Name: ConvTranspose2d() in PyTorch
Rating: 4.4 (2040 reviews)
Author: hyperkai

*Memos:

My post explains Transposed Convolutional Layer.
My post explains ConvTranspose1d().
My post explains ConvTranspose3d().
My post explains manual_seed().
My post explains requires_grad.

ConvTranspose2d() can get the 3D or 4D tensor of the one or more elements computed by 2D transposed convolution from the 3D or 4D tensor of one or more elements as shown below:

*Memos:

The 1st argument for initialization is in_channels(Required-Type:float). *It must be 1 <= x.
The 2nd argument for initialization is out_channels(Required-Type:float). *It must be 1 <= x.
The 3rd argument for initialization is kernel_size(Required-Type:int or tuple or list of int). *It must be 1 <= x.
The 4th argument for initialization is stride(Optional-Default:1-Type:int or tuple or list of int). *It must be 1 <= x.
The 5th argument for initialization is padding(Optional-Default:0-Type:int or tuple or list of int). *It must be 0 <= x.
The 6th argument for initialization is output_padding(Optional-Default:0-Type:int, tuple or list of int). *It must be 0 <= x.
The 7th argument for initialization is groups(Optional-Default:1-Type:int). *It must be 1 <= x.
The 8th argument for initialization is bias(Optional-Default:True-Type:bool). *If it's False, None is set.
The 9th argument for initialization is dilation(Optional-Default:1-Type:int or tuple or list of int). *It must be 1 <= x.
The 10th argument for initialization is padding_mode(Optional-Default:'zeros'-Type:str). *Only 'zeros' can be selected.
The 11th argument for initialization is device(Optional-Type:str, int or device()). *Memos:
- If device is not given, get_default_device() is used. *My post explains get_default_device() and set_default_device().
- device= can be omitted.
- My post explains device argument.
The 12th argument for initialization is dtype(Optional-Type:int). *Memos:
- If dtype is not given, get_default_dtype() is used. *My post explains get_default_dtype() and set_default_dtype().
- dtype= can be omitted.
- My post explains dtype argument.
The 1st argument is input(Required-Type:tensor of float or complex). *complex must be set to dtype of ConvTranspose2d() to use a complex tensor.
The tensor's requires_grad which is False by default is set to True by ConvTranspose2d().
Input tensor's device and dtype must be same as ConvTranspose2d()'s device and dtype respectively.
convtran2d1.device and convtran2d1.dtype don't work.

import torch
from torch import nn

tensor1 = torch.tensor([[[8., -3., 0., 1., 5., -2.]]])

tensor1.requires_grad
# False

torch.manual_seed(42)

convtran2d1 = nn.ConvTranspose2d(in_channels=1, out_channels=3, kernel_size=1)
tensor2 = convtran2d1(input=tensor1)
tensor2
# tensor([[[4.0616, -0.7939, 0.5304, 0.9718, 2.7374, -0.3525]],
#         [[3.7071, -1.5641, -0.1265, 0.3527, 2.2695, -1.0849]],
#         [[-0.9656, 0.5223, 0.1165, -0.0188, -0.5598, 0.3870]]],
#        grad_fn=<SqueezeBackward1>)

tensor2.requires_grad
# True

convtran2d1
# ConvTranspose2d(1, 3, kernel_size=(1, 1), stride=(1, 1))

convtran2d1.in_channels
# 1

convtran2d1.out_channels
# 3

convtran2d1.kernel_size
# (1, 1)

convtran2d1.stride
# (1, 1)

convtran2d1.padding
# (0, 0)

convtran2d1.output_padding
# (0, 0)

convtran2d1.groups
# 1

convtran2d1.bias
# Parameter containing:
# tensor([0.5304, -0.1265, 0.1165], requires_grad=True)

convtran2d1.dilation
# (1, 1)

convtran2d1.padding_mode
# 'zeros'

convtran2d1.weight
# Parameter containing:
# tensor([[[0.4414], [0.4792], [-0.1353]]], requires_grad=True)

torch.manual_seed(42)

convtran2d2 = nn.ConvTranspose2d(in_channels=3, out_channels=3, kernel_size=1)
convtran2d2(input=tensor2)
# tensor([[[3.6068, -1.7503, -0.2893, 0.1977, 2.1458, -1.2633]],
#         [[1.6518, 0.4964, 0.8115, 0.9165, 1.3367, 0.6014]],
#         [[-0.5008, 0.2990, 0.0809, 0.0082, -0.2827, 0.2263]]],
#        grad_fn=<SqueezeBackward1>)

torch.manual_seed(42)

convtran2d = nn.ConvTranspose2d(in_channels=1, out_channels=3, 
             kernel_size=1, stride=1, padding=0, output_padding=0, 
             groups=1, bias=True, dilation=1, padding_mode='zeros', 
             device=None, dtype=None)
convtran2d(input=tensor1)
# tensor([[[4.0616, -0.7939, 0.5304, 0.9718, 2.7374, -0.3525]],
#         [[3.7071, -1.5641, -0.1265, 0.3527, 2.2695, -1.0849]],
#         [[-0.9656, 0.5223, 0.1165, -0.0188, -0.5598, 0.3870]]],
#        grad_fn=<SqueezeBackward1>)

my_tensor = torch.tensor([[[8., -3., 0.],
                           [1., 5., -2.]]])
torch.manual_seed(42)

convtran2d = nn.ConvTranspose2d(in_channels=1, out_channels=3,
                                kernel_size=1)
convtran2d(input=my_tensor)
# tensor([[[4.0616, -0.7939, 0.5304], [0.9718, 2.7374, -0.3525]],
#         [[3.7071, -1.5641, -0.1265], [0.3527, 2.2695, -1.0849]],
#         [[-0.9656, 0.5223, 0.1165], [-0.0188, -0.5598, 0.3870]]],
#        grad_fn=<SqueezeBackward1>)

my_tensor = torch.tensor([[[8.], [-3.], [0.],
                           [1.], [5.], [-2.]]])
torch.manual_seed(42)

convtran2d = nn.ConvTranspose2d(in_channels=1, out_channels=3, kernel_size=1)
convtran2d(input=my_tensor)
# tensor([[[4.0616], [-0.7939], [0.5304], [0.9718], [2.7374], [-0.3525]],
#         [[3.7071], [-1.5641], [-0.1265], [0.3527], [2.2695], [-1.0849]],
#         [[-0.9656], [0.5223], [0.1165], [-0.0188], [-0.5598], [0.3870]]],
#        grad_fn=<SqueezeBackward1>)

my_tensor = torch.tensor([[[[8.], [-3.], [0.]],
                           [[1.], [5.], [-2.]]]])
torch.manual_seed(42)

convtran2d = nn.ConvTranspose2d(in_channels=2, out_channels=3, kernel_size=1)
convtran2d(input=my_tensor)
# tensor([[[[3.7805], [1.0465], [-1.3418]],
#          [[4.0462], [-1.7310], [0.5921]],
#          [[-0.4566], [1.4973], [0.2760]]]],
#        grad_fn=<ConvolutionBackward0>)

my_tensor = torch.tensor([[[[8.+0.j], [-3.+0.j], [0.+0.j]],
                           [[1.+0.j], [5.+0.j], [-2.+0.j]]]])
torch.manual_seed(42)

convtran2d = nn.ConvTranspose2d(in_channels=2, out_channels=3,
                                kernel_size=1, dtype=torch.complex64)
convtran2d(input=my_tensor)
# tensor([[[[3.6767+4.2509j], [-2.3031+0.3359j], [0.9887-0.5999j]],
#         [[-0.2947+3.7378j], [3.2290-3.7904j], [-0.7395+0.7656j]],
#         [[-0.0651+1.1254j], [3.3337+0.2761j], [-0.5586-0.1308j]]]],
#        grad_fn=<AddBackward0>)