Gaudi2 provides hardware based acceleration of inputs pre-processing for deep learning frameworks. It also provides Tensor Processing Core based, highly optimized operators generally used for data augmentation and image manipulation in input media processing.
Media pipe provides an interface which allows users to implement media operations on input elements, such as images. The purpose of this pipeline is to prepare batches of processed and augmented images as well as labels to be fed into training or inference models. Input data can be accompanied by additional input information such as labels for classes or bounding boxes. Most of the operations are implemented on HPU, enabling accelerated execution compared to operations on CPU. See Operators section for more details.
Supports JPEG images.
Scalable across multiple cards.
Portable across TensorFlow and PyTorch frameworks.
Accelerates image classification (ResNet-50).
Supports multiple functions such as reading and decoding as well as various data transformations such as image cropping or flipping.
For further details, see the below sections: