There is accuracy test, performance test and test data in opencv_extra repository, if applicable
Patch to opencv_extra has the same branch name.

@dkurt , do both opencv and opencv_extra branches need to have same name ?

unknown file: Failure
C++ exception with description "OpenCV(3.4.9-dev) /build/precommit_macosx/3.4/opencv/modules/ts/src/ts.cpp:1052: error: (-2:Unspecified error) OpenCV tests: Can't find required data file: dnn/onnx/models/upsample_unfused.onnx in function 'findData'
" thrown in the test body.

@ashishkrshrivastava, by default yes. If you already have one - add it's name to PR description

@dkurt

node = onnx.helper.make_node(
    'Resize',
    inputs=['X', 'roi', 'scales'],
    outputs=['Y'],
    mode='cubic',
)

data = np.array([[[
    [1, 2, 3, 4],
    [5, 6, 7, 8],
    [9, 10, 11, 12],
    [13, 14, 15, 16],
]]], dtype=np.float32)

roi = np.array([], dtype=np.float32)
scales = np.array([1.0, 1.0, 0.8, 0.8], dtype=np.float32)

# [[[[ 1.47119141  2.78125     4.08251953]
#    [ 6.71142578  8.02148438  9.32275391]
#    [11.91650391 13.2265625  14.52783203]]]]
output = interpolate_nd(
    data, cubic_coeffs, scale_factors=scales).astype(np.float32)

expect(node, inputs=[data, roi, scales], outputs=[output],

Here, In ONNX resize node example ROI is empty, and hence ONNX model can contain an empty constant node. I am getting an empty node when exporting pytorch model containing upsample layer using ONNX opset version 11. I want your suggestion regarding this empty constant node.
It need to be handled otherwise it would create problem here,

Mat getMatFromTensor(opencv_onnx::TensorProto& tensor_proto)
{
    CV_Assert(!tensor_proto.raw_data().empty() || !tensor_proto.float_data().empty()
                    || !tensor_proto.double_data().empty() || !tensor_proto.int64_data().empty());

I think we can add some dummy data or can remove the node.

[SOLVED]

i have removed the node at finalize function.

Conflicting files

Please rebase patch on latest commits from 3.4 branch:

git fetch updstream
git rebase -i upstream/3.4

( looks like it conflicts with #16472 - also please split related "Resize" test on smaller parts )

Conflicting files

Please rebase patch on latest commits from 3.4 branch:

Okay, thanks. Let me do that.

@dkurt , I just encountered a problem while loading
net = cv2.dnn.readNetFromONNX("/home/ashish/Downloads/Object Detection & Image Segmentation/yolov3/yolov3.onnx")
Here is the model.
It attempted to match SoftmaxSubgraph ( div node ).

cv2.error: OpenCV(4.2.0-dev) /home/ashish/opencv/modules/dnn/src/graph_simplifier.cpp:84: error: (-212:Parsing error) Input node with name const_fold_opt__570 not found in function 'getInputNodeId'

Here, const_fold_opt__570 is initializer.

So, it is neither node nor output of some node, I guess. What do you suggest about it?

@ashishkrshrivastava, I'd like to recommend to move it to separate issue / PR if you do not mind.

@dkurt , I just encountered a problem while loading
net = cv2.dnn.readNetFromONNX("/home/ashish/Downloads/Object Detection & Image Segmentation/yolov3/yolov3.onnx")
Here is the model.
It attempted to match SoftmaxSubgraph ( div node ).

cv2.error: OpenCV(4.2.0-dev) /home/ashish/opencv/modules/dnn/src/graph_simplifier.cpp:84: error: (-212:Parsing error) Input node with name const_fold_opt__570 not found in function 'getInputNodeId'

Here, const_fold_opt__570 is initializer.

So, it is neither node nor output of some node, I guess. What do you suggest about it?

@dkurt, Are you talking about moving only this issue to another issue/PR or everything ?

Yeah. So if this PR can already fix two mentioned issues (#16323 and #16475) - let's finish it. Or they also require this SoftMax fusion?

okay. for these two issues we dont need to worry about softamx fusion.

@ashishkrshrivastava, thank you!

@dkurt , In issue 16475 Unet model is obtained using Pytorch version (1.3) while modifications in this PR are according to Pytorch version (1.4). should I make changes according to version 1.3 or it is fine to continue with version 1.4(latest) ? I am able to load Unet exported using version 1.4 with current modifications.

@ashishkrshrivastava, If this PR can fix the issues - let's focus on current state. We need to be user oriented and if one have 1.4 version - we need to fix it here.

@ashishkrshrivastava excuse me, can u explain how can i use your fix in OpenCV 4.2.0 ?

@BadMachine, you need to build OpenCV from source. The fix has been merged after OpenCV 4.2.0

@dkurt Thanks for your answer!
But source code now build with errors...

PS: I know it should be in issue section, just for u know

@BadMachine, Please do not use closed PRs for bug reports - open a new issue including build flags and error logs (but not a screenshot).