如何用keras functional查看神经网络的摘要api

Question

我有一个非常大的神经网络，我正在使用 keras 函数 api 制作它。我想监视在尚未定义的模型中添加的层的参数和形状，例如 model.summary().

如果我有这样的模型

input_img = Input(shape=(256, 256, 3))

tower_1 = Conv2D(64, (1, 1), padding='same', activation='relu')(input_img)
tower_1 = Conv2D(64, (3, 3), padding='same', activation='relu')(tower_1)
#stage1

tower_2 = Conv2D(64, (1, 1), padding='same', activation='relu')(input_img)
tower_2 = Conv2D(64, (5, 5), padding='same', activation='relu')(tower_2)
#stage2

tower_3 = MaxPooling2D((3, 3), strides=(1, 1), padding='same')(input_img)
tower_3 = Conv2D(64, (1, 1), padding='same', activation='relu')(tower_3)
#stage3

output = keras.layers.concatenate([tower_1, tower_2, tower_3], axis=1)

我想要 summary() 这个不断发展的模型在这些不同的阶段。我知道我们可以通过定义 model=Model(input,output) 来做到 model.summary() 但是我们可以在层层推进的过程中做到这一点吗？

Answer 1

一个丑陋的答案（因为我没有漂亮的答案）

您可以定义任意多个模型并对其进行总结：

input_img = Input(shape=(256, 256, 3))

tower_1 = Conv2D(64, (1, 1), padding='same', activation='relu')(input_img)
tower_2 = Conv2D(64, (1, 1), padding='same', activation='relu')(input_img)
tower_3 = MaxPooling2D((3, 3), strides=(1, 1), padding='same')(input_img)

model = Model(input_img, [tower1,tower2,tower3])
model.summary()


tower_1 = Conv2D(64, (3, 3), padding='same', activation='relu')(tower_1)
tower_2 = Conv2D(64, (5, 5), padding='same', activation='relu')(tower_2)
tower_3 = Conv2D(64, (1, 1), padding='same', activation='relu')(tower_3)

model = Model(input_img, [tower1,tower2,tower3])
model.summary()

output = keras.layers.concatenate([tower_1, tower_2, tower_3], axis=1)

model = Model(input_img, output)
model.summary()

Answer 2

没有太多的声望点可以发表评论。

您必须参考 TENSORBOARD 以获得与深度学习相关的各种可视化并且也易于使用。

Answer 3

您可以使用 _keras_shape 成员变量轻松获得任何 Keras 张量的编译时形状，例如：

input_img = Input(shape=(256, 256, 3))

tower_1 = Conv2D(64, (1, 1), padding='same', activation='relu')(input_img)
tower_1 = Conv2D(64, (3, 3), padding='same', activation='relu')(tower_1)
#stage1

tower_2 = Conv2D(64, (1, 1), padding='same', activation='relu')(input_img)
tower_2 = Conv2D(64, (5, 5), padding='same', activation='relu')(tower_2)
#stage2

tower_3 = MaxPooling2D((3, 3), strides=(1, 1), padding='same')(input_img)
tower_3 = Conv2D(64, (1, 1), padding='same', activation='relu')(tower_3)
#stage3

output = keras.layers.concatenate([tower_1, tower_2, tower_3], axis=1)

print("Output shape is: {}".format(output._keras_shape))

只要您有 TensorVariable（层的输出），您就可以在计算的任何时候执行此操作。它与完整摘要不同，但它对调试有很大帮助。

如何用keras functional查看神经网络的摘要api

How to view the summary of neural networks with keras functional api

python-2.7

keras

keras-layer