Select#

Versioned name: Select-1

Category: Condition

Short description: Select returns a tensor filled with the elements from the second or the third input, depending on the condition (the first input) value.

Detailed description: Select takes elements from then input tensor or the else input tensor based on a condition mask provided in the first input cond. Before performing selection, input tensors then and else are broadcasted to each other if their shapes are different and auto_broadcast attributes is not none. Then the cond tensor is one-way broadcasted to the resulting shape of broadcasted then and else. Broadcasting is performed according to auto_broadcast value.

\[o_{i} = cond_{i} ? then_{i} : else_{i}\]

Broadcasting rules:

If auto_broadcast attribute is not none, select operation takes a two-step broadcast before performing selection:

  • Step 1: input tensors then and else are broadcasted to each other if their shapes are different

  • Step 2: then the cond tensor will be one-way broadcasted to the resulting shape of broadcasted then and else. To be more specific, we align the two shapes to the right and compare them from right to left. Each dimension should be either a common length or the dimension of cond should be 1.

  • example:

    • cond={4, 5}, output_shape={2, 3, 4, 5} => result = {2, 3, 4, 5}

    • cond={3, 1, 5}, output_shape={2, 3, 4, 5} => result = {2, 3, 4, 5}

    • cond={3,5}, output_shape={2, 3, 4, 5} => result = invalid_shape

OpenVINO description: This OP is as same as OpenVINO OP

Attributes:

  • auto_broadcast

    • Description: specifies rules used for auto-broadcasting of input tensors.

    • Range of values:

      • none - no auto-broadcasting is allowed, all input shapes should match

      • numpy - numpy broadcasting rules, aligned with ONNX Broadcasting. Description is available in ONNX docs.

        Notion: the broadcasting here includes two different steps, multi-directional and uni-directional respectively. numpy here only refers to the first multi-directional step. Please refer to the broadcasting rules section for more details for the second step.

    • Type: string

    • Default value: “numpy”

    • Required: no

Inputs

  • 1: cond tensor with selection mask of type boolean. Required.

    • Type: T1

  • 2: then the tensor with elements to take where the corresponding element in cond is true. Required.

    • Type: T2

  • 3: else the tensor with elements to take where the corresponding element in cond is false. Required.

    • Type: T2

Outputs

  • 1: blended output tensor that is tailored from values of inputs tensors then and else based on cond and broadcasting rules.

    • Type: T2

Types:

  • T1: boolean.

  • T2: f32, f16, bf16.

  • Note: else and then input tensor should have the same data type denoted by T2. cond input tensor should be boolean data type denoted by T1. The output tensor has the same data type of elements as then and else