The Silicon Trend Tech Bulletin

Icon Collap
Home / AI / ML / ML Libraries Are Open-Sourced by Facebook

ML Libraries Are Open-Sourced by Facebook

Published Mon, May 10 2021 16:28 pm
by The Silicon Trend

markImage Source:


Tech giants such as Microsoft, Google & Facebook have been domination deep learning frameworks & tools that AI researchers globally use. Many of their open-sourced libraries are now gaining popularity on GitHub, helping budding AI creators across the globe build scalable & flexible ML models. From autonomous vehicles, a conversational chatbot to the recommendation & weather forecast systems, AI creators are testing with numerous hyperparameters, neural network architecture & other facets to suit the hardware constraints of edge platforms.

Some popular deep learning frameworks are Facebook's Caffe2, Torchcraft AI, Hydra, PyTorch, Google's TensorFlow, etc. According to Statista, AI deal operations global revenue is expected to reach $10.8Bn by 2023, & the natural language processing (NLP) market size is estimated to touch $43.3Bn by 2025.

Artificial Intelligence-1

Latest Open-source Libraries, Architecture & Tools by Facebook

  • Opacus: An open-source high-speed library to train PyTorch models with differential privacy (DP). It is claimed to be more scalable than current methods, supporting minimal code changes & has little effect on training performance. It helps researchers to track the privacy budget expended at any given time. 
  •  PyTorch: Widely used as a deep learning framework, besides Hydra & Caffe2, which aids researchers to build flexible ML models. It offers a Python package for top-level facets like NumPy (tensor computation) with robust TorchScript & GPU acceleration for a seamless transition between graph & eager mode. Its latest release offers - distributed training, mobile deployment, graph-based execution, etc.


You may read: Europe Solicits to Control the Use of AI in Society


  • Flashlight: An open-source ML library that allows users to execute ML/AI applications using C++ API. It doesn't need external bindings or figures to perform - memory mapping, threading, or interoperating with low-level hardware, making code integration direct, fast & straightforward.
  • Detection: An open-source software architecture that executes object detection algorithms like Mask R-CNN. The software has been created on Python & powered by the Caffe2. It has enabled numerous research work at Facebook such as - Mask R-CNN, data distillation: towards Omni-supervised learning, DensePose: dense human pose estimation in the wild, non-local neural networks & others.
  •  PyTorch3D: A highly optimized & modular library offering reusable, efficient components for 3D computer vision (CV) research with the PyTorch. It is designed to smoothly integrate with deep learning methods for manipulating & predicting 3D data.


You may read: Europe Solicits to Control the Use of AI in Society


  • Detectron2: A next-generation library that offers segmentation & detection algorithms. It is a hybrid of Mask C-RNN & Detectron. It supports several CV applications & research works. It can be used on Faster R-CNN, TensorMask, RPN, RetinaNet.
  •  BoTorch: A library for Bayesian optimization built on PyTorch framework. Bayesian optimization is a sequence strategy design for machines that don't assume any functional forms. BoTorch quickly offers a modular & extensible interface for composing Bayesian optimization primitives - acquisition functions, optimizers, probabilistic models, etc. It also enables easy integration with convolutional architectures in PyTorch.
  • Prophet: An open-source architecture released by FB's core data science team. It's a method to forecast time series data based on an additive model & works best with time-series data with numerous seasons of historical data - patient health evolution metrics, weather records & economic indicators.
  •  Classy Vision: A novel end-to-end PyTorch-based framework for massive training of video & image classification models. Unlike other CV libraries, this claims to provide flexibility for researchers. Most CV libraries lead to duplicate efforts & need users to transfer research between frameworks & re-learn the minutiae of efficient distributed training & data loading. In contrast, FB's PyTorch-based CV framework claimed to provide a better solution for training at scale & installing to production.
  • Tensor Comprehensions (TC): A fully functional C++ library that automatically synthesizes top-performance ML kernels using LLVM, Halide, ISL, or NVRTC. It can seamlessly integrate with PyTorch & Caffe2, designed to be highly portable & ML framework agnostic. It needs a simple tensor library with memory allocations, synchronization potentials & offloading.
  • FastText: An open-source library for efficient representation learning & text classification. It operates on generic & standard hardware. ML models can be further minimized on mobile devices as well.