Add the `TabDDPM` generator algorithm to plugins about synthcity

[tztsai]

Hello! I am attempting to add the TabDDPM generator to plugins by adapting the code in https://github.com/rotot0/tab-ddpm. I have forked the repository and added a branch tab_ddpm. I put the plugin in generic/plugin_ddpm.py and the DDPM model in core/models/tabular_ddpm. This is only a quick and dirty attempt, and I will make further updates soon.

[bcebere]

Hello @tztsai

Thank you for working on this.

Regarding your questions:

1. You can extend the TabularEncoder with a parameter for the continuous variables, to switch between the QuantileTransformer and BayesianGMM. It would be better to follow the original paper, and maybe other models will need to encode using QuantileTransformer in the future.
2. Sure. You can assume by default, that the labels are the cond for your model, but give the option to the power users to provide custom conditionals, as long as they understand what is happening. If cond is none, you use the labels in the dataset

[bcebere]

Hello @tztsai

Thank you for working on this. Please raise a PR, and make sure you added tests for your changes.

At the same time, there is an important problem with https://github.com/rotot0/tab-ddpm .
It has no LICENSE next to the code, meaning we cannot directly integrate or copy that code into our codebase.

So, for a successful PR, you must reimplement the relevant parts(and drop the configs, requirements etc.),, before adding them to the synthcity codebase.

Please let us know if you have any questions.

cc @robsdavis

[tztsai]

Hello @bcebere

Thank you for your reply! I am reimplementing it now. I have two questions:

1. Is it always required to use the TabularEncoder to preprocess the input data? In the paper of TabDDPM, the numerical features are preprocessed by a Gaussian quantile transformer and the categorical features by a one-hot encoder. Although the transformation of categorical features is the same, the TabularEncoder uses BayesianGMM to transform the numerical features. Is it OK to directly feed the data encoded by the TabularEncoder into the diffusion model and ignore the original preprocessing procedures in the original paper?
2. In the original paper, for a classification task, the conditional probability distributions $P(x_t\mid x_{t-1}, y),t=1,\ldots,T$ are learned, where $y$ is the label. To implement the fit method of a plugin for a classification task, do we provide the labels as the keyword argument cond?

[tztsai]

Hello,

I found there is no log uniform distribution of float numbers when I was implementing the search space of TabDDPM. Could it be added in the framework?

[tztsai]

Hello @bcebere @gsel9

I think I have finished the reimplementation of TabDDPM. The current version supports both classification and regression, and supports using no conditionals, using the target as conditionals, and using user-provided conditionals to train and generate data. Since all tests were passed, I have created a PR. Please check that out and do not hesitate to inform me if there are any problems.