Welcome to UNaIVERSE ~ https://unaiverse.io

Welcome to a new "UN(a)IVERSE," where humans and artificial agents coexist, interact, learn from each other, grow together, in a privacy and low-energy oriented reality.

UNaIVERSE is a project framed in the context of Collectionless AI, our perspective on Artificial Intelligence rooted in privacy, low energy consumption, and, more importantly, a decentralized model.

UN(a)IVERSE is a peer-to-peer network, aiming to become the new incarnation of the Web, combining (in the long run) the principles of Social Networks and AI under a privacy lens—a perspective that is crucial given how the Web, especially Social Networks, and AI are used today by both businesses and individual users.

• Enter UNaIVERSE: UNaIVERSE portal (login/register)
• Check our preprint of Collectionless AI & UNaIVERSE, to explore UNaIVERSE features
• Read more on our ideas: Collectionless AI website
• Read more on our business: UNaIVERSE company website

🚀 Features

Check our presentation, starting from Collectionless AI and ending up in UNaIVERSE and its features.

UNaIVERSE is a peer-to-peer network where each node is either a world or an agent. What can you do?

• You can create your own agents, based on PyTorch modules, and, in function of their capabilities, they are ready to join the existing worlds and interact with others. Feel free to join a world, stay there for a while, leave it and join another one! They can also just showcase your technology, hence not join any worlds, becoming what we call lone wolves.
• You can create your own worlds as well. Different worlds are about different topics, tasks, whatever (think about a school, a shop, a chat room, an industrial plant, ...), and you don't have to write any code to let your agent participate in a world! It is the world designer that defines the expected roles and corresponding agent behaviors (special State Machines): join a world, get a role, and you are ready to behave coherently with your role!
• In UNaIVERSE, you, as human, are an agent as the other ones. The browser is your interface to UNaIVERSE, and you are already set up! No need to install anything, just jump into the UNaIVERSE portal, login, and you are a citizen of UNaIVERSE.

Remarks:

• Are you a researcher? This is perfect to study models that learn over time (Lifelong/Continual Learning), and social dynamics of different categories of models! Feel free to propose novel ideas to exploit UNaIVERSE in your research!
• Are you in the industry or, more generally, business oriented? Think about privacy-oriented solutions that we can build over this new UN(a)IVERSE! Check the UNaIVERSE company website

⚡ Status

• We think it will always stay alpha/beta/whatever 😎, but right now there are many features we plan to add and several parts to improve, thanks to your feedback!

📦 Installation

Jump to https://unaiverse.io, create a new account (free!) or log in with an existing one. If you did not already do it, click on the top-right icon with "a person" on it: Then click on "Security" and "Regenerate" a new token (aka "key"):

COPY THE TOKEN/KEY, you won't be able to see it twice! Now, let's focus on Python:

pip install unaiverse

That's it. Of course, if you want to dive into details, you find the source code here in this repo.

🛠 Getting Started

See https://docs.unaiverse.io, and follow the path that better suits your needs (e.g., Quickstart). Here we are also providing an additional example but, again, your actual starting point is https://docs.unaiverse.io.

The simplest usage you can think of is the one which does not exploit the real features of UNaIVERSE, but it is so simple that is a good way to put you in touch with UNaIVERSE itself.

You can showcase your PyTorch networks (actually, it can be every kind of model son of the PyTorch torch.nn.Module class) as follows. Let's focus on ResNet for simplicity.

Alright, let's discuss the code in the assets/tutorial folder of this repo, composed of numbered scripts.

Step A1. Do you know how to set up a network in PyTorch?

Let us set up a ResNet50 in the most basic PyTorch manner. The code is composed of a generator of tensors interpreted as pictures (actually, an ugly tensor with randomly colored pixels) and a pretrained resnet classifier which classifies the pictures generating a probability distribution over 1,000 classes. Try to run script 1 from the assets/tutorial folder. We report it here, carefully read the comments!

import torch
import torchvision

# Downloading PyTorch module (ResNet)
net = torchvision.models.resnet50(weights="IMAGENET1K_V1").eval()

# Generating a random image (don't care about it, it is just a toy example,
# think it is a nice image!)
inp = torch.rand((1, 3, 224, 224), dtype=torch.float32)

# Inference: expects as input a tensor of type torch.float32, custom width and
# height, but 3 channels and batch dimension must be there; the output is a
# tensor with shape (1, 1000), i.e., a tensor in which batch dimension is
# present and then 1000 elements.
out = net(inp)

# Print shapes
print(f"Input shape: {tuple(inp.shape)}, dtype: {inp.dtype}")
print(f"Output shape: {tuple(out.shape)}, dtype: {out.dtype}")

Step A2. Let's create UNaIVERSE agents!

We are going to create two agents, independently running and possibly located in different places/machines.

• One is based on the resnet classifier, waiting to be asked (by some other agents) for a prediction about a given image.
• The other is the generator of tensors, ready to generate a tensor (representation of a picture) and ask another agent to classify it.

Here is the resnet classifier agent, running forever and waiting for somebody to ask for a prediction, taken from script 2 in the assets/tutorial folder:

import torch
import torchvision
from unaiverse.agent import Agent
from unaiverse.streams.dataprops import StreamType
from unaiverse.networking.node.node import Node

# Downloading PyTorch module (ResNet)
net = torchvision.models.resnet50(weights="IMAGENET1K_V1").eval()

# Agent: we pass the network as "processor".
# Check the input and output properties of the processor, they are coherent with the
# input and output shapes of ResNet; here "None" means "whatever, but this axis must be
# there!". By default, this agent will act as a serving "lone wolf", serving whoever asks for
# a prediction.
agent = Agent(proc=net,
              proc_inputs=[StreamType(data_type="tensor", tensor_shape=(None, 3, None, None),
                                      tensor_dtype=torch.float32)],
              proc_outputs=[StreamType(data_type="tensor", tensor_shape=(None, 1000),
                                       tensor_dtype=torch.float32)])

# Node hosting agent: a node will be created in your account with this name, if not
# existing; it is "hidden" meaning that only you can see it in UNaIVERSE (since it is
# just a test!); the clock speed can be tuned accordingly to your needed and computing
# power.
node = Node(node_name="Test0", hosted=agent, hidden=True, clock_delta=1. / 5.)

# Running node (forever)
node.run()

Run it. Now, here is the agent capable of generating tensors (let's say images), which is asked to get in touch with the resnet agent, taken from script 3 in the assets/tutorial folder:

import torch
from unaiverse.agent import Agent
from unaiverse.streams.dataprops import StreamType
from unaiverse.networking.node.node import Node


# Custom generator network: a module that simply generates an image with
# "random" pixel intensities; we will use this as processor of our new agent.
class Net(torch.nn.Module):
    def __init__(self):
        super().__init__()

    # The input will be ignored, and a default None value is needed
    def forward(self, x: torch.Tensor | None = None):
        inp = torch.rand((1, 3, 224, 224), dtype=torch.float32)
        print(f"Generated data shape: {tuple(inp.shape)}, dtype: {inp.dtype}")
        return inp


# Agent: we use the generator as processor.
agent = Agent(proc=Net(),
              proc_inputs=[StreamType(data_type="all")],  # Able to get every type of data (since it won't use it :))
              proc_outputs=[StreamType(data_type="tensor", tensor_shape=(1, 3, 224, 224),
                                       tensor_dtype="torch.float32")],  # These are the properties of generator output
              )


# To retrieve the result we got from the ResNet agent, we define a hook
# that will be called at the end of every run cycle
def hook(_node: Node):
    # Printing the last received data from the ResNet agent
    _out = _node.agent.get_last_streamed_data('Test0')[0]
    if _out is not None:
        _node.agent.print(f"Received data shape: {tuple(_out.shape)}, dtype: {_out.dtype}")


# Node hosting agent
node = Node(node_name="Test1", hosted=agent, hidden=True, clock_delta=1. / 5., run_hook=hook)

# Running node for 10 seconds
node.run(get_in_touch="Test0", max_time=10.0)

Run this script as well, and what will happen is that the generator will send its picture through the peer-to-peer network, reaching the resnet agent, and getting back a prediction.

Step B1. Embellishment

We can upgrade the resnet agent to take real-world images as input, instead of random tensors, and to output class names (text) instead of a probability distribution. All we need to do is to re-define the properties of the inputs/outputs of the agent processor, and add transformations. Dive into script 4:

import torchvision
import urllib.request
from unaiverse.agent import Agent
from unaiverse.streams.dataprops import StreamType
from unaiverse.networking.node.node import Node

# Downloading PyTorch module (ResNet)
net = torchvision.models.resnet50(weights="IMAGENET1K_V1").eval()

# Getting input transforms from PyTorch model
transforms = torchvision.transforms.Compose([
    torchvision.transforms.Lambda(lambda x: x.convert("RGB")),
    torchvision.models.ResNet50_Weights.IMAGENET1K_V1.transforms(),
    torchvision.transforms.Lambda(lambda x: x.unsqueeze(0))
])

# Getting output class names
with urllib.request.urlopen("https://raw.githubusercontent.com/pytorch/hub/master/imagenet_classes.txt") as f:
    c_names = [line.strip().decode('utf-8') for line in f.readlines()]

# Agent: we change the data type, to be able to handle stream of images (instead of tensors).
# We can customize the transformations from the streamed format to the processor inference
# format (every callable function is fine!). Similarly, we can customize the way we go from
# the actual output of the processor and what will be streamed (here we go from class
# probabilities to winning class name).
agent = Agent(proc=net,
              proc_inputs=[StreamType(data_type="img", stream_to_proc_transforms=transforms)],
              proc_outputs=[StreamType(data_type="text", proc_to_stream_transforms=lambda p: c_names[p.argmax(1)[0]])])

# Node hosting agent
node = Node(node_name="Test0", hosted=agent, hidden=True, clock_delta=1. / 5.)

# Running node
node.run()

Now let us promote the generator to an agent that downloads and offers a picture of a cat and expects to get back a text description of it (the class name in this case - this is script 5):

import torch
import urllib.request
from PIL import Image
from io import BytesIO
from unaiverse.agent import Agent
from unaiverse.streams.dataprops import StreamType
from unaiverse.networking.node.node import Node


# Image offering network: a module that simpy downloads and offers an image as its output
class Net(torch.nn.Module):
    def __init__(self):
        super().__init__()

    def forward(self, x: torch.Tensor | None = None):
        with urllib.request.urlopen("https://cataas.com/cat") as response:
            inp = Image.open(BytesIO(response.read()))
            # inp.show()  # Let's see the pic (watch out: random pic with a cat somewhere)
            print(f"Downloaded image shape {inp.size}, type: {type(inp)}, expected-content: cat")
        return inp


# Agent
agent = Agent(proc=Net(),
              proc_inputs=[StreamType(data_type="all")],
              proc_outputs=[StreamType(data_type="img")],  # A PIL image is being "generated" here
              behav_lone_wolf="ask")


# To retrieve the result we got from the ResNet agent, we define a hook
# that will be called at the end of every run cycle
def hook(_node: Node):
    # Printing the last received data from the ResNet agent
    out = _node.agent.get_last_streamed_data('Test0')[0]
    _node.agent.print(f"Received response: {out}")  # Now we expect a textual response
    _node.agent.print("")
    _node.agent.print(f"Notice: instead of using this agent, you can also: search for the ResNet node (ResNetAgent) "
                      f"in the UNaIVERSE portal, connect to it using our in-browser agent, select a picture from "
                      f"your disk, send it to the agent, get back the text response!")


# Node hosting agent
node = Node(node_name="Test1", hosted=agent, hidden=True, clock_delta=1. / 5., run_hook=hook)

# Running node for 45 seconds
node.run(max_time=45.0, get_in_touch="Test0")

Step B2. Connect to your ResNet agent by means of a browser running agent!

Instead of using the artificial generator agent, you can become the generator agent! Search for the ResNet node (ResNetAgent) in the UNaIVERSE portal, connect to it using the in-browser agent, select a picture from your disk, send it to the agent, get back the text response!

Step C. Unleash UNaIVERSE!

What you did so far is just to showcase your model. UNaIVERSE is composed of several worlds that you can create and customize. Your agent can enter one world at a time, stay there, leave it, enter another, and so on. Agents will behave according to what the world indicates, and you don't have to write any extra code to act in worlds you have never been into!

Alright, there are so many things to say, but examples are always a good thing! We prepared a repository with examples of many worlds and different lone wolves, go there in order to continue your journey into UNaIVERSE!

THE TUTORIAL CONTINUES WITH EXAMPLES ON HOW TO BUILD WORLDS: https://github.com/collectionlessai/unaiverse-examples

Check also https://docs.unaiverse.io, and follow the path that better suits your needs.

See you in our UNaIVERSE!

📄 License

This project is licensed under the Apache 2.0 License. Commercial licenses can be provided. See the LICENSE file for details (research, etc.). See the Contributor License Agreement CLA.md if you want to contribute. This project includes third-party libraries. See THIRD_PARTY_LICENSES.md for details.

📚 Documentation

You can find an API reference and several extremely useful information here: https://docs.unaiverse.io

🤝 Contributing

Contributions are welcome!

Please contact us in order to suggest changes, report bugs, and suggest ideas for novel applications based on UNaIVERSE!

👨‍💻 Main Authors

• Stefano Melacci (Project Leader) stefano.melacci@unisi.it
• Christian Di Maio christian.dimaio@phd.unipi.it
• Tommaso Guidi tommaso.guidi.1998@gmail.com
• Marco Gori (Scientific Advisor) marco.gori@unisi.it