# Miscellaneous Ops¶

The pyro.ops module implements tensor utilities that are mostly independent of the rest of Pyro.

## Utilities for HMC¶

class DualAveraging(prox_center=0, t0=10, kappa=0.75, gamma=0.05)[source]

Bases: object

Dual Averaging is a scheme to solve convex optimization problems. It belongs to a class of subgradient methods which uses subgradients to update parameters (in primal space) of a model. Under some conditions, the averages of generated parameters during the scheme are guaranteed to converge to an optimal value. However, a counter-intuitive aspect of traditional subgradient methods is “new subgradients enter the model with decreasing weights” (see $$$$). Dual Averaging scheme solves that phenomenon by updating parameters using weights equally for subgradients (which lie in a dual space), hence we have the name “dual averaging”.

This class implements a dual averaging scheme which is adapted for Markov chain Monte Carlo (MCMC) algorithms. To be more precise, we will replace subgradients by some statistics calculated during an MCMC trajectory. In addition, introducing some free parameters such as t0 and kappa is helpful and still guarantees the convergence of the scheme.

References

 Primal-dual subgradient methods for convex problems, Yurii Nesterov

 The No-U-turn sampler: adaptively setting path lengths in Hamiltonian Monte Carlo, Matthew D. Hoffman, Andrew Gelman

Parameters: prox_center (float) – A “prox-center” parameter introduced in $$$$ which pulls the primal sequence towards it. t0 (float) – A free parameter introduced in $$$$ that stabilizes the initial steps of the scheme. kappa (float) – A free parameter introduced in $$$$ that controls the weights of steps of the scheme. For a small kappa, the scheme will quickly forget states from early steps. This should be a number in $$(0.5, 1]$$. gamma (float) – A free parameter which controls the speed of the convergence of the scheme.
reset()[source]
step(g)[source]

Updates states of the scheme given a new statistic/subgradient g.

Parameters: g (float) – A statistic calculated during an MCMC trajectory or subgradient.
get_state()[source]

Returns the latest $$x_t$$ and average of $$\left\{x_i\right\}_{i=1}^t$$ in primal space.

velocity_verlet(z, r, potential_fn, kinetic_grad, step_size, num_steps=1, z_grads=None)[source]

Second order symplectic integrator that uses the velocity verlet algorithm.

Parameters: Return tuple (z_next, r_next, z_grads, potential_energy): z (dict) – dictionary of sample site names and their current values (type Tensor). r (dict) – dictionary of sample site names and corresponding momenta (type Tensor). potential_fn (callable) – function that returns potential energy given z for each sample site. The negative gradient of the function with respect to z determines the rate of change of the corresponding sites’ momenta r. kinetic_grad (callable) – a function calculating gradient of kinetic energy w.r.t. momentum variable. step_size (float) – step size for each time step iteration. num_steps (int) – number of discrete time steps over which to integrate. z_grads (torch.Tensor) – optional gradients of potential energy at current z. next position and momenta, together with the potential energy and its gradient w.r.t. z_next.
potential_grad(potential_fn, z)[source]

Gradient of potential_fn w.r.t. parameters z.

Parameters: potential_fn – python callable that takes in a dictionary of parameters and returns the potential energy. z (dict) – dictionary of parameter values keyed by site name. tuple of (z_grads, potential_energy), where z_grads is a dictionary with the same keys as z containing gradients and potential_energy is a torch scalar.
class WelfordCovariance(diagonal=True)[source]

Bases: object

Implements Welford’s online scheme for estimating (co)variance (see $$$$). Useful for adapting diagonal and dense mass structures for HMC.

References

 The Art of Computer Programming, Donald E. Knuth

reset()[source]
update(sample)[source]
get_covariance(regularize=True)[source]
class WelfordArrowheadCovariance(head_size=0)[source]

Bases: object

Likes WelfordCovariance but generalized to the arrowhead structure.

reset()[source]
update(sample)[source]
get_covariance(regularize=True)[source]

## Newton Optimizers¶

newton_step(loss, x, trust_radius=None)[source]

Performs a Newton update step to minimize loss on a batch of variables, optionally constraining to a trust region .

This is especially usful because the final solution of newton iteration is differentiable wrt the inputs, even when all but the final x is detached, due to this method’s quadratic convergence . loss must be twice-differentiable as a function of x. If loss is 2+d-times differentiable, then the return value of this function is d-times differentiable.

When loss is interpreted as a negative log probability density, then the return values mode,cov of this function can be used to construct a Laplace approximation MultivariateNormal(mode,cov).

Warning

Take care to detach the result of this function when used in an optimization loop. If you forget to detach the result of this function during optimization, then backprop will propagate through the entire iteration process, and worse will compute two extra derivatives for each step.

Example use inside a loop:

x = torch.zeros(1000, 2)  # arbitrary initial value
for step in range(100):
x = x.detach()          # block gradients through previous steps
x.requires_grad = True  # ensure loss is differentiable wrt x
loss = my_loss_function(x)
# the final x is still differentiable

 Yuan, Ya-xiang. Iciam. Vol. 99. 2000.
“A review of trust region algorithms for optimization.” ftp://ftp.cc.ac.cn/pub/yyx/papers/p995.pdf
 Christianson, Bruce. Optimization Methods and Software 3.4 (1994)
“Reverse accumulation and attractive fixed points.” http://uhra.herts.ac.uk/bitstream/handle/2299/4338/903839.pdf
Parameters: loss (torch.Tensor) – A scalar function of x to be minimized. x (torch.Tensor) – A dependent variable of shape (N, D) where N is the batch size and D is a small number. trust_radius (float) – An optional trust region trust_radius. The updated value mode of this function will be within trust_radius of the input x. A pair (mode, cov) where mode is an updated tensor of the same shape as the original value x, and cov is an esitmate of the covariance DxD matrix with cov.shape == x.shape[:-1] + (D,D). tuple
newton_step_1d(loss, x, trust_radius=None)[source]

Performs a Newton update step to minimize loss on a batch of 1-dimensional variables, optionally regularizing to constrain to a trust region.

See newton_step() for details.

Parameters: loss (torch.Tensor) – A scalar function of x to be minimized. x (torch.Tensor) – A dependent variable with rightmost size of 1. trust_radius (float) – An optional trust region trust_radius. The updated value mode of this function will be within trust_radius of the input x. A pair (mode, cov) where mode is an updated tensor of the same shape as the original value x, and cov is an esitmate of the covariance 1x1 matrix with cov.shape == x.shape[:-1] + (1,1). tuple
newton_step_2d(loss, x, trust_radius=None)[source]

Performs a Newton update step to minimize loss on a batch of 2-dimensional variables, optionally regularizing to constrain to a trust region.

See newton_step() for details.

Parameters: loss (torch.Tensor) – A scalar function of x to be minimized. x (torch.Tensor) – A dependent variable with rightmost size of 2. trust_radius (float) – An optional trust region trust_radius. The updated value mode of this function will be within trust_radius of the input x. A pair (mode, cov) where mode is an updated tensor of the same shape as the original value x, and cov is an esitmate of the covariance 2x2 matrix with cov.shape == x.shape[:-1] + (2,2). tuple
newton_step_3d(loss, x, trust_radius=None)[source]

Performs a Newton update step to minimize loss on a batch of 3-dimensional variables, optionally regularizing to constrain to a trust region.

See newton_step() for details.

Parameters: loss (torch.Tensor) – A scalar function of x to be minimized. x (torch.Tensor) – A dependent variable with rightmost size of 2. trust_radius (float) – An optional trust region trust_radius. The updated value mode of this function will be within trust_radius of the input x. A pair (mode, cov) where mode is an updated tensor of the same shape as the original value x, and cov is an esitmate of the covariance 3x3 matrix with cov.shape == x.shape[:-1] + (3,3). tuple

## Special Functions¶

safe_log(x)[source]

Like torch.log() but avoids infinite gradients at log(0) by clamping them to at most 1 / finfo.eps.

log_beta(x, y, tol=0.0)[source]

Computes log Beta function.

When tol >= 0.02 this uses a shifted Stirling’s approximation to the log Beta function. The approximation adapts Stirling’s approximation of the log Gamma function:

lgamma(z) ≈ (z - 1/2) * log(z) - z + log(2 * pi) / 2


to approximate the log Beta function:

log_beta(x, y) ≈ ((x-1/2) * log(x) + (y-1/2) * log(y)
- (x+y-1/2) * log(x+y) + log(2*pi)/2)


The approximation additionally improves accuracy near zero by iteratively shifting the log Gamma approximation using the recursion:

lgamma(x) = lgamma(x + 1) - log(x)


If this recursion is applied n times, then absolute error is bounded by error < 0.082 / n < tol, thus we choose n based on the user provided tol.

Parameters: x (torch.Tensor) – A positive tensor. y (torch.Tensor) – A positive tensor. tol (float) – Bound on maximum absolute error. Defaults to 0.1. For very small tol, this function simply defers to log_beta(). torch.Tensor
log_binomial(n, k, tol=0.0)[source]

Computes log binomial coefficient.

When tol >= 0.02 this uses a shifted Stirling’s approximation to the log Beta function via log_beta().

Parameters: n (torch.Tensor) – A nonnegative integer tensor. k (torch.Tensor) – An integer tensor ranging in [0, n]. torch.Tensor

## Tensor Utilities¶

block_diag_embed(mat)[source]

Takes a tensor of shape (…, B, M, N) and returns a block diagonal tensor of shape (…, B x M, B x N).

Parameters: Returns torch.Tensor: mat (torch.Tensor) – an input tensor with 3 or more dimensions a block diagonal tensor with dimension m.dim() - 1
block_diagonal(mat, block_size)[source]

Takes a block diagonal tensor of shape (…, B x M, B x N) and returns a tensor of shape (…, B, M, N).

Parameters: Returns torch.Tensor: mat (torch.Tensor) – an input tensor with 2 or more dimensions block_size (int) – the number of blocks B. a tensor with dimension mat.dim() + 1
periodic_repeat(tensor, size, dim)[source]

Repeat a period-sized tensor up to given size. For example:

>>> x = torch.tensor([[1, 2, 3], [4, 5, 6]])
>>> periodic_repeat(x, 4, 0)
tensor([[1, 2, 3],
[4, 5, 6],
[1, 2, 3],
[4, 5, 6]])
>>> periodic_repeat(x, 4, 1)
tensor([[1, 2, 3, 1],
[4, 5, 6, 4]])


This is useful for computing static seasonality in time series models.

Parameters: tensor (torch.Tensor) – A tensor of differences. size (int) – Desired size of the result along dimension dim. dim (int) – The tensor dimension along which to repeat.
periodic_cumsum(tensor, period, dim)[source]

Compute periodic cumsum along a given dimension. For example if dim=0:

for t in range(period):
assert result[t] == tensor[t]
for t in range(period, len(tensor)):
assert result[t] == tensor[t] + result[t - period]


This is useful for computing drifting seasonality in time series models.

Parameters: tensor (torch.Tensor) – A tensor of differences. period (int) – The period of repetition. dim (int) – The tensor dimension along which to accumulate.
periodic_features(duration, max_period=None, min_period=None, **options)[source]

Create periodic (sin,cos) features from max_period down to min_period.

This is useful in time series models where long uneven seasonality can be treated via regression. When only max_period is specified this generates periodic features at all length scales. When also min_period is specified this generates periodic features at large length scales, but omits high frequency features. This is useful when combining regression for long seasonality with other techniques like periodic_repeat() and periodic_cumsum() for short time scales. For example, to combine regress yearly seasonality down to the scale of one week one could set max_period=365.25 and min_period=7.

Parameters: duration (int) – Number of discrete time steps. max_period (float) – Optional max period, defaults to duration. min_period (float) – Optional min period (exclusive), defaults to 2 = Nyquist cutoff. **options – Tensor construction options, e.g. dtype and device. A (duration, 2 * ceil(max_period / min_period) - 2)-shaped tensor of features normalized to lie in [-1,1]. Tensor
next_fast_len(size)[source]

Returns the next largest number n >= size whose prime factors are all 2, 3, or 5. These sizes are efficient for fast fourier transforms. Equivalent to scipy.fftpack.next_fast_len().

Parameters: size (int) – A positive number. A possibly larger number.
convolve(signal, kernel, mode='full')[source]

Computes the 1-d convolution of signal by kernel using FFTs. The two arguments should have the same rightmost dim, but may otherwise be arbitrarily broadcastable.

Parameters: signal (torch.Tensor) – A signal to convolve. kernel (torch.Tensor) – A convolution kernel. mode (str) – One of: ‘full’, ‘valid’, ‘same’. A tensor with broadcasted shape. Letting m = signal.size(-1) and n = kernel.size(-1), the rightmost size of the result will be: m + n - 1 if mode is ‘full’; max(m, n) - min(m, n) + 1 if mode is ‘valid’; or max(m, n) if mode is ‘same’.
repeated_matmul(M, n)[source]

Takes a batch of matrices M as input and returns the stacked result of doing the n-many matrix multiplications $$M$$, $$M^2$$, …, $$M^n$$. Parallel cost is logarithmic in n.

Parameters: Returns torch.Tensor: M (torch.Tensor) – A batch of square tensors of shape (…, N, N). n (int) – The order of the largest product $$M^n$$ A batch of square tensors of shape (n, …, N, N)
dct(x, dim=-1)[source]

Discrete cosine transform of type II, scaled to be orthonormal.

This is the inverse of idct_ii() , and is equivalent to scipy.fftpack.dct() with norm="ortho".

Parameters: x (Tensor) – The input signal. dim (int) – Dimension along which to compute DCT. Tensor
idct(x, dim=-1)[source]

Inverse discrete cosine transform of type II, scaled to be orthonormal.

This is the inverse of dct_ii() , and is equivalent to scipy.fftpack.idct() with norm="ortho".

Parameters: x (Tensor) – The input signal. dim (int) – Dimension along which to compute DCT. Tensor
haar_transform(x)[source]

Discrete Haar transform.

Performs a Haar transform along the final dimension. This is the inverse of inverse_haar_transform().

Parameters: x (Tensor) – The input signal. Tensor
inverse_haar_transform(x)[source]

Performs an inverse Haar transform along the final dimension. This is the inverse of haar_transform().

Parameters: x (Tensor) – The input signal. Tensor
cholesky(x)[source]
cholesky_solve(x, y)[source]
matmul(x, y)[source]
matvecmul(x, y)[source]
triangular_solve(x, y, upper=False, transpose=False)[source]
precision_to_scale_tril(P)[source]

## Tensor Indexing¶

index(tensor, args)[source]

Indexing with nested tuples.

See also the convenience wrapper Index.

This is useful for writing indexing code that is compatible with multiple interpretations, e.g. scalar evaluation, vectorized evaluation, or reshaping.

For example suppose x is a parameter with x.dim() == 2 and we wish to generalize the expression x[..., t] where t can be any of:

• a scalar t=1 as in x[..., 1];
• a slice t=slice(None) equivalent to x[..., :]; or
• a reshaping operation t=(Ellipsis, None) equivalent to x.unsqueeze(-1).

While naive indexing would work for the first two , the third example would result in a nested tuple (Ellipsis, (Ellipsis, None)). This helper flattens that nested tuple and combines consecutive Ellipsis.

Parameters: tensor (torch.Tensor) – A tensor to be indexed. args (tuple) – An index, as args to __getitem__. A flattened interpetation of tensor[args]. torch.Tensor
class Index(tensor)[source]

Bases: object

Convenience wrapper around index().

The following are equivalent:

Index(x)[..., i, j, :]
index(x, (Ellipsis, i, j, slice(None)))

Parameters: tensor (torch.Tensor) – A tensor to be indexed. An object with a special __getitem__() method.
vindex(tensor, args)[source]

See also the convenience wrapper Vindex.

This is useful for writing indexing code that is compatible with batching and enumeration, especially for selecting mixture components with discrete random variables.

For example suppose x is a parameter with x.dim() == 3 and we wish to generalize the expression x[i, :, j] from integer i,j to tensors i,j with batch dims and enum dims (but no event dims). Then we can write the generalize version using Vindex

xij = Vindex(x)[i, :, j]

event_shape = (x.size(1),)
assert xij.shape == batch_shape + event_shape


To handle the case when x may also contain batch dimensions (e.g. if x was sampled in a plated context as when using vectorized particles), vindex() uses the special convention that Ellipsis denotes batch dimensions (hence ... can appear only on the left, never in the middle or in the right). Suppose x has event dim 3. Then we can write:

old_batch_shape = x.shape[:-3]
old_event_shape = x.shape[-3:]

xij = Vindex(x)[..., i, :, j]   # The ... denotes unknown batch shape.

new_event_shape = (x.size(1),)
assert xij.shape = new_batch_shape + new_event_shape


Note that this special handling of Ellipsis differs from the NEP .

Formally, this function assumes:

1. Each arg is either Ellipsis, slice(None), an integer, or a batched torch.LongTensor (i.e. with empty event shape). This function does not support Nontrivial slices or torch.BoolTensor masks. Ellipsis can only appear on the left as args.
2. If args is not Ellipsis then tensor is not batched, and its event dim is equal to len(args).
3. If args is Ellipsis then tensor is batched and its event dim is equal to len(args[1:]). Dims of tensor to the left of the event dims are considered batch dims and will be broadcasted with dims of tensor args.

Note that if none of the args is a tensor with .dim() > 0, then this function behaves like standard indexing:

if not any(isinstance(a, torch.Tensor) and a.dim() for a in args):
assert Vindex(x)[args] == x[args]


References

introduces vindex as a helper for vectorized indexing. The Pyro implementation is similar to the proposed notation x.vindex[] except for slightly different handling of Ellipsis.
Parameters: tensor (torch.Tensor) – A tensor to be indexed. args (tuple) – An index, as args to __getitem__. A nonstandard interpetation of tensor[args]. torch.Tensor
class Vindex(tensor)[source]

Bases: object

Convenience wrapper around vindex().

The following are equivalent:

Vindex(x)[..., i, j, :]
vindex(x, (Ellipsis, i, j, slice(None)))

Parameters: tensor (torch.Tensor) – A tensor to be indexed. An object with a special __getitem__() method.

## Tensor Contraction¶

contract_expression(equation, *shapes, **kwargs)[source]

Wrapper around opt_einsum.contract_expression() that optionally uses Pyro’s cheap optimizer and optionally caches contraction paths.

Parameters: cache_path (bool) – whether to cache the contraction path. Defaults to True.
contract(equation, *operands, **kwargs)[source]

Wrapper around opt_einsum.contract() that optionally uses Pyro’s cheap optimizer and optionally caches contraction paths.

Parameters: cache_path (bool) – whether to cache the contraction path. Defaults to True.
einsum(equation, *operands, **kwargs)[source]

Generalized plated sum-product algorithm via tensor variable elimination.

This generalizes contract() in two ways:

1. Multiple outputs are allowed, and intermediate results can be shared.
2. Inputs and outputs can be plated along symbols given in plates; reductions along plates are product reductions.

The best way to understand this function is to try the examples below, which show how einsum() calls can be implemented as multiple calls to contract() (which is generally more expensive).

To illustrate multiple outputs, note that the following are equivalent:

z1, z2, z3 = einsum('ab,bc->a,b,c', x, y)  # multiple outputs

z1 = contract('ab,bc->a', x, y)
z2 = contract('ab,bc->b', x, y)
z3 = contract('ab,bc->c', x, y)


To illustrate plated inputs, note that the following are equivalent:

assert len(x) == 3 and len(y) == 3
z = einsum('ab,ai,bi->b', w, x, y, plates='i')

z = contract('ab,a,a,a,b,b,b->b', w, *x, *y)


When a sum dimension a always appears with a plate dimension i, then a corresponds to a distinct symbol for each slice of a. Thus the following are equivalent:

assert len(x) == 3 and len(y) == 3
z = einsum('ai,ai->', x, y, plates='i')

z = contract('a,b,c,a,b,c->', *x, *y)


When such a sum dimension appears in the output, it must be accompanied by all of its plate dimensions, e.g. the following are equivalent:

assert len(x) == 3 and len(y) == 3
z = einsum('abi,abi->bi', x, y, plates='i')

z = torch.stack([z0, z1, z2])


Note that each plate slice through the output is multilinear in all plate slices through all inptus, thus e.g. batch matrix multiply would be implemented without plates, so the following are all equivalent:

xy = einsum('abc,acd->abd', x, y, plates='')
xy = torch.stack([xa.mm(ya) for xa, ya in zip(x, y)])
xy = torch.bmm(x, y)


Among all valid equations, some computations are polynomial in the sizes of the input tensors and other computations are exponential in the sizes of the input tensors. This function raises NotImplementedError whenever the computation is exponential.

Parameters: equation (str) – An einsum equation, optionally with multiple outputs. operands (torch.Tensor) – A collection of tensors. plates (str) – An optional string of plate symbols. backend (str) – An optional einsum backend, defaults to ‘torch’. cache (dict) – An optional shared_intermediates() cache. modulo_total (bool) – Optionally allow einsum to arbitrarily scale each result plate, which can significantly reduce computation. This is safe to set whenever each result plate denotes a nonnormalized probability distribution whose total is not of interest. a tuple of tensors of requested shape, one entry per output. tuple ValueError – if tensor sizes mismatch or an output requests a plated dim without that dim’s plates. NotImplementedError – if contraction would have cost exponential in the size of any input tensor.
ubersum(equation, *operands, **kwargs)[source]

Deprecated, use einsum() instead.

## Gaussian Contraction¶

class Gaussian(log_normalizer, info_vec, precision)[source]

Bases: object

Non-normalized Gaussian distribution.

This represents an arbitrary semidefinite quadratic function, which can be interpreted as a rank-deficient scaled Gaussian distribution. The precision matrix may have zero eigenvalues, thus it may be impossible to work directly with the covariance matrix.

Parameters: log_normalizer (torch.Tensor) – a normalization constant, which is mainly used to keep track of normalization terms during contractions. info_vec (torch.Tensor) – information vector, which is a scaled version of the mean info_vec = precision @ mean. We use this represention to make gaussian contraction fast and stable. precision (torch.Tensor) – precision matrix of this gaussian.
dim()[source]
batch_shape[source]
expand(batch_shape)[source]
reshape(batch_shape)[source]
__getitem__(index)[source]

Index into the batch_shape of a Gaussian.

static cat(parts, dim=0)[source]

Concatenate a list of Gaussians along a given batch dimension.

event_pad(left=0, right=0)[source]

event_permute(perm)[source]

Permute along event dimension.

__add__(other)[source]

Adds two Gaussians in log-density space.

log_density(value)[source]

Evaluate the log density of this Gaussian at a point value:

-0.5 * value.T @ precision @ value + value.T @ info_vec + log_normalizer


This is mainly used for testing.

rsample(sample_shape=torch.Size([]))[source]

Reparameterized sampler.

condition(value)[source]

Condition this Gaussian on a trailing subset of its state. This should satisfy:

g.condition(y).dim() == g.dim() - y.size(-1)


Note that since this is a non-normalized Gaussian, we include the density of y in the result. Thus condition() is similar to a functools.partial binding of arguments:

left = x[..., :n]
right = x[..., n:]
g.log_density(x) == g.condition(right).log_density(left)

left_condition(value)[source]

Condition this Gaussian on a leading subset of its state. This should satisfy:

g.condition(y).dim() == g.dim() - y.size(-1)


Note that since this is a non-normalized Gaussian, we include the density of y in the result. Thus condition() is similar to a functools.partial binding of arguments:

left = x[..., :n]
right = x[..., n:]
g.log_density(x) == g.left_condition(left).log_density(right)

marginalize(left=0, right=0)[source]

Marginalizing out variables on either side of the event dimension:

g.marginalize(left=n).event_logsumexp() = g.logsumexp()
g.marginalize(right=n).event_logsumexp() = g.logsumexp()


and for data x:

g.condition(x).event_logsumexp()
= g.marginalize(left=g.dim() - x.size(-1)).log_density(x)
event_logsumexp()[source]

Integrates out all latent state (i.e. operating on event dimensions).

class AffineNormal(matrix, loc, scale)[source]

Bases: object

Represents a conditional diagonal normal distribution over a random variable Y whose mean is an affine function of a random variable X. The likelihood of X is thus:

AffineNormal(matrix, loc, scale).condition(y).log_density(x)


which is equivalent to:

Normal(x @ matrix + loc, scale).to_event(1).log_prob(y)

Parameters: matrix (torch.Tensor) – A transformation from X to Y. Should have rightmost shape (x_dim, y_dim). loc (torch.Tensor) – A constant offset for Y’s mean. Should have rightmost shape (y_dim,). scale (torch.Tensor) – Standard deviation for Y. Should have rightmost shape (y_dim,).
batch_shape[source]
condition(value)[source]
left_condition(value)[source]

If value.size(-1) == x_dim, this returns a Normal distribution with event_dim=1. After applying this method, the cost to draw a sample is O(y_dim) instead of O(y_dim ** 3).

rsample(sample_shape=torch.Size([]))[source]

Reparameterized sampler.

to_gaussian()[source]
expand(batch_shape)[source]
reshape(batch_shape)[source]
__getitem__(index)[source]
event_permute(perm)[source]
__add__(other)[source]
marginalize(left=0, right=0)[source]
mvn_to_gaussian(mvn)[source]

Convert a MultivariateNormal distribution to a Gaussian.

Parameters: mvn (MultivariateNormal) – A multivariate normal distribution. An equivalent Gaussian object. Gaussian
matrix_and_mvn_to_gaussian(matrix, mvn)[source]

Convert a noisy affine function to a Gaussian. The noisy affine function is defined as:

y = x @ matrix + mvn.sample()

Parameters: matrix (Tensor) – A matrix with rightmost shape (x_dim, y_dim). mvn (MultivariateNormal) – A multivariate normal distribution. A Gaussian with broadcasted batch shape and .dim() == x_dim + y_dim. Gaussian
gaussian_tensordot(x, y, dims=0)[source]

Computes the integral over two gaussians:

(x @ y)(a,c) = log(integral(exp(x(a,b) + y(b,c)), b)),

where x is a gaussian over variables (a,b), y is a gaussian over variables (b,c), (a,b,c) can each be sets of zero or more variables, and dims is the size of b.

Parameters: x – a Gaussian instance y – a Gaussian instance dims – number of variables to contract

## Statistical Utilities¶

gelman_rubin(input, chain_dim=0, sample_dim=1)[source]

Computes R-hat over chains of samples. It is required that input.size(sample_dim) >= 2 and input.size(chain_dim) >= 2.

Parameters: Returns torch.Tensor: input (torch.Tensor) – the input tensor. chain_dim (int) – the chain dimension. sample_dim (int) – the sample dimension. R-hat of input.
split_gelman_rubin(input, chain_dim=0, sample_dim=1)[source]

Computes R-hat over chains of samples. It is required that input.size(sample_dim) >= 4.

Parameters: Returns torch.Tensor: input (torch.Tensor) – the input tensor. chain_dim (int) – the chain dimension. sample_dim (int) – the sample dimension. split R-hat of input.
autocorrelation(input, dim=0)[source]

Computes the autocorrelation of samples at dimension dim.

Parameters: Returns torch.Tensor: input (torch.Tensor) – the input tensor. dim (int) – the dimension to calculate autocorrelation. autocorrelation of input.
autocovariance(input, dim=0)[source]

Computes the autocovariance of samples at dimension dim.

Parameters: Returns torch.Tensor: input (torch.Tensor) – the input tensor. dim (int) – the dimension to calculate autocorrelation. autocorrelation of input.
effective_sample_size(input, chain_dim=0, sample_dim=1)[source]

Computes effective sample size of input.

Reference:

 Introduction to Markov Chain Monte Carlo,
Charles J. Geyer
 Stan Reference Manual version 2.18,
Stan Development Team
Parameters: Returns torch.Tensor: input (torch.Tensor) – the input tensor. chain_dim (int) – the chain dimension. sample_dim (int) – the sample dimension. effective sample size of input.
resample(input, num_samples, dim=0, replacement=False)[source]

Draws num_samples samples from input at dimension dim.

Parameters: Returns torch.Tensor: input (torch.Tensor) – the input tensor. num_samples (int) – the number of samples to draw from input. dim (int) – dimension to draw from input. samples drawn randomly from input.
quantile(input, probs, dim=0)[source]

Computes quantiles of input at probs. If probs is a scalar, the output will be squeezed at dim.

Parameters: Returns torch.Tensor: input (torch.Tensor) – the input tensor. probs (list) – quantile positions. dim (int) – dimension to take quantiles from input. quantiles of input at probs.
pi(input, prob, dim=0)[source]

Computes percentile interval which assigns equal probability mass to each tail of the interval.

Parameters: Returns torch.Tensor: input (torch.Tensor) – the input tensor. prob (float) – the probability mass of samples within the interval. dim (int) – dimension to calculate percentile interval from input. quantiles of input at probs.
hpdi(input, prob, dim=0)[source]

Computes “highest posterior density interval” which is the narrowest interval with probability mass prob.

Parameters: Returns torch.Tensor: input (torch.Tensor) – the input tensor. prob (float) – the probability mass of samples within the interval. dim (int) – dimension to calculate percentile interval from input. quantiles of input at probs.
waic(input, log_weights=None, pointwise=False, dim=0)[source]

Computes “Widely Applicable/Watanabe-Akaike Information Criterion” (WAIC) and its corresponding effective number of parameters.

Reference:

 WAIC and cross-validation in Stan, Aki Vehtari, Andrew Gelman

Parameters: input (torch.Tensor) – the input tensor, which is log likelihood of a model. log_weights (torch.Tensor) – weights of samples along dim. dim (int) – the sample dimension of input. tuple of WAIC and effective number of parameters.
fit_generalized_pareto(X)[source]

Given a dataset X assumed to be drawn from the Generalized Pareto Distribution, estimate the distributional parameters k, sigma using a variant of the technique described in reference , as described in reference .

References  ‘A new and efficient estimation method for the generalized Pareto distribution.’ Zhang, J. and Stephens, M.A. (2009).  ‘Pareto Smoothed Importance Sampling.’ Aki Vehtari, Andrew Gelman, Jonah Gabry

Parameters: torch.Tensor – the input data X tuple of floats (k, sigma) corresponding to the fit parameters
crps_empirical(pred, truth)[source]

Computes negative Continuous Ranked Probability Score CRPS*  between a set of samples pred and true data truth. This uses an n log(n) time algorithm to compute a quantity equal that would naively have complexity quadratic in the number of samples n:

CRPS* = E|pred - truth| - 1/2 E|pred - pred'|
= (pred - truth).abs().mean(0)
- (pred - pred.unsqueeze(1)).abs().mean([0, 1]) / 2


Note that for a single sample this reduces to absolute error.

References

 Tilmann Gneiting, Adrian E. Raftery (2007)
Strictly Proper Scoring Rules, Prediction, and Estimation https://www.stat.washington.edu/raftery/Research/PDF/Gneiting2007jasa.pdf
Parameters: pred (torch.Tensor) – A set of sample predictions batched on rightmost dim. This should have shape (num_samples,) + truth.shape. truth (torch.Tensor) – A tensor of true observations. A tensor of shape truth.shape. torch.Tensor

## State Space Model and GP Utilities¶

class MaternKernel(nu=1.5, num_gps=1, length_scale_init=None, kernel_scale_init=None)[source]

Provides the building blocks for representing univariate Gaussian Processes (GPs) with Matern kernels as state space models.

Parameters: nu (float) – The order of the Matern kernel (one of 0.5, 1.5 or 2.5) num_gps (int) – the number of GPs length_scale_init (torch.Tensor) – optional num_gps-dimensional vector of initializers for the length scale kernel_scale_init (torch.Tensor) – optional num_gps-dimensional vector of initializers for the kernel scale

References

 Kalman Filtering and Smoothing Solutions to Temporal Gaussian Process Regression Models,
Jouni Hartikainen and Simo Sarkka.
 Stochastic Differential Equation Methods for Spatio-Temporal Gaussian Process Regression,
Arno Solin.
transition_matrix(dt)[source]

Compute the (exponentiated) transition matrix of the GP latent space. The resulting matrix has layout (num_gps, old_state, new_state), i.e. this matrix multiplies states from the right.

See section 5 in reference  for details.

Parameters: Returns torch.Tensor: dt (float) – the time interval over which the GP latent space evolves. a 3-dimensional tensor of transition matrices of shape (num_gps, state_dim, state_dim).
stationary_covariance()[source]

Compute the stationary state covariance. See Eqn. 3.26 in reference .

Returns torch.Tensor:
a 3-dimensional tensor of covariance matrices of shape (num_gps, state_dim, state_dim).
process_covariance(A)[source]

Given a transition matrix A computed with transition_matrix compute the the process covariance as described in Eqn. 3.11 in reference .

Returns torch.Tensor:
a batched covariance matrix of shape (num_gps, state_dim, state_dim)
transition_matrix_and_covariance(dt)[source]

Get the transition matrix and process covariance corresponding to a time interval dt.

Parameters: dt (float) – the time interval over which the GP latent space evolves. (transition_matrix, process_covariance) both 3-dimensional tensors of shape (num_gps, state_dim, state_dim)