Related Experiment Video
Updated: Sep 3, 2025
10:25
Deep Learning-Based Segmentation of Cryo-Electron Tomograms
Published on: November 11, 2022
9.3K
Training of Deep Learning Pipelines on Memory-Constrained GPUs via Segmented Fused-Tiled Execution
Yufan Xu1, Gerald Sabin2, Saurabh Raje1
1University of Utah, Salt Lake City, Utah, USA.
Compiler Construction : ... International Conference, CC ... : Proceedings. CC (Conference)
|July 25, 2022
Summary
Training deep learning models with large images like those in computational pathology and neuroscience is challenging. This study introduces a tiling and checkpointing method to process arbitrarily large images directly, overcoming GPU memory limitations.
Area of Science:
- Deep Learning
- Computational Pathology
- Computational Neuroscience
Background:
- Training deep learning models on massive datasets, such as Whole Slide Imaging (WSI) and fMRI data, faces significant memory challenges.
- Current GPU memory constraints necessitate suboptimal patch-based approaches for processing large-scale image data.
Purpose of the Study:
- To present a novel method for directly processing arbitrarily large images in deep learning pipelines.
- To overcome the limitations of GPU memory size and number of GPUs in training and inference.
Main Methods:
- Employing a combination of tiling and checkpointing techniques.
- Enabling direct processing of large images irrespective of GPU memory capacity.
Main Results:
- Experimental results using PyTorch demonstrate enhanced functionality and performance compared to existing frameworks.
- The proposed method allows for direct processing of large images, eliminating the need for patch-based segmentation.
Conclusions:
- The tiling and checkpointing approach effectively addresses the memory limitations in deep learning for large-scale image analysis.
- This method enhances the capability to process massive datasets in fields like computational pathology and neuroscience.
