Applied Deep Learning : Convolution Neural Network

         

This repo contains source codes for, Building Intuition for Convolution Neural Networks step by step.

Session-01

• Machine Learning Intuition, Background & Basics
• Python 101 for Machine Learning
• blog

Session-02

• Convolutions, Pooling Operations & Channels
• Pytorch 101 for Vision Machine Learning
• blog

Session-03

• Kernels, Activations and Layers
• blog

Session-04

• Architectural Basics suitable for our objective
• MNIST model training
  • parameters used 13,402
  • epochs=20
  • highest test accuracy = 99.46%, epoch = 19th
  • notebook link

Session-05

• Receptive Field : core fundamental concept
• MNIST model training
  • parameters used 7808
  • epochs=15
  • highest test accuracy = 99.43%, epoch = 11th
  • notebook link

Session-06

• BN, Kernels & Regularization
• MNIST model training
  • using L1/L2 regularization with BN/GBN
  • BN : batch normalization
  • GBN : ghost batch normalization
  • best model : BN with L2
    • parameters used 7808
    • epochs=25
    • highest test accuaracy = 99.54%, epoch = 21st
  • notebook link

Session-07

• Advanced Convolution
• Achieve an accuracy of greater than 80% on CIFAR-10 dataset
  • architecture to C1C2C3C40 (basically 3 MPs)
  • total params to be less than 1M
  • RF must be more than 44
  • one of the layers must use Depthwise Separable Convolution
  • one of the layers must use Dilated Convolution
  • use GAP
• Result
  • parameters used 220,778
  • epochs = 20
  • highest test acc = 85.55%
  • notebook link

Session-08

• Receptive Fields and Network Architectures
• Achieve an accuracy of greater than 85% on CIFAR-10 dataset
  • architecture ResNet18
• Result
  • parameters : 11,173,962
  • epoch : 50
  • training acc : 98.65%
  • testing acc : 89.78%
  • notebook link

Session-09

• Data Augmentation
• Achieve an accuracy of greater than 87% on CIFAR-10 dataset
  • Move transformations to Albumentations.
  • Implement GradCam function.
• Result
  • parameters : 11,173,962
  • epoch : 50
  • testing acc : 92.17%
  • work link

Session-10

• Advanced Concepts : Training and Learning Rates
• Achieve an accuracy of greater than 88% on CIFAR-10 dataset
  • Add CutOut augmentation
  • Implement LR Finder (for SGD, not for ADAM)
  • Implement ReduceLROnPlateau
• Result
  • parameters : 11,173,962
  • epoch : 50
  • testing acc : 89.80%
  • work link

Session-11

• Super Convergence
• Achieve an accuracy of greater than 90% on CIFAR-10 dataset
  • 3Layer-Network
  • Implement One Cycle Policy
• Result
  • parameters : 6,573,130
  • epoch : 24
  • testing acc : 91.02%
  • work link

Session-12

• Object Localization : YOLO (Recognition)
• Use TinyImageNet dataset, create custom data loader with 70/30 split.
• Achieve an accuracy of greater than 50% on TinyImageNet dataset
  • ResNet18
  • One Cycle Policy
• Result
  • parameters : 11,173,962
  • epoch : 30
  • testing acc : 58.35%
  • work link