Help Needed with MaxPool3D and QuantConv3D in FINN for 5D Tensor Input

[PedroAntunes178]

Hi everyone,

Has anyone worked with MaxPool3D and QuantConv3D in FINN? I'm currently implementing a neural network with a 5-dimensional input tensor (5D tensor) and encountered an issue when running the following command: build.build_dataflow_cfg(model_filename, cfg_estimates)

I receive the following error:

Running step: step_streamline [3/10]
> /home/pantunes/sandbox/finn/deps/qonnx/src/qonnx/transformation/infer_data_layouts.py(139)apply()
    137                 model.set_tensor_layout(inp_name, DataLayout.NC)
    138             else:
--> 139                 raise Exception(
    140                     """Unknown number of dims for input, don't know
    141                 how to annotate"""

It seems this error is caused by an undefined DataLayout. From what I understand, FINN does not support tensors with dimensions ["N", "C", "H", "W", "L"].

Can anyone confirm if FINN indeed lacks support for this data layout? Additionally, is there a workaround or recommended approach for using a 5D tensor as input to a neural network in FINN?

Any guidance or advice would be greatly appreciated.

Thank you!
Pedro