Beispiel #1
0
    def __init__(self,
                 height_lr=24,
                 width_lr=24,
                 channels=3,
                 upscaling_factor=4,
                 gen_lr=1e-4,
                 dis_lr=1e-4,
                 loss_weights=[0.006, 1e-4],
                 training_mode=True,
                 colorspace='RGB'):
        """        
        :param int height_lr: Height of low-resolution images
        :param int width_lr: Width of low-resolution images
        :param int channels: Image channels
        :param int upscaling_factor: Up-scaling factor
        :param int gen_lr: Learning rate of generator
        :param int dis_lr: Learning rate of discriminator
        """

        # Low-resolution image dimensions
        self.height_lr = height_lr
        self.width_lr = width_lr
        self.training_mode = training_mode

        # High-resolution image dimensions
        if upscaling_factor not in [2, 4, 8]:
            raise ValueError(
                'Upscaling factor must be either 2, 4, or 8. You chose {}'.
                format(upscaling_factor))
        self.upscaling_factor = upscaling_factor
        self.height_hr = int(self.height_lr * self.upscaling_factor)
        self.width_hr = int(self.width_lr * self.upscaling_factor)

        # Low-resolution and high-resolution shapes
        self.channels = channels
        self.colorspace = colorspace
        self.shape_lr = (self.height_lr, self.width_lr, self.channels)
        self.shape_hr = (self.height_hr, self.width_hr, self.channels)

        # Learning rates
        self.gen_lr = gen_lr
        self.dis_lr = dis_lr

        # Gan setup settings
        self.loss_weights = loss_weights
        self.VGGLoss = VGGLoss(self.shape_hr)
        self.gen_loss = 'mse'
        self.content_loss = self.VGGLoss.content_loss  # self.VGGLoss.euclidean_content_loss
        self.adversarial_loss = 'binary_crossentropy'

        # Build & compile the generator network
        self.generator = self.build_generator()
        self.compile_generator(self.generator)

        # If training, build rest of GAN network
        if training_mode:
            self.discriminator = self.build_discriminator()
            self.compile_discriminator(self.discriminator)
            self.srgan = self.build_srgan()
            self.compile_srgan(self.srgan)
Beispiel #2
0
    # loss 関数の設定
    #======================================================================
    # Adv loss
    if (args.gan_type == "vanilla"):
        loss_gan_fn = VanillaGANLoss(device, w_sigmoid_D=False)
    elif (args.gan_type == "lsgan"):
        loss_gan_fn = LSGANLoss(device)
    else:
        raise NotImplementedError('gan_type %s not implemented' %
                                  args.gan_type)

    # feature maching loss
    loss_feat_fn = FeatureMatchingLoss(device, n_dis=args.n_dis, n_layers_D=3)

    # vgg perceptual loss
    loss_vgg_fn = VGGLoss(device)

    #======================================================================
    # モデルの学習処理
    #======================================================================
    # 学習中の生成画像の履歴
    fake_images_historys = []

    print("Starting Training Loop...")
    total_steps = 0
    iterations = 0  # 学習処理のイテレーション回数
    n_print = 1
    #-----------------------------
    # エポック数分トレーニング
    #-----------------------------
    for epoch in tqdm(range(args.n_epoches), desc="Epoches"):
    def __init__(self,
                 height_lr=64,
                 width_lr=64,
                 channels=3,
                 upscaling_factor=2,
                 gen_lr=1e-3,
                 dis_lr=1e-4,
                 ra_lr=1e-4,
                 loss_weights=[1e-3, 5e-3, 1e-2],
                 training_mode=True,
                 colorspace='RGB'):
        """
            height_lr: height of the lr image
            width_lr: width of the lr image 
            channels: number of channel of the image
            upscaling_factor= factor upscaling
            lr = learning rate
            training_mode: True or False
            colorspace: 'RGB' or 'YCbCr'
        """

        # Low-resolution image dimensions
        self.height_lr = height_lr
        self.width_lr = width_lr
        self.training_mode = training_mode

        # High-resolution image dimensions
        if upscaling_factor not in [2, 3, 4, 8]:
            raise ValueError(
                'Upscaling factor must be either 2, 4, or 8. You chose {}'.
                format(upscaling_factor))
        self.upscaling_factor = upscaling_factor
        self.height_hr = int(self.height_lr * self.upscaling_factor)
        self.width_hr = int(self.width_lr * self.upscaling_factor)

        # Low-resolution and high-resolution shapes
        self.channels = channels
        self.colorspace = colorspace
        self.shape_lr = (self.height_lr, self.width_lr, self.channels)
        self.shape_hr = (self.height_hr, self.width_hr, self.channels)

        # Learning rates
        self.gen_lr = gen_lr
        self.dis_lr = dis_lr
        self.ra_lr = ra_lr

        # Gan setup settings
        self.loss_weights = loss_weights
        self.VGGLoss = VGGLoss(self.shape_hr)
        self.gen_loss = euclidean  #euclidean, charbonnier, mae,mse,L2Loss,L1Loss,L1L2EucliLoss
        self.content_loss = L1Loss  #self.VGGLoss.content_loss
        self.adversarial_loss = binary_crossentropy
        self.ra_loss = binary_crossentropy

        # Build & compile the generator network
        self.generator = self.build_generator()
        self.compile_generator(self.generator)

        # If training, build rest of GAN network
        if training_mode:
            self.discriminator = self.build_discriminator()
            self.compile_discriminator(self.discriminator)
            self.ra_discriminator = self.build_ra_discriminator()
            self.compile_ra_discriminator(self.ra_discriminator)
            self.rtvsrgan = self.build_rtvsrgan()
            self.compile_rtvsrgan(self.rtvsrgan)
Beispiel #4
0
 def __init__(self, upscale_factor=4, generator_weights=None, **kwargs):
     super(SRGan, self).__init__(**kwargs)
     self.generator = Generator(weights=generator_weights)
     self.discriminator = Discriminator()
     self.vgg_loss = VGGLoss()
Beispiel #5
0
        torch.save(
            {
                'epoch': epoch,
                'gen': net.state_dict(),
                'gen_opt': hsid_optimizer.state_dict()
            }, os.path.join('./checkpoints', "model_latest.pth"))

    tb_writer.close()


from model_refactored import HSIDDenseNetTwoStage
from Discriminator_copy import DiscriminatorABC
adv_criterion = nn.BCEWithLogitsLoss()
recon_criterion = nn.L1Loss()
lambda_recon = 10
vgg_loss = VGGLoss().cuda(device=DEVICE)


def train_model_residual_lowlight_twostage_gan_best():

    start_epoch = 1

    device = DEVICE
    #准备数据
    train_set = HsiCubicTrainDataset('./data/train_lowlight/')
    print('total training example:', len(train_set))

    train_loader = DataLoader(dataset=train_set,
                              batch_size=BATCH_SIZE,
                              shuffle=True)
Beispiel #6
0
    def build_model(self):

        # Generator = Scene Graph Model
        kwargs = {
            'vocab': self.vocab,
            'image_size': self.image_size,
            'embedding_dim': self.embedding_dim,
            'gconv_dim': self.gconv_dim,
            'gconv_hidden_dim': self.gconv_hidden_dim,
            'gconv_num_layers': self.gconv_num_layers,
            'mlp_normalization': self.mlp_normalization,
            'normalization': self.normalization,
            'activation': self.activation,
            'mask_size': self.mask_size,
            'layout_noise_dim': self.layout_noise_dim,
            'netG': self.netG,
            'ngf': self.ngf,
            'n_downsample_global': self.n_downsample_global,
            'n_blocks_global': self.n_blocks_global,
            'n_blocks_local': self.n_blocks_local,
            'n_local_enhancers': self.n_local_enhancers,
        }

        self.generator = Sg2ImModel(**kwargs)
        # Discriminators
        # OBJ Discriminator
        # self.obj_discriminator = AcCropDiscriminator(vocab=self.vocab,
        #                                              arch=self.d_obj_arch,
        #                                              normalization=self.d_normalization,
        #                                              activation=self.d_activation,
        #                                              padding=self.d_padding,
        #                                              object_size=self.crop_size)

        self.obj_discriminator = ref_network.define_obj_D(
            vocab=self.vocab,
            input_nc=3,
            crop_size=self.crop_size,
            ndf=self.ndf,
            n_layers_D=self.n_layers_D_obj,
            norm=self.normalization,
            num_D=self.num_D_obj,
            getIntermFeat=not self.no_objFeat_loss)

        # IMG Discriminator
        self.img_discriminator = ref_network.define_img_D(
            input_nc=3,
            ndf=self.ndf,
            n_layers_D=self.n_layers_D_img,
            norm=self.normalization,
            num_D=self.num_D_img,
            getIntermFeat=not self.no_imgFeat_loss)

        # Loss
        self.gan_g_loss, self.gan_d_loss = get_gan_losses(self.gan_loss_type)
        self.vgg_loss = VGGLoss()

        # Weights
        # feature matching loss weights
        self.img_feat_weights = 4.0 / (self.n_layers_D_img + 1)
        self.img_D_weights = 1.0 / self.num_D_img
        self.obj_feat_weights = 4.0 / (self.n_layers_D_obj + 1)
        self.obj_D_weights = 1.0 / self.num_D_obj

        # Mode of Model
        if self.init_epochs >= self.eval_mode_after:
            print("Setting Generator to Eval Mode")
            self.generator.eval()
        else:
            print("Setting Generator to Train Mode")
            self.generator.train()

        self.obj_discriminator.train()
        self.img_discriminator.train()

        # Optimizers
        self.gen_optimizer = torch.optim.Adam(self.generator.parameters(),
                                              lr=self.learning_rate)
        self.dis_obj_optimizer = torch.optim.Adam(
            self.obj_discriminator.parameters(), lr=self.learning_rate)
        self.dis_img_optimizer = torch.optim.Adam(
            self.img_discriminator.parameters(), lr=self.learning_rate)

        # Others
        self.ac_ind = 1
        self.obj_ind = 0
        if not self.no_objFeat_loss:
            self.ac_ind = 2
            self.obj_ind = 1

        # Print networks
        self.print_network(self.generator, 'Generator')
        self.print_network(self.obj_discriminator, 'Object Discriminator')
        self.print_network(self.img_discriminator, 'Image Discriminator')

        if torch.cuda.is_available():
            self.generator.cuda()
            self.obj_discriminator.cuda()
            self.img_discriminator.cuda()