Reproducing baseline results

[dylanponsford]

Hi, are you still planning to release the code you used to used to run the baseline methods? I am interested in reproducing some of the results you have in your paper (specifically, for GOGGLE and TabDDPM). Many thanks!

[juntongshi48]

Hi dylanxp,

In our paper, the results of the other baseline methods are either taken from the TabSyn paper [https://arxiv.org/abs/2310.09656] or evaluated using their codebase [https://github.com/amazon-science/tabsyn].

At the current moment, I only have synthetic data generated by Stasy, Tabddpm, and Codi on Adult and Diabetes on my end. Please access them with the following link:
https://drive.google.com/drive/folders/1R2M2EMIY_v2009yYjYC08g1JUDP8gS7u?usp=sharing.

Those of other methods and datasets should be fairly quick to obtain by following the instructions in the TabSyn codebase.

Hope this helps!
Juntong

[dylanponsford]

Hi @juntongshi48 , thanks for your quick reply and advice!

So just to check, there were no changes made to the code for running the baselines? I am just curious because when I use the TabSyn implementation, GOGGLE goes OOM on my GPU for diabetes, which has more memory than you reported using for yours (48GB vs 16GB?) - but your paper still reports metric results for GOGGLE trained on diabetes. I was just wondering if any adaptions were made in this case?

Many thanks,
Dylan