zero.data.IndexLoader.__init__

IndexLoader.__init__(size, *args, device='cpu', **kwargs)[source]

Initialize self.

Parameters

  • size (int) – the number of items (for example, len(dataset))

  • *args – positional arguments for torch.utils.data.DataLoader

  • device (Union[int, str, torch.device]) – if not CPU, then all indices are materialized and moved to the device at the beginning of every loop. It can be useful when the indices are applied to non-CPU data (e.g. CUDA-tensors) and moving data between devices takes non-negligable time (which can happen in the case of simple and fast models like MLPs).

  • **kwargs – keyword arguments for torch.utils.data.DataLoader

Raises

AssertionError – if size is not positive

Return type

None