TensorFlowScorer in ML.NET does not support input tensor of unknown shape

[CESARDELATORRE]

Using Microsoft.ML 0.5.0
Yael told me she's already working on a PR to fix this issue. 👍

Issue/Bug:
We’d like to extend your test case TensorFlowTransformCifarLearningPipelineTest for making a demo, but instead of loading your “cifar” frozen model, we’ll be using an inception model trained for imagenet (downloaded from https://storage.googleapis.com/download.tensorflow.org/models/inception5h.zip ).

The main issue is that when executing the Train method of the pipeline, we get an error related to the initialization of the TensorFlowScorer, which I believe is related with using an input tensor of unknown shape.

Details of the bug/issue are below, but the main question is if we’ll be able to use an input tensor of unknown shape.

var shape = tfShapes[i].ToIntArray().Skip(tfShapes[i][0] == -1 ? BatchSize : 0)

Former code is executed in the method TensorFlowMapper.GetInputMetaData(). tfShapes[i] is an object of type TFShape; if the shape is unknown, accessing by the indexer to the first dimension ( tfSfapes[i][0] ) throws a null exception. I checked that most of the methods in TFShape have a guard condition on dims == null, but the overridden indexer doesn’t have this guard, what could be the culprit of the NullReferenceException

But even if you fix the indexer, then the method tfShapes[i].ToIntArray() it will return null for an unknown tensor, so this will throw again an exception.

Full source code of the demo (and the crash) is available here:
https://github.com/CESARDELATORRE/TensorFlowMLNETSamples/blob/master/src/TensorFlowMLNETInceptionv3ModelScoring/Program.cs