Canonization Experiments

This repository contains two sets of experiments:

1. ModelNet
2. rotatedMNIST

Setup

Clone the repository:

git clone https://github.com/yonatansverdlov/Canonization.git
cd Canonization

Create the conda environment:

conda create -n canon python=3.10 pip -y

Activate it:

conda activate canon

Install PyTorch with CUDA 12.1:

python -m pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121

Install the remaining required packages:

python -m pip install torch-geometric lightning kornia numpy scipy hilbertcurve h5py tqdm 

Install space-filling-pytorch from source:

cd /tmp
git clone https://github.com/Kitsunetic/space-filling-pytorch.git
cd space-filling-pytorch
python -m pip install -e .
cd ~/Canonization

Experiments

ModelNet

The ModelNet experiments are divided into two parts:

1. Covering number experiments
2. Training experiments

Covering Number Experiments

First create the data.

From the repository root, run:

cd ModelNet/data_creation
python create_data.py --dataset_name 10 --P 256

For ModelNet40, run:

python create_data.py --dataset_name 40 --P 256

Here:

dataset_name = 10 or 40
P = number of sampled points per shape

Then run the distance computation.

From the repository root, run:

cd ../compute_distances
python compute_distance.py --dataset_name 10 --P 256

For ModelNet40, run:

python compute_distance.py --dataset_name 40 --P 256

The script compares:

regular distance
group distance
sort distance
Hilbert distance

Training Experiments

The training code uses the HDF5 ModelNet40 point-cloud dataset.

Download the data from Kaggle:

ModelNet40 Kaggle Dataset

Put the extracted data here:

ModelNet/training/data/

To create ModelNet10 from ModelNet40, run:

cd ModelNet/training
python create_h5_modelnet10_from_modelnet40.py

In order to train ModelNet10:

python train.py --dataset modelnet10 --ordering hilbert --run_5_seeds true
python train.py --dataset modelnet10 --ordering lex --run_5_seeds true
python train.py --dataset modelnet10 --ordering ply --run_5_seeds true

In order to train ModelNet40:

python train.py --dataset modelnet40 --ordering hilbert --run_5_seeds true
python train.py --dataset modelnet40 --ordering lex --run_5_seeds true
python train.py --dataset modelnet40 --ordering ply --run_5_seeds true

The training script reports train/test accuracy mean/std over all seeds.

Rotation / Frame Canonization Training

The models are:

1  Pure PCA
2  Frame Averaging
3  Skewness Canonization
4  Random Frame

Run from:

cd ModelNet/training

Single seed:

python train_rot.py --model PurePCA
python train_rot.py --model FrameAveraging
python train_rot.py --model Skewness
python train_rot.py --model RandomFrame

Five seeds:

python train_rot.py --model PurePCA --run_5_seeds true
python train_rot.py --model FrameAveraging --run_5_seeds true
python train_rot.py --model Skewness --run_5_seeds true
python train_rot.py --model RandomFrame --run_5_seeds true

rotatedMNIST

The rotatedMNIST experiments contain two scripts:

1. Training
2. Distance computation

Training

From the repository root, run:

cd mnist
python train.py --model_type cnn

The available model types are:

cnn
average
learned_can

By default, the script runs with one seed.

To reproduce the paper setting, use five seeds:

python train.py --model_type cnn --num_seeds 5
python train.py --model_type average --num_seeds 5
python train.py --model_type learned_can --num_seeds 5

Distance Computation

From the repository root, run:

cd mnist
python compute_distance.py