Extensive AI Reimagined Advanced

This repository contains a collection of advanced AI and Machine Learning projects, showcasing various techniques and applications. Below is a summary of key folders and their contents:

Session 07: MNIST Digit Classification

Iterative optimization of a CNN model for MNIST digit classification

Technologies:

• Framework: PyTorch
• Key libraries: torchvision, matplotlib

Methodologies:

• Convolutional Neural Networks
• Progressive model refinement
• Regularization techniques (Batch Normalization, Dropout)
• Data augmentation
• Learning rate optimization
• Architecture modifications (Global Average Pooling)

Focus: Achieving high accuracy (>98%) with a compact model (<8K parameters) through systematic iterations and performance analysis.

Session 08: Image Classification

To study and compare different normalization techniques (Group Normalization, Layer Normalization, Batch Normalization) applied to convolutional neural networks (CNNs) for image classification on the CIFAR-10 dataset.

Technologies:

• Framework: PyTorch
• Key libs: torchvision, matplotlib, numpy

Methodology:

• Implemented custom CNN architectures with various normalization layers
• Trained models using SGD optimizer and CrossEntropyLoss
• Evaluated performance using accuracy metrics and loss curves
• Visualized misclassified images for error analysis

Focus:

• Comparing the effectiveness of different normalization techniques
• Optimizing model architectures to achieve >70% accuracy with <50,000 parameters
• Analyzing the impact of normalization on model convergence and generalization
• The project demonstrates a systematic approach to evaluating normalization techniques in deep learning, with a focus on practical implementation and performance analysis.

Session 11: CIFAR10 Image Classification with ResNet18

Objective: Implement CIFAR10 image classification by importing ResNet18 model from another repo.

Key Components:

• Model: ResNet18 architecture
• Dataset: CIFAR10
• Training: Custom training loop with 20 epochs
• Evaluation: Test set accuracy and loss tracking
• Visualization: Loss plots and misclassified images display
• Model Persistence: Saving and loading capabilities

Tech:

• PyTorch

Methodology:

• Utilizes a pre-defined ResNet18 architecture
• Trains the model on CIFAR10 data
• Implements custom training and evaluation loops

The project demonstrates a complete machine learning workflow for image classification, from model training to result visualization and model persistence.

Session 13: Pytorch-Lightning on CIFAR10

The code demonstrates image classification on CIFAR10 dataset using PyTorch Lightning

Technology:

• PyTorch Lightning

Methodology:

• Interactive data visualization and model training
• Utilizes progress bars and other UI elements for real-time feedback
• Modular approach with extensive use of widget models and layouts

Focus:

• Creating an interactive and user-friendly interface for machine learning tasks
• Emphasis on real-time progress tracking and data presentation
• Involves training, evaluating models on image datasets

The code tries to create a rich, interactive environment for machine learning experiments or demonstrations.

Session 19: Building GPT: Text generator

Implement a small-scale GPT (Generative Pre-trained Transformer) model for text generation, trained on Shakespeare's works.

Technology:

• PyTorch

Methodology:

• Character-level tokenization
• Transformer architecture with self-attention mechanisms
• Training on a corpus of Shakespeare's text

Key Components:

• Data preprocessing and tokenization
• Custom PyTorch modules for the transformer architecture
• Training loop with gradient descent optimization
• Text generation functionality

Focus:

The project emphasizes understanding and implementing core concepts of transformer-based language models on a smaller scale. It serves as an educational tool for learning about GPT-like architectures and their application to text generation tasks.