深入浅出 diffusion(4)：pytorch 实现简单 diffusion

 1. 训练和采样流程

 2. 无条件实现

import torch, time, os
import numpy as np
import torch.nn as nn
import torch.optim as optim
from torchvision.datasets import MNIST
from torchvision import transforms
from torch.utils.data import DataLoader
from torchvision.utils import save_image
import torch.nn.functional as Fclass ResidualConvBlock(nn.Module):def __init__(self, in_channels: int, out_channels: int, is_res: bool = False) -> None:super().__init__()'''standard ResNet style convolutional block'''self.same_channels = in_channels==out_channelsself.is_res = is_resself.conv1 = nn.Sequential(nn.Conv2d(in_channels, out_channels, 3, 1, 1),nn.BatchNorm2d(out_channels),nn.GELU(),)self.conv2 = nn.Sequential(nn.Conv2d(out_channels, out_channels, 3, 1, 1),nn.BatchNorm2d(out_channels),nn.GELU(),)def forward(self, x: torch.Tensor) -> torch.Tensor:if self.is_res:x1 = self.conv1(x)x2 = self.conv2(x1)# this adds on correct residual in case channels have increasedif self.same_channels:out = x + x2else:out = x1 + x2return out / 1.414else:x1 = self.conv1(x)x2 = self.conv2(x1)return x2class UnetDown(nn.Module):def __init__(self, in_channels, out_channels):super(UnetDown, self).__init__()'''process and downscale the image feature maps'''layers = [ResidualConvBlock(in_channels, out_channels), nn.MaxPool2d(2)]self.model = nn.Sequential(*layers)def forward(self, x):return self.model(x)class UnetUp(nn.Module):def __init__(self, in_channels, out_channels):super(UnetUp, self).__init__()'''process and upscale the image feature maps'''layers = [nn.ConvTranspose2d(in_channels, out_channels, 2, 2),ResidualConvBlock(out_channels, out_channels),ResidualConvBlock(out_channels, out_channels),]self.model = nn.Sequential(*layers)def forward(self, x, skip):x = torch.cat((x, skip), 1)x = self.model(x)return xclass EmbedFC(nn.Module):def __init__(self, input_dim, emb_dim):super(EmbedFC, self).__init__()'''generic one layer FC NN for embedding things  '''self.input_dim = input_dimlayers = [nn.Linear(input_dim, emb_dim),nn.GELU(),nn.Linear(emb_dim, emb_dim),]self.model = nn.Sequential(*layers)def forward(self, x):x = x.view(-1, self.input_dim)return self.model(x)
class Unet(nn.Module):def __init__(self, in_channels, n_feat=256):super(Unet, self).__init__()self.in_channels = in_channelsself.n_feat = n_featself.init_conv = ResidualConvBlock(in_channels, n_feat, is_res=True)self.down1 = UnetDown(n_feat, n_feat)self.down2 = UnetDown(n_feat, 2 * n_feat)self.to_vec = nn.Sequential(nn.AvgPool2d(7), nn.GELU())self.timeembed1 = EmbedFC(1, 2 * n_feat)self.timeembed2 = EmbedFC(1, 1 * n_feat)self.up0 = nn.Sequential(# nn.ConvTranspose2d(6 * n_feat, 2 * n_feat, 7, 7), # when concat temb and cemb end up w 6*n_featnn.ConvTranspose2d(2 * n_feat, 2 * n_feat, 7, 7),  # otherwise just have 2*n_featnn.GroupNorm(8, 2 * n_feat),nn.ReLU(),)self.up1 = UnetUp(4 * n_feat, n_feat)self.up2 = UnetUp(2 * n_feat, n_feat)self.out = nn.Sequential(nn.Conv2d(2 * n_feat, n_feat, 3, 1, 1),nn.GroupNorm(8, n_feat),nn.ReLU(),nn.Conv2d(n_feat, self.in_channels, 3, 1, 1),)def forward(self, x, t):'''输入加噪图像和对应的时间step，预测反向噪声的正态分布:param x: 加噪图像:param t: 对应step:return: 正态分布噪声'''x = self.init_conv(x)down1 = self.down1(x)down2 = self.down2(down1)hiddenvec = self.to_vec(down2)# embed time steptemb1 = self.timeembed1(t).view(-1, self.n_feat * 2, 1, 1)temb2 = self.timeembed2(t).view(-1, self.n_feat, 1, 1)# 将上采样输出与step编码相加，输入到下一个上采样层up1 = self.up0(hiddenvec)up2 = self.up1(up1 + temb1, down2)up3 = self.up2(up2 + temb2, down1)out = self.out(torch.cat((up3, x), 1))return outclass DDPM(nn.Module):def __init__(self, model, betas, n_T, device):super(DDPM, self).__init__()self.model = model.to(device)# register_buffer 可以提前保存alpha相关，节约时间for k, v in self.ddpm_schedules(betas[0], betas[1], n_T).items():self.register_buffer(k, v)self.n_T = n_Tself.device = deviceself.loss_mse = nn.MSELoss()def ddpm_schedules(self, beta1, beta2, T):'''提前计算各个step的alpha，这里beta是线性变化:param beta1: beta的下限:param beta2: beta的下限:param T: 总共的step数'''assert beta1 < beta2 < 1.0, "beta1 and beta2 must be in (0, 1)"beta_t = (beta2 - beta1) * torch.arange(0, T + 1, dtype=torch.float32) / T + beta1 # 生成beta1-beta2均匀分布的数组sqrt_beta_t = torch.sqrt(beta_t)alpha_t = 1 - beta_tlog_alpha_t = torch.log(alpha_t)alphabar_t = torch.cumsum(log_alpha_t, dim=0).exp() # alpha累乘sqrtab = torch.sqrt(alphabar_t) # 根号alpha累乘oneover_sqrta = 1 / torch.sqrt(alpha_t) # 1 / 根号alphasqrtmab = torch.sqrt(1 - alphabar_t) # 根号下（1-alpha累乘）mab_over_sqrtmab_inv = (1 - alpha_t) / sqrtmabreturn {"alpha_t": alpha_t,  # \alpha_t"oneover_sqrta": oneover_sqrta,  # 1/\sqrt{\alpha_t}"sqrt_beta_t": sqrt_beta_t,  # \sqrt{\beta_t}"alphabar_t": alphabar_t,  # \bar{\alpha_t}"sqrtab": sqrtab,  # \sqrt{\bar{\alpha_t}} # 加噪标准差"sqrtmab": sqrtmab,  # \sqrt{1-\bar{\alpha_t}}  # 加噪均值"mab_over_sqrtmab": mab_over_sqrtmab_inv,  # (1-\alpha_t)/\sqrt{1-\bar{\alpha_t}}}def forward(self, x):"""训练过程中, 随机选择step和生成噪声"""# 随机选择step_ts = torch.randint(1, self.n_T + 1, (x.shape[0],)).to(self.device)  # t ~ Uniform(0, n_T)# 随机生成正态分布噪声noise = torch.randn_like(x)  # eps ~ N(0, 1)# 加噪后的图像x_tx_t = (self.sqrtab[_ts, None, None, None] * x+ self.sqrtmab[_ts, None, None, None] * noise)# 将unet预测的对应step的正态分布噪声与真实噪声做对比return self.loss_mse(noise, self.model(x_t, _ts / self.n_T))def sample(self, n_sample, size, device):# 随机生成初始噪声图片 x_T ~ N(0, 1)x_i = torch.randn(n_sample, *size).to(device)for i in range(self.n_T, 0, -1):t_is = torch.tensor([i / self.n_T]).to(device)t_is = t_is.repeat(n_sample, 1, 1, 1)z = torch.randn(n_sample, *size).to(device) if i > 1 else 0eps = self.model(x_i, t_is)x_i = x_i[:n_sample]x_i = self.oneover_sqrta[i] * (x_i - eps * self.mab_over_sqrtmab[i]) + self.sqrt_beta_t[i] * zreturn x_iclass ImageGenerator(object):def __init__(self):'''初始化，定义超参数、数据集、网络结构等'''self.epoch = 20self.sample_num = 100self.batch_size = 256self.lr = 0.0001self.n_T = 400self.device = 'cuda' if torch.cuda.is_available() else 'cpu'self.init_dataloader()self.sampler = DDPM(model=Unet(in_channels=1), betas=(1e-4, 0.02), n_T=self.n_T, device=self.device).to(self.device)self.optimizer = optim.Adam(self.sampler.model.parameters(), lr=self.lr)def init_dataloader(self):'''初始化数据集和dataloader'''tf = transforms.Compose([transforms.ToTensor(),])train_dataset = MNIST('./data/',train=True,download=True,transform=tf)self.train_dataloader = DataLoader(train_dataset, batch_size=self.batch_size, shuffle=True, drop_last=True)val_dataset = MNIST('./data/',train=False,download=True,transform=tf)self.val_dataloader = DataLoader(val_dataset, batch_size=self.batch_size, shuffle=False)def train(self):self.sampler.train()print('训练开始!!')for epoch in range(self.epoch):self.sampler.model.train()loss_mean = 0for i, (images, labels) in enumerate(self.train_dataloader):images, labels = images.to(self.device), labels.to(self.device)# 将latent和condition拼接后输入网络loss = self.sampler(images)loss_mean += loss.item()self.optimizer.zero_grad()loss.backward()self.optimizer.step()train_loss = loss_mean / len(self.train_dataloader)print('epoch:{}, loss:{:.4f}'.format(epoch, train_loss))self.visualize_results(epoch)@torch.no_grad()def visualize_results(self, epoch):self.sampler.eval()# 保存结果路径output_path = 'results/Diffusion'if not os.path.exists(output_path):os.makedirs(output_path)tot_num_samples = self.sample_numimage_frame_dim = int(np.floor(np.sqrt(tot_num_samples)))out = self.sampler.sample(tot_num_samples, (1, 28, 28), self.device)save_image(out, os.path.join(output_path, '{}.jpg'.format(epoch)), nrow=image_frame_dim)if __name__ == '__main__':generator = ImageGenerator()generator.train()

3. 有条件实现

import torch, time, os
import numpy as np
import torch.nn as nn
import torch.optim as optim
from torchvision.datasets import MNIST
from torchvision import transforms
from torch.utils.data import DataLoader
from torchvision.utils import save_image
import torch.nn.functional as Fclass ResidualConvBlock(nn.Module):def __init__(self, in_channels: int, out_channels: int, is_res: bool = False) -> None:super().__init__()'''standard ResNet style convolutional block'''self.same_channels = in_channels==out_channelsself.is_res = is_resself.conv1 = nn.Sequential(nn.Conv2d(in_channels, out_channels, 3, 1, 1),nn.BatchNorm2d(out_channels),nn.GELU(),)self.conv2 = nn.Sequential(nn.Conv2d(out_channels, out_channels, 3, 1, 1),nn.BatchNorm2d(out_channels),nn.GELU(),)def forward(self, x: torch.Tensor) -> torch.Tensor:if self.is_res:x1 = self.conv1(x)x2 = self.conv2(x1)# this adds on correct residual in case channels have increasedif self.same_channels:out = x + x2else:out = x1 + x2return out / 1.414else:x1 = self.conv1(x)x2 = self.conv2(x1)return x2class UnetDown(nn.Module):def __init__(self, in_channels, out_channels):super(UnetDown, self).__init__()'''process and downscale the image feature maps'''layers = [ResidualConvBlock(in_channels, out_channels), nn.MaxPool2d(2)]self.model = nn.Sequential(*layers)def forward(self, x):return self.model(x)class UnetUp(nn.Module):def __init__(self, in_channels, out_channels):super(UnetUp, self).__init__()'''process and upscale the image feature maps'''layers = [nn.ConvTranspose2d(in_channels, out_channels, 2, 2),ResidualConvBlock(out_channels, out_channels),ResidualConvBlock(out_channels, out_channels),]self.model = nn.Sequential(*layers)def forward(self, x, skip):x = torch.cat((x, skip), 1)x = self.model(x)return xclass EmbedFC(nn.Module):def __init__(self, input_dim, emb_dim):super(EmbedFC, self).__init__()'''generic one layer FC NN for embedding things  '''self.input_dim = input_dimlayers = [nn.Linear(input_dim, emb_dim),nn.GELU(),nn.Linear(emb_dim, emb_dim),]self.model = nn.Sequential(*layers)def forward(self, x):x = x.view(-1, self.input_dim)return self.model(x)
class Unet(nn.Module):def __init__(self, in_channels, n_feat=256, n_classes=10):super(Unet, self).__init__()self.in_channels = in_channelsself.n_feat = n_featself.init_conv = ResidualConvBlock(in_channels, n_feat, is_res=True)self.down1 = UnetDown(n_feat, n_feat)self.down2 = UnetDown(n_feat, 2 * n_feat)self.to_vec = nn.Sequential(nn.AvgPool2d(7), nn.GELU())self.timeembed1 = EmbedFC(1, 2 * n_feat)self.timeembed2 = EmbedFC(1, 1 * n_feat)self.conditionembed1 = EmbedFC(n_classes, 2 * n_feat)self.conditionembed2 = EmbedFC(n_classes, 1 * n_feat)self.up0 = nn.Sequential(# nn.ConvTranspose2d(6 * n_feat, 2 * n_feat, 7, 7), # when concat temb and cemb end up w 6*n_featnn.ConvTranspose2d(2 * n_feat, 2 * n_feat, 7, 7),  # otherwise just have 2*n_featnn.GroupNorm(8, 2 * n_feat),nn.ReLU(),)self.up1 = UnetUp(4 * n_feat, n_feat)self.up2 = UnetUp(2 * n_feat, n_feat)self.out = nn.Sequential(nn.Conv2d(2 * n_feat, n_feat, 3, 1, 1),nn.GroupNorm(8, n_feat),nn.ReLU(),nn.Conv2d(n_feat, self.in_channels, 3, 1, 1),)def forward(self, x, c, t):'''输入加噪图像和对应的时间step，预测反向噪声的正态分布:param x: 加噪图像:param c: contition向量:param t: 对应step:return: 正态分布噪声'''x = self.init_conv(x)down1 = self.down1(x)down2 = self.down2(down1)hiddenvec = self.to_vec(down2)# embed time steptemb1 = self.timeembed1(t).view(-1, self.n_feat * 2, 1, 1)temb2 = self.timeembed2(t).view(-1, self.n_feat, 1, 1)cemb1 = self.conditionembed1(c).view(-1, self.n_feat * 2, 1, 1)cemb2 = self.conditionembed2(c).view(-1, self.n_feat, 1, 1)# 将上采样输出与step编码相加，输入到下一个上采样层up1 = self.up0(hiddenvec)up2 = self.up1(cemb1 * up1 + temb1, down2)up3 = self.up2(cemb2 * up2 + temb2, down1)out = self.out(torch.cat((up3, x), 1))return outclass DDPM(nn.Module):def __init__(self, model, betas, n_T, device):super(DDPM, self).__init__()self.model = model.to(device)# register_buffer 可以提前保存alpha相关，节约时间for k, v in self.ddpm_schedules(betas[0], betas[1], n_T).items():self.register_buffer(k, v)self.n_T = n_Tself.device = deviceself.loss_mse = nn.MSELoss()def ddpm_schedules(self, beta1, beta2, T):'''提前计算各个step的alpha，这里beta是线性变化:param beta1: beta的下限:param beta2: beta的下限:param T: 总共的step数'''assert beta1 < beta2 < 1.0, "beta1 and beta2 must be in (0, 1)"beta_t = (beta2 - beta1) * torch.arange(0, T + 1, dtype=torch.float32) / T + beta1 # 生成beta1-beta2均匀分布的数组sqrt_beta_t = torch.sqrt(beta_t)alpha_t = 1 - beta_tlog_alpha_t = torch.log(alpha_t)alphabar_t = torch.cumsum(log_alpha_t, dim=0).exp() # alpha累乘sqrtab = torch.sqrt(alphabar_t) # 根号alpha累乘oneover_sqrta = 1 / torch.sqrt(alpha_t) # 1 / 根号alphasqrtmab = torch.sqrt(1 - alphabar_t) # 根号下（1-alpha累乘）mab_over_sqrtmab_inv = (1 - alpha_t) / sqrtmabreturn {"alpha_t": alpha_t,  # \alpha_t"oneover_sqrta": oneover_sqrta,  # 1/\sqrt{\alpha_t}"sqrt_beta_t": sqrt_beta_t,  # \sqrt{\beta_t}"alphabar_t": alphabar_t,  # \bar{\alpha_t}"sqrtab": sqrtab,  # \sqrt{\bar{\alpha_t}} # 加噪标准差"sqrtmab": sqrtmab,  # \sqrt{1-\bar{\alpha_t}}  # 加噪均值"mab_over_sqrtmab": mab_over_sqrtmab_inv,  # (1-\alpha_t)/\sqrt{1-\bar{\alpha_t}}}def forward(self, x, c):"""训练过程中, 随机选择step和生成噪声"""# 随机选择step_ts = torch.randint(1, self.n_T + 1, (x.shape[0],)).to(self.device)  # t ~ Uniform(0, n_T)# 随机生成正态分布噪声noise = torch.randn_like(x)  # eps ~ N(0, 1)# 加噪后的图像x_tx_t = (self.sqrtab[_ts, None, None, None] * x+ self.sqrtmab[_ts, None, None, None] * noise)# 将unet预测的对应step的正态分布噪声与真实噪声做对比return self.loss_mse(noise, self.model(x_t, c, _ts / self.n_T))def sample(self, n_sample, c, size, device):# 随机生成初始噪声图片 x_T ~ N(0, 1)x_i = torch.randn(n_sample, *size).to(device)for i in range(self.n_T, 0, -1):t_is = torch.tensor([i / self.n_T]).to(device)t_is = t_is.repeat(n_sample, 1, 1, 1)z = torch.randn(n_sample, *size).to(device) if i > 1 else 0eps = self.model(x_i, c, t_is)x_i = x_i[:n_sample]x_i = self.oneover_sqrta[i] * (x_i - eps * self.mab_over_sqrtmab[i]) + self.sqrt_beta_t[i] * zreturn x_iclass ImageGenerator(object):def __init__(self):'''初始化，定义超参数、数据集、网络结构等'''self.epoch = 20self.sample_num = 100self.batch_size = 256self.lr = 0.0001self.n_T = 400self.device = 'cuda' if torch.cuda.is_available() else 'cpu'self.init_dataloader()self.sampler = DDPM(model=Unet(in_channels=1), betas=(1e-4, 0.02), n_T=self.n_T, device=self.device).to(self.device)self.optimizer = optim.Adam(self.sampler.model.parameters(), lr=self.lr)def init_dataloader(self):'''初始化数据集和dataloader'''tf = transforms.Compose([transforms.ToTensor(),])train_dataset = MNIST('./data/',train=True,download=True,transform=tf)self.train_dataloader = DataLoader(train_dataset, batch_size=self.batch_size, shuffle=True, drop_last=True)val_dataset = MNIST('./data/',train=False,download=True,transform=tf)self.val_dataloader = DataLoader(val_dataset, batch_size=self.batch_size, shuffle=False)def train(self):self.sampler.train()print('训练开始!!')for epoch in range(self.epoch):self.sampler.model.train()loss_mean = 0for i, (images, labels) in enumerate(self.train_dataloader):images, labels = images.to(self.device), labels.to(self.device)labels = F.one_hot(labels, num_classes=10).float()# 将latent和condition拼接后输入网络loss = self.sampler(images, labels)loss_mean += loss.item()self.optimizer.zero_grad()loss.backward()self.optimizer.step()train_loss = loss_mean / len(self.train_dataloader)print('epoch:{}, loss:{:.4f}'.format(epoch, train_loss))self.visualize_results(epoch)@torch.no_grad()def visualize_results(self, epoch):self.sampler.eval()# 保存结果路径output_path = 'results/Diffusion'if not os.path.exists(output_path):os.makedirs(output_path)tot_num_samples = self.sample_numimage_frame_dim = int(np.floor(np.sqrt(tot_num_samples)))labels = F.one_hot(torch.Tensor(np.repeat(np.arange(10), 10)).to(torch.int64), num_classes=10).to(self.device).float()out = self.sampler.sample(tot_num_samples, labels, (1, 28, 28), self.device)save_image(out, os.path.join(output_path, '{}.jpg'.format(epoch)), nrow=image_frame_dim)if __name__ == '__main__':generator = ImageGenerator()generator.train()

本文来自互联网用户投稿，该文观点仅代表作者本人，不代表本站立场。本站仅提供信息存储空间服务，不拥有所有权，不承担相关法律责任。如若转载，请注明出处：http://www.mzph.cn/news/651600.shtml

如若内容造成侵权/违法违规/事实不符，请联系多彩编程网进行投诉反馈email:809451989@qq.com，一经查实，立即删除！

相关文章

LayoutInflater.inflate全面解读

LayoutInflater.inflate全面解读

方法解析 LayoutInflater.inflate() 是 Android 系统中用于将 XML 布局文件转换成相应的 View 的方法。在 Android 开发中&#xff0c;我们经常使用此方法来动态创建和填充布局。 public View inflate(LayoutRes int resource, Nullable ViewGroup root, boolean attachToRoo…
阅读更多...
LVGL v9学习笔记 | 12 - 弧形控件的使用方法（arc）

LVGL v9学习笔记 | 12 - 弧形控件的使用方法（arc）

一、arc控件 arc控件的API在lvgl/src/widgets/arc/lv_arc.h 中声明，以lv_arc_xxx命名。 arc控件由背景圆弧和前景圆弧组成，前景圆弧的末端有一个旋钮，前景圆弧可以被触摸调节。 1. 创建arc对象 /*** Create an arc object* @param parent pointer to an object, it w…
阅读更多...
Pyecharts 风采：从基础到高级，打造炫酷象形柱状图的完整指南【第40篇—python：象形柱状图】

Pyecharts 风采：从基础到高级，打造炫酷象形柱状图的完整指南【第40篇—python：象形柱状图】

文章目录 引言安装PyechartsPyecharts象形柱状图参数详解1. Bar 类的基本参数2. 自定义图表样式3. 添加标签和提示框 代码实战&#xff1a;绘制多种炫酷象形柱状图进阶技巧&#xff1a;动态数据更新与交互性1. 动态数据更新2. 交互性设计 拓展应用&#xff1a;结合其他图表类型…
阅读更多...
深度学习-使用Labelimg数据标注

深度学习-使用Labelimg数据标注

数据标注是计算机视觉和机器学习项目中至关重要的一步&#xff0c;而使用工具进行标注是提高效率的关键。本文介绍了LabelImg&#xff0c;一款常用的开源图像标注工具。用户可以在图像中方便而准确地标注目标区域&#xff0c;为训练机器学习模型提供高质量的标注数据。LabelImg…
阅读更多...
Unity中URP下逐顶点光照

Unity中URP下逐顶点光照

文章目录 前言一、之前额外灯逐像素光照的数据准备好后&#xff0c;还有最后的处理二、额外灯的逐顶点光照1、逐顶点额外灯的光照颜色2、inputData.vertexLighting3、surfaceData.albedo 前言 在上篇文章中&#xff0c;我们分析了Unity中URP下额外灯&#xff0c;逐像素光照中聚…
阅读更多...
vue3 codemirror关于 sql 和 json格式化的使用以及深入了解codemirror 使用json格式化提示错误的关键代码

vue3 codemirror关于 sql 和 json格式化的使用以及深入了解codemirror 使用json格式化提示错误的关键代码

文章目录 需求说明0、安装1. 导入js脚本2.配置3.html处使用4.js处理数据&#xff08;1&#xff09;json格式化处理&#xff08;2&#xff09;sql格式化处理 5. 解决问题1:json格式化错误提示报错&#xff08;1&#xff09;打开官网&#xff08;2&#xff09;打开官网&#xff0…
阅读更多...
qt学习：http+访问百度智能云api实现人脸识别

qt学习：http+访问百度智能云api实现人脸识别

目录 登录到百度智能云，找到人脸识别 完成操作指引 开通 添加人脸库 查看人脸搜索与库管理的api文档 ​编辑 查看自己应用的api key 查看回应的数据格式 编程实战 配置ui界面 添加模块，头文件和定义变量
阅读更多...
C#中类型装换

C#中类型装换

在C#中&#xff0c;可以使用Convert.ChangeType()方法进行类型转换。这个方法可以将一个对象转换为指定的类型。 以下是使用Convert.ChangeType()方法的示例&#xff1a; using System;public class MyClass {public int MyProperty { get; set; } }public class Program {pu…
阅读更多...
【机器学习笔记】1 线性回归

【机器学习笔记】1 线性回归

回归的概念 二分类问题可以用1和0来表示 线性回归&#xff08;Linear Regression&#xff09;的概念 是一种通过属性的线性组合来进行预测的线性模型&#xff0c;其目的是找到一条直线或者一个平面或者更高维的超平面&#xff0c;使得预测值与真实值之间的误差最小化&#x…
阅读更多...
ppt背景图片怎么设置？让你的演示更加出彩！

ppt背景图片怎么设置？让你的演示更加出彩！

PowerPoint是一款广泛应用于演示文稿制作的软件&#xff0c;而背景图片是演示文稿中不可或缺的一部分。一个好的背景图片能够提升演示文稿的整体效果&#xff0c;使观众更加关注你的演示内容。可是ppt背景图片怎么设置呢&#xff1f;本文将介绍ppt背景图片设置的三个方法&#…
阅读更多...
数据库 sql select *from account where name=‘张三‘ 执行过程

数据库 sql select *from account where name=‘张三‘ 执行过程

select *from account where name张三分析上面语句的执行过程 用到了索引 由于是根据 1.name字段进行查询&#xff0c;所以先根据name张三’到name字段的二级索引中进行匹配查 找。但是在二级索引中只能查找到 Arm 对应的主键值 10。 2.由于查询返回的数据是*&#xff0c…
阅读更多...
5.Hive表修改Location，一次讲明白

5.Hive表修改Location，一次讲明白

Hive表修改Loction 一、Hive中修改Location语句二、方案1 删表重建1. 创建表&#xff0c;写错误的Location2. 查看Location3. 删表4. 创建表&#xff0c;写正确的Location5. 查看Location 三、方案2 直接修改Location并恢复数据1.建表&#xff0c;指定错误的Location&#xff0…
阅读更多...
pytorch 卷积神经网络CNN

pytorch 卷积神经网络CNN

www.ai-xlab.com AI MOOC 卷积神经网络CNN 1 www.ai-xlab.com AI MOOC 卷积神经网络是近年发展起来&#xff0c;并广泛应用于图像处理&#xff0c;NLP等领域的一 种多层神经网络。 传统BP处理图像时的问题&#xff1a; 1.权值太多&#xff0c;计算量太大 2.权值太多&a…
阅读更多...
ReactHooks 官网文档翻译

ReactHooks 官网文档翻译

useCallback(fn, dependencies) useCallback是一个React Hook&#xff0c;它允许您在重新渲染之间缓存函数定义。 const cachedFn useCallback(fn, dependencies)1、参数&#xff1a; fn&#xff1a;要缓存的函数值。它可以接受任何参数并返回任何值。React将在初始渲染期间…
阅读更多...
【CSS】实现鼠标悬停图片放大的几种方法

【CSS】实现鼠标悬停图片放大的几种方法

1.背景图片放大 使用css设置背景图片大小100%&#xff0c;同时设置位置和过渡效果&#xff0c;然后使用&#xff1a;hover设置当鼠标悬停时修改图片大小&#xff0c;实现悬停放大效果。 <!DOCTYPE html> <html lang"en"> <head><meta charset…
阅读更多...
vue项目使用element-plus

vue项目使用element-plus

介绍 1.element Plus 是一套基于 Vue.js 的组件库&#xff0c;是对饿了么团队的 Element UI 组件库的升级版本。Element Plus 的目标是提供一套更为现代、更好用的 Vue.js UI 组件。 导入 1.1 执行命令: npm install element-plus --save 1.2 在main.js中做如下配置import E…
阅读更多...
###C语言程序设计-----C语言学习（4）#

###C语言程序设计-----C语言学习（4）#

前言&#xff1a;感谢老铁的浏览&#xff0c;希望老铁可以一键三连加个关注&#xff0c;您的支持和鼓励是我前进的动力&#xff0c;后续会分享更多学习编程的内容。现在开始今天的内容&#xff1a; 一. 主干知识的学习 1.字符型数据 &#xff08;1&#xff09;字符型常量 字…
阅读更多...
Leetcode541反转字符串Ⅱ（java实现）

Leetcode541反转字符串Ⅱ（java实现）

我们今天分享的题目是字符串反转的进阶版反转字符串Ⅱ。 我们首先来看题目描述&#xff1a; 乍一看题目&#xff0c;有种懵逼的感觉&#xff0c;不要慌&#xff0c;博主来带着你分析题目&#xff0c;题目要求&#xff1a; 1. 每隔2k个字符&#xff0c;就对2k字符中的前k个字符…
阅读更多...
自监督学习

自监督学习

自监督学习指的是不依赖标注数据&#xff0c;模型直接从无标注的大量数据中进行学习。【属于无监督学习的一部分&#xff0c;它在训练过程中有正负样本&#xff0c;但是正负样本不是人工标号得来的&#xff0c;而是通过自己产生的。】 自监督学习总体分成两类&#xff1a;生成式…
阅读更多...
muduo源码阅读笔记（11、TcpClient）

muduo源码阅读笔记（11、TcpClient）

muduo源码阅读笔记&#xff08;11、TcpClient&#xff09; Muduo源码笔记系列&#xff1a; muduo源码阅读笔记&#xff08;0、下载编译muduo&#xff09; muduo源码阅读笔记&#xff08;1、同步日志&#xff09; muduo源码阅读笔记&#xff08;2、对C语言原生的线程安全以及…
阅读更多...
最新文章

Copyright @ 2022~2023