From repvgg to mobileone, including mobileone code

The idea of re parameterization is essentially to use the additivity of linear models . In industry, it is conv and bn Layer fusion is applied . In recent years, there have been re-parameter The job of .RepVGG It is a good application .

VGG It is a straight cylinder model , Because no skip connnection, Lead to deeper training vgg There will be degradation of the model . However, such as resnet etc. ,skip connnect It increases the running time for the end-to-end equipment , There is a lot of consumption in data access .
therefore RepVGG, Integrate the idea of heavy parameters , During the training , by vgg Introduced skip connection, During the test skip connection Merge with the main branch , Get one convbnrelu block, Then it becomes a straight cylinder model .
But it's different from resnet,RepVgg In order to ensure the additive principle of linear model , In two relu Between , Use residual Branch . Because any linear layer between nonlinear layers , Can be merged .

About how to integrate , You can see it in the reference link , It's very detailed

and MobileOne It is in RepVGG On the basis of , Realize that the straight cylinder model can bring a good acceleration to the end model . So it was decided that RepVgg Transformed into a lightweight model , Convolute depth separation , Use some well-designed training strategies , Make the straight tube model on the light-weight model , On a par mobilenet And other well-known lightweight models , And it is better in running speed .

I reappeared mobileone Of s0 edition , The effect is almost the same as that of the paper , Interest can be found in my github see
MobileOne code

Reference resources

In depth reading repvgg