Tiling

The satrain.tiling module provides functionality for tiling datasets and managing spatial data organization.

ipwgml.tiling

Provides functionality for tiling input data and assembling tiled results.

class DatasetTiler(dataset: Dataset, tile_size: int | None = 512, overlap: int = 32, spatial_dims: Tuple[str, str] | None = None)

Bases: object

This tiler provides functionality for tiling an xarray.Dataset into equal-sized tiles.

Parameters:

• dataset – List of input tensors for the retrieval.

• tile_size – The size of a single tile. If this is None the tiler returns a single tile extending over the full spatial extent of the dataset.

• overlap – The overlap between two subsequent tiles.

• spatial_dims – A tuple containing the names of the spatial dimensions along which to tile the dataset. If not set, will use the two lattermost dimensions in the datset.

assemble(tiles)

Assemble slices back to original shape using linear interpolation in overlap regions.

Parameters:

tiles – A list of lists containing the results for each tile.

Returns:

The data in ‘tiles’ reassembled along the last to dimensions to match the initial dimenions of the input.

assemble_tile(row_index, col_index, results, results_t)

Assembles results from a single tile into the assembled result containers in ‘results’.

Parameters:

• row_index – The row index identifying the current tile.

• col_index – The column index identifying the current tile.

• results – Container for the assembled results.

• results_t – Results for the current tile.

get_slices(row_ind: int, col_ind) → Tuple[slice, slice]

Return slices for the clipping of the tiles.

Parameters:

• row_ind – The 0-based row index of the tile.

• col_ind – The 0-based column index of the tile.

Returns:

Tuple of slices that can be used to clip the retrieval results to obtain non-overlapping tiles.

get_tile(row_ind: int, col_ind: int) → Dataset

Get tile in the ‘row_ind’th row and ‘col_ind’th column of the two dimensional tiling.

Parameters:

• row_ind – The 0-based row index of the tile.

• col_ind – The 0-based column index of the tile.

Returns:

An xarray.Dataset containing the requested tile.

get_weights(row_ind: int, col_ind, like: ndarray | None = None) → ndarray

Get weights to reassemble results.

Parameters:

• row_ind – Row-index of the tile.

• col_ind – Column-index of the tile.

• like – An optional numpy.ndarray to infer the dtype of the results.

Returns:

Numpy array containing weights for the corresponding tile.

initialize_results(results_t)

Initialize containers for assembled results from the results from the first tile.

Parameters:

results_t – Retrieval results returned from the first tile.

Returns:

Depending of the structure of ‘results_t’, a single numpy.ndarray, or a (potentially nested) list or dict of numpy.ndarrays.

predict(predict_fun)

Applies a prediction function to all tiles in the input and assembles the results.

Parameters:

predict_fun – A callable that takes the input from a single tile and returns the corresponding predicted results.

Returns:

The tile-wise results from ‘predict_fun’ assembled to the original size.

get_starts_and_clips(extent: int, tile_size: int, overlap: int) → Tuple[List[int], List[int]]

Calculate start indices and numbers of clipped pixels for a given side length, tile size and overlap.

Parameters:

• extent – The extent of the dimension to tile.

• tile_size – The size of each tile.

• overlap – The number of pixels of overlap.

• soft_end – Allow the last tile to go beyond n, see notes for details

Returns:

A tuple (start, clip) containing the start indices of each tile and the number of pixels to clip between each neighboring tiles.