`OneHotEncoder` causing shape misalignment in `CopulaGAN`, `CTGAN`, and `TVAE`

[sarahmish]

Environment Details

Please indicate the following details about the environment in which you found the bug:

• RDT version: 0.5.2.dev0
• Python version: 3.8
• Operating System: macOS 11.5.1

Error Description

After optimizing the OneHotEncoder Transformer, CopulaGAN, CTGAN, and TVAE start failing in SDV. The primary reason for this failure is caused by the line:

https://github.com/sdv-dev/CTGAN/blob/2dbbc18770d0dc17e7c5e83da6980ace424e9326/ctgan/data_transformer.py#L57

where DataTransformer access ohe.dummies. In PR #186, we altered the contents of self.dummies to not include NaN values (because pd.Categorical doesn't accept it as a category), which inconsequence makes us initialize weights of the the wrong dimensions in consequent parts of the code because we are missing the NaN value in ohe.dummies.

Steps to reproduce

Initialize any dataframe with missing values and it will crash

import pandas as pd
from sdv.tabular import CTGAN

data = pd.DataFrame({"category": ["1", "2", "3", None] * 4})
ctgan = CTGAN()
ctgan.fit(data)

the code crashes with the following message

AssertionError                            Traceback (most recent call last)
<ipython-input-8-8847fc3cae96> in <module>
----> 1 ctgan.fit(data)

~/Downloads/repos/SDV/sdv/tabular/base.py in fit(self, data)
    140             LOGGER.debug(
    141                 'Fitting %s model to table %s', self.__class__.__name__, self._metadata.name)
--> 142             self._fit(transformed)
    143 
    144     def get_metadata(self):

~/Downloads/repos/SDV/sdv/tabular/ctgan.py in _fit(self, table_data)
     55                     categoricals.append(field)
     56 
---> 57         self._model.fit(
     58             table_data,
     59             discrete_columns=categoricals

~/opt/anaconda3/envs/rdt/lib/python3.8/site-packages/ctgan/synthesizers/ctgan.py in fit(self, train_data, discrete_columns, epochs)
    296         train_data = self._transformer.transform(train_data)
    297 
--> 298         self._data_sampler = DataSampler(
    299             train_data,
    300             self._transformer.output_info_list,

~/opt/anaconda3/envs/rdt/lib/python3.8/site-packages/ctgan/data_sampler.py in __init__(self, data, output_info, log_frequency)
     37             else:
     38                 st += sum([span_info.dim for span_info in column_info])
---> 39         assert st == data.shape[1]
     40 
     41         # Prepare an interval matrix for efficiently sample conditional vector

AssertionError: 

Steps to fix

1. revisit the variables in OneHotEncoder and let dummies contain np.nan when that's the case.
2. add a unit test to confirm np.nan is part of dummies when dummy_na is set to True.