.index_copy_()

The .index_copy_() method in PyTorch is an in-place operation that copies values from a source tensor into the specified indices of an input tensor along the given dimension, modifying the input tensor directly.

Syntax

input_tensor.index_copy_(dim, index, source_tensor)

• input_tensor: The tensor to copy values into.
• dim: The dimension along which to copy values.
• index: A 1D tensor specifying the indices to copy the values to.
• source_tensor: The tensor containing the values to be copied. Its size along the specified dimension must match the size of the index tensor.

Example

The following example illustrates the usage of the .index_copy_() method:

import torch

# Case 1: Copying values along rows
input_tensor = torch.zeros(3, 3)
print("Input Tensor:\n", input_tensor)

index_row = torch.tensor([0, 2])
source_tensor_row = torch.tensor([[1, 1, 1], [3, 3, 3]], dtype=torch.float)

input_tensor.index_copy_(0, index_row, source_tensor_row)
print("\nResult Tensor (Row Copy):\n", input_tensor)

# Case 2: Copying values along columns
input_tensor = torch.zeros(3, 3)  # Reset the tensor
index_col = torch.tensor([0, 2])
source_tensor_col = torch.tensor([[1, 3], [1, 3], [1, 3]], dtype=torch.float)

input_tensor.index_copy_(1, index_col, source_tensor_col)
print("\nResult Tensor (Column Copy):\n", input_tensor)

The above program gives the following output:

Input Tensor:
 tensor([[0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.]])

Result Tensor (Row Copy):
 tensor([[1., 1., 1.],
        [0., 0., 0.],
        [3., 3., 3.]])

Result Tensor (Column Copy):
 tensor([[1., 0., 3.],
        [1., 0., 3.],
        [1., 0., 3.]])