Learning Distributions on Manifolds with Free-Form Flows

Authors: Peter Sorrenson, Felix Draxler, Armand Rousselot, Sander Hummerich, Ullrich Köthe

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We provide our code at https://github.com/vislearn/FFF. ... M-FFF consistently matches or outperforms previous single-step methods specialized to specific manifolds. It is typically two orders of magnitude faster than multi-step methods based on diffusion or flow matching, achieving better likelihoods in several experiments. ... We now demonstrate the practical performance of manifold free-form flows on various manifolds. We choose established experiments to ensure comparability with previous methods
Researcher Affiliation Academia Computer Vision and Learning Lab, Heidelberg University
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes We provide our code at https://github.com/vislearn/FFF.
Open Datasets Yes We use four established datasets compiled by Mathieu and Nickel [2020] for density estimation on S2: Volcanic eruptions [NGDC/WDS, 2022b], earthquakes [NGDC/WDS, 2022a], floods [Brakenridge, 2017] and wildfires [EOSDIS, 2020].
Dataset Splits Yes We follow previous works and use a dataset split of 80% for training, 10% for validation and 10% for testing.
Hardware Specification Yes one training run takes approximately 7 hours on a NVIDIA A40. ... Each model trained around 20h on a compute cluster using a single NVIDIA A40. ... takes 2.5 to 3 hours on a NVIDIA Ge Force RTX 2070 graphics card.
Software Dependencies No The paper mentions software packages like Py Torch, Py Torch Lightning, Numpy, Matplotlib, Pandas, and geomstats but does not provide specific version numbers for these dependencies.
Experiment Setup Yes In all cases, we selected hyperparameter using the performance on the validation data. ... (Table 9: Hyperparameter choices for the rotation experiments. Layer type Res Net, Residual blocks 2, Inner depth 5, Inner width 512, Activation Re LU, βx R 500, βz R 0, βU 10, βP 10, Latent distribution uniform, Optimizer Adam, Learning rate 5 10 3, Scheduler Exponential w/ γ = 1 10 5, Gradient clipping 1.0, Weight decay 3 10 5, Batch size 1,024, Step count 585,600, #Repetitions 3)