kaolin.datasets.shapenet

class ShapeNet_Meshes(root: str, categories: list = ['chair'], train: bool = True, split: float = 0.7, no_progress: bool = False)[source]

ShapeNet Dataset class for meshes.

Parameters

  • root (str) – Path to the root directory of the ShapeNet dataset.

  • categories (str) – List of categories to load from ShapeNet. This list may contain synset ids, class label names (for ShapeNetCore classes), or a combination of both.

  • train (bool) – return the training set else the test set

  • split (float) – amount of dataset that is training out of 1

  • no_progress (bool) – if True, disables progress bar

Returns

dict: {
    attributes: {name: str, path: str, synset: str, label: str},
    data: {vertices: torch.Tensor, faces: torch.Tensor}
}

Example

>>> meshes = ShapeNet_Meshes(root='../data/ShapeNet/')
>>> obj = next(iter(meshes))
>>> obj['data']['vertices'].shape
torch.Size([2133, 3])
>>> obj['data']['faces'].shape
torch.Size([1910, 3])
class ShapeNet_Images(root: str, categories: list = ['chair'], train: bool = True, split: float = 0.7, views: int = 24, transform=None)[source]

ShapeNet Dataset class for images.

Parameters

  • root (str) – Path to the root directory of the ShapeNet dataset.

  • categories (str) – List of categories to load from ShapeNet. This list may contain synset ids, class label names (for ShapeNetCore classes), or a combination of both.

  • train (bool) – if true use the training set, else use the test set

  • split (float) – amount of dataset that is training out of

  • views (int) – number of viewpoints per object to load

  • transform (torchvision.transforms) – transformation to apply to images

  • no_progress (bool) – if True, disables progress bar

Returns

dict: {
    attributes: {name: str, path: str, synset: str, label: str},
    data: {vertices: torch.Tensor, faces: torch.Tensor}
    params: {
        cam_mat: torch.Tensor,
        cam_pos: torch.Tensor,
        azi: float,
        elevation: float,
        distance: float
    }
}

Example

>>> from torch.utils.data import DataLoader
>>> images = ShapeNet_Images(root='../data/ShapeNetImages')
>>> train_loader = DataLoader(images, batch_size=10, shuffle=True, num_workers=8)
>>> obj = next(iter(train_loader))
>>> image = obj['data']['imgs']
>>> image.shape
torch.Size([10, 4, 137, 137])
class ShapeNet_Voxels(root: str, cache_dir: str, categories: list = ['chair'], train: bool = True, split: float = 0.7, resolutions=[128, 32], no_progress: bool = False)[source]

ShapeNet Dataset class for voxels.

Parameters

  • root (str) – Path to the root directory of the ShapeNet dataset.

  • cache_dir (str) – Path to save cached converted representations.

  • categories (str) – List of categories to load from ShapeNet. This list may contain synset ids, class label names (for ShapeNetCore classes), or a combination of both.

  • train (bool) – return the training set else the test set

  • split (float) – amount of dataset that is training out of 1

  • resolutions (list) – list of resolutions to be returned

  • no_progress (bool) – if True, disables progress bar

Returns

dict: {
    attributes: {name: str, synset: str, label: str},
    data: {[res]: torch.Tensor}
}

Example

>>> from torch.utils.data import DataLoader
>>> voxels = ShapeNet_Voxels(root='../data/ShapeNet/', cache_dir='cache/')
>>> train_loader = DataLoader(voxels, batch_size=10, shuffle=True, num_workers=8 )
>>> obj = next(iter(train_loader))
>>> obj['data']['128'].shape
torch.Size([10, 128, 128, 128])
class ShapeNet_Surface_Meshes(root: str, cache_dir: str, categories: list = ['chair'], train: bool = True, split: float = 0.7, resolution: int = 100, smoothing_iterations: int = 3, mode='Tri', no_progress: bool = False)[source]

ShapeNet Dataset class for watertight meshes with only the surface preserved.

Parameters

  • root (str) – Path to the root directory of the ShapeNet dataset.

  • cache_dir (str) – Path to save cached converted representations.

  • categories (str) – List of categories to load from ShapeNet. This list may contain synset ids, class label names (for ShapeNetCore classes), or a combination of both.

  • train (bool) – return the training set else the test set

  • split (float) – amount of dataset that is training out of 1

  • resolution (int) – resolution of voxel object to use when converting

  • smoothing_iteration (int) – number of applications of laplacian smoothing

  • no_progress (bool) – if True, disables progress bar

Returns

dict: {
    attributes: {name: str, synset: str, label: str},
    data: {vertices: torch.Tensor, faces: torch.Tensor}
}

Example

>>> surface_meshes = ShapeNet_Surface_Meshes(root='../data/ShapeNet', cache_dir='cache/')
>>> obj = next(iter(surface_meshes))
>>> obj['data']['vertices'].shape
torch.Size([11617, 3])
>>> obj['data']['faces'].shape
torch.Size([23246, 3])
class ShapeNet_Points(root: str, cache_dir: str, categories: list = ['chair'], train: bool = True, split: float = 0.7, num_points: int = 5000, smoothing_iterations=3, surface=True, resolution=100, normals=True, no_progress: bool = False)[source]

ShapeNet Dataset class for sampled point cloud from each object.

Parameters

  • root (str) – Path to the root directory of the ShapeNet dataset.

  • cache_dir (str) – Path to save cached converted representations.

  • categories (str) – List of categories to load from ShapeNet. This list may contain synset ids, class label names (for ShapeNetCore classes), or a combination of both.

  • train (bool) – return the training set else the test set

  • split (float) – amount of dataset that is training out of 1

  • num_points (int) – number of point sampled on mesh

  • smoothing_iteration (int) – number of application of laplacian smoothing

  • surface (bool) – if only the surface of the original mesh should be used

  • resolution (int) – resolution of voxel object to use when converting

  • normals (bool) – should the normals of the points be saved

  • no_progress (bool) – if True, disables progress bar

Returns

dict: {
    attributes: {name: str, synset: str, label: str},
    data: {points: torch.Tensor, normals: torch.Tensor}
}

Example

>>> from torch.utils.data import DataLoader
>>> points = ShapeNet_Points(root='../data/ShapeNet', cache_dir='cache/')
>>> train_loader = DataLoader(points, batch_size=10, shuffle=True, num_workers=8)
>>> obj = next(iter(train_loader))
>>> obj['data']['points'].shape
torch.Size([10, 5000, 3])
class ShapeNet_SDF_Points(root: str, cache_dir: str, categories: list = ['chair'], train: bool = True, split: float = 0.7, resolution: int = 100, num_points: int = 5000, occ: bool = False, smoothing_iterations: int = 3, sample_box=True, no_progress: bool = False)[source]

ShapeNet Dataset class for signed distance functions.

Parameters

  • root (str) – Path to the root directory of the ShapeNet dataset.

  • cache_dir (str) – Path to save cached converted representations.

  • categories (str) – List of categories to load from ShapeNet. This list may contain synset ids, class label names (for ShapeNetCore classes), or a combination of both.

  • train (bool) – return the training set else the test set

  • split (float) – amount of dataset that is training out of 1

  • resolution (int) – resolution of voxel object to use when converting

  • num_points (int) – number of sdf points sampled on mesh

  • occ (bool) – should only occupancy values be returned instead of distances

  • smoothing_iteration (int) – number of application of laplacian smoothing

  • sample_box (bool) – whether to sample only from within mesh extents

  • no_progress (bool) – if True, disables progress bar

Returns

dict: {
    attributes: {name: str, synset: str, label: str},
    data: {
        Union['occ_values', 'sdf_distances']: torch.Tensor,
        Union['occ_points, 'sdf_points']: torch.Tensor}
}

Example

>>> from torch.utils.data import DataLoader
>>> sdf_points = ShapeNet_SDF_Points(root='../data/ShapeNet', cache_dir='cache/')
>>> train_loader = DataLoader(sdf_points, batch_size=10, shuffle=True, num_workers=8)
>>> obj = next(iter(train_loader))
>>> obj['data']['sdf_points'].shape
torch.Size([10, 5000, 3])
class ShapeNet_Tags(dataset, tag_aug=True)[source]

ShapeNet Dataset class for tags.

Parameters

  • dataset (kal.dataloader.shapenet.ShapeNet) – One of the ShapeNet datasets

  • download (bool) – If True will load taxonomy of objects if it is not loaded yet

  • transform – transformation to apply to tags

Returns

Dictionary with key for the input tags encod and : ‘tag_enc’:

Return type

dict

Example

>>> from torch.utils.data import DataLoader
>>> meshes = ShapeNet_Meshes(root='../data/ShapeNet/', cache_dir='cache/')
>>> tags = ShapeNet_Tags(meshes)
>>> train_loader = DataLoader(tags, batch_size=10, shuffle=True, num_workers=8 )
>>> obj = next(iter(train_loader))
>>> obj['data']['tag_enc'].shape
torch.Size([10, N])
class ShapeNet_Combination(datasets)[source]

ShapeNet Dataset class for combinations of representations.

Parameters

  • dataset (list) – List of datasets to be combined

  • categories (str) – List of categories to load from ShapeNet. This list may contain synset ids, class label names (for ShapeNetCore classes), or a combination of both.

  • root (str) – Path to the root directory of the ShapeNet dataset.

  • train (bool) – if true use the training set, else use the test set

Returns

Dictionary with keys indicated by passed datasets

Return type

dict

Example

>>> from torch.utils.data import DataLoader
>>> shapenet = ShapeNet_Meshes(root='../data/ShapeNet', cache_dir='cache/')
>>> voxels = ShapeNet_Voxels(root='../data/ShapeNet', cache_dir='cache/')
>>> images = ShapeNet_Images(root='../data/ShapeNet', cache_dir='cache/')
>>> points = ShapeNet_Points(root='../data/ShapeNet', cache_dir='cache/')
>>> dataset = ShapeNet_Combination([voxels, images, points])
>>> train_loader = DataLoader(dataset, batch_size=10, shuffle=True, num_workers=8)
>>> obj = next(iter(train_loader))
>>> for key in obj['data']:
...     print (key)
...
params
128
32
imgs
cam_mat
cam_pos
azi
elevation
distance
points
normals