MaxPool#

MaxPool - 12 #

Version

name: MaxPool (GitHub)
domain: main
since_version: 12
function: False
support_level: SupportType.COMMON
shape inference: True

This version of the operator has been available since version 12.

Summary

MaxPool consumes an input tensor X and applies max pooling across the tensor according to kernel sizes, stride sizes, and pad lengths. max pooling consisting of computing the max on all values of a subset of the input tensor according to the kernel size and downsampling the data into the output tensor Y for further processing. The output spatial shape will be following:

output_spatial_shape[i] = floor((input_spatial_shape[i] + pad_shape[i] - ((kernel_spatial_shape[i] - 1) * dilations[i] + 1)) / strides_spatial_shape[i] + 1)

or #

output_spatial_shape[i] = ceil((input_spatial_shape[i] + pad_shape[i] - ((kernel_spatial_shape[i] - 1) * dilations[i] + 1)) / strides_spatial_shape[i] + 1)

if ceil_mode is enabled

* pad_shape[i] is sum of pads along axis i

auto_pad is a DEPRECATED attribute. If you are using them currently, the output spatial shape will be following:

VALID: output_spatial_shape[i] = ceil((input_spatial_shape[i] - ((kernel_spatial_shape[i] - 1) * dilations[i] + 1) + 1) / strides_spatial_shape[i])
SAME_UPPER or SAME_LOWER: output_spatial_shape[i] = ceil(input_spatial_shape[i] / strides_spatial_shape[i])

And pad shape will be following if SAME_UPPER or SAME_LOWER:

pad_shape[i] = (output_spatial_shape[i] - 1) * strides_spatial_shape[i] + ((kernel_spatial_shape[i] - 1) * dilations[i] + 1) - input_spatial_shape[i]

The output of each pooling window is maximum number of elements exclude pad.

Attributes

auto_pad: auto_pad must be either NOTSET, SAME_UPPER, SAME_LOWER or VALID. Where default value is NOTSET, which means explicit padding is used. SAME_UPPER or SAME_LOWER mean pad the input so that output_shape[i] = ceil(input_shape[i] / strides[i]) for each axis i. The padding is split between the two sides equally or almost equally (depending on whether it is even or odd). In case the padding is an odd number, the extra padding is added at the end for SAME_UPPER and at the beginning for SAME_LOWER. Default value is 'NOTSET'.
ceil_mode: Whether to use ceil or floor (default) to compute the output shape. Default value is 0.
dilations: Dilation value along each spatial axis of filter. If not present, the dilation defaults to 1 along each spatial axis.
kernel_shape (required): The size of the kernel along each axis.
pads: Padding for the beginning and ending along each spatial axis, it can take any value greater than or equal to 0. The value represent the number of pixels added to the beginning and end part of the corresponding axis. pads format should be as follow [x1_begin, x2_begin…x1_end, x2_end,…], where xi_begin the number of pixels added at the beginning of axis i and xi_end, the number of pixels added at the end of axis i. This attribute cannot be used simultaneously with auto_pad attribute. If not present, the padding defaults to 0 along start and end of each spatial axis.
storage_order: The storage order of the tensor. 0 is row major, and 1 is column major. This attribute is used only to convert an n-tuple index value into a single integer value for producing the second output. Default value is 0.
strides: Stride along each spatial axis. If not present, the stride defaults to 1 along each spatial axis.

Inputs

X (heterogeneous) - T: Input data tensor from the previous operator; dimensions for image case are (N x C x H x W), where N is the batch size, C is the number of channels, and H and W are the height and the width of the data. For non image case, the dimensions are in the form of (N x C x D1 x D2 … Dn), where N is the batch size. Optionally, if dimension denotation is in effect, the operation expects the input data tensor to arrive with the dimension denotation of [DATA_BATCH, DATA_CHANNEL, DATA_FEATURE, DATA_FEATURE …].

Outputs

Between 1 and 2 outputs.

Y (heterogeneous) - T: Output data tensor from average or max pooling across the input tensor. Dimensions will vary based on various kernel, stride, and pad sizes. Floor value of the dimension is used
Indices (optional, heterogeneous) - I: Indices tensor from max pooling across the input tensor. The dimensions of indices are the same as output tensor. The values in indices of are the indices of the selected values during pooling. The indices are computed as flatten 1-D tensor, and the indices do not consider padding. So the values in indices are in [0, N x C x D1 x … x Dn).

Type Constraints

T in ( tensor(double), tensor(float), tensor(float16), tensor(int8), tensor(uint8) ): Constrain input and output types to float and 8 bit tensors.
I in ( tensor(int64) ): Constrain index tensor to int64

Examples

_maxpool_2d_uint8

"""
input_shape: [1, 1, 5, 5]
output_shape: [1, 1, 5, 5]
pad_shape: [4, 4] -> [2, 2, 2, 2] by axis
"""
node = onnx.helper.make_node(
    "MaxPool",
    inputs=["x"],
    outputs=["y"],
    kernel_shape=[5, 5],
    pads=[2, 2, 2, 2],
)
x = np.array(
    [
        [
            [
                [1, 2, 3, 4, 5],
                [6, 7, 8, 9, 10],
                [11, 12, 13, 14, 15],
                [16, 17, 18, 19, 20],
                [21, 22, 23, 24, 25],
            ]
        ]
    ]
).astype(np.uint8)
y = np.array(
    [
        [
            [
                [13, 14, 15, 15, 15],
                [18, 19, 20, 20, 20],
                [23, 24, 25, 25, 25],
                [23, 24, 25, 25, 25],
                [23, 24, 25, 25, 25],
            ]
        ]
    ]
).astype(np.uint8)

expect(node, inputs=[x], outputs=[y], name="test_maxpool_2d_uint8")

_maxpool_2d_precomputed_pads

"""
input_shape: [1, 1, 5, 5]
output_shape: [1, 1, 5, 5]
pad_shape: [4, 4] -> [2, 2, 2, 2] by axis
"""
node = onnx.helper.make_node(
    "MaxPool",
    inputs=["x"],
    outputs=["y"],
    kernel_shape=[5, 5],
    pads=[2, 2, 2, 2],
)
x = np.array(
    [
        [
            [
                [1, 2, 3, 4, 5],
                [6, 7, 8, 9, 10],
                [11, 12, 13, 14, 15],
                [16, 17, 18, 19, 20],
                [21, 22, 23, 24, 25],
            ]
        ]
    ]
).astype(np.float32)
y = np.array(
    [
        [
            [
                [13, 14, 15, 15, 15],
                [18, 19, 20, 20, 20],
                [23, 24, 25, 25, 25],
                [23, 24, 25, 25, 25],
                [23, 24, 25, 25, 25],
            ]
        ]
    ]
).astype(np.float32)

expect(node, inputs=[x], outputs=[y], name="test_maxpool_2d_precomputed_pads")

_maxpool_with_argmax_2d_precomputed_pads

"""
input_shape: [1, 1, 5, 5]
output_shape: [1, 1, 5, 5]
pad_shape: [4, 4] -> [2, 2, 2, 2] by axis
"""
node = onnx.helper.make_node(
    "MaxPool",
    inputs=["x"],
    outputs=["y", "z"],
    kernel_shape=[5, 5],
    pads=[2, 2, 2, 2],
)
x = np.array(
    [
        [
            [
                [1, 2, 3, 4, 5],
                [6, 7, 8, 9, 10],
                [11, 12, 13, 14, 15],
                [16, 17, 18, 19, 20],
                [21, 22, 23, 24, 25],
            ]
        ]
    ]
).astype(np.float32)
y = np.array(
    [
        [
            [
                [13, 14, 15, 15, 15],
                [18, 19, 20, 20, 20],
                [23, 24, 25, 25, 25],
                [23, 24, 25, 25, 25],
                [23, 24, 25, 25, 25],
            ]
        ]
    ]
).astype(np.float32)
z = np.array(
    [
        [
            [
                [12, 13, 14, 14, 14],
                [17, 18, 19, 19, 19],
                [22, 23, 24, 24, 24],
                [22, 23, 24, 24, 24],
                [22, 23, 24, 24, 24],
            ]
        ]
    ]
).astype(np.int64)

expect(
    node,
    inputs=[x],
    outputs=[y, z],
    name="test_maxpool_with_argmax_2d_precomputed_pads",
)

_maxpool_2d_precomputed_strides

"""
input_shape: [1, 1, 5, 5]
output_shape: [1, 1, 2, 2]
"""
node = onnx.helper.make_node(
    "MaxPool", inputs=["x"], outputs=["y"], kernel_shape=[2, 2], strides=[2, 2]
)
x = np.array(
    [
        [
            [
                [1, 2, 3, 4, 5],
                [6, 7, 8, 9, 10],
                [11, 12, 13, 14, 15],
                [16, 17, 18, 19, 20],
                [21, 22, 23, 24, 25],
            ]
        ]
    ]
).astype(np.float32)
y = np.array([[[[7, 9], [17, 19]]]]).astype(np.float32)

expect(
    node, inputs=[x], outputs=[y], name="test_maxpool_2d_precomputed_strides"
)

_maxpool_with_argmax_2d_precomputed_strides

"""
input_shape: [1, 1, 5, 5]
output_shape: [1, 1, 2, 2]
"""
node = onnx.helper.make_node(
    "MaxPool",
    inputs=["x"],
    outputs=["y", "z"],
    kernel_shape=[2, 2],
    strides=[2, 2],
    storage_order=1,
)
x = np.array(
    [
        [
            [
                [1, 2, 3, 4, 5],
                [6, 7, 8, 9, 10],
                [11, 12, 13, 14, 15],
                [16, 17, 18, 19, 20],
                [21, 22, 23, 24, 25],
            ]
        ]
    ]
).astype(np.float32)
y = np.array([[[[7, 9], [17, 19]]]]).astype(np.float32)
z = np.array([[[[6, 16], [8, 18]]]]).astype(np.int64)

expect(
    node,
    inputs=[x],
    outputs=[y, z],
    name="test_maxpool_with_argmax_2d_precomputed_strides",
)

_maxpool_2d_precomputed_same_upper

"""
input_shape: [1, 1, 5, 5]
output_shape: [1, 1, 3, 3]
pad_shape: [2, 2] -> [1, 1, 1, 1] by axis
"""
node = onnx.helper.make_node(
    "MaxPool",
    inputs=["x"],
    outputs=["y"],
    kernel_shape=[3, 3],
    strides=[2, 2],
    auto_pad="SAME_UPPER",
)
x = np.array(
    [
        [
            [
                [1, 2, 3, 4, 5],
                [6, 7, 8, 9, 10],
                [11, 12, 13, 14, 15],
                [16, 17, 18, 19, 20],
                [21, 22, 23, 24, 25],
            ]
        ]
    ]
).astype(np.float32)
y = np.array([[[[7, 9, 10], [17, 19, 20], [22, 24, 25]]]]).astype(np.float32)

expect(
    node, inputs=[x], outputs=[y], name="test_maxpool_2d_precomputed_same_upper"
)

_maxpool_1d_default

"""
input_shape: [1, 3, 32]
output_shape: [1, 3, 31]
"""
node = onnx.helper.make_node(
    "MaxPool",
    inputs=["x"],
    outputs=["y"],
    kernel_shape=[2],
)
x = np.random.randn(1, 3, 32).astype(np.float32)
x_shape = np.shape(x)
kernel_shape = [2]
strides = [1]
out_shape = get_output_shape("VALID", x_shape[2:], kernel_shape, strides)
padded = x
y = pool(padded, x_shape, kernel_shape, strides, out_shape, [0], "MAX")

expect(node, inputs=[x], outputs=[y], name="test_maxpool_1d_default")

_maxpool_2d_default

"""
input_shape: [1, 3, 32, 32]
output_shape: [1, 3, 31, 31]
"""
node = onnx.helper.make_node(
    "MaxPool",
    inputs=["x"],
    outputs=["y"],
    kernel_shape=[2, 2],
)
x = np.random.randn(1, 3, 32, 32).astype(np.float32)
x_shape = np.shape(x)
kernel_shape = (2, 2)
strides = (1, 1)
out_shape = get_output_shape("VALID", x_shape[2:], kernel_shape, strides)
padded = x
y = pool(padded, x_shape, kernel_shape, strides, out_shape, (0, 0), "MAX")

expect(node, inputs=[x], outputs=[y], name="test_maxpool_2d_default")

_maxpool_3d_default

"""
input_shape: [1, 3, 32, 32, 32]
output_shape: [1, 3, 31, 31, 31]
"""
node = onnx.helper.make_node(
    "MaxPool",
    inputs=["x"],
    outputs=["y"],
    kernel_shape=[2, 2, 2],
)
x = np.random.randn(1, 3, 32, 32, 32).astype(np.float32)
x_shape = np.shape(x)
kernel_shape = [2, 2, 2]
strides = [1, 1, 1]
out_shape = get_output_shape("VALID", x_shape[2:], kernel_shape, strides)
padded = x
y = pool(padded, x_shape, kernel_shape, strides, out_shape, [0, 0, 0], "MAX")

expect(node, inputs=[x], outputs=[y], name="test_maxpool_3d_default")

_maxpool_2d_same_upper

"""
input_shape: [1, 3, 32, 32]
output_shape: [1, 3, 32, 32]
pad_shape: [1, 1] -> [0, 1, 0, 1] by axis
"""
node = onnx.helper.make_node(
    "MaxPool",
    inputs=["x"],
    outputs=["y"],
    kernel_shape=[2, 2],
    auto_pad="SAME_UPPER",
)
x = np.random.randn(1, 3, 32, 32).astype(np.float32)
x_shape = np.shape(x)
kernel_shape = (2, 2)
strides = (1, 1)
out_shape = get_output_shape("SAME_UPPER", x_shape[2:], kernel_shape, strides)
pad_shape = get_pad_shape(
    "SAME_UPPER", x_shape[2:], kernel_shape, strides, out_shape
)
pad_top = pad_shape[0] // 2
pad_bottom = pad_shape[0] - pad_top
pad_left = pad_shape[1] // 2
pad_right = pad_shape[1] - pad_left
padded = np.pad(
    x,
    ((0, 0), (0, 0), (pad_top, pad_bottom), (pad_left, pad_right)),
    mode="constant",
    constant_values=np.nan,
)
y = pool(padded, x_shape, kernel_shape, strides, out_shape, pad_shape, "MAX")

expect(node, inputs=[x], outputs=[y], name="test_maxpool_2d_same_upper")

_maxpool_2d_same_lower

"""
input_shape: [1, 3, 32, 32]
output_shape: [1, 3, 32, 32]
pad_shape: [1, 1] -> [1, 0, 1, 0] by axis
"""
node = onnx.helper.make_node(
    "MaxPool",
    inputs=["x"],
    outputs=["y"],
    kernel_shape=[2, 2],
    auto_pad="SAME_LOWER",
)
x = np.random.randn(1, 3, 32, 32).astype(np.float32)
x_shape = np.shape(x)
kernel_shape = (2, 2)
strides = (1, 1)
out_shape = get_output_shape("SAME_LOWER", x_shape[2:], kernel_shape, strides)
pad_shape = get_pad_shape(
    "SAME_LOWER", x_shape[2:], kernel_shape, strides, out_shape
)
pad_bottom = pad_shape[0] // 2
pad_top = pad_shape[0] - pad_bottom
pad_right = pad_shape[1] // 2
pad_left = pad_shape[1] - pad_right
padded = np.pad(
    x,
    ((0, 0), (0, 0), (pad_top, pad_bottom), (pad_left, pad_right)),
    mode="constant",
    constant_values=np.nan,
)
y = pool(padded, x_shape, kernel_shape, strides, out_shape, pad_shape, "MAX")

expect(node, inputs=[x], outputs=[y], name="test_maxpool_2d_same_lower")

_maxpool_2d_pads

"""
input_shape: [1, 3, 28, 28]
output_shape: [1, 3, 30, 30]
pad_shape: [4, 4] -> [2, 2, 2, 2] by axis
"""
node = onnx.helper.make_node(
    "MaxPool",
    inputs=["x"],
    outputs=["y"],
    kernel_shape=[3, 3],
    pads=[2, 2, 2, 2],
)
x = np.random.randn(1, 3, 28, 28).astype(np.float32)
x_shape = np.shape(x)
kernel_shape = (3, 3)
strides = (1, 1)
pad_bottom = pad_top = pad_right = pad_left = 2
pad_shape = [pad_top + pad_bottom, pad_left + pad_right]
out_shape = get_output_shape(
    "VALID", np.add(x_shape[2:], pad_shape), kernel_shape, strides
)
padded = np.pad(
    x,
    ((0, 0), (0, 0), (pad_top, pad_bottom), (pad_left, pad_right)),
    mode="constant",
    constant_values=np.nan,
)
y = pool(padded, x_shape, kernel_shape, strides, out_shape, pad_shape, "MAX")

expect(node, inputs=[x], outputs=[y], name="test_maxpool_2d_pads")

_maxpool_2d_strides

"""
input_shape: [1, 3, 32, 32]
output_shape: [1, 3, 10, 10]
"""
node = onnx.helper.make_node(
    "MaxPool", inputs=["x"], outputs=["y"], kernel_shape=[5, 5], strides=[3, 3]
)
x = np.random.randn(1, 3, 32, 32).astype(np.float32)
x_shape = np.shape(x)
kernel_shape = (5, 5)
strides = (3, 3)
out_shape = get_output_shape("VALID", x_shape[2:], kernel_shape, strides)
padded = x
y = pool(padded, x_shape, kernel_shape, strides, out_shape, (0, 0), "MAX")

expect(node, inputs=[x], outputs=[y], name="test_maxpool_2d_strides")

_maxpool_2d_ceil

"""
input_shape: [1, 1, 4, 4]
output_shape: [1, 1, 2, 2]
"""
node = onnx.helper.make_node(
    "MaxPool",
    inputs=["x"],
    outputs=["y"],
    kernel_shape=[3, 3],
    strides=[2, 2],
    ceil_mode=True,
)
x = np.array(
    [
        [
            [
                [1, 2, 3, 4],
                [5, 6, 7, 8],
                [9, 10, 11, 12],
                [13, 14, 15, 16],
            ]
        ]
    ]
).astype(np.float32)
y = np.array([[[[11, 12], [15, 16]]]]).astype(np.float32)

expect(node, inputs=[x], outputs=[y], name="test_maxpool_2d_ceil")

_maxpool_2d_dilations

"""
input_shape: [1, 1, 4, 4]
output_shape: [1, 1, 2, 2]
"""
node = onnx.helper.make_node(
    "MaxPool",
    inputs=["x"],
    outputs=["y"],
    kernel_shape=[2, 2],
    strides=[1, 1],
    dilations=[2, 2],
)
x = np.array(
    [
        [
            [
                [1, 2, 3, 4],
                [5, 6, 7, 8],
                [9, 10, 11, 12],
                [13, 14, 15, 16],
            ]
        ]
    ]
).astype(np.float32)
y = np.array([[[[11, 12], [15, 16]]]]).astype(np.float32)

expect(node, inputs=[x], outputs=[y], name="test_maxpool_2d_dilations")

Differences

MaxPool - 11 #

Version

name: MaxPool (GitHub)
domain: main
since_version: 11
function: False
support_level: SupportType.COMMON
shape inference: True

This version of the operator has been available since version 11.

Summary

MaxPool consumes an input tensor X and applies max pooling across the tensor according to kernel sizes, stride sizes, and pad lengths. max pooling consisting of computing the max on all values of a subset of the input tensor according to the kernel size and downsampling the data into the output tensor Y for further processing. The output spatial shape will be following:

output_spatial_shape[i] = floor((input_spatial_shape[i] + pad_shape[i] - ((kernel_spatial_shape[i] - 1) * dilations[i] + 1)) / strides_spatial_shape[i] + 1)

or #

output_spatial_shape[i] = ceil((input_spatial_shape[i] + pad_shape[i] - ((kernel_spatial_shape[i] - 1) * dilations[i] + 1)) / strides_spatial_shape[i] + 1)

if ceil_mode is enabled

* pad_shape[i] is sum of pads along axis i

auto_pad is a DEPRECATED attribute. If you are using them currently, the output spatial shape will be following:

VALID: output_spatial_shape[i] = ceil((input_spatial_shape[i] - ((kernel_spatial_shape[i] - 1) * dilations[i] + 1) + 1) / strides_spatial_shape[i])
SAME_UPPER or SAME_LOWER: output_spatial_shape[i] = ceil(input_spatial_shape[i] / strides_spatial_shape[i])

And pad shape will be following if SAME_UPPER or SAME_LOWER:

pad_shape[i] = (output_spatial_shape[i] - 1) * strides_spatial_shape[i] + ((kernel_spatial_shape[i] - 1) * dilations[i] + 1) - input_spatial_shape[i]

The output of each pooling window is maximum number of elements exclude pad.

Attributes

auto_pad: auto_pad must be either NOTSET, SAME_UPPER, SAME_LOWER or VALID. Where default value is NOTSET, which means explicit padding is used. SAME_UPPER or SAME_LOWER mean pad the input so that the output spatial size match the input.In case of odd number add the extra padding at the end for SAME_UPPER and at the beginning for SAME_LOWER. VALID mean no padding. Default value is 'NOTSET'.
ceil_mode: Whether to use ceil or floor (default) to compute the output shape. Default value is 0.
dilations: Dilation value along each spatial axis of filter. If not present, the dilation defaults to 1 along each spatial axis.
kernel_shape (required): The size of the kernel along each axis.
pads: Padding for the beginning and ending along each spatial axis, it can take any value greater than or equal to 0. The value represent the number of pixels added to the beginning and end part of the corresponding axis. pads format should be as follow [x1_begin, x2_begin…x1_end, x2_end,…], where xi_begin the number of pixels added at the beginning of axis i and xi_end, the number of pixels added at the end of axis i. This attribute cannot be used simultaneously with auto_pad attribute. If not present, the padding defaults to 0 along start and end of each spatial axis.
storage_order: The storage order of the tensor. 0 is row major, and 1 is column major. Default value is 0.
strides: Stride along each spatial axis. If not present, the stride defaults to 1 along each spatial axis.

Inputs

X (heterogeneous) - T: Input data tensor from the previous operator; dimensions for image case are (N x C x H x W), where N is the batch size, C is the number of channels, and H and W are the height and the width of the data. For non image case, the dimensions are in the form of (N x C x D1 x D2 … Dn), where N is the batch size. Optionally, if dimension denotation is in effect, the operation expects the input data tensor to arrive with the dimension denotation of [DATA_BATCH, DATA_CHANNEL, DATA_FEATURE, DATA_FEATURE …].

Outputs

Between 1 and 2 outputs.

Y (heterogeneous) - T: Output data tensor from average or max pooling across the input tensor. Dimensions will vary based on various kernel, stride, and pad sizes. Floor value of the dimension is used
Indices (optional, heterogeneous) - I: Indices tensor from max pooling across the input tensor. The dimensions of indices are the same as output tensor. The values in indices of are the indices of the selected values during pooling. The indices are computed as flatten 1-D tensor, and the indices do not consider padding. So the values in indices are in [0, N x C x D1 x … x Dn).

Type Constraints

T in ( tensor(double), tensor(float), tensor(float16) ): Constrain input and output types to float tensors.
I in ( tensor(int64) ): Constrain index tensor to int64

Differences

MaxPool - 10 #

Version

name: MaxPool (GitHub)
domain: main
since_version: 10
function: False
support_level: SupportType.COMMON
shape inference: True

This version of the operator has been available since version 10.

Summary

MaxPool consumes an input tensor X and applies max pooling across the tensor according to kernel sizes, stride sizes, and pad lengths. max pooling consisting of computing the max on all values of a subset of the input tensor according to the kernel size and downsampling the data into the output tensor Y for further processing. The output spatial shape will be following:

output_spatial_shape[i] = floor((input_spatial_shape[i] + pad_shape[i] - ((kernel_spatial_shape[i] - 1) * dilations[i] + 1)) / strides_spatial_shape[i] + 1)

or #

output_spatial_shape[i] = ceil((input_spatial_shape[i] + pad_shape[i] - ((kernel_spatial_shape[i] - 1) * dilations[i] + 1)) / strides_spatial_shape[i] + 1)

if ceil_mode is enabled

* pad_shape[i] is sum of pads along axis i

auto_pad is a DEPRECATED attribute. If you are using them currently, the output spatial shape will be following:

VALID: output_spatial_shape[i] = ceil((input_spatial_shape[i] - ((kernel_spatial_shape[i] - 1) * dilations[i] + 1) + 1) / strides_spatial_shape[i])
SAME_UPPER or SAME_LOWER: output_spatial_shape[i] = ceil(input_spatial_shape[i] / strides_spatial_shape[i])

And pad shape will be following if SAME_UPPER or SAME_LOWER:

pad_shape[i] = (output_spatial_shape[i] - 1) * strides_spatial_shape[i] + ((kernel_spatial_shape[i] - 1) * dilations[i] + 1) - input_spatial_shape[i]

The output of each pooling window is maximum number of elements exclude pad.

Attributes

auto_pad: auto_pad must be either NOTSET, SAME_UPPER, SAME_LOWER or VALID. Where default value is NOTSET, which means explicit padding is used. SAME_UPPER or SAME_LOWER mean pad the input so that the output spatial size match the input.In case of odd number add the extra padding at the end for SAME_UPPER and at the beginning for SAME_LOWER. VALID mean no padding. Default value is 'NOTSET'.
ceil_mode: Whether to use ceil or floor (default) to compute the output shape. Default value is 0.
dilations: Dilation value along each spatial axis of filter.
kernel_shape (required): The size of the kernel along each axis.
pads: Padding for the beginning and ending along each spatial axis, it can take any value greater than or equal to 0. The value represent the number of pixels added to the beginning and end part of the corresponding axis. pads format should be as follow [x1_begin, x2_begin…x1_end, x2_end,…], where xi_begin the number of pixels added at the beginning of axis i and xi_end, the number of pixels added at the end of axis i. This attribute cannot be used simultaneously with auto_pad attribute. If not present, the padding defaults to 0 along start and end of each spatial axis.
storage_order: The storage order of the tensor. 0 is row major, and 1 is column major. Default value is 0.
strides: Stride along each spatial axis.

Inputs

X (heterogeneous) - T: Input data tensor from the previous operator; dimensions for image case are (N x C x H x W), where N is the batch size, C is the number of channels, and H and W are the height and the width of the data. For non image case, the dimensions are in the form of (N x C x D1 x D2 … Dn), where N is the batch size. Optionally, if dimension denotation is in effect, the operation expects the input data tensor to arrive with the dimension denotation of [DATA_BATCH, DATA_CHANNEL, DATA_FEATURE, DATA_FEATURE …].

Outputs

Between 1 and 2 outputs.

Y (heterogeneous) - T: Output data tensor from average or max pooling across the input tensor. Dimensions will vary based on various kernel, stride, and pad sizes. Floor value of the dimension is used
Indices (optional, heterogeneous) - I: Indices tensor from max pooling across the input tensor. The dimensions of indices are the same as output tensor. The values in indices of are the indices of the selected values during pooling. The indices are computed as flatten 1-D tensor, and the indices do not consider padding. So the values in indices are in [0, N x C x D1 x … x Dn).

Type Constraints

T in ( tensor(double), tensor(float), tensor(float16) ): Constrain input and output types to float tensors.
I in ( tensor(int64) ): Constrain index tensor to int64

Differences

MaxPool - 8 #

Version

name: MaxPool (GitHub)
domain: main
since_version: 8
function: False
support_level: SupportType.COMMON
shape inference: True

This version of the operator has been available since version 8.

Summary

MaxPool consumes an input tensor X and applies max pooling across the tensor according to kernel sizes, stride sizes, and pad lengths. max pooling consisting of computing the max on all values of a subset of the input tensor according to the kernel size and downsampling the data into the output tensor Y for further processing. The output spatial shape will be following:

output_spatial_shape[i] = floor((input_spatial_shape[i] + pad_shape[i] - kernel_spatial_shape[i]) / strides_spatial_shape[i] + 1)

* pad_shape[i] is sum of pads along axis i

auto_pad is a DEPRECATED attribute. If you are using them currently, the output spatial shape will be following:

VALID: output_spatial_shape[i] = ceil((input_spatial_shape[i] - kernel_spatial_shape[i] + 1) / strides_spatial_shape[i])
SAME_UPPER or SAME_LOWER: output_spatial_shape[i] = ceil(input_spatial_shape[i] / strides_spatial_shape[i])

And pad shape will be following if SAME_UPPER or SAME_LOWER:

pad_shape[i] = (output_spatial_shape[i] - 1) * strides_spatial_shape[i] + kernel_spatial_shape[i] - input_spatial_shape[i]

The output of each pooling window is maximum number of elements exclude pad.

Attributes

auto_pad: auto_pad must be either NOTSET, SAME_UPPER, SAME_LOWER or VALID. Where default value is NOTSET, which means explicit padding is used. SAME_UPPER or SAME_LOWER mean pad the input so that the output spatial size match the input.In case of odd number add the extra padding at the end for SAME_UPPER and at the beginning for SAME_LOWER. VALID mean no padding. Default value is 'NOTSET'.
kernel_shape (required): The size of the kernel along each axis.
pads: Padding for the beginning and ending along each spatial axis, it can take any value greater than or equal to 0. The value represent the number of pixels added to the beginning and end part of the corresponding axis. pads format should be as follow [x1_begin, x2_begin…x1_end, x2_end,…], where xi_begin the number of pixels added at the beginning of axis i and xi_end, the number of pixels added at the end of axis i. This attribute cannot be used simultaneously with auto_pad attribute. If not present, the padding defaults to 0 along start and end of each spatial axis.
storage_order: The storage order of the tensor. 0 is row major, and 1 is column major. Default value is 0.
strides: Stride along each spatial axis.

Inputs

X (heterogeneous) - T: Input data tensor from the previous operator; dimensions for image case are (N x C x H x W), where N is the batch size, C is the number of channels, and H and W are the height and the width of the data. For non image case, the dimensions are in the form of (N x C x D1 x D2 … Dn), where N is the batch size. Optionally, if dimension denotation is in effect, the operation expects the input data tensor to arrive with the dimension denotation of [DATA_BATCH, DATA_CHANNEL, DATA_FEATURE, DATA_FEATURE …].

Outputs

Between 1 and 2 outputs.

Y (heterogeneous) - T: Output data tensor from average or max pooling across the input tensor. Dimensions will vary based on various kernel, stride, and pad sizes. Floor value of the dimension is used
Indices (optional, heterogeneous) - I: Indices tensor from max pooling across the input tensor. The dimensions of indices are the same as output tensor. The values in indices of are the indices of the selected values during pooling. The indices are computed as flatten 1-D tensor, and the indices do not consider padding. So the values in indices are in [0, N x C x D1 x … x Dn).

Type Constraints

T in ( tensor(double), tensor(float), tensor(float16) ): Constrain input and output types to float tensors.
I in ( tensor(int64) ): Constrain index tensor to int64

Differences

MaxPool - 1 #

Version

name: MaxPool (GitHub)
domain: main
since_version: 1
function: False
support_level: SupportType.COMMON
shape inference: True

This version of the operator has been available since version 1.

Summary

MaxPool consumes an input tensor X and applies max pooling across the tensor according to kernel sizes, stride sizes, and pad lengths. max pooling consisting of computing the max on all values of a subset of the input tensor according to the kernel size and downsampling the data into the output tensor Y for further processing. The output spatial shape will be following:

output_spatial_shape[i] = floor((input_spatial_shape[i] + pad_shape[i] - kernel_spatial_shape[i]) / strides_spatial_shape[i] + 1)

* pad_shape[i] is sum of pads along axis i

auto_pad is a DEPRECATED attribute. If you are using them currently, the output spatial shape will be following:

VALID: output_spatial_shape[i] = ceil((input_spatial_shape[i] - kernel_spatial_shape[i] + 1) / strides_spatial_shape[i])
SAME_UPPER or SAME_LOWER: output_spatial_shape[i] = ceil(input_spatial_shape[i] / strides_spatial_shape[i])

And pad shape will be following if SAME_UPPER or SAME_LOWER:

pad_shape[i] = (output_spatial_shape[i] - 1) * strides_spatial_shape[i] + kernel_spatial_shape[i] - input_spatial_shape[i]

The output of each pooling window is maximum number of elements exclude pad.

Attributes

auto_pad: auto_pad must be either NOTSET, SAME_UPPER, SAME_LOWER or VALID. Where default value is NOTSET, which means explicit padding is used. SAME_UPPER or SAME_LOWER mean pad the input so that the output spatial size match the input.In case of odd number add the extra padding at the end for SAME_UPPER and at the beginning for SAME_LOWER. VALID mean no padding. Default value is 'NOTSET'.
kernel_shape (required): The size of the kernel along each axis.
pads: Padding for the beginning and ending along each spatial axis, it can take any value greater than or equal to 0. The value represent the number of pixels added to the beginning and end part of the corresponding axis. pads format should be as follow [x1_begin, x2_begin…x1_end, x2_end,…], where xi_begin the number of pixels added at the beginning of axis i and xi_end, the number of pixels added at the end of axis i. This attribute cannot be used simultaneously with auto_pad attribute. If not present, the padding defaults to 0 along start and end of each spatial axis.
strides: Stride along each spatial axis.

Inputs

X (heterogeneous) - T: Input data tensor from the previous operator; dimensions for image case are (N x C x H x W), where N is the batch size, C is the number of channels, and H and W are the height and the width of the data. For non image case, the dimensions are in the form of (N x C x D1 x D2 … Dn), where N is the batch size. Optionally, if dimension denotation is in effect, the operation expects the input data tensor to arrive with the dimension denotation of [DATA_BATCH, DATA_CHANNEL, DATA_FEATURE, DATA_FEATURE …].

Outputs

Y (heterogeneous) - T: Output data tensor from average or max pooling across the input tensor. Dimensions will vary based on various kernel, stride, and pad sizes. Floor value of the dimension is used

Type Constraints

T in ( tensor(double), tensor(float), tensor(float16) ): Constrain input and output types to float tensors.

`0`	`0`	`MaxPool consumes an input tensor X and applies max pooling across`	`MaxPool consumes an input tensor X and applies max pooling across`
`1`	`1`	`the tensor according to kernel sizes, stride sizes, and pad lengths.`	`the tensor according to kernel sizes, stride sizes, and pad lengths.`
`2`	`2`	`max pooling consisting of computing the max on all values of a`	`max pooling consisting of computing the max on all values of a`
`3`	`3`	`subset of the input tensor according to the kernel size and downsampling the`	`subset of the input tensor according to the kernel size and downsampling the`
`4`	`4`	`data into the output tensor Y for further processing. The output spatial shape will be following:`	`data into the output tensor Y for further processing. The output spatial shape will be following:`
`5`	`5`	`::`	`::`
`6`	`6`
`7`	`7`	`output_spatial_shape[i] = floor((input_spatial_shape[i] + pad_shape[i] - ((kernel_spatial_shape[i] - 1) * dilations[i] + 1)) / strides_spatial_shape[i] + 1)`	`output_spatial_shape[i] = floor((input_spatial_shape[i] + pad_shape[i] - ((kernel_spatial_shape[i] - 1) * dilations[i] + 1)) / strides_spatial_shape[i] + 1)`
`8`	`8`
`9`	`9`	`or`	`or`
`10`	`10`	`::`	`::`
`11`	`11`
`12`	`12`	`output_spatial_shape[i] = ceil((input_spatial_shape[i] + pad_shape[i] - ((kernel_spatial_shape[i] - 1) * dilations[i] + 1)) / strides_spatial_shape[i] + 1)`	`output_spatial_shape[i] = ceil((input_spatial_shape[i] + pad_shape[i] - ((kernel_spatial_shape[i] - 1) * dilations[i] + 1)) / strides_spatial_shape[i] + 1)`
`13`	`13`
`14`	`14`	`if ceil_mode is enabled`	`if ceil_mode is enabled`
`15`	`15`
`16`	`16`	`::`	`::`
`17`	`17`
`18`	`18`	`* pad_shape[i] is sum of pads along axis i`	`* pad_shape[i] is sum of pads along axis i`
`19`	`19`
`20`	`20`	`auto_pad is a DEPRECATED attribute. If you are using them currently, the output spatial shape will be following:`	`auto_pad is a DEPRECATED attribute. If you are using them currently, the output spatial shape will be following:`
`21`	`21`	`::`	`::`
`22`	`22`
`23`	`23`	`VALID: output_spatial_shape[i] = ceil((input_spatial_shape[i] - ((kernel_spatial_shape[i] - 1) * dilations[i] + 1) + 1) / strides_spatial_shape[i])`	`VALID: output_spatial_shape[i] = ceil((input_spatial_shape[i] - ((kernel_spatial_shape[i] - 1) * dilations[i] + 1) + 1) / strides_spatial_shape[i])`
`24`	`24`	`SAME_UPPER or SAME_LOWER: output_spatial_shape[i] = ceil(input_spatial_shape[i] / strides_spatial_shape[i])`	`SAME_UPPER or SAME_LOWER: output_spatial_shape[i] = ceil(input_spatial_shape[i] / strides_spatial_shape[i])`
`25`	`25`
`26`	`26`	`And pad shape will be following if SAME_UPPER or SAME_LOWER:`	`And pad shape will be following if SAME_UPPER or SAME_LOWER:`
`27`	`27`	`::`	`::`
`28`	`28`
`29`	`29`	`pad_shape[i] = (output_spatial_shape[i] - 1) * strides_spatial_shape[i] + ((kernel_spatial_shape[i] - 1) * dilations[i] + 1) - input_spatial_shape[i]`	`pad_shape[i] = (output_spatial_shape[i] - 1) * strides_spatial_shape[i] + ((kernel_spatial_shape[i] - 1) * dilations[i] + 1) - input_spatial_shape[i]`
`30`	`30`
`31`	`31`	`The output of each pooling window is maximum number of elements exclude pad.`	`The output of each pooling window is maximum number of elements exclude pad.`
`32`	`32`
`33`	`33`	`Attributes`	`Attributes`
`34`	`34`
`35`	`35`	`* auto_pad:`	`* auto_pad:`
`36`	`36`	`auto_pad must be either NOTSET, SAME_UPPER, SAME_LOWER or VALID.`	`auto_pad must be either NOTSET, SAME_UPPER, SAME_LOWER or VALID.`
`37`	`37`	`Where default value is NOTSET, which means explicit padding is used.`	`Where default value is NOTSET, which means explicit padding is used.`
`38`	`38`	`SAME_UPPER or SAME_LOWER mean pad the input so that the output`	`SAME_UPPER or SAME_LOWER mean pad the input so that output_shape[i]`
`39`		`spatial size match the input.In case of odd number add the extra`
	`39`		`= ceil(input_shape[i] / strides[i]) for each axis i. The padding`
	`40`		`is split between the two sides equally or almost equally (depending`
	`41`		`on whether it is even or odd). In case the padding is an odd number,`
`40`	`42`	`padding at the end for SAME_UPPER and at the beginning for`	`the extra padding is added at the end for SAME_UPPER and at the`
`41`	`43`	`SAME_LOWER. VALID mean no padding. Default value is 'NOTSET'.`	`beginning for SAME_LOWER. Default value is 'NOTSET'.`
`42`	`44`	`* ceil_mode:`	`* ceil_mode:`
`43`	`45`	`Whether to use ceil or floor (default) to compute the output shape. Default value is 0.`	`Whether to use ceil or floor (default) to compute the output shape. Default value is 0.`
`44`	`46`	`* dilations:`	`* dilations:`
`45`	`47`	`Dilation value along each spatial axis of filter. If not present,`	`Dilation value along each spatial axis of filter. If not present,`
`46`	`48`	`the dilation defaults to 1 along each spatial axis.`	`the dilation defaults to 1 along each spatial axis.`
`47`	`49`	`* kernel_shape (required):`	`* kernel_shape (required):`
`48`	`50`	`The size of the kernel along each axis.`	`The size of the kernel along each axis.`
`49`	`51`	`* pads:`	`* pads:`
`50`	`52`	`Padding for the beginning and ending along each spatial axis, it can`	`Padding for the beginning and ending along each spatial axis, it can`
`51`	`53`	`take any value greater than or equal to 0. The value represent the`	`take any value greater than or equal to 0. The value represent the`
`52`	`54`	`number of pixels added to the beginning and end part of the`	`number of pixels added to the beginning and end part of the`
`53`	`55`	`corresponding axis. pads format should be as follow [x1_begin,`	`corresponding axis. pads format should be as follow [x1_begin,`
`54`	`56`	`x2_begin...x1_end, x2_end,...], where xi_begin the number of pixels`	`x2_begin...x1_end, x2_end,...], where xi_begin the number of pixels`
`55`	`57`	`added at the beginning of axis i and xi_end, the number of pixels`	`added at the beginning of axis i and xi_end, the number of pixels`
`56`	`58`	`added at the end of axis i. This attribute cannot be used`	`added at the end of axis i. This attribute cannot be used`
`57`	`59`	`simultaneously with auto_pad attribute. If not present, the padding`	`simultaneously with auto_pad attribute. If not present, the padding`
`58`	`60`	`defaults to 0 along start and end of each spatial axis.`	`defaults to 0 along start and end of each spatial axis.`
`59`	`61`	`* storage_order:`	`* storage_order:`
`60`	`62`	`The storage order of the tensor. 0 is row major, and 1 is column`	`The storage order of the tensor. 0 is row major, and 1 is column`
`61`	`63`	`major. Default value is 0.`	`major. This attribute is used only to convert an n-tuple index value`
	`64`		`into a single integer value for producing the second output. Default value is 0.`
`62`	`65`	`* strides:`	`* strides:`
`63`	`66`	`Stride along each spatial axis. If not present, the stride defaults`	`Stride along each spatial axis. If not present, the stride defaults`
`64`	`67`	`to 1 along each spatial axis.`	`to 1 along each spatial axis.`
`65`	`68`
`66`	`69`	`Inputs`	`Inputs`
`67`	`70`
`68`	`71`	`* X (heterogeneous) - T:`	`* X (heterogeneous) - T:`
`69`	`72`	`Input data tensor from the previous operator; dimensions for image`	`Input data tensor from the previous operator; dimensions for image`
`70`	`73`	`case are (N x C x H x W), where N is the batch size, C is the number`	`case are (N x C x H x W), where N is the batch size, C is the number`
`71`	`74`	`of channels, and H and W are the height and the width of the data.`	`of channels, and H and W are the height and the width of the data.`
`72`	`75`	`For non image case, the dimensions are in the form of (N x C x D1 x`	`For non image case, the dimensions are in the form of (N x C x D1 x`
`73`	`76`	`D2 ... Dn), where N is the batch size. Optionally, if dimension`	`D2 ... Dn), where N is the batch size. Optionally, if dimension`
`74`	`77`	`denotation is in effect, the operation expects the input data tensor`	`denotation is in effect, the operation expects the input data tensor`
`75`	`78`	`to arrive with the dimension denotation of [DATA_BATCH,`	`to arrive with the dimension denotation of [DATA_BATCH,`
`76`	`79`	`DATA_CHANNEL, DATA_FEATURE, DATA_FEATURE ...].`	`DATA_CHANNEL, DATA_FEATURE, DATA_FEATURE ...].`
`77`	`80`
`78`	`81`	`Outputs`	`Outputs`
`79`	`82`
`80`	`83`	`Between 1 and 2 outputs.`	`Between 1 and 2 outputs.`
`81`	`84`
`82`	`85`	`* Y (heterogeneous) - T:`	`* Y (heterogeneous) - T:`
`83`	`86`	`Output data tensor from average or max pooling across the input`	`Output data tensor from average or max pooling across the input`
`84`	`87`	`tensor. Dimensions will vary based on various kernel, stride, and`	`tensor. Dimensions will vary based on various kernel, stride, and`
`85`	`88`	`pad sizes. Floor value of the dimension is used`	`pad sizes. Floor value of the dimension is used`
`86`	`89`	`* Indices (optional, heterogeneous) - I:`	`* Indices (optional, heterogeneous) - I:`
`87`	`90`	`Indices tensor from max pooling across the input tensor. The`	`Indices tensor from max pooling across the input tensor. The`
`88`	`91`	`dimensions of indices are the same as output tensor. The values in`	`dimensions of indices are the same as output tensor. The values in`
`89`	`92`	`indices of are the indices of the selected values during pooling.`	`indices of are the indices of the selected values during pooling.`
`90`	`93`	`The indices are computed as flatten 1-D tensor, and the indices do`	`The indices are computed as flatten 1-D tensor, and the indices do`
`91`	`94`	`not consider padding. So the values in indices are in [0, N x C x D1`	`not consider padding. So the values in indices are in [0, N x C x D1`
`92`	`95`	`x ... x Dn).`	`x ... x Dn).`
`93`	`96`
`94`	`97`	`Type Constraints`	`Type Constraints`
`95`	`98`
`96`	`99`	`* T in (`	`* T in (`
`97`	`100`	`tensor(double),`	`tensor(double),`
`98`	`101`	`tensor(float),`	`tensor(float),`
`99`	`102`	`tensor(float16)`	`tensor(float16),`
	`103`		`tensor(int8),`
	`104`		`tensor(uint8)`
`100`	`105`	`):`	`):`
`101`	`106`	`Constrain input and output types to float tensors.`	`Constrain input and output types to float and 8 bit tensors.`
`102`	`107`	`* I in (`	`* I in (`
`103`	`108`	`tensor(int64)`	`tensor(int64)`
`104`	`109`	`):`	`):`
`105`	`110`	`Constrain index tensor to int64`	`Constrain index tensor to int64`