search

ailab26 / pfld-lite

A re-implementation of PFLD, https://arxiv.org/abs/1902.10859
45 stars 12 forks source link

angle loss #2

Open DayBreak-u opened 5 years ago

DayBreak-u commented 5 years ago

hello,how use the angel loss,Is it necessary to add auxiliary network of Angle prediction？

xindongzhang commented 5 years ago

Thanks for your feedback, I will soon upload the full version of PFLD. In my experiments, it is not necessary to add auxiliary network. But from theoretical perspective, adding auxiliary network to weight the importance of the batch training can ease the data imbalance in facial landmark detection.

xindongzhang commented 5 years ago

@ouyanghuiyu FYI

DayBreak-u commented 5 years ago

thanks! convert mxnet to caffe is easy,why try it ! In that case, you can use residuals！

xindongzhang commented 5 years ago

thanks! convert mxnet to caffe is easy,why try it ! In that case, you can use residuals！

Yep, caffe seems like a good alternative to ONNX representation, thanks for your advice.

DayBreak-u commented 5 years ago

Convert to onnx is due to use mnn? caffe2mnn is available

xindongzhang commented 5 years ago

Convert to onnx is due to use mnn? caffe2mnn is available

Honestly, I just wanna get through the onnx-mnn pipeline, but caffe-mnn、pb-mnn seem more stable than onnx-mnn when deploying models to edge-devices.