Thank you!

Can you please for each fixed case provide a simple one-layer test by the following script: https://github.com/opencv/opencv_extra/blob/master/testdata/dnn/onnx/generate_onnx_models.py?

Hi, @dkurt
I was trying to import the ONNX model after making a fix to split layer. However, other layers were not processed correctly like, when layer_type is Constant, attribute_proto.t() is returning an empty tensor. I suspect there is a problem with .onnx file shared in #16370 (comment) as the visualization of the model doesn't have any Constant layer but the graph_proto is showing to have it and that too with string data type which is not supported by getMatFromTensors.
I am still trying to figure out what's not working correctly. I will add the tests once the model is processed correctly. Thank you.

Update:
Hi,
I will be having my mid semester examinations till 7th February(IST). So the work on this may get slowed down. Please do not close this PR until then. I am on it and I will put efforts to complete it from 7th Feb. Thank you and apologies.

Hi,
I am unable to figure out how to run the tests for this PR locally.
I tried running, python3 generate_onnx_models.py under /opencv_project/opencv_extra/testdata/dnn/onnx but got,

maxpooling input has sizes torch.Size([1, 3, 10, 9])
maxpooling output has sizes torch.Size([1, 3, 3, 3])

Segmentation fault (core dumped)

Do we have to build the code after generating models or they are generated automatically during the build?
I am unable to see logs at https://pullrequest.opencv.org/buildbot/builders/precommit_linux64/builds/24598/steps/test_dnn

unable to see logs

These logs were expired.
I rescheduled new builds, please check.

Thanks @alalek
After fixing a probable bug in Constant layer, I observed that ArgMax hasn't been handled. I got the following exception,

 what():  OpenCV(3.4.9-dev) /home/czgdp1807/opencv_project/opencv/modules/dnn/src/dnn.cpp:561: error: (-2:Unspecified error) Can't create layer "LabelId" of type "ArgMax" in function 'getLayerInstance'

Should I add a new file under layers for this and some more uncovered layers(ArgMin)?

@czgdp1807, Can you please dump a subgraph from the model which includes ArgMin node? Just to understand which layer it tries to perform.

I have attached a screenshot below of the last few layers. The ArgMax isn't created as the case for it isn't covered as far as I am able to figure out.

Got it, thanks! This is training only nodes so author should exclude them from deployment model (you may try to load and save model but before export call .eval() method or specify SoftMax as output.

So let's focus on Split layer but ask author to try to remove training nodes.

Thanks. IMO, this could be one of the ways as the last few layers(the layers which are extracting the label on the left side.) aren't making any difference to the result, though the interpretation would be different.
I will drop the last commit from here and will add test files for Split layer in my other PR.

Please restart the tests here as I have made an update to opencv_extra.
Thanks.

Please restart the tests here as I have made an update to opencv_extra.
Thanks.

+1

Amend commit (no changes are required, just update commit date) and push "new" changes to this PR:

git commit --amend
git push <remote> <branch> --force

Done. Locally the tests were passing. Let's see how it goes here.

I have applied the comments and reviews. Let me know if anything else is to be done. Tests should pass here.

The tests passed. Is it good to go?

@czgdp1807, some of test from Inference Engine backend are failed:

[ RUN      ] Test_ONNX_layers.Split/1, where GetParam() = NGRAPH/CPU
/build/precommit_custom_mac/3.4/opencv/modules/dnn/test/test_common.impl.hpp:66: Failure
Expected: (normL1) <= (l1), actual: 0.5 vs 1e-05
/build/precommit_custom_mac/3.4/opencv/modules/dnn/test/test_common.impl.hpp:69: Failure
Expected: (normInf) <= (lInf), actual: 1 vs 0.0001
FALLBACK: Layer [Slice]:[1] is expected to has backend implementation
/build/precommit_custom_mac/3.4/opencv/modules/dnn/test/test_common.impl.hpp:66: Failure
Expected: (normL1) <= (l1), actual: 0.5 vs 1e-05
/build/precommit_custom_mac/3.4/opencv/modules/dnn/test/test_common.impl.hpp:69: Failure
Expected: (normInf) <= (lInf), actual: 1 vs 0.0001
FALLBACK: Layer [Slice]:[1] is expected to has backend implementation
/build/precommit_custom_mac/3.4/opencv/modules/dnn/test/test_common.impl.hpp:66: Failure
Expected: (normL1) <= (l1), actual: 0.5 vs 1e-05
/build/precommit_custom_mac/3.4/opencv/modules/dnn/test/test_common.impl.hpp:69: Failure
Expected: (normInf) <= (lInf), actual: 1 vs 0.0001
FALLBACK: Layer [Slice]:[1] is expected to has backend implementation
/build/precommit_custom_mac/3.4/opencv/modules/dnn/test/test_common.impl.hpp:66: Failure
Expected: (normL1) <= (l1), actual: 0.5 vs 1e-05
/build/precommit_custom_mac/3.4/opencv/modules/dnn/test/test_common.impl.hpp:69: Failure
Expected: (normInf) <= (lInf), actual: 1 vs 0.0001
FALLBACK: Layer [Slice]:[1] is expected to has backend implementation

I can take a look and tell how to fix it

On the other hand, other Split tests are skipped.

So please add these lines at the beginning of test:

    if (backend == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019)
        applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_NN_BUILDER);
    if (backend == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH)
        applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_NGRAPH);

How to run the failing tests locally, as no test was failing on, ./opencv_test_dnn.

@czgdp1807, it's optional backend for deep learning module: https://github.com/opencv/opencv/wiki/Intel%27s-Deep-Learning-Inference-Engine-backend (OpenVINO). Anyway there are other tests for split which are failed as well. So let's skip it for now but fix them all separately.

Sure. Can we open a separate issue for the remaining failures in Split layer? I will work on it in a different PR.

Yep, let's open a separate issue to not to miss. Thanks!

I have updated the code and opened a separate issue. Please let me know, if any more update is to be done in this PR.