whai362 / pan_pp.pytorch

Official implementations of PSENet, PAN and PAN++.
Apache License 2.0
439 stars 90 forks source link

Training from scratch 达不到论文PAN在Total Text上汇报的结果 #33

Closed Pengfei-Hu closed 3 years ago

Pengfei-Hu commented 3 years ago

config file:config/pan/pan_r18_tt.py 配置文件与仓库中一致,只不过用了两张GPU训练。 最后的600epoch的模型测试结果为Precision:_0.834923757993___/Recall:_0.771597096189/Hmean:_0.802012305697,达不到论文中的结果precision 88, recall79.5,hmean83.5 我训练时使用的TotalText的GroundTruth为新发布的,可参考https://github.com/cs-chan/Total-Text-Dataset/tree/master/Groundtruth/Text 请问您可以提供更多的实现细节吗?

whai362 commented 3 years ago

感觉有点奇怪啊,我觉得有可能是以下原因: (1)我用的是旧版tt,https://drive.google.com/file/d/19quCaJGePvTc3yPZ7MAGNijjKfy77-ke/view?usp=sharing,新版tt的点我没有试过; (2)你试试看4卡训练?2卡和4卡的bn有些不一样。但是应该不会造成掉点。 (3)tt的验证集有点小,可能测试结果有一定随机性。 个人认为可能是新版数据集(即第一点)造成的。