triplet loss about alignedreid-re-production-pytorch

[Gavin666Github]

Combining softmax loss with metric learning loss to speed up the convergence is also a popular method.

[huanghoujing]

open-reid，batch size 256可以跑到rank-1 84.5%，参考这里. 印象中应该是在open-reid划分的Market1501上跑的实验，而不是在标准划分上跑的。

虽然我用Python 2跑的，你可以换成Python 3试一下，如果还不行那有可能跑出来就是那样吧。

也可以参考一下我这个工程里的一些设定，和open-reid稍有不同。进一步我做了一个小trick，可以跑到89%。

分类loss和triplet loss结合我还是没解决，你什么时候解决了跟我说一下，3q : )

[Niculuse]

我试过2块显卡跑256batch size 还是低了几点，跑不到那个水平。
谢谢你了！

[Phoebe-star]

hi
how about your triplet loss?

above is my loss result , is it right ? the value is small?
Train loss is triplet loss

Step: 79, Learning rate: 0.009942, Train loss: 0.157365
('pos_loss : ', 0.80680853)
('neg_loss : ', 1.0257512)

Step: 5409, Learning rate: 0.007231, Train loss: 0.057204
('pos_loss : ', 1.3745451)
('neg_loss : ', 1.9384971)

[liliangqi]

@huanghoujing 您这里的trick具体是指什么呢

[huanghoujing]

@liliangqi 这里的trick就是把残差网络conv5的stride从2改为1，这样改过之后average pooling之前特征的分辨率就加倍了. 这个增加分辨率的做法是从文章Beyond Part Models: Person Retrieval with Refined Part Pooling看到的。

[huanghoujing]

@Phoebe-star I think your loss value 0.057204 is not small enough yet. You can also calculate how many triplets satisfy the margin inequality distance(anchor, positive) + margin < distance(anchor, negative), which I think is a better indicator.

[Niculuse]

@huanghoujing 大神有没有看过Harmonious Attention Network for Person Re-identification 这篇文章，如果有的话，我想请教些问题。

[Phoebe-star]

Step: 5409, Learning rate: 0.007231, Local Train loss: 0.29
('local pos_loss : ', 12.745451)
('local neg_loss : ', 11.384971)

I find the local feature doesn't effect .
the local feature use the

S (7 ,7) is the final distance , but its value is big , like 12.46782 or 11.14258
I think your local feature ,it is doing 7x7 =49 times, because you use the " for loop" (python) ,

but I watch the image in the paper , "AlignedReID: Surpassing Human-Level Performance in Person Re-Identificat"

it seem like do 7+6+5+4+3+2+1=28 times

answer can English or 中文 ,thanks

[huanghoujing]

@Phoebe-star The paper explains how to calculate the shortest distance well, and I just follow the paper. So any other explanation I would like to give here is just a duplicate. Maybe I wouldn't be able to get things right for you. Sorry for that.

[huanghoujing]

@Niculuse 不要称呼大神（逃
我觉得你直接问论文作者比较好，由于论文篇幅有限，作者肯定也有很多细节没机会详细解释，所以正是可以请教的地方。问我的话，我只能乱猜。

[Niculuse]

@huanghoujing 好的，感谢！

[Phoebe-star]

ok, thank you ,huanghoujing
but my local loss is right? It is a big value 12.745451 , the training step is about 5409 or 10000
local pos_loss : ', 12.745451)

sorry,because I can not test your network, I can not see your local value.

[huanghoujing]

If this value 12.745451 is loss, then it's large.

[Phoebe-star]

how about your local loss?

[huanghoujing]

Similar to global loss, local loss can also decreases to near 0, e.g. 0.0003.

[Phoebe-star]

I know the local loss will be decreases to near 0. how about the positive local loss and the negative local loss?
where is the local loss compute method ? I find you design two method . one is numpy ,another is torch

[Phoebe-star]

#2
hi ,you say " distance of two unit-length vectors falls in range [0, 2], "
why not in the [0,1]? because normalization

and why after Equation (1), the distance is in range [0, 0.76]?

thanks you , god

[JackeyWang777]

您好，请问一下分类loss和triplet loss结合会遇到什么问题呢？

[huanghoujing]

两个loss结合理论上会更好。但是我在Triplet Loss的performance已经很高的设定下再加分类loss，结果并没有提升，我弄不好这个。