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:
UPDATED: Please see the updated documentation for an API reference and tutorials/demos including a Quickstart tutorial and MNIST demo.
Environment setup instructions for:
- Docker or Singularity containers (recommended)
- Linux Setup (Ubuntu)
- 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)
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
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
cython is required to compile
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:
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
General coding guidelines
Please read through the following high-level guidelines:
- Zen of Python - https://www.python.org/dev/peps/pep-0020/
- Python style guidelines - https://www.python.org/dev/peps/pep-0008/
scikit-learncoding guidelines - http://scikit-learn.org/stable/developers/contributing.html#coding-guidelines
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
fcd. Another example,
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
Use variable names
n_components(number of mixture components, number of PCA vectors, etc) and
n_layersinstead of ambiguous single letter variable names like
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
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__.pyredirects the logging output to
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