PyTorch(二次元头像GAN实战)

1.预处理函数&＃xff08;这里有一个问题&＃xff0c;transforms.Normalize&＃xff08;&＃xff09;函数的使用&＃xff09;

有博主写了 https://blog.csdn.net/xys430381_1/article/details/85724668

2.定义鉴别器

计算方式参考上一张 迁移学习的内容。最后输出向量的类型为256*6*6.

同时每次卷积过后&＃xff0c;进行批归一化&＃xff08;这点与传统的方式不一样&＃xff09;&＃43;LeaKy ReLU

 3.定义生成器&＃xff08;理解为鉴别器的逆过程&＃xff0c;反卷积的过程&＃xff09;

4.训练&＃xff08;初始化参数&＃xff09;

def train(d,g,criterion,d_optimizer,g_optimizer,epochs&＃61;1,show_every&＃61;1000,print_every&＃61;10):iter_count &＃61; 0for epoch in range(epochs):for inputs,_ in trainloader:real_inputs &＃61; inputsfake_inputs &＃61; g(torch.randn(5,100))real_labels &＃61; torch.ones(real_inputs.size(0))fake_labels &＃61; torch.zeros(5)real_outputs &＃61; d(real_inputs)d_loss_real &＃61; criterion(real_outputs,real_labels)real_scores &＃61; real_outputsfake_outputs &＃61; d(fake_inputs)d_loss_fake &＃61; criterion(fake_outputs,fake_labels)fake_scores &＃61; fake_outputsd_loss &＃61; d_loss_real&＃43;d_loss_faked_optimizer.zero_grad()d_loss.backward()d_optimizer.step()fake_inputs &＃61; g(torch.randn(5,100))outputs &＃61; d(fake_inputs)real_labels &＃61; torch.ones(outputs.size(0))g_loss &＃61; criterion(outputs,real_labels)g_optimizer.zero_grad()g_loss.backward()g_optimizer.step()if (iter_count % show_every &＃61;&＃61; 0):print(&＃39;Epoch:{},Iter: {}, D: {:.4}, G:{:.4}&＃39;.format(epoch,iter_count, d_loss.item(), g_loss.item()))picname &＃61; "Epoch_"&＃43;str(epoch)&＃43;"Iter_"&＃43;str(iter_count)imshow(torchvision.utils.make_grid(fake_inputs.data),picname)save_param(d,&＃39;d_model.pkl&＃39;)save_param(g,&＃39;g_model.pkl&＃39;)if (iter_count%print_every &＃61;&＃61; 0):print(&＃39;Epoch:{},Iter: {}, D: {:.4}, G:{:.4}&＃39;.format(epoch,iter_count, d_loss.item(), g_loss.item()))iter_count &＃43;&＃61; 1print(&＃39;Finished Training&＃39;)

这里就是训练的过程&＃xff1a;

最后训练就完事了&＃xff0c;结果我没有完全跑出来&＃xff0c;时间太长不想跑。

跑了一部分结果出来&＃xff0c;大概是这样的效果啦&＃xff01;&＃xff01;