如何修复“模块‘tensorflow’没有属性‘estimator’”错误

Question

我正在使用 conda（通过 YAML 创建的 env）+ pip 在我的 Linux Mint 机器上设置 Tensorflow v1.13.1 环境。设置后，每当我尝试导入时，tf.estimator我都会收到AttributeError标题中描述的信息：

AttributeError: module 'tensorflow' has no attribute 'estimator'

• 我没有在 conda 环境（或在 pip 上）之外安装任何版本的 TF。
• 我确实有另一个以相同方式设置的 conda 环境，但使用 Tensorflow 2.0 alpha，它可以tf.estimator很好地导入。
• 我尝试删除环境并重新构建它
• 我试过重新启动:(
• 出于某种原因，我无法将 conda 更新到 4.6：

$ conda update -n base -c defaults conda

# >>>>>>>>>>>>>>>>>>>>>> ERROR REPORT <<<<<<<<<<<<<<<<<<<<<<

    Traceback (most recent call last):
      File "/usr/share/anaconda3/lib/python3.7/site-packages/conda/exceptions.py", line 819, in __call__
        return func(*args, **kwargs)
      File "/usr/share/anaconda3/lib/python3.7/site-packages/conda/cli/main.py", line 78, in _main
        exit_code = do_call(args, p)
      File "/usr/share/anaconda3/lib/python3.7/site-packages/conda/cli/conda_argparse.py", line 77, in do_call
        exit_code = getattr(module, func_name)(args, parser)
      File "/usr/share/anaconda3/lib/python3.7/site-packages/conda/cli/main_update.py", line 14, in execute
        install(args, parser, 'update')
      File "/usr/share/anaconda3/lib/python3.7/site-packages/conda/cli/install.py", line 253, in install
        handle_txn(unlink_link_transaction, prefix, args, newenv)
      File "/usr/share/anaconda3/lib/python3.7/site-packages/conda/cli/install.py", line 282, in handle_txn
        unlink_link_transaction.execute()
      File "/usr/share/anaconda3/lib/python3.7/site-packages/conda/core/link.py", line 223, in execute
        self.verify()
      File "/usr/share/anaconda3/lib/python3.7/site-packages/conda/common/io.py", line 46, in decorated
        return f(*args, **kwds)
      File "/usr/share/anaconda3/lib/python3.7/site-packages/conda/core/link.py", line 200, in verify
        self.prepare()
      File "/usr/share/anaconda3/lib/python3.7/site-packages/conda/core/link.py", line 192, in prepare
        stp.remove_specs, stp.update_specs)
      File "/usr/share/anaconda3/lib/python3.7/site-packages/conda/core/link.py", line 282, in _prepare
        mkdir_p(transaction_context['temp_dir'])
      File "/usr/share/anaconda3/lib/python3.7/site-packages/conda/gateways/disk/__init__.py", line 60, in mkdir_p
        makedirs(path)
      File "/usr/share/anaconda3/lib/python3.7/os.py", line 221, in makedirs
        mkdir(name, mode)
    PermissionError: [Errno 13] Permission denied: '/usr/share/anaconda3/.condatmp'

yml 文件如下所示：

dependencies:
- python
- numpy
- tensorflow
- cudatoolkit==9.0
...

从有问题的环境内部：

• $ conda list tensorflow

    # packages in environment at /home/cjs/.conda/envs/my-env:
    # 
    # Name                    Version                   Build  Channel
    tensorflow                1.13.1          mkl_py37h54b294f_0
    tensorflow-base           1.13.1          mkl_py37h7ce6ba3_0
    tensorflow-estimator      1.13.0                     py_0

• $ pip list | grep tensorflow

    tensorflow                  1.13.1
    tensorflow-estimator        1.13.0

• $ which pip

    /home/cjs/.conda/envs/my-env/bin/pip

• $ conda --version

    conda 4.5.11

• $ pip --version

    pip 19.0.3 from /home/cjs/.local/lib/python3.7/site-packages/pip (python 3.7)

这是该问题的一个最小示例。如您所见，这只发生在调用 tf.estimator 的地方，所有其他 Tensorflow 属性都按预期运行：

Python 3.7.3 (default, Mar 27 2019, 22:11:17)
[GCC 7.3.0] :: Anaconda, Inc. on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import tensorflow as tf
>>> tf.__version__
'1.13.1'
>>> tf.estimator
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
AttributeError: module 'tensorflow' has no attribute 'estimator'
>>> tf.estimator.Estimator()
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
AttributeError: module 'tensorflow' has no attribute 'estimator'
>>> from tensorflow import estimator
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
ImportError: cannot import name 'estimator' from 'tensorflow' (/home/cjs/.conda/envs/my-env/lib/python3.7/site-packages/tensorflow/__init__.py)
>>> tf.Variable
<class 'tensorflow.python.ops.variables.VariableV1'>
>>> tf.keras
<module 'tensorflow._api.v1.keras' from '/home/cjs/.conda/envs/my-env/lib/python3.7/site-packages/tensorflow/_api/v1/keras/__init__.py'>
>>> tf.constant
<function constant_v1 at 0x7fb25ea24950>

---

更新

根据https://docs.nvidia.com/deploy/cuda-compatibility/index.html#binary-compatibility__table-toolkit-driver我发现我的 nvidia 驱动程序和 cudatoolkit 版本不同步（390.46 vs 9.0 ）。

我现在已经将我的 NVIDIA 驱动程序更新到 v418，并且能够将我的 conda 版本更新到 4.16.14。我将上面显示的 environment.yml 更新为cudatoolkit==10.1，但我似乎无法弄清楚如何实际安装它。

我的numba -s输出包括这一部分，这让我认为从一开始的整个问题就是 cuda 没有找到我的 GPU（或者无法连接到它？）。

__CUDA Information__
Error: CUDA device intialisation problem. Message:Error at driver init:
[100] Call to cuInit results in CUDA_ERROR_NO_DEVICE:
Error class: <class 'numba.cuda.cudadrv.error.CudaSupportError'>

---

（次要）更新

能够确定 numba 问题的原因是自从更新 GPU 驱动程序后我没有重新启动（废话）。

不过，并没有完全摆脱困境。新问题如下：

__CUDA Information__
Found 1 CUDA devices
id 0          b'Quadro K620'                              [SUPPORTED]
                      compute capability: 5.0
                           pci device id: 0
                              pci bus id: 1
Summary:
        1/1 devices are supported
CUDA driver version                 : 10010
CUDA libraries:
Finding cublas
        ERROR: can't locate lib
Finding cusparse
        ERROR: can't locate lib
Finding cufft
        ERROR: can't locate lib
Finding curand
        ERROR: can't locate lib
Finding nvvm
        ERROR: can't locate lib
        finding libdevice for compute_20...     ERROR: can't open libdevice for compute_20
        finding libdevice for compute_30...     ERROR: can't open libdevice for compute_30
        finding libdevice for compute_35...     ERROR: can't open libdevice for compute_35
        finding libdevice for compute_50...     ERROR: can't open libdevice for compute_50

Answer 1

终于找到问题了。我仍然安装了一些本地（非 Conda）Tensorflow 软件包，我猜它们在 python 环境中具有更高的优先级。

此链接解决了我的问题： https ://github.com/tensorflow/tensorboard/issues/2067

• 卸载tensorflow、tensorboard
• 卸载 tb-nightly（如果已安装）
• 使用“pip freeze | grep tensorflow”检查tensorflow-estimator包是否已安装。如果有，请将其卸载。
• 转到 site-packages 并删除与tensorflow、tensorboard、tensorflow-estimator 等相关的所有tensorflow 文件夹
• 重新安装最新版本的tensorflow和tensorboard

我的问题的关键是站点包，可以在两个位置找到

• ~/.conda/envs/<my-env>/lib/python3.<xx>/site-packages
• ~/.local/lib/python3.<xx>/site-packages

<my-env>你的conda环境在哪里，<xx>你的python版本在哪里。

只需将库rm -r <path to package>中的每个tensorflow包~/.local/重新安装conda环境即可。