CUDA out of memory about ctgan HOT 6 CLOSED

It depends on the meaning of the column.

If the column is a unique identifier, it should be removed.
If the column is an ordinal column, converting categories to integers may make sense. For example, ["elementary school", "middle school", "college"] --> [1, 2, 3] is reasonable. But ["Massachusetts", "New York", "California] --> [1, 2, 3] does not make much sense.

from ctgan.

When there’s a discrete column with so many categories, it usually means that there are fewer examples for each category. So the learning task becomes much more difficult. The solution to this problem depends on the data and use case. For example, when there're sufficient data, training multiple models on subsets of the data is helpful. It's also possible to cluster the categories. In your example, you can replace cities with states so that the number of categories becomes much smaller.

from ctgan.

Hi,

This could happen when a categorical column has too many categorical values. In the first 10k rows, there are say 1000 different categories, the model is still small and can fit into the GPU memory. When there are 100k rows, there could a possibility that there are more different categories and it runs out of memory. One typical example of such a column is unique id. Unique id columns are not supported by our model.

Is there a categorical column in your data that has a large set of catagorical values?

from ctgan.

Thank you! I had a few columns that had 20k unique and 300k unique categorical values. I am removing those now.

What about taking a high dimensionality categorical column then remapping it as an integer and running the model. Then the model will learn an integer (I can round it if it comes out as float) and try and remap back

e.g.
['A,'B',...,'Z'] --> [1,2,...26]

from ctgan.

from ctgan.

Closing this issue, as it seems all questions have been answered.

from ctgan.