Esempio n. 1
0
if not os.path.exists(check_root):
    os.mkdir(check_root)

img_size = 128
bsize = 512
nz = 100
ngf = 64
ndf = 64
nc = 3
l = 5

net_g = Net_G(nz, ngf, nc, l)
net_g.cuda()

net_d = Net_D(ndf, nc, l)
net_d.cuda()

dataset = dset.ImageFolder(root=data_root,
                           transform=transforms.Compose([
                               transforms.Resize(img_size),
                               transforms.CenterCrop(img_size),
                               transforms.ToTensor(),
                               transforms.Normalize((0.5, 0.5, 0.5),
                                                    (0.5, 0.5, 0.5)),
                           ]))

loader = torch.utils.data.DataLoader(dataset,
                                     batch_size=bsize,
                                     shuffle=True,
                                     num_workers=4)
Esempio n. 2
0
class Solver(object):
    def __init__(self, config, source_loader, target_loader,
                 target_val_loader):
        self.source_loader = source_loader
        self.target_loader = target_loader
        self.target_val_loader = target_val_loader
        self.net = None
        self.net_optimizer = None
        self.net_d = None
        self.net_optimizer_d = None
        self.beta1 = config.beta1
        self.beta2 = config.beta2
        self.train_iters = config.train_iters
        self.pretrain_iters = config.pretrain_iters
        self.batch_size = config.batch_size
        self.lr = config.lr
        self.lr_d = config.lr_d
        self.alpha_s = config.alpha_s
        self.alpha_t = config.alpha_t
        self.beta_c = config.beta_c
        self.beta_sep = config.beta_sep
        self.beta_p = config.beta_p
        self.log_step = config.log_step
        self.model_path = config.model_path
        self.num_classes = config.num_classes
        self.build_model()

    def build_model(self):
        """Builds a generator and a discriminator."""
        self.net = Net()
        self.net_d = Net_D()

        net_params = list(self.net.parameters())
        net_d_params = list(self.net_d.parameters())
        self.net_optimizer = optim.Adam(net_params, self.lr,
                                        [self.beta1, self.beta2])
        self.net_optimizer_d = optim.Adam(net_d_params, self.lr_d,
                                          [self.beta1, self.beta2])

        if torch.cuda.is_available():
            self.net.cuda()
            self.net_d.cuda()

    def to_var(self, x):
        """Converts numpy to variable."""
        if torch.cuda.is_available():
            x = x.cuda()
        return Variable(x, requires_grad=False)

    def to_data(self, x):
        """Converts variable to numpy."""
        if torch.cuda.is_available():
            x = x.cpu()
        return x.data.numpy()

    def reset_grad(self):
        """Zeros the gradient buffers."""
        self.net_optimizer.zero_grad()
        self.net_optimizer_d.zero_grad()

    def separability_loss(self, labels, latents, imbalance_parameter=1):
        criteria = torch.nn.modules.loss.CosineEmbeddingLoss()
        loss_up = 0
        one_cuda = torch.ones(1).cuda()
        mean = torch.mean(latents, dim=0).cuda().view(1, -1)
        loss_down = 0
        for i in range(self.num_classes):
            indexes = labels.eq(i)
            mean_i = torch.mean(latents[indexes], dim=0).view(1, -1)
            if str(mean_i.norm().item()) != 'nan':
                for latent in latents[indexes]:
                    loss_up += criteria(latent.view(1, -1), mean_i, one_cuda)
                loss_down += criteria(mean, mean_i, one_cuda)
        loss = (loss_up / loss_down) * imbalance_parameter
        return loss

    def initialisation(self):
        self.net.apply(xavier_weights_init)
        self.net_d.apply(xavier_weights_init)
        source_iter = iter(self.source_loader)
        target_iter = iter(self.target_loader)
        target_val_iter = iter(self.target_val_loader)
        source_per_epoch = len(source_iter)
        target_per_epoch = len(target_iter)
        targetval_per_epoch = len(target_val_iter)
        print(source_per_epoch, target_per_epoch, targetval_per_epoch)

        criterion = nn.CrossEntropyLoss()

        f_labels = torch.LongTensor(128)
        f_labels[...] = 10

        t_labels = torch.LongTensor(128)
        t_labels[...] = 1

        # pretrain
        log_pre = 50
        source_iter = iter(self.source_loader)
        target_iter = iter(self.target_loader)
        return criterion, source_per_epoch, target_per_epoch, target_iter, source_iter, log_pre

    def train(self):
        criterion, source_per_epoch, target_per_epoch, target_iter, source_iter, log_pre = self.initialisation(
        )

        pre_train = not os.path.exists(
            os.path.join(self.model_path, 'pre_train.pth'))

        print("Pretrain:\n*********")
        if pre_train:
            for step in range(self.pretrain_iters + 1):
                # ============ Initialization ============#
                # refresh
                if (step + 1) % (source_per_epoch) == 0:
                    source_iter = iter(self.source_loader)
                if (step + 1) % (target_per_epoch) == 0:
                    target_iter = iter(self.target_loader)
                # load the data
                source, s_labels = source_iter.next()
                target, t_labels = target_iter.next()
                target_rgb = target
                target, t_labels = self.to_var(target_rgb), self.to_var(
                    t_labels).long().squeeze()
                source, s_labels = self.to_var(source), self.to_var(
                    s_labels).long().squeeze()

                # ============ Training ============ #
                self.reset_grad()
                # forward
                latent, c = self.net(source)
                # loss
                loss_source_class = criterion(c, s_labels)

                # one step
                loss_source_class.backward()
                self.net_optimizer.step()
                self.reset_grad()
                # ============ Validation ============ #
                if (step + 1) % log_pre == 0:
                    _, c_source = self.net(source)
                    _, c_target = self.net(target)
                    print("[%d/20000] classification loss: %.4f" %
                          (step + 1, loss_source_class.item()))
                    print("source accuracy  %.4f;  target accuracy %.4f" %
                          (accuracy(s_labels,
                                    c_source), accuracy(t_labels, c_target)))

            self.save_model()
        else:
            self.load_model()

        # ============ Initialization ============ #
        source_iter = iter(self.source_loader)
        target_iter = iter(self.target_loader)
        maxacc = 0.0
        maximum_acc = 0.0
        max_iter = 0
        net_params = list(self.net.parameters())
        net_d_params = list(self.net_d.parameters())

        self.net_optimizer = optim.Adam(net_params, self.lr,
                                        [self.beta1, self.beta2])
        self.net_optimizer_d = optim.Adam(net_d_params, self.lr_d,
                                          [self.beta1, self.beta2])
        print("Second:\n******")
        self.validate()

        for step in range(self.train_iters):
            # ============ Initialization ============#

            # refresh
            if (step + 1) % (target_per_epoch) == 0:
                target_iter = iter(self.target_loader)
            if (step + 1) % (source_per_epoch) == 0:
                source_iter = iter(self.source_loader)
            # load the data
            source, s_labels = source_iter.next()
            source, s_labels = self.to_var(source), self.to_var(
                s_labels).long().squeeze()  # must squeeze
            target, t_labels = target_iter.next()
            target_rgb = target
            target, t_labels = self.to_var(target_rgb), self.to_var(
                t_labels).long().squeeze()

            # ============ train D ============#
            self.reset_grad()

            latent_source, c = self.net(source)
            d = self.net_d(latent_source)
            loss_d_s1 = F.binary_cross_entropy(
                d, torch.ones_like(d, dtype=torch.float32))
            loss_d_s0 = F.binary_cross_entropy(
                d, torch.zeros_like(d, dtype=torch.float32))
            loss_c_source = criterion(c, s_labels)

            latent_target, c = self.net(target)
            d = self.net_d(latent_target)
            loss_d_t0 = F.binary_cross_entropy(
                d, torch.zeros_like(d, dtype=torch.float32))

            loss_p = loss_d_s0
            loss_d = loss_d_s1 + loss_d_t0
            # ============ train pseudo labeling ============#

            chosen_target, pseudo_labels, indexes, imbalance_parameter = pseudo_labeling(
                target, c)

            if chosen_target is not None:
                loss_c_target = criterion(c[indexes], pseudo_labels)

                latent_target = latent_target[indexes]
                # ============ class loss  ============#
                loss_sep = self.separability_loss(
                    torch.cat((s_labels, pseudo_labels)),
                    torch.cat((latent_source, latent_target)),
                    imbalance_parameter=imbalance_parameter)
            else:
                loss_c_target = 0
                loss_sep = 0
            loss = self.beta_c * (self.alpha_s * loss_c_source + self.alpha_t * loss_c_target) + \
                   self.beta_p * loss_p + \
                   self.beta_sep * loss_sep

            loss.backward(retain_graph=True)
            self.net_optimizer.step()

            loss_d.backward()
            self.net_optimizer_d.step()

            self.reset_grad()

            # ============ Validation ============ #
            if (step + 1) % self.log_step == 0:
                print("max accuracy:", colored(maximum_acc, "green"),
                      "iteration", max_iter)
                _, c_source = self.net(source)
                _, c_target = self.net(target)
                print("source  accuracy  %.4f;  target accuracy %.4f" %
                      (accuracy(s_labels,
                                c_source), accuracy(t_labels, c_target)))
                acc = self.validate()
                if acc > maximum_acc:
                    maximum_acc = acc
                    max_iter = step

                if acc > maxacc:
                    maxacc = acc
                    torch.save(
                        self.net,
                        "./model_c_" + str(step) + '_' + str(acc) + ".pth")
                    torch.save(
                        self.net_d,
                        "./model_d_" + str(step) + '_' + str(acc) + ".pth")
            self.reset_grad()
        # ============ Save the model ============ #
        torch.save(self.net, "./model_c_final.pth")
        torch.save(self.net_d, "./model_d_final.pth")

    def validate(self, ):
        class_correct = [0] * self.num_classes
        class_total = [0.] * self.num_classes
        classes = [str(i) for i in range(self.num_classes)]
        self.net.eval()  # prep model for evaluation

        for data, target in self.target_val_loader:

            # forward pass: compute predicted outputs by passing inputs to the model
            data, target = self.to_var(data), self.to_var(
                target).long().squeeze()

            data = data.cuda()
            target = target.cuda()

            latent, output = self.net(data)
            _, pred = torch.max(output, 1)
            correct = np.squeeze(pred.eq(target.data.view_as(pred)))
            # calculate test accuracy for each object class
            for i in range(len(target.data)):
                label = target.data[i]
                class_correct[label] += correct[i].item()
                class_total[label] += 1

        for i in range(self.num_classes):
            if class_total[i] > 0:
                print('Test Accuracy of %5s: %2d%% (%2d/%2d)' %
                      (str(i), 100 * class_correct[i] / class_total[i],
                       np.sum(class_correct[i]), np.sum(class_total[i])))
            else:
                print('Test Accuracy of %5s: N/A (no training examples)' %
                      (classes[i]))

        print("\nTest Accuracy (Overall): ", end="")
        print(colored(
            '%2d%% ' % (100. * np.sum(class_correct) / np.sum(class_total)),
            "red"),
              end="")
        print("(", end="")
        print(colored(str(int(np.sum(class_correct))), "red"), end=" ")
        print('/%2d)' % (np.sum(class_total)))
        self.net.train()
        return 100. * np.sum(class_correct) / np.sum(class_total)

    def save_model(self):
        torch.save(self.net, os.path.join(self.model_path, 'pre_train.pth'))

    def load_model(self):
        self.net = torch.load(os.path.join(self.model_path, 'pre_train.pth'))