search

AaronJackson / vrn

:man: Code for "Large Pose 3D Face Reconstruction from a Single Image via Direct Volumetric CNN Regression"
http://aaronsplace.co.uk/papers/jackson2017recon/
MIT License
4.52k stars 747 forks source link

About the training and testing data #126

Closed alalba221 closed 5 years ago

alalba221 commented 5 years ago

Hi Aaron, you said in #12 that you didn't train the network on HELEN and LFPW , so I trained the network on IBUG and AFW using 6000 images for training and about 800 for testing. Here is the loss a got: epoch 0 : train_loss 0.352367, eval_loss 0.387095 epoch 1 : train_loss 0.278556, eval_loss 0.420478 epoch 2 : train_loss 0.239114, eval_loss 0.385385 epoch 3 : train_loss 0.211572, eval_loss 0.436594 epoch 4 : train_loss 0.190551, eval_loss 0.464007 epoch 5 : train_loss 0.175974, eval_loss 0.484641 epoch 6 : train_loss 0.163320, eval_loss 0.493853 epoch 7 : train_loss 0.153328, eval_loss 0.565732 epoch 8 : train_loss 0.145184, eval_loss 0.531872 epoch 9 : train_loss 0.138747, eval_loss 0.520749 epoch 10 : train_loss 0.132051, eval_loss 0.595860 epoch 11 : train_loss 0.126847, eval_loss 0.558690 epoch 12 : train_loss 0.121628, eval_loss 0.688304 epoch 13 : train_loss 0.117325, eval_loss 0.581210 epoch 14 : train_loss 0.113913, eval_loss 0.597701 epoch 15 : train_loss 0.109812, eval_loss 0.629373 epoch 16 : train_loss 0.107197, eval_loss 0.650522 epoch 17 : train_loss 0.103911, eval_loss 0.615461 epoch 18 : train_loss 0.100986, eval_loss 0.663053 epoch 19 : train_loss 0.098672, eval_loss 0.642940 epoch 20 : train_loss 0.096295, eval_loss 0.654145 epoch 21 : train_loss 0.094171, eval_loss 0.625737 epoch 22 : train_loss 0.091768, eval_loss 0.762728 epoch 23 : train_loss 0.089807, eval_loss 0.760316 epoch 24 : train_loss 0.088324, eval_loss 0.612928 epoch 25 : train_loss 0.085970, eval_loss 0.817148 epoch 26 : train_loss 0.084619, eval_loss 0.792491 epoch 27 : train_loss 0.083374, eval_loss 0.687534

as the training loss decreasing, the testing loss keeps increasing. I am not sure if it is because I didn't adopt the Flips images as training data or something else. So would you tel me which folder you used for training and which for testing? Thanks

AaronJackson commented 5 years ago

I can't remember exactly. The exact sets we used are described in the paper. If your testing error is going up, perhaps you have generated the two sets of volumes with different scales or something.