zarr-python-reshape

A reshape codec for zarr-python implementing the Zarr reshape codec extension.

Installation

pip install zarr-reshape

Usage

import zarr
import zarr_reshape  # Registers the codec

# Create an array with the reshape codec
# This reshapes chunks from (100, 50, 64, 3) to (5000, 64, 3)
arr = zarr.create_array(
    store="my_array.zarr",
    shape=(1000, 500, 64, 3),
    chunks=(100, 50, 64, 3),
    dtype="float32",
    filters=[
        zarr_reshape.ReshapeCodec(shape=[[0, 1], [2], 3])
    ],
)

Shape Configuration

The shape configuration parameter specifies how to compute the output shape from the input shape. Each element can be:

• A positive integer: An explicit size for that dimension.
• A list of integers: Input dimension indices whose product determines the output dimension size.
• The value -1: Inferred automatically to satisfy prod(output_shape) == prod(input_shape). May occur at most once.

Examples

For an input chunk of shape (100, 50, 64, 3):

Shape Configuration	Output Shape	Description
[[0, 1], [2], 3]	(5000, 64, 3)	Merge first two dimensions
[10, 10, -1]	(10, 10, 96000)	Fixed first two, infer third
[[0], [1], [2], [3]]	(100, 50, 64, 3)	Identity reshape
[-1]	(960000,)	Flatten to 1D

Constraints

• Input dimensions referenced in lists must be in strictly monotonically increasing order.
• The invariant prod(output_shape) == prod(input_shape) must be satisfiable.

AI statement

Claude Code was used in the initial creation of this package.