Practical Synthesis of Mixed-Tailed Data with Normalizing Flows
Authors: Saba Amiri, Eric Nalisnick, Adam Belloum, Sander Klous, Leon Gommans
TMLR 2024 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We support our claim with extensive experiments on synthetic and real world data. ... We experiment using a range of simulated and tabular data. ... Empirical results showing generally favorable performance compared to other tail-adaptive flow methods. |
| Researcher Affiliation | Collaboration | Saba Amiri EMAIL Informatics Institute University of Amsterdam ... Eric Nalisnick EMAIL Department of Computer Science Johns Hopkins University ... Leon Gommans EMAIL Air France KLM |
| Pseudocode | Yes | Algorithm 1: Training Algorithm for tail-adaptive normalizing flows with enhanced selective generation |
| Open Source Code | No | The paper does not provide an explicit statement about releasing source code for the methodology described, nor does it include a link to a code repository. |
| Open Datasets | Yes | We train our model on four tabular datasets from the UCI repository, namely Power, Gas, Miniboone and Hepmass. ... The POWER dataset (Hebrail & Berard, 2012)... The GAS dataset (Fonollosa et al., 2015)... The HEPMASS dataset (Baldi et al., 2016)... The MINIBOONE dataset (Roe et al., 2005)... |
| Dataset Splits | No | The paper refers to a 'test dataset' and 'synthetic dataset' when reporting metrics, but does not explicitly provide details about how the original datasets were split into training, validation, and test sets, such as percentages, sample counts, or splitting methodology. |
| Hardware Specification | Yes | Our training was done on two types of nodes depending on availability, with either one NVIDIA A40 or one A100 GPU. |
| Software Dependencies | No | The paper mentions using Adam optimization and specific types of normalizing flows (RNVP and MAF), but does not provide specific version numbers for any software dependencies or libraries used for implementation. |
| Experiment Setup | Yes | Minibatch size is 1024 and we perform Adam optimization (lr = 10 5) All trainings are over 1000 iterations. ... The shift and scale operations of RNVPs are modelled by an feed forward network with one hidden layer the width of 1024. ... Each flow model is 12 steps deep, with the MAFs having 12 autoregressive layers which use a stacked transformation with 1024 hidden units and 2 blocks to model the conditional dependencies of the variables. |