I’m trying to load and process an ONNX model. My processing supports NHWC inputs and operations, however, the model input, and operations are all in NCHW.
I know I can use ConvertLayout pass to change the data layout and kernel layout of certain operations, but that doesn’t change the input layout.
This is how I do that: model_obj = onnx.load(model_path) relay_module, params = tvm.relay.frontend.from_onnx(model_obj)
and this is the input shape of the main function: def @main(%Input3: Tensor[(1, 1, 28, 28), float32])
and I would like to have it like this: def @main(%Input3: Tensor[(1, 28, 28, 1), float32])
Can anyone help me with that?