def get_libodbc_path():
"""
Get the first path that GraphLab Create will search for libodbc.so.
"""
c = gl.get_runtime_config()
return c['GRAPHLAB_LIBODBC_PREFIX']
def create(data=None,
target=None,
features=None,
method='auto',
validation_set=None):
"""
Create a model that trains a classifier in order to perform sentiment
analysis on a collection of documents.
When a target column name is not provided, a pretrained model will be used.
The model was trained on a large collection of product review data from both
Amazon and Yelp datasets. Predicted scores are between 0 and 1, where
higher scores indicate more positive predicted sentiment. The model is
a :class:`~graphlab.logistic_classifier.LogisticClassifier` model trained
using a bag-of-words representation of
the text data, using ratings less than 3 as negative sentiment and
ratings of more than 3 as positive sentiment.
Parameters
----------
data: SFrame, optional
Contains at least one column that contains the text data of interest.
This can be unstructured text data, such as that appearing in forums,
user-generated reviews, and so on. This is not required when using a
pre-trained model.
target: str, optional
The column name containing numeric sentiment scores for each document.
If provided, a sentiment model will be trained for the provided data set.
If not provided, a pre-trained model will be used.
features: list of str, optional
The column names of interest containing text data. Each provided column
must be str type. Defaults to using all columns of type str.
method: str, optional
Method to use for feature engineering and modeling. Currently only
bag-of-words and logistic classifier ('bow-logistic') is available.
validation_set : SFrame, optional
A dataset for monitoring the model's generalization performance.
This is ignored if no value is provided to the `target` argument.
Returns
-------
out : :class:`~SentimentAnalysisModel`
Examples
--------
You can train a sentiment analysis classifier on text data when you have
ratings data available.
>>> import graphlab as gl
>>> data = gl.SFrame({'rating': [1, 5], 'text': ['hate it', 'love it']})
>>> m = gl.sentiment_analysis.create(data, 'rating', features=['text'])
>>> m.predict_row({'text': 'really love it'})
>>> m.predict_row({'text': 'really hate it'})
If you do not have ratings data, we provide a pretrained model for you to
use as a starting point.
>>> m = gl.sentiment_analysis.create(data, features=['text'])
>>> m.predict(data)
You may also evaluate predictions against known sentiment scores.
>>> m.evaluate(data)
"""
_mt._get_metric_tracker().track('{}.create'.format(__name__))
logger = _logging.getLogger(__name__)
# Validate method.
if method == 'auto':
method = 'bow-logistic'
if method != 'bow-logistic':
raise ValueError("Unsupported method provided.")
# Check if pretrained
if target is None:
# Name of pretrained model: format is [name]/[version].
model_name = 'sentiment-combined/1'
# Download if model is not present in [tmp dir]/model_cache/.
tmp_dir = _gl.get_runtime_config()['GRAPHLAB_CACHE_FILE_LOCATIONS']
model_local_path = _os.path.join(tmp_dir, 'model_cache', model_name)
model_remote_path = 'https://static.turi.com/products/graphlab-create/resources/models/python2.7/sentiment-analysis/' + model_name
feature_extractor = _feature_extractor_for_pretrained
if not _os.path.exists(model_local_path):
logger.info('Downloading pretrained model...')
m = _gl.load_model(model_remote_path)
m.save(model_local_path)
else:
m = _gl.load_model(model_local_path)
num_rows = 0
else:
if data is None:
raise ValueError(
"The data argument is required when a target column is provided."
)
# Validate data
# Validate features. Use all columns by default
if features is None:
features = data.column_names()
# Remove target column from list of feature columns.
features = [f for f in features if f != target]
# Transform the target column and create the training set.
_target = 'like'
train = _transform_with_target(data, target, _target)
# Process training set using the default feature extractor.
feature_extractor = _default_feature_extractor
train = feature_extractor(train)
# Check for a validation set.
kwargs = {}
if validation_set is not None:
validation_set = _transform_with_target(validation_set, target,
_target)
validation_set = feature_extractor(validation_set)
kwargs['validation_set'] = validation_set
m = _gl.logistic_classifier.create(train,
target=_target,
features=features,
l2_penalty=.2,
**kwargs)
num_rows = data.num_rows()
model = SentimentAnalysisModel()
model.__proxy__.update({
'target': target,
'features': features,
'method': method,
'num_rows': num_rows,
'feature_extractor': feature_extractor,
'classifier': m
})
return model
def get_libodbc_path():
"""
Get the first path that GraphLab Create will search for libodbc.so.
"""
c = gl.get_runtime_config()
return c['GRAPHLAB_LIBODBC_PREFIX']
def create(data, target=None, features=None, method='auto', validation_set=None):
"""
Create a model that trains a classifier in order to perform sentiment
analysis on a collection of documents.
When a target column name is not provided, a pretrained model will be used.
The model was trained on a large collection of product review data from both
Amazon and Yelp datasets. Predicted scores are between 0 and 1, where
higher scores indicate more positive predicted sentiment. The model is
a :class:`~graphlab.logistic_classifier.LogisticClassifier` model trained
using a bag-of-words representation of
the text data, using ratings less than 3 as negative sentiment and
ratings of more than 3 as positive sentiment.
.. warning::
This toolkit is currently in beta, and feedback is
welcome! Please send comments to [email protected].
Parameters
----------
data: SFrame
Contains at least one column that contains the text data of interest.
This can be unstructured text data, such as that appearing in forums,
user-generated reviews, and so on.
target: str, optional
The column name containing numeric sentiment scores for each document.
If provided, a sentiment model will be trained for this data set.
If not provided, a pre-trained model will be used.
features: list of str, optional
The column names of interest containing text data. Each provided column
must be str type. Defaults to using all columns of type str.
method: str, optional
Method to use for feature engineering and modeling. Currently only
bag-of-words and logistic classifier ('bow-logistic') is available.
validation_set : SFrame, optional
A dataset for monitoring the model's generalization performance.
This is ignored if no value is provided to the `target` argument.
Returns
-------
out : :class:`~SentimentAnalysisModel`
Examples
--------
You can train a sentiment analysis classifier on text data when you have
ratings data available.
>>> import graphlab as gl
>>> data = gl.SFrame({'rating': [1, 5], 'text': ['hate it', 'love it']})
>>> m = gl.sentiment_analysis.create(data, 'rating', features=['text'])
>>> m.predict_row({'text': 'really love it'})
>>> m.predict_row({'text': 'really hate it'})
If you do not have ratings data, we provide a pretrained model for you to
use as a starting point.
>>> m = gl.sentiment_analysis.create(data, features=['text'])
>>> m.predict(data)
You may also evaluate predictions against known sentiment scores.
>>> m.evaluate(data)
"""
_mt._get_metric_tracker().track('{}.create'.format(__name__))
logger = _logging.getLogger(__name__)
# Validate method.
if method == 'auto':
method = 'bow-logistic'
if method != 'bow-logistic':
raise ValueError("Unsupported method provided.")
# Validate features. Use all columns by default
if features is None:
features = data.column_names()
# Check if pretrained
if target is None:
# Name of pretrained model: format is [name]/[version].
model_name = 'sentiment-combined/1'
# Download if model is not present in [tmp dir]/model_cache/.
tmp_dir = _gl.get_runtime_config()['GRAPHLAB_CACHE_FILE_LOCATIONS']
model_local_path = _os.path.join(tmp_dir, 'model_cache', model_name)
model_remote_path = 's3://dato-models/sentiment-analysis/' + model_name
feature_extractor = _feature_extractor_for_pretrained
if not _os.path.exists(model_local_path):
logger.info('Downloading pretrained model...')
m = _gl.load_model(model_remote_path)
m.save(model_local_path)
else:
m = _gl.load_model(model_local_path)
else:
# Remove target column from list of feature columns.
features = [f for f in features if f != target]
# Transform the target column and create the training set.
_target = 'like'
train = _transform_with_target(data, target, _target)
# Process training set using the default feature extractor.
feature_extractor = _default_feature_extractor
train = feature_extractor(train)
# Check for a validation set.
kwargs = {}
if validation_set is not None:
validation_set = _transform_with_target(validation_set, target, _target)
validation_set = feature_extractor(validation_set)
kwargs['validation_set'] = validation_set
m = _gl.logistic_classifier.create(train,
target=_target,
features=features,
l2_penalty=.2,
**kwargs)
model = SentimentAnalysisModel()
model._state.update(
{'target': target,
'features': features,
'method': method,
'num_rows': data.num_rows(),
'feature_extractor': feature_extractor,
'classifier': m})
return model
