Can allocation be optional for XGDMatrixCreateFromMat

#1

My use case is latency sensitive and allocating in my prediction path is not an option. Looking at the C code for XGDMatrixCreateFromMat it appears it allocates with is surprising to me. Can we make allocation optional? Is there already a way to satisfy my use-case?

image
image703×627 57.5 KB

#2

After digging a little deeper it appears we allocate at least 3 times. Interesting… Once for the shared_ptr? Once for the object the shared_ptr points to and an unknown number of times from constructing & populating a vector.

#3

You can use in-place prediction via XGBoosterPredictFromDense(). Take a look at the example at https://github.com/dmlc/xgboost/blob/118192f11678aeaab510eaa725e815e4e706279e/demo/c-api/inference/inference.c#L172-L202

#4

Hi,

I have tried to follow the example but haven’t been able to get it working. Here is my code along with the error message.

image
image1233×684 143 KB

#5

You need to first build the array interface, which is a JSON string. Take a look at Matrix_ArrayInterface() from the linked example.