Workflow:NVIDIA DALI Image Preprocessing Pipeline
| Knowledge Sources | |
|---|---|
| Domains | Data_Loading, Image_Processing, GPU_Computing |
| Last Updated | 2026-02-08 17:00 GMT |
Overview
End-to-end process for building a GPU-accelerated image decoding and transformation pipeline using NVIDIA DALI, either as a standalone processing tool or integrated with deep learning frameworks.
Description
This workflow defines the foundational pattern for using DALI as a high-performance image preprocessing engine. It covers both the Pipeline API (graph-based, optimized for throughput in production training) and the Dynamic API (imperative, flexible for experimentation). The pipeline reads encoded image data from files or external sources, decodes them using GPU-accelerated hardware decoders (nvJPEG, nvImageCodec), applies a chain of image transformations (resize, crop, color adjustment, normalization), and outputs processed tensors that can be consumed by any downstream application or deep learning framework.
Usage
Execute this workflow when you need to build a fast image preprocessing pipeline for any purpose: training data preparation, inference preprocessing, batch image processing, or data exploration. This is the starting point for any DALI usage and serves as the foundation for more specialized workflows (classification, detection, segmentation).
Execution Steps
Step 1: Choose the Execution Mode
Select between Pipeline API mode and Dynamic API mode based on your requirements. Pipeline mode defines a static computational graph that is optimized for maximum throughput with prefetching and parallel execution. Dynamic mode calls operators directly in an imperative style for flexibility and rapid iteration.
Key considerations:
- Pipeline mode uses @pipeline_def to declare a processing graph, then builds and runs it
- Dynamic mode uses nvidia.dali.experimental.dynamic (ndd) to call operators directly
- Pipeline mode is recommended for production training workloads
- Dynamic mode is recommended for data exploration, debugging, and experimentation
- Both modes use the same underlying GPU-accelerated operators
Step 2: Configure the Data Source
Set up the data reader or external source that feeds encoded image data into the pipeline. DALI supports multiple input methods: file system readers, external sources for programmatic feeding, and format-specific readers for TFRecord, LMDB, RecordIO, COCO, WebDataset, and NumPy files.
Key considerations:
- fn.readers.file reads images from a directory tree, returning encoded bytes and labels
- fn.external_source accepts data provided by the user at each iteration from Python
- Format-specific readers (fn.readers.tfrecord, fn.readers.coco, etc.) handle structured datasets
- Configure prefetch queue depth for optimal overlap between I/O and compute
Step 3: Decode Images on GPU
Decode encoded image bytes (JPEG, PNG, TIFF, BMP, JPEG2000, WebP) into pixel tensors using DALI's hardware-accelerated decoders. The mixed device mode reads encoded data from CPU memory and decodes on GPU using NVDEC or nvJPEG hardware.
Key considerations:
- Use fn.decoders.image for standard decoding to full resolution
- Use fn.decoders.image_random_crop for fused decode-and-crop (more efficient than separate operations)
- Set device="mixed" for GPU-accelerated decoding
- Specify output_type (RGB, BGR, GRAY) as needed
- Hardware JPEG decoding provides the highest throughput for JPEG inputs
Step 4: Apply Image Transformations
Chain together image processing operators to transform the decoded images. Common operations include resizing, cropping, color space conversion, brightness/contrast adjustment, rotation, and geometric warping. Each operator runs on GPU for maximum throughput.
Key considerations:
- fn.resize scales images to target dimensions with configurable interpolation
- fn.crop extracts a region of interest; fn.crop_mirror_normalize fuses crop, flip, and normalize
- fn.color_twist adjusts brightness, contrast, saturation, and hue
- fn.rotate applies rotation with configurable fill values
- fn.warp_affine applies arbitrary affine transformations
- Operators can be chained; DALI optimizes the graph to minimize memory allocations
Step 5: Output Processed Data
Retrieve processed image tensors from the pipeline. In Pipeline mode, call pipeline.run() or use a framework iterator. In Dynamic mode, results are returned directly as DALI tensor objects that can be converted to framework-native tensors.
Key considerations:
- In Pipeline mode, use DALIGenericIterator (PyTorch), DALIDataset (TensorFlow), or direct pipeline.run()
- In Dynamic mode, use torch.as_tensor() or equivalent to convert DALI tensors
- Results reside on GPU by default; use .as_cpu() if CPU tensors are needed
- The pipeline handles batching, prefetching, and memory management automatically