Training of Deep Learning Pipelines on Memory-Constrained GPUs via Segmented Fused-Tiled Execution.

Training models with massive inputs is a significant challenge in the development of Deep Learning pipelines to process very large digital image datasets as required by Whole Slide Imaging (WSI) in computational pathology and analysis of brain fMRI images in computational neuroscience. Graphics Processing Units (GPUs) represent the primary workhorse in training and inference of Deep Learning models. In order to use GPUs to run inference or training on a neural network pipeline, state-of-the-art machine learning frameworks like PyTorch and TensorFlow currently require that the collective memory on the GPUs must be larger than the size of the activations at any stage in the pipeline. Therefore, existing Deep Learning pipelines for these use cases have been forced to develop sub-optimal "patch-based" modeling approaches, where images are processed in small segments of an image. In this paper, we present a solution to this problem by employing tiling in conjunction with check-pointing, thereby enabling arbitrarily large images to be directly processed, irrespective of the size of global memory on a GPU and the number of available GPUs. Experimental results using PyTorch demonstrate enhanced functionality/performance over existing frameworks.

Retrofitting impervious urban infrastructure with green technology for rainfall-runoff restoration, indirect reuse and pollution load reduction.

The built environs alter hydrology and water resource chemistry. Florida is subject to nutrient criteria and is promulgating "no-net-load-increase" criteria for runoff and constituents (nutrients and particulate matter, PM). With such criteria, green infrastructure, hydrologic restoration, indirect reuse and source control are potential design solutions. The study simulates runoff and constituent load control through urban source area re-design to provide long-term "no-net-load-increases". A long-term continuous simulation of pre- and post-development response for an existing surface parking facility is quantified. Retrofits include a biofiltration area reactor (BAR) for hydrologic and denitrification control. A linear infiltration reactor (LIR) of cementitious permeable pavement (CPP) provides infiltration, adsorption and filtration. Pavement cleaning provided source control. Simulation of climate and source area data indicates re-design achieves "no-net-load-increases" at lower costs compared to standard construction. The retrofit system yields lower cost per nutrient load treated compared to Best Management Practices (BMPs).