Skip to content

Latest commit

 

History

History
104 lines (90 loc) · 4.27 KB

tf_version.md

File metadata and controls

104 lines (90 loc) · 4.27 KB

tf version

1.10.0

  • official compile envionment
    • Cmake v3.6.3 | Bazel 0.15.0 (last official release using cmake on Windows)
    • MSVC 2015 update 3 | GCC 4.8
    • CUDA 9, cuDNN 7
  • tf 1.10.1, No new binaries will be built for Windows(???)

1.12.0

  • official compile envionment
    • Bazel 0.15.0
    • MSVC 2015 update 3 | GCC 4.8
    • CUDA 9, cuDNN 7

1.13.1

  • official compile envionment
    • Bazel 0.19.0-0.21.0 | Bazel 0.19.2
    • MSVC 2015 update 3 | GCC 4.8
    • CUDA 10, cuDNN 7.4 (1st release using CUDA 10)
    • TensorRT 5.0 (optional)

1.14.0

  • 1st release contraining compat.v2 module
  • turn on MKL-DNN by default
  • Non-Windows system libraries are now versioned. This should be a no-op for most users as it affects only package maintainers or those building extensions to TF
    • Pip packages contain one library file(Linux: libtensorflow_framework.so.1)
    • 译, Windows pip 包中不含有 tf 相关 dll(Linux pip 包中含有 libtensorflow_framework.so.1), 这意味着在 Windows 下无法通过链接 lib 方式添加 custom op
  • official compile envionment
    • Bazel 0.24.1-0.25.2 | Bazel 0.24.1
    • MSVC 2017 | GCC 4.8
    • CUDA 10, cuDNN 7.4
  • build cxx lib for Windows

1.15.0

  • last 1.x release for TF
  • add fftshift op
  • auto Mixed-Precision graph optimizer simplifies converting models to float16.
    • enabled by wrapping an optimizer class with tf.train.experimental.enable_mixed_precision_graph_rewrite()
  • should be last release has contrib directory in release code
  • some TensorRT support
  • NO official compile envionment on official web

2.0.0

  • eager, keras, ..., and all features in tf 1.x
  • removed freeze_graph command line tool; SavedModel should be used in place of frozen graphs
  • still has contrib directory in release code, but not in branch r2.1 or master
    • maybe tf2 can be built with cmake
    • tf.contrib moved to tensorflow/addons or tensorflow/io, or removed
  • official compile envionment
    • Bazel 0.26.1
    • MSVC 2017 | GCC 7.3.1
    • CUDA 10, cuDNN 7.4

2.1.0

  • pip package is built with CUDA 10.1 and cuDNN 7.6
  • Windows pip package is built with VS2019
  • experimental support for mixed precision
  • many keras implemented models
  • tf.debugging.enable_check_numerics() to help debugging the root causes of issues involving Inf and NaN
  • TensorRT 6.0 is supported and enabled by default
  • from v2.1.0, contrib folder was removed from src

2.2.0

  • replace std::string with tensorflow::tstring, now ABI is stable
  • add a new Profiler of TF2, for CPU/GPU/TPU
  • export C++ using pybind11, SWIG is deprecated
  • add SyncBatchNormalization layer for multi-GPUs
  • custom training logic by overriding Model.train_step
  • XLA now builds and works on Windows

To be built on Windows

  • tf 1.10.0, CUDA 9

    • 官方支持的最后一个cmake版本, 可代替目前针对 MR 发布的 tf lib

  • tf 1.12.0, CUDA 9

    • 尝试cmake编译, 如成功，可成为支持 CUDA 9 的最后一个版本

  • tf 1.13.1, CUDA 10

    • 尝试 cmake 编译, 如成功, 可成为支持 CUDA 10 的第一个版本
    • 尝试 bazel 编译, 已有成功案例, 但有较多的不确定因素

  • tf 1.14.0, CUDA 10

    • 尝试 cmake 编译, 如成功, 可成为第一个支持 vs2017 版本
    • 尝试 bazel 编译, 尚无可参考的案例

  • tf 1.15.0, CUDA 10

    • 尝试 cmake 编译, 如成功, 可成为最后一个 tf1.x 版本
    • 尝试 bazel 编译, 尚无可参考的案例

  • tf 2.0.0, CUDA 10

    • 尝试 cmake 编译, 如成功, 可成为第一个 tf2 版本
    • 尝试 bazel 编译, 尚无可参考的案例

  • tf 2.1.0, CUDA 10.1

    • 尝试 bazel 编译, 尚无可参考的案例

  • 在无需 custom op 的情况下, 使用 cmake 构建或成为首选

To be done

  • tf2 overview, 看一遍官网guide
  • SavedModel
  • 看fftshift具体实现
  • use clion bazel to view tf code