search

Vanint / SADE-AgnosticLT

This repository is the official Pytorch implementation of Self-Supervised Aggregation of Diverse Experts for Test-Agnostic Long-Tailed Recognition (NeurIPS 2022).
MIT License
146 stars 20 forks source link

variable s in ResNet_s #6

Closed sunhappy6900 closed 2 years ago

sunhappy6900 commented 2 years ago

Hi, Why did you add a variable s(s=30) in ResNet_s

Vanint commented 2 years ago

Hi, the main function of "s" is to adjust the logit magnitude of the cosine classifier if you use such a classifier. In contrast, if you use a linear classifier, "s" would be adjusted to 1 according to the code.

Moreover, this architecture is following RIDE (please refer to https://github.com/frank-xwang/RIDE-LongTailRecognition/blob/main/model/ldam_drw_resnets/ride_resnet_cifar.py). For simplicity, we do not change this hyper-parameter.

sunhappy6900 commented 2 years ago

thank you~