Different output for large YOLO ONNX model in Python (correct) and C++ (incorrect)

8

hi @ryan-csm,

Could you try running with -link-params as was suggested on the other post? To do so:

  1. add -link-params to the Target string.
  2. if you’re just doing the following:
    compiled = tvm.relay.build(...)
compiled.export_library("foo.so")
    the GraphExecutorFactory will shoot you in the foot at runtime as you observed. Instead, you need to export like so: compiled.get_lib().export_library("foo.so"). you then also need to export the Graph JSON separately: open("graph.json", "w").write(compiled.get_json()).
  3. then, at runtime, load the library and JSON and instantiate. i’ll be interested to see what you get.