常量性能的Tensorflow乘法比tf.random慢

Question

我正在使用Tensorflow进行一些非DL计算，并且遇到了我不了解的行为。 我正在tf.matmul(a,a)测试方阵的乘法： tf.matmul(a,a) ：

1. 用tf.constant创建矩阵时
2. 每次运行随机初始化矩阵时

我的期望是，第一种情况应该有一些传输初始数据的开销，即100 MB（使用float32的5000x5000矩阵），但是第二种情况的执行由于每次运行中的随机初始化而会稍微慢一些。

但是，我看到的是，即使在同一会话中连续运行，常数的乘法也要慢得多。

编码

import tensorflow as tf
import numpy as np
from timeit import timeit
import os

os.environ["TF_CPP_MIN_LOG_LEVEL"]="2"  # nospam
SIZE = 5000
NUM_RUNS = 10

a = np.random.random((SIZE, SIZE))
_const_a = tf.constant(a, dtype=tf.float32, name="Const_A")
_mul_const_a = tf.matmul(_const_a, _const_a, name="Mul_Const")

_random_a = tf.random_uniform((SIZE, SIZE), dtype=tf.float32, name="Random_A")
_mul_random_a = tf.matmul(_random_a, _random_a, name="Mul_Random")

with tf.Session(config=tf.ConfigProto(log_device_placement=True)) as s:
    # Run once to make sure everything is initialised
    s.run((_const_a, _mul_const_a, _random_a, _mul_random_a))

    # timeit
    print("TF with const\t", timeit(lambda: s.run((_mul_const_a.op)), number=NUM_RUNS))
    print("TF with random\t", timeit(lambda: s.run((_mul_random_a.op)), number=NUM_RUNS))

输出

Device mapping:
/job:localhost/replica:0/task:0/device:GPU:0 -> device: 0, name: GeForce GTX 1070, pci bus id: 0000:01:00.0, compute capability: 6.1
Random_A/sub: (Sub): /job:localhost/replica:0/task:0/device:GPU:0
Random_A/RandomUniform: (RandomUniform): /job:localhost/replica:0/task:0/device:GPU:0
Random_A/mul: (Mul): /job:localhost/replica:0/task:0/device:GPU:0
Random_A: (Add): /job:localhost/replica:0/task:0/device:GPU:0
Mul_Random: (MatMul): /job:localhost/replica:0/task:0/device:GPU:0
Mul_Const: (MatMul): /job:localhost/replica:0/task:0/device:GPU:0
Random_A/max: (Const): /job:localhost/replica:0/task:0/device:GPU:0
Random_A/min: (Const): /job:localhost/replica:0/task:0/device:GPU:0
Random_A/shape: (Const): /job:localhost/replica:0/task:0/device:GPU:0
Const_A: (Const): /job:localhost/replica:0/task:0/device:GPU:0
TF with const    2.9953213009994215
TF with random   0.513827863998813

Answer 1

YMMV，我在适度的K1100M上获得相反的结果。

Device mapping:
/job:localhost/replica:0/task:0/device:GPU:0 -> device: 0, name: Quadro K1100M, pci bus id: 0000:01:00.0, compute capability: 3.0
Random_A/sub: (Sub): /job:localhost/replica:0/task:0/device:GPU:0
Random_A/RandomUniform: (RandomUniform): /job:localhost/replica:0/task:0/device:GPU:0
Random_A/mul: (Mul): /job:localhost/replica:0/task:0/device:GPU:0
Random_A: (Add): /job:localhost/replica:0/task:0/device:GPU:0
Mul_Random: (MatMul): /job:localhost/replica:0/task:0/device:GPU:0
Mul_Const: (MatMul): /job:localhost/replica:0/task:0/device:GPU:0
Random_A/max: (Const): /job:localhost/replica:0/task:0/device:GPU:0
Random_A/min: (Const): /job:localhost/replica:0/task:0/device:GPU:0
Random_A/shape: (Const): /job:localhost/replica:0/task:0/device:GPU:0
Const_A: (Const): /job:localhost/replica:0/task:0/device:GPU:0
TF with const    4.3167382130868175
TF with random   9.889055849542306

Answer 2

在tensorflow中对session.run（）的第一次调用是不合理的昂贵。 如果要进行基准测试，请记住反复调用它。

不过，在您的情况下，除非禁用常量折叠，否则在常量情况下几乎不会花费任何时间，因为您的图形只会获取常量。