Paddlepaddle 29 dynamically modify the network structure without model definition code (relu changes to prelu, conv2d changes to conv3d, 2D semantic segmentation model changes to 3D semantic segmentat

paddle The framework provides many models , But in the endless engineering needs , This is far from enough . Sometimes we need to modify the network structure according to our own situation , For example, single input becomes double input , Single output becomes multiple output . This is just the adjustment of input and output , In addition, there is the adjustment of network details , For example, change the activation function of the network from relu Change it to prelu、elu、switch etc. ; Such as the 2d The semantic segmentation model of is changed to 3d Segmentation model . If there is the source code of the model structure definition , Modifying the network structure is naturally very simple . however , Many times, the model definition code is too complex , And part of the library code , It cannot be easily modified , To this end, bloggers have achieved without changing the model definition code , Modify the model structure . These operations belong to the scope of network structure magic change , About this operation , For more detailed reference tutorials, please refer to the following links .pytorch 26 Scientific alchemy （ Training ） In this way —— Magic changes the Internet （ Surgery on the model , Realization module The deletion of 、 Modification and addition ,forward Process modification ）_ A flash of hope to my blog -CSDN Blog _ Magic changes the Internet Network structures emerge one after another , But the classic network is always everyone's favorite in Engineering . Sometimes you need to micro operate the network , First, the workload is too large ;