def hessian_vector(fn=None, **kwargs):
"""
Wraps a function with an AutoDiff HessianVector instance, converting it to
a symbolic representation that returns the result with respect to either
all inputs or a subset (if specified with the 'wrt' keyword). A tuple of
the required vectors must be passed to the resulting function with the
keyword '_vectors'.
The function is compiled the first time it is called.
Use:
@gradient
def python_function(...):
return do_something()
python_function(...) # returns the gradient of python_function
Pass keywords to Gradient:
@gradient(wrt = ['x', 'y'])
def python_function(x=1, y=2):
return do_something()
"""
if callable(fn):
return HessianVector(fn, **kwargs)
else:
def hv_wrapper(pyfn):
return HessianVector(pyfn, **kwargs)
return hv_wrapper
def test_hv(self):
def fn(x):
return np.dot(x, x).sum()
x = np.ones((3, 3))
F = HessianVector(fn)
self.assertTrue(np.allclose(x * 6, F(x, vectors=x)))
self.assertTrue(np.allclose(x * 2, F(x[0], vectors=x[0])))
def hv_wrapper(pyfn):
return HessianVector(pyfn, **kwargs)
def test_hv_no_scalar(self):
def fn(x):
return np.dot(x, x)
x = np.ones((3, 3))
F = HessianVector(fn)
self.assertRaises(TypeError, F, x, vectors=x[0])
def test_hv_missing_vectors(self):
def fn(x):
return x
F = HessianVector(fn)
self.assertRaises(ValueError, F, np.array([[1, 1]]))