Throw for scalar input to Shape.

[skottmckay]

Description

Shape returns an empty tensor with shape {0} for a scalar, which is incorrect as there is a single data element. Throwing will at least make the cause of the failure clear.

Expected might be for it to return a shape of {}, but ONNX shape inferencing doesn't support scalar either as it calls add_dim on the output_shape at the start so result will always be rank 1 or higher.

https://github.com/onnx/onnx/blob/b60f69412abb5393ab819b936b473f83867f6c87/onnx/defs/tensor/defs.cc#L496

The ONNX spec doesn't say scalars aren't supported. @gramalingam should scalar input to Shape return an empty shape?

Motivation and Context

Model failed downstream due to nullptr in output from Shape when called with a scalar.

[gramalingam]

The shape of a scalar is {}. The question is its rank. It makes sense that Shape returns a 1-dimensional tensor with k elements, where k could be zero as a special case. But, {0} is incorrect (for a scalar input). (Though it is valid return value for a 1-dimensional tensor with zero elements.) So, the onnx shape inference seems correct to me.

I don't understand your concern about "so result will always be rank 1 or higher". It will be rank 1, but with zero elements, and that is fine?

[skottmckay]

Sorry - I misinterpreted what the shape inferencing was doing due to misconfiguring my unit test and confused myself between empty shapes and shapes for input with no data.

I think after all that it's expected to return nullptr in this case as there is an empty shape for a scalar - so there's no shape info to return.

Slightly confusingly (to me anyway) for an input with no actual data, using input with a shape of {0} as a simple example, it will return a 1D tensor a shape of {1} that has a value of 0 in it.

So we get nullptr output for a scalar input, and 1D output for an input which had a shape but no data due to a dim having a value of 0.

[gramalingam]

I still don't understand the root-cause for the model that failed downstream. Is that resolved or not? I don't understand the nullptr issue. I think that onnxruntime needs to be able to handle the special case of tensors that have size 0, but a very specific shape: for example a tensor of shape 128 x 0 x 256, say. A tensor-representation combines shape information ({120, 0, 256} in this example, with data information (no data in this example). It seems fine if the dataptr is null in this special case, but the tensor-pointer will be non-null.

Is my understanding correct? So, is the problem that something downstream assumes dataptr is non-null, when it is null?