[Bug] InstanceNorm2d with relax not supported

[sidsingla]

Hi,
TVM version: 0.20.dev0

Here is the Traceback:

Traceback (most recent call last):

File "/home/sidharth_modiface_com/looks_factory_tvm/looks-factory-transfer/convert_to_tvm.py", line 93, in <module>
  mod = from_fx(
        ^^^^^^^^
File "/opt/conda/envs/py311-looks-factory-tvm/lib/python3.11/site-packages/tvm/relax/frontend/torch/fx_translator.py", line 1064, in from_fx
  return TorchFXImporter().from_fx(
         ^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/opt/conda/envs/py311-looks-factory-tvm/lib/python3.11/site-packages/tvm/relax/frontend/torch/fx_translator.py", line 935, in from_fx
  assert (
AssertionError: Unsupported module type <class 'torch.nn.modules.instancenorm.InstanceNorm2d'>

Code snippet:

from tvm.relax.frontend.torch import from_fx

mod = from_fx(
        fx_graph,
        [((1, 3, 256, 256), "float32"), ((1, 1, 256, 256), "float32")],
        keep_params_as_input=True,
    )

Model is mostly based on pix2pix GAN model.
Thanks!

[kavin-sai-krishna]

@mshr-h @tlopex @Hzfengsy
I’ve added the Relax-level instance_norm operator, following standard semantic checks and layout inference. Since topi.instance_norm already exists, I tried using it for legalization, but it produced incorrect results.

On debugging, I found that it reuses layer_norm, which expects gamma and beta to match an n-dimensional normalized_shape, whereas instance_norm requires them to be 1D with shape [C].

In this PR #17885 , I’ve implemented the logic assuming gamma and beta have shape {C}.

Could you please confirm if this approach is correct or if I’ve missed anything? Thank you.