ModelSplit

Try to Split one model to many smaller models

Time Test

shape=(1,3,224,224)

• dell04
  • cpu:
    • 0.1935865879058838
    • 0.17652320861816406
  • gpu
    • 0.4194180965423584
    • 0.005059957504272461
• P1000
  • 0.030528545379638672
  • 0.0256803035736084

	dell04-gpu(2080Ti)	dell04-cpu	P1000	开发板A9
冷启动：第一次运行	0.4194180965423584	0.1935865879058838	0.030528545379638672	4.442520
第二次+	0.005059957504272461	0.17652320861816406	0.0256803035736084	2.276656

Support

only support to parse & split IRModule text without meta[Constant] value

• DL model
  • googlenet
  • resnet50
  • squeezenet
  • vgg19
  • yolov2
• convergence node
  • 2 limits
    • a node has only one next node, and the next node only has one prior node, too.
    • there is no data flow across from priors to next
      • priors: the total of in degree(include convergence)==the total of out degree((exclude convergence))
• split model steps
  • init: get a IRModule text file, with a value shape
  • parse IRModule text, build graph with the relationship of operators, find convergence node points
  • choose one or more convergence node points as the split points of the model
  • split the IRModule text as two or more parts, store as some split-IRModule text and a params json file
    • split-IRModule text need to handle input(add new input & remove (include redundant input/params)
    • params json file record the input & ouput of each split IRModule
  • export python file with split-IRModule text
    • the python file consists of a function that return RelayExpr, so you can just import this function
    • use tvm.IRModule.from_expr to translate RelayExpr to IRModule