Contido da visión xeral
- básicos
- Sobre as formas
- Índice
- Manipulación de formas
- Máis sobre DTypes
- emisións
- tf.convert_to para tensor
- Tensores xigantes
- Tensores de liña
- Tensores de aforro
Tensores son arreglos multidimensionais cun tipo uniforme (chamadodtype
Podes ver todos os apoiadosdtypes
aotf.dtypes
.
Se estás familiarizado coaNúmeroOs espíritos son comonp.arrays
.
Todos os tensores son inmutables como números e cadeas de Python: nunca se pode actualizar o contido dun tensor, só se pode crear un novo.
import tensorflow as tf
import numpy as np
2024-08-15 03:05:18.327501: E external/local_xla/xla/stream_executor/cuda/cuda_fft.cc:485] Unable to register cuFFT factory: Attempting to register factory for plugin cuFFT when one has already been registered
2024-08-15 03:05:18.348450: E external/local_xla/xla/stream_executor/cuda/cuda_dnn.cc:8454] Unable to register cuDNN factory: Attempting to register factory for plugin cuDNN when one has already been registered
2024-08-15 03:05:18.354825: E external/local_xla/xla/stream_executor/cuda/cuda_blas.cc:1452] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered
básicos
En primeiro lugar, crear algúns tensores básicos.
Aquí está un tensor "escalar" ou "rank-0".Un escalar contén un único valor, e ningún "eixo".
# This will be an int32 tensor by default; see "dtypes" below.
rank_0_tensor = tf.constant(4)
print(rank_0_tensor)
tf.Tensor(4, shape=(), dtype=int32)
WARNING: All log messages before absl::InitializeLog() is called are written to STDERR
I0000 00:00:1723691120.932442 176945 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723691120.936343 176945 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723691120.940040 176945 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723691120.943264 176945 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723691120.954872 176945 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723691120.958376 176945 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723691120.961894 176945 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723691120.964843 176945 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723691120.967730 176945 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723691120.971300 176945 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723691120.974711 176945 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723691120.977717 176945 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723691122.208679 176945 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723691122.210786 176945 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723691122.212791 176945 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723691122.214776 176945 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723691122.216798 176945 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723691122.218734 176945 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723691122.220650 176945 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723691122.222554 176945 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723691122.224486 176945 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723691122.226429 176945 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723691122.228329 176945 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723691122.230251 176945 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723691122.269036 176945 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723691122.271069 176945 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723691122.273006 176945 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723691122.274956 176945 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723691122.276917 176945 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723691122.278854 176945 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723691122.280754 176945 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723691122.282664 176945 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723691122.284613 176945 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723691122.287058 176945 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723691122.289508 176945 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723691122.291891 176945 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
Un tensor "vector" ou "rank-1" é como unha lista de valores.
# Let's make this a float tensor.
rank_1_tensor = tf.constant([2.0, 3.0, 4.0])
print(rank_1_tensor)
tf.Tensor([2. 3. 4.], shape=(3,), dtype=float32)
A "matrix" or "rank-2" tensor has two axes:
# If you want to be specific, you can set the dtype (see below) at creation time
rank_2_tensor = tf.constant([[1, 2],
[3, 4],
[5, 6]], dtype=tf.float16)
print(rank_2_tensor)
tf.Tensor(
[[1. 2.]
[3. 4.]
[5. 6.]], shape=(3, 2), dtype=float16)
A scalar, shape: |
A vector, shape: |
A matrix, shape: |
---|---|---|
|
|
|
Tensors may have more axes; here is a tensor with three axes:
# There can be an arbitrary number of
# axes (sometimes called "dimensions")
rank_3_tensor = tf.constant([
[[0, 1, 2, 3, 4],
[5, 6, 7, 8, 9]],
[[10, 11, 12, 13, 14],
[15, 16, 17, 18, 19]],
[[20, 21, 22, 23, 24],
[25, 26, 27, 28, 29]],])
print(rank_3_tensor)
tf.Tensor(
[[[ 0 1 2 3 4]
[ 5 6 7 8 9]]
[[10 11 12 13 14]
[15 16 17 18 19]]
[[20 21 22 23 24]
[25 26 27 28 29]]], shape=(3, 2, 5), dtype=int32)
Hai moitas formas de visualizar un tensor con máis de dous eixos.
A 3-axis tensor, shape: |
|
|
---|---|---|
|
|
|
Pode converter un tensor a unha matriz NumPy usandonp.array
ou otensor.numpy
O método:
np.array(rank_2_tensor)
array([[1., 2.],
[3., 4.],
[5., 6.]], dtype=float16)
rank_2_tensor.numpy()
array([[1., 2.],
[3., 4.],
[5., 6.]], dtype=float16)
Os sensores a miúdo conteñen flotadores e incs, pero teñen moitos outros tipos, incluíndo:
- Números complexos
- Strings
A basetf.Tensor
A clase require que os tensores sexan "retangulares"---é dicir, ao longo de cada eixe, cada elemento é do mesmo tamaño.
- Tensores Rápidos
- Tensores de aforro
Podes facer matemáticas básicas sobre tensores, incluíndo adición, multiplicación elemental e multiplicación de matriz.
a = tf.constant([[1, 2],
[3, 4]])
b = tf.constant([[1, 1],
[1, 1]]) # Could have also said `tf.ones([2,2], dtype=tf.int32)`
print(tf.add(a, b), "\n")
print(tf.multiply(a, b), "\n")
print(tf.matmul(a, b), "\n")
tf.Tensor(
[[2 3]
[4 5]], shape=(2, 2), dtype=int32)
tf.Tensor(
[[1 2]
[3 4]], shape=(2, 2), dtype=int32)
tf.Tensor(
[[3 3]
[7 7]], shape=(2, 2), dtype=int32)
print(a + b, "\n") # element-wise addition
print(a * b, "\n") # element-wise multiplication
print(a @ b, "\n") # matrix multiplication
tf.Tensor(
[[2 3]
[4 5]], shape=(2, 2), dtype=int32)
tf.Tensor(
[[1 2]
[3 4]], shape=(2, 2), dtype=int32)
tf.Tensor(
[[3 3]
[7 7]], shape=(2, 2), dtype=int32)
Os sensores utilízanse en todo tipo de operacións (ou "Ops").
c = tf.constant([[4.0, 5.0], [10.0, 1.0]])
# Find the largest value
print(tf.reduce_max(c))
# Find the index of the largest value
print(tf.math.argmax(c))
# Compute the softmax
print(tf.nn.softmax(c))
tf.Tensor(10.0, shape=(), dtype=float32)
tf.Tensor([1 0], shape=(2,), dtype=int64)
tf.Tensor(
[[2.6894143e-01 7.3105854e-01]
[9.9987662e-01 1.2339458e-04]], shape=(2, 2), dtype=float32)
Nota: Normalmente, onde unha función TensorFlow espera un Tensor como entrada, a función tamén aceptará calquera cousa que poida ser convertida a un Tensor usando tf.convert_to_tensor.
Note:Normalmente, en calquera lugar onde unha función TensorFlow espera unhaTensor
como entrada, a función tamén aceptará calquera cousa que poida converterse enTensor
Utilizacióntf.convert_to_tensor
Vexa abaixo para un exemplo.
tf.convert_to_tensor([1,2,3])
<tf.Tensor: shape=(3,), dtype=int32, numpy=array([1, 2, 3], dtype=int32)>
tf.reduce_max([1,2,3])
<tf.Tensor: shape=(), dtype=int32, numpy=3>
tf.reduce_max(np.array([1,2,3]))
<tf.Tensor: shape=(), dtype=int64, numpy=3>
Sobre as formas
Algúns textos teñen vocabulario:
- Forma: A lonxitude (número de elementos) de cada un dos eixos dun tensor.
- Un escalar ten o rango 0, un vector ten o rango 1, unha matriz ten o rango 2.
- Eixo ou dimensión: unha dimensión particular dun tensor.
- Tamaño: O número total de elementos no tensor, o produto dos elementos do vector de forma.
Nota: Aínda que pode ver unha referencia a un "tensor de dúas dimensións", un tensor de posición 2 normalmente non describe un espazo 2D.
Note:Aínda que se pode ver a referencia a un "tensor de dúas dimensións", un tensor de posición 2 normalmente non describe un espazo 2D.
Tensores etf.TensorShape
Os obxectos teñen propiedades convenientes para acceder a estes:
rank_4_tensor = tf.zeros([3, 2, 4, 5])
|
A rank-4 tensor, shape: |
---|---|
|
|
print("Type of every element:", rank_4_tensor.dtype)
print("Number of axes:", rank_4_tensor.ndim)
print("Shape of tensor:", rank_4_tensor.shape)
print("Elements along axis 0 of tensor:", rank_4_tensor.shape[0])
print("Elements along the last axis of tensor:", rank_4_tensor.shape[-1])
print("Total number of elements (3*2*4*5): ", tf.size(rank_4_tensor).numpy())
Type of every element: <dtype: 'float32'>
Number of axes: 4
Shape of tensor: (3, 2, 4, 5)
Elements along axis 0 of tensor: 3
Elements along the last axis of tensor: 5
Total number of elements (3*2*4*5): 120
Pero teña en conta que oTensor.ndim
eTensor.shape
Os atributos non regresanTensor
obxectos. se precisas unhaTensor
Utiliza otf.rank
outf.shape
Esta diferenza é sutil, pero pode ser importante ao construír gráficos (máis tarde).
tf.rank(rank_4_tensor)
<tf.Tensor: shape=(), dtype=int32, numpy=4>
tf.shape(rank_4_tensor)
<tf.Tensor: shape=(4,), dtype=int32, numpy=array([3, 2, 4, 5], dtype=int32)>
Moitas veces os eixos están ordenados de global a local: o eixo de lote primeiro, seguido por dimensións espaciais, e características para cada lugar por último.
Typical axis order |
---|
|
Índice
Indicadores de eixo único
TensorFlow segue as regras de indexación estándar de Python, similares aIndexar unha lista ou cadea en Python, e as regras básicas para a indexación NumPy.
- Os índices comezan en 0
- os índices negativos contan cara atrás desde o final
- As seguintes páxinas ligan con Start:Stop:Step
rank_1_tensor = tf.constant([0, 1, 1, 2, 3, 5, 8, 13, 21, 34])
print(rank_1_tensor.numpy())
[ 0 1 1 2 3 5 8 13 21 34]
A indexación cun escalar elimina o eixe:
print("First:", rank_1_tensor[0].numpy())
print("Second:", rank_1_tensor[1].numpy())
print("Last:", rank_1_tensor[-1].numpy())
First: 0
Second: 1
Last: 34
Índice a:
A liña mantén o eixe:
print("Everything:", rank_1_tensor[:].numpy())
print("Before 4:", rank_1_tensor[:4].numpy())
print("From 4 to the end:", rank_1_tensor[4:].numpy())
print("From 2, before 7:", rank_1_tensor[2:7].numpy())
print("Every other item:", rank_1_tensor[::2].numpy())
print("Reversed:", rank_1_tensor[::-1].numpy())
Everything: [ 0 1 1 2 3 5 8 13 21 34]
Before 4: [0 1 1 2]
From 4 to the end: [ 3 5 8 13 21 34]
From 2, before 7: [1 2 3 5 8]
Every other item: [ 0 1 3 8 21]
Reversed: [34 21 13 8 5 3 2 1 1 0]
Indicacións multiaxis
Os tensores de maior rango son indexados ao pasar varios índices.
As mesmas regras que no caso de un único eixe aplícanse a cada eixe de forma independente.
print(rank_2_tensor.numpy())
[[1. 2.]
[3. 4.]
[5. 6.]]
Pasando un enteiro por cada índice, o resultado é un escalar.
# Pull out a single value from a 2-rank tensor
print(rank_2_tensor[1, 1].numpy())
You can index using any combination of integers and slices:
# Get row and column tensors
print("Second row:", rank_2_tensor[1, :].numpy())
print("Second column:", rank_2_tensor[:, 1].numpy())
print("Last row:", rank_2_tensor[-1, :].numpy())
print("First item in last column:", rank_2_tensor[0, -1].numpy())
print("Skip the first row:")
print(rank_2_tensor[1:, :].numpy(), "\n")
Second row: [3. 4.]
Second column: [2. 4. 6.]
Last row: [5. 6.]
First item in last column: 2.0
Skip the first row:
[[3. 4.]
[5. 6.]]
Here is an example with a 3-axis tensor:
print(rank_3_tensor[:, :, 4])
tf.Tensor(
[[ 4 9]
[14 19]
[24 29]], shape=(3, 2), dtype=int32)
|
Selecting the last feature across all locations in each example in the batch |
---|---|
|
|
Ler oGuía de cortes de tensiónpara aprender como aplicar a indexación para manipular elementos individuais nos seus tensores.
Manipulación de formas
Reforzar un tensor é de gran utilidade.
# Shape returns a `TensorShape` object that shows the size along each axis
x = tf.constant([[1], [2], [3]])
print(x.shape)
(3, 1)
# You can convert this object into a Python list, too
print(x.shape.as_list())
[3, 1]
Pódese transformar un tecido nunha nova forma: otf.reshape
A operación é rápida e barata xa que os datos subxacentes non necesitan ser duplicados.
# You can reshape a tensor to a new shape.
# Note that you're passing in a list
reshaped = tf.reshape(x, [1, 3])
print(x.shape)
print(reshaped.shape)
(3, 1)
(1, 3)
Os datos manteñen o seu deseño na memoria e créase un novo tensor, coa forma solicitada, apuntando aos mesmos datos.TensorFlow utiliza a orde de memoria "row-major" de estilo C, onde incrementar o índice máis dereito corresponde a un único paso na memoria.
print(rank_3_tensor)
tf.Tensor(
[[[ 0 1 2 3 4]
[ 5 6 7 8 9]]
[[10 11 12 13 14]
[15 16 17 18 19]]
[[20 21 22 23 24]
[25 26 27 28 29]]], shape=(3, 2, 5), dtype=int32)
Se planeas un tensor podes ver en que orde está disposto na memoria.
# A `-1` passed in the `shape` argument says "Whatever fits".
print(tf.reshape(rank_3_tensor, [-1]))
tf.Tensor(
[ 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
24 25 26 27 28 29], shape=(30,), dtype=int32)
O único uso razoable datf.reshape
é para combinar ou dividir eixos adxacentes (ou engadir / eliminar1
c) A súa
Para este tensor 3x2x5, a remodelación a (3x2)x5 ou 3x(2x5) son dúas cousas razoables que facer, xa que as fendas non se mesturan:
print(tf.reshape(rank_3_tensor, [3*2, 5]), "\n")
print(tf.reshape(rank_3_tensor, [3, -1]))
tf.Tensor(
[[ 0 1 2 3 4]
[ 5 6 7 8 9]
[10 11 12 13 14]
[15 16 17 18 19]
[20 21 22 23 24]
[25 26 27 28 29]], shape=(6, 5), dtype=int32)
tf.Tensor(
[[ 0 1 2 3 4 5 6 7 8 9]
[10 11 12 13 14 15 16 17 18 19]
[20 21 22 23 24 25 26 27 28 29]], shape=(3, 10), dtype=int32)
|
|
Some good reshapes. |
---|---|---|
|
|
|
A remodelación "traballará" para calquera nova forma co mesmo número total de elementos, pero non fará nada útil se non respectas a orde dos eixos.
Cambio de axentes entf.reshape
non funciona; ten quetf.transpose
para iso.
# Bad examples: don't do this
# You can't reorder axes with reshape.
print(tf.reshape(rank_3_tensor, [2, 3, 5]), "\n")
# This is a mess
print(tf.reshape(rank_3_tensor, [5, 6]), "\n")
# This doesn't work at all
try:
tf.reshape(rank_3_tensor, [7, -1])
except Exception as e:
print(f"{type(e).__name__}: {e}")
tf.Tensor(
[[[ 0 1 2 3 4]
[ 5 6 7 8 9]
[10 11 12 13 14]]
[[15 16 17 18 19]
[20 21 22 23 24]
[25 26 27 28 29]]], shape=(2, 3, 5), dtype=int32)
tf.Tensor(
[[ 0 1 2 3 4 5]
[ 6 7 8 9 10 11]
[12 13 14 15 16 17]
[18 19 20 21 22 23]
[24 25 26 27 28 29]], shape=(5, 6), dtype=int32)
InvalidArgumentError: { {function_node __wrapped__Reshape_device_/job:localhost/replica:0/task:0/device:GPU:0} } Input to reshape is a tensor with 30 values, but the requested shape requires a multiple of 7 [Op:Reshape]
|
|
Some bad reshapes. |
---|---|---|
|
|
|
Pode executar sobre formas non especificadas. Ou a forma contén unhaNone
(a lonxitude dun eixe é descoñecida) ou a forma enteira éNone
(O rango do tensor é descoñecido).
Excepto para tf.RaggedTensor, tales formas só ocorrerán no contexto das APIs simbólicas de construción de gráficos de TensorFlow:
- F. Funcións
- O API funcional duro.
Máis enDTypes
Dtítulos
Para comprobar atf.Tensor
Os tipos de datos utilizan oTensor.dtype
Propiedade .
Ao crear unhatf.Tensor
desde un obxecto Python pode especificar opcionalmente o tipo de datos.
Se non, TensorFlow elixe un tipo de datos que pode representar os seus datos.tf.int32
e Python os números de puntos flotantes paratf.float32
Se non, TensorFlow usa as mesmas regras que NumPy utiliza ao converter en array.
Pódese facer de tipo a tipo.
the_f64_tensor = tf.constant([2.2, 3.3, 4.4], dtype=tf.float64)
the_f16_tensor = tf.cast(the_f64_tensor, dtype=tf.float16)
# Now, cast to an uint8 and lose the decimal precision
the_u8_tensor = tf.cast(the_f16_tensor, dtype=tf.uint8)
print(the_u8_tensor)
tf.Tensor([2 3 4], shape=(3,), dtype=uint8)
emisións
A emisión é un concepto emprestado daFuncións equivalentes en NumPyEn resumo, baixo certas condicións, os tensores máis pequenos son "estendidos" automaticamente para encaixar os tensores máis grandes cando se executan operacións combinadas neles.
O caso máis sinxelo e máis común é cando intenta multiplicar ou engadir un tensor a un escalar.
x = tf.constant([1, 2, 3])
y = tf.constant(2)
z = tf.constant([2, 2, 2])
# All of these are the same computation
print(tf.multiply(x, 2))
print(x * y)
print(x * z)
tf.Tensor([2 4 6], shape=(3,), dtype=int32)
tf.Tensor([2 4 6], shape=(3,), dtype=int32)
tf.Tensor([2 4 6], shape=(3,), dtype=int32)
Do mesmo xeito, os eixos de lonxitude 1 poden estirarse para coincidir cos outros argumentos.
Neste caso, unha matriz 3x1 é elementalmente multiplicada por unha matriz 1x4 para producir unha matriz 3x4.[4]
.
# These are the same computations
x = tf.reshape(x,[3,1])
y = tf.range(1, 5)
print(x, "\n")
print(y, "\n")
print(tf.multiply(x, y))
tf.Tensor(
[[1]
[2]
[3]], shape=(3, 1), dtype=int32)
tf.Tensor([1 2 3 4], shape=(4,), dtype=int32)
tf.Tensor(
[[ 1 2 3 4]
[ 2 4 6 8]
[ 3 6 9 12]], shape=(3, 4), dtype=int32)
A broadcasted add: a |
---|
|
Aquí está a mesma operación sen transmisión:
x_stretch = tf.constant([[1, 1, 1, 1],
[2, 2, 2, 2],
[3, 3, 3, 3]])
y_stretch = tf.constant([[1, 2, 3, 4],
[1, 2, 3, 4],
[1, 2, 3, 4]])
print(x_stretch * y_stretch) # Again, operator overloading
tf.Tensor(
[[ 1 2 3 4]
[ 2 4 6 8]
[ 3 6 9 12]], shape=(3, 4), dtype=int32)
A maior parte do tempo, a transmisión é eficiente tanto no tempo como no espazo, xa que a operación de transmisión nunca materializa os tensores expandidos na memoria.
Vexa como funciona a emisióntf.broadcast_to
.
print(tf.broadcast_to(tf.constant([1, 2, 3]), [3, 3]))
tf.Tensor(
[[1 2 3]
[1 2 3]
[1 2 3]], shape=(3, 3), dtype=int32)
A diferenza das matemáticas, por exemplo,broadcast_to
Non fai nada especial para salvar a memoria. aquí, estás materializando o tensor.
Pode ser aínda máis complicado.Esta secciónDescargar cancións de Jake VanderPlasPython Manual de Ciencia de Datosmostra máis trucos de transmisión (de novo en NumPy).
tf.convert_to para tensor
A maioría das opcións, comotf.matmul
etf.reshape
Argumentos de clasetf.Tensor
Non obstante, notarás no caso anterior que se aceptan obxectos de Python con forma de tensores.
A maioría, pero non todos, chamadas de opsconvert_to_tensor
Hai un rexistro de conversións, e a maioría das clases de obxectos como a de NumPyndarray
,TensorShape
Listas de Python, etf.Variable
Todos se converterán automaticamente.
quetf.register_tensor_conversion_function
para máis detalles, e se ten o seu propio tipo que quere converter automaticamente a un tensor.
Tensores xigantes
Un tensor con números variables de elementos ao longo dalgún eixe chámase "ragged".tf.ragged.RaggedTensor
para os datos recollidos.
Por exemplo, Isto non pode ser representado como un tensor regular:
A |
---|
|
ragged_list = [
[0, 1, 2, 3],
[4, 5],
[6, 7, 8],
[9]]
try:
tensor = tf.constant(ragged_list)
except Exception as e:
print(f"{type(e).__name__}: {e}")
ValueError: Can't convert non-rectangular Python sequence to Tensor.
En vez de crear unhatf.RaggedTensor
Utilizacióntf.ragged.constant
:
ragged_tensor = tf.ragged.constant(ragged_list)
print(ragged_tensor)
<tf.RaggedTensor [[0, 1, 2, 3], [4, 5], [6, 7, 8], [9]]>
A forma dunhatf.RaggedTensor
contén algúns eixos con lonxitudes descoñecidas:
print(ragged_tensor.shape)
(4, None)
Tensores de liña
tf.string
é adtype
, é dicir, pode representar datos como cadeas (arados de bytes de lonxitude variable) en tensores.
As cadeas son atómicas e non poden ser indexadas como as cadeas de Python. A lonxitude da cadea non é un dos eixos do tensor.tf.strings
para manipular as súas funcións.
Aquí está un tensor de cadea escalar:
# Tensors can be strings, too here is a scalar string.
scalar_string_tensor = tf.constant("Gray wolf")
print(scalar_string_tensor)
tf.Tensor(b'Gray wolf', shape=(), dtype=string)
Un vector de cadeas:
A vector of strings, shape: |
---|
|
# If you have three string tensors of different lengths, this is OK.
tensor_of_strings = tf.constant(["Gray wolf",
"Quick brown fox",
"Lazy dog"])
# Note that the shape is (3,). The string length is not included.
print(tensor_of_strings)
tf.Tensor([b'Gray wolf' b'Quick brown fox' b'Lazy dog'], shape=(3,), dtype=string)
En primeiro lugar, a impresión dob
Prefixo indica quetf.string
dtype non é unha cadea de unicode, senón unha cadea de byte.Unicode TutoriaisPara obter máis información sobre como traballar con texto de unicode en TensorFlow.
Se pasas os caracteres de unicode son UTF-8 codificados.
tf.constant("🥳👍")
<tf.Tensor: shape=(), dtype=string, numpy=b'\xf0\x9f\xa5\xb3\xf0\x9f\x91\x8d'>
Algunhas funcións básicas con cordas pódense atopar entf.strings
incluíndotf.strings.split
.
# You can use split to split a string into a set of tensors
print(tf.strings.split(scalar_string_tensor, sep=" "))
tf.Tensor([b'Gray' b'wolf'], shape=(2,), dtype=string)
# ...but it turns into a `RaggedTensor` if you split up a tensor of strings,
# as each string might be split into a different number of parts.
print(tf.strings.split(tensor_of_strings))
<tf.RaggedTensor [[b'Gray', b'wolf'], [b'Quick', b'brown', b'fox'], [b'Lazy', b'dog']]>
Three strings split, shape: |
---|
|
etf.strings.to_number
:
text = tf.constant("1 10 100")
print(tf.strings.to_number(tf.strings.split(text, " ")))
tf.Tensor([ 1. 10. 100.], shape=(3,), dtype=float32)
Aínda que non podes usartf.cast
Para converter un tensor de cadea en números, pode convertelo en bytes e despois en números.
byte_strings = tf.strings.bytes_split(tf.constant("Duck"))
byte_ints = tf.io.decode_raw(tf.constant("Duck"), tf.uint8)
print("Byte strings:", byte_strings)
print("Bytes:", byte_ints)
Byte strings: tf.Tensor([b'D' b'u' b'c' b'k'], shape=(4,), dtype=string)
Bytes: tf.Tensor([ 68 117 99 107], shape=(4,), dtype=uint8)
# Or split it up as unicode and then decode it
unicode_bytes = tf.constant("アヒル 🦆")
unicode_char_bytes = tf.strings.unicode_split(unicode_bytes, "UTF-8")
unicode_values = tf.strings.unicode_decode(unicode_bytes, "UTF-8")
print("\nUnicode bytes:", unicode_bytes)
print("\nUnicode chars:", unicode_char_bytes)
print("\nUnicode values:", unicode_values)
Unicode bytes: tf.Tensor(b'\xe3\x82\xa2\xe3\x83\x92\xe3\x83\xab \xf0\x9f\xa6\x86', shape=(), dtype=string)
Unicode chars: tf.Tensor([b'\xe3\x82\xa2' b'\xe3\x83\x92' b'\xe3\x83\xab' b' ' b'\xf0\x9f\xa6\x86'], shape=(5,), dtype=string)
Unicode values: tf.Tensor([ 12450 12498 12523 32 129414], shape=(5,), dtype=int32)
A súatf.string
dtype utilízase para todos os datos de bytes en bruto en TensorFlow.tf.io
O módulo contén funcións para converter datos a e de bytes, incluíndo a decodificación de imaxes e a análise de csv.
Tensores de aforro
Ás veces, os seus datos son escasos, como un espazo de embalaxe moi amplo.tf.sparse.SparseTensor
e operacións relacionadas para almacenar datos escasos de forma eficiente.
A |
---|
|
# Sparse tensors store values by index in a memory-efficient manner
sparse_tensor = tf.sparse.SparseTensor(indices=[[0, 0], [1, 2]],
values=[1, 2],
dense_shape=[3, 4])
print(sparse_tensor, "\n")
# You can convert sparse tensors to dense
print(tf.sparse.to_dense(sparse_tensor))
SparseTensor(indices=tf.Tensor(
[[0 0]
[1 2]], shape=(2, 2), dtype=int64), values=tf.Tensor([1 2], shape=(2,), dtype=int32), dense_shape=tf.Tensor([3 4], shape=(2,), dtype=int64))
tf.Tensor(
[[1 0 0 0]
[0 0 2 0]
[0 0 0 0]], shape=(3, 4), dtype=int32)
Publicado orixinalmente no sitio web de TensorFlow, este artigo aparece aquí baixo un novo título e está licenciado baixo CC BY 4.0.
Publicado orixinalmente no sitio web de TensorFlow, este artigo aparece aquí baixo un novo título e está licenciado baixo CC BY 4.0.
Páxina web de TensorFlow