cannot access GPU with tensorflow-gpu 1.8.0 conda package

0
votes

I have the tensorflow and tensorflow-gpu 1.8.0 conda (not pip) packages installed in a conda environment on Ubuntu 16.04.4:

conda list t.*flow
# packages in environment at /home/lebedov/miniconda3/envs/TF: 
#
# Name                    Version                   Build  Channel
_tflow_180_select         1.0                         gpu  
tensorflow                1.8.0                    py36_1    conda-forge
tensorflow-gpu            1.8.0                h7b35bdc_0

I have CUDA 9.0 installed on my system, which has a Quadro M2200 GPU. I can see the GPU listed in the output of nvidia-smi and can also access the GPU using other deep learning frameworks such as PyTorch 0.4.0, but for some reason TensorFlow doesn't seem to see it:

Python 3.6.5 | packaged by conda-forge | (default, Apr  6 2018, 13:39:56) 
Type 'copyright', 'credits' or 'license' for more information
IPython 6.4.0 -- An enhanced Interactive Python. Type '?' for help.

In [1]: import tensorflow as tf
   ...: sess = tf.Session(config=tf.ConfigProto(log_device_placement=True))
2018-07-11 23:21:11.827064: I tensorflow/core/platform/cpu_feature_guard.cc:140] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
Device mapping: no known devices.
2018-07-11 23:21:11.827942: I tensorflow/core/common_runtime/direct_session.cc:284] Device mapping:

If I downgrade to tensorflow-gpu 1.7.0, however, I can see the GPU. Any thoughts as to why the GPU isn't being detected by TensorFlow 1.8.0?

2
tensorflowgpu
You have both CPU and GPU tensorflow installed? How can that be? - Dr. Snoopy
That's apparently how the tensorflow 1.8.0 conda packages are structured; see github.com/AnacondaRecipes/tensorflow_recipes - lebedov
To qualify my previous comment, the anaconda and conda-forge TensorFlow packages are evidently structured differently and might not necessarily get along. See my answer to this question. - lebedov

2 Answers

1
votes

The tensorflow 1.8.0 packages from Anaconda appear to support GPUs properly, but those from conda-forge do not; see this issue.

1
votes

I also had similar instances where the tensorflow-gpu won't run on default.

As for me, I solved it by just uninstalling tensorflow and only keeping tensorflow-gpu. That way, it would always run with GPU as there was no other option.

As far as compatibility is concerned, I would recommend the new conda way of installing tensorflow-gpu, which automatically installs the relevant cudnn files for you. The code would be as follows:

conda create -n [EnvironmentName] python=3.6

conda activate [EnvironmentName]

conda install -c conda-forge tensorflow-gpu==1.14

it will assess which version (CUDA,CUDNN, etc.) you require and download and install it directly to your environment. Then run your python file from this environment.

Do check afterwards with "conda list" to ensure that tensorflow-gpu is installed for you.