As the title said. I have a MobileNetV2-SSDLite model trained on the GTSDB dataset using RMSprop with Caffe whose validation accuracy is around 0.75 mAP. When I deploy it to Arria 10 using OpenVINO, it gives ~0.75 mAP with float precision and FP16, but only gives ~0.2 mAP with other combinations, e.g. half precision and FP11.
I'm still trying to train MobileNetV2-SSDLite using SGD. Based on my previous experience on MobileNetV1-SSD, the same bug of RMSprop exists but it's gone if I train the model using SGD. I search a bit and find this paper: https://arxiv.org/pdf/1705.08292.pdf saying RMSprop-trained models don't generalize so well as SGD-trained ones, but I'm not sure if that's the explanation to my case.
Could somebody try to help me out of this? Thanks in advance!
( Pretrained models (V1-SSD (SGD) and V2-SSDLite (RMSprop)) and test scripts: https://drive.google.com/open?id=16mW-f-LQ90zQQeiqA-sCgIJ9QhZismFN
original training/testing and dataset preprocessing code: https://github.com/moderato/caffe/tree/ssd )