【神经网络结构可视化】使用 Visualkeras 可视化 Keras / TensorFlow 神经网络结构

---

Visualkeras介绍

Visualkeras是一个Python包，用于帮助可视化Keras（独立或包含在tensorflow中）神经网络架构。它允许简单的造型来满足大多数需求。该模块支持分层风格的架构生成，这对CNN（卷积神经网络）非常有用。

---

下载安装

Visualkeras源代码链接：https://github.com/paulgavrikov/visualkeras

使用清华源安装Visualkeras

pip install visualkeras -i https://pypi.tuna.tsinghua.edu.cn/simple

---

代码示例

使用CNN经典网络VGG16作为示例，可视化神经网络结构。

1、导入必要的库

from tensorflow.python.keras.models import Sequential
from tensorflow.python.keras.layers import Dense, Flatten, Conv2D, Dropout, MaxPooling2D, InputLayer, ZeroPadding2D
from collections import defaultdict
import visualkeras
from PIL import ImageFont

2、创建VGG16神经网络模型

# create VGG16
image_size = 224
model = Sequential()
model.add(InputLayer(input_shape=(image_size, image_size, 3)))model.add(ZeroPadding2D((1, 1)))
model.add(Conv2D(64, activation='relu', kernel_size=(3, 3)))
model.add(ZeroPadding2D((1, 1)))
model.add(Conv2D(64, activation='relu', kernel_size=(3, 3)))
model.add(visualkeras.SpacingDummyLayer())model.add(MaxPooling2D((2, 2), strides=(2, 2)))
model.add(ZeroPadding2D((1, 1)))
model.add(Conv2D(128, activation='relu', kernel_size=(3, 3)))
model.add(ZeroPadding2D((1, 1)))
model.add(Conv2D(128, activation='relu', kernel_size=(3, 3)))
model.add(visualkeras.SpacingDummyLayer())model.add(MaxPooling2D((2, 2), strides=(2, 2)))
model.add(ZeroPadding2D((1, 1)))
model.add(Conv2D(256, activation='relu', kernel_size=(3, 3)))
model.add(ZeroPadding2D((1, 1)))
model.add(Conv2D(256, activation='relu', kernel_size=(3, 3)))
model.add(ZeroPadding2D((1, 1)))
model.add(Conv2D(256, activation='relu', kernel_size=(3, 3)))
model.add(visualkeras.SpacingDummyLayer())model.add(MaxPooling2D((2, 2), strides=(2, 2)))
model.add(ZeroPadding2D((1, 1)))
model.add(Conv2D(512, activation='relu', kernel_size=(3, 3)))
model.add(ZeroPadding2D((1, 1)))
model.add(Conv2D(512, activation='relu', kernel_size=(3, 3)))
model.add(ZeroPadding2D((1, 1)))
model.add(Conv2D(512, activation='relu', kernel_size=(3, 3)))
model.add(visualkeras.SpacingDummyLayer())model.add(MaxPooling2D((2, 2), strides=(2, 2)))
model.add(ZeroPadding2D((1, 1)))
model.add(Conv2D(512, activation='relu', kernel_size=(3, 3)))
model.add(ZeroPadding2D((1, 1)))
model.add(Conv2D(512, activation='relu', kernel_size=(3, 3)))
model.add(ZeroPadding2D((1, 1)))
model.add(Conv2D(512, activation='relu', kernel_size=(3, 3)))
model.add(MaxPooling2D())
model.add(visualkeras.SpacingDummyLayer())model.add(Flatten())model.add(Dense(4096, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(4096, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(1000, activation='softmax'))

3、可视化神经网络结构

# Now visualize the model!color_map = defaultdict(dict)
color_map[Conv2D]['fill'] = 'orange'
color_map[ZeroPadding2D]['fill'] = 'gray'
color_map[Dropout]['fill'] = 'pink'
color_map[MaxPooling2D]['fill'] = 'red'
color_map[Dense]['fill'] = 'green'
color_map[Flatten]['fill'] = 'teal'font = ImageFont.truetype("./Arial.ttf", 32)visualkeras.layered_view(model, to_file='./figures/vgg16.png', type_ignore=[visualkeras.SpacingDummyLayer])
visualkeras.layered_view(model, to_file='./figures/vgg16_legend.png', type_ignore=[visualkeras.SpacingDummyLayer],legend=True, font=font)
visualkeras.layered_view(model, to_file='./figures/vgg16_spacing_layers.png', spacing=0)
visualkeras.layered_view(model, to_file='./figures/vgg16_type_ignore.png',type_ignore=[ZeroPadding2D, Dropout, Flatten, visualkeras.SpacingDummyLayer])
visualkeras.layered_view(model, to_file='./figures/vgg16_color_map.png',color_map=color_map, type_ignore=[visualkeras.SpacingDummyLayer])
visualkeras.layered_view(model, to_file='./figures/vgg16_flat.png',draw_volume=False, type_ignore=[visualkeras.SpacingDummyLayer])
visualkeras.layered_view(model, to_file='./figures/vgg16_scaling.png',scale_xy=1, scale_z=1, max_z=1000, type_ignore=[visualkeras.SpacingDummyLayer])

4、完整代码

from tensorflow.python.keras.models import Sequential
from tensorflow.python.keras.layers import Dense, Flatten, Conv2D, Dropout, MaxPooling2D, InputLayer, ZeroPadding2D
from collections import defaultdict
import visualkeras
from PIL import ImageFont# create VGG16
image_size = 224
model = Sequential()
model.add(InputLayer(input_shape=(image_size, image_size, 3)))model.add(ZeroPadding2D((1, 1)))
model.add(Conv2D(64, activation='relu', kernel_size=(3, 3)))
model.add(ZeroPadding2D((1, 1)))
model.add(Conv2D(64, activation='relu', kernel_size=(3, 3)))
model.add(visualkeras.SpacingDummyLayer())model.add(MaxPooling2D((2, 2), strides=(2, 2)))
model.add(ZeroPadding2D((1, 1)))
model.add(Conv2D(128, activation='relu', kernel_size=(3, 3)))
model.add(ZeroPadding2D((1, 1)))
model.add(Conv2D(128, activation='relu', kernel_size=(3, 3)))
model.add(visualkeras.SpacingDummyLayer())model.add(MaxPooling2D((2, 2), strides=(2, 2)))
model.add(ZeroPadding2D((1, 1)))
model.add(Conv2D(256, activation='relu', kernel_size=(3, 3)))
model.add(ZeroPadding2D((1, 1)))
model.add(Conv2D(256, activation='relu', kernel_size=(3, 3)))
model.add(ZeroPadding2D((1, 1)))
model.add(Conv2D(256, activation='relu', kernel_size=(3, 3)))
model.add(visualkeras.SpacingDummyLayer())model.add(MaxPooling2D((2, 2), strides=(2, 2)))
model.add(ZeroPadding2D((1, 1)))
model.add(Conv2D(512, activation='relu', kernel_size=(3, 3)))
model.add(ZeroPadding2D((1, 1)))
model.add(Conv2D(512, activation='relu', kernel_size=(3, 3)))
model.add(ZeroPadding2D((1, 1)))
model.add(Conv2D(512, activation='relu', kernel_size=(3, 3)))
model.add(visualkeras.SpacingDummyLayer())model.add(MaxPooling2D((2, 2), strides=(2, 2)))
model.add(ZeroPadding2D((1, 1)))
model.add(Conv2D(512, activation='relu', kernel_size=(3, 3)))
model.add(ZeroPadding2D((1, 1)))
model.add(Conv2D(512, activation='relu', kernel_size=(3, 3)))
model.add(ZeroPadding2D((1, 1)))
model.add(Conv2D(512, activation='relu', kernel_size=(3, 3)))
model.add(MaxPooling2D())
model.add(visualkeras.SpacingDummyLayer())model.add(Flatten())model.add(Dense(4096, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(4096, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(1000, activation='softmax'))# Now visualize the model!color_map = defaultdict(dict)
color_map[Conv2D]['fill'] = 'orange'
color_map[ZeroPadding2D]['fill'] = 'gray'
color_map[Dropout]['fill'] = 'pink'
color_map[MaxPooling2D]['fill'] = 'red'
color_map[Dense]['fill'] = 'green'
color_map[Flatten]['fill'] = 'teal'font = ImageFont.truetype("./Arial.ttf", 32)visualkeras.layered_view(model, to_file='./figures/vgg16.png', type_ignore=[visualkeras.SpacingDummyLayer])
visualkeras.layered_view(model, to_file='./figures/vgg16_legend.png', type_ignore=[visualkeras.SpacingDummyLayer],legend=True, font=font)
visualkeras.layered_view(model, to_file='./figures/vgg16_spacing_layers.png', spacing=0)
visualkeras.layered_view(model, to_file='./figures/vgg16_type_ignore.png',type_ignore=[ZeroPadding2D, Dropout, Flatten, visualkeras.SpacingDummyLayer])
visualkeras.layered_view(model, to_file='./figures/vgg16_color_map.png',color_map=color_map, type_ignore=[visualkeras.SpacingDummyLayer])
visualkeras.layered_view(model, to_file='./figures/vgg16_flat.png',draw_volume=False, type_ignore=[visualkeras.SpacingDummyLayer])
visualkeras.layered_view(model, to_file='./figures/vgg16_scaling.png',scale_xy=1, scale_z=1, max_z=1000, type_ignore=[visualkeras.SpacingDummyLayer])

5、使用教程

• 创建一个项目文件夹（例如：Project）
• 在创建的项目文件夹Project 中新建一个文件夹（文件夹名为 figures ）
• 通过链接（https://ultralytics.com/assets/Arial.ttf）下载 Arial.ttf 字体文件
• 将下载的 Arial.ttf 字体文件 放在 项目文件夹Project 下
• 在 项目文件夹Project 下新建一个py文件（如：examples.py）
• 将上述的完整代码复制到 examples.py 中
• 运行examples.py
• 在 figures文件夹中查看生成的可视化图
• vgg16.png
  
• vgg16_legend.png
  

---

可视化自己创建的神经网络结构

1、导入要的库

from tensorflow.keras.preprocessing.image import ImageDataGenerator
from tensorflow.keras import models,layers
from tensorflow.keras.layers import Conv2D, MaxPooling2D, BatchNormalization, Flatten, Dense
from tensorflow.keras.callbacks import Callback, ModelCheckpoint
import visualkeras

2、创建自己的神经网络模型

将以下代码替换为自己的Keras / TensorFlow 神经网络结构。

model = models.Sequential()
# 第一层卷积层
model.add(layers.Conv2D(32, (3, 3), activation='relu', input_shape=(48, 48, 1)))  # 假设输入图像大小为48x48，1为灰度图
model.add(layers.MaxPooling2D((2, 2)))
# 第二层卷积层
model.add(layers.Conv2D(64, (3, 3), activation='relu'))
model.add(layers.MaxPooling2D((2, 2)))
# 展平层
model.add(layers.Flatten())
# 全连接层
model.add(layers.Dense(64, activation='relu'))
# 输出层，假设分类任务有7个类别
model.add(layers.Dense(7, activation='softmax'))

3、可视化神经网络结构图

显示层风格图

visualkeras.layered_view(model).show() # 只显示图
# visualkeras.layered_view(model, to_file='output.png').show() # 保存和显示图

显示带有标签的层风格图

from PIL import ImageFont
font = ImageFont.truetype("./Arial.ttf", 32)visualkeras.layered_view(model, legend=True, font=font).show() # 只显示图
# visualkeras.layered_view(model, to_file='output_legend.png', legend=True, font=font).show()  # 保存和显示图

4、完整代码

from tensorflow.keras.preprocessing.image import ImageDataGenerator
from tensorflow.keras import models,layers
from tensorflow.keras.layers import Conv2D, MaxPooling2D, BatchNormalization, Flatten, Dense
from tensorflow.keras.callbacks import Callback, ModelCheckpoint
import visualkeras# 可以将下面这部分创建模型的代码更换你自己的神经网络结构
model = models.Sequential()
# 第一层卷积层
model.add(layers.Conv2D(32, (3, 3), activation='relu', input_shape=(48, 48, 1)))  # 假设输入图像大小为48x48，1为灰度图
model.add(layers.MaxPooling2D((2, 2)))
# 第二层卷积层
model.add(layers.Conv2D(64, (3, 3), activation='relu'))
model.add(layers.MaxPooling2D((2, 2)))
# 展平层
model.add(layers.Flatten())
# 全连接层
model.add(layers.Dense(64, activation='relu'))
# 输出层，假设分类任务有7个类别
model.add(layers.Dense(7, activation='softmax'))visualkeras.layered_view(model).show() # 只显示图
# visualkeras.layered_view(model, to_file='output.png').show() # 保存和显示图from PIL import ImageFont
font = ImageFont.truetype("./Arial.ttf", 32)visualkeras.layered_view(model, legend=True, font=font).show() # 只显示图
# visualkeras.layered_view(model, to_file='output_legend.png', legend=True, font=font).show()  # 保存和显示图