Default axis being 1 of batchNormalization operation is invalid for 1D input tensor

[BruceDai]

The input of batchNormalization operation would be ND tensor, when it's 1D tensor, the axis as a long scalar would be 0 or -1, it couldn't be 1, right?

And current three operations use axis which are of the same long type, but explanations for axis value are not enough clear, and could axis of concat be negative value?

Op	Axis	Type	Value description
batchNormalization	axis is an option of the optional options parameter	long	When it’s not specified, the default value is 1.When input is a 4-D tensor of the "nchw" or "nhwc" layout, options.axis should be set to 1 or 3 respectively.
concat	axis is the second parameter	long	with the value in the interval [0, N) where N is the rank of all the inputs
split	axis is an option of the optional options parameter	long	Default to 0. A negative value is interpreted as counting back from the end

@huningxin @wchao1115 PTAL, thanks.

[fdwr]

Yeah, a default axis = 1 value would be invalid for a 1D input tensor.

[wchao1115]

I don't understand how a batchNormalization input can be 1D. I think the wording can probably be improved to be more explicit but ND generally implies not 1D in most other specs.

[fdwr]

1D tensors are valid in ONNX BatchNormalization-15. "The op also accepts single dimension input of size N in which case C is assumed to be 1". In 1D, elements are normalized across batches (no spatial dimensions), applying Y = (X - input_mean) / sqrt(input_var + epsilon) * scale + B for each 1D element. I think we should just treat it the same way here, and require an explicit axis (0 or -1) since the default value of 1 is in inappropriate.

[inexorabletash]

Enumerating options:

1. Leave [EnforceRange] unsigned long axis = 1; in the IDL; if the input's rank is 1 then implicitly {axis: 0} must be specified by the caller, otherwise it's an error. This is the behavior specified today, so no spec change needed.
2. Change the IDL to [EnforceRange] unsigned long axis; removing the default value in the IDL, and add a default via prose e.g. "If options.axis does not exist, then set options.axis to 1 if input's rank is 2 or greater, or 0 otherwise"
3. Change the IDL to make axis a positional argument instead of an option; caller must specify.

Any preferences?

[zolkis]

Option 2 sounds like the best way to me.