Enable to load models with low opset version

[sungchul2]

Changes

• Enable to load models with low opset version
• Add to change tensor shape in Accuracy

Reason for changes

1. The models with opset==7, 8, 9 have not only input tensor, but also every weight and bias in original_model.graph.input. But the model with opset==12, which have only input tensor in original_model.graph.input, can be only loaded. <opset==7, 8, 9> <opset==12>

2. densenets output a 4d tensor , but Accuracy metric cannot use it because it can only handle a 2d tensor.

Related tickets

88636

Tests

[kshpv]

@sungchul2 thanks! Could you also add a test that will failed without your fix and pass with it?

[vinnamkim]

@sungchul2 Please rebase this PR on the latest develop branch and push again. This is because of https://github.com/openvinotoolkit/nncf/pull/1248.

[sungchul2]

run tensorflow pre-commit tests

[sungchul2]

run pytorch pre-commit tests

[MaximProshin]

@sungchul2 , do you plan to finalize this PR?

[sungchul2]

run pytorch pre-commit tests

[sungchul2]

@kshpv I saw models from torchvision just have single input, not multiple inputs, and it is not the purpose of this ticket. So, I changed them to models from onnx model zoo, which directly have had this issue, such as densenet.

And I added Mask R-CNN and yolov3 for test as you requested. These models have dynamic input_shape and have been converted with high opset version (10, 12). So, I separately added a test for them and renamed the test file.

[kshpv]

@sungchul2 thank you for the PR. It looks very good!

[sungchul2]

retest this please