fleqx 0.0.2__py3-none-any.whl

fleqx/__init__.py

@@ -0,0 +1,12 @@
+"""fleqx: normalizing flows for JAX, with a distreqx-native API."""
+
+from . import bijectors as bijectors
+from . import flows as flows
+from . import train as train
+
+try:  # pragma: no cover
+    import importlib.metadata
+
+    __version__ = importlib.metadata.version("fleqx")
+except importlib.metadata.PackageNotFoundError:  # pragma: no cover
+    __version__ = "0.0.1"

fleqx/bijectors/__init__.py

@@ -0,0 +1,13 @@
+"""Bijectors used to build normalizing flows.
+
+Every class here is a genuine `distreqx.bijectors.AbstractBijector`, implemented
+directly against the distreqx API with no third-party flow library involved. Most
+users should build flows with a constructor from [`fleqx.flows`][] instead; this
+module is for composing layers by hand.
+"""
+
+from ._coupling import Coupling as Coupling
+from ._invert import Invert as Invert
+from ._masked_autoregressive import MaskedAutoregressive as MaskedAutoregressive
+from ._permute import Permute as Permute
+from ._planar import Planar as Planar

fleqx/bijectors/_affine_common.py

@@ -0,0 +1,31 @@
+"""Shared helper for bijectors parameterising an elementwise affine transform."""
+
+from __future__ import annotations
+
+import jax.nn as jnn
+import jax.numpy as jnp
+from jaxtyping import Array
+
+
+def positive_scale(raw_scale: Array, min_scale: float) -> Array:
+    """Maps an unconstrained array to strictly positive values via `softplus`.
+
+    `min_scale` is added as a floor, both for numerical stability (avoiding
+    `log(scale)` blowing up as `scale -> 0`) and to guarantee positivity outright.
+    """
+    return jnn.softplus(raw_scale) + min_scale
+
+
+def identity_init_offset(min_scale: float) -> Array:
+    """Offset added to a conditioner's raw scale output before `positive_scale`.
+
+    A freshly initialised conditioner network outputs values close to zero, and
+    `positive_scale(0, min_scale)` is around 0.7, not 1 -- so without this offset, a
+    layer starts noticeably *shrinking* its input rather than close to the identity.
+    That compounds across layers (e.g. 0.7^8 ~ 0.06), making a deep flow start far
+    from the base distribution's scale and harder to train. Adding this offset before
+    `positive_scale` makes `positive_scale(0 + offset, min_scale) == 1` exactly, so a
+    freshly initialised layer starts at (approximately) the identity.
+    """
+    target = 1.0 - min_scale
+    return jnp.log(-jnp.expm1(-target)) + target

fleqx/bijectors/_coupling.py

@@ -0,0 +1,104 @@
+"""Coupling bijector."""
+
+from __future__ import annotations
+
+from collections.abc import Callable
+
+import equinox as eqx
+import jax
+import jax.numpy as jnp
+from distreqx.bijectors import (
+    AbstractBijector,
+    AbstractForwardInverseBijector,
+    AbstractFwdLogDetJacBijector,
+    AbstractInvLogDetJacBijector,
+)
+from jaxtyping import Array, PRNGKeyArray
+
+from ._affine_common import identity_init_offset, positive_scale
+
+
+class Coupling(
+    AbstractForwardInverseBijector,
+    AbstractFwdLogDetJacBijector,
+    AbstractInvLogDetJacBijector,
+):
+    """Coupling layer with an elementwise affine transform.
+
+    See [Dinh et al., 2016](https://arxiv.org/abs/1605.08803). Splits the input into
+    two parts. The first `untransformed_dim` elements are left untouched and fed
+    through an MLP conditioner to produce a per-element shift and (positive) scale,
+    which affinely transform the remaining elements.
+    """
+
+    untransformed_dim: int
+    dim: int
+    min_scale: float
+    conditioner: eqx.nn.MLP
+    _is_constant_jacobian: bool
+    _is_constant_log_det: bool
+
+    def __init__(
+        self,
+        key: PRNGKeyArray,
+        *,
+        untransformed_dim: int,
+        dim: int,
+        nn_width: int,
+        nn_depth: int,
+        nn_activation: Callable = jax.nn.relu,
+        min_scale: float = 1e-2,
+    ):
+        """Initializes a `Coupling` bijector.
+
+        **Arguments:**
+
+        - `key`: JAX random key used to initialise the conditioner MLP.
+        - `untransformed_dim`: Number of leading elements left untouched.
+        - `dim`: Total dimension.
+        - `nn_width`: Conditioner hidden layer width.
+        - `nn_depth`: Conditioner hidden layer depth.
+        - `nn_activation`: Conditioner activation function. Defaults to
+            `jax.nn.relu`.
+        - `min_scale`: Lower bound added to the transform's scale, for numerical
+            stability (also keeps the scale strictly positive). Defaults to 0.01.
+        """
+        self.untransformed_dim = untransformed_dim
+        self.dim = dim
+        self.min_scale = min_scale
+        transform_dim = dim - untransformed_dim
+        self.conditioner = eqx.nn.MLP(
+            in_size=untransformed_dim,
+            out_size=2 * transform_dim,
+            width_size=nn_width,
+            depth=nn_depth,
+            activation=nn_activation,
+            key=key,
+        )
+        self._is_constant_jacobian = False
+        self._is_constant_log_det = False
+
+    def _shift_and_scale(self, x_cond: Array) -> tuple[Array, Array]:
+        # Each dimension's (shift, raw_scale) pair is adjacent in the conditioner's
+        # output, not grouped by kind -- i.e. [shift_0, scale_0, shift_1, scale_1, ...].
+        shift, raw_scale = jnp.reshape(self.conditioner(x_cond), (-1, 2)).T
+        offset = identity_init_offset(self.min_scale)
+        return shift, positive_scale(raw_scale + offset, self.min_scale)
+
+    def forward_and_log_det(self, x: Array) -> tuple[Array, Array]:
+        """Computes y = f(x) and log|det J(f)(x)|."""
+        x_cond, x_trans = x[: self.untransformed_dim], x[self.untransformed_dim :]
+        shift, scale = self._shift_and_scale(x_cond)
+        y = jnp.concatenate([x_cond, x_trans * scale + shift])
+        return y, jnp.sum(jnp.log(scale))
+
+    def inverse_and_log_det(self, y: Array) -> tuple[Array, Array]:
+        """Computes x = f^{-1}(y) and log|det J(f^{-1})(y)|."""
+        y_cond, y_trans = y[: self.untransformed_dim], y[self.untransformed_dim :]
+        shift, scale = self._shift_and_scale(y_cond)
+        x = jnp.concatenate([y_cond, (y_trans - shift) / scale])
+        return x, -jnp.sum(jnp.log(scale))
+
+    def same_as(self, other: AbstractBijector) -> bool:
+        """Returns True if this bijector is guaranteed to be the same as `other`."""
+        return self is other

fleqx/bijectors/_invert.py

@@ -0,0 +1,62 @@
+"""Bijector inversion."""
+
+from __future__ import annotations
+
+from distreqx.bijectors import AbstractBijector
+from jaxtyping import Array
+
+
+class Invert(AbstractBijector):
+    """Inverts a bijector, swapping its forward and inverse directions.
+
+    Bijectors are usually implemented with their cheaper direction as "forward". For
+    a coupling flow, the inverse transform is used by `log_prob` and the forward
+    transform by `sample`, so wrapping the flow's bijector in `Invert` swaps which of
+    the two is fast -- useful for prioritising fast maximum-likelihood training,
+    which repeatedly evaluates `log_prob`.
+    """
+
+    bijector: AbstractBijector
+    _is_constant_jacobian: bool
+    _is_constant_log_det: bool
+
+    def __init__(self, bijector: AbstractBijector):
+        """Initializes an `Invert` bijector.
+
+        **Arguments:**
+
+        - `bijector`: The bijector to invert.
+        """
+        self.bijector = bijector
+        self._is_constant_jacobian = bijector.is_constant_jacobian
+        self._is_constant_log_det = bijector.is_constant_log_det
+
+    def forward(self, x: Array) -> Array:
+        """Computes y = f(x)."""
+        return self.bijector.inverse(x)
+
+    def inverse(self, y: Array) -> Array:
+        """Computes x = f^{-1}(y)."""
+        return self.bijector.forward(y)
+
+    def forward_log_det_jacobian(self, x: Array) -> Array:
+        """Computes log|det J(f)(x)|."""
+        return self.bijector.inverse_log_det_jacobian(x)
+
+    def inverse_log_det_jacobian(self, y: Array) -> Array:
+        """Computes log|det J(f^{-1})(y)|."""
+        return self.bijector.forward_log_det_jacobian(y)
+
+    def forward_and_log_det(self, x: Array) -> tuple[Array, Array]:
+        """Computes y = f(x) and log|det J(f)(x)|."""
+        return self.bijector.inverse_and_log_det(x)
+
+    def inverse_and_log_det(self, y: Array) -> tuple[Array, Array]:
+        """Computes x = f^{-1}(y) and log|det J(f^{-1})(y)|."""
+        return self.bijector.forward_and_log_det(y)
+
+    def same_as(self, other: AbstractBijector) -> bool:
+        """Returns True if this bijector is guaranteed to be the same as `other`."""
+        if type(other) is Invert:
+            return self.bijector.same_as(other.bijector)
+        return False

fleqx/bijectors/_masked_autoregressive.py

@@ -0,0 +1,132 @@
+"""Masked autoregressive bijector."""
+
+from __future__ import annotations
+
+from collections.abc import Callable
+
+import equinox as eqx
+import jax
+import jax.numpy as jnp
+from distreqx.bijectors import (
+    AbstractBijector,
+    AbstractForwardInverseBijector,
+    AbstractFwdLogDetJacBijector,
+    AbstractInvLogDetJacBijector,
+)
+from jaxtyping import Array, PRNGKeyArray
+
+from ._affine_common import identity_init_offset, positive_scale
+from ._masks import rank_based_mask
+
+
+class MaskedAutoregressive(
+    AbstractForwardInverseBijector,
+    AbstractFwdLogDetJacBijector,
+    AbstractInvLogDetJacBijector,
+):
+    """Masked autoregressive bijector with an elementwise affine transform.
+
+    See [Papamakarios et al., 2017](https://arxiv.org/abs/1705.07057). An MLP with
+    weights masked in a MADE-style (Germain et al., 2015) pattern maps `x` to a
+    per-element shift and (positive) scale for each dimension, such that dimension
+    `i`'s transform depends only on `x[:i]`. This makes evaluating the whole forward
+    transform (and hence `log_prob`) a single parallel pass, at the cost of the
+    inverse (hence `sample`) needing a `dim`-step sequential loop, one dimension at a
+    time -- the opposite trade-off to [`fleqx.bijectors.Coupling`][].
+    """
+
+    layers: tuple[eqx.nn.Linear, ...]
+    masks: tuple[Array, ...]
+    activation: Callable
+    dim: int
+    min_scale: float
+    _is_constant_jacobian: bool
+    _is_constant_log_det: bool
+
+    def __init__(
+        self,
+        key: PRNGKeyArray,
+        *,
+        dim: int,
+        nn_width: int,
+        nn_depth: int,
+        nn_activation: Callable = jax.nn.relu,
+        min_scale: float = 1e-2,
+    ):
+        """Initializes a `MaskedAutoregressive` bijector.
+
+        **Arguments:**
+
+        - `key`: JAX random key used to initialise the conditioner MLP.
+        - `dim`: Total dimension.
+        - `nn_width`: Conditioner hidden layer width.
+        - `nn_depth`: Conditioner hidden layer depth.
+        - `nn_activation`: Conditioner activation function. Defaults to
+            `jax.nn.relu`.
+        - `min_scale`: Lower bound added to the transform's scale, for numerical
+            stability (also keeps the scale strictly positive). Defaults to 0.01.
+        """
+        mlp = eqx.nn.MLP(
+            in_size=dim,
+            out_size=2 * dim,
+            width_size=nn_width,
+            depth=nn_depth,
+            activation=nn_activation,
+            key=key,
+        )
+        in_ranks = jnp.arange(dim)
+        # For dim==1 this would divide by zero; MAF has nothing to condition on
+        # there regardless (all weights end up masked out below).
+        hidden_ranks = jnp.arange(nn_width) % max(dim - 1, 1)
+        out_ranks = jnp.repeat(jnp.arange(dim), 2)
+        ranks = [in_ranks, *([hidden_ranks] * nn_depth), out_ranks]
+
+        self.layers = mlp.layers
+        self.masks = tuple(
+            rank_based_mask(ranks[i], ranks[i + 1], eq=i != len(mlp.layers) - 1)
+            for i in range(len(mlp.layers))
+        )
+        self.activation = nn_activation
+        self.dim = dim
+        self.min_scale = min_scale
+        self._is_constant_jacobian = False
+        self._is_constant_log_det = False
+
+    def _conditioner(self, x: Array) -> Array:
+        h = x
+        for i, (layer, mask) in enumerate(zip(self.layers, self.masks, strict=True)):
+            h = jnp.where(mask, layer.weight, 0.0) @ h + layer.bias
+            if i < len(self.layers) - 1:
+                h = self.activation(h)
+        return h
+
+    def _shift_and_scale(self, params: Array) -> tuple[Array, Array]:
+        shift, raw_scale = jnp.reshape(params, (self.dim, 2)).T
+        offset = identity_init_offset(self.min_scale)
+        return shift, positive_scale(raw_scale + offset, self.min_scale)
+
+    def forward_and_log_det(self, x: Array) -> tuple[Array, Array]:
+        """Computes y = f(x) and log|det J(f)(x)|."""
+        shift, scale = self._shift_and_scale(self._conditioner(x))
+        y = x * scale + shift
+        return y, jnp.sum(jnp.log(scale))
+
+    def inverse_and_log_det(self, y: Array) -> tuple[Array, Array]:
+        """Computes x = f^{-1}(y) and log|det J(f^{-1})(y)|.
+
+        Sequential in `dim`: recovering `x[i]` needs `x[:i]`, which is only
+        available once earlier iterations have filled it in.
+        """
+
+        def step(x: Array, i: Array) -> tuple[Array, None]:
+            shift, scale = self._shift_and_scale(self._conditioner(x))
+            x = x.at[i].set((y[i] - shift[i]) / scale[i])
+            return x, None
+
+        x, _ = jax.lax.scan(step, jnp.zeros_like(y), jnp.arange(self.dim))
+        _, scale = self._shift_and_scale(self._conditioner(x))
+        return x, -jnp.sum(jnp.log(scale))
+
+    def same_as(self, other: AbstractBijector) -> bool:
+        """Returns True if this bijector is guaranteed to be the same as `other`."""
+        return self is other

fleqx/bijectors/_masks.py

@@ -0,0 +1,26 @@
+"""Masks used to build autoregressive neural networks."""
+
+from __future__ import annotations
+
+import operator
+
+import jax.numpy as jnp
+from jaxtyping import Array, Bool, Int
+
+
+def rank_based_mask(
+    in_ranks: Int[Array, " a"], out_ranks: Int[Array, " b"], *, eq: bool = False
+) -> Bool[Array, "b a"]:
+    """A mask with entries `out_ranks > in_ranks` (or `>=`, if `eq`).
+
+    Used to build a MADE-style masked MLP (Germain et al., 2015): zeroing a weight
+    matrix with this mask ensures the layer's output at rank `r` only depends on
+    inputs with a strictly lower rank, which is what makes the resulting network
+    autoregressive.
+
+    **Returns:**
+
+    - A boolean array of shape `(len(out_ranks), len(in_ranks))`.
+    """
+    op = operator.ge if eq else operator.gt
+    return op(out_ranks[:, None], in_ranks)

fleqx/bijectors/_permute.py

@@ -0,0 +1,57 @@
+"""Fixed-permutation bijector."""
+
+from __future__ import annotations
+
+import jax.numpy as jnp
+from distreqx.bijectors import (
+    AbstractBijector,
+    AbstractForwardInverseBijector,
+    AbstractFwdLogDetJacBijector,
+    AbstractInvLogDetJacBijector,
+)
+from jaxtyping import Array
+
+
+class Permute(
+    AbstractForwardInverseBijector,
+    AbstractFwdLogDetJacBijector,
+    AbstractInvLogDetJacBijector,
+):
+    """Reorders the elements of a vector according to a fixed permutation.
+
+    A permutation has a constant Jacobian determinant of magnitude 1, so it
+    contributes nothing to a flow's log-density -- it exists purely to let later
+    layers mix over dimensions left untouched by earlier ones.
+    """
+
+    permutation: Array
+    inverse_permutation: Array
+    _is_constant_jacobian: bool
+    _is_constant_log_det: bool
+
+    def __init__(self, permutation: Array):
+        """Initializes a `Permute` bijector.
+
+        **Arguments:**
+
+        - `permutation`: A 1-D integer array containing a permutation of
+            `arange(len(permutation))`.
+        """
+        self.permutation = jnp.asarray(permutation)
+        self.inverse_permutation = jnp.argsort(self.permutation)
+        self._is_constant_jacobian = True
+        self._is_constant_log_det = True
+
+    def forward_and_log_det(self, x: Array) -> tuple[Array, Array]:
+        """Computes y = f(x) and log|det J(f)(x)|."""
+        return x[self.permutation], jnp.zeros(())
+
+    def inverse_and_log_det(self, y: Array) -> tuple[Array, Array]:
+        """Computes x = f^{-1}(y) and log|det J(f^{-1})(y)|."""
+        return y[self.inverse_permutation], jnp.zeros(())
+
+    def same_as(self, other: AbstractBijector) -> bool:
+        """Returns True if this bijector is guaranteed to be the same as `other`."""
+        if type(other) is Permute:
+            return self.permutation is other.permutation
+        return False

fleqx/bijectors/_planar.py

@@ -0,0 +1,95 @@
+"""Planar bijector."""
+
+from __future__ import annotations
+
+import jax
+import jax.numpy as jnp
+import jax.random as jr
+from distreqx.bijectors import (
+    AbstractBijector,
+    AbstractForwardInverseBijector,
+    AbstractFwdLogDetJacBijector,
+    AbstractInvLogDetJacBijector,
+)
+from jaxtyping import Array, PRNGKeyArray
+
+
+class Planar(
+    AbstractForwardInverseBijector,
+    AbstractFwdLogDetJacBijector,
+    AbstractInvLogDetJacBijector,
+):
+    r"""Planar bijector: $y = x + u \cdot \text{leaky\_relu}(w^T x + b)$.
+
+    See [Rezende and Mohamed, 2015](https://arxiv.org/abs/1505.05770). The original
+    paper uses a $\tanh$ activation, but that has no closed-form inverse; this uses a
+    leaky ReLU instead (as the paper's appendix suggests), which does, at the cost of
+    slightly less smooth transformed densities.
+    """
+
+    weight: Array
+    act_scale: Array
+    bias: Array
+    negative_slope: float
+    _is_constant_jacobian: bool
+    _is_constant_log_det: bool
+
+    def __init__(self, key: PRNGKeyArray, *, dim: int, negative_slope: float = 1e-2):
+        """Initializes a `Planar` bijector.
+
+        **Arguments:**
+
+        - `key`: JAX random key used to initialise `w`, `u` and `b`.
+        - `dim`: Dimension of the bijection.
+        - `negative_slope`: The leaky ReLU's negative slope, in `(0, 1)`. Defaults to
+            0.01.
+        """
+        if not 0 < negative_slope < 1:
+            raise ValueError("`negative_slope` must be in (0, 1).")
+        w_key, u_key, b_key = jr.split(key, 3)
+        self.weight = 1e-2 * jr.normal(w_key, (dim,))
+        self.act_scale = 1e-2 * jr.normal(u_key, (dim,))
+        self.bias = 1e-2 * jr.normal(b_key, ())
+        self.negative_slope = negative_slope
+        self._is_constant_jacobian = False
+        self._is_constant_log_det = False
+
+    def _invertible_act_scale(self) -> Array:
+        """Constrains `u` so that `w^T u >= -1`, guaranteeing invertibility.
+
+        See Appendix A.1 of Rezende and Mohamed, 2015.
+        """
+        wtu = self.act_scale @ self.weight
+        m_wtu = -1 + jnp.log1p(jax.nn.softplus(wtu))
+        return self.act_scale + (m_wtu - wtu) * self.weight / jnp.sum(self.weight**2)
+
+    def forward_and_log_det(self, x: Array) -> tuple[Array, Array]:
+        """Computes y = f(x) and log|det J(f)(x)|."""
+        u = self._invertible_act_scale()
+        pre_act = x @ self.weight + self.bias
+        act = jax.nn.leaky_relu(pre_act, negative_slope=self.negative_slope)
+        y = x + u * act
+        slope = jnp.where(pre_act < 0, self.negative_slope, 1.0)
+        log_det = jnp.log(jnp.abs(1 + u @ (slope * self.weight)))
+        return y, log_det
+
+    def inverse_and_log_det(self, y: Array) -> tuple[Array, Array]:
+        """Computes x = f^{-1}(y) and log|det J(f^{-1})(y)|.
+
+        Derivation: let $z = w^Tx + b$. Since $x = y - u \\cdot \\text{lrelu}(z)$,
+        substituting into $z = w^Tx + b$ and solving gives
+        $z = (w^Ty + b) / (1 + w^Tus)$, where $s$ is the leaky ReLU's slope at $z$
+        (found from the sign of $w^Ty + b$, since the denominator is positive).
+        """
+        u = self._invertible_act_scale()
+        numerator = self.weight @ y + self.bias
+        slope = jnp.where(numerator < 0, self.negative_slope, 1.0)
+        us = u * slope
+        denominator = 1 + self.weight @ us
+        x = y - us * (numerator / denominator)
+        log_det = -jnp.log(jnp.abs(1 + us @ self.weight))
+        return x, log_det
+
+    def same_as(self, other: AbstractBijector) -> bool:
+        """Returns True if this bijector is guaranteed to be the same as `other`."""
+        return self is other

fleqx/flows.py

@@ -0,0 +1,227 @@
+"""Constructors for normalizing-flow distributions.
+
+Each constructor returns a plain `distreqx.distributions.Transformed` -- a base
+distribution pushed through a bijector, no flow-specific wrapper class. As with any
+distreqx distribution, `log_prob` and `sample` take a single event; use `jax.vmap`
+for batches, e.g. `jax.vmap(flow.log_prob)(xs)`.
+"""
+
+from __future__ import annotations
+
+from collections.abc import Callable
+
+import jax.numpy as jnp
+import jax.random as jr
+from distreqx.bijectors import AbstractBijector, Chain, ScalarAffine
+from distreqx.distributions import Independent, Normal, Transformed
+from jax.nn import relu
+from jaxtyping import Array, Float, PRNGKeyArray
+
+from .bijectors import Coupling, Invert, MaskedAutoregressive, Permute, Planar
+
+__all__ = ["coupling_flow", "masked_autoregressive_flow", "planar_flow"]
+
+
+def _default_permute(dim: int, key: PRNGKeyArray) -> AbstractBijector | None:
+    """A permutation to mix dimensions between layers.
+
+    `None` for `dim == 1`. For `dim == 2`, always swaps rather than drawing a random
+    permutation, since a random permutation of two elements is the identity half the
+    time.
+    """
+    if dim == 1:
+        return None
+    if dim == 2:
+        return Permute(jnp.array([1, 0]))
+    return Permute(jr.permutation(key, dim))
+
+
+def _standardizing_bijector(data: Float[Array, "n dim"]) -> ScalarAffine:
+    loc = jnp.mean(data, axis=0)
+    scale = jnp.std(data, axis=0)
+    return ScalarAffine(shift=loc, scale=scale)
+
+
+def _finalize(
+    bijector: AbstractBijector, *, dim: int, data: Float[Array, "n dim"] | None
+) -> Transformed:
+    if data is not None:
+        bijector = Chain([_standardizing_bijector(jnp.asarray(data)), bijector])
+    base = Independent(Normal(loc=jnp.zeros(dim), scale=jnp.ones(dim)))
+    return Transformed(base, bijector)
+
+
+def _stack_layers(
+    key: PRNGKeyArray, flow_layers: int, make_layer: Callable[[PRNGKeyArray], AbstractBijector]
+) -> AbstractBijector:
+    layer_keys = jr.split(key, flow_layers)
+    layers = [make_layer(k) for k in layer_keys]
+    return Chain(list(reversed(layers)))
+
+
+def coupling_flow(
+    key: PRNGKeyArray,
+    *,
+    dim: int,
+    flow_layers: int = 8,
+    nn_width: int = 50,
+    nn_depth: int = 1,
+    nn_activation: Callable = relu,
+    invert: bool = True,
+    data: Float[Array, "n dim"] | None = None,
+) -> Transformed:
+    """Coupling flow ([Dinh et al., 2016](https://arxiv.org/abs/1605.08803)).
+
+    A stack of affine coupling layers ([`fleqx.bijectors.Coupling`][]), each followed
+    by a permutation, applied to a learnable diagonal-Gaussian base distribution.
+
+    **Arguments:**
+
+    - `key`: JAX random key.
+    - `dim`: Dimensionality of the distribution.
+    - `flow_layers`: Number of coupling layers. Defaults to 8.
+    - `nn_width`: Conditioner hidden layer width. Defaults to 50.
+    - `nn_depth`: Conditioner depth. Defaults to 1.
+    - `nn_activation`: Conditioner activation function. Defaults to `jax.nn.relu`.
+    - `invert`: If `True` (default), `log_prob` is the fast direction; if `False`,
+        `sample` is.
+    - `data`: Optional array of shape `(n, dim)`. If given, an extra affine
+        layer maps the flow's output to `data`'s mean and standard deviation
+        from the start, rather than the base distribution learning this during
+        training. Gives training a head start; useful when training for few
+        epochs. Defaults to `None`.
+
+    **Returns:**
+
+    A `distreqx.distributions.Transformed` distribution.
+    """
+    if dim < 1:
+        raise ValueError(f"`dim` must be a positive integer, got {dim}.")
+
+    def make_layer(k: PRNGKeyArray) -> AbstractBijector:
+        bij_key, perm_key = jr.split(k)
+        coupling = Coupling(
+            bij_key,
+            untransformed_dim=dim // 2,
+            dim=dim,
+            nn_width=nn_width,
+            nn_depth=nn_depth,
+            nn_activation=nn_activation,
+        )
+        permute = _default_permute(dim, perm_key)
+        return coupling if permute is None else Chain([permute, coupling])
+
+    bijector = _stack_layers(key, flow_layers, make_layer)
+    if invert:
+        bijector = Invert(bijector)
+    return _finalize(bijector, dim=dim, data=data)
+
+
+def masked_autoregressive_flow(
+    key: PRNGKeyArray,
+    *,
+    dim: int,
+    flow_layers: int = 8,
+    nn_width: int = 50,
+    nn_depth: int = 1,
+    nn_activation: Callable = relu,
+    invert: bool = True,
+    data: Float[Array, "n dim"] | None = None,
+) -> Transformed:
+    """Masked autoregressive flow ([Papamakarios et al., 2017](https://arxiv.org/abs/1705.07057)).
+
+    A stack of masked autoregressive layers ([`fleqx.bijectors.MaskedAutoregressive`][]),
+    each followed by a permutation, applied to a learnable diagonal-Gaussian base
+    distribution.
+
+    Unlike [`coupling_flow`][fleqx.flows.coupling_flow], the two transform directions
+    genuinely differ in cost: one is a single parallel pass, the other a `dim`-step
+    sequential loop. `invert` picks which direction `log_prob` gets.
+
+    **Arguments:**
+
+    - `key`: JAX random key.
+    - `dim`: Dimensionality of the distribution.
+    - `flow_layers`: Number of masked autoregressive layers. Defaults to 8.
+    - `nn_width`: Conditioner hidden layer width. Defaults to 50.
+    - `nn_depth`: Conditioner depth. Defaults to 1.
+    - `nn_activation`: Conditioner activation function. Defaults to `jax.nn.relu`.
+    - `invert`: If `True` (default), `log_prob` is the fast (parallel) direction --
+        recommended for maximum-likelihood training. If `False`, `sample` is fast
+        instead.
+    - `data`: Optional array of shape `(n, dim)`. If given, an extra affine
+        layer maps the flow's output to `data`'s mean and standard deviation
+        from the start, rather than the base distribution learning this during
+        training. Gives training a head start; useful when training for few
+        epochs. Defaults to `None`.
+
+    **Returns:**
+
+    A `distreqx.distributions.Transformed` distribution.
+    """
+    if dim < 1:
+        raise ValueError(f"`dim` must be a positive integer, got {dim}.")
+
+    def make_layer(k: PRNGKeyArray) -> AbstractBijector:
+        bij_key, perm_key = jr.split(k)
+        maf = MaskedAutoregressive(
+            bij_key,
+            dim=dim,
+            nn_width=nn_width,
+            nn_depth=nn_depth,
+            nn_activation=nn_activation,
+        )
+        permute = _default_permute(dim, perm_key)
+        return maf if permute is None else Chain([permute, maf])
+
+    bijector = _stack_layers(key, flow_layers, make_layer)
+    if invert:
+        bijector = Invert(bijector)
+    return _finalize(bijector, dim=dim, data=data)
+
+
+def planar_flow(
+    key: PRNGKeyArray,
+    *,
+    dim: int,
+    flow_layers: int = 8,
+    negative_slope: float = 1e-2,
+    invert: bool = True,
+    data: Float[Array, "n dim"] | None = None,
+) -> Transformed:
+    """Planar flow ([Rezende and Mohamed, 2015](https://arxiv.org/abs/1505.05770)).
+
+    A stack of planar layers ([`fleqx.bijectors.Planar`][]) applied to a learnable
+    diagonal-Gaussian base distribution. No permutation is needed between layers,
+    since each planar layer already depends on every dimension.
+
+    **Arguments:**
+
+    - `key`: JAX random key.
+    - `dim`: Dimensionality of the distribution.
+    - `flow_layers`: Number of planar layers. Defaults to 8.
+    - `negative_slope`: Negative slope of the leaky ReLU used within each layer (see
+        [`fleqx.bijectors.Planar`][]), in `(0, 1)`. Defaults to 0.01.
+    - `invert`: If `True` (default), `log_prob` is the fast direction; if `False`,
+        `sample` is.
+    - `data`: Optional array of shape `(n, dim)`. If given, an extra affine
+        layer maps the flow's output to `data`'s mean and standard deviation
+        from the start, rather than the base distribution learning this during
+        training. Gives training a head start; useful when training for few
+        epochs. Defaults to `None`.
+
+    **Returns:**
+
+    A `distreqx.distributions.Transformed` distribution.
+    """
+    if dim < 1:
+        raise ValueError(f"`dim` must be a positive integer, got {dim}.")
+
+    def make_layer(k: PRNGKeyArray) -> AbstractBijector:
+        return Planar(k, dim=dim, negative_slope=negative_slope)
+
+    bijector = _stack_layers(key, flow_layers, make_layer)
+    if invert:
+        bijector = Invert(bijector)
+    return _finalize(bijector, dim=dim, data=data)
+

fleqx/train/__init__.py

@@ -0,0 +1,4 @@
+"""Training utilities for fleqx distributions."""
+
+from ._losses import MaximumLikelihoodLoss as MaximumLikelihoodLoss
+from ._loops import fit as fit

fleqx/train/_loops.py

@@ -0,0 +1,118 @@
+"""Training loops."""
+
+from __future__ import annotations
+
+from collections.abc import Callable
+
+import equinox as eqx
+import jax.numpy as jnp
+import jax.random as jr
+import optax
+from distreqx.distributions import AbstractDistribution
+from jaxtyping import Array, Float, PRNGKeyArray, Scalar
+from tqdm import tqdm
+
+from ._losses import MaximumLikelihoodLoss
+from ._train_utils import count_fruitless, get_batches, step, train_val_split
+
+
+def fit(
+    key: PRNGKeyArray,
+    dist: AbstractDistribution,
+    data: Float[Array, "n dim"],
+    *,
+    loss_fn: Callable[..., Scalar] | None = None,
+    learning_rate: float = 5e-4,
+    optimizer: optax.GradientTransformation | None = None,
+    max_epochs: int = 100,
+    max_patience: int = 5,
+    batch_size: int = 100,
+    val_prop: float = 0.1,
+    return_best: bool = True,
+    show_progress: bool = True,
+) -> tuple[AbstractDistribution, dict[str, list[float]]]:
+    """Fit a distribution to samples by minimising a loss (maximum likelihood by default).
+
+    A held-out validation split is used for early stopping: training stops once
+    `max_patience` consecutive epochs pass without an improvement in validation loss.
+    The last batch in each epoch is dropped if it would be truncated, to avoid
+    recompilation.
+
+    **Arguments:**
+
+    - `key`: JAX random key controlling the train/validation split, shuffling, and
+        (if applicable) any stochasticity in `loss_fn`.
+    - `dist`: The distribution to train (as returned by e.g.
+        [`fleqx.flows.coupling_flow`][]).
+    - `data`: Array of observations with shape ``(n, dim)``.
+    - `loss_fn`: Loss with signature ``(params, static, x, key)``. Defaults to
+        [`fleqx.train.MaximumLikelihoodLoss`][].
+    - `learning_rate`: Adam learning rate. Ignored if `optimizer` is given.
+    - `optimizer`: An optax optimizer. Defaults to ``optax.adam(learning_rate)``.
+    - `max_epochs`: Maximum number of passes over the data. Defaults to 100.
+    - `max_patience`: Number of consecutive epochs with no validation-loss
+        improvement after which training stops early. Defaults to 5.
+    - `batch_size`: Mini-batch size. Defaults to 100.
+    - `val_prop`: Proportion of `data` held out for validation. Defaults to 0.1.
+    - `return_best`: Whether to return the parameters from the epoch with the lowest
+        validation loss (`True`), or the parameters after the final update (`False`).
+        Defaults to `True`.
+    - `show_progress`: Whether to display a progress bar. Defaults to `True`.
+
+    **Returns:**
+
+    - A tuple ``(trained_dist, losses)``, where ``losses`` is a dict with ``"train"``
+        and ``"val"`` keys, each holding the mean loss per epoch.
+    """
+    if loss_fn is None:
+        loss_fn = MaximumLikelihoodLoss()
+    if optimizer is None:
+        optimizer = optax.adam(learning_rate)
+
+    data = (jnp.asarray(data),)
+    params, static = eqx.partition(dist, eqx.is_inexact_array)
+    best_params = params
+    opt_state = optimizer.init(params)
+
+    key, subkey = jr.split(key)
+    train_data, val_data = train_val_split(subkey, data, val_prop=val_prop)
+    losses: dict[str, list[float]] = {"train": [], "val": []}
+
+    loop = tqdm(range(max_epochs), disable=not show_progress)
+    for _ in loop:
+        key, *subkeys = jr.split(key, 3)
+        train_data = [jr.permutation(subkeys[0], a) for a in train_data]
+        val_data = [jr.permutation(subkeys[1], a) for a in val_data]
+
+        batch_losses = []
+        for batch in zip(*get_batches(train_data, batch_size), strict=True):
+            key, subkey = jr.split(key)
+            params, opt_state, loss_i = step(
+                params,
+                static,
+                *batch,
+                optimizer=optimizer,
+                opt_state=opt_state,
+                loss_fn=loss_fn,
+                key=subkey,
+            )
+            batch_losses.append(loss_i)
+        losses["train"].append((sum(batch_losses) / len(batch_losses)).item())
+
+        batch_losses = []
+        for batch in zip(*get_batches(val_data, batch_size), strict=True):
+            key, subkey = jr.split(key)
+            loss_i = eqx.filter_jit(loss_fn)(params, static, *batch, key=subkey)
+            batch_losses.append(loss_i)
+        losses["val"].append((sum(batch_losses) / len(batch_losses)).item())
+
+        loop.set_postfix({k: v[-1] for k, v in losses.items()})
+        if losses["val"][-1] == min(losses["val"]):
+            best_params = params
+        elif count_fruitless(losses["val"]) > max_patience:
+            loop.set_postfix_str(f"{loop.postfix} (Max patience reached)")
+            break
+
+    params = best_params if return_best else params
+    trained = eqx.combine(params, static)
+    return trained, losses

fleqx/train/_losses.py

@@ -0,0 +1,32 @@
+"""Loss functions for training distributions."""
+
+from __future__ import annotations
+
+import equinox as eqx
+import jax
+from distreqx.distributions import AbstractDistribution
+from jaxtyping import Array, Float, PRNGKeyArray, PyTree
+
+
+class MaximumLikelihoodLoss:
+    """Negative log-likelihood loss for fitting a distribution by maximum likelihood.
+
+    The call signature ``(params, static, x, key)`` matches what
+    [`fleqx.train.fit`][] expects: ``params`` and ``static`` are the two halves of an
+    ``equinox.partition`` of the distribution, ``x`` is a batch of observations, and
+    ``key`` is accepted (and ignored) for API consistency with stochastic losses.
+    """
+
+    @eqx.filter_jit
+    def __call__(
+        self,
+        params: PyTree,
+        static: PyTree,
+        x: Float[Array, "batch dim"],
+        key: PRNGKeyArray | None = None,
+    ) -> Float[Array, ""]:
+        """Return the mean negative log-likelihood of ``x`` under the distribution."""
+        dist: AbstractDistribution = eqx.combine(params, static)
+        # `log_prob` takes a single, unbatched event -- like any distreqx
+        # distribution -- so batches of `x` must be vmapped explicitly.
+        return -jax.vmap(dist.log_prob)(x).mean()

fleqx/train/_train_utils.py

@@ -0,0 +1,71 @@
+"""Generic training utilities: batching, splitting, and an optax step.
+
+Nothing here is specific to distreqx or to flows -- these operate on any pytree and
+loss function.
+"""
+
+from __future__ import annotations
+
+from collections.abc import Callable, Sequence
+from functools import partial
+
+import equinox as eqx
+import jax.numpy as jnp
+import jax.random as jr
+import optax
+from jax import jit
+from jaxtyping import Array, PRNGKeyArray, PyTree, Scalar
+
+
+@eqx.filter_jit
+def step(
+    params: PyTree,
+    *args,
+    optimizer: optax.GradientTransformation,
+    opt_state: PyTree,
+    loss_fn: Callable[..., Scalar],
+    **kwargs,
+) -> tuple[PyTree, PyTree, Scalar]:
+    """A single optimisation step."""
+    loss_val, grads = eqx.filter_value_and_grad(loss_fn)(params, *args, **kwargs)
+    updates, opt_state = optimizer.update(grads, opt_state, params=params)
+    params = eqx.apply_updates(params, updates)
+    return params, opt_state, loss_val
+
+
+def train_val_split(
+    key: PRNGKeyArray,
+    arrays: Sequence[Array],
+    val_prop: float = 0.1,
+) -> tuple[list[Array], list[Array]]:
+    """Random train/validation split for a sequence of arrays sharing axis-0 size."""
+    if not 0 <= val_prop <= 1:
+        raise ValueError("val_prop should be between 0 and 1.")
+    num_samples = arrays[0].shape[0]
+    n_train = num_samples - round(val_prop * num_samples)
+    arrays = [jr.permutation(key, a) for a in arrays]
+    train_arrays = [arr[:n_train] for arr in arrays]
+    val_arrays = [arr[n_train:] for arr in arrays]
+    return train_arrays, val_arrays
+
+
+@partial(jit, static_argnums=1)
+def get_batches(arrays: Sequence[Array], batch_size: int) -> tuple[Array, ...]:
+    """Reshape arrays with shape ``(n, ...)`` to ``(n // batch_size, batch_size, ...)``.
+
+    The trailing partial batch is dropped if truncated (to avoid recompilation), and
+    `batch_size` is capped at the array length.
+    """
+    return tuple(_add_batch(arr, batch_size) for arr in arrays)
+
+
+def _add_batch(arr: Array, batch_size: int) -> Array:
+    batch_size = min(batch_size, arr.shape[0])
+    n_batches = arr.shape[0] // batch_size
+    return arr[: n_batches * batch_size].reshape(n_batches, batch_size, *arr.shape[1:])
+
+
+def count_fruitless(losses: list[float]) -> int:
+    """Number of epochs since the minimum loss in a list of losses."""
+    min_idx = jnp.argmin(jnp.array(losses)).item()
+    return len(losses) - min_idx - 1

fleqx-0.0.2.dist-info/METADATA

@@ -0,0 +1,92 @@
+Metadata-Version: 2.4
+Name: fleqx
+Version: 0.0.2
+Summary: Normalizing flows for JAX, with a distreqx-native API.
+Author: fleqx contributors
+License: Apache-2.0
+Project-URL: homepage, https://gvcallen.github.io/fleqx
+Project-URL: documentation, https://gvcallen.github.io/fleqx
+Project-URL: repository, https://github.com/gvcallen/fleqx
+Keywords: jax,normalizing-flows,distreqx,equinox,machine-learning
+Classifier: Development Status :: 3 - Alpha
+Classifier: Intended Audience :: Science/Research
+Classifier: License :: OSI Approved :: Apache Software License
+Classifier: Programming Language :: Python :: 3
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Classifier: Topic :: Scientific/Engineering :: Mathematics
+Requires-Python: >=3.11
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: jax>=0.7.0
+Requires-Dist: equinox>=0.11
+Requires-Dist: jaxtyping
+Requires-Dist: optax
+Requires-Dist: tqdm
+Requires-Dist: distreqx>=0.0.3
+Provides-Extra: test
+Requires-Dist: pytest; extra == "test"
+Requires-Dist: jax[cpu]; extra == "test"
+Provides-Extra: docs
+Requires-Dist: mkdocs>=1.6; extra == "docs"
+Requires-Dist: mkdocs-material>=9.5; extra == "docs"
+Requires-Dist: mkdocstrings[python]>=0.26; extra == "docs"
+Requires-Dist: pymdown-extensions>=10.7; extra == "docs"
+Dynamic: license-file
+
+# fleqx
+
+Normalizing flows for JAX, built directly on [distreqx](https://github.com/lockwo/distreqx).
+
+A flow from `fleqx` is a plain `distreqx.distributions.Transformed` — there's no
+flow-specific wrapper, so `log_prob`, `sample`, `optax` training, etc. all work
+exactly as they would for any other distreqx distribution. As with any distreqx
+distribution, batches are handled with `jax.vmap` rather than by passing in arrays
+with a leading batch axis.
+
+Three flow types are implemented so far: coupling, masked autoregressive, and planar.
+
+## Example
+
+```python
+import jax.numpy as jnp
+import jax.random as jr
+
+import fleqx
+
+key = jr.key(0)
+flow = fleqx.flows.coupling_flow(key, dim=2)
+
+sample = flow.sample(jr.key(1))
+log_p = flow.log_prob(sample)
+
+data = jr.normal(jr.key(2), (1000, 2))
+flow, losses = fleqx.train.fit(jr.key(3), flow, data)
+```
+
+## Installation
+
+```
+pip install fleqx
+```
+
+See the [documentation](https://gvcallen.github.io/fleqx) for the full API.
+
+## Acknowledgements
+
+Built on [distreqx](https://github.com/lockwo/distreqx) (Owen Lockwood). The
+bijectors were ported from [flowjax](https://github.com/danielward27/flowjax)
+(Daniel Ward), a more complete flows library that's worth using directly if you don't
+specifically need the distreqx API.
+
+## Development
+
+```
+pip install -e ".[test]"
+pytest
+```
+
+---
+
+This library was written by Claude, porting the bijectors directly from flowjax.
+Behaviour should be nearly identical to flowjax's, though minor differences may
+remain, and the code hasn't yet had a full human review.

fleqx-0.0.2.dist-info/RECORD

@@ -0,0 +1,20 @@
+fleqx/__init__.py,sha256=AVibg38rRDU1xNlFZQIQuANPbvN1B3bbijueh2qaQVA,369
+fleqx/flows.py,sha256=sVXJCl0vCOu3CQEbOCg56zDpJTB7nldERNCmfoTzO4c,8407
+fleqx/py.typed,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+fleqx/bijectors/__init__.py,sha256=D-fQ8bFuxSEUT83LrX5Kpi0pcBTe3NfG48tlBIbOKZI,578
+fleqx/bijectors/_affine_common.py,sha256=IgWHBNi6JhmH5esB496JYrjzaJOLY5XAbuRMfvBMfwQ,1351
+fleqx/bijectors/_coupling.py,sha256=WHvW8LWYebmQHcIyzBTx7RE_7vs-vVKEtSIGtoiQu7g,3812
+fleqx/bijectors/_invert.py,sha256=2UT29cWx0qFRMW7JSurcD_nYt15fwe-67ONHUBgPhDk,2243
+fleqx/bijectors/_masked_autoregressive.py,sha256=ivkSMGMsH3g58-PCPH9Dq1yDmG1g48OFKqjqPwXlxK0,4887
+fleqx/bijectors/_masks.py,sha256=PxZ6QVOCj6YSuDWONsojjc3z1FqvAODVci1vxNBxh10,821
+fleqx/bijectors/_permute.py,sha256=4cNbaz4fbVnw9UR3ZN0zwqZDS_T6SLb9hLIro1f6bOc,1889
+fleqx/bijectors/_planar.py,sha256=HE466iWfBEm9I84NPLICVBphczZj1uTPtJrljj5qeHc,3620
+fleqx/train/__init__.py,sha256=PKOv80KWtYiy8MjbjwCbFzSGLJRcAuHI5jqG6S95Xio,150
+fleqx/train/_loops.py,sha256=obj6951wSvAHj_9IV0HZXec9uDKlQZmPJqcBKbRHHaQ,4720
+fleqx/train/_losses.py,sha256=hFvgXFTVNo-_BrFj9wQDn8SIsz01GBn-3w85TyylPwQ,1231
+fleqx/train/_train_utils.py,sha256=WOmWgDGhVMVkB0uR36QwOB1I7VLcPxglj4PQ-6OUbRc,2409
+fleqx-0.0.2.dist-info/licenses/LICENSE,sha256=9z5yyVJQq21eh9ss3dEtU4t_x6Z9cFfE7JXB0-jYUvo,11348
+fleqx-0.0.2.dist-info/METADATA,sha256=egzx0qh7NG6JraA9MObRc_m6dwgM7zIzWK7mJ6QQA3s,2885
+fleqx-0.0.2.dist-info/WHEEL,sha256=aeYiig01lYGDzBgS8HxWXOg3uV61G9ijOsup-k9o1sk,91
+fleqx-0.0.2.dist-info/top_level.txt,sha256=OPrd8lMWcMxFUMlAtrWPASjjTfLEFMoWmF6F_xHbLJ0,6
+fleqx-0.0.2.dist-info/RECORD,,

fleqx-0.0.2.dist-info/WHEEL

@@ -0,0 +1,5 @@
+Wheel-Version: 1.0
+Generator: setuptools (82.0.1)
+Root-Is-Purelib: true
+Tag: py3-none-any
+

fleqx-0.0.2.dist-info/licenses/LICENSE

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright 2026 fleqx contributors
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

fleqx-0.0.2.dist-info/top_level.txt

@@ -0,0 +1 @@
+fleqx