PointCloudNet

This repository is for PointCloudNet, a method designed to learn from high-dimensional point cloud data using multiple graph embeddings and graph wavelet transforms.

Overview

The provided Python script trains PointCloudNet on a given dataset of point features and associated labels. It uses:

• PyTorch for model definition and training.
• wandb (Weights & Biases) for experiment tracking.
• scikit-learn for data splitting.
• tqdm for progress visualization.

The script:

1. Loads and processes your input data.
2. Initializes the model and MLP classifier.
3. Trains the model, logging training progress and metrics to wandb.
4. Saves the best model weights for reproducibility.

Requirements

• Python 3.7 or later
• PyTorch (compatible with CUDA if GPU training is desired)
• NumPy
• scikit-learn
• tqdm
• wandb

Install all requirements using:

pip install torch numpy scikit-learn tqdm wandb

Arguments

You can specify various arguments to customize training:

• --raw_dir (str): Directory containing the raw data. Default: melanoma_data_full
• --full (flag): If provided, may indicate the use of a full dataset variant.
• --num_weights (int): Number of weights (features dimensions) to learn. Default: 2
• --threshold (float): Threshold used for graph creation. Default: 5e-5
• --hidden_dim (int): Hidden dimension size for the MLP. Default: 50
• --num_layers (int): Number of MLP layers. Default: 3
• --lr (float): Learning rate. Default: 0.03
• --wd (float): Weight decay. Default: 3e-3
• --num_epochs (int): Number of training epochs. Default: 100
• --batch_size (int): Batch size for training. Default: 128
• --gpu (int): GPU index to use. Set to -1 for CPU-only. Default: 0

Running the Script

Before running, ensure that raw_dir points to a directory containing compatible data files. The data loading and preparation code is assumed to be handled within the PointCloudFeatLearning class. Consult that class for specifics on required data format.

Run the script:

python train_pointcloudnet.py --raw_dir path_to_data --num_weights 2 --threshold 0.00005 --gpu 0

Adjust parameters as needed. For example:

• To train on CPU:

  python train_pointcloudnet.py --gpu -1

• To change the learning rate and number of epochs:

  python train_pointcloudnet.py --lr 0.01 --num_epochs 200

Weights & Biases Integration

The script automatically logs metrics to Weights & Biases if you have an account and have run wandb login locally. If you do not want to use wandb, remove or comment out the wandb lines in the code.

Outputs

• Model Checkpoints: The best performing model weights will be saved as bestalpha_{num_weights} and bestmlp_{num_weights}.
• Alpha Weights: The learned alpha weights for feature importance are saved as bestweights_{num_weights}.pt.

These files can be used to reproduce results or for downstream analysis.