Hypersolvers: Toward Fast Continuous-Depth Models

Authors: Michael Poli, Stefano Massaroli, Atsushi Yamashita, Hajime Asama, Jinkyoo Park

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental evaluations on standard benchmarks, such as sampling for continuous normalizing flows, reveal consistent pareto efficiency over classical numerical methods.
Researcher Affiliation Academia Michael Poli KAIST, Diff Eq ML poli_m@kaist.ac.kr Stefano Massaroli The University of Tokyo, Diff Eq ML massaroli@robot.t.u-tokyo.ac.jp Atsushi Yamashita The University of Tokyo yamashita@robot.t.u-tokyo.ac.jp Hajime Asama The University of Tokyo asama@robot.t.u-tokyo.ac.jp Jinkyoo Park KAIST jinkyoo.park@kaist.ac.kr
Pseudocode No The paper provides mathematical formulations and equations but does not include structured pseudocode or an algorithm block.
Open Source Code Yes Supporting reproducibility code is at https://github.com/Diff Eq ML/diffeqml-research/tree/master/hypersolver
Open Datasets Yes We train standard convolutional Neural ODEs with input layer augmentation (Massaroli et al., 2020b) on MNIST and CIFAR10 datasets.
Dataset Splits No The paper mentions using 'training dataset' and 'test data' but does not explicitly specify a validation set or detailed split percentages (e.g., 80/10/10) needed for reproduction.
Hardware Specification Yes The measurements presented are collected on a single V100 GPU.
Software Dependencies No The paper mentions 'Torch Dyn (Poli et al., 2020) library' and 'Py Torch (Paszke et al., 2017) module implementation' but does not specify their version numbers.
Experiment Setup Yes Following this initial optimization step, 2 layer convolutional Euler hypersolvers, Hyper Euler, (4) are trained by residual fitting (6) on 10 epochs of the training dataset with solution mesh length set to K = 10. As ground truth labels, we utilize the solutions obtained via dopri5 with absolute and relative tolerances set to 10 4 on the same data.