def process(self):
tok = Tokenizer()
# consider entire corpus as text ( train + test text columns )
if self.test_csv:
text = list(self.df.loc[:, self.text_cols].values) + list(
self.test_df.loc[:, self.text_cols])
else:
text = list(self.df.loc[:, self.text_cols].values)
self.tokens = [tok.tokenizer(x) for x in text]
self.vocab = Vocab.create(self.tokens, self.max_vocab, self.min_freq)
self.ntokens = [self.vocab.numericalize(t) for t in self.tokens]
# only full training
if self.valid_pct == 0 and self.test_csv is None:
self.trn_ds = (self.ntokens, self.df.loc[:,
self.label_cols].values)
self.vld_ds = ([], [])
self.test_ds = ([], [])
# holdout
elif self.valid_pct > 0 and self.test_csv is None:
self.trn_ds = (self.ntokens[self.cut:],
self.df.loc[:, self.label_cols].values[self.cut:])
self.vld_ds = (self.ntokens[:self.cut],
self.df.loc[:, self.label_cols].values[:self.cut])
self.tst_ds = ([], [])
# holdout and test prediction
elif self.valid_pct > 0 and self.test_csv is not None:
self.trn_tokens = self.ntokens[:len(self.df)]
self.tst_ds = (self.ntokens[len(self.df):], [])
trn_tokens = self.trn_tokens[self.cut:]
vld_tokens = self.trn_tokens[:self.cut]
self.trn_ds = (trn_tokens,
self.df.loc[:, self.label_cols].values[self.cut:])
self.vld_ds = (vld_tokens,
self.df.loc[:, self.label_cols].values[:self.cut])
# full training and test prediction
else:
self.trn_ds = (self.ntokens[:len(self.df)],
self.df.loc[:, self.label_cols].values)
self.vld_ds = ([], [])
self.tst_ds = (self.ntokens[len(self.df):], [])
return self.vocab, self.trn_ds, self.vld_ds, self.tst_ds
def __init__(self, dataset_name, tokenizer, is_train:bool, data_path, read_data_func=None, is_to_tokens=True):
self.is_train = is_train
self.dataset_name = dataset_name
self.labels_num = config_data[dataset_name].labels_num
self.data = []
self.labels = []
self.data_token = []
self.data_seq = []
self.labels_tensor = []
self.vocab = None
self.tokenizer = tokenizer if tokenizer else Tokenizer('normal', remove_stop_words=False)
self.maxlen = None
if isinstance(data_path, str):
data_path = [data_path]
for path in data_path:
if read_data_func is not None: td, tl = read_data_func(path)
else: td, tl = read_standard_data(path)
self.data += td
self.labels += tl
if is_to_tokens:
self.data2token()
n_h = 1000 ## number of hidden nodes in encoder
n_d = 1000 ## number of hidden nodes in decoder
n_y = dim_word
stochastic = False
verbose = 1
## tokenize text, change to matrix
text = []
with open("data/TED2013.raw.en") as f:
for line in f:
text.append(line)
#text.append(korean_morph(line))
input = Tokenizer(n_words_x)
input.fit_on_texts(text)
seq = input.texts_to_sequences(text, n_sentence, n_maxlen)
n_words_x = input.nb_words
'''
text=[]
with open("data/TED2013.raw.en") as f:
for line in f:
text.append(line)
output=Tokenizer(n_words)
output.fit_on_texts(text)
'''
output = input
#targets=output.texts_to_sequences(text,n_sentence,n_maxlen)
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn import metrics
import matplotlib.pyplot as plt
from preprocess import Tokenizer
import sklearn.cross_validation
from sklearn.linear_model import Perceptron
import reading as rd
# Random Forest # other
score = 0.0
vectorizer_tfidf = TfidfVectorizer(sublinear_tf=True,
max_df=0.5,
tokenizer=Tokenizer(),
lowercase=True,
strip_accents='unicode',
stop_words='english',
ngram_range=(1, 3))
for i in range(0, 5):
Xtrain, Xtest, y_train, y_test = sklearn.cross_validation.train_test_split(
rd.dataset, rd.target, test_size=0.2)
X_train = vectorizer_tfidf.fit_transform(Xtrain)
X_test = vectorizer_tfidf.transform(Xtest)
clf = Perceptron(n_iter=50)
clf.fit(X_train, y_train)
pred = clf.predict(X_test)
from sklearn.feature_extraction.text import HashingVectorizer
#from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn import metrics
import matplotlib.pyplot as plt
from sklearn.linear_model import RidgeClassifier
import sklearn.cross_validation
import reading as rd
from preprocess import Tokenizer
vectorizer_hash = HashingVectorizer(tokenizer=Tokenizer(),
lowercase=True,
strip_accents='unicode',
stop_words='english',
ngram_range=(1, 3))
score = 0.0
for i in range(0, 5):
Xtrain, Xtest, y_train, y_test = sklearn.cross_validation.train_test_split(
rd.dataset, rd.target, test_size=0.2)
X_train = vectorizer_hash.transform(Xtrain)
X_test = vectorizer_hash.transform(Xtest)
#SVM with SGD
clf = RidgeClassifier(tol=1e-2, solver="lsqr")
clf.fit(X_train, y_train)
pred = clf.predict(X_test)
print(metrics.confusion_matrix(y_test, pred))
n_d = 1000 ## number of hidden nodes in decoder
n_y = dim_word
stochastic=False
verbose=1
## tokenize text, change to matrix
text=[]
with open("data/TED2013.raw.en") as f:
for line in f:
text.append(line)
#text.append(korean_morph(line))
input=Tokenizer(n_words)
input.fit_on_texts(text)
seq=input.texts_to_sequences(text,n_sentence,n_maxlen)
n_words_x=input.nb_words
text=[]
with open("data/TED2013.raw.en") as f:
for line in f:
text.append(line)
output=Tokenizer(n_words)
output.fit_on_texts(text)
targets=output.texts_to_sequences(text,n_sentence,n_maxlen)
n_words_y=output.nb_words
n_d = 1000 ## number of hidden nodes in decoder
n_y = dim_word
stochastic=False
verbose=1
## tokenize text, change to matrix
text=[]
with open("data/TED2013.raw.en") as f:
for line in f:
text.append(line)
#text.append(korean_morph(line))
input=Tokenizer(n_words_x)
input.fit_on_texts(text)
seq=input.texts_to_sequences(text,n_sentence,n_maxlen)
n_words_x=input.nb_words
'''
text=[]
with open("data/TED2013.raw.en") as f:
for line in f:
text.append(line)
output=Tokenizer(n_words)
output.fit_on_texts(text)
'''
output=input
#targets=output.texts_to_sequences(text,n_sentence,n_maxlen)
