nvidia.dali.fn.external_source(source=None, num_outputs=None, *, cycle=None, name=None, device='cpu', layout=None, dtype=None, ndim=None, cuda_stream=None, use_copy_kernel=None, batch=True, **kwargs)

Creates a data node which is populated with data from a Python source. The data can be provided by the source function or iterable, or it can be provided by pipeline.feed_input(name, data, layout, cuda_stream) inside pipeline.iter_setup.

In the case of the GPU input, it is the user responsibility to modify the provided GPU memory content only using provided stream (DALI schedules a copy on it and all work is properly queued). If no stream is provided feeding input blocks until the provided memory is copied to the internal buffer.


nvidia.dali.fn.external_source() operator is partially compatible with TensorFlow integration via nvidia.dali.plugin.tf.experimental.DALIDatasetWithInputs(). Please refer to its documentation for details.


To return a batch of copies of the same tensor, use nvidia.dali.types.Constant(), which is more performant.

  • source (callable or iterable) –

    The source of the data.

    The source is polled for data (via a call source() or next(source)) when the pipeline needs input for the next iteration. Depending on the value of num_outputs, the source can supply one or more data items. The data item can be a whole batch (default) or a single batch entry (when batch==False). If num_outputs is not set, the source is expected to return one item (a batch or a sample). If this value is specified (even if its value is 1), the data is expected to a be tuple, or list, where each element corresponds to respective return value of the external_source.

    The data samples must be in one of the compatible array types:

    • NumPy ndarray (CPU)

    • MXNet ndarray (CPU)

    • PyTorch tensor (CPU or GPU)

    • CuPy array (GPU)

    • objects implementing __cuda_array_interface__

    • DALI Tensor object

    Batch sources must produce entire batches of data. This can be achieved either by adding a new outermost dimension to an array or by returning a list of arrays (in which case they can be of different size, but must have the same rank and element type). A batch source can also produce a DALI TensorList object, which can be an output of another DALI pipeline.

    A per-batch source may accept one positional argument. If it does, it is the index of current iteration within epoch and consecutive calls will be source(0), source(1), and so on. If batch_info is set to True, instance of nvidia.dali.types.BatchInfo will be passed to the source, instead of a plain index.

    A per-sample source may accept one positional argument of type nvidia.dali.types.SampleInfo, which contains index of the sample in current epoch and in the batch, as well as current iteration number.

    If the source is a generator function, the function is invoked and treated as an iterable. However, unlike a generator, the function can be used with cycle. In this case, the function will be called again when the generator reaches the end of iteration.

    For GPU inputs, it is a user’s responsibility to modify the provided GPU memory content only in the provided stream. DALI schedules a copy on this stream, and all work is properly queued. If no stream is provided, DALI will use a default, with a best-effort approach at correctness. See the cuda_stream argument documentation for more information.

  • num_outputs (int, optional) –

    If specified, denotes the number of TensorLists that are produced by the source function.

    If set, the operator returns a list of DataNode objects, otherwise a single DataNode object is returned.

Keyword Arguments
  • cycle (string or bool, optional) –

    Specifies if and how to cycle through the source. It can be one of the following values:

    • "no", False or None - don’t cycle; StopIteration is raised whe end of data is reached; this is the default behavior

    • "quiet" or True - the data is repeated indefinitely,

    • "raise" - when the end of data is reached, StopIteration is raised, but the iteration is restarted on subsequent call.

    This flag requires that the source is a collection, for example, an iterable object where iter(source) returns a fresh iterator on each call, or a generator function. In the latter case, the generator function is called again when more data than was yielded by the function is requested.

    Specifying "raise" can be used with DALI iterators to create a notion of epoch.

  • name (str, optional) –

    The name of the data node.

    Used when feeding the data in iter_setup and can be omitted if the data is provided by source.

  • layout (layout str or list/tuple thereof, optional) –

    If provided, sets the layout of the data.

    When num_outputs > 1, the layout can be a list that contains a distinct layout for each output. If the list has fewer than num_outputs elements, only the first outputs have the layout set, the rest of the outputs don’t have a layout set.

  • dtype (nvidia.dali.types.DALIDataType or list/tuple thereof, optional) –

    Input data type.

    When num_outputs > 1, the dtype can be a list that contains a distinct value for each output.

    The operator will validate that the fetched data is of the provided type. If the argument is omitted or DALIDataType.NO_TYPE is passed, the operator will infer the type from the provided data.

    This argument will be required starting from DALI 2.0.

  • ndim (int or list/tuple thereof, optional) –

    Number of dimensions in the input data.

    When num_outputs > 1, the ndim can be a list that contains a distinct value for each output.

    The dimensionality of the data provided to the operator will be verified against this value. Number of dimensions can be also inferred from the layout argument if provided.

    If the layout argument is provided, the ndim must match the number of dimensions in the layout.

    Specifying the input dimensionality will be required starting from DALI 2.0

  • cuda_stream (optional, cudaStream_t or an object convertible to cudaStream_t, such as cupy.cuda.Stream or torch.cuda.Stream) –

    The CUDA stream is used to copy data to the GPU or from a GPU source.

    If this parameter is not set, a best-effort will be taken to maintain correctness. That is, if the data is provided as a tensor/array from a recognized library such as CuPy or PyTorch, the library’s current stream is used. Although this approach works in typical scenarios, with advanced use cases, and code that uses unsupported libraries, you might need to explicitly supply the stream handle.

    This argument has two special values:

    • 0 - Use the default CUDA stream

    • 1 - Use DALI’s internal stream

    If internal stream is used, the call to feed_input will block until the copy to internal buffer is complete, since there’s no way to synchronize with this stream to prevent overwriting the array with new data in another stream.

  • use_copy_kernel (bool, optional) –

    If set to True, DALI will use a CUDA kernel to feed the data instead of cudaMemcpyAsync (default).


    This is applicable only when copying data to and from GPU memory.

  • blocking (bool, optional) – Determines whether the external source should wait until data is available or just fail when the data is not available.

  • no_copy (bool, optional) –

    Determines whether DALI should copy the buffer when feed_input is called.

    If set to True, DALI passes the user memory directly to the pipeline, instead of copying it. It is the user responsibility to keep the buffer alive and unmodified until it is consumed by the pipeline.

    The buffer can be modified or freed again after the output of the relevant iterations has been consumed. Effectively, it happens after Pipeline’s prefetch_queue_depth or cpu_queue_depth * gpu_queue_depth (when they are not equal) iterations following the feed_input call.

    The memory location must match the specified device parameter of the operator. For the CPU, the provided memory can be one contiguous buffer or a list of contiguous Tensors. For the GPU, to avoid extra copy, the provided buffer must be contiguous. If you provide a list of separate Tensors, there will be an additional copy made internally, consuming both memory and bandwidth.

    Automatically set to True when parallel=True

  • batch (bool, optional) –

    If set to True or None, the source is expected to produce an entire batch at once. If set to False, the source is called per-sample.

    Setting parallel to True automatically sets batch to False if it was not provided.

  • batch_info (bool, optional, default = False) – Controls if a callable source that accepts an argument and returns batches should receive BatchInfo instance or just an integer representing the iteration number. If set to False (the default), only the integer is passed. If source is not callable, does not accept arguments or batch is set to False, setting this flag has no effect.

  • parallel (bool, optional, default = False) –

    If set to True, the corresponding pipeline will start a pool of Python workers to run the callback in parallel. You can specify the number of workers by passing py_num_workers into pipeline’s constructor.

    When parallel is set to True, samples returned by source must be NumPy/MXNet/PyTorch CPU arrays or TensorCPU instances.

    Acceptable sources depend on the value specified for batch parameter.

    If batch is set to False, the source must be:

    • a callable (a function or an object with __call__ method) that accepts exactly one argument (SampleInfo instance that represents the index of the requested sample).

    If batch is set to True, the source can be either:

    • a callable that accepts exactly one argument (either BatchInfo instance or an integer - see batch_info for details)

    • an iterable,

    • a generator function.


    Irrespective of batch value, callables should be stateless - they should produce requested sample or batch solely based on the SampleInfo/BatchInfo instance or index in batch, so that they can be run in parallel in a number of workers.

    The source callback must raise a StopIteration when the end of the data is reached. Note, that due to prefetching, the callback may be invoked with a few iterations past the end of dataset - make sure it consistently raises a StopIteration in that case.


    Callable source can be run in parallel by multiple workers. For batch=True multiple batches can be prepared in parallel, with batch=False it is possible to parallelize computation within the batch.

    When batch=True, callables performance might especially benefit from increasing prefetch_queue_depth so that a few next batches can be computed in parallel.


    Iterator or generator function will be assigned to a single worker that will iterate over them. The main advantage is execution in parallel to the main Python process, but due to their state it is not possible to calculate more than one batch at a time.

  • prefetch_queue_depth (int, option, default = 1) – When run in parallel=True mode, specifies the number of batches to be computed in advance and stored in the internal buffer, otherwise parameter is ignored.