KDE - method

Installation

In order to setup the environment, first create a virtual environement from the root of the project:

python -m venv .venv && source .venv/bin/activate

Then, in order to install the differente libraries, two requirements.txt files are at disposal depending on the version of CUDA that you are using. Whether you are on windows or linux, you can find the version of CUDA you are using with the command:

nvcc --version

Then, you can follow one of the following cases:

• if you have a version newer or equal to CUDA 11.8, you can try and directly install all the libraries by running the following:

pip install -r requirements_cuda12.txt

• if you have an older version or if for some reasons the first requirement file doesn't work, you can install the addequate version of pytorch and install the rest of the libraries by runnign the following:

pip install -r requirement_wo_torch.txt

Additional note:

if, during the inference, all samples are marked as failed, it might be due to an incompatible version of numpy with the one of pytorch installed. This issue can be solved by downgrading the version of numpy to the following: pip install numpy==1.26.4

Introduction

The concept of this model is based on Kernel Density Estimator. Each sample is mapped to a 3D normalized grid. This grid is then filtered through a process where a 3D kernel is matched to each point and discretized over a fixed number of neighbour cells. For each position in the grid, the values of all the kernels that overlap on it are then added up.

Continue	Discrete

Architecture

The architecture of this method is as following:

• data : this folder needs to be created since it is in the .gitignore file. The dataset for the training and grid-search can then be placed into it and will need to be referenced in the corresponding files.
• inference: This folder contains everything linked with inference (dataset, results)
• log: This folder contains all the results of the different trainings and grid-searchs. For each new one, a subfolder is created containing all the results. (logs, confusion matrix, best model, ...)
• src: This folder contains the different files for preprocessing, results visualization, dataset class, etc
• models: This folder contains the model implementation and a subfolder in which a .tar file can be placed to train from existing model or for inference.
• gridsearch.py, train.py, inference.py are the script to run for the corresponding tasks

dataset

Each dataset needs to be paced in the root of the folder ./data/. The structure of the dataset needs to be as following:

• dataset_name/
  • Single/
    • *.pcd
  • Multi/
    • *.pcd
  • Garbage/
    • *.pcd
  • modeltrees_shape_names.txt (name of the classes)

Important note: The file modeltrees_shape_names.txt needs to be at the root of each dataset. It can be found in ./inference/

Training

In order to train a model, the script train.py needs to be runned. The hyperparameters are all at the beggining of the file. If the dataset is new, the flag do_preprocess needs to be True. It will trigger the creation of the .csv files and the list of classes inside the foler of the dataset. If the dataset has not been trained on yet, the flage do_update_caching needs to be True. It will trigger the embedding of each .pcd file into a KDE grid saved into a .pickle file.

Grid-Search

The grid-search is done by running the gridsearch.py script. It will loop into a range of values for the kernel size and number of repetition of the kernel. All the hyperparameters can be specified at the beggining of the main() and the range of kernel size and number of repetition of the kernel are specified by the variables lst_kernel_sizes and lst_repeat_kernel.

Inference

In order to do inference, the samples to processed needs to be placed in the folder ./inference/<SRC_INF_DATA>/ (where SRC_INF_DATA is given in the HYPERPARAMETERS part of the inference.py script). The inference is then done by running the script inference.py. It will process each sample with the model saved in <SRC_MODEL>. The samples will be copied in the folder ./inference/results/ inside the folder corresponding to their respective class and a file results.csv will be created containing the class assignation per sample. Moreover, the failing samples will be stored in a corresponding folder and file failed_data.csv will be created in the results.