Support for transposed conv3d

[devbanu]

Hi,

I was trying to convert a PyTorch model for 3D segmentation based on the UNETR architecture.

UNETR paper: https://arxiv.org/abs/2103.10504

UNETR documentation in MONAI framework: https://monai.readthedocs.io/en/stable/networks.html#monai.networks.nets.UNETR

An example model with weights can be found here: https://github.com/lab-smile/GRACE

I am converting following the steps in the documentation:

import coreai
import coreai_torch
from coreai.runtime import NDArray

converter = coreai_torch.TorchConverter()

exported_program = exported_program.run_decompositions(coreai_torch.get_decomp_table())

converter.add_exported_program( exported_program, input_names=["features"], output_names=["logits"])

core_ai_program = converter.to_coreai()

I am hitting the case here:
https://github.com/apple/coreai-torch/blob/main/coreai_torch/_aten_to_core.py#L1179

It seems transposed 3D convolutions are not yet supported: Transposed conv3d is not yet supported.

Is there a workaround I can use here, even if performance is not the best?

coreai-core==1.0.0b1
coreai-torch==0.4.0