Skip to content

README.md

Latest commit

 

History

History
134 lines (112 loc) · 5.42 KB

README.md

File metadata and controls

134 lines (112 loc) · 5.42 KB

Pointcloud Classification

We describe the procedures to re-produce the experiments for Point-BERT+LLaMA in the paper.
We build our model based on PointBERT official implementations.

1. Environment

Install Basic Packages

conda create -n pointllama python=3.8
conda activate pointllama

conda install pytorch==1.21.1 torchvision==0.13.1 -c pytorch
pip install -r requirements.txt

Building Pytorch Extensions for Chamfer Distance, PointNet++ and kNN

# Chamfer Distance
bash install.sh
# PointNet++
pip install "git+git://github.com/erikwijmans/Pointnet2_PyTorch.git#egg=pointnet2_ops&subdirectory=pointnet2_ops_lib"
# GPU kNN
pip install --upgrade https://github.com/unlimblue/KNN_CUDA/releases/download/0.2/KNN_CUDA-0.2-py3-none-any.whl

2. Experiments Preparation

2.1 Data

Follow the instructions in PointBERT to download and prepare the preprocessed data.

2.2 PointBERT Pretrained Checkpoints

Download the ShapeNet pretrained Point-BERT.pth checkpoints and put it under checkpoints/pointbert.

2.3 LLaMA Checkpoints

Download the LLaMA-7B checkpoint from LLaMA-v1 and put it under checkpoints/llama.

The final directory structure should look like this:

checkpoints
│   ├── llama 
│   │   ├── checklist.chk
│   │   ├── consolidated.00.pth
│   │   └── params.json
│   └── pointbert
│       └── Point-BERT.pth

3. Running Experiments

Training

Single GPU: If you are using a single GPU, run the following command:

bash scripts/train.sh [CONFIG_PATH] [EXP_NAME]

Slurm: If you are using slurm for multi-gpus training, run the following command:

sbatch scripts/train_slurm.sh [CONFIG_PATH] [EXP_NAME]

Replace [CONFIG_PATH] with the path to the config file, and [EXP_NAME] with the name of the experiment.

The training logs and checkpoints will be saved under experiments/[EXP_NAME]

ScanObjectNN-hardest Training Example

We provide an example of how to train our model on ScanObjectNN dataset.

Train our Point-BERT+LLaMA model on ScanObjectNN-hardest split:

// Single GPU
bash scripts/train.sh configs/ScanObjectNN_models/PointLLaMa_hardest.yaml PointTransformer_LLaMA
// Slurm
sbatch scripts/train_slurm.sh configs/ScanObjectNN_models/PointLLaMa_hardest.yaml PointTransformer_LLaMA

Evaluation

You can evaluate the model by checking the log file saved during training or running the following command with the checkpoint:

bash scripts/eval.sh [CONFIG_PATH] [CKPT_PATH]

4. Model Zoo

ScanObjectNN Dataset

Model Split Checkpoint Config Acc
PointBERT hardest model config 83.07
PointBERT objectbg model config 87.43
PointBERT objectonly model config 88.12
PointBERT+LLaMA hardest log / model config 83.87
PointBERT+LLaMA objectbg log / model config 88.64
PointBERT+LLaMA objectonly log / model config 88.81

ModelNet Dataset: TBD

5. Key Places to Watch

  • In llama.py, we re-write LLaMA's code by removing positional embedding and auto-regressive attention masks.
  • In Point_BERT.py, we add the LLaMA layer to Point-BERT according to user's config.

Initialize LLaMa Model

# Line 154:
if config.use_llama:
    llama_default_config = dict(config.llama_cfg)
    self.llama = LLaMATransformer(llama_default_config)
    for param in self.llama.parameters():
        param.requires_grad = False
    self.llama_dim_mapper1 = nn.Linear(config.trans_dim, 4096, bias=False)
    self.llama_dim_mapper2 = nn.Linear(4096, config.trans_dim, bias=False)

Load LLaMa Checkpoint

# Line 215:
if self.config.use_llama:
    print_log("Loading LLaMA checkpoints", logger = 'LLaMA')
    checkpoints = sorted(Path(self.config.llama_path).glob("*.pth"))
    ckpt_path = checkpoints[0]
    checkpoint = torch.load(ckpt_path, map_location="cpu")
    self.llama.custom_load_state_dict(checkpoint, tail=True, strict=False)
    print_log("Loading LLaMA Done", logger = 'LLaMA')

Forward LLaMa Model

# Line 241:
# ...
# x = self.blocks(x, pos)
if self.config.use_llama:
    x = self.llama_dim_mapper1(x)
    x = self.llama(x)
    x = self.llama_dim_mapper2(x)
# x = self.norm(x)