To test model's performance, I feed a series of video frames one by one and check model's output confidence.
And I discover that if I feed frames from frame 5, the model output confidence 0.797 in frame 60; however, if I feed frames from frame 20, the model output confidence 0.831 of frame 60, the same frame with the previous case.
This difference occurs only on android devices. PC doesn't have this problem. And on android devices, this difference occurs no matter what processor the model runs on, CPU, GPU and DSP.
Then I do another experiment: I create a new model object every time when I feed a new frame and then the difference disappears. Are some bugs in model's execute() on Android implementation?
My SNPE library's version is 1.40. My dlc model is converted from tensorflow's pb file by snpe-tensorflow-to-dlc.