osprey 1.1.0

|Build Status| |Coverage Status| |PyPi version| [|License|] (http://www.apache.org/licenses/LICENSE-2.0) |DOI| [|Documentation|] (http://msmbuilder.org/osprey)

[] (http://www.apache.org/licenses/LICENSE-2.0) [] (http://msmbuilder.org/osprey)

<figcaption>

Logo

</figcaption>

Osprey is an easy-to-use tool for hyperparameter optimization for machine learning algorithms in python using scikit-learn (or using scikit-learn compatible APIs).

Each Osprey experiment combines an dataset, an estimator, a search space (and engine), cross validation and asynchronous serialization for distributed parallel optimization of model hyperparameters.

Documentation

For full documentation, please visit the Osprey homepage.

Installation

If you have an Anaconda Python distribution, installation is as easy as:

$ conda install -c omnia osprey

You can also install Osprey with pip:

$ pip install osprey

Alternatively, you can install directly from this GitHub repo:

$ git clone https://github.com/msmbuilder/osprey.git
$ cd osprey && git checkout 1.1.0
$ python setup.py install

Example using MSMBuilder

Below is an example of an osprey config file to cross validate Markov state models based on varying the number of clusters and dihedral angles used in a model:

estimator:
  eval_scope: msmbuilder
  eval: |
    Pipeline([
        ('featurizer', DihedralFeaturizer(types=['phi', 'psi'])),
        ('cluster', MiniBatchKMeans()),
        ('msm', MarkovStateModel(n_timescales=5, verbose=False)),
    ])

search_space:
  cluster__n_clusters:
    min: 10
    max: 100
    type: int
  featurizer__types:
    choices:
      - ['phi', 'psi']
      - ['phi', 'psi', 'chi1']
   type: enum

cv: 5

dataset_loader:
  name: mdtraj
  params:
    trajectories: ~/local/msmbuilder/Tutorial/XTC/*/*.xtc
    topology: ~/local/msmbuilder/Tutorial/native.pdb
    stride: 1

trials:
    uri: sqlite:///osprey-trials.db

Then run osprey worker. You can run multiple parallel instances of osprey worker simultaneously on a cluster too.

$ osprey worker config.yaml

...

----------------------------------------------------------------------
Beginning iteration                                              1 / 1
----------------------------------------------------------------------
History contains: 0 trials
Choosing next hyperparameters with random...
  {'cluster__n_clusters': 20, 'featurizer__types': ['phi', 'psi']}

Fitting 5 folds for each of 1 candidates, totalling 5 fits
[Parallel(n_jobs=1)]: Done   1 jobs       | elapsed:    0.3s
[Parallel(n_jobs=1)]: Done   5 out of   5 | elapsed:    1.8s finished
---------------------------------
Success! Model score = 4.080646
(best score so far   = 4.080646)
---------------------------------

1/1 models fit successfully.
time:         October 27, 2014 10:44 PM
elapsed:      4 seconds.
osprey worker exiting.

You can dump the database to JSON or CSV with osprey dump.

Dependencies

• python>=2.7.11

• six>=1.10.0

• pyyaml>=3.11

• numpy>=1.10.4

• scipy>=0.17.0

• scikit-learn>=0.17.0

• sqlalchemy>=1.0.10

• bokeh>=0.12.0

• matplotlib>=1.5.0

• pandas>=0.18.0

• GPy (optional, required for gp strategy)

• hyperopt (optional, required for hyperopt_tpe strategy)

• nose (optional, for testing)

Contributing

In case you encounter any issues with this package, please consider submitting a ticket to the GitHub Issue Tracker. We also welcome any feature requests and highly encourage users to submit pull requests for bug fixes and improvements.

For more detailed information, please refer to our documentation.

Citing

If you use Osprey in your research, please cite:

@misc{osprey,
  author       = {Robert T. McGibbon and
                  Carlos X. Hernández and
                  Matthew P. Harrigan and
                  Steven Kearnes and
                  Mohammad M. Sultan and
                  Stanislaw Jastrzebski and
                  Brooke E. Husic and
                  Vijay S. Pande},
  title        = {Osprey 1.0.0},
  month        = jun,
  year         = 2016,
  doi          = {10.5281/zenodo.56251},
  url          = {http://dx.doi.org/10.5281/zenodo.56251}
}