ScatterElementsUpdate#

Versioned name: ScatterElementsUpdate-12

Category: Data movement

Short description: Creates a copy of the first input tensor with elements from updates input applied according to the logic specified by reduction attribute and indices along the axis.

Detailed description: Creates copy of the first input tensor, and applies the elements of updates according to the logic specified by reduction attribute. For each element of updates, at the same index there is corresponding value of indices, which is an index along dimension specified by axis. The index for dimension pointed by axis is provided by values of indices input, otherwise, the index is the same as the index of the entry itself.

The dimensions of updates tensor are allowed to be less or equal to the corresponding dimensions of data tensor, but the dimension pointed by axis can be also greater (especially if the indices input contains duplicated values).

The operation to perform between the corresponding elements is specified by reduction attribute, by default the elements of data tensor are simply overwritten by the values from updates input.

Additionally, use_init_val attribute can be used to control whether the elements from the data input tensor are used as initial value (enabled by default).

General logic of output values calculations is presented below for 1D tensor case, the element corresponding to the [i] is performed as:

output[indices[i]] = reduction(updates[i], output[indices[i]]), axis = 0

  • Overwrite without additional operation, reduction = “none”

output[indices[i]] = updates[i], axis = 0

  • Update by adding corresponding elements, reduction = “sum”

output[indices[i]] += updates[i], axis = 0

  • Update by multiplication of the corresponding elements, reduction = “prod”

output[indices[i]] *= updates[i], axis = 0

  • Update with minimum value of the corresponding elements, reduction = “min”

output[indices[i]] = min(updates[i], output[indices[i]]) axis = 0

  • Update with maximum value of the corresponding elements, reduction = “max”

output[indices[i]] = max(updates[i], output[indices[i]]) axis = 0

  • Update with mean value of the corresponding elements, reduction = “mean”. For integer types the calculated mean is rounded down (towards negative infinity). This reduction type is not supported for the boolean data type.

output[indices[i]] = mean(updates[i], output[indices[i]]) axis = 0

For 2D tensor case, the update of the element corresponding to the [i][j] is performed as:

output[indices[i][j]][j] = reduction(updates[i][j], output[indices[i][j]][j]) if axis = 0
output[i][indices[i][j]] = reduction(updates[i][j], output[indices[i][j]][j]) if axis = 1

Accordingly for 3D tensor case, the update of the element corresponding to the [i][j][k] is performed as:

output[indices[i][j][k]][j][k] = reduction(updates[i][j][k], output[indices[i][j][k]][j][k]) if axis = 0
output[i][indices[i][j][k]][k] = reduction(updates[i][j][k], output[i][indices[i][j][k]][k]) if axis = 1
output[i][j][indices[i][j][k]] = reduction(updates[i][j][k], output[i][j][indices[i][j][k]]) if axis = 2

Attributes:

  • reduction

    • Description: The type of operation to perform on the inputs.

    • Range of values: one of none, sum, prod, min, max, mean

    • Type: string

    • Default value: none

    • Required: no

  • use_init_val

    • Description: Controls whether the elements in the data input tensor are used as init value for reduce operations.

    • Range of values: * true - data input elements are used * false - data input elements are not used

    • Type: boolean

    • Default value: true

    • Required: no

    • Note: The attribute has no effect for reduction == “none”

Inputs:

  • 1: data tensor of arbitrary rank r and of type T. Required.

  • 2: indices tensor with indices of type T_IND. The rank of the tensor is equal to the rank of data tensor. All index values are expected to be within bounds [-d, d - 1] along dimension d pointed by axis. If multiple indices point to the same output location then the order of updating the values is undefined. Negative value of index means reverse indexing and will be normalized to value len(data.shape[axis] + index). If an index points to non-existing element then exception is raised. Required.

  • 3: updates tensor of shape equal to the shape of indices tensor and of type T. Required.

  • 4: axis tensor with scalar or 1D tensor with one element of type T_AXIS specifying axis for scatter. Negative axis means reverse indexing and will be normalized to value axis = data.rank + axis. The value can be in range [-r, r - 1] where r is the rank of data. Required.

Outputs:

  • 1: Tensor with shape equal to data tensor of the type T.

Types

  • T: any supported type.

  • T_IND: any integer numeric type.

  • T_AXIS: any integer numeric type.

  • For boolean type of data input, reduction sum, prod behaves like logical OR, AND accordingly, but there is no implementation for boolean data type and reduction mean.

Example

Example 1

<layer ... use_init_val="true" reduction="sum" type="ScatterElementsUpdate">
    <input>
        <port id="0">>  <!-- data -->
            <dim>4</dim>  <!-- values: [2, 3, 4, 6] -->
        </port>
        <port id="1">  <!-- indices (negative values allowed) -->
            <dim>6</dim>  <!-- values: [1, 0, 0, -2, -1, 2] -->
        </port>
        <port id="2">>  <!-- updates -->
            <dim>6</dim>  <!-- values: [10, 20, 30, 40, 70, 60] -->
        </port>
        <port id="3">     <!-- values: [0] -->
            <dim>1</dim>
        </port>
    </input>
    <output>
        <port id="4" precision="FP32">
            <dim>4</dim>  <!-- values: [52, 13, 104, 76] -->
        </port>
    </output>
</layer>

Example 2

<layer ... use_init_val="false" reduction="sum" type="ScatterElementsUpdate">
    <input>
        <port id="0">>  <!-- data -->
            <dim>4</dim>  <!-- values: [2, 3, 4, 6] -->
        </port>
        <port id="1">  <!-- indices -->
            <dim>6</dim>  <!-- values: [1, 0, 0, 2, 3, 2] -->
        </port>
        <port id="2">>  <!-- updates -->
            <dim>6</dim>  <!-- values: [10, 20, 30, 40, 70, 60] -->
        </port>
        <port id="3">     <!-- values: [0] -->
            <dim>1</dim>
        </port>
    </input>
    <output>
        <port id="4" precision="FP32">
            <dim>4</dim>  <!-- values: [50, 10, 100, 70] -->
        </port>
    </output>
</layer>

Example 3

<layer ... use_init_val="true" reduction="none" type="ScatterElementsUpdate">
    <input>
        <port id="0">>  <!-- data -->
            <dim>3</dim>
            <dim>4</dim>  <!-- values: [[0, 0, 0, 0],
                                         [0, 0, 0, 0],
                                         [0, 0, 0, 0]] -->
        </port>
        <port id="1">  <!-- indices -->
            <dim>2</dim>
            <dim>2</dim>  <!-- values: [[1, 2],
                                         [0, 3]] -->
        </port>
        <port id="2">>  <!-- updates -->
            <dim>2</dim>
            <dim>2</dim>  <!-- values: [[11, 12],
                                         [13, 14]]) -->
        </port>
        <port id="3">     <!-- values: [1] -->
            <dim>1</dim>
        </port>
    </input>
    <output>
        <port id="4" precision="I32">
            <dim>3</dim>
            <dim>4</dim>  <!-- values:  [[ 0, 11, 12,  0],
                                          [13,  0,  0, 14],
                                          [ 0,  0,  0,  0]] -->
        </port>
    </output>
</layer>

Example 4

<layer ... use_init_val="true" reduction="sum" type="ScatterElementsUpdate">
    <input>
        <port id="0">>  <!-- data -->
            <dim>3</dim>
            <dim>4</dim>  <!-- values: [[1, 1, 1, 1],
                                         [1, 1, 1, 1],
                                         [1, 1, 1, 1]] -->
        </port>
        <port id="1">  <!-- indices -->
            <dim>2</dim>
            <dim>2</dim>  <!-- values: [[1, 1],
                                         [0, 3]] -->
        </port>
        <port id="2">>  <!-- updates -->
            <dim>2</dim>
            <dim>2</dim>  <!-- values: [[11, 12],
                                         [13, 14]]) -->
        </port>
        <port id="3">     <!-- values: [1] -->
            <dim>1</dim>
        </port>
    </input>
    <output>
        <port id="4" precision="I32">
            <dim>3</dim>
            <dim>4</dim>  <!-- values: [[ 1, 24,  1,  1],
                                         [14,  1,  1, 15],
                                         [ 1,  1,  1,  1]] -->
        </port>
    </output>
</layer>

Example 5

<layer ... use_init_val="true" reduction="prod" type="ScatterElementsUpdate">
    <input>
        <port id="0">>  <!-- data -->
            <dim>3</dim>
            <dim>4</dim>  <!-- values: [[2, 2, 2, 2],
                                         [2, 2, 2, 2],
                                         [2, 2, 2, 2]] -->
        </port>
        <port id="1">  <!-- indices -->
            <dim>2</dim>
            <dim>2</dim>  <!-- values: [[1, 1],
                                         [0, 3]] -->
        </port>
        <port id="2">>  <!-- updates -->
            <dim>2</dim>
            <dim>2</dim>  <!-- values: [[11, 12],
                                         [13, 14]]) -->
        </port>
        <port id="3">     <!-- values: [1] -->
            <dim>1</dim>
        </port>
    </input>
    <output>
        <port id="4" precision="I32">
            <dim>3</dim>
            <dim>4</dim>  <!-- values: [[  2, 264,   2,   2],
                                         [ 26,   2,   2,  28],
                                         [  2,   2,   2,   2]] -->
        </port>
    </output>
</layer>

Example 6

<layer ... type="ScatterElementsUpdate">
    <input>
        <port id="0">
            <dim>1000</dim>
            <dim>256</dim>
            <dim>7</dim>
            <dim>7</dim>
        </port>
        <port id="1">
            <dim>125</dim>
            <dim>20</dim>
            <dim>7</dim>
            <dim>6</dim>
        </port>
        <port id="2">
            <dim>125</dim>
            <dim>20</dim>
            <dim>7</dim>
            <dim>6</dim>
        </port>
        <port id="3">     <!-- values: [0] -->
            <dim>1</dim>
        </port>
    </input>
    <output>
        <port id="4" precision="FP32">
            <dim>1000</dim>
            <dim>256</dim>
            <dim>7</dim>
            <dim>7</dim>
        </port>
    </output>
</layer>