转载请注明出处：
http://www.cnblogs.com/darkknightzh/p/7608916.html
参考网址：
https://stackoverflow.com/questions/39758094/clearing-tensorflow-gpu-memory-after-model-execution
https://github.com/tensorflow/tensorflow/issues/1727#issuecomment-285815312s
tensorflow中，在一个函数内配置完GPU，tf分配了显存，等函数执行完，显存不会释放（貌似torch7中也一样。。。）。第二个参考网址指出：
As for the original problem, currently the Allocator in the GPUDevice belongs to the ProcessState, which is essentially a global singleton. The first session using GPU initializes it, and frees itself when the process shuts down. Even if a second session chooses a different GPUOptions, it would not take effect.
第一个session对GPU初始化后，即便释放了显存，第二个sess使用不同的GPU选项来初始化GPU，也不会起效。
第一个网址Oli Blum指出，use processes and shut them down after the computation才能释放显存。具体代码如下（可以参考第一个网址）：
1 import tensorflow as tf
2 import multiprocessing
3 import numpy as np
4
5 def run_tensorflow():
6
7 n_input = 10000
8 n_classes = 1000
9
10 # Create model
11 def multilayer_perceptron(x, weight):
12 # Hidden layer with RELU activation
13 layer_1 = tf.matmul(x, weight)
14 return layer_1
15
16 # Store layers weight & bias
17 weights = tf.Variable(tf.random_normal([n_input, n_classes]))
18
19
20 x = tf.placeholder("float", [None, n_input])
21 y = tf.placeholder("float", [None, n_classes])
22 pred = multilayer_perceptron(x, weights)
23
24 cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=pred, labels=y))
25 optimizer = tf.train.AdamOptimizer(learning_rate=0.001).minimize(cost)
26
27 init = tf.global_variables_initializer()
28
29 with tf.Session() as sess:
30 sess.run(init)
31
32 for i in range(100):
33 batch_x = np.random.rand(10, 10000)
34 batch_y = np.random.rand(10, 1000)
35 sess.run([optimizer, cost], feed_dict={x: batch_x, y: batch_y})
36
37 print "finished doing stuff with tensorflow!"
38
39
40 if __name__ == "__main__":
41
42 # option 1: execute code with extra process
43 p = multiprocessing.Process(target=run_tensorflow)
44 p.start()
45 p.join()
46
47 # wait until user presses enter key
48 raw_input()
49
50 # option 2: just execute the function
51 run_tensorflow()
52
53 # wait until user presses enter key
54 raw_input()
使用multiprocessing.Process运行run_tensorflow后，显存会自动释放，但是如果直接执行run_tensorflow，显存不会自动释放。当然，该函数计算量较小，如果显卡太好，可能看不到运行multiprocessing.Process后，显存分配、计算并释放的过程，感觉就像没有运行一样。。。