コード例 #1
0
    dataset = ClassImageLoader(paths=sep_data, transform=transform, inf=True)

    loader = torch.utils.data.DataLoader(dataset,
                                         batch_size=args.batch_size,
                                         num_workers=args.num_workers,
                                         drop_last=True)
    random_loader = torch.utils.data.DataLoader(dataset,
                                                batch_size=args.batch_size,
                                                num_workers=args.num_workers,
                                                drop_last=True)

    # load model
    transfer = Conditional_UNet(num_classes=args.num_classes)
    sd = torch.load(args.cp_path)
    transfer.load_state_dict(sd['inference'])
    transfer.eval()

    classifer = torch.load(args.classifer_path)
    classifer = nn.Sequential(classifer, nn.Softmax(dim=1))
    classifer.eval()

    transfer.cuda()
    classifer.cuda()

    bs = args.batch_size
    labels = torch.as_tensor(np.arange(args.num_classes, dtype=np.int64))
    onehot = torch.eye(args.num_classes)[labels].to('cuda')

    cls_li = []
    vec_li = []
コード例 #2
0
class WeatherTransfer(object):
    def __init__(self, args):

        self.args = args
        self.batch_size = args.batch_size
        self.global_step = 0

        os.makedirs(os.path.join(args.save_dir, args.name), exist_ok=True)
        comment = '_lr-{}_bs-{}_ne-{}_name-{}'.format(args.lr, args.batch_size,
                                                      args.num_epoch,
                                                      args.name)
        self.writer = SummaryWriter(comment=comment)

        # Consts
        self.real = Variable_Float(1., self.batch_size)
        self.fake = Variable_Float(0., self.batch_size)
        self.lmda = 0.

        if args.augmentation:
            train_transform = transforms.Compose([
                transforms.RandomRotation(10),
                transforms.RandomResizedCrop(args.input_size),
                transforms.RandomHorizontalFlip(),
                transforms.ColorJitter(brightness=0.5,
                                       contrast=0.3,
                                       saturation=0.3,
                                       hue=0),
                transforms.ToTensor(),
                transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
            ])
        else:
            train_transform = transforms.Compose([
                transforms.Resize((args.input_size, ) * 2),
                transforms.RandomRotation(10),
                transforms.RandomHorizontalFlip(),
                transforms.ToTensor(),
                transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
            ])

        test_transform = transforms.Compose([
            transforms.Resize((args.input_size, ) * 2),
            transforms.ToTensor(),
            transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
        ])

        self.cols = ['clouds', 'temp', 'humidity', 'pressure', 'windspeed']
        self.num_classes = len(self.cols)

        self.transform = {'train': train_transform, 'test': test_transform}
        self.train_set, self.test_set = self.load_data(
            varbose=True, image_only=args.image_only)

        self.build()

    def load_data(self, varbose=False, image_only=False, train_data_rate=0.7):

        print('Start loading image files...')
        if image_only:
            paths = glob(os.path.join(args.image_root, '*'))
            print('loaded {} data'.format(len(paths)))
            pivot = int(len(paths) * train_data_rate)
            paths_sep = {'train': paths[:pivot], 'test': paths[pivot:]}
            loader = lambda s: ImageLoader(paths_sep[s],
                                           transform=self.transform[s])

        else:
            df = pd.read_pickle(args.pkl_path)

            temp = pd.read_pickle(
                '/mnt/fs2/2019/okada/from_nitta/parm_0.3/sepalated_data_wo-outlier.pkl'
            )
            df_ = temp.loc[:, self.cols].fillna(0)
            df_mean = df_.mean()
            df_std = df_.std()

            df.loc[:, self.cols] = (df.loc[:, self.cols].fillna(0) -
                                    df_mean) / df_std

            print('loaded {} signals data'.format(len(df)))
            df_shuffle = df.sample(frac=1)
            df_sep = {
                'train': df_shuffle[df_shuffle['mode'] == 'train'],
                'test': df_shuffle[df_shuffle['mode'] == 'test']
            }
            del df, df_shuffle, temp
            loader = lambda s: FlickrDataLoader(args.image_root,
                                                df_sep[s],
                                                self.cols,
                                                transform=self.transform[s])

        train_set = loader('train')
        test_set = loader('test')
        print('train:{} test:{} sets have already loaded.'.format(
            len(train_set), len(test_set)))
        return train_set, test_set

    def build(self):
        args = self.args

        # Models
        print('Build Models...')
        self.inference = Conditional_UNet(num_classes=self.num_classes)
        self.discriminator = SNDisc(num_classes=self.num_classes)
        exist_cp = sorted(glob(os.path.join(args.save_dir, args.name, '*')))
        if len(exist_cp) != 0:
            print('Load checkpoint:{}'.format(exist_cp[-1]))
            sd = torch.load(exist_cp[-1])
            self.inference.load_state_dict(sd['inference'])
            self.discriminator.load_state_dict(sd['discriminator'])
            self.epoch = sd['epoch']
            self.global_step = sd['global_step']
            print('Success checkpoint loading!')
        else:
            print('Initialize training status.')
            self.epoch = 0
            self.global_step = 0

        self.estimator = torch.load(args.estimator_path)
        self.estimator.eval()

        # Models to CUDA
        [
            i.cuda()
            for i in [self.inference, self.discriminator, self.estimator]
        ]

        # Optimizer
        self.g_opt = torch.optim.Adam(self.inference.parameters(),
                                      lr=args.lr,
                                      betas=(0.0, 0.999),
                                      weight_decay=args.lr / 20)
        self.d_opt = torch.optim.Adam(self.discriminator.parameters(),
                                      lr=args.lr,
                                      betas=(0.0, 0.999),
                                      weight_decay=args.lr / 20)

        # これらのloaderにsamplerは必要ないのか?
        self.train_loader = torch.utils.data.DataLoader(
            self.train_set,
            batch_size=args.batch_size,
            shuffle=True,
            drop_last=True,
            num_workers=args.num_workers)

        if args.sampler:
            self.random_loader = torch.utils.data.DataLoader(
                self.train_set,
                batch_size=args.batch_size,
                sampler=ImbalancedDatasetSampler(self.train_set),
                drop_last=True,
                num_workers=args.num_workers)
        else:
            self.random_loader = torch.utils.data.DataLoader(
                self.train_set,
                batch_size=args.batch_size,
                shuffle=True,
                drop_last=True,
                num_workers=args.num_workers)

        if not args.image_only:
            self.test_loader = torch.utils.data.DataLoader(
                self.test_set,
                batch_size=args.batch_size,
                shuffle=True,
                drop_last=True,
                num_workers=args.num_workers)
            test_data_iter = iter(self.test_loader)
            self.test_random_sample = [
                tuple(d.to('cuda') for d in test_data_iter.next())
                for i in range(2)
            ]
            del test_data_iter, self.test_loader

        self.scalar_dict = {}
        self.image_dict = {}
        self.shift_lmda = lambda a, b: (1. - self.lmda) * a + self.lmda * b
        print('Build has been completed.')

    def update_inference(self, images, r_labels):
        # --- UPDATE(Inference) --- #
        self.g_opt.zero_grad()

        # for real
        pred_labels = self.estimator(images).detach()
        # real_res = self.discriminator(images, pred_labels)

        # real_d_out = real_res[0]
        # real_feat = real_res[3]

        fake_out = self.inference(images, r_labels)
        fake_c_out = self.estimator(fake_out)
        fake_res = self.discriminator(fake_out, r_labels)
        fake_d_out = fake_res[0]
        # fake_feat = fake_res[3]

        # Calc Generator Loss
        g_loss_adv = gen_hinge(fake_d_out)  # Adversarial loss
        g_loss_l1 = l1_loss(fake_out, images)
        g_loss_w = pred_loss(fake_c_out, r_labels)  # Weather prediction

        # abs_loss = torch.mean(torch.abs(fake_out - images), [1, 2, 3])
        diff = torch.mean(torch.abs(fake_out - images), [1, 2, 3])
        lmda = torch.mean(torch.abs(pred_labels - r_labels), 1)
        loss_con = torch.mean(diff / (lmda + 1e-7))  # Reconstraction loss

        lmda_con, lmda_w = (1, 1)

        g_loss = g_loss_adv + lmda_con * loss_con + lmda_w * g_loss_w

        g_loss.backward()
        self.g_opt.step()

        self.scalar_dict.update({
            'losses/g_loss/train': g_loss.item(),
            'losses/g_loss_adv/train': g_loss_adv.item(),
            'losses/g_loss_l1/train': g_loss_l1.item(),
            'losses/g_loss_w/train': g_loss_w.item(),
            'losses/loss_con/train': loss_con.item(),
            'variables/lmda': self.lmda
        })

        self.image_dict.update({
            'io/train':
            torch.cat([images, fake_out], dim=3),
        })

    def update_discriminator(self, images, labels):

        # --- UPDATE(Discriminator) ---#
        self.d_opt.zero_grad()

        # for real
        real_c_out = self.estimator(images)
        pred_labels = real_c_out.detach()

        real_d_out_pred = self.discriminator(images, pred_labels)[0]

        # for fake
        fake_out = self.inference(images, labels)
        fake_d_out = self.discriminator(fake_out.detach(), labels)[0]

        d_loss = dis_hinge(fake_d_out, real_d_out_pred)

        d_loss.backward()
        self.d_opt.step()

        self.scalar_dict.update({'losses/d_loss/train': d_loss.item()})

    def evaluation(self):
        g_loss_l1_ = []
        g_loss_adv_ = []
        g_loss_w_ = []
        fake_out_li = []
        d_loss_ = []
        # loss_con_ = []

        images, labels = self.test_random_sample[0]
        # if not args.supervised:
        #     labels_ = self.estimator(images).detach()
        blank = torch.zeros_like(images[0]).unsqueeze(0)
        ref_images, ref_labels = self.test_random_sample[1]

        for i in range(self.batch_size):
            with torch.no_grad():
                if ref_labels is None:
                    ref_labels = self.estimator(ref_images)
                ref_labels_expand = torch.cat([ref_labels[i]] *
                                              self.batch_size).view(
                                                  -1, self.num_classes)
                fake_out_ = self.inference(images, ref_labels_expand)
                fake_c_out_ = self.estimator(fake_out_)
                # fake_d_out_ = self.discriminator(fake_out_, labels)[0]  # Dへの入力はfake_out_ と re_labels_expandではないのか?
                real_d_out_ = self.discriminator(images, labels)[0]
                fake_d_out_ = self.discriminator(fake_out_,
                                                 ref_labels_expand)[0]

            # diff = torch.mean(torch.abs(fake_out_ - images), [1, 2, 3])
            # lmda = torch.mean(torch.abs(pred_labels_ - ref_labels_expand), 1)
            # loss_con_ = torch.mean(diff / (lmda + 1e-7))

            fake_out_li.append(fake_out_)
            g_loss_adv_.append(gen_hinge(fake_d_out_).item())
            g_loss_l1_.append(l1_loss(fake_out_, images).item())
            g_loss_w_.append(pred_loss(fake_c_out_, ref_labels_expand).item())
            d_loss_.append(dis_hinge(fake_d_out_, real_d_out_).item())
            # loss_con_.append(torch.mean(diff / (lmda + 1e-7).item())

        # --- WRITING SUMMARY ---#
        self.scalar_dict.update({
            'losses/g_loss_adv/test': np.mean(g_loss_adv_),
            'losses/g_loss_l1/test': np.mean(g_loss_l1_),
            'losses/g_loss_w/test': np.mean(g_loss_w_),
            'losses/d_loss/test': np.mean(d_loss_)
        })
        ref_img = torch.cat([blank] + list(torch.split(ref_images, 1)), dim=3)
        in_out_img = torch.cat([images] + fake_out_li, dim=3)
        res_img = torch.cat([ref_img, in_out_img], dim=0)

        self.image_dict.update({
            'images/test': res_img,
        })

    def update_summary(self):
        # Summarize
        for k, v in self.scalar_dict.items():
            spk = k.rsplit('/', 1)
            self.writer.add_scalars(spk[0], {spk[1]: v}, self.global_step)
        for k, v in self.image_dict.items():
            grid = make_grid(v, nrow=1, normalize=True, scale_each=True)
            self.writer.add_image(k, grid, self.global_step)

    def train(self):
        args = self.args

        # train setting
        eval_per_step = 1000
        display_per_step = 1000
        save_per_epoch = 5

        self.all_step = args.num_epoch * len(self.train_set) // self.batch_size

        tqdm_iter = trange(args.num_epoch, desc='Training', leave=True)
        for epoch in tqdm_iter:
            if epoch > 0:
                self.epoch += 1

            for i, (data, rand_data) in enumerate(
                    zip(self.train_loader, self.random_loader)):
                self.global_step += 1

                if self.global_step % eval_per_step == 0:
                    out_path = os.path.join(args.save_dir, args.name,
                                            (args.name +
                                             '_e{:04d}_s{}.pt').format(
                                                 self.epoch, self.global_step))
                    state_dict = {
                        'inference': self.inference.state_dict(),
                        'discriminator': self.discriminator.state_dict(),
                        'epoch': self.epoch,
                        'global_step': self.global_step
                    }
                    torch.save(state_dict, out_path)

                tqdm_iter.set_description('Training [ {} step ]'.format(
                    self.global_step))
                if args.lmda:
                    self.lmda = args.lmda
                else:
                    self.lmda = self.global_step / self.all_step

                images, d_ = (d.to('cuda') for d in data)
                rand_images, r_ = (d.to('cuda') for d in rand_data)

                rand_labels = self.estimator(rand_images).detach()

                if images.size(0) != self.batch_size:
                    continue

                self.update_discriminator(images, rand_labels)
                if self.global_step % args.GD_train_ratio == 0:
                    self.update_inference(images, rand_labels)

                # --- EVALUATION ---#
                if (self.global_step % eval_per_step
                        == 0) and not args.image_only:
                    self.evaluation()

                # --- UPDATE SUMMARY ---#
                if self.global_step % display_per_step == 0:
                    self.update_summary()
        print('Done: training')