Clarify whether `Unsqueeze` should accept scalar `axes`

[robertknight]

The specifications for Squeeze and Unsqueeze say that the axes input should be a "list of integers", which I took to mean a 1D tensor. ONNX Runtime expects the axes input to be a 1D tensor for Squeeze and Reduce* ops, but for Unsqueeze specifically a change was made in microsoft/onnxruntime#5665 to support scalars as well. The current test cases for Unsqueeze in the onnx repo always pass a vector for the axes input, although the reference implementation will accept either since numpy functions often allow axis arguments to be either a scalar or tuple.

I'm not sure what the preferred change is here, but if Unsqueeze should accept scalar axes, then the spec should make this clear.

[justinchuby]

I don't think it should. In ONNX axes is always 1d and axis is always scalar.

[justinchuby]

#6888

[robertknight]

Thanks! I forgot to mention, the context where this issue came up was exporting a Swin Transformer model from timm (v1.0.15), using the new TorchDynamo exporter (PyTorch v2.6.0):

import timm
import torch

# https://huggingface.co/timm/swin_tiny_patch4_window7_224.ms_in1k
model = timm.create_model("swin_tiny_patch4_window7_224.ms_in1k", pretrained=True)
model = model.eval()
data_config = timm.data.resolve_model_data_config(model)
transforms = timm.data.create_transform(**data_config, is_training=False)
img = torch.zeros([1, 3, 224, 224])

onnx_prog = torch.onnx.export(model, img, dynamo=True)
onnx_prog.optimize()
onnx_prog.save("swin.onnx")

The resulting model worked in ORT but not some other runtimes that I tested, due to an Unsqueeze node (node_Unsqueeze_43) with scalar axes input.

[justinchuby]

I see. That we should fix. Thanks!

[justinchuby]

Fixed in microsoft/onnxscript#2186