gnp.dataset module

gnp.dataset.patch

class gnp.dataset.patch.PatchData(x: Tensor | None = None, **kwargs)

Bases: Data

A PyTorch Geometric Data object specialized for patch-based point cloud data.

This class extends the base Data object to include patch-specific attributes and a utility for iterating over batches of patches.

batch_iterator(batch_size: int)

Create a generator to iterate over batches of patches.

This method yields smaller PatchData objects, each containing a subset of the patches. It assumes that the patch data is sorted by patch_number.

Parameters:

batch_size (int) -- The maximum number of patches in each batch.

Yields:

PatchData -- A new PatchData object representing a batch of patches.

property num_patches

Returns the number of patches in the dataset.

class gnp.dataset.patch.PatchTensor(data: dict, k: int = 30, mode: str = 'test', pca: bool = True, scale: bool = True, min_z_scale: float = 0.005, basis: str = 'legendre', basis_degree: int = 3, num_training_patches: int = 1024, device: str = 'cpu')

Bases: object

Processes a point cloud into a collection of overlapping patches.

This class handles the entire pipeline of patchifying a point cloud. It selects patch centers, finds neighboring points for each patch, computes local coordinate systems using PCA, and scales the coordinates. The final output is a PatchData object ready for use in a model.

Parameters:

• data (dict) -- A dictionary containing the point cloud data, requires at least an 'x' key with a tensor of shape (N, 3).

• k (int, optional) -- Number of nearest neighbors to consider for various calculations, by default 30.

• mode (str, optional) -- The mode for center selection ('train', 'test', or 'gmls'), by default "test".

• pca (bool, optional) -- Whether to use PCA to determine local coordinate systems, by default True.

• scale (bool, optional) -- Whether to scale the local coordinates, by default True.

• min_z_scale (float, optional) -- The minimum value for z-scaling, by default 5e-3.

• basis (str, optional) -- The basis to use, by default "legendre".

• basis_degree (int, optional) -- The degree of the basis, by default 3.

• num_training_patches (int, optional) -- Number of patches to sample in 'train' mode, by default 1024.

• device (str, optional) -- The device to perform computations on, by default "cpu".

as_patch_data() → PatchData

Assemble and return the final PatchData object.

This method collects all computed attributes (patch indices, local coordinates, PCA vectors, etc.) and any additional data from the input dictionary into a single PatchData object.

Returns:

The fully processed patch data object.

Return type:

PatchData

get_centers() → tuple[Tensor, Tensor, Tensor]

Dispatch method to get patch centers based on the current mode.

Returns:

A tuple containing: - The indices of the center points. - The cluster assignment for each point in the cloud. - The k-NN distance for each center, used to determine patch radius.

Return type:

tuple[torch.Tensor, torch.Tensor, torch.Tensor]

get_gmls_centers() → tuple[Tensor, Tensor, Tensor]

Select every point as a patch center for GMLS.

In GMLS mode, a patch is centered at every single point in the cloud.

Returns:

Center indices, cluster assignments, and k-NN distances.

Return type:

tuple[torch.Tensor, torch.Tensor, torch.Tensor]

get_test_centers() → tuple[Tensor, Tensor, Tensor]

Select patch centers using a greedy covering strategy for testing.

This method iterates through shuffled points and selects a point as a center if it's not already covered by an existing patch, effectively creating a set of patches that cover the entire point cloud.

Returns:

Center indices, cluster assignments, and k-NN distances.

Return type:

tuple[torch.Tensor, torch.Tensor, torch.Tensor]

get_train_centers() → tuple[Tensor, Tensor, Tensor]

Select patch centers by random sampling for training.

Returns:

Center indices, cluster assignments (dummy), and k-NN distances.

Return type:

tuple[torch.Tensor, torch.Tensor, torch.Tensor

property local_coordinates

The scaled local coordinates of points within their respective patches.

property local_coordinates_original

The scaled local coordinates of the 'original' points.

property pca_vectors

The PCA vectors (local basis) for each patch.

property scaling

The combined (x, y, z) scaling vector for each patch.

property tensor_centered

Dense tensor of patch points, centered by subtracting the patch mean.

Returns:

A tensor of shape (num_patches, max_points_in_patch, 3).

Return type:

torch.Tensor

property tensor_local

Dense tensor of patch points in local coordinates.

property x_local

All points transformed into the local coordinate system of their assigned patch.

property xy_scale

The xy-plane scaling factor for each patch.

property z_scale

The z-axis scaling factor for each patch.

Module contents