kendryte
Face Detection stuck at kpu_run_kmodel function
Hi,
I am trying to run the face detection demo. Unfortunately, the code is stuck at kpu_run_model function when being executed. The details are as follows.
I adopted the current src code in the repo to fit my needs and created my own .bin file. Now using my bin file, the kmodel file provided in the repo, and a simple flash-list.json, I was able to flash the model onto the chip and run the demo. Everything went smoothly.
Next, I tried replicating the pipeline from the start, training my own model from the original repo, converting it to caffe, and then finally converting it to kmodel. I did not get any form of error during the conversion or even during execution, however while running the demo the C++ code is unfortunately stuck at kpu_run_kmodel function, and the g_ai_done_flag never becomes 1.
I was able to pinpoint the issue to conversion from caffemodel to kmodel. Even when I simply use the caffemodel provided by you in the repo and try to convert it to kmodel through nncase, I still face the exact same issues. I built nncase from the latest repo commit, so I thought that might be the issue, and instead decided to use the latest release provided. But I actually got a segmentation fault error in that case while doing the conversion.
I have attached the kmodel provided by you (slim-320.kmodel), caffemodel provided by you (slim-320.caffemodel), kmodel I got from your caffemodel (slim-320-new.kmodel) and my bin file (ultralite_face.bin).
The combination of slim-320.kmodel and ultralite_face.bin works flawlessly.
The combination of slim-320-new.kmodel and ultralite_face.bin causes the issue I mentioned above.
