search

xmed-lab / DHC

MICCAI 2023: DHC: Dual-debiased Heterogeneous Co-training Framework for Class-imbalanced Semi-supervised Medical Image Segmentation
MIT License
52 stars 3 forks source link

about epoch #5

Closed yinguanchun closed 10 months ago

yinguanchun commented 10 months ago

Hello, the training epoch of synapse in code is 500，but it is 300 in the paper. I want to know the epoch of Synapse and AMOS. Tranks!

yinguanchun commented 10 months ago

In addition, I noticed the following code during the loading of the data: QQ图片20240122000444 However, the synapse dataset has a wide range of values, I wonder if there's a problem with that.

yinguanchun commented 10 months ago

if my understanding of the code is not wrong, I find that the size of the data does not change during the training phase, but the height and width become half of the original when interfering, is there any reason for this operation？

McGregorWwww commented 10 months ago

Hi,

  1. for the number of epochs, there's no major difference since we use early stopping strategy, the training will be ended around 200 epochs.
  2. we use the clip range following previous works such TransUNet.
  3. the size of the data is also changed in augmentations during training, it's for reducing the GPU costs.