LLaVA-Next-3D/README.md

Installation

1. Clone this repository:

git clone https://github.com/ZCMax/LLaVA-Next-3D.git
cd LLaVA-Next-3D

2. Create the conda environment:

conda create -n llavanext3d python=3.10 -y
conda activate llavanext3d
pip install --upgrade pip  # Enable PEP 660 support.
pip install -e ".[train]"
pip install flash-attn --no-build-isolation     # install flash attention

Data Preparation

The directory should be orgainized as:

LLaVA-3D-Next # project root
├── data
│   ├── scannet
│   │   ├── scans
│   │   ├── posed_images
│   │   ├── pcd_with_object_aabbs
│   │   └── mask
│   ├── embodiedscan
│   │   ├── embodiedscan_infos_full_llava3d_v2.json
│   ├── metadata
│   │   ├── scannet_select_frames.json
│   │   ├── pcd_discrete_0.1.pkl
│   │   ├── scannet_train_gt_box.json
│   │   └── scannet_val_pred_box.json
│   ├── prcoessed
│   │   ├── multi3drefer_train_llava_style.json
│   │   ├── multi3drefer_val_llava_style.json
│   │   ├── ...

We have prepared the well organzied data under /mnt/hwfile/openmmlab/zhuchenming/llava-next-3d-data, you can directly link this to your data. Currently we only support training on ScanNet.

Training & Inference

Full-finetuned Training

You can use sbatch to launch the multi-node script:

sh scripts/3d/train/train_16gpu_sbatch.sh

Inference

sh scripts/3d/eval/eval_scanrefer.sh $CKPT_NAME uniform 32