DL: Add custom loss function support for DL module

This commit adds support for custom loss function to DL fit, evaluate,
fit_multiple. Following are the changes:

1. fit: A new optional param for passing in the object_table name
object_table (optional) VARCHAR: Name of the table containing Python
objects in the case that custom loss functions or custom metrics are
specified in the parameter `compile_params`
```
madlib_keras_fit(
    source_table,
    model,
    model_arch_table,
    model_id,
    compile_params,
    fit_params,
    num_iterations,
    use_gpus,
    validation_table,
    metrics_compute_frequency,
    warm_start,
    name,
    description,
    object_table  -- new parameter
    )
```

This new param is also outputed in the output summary table.

2. Adding helper functions to parse custom loss functions, query their
definitions from the object_table and create an object of a dictionary
of {'fn_name': 'fn_object'} to be passed to the fit functions, where it
is read and passed as a python object to keras.

3. Evaluate: No change to the madlib_keras_evaluate() function. Reads object_table
information from the fit/fit_multiple output model table.
Output table adds a new column:
loss_type: Type of loss used that was used in the training step
	   If a custom loss or metric is used, we should give the name of it.
	   Otherwise list the built-in one used

4. fit_multiple: No change to the madlib_keras_fit_multiple_model() function. Reads
object_table information from the model_selection table.  The mst keys
having  are populated by None object_map by default. If the object_table
exists, the helper function to parse custom loss functions from the
compile_params is called to get all the custom_function names. Once we
get all the custom function names, we query their definitions from the
object_table and create a single dictionary of {{'fn_name1':
'fn_object1'}, {'fn_name2': 'fn_object2'}...} and pass it to the fit
multiple functions, where it is read and the corresponding function
definition is passed as a python object to keras.

A summary table named <model>_summary is also created, which has the
following new columns:
model_selection_table: 	Name of the table containing model selection
			parameters to be tried.
object_table:  	        Name of the object table containing the serialized
			Python objects for custom loss functions and custom metrics (read from
			the mst_summary table).

5. Adding corresponding unit tests and dev-check tests
13 files changed
tree: 7e53538cbe39ca448b6ee324aecc7297343388e6
  1. .github/
  2. cmake/
  3. deploy/
  4. doc/
  5. examples/
  6. licenses/
  7. methods/
  8. src/
  9. tool/
  10. .gitignore
  11. CMakeLists.txt
  12. configure
  13. LICENSE
  14. NOTICE
  15. pom.xml
  16. README.md
  17. ReadMe_Build.txt
  18. RELEASE_NOTES
  19. Release_Review_HOWTO.txt
README.md

MADlib® is an open-source library for scalable in-database analytics. It provides data-parallel implementations of mathematical, statistical and machine learning methods for structured and unstructured data.

master build status

Installation and Contribution

See the project website MADlib Home for links to the latest binary and source packages.

We appreciate all forms of project contributions to MADlib including bug reports, providing help to new users, documentation, or code patches. Please refer to Contribution Guidelines for instructions.

For more installation and contribution guides, please refer to the MADlib Wiki.

Compiling from source on Linux details are also on the wiki.

Development with Docker

We provide a Docker image with necessary dependencies required to compile and test MADlib on PostgreSQL 10.5. You can view the dependency Docker file at ./tool/docker/base/Dockerfile_ubuntu16_postgres10. The image is hosted on Docker Hub at madlib/postgres_10:latest. Later we will provide a similar Docker image for Greenplum Database.

We provide a script to quickly run this docker image at ./tool/docker_start.sh, which will mount your local madlib directory, build MADlib and run install check on this Docker image. At the end, it will docker exec as postgres user. Note that you have to run this script from inside your madlib directory, and you can specify your docker CONTAINER_NAME (default is madlib) and IMAGE_TAG (default is latest). Here is an example:

CONTAINER_NAME=my_madlib IMAGE_TAG=LaTex ./tool/docker_start.sh

Notice that this script only needs to be run once. After that, you will have a local docker container with CONTAINER_NAME running. To get access to the container, run the following command and you can keep working on it.

docker exec -it CONTAINER_NAME bash

To kill this docker container, run:

docker kill CONTAINER_NAME
docker rm CONTAINER_NAME

You can also manually run those commands to do the same thing:

## 1) Pull down the `madlib/postgres_10:latest` image from docker hub:
docker pull madlib/postgres_10:latest

## 2) Launch a container corresponding to the MADlib image, name it
##    madlib, mounting the source code folder to the container:
docker run -d -it --name madlib \
    -v (path to madlib directory):/madlib/ madlib/postgres_10
# where madlib is the directory where the MADlib source code resides.

################################# * WARNING * #################################
# Please be aware that when mounting a volume as shown above, any changes you
# make in the "madlib" folder inside the Docker container will be
# reflected on your local disk (and vice versa). This means that deleting data
# in the mounted volume from a Docker container will delete the data from your
# local disk also.
###############################################################################

## 3) When the container is up, connect to it and build MADlib:
docker exec -it madlib bash
mkdir /madlib/build_docker
cd /madlib/build_docker
cmake ..
make
make doc
make install

## 4) Install MADlib:
src/bin/madpack -p postgres -c postgres/postgres@localhost:5432/postgres install

## 5) Several other commands can now be run, such as:
# Run install check, on all modules:
src/bin/madpack -p postgres -c postgres/postgres@localhost:5432/postgres install-check
# Run install check, on a specific module, say svm:
src/bin/madpack -p postgres -c postgres/postgres@localhost:5432/postgres install-check -t svm
# Reinstall MADlib:
src/bin/madpack -p postgres -c postgres/postgres@localhost:5432/postgres reinstall

## 6) Kill and remove containers (after exiting the container):
docker kill madlib
docker rm madlib

Instruction for building design pdf on Docker:

For users who wants to build design pdf, make sure you use the IMAGE_TAG=LaTex parameter when running the script. After launching your docker container, run the following to get design.pdf:

cd /madlib/build_docker
make design_pdf
cd doc/design

Detailed build instructions are available in ReadMe_Build.txt

User and Developer Documentation

The latest documentation of MADlib modules can be found at MADlib Docs.

Architecture

The following block-diagram gives a high-level overview of MADlib's architecture.

MADlib Architecture

Third Party Components

MADlib incorporates software from the following third-party components. Bundled with source code:

  1. libstemmer “small string processing language”
  2. m_widen_init “allows compilation with recent versions of gcc with runtime dependencies from earlier versions of libstdc++”
  3. argparse 1.2.1 “provides an easy, declarative interface for creating command line tools”
  4. PyYAML 3.10 “YAML parser and emitter for Python”
  5. UseLATEX.cmake “CMAKE commands to use the LaTeX compiler”

Downloaded at build time (or supplied as build dependencies):

  1. Boost 1.61.0 (or newer) “provides peer-reviewed portable C++ source libraries”
  2. PyXB 1.2.6 “Python library for XML Schema Bindings”
  3. Eigen 3.2.2 “C++ template library for linear algebra”

Licensing

Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this project to You under the Apache License, Version 2.0 (the “License”); you may not use this project except in compliance with the License. You may obtain a copy of the License at LICENSE.

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an “AS IS” BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.

As specified in LICENSE additional license information regarding included third-party libraries can be found inside the licenses directory.

Release Notes

Changes between MADlib versions are described in the ReleaseNotes.txt file.

Papers and Talks

Related Software

  • PivotalR - PivotalR also lets the user run the functions of the open-source big-data machine learning package MADlib directly from R.
  • PyMADlib - PyMADlib is a python wrapper for MADlib, which brings you the power and flexibility of python with the number crunching power of MADlib.