[JS API] Fix inference for outputs with missing names

[Retribution98]

Details:

• Output.getAnyName() returns the unique name of the node if the tensor has no name.
• The unique name is also used in InferRequest.infer(...) to get the result if the tensor has no name.

Tickets:

• CVS-155526

[vishniakov-nikolai]

Not sure that we want replace default behavior.
Better expose get_names method to allow user perform this check

[Retribution98]

I agree that it is not the best solution, I prefer to use core implementation for Output.get_any_name(). I hope it will be done in next releases.

But now we have this code in samples (hello-world.nnb):

const result = inferRequest.infer([tensor]);
const outputData = result[outputLayer].data;

If we don't add this change the current API will be broken for model with unnamed outputs.
I believe we need to choose one of the options:

• Change the API for infer result, to avoid cases with missing or duplicate names
• Use this fix until it is fixed in the core
• Change our tutorial to avoid this use case and mark this functionality as deprecated.

[vishniakov-nikolai]

API won't be broken, this fix allows to get results by name anyhow. And it's better than exception.
Yes, agree that:

• we must add detailed explanation into documentation about what user can expect to decode infer results
• create task in backlog to change this behavior when fix in the core appears
• leave comments in the samples with documentation reference

[praasz]

Consider to make this behaviour only for model outputs. Generate this way name for tensor can make same name for different outputs.

[vishniakov-nikolai]

The same here

[vishniakov-nikolai]

There is can be name collision. I'm not sure that node name should be unique, across the possible output nodes names.
Hi @praasz, could you please tell, can more than one nodes of outputs have the same name?

[praasz]

The ov::Node has name and friendly_name. The name is generated automatically as unique names, friendly name can be set by user (is name if not set). There is transformation which check/resolve nodes name issues ResolveNameCollisions.

The tensor names are optional and not validated.
The proposed used node names as tensor names is try to use name which not exist in model but can be fine for JS API (it would be good to indicated for user that this name is not real ov model name)

In case of Model outputs (Result nodes) It should be fine as Result has only one output.
In case used for node like Split which can have many outputs using this solution will give each output the same same "virtual" name.
There can be a case that user set some names inside the model which can be same as this generated virtual name.

Solution to consider:

• not use tensor name as key for mapping (use port id instead)
• map only named tensors and skip not named (warning to user that there are not mapped tensors by names)
• map by names and create virtual names but indicate for user that this names are not real and not existing in the model

[praasz]

Suggested change

	std::string name;
	if (_output.get_names().empty()) {
	name = _output.get_node()->get_name();
	} else {
	name = _output.get_any_name();
	}
	const std::string& js_model_output_name(output){
	return output.get_names().empty() ? output.get_node()->get_name() : output.get_any_name()
	}

Suggest to make js internal helper to get this name. If there will be change it will be easier to replace or update it.
Also tak into consideration that will work only for OV output which belongs node with single output. for Nodes which have got many outputs the same tensor name will be generated.

[praasz]

Here is better to use model output ID (number) to show all outputs.

In case model has 3 outputs but output 2,3 have same tensor (same names) then this map shows only two outputs but model has got three.