sekioo

Results 6 comments of sekioo

While fine tuning the Swin UNETR, Training loss is not decreasing and training gets crashed after 10 epohs

> [@Mgithus](https://github.com/Mgithus) it shows ET as 0 because the segmentation labels are different. BRATS labelling "ConvertMultiChannel..." takes 0, 1, 2, 4 based on previous datasets segmentation labels (..., 2021). The...

Epoch 556/1000, Iter 182/190, Loss 2.9084, fusecross:0.0088, fusedice:0.1742,sepcross:0.0643, sepdice:1.1391,prmcross:0.0691, prmdice:1.4530

![微信截图_20241127204510](https://github.com/user-attachments/assets/3e482791-c490-4db6-a060-8b918e295de3) 我读取了hetero_training.txt文件，从图片上看Loss到epoch688时还在剧烈震荡

Epoch 556/1000, Iter 182/190, Loss 2.9084, fusecross:0.0088, fusedice:0.1742,sepcross:0.0643, sepdice:1.1391,prmcross:0.0691, prmdice:1.4530

![微信截图_20241128175912](https://github.com/user-attachments/assets/c9388ae1-c211-4530-9382-0390ac82a330) 谢谢您的回答，我是这方面的新手，我想问下1000次训练后，还是发现sepdice和prmdice这两个损失普遍偏高，请问下这是正常的吗

Epoch 556/1000, Iter 182/190, Loss 2.9084, fusecross:0.0088, fusedice:0.1742,sepcross:0.0643, sepdice:1.1391,prmcross:0.0691, prmdice:1.4530

作者我还有个问题请教，就是这里打印的是loss的值，那我该如何去计算dice或者重合度、准确率这写值呢

Epoch 556/1000, Iter 182/190, Loss 2.9084, fusecross:0.0088, fusedice:0.1742,sepcross:0.0643, sepdice:1.1391,prmcross:0.0691, prmdice:1.4530

谢谢您的细心回答，请问下fusecross:0.1324, fusedice:0.4548,sepcross:0.6408, sepdice:2.0106,prmcross:0.5424, prmdice:2.1771,这几个loss分别表示什么意思呢

Epoch 556/1000, Iter 182/190, Loss 2.9084, fusecross:0.0088, fusedice:0.1742,sepcross:0.0643, sepdice:1.1391,prmcross:0.0691, prmdice:1.4530

谢谢您的细心回答，请问下fusecross:0.1324, fusedice:0.4548,sepcross:0.6408, sepdice:2.0106,prmcross:0.5424, prmdice:2.1771,这几个loss分别表示什么意思呢