Why are the output parameters of tvm.relay.optimize empty?

Questions
#1

I am a beginner in using TVM. I have traced the resnet50. Now, I want to get the optimized modules using: model = ResNet50(10, 3) model.to(dtype)

    trace_input = torch.rand(1, 3, 224, 224)
    traced_model = torch.jit.trace(model, trace_input)
    for p in traced_model.parameters():
        p.requires_grad_(False)
    
    shape_list = [(i.debugName().split('.')[0], i.type().sizes()) 
        for i in  list(traced_model.graph.inputs())[1:]]
    mod, mod_params = tvm.relay.frontend.from_pytorch(traced_model, shape_list, default_dtype=dtype_str)
    mod_optimized, mod_params_optimized = tvm.relay.optimize(mod, target, params=mod_params)

However, the length of the mode_params_optimized dictionary is zero. I would appreciate it if anyone could suggest to me how to find the problem.

#2

This is expected, during optimize parameters are bound to the module as constants. If you print mod_optimized you will see a lot of constants embedded directly in the module.

1 Like