classifier.py

from . import utils as U
from .imports import *

__all__ = ["NBSVM"]


class Classifier:
    def __init__(self, model=None):
        """
        instantiate a classifier with an optional previously-saved model
        """
        self.model = None

    def create_model(self, ctype, texts, use_tfidf=False, **kwargs):
        """
        ```
        create a model
        Args:
          ctype(str): one of {'nbsvm', 'logreg', 'sgdclassifier'}
          texts(list): list of texts
          kwargs(dict):   additional parameters should have one of the following prefixes:
                          vec__ :  hyperparameters to CountVectorizer (e.g., vec__max_features=10000)
                          tfidf__ :  hyperparameters to TfidfTransformer
                          clf__:   hyperparameters to classifier (specific to ctype).
                                   If ctype='logreg', then an example is clf__solver='liblinear'.
        ```
        """
        if ctype == "nbsvm":
            if kwargs.get("vec__binary", False) is False:
                warnings.warn("nbsvm must use binary=True - changing automatically")
            if use_tfidf:
                warnings.warn("nbsvm must use use_tfidf=False = changing automatically")
        vec_kwargs = dict(
            (k.replace("vec__", ""), kwargs[k]) for k in kwargs if k.startswith("vec__")
        )
        tfidf_kwargs = dict(
            (k.replace("tfidf__", ""), kwargs[k])
            for k in kwargs
            if k.startswith("tfidf__")
        )
        clf_kwargs = dict(
            (k.replace("clf__", ""), kwargs[k]) for k in kwargs if k.startswith("clf__")
        )

        lang = U.detect_lang(texts)
        if U.is_chinese(lang) and not vec_kwargs.get("token_pattern", None):
            vec_kwargs["token_pattern"] = r"(?u)\b\w+\b"
        elif not kwargs.get("vec__token_pattern", None):
            vec_kwargs["token_pattern"] = r"\w+|[%s]" % string.punctuation

        if ctype == "nbsvm":
            clf = NBSVM(**clf_kwargs)
        elif ctype == "logreg":
            clf = LogisticRegression(**clf_kwargs)
        elif ctype == "sgdclassifier":
            clf = SGDClassifier(**clf_kwargs)
        else:
            raise ValueError("Unknown ctype: %s" % (ctype))

        pipeline = [("vect", CountVectorizer(**vec_kwargs))]
        if use_tfidf:
            pipeline.append(("tfidf", TfidfTransformer(**tfidf_kwargs)))
        pipeline.append(("clf", clf))
        self.model = Pipeline(pipeline)
        return

    @classmethod
    def load_texts_from_folder(
        cls, folder_path, subfolders=None, shuffle=True, encoding=None
    ):
        """
        ```
        load text files from folder

        Args:
          folder_path(str): path to folder containing documents
                            The supplied folder should contain a subfolder
                            for each category, which will be used as the class label
          subfolders(list): list of subfolders under folder_path to consider
                            Example: If folder_path contains subfolders pos, neg, and
                            unlabeled, then unlabeled folder can be ignored by
                            setting subfolders=['pos', 'neg']
          shuffle(bool):  If True, list of texts will be shuffled
          encoding(str): encoding to use.  default:None (auto-detected)
        Returns:
          tuple: (texts, labels, label_names)
        ```
        """
        bunch = load_files(folder_path, categories=subfolders, shuffle=shuffle)
        texts = bunch.data
        labels = bunch.target
        label_names = bunch.target_names
        # print('target names:')
        # for idx, label_name in enumerate(bunch.target_names):
        # print('\t%s:%s' % (idx, label_name))

        # decode based on supplied encoding
        if encoding is None:
            encoding = U.detect_encoding(texts)
            if encoding != "utf-8":
                print("detected encoding: %s" % (encoding))

        try:
            texts = [text.decode(encoding) for text in texts]
        except:
            print(
                "Decoding with %s failed 1st attempt - using %s with skips"
                % (encoding, encoding)
            )
            texts = U.decode_by_line(texts, encoding=encoding)
        return (texts, labels, label_names)

    @classmethod
    def load_texts_from_csv(
        cls,
        csv_filepath,
        text_column="text",
        label_column="label",
        sep=",",
        encoding=None,
    ):
        """
        ```
        load text files from csv file
        CSV should have at least two columns.
        Example:
        Text               | Label
        I love this movie. | positive
        I hated this movie.| negative


        Args:
          csv_filepath(str): path to CSV file
          text_column(str): name of column containing the texts. default:'text'
          label_column(str): name of column containing the labels in string format
                             default:'label'
          sep(str): character that separates columns in CSV. default:','
          encoding(str): encoding to use. default:None (auto-detected)
        Returns:
          tuple: (texts, labels, label_names)
        ```
        """
        if encoding is None:
            with open(csv_filepath, "rb") as f:
                encoding = U.detect_encoding([f.read()])
                if encoding != "utf-8":
                    print("detected encoding: %s (if wrong, set manually)" % (encoding))
        import pandas as pd

        df = pd.read_csv(csv_filepath, encoding=encoding, sep=sep)
        texts = df[text_column].fillna("fillna").values
        labels = df[label_column].values
        le = LabelEncoder()
        le.fit(labels)
        labels = le.transform(labels)
        return (texts, labels, le.classes_)

    def fit(self, x_train, y_train, ctype="logreg"):
        """
        ```
        train a classifier
        Args:
          x_train(list or np.ndarray):  training texts
          y_train(np.ndarray):  training labels
          ctype(str):  One of {'logreg', 'nbsvm', 'sgdclassifier'}.  default:nbsvm
        ```
        """
        lang = U.detect_lang(x_train)
        if U.is_chinese(lang):
            x_train = U.split_chinese(x_train)
        if self.model is None:
            self.create_model(ctype, x_train)
        self.model.fit(x_train, y_train)
        return self

    def predict(self, x_test, return_proba=False):
        """
        ```
        make predictions on text data
        Args:
          x_test(list or np.ndarray or str): array of texts on which to make predictions or a string representing text
        ```
        """
        if return_proba and not hasattr(self.model["clf"], "predict_proba"):
            raise ValueError(
                "%s does not support predict_proba" % (type(self.model["clf"]).__name__)
            )
        if isinstance(x_test, str):
            x_test = [x_test]
        lang = U.detect_lang(x_test)
        if U.is_chinese(lang):
            x_test = U.split_chinese(x_test)
        if self.model is None:
            raise ValueError("model is None - call fit or load to set the model")
        if return_proba:
            predicted = self.model.predict_proba(x_test)
        else:
            predicted = self.model.predict(x_test)
        if len(predicted) == 1:
            predicted = predicted[0]
        return predicted

    def predict_proba(self, x_test):
        """
        predict_proba
        """
        return self.predict(x_test, return_proba=True)

    def evaluate(self, x_test, y_test):
        """
        ```
        evaluate
        Args:
          x_test(list or np.ndarray):  training texts
          y_test(np.ndarray):  training labels
        ```
        """
        predicted = self.predict(x_test)
        return np.mean(predicted == y_test)

    def save(self, filename):
        """
        save model
        """
        dump(self.model, filename)

    def load(self, filename):
        """
        load model
        """
        self.model = load(filename)

    def grid_search(self, params, x_train, y_train, n_jobs=-1):
        """
        ```
        Performs grid search to find optimal set of hyperparameters
        Args:
          params (dict):  A dictionary defining the space of the search.
                          Example for finding optimal value of alpha in NBSVM:
                        parameters = {
                                      #'clf__C': (1e0, 1e-1, 1e-2),
                                      'clf__alpha': (0.1, 0.2, 0.4, 0.5, 0.75, 0.9, 1.0),
                                      #'clf__fit_intercept': (True, False),
                                      #'clf__beta' : (0.1, 0.25, 0.5, 0.9)
                                      }
          n_jobs(int): number of jobs to run in parallel.  default:-1 (use all processors)
        ```
        """
        gs_clf = GridSearchCV(self.model, params, n_jobs=n_jobs)
        gs_clf = gs_clf.fit(x_train, y_train)
        # gs_clf.best_score_
        for param_name in sorted(params.keys()):
            print("%s: %r" % (param_name, gs_clf.best_params_[param_name]))
        return


class NBSVM(BaseEstimator, LinearClassifierMixin, SparseCoefMixin):
    def __init__(self, alpha=0.75, C=0.01, beta=0.25, fit_intercept=False):
        self.alpha = alpha
        self.C = C
        self.beta = beta
        self.fit_intercept = fit_intercept

    def fit(self, X, y):
        self.classes_ = np.unique(y)
        if len(self.classes_) == 2:
            coef_, intercept_ = self._fit_binary(X, y)
            self.coef_ = coef_
            self.intercept_ = intercept_
        else:
            coef_, intercept_ = zip(
                *[self._fit_binary(X, y == class_) for class_ in self.classes_]
            )
            self.coef_ = np.concatenate(coef_)
            self.intercept_ = np.array(intercept_).flatten()
        return self

    def _fit_binary(self, X, y):
        p = np.asarray(self.alpha + X[y == 1].sum(axis=0)).flatten()
        q = np.asarray(self.alpha + X[y == 0].sum(axis=0)).flatten()
        r = np.log(p / np.abs(p).sum()) - np.log(q / np.abs(q).sum())
        b = np.log((y == 1).sum()) - np.log((y == 0).sum())

        if isinstance(X, spmatrix):
            indices = np.arange(len(r))
            r_sparse = coo_matrix((r, (indices, indices)), shape=(len(r), len(r)))
            X_scaled = X * r_sparse
        else:
            X_scaled = X * r

        lsvc = LinearSVC(
            C=self.C, fit_intercept=self.fit_intercept, max_iter=10000, dual=True
        ).fit(X_scaled, y)

        mean_mag = np.abs(lsvc.coef_).mean()

        coef_ = (1 - self.beta) * mean_mag * r + self.beta * (r * lsvc.coef_)

        intercept_ = (1 - self.beta) * mean_mag * b + self.beta * lsvc.intercept_

        return coef_, intercept_