Build tensorflow on Mac

Posted by chunyang on June 20, 2020

Building tensorflow from source on mac is really not an easy thing at first. This blog introduces problems met during building.

Bazel installation

On mac, you can install lots of command line tools using Homebrew.

brew install bazel

It is a pity that I cannot specify the version like brew install [email protected].

Why I want to install version 0.21.0? Please refer to Github issue. Different versions of tensorflow have different requirements. Currently I am building r1.13. It requires bazel with a version lower than 0.21.0.

Go to Github bazel release page, find the version you need and download it. BTW, (releases/tag/${version}) will directly go to the corresponding page.

Build command

Configure

./configure

Select functions you care and you are ready to go.

Build

Install the python packages needed by tensorflow. It is better that we use virtualenv or anaconda to manage the environment.

pip install numpy scipy matplotlib ipython jupyter pandas sympy nose
bazel build -c opt //tensorflow/tools/pip_package:build_pip_package
bazel-bin/tensorflow/tools/pip_package/build_pip_package /tmp/tensorflow_pkg

The first command sometimes fails with no reasons, you just keep retry it (Oops!).

The second command will generate the wheel file you need to install.

@org_python_license

The checksum is not correct. Just modify the file in tensorflow/tensorflow/workspace.bzl and change the sha1 value in sha256_urls.

Reference in Chinese

Generate package

google/protobuf/port_def.inc not found. Github issue

It seems that tensorflow misses files. Way to fix Commit

Run tensorflow

A annoying warning:

>>> import tensorflow
/Users/chunyangwen/Documents/envs/python3.7.7/lib/python3.7/site-packages/tensorflow/python/framework/dtypes.py:526: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) / '(1,)type'.
  _np_qint8 = np.dtype([("qint8", np.int8, 1)])
/Users/chunyangwen/Documents/envs/python3.7.7/lib/python3.7/site-packages/tensorflow/python/framework/dtypes.py:527: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) / '(1,)type'.
  _np_quint8 = np.dtype([("quint8", np.uint8, 1)])
/Users/chunyangwen/Documents/envs/python3.7.7/lib/python3.7/site-packages/tensorflow/python/framework/dtypes.py:528: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) / '(1,)type'.
  _np_qint16 = np.dtype([("qint16", np.int16, 1)])
/Users/chunyangwen/Documents/envs/python3.7.7/lib/python3.7/site-packages/tensorflow/python/framework/dtypes.py:529: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) / '(1,)type'.
  _np_quint16 = np.dtype([("quint16", np.uint16, 1)])
/Users/chunyangwen/Documents/envs/python3.7.7/lib/python3.7/site-packages/tensorflow/python/framework/dtypes.py:530: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) / '(1,)type'.
  _np_qint32 = np.dtype([("qint32", np.int32, 1)])
/Users/chunyangwen/Documents/envs/python3.7.7/lib/python3.7/site-packages/tensorflow/python/framework/dtypes.py:535: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) / '(1,)type'.
  np_resource = np.dtype([("resource", np.ubyte, 1)])

Please change the file tensorflow/python/framework/dtypes.py. Replace the number from 526 to 535 with a pair of parentheses and a comma.

# before
_np_qint8 = np.dtype([("qint8", np.int8, 1)])
# after
_np_qint8 = np.dtype([("qint8", np.int8, (1,))])

Have fun!