更新 10月25日本周科研工作进展.md
parent
df4db30c41
commit
ef07d9412d
|
@ -39,6 +39,6 @@ loss_proto /= len(args.num_cluster)
|
|||
|
||||
### 解决办法与下周安排:
|
||||
|
||||
1.将实验视为loss相互竞争的多任务学习,考虑采用梯度标准化多目标优化(GradNorm: Gradient Normalization for Adaptive Loss Balancing in Deep Multitask Networks)或多目标优化(Multi-task learning as multi-objective optimization)的方式修改代码。
|
||||
1.将实验视为loss相互竞争的多任务学习,考虑采用梯度标准化(GradNorm: Gradient Normalization for Adaptive Loss Balancing in Deep Multitask Networks)或多目标优化(Multi-task learning as multi-objective optimization)的方式修改代码。
|
||||
|
||||
2.增加下游任务实验,观察不同结果的模型在图像分类/目标检测/聚类评估的任务中的发挥。
|
Loading…
Reference in New Issue