Data Readers Reference

This page documents the data readers available in DALI Dynamic.

Once you create an instance of a reader, you can use it to iterate over the data multiple times. To start a new iteration over the data, you can call the next_epoch method, which will return an iterator that yields data.

If next_epoch is called without batch_size argument specified, the reader will return individual samples. Otherwise, it will return batches of the specified size. The batch size can differ from epoch to epoch.

import nvidia.dali.experimental.dynamic as ndd
reader = ndd.readers.File(file_root=images_dir)
for batch in reader.next_epoch(batch_size=16):
    images_batch, labels_batch = batch
    # process the batch

for batch in reader.next_epoch():
    image, label = batch
    # process the single sample

The table below lists the available readers.

Function	Device support	Short description
experimental.readers.Fits	CPU, GPU	Reads Fits image HDUs from a directory.
experimental.readers.Video	CPU, GPU	Loads and decodes video files from disk.
readers.Caffe	CPU	Reads (Image, label) pairs from a Caffe LMDB.
readers.Caffe2	CPU	Reads sample data from a Caffe2 Lightning Memory-Mapped Database (LMDB).
readers.COCO	CPU	Reads data from a COCO dataset that is composed of a directory with images and annotation JSON files.
readers.File	CPU	Reads file contents and returns file-label pairs.
readers.MXNet	CPU	Reads the data from an MXNet RecordIO.
readers.NemoAsr	CPU	Reads automatic speech recognition (ASR) data (audio, text) from an NVIDIA NeMo compatible manifest.
readers.Numpy	CPU, GPU	Reads Numpy arrays from a directory.
readers.Sequence	CPU	Reads [Frame] sequences from a directory representing a collection of streams.
readers.TFRecord	CPU	Reads samples from a TensorFlow TFRecord file.
readers.Video	GPU	Loads and decodes video files using FFmpeg and NVDECODE, which is the hardware-accelerated video decoding feature in the NVIDIA(R) GPU.
readers.VideoResize	GPU	Loads, decodes and resizes video files with FFmpeg and NVDECODE, which is NVIDIA GPU’s hardware-accelerated video decoding.
readers.Webdataset	CPU	A reader for the webdataset format.