Open AustinYuAo opened 9 months ago
@tangy5 I'm sorry for bothering you. Can you help me with this confusion? Thank you.
Hi @AustinYuAo which subdirectory are you seeing this in specifically? We can look at who contributed specific code you're looking at and ask them as well.
I haven't been involved in these specific contributions myself but I can only guess that the differing training requirements for transformer networks versus CNNs is the cause here. Transforms tend to take a long time to train correctly and would take differing amounts for different tasks such that the epoch counts would be determined experimentally. CNNs like UNet are perhaps easier to train in this regard, or one can get good results for all tasks with one fixed and sufficiently high epoch count.
Hi @ericspod , I apologize for the delayed response. I specifically saw it in this link: https://github.com/Project-MONAI/research-contributions/tree/main/SwinUNETR/BTCV.
Hi @AustinYuAo I'm not sure from this what I can say to help. @ahatamiz this repository was initially your contribution, would you be able to comment here? Thanks.
Hello, sorry to bother you. When comparing Swin-UNetR and nnU-Net, I noticed that Swin-UNetR requires a significantly different maximum number of epochs for different segmentation tasks, while nnU-Net uses the same maximum epoch value. Can you help me resolve this confusion?