Variable

class csdl_alpha.src.graph.variable.Variable(shape=None, name=None, value=None, tags=None, hierarchy=None)[source]

Attributes

value: The value of the variable used for inline evaluation

Methods

`T`()	Invert the axes of a tensor.
`dtype`	alias of `float64`
`expand`(out_shape[, action])	Expands the input scalar/tensor to the specified out_shape by repeating the tensor along certain axes determined fom the action argument.
`flatten`()	Returns a 1D version of the variable.
`get`(slices)	Similar to __getitem__ but only accepts a Slice object.
`info`()	returns a string containing information about the node
`inner`(other)	Inner product of two tensors x and y.
`print_on_update`([string])	Prints the variable value when the value is updated along with the string provided.
`print_trace`([tab])	Prints the trace of the node.
`reshape`(*shape)	Returns a reshaped version of the variable.
`save`()	Sets variable to be saved
`set`(slices, value)	Sets a sliced selection of the variable to a new value.
`set_as_design_variable`([upper, lower, ...])
`set_hierarchy`(hierarchy)
`set_value`(value)	Sets the value of a variable.

Slice
add_name
add_tag
get_base_str
post_init
set_as_constraint
set_as_objective

T()[source]

Invert the axes of a tensor. The shape of the output is the reverse of the input shape.

Parameters

xVariableLike

Returns

out: Variable

Examples

>>> recorder = csdl.Recorder(inline = True)
>>> recorder.start()
>>> x = csdl.Variable(value = np.array([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]]))
>>> csdl.transpose(x).value
array([[1., 4.],
       [2., 5.],
       [3., 6.]])
>>> x.T().value # equivalent to the above
array([[1., 4.],
       [2., 5.],
       [3., 6.]])

__getitem__(indices)[source]

Returns a sliced selection of the variable as a new variable. The slicing can be specified by a csdl Slice object or a tuple of lists of integers, integers, and slices. The slicing rules are similar to Numpy’s tensor indexing rules with some restrictions. See examples for more information.

Parameters

indicesUnion[Slice, tuple[list[int], int, slice]]: The indices to slice the variable by. See examples for more information.

Returns

out: Variable: a new variable that is a indexed selection of the original variable.

Examples

Integer indexing allows a selection of a single element in a dimension and removes that dimension in the output.

>>> recorder = csdl.Recorder(inline = True)
>>> recorder.start()
>>> x = csdl.Variable(value = np.array([1.0, 2.0, 3.0]))
>>> x[0].shape
(1,)
>>> x[0].value
array([1.])
>>> x = csdl.Variable(value = np.arange(6).reshape(2,3))
>>> x[0].shape # removes the first dimension in the output
(3,)
>>> x[0].value
array([0., 1., 2.])
>>> x[1,2].shape # returns a single element
(1,)
>>> x[1,2].value
array([5.])

Slicing allows a selection of a range of elements in a dimension using slice notation and keeps that dimension in the output.

>>> x[1:2].shape # keeps the first dimension in the output
(1, 3)
>>> x[1:2].value
array([[3., 4., 5.]])
>>> x[:].shape
(2, 3)
>>> np.all(x[:].value == x.value)
True
>>> x[1:2,:-1].shape
(1, 2)
>>> x[1:2,:-1].value
array([[3., 4.]])

Integer lists allows for selecting a coordinate of elements across multiple dimensions and compresses them to one one dimension.

>>> x[[0,1]].shape
(2, 3)
>>> x[[0,1]].value
array([[0., 1., 2.],
       [3., 4., 5.]])
>>> x[[0,1],[0,1]].shape # outputs x[0,0] and x[1,1] in a 1D array
(2,)
>>> x[[0,1],[0,1]].value # outputs x[0,0] and x[1,1] in a 1D array
array([0., 4.])

All three types of indexing can be combined.

>>> x = csdl.Variable(value = np.arange(24).reshape(4,2,3))
>>> x[[0,1],1:].shape
(2, 1, 3)
>>> x[[0,1],1:].value
array([[[ 3.,  4.,  5.]],

       [[ 9., 10., 11.]]])
>>> x[0,[0,1],[1,0]].shape
(2,)
>>> x[0,[0,1],[1,0]].value
array([1., 3.])

__init__(shape=None, name=None, value=None, tags=None, hierarchy=None)[source]

Initialize a Variable object.

Parameters

shapetuple, optional: The shape of the variable. If not provided, it will be inferred from the value.
namestr, optional: The name of the variable.
valueUnion[np.ndarray, float, int], optional: The initial value of the variable.
tagslist[str], optional: A list of tags associated with the variable.
hierarchyint, optional: The hierarchy level of the variable.

Attributes

hierarchyint: The hierarchy level of the variable.
shapetuple: The shape of the variable.
sizeint: The size of the variable.
nameslist[str]: A list of names associated with the variable.
valueUnion[np.ndarray, float, int]: The value of the variable.
tagslist[str]: A list of tags associated with the variable.

dtype: alias of float64

expand(out_shape, action=None)[source]

Expands the input scalar/tensor to the specified out_shape by repeating the tensor along certain axes determined fom the action argument. For example, action=’i->ijk’ will expand a 1D tensor to a 3D tensor by repeating the input tensor along two new axes. The action argument is optional if the input is a scalar since the scalar will be simply broadcasted to the specified out_shape.

Parameters

xVariableLike: Input scalar/tensor that needs to be expanded.
out_shapetuple of int: Desired shape of the expanded output tensor.
actionstr, default=None: Specifies the action to be taken when expanding the tensor, e.g.,`’i->ij’` expands a vector to a matrix by repeating the input vector rowwise.

Returns

Variable: Expanded output tensor as per the specified out_shape and action.

Examples

>>> recorder = csdl.Recorder(inline = True)
>>> recorder.start()
>>> x = csdl.Variable(value = 3.0)
>>> y1 = csdl.expand(x, out_shape=(2,3))
>>> y1.value
array([[3., 3., 3.],
       [3., 3., 3.]])
>>> x = csdl.Variable(value = np.array([1.0, 2.0, 3.0]))
>>> y2 = csdl.expand(x, out_shape=(2,3), action='i->ji')
>>> y2.value
array([[1., 2., 3.],
       [1., 2., 3.]])
>>> y3 = csdl.expand(x, out_shape=(3,2), action='i->ij')
>>> y3.value
array([[1., 1.],
       [2., 2.],
       [3., 3.]])
>>> y4 = csdl.expand(x, out_shape=(4,3,2), action='i->lij')
>>> y4.value
array([[[1., 1.],
        [2., 2.],
        [3., 3.]],

       [[1., 1.],
        [2., 2.],
        [3., 3.]],

       [[1., 1.],
        [2., 2.],
        [3., 3.]],

       [[1., 1.],
        [2., 2.],
        [3., 3.]]])

flatten()[source]

Returns a 1D version of the variable.

Parameters

selfVariable

Returns

out: Variable

Examples

>>> recorder = csdl.Recorder(inline = True)
>>> recorder.start()
>>> x = csdl.Variable(value = np.array([1.0, 2.0, 3.0, 4.0]))
>>> x.flatten().value # reshapes to 1 dimension
array([1., 2., 3., 4.])

get(slices)[source]

Similar to __getitem__ but only accepts a Slice object.

Parameters

slicesSlice

Returns

out: Variable

inner(other)[source]

Inner product of two tensors x and y. The result is a scalar of shape (1,). The input tensors must have the same shape.

Parameters

selfVariable: First input tensor.
otherVariableLike: Second input tensor.

Returns

out: Variable: Scalar inner product of x and y.

Examples

>>> recorder = csdl.Recorder(inline = True)
>>> recorder.start()
>>> x = csdl.Variable(value = np.array([1, 2, 3]))
>>> y = csdl.Variable(value = np.array([4, 5, 6]))
>>> x.inner(y).value
array([32.])
>>> a = csdl.Variable(value = np.array([[1, 2], [3, 4]]))
>>> b = csdl.Variable(value = np.array([[5, 6], [7, 8]]))
>>> a.inner(b).value
array([70.])

print_on_update(string=None)[source]

Prints the variable value when the value is updated along with the string provided.

Parameters

stringstr, optional: additional string to print along with the value, by default prints the name of the node

reshape(*shape)[source]

Returns a reshaped version of the variable.

Parameters

selfVariable

Returns

out: Variable

Examples

>>> recorder = csdl.Recorder(inline = True)
>>> recorder.start()
>>> x = csdl.Variable(value = np.array([1.0, 2.0, 3.0, 4.0]))
>>> csdl.reshape(x, (2,2)).value
array([[1., 2.],
       [3., 4.]])
>>> x.reshape((2,2)).value # same thing as above
array([[1., 2.],
       [3., 4.]])
>>> x.reshape(2,2).value # same thing as above
array([[1., 2.],
       [3., 4.]])
>>> x.reshape(4).value # optionally pass in integers
array([1., 2., 3., 4.])

save()[source]: Sets variable to be saved

set(slices, value)[source]

Sets a sliced selection of the variable to a new value. The slicing must be specified by a csdl Slice object. See examples for more information.

Parameters

indicesSlice: The indices to slice the variable by. See examples for more information.
valueVariableLike: The value to set the sliced selection of the variable to.

Returns

out: Variable: A new variable that represents the original variable with the sliced selection set to the new value.

Examples

The set method creates a new variable with the sliced selection set to the new value. The original variable is not modified.

>>> recorder = csdl.Recorder(inline = True)
>>> recorder.start()
>>> x = csdl.Variable(value = np.array([1.0, 2.0, 3.0]))
>>> x1 = x.set(csdl.slice[0], 0.0)
>>> x1.value
array([0., 2., 3.])

Use the csdl.slice slicer object when using slices.

>>> x1 = x.set(csdl.slice[1:3], csdl.Variable(value = np.array([4.0, 5.0])))
>>> x1.value
array([1., 4., 5.])
>>> x = csdl.Variable(value = np.array([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]]))
>>> x1 = x.set(csdl.slice[1, 1:3], csdl.Variable(value = np.array([10.0, 11.0])))
>>> x1.value
array([[ 1.,  2.,  3.],
       [ 4., 10., 11.]])

The slicing conventions are identical to those in the __getitem__ method and broadcasting from a scalar is also supported.

>>> x1 = x.set(csdl.slice[[0,1], [1,2]], csdl.Variable(value = np.array([10.0, 11.0])))
>>> x1.value
array([[ 1., 10.,  3.],
       [ 4.,  5., 11.]])
>>> x1 = x.set(csdl.slice[0, 1:], 10.0)
>>> x1.value
array([[ 1., 10., 10.],
       [ 4.,  5.,  6.]])
>>> x1 = x.set(csdl.slice[:, [0, 2]], 11.0)
>>> x1.value
array([[11.,  2., 11.],
       [11.,  5., 11.]])

Slicing with CSDL variables is also supported in some cases when the slice size is constant.

>>> start = csdl.Variable(value = 0)
>>> x1 = x.set(csdl.slice[start:start+2, 1], 10.0)
>>> x1.value
array([[ 1., 10.,  3.],
       [ 4., 10.,  6.]])

Get the same behaviour of in-place modification by returning the same variable (it is recommended to combine slices to one .set call when possible to reduce the number of operations).

>>> x = x.set(csdl.slice[0,0], 10.0)
>>> x = x.set(csdl.slice[1,0], 11.0)
>>> x = x.set(csdl.slice[1,2], 12.0)
>>> x.value
array([[10.,  2.,  3.],
       [11.,  5., 12.]])

set_hierarchy(hierarchy)[source]: Warning

This function should not need to be called by the user

set_value(value)[source]

Sets the value of a variable.

Parameters

valueUnion[np.ndarray, float, int]: Value for the variable

property value: The value of the variable used for inline evaluation