为什么我的keras/tensorflow不使用GPU

我看过其他一些关于stackoverflow的帖子，但似乎没有任何帮助。我只安装了tensorflow gpu版本和keras

我检查了gpu是否被识别，它是：

print(device_lib.list_local_devices())
print('Tensorflow: ', tf.__version__)

出去

<>但是当我用TysFROW后端运行KARAS模型时，它看起来并没有比在英伟达CPU上运行我的另一台计算机时更快速，没有NVIDIA。同时看看任务管理器，我看到了CPU使用率的峰值。我对这一点还不熟悉，并试图弄清楚这一切，但模型/训练是否应该运行得更快？如果使用GPU，我应该不会看到CPU使用率的上升吗

如果有人能帮忙，我们将不胜感激。我按照所有步骤安装tensorflow gpu、CUDA、cudnn等

以下是我所拥有的：

+-----------------------------------------------------------------------------+
| NVIDIA-SMI 387.92                 Driver Version: 387.92                    |
|-------------------------------+----------------------+----------------------+
| GPU  Name            TCC/WDDM | Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. | 


|   0  GeForce GTX 1070   WDDM  | 00000000:01:00.0  On |                  N/A |
| N/A   49C    P8     9W /  N/A |   7028MiB /  8192MiB |      0%      Default |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                       GPU Memory |
|  GPU       PID   Type   Process name                             Usage      |


|    0      3740      C   C:\Users\Jason\Anaconda3\pythonw.exe       N/A      |
|    0      3968    C+G   ...dows.Cortana_cw5n1h2txyewy\SearchUI.exe N/A      |
|    0      4392    C+G   C:\Windows\explorer.exe                    N/A      |
|    0      7760    C+G   ...0.0_x64__8wekyb3d8bbwe\WinStore.App.exe N/A      |
|    0      8976    C+G   ...6)\Google\Chrome\Application\chrome.exe N/A      |
|    0      9932    C+G   ...rosoft Office\root\Office16\WINWORD.EXE N/A      |
|    0     12632    C+G   Insufficient Permissions                   N/A      |
|    0     14168    C+G   Insufficient Permissions                   N/A      |
|    0     14548    C+G   ...t_cw5n1h2txyewy\ShellExperienceHost.exe N/A      |
|    0     15648    C+G   Insufficient Permissions                   N/A      |
+-----------------------------------------------------------------------------+

在培训时：

C:\Program Files\NVIDIA Corporation\NVSMI>nvidia-smi
Fri Oct 13 13:18:23 2017
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 387.92                 Driver Version: 387.92                    |
|-------------------------------+----------------------+----------------------+
| GPU  Name            TCC/WDDM | Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |

|   0  GeForce GTX 1070   WDDM  | 00000000:01:00.0  On |                  N/A |
| N/A   51C    P2    34W /  N/A |   7126MiB /  8192MiB |     15%      Default |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                       GPU Memory |
|  GPU       PID   Type   Process name                             Usage      |

|    0      3740      C   C:\Users\Jason\Anaconda3\pythonw.exe       N/A      |
|    0      3968    C+G   ...dows.Cortana_cw5n1h2txyewy\SearchUI.exe N/A      |
|    0      4392    C+G   C:\Windows\explorer.exe                    N/A      |
|    0      7760    C+G   ...0.0_x64__8wekyb3d8bbwe\WinStore.App.exe N/A      |
|    0      8976    C+G   ...6)\Google\Chrome\Application\chrome.exe N/A      |
|    0      9932    C+G   ...rosoft Office\root\Office16\WINWORD.EXE N/A      |
|    0     12632    C+G   Insufficient Permissions                   N/A      |
|    0     14168    C+G   Insufficient Permissions                   N/A      |
|    0     14548    C+G   ...t_cw5n1h2txyewy\ShellExperienceHost.exe N/A      |
|    0     15648    C+G   Insufficient Permissions                   N/A      |
+-----------------------------------------------------------------------------+

蟒蛇3.6； Tensorflow gpu 1.3

下载并安装Visual Studio 15（因为17不支持CUDA 8， 需要（而不是15）

CUDA v8.0：设置路径环境变量

更新的Nvidia驱动程序

已下载cudnn-v6.0以替代CUDA 8文件 第5.1条。因为Tensorflow 1.3支持Cudnn-v6.0：为那些

更新：

这是我正在运行的代码。我正在尝试为历代和批次获取最佳超参数。该模型采用Keras后端tensorflow

import numpy as np
from keras import models
from keras import layers
from keras.wrappers.scikit_learn import KerasRegressor
from sklearn.model_selection import GridSearchCV
import datetime

# Set random seed
np.random.seed(7)

# Create function returning a compiled network
def create_network(optimizer='adam',activation='relu'):

# Start neural network
network = models.Sequential()

# Add fully connected layer with a ReLU activation function
network.add(layers.Dense(units=dim, activation=activation, input_shape=(dim,)))

# Add fully connected layer with a ReLU activation function
network.add(layers.Dense(units=20, activation=activation))

# Add fully connected layer with a sigmoid activation function
network.add(layers.Dense(units=1, activation='linear'))

# Compile neural network
network.compile(loss='mse', # 
                optimizer=optimizer, # Optimizer
                metrics=['mae']) # Accuracy performance metric

# Return compiled network
return network

# Wrap Keras model so it can be used by scikit-learn
neural_network = KerasRegressor(build_fn=create_network, verbose=2)


# Create hyperparameter space
epochs = [5, 50, 100, 150]
batches = [1, 5, 10, 100]



# Create hyperparameter options
hyperparameters = dict(epochs=epochs, batch_size=batches)


# Time the process
start_time = datetime.datetime.now()

# Create grid search
grid = GridSearchCV(estimator=neural_network, param_grid=hyperparameters)

# Fit grid search
grid_result = grid.fit(X, Y)


# View hyperparameters of best neural network
grid_result.best_params_

# summarize results
print("Best: %f using %s" % (grid_result.best_score_, 
grid_result.best_params_))
means = grid_result.cv_results_['mean_test_score']
stds = grid_result.cv_results_['std_test_score']
params = grid_result.cv_results_['params']
for mean, stdev, param in zip(means, stds, params):
    print("%f (%f) with: %r" % (mean, stdev, param))

end_time = datetime.datetime.now()
processing_time = end_time - start_time
print(str(processing_time).split('.')[0])

更新2： 我一直在寻找答案，也许找到了什么…也许这就是原因。。。有人能证明谁知道的比我多吗

---

我正在使用scikit学习实现GridSearchCV。即使我在tensorflow后端使用keras，scikit learn也不使用GPU，这可能是原因吗？因此，除非我独家使用keras，否则我不会看到GPU启动？

当您进行培训时，

nvidia smi

的输出是什么？好的，我知道了如何看到nvidia smi（就像我说的，我是新手）。因此，它在训练时显示出0%到15%的差异。但是它的运行速度并没有明显加快。你有没有考虑过这个模型可能太小而无法从GPU加速中获益？Matias，我没有考虑过这一点。我一直将数据过滤（因为它无法处理所有数据，所以将数据过滤成更小的大小。我还试图看看过滤是否会改进模型）。但由于对TF的使用还不熟悉，只是不知道会发生什么。但是，如果你有经验，并且说如果两个（CPU和GPU）上的型号没有显著的差异，那么这就是为什么我在这里问。。。因此，我将尝试扩展它以看到不同之处。当我设想GPU将完成大部分工作时，我突然发现CPU利用率急剧上升。@chitown88我们还没有看到你的模型，所以我说的只是一个假设。

import numpy as np
from keras import models
from keras import layers
from keras.wrappers.scikit_learn import KerasRegressor
from sklearn.model_selection import GridSearchCV
import datetime

# Set random seed
np.random.seed(7)

# Create function returning a compiled network
def create_network(optimizer='adam',activation='relu'):

# Start neural network
network = models.Sequential()

# Add fully connected layer with a ReLU activation function
network.add(layers.Dense(units=dim, activation=activation, input_shape=(dim,)))

# Add fully connected layer with a ReLU activation function
network.add(layers.Dense(units=20, activation=activation))

# Add fully connected layer with a sigmoid activation function
network.add(layers.Dense(units=1, activation='linear'))

# Compile neural network
network.compile(loss='mse', # 
                optimizer=optimizer, # Optimizer
                metrics=['mae']) # Accuracy performance metric

# Return compiled network
return network

# Wrap Keras model so it can be used by scikit-learn
neural_network = KerasRegressor(build_fn=create_network, verbose=2)


# Create hyperparameter space
epochs = [5, 50, 100, 150]
batches = [1, 5, 10, 100]



# Create hyperparameter options
hyperparameters = dict(epochs=epochs, batch_size=batches)


# Time the process
start_time = datetime.datetime.now()

# Create grid search
grid = GridSearchCV(estimator=neural_network, param_grid=hyperparameters)

# Fit grid search
grid_result = grid.fit(X, Y)


# View hyperparameters of best neural network
grid_result.best_params_

# summarize results
print("Best: %f using %s" % (grid_result.best_score_, 
grid_result.best_params_))
means = grid_result.cv_results_['mean_test_score']
stds = grid_result.cv_results_['std_test_score']
params = grid_result.cv_results_['params']
for mean, stdev, param in zip(means, stds, params):
    print("%f (%f) with: %r" % (mean, stdev, param))

end_time = datetime.datetime.now()
processing_time = end_time - start_time
print(str(processing_time).split('.')[0])