我正在使用Keras使用fit_generator函数训练卷积神经网络，因为图像存储在.h5文件中并且不适合内存。在大多数情况下，由于模型卡在第一个时期的中间，我无法训练模型，否则会崩溃，并说“ GPU同步失败”或“ CUDA_ERROR_LAUNCH_FAILED”（请参阅下面的日志）。使用CPU的训练效果很好，但是当然会慢一些。我使用的是两台不同的机器，并且都有相同的问题。我的猜测是，这是与安装/配置有关的问题，但我不知道如何解决。
在两台机器上均按以下说明安装了Tensorflow：https://www.anaconda.com/blog/developer-blog/tensorflow-in-anaconda/
我已使用此脚本https://github.com/tensorflow/tensorflow/blob/master/tools/tf_env_collect.sh来收集以下信息。
在这里tf_env.txt
First machine:
Keras 2.2.4.
== cat /etc/issue ===============================================
Linux liph02.novalocal 3.10.0-862.9.1.el7.x86_64 #1 SMP Mon Jul 16 16:29:36 UTC 2018 x86_64 x86_64 x86_64 GNU/Linux
VERSION="7 (Core)"
VERSION_ID="7"
CENTOS_MANTISBT_PROJECT_VERSION="7"
REDHAT_SUPPORT_PRODUCT_VERSION="7"
== are we in docker =============================================
No
== compiler =====================================================
c++ (GCC) 4.8.5 20150623 (Red Hat 4.8.5-28)
Copyright (C) 2015 Free Software Foundation, Inc.
This is free software; see the source for copying conditions. There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
== uname -a =====================================================
Linux liph02.novalocal 3.10.0-862.9.1.el7.x86_64 #1 SMP Mon Jul 16 16:29:36 UTC 2018 x86_64 x86_64 x86_64 GNU/Linux
== check pips ===================================================
numpy 1.15.4
numpydoc 0.8.0
protobuf 3.6.1
tensorflow 1.12.0
== check for virtualenv =========================================
False
== tensorflow import ============================================
tf.VERSION = 1.12.0
tf.GIT_VERSION = b'unknown'
tf.COMPILER_VERSION = b'unknown'
Sanity check: array([1], dtype=int32)
== env ==========================================================
LD_LIBRARY_PATH /usr/local/cuda-9.2/lib64
DYLD_LIBRARY_PATH is unset
== nvidia-smi ===================================================
Fri Dec 28 16:13:39 2018
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 410.78 Driver Version: 410.78 CUDA Version: 10.0 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
|===============================+======================+======================|
| 0 TITAN Xp Off | 00000000:00:06.0 Off | N/A |
| 22% 38C P0 57W / 250W | 0MiB / 12196MiB | 0% Default |
+-------------------------------+----------------------+----------------------+
+-----------------------------------------------------------------------------+
| Processes: GPU Memory |
| GPU PID Type Process name Usage |
|=============================================================================|
| No running processes found |
+-----------------------------------------------------------------------------+
== cuda libs ===================================================
/usr/local/Wolfram/Mathematica/11.3/SystemFiles/Components/MXNetLink/LibraryResources/Linux-x86-64/libcudart.so.9.1
Second machine:
Keras 2.2.4.
== cat /etc/issue ===============================================
Linux liph01.novalocal 3.10.0-862.14.4.el7.x86_64 #1 SMP Wed Sep 26 15:12:11 UTC 2018 x86_64 x86_64 x86_64 GNU/Linux
VERSION="7 (Core)"
VERSION_ID="7"
CENTOS_MANTISBT_PROJECT_VERSION="7"
REDHAT_SUPPORT_PRODUCT_VERSION="7"
== are we in docker =============================================
No
== compiler =====================================================
c++ (GCC) 7.3.0
Copyright (C) 2017 Free Software Foundation, Inc.
This is free software; see the source for copying conditions. There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
== uname -a =====================================================
Linux liph01.novalocal 3.10.0-862.14.4.el7.x86_64 #1 SMP Wed Sep 26 15:12:11 UTC 2018 x86_64 x86_64 x86_64 GNU/Linux
== check pips ===================================================
msgpack-numpy 0.4.3.2
numpy 1.15.3
numpydoc 0.8.0
protobuf 3.6.0
tensorflow 1.11.0
== check for virtualenv =========================================
False
== tensorflow import ============================================
tf.VERSION = 1.11.0
tf.GIT_VERSION = b'unknown'
tf.COMPILER_VERSION = b'unknown'
== env ==========================================================
LD_LIBRARY_PATH is unset
DYLD_LIBRARY_PATH is unset
== nvidia-smi ===================================================
Thu Jan 3 17:38:44 2019
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 410.48 Driver Version: 410.48 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
|===============================+======================+======================|
| 0 TITAN Xp Off | 00000000:00:07.0 Off | N/A |
| 40% 65C P2 94W / 250W | 11747MiB / 12196MiB | 90% Default |
+-------------------------------+----------------------+----------------------+
+-----------------------------------------------------------------------------+
| Processes: GPU Memory |
| GPU PID Type Process name Usage |
|=============================================================================|
| 0 16991 C python 11737MiB |
+-----------------------------------------------------------------------------+
== cuda libs ===================================================
/usr/local/cuda-9.2/targets/x86_64-linux/lib/libcudart_static.a
/usr/local/cuda-9.2/targets/x86_64-linux/lib/libcudart.so.9.2.148
/usr/local/cuda-9.2/doc/man/man7/libcudart.7
/usr/local/cuda-9.2/doc/man/man7/libcudart.so.7
这是两个堆栈跟踪
(dev) -bash-4.2$ python classifier_training.py --dirs /data/simulations/Paranal_gam/ /data/simulations/Paranal_prot/ --epochs 1 --batch_size 32 --workers 16 --model ClassifierV2 --patience 1
Using TensorFlow backend.
ClassifierV2
Building training generator...
Building validation generator...
2018-12-18 12:15:19.553286: I tensorflow/core/platform/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: SSE4.1 SSE4.2 AVX AVX2 AVX512F FMA
2018-12-18 12:15:20.043811: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:964] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2018-12-18 12:15:20.047991: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1432] Found device 0 with properties:
name: TITAN Xp major: 6 minor: 1 memoryClockRate(GHz): 1.582
pciBusID: 0000:00:06.0
totalMemory: 11.91GiB freeMemory: 11.75GiB
2018-12-18 12:15:20.048093: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1511] Adding visible gpu devices: 0
Traceback (most recent call last):
File "classifier_training.py", line 122, in <module>
model = class_v2.get_model()
File "/data/ctasoft/cta-lstchain/cnn/classifiers.py", line 40, in get_model
self.model.add(Conv2D(16, kernel_size=(3, 3), input_shape=(1, self.img_rows, self.img_cols), data_format='channels_first', activation='relu'))
File "/data/ctasoft/anaconda3/envs/cta-dev/lib/python3.6/site-packages/keras/engine/sequential.py", line 165, in add
layer(x)
File "/data/ctasoft/anaconda3/envs/cta-dev/lib/python3.6/site-packages/keras/engine/base_layer.py", line 457, in __call__
output = self.call(inputs, **kwargs)
File "/data/ctasoft/anaconda3/envs/cta-dev/lib/python3.6/site-packages/keras/layers/convolutional.py", line 171, in call
dilation_rate=self.dilation_rate)
File "/data/ctasoft/anaconda3/envs/cta-dev/lib/python3.6/site-packages/keras/backend/tensorflow_backend.py", line 3641, in conv2d
x, tf_data_format = _preprocess_conv2d_input(x, data_format)
File "/data/ctasoft/anaconda3/envs/cta-dev/lib/python3.6/site-packages/keras/backend/tensorflow_backend.py", line 3521, in _preprocess_conv2d_input
if not _has_nchw_support() or force_transpose:
File "/data/ctasoft/anaconda3/envs/cta-dev/lib/python3.6/site-packages/keras/backend/tensorflow_backend.py", line 292, in _has_nchw_support
gpus_available = len(_get_available_gpus()) > 0
File "/data/ctasoft/anaconda3/envs/cta-dev/lib/python3.6/site-packages/keras/backend/tensorflow_backend.py", line 278, in _get_available_gpus
_LOCAL_DEVICES = get_session().list_devices()
File "/data/ctasoft/anaconda3/envs/cta-dev/lib/python3.6/site-packages/keras/backend/tensorflow_backend.py", line 186, in get_session
_SESSION = tf.Session(config=config)
File "/data/ctasoft/anaconda3/envs/cta-dev/lib/python3.6/site-packages/tensorflow/python/client/session.py", line 1551, in __init__
super(Session, self).__init__(target, graph, config=config)
File "/data/ctasoft/anaconda3/envs/cta-dev/lib/python3.6/site-packages/tensorflow/python/client/session.py", line 676, in __init__
self._session = tf_session.TF_NewSessionRef(self._graph._c_graph, opts)
tensorflow.python.framework.errors_impl.InternalError: CUDA runtime implicit initialization on GPU:0 failed. Status: unspecified launch failure
(dev) -bash-4.2$ python classifier_training.py --dirs /data/simulations/Paranal_gam /data/simulations/Paranal_prot --workers 1 --epochs 10 --batch_size 16 --model ClassifierV2 --patience 9
Using TensorFlow backend.
ClassifierV2
Building training generator...
Building validation generator...
2018-12-29 19:29:11.142008: I tensorflow/core/platform/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: SSE4.1 SSE4.2 AVX AVX2 AVX512F FMA
2018-12-29 19:29:11.892617: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:964] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning
NUMA node zero
2018-12-29 19:29:11.896828: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1432] Found device 0 with properties:
name: TITAN Xp major: 6 minor: 1 memoryClockRate(GHz): 1.582
pciBusID: 0000:00:06.0
totalMemory: 11.91GiB freeMemory: 11.75GiB
2018-12-29 19:29:11.896880: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1511] Adding visible gpu devices: 0
2018-12-29 19:29:12.960736: I tensorflow/core/common_runtime/gpu/gpu_device.cc:982] Device interconnect StreamExecutor with strength 1 edge matrix:
2018-12-29 19:29:12.960804: I tensorflow/core/common_runtime/gpu/gpu_device.cc:988] 0
2018-12-29 19:29:12.960819: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1001] 0: N
2018-12-29 19:29:12.961681: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1115] Created TensorFlow device (/job:localhost/replica:0/task:0/device:GPU:0 with 11366 MB memory) -> physical GPU (device:
0, name: TITAN Xp, pci bus id: 0000:00:06.0, compute capability: 6.1)
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d_1 (Conv2D) (None, 16, 98, 98) 160
_________________________________________________________________
conv2d_2 (Conv2D) (None, 16, 96, 96) 2320
_________________________________________________________________
average_pooling2d_1 (Average (None, 16, 48, 48) 0
_________________________________________________________________
dropout_1 (Dropout) (None, 16, 48, 48) 0
_________________________________________________________________
conv2d_3 (Conv2D) (None, 32, 46, 46) 4640
_________________________________________________________________
conv2d_4 (Conv2D) (None, 32, 44, 44) 9248
_________________________________________________________________
average_pooling2d_2 (Average (None, 32, 22, 22) 0
_________________________________________________________________
dropout_2 (Dropout) (None, 32, 22, 22) 0
_________________________________________________________________
flatten_1 (Flatten) (None, 15488) 0
_________________________________________________________________
dense_1 (Dense) (None, 128) 1982592
_________________________________________________________________
dropout_3 (Dropout) (None, 128) 0
_________________________________________________________________
dense_2 (Dense) (None, 256) 33024
_________________________________________________________________
dropout_4 (Dropout) (None, 256) 0
_________________________________________________________________
dense_3 (Dense) (None, 1) 257
=================================================================
Total params: 2,032,241
Trainable params: 2,032,241
Non-trainable params: 0
_________________________________________________________________
Epoch 1/10
4/8065 [..............................] - ETA: 1:52:06 - loss: 0.9940 - acc: 0.4531 - precision: 0.4947 - recall: 0.71882018-12-29 19:29:54.459471: E tensorflow/stream_executor/cuda/cuda_event.cc:48] E
rror polling for event status: failed to query event: CUDA_ERROR_LAUNCH_FAILED: unspecified launch failure
2018-12-29 19:29:54.459645: F tensorflow/core/common_runtime/gpu/gpu_event_mgr.cc:274] Unexpected Event status: 1
Aborted
答案 0 :(得分：0)
看起来在第一台机器上，CUDA 版本不匹配，请确保使用单一版本的 CUDA，而在第二台机器上，CUDA 和 cuDNN 的变量设置不正确。按照 Tensorflow 中提到的说明进行 GPU 支持。
还要检查 NVIDIA 驱动程序计算能力并相应地安装 CUDA。