Best preprocessing for presentation datasets on NVIDIA GPUs

AI & Data Science Deep Learning (Training & Inference) TensorRT
1

Hi,

I am working on a small deep learning dataset and most of my source material is presentation files in PPT/PDF format.

Some of these files are public presentations that I downloaded from SlideShare using a SlideShare Downloader. The problem is that after converting the files to text/images, some documents are fine but some have missing text, strange characters or different page layouts.

I am planning to use the extracted data for training, so I want to make the preprocessing more consistent before I start the actual training.

My current workflow is:

PPT/PDF → extract content → clean/preprocess → dataset → deep learning training

I am not sure if I should first convert every presentation page to an image and process it with a vision model, or extract the text and keep it as text data.

For this type of presentation dataset, what preprocessing approach would you recommend on NVIDIA GPUs?

Also, is there any NVIDIA tool or library that would make this PDF/image preprocessing faster?

Thanks.

4

Hi @JosePhillip,

You didn’t have a Topic title. I took the liberty and generated one for you.

Thanks!

5

Thanks for fixing the topic title. I appreciate your help!

6

I am also checking if the issue comes from the downloaded presentation files or the preprocessing step. Files from SlideShare Downloader open fine, but some text gets messy after extraction. Maybe someone in this NVIDIA community has faced something similar with PDF or PPT datasets.

7

For presentation datasets with mixed text and layout issues, NVIDIA recommends a hybrid extraction approach using the NeMo Retriever library, which preserves document structure (tables remain tables, figures stay positioned) while converting PDFs to structured, queryable units. This is more reliable than raw text extraction for presentations with complex layouts.

When to use each approach:

  • Image-based (vision models): Use when presentations contain critical visual elements (charts, diagrams, handwritten notes) or when text extraction produces garbled output. Convert problematic pages to images at high DPI (e.g., 300 DPI) and process with vision models.
  • Text extraction: Use for text-heavy presentations where layout preservation is less critical.
  • Hybrid (recommended): Extract text where reliable using NeMo Retriever; convert only problematic pages to images for vision-based processing.

NVIDIA tools for acceleration:

  • NVIDIA DALI (Data Loading Library): GPU-accelerated library for data loading and preprocessing. It provides optimized building blocks for loading and processing image, video, and audio data, and can replace standard data loaders in PyTorch and TensorFlow pipelines.
  • NeMo Retriever: Specifically designed for document extraction and structure preservation, handling complex PDFs and presentations.

Recommended workflow:

  1. Use NeMo Retriever to extract and structure your presentation content
  2. Apply DALI for GPU-accelerated image preprocessing on converted pages
  3. Profile your pipeline to identify bottlenecks, then apply GPU acceleration selectively

Start with your current workflow and measure where time is spent; this will guide which components benefit most from GPU acceleration.