search

Project-MONAI / tutorials

MONAI Tutorials
https://monai.io/started.html
Apache License 2.0
1.85k stars 681 forks source link

UNETR‘s training accuracy is always below value in the paper #642

Closed AustinYuAo closed 2 years ago

AustinYuAo commented 2 years ago

Hi, I attempted to train UNETR, but I found the accuracy value always at about 0.80. I know the Vit was used in the encoder, Do I need to load the pre-training model when I use it? What can I do for training to make the accuracy value improve?

Nic-Ma commented 2 years ago

Hi @ahatamiz ,

Could you please help share some best practices or suggestions?

Thanks in advance.

ahatamiz commented 2 years ago

Hi @aubest , values reported in the paper are from 5 fold cross-validation, hence being better than any single fold results. In addition, we use a cohort of 80 images, which come from a privately-held dataset, for training. Please see section 4.3. I recommend using the pre-trained weights as provided in the Research Contribution Repository for fine-tuning.

AustinYuAo commented 2 years ago

@Nic-Ma, Thank you very much!

AustinYuAo commented 2 years ago

Hi@ahatamiz @achen46, Thanks for your reply. Your advice is very useful and it has enlightened me. I will try to train it again. I am hoping to communicate with you in the future. Thank you！