destimator makes it easy to store trained scikit-learn
estimators together with their metadata (training data, package versions, performance numbers etc.). This makes it much safer to store already-trained classifiers/regressors and allows for better reproducibility (see this talk by Alex Gaynor for some rationale).
Specifically, the DescribedEstimator
class proxies most calls to the original Estimator
it is wrapping, but also contains the following information:
- training and test (validation) data (
features_train
,labels_train
,features_test
,labels_test
) - creation date (
created_at
) - feature names (
feature_names
) - performance numbers on the test set (
precision
,recall
,fscore
,support
via sklearn) - distribution info (
distribution_info
; python distribution and versions of all installed packages) - VCS hash (
vcs_hash
, if used inside a git repository, otherwise and empty string).
An instantiated DescribedEstimator
can be easily serialized using the .save()
method and deserialized using either .from_file()
or .from_url()
. Did you ever want to store your models in S3? Now it's easy!
DescribedEstimator
can be used as follows:
import numpy as np
from sklearn.datasets import load_iris
from sklearn.ensemble import RandomForestClassifier
from sklearn.cross_validation import train_test_split
from sklearn.metrics import precision_recall_fscore_support
from destimator import DescribedEstimator
# get some data
iris = load_iris()
features = iris.data
labels = iris.target
features_train, features_test, labels_train, labels_test = train_test_split(features, labels, test_size=0.1)
# train an estimator as usual (in this case a RandomForestClassifier)
clf = RandomForestClassifier(n_estimators=10, max_depth=None, min_samples_split=10, random_state=0)
clf.fit(features_train, labels_train)
# wrap the estimator in the DescribedEstimator class giving it all the training and test (validation) data
dclf = DescribedEstimator(
clf,
features_train,
labels_train,
features_test,
labels_test,
iris.feature_names
)
Now you can use the classifier as usual:
print(dclf.predict(features_test))
> [2 1 2 2 0 1 0 2 2 1 2 0 2 1 2]
and you can also access a bunch of other properties, such as the training data you supplied:
print(dclf.feature_names)
> ['sepal length (cm)', 'sepal width (cm)', 'petal length (cm)', 'petal width (cm)']
print(dclf.features_test)
> [[ 6.3 2.8 5.1 1.5]
[ 5.6 3. 4.5 1.5]
[ 6.7 3.1 5.6 2.4]
[ 6. 2.7 5.1 1.6]
[ 4.9 3.1 1.5 0.1]
[ 6.2 2.2 4.5 1.5]
[ 4.7 3.2 1.6 0.2]
[ 6.9 3.1 5.1 2.3]
[ 7.7 2.6 6.9 2.3]
[ 5.8 2.6 4. 1.2]
[ 7.2 3. 5.8 1.6]
[ 5.4 3.7 1.5 0.2]
[ 7.2 3.2 6. 1.8]
[ 6.3 3.3 4.7 1.6]
[ 6.8 3.2 5.9 2.3]]
print(dclf.labels_test)
> [2 1 2 1 0 1 0 2 2 1 2 0 2 1 2]
the performance numbers:
print('precision: %s' % (dclf.precision))
> precision: [1.0, 1.0, 0.875]
print('recall: %s' % (dclf.recall))
> recall: [1.0, 0.8, 1.0]
print('fscore: %s' % (dclf.fscore))
> fscore: [1.0, 0.888888888888889, 0.9333333333333333]
print('support: %s' % (dclf.support))
> support: [3, 5, 7]
print('roc_auc: %s' % (dclf.roc_auc))
> roc_auc: 0.5
or information about the Python distribution used for training:
from pprint import pprint
pprint(dclf.distribution_info)
> {'packages': ['appnope==0.1.0',
'decorator==4.0.4',
'destimator==0.0.0.dev3',
'gnureadline==6.3.3',
'ipykernel==4.2.1',
'ipython-genutils==0.1.0',
'ipython==4.0.1',
'ipywidgets==4.1.1',
'jinja2==2.8',
'jsonschema==2.5.1',
'jupyter-client==4.1.1',
'jupyter-console==4.0.3',
'jupyter-core==4.0.6',
'jupyter==1.0.0',
'markupsafe==0.23',
'mistune==0.7.1',
'nbconvert==4.1.0',
'nbformat==4.0.1',
'notebook==4.0.6',
'numpy==1.10.1',
'path.py==8.1.2',
'pexpect==4.0.1',
'pickleshare==0.5',
'pip==7.1.2',
'ptyprocess==0.5',
'pygments==2.0.2',
'pyzmq==15.1.0',
'qtconsole==4.1.1',
'requests==2.8.1',
'scikit-learn==0.17',
'scipy==0.16.1',
'setuptools==18.2',
'simplegeneric==0.8.1',
'terminado==0.5',
'tornado==4.3',
'traitlets==4.0.0',
'wheel==0.24.0'],
'python': '3.5.0 (default, Sep 14 2015, 02:37:27) \n'
'[GCC 4.2.1 Compatible Apple LLVM 6.1.0 (clang-602.0.53)]'}
Finally, the object can be serialized to a zip file containing all the above data:
dclf.save('./classifiers', 'dclf')
and deserialized either from a file,
dclf = DescribedEstimator.from_file('./classifiers/dclf.zip')
or from a URL:
dclf = DescribedEstimator.from_url('http://localhost/dclf.zip')