Azure ML Studio笔记本导入MobileNetV2模型触发AttributeError报错

问题现象 #

在Azure ML Studio笔记本环境中导入Keras内置的MobileNetV2模型时触发报错，完全相同的代码在本地机器Jupyter Notebook中可正常运行。
报错触发阶段为模型权重加载环节，核心异常为：
AttributeError: 'str' object has no attribute 'decode'

复现代码 #

import tensorflow as tf
from tensorflow import keras

model = tf.keras.applications.MobileNetV2()

完整报错日志 #

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
Input In [23], in <cell line: 1>()
----> 1 model = tf.keras.applications.MobileNetV2()

File /anaconda/envs/azureml_py38/lib/python3.8/site-packages/tensorflow/python/keras/applications/mobilenet_v2.py:405, in MobileNetV2(input_shape, alpha, include_top, weights, input_tensor, pooling, classes, classifier_activation, **kwargs)
    402     weight_path = BASE_WEIGHT_PATH + model_name
    403     weights_path = data_utils.get_file(
    404         model_name, weight_path, cache_subdir='models')
--> 405   model.load_weights(weights_path)
    406 elif weights is not None:
    407   model.load_weights(weights)

File /anaconda/envs/azureml_py38/lib/python3.8/site-packages/tensorflow/python/keras/engine/training.py:250, in Model.load_weights(self, filepath, by_name, skip_mismatch)
    246   if (self._distribution_strategy.extended.steps_per_run > 1 and
    247       (not network._is_hdf5_filepath(filepath))):  # pylint: disable=protected-access
    248     raise ValueError('Load weights is not yet supported with TPUStrategy '
    249                      'with steps_per_run greater than 1.')
--> 250 return super(Model, self).load_weights(filepath, by_name, skip_mismatch)

File /anaconda/envs/azureml_py38/lib/python3.8/site-packages/tensorflow/python/keras/engine/network.py:1266, in Network.load_weights(self, filepath, by_name, skip_mismatch)
   1263   hdf5_format.load_weights_from_hdf5_group_by_name(
   1264       f, self.layers, skip_mismatch=skip_mismatch)
   1265 else:
-> 1266   hdf5_format.load_weights_from_hdf5_group(f, self.layers)

File /anaconda/envs/azureml_py38/lib/python3.8/site-packages/tensorflow/python/keras/saving/hdf5_format.py:659, in load_weights_from_hdf5_group(f, layers)
    648 """Implements topological (order-based) weight loading.
    649 
    650 Arguments:
   (...)
    656         and weights file.
    657 """
    658 if 'keras_version' in f.attrs:
--> 659   original_keras_version = f.attrs['keras_version'].decode('utf8')
    660 else:
    661   original_keras_version = '1'

AttributeError: 'str' object has no attribute 'decode'

故障原因 #

这是典型的依赖版本不兼容问题：

• Azure ML默认的azureml_py38环境中预装的h5py版本为3.x系列，该版本读取HDF5文件属性时会直接返回字符串类型，不需要额外做字节转字符串的decode操作
• 环境中预装的TensorFlow为2.5及更早版本，这部分版本的Keras权重加载逻辑是适配h5py 2.x版本编写的，默认读取到的属性是字节类型，会主动调用.decode('utf8')做转码，遇到3.x版本返回的字符串时就会触发属性不存在的报错
  本地环境要么h5py停留在2.x版本，要么TensorFlow版本已经升级到适配h5py3.x的新版本，所以不会出现该问题。

解决方法 #

二选一即可，优先推荐第一种，和Azure ML默认环境的其他依赖冲突最少：

• 方案1：降级h5py到兼容的2.x版本，在Notebook单元格中执行以下命令，安装完成后重启内核再运行代码即可

  !pip install h5py==2.10.0
  

• 方案2：升级TensorFlow到2.6及以上版本，新版本已经修复了h5py 3.x的兼容问题，执行以下命令后重启内核即可

  !pip install --upgrade tensorflow
  

内容的提问来源于stack exchange，提问作者Skyward