如何将[N]维度张量转换为[N,D]维度矩阵（每列为原向量）？

Got it, let's walk through how to convert your [N]-dimensional tensor into an [N,D] matrix where every column is identical to the original vector. I'll cover the most common frameworks since implementation varies a bit depending on what you're using:

NumPy #

If you're working with NumPy arrays, you've got a couple straightforward options:

• Using np.tile(): First expand your original tensor to a [N,1] shape, then tile it across the column dimension:

  import numpy as np
  
  original_tensor = np.array([1, 2, 3])  # Shape (3,)
  D = 4
  # Expand to (3,1) then tile 1 time along rows, D times along columns
  result = np.tile(original_tensor[:, np.newaxis], (1, D))  # Shape (3,4)
  

• Using broadcasting (memory-efficient): Broadcasting avoids explicit memory duplication, which is great if D is large:

  result = original_tensor[:, np.newaxis] * np.ones((1, D))
  # Or using np.expand_dims + repeat
  result = np.expand_dims(original_tensor, axis=1).repeat(D, axis=1)
  

PyTorch #

For PyTorch tensors, the approach is similar with framework-specific methods:

• Using repeat():

  import torch
  
  original_tensor = torch.tensor([1, 2, 3])  # Shape (3,)
  D = 4
  # Unsqueeze to add column dimension, then repeat
  result = original_tensor.unsqueeze(1).repeat(1, D)  # Shape (3,4)
  

• Using torch.tile() (available in PyTorch 1.10+):

  result = torch.tile(original_tensor.unsqueeze(1), (1, D))
  

• Broadcasting trick: This creates a view without copying memory (until you modify it):

  result = original_tensor.unsqueeze(1) + torch.zeros(1, D, device=original_tensor.device)
  

TensorFlow #

In TensorFlow, you can use tf.tile or tf.repeat to achieve this:

• Using tf.tile():

  import tensorflow as tf
  
  original_tensor = tf.constant([1, 2, 3])  # Shape (3,)
  D = 4
  expanded = tf.expand_dims(original_tensor, axis=1)  # Shape (3,1)
  result = tf.tile(expanded, [1, D])  # Shape (3,4)
  

• Using tf.repeat():

  result = tf.repeat(tf.expand_dims(original_tensor, 1), D, axis=1)
  

Quick Note #

If you ever need to replicate the original vector across rows instead of columns, just adjust the axis you expand and repeat/tile along (e.g., unsqueeze(0) instead of unsqueeze(1)).

内容的提问来源于stack exchange，提问作者Pablo Sanchez