Permalink
b698744 Jan 10, 2017
@arnocandel @thirdwing @Mistobaan @mmalohlava @surekhajadhwani
153 lines (109 sloc) 6.61 KB

Deep Water

What it is

  • Native implementation of Deep Learning models for GPU-optimized backends (MXNet, Caffe, TensorFlow, etc.)
  • State-of-the-art Deep Learning models trained from the H2O Platform
  • Train user-defined or pre-defined deeplearning models for image/text/H2OFrame classification from Flow, R, Python, Java, Scala or REST API
  • Behaves just like any other H2O model (Flow, cross-validation, early stopping, hyper-parameter search, etc.)
  • The next best thing after sliced bread
  • Under development

What it is not

  • An oil drilling platform

Python/R Jupyter Notebooks

Check out a sample of cool Deep Learning Jupyter notebooks!

PreRelease Downloads

For the following system dependencies, we provide recent builds for your convenience.

  • Ubuntu 16.04 LTS
  • Latest NVIDIA Display driver
  • CUDA 8 (latest available) in /usr/local/cuda
  • CUDNN 5.1 (placed inside of lib and include directories in /usr/local/cuda/)

In the future, we'll have more pre-built jars for more OS/CUDA combinations.

If you are interested in running H2O Deep Water on a different infrastructure, see the DIY build instructions below

PreRelease Amazon AWS Image

For your convenience, here's a pre-built image for Amazon's EC2 environment, based off our recent H2O Open Tour Hands-On Deep Water workshop (recording coming soon).

  • AMI ID: ami-10bd9607
  • AMI Name: deepwater-dallas-v4
  • Recommended instance types: g2.2xlarge or p2.xlarge
  • After launching the instance, you can connect to port 8888 (Jupyter Notebook) or port 54321 (H2O Flow).

Refer to the Deep Water Workshop on EC2 Amazon AMI document for additional information on how to run this AMI.

Roadmap, Architecture and Demo

Download the Deep Water overview slides.

architecture architecture architecture architecture

DIY Build Instructions:

If you want to use Deep Water in H2O-3, you'll need to have a .jar file that includes backend support for at least one of MXNet, Caffe or TensorFlow.

1. Build MXNet

Instructions to build MXNet

2. Build TensorFlow

Instructions to build TensorFlow

3. Build Caffe

Coming soon.

4. Build H2O Backend Connectors

From the top-level of the deepwater repository, do

./gradlew build -x test

This will create the following file: build/libs/deepwater-all.jar

5. Add DeepWater support to H2O-3

You need to check out the h2o-3. Copy the freshly created jar file build/libs/deepwater-all.jar from the previous step to H2O-3's library h2o-3/lib/deepwater-all.jar (create the directory if it's not there) and you're done!

Build H2O-3 as usual:
./gradlew build -x test

This H2O version will now have GPU Deep Learning support!

To use the GPU, please make sure to set your path to your CUDA installation:

export CUDA_PATH=/usr/local/cuda
Install the Python wheel:
sudo pip install h2o-3/h2o-py/dist/h2o-3.11.0.99999-py2.py3-none-any.whl
(Optional) Install the Python egg for MXNet

If you want to build your own MXNet models (from Python so far), install the MXNet wheel (which was built together with MXNet above):

sudo easy_install deepwater/thirdparty/mxnet/python/dist/mxnet-0.7.0-py2.7.egg

Running GPU enabled Deep Water in H2O

(Optional) Launch H2O by hand and build Deep Water models from Flow (localhost:54321)

java -jar h2o.jar

Java example use cases

Example Java GPU-enabled unit tests.

Python example use cases

Example Python GPU-enabled unit tests. Check out a sample of cool Deep Learning Python Jupyter notebooks!

R example use cases

Check out a sample of cool Deep Learning R Jupyter notebooks!

Scala / Sparkling Water example use cases

Coming soon.

Releasing

The release process bundles all defined submodules and push them into Maven central via Sonatype repository provider. The released artifacts are Java 6 compatible.

The release can be invoked for all modules by:

./gradlew -PdoRelease -PbuildOnlyBackendApi -PdoJava6Bytecode=true -Prelease.useAutomaticVersion=true release

The process performs the following steps:

  • update gradle.properties and remove SNAPSHOT and increase minor version (can be changed)
  • create a new release commit and tag it with release tag (see gradle/release.gradle file to override default template)
  • build
  • verification of compatibility of used API with Java 6 API
  • bytecode rewrite to be compatible with Java 6
  • generation of artifact metadata
  • push of artifacts into staging area at https://oss.sonatype.org/

The process needs to be finished manually by:

Note: The release process creates two new commits and a new tag with release version. However, the process does not push it to a remote repository and it is necessary to perform remote update manually by git push --tags or update gradle/release.gradle settings and remove --dry-run option from pushOptions field.