SparseFillEmptyRows

Versioned name: SparseFillEmptyRows-16

Category: Sparse

Short description: Fills empty rows of an input sparse tensor with a default value.

Detailed description:

Operation SparseFillEmptyRows is an implementation of tf.raw_ops.SparseFillEmptyRows for 2D sparse tensors only.

The input sparse tensor is represented by the three inputs:

• indices

• values

• dense_shape

For each row in the input 2D sparse tensor, this operator checks if the row is empty. If the row is empty, the operator adds an entry with the specified default value at index [row, 0]. The input may have empty columns at the end, which will not be affected by this operation.

The output sparse tensor will be in row-major order and will have the same dense shape as the dense_shape input, but with updated output_indices and output_values.

This operator also returns a boolean vector indicating which rows were filled with the default value: empty_row_indicator[i] = True if row i was an empty row.

Attributes: SparseFillEmptyRows-16 operation has no attributes.

Inputs:

• 1: values 1D tensor containing the values of type T to be inserted at the specified indices. Required.

• 2: dense_shape 1D tensor of type T_IDX indicating the shape of the 2D dense tensor. Required.

• 3: indices 2D tensor of type T_IDX and non-negative values indicating the positions at which values are placed in the sparse tensor. Required.

  It is of shape [M, 2], where:

  • M is the same as the length of the values input.

  • The second dimension is always 2, as only 2D sparse tensors are supported.

• 4: default_value a scalar of type T to be inserted into the empty rows. Required.

Outputs:

• 1: output_indices 2D tensor of type T_IDX indicating the positions at which output_values are placed in the sparse tensor.

  It is of shape [M', 2], where:

  • M' is the length of the updated output_values.

  • The second dimension is always 2, as only 2D sparse tensors are supported.

• 2: output_values 1D tensor containing the values of type T to be inserted at the specified indices.

• 3: empty_row_indicator 1D tensor of type boolean indicating True for rows which were empty before executing the operation.

Types

• T: any numeric type.

• T_IDX: int32 or int64.

Example

Example 1: sparse tensor input with shape [5, 6].

Input sparse tensor:

• indices = [[0, 1], [0, 3], [2, 0], [3, 1]]

• values = [a, b, c, d]

• dense_shape = [5, 6]

Rows 1 and 4 are empty. The output sparse tensor will be:

• output_indices = [[0, 1], [0, 3], [1, 0], [2, 0], [3, 1], [4, 0]]

• output_values = [a, b, default_value, c, d, default_value]

• empty_row_indicator = [False, True, False, False, True]

The output sparse tensor will be in row-major order and will have the same dense shape as the dense_shape input.

<layer ... type="SparseFillEmptyRows" ... >
    <input>
        <port id="0" precision="FP32">      <!-- values are: [1, 3] -->
            <dim>2</dim>
        </port>
        <port id="1" precision="I32">       <!-- dense_shape value is: [3, 3] -->
            <dim>2</dim>
        </port>
        <port id="2" precision="I32">       <!-- indices value is: [[0, 0], [2, 2]] -->
            <dim>2</dim>
            <dim>2</dim>
        </port>
        <port id="3" precision="FP32">      <!-- default_value is: 42 -->
            <dim>0</dim>
        </port>
    </input>
    <output>
        <port id="4" precision="I32">       <!-- output_indices -->
            <dim>3</dim>
            <dim>2</dim>
        </port>
        <port id="5" precision="FP32">      <!-- output_values -->
            <dim>3</dim>
        </port>
        <port id="6" precision="BOOL">      <!-- empty_row_indicator -->
            <dim>3</dim>
        </port>
    </output>
</layer>