我正在尝试实现 Maxent 分类器，但在使用时遇到问题 iis算法。以下代码适用于gis算法。
import nltk
from nltk.classify import MaxentClassifier, accuracy
from featx import split_label_feats, label_feats_from_corpus
from nltk.corpus import movie_reviews
from nltk.classify import megam
from openpyxl import load_workbook
from featx import bag_of_non_words
from nltk.tokenize import word_tokenize
movie_reviews.categories()
lfeats = label_feats_from_corpus(movie_reviews)
lfeats.keys()
train_feats, test_feats = split_label_feats(lfeats)
me_classifier = nltk.MaxentClassifier.train(train_feats, algorithm='iis', trace=0, max_iter=3)
print accuracy(me_classifier, test_feats)
我正在 WIN32 机器上工作，上面的代码来自 Jacob Perkins 的 NLTK 书。 它抛出的警告是
C:\Python27\lib\site-packages\nltk\classify\maxent.py:1308: RuntimeWarning: invalid value encountered in multiply
sum1 = numpy.sum(exp_nf_delta * A, axis=0)
C:\Python27\lib\site-packages\nltk\classify\maxent.py:1309: RuntimeWarning: invalid value encountered in multiply
sum2 = numpy.sum(nf_exp_nf_delta * A, axis=0)
C:\Python27\lib\site-packages\nltk\classify\maxent.py:1315: RuntimeWarning: invalid value encountered in divide
deltas -= (ffreq_empirical - sum1) / -sum2
然后计算机挂起。所以我必须停止执行。
.
最佳答案
首先，您导入未排序的库的方式太困惑了。还有很多未使用的进口产品。经过一番谷歌搜索后，让我们减少导入并坚持这样做:
from collections import defaultdict
import nltk
from nltk.classify import MaxentClassifier, accuracy
from nltk.corpus import movie_reviews
然后我发现 featx 是 Jacob Perkins 在他的书中使用的一些示例模块，这是一个更好的来源( https://github.com/sophist114/Python/blob/master/EmotionAnalysis.py )。因此，我们这里有一个文档版本，其中包含一些函数功能的解释:
def bag_of_words(words):
"""
Change a document into a BOW feature vector represented by a dict object.
"""
return dict([(word, True) for word in words])
def label_feats_from_corpus(corp, feature_detector=bag_of_words):
"""
Change the corpus into a feature matrix. Sometimes the proceess is
known as vectorization. The default is the use BOW features.
"""
label_feats = defaultdict(list)
for label in corp.categories():
for fileid in corp.fileids(categories=[label]):
feats = feature_detector(corp.words(fileids=[fileid]))
label_feats[label].append(feats)
return label_feats
def split_label_feats(lfeats, split=0.75):
"""
Splits corpus into train and test portion.
This module is used after using `label_feats_from_corpus`.
"""
train_feats = []
test_feats = []
for label, feats in lfeats.iteritems():
cutoff = int(len(feats) * split)
train_feats.extend([(feat, label) for feat in feats[:cutoff]])
test_feats.extend([(feat, label) for feat in feats[cutoff:]])
return train_feats, test_feats
现在让我们回顾一下模型的训练和测试过程，首先是特征提取:
# Extract features from corpus and for each document label it with the appropriate labels.
label_feats = label_feats_from_corpus(movie_reviews)
让我们看看调用 label_feats_from_corpus 后会得到什么:
for label in label_feats:
for document in label_feats[label]:
print label, document
break
break
[输出]:
neg {u'all': True, u'concept': True, u'skip': True, u'go': True, u'seemed': True, u'suits': True, u'presents': True, u'to': True, u'sitting': True, u'very': True, u'horror': True, u'continues': True, u'every': True, u'exact': True, u'cool': True, u'entire': True, u'did': True, u'dig': True, u'flick': True, u'neighborhood': True, u'crow': True, u'street': True, u'video': True, u'further': True, u'even': True, u'what': True, u'hide': True, u'giving': True, u'new': True, u'ever': True, u'here': True, u'understanding': True, u'entertain': True, u'studio': True, u'others': True, u'kudos': True, u'weird': True, u'makes': True, u'explained': True, u'rarely': True, u'plot': True, u'fed': True, u'disappearances': True, u'from': True, u'would': True, u'&': True, u'two': True, u'music': True, u'films': True, u'themselves': True, u'until': True, u'more': True, u'teen': True, u'clue': True, u'stick': True, u'given': True, u'me': True, u'this': True, u'package': True, u'movies': True, u'making': True, u'my': True, u'give': True, u'fuck': True, u'want': True, u'sense': True, u'!': True, u'holds': True, u'write': True, u'how': True, u'hot': True, u'stir': True, u'okay': True, u'beauty': True, u'mess': True, u'overall': True, u'after': True, u'coming': True, u'such': True, u'guys': True, u'types': True, u'a': True, u'downshifts': True, u'chasing': True, u'redundant': True, u'so': True, u'enter': True, u'playing': True, u'executed': True, u'over': True, u'insight': True, u'years': True, u'still': True, u'its': True, u'before': True, u'thrilling': True, u'somewhere': True, u',': True, u'actually': True, u'meantime': True, u'production': True, u'main': True, u'might': True, u'then': True, u'good': True, u'break': True, u'they': True, u'half': True, u'not': True, u'now': True, u'always': True, u'didn': True, u'arrow': True, u'mean': True, u'bentley': True, u'generation': True, u'idea': True, u'engaging': True, u'happen': True, u'out': True, u"'": True, u'since': True, u'7': True, u'got': True, u'highway': True, u'shows': True, u'blair': True, u'turning': True, u'little': True, u'completely': True, u'shelves': True, u'starts': True, u'terribly': True, u'american': True, u'jumbled': True, u'chopped': True, u'one': True, u'fantasy': True, u'visions': True, u'guess': True, u'"': True, u'2': True, u'too': True, u'wrapped': True, u'final': True, u'slasher': True, u'that': True, u'explanation': True, u'took': True, u'part': True, u'attempt': True, u'10': True, u'kind': True, u'scenes': True, u'feeling': True, u'and': True, u'mind': True, u'sad': True, u'have': True, u'need': True, u'seem': True, u'apparently': True, u'-': True, u'also': True, u'which': True, u'sure': True, u'normal': True, u'who': True, u'most': True, u'don': True, u'drive': True, u'ways': True, u'entertaining': True, u'review': True, u'came': True, u'ending': True, u'find': True, u'touches': True, u'craziness': True, u'(': True, u'should': True, u'only': True, u'going': True, u'pretty': True, u'joblo': True, u'folks': True, u'8': True, u'do': True, u'his': True, u'get': True, u'watch': True, u'feels': True, u'despite': True, u'him': True, u'bad': True, u'where': True, u'lazy': True, u'see': True, u'decided': True, u'are': True, u'sorta': True, u'movie': True, u'nightmare': True, u'3': True, u'unravel': True, u'melissa': True, u'correctly': True, u'flicks': True, u'we': True, u'packaged': True, u'nightmares': True, u'genre': True, u'20': True, u'memento': True, u'both': True, u'accident': True, u's': True, u'witch': True, u'point': True, u'character': True, u'whatever': True, u'tons': True, u'simply': True, u'church': True, u'throughout': True, u'decent': True, u'been': True, u'.': True, u'secret': True, u'life': True, u'kids': True, u'personally': True, u'look': True, u'these': True, u'plain': True, u'harder': True, u'apparitions': True, u'while': True, u'neat': True, u've': True, u'is': True, u'it': True, u'couples': True, u'someone': True, u'in': True, u'chase': True, u'different': True, u')': True, u'things': True, u'make': True, u'same': True, u'member': True, u'strange': True, u'9': True, u'party': True, u'applaud': True, u'drink': True, u'director': True, u'running': True, u'characters': True, u'off': True, u'i': True, u'salvation': True, u'well': True, u'obviously': True, u'edge': True, u'echoes': True, u'the': True, u'away': True, u'just': True, u'generally': True, u'elm': True, u'excites': True, u'seems': True, u'snag': True, u'wes': True, u'4': True, u'has': True, u'big': True, u'showing': True, u'five': True, u'know': True, u'world': True, u'bit': True, u'password': True, u'dreams': True, u'like': True, u'lost': True, u'audience': True, u't': True, u'looooot': True, u'because': True, u'deal': True, u'people': True, u'back': True, u'dead': True, u'unraveling': True, u'critique': True, u'confusing': True, u'for': True, u'bottom': True, u'/': True, u'does': True, u'assuming': True, u'?': True, u'be': True, u'although': True, u'by': True, u'on': True, u'about': True, u'oh': True, u'of': True, u'runtime': True, u'or': True, u'own': True, u'strangeness': True, u'into': True, u'down': True, u'your': True, u'her': True, u'there': True, u'start': True, u'way': True, u'biggest': True, u':': True, u'head': True, u'offering': True, u'but': True, u'taken': True, u'line': True, u'trying': True, u'with': True, u'he': True, u'up': True, u'us': True, u'problem': True, u'minutes': True, u'figured': True, u'doesn': True, u'an': True, u'as': True, u'girlfriend': True, u'mold': True, u'sagemiller': True, u'film': True, u'again': True, u'no': True, u'when': True, u'actors': True, u'you': True, u'really': True, u'dies': True, u'problems': True, u'ago': True}
因此，我们得到一个带有 neg 标签的文档，对于文档中的每个单词，我们看到所有单词都是 True。目前每个文档仅包含它所具有的特征(即单词)。
让我们继续:
# Let's split the data up into train and test.
train_feats, test_feats = split_label_feats(label_feat)
现在我们看到 split_label_feats 更改了键值结构，这样 train_feats 的每次迭代都会为我们提供一个包含 (features, label) 元组的文档
for features, label in train_documents:
label, features
break
print len(train_documents)
print len(test_documents)
# Get the number of documents in movie_review corpus
num_docs_in_corpus = len(list(chain(*[movie_reviews.fileids(categories=[cat]) for cat in movie_reviews.categories()])))
print len(train_documents) + len(test_documents) == num_docs_in_corpus
[输出]:
1500
500
True
因此，当您运行以下行时，该错误似乎只能由最后两行代码引起:
# To train the tagger.
me_classifier = nltk.MaxentClassifier.train(train_documents, algorithm='iis', trace=0, max_iter=3)
您会收到这些警告，但请注意，代码仍在构建模型!!!!所以这只是由于下溢而发出的警告，请参阅 What are arithmetic underflow and overflow in C?
构建分类器需要一段时间，但不用担心，只需等待它完成，不要ctr + c结束Python进程。如果你杀死了这个过程，你会看到这个:
Training stopped: keyboard interrupt
那么让我们了解为什么会出现警告，给出了 4 个警告:
/usr/local/lib/python2.7/dist-packages/nltk/classify/maxent.py:1306: RuntimeWarning: overflow encountered in power
exp_nf_delta = 2 ** nf_delta
/usr/local/lib/python2.7/dist-packages/nltk/classify/maxent.py:1308: RuntimeWarning: invalid value encountered in multiply
sum1 = numpy.sum(exp_nf_delta * A, axis=0)
/usr/local/lib/python2.7/dist-packages/nltk/classify/maxent.py:1309: RuntimeWarning: invalid value encountered in multiply
sum2 = numpy.sum(nf_exp_nf_delta * A, axis=0)
/usr/local/lib/python2.7/dist-packages/nltk/classify/maxent.py:1315: RuntimeWarning: invalid value encountered in divide
deltas -= (ffreq_empirical - sum1) / -sum2
它们都指向 NLTk maxent 实现中用于计算 delta 的相同函数，即 https://github.com/nltk/nltk/blob/develop/nltk/classify/maxent.py#L1208 。您会发现这个增量计算是特定于 IIS(改进的迭代缩放)算法的。
此时，你需要了解机器学习和监督学习，https://en.wikipedia.org/wiki/Supervised_learning
回答你的问题，变暖只是表明在某些时候δ值很难计算，但处理起来仍然合理，可能是因为计算δ值时出现了一些超小的值。 算法正在运行。不是挂，是训练。
为了欣赏 NLTK 中 MaxEnt 的简洁实现，我建议您学习本类(class) https://www.youtube.com/playlist?list=PL6397E4B26D00A269或者了解更多核心机器学习类(class)，请访问https://www.coursera.org/course/ml
训练分类器需要时间和计算能力，在您等待足够长的时间后，您应该会看到它确实如此:
print accuracy(me_classifier, test_feats)
[输出]:
0.5
您可以看到准确度很差，正如预期的那样，因为增量计算太过分了，0.5 是您的基线。完成上面列出的类(class)，在了解它们是如何产生以及如何调整它们后，您应该能够生成更好的分类器。
顺便说一句，记得腌制你的分类器，这样下次就不必重新训练它，请参阅 Save Naive Bayes Trained Classifier in NLTK和 Pickling a trained classifier yields different results from the results obtained directly from a newly but identically trained classifier
完整代码如下:
from itertools import chain
from collections import defaultdict
import nltk
from nltk.classify import MaxentClassifier, accuracy
from nltk.corpus import movie_reviews
def bag_of_words(words):
"""
Change a document into a BOW feature vector represented by a dict object.
"""
return dict([(word, True) for word in words])
def label_feats_from_corpus(corp, feature_detector=bag_of_words):
"""
Change the corpus into a feature matrix. Sometimes the proceess is
known as vectorization. The default is the use BOW features.
"""
label_feats = defaultdict(list)
for label in corp.categories():
for fileid in corp.fileids(categories=[label]):
feats = feature_detector(corp.words(fileids=[fileid]))
label_feats[label].append(feats)
return label_feats
def split_label_feats(lfeats, split=0.75):
"""
Splits corpus into train and test portion.
This module is used after using `label_feats_from_corpus`.
"""
train_feats = []
test_feats = []
for label, feats in lfeats.iteritems():
cutoff = int(len(feats) * split)
train_feats.extend([(feat, label) for feat in feats[:cutoff]])
test_feats.extend([(feat, label) for feat in feats[cutoff:]])
return train_feats, test_feats
# Extract features from corpus and for each document label it with the appropriate labels.
label_feats = label_feats_from_corpus(movie_reviews)
'''
for label in label_feats:
for document in label_feats[label]:
print label, document
break
break
'''
# Let's split the data up into train and test.
train_documents, test_documents = split_label_feats(label_feats)
'''
# Now we see that the `split_label_feats` change the key value structure such that each iteration of train_feats gives us a document with a tuple of the (features, label)
for features, label in train_documents:
print label, features
break
print len(train_documents)
print len(test_documents)
# Get the number of documents in movie_review corpus
num_docs_in_corpus = len(list(chain(*[movie_reviews.fileids(categories=[cat]) for cat in movie_reviews.categories()])))
print len(train_documents) + len(test_documents) == num_docs_in_corpus
'''
# To train the tagger.
me_classifier = nltk.MaxentClassifier.train(train_documents, algorithm='iis', trace=0, max_iter=3)
print accuracy(me_classifier, test_feats)
关于python-2.7 - 我的 Maxent 分类器适用于 gis 算法，但不适用于 iis 算法。它没有抛出任何错误，只是一些警告，我们在Stack Overflow上找到一个类似的问题： https://stackoverflow.com/questions/27246917/