ngio 0.5.0b6__py3-none-any.whl

ngio/common/_pyramid.py

@@ -0,0 +1,408 @@
+from collections.abc import Callable, Mapping, Sequence
+from typing import Any, Literal
+
+import dask.array as da
+import numpy as np
+import zarr
+from pydantic import BaseModel, ConfigDict, model_validator
+
+from ngio.common._zoom import (
+    InterpolationOrder,
+    _zoom_inputs_check,
+    dask_zoom,
+    numpy_zoom,
+)
+from ngio.utils import (
+    NgioValueError,
+)
+
+
+def _on_disk_numpy_zoom(
+    source: zarr.Array,
+    target: zarr.Array,
+    order: InterpolationOrder,
+) -> None:
+    source_array = source[...]
+    if not isinstance(source_array, np.ndarray):
+        raise NgioValueError("source zarr array could not be read as a numpy array")
+    target[...] = numpy_zoom(source_array, target_shape=target.shape, order=order)
+
+
+def _on_disk_dask_zoom(
+    source: zarr.Array,
+    target: zarr.Array,
+    order: InterpolationOrder,
+) -> None:
+    source_array = da.from_zarr(source)
+    target_array = dask_zoom(source_array, target_shape=target.shape, order=order)
+
+    # This is a potential fix for Dask 2025.11
+    # import dask.config
+    # chunk_size_bytes = np.prod(target.chunks) * target_array.dtype.itemsize
+    # current_chunk_size = dask.config.get("array.chunk-size")
+    # Increase the chunk size to avoid dask potentially creating
+    # corrupted chunks when writing chunks that are not multiple of the
+    # target chunk size
+    # dask.config.set({"array.chunk-size": f"{chunk_size_bytes}B"})
+    target_array = target_array.rechunk(target.chunks)
+    target_array = target_array.compute_chunk_sizes()
+    target_array.to_zarr(target)
+    # Restore previous chunk size
+    # dask.config.set({"array.chunk-size": current_chunk_size})
+
+
+def _on_disk_coarsen(
+    source: zarr.Array,
+    target: zarr.Array,
+    order: InterpolationOrder = "linear",
+    aggregation_function: Callable | None = None,
+) -> None:
+    """Apply a coarsening operation from a source zarr array to a target zarr array.
+
+    Args:
+        source (zarr.Array): The source array to coarsen.
+        target (zarr.Array): The target array to save the coarsened result to.
+        order (InterpolationOrder): The order of interpolation is not really implemented
+            for coarsening, but it is kept for compatibility with the zoom function.
+            order="linear" -> linear interpolation ~ np.mean
+            order="nearest" -> nearest interpolation ~ np.max
+        aggregation_function (np.ufunc): The aggregation function to use.
+    """
+    source_array = da.from_zarr(source)
+
+    _scale, _target_shape = _zoom_inputs_check(
+        source_array=source_array, scale=None, target_shape=target.shape
+    )
+
+    assert _target_shape == target.shape, (
+        "Target shape must match the target array shape"
+    )
+
+    if aggregation_function is None:
+        if order == "linear":
+            aggregation_function = np.mean
+        elif order == "nearest":
+            aggregation_function = np.max
+        elif order == "cubic":
+            raise NgioValueError("Cubic interpolation is not supported for coarsening.")
+        else:
+            raise NgioValueError(
+                f"Aggregation function must be provided for order {order}"
+            )
+
+    coarsening_setup = {}
+    for i, s in enumerate(_scale):
+        coarsening_setup[i] = int(np.round(1 / s))
+
+    out_target = da.coarsen(
+        aggregation_function, source_array, coarsening_setup, trim_excess=True
+    )
+    out_target = out_target.rechunk(target.chunks)
+    out_target.to_zarr(target)
+
+
+def on_disk_zoom(
+    source: zarr.Array,
+    target: zarr.Array,
+    order: InterpolationOrder = "linear",
+    mode: Literal["dask", "numpy", "coarsen"] = "dask",
+) -> None:
+    """Apply a zoom operation from a source zarr array to a target zarr array.
+
+    Args:
+        source (zarr.Array): The source array to zoom.
+        target (zarr.Array): The target array to save the zoomed result to.
+        order (InterpolationOrder): The order of interpolation. Defaults to "linear".
+        mode (Literal["dask", "numpy", "coarsen"]): The mode to use. Defaults to "dask".
+    """
+    if not isinstance(source, zarr.Array):
+        raise NgioValueError("source must be a zarr array")
+
+    if not isinstance(target, zarr.Array):
+        raise NgioValueError("target must be a zarr array")
+
+    if source.dtype != target.dtype:
+        raise NgioValueError("source and target must have the same dtype")
+
+    match mode:
+        case "numpy":
+            return _on_disk_numpy_zoom(source, target, order)
+        case "dask":
+            return _on_disk_dask_zoom(source, target, order)
+        case "coarsen":
+            return _on_disk_coarsen(
+                source,
+                target,
+            )
+        case _:
+            raise NgioValueError("mode must be either 'dask', 'numpy' or 'coarsen'")
+
+
+def _find_closest_arrays(
+    processed: list[zarr.Array], to_be_processed: list[zarr.Array]
+) -> tuple[np.intp, np.intp]:
+    dist_matrix = np.zeros((len(processed), len(to_be_processed)))
+    for i, arr_to_proc in enumerate(to_be_processed):
+        for j, proc_arr in enumerate(processed):
+            dist_matrix[j, i] = np.sqrt(
+                np.sum(
+                    [
+                        (s1 - s2) ** 2
+                        for s1, s2 in zip(
+                            arr_to_proc.shape, proc_arr.shape, strict=False
+                        )
+                    ]
+                )
+            )
+
+    indices = np.unravel_index(dist_matrix.argmin(), dist_matrix.shape)
+    assert len(indices) == 2, "Indices must be of length 2"
+    return indices
+
+
+def consolidate_pyramid(
+    source: zarr.Array,
+    targets: list[zarr.Array],
+    order: InterpolationOrder = "linear",
+    mode: Literal["dask", "numpy", "coarsen"] = "dask",
+) -> None:
+    """Consolidate the Zarr array."""
+    processed = [source]
+    to_be_processed = targets
+
+    while to_be_processed:
+        source_id, target_id = _find_closest_arrays(processed, to_be_processed)
+
+        source_image = processed[source_id]
+        target_image = to_be_processed.pop(target_id)
+
+        on_disk_zoom(
+            source=source_image,
+            target=target_image,
+            mode=mode,
+            order=order,
+        )
+        processed.append(target_image)
+
+
+################################################
+#
+# Builders for image pyramids
+#
+################################################
+
+ChunksLike = tuple[int, ...] | Literal["auto"]
+ShardsLike = tuple[int, ...] | Literal["auto"]
+
+
+def shapes_from_scaling_factors(
+    base_shape: tuple[int, ...],
+    scaling_factors: tuple[float, ...],
+    num_levels: int,
+) -> list[tuple[int, ...]]:
+    """Compute the shapes of each level in the pyramid from scaling factors.
+
+    Args:
+        base_shape (tuple[int, ...]): The shape of the base level.
+        scaling_factors (tuple[float, ...]): The scaling factors between levels.
+        num_levels (int): The number of levels in the pyramid.
+
+    Returns:
+        list[tuple[int, ...]]: The shapes of each level in the pyramid.
+    """
+    shapes = []
+    current_shape = base_shape
+    for _ in range(num_levels):
+        shapes.append(current_shape)
+        current_shape = tuple(
+            max(1, int(s / f))
+            for s, f in zip(current_shape, scaling_factors, strict=True)
+        )
+    return shapes
+
+
+def _check_order(shapes: Sequence[tuple[int, ...]]):
+    """Check if the shapes are in decreasing order."""
+    num_pixels = [np.prod(shape) for shape in shapes]
+    for i in range(1, len(num_pixels)):
+        if num_pixels[i] >= num_pixels[i - 1]:
+            raise NgioValueError("Shapes are not in decreasing order.")
+
+
+class PyramidLevel(BaseModel):
+    path: str
+    shape: tuple[int, ...]
+    scale: tuple[float, ...]
+    chunks: ChunksLike = "auto"
+    shards: ShardsLike | None = None
+
+    @model_validator(mode="after")
+    def _model_validation(self) -> "PyramidLevel":
+        # Same length as shape
+        if len(self.scale) != len(self.shape):
+            raise NgioValueError(
+                "Scale must have the same length as shape "
+                f"({len(self.shape)}), got {len(self.scale)}"
+            )
+        if any(isinstance(s, float) and s < 0 for s in self.scale):
+            raise NgioValueError("Scale values must be positive.")
+
+        if isinstance(self.chunks, tuple):
+            if len(self.chunks) != len(self.shape):
+                raise NgioValueError(
+                    "Chunks must have the same length as shape "
+                    f"({len(self.shape)}), got {len(self.chunks)}"
+                )
+            normalized_chunks = []
+            for dim_size, chunk_size in zip(self.shape, self.chunks, strict=True):
+                normalized_chunks.append(min(dim_size, chunk_size))
+            self.chunks = tuple(normalized_chunks)
+
+        if isinstance(self.shards, tuple):
+            if len(self.shards) != len(self.shape):
+                raise NgioValueError(
+                    "Shards must have the same length as shape "
+                    f"({len(self.shape)}), got {len(self.shards)}"
+                )
+            normalized_shards = []
+            for dim_size, shard_size in zip(self.shape, self.shards, strict=True):
+                normalized_shards.append(min(dim_size, shard_size))
+            self.shards = tuple(normalized_shards)
+        return self
+
+
+class ImagePyramidBuilder(BaseModel):
+    levels: list[PyramidLevel]
+    axes: tuple[str, ...]
+    data_type: str = "uint16"
+    dimension_separator: Literal[".", "/"] = "/"
+    compressors: Any = "auto"
+    zarr_format: Literal[2, 3] = 2
+    other_array_kwargs: Mapping[str, Any] = {}
+
+    model_config = ConfigDict(arbitrary_types_allowed=True)
+
+    @classmethod
+    def from_scaling_factors(
+        cls,
+        levels_paths: tuple[str, ...],
+        scaling_factors: tuple[float, ...],
+        base_shape: tuple[int, ...],
+        base_scale: tuple[float, ...],
+        axes: tuple[str, ...],
+        chunks: ChunksLike = "auto",
+        shards: ShardsLike | None = None,
+        data_type: str = "uint16",
+        dimension_separator: Literal[".", "/"] = "/",
+        compressors: Any = "auto",
+        zarr_format: Literal[2, 3] = 2,
+        other_array_kwargs: Mapping[str, Any] | None = None,
+    ) -> "ImagePyramidBuilder":
+        shapes = shapes_from_scaling_factors(
+            base_shape=base_shape,
+            scaling_factors=scaling_factors,
+            num_levels=len(levels_paths),
+        )
+        return cls.from_shapes(
+            shapes=shapes,
+            base_scale=base_scale,
+            axes=axes,
+            levels_paths=levels_paths,
+            chunks=chunks,
+            shards=shards,
+            data_type=data_type,
+            dimension_separator=dimension_separator,
+            compressors=compressors,
+            zarr_format=zarr_format,
+            other_array_kwargs=other_array_kwargs,
+        )
+
+    @classmethod
+    def from_shapes(
+        cls,
+        shapes: Sequence[tuple[int, ...]],
+        base_scale: tuple[float, ...],
+        axes: tuple[str, ...],
+        levels_paths: Sequence[str] | None = None,
+        chunks: ChunksLike = "auto",
+        shards: ShardsLike | None = None,
+        data_type: str = "uint16",
+        dimension_separator: Literal[".", "/"] = "/",
+        compressors: Any = "auto",
+        zarr_format: Literal[2, 3] = 2,
+        other_array_kwargs: Mapping[str, Any] | None = None,
+    ) -> "ImagePyramidBuilder":
+        levels = []
+        if levels_paths is None:
+            levels_paths = tuple(str(i) for i in range(len(shapes)))
+        _check_order(shapes)
+        scale_ = base_scale
+        for i, (path, shape) in enumerate(zip(levels_paths, shapes, strict=True)):
+            levels.append(
+                PyramidLevel(
+                    path=path,
+                    shape=shape,
+                    scale=scale_,
+                    chunks=chunks,
+                    shards=shards,
+                )
+            )
+            if i + 1 < len(shapes):
+                # This only works for downsampling pyramids
+                # The _check_order function ensures that
+                # shapes are decreasing
+                next_shape = shapes[i + 1]
+                scaling_factor = tuple(
+                    s1 / s2
+                    for s1, s2 in zip(
+                        shape,
+                        next_shape,
+                        strict=True,
+                    )
+                )
+                scale_ = tuple(
+                    s * f for s, f in zip(scale_, scaling_factor, strict=True)
+                )
+        other_array_kwargs = other_array_kwargs or {}
+        return cls(
+            levels=levels,
+            axes=axes,
+            data_type=data_type,
+            dimension_separator=dimension_separator,
+            compressors=compressors,
+            zarr_format=zarr_format,
+            other_array_kwargs=other_array_kwargs,
+        )
+
+    def to_zarr(self, group: zarr.Group) -> None:
+        """Save the pyramid specification to a Zarr group.
+
+        Args:
+            group (zarr.Group): The Zarr group to save the pyramid specification to.
+        """
+        array_static_kwargs = {
+            "dtype": self.data_type,
+            "overwrite": True,
+            "compressors": self.compressors,
+            **self.other_array_kwargs,
+        }
+
+        if self.zarr_format == 2:
+            array_static_kwargs["chunk_key_encoding"] = {
+                "name": "v2",
+                "separator": self.dimension_separator,
+            }
+        else:
+            array_static_kwargs["chunk_key_encoding"] = {
+                "name": "default",
+                "separator": self.dimension_separator,
+            }
+            array_static_kwargs["dimension_names"] = self.axes
+        for p_level in self.levels:
+            group.create_array(
+                name=p_level.path,
+                shape=tuple(p_level.shape),
+                chunks=p_level.chunks,
+                shards=p_level.shards,
+                **array_static_kwargs,
+            )