abhishek-kathuria / CIFAR100-Image-Classification

Classify CIFAR-100 images using CNN, ResNet and transfer learning using PyTorch

CIFAR 100 Image Classification using Deep Learning

This repository contains the code for implementation of CIFAR-100 classification using various models such as Linear Classifier, Convolutional Neural Net (CNNs) from scratch, Residual Neural Nets (ResNet) from scratch, Pretrained ResNet-18 and Pretrained ResNet-50.

Requirements and How to use

The following are the basic requirements needed:

• Anaconda
• Pytorch
  

For Executing the code directly from the Github Repo:

• Download the repo (make sure you have fulfilled the requirements)
• You can either run the code on python 3.6 or on the Google Colab GPU.
• The Dataset will be automatically downloaded from the torchvision.datasets.CIFAR100 library

Files

The usage of the following python files is given as follows:

• Linear_Classifier contains an Artificial Neural Net with two hidden layers.
• CNN contains a four layered CNN built from scratch.
• RESNET contains a custom ResNet with two residual blocks.
• PRETRAINED_RESNET contains pretrained ResNet-18 and ResNet-50 architectures fine tuned on the given dataset.

Key Learnings

Through this project, I have learned the following:

• Building an ANN and a Convolutional NN from scratch.
• Build custom ResNet architectures from scratch.
• Pretraining and transfer learning.
• Identifying the maximum learning rate.
• Learning Schedulers such as OneCycleLR and CosineAnnealingLR and their needs.
• Usage of different optimizers such as SGD and ADAM.
• Various Data Augmentation techniques such as random crop, random split and random rotate.
• Regularization methods such as L1 and L2 regularization, dropout, and early stopping.