TensorFlow 数据类型

Reference


list

1
2
[1, 1.2, 'Hello', (1, 2), layers]
[64, 32, 32, 3]

np.array

1
[64, 224, 224, 3]

解决同类型数据运算。

深度计算之前已经设计好的科学计算库,不支持 GPU,不支持自动求导。TensorFlow 解决了这一问题。


tf.Tensor

int, float, double, bool, string

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
a = tf.constant(1)
print(a) # tf.Tensor(1, shape=(), dtype=int32)
a = tf.constant(1.)
print(a) # tf.Tensor(1.0, shape=(), dtype=float32)
a = tf.constant(2.2, dtype=tf.int32)
print(a) # TypeError: Cannot convert provided value to EagerTensor. Provided value: 2.2 Requested dtype: int32
a = tf.constant(2.2, dtype=tf.double)
print(a) # tf.Tensor(2.2, shape=(), dtype=float64)
a = tf.constant([True, False])
print(a) # tf.Tensor([ True False], shape=(2,), dtype=bool)
a = tf.constant('Hello, world!')
print(a) # tf.Tensor(b'Hello, world!', shape=(), dtype=string)

scalar

1
1.1

vector

1
[1.1], [1.1, 1.2]

matrix

1
[[1.1, 2.2], [3.3, 4.4], [5.5, 6.6]]

tensor

rank > 2


Tensor Property

切换 Tensor 到 CPU 或 GPU

1
2
3
4
5
6
7
8
9
10
11
with tf.device("cpu"):
a = tf.constant([1])
with tf.device("gpu"):
b = tf.range(4)
print(a.device) # /job:localhost/replica:0/task:0/device:CPU:0
print(b.device) # /job:localhost/replica:0/task:0/device:GPU:0
aa = a.gpu()
bb = b.cpu()
print(aa.device) # /job:localhost/replica:0/task:0/device:GPU:0
print(bb.device) # /job:localhost/replica:0/task:0/device:CPU:0

Tensor 转 numpy

1
2
3
4
b = tf.range(4)
bb = b.numpy()
print(b) # tf.Tensor([0 1 2 3], shape=(4,), dtype=int32)
print(bb) # [0 1 2 3]

numpy 转 Tensor

1
2
3
4
5
6
7
a = np.arange(5)
a1 = a.dtype
a2 = tf.convert_to_tensor(a)
a3 = tf.convert_to_tensor(a, dtype=tf.int64)
print(a1) # int32
print(a2) # tf.Tensor([0 1 2 3 4], shape=(5,), dtype=int32)
print(a3) # tf.Tensor([0 1 2 3 4], shape=(5,), dtype=int64)

int32 转 float32/double/int64

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
b = tf.range(4)
b1 = tf.cast(b, dtype=tf.float32)
b2 = tf.cast(b, dtype=tf.double)
b3 = tf.cast(b, dtype=tf.int64)
print(b) # tf.Tensor([0 1 2 3], shape=(4,), dtype=int32)
print(b1) # tf.Tensor([0. 1. 2. 3.], shape=(4,), dtype=float32)
print(b2) # tf.Tensor([0. 1. 2. 3.], shape=(4,), dtype=float64)
print(b3) # tf.Tensor([0 1 2 3], shape=(4,), dtype=int64)
a = tf.ones([])
a1 = int(a)
a2 = float(a)
print(a) # tf.Tensor(1.0, shape=(), dtype=float32)
print(a1) # 1
print(a2) # 1.0

int32 和 bool 的相互转换

1
2
3
4
5
b = tf.constant([0, 1])
b1 = tf.cast(b, dtype=tf.bool)
b2 = tf.cast(b1, dtype=tf.int32)
print(b1) # tf.Tensor([False True], shape=(2,), dtype=bool)
print(b2) # tf.Tensor([0 1], shape=(2,), dtype=int32)

Tensor 转 Variable

1
2
3
4
5
6
7
8
9
10
a = tf.range(5)
b = tf.Variable(a)
print(b.trainable) # True
b1 = isinstance(b, tf.Tensor)
b2 = isinstance(b, tf.Variable)
b3 = tf.is_tensor(b)
print(b1) # False
print(b2) # True
print(b3) # True

获取 Tensor 的 rank

1
2
3
4
5
6
a = tf.constant(1.1)
b = tf.constant([1.1])
print(a.ndim) # 0
print(b.ndim) # 1
print(tf.rank(a)) # tf.Tensor(0, shape=(), dtype=int32)
print(tf.rank(b)) # tf.Tensor(1, shape=(), dtype=int32)

判断是否为 Tensor

1
2
3
4
5
6
7
8
9
10
a = tf.constant([1.])
d = np.arange(4)
a1 = isinstance(a, tf.Tensor)
a2 = tf.is_tensor(a)
d1 = isinstance(d, tf.Tensor)
d2 = tf.is_tensor(d)
print(a1) # True
print(a2) # True
print(d1) # False
print(d2) # False

获取 Tensor 的数据类型

1
2
3
4
a = tf.constant([1.])
d = np.arange(4)
print(a.dtype) # <dtype: 'float32'>
print(d.dtype) # int32

判断 Tensor 是否为指定的数据类型

1
2
3
4
a = tf.constant([1.])
d = np.arange(4)
print(a.dtype == tf.float32) # True
print(d.dtype == np.int32) # True

Tensor 初始化

from numpy, list

1
2
3
4
5
6
t1 = tf.convert_to_tensor(np.ones([2, 3]))
t2 = tf.convert_to_tensor([2, 3])
print(t1) # tf.Tensor(
# [[1. 1. 1.]
# [1. 1. 1.]], shape=(2, 3), dtype=float64)
print(t2) # tf.Tensor([2 3], shape=(2,), dtype=int32)

zeros

1
2
3
4
t1 = tf.zeros([2, 3])
print(t1) # tf.Tensor(
# [[0. 0. 0.]
# [0. 0. 0.]], shape=(2, 3), dtype=float32)

ones

1
2
3
4
t1 = tf.ones([2, 3])
print(t1) # tf.Tensor(
# [[1. 1. 1.]
# [1. 1. 1.]], shape=(2, 3), dtype=float32)

fill

1
2
3
4
t1 = tf.fill([2, 3], 3.14)
print(t1) # tf.Tensor(
# [[3.14 3.14 3.14]
# [3.14 3.14 3.14]], shape=(2, 3), dtype=float32)

random

normal, truncated_normal

1
2
3
4
5
6
7
8
9
10
11
12
t1 = tf.random.normal([2, 3], mean=1, stddev=1)
t2 = tf.random.normal([2, 3], mean=1, stddev=1)
t3 = tf.random.truncated_normal([2, 3], mean=1, stddev=1)
print(t1) # tf.Tensor(
# [[1.2308737 0.8250756 2.6446304]
# [2.714213 2.0199542 1.3591613]], shape=(2, 3), dtype=float32)
print(t2) # tf.Tensor(
# [[ 1.5890408 -0.8608984 1.7205659 ]
# [ 1.9205048 0.7741975 0.73170424]], shape=(2, 3), dtype=float32)
print(t3) # tf.Tensor(
# [[ 1.3489898 -0.4739654 0.5768948]
# [ 2.3000512 0.7400613 1.8497746]], shape=(2, 3), dtype=float32)

uniform

1
2
3
4
t1 = tf.random.uniform([2, 3], minval=0, maxval=100)
print(t1) # tf.Tensor(
# [[19.821571 44.24106 29.511225 ]
# [ 0.7015109 7.66474 5.251384 ]], shape=(2, 3), dtype=float32)

shuffle

1
2
3
4
id = tf.range(10)
id1 = tf.random.shuffle(id)
print(id) # tf.Tensor([0 1 2 3 4 5 6 7 8 9], shape=(10,), dtype=int32)
print(id1) # tf.Tensor([8 1 4 0 7 2 5 3 9 6], shape=(10,), dtype=int32)

constant

1
2
t1 = tf.constant(1)
print(t1) # tf.Tensor(1, shape=(), dtype=int32)

range

1
2
3
4
5
6
7
8
t1 = tf.range(4)
t2 = tf.one_hot(t1, 6)
print(t1) # tf.Tensor([0 1 2 3], shape=(4,), dtype=int32)
print(t2) # tf.Tensor(
# [[1. 0. 0. 0. 0. 0.]
# [0. 1. 0. 0. 0. 0.]
# [0. 0. 1. 0. 0. 0.]
# [0. 0. 0. 1. 0. 0.]], shape=(4, 6), dtype=float32)

本文标题:TensorFlow 数据类型

文章作者:魏超

发布时间:2019年05月06日 - 09:05

最后更新:2019年05月20日 - 09:05

原始链接:http://www.weichao.io/2019/05/06/TensorFlow-数据类型/

许可协议: 署名-非商业性使用-禁止演绎 4.0 国际 转载请保留原文链接及作者。

---------------------本文结束---------------------