gnp.dataset module

gnp.dataset.patch

class gnp.dataset.patch.PatchGenerator(data: dict, graph_radius: float, batch_size: int = 1, center: str = 'tree', shuffle_patches: bool = True, knn: int = 50, min_radius: float = 0.01, pca: bool = True, degree: int = 3, orientation: Tensor | None = None, min_z_scale: float = 0.001, max_num_neighbors: int = 32, device: str = 'cpu')

Bases: object

PatchLoader class that generates patches from a point cloud.

Parameters:

• data (dict) -- Point cloud data. Must include keys 'x' for point cloud and 'normals' which must have the correct orientation of the pointcloud. The normals do not need to be accurate other than the orientation.

• graph_radius (float) -- Radius for graph construction.

• batch_size (int, optional) -- Batch size. Defaults to 1.

• center (str, optional) -- Method for choosing patch centers. If 'tree', a KDTree is used to ensure patches completely cover the point cloud. If 'gmls', there is one patch for every point in the data. Defaults to 'tree'.

• shuffle_patches (bool, optional) -- Whether to shuffle patches. Defaults to True.

• knn (int, optional) -- Number of neighbors to choose for patch construction. Defaults to 50.

• min_radius (float, optional) -- Minimum patch radius. Defaults to 0.01.

• pca (bool, optional) -- If True, returns patch in local PCA coordinates. Defaults to True.

• degree (int, optional) -- Maximum degree of 1D Legendre basis if pca is True. Defaults to 3.

• orientation (Optional[torch.Tensor], optional) -- Orientation used if pca is True. Defaults to None.

• scale (bool, optional) -- If True, scaling is used if pca is also True. Defaults to True.

• include_scale (bool, optional) -- If True, includes scale in the output. Defaults to True.

• min_z_scale (float, optional) -- Minimum scale used in pca rescaling. Defaults to 1e-3.

• max_num_neighbors (int, optional) -- Maximum neighbors used in graph construction. Defaults to 32.

• device (str, optional) -- Device to use. Defaults to 'cpu'.

get_centers() → LongTensor

Choose the centers of the patches. If self.center is 'tree', the centers are chosen so that there is sufficient coverage of the data. If self.center is 'gmls', there is one patch center :returns: Indices of the patch centers. :rtype: torch.LongTensor

get_patch(idx: int) → Data

Returns a single patch in the form of a torch_geometric Data object.

Parameters:

idx (int) -- Index of the patch to return.

Returns:

torch_geometric Data object containing the patch and all the associated data.

Return type:

Data

get_patch_indices() → tuple[list[int], list[int]]

Returns indices of the patches and buffer regions. :returns:

patch_idxlist of int

List of indices of the patches.

buffer_idxlist of int

List of indices of the buffer regions.

Return type:

tuple of list of int

class gnp.dataset.patch.PatchLoader(graph_radius: float, batch_size: int = 1, center: str = 'tree', shuffle_patches: bool = True, knn: int = 50, min_radius: float = 0.01, pca: bool = True, degree: int = 3, min_z_scale: float = 0.001, max_num_neighbors: int = 32, device: str = 'cpu')

Bases: object

PatchLoader class that outputs a PatchGenerator instance when given PCD data.

Parameters

graph_radiusfloat

Radius for graph construction.

batch_sizeint, optional

Batch size. Defaults to 1.

centerstr, optional

Method for choosing patch centers. If 'tree', a KDTree is used to ensure patches completely cover the point cloud. If 'gmls', there is one patch for every point in the data. Defaults to 'tree'.

shuffle_patchesbool, optional

Whether to shuffle patches. Defaults to True.

knnint, optional

Number of neighbors to choose for patch construction. Defaults to 50.

min_radiusfloat, optional

Minimum patch radius. Defaults to 0.01.

pcabool, optional

If True, returns patch in local PCA coordinates. Defaults to True.

degreeint, optional

Maximum degree of 1D Legendre basis if pca is True. Defaults to 3.

min_z_scalefloat, optional

Minimum scale used in pca rescaling. Defaults to 1e-3.

max_num_neighborsint, optional

Maximum neighbors used in graph construction. Defaults to 32.

devicestr, optional

Device to use. Defaults to 'cpu'.

gnp.dataset.patch.graph_from_dict(sample: dict, radius: float, max_num_neighbors: int = 32)

Construct a graph from a dictionary of node features and other data.

Parameters:

• sample (dict) -- Dictionary containing the node features and other data.

• radius (float) -- Radius for the radius graph.

• max_num_neighbors (int, optional) -- Maximum number of neighbors per node. Defaults to 32.

Returns:

A graph data object containing node features, edge indices, and edge attributes.

Return type:

torch_geometric.data.Data

gnp.dataset.pca

class gnp.dataset.pca.PCABatch(num_graphs: int, **kwargs)

Bases: object

Class for handling multiple PCA patches.

to_pca(data: Batch) → Batch

Convert the input batch to PCA basis.

Parameters:

data (Batch) -- Input batch of data.

Returns:

Batch of data in PCA basis.

Return type:

Batch

class gnp.dataset.pca.PCAPatch(data: dict | None = None, degree: int = 3, min_z_scale: float = 0.001)

Bases: object

initialize(data: dict)

Initialize the PCA basis for the input patch

Parameters:

data (dict) -- Dictionary containing the input patch data

to_pca() → Data

Perform change of coordinates to local PCA basis for input patch.

Returns:

Data object of patch in PCA basis.

Return type:

Data

Module contents