Profiling with TVM 0.9

Troubleshooting
#1

Hello,

I am trying to profile my NNs in C++ using TVM 0.9. However, to profile, we make use of the old way of building the models (with separate graph_json, lib, params) and in particular the graph in JSON. However, I am using the new way (i.e. with the .so file that contains the graph executor). Is there a way/tutorial that indicates how to get the JSON from the .so file to then profile the NN ?

1 Like
#2

@Upscale4152 you can call get_graph_json() on the returned Module from load_library() to retrieve the JSON.

2 Likes
#3

Thank you very much ! Everything works now. I am wondering, it is mentionned somewhere on the docs? I’m pretty sure I searched everywhere and didn’t find mentions about it.

#4

@Upscale4152 I don’t actually see GraphExecutorFactory in our docs. I’ll try to fix that soon.