[BUG] An initializer can be declared without previous reference in input/output/value

[owulveryck]

Trying to execute the yolo v3 model raise this error:

model_zoo_executor git:(master) ✗  MODELDIR=models/yolov3 go test -failfast | more
--- FAIL: TestModel (1.37s)
    --- FAIL: TestModel/Unmarshal (1.37s)
        main_test.go:56: invalid model: initializer has not been defined in input, output or value
FAIL
exit status 1
FAIL    github.com/owulveryck/onnx-go/examples/model_zoo_executor       2.306s

This is triggered by

onnx-go/decoder.go

Lines 137 to 145 in 1681b26

	for _, tensorProto := range model.Graph.GetInitializer() {
	name := tensorProto.GetName()
	if name == "" {
	return errors.New("initializer should have a name")
	}
	n, ok := m.dbByName[name]
	if !ok {
	return errors.New("invalid model: initializer has not been defined in input, output or value")
	}

This error is triggered if the initializer (referenced by its name) has not been defined previously.

However, a recent change in the IR doc of onnx now specifies:

When an initializer has the same name as a graph input, it specifies a default value for that input. When an initializer has a name different from all graph inputs, it specifies a constant value.

Hence, the onnx-go parser must be fixed to create a new node in the graph if the initializer has not been defined before.

Note: by now, onnx-go only handles two types of nodes:

• Operator
• tensor
  We may use tensor to represent the initializer, but it will silently discard the fact that this node is constant. I propose to use the YAGNI principle and postpone the decision to mark the node as a constant.