python - 为什么 TensorFlow WHL 文件不包含所有必需的依赖项？

Question

系统:windows 10 64 位，python 3.6.5 64 位。使用 python 。 CUDA 10.0，安装了 cuddn。

美好的一天。出于安全原因，我们公司拒绝访问互联网。我从官网下载的TF tensorFlow whl包:

https: //storage.googleapis.com/tensorflow/windows/gpu/tensorflow_gpu-1.14.0-cp36-cp36m-win_amd64.whl

我尝试使用 pip 安装它:

pip install --upgrade D:\tensorflow_gpu-1.14.0-cp36-cp36m-win_amd64.whl

我收到以下消息:

collecting keras-preprocessing>=1.0.5 (from tensorflow-gpu==1.14.0)
Retrying ....after connection broken by 'NewConnectionError'
......
could not find a version that satifies the requirement keras-preprocessing=>1.0.5.....
no matching distribution found for keras-preprocessing=>1.0.5 (from tensorflow-gpu==1.14.0)

重启时，会找不到其他一些TensorFlow依赖，据我了解，pip根本看不到任何依赖

我认为 WHL 包具有所有必要的依赖项，但是在安装 pip 时它会尝试从 Internet 下载一些东西。告诉我我做错了什么？

pip 检查 tensorflow 结果:

(base) C:\Windows\system32>pip check tensorflow

tensorflow 1.10.0 requires absl-py, which is not installed.
tensorflow 1.10.0 requires astor, which is not installed.
tensorflow 1.10.0 requires gast, which is not installed.
tensorflow 1.10.0 requires grpcio, which is not installed.
tensorflow 1.10.0 requires protobuf, which is not installed.
tensorflow 1.10.0 requires tensorboard, which is not installed.
tensorflow 1.10.0 requires termcolor, which is not installed.
tensorflow-gpu 1.14.0 requires absl-py, which is not installed.
tensorflow-gpu 1.14.0 requires astor, which is not installed.
tensorflow-gpu 1.14.0 requires gast, which is not installed.
tensorflow-gpu 1.14.0 requires google-pasta, which is not installed.
tensorflow-gpu 1.14.0 requires grpcio, which is not installed.
tensorflow-gpu 1.14.0 requires keras-applications, which is not installed.
tensorflow-gpu 1.14.0 requires keras-preprocessing, which is not installed.
tensorflow-gpu 1.14.0 requires protobuf, which is not installed.
tensorflow-gpu 1.14.0 requires tensorboard, which is not installed.
tensorflow-gpu 1.14.0 requires tensorflow-estimator, which is not installed.
tensorflow-gpu 1.14.0 requires termcolor, which is not installed.
distributed 1.21.8 requires msgpack, which is not installed.
tensorflow-gpu 1.14.0 has requirement numpy<2.0,>=1.14.5, but you have numpy 1.14.3.
tensorflow-gpu 1.14.0 has requirement wrapt>=1.11.1, but you have wrapt 1.10.11.

Answer 1

I thought that the WHL package has all the necessary dependencies

没有。轮子不包含所有依赖项，它们也不应该包含，否则每个轮子都会非常臃肿。如果您必须在无法访问 Internet 的机器上安装 tensorflow-gpu(或任何其他轮子)，请执行以下步骤:

1. 在一台有互联网访问的机器上，运行:

   \> pip download tensorflow-gpu --dest some\dir
   

   (您可以通过将包含下载的 tensorflow_gpu-1.14.0-cp36-cp36m-win_amd64.whl 的目录作为 --dest 来运行此命令，这样您不用重新下载)

   这将下载 tensorflow-gpu 及其所有依赖树。

2. 现在将目录(已下载的所有文件)传输到计算机没有互联网访问并运行:

   \> pip install tensorflow-gpu --no-index --find-links some\dir
   

   --no-index 将指示pip 不在线查找包，--find-links 将指向本地目录在哪里可以找到安装包。