PyTorch Usage Guide

DALI Dynamic provides interoperability with PyTorch, allowing you to use DALI operators for high-performance data loading and preprocessing in your PyTorch workflows.

Basic usage

The following example reads images from disk, decodes and preprocesses them using DALI Dynamic operators, and converts the results to PyTorch tensors:

import nvidia.dali.experimental.dynamic as ndd
import nvidia.dali.types as types

reader = ndd.readers.File(file_root=image_dir)
for jpegs, labels in reader.next_epoch(batch_size=16):
    images = ndd.decoders.image(jpegs, device="gpu")
    images = ndd.crop_mirror_normalize(
        images,
        dtype=types.FLOAT,
        output_layout="CHW",
        crop=(224, 224),
        mean=[0.485 * 255, 0.456 * 255, 0.406 * 255],
        std=[0.229 * 255, 0.224 * 255, 0.225 * 255],
    )
    images_torch = images.torch()
    labels_torch = labels.gpu().torch()

The conversion from DALI tensors to PyTorch tensors uses DLPack under the hood to avoid copies when possible See Tensor.torch() and Batch.torch() in the API reference for details.

Next steps

• TorchData Integration Reference: Build composable data loading pipelines with torchdata.nodes.

• Training a neural network with DALI and TorchData: End-to-end training example using DALI Dynamic with PyTorch.