Marat Dukhan

@MaratDukhan

Performance optimization ninja @ Google Opinions are my own

SF bay area, CA
Vrijeme pridruživanja: listopad 2013.

Tweetovi

Blokirali ste korisnika/cu @MaratDukhan

Jeste li sigurni da želite vidjeti te tweetove? Time nećete deblokirati korisnika/cu @MaratDukhan

  1. proslijedio/la je Tweet
    30. sij

    Pay attention to this. We’re seeing up to 3X speed ups for real world models in the tfjs WASM backend with SIMD128 enabled!

    Poništi
  2. 28. sij

    New goodies from colleagues at ! All effects - edge detection, face detection, hair segmentation, and hand tracking - now can run inside a Web browser, powered by XNNPACK and

    Poništi
  3. proslijedio/la je Tweet
    6. sij

    Really excited about this model, which runs in real-time on Pixel and iPhone on our WebGL and WASM backends! Really nice work and !

    Poništi
  4. 23. pro 2019.

    TensorFlow.js on CPU now faster with an XNNPACK-powered backend! Whopping 4-20x over previous TF.js CPU backend in pure JavaScript🚀, near-universal coverage, and Node.js compatibility - available right now in the Alpha release

    Poništi
  5. 27. stu 2019.

    The micro-kernels for sparse inference on ARM64 and WebAssembly are already open-sourced in XNNPACK [], and so are pre-trained sparse models [] [4/4]

    Prikaži ovu nit
    Poništi
  6. 27. stu 2019.

    Our recent work [] with colleagues from and demonstrates that with a right layout and optimizations sparse inference delivers practical and non-negligible speedups of 1.3X-2.4X on a range of MobileNet and EfficientNet models. [3/4]

    Prikaži ovu nit
    Poništi
  7. 27. stu 2019.

    Computations in sparsified models involve many multiplications by zeroes, which can be skipped in theory, but common wisdom suggested that it is impractical in software inference implementations. [2/4]

    Prikaži ovu nit
    Poništi
  8. 27. stu 2019.

    Sparsification, or pruning of weights, in convolutional neural networks has a long history as a compression technique, and good support in deep learning frameworks, e.g. Model Optimization Toolkit in TensorFlow. [1/4]

    Prikaži ovu nit
    Poništi
  9. proslijedio/la je Tweet
    26. stu 2019.

    “Fast Sparse ConvNets”, a collaboration w/ [], implements fast Sparse Matrix-Matrix Multiplication to replace dense 1x1 convolutions in MobileNet architectures. The sparse networks are 66% the size and 1.5-2x faster than their dense equivalents.

    Prikaži ovu nit
    Poništi
  10. 13. stu 2019.

    Live in-browser Hand Tracking + Landmark Detection demos:

    Prikaži ovu nit
    Poništi
  11. 13. stu 2019.

    A preview of in-browser machine learning-based demos by our group in a great talk by and Powered By XNNPACK, MediaPipe, and (+SIMD)

    Prikaži ovu nit
    Poništi
  12. 9. lis 2019.

    Third-generation NNPACK-family library is here, at ! This time the focus is on accelerating FP32 models in NHWC layout, and it supports both mobile and Web platforms.

    Poništi
  13. 29. lis 2018.

    Today Facebook publicly released QNNPACK, open source library for low-precision neural network computations on mobile. Caffe2+QNNPACK = 2x speedup over TFLite + support for grouped conv (CondenseNet, ShuffleNet, RexNeXt).

    Poništi
  14. proslijedio/la je Tweet
    5. tra 2018.

    We've just released our new tutorial on how to interface it with Numpy and use it for in . Go and check it out!

    Prikaži ovu nit
    Poništi
  15. proslijedio/la je Tweet
    5. sij 2018.

    The security flaw suddenly casts 's and 's 2015 idea of avoiding branches altogether in a new light:

    Poništi
  16. 14. svi 2017.

    demo, that won the 1st place on the AI Hackathon in Minsk, is now live on

    Poništi
  17. proslijedio/la je Tweet
    19. tra 2017.

    Andrew Tulloch and Yangqing Jia of Facebook give a shout out to and NNPACK at

    Poništi
  18. proslijedio/la je Tweet
    17. tra 2017.

    of demos NNPACK at the launch event

    Poništi
  19. 13. ožu 2017.
    Odgovor korisniku/ci

    A 1x1 layer computes a linear combination of per-channel images. But if the layer has 2x2 stride, it wastes 75% of computed pixels. (2/2)

    Poništi
  20. 13. ožu 2017.

    Tonight I learnt: there are 1x1 convolutional layers with non-unit stride, e.g. in ResNet models. They make no sense, yet they exist. (1/2)

    Poništi

Čini se da učitavanje traje već neko vrijeme.

Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.

    Možda bi vam se svidjelo i ovo:

    ·