search

thtang / CheXNet-with-localization

Weakly Supervised Learning for Findings Detection in Medical Images
https://www.csie.ntu.edu.tw/~yvchen/f106-adl/doc/HTCMedical.pdf
GNU General Public License v3.0
316 stars 108 forks source link

Weights not used in training batch #14

Open hugokitano opened 4 years ago

hugokitano commented 4 years ago

Hi! Thanks for this project. I noticed the weights for each training batch are not used. How did you incorporate them in training? I read in your report that your mean AUC ended up being better with them.

thtang commented 4 years ago

What do you mean weights for each training batch ?

hugokitano commented 4 years ago

image

In train.py, you define weights_sub, but they are not used for anything in the loss function.

thtang commented 4 years ago

Yup, the weights are originally used for loss function as described in the paper. However, there is no improvement regarding AUC score but makes the training code messy. If you've read different versions of the paper on arXiv, you will find out that they obtained the definitely same AUC number using different loss functions.