Mac使用tensorflow GPU

主要ref：苹果开发者官网 - tensorflow配置
重点是注意tensorflow-metal、tensorflow和python的版本!

tensorflow-metal的pypi仓库

tensorflow	tensorflow-metal	MacOs	features
v2.5	v0.1.2	12.0+	Pluggable device
v2.6	v0.2.0	12.0+	Variable seq. length RNN
v2.7	v0.3.0	12.0+	Custom op support
v2.8	v0.4.0	12.0+	RNN perf. improvements
v2.9	v0.5.0	12.1+	Distributed training
v2.10	v0.6.0	12.1+
v2.11	v0.7.0	12.1+
v2.12	v0.8.0	12.1+
v2.13	v1.0.0	12.1+	FP16 and BF16 support
v2.14	v1.1.0	12.1+

tensorflow和python对应版本(后面测试我高版本python3.11也可以)

1. 安装conda(省略，也可以不使用conda)
2. conda查看可以安装的python版本conda search python
   • 如果没有想要的太低版本，需要添加conda源啥的；
   • 我Macbook M3 pro测试了python3.11 + tf 2.14 + tf-metal v.1.1.0是可以运行的
3. conda创造指定python版本的环境：conda create -n tf-metal python=3.11.8
4. conda选择刚创建的环境：conda activate tf-metal
5. pip下载tf：pip install tensorflow==2.14
6. pip下载tf-metal：pip install tensorflow-metal==1.1.0
7. 下载jupyter notebook(或者直接跳到step10运行python代码)：conda install jupyter notebook
8. 运行jupyter notebook：jupyter notebook
   • 如果运行失败显示了下面的错误，安装chardet：pip install chardet

     Traceback (most recent call last): File
     “/opt/miniconda3/envs/tf-metal/lib/python3.11/site-packages/requests/compat.py”,
     line 11, in
     import chardet ModuleNotFoundError: No module named ‘chardet’

9. 在jupyter notebook（或者直接python脚本）进行测试：

• 显示信息 ref: github参考

  import numpy as np
  import pandas as pd
  import tensorflow as tf
  
  # Check for TensorFlow GPU access
  print(f"TensorFlow has access to the following devices:\n{tf.config.list_physical_devices()}")
  
  # See TensorFlow version
  print(f"TensorFlow version: {tf.__version__}")
  

  

• 训练测试

  import tensorflow as tf
  
  cifar = tf.keras.datasets.cifar100
  (x_train, y_train), (x_test, y_test) = cifar.load_data()
  model = tf.keras.applications.ResNet50(
      include_top=True,
      weights=None,
      input_shape=(32, 32, 3),
      classes=100,)
  
  loss_fn = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=False)
  model.compile(optimizer="adam", loss=loss_fn, metrics=["accuracy"])
  model.fit(x_train, y_train, epochs=5, batch_size=64) 
  

  
  ⚠️提示了：Could not identify NUMA node of platform GPU ID 0, defaulting to 0. Your kernel may not have been built with NUMA support.
  🙆The NUMA error message for an Apple Silicon computer is benign and can be ignored. Apple silicon memory is UMA (unified memory architecture) not NUMA. — from 苹果开发者论坛讨论帖

• 依旧担心上面的提示？使用矩阵乘法对比gpu和cpu

  import tensorflow as tf
  import time
  
  # 创建一个随机的大型矩阵
  matrix_size = 10000
  matrix_a = tf.random.normal((matrix_size, matrix_size))
  matrix_b = tf.random.normal((matrix_size, matrix_size))
  
  # 使用CPU执行矩阵乘法
  with tf.device('/CPU:0'):
      start_time = time.time()
      result_cpu = tf.matmul(matrix_a, matrix_b)
      end_time = time.time()
      print("使用CPU执行矩阵乘法所需时间：", end_time - start_time, "秒")
  
  # 使用GPU执行矩阵乘法
  with tf.device('/GPU:0'):
      start_time = time.time()
      result_gpu = tf.matmul(matrix_a, matrix_b)
      end_time = time.time()
      print("使用GPU执行矩阵乘法所需时间：", end_time - start_time, "秒")
  

  🎉 结果对比超明显