如何查网站是哪家公司做的,物流案例 网站,网络优化网络维护网络建站,广州搬家公司关于 近年来#xff0c;基于卷积网络#xff08;CNN#xff09;的监督学习已经 在计算机视觉应用中得到了广泛的采用。相比之下#xff0c;无监督 使用 CNN 进行学习受到的关注较少。在这项工作中#xff0c;我们希望能有所帮助 缩小了 CNN 在监督学习和无监督学习方面的成…关于 近年来基于卷积网络CNN的监督学习已经 在计算机视觉应用中得到了广泛的采用。相比之下无监督 使用 CNN 进行学习受到的关注较少。在这项工作中我们希望能有所帮助 缩小了 CNN 在监督学习和无监督学习方面的成功之间的差距。我们介绍一类称为深度卷积生成的 CNN 对抗性网络DCGAN具有一定的架构限制以及 证明他们是无监督学习的有力候选人。训练 在各种图像数据集上我们展示了令人信服的证据表明我们的深度卷积对抗对学习了从对象部分到 生成器和鉴别器中的场景。此外我们使用学到的 新任务的特征 - 证明它们作为一般图像表示的适用性。https://arxiv.org/pdf/1511.06434.pdf
工具 数据集 方法实现
加载必要的库函数和自定义函数
import torch
import torchvision
import torch.nn as nn
import torch.nn.functional as Ffrom torch.utils.data import DataLoader
from torchvision import datasets
from torchvision import transforms
from torchvision.utils import save_image
def get_sample_image(G, n_noise):save sample 100 imagesz torch.randn(100, n_noise).to(DEVICE)y_hat G(z).view(100, 28, 28) # (100, 28, 28)result y_hat.cpu().data.numpy()img np.zeros([280, 280])for j in range(10):img[j*28:(j1)*28] np.concatenate([x for x in result[j*10:(j1)*10]], axis-1)return img
定义判别模型
class Discriminator(nn.Module):Convolutional Discriminator for MNISTdef __init__(self, in_channel1, num_classes1):super(Discriminator, self).__init__()self.conv nn.Sequential(# 28 - 14nn.Conv2d(in_channel, 512, 3, stride2, padding1, biasFalse),nn.BatchNorm2d(512),nn.LeakyReLU(0.2),# 14 - 7nn.Conv2d(512, 256, 3, stride2, padding1, biasFalse),nn.BatchNorm2d(256),nn.LeakyReLU(0.2),# 7 - 4nn.Conv2d(256, 128, 3, stride2, padding1, biasFalse),nn.BatchNorm2d(128),nn.LeakyReLU(0.2),nn.AvgPool2d(4),)self.fc nn.Sequential(# reshape input, 128 - 1nn.Linear(128, 1),nn.Sigmoid(),)def forward(self, x, yNone):y_ self.conv(x)y_ y_.view(y_.size(0), -1)y_ self.fc(y_)return y_
定义生成模型
class Generator(nn.Module):Convolutional Generator for MNISTdef __init__(self, input_size100, num_classes784):super(Generator, self).__init__()self.fc nn.Sequential(nn.Linear(input_size, 4*4*512),nn.ReLU(),)self.conv nn.Sequential(# input: 4 by 4, output: 7 by 7nn.ConvTranspose2d(512, 256, 3, stride2, padding1, biasFalse),nn.BatchNorm2d(256),nn.ReLU(),# input: 7 by 7, output: 14 by 14nn.ConvTranspose2d(256, 128, 4, stride2, padding1, biasFalse),nn.BatchNorm2d(128),nn.ReLU(),# input: 14 by 14, output: 28 by 28nn.ConvTranspose2d(128, 1, 4, stride2, padding1, biasFalse),nn.Tanh(),)def forward(self, x, yNone):x x.view(x.size(0), -1)y_ self.fc(x)y_ y_.view(y_.size(0), 512, 4, 4)y_ self.conv(y_)return y_ 模型超参数定义配置
batch_size 64criterion nn.BCELoss()
D_opt torch.optim.Adam(D.parameters(), lr0.001, betas(0.5, 0.999))
G_opt torch.optim.Adam(G.parameters(), lr0.001, betas(0.5, 0.999))max_epoch 30 # need more than 20 epochs for training generator
step 0
n_critic 1 # for training more k steps about Discriminator
n_noise 100D_labels torch.ones([batch_size, 1]).to(DEVICE) # Discriminator Label to real
D_fakes torch.zeros([batch_size, 1]).to(DEVICE) # Discriminator Label to fake 模型训练
for epoch in range(max_epoch):for idx, (images, labels) in enumerate(data_loader):# Training Discriminatorx images.to(DEVICE)x_outputs D(x)D_x_loss criterion(x_outputs, D_labels)z torch.randn(batch_size, n_noise).to(DEVICE)z_outputs D(G(z))D_z_loss criterion(z_outputs, D_fakes)D_loss D_x_loss D_z_lossD.zero_grad()D_loss.backward()D_opt.step()if step % n_critic 0:# Training Generatorz torch.randn(batch_size, n_noise).to(DEVICE)z_outputs D(G(z))G_loss criterion(z_outputs, D_labels)D.zero_grad()G.zero_grad()G_loss.backward()G_opt.step()if step % 500 0:print(Epoch: {}/{}, Step: {}, D Loss: {}, G Loss: {}.format(epoch, max_epoch, step, D_loss.item(), G_loss.item()))if step % 1000 0:G.eval()img get_sample_image(G, n_noise)imsave(./{}_step{}.jpg.format(MODEL_NAME, str(step).zfill(3)), img, cmapgray)G.train()step 1 测试生成效果
# generation to image
G.eval()
imshow(get_sample_image(G, n_noise), cmapgray) 模型和状态参量保存
def save_checkpoint(state, file_namecheckpoint.pth.tar):torch.save(state, file_name)# Saving params.
# torch.save(D.state_dict(), D_c.pkl)
# torch.save(G.state_dict(), G_c.pkl)
save_checkpoint({epoch: epoch 1, state_dict:D.state_dict(), optimizer : D_opt.state_dict()}, D_dc.pth.tar)
save_checkpoint({epoch: epoch 1, state_dict:G.state_dict(), optimizer : G_opt.state_dict()}, G_dc.pth.tar)
应用
DCGAN作为一个成熟的生成模型在自然图像医学图像医学电生理信号数据分析中都可以用来实现数据的合成达到数据增强的目的同时如何减少增强数据对于后端任务的不利干扰也是一个需要关注的方面。
