I tried to get information about which functions have been fused in my compiled IRModule.
The network is a MXNet version of ResNet-18 and has been compiled with relay.build_module.build(mod, target, params=params).
When I list the functions of the IRModule, I get only a single function with 103 input arguments. The output of relay.analysis.extract_fused_functions(mod) is a dictionary with 26 relay functions. However, I do not understand what is being returned here.
The functions have a body, an op and a name. The names are usually “nn.relu”, “nn.global_avg”, and similar. Convolutional layers are not in there.
Is there a way to correspond these layers/groups with the PackedFunctions that are executed by the graph_executor? These usually have namens like “fused_nn_contrib_conv2d_NCHWc_add_add_nn_relu_4”