This repository contains the code for the paper "Accurate and Diverse Sampling of Sequences based on a “Best of Many” Sample Objective" (https://arxiv.org/abs/1806.07772). Currently only the code corresponding only to the experiments with the MNIST Sequence Dataset is available. Will be expanded in the near future.
The code was tested on,
- Tensorflow - 1.3.0
- Keras - 2.1.5
Aditional requirements,
- Matplotlib for plotting.
- Tqdm for progress bars.
The MNIST Sequence dataset can be found at: (https://edwin-de-jong.github.io/blog/mnist-sequence-data/). Please download and extract the sequences and place them under ./data/MNIST/sequences/.
To train a CGM model using the "Best of Many" objective use,
python mnist.py
The (negative) CLL and the KL divergence between the true and learned latent distributions is printed.
To train without plotting use,
python mnist.py --plot False
@inproceedings{apratim18cvpr2,
title = {Accurate and Diverse Sampling of Sequences based on a {\textquotedblleft}Best of Many{\textquotedblright} Sample Objective},
author = {Bhattacharyya, Apratim and Fritz, Mario and Schiele, Bernt},
year = {2018},
booktitle = {31st IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2018)},
}