### Running TorchSparse Plugin Demo Container

Source: https://github.com/mit-han-lab/torchsparse/blob/master/examples/README.md

This command starts a Docker container with all necessary dependencies installed for TorchSparse plugins. It mounts the KITTI and NuScenes datasets into the container for use in the demos.

```bash
docker run -it --gpus all --mount type=bind,source=<kitti_dataset_root>,target=/root/data/kitti --mount type=bind,source=<nuscenes_dataset_root>,target=/root/data/nuscenes ioeddk/torchsparse_plugin_demo:latest
```

--------------------------------

### Install TorchSparse using Installation Script

Source: https://github.com/mit-han-lab/torchsparse/blob/master/README.md

Install TorchSparse by running the provided Python script, which automatically handles dependency checks and installation.

```bash
python -c "$(curl -fsSL https://raw.githubusercontent.com/mit-han-lab/torchsparse/master/install.py)"
```

--------------------------------

### Install TorchSparse from Source

Source: https://github.com/mit-han-lab/torchsparse/blob/master/README.md

Install the library directly from the source code using setup.py. This method is useful if pre-built packages are not suitable.

```bash
python setup.py install
```

--------------------------------

### Install TorchSparse using Pip from Git Repository

Source: https://github.com/mit-han-lab/torchsparse/blob/master/README.md

Install TorchSparse directly from its GitHub repository using pip, without needing to clone the repository locally.

```bash
pip install git+https://github.com/mit-han-lab/torchsparse.git
```

--------------------------------

### Check CUDA Installations

Source: https://github.com/mit-han-lab/torchsparse/blob/master/docs/FAQ.md

Lists directories related to CUDA installations to help identify potential conflicts or multiple versions.

```bash
ls /usr/local/cuda* -d
```

--------------------------------

### Check CUDA Installation for PyTorch

Source: https://github.com/mit-han-lab/torchsparse/blob/master/README.md

Ensure PyTorch was installed with CUDA support if GPU acceleration is required for TorchSparse.

```bash
python -c "import torch; print(torch.version.cuda)"
>>> 11.3
```

--------------------------------

### Check PyTorch Version

Source: https://github.com/mit-han-lab/torchsparse/blob/master/README.md

Verify that PyTorch version 1.9.0 or higher is installed. This is a prerequisite for installing TorchSparse.

```bash
python -c "import torch; print(torch.__version__)"
>>> 1.10.0
```

--------------------------------

### Compile TorchSparse for Specific GPU Architectures

Source: https://github.com/mit-han-lab/torchsparse/blob/master/docs/FAQ.md

Installs or upgrades TorchSparse, specifying the target CUDA compute architectures using the TORCH_CUDA_ARCH_LIST environment variable. Useful for cross-compilation or targeting specific GPU types.

```bash
TORCH_CUDA_ARCH_LIST="7.0;7.5" pip install --upgrade git+https://github.com/mit-han-lab/torchsparse.git
```

--------------------------------

### Verify CUDA and PyTorch Compatibility

Source: https://github.com/mit-han-lab/torchsparse/blob/master/docs/FAQ.md

Checks if the installed NVCC (CUDA compiler) and PyTorch are using the same CUDA version.

```bash
nvcc --version
python -c "import torch; print(torch.version.cuda);"
```

--------------------------------

### Limit Parallel Compilation to Prevent Memory Errors

Source: https://github.com/mit-han-lab/torchsparse/blob/master/docs/FAQ.md

Installs or upgrades TorchSparse while limiting the number of parallel compilation jobs using the MAX_JOBS environment variable. This helps prevent 'Killed' errors due to out-of-memory issues during compilation.

```bash
MAX_JOBS=2 pip install --upgrade git+https://github.com/mit-han-lab/torchsparse.git
```

--------------------------------

### Run SECOND Demo

Source: https://github.com/mit-han-lab/torchsparse/blob/master/examples/openpcdet/demo.ipynb

Configures and runs the SECOND model demo. Prints standard output from the inference process.

```python
second_registry = {
    'ckpt_before': 'SECOND/second_7862.pth',
    'ckpt_after': 'SECOND/second_7862.pth',
    'openpc_cfg_path': 'kitti_models/second.yaml',
    'torchsparse_cfg_path': 'kitti_models/second_plugin.yaml',
    'v_spconv': 1
}

second_results = openpc_single_demo(second_registry, base_paths, convert=True)
print(second_results.stdout)
```

--------------------------------

### Run VoxelNeXt Demo with Conversion

Source: https://github.com/mit-han-lab/torchsparse/blob/master/examples/openpcdet/demo.ipynb

Prepares paths and runs the VoxelNeXt model demo using a specialized converter. Prints standard error from the inference process.

```python
from converter_voxelnext import convert_voxelnext

voxelnext_registry = {
    'ckpt_before': 'VoxelNeXt/voxelnext_nuscenes_kernel1.pth',
    'ckpt_after': 'VoxelNeXt/voxelnext_nuscenes_kernel1.pth',
    'openpc_cfg_path': 'nuscenes_models/cbgs_voxel0075_voxelnext.yaml',
    'torchsparse_cfg_path': 'nuscenes_models/cbgs_voxel0075_voxelnext.yaml',
    'v_spconv': 2
}

# pre-process paths
openpc_cfg_path = os.path.join(base_paths['openpc_cfg_base_path'], voxelnext_registry['openpc_cfg_path'])
torchsparse_cfg_path = os.path.join(base_paths['torchsparse_cfg_base_path'], voxelnext_registry['torchsparse_cfg_path'])
test_file_path = os.path.join(base_paths['openpc_path'], "tools/test.py")
openpc_model_path = os.path.join(base_paths['openpc_model_base_path'], voxelnext_registry['ckpt_before'])
torchsparse_model_path = os.path.join(base_paths['torchsparse_model_base_path'], voxelnext_registry['ckpt_after'])

# convert_voxelnext(
#     ckpt_before=openpc_model_path,
#     ckpt_after=torchsparse_model_path,
# )

command = f'bash -c "conda activate torchsparse; python {test_file_path} --cfg_file {torchsparse_cfg_path} --ckpt {torchsparse_model_path}"'
voxelnext_results = subprocess.run(command, capture_output=True, text=True, shell=True, executable='/bin/bash')

print(voxelnext_results.stderr)
```

--------------------------------

### PV-RCNN Configuration and Demo

Source: https://github.com/mit-han-lab/torchsparse/blob/master/examples/mmdetection3d/demo.ipynb

Sets up the configuration for the PV-RCNN model and runs a single inference demo. This snippet is useful for performing object detection with PV-RCNN on a dataset.

```python
pv_rcnn_config = {
    "ckpt_before": "PV-RCNN/pv_rcnn_8xb2-80e_kitti-3d-3class_20221117_234428-b384d22f.pth",
    "ckpt_after": "PV-RCNN/pv_rcnn_8xb2-80e_kitti-3d-3class_20221117_234428-b384d22f.pth",
    "cfg_path": "pv_rcnn/pv_rcnn_8xb2-80e_kitti-3d-3class.py",
    "v_spconv": 1,
    "cfg_options": "--cfg-options test_evaluator.pklfile_prefix=outputs/torchsparse/pv_rcnn --cfg-options model.middle_encoder.type=SparseEncoderTS --cfg-options model.points_encoder.type=VoxelSetAbstractionTS"
}

pv_rcnn_results = mmdet3d_single_demo(pv_rcnn_config, base_paths, convert=True)
print(pv_rcnn_results.stderr)
```

--------------------------------

### Run MMDetection3D Demo with TorchSparse

Source: https://github.com/mit-han-lab/torchsparse/blob/master/examples/mmdetection3d/README.md

Execute the MMDetection3D test script with TorchSparse model checkpoints and configuration options. Ensure all paths and options are correctly specified for your model.

```bash
python <test_file_path> <cfg_path> <torchsparse_model_path> <cfg_options> --task lidar_det
```

--------------------------------

### Run PV-RCNN Demo

Source: https://github.com/mit-han-lab/torchsparse/blob/master/examples/openpcdet/demo.ipynb

Configures and runs the PV-RCNN model demo. Prints standard output from the inference process.

```python
pv_rcnn_registry = {
    'ckpt_before': 'PV-RCNN/pv_rcnn_8369.pth',
    'ckpt_after': 'PV-RCNN/pv_rcnn_8369.pth',
    'openpc_cfg_path': 'kitti_models/pv_rcnn.yaml',
    'torchsparse_cfg_path': 'kitti_models/pv_rcnn_plugin.yaml',
    'v_spconv': 1
}

pv_rcnn_results = openpc_single_demo(pv_rcnn_registry, base_paths, convert=True)
print(pv_rcnn_results.stdout)
```

--------------------------------

### Run OpenPCDet Evaluation with TorchSparse Model

Source: https://github.com/mit-han-lab/torchsparse/blob/master/examples/openpcdet/README.md

Execute the evaluation script after converting weights and setting up the environment. Specify the paths to the OpenPCDet test file, TorchSparse config, and the converted model checkpoint.

```bash
python <test_file_path> --cfg_file <torchsparse_cfg_path> --ckpt <torchsparse_model_path>
```

--------------------------------

### Load TorchSparse Weight Conversion Module

Source: https://github.com/mit-han-lab/torchsparse/blob/master/examples/mmdetection3d/demo.ipynb

Loads the weight conversion module from a local file. Ensure 'converter.py' is in the parent directory.

```python
import importlib.util
import sys, os
from pathlib import Path
import subprocess

# Define the relative path to the file
relative_path = "../converter.py"
file_path = Path().resolve() / relative_path

# Add the directory containing the file to sys.path
sys.path.append(str(file_path.parent))

# Load the module
spec = importlib.util.spec_from_file_location("convert_weights", str(file_path))
converter = importlib.util.module_from_spec(spec)
spec.loader.exec_module(converter)

converter = getattr(converter, "convert_weights")
```

--------------------------------

### Define Model Configuration for SECOND

Source: https://github.com/mit-han-lab/torchsparse/blob/master/examples/mmdetection3d/demo.ipynb

Configuration dictionary for the SECOND 3D car detection model, specifying checkpoint paths, configuration file, and TorchSparse specific settings.

```python
second_3d_car = {
    'ckpt_before': 'SECOND/second_hv_secfpn_8xb6-80e_kitti-3d-car-75d9305e.pth',
    'ckpt_after': 'SECOND/second_hv_secfpn_8xb6-80e_kitti-3d-car-75d9305e.pth',
    'cfg_path': 'second/second_hv_secfpn_8xb6-80e_kitti-3d-car.py',
    'v_spconv': 2,
    'cfg_options': "--cfg-options test_evaluator.pklfile_prefix=outputs/torchsparse/second --cfg-options model.middle_encoder.type=SparseEncoderTS"
}
```

--------------------------------

### Run Part-A2 Demo

Source: https://github.com/mit-han-lab/torchsparse/blob/master/examples/openpcdet/demo.ipynb

Configures and runs the Part-A2 model demo. Prints standard error from the inference process.

```python
part_a2_registry = {
    'ckpt_before': 'Part-A2/PartA2_7940.pth',
    'ckpt_after': 'Part-A2/PartA2_7940.pth',
    'openpc_cfg_path': 'kitti_models/PartA2.yaml',
    'torchsparse_cfg_path': 'kitti_models/PartA2_plugin.yaml',
    'v_spconv': 1
}

part_a2_results = openpc_single_demo(part_a2_registry, base_paths, convert=True)
print(part_a2_results.stderr)
```

--------------------------------

### Specify Paths and Environment Parameters

Source: https://github.com/mit-han-lab/torchsparse/blob/master/examples/mmdetection3d/demo.ipynb

Sets up base paths for MMDetection3D, model weights (input and output), and configuration files, along with the conda environment name. Customize these paths according to your project structure.

```python
env_name = "torchsparse"

# Please complete the following base paths. 
base_paths = {
    'mmdet3d_path': None,
    'mmdet3d_model_base_path': None,
    'torchsparse_model_base_path': os.path.join(os.path.abspath(''), "converted_models"),
    'mmdet3d_cfg_base_path': None
}

```

--------------------------------

### Run SECOND Model Demo

Source: https://github.com/mit-han-lab/torchsparse/blob/master/examples/mmdetection3d/demo.ipynb

Executes the SECOND 3D car detection model demo using the `mmdet3d_single_demo` function. Prints the standard error and standard output from the evaluation process.

```python
second_results = mmdet3d_single_demo(second_3d_car, base_paths, convert=True)
print(second_results.stderr)
print(second_results.stdout)
```

--------------------------------

### Dummy Check for Weight Converter Loading

Source: https://github.com/mit-han-lab/torchsparse/blob/master/examples/mmdetection3d/demo.ipynb

Prints the loaded weight converter object to verify successful loading.

```python
# Dummy check for whether the weight converter is successfully loaded. 
print(converter)
```

--------------------------------

### Base Paths Dictionary

Source: https://github.com/mit-han-lab/torchsparse/blob/master/examples/mmdetection3d/demo.ipynb

A dictionary containing essential base paths for MMDetection3D and TorchSparse, including paths to the MMDetection3D repository, model directories, and configuration files.

```python
base_paths

```

--------------------------------

### Running Single Model Demo with OpenPCDet and TorchSparse

Source: https://github.com/mit-han-lab/torchsparse/blob/master/examples/openpcdet/demo.ipynb

This function handles the conversion of model weights and subsequent evaluation using OpenPCDet's test script. It includes assertions to validate path existence and constructs a bash command to activate a conda environment and run the evaluation.

```python
def openpc_single_demo(registry_entry, base_paths, convert=True):

    assert os.path.isdir(base_paths['openpc_model_base_path']), f"OpenPCDet model base path {base_paths['openpc_model_base_path']} does not exist."
    assert os.path.isdir(base_paths['torchsparse_model_base_path']), f"TorchSparse model base path {base_paths['torchsparse_model_base_path']} does not exist."
    assert os.path.isdir(base_paths['openpc_cfg_base_path']), f"OpenPCDet cfg base path {base_paths['openpc_cfg_base_path']} does not exist."
    assert os.path.isdir(base_paths['torchsparse_cfg_base_path']), f"TorchSparse cfg base path {base_paths['torchsparse_cfg_base_path']} does not exist."
    assert os.path.isdir(base_paths['openpc_path']), f"OpenPCDet path {base_paths['openpc_path']} does not exist."

    # pre-process paths
    openpc_cfg_path = os.path.join(base_paths['openpc_cfg_base_path'], registry_entry['openpc_cfg_path'])
    torchsparse_cfg_path = os.path.join(base_paths['torchsparse_cfg_base_path'], registry_entry['torchsparse_cfg_path'])
    test_file_path = os.path.join(base_paths['openpc_path'], "tools/test.py")
    openpc_model_path = os.path.join(base_paths['openpc_model_base_path'], registry_entry['ckpt_before'])
    assert os.path.isdir(base_paths['torchsparse_model_base_path']), "Please create the directory for the converted model."
    torchsparse_model_path = os.path.join(base_paths['torchsparse_model_base_path'], registry_entry['ckpt_after'])

    # convert the model
    if convert:
        parent_dir = os.path.dirname(torchsparse_model_path)
        if not os.path.exists(parent_dir):
            os.makedirs(parent_dir)
        converter(
            ckpt_before=openpc_model_path,
            ckpt_after=torchsparse_model_path,
            cfg_path=openpc_cfg_path,
            v_spconv = registry_entry['v_spconv'],
            framework='openpc'
        )

    command = f'bash -c "conda activate {env_name}; python {test_file_path} --cfg_file {torchsparse_cfg_path} --ckpt {torchsparse_model_path}"'
    print(command)
    result = subprocess.run(command, capture_output=True, text=True, shell=True, executable='/bin/bash')
    return result  # result have .stdout and .stderr attributes to get the output.
    
```

--------------------------------

### CenterPoint Configuration and Demo

Source: https://github.com/mit-han-lab/torchsparse/blob/master/examples/mmdetection3d/demo.ipynb

Configures and runs a single inference demo for the CenterPoint model with circular NMS on NuScenes dataset. Ensure the dataset path is correctly set in the mmdetection3d config.

```python
centerpoint_config = {
    "ckpt_before": "CenterPoint/centerpoint_01voxel_second_secfpn_circlenms_4x8_cyclic_20e_nus_20220810_030004-9061688e.pth",
    "ckpt_after": "CenterPoint/centerpoint_01voxel_second_secfpn_circlenms_4x8_cyclic_20e_nus_20220810_030004-9061688e.pth",
    "cfg_path": "centerpoint/centerpoint_voxel01_second_secfpn_head-circlenms_8xb4-cyclic-20e_nus-3d.py",
    "v_spconv": 1,
    "cfg_options": "--cfg-options model.pts_middle_encoder.type=SparseEncoderTS"
}

centerpoint_results = mmdet3d_single_demo(centerpoint_config, base_paths, convert=True)
print(centerpoint_results.stderr)
```

--------------------------------

### Convert SpConv Checkpoint to TorchSparse

Source: https://github.com/mit-han-lab/torchsparse/blob/master/examples/openpcdet/README.md

Use this command-line tool to convert model weights from SpConv format to TorchSparse format. Ensure all paths and SpConv version are correctly specified.

```bash
python examples/converter.py --ckpt_before ../OpenPCDet/models/SECOND/second_7862.pth --cfg_path ../OpenPCDet/tools/cfgs/kitti_models/second.yaml --ckpt_after ./converted/SECOND/second_7862.pth --v_spconv 1 --framework openpc
```

--------------------------------

### SECOND Model Configuration Options

Source: https://github.com/mit-han-lab/torchsparse/blob/master/examples/mmdetection3d/README.md

Specify configuration options for the SECOND model to use TorchSparse layers. This includes setting the evaluation pickle file prefix and the middle encoder type.

```bash
--cfg-options test_evaluator.pklfile_prefix=outputs/torchsparse/second --cfg-options model.middle_encoder.type=SparseEncoderTS
```

--------------------------------

### Setting Base Paths for Model Conversion

Source: https://github.com/mit-han-lab/torchsparse/blob/master/examples/openpcdet/demo.ipynb

Defines the base directory paths required for OpenPCDet and TorchSparse model conversion and configuration. Ensure these paths are correctly set before running conversion scripts.

```python
env_name = "torchsparse"

base_paths = {
    'openpc_path': None,
    'openpc_model_base_path': None,
    'torchsparse_model_base_path': os.path.joinin(os.path.abspath(''), "converted_models"),
    'openpc_cfg_base_path': None,
    'torchsparse_cfg_base_path': os.path.joinin(os.path.abspath(''), "cfgs")
}
```

--------------------------------

### Convert MMDetection3D Model Weights

Source: https://github.com/mit-han-lab/torchsparse/blob/master/examples/README.md

Converts pre-trained MMDetection3D model weights to a format compatible with TorchSparse. Ensure the paths to the checkpoint, config file, and output directory are correct.

```bash
python examples/converter.py --ckpt_before ../mmdetection3d/models/PV-RCNN/pv_rcnn_8xb2-80e_kitti-3d-3class_20221117_234428-b384d22f.pth --cfg_path ../mmdetection3d/pv_rcnn/pv_rcnn_8xb2-80e_kitti-3d-3class.py --ckpt_after ./converted/PV-RCNN/pv_rcnn_8xb2-80e_kitti-3d-3class_20221117_234428-b384d22f.pth --v_spconv 1 --framework mmdet3d
```

--------------------------------

### Verify Model Converter Loading

Source: https://github.com/mit-han-lab/torchsparse/blob/master/examples/openpcdet/demo.ipynb

Prints the loaded model converter object to verify that the module was successfully imported and is accessible.

```python
# Check that the model converter is successfully loaded
print(converter)
```

--------------------------------

### PV-RCNN Model Configuration Options

Source: https://github.com/mit-han-lab/torchsparse/blob/master/examples/mmdetection3d/README.md

Specify configuration options for the PV-RCNN model to use TorchSparse layers. This includes setting the evaluation pickle file prefix, the middle encoder type, and the points encoder type.

```bash
--cfg-options test_evaluator.pklfile_prefix=outputs/torchsparse/pv_rcnn --cfg-options model.middle_encoder.type=SparseEncoderTS --cfg-options model.points_encoder.type=VoxelSetAbstractionTS
```

--------------------------------

### Result of Base Paths

Source: https://github.com/mit-han-lab/torchsparse/blob/master/examples/mmdetection3d/demo.ipynb

The output showing the resolved base paths for MMDetection3D and TorchSparse, indicating the directory structure for models and configurations.

```python
{
'mmdet3d_path': '/home/yingqi/repo/mmdetection3d',
'mmdet3d_model_base_path': '/home/yingqi/repo/mmdetection3d/models',
'torchsparse_model_base_path': '/home/yingqi/repo/torchsparse-dev/examples/mmdetection3d/converted_models',
'mmdet3d_cfg_base_path': '/home/yingqi/repo/mmdetection3d/configs'
}
```

--------------------------------

### PCEngine Citation (MLSys 2023)

Source: https://github.com/mit-han-lab/torchsparse/blob/master/README.md

BibTeX entry for the PCEngine paper accepted by MLSys 2023. This relates to efficient sparse convolution kernels.

```bibtex
@inproceedings{hong2023pcengine,
  title={{Exploiting Hardware Utilization and Adaptive Dataflow for Efficient Sparse Convolution in 3D Point Clouds}},
  author={Hong, Ke and Yu, Zhongming and Dai, Guohao and Yang, Xinhao and Lian, Yaoxiu and Liu, Zehao and Xu, Ningyi and Wang, Yu},
  booktitle={Sixth Conference on Machine Learning and Systems (MLSys)},
  year={2023}
}
```

--------------------------------

### TorchSparse++ Citation (CVPR Workshops 2023)

Source: https://github.com/mit-han-lab/torchsparse/blob/master/README.md

BibTeX entry for the preliminary version of TorchSparse++ presented at CVPR Workshops 2023. This cites an earlier release.

```bibtex
@inproceedings{tangandyang2023torchsparse++,
  title = {{TorchSparse++: Efficient Point Cloud Engine}},
  author = {Tang, Haotian and Yang, Shang and Liu, Zhijian and Hong, Ke and Yu, Zhongming and Li, Xiuyu and Dai, Guohao and Wang, Yu and Han, Song},
  booktitle = {Computer Vision and Pattern Recognition Workshops (CVPRW)},
  year = {2023}
}
```

--------------------------------

### MMDetection3D Single Model Demo Function

Source: https://github.com/mit-han-lab/torchsparse/blob/master/examples/mmdetection3d/demo.ipynb

Python function to run a single MMDetection3D model demo. It handles model conversion (optional) and executes the evaluation using `tools/test.py`.

```python
def mmdet3d_single_demo(registry_entry, base_paths, convert=True):
    """Run Single Model Demo

    :param registry_entry: the model demo registry. 
    :type registry_entry: dict
    :param base_paths: the base paths. 
    :type base_paths: dict
    :param convert: whether to convert the model. If set to false, it skip the model conversion and use the provided checkpoint to run model evaluation directly. Defaults to True.
    :type convert: bool, optional
    :return: return the process object that used to run the demo. 
    :rtype: CompletedProcess
    """

    assert os.path.isdir(base_paths['mmdet3d_path']), "Please specify the mmdet3d_path in the base_paths."
    assert os.path.isdir(base_paths['mmdet3d_model_base_path']), "Please specify the mmdet3d_model_base_path in the base_paths."
    assert os.path.isdir(base_paths['torchsparse_model_base_path']), "Please specify the torchsparse_model_base_path in the base_paths."
    assert os.path.isdir(base_paths['mmdet3d_cfg_base_path']), "Please specify the mmdet3d_cfg_base_path in the base_paths."

    # pre-process paths
    cfg_path = os.path.join(base_paths['mmdet3d_cfg_base_path'], registry_entry['cfg_path'])
    test_file_path = os.path.join(base_paths['mmdet3d_path'], "tools/test.py")
    mmdet3d_model_path = os.path.join(base_paths['mmdet3d_model_base_path'], registry_entry['ckpt_before'])
    assert os.path.isdir(base_paths['torchsparse_model_base_path']), "Please create the directory for the converted model."
    torchsparse_model_path = os.path.join(base_paths['torchsparse_model_base_path'], registry_entry['ckpt_after'])
    
    cfg_options = registry_entry['cfg_options']
    # convert the model
    if convert:
        parent_dir = os.path.dirname(torchsparse_model_path)
        if not os.path.exists(parent_dir):
            os.makedirs(parent_dir)
        converter(
            ckpt_before=mmdet3d_model_path,
            ckpt_after=torchsparse_model_path,
            cfg_path=cfg_path,
            v_spconv = registry_entry['v_spconv']
        )

    command = f'bash -c "conda activate {env_name}; python {test_file_path} {cfg_path} {torchsparse_model_path} {cfg_options} --task lidar_det"'
    print(command)
    result = subprocess.run(command, capture_output=True, text=True, shell=True, executable='/bin/bash')
    return result  # result have .stdout and .stderr attributes to get the output. 
    
```

--------------------------------

### Load Model Converter Module

Source: https://github.com/mit-han-lab/torchsparse/blob/master/examples/openpcdet/demo.ipynb

Loads a custom model converter module from a relative path, making it available for use in the current environment. Ensure the 'converter.py' file exists in the specified relative path.

```python
import importlib.util
import sys, os
from pathlib import Path
import subprocess

# Loading the model converter to this notebook. 
# Define the relative path to the file
relative_path = "../converter.py"
file_path = Path().resolve() / relative_path

# Add the directory containing the file to sys.path
sys.path.append(str(file_path.parent))

# Load the module
spec = importlib.util.spec_from_file_location("convert_weights", str(file_path))
converter = importlib.util.module_from_spec(spec)
spec.loader.exec_module(converter)

converter = getattr(converter, "convert_weights")
```

--------------------------------

### PV-RCNN Demo Output

Source: https://github.com/mit-han-lab/torchsparse/blob/master/examples/mmdetection3d/demo.ipynb

Prints the standard output from the PV-RCNN demo, which typically contains detection results or processing information.

```python
print(pv_rcnn_results.stdout)
```

--------------------------------

### Disable PyTorch JIT Compile

Source: https://github.com/mit-han-lab/torchsparse/blob/master/examples/openpcdet/README.md

Add this to your import section if you encounter PyTorch JIT compile errors.

```python
import torch
torch.jit._state.disable()
```

--------------------------------

### Print SECOND Demo Stderr

Source: https://github.com/mit-han-lab/torchsparse/blob/master/examples/openpcdet/demo.ipynb

Prints standard error from the SECOND model inference process.

```python
print(second_results.stderr)
```

--------------------------------

### TorchSparse++ Citation (MICRO 2023)

Source: https://github.com/mit-han-lab/torchsparse/blob/master/README.md

BibTeX entry for the TorchSparse++ paper presented at MICRO 2023. Use this for citing the latest version of the framework.

```bibtex
@inproceedings{tangandyang2023torchsparse,  
title={TorchSparse++: Efficient Training and Inference Framework for Sparse Convolution on GPUs},
  author={Tang, Haotian and Yang, Shang and Liu, Zhijian and Hong, Ke and Yu, Zhongming and Li, Xiuyu and Dai, Guohao and Wang, Yu and Han, Song},
  booktitle={IEEE/ACM International Symposium on Microarchitecture (MICRO)},
  year={2023}
}
```

--------------------------------

### SECOND Model Evaluation Results (AP40)

Source: https://github.com/mit-han-lab/torchsparse/blob/master/examples/mmdetection3d/demo.ipynb

Expected output showing AP40 evaluation results for the SECOND model on the Car class, including bounding box, Bird's Eye View (BEV), 3D, and Angle-of-View (AOS) metrics at different IoU thresholds.

```text
----------- AP40 Results ------------

Car AP40@0.70, 0.70, 0.70:
bbox AP40:97.4063, 92.4550, 89.2481
bev  AP40:92.6387, 88.4049, 85.2355
3d   AP40:90.4511, 81.3433, 76.1927
aos  AP40:97.13, 91.81, 88.42
Car AP40@0.70, 0.50, 0.50:
bbox AP40:97.4063, 92.4550, 89.2481
bev  AP40:97.5160, 94.7415, 91.7295
3d   AP40:97.3701, 94.5687, 91.4920
aos  AP40:97.13, 91.81, 88.42
```

--------------------------------

### SECOND Model Evaluation Results (AP11)

Source: https://github.com/mit-han-lab/torchsparse/blob/master/examples/mmdetection3d/demo.ipynb

Expected output showing AP11 evaluation results for the SECOND model on the Car class, including bounding box, Bird's Eye View (BEV), 3D, and Angle-of-View (AOS) metrics at different IoU thresholds.

```text
----------- AP11 Results ------------

Car AP11@0.70, 0.70, 0.70:
bbox AP11:95.2015, 89.6519, 88.0073
bev  AP11:89.9621, 87.2725, 84.2825
3d   AP11:88.3629, 78.2199, 76.0327
aos  AP11:94.94, 89.08, 87.23
Car AP11@0.70, 0.50, 0.50:
bbox AP11:95.2015, 89.6519, 88.0073
bev  AP11:95.3329, 89.9520, 88.7400
3d   AP11:95.2805, 89.8595, 88.5336
aos  AP11:94.94, 89.08, 87.23
```

--------------------------------

### CenterPoint Demo Output

Source: https://github.com/mit-han-lab/torchsparse/blob/master/examples/mmdetection3d/demo.ipynb

Prints the standard output from the CenterPoint demo, which includes evaluation metrics and per-class results.

```python
print(centerpoint_results.stdout)
```

--------------------------------

### TorchSparse Citation (MLSys 2022)

Source: https://github.com/mit-han-lab/torchsparse/blob/master/README.md

BibTeX entry for the TorchSparse paper presented at MLSys 2022. Use this for citing the initial release of the TorchSparse framework.

```bibtex
@inproceedings{tang2022torchsparse,
  title = {{TorchSparse: Efficient Point Cloud Inference Engine}},
  author = {Tang, Haotian and Liu, Zhijian and Li, Xiuyu and Lin, Yujun and Han, Song},
  booktitle = {Conference on Machine Learning and Systems (MLSys)},
  year = {2022}
}
```

--------------------------------

### Print PV-RCNN Demo Stderr

Source: https://github.com/mit-han-lab/torchsparse/blob/master/examples/openpcdet/demo.ipynb

Prints standard error from the PV-RCNN model inference process.

```python
print(pv_rcnn_results.stderr)
```

--------------------------------

### SPVNAS Paper Citation (ECCV 2020)

Source: https://github.com/mit-han-lab/torchsparse/blob/master/README.md

BibTeX entry for the SPVNAS paper at ECCV 2020, which includes the initial version of TorchSparse. This citation is for the foundational work.

```bibtex
@inproceedings{tang2020searching,
  title = {{Searching Efficient 3D Architectures with Sparse Point-Voxel Convolution}},
  author = {Tang, Haotian and Liu, Zhijian and Zhao, Shengyu and Lin, Yujun and Lin, Ji and Wang, Hanrui and Han, Song},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year = {2020}
}
```

--------------------------------

### CenterPoint Voxel 0.1 Circular NMS Configuration Options

Source: https://github.com/mit-han-lab/torchsparse/blob/master/examples/mmdetection3d/README.md

Specify configuration options for CenterPoint Voxel 0.1 Circular NMS to use TorchSparse layers. This involves updating the path in the dataset config and setting the points middle encoder type.

```bash
--cfg-options model.pts_middle_encoder.type=SparseEncoderTS
```

=== COMPLETE CONTENT === This response contains all available snippets from this library. No additional content exists. Do not make further requests.