How dasotse/deeplearning works
A. Project Type and Purpose
This is an educational Deep Learning module with two layers
• A set of teaching notebooks (introductory → advanced) that guide students through AI, neural networks, and advanced topics.
• A final “Práctica Final” project: a hybrid CNN+metadata model to predict tourist POI engagement.
B. File Structure Analysis
Top-level
• README.md – project overview, goals, deliverables
• Proyecto.zip – bundled dataset and code scaffolding for the final exercise
• notebooks/ – sequential Jupyter notebooks organized into three folders:
• 1. Introducción
• 2. Redes Neuronales
• 3. Redes Neuronales Avanzadas (plus module_utils.py)
C. Key Insights from README
Data: Images + structured metadata + engagement metrics
3. Deliverables: reproducible notebook (+ requirements), PDF report, trained model, preprocessing scripts
Compute Environment
• Google Colab / Local Jupyter + GPU
• Dependencies (requirements.txt / environment.yml)
Data Ingestion & Storage
• Proyecto.zip → unzipped into folders: images/, metadata.csv, engagement.csv
Data Preparation & EDA
• Preprocessing scripts (module_utils.py and student scripts)
• Image transformations (resize, normalize)
• Metadata cleaning, feature encoding
Feature Extraction Pipelines
• Visual branch: CNN backbone (e.g., pre-trained model, custom layers)
• Contextual branch: MLP or embedding layers for metadata
Model Architecture
• Fusion layer combining CNN outputs + metadata embeddings
• Classifier head → high vs. low engagement
Training & Optimization
• Training loop (batch generation, loss & optimizer, callbacks)
• Anti-overfitting (dropout, early stopping)
Evaluation & Analysis
• Metrics calculation (accuracy, confusion matrix, ROC)
• Visualization notebooks for results
Outputs & Deliverables
• Final model weights (.h5 or .pt)
• Reproducible notebook(s) with seeded experiments
• Technical report (PDF)
• Utility scripts
E. Relationships and Data Flows
• Convergence into the Model component
• Feedback loop from Evaluation back to Training (hyperparameter tuning)
F. Architectural Patterns & Design Principles
• Pipeline/Layered architecture: clear separation of ingestion, preprocessing, modeling, evaluation
• Modularity: notebooks vs. utility scripts vs. data
• Reproducibility: environment files, random seeds, documented pipeline steps
External dependencies: Google Colab and Google Drive for dataset access.
H. Diagramming Guidelines
Decisions: e.g., “Transfer Learning CNN,” “One-hot encoding.”
I. Emphasis
– Don’t over-complicate: model the project as discrete stages (Ingest → Prep → Feature → Train → Evaluate → Deliver).