System information (version)

• OpenCV => 4.4.0 pre
• Operating System / Platform => Windows 64 Bit
• Compiler => Visual Studio 2019

Detailed description

We have a custom ONNX model wich can be load and inference with opencv 4.1.1 but currently can not do it with opencv 4.4.0 pre. We received the error below:

OpenCV(4.4.0-pre) D:\opencv-master\modules\dnn\src\graph_simplifier.cpp:79: error: (-212:Parsing error) Input node with name StatefulPartitionedCall/sequential/dense/MatMul not found in function 'cv::dnn::Subgraph::getInputNodeId'

the network structure is:

keras.layers.Conv2D(128,(5,5),input_shape=(height,width,1)),
    keras.layers.MaxPooling2D(pool_size=(2,2)),
    keras.layers.Dropout(0.49),
    keras.layers.Conv2D(64,(5,5)),
    keras.layers.MaxPooling2D(pool_size=(2,2)),
    keras.layers.Dropout(0.49),
    keras.layers.Flatten(),
    keras.layers.Dense(64, activation='relu'),
    keras.layers.Dense(64, activation='relu'),
    keras.layers.Flatten(),
    keras.layers.Dropout(0.45),
    keras.layers.Dense(36, activation='softmax')

The model is saved as TF saved_model and exported to ONNX with tf2onnx package.

Steps to reproduce

The model can be downloaded from here:
ONNX model

I try to load it with cv::dnn::readNetFromONNX but it cause the error above altough it was successfully loaded with previous opencv version (4.1.1).