Accuracy decrease after use the delegate

[ZIFENG278]

hi,

i used the edgeai-tidl-tools(08_02_00_05) to compile my speech recognize model(CNN), after i product the artifacts, i deployed it on the Beaglebone AI-64(TDA4VM), and run the infer with hardware acceleration, but the accuracy is super bad, it look like all wrong, but run without the delegate, the accuracy is perfect.

I want to ask it is when compile the model by OSRT just need one data(from the example of osrt just use two picture to compile and generate the artifacts) or need the full dataset to compile?

best,