def matrix_rank(a, tol=None, validate_args=False, name=None):
"""Compute the matrix rank; the number of non-zero SVD singular values.
Arguments:
a: (Batch of) `float`-like matrix-shaped `Tensor`(s) which are to be
pseudo-inverted.
tol: Threshold below which the singular value is counted as 'zero'.
Default value: `None` (i.e., `eps * max(rows, cols) * max(singular_val)`).
validate_args: When `True`, additional assertions might be embedded in the
graph.
Default value: `False` (i.e., no graph assertions are added).
name: Python `str` prefixed to ops created by this function.
Default value: 'matrix_rank'.
Returns:
matrix_rank: (Batch of) `int32` scalars representing the number of non-zero
singular values.
"""
with tf.name_scope(name or 'matrix_rank'):
a = tf.convert_to_tensor(a, dtype_hint=tf.float32, name='a')
assertions = _maybe_validate_matrix(a, validate_args)
if assertions:
with tf.control_dependencies(assertions):
a = tf.identity(a)
s = tf.linalg.svd(a, compute_uv=False)
if tol is None:
if tensorshape_util.is_fully_defined(a.shape[-2:]):
m = np.max(a.shape[-2:].as_list())
else:
m = tf.reduce_max(tf.shape(a)[-2:])
eps = np.finfo(dtype_util.as_numpy_dtype(a.dtype)).eps
tol = (eps * tf.cast(m, a.dtype) *
tf.reduce_max(s, axis=-1, keepdims=True))
return tf.reduce_sum(tf.cast(s > tol, tf.int32), axis=-1)
def _entropy(self):
if self._logits is None:
# If we only have probs, there's not much we can do to ensure numerical
# precision.
probs = tf.convert_to_tensor(self._probs)
return -tf.reduce_sum(
tf.math.multiply_no_nan(tf.math.log(probs), probs), axis=-1)
# The following result can be derived as follows. Write log(p[i]) as:
# s[i]-m-lse(s[i]-m) where m=max(s), then you have:
# sum_i exp(s[i]-m-lse(s-m)) (s[i] - m - lse(s-m))
# = -m - lse(s-m) + sum_i s[i] exp(s[i]-m-lse(s-m))
# = -m - lse(s-m) + (1/exp(lse(s-m))) sum_i s[i] exp(s[i]-m)
# = -m - lse(s-m) + (1/sumexp(s-m)) sum_i s[i] exp(s[i]-m)
# Write x[i]=s[i]-m then you have:
# = -m - lse(x) + (1/sum_exp(x)) sum_i s[i] exp(x[i])
# Negating all of this result is the Shanon (discrete) entropy.
logits = tf.convert_to_tensor(self._logits)
m = tf.reduce_max(logits, axis=-1, keepdims=True)
x = logits - m
lse_logits = m[..., 0] + tf.reduce_logsumexp(x, axis=-1)
sum_exp_x = tf.reduce_sum(tf.math.exp(x), axis=-1)
return lse_logits - tf.reduce_sum(tf.math.multiply_no_nan(
logits, tf.math.exp(x)),
axis=-1) / sum_exp_x
def cond(m, pchol, perm, matrix_diag):
"""Condition for `tf.while_loop` continuation."""
del pchol
del perm
error = tf.linalg.norm(matrix_diag, ord=1, axis=-1)
max_err = tf.reduce_max(error / orig_error)
return (m < max_rank) & (tf.equal(m, 0) | (max_err > diag_rtol))
def forward_step(previous_step_pair,
log_prob_observation):
log_prob_previous = previous_step_pair[0]
log_prob = (
log_prob_previous[..., tf.newaxis] +
self._log_trans +
log_prob_observation[..., tf.newaxis, :])
most_likely_given_successor = tf.argmax(log_prob,
axis=-2)
max_log_p_given_successor = tf.reduce_max(
input_tensor=log_prob, axis=-2)
return (max_log_p_given_successor,
most_likely_given_successor)
def pinv(a, rcond=None, validate_args=False, name=None):
"""Compute the Moore-Penrose pseudo-inverse of a matrix.
Calculate the [generalized inverse of a matrix](
https://en.wikipedia.org/wiki/Moore%E2%80%93Penrose_inverse) using its
singular-value decomposition (SVD) and including all large singular values.
The pseudo-inverse of a matrix `A`, is defined as: 'the matrix that 'solves'
[the least-squares problem] `A @ x = b`,' i.e., if `x_hat` is a solution, then
`A_pinv` is the matrix such that `x_hat = A_pinv @ b`. It can be shown that if
`U @ Sigma @ V.T = A` is the singular value decomposition of `A`, then
`A_pinv = V @ inv(Sigma) U^T`. [(Strang, 1980)][1]
This function is analogous to [`numpy.linalg.pinv`](
https://docs.scipy.org/doc/numpy/reference/generated/numpy.linalg.pinv.html).
It differs only in default value of `rcond`. In `numpy.linalg.pinv`, the
default `rcond` is `1e-15`. Here the default is
`10. * max(num_rows, num_cols) * np.finfo(dtype).eps`.
Args:
a: (Batch of) `float`-like matrix-shaped `Tensor`(s) which are to be
pseudo-inverted.
rcond: `Tensor` of small singular value cutoffs. Singular values smaller
(in modulus) than `rcond` * largest_singular_value (again, in modulus) are
set to zero. Must broadcast against `tf.shape(a)[:-2]`.
Default value: `10. * max(num_rows, num_cols) * np.finfo(a.dtype).eps`.
validate_args: When `True`, additional assertions might be embedded in the
graph.
Default value: `False` (i.e., no graph assertions are added).
name: Python `str` prefixed to ops created by this function.
Default value: 'pinv'.
Returns:
a_pinv: The pseudo-inverse of input `a`. Has same shape as `a` except
rightmost two dimensions are transposed.
Raises:
TypeError: if input `a` does not have `float`-like `dtype`.
ValueError: if input `a` has fewer than 2 dimensions.
#### Examples
```python
from tensorflow_probability.python.internal.backend import jax as tf
import tensorflow_probability as tfp; tfp = tfp.experimental.substrates.jax
a = tf.constant([[1., 0.4, 0.5],
[0.4, 0.2, 0.25],
[0.5, 0.25, 0.35]])
tf.matmul(tfp.math.pinv(a), a)
# ==> array([[1., 0., 0.],
[0., 1., 0.],
[0., 0., 1.]], dtype=float32)
a = tf.constant([[1., 0.4, 0.5, 1.],
[0.4, 0.2, 0.25, 2.],
[0.5, 0.25, 0.35, 3.]])
tf.matmul(tfp.math.pinv(a), a)
# ==> array([[ 0.76, 0.37, 0.21, -0.02],
[ 0.37, 0.43, -0.33, 0.02],
[ 0.21, -0.33, 0.81, 0.01],
[-0.02, 0.02, 0.01, 1. ]], dtype=float32)
```
#### References
[1]: G. Strang. 'Linear Algebra and Its Applications, 2nd Ed.' Academic Press,
Inc., 1980, pp. 139-142.
"""
with tf.name_scope(name or 'pinv'):
a = tf.convert_to_tensor(a, name='a')
assertions = _maybe_validate_matrix(a, validate_args)
if assertions:
with tf.control_dependencies(assertions):
a = tf.identity(a)
dtype = dtype_util.as_numpy_dtype(a.dtype)
if rcond is None:
def get_dim_size(dim):
if tf.compat.dimension_value(a.shape[dim]) is not None:
return tf.compat.dimension_value(a.shape[dim])
return tf.shape(a)[dim]
num_rows = get_dim_size(-2)
num_cols = get_dim_size(-1)
if isinstance(num_rows, int) and isinstance(num_cols, int):
max_rows_cols = float(max(num_rows, num_cols))
else:
max_rows_cols = tf.cast(tf.maximum(num_rows, num_cols), dtype)
rcond = 10. * max_rows_cols * np.finfo(dtype).eps
rcond = tf.convert_to_tensor(rcond, dtype=dtype, name='rcond')
# Calculate pseudo inverse via SVD.
# Note: if a is symmetric then u == v. (We might observe additional
# performance by explicitly setting `v = u` in such cases.)
[
singular_values, # Sigma
left_singular_vectors, # U
right_singular_vectors, # V
] = tf.linalg.svd(a, full_matrices=False, compute_uv=True)
# Saturate small singular values to inf. This has the effect of make
# `1. / s = 0.` while not resulting in `NaN` gradients.
cutoff = rcond * tf.reduce_max(singular_values, axis=-1)
singular_values = tf.where(singular_values > cutoff[..., tf.newaxis],
singular_values, np.array(np.inf, dtype))
# Although `a == tf.matmul(u, s * v, transpose_b=True)` we swap
# `u` and `v` here so that `tf.matmul(pinv(A), A) = tf.eye()`, i.e.,
# a matrix inverse has 'transposed' semantics.
a_pinv = tf.matmul(right_singular_vectors /
singular_values[..., tf.newaxis, :],
left_singular_vectors,
adjoint_b=True)
if tensorshape_util.rank(a.shape) is not None:
a_pinv.set_shape(a.shape[:-2].concatenate(
[a.shape[-1], a.shape[-2]]))
return a_pinv
def pivoted_cholesky(matrix, max_rank, diag_rtol=1e-3, name=None):
"""Computes the (partial) pivoted cholesky decomposition of `matrix`.
The pivoted Cholesky is a low rank approximation of the Cholesky decomposition
of `matrix`, i.e. as described in [(Harbrecht et al., 2012)][1]. The
currently-worst-approximated diagonal element is selected as the pivot at each
iteration. This yields from a `[B1...Bn, N, N]` shaped `matrix` a `[B1...Bn,
N, K]` shaped rank-`K` approximation `lr` such that `lr @ lr.T ~= matrix`.
Note that, unlike the Cholesky decomposition, `lr` is not triangular even in
a rectangular-matrix sense. However, under a permutation it could be made
triangular (it has one more zero in each column as you move to the right).
Such a matrix can be useful as a preconditioner for conjugate gradient
optimization, i.e. as in [(Wang et al. 2019)][2], as matmuls and solves can be
cheaply done via the Woodbury matrix identity, as implemented by
`tf.linalg.LinearOperatorLowRankUpdate`.
Args:
matrix: Floating point `Tensor` batch of symmetric, positive definite
matrices.
max_rank: Scalar `int` `Tensor`, the rank at which to truncate the
approximation.
diag_rtol: Scalar floating point `Tensor` (same dtype as `matrix`). If the
errors of all diagonal elements of `lr @ lr.T` are each lower than
`element * diag_rtol`, iteration is permitted to terminate early.
name: Optional name for the op.
Returns:
lr: Low rank pivoted Cholesky approximation of `matrix`.
#### References
[1]: H Harbrecht, M Peters, R Schneider. On the low-rank approximation by the
pivoted Cholesky decomposition. _Applied numerical mathematics_,
62(4):428-440, 2012.
[2]: K. A. Wang et al. Exact Gaussian Processes on a Million Data Points.
_arXiv preprint arXiv:1903.08114_, 2019. https://arxiv.org/abs/1903.08114
"""
with tf.name_scope(name or 'pivoted_cholesky'):
dtype = dtype_util.common_dtype([matrix, diag_rtol],
dtype_hint=tf.float32)
matrix = tf.convert_to_tensor(matrix, name='matrix', dtype=dtype)
if tensorshape_util.rank(matrix.shape) is None:
raise NotImplementedError(
'Rank of `matrix` must be known statically')
max_rank = tf.convert_to_tensor(max_rank,
name='max_rank',
dtype=tf.int64)
max_rank = tf.minimum(
max_rank,
prefer_static.shape(matrix, out_type=tf.int64)[-1])
diag_rtol = tf.convert_to_tensor(diag_rtol,
dtype=dtype,
name='diag_rtol')
matrix_diag = tf.linalg.diag_part(matrix)
# matrix is P.D., therefore all matrix_diag > 0, so we don't need abs.
orig_error = tf.reduce_max(matrix_diag, axis=-1)
def cond(m, pchol, perm, matrix_diag):
"""Condition for `tf.while_loop` continuation."""
del pchol
del perm
error = tf.linalg.norm(matrix_diag, ord=1, axis=-1)
max_err = tf.reduce_max(error / orig_error)
return (m < max_rank) & (tf.equal(m, 0) | (max_err > diag_rtol))
batch_dims = tensorshape_util.rank(matrix.shape) - 2
def batch_gather(params, indices, axis=-1):
return tf.gather(params, indices, axis=axis, batch_dims=batch_dims)
def body(m, pchol, perm, matrix_diag):
"""Body of a single `tf.while_loop` iteration."""
# Here is roughly a numpy, non-batched version of what's going to happen.
# (See also Algorithm 1 of Harbrecht et al.)
# 1: maxi = np.argmax(matrix_diag[perm[m:]]) + m
# 2: maxval = matrix_diag[perm][maxi]
# 3: perm[m], perm[maxi] = perm[maxi], perm[m]
# 4: row = matrix[perm[m]][perm[m + 1:]]
# 5: row -= np.sum(pchol[:m][perm[m + 1:]] * pchol[:m][perm[m]]], axis=-2)
# 6: pivot = np.sqrt(maxval); row /= pivot
# 7: row = np.concatenate([[[pivot]], row], -1)
# 8: matrix_diag[perm[m:]] -= row**2
# 9: pchol[m, perm[m:]] = row
# Find the maximal position of the (remaining) permuted diagonal.
# Steps 1, 2 above.
permuted_diag = batch_gather(matrix_diag, perm[..., m:])
maxi = tf.argmax(permuted_diag, axis=-1,
output_type=tf.int64)[..., tf.newaxis]
maxval = batch_gather(permuted_diag, maxi)
maxi = maxi + m
maxval = maxval[..., 0]
# Update perm: Swap perm[...,m] with perm[...,maxi]. Step 3 above.
perm = _swap_m_with_i(perm, m, maxi)
# Step 4.
row = batch_gather(matrix, perm[..., m:m + 1], axis=-2)
row = batch_gather(row, perm[..., m + 1:])
# Step 5.
prev_rows = pchol[..., :m, :]
prev_rows_perm_m_onward = batch_gather(prev_rows, perm[...,
m + 1:])
prev_rows_pivot_col = batch_gather(prev_rows, perm[..., m:m + 1])
row -= tf.reduce_sum(prev_rows_perm_m_onward * prev_rows_pivot_col,
axis=-2)[..., tf.newaxis, :]
# Step 6.
pivot = tf.sqrt(maxval)[..., tf.newaxis, tf.newaxis]
# Step 7.
row = tf.concat([pivot, row / pivot], axis=-1)
# TODO(b/130899118): Pad grad fails with int64 paddings.
# Step 8.
paddings = tf.concat([
tf.zeros([prefer_static.rank(pchol) - 1, 2], dtype=tf.int32),
[[tf.cast(m, tf.int32), 0]]
],
axis=0)
diag_update = tf.pad(row**2, paddings=paddings)[..., 0, :]
reverse_perm = _invert_permutation(perm)
matrix_diag -= batch_gather(diag_update, reverse_perm)
# Step 9.
row = tf.pad(row, paddings=paddings)
# TODO(bjp): Defer the reverse permutation all-at-once at the end?
row = batch_gather(row, reverse_perm)
pchol_shape = pchol.shape
pchol = tf.concat([pchol[..., :m, :], row, pchol[..., m + 1:, :]],
axis=-2)
tensorshape_util.set_shape(pchol, pchol_shape)
return m + 1, pchol, perm, matrix_diag
m = np.int64(0)
pchol = tf.zeros_like(matrix[..., :max_rank, :])
matrix_shape = prefer_static.shape(matrix, out_type=tf.int64)
perm = tf.broadcast_to(prefer_static.range(matrix_shape[-1]),
matrix_shape[:-1])
_, pchol, _, _ = tf.while_loop(cond=cond,
body=body,
loop_vars=(m, pchol, perm, matrix_diag))
pchol = tf.linalg.matrix_transpose(pchol)
tensorshape_util.set_shape(
pchol, tensorshape_util.concatenate(matrix_diag.shape, [None]))
return pchol
def reduce_weighted_logsumexp(logx,
w=None,
axis=None,
keep_dims=False,
return_sign=False,
name=None):
"""Computes `log(abs(sum(weight * exp(elements across tensor dimensions))))`.
If all weights `w` are known to be positive, it is more efficient to directly
use `reduce_logsumexp`, i.e., `tf.reduce_logsumexp(logx + tf.log(w))` is more
efficient than `du.reduce_weighted_logsumexp(logx, w)`.
Reduces `input_tensor` along the dimensions given in `axis`.
Unless `keep_dims` is true, the rank of the tensor is reduced by 1 for each
entry in `axis`. If `keep_dims` is true, the reduced dimensions
are retained with length 1.
If `axis` has no entries, all dimensions are reduced, and a
tensor with a single element is returned.
This function is more numerically stable than log(sum(w * exp(input))). It
avoids overflows caused by taking the exp of large inputs and underflows
caused by taking the log of small inputs.
For example:
```python
x = tf.constant([[0., 0, 0],
[0, 0, 0]])
w = tf.constant([[-1., 1, 1],
[1, 1, 1]])
du.reduce_weighted_logsumexp(x, w)
# ==> log(-1*1 + 1*1 + 1*1 + 1*1 + 1*1 + 1*1) = log(4)
du.reduce_weighted_logsumexp(x, w, axis=0)
# ==> [log(-1+1), log(1+1), log(1+1)]
du.reduce_weighted_logsumexp(x, w, axis=1)
# ==> [log(-1+1+1), log(1+1+1)]
du.reduce_weighted_logsumexp(x, w, axis=1, keep_dims=True)
# ==> [[log(-1+1+1)], [log(1+1+1)]]
du.reduce_weighted_logsumexp(x, w, axis=[0, 1])
# ==> log(-1+5)
```
Args:
logx: The tensor to reduce. Should have numeric type.
w: The weight tensor. Should have numeric type identical to `logx`.
axis: The dimensions to reduce. If `None` (the default), reduces all
dimensions. Must be in the range `[-rank(input_tensor),
rank(input_tensor))`.
keep_dims: If true, retains reduced dimensions with length 1.
return_sign: If `True`, returns the sign of the result.
name: A name for the operation (optional).
Returns:
lswe: The `log(abs(sum(weight * exp(x))))` reduced tensor.
sign: (Optional) The sign of `sum(weight * exp(x))`.
"""
with tf.name_scope(name or 'reduce_weighted_logsumexp'):
logx = tf.convert_to_tensor(logx, name='logx')
if w is None:
lswe = tf.reduce_logsumexp(logx, axis=axis, keepdims=keep_dims)
if return_sign:
sgn = tf.ones_like(lswe)
return lswe, sgn
return lswe
w = tf.convert_to_tensor(w, dtype=logx.dtype, name='w')
log_absw_x = logx + tf.math.log(tf.abs(w))
max_log_absw_x = tf.reduce_max(log_absw_x, axis=axis, keepdims=True)
# If the largest element is `-inf` or `inf` then we don't bother subtracting
# off the max. We do this because otherwise we'd get `inf - inf = NaN`. That
# this is ok follows from the fact that we're actually free to subtract any
# value we like, so long as we add it back after taking the `log(sum(...))`.
max_log_absw_x = tf.where(tf.math.is_inf(max_log_absw_x),
tf.zeros([], max_log_absw_x.dtype),
max_log_absw_x)
wx_over_max_absw_x = (tf.sign(w) * tf.exp(log_absw_x - max_log_absw_x))
sum_wx_over_max_absw_x = tf.reduce_sum(wx_over_max_absw_x,
axis=axis,
keepdims=keep_dims)
if not keep_dims:
max_log_absw_x = tf.squeeze(max_log_absw_x, axis)
sgn = tf.sign(sum_wx_over_max_absw_x)
lswe = max_log_absw_x + tf.math.log(sgn * sum_wx_over_max_absw_x)
if return_sign:
return lswe, sgn
return lswe