Failed to install from pip/pypi?

[Grant-Tao]

where is 'generate_dist'?

and when I use 'pip install tensor flow-gpu-macosx, I got the following error:

conf = open(pwd + '/.tf_configure.bazelrc', 'r')
FileNotFoundError: [Errno 2] No such file or directory: '/.tf_configure.bazelrc'

thanks

[zylo117]

Hi, I think I didn't update the readme file after I update the source code obviously. I don't recommend you use pip install it from pypi. Instead, you should just download the newest release .whl from Release. Then 'pip install xxx.whl'.

I'd love to upload the whole whl to pypi but they forbid me because it's too big.
So I uploaded a source code and the script to compile, which will take more than an hour of your time.

[Grant-Tao]

thank you very much!
I've successfully installed the whl release (1.8.0), but :
import tensorflow as tf: success
sess = tf.Session(): Error prompt:
E tensorflow/core/grappler/clusters/utils.cc:127] Not found: TF GPU device with id 0 was not registered
yet, this error does not influence the python program, and it is verified, the GPU works normally.

My environment:
MacOS X 10.13.4
CUDA 9.1
Cudnn 7.0

I compared the efficiency of the same training program (mnist example) on macosx and on ubuntu 18.04, the program on Ubuntu platform runs much faster (roughly 30%) than that on MacOSX, why?

[zylo117]

This warning does exist and has no effects for me for now. As for the performance issue, Tensorflow-gpu was never meant to run on Mac, all I did was fixing the compile errors directly. Considering that, not all source code was being compiled in the official way, having performance issue is understandable. Also, are you sure the test was under the same hardware and CUDA toolkit version? BTW, nvidia driver on Mac sucks.