Image postprocessor

[jonpsy]

Continuation of #679. @lu-wang-g

[lu-wang-g]

Simplify this by something like:

ASSIGN_OR_RETURN(auto normalization_options, GetNormalizationOptionsIfAny(*GetTensorMetadata()));

if (!normalization_options.has_value() && input_index>-1) {
  ASSIGN_OR_RETURN(normalization_options, GetNormalizationOptionsIfAny(*input_metadata));
}

[jonpsy]

Nice one! 👍

[lu-wang-g]

This will break if we use ImagePreprocessor and ImagePostprocessor on a model with multiple inputs. We should remove these check.

[lu-wang-g]

metadata_extractor.GetOutputTensorCount() returns the output tensor count of the tflite model, which can be greater than one. For example, a model may takes two input image tensors. You can bind an ImagePreprocessor to each of the input image tensor, respectively, but not both input tensors to one ImagePreprocessor (that's what we've agreed in ImagePre(/Post)processor). However, the code will fail here, because metadata_extractor.GetOutputTensorCount() returns 2.

This is a bug in our existing code, it's not discovered before, because all models we supported so far have only one input tensor of image. But we need to correct the logic.

[lu-wang-g]

And if you pass in TensorMetadata, it will be garnered that the number of tensor is 1 for the ImagePostprocessor.

[lu-wang-g]

Should get TensorMetadata in ImagePreprocessor and ImagePostprocessor respectively, and pass in TensorMetadata instead of metadata_extractor into BuildImageTensorSpecs. This helps to avoid the if else statements in GetTensorMetadataIfAny to check input or output.

[jonpsy]

@lu-wang-g Okay but we need to use GetTensorMetadataIfAny for these checks:

  if (metadata_extractor.GetModelMetadata() == nullptr ||
      metadata_extractor.GetModelMetadata()->subgraph_metadata() == nullptr) 

and it requires metadata_extractor. So probably we should keep it?

[lu-wang-g]

And if you pass in TensorMetadata, it will be garnered that the number of tensor is 1 for the ImagePostprocessor.

[jonpsy]

Made substantial changes. To summarize:

a) I'm storing output_tensor_ in Postprocessor class. So that Postprocess() can access the corresponding output tensor and also find tensor_type.
b) If output metadata doesn't have norm params, fallback to input metadata in Init(). This check is done via FindFirstNormProcessingUnit.

TODO:
Currently, we pass first input tensor into BuildImageTensorSpecs, instead we should get the input_index similar to what was done here. That is out of the scope of this PR, we can create a follow up.

Also need to work on some docs, any suggestions?