在 Arm MBP 上的 macOS 上安装 Tensorflow

Question

当尝试在 ARM M1 MacBookPro 上安装 TensorFlow 时，我遇到以下问题：

% pip install tensorflow
ERROR: Could not find a version that satisfies the requirement tensorflow (from versions: none)
ERROR: No matching distribution found for tensorflow

经过搜索后，我发现一个名为tensorflow-macos的包会产生以下问题：

clang: error: the clang compiler does not support 'faltivec', please use -maltivec and include altivec.h explicitly
    error: Command "gcc -Wno-unused-result -Wsign-compare -Wunreachable-code -fno-common -dynamic -DNDEBUG -g -fwrapv -O3 -Wall -arch arm64 -arch x86_64 -g -DNPY_INTERNAL_BUILD=1 -DHAVE_NPY_CONFIG_H=1 -D_FILE_OFFSET_BITS=64 -D_LARGEFILE_SOURCE=1 -D_LARGEFILE64_SOURCE=1 -DNO_ATLAS_INFO=3 -DHAVE_CBLAS -Ibuild/src.macosx-10.9-universal2-3.9/numpy/core/src/umath -Ibuild/src.macosx-10.9-universal2-3.9/numpy/core/src/npymath -Ibuild/src.macosx-10.9-universal2-3.9/numpy/core/src/common -Inumpy/core/include -Ibuild/src.macosx-10.9-universal2-3.9/numpy/core/include/numpy -Inumpy/core/src/common -Inumpy/core/src -Inumpy/core -Inumpy/core/src/npymath -Inumpy/core/src/multiarray -Inumpy/core/src/umath -Inumpy/core/src/npysort -I/Users/thing/dev/SOTAMoon/venv/include -I/Library/Frameworks/Python.framework/Versions/3.9/include/python3.9 -Ibuild/src.macosx-10.9-universal2-3.9/numpy/core/src/common -Ibuild/src.macosx-10.9-universal2-3.9/numpy/core/src/npymath -c numpy/core/src/multiarray/descriptor.c -o build/temp.macosx-10.9-universal2-3.9/numpy/core/src/multiarray/descriptor.o -MMD -MF build/temp.macosx-10.9-universal2-3.9/numpy/core/src/multiarray/descriptor.o.d -faltivec -I/System/Library/Frameworks/vecLib.framework/Headers -std=c99" failed with exit status 1
    ----------------------------------------
    ERROR: Failed building wheel for numpy
  Failed to build numpy
  ERROR: Could not build wheels for numpy which use PEP 517 and cannot be installed directly
  ----------------------------------------
WARNING: Discarding https://files.pythonhosted.org/packages/a7/81/20d5d994c91ed8347efda90d32c396ea28254fd8eb9e071e28ee5700ffd5/h5py-3.1.0.tar.gz#sha256=1e2516f190652beedcb8c7acfa1c6fa92d99b42331cbef5e5c7ec2d65b0fc3c2 (from https://pypi.org/simple/h5py/) (requires-python:>=3.6). Command errored out with exit status 1: /Users/thing/dev/SOTAMoon/venv/bin/python3 /private/var/folders/8k/z291bhgd5gs06tp9b4j6_bb40000gn/T/pip-standalone-pip-8qa87uy0/__env_pip__.zip/pip install --ignore-installed --no-user --prefix /private/var/folders/8k/z291bhgd5gs06tp9b4j6_bb40000gn/T/pip-build-env-bxy_nyoo/normal --no-warn-script-location --no-binary :none: --only-binary :none: -i https://pypi.org/simple -- 'numpy==1.12; python_version == "3.6"' 'numpy==1.19.3; python_version >= "3.9"' 'numpy==1.14.5; python_version == "3.7"' pkgconfig 'Cython>=0.29; python_version < "3.8"' 'Cython>=0.29.14; python_version >= "3.8"' 'numpy==1.17.5; python_version == "3.8"' Check the logs for full command output.
ERROR: Could not find a version that satisfies the requirement h5py~=3.1.0 (from tensorflow-macos) (from versions: 2.2.1, 2.3.0b1, 2.3.0, 2.3.1, 2.4.0b1, 2.4.0, 2.5.0, 2.6.0, 2.7.0rc2, 2.7.0, 2.7.1, 2.8.0rc1, 2.8.0, 2.9.0rc1, 2.9.0, 2.10.0, 3.0.0rc1, 3.0.0, 3.1.0, 3.2.0, 3.2.1, 3.3.0)
ERROR: No matching distribution found for h5py~=3.1.0

不幸的是，我无法找到“faltivec”问题的解决方案。我还发现了Apple 制作的一个 Tensorflow 包，似乎专门针对 M1，但是按照本教程会产生以下问题：

% pip install --upgrade --force --no-dependencies https://github.com/apple/tensorflow_macos/releases/download/v0.1alpha3/tensorflow_addons_macos-0.1a3-cp38-cp38-macosx_11_0_arm64.whl https://github.com/apple/tensorflow_macos/releases/download/v0.1alpha3/tensorflow_macos-0.1a3-cp38-cp38-macosx_11_0_arm64.whl
ERROR: tensorflow_addons_macos-0.1a3-cp38-cp38-macosx_11_0_arm64.whl is not a supported wheel on this platform.

在此之后，我有点卡住了。我已经将 pip 升级到21.1.3，我的 python 版本是Python 3.9.6.

Answer 1

Tensorflow-macos 和 Tensorflow-metal 安装

目前，要利用 M1 GPU，您需要安装Tensorflow-macos，与TensorFlow-metal此相反Tensorflow，安装步骤详细介绍如下，使用 mini-forge 可以将其总结如下：

conda create -n tf python=3.8 -y
conda activate tf
conda install -c apple tensorflow-deps -y
# Navigate the issue with conda environments
# built against pre-macOS 11 SDK use (Just in case)
# this solves the Intel / M1 install issues:
SYSTEM_VERSION_COMPAT=0 pip install tensorflow-macos 
SYSTEM_VERSION_COMPAT=0 pip install tensorflow-metal

为了测试一切是否正常，我们可以运行：

#!/usr/bin/env bash

PYCMD=$(cat <<EOF
import tensorflow as tf

tf.config.list_physical_devices()
with tf.device('/GPU'):
  a = tf.random.normal(shape=(2,), dtype=tf.float32)
  b = tf.nn.relu(a)

print(a)
print(b)

EOF
)

python3 -c "$PYCMD" 2>/dev/null

我的测试机的预期输出：

>./val.sh
Metal device set to: Apple M1

systemMemory: 16.00 GB
maxCacheSize: 5.33 GB

tf.Tensor([-0.4044231   0.08157467], shape=(2,), dtype=float32)
tf.Tensor([0.         0.08157467], shape=(2,), dtype=float32)

好吧，一切都（假设）正常工作......

旁注：软件包安装问题：

需要注意的一件事是，依赖的软件包tensorflow将尝试安装tensorflow，因为它们可能不会显式依赖于tensorflow-macos或tensorflow-metal。

当您安装tensorflow依赖包或代码时，您可能需要requirements.txt file在运行之前检查：

python3 pip install -r requirements.txt

它将requirements.txt尝试安装该tensorflow软件包，而不是安装tensorflow-macosx或tensorflow-metal软件包，如果您不加更改地运行该文件，则会导致许多问题requirements.txt。该包和显式版本的包==也可能导致问题。

希望这能为您指明正确的方向，帮助您思考使用 TensorFlow Metal 插件利用 Python 与 Apple-M1 笔记本电脑配合使用 M1 GPU 功能所遇到的障碍。

保持安全和健康。