Deep Learning Challenges Repository

This repository houses challenges from the Artificial Neural Networks and Deep Learning course at Politecnico di Milano, Italy. It contains two main challenges: Leaf Health Classification and Time Series Forecasting, each stored in separate directories.

Repository Structure

• challenge-1/: Leaf Health Classification
  • src/: Source code including Jupyter notebooks and Python scripts for leaf health classification.
  • docs/: Report and related imagery for the leaf classification project.
• challenge-2/: Time Series Forecasting
  • src/: Jupyter notebooks detailing various advanced deep learning models for time series forecasting.
  • docs/: Comprehensive report detailing methodologies and findings of the time series project.

Each folder is structured to contain all necessary code and documentation relevant to each challenge.

Challenge 1: Leaf Health Classification

This challenge involves classifying images of leaves as healthy or unhealthy using convolutional neural networks and other advanced deep learning techniques. Techniques such as data augmentation, various CNN architectures, and ensemble models were explored.

Highlights

• Use of advanced image preprocessing and augmentation techniques to improve model performance.
• Exploration of various CNN architectures including MobileNet, EfficientNet, and ConvNeXt.
• Implementation of test-time augmentation and ensemble methods to enhance prediction accuracy.

Challenge 2: Time Series Forecasting

The second challenge focuses on forecasting future values in time series data across various categories using state-of-the-art models like LSTMs, GRUs, Attention mechanisms, Transformers, and TSMixer architectures.

Highlights

• Detailed analysis and preprocessing of time series data to understand underlying patterns and structures.
• Experimentation with various deep learning models including bidirectional LSTMs, GRUs, and Transformer-based models.
• Best performing model utilized the TSMixer architecture, showcasing advanced approaches in handling time series data.

Installation and Usage

Instructions for setting up and running the projects:

git clone [email protected]:biromiro/polimi-an2dl-challenges.git
cd polimi-an2dl-challenges

Contributors

• Francesco Caserta - [email protected]
• Nuno Costa - [email protected]
• Shodai Fujimoto - [email protected]
• Rio Ishibashi - [email protected]

Make sure to use the correct email addresses based on your institution's or project's email scheme or the personal preferences of the contributors. If the actual email addresses differ, replace them accordingly. Each team member contributed across all aspects of both challenges, ensuring a collaborative effort that enriched the learning experience.

License

This project is licensed under the MIT License - see the LICENSE.md file for details.

Acknowledgements

• Course instructors and TA’s at Politecnico di Milano.
• Open-source libraries and frameworks that facilitated the model developments.