使用 Keras 绘制神经网络结构图¶
使用keras模块的plot_model绘制神经网络结构图，将神经网络可视化能够帮助理解模型是如何运行的。
基本代码¶
导入搭建模型和绘制神经网络结构的包¶
Python
from keras.utils import plot_model
from tensorflow.keras import backend as Keras
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import (
Conv2D,
MaxPool2D,
Activation,
Dropout,
Flatten,
Dense,
)
定义搭建模型的函数¶
Python
# 定义添加卷积层的函数，可以添加 input_shape 参数
def add_conv_layer(model, filters, kernel_size, input_shape=None):
if input_shape:
model.add(
Conv2D(
filters=filters,
kernel_size=kernel_size,
padding="same",
activation="relu",
input_shape=input_shape,
)
)
else:
model.add(
Conv2D(
filters=filters,
kernel_size=kernel_size,
padding="same",
activation="relu",
)
)
return model
# 定义搭建模型的函数
def build_model(
block_num,
layer_num,
filters_in_each_block,
kernel_size,
pool_size,
strides,
dropout_rate,
x_train,
):
# 清除 Keras 的 session
Keras.clear_session()
# 定义一个空的 Sequential 模型
model = Sequential()
# 添加所有的 block
for block in range(1, block_num + 1):
# 第一个 Block，需要指定输入的形状
if block == 1:
# 该 block 的 filters
filters = filters_in_each_block[block - 1]
# 添加第一个卷积层，需要指定输入的形状
model = add_conv_layer(
model,
filters=filters,
kernel_size=kernel_size,
input_shape=x_train.shape[1:],
)
# 添加除第一个卷积层之外的其他卷积层
for layer in range(2, layer_num + 1):
model = add_conv_layer(model, filters=filters, kernel_size=kernel_size)
# 添加池化层
model.add(MaxPool2D(pool_size=pool_size, strides=strides))
# 其他 Block
else:
# 该 block 的 filters
filters = filters_in_each_block[block - 1]
# 添加所有卷积层
for layer in range(1, layer_num + 1):
model = add_conv_layer(model, filters=filters, kernel_size=kernel_size)
# 添加池化层
model.add(MaxPool2D(pool_size=pool_size, strides=strides))
# 添加 Flatten 层
model.add(Flatten())
# 添加 Dropout 层
model.add(Dropout(dropout_rate))
# 添加 Dense 层
model.add(Dense(7, activation="softmax"))
return model
搭建模型¶
Python
model = build_model(
block_num=4,
layer_num=2,
filters_in_each_block=[64, 128, 256, 512],
kernel_size=3,
pool_size=4,
strides=2,
dropout_rate=0.2,
x_train=x_train,
)
使用model.summary()查看网络结构¶
Text Only
Model: "sequential"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d (Conv2D) (None, 48, 48, 64) 640
conv2d_1 (Conv2D) (None, 48, 48, 64) 36928
max_pooling2d (MaxPooling2D (None, 23, 23, 64) 0
)
conv2d_2 (Conv2D) (None, 23, 23, 128) 73856
conv2d_3 (Conv2D) (None, 23, 23, 128) 147584
max_pooling2d_1 (MaxPooling (None, 10, 10, 128) 0
2D)
conv2d_4 (Conv2D) (None, 10, 10, 256) 295168
conv2d_5 (Conv2D) (None, 10, 10, 256) 590080
max_pooling2d_2 (MaxPooling (None, 4, 4, 256) 0
2D)
conv2d_6 (Conv2D) (None, 4, 4, 512) 1180160
conv2d_7 (Conv2D) (None, 4, 4, 512) 2359808
max_pooling2d_3 (MaxPooling (None, 1, 1, 512) 0
2D)
flatten (Flatten) (None, 512) 0
dropout (Dropout) (None, 512) 0
dense (Dense) (None, 7) 3591
=================================================================
Total params: 4,687,815
Trainable params: 4,687,815
Non-trainable params: 0
_________________________________________________________________
我们已经得到了表格样式的模型结构，可以清楚地知道神经网络有多少层、每一层的输入输出形状这些关键信息。下面用plot_model()绘制神经网络结构，通过可视化的方式进一步帮助我们理解神经网络。
使用plot_model()绘制神经网络结构¶
指定输出网络形状，且输出到本地文件¶
报错解决方法¶
如果提示
Text Only
('You must install pydot (`pip install pydot`) and install graphviz (see instructions at https://graphviz.gitlab.io/download/) ', 'for plot_model/model_to_dot to work.')
可以将keras绘图的源代码中的pydot改成pydotplus。具体来说，需要找到以下地址：
进入vis_utils.py，如下图所示：
将pydot基本上全部改成pydotplus（除了开始的少数代码不用改，因为有些代码原来就是pydotplus，最好保留）之后，即可正常导入包。
本报错解决方法参考了CSDN 的文章。