I trained a segmentation model (jsegnet21v2) using the Caffe Jacinto framework and it performs well on the validation set. I used the TIDL import tool to convert it to TIDL format and the results are noticeably worse. I am running the TIDL model on an AM57x machine and tried changing import parameters such as quantRoundAdd and numParamBits as suggested in other forum posts, but the results were still not close enough to the original Caffe Jacinto model. How else can I minimize the optimizations happening in the import tool? Accuracy is important for my application rather than speed.
Additionally, other forum posts mentioned parameters such as createParams.quantHistoryParam1, createParams.quantHistoryParam2, and createParams.quantMargin, but I don't see the documentation for them in http://software-dl.ti.com/processor-sdk-linux/esd/docs/05_00_00_15/linux/Foundational_Components_TIDL.html#import-process
