Ensemble model and Triplet about angular-triplet-loss HOT 6 OPEN

@MartinDione
ensemble means train with different dataset and different network and concat feature togeter.
for triplet there is no need to choose 3 image,i implement a method called 'easy triplet' which you can look up for details in "MarginInnerProduct.cpp/.cu"

from angular-triplet-loss.

Thank you for your answer. Can you tell me about your ensemble network what did you use for the network and dataset?

from angular-triplet-loss.

usually i use resface28 network but with differenct dataset like ms_celeb and vggface2.

from angular-triplet-loss.

@KaleidoZhouYN 非常感谢作者，我这边使用您的代码之后，模型性能的确提升了，但是理解代码的时候还是有想不明白的地方。比如418行到420行，420行我知道是直接把传进来的样本当成是新的类中心来做分类，但是前两行就不明白有什么作用了， ran_flag 的范围应该是 0至 9 的整数，(ran_flag > 4)应该是一半的概率，并没有体现 trinlet的**。 有点不太明白
int ran_flag = caffe_rng_rand() % int(10);
if (ran_flag > 4)
caffe_copy(K_,(Dtype*)&bottom_data[i],(Dtype*)&m_weight[label_i]);

from angular-triplet-loss.

Can you please tell me about how to train Ensemble models(please provide me with details about this)?

And how to choose the images for triplet as we need for 3 images is that right in your method?

Thanks.

do you understand the easy triplet , I`ve read code, but still didnt get it

from angular-triplet-loss.

@ly0303521 我也没看明白。
首先他 cpp 代码里只在backward里写了triplet和semi-hard相关的内容，这个倒是可以训练。

问题是他cpp代码里面L405之后，关于semi-hard 和 triplet 相关的代码，并没有对back-prop的梯度 bottom_diff 以及参数的导数 m_weight_diff 做任何赋值，除了置0。这一点好奇怪。
而且作者好像基本不response了。。。

from angular-triplet-loss.