tf.session.run()单函数运行和多函数运行区别
觉得有用的话,欢迎一起讨论相互学习~
problem introduction
sess.run([a,b]) # (1)同时运行a,b两个函数
sess.run(a)
sess.run(b) # (2)运行完a函数后再运行b函数
- 这两个语句初看时没有任何区别,但是如果a,b函数恰好是读取example_batch和label_batch这种需要使用到 数据批次输入输出函数时 例如(tf.train.shuffle_batch.tf.reader.read).
- (1)式只会调用一次输入数据函数,则得到的example_batch和label_batch来自同一批次。 (2)式会单独调用两次输入数据函数,则得到的example_batch来自上一批次而label_batch来自下一批次。
- 这个需要十分注意,因为如果我们想要实时打印出label_batch和inference(example_batch)时,即将输入数据的标签和经过模型预测推断的结果进行比较时.如果我们使用(2)中的写法,则label_batch和inference(example_batch)并不是来自与同一批次数据。
example code
这里我们分别使用两种不同的代码,读取csv文件中的数据。我们观察这两种方式读取的数据有什么不同。
源程序文件下载
test_tf_train_batch.csv
import tensorflow as tf
BATCH_SIZE = 400
NUM_THREADS = 2
MAX_NUM = 5
def read_data(file_queue):
reader = tf.TextLineReader(skip_header_lines=1)
key, value = reader.read(file_queue)
defaults = [[0], [0.], [0.]]
NUM, C, Tensile = tf.decode_csv(value, defaults)
vertor_example = tf.stack([C])
vertor_label = tf.stack([Tensile])
vertor_num = tf.stack([NUM])
return vertor_example, vertor_label, vertor_num
def create_pipeline(filename, batch_size, num_threads):
file_queue = tf.train.string_input_producer([filename]) # 设置文件名队列
example, label, no = read_data(file_queue) # 读取数据和标签
example_batch, label_batch, no_batch = tf.train.batch(
[example, label, no], batch_size=batch_size, num_threads=num_threads, capacity=MAX_NUM)
return example_batch, label_batch, no_batch
x_train_batch, y_train_batch, no_train_batch = create_pipeline('test_tf_train_batch.csv', batch_size=BATCH_SIZE,
num_threads=NUM_THREADS)
init_op = tf.global_variables_initializer()
local_init_op = tf.local_variables_initializer()
with tf.Session() as sess:
sess.run(local_init_op)
sess.run(init_op)
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(coord=coord)
# 同时运行的方式
example, label, num = sess.run([x_train_batch, y_train_batch, no_train_batch])
print('The first mode to load data')
print('example', example)
print('label', label)
print('num', num)
# 分别运行的方式
# example = sess.run(x_train_batch)
# label = sess.run(y_train_batch)
# num = sess.run(no_train_batch)
# print('The second mode to load data')
# print('example', example)
# print('label', label)
# print('num', num)
coord.request_stop()
coord.join(threads)
Result
Run at the same time
example, label, num = sess.run([x_train_batch, y_train_batch, no_train_batch])
print('The first mode to load data')
print('example', example)
print('label', label)
print('num', num)
example |
label |
num |
[ 0.294 ] |
[ 0.59821427] |
[1] |
[ 0.31 ] |
[ 0.51785713] |
[2] |
[ 0.2 ] |
[ 0.79464287] |
[3] |
[ 0.30000001] |
[ 0.4732143 ] |
[4] |
[ 0.36000001] |
[ 0.6964286 ] |
[5] |
Run respectively
example = sess.run(x_train_batch)
label = sess.run(y_train_batch)
num = sess.run(no_train_batch)
print('The second mode to load data')
print('example
', example)
print('label
', label)
print('num
', num)
经过对比原始数据,我们发现采用单独运行的方式读取的example来自第一个batch,label来自下一个batch,而num来自第三个batch.也就是说其实我们单独运行了三次文件输入的程序。虽然是个小事,但是有些方面不注意,我们会酿成大错
example |
label |
num |
[ 0.294 ] |
[ 0.5625 ] |
[11] |
[ 0.31 ] |
[ 0.3482143 ] |
[13] |
[ 0.2 ] |
[ 0.5535714 ] |
[12] |
[ 0.30000001] |
[ 0.5714286 ] |
[14] |
[ 0.36000001] |
[ 0.48214287] |
[15] |
C |
tensile |
NUM |
0.294 |
0.598214286 |
1 |
0.31 |
0.517857143 |
2 |
0.2 |
0.794642857 |
3 |
0.3 |
0.473214286 |
4 |
0.36 |
0.696428571 |
5 |
0.28 |
0.5625 |
6 |
0.2 |
0.348214286 |
7 |
0.284 |
0.553571429 |
8 |
0.38 |
0.482142857 |
9 |
0.44 |
0.571428571 |
10 |
0.214 |
0.660714286 |
11 |
0.72 |
0.589285714 |
12 |
0.38 |
0.616071429 |
13 |
0.266 |
0.5 |
14 |
0.46 |
0.642857143 |
15 |