Design
Visually design and train machine learning (ML) models in an intuitive studio without writing a line of code.
Deploy
Deploy any ML model to production using fully managed infrastructure and create secure and scalable inference endpoints for predictions.
Explore
Explore datasets and POC models using Jupyter Notebooks and a dedicated JupyterHub interface.
Scale
Automate the life-cycle management of your train and test datasets and your models for explainable and reproductible AI at scale.
Who is this for?
Data Scientists
Explore in notebooks then design and train any machine learning model using the visual studio and managed GPU resources of ML pipelines.
ML Engineers
Deploy models trained by your teammates either on or out of ForePaaS as simple prediction endpoints.
Machine Learning Manager features
Do your initial discovery and exploration with notebooks
ForePaaS Machine Learning Manager comes with a JupyterHub and Jupyter Notebooks integration to allow you to explore datasets and start creating models. Get started with an pre-made image loaded with the most popular ML packages. Leverage built-in integrations to connect with other components like Buckets, Data Manager datasets, Data Processing Engine and the ML orchestrator. Export your notebooks in a few clicks once you are ready to move to production.
Create and design machine learning models
Machine Learning Manager offers a complete ML training studio to visually design, create and deploy AI models from scratch in pipelines, a point-and-click design UI. Pipelines unify the typical ML framework in one single interface: set up your train and test datasets and how they evolve in time, choose your validation framework, define your estimator, fine-tune hyper-parameters, train and validate models, and select the best one for deployment. A rich catalog of off-the-shelf estimators enables you to get started with the most popular estimators from the scikit-learn ecosystem: random forests, SVM, etc — while data scientists can leverage a full IDE to write and run their own code using any of the supported frameworks: Keras, PyTorch, scikit-learn. Use a full hyper-parameter studio to optimize for parameters with a grid search.
Train models using a fully managed and automated infrastructure
Use managed ForePaaS infrastructure to run your code and pipelines and train models. Use general purpose CPU for smaller models or supercharge your trainings with managed GPU resources. Monitor pipeline job executions, and troubleshoot in real-time using metrics and logs. Schedule all or part of a pipeline for automation with triggers: you can decouple the dataset preprocessing and train-test split, from the actual model training and validation, its scoring, and its deployment. Handle the life-cycle of models yourself, or use Machine Learning Manager managed dataset generation feature which handles the life-cycle to ensure models are never tested on train data. Automate your model deployment to production with deployment criteria.
Easily deploy and invoke any model on dedicated infrastructure
Deploy in one click any model manually or automatically as a model API used for inference. The infrastructure, CPU or GPU, is fully managed by the platform, ensuring total security as your number of predictions scales. Activate internal and external model consumers : write batch predictions on large datasets by consuming the model from a ForePaaS Data Processing Engine action, or make near-real-time inferencing using external endpoints from a ForePaaS API Gateway. Endpoints allow you to expose your model to the Internet, making predictions by calling the endpoint with your payload to predict. Authentication and authorization is managed by the ForePaaS IAM.
Bring-your-own ML models to production
ForePaaS Machine Learning Manager supports externally trained models for deployment too. Use existing models from industry-standard frameworks or upload your own model files to publish your work in production in a few clicks. ForePaaS Machine Learning Manager offers a versioned MLops environment — empowering data scientists to deploy their own work and get value from their models without managing infrastructure. Choose among popular frameworks, including Dataiku, Keras and others, to set up your production-grade environments with all necessary system requirements and code libraries.
53%
of data science projects never get fully deployed
Source: Gartner Research 2018-2019
36%
of data scientists report dirty data as their main challenge
Source: Kaggle 2017 State of Data Science
4%
of a data scientist’s time actually spent on refining algorithms
Source: Crowdflower 2016 Data Science Report