This fix is intended to solve bug of binning:
There are times in which the float_values are NaN which in turn cause the gradient to become NaN (see log: NAN prediction in example X, forcing 0.0).
So in order to avoid it, we define the src features values (val1, val2) to be zero and then reassign them to be equal to float_values if its not NaN.
This fix is intended to solve bug of binning: There are times in which the float_values are NaN which in turn cause the gradient to become NaN (see log: NAN prediction in example X, forcing 0.0). So in order to avoid it, we define the src features values (val1, val2) to be zero and then reassign them to be equal to float_values if its not NaN.