AdaptiveAvgPool#

Versioned name: AdaptiveAvgPool-8

Category: Pooling

Short description: Applies average pooling with adaptive kernel size over the input.

Detailed description: This operation calculates the output based on the first input and output_size determined by the second input. The kernel dimensions are calculated using the following formulae for the NCDHW input case:

\begin{array}{r} \begin{array}{lcl} d_{s t a r t} & = & ⌊ i \cdot \frac{D_{i n}}{D_{o u t}} ⌋ \\ d_{e n d} & = & ⌈ (i + 1) \cdot \frac{D_{i n}}{D_{o u t}} ⌉ \\ h_{s t a r t} & = & ⌊ j \cdot \frac{H_{i n}}{H_{o u t}} ⌋ \\ h_{e n d} & = & ⌈ (j + 1) \cdot \frac{H_{i n}}{H_{o u t}} ⌉ \\ w_{s t a r t} & = & ⌊ k \cdot \frac{W_{i n}}{W_{o u t}} ⌋ \\ w_{e n d} & = & ⌈ (k + 1) \cdot \frac{W_{i n}}{W_{o u t}} ⌉ \end{array} \end{array}

The output is calculated with the following formula:

O u t p u t (i, j, k) = \frac{I n p u t [d_{s t a r t} : d_{e n d}, h_{s t a r t} : h_{e n d}, w_{s t a r t} : w_{e n d}]}{(d_{e n d} - d_{s t a r t}) \cdot (h_{e n d} - h_{s t a r t}) \cdot (w_{e n d} - w_{s t a r t})}

Inputs:

1: 3D, 4D, or 5D input tensor of shape [N, C, H], [N, C, H, W] or [N, C, D, H, W] and type T. Required.
2: 1D tensor describing output shape for spatial dimensions. Can be [H_out] for 3D input, [H_out, W_out] for 4D input, [D_out, H_out, W_out] for 5D input and of type T_SHAPE. Required.

Outputs:

1: Output of type T and shape [N, C, H_out], [N, C, H_out, W_out] or [N, C, D_out, H_out, W_out].

Types

T: floating-point type.
T_SHAPE: int32 or int64.

Examples

<layer ... type="AdaptiveAvgPool" ... >
    <data output_type="i64"/>
    <input>
        <port id="0">
            <dim>1</dim>
            <dim>3</dim>
            <dim>32</dim>
            <dim>32</dim>
        </port>
    </input>
    <input>
        <port id="1">
            <dim>2</dim>
        </port>
    </input>
    <output>
        <port id="2">
            <dim>1</dim>
            <dim>3</dim>
            <dim>16</dim>
            <dim>16</dim>
        </port>
    </output>
</layer>