Created
September 12, 2021 17:22
-
-
Save discort/1036c9d5dcf242066480ab613c53c7c4 to your computer and use it in GitHub Desktop.
Online hard example mining
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
import torch | |
from torch import Tensor | |
import torch.nn as nn | |
class OHEM(nn.Module): | |
""" | |
Online hard example mining. | |
Details: <https://arxiv.org/pdf/1604.03540.pdf> | |
""" | |
def __init__(self, | |
loss_fn: nn.Module, | |
ratio: float = 0.7, | |
reduction='mean'): | |
super(OHEM, self).__init__() | |
self.ratio = ratio | |
self.loss_fn = loss_fn | |
self.loss_fn.reduction = 'none' | |
def forward(self, | |
pred: Tensor, | |
target: Tensor, | |
dim: int = 1) -> Tensor: | |
loss = self.loss_fn(pred, target) | |
# if self.ratio == 1 or dim is None: | |
# return torch.mean(loss) | |
_, idxs = torch.topk(loss[:, dim], int(self.ratio * loss.size(0))) | |
return torch.mean(loss[idxs]) |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment