本文介绍一下tensorflow(以下称TF)项目的主要文件目录结构。信息来自tensorflow的google group讨论。希望对想优化TF，或者想搞清楚TF底层细节的朋友有所帮助 _。
首先看一下0.12版本的文件目录截图
根据讨论组提供的信息，其项目文件的实现功能主要为以下内容
core/ contains the main C++ code and runtimes.
core/ops/ contains the "signatures" of the operations
core/kernels/ contains the "implementations" of the operations (including CPU and CUDA kernels)
core/framework/ contains the main abstract graph computation and other useful libraries
core/platform/ contains code that abstracts away the platform and other imported libraries (protobuf, etc)
TensorFlow relies heavily on the Eigen library for both CPU and GPU calculations. Though some GPU kernels are implemented directly with CUDA code.
bazel builds certain C++ code using gcc/clang, and certain CUDA code (files with extension .cu.cc) with nvcc.
python/ops/ contain the core python interface
python/kernel_tests/ contain the unit tests and lots of example code
python/framework/ contains the python abstractions of graph, etc, a lot of which get serialized down to proto and/or get passed to swigged session calls.
python/platform/ is similar to the C++ platform, adding lightweight wrappers for python I/O, unit testing, etc.
contrib/*/ directories generally mimic the root tensorflow path (i.e., they have core/ops/, etc)
翻译过来就是
后台实现部分
这部分的代码都在core/目录下
core/ 目录里面是C++代码和运行时，所以是核心代码。
core/ops/ 目录下是各种operation的签名'signature'。
core/kernels/ 目录下是各种operation的实现(包括CPU kernel和CUDA kernel)。
core/framework/ 目录下是计算图的抽象和一些其他有用的库。
core/platform/ 目录下对其他平台和导入库(protobuf等)的抽象代码。
TF的CPU和GPU计算严重依赖于Eigen库，尽管一些GPU kernel直接使用CUDA代码实现的。
bazel编译时，一部分C++代码用的是gcc或clang，一部分CUDA代码(后缀名是.cu.cc)用的是nvcc。
前台Python接口部分
python/ops/ 核心python接口
python/kernel_tests/ 单元测试代码和示例代码
python/framework/ python抽象图, a lot of which get serialized down to proto and/or get passed to swigged session calls.
python/platform/ 和上面C++部分的platform差不多, 对python I/O、单元测试等做了轻量级的包装。
