placeholder和assign速度对比

在CPU上，使用variable和placeholder效果差不多
在GPU上，使用variable要比每次都传placeholder快得多3:2
使用GPU的瓶颈主要在于GPU和内存之间的复制操作

"""
place_holder和variable速度对比
"""
import time

import numpy as np
import tensorflow as tf

M = 4096
N = 4096
K = 4096
A = np.random.random((N, M))
B = np.random.random((M, K))
a = tf.placeholder(dtype=tf.float32, shape=(None, M))
b = tf.placeholder(dtype=tf.float32, shape=(None, N))
c = tf.Variable(initial_value=A, dtype=tf.float32)
pro = a @ b
use_assign = c @ b
with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    beg_time = time.time()
    for i in range(5):
        sess.run(use_assign, feed_dict={
            b: B
        })
    print("use variable", time.time() - beg_time)
    beg_time = time.time()
    for i in range(5):
        sess.run(pro, feed_dict={
            a: A,
            b: B
        })
    print("use placeholder", time.time() - beg_time)

转载于:https://www.cnblogs.com/weiyinfu/p/11264639.html