the on the fly sklearn dictionary vectorizer and SGD classifier and GridSearchCV and RandomSearchCV.
It is always painful to generate dictinaries for SGD algorithms. Why not use them on the fly.
Parameter in GridSearch, RandomSearch are added to make Rdd distributed again, (splits/duplicate raw data and get result on the fly)
Example
For steaming dictionaries/jsons
from on_the_fly import FlyVectorizer, FlyClassifier
vec = FlyVectorizer()
clf = FlyClassifier()
features = ['name', 'age', 'stuff..']
label = ['gender']
for batch_data_in_dict in iterator_of_data_in_dict:
batch_data = vec.partial_fit_transform(batch_data_in_dict)
feature_dimension = vec.subset_features(features)
label_dimension = vec.subset_features(label)
batch_X = batch_data[:, feature_dimension]
batch_y = batch_data[:, label_dimension]
clf.partial_fit(batch_X, batch_y)For spark rdd of dictionaries
from on_the_fly import FlyClassifier, RddVectorizer, RddClassifier
vec = RddVectorizer(features=['name', 'age', 'stuff'], label='gender')
base_clf = FlyClassifier(loss='log')
clf = RddClassifier(base_clf)
training_design_matrix = vec.fit_transform(trainning_rdd_of_dicts)
clf.fit(training_design_matrix)
testing_design_matrix = vec.transform(testing_rdd_of_dicts)
clf.score(testing_design_matrix)
