Question

这是我的代码加载预训练的权重和嵌入矩阵

from __future__ import print_function
import numpy as np
import pandas as pd
import csv, datetime, time, json
from zipfile import ZipFile
from os.path import expanduser, exists
from keras.preprocessing.text import Tokenizer
 from keras.preprocessing.sequence import pad_sequences
  from keras.models import Model
  from keras.layers import Input, TimeDistributed, Dense, Lambda, 
concatenate, Dropout, BatchNormalization
 from keras.layers.embeddings import Embedding
from keras.regularizers import l2
from keras.callbacks import Callback, ModelCheckpoint
from keras.utils.data_utils import get_file
from keras import backend as K
from sklearn.model_selection import train_test_split


**Initialize global variables**
 KERAS_DATASETS_DIR = expanduser('~/.keras/datasets/')
 QUESTION_PAIRS_FILE_URL = 
 'http://qim.ec.quoracdn.net/quora_duplicate_questions.tsv'
 QUESTION_PAIRS_FILE = 'test.csv'
 GLOVE_ZIP_FILE_URL = 'http://nlp.stanford.edu/data/glove.840B.300d.zip'
 GLOVE_ZIP_FILE = 'glove.840B.300d.zip'
 GLOVE_FILE = 'glove.840B.300d.txt'
 Q1_TRAINING_DATA_FILE = 'q1_train.npy'
 Q2_TRAINING_DATA_FILE = 'q2_train.npy'
 LABEL_TRAINING_DATA_FILE = 'label_train.npy'
 WORD_EMBEDDING_MATRIX_FILE = 'word_embedding_matrix.npy'
 NB_WORDS_DATA_FILE = 'nb_words.json'
 MAX_NB_WORDS = 200000
 MAX_SEQUENCE_LENGTH = 25
 EMBEDDING_DIM = 300
 MODEL_WEIGHTS_FILE = 'question_pairs_weights.h5'
 VALIDATION_SPLIT = 0.1
 TEST_SPLIT = 0.1
 RNG_SEED = 13371447
 NB_EPOCHS = 1
 DROPOUT = 0.1
 BATCH_SIZE = 32
 OPTIMIZER = 'adam'



word_embedding_matrix = np.load(open(WORD_EMBEDDING_MATRIX_FILE, 'rb'))
with open(NB_WORDS_DATA_FILE, 'r') as f:
nb_words = json.load(f)['nb_words']

print("Processing", QUESTION_PAIRS_FILE)
question1 = []
question2 = []

with open(KERAS_DATASETS_DIR + QUESTION_PAIRS_FILE, encoding='utf-8') as 
csvfile:
reader = csv.DictReader(csvfile, delimiter=',')
for row in reader:
    question1.append(row['question1'])
    question2.append(row['question2'])

 print('Question pairs: %d' % len(question1))
 T1=len(question1)
 print(T1)
 **Build tokenized word index**
 questions = question1 + question2
 tokenizer = Tokenizer(num_words=MAX_NB_WORDS)
 tokenizer.fit_on_texts(questions)
 question1_word_sequences = tokenizer.texts_to_sequences(question1)
 question2_word_sequences = tokenizer.texts_to_sequences(question2)
 word_index = tokenizer.word_index

 print("Words in index: %d" % len(word_index))

  Prepare word embedding matrix

 Prepare training data tensors
 q1_data = pad_sequences(question1_word_sequences, maxlen=MAX_SEQUENCE_LENGTH)
 q2_data = pad_sequences(question2_word_sequences, maxlen=MAX_SEQUENCE_LENGTH)

  print('Shape of question1 data tensor:', q1_data.shape)
  print('Shape of question2 data tensor:', q2_data.shape)




 **Define the model**
question1 = Input(shape=(MAX_SEQUENCE_LENGTH,))
question2 = Input(shape=(MAX_SEQUENCE_LENGTH,))


q1 = Embedding(nb_words + 1, 
             EMBEDDING_DIM, 
             weights=[word_embedding_matrix], 
             input_length=MAX_SEQUENCE_LENGTH, 
             trainable=False)(question1)
q1 = TimeDistributed(Dense(EMBEDDING_DIM, activation='relu'))(q1)
q1 = Lambda(lambda x: K.max(x, axis=1), output_shape=(EMBEDDING_DIM, ))(q1)

q2 = Embedding(nb_words + 1, 
             EMBEDDING_DIM, 
             weights=[word_embedding_matrix], 
             input_length=MAX_SEQUENCE_LENGTH, 
             trainable=False)(question2)
q2 = TimeDistributed(Dense(EMBEDDING_DIM, activation='relu'))(q2)
q2 = Lambda(lambda x: K.max(x, axis=1), output_shape=(EMBEDDING_DIM, ))(q2)

merged = concatenate([q1,q2])
merged = Dense(200, activation='relu')(merged)
merged = Dropout(DROPOUT)(merged)
merged = BatchNormalization()(merged)
merged = Dense(200, activation='relu')(merged)
merged = Dropout(DROPOUT)(merged)
merged = BatchNormalization()(merged)
merged = Dense(200, activation='relu')(merged)
merged = Dropout(DROPOUT)(merged)
merged = BatchNormalization()(merged)
merged = Dense(200, activation='relu')(merged)
merged = Dropout(DROPOUT)(merged)
merged = BatchNormalization()(merged)

is_duplicate = Dense(1, activation='sigmoid')(merged)

model = Model(inputs=[question1,question2], outputs=is_duplicate)
 model.compile(loss='binary_crossentropy', optimizer=OPTIMIZER, metrics= 
 ['accuracy'])
 model.load_weights(MODEL_WEIGHTS_FILE)

 temp= model.predict([q1_data, q2_data])

 df = pd.DataFrame(temp)
 df.to_csv("hero.csv",header=['is_duplicate'])

它为我提供了这样的输出

test id is_duplicate
0   0.585984
1   0.13437697
2   0.7458109
3   0.6282846
4   0.44692168

但我需要布尔值

 test id    is_duplicate
 0       1
 1       0
 2       1
 3       1

不要告诉我整理价值

我希望神经网络本身给我布尔值

是否有可能像那样训练网络

如果是，请建议我可以在代码中包含的内容

提前致谢

Answer 1

尝试：

temp= model.predict([q1_data, q2_data])
y_classes = temp.argmax(axis=-1)
df = pd.DataFrame(y_classes)

如下文所述：

Get class labels from Keras functional model

另外，为什么你不能围绕价值观？

Answer 2

网络无法输出普通布尔值，您只需要对网络输出进行阈值处理，因为它具有sigmoid激活。假设您将阈值设置为0.5，任何超过0.5的值都将为1级，而低于0.5的输出将为0级。

此外，网络输出值可以让您对网络输出的不确定性进行“置信度”估计。

keras神经网络给出浮动输出，但我需要布尔输出

2 个答案: