Questions about data generation

[Yanboliu123]

I used the code given on the official website for testing, as follows:

import numpy as np
import torch
from sklearn.datasets import load_breast_cancer
from sklearn.model_selection import train_test_split
from tabpfn_extensions import TabPFNClassifier, TabPFNRegressor
from tabpfn_extensions import unsupervised

df = load_breast_cancer(return_X_y=False)
X, y = df["data"], df["target"]
attribute_names = df["feature_names"]

X_train, X_test, y_train, y_test = train_test_split(
X, y, test_size=0.5, random_state=42
)

clf = TabPFNClassifier(n_estimators=3, device='cuda')
reg = TabPFNRegressor(n_estimators=3, device='cuda')
model_unsupervised = unsupervised.TabPFNUnsupervisedModel(
tabpfn_clf=clf, tabpfn_reg=reg
)

feature_indices = [0, 1]

exp_synthetic = unsupervised.experiments.GenerateSyntheticDataExperiment(
task_type="unsupervised"
)

X_tensor = torch.tensor(X_train, dtype=torch.float32).reshape(-1, X_train.shape[1])
y_tensor = torch.tensor(y_train, dtype=torch.float32)

results = exp_synthetic.run(
tabpfn=model_unsupervised,
X=X_tensor,
y=y_tensor,
attribute_names=attribute_names,
temp=0.8,
n_samples=X_train.shape[0] * 3,
indices=feature_indices,
)

if results is None:
raise ValueError("Build failed, please check the library version or parameters!")

After I finished running it, I found that there was an error after running it, that is, there was no new data generated in the "results". I would like to ask you, if I want to save new data as a csv file, how do I do it?

[LeoGrin]

Thanks for the report, and sorry for the late reply. The tabpfn-extensions package has been recently refactored, could you try again with the new 0.1 version (available from source here: https://github.com/PriorLabs/tabpfn-extensions)? Thanks!

[sebhaan]

Just wanted to mention another option - TabPFGen, which is built on TabPFN too and focuses on synthetic data generation with some visualization. Might be worth a look if you're still having trouble.

[PangLuo]

I've taken a look at the code src/tabpfn_extensions/unsupervised/experiments.py from https://github.com/PriorLabs/tabpfn-extensions.

The returned value of the run method is always None. You can access exp_synthetic.data_real, exp_synthetic.data_synthetic and exp_synthetic.data to get the real data, synthetic data and the combination respectively (see the code below).

I suggest updating the Data Generation of the notebook, referred to as the Interactive Colab Tutorial in the README, to avoid confusion.

            if self.data_real.shape[0] < self.data_synthetic.shape[0]:
                self.data_real = self.data_real.sample(
                    n=self.data_synthetic.shape[0],
                    replace=True,
                )
            elif self.data_synthetic.shape[0] < self.data_real.shape[0]:
                self.data_synthetic = self.data_synthetic.sample(
                    n=self.data_real.shape[0],
                    replace=True,
                )
            self.data = pd.concat([self.data_real, self.data_synthetic])

[noahho]

Done, thank you!