This project is an educational project made to discover p5.js and tensorflow.js environment.
The project contains two main pages :
- /train.html : Train neural network using prepared dataset. Tensorflow-vis allows to follow neural network loss per batch and per epoch.
- /map.html : Get lastest daily propagation data, use trained model to predict for the next and display probabilities on map.
- The dataset file (sampled_covid_dataset) has been created with scripts/get_dataset.py. This script get historical data from API (on 120 days) and construct a dataset containing the 14th previous hospitalizations values and the direction of the next step (1 or -1). The dataset needed some resampling because it was largely imbalanced using only 120 days.
- The neural network is equivalent to a simple Logistic Regression using 14 variables, with Gradient Descent Optimization.
- The implemented algorithm accuracy is near random.
- Get larger dataset.
- Use LSTM cells / Conv-LSTM architecture to detect long term patterns
- Store propagation data in database
- Use train / test sets for training and store model metrics for optimization
- Display only France map and not all world.
- Use different data sources (population movements, temperature, news feeds...)
- Handle Covid French API no-cors
- Change starting day at 20:00 pm for API calls