def train_model(self, data, model_library, model):
"""data is an instance of DragnetModelData
model_library is a list of model definitions as input to
run_train_models
model provides model.make_features to make the features
"""
from mozsci.map_train import run_train_models
# to train the model need a set of all the features and their labels
# the features + labels are block level
# get the features from the first document to see how many features we have
features, labels, weights = self.make_features_from_data(data, model)
# cap weights!
weights = np.minimum(weights, 200)
# do kfold cross validation
folds = cv_kfold(len(labels), self.kfolds, seed=2)
if self.weighted:
errors = run_train_models(processes=4, model_library=model_library,
X=features, y=labels, folds=folds, weights=weights)
else:
errors = run_train_models(processes=4, model_library=model_library,
X=features, y=labels, folds=folds)
return errors, features, labels, weights, folds
def setUp(self):
np.random.seed(5)
self.X = np.linspace(0, 1, 100).reshape(100, 1)
self.y = (5 * self.X.reshape(100, ) - 2 + np.random.rand(100) >
0).astype(np.int)
self.folds = cv_kfold(100, 4, seed=2)
def train_model(self, data, model_library, features_to_use):
"""data is an instance of DragnetModelData
model_library: the block_models to train as a list of model
definitions as input to run_train_models
features_to_use = a list of the features to use. Must be one of
the features known by AllFeatures
"""
from . import AllFeatures
from .blocks import TagCountReadabilityBlockifier as Blkr
from mozsci.map_train import run_train_models
# assemble the features
feature_instances = []
for f in features_to_use:
feature_instances.append(AllFeatures.get(f))
# do feature centering
print("Initializing features")
for f in feature_instances:
# check to see if this feature needs to be init
# if so, then init it, take the return object and serialize to json
if hasattr(f, 'init_params'):
# initialize it
model_init = ContentExtractionModel(Blkr, [f], None)
features, labels, weights = self.make_features_from_data(
data, model_init, train=True)
mean_std = f.init_params(features)
f.set_params(mean_std)
model_to_train = ContentExtractionModel(Blkr, feature_instances, None)
# train the model
print("Training the model")
features, labels, weights = self.make_features_from_data(
data, model_to_train, training_or_test='training')
# cap weights!
weights = np.minimum(weights, 200)
# do kfold cross validation
if self.kfolds > 1:
folds = cv_kfold(len(labels), self.kfolds, seed=2)
else:
folds = None
if self.weighted:
errors = run_train_models(
processes=1, model_library=model_library,
X=features, y=labels, folds=folds, weights=weights)
else:
errors = run_train_models(
processes=1, model_library=model_library,
X=features, y=labels, folds=folds)
return errors, features, labels, weights, folds
def train_model(self, data, model_library, features_to_use):
"""data is an instance of DragnetModelData
model_library: the block_models to train as a list of model
definitions as input to run_train_models
features_to_use = a list of the features to use. Must be one of
the features known by AllFeatures
"""
from . import AllFeatures
from .blocks import TagCountReadabilityBlockifier as Blkr
from mozsci.map_train import run_train_models
# assemble the features
feature_instances = []
for f in features_to_use:
feature_instances.append(AllFeatures.get(f))
# do feature centering
print("Initializing features")
for f in feature_instances:
# check to see if this feature needs to be init
# if so, then init it, take the return object and serialize to json
if hasattr(f, 'init_params'):
# initialize it
model_init = ContentExtractionModel(Blkr, [f], None)
features, labels, weights = self.make_features_from_data(
data, model_init, train=True)
mean_std = f.init_params(features)
f.set_params(mean_std)
model_to_train = ContentExtractionModel(Blkr, feature_instances, None)
# train the model
print("Training the model")
features, labels, weights = self.make_features_from_data(
data, model_to_train, training_or_test='training')
# cap weights!
weights = np.minimum(weights, 200)
# do kfold cross validation
if self.kfolds > 1:
folds = cv_kfold(len(labels), self.kfolds, seed=2)
else:
folds = None
if self.weighted:
errors = run_train_models(
processes=1, model_library=model_library,
X=features, y=labels, folds=folds, weights=weights)
else:
errors = run_train_models(
processes=1, model_library=model_library,
X=features, y=labels, folds=folds)
return errors, features, labels, weights, folds
def test_cv_kfold(self):
folds = cross_validate.cv_kfold(20, 4, seed=2)
sum_training = np.sum([len(ele[0]) for ele in folds])
self.assertTrue(sum_training == 3 * 20)
sum_training = np.sum([len(ele[1]) for ele in folds])
self.assertTrue(sum_training == 20)
actual_folds = [
[[0, 3, 4, 5, 8, 9, 17, 2, 7, 10, 11, 13, 15, 16, 18], [1, 6, 12, 14, 19]],
[[1, 6, 12, 14, 19, 2, 7, 10, 11, 13, 15, 16, 18], [0, 3, 4, 5, 8, 9, 17]],
[[1, 6, 12, 14, 19, 0, 3, 4, 5, 8, 9, 17, 15, 16, 18], [2, 7, 10, 11, 13]],
[[1, 6, 12, 14, 19, 0, 3, 4, 5, 8, 9, 17, 2, 7, 10, 11, 13], [15, 16, 18]]]
self.assertEqual(actual_folds, folds)
def test_cv_kfold(self):
folds = cross_validate.cv_kfold(20, 4, seed=2)
sum_training = np.sum([len(ele[0]) for ele in folds])
self.assertTrue(sum_training == 3 * 20)
sum_training = np.sum([len(ele[1]) for ele in folds])
self.assertTrue(sum_training == 20)
actual_folds = [
[[0, 3, 4, 5, 8, 9, 17, 2, 7, 10, 11, 13, 15, 16, 18],
[1, 6, 12, 14, 19]],
[[1, 6, 12, 14, 19, 2, 7, 10, 11, 13, 15, 16, 18],
[0, 3, 4, 5, 8, 9, 17]],
[[1, 6, 12, 14, 19, 0, 3, 4, 5, 8, 9, 17, 15, 16, 18],
[2, 7, 10, 11, 13]],
[[1, 6, 12, 14, 19, 0, 3, 4, 5, 8, 9, 17, 2, 7, 10, 11, 13],
[15, 16, 18]]
]
self.assertEqual(actual_folds, folds)
def setUp(self):
np.random.seed(5)
self.X = np.linspace(0, 1, 100).reshape(100, 1)
self.y = (5 * self.X.reshape(100, ) - 2 + np.random.rand(100) > 0).astype(np.int)
self.folds = cv_kfold(100, 4, seed=2)
