Skip to content

Conversation

@jmanhype
Copy link

@jmanhype jmanhype commented Mar 7, 2025

This PR adds Docker support for containerized training

@jmanhype
Copy link
Author

jmanhype commented Mar 7, 2025

Docker Setup for InsTaG Training Framework

English

This pull request provides a complete Docker-based environment for the InsTaG training framework. It addresses several setup challenges documented in the issues by providing a consistent, containerized environment.

Key Features:

  1. Dual Container Architecture:

    • Main container (CUDA 11.7, Python 3.9) for training and inference
    • Separate Sapiens container (CUDA 12.1, Python 3.10) for geometry priors
  2. Helper Scripts:

    • docker-run.sh - Simplifies common operations
    • setup-docker.sh - Automates initial setup and dependency installation
  3. Comprehensive Documentation:

    • Complete workflow examples
    • Detailed troubleshooting guidance
    • Support for different audio feature extractors (DeepSpeech, Wav2Vec, AVE, HuBERT)
  4. Automated Setup:

    • OpenFace integration for facial AU extraction
    • EasyPortrait model download
    • Sapiens model download
  5. Workflow Improvements:

    • No manual environment conflicts
    • Simplified audio feature extraction
    • Streamlined teeth mask generation
    • Container-based geometry prior generation

The documentation includes examples for both short-video adaptation (with geometry priors) and long-video training, making it easier to use the framework in various scenarios.


中文

此 Pull Request 为 InsTaG 训练框架提供了完整的基于 Docker 的环境。它通过提供一致的容器化环境解决了 issues 中记录的几个设置挑战。

主要特点:

  1. 双容器架构:

    • 主容器(CUDA 11.7,Python 3.9)用于训练和推理
    • 单独的 Sapiens 容器(CUDA 12.1,Python 3.10)用于几何先验生成
  2. 辅助脚本:

    • docker-run.sh - 简化常见操作
    • setup-docker.sh - 自动化初始设置和依赖安装
  3. 全面的文档:

    • 完整的工作流示例
    • 详细的故障排除指南
    • 支持不同的音频特征提取器(DeepSpeech、Wav2Vec、AVE、HuBERT)
  4. 自动化设置:

    • OpenFace 集成用于面部 AU 提取
    • EasyPortrait 模型下载
    • Sapiens 模型下载
  5. 工作流改进:

    • 没有手动环境冲突
    • 简化的音频特征提取
    • 简化的牙齿遮罩生成
    • 基于容器的几何先验生成

文档包括短视频适应(带几何先验)和长视频训练的示例,使框架在各种场景中更易于使用。

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant