train once for all OFA with paddleslim

we want to train a neural network that is compressed by OFA with paddle with part of imageNet , subtract a sub-net, and train it with a new dataset finally, design a header for implementation