OT-Flow: Fast and Accurate Continuous Normalizing Flows via Optimal Transport

Authors: Derek Onken, Samy Wu Fung, Xingjian Li, Lars Ruthotto9223-9232

AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental On five high-dimensional density estimation and generative modeling tasks, OT-Flow performs competitively to state-of-the-art CNFs while on average requiring one-fourth of the number of weights with an 8x speedup in training time and 24x speedup in inference.
Researcher Affiliation Academia Derek Onken,1 Samy Wu Fung,2 Xingjian Li,3 Lars Ruthotto,3,1 1 Department of Computer Science, Emory University 2 Department of Mathematics, University of California, Los Angeles 3 Department of Mathematics, Emory University donken@emory.edu, swufung@math.ucla.edu, xingjian.li@emory.edu, lruthotto@emory.edu
Pseudocode No No pseudocode or algorithm blocks are explicitly labeled or presented.
Open Source Code Yes 1Code is available at https://github.com/EmoryMLIP/OT-Flow .
Open Datasets Yes We perform density estimation on seven two-dimensional toy problems and five high-dimensional problems from real data sets. ...We compare our model s performance on real data sets (POWER, GAS, HEPMASS, MINIBOONE) from the University of California Irvine (UCI) machine learning data repository and the BSDS300 data set containing natural image patches. ...We also show OT-Flow s generative abilities on MNIST.
Dataset Splits No The model using the exact trace (OT-Flow) converges more quickly and to a lower validation loss, while its training loss has less variance (Fig. 3). ...For validation and testing, we use more time steps than for training, which allows for higher precision and a check that our discrete OT-Flow still approximates the continuous object.
Hardware Specification Yes All values are the average across three runs on a single NVIDIA TITAN X GPU with 12GB RAM.
Software Dependencies No Our Py Torch implementation1 of OT-Flow produces results of similar quality to state-of-the-art CNFs at 8x training and 24x inference speedups on average (see Numerical Experiments).
Experiment Setup No The number of time steps is a hyperparameter. ...We tune the number of training time steps so that validation and training loss are similar with low computational cost.