我想在Keras中创建一个自定义图层。 在此示例中,我使用变量将张量相乘,但出现了
错误在/keras/engine/training_arrays.py中,第304行,在predict_loop中 outs [i] [batch_start:batch_end] = batch_out ValueError:无法将输入数组从形状(36)广播到形状(2)。
实际上我已经检查了这个文件,但是我什么也没得到。我的自定义层有问题吗?
#the definition of mylayer.
from keras import backend as K
import keras
from keras.engine.topology import Layer
class mylayer(Layer):
def __init__(self, output_dim, **kwargs):
self.output_dim = output_dim
super(mylayer, self).__init__(**kwargs)
def build(self, input_shape):
self.kernel = self.add_weight(name = 'kernel',
shape=(1,),dtype='float32',trainable=True,initializer='uniform')
super(mylayer, self).build(input_shape)
def call(self, inputs, **kwargs):
return self.kernel * inputs[0]
def compute_output_shape(self, input_shape):
return (input_shape[0], input_shape[1])
#the test of mylayer.
from mylayer import mylayer
from tensorflow import keras as K
import numpy as np
from keras.layers import Input, Dense, Flatten
from keras.models import Model
x_train = np.random.random((2, 3, 4, 3))
y_train = np.random.random((2, 36))
print(x_train)
x = Input(shape=(3, 4, 3))
y = Flatten()(x)
output = mylayer((36, ))(y)
model = Model(inputs=x, outputs=output)
model.summary()
model.compile(optimizer='Adam',loss='categorical_crossentropy',metrics=['accuracy'])
model.fit(x_train, y_train, epochs=2)
hist = model.predict(x_train,batch_size=2)
print(hist)
print(model.get_layer(index=1).get_weights())
#So is there some wrong in my custom error?
特别是,当我训练此网络时,没关系,但是当我尝试使用“ prdict”时,这是错误的。
答案 0 :(得分:0)
您的self.kernel * inputs[0]
形状为(36,)
,但您的期望为(?,36)
。更改它:
def call(self, inputs, **kwargs):
return self.kernel * inputs
如果要输出mylayer
的权重,则应设置index=2
。