[Frontend] Fix shape init for ONNX frontend

[dlexplorer]

The model https://github.com/onnx/models/tree/master/vision/classification/inception_and_googlenet/inception_v1 has 0th input pointing to the weights. The change repeats onnx frontend tests approach to shapes for all inputs

[giuseros]

Hi @dlexplorer ,
Would it be possible to add a test stressing the onnx case?

[dlexplorer]

Hi @dlexplorer ,
Would it be possible to add a test stressing the onnx case?

to stress test of tvmc with onnx frontend? We can add inception v1 into tests, I see some a_model.onnx is used in tvmc frontend tests already, but do not know how to add assets for testing and layout of final test package.

BTW, I was wrong saying in the desc that the same approach was used in the tests. It is used in another app: https://github.com/apache/tvm/blob/main/apps/wasm-standalone/wasm-graph/tools/build_graph_lib.py#L33

[leandron]

Hi @dlexplorer, thanks for the PR. I got a comment, maybe to generate a quick discussion.

Also cc @comaniac.

[comaniac]

LGTM

[mbrookhart]

I agree with the overall design we've reached here, I think it's the right call. I think there's a bug with dynamic onnx inputs

Can we add a unit test that hits this import path?

[mbrookhart]

I think we need to edit this logic just a little bit. If we have an unspecified input shape in the onnx model, it will show up as a variable, I think we need to identify that case and convert it to relay.Any() here.

For instance, this version of BERT has for one of it's input shapes [input_ids_dynamic_axes_1,384] I believe this would fail on those inputs.

[dlexplorer]

Can we add a unit test that hits this import path?

Could you specify usecases to cover in the test?
I am aware about googlenet from onnx models having input not in the 0th place.
Or do you mean test for dynamic input? The initially PR wanted to cover only googlenet case, but yes, it might be good idea to add here case of dynamic shape as well.

[mbrookhart]

I just want to hit the code, even creating a silly 1 op onnx model and parsing without providing shapes would be sufficient.

[leandron]

I just want to hit the code, even creating a silly 1 op onnx model and parsing without providing shapes would be sufficient.

There is this one (below) that compiles a network using TVMC (and hence this code), without providing shape.

Are you talking about something different?

tvm/tests/python/driver/tvmc/test_compiler.py

Line 117 in dc51f17

def test_compile_onnx_module(onnx_resnet50):

[mbrookhart]

That works for me.