Add a test case for handling output_shape property for deconv layer.

[kohei-us]

This is related to opencv/opencv#13989.

[dkurt]

@kohei-us, Please show us how these test files are generated. Add a modified https://github.com/opencv/opencv_extra/blob/master/testdata/dnn/onnx/generate_onnx_models.py

[kohei-us]

@kurt I used Keras to create a model and then convert it to onnx. I noticed that the script uses PyTorch to generate models. Would a use of Keras be okay in that script, or is PyTorch required?

[dkurt]

@kohei-us, If it's possible to do all the conversions in python I'd like to suggest to add Keras model in this script. Please use try except scope to check if TensorFlow exists.

[kohei-us]

I've now stumbled upon one major issue.

Originally, I created a keras model and converted to onnx using an older version of onnxmltools, and that converter added the output_shape attribute to ConvTranspose, which is what this test case is designed to test.

However, since then the onnxmltools has changed quite a bit, and when using their latest code (it's now been split into a separate project called keras-onnx), the converted onnx model no longer contains the output_shape attribute... :-(

I have also tried using PyTorch to create an onnx model with ConvTranspose, but that one doesn't have output_shape either.

Hmm...

[dkurt]

@kohei-us, May I ask you to add a line

std::cout << layerParams << std::endl;

at
https://github.com/opencv/opencv/blob/c3cf35ab63c04fb1d7b2f6760128f42c20cac0e1/modules/dnn/src/onnx/onnx_importer.cpp#L585-L594?

So we can check all the parameters except output_shape. Please show the output.

[kohei-us]

@dkurt Sure thing. Here is the output for the new test data I've added:

dilation_h : 1
dilation_w : 1
group : 1
kernel_h : 4
kernel_w : 4
output_shape : -1, 1, 83, 83
pad_mode : SAME
stride_h : 3
stride_w : 3

[dkurt]

@kohei-us, May I also ask you to specify which size has input blob for this layer?

[kohei-us]

@dkurt Sure. It's (1, 1, 28, 28).

[dkurt]

@kohei-us, This way deconvolution layer uses the following formula to compute output shape:

        else if (padMode == "SAME")
        {
            outH = stride.height * (inpH - 1) + 1 + adjustPad.height;
            outW = stride.width * (inpW - 1) + 1 + adjustPad.width;
        }

So

3 * (28 - 1) + 1 + 0 = 82

However user specified output shape 83 that means we need to add extra padding called adjust padding:

3 * (28 - 1) + 1 + 1 = 83

May I ask you to check how OpenCV manage it for TensorFlow models: https://github.com/opencv/opencv/blob/master/modules/dnn/src/tensorflow/tf_importer.cpp:

            const int outH = outShape.at<int>(1);
            const int outW = outShape.at<int>(2);
            if (layerParams.get<String>("pad_mode") == "SAME")
            {
                layerParams.set("adj_w", (outW - 1) % strideX);
                layerParams.set("adj_h", (outH - 1) % strideY);
            }
            else if (layerParams.get<String>("pad_mode") == "VALID")
            {
                layerParams.set("adj_w", (outW - kernelW) % strideX);
                layerParams.set("adj_h", (outH - kernelH) % strideY);
}

So you can just read output shape and compute adjust padding:

                layerParams.set("adj_w", (83 - 1) % 3);  // == 1
                layerParams.set("adj_h", (83 - 1) % 3);