Copulas as High-Dimensional Generative Models: Vine Copula Autoencoders
Authors: Natasa Tagasovska, Damien Ackerer, Thibault Vatter
NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on MNIST, Street View House Numbers and Large-Scale Celeb Faces Attributes datasets show that VCAEs can achieve competitive results to standard baselines. |
| Researcher Affiliation | Collaboration | Natasa Tagasovska Department of Information Systems HEC Lausanne, Switzerland natasa.tagasovska@unil.ch Damien Ackerer Swissquote Bank Gland, Switzerland damien.ackerer@swissquote.ch Thibault Vatter Department of Statistics Columbia University, New York, USA thibault.vatter@columbia.edu |
| Pseudocode | Yes | pseudo-code for the VCAE algorithm can be found in Appendix B. |
| Open Source Code | Yes | We used Pythorch 4.1 [58], and we provide our code in Appendix E. |
| Open Datasets | Yes | We explore three real-world datasets: two small scale MNIST [40] and Street View House Numbers (SVNH) [57], and one large scale Celeb A [44]. |
| Dataset Splits | No | All models were trained on a separate train set, and evaluated on hold out test sets of 2000 samples, which is the evaluation size used in [82]. While train and test sets are mentioned, no explicit validation split percentages or counts are provided. |
| Hardware Specification | Yes | All experiments were executed on an AWS instance p2.xlarge with an NVIDIA K80 GPU, 4 CPUs and 61 GB of RAM. |
| Software Dependencies | Yes | We used Pythorch 4.1 [58] |
| Experiment Setup | Yes | Parameters of the optimizers and other hyperparameters are fixed as follows. Unless stated otherwise, all experiments were run with nonparametric vines and truncated after 5 trees. For all AE-based methods, we use the Adam optimizer with learning rate 0.005 and weight decay 0.001 for all the natural image experiments, and 0.001 for both parameters on MNIST. For DCGAN, we use the recommended learning rate 0.0002 and β1 = 0.5 for Adam. The size of the latent spaces z was selected depending on the dataset s size and complexity. For MNIST, we present results with z = 10, SVHN z = 20 and for Celeb A z = 100. For MNIST, we used batch size of 128, for SVHN 32, and for Celeb A batches of 100 samples for training. |