This is the repository for the paper of "Well-classified Examples are Underestimated in Classification with Deep Neural Networks"
In this paper, we find that the cross-entropy loss hinders representation learning, energy optimization, and margin growth, and well-classified examples play a vital role to dealing with these issues. We support this finding by both theoretical analysis and empirical results.
You can find implementation and scripts (readme.sh) in the corresponding directory for each task.
Our modification is mainly around the el.py in each task.
We give the code for a conterexample (encouraging loss) below.
import torch
import torch.nn as nn
from torch.nn import functional as F
class EncouragingLoss(nn.Module):
def __init__(self, log_end=0.75, reduction='mean'):
super(EncouragingLoss, self).__init__()
self.log_end = log_end # 1 refers to the normal bonus, but 0.75 can easily work in existing optimization systems, 0.5 work for all settings we tested, recommend LE=0.75 for high accuracy scenarios and low LE for low accuracy scenarios.
self.reduction = reduction
def forward(self, input, target):
lprobs = F.log_softmax(input) # logp
probs = torch.exp(lprobs)
bonus = torch.log(torch.clamp((torch.ones_like(probs) - probs), min=1e-5)) # log(1-p)
if self.log_end != 1.0: # end of the log curve in conservative bonus
log_end = self.log_end
y_log_end = torch.log(torch.ones_like(probs) - log_end)
bonus_after_log_end = 1/(log_end - torch.ones_like(probs)) * (probs-log_end) + y_log_end
bonus = torch.where(probs > log_end, bonus_after_log_end, bonus)
loss = F.nll_loss(lprobs-bonus, target.view(-1), reduction=self.reduction)
return lossFor the label smoothed version, you can refer to label_smoothed_encouraging_loss_fairseq.py and label_smoothed_el_timm.py