效果展示
效果如下所示
实现步骤
1.爬取犬夜叉数据
2.截取脸部数据
3.构造网络训练。
其中大多数代码为参考已有的博客，只在调参部分有所不同。
当我按照原博调参时，判断器被训练的太强，导致生成器生成出所有的都是被一眼顶针-鉴定为假，所以，我选择，当假图片平均分大于0.5 或者 真图片平均分小于 0.5 时，训练判断器，当假图片得分小于0.3时训练生成网络。
另外，特别注意，谷歌图片质量明显高于百度，有条件优先爬谷歌
附代码
import globimport osimport cv2import visiualimport torchvisionimport torchfrom torch import nnfrom torch.utils.data import Datasetfrom torch.autograd import Variableimport torch.utils.data as Dataimport torchvision.transforms as transformersimport torchvision.transforms as tfsfrom torch.utils.data import DataLoader, samplerfrom torchvision.datasets import MNISTimport PIL.Image as Imageimport numpy as npimport torchvision.transforms as transformsimport matplotlib.pyplot as pltimport matplotlib.gridspec as gridspecplt.rcParams['figure.figsize'] = (100.0, 80.0) # 设置画图的尺寸plt.rcParams['image.interpolation'] = 'nearest'plt.rcParams['image.cmap'] = 'gray'class MyDataset(Dataset):def __init__(self,root_path):self.path = root_path# data = []# images = glob.glob((path+'*'))#取出所有照片名称self.img_path = os.listdir(self.path)def __len__(self):return len(self.img_path)def __getitem__(self, index):img_name =self.img_path[index]img_item_path = os.path.join(self.path,img_name)img = Image.open(img_item_path)# img = img.resize((500,500))#用来改为指定大小img = np.array(img)if img.shape[2]==4:img= img[:,:,:3]label =1return img, labeldef show_images(images): # 定义画图工具images = np.reshape(images, [images.shape[0], -1])sqrtn = int(np.ceil(np.sqrt(images.shape[0])))sqrtimg = int(np.ceil(np.sqrt(images.shape[1])))fig = plt.figure(figsize=(sqrtn, sqrtn))gs = gridspec.GridSpec(sqrtn, sqrtn)gs.update(wspace=0.05, hspace=0.05)for i, img in enumerate(images):ax = plt.subplot(gs[i])plt.axis('off')ax.set_xticklabels([])ax.set_yticklabels([])ax.set_aspect('equal')# plt.imshow(img.reshape([sqrtimg, sqrtimg]))plt.savefig('./pic/data.png')plt.close()returndef preprocess_img(x):x = tfs.ToTensor()(x)return (x - 0.5) / 0.5def deprocess_img(x):# return (x + 1.0) / 2.0return xclass ChunkSampler(sampler.Sampler): # 定义一个取样的函数"""Samples elements sequentially from some offset.Arguments:num_samples: # of desired datapointsstart: offset where we should start selecting from"""def __init__(self, num_samples, start=0):self.num_samples = num_samplesself.start = startdef __iter__(self):return iter(range(self.start, self.start + self.num_samples))def __len__(self):return self.num_samples# 判决网络# 判别器class NetD(nn.Module):#opt 是通道数 先是个5def __init__(self, opt):super(NetD, self).__init__()ndf = opt.ndfself.main = nn.Sequential(# 输入3*96*96即生成器生成的图片nn.Conv2d(3, ndf, kernel_size=5, stride=3, padding=1, bias=False), # 卷积，下采样，也是编码的过程nn.LeakyReLU(0.2, inplace=True),# 输出 ndf x 32 x32nn.Conv2d(ndf, ndf * 2, kernel_size=4, stride=2, padding=1, bias=False), # 正好将feature map图片大小缩小一半nn.BatchNorm2d(ndf * 2),nn.LeakyReLU(0.2, inplace=True),# 输出 (ndf*2) x 16 x 16nn.Conv2d(ndf * 2, ndf * 4, kernel_size=4, stride=2, padding=1, bias=False),nn.BatchNorm2d(ndf * 4),nn.LeakyReLU(0.2, inplace=True),# 输出 (ndf*4) x 8 x 8nn.Conv2d(ndf * 4, ndf * 8, kernel_size=4, stride=2, padding=1, bias=False),nn.BatchNorm2d(ndf * 8),nn.LeakyReLU(0.2, inplace=True),# 输出 (ndf*8) x 4 x 4nn.Conv2d(ndf * 8, 1, kernel_size=4, stride=1, padding=0, bias=False), # 最后编码成为一个维度为1的向量nn.Sigmoid() # 最后用Sigmoid作为分类，使得判别器成为一个判断二分类问题的模型，实际上判别器也是做一个二分类任务，判断是否为原图输出0或1)def forward(self, input):return self.main(input).view(-1) # 转成一个列向量，即sigmoid的结果在更前面的维度# 生成网络# 生成器class OPT():def __init__(self,):self.ndf =64self.ngf = 64self.nz =100class NetG(nn.Module):def __init__(self, opt):super(NetG, self).__init__()ngf = opt.ngf # 生成器feature map数# 生成器主要的网络模块self.main = nn.Sequential(# 输入是一个nz维的噪音，是一个随机生成的张量，可以认为是大小为1x1的feature ampnn.ConvTranspose2d(opt.nz, ngf * 8, kernel_size=4, stride=1, padding=0, bias=False), # 反卷积，做上采样nn.BatchNorm2d(ngf * 8),nn.ReLU(True),# 上一步的输出形状：(ngf*8) x 4 x 4nn.ConvTranspose2d(ngf * 8, ngf * 4, kernel_size=4, stride=2, padding=1, bias=False), # 继续上采样，不断减小图片维度nn.BatchNorm2d(ngf * 4),nn.ReLU(True),# 上一步的输出形状： (ngf*4) x 8 x 8nn.ConvTranspose2d(ngf * 4, ngf * 2, kernel_size=4, stride=2, padding=1, bias=False),nn.BatchNorm2d(ngf * 2),nn.ReLU(True),# 上一步的输出形状： (ngf*2) x 16 x 16nn.ConvTranspose2d(ngf * 2, ngf, kernel_size=4, stride=2, padding=1, bias=False),nn.BatchNorm2d(ngf),nn.ReLU(True),# 上一步的输出形状：(ngf) x 32 x 32nn.ConvTranspose2d(ngf, 3, kernel_size=5, stride=3, padding=1, bias=False),nn.Tanh() # 输出范围固定在 -1 ~ 1故而采用Tanh# 输出形状：3 x 96 x 96# feature map经过解码过程，最后生成一个图片)def forward(self, input):return self.main(input)bce_loss = nn.BCEWithLogitsLoss()def discriminator_loss(logits_real, logits_fake): # 判别器的 losssize = logits_real.shape[0]true_labels = Variable(torch.ones(size, 1)).float().cuda()false_labels = Variable(torch.zeros(size, 1)).float().cuda()loss = bce_loss(logits_real, true_labels) + bce_loss(logits_fake, false_labels)return lossdef generator_loss(logits_fake): # 生成器的 losssize = logits_fake.shape[0]true_labels = Variable(torch.ones(size, 1)).float().cuda()loss = bce_loss(logits_fake, true_labels)return loss# 使用 adam 来进行训练，学习率是 3e-4, beta1 是 0.5, beta2 是 0.999def get_optimizer(net,g=True):if g==True:optimizer = torch.optim.Adam(net.parameters(), lr=3e-5,betas=(0.5, 0.999))#) betas=(0.5, 0.999))else:optimizer = torch.optim.Adam(net.parameters(), lr=3e-5,betas=(0.5, 0.999))#, betas=(0.5, 0.999))return optimizerdef train_a_gan(D_net, G_net, D_optimizer, G_optimizer, discriminator_loss, generator_loss, show_every=10,noise_size=96, num_epochs=8000):iter_count = 0for epoch in range(num_epochs):g_loss =[]d_loss = []g_error, d_total_error = 0,0for i, (x, _) in enumerate(train_data):# x = x.permute(0,3,1,2)x = x.type(torch.cuda.FloatTensor)bs = x.shape[0]# 判别网络# real_data = Variable(x).view(bs, -1).cuda() # 真实数据logits_real = D_net(x) .unsqueeze(1) # 判别网络得分sample_noise = (torch.rand((bs, opt.nz,1,1)) - 0.5) / 0.5 # -1 ~ 1 的均匀分布g_fake_seed = Variable(sample_noise).cuda()fake_images = G_net(g_fake_seed) # 生成的假的数据logits_fake = D_net(fake_images).unsqueeze(1) # 判别网络得分if (i%1==0 and float(torch.mean(logits_real).cpu())<0.5) or float(torch.mean(logits_fake).cpu())>0.5 :d_total_error = discriminator_loss(logits_real, logits_fake) # 判别器的 lossD_optimizer.zero_grad()d_total_error.backward()D_optimizer.step() # 优化判别网络d_loss.append(float(d_total_error.cpu()))if np.random.rand()<0.011:# print('logits_real{:.2f}, logits_fake{:.2f}'.format(float(torch.mean(logits_real).cpu()),float(torch.mean(logits_fake).cpu())))print('logits_real{} , logits_fake{} '.format( torch.mean(logits_real).cpu() ,torch.mean(logits_fake).cpu()) )# 生成网络g_fake_seed = Variable(sample_noise).cuda()fake_images = G_net(g_fake_seed) # 生成的假的数据gen_logits_fake = D_net(fake_images).unsqueeze(1)if i %5==0 and float(torch.mean(logits_real).cpu())>0.3:# print('2')g_error = generator_loss(gen_logits_fake) # 生成网络的 lossG_optimizer.zero_grad()g_error.backward()G_optimizer.step() # 优化生成网络g_loss.append(float(g_error.cpu()))if (epoch % 50 == 0):torch.save(G_net,'params_epoch/2/'+'G_net'+str(epoch)+'.pkl')torch.save(D_net, 'params_epoch/2/' +'N_net'+ str(epoch) + '.pkl')if (epoch %5 == 0):fake_face= (((fake_images[np.random.randint(0,3), :, :, :] * 0.5 + 0.5) * 255).type(torch.int16)).cpu().permute(1,2,0).numpy()fake_face=fake_face.astype(np.uint8)# cv2.imshow('face',fake_face)cv2.imwrite('result/2/'+str(epoch)+'.jpg',fake_face)iter_count += 1print('epoch {} Gloss {} , Dloss_D {}'.format(epoch, np.mean(g_loss), np.mean(d_loss)))img_transforms=pose([transforms.ToTensor(),transformers.RandomHorizontalFlip(),transformers.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))# transforms.Grayscale(num_output_channels=1)])NUM_TRAIN = 50000NUM_VAL = 5000NOISE_DIM = 96batch_size = 8#数据增强# train_set = MyDataset('Inuyasha/data/0/')train_set = torchvision.datasets.ImageFolder('Inuyasha/data',transform=img_transforms)train_data = DataLoader(train_set, batch_size=batch_size)# train_data = DataLoader(train_set, batch_size=batch_size, sampler=ChunkSampler(NUM_TRAIN, 0))# val_set = MyDataset('Inuyasha/data')# val_data = DataLoader(val_set, batch_size=batch_size)# imgs = deprocess_img(train_data.__iter__().next()[0].view(batch_size, -1)).numpy().squeeze() # 可视化图片效果# show_images(imgs)# experiment6/params_epoch/1/G_net0.pklopt = OPT()D=torch.load('params_epoch/1/N_net0.pkl')G=torch.load('params_epoch/1/G_net0.pkl')# D = NetD(opt=opt).cuda()# G =NetG(opt).cuda()D_optim = get_optimizer(D,False)G_optim = get_optimizer(G)# vis = visiual.Visualizer()train_a_gan(D, G, D_optim, G_optim, discriminator_loss, generator_loss)
