MobileNet ShuffleNet & yolov5 replace backbone

The following content refers to reference

MobileNet V1

Highlights:
(1) The use of separable convolution kernels greatly reduces the amount of parameters
(2) Increases hyperparameters a, b
a: magnification of the number of convolution kernels
b: resolution level

MobileNet V2

Highlights:
(1) Use inverted structure residuals
Residual structure: thick on both sides, thin in the middle
Inverted residual structure: thin on both sides, thick in the middle

MobileNet V3

Highlights:
(1) Update block
①Add SE module
②Replace activation function

(2) Redesign the time-consuming layer structure
①Reduce the number of convolution kernels in the first convolutional layer
②Simplify the last stage: delete the part that does not significantly improve performance

(3) Redesign the activation function
swishx is complicated, replace it with h-switch with similar effect

ShuffleNetV1

Highlights:
①Using group convolution to greatly reduce the amount of parameters
②Using Shuffle layer after group convolution is beneficial to information sharing between different groups

ShuffleNetV2

Notes from here!
ShuffleNetV2 directly tests how fast the network runs on the device by controlling different environments, rather than judging by FLOPs.

(1) Four Principles

1. When the input feature matrix of the convolutional layer is equal to the output feature matrix channel, the MAC is the smallest (keeping FLOPs unchanged)

• MAC:Memory access cost

2. When the GConv groups increase (while keeping the FLOPs unchanged), the MAC also increases

• GConv: Group ConvolutionThe introduction of group convolution is here

GroupConv divides the convolutional Filters into G groups, and the input feature map channels are also divided into G groups, each group of convolutional Filters processes a corresponding set of input feature mapsaisle.Since each group of convolution filters is only applied to the corresponding input channel group, the computational cost of convolution is significantly reduced.

There are also disadvantages:

Channel information is not shared between different groups, that is, the output Feature map channels of different groups only receive information from the input channels of the corresponding group.This hinders the information flow between different group channels and reduces the feature extraction ability of GroupConv.

---

On the one hand, it allows more channels to be used under a fixed FLOPs, and increases network capacity (and thus accuracy).However, on the other hand, increased number of channels results in more MACs

3. The more fragmented the network design, the slower the speed

4. The impact of Element-wise operation cannot be ignored

Practice

Someone else wrote it so well!!!

Actually, the replacement of a module is nothing more than three steps:
①Define the module
②Register the module in yolo.py
In other words, registration may not necessarily use

③Replace the corresponding layer structure in the yaml file

---

Shufflenetv2

---

Mobilenetv3