# 特徵擷取網路架構


先固定encoder，train discriminator和classsifer，然後再train encoder，藉由min||v1-v2||，再back回encoder更新參數

資料夾：先放人，再放速度

-------------------------------------------


* 先train discriminator和action classifier，後再train encoder和classifier
* encoder出來的會是vector，上半
*
-------------------------------------------
categorical_crossentropy + mse[1,0.05]


categorical_crossentropy + kullback_leibler_divergence[1,0.05]

categorical_crossentropy + mse[1,0.003]


origin+offset latent_dim=200