jxgu1016/MNIST_center_loss_pytorch

question about center loss item

Closed this issue · 3 comments

Excuse me, the formula says that Xi is a feature dimension that changes with depth. How does the code reflect it? I can't see that. Is feat_dim set as the input dimension, and can it ensure that the optimized output dimension is the final one? Please answer my question, thank you.

Sorry, cannot understand your questions.

不好意思,我表达的不清楚,我用中文说一下问题。center loss公式里的Xi是随深度变化的特征,代码中的feat_dim和feat指的是输入时的特征维度吗?怎么保证随深度变化我没在代码中看出来,也可能是我理解的不对。而且loss优化的不应该是输出的特征吗?我的问题和自己的理解不到位有所相关,请指教一下。

论文里面只在fc上加了center loss,维度需要手工设置。