Firstly, the OpenVINO toolkit is amazing! I successfully converted my custom YOLOv3-based model from TensorFlow 2.0 framework to OpenVINO, in all three FP32/FP16/INT8 precision. The model is deployed successfully in intel CPU and iGPU.
However, my custom model could not be deployed in MYRIAD device. I encounter the error stated below:
/home/pi/openvino/inference-engine/src/vpu/graph_transformer/src/stages/static_shape_nms.cpp:126 StaticShapeNMS with name StaticShapeNonMaxSuppression_7550 parsing failed: the only supported value for softNMSSigma is 0, while it actually equal to 0.5
Based on the above error code, I suspect the error is due to the "tf.image.non_max_suppression_with_scores" in my code. So, I change the softNMSSigma to 0 in my model. After convert the edited model to OpenVINO IR format, I faced a new error code, as shown below:
/home/pi/openvino/inference-engine/src/vpu/graph_transformer/src/stages/static_shape_nms.cpp:138 StaticShapeNMS with name StaticShapeNonMaxSuppression_7542 parsing failed: selected_scores output is not supported
My speculation is that:
I changed the softNMSSigma to 0.0 instead of 0, because the function expects me to pass a float instead of int value. Is it due to this issue?
FYI, I'm using Ubuntu 18 + TensorFlow 2.5 version + OpenVINO 2021.1
Any suggestion to solve this? Thanks in advance!
