针对图像的经典卷积网络结构进化史及可视化
运行系统：macOS Sequoia 15.0
Python编译器：PyCharm 2024.1.4 (Community Edition)
Python版本：3.12
TensorFlow版本：2.17.0
Pytorch版本：2.4.1
往期链接：
| 1-5 | 6-10 | 11-20 | 21-30 | 31-40 | 41-50 |
|---|
| 51-60：函数 | 61-70：类 | 71-80：编程范式及设计模式 |
|---|
| 81-90：Python编码规范 | 91-100：Python自带常用模块-1 |
|---|
| 101-105：Python自带模块-2 | 106-110：Python自带模块-3 |
|---|
| 111-115：Python常用第三方包-频繁使用 | 116-120：Python常用第三方包-深度学习 |
|---|
| 121-125：Python常用第三方包-爬取数据 | 126-130：Python常用第三方包-为了乐趣 |
|---|
| 131-135：Python常用第三方包-拓展工具1 | 136-140：Python常用第三方包-拓展工具2 |
|---|
Python项目实战
| 141-145 | 146-150 | 151-155 | 156-160 | 161-165 | 166-170 | 171-175 | 176-180 |
|---|
针对图像的经典卷积网络结构进化史及可视化（续）
P181–MobileNet【2017】
模型结构及创新性说明
MobileNet是一系列为移动和嵌入式视觉应用设计的轻量级卷积神经网络。以下是MobileNet各个版本的的主要特点：
（1）MobileNetV1版本
主要特点
- 引入深度可分离卷积（Depthwise Separable Convolution）
- 使用宽度乘子（Width Multiplier）和分辨率乘子（Resolution Multiplier）调整模型大小和复杂度
创新点
- 深度可分离卷积将标准卷积分解为深度卷积和逐点卷积，大大减少了计算量
- 使用ReLU6作为激活函数，有利于低精度计算
（2）MobileNetV2版本
主要特点
- 引入倒置残差结构（Inverted Residual Structure）
- 设计线性瓶颈（Linear Bottleneck）
创新点
- 倒置残差结构先扩展通道数，再做深度卷积，最后压缩回原来的通道数
- 去掉了最后一个ReLU，使用线性激活，有助于保留低维特征
（3）MobileNetV3
主要特点
- 网络结构搜索（NAS）优化的网络架构
- 引入新的激活函数：h-swish
- 集成Squeeze-and-Excitation (SE) 模块
- 提供Small和Large两个版本
创新点
- 使用NAS自动搜索最优网络结构
- h-swish激活函数提高了精度，同时计算效率高
- SE模块增强了特征的表达能力
- 优化了网络的首尾层，进一步提高效率
模型结构代码
MobileNet V1版本
import tensorflow as tf
from tensorflow.keras import layers, models
def depthwise_conv_block(inputs, pointwise_conv_filters, alpha,
depth_multiplier=1, strides=(1, 1), block_id=1):
"""Adds a depthwise convolution block.
A depthwise convolution block consists of a depthwise conv,
batch normalization, ReLU6, pointwise convolution,
batch normalization and ReLU6 activation.
"""
channel_axis = -1
pointwise_conv_filters = int(pointwise_conv_filters * alpha)
x = layers.DepthwiseConv2D((3, 3),
padding='same',
depth_multiplier=depth_multiplier,
strides=strides,
use_bias=False,
name='conv_dw_%d' % block_id)(inputs)
x = layers.BatchNormalization(axis=channel_axis, name='conv_dw_%d_bn' % block_id)(x)
x = layers.ReLU(6., name='conv_dw_%d_relu' % block_id)(x)
x = layers.Conv2D(pointwise_conv_filters, (1, 1),
padding='same',
use_bias=False,
strides=(1, 1),
name='conv_pw_%d' % block_id)(x)
x = layers.BatchNormalization(axis=channel_axis, name='conv_pw_%d_bn' % block_id)(x)
return layers.ReLU(6., name='conv_pw_%d_relu' % block_id)(x)
def MobileNetV1(input_shape=(224, 224, 3),
alpha=1.0,
depth_multiplier=1,
dropout=1e-3,
classes=1000):
"""Instantiates the MobileNet architecture.
Arguments:
input_shape: Optional shape tuple, to be specified if you would
like to use a model with an input img resolution that is not
(224, 224, 3).
alpha: Controls the width of the network. This is known as the
width multiplier in the MobileNet paper.
- If `alpha` < 1.0, proportionally decreases the number
of filters in each layer.
- If `alpha` > 1.0, proportionally increases the number
of filters in each layer.
- If `alpha` = 1, default number of filters from the paper
are used at each layer.
depth_multiplier: Depth multiplier for depthwise convolution.
This is called the resolution multiplier in the MobileNet paper.
dropout: Dropout rate.
classes: Optional number of classes to classify images into.
Returns:
A Keras model instance.
"""
img_input = layers.Input(shape=input_shape)
x = layers.Conv2D(int(32 * alpha), (3, 3),
strides=(2, 2),
padding='same',
use_bias=False,
name='conv1')(img_input)
x = layers.BatchNormalization(axis=-1, name='conv1_bn')(x)
x = layers.ReLU(6., name='conv1_relu')(x)
x = depthwise_conv_block(x, 64, alpha, depth_multiplier, block_id=1)
x = depthwise_conv_block(x, 128, alpha, depth_multiplier, strides=(2, 2), block_id=2)
x = depthwise_conv_block(x, 128, alpha, depth_multiplier, block_id=3)
x = depthwise_conv_block(x, 256, alpha, depth_multiplier, strides=(2, 2), block_id=4)
x = depthwise_conv_block(x, 256, alpha, depth_multiplier, block_id=5)
x = depthwise_conv_block(x, 512, alpha, depth_multiplier, strides=(2, 2), block_id=6)
x = depthwise_conv_block(x, 512, alpha, depth_multiplier, block_id=7)
x = depthwise_conv_block(x, 512, alpha, depth_multiplier, block_id=8)
x = depthwise_conv_block(x, 512, alpha, depth_multiplier, block_id=9)
x = depthwise_conv_block(x, 512, alpha, depth_multiplier, block_id=10)
x = depthwise_conv_block(x, 512, alpha, depth_multiplier, block_id=11)
x = depthwise_conv_block(x, 1024, alpha, depth_multiplier, strides=(2, 2), block_id=12)
x = depthwise_conv_block(x, 1024, alpha, depth_multiplier, block_id=13)
x = layers.GlobalAveragePooling2D()(x)
x = layers.Reshape((1, 1, int(1024 * alpha)))(x)
x = layers.Dropout(dropout, name='dropout')(x)
x = layers.Conv2D(classes, (1, 1),
padding='same',
name='conv_preds')(x)
x = layers.Reshape((classes,), name='reshape_2')(x)
x = layers.Activation('softmax', name='act_softmax')(x)
model = models.Model(img_input, x, name='mobilenet_v1')
return model
# 创建MobileNet V1模型
model = MobileNetV1(input_shape=(224, 224, 3), classes=1000)
# 打印模型摘要
model.summary()
可以通过调整alpha参数来创建不同大小的MobileNetV1模型：
custom_model = MobileNetV1(input_shape=(224, 224, 3), classes=10, alpha=0.75)
custom_model.summary()
这将创建一个稍微窄一些（alpha=0.75）的MobileNet模型，用于10类分类任务。
MobileNet V2版本
import tensorflow as tf
from tensorflow.keras import layers, models
def inverted_residual_block(inputs, filters, stride, expand_ratio, alpha):
input_channels = inputs.shape[-1]
pointwise_filters = int(filters * alpha)
# Expansion phase
x = layers.Conv2D(int(input_channels * expand_ratio), kernel_size=1, padding='same', use_bias=False)(inputs)
x = layers.BatchNormalization()(x)
x = layers.ReLU(6.)(x)
# Depthwise Convolution
x = layers.DepthwiseConv2D(kernel_size=3, strides=stride, padding='same', use_bias=False)(x)
x = layers.BatchNormalization()(x)
x = layers.ReLU(6.)(x)
# Projection
x = layers.Conv2D(pointwise_filters, kernel_size=1, padding='same', use_bias=False)(x)
x = layers.BatchNormalization()(x)
# Residual connection if possible
if stride == 1 and input_channels == pointwise_filters:
return layers.Add()([inputs, x])
return x
def MobileNetV2(input_shape=(224, 224, 3), num_classes=1000, alpha=1.0, include_top=True):
inputs = layers.Input(shape=input_shape)
# First Convolution Layer
x = layers.Conv2D(int(32 * alpha), kernel_size=3, strides=(2, 2), padding='same', use_bias=False)(inputs)
x = layers.BatchNormalization()(x)
x = layers.ReLU(6.)(x)
# Inverted Residual Blocks
x = inverted_residual_block(x, filters=16, stride=1, expand_ratio=1, alpha=alpha)
x = inverted_residual_block(x, filters=24, stride=2, expand_ratio=6, alpha=alpha)
x = inverted_residual_block(x, filters=24, stride=1, expand_ratio=6, alpha=alpha)
x = inverted_residual_block(x, filters=32, stride=2, expand_ratio=6, alpha=alpha)
x = inverted_residual_block(x, filters=32, stride=1, expand_ratio=6, alpha=alpha)
x = inverted_residual_block(x, filters=32, stride=1, expand_ratio=6, alpha=alpha)
x = inverted_residual_block(x, filters=64, stride=2, expand_ratio=6, alpha=alpha)
x = inverted_residual_block(x, filters=64, stride=1, expand_ratio=6, alpha=alpha)
x = inverted_residual_block(x, filters=64, stride=1, expand_ratio=6, alpha=alpha)
x = inverted_residual_block(x, filters=64, stride=1, expand_ratio=6, alpha=alpha)
x = inverted_residual_block(x, filters=96, stride=1, expand_ratio=6, alpha=alpha)
x = inverted_residual_block(x, filters=96, stride=1, expand_ratio=6, alpha=alpha)
x = inverted_residual_block(x, filters=96, stride=1, expand_ratio=6, alpha=alpha)
x = inverted_residual_block(x, filters=160, stride=2, expand_ratio=6, alpha=alpha)
x = inverted_residual_block(x, filters=160, stride=1, expand_ratio=6, alpha=alpha)
x = inverted_residual_block(x, filters=160, stride=1, expand_ratio=6, alpha=alpha)
x = inverted_residual_block(x, filters=320, stride=1, expand_ratio=6, alpha=alpha)
# Last Convolution Layer
x = layers.Conv2D(int(1280 * alpha), kernel_size=1, use_bias=False)(x)
x = layers.BatchNormalization()(x)
x = layers.ReLU(6.)(x)
if include_top:
x = layers.GlobalAveragePooling2D()(x)
x = layers.Dense(num_classes, activation='softmax')(x)
model = models.Model(inputs, x, name='MobileNetV2')
return model
# 创建MobileNet V2模型
model = MobileNetV2(input_shape=(224, 224, 3), num_classes=1000)
# 打印模型摘要
model.summary()
MobileNet V3 版本
Small版本
import tensorflow as tf
from tensorflow.keras import layers, models
class HSwish(layers.Layer):
def call(self, x):
return x * tf.nn.relu6(x + 3) / 6
class HSigmoid(layers.Layer):
def call(self, x):
return tf.nn.relu6(x + 3) / 6
def squeeze_excite_block(inputs, se_ratio=0.25):
x = layers.GlobalAveragePooling2D()(inputs)
filters = inputs.shape[-1]
x = layers.Dense(max(1, int(filters * se_ratio)), activation='relu')(x)
x = layers.Dense(filters, activation=HSigmoid())(x)
x = layers.Reshape((1, 1, filters))(x)
return layers.multiply([inputs, x])
def bneck(inputs, out_channels, exp_channels, kernel_size, stride, se_ratio, activation, alpha=1.0):
x = layers.Conv2D(int(exp_channels * alpha), 1, padding='same', use_bias=False)(inputs)
x = layers.BatchNormalization()(x)
x = activation(x)
x = layers.DepthwiseConv2D(kernel_size, stride, padding='same', use_bias=False)(x)
x = layers.BatchNormalization()(x)
x = activation(x)
if se_ratio:
x = squeeze_excite_block(x, se_ratio)
x = layers.Conv2D(int(out_channels * alpha), 1, padding='same', use_bias=False)(x)
x = layers.BatchNormalization()(x)
if stride == 1 and inputs.shape[-1] == int(out_channels * alpha):
return layers.Add()([inputs, x])
return x
def MobileNetV3Small(input_shape=(224, 224, 3), num_classes=1000, alpha=1.0, include_top=True):
inputs = layers.Input(shape=input_shape)
x = layers.Conv2D(16, 3, strides=2, padding='same', use_bias=False)(inputs)
x = layers.BatchNormalization()(x)
x = HSwish()(x)
x = bneck(x, 16, 16, 3, 2, 0.25, layers.ReLU(), alpha)
x = bneck(x, 24, 72, 3, 2, None, layers.ReLU(), alpha)
x = bneck(x, 24, 88, 3, 1, None, layers.ReLU(), alpha)
x = bneck(x, 40, 96, 5, 2, 0.25, HSwish(), alpha)
x = bneck(x, 40, 240, 5, 1, 0.25, HSwish(), alpha)
x = bneck(x, 40, 240, 5, 1, 0.25, HSwish(), alpha)
x = bneck(x, 48, 120, 5, 1, 0.25, HSwish(), alpha)
x = bneck(x, 48, 144, 5, 1, 0.25, HSwish(), alpha)
x = bneck(x, 96, 288, 5, 2, 0.25, HSwish(), alpha)
x = bneck(x, 96, 576, 5, 1, 0.25, HSwish(), alpha)
x = bneck(x, 96, 576, 5, 1, 0.25, HSwish(), alpha)
x = layers.Conv2D(int(576 * alpha), 1, padding='same', use_bias=False)(x)
x = layers.BatchNormalization()(x)
x = HSwish()(x)
x = layers.GlobalAveragePooling2D()(x)
x = layers.Reshape((1, 1, int(576 * alpha)))(x)
x = layers.Conv2D(int(1024 * alpha), 1, padding='same')(x)
x = HSwish()(x)
if include_top:
x = layers.Conv2D(num_classes, 1, padding='same', activation='softmax')(x)
x = layers.Reshape((num_classes,))(x)
model = models.Model(inputs, x, name='MobileNetV3Small')
return model
# 创建MobileNet V3 Small模型
model = MobileNetV3Small(input_shape=(224, 224, 3), num_classes=1000)
# 打印模型摘要
model.summary()
Large版本
import tensorflow as tf
from tensorflow.keras import layers, models
class HSwish(layers.Layer):
def call(self, x):
return x * tf.nn.relu6(x + 3) / 6
class HSigmoid(layers.Layer):
def call(self, x):
return tf.nn.relu6(x + 3) / 6
def squeeze_excite_block(inputs, se_ratio=0.25):
x = layers.GlobalAveragePooling2D()(inputs)
filters = inputs.shape[-1]
x = layers.Dense(max(1, int(filters * se_ratio)), activation='relu')(x)
x = layers.Dense(filters, activation=HSigmoid())(x)
x = layers.Reshape((1, 1, filters))(x)
return layers.multiply([inputs, x])
def bneck(inputs, out_channels, exp_channels, kernel_size, stride, se_ratio, activation, alpha=1.0):
x = layers.Conv2D(int(exp_channels * alpha), 1, padding='same', use_bias=False)(inputs)
x = layers.BatchNormalization()(x)
x = activation(x)
x = layers.DepthwiseConv2D(kernel_size, stride, padding='same', use_bias=False)(x)
x = layers.BatchNormalization()(x)
x = activation(x)
if se_ratio:
x = squeeze_excite_block(x, se_ratio)
x = layers.Conv2D(int(out_channels * alpha), 1, padding='same', use_bias=False)(x)
x = layers.BatchNormalization()(x)
if stride == 1 and inputs.shape[-1] == int(out_channels * alpha):
return layers.Add()([inputs, x])
return x
def MobileNetV3Large(input_shape=(224, 224, 3), num_classes=1000, alpha=1.0, include_top=True):
inputs = layers.Input(shape=input_shape)
x = layers.Conv2D(16, 3, strides=2, padding='same', use_bias=False)(inputs)
x = layers.BatchNormalization()(x)
x = HSwish()(x)
x = bneck(x, 16, 16, 3, 1, None, layers.ReLU(), alpha)
x = bneck(x, 24, 64, 3, 2, None, layers.ReLU(), alpha)
x = bneck(x, 24, 72, 3, 1, None, layers.ReLU(), alpha)
x = bneck(x, 40, 72, 5, 2, 0.25, layers.ReLU(), alpha)
x = bneck(x, 40, 120, 5, 1, 0.25, layers.ReLU(), alpha)
x = bneck(x, 40, 120, 5, 1, 0.25, layers.ReLU(), alpha)
x = bneck(x, 80, 240, 3, 2, None, HSwish(), alpha)
x = bneck(x, 80, 200, 3, 1, None, HSwish(), alpha)
x = bneck(x, 80, 184, 3, 1, None, HSwish(), alpha)
x = bneck(x, 80, 184, 3, 1, None, HSwish(), alpha)
x = bneck(x, 112, 480, 3, 1, 0.25, HSwish(), alpha)
x = bneck(x, 112, 672, 3, 1, 0.25, HSwish(), alpha)
x = bneck(x, 160, 672, 5, 2, 0.25, HSwish(), alpha)
x = bneck(x, 160, 960, 5, 1, 0.25, HSwish(), alpha)
x = bneck(x, 160, 960, 5, 1, 0.25, HSwish(), alpha)
x = layers.Conv2D(int(960 * alpha), 1, padding='same', use_bias=False)(x)
x = layers.BatchNormalization()(x)
x = HSwish()(x)
x = layers.GlobalAveragePooling2D()(x)
x = layers.Reshape((1, 1, int(960 * alpha)))(x)
x = layers.Conv2D(int(1280 * alpha), 1, padding='same')(x)
x = HSwish()(x)
if include_top:
x = layers.Conv2D(num_classes, 1, padding='same', activation='softmax')(x)
x = layers.Reshape((num_classes,))(x)
model = models.Model(inputs, x, name='MobileNetV3Large')
return model
# 创建MobileNet V3 Large模型
model = MobileNetV3Large(input_shape=(224, 224, 3), num_classes=1000)
# 打印模型摘要
model.summary()
P182–EfficientNet【2019】
模型结构及创新性说明
EfficientNet是由Google研究人员在2019年提出的一系列卷积神经网络模型，旨在提高模型效率和准确性。以下是EfficientNet的主要特点：
模型结构
- 基于MobileNetV2的倒置残差结构
- 使用Squeeze-and-Excitation (SE) 块
- 采用复合缩放方法
创新性：
- 提出了复合缩放方法，同时缩放网络的宽度、深度和分辨率
- 通过神经架构搜索(NAS)优化基础网络结构
- 在同等计算资源下，实现了更高的准确率
模型结构代码
B0版本
import matplotlib.pyplot as plt
import tensorflow as tf
from keras.utils import plot_model
from tensorflow.keras import layers, models
# macos系统显示中文
plt.rcParams['font.sans-serif'] = ['Arial Unicode MS']
def swish(x):
return x * tf.nn.sigmoid(x)
def se_block(inputs, se_ratio):
channels = inputs.shape[-1]
x = layers.GlobalAveragePooling2D()(inputs)
x = layers.Dense(max(1, int(channels * se_ratio)), activation=swish)(x)
x = layers.Dense(channels, activation='sigmoid')(x)
return layers.Multiply()([inputs, x])
def mbconv_block(inputs, out_channels, expand_ratio, stride, kernel_size, se_ratio):
channels = inputs.shape[-1]
x = inputs
# Expansion phase
if expand_ratio != 1:
expand_channels = channels * expand_ratio
x = layers.Conv2D(expand_channels, 1, padding='same', use_bias=False)(x)
x = layers.BatchNormalization()(x)
x = layers.Activation(swish)(x)
# Depthwise Conv
x = layers.DepthwiseConv2D(kernel_size, stride, padding='same', use_bias=False)(x)
x = layers.BatchNormalization()(x)
x = layers.Activation(swish)(x)
# Squeeze and Excitation
if se_ratio:
x = se_block(x, se_ratio)
# Output phase
x = layers.Conv2D(out_channels, 1, padding='same', use_bias=False)(x)
x = layers.BatchNormalization()(x)
if stride == 1 and channels == out_channels:
x = layers.Add()([inputs, x])
return x
def efficientnet(width_coefficient, depth_coefficient, resolution, dropout_rate):
base_architecture = [
# expansion, channels, repeats, stride, kernel_size
[1, 16, 1, 1, 3],
[6, 24, 2, 2, 3],
[6, 40, 2, 2, 5],
[6, 80, 3, 2, 3],
[6, 112, 3, 1, 5],
[6, 192, 4, 2, 5],
[6, 320, 1, 1, 3]
]
inputs = layers.Input(shape=(resolution, resolution, 3))
x = layers.Conv2D(32, 3, strides=2, padding='same', use_bias=False)(inputs)
x = layers.BatchNormalization()(x)
x = layers.Activation(swish)(x)
for i, (expansion, channels, repeats, stride, kernel_size) in enumerate(base_architecture):
channels = int(channels * width_coefficient)
repeats = int(repeats * depth_coefficient)
for j in range(repeats):
x = mbconv_block(x, channels, expansion, stride if j == 0 else 1, kernel_size, se_ratio=0.25)
x = layers.Conv2D(1280, 1, padding='same', use_bias=False)(x)
x = layers.BatchNormalization()(x)
x = layers.Activation(swish)(x)
x = layers.GlobalAveragePooling2D()(x)
if dropout_rate > 0:
x = layers.Dropout(dropout_rate)(x)
outputs = layers.Dense(1000, activation='softmax')(x)
model = tf.keras.Model(inputs, outputs)
return model
# EfficientNet-B0 configuration
def efficientnet_b0():
return efficientnet(
width_coefficient=1.0,
depth_coefficient=1.0,
resolution=224,
dropout_rate=0.2
)
# Create the model
model_b0 = efficientnet_b0()
# Print model summary
model_b0.summary()
# 将模型结构输出到pdf
plot_model(model_b0, to_file='model_b0.pdf', show_shapes=True,
show_layer_names=True)
B1–B7版本
def efficientnet_b1():
return efficientnet(width_coefficient=1.0, depth_coefficient=1.1, resolution=240, dropout_rate=0.2)
def efficientnet_b2():
return efficientnet(width_coefficient=1.1, depth_coefficient=1.2, resolution=260, dropout_rate=0.3)
def efficientnet_b3():
return efficientnet(width_coefficient=1.2, depth_coefficient=1.4, resolution=300, dropout_rate=0.3)
def efficientnet_b4():
return efficientnet(width_coefficient=1.4, depth_coefficient=1.8, resolution=380, dropout_rate=0.4)
def efficientnet_b5():
return efficientnet(width_coefficient=1.6, depth_coefficient=2.2, resolution=456, dropout_rate=0.4)
def efficientnet_b6():
return efficientnet(width_coefficient=1.8, depth_coefficient=2.6, resolution=528, dropout_rate=0.5)
def efficientnet_b7():
return efficientnet(width_coefficient=2.0, depth_coefficient=3.1, resolution=600, dropout_rate=0.5)
扫一扫
22万+
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
