SegmentMax

Versioned name: SegmentMax-16

Category: Arithmetic

Short description: SegmentMax-16 operation finds the maximum value in each specified segment of the input tensor.

Detailed description

For each index in segment_ids the operator gets values from data input tensor and calculates the maximum value for each segment.

For example segments_ids with value [0,0,0,1,1,3,5,5] defines 4 non-empty segments. num_segments is not given. When coupled with a 1D data tensor data, the segments are as follows:

• Segment_0: [data[0], data[1], data[2]]

• Segment_1: [data[3], data[4]]

• Segment_2: []

• Segment_3: [data[5]]

• Segment_4: []

• Segment_5: [data[6], data[7]]

When there are no values in a segment, output[segment] is defined by fill_mode attribute.

For fill_mode equal to ZERO , the operation output would be [max(Segment_0), max(Segment_1), 0, max(Segment_3), 0, max(Segment_5)].

Attributes:

• 1: fill_mode

  • Description: Responsible for the value assigned to segments which are empty. Required.

  • Range of values: Name of the mode in string format:

    • ZERO - the empty segments are filled with zeros.

    • LOWEST - the empty segments are filled with the lowest value of the data type T.

  • Type: string

Inputs

• 1: data - ND tensor of type T, the numerical data on which SegmentMax operation will be performed. Required.

• 2: segment_ids - 1D Tensor of sorted non-negative numbers of type T_IDX1. Its size is equal to the size of the first dimension of the data input tensor. Required.

• 3: num_segments - A scalar value of type T_IDX2 representing the segments count. If num_segments < max(segment_ids) + 1 then the extra segments defined in segment_ids are not included in the output. If If num_segments > max(segment_ids) + 1 then the output is padded with empty segments. Defaults to max(segment_ids) + 1. Optional.

Outputs

• 1: The output tensor has same rank and dimensions as the data input tensor except for the first dimension which is equal to the value of num_segments.

Types

• T: any supported numerical data type.

• T_IDX1, T_IDX2: int64 or int32.

Examples

Example 1: num_segments < max(segment_ids) + 1

 <layer ... type="SegmentMax" ... >
     <data empty_segment_value="ZERO">
     <input>
         <port id="0" precision="F32">   <!-- data -->
             <dim>5</dim>
         </port>
         <port id="1" precision="I32">   <!-- segment_ids with 4 segments: [0, 0, 2, 3, 3] -->
             <dim>5</dim>
         </port>
         <port id="2" precision="I64">   <!-- number of segments: 2 -->
             <dim>0</dim>
         </port>
     </input>
     <output>
         <port id="3" precision="F32">
             <dim>2</dim>
         </port>
     </output>
 </layer>

Example 2: num_segments > max(segment_ids) + 1

 <layer ... type="SegmentMax" ... >
     <data empty_segment_value="ZERO">
     <input>
         <port id="0" precision="F32">   <!-- data -->
             <dim>5</dim>
         </port>
         <port id="1" precision="I32">   <!-- segment_ids with 4 segments: [0, 0, 2, 3, 3] -->
             <dim>5</dim>
         </port>
         <port id="2" precision="I64">   <!-- number of segments: 8 -->
             <dim>0</dim>
         </port>
     </input>
     <output>
         <port id="3" precision="F32">
             <dim>8</dim>
         </port>
     </output>
 </layer>

Example 3: 2D input data, no num_segments

 <layer ... type="SegmentMax" ... >
     <data empty_segment_value="LOWEST">
     <input>
         <port id="0" precision="I32">   <!-- data -->
             <dim>3</dim>
             <dim>4</dim>
         </port>
         <port id="1" precision="I64">   <!-- segment_ids with 2 segments: [0, 1, 1] -->
             <dim>3</dim>
         </port>
     </input>
     <output>
         <port id="2" precision="I32">
             <dim>2</dim>
             <dim>4</dim>
         </port>
     </output>
 </layer>