[Experimental] DALI Dynamic

DALI Dynamic extends NVIDIA DALI by introducing an imperative execution model with lazy evaluation. It complements the existing graph-based pipeline execution, and its main goal is to enable seamless integration into Python workflows. This results in easier debugging and opens a path to rapid prototyping of pre-processing pipelines.

import nvidia.dali.experimental.dynamic as ndd

model = MyModel(...)
flip_horizontal = True
flip_vertical = False
dataset = ndd.readers.File(file_root=images_dir)
for batch in dataset.next_epoch(batch_size=16):
    img = ndd.decoders.image(batch, device="mixed")
    flipped = ndd.flip(img, horizontal=flip_horizontal, vertical=flip_vertical)
    model((flipped, img))

See Overview to learn more.

• Overview
  • How it works
• API Reference
  • Tensor and Batch objects
  • Execution context
  • Random state
• Data Readers Reference
• Operation Reference