search

facebookresearch / Detic

Code release for "Detecting Twenty-thousand Classes using Image-level Supervision".
Apache License 2.0
1.86k stars 211 forks source link

Ablation for softmax ce using federated loss #35

Open jihwanp opened 2 years ago

jihwanp commented 2 years ago

Hi Is there any ablation that use softmax ce using federeated loss? CenterNet2 and Detic demonstrated that using federated loss is crucial for long tail distributed dataset like LVIS, and this sampling strategy has been used only for sigmoid ce. I found that you've implemented the federated loss for softmax ce version, so I wonder how much gap between sigmoid ce and softmax ce. Thanks

xingyizhou commented 2 years ago

Hi,

Great question! I haven't played a lot with it (I always prefer sigmoid CE in a large vocabulary). In my preliminary runs, Federated loss improves less in softmax CE than sigmoid CE. However, I do hear from my labmate that Federated loss could perform similarly in softmax CE under proper settings. I'll get back to you later.

Best, Xingyi