search

yewsiang / ConceptBottleneck

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

Questions about N_ATTRIBUTES for inception #13

Closed kittykg closed 2 years ago

kittykg commented 2 years ago

Hi guys, I'm looking at the code and I'm just curious why n_attribute is set to be 112 for the inception model. The CUB_processed uploaded on Codalab only contains 112 elements in attribute_label. However, the CUB 200_2011 dataset contains 312 attributes. Looking at CUB/data_processing.py, there seems no attribute selection process. Could you please explain the reason for the difference between the number of attributes used in the inception model and the dataset? Thanks a lot!

thaonguyen19 commented 2 years ago

The class-level attribute aggregation process (which reduces the total number of attributes used from 312 to 112) is described in the paper. Further code and data processing instructions can be found in the README https://github.com/yewsiang/ConceptBottleneck/blob/master/CUB/README.md. In particular, CUB/generate_new_data.py another file that you should consider running, besides CUB/data_processing.py. Hope that helps!

kittykg commented 2 years ago

The class-level attribute aggregation process (which reduces the total number of attributes used from 312 to 112) is described in the paper. Further code and data processing instructions can be found in the README https://github.com/yewsiang/ConceptBottleneck/blob/master/CUB/README.md. In particular, CUB/generate_new_data.py another file that you should consider running, besides CUB/data_processing.py. Hope that helps!

Yeah I checked the appendix, I missed it when reading the paper. I found the 112 attributes in the code now, thanks a lot! :)