NCHW conversion not supported for 3D models

[alkamid]

Describe the bug
I am trying to convert a TF-trained 3DUNet. Below is the head and the tail of the model's summary:

 Layer (type)                 Output Shape                  Param #   Connected to
====================================================================================================
 input_1 (InputLayer)         [(None, 100, 100, 100, 1)]    0         []

...

 activation (Activation)      (None, 100, 100, 100, 1)      0         ['conv3d_18[0][0]']

Conversion command:

python -m tf2onnx.convert --saved-model /data/model.saved_model --output /data/model.onnx --tag serve --signature_def serving_default --opset 18 --inputs-as-nchw input_1 --outputs-as-nchw activation

model.onnx is written, but I'm getting the following two warnings in the output:

WARNING - transpose_input for input_1: shape must be rank 4, ignored
WARNING - transpose_output for activation: shape must be rank 4, ignored

NCHW arguments indeed don't seem to have any effect. When I run the model in PyTorch on a (1, 1, 100, 100, 100) input, the output shape is (1, 100, 100, 100, 1). It's unclear to me why the input is accepted in NCHW format despite the warnings above.

What I expect is the converted model to take (batch_size, 1, 100, 100, 100) inputs and output (batch_size, 1, 100, 100, 100) tensors.

System information

• OS Platform and Distribution (e.g., Linux Ubuntu 18.04*): Ubuntu 20.04
• TensorFlow Version: 2.13.1
• Python version: 3.11.9
• ONNX version (if applicable, e.g. 1.11*): 1.16.2
• ONNXRuntime version (if applicable, e.g. 1.11*):

To reproduce

I am new to tf2onnx so I would first like to establish if this is the intended behaviour. If not, I'll try to create a MWE.