def get_hvp(v):
"""Get the Hessian-vector product of v and Hessian in loss_function.
Args:
v (vector): a (n * p,)-shaped vector we want to multiply with H.
Returns:
hvp (vector): a (n, p)-shaped matrix representing Hv.
"""
v = tf.reshape(v, concat_vec.shape)
v = tensor_utils.reshape_vector_as([el[0] for el in vec], v)
v_hvp = calculate_influence.hvp(v,
itr,
loss_function,
gradient_function,
map_gradient_function,
n_samples=approx_params['hvp_samples'])
if approx_params['squared']:
v_hvp = calculate_influence.hvp(
v_hvp,
itr,
loss_function,
gradient_function,
map_gradient_function,
n_samples=approx_params['hvp_samples'])
return tensor_utils.flat_concat(v_hvp)
def hessian_vector_product(vec):
"""Takes the Hessian-vector product (HVP) of vec with model.
Args:
vec (vector): a (possibly batched) vector.
Returns:
flat_v_hvp: a (possibly batched) HVP of H(model) * vec.
"""
weight_shaped_vector = tensor_utils.reshape_vector_as(model.weights, vec)
v_hvp_total = 0.
for _ in range(n_samples):
v_hvp = ci.hvp(weight_shaped_vector, itr,
loss_fn, grad_fn, map_grad_fn)
flat_v_hvp = tensor_utils.flat_concat(v_hvp)
v_hvp_total += flat_v_hvp
return v_hvp_total / float(n_samples)