TensorFlow >2.1 を試したい方は，こちらもご参照ください：
- https://qiita.com/gakyo/items/b4343e5c67e4fcaeed8e

ポイント

わりと大胆に conda に任せられる．お手軽
CUDA やら CUDNN やらの個別インストールは不要
なんだったら管理者権限もいらなかった
ただし，この方法はいまのところTF2.1までのみ対応
- 2.3とか2.4とかは Anaconda 様が対応してくれるのを待ちましょう

環境

Windows 10 Pro (20H2, 10.0.19042)
nVidia GeForce RTX 3090 (457.51)

(base) PS C:\Users\****> Get-WmiObject Win32_OperatingSystem


SystemDirectory : C:\Windows\system32
Organization    :
BuildNumber     : 19042
RegisteredUser  : ****
SerialNumber    : 
Version         : 10.0.19042


(base) PS C:\Users\****>nvidia-smi
Tue May 18 23:05:02 2021
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 457.51       Driver Version: 457.51       CUDA Version: 11.1     |
|-------------------------------+----------------------+----------------------+
| GPU  Name            TCC/WDDM | Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  GeForce RTX 3090   WDDM  | 00000000:01:00.0  On |                  N/A |
| 44%   59C    P2   113W / 350W |    892MiB / 24576MiB |      1%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
|    0   N/A  N/A      1360    C+G   Insufficient Permissions        N/A      |
|    0   N/A  N/A      1560    C+G   Insufficient Permissions        N/A      |
|    0   N/A  N/A      3368    C+G   Insufficient Permissions        N/A      |
|    0   N/A  N/A      7884    C+G   Insufficient Permissions        N/A      |
|    0   N/A  N/A      8384    C+G   C:\Windows\explorer.exe         N/A      |
|    0   N/A  N/A      9208    C+G   ...ekyb3d8bbwe\YourPhone.exe    N/A      |
|    0   N/A  N/A      9952    C+G   ...5n1h2txyewy\SearchApp.exe    N/A      |
|    0   N/A  N/A     10400    C+G   ...nputApp\TextInputHost.exe    N/A      |
|    0   N/A  N/A     11312      C   ...vs\py37tf21gpu\python.exe    N/A      |
+-----------------------------------------------------------------------------+

手順

GeForceドライバをできればアップデートしておく¹
condaを使えるようにする²
Python 3.7の環境を conda で作成する³
```
conda create -n py37tf21gpu python=3.7
```
TensorFlow 2.1, CUDA toolkit, CuDNN などを conda で仮想環境にインストールする
```
conda install tensorflow=2.1.0=gpu_py37h7db9008_0 cudnn cudatoolkit jupyter matplotlib
```
- ここで重要なのは tensorflow のバージョンをビルドまで含めて指定すること⁴
  - Windows向けにGPU版が提供されているものの2021/5/18現在の最新版はTF 2.1.0 gpu_py37h7db9008_0．最新版は conda search tensorflow とかでビルドに gpu がついてるやつを探せばよい．
- cudnn と cuda toolkit は特に未指定でも適切なものが選択されるっぽい → 今回は cudatoolkit 10.1.243 h74a9793_0 と cudnn 7.6.5 cuda10.1_0 が勝手に入った．
適当なコードで動作確認する．
- 今回は，TensorFlow2の動作確認用のコードサンプル（非チュートリアル）を利用させてもらう
- 学習がはじまるまですごーく待たされるが，はじまると一瞬でおわる．すげーこんなんで構築終わって動くのか．なんか騙されてないか？⁵
- エポック数を100回くらいにして別窓でnvidia-smiを連打していると GPU 使用率 9% くらいでした．

(base) C:\Users\****>conda activate py37tf21gpu

(py37tf21gpu) C:\Users\****>ipython
Python 3.7.10 (default, Feb 26 2021, 13:06:18) [MSC v.1916 64 bit (AMD64)]
Type 'copyright', 'credits' or 'license' for more information
IPython 7.22.0 -- An enhanced Interactive Python. Type '?' for help.

In [1]: import numpy as np
   ...: import tensorflow as tf
   ...:
   ...: print(tf.__version__)
   ...:
   ...: gpus = tf.config.experimental.list_physical_devices('GPU')
   ...: if gpus:
   ...:     logical_gpus = tf.config.experimental.list_logical_devices('GPU')
   ...:     print("Physical GPUs: {}, Logical GPUs: {}".format(len(gpus), len(logical_gpus)))
   ...: else:
   ...:     print("CPU only")
   ...:
   ...: x = np.arange(-1, 1, 0.0001)
   ...: y = 0.8 * x + 0.2
   ...:
   ...: model = tf.keras.Sequential([tf.keras.layers.Dense(1, activation=None)])
   ...: model.compile("sgd", "mse")
   ...: model.build(input_shape=(0,1))
   ...: model.summary()
   ...: model.fit(x, y, epochs=5)
   ...:
   ...: print("ground truth: 0.8, 0.2")
   ...: print("estimated: ", model.variables[0][0,0].numpy(), model.variables[1][0].numpy())
2021-05-18 23:03:01.111120: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library cudart64_101.dll
2.1.0
2021-05-18 23:03:03.492644: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library nvcuda.dll
2021-05-18 23:03:03.621490: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1555] Found device 0 with properties:
pciBusID: 0000:01:00.0 name: GeForce RTX 3090 computeCapability: 8.6
coreClock: 1.725GHz coreCount: 82 deviceMemorySize: 24.00GiB deviceMemoryBandwidth: 871.81GiB/s
2021-05-18 23:03:03.621670: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library cudart64_101.dll
2021-05-18 23:03:03.696771: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library cublas64_10.dll
2021-05-18 23:03:03.733664: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library cufft64_10.dll
2021-05-18 23:03:03.746518: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library curand64_10.dll
2021-05-18 23:03:03.781143: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library cusolver64_10.dll
2021-05-18 23:03:03.800854: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library cusparse64_10.dll
2021-05-18 23:03:03.857241: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library cudnn64_7.dll
2021-05-18 23:03:03.857677: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1697] Adding visible gpu devices: 0
2021-05-18 23:03:03.860583: I tensorflow/core/platform/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX AVX2
2021-05-18 23:03:03.865619: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1555] Found device 0 with properties:
pciBusID: 0000:01:00.0 name: GeForce RTX 3090 computeCapability: 8.6
coreClock: 1.725GHz coreCount: 82 deviceMemorySize: 24.00GiB deviceMemoryBandwidth: 871.81GiB/s
2021-05-18 23:03:03.865659: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library cudart64_101.dll
2021-05-18 23:03:03.865682: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library cublas64_10.dll
2021-05-18 23:03:03.865702: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library cufft64_10.dll
2021-05-18 23:03:03.865723: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library curand64_10.dll
2021-05-18 23:03:03.865743: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library cusolver64_10.dll
2021-05-18 23:03:03.865764: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library cusparse64_10.dll
2021-05-18 23:03:03.865784: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library cudnn64_7.dll
2021-05-18 23:03:03.865825: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1697] Adding visible gpu devices: 0
2021-05-18 23:05:25.974908: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1096] Device interconnect StreamExecutor with strength 1 edge matrix:
2021-05-18 23:05:25.974975: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1102]      0
2021-05-18 23:05:25.975247: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1115] 0:   N
2021-05-18 23:05:25.976321: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1241] Created TensorFlow device (/job:localhost/replica:0/task:0/device:GPU:0 with 22065 MB memory) -> physical GPU (device: 0, name: GeForce RTX 3090, pci bus id: 0000:01:00.0, compute capability: 8.6)
Physical GPUs: 1, Logical GPUs: 1
Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #
=================================================================
dense (Dense)                multiple                  2
=================================================================
Total params: 2
Trainable params: 2
Non-trainable params: 0
_________________________________________________________________
Train on 20000 samples
Epoch 1/5
2021-05-18 23:05:26.881185: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library cublas64_10.dll
20000/20000 [==============================] - 64s 3ms/sample - loss: 0.0081
Epoch 2/5
20000/20000 [==============================] - 0s 22us/sample - loss: 1.5123e-06
Epoch 3/5
20000/20000 [==============================] - 0s 21us/sample - loss: 3.5227e-10
Epoch 4/5
20000/20000 [==============================] - 0s 21us/sample - loss: 3.3563e-12
Epoch 5/5
20000/20000 [==============================] - 0s 22us/sample - loss: 3.0308e-12
ground truth: 0.8, 0.2
estimated:  0.80000293 0.19999996

In [2]:

最新版とまではいかずとも，とりあえず試す分にはお手軽でいいんじゃないでしょうか．

互換性がおおむね保たれている範囲では大丈夫なんでしょう． ↩
今回はユーザーモードで miniconda をインストール・利用した．具体的には https://repo.anaconda.com/miniconda/Miniconda3-py39_4.9.2-Windows-x86_64.exe． ↩
後述のように GPU対応版バイナリは TF2.1 のみが提供されているのでそれに応じて Python 3.7 にしておいた……が，次のステップでの tensorflow のバージョン・ビルド指定でのインストールのふるまいをみると Python 3.7にしなくてもよいのかもしれない． ↩
バージョン未指定だと Windows の最新バイナリの 2.3 などがインストールされる，がこれには gpu 版がないので，CPUのみのMKL版がインストールされる．そして巷間流布されているように別途 cuda やら cudnn やら tensorflow やら入れないと動かなかったりする．入れてもバージョン不適合で動かなかったりもする． ↩
ごりごり使っているわけではないので，何か不具合があっても不思議には思わない ↩

Windows PC に conda だけでGPU計算可能な TensorFlow 2.1 をセットアップする

ポイント

環境

手順