这时候思路就很明显了,要想“软化”这个 loss,就得“软化”θ(x),而软化它就再容易不过,它就是 sigmoid 函数。我们有: 所以很显然,我们将θ(x)替换为σ(Kx)即可: 这就是我昨晚思考得到的 loss 了,显然实现上也是很容易的。 现在跟 Focal Loss Focal Loss Kaiming 大神的 Focal Loss 如果落实到ŷ =...
03/18 06:15:39 - mmengine - INFO - Epoch(train) [1][50/70] base_lr: 2.0000e-04 lr: 9.8000e-06 eta: 1:58:38 time: 1.2826 data_time: 0.0524 memory: 11573 grad_norm: nan loss: 331.5620 loss_cls: 331.5620 loss_bbox: 0.0000 loss_dfl: 0.0000 具体配置文件为 Mar 18, 2024 Author...
我还想问一个问题,我现在转coco_format的json用的类别和我输入的metainfo参数是保持一致的,都是类别对应的中文首字母简写,但是我的class_text_path对每个类别给了更详细的语义描述,模型在正常收敛,loss也正常,class_text_path和coco_format的json不一致会影响开放词检测的能力吗? Unicorn123455678commentedMay 23, 20...
P-TileHostVM 0(Key Client)VM 1(Key Server)MCDMAChannel Decoder (CH0-PF 0, CH1-PF 1)Host InterfacePIOAVSTAXI-STAXI-AVSTAVSTUncontrolledPortsMACSec0CSRCSRCSRE-TileTXRXMACQSFPLoopback*EAPOLPacketGenerator& CheckerCSRPacketGenerator& Checker[vf_active, clog2 (PF_NUM),clog2 (VF_NUM), PIO ...
Regarding the classification loss value you're seeing (5.555e-05): This is not necessarily an issue. The value of the classification loss largely depends on your dataset, model, and training progress. If the model is learning well, it's possible that your classification loss becomes quite smal...