Landmark accuracy

[cholihao]

I tried to follow your instructions to modify config.py and the result shows quite different from the pretrained model. The sshb result looks quite good but the e2e RPNL1Loss seems incorrect. The fixed params is set to ['^conv', '^ssh', '^rpn_cls', '^rpn_bbox',] and bounding box is set to new list_bbox_celeba.txt.

INFO:root:Epoch[14] Train-RPNAcc=0.999846 INFO:root:Epoch[14] Train-RPNAcc_BG=0.999957
INFO:root:Epoch[14] Train-RPNAcc_FG=0.992340 INFO:root:Epoch[14] Train-RPNL1Loss=0.071680 INFO:root:Epoch[14] Train-RPNAcc=0.999226
INFO:root:Epoch[14] Train-RPNAcc_BG=0.999249
INFO:root:Epoch[14] Train-RPNAcc_FG=0.998211 INFO:root:Epoch[14] Train-RPNL1Loss=0.530812 INFO:root:Epoch[14] Train-RPNAcc=0.999431
INFO:root:Epoch[14] Train-RPNAcc_BG=0.999443 INFO:root:Epoch[14] Train-RPNAcc_FG=0.985772 INFO:root:Epoch[14] Train-RPNL1Loss=1.053143
INFO:root:Epoch[14] Time cost=9396.422
lr change to 4e-05 in batch 47484
saving final checkpoint 47484
INFO:root:Saved checkpoint to "model/e2e-0000.params"

Could you tell what might be the reason?