如何在 TensorFlow 中使用 GPU？

Question

我的电脑

Microsoft Windows [Version 10.0.22621.963]
(c) Microsoft Corporation. All rights reserved.

C:\Users\donhu>nvcc -V
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2022 NVIDIA Corporation
Built on Tue_May__3_19:00:59_Pacific_Daylight_Time_2022
Cuda compilation tools, release 11.7, V11.7.64
Build cuda_11.7.r11.7/compiler.31294372_0

C:\Users\donhu>nvidia-smi
Sat Dec 17 23:40:44 2022
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 512.77       Driver Version: 512.77       CUDA Version: 11.6     |
|-------------------------------+----------------------+----------------------+
| 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  NVIDIA GeForce ... WDDM  | 00000000:01:00.0  On |                  N/A |
| 34%   31C    P8    16W / 125W |   1377MiB /  6144MiB |      4%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
|    0   N/A  N/A      3392    C+G   C:\Windows\explorer.exe         N/A      |
|    0   N/A  N/A      4484    C+G   ...artMenuExperienceHost.exe    N/A      |
|    0   N/A  N/A      6424    C+G   ...n1h2txyewy\SearchHost.exe    N/A      |
|    0   N/A  N/A      6796    C+G   ...lPanel\SystemSettings.exe    N/A      |
|    0   N/A  N/A      7612    C+G   ...8bbwe\WindowsTerminal.exe    N/A      |
|    0   N/A  N/A      9700    C+G   ...8bbwe\WindowsTerminal.exe    N/A      |
|    0   N/A  N/A     10624    C+G   ...perience\NVIDIA Share.exe    N/A      |
|    0   N/A  N/A     10728    C+G   ...er Java\jre\bin\javaw.exe    N/A      |
|    0   N/A  N/A     13064    C+G   ...8bbwe\WindowsTerminal.exe    N/A      |
|    0   N/A  N/A     14496    C+G   ...462.46\msedgewebview2.exe    N/A      |
|    0   N/A  N/A     17124    C+G   ...ooting 2\BugShooting2.exe    N/A      |
|    0   N/A  N/A     19064    C+G   ...8bbwe\Notepad\Notepad.exe    N/A      |
|    0   N/A  N/A     19352    C+G   ...8bbwe\WindowsTerminal.exe    N/A      |
|    0   N/A  N/A     20920    C+G   ...y\ShellExperienceHost.exe    N/A      |
|    0   N/A  N/A     21320    C+G   ...e\PhoneExperienceHost.exe    N/A      |
|    0   N/A  N/A     21368    C+G   ...me\Application\chrome.exe    N/A      |
+-----------------------------------------------------------------------------+

C:\Users\donhu>

我训练模型

from tensorflow import keras
from tensorflow.keras import layers

def get_model():
    model = keras.Sequential([
        layers.Dense(512, activation="relu"),
        layers.Dense(10, activation="softmax")
    ])
    model.compile(optimizer="rmsprop",
                  loss="sparse_categorical_crossentropy",
                  metrics=["accuracy"])
    return model

model = get_model()
history_noise = model.fit(
    train_images_with_noise_channels, train_labels,
    epochs=10,
    batch_size=128,
    validation_split=0.2)

model = get_model()
history_zeros = model.fit(
    train_images_with_zeros_channels, train_labels,
    epochs=10,
    batch_size=128,
    validation_split=0.2)

源代码https://github.com/donhuvy/deep-learning-with-python-notebooks/blob/master/chapter05_fundamentals-of-ml.ipynb

如何在 TensorFlow 中使用 GPU？

Answer 1

对于 Tensorflow，使用 GPU 应该是自动的，您似乎缺少一些必需的组件（引用 Tensorflow 网页）：

The following NVIDIA® software are only required for GPU support.

NVIDIA® GPU drivers version 450.80.02 or higher.
CUDA® Toolkit 11.2.
cuDNN SDK 8.1.0.
(Optional) TensorRT to improve latency and throughput for inference.

在此处查看完整列表： https ://www.tensorflow.org/install/pip#software_requirements

安装所有这些后，Tensorflow 应该可以正常工作并显示它找到了可用的 GPU 设备。 另请注意，在下载软件包时，您需要相互匹配的 CUDA 和 cuDNN 版本。

Answer 2

如果使用虚拟环境 CMD 以管理员身份运行（重要：以管理员角色运行）

cd /d D:\temp20221103\
py --list
py -3.10 -m venv vy310
vy310\Scripts\activate
py -V

jupyter lab

!pip install tensorflow
!pip install cuda-python
!pip install nvidia-pyindex
!pip install nvidia-cudnn
!pip install tensorflow-gpu

import tensorflow as tf
tf.sysconfig.get_build_info()

tf.sysconfig.get_build_info()["cuda_version"]

结果

OrderedDict([('cpu_compiler',
              'C:/Program Files (x86)/Microsoft Visual Studio/2019/Community/VC/Tools/MSVC/14.29.30133/bin/HostX64/x64/cl.exe'),
             ('cuda_compute_capabilities',
              ['sm_35', 'sm_50', 'sm_60', 'sm_70', 'sm_75', 'compute_80']),
             ('cuda_version', '64_112'),
             ('cudart_dll_name', 'cudart64_112.dll'),
             ('cudnn_dll_name', 'cudnn64_8.dll'),
             ('cudnn_version', '64_8'),
             ('is_cuda_build', True),
             ('is_rocm_build', False),
             ('is_tensorrt_build', False),
             ('msvcp_dll_names', 'msvcp140.dll,msvcp140_1.dll'),
             ('nvcuda_dll_name', 'nvcuda.dll')])

作为屏幕截图，您正在使用 Anaconda。 需要安装

cudatoolkit
cudnn

然后

tf.debugging.set_log_device_placement(True)

import tensorflow as tf

tf.debugging.set_log_device_placement(False)
# mirrored_strategy = tf.distribute.MirroredStrategy(devices=["/cpu:0", "/gpu:1"])
mirrored_strategy = tf.distribute.MirroredStrategy(devices=["/gpu:0"])

结果