do you have special training strategies? error:Yaw: 8.7961, Pitch: 6.7705, Roll: 5.8057, my training models can not reach your results as paper said

natanielruiz / deep-head-pose

:fire::fire: Deep Learning Head Pose Estimation using PyTorch.

Other

1.58k stars 367 forks source link

do you have special training strategies? error:Yaw: 8.7961, Pitch: 6.7705, Roll: 5.8057, my training models can not reach your results as paper said #64

Open maliho0803 opened 5 years ago

maliho0803 commented 5 years ago

below is my best results. resnet_multi-loss: Test error in degrees of the model on the 1969 test images. Yaw: 8.7961, Pitch: 6.7705, Roll: 5.8057

resnet50 only: Test error in degrees of the model on the 1969 test images. Yaw: 15.6866, Pitch: 7.1914, Roll: 6.2253

natanielruiz commented 5 years ago

Not really. I used the exact code that is online to achieve the results in the paper. Try to monitor your validation loss and do early stopping when it is minimal.

maliho0803 commented 5 years ago

thank you for your reply!!! I use batch_size of 64, maybe this will decress the accuracy?

One more question, I validate AFLW2000 during training every epoch, but it get different results as I test AFLW2000 just using testing code with the same model, I check the weight is the same and I have set model.eval(), do you know why I get such difference?

tx1994108 commented 5 years ago

@maliho0803 .hello, I want to know what the parameters of your training are. I still have 10 losses. And the loss is still very volatile. thank you.