Distributed TensorFlow training (Google I/O '18)

แชร์
ฝัง

ความคิดเห็น • 24

  • @junkmail75034
    @junkmail75034 4 ปีที่แล้ว +4

    I can't believe it. It's now 2020 and this presentation is still relevant. This is now a classic.

  • @RahulHuilgol
    @RahulHuilgol 6 ปีที่แล้ว +12

    The whole talk is about single node training and the title of distributed training is misleading. It barely mentions distributed at the end. Also only mentions async parameter server approach.

  • @sanyamjaincs1
    @sanyamjaincs1 3 ปีที่แล้ว

    the one which they are calling as Sync All reduce architecture can also be called as swarm ?

  • @pranavs430
    @pranavs430 6 ปีที่แล้ว

    At 15:43 you are saying that adding the distribution strategy to the runconfig is sufficient to run gpus as well. But when I run this with tensorflow-gpu version 1.10, I am getting the error that optimizer.minimize uses apply_gradient() which doesnt work and that I should use _distributed_apply(). Is this because I am not using the proper versions or something?

  • @PunitaOjha01
    @PunitaOjha01 3 ปีที่แล้ว

    Can I find the ppt somewhere, it will be handy for a quick revision

  • @balarampanda7401
    @balarampanda7401 6 ปีที่แล้ว +1

    What is the difference between Distributed Tensorflow with GPU and TPU ? Can Tensorflow with GPU able to produce same performance as TPU ?

  • @jinojossy93
    @jinojossy93 6 ปีที่แล้ว

    That's cool. Definitely gonna try that. Would you please point too some tutorial which will help me to try out using my CPU/GPU for building ML models?

  • @AliciaMoralesCarrasco
    @AliciaMoralesCarrasco 6 ปีที่แล้ว

    I'm trying to build this GoogleCloud Compute Engine but no way to get 8 NVIDIA V100 nor 4 NVIDIA 100P :(

    • @biologicalstatistics3320
      @biologicalstatistics3320 4 ปีที่แล้ว

      you can try a cluster of single board computers with cuda cores like the Jetson Nano.

  • @Donaldo
    @Donaldo 6 ปีที่แล้ว

    At 22:28 she says the training is more effective with more GPUs, but the graph is in step mode. Why are more GPUs more effective per step? I would have only expected it to be more effective per unit time.

    • @donm7906
      @donm7906 6 ปีที่แล้ว

      my thoughts: more GPUs -> bigger batch size -> learning from more data per step -> more effective training. ( Although in practice bigger batch doesn't always yield better performance, but in this case it seems like so)

  • @iulianlg
    @iulianlg 6 ปีที่แล้ว +2

    Perfect! Big Thanks for this presentation

  • @Thegreatindiaexpedition
    @Thegreatindiaexpedition 6 ปีที่แล้ว +2

    This one was amazing

  • @3pmahadevgroups26
    @3pmahadevgroups26 3 ปีที่แล้ว +1

    I am the reason why you have to be successful

  • @pradeep422
    @pradeep422 6 ปีที่แล้ว +1

    lol why all those empty seats

    • @MohdAkmalZakiIO
      @MohdAkmalZakiIO 6 ปีที่แล้ว +1

      Because the main things that make the developers come to IO is about app and web development. Not everyone really into AI stuff...

  • @sandipansarkar9211
    @sandipansarkar9211 3 ปีที่แล้ว

    finished watching

  • @salims3660
    @salims3660 6 ปีที่แล้ว

    Tesla P100D is GPU really

  • @ProfessionalTycoons
    @ProfessionalTycoons 6 ปีที่แล้ว

    great talk really amazing

  • @hemalmamtora3676
    @hemalmamtora3676 5 ปีที่แล้ว

    Thanks !

  • @rahulsaha4439
    @rahulsaha4439 6 ปีที่แล้ว

    That’s cool 😎

  • @kimchi_taco
    @kimchi_taco 6 ปีที่แล้ว

    nice talk

  • @thebrokeartist1272
    @thebrokeartist1272 2 ปีที่แล้ว +1

    Priys Mam is an IIT graduate with gold medal and was also second rank in IIT JEE exam, one of the toughest exams in the world. She is such a legend ❤️❤️❤️❤️💓💓