Stateoftheart AI

Welcome! The objective of this repository is to simplify the usage of the many machine learning libraries available out there.

How do we do this?

• We collect all models and datasets into a single, centralized place (here).
• We offer a same and single function to instantiate a model or a dataset, irrespective of the source library (load_model and load_dataset).
• We provide the functionality to connect any model/dataset pair for evaluation and training purposes via a model_to_dataset function, irrespective of the source library for either of them (naturally with intrinsic limitations, e.g., cannot pair a CV model with a RL environement).
• We provide the functionality to obtain predictions i.e. evaluating a model on a given dataset predictions = model(dataset)

---

Overview of the platform

We provide the following main functions to train any model on any available dataset and get predictions:

• model = load_model('model-name'): creates an instance of a model from any source library.
• dataset = load_dataset('dataset-name'): creates an instance of a dataset from any source library.
• model, dataset = model_to_dataset(model,dataset): given a model and a dataset, this pairing function adjusts both objects to ensure they are compatible.
• predictions = model(dataset): each model instance is callable, thus you can get predictions by calling the model and passing the dataset to get predictions from.

Depending on the type of model, dataset, area, and source library, the inner workings of the aforementioned functions can wildly vary. Additionally, extra functions might be needed to fully establish the pipeline between models and datasets (e.g., in NLP where we need to account for tokenizers and embeddings).

The main contribution of this library is the development of these interfaces between external libraries warranted for the interconnection of models from one source and datasets from another, with the ultimate objective being to obtain a unique, simplified and unified approach to interacting with and leveraging all of the existing machine learning libraries.

In the following sections we document the progress of this massive undertaking, explicitly showing where the connections between libraries have already been established, and where is more work needed.

---

Areas Supported

Currently we support the areas listed below. To see more information and documentation of a specific area feel free to navigate to the respective section:

• Computer Vision
• Natural Language Processing [WORK IN PROGRESS]
• Neurosymbolic Reasoning [WORK IN PROGRESS]
• Reinforcement Learning [WORK IN PROGRESS]
• Robotics [WORK IN PROGRESS]

---

Installation

Requirements

• Python3 is required. Recommended to use Python <= 3.8.0.
• As a best practice, we recommend the use of pyenv, this way you can have different versions of Python installed in your machine and switch among them, being able to meet any Python version requirement.
• Regarding the installation process, we have detected some issues with the pip version. As of now the library installation has been fully tested and proved working with the following versions of pip: 19.2.3, 21.0.1 (latest as of March 2021)

PIP Installation

The library can be installed through pip. However, since sotaai depends on many libraries, we split the library into subpackages so that end users can install only the subpackage and dependencies they require, to install the desired subpackage use the following instructions:

For Computer Vision:

pip install sotaai[cv]

For Natural Language Processing:

pip install sotaai[nlp]

For Neurosymbolic Reasoning:

pip install sotaai[neuro]

For Reinforcement Learning:

pip install sotaai[rl]

For Reinforcement Learning it is also necessary to install some dependencies independently:

pip install pybullet-gym@git+https://github.com/benelot/pybullet-gym

Important: if you make use of zsh you have to add a \ before each bracket ([]) in order for zsh to understand the command e.g. pip install sotaai\[cv\].

Installation Issues

Tensorflow SSL: CERTIFICATE_VERIFY_FAILED

This is a known issue in Tensorlow. It is documented to occur in MacOSX. As of now, a working solution is to manually install certificates by running the Install Certificates.command that exists inside the python installation directory:

Applications/Python X.Y/Install Certificates.command

Replace Python X.Y with your installed version e.g. Python 3.8. Take into account that the location of this file might depend on your Python and/or OS version.