summaries, and models are saved to the same folder using Keras.

Loading models

Existing models can be loaded by setting LOAD_LAST_WEIGHTS = True. If you don't want to train remember to set TRAIN = False. Models are loaded from the folder specified by the input_dir variable, which by default points to the output folder. You can either load models you've trained yourself, or download pre-trained model listed below.

Pre-trained models

To avoid having to train all the different models in the notebook the following is provided. Simply extract the output folder in the repository root and load the models as described above.

• Un- and semi-supservised models and training data - 20 epochs
• Un- and semi-supservised models and training data - 50 epochs