IMPORTANT

My VASA hack project https://github.com/johndpope/vasa-1-hack has running /training code stage 1 (megaportraits) - with hot fixes https://github.com/johndpope/VASA-1-hack/blob/main/train_stage_1.py

MegaPortrait - SamsungLabs AI - Russia

Implementation of Megaportrait using Claude Opus

All models / code is in model.py

memory debug

    mprof run train.py

or just

    python train.py

UPDATES

• Save / restore checkpoint) specify in config ./configs/training/stage10base.yaml to restore checkpoint
• auto crop video frames to sweet spot
• tensorboard losses
• LPIPS
• additional imagepyramide from one shot view code for loss - (this broke things..)

EmoDataset

warp / crop / spline / remove background / transforms

Training Data (☢️ dont need this yet.)

• Total Videos: 35,000 facial videos
• Total Size: 40GB

Training Strategy

for now - to simplify problem - use the 4 videos in junk folder. once models are validated - can point the video_dir to above torrent

 # video_dir:  '/Downloads/CelebV-HQ/celebvhq/35666'  
  video_dir: './junk'

the preprocessing is taking 1-2 mins for each video - I add some saving to npz format for faster reloading.

Torrent Download

You can download the dataset via the provided magnet link or by visiting Academic Torrents.

magnet:?xt=urn:btih:843b5adb0358124d388c4e9836654c246b988ff4&dn=CelebV-HQ&tr=https%3A%2F%2Facademictorrents.com%2Fannounce.php&tr=https%3A%2F%2Fipv6.academictorrents.com%2Fannounce.php

Implemented Functionality / Descriptions

Base Model (Gbase)

• Description: Responsible for creating the foundational neural head avatar at a medium resolution of (512 x 512). Uses volumetric features to encode appearance and latent descriptors to encode motion.
• Components:
  • Appearance Encoder (Eapp): Encodes the appearance of the source frame into volumetric features and a global descriptor.

    class Eapp(nn.Module):
    # Architecture details omitted for brevity

  • Motion Encoder (Emtn): Encodes the motion from both source and driving images into head rotations, translations, and latent expression descriptors.

    class Emtn(nn.Module):
    # Architecture details omitted for brevity

  • Warping Generators (Wsrc_to_can and Wcan_to_drv): Removes motion from the source and imposes driver motion onto canonical features.

    class WarpGenerator(nn.Module):
    # Architecture details omitted for brevity

  • 3D Convolutional Network (G3D): Processes canonical volumetric features.

    class G3D(nn.Module):
    # Architecture details omitted for brevity

  • 2D Convolutional Network (G2D): Projects 3D features into 2D and generates the output image.

    class G2D(nn.Module):
    # Architecture details omitted for brevity

High-Resolution Model (GHR)

• Description: Enhances the resolution of the base model output from (512 \times 512) to (1024 \times 1024) using a high-resolution dataset of photographs.
• Components:
  • Encoder: Takes the base model output and produces a 3D feature tensor.

    class EncoderHR(nn.Module):
    # Architecture details omitted for brevity

  • Decoder: Converts the 3D feature tensor to a high-resolution image.

    class DecoderHR(nn.Module):
    # Architecture details omitted for brevity

Student Model (Student)

• Description: A distilled version of the high-resolution model for real-time applications. Trained to mimic the full model’s predictions but runs faster and is limited to a predefined number of avatars.
• Components:
  • ResNet18 Encoder: Encodes the input image.

    class ResNet18(nn.Module):
    # Architecture details omitted for brevity

  • Generator with SPADE Normalization Layers: Generates the final output image. Each SPADE block uses tensors specific to an avatar.

    class SPADEGenerator(nn.Module):
    # Architecture details omitted for brevity

Gaze and Blink Loss Model

• Description: Computes the gaze and blink loss using a pretrained face mesh from MediaPipe and a custom network. The gaze loss uses MAE and MSE, while the blink loss uses binary cross-entropy.
• Components:
  • Backbone (VGG16): Extracts features from the eye images.

    class VGG16Backbone(nn.Module):
    # Architecture details omitted for brevity

  • Keypoint Network: Processes 2D keypoints.

    class KeypointNet(nn.Module):
    # Architecture details omitted for brevity

  • Gaze Head: Predicts gaze direction.

    class GazeHead(nn.Module):
    # Architecture details omitted for brevity

  • Blink Head: Predicts blink probability.

    class BlinkHead(nn.Module):
    # Architecture details omitted for brevity

Training Functions

• train_base(cfg, Gbase, Dbase, dataloader): Trains the base model using perceptual, adversarial, and cycle consistency losses.

  def train_base(cfg, Gbase, Dbase, dataloader):
    # Training code omitted for brevity

• train_hr(cfg, GHR, Dhr, dataloader): Trains the high-resolution model using super-resolution objectives and adversarial losses.

  def train_hr(cfg, GHR, Dhr, dataloader):
    # Training code omitted for brevity

• train_student(cfg, Student, GHR, dataloader): Distills the high-resolution model into a student model for faster inference.

  def train_student(cfg, Student, GHR, dataloader):
    # Training code omitted for brevity

Training Pipeline

• Data Augmentation: Applies random horizontal flips, color jitter, and other augmentations to the input images.
• Optimizers: Uses AdamW optimizer with cosine learning rate scheduling for both base and high-resolution models.
• Losses:
  • Perceptual Loss: Matches the content and facial appearance between predicted and ground-truth images.
  • Adversarial Loss: Ensures the realism of predicted images using a multi-scale patch discriminator.
  • Cycle Consistency Loss: Prevents appearance leakage through the motion descriptor.

Main Function

• Description: Sets up the dataset and data loaders, initializes the models, and calls the training functions for base, high-resolution, and student models.

• Implementation:

  def main(cfg: OmegaConf) -> None:
    use_cuda = torch.cuda.is_available()
    device = torch.device("cuda" if use_cuda else "cpu")
  
    transform = transforms.Compose([
        transforms.ToTensor(),
        transforms.Normalize([0.5], [0.5]),
        transforms.RandomHorizontalFlip(),
        transforms.ColorJitter()
    ])
  
    dataset = EMODataset(
        use_gpu=use_cuda,
        width=cfg.data.train_width,
        height=cfg.data.train_height,
        n_sample_frames=cfg.training.n_sample_frames,
        sample_rate=cfg.training.sample_rate,
        img_scale=(1.0, 1.0),
        video_dir=cfg.training.video_dir,
        json_file=cfg.training.json_file,
        transform=transform
    )
  
    dataloader = DataLoader(dataset, batch_size=4, shuffle=True, num_workers=4)
  
    Gbase = model.Gbase()
    Dbase = model.Discriminator()
    train_base(cfg, Gbase, Dbase, dataloader)
  
    GHR = model.GHR()
    GHR.Gbase.load_state_dict(Gbase.state_dict())
    Dhr = model.Discriminator()
    train_hr(cfg, GHR, Dhr, dataloader)
  
    Student = model.Student(num_avatars=100)
    train_student(cfg, Student, GHR, dataloader)
  
    torch.save(Gbase.state_dict(), 'Gbase.pth')
    torch.save(GHR.state_dict(), 'GHR.pth')
    torch.save(Student.state_dict(), 'Student.pth')
  
  if __name__ == "__main__":
    config = OmegaConf.load("./configs/training/stage1-base.yaml")
    main(config)

rome/losses - cherry picked from https://github.com/SamsungLabs/rome

wget 'https://download.pytorch.org/models/resnet18-5c106cde.pth' extract to state_dicts

RT-GENE (Real-Time Gaze Estimation) - couldn't get working

git clone https://github.com/Tobias-Fischer/rt_gene.git
cd rt_gene/rt_gene
pip install .