Phil 3.30.18

TF Dev Sumit

Highlights blog post from the TF product manager

Keynote

  • Connecterra tracking cows
  • Google is an AI – first company. All products are being influenced. TF is the dogfood that everyone is eating at google.

Rajat Monga

  • Last year has been focussed on making TF easy to use
  • 11 million downloads
  • blog.tensorflow.org
  • youtube.com/tensorflow
  • tensorflow.org/ub
  • tf.keras – full implementation.
  • Premade estimators
  • three line training from reading to model? What data formats?
  • Swift and tensorflow.js

Megan

  • Real-world data and time-to-accuracy
  • Fast version is the pretty version
  • TensorflowLite is 300% speedup in inference? Just on mobile(?)
  • Training speedup is about 300% – 400% anually
  • Cloud TPUs are available in V2. 180 TF computation
  • github.com/tensorflow/tpu
  • ResNet-50 on Cloud TPU in < 15

Jeff Dean

  • Grand Engineering challenges as a list of  ML goals
  • Engineer the tools for scientific discovery
  • AutoML – Hyperparameter tuning
  • Less expertise (What about data cleaning?)
    • Neural architecture search
    • Cloud Automl for computer vision (for now – more later)
  • Retinal data is being improved as the data labeling improves. The trained human trains the system proportionally
  • Completely new, novel scientific discoveries – machine scan explore horizons in different ways from humans
  • Single shot detector

Derrek Murray @mrry (tf.data)

  • Core TF team
  • tf.data  –
  • Fast, Flexible, and Easy to use
    • ETL for TF
    • tensorflow.org/performance/datasets_performance
    • Dataset tf.SparseTensor
    • Dataset.from_generator – generates graphs from numpy arrays
    • for batch in dataset: train_model(batch)
    • 1.8 will read in CSV
    • tf.contrib.data.make_batched_features_dataset
    • tf.contrib.data.make_csv_dataset()
    • Figures out types from column names

Alexandre Passos (Eager Execution)

  • Eager Execution
  • Automatic differentiation
  • Differentiation of graphs and code <- what does this mean?
  • Quick iterations without building graphs
  • Deep inspection of running models
  • Dynamic models with complex control flows
  • tf.enable_eager_execution()
  • immediately run the tf code that can then be conditional
  • w = tfe.variables([[1.0]])
  • tape to record actions, so it’s possible to evaluate a variety of approaches as functions
  • eager supports debugging!!!
  • And profilable…
  • Google collaboratory for Jupyter
  • Customizing gradient, clipping to keep from exploding, etc
  • tf variables are just python objects.
  • tfe.metrics
  • Object oriented savings of TF models Kind of like pickle, in that associated variables are saved as well
  • Supports component reuse?
  • Single GPU is competitive in speed
  • Interacting with graphs: Call into graphs Also call into eager from a graph
  • Use tf.keras.layers, tf.keras.Model, tf.contribs.summary, tfe.metrics, and object-based saving
  • Recursive RNNs work well in this
  • Live demo goo.gl/eRpP8j
  • getting started guide tensorflow.org/programmers_guide/eager
  • example models goo.gl/RTHJa5

Daniel Smilkov (@dsmilkov) Nikhl Thorat (@nsthorat)

  • In-Browser ML (No drivers, no installs)
  • Interactive
  • Browsers have access to sensors
  • Data stays on the client (preprocessing stage)
  • Allows inference and training entirely in the browser
  • Tensorflow.js
    • Author models directly in the browser
    • import pre-trained models for inference
    • re-train imported models (with private data)
    • Layers API, (Eager) Ops API
    • Can port keras or TF morel
  • Can continue to train a model that is downloaded from the website
  • This is really nice for accessibility
  • js.tensorflow.org
  • github.com/tensorflow/tfjs
  • Mailing list: goo.gl/drqpT5

Brennen Saeta

  • Performance optimization
  • Need to be able to increase performance exponentially to be able to train better
  • tf.data is the way to load data
  • Tensorboard profiling tools
  • Trace viewer within Tensorboard
  • Map functions seem to take a long time?
  • dataset.map(Parser_fn, num_parallel_calls = 64)) <- multithreading
  • Software pipelining
  • Distributed datasets are becoming critical. They will not fit on a single instance
  • Accelerators work in a variety of ways, so optimizing is hardware dependent For example, lower precision can be much faster
  • bfloat16 brain floating point format. Better for vanishing and exploding gradients
  • Systolic processors load the hardware matrix while it’s multiplying, since you start at the upper left corner…
  • Hardware is becoming harder and harder to do apples-to apples. You need to measure end-to-end on your own workloads. As a proxy, Stanford’s DAWNBench
  • Two frameworks XLA nd Graph

Mustafa Ispir (tf.estimator, high level modules for experiments and scaling)

  • estimators fill in the model, based on Google experiences
  • define as an ml problem
  • pre made estimators
  • reasonable defaults
  • feature columns – bucketing, embedding, etc
  • estimator = model_to_estimator
  • image = hum.image_embedding_column(…)
  • supports scaling
  • export to production
  • estimator.export_savemodel()
  • Feature columns (from csv, etc) intro, goo.gl/nMEPBy
  • Estimators documentation, custom estimators
  • Wide-n-deep (goo.gl/l1cL3N from 2017)
  • Estimators and Keras (goo.gl/ito9LE Effective TensorFlow for Non-Experts)

Igor Sapirkin

  • distributed tensorflow
  • estimator is TFs highest level of abstraction in the API google recommends using the highest level of abstraction you can be effective in
  • Justine debugging with Tensorflow Debugger
  • plugins are how you add features
  • embedding projector with interactive label editing

Sarah Sirajuddin, Andrew Selle (TensorFlow Lite) On-device ML

  • TF Lite interpreter is only 75 kilobytes!
  • Would be useful as a biometric anonymizer for trustworthy anonymous citizen journalism. Maybe even adversarial recognition
  • Introduction to TensorFlow Lite → https://goo.gl/8GsJVL
  • Take a look at this article “Using TensorFlow Lite on Android” → https://goo.gl/J1ZDqm

Vijay Vasudevan AutoML @spezzer

  • Theory lags practice in valuable discipline
  • Iteration using human input
  • Design your code to be tunable at all levels
  • Submit your idea to an idea bank

Ian Langmore

  • Nuclear Fusion
  • TF for math, not ML

Cory McLain

  • Genomics
  • Would this be useful for genetic algorithms as well?

Ed Wilder-James

  • Open source TF community
  • Developers mailing list developers@tensorflow.org
  • tensorflow.org/community
  • SIGs SIGBuild, other coming up
  • SIG Tensorboard <- this

Chris Lattner

  • Improved usability of TF
  • 2 approaches, Graph and Eager
  • Compiler analysis?
  • Swift language support as a better option than Python?
  • Richard Wei
  • Did not actually see the compilation process with error messages?

TensorFlow Hub Andrew Gasparovic and Jeremiah Harmsen

  • Version control for ML
  • Reusable module within the hub. Less than a model, but shareable
  • Retrainable and backpropagateable
  • Re-use the architecture and trained weights (And save, many, many, many hours in training)
  • tensorflow.org/hub
  • module = hub.Module(…., trainable = true)
  • Pretrained and ready to use for classification
  • Packages the graph and the data
  • Universal Sentence Encodings semantic similarity, etc. Very little training data
  • Lower the learning rate so that you don’t ruin the existing rates
  • tfhub.dev
  • modules are immutable
  • Colab notebooks
  • use #tfhub when modules are completed
  • Try out the end-to-end example on GitHub → https://goo.gl/4DBvX7

TF Extensions Clemens Mewald and Raz Mathias

  • TFX is developed to support lifecycle from data gathering to production
  • Transform: Develop training model and serving model during development
  • Model takes a raw data model as the request. The transform is being done in the graph
  • RESTful API
  • Model Analysis:
  • ml-fairness.com – ROC curve for every group of users
  • github.com/tensorflow/transform

Project Magenta (Sherol Chen)

People:

  • Suharsh Sivakumar – Google
  • Billy Lamberta (documentation?) Google
  • Ashay Agrawal Google
  • Rajesh Anantharaman Cray
  • Amanda Casari Concur Labs
  • Gary Engler Elemental Path
  • Keith J Bennett (bennett@bennettresearchtech.com – ask about rover decision transcripts)
  • Sandeep N. Gupta (sandeepngupta@google.com – ask about integration of latent variables into TF usage as a way of understanding the space better)
  • Charlie Costello (charlie.costello@cloudminds.com – human robot interaction communities)
  • Kevin A. Shaw (kevin@algoint.com data from elderly to infer condition)

 

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google+ photo

You are commenting using your Google+ account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s

This site uses Akismet to reduce spam. Learn how your comment data is processed.