Тензорная нарезка и вставка данных
data-science tensors tensor-slicing tensor-data-insertion tensorflow-tensor-slicing tf.gather_nd-tutorial tensorflow-scatter-insert tensorflow-tensor-operations insert-data-into-tensor

Тензорная нарезка и вставка данных

19 июля 2025 г.

Обзор контента

  • Введение
  • Извлеките тензорные ломтики
  • Вставить данные в тензоры
  • Дальнейшее чтение и ресурсы

При работе над ML-приложениями, такими как обнаружение объектов и NLP, иногда необходимо работать с подразделы (срезы) тензоров. Например, если ваша модель архитектура включает маршрутизацию, где один слой может контролировать, какой пример обучения направляется на следующий уровень. В этом случае вы можете использовать тензорные нарезки OPS, чтобы разделить тензоры и собрать их вместе в правильном порядке.

В приложениях NLP вы можете использовать тензорную нарезку для выполнения маскировки слов во время обучения. Например, вы можете генерировать учебные данные из списка предложений, выбрав индекс слов для маски в каждом предложении, доставая слово в качестве метки, а затем заменив выбранное слово в токен маски.

В этом руководстве вы узнаете, как использовать API Tensorflow для:

  • Извлекать ломтики из тензора
  • Вставить данные в определенных индексах в тензор

Это руководство предполагает знакомство с индексацией тензора. Прочитайте разделы индексацииТензориTensorflow NumpyРуководства перед началом работы с этим руководством.

Настраивать

import tensorflow as tf
import numpy as np


2024-08-15 02:59:22.387959: 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 02:59:22.409014: 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 02:59:22.415391: 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

Извлеките тензорные ломтики

Выполните нарезку для тензора, похожих на Numpy, используяtf.sliceПолем

t1 = tf.constant([0, 1, 2, 3, 4, 5, 6, 7])

print(tf.slice(t1,
               begin=[1],
               size=[3]))


tf.Tensor([1 2 3], shape=(3,), dtype=int32)
WARNING: All log messages before absl::InitializeLog() is called are written to STDERR
I0000 00:00:1723690765.029742  169903 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 00:00:1723690765.033507  169903 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 00:00:1723690765.037154  169903 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 00:00:1723690765.040869  169903 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 00:00:1723690765.052765  169903 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 00:00:1723690765.056359  169903 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 00:00:1723690765.059719  169903 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 00:00:1723690765.063138  169903 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 00:00:1723690765.066532  169903 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 00:00:1723690765.070040  169903 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 00:00:1723690765.073650  169903 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 00:00:1723690765.077235  169903 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 00:00:1723690766.300773  169903 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 00:00:1723690766.302797  169903 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 00:00:1723690766.304766  169903 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 00:00:1723690766.306851  169903 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 00:00:1723690766.308843  169903 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 00:00:1723690766.310723  169903 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 00:00:1723690766.312589  169903 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 00:00:1723690766.314574  169903 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 00:00:1723690766.316423  169903 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 00:00:1723690766.318291  169903 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 00:00:1723690766.320168  169903 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 00:00:1723690766.322150  169903 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 00:00:1723690766.360968  169903 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 00:00:1723690766.362946  169903 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 00:00:1723690766.364849  169903 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 00:00:1723690766.366867  169903 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 00:00:1723690766.368710  169903 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 00:00:1723690766.370591  169903 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 00:00:1723690766.372447  169903 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 00:00:1723690766.374435  169903 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 00:00:1723690766.376302  169903 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 00:00:1723690766.378677  169903 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 00:00:1723690766.380966  169903 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 00:00:1723690766.383365  169903 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

В качестве альтернативы, вы можете использовать более питонический синтаксис. Обратите внимание, что тензорные ломтики равномерно расположены в диапазоне стартовой стоп.

print(t1[1:4])


tf.Tensor([1 2 3], shape=(3,), dtype=int32)


print(t1[-3:])


tf.Tensor([5 6 7], shape=(3,), dtype=int32)

Для двухмерных тензоров вы можете использовать что-то вроде:

t2 = tf.constant([[0, 1, 2, 3, 4],
                  [5, 6, 7, 8, 9],
                  [10, 11, 12, 13, 14],
                  [15, 16, 17, 18, 19]])

print(t2[:-1, 1:3])


tf.Tensor(
[[ 1  2]
 [ 6  7]
 [11 12]], shape=(3, 2), dtype=int32)

Вы можете использоватьtf.sliceНа более высоких тенденциях.

t3 = tf.constant([[[1, 3, 5, 7],
                   [9, 11, 13, 15]],
                  [[17, 19, 21, 23],
                   [25, 27, 29, 31]]
                  ])

print(tf.slice(t3,
               begin=[1, 1, 0],
               size=[1, 1, 2]))


tf.Tensor([[[25 27]]], shape=(1, 1, 2), dtype=int32)

Вы также можете использоватьtf.strided_sliceЧтобы извлечь кусочки тензоров путем «ходатайства» по тензорным размерам.

Использоватьtf.gatherЧтобы извлечь определенные индексы из одной оси тензора.

print(tf.gather(t1,
                indices=[0, 3, 6]))

# This is similar to doing

t1[::3]


tf.Tensor([0 3 6], shape=(3,), dtype=int32)
<tf.Tensor: shape=(3,), dtype=int32, numpy=array([0, 3, 6], dtype=int32)>

tf.gatherне требует, чтобы показатели были равномерно распределены.

alphabet = tf.constant(list('abcdefghijklmnopqrstuvwxyz'))

print(tf.gather(alphabet,
                indices=[2, 0, 19, 18]))


tf.Tensor([b'c' b'a' b't' b's'], shape=(4,), dtype=string)

Чтобы извлечь срезы из нескольких оси тензора, используйтеtf.gather_ndПолем Это полезно, когда вы хотите собрать элементы матрицы, а не только его ряды или столбцы.

t4 = tf.constant([[0, 5],
                  [1, 6],
                  [2, 7],
                  [3, 8],
                  [4, 9]])

print(tf.gather_nd(t4,
                   indices=[[2], [3], [0]]))


tf.Tensor(
[[2 7]
 [3 8]
 [0 5]], shape=(3, 2), dtype=int32)


t5 = np.reshape(np.arange(18), [2, 3, 3])

print(tf.gather_nd(t5,
                   indices=[[0, 0, 0], [1, 2, 1]]))


tf.Tensor([ 0 16], shape=(2,), dtype=int64)


# Return a list of two matrices

print(tf.gather_nd(t5,
                   indices=[[[0, 0], [0, 2]], [[1, 0], [1, 2]]]))


tf.Tensor(
[[[ 0  1  2]
  [ 6  7  8]]

 [[ 9 10 11]
  [15 16 17]]], shape=(2, 2, 3), dtype=int64)


# Return one matrix

print(tf.gather_nd(t5,
                   indices=[[0, 0], [0, 2], [1, 0], [1, 2]]))


tf.Tensor(
[[ 0  1  2]
 [ 6  7  8]
 [ 9 10 11]
 [15 16 17]], shape=(4, 3), dtype=int64)

Вставить данные в тензоры

Использоватьtf.scatter_ndЧтобы вставить данные в определенных срезах/индексах тензора. Обратите внимание, что тензор, в который вы вставляете значения, нулевой инициализирован.

t6 = tf.constant([10])
indices = tf.constant([[1], [3], [5], [7], [9]])
data = tf.constant([2, 4, 6, 8, 10])

print(tf.scatter_nd(indices=indices,
                    updates=data,
                    shape=t6))


tf.Tensor([ 0  2  0  4  0  6  0  8  0 10], shape=(10,), dtype=int32)

Методы, какtf.scatter_ndкоторые требуют нулевых тензоров, аналогичны разреженным тензорным инициализаторам. Вы можете использоватьtf.gather_ndиtf.scatter_ndимитировать поведение редких тензоров.

Рассмотрим пример, в котором вы строите скучный тензор, используя эти два метода в сочетании.

# Gather values from one tensor by specifying indices

new_indices = tf.constant([[0, 2], [2, 1], [3, 3]])
t7 = tf.gather_nd(t2, indices=new_indices)


# Add these values into a new tensor

t8 = tf.scatter_nd(indices=new_indices, updates=t7, shape=tf.constant([4, 5]))

print(t8)


tf.Tensor(
[[ 0  0  2  0  0]
 [ 0  0  0  0  0]
 [ 0 11  0  0  0]
 [ 0  0  0 18  0]], shape=(4, 5), dtype=int32)

Это похоже на:

t9 = tf.SparseTensor(indices=[[0, 2], [2, 1], [3, 3]],
                     values=[2, 11, 18],
                     dense_shape=[4, 5])

print(t9)


SparseTensor(indices=tf.Tensor(
[[0 2]
 [2 1]
 [3 3]], shape=(3, 2), dtype=int64), values=tf.Tensor([ 2 11 18], shape=(3,), dtype=int32), dense_shape=tf.Tensor([4 5], shape=(2,), dtype=int64))


# Convert the sparse tensor into a dense tensor

t10 = tf.sparse.to_dense(t9)

print(t10)


tf.Tensor(
[[ 0  0  2  0  0]
 [ 0  0  0  0  0]
 [ 0 11  0  0  0]
 [ 0  0  0 18  0]], shape=(4, 5), dtype=int32)

Чтобы вставить данные в тензор с ранее существовавшими значениями, используйтеtf.tensor_scatter_nd_addПолем

t11 = tf.constant([[2, 7, 0],
                   [9, 0, 1],
                   [0, 3, 8]])

# Convert the tensor into a magic square by inserting numbers at appropriate indices

t12 = tf.tensor_scatter_nd_add(t11,
                               indices=[[0, 2], [1, 1], [2, 0]],
                               updates=[6, 5, 4])

print(t12)


tf.Tensor(
[[2 7 6]
 [9 5 1]
 [4 3 8]], shape=(3, 3), dtype=int32)

Точно так же использоватьtf.tensor_scatter_nd_subвычесть значения из тензора с ранее существовавшими значениями.

# Convert the tensor into an identity matrix

t13 = tf.tensor_scatter_nd_sub(t11,
                               indices=[[0, 0], [0, 1], [1, 0], [1, 1], [1, 2], [2, 1], [2, 2]],
                               updates=[1, 7, 9, -1, 1, 3, 7])

print(t13)


tf.Tensor(
[[1 0 0]
 [0 1 0]
 [0 0 1]], shape=(3, 3), dtype=int32)

Использоватьtf.tensor_scatter_nd_minскопировать минимальные значения с помощью элементов из одного тензора к другому.

t14 = tf.constant([[-2, -7, 0],
                   [-9, 0, 1],
                   [0, -3, -8]])

t15 = tf.tensor_scatter_nd_min(t14,
                               indices=[[0, 2], [1, 1], [2, 0]],
                               updates=[-6, -5, -4])

print(t15)


tf.Tensor(
[[-2 -7 -6]
 [-9 -5  1]
 [-4 -3 -8]], shape=(3, 3), dtype=int32)

Точно так же использоватьtf.tensor_scatter_nd_maxскопировать максимальные значения с помощью элемента от одного тензора к другому.

t16 = tf.tensor_scatter_nd_max(t14,
                               indices=[[0, 2], [1, 1], [2, 0]],
                               updates=[6, 5, 4])

print(t16)


tf.Tensor(
[[-2 -7  6]
 [-9  5  1]
 [ 4 -3 -8]], shape=(3, 3), dtype=int32)

Дальнейшее чтение и ресурсы

В этом руководстве вы узнали, как использовать тензоры нарезки, доступные с TensorFlow, чтобы обеспечить более тонкий контроль над элементами в ваших тензорах.

  • Проверьте нарезки OPS с TensorFlow Numpy, напримерtf.experimental.numpy.take_along_axisиtf.experimental.numpy.takeПолем
  • Также проверьтеТенсорский гидиРуководство с переменнойПолем

Первоначально опубликовано наTensorflowВеб -сайт, эта статья появляется здесь под новым заголовком и имеет лицензию в CC на 4.0. Образцы кода, разделенные по лицензии Apache 2.0.

Оригинал

Recent Post

Categories

PREVIOUS ARTICLE
NEXT ARTICLE