I'm trying to test object detection with openvino, I was able to convert using model optimizer, however, I am unable to calibrate the converted IR model to INT8. I'm trying following command. Is this model not supported yet? Is there a way around? I am on release - 2019.3.376
For now, I am only trying simplified mode, ideally I would like to run accuracy checker as well.
python3 /opt/intel/openvino/deployment_tools/tools/calibration_tool/calibrate.py -m ./VGG16_faster_rcnn_final.xml -sm -s /media/amiya/KINGSTON/VOC2017_sub/. -o ./out/
[ INFO ] Collect inputs...
terminate called after throwing an instance of 'InferenceEngine::details::InferenceEngineException'
what(): Just networks with one input are supported.
I have also tried ignoring im_info layer with -ignore_layer_names="im_info", but that did not help.
Hi, trivedi, amiya
Can you let me know which pre-trained model you are using from which github repository? Assuming you are not downloading it from tensorflow object detection api model zoo, right? Also please provide your conversion command as well, thank you.
Thanks for your reply, the model is downloaded from https://dl.dropboxusercontent.com/s/o6ii098bu51d139/faster_rcnn_models.tgz?dl=0
The command for model optimizer was something like -
/opt/intel/openvino/deployment_tools/model_optimizer/mo_caffe.py --input_model VGG16_faster_rcnn_final.caffemodel --input_proto test.prototxt
Hi, Trivedi, amiya,
Sorry for late response, I was occupied by other urgent task. I've downloaded your model, but I didn't see the prototxt inside. I don't know which model repository did you use from your dropbox link, so could you give me the prototxt you are using? Currently I assume you use the prototxt from https://raw.githubusercontent.com/rbgirshick/py-faster-rcnn/master/models/pascal_voc/VGG16/faster_rcnn_end2end/test.prototxt
From my test, the model can be converted and run with correct result, but after I run with the calibration tool with the configuration revised, I still can't go through the calibration process, I have reached inside to know where the problem is. Attached the yaml file I used first and will keep you updated once I got the response.
Thanks for your response.
Yes, I used the same test.prototxt as it was pointed to by the README inside object_detection_demo_faster_rcnn.
I did not use any yml as I used the simplified mode calibration command as mentioned in my post before..
Can you please share your calibration command line?
It is not clear from your response whether you got same error or different, please let me know your exact test result.
Hope to hear back soon.
I see, did you try simplified mode command ( -sm ) as in my first post? Just curious if you get the same result. It would be good to get accuracy checker to also work though, but I was testing simplified mode first just check basic operation.
I don't think the simplified mode works for the faster_rcnn model for now, since it has two inputs, and I didn't see where two inputs is supported by simplified mode configuration compare to the standard mode. As you can see from the yaml file I provided, I specify the im_info as "IMAGE_INFO" which is one of the mechanism supported from standard calibration, you can refer to section Specifying model inputs in config in "https://docs.openvinotoolkit.org/latest/_tools_accuracy_checker_accuracy_checker_launcher_dlsdk_launcher_readme.html "