# Copyright OTT-JAX
#
# 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
#
# https://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.
import dataclasses
import functools
from typing import Any, Iterable, Iterator, Literal, Optional, Tuple
import jax
import jax.random as jr
from ott.geometry import costs, pointcloud
from ott.solvers import linear
__all__ = ["LinearOTDataloader"]
[docs]
@dataclasses.dataclass(frozen=False, repr=False)
class LinearOTDataloader:
"""Linear OT dataloader.
This dataloader wraps a dataloader that generates ``(source, target)``
arrays with shape ``[batch, ...]`` and aligns them
using the :class:`~ott.solvers.linear.sinkhorn.Sinkhorn` algorithm.
Args:
rng: Random number generator.
dataset: Iterable dataset which yields a tuple of source and target arrays
of shape ``[batch, ...]``.
epsilon: Epsilon regularization.
See :class:`~ott.geometry.geometry.Geometry` for more information.
relative_epsilon: Whether ``epsilon`` refers to a fraction of the
:attr:`~ott.geometry.pointcloud.PointCloud.mean_cost_matrix` or
:attr:`~ott.geometry.pointcloud.PointCloud.std_cost_matrix`.
cost_fn: Cost function between two points.
threshold: Convergence threshold for
:class:`~ott.solvers.linear.sinkhorn.Sinkhorn`.
max_iterations: Maximum number of Sinkhorn iterations.
replace: Whether to sample with replacement.
shardings: Input and output shardings for the source and target arrays.
"""
rng: jax.Array
dataset: Iterable[Tuple[jax.Array, jax.Array]]
epsilon: Optional[float] = None
relative_epsilon: Optional[Literal["mean", "std"]] = None
cost_fn: Optional[costs.CostFn] = None
threshold: float = 1e-3
max_iterations: int = 2000
replace: bool = True
shardings: Optional[jax.sharding.Sharding] = None
def __post_init__(self) -> None:
self._align_fn = jax.jit(
functools.partial(
_align,
threshold=self.threshold,
max_iterations=self.max_iterations
),
static_argnames=["cost_fn", "epsilon", "relative_epsilon", "replace"],
in_shardings=(None, self.shardings, self.shardings),
out_shardings=(self.shardings, self.shardings),
)
self._data_it: Optional[Iterator[Tuple[jax.Array, jax.Array]]] = None
self._rng_it: Optional[jax.Array] = None
def __iter__(self) -> "LinearOTDataloader":
"""Return self."""
self._data_it = iter(self.dataset)
self._rng_it = self.rng
return self
def __next__(self) -> Tuple[jax.Array, jax.Array]:
"""Align source and target samples in a batch.
Returns:
The aligned source and target arrays of shape ``[batch, ...]``.
"""
assert self._data_it is not None, "Please call `iter()` first."
assert self._rng_it is not None, "Please call `iter()` first."
self._rng_it, rng_sample = jr.split(self._rng_it, 2)
x, y = next(self._data_it)
return self._align_fn(
rng_sample,
x,
y,
self.cost_fn,
self.epsilon,
self.relative_epsilon,
self.replace,
)
def _align(
rng: jax.Array,
x: jax.Array,
y: jax.Array,
cost_fn: costs.CostFn,
epsilon: Optional[float],
relative_epsilon: Optional[Literal["mean", "std"]],
replace: bool,
**kwargs: Any,
) -> Tuple[jax.Array, jax.Array]:
geom = pointcloud.PointCloud(
x,
y,
cost_fn=cost_fn,
epsilon=epsilon,
relative_epsilon=relative_epsilon,
)
out = linear.solve(geom, **kwargs)
n, m = geom.shape
probs = out.matrix.ravel()
probs = probs / probs.sum()
ixs = jr.choice(rng, n * m, shape=(n,), p=probs, replace=replace)
row_ixs, col_ixs = ixs // m, ixs % m
return x[row_ixs], y[col_ixs]
