Note

――I hope you can think of it as a personal memo when building the environment again someday. ――Since the evolution is fast, the version may be different from the current one. ――We cannot take responsibility for "it broke when you tried it", so thank you. .. ..

The site I used as a reference, qiita

-Install TensorFlow 2.2 (GPU compatible) (on Windows) -Installing NVIDIA CUDA Toolkit 10.1, 10.0 (on Windows) -Install Visual C ++ Build Tools 2019 (Build Tools for Visual Studio 2019) (on Windows) -[Solving the problem that GPU is not recognized even though it is Cuda 10.1 in TensorFlow 2.1](https://medium.com/lsc-psd/tensorflow2-1%E3%81%A7cuda10-1%E3%81%AA% E3% 81% AE% E3% 81% ABgpu% E3% 81% 8C% E8% AA% 8D% E8% AD% 98% E3% 81% 95% E3% 82% 8C% E3% 81% AA% E3% 81% 84% E5% 95% 8F% E9% A1% 8C% E3% 81% AE% E8% A7% A3% E6% B1% BA% E6% B3% 95-6be5137ec216) -Summary of CUDA + cuDNN installation

environment

windows10
NVIDIA GeForce GTX 1660 SUPER
CUDA 10.1
cuDNNv 7.6.5
python 3.7.9
tensorflow 2.1.0
keras 2.3.1
Visual Studio Community 2019 version16.2

1. Install Visual Studio

-Download from here --Required to use the CUDA toolkit described below --Select "Desktop development with C ++" for the workload on the setting screen after download.

2. Install CUDA toolkit

-Download version 10.1 (when using tensorflow2.x) from here --Check the version with nvcc -V

3. Install cuDNN

-Download version 7.6 from here (when using tensorflow2.x) --Membership registration is required.

For the installation of CUDA toolkit and cuDNN, please refer to the reference site, which is very carefully summarized.

4. Enter the anaconda virtual environment and check the following

――This time we created a new virtual environment. --Check if the result cudnn 7.6.5 cuda10.1_0 is returned by conda list cudnn. --If the above result is not obtained, execute conda install cudnn = 7.6.5 = cuda10.1_0.

5. Install TensorFlow 2.1

conda install tensorflow-gpu --python -c "import tensorflow as tf; print (tf.__version__)" to check the version

6. Install keras 2.3.1.

conda install keras==2.3.1 --Check the version with python -c" import keras; print (keras .__ version__) "

7. Check if it can be used with jupyter notebook

Is the GPU recognized?

--If the GPU is recognized, the result will include the description device_type:" GPU "

import temsorflow
from tensorflow.python.client import device_lib
device_lib.list_local_devices()

――After that, if you import tensorflow and keras, you can use it in my environment.

[Environment construction] @anaconda that runs keras / tensorflow on GPU