| DataLoaders | |
| =========== | |
| After running a preprocessing pipeline and writing the resulting ``.h5path`` file to disk, the next step is to | |
| create a DataLoader for feeding tiles into a machine learning model in PyTorch. | |
| To do this, use the :class:`~pathml.ml.dataset.TileDataset` class and then wrap it in a PyTorch DataLoader: | |
| .. code-block:: | |
| dataset = TileDataset("/path/to/file.h5path") | |
| dataloader = torch.utils.data.DataLoader(dataset, batch_size = 16, shuffle = True, num_workers = 4) | |
| .. note:: | |
| Label dictionaries are not standardized, as users are free to store whatever labels they want. | |
| For that reason, PyTorch cannot automatically stack labels into batches. | |
| It may therefore be necessary to create a custom ``collate_fn`` to specify how to create batches of labels. | |
| See `here <https://discuss.pytorch.org/t/how-to-use-collate-fn/27181>`_. | |
| This provides an interface between PathML and the broader ecosystem of machine learning tools built on PyTorch. | |
| For more information on how to use Datasets and DataLoaders, please see the PyTorch | |
| `documentation <https://pytorch.org/docs/stable/data.html>`_ and | |
| `tutorials <https://pytorch.org/tutorials/beginner/basics/data_tutorial.html>`_. | |