models/official/cv/PointNet2

• Contents
• PointNet2 Description
• Model Architecture
• Dataset
• Environment Requirements
• Quick Start
• Script Description
• Script and Sample Code
• Script Parameters
• Training Process
• • Training
• Evaluation Process
• • Evaluation
• Model Description
• • Performance
  • Training Performance
  • Inference Performance
• Description of Random Situation
• ModelZoo Homepage

Contents

• Contents
• PointNet2 Description
• Model Architecture
• Dataset
• Environment Requirements
• Quick Start
• Script Description
• Script and Sample Code
• Script Parameters
• Training Process
  • Training
• Evaluation Process
  • Evaluation
• Model Description
  • Performance
  • Training Performance
  • Inference Performance
• Description of Random Situation
• ModelZoo Homepage

PointNet2 Description

PointNet++ was proposed in 2017, it is a hierarchical neural network that applies PointNet recursively on a nested
partitioning of the input point set. By exploiting metric space distances, this network is able to learn local features
with increasing contextual scales. Experiments show that our network called PointNet++ is able to learn deep point set
features efficiently and robustly.

Paper: Qi, Charles R., et al. "Pointnet++: Deep hierarchical feature learning on
point sets in a metric space." arXiv preprint arXiv:1706.02413 (2017).

Model Architecture

The hierarchical structure of PointNet++ is composed by a number of set abstraction levels. At each level, a set of
points is processed and abstracted to produce a new set with fewer elements. The set abstraction level is made of three
key layers: Sampling layer, Grouping layer and PointNet layer. The Sampling layer selects a set of points from
input points, which defines the centroids of local regions. Grouping layer then constructs local region sets by
finding “neighboring” points around the centroids. PointNet layer uses a mini-PointNet to encode local region
patterns into feature vectors.

Dataset

Dataset used: alignment ModelNet40

• Dataset size：6.48G，Each point cloud contains 2048 points uniformly sampled from a shape surface. Each cloud is
  zero-mean and normalized into a unit sphere.
  • Train：5.18G, 9843 point clouds
  • Test：1.3G, 2468 point clouds
• Data format：txt files
  • Note：Data will be processed in src/dataset.py

Environment Requirements

• Hardware（Ascend）
  • Prepare hardware environment with Ascend processor.
• Framework
  • MindSpore
• For more information, please check the resources below：
  • MindSpore Tutorials
  • MindSpore Python API

Quick Start

After installing MindSpore via the official website, you can start training and evaluation as follows:

# Run stand-alone training
bash scripts/run_standalone_train.sh [DATA_PATH] [SAVE_DIR] [PRETRAINDE_CKPT(optional)]
# example:
bash scripts/run_standalone_train.sh modelnet40_normal_resampled save pointnet2.ckpt

# Run distributed training
bash scripts/run_distributed_train.sh [RANK_TABLE_FILE] [DATA_PATH] [SAVE_DIR] [PRETRAINDE_CKPT(optional)]
# example:
bash scripts/run_distributed_train.sh hccl_8p_01234567_127.0.0.1.json modelnet40_normal_resampled save pointnet2.ckpt

# Evaluate
bash scripts/run_eval.sh [DATA_PATH] [CKPT_NAME]
# example:
bash scripts/run_eval.sh modelnet40_normal_resampled pointnet2.ckpt

if you want to run this model on GPU

# Run stand-alone training
bash scripts/run_standalone_train_gpu.sh [DATA_PATH] [SAVE_DIR] [PRETRAINDE_CKPT(optional)]
# example:
bash scripts/run_standalone_train_gpu.sh modelnet40_normal_resampled save pointnet2.ckpt

# Run distributed training
bash scripts/run_distributed_train_gpu.sh [DEVICE_NUM] [DATA_PATH] [SAVE_DIR] [PRETRAINDE_CKPT(optional)]
# example:
bash scripts/run_distributed_train_gpu.sh 8 modelnet40_normal_resampled save pointnet2.ckpt

# Evaluate
bash scripts/run_eval_gpu.sh [DATA_PATH] [CKPT_NAME]
# example:
bash scripts/run_eval_gpu.sh modelnet40_normal_resampled pointnet2.ckpt

Script Description

Script and Sample Code

├── PointNet2
    ├── eval.py                             # eval net
    ├── export.py
    ├── README.md
    ├── requirements.txt
    ├── scripts
    │   ├── run_distributed_train_gpu.sh    # launch distributed training with GPU platform (8p)
    │   ├── run_distributed_train.sh        # launch distributed training with ascend platform (8p)
    │   ├── run_eval_gpu.sh                 # launch evaluating with GPU platform
    │   ├── run_eval.sh                     # launch evaluating with ascend platform
    │   ├── run_standalone_train_gpu.sh     # launch standalone training with GPU platform (1p)
    │   └── run_standalone_train.sh         # launch standalone training with ascend platform (1p)
    ├── src
    │   ├── callbacks.py                    # callbacks definition
    │   ├── dataset.py                      # data preprocessing
    │   ├── layers.py                       # network layers initialization
    │   ├── lr_scheduler.py                 # learning rate scheduler
    │   ├── pointnet2.py                    # network definition
    │   ├── pointnet2_utils.py              # network definition utils
    │   ├── provider.py                     # data preprocessing for training
    │
    └── train.py                            # train net

Script Parameters

Major parameters in train.py are as follows:
--batch_size        # Training batch size.
--epoch             # Total training epochs.
--learning_rate     # Training learning rate.
--optimizer         # Optimizer for training. Optional values are "Adam", "SGD".
--data_path         # The path to the train and evaluation datasets.
--loss_per_epoch    # The times to print loss value per epoch.
--save_dir          # The path to save files generated during training.
--use_normals       # Whether to use normals data in training.
--pretrained_ckpt   # The file path to load checkpoint.
--enable_modelarts         # Whether to use modelarts.

Training Process

Training

• running on Ascend

# Run stand-alone training
bash scripts/run_standalone_train.sh [DATA_PATH] [SAVE_DIR] [PRETRAINDE_CKPT(optional)]
# example:
bash scripts/run_standalone_train.sh modelnet40_normal_resampled save pointnet2.ckpt

# Run distributed training
bash scripts/run_distributed_train.sh [RANK_TABLE_FILE] [DATA_PATH] [SAVE_DIR] [PRETRAINDE_CKPT(optional)]
# example:
bash scripts/run_distributed_train.sh hccl_8p_01234567_127.0.0.1.json modelnet40_normal_resampled save pointnet2.ckpt

• running on GPU

  # Run stand-alone training
  bash scripts/run_standalone_train_gpu.sh [DATA_PATH] [SAVE_DIR] [PRETRAINDE_CKPT(optional)]
  # example:
  bash scripts/run_standalone_train_gpu.sh modelnet40_normal_resampled save pointnet2.ckpt
  
  # Run distributed training
  bash scripts/run_distributed_train_gpu.sh [DEVICE_NUM] [DATA_PATH] [SAVE_DIR] [PRETRAINDE_CKPT(optional)]
  # example:
  bash scripts/run_distributed_train_gpu.sh 8 modelnet40_normal_resampled save pointnet2.ckpt
  

After training, the loss value will be achieved as follows:

# train log
epoch: 1 step: 410, loss is 1.4731973
epoch time: 704454.051 ms, per step time: 1718.181 ms
epoch: 2 step: 410, loss is 1.0621885
epoch time: 471478.224 ms, per step time: 1149.947 ms
epoch: 3 step: 410, loss is 1.176581
epoch time: 471530.000 ms, per step time: 1150.073 ms
epoch: 4 step: 410, loss is 1.0118457
epoch time: 471498.514 ms, per step time: 1149.996 ms
epoch: 5 step: 410, loss is 0.47454038
epoch time: 471535.602 ms, per step time: 1150.087 ms
...

The model checkpoint will be saved in the 'SAVE_DIR' directory.

Evaluation Process

Evaluation

Before running the command below, please check the checkpoint path used for evaluation.

• running on Ascend

  # Evaluate
  bash scripts/run_eval.sh [DATA_PATH] [CKPT_NAME]
  # example:
  bash scripts/run_eval.sh modelnet40_normal_resampled pointnet2.ckpt
  

• running on GPU

  # Evaluate
  bash scripts/run_eval_gpu.sh [DATA_PATH] [CKPT_NAME]
  # example:
  bash scripts/run_eval_gpu.sh modelnet40_normal_resampled pointnet2.ckpt
  

You can view the results through the file "eval.log". The accuracy of the test dataset will be as follows:

# grep "Accuracy: " eval.log
'Accuracy': 0.916

Model Description

Performance

Training Performance

Parameters	Ascend	GPU
Model Version	PointNet++	PointNet++
Resource	Ascend 910; CPU 24cores; Memory 256G; OS Euler2.8	RTX 3090; GPU Memory 24268MiB; Ubuntu
uploaded Date	08/31/2021 (month/day/year)	12/20/2021 (month/day/year)
MindSpore Version	1.3.0	1.5.0rc1
Dataset	ModelNet40	ModelNet40
Training Parameters	epoch=200, steps=82000, batch_size=24, lr=0.001	epoch=200, steps=82000, batch_size=24, lr=0.001
Optimizer	Adam	Adam
Loss Function	NLLLoss	NLLLoss
outputs	probability	probability
Loss	0.01	0.01
Speed	1.2 s/step (1p)	390 ms/step (1p)
Total time	27.3 h (1p)	8h8m6s (1p)
Checkpoint for Fine tuning	17 MB (.ckpt file)	17 MB (.ckpt file)

Inference Performance

Parameters	Ascend	GPU
Model Version	PointNet++	PointNet++
Resource	Ascend 910; CPU 24cores; Memory 256G; OS Euler2.8	RTX 3090; GPU Memory 24268MiB; Ubuntu
Uploaded Date	08/31/2021 (month/day/year)	12/20/2021 (month/day/year)
MindSpore Version	1.3.0	1.3.0
Dataset	ModelNet40	ModelNet40
Batch_size	24	24
Outputs	probability	probability
Accuracy	91.5% (1p)	91.42% (1p)
Total time	2.5 min	1.13 min

Description of Random Situation

We use random seed in dataset.py, provider.py and pointnet2_utils.py.

ModelZoo Homepage

Please check the official homepage.