tlc.core.objects.tables.from_python_object.table_from_torch_dataset#

A table populated from a Torch dataset.

Module Contents#

Classes#

Class

Description

TableFromTorchDataset

A table populated from a Torch dataset.

API#

class tlc.core.objects.tables.from_python_object.table_from_torch_dataset.TableFromTorchDataset(url: tlc.core.url.Url | None = None, created: str | None = None, row_cache_url: tlc.core.url.Url | None = None, row_cache_populated: bool | None = None, override_table_rows_schema: tlc.core.schema.Schema | None = None, init_parameters: Any | None = None, input_dataset: torch.utils.data.Dataset | None = None, all_arrays_are_fixed_size: bool = False)#

Bases: tlc.core.objects.table.Table

A table populated from a Torch dataset.

When creating a TableFromTorchDataset, the row schema specified by the override_table_rows_schema parameter (if provided) must match the structure of the samples in the provided input_dataset. This row schema is used to convert the samples in the input_dataset to rows in the table, and to make sure that samples returned by the table’s __getitem__ method match the structure of the samples in the original input_dataset.

If the input_dataset is a torchvision.datasets.DatasetFolder, the TableFromTorchDataset will use the torchvision.datasets.folder.default_loader to load the images. This loader will be replaced with a 3LC loader that does not copy the images, but instead returns a PIL image with a filename attribute that contains the absolute path to the image. This allows the front-end to access the images in the table without copying them.

If the input_dataset is a torchvision.datasets.VisionDataset, the TableFromTorchDataset will also remove any transforms from the input_dataset before serializing it to the table, but will recreate the transforms on the VisionDataset after serialization. Any transforms defined on the input_dataset are added to the list of map_functions of the Table, and will be reflected in the samples returned by the Table’s __getitem__ method, but not serialized to the Table’s json file.

Create a Table from a Torch dataset.

Parameters:
  • url – The URL of the table.

  • created – The date and time the table was created.

  • dataset_name – The name of the dataset.

  • project_name – The name of the project.

  • row_cache_url – The URL of the row cache.

  • row_cache_populated – Whether the row cache has been populated.

  • override_table_rows_schema – The schema of the table rows.

  • init_parameters – The parameters used to initialize the table.

  • input_dataset – The Torch dataset to use to populate the table.