kumarchandan / EVA4-TSAI

Solutions to the assignments in the EVA4-TSAI course

EVA4 - TSAI

This repository contains the solutions to the assignments of the EVA4 course conducted by The School of AI

Contents

Session 4 - Architectural Basics

Achieve 99.4% test accuracy on the MNIST dataset with a CNN using less than 20,000 parameters and 20 epochs.

Session 5 - Coding Drill Down

Achieve 99.4% test accuracy on the MNIST dataset with a CNN using less than 10,000 parameters and 15 epochs.

Session 6 - Batch Normalization and Regularization

Apply L1 and L2 regularization on Session 5 code and compare the models.

Session 7 - Advanced Convolutions

Achieve 80% test accuracy on the CIFAR-10 dataset with a CNN using less than 1M parameters and make use of advanced convolutions.

Session 8 - Receptive Fields and Network Architecture

Achieve 85% test accuracy on the CIFAR-10 dataset with ResNet18 using image augmentation.

Session 9 - Data Augmentation and GradCAM

Achieve 87% test accuracy on the CIFAR-10 dataset with ResNet18 using albumentations for image augmentation. Implement GradCAM to visualize heatmaps.

Session 10 - Training and Learning Rates

Achieve 88% test accuracy on the CIFAR-10 dataset with ResNet18. Implement Leslie Smith LR finder to find the best starting learning rate and use ReduceLROnPlateau.

Session 11 - Super Convergence

Achieve 90% test accuracy on the CIFAR-10 dataset with a custom ResNet architecture and One Cycle Policy for faster convergence.

Session 12 - Object Localization

Achieve 50% test accuracy on the tiny-imagenet-200 dataset. Make a custom dataset of 50 dogs and find optimal template anchor boxes.

Session 13 - YoloV2&3

Perform object detection using OpenCV Yolo implementation. Create a custom dataset of a unique object not present in the existing COCO classess and train YoloV3 on the custom class.

Session 14-15 - RCNN

Create a custom dataset for monocular depth estimation and segmentation simultaneously.

Session 15 - Transfer Learning

Use the dataset created in Session 14 and build a monocular depth estimation and segmentation model.