Add dockerized installation (pytorch 1.10.1, CUDA 11.3.1, ubuntu 20.04)

Dockerfile

@@ -0,0 +1,33 @@
+FROM cnstark/pytorch:1.10.2-py3.9.12-cuda11.3.1-devel-ubuntu20.04
+
+RUN apt update && apt upgrade -y &&\
+ DEBIAN_FRONTEND=noninteractive apt install -y --no-install-recommends \
+ tzdata git libsparsehash-dev unzip wget vim tmux ffmpeg libsm6 libxext6
+
+WORKDIR /home/pcseg
+COPY ./ .
+
+# package installation
+ENV PATH "/usr/.local/bin:${PATH}"
+RUN pip3 install -r requirements.txt
+RUN cd /home/pcseg/package &&\
+ mkdir torchsparse_dir/ &&\
+ unzip -o sparsehash.zip -d sparsehash &&\
+ unzip -o torchsparse.zip &&\
+ unzip -o range_lib.zip
+RUN cd /home/pcseg/package/sparsehash/sparsehash-master &&\
+ ./configure --prefix=/home/pcseg/package/torchsparse_dir/spash &&\
+ make &&\
+ make install
+RUN pip3 install -e package/range_lib &&\
+ pip3 install -e package/torchsparse
+
+
+# filesystem and permissions setup
+ARG UNAME=pcseg
+ARG UID=1000
+ARG GID=1000
+RUN groupadd -g $GID -o $UNAME &&\
+ useradd -m -u $UID -g $GID -o -s /bin/bash $UNAME &&\
+ chown -R $UNAME:$UNAME /home/pcseg
+USER $UNAME

dist_train.sh

@@ -2,7 +2,7 @@
 
 set -x
 NGPUS=$1
-PY_ARGS=${@:2}
+PY_ARGS=${*:2}
 
 while true
 do
@@ -16,7 +16,5 @@ echo $PORT
 
 python -m torch.distributed.launch --nproc_per_node=${NGPUS} \
  --rdzv_endpoint=localhost:${PORT} train.py --launcher pytorch --amp \
- ${PY_ARGS}
-
-echo 'dist_train finished!'
+ ${PY_ARGS} && echo 'dist_train finished!'
 

docs/GET_STARTED.md

@@ -14,7 +14,6 @@ CUDA_VISIBLE_DEVICES=4,5 sh dist_train.sh 2 --cfg_file tools/cfgs/fusion/semanti
 CUDA_VISIBLE_DEVICES=6,7 sh dist_train.sh 2 --cfg_file tools/cfgs/fusion/semantic_kitti/rpvnet_mk18_cr10.yaml
 ```
 
-
 ### Waymo Open Dataset
 
 For example, if you want to train the following models with 2 GPUs:
@@ -24,3 +23,15 @@ CUDA_VISIBLE_DEVICES=0,1 sh dist_train.sh 2 --cfg_file tools/cfgs/voxel/waymo/mi
 
 CUDA_VISIBLE_DEVICES=2,3 sh dist_train.sh 2 --cfg_file tools/cfgs/voxel/waymo/cylinder_cy480_cr10.yaml
 ```
+
+### Docker
+
+#### Docker
+
+After building according to [installation instructions](./INSTALL.md), run the container while mounting the dataset, e.g. SemanticKITTI:
+
+```bash
+docker run -it --rm --gpus all -v /path/to/logs:/home/pcseg -v /path/to/kitti/semanticKITTI/dataset:/home/pcseg/data_root/SemanticKITTI/ pcseg
+```
+
+and follow the train instructions above

docs/INSTALL.md

@@ -2,8 +2,8 @@
 
 ### General Requirements
 
-This codebase is tested with `torch==1.10.0` and `torchvision==0.11.0`, with `CUDA 11.3` and `gcc 7.3.0`. In order to successfully reproduce the results reported in our paper, we recommend you to follow the exact same configuation with us. However, similar versions that came out lately should be good as well.
-
+* This codebase is tested with `torch==1.10.0` and `torchvision==0.11.0`, with `CUDA 11.3` and `gcc 7.3.0`. In order to successfully reproduce the results reported in our paper, we recommend you to follow the exact same configuation with us. However, similar versions that came out lately should be good as well.
+* OS with `ffmpeg libsm6 libxext6 libsparsehash-dev` installed
 
 ### Step 1: Create Enviroment
 ```Shell
@@ -90,7 +90,7 @@ Finished processing dependencies for rangelib==1.0.0
 ```
 #### 4.5 - Other Packages
 ```Shell
-pip install pyyaml easydict numba torchpack strictyaml llvmlite easydict scikit-image tqdm SharedArray prettytable opencv-python
+pip install pyyaml easydict numba torchpack strictyaml llvmlite easydict scikit-image tqdm SharedArray prettytable opencv-python tensorboard
 ```
 ```Shell
 pip uninstall setuptools
@@ -105,3 +105,20 @@ Inside `PCSeg` directory:
 python setup.py develop
 ```
 
+## Docker
+
+### Prerequisites
+
+Make sure you have applied the following steps before building the docker image:
+
+* Install `nvidia-docker`: `sudo apt-get install -y nvidia-container-toolkit`
+* Configure docker runtime, according to [this answer](https://stackoverflow.com/a/61737404) while [this addition](https://stackoverflow.com/a/75629058) may be necessary as of 03/2023
+* Restart docker: `sudo systemctl restart docker`
+
+### Build docker image
+
+```Shell
+docker build --build-arg USER_UID=$(id -u) --build-arg USER_GID=$(id -g) -t pcseg -f Dockerfile .
+```
+
+**Note:** `DOCKER_BUILDKIT=0 docker build ...` may be necessary. See Prerequisites

pcseg/data/dataset/semantickitti/laserscan.py

@@ -299,11 +299,11 @@ def reset(self):
 
  # projection color with semantic labels
  self.proj_sem_label = np.zeros((self.proj_H, self.proj_W), dtype=np.int32) # [H, W]: label
- self.proj_sem_color = np.zeros((self.proj_H, self.proj_W, 3), dtype=np.float) # [H, W, 3]: color
+ self.proj_sem_color = np.zeros((self.proj_H, self.proj_W, 3), dtype=np.float32) # [H, W, 3]: color
 
  # projection color with instance labels
  self.proj_inst_label = np.zeros((self.proj_H, self.proj_W), dtype=np.int32) # [H, W]: label
- self.proj_inst_color = np.zeros((self.proj_H, self.proj_W, 3), dtype=np.float) # [H, W, 3]: color
+ self.proj_inst_color = np.zeros((self.proj_H, self.proj_W, 3), dtype=np.float32) # [H, W, 3]: color
 
 
  def open_label(self, filename):

pcseg/data/dataset/semantickitti/semantickitti_rv.py

@@ -3,6 +3,7 @@
 import glob
 import random
 import yaml
+from itertools import chain
 
 import numpy as np
 import torch
@@ -80,15 +81,26 @@ def __init__(
  if self.split == 'train': folders = ['00', '01', '02', '03', '04', '05', '06', '07', '09', '10']
  elif self.split == 'val': folders = ['08']
  elif self.split == 'test': folders = ['11', '12', '13', '14', '15', '16', '17', '18', '19', '20', '21']
-
- self.lidar_list = []
- for folder in folders:
- self.lidar_list += glob.glob(self.root + 'sequences/' + folder + '/velodyne/*.bin') 
- print("Loading '{}' samples from SemanticKITTI under '{}' split".format(len(self.lidar_list), self.split))
+
+ if not os.path.exists(self.root):
+ raise ValueError(f"SemanticKITTI root dir {self.root} doesn't exist!")
+
+ if not os.path.exists(self.root + "sequences"):
+ raise ValueError("Please use SemanticKITTI root directory, e.g. /path/to/SemanticKITTI/dataset")
+
+ self.lidar_list = list(
+ chain.from_iterable(glob.glob(self.root + 'sequences/' + f + '/velodyne/*.bin') for f in folders))
+ if len(self.lidar_list) == 0:
+ raise ValueError(f"Coudn't read pointclouds. Found {len(self.lidar_list)} clouds")
 
  self.label_list = [i.replace("velodyne", "labels") for i in self.lidar_list]
  self.label_list = [i.replace("bin", "label") for i in self.label_list]
 
+ if not any([os.path.exists(path) for path in self.label_list]):
+ raise ValueError("Couldn't find matching labels")
+
+ print("Loading {} samples from SemanticKITTI under {} split".format(len(self.lidar_list), self.split))
+
  if self.split == 'train_test':
  root_psuedo_labels = '/mnt/lustre/konglingdong/data/sets/sequences/'
  folders_test = ['11', '12', '13', '14', '15', '16', '17', '18', '19', '20', '21']

pcseg/model/segmentor/fusion/rpvnet/rpvnet.py

@@ -23,7 +23,7 @@
 from pcseg.loss import Losses
 
 
-import range_utils.nn.functional as rnf
+from .range_lib.range_utils.nn import functional as rnf
 
 
 __all__ = ['RPVNet', 'SalsaNext']

requirements.txt

@@ -0,0 +1,13 @@
+torch-scatter==2.1.1
+tensorboard==2.12.3
+opencv_python==4.7.0.72
+pyyaml
+easydict
+numba
+strictyaml
+llvmlite
+scikit-image
+tqdm
+sharedarray
+prettytable
+setuptools==59.5.0