Families - Quivers

ConditionalNormal ¶

ConditionalNormal(domain: AnySpace, codomain: ContinuousSpace, hidden_dim: int = 64)

Bases: ContinuousMorphism

Conditional normal (Gaussian) distribution.

For each input x, produces an independent normal distribution on each dimension of the codomain:

y_i ~ Normal(mu_i(x), sigma_i(x))

Parameters are learnable: mu and log(sigma) are functions of x, implemented as lookup tables (discrete domain) or neural networks (continuous domain).

PARAMETER	DESCRIPTION
`domain`	Source space. TYPE: `SetObject or ContinuousSpace`
`codomain`	Target space. TYPE: `Euclidean`
`hidden_dim`	Hidden layer width for neural parameter source. TYPE: `int` DEFAULT: `64`

Examples:

>>> from quivers import FinSet
>>> from quivers.continuous.spaces import Euclidean
>>> A = FinSet(name="context", cardinality=5)
>>> Y = Euclidean(name="response", dim=3)
>>> f = ConditionalNormal(A, Y)
>>> x = torch.tensor([0, 1, 2])
>>> samples = f.rsample(x)  # shape (3, 3)

Source code in src/quivers/continuous/families.py

def __init__(
    self,
    domain: AnySpace,
    codomain: ContinuousSpace,
    hidden_dim: int = 64,
) -> None:
    super().__init__(domain, codomain)
    d = codomain.dim

    # param_dim = d (mu) + d (log_sigma)
    self.param_source = _make_source(domain, 2 * d, hidden_dim)
    self._d = d

ConditionalLogitNormal ¶

ConditionalLogitNormal(domain: AnySpace, codomain: ContinuousSpace, hidden_dim: int = 64)

Bases: ContinuousMorphism

Conditional logit-normal distribution on (0, 1)^d.

If z ~ Normal(mu(x), sigma(x)), then y = sigmoid(z) ~ LogitNormal. Useful for modeling probabilities and bounded quantities.

PARAMETER	DESCRIPTION
`domain`	Source space. TYPE: `SetObject or ContinuousSpace`
`codomain`	Target space (should have bounds [0, 1]). TYPE: `Euclidean`
`hidden_dim`	Hidden layer width for neural parameter source. TYPE: `int` DEFAULT: `64`

Source code in src/quivers/continuous/families.py

def __init__(
    self,
    domain: AnySpace,
    codomain: ContinuousSpace,
    hidden_dim: int = 64,
) -> None:
    super().__init__(domain, codomain)
    d = codomain.dim
    self.param_source = _make_source(domain, 2 * d, hidden_dim)
    self._d = d

ConditionalBeta ¶

ConditionalBeta(domain: AnySpace, codomain: ContinuousSpace, hidden_dim: int = 64)

Bases: ContinuousMorphism

Conditional beta distribution on (0, 1)^d.

For each input x, produces an independent Beta(alpha_i(x), beta_i(x)) on each dimension of the codomain.

PARAMETER	DESCRIPTION
`domain`	Source space. TYPE: `SetObject or ContinuousSpace`
`codomain`	Target space (should have bounds [0, 1]). TYPE: `Euclidean`
`hidden_dim`	Hidden layer width for neural parameter source. TYPE: `int` DEFAULT: `64`

Source code in src/quivers/continuous/families.py

def __init__(
    self,
    domain: AnySpace,
    codomain: ContinuousSpace,
    hidden_dim: int = 64,
) -> None:
    super().__init__(domain, codomain)
    d = codomain.dim
    self.param_source = _make_source(domain, 2 * d, hidden_dim)
    self._d = d

ConditionalTruncatedNormal ¶

ConditionalTruncatedNormal(domain: AnySpace, codomain: Euclidean, hidden_dim: int = 64)

Bases: ContinuousMorphism

Conditional truncated normal on [low, high]^d.

A normal distribution restricted to a bounded interval. Uses rejection-free sampling via the inverse CDF (Phi-based) method.

PARAMETER	DESCRIPTION
`domain`	Source space. TYPE: `SetObject or ContinuousSpace`
`codomain`	Target space (must be bounded). TYPE: `Euclidean`
`hidden_dim`	Hidden layer width for neural parameter source. TYPE: `int` DEFAULT: `64`

Source code in src/quivers/continuous/families.py

def __init__(
    self,
    domain: AnySpace,
    codomain: Euclidean,
    hidden_dim: int = 64,
) -> None:
    if codomain.low is None or codomain.high is None:
        raise ValueError("ConditionalTruncatedNormal requires a bounded codomain")

    super().__init__(domain, codomain)
    d = codomain.dim
    self.param_source = _make_source(domain, 2 * d, hidden_dim)
    self._d = d
    self._low = codomain.low
    self._high = codomain.high

ConditionalDirichlet ¶

ConditionalDirichlet(domain: AnySpace, codomain: ContinuousSpace, hidden_dim: int = 64)

Bases: ContinuousMorphism

Conditional Dirichlet distribution on a probability simplex.

For each input x, produces a Dirichlet(alpha(x)) distribution on the simplex.

PARAMETER	DESCRIPTION
`domain`	Source space. TYPE: `SetObject or ContinuousSpace`
`codomain`	Target simplex. TYPE: `Simplex`
`hidden_dim`	Hidden layer width for neural parameter source. TYPE: `int` DEFAULT: `64`

Source code in src/quivers/continuous/families.py

def __init__(
    self,
    domain: AnySpace,
    codomain: ContinuousSpace,
    hidden_dim: int = 64,
) -> None:
    super().__init__(domain, codomain)
    d = codomain.dim
    self.param_source = _make_source(domain, d, hidden_dim)
    self._d = d

ConditionalUniform ¶

ConditionalUniform(domain: AnySpace, codomain: ContinuousSpace, hidden_dim: int = 64)

Bases: ContinuousMorphism

Conditional uniform distribution on a learnable interval.

Parameterized as Uniform(loc - width/2, loc + width/2) where loc is unconstrained and width is positive. This ensures low < high is always satisfied.

PARAMETER	DESCRIPTION
`domain`	Source space. TYPE: `SetObject or ContinuousSpace`
`codomain`	Target space. TYPE: `ContinuousSpace`
`hidden_dim`	Hidden layer width for neural parameter source. TYPE: `int` DEFAULT: `64`

Source code in src/quivers/continuous/families.py

def __init__(
    self,
    domain: AnySpace,
    codomain: ContinuousSpace,
    hidden_dim: int = 64,
) -> None:
    super().__init__(domain, codomain)
    d = codomain.dim
    # param_dim = d (loc) + d (raw_width)
    self.param_source = _make_source(domain, 2 * d, hidden_dim)
    self._d = d
    # The bounds are data-dependent, so we cannot pin a single
    # interval at construction time; advertise the codomain's
    # declared bounds when available, otherwise fall back to the
    # real line.
    low, high = getattr(codomain, "low", None), getattr(codomain, "high", None)
    if low is not None and high is not None:
        self._support_cache: _constraints.Constraint = _constraints.interval(
            float(low), float(high)
        )
    else:
        self._support_cache = _constraints.real

ConditionalMultivariateNormal ¶

ConditionalMultivariateNormal(domain: AnySpace, codomain: ContinuousSpace, hidden_dim: int = 64)

Bases: ContinuousMorphism

Conditional multivariate normal with full covariance.

Parameterized via Cholesky factor: the parameter source outputs loc (d values) and the lower-triangular entries of L (d*(d+1)/2 values), where Sigma = L @ L^T.

PARAMETER	DESCRIPTION
`domain`	Source space. TYPE: `SetObject or ContinuousSpace`
`codomain`	Target space (d-dimensional). TYPE: `ContinuousSpace`
`hidden_dim`	Hidden layer width for neural parameter source. TYPE: `int` DEFAULT: `64`

Source code in src/quivers/continuous/families.py

def __init__(
    self,
    domain: AnySpace,
    codomain: ContinuousSpace,
    hidden_dim: int = 64,
) -> None:
    super().__init__(domain, codomain)
    d = codomain.dim
    n_tril = d * (d + 1) // 2
    self.param_source = _make_source(domain, d + n_tril, hidden_dim)
    self._d = d
    self._n_tril = n_tril

ConditionalLowRankMVN ¶

ConditionalLowRankMVN(domain: AnySpace, codomain: ContinuousSpace, rank: int = 2, hidden_dim: int = 64)

Bases: ContinuousMorphism

Conditional low-rank multivariate normal.

Parameterized as loc + low-rank factor + diagonal: Sigma = W @ W^T + diag(d)

This is more parameter-efficient than full MVN for high dimensions.

PARAMETER	DESCRIPTION
`domain`	Source space. TYPE: `SetObject or ContinuousSpace`
`codomain`	Target space (d-dimensional). TYPE: `ContinuousSpace`
`rank`	Rank of the low-rank factor W. TYPE: `int` DEFAULT: `2`
`hidden_dim`	Hidden layer width for neural parameter source. TYPE: `int` DEFAULT: `64`

Source code in src/quivers/continuous/families.py

def __init__(
    self,
    domain: AnySpace,
    codomain: ContinuousSpace,
    rank: int = 2,
    hidden_dim: int = 64,
) -> None:
    super().__init__(domain, codomain)
    d = codomain.dim
    self._d = d
    self._rank = rank

    # loc (d) + factor (d * rank) + diag (d)
    total = d + d * rank + d
    self.param_source = _make_source(domain, total, hidden_dim)

ConditionalRelaxedBernoulli ¶

ConditionalRelaxedBernoulli(domain: AnySpace, codomain: ContinuousSpace, temperature: float = 0.5, hidden_dim: int = 64)

Bases: ContinuousMorphism

Conditional relaxed Bernoulli (concrete) distribution.

Outputs continuous values in (0, 1) that approximate Bernoulli samples. The temperature controls the relaxation: lower temperature = closer to discrete.

PARAMETER	DESCRIPTION
`domain`	Source space. TYPE: `SetObject or ContinuousSpace`
`codomain`	Target space (should be 1-d per Bernoulli component). TYPE: `ContinuousSpace`
`temperature`	Relaxation temperature. TYPE: `float` DEFAULT: `0.5`
`hidden_dim`	Hidden layer width for neural parameter source. TYPE: `int` DEFAULT: `64`

Source code in src/quivers/continuous/families.py

def __init__(
    self,
    domain: AnySpace,
    codomain: ContinuousSpace,
    temperature: float = 0.5,
    hidden_dim: int = 64,
) -> None:
    super().__init__(domain, codomain)
    d = codomain.dim
    self.param_source = _make_source(domain, d, hidden_dim)
    self._d = d
    self._temperature = temperature

ConditionalRelaxedOneHotCategorical ¶

ConditionalRelaxedOneHotCategorical(domain: AnySpace, codomain: ContinuousSpace, temperature: float = 0.5, hidden_dim: int = 64)

Bases: ContinuousMorphism

Conditional relaxed one-hot categorical (Gumbel-Softmax).

Outputs continuous vectors on the simplex that approximate one-hot categorical samples.

PARAMETER	DESCRIPTION
`domain`	Source space. TYPE: `SetObject or ContinuousSpace`
`codomain`	Target space (simplex or d-dimensional). TYPE: `ContinuousSpace`
`temperature`	Relaxation temperature. TYPE: `float` DEFAULT: `0.5`
`hidden_dim`	Hidden layer width for neural parameter source. TYPE: `int` DEFAULT: `64`

Source code in src/quivers/continuous/families.py

def __init__(
    self,
    domain: AnySpace,
    codomain: ContinuousSpace,
    temperature: float = 0.5,
    hidden_dim: int = 64,
) -> None:
    super().__init__(domain, codomain)
    d = codomain.dim
    self.param_source = _make_source(domain, d, hidden_dim)
    self._d = d
    self._temperature = temperature

ConditionalWishart ¶

ConditionalWishart(domain: AnySpace, codomain: ContinuousSpace, hidden_dim: int = 64)

Bases: ContinuousMorphism

Conditional Wishart distribution over positive-definite matrices.

Produces random d x d positive-definite matrices. Parameterized by degrees of freedom df(x) and a scale matrix V(x).

The codomain dimension is interpreted as d, and outputs are d x d matrices flattened to d^2.

PARAMETER	DESCRIPTION
`domain`	Source space. TYPE: `SetObject or ContinuousSpace`
`codomain`	Target space. dim is the matrix size d (output is d x d). TYPE: `ContinuousSpace`
`hidden_dim`	Hidden layer width for neural parameter source. TYPE: `int` DEFAULT: `64`

Source code in src/quivers/continuous/families.py

def __init__(
    self,
    domain: AnySpace,
    codomain: ContinuousSpace,
    hidden_dim: int = 64,
) -> None:
    super().__init__(domain, codomain)
    d = codomain.dim
    n_tril = d * (d + 1) // 2
    # df (1) + lower-triangular scale (n_tril)
    self.param_source = _make_source(domain, 1 + n_tril, hidden_dim)
    self._d = d
    self._n_tril = n_tril

ConditionalMatrixNormal ¶

ConditionalMatrixNormal(domain: AnySpace, codomain: ContinuousSpace, rows: int, cols: int, hidden_dim: int = 64)

Bases: ContinuousMorphism

Conditional Matrix-Normal :math:\mathcal{MN}(M, U, V).

The matrix-Normal distribution on :math:\mathbb{R}^{n \times p} factorises with a Kronecker-product covariance: if :math:X \sim \mathcal{MN}(M, U, V) then :math:\mathrm{vec}(X) \sim \mathcal{N}(\mathrm{vec}(M), V \otimes U) with :math:U \in \mathbb{R}^{n \times n} the row covariance and :math:V \in \mathbb{R}^{p \times p} the column covariance.

Categorically, the Kronecker structure is the tensor product of two Gaussians; it is strictly more constrained than the flat :math:np-dim MVN that the same parameter tensor could carry, so the surface keeps the two families distinct (no auto-substitution). Use this when the prior should express independent row and column correlation structure.

The codomain's product factorisation supplies the row and column dimensions. The grammar surface ~ MatrixNormal(loc, row_scale, col_scale) over (dom, cod) binds the first axis listed in over to the row covariance and the second to the column covariance.

PARAMETER	DESCRIPTION
`domain`	Source space. TYPE: `AnySpace`
`codomain`	Target space whose factorisation supplies `(n, p)`. Must carry a product structure of two factors; the first is the row axis, the second the column. TYPE: `ContinuousSpace`
`rows`	Row dimension :math:`n`. TYPE: `int`
`cols`	Column dimension :math:`p`. TYPE: `int`
`hidden_dim`	Hidden layer width for the parameter network. TYPE: `int` DEFAULT: `64`

Source code in src/quivers/continuous/families.py

def __init__(
    self,
    domain: AnySpace,
    codomain: ContinuousSpace,
    rows: int,
    cols: int,
    hidden_dim: int = 64,
) -> None:
    super().__init__(domain, codomain)
    self._rows = int(rows)
    self._cols = int(cols)
    n_loc = self._rows * self._cols
    n_row_tril = self._rows * (self._rows + 1) // 2
    n_col_tril = self._cols * (self._cols + 1) // 2
    self._n_loc = n_loc
    self._n_row_tril = n_row_tril
    self._n_col_tril = n_col_tril
    self.param_source = _make_source(
        domain, n_loc + n_row_tril + n_col_tril, hidden_dim
    )

ConditionalInverseWishart ¶

ConditionalInverseWishart(domain: AnySpace, codomain: ContinuousSpace, hidden_dim: int = 64)

Bases: ContinuousMorphism

Conditional Inverse-Wishart over positive-definite matrices.

Conjugate prior for the covariance of a multivariate normal. Realised as a deterministic inversion of a Wishart sample so autograd flows; equivalent in distribution to drawing :math:\Sigma^{-1} \sim \mathcal{W}(\nu, V^{-1}) and inverting. See Gelman et al. (2013) §3.6 for the conjugacy statement.

PARAMETER	DESCRIPTION
`domain`	Source space. TYPE: `AnySpace`
`codomain`	Target space whose `dim` is the matrix size :math:`d`. TYPE: `ContinuousSpace`
`hidden_dim`	Hidden layer width for the parameter network. TYPE: `int` DEFAULT: `64`

Source code in src/quivers/continuous/families.py

def __init__(
    self,
    domain: AnySpace,
    codomain: ContinuousSpace,
    hidden_dim: int = 64,
) -> None:
    super().__init__(domain, codomain)
    d = codomain.dim
    n_tril = d * (d + 1) // 2
    self._d = d
    self._n_tril = n_tril
    self.param_source = _make_source(domain, 1 + n_tril, hidden_dim)

ConditionalBernoulli ¶

ConditionalBernoulli(domain: AnySpace, codomain: AnySpace, hidden_dim: int = 64)

Bases: ContinuousMorphism

Conditional Bernoulli: continuous probability -> discrete truth value.

Takes a continuous input x and produces learnable logits that parameterize a Bernoulli distribution. The output is a discrete sample in {0, 1}, returned as a LongTensor.

This is the key bridge used in PDS (Grove & White) for the Bern x pattern, where a LogitNormal draw x in (0,1) parameterizes a Bernoulli over truth values.

The codomain must be a FinSet of size 2 (representing {False, True} or {0, 1}).

Note

Sampling from Bernoulli is NOT reparameterizable. Gradients do not flow through the discrete samples. Use score function estimators (REINFORCE) or the Gumbel-Softmax trick if differentiable samples are needed.

PARAMETER	DESCRIPTION
`domain`	Source space (typically UnitInterval or a FinSet). TYPE: `SetObject or ContinuousSpace`
`codomain`	Target FinSet of size 2. TYPE: `SetObject`
`hidden_dim`	Hidden layer width for neural parameter source. TYPE: `int` DEFAULT: `64`

Source code in src/quivers/continuous/families.py

def __init__(
    self,
    domain: AnySpace,
    codomain: AnySpace,
    hidden_dim: int = 64,
) -> None:
    from quivers.core.objects import SetObject

    if not isinstance(codomain, SetObject) or codomain.size != 2:
        raise ValueError(
            f"ConditionalBernoulli requires a FinSet(2) codomain, got {codomain!r}"
        )

    super().__init__(domain, codomain)

    # one logit per input
    self.param_source = _make_source(domain, 1, hidden_dim)

log_prob ¶

log_prob(x: Tensor, y: Tensor) -> Tensor

Log-probability of discrete output y given input x.

PARAMETER	DESCRIPTION
`x`	Input tensor. TYPE: `Tensor`
`y`	Discrete output in {0, 1}. Shape (batch,). TYPE: `Tensor`

RETURNS	DESCRIPTION
`Tensor`	Log-probabilities. Shape (batch,).

Source code in src/quivers/continuous/families.py

def log_prob(self, x: torch.Tensor, y: torch.Tensor) -> torch.Tensor:
    """Log-probability of discrete output y given input x.

    Parameters
    ----------
    x : torch.Tensor
        Input tensor.
    y : torch.Tensor
        Discrete output in {0, 1}. Shape (batch,).

    Returns
    -------
    torch.Tensor
        Log-probabilities. Shape (batch,).
    """
    probs = self._get_probs(x)
    dist = D.Bernoulli(probs=probs)
    return dist.log_prob(y.float())

rsample ¶

rsample(x: Tensor, sample_shape: Size = Size()) -> Tensor

Sample from Bernoulli (not reparameterizable).

PARAMETER	DESCRIPTION
`x`	Input tensor. TYPE: `Tensor`
`sample_shape`	Additional leading sample dimensions. TYPE: `Size` DEFAULT: `Size()`

RETURNS	DESCRIPTION
`Tensor`	Discrete samples in {0, 1}. Shape (*sample_shape, batch).

Source code in src/quivers/continuous/families.py

def rsample(
    self,
    x: torch.Tensor,
    sample_shape: torch.Size = torch.Size(),
) -> torch.Tensor:
    """Sample from Bernoulli (not reparameterizable).

    Parameters
    ----------
    x : torch.Tensor
        Input tensor.
    sample_shape : torch.Size
        Additional leading sample dimensions.

    Returns
    -------
    torch.Tensor
        Discrete samples in {0, 1}. Shape (*sample_shape, batch).
    """
    probs = self._get_probs(x)
    dist = D.Bernoulli(probs=probs)
    return dist.sample(sample_shape).long()

ConditionalCategorical ¶

ConditionalCategorical(domain: AnySpace, codomain: AnySpace, hidden_dim: int = 64)

Bases: ContinuousMorphism

Conditional Categorical: continuous input -> discrete category.

Generalizes ConditionalBernoulli to k > 2 categories. Takes a continuous input and produces learnable logits over k categories. The output is a discrete sample in {0, ..., k-1}.

PARAMETER	DESCRIPTION
`domain`	Source space. TYPE: `SetObject or ContinuousSpace`
`codomain`	Target FinSet of size k. TYPE: `SetObject`
`hidden_dim`	Hidden layer width for neural parameter source. TYPE: `int` DEFAULT: `64`

Source code in src/quivers/continuous/families.py

def __init__(
    self,
    domain: AnySpace,
    codomain: AnySpace,
    hidden_dim: int = 64,
) -> None:
    from quivers.core.objects import SetObject

    if not isinstance(codomain, SetObject):
        raise ValueError(
            f"ConditionalCategorical requires a FinSet codomain, got {codomain!r}"
        )

    super().__init__(domain, codomain)
    self._k = codomain.size
    self.param_source = _make_source(domain, self._k, hidden_dim)

support `property` ¶

support

Discrete-integer support over {0, …, k-1}.

log_prob ¶

log_prob(x: Tensor, y: Tensor) -> Tensor

Log-probability of discrete output y given input x.

PARAMETER	DESCRIPTION
`x`	Input tensor. TYPE: `Tensor`
`y`	Discrete output in {0, ..., k-1}. Shape (batch,). TYPE: `Tensor`

RETURNS	DESCRIPTION
`Tensor`	Log-probabilities. Shape (batch,).

Source code in src/quivers/continuous/families.py

def log_prob(self, x: torch.Tensor, y: torch.Tensor) -> torch.Tensor:
    """Log-probability of discrete output y given input x.

    Parameters
    ----------
    x : torch.Tensor
        Input tensor.
    y : torch.Tensor
        Discrete output in {0, ..., k-1}. Shape (batch,).

    Returns
    -------
    torch.Tensor
        Log-probabilities. Shape (batch,).
    """
    logits = self._get_logits(x)
    dist = D.Categorical(logits=logits)
    return dist.log_prob(y.long())

rsample ¶

rsample(x: Tensor, sample_shape: Size = Size()) -> Tensor

Sample from Categorical (not reparameterizable).

PARAMETER	DESCRIPTION
`x`	Input tensor. TYPE: `Tensor`
`sample_shape`	Additional leading sample dimensions. TYPE: `Size` DEFAULT: `Size()`

RETURNS	DESCRIPTION
`Tensor`	Discrete samples in {0, ..., k-1}. Shape (*sample_shape, batch).

Source code in src/quivers/continuous/families.py

def rsample(
    self,
    x: torch.Tensor,
    sample_shape: torch.Size = torch.Size(),
) -> torch.Tensor:
    """Sample from Categorical (not reparameterizable).

    Parameters
    ----------
    x : torch.Tensor
        Input tensor.
    sample_shape : torch.Size
        Additional leading sample dimensions.

    Returns
    -------
    torch.Tensor
        Discrete samples in {0, ..., k-1}. Shape (*sample_shape, batch).
    """
    logits = self._get_logits(x)
    dist = D.Categorical(logits=logits)
    return dist.sample(sample_shape).long()

ConditionalBinomial ¶

ConditionalBinomial(domain: AnySpace, codomain: ContinuousSpace, total_count: int = 1, hidden_dim: int = 64)

Bases: ContinuousMorphism

Conditional Binomial(total_count, probs(x)).

The total_count (number of trials) is a fixed hyperparameter set at construction time — typical for binomial likelihoods where n is known per observation. Only the probs parameter is learnable.

Outputs integer counts in {0, 1, ..., total_count}.

PARAMETER	DESCRIPTION
`domain`	Source space. TYPE: `SetObject or ContinuousSpace`
`codomain`	Target space. TYPE: `ContinuousSpace`
`total_count`	Number of Bernoulli trials per observation. TYPE: `int` DEFAULT: `1`
`hidden_dim`	Hidden layer width for the parameter source. TYPE: `int` DEFAULT: `64`

Source code in src/quivers/continuous/families.py

def __init__(
    self,
    domain: AnySpace,
    codomain: ContinuousSpace,
    total_count: int = 1,
    hidden_dim: int = 64,
) -> None:
    if total_count < 1:
        raise ValueError(
            f"ConditionalBinomial: total_count must be >= 1, got {total_count}"
        )
    super().__init__(domain, codomain)
    d = codomain.dim
    self._d = d
    self._total_count = int(total_count)
    self.param_source = _make_source(domain, d, hidden_dim)

ConditionalLogisticNormal ¶

ConditionalLogisticNormal(domain: AnySpace, codomain: ContinuousSpace, hidden_dim: int = 64)

Bases: ContinuousMorphism

Conditional LogisticNormal on the simplex.

Pushes a Normal(loc(x), scale(x)) draw through the softmax transform to produce a simplex-valued sample. Multivariate analogue of ConditionalLogitNormal. Useful as an alternative to ConditionalDirichlet when the underlying simplex distribution should be Gaussian in logit space rather than Beta-shaped.

PARAMETER	DESCRIPTION
`domain`	Source space. TYPE: `SetObject or ContinuousSpace`
`codomain`	Target space; `codomain.dim` is the simplex dimension. TYPE: `ContinuousSpace`
`hidden_dim`	Hidden layer width for the parameter source. TYPE: `int` DEFAULT: `64`

Source code in src/quivers/continuous/families.py

def __init__(
    self,
    domain: AnySpace,
    codomain: ContinuousSpace,
    hidden_dim: int = 64,
) -> None:
    super().__init__(domain, codomain)
    d = codomain.dim
    # We use a Normal in (d-1)-dim space and the
    # StickBreakingTransform to land on the d-simplex.
    # torch.distributions.LogisticNormal handles this.
    self.param_source = _make_source(domain, 2 * (d - 1), hidden_dim)
    self._d = d

ConditionalOneHotCategorical ¶

ConditionalOneHotCategorical(domain: AnySpace, codomain: ContinuousSpace, hidden_dim: int = 64)

Bases: ContinuousMorphism

Conditional OneHotCategorical(probs(x)).

Generalises ConditionalCategorical to one-hot encoded outputs (vector of zeros with a single one). Useful as a discrete-output observation kernel where downstream code wants a vector rather than an integer index.

PARAMETER	DESCRIPTION
`domain`	Source space. TYPE: `SetObject or ContinuousSpace`
`codomain`	Target space; `codomain.dim` is the number of categories. TYPE: `ContinuousSpace`
`hidden_dim`	Hidden layer width for the parameter source. TYPE: `int` DEFAULT: `64`

Source code in src/quivers/continuous/families.py

def __init__(
    self,
    domain: AnySpace,
    codomain: ContinuousSpace,
    hidden_dim: int = 64,
) -> None:
    super().__init__(domain, codomain)
    d = codomain.dim
    self._d = d
    self.param_source = _make_source(domain, d, hidden_dim)

ConditionalMixture ¶

ConditionalMixture(domain: AnySpace, codomain: ContinuousSpace, component_class: type, num_components: int = 4, hidden_dim: int = 64)

Bases: ContinuousMorphism

K-component mixture of a conditional family.

Wraps a single conditional family class (one of the registered ConditionalX types) and gives it K independent parameterizations plus learnable mixture logits, producing

p(y | x) = sum_k pi_k(x) * f_k(y | x)

where each f_k is an instance of the component class and pi is the softmax of K learnable logits.

Sampling is via ancestral simulation (Categorical pick + the chosen component's rsample). log_prob evaluates the log-sum-exp of the per-component log-densities. The Categorical pick is non-reparameterizable; gradient flow through the weights uses the score-function path (which higher-level objectives like IWAE can route through).

PARAMETER	DESCRIPTION
`domain`	Source space. TYPE: `SetObject or ContinuousSpace`
`codomain`	Target space; matches the component family's codomain. TYPE: `ContinuousSpace`
`component_class`	A `ConditionalX` class accepting `(domain, codomain, hidden_dim)` constructor args. TYPE: `type`
`num_components`	Number of mixture components. TYPE: `int` DEFAULT: `4`
`hidden_dim`	Hidden width for both the mixture-logit MLP and each component's parameter source. TYPE: `int` DEFAULT: `64`

Source code in src/quivers/continuous/families.py

def __init__(
    self,
    domain: AnySpace,
    codomain: ContinuousSpace,
    component_class: type,
    num_components: int = 4,
    hidden_dim: int = 64,
) -> None:
    if num_components < 2:
        raise ValueError(
            f"ConditionalMixture: num_components must be >= 2, got {num_components}"
        )
    super().__init__(domain, codomain)
    self._K = int(num_components)
    self._components = torch.nn.ModuleList(
        [component_class(domain, codomain, hidden_dim) for _ in range(self._K)]
    )
    self.mixture_logits = _make_source(domain, self._K, hidden_dim)

ConditionalIndependent ¶

ConditionalIndependent(base: ContinuousMorphism)

Bases: ContinuousMorphism

Reinterpret the trailing batch dimension of a base conditional family as an event dimension.

Equivalent to wrapping the base distribution in torch.distributions.Independent with reinterpreted_batch_ndims = 1. Used to make per-element independence explicit when a downstream guide wants to score a vector-valued observation as a single event.

PARAMETER	DESCRIPTION
`base`	The base conditional family. The wrapped distribution sums the base's per-element log-probabilities along the last axis to score a vector-valued observation. TYPE: `ContinuousMorphism`

Source code in src/quivers/continuous/families.py

def __init__(self, base: ContinuousMorphism) -> None:
    super().__init__(base.domain, base.codomain)
    self._base = base

ConditionalTransformed ¶

ConditionalTransformed(base: ContinuousMorphism, transforms: list)

Bases: ContinuousMorphism

A base conditional family composed with a chain of bijectors.

Equivalent to torch.distributions.TransformedDistribution lifted to ContinuousMorphism. The transforms are applied in forward order to rsample outputs; log_prob includes the log-determinant Jacobian correction.

PARAMETER	DESCRIPTION
`base`	Base conditional family. TYPE: `ContinuousMorphism`
`transforms`	Bijectors applied in forward order. Each must implement `__call__`, `inv`, and `log_abs_det_jacobian`. TYPE: `list of torch.distributions.Transform`

Source code in src/quivers/continuous/families.py

def __init__(
    self,
    base: ContinuousMorphism,
    transforms: list,
) -> None:
    super().__init__(base.domain, base.codomain)
    self._base = base
    self._transforms = list(transforms)

ConditionalLKJCholesky ¶

ConditionalLKJCholesky(domain: AnySpace, codomain: ContinuousSpace, hidden_dim: int = 64)

Bases: ContinuousMorphism

Conditional LKJCholesky(dim, concentration(x)).

Produces lower-triangular Cholesky factors of correlation matrices on the LKJ distribution (Lewandowski-Kurowicka-Joe 2009, doi:10.1016/j.jmva.2009.04.008). The matrix dimension is taken from codomain.dim; only the concentration parameter is learnable.

PARAMETER	DESCRIPTION
`domain`	Source space. TYPE: `SetObject or ContinuousSpace`
`codomain`	Target space; `codomain.dim` is the correlation-matrix size. TYPE: `ContinuousSpace`
`hidden_dim`	Hidden layer width for the parameter source. TYPE: `int` DEFAULT: `64`

Source code in src/quivers/continuous/families.py

def __init__(
    self,
    domain: AnySpace,
    codomain: ContinuousSpace,
    hidden_dim: int = 64,
) -> None:
    super().__init__(domain, codomain)
    self._matrix_dim = codomain.dim
    self.param_source = _make_source(domain, 1, hidden_dim)

ConditionalGaussianProcess ¶

ConditionalGaussianProcess(domain: AnySpace, codomain: ContinuousSpace, kernel: str = 'rbf', length_scale: float = 1.0, amplitude: float = 1.0, jitter: float = 1e-06)

Bases: ContinuousMorphism

Gaussian process prior with mean zero and a chosen covariance kernel.

A Gaussian process is a Markov kernel X^N -> G(R^N) whose value at a finite set of input locations x_1, ..., x_N follows a multivariate Normal with covariance matrix K(x_i, x_j). Unlike the parametric families that derive their distribution parameters from a neural network on the input, the GP's "parameters" are the input locations themselves: the kernel function evaluated on the inputs produces the covariance directly.

Reference: Rasmussen & Williams (2006), Gaussian Processes for Machine Learning.

PARAMETER	DESCRIPTION
`domain`	Source space. Its `dim` is the per-location feature dimensionality `D` of the inputs. TYPE: `SetObject or ContinuousSpace`
`codomain`	Target space. Its `dim` is the number of input locations `N` at which the GP is evaluated. TYPE: `ContinuousSpace`
`kernel`	Covariance kernel. `"rbf"` is the squared-exponential kernel; `"matern52"` is the Matern kernel with smoothness `nu = 5/2`; `"linear"` is the inner-product kernel. TYPE: `('rbf', 'matern52', 'linear')` DEFAULT: `"rbf"`
`length_scale`	Initial length scale of the kernel (positive; learnable). Ignored by the linear kernel. TYPE: `float` DEFAULT: `1.0`
`amplitude`	Initial amplitude (positive; learnable). Multiplies the kernel by `amplitude^2`. TYPE: `float` DEFAULT: `1.0`
`jitter`	Diagonal regulariser added to `K` for numerical positive-definiteness of the Cholesky factorisation. TYPE: `float` DEFAULT: `1e-06`

Source code in src/quivers/continuous/families.py

def __init__(
    self,
    domain: AnySpace,
    codomain: ContinuousSpace,
    kernel: str = "rbf",
    length_scale: float = 1.0,
    amplitude: float = 1.0,
    jitter: float = 1e-6,
) -> None:
    super().__init__(domain, codomain)
    if kernel not in _GP_KERNEL_CHOICES:
        raise ValueError(
            f"ConditionalGaussianProcess: unknown kernel {kernel!r}; "
            f"valid choices: {_GP_KERNEL_CHOICES}"
        )
    if length_scale <= 0.0:
        raise ValueError(
            f"ConditionalGaussianProcess: length_scale must be > 0, got {length_scale!r}"
        )
    if amplitude <= 0.0:
        raise ValueError(
            f"ConditionalGaussianProcess: amplitude must be > 0, got {amplitude!r}"
        )
    self._kernel = kernel
    self._jitter = jitter
    self._n = codomain.dim
    self._d = getattr(domain, "dim", None)
    # Store raw (pre-softplus) parameters so the transformed value
    # is strictly positive and the optimiser sees unconstrained
    # variables.
    inv_softplus_ls = math.log(math.expm1(length_scale))
    inv_softplus_amp = math.log(math.expm1(amplitude))
    self._raw_length_scale = torch.nn.Parameter(
        torch.tensor(inv_softplus_ls, dtype=torch.get_default_dtype())
    )
    self._raw_amplitude = torch.nn.Parameter(
        torch.tensor(inv_softplus_amp, dtype=torch.get_default_dtype())
    )

length_scale `property` ¶

length_scale: Tensor

Current (positive) length scale of the kernel.

amplitude `property` ¶

amplitude: Tensor

Current (positive) amplitude of the kernel.

ConditionalHorseshoe ¶

ConditionalHorseshoe(domain: AnySpace, codomain: ContinuousSpace, scale: float = 1.0)

Bases: ContinuousMorphism

Carvalho-Polson-Scott horseshoe prior.

The horseshoe prior places a global-local shrinkage structure on each coordinate:

.. code-block:: text

tau ~ HalfCauchy(scale) lambda_d ~ HalfCauchy(1) for d = 1, ..., D beta_d | tau, lambda_d ~ Normal(0, (tau * lambda_d)^2)

The marginal density of beta_d after integrating the local scale lambda_d has no closed form; this implementation uses a 16-point Gauss-Legendre quadrature after mapping the half-line lambda in (0, inf) to t in (0, 1) via the change of variables lambda = tan(pi * t / 2), whose Jacobian is (pi / 2) * sec^2(pi * t / 2).

Reference: Carvalho, Polson & Scott (2010), The horseshoe estimator for sparse signals.

PARAMETER	DESCRIPTION
`domain`	Source space. The prior is conditionally independent of `x`; `x` only carries the batch shape. TYPE: `SetObject or ContinuousSpace`
`codomain`	Target space. Its `dim` is the coordinate count `D`. TYPE: `ContinuousSpace`
`scale`	Initial global shrinkage `tau` (positive; learnable). TYPE: `float` DEFAULT: `1.0`

Source code in src/quivers/continuous/families.py

def __init__(
    self,
    domain: AnySpace,
    codomain: ContinuousSpace,
    scale: float = 1.0,
) -> None:
    super().__init__(domain, codomain)
    if scale <= 0.0:
        raise ValueError(f"ConditionalHorseshoe: scale must be > 0, got {scale!r}")
    self._d = codomain.dim
    inv_softplus_scale = math.log(math.expm1(scale))
    self._raw_scale = torch.nn.Parameter(
        torch.tensor(inv_softplus_scale, dtype=torch.get_default_dtype())
    )

scale `property` ¶

scale: Tensor

Current (positive) global shrinkage tau.

ConditionalGeneralizedPareto ¶

ConditionalGeneralizedPareto(domain: AnySpace, codomain: ContinuousSpace, hidden_dim: int = 64)

Bases: ContinuousMorphism

Conditional generalized Pareto distribution.

PARAMETER	DESCRIPTION
`domain`	Source space. TYPE: `SetObject or ContinuousSpace`
`codomain`	Target space. TYPE: `ContinuousSpace`
`hidden_dim`	Hidden layer width for neural parameter source. TYPE: `int` DEFAULT: `64`

Source code in src/quivers/continuous/families.py

def __init__(
    self,
    domain: AnySpace,
    codomain: ContinuousSpace,
    hidden_dim: int = 64,
) -> None:
    super().__init__(domain, codomain)
    d = codomain.dim
    # loc + scale + concentration
    self.param_source = _make_source(domain, 3 * d, hidden_dim)
    self._d = d

LKJCorrelationFactor ¶

LKJCorrelationFactor(dim: int, eta: float, domain: AnySpace)

Bases: ContinuousMorphism

LKJ prior on Cholesky factors LKJ(K, η) over CholeskyFactor(K).

Density on the Cholesky factor:

.. math::

p(L) \propto \prod_{k=2}^{K} L_{kk}^{K - k + 2(\eta - 1)}.

A higher concentration :math:\eta > 1 pulls toward the identity correlation; :math:\eta = 1 is uniform on correlations. Sampling uses the onion method of Lewandowski-Kurowicka-Joe 2009: draw row-norm partial correlations from Beta distributions and form :math:L row-by-row.

PARAMETER	DESCRIPTION
`dim`	Correlation-matrix size :math:`K \ge 2`. TYPE: `int`
`eta`	Concentration :math:`\eta > 0`. TYPE: `float`
`domain`	The morphism's source (parameter conditioning); typically the program's input space. The LKJ prior itself does not consume per-observation conditioning, so the rsample path broadcasts the prior across the batch dimension. TYPE: `AnySpace`

Source code in src/quivers/continuous/families.py

def __init__(self, dim: int, eta: float, domain: AnySpace) -> None:
    if dim < 2:
        raise ValueError(f"LKJ requires dim >= 2; got {dim}")
    if eta <= 0:
        raise ValueError(f"LKJ requires eta > 0; got {eta}")
    codomain = CholeskyFactor(name=f"L({dim})", dim=dim)
    super().__init__(domain, codomain)
    self._dim = dim
    self._eta = float(eta)

log_prob ¶

log_prob(x: Tensor, y: Tensor) -> Tensor

Log-density of the LKJ prior at the Cholesky factor y.

Up to a normalizing constant that doesn't depend on :math:L, :math:\log p(L) = \sum_{k=2}^{K} (K-k+2(\eta-1)) \log L_{kk}. The diagonal entries are extracted from the flattened representation.

Source code in src/quivers/continuous/families.py

def log_prob(self, x: torch.Tensor, y: torch.Tensor) -> torch.Tensor:
    """Log-density of the LKJ prior at the Cholesky factor ``y``.

    Up to a normalizing constant that doesn't depend on
    :math:`L`, :math:`\\log p(L) = \\sum_{k=2}^{K} (K-k+2(\\eta-1))
    \\log L_{kk}`. The diagonal entries are extracted from the
    flattened representation.
    """
    batch = y.shape[0]
    K = self._dim
    L = y.reshape(batch, K, K)
    diag = torch.diagonal(L, dim1=-2, dim2=-1)  # (batch, K)
    # Coefficients per diagonal entry (Stan's lkj_corr_cholesky_lpdf):
    # log_jac_term[k] = (K - k + 2*(eta - 1)) * log(L_kk)  for k = 2..K
    # Pre-K-indexed: power[0..K-1] where power[k] = (K-1-k) + 2*(eta-1).
    # The first diagonal is fixed at 1 so log(1)=0 contributes nothing.
    ks = torch.arange(K, device=y.device, dtype=y.dtype)
    powers = (K - 1 - ks) + 2.0 * (self._eta - 1.0)
    log_diag = torch.log(diag.clamp(min=1e-30))
    return (powers * log_diag).sum(dim=-1)

Truncated ¶

Truncated(base: ContinuousMorphism, lower: float | None = None, upper: float | None = None, max_rejection_iterations: int = 64)

Bases: ContinuousMorphism

Truncate a base family to an interval :math:[a, b].

Categorical denotation: given a base family :math:F : \Theta \to \mathcal{G}(\mathbb{R}) and constants :math:a, b \in \bar{\mathbb{R}} with :math:a < b, the truncated family has density

.. math::

p_{F_{|[a,b]}}(x) = \frac{p_F(x)}{F_{\text{cdf}}(b)
- F_{\text{cdf}}(a)} \cdot \mathbb{1}_{[a,b]}(x)

and the morphism :math:F_{|[a,b]} : \Theta \to \mathcal{G}([a,b]). Sampling uses inverse-CDF when base supports it; otherwise rejection sampling.

PARAMETER	DESCRIPTION
`base`	The base distribution-family morphism. Must expose `log_prob` and `rsample` plus an `icdf` method or a `base_distribution` torch `Distribution` for inverse-CDF sampling. Falls back to rejection sampling otherwise. TYPE: `ContinuousMorphism`
`lower`	Lower bound :math:`a`. `None` means :math:`-\infty`. TYPE: `float or None` DEFAULT: `None`
`upper`	Upper bound :math:`b`. `None` means :math:`+\infty`. TYPE: `float or None` DEFAULT: `None`
`max_rejection_iterations`	Cap on rejection-sampling attempts before raising. TYPE: `int` DEFAULT: `64`

Source code in src/quivers/continuous/families.py

def __init__(
    self,
    base: ContinuousMorphism,
    lower: float | None = None,
    upper: float | None = None,
    max_rejection_iterations: int = 64,
) -> None:
    super().__init__(base.domain, base.codomain)
    if lower is None and upper is None:
        raise ValueError(
            "Truncated requires at least one of lower / upper to be finite; "
            "without truncation, use the base family directly"
        )
    if lower is not None and upper is not None and not (lower < upper):
        raise ValueError(
            f"Truncated requires lower < upper; got lower={lower}, upper={upper}"
        )
    self._base = base
    self._lower = lower
    self._upper = upper
    self._max_iters = max_rejection_iterations
    # Attach so the parent nn.Module tracks parameters.
    self._base_mod = base

Continuous Families¶

families ¶

ConditionalNormal ¶

ConditionalLogitNormal ¶

ConditionalBeta ¶

ConditionalTruncatedNormal ¶

ConditionalDirichlet ¶

ConditionalUniform ¶

ConditionalMultivariateNormal ¶

ConditionalLowRankMVN ¶

ConditionalRelaxedBernoulli ¶

ConditionalRelaxedOneHotCategorical ¶

ConditionalWishart ¶

ConditionalMatrixNormal ¶

ConditionalInverseWishart ¶

ConditionalBernoulli ¶

log_prob ¶

rsample ¶

ConditionalCategorical ¶

support `property` ¶

log_prob ¶

rsample ¶

ConditionalBinomial ¶

ConditionalLogisticNormal ¶

ConditionalOneHotCategorical ¶

ConditionalMixture ¶

ConditionalIndependent ¶

ConditionalTransformed ¶

ConditionalLKJCholesky ¶

ConditionalGaussianProcess ¶

length_scale `property` ¶

amplitude `property` ¶

ConditionalHorseshoe ¶

scale `property` ¶

ConditionalGeneralizedPareto ¶

LKJCorrelationFactor ¶

log_prob ¶

Truncated ¶

Continuous Families¶

families ¶

ConditionalNormal ¶

ConditionalLogitNormal ¶

ConditionalBeta ¶

ConditionalTruncatedNormal ¶

ConditionalDirichlet ¶

ConditionalUniform ¶

ConditionalMultivariateNormal ¶

ConditionalLowRankMVN ¶

ConditionalRelaxedBernoulli ¶

ConditionalRelaxedOneHotCategorical ¶

ConditionalWishart ¶

ConditionalMatrixNormal ¶

ConditionalInverseWishart ¶

ConditionalBernoulli ¶

log_prob ¶

rsample ¶

ConditionalCategorical ¶

support property ¶

log_prob ¶

rsample ¶

ConditionalBinomial ¶

ConditionalLogisticNormal ¶

ConditionalOneHotCategorical ¶

ConditionalMixture ¶

ConditionalIndependent ¶

ConditionalTransformed ¶

ConditionalLKJCholesky ¶

ConditionalGaussianProcess ¶

length_scale property ¶

amplitude property ¶

ConditionalHorseshoe ¶

scale property ¶

ConditionalGeneralizedPareto ¶

LKJCorrelationFactor ¶

log_prob ¶

Truncated ¶

support `property` ¶

length_scale `property` ¶

amplitude `property` ¶

scale `property` ¶