Deep Learning for Time Series: LSST Classification

Project Description

This repository contains the source code for our evaluation project for the "Deep Learning for Time Series" course.

The main objective is to perform a Time Series Classification task on the multivariate astronomical dataset LSST (Large Synoptic Survey Telescope).

In accordance with the project guidelines, we chose the following setting:

1. Pre-training (Forecasting): Training an encoder model on a forecasting task using exclusively the Informer family datasets (ETT, Electricity, Weather, etc.).
2. Adaptation (Classification): Transferring the pre-trained encoder and fine-tuning it on the LSST dataset for the 14-class classification task.

💡 Where to start: > We highly recommend reviewing the main_LSST_classification.ipynb notebook first. It provides a step-by-step walkthrough of our scientific approach, including exploratory data analysis, the ROCKET baseline evaluation, and the final results of our Transfer Learning architecture.

Repository Structure

LSST-TimeSeries-DL/
│
├── data/
│   └── informer/                    
│                  
├── src/                          
│   ├── __init__.py           
│   ├── dataset.py              
│   ├── model.py                   
│   └── preprocessing.py           
│
├── baseline.py                     
├── pretrain.py                
├── finetune.py                       
│
├── main_LSST_classification.ipynb  
│
├── pretrained_encoder.pth         
├── requirements.txt               
├── report.pdf        
└── README.md