217 - 9 steps to installing TensorFlow GPU on Windows 10

195 Aufrufe

  • Danke für die Bewertung! Teile es mit Deinen Freunden!

    Danke für Deine Bewertung!

admin
Published
Step 1: Install Visual Studio community edition: I installed VS2019
https://visualstudio.microsoft.com/vs/community/

Step 2: Install your IDE if you haven't already done so. I installed Spyder 4.1.5 via Anaconda individual edition.
https://www.anaconda.com/products/individual

Step 3: Setup a separate environment for your IDE. I created an env. with python 3.7
I skipped 3.8 as it is relatively new and some of my libraries may not be tested or available.
On Anaconda you can use Anaconda Navigaor to create a new environment.
If you only see 3.8 as option on Navigator you can create new env from anaconda prompt:
conda create -n py37gpu python=3.7 anaconda

For the following steps verify instructions from www.tensorflow.org/install/gpu

Step 4: Verify your GPU is actually supported for deep learning:
https://developer.nvidia.com/cuda-gpus

Step 5: Figure out your GPU model: I have P5000 Quadro
and Update the GPU driver: For me it updated to Version 461.72
https://www.nvidia.com/download/index.aspx?lang=en-us

Step 6: Download and install CUDA Toolkit. I installed Version 11.0
https://developer.nvidia.com/cuda-toolkit-archive

Step 7: Download cuDNN. I downloaded version 8.0.4
You need to sign up for Nvidia developer program (free)
Extract all folder contents from cudnn you just downloaded to
C:\program files\Nvidia GPU computing toolkit\CUDA\v11.0

Step 8: Install tensorflow
Open spyder in the new environment that got named as: py37gpu (or whatever name you assigned)
- pip install tensorflow

Step 9: Verify the installation
Create a new python file and run these lines to test if GPU is recognized by tensorflow.
import tensorflow as tf
tf.test.is_gpu_available(
cuda_only=False, min_cuda_compute_capability=None
)

The last few lines of my output shows...
physical GPU (device: 0, name: Quadro P5000, pci bus id: 0000:03:00.0, compute capability: 6.1)
physical GPU (device: 1, name: Quadro P5000, pci bus id: 0000:a1:00.0, compute capability: 6.1)

I have 2 GPUs.

But on Windows you will not be able to efficiently use both GPUs for training.

#On windows sysems you cannot install NCCL that is required for multi GPU
#So we need to follow Heirarchial copy method or reduce to single GPU.

strategy = tf.distribute.MirroredStrategy(devices=devices_names[:n_gpus],
cross_device_ops=tf.distribute.HierarchicalCopyAllReduce())

#The following will raise an error on Windows about NCCL.
#strategy = tf.distribute.MirroredStrategy(devices=devices_names[:n_gpus])
Kategorien
PC (Windows/Mac/Linux) Anleitungen
Kommentare deaktiviert.

Als nächstes

Яндекс.Метрика