比较与torch.nn.ConvTranspose2d的差异
torch.nn.ConvTranspose2d
class torch.nn.ConvTranspose2d(
in_channels,
out_channels,
kernel_size,
stride=1,
padding=0,
output_padding=0,
groups=1,
bias=True,
dilation=1,
padding_mode='zeros'
)(input) -> Tensor
更多内容详见torch.nn.ConvTranspose2d。
mindspore.nn.Conv2dTranspose
class mindspore.nn.Conv2dTranspose(
in_channels,
out_channels,
kernel_size,
stride=1,
pad_mode='same',
padding=0,
dilation=1,
group=1,
has_bias=False,
weight_init=None,
bias_init=None
)(x) -> Tensor
更多内容详见mindspore.nn.Conv2dTranspose。
差异对比
PyTorch:计算二维转置卷积,可以视为Conv2d对输入求梯度,也称为反卷积(实际不是真正的反卷积)。输入的shape通常是\((N,C_{in},H_{in},W_{in})\),其中\(N\)是batch size,\(C\)是空间维度,\(H_{in},W_{in}\)分别为高度和宽度。输出的shape为\((N,C_{out},H_{out},W_{out})\),高度和宽度分别为: \(H_{out}=(H_{in}−1)×stride[0]−2×padding[0]+dilation[0]×(kernel\underline{ }size[0]−1)+output\underline{ }padding[0]+1\) \(W_{out}=(W_{in}−1)×stride[1]−2×padding[1]+dilation[1]×(kernel\underline{ }size[1]−1)+output\underline{ }padding[1]+1\)
MindSpore:MindSpore此API实现功能与PyTorch基本一致,新增了填充模式参数”pad_mode”,当”pad_mode” = “pad”时与PyTorch默认方式相同,利用weight_init和bias_init参数可以配置初始化方式。此外,torch.nn.ConvTranspose2d有一个output_padding参数,其功能是指对反卷积后的特征图进行单侧补零(右侧和下侧),而mindspore.nn.Conv2dTranspose中目前没有该参数,可以对输出结果使用nn.Pad进行补维来代替。
权重初始化差异
mindspore.nn.Conv2dTranspose的 weight_init
是 None
时,权重使用HeUniform初始化。此时和PyTorch权重初始化方式一致。
mindspore.nn.Conv2dTranspose的 bias_init
是 None
时,偏差使用Uniform初始化。此时和PyTorch偏差初始化方式一致。
分类 |
子类 |
PyTorch |
MindSpore |
差异 |
---|---|---|---|---|
参数 |
参数1 |
in_channels |
in_channels |
- |
参数2 |
out_channels |
out_channels |
- |
|
参数3 |
kernel_size |
kernel_size |
- |
|
参数4 |
stride |
stride |
- |
|
参数5 |
padding |
padding |
功能一致,PyTorch中只能在两个维度的两侧分别填充相同的值,可为长度为2的tuple。MindSpore中可以分别设置顶部、底部、左边和右边的填充数量,可为长度为4的tuple |
|
参数6 |
output_padding |
- |
对反卷积后的特征图进行单侧补零(右侧和下侧),通常在stride > 1的前提下使用,用来调整output shapes。例如,通常将padding设置为(kernel_size - 1)/2,此时设置output_padding = (stride - 1)可确保input shapes/output shapes = stride,MindSpore无此参数 |
|
参数7 |
groups |
group |
功能一致,参数名不同 |
|
参数8 |
bias |
has_bias |
PyTorch默认为True,MindSpore默认为False |
|
参数9 |
dilation |
dilation |
- |
|
参数10 |
padding_mode |
- |
数值填充模式,只支持”zeros”即填充0。MindSpore无此参数,但默认填充0 |
|
参数11 |
- |
pad_mode |
指定填充模式。可选值为”same”、”valid”、”pad”,在”same”和”valid”模式下,padding必须设置为0,默认为”same”,PyTorch无此参数 |
|
参数12 |
- |
weight_init |
权重参数的初始化方法。可为Tensor、str、Initializer或numbers.Number。当使用str时,可选”TruncatedNormal” 、”Normal” 、”Uniform” 、”HeUniform”和”XavierUniform”分布以及常量”One”和”Zero”分布的值,默认为None,PyTorch无此参数 |
|
参数13 |
- |
bias_init |
偏置参数的初始化方法。初始化可选参数与”weight_init”相同,默认为None,PyTorch无此参数 |
|
输入 |
单输入 |
input |
x |
功能一致,参数名不同 |
代码示例1
两API都是实现二维转置卷积运算,使用时需先进行实例化。PyTorch中高度和宽度的padding值在同一方向上相同,如padding设为(2,4)表示分别在高度和宽度的两侧填充2行和4列0,对应在MindSpore中将pad_mode设为”pad”,padding设置为(2,2,4,4)。PyTorch中利用net.weight.data = torch.ones()的方式将权重初始化为1,shape为\((in\underline{ }channels,\frac {out\underline{ }channels}{groups} , kernel\underline{ }size[0], kernel\underline{ }size[1])\),MindSpore直接设置参数weight_init = “ones”。
# PyTorch
import torch
from torch import tensor
import torch.nn as nn
import numpy as np
k = 4
x_ = np.ones([1, 3, 16, 50])
x = tensor(x_, dtype=torch.float32)
net = nn.ConvTranspose2d(3, 64, kernel_size=k, stride=1, padding=(2, 4), output_padding=0, bias=False)
net.weight.data = torch.ones(3, 64, k, k)
output = net(x).detach().numpy()
print(output.shape)
# (1, 64, 15, 45)
# MindSpore
import mindspore as ms
import mindspore.nn as nn
import numpy as np
k = 4
x_ = np.ones([1, 3, 16, 50])
x = ms.Tensor(x_, ms.float32)
net = nn.Conv2dTranspose(3, 64, kernel_size=k, weight_init='ones', pad_mode='pad', padding=(2, 2, 4, 4))
output = net(x)
print(output.shape)
# (1, 64, 15, 45)
代码示例2
为使输出的宽度与输入整除stride后的值相同,PyTorch中设置output_padding = stride - 1,padding设置为(kernel_size - 1)/2。MindSpore则设置pad_mode = “same”,同时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, 16, 50])
x = tensor(x_, dtype=torch.float32)
net = nn.ConvTranspose2d(3, 64, kernel_size=k, stride=s, padding=(k-1)//2, output_padding=s-1, bias=False)
net.weight.data = torch.ones(3, 64, k, k)
output = net(x).detach().numpy()
print(output.shape)
# (1, 64, 48, 150)
# MindSpore
import mindspore as ms
import mindspore.nn as nn
import numpy as np
k = 5
s = 3
x_ = np.ones([1, 3, 16, 50])
x = ms.Tensor(x_, ms.float32)
net = nn.Conv2dTranspose(3, 64, kernel_size=k, stride=s, weight_init='ones', pad_mode='same', padding=0)
output = net(x)
print(output.shape)
# (1, 64, 48, 150)
代码示例3
若不在原有图像上做任何填充,在stride>1的情况下可能舍弃一部分数据,在PyTorch中将padding和output_padding设为0,MindSpore中设置pad_mode = “valid”,同时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, 16, 50])
x = tensor(x_, dtype=torch.float32)
net = nn.ConvTranspose2d(3, 64, kernel_size=k, stride=s, padding=0, output_padding=0, bias=False)
net.weight.data = torch.ones(3, 64, k, k)
output = net(x).detach().numpy()
print(output.shape)
# (1, 64, 50, 152)
# MindSpore
import mindspore as ms
import mindspore.nn as nn
import numpy as np
k = 5
s = 3
x_ = np.ones([1, 3, 16, 50])
x = ms.Tensor(x_, ms.float32)
net = nn.Conv2dTranspose(3, 64, kernel_size=k, stride=s, weight_init='ones', pad_mode='valid', padding=0)
output = net(x)
print(output.shape)
# (1, 64, 50, 152)
代码示例4
下面的示例实现了对输入tensor进行反卷积,并且输出反卷积后的特征图尺寸,其中PyTorch可以通过设置output_padding的值来对反卷积后的输出图像进行右侧和下侧补维,用于弥补stride大于1带来的缺失。MindSpore暂时不支持output_padding参数,需要对输出结果再使用nn.Pad进行单侧补维。
# PyTorch
import torch
import torch.nn as nn
import numpy as np
m = nn.ConvTranspose2d(in_channels=3, out_channels=32,
kernel_size=3,
stride=2,
padding=1,
output_padding=1,
bias=False)
input = torch.tensor(np.ones([1, 3, 48, 48]), dtype=torch.float32)
output = m(input).detach().numpy()
print(output.shape)
#(1, 32, 96, 96)
# MindSpore
import mindspore as ms
import mindspore.nn as nn
import numpy as np
input = ms.Tensor(np.ones([1, 3, 48, 48]), dtype=ms.float32)
m = nn.Conv2dTranspose(in_channels=3,
out_channels=32,
kernel_size=3,
stride=2,
padding=1,
pad_mode="pad",
has_bias=False)
output = m(input)
pad = nn.Pad(paddings=((0, 0), (0, 0), (0, 1), (0, 1)), mode="CONSTANT")
output = pad(output)
print(output.shape)
#(1, 32, 96, 96)