how to build this

[GuitarZhang]

how to build this

[luhaofang]

Hi,
Firstly you should makesure your caffe is built for Python layers. If not open "Makefile.config" in caffe root and uncomment the line “WITH_PYTHON_LAYER := 1”.Rebuild caffe.

Go to tripletloss root path. Add tripletloss root path to train.py.

Change the configs in config.py then caffe can create the network.

Note :Training a network you should read the code first.

[GuitarZhang]

When I run train.py , an error occurred, " No module named tripletloss.datalayer"。
You said "Add tripletloss root path to train.py", I only changed pathes in main function, is it right?

[GuitarZhang]

" No module named tripletloss.datalayer"这个错误解决了，把module设置为datalayer就好了。
出现了另外一个问题，“AttributeError: 'DataLayer' object has no attribute 'param_str_'”，我查了caffe的源码也没有找到这个param_str_，这是哪里错了呢？

[luhaofang]

你的python路径没有找到，这几个datalayer都是python写的，需要添加工程的根路径到你的python中

[GuitarZhang]

不知道楼主的比对结果是多少呢？我看tripletselectlayer里面是把if an > ap:记下来了，应该是ap>an吧？

[luhaofang]

semi-hard的筛选请参见paper的内容。
个人感觉这里应该有一些trick，似乎可以通过设置区间参数来调整样本的筛选，本人也做了些调整，但目前并未发现选择 an > ap + beta 区间，对结果产生的差异性表现

[GuitarZhang]

恩，恩论文中看到了。tripletselectlayer这一层需要backward吗？总感觉不backward，loss像传不到前面似的。

[GuitarZhang]

我运行的时候所有的都显示not need backward computation，不知哪里出了问题？
tripletloss does not need backward computation.
I0504 18:03:40.812458 66083 net.cpp:228] triplet_select does not need backward computation.
I0504 18:03:40.812466 66083 net.cpp:228] norm2 does not need backward computation.
I0504 18:03:40.812474 66083 net.cpp:228] drop8 does not need backward computation.
I0504 18:03:40.812482 66083 net.cpp:228] relu8 does not need backward computation.
I0504 18:03:40.812489 66083 net.cpp:228] fc9 does not need backward computation.
I0504 18:03:40.812497 66083 net.cpp:228] drop7 does not need backward computation.
I0504 18:03:40.812505 66083 net.cpp:228] relu7 does not need backward computation.
I0504 18:03:40.812512 66083 net.cpp:228] fc7 does not need backward computation.
I0504 18:03:40.812520 66083 net.cpp:228] drop6 does not need backward computation.
I0504 18:03:40.812528 66083 net.cpp:228] relu6 does not need backward computation.
I0504 18:03:40.812536 66083 net.cpp:228] fc6 does not need backward computation.
I0504 18:03:40.812544 66083 net.cpp:228] pool5 does not need backward computation.
I0504 18:03:40.812557 66083 net.cpp:228] relu5_3 does not need backward computation.
I0504 18:03:40.812566 66083 net.cpp:228] conv5_3 does not need backward computation.
I0504 18:03:40.812573 66083 net.cpp:228] relu5_2 does not need backward computation.
I0504 18:03:40.812582 66083 net.cpp:228] conv5_2 does not need backward computation.
I0504 18:03:40.812590 66083 net.cpp:228] relu5_1 does not need backward computation.
I0504 18:03:40.812597 66083 net.cpp:228] conv5_1 does not need backward computation.
I0504 18:03:40.812605 66083 net.cpp:228] pool4 does not need backward computation.
I0504 18:03:40.812613 66083 net.cpp:228] relu4_3 does not need backward computation.
I0504 18:03:40.812621 66083 net.cpp:228] conv4_3 does not need backward computation.
I0504 18:03:40.812629 66083 net.cpp:228] relu4_2 does not need backward computation.
I0504 18:03:40.812638 66083 net.cpp:228] conv4_2 does not need backward computation.
I0504 18:03:40.812662 66083 net.cpp:228] relu4_1 does not need backward computation.
I0504 18:03:40.812672 66083 net.cpp:228] conv4_1 does not need backward computation.
I0504 18:03:40.812680 66083 net.cpp:228] pool3 does not need backward computation.
I0504 18:03:40.812688 66083 net.cpp:228] relu3_3 does not need backward computation

[louielu1027]

@GuitarZhang 您好，我也遇到了您的问题，AttributeError: 'DataLayer' object has no attribute 'param_str_；请问您如何解决这个问题的？具体过程有吗？