OptML.md

OptML.md

Optimization frameworks for machine learning

Low-level features (e.g. optimized for something special)

• https://www.tensorflow.org/lite

  • Small version of TF, targets mobile devices on Android and IoT

• https://www.tensorflow.org/performance/xla/

  • XLA (Accelerated Linear Algebra) is a domain-specific compiler for linear algebra that optimizes TensorFlow computations.
  • https://haosdent.gitbooks.io/tensorflow-document/content/resources/xla_prerelease.html
  • Has two modes: JIT and AOT. JIT is more like TFlite. AOT resembles TVM.

• http://ngraph.nervanasys.com/docs/latest/

  • nGraph (Intel, make business with PlaidML)

• https://software.intel.com/en-us/openvino-toolkit

  • OpenVINO (Intel).

• https://github.com/xiaomi/mace

  • 2018, MACE is a deep learning inference framework optimized for mobile heterogeneous computing platforms

• https://github.com/Tencent/PocketFlow

  • PocketFlow, a kind of Autotuner for TF?

• https://github.com/Tiramisu-Compiler/tiramisu

  • Tiramisu, Polyhedral compiler
  • Paper https://arxiv.org/pdf/1804.10694.pdf
    • 2018, Tiramisu: A Code Optimization Framework for High Performance Systems
  • Not mentioned:
    • Autodiff
    • Quantization
    • Python API

• http://eyeriss.mit.edu/

  • CNN enegry-efficient optimizations, by MIT

• https://github.com/Tencent/ncnn

  • Something for fast inference on ARMs. Often used as a baseline for internal Huawei projects.

• https://github.com/snipsco/tract

  • NLP-optimized (?) inference framework for IoT
  • Well, SNIPS is about NLP, but Track targets MobileNet. Strange.
  • Claims that is optimized for streaming.

High-level features

• https://github.com/google/jax

  • Another project of Google
  • Has Autodiff through numpy, AFAIK, PyTorch-style.
  • Backed by XLA.

• https://tvm.ai

  • TVM itself. main feature is the Tensor-Expression language (with both auto- and manual-schedulings). Supports many frontends and backends, from Web(sic!) to GPUs to FPGAs and IoT. Level of support varies.
  • Has Relay for static typing and high-level frontend.
  • Has muTvm for IoT backends and bare-metal systems.
  • Autodiff is in PRs.

• http://halide-lang.org/

  • Halide is a source of inspiration for TVM. AFAIK, invented Tensor-Expression language for image processing.
  • Has autodiff since recently (merged or not?)
  • Initially was not intended for machine learning, but who knows..

• https://github.com/plaidml/plaidml

  • PlaidML - "PlaidML is the easiest, fastest way to learn and deploy deep learning on any device, especially those running macOS or Windows"
  • Main feature is AutoTuning (Search of optimizations)
  • (RIP?) Obtained by Intel

• http://dlvm.org/

  • Paper https://arxiv.org/pdf/1711.03016 DLVM: A modern compiler infrastructure for deep learning systems (autodiff)
  • RIP in favor of TF-Swift

• https://github.com/vgvassilev/clad

  • Posters https://llvm.org/devmtg/2013-11/slides/Vassilev-Poster.pdf clad - Automatic Differentiation using Clang
  • RIP?

• https://github.com/facebookresearch/TensorComprehensions

  • A domain specific language to express machine learning workloads.
  • https://arxiv.org/abs/1802.04730 Tensor Comprehensions: Framework-Agnostic High-Performance Machine Learning Abstractions
  • RIP?
  • Integrated into PyTorch
  • Like TVM, has Tensor-Expression language (with auto-schedulers)