Converting a GPU DF to DMatrix doesn't handle uint8

[raghavmi]

Reporting a bug

Converting a GPU DF to Dmatrix throws an error if some of the columns are of type uint8.

cuDF's one_hot_encoding function automatically creates columns of type uint8. But these columns then throw an error when you try to convert it to a DMatrix.

You can reproduce this issue on the latest RAPIDS Docker image using the attached Python code .

GPU DF to DMatrix Conversion Error #1.txt

[kkraus14]

@raghavmi I don't believe we support uint dtypes yet, have you tried with an int8 or int16? Does it still show this issue?

[raghavmi]

It works when I remove the int8 columns.

How do I convert a column dtype in cuDF? I don't see an option like Pandas' astype or to_numeric.

I created this issue to capture the lack of support for uint8. Is that already being tracked elsewhere ?

Note, uint8 is the default type that cuDF's one hot encoding generates. So people are likely to run into it frequently.

There's a separate issue when you get past the uint8 issue, that I've it documented here:
#469

[kkraus14]

astype is supported you should be able to specify the dtype as int8 or int16 instead of uint8 as well

[raghavmi]

My apologies. I found the documentation for it now. I was looking for it as a function on the dataframe object itself.

I still see some weirdness. My attempt to change the dtype doesn't quite seem to work.

In my case, APPT_WEEKDAY_1 is a column of dtype uint8.

print(X_train.APPT_WEEKDAY_1.dtype)

uint8

1. I try to convert it to int8 and it looks like it worked.

X_train.APPT_WEEKDAY_1 = X_train.APPT_WEEKDAY_1.astype(np.int8)
print(X_train.APPT_WEEKDAY_1.dtype)

int8

2. But when I print all dtypes, it still shows APPT_WEEKDAY_1 has a dtype of uint8

print(X_train.dtypes)

AGE int64
DISTANCE int64
APPT_WEEKDAY_1 uint8

3. Finally, when I try to create a DMatrix, it still errors out with a key error for numpy.unit8.

[kkraus14]

Should be fixed in latest release which had numerous dtype handling fixes.