NotAllNeighborsMatter

Environments

- gcc-7
- cuda 11.2

Setup

conda create -n NotAllNeighborsMatter python=3.8
conda activate NotAllNeighborsMatter
pip install torch==1.8.1+cu111 torchvision==0.9.1+cu111 torchaudio==0.8.1 -f <https://download.pytorch.org/whl/torch_stable.html>
pip install ninja
cd s3dis_scannet
pip install -r requirements.txt
pip install protobuf==3.20.0
pip install scikit-learn==1.0.2
cd ../libraries/MinkowskiEngine
python setup.py install
cd ../libraries/torchsparse
python setup.py install
cd ../libraries/cumm
pip install --editable .
pip install ccimport==0.3.7
python setup.py develop
cd ../libraries/spconv
python setup.py develop
python (Open python prompt)
import spconv (This will build cumm and spconv)

S3DIS

Prepare Dataset (Brought from here)

1. Download the stanford 3d dataset (Stanford3dDataset_v1.2.zip) from the website to ./s3dis_scannet/dataset/s3dis
2. Preprocess
   • Modify the input and output directory accordingly in lib/datasets/preprocessing/stanford.py
   • And run python -m lib.datasets.preprocessing.stanford
3. Train ./scripts/train_stanford.sh 0 "-default" "--stanford3d_path ./dataset/s3dis/preprocessing"

Train

To train baseline model from scratch, run below command.

We also provide checkpoints of baseline models in ./checkpoints.

LOG_DIR=PATH_TO_OUTPUT_DIR ./scripts/train_stanford.sh $GPU_ID --backend=[spconv, mink, torchsparse]

• If you would like to resume training from the latest checkpoint, add -resume=PATH_TO_OUTPUT_DIR following the train command

Retrain

For given prune_edge configuration, run below command to retrain baseline checkpoint to recover the accuracy loss.

./scripts/retrain_stanford.sh $GPU_ID --backend=[spconv, mink, torchsparse]  --weights=PATH_TO_BASELINE_CKPT --prune_edge "{'0':0,'1':0,'2':0,'3':0,'4':0,'5':0,'6':0,'7':0}"

• e.g., To retrain pruned baseline model with --prune_edge "{'0':4,'1':3,'2':2,'3':2,'4':1,'5':0,'6':0,'7':0}" with spconv backend, run

  ./scripts/retrain_scannet.sh 0 --backend=spconv --weights=PATH_TO_BASELINE_CKPT --prune_edge "{'0':4,'1':3,'2':2,'3':2,'4':1,'5':0,'6':0,'7':0}"
  

Test

./scripts/test_stanford.sh $GPU_ID --backend=[spconv, mink, torchsparse] --weights=PATH_TO_CKPT --prune_edge "{'0':0,'1':0,'2':0,'3':0,'4':0,'5':0,'6':0,'7':0}"

• To run model with half precision, add -fp16 following the test command

ScannetV2

Prepare Dataset (Brought from here)

1. Download the ScanNet dataset from the official website to ./dataset/scannet. You need to sign the terms of use.
2. Next, preprocess all scannet raw point cloud with the following command after you set the path correctly. python -m lib.datasets.preprocessing.scannet
3. Train the network with export BATCH_SIZE=N; ./scripts/train_scannet.sh 0 -default "--scannet_path ./dataset/scannet/preprocessing"
   • Modify the BATCH_SIZE accordingly.
   • The first argument is the GPU id and the second argument is the path postfix and the last argument is the miscellaneous arguments.

Train

To train baseline model from scratch, run below command.

We also provide checkpoints of baseline models in ./checkpoints.

./scripts/train_scannet.sh $GPU_ID --backend=[spconv, mink, torchsparse] PATH_TO_OUTPUT_DIR

• If you would like to resume training from the latest checkpoint, add -resume PATH_TO_OUTPUT_DIR following the train command

Retrain

For given prune_edge configuration, run below command to retrain baseline checkpoint to recover the accuracy loss.

./scripts/retrain_scannet.sh $GPU_ID --backend=[spconv, mink, torchsparse] --weights=PATH_TO_BASELINE_CKPT --prune_edge "{'0':0,'1':0,'2':0,'3':0,'4':0,'5':0,'6':0,'7':0}"

• e.g., To retrain pruned baseline model with --prune_edge "{'0':4,'1':3,'2':2,'3':2,'4':1,'5':0,'6':0,'7':0}" with spconv backend, run

  ./scripts/retrain_scannet.sh 0 --backend=spconv --weights=PATH_TO_BASELINE_CKPT --prune_edge "{'0':4,'1':3,'2':2,'3':2,'4':1,'5':0,'6':0,'7':0}"
  

Test

./scripts/test_scannet.sh $GPU_ID --backend=[spconv, mink, torchsparse] --weights=PATH_TO_CKPT --prune_edge "{'0':0,'1':0,'2':0,'3':0,'4':0,'5':0,'6':0,'7':0}"

• To run model with half precision, add -fp16 following the test command

Checkpoints

Checkpoints are uploaded in ./checkpoints.

Dataset	Library	Baseline	Lossless	Lossy
S3DIS	Spconv	63.107	63.464	62.052
S3DIS	MinkowskiEngine	63.547	63.471	62.801
ScannetV2	Spconv	72.354	72.413	71.461
ScannetV2	MinkowskiEngine	72.341	72.108	71.340