计算keras LSTM中矢量之间的距离和角度

I am using keras to detect similarity between question pairs. The model structure seems to be working fine but it gives me error on model.fit function. I even checked the datatype of my input data and it is numpy.ndarray. Any pointers in this regard and I'll be grateful.

我正在使用keras来检测问题对之间的相似性。模型结构似乎工作正常，但它给了我model.fit函数的错误。我甚至检查了输入数据的数据类型，它是numpy.ndarray。在这方面有任何指示，我将不胜感激。

ValueError: Error when checking model input: the list of Numpy arrays that you are passing to your model is not the size the model expected. Expected to see 1 arrays but instead got the following list of 2 arrays: [array([[ 0, 0, 0, ..., 251, 46, 50], [ 0, 0, 0, ..., 7, 40, 6935], [ 0, 0, 0, ..., 17, 314, 2317], ..., [ 0, ...

ValueError：检查模型输入时出错：您传递给模型的Numpy数组列表不是模型预期的大小。预计会看到1个数组但是得到以下2个数组的列表：[array（[[0,0,0，...，251,46,50]，[0,0,0，...，7， 40,6935]，[0,0,0，...，17,314,2317]，...，[0，...

def Angle(inputs):

    length_input_1=K.sqrt(K.sum(tf.pow(inputs[0],2),axis=1,keepdims=True))
    length_input_2=K.sqrt(K.sum(tf.pow(inputs[1],2),axis=1,keepdims=True))
    result=K.batch_dot(inputs[0],inputs[1],axes=1)/(length_input_1*length_input_2)
    angle = tf.acos(result)
    return angle

def Distance(inputs):

    s = inputs[0] - inputs[1]
    output = K.sum(s ** 2,axis=1,keepdims=True)
    return output    




y=data.is_duplicate.values
tk=text.Tokenizer()
tk.fit_on_texts(list(data.question1.values)+list(data.question2.values))


question1 = tk.texts_to_sequences(data.question1.values)
question1 = sequence.pad_sequences(question1,maxlen=MAX_LEN)

question2 = tk.texts_to_sequences(data.question2.values)
question2 = sequence.pad_sequences(question2,maxlen=MAX_LEN)


word_index = tk.word_index
logging.basicConfig(format='%(asctime)s : %(levelname)s : %(message)s', level=logging.INFO)
num_features = 300
num_workers = multiprocessing.cpu_count()
context_size = 5
downsampling = 7.5e-06
seed = 1
min_word_count = 5
hs = 1
negative = 5


Quora_word2vec = gensim.models.Word2Vec(

    sg=0,
    seed=1,
    workers=num_workers,
    min_count=min_word_count,
    size=num_features,
    window=context_size,  # (2 and 5)
    hs=hs,  # (1 and 0)
    negative=negative,  # (5 and 10)
    sample=downsampling  # (range (0, 1e-5). )

)

Quora_word2vec = gensim.models.KeyedVectors.load_word2vec_format('GoogleNews-vectors-negative300.bin',binary=True)
embedding_matrix=np.zeros((len(word_index)+1,300))

for word , i in tqdm(word_index.items()): #i is index

    try:

        embedding_vector =  Quora_word2vec[word] #Exception is thrown if there is key error
        embedding_matrix[i] = embedding_vector

    except Exception as e:  #If word is not found continue

        continue

--------question1--------

model1 = Sequential()
print "Build Model"

model1.add(Embedding(
    len(word_index)+1,
    300,
    weights=[embedding_matrix],
    input_length=MAX_LEN
    ))

model1.add(SpatialDropout1D(0.2))
model1.add(TimeDistributed(Dense(300, activation='relu')))
model1.add(Lambda(lambda x: K.sum(x, axis=1), output_shape=(300,)))

print model1.summary()

#---------question2-------#

＃ - - - - -问题2 - - - -＃

model2=Sequential()

model2.add(Embedding(
    len(word_index) + 1,
    300,
    weights=[embedding_matrix],
    input_length=MAX_LEN
    ))  # Embedding layer
model2.add(SpatialDropout1D(0.2))
model2.add(TimeDistributed(Dense(300, activation='relu')))
model2.add(Lambda(lambda x: K.sum(x, axis=1), output_shape=(300,)))


print model2.summary()


#---------Merged------#

#Here you get question embedding

#Calculate distance between vectors
Distance_merged_model=Sequential()
Distance_merged_model.add(Merge(layers=[model1, model2], mode=Distance, output_shape=(1,)))

print Distance_merged_model.summary()

#Calculate Angle between vectors

Angle_merged_model=Sequential()
Angle_merged_model.add(Merge(layers=[model1,model2],mode=Angle,output_shape=(1,)))

print Angle_merged_model.summary()


neural_network = Sequential()
neural_network.add(Dense(2,input_shape=(1,)))
neural_network.add(Dense(1))
neural_network.add(Activation('sigmoid'))

print neural_network.summary()


neural_network.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])

checkpoint = ModelCheckpoint('weights.h5', monitor='val_acc', save_best_only=True, verbose=2)

print type(question1)
print type(question2)
neural_network.fit([question1,question2],y=y, batch_size=384, epochs=10,
                 verbose=1, validation_split=0.3, shuffle=True, callbacks=[checkpoint])

1 个解决方案

#1

You're not connecting your final two layers into the dense and just having your neural network be the only network having data passed into since you are compiling and fitting on that layer with out having the distance and angle networks connecting to your final dense.

您没有将最后两层连接到密集区域，只是让您的神经网络成为唯一传入数据的网络，因为您正在编译和拟合该层，而不会将距离和角度网络连接到最终密集区域。

Everything with getting network #1 and #2 seem correct into the merge layers but you need to do something similar to:

获取网络＃1和＃2的所有内容在合并层中都是正确的，但您需要执行类似的操作：

neural_network = Sequential()
neural_network.add(Merge([Distance_merged_model, Angle_merged_model], mode='concat'))
neural_network.add(Dense(2,input_shape=(1,)))
neural_network.add(Dense(1))
neural_network.add(Activation('sigmoid'))