search

yewsiang / ConceptBottleneck

Concept Bottleneck Models, ICML 2020
MIT License
174 stars 30 forks source link

Model choice for CUB #21

Closed ThomasNorr closed 2 years ago

ThomasNorr commented 2 years ago

Hello,

thanks for your interesting work. I was wondering why you chose InceptionV3 for the experiments on CUB rather than Resnet18/50 which seems to be more common in the literature?

I am grateful for some help. Greetings

kohpangwei commented 2 years ago

Hi! We had no particular reason -- I think we were just trying to match previous work which used InceptionV3.

ThomasNorr commented 2 years ago

Thanks for the answer. Do you by any chance have tried resnet50 and have some hyperparameters/results for them? I am trying to compare my work to yours on the basis of resnet50 too, as a lot of the competitors have used that (and my method seems to be benefit from resnet).

ThomasNorr commented 2 years ago

Additionally, I think there is an error in CUB/scripts/experiment.sh: For the third seed of the joint model with lambda=0.01 the scheduler step is 20 instead of 1000

kohpangwei commented 2 years ago

Oh, thank you for catching that error! I've fixed it. I'm sorry but we didn't try resnet50, so we don't have any hyperparameters or results for them.