Apple M2 Pro运行TensorFlow测试脚本出现总线错误求助

在Apple M2 Pro设备上运行TensorFlow测试脚本时触发总线错误（zsh: bus error）。

测试脚本代码：

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=True)
model.compile(optimizer="adam", loss=loss_fn, metrics=["accuracy"])
model.fit(x_train, y_train, epochs=5, batch_size=4)

终端输出信息：

Metal device set to: Apple M2 Pro

systemMemory: 16.00 GB
maxCacheSize: 5.33 GB

2023-03-23 00:26:32.203361: I tensorflow/core/common_runtime/pluggable_device/pluggable_device_factory.cc:305] Could not identify NUMA node of platform GPU ID 0, defaulting to 0. Your kernel may not have been built with NUMA support.
2023-03-23 00:26:32.203521: I tensorflow/core/common_runtime/pluggable_device/pluggable_device_factory.cc:271] Created TensorFlow device (/job:localhost/replica:0/task:0/device:GPU:0 with 0 MB memory) -> physical PluggableDevice (device: 0, name: METAL, pci bus id: <undefined>)
zsh: bus error  python3 app/model/tf_verify.py

• TensorFlow Metal插件兼容性bug：Apple Silicon设备上的TensorFlow依赖tensorflow-metal实现GPU加速，早期版本的插件在处理ResNet50这类模型时，存在底层内存访问或指令集适配问题，直接触发总线错误。
• 输入尺寸不匹配：ResNet50原生设计的输入尺寸为224×224，强行传入32×32的CIFAR-100数据，会导致模型内部部分层的张量计算出现内存越界，引发总线错误。
• GPU内存识别异常：终端输出显示GPU设备内存为0MB，说明TensorFlow对Metal设备的内存识别存在问题，进而导致内存分配失败触发总线错误。

• 升级TensorFlow及Metal插件：安装最新版的tensorflow-macos和tensorflow-metal，新版本会修复Apple Silicon平台的兼容性问题。执行以下命令更新：

  pip install --upgrade tensorflow-macos tensorflow-metal
  

• 调整输入尺寸至模型适配规格：将CIFAR-100的32×32图片resize为ResNet50适配的224×224，修改脚本如下：

  import tensorflow as tf
  cifar = tf.keras.datasets.cifar100
  (x_train, y_train), (x_test, y_test) = cifar.load_data()
  # 调整输入尺寸到224×224
  x_train = tf.image.resize(x_train, (224, 224))
  x_test = tf.image.resize(x_test, (224, 224))
  model = tf.keras.applications.ResNet50(
      include_top=True,
      weights=None,
      input_shape=(224, 224, 3),
      classes=100,)
  loss_fn = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
  model.compile(optimizer="adam", loss=loss_fn, metrics=["accuracy"])
  model.fit(x_train, y_train, epochs=5, batch_size=4)
  

• 强制使用CPU运行：如果升级插件后问题仍存在，可临时禁用GPU，让TensorFlow使用CPU执行，规避Metal插件的适配问题。在脚本开头添加：

  import os
  os.environ["CUDA_VISIBLE_DEVICES"] = "-1"
  

• 进一步降低batch size：将batch size从4调整为2，减少单步训练的内存占用，避免内存分配异常。

内容的提问来源于stack exchange，提问作者Josh Purtell