search

ArchipLab-LinfengZhang / Object-Detection-Knowledge-Distillation-ICLR2021

The official implementation of ICLR2021 paper "Improve Object Detection with Feature-based Knowledge Distillation: Towards Accurate and Efficient Detectors".
MIT License
57 stars 6 forks source link

some difference between the paper and the provided code #5

Open jy-Hamlet opened 3 years ago

jy-Hamlet commented 3 years ago

Hello I find some settings described in the paper are different from that in the code. In the paper, they are:

model α β γ T
one-stage 4e-4 2e-2 4e-4 0.5
two-stage 7e-5 4e-3 7e-5 0.1

But in the code, they are:

model α β γ T
one-stage 2e-2 4e-4 4e-4 0.1
two-stage 4e-3 7e-5 7e-5 0.5

Which version is accurate?