1. Johngo学长首页
  2. 技术点详细复盘专题
  3. 人工智能

DDPM代码详细解读(2)：Unet结构、正向和逆向过程、IS和FID测试、EMA优化

• 人工智能 • 阅读 54

以下是将 Unet_和门 _结构_结合的 _PyTorch 代码： import torch import torch.nn as nn import torch.nn.functional as F class ConvBlock(nn.Module): def __init__(self, in_channels, out_channels): super(ConvBlock, self).__init__() self.conv1 = nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1) self.bn1 = nn.BatchNorm2d(out_channels) self.conv2 = nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1) self.bn2 = nn.BatchNorm2d(out_channels) def forward(self, x): x = self.conv1(x) x = self.bn1(x) x = F.relu(x) x = self.conv2(x) x = self.bn2(x) x = F.relu(x) return x class _UNet_(nn.Module): def __init__(self, in_channels=3, out_channels=1): super(_UNet_, self).__init__() self.down1 = ConvBlock(in_channels, 64) self.pool1 = nn.MaxPool2d(kernel_size=2, stride=2) self.down2 = ConvBlock(64, 128) self.pool2 = nn.MaxPool2d(kernel_size=2, stride=2) self.down3 = ConvBlock(128, 256) self.pool3 = nn.MaxPool2d(kernel_size=2, stride=2) self.down4 = ConvBlock(256, 512) self.pool4 = nn.MaxPool2d(kernel_size=2, stride=2) self.bottom = ConvBlock(512, 1024) self.up4 = nn.ConvTranspose2d(1024, 512, kernel_size=2, stride=2) self.upconv4 = ConvBlock(1024, 512) self.up3 = nn.ConvTranspose2d(512, 256, kernel_size=2, stride=2) self.upconv3 = ConvBlock(512, 256) self.up2 = nn.ConvTranspose2d(256, 128, kernel_size=2, stride=2) self.upconv2 = ConvBlock(256, 128) self.up1 = nn.ConvTranspose2d(128, 64, kernel_size=2, stride=2) self.upconv1 = ConvBlock(128, 64) self.outconv = nn.Conv2d(64, out_channels, kernel_size=1) # 定义门 _结构_的两个分支 self.branch1 = nn.Sequential( nn.Conv2d(1024, 1024, kernel_size=3, padding=1), nn.BatchNorm2d(1024), nn.ReLU(inplace=True), nn.Conv2d(1024, 1, kernel_size=1) ) self.branch2 = nn.Sequential( nn.Conv2d(1024, 1024, kernel_size=3, padding=1), nn.BatchNorm2d(1024), nn.ReLU(inplace=True), nn.Conv2d(1024, 1, kernel_size=1) ) def forward(self, x): # 下采样 x1 = self.down1(x) x2 = self.pool1(x1) x3 = self.down2(x2) x4 = self.pool2(x3) x5 = self.down3(x4) x6 = self.pool3(x5) x7 = self.down4(x6) x8 = self.pool4(x7) # 中间部分 x9 = self.bottom(x8) # 上采样 x10 = self.up4(x9) x10 = torch.cat([x10, x7], dim=1) x11 = self.upconv4(x10) x12 = self.up3(x11) x12 = torch.cat([x12, x5], dim=1) x13 = self.upconv3(x12) x14 = self.up2(x13) x14 = torch.cat([x14, x3], dim=1) x15 = self.upconv2(x14) x16 = self.up1(x15) x16 = torch.cat([x16, x1], dim=1) x17 = self.upconv1(x16) # 输出结果 x18 = self.outconv(x17) # 计算门控信号 branch1 = self.branch1(x9) branch2 = self.branch2(x9) gate = F.sigmoid(branch2) x18 = gate*x18 + (1-gate)*branch1 return x18 上述 _代码_定义了一个 _Unet_模型和一个门 _结构_模块，并将它们结合起来。具体来说， _Unet_模型由一系列卷积块和反卷积块组成，使用了下采样和上采样操作实现了图像分割。门 _结构_模块由两个分支组成，它们分别计算了输入特征图的重要程度，用sigmoid函数将它们结合起来得到一个门控信号，然后将这个信号应用于 _Unet_模型输出的结果和门 _结构_模块的输出结果来获得最终的输出。

Original: https://blog.csdn.net/qq_41895747/article/details/123678236
Author: 沉迷单车的追风少年
Title: DDPM代码详细解读(2)：Unet结构、正向和逆向过程、IS和FID测试、EMA优化

原创文章受到原创版权保护。转载请注明出处：https://www.johngo689.com/717459/

转载文章受原作者版权保护。转载请注明原作者出处！

赞 (0)
0

【自取】最近整理的，有需要可以领取学习：



Linux核心资料大放送~

全栈面试题汇总（持续更新&可下载）

一个提高学习100%效率的工具！

【超详细】深度学习面试题目！

LeetCode Python刷题答案下载！

LeetCode Java版刷题答案下载！

LeetCode C++ 版本，抓紧保存！

LeetCode GO语言 刷题答案下载！

大家都在看

亲爱的 Coder【最近整理，可免费获取】👉 最新必读书单  | 👏 面试题下载  | 🌎 免费的AI知识星球