python - Scikit-learn out-of-core 文本分类内存消耗

Question

我正在尝试使用 scikit-learn 对大量文本文档进行分类，尽管我使用的是核心外功能(使用 SGDClassifier 和 HashingVectorizer) 该程序似乎消耗了大量 RAM (>10GB)。在此之前，我执行了词形还原并从文本数据中删除了停用词。我觉得我在这里错过了一些重要的东西。你能发现我的代码中的错误吗？

非常感谢您的任何建议!

这是我的 python 代码:

import time
import numpy as np
import os
import re
import pyprind
from sklearn.feature_extraction.text import HashingVectorizer
from sklearn.linear_model import SGDClassifier
from sklearn.naive_bayes import MultinomialNB
from sklearn import metrics

directory = "mydirectory"
batch_size = 1000
n_batches = 44
pbar = pyprind.ProgBar(n_batches)

class Doc_Iterable:
    def __init__(self, file):
        self.file = file
    def __iter__(self):
        for line in self.file:
            line = re.sub('[^\w\s]|(.\d{1,4}[\./]\d{1,2}[\./]\d{1,4})|(\s\d{1,})', '', line)
            yield line


def stream_docs(path, texts_file, labels_file):
    with open(path + texts_file, 'r') as fX, open(path + labels_file, 'r') as fy:
        for text in fX:
            label = next(fy)
            text = re.sub('[^\w\s]|(.\d{1,4}[\./]\d{1,2}[\./]\d{1,4})|(\s\d{1,})', '', text)
            yield text, label

def get_minibatch(doc_stream, size):
    X, y = [], []
    for _ in range(size):
        text, label = next(doc_stream)
        X.append(text)
        y.append(label)
    return X, y


classes = set()
for label in open(directory + 'y_train', 'r'):
    classes.add(label)
for label in open(directory + 'y_test', 'r'):
    classes.add(label)
classes = list(classes)

validation_scores = []
training_set_size = []

h_vectorizer = HashingVectorizer(lowercase=True, ngram_range=(1,1))
clf = SGDClassifier(loss='hinge', n_iter=5, alpha=1e-4, shuffle=True)

doc_stream = stream_docs(path=directory, texts_file='X_train', labels_file='y_train')
n_samples = 0
iteration = 0

for _ in range(n_batches):
    print("Training with batch nr.", iteration)
    iteration += 1

    X_train, y_train = get_minibatch(doc_stream, size=batch_size)

    n_samples += len(X_train)

    X_train = h_vectorizer.transform(X_train)

    clf.partial_fit(X_train, y_train, classes=classes)

    pbar.update()


del X_train
del y_train
print("Training complete. Classifier trained with " + str(n_samples) + " samples.")
print()
print("Testing...")
print()
X_test = h_vectorizer.transform(Doc_Iterable(open(directory + 'X_test')))
y_test = np.genfromtxt(directory + 'y_test', dtype=None, delimiter='|').astype(str)
prediction = clf.predict(X_test)
score = metrics.accuracy_score(y_test, prediction)
print("Accuracy: ", score)
print()

Answer 1

尝试调整HashingVectorizer中的n_features，例如:

h_vectorizer = HashingVectorizer(n_features=10000, lowercase=True, ngram_range=(1,1))

使用默认参数 (n_features=1048576)，您可以预期转换后的矩阵最多具有:

1048576(features) x 1000(mini batch size) x 8 bytes = 8.4 GB

由于稀疏性，它会小于那个值，但分类器的系数会相加:

1048576(features) x len(classes) * 8 bytes

这样可以解释您当前的内存使用情况。