Destructive Deep Learning (ddl) README

Destructive deep learning estimators and functions. Estimators are compatible with scikit-learn. Source code is distributed under the BSD 3-clause license.

Please cite the following paper if you use this code:

Deep Density Destructors (pdf)

David I. Inouye, Pradeep Ravikumar

In International Conference on Machine Learning (ICML), 2018.

Documentation

UPDATED: Please see the updated documentation for an API reference and tutorials/demos including a Quickstart tutorial and MNIST demo.

Environment Setup

Environment setup instructions for:

1. Docker or Singularity containers (recommended)
2. Linux Setup (Ubuntu)
3. Mac OSX (unsupported)

1. Docker or Singularity Setup

Because MLPACK is required for the tree density destructors used in the experiments, the suggested installation method is to download and start a shell in a Docker or Singularity container as below. (If you are using Docker for Mac or Docker for Windows, you will probably have to increase the available memory to Docker for these experiments. See Docker documentation.) For Docker (recommended if available):

docker run -it davidinouye/destructive-deep-learning /bin/bash

Or, for Singularity:

singularity shell -s /bin/bash shub://davidinouye/destructive-deep-learning

2. Linux Setup (Ubuntu)

Install build essentials and cmake (needed for building mlpack destructors), laplack and blas (for fast linear operations), boost and armadillo libraries (required to build mlpack).

apt-get update && apt-get install \
    build-essential \
    cmake \
    liblapack-dev \
    libblas-dev \
    libboost-math-dev \
    libboost-program-options-dev \
    libboost-test-dev \
    libboost-serialization-dev \
    libarmadillo-dev

3. Mac OSX (unsupported)

Install homebrew as per homebrew documentation (note that this also install the required xcode tools). Then, install cmake, armadillo and boost (required to build mlpack) and llvm (required for openmp support for mlpack):

brew update && brew install \
    cmake \
    armadillo \
    boost \
    llvm

Installation

Once your environment is setup via one of the methods described above, download and compile the code to link to MLPACK. The first pip install is for scikit-learn, cython is required to compile pot and ddl, and pot and nose are used in ddl tests.

pip install numpy scipy scikit-learn
pip install setuptools Cython
git clone https://github.com/davidinouye/destructive-deep-learning.git
cd destructive-deep-learning
pip install .[test]

To run tests (which uses pytest), execute:

make test

Reproduce experiments from ICML 2018 paper

NOTE: MLPACK is required to reproduce experiments, please see installation instructions.

To reproduce the 2D experiment in the paper and generate the paper figures open and run the notebook notebooks/demo_toy_experiment.ipynb or run the notebook from the command line. Note that this notebook may take a while to run. Also, if the command below is interrupted with Ctrl+C, the underlying python process may need to be killed manually.

jupyter nbconvert --ExecutePreprocessor.timeout=-1 --to notebook --execute notebooks/demo_toy_experiment.ipynb

To reproduce the MNIST and CIFAR-10 experiments execute the command below. Note that this script will download the MNIST and CIFAR-10 datasets into data/download_cache if not downloaded already. The results are stored in data/results both the log files and pickle files that include the fitted models. Note that the log files will always append to the previous log file rather than overwriting the existing log file.

# Download data cache directly since mldata.org is sometimes down
wget http://www.cs.cmu.edu/~dinouye/data/data-icml2018.tar.gz && tar -xzvf data-icml2018.tar.gz && rm data-icml2018.tar.gz

# Example command for deep copula model and MNIST data
python scripts/icml_2018_experiment.py --model_names=deep-copula --data_names=mnist

# View tail of output log files
tail data/results/data-mnist_model-deep-copula_n_jobs-1.log

# Command for all models and datasets (using commas to separate)
python scripts/icml_2018_experiment.py --model_names=deep-copula,image-pairs-copula,image-pairs-tree --data_names=mnist,cifar10

# Command to run all experiments in parallel using subprocesses
python scripts/icml_2018_experiment.py --model_names=deep-copula,image-pairs-copula,image-pairs-tree --data_names=mnist,cifar10 --parallel_subprocesses=True

Contributing

General coding guidelines

Please read through the following high-level guidelines:

1. Zen of Python - https://www.python.org/dev/peps/pep-0020/
2. Python style guidelines - https://www.python.org/dev/peps/pep-0008/
3. scikit-learn coding guidelines - http://scikit-learn.org/stable/developers/contributing.html#coding-guidelines

Project-specific guidelies

For this particular project, please follow these additional guidelines:

• Use lower case with underscores for variable names and functions.

• Please use longer names with full spellings especially for public interfaces to allow for super lightweight documentation. The variable names should be descriptive of its function. For example, a constructor name should be fitted_canonical_destructor rather than fitted_destructor or destructor or fit_canon_destr or fcd. Another example, univariate_estimators rather than univ_est or univariate_est or uest. It is much easier to change a long variable name to short one than the other way around.

• Methods should generally be private designated by underscore prefix unless sure the method should be exposed publicly.

• For non-negative integer count variables prefix with n_ rather than num_ or number_of_

• Use variable names n_samples, n_features, and n_components (number of mixture components, number of PCA vectors, etc) and n_layers instead of ambiguous single letter variable names like n, p or k.

• In the library and tests, please use the logging API instead of print statements. In particular, create a logger for each module and call the appropriate logging function (usually logger.debug(message))

  import logging
  logger = logging.getLogger(__name__)
  def foo():
       logger.debug('Checking inside foo')

• To avoid the module from outputing anything unless requested, the root module file __init__.py redirects the logging output to None as follows:

  import logging
  from logging import NullHandler
  logging.getLogger(__name__).addHandler(NullHandler())

• Thus, to view these logs when executing a program and capture warnings as logs for a particular module you must setup logging to output to standard out (and/or a local file). For example, you could write:

  logging.basicConfig(stream=sys.stdout)  # Push towards stdout instead of null handler
  logging.captureWarnings(True)  # Capture warnings in loggers
  logging.getLogger('ddl').setLevel(logging.DEBUG)  # Show everything above DEBUG level for the root ddl module