kaleidozhouyn / angular-triplet-loss Goto Github PK

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.

大佬好，最近在尝试用这个loss进行训练，关于这个loss添加的在prototxt中内容不是太明确，请问我像下面这样可以么？
layer {
name: "fc6"
type: "MarginInnerProduct"
bottom: "fc5"
bottom: "label"
top: "fc6"
top: "lambda"
param: {
lr_mult: 1
decay_mult: 1
}
margin_inner_product_param {
num_output: 70116
type: SINGLE
weight_filler {
type: "xavier"
}
base: 0
gamma: 0.12
power: 1
lambda_min: 3
iteration: 0
triplet: true
semihard: true
}
include {
phase: TRAIN
}
}
不知道有没有什么参数需要修改，训练集是清理过的Ms_celeb_1M，batch size=64这样子，跪求大佬解答，谢谢啦～
还有读取数据的时候需不需要预处理，减均值归一化之类的，我的是下面这样，再一次跪求指导～
layer {
name: "data"
type: "ImageData"
top: "data"
top: "label"
include {
phase: TRAIN
}
transform_param {
mean_value: 104
mean_value: 117
mean_value: 123
scale: 0.00390625
mirror: true
}
image_data_param {
new_width:112
new_height:112
source: "/home/lmd/data/ms_celeb_1m_cropped_alignment/ms_train_list.txt"
root_folder : "/home/lmd/data/ms_celeb_1m_cropped_alignment/ms_celeb_1m_face_image_data/"
batch_size: 64
shuffle: true
}
}

Did you use Triplet with ASoftmax? I think it worth to test it.

Why do think we should do the training using pre-trained model not a new training from scratch?!
"method 1:step
step 1:
train a pre-trained model use sphereface"

margin_inner_product_param{
num_output:70609
type: SINGLE
weight_filler{
type: "xavier"
}
base: 0
gamma: 0.12
power: 1
lambda_min: 3
iteration: 0
triplet: true
semihard: true
}

809-812行少了一个 optional int32 axis = 2 [default = 1]; // axis of prob
message InfogainLossParameter {

// Specify the infogain matrix source.

optional string source = 1;

}

I think it should be softmax loss layer, beacuse for triplet layer the dim of fc6 is quite large, but still not sure, it is not mentioned in the project.

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]);

Hi, I wonder how different align methods can affect the results in your experiments, thanks

I have train a batch triplet loss on my own dataset about retail goods,but I do not have a good result,so I want to know how to train the A-T-loss,specifically,for the dataset,whether need split group for triplet selection??the loss layer just choose triplet loss?
so you mainly change the embedding layer('fc6'?)and other setting like original triplet loss?
I hope you ca provide an example for me,if you are uncomfortable to publish,please contact me by [email protected] to give me some instructions.
thanks a lot!