解決TensorFlow程序無限制占用GPU

今天遇到一個奇怪的現象，使用tensorflow-gpu的時候，出現內存超額~~如果我訓練什么大型數據也就算了，關鍵我就寫了一個y=W*x.......顯示如下圖所示：

程序如下：

import tensorflow as tf

w = tf.Variable([[1.0,2.0]])
b = tf.Variable([[2.],[3.]])

y = tf.multiply(w,b)

init_op = tf.global_variables_initializer()

with tf.Session() as sess:
    sess.run(init_op)
    print(sess.run(y))

出錯提示：

• 占用的內存越來越多，程序崩潰之后，整個電腦都奔潰了，因為整個顯卡全被吃了

2018-06-10 18:28:00.263424: I T:\src\github\tensorflow\tensorflow\core\platform\cpu_feature_guard.cc:140] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2
2018-06-10 18:28:00.598075: I T:\src\github\tensorflow\tensorflow\core\common_runtime\gpu\gpu_device.cc:1356] Found device 0 with properties: 
name: GeForce GTX 1060 major: 6 minor: 1 memoryClockRate(GHz): 1.6705
pciBusID: 0000:01:00.0
totalMemory: 6.00GiB freeMemory: 4.97GiB
2018-06-10 18:28:00.598453: I T:\src\github\tensorflow\tensorflow\core\common_runtime\gpu\gpu_device.cc:1435] Adding visible gpu devices: 0
2018-06-10 18:28:01.265600: I T:\src\github\tensorflow\tensorflow\core\common_runtime\gpu\gpu_device.cc:923] Device interconnect StreamExecutor with strength 1 edge matrix:
2018-06-10 18:28:01.265826: I T:\src\github\tensorflow\tensorflow\core\common_runtime\gpu\gpu_device.cc:929]      0 
2018-06-10 18:28:01.265971: I T:\src\github\tensorflow\tensorflow\core\common_runtime\gpu\gpu_device.cc:942] 0:   N 
2018-06-10 18:28:01.266220: I T:\src\github\tensorflow\tensorflow\core\common_runtime\gpu\gpu_device.cc:1053] Created TensorFlow device (/job:localhost/replica:0/task:0/device:GPU:0 with 4740 MB memory) -> physical GPU (device: 0, name: GeForce GTX 1060, pci bus id: 0000:01:00.0, compute capability: 6.1)
2018-06-10 18:28:01.331056: E T:\src\github\tensorflow\tensorflow\stream_executor\cuda\cuda_driver.cc:936] failed to allocate 4.63G (4970853120 bytes) from device: CUDA_ERROR_OUT_OF_MEMORY
2018-06-10 18:28:01.399111: E T:\src\github\tensorflow\tensorflow\stream_executor\cuda\cuda_driver.cc:936] failed to allocate 4.17G (4473767936 bytes) from device: CUDA_ERROR_OUT_OF_MEMORY
2018-06-10 18:28:01.468293: E T:\src\github\tensorflow\tensorflow\stream_executor\cuda\cuda_driver.cc:936] failed to allocate 3.75G (4026391040 bytes) from device: CUDA_ERROR_OUT_OF_MEMORY
2018-06-10 18:28:01.533138: E T:\src\github\tensorflow\tensorflow\stream_executor\cuda\cuda_driver.cc:936] failed to allocate 3.37G (3623751936 bytes) from device: CUDA_ERROR_OUT_OF_MEMORY
2018-06-10 18:28:01.602452: E T:\src\github\tensorflow\tensorflow\stream_executor\cuda\cuda_driver.cc:936] failed to allocate 3.04G (3261376768 bytes) from device: CUDA_ERROR_OUT_OF_MEMORY
2018-06-10 18:28:01.670225: E T:\src\github\tensorflow\tensorflow\stream_executor\cuda\cuda_driver.cc:936] failed to allocate 2.73G (2935238912 bytes) from device: CUDA_ERROR_OUT_OF_MEMORY
2018-06-10 18:28:01.733120: E T:\src\github\tensorflow\tensorflow\stream_executor\cuda\cuda_driver.cc:936] failed to allocate 2.46G (2641714944 bytes) from device: CUDA_ERROR_OUT_OF_MEMORY
2018-06-10 18:28:01.800101: E T:\src\github\tensorflow\tensorflow\stream_executor\cuda\cuda_driver.cc:936] failed to allocate 2.21G (2377543424 bytes) from device: CUDA_ERROR_OUT_OF_MEMORY
2018-06-10 18:28:01.862064: E T:\src\github\tensorflow\tensorflow\stream_executor\cuda\cuda_driver.cc:936] failed to allocate 1.99G (2139789056 bytes) from device: CUDA_ERROR_OUT_OF_MEMORY
2018-06-10 18:28:01.925434: E T:\src\github\tensorflow\tensorflow\stream_executor\cuda\cuda_driver.cc:936] failed to allocate 1.79G (1925810176 bytes) from device: CUDA_ERROR_OUT_OF_MEMORY
2018-06-10 18:28:01.986180: E T:\src\github\tensorflow\tensorflow\stream_executor\cuda\cuda_driver.cc:936] failed to allocate 1.61G (1733229056 bytes) from device: CUDA_ERROR_OUT_OF_MEMORY
2018-06-10 18:28:02.043456: E T:\src\github\tensorflow\tensorflow\stream_executor\cuda\cuda_driver.cc:936] failed to allocate 1.45G (1559906048 bytes) from device: CUDA_ERROR_OUT_OF_MEMORY
2018-06-10 18:28:02.103531: E T:\src\github\tensorflow\tensorflow\stream_executor\cuda\cuda_driver.cc:936] failed to allocate 1.31G (1403915520 bytes) from device: CUDA_ERROR_OUT_OF_MEMORY
2018-06-10 18:28:02.168973: E T:\src\github\tensorflow\tensorflow\stream_executor\cuda\cuda_driver.cc:936] failed to allocate 1.18G (1263524096 bytes) from device: CUDA_ERROR_OUT_OF_MEMORY
2018-06-10 18:28:02.229387: E T:\src\github\tensorflow\tensorflow\stream_executor\cuda\cuda_driver.cc:936] failed to allocate 1.06G (1137171712 bytes) from device: CUDA_ERROR_OUT_OF_MEMORY
2018-06-10 18:28:02.292997: E T:\src\github\tensorflow\tensorflow\stream_executor\cuda\cuda_driver.cc:936] failed to allocate 976.04M (1023454720 bytes) from device: CUDA_ERROR_OUT_OF_MEMORY
2018-06-10 18:28:02.356714: E T:\src\github\tensorflow\tensorflow\stream_executor\cuda\cuda_driver.cc:936] failed to allocate 878.44M (921109248 bytes) from device: CUDA_ERROR_OUT_OF_MEMORY
2018-06-10 18:28:02.418167: E T:\src\github\tensorflow\tensorflow\stream_executor\cuda\cuda_driver.cc:936] failed to allocate 790.59M (828998400 bytes) from device: CUDA_ERROR_OUT_OF_MEMORY
2018-06-10 18:28:02.482394: E T:\src\github\tensorflow\tensorflow\stream_executor\cuda\cuda_driver.cc:936] failed to allocate 711.54M (746098688 bytes) from device: CUDA_ERROR_OUT_OF_MEMORY

分析原因：

• 1. 顯卡驅動不是最新版本，用__驅動軟件__更新一下驅動，或者自己去下載更新。
• 1. TF運行太多，注銷全部程序沖洗打開。
• 1. 由於TF內核編寫的原因，默認占用全部的GPU去訓練自己的東西，也就是像meiguo一樣優先政策吧

  這個時候我們得設置兩個方面：

  1. 選擇什么樣的占用方式？優先占用__還是__按需占用
  2. 選擇最大占用多少GPU，因為占用過大GPU會導致其它程序奔潰。最好在0.7以下

先更新驅動：

再設置TF程序：

注意：單獨設置一個不行！按照網上大神博客試了，結果效果還是很差（占用很多GPU）

設置TF：

• 按需占用
• 最大占用70%GPU

修改代碼如下：

import tensorflow as tf

w = tf.Variable([[1.0,2.0]])
b = tf.Variable([[2.],[3.]])

y = tf.multiply(w,b)

init_op = tf.global_variables_initializer()

config = tf.ConfigProto(allow_soft_placement=True)
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.7)
config.gpu_options.allow_growth = True
with tf.Session(config=config) as sess:
    sess.run(init_op)
    print(sess.run(y))

成功解決：

2018-06-10 18:21:17.532630: I T:\src\github\tensorflow\tensorflow\core\platform\cpu_feature_guard.cc:140] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2
2018-06-10 18:21:17.852442: I T:\src\github\tensorflow\tensorflow\core\common_runtime\gpu\gpu_device.cc:1356] Found device 0 with properties: 
name: GeForce GTX 1060 major: 6 minor: 1 memoryClockRate(GHz): 1.6705
pciBusID: 0000:01:00.0
totalMemory: 6.00GiB freeMemory: 4.97GiB
2018-06-10 18:21:17.852817: I T:\src\github\tensorflow\tensorflow\core\common_runtime\gpu\gpu_device.cc:1435] Adding visible gpu devices: 0
2018-06-10 18:21:18.511176: I T:\src\github\tensorflow\tensorflow\core\common_runtime\gpu\gpu_device.cc:923] Device interconnect StreamExecutor with strength 1 edge matrix:
2018-06-10 18:21:18.511397: I T:\src\github\tensorflow\tensorflow\core\common_runtime\gpu\gpu_device.cc:929]      0 
2018-06-10 18:21:18.511544: I T:\src\github\tensorflow\tensorflow\core\common_runtime\gpu\gpu_device.cc:942] 0:   N 
2018-06-10 18:21:18.511815: I T:\src\github\tensorflow\tensorflow\core\common_runtime\gpu\gpu_device.cc:1053] Created TensorFlow device (/job:localhost/replica:0/task:0/device:GPU:0 with 4740 MB memory) -> physical GPU (device: 0, name: GeForce GTX 1060, pci bus id: 0000:01:00.0, compute capability: 6.1)
[[2. 4.]
 [3. 6.]]

參考資料：

主要參考博客

錯誤實例