Name: Unflatten in PyTorch
Rating: 1.3 (3065 reviews)
Author: hyperkai

*Memos:

My post explains unflatten().
My post explains flatten() and ravel().
My post explains Flatten().

Unflatten() can add zero or more dimensions to the 1D or more D tensor of zero or more elements, getting the 1D or more D tensor of zero or more elements as shown below:

*Memos:

The 1st argument for initialization is dim(Required-Type:int).
The 2nd argument for initialization is unflattened_size(Required-Type:tuple or list of int).
The 1st argument is input(Required-Type:tensor of int, float, complex or bool). *-1 infers and adjust the size.
The difference between Unflatten() and unflatten() is:
- Unflatten() has unflattened_size argument which is identical to sizes argument of unflatten().
- Basically, Unflatten() is used to define a model while unflatten() is not used to define a model.

import torch
from torch import nn

unflatten = nn.Unflatten()
unflatten
# Unflatten(dim=0, unflattened_size=(6,))

unflatten.dim
# 0

unflatten.unflattened_size
# (6,)

my_tensor = torch.tensor([7, 1, -8, 3, -6, 0])

unflatten = nn.Unflatten(dim=0, unflattened_size=(6,))
unflatten = nn.Unflatten(dim=0, unflattened_size=(-1,))
unflatten = nn.Unflatten(dim=-1, unflattened_size=(6,))
unflatten = nn.Unflatten(dim=-1, unflattened_size=(-1,))
unflatten(input=my_tensor)
# tensor([7, 1, -8, 3, -6, 0])

unflatten = nn.Unflatten(dim=0, unflattened_size=(1, 6))
unflatten = nn.Unflatten(dim=0, unflattened_size=(-1, 6))
unflatten = nn.Unflatten(dim=0, unflattened_size=(1, -1))
unflatten = nn.Unflatten(dim=-1, unflattened_size=(1, 6))
unflatten = nn.Unflatten(dim=-1, unflattened_size=(-1, 6))
unflatten = nn.Unflatten(dim=-1, unflattened_size=(1, -1))
unflatten(input=my_tensor)
# tensor([[7, 1, -8, 3, -6, 0]])

unflatten = nn.Unflatten(dim=0, unflattened_size=(2, 3))
unflatten = nn.Unflatten(dim=0, unflattened_size=(2, -1))
unflatten = nn.Unflatten(dim=-1, unflattened_size=(2, 3))
unflatten = nn.Unflatten(dim=-1, unflattened_size=(2, -1))
unflatten(input=my_tensor)
# tensor([[7, 1, -8], [3, -6, 0]])

unflatten = nn.Unflatten(dim=0, unflattened_size=(3, 2))
unflatten = nn.Unflatten(dim=0, unflattened_size=(3, -1))
unflatten = nn.Unflatten(dim=-1, unflattened_size=(3, 2))
unflatten = nn.Unflatten(dim=-1, unflattened_size=(3, -1))
unflatten(input=my_tensor)
# tensor([[7, 1], [-8, 3], [-6, 0]])

unflatten = nn.Unflatten(dim=0, unflattened_size=(6, 1))
unflatten = nn.Unflatten(dim=0, unflattened_size=(6, -1))
unflatten = nn.Unflatten(dim=-1, unflattened_size=(6, 1))
unflatten = nn.Unflatten(dim=-1, unflattened_size=(6, -1))
unflatten(input=my_tensor)
# tensor([[7], [1], [-8], [3], [-6], [0]])

unflatten = nn.Unflatten(dim=0, unflattened_size=(1, 2, 3))
unflatten = nn.Unflatten(dim=0, unflattened_size=(-1, 2, 3))
unflatten = nn.Unflatten(dim=0, unflattened_size=(1, -1, 3))
unflatten = nn.Unflatten(dim=0, unflattened_size=(1, 2, -1))
unflatten = nn.Unflatten(dim=-1, unflattened_size=(1, 2, 3))
unflatten = nn.Unflatten(dim=-1, unflattened_size=(-1, 2, 3))
unflatten = nn.Unflatten(dim=-1, unflattened_size=(1, -1, 3))
unflatten = nn.Unflatten(dim=-1, unflattened_size=(1, 2, -1))
unflatten(input=my_tensor)
# tensor([[[7, 1, -8], [3, -6, 0]]])
etc

my_tensor = torch.tensor([[7, 1, -8], [3, -6, 0]])

unflatten = nn.Unflatten(dim=0, unflattened_size=(2,))
unflatten = nn.Unflatten(dim=0, unflattened_size=(-1,))
unflatten = nn.Unflatten(dim=1, unflattened_size=(3,))
unflatten = nn.Unflatten(dim=1, unflattened_size=(-1,))
unflatten = nn.Unflatten(dim=-1, unflattened_size=(3,))
unflatten = nn.Unflatten(dim=-1, unflattened_size=(-1,))
unflatten = nn.Unflatten(dim=-2, unflattened_size=(2,))
unflatten = nn.Unflatten(dim=-2, unflattened_size=(-1,))
unflatten(input=my_tensor)
# tensor([[7, 1, -8], [3, -6, 0]])

unflatten = nn.Unflatten(dim=0, unflattened_size=(1, 2))
unflatten = nn.Unflatten(dim=0, unflattened_size=(-1, 2))
unflatten = nn.Unflatten(dim=-2, unflattened_size=(1, 2))
unflatten = nn.Unflatten(dim=-2, unflattened_size=(-1, 2))
unflatten(input=my_tensor)
# tensor([[[7, 1, -8], [3, -6, 0]]])

unflatten = nn.Unflatten(dim=0, unflattened_size=(2, 1))
unflatten = nn.Unflatten(dim=0, unflattened_size=(2, -1))
unflatten = nn.Unflatten(dim=1, unflattened_size=(1, 3))
unflatten = nn.Unflatten(dim=1, unflattened_size=(-1, 3))
unflatten = nn.Unflatten(dim=-1, unflattened_size=(1, 3))
unflatten = nn.Unflatten(dim=-1, unflattened_size=(-1, 3))
unflatten = nn.Unflatten(dim=-2, unflattened_size=(2, 1))
unflatten = nn.Unflatten(dim=-2, unflattened_size=(2, -1))
unflatten(input=my_tensor)
# tensor([[[7, 1, -8]], [[3, -6, 0]]])

unflatten = nn.Unflatten(dim=1, unflattened_size=(3, 1))
unflatten = nn.Unflatten(dim=1, unflattened_size=(3, -1))
unflatten = nn.Unflatten(dim=-1, unflattened_size=(3, 1))
unflatten = nn.Unflatten(dim=-1, unflattened_size=(3, -1))
unflatten(input=my_tensor)
# tensor([[[7], [1], [-8]], [[3], [-6], [0]]])

unflatten = nn.Unflatten(dim=0, unflattened_size=(1, 1, 2))
unflatten = nn.Unflatten(dim=0, unflattened_size=(-1, 1, 2))
unflatten = nn.Unflatten(dim=0, unflattened_size=(1, -1, 2))
unflatten = nn.Unflatten(dim=0, unflattened_size=(1, 1, -1))
unflatten = nn.Unflatten(dim=-2, unflattened_size=(1, 1, 2))
unflatten = nn.Unflatten(dim=-2, unflattened_size=(-1, 1, 2))
unflatten = nn.Unflatten(dim=-2, unflattened_size=(1, -1, 2))
unflatten = nn.Unflatten(dim=-2, unflattened_size=(1, 1, -1))
unflatten(input=my_tensor)
# tensor([[[[7, 1, -8], [3, -6, 0]]]])

unflatten = nn.Unflatten(dim=1, unflattened_size=(1, 1, 3))
unflatten = nn.Unflatten(dim=1, unflattened_size=(-1, 1, 3))
unflatten = nn.Unflatten(dim=1, unflattened_size=(1, -1, 3))
unflatten = nn.Unflatten(dim=1, unflattened_size=(1, 1, -1))
unflatten = nn.Unflatten(dim=-1, unflattened_size=(1, 1, 3))
unflatten = nn.Unflatten(dim=-1, unflattened_size=(-1, 1, 3))
unflatten = nn.Unflatten(dim=-1, unflattened_size=(1, -1, 3))
unflatten = nn.Unflatten(dim=-1, unflattened_size=(1, 1, -1))
unflatten(input=my_tensor)
# tensor([[[[7, 1, -8]]], [[[3, -6, 0]]]])

my_tensor = torch.tensor([[7., 1., -8.], [3., -6., 0.]])

unflatten = nn.Unflatten(dim=0, unflattened_size=(2,))
unflatten(input=my_tensor)
# tensor([[7., 1., -8.], [3., -6., 0.]])

my_tensor = torch.tensor([[7.+0.j, 1.+0.j, -8.+0.j],
                          [3.+0.j, -6.+0.j, 0.+0.j]])
unflatten = nn.Unflatten(dim=0, unflattened_size=(2,))
unflatten(input=my_tensor)
# tensor([[7.+0.j, 1.+0.j, -8.+0.j],
#         [3.+0.j, -6.+0.j, 0.+0.j]])

my_tensor = torch.tensor([[True, False, True], [False, True, False]])

unflatten = nn.Unflatten(dim=0, unflattened_size=(2,))
unflatten(input=my_tensor)
# tensor([[True, False, True], [False, True, False]])