`downsample` Driver¶

Virtual read-only view that performs downsampling.

Downsampling is performed on-the-fly to compute exactly the positions of the downsampled view that are read.

json driver/downsample : object¶

Extends:¶

TensorStore — Specifies a TensorStore to open/create.

Required members:¶

driver : "downsample"¶

base : TensorStore¶

Base TensorStore to downsample.

Downsampling is always performed with respect to an origin of 0 in all dimensions in the base TensorStore. To use a different origin, translate the domain by specifying an IndexTransform via the TensorStore.transform property of base.

Translating the domain of the downsampled view by a given offset is equivalent to translating the domain of base by offset * downsample_factors. However, there is no translation of the downsampled view that is exactly equivalent to translating the domain of base by an offset that is not a multiple of downsample_factors.

downsample_factors : array of integer_{[1, +∞)}¶

Factors by which to downsample each dimension of base. The length must match the rank of base. Specifying a factor of 1 indicates not to downsample a given dimension.

Example

[2, 2]

downsample_method : DownsampleMethod¶

Optional members:¶

context : Context¶: Specifies context resources that augment/override the parent context.

dtype : dtype¶: Specifies the data type.

rank : integer_{[0, 32]}¶: Specifies the rank of the TensorStore.

If transform is also specified, the input rank must match. Otherwise, the rank constraint applies to the driver directly.

transform : IndexTransform¶: Specifies a transform.

schema : Schema¶: Specifies constraints on the schema.

When opening an existing array, specifies constraints on the existing schema; opening will fail if the constraints do not match. Any soft constraints specified in the chunk_layout are ignored. When creating a new array, a suitable schema will be selected automatically based on the specified schema constraints in combination with any driver-specific constraints.

Example

{
  "driver": "downsample",
  "downsample_factors": [1, 2, 2],
  "downsample_method": "mean",
  "base": {"driver": "zarr", "kvstore": {"driver": "gcs", "bucket": "my-bucket"}}
}

The following downsampling methods are supported:

Downsample method

One of:¶

"stride"¶

Downsampling is performed by striding: reading a given position position in the downsampled view is exactly equivalent to reading position * downsample_factors in the original view. This method supports all data types and is quite efficient (there is no additional buffering of data). It is equivalent to applying a striding index transform to driver/downsample.base; while it offers no additional functionality, it offers uniformity with other methods and may be simpler to specify.

Note that striding is always with respect to an origin of 0 in all dimensions. For example, downsampling a 1-dimensional TensorStore with a factor of 3 reads from positions 0, 3, 6, etc. To use a different offset, translate the domain of driver/downsample.base.

"median"¶: Downsampling computes the median value: reading a given position in the downsampled view returns the median value within the rectangular region [position * downsample_factors, (position + 1) * downsample_factors) of the original view. Only non-complex numeric data types and bool are supported. The computed median value is always a value present in the original data. In the case that the rectangular region contains an odd number of elements, the lower value is always used. For example, the median computed for [4.0, 3.0, 1.0, 2.0] is 2.0.

"mode"¶: Downsampling computes the mode (most frequent value): reading a given position in the downsampled view returns the mode within the rectangular region [position * downsample_factors, (position + 1) * downsample_factors) of the original view. All data types are supported. In the case of a tie, for non-complex numeric data types and bool, the lowest value (among the set of common frequent values) is chosen. For strings, the lexicographically smallest value is chosen. For other data types, an unspecified (but deterministic for a given version of TensorStore) most frequent value is chosen.

"mean"¶

Downsampling is performed by arithmetic mean (averaging): reading a given position in the downsampled view returns the arithmetic mean within the rectangular region :python`[position * downsample_factors, (position + 1) * downsample_factors)` of the original view, where the average is computed only over positions that are within the bounds of the original view. Only numeric data types and "bool" are supported. For integer data types, the means are computed exactly and rounded to the nearest integer, and rounded to even in the case of a tie. For the "bool" data type, the mode is chosen.

Note that the blocks used for averaging start with an origin of 0 in all dimensions. For example, downsampling a 1-dimensional TensorStore with a downsample factor of 3 averages the intervals [0, 3), [3, 6), [6, 9), etc. To use a different offset, translate the domain of driver/downsample.base.

"min"¶: Downsampling computes the minimum value: reading a given position in the downsampled view returns the minimum value within the rectangular region [position * downsample_factors, (position + 1) * downsample_factors) of the original view. Only non-complex numeric data types and bool are supported.

"max"¶: Downsampling computes the maximum value: reading a given position in the downsampled view returns the maximum value within the rectangular region [position * downsample_factors, (position + 1) * downsample_factors) of the original view. Only non-complex numeric data types and bool are supported.

downsample Driver¶

`downsample` Driver¶