Я построил модель классификации текста для несбалансированных данных классификации классов.Вместо использования вектора слов keras я использовал встраивание, используя вектор googlenews word2vec в качестве базовой линии в слое встраивания.
import pandas as pd
from keras.models import Sequential, Model
from keras.layers import Dense, Dropout, Activation, Embedding, SpatialDropout1D, Bidirectional, LSTM, Input, concatenate, Conv1D, GlobalMaxPooling1D, BatchNormalization
from keras.optimizers import SGD, Adam
from sklearn.model_selection import train_test_split
from keras.preprocessing.text import Tokenizer
from keras.preprocessing import sequence
import keras.backend as K
from keras import backend as K
from keras import metrics
import numpy as np
from itertools import chain
from collections import Counter
from sklearn.utils import shuffle
import nltk
import gensim
from gensim.models import KeyedVectors
from sklearn.utils import class_weight
dat = pd.read_csv('/home/data.csv',encoding='latin',delimiter='\t')
dat = shuffle(dat)
dat.reset_index(drop=True,inplace=True)
Поскольку это проблема дисбаланса классов, я использовал метрику f1.
def f1_metric(y_true, y_pred):
def recall(y_true, y_pred):
"""Recall metric.
Only computes a batch-wise average of recall.
Computes the recall, a metric for multi-label classification of
how many relevant items are selected.
"""
true_positives = K.sum(K.round(K.clip(y_true * y_pred, 0, 1)))
possible_positives = K.sum(K.round(K.clip(y_true, 0, 1)))
recall = true_positives / (possible_positives + K.epsilon())
return recall
def precision(y_true, y_pred):
"""Precision metric.
Only computes a batch-wise average of precision.
Computes the precision, a metric for multi-label classification of
how many selected items are relevant.
"""
true_positives = K.sum(K.round(K.clip(y_true * y_pred, 0, 1)))
predicted_positives = K.sum(K.round(K.clip(y_pred, 0, 1)))
precision = true_positives / (predicted_positives + K.epsilon())
return precision
precision = precision(y_true, y_pred)
recall = recall(y_true, y_pred)
return 2*((precision*recall)/(precision+recall+K.epsilon()))
Я обработал текст и создал вектор слов, как показано ниже
def preprocess(dat):
return [nltk.word_tokenize(row) for row in dat]
x_train, x_test, y_train, y_test= train_test_split(dat.text,dat.labels,test_size=0.20)
X = preprocess(x_train)
model = KeyedVectors.load_word2vec_format('/home/user/Downloads/GoogleNews-vectors-negative300.bin', binary=True,limit=100000)
Я использую эту функцию для преобразования массива текста в числовые значения из модели word2vec.
def word2idx(word):
return model.wv.vocab[word].index
vocab_size, emdedding_size = model.wv.syn0.shape
pretrained_weights = model.wv.syn0
print(vocab_size, emdedding_size)
100000 300
Я создалматрица
max_sentence_len = 50
train_x = np.zeros([len(X), max_sentence_len], dtype=np.int32)
И заменить 0 на значения индекса из модели word2vec для соответствующих токенизированных слов, максимум до 50 слов.
for i in range(len(X)):
for j in range(len(X[i])):
try:
train_x[i][j] = word2idx(X[i][j])
except:
pass
Я вычислил вес класса, используя функцию sklearn, так как этопроблема дисбаланса класса.
class_weights = class_weight.compute_class_weight('balanced',np.unique(y_train),y_train)
Это функция для создания модели multiConvnet.
def model_architecture(vocab_size,emdedding_size,pretrained_weights):
# vector-space embedding:
n_dim = 64
n_unique_words = 5000
max_review_length = 50
pad_type = trunc_type = 'pre'
drop_embed = 0.2
# convolutional layer architecture:
n_conv_1 = n_conv_2 = n_conv_3 = n_conv_4= 256
k_conv_1 = 3
k_conv_2 = 2
k_conv_3 = 4
k_conv_4 = 5
# dense layer architecture:
n_dense = 256
dropout = 0.2
input_layer = Input(shape=(max_review_length,), dtype='int16', name='input') # supports integers +/- 32.7k
# embedding_layer = Embedding(n_unique_words, n_dim, input_length=max_review_length, name='embedding')(input_layer)
embedding_layer = Embedding(input_dim=vocab_size, output_dim=emdedding_size, weights=[pretrained_weights], name='embedding')(input_layer)
drop_embed_layer = SpatialDropout1D(drop_embed, name='drop_embed')(embedding_layer)
conv_1 = Conv1D(n_conv_1, k_conv_1, activation='relu', name='conv_1')(drop_embed_layer)
maxp_1 = GlobalMaxPooling1D(name='maxp_1')(conv_1)
conv_2 = Conv1D(n_conv_2, k_conv_2, activation='relu', name='conv_2')(drop_embed_layer)
maxp_2 = GlobalMaxPooling1D(name='maxp_2')(conv_2)
conv_3 = Conv1D(n_conv_3, k_conv_3, activation='relu', name='conv_3')(drop_embed_layer)
maxp_3 = GlobalMaxPooling1D(name='maxp_3')(conv_3)
concat = concatenate([maxp_1, maxp_2, maxp_3])
dense_layer = Dense(n_dense, activation='relu', name='dense')(concat)
drop_dense_layer = Dropout(dropout, name='drop_dense')(dense_layer)
dense_2 = Dense(64, activation='relu', name='dense_2')(drop_dense_layer)
dropout_2 = Dropout(dropout, name='drop_dense_2')(dense_2)
predictions = Dense(units=1, activation='sigmoid', name='output')(dropout_2)
model = Model(input_layer, predictions)
model.compile(loss='binary_crossentropy', optimizer='adam', metrics=[f1_metric])
return model
Моя модель ниже
mod_keras = model_architecture(vocab_size,emdedding_size,pretrained_weights)
mod_keras.fit(train_x,y_train,batch_size=32,epochs=2,verbose=1,validation_split=0.2,class_weight=class_weights)
, когда я ее запускаю,я получаю ошибку ниже.
Train on 287895 samples, validate on 71974 samples
Epoch 1/2
---------------------------------------------------------------------------
InvalidArgumentError Traceback (most recent call last)
<ipython-input-25-fcb6fa008311> in <module>
----> 1 mod_Access.fit(train_x,y_train_Access,batch_size=32,epochs=2,verbose=1,validation_split=0.2,class_weight=class_weights)
~/.local/lib/python3.5/site-packages/keras/engine/training.py in fit(self, x, y, batch_size, epochs, verbose, callbacks, validation_split, validation_data, shuffle, class_weight, sample_weight, initial_epoch, steps_per_epoch, validation_steps, **kwargs)
1037 initial_epoch=initial_epoch,
1038 steps_per_epoch=steps_per_epoch,
-> 1039 validation_steps=validation_steps)
1040
1041 def evaluate(self, x=None, y=None,
~/.local/lib/python3.5/site-packages/keras/engine/training_arrays.py in fit_loop(model, f, ins, out_labels, batch_size, epochs, verbose, callbacks, val_f, val_ins, shuffle, callback_metrics, initial_epoch, steps_per_epoch, validation_steps)
197 ins_batch[i] = ins_batch[i].toarray()
198
--> 199 outs = f(ins_batch)
200 outs = to_list(outs)
201 for l, o in zip(out_labels, outs):
~/.local/lib/python3.5/site-packages/keras/backend/tensorflow_backend.py in __call__(self, inputs)
2713 return self._legacy_call(inputs)
2714
-> 2715 return self._call(inputs)
2716 else:
2717 if py_any(is_tensor(x) for x in inputs):
~/.local/lib/python3.5/site-packages/keras/backend/tensorflow_backend.py in _call(self, inputs)
2673 fetched = self._callable_fn(*array_vals, run_metadata=self.run_metadata)
2674 else:
-> 2675 fetched = self._callable_fn(*array_vals)
2676 return fetched[:len(self.outputs)]
2677
~/.local/lib/python3.5/site-packages/tensorflow/python/client/session.py in __call__(self, *args, **kwargs)
1437 ret = tf_session.TF_SessionRunCallable(
1438 self._session._session, self._handle, args, status,
-> 1439 run_metadata_ptr)
1440 if run_metadata:
1441 proto_data = tf_session.TF_GetBuffer(run_metadata_ptr)
~/.local/lib/python3.5/site-packages/tensorflow/python/framework/errors_impl.py in __exit__(self, type_arg, value_arg, traceback_arg)
526 None, None,
527 compat.as_text(c_api.TF_Message(self.status.status)),
--> 528 c_api.TF_GetCode(self.status.status))
529 # Delete the underlying status object from memory otherwise it stays alive
530 # as there is a reference to status from this from the traceback due to
InvalidArgumentError: indices[26,0] = -3338 is not in [0, 100000)
[[{{node embedding/embedding_lookup}} = GatherV2[Taxis=DT_INT32, Tindices=DT_INT32, Tparams=DT_FLOAT, _class=["loc:@training/Adam/Assign_2"], _device="/job:localhost/replica:0/task:0/device:CPU:0"](embedding/embeddings/read, embedding/Cast, training/Adam/gradients/embedding/embedding_lookup_grad/concat/axis)]]
Я прочитал это сообщение InvalidArgumentError (см. выше для отслеживания): индексы [1] = 10 не в [0, 10)
Согласно этому посту внужно установить словарный запас.В моем случае это именно то, что я сделал, используя параметр vocab_size.