Overview of the my Updates

The architecture of the ImpactNet

  • Focusing on MobileNet
  • And Masking of Soiling Using Basic Image Processing Techniques instead of MFCNN

Edge Computing-Models Overview

  • Ran Resnet on CPU, GPU and pi
  • Ran MobileNet on CPU and GPU
  • Benchmarked all results
  • To Showcase why MobileNet is the best
Type of Model Device Time Taken for Inferencems Time Taken for PreProcessing -ms Total Time
ms
ResNet-50 CPU 119.603 1.639 121.289
Resnet-50 GPU 9.84 1.624 11.45
MobileNet CPU 43.9 1.646 45.549
MobileNet GPU 3.75 1.58 5.33
ResNet Pi 2.26 0.774 3.03

Using FPGAs

  • Go behind a much more comprehensive edge computing-based project 
  • Initial results of running image processing algorithms
  • Problems being faced by FPGA
    • ARM+PL side
    • PL side stopped working
    • Looking into new alternatives like Intel Devcloud

Edge TPU-Coral Board by Google 

  • It is Google's ASIC board which uses TPU 
  • Utilises quantisation of models to optimise for the edge 
  • Ran MobileNet and ResNet using edgeTPU
  • Results:
Name of the Model Total Inference Time-ms For PreProcessing-ms Total Time-ms
ResNet-50 0.537 0.277 0.914
MobileNet

Contents of the paper

  • Introduction 
  • Previous Work
    • Power Loss Estimation
    • Economic Solutions to monitor a panel
    • Hotspot Formation and its effects
  • Data
  • Deep Learning
    • Models Used
  • Edge Computing
    • Use cases and Advantages
    • Edge Devices on Solar Farms
    • Pretrained Models for the Edge
      • MobileNet
      • Resnet

Ideal Timeline of the Project

1- Finish Writing the Paper-by Dec end-Jan first week

2-Masking

3- Editing?

4- FPGA Part?

Resync-Work Updates

  • Built models for daily, hourly, weekly predictions
  • Built Models for hourly and 15 min data
  • Found LSTM models to be performing the best
  • Trying to understand how the model learns-attention applied-did it with FNN
  • Transformer based architecture for energy-based system-Could this be my work ahead?

Other Updates

  • Google DevFest Bandung, Indonesia
  • Instructing a course for Udacity and Intel- AI on the edge-Link
Made with Slides.com