How fab16bsb/image_classification works
Project Type and Purpose
• This is a Machine Learning (Deep Learning) Pipeline for image classification using Convolutional Neural Networks (CNNs), built with Python/Keras.
• Core use cases:
• Data preparation (scraping, organizing images)
• Model creation & training
• Model loading & inference
• Explainability (feature maps, saliency maps)
File Structure
• Top–level directories:
• code/: scripts & notebooks for training (Creer_model.ipynb, creer_model.py) and inference (Load_model.ipynb, load_model.py)
• data/: three dataset subfolders (dragonball, fairytail, pokemon), each with train/, validation/, predict/
• modele/: stored model artifacts (architecture JSON, weights .h5, labels.txt)
• Other files:
• requirement.txt: dependency list
• README files, LICENSE
Key Insights from the README
• Three datasets, four classes each
• Data augmentation steps using Keras’ ImageDataGenerator
• CNN architecture: convolutional layers → pooling → dense layers
• Explainability modules: feature‐map and saliency‐map generation
• No external web service or UI—everything runs locally via CLI or notebooks
Main Components and Their Relationships
a. Data Sources & Storage
• Raw images on local filesystem (data/<dataset>/{train,validation,predict})
• Data augmentation module reads from disk, produces augmented batches
b. Training Pipeline
• Data Loader & Augmentor (creer_model.py / Creer_model.ipynb)
• CNN Trainer (Keras model definition → compile → fit)
• Model Export: save architecture JSON + weights .h5 + labels.txt → modele/<dataset>_model/
c. Inference Pipeline
• Model Loader (load_model.py / Load_model.ipynb)
• Prediction Module: loads saved model + labels → runs inference on data/<dataset>/predict → writes results to result/
d. Explainability Module
• Hooks into trained model to extract feature maps & saliency maps → outputs visualization images under result/
e. Environment & Dependencies
• Python 3.12, Keras, listed in requirements.txt
• (Optional) GPU accelerator block for model training
Architectural Patterns & Design Principles
• Layered Architecture: clearly separate data ingestion, training logic, inference logic, and visualization.
• Modularity: each dataset (Dragon Ball, Fairy Tail, Pokémon) follows the same pipeline steps—use a parameterized script.
• Artifact Repository: local file‐based model registry under modele/, versionable via Git.