# onnx/onnx

Switch branches/tags
Nothing to show
96758c9 Oct 16, 2018
55 contributors

### Users who have contributed to this file

11859 lines (8772 sloc) 321 KB

## Operator Schemas

This file is automatically generated from the def files via this script. Do not modify directly and instead edit operator definitions.

• ai.onnx (default)

## ai.onnx (default)

### Abs

Absolute takes one input data (Tensor) and produces one output data (Tensor) where the absolute is, y = abs(x), is applied to the tensor elementwise.

#### Version

This version of the operator has been available since version 6 of the default ONNX operator set.

Other versions of this operator: Abs-1

X : T
Input tensor

Y : T
Output tensor

#### Type Constraints

T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double)
Constrain input and output types to all numeric tensors.

#### Examples

abs
node = onnx.helper.make_node(
'Abs',
inputs=['x'],
outputs=['y'],
)
x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.abs(x)

expect(node, inputs=[x], outputs=[y],
name='test_abs')

### Acos

Calculates the arccosine (inverse of cosine) of the given input tensor, element-wise.

#### Version

This version of the operator has been available since version 7 of the default ONNX operator set.

input : T
Input tensor

#### Outputs

output : T
The arccosine of the input tensor computed element-wise

#### Type Constraints

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

#### Examples

acos
node = onnx.helper.make_node(
'Acos',
inputs=['x'],
outputs=['y'],
)

x = np.array([-0.5, 0, 0.5]).astype(np.float32)
y = np.arccos(x)
expect(node, inputs=[x], outputs=[y],
name='test_acos_example')

x = np.random.rand(3, 4, 5).astype(np.float32)
y = np.arccos(x)
expect(node, inputs=[x], outputs=[y],
name='test_acos')

Performs element-wise binary addition (with Numpy-style broadcasting support).

This operator supports multidirectional (i.e., Numpy-style) broadcasting; for more details please check the doc.

#### Version

This version of the operator has been available since version 7 of the default ONNX operator set.

Other versions of this operator: Add-1, Add-6

A : T
First operand.
B : T
Second operand.

#### Outputs

C : T
Result, has same element type as two inputs

#### Type Constraints

T : tensor(uint32), tensor(uint64), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double)
Constrain input and output types to high-precision numeric tensors.

#### Examples

node = onnx.helper.make_node(
inputs=['x', 'y'],
outputs=['sum'],
)

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.random.randn(3, 4, 5).astype(np.float32)
expect(node, inputs=[x, y], outputs=[x + y],
name='test_add')
node = onnx.helper.make_node(
inputs=['x', 'y'],
outputs=['sum'],
)

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.random.randn(5).astype(np.float32)
expect(node, inputs=[x, y], outputs=[x + y],
name='test_add_bcast')

### And

Returns the tensor resulted from performing the and logical operation elementwise on the input tensors A and B (with Numpy-style broadcasting support).

This operator supports multidirectional (i.e., Numpy-style) broadcasting; for more details please check the doc.

#### Version

This version of the operator has been available since version 7 of the default ONNX operator set.

Other versions of this operator: And-1

#### Inputs

A : T
First input operand for the logical operator.
B : T
Second input operand for the logical operator.

C : T1
Result tensor.

#### Type Constraints

T : tensor(bool)
Constrains input to boolean tensor.
T1 : tensor(bool)
Constrains output to boolean tensor.

#### Examples

and
node = onnx.helper.make_node(
'And',
inputs=['x', 'y'],
outputs=['and'],
)

# 2d
x = (np.random.randn(3, 4) > 0).astype(np.bool)
y = (np.random.randn(3, 4) > 0).astype(np.bool)
z = np.logical_and(x, y)
expect(node, inputs=[x, y], outputs=[z],
name='test_and2d')

# 3d
x = (np.random.randn(3, 4, 5) > 0).astype(np.bool)
y = (np.random.randn(3, 4, 5) > 0).astype(np.bool)
z = np.logical_and(x, y)
expect(node, inputs=[x, y], outputs=[z],
name='test_and3d')

# 4d
x = (np.random.randn(3, 4, 5, 6) > 0).astype(np.bool)
y = (np.random.randn(3, 4, 5, 6) > 0).astype(np.bool)
z = np.logical_and(x, y)
expect(node, inputs=[x, y], outputs=[z],
name='test_and4d')
node = onnx.helper.make_node(
'And',
inputs=['x', 'y'],
outputs=['and'],
)

# 3d vs 1d
x = (np.random.randn(3, 4, 5) > 0).astype(np.bool)
y = (np.random.randn(5) > 0).astype(np.bool)
z = np.logical_and(x, y)
expect(node, inputs=[x, y], outputs=[z],
name='test_and_bcast3v1d')

# 3d vs 2d
x = (np.random.randn(3, 4, 5) > 0).astype(np.bool)
y = (np.random.randn(4, 5) > 0).astype(np.bool)
z = np.logical_and(x, y)
expect(node, inputs=[x, y], outputs=[z],
name='test_and_bcast3v2d')

# 4d vs 2d
x = (np.random.randn(3, 4, 5, 6) > 0).astype(np.bool)
y = (np.random.randn(5, 6) > 0).astype(np.bool)
z = np.logical_and(x, y)
expect(node, inputs=[x, y], outputs=[z],
name='test_and_bcast4v2d')

# 4d vs 3d
x = (np.random.randn(3, 4, 5, 6) > 0).astype(np.bool)
y = (np.random.randn(4, 5, 6) > 0).astype(np.bool)
z = np.logical_and(x, y)
expect(node, inputs=[x, y], outputs=[z],
name='test_and_bcast4v3d')

# 4d vs 4d
x = (np.random.randn(1, 4, 1, 6) > 0).astype(np.bool)
y = (np.random.randn(3, 1, 5, 6) > 0).astype(np.bool)
z = np.logical_and(x, y)
expect(node, inputs=[x, y], outputs=[z],
name='test_and_bcast4v4d')

### ArgMax

Computes the indices of the max elements of the input tensor's element along the provided axis. The resulted tensor has the same rank as the input if keepdims equal 1. If keepdims equal 0, then the resulted tensor have the reduced dimension pruned. The type of the output tensor is integer.

#### Version

This version of the operator has been available since version 1 of the default ONNX operator set.

#### Attributes

axis : int (default is 0)
The axis in which to compute the arg indices.
keepdims : int (default is 1)
Keep the reduced dimension or not, default 1 mean keep reduced dimension.

data : T
An input tensor.

#### Outputs

reduced : tensor(int64)
Reduced output tensor with integer data type.

#### Type Constraints

T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double)
Constrain input and output types to all numeric tensors.

#### Examples

default_axes_keepdims
data = np.array([[2, 1], [3, 10]], dtype=np.float32)
keepdims = 1
node = onnx.helper.make_node(
'ArgMax',
inputs=['data'],
outputs=['result'],
keepdims=keepdims)

# result: [[1], [1]]
result = argmax_use_numpy(data, keepdims=keepdims)
expect(node, inputs=[data], outputs=[result], name='test_argmax_default_axis_example')

data = np.random.uniform(-10, 10, [2, 3, 4]).astype(np.float32)
# result's shape: [1, 3, 4]
result = argmax_use_numpy(data, keepdims=keepdims)
expect(node, inputs=[data], outputs=[result], name='test_argmax_default_axis_random')
keepdims
data = np.array([[2, 1], [3, 10]], dtype=np.float32)
axis = 1
keepdims = 1
node = onnx.helper.make_node(
'ArgMax',
inputs=['data'],
outputs=['result'],
axis=axis,
keepdims=keepdims)
# result: [[0], [1]]
result = argmax_use_numpy(data, axis=axis, keepdims=keepdims)
expect(node, inputs=[data], outputs=[result], name='test_argmax_keepdims_example')

data = np.random.uniform(-10, 10, [2, 3, 4]).astype(np.float32)
# result's shape: [2, 1, 4]
result = argmax_use_numpy(data, axis=axis, keepdims=keepdims)
expect(node, inputs=[data], outputs=[result], name='test_argmax_keepdims_random')
no_keepdims
data = np.array([[2, 1], [3, 10]], dtype=np.float32)
axis = 1
keepdims = 0
node = onnx.helper.make_node(
'ArgMax',
inputs=['data'],
outputs=['result'],
axis=axis,
keepdims=keepdims)
# result: [[0, 1]]
result = argmax_use_numpy(data, axis=axis, keepdims=keepdims)
expect(node, inputs=[data], outputs=[result], name='test_argmax_no_keepdims_example')

data = np.random.uniform(-10, 10, [2, 3, 4]).astype(np.float32)
# result's shape: [2, 4]
result = argmax_use_numpy(data, axis=axis, keepdims=keepdims)
expect(node, inputs=[data], outputs=[result], name='test_argmax_no_keepdims_random')

### ArgMin

Computes the indices of the min elements of the input tensor's element along the provided axis. The resulted tensor has the same rank as the input if keepdims equal 1. If keepdims equal 0, then the resulted tensor have the reduced dimension pruned. The type of the output tensor is integer.

#### Version

This version of the operator has been available since version 1 of the default ONNX operator set.

#### Attributes

axis : int (default is 0)
The axis in which to compute the arg indices.
keepdims : int (default is 1)
Keep the reduced dimension or not, default 1 mean keep reduced dimension.

data : T
An input tensor.

#### Outputs

reduced : tensor(int64)
Reduced output tensor with integer data type.

#### Type Constraints

T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double)
Constrain input and output types to all numeric tensors.

#### Examples

default_axes_keepdims
data = np.array([[2, 1], [3, 10]], dtype=np.float32)
keepdims = 1
node = onnx.helper.make_node(
'ArgMin',
inputs=['data'],
outputs=['result'],
keepdims=keepdims)

# result: [[0], [0]]
result = argmin_use_numpy(data, keepdims=keepdims)
expect(node, inputs=[data], outputs=[result], name='test_argmin_default_axis_example')

data = np.random.uniform(-10, 10, [2, 3, 4]).astype(np.float32)
# result's shape: [1, 3, 4]
result = argmin_use_numpy(data, keepdims=keepdims)
expect(node, inputs=[data], outputs=[result], name='test_argmin_default_axis_random')
keepdims
data = np.array([[2, 1], [3, 10]], dtype=np.float32)
axis = 1
keepdims = 1
node = onnx.helper.make_node(
'ArgMin',
inputs=['data'],
outputs=['result'],
axis=axis,
keepdims=keepdims)
# result: [[1], [0]]
result = argmin_use_numpy(data, axis=axis, keepdims=keepdims)
expect(node, inputs=[data], outputs=[result], name='test_argmin_keepdims_example')

data = np.random.uniform(-10, 10, [2, 3, 4]).astype(np.float32)
# result's shape: [2, 1, 4]
result = argmin_use_numpy(data, axis=axis, keepdims=keepdims)
expect(node, inputs=[data], outputs=[result], name='test_argmin_keepdims_random')
no_keepdims
data = np.array([[2, 1], [3, 10]], dtype=np.float32)
axis = 1
keepdims = 0
node = onnx.helper.make_node(
'ArgMin',
inputs=['data'],
outputs=['result'],
axis=axis,
keepdims=keepdims)
# result: [[1, 0]]
result = argmin_use_numpy(data, axis=axis, keepdims=keepdims)
expect(node, inputs=[data], outputs=[result], name='test_argmin_no_keepdims_example')

data = np.random.uniform(-10, 10, [2, 3, 4]).astype(np.float32)
# result's shape: [2, 4]
result = argmin_use_numpy(data, axis=axis, keepdims=keepdims)
expect(node, inputs=[data], outputs=[result], name='test_argmin_no_keepdims_random')

### Asin

Calculates the arcsine (inverse of sine) of the given input tensor, element-wise.

#### Version

This version of the operator has been available since version 7 of the default ONNX operator set.

input : T
Input tensor

#### Outputs

output : T
The arcsine of the input tensor computed element-wise

#### Type Constraints

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

#### Examples

asin
node = onnx.helper.make_node(
'Asin',
inputs=['x'],
outputs=['y'],
)

x = np.array([-0.5, 0, 0.5]).astype(np.float32)
y = np.arcsin(x)
expect(node, inputs=[x], outputs=[y],
name='test_asin_example')

x = np.random.rand(3, 4, 5).astype(np.float32)
y = np.arcsin(x)
expect(node, inputs=[x], outputs=[y],
name='test_asin')

### Atan

Calculates the arctangent (inverse of tangent) of the given input tensor, element-wise.

#### Version

This version of the operator has been available since version 7 of the default ONNX operator set.

input : T
Input tensor

#### Outputs

output : T
The arctangent of the input tensor computed element-wise

#### Type Constraints

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

#### Examples

atan
node = onnx.helper.make_node(
'Atan',
inputs=['x'],
outputs=['y'],
)

x = np.array([-1, 0, 1]).astype(np.float32)
y = np.arctan(x)
expect(node, inputs=[x], outputs=[y],
name='test_atan_example')

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.arctan(x)
expect(node, inputs=[x], outputs=[y],
name='test_atan')

### AveragePool

AveragePool consumes an input tensor X and applies average pooling across the tensor according to kernel sizes, stride sizes, and pad lengths. average pooling consisting of computing the average 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 divided by the number of elements (exclude pad when attribute count_include_pad is zero).

#### Version

This version of the operator has been available since version 7 of the default ONNX operator set.

Other versions of this operator: AveragePool-1

#### Attributes

auto_pad : string (default is NOTSET)
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 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. DEPRECATION NOTE: auto_pad is only intended to support legacy uses, and for framework authors, one is explicitly encouraged to use explicit padding specified in the pads attribute.
count_include_pad : int (default is 0)
Whether include pad pixels when calculating values for the edges. Default is 0, doesn't count include pad.
kernel_shape : list of ints (required)
The size of the kernel along each axis.
pads : list of ints
Padding for the beginning and ending along each 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 axis.
strides : list of ints
Stride along each axis.

#### Inputs

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

#### Examples

averagepool_1d_default
"""
input_shape: [1, 3, 32]
output_shape: [1, 3, 31]
"""
node = onnx.helper.make_node(
'AveragePool',
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)
y = pool(padded, x_shape, kernel_shape, strides, out_shape, [0], 'AVG')

expect(node, inputs=[x], outputs=[y], name='test_averagepool_1d_default')
averagepool_2d_default
"""
input_shape: [1, 3, 32, 32]
output_shape: [1, 3, 31, 31]
"""
node = onnx.helper.make_node(
'AveragePool',
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)
y = pool(padded, x_shape, kernel_shape, strides, out_shape, (0, 0), 'AVG')

expect(node, inputs=[x], outputs=[y], name='test_averagepool_2d_default')
"""
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(
'AveragePool',
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)
out_shape = get_output_shape('VALID', np.add(x_shape[2:], pad_shape), kernel_shape, strides)
constant_values=np.nan)
y = pool(padded, x_shape, kernel_shape, strides, out_shape, pad_shape, 'AVG')

expect(node, inputs=[x], outputs=[y], name='test_averagepool_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(
'AveragePool',
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)
out_shape = get_output_shape('VALID', np.add(x_shape[2:], pad_shape), kernel_shape, strides)
constant_values=0)
y = pool(padded, x_shape, kernel_shape, strides, out_shape, pad_shape, 'AVG', count_include_pad=1)

expect(node, inputs=[x], outputs=[y], name='test_averagepool_2d_pads_count_include_pad')
"""
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(
'AveragePool',
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([[[[7, 7.5, 8, 8.5, 9],
[9.5, 10, 10.5, 11, 11.5],
[12, 12.5, 13, 13.5, 14],
[14.5, 15, 15.5, 16, 16.5],
[17, 17.5, 18, 18.5, 19]]]]).astype(np.float32)

expect(node, inputs=[x], outputs=[y], name='test_averagepool_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(
'AveragePool',
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([[[[2.5200, 3.6000, 4.8000, 4.0800, 3.2400],
[4.5600, 6.4000, 8.4000, 7.0400, 5.5200],
[7.2000, 10.0000, 13.0000, 10.8000, 8.4000],
[6.9600, 9.6000, 12.4000, 10.2400, 7.9200],
[6.1200, 8.4000, 10.8000, 8.8800, 6.8400]]]]).astype(np.float32)

expect(node, inputs=[x], outputs=[y], name='test_averagepool_2d_precomputed_pads_count_include_pad')
averagepool_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(
'AveragePool',
inputs=['x'],
outputs=['y'],
kernel_shape=[3, 3],
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([[[[4, 5.5, 7],
[11.5, 13, 14.5],
[19, 20.5, 22]]]]).astype(np.float32)

expect(node, inputs=[x], outputs=[y], name='test_averagepool_2d_precomputed_same_upper')
averagepool_2d_precomputed_strides
"""
input_shape: [1, 1, 5, 5]
output_shape: [1, 1, 2, 2]
"""
node = onnx.helper.make_node(
'AveragePool',
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([[[[4, 6],
[14, 16]]]]).astype(np.float32)

expect(node, inputs=[x], outputs=[y], name='test_averagepool_2d_precomputed_strides')
averagepool_2d_same_lower
"""
iutput_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(
'AveragePool',
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('SAME_LOWER', x_shape[2:], kernel_shape, strides)
pad_shape = get_pad_shape('SAME_LOWER', x_shape[2:], kernel_shape, strides, out_shape)
constant_values=np.nan)
y = pool(padded, x_shape, kernel_shape, strides, out_shape, pad_shape, 'AVG')

expect(node, inputs=[x], outputs=[y], name='test_averagepool_2d_same_lower')
averagepool_2d_same_upper
"""
iutput_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(
'AveragePool',
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('SAME_UPPER', x_shape[2:], kernel_shape, strides)
pad_shape = get_pad_shape('SAME_UPPER', x_shape[2:], kernel_shape, strides, out_shape)
constant_values=np.nan)
y = pool(padded, x_shape, kernel_shape, strides, out_shape, pad_shape, 'AVG')

expect(node, inputs=[x], outputs=[y], name='test_averagepool_2d_same_upper')
averagepool_2d_strides
"""
input_shape: [1, 3, 32, 32]
output_shape: [1, 3, 10, 10]
"""
node = onnx.helper.make_node(
'AveragePool',
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)
y = pool(padded, x_shape, kernel_shape, strides, out_shape, (0, 0), 'AVG')

expect(node, inputs=[x], outputs=[y], name='test_averagepool_2d_strides')
averagepool_3d_default
"""
input_shape: [1, 3, 32, 32, 32]
output_shape: [1, 3, 31, 31, 31]
"""
node = onnx.helper.make_node(
'AveragePool',
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)
y = pool(padded, x_shape, kernel_shape, strides, out_shape, [0, 0, 0], 'AVG')

expect(node, inputs=[x], outputs=[y], name='test_averagepool_3d_default')

### BatchNormalization

Carries out batch normalization as described in the paper https://arxiv.org/abs/1502.03167. Depending on the mode it is being run, there are multiple cases for the number of outputs, which we list below:

Output case #1: Y, mean, var, saved_mean, saved_var (training mode) Output case #2: Y (test mode) This operator has optional inputs/outputs. See the doc for more details about the representation of optional arguments. An empty string may be used in the place of an actual argument's name to indicate a missing argument. Trailing optional arguments (those not followed by an argument that is present) may also be simply omitted.

#### Version

This version of the operator has been available since version 7 of the default ONNX operator set.

Other versions of this operator: BatchNormalization-1, BatchNormalization-6

#### Attributes

epsilon : float (default is 1e-05)
The epsilon value to use to avoid division by zero.
momentum : float (default is 0.9)
Factor used in computing the running mean and variance.e.g., running_mean = running_mean * momentum + mean * (1 - momentum).
spatial : int (default is 1)
If true, compute the mean and variance across all spatial elements If false, compute the mean and variance across per feature.

#### Inputs

X : 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.
scale : T
The scale as a 1-dimensional tensor of size C to be applied to the output.
B : T
The bias as a 1-dimensional tensor of size C to be applied to the output.
mean : T
The running mean (training) or the estimated mean (testing) as a 1-dimensional tensor of size C.
var : T
The running variance (training) or the estimated variance (testing) as a 1-dimensional tensor of size C.

#### Outputs (1 - 5)

Y : T
The output tensor of the same shape as X.
mean (optional) : T
The running mean after the BatchNormalization operator.
var (optional) : T
The running variance after the BatchNormalization operator.
saved_mean (optional) : T
Saved mean used during training to speed up gradient computation.
saved_var (optional) : T
Saved variance used during training to speed up gradient computation.

#### Type Constraints

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

#### Examples

batchnormalization
def _batchnorm_test_mode(x, s, bias, mean, var, epsilon=1e-5):  # type: ignore
dims_x = len(x.shape)
dim_ones = (1,) * (dims_x - 2)
s = s.reshape(-1, *dim_ones)
bias = bias.reshape(-1, *dim_ones)
mean = mean.reshape(-1, *dim_ones)
var = var.reshape(-1, *dim_ones)
return s * (x - mean) / np.sqrt(var + epsilon) + bias

# input size: (1, 2, 1, 3)
x = np.array([[[[-1, 0, 1]], [[2, 3, 4]]]]).astype(np.float32)
s = np.array([1.0, 1.5]).astype(np.float32)
bias = np.array([0, 1]).astype(np.float32)
mean = np.array([0, 3]).astype(np.float32)
var = np.array([1, 1.5]).astype(np.float32)
y = _batchnorm_test_mode(x, s, bias, mean, var).astype(np.float32)

node = onnx.helper.make_node(
'BatchNormalization',
inputs=['x', 's', 'bias', 'mean', 'var'],
outputs=['y'],
)

# output size: (1, 2, 1, 3)
expect(node, inputs=[x, s, bias, mean, var], outputs=[y],
name='test_batchnorm_example')

# input size: (2, 3, 4, 5)
x = np.random.randn(2, 3, 4, 5).astype(np.float32)
s = np.random.randn(3).astype(np.float32)
bias = np.random.randn(3).astype(np.float32)
mean = np.random.randn(3).astype(np.float32)
var = np.random.rand(3).astype(np.float32)
epsilon = 1e-2
y = _batchnorm_test_mode(x, s, bias, mean, var, epsilon).astype(np.float32)

node = onnx.helper.make_node(
'BatchNormalization',
inputs=['x', 's', 'bias', 'mean', 'var'],
outputs=['y'],
epsilon=epsilon,
)

# output size: (2, 3, 4, 5)
expect(node, inputs=[x, s, bias, mean, var], outputs=[y],
name='test_batchnorm_epsilon')

### Cast

The operator casts the elements of a given input tensor to a data type specified by the 'to' argument and returns an output tensor of the same size in the converted type. The 'to' argument must be one of the data types specified in the 'DataType' enum field in the TensorProto message. NOTE: Casting to and from strings is not supported yet.

#### Version

This version of the operator has been available since version 6 of the default ONNX operator set.

Other versions of this operator: Cast-1

#### Attributes

to : int (required)
The data type to which the elements of the input tensor are cast.Strictly must be one of the types from DataType enum in TensorProto

#### Inputs

input : T1
Input tensor to be cast.

#### Outputs

output : T2
Output tensor with the same shape as input with type specified by the 'to' argument

#### Type Constraints

T1 : tensor(float16), tensor(float), tensor(double), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(bool)
Constrain input types. Casting from strings and complex are not supported.
T2 : tensor(float16), tensor(float), tensor(double), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(bool)
Constrain output types. Casting to strings and complex are not supported.

#### Examples

cast
shape = (3, 4)
test_cases = [
('FLOAT', 'FLOAT16'),
('FLOAT', 'DOUBLE'),
('FLOAT16', 'FLOAT'),
('FLOAT16', 'DOUBLE'),
('DOUBLE', 'FLOAT'),
('DOUBLE', 'FLOAT16'),
]

for from_type, to_type in test_cases:
input = np.random.random_sample(shape).astype(
TENSOR_TYPE_TO_NP_TYPE[getattr(TensorProto, from_type)])
node = onnx.helper.make_node(
'Cast',
inputs=['input'],
outputs=['output'],
to=getattr(TensorProto, to_type),
)
output = input.astype(TENSOR_TYPE_TO_NP_TYPE[getattr(TensorProto, to_type)])

expect(node, inputs=[input], outputs=[output],
name='test_cast_' + from_type + '_to_' + to_type)

### Ceil

Ceil takes one input data (Tensor) and produces one output data (Tensor) where the ceil is, y = ceil(x), is applied to the tensor elementwise.

#### Version

This version of the operator has been available since version 6 of the default ONNX operator set.

Other versions of this operator: Ceil-1

X : T
Input tensor

Y : T
Output tensor

#### Type Constraints

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

#### Examples

ceil
node = onnx.helper.make_node(
'Ceil',
inputs=['x'],
outputs=['y'],
)

x = np.array([-1.5, 1.2]).astype(np.float32)
y = np.ceil(x)  # expected output [-1., 2.]
expect(node, inputs=[x], outputs=[y],
name='test_ceil_example')

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.ceil(x)
expect(node, inputs=[x], outputs=[y],
name='test_ceil')

### Clip

Clip operator limits the given input within an interval. The interval is specified with arguments 'min' and 'max'. They default to numeric_limits::lowest() and numeric_limits::max() respectively.

#### Version

This version of the operator has been available since version 6 of the default ONNX operator set.

Other versions of this operator: Clip-1

#### Attributes

max : float (default is 3.4028234663852886e+38)
Maximum value, above which element is replaced by max
min : float (default is -3.4028234663852886e+38)
Minimum value, under which element is replaced by min

#### Inputs

input : T
Input tensor whose elements to be clipped

#### Outputs

output : T
Output tensor with clipped input elements

#### Type Constraints

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

#### Examples

clip
node = onnx.helper.make_node(
'Clip',
inputs=['x'],
outputs=['y'],
min=-1.0,
max=1.0
)

x = np.array([-2, 0, 2]).astype(np.float32)
y = np.clip(x, -1, 1)  # expected output [-1., 0., 1.]
expect(node, inputs=[x], outputs=[y],
name='test_clip_example')

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.clip(x, -1.0, 1.0)
expect(node, inputs=[x], outputs=[y],
name='test_clip')
node = onnx.helper.make_node(
'Clip',
inputs=['x'],
outputs=['y'],
min=-5.0,
max=5.0,
)

x = np.array([-1, 0, 1]).astype(np.float32)
y = np.array([-1, 0, 1]).astype(np.float32)
expect(node, inputs=[x], outputs=[y],
name='test_clip_inbounds')

x = np.array([-6, 0, 6]).astype(np.float32)
y = np.array([-5, 0, 5]).astype(np.float32)
expect(node, inputs=[x], outputs=[y],
name='test_clip_outbounds')

x = np.array([-1, 0, 6]).astype(np.float32)
y = np.array([-1, 0, 5]).astype(np.float32)
expect(node, inputs=[x], outputs=[y],
name='test_clip_splitbounds')
clip_default
node = onnx.helper.make_node(
'Clip',
inputs=['x'],
outputs=['y'],
min=0.0
)
x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.clip(x, 0.0, np.inf)
expect(node, inputs=[x], outputs=[y],
name='test_clip_default_min')

node = onnx.helper.make_node(
'Clip',
inputs=['x'],
outputs=['y'],
max=0.0
)
x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.clip(x, -np.inf, 0.0)
expect(node, inputs=[x], outputs=[y],
name='test_clip_default_max')
node = onnx.helper.make_node(
'Clip',
inputs=['x'],
outputs=['y'],
)

x = np.array([-1, 0, 1]).astype(np.float32)
y = np.array([-1, 0, 1]).astype(np.float32)
expect(node, inputs=[x], outputs=[y],
name='test_clip_default_inbounds')

### Concat

Concatenate a list of tensors into a single tensor

#### Version

This version of the operator has been available since version 4 of the default ONNX operator set.

Other versions of this operator: Concat-1

#### Attributes

axis : int (required)
Which axis to concat on

#### Inputs (1 - ∞)

inputs (variadic) : T
List of tensors for concatenation

#### Outputs

concat_result : T
Concatenated tensor

#### Type Constraints

T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double), tensor(string), tensor(bool), tensor(complex64), tensor(complex128)
Constrain output types to any tensor type.

#### Examples

concat
test_cases = {
'1d': ([1, 2],
[3, 4]),
'2d': ([[1, 2], [3, 4]],
[[5, 6], [7, 8]]),
'3d': ([[[1, 2], [3, 4]], [[5, 6], [7, 8]]],
[[[9, 10], [11, 12]], [[13, 14], [15, 16]]])
}  # type: Dict[Text, Sequence[Any]]

for test_case, values_ in test_cases.items():
values = [np.asarray(v, dtype=np.float32) for v in values_]
for i in range(len(values[0].shape)):
in_args = ['value' + str(k) for k in range(len(values))]
node = onnx.helper.make_node(
'Concat',
inputs=[s for s in in_args],
outputs=['output'],
axis=i
)
output = np.concatenate(values, i)
expect(node, inputs=[v for v in values], outputs=[output],
name='test_concat_' + test_case + '_axis_' + str(i))

### Constant

A constant tensor.

#### Version

This version of the operator has been available since version 9 of the default ONNX operator set.

Other versions of this operator: Constant-1

#### Attributes

value : tensor (required)
The value for the elements of the output tensor.

#### Outputs

output : T
Output tensor containing the same value of the provided tensor.

#### Type Constraints

T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double), tensor(string), tensor(bool), tensor(complex64), tensor(complex128)
Constrain input and output types to all tensor types.

#### Examples

constant
values = np.random.randn(5, 5).astype(np.float32)
node = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=['values'],
value=onnx.helper.make_tensor(
name='const_tensor',
data_type=onnx.TensorProto.FLOAT,
dims=values.shape,
vals=values.flatten().astype(float),
),
)

expect(node, inputs=[], outputs=[values],
name='test_constant')

### ConstantLike

Generate a tensor with specific constant value. The value can be specified by the 'value' attribute. The shape of the output tensor is the same as the input tensor, if the input tensor is provided, or the shape provided in the 'shape' attribute (if both are provided, the input tensor shape takes precendence). The data type can be specified by the 'dtype' argument. If 'dtype' is not specified, then the type of input tensor is used. If input tensor is also not specified, then the type defaults to 'float'.

The 'dtype' argument must be one of the data types specified in the 'DataType' enum field in the TensorProto message and be valid as an output type.

#### Version

This version of the operator has been available since version 9 of the default ONNX operator set.

#### Attributes

dtype : int
(Optional) The data type for the elements of the output tensor. If not specified,the data type of the input tensor T1 is used. If input tensor T1 is also not specified, then output tensor type defaults to 'float'.
shape : list of ints
(Optional) The shape of the output tensor. If input tensor T1 is provided, then 'shape' attribute is ignored and the output follows the shape of the input. One of either input tensor T1 or 'shape' attribute must be provided.
value : float (default is 0.0)
(Optional) The value for the elements of the output tensor.

#### Inputs (0 - 1)

input (optional) : T1
Input tensor to copy shape, and optionally, type information from. One of either input tensor T1 or 'shape' attribute must be provided.

#### Outputs

output : T2
Output tensor, same shape as input tensor T1.

#### Type Constraints

T1 : tensor(float16), tensor(float), tensor(double), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(bool)
Constrain input types. Strings and complex are not supported.
T2 : tensor(float16), tensor(float), tensor(double), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(bool)
Constrain output types. Strings and complex are not supported.

#### Examples

ones_with_input
shape = (4, 3, 2)
node = onnx.helper.make_node(
'ConstantLike',
inputs=['x'],
outputs=['y'],
value=1.0,
)
x = np.random.randint(0, 100, size=shape, dtype=np.int32)
y = np.ones(shape, dtype=np.int32)
expect(node, inputs=[x], outputs=[y], name='test_constantlike_ones_with_input')
threes_with_shape_and_dtype
shape = (3, 4)
node = onnx.helper.make_node(
'ConstantLike',
shape=shape,
inputs=[],
outputs=['y'],
value=3.0,
dtype=onnx.TensorProto.DOUBLE,  # 11: DOUBLE
)

y = 3.0 * np.ones(shape, dtype=np.float64)
expect(node, inputs=[], outputs=[y], name='test_constantlike_threes_with_shape_and_dtype')
zeros_without_input_dtype
shape = (2, 5, 1)
node = onnx.helper.make_node(
'ConstantLike',
inputs=[],
outputs=['y'],
shape=shape,
)
y = np.zeros(shape, dtype=np.float32)
expect(node, inputs=[], outputs=[y], name='test_constantlike_zeros_without_input_dtype')

### Conv

The convolution operator consumes an input tensor and a filter, and computes the output.

#### Version

This version of the operator has been available since version 1 of the default ONNX operator set.

#### Attributes

auto_pad : string (default is NOTSET)
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 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. DEPRECATION NOTE: auto_pad is only intended to support legacy uses, and for framework authors, one is explicitly encouraged to use explicit padding specified in the pads attribute.
dilations : list of ints
dilation value along each axis of the filter.
group : int (default is 1)
number of groups input channels and output channels are divided into.
kernel_shape : list of ints
The shape of the convolution kernel. If not present, should be inferred from input W.
pads : list of ints
Padding for the beginning and ending along each 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 axis.
strides : list of ints
Stride along each axis.

#### Inputs (2 - 3)

X : T
Input data tensor from previous layer; has size (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 width. Note that this is for the 2D image. Otherwise the size is (N x C x D1 x D2 ... x Dn). Optionally, if dimension denotation is in effect, the operation expects input data tensor to arrive with the dimension denotation of [DATA_BATCH, DATA_CHANNEL, DATA_FEATURE, DATA_FEATURE ...].
W : T
The weight tensor that will be used in the convolutions; has size (M x C/group x kH x kW), where C is the number of channels, and kH and kW are the height and width of the kernel, and M is the number of feature maps. For more than 2 dimensions, the kernel shape will be (M x C/group x k1 x k2 x ... x kn), where (k1 x k2 x ... kn) is the dimension of the kernel. Optionally, if dimension denotation is in effect, the operation expects the weight tensor to arrive with the dimension denotation of [FILTER_OUT_CHANNEL, FILTER_IN_CHANNEL, FILTER_SPATIAL, FILTER_SPATIAL ...]. X.shape[1] == (W.shape[1] * group) == C (assuming zero based indices for the shape array). Or in other words FILTER_IN_CHANNEL should be equal to DATA_CHANNEL.
B (optional) : T
Optional 1D bias to be added to the convolution, has size of M.

#### Outputs

Y : T
Output data tensor that contains the result of the convolution. The output dimensions are functions of the kernel size, stride size, and pad lengths.

#### Type Constraints

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

#### Examples

conv
x = np.array([[[[0., 1., 2., 3., 4.],  # (1, 1, 5, 5) input tensor
[5., 6., 7., 8., 9.],
[10., 11., 12., 13., 14.],
[15., 16., 17., 18., 19.],
[20., 21., 22., 23., 24.]]]]).astype(np.float32)
W = np.array([[[[1., 1., 1.],  # (1, 1, 3, 3) tensor for convolution weights
[1., 1., 1.],
[1., 1., 1.]]]]).astype(np.float32)

# Convolution with padding
'Conv',
inputs=['x', 'W'],
outputs=['y'],
kernel_shape=[3, 3],
# Default values for other attributes: strides=[1, 1], dilations=[1, 1], groups=1
pads=[1, 1, 1, 1],
)
y_with_padding = np.array([[[[12., 21., 27., 33., 24.],  # (1, 1, 5, 5) output tensor
[33., 54., 63., 72., 51.],
[63., 99., 108., 117., 81.],
[93., 144., 153., 162., 111.],
[72., 111., 117., 123., 84.]]]]).astype(np.float32)

# Convolution without padding
'Conv',
inputs=['x', 'W'],
outputs=['y'],
kernel_shape=[3, 3],
# Default values for other attributes: strides=[1, 1], dilations=[1, 1], groups=1
pads=[0, 0, 0, 0],
)
y_without_padding = np.array([[[[54., 63., 72.],  # (1, 1, 3, 3) output tensor
[99., 108., 117.],
[144., 153., 162.]]]]).astype(np.float32)
name='test_basic_conv_without_padding')
conv_with_strides
x = np.array([[[[0., 1., 2., 3., 4.],  # (1, 1, 7, 5) input tensor
[5., 6., 7., 8., 9.],
[10., 11., 12., 13., 14.],
[15., 16., 17., 18., 19.],
[20., 21., 22., 23., 24.],
[25., 26., 27., 28., 29.],
[30., 31., 32., 33., 34.]]]]).astype(np.float32)
W = np.array([[[[1., 1., 1.],  # (1, 1, 3, 3) tensor for convolution weights
[1., 1., 1.],
[1., 1., 1.]]]]).astype(np.float32)

# Convolution with strides=2 and padding
'Conv',
inputs=['x', 'W'],
outputs=['y'],
kernel_shape=[3, 3],
pads=[1, 1, 1, 1],
strides=[2, 2],  # Default values for other attributes: dilations=[1, 1], groups=1
)
y_with_padding = np.array([[[[12., 27., 24.],  # (1, 1, 4, 3) output tensor
[63., 108., 81.],
[123., 198., 141.],
[112., 177., 124.]]]]).astype(np.float32)

# Convolution with strides=2 and no padding
'Conv',
inputs=['x', 'W'],
outputs=['y'],
kernel_shape=[3, 3],
pads=[0, 0, 0, 0],
strides=[2, 2],  # Default values for other attributes: dilations=[1, 1], groups=1
)
y_without_padding = np.array([[[[54., 72.],  # (1, 1, 3, 2) output tensor
[144., 162.],
[234., 252.]]]]).astype(np.float32)

# Convolution with strides=2 and padding only along one dimension (the H dimension in NxCxHxW tensor)
'Conv',
inputs=['x', 'W'],
outputs=['y'],
kernel_shape=[3, 3],
pads=[1, 0, 1, 0],
strides=[2, 2],  # Default values for other attributes: dilations=[1, 1], groups=1
)
y_with_asymmetric_padding = np.array([[[[21., 33.],  # (1, 1, 4, 2) output tensor
[99., 117.],
[189., 207.],
[171., 183.]]]]).astype(np.float32)
name='test_conv_with_strides_and_asymmetric_padding')

### ConvTranspose

The convolution transpose operator consumes an input tensor and a filter, and computes the output.

If the pads parameter is provided the shape of the output is calculated via the following equation:

output_shape[i] = stride[i] * (input_size[i] - 1) + output_padding[i] + kernel_shape[i] - pads[start_i] - pads[end_i]


output_shape can also be explicitly specified in which case pads values are auto generated using these equations:

total_padding[i] = stride[i] * (input_size[i] - 1) + output_padding[i] + kernel_shape[i] - output_shape[i]


#### Version

This version of the operator has been available since version 1 of the default ONNX operator set.

#### Attributes

auto_pad : string (default is NOTSET)
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 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. DEPRECATION NOTE: auto_pad is only intended to support legacy uses, and for framework authors, one is explicitly encouraged to use explicit padding specified in the pads attribute.
dilations : list of ints
dilation value along each axis of the filter.
group : int (default is 1)
number of groups input channels and output channels are divided into.
kernel_shape : list of ints
The shape of the convolution kernel. If not present, should be inferred from input W.
output_padding : list of ints
The zero-padding added to one side of the output. This is also called adjs/adjustment in some frameworks.
output_shape : list of ints
The shape of the output can be explicitly set which will cause pads values to be auto generated. If output_shape is specified pads values are ignored. See doc for details for equations to generate pads
pads : list of ints
Padding for the beginning and ending along each 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 axis.
strides : list of ints
Stride along each axis.

#### Inputs (2 - 3)

X : T
Input data tensor from previous layer; has size (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 width. Note that this is for the 2D image. Otherwise the size is (N x C x D1 x D2 ... x Dn)
W : T
The weight tensor that will be used in the convolutions; has size (C x M/group x kH x kW), where C is the number of channels, and kH and kW are the height and width of the kernel, and M is the number of feature maps. For more than 2 dimensions, the weight shape will be (C x M/group x k1 x k2 x ... x kn), where (k1 x k2 x ... x kn) is the dimension of the kernel. The number of channels in the output should be equal to W.shape[1] * group (assuming zero based indices of the shape array)
B (optional) : T
Optional 1D bias to be added to the convolution, has size of C.

#### Outputs

Y : T
Output data tensor that contains the result of the convolution. The output dimensions are functions of the kernel size, stride size, pad lengths and group count. The number of channels in the output should be equal to W.shape[1] * group (assuming zero based indices of the shape array)

#### Type Constraints

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

#### Examples

convtranspose
x = np.array([[[[0., 1., 2.],  # (1, 1, 3, 3)
[3., 4., 5.],
[6., 7., 8.]]]]).astype(np.float32)

W = np.array([[[[1., 1., 1.],  # (1, 2, 3, 3)
[1., 1., 1.],
[1., 1., 1.]],
[[1., 1., 1.],
[1., 1., 1.],
[1., 1., 1.]]]]).astype(np.float32)

node = onnx.helper.make_node("ConvTranspose", ["X", "W"], ["Y"])

y = np.array([[[[0., 1., 3., 3., 2.],  # (1, 2, 5, 5)
[3., 8., 15., 12., 7.],
[9., 21., 36., 27., 15.],
[9., 20., 33., 24., 13.],
[6., 13., 21., 15., 8.]],

[[0., 1., 3., 3., 2.],
[3., 8., 15., 12., 7.],
[9., 21., 36., 27., 15.],
[9., 20., 33., 24., 13.],
[6., 13., 21., 15., 8.]]]]).astype(np.float32)

expect(node, inputs=[x, W], outputs=[y], name='test_convtranspose')
convtranspose_1d
x = np.array([[[0., 1., 2.]]]).astype(np.float32)  # (1, 1, 3)

W = np.array([[[1., 1., 1.],  # (1, 2, 3)
[1., 1., 1.]]]).astype(np.float32)

node = onnx.helper.make_node("ConvTranspose", ["X", "W"], ["Y"])

y = np.array([[[0., 1., 3., 3., 2.],  # (1, 2, 5)
[0., 1., 3., 3., 2.]]]).astype(np.float32)

expect(node, inputs=[x, W], outputs=[y], name='test_convtranspose_1d')
convtranspose_3d
x = np.array([[[[[0., 1., 2., 3., 4.],  # (1, 1, 3, 4, 5)
[5., 6., 7., 8., 9.],
[10., 11., 12., 13., 14.],
[15., 16., 17., 18., 19.]],
[[20., 21., 22., 23., 24.],
[25., 26., 27., 28., 29.],
[30., 31., 32., 33., 34.],
[35., 36., 37., 38., 39.]],
[[40., 41., 42., 43., 44.],
[45., 46., 47., 48., 49.],
[50., 51., 52., 53., 54.],
[55., 56., 57., 58., 59.]]]]]).astype(np.float32)

W = np.array([[[[[1., 1., 1.],  # (1, 2, 3, 3, 3)
[1., 1., 1.],
[1., 1., 1.]],
[[1., 1., 1.],
[1., 1., 1.],
[1., 1., 1.]],
[[1., 1., 1.],
[1., 1., 1.],
[1., 1., 1.]]],
[[[1., 1., 1.],
[1., 1., 1.],
[1., 1., 1.]],
[[1., 1., 1.],
[1., 1., 1.],
[1., 1., 1.]],
[[1., 1., 1.],
[1., 1., 1.],
[1., 1., 1.]]]]]).astype(np.float32)

node = onnx.helper.make_node("ConvTranspose", ["X", "W"], ["Y"])

y = np.array([[[[[0., 1., 3., 6., 9., 7., 4.],  # (1, 2, 5, 6, 7)
[5., 12., 21., 27., 33., 24., 13.],
[15., 33., 54., 63., 72., 51., 27.],
[30., 63., 99., 108., 117., 81., 42.],
[25., 52., 81., 87., 93., 64., 33.],
[15., 31., 48., 51., 54., 37., 19.]],

[[20., 42., 66., 72., 78., 54., 28.],
[50., 104., 162., 174., 186., 128., 66.],
[90., 186., 288., 306., 324., 222., 114.],
[120., 246., 378., 396., 414., 282., 144.],
[90., 184., 282., 294., 306., 208., 106.],
[50., 102., 156., 162., 168., 114., 58.]],

[[60., 123., 189., 198., 207., 141., 72.],
[135., 276., 423., 441., 459., 312., 159.],
[225., 459., 702., 729., 756., 513., 261.],
[270., 549., 837., 864., 891., 603., 306.],
[195., 396., 603., 621., 639., 432., 219.],
[105., 213., 324., 333., 342., 231., 117.]],

[[60., 122., 186., 192., 198., 134., 68.],
[130., 264., 402., 414., 426., 288., 146.],
[210., 426., 648., 666., 684., 462., 234.],
[240., 486., 738., 756., 774., 522., 264.],
[170., 344., 522., 534., 546., 368., 186.],
[90., 182., 276., 282., 288., 194., 98.]],

[[40., 81., 123., 126., 129., 87., 44.],
[85., 172., 261., 267., 273., 184., 93.],
[135., 273., 414., 423., 432., 291., 147.],
[150., 303., 459., 468., 477., 321., 162.],
[105., 212., 321., 327., 333., 224., 113.],
[55., 111., 168., 171., 174., 117., 59.]]],

[[[0., 1., 3., 6., 9., 7., 4.],
[5., 12., 21., 27., 33., 24., 13.],
[15., 33., 54., 63., 72., 51., 27.],
[30., 63., 99., 108., 117., 81., 42.],
[25., 52., 81., 87., 93., 64., 33.],
[15., 31., 48., 51., 54., 37., 19.]],

[[20., 42., 66., 72., 78., 54., 28.],
[50., 104., 162., 174., 186., 128., 66.],
[90., 186., 288., 306., 324., 222., 114.],
[120., 246., 378., 396., 414., 282., 144.],
[90., 184., 282., 294., 306., 208., 106.],
[50., 102., 156., 162., 168., 114., 58.]],

[[60., 123., 189., 198., 207., 141., 72.],
[135., 276., 423., 441., 459., 312., 159.],
[225., 459., 702., 729., 756., 513., 261.],
[270., 549., 837., 864., 891., 603., 306.],
[195., 396., 603., 621., 639., 432., 219.],
[105., 213., 324., 333., 342., 231., 117.]],

[[60., 122., 186., 192., 198., 134., 68.],
[130., 264., 402., 414., 426., 288., 146.],
[210., 426., 648., 666., 684., 462., 234.],
[240., 486., 738., 756., 774., 522., 264.],
[170., 344., 522., 534., 546., 368., 186.],
[90., 182., 276., 282., 288., 194., 98.]],

[[40., 81., 123., 126., 129., 87., 44.],
[85., 172., 261., 267., 273., 184., 93.],
[135., 273., 414., 423., 432., 291., 147.],
[150., 303., 459., 468., 477., 321., 162.],
[105., 212., 321., 327., 333., 224., 113.],
[55., 111., 168., 171., 174., 117., 59.]]]]]).astype(np.float32)

expect(node, inputs=[x, W], outputs=[y], name='test_convtranspose_3d')
convtranspose_attributes
x = np.array([[[[0., 1., 2.],  # (1, 1, 3, 3)
[3., 4., 5.],
[6., 7., 8.]]]]).astype(np.float32)

W = np.array([[[[1., 1., 1.],  # (1, 2, 3, 3)
[1., 1., 1.],
[1., 1., 1.]],
[[1., 1., 1.],
[1., 1., 1.],
[1., 1., 1.]]]]).astype(np.float32)

y = np.array([[[[0., 0., 1., 1., 3., 2., 2., 0.],  # (1, 2, 10, 8)
[0., 0., 1., 1., 3., 2., 2., 0.],
[0., 0., 1., 1., 3., 2., 2., 0.],
[3., 3., 7., 4., 9., 5., 5., 0.],
[3., 3., 7., 4., 9., 5., 5., 0.],
[3., 3., 7., 4., 9., 5., 5., 0.],
[6., 6., 13., 7., 15., 8., 8., 0.],
[6., 6., 13., 7., 15., 8., 8., 0.],
[6., 6., 13., 7., 15., 8., 8., 0.],
[0., 0., 0., 0., 0., 0., 0., 0.]],

[[0., 0., 1., 1., 3., 2., 2., 0.],
[0., 0., 1., 1., 3., 2., 2., 0.],
[0., 0., 1., 1., 3., 2., 2., 0.],
[3., 3., 7., 4., 9., 5., 5., 0.],
[3., 3., 7., 4., 9., 5., 5., 0.],
[3., 3., 7., 4., 9., 5., 5., 0.],
[6., 6., 13., 7., 15., 8., 8., 0.],
[6., 6., 13., 7., 15., 8., 8., 0.],
[6., 6., 13., 7., 15., 8., 8., 0.],
[0., 0., 0., 0., 0., 0., 0., 0.]]]]).astype(np.float32)

node = onnx.helper.make_node("ConvTranspose", ["X", "W"], ["Y"],
strides=[3, 2],
output_shape=[10, 8])
expect(node, inputs=[x, W], outputs=[y], name='test_convtranspose_output_shape')

node = onnx.helper.make_node("ConvTranspose", ["X", "W"], ["Y"],
strides=[3, 2],
expect(node, inputs=[x, W], outputs=[y], name='test_convtranspose_pad')

node = onnx.helper.make_node(
'ConvTranspose', ['X', 'W'], ['Y'],
name='test',
strides=[3, 2],
output_shape=[10, 8],
kernel_shape=[3, 3],
)
expect(node, inputs=[x, W], outputs=[y],
name='test_convtranspose_kernel_shape')
x = np.array([[[[0., 1., 2.],  # (1, 1, 3, 3)
[3., 4., 5.],
[6., 7., 8.]]]]).astype(np.float32)

W = np.array([[[[1., 1., 1.],  # (1, 2, 3, 3)
[1., 1., 1.],
[1., 1., 1.]],
[[1., 1., 1.],
[1., 1., 1.],
[1., 1., 1.]]]]).astype(np.float32)

node = onnx.helper.make_node("ConvTranspose", ["X", "W"], ["Y"],
strides=[3, 2],
pads=[1, 2, 1, 2])

y = np.array([[[[1., 1., 3.],  # (1, 2, 7, 3)
[1., 1., 3.],
[7., 4., 9.],
[7., 4., 9.],
[7., 4., 9.],
[13., 7., 15.],
[13., 7., 15.]],

[[1., 1., 3.],
[1., 1., 3.],
[7., 4., 9.],
[7., 4., 9.],
[7., 4., 9.],
[13., 7., 15.],
[13., 7., 15.]]]]).astype(np.float32)

expect(node, inputs=[x, W], outputs=[y], name='test_convtranspose_pads')

### Cos

Calculates the cosine of the given input tensor, element-wise.

#### Version

This version of the operator has been available since version 7 of the default ONNX operator set.

input : T
Input tensor

#### Outputs

output : T
The cosine of the input tensor computed element-wise

#### Type Constraints

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

#### Examples

cos
node = onnx.helper.make_node(
'Cos',
inputs=['x'],
outputs=['y'],
)

x = np.array([-1, 0, 1]).astype(np.float32)
y = np.cos(x)
expect(node, inputs=[x], outputs=[y],
name='test_cos_example')

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.cos(x)
expect(node, inputs=[x], outputs=[y],
name='test_cos')

### DepthToSpace

DepthToSpace rearranges (permutes) data from depth into blocks of spatial data. This is the reverse transformation of SpaceToDepth. More specifically, this op outputs a copy of the input tensor where values from the depth dimension are moved in spatial blocks to the height and width dimensions.

#### Version

This version of the operator has been available since version 1 of the default ONNX operator set.

#### Attributes

blocksize : int (required)
Blocks of [blocksize, blocksize] are moved.

#### Inputs

input : T
Input tensor of [N,C,H,W], where N is the batch axis, C is the channel or depth, H is the height and W is the width.

#### Outputs

output : T
Output tensor of [N, C/(blocksize * blocksize), H * blocksize, W * blocksize].

#### Type Constraints

T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double), tensor(string), tensor(bool), tensor(complex64), tensor(complex128)
Constrain input and output types to all tensor types.

#### Examples

depthtospace
b, c, h, w = shape = (2, 8, 3, 3)
blocksize = 2
node = onnx.helper.make_node(
'DepthToSpace',
inputs=['x'],
outputs=['y'],
blocksize=blocksize,
)
x = np.random.random_sample(shape).astype(np.float32)
tmp = np.reshape(x, [b, blocksize, blocksize, c // (blocksize**2), h, w])
tmp = np.transpose(tmp, [0, 3, 4, 1, 5, 2])
y = np.reshape(tmp, [b, c // (blocksize**2), h * blocksize, w * blocksize])
expect(node, inputs=[x], outputs=[y],
name='test_depthtospace')
example
node = onnx.helper.make_node(
'DepthToSpace',
inputs=['x'],
outputs=['y'],
blocksize=2,
)

# (1, 4, 2, 3) input tensor
x = np.array([[[[0, 1, 2],
[3, 4, 5]],
[[6, 7, 8],
[9, 10, 11]],
[[12, 13, 14],
[15, 16, 17]],
[[18, 19, 20],
[21, 22, 23]]]]).astype(np.float32)

# (1, 1, 4, 6) output tensor
y = np.array([[[[0, 6, 1, 7, 2, 8],
[12, 18, 13, 19, 14, 20],
[3, 9, 4, 10, 5, 11],
[15, 21, 16, 22, 17, 23]]]]).astype(np.float32)
expect(node, inputs=[x], outputs=[y],
name='test_depthtospace_example')

### Div

Performs element-wise binary division (with Numpy-style broadcasting support).

This operator supports multidirectional (i.e., Numpy-style) broadcasting; for more details please check the doc.

#### Version

This version of the operator has been available since version 7 of the default ONNX operator set.

Other versions of this operator: Div-1, Div-6

A : T
First operand.
B : T
Second operand.

#### Outputs

C : T
Result, has same element type as two inputs

#### Type Constraints

T : tensor(uint32), tensor(uint64), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double)
Constrain input and output types to high-precision numeric tensors.

#### Examples

div
node = onnx.helper.make_node(
'Div',
inputs=['x', 'y'],
outputs=['z'],
)

x = np.array([3, 4]).astype(np.float32)
y = np.array([1, 2]).astype(np.float32)
z = x / y  # expected output [3., 2.]
expect(node, inputs=[x, y], outputs=[z],
name='test_div_example')

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.random.rand(3, 4, 5).astype(np.float32) + 1.0
z = x / y
expect(node, inputs=[x, y], outputs=[z],
name='test_div')
node = onnx.helper.make_node(
'Div',
inputs=['x', 'y'],
outputs=['z'],
)

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.random.rand(5).astype(np.float32) + 1.0
z = x / y
expect(node, inputs=[x, y], outputs=[z],
name='test_div_bcast')

### Dropout

Dropout takes one input data (Tensor) and produces two Tensor outputs, output (Tensor) and mask (Tensor). Depending on whether it is in test mode or not, the output Y will either be a random dropout, or a simple copy of the input. Note that our implementation of Dropout does scaling in the training phase, so during testing nothing needs to be done. This operator has optional inputs/outputs. See the doc for more details about the representation of optional arguments. An empty string may be used in the place of an actual argument's name to indicate a missing argument. Trailing optional arguments (those not followed by an argument that is present) may also be simply omitted.

#### Version

This version of the operator has been available since version 7 of the default ONNX operator set.

Other versions of this operator: Dropout-1, Dropout-6

#### Attributes

ratio : float (default is 0.5)
The ratio of random dropout

#### Inputs

data : T
The input data as Tensor.

#### Outputs (1 - 2)

output : T
The output.
mask (optional) : T

#### Type Constraints

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

#### Examples

default
node = onnx.helper.make_node(
'Dropout',
inputs=['x'],
outputs=['y'],
)

x = np.array([-1, 0, 1]).astype(np.float32)
y = x
expect(node, inputs=[x], outputs=[y],
name='test_dropout_default')
random
node = onnx.helper.make_node(
'Dropout',
inputs=['x'],
outputs=['y'],
ratio=.2,
)

x = np.random.randn(3, 4, 5).astype(np.float32)
y = x
expect(node, inputs=[x], outputs=[y],
name='test_dropout_random')

### Elu

Elu takes one input data (Tensor) and produces one output data (Tensor) where the function f(x) = alpha * (exp(x) - 1.) for x < 0, f(x) = x for x >= 0., is applied to the tensor elementwise.

#### Version

This version of the operator has been available since version 6 of the default ONNX operator set.

Other versions of this operator: Elu-1

#### Attributes

alpha : float (default is 1.0)
Coefficient of ELU.

X : T
1D input tensor

Y : T
1D input tensor

#### Type Constraints

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

#### Examples

elu
node = onnx.helper.make_node(
'Elu',
inputs=['x'],
outputs=['y'],
alpha=2.0
)

x = np.array([-1, 0, 1]).astype(np.float32)
# expected output [-1.2642411, 0., 1.]
y = np.clip(x, 0, np.inf) + (np.exp(np.clip(x, -np.inf, 0)) - 1) * 2.0
expect(node, inputs=[x], outputs=[y],
name='test_elu_example')

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.clip(x, 0, np.inf) + (np.exp(np.clip(x, -np.inf, 0)) - 1) * 2.0
expect(node, inputs=[x], outputs=[y],
name='test_elu')
elu_default
default_alpha = 1.0
node = onnx.helper.make_node(
'Elu',
inputs=['x'],
outputs=['y'],
)
x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.clip(x, 0, np.inf) + (np.exp(np.clip(x, -np.inf, 0)) - 1) * default_alpha
expect(node, inputs=[x], outputs=[y],
name='test_elu_default')

### Equal

Returns the tensor resulted from performing the equal logical operation elementwise on the input tensors A and B (with Numpy-style broadcasting support).

This operator supports multidirectional (i.e., Numpy-style) broadcasting; for more details please check the doc.

#### Version

This version of the operator has been available since version 7 of the default ONNX operator set.

Other versions of this operator: Equal-1

#### Inputs

A : T
First input operand for the logical operator.
B : T
Second input operand for the logical operator.

C : T1
Result tensor.

#### Type Constraints

T : tensor(bool), tensor(int32), tensor(int64)
Constrains input to integral tensors.
T1 : tensor(bool)
Constrains output to boolean tensor.

#### Examples

equal
node = onnx.helper.make_node(
'Equal',
inputs=['x', 'y'],
outputs=['z'],
)

x = (np.random.randn(3, 4, 5) * 10).astype(np.int32)
y = (np.random.randn(3, 4, 5) * 10).astype(np.int32)
z = np.equal(x, y)
expect(node, inputs=[x, y], outputs=[z],
name='test_equal')
node = onnx.helper.make_node(
'Equal',
inputs=['x', 'y'],
outputs=['z'],
)

x = (np.random.randn(3, 4, 5) * 10).astype(np.int32)
y = (np.random.randn(5) * 10).astype(np.int32)
z = np.equal(x, y)
expect(node, inputs=[x, y], outputs=[z],
name='test_equal_bcast')

### Exp

Calculates the exponential of the given input tensor, element-wise.

#### Version

This version of the operator has been available since version 6 of the default ONNX operator set.

Other versions of this operator: Exp-1

input : T
Input tensor

#### Outputs

output : T
The exponential of the input tensor computed element-wise

#### Type Constraints

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

#### Examples

exp
node = onnx.helper.make_node(
'Exp',
inputs=['x'],
outputs=['y'],
)

x = np.array([-1, 0, 1]).astype(np.float32)
y = np.exp(x)  # expected output [0.36787945, 1., 2.71828175]
expect(node, inputs=[x], outputs=[y],
name='test_exp_example')

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.exp(x)
expect(node, inputs=[x], outputs=[y],
name='test_exp')

### Expand

Broadcast the input tensor following the given shape and the broadcast rule. The broadcast rule is similar to numpy.array(input) * numpy.ones(shape): Dimensions are right alignment; Two corresponding dimension must have the same value, or one of them is equal to 1. Also, this operator is similar to numpy.broadcast_to(input, shape), but the major difference is numpy.broadcast_to() does not allow shape to be smaller than input.size(). It is possible that the output.shape is not equal to shape, when some dimensions in shape is equal to 1, or the shape.ndim < input.shape.ndim.

#### Version

This version of the operator has been available since version 8 of the default ONNX operator set.

#### Inputs

input : T
Input tensor
shape : tensor(int64)
A 1-D tensor indicates the shape you want to expand to, following the broadcast rule

output : T
Output tensor

#### Type Constraints

T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double), tensor(string), tensor(bool), tensor(complex64), tensor(complex128)
Constrain input and output types to all tensors.

#### Examples

dim_changed
node = onnx.helper.make_node(
'Expand',
inputs=['data', 'new_shape'],
outputs=['expanded'],
)
shape = [3, 1]
data = np.reshape(np.arange(1, np.prod(shape) + 1, dtype=np.float32), shape)
#print(data)
#[[1.], [2.], [3.]]
new_shape = [2, 1, 6]
expanded = data * np.ones(new_shape, dtype=np.float32)
#print(expanded)
#[[[1., 1., 1., 1., 1., 1.],
#  [2., 2., 2., 2., 2., 2.],
#  [3., 3., 3., 3., 3., 3.]],
#
# [[1., 1., 1., 1., 1., 1.],
#  [2., 2., 2., 2., 2., 2.],
#  [3., 3., 3., 3., 3., 3.]]]
new_shape = np.array(new_shape, dtype=np.int64)
expect(node, inputs=[data, new_shape], outputs=[expanded],
name='test_expand_dim_changed')
dim_unchanged
node = onnx.helper.make_node(
'Expand',
inputs=['data', 'new_shape'],
outputs=['expanded'],
)
shape = [3, 1]
new_shape = [3, 4]
data = np.reshape(np.arange(1, np.prod(shape) + 1, dtype=np.float32), shape)
#print(data)
#[[1.], [2.], [3.]]
expanded = np.tile(data, 4)
#print(expanded)
#[[1., 1., 1., 1.],
# [2., 2., 2., 2.],
# [3., 3., 3., 3.]]
new_shape = np.array(new_shape, dtype=np.int64)
expect(node, inputs=[data, new_shape], outputs=[expanded],
name='test_expand_dim_unchanged')

### EyeLike

Generate a 2D tensor (matrix) with ones on the diagonal and zeros everywhere else. Only 2D tensors are supported, i.e. input T1 must be of rank 2. The shape of the output tensor is the same as the input tensor. The data type can be specified by the 'dtype' argument. If 'dtype' is not specified, then the type of input tensor is used. By default, the main diagonal is populated with ones, but attribute 'k' can be used to populate upper or lower diagonals.

The 'dtype' argument must be one of the data types specified in the 'DataType' enum field in the TensorProto message and be valid as an output type.

#### Version

This version of the operator has been available since version 9 of the default ONNX operator set.

#### Attributes

dtype : int
(Optional) The data type for the elements of the output tensor. If not specified,the data type of the input tensor T1 is used. If input tensor T1 is also notspecified, then type defaults to 'float'.
k : int (default is 0)
(Optional) Index of the diagonal to be populated with ones. Default is 0. If T2 is the output, this op sets T2[i, i+k] = 1. k = 0 populates the main diagonal, k > 0 populates an upper diagonal, and k < 0 populates a lower diagonal.

#### Inputs

input : T1
2D input tensor to copy shape, and optionally, type information from.

#### Outputs

output : T2
Output tensor, same shape as input tensor T1.

#### Type Constraints

T1 : tensor(float16), tensor(float), tensor(double), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(bool)
Constrain input types. Strings and complex are not supported.
T2 : tensor(float16), tensor(float), tensor(double), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(bool)
Constrain output types. Strings and complex are not supported.

#### Examples

populate_off_main_diagonal
shape = (4, 5)
off_diagonal_offset = 1
node = onnx.helper.make_node(
'EyeLike',
inputs=['x'],
outputs=['y'],
k=off_diagonal_offset,
dtype=onnx.TensorProto.FLOAT,
)

x = np.random.randint(0, 100, size=shape, dtype=np.int32)
y = np.eye(shape[0], shape[1], k=off_diagonal_offset, dtype=np.float32)
expect(node, inputs=[x], outputs=[y], name='test_eyelike_populate_off_main_diagonal')
with_dtype
shape = (3, 4)
node = onnx.helper.make_node(
'EyeLike',
inputs=['x'],
outputs=['y'],
dtype=onnx.TensorProto.DOUBLE,
)

x = np.random.randint(0, 100, size=shape, dtype=np.int32)
y = np.eye(shape[0], shape[1], dtype=np.float64)
expect(node, inputs=[x], outputs=[y], name='test_eyelike_with_dtype')
without_dtype
shape = (4, 4)
node = onnx.helper.make_node(
'EyeLike',
inputs=['x'],
outputs=['y'],
)

x = np.random.randint(0, 100, size=shape, dtype=np.int32)
y = np.eye(shape[0], shape[1], dtype=np.int32)
expect(node, inputs=[x], outputs=[y], name='test_eyelike_without_dtype')

### Flatten

Flattens the input tensor into a 2D matrix. If input tensor has shape (d_0, d_1, ... d_n) then the output will have shape (d_0 X d_1 ... d_(axis-1), d_axis X d_(axis+1) ... X dn).

#### Version

This version of the operator has been available since version 9 of the default ONNX operator set.

Other versions of this operator: Flatten-1

#### Attributes

axis : int (default is 1)
Indicate up to which input dimensions (exclusive) should be flattened to the outer dimension of the output. The value for axis must be in the range [0, R], where R is the rank of the input tensor. When axis = 0, the shape of the output tensor is (1, (d_0 X d_1 ... d_n), where the shape of the input tensor is (d_0, d_1, ... d_n).

#### Inputs

input : T
A tensor of rank >= axis.

#### Outputs

output : T
A 2D tensor with the contents of the input tensor, with input dimensions up to axis flattened to the outer dimension of the output and remaining input dimensions flattened into the inner dimension of the output.

#### Type Constraints

T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double), tensor(string), tensor(bool), tensor(complex64), tensor(complex128)
Constrain input and output to all tensor types.

#### Examples

flatten
shape = (2, 3, 4, 5)
a = np.random.random_sample(shape).astype(np.float32)

for i in range(len(shape)):
node = onnx.helper.make_node(
'Flatten',
inputs=['a'],
outputs=['b'],
axis=i,
)

new_shape = (1, -1) if i == 0 else (np.prod(shape[0:i]).astype(int), -1)
b = np.reshape(a, new_shape)
expect(node, inputs=[a], outputs=[b],
name='test_flatten_axis' + str(i))
flatten_with_default_axis
node = onnx.helper.make_node(
'Flatten',
inputs=['a'],
outputs=['b'],  # Default value for axis: axis=1
)

shape = (5, 4, 3, 2)
a = np.random.random_sample(shape).astype(np.float32)
new_shape = (5, 24)
b = np.reshape(a, new_shape)
expect(node, inputs=[a], outputs=[b],
name='test_flatten_default_axis')

### Floor

Floor takes one input data (Tensor) and produces one output data (Tensor) where the floor is, y = floor(x), is applied to the tensor elementwise.

#### Version

This version of the operator has been available since version 6 of the default ONNX operator set.

Other versions of this operator: Floor-1

X : T
Input tensor

Y : T
Output tensor

#### Type Constraints

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

#### Examples

floor
node = onnx.helper.make_node(
'Floor',
inputs=['x'],
outputs=['y'],
)

x = np.array([-1.5, 1.2, 2]).astype(np.float32)
y = np.floor(x)  # expected output [-2., 1., 2.]
expect(node, inputs=[x], outputs=[y],
name='test_floor_example')

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.floor(x)
expect(node, inputs=[x], outputs=[y],
name='test_floor')

### GRU

Computes an one-layer GRU. This operator is usually supported via some custom implementation such as CuDNN.

Notations:

X - input tensor

z - update gate

r - reset gate

h - hidden gate

t - time step (t-1 means previous time step)

W[zrh] - W parameter weight matrix for update, reset, and hidden gates

R[zrh] - R recurrence weight matrix for update, reset, and hidden gates

Wb[zrh] - W bias vectors for update, reset, and hidden gates

Rb[zrh] - R bias vectors for update, reset, and hidden gates

WB[zrh] - W parameter weight matrix for backward update, reset, and hidden gates

RB[zrh] - R recurrence weight matrix for backward update, reset, and hidden gates

WBb[zrh] - W bias vectors for backward update, reset, and hidden gates

RBb[zrh] - R bias vectors for backward update, reset, and hidden gates

H - Hidden state

num_directions - 2 if direction == bidirectional else 1

Activation functions:

Relu(x)                - max(0, x)

Tanh(x)                - (1 - e^{-2x})/(1 + e^{-2x})

Sigmoid(x)             - 1/(1 + e^{-x})

(NOTE: Below are optional)

Affine(x)              - alpha*x + beta

LeakyRelu(x)           - x if x >= 0 else alpha * x

ThresholdedRelu(x)     - x if x >= alpha else 0

ScaledTanh(x)          - alpha*Tanh(beta*x)

HardSigmoid(x)         - min(max(alpha*x + beta, 0), 1)

Elu(x)                 - x if x >= 0 else alpha*(e^x - 1)

Softsign(x)            - x/(1 + |x|)

Softplus(x)            - log(1 + e^x)


Equations (Default: f=Sigmoid, g=Tanh):

- zt = f(Xt*(Wz^T) + Ht-1*(Rz^T) + Wbz + Rbz)

- rt = f(Xt*(Wr^T) + Ht-1*(Rr^T) + Wbr + Rbr)

- ht = g(Xt*(Wh^T) + (rt (.) Ht-1)*(Rh^T) + Rbh + Wbh) # default, when linear_before_reset = 0

- ht = g(Xt*(Wh^T) + (rt (.) (Ht-1*(Rh^T) + Rbh)) + Wbh) # when linear_before_reset != 0

- Ht = (1 - zt) (.) ht + zt (.) Ht-1


This operator has optional inputs/outputs. See the doc for more details about the representation of optional arguments. An empty string may be used in the place of an actual argument's name to indicate a missing argument. Trailing optional arguments (those not followed by an argument that is present) may also be simply omitted.

#### Version

This version of the operator has been available since version 7 of the default ONNX operator set.

Other versions of this operator: GRU-1, GRU-3

#### Attributes

activation_alpha : list of floats
Optional scaling values used by some activation functions. The values are consumed in the order of activation functions, for example (f, g, h) in LSTM. Default values are the same as of corresponding ONNX operators.For example with LeakyRelu, the default alpha is 0.01.
activation_beta : list of floats
Optional scaling values used by some activation functions. The values are consumed in the order of activation functions, for example (f, g, h) in LSTM. Default values are the same as of corresponding ONNX operators.
activations : list of strings
A list of 2 (or 4 if bidirectional) activation functions for update, reset, and hidden gates. The activation functions must be one of the activation functions specified above. Optional: See the equations for default if not specified.
clip : float
Cell clip threshold. Clipping bounds the elements of a tensor in the range of [-threshold, +threshold] and is applied to the input of activations. No clip if not specified.
direction : string (default is forward)
Specify if the RNN is forward, reverse, or bidirectional. Must be one of forward (default), reverse, or bidirectional.
hidden_size : int
Number of neurons in the hidden layer
linear_before_reset : int (default is 0)
When computing the output of the hidden gate, apply the linear transformation before multiplying by the output of the reset gate.

#### Inputs (3 - 6)

X : T
The input sequences packed (and potentially padded) into one 3-D tensor with the shape of [seq_length, batch_size, input_size].
W : T
The weight tensor for the gates. Concatenation of W[zrh] and WB[zrh] (if bidirectional) along dimension 0. This tensor has shape [num_directions, 3*hidden_size, input_size].
R : T
The recurrence weight tensor. Concatenation of R[zrh] and RB[zrh] (if bidirectional) along dimension 0. This tensor has shape [num_directions, 3*hidden_size, hidden_size].
B (optional) : T
The bias tensor for the gates. Concatenation of [Wb[zrh], Rb[zrh]] and [WBb[zrh], RBb[zrh]] (if bidirectional) along dimension 0. This tensor has shape [num_directions, 6*hidden_size]. Optional: If not specified - assumed to be 0
sequence_lens (optional) : T1
Optional tensor specifying lengths of the sequences in a batch. If not specified - assumed all sequences in the batch to have length seq_length. It has shape [batch_size].
initial_h (optional) : T
Optional initial value of the hidden. If not specified - assumed to be 0. It has shape [num_directions, batch_size, hidden_size].

#### Outputs (0 - 2)

Y (optional) : T
A tensor that concats all the intermediate output values of the hidden. It has shape [seq_length, num_directions, batch_size, hidden_size].
Y_h (optional) : T
The last output value of the hidden. It has shape [num_directions, batch_size, hidden_size].

#### Type Constraints

T : tensor(float16), tensor(float), tensor(double)
Constrain input and output types to float tensors.
T1 : tensor(int32)
Constrain seq_lens to integer tensor.

#### Examples

defaults
input = np.array([[[1., 2.], [3., 4.], [5., 6.]]]).astype(np.float32)

input_size = 2
hidden_size = 5
weight_scale = 0.1
number_of_gates = 3

node = onnx.helper.make_node(
'GRU',
inputs=['X', 'W', 'R'],
outputs=['', 'Y'],
hidden_size=hidden_size
)

W = weight_scale * np.ones((1, number_of_gates * hidden_size, input_size)).astype(np.float32)
R = weight_scale * np.ones((1, number_of_gates * hidden_size, hidden_size)).astype(np.float32)

gru = GRU_Helper(X=input, W=W, R=R)
_, Y_h = gru.step()
expect(node, inputs=[input, W, R], outputs=[Y_h.astype(np.float32)], name='test_gru_defaults')
initial_bias
input = np.array([[[1., 2., 3.], [4., 5., 6.], [7., 8., 9.]]]).astype(np.float32)

input_size = 3
hidden_size = 3
weight_scale = 0.1
custom_bias = 0.1
number_of_gates = 3

node = onnx.helper.make_node(
'GRU',
inputs=['X', 'W', 'R', 'B'],
outputs=['', 'Y'],
hidden_size=hidden_size
)

W = weight_scale * np.ones((1, number_of_gates * hidden_size, input_size)).astype(np.float32)
R = weight_scale * np.ones((1, number_of_gates * hidden_size, hidden_size)).astype(np.float32)

# Adding custom bias
W_B = custom_bias * np.ones((1, number_of_gates * hidden_size)).astype(np.float32)
R_B = np.zeros((1, number_of_gates * hidden_size)).astype(np.float32)
B = np.concatenate((W_B, R_B), axis=1)

gru = GRU_Helper(X=input, W=W, R=R, B=B)
_, Y_h = gru.step()
expect(node, inputs=[input, W, R, B], outputs=[Y_h.astype(np.float32)], name='test_gru_with_initial_bias')
seq_length
input = np.array([[[1., 2., 3.], [4., 5., 6.], [7., 8., 9.]],
[[10., 11., 12.], [13., 14., 15.], [16., 17., 18.]]]).astype(np.float32)

input_size = 3
hidden_size = 5
number_of_gates = 3

node = onnx.helper.make_node(
'GRU',
inputs=['X', 'W', 'R', 'B'],
outputs=['', 'Y'],
hidden_size=hidden_size
)

W = np.random.randn(1, number_of_gates * hidden_size, input_size).astype(np.float32)
R = np.random.randn(1, number_of_gates * hidden_size, hidden_size).astype(np.float32)

# Adding custom bias
W_B = np.random.randn(1, number_of_gates * hidden_size).astype(np.float32)
R_B = np.random.randn(1, number_of_gates * hidden_size).astype(np.float32)
B = np.concatenate((W_B, R_B), axis=1)

gru = GRU_Helper(X=input, W=W, R=R, B=B)
_, Y_h = gru.step()
expect(node, inputs=[input, W, R, B], outputs=[Y_h.astype(np.float32)], name='test_gru_seq_length')

### Gather

Given data tensor of rank r >= 1, and indices tensor of rank q, gather entries of the axis dimension of data (by default outer-most one as axis=0) indexed by indices, and concatenates them in an output tensor of rank q + (r - 1). Example 1: data = [ [1.0, 1.2], [2.3, 3.4], [4.5, 5.7], ] indices = [ [0, 1], [1, 2], ] output = [ [ [1.0, 1.2], [2.3, 3.4], ], [ [2.3, 3.4], [4.5, 5.7], ], ] Example 2: data = [ [1.0, 1.2, 1.9], [2.3, 3.4, 3.9], [4.5, 5.7, 5.9], ] indices = [ [0, 2], ] axis = 1, output = [ [ [1.0, 1.9], [2.3, 3.9], [4.5, 5.9], ], ]

#### Version

This version of the operator has been available since version 1 of the default ONNX operator set.

#### Attributes

axis : int (default is 0)
Which axis to gather on. Negative value means counting dimensions from the back. Accepted range in [-r, r-1]

#### Inputs

data : T
Tensor of rank r >= 1.
indices : Tind
Tensor of int32/int64 indices, of any rank q.

#### Outputs

output : T
Tensor of rank q + (r - 1).

#### Type Constraints

T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double), tensor(string), tensor(bool), tensor(complex64), tensor(complex128)
Constrain input and output types to any tensor type.
Tind : tensor(int32), tensor(int64)
Constrain indices to integer types

#### Examples

gather_0
node = onnx.helper.make_node(
'Gather',
inputs=['data', 'indices'],
outputs=['y'],
axis=0,
)
data = np.random.randn(5, 4, 3, 2).astype(np.float32)
indices = np.array([0, 1, 3])
y = np.take(data, indices, axis=0)

expect(node, inputs=[data, indices.astype(np.int64)], outputs=[y],
name='test_gather_0')
gather_1
node = onnx.helper.make_node(
'Gather',
inputs=['data', 'indices'],
outputs=['y'],
axis=1,
)
data = np.random.randn(5, 4, 3, 2).astype(np.float32)
indices = np.array([0, 1, 3])
y = np.take(data, indices, axis=1)

expect(node, inputs=[data, indices.astype(np.int64)], outputs=[y],
name='test_gather_1')

### Gemm

General Matrix multiplication: https://en.wikipedia.org/wiki/Basic_Linear_Algebra_Subprograms#Level_3

A' = transpose(A) if transA else A

B' = transpose(B) if transB else B

Compute Y = alpha * A' * B' + beta * C, where input tensor A has shape (M, K) or (K, M), input tensor B has shape (K, N) or (N, K), input tensor C is broadcastable to shape (M, N), and output tensor Y has shape (M, N). A will be transposed before doing the computation if attribute transA is non-zero, same for B and transB. This operator supports unidirectional broadcasting (tensor C should be unidirectional broadcastable to tensor A * B); for more details please check the doc.

#### Version

This version of the operator has been available since version 9 of the default ONNX operator set.

Other versions of this operator: Gemm-1, Gemm-6, Gemm-7

#### Attributes

alpha : float (default is 1.0)
Scalar multiplier for the product of input tensors A * B.
beta : float (default is 1.0)
Scalar multiplier for input tensor C.
transA : int (default is 0)
Whether A should be transposed
transB : int (default is 0)
Whether B should be transposed

#### Inputs

A : T
Input tensor A. The shape of A should be (M, K) if transA is 0, or (K, M) if transA is non-zero.
B : T
Input tensor B. The shape of B should be (K, N) if transB is 0, or (N, K) if transB is non-zero.
C : T
Input tensor C. The shape of C should be unidirectional broadcastable to (M, N).

#### Outputs

Y : T
Output tensor of shape (M, N).

#### Type Constraints

T : tensor(float16), tensor(float), tensor(double), tensor(uint32), tensor(uint64), tensor(int32), tensor(int64)
Constrain input and output types to float/int tensors.

#### Examples

notranspose
node = onnx.helper.make_node(
'Gemm',
inputs=['a', 'b', 'c'],
outputs=['y'],
alpha=0.5,
beta=0.5
)
a = np.random.ranf([3, 6]).astype(np.float32)
b = np.random.ranf([6, 4]).astype(np.float32)
c = np.random.ranf([3, 4]).astype(np.float32)
y = 0.5 * np.dot(a, b) + 0.5 * c
expect(node, inputs=[a, b, c], outputs=[y],
name='test_gemm_nobroadcast')
transpose
node = onnx.helper.make_node(
'Gemm',
inputs=['a', 'b', 'c'],
outputs=['y'],
alpha=0.5,
beta=0.5,
transA=1,
transB=1
)
a = np.random.ranf([6, 3]).astype(np.float32)
b = np.random.ranf([4, 6]).astype(np.float32)
c = np.random.ranf([1, 1]).astype(np.float32)
y = 0.5 * np.dot(a.T, b.T) + 0.5 * c
expect(node, inputs=[a, b, c], outputs=[y],
name='test_gemm_broadcast')

### GlobalAveragePool

GlobalAveragePool consumes an input tensor X and applies average pooling across the values in the same channel. This is equivalent to AveragePool with kernel size equal to the spatial dimension of input tensor.

#### Version

This version of the operator has been available since version 1 of the default ONNX operator set.

#### Inputs

X : 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.

#### Outputs

Y : T
Output data tensor from pooling across the input tensor. Dimensions will be N x C x 1 x 1

#### Type Constraints

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

#### Examples

globalaveragepool
node = onnx.helper.make_node(
'GlobalAveragePool',
inputs=['x'],
outputs=['y'],
)
x = np.random.randn(1, 3, 5, 5).astype(np.float32)
spatial_shape = np.ndim(x) - 2
y = np.average(x, axis=tuple(range(spatial_shape, spatial_shape + 2)))
for _ in range(spatial_shape):
y = np.expand_dims(y, -1)
expect(node, inputs=[x], outputs=[y], name='test_globalaveragepool')
globalaveragepool_precomputed
node = onnx.helper.make_node(
'GlobalAveragePool',
inputs=['x'],
outputs=['y'],
)
x = np.array([[[
[1, 2, 3],
[4, 5, 6],
[7, 8, 9],
]]]).astype(np.float32)
y = np.array([[[[5]]]]).astype(np.float32)
expect(node, inputs=[x], outputs=[y], name='test_globalaveragepool_precomputed')

### GlobalLpPool

GlobalLpPool consumes an input tensor X and applies lp pool pooling across the values in the same channel. This is equivalent to LpPool with kernel size equal to the spatial dimension of input tensor.

#### Version

This version of the operator has been available since version 2 of the default ONNX operator set.

Other versions of this operator: GlobalLpPool-1

#### Attributes

p : int (default is 2)
p value of the Lp norm used to pool over the input data.

#### Inputs

X : 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.

#### Outputs

Y : T
Output data tensor from pooling across the input tensor. Dimensions will be N x C x 1 x 1

#### Type Constraints

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

### GlobalMaxPool

GlobalMaxPool consumes an input tensor X and applies max pooling across the values in the same channel. This is equivalent to MaxPool with kernel size equal to the spatial dimension of input tensor.

#### Version

This version of the operator has been available since version 1 of the default ONNX operator set.

#### Inputs

X : 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.

#### Outputs

Y : T
Output data tensor from pooling across the input tensor. Dimensions will be N x C x 1 x 1

#### Type Constraints

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

#### Examples

globalmaxpool
node = onnx.helper.make_node(
'GlobalMaxPool',
inputs=['x'],
outputs=['y'],
)
x = np.random.randn(1, 3, 5, 5).astype(np.float32)
spatial_shape = np.ndim(x) - 2
y = np.max(x, axis=tuple(range(spatial_shape, spatial_shape + 2)))
for _ in range(spatial_shape):
y = np.expand_dims(y, -1)
expect(node, inputs=[x], outputs=[y], name='test_globalmaxpool')
globalmaxpool_precomputed
node = onnx.helper.make_node(
'GlobalMaxPool',
inputs=['x'],
outputs=['y'],
)
x = np.array([[[
[1, 2, 3],
[4, 5, 6],
[7, 8, 9],
]]]).astype(np.float32)
y = np.array([[[[9]]]]).astype(np.float32)
expect(node, inputs=[x], outputs=[y], name='test_globalmaxpool_precomputed')

### Greater

Returns the tensor resulted from performing the greater logical operation elementwise on the input tensors A and B (with Numpy-style broadcasting support).

This operator supports multidirectional (i.e., Numpy-style) broadcasting; for more details please check the doc.

#### Version

This version of the operator has been available since version 9 of the default ONNX operator set.

Other versions of this operator: Greater-1, Greater-7

#### Inputs

A : T
First input operand for the logical operator.
B : T
Second input operand for the logical operator.

C : T1
Result tensor.

#### Type Constraints

T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double)
Constrains input to float tensors.
T1 : tensor(bool)
Constrains output to boolean tensor.

#### Examples

greater
node = onnx.helper.make_node(
'Greater',
inputs=['x', 'y'],
outputs=['greater'],
)

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.random.randn(3, 4, 5).astype(np.float32)
z = np.greater(x, y)
expect(node, inputs=[x, y], outputs=[z],
name='test_greater')
node = onnx.helper.make_node(
'Greater',
inputs=['x', 'y'],
outputs=['greater'],
)

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.random.randn(5).astype(np.float32)
z = np.greater(x, y)
expect(node, inputs=[x, y], outputs=[z],
name='test_greater_bcast')

### HardSigmoid

HardSigmoid takes one input data (Tensor) and produces one output data (Tensor) where the HardSigmoid function, y = max(0, min(1, alpha * x + beta)), is applied to the tensor elementwise.

#### Version

This version of the operator has been available since version 6 of the default ONNX operator set.

Other versions of this operator: HardSigmoid-1

#### Attributes

alpha : float (default is 0.2)
Value of alpha.
beta : float (default is 0.5)
Value of beta.

X : T
Input tensor

Y : T
Output tensor

#### Type Constraints

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

#### Examples

hardsigmoid
node = onnx.helper.make_node(
'HardSigmoid',
inputs=['x'],
outputs=['y'],
alpha=0.5,
beta=0.6
)

x = np.array([-1, 0, 1]).astype(np.float32)
y = np.clip(x * 0.5 + 0.6, 0, 1)  # expected output [0.1, 0.6, 1.]
expect(node, inputs=[x], outputs=[y],
name='test_hardsigmoid_example')

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.clip(x * 0.5 + 0.6, 0, 1)
expect(node, inputs=[x], outputs=[y],
name='test_hardsigmoid')
hardsigmoid_default
default_alpha = 0.2
default_beta = 0.5
node = onnx.helper.make_node(
'HardSigmoid',
inputs=['x'],
outputs=['y'],
)
x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.clip(x * default_alpha + default_beta, 0, 1)
expect(node, inputs=[x], outputs=[y],
name='test_hardsigmoid_default')

### Hardmax

The operator computes the hardmax (1 for the first maximum value, and 0 for all others) values for each layer in the batch of the given input. The input is a 2-D tensor (Tensor) of size (batch_size x input_feature_dimensions). The output tensor has the same shape and contains the hardmax values of the corresponding input.

Input does not need to explicitly be a 2D vector; rather, it will be coerced into one. For an arbitrary n-dimensional tensor input \in [a_0, a_1, ..., a_{k-1}, a_k, ..., a_{n-1}] and k is the axis provided, then input will be coerced into a 2-dimensional tensor with dimensions [a_0 * ... * a_{k-1}, a_k * ... * a_{n-1}]. For the default case where axis=1, this means the input tensor will be coerced into a 2D tensor of dimensions [a_0, a_1 * ... * a_{n-1}], where a_0 is often the batch size. In this situation, we must have a_0 = N and a_1 * ... * a_{n-1} = D. Each of these dimensions must be matched correctly, or else the operator will throw errors.

#### Version

This version of the operator has been available since version 1 of the default ONNX operator set.

#### Attributes

axis : int (default is 1)
Describes the axis of the inputs when coerced to 2D; defaults to one because the 0th axis most likely describes the batch_size

#### Inputs

input : T
The input tensor that's coerced into a 2D matrix of size (NxD) as described above.

#### Outputs

output : T
The output values with the same shape as input tensor.

#### Type Constraints

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

#### Examples

hardmax
node = onnx.helper.make_node(
'Hardmax',
inputs=['x'],
outputs=['y'],
)

x = np.array([[3, 0, 1, 2], [2, 5, 1, 0], [0, 1, 3, 2], [0, 1, 2, 3]]).astype(np.float32)
y = np.array([[1, 0, 0, 0], [0, 1, 0, 0], [0, 0, 1, 0], [0, 0, 0, 1]]).astype(np.float32)
expect(node, inputs=[x], outputs=[y],
name='test_hardmax_example')

# For multiple occurrances of the maximal values, the first occurrence is selected for one-hot output
x = np.array([[3, 3, 3, 1]]).astype(np.float32)
y = np.array([[1, 0, 0, 0]]).astype(np.float32)
expect(node, inputs=[x], outputs=[y],
name='test_hardmax_one_hot')
hardmax_axis
def hardmax_2d(x):  # type: (np.ndarray) -> np.ndarray
return np.eye(x.shape[1], dtype=x.dtype)[np.argmax(x, axis=1)]

x = np.random.randn(3, 4, 5).astype(np.float32)
node = onnx.helper.make_node(
'Hardmax',
inputs=['x'],
outputs=['y'],
axis=0,
)
y = hardmax_2d(x.reshape(1, 60)).reshape(3, 4, 5)
expect(node, inputs=[x], outputs=[y],
name='test_hardmax_axis_0')

node = onnx.helper.make_node(
'Hardmax',
inputs=['x'],
outputs=['y'],
axis=1,
)
y = hardmax_2d(x.reshape(3, 20)).reshape(3, 4, 5)
expect(node, inputs=[x], outputs=[y],
name='test_hardmax_axis_1')

# default axis is 1
node = onnx.helper.make_node(
'Hardmax',
inputs=['x'],
outputs=['y'],
)
expect(node, inputs=[x], outputs=[y],
name='test_hardmax_default_axis')

node = onnx.helper.make_node(
'Hardmax',
inputs=['x'],
outputs=['y'],
axis=2,
)
y = hardmax_2d(x.reshape(12, 5)).reshape(3, 4, 5)
expect(node, inputs=[x], outputs=[y],
name='test_hardmax_axis_2')

### Identity

Identity operator

#### Version

This version of the operator has been available since version 1 of the default ONNX operator set.

input : T
Input tensor

#### Outputs

output : T
Tensor to copy input into.

#### Type Constraints

T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double), tensor(string), tensor(bool), tensor(complex64), tensor(complex128)
Constrain input and output types to all tensor types.

#### Examples

identity
node = onnx.helper.make_node(
'Identity',
inputs=['x'],
outputs=['y'],
)

data = np.array([[[
[1, 2],
[3, 4],
]]], dtype=np.float32)

expect(node, inputs=[data], outputs=[data],
name='test_identity')

### If

If conditional

#### Version

This version of the operator has been available since version 1 of the default ONNX operator set.

#### Attributes

else_branch : graph (required)
Graph to run if condition is false. Has N outputs: values you wish to be live-out to the enclosing scope. The number of outputs must match the number of outputs in the then_branch.
then_branch : graph (required)
Graph to run if condition is true. Has N outputs: values you wish to be live-out to the enclosing scope. The number of outputs must match the number of outputs in the else_branch.

#### Inputs

cond : B
Condition for the if

#### Outputs (1 - ∞)

outputs (variadic) : V
Values that are live-out to the enclosing scope. The return values in the then_branch and else_branch must be of the same shape and same data type.

#### Type Constraints

V : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double), tensor(string), tensor(bool), tensor(complex64), tensor(complex128)
All Tensor types
B : tensor(bool)
Only bool

### InstanceNormalization

Carries out instance normalization as described in the paper https://arxiv.org/abs/1607.08022.

y = scale * (x - mean) / sqrt(variance + epsilon) + B, where mean and variance are computed per instance per channel.

#### Version

This version of the operator has been available since version 6 of the default ONNX operator set.

Other versions of this operator: InstanceNormalization-1

#### Attributes

epsilon : float (default is 1e-05)
The epsilon value to use to avoid division by zero.

#### Inputs

input : 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.
scale : T
The input 1-dimensional scale tensor of size C.
B : T
The input 1-dimensional bias tensor of size C.

#### Outputs

output : T
The output tensor of the same shape as input.

#### Type Constraints

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

#### Examples

instancenormalization
def _instancenorm_test_mode(x, s, bias, epsilon=1e-5):  # type: ignore
dims_x = len(x.shape)
axis = tuple(range(2, dims_x))
mean = np.mean(x, axis=axis, keepdims=True)
var = np.var(x, axis=axis, keepdims=True)
dim_ones = (1,) * (dims_x - 2)
s = s.reshape(-1, *dim_ones)
bias = bias.reshape(-1, *dim_ones)
return s * (x - mean) / np.sqrt(var + epsilon) + bias

# input size: (1, 2, 1, 3)
x = np.array([[[[-1, 0, 1]], [[2, 3, 4]]]]).astype(np.float32)
s = np.array([1.0, 1.5]).astype(np.float32)
bias = np.array([0, 1]).astype(np.float32)
y = _instancenorm_test_mode(x, s, bias).astype(np.float32)

node = onnx.helper.make_node(
'InstanceNormalization',
inputs=['x', 's', 'bias'],
outputs=['y'],
)

# output size: (1, 2, 1, 3)
expect(node, inputs=[x, s, bias], outputs=[y],
name='test_instancenorm_example')

# input size: (2, 3, 4, 5)
x = np.random.randn(2, 3, 4, 5).astype(np.float32)
s = np.random.randn(3).astype(np.float32)
bias = np.random.randn(3).astype(np.float32)
epsilon = 1e-2
y = _instancenorm_test_mode(x, s, bias, epsilon).astype(np.float32)

node = onnx.helper.make_node(
'InstanceNormalization',
inputs=['x', 's', 'bias'],
outputs=['y'],
epsilon=epsilon,
)

# output size: (2, 3, 4, 5)
expect(node, inputs=[x, s, bias], outputs=[y],
name='test_instancenorm_epsilon')

### LRN

Local Response Normalization proposed in the AlexNet paper. It normalizes over local input regions. The local region is defined across the channels. For an element X[n, c, d1, ..., dk] in a tensor of shape (N x C x D1 x D2, ..., Dk), its region is {X[n, i, d1, ..., dk] | max(0, c - floor((size - 1) / 2)) <= i <= min(C - 1, c + ceil((size - 1) / 2))}.

square_sum[n, c, d1, ..., dk] = sum(X[n, i, d1, ..., dk] ^ 2), where max(0, c - floor((size - 1) / 2)) <= i <= min(C - 1, c + ceil((size - 1) / 2)).

Y[n, c, d1, ..., dk] = X[n, c, d1, ..., dk] / (bias + alpha / size * square_sum[n, c, d1, ..., dk] ) ^ beta

#### Version

This version of the operator has been available since version 1 of the default ONNX operator set.

#### Attributes

alpha : float (default is 0.0001)
Scaling parameter.
beta : float (default is 0.75)
The exponent.
bias : float (default is 1.0)
size : int (required)
The number of channels to sum over

#### Inputs

X : 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 : T
Output tensor, which has the shape and type as input tensor

#### Type Constraints

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

#### Examples

default
alpha = 0.0001
beta = 0.75
bias = 1.0
nsize = 3
node = onnx.helper.make_node(
'LRN',
inputs=['x'],
outputs=['y'],
size=3
)
x = np.random.randn(5, 5, 5, 5).astype(np.float32)
square_sum = np.zeros((5, 5, 5, 5)).astype(np.float32)
for n, c, h, w in np.ndindex(x.shape):
square_sum[n, c, h, w] = sum(x[n,
max(0, c - int(math.floor((nsize - 1) / 2))):min(5, c + int(math.ceil((nsize - 1) / 2)) + 1),
h,
w] ** 2)
y = x / ((bias + (alpha / nsize) * square_sum) ** beta)
expect(node, inputs=[x], outputs=[y],
name='test_lrn_default')
lrn
alpha = 0.0002
beta = 0.5
bias = 2.0
nsize = 3
node = onnx.helper.make_node(
'LRN',
inputs=['x'],
outputs=['y'],
alpha=alpha,
beta=beta,
bias=bias,
size=nsize
)
x = np.random.randn(5, 5, 5, 5).astype(np.float32)
square_sum = np.zeros((5, 5, 5, 5)).astype(np.float32)
for n, c, h, w in np.ndindex(x.shape):
square_sum[n, c, h, w] = sum(x[n,
max(0, c - int(math.floor((nsize - 1) / 2))):min(5, c + int(math.ceil((nsize - 1) / 2)) + 1),
h,
w] ** 2)
y = x / ((bias + (alpha / nsize) * square_sum) ** beta)
expect(node, inputs=[x], outputs=[y],
name='test_lrn')

### LSTM

Computes an one-layer LSTM. This operator is usually supported via some custom implementation such as CuDNN.

Notations:

X - input tensor

i - input gate

o - output gate

f - forget gate

c - cell gate

t - time step (t-1 means previous time step)

W[iofc] - W parameter weight matrix for input, output, forget, and cell gates

R[iofc] - R recurrence weight matrix for input, output, forget, and cell gates

Wb[iofc] - W bias vectors for input, output, forget, and cell gates

Rb[iofc] - R bias vectors for input, output, forget, and cell gates

P[iof] - P peephole weight vector for input, output, and forget gates

WB[iofc] - W parameter weight matrix for backward input, output, forget, and cell gates

RB[iofc] - R recurrence weight matrix for backward input, output, forget, and cell gates

WBb[iofc] - W bias vectors for backward input, output, forget, and cell gates

RBb[iofc] - R bias vectors for backward input, output, forget, and cell gates

PB[iof] - P peephole weight vector for backward input, output, and forget gates

H - Hidden state

num_directions - 2 if direction == bidirectional else 1

Activation functions:

Relu(x)                - max(0, x)

Tanh(x)                - (1 - e^{-2x})/(1 + e^{-2x})

Sigmoid(x)             - 1/(1 + e^{-x})

(NOTE: Below are optional)

Affine(x)              - alpha*x + beta

LeakyRelu(x)           - x if x >= 0 else alpha * x

ThresholdedRelu(x)     - x if x >= alpha else 0

ScaledTanh(x)          - alpha*Tanh(beta*x)

HardSigmoid(x)         - min(max(alpha*x + beta, 0), 1)

Elu(x)                 - x if x >= 0 else alpha*(e^x - 1)

Softsign(x)            - x/(1 + |x|)

Softplus(x)            - log(1 + e^x)


Equations (Default: f=Sigmoid, g=Tanh, h=Tanh):

- it = f(Xt*(Wi^T) + Ht-1*(Ri^T) + Pi (.) Ct-1 + Wbi + Rbi)

- ft = f(Xt*(Wf^T) + Ht-1*(Rf^T) + Pf (.) Ct-1 + Wbf + Rbf)

- ct = g(Xt*(Wc^T) + Ht-1*(Rc^T) + Wbc + Rbc)

- Ct = ft (.) Ct-1 + it (.) ct

- ot = f(Xt*(Wo^T) + Ht-1*(Ro^T) + Po (.) Ct + Wbo + Rbo)

- Ht = ot (.) h(Ct)


This operator has optional inputs/outputs. See the doc for more details about the representation of optional arguments. An empty string may be used in the place of an actual argument's name to indicate a missing argument. Trailing optional arguments (those not followed by an argument that is present) may also be simply omitted.

#### Version

This version of the operator has been available since version 7 of the default ONNX operator set.

Other versions of this operator: LSTM-1

#### Attributes

activation_alpha : list of floats
Optional scaling values used by some activation functions. The values are consumed in the order of activation functions, for example (f, g, h) in LSTM. Default values are the same as of corresponding ONNX operators.For example with LeakyRelu, the default alpha is 0.01.
activation_beta : list of floats
Optional scaling values used by some activation functions. The values are consumed in the order of activation functions, for example (f, g, h) in LSTM. Default values are the same as of corresponding ONNX operators.
activations : list of strings
A list of 3 (or 6 if bidirectional) activation functions for input, output, forget, cell, and hidden. The activation functions must be one of the activation functions specified above. Optional: See the equations for default if not specified.
clip : float
Cell clip threshold. Clipping bounds the elements of a tensor in the range of [-threshold, +threshold] and is applied to the input of activations. No clip if not specified.
direction : string (default is forward)
Specify if the RNN is forward, reverse, or bidirectional. Must be one of forward (default), reverse, or bidirectional.
hidden_size : int
Number of neurons in the hidden layer
input_forget : int (default is 0)
Couple the input and forget gates if 1.

#### Inputs (3 - 8)

X : T
The input sequences packed (and potentially padded) into one 3-D tensor with the shape of [seq_length, batch_size, input_size].
W : T
The weight tensor for the gates. Concatenation of W[iofc] and WB[iofc] (if bidirectional) along dimension 0. The tensor has shape [num_directions, 4*hidden_size, input_size].
R : T
The recurrence weight tensor. Concatenation of R[iofc] and RB[iofc] (if bidirectional) along dimension 0. This tensor has shape [num_directions, 4*hidden_size, hidden_size].
B (optional) : T
The bias tensor for input gate. Concatenation of [Wb[iofc], Rb[iofc]], and [WBb[iofc], RBb[iofc]] (if bidirectional) along dimension 0. This tensor has shape [num_directions, 8*hidden_size]. Optional: If not specified - assumed to be 0.
sequence_lens (optional) : T1
Optional tensor specifying lengths of the sequences in a batch. If not specified - assumed all sequences in the batch to have length seq_length. It has shape [batch_size].
initial_h (optional) : T
Optional initial value of the hidden. If not specified - assumed to be 0. It has shape [num_directions, batch_size, hidden_size].
initial_c (optional) : T
Optional initial value of the cell. If not specified - assumed to be 0. It has shape [num_directions, batch_size, hidden_size].
P (optional) : T
The weight tensor for peepholes. Concatenation of P[iof] and PB[iof] (if bidirectional) along dimension 0. It has shape [num_directions, 3*hidde_size]. Optional: If not specified - assumed to be 0.

#### Outputs (0 - 3)

Y (optional) : T
A tensor that concats all the intermediate output values of the hidden. It has shape [seq_length, num_directions, batch_size, hidden_size].
Y_h (optional) : T
The last output value of the hidden. It has shape [num_directions, batch_size, hidden_size].
Y_c (optional) : T
The last output value of the cell. It has shape [num_directions, batch_size, hidden_size].

#### Type Constraints

T : tensor(float16), tensor(float), tensor(double)
Constrain input and output types to float tensors.
T1 : tensor(int32)
Constrain seq_lens to integer tensor.

#### Examples

defaults
input = np.array([[[1., 2.], [3., 4.], [5., 6.]]]).astype(np.float32)

input_size = 2
hidden_size = 3
weight_scale = 0.1
number_of_gates = 4

node = onnx.helper.make_node(
'LSTM',
inputs=['X', 'W', 'R'],
outputs=['', 'Y'],
hidden_size=hidden_size
)

W = weight_scale * np.ones((1, number_of_gates * hidden_size, input_size)).astype(np.float32)
R = weight_scale * np.ones((1, number_of_gates * hidden_size, hidden_size)).astype(np.float32)

lstm = LSTM_Helper(X=input, W=W, R=R)
_, Y_h = lstm.step()
expect(node, inputs=[input, W, R], outputs=[Y_h.astype(np.float32)], name='test_lstm_defaults')
initial_bias
input = np.array([[[1., 2., 3.], [4., 5., 6.], [7., 8., 9.]]]).astype(np.float32)

input_size = 3
hidden_size = 4
weight_scale = 0.1
custom_bias = 0.1
number_of_gates = 4

node = onnx.helper.make_node(
'LSTM',
inputs=['X', 'W', 'R', 'B'],
outputs=['', 'Y'],
hidden_size=hidden_size
)

W = weight_scale * np.ones((1, number_of_gates * hidden_size, input_size)).astype(np.float32)
R = weight_scale * np.ones((1, number_of_gates * hidden_size, hidden_size)).astype(np.float32)

# Adding custom bias
W_B = custom_bias * np.ones((1, number_of_gates * hidden_size)).astype(np.float32)
R_B = np.zeros((1, number_of_gates * hidden_size)).astype(np.float32)
B = np.concatenate((W_B, R_B), 1)

lstm = LSTM_Helper(X=input, W=W, R=R, B=B)
_, Y_h = lstm.step()
expect(node, inputs=[input, W, R, B], outputs=[Y_h.astype(np.float32)], name='test_lstm_with_initial_bias')
peepholes
input = np.array([[[1., 2., 3., 4.], [5., 6., 7., 8.]]]).astype(np.float32)

input_size = 4
hidden_size = 3
weight_scale = 0.1
number_of_gates = 4
number_of_peepholes = 3

node = onnx.helper.make_node(
'LSTM',
inputs=['X', 'W', 'R', 'B', 'sequence_lens', 'initial_h', 'initial_c', 'P'],
outputs=['', 'Y'],
hidden_size=hidden_size
)

# Initializing Inputs
W = weight_scale * np.ones((1, number_of_gates * hidden_size, input_size)).astype(np.float32)
R = weight_scale * np.ones((1, number_of_gates * hidden_size, hidden_size)).astype(np.float32)
B = np.zeros((1, 2 * number_of_gates * hidden_size)).astype(np.float32)
seq_lens = np.repeat(input.shape[0], input.shape[1]).astype(np.int32)
init_h = np.zeros((1, input.shape[1], hidden_size)).astype(np.float32)
init_c = np.zeros((1, input.shape[1], hidden_size)).astype(np.float32)
P = weight_scale * np.ones((1, number_of_peepholes * hidden_size)).astype(np.float32)

lstm = LSTM_Helper(X=input, W=W, R=R, B=B, P=P, initial_c=init_c, initial_h=init_h)
_, Y_h = lstm.step()
expect(node, inputs=[input, W, R, B, seq_lens, init_h, init_c, P], outputs=[Y_h.astype(np.float32)],
name='test_lstm_with_peepholes')

### LeakyRelu

LeakyRelu takes input data (Tensor) and an argument alpha, and produces one output data (Tensor) where the function f(x) = alpha * x for x < 0, f(x) = x for x >= 0, is applied to the data tensor elementwise.

#### Version

This version of the operator has been available since version 6 of the default ONNX operator set.

Other versions of this operator: LeakyRelu-1

#### Attributes

alpha : float (default is 0.01)
Coefficient of leakage.

X : T
Input tensor

Y : T
Output tensor

#### Type Constraints

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

#### Examples

leakyrelu
node = onnx.helper.make_node(
'LeakyRelu',
inputs=['x'],
outputs=['y'],
alpha=0.1
)

x = np.array([-1, 0, 1]).astype(np.float32)
# expected output [-0.1, 0., 1.]
y = np.clip(x, 0, np.inf) + np.clip(x, -np.inf, 0) * 0.1
expect(node, inputs=[x], outputs=[y],
name='test_leakyrelu_example')

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.clip(x, 0, np.inf) + np.clip(x, -np.inf, 0) * 0.1
expect(node, inputs=[x], outputs=[y],
name='test_leakyrelu')
leakyrelu_default
default_alpha = 0.01
node = onnx.helper.make_node(
'LeakyRelu',
inputs=['x'],
outputs=['y'],
)
x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.clip(x, 0, np.inf) + np.clip(x, -np.inf, 0) * default_alpha
expect(node, inputs=[x], outputs=[y],
name='test_leakyrelu_default')

### Less

Returns the tensor resulted from performing the less logical operation elementwise on the input tensors A and B (with Numpy-style broadcasting support).

This operator supports multidirectional (i.e., Numpy-style) broadcasting; for more details please check the doc.

#### Version

This version of the operator has been available since version 9 of the default ONNX operator set.

Other versions of this operator: Less-1, Less-7

#### Inputs

A : T
First input operand for the logical operator.
B : T
Second input operand for the logical operator.

C : T1
Result tensor.

#### Type Constraints

T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double)
Constrains input to float tensors.
T1 : tensor(bool)
Constrains output to boolean tensor.

#### Examples

less
node = onnx.helper.make_node(
'Less',
inputs=['x', 'y'],
outputs=['less'],
)

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.random.randn(3, 4, 5).astype(np.float32)
z = np.less(x, y)
expect(node, inputs=[x, y], outputs=[z],
name='test_less')
node = onnx.helper.make_node(
'Less',
inputs=['x', 'y'],
outputs=['less'],
)

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.random.randn(5).astype(np.float32)
z = np.less(x, y)
expect(node, inputs=[x, y], outputs=[z],
name='test_less_bcast')

### Log

Calculates the natural log of the given input tensor, element-wise.

#### Version

This version of the operator has been available since version 6 of the default ONNX operator set.

Other versions of this operator: Log-1

input : T
Input tensor

#### Outputs

output : T
The natural log of the input tensor computed element-wise

#### Type Constraints

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

#### Examples

log
node = onnx.helper.make_node(
'Log',
inputs=['x'],
outputs=['y'],
)

x = np.array([1, 10]).astype(np.float32)
y = np.log(x)  # expected output [0., 2.30258512]
expect(node, inputs=[x], outputs=[y],
name='test_log_example')

x = np.exp(np.random.randn(3, 4, 5).astype(np.float32))
y = np.log(x)
expect(node, inputs=[x], outputs=[y],
name='test_log')

### LogSoftmax

The operator computes the logsoftmax (log of softmax) values for each layer in the batch of the given input. The input is a 2-D tensor (Tensor) of size (batch_size x input_feature_dimensions). The output tensor has the same shape and contains the logsoftmax values of the corresponding input.

Input does not need to explicitly be a 2D vector; rather, it will be coerced into one. For an arbitrary n-dimensional tensor input \in [a_0, a_1, ..., a_{k-1}, a_k, ..., a_{n-1}] and k is the axis provided, then input will be coerced into a 2-dimensional tensor with dimensions [a_0 * ... * a_{k-1}, a_k * ... * a_{n-1}]. For the default case where axis=1, this means the input tensor will be coerced into a 2D tensor of dimensions [a_0, a_1 * ... * a_{n-1}], where a_0 is often the batch size. In this situation, we must have a_0 = N and a_1 * ... * a_{n-1} = D. Each of these dimensions must be matched correctly, or else the operator will throw errors.

#### Version

This version of the operator has been available since version 1 of the default ONNX operator set.

#### Attributes

axis : int (default is 1)
Describes the axis of the inputs when coerced to 2D; defaults to one because the 0th axis most likely describes the batch_size

#### Inputs

input : T
The input tensor that's coerced into a 2D matrix of size (NxD) as described above.

#### Outputs

output : T
The output values with the same shape as input tensor.

#### Type Constraints

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

#### Examples

logsoftmax
node = onnx.helper.make_node(
'LogSoftmax',
inputs=['x'],
outputs=['y'],
)
x = np.array([[-1, 0, 1]]).astype(np.float32)
# expected output [[-2.40760589, -1.40760589, -0.40760589]]
y = x - np.log(np.sum(np.exp(x), axis=1))
expect(node, inputs=[x], outputs=[y],
name='test_logsoftmax_example_1')
logsoftmax_axis
def logsoftmax_2d(x):  # type: (np.ndarray) -> np.ndarray
max_x = np.max(x, axis=1).reshape((-1, 1))
exp_x = np.exp(x - max_x)
return x - max_x - np.log(np.sum(exp_x, axis=1).reshape((-1, 1)))

x = np.array([[0, 1, 2, 3], [10000, 10001, 10002, 10003]]).astype(np.float32)
# expected output [[-3.4401896, -2.4401896, -1.44018972, -0.44018969],
#                 [-3.4401896, -2.4401896, -1.44018972, -0.44018969]]
y = logsoftmax_2d(x)

node = onnx.helper.make_node(
'LogSoftmax',
inputs=['x'],
outputs=['y'],
)
expect(node, inputs=[x], outputs=[y],
name='test_logsoftmax_large_number')

x = np.abs(np.random.randn(3, 4, 5).astype(np.float32))
node = onnx.helper.make_node(
'LogSoftmax',
inputs=['x'],
outputs=['y'],
axis=0,
)
y = logsoftmax_2d(x.reshape(1, 60)).reshape(3, 4, 5)
expect(node, inputs=[x], outputs=[y],
name='test_logsoftmax_axis_0')

node = onnx.helper.make_node(
'LogSoftmax',
inputs=['x'],
outputs=['y'],
axis=1,
)
y = logsoftmax_2d(x.reshape(3, 20)).reshape(3, 4, 5)
expect(node, inputs=[x], outputs=[y],
name='test_logsoftmax_axis_1')

# default axis is 1
node = onnx.helper.make_node(
'LogSoftmax',
inputs=['x'],
outputs=['y'],
)
expect(node, inputs=[x], outputs=[y],
name='test_logsoftmax_default_axis')

node = onnx.helper.make_node(
'LogSoftmax',
inputs=['x'],
outputs=['y'],
axis=2,
)
y = logsoftmax_2d(x.reshape(12, 5)).reshape(3, 4, 5)
expect(node, inputs=[x], outputs=[y],
name='test_logsoftmax_axis_2')

### Loop

Generic Looping construct. This loop has multiple termination conditions:

1. Trip count. Iteration count specified at runtime. Set by specifying the input M. Optional. Set to empty string to omit. Note that a static trip count (specified at graph construction time) can be specified by passing in a constant node for input M.
2. Loop termination condition. This is an input to the op that determines whether to run the first iteration and also a loop-carried dependency for the body graph. The body graph must yield a value for the condition variable, whether this input is provided or not.

This table summarizes the operating modes of this operator with equivalent C-style code:

  Operator inputs defined as (max_trip_count, condition_var).

input ("", ""):
for (int i=0; ; ++i) {
cond = ... // Note this value is ignored, but is required in the body
}

input ("", cond) // Note this is analogous to a while loop
bool cond = ...;
for (int i=0; cond; ++i) {
cond = ...;
}

input ("", 1) // Note this is analogous to a do-while loop
bool cond = true
for (int i=0; cond; ++i) {
cond = ...;
}

input (trip_count, "") // Note this is analogous to a for loop
int trip_count = ...
for (int i=0; i < trip_count; ++i) {
cond = ...; // ignored
}

input (trip_count, cond)
int trip_count = ...;
bool cond = ...;
for (int i=0; i < trip_count && cond; ++i) {
cond = ...;
}


Sample usage - cond as well as trip count

  graph predict-net {
%a = Constant[value = <Scalar Tensor [3]>]()
%b = Constant[value = <Scalar Tensor [6]>]()
%keepgoing = Constant[value = <Scalar Tensor [1]>]()
%max_trip_count = Constant[value = <Scalar Tensor [10]>]()
%keepgoing_out, %b_out, %user_defined_vals = Loop[body = <graph body-net>](%max_trip_count, %keepgoing, %b)
return
}

graph body-net (
%i[INT32, scalar]
%keepgoing[BOOL, scalar]
%b[INT32, scalar]
) {
%my_local = Add(%a, %b)
%b_out = Sub(%a, %b)
%keepgoing_out = Greater(%my_local, %b_out)
%user_defined_vals = Add(%b, %b)
return %keepgoing_out, %b_out, %user_defined_vals
}


Sample equivalent C code

  {
/* User-defined code (enclosing scope) */
int a = 3, b = 6;
bool keepgoing = true; // Analogous to input cond
/* End user-defined code */

/* Implicitly-defined code */
const int max_trip_count = 10; // Analogous to input M
int user_defined_vals[]; // Imagine this is resizable
/* End implicitly-defined code */
for (int i=0; i < max_trip_count && keepgoing; ++i) {
/* User-defined code (loop body) */
int my_local = a + b; // Reading values in the enclosing scope is fine
b = a - b; // writes fine if we specify b as a loop-carried dependency
keepgoing = my_local > b; // keepgoing is a loop-carried dependency
user_defined_vals[i] = b + b;
/* End user-defined code */
}
// my_local = 123; // Can't do this. my_local was defined in the the body

// These below values are live-out from the loop and therefore accessible
b_out; user_defined_vals; keepgoing_out;
}


There are several things of note in this code snippet:

1. Values from the enclosing scope (i.e. variable a here) are in scope and can be referenced in the inputs of the loop.
2. Any variables which you wish to make available in the enclosing scope (i.e. the variables b and keepgoing) must be declared as either loop-carried dependencies (both at the op inputs and output and at the body net input and output) or scan_outputs.
3. Values created in the body cannot be accessed in the enclosing scope.

Note that the semantics of this op support "diagonal" or "wavefront" execution. (See Step 3 here for an example: https://devblogs.nvidia.com/optimizing-recurrent-neural-networks-cudnn-5/). Frontends should emit multi-layer RNNs as a series of While operators (with time being the inner looping dimension), with each successive layer consuming the scan_outputs from the previous layer, possibly going through several point-wise operators (e.g. dropout, residual connections, linear layer).

#### Version

This version of the operator has been available since version 1 of the default ONNX operator set.

#### Attributes

body : graph (required)
The graph run each iteration. It has 2+N inputs: (iteration_num, condition, loop carried dependencies...). It has 1+N+K outputs: (condition, loop carried dependencies..., scan_outputs...). Each scan_output is created by concatenating the value of the specified output value at the end of each iteration of the loop. It is an error if the dimensions or data type of these scan_outputs change across loop iterations.

#### Inputs (3 - ∞)

M : I
A maximum trip-count for the loop specified at runtime. Optional. pass empty string to skip.
cond : B
A boolean termination condition. Pass empty string to skip.
v_initial (variadic) : V
The initial values of any loop-carried dependencies (values that change across loop iterations)

#### Outputs (1 - ∞)

v_final_and_scan_outputs (variadic) : V
Final N loop carried dependency values then K scan_outputs

#### Type Constraints

V : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double), tensor(string), tensor(bool), tensor(complex64), tensor(complex128)
All Tensor types
I : int64
Only int64
B : bool
Only bool

### LpNormalization

Given a matrix, apply Lp-normalization along the provided axis.

#### Version

This version of the operator has been available since version 1 of the default ONNX operator set.

#### Attributes

axis : int (default is -1)
The axis on which to apply normalization, -1 mean last axis.
p : int (default is 2)
The order of the normalization, only 1 or 2 are supported.

input : T
Input matrix

#### Outputs

output : T
Matrix after normalization

#### Type Constraints

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

### LpPool

LpPool consumes an input tensor X and applies Lp pooling across the tensor according to kernel sizes, stride sizes, and pad lengths. Lp pooling consisting of computing the Lp norm 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.

#### Version

This version of the operator has been available since version 2 of the default ONNX operator set.

Other versions of this operator: LpPool-1

#### Attributes

auto_pad : string (default is NOTSET)
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 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. DEPRECATION NOTE: auto_pad is only intended to support legacy uses, and for framework authors, one is explicitly encouraged to use explicit padding specified in the pads attribute.
kernel_shape : list of ints (required)
The size of the kernel along each axis.
p : int (default is 2)
p value of the Lp norm used to pool over the input data.
pads : list of ints
Padding for the beginning and ending along each 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 axis.
strides : list of ints
Stride along each axis.

#### Inputs

X : 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.

#### Outputs

Y : T
Output data tensor from Lp pooling across the input tensor. Dimensions will vary based on various kernel, stride, and pad sizes.

#### Type Constraints

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

### MatMul

Matrix product that behaves like numpy.matmul: https://docs.scipy.org/doc/numpy-1.13.0/reference/generated/numpy.matmul.html

#### Version

This version of the operator has been available since version 9 of the default ONNX operator set.

Other versions of this operator: MatMul-1

#### Inputs

A : T
N-dimensional matrix A
B : T
N-dimensional matrix B

#### Outputs

Y : T
Matrix multiply results from A * B

#### Type Constraints

T : tensor(float16), tensor(float), tensor(double), tensor(uint32), tensor(uint64), tensor(int32), tensor(int64)
Constrain input and output types to float/int tensors.

#### Examples

matmul
node = onnx.helper.make_node(
'MatMul',
inputs=['a', 'b'],
outputs=['c'],
)

# 2d
a = np.random.randn(3, 4).astype(np.float32)
b = np.random.randn(4, 3).astype(np.float32)
c = np.matmul(a, b)
expect(node, inputs=[a, b], outputs=[c],
name='test_matmul_2d')

# 3d
a = np.random.randn(2, 3, 4).astype(np.float32)
b = np.random.randn(2, 4, 3).astype(np.float32)
c = np.matmul(a, b)
expect(node, inputs=[a, b], outputs=[c],
name='test_matmul_3d')

# 4d
a = np.random.randn(1, 2, 3, 4).astype(np.float32)
b = np.random.randn(1, 2, 4, 3).astype(np.float32)
c = np.matmul(a, b)
expect(node, inputs=[a, b], outputs=[c],
name='test_matmul_4d')

### Max

Element-wise max of each of the input tensors (with Numpy-style broadcasting support). All inputs and outputs must have the same data type. This operator supports multidirectional (i.e., Numpy-style) broadcasting; for more details please check the doc.

#### Version

This version of the operator has been available since version 8 of the default ONNX operator set.

Other versions of this operator: Max-1, Max-6

#### Inputs (1 - ∞)

data_0 (variadic) : T
List of tensors for max.

max : T
Output tensor.

#### Type Constraints

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

#### Examples

max
data_0 = np.array([3, 2, 1]).astype(np.float32)
data_1 = np.array([1, 4, 4]).astype(np.float32)
data_2 = np.array([2, 5, 3]).astype(np.float32)
result = np.array([3, 5, 4]).astype(np.float32)
node = onnx.helper.make_node(
'Max',
inputs=['data_0', 'data_1', 'data_2'],
outputs=['result'],
)
expect(node, inputs=[data_0, data_1, data_2], outputs=[result],
name='test_max_example')

node = onnx.helper.make_node(
'Max',
inputs=['data_0'],
outputs=['result'],
)
expect(node, inputs=[data_0], outputs=[data_0],
name='test_max_one_input')

result = np.maximum(data_0, data_1)
node = onnx.helper.make_node(
'Max',
inputs=['data_0', 'data_1'],
outputs=['result'],
)
expect(node, inputs=[data_0, data_1], outputs=[result],
name='test_max_two_inputs')

### MaxPool

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.

#### Version

This version of the operator has been available since version 8 of the default ONNX operator set.

Other versions of this operator: MaxPool-1

#### Attributes

auto_pad : string (default is NOTSET)
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 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. DEPRECATION NOTE: auto_pad is only intended to support legacy uses, and for framework authors, one is explicitly encouraged to use explicit padding specified in the pads attribute.
kernel_shape : list of ints (required)
The size of the kernel along each axis.
pads : list of ints
Padding for the beginning and ending along each 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 axis.
storage_order : int (default is 0)
The storage order of the tensor. 0 is row major, and 1 is column major.
strides : list of ints
Stride along each axis.

#### Inputs

X : 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 (1 - 2)

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

#### Examples

maxpool_1d_default
"""
iutput_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)
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
"""
iutput_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)
y = pool(padded, x_shape, kernel_shape, strides, out_shape, (0, 0), 'MAX')

expect(node, inputs=[x], outputs=[y], name='test_maxpool_2d_default')
"""
iutput_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)
out_shape = get_output_shape('VALID', np.add(x_shape[2:], pad_shape), kernel_shape, strides)
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')
"""
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_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],
)
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_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_2d_same_lower
"""
iutput_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],
)
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)
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_same_upper
"""
iutput_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],
)
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)
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_strides
"""
iutput_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)
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_3d_default
"""
iutput_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)
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')
"""
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_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')

### MaxRoiPool

ROI max pool consumes an input tensor X and region of interests (RoIs) to apply max pooling across each RoI, to produce output 4-D tensor of shape (num_rois, channels, pooled_shape[0], pooled_shape[1]).

#### Version

This version of the operator has been available since version 1 of the default ONNX operator set.

#### Attributes

pooled_shape : list of ints (required)
ROI pool output shape (height, width).
spatial_scale : float (default is 1.0)
Multiplicative spatial scale factor to translate ROI coordinates from their input scale to the scale used when pooling.

#### Inputs

X : 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.
rois : T
RoIs (Regions of Interest) to pool over. Should be a 2-D tensor of shape (num_rois, 5) given as [[batch_id, x1, y1, x2, y2], ...].

#### Outputs

Y : T
RoI pooled output 4-D tensor of shape (num_rois, channels, pooled_shape[0], pooled_shape[1]).

#### Type Constraints

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

### MaxUnpool

MaxUnpool essentially computes the partial inverse of the MaxPool op. The input information to this op is typically the the output information from a MaxPool op. The first input tensor X is the tensor that needs to be unpooled, which is typically the pooled tensor (first output) from MaxPool. The second input tensor, I, contains the indices to the (locally maximal) elements corrsponding to the elements in the first input tensor X. Input tensor I is typically the second output of the MaxPool op. The third (optional) input is a tensor that specifies the output size of the unpooling operation.

MaxUnpool is intended to do 'partial' inverse of the MaxPool op. 'Partial' because all the non-maximal values from the original input to MaxPool are set to zero in the output of the MaxUnpool op. Pooling the result of an unpooling operation should give back the original input to the unpooling op.

MaxUnpool can produce the same output size for several input sizes, which makes unpooling op ambiguous. The third input argument, output_size, is meant to disambiguate the op and produce output tensor of known/predictable size.

In addition to the inputs, MaxUnpool takes three attributes, namely kernel_shape, strides, and pads, which define the exact unpooling op. The attributes typically have the same values as the corrsponding pooling op that the unpooling op is trying to invert.

#### Version

This version of the operator has been available since version 9 of the default ONNX operator set.

#### Attributes

kernel_shape : list of ints (required)
The size of the kernel along each axis.
pads : list of ints
Padding for the beginning and ending along each 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 axis.
strides : list of ints
Stride along each axis.

#### Inputs (2 - 3)

X : T1
Input data tensor that has to be unpooled. This tensor is typically the first output of the MaxPool op.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 ...].
I : T2
Input data tensor containing the indices corresponding to elements in the first input tensor X.This tensor is typically the second output of the MaxPool op.Dimensions must be the same as input tensor X. The indices are linear, i.e. computed considering the tensor as flattened 1-D tensor, assuming row-major storage. Also, the linear indices should not consider padding. So the values in indices are in the range [0, N x C x D1 x ... x Dn).
output_shape (optional) : T2
The shape of the output can be explicitly set which will cause pads values to be auto generated. If 'output_shape' is specified, 'pads' values are ignored.

#### Outputs

output : T2
Output data tensor that contains the result of the unpooling.

#### Type Constraints

T1 : tensor(float16), tensor(float), tensor(double)
Constrain input and output types to float tensors.
T2 : tensor(int64)
Constrain index tensor to int64

#### Examples

with_output_shape
node = onnx.helper.make_node(
'MaxUnpool',
inputs=['xT', 'xI', 'output_shape'],
outputs=['y'],
kernel_shape=[2, 2],
strides=[2, 2]
)
xT = np.array([[[[5, 6],
[7, 8]]]], dtype=np.float32)
xI = np.array([[[[5, 7],
[13, 15]]]], dtype=np.int64)
output_shape = np.array((1, 1, 5, 5), dtype=np.int64)
y = np.array([[[[0, 0, 0, 0, 0],
[0, 5, 0, 6, 0],
[0, 0, 0, 0, 0],
[0, 7, 0, 8, 0],
[0, 0, 0, 0, 0]]]], dtype=np.float32)
expect(node, inputs=[xT, xI, output_shape], outputs=[y], name='test_maxunpool_export_with_output_shape')
without_output_shape
node = onnx.helper.make_node(
'MaxUnpool',
inputs=['xT', 'xI'],
outputs=['y'],
kernel_shape=[2, 2],
strides=[2, 2]
)
xT = np.array([[[[1, 2],
[3, 4]]]], dtype=np.float32)
xI = np.array([[[[5, 7],
[13, 15]]]], dtype=np.int64)
y = np.array([[[[0, 0, 0, 0],
[0, 1, 0, 2],
[0, 0, 0, 0],
[0, 3, 0, 4]]]], dtype=np.float32)
expect(node, inputs=[xT, xI], outputs=[y], name='test_maxunpool_export_without_output_shape')

### Mean

Element-wise mean of each of the input tensors (with Numpy-style broadcasting support). All inputs and outputs must have the same data type. This operator supports multidirectional (i.e., Numpy-style) broadcasting; for more details please check the doc.

#### Version

This version of the operator has been available since version 8 of the default ONNX operator set.

Other versions of this operator: Mean-1, Mean-6

#### Inputs (1 - ∞)

data_0 (variadic) : T
List of tensors for mean.

mean : T
Output tensor.

#### Type Constraints

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

#### Examples

mean
data_0 = np.array([3, 0, 2]).astype(np.float32)
data_1 = np.array([1, 3, 4]).astype(np.float32)
data_2 = np.array([2, 6, 6]).astype(np.float32)
result = np.array([2, 3, 4]).astype(np.float32)
node = onnx.helper.make_node(
'Mean',
inputs=['data_0', 'data_1', 'data_2'],
outputs=['result'],
)
expect(node, inputs=[data_0, data_1, data_2], outputs=[result],
name='test_mean_example')

node = onnx.helper.make_node(
'Mean',
inputs=['data_0'],
outputs=['result'],
)
expect(node, inputs=[data_0], outputs=[data_0],
name='test_mean_one_input')

result = np.divide(np.add(data_0, data_1), 2.)
node = onnx.helper.make_node(
'Mean',
inputs=['data_0', 'data_1'],
outputs=['result'],
)
expect(node, inputs=[data_0, data_1], outputs=[result],
name='test_mean_two_inputs')

### Min

Element-wise min of each of the input tensors (with Numpy-style broadcasting support). All inputs and outputs must have the same data type. This operator supports multidirectional (i.e., Numpy-style) broadcasting; for more details please check the doc.

#### Version

This version of the operator has been available since version 8 of the default ONNX operator set.

Other versions of this operator: Min-1, Min-6

#### Inputs (1 - ∞)

data_0 (variadic) : T
List of tensors for min.

min : T
Output tensor.

#### Type Constraints

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

#### Examples

min
data_0 = np.array([3, 2, 1]).astype(np.float32)
data_1 = np.array([1, 4, 4]).astype(np.float32)
data_2 = np.array([2, 5, 0]).astype(np.float32)
result = np.array([1, 2, 0]).astype(np.float32)
node = onnx.helper.make_node(
'Min',
inputs=['data_0', 'data_1', 'data_2'],
outputs=['result'],
)
expect(node, inputs=[data_0, data_1, data_2], outputs=[result],
name='test_min_example')

node = onnx.helper.make_node(
'Min',
inputs=['data_0'],
outputs=['result'],
)
expect(node, inputs=[data_0], outputs=[data_0],
name='test_min_one_input')

result = np.minimum(data_0, data_1)
node = onnx.helper.make_node(
'Min',
inputs=['data_0', 'data_1'],
outputs=['result'],
)
expect(node, inputs=[data_0, data_1], outputs=[result],
name='test_min_two_inputs')

### Mul

Performs element-wise binary multiplication (with Numpy-style broadcasting support).

This operator supports multidirectional (i.e., Numpy-style) broadcasting; for more details please check the doc.

#### Version

This version of the operator has been available since version 7 of the default ONNX operator set.

Other versions of this operator: Mul-1, Mul-6

A : T
First operand.
B : T
Second operand.

#### Outputs

C : T
Result, has same element type as two inputs

#### Type Constraints

T : tensor(uint32), tensor(uint64), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double)
Constrain input and output types to high-precision numeric tensors.

#### Examples

mul
node = onnx.helper.make_node(
'Mul',
inputs=['x', 'y'],
outputs=['z'],
)

x = np.array([1, 2, 3]).astype(np.float32)
y = np.array([4, 5, 6]).astype(np.float32)
z = x * y  # expected output [4., 10., 18.]
expect(node, inputs=[x, y], outputs=[z],
name='test_mul_example')

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.random.randn(3, 4, 5).astype(np.float32)
z = x * y
expect(node, inputs=[x, y], outputs=[z],
name='test_mul')
node = onnx.helper.make_node(
'Mul',
inputs=['x', 'y'],
outputs=['z'],
)

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.random.randn(5).astype(np.float32)
z = x * y
expect(node, inputs=[x, y], outputs=[z],
name='test_mul_bcast')

### Multinomial

Generate a tensor of samples from a multinomial distribution according to the probabilities of each of the possible outcomes.

#### Version

This version of the operator has been available since version 7 of the default ONNX operator set.

#### Attributes

dtype : int (default is 6)
(Optional) The data type for the elements of the output tensor, if not specified, we will use int32.
sample_size : int (default is 1)
Number of times to sample.
seed : float
(Optional) Seed to the random generator, if not specified we will auto generate one.

#### Inputs

input : T1
Input tensor with shape [batch_size, class_size], where class_size is the number of all possible outcomes. Each value along the axis zero represents the unnormalized log-probability of each corresponding outcome in a batch.

#### Outputs

output : T2
Output tensor with shape [batch_size, sample_size], where sample_size is the number of times to sample. Each value along the axis zero represents the outcome of the corresponding sample in a batch.

#### Type Constraints

T1 : tensor(float16), tensor(float), tensor(double)
Constrain input types to float tensors.
T2 : tensor(int32), tensor(int64)
Constrain output types to integral tensors.

### Neg

Neg takes one input data (Tensor) and produces one output data (Tensor) where each element flipped sign, y = -x, is applied to the tensor elementwise.

#### Version

This version of the operator has been available since version 6 of the default ONNX operator set.

Other versions of this operator: Neg-1

X : T
Input tensor

Y : T
Output tensor

#### Type Constraints

T : tensor(float), tensor(int32), tensor(int8), tensor(int16), tensor(int64), tensor(float16), tensor(double)
Constrain input and output types to signed numeric tensors.

#### Examples

neg
node = onnx.helper.make_node(
'Neg',
inputs=['x'],
outputs=['y'],
)

x = np.array([-4, 2]).astype(np.float32)
y = np.negative(x)  # expected output [4., -2.],
expect(node, inputs=[x], outputs=[y],
name='test_neg_example')

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.negative(x)
expect(node, inputs=[x], outputs=[y],
name='test_neg')

### Not

Returns the negation of the input tensor element-wise.

#### Version

This version of the operator has been available since version 1 of the default ONNX operator set.

X : T
Input tensor

Y : T
Output tensor

#### Type Constraints

T : tensor(bool)
Constrains input/output to boolean tensors.

#### Examples

not
node = onnx.helper.make_node(
'Not',
inputs=['x'],
outputs=['not'],
)

# 2d
x = (np.random.randn(3, 4) > 0).astype(np.bool)
expect(node, inputs=[x], outputs=[np.logical_not(x)],
name='test_not_2d')

# 3d
x = (np.random.randn(3, 4, 5) > 0).astype(np.bool)
expect(node, inputs=[x], outputs=[np.logical_not(x)],
name='test_not_3d')

# 4d
x = (np.random.randn(3, 4, 5, 6) > 0).astype(np.bool)
expect(node, inputs=[x], outputs=[np.logical_not(x)],
name='test_not_4d')

### Or

Returns the tensor resulted from performing the or logical operation elementwise on the input tensors A and B (with Numpy-style broadcasting support).

This operator supports multidirectional (i.e., Numpy-style) broadcasting; for more details please check the doc.

#### Version

This version of the operator has been available since version 7 of the default ONNX operator set.

Other versions of this operator: Or-1

#### Inputs

A : T
First input operand for the logical operator.
B : T
Second input operand for the logical operator.

C : T1
Result tensor.

#### Type Constraints

T : tensor(bool)
Constrains input to boolean tensor.
T1 : tensor(bool)
Constrains output to boolean tensor.

#### Examples

or
node = onnx.helper.make_node(
'Or',
inputs=['x', 'y'],
outputs=['or'],
)

# 2d
x = (np.random.randn(3, 4) > 0).astype(np.bool)
y = (np.random.randn(3, 4) > 0).astype(np.bool)
z = np.logical_or(x, y)
expect(node, inputs=[x, y], outputs=[z],
name='test_or2d')

# 3d
x = (np.random.randn(3, 4, 5) > 0).astype(np.bool)
y = (np.random.randn(3, 4, 5) > 0).astype(np.bool)
z = np.logical_or(x, y)
expect(node, inputs=[x, y], outputs=[z],
name='test_or3d')

# 4d
x = (np.random.randn(3, 4, 5, 6) > 0).astype(np.bool)
y = (np.random.randn(3, 4, 5, 6) > 0).astype(np.bool)
z = np.logical_or(x, y)
expect(node, inputs=[x, y], outputs=[z],
name='test_or4d')
node = onnx.helper.make_node(
'Or',
inputs=['x', 'y'],
outputs=['or'],
)

# 3d vs 1d
x = (np.random.randn(3, 4, 5) > 0).astype(np.bool)
y = (np.random.randn(5) > 0).astype(np.bool)
z = np.logical_or(x, y)
expect(node, inputs=[x, y], outputs=[z],
name='test_or_bcast3v1d')

# 3d vs 2d
x = (np.random.randn(3, 4, 5) > 0).astype(np.bool)
y = (np.random.randn(4, 5) > 0).astype(np.bool)
z = np.logical_or(x, y)
expect(node, inputs=[x, y], outputs=[z],
name='test_or_bcast3v2d')

# 4d vs 2d
x = (np.random.randn(3, 4, 5, 6) > 0).astype(np.bool)
y = (np.random.randn(5, 6) > 0).astype(np.bool)
z = np.logical_or(x, y)
expect(node, inputs=[x, y], outputs=[z],
name='test_or_bcast4v2d')

# 4d vs 3d
x = (np.random.randn(3, 4, 5, 6) > 0).astype(np.bool)
y = (np.random.randn(4, 5, 6) > 0).astype(np.bool)
z = np.logical_or(x, y)
expect(node, inputs=[x, y], outputs=[z],
name='test_or_bcast4v3d')

# 4d vs 4d
x = (np.random.randn(1, 4, 1, 6) > 0).astype(np.bool)
y = (np.random.randn(3, 1, 5, 6) > 0).astype(np.bool)
z = np.logical_or(x, y)
expect(node, inputs=[x, y], outputs=[z],
name='test_or_bcast4v4d')

### PRelu

PRelu takes input data (Tensor) and slope tensor as input, and produces one output data (Tensor) where the function f(x) = slope * x for x < 0, f(x) = x for x >= 0., is applied to the data tensor elementwise. This operator supports unidirectional broadcasting (tensor slope should be unidirectional broadcastable to input tensor X); for more details please check the doc.

#### Version

This version of the operator has been available since version 9 of the default ONNX operator set.

Other versions of this operator: PRelu-1, PRelu-6, PRelu-7

#### Inputs

X : T
Input tensor
slope : T
Slope tensor. The shape of slope can be smaller then first input X; if so, its shape must be unidirectional broadcastable to X

#### Outputs

Y : T
Output tensor (same size as X)

#### Type Constraints

T : tensor(float16), tensor(float), tensor(double), tensor(uint32), tensor(uint64), tensor(int32), tensor(int64)
Constrain input and output types to float/int tensors.

#### Examples

prelu
node = onnx.helper.make_node(
'PRelu',
inputs=['x', 'slope'],
outputs=['y'],
)

x = np.random.randn(3, 4, 5).astype(np.float32)
slope = np.random.randn(3, 4, 5).astype(np.float32)
y = np.clip(x, 0, np.inf) + np.clip(x, -np.inf, 0) * slope

expect(node, inputs=[x, slope], outputs=[y],
name='test_prelu_example')
node = onnx.helper.make_node(
'PRelu',
inputs=['x', 'slope'],
outputs=['y'],
)

x = np.random.randn(3, 4, 5).astype(np.float32)
slope = np.random.randn(5).astype(np.float32)
y = np.clip(x, 0, np.inf) + np.clip(x, -np.inf, 0) * slope

expect(node, inputs=[x, slope], outputs=[y],
name='test_prelu_broadcast')

Given data tensor, pads, mode, and value. Example: Insert 0 pads to the beginning of the second dimension. data = [ [1.0, 1.2], [2.3, 3.4], [4.5, 5.7], ] pads = [0, 2, 0, 0] output = [ [ [0.0, 0.0, 1.0, 1.2], [0.0, 0.0, 2.3, 3.4], [0.0, 0.0, 4.5, 5.7], ], ]

#### Version

This version of the operator has been available since version 2 of the default ONNX operator set.

Other versions of this operator: Pad-1

#### Attributes

mode : string (default is constant)
Three modes: constant(default), reflect, edge
pads : list of ints (required)
List of integers indicating the number of padding elements to add or remove (if negative) at the beginning and end of each axis. For 2D it is the number of pixels. pads rank should be double of the input's rank. 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.
value : float (default is 0.0)
One float, indicates the value to be filled.

data : T
Input tensor.

output : T

#### Type Constraints

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

#### Examples

node = onnx.helper.make_node(
inputs=['x'],
outputs=['y'],
mode='constant',
value=1.2,
pads=[0, 0, 1, 3, 0, 0, 2, 4],
)
x = np.random.randn(1, 3, 4, 5).astype(np.float32)
x,
pad_width=((0, 0), (0, 0), (1, 2), (3, 4)),
mode='constant',
constant_values=1.2,
)

expect(node, inputs=[x], outputs=[y],
name='test_constant_pad')
for mode in ['edge', 'reflect']:
node = onnx.helper.make_node(
inputs=['x'],
outputs=['y'],
mode=mode,
pads=[0, 0, 1, 1, 0, 0, 1, 1]
)
x = np.random.randn(1, 3, 4, 5).astype(np.float32)
x,
pad_width=((0, 0), (0, 0), (1, 1), (1, 1)),
mode=mode,
)

expect(node, inputs=[x], outputs=[y],
name='test_{}_pad'.format(mode))

### Pow

Pow takes input data (Tensor) and exponent Tensor, and produces one output data (Tensor) where the function f(x) = x^exponent, is applied to the data tensor elementwise. This operator supports multidirectional (i.e., Numpy-style) broadcasting; for more details please check the doc.

#### Version

This version of the operator has been available since version 7 of the default ONNX operator set.

Other versions of this operator: Pow-1

#### Inputs

X : T
First operand, base of the exponent.
Y : T
Second operand, power of the exponent.

#### Outputs

Z : T
Output tensor (same size as X)

#### Type Constraints

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

#### Examples

pow
node = onnx.helper.make_node(
'Pow',
inputs=['x', 'y'],
outputs=['z'],
)

x = np.array([1, 2, 3]).astype(np.float32)
y = np.array([4, 5, 6]).astype(np.float32)
z = np.power(x, y)  # expected output [1., 32., 729.]
expect(node, inputs=[x, y], outputs=[z],
name='test_pow_example')

x = np.arange(60).reshape(3, 4, 5).astype(np.float32)
y = np.random.randn(3, 4, 5).astype(np.float32)
z = np.power(x, y)
expect(node, inputs=[x, y], outputs=[z],
name='test_pow')
node = onnx.helper.make_node(
'Pow',
inputs=['x', 'y'],
outputs=['z'],
)

x = np.array([1, 2, 3]).astype(np.float32)
y = np.array(2).astype(np.float32)
z = np.power(x, y)  # expected output [1., 4., 9.]
expect(node, inputs=[x, y], outputs=[z],
name='test_pow_bcast_scalar')

node = onnx.helper.make_node(
'Pow',
inputs=['x', 'y'],
outputs=['z'],
)
x = np.array([[1, 2, 3], [4, 5, 6]]).astype(np.float32)
y = np.array([1, 2, 3]).astype(np.float32)
# expected output [[1, 4, 27], [4, 25, 216]]
z = np.power(x, y).astype(np.float32)
expect(node, inputs=[x, y], outputs=[z],
name='test_pow_bcast_array')

### RNN

Computes an one-layer simple RNN. This operator is usually supported via some custom implementation such as CuDNN.

Notations:

X - input tensor

i - input gate

t - time step (t-1 means previous time step)

Wi - W parameter weight matrix for input gate

Ri - R recurrence weight matrix for input gate

Wbi - W parameter bias vector for input gate

Rbi - R parameter bias vector for input gate

WBi - W parameter weight matrix for backward input gate

RBi - R recurrence weight matrix for backward input gate

WBbi - WR bias vectors for backward input gate

RBbi - RR bias vectors for backward input gate

H - Hidden state

num_directions - 2 if direction == bidirectional else 1

Activation functions:

Relu(x)                - max(0, x)

Tanh(x)                - (1 - e^{-2x})/(1 + e^{-2x})

Sigmoid(x)             - 1/(1 + e^{-x})

(NOTE: Below are optional)

Affine(x)              - alpha*x + beta

LeakyRelu(x)           - x if x >= 0 else alpha * x

ThresholdedRelu(x)     - x if x >= alpha else 0

ScaledTanh(x)          - alpha*Tanh(beta*x)

HardSigmoid(x)         - min(max(alpha*x + beta, 0), 1)

Elu(x)                 - x if x >= 0 else alpha*(e^x - 1)

Softsign(x)            - x/(1 + |x|)

Softplus(x)            - log(1 + e^x)


Equations (Default: f=Tanh):

- Ht = f(Xt*(Wi^T) + Ht-1*(Ri^T) + Wbi + Rbi)


This operator has optional inputs/outputs. See the doc for more details about the representation of optional arguments. An empty string may be used in the place of an actual argument's name to indicate a missing argument. Trailing optional arguments (those not followed by an argument that is present) may also be simply omitted.

#### Version

This version of the operator has been available since version 7 of the default ONNX operator set.

Other versions of this operator: RNN-1

#### Attributes

activation_alpha : list of floats
Optional scaling values used by some activation functions. The values are consumed in the order of activation functions, for example (f, g, h) in LSTM. Default values are the same as of corresponding ONNX operators.For example with LeakyRelu, the default alpha is 0.01.
activation_beta : list of floats
Optional scaling values used by some activation functions. The values are consumed in the order of activation functions, for example (f, g, h) in LSTM. Default values are the same as of corresponding ONNX operators.
activations : list of strings (default is ['Tanh', 'Tanh'])
One (or two if bidirectional) activation function for input gate. The activation function must be one of the activation functions specified above. Optional: Default Tanh if not specified.
clip : float
Cell clip threshold. Clipping bounds the elements of a tensor in the range of [-threshold, +threshold] and is applied to the input of activations. No clip if not specified.
direction : string (default is forward)
Specify if the RNN is forward, reverse, or bidirectional. Must be one of forward (default), reverse, or bidirectional.
hidden_size : int
Number of neurons in the hidden layer

#### Inputs (3 - 6)

X : T
The input sequences packed (and potentially padded) into one 3-D tensor with the shape of [seq_length, batch_size, input_size].
W : T
The weight tensor for input gate. Concatenation of Wi and WBi (if bidirectional). The tensor has shape [num_directions, hidden_size, input_size].
R : T
The recurrence weight tensor. Concatenation of Ri and RBi (if bidirectional). The tensor has shape [num_directions, hidden_size, hidden_size].
B (optional) : T
The bias tensor for input gate. Concatenation of [Wbi, Rbi] and [WBbi, RBbi] (if bidirectional). The tensor has shape [num_directions, 2*hidden_size]. Optional: If not specified - assumed to be 0.
sequence_lens (optional) : T1
Optional tensor specifying lengths of the sequences in a batch. If not specified - assumed all sequences in the batch to have length seq_length. It has shape [batch_size].
initial_h (optional) : T
Optional initial value of the hidden. If not specified - assumed to be 0. It has shape [num_directions, batch_size, hidden_size].

#### Outputs (0 - 2)

Y (optional) : T
A tensor that concats all the intermediate output values of the hidden. It has shape [seq_length, num_directions, batch_size, hidden_size].
Y_h (optional) : T
The last output value of the hidden. It has shape [num_directions, batch_size, hidden_size].

#### Type Constraints

T : tensor(float16), tensor(float), tensor(double)
Constrain input and output types to float tensors.
T1 : tensor(int32)
Constrain seq_lens to integer tensor.

#### Examples

defaults
input = np.array([[[1., 2.], [3., 4.], [5., 6.]]]).astype(np.float32)

input_size = 2
hidden_size = 4
weight_scale = 0.1

node = onnx.helper.make_node(
'RNN',
inputs=['X', 'W', 'R'],
outputs=['', 'Y'],
hidden_size=hidden_size
)

W = weight_scale * np.ones((1, hidden_size, input_size)).astype(np.float32)
R = weight_scale * np.ones((1, hidden_size, hidden_size)).astype(np.float32)

rnn = RNN_Helper(X=input, W=W, R=R)
_, Y_h = rnn.step()
expect(node, inputs=[input, W, R], outputs=[Y_h.astype(np.float32)], name='test_simple_rnn_defaults')
initial_bias
input = np.array([[[1., 2., 3.], [4., 5., 6.], [7., 8., 9.]]]).astype(np.float32)

input_size = 3
hidden_size = 5
custom_bias = 0.1
weight_scale = 0.1

node = onnx.helper.make_node(
'RNN',
inputs=['X', 'W', 'R', 'B'],
outputs=['', 'Y'],
hidden_size=hidden_size
)

W = weight_scale * np.ones((1, hidden_size, input_size)).astype(np.float32)
R = weight_scale * np.ones((1, hidden_size, hidden_size)).astype(np.float32)

# Adding custom bias
W_B = custom_bias * np.ones((1, hidden_size)).astype(np.float32)
R_B = np.zeros((1, hidden_size)).astype(np.float32)
B = np.concatenate((W_B, R_B), axis=1)

rnn = RNN_Helper(X=input, W=W, R=R, B=B)
_, Y_h = rnn.step()
expect(node, inputs=[input, W, R, B], outputs=[Y_h.astype(np.float32)],
name='test_simple_rnn_with_initial_bias')
seq_length
input = np.array([[[1., 2., 3.], [4., 5., 6.], [7., 8., 9.]],
[[10., 11., 12.], [13., 14., 15.], [16., 17., 18.]]]).astype(np.float32)

input_size = 3
hidden_size = 5

node = onnx.helper.make_node(
'RNN',
inputs=['X', 'W', 'R', 'B'],
outputs=['', 'Y'],
hidden_size=hidden_size
)

W = np.random.randn(1, hidden_size, input_size).astype(np.float32)
R = np.random.randn(1, hidden_size, hidden_size).astype(np.float32)

# Adding custom bias
W_B = np.random.randn(1, hidden_size).astype(np.float32)
R_B = np.random.randn(1, hidden_size).astype(np.float32)
B = np.concatenate((W_B, R_B), axis=1)

rnn = RNN_Helper(X=input, W=W, R=R, B=B)
_, Y_h = rnn.step()
expect(node, inputs=[input, W, R, B], outputs=[Y_h.astype(np.float32)], name='test_rnn_seq_length')

### RandomNormal

Generate a tensor with random values drawn from a normal distribution. The shape of the tensor is specified by the shape argument and the parameter of the normal distribution specified by mean and scale.

The data type is specified by the 'dtype' argument. The 'dtype' argument must be one of the data types specified in the 'DataType' enum field in the TensorProto message.

#### Version

This version of the operator has been available since version 1 of the default ONNX operator set.

#### Attributes

dtype : int (default is 1)
The data type for the elements of the output tensor. Default is TensorProto::FLOAT.
mean : float (default is 0.0)
The mean of the normal distribution.
scale : float (default is 1.0)
The standard deviation of the normal distribution.
seed : float
(Optional) Seed to the random generator, if not specified we will auto generate one.
shape : list of ints (required)
The shape of the output tensor.

#### Outputs

output : T
Output tensor of random values drawn from normal distribution

#### Type Constraints

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

### RandomNormalLike

Generate a tensor with random values drawn from a normal distribution. The shape of the output tensor is copied from the shape of the input tensor, and the parameters of the normal distribution are specified by mean and scale.

The data type is specified by the 'dtype' argument, or copied from the input tensor if not provided. The 'dtype' argument must be one of the data types specified in the 'DataType' enum field in the TensorProto message, and be valid as an output type.

#### Version

This version of the operator has been available since version 1 of the default ONNX operator set.

#### Attributes

dtype : int
(Optional) The data type for the elements of the output tensor, if not specified, we will usethe data type of the input tensor.
mean : float (default is 0.0)
The mean of the normal distribution.
scale : float (default is 1.0)
The standard deviation of the normal distribution.
seed : float
(Optional) Seed to the random generator, if not specified we will auto generate one.

#### Inputs

input : T1
Input tensor to copy shape and optionally type information from.

#### Outputs

output : T2
Output tensor of random values drawn from normal distribution

#### Type Constraints

T1 : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double), tensor(string), tensor(bool), tensor(complex64), tensor(complex128)
Constrain to any tensor type. If the dtype attribute is not provided this must be a valid output type.
T2 : tensor(float16), tensor(float), tensor(double)
Constrain output types to float tensors.

### RandomUniform

Generate a tensor with random values drawn from a uniform distribution. The shape of the tensor is specified by the shape argument and the range by low and high.

The data type is specified by the 'dtype' argument. The 'dtype' argument must be one of the data types specified in the 'DataType' enum field in the TensorProto message.

#### Version

This version of the operator has been available since version 1 of the default ONNX operator set.

#### Attributes

dtype : int (default is 1)
The data type for the elements of the output tensor. If not specified, default is TensorProto::FLOAT.
high : float (default is 1.0)
Upper boundary of the output values.
low : float (default is 0.0)
Lower boundary of the output values.
seed : float
(Optional) Seed to the random generator, if not specified we will auto generate one.
shape : list of ints (required)
The shape of the output tensor.

#### Outputs

output : T
Output tensor of random values drawn from uniform distribution

#### Type Constraints

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

### RandomUniformLike

Generate a tensor with random values drawn from a uniform distribution. The shape of the output tensor is copied from the shape of the input tensor, and the parameters of the uniform distribution are specified by low and high.

The data type is specified by the 'dtype' argument, or copied from the input tensor if not provided. The 'dtype' argument must be one of the data types specified in the 'DataType' enum field in the TensorProto message and be valid as an output type.

#### Version

This version of the operator has been available since version 1 of the default ONNX operator set.

#### Attributes

dtype : int
(Optional) The data type for the elements of the output tensor, if not specified, we will usethe data type of the input tensor.
high : float (default is 1.0)
Upper boundary of the output values.
low : float (default is 0.0)
Lower boundary of the output values.
seed : float
(Optional) Seed to the random generator, if not specified we will auto generate one.

#### Inputs

input : T1
Input tensor to copy shape and optionally type information from.

#### Outputs

output : T2
Output tensor of random values drawn from uniform distribution

#### Type Constraints

T1 : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double), tensor(string), tensor(bool), tensor(complex64), tensor(complex128)
Constrain to any tensor type. If the dtype attribute is not provided this must be a valid output type.
T2 : tensor(float16), tensor(float), tensor(double)
Constrain output types to float tensors.

### Reciprocal

Reciprocal takes one input data (Tensor) and produces one output data (Tensor) where the reciprocal is, y = 1/x, is applied to the tensor elementwise.

#### Version

This version of the operator has been available since version 6 of the default ONNX operator set.

Other versions of this operator: Reciprocal-1

X : T
Input tensor

Y : T
Output tensor

#### Type Constraints

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

#### Examples

reciprocal
node = onnx.helper.make_node(
'Reciprocal',
inputs=['x'],
outputs=['y'],
)

x = np.array([-4, 2]).astype(np.float32)
y = np.reciprocal(x)  # expected output [-0.25, 0.5],
expect(node, inputs=[x], outputs=[y],
name='test_reciprocal_example')

x = np.random.rand(3, 4, 5).astype(np.float32) + 0.5
y = np.reciprocal(x)
expect(node, inputs=[x], outputs=[y],
name='test_reciprocal')

### ReduceL1

Computes the L1 norm of the input tensor's element along the provided axes. The resulted tensor has the same rank as the input if keepdims equal 1. If keepdims equal 0, then the resulted tensor have the reduced dimension pruned.

The above behavior is similar to numpy, with the exception that numpy default keepdims to False instead of True.

#### Version

This version of the operator has been available since version 1 of the default ONNX operator set.

#### Attributes

axes : list of ints
A list of integers, along which to reduce. The default is to reduce over all the dimensions of the input tensor.
keepdims : int (default is 1)
Keep the reduced dimension or not, default 1 mean keep reduced dimension.

data : T
An input tensor.

#### Outputs

reduced : T
Reduced output tensor.

#### Type Constraints

T : tensor(uint32), tensor(uint64), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double)
Constrain input and output types to high-precision numeric tensors.

#### Examples

default_axes_keepdims
shape = [3, 2, 2]
axes = None
keepdims = 1

node = onnx.helper.make_node(
'ReduceL1',
inputs=['data'],
outputs=['reduced'],
keepdims=keepdims
)

data = np.reshape(np.arange(1, np.prod(shape) + 1, dtype=np.float32), shape)
#print(data)
#[[[1., 2.], [3., 4.]], [[5., 6.], [7., 8.]], [[9., 10.], [11., 12.]]]

reduced = np.sum(a=np.abs(data), axis=axes, keepdims=keepdims == 1)
#print(reduced)
#[[[78.]]]

expect(node, inputs=[data], outputs=[reduced],
name='test_reduce_l1_default_axes_keepdims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.sum(a=np.abs(data), axis=axes, keepdims=keepdims == 1)

expect(node, inputs=[data], outputs=[reduced],
name='test_reduce_l1_default_axes_keepdims_random')
do_not_keepdims
shape = [3, 2, 2]
axes = [2]
keepdims = 0

node = onnx.helper.make_node(
'ReduceL1',
inputs=['data'],
outputs=['reduced'],
axes=axes,
keepdims=keepdims
)

data = np.reshape(np.arange(1, np.prod(shape) + 1, dtype=np.float32), shape)
#print(data)
#[[[1., 2.], [3., 4.]], [[5., 6.], [7., 8.]], [[9., 10.], [11., 12.]]]

reduced = np.sum(a=np.abs(data), axis=tuple(axes), keepdims=keepdims == 1)
#print(reduced)
#[[3., 7.], [11., 15.], [19., 23.]]

expect(node, inputs=[data], outputs=[reduced],
name='test_reduce_l1_do_not_keepdims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.sum(a=np.abs(data), axis=tuple(axes), keepdims=keepdims == 1)

expect(node, inputs=[data], outputs=[reduced],
name='test_reduce_l1_do_not_keepdims_random')
keepdims
shape = [3, 2, 2]
axes = [2]
keepdims = 1

node = onnx.helper.make_node(
'ReduceL1',
inputs=['data'],
outputs=['reduced'],
axes=axes,
keepdims=keepdims
)

data = np.reshape(np.arange(1, np.prod(shape) + 1, dtype=np.float32), shape)
#print(data)
#[[[1., 2.], [3., 4.]], [[5., 6.], [7., 8.]], [[9., 10.], [11., 12.]]]

reduced = np.sum(a=np.abs(data), axis=tuple(axes), keepdims=keepdims == 1)
#print(reduced)
#[[[3.], [7.]], [[11.], [15.]], [[19.], [23.]]]

expect(node, inputs=[data], outputs=[reduced],
name='test_reduce_l1_keep_dims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.sum(a=np.abs(data), axis=tuple(axes), keepdims=keepdims == 1)

expect(node, inputs=[data], outputs=[reduced],
name='test_reduce_l1_keep_dims_random')

### ReduceL2

Computes the L2 norm of the input tensor's element along the provided axes. The resulted tensor has the same rank as the input if keepdims equal 1. If keepdims equal 0, then the resulted tensor have the reduced dimension pruned.

The above behavior is similar to numpy, with the exception that numpy default keepdims to False instead of True.

#### Version

This version of the operator has been available since version 1 of the default ONNX operator set.

#### Attributes

axes : list of ints
A list of integers, along which to reduce. The default is to reduce over all the dimensions of the input tensor.
keepdims : int (default is 1)
Keep the reduced dimension or not, default 1 mean keep reduced dimension.

data : T
An input tensor.

#### Outputs

reduced : T
Reduced output tensor.

#### Type Constraints

T : tensor(uint32), tensor(uint64), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double)
Constrain input and output types to high-precision numeric tensors.

#### Examples

default_axes_keepdims
shape = [3, 2, 2]
axes = None
keepdims = 1

node = onnx.helper.make_node(
'ReduceL2',
inputs=['data'],
outputs=['reduced'],
keepdims=keepdims
)

data = np.reshape(np.arange(1, np.prod(shape) + 1, dtype=np.float32), shape)
#print(data)
#[[[1., 2.], [3., 4.]], [[5., 6.], [7., 8.]], [[9., 10.], [11., 12.]]]

reduced = np.sqrt(np.sum(
a=np.square(data), axis=axes, keepdims=keepdims == 1))
#print(reduced)
#[[[25.49509757]]]

expect(node, inputs=[data], outputs=[reduced],
name='test_reduce_l2_default_axes_keepdims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.sqrt(np.sum(
a=np.square(data), axis=axes, keepdims=keepdims == 1))

expect(node, inputs=[data], outputs=[reduced],
name='test_reduce_l2_default_axes_keepdims_random')
do_not_keepdims
shape = [3, 2, 2]
axes = [2]
keepdims = 0

node = onnx.helper.make_node(
'ReduceL2',
inputs=['data'],
outputs=['reduced'],
axes=axes,
keepdims=keepdims
)

data = np.reshape(np.arange(1, np.prod(shape) + 1, dtype=np.float32), shape)
#print(data)
#[[[1., 2.], [3., 4.]], [[5., 6.], [7., 8.]], [[9., 10.], [11., 12.]]]

reduced = np.sqrt(np.sum(
a=np.square(data), axis=tuple(axes), keepdims=keepdims == 1))
#print(reduced)
#[[2.23606798, 5.],
# [7.81024968, 10.63014581],
# [13.45362405, 16.2788206]]

expect(node, inputs=[data], outputs=[reduced],
name='test_reduce_l2_do_not_keepdims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.sqrt(np.sum(
a=np.square(data), axis=tuple(axes), keepdims=keepdims == 1))

expect(node, inputs=[data], outputs=[reduced],
name='test_reduce_l2_do_not_keepdims_random')
keepdims
shape = [3, 2, 2]
axes = [2]
keepdims = 1

node = onnx.helper.make_node(
'ReduceL2',
inputs=['data'],
outputs=['reduced'],
axes=axes,
keepdims=keepdims
)

data = np.reshape(np.arange(1, np.prod(shape) + 1, dtype=np.float32), shape)
#print(data)
#[[[1., 2.], [3., 4.]], [[5., 6.], [7., 8.]], [[9., 10.], [11., 12.]]]

reduced = np.sqrt(np.sum(
a=np.square(data), axis=tuple(axes), keepdims=keepdims == 1))
#print(reduced)
#[[[2.23606798], [5.]]
# [[7.81024968], [10.63014581]]
# [[13.45362405], [16.2788206 ]]]

expect(node, inputs=[data], outputs=[reduced],
name='test_reduce_l2_keep_dims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.sqrt(np.sum(
a=np.square(data), axis=tuple(axes), keepdims=keepdims == 1))

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_l2_keep_dims_random')

### ReduceLogSum

Computes the log sum of the input tensor's element along the provided axes. The resulted tensor has the same rank as the input if keepdims equal 1. If keepdims equal 0, then the resulted tensor have the reduced dimension pruned.

The above behavior is similar to numpy, with the exception that numpy default keepdims to False instead of True.

#### Version

This version of the operator has been available since version 1 of the default ONNX operator set.

#### Attributes

axes : list of ints
A list of integers, along which to reduce. The default is to reduce over all the dimensions of the input tensor.
keepdims : int (default is 1)
Keep the reduced dimension or not, default 1 mean keep reduced dimension.

data : T
An input tensor.

#### Outputs

reduced : T
Reduced output tensor.

#### Type Constraints

T : tensor(uint32), tensor(uint64), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double)
Constrain input and output types to high-precision numeric tensors.

#### Examples

keepdims
node = onnx.helper.make_node(
'ReduceLogSum',
inputs=['data'],
outputs=["reduced"]
)
data = np.random.ranf([3, 4, 5]).astype(np.float32)
reduced = np.log(np.sum(data, keepdims=True))
expect(node, inputs=[data], outputs=[reduced],
name='test_reduce_log_sum_default')
nokeepdims
node = onnx.helper.make_node(
'ReduceLogSum',
inputs=['data'],
outputs=["reduced"],
axes=[2, 1],
keepdims=0
)
data = np.random.ranf([3, 4, 5]).astype(np.float32)
reduced = np.log(np.sum(data, axis=(2, 1), keepdims=False))
expect(node, inputs=[data], outputs=[reduced],
name='test_reduce_log_sum_desc_axes')

node = onnx.helper.make_node(
'ReduceLogSum',
inputs=['data'],
outputs=["reduced"],
axes=[0, 1],
keepdims=0
)
data = np.random.ranf([3, 4, 5]).astype(np.float32)
reduced = np.log(np.sum(data, axis=(0, 1), keepdims=False))
expect(node, inputs=[data], outputs=[reduced],
name='test_reduce_log_sum_asc_axes')

### ReduceLogSumExp

Computes the log sum exponent of the input tensor's element along the provided axes. The resulted tensor has the same rank as the input if keepdims equal 1. If keepdims equal 0, then the resulted tensor have the reduced dimension pruned.

The above behavior is similar to numpy, with the exception that numpy default keepdims to False instead of True.

#### Version

This version of the operator has been available since version 1 of the default ONNX operator set.

#### Attributes

axes : list of ints
A list of integers, along which to reduce. The default is to reduce over all the dimensions of the input tensor.
keepdims : int (default is 1)
Keep the reduced dimension or not, default 1 mean keep reduced dimension.

data : T
An input tensor.

#### Outputs

reduced : T
Reduced output tensor.

#### Type Constraints

T : tensor(uint32), tensor(uint64), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double)
Constrain input and output types to high-precision numeric tensors.

#### Examples

default_axes_keepdims
shape = [3, 2, 2]
axes = None
keepdims = 1

node = onnx.helper.make_node(
'ReduceLogSumExp',
inputs=['data'],
outputs=['reduced'],
keepdims=keepdims
)

data = np.array(
[[[5, 1], [20, 2]], [[30, 1], [40, 2]], [[55, 1], [60, 2]]],
dtype=np.float32)
reduced = np.log(np.sum(np.exp(data),
axis=axes,
keepdims=keepdims == 1))
# print(reduced)
# [[[60.00671387]]]

expect(node, inputs=[data], outputs=[reduced],
name='test_reduce_log_sum_exp_default_axes_keepdims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.log(np.sum(np.exp(data),
axis=axes,
keepdims=keepdims == 1))
expect(node, inputs=[data], outputs=[reduced],
name='test_reduce_log_sum_exp_default_axes_keepdims_random')
do_not_keepdims
shape = [3, 2, 2]
axes = [1]
keepdims = 0
node = onnx.helper.make_node(
'ReduceLogSumExp',
inputs=['data'],
outputs=['reduced'],
axes=axes,
keepdims=keepdims
)

data = np.array(
[[[5, 1], [20, 2]], [[30, 1], [40, 2]], [[55, 1], [60, 2]]],
dtype=np.float32)
reduced = np.log(np.sum(
np.exp(data), axis=tuple(axes), keepdims=keepdims == 1))
# print(reduced)
#[[20., 2.31326175]
# [40.00004578, 2.31326175]
# [60.00671387, 2.31326175]]

expect(node, inputs=[data], outputs=[reduced],
name='test_reduce_log_sum_exp_do_not_keepdims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.log(np.sum(
np.exp(data), axis=tuple(axes), keepdims=keepdims == 1))

expect(node, inputs=[data], outputs=[reduced],
name='test_reduce_log_sum_exp_do_not_keepdims_random')
keepdims
shape = [3, 2, 2]
axes = [1]
keepdims = 1
node = onnx.helper.make_node(
'ReduceLogSumExp',
inputs=['data'],
outputs=['reduced'],
axes=axes,
keepdims=keepdims
)

data = np.array(
[[[5, 1], [20, 2]], [[30, 1], [40, 2]], [[55, 1], [60, 2]]],
dtype=np.float32)
reduced = np.log(np.sum(np.exp(data),
axis=tuple(axes),
keepdims=keepdims == 1))
# print(reduced)
# [[[20., 2.31326175]]
# [[40.00004578, 2.31326175]]
# [[60.00671387, 2.31326175]]]

expect(node, inputs=[data], outputs=[reduced],
name='test_reduce_log_sum_exp_keepdims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.log(np.sum(np.exp(data),
axis=tuple(axes),
keepdims=keepdims == 1))

expect(node, inputs=[data], outputs=[reduced],
name='test_reduce_log_sum_exp_keepdims_random')

### ReduceMax

Computes the max of the input tensor's element along the provided axes. The resulted tensor has the same rank as the input if keepdims equal 1. If keepdims equal 0, then the resulted tensor have the reduced dimension pruned.

The above behavior is similar to numpy, with the exception that numpy default keepdims to False instead of True.

#### Version

This version of the operator has been available since version 1 of the default ONNX operator set.

#### Attributes

axes : list of ints
A list of integers, along which to reduce. The default is to reduce over all the dimensions of the input tensor.
keepdims : int (default is 1)
Keep the reduced dimension or not, default 1 mean keep reduced dimension.

data : T
An input tensor.

#### Outputs

reduced : T
Reduced output tensor.

#### Type Constraints

T : tensor(uint32), tensor(uint64), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double)
Constrain input and output types to high-precision numeric tensors.

#### Examples

default_axes_keepdims
shape = [3, 2, 2]
axes = None
keepdims = 1
node = onnx.helper.make_node(
'ReduceMax',
inputs=['data'],
outputs=['reduced'],
keepdims=keepdims)

data = np.array([[[5, 1], [20, 2]], [[30, 1], [40, 2]], [[55, 1], [60, 2]]], dtype=np.float32)
reduced = np.maximum.reduce(data, axis=axes, keepdims=keepdims == 1)
#print(reduced)
[[[60.]]]

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_max_default_axes_keepdim_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.maximum.reduce(data, axis=axes, keepdims=keepdims == 1)

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_max_default_axes_keepdims_random')
do_not_keepdims
shape = [3, 2, 2]
axes = [1]
keepdims = 0

node = onnx.helper.make_node(
'ReduceMax',
inputs=['data'],
outputs=['reduced'],
axes=axes,
keepdims=keepdims)

data = np.array([[[5, 1], [20, 2]], [[30, 1], [40, 2]], [[55, 1], [60, 2]]], dtype=np.float32)
reduced = np.maximum.reduce(data, axis=tuple(axes), keepdims=keepdims == 1)
#print(reduced)
#[[20., 2.]
# [40., 2.]
# [60., 2.]]

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_max_do_not_keepdims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.maximum.reduce(data, axis=tuple(axes), keepdims=keepdims == 1)

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_max_do_not_keepdims_random')
keepdims
shape = [3, 2, 2]
axes = [1]
keepdims = 1

node = onnx.helper.make_node(
'ReduceMax',
inputs=['data'],
outputs=['reduced'],
axes=axes,
keepdims=keepdims)

data = np.array([[[5, 1], [20, 2]], [[30, 1], [40, 2]], [[55, 1], [60, 2]]], dtype=np.float32)
reduced = np.maximum.reduce(data, axis=tuple(axes), keepdims=keepdims == 1)
#print(reduced)
#[[[20., 2.]]
# [[40., 2.]]
# [[60., 2.]]]

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_max_keepdims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.maximum.reduce(data, axis=tuple(axes), keepdims=keepdims == 1)

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_max_keepdims_random')

### ReduceMean

Computes the mean of the input tensor's element along the provided axes. The resulted tensor has the same rank as the input if keepdims equal 1. If keepdims equal 0, then the resulted tensor have the reduced dimension pruned.

The above behavior is similar to numpy, with the exception that numpy default keepdims to False instead of True.

#### Version

This version of the operator has been available since version 1 of the default ONNX operator set.

#### Attributes

axes : list of ints
A list of integers, along which to reduce. The default is to reduce over all the dimensions of the input tensor.
keepdims : int (default is 1)
Keep the reduced dimension or not, default 1 mean keep reduced dimension.

data : T
An input tensor.

#### Outputs

reduced : T
Reduced output tensor.

#### Type Constraints

T : tensor(uint32), tensor(uint64), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double)
Constrain input and output types to high-precision numeric tensors.

#### Examples

default_axes_keepdims
shape = [3, 2, 2]
axes = None
keepdims = 1

node = onnx.helper.make_node(
'ReduceMean',
inputs=['data'],
outputs=['reduced'],
keepdims=keepdims)

data = np.array([[[5, 1], [20, 2]], [[30, 1], [40, 2]], [[55, 1], [60, 2]]], dtype=np.float32)
reduced = np.mean(data, axis=axes, keepdims=keepdims == 1)
#print(reduced)
#[[[18.25]]]

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_mean_default_axes_keepdims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.mean(data, axis=axes, keepdims=keepdims == 1)

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_mean_default_axes_keepdims_random')
do_not_keepdims
shape = [3, 2, 2]
axes = [1]
keepdims = 0

node = onnx.helper.make_node(
'ReduceMean',
inputs=['data'],
outputs=['reduced'],
axes=axes,
keepdims=keepdims)

data = np.array([[[5, 1], [20, 2]], [[30, 1], [40, 2]], [[55, 1], [60, 2]]], dtype=np.float32)
reduced = np.mean(data, axis=tuple(axes), keepdims=keepdims == 1)
#print(reduced)
#[[12.5, 1.5]
# [35., 1.5]
# [57.5, 1.5]]

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_mean_do_not_keepdims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.mean(data, axis=tuple(axes), keepdims=keepdims == 1)

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_mean_do_not_keepdims_random')
keepdims
shape = [3, 2, 2]
axes = [1]
keepdims = 1

node = onnx.helper.make_node(
'ReduceMean',
inputs=['data'],
outputs=['reduced'],
axes=axes,
keepdims=keepdims)

data = np.array([[[5, 1], [20, 2]], [[30, 1], [40, 2]], [[55, 1], [60, 2]]], dtype=np.float32)
reduced = np.mean(data, axis=tuple(axes), keepdims=keepdims == 1)
#print(reduced)
#[[[12.5, 1.5]]
# [[35., 1.5]]
# [[57.5, 1.5]]]

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_mean_keepdims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.mean(data, axis=tuple(axes), keepdims=keepdims == 1)

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_mean_keepdims_random')

### ReduceMin

Computes the min of the input tensor's element along the provided axes. The resulted tensor has the same rank as the input if keepdims equal 1. If keepdims equal 0, then the resulted tensor have the reduced dimension pruned.

The above behavior is similar to numpy, with the exception that numpy default keepdims to False instead of True.

#### Version

This version of the operator has been available since version 1 of the default ONNX operator set.

#### Attributes

axes : list of ints
A list of integers, along which to reduce. The default is to reduce over all the dimensions of the input tensor.
keepdims : int (default is 1)
Keep the reduced dimension or not, default 1 mean keep reduced dimension.

data : T
An input tensor.

#### Outputs

reduced : T
Reduced output tensor.

#### Type Constraints

T : tensor(uint32), tensor(uint64), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double)
Constrain input and output types to high-precision numeric tensors.

#### Examples

default_axes_keepdims
shape = [3, 2, 2]
axes = None
keepdims = 1

node = onnx.helper.make_node(
'ReduceMin',
inputs=['data'],
outputs=['reduced'],
keepdims=keepdims)

data = np.array([[[5, 1], [20, 2]], [[30, 1], [40, 2]], [[55, 1], [60, 2]]], dtype=np.float32)
reduced = np.minimum.reduce(data, axis=axes, keepdims=keepdims == 1)
#print(reduced)
#[[[1.]]]

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_min_default_axes_keepdims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.minimum.reduce(data, axis=axes, keepdims=keepdims == 1)

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_min_default_axes_keepdims_random')
do_not_keepdims
shape = [3, 2, 2]
axes = [1]
keepdims = 0

node = onnx.helper.make_node(
'ReduceMin',
inputs=['data'],
outputs=['reduced'],
axes=axes,
keepdims=keepdims)

data = np.array([[[5, 1], [20, 2]], [[30, 1], [40, 2]], [[55, 1], [60, 2]]], dtype=np.float32)
reduced = np.minimum.reduce(data, axis=tuple(axes), keepdims=keepdims == 1)
#print(reduced)
#[[5., 1.]
# [30., 1.]
# [55., 1.]]

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_min_do_not_keepdims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.minimum.reduce(data, axis=tuple(axes), keepdims=keepdims == 1)

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_min_do_not_keepdims_random')
keepdims
shape = [3, 2, 2]
axes = [1]
keepdims = 1

node = onnx.helper.make_node(
'ReduceMin', inputs=['data'],
outputs=['reduced'],
axes=axes,
keepdims=keepdims)

data = np.array([[[5, 1], [20, 2]], [[30, 1], [40, 2]], [[55, 1], [60, 2]]], dtype=np.float32)
reduced = np.minimum.reduce(data, axis=tuple(axes), keepdims=keepdims == 1)
#print(reduced)
#[[[5., 1.]]
# [[30., 1.]]
# [[55., 1.]]]

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_min_keepdims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.minimum.reduce(data, axis=tuple(axes), keepdims=keepdims == 1)

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_min_keepdims_random')

### ReduceProd

Computes the product of the input tensor's element along the provided axes. The resulted tensor has the same rank as the input if keepdims equal 1. If keepdims equal 0, then the resulted tensor have the reduced dimension pruned.

The above behavior is similar to numpy, with the exception that numpy default keepdims to False instead of True.

#### Version

This version of the operator has been available since version 1 of the default ONNX operator set.

#### Attributes

axes : list of ints
A list of integers, along which to reduce. The default is to reduce over all the dimensions of the input tensor.
keepdims : int (default is 1)
Keep the reduced dimension or not, default 1 mean keep reduced dimension.

data : T
An input tensor.

#### Outputs

reduced : T
Reduced output tensor.

#### Type Constraints

T : tensor(uint32), tensor(uint64), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double)
Constrain input and output types to high-precision numeric tensors.

#### Examples

default_axes_keepdims
shape = [3, 2, 2]
axes = None
keepdims = 1

node = onnx.helper.make_node(
'ReduceProd',
inputs=['data'],
outputs=['reduced'],
keepdims=keepdims)

data = np.array([[[1, 2], [3, 4]], [[5, 6], [7, 8]], [[9, 10], [11, 12]]], dtype=np.float32)
reduced = np.prod(data, axis=axes, keepdims=keepdims == 1)
#print(reduced)
#[[[4.790016e+08]]]

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_prod_default_axes_keepdims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.prod(data, axis=axes, keepdims=keepdims == 1)
expect(node, inputs=[data], outputs=[reduced], name='test_reduce_prod_default_axes_keepdims_random')
do_not_keepdims
shape = [3, 2, 2]
axes = [1]
keepdims = 0

node = onnx.helper.make_node(
'ReduceProd',
inputs=['data'],
outputs=['reduced'],
axes=axes,
keepdims=keepdims)

data = np.array([[[1, 2], [3, 4]], [[5, 6], [7, 8]], [[9, 10], [11, 12]]], dtype=np.float32)
reduced = np.prod(data, axis=tuple(axes), keepdims=keepdims == 1)
#print(reduced)
#[[3., 8.]
# [35., 48.]
# [99., 120.]]

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_prod_do_not_keepdims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.prod(data, axis=tuple(axes), keepdims=keepdims == 1)
expect(node, inputs=[data], outputs=[reduced], name='test_reduce_prod_do_not_keepdims_random')
keepdims
shape = [3, 2, 2]
axes = [1]
keepdims = 1

node = onnx.helper.make_node(
'ReduceProd',
inputs=['data'],
outputs=['reduced'],
axes=axes,
keepdims=keepdims)

data = np.array([[[1, 2], [3, 4]], [[5, 6], [7, 8]], [[9, 10], [11, 12]]], dtype=np.float32)
reduced = np.prod(data, axis=tuple(axes), keepdims=keepdims == 1)
#print(reduced)
#[[[3., 8.]]
# [[35., 48.]]
# [[99., 120.]]]

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_prod_keepdims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.prod(data, axis=tuple(axes), keepdims=keepdims == 1)
expect(node, inputs=[data], outputs=[reduced], name='test_reduce_prod_keepdims_random')

### ReduceSum

Computes the sum of the input tensor's element along the provided axes. The resulted tensor has the same rank as the input if keepdims equal 1. If keepdims equal 0, then the resulted tensor have the reduced dimension pruned.

The above behavior is similar to numpy, with the exception that numpy default keepdims to False instead of True.

#### Version

This version of the operator has been available since version 1 of the default ONNX operator set.

#### Attributes

axes : list of ints
A list of integers, along which to reduce. The default is to reduce over all the dimensions of the input tensor.
keepdims : int (default is 1)
Keep the reduced dimension or not, default 1 mean keep reduced dimension.

data : T
An input tensor.

#### Outputs

reduced : T
Reduced output tensor.

#### Type Constraints

T : tensor(uint32), tensor(uint64), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double)
Constrain input and output types to high-precision numeric tensors.

#### Examples

default_axes_keepdims
shape = [3, 2, 2]
axes = None
keepdims = 1

node = onnx.helper.make_node(
'ReduceSum',
inputs=['data'],
outputs=['reduced'],
keepdims=keepdims)

data = np.array([[[1, 2], [3, 4]], [[5, 6], [7, 8]], [[9, 10], [11, 12]]], dtype=np.float32)
reduced = np.sum(data, axis=axes, keepdims=keepdims == 1)
#print(reduced)
#[[[78.]]]

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_sum_default_axes_keepdims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.sum(data, axis=axes, keepdims=keepdims == 1)

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_sum_default_axes_keepdims_random')
do_not_keepdims
shape = [3, 2, 2]
axes = [1]
keepdims = 0

node = onnx.helper.make_node(
'ReduceSum',
inputs=['data'],
outputs=['reduced'],
axes=axes,
keepdims=keepdims)

data = np.array([[[1, 2], [3, 4]], [[5, 6], [7, 8]], [[9, 10], [11, 12]]], dtype=np.float32)
reduced = np.sum(data, axis=tuple(axes), keepdims=keepdims == 1)
#print(reduced)
#[[4., 6.]
# [12., 14.]
# [20., 22.]]

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_sum_do_not_keepdims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.sum(data, axis=tuple(axes), keepdims=keepdims == 1)

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_sum_do_not_keepdims_random')
keepdims
shape = [3, 2, 2]
axes = [1]
keepdims = 1

node = onnx.helper.make_node(
'ReduceSum',
inputs=['data'],
outputs=['reduced'],
axes=axes,
keepdims=keepdims)

data = np.array([[[1, 2], [3, 4]], [[5, 6], [7, 8]], [[9, 10], [11, 12]]], dtype=np.float32)
reduced = np.sum(data, axis=tuple(axes), keepdims=keepdims == 1)
#print(reduced)
#[[[4., 6.]]
# [[12., 14.]]
# [[20., 22.]]]

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_sum_keepdims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.sum(data, axis=tuple(axes), keepdims=keepdims == 1)

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_sum_keepdims_random')

### ReduceSumSquare

Computes the sum square of the input tensor's element along the provided axes. The resulted tensor has the same rank as the input if keepdims equal 1. If keepdims equal 0, then the resulted tensor have the reduced dimension pruned.

The above behavior is similar to numpy, with the exception that numpy default keepdims to False instead of True.

#### Version

This version of the operator has been available since version 1 of the default ONNX operator set.

#### Attributes

axes : list of ints
A list of integers, along which to reduce. The default is to reduce over all the dimensions of the input tensor.
keepdims : int (default is 1)
Keep the reduced dimension or not, default 1 mean keep reduced dimension.

data : T
An input tensor.

#### Outputs

reduced : T
Reduced output tensor.

#### Type Constraints

T : tensor(uint32), tensor(uint64), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double)
Constrain input and output types to high-precision numeric tensors.

#### Examples

default_axes_keepdims
shape = [3, 2, 2]
axes = None
keepdims = 1

node = onnx.helper.make_node(
'ReduceSumSquare',
inputs=['data'],
outputs=['reduced'],
keepdims=keepdims)

data = np.array([[[1, 2], [3, 4]], [[5, 6], [7, 8]], [[9, 10], [11, 12]]], dtype=np.float32)
reduced = np.sum(np.square(data), axis=axes, keepdims=keepdims == 1)
#print(reduced)
#[[[650.]]]

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_sum_square_default_axes_keepdims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.sum(np.square(data), axis=axes, keepdims=keepdims == 1)

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_sum_square_default_axes_keepdims_random')
do_not_keepdims
shape = [3, 2, 2]
axes = [1]
keepdims = 0

node = onnx.helper.make_node(
'ReduceSumSquare',
inputs=['data'],
outputs=['reduced'],
axes=axes,
keepdims=keepdims)

data = np.array([[[1, 2], [3, 4]], [[5, 6], [7, 8]], [[9, 10], [11, 12]]], dtype=np.float32)
reduced = np.sum(np.square(data), axis=tuple(axes), keepdims=keepdims == 1)
#print(reduced)
#[[10., 20.]
# [74., 100.]
# [202., 244.]]

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_sum_square_do_not_keepdims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.sum(np.square(data), axis=tuple(axes), keepdims=keepdims == 1)

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_sum_square_do_not_keepdims_random')
keepdims
shape = [3, 2, 2]
axes = [1]
keepdims = 1

node = onnx.helper.make_node(
'ReduceSumSquare',
inputs=['data'],
outputs=['reduced'],
axes=axes,
keepdims=keepdims)

data = np.array([[[1, 2], [3, 4]], [[5, 6], [7, 8]], [[9, 10], [11, 12]]], dtype=np.float32)
reduced = np.sum(np.square(data), axis=tuple(axes), keepdims=keepdims == 1)
#print(reduced)
#[[[10., 20.]]
# [[74., 100.]]
# [[202., 244.]]]

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_sum_square_keepdims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.sum(np.square(data), axis=tuple(axes), keepdims=keepdims == 1)

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_sum_square_keepdims_random')

### Relu

Relu takes one input data (Tensor) and produces one output data (Tensor) where the rectified linear function, y = max(0, x), is applied to the tensor elementwise.

#### Version

This version of the operator has been available since version 6 of the default ONNX operator set.

Other versions of this operator: Relu-1

X : T
Input tensor

Y : T
Output tensor

#### Type Constraints

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

#### Examples

relu
node = onnx.helper.make_node(
'Relu',
inputs=['x'],
outputs=['y'],
)
x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.clip(x, 0, np.inf)

expect(node, inputs=[x], outputs=[y],
name='test_relu')

### Reshape

Reshape the input tensor similar to numpy.reshape. First input is the data tensor, second input is a shape tensor which specifies the output shape. It outputs the reshaped tensor. At most one dimension of the new shape can be -1. In this case, the value is inferred from the size of the tensor and the remaining dimensions. A dimension could also be 0, in which case the actual dimension value is unchanged (i.e. taken from the input tensor).

#### Version

This version of the operator has been available since version 5 of the default ONNX operator set.

Other versions of this operator: Reshape-1

#### Inputs

data : T
An input tensor.
shape : tensor(int64)
Specified shape for output.

reshaped : T
Reshaped data.

#### Type Constraints

T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double), tensor(string), tensor(bool), tensor(complex64), tensor(complex128)
Constrain input and output types to all tensor types.

#### Examples

reshape
original_shape = [2, 3, 4]
test_cases = {
'reordered_dims': np.array([4, 2, 3], dtype=np.int64),
'reduced_dims': np.array([3, 8], dtype=np.int64),
'extended_dims': np.array([3, 2, 2, 2], dtype=np.int64),
'one_dim': np.array([24], dtype=np.int64),
'negative_dim': np.array([6, -1, 2], dtype=np.int64),
}
data = np.random.random_sample(original_shape).astype(np.float32)

for test_name, shape in test_cases.items():
node = onnx.helper.make_node(
'Reshape',
inputs=['data', 'shape'],
outputs=['reshaped'],
)

reshaped = np.reshape(data, shape)
expect(node, inputs=[data, shape], outputs=[reshaped],
name='test_reshape_' + test_name)

### Scan

Scan can be used to iterate over one or more scan_input tensors, constructing zero or more scan_output tensors. It combines ideas from general recurrences, functional programming constructs such as scan, fold, map, and zip and is intended to enable generalizations of RNN-like constructs for sequence-to-sequence processing. Other tensors (referred to as state_variables here) can be used to carry a state when iterating from one element to another (similar to hidden-state in RNNs, also referred to as loop-carried dependences in the context of loops). All these tensors are required to have the same shape in each iteration of the loop (a restriction imposed to enable efficient memory allocation). Many common usages involve a single scan_input tensor (where functionality similar to scan, fold and map can be obtained). When more than one scan_input is used, a behavior similar to zip is obtained.

The attribute body must be a graph, specifying the computation to be performed in every iteration. It takes as input the current values of the state_variables and the current iterated element of the scan_inputs. It must return the (updated) values of the state_variables and zero or more scan_output_element tensors. The values of the scan_output_element tensors are concatenated over all the iterations to produce the scan_output values of the scan construct (similar to the concatenated intermediate hidden-state values of RNN-like constructs).

The scan operation returns the final values of the state_variables as well as the scan_outputs.

The operation supports batching, and the batch-axis is required to be 0. When multiple scan_input tensors are used, they must all have the same batch-size, and they must all have the same maximum-sequence-length (the dimensionality of the sequence axis or scan axis). The sequence axis or scan axis is required to be 1.

The operation has an optional sequence_lens input (of shape [BATCH_SIZE]) to allow variable length sequences of length <= the maximum-sequence-length. If this input is not specified, all sequences are assumed to be of length equal to maximum-sequence-length. For variable length input sequences, the scan_outputs will consist of a sequence of same length as the input, padded to the maximum-sequence-length.

The optional attribute directions can be used to scan a sequence in the reverse direction. If this attribute is omitted, all sequences are scanned in the forward direction. A bidirectional scan be performed by specifying the same tensor input twice in the scan_inputs, once with a forward direction, and once with a backward direction.

Note that because of the ONNX restriction that only the last parameter of an operator can be variadic, the initial-states and scan-inputs are listed together as one input parameter. Similarly, the final-states and scan-outputs are listed together as one output parameter. The attribute num_scan_inputs indicates the number M of scan-inputs.

The behavior of

  Scan <
num_scan_inputs = m,
body = loop-body
> (sequence_lengths, init_1, ..., init_n, scan_1, ..., scan_m)


is equivalent to the following pseudo-code:

  // T.shape[0] denotes the batch-size of T
// The batch-size of scan_1, ..., scan_m are all required to be equal
batch_size = scan_1.shape[0];

// scan_i.shape[1] denotes the (max) sequence-length of scan_i
// scan_i.shape[1] is required to be equal to scan_j.shape[1] for all i,j.
max_sequence_length = scan_1.shape[1];

for (int batch = 0; batch < batch_size; ++batch) {
// initialize state-variables
st_1 = init_1; ... st_n = init_n;
// initialize scan-output variables: [] denotes an empty tensor
scan_out_1 = []; ...; scan_out_k = [];
// identify number of iterations:
N = (sequence_lengths specified) ? sequence_lengths[batch] : max_sequence_length;

// execute loop
for (int t = 0; t < N; ++t) {
// generate the scan-input elements: the notation T<axis=k>[t] indicates the sub-tensor
// of rank one less than T obtained by indexing T at position t along axis k.
si_1 = (scan_1<axis=0>[batch])<axis=1>[t];
... ;
si_m = (scan_m<axis=0>[batch])<axis=1>[t];
// execute loop-body
st_1, ..., st_n, so_1, ..., so_k = loop-body(st_1, ..., st_n, si_1, ..., si_m)
// accumulate the scan-output elements
scan_out_1 = Concat<axis=0>(scan_out_1, so_1); ... ; scan_out_k = Concat<axis=0>(scan_out_k, so_k);
}
// accumulate the outputs for this batch:
bst_1[batch] = st_1; ..., bst_n[batch] = st_n;
// Note scan-outputs will have size max_sequence_length, but only first N values will be meaningful.
// The remaining values have an undefined value.
b_scan_out_1[batch] = scan_out_1; ...; b_scan_out_k[batch] = scan_out_k;
}
return bst_1, ..., bst_n, b_scan_out_1, ..., b_scan_out_k;


Sample usage: Encoding RNN using a Scan The following example shows how a simple RNN over an input tensor %X, with weight tensor %Wi, recurrence weight tensor %Ri, bias tensors %Wbi and %Rbi, and initial hidden-state %H_0 can be encoded as a ScanLoop. Note that the loop-body is a nested graph, and it directly computes %Wi, %Ri, %Wbi, and %Rbi (typically constants or initializers in the body graph). If these values are computed in the outer graph, they need to be passed in as extra state_variables.

  graph rnn-encoding {
%H_0 = ...
%X = ...
%Y_h, %Y = Scan[body = <graph rnn-cell-1>, num_scan_inputs=1]("", %H_0, %X)
return %Y, %Y_h
}

graph rnn-cell-1 (
%H_tminus1[FLOAT, tensor]
%X_t[FLOAT, tensor]
) {
%Wi = ...
%Ri = ...
%Wbi = ...
%Rbi = ...
%t1 = X_t * (Wi^T)
%t2 = H_tminus1*(Ri^T)
%t3 = Add(%t1, %t2)
%t4 = Add(%t3, %Wbi)
%t5 = Add(%t4, %Rbi)
%Ht = Tanh(%t5)
%Accumulate = Identity(%Ht)
return %Ht, %Accumulate
}


#### Version

This version of the operator has been available since version 8 of the default ONNX operator set.

#### Attributes

body : graph (required)
The graph run each iteration. It has N+M inputs: (loop state variables..., scan_input_elts...). It has N+K outputs: (loop state variables..., scan_output_elts...). Each scan_output is created by concatenating the value of the specified scan_output_elt value at the end of each iteration of the loop. It is an error if the dimensions of these values change across loop iterations.
directions : list of ints
An optional list of M flags. The i-th element of the list specifies the direction to be scanned for the i-th scan_input tensor: 0 indicates forward direction and 1 indicates reverse direction. If omitted, all scan_input tensors will be scanned in the forward direction.
num_scan_inputs : int (required)
An attribute specifying the number of scan_inputs M.

#### Inputs (2 - ∞)

sequence_lens (optional) : I
Optional tensor specifying lengths of the sequences in a batch. If this input is not specified, all sequences are assumed to be of the maximum sequence length (the dimension of the sequence axis of the scan_input tensors).
initial_state_and_scan_inputs (variadic) : V
Initial values of the loop's N state variables followed by M scan_inputs

#### Outputs (1 - ∞)

final_state_and_scan_outputs (variadic) : V
Final values of the loop's N state variables followed by K scan_outputs

#### Type Constraints

I : tensor(int64)
Int64 tensor
V : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double), tensor(string), tensor(bool), tensor(complex64), tensor(complex128)
All Tensor types

### Selu

Selu takes one input data (Tensor) and produces one output data (Tensor) where the scaled exponential linear unit function, y = gamma * (alpha * e^x - alpha) for x <= 0, y = gamma * x for x > 0, is applied to the tensor elementwise.

#### Version

This version of the operator has been available since version 6 of the default ONNX operator set.

Other versions of this operator: Selu-1

#### Attributes

alpha : float (default is 1.67326)
Coefficient of SELU default to 1.67326319217681884765625 (i.e., float32 approximation of 1.6732632423543772848170429916717).
gamma : float (default is 1.0507)
Coefficient of SELU default to 1.05070102214813232421875 (i.e., float32 approximation of 1.0507009873554804934193349852946).

X : T
Input tensor

Y : T
Output tensor

#### Type Constraints

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

#### Examples

selu
node = onnx.helper.make_node(
'Selu',
inputs=['x'],
outputs=['y'],
alpha=2.0,
gamma=3.0
)

x = np.array([-1, 0, 1]).astype(np.float32)
# expected output [-3.79272318, 0., 3.]
y = np.clip(x, 0, np.inf) * 3.0 + (np.exp(np.clip(x, -np.inf, 0)) - 1) * 2.0 * 3.0
expect(node, inputs=[x], outputs=[y],
name='test_selu_example')

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.clip(x, 0, np.inf) * 3.0 + (np.exp(np.clip(x, -np.inf, 0)) - 1) * 2.0 * 3.0
expect(node, inputs=[x], outputs=[y],
name='test_selu')
selu_default
default_alpha = 1.67326319217681884765625
default_gamma = 1.05070102214813232421875
node = onnx.helper.make_node(
'Selu',
inputs=['x'],
outputs=['y'],
)
x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.clip(x, 0, np.inf) * default_gamma + \
(np.exp(np.clip(x, -np.inf, 0)) - 1) * default_alpha * default_gamma
expect(node, inputs=[x], outputs=[y],
name='test_selu_default')

### Shape

Takes a tensor as input and outputs an 1D int64 tensor containing the shape of the input tensor.

#### Version

This version of the operator has been available since version 1 of the default ONNX operator set.

data : T
An input tensor.

#### Outputs

shape : T1
Shape of the input tensor

#### Type Constraints

T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double), tensor(string), tensor(bool), tensor(complex64), tensor(complex128)
Input tensor can be of arbitrary type.
T1 : tensor(int64)
Constrain output to int64 tensor.

#### Examples

shape
node = onnx.helper.make_node(
'Shape',
inputs=['x'],
outputs=['y'],
)

x = np.array([
[1, 2, 3],
[4, 5, 6],
]).astype(np.float32)
y = np.array([
2, 3,
]).astype(np.int64)

expect(node, inputs=[x], outputs=[y],
name='test_shape_example')

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.array(x.shape).astype(np.int64)

expect(node, inputs=[x], outputs=[y],
name='test_shape')

### Sigmoid

Sigmoid takes one input data (Tensor) and produces one output data (Tensor) where the sigmoid function, y = 1 / (1 + exp(-x)), is applied to the tensor elementwise.

#### Version

This version of the operator has been available since version 6 of the default ONNX operator set.

Other versions of this operator: Sigmoid-1

X : T
Input tensor

Y : T
Output tensor

#### Type Constraints

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

#### Examples

sigmoid
node = onnx.helper.make_node(
'Sigmoid',
inputs=['x'],
outputs=['y'],
)

x = np.array([-1, 0, 1]).astype(np.float32)
y = 1.0 / (1.0 + np.exp(np.negative(x)))  # expected output [0.26894143, 0.5, 0.7310586]
expect(node, inputs=[x], outputs=[y],
name='test_sigmoid_example')

x = np.random.randn(3, 4, 5).astype(np.float32)
y = 1.0 / (1.0 + np.exp(np.negative(x)))
expect(node, inputs=[x], outputs=[y],
name='test_sigmoid')

### Sin

Calculates the sine of the given input tensor, element-wise.

#### Version

This version of the operator has been available since version 7 of the default ONNX operator set.

input : T
Input tensor

#### Outputs

output : T
The sine of the input tensor computed element-wise

#### Type Constraints

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

#### Examples

sin
node = onnx.helper.make_node(
'Sin',
inputs=['x'],
outputs=['y'],
)

x = np.array([-1, 0, 1]).astype(np.float32)
y = np.sin(x)
expect(node, inputs=[x], outputs=[y],
name='test_sin_example')

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.sin(x)
expect(node, inputs=[x], outputs=[y],
name='test_sin')

### Size

Takes a tensor as input and outputs a int64 scalar that equals to the total number of elements of the input tensor.

#### Version

This version of the operator has been available since version 1 of the default ONNX operator set.

data : T
An input tensor.

#### Outputs

size : T1
Total number of elements of the input tensor

#### Type Constraints

T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double), tensor(string), tensor(bool), tensor(complex64), tensor(complex128)
Input tensor can be of arbitrary type.
T1 : tensor(int64)
Constrain output to int64 tensor, which should be a scalar though.

#### Examples

size
node = onnx.helper.make_node(
'Size',
inputs=['x'],
outputs=['y'],
)

x = np.array([
[1, 2, 3],
[4, 5, 6],
]).astype(np.float32)
y = np.array(6).astype(np.int64)

expect(node, inputs=[x], outputs=[y],
name='test_size_example')

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.array(x.size).astype(np.int64)

expect(node, inputs=[x], outputs=[y],
name='test_size')

### Slice

Produces a slice of the input tensor along multiple axes. Similar to numpy: https://docs.scipy.org/doc/numpy/reference/arrays.indexing.html Slices uses axes, starts and ends attributes to specify the start and end dimension for each axis in the list of axes, it uses this information to slice the input data tensor. If a negative value is passed for any of the start or end indices, it represent number of elements before the end of that dimension. If the value passed to start or end is larger than the n (the number of elements in this dimension), it represents n. For slicing to the end of a dimension with unknown size, it is recommended to pass in INT_MAX. If axes are omitted, they are set to [0, ..., ndim-1]. Example 1: data = [ [1, 2, 3, 4], [5, 6, 7, 8], ] axes = [0, 1] starts = [1, 0] ends = [2, 3] result = [ [5, 6, 7], ] Example 2: data = [ [1, 2, 3, 4], [5, 6, 7, 8], ] starts = [0, 1] ends = [-1, 1000] result = [ [2, 3, 4], ]

#### Version

This version of the operator has been available since version 1 of the default ONNX operator set.

#### Attributes

axes : list of ints
Axes that starts and ends apply to. It's optional. If not present, will be treated as [0, 1, ..., len(starts) - 1].
ends : list of ints (required)
Ending indices (exclusive) of corresponding axis in axes
starts : list of ints (required)
Starting indices of corresponding axis in axes

#### Inputs

data : T
Tensor of data to extract slices from.

#### Outputs

output : T
Sliced data tensor.

#### Type Constraints

T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double), tensor(string), tensor(bool), tensor(complex64), tensor(complex128)
Constrain input and output types to all tensor types.

#### Examples

slice
node = onnx.helper.make_node(
'Slice',
inputs=['x'],
outputs=['y'],
axes=[0, 1],
starts=[0, 0],
ends=[3, 10],
)

x = np.random.randn(20, 10, 5).astype(np.float32)
y = x[0:3, 0:10]

expect(node, inputs=[x], outputs=[y],
name='test_slice')
slice_default_axes
node = onnx.helper.make_node(
'Slice',
inputs=['x'],
outputs=['y'],
starts=[0, 0, 3],
ends=[20, 10, 4],
)

x = np.random.randn(20, 10, 5).astype(np.float32)
y = x[:, :, 3:4]

expect(node, inputs=[x], outputs=[y],
name='test_slice_default_axes')
slice_end_out_of_bounds
node = onnx.helper.make_node(
'Slice',
inputs=['x'],
outputs=['y'],
axes=[1],
starts=[1],
ends=[1000],
)

x = np.random.randn(20, 10, 5).astype(np.float32)
y = x[:, 1:1000]

expect(node, inputs=[x], outputs=[y],
name='test_slice_end_out_of_bounds')
slice_neg
node = onnx.helper.make_node(
'Slice',
inputs=['x'],
outputs=['y'],
axes=[1],
starts=[0],
ends=[-1],
)

x = np.random.randn(20, 10, 5).astype(np.float32)
y = x[:, 0:-1]

expect(node, inputs=[x], outputs=[y],
name='test_slice_neg')
slice_start_out_of_bounds
node = onnx.helper.make_node(
'Slice',
inputs=['x'],
outputs=['y'],
axes=[1],
starts=[1000],
ends=[1000],
)

x = np.random.randn(20, 10, 5).astype(np.float32)
y = x[:, 1000:1000]

expect(node, inputs=[x], outputs=[y],
name='test_slice_start_out_of_bounds')

### Softmax

The operator computes the softmax (normalized exponential) values for each layer in the batch of the given input. The input is a 2-D tensor (Tensor) of size (batch_size x input_feature_dimensions). The output tensor has the same shape and contains the softmax values of the corresponding input.

Input does not need to explicitly be a 2D vector; rather, it will be coerced into one. For an arbitrary n-dimensional tensor input \in [a_0, a_1, ..., a_{k-1}, a_k, ..., a_{n-1}] and k is the axis provided, then input will be coerced into a 2-dimensional tensor with dimensions [a_0 * ... * a_{k-1}, a_k * ... * a_{n-1}]. For the default case where axis=1, this means the input tensor will be coerced into a 2D tensor of dimensions [a_0, a_1 * ... * a_{n-1}], where a_0 is often the batch size. In this situation, we must have a_0 = N and a_1 * ... * a_{n-1} = D. Each of these dimensions must be matched correctly, or else the operator will throw errors.

#### Version

This version of the operator has been available since version 1 of the default ONNX operator set.

#### Attributes

axis : int (default is 1)
Describes the axis of the inputs when coerced to 2D; defaults to one because the 0th axis most likely describes the batch_size

#### Inputs

input : T
The input tensor that's coerced into a 2D matrix of size (NxD) as described above.

#### Outputs

output : T
The output values with the same shape as input tensor.

#### Type Constraints

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

#### Examples

softmax
node = onnx.helper.make_node(
'Softmax',
inputs=['x'],
outputs=['y'],
)
x = np.array([[-1, 0, 1]]).astype(np.float32)
# expected output [[0.09003058, 0.24472848, 0.66524094]]
y = np.exp(x) / np.sum(np.exp(x), axis=1)
expect(node, inputs=[x], outputs=[y],
name='test_softmax_example')
softmax_axis
def softmax_2d(x):  # type: (np.ndarray) -> np.ndarray
max_x = np.max(x, axis=1).reshape((-1, 1))
exp_x = np.exp(x - max_x)
return exp_x / np.sum(exp_x, axis=1).reshape((-1, 1))

x = np.array([[0, 1, 2, 3], [10000</`