I'm having trouble with predictions using a custom-trained graph converted from tensorflow tiny-yolo-v2 to NCS graph.
I trained a 1-class classifier and I modified the run.py example and adapt it from 20 to 1 class. (Only modified all the 125 instances to 30 for filters and the value 20 to 1 for classes)
The example is running but it returns a very bizarre response.
I tried the custom-trained model before conversion and is working perfectly. It returns a very good result on predictions, the compiled graph is giving hundred of invalid boxes and not a single one corresponds to ground-truth.
Is there anything that has to be done before compiling? Why is my model not working on the graph. Why is the graph not working but the original .pb does.
I used this command to compile
mvNCCompile -o output/ncs_tiny_yolo_v2.graph built_graph/tiny-yolo-voc-1c.pb -s 12
@frank.aislab Have you found any solution.I am having a similar issue. I trained a model with tiny-yolo-voc weights to detect number plates. it works fine with .pb and meta files. But with graph files i am not able to generate bounding boxes from the graph''s output. I referred this work ( https://github.com/fernandodelacalle/yolo-darkflow-movidius/blob/master/detection_camera.py, https://github.com/fernandodelacalle/yolo-darkflow-movidius/blob/8a2c07878e98fe370c72f8a6e951041fbb5... ) for generating bbox from graph's output.