# Function Differences with torch.nn.ConvTranspose3d

torch.nn.ConvTranspose3d

class torch.nn.ConvTranspose3d(
    in_channels,
    out_channels,
    kernel_size,
    stride=1,
    padding=0,
    output_padding=0,
    groups=1,
    bias=True,
    dilation=1,
    padding_mode='zeros'
)(input) -> Tensor

For more information, see torch.nn.ConvTranspose3d.

mindspore.nn.Conv3dTranspose

class mindspore.nn.Conv3dTranspose(
    in_channels,
    out_channels,
    kernel_size,
    stride=1,
    pad_mode='same',
    padding=0,
    dilation=1,
    group=1,
    output_padding=0,
    has_bias=False,
    weight_init='normal',
    bias_init='zeros',
    data_format='NCDHW'
)(x) -> Tensor

For more information, see mindspore.nn.Conv3dTranspose.

Differences

PyTorch: Computing the 3D transposed convolution can be thought of as Conv3d solving for the gradient of the input, also known as deconvolution (which is not really deconvolution). The input shape is usually $(N,C_{in},D_{in},H_{in},W_{in})$, where $N$ is the batch size, $C$ is the spatial dimension, and $D_{in},H_{in},W_{in}$ are the depth, height and width of the feature layer, respectively. The output shape is $(N,C_{out},D_{out},H_{out},W_{out})$, which is calculated as follows. $D_{out}=(D_{in}-1)×stride[0]-2×padding[0]+dilation[0]×(kernel\underline{ }size[0]-1)+output\\ underline{ }padding[0]+1$ $H_{out}=(H_{in}−1)×stride[1]−2×padding[1]+dilation[1]×(kernel\underline{ }size[1]−1)+output\underline{ }padding[1]+1$ $W_{out}=(W_{in}−1)×stride[2]−2×padding[2]+dilation[2]×(kernel\underline{ }size[2]−1)+output\underline{ }padding[2]+1$

MindSpore: MindSpore API implements essentially the same function as PyTorch, but with the new "pad_mode" parameter. When "pad_mode" = "pad", it is the same as the PyTorch default, and the weight_init and bias_init parameters can be used to configure the initialization method.

Code Example 1

Both APIs implement 3D transposed convolutional operations and need to be instantiated first when used. To make the output the same width as the input after dividing stride, PyTorch sets output_padding = stride - 1 and padding to (kernel_size - 1)/2. MindSpore sets pad_mode = "same" and padding = 0.

# PyTorch
import torch
from torch import tensor
import torch.nn as nn
import numpy as np

k = 5
s = 3
x_ = np.ones([1, 3, 4, 9, 16])
x = tensor(x_, dtype=torch.float32)
net = nn.ConvTranspose3d(3, 32, kernel_size=k, stride=s, padding=(k-1)//2, output_padding=s-1, bias=False)
net.weight.data = torch.ones(3, 32, k, k, k)
output = net(x).detach().numpy()
print(output.shape)
# (1, 32, 12, 27, 48)


# MindSpore
import mindspore as ms
import mindspore.nn as nn
import numpy as np

k = 5
s = 3
x_ = np.ones([1, 3, 4, 9, 16])
x = ms.Tensor(x_, ms.float32)
net = nn.Conv3dTranspose(3, 32, kernel_size=k, stride=s, weight_init='ones', pad_mode='same')
output = net(x)
print(output.shape)
# (1, 32, 12, 27, 48)

Code Example 2

Both APIs implement 3D transposed convolutional operations and need to be instantiated first when used. If you do not do any padding on the original image, you may discard part of the data if stride > 1. Set padding and output_padding to 0 in PyTorch, and set pad_mode = "valid" in MindSpore, while padding = 0.

# PyTorch
import torch
from torch import tensor
import torch.nn as nn
import numpy as np

k = 5
s = 3
x_ = np.ones([1, 3, 4, 9, 16])
x = tensor(x_, dtype=torch.float32)
net = nn.ConvTranspose3d(3, 32, kernel_size=k, stride=s, bias=False)
net.weight.data = torch.ones(3, 32, k, k, k)
output = net(x).detach().numpy()
print(output.shape)
# (1, 32, 14, 29, 50)


# MindSpore
import mindspore as ms
import mindspore.nn as nn
import numpy as np

k = 5
s = 3
x_ = np.ones([1, 3, 4, 9, 16])
x = ms.Tensor(x_, ms.float32)
net = nn.Conv3dTranspose(3, 32, kernel_size=k, stride=s, weight_init='ones', pad_mode='valid')
output = net(x)
print(output.shape)
# (1, 32, 14, 29, 50)