1.预处理函数（这里有一个问题，transforms.Normalize（）函数的使用）
有博主写了/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，进行批归一化（这点与传统的方式不一样）+LeaKy ReLU
3.定义生成器（理解为鉴别器的逆过程，反卷积的过程）
4.训练（初始化参数）
def train(d,g,criterion,d_optimizer,g_optimizer,epochs=1,show_every=1000,print_every=10):iter_count = 0for epoch in range(epochs):for inputs,_ in trainloader:real_inputs = inputsfake_inputs = g(torch.randn(5,100))real_labels = torch.ones(real_inputs.size(0))fake_labels = torch.zeros(5)real_outputs = d(real_inputs)d_loss_real = criterion(real_outputs,real_labels)real_scores = real_outputsfake_outputs = d(fake_inputs)d_loss_fake = criterion(fake_outputs,fake_labels)fake_scores = fake_outputsd_loss = d_loss_real+d_loss_faked_optimizer.zero_grad()d_loss.backward()d_optimizer.step()fake_inputs = g(torch.randn(5,100))outputs = d(fake_inputs)real_labels = torch.ones(outputs.size(0))g_loss = criterion(outputs,real_labels)g_optimizer.zero_grad()g_loss.backward()g_optimizer.step()if (iter_count % show_every == 0):print('Epoch:{},Iter: {}, D: {:.4}, G:{:.4}'.format(epoch,iter_count, d_loss.item(), g_loss.item()))picname = "Epoch_"+str(epoch)+"Iter_"+str(iter_count)imshow(torchvision.utils.make_grid(fake_inputs.data),picname)save_param(d,'d_model.pkl')save_param(g,'g_model.pkl')if (iter_count%print_every == 0):print('Epoch:{},Iter: {}, D: {:.4}, G:{:.4}'.format(epoch,iter_count, d_loss.item(), g_loss.item()))iter_count += 1print('Finished Training')
这里就是训练的过程：
最后训练就完事了，结果我没有完全跑出来，时间太长不想跑。
跑了一部分结果出来，大概是这样的效果啦！！
