PARSynthesizer crashes during sample if there was an all-null column in the input data

[npatki]

Environment Details

• SDV version: 1.20.0
• Python version: 3.11
• Operating System: Linux (Colab Notebook)

Error Description

If my input data (training data) has a column that is all-null, then the PARSynthesizer crashes during sample with a KeyError. This happens whenever the sdtype of this column indicates that it should be statistically modeled -- for eg. numerical or categorical.

Steps to reproduce

import numpy as np
import pandas as pd

from sdv.sequential import PARSynthesizer
from sdv.metadata import Metadata

data = pd.DataFrame(data={
    'sequence_key': ['sequence-' + str(int(i/5)) for i in range(100)],
    'numerical_col': np.random.randint(low=0, high=100, size=100),
    'categorical_col': np.random.choice(['A', 'B', 'C'], size=100),
    'all_null_col': [np.nan]*100
})

metadata = Metadata.load_from_dict({
    'tables': {
        'table': {
            'columns': {
                'sequence_key': { 'sdtype': 'id' },
                'numerical_col': { 'sdtype': 'numerical' },
                'categorical_col': { 'sdtype': 'categorical' },
                'all_null_col': { 'sdtype': 'numerical' }
            },
            'sequence_key': 'sequence_key'
        }
    }
})

synthesizer = PARSynthesizer(metadata, epochs=1)
synthesizer.fit(data)
synthesizer.sample(num_sequences=2)

KeyError: "['all_null_col'] not in index"

Stack trace provided below.

stack_trace.txt

Expected Behavior

I expect that -- no matter what the sdtype of the original column -- PARSynthesizer should be able to understand that the column is all-null and therefore produce synthetic data where the column is all-null.

This is the behavior for all our other single-table and multi-table synthesizers.

Workaround

Until we fix this bug, the recommended workaround is to replace the missing values in the column with a static, numerical value such as 0.

data['all_null_col'] = data['all_null_col'].fillna(0)

Now PARSynthesizer will no longer crash. The synthetic data column will contain a static value (0) that you can change back to null if you'd like.