Efficient Distributed Inference of Deep Neural Networks via Restructuring and Pruning

Authors: Afshin Abdi, Saeed Rashidi, Faramarz Fekri, Tushar Krishna

AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental To evaluate the performance of the Re Purpose framework, we consider different DNN architectures and compare the accuracy, communication, and wall-clock times of the proposed framework to the following approaches; (1) naive implementation where the input data (or locally computed features) are communicated to all nodes in the network, so they all have the entire input data and process the entire deep model locally. [...] The results are shown in Fig. 7 for P = 6 sensors. [...] Figures 7a and 8a compare the performance of Re Purpose with the baseline, naive sparsification, and model distillation.
Researcher Affiliation Academia School of Electrical and Computer Engineering, Georgia Institute of Technology
Pseudocode Yes Algorithm 1: Re Purpose algorithm for a single layer
Open Source Code No The paper mentions ASTRA-sim is 'an open-source distributed Deep Learning platform simulator' but does not provide a statement or link for the open-sourcing of their own proposed methodology (Re Purpose) code.
Open Datasets Yes Next, we consider a network of P sensors where each sensor observes an image of a digit xi (from MNIST dataset)
Dataset Splits No The paper does not explicitly provide specific percentages or counts for training, validation, and test dataset splits. It mentions 'fine-tuning' or 'post-training' but without specifying how the data was partitioned for these stages.
Hardware Specification Yes Compute Memory Bandwidth Datacenter 125 TOPS 32GB 150 GB/s (NVLink) Edge 0.5 TOPS 1GB 100 MB/s (Ethernet)
Software Dependencies No The paper mentions 'ASTRA-sim' and 'NVIDIA NCCL' but does not provide specific version numbers for these software components or any other libraries or frameworks used in the experiments.
Experiment Setup Yes We applied Re Purpose with η1 = 0, η2 {0.01, 0.1} (figures 4(c)-(d)).