PATECTGAN performs poorly on categorical

[joshua-oss]

PATECTGAN performs poorly with categorical values, and seems to have been broken since at least 0.2.1. Continuous values work OK in PATECTGAN, and categorical and continuous both work OK with DPCTGAN, so this doesn't appear to be a bug in the conditional vector sampler.

repro code:

import pandas as pd
import numpy as np
import seaborn as sns
import matplotlib.pyplot as plt
from snsynth.pytorch.nn import PATECTGAN

size = 10000
eps = 3.0

# Two columns with 5 categories each, all mass is on diagonal
np_data_xy = (
    np.array([np.arange(0, size) % 5, (np.arange(0, size) % 5) * 10]).astype(np.int16).T
)
df = pd.DataFrame(np_data_xy, columns=["x", "y"])
sns.scatterplot(data=df, x="x", y="y")
plt.show()

# train and synthesize
dpgan = PATECTGAN(epsilon=eps)
dpgan.train(np_data_xy, categorical_columns=[0, 1])
synth_data = dpgan.generate(size)
synth_df = pd.DataFrame(synth_data, columns=["x", "y"])

sns.scatterplot(data=synth_df, x="x", y="y")
sns.kdeplot(data=synth_df, x="x", y="y", levels=5, alpha=0.5, fill=True)
plt.show()

expected: density plot with most mass on the diagonal

observed: density plot with mass evenly spread across all combinations