System Information

OpenCV version: 4.7 release
OpenCV Python version: 4.7.0.68
OS: macOS 13.0.1
Python version: 3.9.6
Compiler: Apple clang version 14.0.0 (clang-1400.0.29.202)

Detailed description

I made two attention models of ONNX format. They have the same structure and the only difference is input shape:

• One is [1, 197, 768], the original shape from a vision transformer; Click to download multi_head_attention.onnx.zip.

• The other one is [1, 5, 12], a shrinked shape for debugging and stuff. Click to download multi_head_attention.onnx.zip.

I found if using Python interfaces with the correct shape set, the one of input shape [1, 197, 768] can run successfully, but the other one of input shape [1, 5, 12] fails:

[ERROR:0@110.035] global net_impl.cpp:1164 getLayerShapesRecursively OPENCV/DNN: [Permute]:(onnx_node_output_0!Transpose.0.out): getMemoryShapes() throws exception. inputs=1 outputs=0/1 blobs=0
[ERROR:0@110.035] global net_impl.cpp:1167 getLayerShapesRecursively     input[0] = [ 1 5 ]
[ERROR:0@110.035] global net_impl.cpp:1177 getLayerShapesRecursively Exception message: OpenCV(4.7.0) /Users/opencv-cn/GHA-OCV-1/_work/opencv-python/opencv-python/opencv/modules/dnn/src/layers/permute_layer.cpp:163: error: (-215:Assertion failed) (int)_numAxes == inputs[0].size() in function 'getMemoryShapes'

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
cv2.error: OpenCV(4.7.0) /Users/opencv-cn/GHA-OCV-1/_work/opencv-python/opencv-python/opencv/modules/dnn/src/layers/permute_layer.cpp:163: error: (-215:Assertion failed) (int)_numAxes == inputs[0].size() in function 'getMemoryShapes'

But if I try to debug in C++, for example adding test cases in test_onnx_importer.cpp, they all pass. So I guess there is something wrong with the python interface. Reproducers are attached below.

Steps to reproduce

Python:

import numpy as np
import cv2 as cv

# This passes!
net = cv.dnn.readNet("original_scale/models/multi_head_attention.onnx")
x = np.random.rand(1, 197, 768).astype(np.float32)
net.setInput(x)
net.forward()

# This fails!
net = cv.dnn.readNet("scale_1_5_12/models/multi_head_attention.onnx")
x = np.random.rand(1, 5, 12).astype(np.float32)
net.setInput(x)
net.forward()

C++: Add the following test cases in test_onnx_importer.cpp

TEST_P(Test_ONNX_layers, MHA_ORIG)
{
    Net net = readNet(findDataFile("original_scale/models/multi_head_attention.onnx", true));

    std::vector<int> blob_shape{1, 197, 768};
    Mat blob(blob_shape, CV_32FC1);
    randn(blob, 1.f, 0.f);

    net.setInput(blob);
    net.forward();
}

TEST_P(Test_ONNX_layers, MHA_SMALL_SCALE)
{
    Net net = readNet(findDataFile("scale_1_5_12/models/multi_head_attention.onnx", true));

    std::vector<int> blob_shape{1, 5, 12};
    Mat blob(blob_shape, CV_32FC1);
    randn(blob, 1.f, 0.f);

    net.setInput(blob);
    net.forward();
}

Issue submission checklist

• I report the issue, it's not a question
• I checked the problem with documentation, FAQ, open issues, forum.opencv.org, Stack Overflow, etc and have not found any solution
• I updated to the latest OpenCV version and the issue is still there
• There is reproducer code and related data files (videos, images, onnx, etc)