xiangly55/lfme

Code for "Learning From Multiple Experts: Self-paced Knowledge Distillation for Long-tailed Classification", ECCV 2020 Spotlight

lfme's Issues

KeyError

Dear author：
使用config\ImageNet_LT\many_shot.py时，报错KeyError: 'distill_shot_phases'

Questions about the equation in the paper

Thanks for your great work and contribution. I have some questions about the equation in the paper.

1. As section 4.3 described (the line behind the equation f(v_i^k)=... ), v_i^{(1)}=pi (Nsl / Nsmin). I'm wondering if the authors made some mistakes. It should be v_i^{(1)}=pi (Nsmin / Nsl)?

2. As described in section 4.3,
   Since the whole dataset is long-tailed, while we select samples from easy to hard, we also wish to select as uniform as possible across all subsets at the beginning of the training, and gradually add more hard samples as the epoch increases. In other words, at the first epoch we wish to select all the samples in the subset with lowest shots Smin (i.e. classes in Smin have the smallest number of samples) and same amount of samples in other subsets, and gradually add more samples until all the samples in all subsets are selected in the last epoch.
   From the equation f(v_i^k)=..., I don't understand how the hard samples and the easy samples in the same subset can be weighted in different ways in the previous and later epoch. In other words, at the 1st epoch, the easy samples have larger weights than the hard samples in the same subset. However, at the E_th epoch, the easy samples also have larger weights than the hard samples. So, I'm confused about how to implement the idea of "gradually add more hard samples as the epoch increases" as described in the paper.

RuntimeError: Found dtype Double but expected Float

I want to use pretrained model, but when I train, it occurs problem, I can't find where the type is Double and change them, please give me some advice, thank you!

Discrepancy of performance loss implementation

Thanks for the work!
在研究代码的过程中发现关于CE loss的计算，似乎原文公式与代码有些差别：
1）原文中对样本求取CEloss之后，再根据每个样本的vi进行加权求和；

2）代码中直接用logits乘以vi，再求取CEloss。

二者在数学上并不等价，请问是否我理解错误？该如何解释？