extend score to take an array of shorttext

[rja172]

Currently, score takes only a single input and as a result, the method is very slow if you are trying to classify thousands of examples. Is there a way you can generate scores for 10K+ samples at the same time.

[stephenhky]

vectorizing the input might works....

[rja172]

do you have an example? If I do

predictions = classifier.score(sampleDF['product_name'].values)

I get following error:

AttributeError                            Traceback (most recent call last)
Input In [10], in <cell line: 2>()
      1 # predictions = model.predict(sampleDF['product_name'].head(100).values)
----> 2 predictions = model.classifier.score(sampleDF['product_name'].values)

File ~/.venv/category-predictor-e7zd-_vR-py3.9/lib/python3.9/site-packages/shorttext/classifiers/embed/nnlib/VarNNEmbedVecClassification.py:229, in VarNNEmbeddedVecClassifier.score(self, shorttext)
    226     raise e.ModelNotTrainedException()
    228 # retrieve vector
--> 229 matrix = np.array([self.shorttext_to_matrix(shorttext)])
    231 # classification using the neural network
    232 predictions = self.model.predict(matrix)

File ~/.venv/category-predictor-e7zd-_vR-py3.9/lib/python3.9/site-packages/shorttext/classifiers/embed/nnlib/VarNNEmbedVecClassification.py:205, in VarNNEmbeddedVecClassifier.shorttext_to_matrix(self, shorttext)
    193 def shorttext_to_matrix(self, shorttext):
    194     """ Convert the short text into a matrix with word-embedding representation.
    195 
    196     Given a short sentence, it converts all the tokens into embedded vectors according to
   (...)
    203     :rtype: numpy.ndarray
    204     """
--> 205     tokens = tokenize(shorttext)
    206     matrix = np.zeros((self.maxlen, self.vecsize))
    207     for i in range(min(self.maxlen, len(tokens))):

File ~/.venv/category-predictor-e7zd-_vR-py3.9/lib/python3.9/site-packages/shorttext/utils/textpreprocessing.py:9, in <lambda>(s)
      6 import snowballstemmer
      8 # tokenizer
----> 9 tokenize = lambda s: s.split(' ')
     12 # stemmer
     13 stemmer = snowballstemmer.stemmer('porter')

AttributeError: 'numpy.ndarray' object has no attribute 'split'

if I try using np.vectorize command (example: predictions = np.vectorize(lambda x: model.classifier.score(x))(sampleDF['product_name'])) It doesn't improve time by a lot

[rja172]

As suggested, vectorizing np. vectorize(self.classifier.score)(data['X']) did help speed up the process. However, inference is still very slow. I think if the internal functions can handle multiple short text at the same time might significantly help speed up things. It is taking about 1 hours to resolve 100K short text.

[stephenhky]

Because the procedure involves tokenizing the string, it takes some time. Yes, doing it in parallel can speed it up. I will need to do some careful changes to the codes.

[ragrawal]

this PR addresses it: #134

[stephenhky]

Release 1.5.6.