python-3.x - 在 keras 中获取权重

Question

当我试图get_weights在 keras 中获取 CNN 图层的权重时，它会导致错误并说：“张量”对象没有属性“权重”。我看到了 keras 文档，它说使用get_weights命令来处理权重。所以我不知道它有什么问题。我也使用 keras 2.2.4

这是我的代码的一部分：

input_layer = Input(shape=(32,32,3))
conv1 = Conv2D(32,(5,5), activation='relu', padding='same')(input_layer)
conv2 = Conv2D(32,(5,5), activation='relu', padding='same')(conv1)
maxpool1 = MaxPool2D(pool_size=2, padding='same')(conv2)
conv3 = Conv2D(32,(5,5), activation='relu', padding='same')(maxpool1)
conv4 = Conv2D(32,(5,5), activation='relu', padding='same')(conv3)
maxpool2 = MaxPool2D(pool_size=2, padding='same')(conv4)
conv5 = Conv2D(32,(5,5), activation='relu', padding='same')(maxpool2)
flatten1 = Flatten()(conv5)
dense1 = Dense(128, kernel_initializer='random_normal', bias_initializer='zeros')(flatten1)
dense2 = Dense(128,kernel_initializer='random_normal', bias_initializer='zeros')(dense1)
output_layer = Dense(10,activation='softmax',kernel_initializer='random_normal', bias_initializer='zeros')(dense2)
Cifar10_CNN = Model(input_layer, output_layer)

print(Cifar10_CNN.summary())

Cifar10_CNN.compile(optimizer=Adam(lr=0.0001), loss=categorical_crossentropy, metrics=['accuracy'])

conv1_weight_visualization = conv1.get_weights()
plt.imshow(conv1_weight_visualization)

标签： python-3.xtensorflowkerasconv-neural-network

Answer 1

for layer in Cifar10_CNN.layers: 
   print(layer.name, np.array(layer.get_weights()))

您可以像这样获得每一层的权重。