search

FlamieZhu / Balanced-Contrastive-Learning

Code Release for “Balanced Contrastive Learning for Long-Tailed Visual Recognition”
MIT License
99 stars 13 forks source link

Could the v1/v2/v3 input data be same? #3

Open whuhxb opened 1 year ago

whuhxb commented 1 year ago

Hi @FlamieZhu

Could the v1/v2/v3 input data be the same for the proposed BCL algorithm? Or must the v2/v3 be with same input different from v1 data augmentation? What's the influence of the same v1/v2/v3 for the proposed BCL loss? Thanks.

Xiaobing

FlamieZhu commented 1 year ago

Hi, Thanks for your interest. Data augmentation v1/v2/v3 can be the same or different. Since contrastive learning is more sensitive to data augmentation, we recommend that v2 and v3 be the same and v1 be different. It will bring a small performance gain.

whuhxb commented 1 year ago

Hi, Thanks for your interest. Data augmentation v1/v2/v3 can be the same or different. Since contrastive learning is more sensitive to data augmentation, we recommend that v2 and v3 be the same and v1 be different. It will bring a small performance gain.

Hi, Thanks for your kind reply. I will try data augmentation with different v2/v3 and v1 to see the performance.