3D-PI0

This repository contains code for training and evaluating 3D-PI0, a vision-language model for 3D understanding and generation.

Setup

1. Create and activate conda environment:

conda create -n cogvideo python=3.10
conda activate cogvideo

2. Install dependencies:

pip install -r requirements.txt

Training

To train the model:

bash slurm/train_multi_gpu.sh

Directory Structure

• slurm/: Training scripts and configurations
• scripts/: Main training code
• src/: Source code
  • cog_video/: Core model implementation

Notes

• The model requires GPU for training
• Logs are stored in logs/ directory
• Model checkpoints will be saved in the specified checkpoint directory

License

MIT