I am stuck at a curious problem with the OpenVINO model optimizer. I have a custom ONNX network which I would like to optimize using the MO. I am using the OpenCV that's shipped with OpenVINO to perform the final inference.
First I convert my onnx network to its equivalent IR representation.
python mo.py --input_model E:\cv_align.dll --framework onnx --output_dir E:\models\b1 --log_level DEBUG > log.txt
The output looks fine.
Common parameters:
- Path to the Input Model: E:\cv_align.dll
- Path for generated IR: E:\models\b1
- IR output name: cv_align
- Log level: DEBUG
- Batch: Not specified, inherited from the model
- Input layers: Not specified, inherited from the model
- Output layers: Not specified, inherited from the model
- Input shapes: Not specified, inherited from the model
- Mean values: Not specified
- Scale values: Not specified
- Scale factor: Not specified
- Precision of IR: FP32
- Enable fusing: True
- Enable grouped convolutions fusing: True
- Move mean values to preprocess section: False
- Reverse input channels: False
ONNX specific parameters:
Model Optimizer version: 2019.2.0-436-gf5827d4
[ SUCCESS ] Generated IR model.
[ SUCCESS ] XML file: E:\models\b1\cv_align.xml
[ SUCCESS ] BIN file: E:\models\b1\cv_align.bin
[ SUCCESS ] Total execution time: 30.65 seconds.
I take the generated xml,bin and load it into my C++ program. That crashes when I forward the model with dummy inputs. Then I take my onnx and generate network directly from onnx, which works as expected.
Like so,
static cv::dnn::Net alignNet;
int main()
{
//initialise
//auto out = Align_init("E:\\cv_align.dll", 1);
auto out = Align_init("E:\\models\\b1\\cv_align.xml",
"E:\\models\\b1\\cv_align.bin", 1);
return 0;
}
///THIS ONE CRASHES AT POINT SHOWN
ALIGN_OUT Align_init(std::string xmlPath, std::string binPath, int batch_size)
{
assert(std::experimental::filesystem::exists(xmlPath));
assert(std::experimental::filesystem::exists(binPath));
alignNet = cv::dnn::readNetFromModelOptimizer(xmlPath, binPath);
alignNet.dumpToFile("E:\\models\\dump.dmp");
alignNet.setPreferableBackend(cv::dnn::DNN_BACKEND_INFERENCE_ENGINE);
alignNet.setPreferableTarget(cv::dnn::DNN_TARGET_OPENCL);
//initialise using dummy vars
std::cout << " -- Dummy image -- " << std::endl;
Image dummyImage1 = generateRandMat();
Image dummyImage2 = generateRandMat();
Image cImage = getCombinedImage(dummyImage1, dummyImage2);
auto dummyInput = imgToBlob(cImage);
alignNet.setInput(dummyInput);
auto dummyProb = alignNet.forward(); <== This statement throws a Microsoft C++ exception: InferenceEngine::details::InferenceEngineException
return ALIGN_OUT();
}
///THIS ONE'S FINE
ALIGN_OUT Align_init(std::string onnx_path, int batch_size)
{
assert(batch_size == 1);
//TODO: other batch sizes if required
alignNet = cv::dnn::readNetFromONNX(onnx_path);
alignNet.dumpToFile("E:\\models\\dump1.dmp");
alignNet.setPreferableBackend(cv::dnn::DNN_BACKEND_INFERENCE_ENGINE);
alignNet.setPreferableTarget(cv::dnn::DNN_TARGET_OPENCL);
//initialise using dummy vars
std::cout << " -- Dummy image -- " << std::endl;
Image dummyImage1 = generateRandMat();
Image dummyImage2 = generateRandMat();
Image cImage = getCombinedImage(dummyImage1, dummyImage2);
auto dummyInput = imgToBlob(cImage);
alignNet.setInput(dummyInput);
auto dummyProb = alignNet.forward(); <== This one's fine
return ALIGN_OUT();
}
Also just in case
typedef cv::Mat Image;
typedef std::vector<Image> Images;
Any idea if I am doing something wrong?
Thanks.
More info from dumps
ONNX dump -
digraph G {
"261" [label="261\nSlice\nOCV/CPU\n" fillcolor="#ffffb3" style=filled shape=box]
"262" [label="262\nConvolution\nkernel_size (HxW): 7 x 7\lstride (HxW): 2 x 2\ldilation (HxW): 1 x 1\lpad (HxW): (3, 3) x (3, 3)\lgroup: 1\lOCV/CPU\n" fillcolor="#ffffb3" style=filled shape=box]
"263" [label="263\nBatchNorm\nOCV/CPU\n" fillcolor="#ffffb3" style=filled shape=box]
"264" [label="264\nRelu\nOCV/CPU\n" fillcolor="#ffffb3" style=filled shape=box]
"265" [label="265\nPooling\nkernel_size (HxW): 3 x 3\lstride (HxW): 2 x 2\lpad (HxW): (1, 1) x (1, 1)\lpool: MAX\lOCV/CPU\n" fillcolor="#ffffb3" style=filled shape=box]
"266" [label="266\nConvolution\nkernel_size (HxW): 3 x 3\lstride (HxW): 1 x 1\ldilation (HxW): 1 x 1\lpad (HxW): (1, 1) x (1, 1)\lgroup: 1\lOCV/CPU\n" fillcolor="#ffffb3" style=filled shape=box]
"267" [label="267\nBatchNorm\nOCV/CPU\n" fillcolor="#ffffb3" style=filled shape=box]
"268" [label="268\nRelu\nOCV/CPU\n" fillcolor="#ffffb3" style=filled shape=box]
"269" [label="269\nConvolution\nkernel_size (HxW): 3 x 3\lstride (HxW): 1 x 1\ldilation (HxW): 1 x 1\lpad (HxW): (1, 1) x (1, 1)\lgroup: 1\lOCV/CPU\n" fillcolor="#ffffb3" style=filled shape=box]
"270" [label="270\nBatchNorm\nOCV/CPU\n" fillcolor="#ffffb3" style=filled shape=box]
"271" [label="271\nEltwise\nOCV/CPU\n" fillcolor="#ffffb3" style=filled shape=box]
"272" [label="272\nRelu\nOCV/CPU\n" fillcolor="#ffffb3" style=filled shape=box]
"273" [label="273\nConvolution\nkernel_size (HxW): 3 x 3\lstride (HxW): 1 x 1\ldilation (HxW): 1 x 1\lpad (HxW): (1, 1) x (1, 1)\lgroup: 1\lOCV/CPU\n" fillcolor="#ffffb3" style=filled shape=box]
"274" [label="274\nBatchNorm\nOCV/CPU\n" fillcolor="#ffffb3" style=filled shape=box]
"275" [label="275\nRelu\nOCV/CPU\n" fillcolor="#ffffb3" style=filled shape=box]
"276" [label="276\nConvolution\nkernel_size (HxW): 3 x 3\lstride (HxW): 1 x 1\ldilation (HxW): 1 x 1\lpad (HxW): (1, 1) x (1, 1)\lgroup: 1\lOCV/CPU\n" fillcolor="#ffffb3" style=filled shape=box]
"277" [label="277\nBatchNorm\nOCV/CPU\n" fillcolor="#ffffb3" style=filled shape=box]
"278" [label="278\nEltwise\nOCV/CPU\n" fillcolor="#ffffb3" style=filled shape=box]
....
...
many more lines
IR dump -
digraph G {
"545" [label="545\n\nDLIE/CPU\n" fillcolor="#fdb462" style=filled shape=box]
"_input" -> "545"
}
Seems like it just skips every layer in the middle somehow.
It was a problem with the OpenCV shipped with OpenVINO. The program works as expected when using Inference Engine Core API.