ivy.unify()#

⚠️ Warning: The tracer and the transpiler are not publicly available yet, so certain parts of this doc won’t work as expected as of now!

Ivy’s Unify function is an alias for ivy.transpile(..., to="ivy", ...). You can know more about the transpiler in the transpile() page.

Unify API#

ivy.unify(*objs, source=None, args=None, kwargs=None, **transpile_kwargs)#

Transpiles an object into Ivy code. It’s an alias to ivy.transpile(..., to="ivy", ...)

Parameters:
  • objs (Callable) – Native callable(s) to transpile.

  • source (Optional[str]) – The framework that obj is from. This must be provided unless obj is a framework-specific module.

  • args (Optional[Tuple]) – If specified, arguments that will be used to unify eagerly.

  • kwargs (Optional[dict]) – If specified, keyword arguments that will be used to unify eagerly.

  • transpile_kwargs – Arbitrary keyword arguments that will be passed to ivy.transpile.

Return type:

Union[Graph, LazyGraph, ModuleType, ivy.Module]

Returns:

A transpiled Graph or a non-initialized LazyGraph. If the object is a native trainable module, the corresponding module in the target framework will be returned. If the object is a ModuleType, the function will return a copy of the module with every method lazily transpiled.

Usage#

As we mentioned, ivy.unify() is an alias for ivy.transpile(..., to="ivy", ...). So you can use it in the same way as ivy.transpile(). In this case, instead of getting a graph composed of functions from the functional API of the target framework, the function will return a graph fully composed of ivy functions, allowing you to run the graph in any framework directly.

import ivy
ivy.set_backend("jax")

def test_fn(x):
    return jax.numpy.sum(x)

x1 = ivy.array([1., 2.])

# transpiled_func and unified_func will have the same result
transpiled_func = ivy.transpile(test_fn, to="ivy", args=(x1,))
unified_func = ivy.unify(test_fn, args=(x1,))

Sharp bits#

ivy.unify() has the same sharp bits as ivy.transpile(). You can know more about them in the overview/one_liners/transpile:Sharp bits section of the transpiler.

Examples#

Below, we will define a function in torch and try to call it with different native arguments.

Here we will define the torch function and unify it:

import ivy
import torch

def normalize(x):
    mean = torch.mean(x)
    std = torch.std(x)
    return torch.div(torch.sub(x, mean), std)

  normalize = ivy.unify(normalize, source="torch")

Now we can call the function with different ivy backends:

import numpy as np
import jax.numpy as jnp
import tensorflow as tf

# create random numpy arrays for testing
x = np.random.uniform(size=10).astype(np.float32)
ivy.set_backend("numpy")
print(normalize(x))

# jax
x_ = jnp.array(x)
ivy.set_backend("jax")
print(normalize(x_))

# tensorflow
x_ = tf.constant(x)
ivy.set_backend("tensorflow")
print(normalize(x_))

# torch
x_ = torch.tensor(x)
ivy.set_backend("torch")
print(normalize(x_))