modules/model.py

from torch import nn
import torch
import torch.nn.functional as F
from modules.util import AntiAliasInterpolation2d, make_coordinate_grid_2d
from torchvision import models
import numpy as np
from torch.autograd import grad
import modules.hopenet as hopenet
from torchvision import transforms

import random
from basicsr.data.degradations import random_add_gaussian_noise_pt, random_add_poisson_noise_pt
from prefetch_generator import BackgroundGenerator
from basicsr.utils.img_process_util import filter2D
from basicsr.utils import DiffJPEG, USMSharp


class Vgg19(torch.nn.Module):
    """
    Vgg19 network for perceptual loss.
    """
    def __init__(self, requires_grad=False):
        super(Vgg19, self).__init__()
        vgg_pretrained_features = models.vgg19(pretrained=True).features
        self.slice1 = torch.nn.Sequential()
        self.slice2 = torch.nn.Sequential()
        self.slice3 = torch.nn.Sequential()
        self.slice4 = torch.nn.Sequential()
        self.slice5 = torch.nn.Sequential()
        for x in range(2):
            self.slice1.add_module(str(x), vgg_pretrained_features[x])
        for x in range(2, 7):
            self.slice2.add_module(str(x), vgg_pretrained_features[x])
        for x in range(7, 12):
            self.slice3.add_module(str(x), vgg_pretrained_features[x])
        for x in range(12, 21):
            self.slice4.add_module(str(x), vgg_pretrained_features[x])
        for x in range(21, 30):
            self.slice5.add_module(str(x), vgg_pretrained_features[x])

        self.mean = torch.nn.Parameter(data=torch.Tensor(np.array([0.485, 0.456, 0.406]).reshape((1, 3, 1, 1))),
                                       requires_grad=False)
        self.std = torch.nn.Parameter(data=torch.Tensor(np.array([0.229, 0.224, 0.225]).reshape((1, 3, 1, 1))),
                                      requires_grad=False)

        if not requires_grad:
            for param in self.parameters():
                param.requires_grad = False

    def forward(self, X):
        X = (X - self.mean) / self.std
        h_relu1 = self.slice1(X)
        h_relu2 = self.slice2(h_relu1)
        h_relu3 = self.slice3(h_relu2)
        h_relu4 = self.slice4(h_relu3)
        h_relu5 = self.slice5(h_relu4)
        out = [h_relu1, h_relu2, h_relu3, h_relu4, h_relu5]
        return out


class ImagePyramide(torch.nn.Module):
    """
    Create image pyramide for computing pyramide perceptual loss.
    """
    def __init__(self, scales, num_channels):
        super(ImagePyramide, self).__init__()
        downs = {}
        for scale in scales:
            downs[str(scale).replace('.', '-')] = AntiAliasInterpolation2d(num_channels, scale)
        self.downs = nn.ModuleDict(downs)

    def forward(self, x):
        out_dict = {}
        for scale, down_module in self.downs.items():
            out_dict['prediction_' + str(scale).replace('-', '.')] = down_module(x)
        return out_dict


class Transform:
    """
    Random tps transformation for equivariance constraints.
    """
    def __init__(self, bs, **kwargs):
        noise = torch.normal(mean=0, std=kwargs['sigma_affine'] * torch.ones([bs, 2, 3]))
        self.theta = noise + torch.eye(2, 3).view(1, 2, 3)
        self.bs = bs

        if ('sigma_tps' in kwargs) and ('points_tps' in kwargs):
            self.tps = True
            self.control_points = make_coordinate_grid_2d((kwargs['points_tps'], kwargs['points_tps']), type=noise.type())
            self.control_points = self.control_points.unsqueeze(0)
            self.control_params = torch.normal(mean=0,
                                               std=kwargs['sigma_tps'] * torch.ones([bs, 1, kwargs['points_tps'] ** 2]))
        else:
            self.tps = False

    def transform_frame(self, frame):
        grid = make_coordinate_grid_2d(frame.shape[2:], type=frame.type()).unsqueeze(0)
        grid = grid.view(1, frame.shape[2] * frame.shape[3], 2)
        grid = self.warp_coordinates(grid).view(self.bs, frame.shape[2], frame.shape[3], 2)
        return F.grid_sample(frame, grid, padding_mode="reflection")

    def warp_coordinates(self, coordinates):
        theta = self.theta.type(coordinates.type())
        theta = theta.unsqueeze(1)
        transformed = torch.matmul(theta[:, :, :, :2], coordinates.unsqueeze(-1)) + theta[:, :, :, 2:]
        transformed = transformed.squeeze(-1)

        if self.tps:
            control_points = self.control_points.type(coordinates.type())
            control_params = self.control_params.type(coordinates.type())
            distances = coordinates.view(coordinates.shape[0], -1, 1, 2) - control_points.view(1, 1, -1, 2)
            distances = torch.abs(distances).sum(-1)

            result = distances ** 2
            result = result * torch.log(distances + 1e-6)
            result = result * control_params
            result = result.sum(dim=2).view(self.bs, coordinates.shape[1], 1)
            transformed = transformed + result

        return transformed

    def jacobian(self, coordinates):
        new_coordinates = self.warp_coordinates(coordinates)
        grad_x = grad(new_coordinates[..., 0].sum(), coordinates, create_graph=True)
        grad_y = grad(new_coordinates[..., 1].sum(), coordinates, create_graph=True)
        jacobian = torch.cat([grad_x[0].unsqueeze(-2), grad_y[0].unsqueeze(-2)], dim=-2)
        return jacobian


def detach_kp(kp):
    return {key: value.detach() for key, value in kp.items()}


def headpose_pred_to_degree(pred):
    device = pred.device
    idx_tensor = [idx for idx in range(66)]
    idx_tensor = torch.FloatTensor(idx_tensor).to(device)
    pred = F.softmax(pred)
    degree = torch.sum(pred*idx_tensor, axis=1) * 3 - 99

    return degree

'''
# beta version
def get_rotation_matrix(yaw, pitch, roll):
    yaw = yaw / 180 * 3.14
    pitch = pitch / 180 * 3.14
    roll = roll / 180 * 3.14

    roll = roll.unsqueeze(1)
    pitch = pitch.unsqueeze(1)
    yaw = yaw.unsqueeze(1)

    roll_mat = torch.cat([torch.ones_like(roll), torch.zeros_like(roll), torch.zeros_like(roll), 
                          torch.zeros_like(roll), torch.cos(roll), -torch.sin(roll),
                          torch.zeros_like(roll), torch.sin(roll), torch.cos(roll)], dim=1)
    roll_mat = roll_mat.view(roll_mat.shape[0], 3, 3)

    pitch_mat = torch.cat([torch.cos(pitch), torch.zeros_like(pitch), torch.sin(pitch), 
                           torch.zeros_like(pitch), torch.ones_like(pitch), torch.zeros_like(pitch),
                           -torch.sin(pitch), torch.zeros_like(pitch), torch.cos(pitch)], dim=1)
    pitch_mat = pitch_mat.view(pitch_mat.shape[0], 3, 3)

    yaw_mat = torch.cat([torch.cos(yaw), -torch.sin(yaw), torch.zeros_like(yaw),  
                         torch.sin(yaw), torch.cos(yaw), torch.zeros_like(yaw),
                         torch.zeros_like(yaw), torch.zeros_like(yaw), torch.ones_like(yaw)], dim=1)
    yaw_mat = yaw_mat.view(yaw_mat.shape[0], 3, 3)

    rot_mat = torch.einsum('bij,bjk,bkm->bim', roll_mat, pitch_mat, yaw_mat)

    return rot_mat
'''

def get_rotation_matrix(yaw, pitch, roll):
    yaw = yaw / 180 * 3.14
    pitch = pitch / 180 * 3.14
    roll = roll / 180 * 3.14

    roll = roll.unsqueeze(1)
    pitch = pitch.unsqueeze(1)
    yaw = yaw.unsqueeze(1)

    pitch_mat = torch.cat([torch.ones_like(pitch), torch.zeros_like(pitch), torch.zeros_like(pitch), 
                          torch.zeros_like(pitch), torch.cos(pitch), -torch.sin(pitch),
                          torch.zeros_like(pitch), torch.sin(pitch), torch.cos(pitch)], dim=1)
    pitch_mat = pitch_mat.view(pitch_mat.shape[0], 3, 3)

    yaw_mat = torch.cat([torch.cos(yaw), torch.zeros_like(yaw), torch.sin(yaw), 
                           torch.zeros_like(yaw), torch.ones_like(yaw), torch.zeros_like(yaw),
                           -torch.sin(yaw), torch.zeros_like(yaw), torch.cos(yaw)], dim=1)
    yaw_mat = yaw_mat.view(yaw_mat.shape[0], 3, 3)

    roll_mat = torch.cat([torch.cos(roll), -torch.sin(roll), torch.zeros_like(roll),  
                         torch.sin(roll), torch.cos(roll), torch.zeros_like(roll),
                         torch.zeros_like(roll), torch.zeros_like(roll), torch.ones_like(roll)], dim=1)
    roll_mat = roll_mat.view(roll_mat.shape[0], 3, 3)

    rot_mat = torch.einsum('bij,bjk,bkm->bim', pitch_mat, yaw_mat, roll_mat)

    return rot_mat

def keypoint_transformation(kp_canonical, he, estimate_jacobian=True):
    kp = kp_canonical['value']    # (bs, k, 3)
    yaw, pitch, roll = he['yaw'], he['pitch'], he['roll']
    t, exp = he['t'], he['exp']
    
    yaw = headpose_pred_to_degree(yaw)
    pitch = headpose_pred_to_degree(pitch)
    roll = headpose_pred_to_degree(roll)

    rot_mat = get_rotation_matrix(yaw, pitch, roll)    # (bs, 3, 3)
    
    # keypoint rotation
    kp_rotated = torch.einsum('bmp,bkp->bkm', rot_mat, kp)

    # keypoint translation
    t = t.unsqueeze_(1).repeat(1, kp.shape[1], 1)
    kp_t = kp_rotated + t

    # add expression deviation 
    exp = exp.view(exp.shape[0], -1, 3)
    kp_transformed = kp_t + exp

    if estimate_jacobian:
        jacobian = kp_canonical['jacobian']   # (bs, k ,3, 3)
        jacobian_transformed = torch.einsum('bmp,bkps->bkms', rot_mat, jacobian)
    else:
        jacobian_transformed = None

    return {'value': kp_transformed, 'jacobian': jacobian_transformed}

class GeneratorFullModel(torch.nn.Module):
    """
    Merge all generator related updates into single model for better multi-gpu usage
    """

    def __init__(self, kp_extractor, he_estimator, generator, discriminator, train_params, estimate_jacobian=True):
        super(GeneratorFullModel, self).__init__()
        self.kp_extractor = kp_extractor
        self.he_estimator = he_estimator
        self.generator = generator
        self.discriminator = discriminator
        self.train_params = train_params
        self.scales = train_params['scales']
        self.disc_scales = self.discriminator.scales
        self.pyramid = ImagePyramide(self.scales, generator.image_channel)
        if torch.cuda.is_available():
            self.pyramid = self.pyramid.cuda()

        self.loss_weights = train_params['loss_weights']

        self.estimate_jacobian = estimate_jacobian

        if sum(self.loss_weights['perceptual']) != 0:
            self.vgg = Vgg19()
            if torch.cuda.is_available():
                self.vgg = self.vgg.cuda()

        self.L1 = nn.L1Loss().cuda()

        if self.loss_weights['headpose'] != 0:
            self.hopenet = hopenet.Hopenet(models.resnet.Bottleneck, [3, 4, 6, 3], 66)
            print('Loading hopenet')
            hopenet_state_dict = torch.load(train_params['hopenet_snapshot'])
            self.hopenet.load_state_dict(hopenet_state_dict)
            if torch.cuda.is_available():
                self.hopenet = self.hopenet.cuda()
                self.hopenet.eval()
        
        self.jpeger = DiffJPEG(differentiable=False).cuda()
        self.usm_sharpener = USMSharp().cuda()

        self.resize_prob = [0.2, 0.7, 0.1]
        self.resize_range= [0.5, 1.2]
        self.noise_range = [1, 10]
        self.poisson_scale_range =[0.05, 1]
        self.jpeg_range = [10, 25]
        
        self.opt_scale = 1

        self.resize_prob2 = [0.3, 0.4, 0.3]
        self.resize_range2= [0.5, 1.2]
        self.noise_range2 = [1, 10]
        self.poisson_scale_range2 = [0.05, 1.0]
        self.jpeg_range2 = [10, 25]


    def forward(self, x, config):
        kp_canonical = self.kp_extractor(x['source'])     # {'value': value, 'jacobian': jacobian}   

        he_source = self.he_estimator(x['source'])        # {'yaw': yaw, 'pitch': pitch, 'roll': roll, 't': t, 'exp': exp}
        he_driving = self.he_estimator(x['driving'])      # {'yaw': yaw, 'pitch': pitch, 'roll': roll, 't': t, 'exp': exp}

        # {'value': value, 'jacobian': jacobian}
        kp_source = keypoint_transformation(kp_canonical, he_source, self.estimate_jacobian)
        kp_driving = keypoint_transformation(kp_canonical, he_driving, self.estimate_jacobian)
        if config['train_params']['low_quality_train']:
            # ----------------------- The first degradation process ----------------------- #
            # blur
            # x_source = self.usm_sharpener(x['source'])
            x_source = x['source']
            x_source = filter2D(x_source, x['kernel'])
            
            # random resize
            updown_type = random.choices(['up', 'down', 'keep'], self.resize_prob)[0]
            if updown_type == 'up':
                scale = np.random.uniform(1, self.resize_range[1])
            elif updown_type == 'down':
                scale = np.random.uniform(self.resize_range[0], 1)
            else:
                scale = 1
            mode = random.choice(['area', 'bilinear', 'bicubic'])
            x_source = F.interpolate(x_source, scale_factor=scale, mode=mode)

            # add noise
            gray_noise_prob = 0.4
            if np.random.uniform() < 0.5:
                x_source = random_add_gaussian_noise_pt(
                    x_source, sigma_range=self.noise_range, clip=True, rounds=False, gray_prob=gray_noise_prob)
            else:
                x_source = random_add_poisson_noise_pt(
                    x_source,
                    scale_range=self.poisson_scale_range,
                    gray_prob=gray_noise_prob,
                    clip=True,
                    rounds=False)
            # JPEG compression
            jpeg_p = x_source.new_zeros(x_source.size(0)).uniform_(*self.jpeg_range)
            x_source = torch.clamp(x_source, 0, 1)  # clamp to [0, 1], otherwise JPEGer will result in unpleasant artifacts
            x_source = self.jpeger(x_source, quality=jpeg_p)
            
            # ----------------------- The second degradation process ----------------------- #
            # blur
            if np.random.uniform() < 0.8:
                x_source = filter2D(x_source, x['kernel2'].cuda())
            # random resize
            updown_type = random.choices(['up', 'down', 'keep'], self.resize_prob2)[0]
            if updown_type == 'up':
                scale = np.random.uniform(1, self.resize_range2[1])
            elif updown_type == 'down':
                scale = np.random.uniform(self.resize_range2[0], 1)
            else:
                scale = 1
            mode = random.choice(['area', 'bilinear', 'bicubic'])
            x_source = F.interpolate(
                x_source, size=(int(config['dataset_params']['frame_shape'][0] / self.opt_scale * scale), int(config['dataset_params']['frame_shape'][1] / self.opt_scale * scale)), mode=mode)
            # add noise
            gray_noise_prob = 0.4
            if np.random.uniform() < 0.5:
                x_source = random_add_gaussian_noise_pt(
                    x_source, sigma_range=self.noise_range2, clip=True, rounds=False, gray_prob=gray_noise_prob)
            else:
                x_source = random_add_poisson_noise_pt(
                    x_source,
                    scale_range=self.poisson_scale_range2,
                    gray_prob=gray_noise_prob,
                    clip=True,
                    rounds=False)

            # JPEG compression + the final sinc filter
            # We also need to resize images to desired sizes. We group [resize back + sinc filter] together
            # as one operation.
            # We consider two orders:
            #   1. [resize back + sinc filter] + JPEG compression
            #   2. JPEG compression + [resize back + sinc filter]
            # Empirically, we find other combinations (sinc + JPEG + Resize) will introduce twisted lines.
            if np.random.uniform() < 0.5:
                # resize back + the final sinc filter
                mode = random.choice(['area', 'bilinear', 'bicubic'])
                x_source = F.interpolate(x_source, size=(config['dataset_params']['frame_shape'][0] // self.opt_scale, config['dataset_params']['frame_shape'][1] // self.opt_scale), mode=mode)
                x_source = filter2D(x_source, x['sinc_kernel'].cuda())
                # JPEG compression
                jpeg_p = x_source.new_zeros(x_source.size(0)).uniform_(*self.jpeg_range2)
                x_source = torch.clamp(x_source, 0, 1)
                x_source = self.jpeger(x_source, quality=jpeg_p)
            else:
                # JPEG compression
                jpeg_p = x_source.new_zeros(x_source.size(0)).uniform_(*self.jpeg_range2)
                x_source = torch.clamp(x_source, 0, 1)
                x_source = self.jpeger(x_source, quality=jpeg_p)
                # resize back + the final sinc filter
                mode = random.choice(['area', 'bilinear', 'bicubic'])
                x_source = F.interpolate(x_source, size=(config['dataset_params']['frame_shape'][0] // self.opt_scale, config['dataset_params']['frame_shape'][1] // self.opt_scale), mode=mode)
                x_source = filter2D(x_source, x['sinc_kernel'].cuda())

            # clamp and round
            lq = torch.clamp((x_source * 255.0).round(), 0, 255) / 255.
            lq_img = lq.contiguous()


            generated = self.generator(lq_img, kp_source=kp_source, kp_driving=kp_driving)
        else:
            generated = self.generator(x['source'], kp_source=kp_source, kp_driving=kp_driving)
        generated.update({'kp_source': kp_source, 'kp_driving': kp_driving})

        loss_values = {}

        pyramide_real = self.pyramid(x['driving'])
        pyramide_generated = self.pyramid(generated['prediction'])

        if sum(self.loss_weights['perceptual']) != 0:
            value_total = 0
            for scale in self.scales:
                x_vgg = self.vgg(pyramide_generated['prediction_' + str(scale)])
                y_vgg = self.vgg(pyramide_real['prediction_' + str(scale)])

                for i, weight in enumerate(self.loss_weights['perceptual']):
                    value = torch.abs(x_vgg[i] - y_vgg[i].detach()).mean()
                    value_total += self.loss_weights['perceptual'][i] * value
            loss_values['perceptual'] = value_total
        
        if self.loss_weights['generator_gan'] != 0:
            discriminator_maps_generated = self.discriminator(pyramide_generated)
            discriminator_maps_real = self.discriminator(pyramide_real)
            value_total = 0
            for scale in self.disc_scales:
                key = 'prediction_map_%s' % scale
                if self.train_params['gan_mode'] == 'hinge':
                    value = -torch.mean(discriminator_maps_generated[key])
                elif self.train_params['gan_mode'] == 'ls':
                    value = ((1 - discriminator_maps_generated[key]) ** 2).mean()
                else:
                    raise ValueError('Unexpected gan_mode {}'.format(self.train_params['gan_mode']))

                value_total += self.loss_weights['generator_gan'] * value
            loss_values['gen_gan'] = value_total

            if sum(self.loss_weights['feature_matching']) != 0:
                value_total = 0
                for scale in self.disc_scales:
                    key = 'feature_maps_%s' % scale
                    for i, (a, b) in enumerate(zip(discriminator_maps_real[key], discriminator_maps_generated[key])):
                        if self.loss_weights['feature_matching'][i] == 0:
                            continue
                        value = torch.abs(a - b).mean()
                        value_total += self.loss_weights['feature_matching'][i] * value
                    loss_values['feature_matching'] = value_total

        if (self.loss_weights['equivariance_value'] + self.loss_weights['equivariance_jacobian']) != 0:
            transform = Transform(x['driving'].shape[0], **self.train_params['transform_params'])
            transformed_frame = transform.transform_frame(x['driving'])

            transformed_he_driving = self.he_estimator(transformed_frame)

            transformed_kp = keypoint_transformation(kp_canonical, transformed_he_driving, self.estimate_jacobian)

            generated['transformed_frame'] = transformed_frame
            generated['transformed_kp'] = transformed_kp

            ## Value loss part
            if self.loss_weights['equivariance_value'] != 0:
                # project 3d -> 2d
                kp_driving_2d = kp_driving['value'][:, :, :2]
                transformed_kp_2d = transformed_kp['value'][:, :, :2]
                value = torch.abs(kp_driving_2d - transform.warp_coordinates(transformed_kp_2d)).mean()
                loss_values['equivariance_value'] = self.loss_weights['equivariance_value'] * value

            ## jacobian loss part
            if self.loss_weights['equivariance_jacobian'] != 0:
                # project 3d -> 2d
                transformed_kp_2d = transformed_kp['value'][:, :, :2]
                transformed_jacobian_2d = transformed_kp['jacobian'][:, :, :2, :2]
                jacobian_transformed = torch.matmul(transform.jacobian(transformed_kp_2d),
                                                    transformed_jacobian_2d)
                
                jacobian_2d = kp_driving['jacobian'][:, :, :2, :2]
                normed_driving = torch.inverse(jacobian_2d)
                normed_transformed = jacobian_transformed
                value = torch.matmul(normed_driving, normed_transformed)

                eye = torch.eye(2).view(1, 1, 2, 2).type(value.type())

                value = torch.abs(eye - value).mean()
                loss_values['equivariance_jacobian'] = self.loss_weights['equivariance_jacobian'] * value

        if self.loss_weights['keypoint'] != 0:
            # print(kp_driving['value'].shape)     # (bs, k, 3)
            value_total = 0
            for i in range(kp_driving['value'].shape[1]):
                for j in range(kp_driving['value'].shape[1]):
                    dist = F.pairwise_distance(kp_driving['value'][:, i, :], kp_driving['value'][:, j, :], p=2, keepdim=True) ** 2
                    dist = 0.1 - dist      # set Dt = 0.1
                    dd = torch.gt(dist, 0) 
                    value = (dist * dd).mean()
                    value_total += value

            kp_mean_depth = kp_driving['value'][:, :, -1].mean(-1)
            value_depth = torch.abs(kp_mean_depth - 0.33).mean()          # set Zt = 0.33

            value_total += value_depth
            loss_values['keypoint'] = self.loss_weights['keypoint'] * value_total

        if self.loss_weights['headpose'] != 0:
            transform_hopenet =  transforms.Compose([transforms.Resize(size=(224, 224)),
                                                     transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])])
            driving_224 = transform_hopenet(x['driving'])

            yaw_gt, pitch_gt, roll_gt = self.hopenet(driving_224)
            yaw_gt = headpose_pred_to_degree(yaw_gt)
            pitch_gt = headpose_pred_to_degree(pitch_gt)
            roll_gt = headpose_pred_to_degree(roll_gt)

            yaw, pitch, roll = he_driving['yaw'], he_driving['pitch'], he_driving['roll']
            yaw = headpose_pred_to_degree(yaw)
            pitch = headpose_pred_to_degree(pitch)
            roll = headpose_pred_to_degree(roll)

            value = torch.abs(yaw - yaw_gt).mean() + torch.abs(pitch - pitch_gt).mean() + torch.abs(roll - roll_gt).mean()
            loss_values['headpose'] = self.loss_weights['headpose'] * value

        if self.loss_weights['expression'] != 0:
            value = torch.norm(he_driving['exp'], p=1, dim=-1).mean()
            loss_values['expression'] = self.loss_weights['expression'] * value

        loss_values['reconstruction'] = self.loss_weights['reconstruction'] * self.L1(generated['prediction'], x['driving'])
        
        return loss_values, generated


class DiscriminatorFullModel(torch.nn.Module):
    """
    Merge all discriminator related updates into single model for better multi-gpu usage
    """

    def __init__(self, kp_extractor, generator, discriminator, train_params):
        super(DiscriminatorFullModel, self).__init__()
        self.kp_extractor = kp_extractor
        self.generator = generator
        self.discriminator = discriminator
        self.train_params = train_params
        self.scales = self.discriminator.scales
        self.pyramid = ImagePyramide(self.scales, generator.image_channel)
        if torch.cuda.is_available():
            self.pyramid = self.pyramid.cuda()

        self.loss_weights = train_params['loss_weights']

        self.zero_tensor = None

    def get_zero_tensor(self, input):
        if self.zero_tensor is None:
            self.zero_tensor = torch.FloatTensor(1).fill_(0).cuda()
            self.zero_tensor.requires_grad_(False)
        return self.zero_tensor.expand_as(input)

    def forward(self, x, generated):
        pyramide_real = self.pyramid(x['driving'])
        pyramide_generated = self.pyramid(generated['prediction'].detach())

        discriminator_maps_generated = self.discriminator(pyramide_generated)
        discriminator_maps_real = self.discriminator(pyramide_real)

        loss_values = {}
        value_total = 0
        for scale in self.scales:
            key = 'prediction_map_%s' % scale
            if self.train_params['gan_mode'] == 'hinge':
                value = -torch.mean(torch.min(discriminator_maps_real[key]-1, self.get_zero_tensor(discriminator_maps_real[key]))) - torch.mean(torch.min(-discriminator_maps_generated[key]-1, self.get_zero_tensor(discriminator_maps_generated[key])))
            elif self.train_params['gan_mode'] == 'ls':
                value = ((1 - discriminator_maps_real[key]) ** 2 + discriminator_maps_generated[key] ** 2).mean()
            else:
                raise ValueError('Unexpected gan_mode {}'.format(self.train_params['gan_mode']))

            value_total += self.loss_weights['discriminator_gan'] * value
        loss_values['disc_gan'] = value_total

        return loss_values