Efficient Deep Learning for Point Clouds

This project is about designing efficient point cloud Deep Neural Networks with pure algorithm (software-level) optimizations. We propose a technique named Delayed-Aggregation, which:

1. reduces redundant computation to achieve workload efficiency;
2. exposes parallelism that can be easily captured by the underlying hardware.

For the background of point cloud neural networks and how our delayed-aggregation helps improves the execution efficiency, see the wiki page.

Networks

Delayed-aggregation applies to a wide range of different point cloud networks. This repo has the implementation for the following five networks:

• PointNet++: Classification, Segmentation
• DGCNN: Classification, Segmentation
• LDGCNN: Classification
• F-PointNet: 3D Detection
• DensePoint: Classification

For each network, we have provided three versions:

1. Baseline: the original networks with implementation optimizations.
2. Limited Delayed-Aggregation: the one with limited delayed-aggregation optimization, which is inspired by some GNNs implementations.
3. Fully Delayed-Aggregation: the one with full delayed-aggregation optimization, i.e., our proposed technique.

For the difference between the three versions, again see the wiki page.

How to Run

We have created a simple PYTHON script to navigate the repository. Run:

$ python launcher.py -h

You will see:

usage: launcher.py [-h] [--compile COMPILE] [--download DOWNLOAD]
                   [--list_models LIST_MODELS] [--run RUN] [--train TRAIN]
                   [--use_baseline USE_BASELINE] [--use_limited USE_LIMITED]
                   [--segmentation SEGMENTATION]

optional arguments:
  -h, --help            show this help message and exit
  --compile COMPILE     Compile libraries in the models, to compile a specific
                        network, use: --compile [NETWORK_NAME] or to compile
                        all models using, --compile all
  --download DOWNLOAD   Download the specific dataset for the models, to
                        download a dataset for a specific network, use:
                        --download [NETWORK_NAME] or to download all datasets
                        using, --download all
  --list_models LIST_MODELS
                        List all model names.
  --run RUN             Evaluate the model with Fully Delayed-Aggregation.
  --train TRAIN         Train the model with Fully Delayed-Aggregation.
  --use_baseline USE_BASELINE
                        Use the baseline without any kind of Delayed-Aggregation.
  --use_limited USE_LIMITED
                        Use Limited Delayed-Aggregation.
  --segmentation SEGMENTATION
                        Execute the segmentation version.

There is a slight naming difference between the actual model name and the name in the code. Make sure you use names in the second column of this table to run the launcher.py.

Actual Model Name	Name in Our Code
PointNet++	pointnet2
DGCNN	dgcnn
LDGCNN	ldgcnn
F-PointNet	frustum-pointnets
DensePoint	DensePoint

Dataset

Datasets shared by multiple networks are placed in the Datasets directory, e.g., ModelNet40 and ShapeNet.
Datasets exclusively used by a network are placed in its directory, e.g., KITTI for F-PointNet.
Use the following command to download dataset:

$ python launcher.py --download [NETWORK]

• Specify [NETWORK] to the name of a network to download the corresponding dataset or all to download all the datasets.
• Add --segmentation True to download segmentation data for pointnet++ and dgcnn.

---

* Make sure to activate the correct environment for each network before running any of the following commands.

---

Compile Customized Operators

Some networks are native Python code and do not need to compile. Others such as pointnet++, f-pointnet, and DensePoint have customized modules that need to be compiled.
To compile, run:

$ python launcher.py --compile [NETWORK]

• [NETWORK] can be pointnet2 (for pointnet++), frustum-pointnets (for f-pointnet), DensePoint (for DensePoint), or simply compile all by using all.
• Please check out the instructions in each network to modify the CUDA_PATH if you encounter any compiling issues.

Training

To train the Baseline version, add flag --use_baseline True:

$ python launcher.py --train [NETWORK] --use_baseline True

To train the Limited Delayed-Aggregation version, add flag --use_limited True:

$ python launcher.py --train [NETWORK] --use_limited True

To train the Fully Delayed-Aggregation version:

$ python launcher.py --train [NETWORK]

To train the segmentation model of pointnet++ and dgcnn, add flag --segmentation True to the above commands:

$ python launcher.py --train [NETWORK] --segmentation True

Evaluation

To evaluate the Baseline version, add flag --use_baseline True:

$ python launcher.py --run [NETWORK] --use_baseline True

To evaluate the Limited Delayed-Aggregation version, add flag --use_limited True:

$ python launcher.py --run [NETWORK] --use_limited True

To evaluate the Fully Delayed-Aggregation version:

$ python launcher.py --run [NETWORK]

To evaluate the segmentation model of pointnet++ and dgcnn, add flag --segmentation True to the above commands:

$ python launcher.py --run [NETWORK] --segmentation True

Publication

This project contains the artifact for our paper Mesorasi: Architecture Support for Point Cloud Analytics via Delayed-Aggregation (MICRO 2020).

@inproceedings{feng2020mesorasi,
  title={Mesorasi: Architecture Support for Point Cloud Analytics via Delayed-Aggregation},
  author={Feng, Yu and Tian, Boyuan and Xu, Tiancheng and Whatmough, Paul and Zhu, Yuhao},
  booktitle={Proceedings of the 53th International Symposium on Microarchitecture},
  year={2020},
  organization={ACM}
}