Question

当其中一个图层具有其尺寸（无，512）而另一个图层具有尺寸（18577,4）时，如何在keras中连接2个图层。我尝试使用Concatenate

concat_layer = Concatenate()([z1,agp]

但这告诉我一个错误：

ValueError: A `Concatenate` layer requires inputs with matching shapes except for the concat axis. Got inputs shapes: [(None, 512), (18577, 4)]

模型看起来像这样：

a1= (Convolution2D(32, filter_dim, activation='linear', 
                    padding='same',kernel_regularizer=regularizers.l2(reg)))(input_img) 
b1 = (BatchNormalization())(a1)
c1 = (PReLU())(b1)
d1 = (Convolution2D(32, filter_dim, activation='linear',kernel_regularizer=regularizers.l2(reg)))(c1)
e1 = (BatchNormalization())(d1)
f1 = (PReLU())(e1)
g1 = (MaxPooling2D(pool_size=(2,2)))(f1)
h1 = (Dropout(0.2))(g1)

i1= (Convolution2D(64, filter_dim, activation='linear', padding='same',kernel_regularizer=regularizers.l2(reg)))(h1)
j1 = (BatchNormalization())(i1)
k1 = (PReLU())(j1)
l1 = (Convolution2D(64, filter_dim, activation='linear',kernel_regularizer=regularizers.l2(reg)))(k1)
m1 = (BatchNormalization())(k1)
n1 = (PReLU())(m1)
o1 = (MaxPooling2D(pool_size=(2,2)))(n1)
p1 = (Dropout(0.2))(o1)

q1= (Convolution2D(128, filter_dim, activation='linear', padding='same',kernel_regularizer=regularizers.l2(reg)))(p1)
r1=q1
s1 = (BatchNormalization())(r1)
t1 = (PReLU())(s1)
u1 = (Convolution2D(128, filter_dim, activation='linear',kernel_regularizer=regularizers.l2(reg)))(t1)
v1 = (BatchNormalization())(u1)
w1 = (PReLU())(v1)
x1 = (MaxPooling2D(pool_size=(3,3)))(w1)
y1 = (Dropout(0.2))(x1)



z1 = (Flatten())(y1)
agp=tf.convert_to_tensor(agp,np.float32)
z1 = Concatenate(axis=1)([z1,agp])

a2 = (Dense(128, activation='linear',kernel_regularizer=regularizers.l2(reg)))(z1)
b2 = (BatchNormalization())(a2)
c2 = (PReLU())(b2)
d2 = (Dropout(0.2))(c2)

e2 = (Dense(32, activation='linear',kernel_regularizer=regularizers.l2(reg)))(d2)
f2 = (BatchNormalization())(e2)
g2 = (PReLU())(f2)
h2 = (Dropout(0.3))(g2)

我的输入图片有尺寸（32,32,3）。我想将z1（无，512）与agp（18577,4）

连接起来

Answer 1

#!/usr/bin/env python


def create_model(nb_classes, input_shape):
    """Create a NN model."""
    # from keras.layers import Dropout
    from keras.layers import Activation, Input
    from keras.layers import Dense, Concatenate
    from keras.models import Model

    input_ = Input(shape=input_shape)
    x = input_

    # Branch in two directions - this can be more
    # complex, of course
    x1 = Dense(512, activation='relu')(x)
    x2 = Dense(4, activation='relu')(x)

    # And this is how you use concatenation
    x = Concatenate(axis=-1)([x1, x2])

    # And then finish it
    x = Dense(nb_classes, activation='softmax')(x)
    model = Model(inputs=input_, outputs=x)
    return model


model = create_model(10, (512, ))
print(model.summary())

给出

Using TensorFlow backend.
____________________________________________________________________________________________________
Layer (type)                     Output Shape          Param #     Connected to                     
====================================================================================================
input_1 (InputLayer)             (None, 512)           0                                            
____________________________________________________________________________________________________
dense_1 (Dense)                  (None, 512)           262656      input_1[0][0]                    
____________________________________________________________________________________________________
dense_2 (Dense)                  (None, 4)             2052        input_1[0][0]                    
____________________________________________________________________________________________________
concatenate_1 (Concatenate)      (None, 516)           0           dense_1[0][0]                    
                                                                   dense_2[0][0]                    
____________________________________________________________________________________________________
dense_3 (Dense)                  (None, 10)            5170        concatenate_1[0][0]              
====================================================================================================
Total params: 269,878
Trainable params: 269,878
Non-trainable params: 0
____________________________________________________________________________________________________
None

在Keras（无，512）和（18577,4）中连接图层

1 个答案: