Upload 168 files

.gitattributes

@@ -32,3 +32,9 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+sup-mat/absolute-demo.gif filter=lfs diff=lfs merge=lfs -text
+sup-mat/face-swap.gif filter=lfs diff=lfs merge=lfs -text
+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

config/bair-256.yaml

@@ -0,0 +1,82 @@
+dataset_params:
+  root_dir: data/bair
+  frame_shape: [256, 256, 3]
+  id_sampling: False
+  augmentation_params:
+    flip_param:
+      horizontal_flip: True
+      time_flip: True
+    jitter_param:
+      brightness: 0.1
+      contrast: 0.1
+      saturation: 0.1
+      hue: 0.1
+
+
+model_params:
+  common_params:
+    num_kp: 10
+    num_channels: 3
+    estimate_jacobian: True
+  kp_detector_params:
+     temperature: 0.1
+     block_expansion: 32
+     max_features: 1024
+     scale_factor: 0.25
+     num_blocks: 5
+  generator_params:
+    block_expansion: 64
+    max_features: 512
+    num_down_blocks: 2
+    num_bottleneck_blocks: 6
+    estimate_occlusion_map: True
+    dense_motion_params:
+      block_expansion: 64
+      max_features: 1024
+      num_blocks: 5
+      scale_factor: 0.25
+  discriminator_params:
+    scales: [1]
+    block_expansion: 32
+    max_features: 512
+    num_blocks: 4
+    sn: True
+
+train_params:
+  num_epochs: 20
+  num_repeats: 1
+  epoch_milestones: [12, 18]
+  lr_generator: 2.0e-4
+  lr_discriminator: 2.0e-4
+  lr_kp_detector: 2.0e-4
+  batch_size: 36
+  scales: [1, 0.5, 0.25, 0.125]
+  checkpoint_freq: 10
+  transform_params:
+    sigma_affine: 0.05
+    sigma_tps: 0.005
+    points_tps: 5
+  loss_weights:
+    generator_gan: 1
+    discriminator_gan: 1
+    feature_matching: [10, 10, 10, 10]
+    perceptual: [10, 10, 10, 10, 10]
+    equivariance_value: 10
+    equivariance_jacobian: 10
+
+reconstruction_params:
+  num_videos: 1000
+  format: '.mp4'
+
+animate_params:
+  num_pairs: 50
+  format: '.mp4'
+  normalization_params:
+    adapt_movement_scale: False
+    use_relative_movement: True
+    use_relative_jacobian: True
+
+visualizer_params:
+  kp_size: 5
+  draw_border: True
+  colormap: 'gist_rainbow'

config/fashion-256.yaml

@@ -0,0 +1,77 @@
+dataset_params:
+  root_dir: data/fashion-png
+  frame_shape: [256, 256, 3]
+  id_sampling: False
+  augmentation_params:
+    flip_param:
+      horizontal_flip: True
+      time_flip: True
+    jitter_param:
+      hue: 0.1
+
+model_params:
+  common_params:
+    num_kp: 10
+    num_channels: 3
+    estimate_jacobian: True
+  kp_detector_params:
+     temperature: 0.1
+     block_expansion: 32
+     max_features: 1024
+     scale_factor: 0.25
+     num_blocks: 5
+  generator_params:
+    block_expansion: 64
+    max_features: 512
+    num_down_blocks: 2
+    num_bottleneck_blocks: 6
+    estimate_occlusion_map: True
+    dense_motion_params:
+      block_expansion: 64
+      max_features: 1024
+      num_blocks: 5
+      scale_factor: 0.25
+  discriminator_params:
+    scales: [1]
+    block_expansion: 32
+    max_features: 512
+    num_blocks: 4
+
+train_params:
+  num_epochs: 100
+  num_repeats: 50
+  epoch_milestones: [60, 90]
+  lr_generator: 2.0e-4
+  lr_discriminator: 2.0e-4
+  lr_kp_detector: 2.0e-4
+  batch_size: 27
+  scales: [1, 0.5, 0.25, 0.125]
+  checkpoint_freq: 50
+  transform_params:
+    sigma_affine: 0.05
+    sigma_tps: 0.005
+    points_tps: 5
+  loss_weights:
+    generator_gan: 1
+    discriminator_gan: 1
+    feature_matching: [10, 10, 10, 10]
+    perceptual: [10, 10, 10, 10, 10]
+    equivariance_value: 10
+    equivariance_jacobian: 10
+
+reconstruction_params:
+  num_videos: 1000
+  format: '.mp4'
+
+animate_params:
+  num_pairs: 50
+  format: '.mp4'
+  normalization_params:
+    adapt_movement_scale: False
+    use_relative_movement: True
+    use_relative_jacobian: True
+
+visualizer_params:
+  kp_size: 5
+  draw_border: True
+  colormap: 'gist_rainbow'

config/mgif-256.yaml

@@ -0,0 +1,84 @@
+dataset_params:
+  root_dir: data/moving-gif
+  frame_shape: [256, 256, 3]
+  id_sampling: False
+  augmentation_params:
+    flip_param:
+      horizontal_flip: True
+      time_flip: True
+    crop_param:
+      size: [256, 256]
+    resize_param:
+      ratio: [0.9, 1.1]
+    jitter_param:
+      hue: 0.5
+
+model_params:
+  common_params:
+    num_kp: 10
+    num_channels: 3
+    estimate_jacobian: True
+  kp_detector_params:
+     temperature: 0.1
+     block_expansion: 32
+     max_features: 1024
+     scale_factor: 0.25
+     num_blocks: 5
+     single_jacobian_map: True
+  generator_params:
+    block_expansion: 64
+    max_features: 512
+    num_down_blocks: 2
+    num_bottleneck_blocks: 6
+    estimate_occlusion_map: True
+    dense_motion_params:
+      block_expansion: 64
+      max_features: 1024
+      num_blocks: 5
+      scale_factor: 0.25
+  discriminator_params:
+    scales: [1]
+    block_expansion: 32
+    max_features: 512
+    num_blocks: 4
+    sn: True
+
+train_params:
+  num_epochs: 100
+  num_repeats: 25
+  epoch_milestones: [60, 90]
+  lr_generator: 2.0e-4
+  lr_discriminator: 2.0e-4
+  lr_kp_detector: 2.0e-4
+ 
+  batch_size: 36
+  scales: [1, 0.5, 0.25, 0.125]
+  checkpoint_freq: 100
+  transform_params:
+    sigma_affine: 0.05
+    sigma_tps: 0.005
+    points_tps: 5
+  loss_weights:
+    generator_gan: 1
+    discriminator_gan: 1
+    feature_matching: [10, 10, 10, 10]
+    perceptual: [10, 10, 10, 10, 10]
+    equivariance_value: 10
+    equivariance_jacobian: 10
+
+reconstruction_params:
+  num_videos: 1000
+  format: '.mp4'
+
+animate_params:
+  num_pairs: 50
+  format: '.mp4'
+  normalization_params:
+    adapt_movement_scale: False
+    use_relative_movement: True
+    use_relative_jacobian: True
+
+visualizer_params:
+  kp_size: 5
+  draw_border: True
+  colormap: 'gist_rainbow'

config/nemo-256.yaml

@@ -0,0 +1,76 @@
+dataset_params:
+  root_dir: data/nemo-png 
+  frame_shape: [256, 256, 3]
+  id_sampling: False
+  augmentation_params:
+    flip_param:
+      horizontal_flip: True
+      time_flip: True
+
+model_params:
+  common_params:
+    num_kp: 10
+    num_channels: 3
+    estimate_jacobian: True
+  kp_detector_params:
+     temperature: 0.1
+     block_expansion: 32
+     max_features: 1024
+     scale_factor: 0.25
+     num_blocks: 5
+  generator_params:
+    block_expansion: 64
+    max_features: 512
+    num_down_blocks: 2
+    num_bottleneck_blocks: 6
+    estimate_occlusion_map: True
+    dense_motion_params:
+      block_expansion: 64
+      max_features: 1024
+      num_blocks: 5
+      scale_factor: 0.25
+  discriminator_params:
+    scales: [1]
+    block_expansion: 32
+    max_features: 512
+    num_blocks: 4
+    sn: True
+
+train_params:
+  num_epochs: 100
+  num_repeats: 8
+  epoch_milestones: [60, 90]
+  lr_generator: 2.0e-4
+  lr_discriminator: 2.0e-4
+  lr_kp_detector: 2.0e-4
+  batch_size: 36
+  scales: [1, 0.5, 0.25, 0.125]
+  checkpoint_freq: 50
+  transform_params:
+    sigma_affine: 0.05
+    sigma_tps: 0.005
+    points_tps: 5
+  loss_weights:
+    generator_gan: 1
+    discriminator_gan: 1
+    feature_matching: [10, 10, 10, 10]
+    perceptual: [10, 10, 10, 10, 10]
+    equivariance_value: 10
+    equivariance_jacobian: 10
+
+reconstruction_params:
+  num_videos: 1000
+  format: '.mp4'
+
+animate_params:
+  num_pairs: 50
+  format: '.mp4'
+  normalization_params:
+    adapt_movement_scale: False
+    use_relative_movement: True
+    use_relative_jacobian: True
+
+visualizer_params:
+  kp_size: 5
+  draw_border: True
+  colormap: 'gist_rainbow'

config/taichi-256.yaml

@@ -0,0 +1,157 @@
+# Dataset parameters
+# Each dataset should contain 2 folders train and test
+# Each video can be represented as:
+#   - an image of concatenated frames
+#   - '.mp4' or '.gif'
+#   - folder with all frames from a specific video
+# In case of Taichi. Same (youtube) video can be splitted in many parts (chunks). Each part has a following
+# format (id)#other#info.mp4. For example '12335#adsbf.mp4' has an id 12335. In case of TaiChi id stands for youtube
+# video id.
+dataset_params:
+  # Path to data, data can be stored in several formats: .mp4 or .gif videos, stacked .png images or folders with frames.
+  root_dir: data/taichi-png
+  # Image shape, needed for staked .png format.
+  frame_shape: [256, 256, 3]
+  # In case of TaiChi single video can be splitted in many chunks, or the maybe several videos for single person.
+  # In this case epoch can be a pass over different videos (if id_sampling=True) or over different chunks (if id_sampling=False)
+  # If the name of the video '12335#adsbf.mp4' the id is assumed to be 12335
+  id_sampling: True
+  # List with pairs for animation, None for random pairs
+  pairs_list: data/taichi256.csv
+  # Augmentation parameters see augmentation.py for all posible augmentations
+  augmentation_params:
+    flip_param:
+      horizontal_flip: True
+      time_flip: True
+    jitter_param:
+      brightness: 0.1
+      contrast: 0.1
+      saturation: 0.1
+      hue: 0.1
+
+# Defines model architecture
+model_params:
+  common_params:
+    # Number of keypoint
+    num_kp: 10
+    # Number of channels per image
+    num_channels: 3
+    # Using first or zero order model
+    estimate_jacobian: True
+  kp_detector_params:
+     # Softmax temperature for keypoint heatmaps
+     temperature: 0.1
+     # Number of features mutliplier
+     block_expansion: 32
+     # Maximum allowed number of features
+     max_features: 1024
+     # Number of block in Unet. Can be increased or decreased depending or resolution.
+     num_blocks: 5
+     # Keypioint is predicted on smaller images for better performance,
+     # scale_factor=0.25 means that 256x256 image will be resized to 64x64
+     scale_factor: 0.25
+  generator_params:
+    # Number of features mutliplier
+    block_expansion: 64
+    # Maximum allowed number of features
+    max_features: 512
+    # Number of downsampling blocks in Jonson architecture.
+    # Can be increased or decreased depending or resolution.
+    num_down_blocks: 2
+    # Number of ResBlocks  in Jonson architecture.
+    num_bottleneck_blocks: 6
+    # Use occlusion map or not
+    estimate_occlusion_map: True
+
+    dense_motion_params:
+      # Number of features mutliplier
+      block_expansion: 64
+      # Maximum allowed number of features
+      max_features: 1024
+      # Number of block in Unet. Can be increased or decreased depending or resolution.
+      num_blocks: 5
+      # Dense motion is predicted on smaller images for better performance,
+      # scale_factor=0.25 means that 256x256 image will be resized to 64x64
+      scale_factor: 0.25
+  discriminator_params:
+    # Discriminator can be multiscale, if you want 2 discriminator on original
+    # resolution and half of the original, specify scales: [1, 0.5]
+    scales: [1]
+    # Number of features mutliplier
+    block_expansion: 32
+    # Maximum allowed number of features
+    max_features: 512
+    # Number of blocks. Can be increased or decreased depending or resolution.
+    num_blocks: 4
+
+# Parameters of training
+train_params:
+  # Number of training epochs 
+  num_epochs: 100
+  # For better i/o performance when number of videos is small number of epochs can be multiplied by this number.
+  # Thus effectivlly with num_repeats=100 each epoch is 100 times larger. 
+  num_repeats: 150
+  # Drop learning rate by 10 times after this epochs 
+  epoch_milestones: [60, 90]
+  # Initial learing rate for all modules
+  lr_generator: 2.0e-4
+  lr_discriminator: 2.0e-4
+  lr_kp_detector: 2.0e-4
+  batch_size: 30
+  # Scales for perceptual pyramide loss. If scales = [1, 0.5, 0.25, 0.125] and image resolution is 256x256,
+  # than the loss will be computer on resolutions 256x256, 128x128, 64x64, 32x32.
+  scales: [1, 0.5, 0.25, 0.125]
+  # Save checkpoint this frequently. If checkpoint_freq=50, checkpoint will be saved every 50 epochs.
+  checkpoint_freq: 50
+  # Parameters of transform for equivariance loss
+  transform_params:
+    # Sigma for affine part
+    sigma_affine: 0.05
+    # Sigma for deformation part
+    sigma_tps: 0.005
+    # Number of point in the deformation grid
+    points_tps: 5
+  loss_weights:
+    # Weight for LSGAN loss in generator, 0 for no adversarial loss.
+    generator_gan: 0
+    # Weight for LSGAN loss in discriminator
+    discriminator_gan: 1
+    # Weights for feature matching loss, the number should be the same as number of blocks in discriminator.
+    feature_matching: [10, 10, 10, 10]
+    # Weights for perceptual loss.
+    perceptual: [10, 10, 10, 10, 10]
+    # Weights for value equivariance.
+    equivariance_value: 10
+    # Weights for jacobian equivariance.
+    equivariance_jacobian: 10
+
+# Parameters of reconstruction
+reconstruction_params:
+  # Maximum number of videos for reconstruction
+  num_videos: 1000
+  # Format for visualization, note that results will be also stored in staked .png.
+  format: '.mp4'
+
+# Parameters of animation
+animate_params:
+  # Maximum number of pairs for animation, the pairs will be either taken from pairs_list or random.
+  num_pairs: 50
+  # Format for visualization, note that results will be also stored in staked .png.
+  format: '.mp4'
+  # Normalization of diriving keypoints
+  normalization_params:
+    # Increase or decrease relative movement scale depending on the size of the object
+    adapt_movement_scale: False
+    # Apply only relative displacement of the keypoint
+    use_relative_movement: True
+    # Apply only relative change in jacobian
+    use_relative_jacobian: True
+
+# Visualization parameters
+visualizer_params:
+  # Draw keypoints of this size, increase or decrease depending on resolution
+  kp_size: 5
+  # Draw white border around images
+  draw_border: True
+  # Color map for keypoints
+  colormap: 'gist_rainbow'

config/taichi-adv-256.yaml

@@ -0,0 +1,150 @@
+# Dataset parameters
+dataset_params:
+  # Path to data, data can be stored in several formats: .mp4 or .gif videos, stacked .png images or folders with frames.
+  root_dir: data/taichi-png
+  # Image shape, needed for staked .png format.
+  frame_shape: [256, 256, 3]
+  # In case of TaiChi single video can be splitted in many chunks, or the maybe several videos for single person.
+  # In this case epoch can be a pass over different videos (if id_sampling=True) or over different chunks (if id_sampling=False)
+  # If the name of the video '12335#adsbf.mp4' the id is assumed to be 12335
+  id_sampling: True
+  # List with pairs for animation, None for random pairs
+  pairs_list: data/taichi256.csv
+  # Augmentation parameters see augmentation.py for all posible augmentations
+  augmentation_params:
+    flip_param:
+      horizontal_flip: True
+      time_flip: True
+    jitter_param:
+      brightness: 0.1
+      contrast: 0.1
+      saturation: 0.1
+      hue: 0.1
+
+# Defines model architecture
+model_params:
+  common_params:
+    # Number of keypoint
+    num_kp: 10
+    # Number of channels per image
+    num_channels: 3
+    # Using first or zero order model
+    estimate_jacobian: True
+  kp_detector_params:
+     # Softmax temperature for keypoint heatmaps
+     temperature: 0.1
+     # Number of features mutliplier
+     block_expansion: 32
+     # Maximum allowed number of features
+     max_features: 1024
+     # Number of block in Unet. Can be increased or decreased depending or resolution.
+     num_blocks: 5
+     # Keypioint is predicted on smaller images for better performance,
+     # scale_factor=0.25 means that 256x256 image will be resized to 64x64
+     scale_factor: 0.25
+  generator_params:
+    # Number of features mutliplier
+    block_expansion: 64
+    # Maximum allowed number of features
+    max_features: 512
+    # Number of downsampling blocks in Jonson architecture.
+    # Can be increased or decreased depending or resolution.
+    num_down_blocks: 2
+    # Number of ResBlocks  in Jonson architecture.
+    num_bottleneck_blocks: 6
+    # Use occlusion map or not
+    estimate_occlusion_map: True
+
+    dense_motion_params:
+      # Number of features mutliplier
+      block_expansion: 64
+      # Maximum allowed number of features
+      max_features: 1024
+      # Number of block in Unet. Can be increased or decreased depending or resolution.
+      num_blocks: 5
+      # Dense motion is predicted on smaller images for better performance,
+      # scale_factor=0.25 means that 256x256 image will be resized to 64x64
+      scale_factor: 0.25
+  discriminator_params:
+    # Discriminator can be multiscale, if you want 2 discriminator on original
+    # resolution and half of the original, specify scales: [1, 0.5]
+    scales: [1]
+    # Number of features mutliplier
+    block_expansion: 32
+    # Maximum allowed number of features
+    max_features: 512
+    # Number of blocks. Can be increased or decreased depending or resolution.
+    num_blocks: 4
+    use_kp: True
+
+# Parameters of training
+train_params:
+  # Number of training epochs 
+  num_epochs: 150
+  # For better i/o performance when number of videos is small number of epochs can be multiplied by this number.
+  # Thus effectivlly with num_repeats=100 each epoch is 100 times larger. 
+  num_repeats: 150
+  # Drop learning rate by 10 times after this epochs 
+  epoch_milestones: []
+  # Initial learing rate for all modules
+  lr_generator: 2.0e-4
+  lr_discriminator: 2.0e-4
+  lr_kp_detector: 0
+  batch_size: 27
+  # Scales for perceptual pyramide loss. If scales = [1, 0.5, 0.25, 0.125] and image resolution is 256x256,
+  # than the loss will be computer on resolutions 256x256, 128x128, 64x64, 32x32.
+  scales: [1, 0.5, 0.25, 0.125]
+  # Save checkpoint this frequently. If checkpoint_freq=50, checkpoint will be saved every 50 epochs.
+  checkpoint_freq: 50
+  # Parameters of transform for equivariance loss
+  transform_params:
+    # Sigma for affine part
+    sigma_affine: 0.05
+    # Sigma for deformation part
+    sigma_tps: 0.005
+    # Number of point in the deformation grid
+    points_tps: 5
+  loss_weights:
+    # Weight for LSGAN loss in generator
+    generator_gan: 1
+    # Weight for LSGAN loss in discriminator
+    discriminator_gan: 1
+    # Weights for feature matching loss, the number should be the same as number of blocks in discriminator.
+    feature_matching: [10, 10, 10, 10]
+    # Weights for perceptual loss.
+    perceptual: [10, 10, 10, 10, 10]
+    # Weights for value equivariance.
+    equivariance_value: 10
+    # Weights for jacobian equivariance.
+    equivariance_jacobian: 10
+
+# Parameters of reconstruction
+reconstruction_params:
+  # Maximum number of videos for reconstruction
+  num_videos: 1000
+  # Format for visualization, note that results will be also stored in staked .png.
+  format: '.mp4'
+
+# Parameters of animation
+animate_params:
+  # Maximum number of pairs for animation, the pairs will be either taken from pairs_list or random.
+  num_pairs: 50
+  # Format for visualization, note that results will be also stored in staked .png.
+  format: '.mp4'
+  # Normalization of diriving keypoints
+  normalization_params:
+    # Increase or decrease relative movement scale depending on the size of the object
+    adapt_movement_scale: False
+    # Apply only relative displacement of the keypoint
+    use_relative_movement: True
+    # Apply only relative change in jacobian
+    use_relative_jacobian: True
+
+# Visualization parameters
+visualizer_params:
+  # Draw keypoints of this size, increase or decrease depending on resolution
+  kp_size: 5
+  # Draw white border around images
+  draw_border: True
+  # Color map for keypoints
+  colormap: 'gist_rainbow'

config/vox-256.yaml

@@ -0,0 +1,83 @@
+dataset_params:
+  root_dir: data/vox-png
+  frame_shape: [256, 256, 3]
+  id_sampling: True
+  pairs_list: data/vox256.csv
+  augmentation_params:
+    flip_param:
+      horizontal_flip: True
+      time_flip: True
+    jitter_param:
+      brightness: 0.1
+      contrast: 0.1
+      saturation: 0.1
+      hue: 0.1
+
+
+model_params:
+  common_params:
+    num_kp: 10
+    num_channels: 3
+    estimate_jacobian: True
+  kp_detector_params:
+     temperature: 0.1
+     block_expansion: 32
+     max_features: 1024
+     scale_factor: 0.25
+     num_blocks: 5
+  generator_params:
+    block_expansion: 64
+    max_features: 512
+    num_down_blocks: 2
+    num_bottleneck_blocks: 6
+    estimate_occlusion_map: True
+    dense_motion_params:
+      block_expansion: 64
+      max_features: 1024
+      num_blocks: 5
+      scale_factor: 0.25
+  discriminator_params:
+    scales: [1]
+    block_expansion: 32
+    max_features: 512
+    num_blocks: 4
+    sn: True
+
+train_params:
+  num_epochs: 100
+  num_repeats: 75
+  epoch_milestones: [60, 90]
+  lr_generator: 2.0e-4
+  lr_discriminator: 2.0e-4
+  lr_kp_detector: 2.0e-4
+  batch_size: 40
+  scales: [1, 0.5, 0.25, 0.125]
+  checkpoint_freq: 50
+  transform_params:
+    sigma_affine: 0.05
+    sigma_tps: 0.005
+    points_tps: 5
+  loss_weights:
+    generator_gan: 0
+    discriminator_gan: 1
+    feature_matching: [10, 10, 10, 10]
+    perceptual: [10, 10, 10, 10, 10]
+    equivariance_value: 10
+    equivariance_jacobian: 10
+
+reconstruction_params:
+  num_videos: 1000
+  format: '.mp4'
+
+animate_params:
+  num_pairs: 50
+  format: '.mp4'
+  normalization_params:
+    adapt_movement_scale: False
+    use_relative_movement: True
+    use_relative_jacobian: True
+
+visualizer_params:
+  kp_size: 5
+  draw_border: True
+  colormap: 'gist_rainbow'

config/vox-adv-256.yaml

@@ -0,0 +1,84 @@
+dataset_params:
+  root_dir: data/vox-png
+  frame_shape: [256, 256, 3]
+  id_sampling: True
+  pairs_list: data/vox256.csv
+  augmentation_params:
+    flip_param:
+      horizontal_flip: True
+      time_flip: True
+    jitter_param:
+      brightness: 0.1
+      contrast: 0.1
+      saturation: 0.1
+      hue: 0.1
+
+
+model_params:
+  common_params:
+    num_kp: 10
+    num_channels: 3
+    estimate_jacobian: True
+  kp_detector_params:
+     temperature: 0.1
+     block_expansion: 32
+     max_features: 1024
+     scale_factor: 0.25
+     num_blocks: 5
+  generator_params:
+    block_expansion: 64
+    max_features: 512
+    num_down_blocks: 2
+    num_bottleneck_blocks: 6
+    estimate_occlusion_map: True
+    dense_motion_params:
+      block_expansion: 64
+      max_features: 1024
+      num_blocks: 5
+      scale_factor: 0.25
+  discriminator_params:
+    scales: [1]
+    block_expansion: 32
+    max_features: 512
+    num_blocks: 4
+    use_kp: True
+
+
+train_params:
+  num_epochs: 150
+  num_repeats: 75
+  epoch_milestones: []
+  lr_generator: 2.0e-4
+  lr_discriminator: 2.0e-4
+  lr_kp_detector: 2.0e-4
+  batch_size: 36
+  scales: [1, 0.5, 0.25, 0.125]
+  checkpoint_freq: 50
+  transform_params:
+    sigma_affine: 0.05
+    sigma_tps: 0.005
+    points_tps: 5
+  loss_weights:
+    generator_gan: 1
+    discriminator_gan: 1
+    feature_matching: [10, 10, 10, 10]
+    perceptual: [10, 10, 10, 10, 10]
+    equivariance_value: 10
+    equivariance_jacobian: 10
+
+reconstruction_params:
+  num_videos: 1000
+  format: '.mp4'
+
+animate_params:
+  num_pairs: 50
+  format: '.mp4'
+  normalization_params:
+    adapt_movement_scale: False
+    use_relative_movement: True
+    use_relative_jacobian: True
+
+visualizer_params:
+  kp_size: 5
+  draw_border: True
+  colormap: 'gist_rainbow'

data/bair256.csv

@@ -0,0 +1,51 @@
+distance,source,driving,frame
+0,000054.mp4,000048.mp4,0
+0,000050.mp4,000063.mp4,0
+0,000073.mp4,000007.mp4,0
+0,000021.mp4,000010.mp4,0
+0,000084.mp4,000046.mp4,0
+0,000031.mp4,000102.mp4,0
+0,000029.mp4,000111.mp4,0
+0,000090.mp4,000112.mp4,0
+0,000039.mp4,000010.mp4,0
+0,000008.mp4,000069.mp4,0
+0,000068.mp4,000076.mp4,0
+0,000051.mp4,000052.mp4,0
+0,000022.mp4,000098.mp4,0
+0,000096.mp4,000032.mp4,0
+0,000032.mp4,000099.mp4,0
+0,000006.mp4,000053.mp4,0
+0,000098.mp4,000020.mp4,0
+0,000029.mp4,000066.mp4,0
+0,000022.mp4,000007.mp4,0
+0,000027.mp4,000065.mp4,0
+0,000026.mp4,000059.mp4,0
+0,000015.mp4,000112.mp4,0
+0,000086.mp4,000123.mp4,0
+0,000103.mp4,000052.mp4,0
+0,000123.mp4,000103.mp4,0
+0,000051.mp4,000005.mp4,0
+0,000062.mp4,000125.mp4,0
+0,000126.mp4,000111.mp4,0
+0,000066.mp4,000090.mp4,0
+0,000075.mp4,000106.mp4,0
+0,000020.mp4,000010.mp4,0
+0,000076.mp4,000028.mp4,0
+0,000062.mp4,000002.mp4,0
+0,000095.mp4,000127.mp4,0
+0,000113.mp4,000072.mp4,0
+0,000027.mp4,000104.mp4,0
+0,000054.mp4,000124.mp4,0
+0,000019.mp4,000089.mp4,0
+0,000052.mp4,000072.mp4,0
+0,000108.mp4,000033.mp4,0
+0,000044.mp4,000118.mp4,0
+0,000029.mp4,000086.mp4,0
+0,000068.mp4,000066.mp4,0
+0,000014.mp4,000036.mp4,0
+0,000053.mp4,000071.mp4,0
+0,000022.mp4,000094.mp4,0
+0,000000.mp4,000121.mp4,0
+0,000071.mp4,000079.mp4,0
+0,000127.mp4,000005.mp4,0
+0,000085.mp4,000023.mp4,0

data/taichi-loading/README.md

@@ -0,0 +1,18 @@
+# TaiChi dataset
+
+The scripst for loading the TaiChi dataset. 
+
+We provide only the id of the corresponding video and the bounding box. Following script will download videos from youtube and crop them according to the provided bounding boxes.
+
+1) Load youtube-dl:
+```
+wget https://yt-dl.org/downloads/latest/youtube-dl -O youtube-dl
+chmod a+rx youtube-dl
+```
+
+2) Run script to download videos, there are 2 formats that can be used for storing videos one is .mp4 and another is folder with .png images. While .png images occupy significantly more space, the format is loss-less and have better i/o performance when training.
+
+```
+python load_videos.py --metadata taichi-metadata.csv --format .mp4 --out_folder taichi --workers 8
+```
+select number of workers based on number of cpu avaliable. Note .png format take aproximatly 80GB.

data/taichi-loading/load_videos.py

@@ -0,0 +1,113 @@
+import numpy as np
+import pandas as pd
+import imageio
+import os
+import subprocess
+from multiprocessing import Pool
+from itertools import cycle
+import warnings
+import glob
+import time
+from tqdm import tqdm
+from argparse import ArgumentParser
+from skimage import img_as_ubyte
+from skimage.transform import resize
+warnings.filterwarnings("ignore")
+
+DEVNULL = open(os.devnull, 'wb')
+
+
+def save(path, frames, format):
+    if format == '.mp4':
+        imageio.mimsave(path, frames)
+    elif format == '.png':
+        if os.path.exists(path):
+            print ("Warning: skiping video %s" % os.path.basename(path))
+            return
+        else:
+            os.makedirs(path)
+        for j, frame in enumerate(frames):
+            imageio.imsave(os.path.join(path, str(j).zfill(7) + '.png'), frames[j]) 
+    else:
+        print ("Unknown format %s" % format)
+        exit()
+
+
+def download(video_id, args):
+    video_path = os.path.join(args.video_folder, video_id + ".mp4")
+    subprocess.call([args.youtube, '-f', "''best/mp4''", '--write-auto-sub', '--write-sub',
+                     '--sub-lang', 'en', '--skip-unavailable-fragments',
+                     "https://www.youtube.com/watch?v=" + video_id, "--output",
+                     video_path], stdout=DEVNULL, stderr=DEVNULL)
+    return video_path
+
+
+def run(data):
+    video_id, args = data
+    if not os.path.exists(os.path.join(args.video_folder, video_id.split('#')[0] + '.mp4')):
+       download(video_id.split('#')[0], args)
+
+    if not os.path.exists(os.path.join(args.video_folder, video_id.split('#')[0] + '.mp4')):
+       print ('Can not load video %s, broken link' % video_id.split('#')[0])
+       return 
+    reader = imageio.get_reader(os.path.join(args.video_folder, video_id.split('#')[0] + '.mp4'))
+    fps = reader.get_meta_data()['fps']
+
+    df = pd.read_csv(args.metadata)
+    df = df[df['video_id'] == video_id]
+    
+    all_chunks_dict = [{'start': df['start'].iloc[j], 'end': df['end'].iloc[j],
+                        'bbox': list(map(int, df['bbox'].iloc[j].split('-'))), 'frames':[]} for j in range(df.shape[0])]
+    ref_fps = df['fps'].iloc[0]
+    ref_height = df['height'].iloc[0]
+    ref_width = df['width'].iloc[0]
+    partition = df['partition'].iloc[0]
+    try:
+        for i, frame in enumerate(reader):
+            for entry in all_chunks_dict:
+                if (i * ref_fps >= entry['start'] * fps) and (i * ref_fps < entry['end'] * fps):
+                    left, top, right, bot = entry['bbox']
+                    left = int(left / (ref_width / frame.shape[1]))
+                    top = int(top / (ref_height / frame.shape[0]))
+                    right = int(right / (ref_width / frame.shape[1]))
+                    bot = int(bot / (ref_height / frame.shape[0]))
+                    crop = frame[top:bot, left:right]
+                    if args.image_shape is not None:
+                       crop = img_as_ubyte(resize(crop, args.image_shape, anti_aliasing=True))
+                    entry['frames'].append(crop)
+    except imageio.core.format.CannotReadFrameError:
+        None
+
+    for entry in all_chunks_dict:
+        first_part = '#'.join(video_id.split('#')[::-1])
+        path = first_part + '#' + str(entry['start']).zfill(6) + '#' + str(entry['end']).zfill(6) + '.mp4'
+        save(os.path.join(args.out_folder, partition, path), entry['frames'], args.format)
+
+
+if __name__ == "__main__":
+    parser = ArgumentParser()
+    parser.add_argument("--video_folder", default='youtube-taichi', help='Path to youtube videos')
+    parser.add_argument("--metadata", default='taichi-metadata-new.csv', help='Path to metadata')
+    parser.add_argument("--out_folder", default='taichi-png', help='Path to output')
+    parser.add_argument("--format", default='.png', help='Storing format')
+    parser.add_argument("--workers", default=1, type=int, help='Number of workers')
+    parser.add_argument("--youtube", default='./youtube-dl', help='Path to youtube-dl')
+ 
+    parser.add_argument("--image_shape", default=(256, 256), type=lambda x: tuple(map(int, x.split(','))),
+                        help="Image shape, None for no resize")
+
+    args = parser.parse_args()
+    if not os.path.exists(args.video_folder):
+        os.makedirs(args.video_folder)
+    if not os.path.exists(args.out_folder):
+        os.makedirs(args.out_folder)
+    for partition in ['test', 'train']:
+        if not os.path.exists(os.path.join(args.out_folder, partition)):
+            os.makedirs(os.path.join(args.out_folder, partition))
+
+    df = pd.read_csv(args.metadata)
+    video_ids = set(df['video_id'])
+    pool = Pool(processes=args.workers)
+    args_list = cycle([args])
+    for chunks_data in tqdm(pool.imap_unordered(run, zip(video_ids, args_list))):
+        None  

data/taichi256.csv

@@ -0,0 +1,51 @@
+distance,source,driving,frame
+3.54437869822485,ab28GAufK8o#000261#000596.mp4,aDyyTMUBoLE#000164#000351.mp4,0
+2.8639053254437887,DMEaUoA8EPE#000028#000354.mp4,0Q914by5A98#010440#010764.mp4,0
+2.153846153846153,L82WHgYRq6I#000021#000479.mp4,0Q914by5A98#010440#010764.mp4,0
+2.8994082840236666,oNkBx4CZuEg#000000#001024.mp4,DMEaUoA8EPE#000028#000354.mp4,0
+3.3905325443786998,ab28GAufK8o#000261#000596.mp4,uEqWZ9S_-Lw#000089#000581.mp4,0
+3.266272189349112,0Q914by5A98#010440#010764.mp4,ab28GAufK8o#000261#000596.mp4,0
+2.7514792899408294,WlDYrq8K6nk#008186#008512.mp4,OiblkvkAHWM#014331#014459.mp4,0
+3.0177514792899407,oNkBx4CZuEg#001024#002048.mp4,aDyyTMUBoLE#000375#000518.mp4,0
+3.4792899408284064,aDyyTMUBoLE#000164#000351.mp4,w2awOCDRtrc#001729#002009.mp4,0
+2.769230769230769,oNkBx4CZuEg#000000#001024.mp4,L82WHgYRq6I#000021#000479.mp4,0
+3.8047337278106514,ab28GAufK8o#000261#000596.mp4,w2awOCDRtrc#001729#002009.mp4,0
+3.4260355029585763,w2awOCDRtrc#001729#002009.mp4,oNkBx4CZuEg#000000#001024.mp4,0
+3.313609467455621,DMEaUoA8EPE#000028#000354.mp4,WlDYrq8K6nk#005943#006135.mp4,0
+3.8402366863905333,oNkBx4CZuEg#001024#002048.mp4,ab28GAufK8o#000261#000596.mp4,0
+3.3254437869822504,aDyyTMUBoLE#000164#000351.mp4,oNkBx4CZuEg#000000#001024.mp4,0
+1.2485207100591724,0Q914by5A98#010440#010764.mp4,aDyyTMUBoLE#000164#000351.mp4,0
+3.804733727810652,OiblkvkAHWM#006251#006533.mp4,aDyyTMUBoLE#000375#000518.mp4,0
+3.662721893491124,uEqWZ9S_-Lw#000089#000581.mp4,DMEaUoA8EPE#000028#000354.mp4,0
+3.230769230769233,A3ZmT97hAWU#000095#000678.mp4,ab28GAufK8o#000261#000596.mp4,0
+3.3668639053254434,w81Tr0Dp1K8#015329#015485.mp4,WlDYrq8K6nk#008186#008512.mp4,0
+3.313609467455621,WlDYrq8K6nk#005943#006135.mp4,DMEaUoA8EPE#000028#000354.mp4,0
+2.7514792899408294,OiblkvkAHWM#014331#014459.mp4,WlDYrq8K6nk#008186#008512.mp4,0
+1.964497041420118,L82WHgYRq6I#000021#000479.mp4,DMEaUoA8EPE#000028#000354.mp4,0
+3.78698224852071,FBuF0xOal9M#046824#047542.mp4,lCb5w6n8kPs#011879#012014.mp4,0
+3.92307692307692,ab28GAufK8o#000261#000596.mp4,L82WHgYRq6I#000021#000479.mp4,0
+3.8402366863905333,ab28GAufK8o#000261#000596.mp4,oNkBx4CZuEg#001024#002048.mp4,0
+3.828402366863905,ab28GAufK8o#000261#000596.mp4,OiblkvkAHWM#006251#006533.mp4,0
+2.041420118343196,L82WHgYRq6I#000021#000479.mp4,aDyyTMUBoLE#000164#000351.mp4,0
+3.2485207100591724,0Q914by5A98#010440#010764.mp4,w2awOCDRtrc#001729#002009.mp4,0
+3.2485207100591746,oNkBx4CZuEg#000000#001024.mp4,0Q914by5A98#010440#010764.mp4,0
+1.964497041420118,DMEaUoA8EPE#000028#000354.mp4,L82WHgYRq6I#000021#000479.mp4,0
+3.5266272189349115,kgvcI9oe3NI#001578#001763.mp4,lCb5w6n8kPs#004451#004631.mp4,0
+3.005917159763317,A3ZmT97hAWU#000095#000678.mp4,0Q914by5A98#010440#010764.mp4,0
+3.230769230769233,ab28GAufK8o#000261#000596.mp4,A3ZmT97hAWU#000095#000678.mp4,0
+3.5266272189349115,lCb5w6n8kPs#004451#004631.mp4,kgvcI9oe3NI#001578#001763.mp4,0
+2.769230769230769,L82WHgYRq6I#000021#000479.mp4,oNkBx4CZuEg#000000#001024.mp4,0
+3.165680473372782,WlDYrq8K6nk#005943#006135.mp4,w81Tr0Dp1K8#001375#001516.mp4,0
+2.8994082840236666,DMEaUoA8EPE#000028#000354.mp4,oNkBx4CZuEg#000000#001024.mp4,0
+2.4556213017751523,0Q914by5A98#010440#010764.mp4,mndSqTrxpts#000000#000175.mp4,0
+2.201183431952659,A3ZmT97hAWU#000095#000678.mp4,VMSqvTE90hk#007168#007312.mp4,0
+3.8047337278106514,w2awOCDRtrc#001729#002009.mp4,ab28GAufK8o#000261#000596.mp4,0
+3.769230769230769,uEqWZ9S_-Lw#000089#000581.mp4,0Q914by5A98#010440#010764.mp4,0
+3.6568047337278102,A3ZmT97hAWU#000095#000678.mp4,aDyyTMUBoLE#000164#000351.mp4,0
+3.7869822485207107,uEqWZ9S_-Lw#000089#000581.mp4,L82WHgYRq6I#000021#000479.mp4,0
+3.78698224852071,lCb5w6n8kPs#011879#012014.mp4,FBuF0xOal9M#046824#047542.mp4,0
+3.591715976331361,nAQEOC1Z10M#020177#020600.mp4,w81Tr0Dp1K8#004036#004218.mp4,0
+3.8757396449704156,uEqWZ9S_-Lw#000089#000581.mp4,aDyyTMUBoLE#000164#000351.mp4,0
+2.45562130177515,aDyyTMUBoLE#000164#000351.mp4,DMEaUoA8EPE#000028#000354.mp4,0
+3.5502958579881647,uEqWZ9S_-Lw#000089#000581.mp4,OiblkvkAHWM#006251#006533.mp4,0
+3.7928994082840224,aDyyTMUBoLE#000375#000518.mp4,ab28GAufK8o#000261#000596.mp4,0

modules/dense_motion.py

@@ -0,0 +1,113 @@
+from torch import nn
+import torch.nn.functional as F
+import torch
+from modules.util import Hourglass, AntiAliasInterpolation2d, make_coordinate_grid, kp2gaussian
+
+
+class DenseMotionNetwork(nn.Module):
+    """
+    Module that predicting a dense motion from sparse motion representation given by kp_source and kp_driving
+    """
+
+    def __init__(self, block_expansion, num_blocks, max_features, num_kp, num_channels, estimate_occlusion_map=False,
+                 scale_factor=1, kp_variance=0.01):
+        super(DenseMotionNetwork, self).__init__()
+        self.hourglass = Hourglass(block_expansion=block_expansion, in_features=(num_kp + 1) * (num_channels + 1),
+                                   max_features=max_features, num_blocks=num_blocks)
+
+        self.mask = nn.Conv2d(self.hourglass.out_filters, num_kp + 1, kernel_size=(7, 7), padding=(3, 3))
+
+        if estimate_occlusion_map:
+            self.occlusion = nn.Conv2d(self.hourglass.out_filters, 1, kernel_size=(7, 7), padding=(3, 3))
+        else:
+            self.occlusion = None
+
+        self.num_kp = num_kp
+        self.scale_factor = scale_factor
+        self.kp_variance = kp_variance
+
+        if self.scale_factor != 1:
+            self.down = AntiAliasInterpolation2d(num_channels, self.scale_factor)
+
+    def create_heatmap_representations(self, source_image, kp_driving, kp_source):
+        """
+        Eq 6. in the paper H_k(z)
+        """
+        spatial_size = source_image.shape[2:]
+        gaussian_driving = kp2gaussian(kp_driving, spatial_size=spatial_size, kp_variance=self.kp_variance)
+        gaussian_source = kp2gaussian(kp_source, spatial_size=spatial_size, kp_variance=self.kp_variance)
+        heatmap = gaussian_driving - gaussian_source
+
+        #adding background feature
+        zeros = torch.zeros(heatmap.shape[0], 1, spatial_size[0], spatial_size[1]).type(heatmap.type())
+        heatmap = torch.cat([zeros, heatmap], dim=1)
+        heatmap = heatmap.unsqueeze(2)
+        return heatmap
+
+    def create_sparse_motions(self, source_image, kp_driving, kp_source):
+        """
+        Eq 4. in the paper T_{s<-d}(z)
+        """
+        bs, _, h, w = source_image.shape
+        identity_grid = make_coordinate_grid((h, w), type=kp_source['value'].type())
+        identity_grid = identity_grid.view(1, 1, h, w, 2)
+        coordinate_grid = identity_grid - kp_driving['value'].view(bs, self.num_kp, 1, 1, 2)
+        if 'jacobian' in kp_driving:
+            jacobian = torch.matmul(kp_source['jacobian'], torch.inverse(kp_driving['jacobian']))
+            jacobian = jacobian.unsqueeze(-3).unsqueeze(-3)
+            jacobian = jacobian.repeat(1, 1, h, w, 1, 1)
+            coordinate_grid = torch.matmul(jacobian, coordinate_grid.unsqueeze(-1))
+            coordinate_grid = coordinate_grid.squeeze(-1)
+
+        driving_to_source = coordinate_grid + kp_source['value'].view(bs, self.num_kp, 1, 1, 2)
+
+        #adding background feature
+        identity_grid = identity_grid.repeat(bs, 1, 1, 1, 1)
+        sparse_motions = torch.cat([identity_grid, driving_to_source], dim=1)
+        return sparse_motions
+
+    def create_deformed_source_image(self, source_image, sparse_motions):
+        """
+        Eq 7. in the paper \hat{T}_{s<-d}(z)
+        """
+        bs, _, h, w = source_image.shape
+        source_repeat = source_image.unsqueeze(1).unsqueeze(1).repeat(1, self.num_kp + 1, 1, 1, 1, 1)
+        source_repeat = source_repeat.view(bs * (self.num_kp + 1), -1, h, w)
+        sparse_motions = sparse_motions.view((bs * (self.num_kp + 1), h, w, -1))
+        sparse_deformed = F.grid_sample(source_repeat, sparse_motions)
+        sparse_deformed = sparse_deformed.view((bs, self.num_kp + 1, -1, h, w))
+        return sparse_deformed
+
+    def forward(self, source_image, kp_driving, kp_source):
+        if self.scale_factor != 1:
+            source_image = self.down(source_image)
+
+        bs, _, h, w = source_image.shape
+
+        out_dict = dict()
+        heatmap_representation = self.create_heatmap_representations(source_image, kp_driving, kp_source)
+        sparse_motion = self.create_sparse_motions(source_image, kp_driving, kp_source)
+        deformed_source = self.create_deformed_source_image(source_image, sparse_motion)
+        out_dict['sparse_deformed'] = deformed_source
+
+        input = torch.cat([heatmap_representation, deformed_source], dim=2)
+        input = input.view(bs, -1, h, w)
+
+        prediction = self.hourglass(input)
+
+        mask = self.mask(prediction)
+        mask = F.softmax(mask, dim=1)
+        out_dict['mask'] = mask
+        mask = mask.unsqueeze(2)
+        sparse_motion = sparse_motion.permute(0, 1, 4, 2, 3)
+        deformation = (sparse_motion * mask).sum(dim=1)
+        deformation = deformation.permute(0, 2, 3, 1)
+
+        out_dict['deformation'] = deformation
+
+        # Sec. 3.2 in the paper
+        if self.occlusion:
+            occlusion_map = torch.sigmoid(self.occlusion(prediction))
+            out_dict['occlusion_map'] = occlusion_map
+
+        return out_dict

modules/discriminator.py

@@ -0,0 +1,95 @@
+from torch import nn
+import torch.nn.functional as F
+from modules.util import kp2gaussian
+import torch
+
+
+class DownBlock2d(nn.Module):
+    """
+    Simple block for processing video (encoder).
+    """
+
+    def __init__(self, in_features, out_features, norm=False, kernel_size=4, pool=False, sn=False):
+        super(DownBlock2d, self).__init__()
+        self.conv = nn.Conv2d(in_channels=in_features, out_channels=out_features, kernel_size=kernel_size)
+
+        if sn:
+            self.conv = nn.utils.spectral_norm(self.conv)
+
+        if norm:
+            self.norm = nn.InstanceNorm2d(out_features, affine=True)
+        else:
+            self.norm = None
+        self.pool = pool
+
+    def forward(self, x):
+        out = x
+        out = self.conv(out)
+        if self.norm:
+            out = self.norm(out)
+        out = F.leaky_relu(out, 0.2)
+        if self.pool:
+            out = F.avg_pool2d(out, (2, 2))
+        return out
+
+
+class Discriminator(nn.Module):
+    """
+    Discriminator similar to Pix2Pix
+    """
+
+    def __init__(self, num_channels=3, block_expansion=64, num_blocks=4, max_features=512,
+                 sn=False, use_kp=False, num_kp=10, kp_variance=0.01, **kwargs):
+        super(Discriminator, self).__init__()
+
+        down_blocks = []
+        for i in range(num_blocks):
+            down_blocks.append(
+                DownBlock2d(num_channels + num_kp * use_kp if i == 0 else min(max_features, block_expansion * (2 ** i)),
+                            min(max_features, block_expansion * (2 ** (i + 1))),
+                            norm=(i != 0), kernel_size=4, pool=(i != num_blocks - 1), sn=sn))
+
+        self.down_blocks = nn.ModuleList(down_blocks)
+        self.conv = nn.Conv2d(self.down_blocks[-1].conv.out_channels, out_channels=1, kernel_size=1)
+        if sn:
+            self.conv = nn.utils.spectral_norm(self.conv)
+        self.use_kp = use_kp
+        self.kp_variance = kp_variance
+
+    def forward(self, x, kp=None):
+        feature_maps = []
+        out = x
+        if self.use_kp:
+            heatmap = kp2gaussian(kp, x.shape[2:], self.kp_variance)
+            out = torch.cat([out, heatmap], dim=1)
+
+        for down_block in self.down_blocks:
+            feature_maps.append(down_block(out))
+            out = feature_maps[-1]
+        prediction_map = self.conv(out)
+
+        return feature_maps, prediction_map
+
+
+class MultiScaleDiscriminator(nn.Module):
+    """
+    Multi-scale (scale) discriminator
+    """
+
+    def __init__(self, scales=(), **kwargs):
+        super(MultiScaleDiscriminator, self).__init__()
+        self.scales = scales
+        discs = {}
+        for scale in scales:
+            discs[str(scale).replace('.', '-')] = Discriminator(**kwargs)
+        self.discs = nn.ModuleDict(discs)
+
+    def forward(self, x, kp=None):
+        out_dict = {}
+        for scale, disc in self.discs.items():
+            scale = str(scale).replace('-', '.')
+            key = 'prediction_' + scale
+            feature_maps, prediction_map = disc(x[key], kp)
+            out_dict['feature_maps_' + scale] = feature_maps
+            out_dict['prediction_map_' + scale] = prediction_map
+        return out_dict

modules/generator.py

@@ -0,0 +1,97 @@
+import torch
+from torch import nn
+import torch.nn.functional as F
+from modules.util import ResBlock2d, SameBlock2d, UpBlock2d, DownBlock2d
+from modules.dense_motion import DenseMotionNetwork
+
+
+class OcclusionAwareGenerator(nn.Module):
+    """
+    Generator that given source image and and keypoints try to transform image according to movement trajectories
+    induced by keypoints. Generator follows Johnson architecture.
+    """
+
+    def __init__(self, num_channels, num_kp, block_expansion, max_features, num_down_blocks,
+                 num_bottleneck_blocks, estimate_occlusion_map=False, dense_motion_params=None, estimate_jacobian=False):
+        super(OcclusionAwareGenerator, self).__init__()
+
+        if dense_motion_params is not None:
+            self.dense_motion_network = DenseMotionNetwork(num_kp=num_kp, num_channels=num_channels,
+                                                           estimate_occlusion_map=estimate_occlusion_map,
+                                                           **dense_motion_params)
+        else:
+            self.dense_motion_network = None
+
+        self.first = SameBlock2d(num_channels, block_expansion, kernel_size=(7, 7), padding=(3, 3))
+
+        down_blocks = []
+        for i in range(num_down_blocks):
+            in_features = min(max_features, block_expansion * (2 ** i))
+            out_features = min(max_features, block_expansion * (2 ** (i + 1)))
+            down_blocks.append(DownBlock2d(in_features, out_features, kernel_size=(3, 3), padding=(1, 1)))
+        self.down_blocks = nn.ModuleList(down_blocks)
+
+        up_blocks = []
+        for i in range(num_down_blocks):
+            in_features = min(max_features, block_expansion * (2 ** (num_down_blocks - i)))
+            out_features = min(max_features, block_expansion * (2 ** (num_down_blocks - i - 1)))
+            up_blocks.append(UpBlock2d(in_features, out_features, kernel_size=(3, 3), padding=(1, 1)))
+        self.up_blocks = nn.ModuleList(up_blocks)
+
+        self.bottleneck = torch.nn.Sequential()
+        in_features = min(max_features, block_expansion * (2 ** num_down_blocks))
+        for i in range(num_bottleneck_blocks):
+            self.bottleneck.add_module('r' + str(i), ResBlock2d(in_features, kernel_size=(3, 3), padding=(1, 1)))
+
+        self.final = nn.Conv2d(block_expansion, num_channels, kernel_size=(7, 7), padding=(3, 3))
+        self.estimate_occlusion_map = estimate_occlusion_map
+        self.num_channels = num_channels
+
+    def deform_input(self, inp, deformation):
+        _, h_old, w_old, _ = deformation.shape
+        _, _, h, w = inp.shape
+        if h_old != h or w_old != w:
+            deformation = deformation.permute(0, 3, 1, 2)
+            deformation = F.interpolate(deformation, size=(h, w), mode='bilinear')
+            deformation = deformation.permute(0, 2, 3, 1)
+        return F.grid_sample(inp, deformation)
+
+    def forward(self, source_image, kp_driving, kp_source):
+        # Encoding (downsampling) part
+        out = self.first(source_image)
+        for i in range(len(self.down_blocks)):
+            out = self.down_blocks[i](out)
+
+        # Transforming feature representation according to deformation and occlusion
+        output_dict = {}
+        if self.dense_motion_network is not None:
+            dense_motion = self.dense_motion_network(source_image=source_image, kp_driving=kp_driving,
+                                                     kp_source=kp_source)
+            output_dict['mask'] = dense_motion['mask']
+            output_dict['sparse_deformed'] = dense_motion['sparse_deformed']
+
+            if 'occlusion_map' in dense_motion:
+                occlusion_map = dense_motion['occlusion_map']
+                output_dict['occlusion_map'] = occlusion_map
+            else:
+                occlusion_map = None
+            deformation = dense_motion['deformation']
+            out = self.deform_input(out, deformation)
+
+            if occlusion_map is not None:
+                if out.shape[2] != occlusion_map.shape[2] or out.shape[3] != occlusion_map.shape[3]:
+                    occlusion_map = F.interpolate(occlusion_map, size=out.shape[2:], mode='bilinear')
+                out = out * occlusion_map
+
+            output_dict["deformed"] = self.deform_input(source_image, deformation)
+
+        # Decoding part
+        out = self.bottleneck(out)
+        for i in range(len(self.up_blocks)):
+            out = self.up_blocks[i](out)
+        out = self.final(out)
+        out = F.sigmoid(out)
+
+        output_dict["prediction"] = out
+
+        return output_dict

modules/keypoint_detector.py

@@ -0,0 +1,75 @@
+from torch import nn
+import torch
+import torch.nn.functional as F
+from modules.util import Hourglass, make_coordinate_grid, AntiAliasInterpolation2d
+
+
+class KPDetector(nn.Module):
+    """
+    Detecting a keypoints. Return keypoint position and jacobian near each keypoint.
+    """
+
+    def __init__(self, block_expansion, num_kp, num_channels, max_features,
+                 num_blocks, temperature, estimate_jacobian=False, scale_factor=1,
+                 single_jacobian_map=False, pad=0):
+        super(KPDetector, self).__init__()
+
+        self.predictor = Hourglass(block_expansion, in_features=num_channels,
+                                   max_features=max_features, num_blocks=num_blocks)
+
+        self.kp = nn.Conv2d(in_channels=self.predictor.out_filters, out_channels=num_kp, kernel_size=(7, 7),
+                            padding=pad)
+
+        if estimate_jacobian:
+            self.num_jacobian_maps = 1 if single_jacobian_map else num_kp
+            self.jacobian = nn.Conv2d(in_channels=self.predictor.out_filters,
+                                      out_channels=4 * self.num_jacobian_maps, kernel_size=(7, 7), padding=pad)
+            self.jacobian.weight.data.zero_()
+            self.jacobian.bias.data.copy_(torch.tensor([1, 0, 0, 1] * self.num_jacobian_maps, dtype=torch.float))
+        else:
+            self.jacobian = None
+
+        self.temperature = temperature
+        self.scale_factor = scale_factor
+        if self.scale_factor != 1:
+            self.down = AntiAliasInterpolation2d(num_channels, self.scale_factor)
+
+    def gaussian2kp(self, heatmap):
+        """
+        Extract the mean and from a heatmap
+        """
+        shape = heatmap.shape
+        heatmap = heatmap.unsqueeze(-1)
+        grid = make_coordinate_grid(shape[2:], heatmap.type()).unsqueeze_(0).unsqueeze_(0)
+        value = (heatmap * grid).sum(dim=(2, 3))
+        kp = {'value': value}
+
+        return kp
+
+    def forward(self, x):
+        if self.scale_factor != 1:
+            x = self.down(x)
+
+        feature_map = self.predictor(x)
+        prediction = self.kp(feature_map)
+
+        final_shape = prediction.shape
+        heatmap = prediction.view(final_shape[0], final_shape[1], -1)
+        heatmap = F.softmax(heatmap / self.temperature, dim=2)
+        heatmap = heatmap.view(*final_shape)
+
+        out = self.gaussian2kp(heatmap)
+
+        if self.jacobian is not None:
+            jacobian_map = self.jacobian(feature_map)
+            jacobian_map = jacobian_map.reshape(final_shape[0], self.num_jacobian_maps, 4, final_shape[2],
+                                                final_shape[3])
+            heatmap = heatmap.unsqueeze(2)
+
+            jacobian = heatmap * jacobian_map
+            jacobian = jacobian.view(final_shape[0], final_shape[1], 4, -1)
+            jacobian = jacobian.sum(dim=-1)
+            jacobian = jacobian.view(jacobian.shape[0], jacobian.shape[1], 2, 2)
+            out['jacobian'] = jacobian
+
+        return out

modules/model.py

@@ -0,0 +1,259 @@
+from torch import nn
+import torch
+import torch.nn.functional as F
+from modules.util import AntiAliasInterpolation2d, make_coordinate_grid
+from torchvision import models
+import numpy as np
+from torch.autograd import grad
+
+
+class Vgg19(torch.nn.Module):
+    """
+    Vgg19 network for perceptual loss. See Sec 3.3.
+    """
+    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. See Sec 3.3
+    """
+    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. See Sec 3.3
+    """
+    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((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(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()}
+
+
+class GeneratorFullModel(torch.nn.Module):
+    """
+    Merge all generator related updates into single model for better multi-gpu usage
+    """
+
+    def __init__(self, kp_extractor, generator, discriminator, train_params):
+        super(GeneratorFullModel, self).__init__()
+        self.kp_extractor = kp_extractor
+        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.num_channels)
+        if torch.cuda.is_available():
+            self.pyramid = self.pyramid.cuda()
+
+        self.loss_weights = train_params['loss_weights']
+
+        if sum(self.loss_weights['perceptual']) != 0:
+            self.vgg = Vgg19()
+            if torch.cuda.is_available():
+                self.vgg = self.vgg.cuda()
+
+    def forward(self, x):
+        kp_source = self.kp_extractor(x['source'])
+        kp_driving = self.kp_extractor(x['driving'])
+
+        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, kp=detach_kp(kp_driving))
+            discriminator_maps_real = self.discriminator(pyramide_real, kp=detach_kp(kp_driving))
+            value_total = 0
+            for scale in self.disc_scales:
+                key = 'prediction_map_%s' % scale
+                value = ((1 - discriminator_maps_generated[key]) ** 2).mean()
+                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_kp = self.kp_extractor(transformed_frame)
+
+            generated['transformed_frame'] = transformed_frame
+            generated['transformed_kp'] = transformed_kp
+
+            ## Value loss part
+            if self.loss_weights['equivariance_value'] != 0:
+                value = torch.abs(kp_driving['value'] - transform.warp_coordinates(transformed_kp['value'])).mean()
+                loss_values['equivariance_value'] = self.loss_weights['equivariance_value'] * value
+
+            ## jacobian loss part
+            if self.loss_weights['equivariance_jacobian'] != 0:
+                jacobian_transformed = torch.matmul(transform.jacobian(transformed_kp['value']),
+                                                    transformed_kp['jacobian'])
+
+                normed_driving = torch.inverse(kp_driving['jacobian'])
+                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
+
+        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.num_channels)
+        if torch.cuda.is_available():
+            self.pyramid = self.pyramid.cuda()
+
+        self.loss_weights = train_params['loss_weights']
+
+    def forward(self, x, generated):
+        pyramide_real = self.pyramid(x['driving'])
+        pyramide_generated = self.pyramid(generated['prediction'].detach())
+
+        kp_driving = generated['kp_driving']
+        discriminator_maps_generated = self.discriminator(pyramide_generated, kp=detach_kp(kp_driving))
+        discriminator_maps_real = self.discriminator(pyramide_real, kp=detach_kp(kp_driving))
+
+        loss_values = {}
+        value_total = 0
+        for scale in self.scales:
+            key = 'prediction_map_%s' % scale
+            value = (1 - discriminator_maps_real[key]) ** 2 + discriminator_maps_generated[key] ** 2
+            value_total += self.loss_weights['discriminator_gan'] * value.mean()
+        loss_values['disc_gan'] = value_total
+
+        return loss_values

modules/util.py

@@ -0,0 +1,245 @@
+from torch import nn
+
+import torch.nn.functional as F
+import torch
+
+from sync_batchnorm import SynchronizedBatchNorm2d as BatchNorm2d
+
+
+def kp2gaussian(kp, spatial_size, kp_variance):
+    """
+    Transform a keypoint into gaussian like representation
+    """
+    mean = kp['value']
+
+    coordinate_grid = make_coordinate_grid(spatial_size, mean.type())
+    number_of_leading_dimensions = len(mean.shape) - 1
+    shape = (1,) * number_of_leading_dimensions + coordinate_grid.shape
+    coordinate_grid = coordinate_grid.view(*shape)
+    repeats = mean.shape[:number_of_leading_dimensions] + (1, 1, 1)
+    coordinate_grid = coordinate_grid.repeat(*repeats)
+
+    # Preprocess kp shape
+    shape = mean.shape[:number_of_leading_dimensions] + (1, 1, 2)
+    mean = mean.view(*shape)
+
+    mean_sub = (coordinate_grid - mean)
+
+    out = torch.exp(-0.5 * (mean_sub ** 2).sum(-1) / kp_variance)
+
+    return out
+
+
+def make_coordinate_grid(spatial_size, type):
+    """
+    Create a meshgrid [-1,1] x [-1,1] of given spatial_size.
+    """
+    h, w = spatial_size
+    x = torch.arange(w).type(type)
+    y = torch.arange(h).type(type)
+
+    x = (2 * (x / (w - 1)) - 1)
+    y = (2 * (y / (h - 1)) - 1)
+
+    yy = y.view(-1, 1).repeat(1, w)
+    xx = x.view(1, -1).repeat(h, 1)
+
+    meshed = torch.cat([xx.unsqueeze_(2), yy.unsqueeze_(2)], 2)
+
+    return meshed
+
+
+class ResBlock2d(nn.Module):
+    """
+    Res block, preserve spatial resolution.
+    """
+
+    def __init__(self, in_features, kernel_size, padding):
+        super(ResBlock2d, self).__init__()
+        self.conv1 = nn.Conv2d(in_channels=in_features, out_channels=in_features, kernel_size=kernel_size,
+                               padding=padding)
+        self.conv2 = nn.Conv2d(in_channels=in_features, out_channels=in_features, kernel_size=kernel_size,
+                               padding=padding)
+        self.norm1 = BatchNorm2d(in_features, affine=True)
+        self.norm2 = BatchNorm2d(in_features, affine=True)
+
+    def forward(self, x):
+        out = self.norm1(x)
+        out = F.relu(out)
+        out = self.conv1(out)
+        out = self.norm2(out)
+        out = F.relu(out)
+        out = self.conv2(out)
+        out += x
+        return out
+
+
+class UpBlock2d(nn.Module):
+    """
+    Upsampling block for use in decoder.
+    """
+
+    def __init__(self, in_features, out_features, kernel_size=3, padding=1, groups=1):
+        super(UpBlock2d, self).__init__()
+
+        self.conv = nn.Conv2d(in_channels=in_features, out_channels=out_features, kernel_size=kernel_size,
+                              padding=padding, groups=groups)
+        self.norm = BatchNorm2d(out_features, affine=True)
+
+    def forward(self, x):
+        out = F.interpolate(x, scale_factor=2)
+        out = self.conv(out)
+        out = self.norm(out)
+        out = F.relu(out)
+        return out
+
+
+class DownBlock2d(nn.Module):
+    """
+    Downsampling block for use in encoder.
+    """
+
+    def __init__(self, in_features, out_features, kernel_size=3, padding=1, groups=1):
+        super(DownBlock2d, self).__init__()
+        self.conv = nn.Conv2d(in_channels=in_features, out_channels=out_features, kernel_size=kernel_size,
+                              padding=padding, groups=groups)
+        self.norm = BatchNorm2d(out_features, affine=True)
+        self.pool = nn.AvgPool2d(kernel_size=(2, 2))
+
+    def forward(self, x):
+        out = self.conv(x)
+        out = self.norm(out)
+        out = F.relu(out)
+        out = self.pool(out)
+        return out
+
+
+class SameBlock2d(nn.Module):
+    """
+    Simple block, preserve spatial resolution.
+    """
+
+    def __init__(self, in_features, out_features, groups=1, kernel_size=3, padding=1):
+        super(SameBlock2d, self).__init__()
+        self.conv = nn.Conv2d(in_channels=in_features, out_channels=out_features,
+                              kernel_size=kernel_size, padding=padding, groups=groups)
+        self.norm = BatchNorm2d(out_features, affine=True)
+
+    def forward(self, x):
+        out = self.conv(x)
+        out = self.norm(out)
+        out = F.relu(out)
+        return out
+
+
+class Encoder(nn.Module):
+    """
+    Hourglass Encoder
+    """
+
+    def __init__(self, block_expansion, in_features, num_blocks=3, max_features=256):
+        super(Encoder, self).__init__()
+
+        down_blocks = []
+        for i in range(num_blocks):
+            down_blocks.append(DownBlock2d(in_features if i == 0 else min(max_features, block_expansion * (2 ** i)),
+                                           min(max_features, block_expansion * (2 ** (i + 1))),
+                                           kernel_size=3, padding=1))
+        self.down_blocks = nn.ModuleList(down_blocks)
+
+    def forward(self, x):
+        outs = [x]
+        for down_block in self.down_blocks:
+            outs.append(down_block(outs[-1]))
+        return outs
+
+
+class Decoder(nn.Module):
+    """
+    Hourglass Decoder
+    """
+
+    def __init__(self, block_expansion, in_features, num_blocks=3, max_features=256):
+        super(Decoder, self).__init__()
+
+        up_blocks = []
+
+        for i in range(num_blocks)[::-1]:
+            in_filters = (1 if i == num_blocks - 1 else 2) * min(max_features, block_expansion * (2 ** (i + 1)))
+            out_filters = min(max_features, block_expansion * (2 ** i))
+            up_blocks.append(UpBlock2d(in_filters, out_filters, kernel_size=3, padding=1))
+
+        self.up_blocks = nn.ModuleList(up_blocks)
+        self.out_filters = block_expansion + in_features
+
+    def forward(self, x):
+        out = x.pop()
+        for up_block in self.up_blocks:
+            out = up_block(out)
+            skip = x.pop()
+            out = torch.cat([out, skip], dim=1)
+        return out
+
+
+class Hourglass(nn.Module):
+    """
+    Hourglass architecture.
+    """
+
+    def __init__(self, block_expansion, in_features, num_blocks=3, max_features=256):
+        super(Hourglass, self).__init__()
+        self.encoder = Encoder(block_expansion, in_features, num_blocks, max_features)
+        self.decoder = Decoder(block_expansion, in_features, num_blocks, max_features)
+        self.out_filters = self.decoder.out_filters
+
+    def forward(self, x):
+        return self.decoder(self.encoder(x))
+
+
+class AntiAliasInterpolation2d(nn.Module):
+    """
+    Band-limited downsampling, for better preservation of the input signal.
+    """
+    def __init__(self, channels, scale):
+        super(AntiAliasInterpolation2d, self).__init__()
+        sigma = (1 / scale - 1) / 2
+        kernel_size = 2 * round(sigma * 4) + 1
+        self.ka = kernel_size // 2
+        self.kb = self.ka - 1 if kernel_size % 2 == 0 else self.ka
+
+        kernel_size = [kernel_size, kernel_size]
+        sigma = [sigma, sigma]
+        # The gaussian kernel is the product of the
+        # gaussian function of each dimension.
+        kernel = 1
+        meshgrids = torch.meshgrid(
+            [
+                torch.arange(size, dtype=torch.float32)
+                for size in kernel_size
+                ]
+        )
+        for size, std, mgrid in zip(kernel_size, sigma, meshgrids):
+            mean = (size - 1) / 2
+            kernel *= torch.exp(-(mgrid - mean) ** 2 / (2 * std ** 2))
+
+        # Make sure sum of values in gaussian kernel equals 1.
+        kernel = kernel / torch.sum(kernel)
+        # Reshape to depthwise convolutional weight
+        kernel = kernel.view(1, 1, *kernel.size())
+        kernel = kernel.repeat(channels, *[1] * (kernel.dim() - 1))
+
+        self.register_buffer('weight', kernel)
+        self.groups = channels
+        self.scale = scale
+        inv_scale = 1 / scale
+        self.int_inv_scale = int(inv_scale)
+
+    def forward(self, input):
+        if self.scale == 1.0:
+            return input
+
+        out = F.pad(input, (self.ka, self.kb, self.ka, self.kb))
+        out = F.conv2d(out, weight=self.weight, groups=self.groups)
+        out = out[:, :, ::self.int_inv_scale, ::self.int_inv_scale]
+
+        return out

share/doc/networkx-3.0/LICENSE.txt

@@ -0,0 +1,37 @@
+NetworkX is distributed with the 3-clause BSD license.
+
+::
+
+   Copyright (C) 2004-2023, NetworkX Developers
+   Aric Hagberg <hagberg@lanl.gov>
+   Dan Schult <dschult@colgate.edu>
+   Pieter Swart <swart@lanl.gov>
+   All rights reserved.
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+     * Redistributions of source code must retain the above copyright
+       notice, this list of conditions and the following disclaimer.
+
+     * Redistributions in binary form must reproduce the above
+       copyright notice, this list of conditions and the following
+       disclaimer in the documentation and/or other materials provided
+       with the distribution.
+
+     * Neither the name of the NetworkX Developers nor the names of its
+       contributors may be used to endorse or promote products derived
+       from this software without specific prior written permission.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

share/doc/networkx-3.0/examples/3d_drawing/README.txt

@@ -0,0 +1,2 @@
+3D Drawing
+----------

share/doc/networkx-3.0/examples/3d_drawing/mayavi2_spring.py

@@ -0,0 +1,43 @@
+"""
+=======
+Mayavi2
+=======
+
+"""
+
+import networkx as nx
+import numpy as np
+from mayavi import mlab
+
+# some graphs to try
+# H=nx.krackhardt_kite_graph()
+# H=nx.Graph();H.add_edge('a','b');H.add_edge('a','c');H.add_edge('a','d')
+# H=nx.grid_2d_graph(4,5)
+H = nx.cycle_graph(20)
+
+# reorder nodes from 0,len(G)-1
+G = nx.convert_node_labels_to_integers(H)
+# 3d spring layout
+pos = nx.spring_layout(G, dim=3, seed=1001)
+# numpy array of x,y,z positions in sorted node order
+xyz = np.array([pos[v] for v in sorted(G)])
+# scalar colors
+scalars = np.array(list(G.nodes())) + 5
+
+mlab.figure()
+
+pts = mlab.points3d(
+    xyz[:, 0],
+    xyz[:, 1],
+    xyz[:, 2],
+    scalars,
+    scale_factor=0.1,
+    scale_mode="none",
+    colormap="Blues",
+    resolution=20,
+)
+
+pts.mlab_source.dataset.lines = np.array(list(G.edges()))
+tube = mlab.pipeline.tube(pts, tube_radius=0.01)
+mlab.pipeline.surface(tube, color=(0.8, 0.8, 0.8))
+mlab.orientation_axes()

share/doc/networkx-3.0/examples/3d_drawing/plot_basic.py

@@ -0,0 +1,51 @@
+"""
+================
+Basic matplotlib
+================
+
+A basic example of 3D Graph visualization using `mpl_toolkits.mplot_3d`.
+
+"""
+
+import networkx as nx
+import numpy as np
+import matplotlib.pyplot as plt
+from mpl_toolkits.mplot3d import Axes3D
+
+# The graph to visualize
+G = nx.cycle_graph(20)
+
+# 3d spring layout
+pos = nx.spring_layout(G, dim=3, seed=779)
+# Extract node and edge positions from the layout
+node_xyz = np.array([pos[v] for v in sorted(G)])
+edge_xyz = np.array([(pos[u], pos[v]) for u, v in G.edges()])
+
+# Create the 3D figure
+fig = plt.figure()
+ax = fig.add_subplot(111, projection="3d")
+
+# Plot the nodes - alpha is scaled by "depth" automatically
+ax.scatter(*node_xyz.T, s=100, ec="w")
+
+# Plot the edges
+for vizedge in edge_xyz:
+    ax.plot(*vizedge.T, color="tab:gray")
+
+
+def _format_axes(ax):
+    """Visualization options for the 3D axes."""
+    # Turn gridlines off
+    ax.grid(False)
+    # Suppress tick labels
+    for dim in (ax.xaxis, ax.yaxis, ax.zaxis):
+        dim.set_ticks([])
+    # Set axes labels
+    ax.set_xlabel("x")
+    ax.set_ylabel("y")
+    ax.set_zlabel("z")
+
+
+_format_axes(ax)
+fig.tight_layout()
+plt.show()

share/doc/networkx-3.0/examples/README.txt

@@ -0,0 +1,8 @@
+.. _examples_gallery:
+
+Gallery
+=======
+
+General-purpose and introductory examples for NetworkX.
+The `tutorial <../tutorial.html>`_ introduces conventions and basic graph
+manipulations.

share/doc/networkx-3.0/examples/algorithms/README.txt

@@ -0,0 +1,2 @@
+Algorithms
+----------