Upload 14 files

.gitattributes

@@ -38,3 +38,4 @@ sup-mat/fashion-teaser.gif filter=lfs diff=lfs merge=lfs -text
 sup-mat/mgif-teaser.gif filter=lfs diff=lfs merge=lfs -text
 sup-mat/relative-demo.gif filter=lfs diff=lfs merge=lfs -text
 sup-mat/vox-teaser.gif filter=lfs diff=lfs merge=lfs -text
+got-05.jpg filter=lfs diff=lfs merge=lfs -text

animate.py

@@ -0,0 +1,101 @@
+import os
+from tqdm import tqdm
+
+import torch
+from torch.utils.data import DataLoader
+
+from frames_dataset import PairedDataset
+from logger import Logger, Visualizer
+import imageio
+from scipy.spatial import ConvexHull
+import numpy as np
+
+from sync_batchnorm import DataParallelWithCallback
+
+
+def normalize_kp(kp_source, kp_driving, kp_driving_initial, adapt_movement_scale=False,
+                 use_relative_movement=False, use_relative_jacobian=False):
+    if adapt_movement_scale:
+        source_area = ConvexHull(kp_source['value'][0].data.cpu().numpy()).volume
+        driving_area = ConvexHull(kp_driving_initial['value'][0].data.cpu().numpy()).volume
+        adapt_movement_scale = np.sqrt(source_area) / np.sqrt(driving_area)
+    else:
+        adapt_movement_scale = 1
+
+    kp_new = {k: v for k, v in kp_driving.items()}
+
+    if use_relative_movement:
+        kp_value_diff = (kp_driving['value'] - kp_driving_initial['value'])
+        kp_value_diff *= adapt_movement_scale
+        kp_new['value'] = kp_value_diff + kp_source['value']
+
+        if use_relative_jacobian:
+            jacobian_diff = torch.matmul(kp_driving['jacobian'], torch.inverse(kp_driving_initial['jacobian']))
+            kp_new['jacobian'] = torch.matmul(jacobian_diff, kp_source['jacobian'])
+
+    return kp_new
+
+
+def animate(config, generator, kp_detector, checkpoint, log_dir, dataset):
+    log_dir = os.path.join(log_dir, 'animation')
+    png_dir = os.path.join(log_dir, 'png')
+    animate_params = config['animate_params']
+
+    dataset = PairedDataset(initial_dataset=dataset, number_of_pairs=animate_params['num_pairs'])
+    dataloader = DataLoader(dataset, batch_size=1, shuffle=False, num_workers=1)
+
+    if checkpoint is not None:
+        Logger.load_cpk(checkpoint, generator=generator, kp_detector=kp_detector)
+    else:
+        raise AttributeError("Checkpoint should be specified for mode='animate'.")
+
+    if not os.path.exists(log_dir):
+        os.makedirs(log_dir)
+
+    if not os.path.exists(png_dir):
+        os.makedirs(png_dir)
+
+    if torch.cuda.is_available():
+        generator = DataParallelWithCallback(generator)
+        kp_detector = DataParallelWithCallback(kp_detector)
+
+    generator.eval()
+    kp_detector.eval()
+
+    for it, x in tqdm(enumerate(dataloader)):
+        with torch.no_grad():
+            predictions = []
+            visualizations = []
+
+            driving_video = x['driving_video']
+            source_frame = x['source_video'][:, :, 0, :, :]
+
+            kp_source = kp_detector(source_frame)
+            kp_driving_initial = kp_detector(driving_video[:, :, 0])
+
+            for frame_idx in range(driving_video.shape[2]):
+                driving_frame = driving_video[:, :, frame_idx]
+                kp_driving = kp_detector(driving_frame)
+                kp_norm = normalize_kp(kp_source=kp_source, kp_driving=kp_driving,
+                                       kp_driving_initial=kp_driving_initial, **animate_params['normalization_params'])
+                out = generator(source_frame, kp_source=kp_source, kp_driving=kp_norm)
+
+                out['kp_driving'] = kp_driving
+                out['kp_source'] = kp_source
+                out['kp_norm'] = kp_norm
+
+                del out['sparse_deformed']
+
+                predictions.append(np.transpose(out['prediction'].data.cpu().numpy(), [0, 2, 3, 1])[0])
+
+                visualization = Visualizer(**config['visualizer_params']).visualize(source=source_frame,
+                                                                                    driving=driving_frame, out=out)
+                visualization = visualization
+                visualizations.append(visualization)
+
+            predictions = np.concatenate(predictions, axis=1)
+            result_name = "-".join([x['driving_name'][0], x['source_name'][0]])
+            imageio.imsave(os.path.join(png_dir, result_name + '.png'), (255 * predictions).astype(np.uint8))
+
+            image_name = result_name + animate_params['format']
+            imageio.mimsave(os.path.join(log_dir, image_name), visualizations)

augmentation.py

@@ -0,0 +1,345 @@
+"""
+Code from https://github.com/hassony2/torch_videovision
+"""
+
+import numbers
+
+import random
+import numpy as np
+import PIL
+
+from skimage.transform import resize, rotate
+from skimage.util import pad
+import torchvision
+
+import warnings
+
+from skimage import img_as_ubyte, img_as_float
+
+
+def crop_clip(clip, min_h, min_w, h, w):
+    if isinstance(clip[0], np.ndarray):
+        cropped = [img[min_h:min_h + h, min_w:min_w + w, :] for img in clip]
+
+    elif isinstance(clip[0], PIL.Image.Image):
+        cropped = [
+            img.crop((min_w, min_h, min_w + w, min_h + h)) for img in clip
+            ]
+    else:
+        raise TypeError('Expected numpy.ndarray or PIL.Image' +
+                        'but got list of {0}'.format(type(clip[0])))
+    return cropped
+
+
+def pad_clip(clip, h, w):
+    im_h, im_w = clip[0].shape[:2]
+    pad_h = (0, 0) if h < im_h else ((h - im_h) // 2, (h - im_h + 1) // 2)
+    pad_w = (0, 0) if w < im_w else ((w - im_w) // 2, (w - im_w + 1) // 2)
+
+    return pad(clip, ((0, 0), pad_h, pad_w, (0, 0)), mode='edge')
+
+
+def resize_clip(clip, size, interpolation='bilinear'):
+    if isinstance(clip[0], np.ndarray):
+        if isinstance(size, numbers.Number):
+            im_h, im_w, im_c = clip[0].shape
+            # Min spatial dim already matches minimal size
+            if (im_w <= im_h and im_w == size) or (im_h <= im_w
+                                                   and im_h == size):
+                return clip
+            new_h, new_w = get_resize_sizes(im_h, im_w, size)
+            size = (new_w, new_h)
+        else:
+            size = size[1], size[0]
+
+        scaled = [
+            resize(img, size, order=1 if interpolation == 'bilinear' else 0, preserve_range=True,
+                   mode='constant', anti_aliasing=True) for img in clip
+            ]
+    elif isinstance(clip[0], PIL.Image.Image):
+        if isinstance(size, numbers.Number):
+            im_w, im_h = clip[0].size
+            # Min spatial dim already matches minimal size
+            if (im_w <= im_h and im_w == size) or (im_h <= im_w
+                                                   and im_h == size):
+                return clip
+            new_h, new_w = get_resize_sizes(im_h, im_w, size)
+            size = (new_w, new_h)
+        else:
+            size = size[1], size[0]
+        if interpolation == 'bilinear':
+            pil_inter = PIL.Image.NEAREST
+        else:
+            pil_inter = PIL.Image.BILINEAR
+        scaled = [img.resize(size, pil_inter) for img in clip]
+    else:
+        raise TypeError('Expected numpy.ndarray or PIL.Image' +
+                        'but got list of {0}'.format(type(clip[0])))
+    return scaled
+
+
+def get_resize_sizes(im_h, im_w, size):
+    if im_w < im_h:
+        ow = size
+        oh = int(size * im_h / im_w)
+    else:
+        oh = size
+        ow = int(size * im_w / im_h)
+    return oh, ow
+
+
+class RandomFlip(object):
+    def __init__(self, time_flip=False, horizontal_flip=False):
+        self.time_flip = time_flip
+        self.horizontal_flip = horizontal_flip
+
+    def __call__(self, clip):
+        if random.random() < 0.5 and self.time_flip:
+            return clip[::-1]
+        if random.random() < 0.5 and self.horizontal_flip:
+            return [np.fliplr(img) for img in clip]
+
+        return clip
+
+
+class RandomResize(object):
+    """Resizes a list of (H x W x C) numpy.ndarray to the final size
+    The larger the original image is, the more times it takes to
+    interpolate
+    Args:
+    interpolation (str): Can be one of 'nearest', 'bilinear'
+    defaults to nearest
+    size (tuple): (widht, height)
+    """
+
+    def __init__(self, ratio=(3. / 4., 4. / 3.), interpolation='nearest'):
+        self.ratio = ratio
+        self.interpolation = interpolation
+
+    def __call__(self, clip):
+        scaling_factor = random.uniform(self.ratio[0], self.ratio[1])
+
+        if isinstance(clip[0], np.ndarray):
+            im_h, im_w, im_c = clip[0].shape
+        elif isinstance(clip[0], PIL.Image.Image):
+            im_w, im_h = clip[0].size
+
+        new_w = int(im_w * scaling_factor)
+        new_h = int(im_h * scaling_factor)
+        new_size = (new_w, new_h)
+        resized = resize_clip(
+            clip, new_size, interpolation=self.interpolation)
+
+        return resized
+
+
+class RandomCrop(object):
+    """Extract random crop at the same location for a list of videos
+    Args:
+    size (sequence or int): Desired output size for the
+    crop in format (h, w)
+    """
+
+    def __init__(self, size):
+        if isinstance(size, numbers.Number):
+            size = (size, size)
+
+        self.size = size
+
+    def __call__(self, clip):
+        """
+        Args:
+        img (PIL.Image or numpy.ndarray): List of videos to be cropped
+        in format (h, w, c) in numpy.ndarray
+        Returns:
+        PIL.Image or numpy.ndarray: Cropped list of videos
+        """
+        h, w = self.size
+        if isinstance(clip[0], np.ndarray):
+            im_h, im_w, im_c = clip[0].shape
+        elif isinstance(clip[0], PIL.Image.Image):
+            im_w, im_h = clip[0].size
+        else:
+            raise TypeError('Expected numpy.ndarray or PIL.Image' +
+                            'but got list of {0}'.format(type(clip[0])))
+
+        clip = pad_clip(clip, h, w)
+        im_h, im_w = clip.shape[1:3]
+        x1 = 0 if h == im_h else random.randint(0, im_w - w)
+        y1 = 0 if w == im_w else random.randint(0, im_h - h)
+        cropped = crop_clip(clip, y1, x1, h, w)
+
+        return cropped
+
+
+class RandomRotation(object):
+    """Rotate entire clip randomly by a random angle within
+    given bounds
+    Args:
+    degrees (sequence or int): Range of degrees to select from
+    If degrees is a number instead of sequence like (min, max),
+    the range of degrees, will be (-degrees, +degrees).
+    """
+
+    def __init__(self, degrees):
+        if isinstance(degrees, numbers.Number):
+            if degrees < 0:
+                raise ValueError('If degrees is a single number,'
+                                 'must be positive')
+            degrees = (-degrees, degrees)
+        else:
+            if len(degrees) != 2:
+                raise ValueError('If degrees is a sequence,'
+                                 'it must be of len 2.')
+
+        self.degrees = degrees
+
+    def __call__(self, clip):
+        """
+        Args:
+        img (PIL.Image or numpy.ndarray): List of videos to be cropped
+        in format (h, w, c) in numpy.ndarray
+        Returns:
+        PIL.Image or numpy.ndarray: Cropped list of videos
+        """
+        angle = random.uniform(self.degrees[0], self.degrees[1])
+        if isinstance(clip[0], np.ndarray):
+            rotated = [rotate(image=img, angle=angle, preserve_range=True) for img in clip]
+        elif isinstance(clip[0], PIL.Image.Image):
+            rotated = [img.rotate(angle) for img in clip]
+        else:
+            raise TypeError('Expected numpy.ndarray or PIL.Image' +
+                            'but got list of {0}'.format(type(clip[0])))
+
+        return rotated
+
+
+class ColorJitter(object):
+    """Randomly change the brightness, contrast and saturation and hue of the clip
+    Args:
+    brightness (float): How much to jitter brightness. brightness_factor
+    is chosen uniformly from [max(0, 1 - brightness), 1 + brightness].
+    contrast (float): How much to jitter contrast. contrast_factor
+    is chosen uniformly from [max(0, 1 - contrast), 1 + contrast].
+    saturation (float): How much to jitter saturation. saturation_factor
+    is chosen uniformly from [max(0, 1 - saturation), 1 + saturation].
+    hue(float): How much to jitter hue. hue_factor is chosen uniformly from
+    [-hue, hue]. Should be >=0 and <= 0.5.
+    """
+
+    def __init__(self, brightness=0, contrast=0, saturation=0, hue=0):
+        self.brightness = brightness
+        self.contrast = contrast
+        self.saturation = saturation
+        self.hue = hue
+
+    def get_params(self, brightness, contrast, saturation, hue):
+        if brightness > 0:
+            brightness_factor = random.uniform(
+                max(0, 1 - brightness), 1 + brightness)
+        else:
+            brightness_factor = None
+
+        if contrast > 0:
+            contrast_factor = random.uniform(
+                max(0, 1 - contrast), 1 + contrast)
+        else:
+            contrast_factor = None
+
+        if saturation > 0:
+            saturation_factor = random.uniform(
+                max(0, 1 - saturation), 1 + saturation)
+        else:
+            saturation_factor = None
+
+        if hue > 0:
+            hue_factor = random.uniform(-hue, hue)
+        else:
+            hue_factor = None
+        return brightness_factor, contrast_factor, saturation_factor, hue_factor
+
+    def __call__(self, clip):
+        """
+        Args:
+        clip (list): list of PIL.Image
+        Returns:
+        list PIL.Image : list of transformed PIL.Image
+        """
+        if isinstance(clip[0], np.ndarray):
+            brightness, contrast, saturation, hue = self.get_params(
+                self.brightness, self.contrast, self.saturation, self.hue)
+
+            # Create img transform function sequence
+            img_transforms = []
+            if brightness is not None:
+                img_transforms.append(lambda img: torchvision.transforms.functional.adjust_brightness(img, brightness))
+            if saturation is not None:
+                img_transforms.append(lambda img: torchvision.transforms.functional.adjust_saturation(img, saturation))
+            if hue is not None:
+                img_transforms.append(lambda img: torchvision.transforms.functional.adjust_hue(img, hue))
+            if contrast is not None:
+                img_transforms.append(lambda img: torchvision.transforms.functional.adjust_contrast(img, contrast))
+            random.shuffle(img_transforms)
+            img_transforms = [img_as_ubyte, torchvision.transforms.ToPILImage()] + img_transforms + [np.array,
+                                                                                                     img_as_float]
+
+            with warnings.catch_warnings():
+                warnings.simplefilter("ignore")
+                jittered_clip = []
+                for img in clip:
+                    jittered_img = img
+                    for func in img_transforms:
+                        jittered_img = func(jittered_img)
+                    jittered_clip.append(jittered_img.astype('float32'))
+        elif isinstance(clip[0], PIL.Image.Image):
+            brightness, contrast, saturation, hue = self.get_params(
+                self.brightness, self.contrast, self.saturation, self.hue)
+
+            # Create img transform function sequence
+            img_transforms = []
+            if brightness is not None:
+                img_transforms.append(lambda img: torchvision.transforms.functional.adjust_brightness(img, brightness))
+            if saturation is not None:
+                img_transforms.append(lambda img: torchvision.transforms.functional.adjust_saturation(img, saturation))
+            if hue is not None:
+                img_transforms.append(lambda img: torchvision.transforms.functional.adjust_hue(img, hue))
+            if contrast is not None:
+                img_transforms.append(lambda img: torchvision.transforms.functional.adjust_contrast(img, contrast))
+            random.shuffle(img_transforms)
+
+            # Apply to all videos
+            jittered_clip = []
+            for img in clip:
+                for func in img_transforms:
+                    jittered_img = func(img)
+                jittered_clip.append(jittered_img)
+
+        else:
+            raise TypeError('Expected numpy.ndarray or PIL.Image' +
+                            'but got list of {0}'.format(type(clip[0])))
+        return jittered_clip
+
+
+class AllAugmentationTransform:
+    def __init__(self, resize_param=None, rotation_param=None, flip_param=None, crop_param=None, jitter_param=None):
+        self.transforms = []
+
+        if flip_param is not None:
+            self.transforms.append(RandomFlip(**flip_param))
+
+        if rotation_param is not None:
+            self.transforms.append(RandomRotation(**rotation_param))
+
+        if resize_param is not None:
+            self.transforms.append(RandomResize(**resize_param))
+
+        if crop_param is not None:
+            self.transforms.append(RandomCrop(**crop_param))
+
+        if jitter_param is not None:
+            self.transforms.append(ColorJitter(**jitter_param))
+
+    def __call__(self, clip):
+        for t in self.transforms:
+            clip = t(clip)
+        return clip

crop-video.py

@@ -0,0 +1,158 @@
+import face_alignment
+import skimage.io
+import numpy
+from argparse import ArgumentParser
+from skimage import img_as_ubyte
+from skimage.transform import resize
+from tqdm import tqdm
+import os
+import imageio
+import numpy as np
+import warnings
+warnings.filterwarnings("ignore")
+
+def extract_bbox(frame, fa):
+    if max(frame.shape[0], frame.shape[1]) > 640:
+        scale_factor =  max(frame.shape[0], frame.shape[1]) / 640.0
+        frame = resize(frame, (int(frame.shape[0] / scale_factor), int(frame.shape[1] / scale_factor)))
+        frame = img_as_ubyte(frame)
+    else:
+        scale_factor = 1
+    frame = frame[..., :3]
+    bboxes = fa.face_detector.detect_from_image(frame[..., ::-1])
+    if len(bboxes) == 0:
+        return []
+    return np.array(bboxes)[:, :-1] * scale_factor
+
+
+
+def bb_intersection_over_union(boxA, boxB):
+    xA = max(boxA[0], boxB[0])
+    yA = max(boxA[1], boxB[1])
+    xB = min(boxA[2], boxB[2])
+    yB = min(boxA[3], boxB[3])
+    interArea = max(0, xB - xA + 1) * max(0, yB - yA + 1)
+    boxAArea = (boxA[2] - boxA[0] + 1) * (boxA[3] - boxA[1] + 1)
+    boxBArea = (boxB[2] - boxB[0] + 1) * (boxB[3] - boxB[1] + 1)
+    iou = interArea / float(boxAArea + boxBArea - interArea)
+    return iou
+
+
+def join(tube_bbox, bbox):
+    xA = min(tube_bbox[0], bbox[0])
+    yA = min(tube_bbox[1], bbox[1])
+    xB = max(tube_bbox[2], bbox[2])
+    yB = max(tube_bbox[3], bbox[3])
+    return (xA, yA, xB, yB)
+
+
+def compute_bbox(start, end, fps, tube_bbox, frame_shape, inp, image_shape, increase_area=0.1):
+    left, top, right, bot = tube_bbox
+    width = right - left
+    height = bot - top
+
+    #Computing aspect preserving bbox
+    width_increase = max(increase_area, ((1 + 2 * increase_area) * height - width) / (2 * width))
+    height_increase = max(increase_area, ((1 + 2 * increase_area) * width - height) / (2 * height))
+
+    left = int(left - width_increase * width)
+    top = int(top - height_increase * height)
+    right = int(right + width_increase * width)
+    bot = int(bot + height_increase * height)
+
+    top, bot, left, right = max(0, top), min(bot, frame_shape[0]), max(0, left), min(right, frame_shape[1])
+    h, w = bot - top, right - left
+
+    start = start / fps
+    end = end / fps
+    time = end - start
+
+    scale = f'{image_shape[0]}:{image_shape[1]}'
+
+    return f'ffmpeg -i {inp} -ss {start} -t {time} -filter:v "crop={w}:{h}:{left}:{top}, scale={scale}" crop.mp4'
+
+
+def compute_bbox_trajectories(trajectories, fps, frame_shape, args):
+    commands = []
+    for i, (bbox, tube_bbox, start, end) in enumerate(trajectories):
+        if (end - start) > args.min_frames:
+            command = compute_bbox(start, end, fps, tube_bbox, frame_shape, inp=args.inp, image_shape=args.image_shape, increase_area=args.increase)
+            commands.append(command)
+    return commands
+
+
+def process_video(args):
+    device = 'cpu' if args.cpu else 'cuda'
+    fa = face_alignment.FaceAlignment(face_alignment.LandmarksType._2D, flip_input=False, device=device)
+    video = imageio.get_reader(args.inp)
+
+    trajectories = []
+    previous_frame = None
+    fps = video.get_meta_data()['fps']
+    commands = []
+    try:
+        for i, frame in tqdm(enumerate(video)):
+            frame_shape = frame.shape
+            bboxes =  extract_bbox(frame, fa)
+            ## For each trajectory check the criterion
+            not_valid_trajectories = []
+            valid_trajectories = []
+
+            for trajectory in trajectories:
+                tube_bbox = trajectory[0]
+                intersection = 0
+                for bbox in bboxes:
+                    intersection = max(intersection, bb_intersection_over_union(tube_bbox, bbox))
+                if intersection > args.iou_with_initial:
+                    valid_trajectories.append(trajectory)
+                else:
+                    not_valid_trajectories.append(trajectory)
+
+            commands += compute_bbox_trajectories(not_valid_trajectories, fps, frame_shape, args)
+            trajectories = valid_trajectories
+
+            ## Assign bbox to trajectories, create new trajectories
+            for bbox in bboxes:
+                intersection = 0
+                current_trajectory = None
+                for trajectory in trajectories:
+                    tube_bbox = trajectory[0]
+                    current_intersection = bb_intersection_over_union(tube_bbox, bbox)
+                    if intersection < current_intersection and current_intersection > args.iou_with_initial:
+                        intersection = bb_intersection_over_union(tube_bbox, bbox)
+                        current_trajectory = trajectory
+
+                ## Create new trajectory
+                if current_trajectory is None:
+                    trajectories.append([bbox, bbox, i, i])
+                else:
+                    current_trajectory[3] = i
+                    current_trajectory[1] = join(current_trajectory[1], bbox)
+
+
+    except IndexError as e:
+        raise (e)
+
+    commands += compute_bbox_trajectories(trajectories, fps, frame_shape, args)
+    return commands
+
+
+if __name__ == "__main__":
+    parser = ArgumentParser()
+
+    parser.add_argument("--image_shape", default=(256, 256), type=lambda x: tuple(map(int, x.split(','))),
+                        help="Image shape")
+    parser.add_argument("--increase", default=0.1, type=float, help='Increase bbox by this amount')
+    parser.add_argument("--iou_with_initial", type=float, default=0.25, help="The minimal allowed iou with inital bbox")
+    parser.add_argument("--inp", required=True, help='Input image or video')
+    parser.add_argument("--min_frames", type=int, default=150,  help='Minimum number of frames')
+    parser.add_argument("--cpu", dest="cpu", action="store_true", help="cpu mode.")
+
+
+    args = parser.parse_args()
+
+    commands = process_video(args)
+    for command in commands:
+        print (command)
+
+        

demo.py

@@ -0,0 +1,157 @@
+import matplotlib
+matplotlib.use('Agg')
+import os, sys
+import yaml
+from argparse import ArgumentParser
+from tqdm import tqdm
+
+import imageio
+import numpy as np
+from skimage.transform import resize
+from skimage import img_as_ubyte
+import torch
+from sync_batchnorm import DataParallelWithCallback
+from modules.generator import OcclusionAwareGenerator
+from modules.keypoint_detector import KPDetector
+from animate import normalize_kp
+from scipy.spatial import ConvexHull
+
+
+if sys.version_info[0] < 3:
+    raise Exception("You must use Python 3 or higher. Recommended version is Python 3.7")
+
+
+def load_checkpoints(config_path, checkpoint_path, cpu=False):
+
+    with open(config_path) as f:
+        config = yaml.load(f)
+
+    generator = OcclusionAwareGenerator(**config['model_params']['generator_params'],
+                                        **config['model_params']['common_params'])
+    if not cpu:
+        generator.cuda()
+
+    kp_detector = KPDetector(**config['model_params']['kp_detector_params'],
+                             **config['model_params']['common_params'])
+    if not cpu:
+        kp_detector.cuda()
+    
+    if cpu:
+        checkpoint = torch.load(checkpoint_path, map_location=torch.device('cpu'))
+    else:
+        checkpoint = torch.load(checkpoint_path)
+ 
+    generator.load_state_dict(checkpoint['generator'])
+    kp_detector.load_state_dict(checkpoint['kp_detector'])
+    
+    if not cpu:
+        generator = DataParallelWithCallback(generator)
+        kp_detector = DataParallelWithCallback(kp_detector)
+
+    generator.eval()
+    kp_detector.eval()
+    
+    return generator, kp_detector
+
+def make_animation(source_image, driving_video, generator, kp_detector, relative=True, adapt_movement_scale=True, cpu=False):
+    with torch.no_grad():
+        predictions = []
+        source = torch.tensor(source_image[np.newaxis].astype(np.float32)).permute(0, 3, 1, 2)
+        if not cpu:
+            source = source.cuda()
+        driving = torch.tensor(np.array(driving_video)[np.newaxis].astype(np.float32)).permute(0, 4, 1, 2, 3)
+        kp_source = kp_detector(source)
+        kp_driving_initial = kp_detector(driving[:, :, 0])
+
+        for frame_idx in tqdm(range(driving.shape[2])):
+            
+            driving_frame = driving[:, :, frame_idx]
+            if not cpu:
+                driving_frame = driving_frame.cuda()
+            kp_driving = kp_detector(driving_frame)
+            kp_norm = normalize_kp(kp_source=kp_source, kp_driving=kp_driving,
+                                   kp_driving_initial=kp_driving_initial, use_relative_movement=relative,
+                                   use_relative_jacobian=relative, adapt_movement_scale=adapt_movement_scale)
+            out = generator(source, kp_source=kp_source, kp_driving=kp_norm)
+
+            predictions.append(np.transpose(out['prediction'].data.cpu().numpy(), [0, 2, 3, 1])[0])
+    return predictions
+
+def find_best_frame(source, driving, cpu=False):
+    import face_alignment
+
+    def normalize_kp(kp):
+        kp = kp - kp.mean(axis=0, keepdims=True)
+        area = ConvexHull(kp[:, :2]).volume
+        area = np.sqrt(area)
+        kp[:, :2] = kp[:, :2] / area
+        return kp
+
+    fa = face_alignment.FaceAlignment(face_alignment.LandmarksType._2D, flip_input=True,
+                                      device='cpu' if cpu else 'cuda')
+    kp_source = fa.get_landmarks(255 * source)[0]
+    kp_source = normalize_kp(kp_source)
+    norm  = float('inf')
+    frame_num = 0
+    for i, image in tqdm(enumerate(driving)):
+        kp_driving = fa.get_landmarks(255 * image)[0]
+        kp_driving = normalize_kp(kp_driving)
+        new_norm = (np.abs(kp_source - kp_driving) ** 2).sum()
+        if new_norm < norm:
+            norm = new_norm
+            frame_num = i
+    return frame_num
+
+if __name__ == "__main__":
+    parser = ArgumentParser()
+    parser.add_argument("--config", required=True, help="path to config")
+    parser.add_argument("--checkpoint", default='vox-cpk.pth.tar', help="path to checkpoint to restore")
+
+    parser.add_argument("--source_image", default='sup-mat/source.png', help="path to source image")
+    parser.add_argument("--driving_video", default='sup-mat/source.png', help="path to driving video")
+    parser.add_argument("--result_video", default='result.mp4', help="path to output")
+ 
+    parser.add_argument("--relative", dest="relative", action="store_true", help="use relative or absolute keypoint coordinates")
+    parser.add_argument("--adapt_scale", dest="adapt_scale", action="store_true", help="adapt movement scale based on convex hull of keypoints")
+
+    parser.add_argument("--find_best_frame", dest="find_best_frame", action="store_true", 
+                        help="Generate from the frame that is the most alligned with source. (Only for faces, requires face_aligment lib)")
+
+    parser.add_argument("--best_frame", dest="best_frame", type=int, default=None,  
+                        help="Set frame to start from.")
+ 
+    parser.add_argument("--cpu", dest="cpu", action="store_true", help="cpu mode.")
+ 
+
+    parser.set_defaults(relative=False)
+    parser.set_defaults(adapt_scale=False)
+
+    opt = parser.parse_args()
+
+    source_image = imageio.imread(opt.source_image)
+    reader = imageio.get_reader(opt.driving_video)
+    fps = reader.get_meta_data()['fps']
+    driving_video = []
+    try:
+        for im in reader:
+            driving_video.append(im)
+    except RuntimeError:
+        pass
+    reader.close()
+
+    source_image = resize(source_image, (256, 256))[..., :3]
+    driving_video = [resize(frame, (256, 256))[..., :3] for frame in driving_video]
+    generator, kp_detector = load_checkpoints(config_path=opt.config, checkpoint_path=opt.checkpoint, cpu=opt.cpu)
+
+    if opt.find_best_frame or opt.best_frame is not None:
+        i = opt.best_frame if opt.best_frame is not None else find_best_frame(source_image, driving_video, cpu=opt.cpu)
+        print ("Best frame: " + str(i))
+        driving_forward = driving_video[i:]
+        driving_backward = driving_video[:(i+1)][::-1]
+        predictions_forward = make_animation(source_image, driving_forward, generator, kp_detector, relative=opt.relative, adapt_movement_scale=opt.adapt_scale, cpu=opt.cpu)
+        predictions_backward = make_animation(source_image, driving_backward, generator, kp_detector, relative=opt.relative, adapt_movement_scale=opt.adapt_scale, cpu=opt.cpu)
+        predictions = predictions_backward[::-1] + predictions_forward[1:]
+    else:
+        predictions = make_animation(source_image, driving_video, generator, kp_detector, relative=opt.relative, adapt_movement_scale=opt.adapt_scale, cpu=opt.cpu)
+    imageio.mimsave(opt.result_video, [img_as_ubyte(frame) for frame in predictions], fps=fps)
+

frames_dataset.py

@@ -0,0 +1,197 @@
+import os
+from skimage import io, img_as_float32
+from skimage.color import gray2rgb
+from sklearn.model_selection import train_test_split
+from imageio import mimread
+
+import numpy as np
+from torch.utils.data import Dataset
+import pandas as pd
+from augmentation import AllAugmentationTransform
+import glob
+
+
+def read_video(name, frame_shape):
+    """
+    Read video which can be:
+      - an image of concatenated frames
+      - '.mp4' and'.gif'
+      - folder with videos
+    """
+
+    if os.path.isdir(name):
+        frames = sorted(os.listdir(name))
+        num_frames = len(frames)
+        video_array = np.array(
+            [img_as_float32(io.imread(os.path.join(name, frames[idx]))) for idx in range(num_frames)])
+    elif name.lower().endswith('.png') or name.lower().endswith('.jpg'):
+        image = io.imread(name)
+
+        if len(image.shape) == 2 or image.shape[2] == 1:
+            image = gray2rgb(image)
+
+        if image.shape[2] == 4:
+            image = image[..., :3]
+
+        image = img_as_float32(image)
+
+        video_array = np.moveaxis(image, 1, 0)
+
+        video_array = video_array.reshape((-1,) + frame_shape)
+        video_array = np.moveaxis(video_array, 1, 2)
+    elif name.lower().endswith('.gif') or name.lower().endswith('.mp4') or name.lower().endswith('.mov'):
+        video = np.array(mimread(name))
+        if len(video.shape) == 3:
+            video = np.array([gray2rgb(frame) for frame in video])
+        if video.shape[-1] == 4:
+            video = video[..., :3]
+        video_array = img_as_float32(video)
+    else:
+        raise Exception("Unknown file extensions  %s" % name)
+
+    return video_array
+
+
+class FramesDataset(Dataset):
+    """
+    Dataset of videos, each video can be represented as:
+      - an image of concatenated frames
+      - '.mp4' or '.gif'
+      - folder with all frames
+    """
+
+    def __init__(self, root_dir, frame_shape=(256, 256, 3), id_sampling=False, is_train=True,
+                 random_seed=0, pairs_list=None, augmentation_params=None):
+        self.root_dir = root_dir
+        self.videos = os.listdir(root_dir)
+        self.frame_shape = tuple(frame_shape)
+        self.pairs_list = pairs_list
+        self.id_sampling = id_sampling
+        if os.path.exists(os.path.join(root_dir, 'train')):
+            assert os.path.exists(os.path.join(root_dir, 'test'))
+            print("Use predefined train-test split.")
+            if id_sampling:
+                train_videos = {os.path.basename(video).split('#')[0] for video in
+                                os.listdir(os.path.join(root_dir, 'train'))}
+                train_videos = list(train_videos)
+            else:
+                train_videos = os.listdir(os.path.join(root_dir, 'train'))
+            test_videos = os.listdir(os.path.join(root_dir, 'test'))
+            self.root_dir = os.path.join(self.root_dir, 'train' if is_train else 'test')
+        else:
+            print("Use random train-test split.")
+            train_videos, test_videos = train_test_split(self.videos, random_state=random_seed, test_size=0.2)
+
+        if is_train:
+            self.videos = train_videos
+        else:
+            self.videos = test_videos
+
+        self.is_train = is_train
+
+        if self.is_train:
+            self.transform = AllAugmentationTransform(**augmentation_params)
+        else:
+            self.transform = None
+
+    def __len__(self):
+        return len(self.videos)
+
+    def __getitem__(self, idx):
+        if self.is_train and self.id_sampling:
+            name = self.videos[idx]
+            path = np.random.choice(glob.glob(os.path.join(self.root_dir, name + '*.mp4')))
+        else:
+            name = self.videos[idx]
+            path = os.path.join(self.root_dir, name)
+
+        video_name = os.path.basename(path)
+
+        if self.is_train and os.path.isdir(path):
+            frames = os.listdir(path)
+            num_frames = len(frames)
+            frame_idx = np.sort(np.random.choice(num_frames, replace=True, size=2))
+            video_array = [img_as_float32(io.imread(os.path.join(path, frames[idx]))) for idx in frame_idx]
+        else:
+            video_array = read_video(path, frame_shape=self.frame_shape)
+            num_frames = len(video_array)
+            frame_idx = np.sort(np.random.choice(num_frames, replace=True, size=2)) if self.is_train else range(
+                num_frames)
+            video_array = video_array[frame_idx]
+
+        if self.transform is not None:
+            video_array = self.transform(video_array)
+
+        out = {}
+        if self.is_train:
+            source = np.array(video_array[0], dtype='float32')
+            driving = np.array(video_array[1], dtype='float32')
+
+            out['driving'] = driving.transpose((2, 0, 1))
+            out['source'] = source.transpose((2, 0, 1))
+        else:
+            video = np.array(video_array, dtype='float32')
+            out['video'] = video.transpose((3, 0, 1, 2))
+
+        out['name'] = video_name
+
+        return out
+
+
+class DatasetRepeater(Dataset):
+    """
+    Pass several times over the same dataset for better i/o performance
+    """
+
+    def __init__(self, dataset, num_repeats=100):
+        self.dataset = dataset
+        self.num_repeats = num_repeats
+
+    def __len__(self):
+        return self.num_repeats * self.dataset.__len__()
+
+    def __getitem__(self, idx):
+        return self.dataset[idx % self.dataset.__len__()]
+
+
+class PairedDataset(Dataset):
+    """
+    Dataset of pairs for animation.
+    """
+
+    def __init__(self, initial_dataset, number_of_pairs, seed=0):
+        self.initial_dataset = initial_dataset
+        pairs_list = self.initial_dataset.pairs_list
+
+        np.random.seed(seed)
+
+        if pairs_list is None:
+            max_idx = min(number_of_pairs, len(initial_dataset))
+            nx, ny = max_idx, max_idx
+            xy = np.mgrid[:nx, :ny].reshape(2, -1).T
+            number_of_pairs = min(xy.shape[0], number_of_pairs)
+            self.pairs = xy.take(np.random.choice(xy.shape[0], number_of_pairs, replace=False), axis=0)
+        else:
+            videos = self.initial_dataset.videos
+            name_to_index = {name: index for index, name in enumerate(videos)}
+            pairs = pd.read_csv(pairs_list)
+            pairs = pairs[np.logical_and(pairs['source'].isin(videos), pairs['driving'].isin(videos))]
+
+            number_of_pairs = min(pairs.shape[0], number_of_pairs)
+            self.pairs = []
+            self.start_frames = []
+            for ind in range(number_of_pairs):
+                self.pairs.append(
+                    (name_to_index[pairs['driving'].iloc[ind]], name_to_index[pairs['source'].iloc[ind]]))
+
+    def __len__(self):
+        return len(self.pairs)
+
+    def __getitem__(self, idx):
+        pair = self.pairs[idx]
+        first = self.initial_dataset[pair[0]]
+        second = self.initial_dataset[pair[1]]
+        first = {'driving_' + key: value for key, value in first.items()}
+        second = {'source_' + key: value for key, value in second.items()}
+
+        return {**first, **second}

logger.py

@@ -0,0 +1,208 @@
+import numpy as np
+import torch
+import torch.nn.functional as F
+import imageio
+
+import os
+from skimage.draw import circle
+
+import matplotlib.pyplot as plt
+import collections
+
+
+class Logger:
+    def __init__(self, log_dir, checkpoint_freq=100, visualizer_params=None, zfill_num=8, log_file_name='log.txt'):
+
+        self.loss_list = []
+        self.cpk_dir = log_dir
+        self.visualizations_dir = os.path.join(log_dir, 'train-vis')
+        if not os.path.exists(self.visualizations_dir):
+            os.makedirs(self.visualizations_dir)
+        self.log_file = open(os.path.join(log_dir, log_file_name), 'a')
+        self.zfill_num = zfill_num
+        self.visualizer = Visualizer(**visualizer_params)
+        self.checkpoint_freq = checkpoint_freq
+        self.epoch = 0
+        self.best_loss = float('inf')
+        self.names = None
+
+    def log_scores(self, loss_names):
+        loss_mean = np.array(self.loss_list).mean(axis=0)
+
+        loss_string = "; ".join(["%s - %.5f" % (name, value) for name, value in zip(loss_names, loss_mean)])
+        loss_string = str(self.epoch).zfill(self.zfill_num) + ") " + loss_string
+
+        print(loss_string, file=self.log_file)
+        self.loss_list = []
+        self.log_file.flush()
+
+    def visualize_rec(self, inp, out):
+        image = self.visualizer.visualize(inp['driving'], inp['source'], out)
+        imageio.imsave(os.path.join(self.visualizations_dir, "%s-rec.png" % str(self.epoch).zfill(self.zfill_num)), image)
+
+    def save_cpk(self, emergent=False):
+        cpk = {k: v.state_dict() for k, v in self.models.items()}
+        cpk['epoch'] = self.epoch
+        cpk_path = os.path.join(self.cpk_dir, '%s-checkpoint.pth.tar' % str(self.epoch).zfill(self.zfill_num)) 
+        if not (os.path.exists(cpk_path) and emergent):
+            torch.save(cpk, cpk_path)
+
+    @staticmethod
+    def load_cpk(checkpoint_path, generator=None, discriminator=None, kp_detector=None,
+                 optimizer_generator=None, optimizer_discriminator=None, optimizer_kp_detector=None):
+        checkpoint = torch.load(checkpoint_path)
+        if generator is not None:
+            generator.load_state_dict(checkpoint['generator'])
+        if kp_detector is not None:
+            kp_detector.load_state_dict(checkpoint['kp_detector'])
+        if discriminator is not None:
+            try:
+               discriminator.load_state_dict(checkpoint['discriminator'])
+            except:
+               print ('No discriminator in the state-dict. Dicriminator will be randomly initialized')
+        if optimizer_generator is not None:
+            optimizer_generator.load_state_dict(checkpoint['optimizer_generator'])
+        if optimizer_discriminator is not None:
+            try:
+                optimizer_discriminator.load_state_dict(checkpoint['optimizer_discriminator'])
+            except RuntimeError as e:
+                print ('No discriminator optimizer in the state-dict. Optimizer will be not initialized')
+        if optimizer_kp_detector is not None:
+            optimizer_kp_detector.load_state_dict(checkpoint['optimizer_kp_detector'])
+
+        return checkpoint['epoch']
+
+    def __enter__(self):
+        return self
+
+    def __exit__(self, exc_type, exc_val, exc_tb):
+        if 'models' in self.__dict__:
+            self.save_cpk()
+        self.log_file.close()
+
+    def log_iter(self, losses):
+        losses = collections.OrderedDict(losses.items())
+        if self.names is None:
+            self.names = list(losses.keys())
+        self.loss_list.append(list(losses.values()))
+
+    def log_epoch(self, epoch, models, inp, out):
+        self.epoch = epoch
+        self.models = models
+        if (self.epoch + 1) % self.checkpoint_freq == 0:
+            self.save_cpk()
+        self.log_scores(self.names)
+        self.visualize_rec(inp, out)
+
+
+class Visualizer:
+    def __init__(self, kp_size=5, draw_border=False, colormap='gist_rainbow'):
+        self.kp_size = kp_size
+        self.draw_border = draw_border
+        self.colormap = plt.get_cmap(colormap)
+
+    def draw_image_with_kp(self, image, kp_array):
+        image = np.copy(image)
+        spatial_size = np.array(image.shape[:2][::-1])[np.newaxis]
+        kp_array = spatial_size * (kp_array + 1) / 2
+        num_kp = kp_array.shape[0]
+        for kp_ind, kp in enumerate(kp_array):
+            rr, cc = circle(kp[1], kp[0], self.kp_size, shape=image.shape[:2])
+            image[rr, cc] = np.array(self.colormap(kp_ind / num_kp))[:3]
+        return image
+
+    def create_image_column_with_kp(self, images, kp):
+        image_array = np.array([self.draw_image_with_kp(v, k) for v, k in zip(images, kp)])
+        return self.create_image_column(image_array)
+
+    def create_image_column(self, images):
+        if self.draw_border:
+            images = np.copy(images)
+            images[:, :, [0, -1]] = (1, 1, 1)
+            images[:, :, [0, -1]] = (1, 1, 1)
+        return np.concatenate(list(images), axis=0)
+
+    def create_image_grid(self, *args):
+        out = []
+        for arg in args:
+            if type(arg) == tuple:
+                out.append(self.create_image_column_with_kp(arg[0], arg[1]))
+            else:
+                out.append(self.create_image_column(arg))
+        return np.concatenate(out, axis=1)
+
+    def visualize(self, driving, source, out):
+        images = []
+
+        # Source image with keypoints
+        source = source.data.cpu()
+        kp_source = out['kp_source']['value'].data.cpu().numpy()
+        source = np.transpose(source, [0, 2, 3, 1])
+        images.append((source, kp_source))
+
+        # Equivariance visualization
+        if 'transformed_frame' in out:
+            transformed = out['transformed_frame'].data.cpu().numpy()
+            transformed = np.transpose(transformed, [0, 2, 3, 1])
+            transformed_kp = out['transformed_kp']['value'].data.cpu().numpy()
+            images.append((transformed, transformed_kp))
+
+        # Driving image with keypoints
+        kp_driving = out['kp_driving']['value'].data.cpu().numpy()
+        driving = driving.data.cpu().numpy()
+        driving = np.transpose(driving, [0, 2, 3, 1])
+        images.append((driving, kp_driving))
+
+        # Deformed image
+        if 'deformed' in out:
+            deformed = out['deformed'].data.cpu().numpy()
+            deformed = np.transpose(deformed, [0, 2, 3, 1])
+            images.append(deformed)
+
+        # Result with and without keypoints
+        prediction = out['prediction'].data.cpu().numpy()
+        prediction = np.transpose(prediction, [0, 2, 3, 1])
+        if 'kp_norm' in out:
+            kp_norm = out['kp_norm']['value'].data.cpu().numpy()
+            images.append((prediction, kp_norm))
+        images.append(prediction)
+
+
+        ## Occlusion map
+        if 'occlusion_map' in out:
+            occlusion_map = out['occlusion_map'].data.cpu().repeat(1, 3, 1, 1)
+            occlusion_map = F.interpolate(occlusion_map, size=source.shape[1:3]).numpy()
+            occlusion_map = np.transpose(occlusion_map, [0, 2, 3, 1])
+            images.append(occlusion_map)
+
+        # Deformed images according to each individual transform
+        if 'sparse_deformed' in out:
+            full_mask = []
+            for i in range(out['sparse_deformed'].shape[1]):
+                image = out['sparse_deformed'][:, i].data.cpu()
+                image = F.interpolate(image, size=source.shape[1:3])
+                mask = out['mask'][:, i:(i+1)].data.cpu().repeat(1, 3, 1, 1)
+                mask = F.interpolate(mask, size=source.shape[1:3])
+                image = np.transpose(image.numpy(), (0, 2, 3, 1))
+                mask = np.transpose(mask.numpy(), (0, 2, 3, 1))
+
+                if i != 0:
+                    color = np.array(self.colormap((i - 1) / (out['sparse_deformed'].shape[1] - 1)))[:3]
+                else:
+                    color = np.array((0, 0, 0))
+
+                color = color.reshape((1, 1, 1, 3))
+
+                images.append(image)
+                if i != 0:
+                    images.append(mask * color)
+                else:
+                    images.append(mask)
+
+                full_mask.append(mask * color)
+
+            images.append(sum(full_mask))
+
+        image = self.create_image_grid(*images)
+        image = (255 * image).astype(np.uint8)
+        return image

reconstruction.py

@@ -0,0 +1,67 @@
+import os
+from tqdm import tqdm
+import torch
+from torch.utils.data import DataLoader
+from logger import Logger, Visualizer
+import numpy as np
+import imageio
+from sync_batchnorm import DataParallelWithCallback
+
+
+def reconstruction(config, generator, kp_detector, checkpoint, log_dir, dataset):
+    png_dir = os.path.join(log_dir, 'reconstruction/png')
+    log_dir = os.path.join(log_dir, 'reconstruction')
+
+    if checkpoint is not None:
+        Logger.load_cpk(checkpoint, generator=generator, kp_detector=kp_detector)
+    else:
+        raise AttributeError("Checkpoint should be specified for mode='reconstruction'.")
+    dataloader = DataLoader(dataset, batch_size=1, shuffle=False, num_workers=1)
+
+    if not os.path.exists(log_dir):
+        os.makedirs(log_dir)
+
+    if not os.path.exists(png_dir):
+        os.makedirs(png_dir)
+
+    loss_list = []
+    if torch.cuda.is_available():
+        generator = DataParallelWithCallback(generator)
+        kp_detector = DataParallelWithCallback(kp_detector)
+
+    generator.eval()
+    kp_detector.eval()
+
+    for it, x in tqdm(enumerate(dataloader)):
+        if config['reconstruction_params']['num_videos'] is not None:
+            if it > config['reconstruction_params']['num_videos']:
+                break
+        with torch.no_grad():
+            predictions = []
+            visualizations = []
+            if torch.cuda.is_available():
+                x['video'] = x['video'].cuda()
+            kp_source = kp_detector(x['video'][:, :, 0])
+            for frame_idx in range(x['video'].shape[2]):
+                source = x['video'][:, :, 0]
+                driving = x['video'][:, :, frame_idx]
+                kp_driving = kp_detector(driving)
+                out = generator(source, kp_source=kp_source, kp_driving=kp_driving)
+                out['kp_source'] = kp_source
+                out['kp_driving'] = kp_driving
+                del out['sparse_deformed']
+                predictions.append(np.transpose(out['prediction'].data.cpu().numpy(), [0, 2, 3, 1])[0])
+
+                visualization = Visualizer(**config['visualizer_params']).visualize(source=source,
+                                                                                    driving=driving, out=out)
+                visualizations.append(visualization)
+
+                loss_list.append(torch.abs(out['prediction'] - driving).mean().cpu().numpy())
+
+            predictions = np.concatenate(predictions, axis=1)
+            imageio.imsave(os.path.join(png_dir, x['name'][0] + '.png'), (255 * predictions).astype(np.uint8))
+
+            image_name = x['name'][0] + config['reconstruction_params']['format']
+            imageio.mimsave(os.path.join(log_dir, image_name), visualizations)
+
+    print("Reconstruction loss: %s" % np.mean(loss_list))

run.py

@@ -0,0 +1,87 @@
+import matplotlib
+
+matplotlib.use('Agg')
+
+import os, sys
+import yaml
+from argparse import ArgumentParser
+from time import gmtime, strftime
+from shutil import copy
+
+from frames_dataset import FramesDataset
+
+from modules.generator import OcclusionAwareGenerator
+from modules.discriminator import MultiScaleDiscriminator
+from modules.keypoint_detector import KPDetector
+
+import torch
+
+from train import train
+from reconstruction import reconstruction
+from animate import animate
+
+if __name__ == "__main__":
+    
+    if sys.version_info[0] < 3:
+        raise Exception("You must use Python 3 or higher. Recommended version is Python 3.7")
+
+    parser = ArgumentParser()
+    parser.add_argument("--config", required=True, help="path to config")
+    parser.add_argument("--mode", default="train", choices=["train", "reconstruction", "animate"])
+    parser.add_argument("--log_dir", default='log', help="path to log into")
+    parser.add_argument("--checkpoint", default=None, help="path to checkpoint to restore")
+    parser.add_argument("--device_ids", default="0", type=lambda x: list(map(int, x.split(','))),
+                        help="Names of the devices comma separated.")
+    parser.add_argument("--verbose", dest="verbose", action="store_true", help="Print model architecture")
+    parser.set_defaults(verbose=False)
+
+    opt = parser.parse_args()
+    with open(opt.config) as f:
+        config = yaml.load(f)
+
+    if opt.checkpoint is not None:
+        log_dir = os.path.join(*os.path.split(opt.checkpoint)[:-1])
+    else:
+        log_dir = os.path.join(opt.log_dir, os.path.basename(opt.config).split('.')[0])
+        log_dir += ' ' + strftime("%d_%m_%y_%H.%M.%S", gmtime())
+
+    generator = OcclusionAwareGenerator(**config['model_params']['generator_params'],
+                                        **config['model_params']['common_params'])
+
+    if torch.cuda.is_available():
+        generator.to(opt.device_ids[0])
+    if opt.verbose:
+        print(generator)
+
+    discriminator = MultiScaleDiscriminator(**config['model_params']['discriminator_params'],
+                                            **config['model_params']['common_params'])
+    if torch.cuda.is_available():
+        discriminator.to(opt.device_ids[0])
+    if opt.verbose:
+        print(discriminator)
+
+    kp_detector = KPDetector(**config['model_params']['kp_detector_params'],
+                             **config['model_params']['common_params'])
+
+    if torch.cuda.is_available():
+        kp_detector.to(opt.device_ids[0])
+
+    if opt.verbose:
+        print(kp_detector)
+
+    dataset = FramesDataset(is_train=(opt.mode == 'train'), **config['dataset_params'])
+
+    if not os.path.exists(log_dir):
+        os.makedirs(log_dir)
+    if not os.path.exists(os.path.join(log_dir, os.path.basename(opt.config))):
+        copy(opt.config, log_dir)
+
+    if opt.mode == 'train':
+        print("Training...")
+        train(config, generator, discriminator, kp_detector, opt.checkpoint, log_dir, dataset, opt.device_ids)
+    elif opt.mode == 'reconstruction':
+        print("Reconstruction...")
+        reconstruction(config, generator, kp_detector, opt.checkpoint, log_dir, dataset)
+    elif opt.mode == 'animate':
+        print("Animate...")
+        animate(config, generator, kp_detector, opt.checkpoint, log_dir, dataset)

train.py

@@ -0,0 +1,87 @@
+from tqdm import trange
+import torch
+
+from torch.utils.data import DataLoader
+
+from logger import Logger
+from modules.model import GeneratorFullModel, DiscriminatorFullModel
+
+from torch.optim.lr_scheduler import MultiStepLR
+
+from sync_batchnorm import DataParallelWithCallback
+
+from frames_dataset import DatasetRepeater
+
+
+def train(config, generator, discriminator, kp_detector, checkpoint, log_dir, dataset, device_ids):
+    train_params = config['train_params']
+
+    optimizer_generator = torch.optim.Adam(generator.parameters(), lr=train_params['lr_generator'], betas=(0.5, 0.999))
+    optimizer_discriminator = torch.optim.Adam(discriminator.parameters(), lr=train_params['lr_discriminator'], betas=(0.5, 0.999))
+    optimizer_kp_detector = torch.optim.Adam(kp_detector.parameters(), lr=train_params['lr_kp_detector'], betas=(0.5, 0.999))
+
+    if checkpoint is not None:
+        start_epoch = Logger.load_cpk(checkpoint, generator, discriminator, kp_detector,
+                                      optimizer_generator, optimizer_discriminator,
+                                      None if train_params['lr_kp_detector'] == 0 else optimizer_kp_detector)
+    else:
+        start_epoch = 0
+
+    scheduler_generator = MultiStepLR(optimizer_generator, train_params['epoch_milestones'], gamma=0.1,
+                                      last_epoch=start_epoch - 1)
+    scheduler_discriminator = MultiStepLR(optimizer_discriminator, train_params['epoch_milestones'], gamma=0.1,
+                                          last_epoch=start_epoch - 1)
+    scheduler_kp_detector = MultiStepLR(optimizer_kp_detector, train_params['epoch_milestones'], gamma=0.1,
+                                        last_epoch=-1 + start_epoch * (train_params['lr_kp_detector'] != 0))
+
+    if 'num_repeats' in train_params or train_params['num_repeats'] != 1:
+        dataset = DatasetRepeater(dataset, train_params['num_repeats'])
+    dataloader = DataLoader(dataset, batch_size=train_params['batch_size'], shuffle=True, num_workers=6, drop_last=True)
+
+    generator_full = GeneratorFullModel(kp_detector, generator, discriminator, train_params)
+    discriminator_full = DiscriminatorFullModel(kp_detector, generator, discriminator, train_params)
+
+    if torch.cuda.is_available():
+        generator_full = DataParallelWithCallback(generator_full, device_ids=device_ids)
+        discriminator_full = DataParallelWithCallback(discriminator_full, device_ids=device_ids)
+
+    with Logger(log_dir=log_dir, visualizer_params=config['visualizer_params'], checkpoint_freq=train_params['checkpoint_freq']) as logger:
+        for epoch in trange(start_epoch, train_params['num_epochs']):
+            for x in dataloader:
+                losses_generator, generated = generator_full(x)
+
+                loss_values = [val.mean() for val in losses_generator.values()]
+                loss = sum(loss_values)
+
+                loss.backward()
+                optimizer_generator.step()
+                optimizer_generator.zero_grad()
+                optimizer_kp_detector.step()
+                optimizer_kp_detector.zero_grad()
+
+                if train_params['loss_weights']['generator_gan'] != 0:
+                    optimizer_discriminator.zero_grad()
+                    losses_discriminator = discriminator_full(x, generated)
+                    loss_values = [val.mean() for val in losses_discriminator.values()]
+                    loss = sum(loss_values)
+
+                    loss.backward()
+                    optimizer_discriminator.step()
+                    optimizer_discriminator.zero_grad()
+                else:
+                    losses_discriminator = {}
+
+                losses_generator.update(losses_discriminator)
+                losses = {key: value.mean().detach().data.cpu().numpy() for key, value in losses_generator.items()}
+                logger.log_iter(losses=losses)
+
+            scheduler_generator.step()
+            scheduler_discriminator.step()
+            scheduler_kp_detector.step()
+            
+            logger.log_epoch(epoch, {'generator': generator,
+                                     'discriminator': discriminator,
+                                     'kp_detector': kp_detector,
+                                     'optimizer_generator': optimizer_generator,
+                                     'optimizer_discriminator': optimizer_discriminator,
+                                     'optimizer_kp_detector': optimizer_kp_detector}, inp=x, out=generated)