Closed liyy201912 closed 5 months ago
The 80% top1 performance is based on distillation. Even train from scratch without distillation, 72.3% is too low (it seems that your may overfit to the training data where there is a large gap between top 1 acc on training and validation). Can you reproduce the mobilenet_v3 baseline which is 75.2%? Yes, sure. we will release the pretrained model in the future.
I tried to reproduce the 80% top1 performance of SlimMobilenet(V5), but can only get around 72.3% top 1 accuracy. Would you release the pretrained model in the future?