PyTorch Plugin API reference

class nvidia.dali.plugin.pytorch.DALIClassificationIterator(pipelines, size=-1, reader_name=None, auto_reset=False, fill_last_batch=None, dynamic_shape=False, last_batch_padded=False, last_batch_policy=<LastBatchPolicy.FILL: 0>, prepare_first_batch=True)

DALI iterator for classification tasks for PyTorch. It returns 2 outputs (data and label) in the form of PyTorch’s Tensor.

Calling

DALIClassificationIterator(pipelines, reader_name)

is equivalent to calling

DALIGenericIterator(pipelines, ["data", "label"], reader_name)
Parameters

  • pipelines (list of nvidia.dali.Pipeline) – List of pipelines to use

  • size (int, default = -1) – Number of samples in the shard for the wrapped pipeline (if there is more than one it is a sum) Providing -1 means that the iterator will work until StopIteration is raised from the inside of iter_setup(). The options last_batch_policy, last_batch_padded and auto_reset don’t work in such case. It works with only one pipeline inside the iterator. Mutually exclusive with reader_name argument

  • reader_name (str, default = None) – Name of the reader which will be queried to the shard size, number of shards and all other properties necessary to count properly the number of relevant and padded samples that iterator needs to deal with. It automatically sets last_batch_policy to PARTIAL when the FILL is used, and last_batch_padded accordingly to match the reader’s configuration

  • auto_reset (bool, optional, default = False) – Whether the iterator resets itself for the next epoch or it requires reset() to be called separately.

  • dynamic_shape (any, optional,) – Parameter used only for backward compatibility.

  • fill_last_batch (bool, optional, default = None) –

    Deprecated Please use last_batch_policy instead

    Whether to fill the last batch with data up to ‘self.batch_size’. The iterator would return the first integer multiple of self._num_gpus * self.batch_size entries which exceeds ‘size’. Setting this flag to False will cause the iterator to return exactly ‘size’ entries.

  • optional, default = LastBatchPolicy.FILL (last_batch_policy,) – What to do with the last batch when there is no enough samples in the epoch to fully fill it. See nvidia.dali.plugin.base_iterator.LastBatchPolicy()

  • last_batch_padded (bool, optional, default = False) – Whether the last batch provided by DALI is padded with the last sample or it just wraps up. In the conjunction with last_batch_policy it tells if the iterator returning last batch with data only partially filled with data from the current epoch is dropping padding samples or samples from the next epoch. If set to False next epoch will end sooner as data from it was consumed but dropped. If set to True next epoch would be the same length as the first one. For this to happen, the option pad_last_batch in the reader needs to be set to True as well. It is overwritten when reader_name argument is provided

  • prepare_first_batch (bool, optional, default = True) – Whether DALI should buffer the first batch right after the creation of the iterator, so one batch is already prepared when the iterator is prompted for the data

Example

With the data set [1,2,3,4,5,6,7] and the batch size 2:

last_batch_policy = LastBatchPolicy.PARTIAL, last_batch_padded = True -> last batch = [7], next iteration will return [1, 2]

last_batch_policy = LastBatchPolicy.PARTIAL, last_batch_padded = False -> last batch = [7], next iteration will return [2, 3]

last_batch_policy = LastBatchPolicy.FILL, last_batch_padded = True -> last batch = [7, 7], next iteration will return [1, 2]

last_batch_policy = LastBatchPolicy.FILL, last_batch_padded = False -> last batch = [7, 1], next iteration will return [2, 3]

last_batch_policy = LastBatchPolicy.DROP, last_batch_padded = True -> last batch = [5, 6], next iteration will return [1, 2]

last_batch_policy = LastBatchPolicy.DROP, last_batch_padded = False -> last batch = [5, 6], next iteration will return [2, 3]

class nvidia.dali.plugin.pytorch.DALIGenericIterator(pipelines, output_map, size=-1, reader_name=None, auto_reset=False, fill_last_batch=None, dynamic_shape=False, last_batch_padded=False, last_batch_policy=<LastBatchPolicy.FILL: 0>, prepare_first_batch=True)

General DALI iterator for PyTorch. It can return any number of outputs from the DALI pipeline in the form of PyTorch’s Tensors.

Parameters

  • pipelines (list of nvidia.dali.Pipeline) – List of pipelines to use

  • output_map (list of str) – List of strings which maps consecutive outputs of DALI pipelines to user specified name. Outputs will be returned from iterator as dictionary of those names. Each name should be distinct

  • size (int, default = -1) – Number of samples in the shard for the wrapped pipeline (if there is more than one it is a sum) Providing -1 means that the iterator will work until StopIteration is raised from the inside of iter_setup(). The options last_batch_policy, last_batch_padded and auto_reset don’t work in such case. It works with only one pipeline inside the iterator. Mutually exclusive with reader_name argument

  • reader_name (str, default = None) – Name of the reader which will be queried to the shard size, number of shards and all other properties necessary to count properly the number of relevant and padded samples that iterator needs to deal with. It automatically sets last_batch_policy to PARTIAL when the FILL is used, and last_batch_padded accordingly to match the reader’s configuration

  • auto_reset (bool, optional, default = False) – Whether the iterator resets itself for the next epoch or it requires reset() to be called separately.

  • dynamic_shape (any, optional,) – Parameter used only for backward compatibility.

  • fill_last_batch (bool, optional, default = None) –

    Deprecated Please use last_batch_policy instead

    Whether to fill the last batch with data up to ‘self.batch_size’. The iterator would return the first integer multiple of self._num_gpus * self.batch_size entries which exceeds ‘size’. Setting this flag to False will cause the iterator to return exactly ‘size’ entries.

  • optional, default = LastBatchPolicy.FILL (last_batch_policy,) – What to do with the last batch when there is no enough samples in the epoch to fully fill it. See nvidia.dali.plugin.base_iterator.LastBatchPolicy()

  • last_batch_padded (bool, optional, default = False) – Whether the last batch provided by DALI is padded with the last sample or it just wraps up. In the conjunction with last_batch_policy it tells if the iterator returning last batch with data only partially filled with data from the current epoch is dropping padding samples or samples from the next epoch. If set to False next epoch will end sooner as data from it was consumed but dropped. If set to True next epoch would be the same length as the first one. For this to happen, the option pad_last_batch in the reader needs to be set to True as well. It is overwritten when reader_name argument is provided

  • prepare_first_batch (bool, optional, default = True) – Whether DALI should buffer the first batch right after the creation of the iterator, so one batch is already prepared when the iterator is prompted for the data

Example

With the data set [1,2,3,4,5,6,7] and the batch size 2:

last_batch_policy = LastBatchPolicy.PARTIAL, last_batch_padded = True -> last batch = [7], next iteration will return [1, 2]

last_batch_policy = LastBatchPolicy.PARTIAL, last_batch_padded = False -> last batch = [7], next iteration will return [2, 3]

last_batch_policy = LastBatchPolicy.FILL, last_batch_padded = True -> last batch = [7, 7], next iteration will return [1, 2]

last_batch_policy = LastBatchPolicy.FILL, last_batch_padded = False -> last batch = [7, 1], next iteration will return [2, 3]

last_batch_policy = LastBatchPolicy.DROP, last_batch_padded = True -> last batch = [5, 6], next iteration will return [1, 2]

last_batch_policy = LastBatchPolicy.DROP, last_batch_padded = False -> last batch = [5, 6], next iteration will return [2, 3]

nvidia.dali.plugin.pytorch.feed_ndarray(dali_tensor, arr, cuda_stream=None)

Copy contents of DALI tensor to PyTorch’s Tensor.

Parameters

  • dali_tensor (nvidia.dali.backend.TensorCPU or nvidia.dali.backend.TensorGPU) – Tensor from which to copy

  • arr (torch.Tensor) – Destination of the copy

  • cuda_stream (torch.cuda.Stream, cudaStream_t or any value that can be cast to cudaStream_t.) – CUDA stream to be used for the copy (if not provided, an internal user stream will be selected) In most cases, using pytorch’s current stream is expected (for example, if we are copying to a tensor allocated with torch.zeros(…))