D-X-Y/landmark-detection

Novice ask a question

Opened this issue · 4 comments

Hello, I am a newcomer in the field of computer vision. According to the instructions you gave on github, I downloaded the pre-trained model in snapshots on Google Cloud and found that NME = 4.04, but it was 3.96 in the training log. I would like to ask what is the reason for this?Is it because the source code of the project is not fully given? I know this question might be idiot, but please forgive me for being a novice.Thanks.

D-X-Y commented

Would you mind to provide the project name (SBR or SAN)? In addition, please provide the Python version and PyTorch version?

I am so sorry. The project is SAN. And Python 3.6.9 with Pytorch 0.4.1

D-X-Y commented

I'm not sure about the reasons, the source codes of SAN are all released.
The log and pre-trained model on Google Drive also uses the same codes as in this repo.
Right now, I'm working on another project. Once I have some spare time, I will look into this problem.

QcQcM commented

I have met the same question when I use SAN code to train and test.
the evaluation result of epoch 99 printed by training is :(nms on 3 datasets is 3.56,8.44,4.51)
2020-12-08 14-35-41 的屏幕截图
but when I load the checkpoint and evaluate, the result is different(nms on 3 datasets is 3.71,8.57,4.66).
2020-12-08 14-40-40 的屏幕截图