/center-loss

Primary LanguagePython

1. raw start

时间关系训了三类0、1、2
ep27左右，acc达到0.9左右，可视化散点图初步可分，但是分类边界不明显
ep30左右，可以观察到散点在聚拢

2. train strategy

后来发现应该先训no_center_loss的模型
然后用center_loss模型fine-tune

3. cls center update

1. using the GT label to embed a (1,2) vector standing for the center
    the center is updated along with the changement of the embedding layers' weights
    the entire layer is a black box to us

2. using the batch center and a learning rate alpha to update the center
    the center is computed through the call function defined in the custom layer
    we know what actually happened through the pipeline

4. learning rate

学习率的选择很重要，大了不收敛，小了收敛贼慢

5. results

raw: 数据可分，但是类内差异较大，因此数据整体分布呈现长椭圆形，raw_ep9
centerloss：相比之下，数据更聚集，custom_ep3

6. todolist

6.1 triplet-center loss: 普通分类只是在找类间边界而没有考虑类内变化，而center loss减小了类内变化，但是考虑类间距离，容易造成类间重叠。

​