Can the data type of Grad and Hess be np.float64?

[caffery-chen]

It seems internally the Grad and Hess are converted to float32 which causes the some gradient in my application becomes Inf.

In my application, I do need grad to be quite large value at the beginning. Is there any possible way to change the underly C code to support float64 in xgboost?

I am customizing the objective function using

def gradient(truth, predt)
def hessian(truth, predt)
def my_loss(truth, predt)

Another questions: Is the output of the hess should be a vector or a matrix is also possible?

[hcho3]

@caffery-chen

Is there any possible way to change the underly C code to support float64 in xgboost?

No, we (XGBoost developers) will not make this change. Using float64 would increase memory usage. However, since XGBoost is an open-source project, you are free to make the change on your own fork.

@caffery-chen

Is the output of the hess should be a vector or a matrix is also possible?

The hess should be a vector.

[caffery-chen]

Thank you. Tried a workaround of my loss function to make it work with float32.