为什么我的 Pytorch 训练会出现错误：“trying to backward through the graph a second time...”

Question

我试过在一些输出张量上调用

.detach()

，但这似乎不起作用。作为参考，我在训练 GAN 时遇到了这个错误。这是代码：

def main():
    # Encoder In: BxDxN
    # Encoder Out: BxNxK
    # Decoder In: BxNxK
    # Decoder Out: BxDxN

    const = Constants()
    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

    # Autoencoder training setup
    AE = USuRPNet(point_per_cloud=43)
        #train_loader = DataLoader(PC_Data("ModelNet10"), batch_size = 128)
    print("The autoencoder will be running on", device, "device")
    AE.to(device)
    sample = torch.randn(2000, 43, 3)
    AE_train_loader = DataLoader(sample, batch_size=const.batch_size)
    AE_loss = chamfer_distance
    AE_optimizer = torch.optim.Adam(AE.parameters(), lr = 0.001, weight_decay=0.0001)
    
    for epoch in range(const.num_epochs):
        run_loss = 0.0

        for i, pc in enumerate(AE_train_loader):
            # Train the autoencoder
            pc = pc.cuda()
            AE_optimizer.zero_grad()
            recon, latent_space = AE(pc)
            loss = AE_loss(recon, pc)[0]
            loss.backward()
            AE_optimizer.step()
            run_loss += loss.item()

        print(f'epoch: {epoch+1} ~ loss: {run_loss}')
        run_loss = 0.0

    # Generative Adversarial Network training setup
    gen = Generator()
    disc = Discriminator()
    print("The GAN will be running on", device, "device")
    gen.to(device)
    disc.to(device)
    _, latent_space = AE(AE_train_loader.dataset.cuda())
    GAN_train_loader = DataLoader(latent_space, batch_size=const.batch_size)
    GAN_loss = nn.BCELoss()
    disc_optimizer = torch.optim.Adam(disc.parameters(), lr=0.001)
    gen_optimizer = torch.optim.Adam(gen.parameters(), lr=0.001)

    for epoch in range(const.num_epochs):
        run_loss = 0.0

        for i, batch in enumerate(GAN_train_loader):
            # Train the discriminator
            real_sample_labels = torch.ones((const.batch_size, 1)).cuda()
            real_samples = batch
            gen_sample_labels = torch.zeros((const.batch_size, 1)).cuda()
            gen_samples = gen(torch.randn((const.batch_size, 32)).cuda())
            all_samples = torch.cat((real_samples, gen_samples))
            all_samples_labels = torch.cat((real_sample_labels, gen_sample_labels))

            disc.zero_grad()
            disc_out = disc(all_samples.cuda())
            disc_loss = GAN_loss(disc_out, all_samples_labels)
            disc_loss.requires_grad = True
            disc_loss.backward()
            disc_optimizer.step()

            # Train the generator
            rand_vecs = torch.rand((const.batch_size, 32))

            gen.zero_grad()
            gen_samples = gen(rand_vecs.cuda())
            out_disc_gen = disc(gen_samples.cuda())
            gen_loss = GAN_loss(out_disc_gen, real_sample_labels)
            gen_loss.backward()
            gen_optimizer.step()

            run_loss+=gen_loss+disc_loss

        print(f'epoch: {epoch+1} ~ loss: {run_loss}')
        run_loss = 0.0

if __name__=="__main__":
    main()

错误出现在GAN模型的第二个大循环中。该站点上的其他一些问题的答案是分离图形或不指定

retain_graph=True

（但我听说这会增加运行时间）。