def body_fn(i, permutation):
j = swaps[..., i]
iotas = np.ix_(*(lax.iota(np.int32, b) for b in batch_dims))
x = permutation[..., i]
y = permutation[iotas + (j, )]
permutation = ops.index_update(permutation, ops.index[..., i], y)
return ops.index_update(permutation, ops.index[iotas + (j, )], x)
def _lu_pivots_body_fn(i, permutation_and_swaps):
permutation, swaps = permutation_and_swaps
batch_dims = swaps.shape[:-1]
j = swaps[..., i]
iotas = np.ix_(*(lax.iota(np.int32, b) for b in batch_dims))
x = permutation[..., i]
y = permutation[iotas + (j, )]
permutation = ops.index_update(permutation, ops.index[..., i], y)
return ops.index_update(permutation, ops.index[iotas + (j, )], x), swaps
def _lu_jvp_rule(primals, tangents):
a, = primals
a_dot, = tangents
lu, pivots = lu_p.bind(a)
if a_dot is ad_util.zero:
return (core.pack(
(lu, pivots)), ad.TangentTuple((ad_util.zero, ad_util.zero)))
a_shape = np.shape(a)
m, n = a_shape[-2:]
dtype = lax.dtype(a)
k = min(m, n)
permutation = lu_pivots_to_permutation(pivots, m)
batch_dims = a_shape[:-2]
iotas = np.ix_(*(lax.iota(np.int32, b) for b in batch_dims + (1, )))
x = a_dot[iotas[:-1] + (permutation, slice(None))]
# Differentiation of Matrix Functionals Using Triangular Factorization
# F. R. De Hoog, R. S. Anderssen, and M. A. Lukas
#
# LU = A
# ==> L'U + LU' = A'
# ==> inv(L) . L' + U' . inv(U) = inv(L) A' inv(U)
# ==> L' = L . tril(inv(L) . A' . inv(U), -1)
# U' = triu(inv(L) . A' . inv(U)) . U
ndims = len(a_shape)
l_padding = [(0, 0, 0)] * ndims
l_padding[-1] = (0, m - k, 0)
zero = np._constant_like(lu, 0)
l = lax.pad(np.tril(lu[..., :, :k], -1), zero, l_padding)
l = l + np.eye(m, m, dtype=dtype)
u_eye = lax.pad(np.eye(n - k, n - k, dtype=dtype), zero,
((k, 0, 0), (k, 0, 0)))
u_padding = [(0, 0, 0)] * ndims
u_padding[-2] = (0, n - k, 0)
u = lax.pad(np.triu(lu[..., :k, :]), zero, u_padding) + u_eye
la = triangular_solve(l,
x,
left_side=True,
transpose_a=False,
lower=True,
unit_diagonal=True)
lau = triangular_solve(u,
la,
left_side=False,
transpose_a=False,
lower=False)
l_dot = np.matmul(l, np.tril(lau, -1))
u_dot = np.matmul(np.triu(lau), u)
lu_dot = l_dot + u_dot
return (lu, pivots), (lu_dot, ad_util.zero)