tensorflow windows 创建自己的插件 [英] tensorflow windows create own plugin
问题描述
我从源代码中使用 Visual Studio 2015 在 Windows 10 上成功构建了 tensorflow+gpu.
I have tensorflow+gpu successfully built on windows 10 with visual studio 2015, from the source code.
结果，我得到了 tensorflow.dll 和 tensorflow.lib.我有 CUDA8.0 和 cudnn 5.0；配备 gtx 1080 gpu.
As a result, I get tensorflow.dll and tensorflow.lib.
I have CUDA8.0 and cudnn 5.0; with a gtx 1080 gpu equipped.
然而，我的问题不是关于构建和编译 tensorflow.这是关于创建 tensorflow 插件.
However, my question is not about building and compiling tensorflow. It's about creating tensorflow plugins.
我按照教程构建了我自己的插件".然后我尝试编译一个windows .dll；所以windows不会自动为我导出符号.然后我先编译一个静态库并使用你的工具
I followed the tutorial to construct my own "plug-in". and then I tried to compile a windows .dll; so windows would not export symbols automatically for me . then I compile a static lib first and used your tools
/tensorflow/contrib/cmake/tools/create_def_file.py
为我创建一个 .def 文件，并最终用它来编译 .dll.
to create a .def file for me and eventually used that to compile the .dll.
但是，在我的python代码中，当我尝试
However, in my python code, when I tried to
correlation = tf.load_op_library(correlation.dll)
我打电话给
correlation.correlation()
使用REGISTER_OP("Correlation")注册相关性；它仍然告诉我
with Correlation registered using REGISTER_OP("Correlation");
It still tells me
AttributeError: module '7b088d8b906b36d3e50721b0adbaaa6a' 没有属性 'correlation'
AttributeError: module '7b088d8b906b36d3e50721b0adbaaa6a' has no attribute 'correlation'
我认为这只是 windows(或 cl 编译器)的问题，也许 REGISTER_OP("Correlation") 所做的只是编译器没有发现，
I think this is just a windows (or cl compiler) issue, maybe what REGISTER_OP("Correlation") did is just not picked up by the compiler,
那么我可以做些什么来使这在 Windows 上发生??
so is there any thing I can do to make this happen on windows??
推荐答案
Windows 不支持通过 tf.load_op_library() 加载自定义操作库(至少在 TensorFlow 1.8 中).解决方法是将您的自定义操作添加到 TensorFlow 库本身中.以tensorflow\tensorflow\core\user_ops\fact.cc中实现的tf.user_ops.my_fact为例:
Loading custom op libraries via tf.load_op_library() is not supported on Windows (at least with TensorFlow 1.8). The workaround is to add your custom op into the TensorFlow library itself. Follow the example of tf.user_ops.my_fact implemented in tensorflow\tensorflow\core\user_ops\fact.cc:
- 将您的 C++ 实现放在 tensorflow\tensorflow\core\user_ops 中
- 在 tensorflow\tensorflow\python\user_ops 中添加 python 绑定
- 编译 TensorFlow(阅读https://github.com/tensorflow/tensorflow/tree/master/tensorflow/contrib/cmake)
- 将 Conda 环境中的 tensorflow 目录替换为张量流\张量流\contrib\cmake\build\tf_python\tensorflow
- 您的新 op 函数将导入 tf.user_ops
这篇关于tensorflow windows 创建自己的插件的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！
