Ivy Array

Here, we explain the ivy.Array class, which is the class used to represent all arrays in Ivy. Every Ivy method returns ivy.Array instances for all returned arrays.

The Array Class

Let’s dive straight in and check out what the ivy.Array constructor looks like.

# ivy/array/array.py
class Array(
    ArrayWithActivations,
    ArrayWithCreation,
    ArrayWithDataTypes,
    ArrayWithDevice,
    ArrayWithElementwise,
    ArrayWithGeneral,
    ArrayWithGradients,
    ArrayWithImage,
    ArrayWithLayers,
    ArrayWithLinearAlgebra,
    ArrayWithLosses,
    ArrayWithManipulation,
    ArrayWithNorms,
    ArrayWithRandom,
    ArrayWithSearching,
    ArrayWithSet,
    ArrayWithSorting,
    ArrayWithStatistical,
    ArrayWithUtility,
):
    def __init__(self, data):
        ArrayWithActivations.__init__(self)
        ArrayWithCreation.__init__(self)
        ArrayWithDataTypes.__init__(self)
        ArrayWithDevice.__init__(self)
        ArrayWithElementwise.__init__(self)
        ArrayWithGeneral.__init__(self)
        ArrayWithGradients.__init__(self)
        ArrayWithImage.__init__(self)
        ArrayWithLayers.__init__(self)
        ArrayWithLinearAlgebra.__init__(self)
        ArrayWithLosses.__init__(self)
        ArrayWithManipulation.__init__(self)
        ArrayWithNorms.__init__(self)
        ArrayWithRandom.__init__(self)
        ArrayWithSearching.__init__(self)
        ArrayWithSet.__init__(self)
        ArrayWithSorting.__init__(self)
        ArrayWithStatistical.__init__(self)
        ArrayWithUtility.__init__(self)
        self._init(data)

    def _init(self, data):
        if ivy.is_ivy_array(data):
            self._data = data.data
        else:
            assert ivy.is_native_array(data)
            self._data = data
        self._shape = self._data.shape
        self._size = (
            functools.reduce(mul, self._data.shape) if len(self._data.shape) > 0 else 0
        )
        self._dtype = ivy.dtype(self._data)
        self._device = ivy.dev(self._data)
        self._dev_str = ivy.as_ivy_dev(self._device)
        self._pre_repr = "ivy."
        if "gpu" in self._dev_str:
            self._post_repr = ", dev={})".format(self._dev_str)
        else:
            self._post_repr = ")"
        self.framework_str = ivy.current_backend_str()

    # Properties #
    # -----------#

    # noinspection PyPep8Naming
    @property
    def mT(self):
        assert len(self._data.shape) >= 2
        return ivy.matrix_transpose(self._data)

    @property
    def data(self):
        return self._data

    @property
    def shape(self):
        return ivy.Shape(self._shape)

We can see that the ivy.Array class is a simple wrapper around an ivy.NativeArray class (such as np.ndarray, torch.Tensor etc), stored in the self._data attribute.

This all makes sense, but the first question you might ask is, why do we need a dedicated ivy.Array class at all?

Can’t we just operate with the native arrays directly such as np.ndarray, torch.Tensor etc. when calling ivy methods?

This is a great question, and has a couple of answers with varying importance. Perhaps the most important motivation for having a dedicated ivy.Array class is the unification of array operators, which we discuss next!

Unifying Operators

Let’s assume that there is no such thing as the ivy.Array class, and we are just returning native arrays from all Ivy methods.

Consider the code below:

ivy.set_backend(...)
x = ivy.array([1, 2, 3])
x[0] = 0
print(x)

Let’s first assume we use numpy in the backend by calling ivy.set_backend('numpy') in the first line. x would then be a np.ndarray instance.

In this case, the code will execute without error, printing array([0, 2, 3]) to the console.

Now consider we use JAX in the backend by calling ivy.set_backend('jax') in the first line. x would then be a jax.numpy.ndarray instance.

The code will now throw the error TypeError: '<class 'jaxlib.xla_extension.DeviceArray'>' object does not support item assignment. JAX arrays are immutable. Instead of x[idx] = y, use x = x.at[idx].set(y) or another .at[] method when we try to set index 0 to the value 0.

As can be seen from the error message, the reason for this is that JAX does not support inplace updates for arrays.

This is a problem. The code written above is pure Ivy code which means it should behave identically irrespective of the backend, but as we’ve just seen it behaves differently with different backends. Therefore, in this case, we could not claim that the Ivy code was truly framework-agnostic.

For the purposes of explanation, we can re-write the above code as follows:

ivy.set_backend(...)
x = ivy.array([1, 2, 3])
x.__setitem__(0, 0)
print(x)

If x is an ivy.NativeArray instance, such as torch.Tensor or np.ndarray, then the __setitem__() method is defined in the native array class, which is completely outside of our control.

However, if x is an ivy.Array instance then the __setitem__() method is defined in the ivy.Array class, which we do have control over.

Let’s take a look at how that method is implemented in the ivy.Array class:

@_native_wrapper
def __setitem__(self, query, val):
    try:
        self._data.__setitem__(query, val)
    except (AttributeError, TypeError):
        self._data = ivy.scatter_nd(
            query, val, tensor=self._data, reduction="replace"
        )._data
        self._dtype = ivy.dtype(self._data)

We can implement inplace updates in the ivy.Array class without requiring inplace updates in the backend array classes. If the backend does not support inplace updates, then we can use the ivy.scatter_nd() method to return a new array and store this in the self._data attribute.

Now, with ivy.Array instances, our code will run without error, regardless of which backend is selected. We can genuinely say our code is fully framework-agnostic.

The same logic applies to all python operators. For example, if x and y are both ivy.NativeArray instances then the following code might execute identically for all backend frameworks:

x = ivy.some_method(...)
y = ivy.some_method(...)
z = ((x + y) * 3) ** 0.5
print(z)

Similarly, for demonstration purposes, this code can be rewritten as:

x = ivy.some_method(...)
y = ivy.some_method(...)
z = x.__add__(y).__mul__(3).__pow__(0.5)
print(z)

Even if this works fine for all backend frameworks now, what if Ivy is updated to support new backends in future, and one of them behaves a little bit differently? For example, maybe one framework makes the strange decision to return rounded integer data types when integer arrays are raised to floating point powers.

Without enforcing the use of the ivy.Array class for arrays returned from Ivy methods, we would have no way to control this behaviour and unify the output z for all backends.

Therefore, with the design of Ivy, we have made the decision to require all arrays returned from Ivy methods to be instances of the ivy.Array class.

API Monkey Patching

All ivy functions with array inputs/outputs have been wrapped to return ivy.Array instances while accepting both ivy.Array and ivy.NativeArray instances. This allows for the control required to provide a unified array interface. For more detail on wrapping, see the the Function Wrapping page in deep dive.

Instance Methods

Taking a look at the class definition, you may wonder why there are so many parent classes! The only reason the Array class derives from so many different Array classes is so we can compartmentalize the different array functions into separate classes for better code readability.

All methods in the Ivy functional API are implemented as public instance methods in the ivy.Array class via inheritance. For example, a few functions in ivy.ArrayWithGeneral are shown below.

# ivy/array/general.py
class ArrayWithGeneral(abc.ABC):

    def reshape(self, newshape):
        return ivy.reshape(self, new_shape)

    def transpose(self, axes=None):
        return ivy.transpose(self, axes)

    def flip(self, axis=None, batch_shape=None):
        return ivy.flip(self, axis, batch_shape)

One benefit of these instance methods is that they can help to tidy up code. For example:

x = ivy.ones((1, 2, 3, 4, 5))

# without ivy.Array
y = ivy.reshape(ivy.flip(ivy.matrix_transpose(
            ivy.reshape(x, (6, 20))), axis=0), (2, 10, 6))

# with ivy.Array
y = x.reshape((6, 20)).matrix_transpose().flip(axis=0).reshape((2, 10, 6))

In the example above, not only is the ivy.Array approach shorter to write, but more importantly there is much better alignment between each function and the function arguments. It’s hard to work out which shape parameters align with which method in the first case, but in the second case this is crystal clear.

In addition to the functions in the topic-specific parent classes, there are about 50 builtin methods implemented directly in the ivy.Array class, most of which directly wrap a method in Ivy’s functional API. Some examples are given below.

# ivy/array/array.py
def __add__(self, other):
    return ivy.add(self, other)

def __sub__(self, other):
    return ivy.sub(self, other)

def __mul__(self, other):
    return ivy.mul(self, other)

Round Up

That should hopefully be enough to get you started with the Ivy Array 😊

Please reach out on discord if you have any questions!