rocAL Python API overview

rocAL Python API overview#

The rocAL Python package has been created using Pybind11 which enables data transfer between the rocAL C++ API and Python API. The rocal_pybind package includes both PyTorch and TensorFlow framework support and support for multiple data readers such as FileReader, COCOReader, and TFRecordReader.

The rocAL data types are defined in amd.rocal.types.

amd.rocal.fn: Contains the image augmentations linked to the rocAL C++ API.
amd.rocal.decoders: Image, video, and audio decoders.
amd.rocal.readers: Image, video, and audio readers.
amd.rocal.pipeline: The pipeline class encapsulates the data needed to build and run a rocAL graph. This includes support for context and graph creation, functions to verify and run the graph, and data transfer functions.
amd.rocal.types: enums exported from the C++ API to Python.
amd.rocal.plugin.pytorch: PyTorch plugin that includes the ROCALGenericIterator for Pytorch. The ROCALClassificationIterator class implements an iterator for image classification that returns labelled images.
amd.rocal.plugin.tf: TensorFlow plugin that includes the readers.tfrecord TensorFlow reader.
amd.rocal.plugin.jax: JAX plugin that includes the ROCALJaxIterator for JAX. The ROCALJaxIterator implements an iterator for running a pipeline over multiple devices.