楼主的model训练时是image输入，但是测试时却是输入image和flipped image的连接

[ntulu]

这样会矛盾吗？以及最终采取slice后取max的步骤，能解释一下吗

[wujiyang]

@ntulu 这是测试时候使用的trick，很多时候测试的时候都是原始图像和水平翻转图像的连接，但这个取的是max，具体可以参看 happynear wangfeng 的最新论文：NormFace

[ntulu]

@wujiyang 编译时出现了好多warning，正常吗 src/caffe/layers/rank_hard_loss_layer.cpp:149:15: warning: unused variable 'bottom_data' [-Wunused-variable] src/caffe/layers/rank_hard_loss_layer.cpp:150:15: warning: unused variable 'label' [-Wunused-variable] src/caffe/layers/rank_hard_loss_layer.cpp:151:6: warning: unused variable 'count' [-Wunused-variable] src/caffe/layers/rank_hard_loss_layer.cpp:153:6: warning: unused variable 'dim' [-Wunused-variable] src/caffe/layers/rank_hard_loss_layer.cpp:157:8: warning: unused variable 'hard_ratio' [-Wunused-variable] src/caffe/layers/rank_hard_loss_layer.cpp:158:8: warning: unused variable 'rand_ratio' [-Wunused-variable]

[wujiyang]

@ntulu 正常，在编程的时候我把全部的层参数都读取了一遍，不过有些没有用到