Skip to main content

Week 3 | Day 10: MNIST Training using HPCC Cluster

 Todo: 

- Train MNIST data on HPCC Cluster using given example code

Jun 5, 2023

Labels:

Comments

Post a Comment

Popular posts from this blog

Enhancing Performance of Distributed Neural Network with GNN Bundle - Project Poster and Videos

Poster contents Background Deep learning implementation demands are rising in various fields, including manufacturing, medicine, art, history, etc. The high demand for powerful graphics cards has led to shortages. Generalized Neural Network (GNN) being one of the choices for distributed neural network training; we want to improve it to have better control over machine utilization and optimize its performance. Within GNN, during backpropagation, the time required for each iteration is equal to the time for the weakest node due to the aggregation of loss required at the end of every iteration Research Questions Can we improve resource usage reduce training time required for deep neural network training using the more powerful machine? Method Adopted the notion of training with a restricted node count Enhanced the Fit function to accommodate nodeNumber parameter. Extracted data aligned with the chosen node count. Conducted training and weight synchronization on chosen nodes. Redistributed...
Post a Comment
Read more

Week 9: Weekly work summary

Project: Distributed Neural Network Training and Prediction Mentors: Roger Dev, Lili Xu Week 9: Weekly work summary (July 17-21) During this week, my primary focus was on conducting a comparative study between nNode and single node optimization methods. To carry out this analysis, I performed training for both classification and mnist tasks with different node configurations, specifically using node=all, 1, 2, and 3. However, upon reviewing the training time data from my local ubuntu machine and university cluster machine, it became evident that the recorded times did not align with our initial expectations. Roger and I collaborated to investigate this issue, but unfortunately, we were unable to pinpoint the exact cause. This matter remains unresolved and requires further attention. Alongside the comparative study, I dedicated time to explore an alternative approach for multi-node neural network training. Our current implementation involves a functional library for multi-node neural ne...
Post a Comment
Read more

Week 10: Weekly work summary

Project: Distributed Neural Network Training and Prediction Mentors: Roger Dev, Lili Xu Week 10: Weekly work summary (July 24-28) During this week, I participated in a short demo on utilizing ecl watch for code analysis, specifically focusing on analyzing time for each segment of the code. Gordon Smith provided us with a helpful walkthrough on effectively using Filters and Cross Tab from Metrics page to gather valuable analytics information. Additionally, he demonstrated the ECL tab, offering a brief overview of the time associated with each code block and its color code. I learned that we have the capability to update graph labels through the Options -> Graph tab. This session was highly informative and enhanced my understanding of the analytics available on ECL Watch. Following the demonstration, I devoted my time to analyzing the performance of the single node and nNodeFit Functions. Notably, I observed that FitBatch did not exhibit any time for a single node. To address this, I ...
Post a Comment
Read more