Tensorflow Dev Summit Recap

Keynote:

  • TF Goal:
    • to introduce ML to everyone
  • TF feature:
    • Scalable
    • Performance
  • Widely usage for TF
  • Multiple device for TF
    • TPU
      • ASIC
      • 10x faster
  • TF for everyone
  • TF 1.0
    • Feature:
      • Fast
      • Flexible
      • Product-Ready
    • More ML
      • K-means
      • SVM
      • Random Forest

XLA

  • What is XLA (Accelerated Linear Algebra)
    • JIT Complication
    • JIT
      • Program built at runtime
      • Low-Overhead complication
    • TF-Level Block Diagram
  • Why excited about XLA
    • Server-side speedups
    • XLA’s JIT compilation and specialization
    • Model-shaped benchmark wins up to 60%
    • SyntaxNet from 200us —> 5 us
    • Mobile footprint reduction
    • Cross-compile for AMR, PPC, x86
    • LSTM model for mobile: 2.6 MB —> 600 KB (4x reduction)
    • XLA’s high-leve optimizer
  • Caveats:
    • Not all Ops compile
    • Not everything is faster
  • JIT :
    • Improvement lot in GPU
    • But still WIP in CPU (slower when use CPU)

TensorFlow High-Level API

  • Original Tensorflow:
    • Flexible,
    • Extensible,
    • Maintainable
    • No Out-of-the-bix algorithm
  • Fast Iteration
    • Estimator could train, fit, predict for models.
  • Encodes best practices
  • Deploy with Tensorflow Serving
  • Distribution
  • High Level API
    • Layer
    • Estimator (1.1)
    • Canned Estimator (1.2)
  • Keras
    • tf.keras (1.2) tf.contrib.keras(1.1)
    • tensorflow.layer and keras.layer is the same
    • run keras on tensorflow help keras user
      • Use distribution training
      • Cloud ML
      • Tensor Serving
  • Wide & Deep Learning

Lightening Talks:

  • Rayan Z
    • How to teach machine learning to non-tech people.
    • GDG
    • Women in Tech
  • Sin c
    • How your data could be trusted.
  • Alex B
  • Donghyun Kwak
    • Policy Learning in Sparkse Reward Home Simulation with Introspec
    • Home Robot
    • 26% improvement
  • Tomoyuki Chikanaga:
    • Magellan block
      • Make GCP service as building BLOCKS
    • Inspection of cloud machine learning hyper parameter tuning
  • Sprawit Saengkyongam (James)
  • Sung Kim
    • 70% code are redundant
    • Ongoing:
      • iOS to Android
      • Auto determine copy homework
  • Jeongkyu Shin
  • Luke (freelancer in Australia)
  • Masahiko Adachi (GDE)
    • Robot with Neural Controller
  • Norihiro Shimoda
    • TFUG (300 member per meetup)
  • Mithuhisa Ohta (deep learning team leader)
    • Image classification by car
    • Obejct Detection
    • Anomaly Detection
    • Robot pickup staff by speaking
  • Karthik Muthuswamy(GCPUG)
    • Object Detection
      • Challenge: Need enough resource
  • Thia Kai Xin
    • Data Scientist SG
    • Big data SG
    • Current situation:
      • Lots of requirement (Data Scientsit) in SG, but no supply.
      • Student could not meet market
  • Andrew Stevens (CTO and Architect : two company)
    • (Security Analysis) Anomalies in time series for RNN
    • Challenge:
      • Data –> Build Data As A Service.
    • Passonate about:
      • Kickoff tensorflow user group in SG and AUS
  • Yoshihiro Sugi
  • Martin Andrews
    • A usb flash drive for jupyter tensorflow 1.0
      • Also include notebook
        • CNN
        • RNN
        • Reinforcement learning
    • github
  • “Ta” ex-facebook data scientist
    • News Feed Tanking with Human-in-the-loop
    • Thai Programmer Association
  • Sujoy Roy (SAP)
    • SAP Clea: Lots of machine learning product (CV, Bot, IOT…)
  • Talha Obaid - Samantec Email Security
    • Semantec
      • Scikit Learn, Spark
    • PyData SG meetup
      • 2.1k members
  • Amit Kapoor (Teaching DS)
    • Teach ML
      • Provide learning path, enable sharing
    • Visualization by markdown (model-vis Approach)
  • Nichal & Raghotham (UnnatiData Labs)
    • PyData
    • Music Generation
  • Christin (Master Student in Malasia)
    • Round text detect (also other direction of text detection)

Evan

Attitude is everything