laverdes / multivariate-time-series-classification Goto Github PK

• re-train model with lenghts and coords features.

The code under the Quick Tour section in the README should be executable, if not, then the smallest necessary changes/additions should be done to the code. A good reference si the notebook tinder use-case

When the notebook is uploaded to Github, plots coming from plotty are not displayed (HTML). This does not happen in the colab notebook. The idea with this Issue is to make the notebook and the colab notebooks work properly displaying the time-series figures coming from plotty. For more details and the link to the colab please write a comment in this Issues' discussion.

• finish README.md sample execution flow, add notebook description
• add application and purpose of this repo
• what matters the most? No fp at all? significance test
• add result plots for tested models and summarize performance analysis with ROC curves.
• add further work as an extra issue or in README.md

• clean blinking values from data. dataset.py should be able to produce several versions of the data.
• apply time-series transformations

• add argument parser to aggregator.py
• clean feature_engineering.py
• debug model.py
• build pipeline in classifier.py
• sample execution with minimum parameters (cpu)

tabularization, rocket and feature-extractor share many of the execution logic. Code can be improved by adding a general training loop and just setting a config for it in each of the ifs in the class method train() in classifying.py.

Add better documentation for the toolbox in README

Some figures in the notebook are displayed twice. The return for plot_outliers_in is a figure that should be wrapped around plt.show() to avoid this.

• torch use deterministic algorithms
• DataLoader use generator_fn
• CUDA force determinism
• (google colab particularities, ...)