The datetime transformers don't give me back the same dtype sometimes

[npatki]

Environment Details

• RDT version: 1.5.0
• Python version: Any
• Operating System: Any

Error Description

In some rare cases, I may have a datetime column that is represented as a float or integer dtype. When this happens, I will always have a datetime_format string that specifies how to read this.

Example:

• A column of all nan values
• A column with format '%Y%m%d' such as [20120212, 20100321, 20091102, ...]

If this happens, then the output dtype (on the reverse_transform) is not the same dtype as the input.

Proposed Fix

On the fit:

• Store the original _dtype of the input column

On the transform:

1. Convert it to a string (astype("string") to make sure the NaNs are properly converted)
2. Parse to pd.datetime to using the provided datetime_format
3. Convert from pd.datetime -> unix timestamp

On the reverse_transform:

1. Convert from unix timestamp -> pd.datetime
2. Convert it to a string (using datetime_format)
3. Convert back to the original _dtype