1）ResNet
（1）Deeper Bottleneck Architectures
bottleneck架构设计的目的就在于减少参数数量，进而缩短训练时间。注意bottleneck只在resnet50、resnet101、resnet152中使用，resnet18与resnet34仍使用两层3*3卷积堆叠的设计。
（2）对于channel不同的卷积层之间，使用步长为2的卷积操作。反之直接使用相加操作。
以keras resnet50实现为例：
def identity_block(input_tensor, kernel_size, filters, stage, block):
"""The identity block is the block that has no conv layer at shortcut.
# Arguments
input_tensor: input tensor
kernel_size: default 3, the kernel size of
middle conv layer at main path
filters: list of integers, the filters of 3 conv layer at main path
stage: integer, current stage label, used for generating layer names
block: 'a','b'..., current block label, used for generating layer names
# Returns
Output tensor for the block.
"""
filters1, filters2, filters3 = filters
if backend.image_data_format() == 'channels_last':
bn_axis = 3
else:
bn_axis = 1
conv_name_base = 'res' + str(stage) + block + '_branch'
bn_name_base = 'bn' + str(stage) + block + '_branch'
x = layers.Conv2D(filters1, (1, 1),
name=conv_name_base + '2a')(input_tensor)
x = layers.BatchNormalization(axis=bn_axis, name=bn_name_base + '2a')(x)
x = layers.Activation('relu')(x)
x = layers.Conv2D(filters2, kernel_size,
padding='same', name=conv_name_base + '2b')(x)
x = layers.BatchNormalization(axis=bn_axis, name=bn_name_base + '2b')(x)
x = layers.Activation('relu')(x)
x = layers.Conv2D(filters3, (1, 1), name=conv_name_base + '2c')(x)
x = layers.BatchNormalization(axis=bn_axis, name=bn_name_base + '2c')(x)
x = layers.add([x, input_tensor])
x = layers.Activation('relu')(x)
return xidentity block是维度一致的res block，最后将输出与input_tensor直接相加。
def conv_block(input_tensor,
kernel_size,
filters,
stage,
block,
strides=(2, 2)):
"""A block that has a conv layer at shortcut.
# Arguments
input_tensor: input tensor
kernel_size: default 3, the kernel size of
middle conv layer at main path
filters: list of integers, the filters of 3 conv layer at main path
stage: integer, current stage label, used for generating layer names
block: 'a','b'..., current block label, used for generating layer names
strides: Strides for the first conv layer in the block.
# Returns
Output tensor for the block.
Note that from stage 3,
the first conv layer at main path is with strides=(2, 2)
And the shortcut should have strides=(2, 2) as well
"""
filters1, filters2, filters3 = filters
if backend.image_data_format() == 'channels_last':
bn_axis = 3
else:
bn_axis = 1
conv_name_base = 'res' + str(stage) + block + '_branch'
bn_name_base = 'bn' + str(stage) + block + '_branch'
x = layers.Conv2D(filters1, (1, 1), strides=strides,
name=conv_name_base + '2a')(input_tensor)
x = layers.BatchNormalization(axis=bn_axis, name=bn_name_base + '2a')(x)
x = layers.Activation('relu')(x)
x = layers.Conv2D(filters2, kernel_size, padding='same',
name=conv_name_base + '2b')(x)
x = layers.BatchNormalization(axis=bn_axis, name=bn_name_base + '2b')(x)
x = layers.Activation('relu')(x)
x = layers.Conv2D(filters3, (1, 1), name=conv_name_base + '2c')(x)
x = layers.BatchNormalization(axis=bn_axis, name=bn_name_base + '2c')(x)
shortcut = layers.Conv2D(filters3, (1, 1), strides=strides,
name=conv_name_base + '1')(input_tensor)
shortcut = layers.BatchNormalization(
axis=bn_axis, name=bn_name_base + '1')(shortcut)
x = layers.add([x, shortcut])
x = layers.Activation('relu')(x)
return x可以看到convblock中shortcut使用了一个在h、w两个方向上strides均为2的1*1卷积。最后同样包括相加操作。
2）mobilenet
（1）V1
将标准卷积分解为深度卷积（depthwise convolution）和逐点卷积（pointwise convolution）。
还是以keras实现的mobilenet为例：
def _depthwise_conv_block(inputs, pointwise_conv_filters, alpha,
depth_multiplier=1, strides=(1, 1), block_id=1):
channel_axis = 1 if backend.image_data_format() == 'channels_first' else -1
pointwise_conv_filters = int(pointwise_conv_filters * alpha)
x = layers.ZeroPadding2D((1, 1), name='conv_pad_%d' % block_id)(inputs)
x = layers.DepthwiseConv2D((3, 3),
padding='valid',
depth_multiplier=depth_multiplier,
strides=strides,
use_bias=False,
name='conv_dw_%d' % block_id)(x)
x = layers.BatchNormalization(
axis=channel_axis, name='conv_dw_%d_bn' % block_id)(x)
x = layers.Activation(relu6, 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.Activation(relu6, name='conv_pw_%d_relu' % block_id)(x)
过程比较明确，depthwiseconv与1*1卷积（逐点卷积）。
（2）V2
a）引入了残差操作。
b）在depthwiseconv之前先使用1*1卷积增加维度。使用inverted residual block模块。
c）pointwise结束之后不在使用relu激活函数，而是使用linear激活函数，来防止relu对特征的破坏。
3）inception
Inception V1——构建了1x1、3x3、5x5的 conv 和3x3的 pooling 的分支网络，同时使用 MLPConv 和全局平均池化，扩宽卷积层网络宽度，增加了网络对尺度的适应性；
Inception V2——提出了 Batch Normalization，代替 Dropout 和 LRN，其正则化的效果让大型卷积网络的训练速度加快很多倍，同时收敛后的分类准确率也可以得到大幅提高，同时学习 VGG 使用两个3´3的卷积核代替5´5的卷积核，在降低参数量同时提高网络学习能力；
Inception V3——引入了 Factorization，将一个较大的二维卷积拆成两个较小的一维卷积，比如将3´3卷积拆成1´3卷积和3´1卷积，一方面节约了大量参数，加速运算并减轻了过拟合，同时增加了一层非线性扩展模型表达能力，除了在 Inception Module 中使用分支，还在分支中使用了分支（Network In Network In Network）；
Inception V4——研究了 Inception Module 结合 Residual Connection，结合 ResNet 可以极大地加速训练，同时极大提升性能，在构建 Inception-ResNet 网络同时，还设计了一个更深更优化的 Inception v4 模型，能达到相媲美的性能。