Implementation:NVIDIA DALI TF Image Processing

Knowledge Sources	NVIDIA_DALI
Domains	Image_Classification, TensorFlow
Last Updated	2026-02-08 16:00 GMT

Overview

Provides image loading, preprocessing, and data pipeline utilities for ResNet training, with support for both native TensorFlow data loading and NVIDIA DALI GPU-accelerated preprocessing.

Description

This module implements the complete image preprocessing pipeline for ImageNet-based ResNet training with two interchangeable backends: a standard TensorFlow `tf.data` pipeline and an NVIDIA DALI GPU-accelerated pipeline. Both pipelines read TFRecord files containing ImageNet data and produce normalized, augmented image batches.

The TensorFlow pipeline deserializes TFRecord examples containing JPEG-encoded images, class labels, bounding boxes, and class text. It applies a chain of preprocessing operations: JPEG decoding, random crop and resize (training) or central crop (evaluation) with configurable aspect ratios and area ranges, optional color distortion (brightness, saturation, hue, contrast jittering), random horizontal flipping, and channel mean subtraction (using ImageNet means: R=123.68, G=116.78, B=103.94).

The DALI pipeline (`get_dali_pipeline`) provides equivalent functionality using GPU-accelerated operators: `fn.readers.tfrecord` for data reading with automatic sharding across GPUs, `fn.decoders.image_random_crop` for fused decode-and-crop (training) or `fn.decoders.image` with `fn.resize` (evaluation), and `fn.crop_mirror_normalize` for normalization and random mirroring. The pipeline supports CPU or mixed (CPU+GPU) decode modes with pre-allocated memory hints for NVJPEG.

The `DALIPreprocessor` class wraps the DALI pipeline for use with TensorFlow, providing both a `DALIIterator`-based interface and a `DALIDataset` interface. The `image_set()` factory function creates the appropriate pipeline based on the `use_dali` flag, supporting 'CPU' and 'GPU' DALI modes or falling back to the TensorFlow pipeline. A `fake_image_set()` function generates synthetic data for benchmarking.

Usage

Use this module in the ResNet-N training pipeline to create training and validation data iterators. Select between DALI and native TF preprocessing by setting the `dali_mode` parameter. Horovod is used for distributed data sharding across multiple GPUs.

Code Reference

Source Location

Repository: NVIDIA_DALI
File: docs/examples/use_cases/tensorflow/resnet-n/nvutils/image_processing.py
Lines: 1-322

Signature

@pipeline_def
def get_dali_pipeline(
    tfrec_filenames, tfrec_idx_filenames,
    height, width, shard_id, num_gpus,
    dali_cpu=True, training=True,
): ...

class DALIPreprocessor(object):
    def __init__(self, filenames, idx_filenames, height, width,
                 batch_size, num_threads, dtype=tf.uint8,
                 dali_cpu=True, deterministic=False, training=False): ...
    def get_device_minibatches(self): ...
    def get_device_dataset(self): ...

def image_set(filenames, batch_size, height, width, training=False,
              distort_color=False, num_threads=10, nsummary=10,
              deterministic=False, use_dali=None, idx_filenames=None): ...

def fake_image_set(batch_size, height, width, with_label=True): ...

def _deserialize_image_record(record): ...
def _decode_jpeg(imgdata, channels=3): ...
def _crop_and_resize_image(image, original_bbox, height, width,
                            deterministic=False, random_crop=False): ...
def _distort_image_color(image, order=0): ...
def _parse_and_preprocess_image_record(record, height, width,
                                        deterministic=False, random_crop=False,
                                        distort_color=False): ...

Import

from nvutils import image_processing

# Create DALI-accelerated dataset
dataset = image_processing.image_set(
    filenames=train_files,
    batch_size=256,
    height=224,
    width=224,
    training=True,
    use_dali="GPU",
    idx_filenames=train_idx_files,
)

I/O Contract

Inputs

Name	Type	Required	Description
filenames	list[str]	Yes	List of TFRecord file paths
idx_filenames	list[str]	No	List of DALI index file paths (required when use_dali is set)
batch_size	int	Yes	Number of images per batch
height	int	Yes	Target image height (e.g., 224)
width	int	Yes	Target image width (e.g., 224)
training	bool	No	Whether to apply training augmentations (default: False)
distort_color	bool	No	Whether to apply color distortion (default: False)
use_dali	str or None	No	DALI mode: 'CPU', 'GPU', or None for TF native pipeline
deterministic	bool	No	Whether to use deterministic random seeds (default: False)

Outputs

Name	Type	Description
images	tf.Tensor	Batch of preprocessed images [batch, height, width, 3] as float32
labels	tf.Tensor	Batch of 0-based class labels [batch] as int32/int64

Usage Examples

DALI Pipeline for Training

from nvutils import image_processing

# Create DALI preprocessor
preprocessor = image_processing.DALIPreprocessor(
    filenames=train_files,
    idx_filenames=train_idx_files,
    height=224, width=224,
    batch_size=256,
    num_threads=4,
    dali_cpu=False,
    training=True,
)

# Get dataset
dataset = preprocessor.get_device_dataset()
for images, labels in dataset:
    # images: [256, 224, 224, 3], labels: [256]
    pass

Related Pages

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment