Hessian computation/Compatibility with dual numbers about smt

[relf]

Thanks, I was not aware of the num_dual library. I can not run the code above with the expected error. What is the type/value of query_points[0]?

Anyway seems it is not compatible with num_dual! Regarding hessians implementations, I would say we are always open to contributions.

[NAThompson]

I can not run the code above with the expected error. What is the type/value of query_points[0]?

Sorry-I introduced a typo in the reproducer while I was trying to pare it down. Fixed now.

[relf]

I had to change the code of f to (try to) make it work (missing ] , try to use item() to get a "scalar"):

    def f(x):
        y = rmtb.predict_values(np.array([[0.0, 1.0]])).item()
        print(y, type(y))
        return y

When I run it:

> python test_num_dual.py
-3.949565971416e-06 <class 'float'>
Traceback (most recent call last):
  File "D:\rlafage\test_num_dual.py", line 52, in <module>
    out = num_dual.hessian(f, query_point)
          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
TypeError: argument 'f' must return a scalar.

I do not understand this error and I do not get yours, anyway still not ready to use num_dual 😉

[NAThompson]

That looks like it may well be a bug in num_dual . . . you've made a pretty compelling case that f returns a scalar . . .