tensorflow - keras.utils.plot_model 生成的图形不正确

Question

我有一个由多个较小模型组成的模型。我想确保一切都正确连接并查看由 keras.utils.plot_model 生成的图表。在那里我偶然发现了一些对我来说不正确的部分。一些连接丢失了，在输入端绘制了一些额外的连接(可能是因为模型的堆叠)。

为什么缺少连接？模型是否正确？图表是否正确？检查信息流是否符合预期的最佳方法是什么？

from tensorflow.keras.utils import plot_model
from tensorflow.keras.models import Model
from tensorflow.keras.layers import Input, Dense, concatenate, Add


n_input_a0 = 10
# A -> n_input_b0 = n_input_c0

n_input_b0 = 20
n_input_b1 = 21
# B -> n_input_c1, n_input_c1

n_input_c0 = 20
n_input_c1 = 31
# C -> n_output

n_output = 3

# A
ip_a0 = Input(shape=(n_input_a0,), name='ip_a0')
x_a = Dense(units=10)(ip_a0)
x_a = Dense(units=n_input_b0)(x_a)
model_a = Model(inputs=ip_a0, outputs=x_a, name='model_a')

# B
ip_b0 = Input(shape=(n_input_b0,), name='ip_b0')
ip_b1 = Input(shape=(n_input_b1,), name='ip_b1')
ip_b0b1 = concatenate([ip_b0, ip_b1])
x_b = Dense(units=10)(ip_b0b1)
x_b = Dense(units=n_output)(x_b)
x_b_left = Dense(units=n_input_c1)(x_b)
x_b_right = Dense(units=n_input_c1)(x_b)
model_b = Model(inputs=[ip_b0, ip_b1], outputs=[x_b_left, x_b_right], name='model_b')

# C
ip_c0 = Input(shape=(n_input_c0,), name='ip_c0')
ip_c1 = Input(shape=(n_input_c1,), name='ip_c1')
ip_c0c1 = concatenate([ip_c0, ip_c1])
x_c = Dense(units=10)(ip_c0c1)
x_c = Dense(units=n_output)(x_c)
model_c = Model(inputs=[ip_c0, ip_c1], outputs=[x_c], name='model_c')


# Combined Model
ip_a0_external = Input(shape=(n_input_a0,), name='ip_a0_external')
ip_b1_external = Input(shape=(n_input_b1,), name='ip_b1_external')

x_a2 = model_a(ip_a0_external)
x_b_left2, x_b_right2 = model_b([x_a2, ip_b1_external])
x_b2 = Add()([x_b_left2, x_b_right2])
x_c2 = model_c([x_a2, x_b2])

model_total = Model(inputs=[ip_a0_external, ip_b1_external], outputs=[x_c2], name='model_total')
plot_model(model_total, expand_nested=True, show_shapes=True, to_file='model.png', dpi=80)

我还查看了 TensorBoard 的输出。结果稍好一些，但也缺少 x_b_left + x_b_right 的添加。

import numpy as np
import tensorflow as tf
tf.compat.v1.disable_eager_execution()
with tf.compat.v1.Session() as sess:
    writer = tf.compat.v1.summary.FileWriter('logs', sess.graph)
    model_total([np.ones((1, n_input_a0)), np.ones((1, n_input_b1))])
    writer.close()

Answer 1

我认为当您将相同的输入传递给模型两次时，您破坏了模型。这是一件奇怪的事情。

• 在这里:x_a2 = model_a(ip_a0)
• 这里:x_b_left2, x_b_right2 = model_b([x_a2, ip_b1])

输入 ip_a0 和 ip_b1 已经是模型的一部分，不应再次调用。

我会这样做(没有检查图表是否得到纠正)

原代码:

from tensorflow.keras.utils import plot_model
from tensorflow.keras.models import Model
from tensorflow.keras.layers import Input, Dense, concatenate, Add


n_input_a0 = 10
# A -> n_input_b0 = n_input_c0

n_input_b0 = 20
n_input_b1 = 21
# B -> n_input_c1, n_input_c1

n_input_c0 = 20
n_input_c1 = 31
# C -> n_output

n_output = 3

# A
ip_a0 = Input(shape=(n_input_a0,), name='ip_a0')
x_a = Dense(units=10)(ip_a0)
x_a = Dense(units=n_input_b0)(x_a)

model_a = Model(inputs=ip_a0, outputs=x_a, name='model_a')

在此处更改:

#don't call the model on its own inputs!!!!
x_a2 = model_a.output

#alternatively, following the same pattern as followed for model B
#ip_a0_external = Input(shape=(n_input_a0,), name='ip_a0_external')
#x_a2 = model_a(ip_a0_external)

原代码:

# B
ip_b0 = Input(shape=(n_input_b0,), name='ip_b0')
ip_b1 = Input(shape=(n_input_b1,), name='ip_b1')
ip_b0b1 = concatenate([ip_b0, ip_b1])
x_b = Dense(units=10)(ip_b0b1)
x_b_temp = Dense(units=n_output)(x_b)
x_b_left = Dense(units=n_input_c1)(x_b_temp)
x_b_right = Dense(units=n_input_c1)(x_b_temp)

model_b = Model(inputs=[ip_b0, ip_b1], outputs=[x_b_left, x_b_right], name='model_b')

在此处更改:

#don't call the model on its own inputs!!
#create a new input for this
ip_b1_external = Input(shape=(n_input_b1,), name='ip_b1_external')
x_b_left2, x_b_right2 = model_b([x_a2, ip_b1_external])

原代码:

x_b2 = Add()([x_b_left2, x_b_right2])

# C
ip_c0 = Input(shape=(n_input_c0,), name='ip_c0')
ip_c1 = Input(shape=(n_input_c1,), name='ip_c1')
ip_c0c1 = concatenate([ip_c0, ip_c1])
x_c = Dense(units=10)(ip_c0c1)
x_c = Dense(units=n_output)(x_c)

model_c = Model(inputs=[ip_c0, ip_c1], outputs=[x_c], name='model_c')

#ok, both inputs are from outside
x_c2 = model_c([x_a2, x_b2])

在此处更改:

# Combined Model
# keep track of what inputs were actual inputs for this model 
# and what inputs were only helpers for the submodels:

model_total = Model(inputs=[model_a.input, ip_b1_external], outputs=[x_c2], name='model_total', )
#alternatively
#model_total = Model(inputs=[ip_a0_external, ip_b1_external], outputs=[x_c2])
plot_model(model_total, expand_nested=True, show_shapes=True, to_file='model.png', dpi=64)