🧠 Computer Vision & NLP

A collection of deep learning projects exploring computer vision and natural language processing (NLP) using PyTorch and advanced data pipelines. Includes a CNN model trained on CIFAR-10 for image classification and a sentiment analysis model built with the Sentiment140 Twitter dataset.

📋 Project Overview

This repository contains two fully independent deep learning projects, each showcasing different areas of AI:

1. CNN for Image Classification (CIFAR-10)

   • Builds and trains a Convolutional Neural Network (CNN) from scratch using PyTorch.
   • Focuses on classifying images into 10 different categories.
   • Uses custom data loading, augmentation, and model evaluation.

2. Twitter Sentiment Analysis

   • Uses the Sentiment140 dataset (1.6M tweets) to classify tweets as Positive or Negative.
   • Implements full NLP preprocessing pipeline: cleaning, tokenization, lemmatization.
   • Applies deep learning for text sentiment classification and evaluates multiple metrics.

🖼 Project 1: CIFAR-10 - CNN Image Classification

Dataset Overview

The CIFAR-10 dataset contains 60,000 32x32 color images divided into 10 classes, with 50,000 training and 10,000 test images.



The project loads CIFAR-10 batches, normalizes images, and implements data augmentation using torchvision.transforms.

Model Architecture

The CNN was built from scratch in PyTorch and includes:

• Convolutional layers with ReLU activation
• MaxPooling for spatial downsampling
• Fully connected layers for classification
• Softmax output layer

Optimizer: Stochastic Gradient Descent (SGD) with CosineAnnealingLR scheduler.

Training Pipeline

• Custom DataLoader with PyTorch
• Data augmentation: random crops, flips, normalization
• Epoch tracking and loss visualization
• Confusion matrix and per-class accuracy

💬 Project 2: Twitter Sentiment Analysis

Dataset Overview

The project uses the Sentiment140 dataset downloaded via kagglehub. It contains 1.6 million tweets, each labeled as:



Columns in the raw CSV:

• target → Sentiment label (0 or 4)
• ids → Tweet ID
• date → Date of tweet
• flag → Query flag
• user → Username
• text → Tweet content

Preprocessing Steps

• Remove special characters, URLs, and mentions (@usernames)
• Convert text to lowercase
• Tokenization with NLTK
• Lemmatization and stemming
• Train/test split with StratifiedShuffleSplit

Model Pipeline

• Built using PyTorch with fully connected layers
• Binary classification (Positive vs Negative)
• Evaluation metrics: Track training/validation & Evaluate on the test set → use eval_on to measure final accuracy.

📊 Results & Metrics

Project	Accuracy
🖼 CIFAR-10 CNN	🎯 94.25%
💬 Twitter Sentiment	🎯 77%

# Author Info

# LinkedIn: https://www.linkedin.com/in/profile-mariana-martins/
# GitHub: https://github.com/marianamartiyns
# Email: marianamartiyns@gmail.com

Note

The notebooks comments are written in Portuguese, while variable names and code remain in English for clarity.