We’re on a mission to unify all ML frameworks 💥 + automate code conversions 🔄. pip install ivy-core 🚀, join our growing community 😊, and lets-unify.ai! 🦾

Contents

• Overview
• Quick Start
• Background
• Design
• Extensions
• Contributing

Overview

Ivy is an ML framework that currently supports JAX, TensorFlow, PyTorch, and Numpy. We’re very excited for you to try it out!

Next on our roadmap is to support automatic code conversions between all frameworks 🔄, and add instant multi-framework support for all open-source libraries with only a few lines of code changed! Read on to learn more 😊

The docs are split into a number of sub-pages explaining different aspects of why we created Ivy, how to use it, what we’ve got planned on our roadmap, and how to contribute! Click on the sub-headings below to check out these pages!

We use 🚧 to indicate that the feature being discussed is in development. We use ✅ to indicate that it is already implemented!

Check out the docs for more info, and check out our Google Colabs for some interactive demos!

🚨 Ivy is still at a relatively early stage of development. Expect breaking changes and sharp edges until we release version 1.2.0 in the next few weeks!

If you would like to contribute, please check out our contributor guide, and take a look at the open tasks if you'd like to dive straight in! 🧑‍💻

Quick Start

Ivy can be installed like so: pip install ivy-core You can immediately use Ivy to train a neural network, using your favorite framework in the background, like so:

import ivy

class MyModel(ivy.Module):
    def __init__(self):
        self.linear0 = ivy.Linear(3, 64)
        self.linear1 = ivy.Linear(64, 1)
        ivy.Module.__init__(self)

    def _forward(self, x):
        x = ivy.relu(self.linear0(x))
        return ivy.sigmoid(self.linear1(x))

ivy.set_backend('torch')  # change to any backend!
model = MyModel()
optimizer = ivy.Adam(1e-4)
x_in = ivy.array([1., 2., 3.])
target = ivy.array([0.])

def loss_fn(v):
    out = model(x_in, v=v)
    return ivy.mean((out - target)**2)

for step in range(100):
    loss, grads = ivy.execute_with_gradients(loss_fn, model.v)
    model.v = optimizer.step(model.v, grads)
    print('step {} loss {}'.format(step, ivy.to_numpy(loss).item()))

print('Finished training!')

This example uses PyTorch as a backend framework, but the backend can easily be changed to your favorite frameworks, such as TensorFlow, or JAX.

Framework Agnostic Functions

In the example below we show how Ivy's concatenation function is compatible with tensors from different frameworks. This is the same for ALL Ivy functions. They can accept tensors from any framework and return the correct result.

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

import ivy

jax_concatted   = ivy.concat((jnp.ones((1,)), jnp.ones((1,))), -1)
tf_concatted    = ivy.concat((tf.ones((1,)), tf.ones((1,))), -1)
np_concatted    = ivy.concat((np.ones((1,)), np.ones((1,))), -1)
torch_concatted = ivy.concat((torch.ones((1,)), torch.ones((1,))), -1)

To see a list of all Ivy methods, type ivy. into a python command prompt and press tab. You should then see output like the following:

ivy.Container(                         ivy.general                               ivy.reduce_min(
ivy.abs(                               ivy.get_device(                           ivy.reduce_prod(
ivy.acos(                              ivy.get_num_dims(                         ivy.reduce_sum(
ivy.acosh(                             ivy.gradient_descent_update(              ivy.reductions
ivy.activations                        ivy.gradient_image(                       ivy.relu(
ivy.arange(                            ivy.gradients                             ivy.reshape(
ivy.argmax(                            ivy.identity(                             ivy.round(
ivy.argmin(                            ivy.image                                 ivy.scatter_nd(
ivy.array(                             ivy.indices_where(                        ivy.seed(
ivy.asin(                              ivy.inv(                                  ivy.shape(
ivy.asinh(                             ivy.layers                                ivy.shuffle(
ivy.atan(                              ivy.leaky_relu(                           ivy.sigmoid(
ivy.atan2(                             ivy.linalg                                ivy.sin(
ivy.atanh(                             ivy.linear(                               ivy.sinh(
ivy.bilinear_resample(                 ivy.linspace(                             ivy.softmax(
ivy.cast(                              ivy.log(                                  ivy.softplus(
ivy.ceil(                              ivy.logic                                 ivy.split(
ivy.clip(                              ivy.logical_and(                          ivy.squeeze(
ivy.concatenate(                       ivy.logical_not(                          ivy.stack(
ivy.container                          ivy.logical_or(                           ivy.stack_images(
ivy.conv2d(                            ivy.math                                  ivy.stop_gradient(
ivy.core                               ivy.matmul(                               ivy.svd(
ivy.cos(                               ivy.maximum(                              ivy.tan(
ivy.cosh(                              ivy.minimum(                              ivy.tanh(
ivy.cross(                             ivy.neural_net                            ivy.tile(
ivy.cumsum(                            ivy.nn                                    ivy.to_list(
ivy.depthwise_conv2d(                  ivy.norm(                                 ivy.to_numpy(
ivy.dtype(                             ivy.one_hot(                              ivy.transpose(
ivy.execute_with_gradients(            ivy.ones(                                 ivy.unstack(
ivy.exp(                               ivy.ones_like(                            ivy.vector_norm(
ivy.expand_dims(                       ivy.pinv(                                 ivy.vector_to_skew_symmetric_matrix(
ivy.flip(                              ivy.randint(                              ivy.verbosity
ivy.floor(                             ivy.random                                ivy.where(
ivy.floormod(                          ivy.random_uniform(                       ivy.zero_pad(
ivy.backend_handler                    ivy.reduce_max(                           ivy.zeros(
ivy.gather_nd(                         ivy.reduce_mean(                          ivy.zeros_like(

Background

(a) ML Explosion

A huge number of ML tools have exploded onto the scene!

(b) Why Unify?

Why should we try to unify them?

(c) Standardization

We’re collaborating with The Consortium for Python Data API Standards

Design

Ivy can fulfill two distinct purposes:

1. Serve as a transpiler between frameworks 🚧

2. Serve as a new ML framework with multi-framework support ✅

The Ivy codebase can then be split into three categories, and can be further split into 8 distinct submodules, each of which falls into one of these three categories as follows:

(a) Building Blocks

Backend functional APIs ✅

Ivy functional API ✅

Backend Handler ✅

Ivy Compiler 🚧

(b) Ivy as a Transpiler

Front-end functional APIs 🚧

(c) Ivy as a Framework

Ivy stateful API ✅

Ivy Container ✅

Ivy Array 🚧

Extensions

(a) Applied Libraries ✅

Ivy libraries in mechanics, vision, robotics, memory, and other areas

(b) Builder [page coming soon!] ✅

ivy.Trainer, ivy.Dataset, ivy.Dataloader and other helpful classes and functions for creating training workflows in only a few lines of code

Contributing

Join our community as a code contributor, and help accelerate our journey to unify all ML frameworks! Check out all of our open tasks, and find out more info in our Contributing guide!

Citation

@article{lenton2021ivy,
  title={Ivy: Templated deep learning for inter-framework portability},
  author={Lenton, Daniel and Pardo, Fabio and Falck, Fabian and James, Stephen and Clark, Ronald},
  journal={arXiv preprint arXiv:2102.02886},
  year={2021}
}