Tensor数据类型
- list: [1,1.2,'hello'] ,内存占用大,处理速度慢
- np.array,数据同类型,但不支持GPU和自动求导
- tf.Tensor,为了弥补numpy的缺点,为深度学习而生,支持GPU与自动求导
- tensor:
- scalar:标量,1.1,dim = 0
- vector:向量,[1.1],[1.1,2.2,...],dim = 1
- matrix: 矩阵,[[1.1,2.2],[3.3,4.4]],dim = 2
- tensor:rank>2 ,dim > 2
- 数据类型:
- Int, float, double
- bool
- string
- 定义tensor
tf.constant(1) # 定义int,普通的tensor
tf.constant(1.) # 定义float
tf.constant([True, False]) # 定义bool
tf.constant('hello nick') # 定义string
属性
with tf.device('cpu'):
a = tf.constant([1]) # 在cpu创建
with tf.device('gpu'):
b = tf.constant([1]) # 在gpu创建
a.device # 设备属性
a.gpu() # cpu转gpu
b.cpu() # gpu转cpu
a.numpy()# 获取numpy数据类型
a.shape # 获取a的属性
a.ndim # 获取维度
tf.rank(a) # 获取维度
a.name # 1.+历史遗留问题
数据类型判断
instance(a,tf.Tensor) # 判断是否为tensor (不推荐)
tf.is_tensor(a) # 判断是否为tensor (推荐)
a.dtype,b.dtype,c.dtype # 判断数据类型
数据类型转换
a = np.arange(5)
aa = tf.convert_to_tensor(a,dtype=tf.int32) # numpy转tensor
tf.cast(aa,dtype=tf.float32) # tensor之间数据类型转换 tf.int32转tf.float32
b = tf.constant([0,1])
tf.cast(b,dtype=tf.bool) # int转bool
# tf.Variable
a = tf.range(5) # 生成一个tensor
b = tf.Variable(a, name='input_data') # tensor转为Variable后具有求导的特性,即自动记录a的梯度相关信息
b.name # input_data:0
b.trainable # True # 可训练
isinstance(b,tf.Tensor) # False
isinstance(b,tf.Variable) # True
tf.is_tensor(b) # True # 推荐使用
tensor转numpy
a= tf.range(5)
a.numpy()
# a必须是scalar
a = tf.ones([])
a.numpy()
int(a)
float(a)
创建Tensor
from numpy, list
* zeros, ones, fill
* random # if big dimension, random initial
* constant
* Application
numpy, list
numpy
import numpy as np
import tensorflow as tf
tf.convert_to_tensor(np.ones([2, 3]))
<tf.Tensor: id=0, shape=(2, 3), dtype=float64, numpy=
array([[1., 1., 1.],
[1., 1., 1.]])>
tf.convert_to_tensor(np.zeros([2, 3]))
<tf.Tensor: id=2, shape=(2, 3), dtype=float64, numpy=
array([[0., 0., 0.],
[0., 0., 0.]])>
list
tf.convert_to_tensor([1, 2])
<tf.Tensor: id=4, shape=(2,), dtype=int32, numpy=array([1, 2], dtype=int32)>
tf.convert_to_tensor([1, 2.])
<tf.Tensor: id=6, shape=(2,), dtype=float32, numpy=array([1., 2.], dtype=float32)>
tf.convert_to_tensor([[1], [2.]])
<tf.Tensor: id=8, shape=(2, 1), dtype=float32, numpy=
array([[1.],
[2.]], dtype=float32)>
zeros, ones, fill
zeros
tf.zeros([])
<tf.Tensor: id=10, shape=(), dtype=float32, numpy=0.0>
tf.zeros([1])
<tf.Tensor: id=14, shape=(1,), dtype=float32, numpy=array([0.], dtype=float32)>
tf.zeros([2, 2])
<tf.Tensor: id=18, shape=(2, 2), dtype=float32, numpy=
array([[0., 0.],
[0., 0.]], dtype=float32)>
tf.zeros([2, 3, 3])
<tf.Tensor: id=22, shape=(2, 3, 3), dtype=float32, numpy=
array([[[0., 0., 0.],
[0., 0., 0.],
[0., 0., 0.]],
[[0., 0., 0.],
[0., 0., 0.],
[0., 0., 0.]]], dtype=float32)>
a = tf.constant([0])
tf.zeros_like(a) # 等同于tf.zeros(a.shape)
<tf.Tensor: id=25, shape=(1,), dtype=int32, numpy=array([0], dtype=int32)>
ones
tf.ones(1)
<tf.Tensor: id=29, shape=(1,), dtype=float32, numpy=array([1.], dtype=float32)>
tf.ones([])
<tf.Tensor: id=31, shape=(), dtype=float32, numpy=1.0>
tf.ones([2])
<tf.Tensor: id=35, shape=(2,), dtype=float32, numpy=array([1., 1.], dtype=float32)>
tf.ones([2, 3])
<tf.Tensor: id=39, shape=(2, 3), dtype=float32, numpy=
array([[1., 1., 1.],
[1., 1., 1.]], dtype=float32)>
a = tf.constant([0])
tf.ones_like(a) # # 等同于tf.ones(a.shape)
<tf.Tensor: id=44, shape=(1,), dtype=int32, numpy=array([1], dtype=int32)>
fill
tf.fill([2, 2], 0)
<tf.Tensor: id=48, shape=(2, 2), dtype=int32, numpy=
array([[0, 0],
[0, 0]], dtype=int32)>
tf.fill([2, 2], 0)
<tf.Tensor: id=52, shape=(2, 2), dtype=int32, numpy=
array([[0, 0],
[0, 0]], dtype=int32)>
tf.fill([2, 2], 1)
<tf.Tensor: id=56, shape=(2, 2), dtype=int32, numpy=
array([[1, 1],
[1, 1]], dtype=int32)>
tf.fill([2, 2], 9)
<tf.Tensor: id=60, shape=(2, 2), dtype=int32, numpy=
array([[9, 9],
[9, 9]], dtype=int32)>
random
# 正态分布,均值为1,方差为1
tf.random.normal([2, 2], mean=1, stddev=1)
<tf.Tensor: id=67, shape=(2, 2), dtype=float32, numpy=
array([[1.0804566, 0.9318387],
[1.0620257, 0.6907253]], dtype=float32)>
tf.random.normal([2, 2])
<tf.Tensor: id=74, shape=(2, 2), dtype=float32, numpy=
array([[-0.06452972, 0.05704789],
[ 0.82857376, 0.71619517]], dtype=float32)>
# 截断的正态分布,
tf.random.truncated_normal([2, 2], mean=0, stddev=1)
<tf.Tensor: id=81, shape=(2, 2), dtype=float32, numpy=
array([[ 0.19161457, -0.820383 ],
[ 0.43668088, -0.3798696 ]], dtype=float32)>
如下图所示为截断正态分布,截掉红色部分,对新的正态分布重新采样。因为sigmoid的和新的正态分布不冲突的地方的区域,对于sigmoid函数来说是近似于平滑的直线,梯度为0,因此会有梯度消失。
# 均匀分布
tf.random.uniform([2, 2], minval=0, maxval=1)
<tf.Tensor: id=89, shape=(2, 2), dtype=float32, numpy=
array([[0.01481438, 0.15071952],
[0.5599004 , 0.59821343]], dtype=float32)>
tf.random.uniform([2, 2], minval=0, maxval=100, dtype=tf.int32)
<tf.Tensor: id=102, shape=(2, 2), dtype=int32, numpy=
array([[51, 9],
[10, 14]], dtype=int32)>
打乱idx后,a和b的索引不变
idx = tf.range(10)
idx = tf.random.shuffle(idx)
idx
<tf.Tensor: id=113, shape=(10,), dtype=int32, numpy=array([0, 8, 4, 9, 6, 7, 5, 2, 1, 3], dtype=int32)>
a = tf.random.normal([10, 784])
b = tf.random.uniform([10], maxval=10, dtype=tf.int32)
b
<tf.Tensor: id=134, shape=(10,), dtype=int32, numpy=array([1, 8, 1, 2, 4, 6, 2, 7, 4, 5], dtype=int32)>
a = tf.gather(a, idx)
b = tf.gather(b, idx)
b
<tf.Tensor: id=147, shape=(10,), dtype=int32, numpy=array([1, 8, 2, 2, 6, 1, 7, 4, 4, 5], dtype=int32)>
constant
tf.constant(1)
<tf.Tensor: id=149, shape=(), dtype=int32, numpy=1>
tf.constant([1])
<tf.Tensor: id=151, shape=(1,), dtype=int32, numpy=array([1], dtype=int32)>
tf.constant([1, 2.])
<tf.Tensor: id=153, shape=(2,), dtype=float32, numpy=array([1., 2.], dtype=float32)>
tf.constant([[1, 2], [3., 2]])
<tf.Tensor: id=156, shape=(2, 2), dtype=float32, numpy=
array([[1., 2.],
[3., 2.]], dtype=float32)>
loss计算
无bias的loss
out = tf.random.uniform([4, 10])
out
<tf.Tensor: id=171, shape=(4, 10), dtype=float32, numpy=
array([[0.67733276, 0.2267617 , 0.21761227, 0.28679788, 0.68864655,
0.21349418, 0.5646602 , 0.8294822 , 0.22094071, 0.20246148],
[0.7940483 , 0.86402774, 0.78399694, 0.80085063, 0.01357341,
0.11889946, 0.89162886, 0.755934 , 0.8058628 , 0.40188062],
[0.115659 , 0.30951428, 0.39866602, 0.5358803 , 0.9163326 ,
0.47557557, 0.9397205 , 0.3110628 , 0.49839914, 0.34321547],
[0.5563061 , 0.78829396, 0.52705276, 0.29077685, 0.35033226,
0.9630101 , 0.338771 , 0.6301584 , 0.7393383 , 0.7073529 ]],
dtype=float32)>
y = tf.range(4)
y = tf.one_hot(y, depth=10)
y
<tf.Tensor: id=188, shape=(4, 10), dtype=float32, numpy=
array([[1., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
[0., 1., 0., 0., 0., 0., 0., 0., 0., 0.],
[0., 0., 1., 0., 0., 0., 0., 0., 0., 0.],
[0., 0., 0., 1., 0., 0., 0., 0., 0., 0.]], dtype=float32)>
loss = tf.keras.losses.mse(y, out)
loss
<tf.Tensor: id=195, shape=(4,), dtype=float32, numpy=array([0.19016443, 0.4096697 , 0.31698173, 0.43206215], dtype=float32)>
loss = tf.reduce_mean(loss)
loss
<tf.Tensor: id=200, shape=(), dtype=float32, numpy=0.3372195>