how do you choose your cuhk03 dataset ?

[Phoebe-star]

how about your gallery and probe ?
and training data

[zlmzju]

The official CUHK03 dataset provides the splits for training and test. I follow the paper of CUHK03 to use their protocol (choosing the gallery and probe for each identity randomly).

[Phoebe-star]

the code can convert to tensorflow ?

and when training , your triplet loss is always small ?
like : Step: 67, Learning rate: 0.009951, Train loss: 0.031987
Step: 3300, Learning rate: 0.007951, Train loss: 0.022587
Step: 5200, Learning rate: 0.006951, Train loss: 0.030987

[Phoebe-star]

how do you do resize images?
you directly resize 224x224 ? not do crop? or others?
and how about your Data Augmentation

[jianwu585218]

I'm trying to convert to tensorflow,but get some troubles.
Did you successfully convert to tensorflow ?

[Phoebe-star]

how about your troubles?

[jianwu585218]

In the Part Map Detector :

I think the mk(x,y) is a 1*1 local convolutional layer which not shared weight , isn't it ?
but i don't know how to make it .
I need your help, thanks very much.

[jianwu585218]

Hi,when i'm training my model by tensorflow, I found the triplet loss is drop hard.I want to know have you success reappearance ? Thank you very much~

[Phoebe-star]

yes, you are right , mk(x,y) is a 1*1 local convolutional layer which not shared weight
have you been used hard example? if you used, the training will be hard.

[jianwu585218]

I have not use the hard example ,I used the random triplet ,and the loss can't be decline

[Phoebe-star]

Do you have use the googlenet pretrain from imagenet? how about your loss?

[jianwu585218]

Thanks for your reply!
I use the vgg16 pretrain model from imagenet , my loss is original triplet loss ,the triplet pair is random generated , such as :

Do you use the hard triplet ?

[Phoebe-star]

I think your vgg have some problem? like the whights don't change . why don't you use the googlenet?
if use original triplet loss ,the loss will be easy to train ,
but the hard triplet , I guess the loss will be difficult to train

[jianwu585218]

Thanks for your reply again~
Because i am used to use the vgg , i have not used googlenet before , and i don't know how to use the local conv layer by tensorflow .
I'm a new hand of tensorflow , I have tried this project for nearly a month.
And can you send your demo to learning ? Although I know this is an impolite request , you can refuse me ~
Thank you very much~

[Phoebe-star]

I also use someone github , but can I point it here?
after all, here is zlmzju 's github

[jianwu585218]

Can you send the github to my email?
This is my email :[email protected].
Thank you very much ,you and zlm is my star😊