Showing results for 
Search instead for 
Did you mean: 
New Contributor I

RMSprop-trained models suffering low detection accuracy with FP11 and/or half precision

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: 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:

original training/testing and dataset preprocessing code: )

0 Kudos
1 Reply
New Contributor I

Problem solved after upgrading to R4.