kwikerrr/point2wss

How kwikerrr/point2wss works

Point2WSS trains and evaluates a cross-modal masked autoencoder that reconstructs WSS roof parameters from LiDAR point clouds. Its data path loads paired cloud/label samples, preprocesses them and tokenizes roof geometry; training feeds both modalities to mutually conditioned branches, with reconstruction and alignment objectives. The graph also includes the documented evaluation entry point and the model’s key internal stages. Relationship evidence is limited to the README and sampled/indexed source references; unsampled wiring is not inferred.

Main entry points: pointcloud_branch_mae.py Point-cloud branch · point2wss_loss.py Multimodal losses · train.py Training entry point · tokenizer.py Multipitch tokenizer

Components 30

Data preparation

Split loader src/datamodules/loaders/txt_loader.py
Synthetic data module src/datamodules/synthetic_datamodule.py
Synthetic dataset src/datasets/synthetic_dataset.py
Grid dataset src/datasets/grid_synthetic_dataset.py
Dataset preprocessing src/datasets/base_dataset.py

Roof representation

Multipitch tokenizer src/tokenization/numerical/tokenizer.py
Sequence grammar src/tokenization/numerical/sequence_format.py
Token vocabulary src/tokenization/numerical/vocabulary.py
Numerical encoding src/tokenization/numerical/encoders.py

Cross-modal model

Point2WSS network src/nets/point2wss.py
Parameter branch src/nets/branches/params_branch.py
Parameter encoder src/nets/encoders/params_encoder.py
Parameter masker src/nets/maskers/params_masker.py
Parameter cross-encoder src/nets/cross_encoders/param_cross_encoder.py
Point-cloud branch src/nets/branches/pointcloud_branch_mae.py
Point-cloud encoder src/nets/encoders/pointcloud_encoder.py
Patch divider src/nets/pointcloud/patch_divider.py
Point-cloud geometry src/nets/pointcloud/geometry.py
Point-cloud masker src/nets/maskers/pointcloud_masker.py
Point-cloud cross-encoder src/nets/cross_encoders/pointcloud_cross_encoder.py
Attention layers src/nets/transformers/attention.py

Training objectives

Multimodal losses src/losses/point2wss_loss.py
Reconstruction loss src/losses/components/reconstruction.py
Chamfer loss src/losses/components/chamfer.py
Contrastive losses src/losses/components/contrastive.py
Cross-modal alignment src/losses/components/cross_contrastive.py

Train and evaluate

Lightning training module src/models/training_module.py
Training entry point src/train.py
Evaluation entry point src/eval.py

Other components

Researcher

Connections 34

Researcher → Training entry point: starts training src/train.py
Training entry point → Multipitch tokenizer: instantiates src/train.py
Training entry point → Synthetic data module: instantiates src/train.py
Training entry point → Lightning training module: instantiates src/train.py
Training entry point → Lightning training module: fits model src/train.py
Synthetic data module → Split loader: loads splits src/datamodules/synthetic_datamodule.py
Synthetic data module → Grid dataset: creates datasets
Synthetic data module → Synthetic dataset: creates datasets src/datamodules/synthetic_datamodule.py
Grid dataset → Dataset preprocessing: inherits preprocessing src/datasets/grid_synthetic_dataset.py
Synthetic dataset → Dataset preprocessing: uses base class src/datasets/synthetic_dataset.py
Grid dataset → Multipitch tokenizer: tokenizes labels src/datasets/grid_synthetic_dataset.py
Synthetic dataset → Multipitch tokenizer: tokenizes labels src/datasets/synthetic_dataset.py
Multipitch tokenizer → Sequence grammar: uses grammar src/tokenization/numerical/tokenizer.py
Multipitch tokenizer → Token vocabulary: uses token IDs src/tokenization/numerical/tokenizer.py
Sequence grammar → Numerical encoding: uses numeric ranks
Lightning training module → Point2WSS network: runs network
Point2WSS network → Parameter branch: dispatches parameters
Point2WSS network → Point-cloud branch: dispatches point cloud
Parameter branch → Parameter encoder: encodes tokens
Parameter branch → Parameter masker: masks tokens
Parameter branch → Parameter cross-encoder: cross-attends
Point-cloud branch → Patch divider: forms patches
Patch divider → Point-cloud geometry: uses sampling src/nets/pointcloud/patch_divider.py
Point-cloud branch → Point-cloud encoder: encodes patches
Point-cloud branch → Point-cloud masker: masks patches
Point-cloud branch → Point-cloud cross-encoder: cross-attends
Parameter cross-encoder → Attention layers: uses attention
Point-cloud cross-encoder → Attention layers: uses attention
Lightning training module → Multimodal losses: computes objectives
Multimodal losses → Reconstruction loss: uses reconstruction
Multimodal losses → Chamfer loss: uses geometry loss
Multimodal losses → Contrastive losses: uses alignment
Multimodal losses → Cross-modal alignment: uses cross alignment
Researcher → Evaluation entry point: starts evaluation

Drawn by GitDiagram on November 11, 2026 from the repository's file tree, README and source files. Read as Markdown ·