This repository contains the code used for the experiments detailed in a forthcoming paper. The paper is available pre-published at arXiv: http://arxiv.org/abs/2001.09136

To train models (python/training):

• TensorFlow (see http://www.tensorflow.org)
• NumPy (see http://www.numpy.org)
• OpenCV (see http://opencv.org)
• At least one GPU

To produce GIFs of the MNIST evaluation digits (python/etc/produce_MNIST_eval_digits.py):

• Pillow (see https://python-pillow.org/)

To extract the scalars from the tensorflow events.out.tfevents file created during training into CSV formatted data (python/etc/extract_scalars_from_logs.py):

• Pandas (see https://pandas.pydata.org)

To evaluate all possible combinations of ensembles of models (C++):

• The NVIDIA CUDA Toolkit (see https://developer.nvidia.com/cuda-toolkit)

...more coming soon...