Add basic tests for ONNX module

[zaleslaw]

Add tests for the following classes:

• ONNXModelHub
• OnnxInferenceModel
• ONNXModels (for each preprocessing)
• SSDObjectDetectionModel
• ONNXModelPreprocessor
• Transpose

All tests should be located in the test directory of the ONNX module

[femialaka]

I would like to work on this one

[zaleslaw]

Great @femialaka fill free to create draft PR and ask some questions or create the list of tests before implementation to discuss here, in the issue's comments.

[femialaka]

For the first 3 tests I have the following suggestions

• ONNXModelHubTest

    	assert that ONNXModelHub creates a directory for a pre-trained onnx model
    	assert that ONNXModelHub can load the model
  




•  OnnxInferenceModelTest

    	assert that the correct onnx model is loaded - is this enough?
  


 

• ONNXModels (for each preprocessing)

    For this test I would like to ask for the best way to test.  For example would it be ok to test the accuracy? 
            What would be the best approach, using the "predict" value?
  

I'll appreciate your guidance on this, thanks.

[zaleslaw]

Hi, @femialaka will try to comment on your ideas.

assert that ONNXModelHub creates a directory for a pre-trained onnx model - better to check it in the temp directory (use @tempdir Junit feature)

assert that ONNXModelHub can load the model - do you want to check the model availability on S3?

assert that the correct onnx model is loaded - is this enough?
In the first, we need to check that the directory is not empty and the appropriate file is created (maybe check fileName, fileSize in bytes, unfortunately, we have no control sum for the files here)
Also, I suppose that the loading of the model is a part of modelHub test cases.

The OnnxInferenceModel has a few properties and methods that should be tested. Use a tiny model (Lenet, for example) for test purposes. If required, it could be added to the testResources in the onnx module.

The accuracy checking could be a good test in the case of ONNX Runtime's ability to predict the same data on the same outputs; I hope so.

In this case ONNXModels (for each preprocessing) I suggest defining tests inputs and outputs for each preprocessing hidden in the ONNXModels.