Andrei C comments

Results 5 comments of


                                            Andrei C

CORAL loss is defined differently from the original paper

tldr, no error, this code is "correct" but the magnitude of the loss is not scaled correctly. 1. deep coral uses squared frobenius loss so sqrt is not necessary; original...

I think you have a error.

tldr, no error, this code is "correct". See my answer for this [issue ](https://github.com/SSARCandy/DeepCORAL/issues/17).

Test results of official trained model(download_PieAPPv0.1_PT_weights.sh) not same with paper

@JiadiMo What SROCC did you achieve? FYI, on my end, after a quick test with the provided pretrained weights and similar (although modified) setup SROCC=0.79185 KROCC=0.60356 PLCC=0.78609 RMSE=0.09523 on PieAPP...

White text change to grey when exported to PDF

Also have this issue. @adithyaappu thank you for sharing a workaround

Ask regarding the dataset

https://codalab.lisn.upsaclay.fr/competitions/1567 It's part of the NTIRE22 competition. Make an account and you can download train/val/test data. Only the training set has publicly available quality labels.