Repeat, tile and repeat in pytorch_ The difference between interleave

torch.Tensor.repeat

repeat It can be understood visually as repeating the existing tensor many times to form “ Block matrix ”.

import torch

""" Example 1 """
t = torch.arange(3)
print(t.repeat((2, )))
# tensor([0, 1, 2, 0, 1, 2])
print(t.repeat((2, 2)))
# tensor([[0, 1, 2, 0, 1, 2],
# [0, 1, 2, 0, 1, 2]])

""" Example 2 """
t = torch.arange(4).reshape(2, 2)
print(t.repeat((2, )))
# RuntimeError: Number of dimensions of repeat dims can not be smaller than number of dimensions of tensor
print(t.repeat((2, 2)))
# tensor([[0, 1, 0, 1],
# [2, 3, 2, 3],
# [0, 1, 0, 1],
# [2, 3, 2, 3]])
print(t.repeat((2, 3, 4)))
# tensor([[[0, 1, 0, 1, 0, 1, 0, 1],
# [2, 3, 2, 3, 2, 3, 2, 3],
# [0, 1, 0, 1, 0, 1, 0, 1],
# [2, 3, 2, 3, 2, 3, 2, 3],
# [0, 1, 0, 1, 0, 1, 0, 1],
# [2, 3, 2, 3, 2, 3, 2, 3]],
# 
# [[0, 1, 0, 1, 0, 1, 0, 1],
# [2, 3, 2, 3, 2, 3, 2, 3],
# [0, 1, 0, 1, 0, 1, 0, 1],
# [2, 3, 2, 3, 2, 3, 2, 3],
# [0, 1, 0, 1, 0, 1, 0, 1],
# [2, 3, 2, 3, 2, 3, 2, 3]]])

It can be seen that repeat Dimensions Not lower than The dimension of tensor itself .

torch.Tensor.tile

In most cases ,tile And repeat It's the same thing , as follows ：

""" Example 1 """
t = torch.arange(3)
print(t.tile((2, )))
# tensor([0, 1, 2, 0, 1, 2])
print(t.tile((2, 2)))
# tensor([[0, 1, 2, 0, 1, 2],
# [0, 1, 2, 0, 1, 2]])

""" Example 2 """
t = torch.arange(4).reshape(2, 2)
print(t.tile((2, )))
# tensor([[0, 1, 0, 1],
# [2, 3, 2, 3]])
print(t.tile((2, 2)))
# tensor([[0, 1, 0, 1],
# [2, 3, 2, 3],
# [0, 1, 0, 1],
# [2, 3, 2, 3]])
print(t.tile((2, 3, 4)))
# tensor([[[0, 1, 0, 1, 0, 1, 0, 1],
# [2, 3, 2, 3, 2, 3, 2, 3],
# [0, 1, 0, 1, 0, 1, 0, 1],
# [2, 3, 2, 3, 2, 3, 2, 3],
# [0, 1, 0, 1, 0, 1, 0, 1],
# [2, 3, 2, 3, 2, 3, 2, 3]],
# 
# [[0, 1, 0, 1, 0, 1, 0, 1],
# [2, 3, 2, 3, 2, 3, 2, 3],
# [0, 1, 0, 1, 0, 1, 0, 1],
# [2, 3, 2, 3, 2, 3, 2, 3],
# [0, 1, 0, 1, 0, 1, 0, 1],
# [2, 3, 2, 3, 2, 3, 2, 3]]])

And repeat The difference is , When the dimension to be repeated is lower than the tensor dimension ,tile Can also deal with , here tile Can use front $1$ Automatically replenish dimensions .

torch.Tensor.repeat_interleave

Previously mentioned repeat and tile All are Repeat the entire tensor , And this time repeat_interleave It is Repeat the elements in the tensor .

The parameters are as follows ：

torch.Tensor.repeat_interleave(repeats, dim=None)

• repeats： Represents the number of times each element in the tensor will be repeated . It can be integer or tensor ;
• dim： Determines which axis to repeat numbers along . By default, the input will be flattened and repeated , Finally, output the flattened tensor .

""" Example 1 """
t = torch.arange(3)
print(t.repeat_interleave(repeats=3))
# tensor([0, 0, 0, 1, 1, 1, 2, 2, 2])

""" Example 2 """
t = torch.arange(4).reshape(2, 2)
print(t.repeat_interleave(repeats=3))
# tensor([0, 0, 0, 1, 1, 1, 2, 2, 2, 3, 3, 3])
print(t.repeat_interleave(repeats=3, dim=0))
# tensor([[0, 1],
# [0, 1],
# [0, 1],
# [2, 3],
# [2, 3],
# [2, 3]])
print(t.repeat_interleave(repeats=3, dim=1))
# tensor([[0, 0, 0, 1, 1, 1],
# [2, 2, 2, 3, 3, 3]])

""" Example 3 """
t = torch.arange(4).reshape(2, 2)
print(t.repeat_interleave(repeats=torch.tensor([2, 3]), dim=0))  # t Repeat the first line of 2 Time , The first 2 Row repeat 3 Time 
# tensor([[0, 1],
# [0, 1],
# [2, 3],
# [2, 3],
# [2, 3]])
print(t.repeat_interleave(repeats=torch.tensor([3, 2]), dim=1))  # t The first column of repeats 3 Time , The first 2 Column repeat 2 Time 
# tensor([[0, 0, 0, 1, 1],
# [2, 2, 2, 3, 3]])