Clarification of Reshape semantics for attribute 'allowzero' NOT set, zero volume about onnx

[justinchuby]

@onnx/sig-operators

[gramalingam]

Good question. Looking at the onnx shape inference implementation, it computes the complete products (not skipping the forwarded dimensions). Hence, it will not be able to infer an output dimension and will flag this as an error here.

However, the more general interpretation could be useful in some situations, I guess ... it may be worth investigating whether backend implementations support it.

Do you see any examples/models where this will be useful? If so, it may be worth updating the spec to allow it.

[ArchRobison]

It seems to be useful for dealing with batch dimensions that might be zero. For example, reshaping from [n,a,b] to [0,-1] where n is the batch dimension.

We (Nvidia TensorRT group) ran into the issue with fasterrcnn_resnet50_fpn.onnx (I think it's derived from here)> and accidentally fed it random data. I'm guessing there's some kind of internal batch dimension there, with a data-dependent length.

On the other hand, the "forwarding 0" is dangerous with networks that contain empty tensors, so there's much to be said for just discouraging "forwarding 0", even if it helps the use of wildcard -1. In retrospect, "forwarding -2" would have been a much better design, but Caffe chose 0.