Code of the paper "Multi variables time series information bottleneck"
Multi variables time series information bottleneck
Denis Ullmann, Olga Taran, Slava Voloshynoskyy
University of Geneva
SNSF NRP75 Grant
This repository contains all the necessary information to replicate the results obtained in the paper "Multi variables time series information bottleneck"
The part of the code for partial convolution layers from https://github.com/MathiasGruber/PConv-Keras/blob/master/requirements.txt is updated to work on TensorFlow2.
The code is dedicated to Multiple Time Series (MTS) forecasting and includes evaluations with several metrics:
- MTS metrics: MAE, MAPE, RMSE
- CV metrics: PSNR, SSIM
- Information Theory metrics: Mutual Information I, KL-divergence when the time profiles of the MTS are already clustered (eg. for IRIS data).
- Astrophysics metrics: profile and activity classification error when the MTS data is already labeled or the time profiles of the MTS are already clustered (eg. for IRIS data).
Different concurrent models are evaluated and their codes are present in the repository:
- IB-MTS that makes use of UNet with partial convolution
- LSTM, GRU
- IB-LSTM, IB-GRU that are made of 2 LSTMs or GRUs to encode and 2 LSTMs or GRUs to decode
- NBeats
- Requirements
- Install
- Download data
- Download weights
- Download test results
- Usage
- Brief explanation
- License
Check your available memory, used and downloaded datasets may be significantly big.
Preprocessings, trainings, testings and figure generation were performed on a machine with more that free 15GB and CUDA capable GPUs with more than 10GB RAM each.
You harware and software should be TensorFlow2-friendly (see requirements at TF installation)
You may create a virtual environnement and use requirements.txt:
$ pip install -r requirements.txt
Use Git to clone this repository into your computer.
git clone https://github.com/DenisUllmann/IB-MTS
Or pip install from Git:
pip install git+https://github.com/DenisUllmann/IB-MTS
Three datasets are used in the paper (preprocessed data available bellow):
To skip the preprocessing part, you may download the preprocessed data instead:
- IRIS preprocessed data (11.9GB) available at Zenodo
- AL preprocessed data (34.5MB) available at Zenodo
- PB preprocessed data (114.2MB) available at Zenodo
You can run download_data.py script:
$ python download_data.py --data=<data name> --dir=<data dir> --prpr
where <data name> is iris, al or pb.
Ensure that <data dir> and dataset_address=<data name>_data parameter of other python files are the same: by default, <data dir>=dataset_address=<data name>_data.
Parameter --prpr is for preprocessed data, use --noprpr if you want to download raw data.
Weights of all forecasting models (IB-MTS, LSTM, IB-LSTM, GRU, IB-GRU, NBeats), of IRIS classifiers and for VGG16 layers used in the loss are available in h5py format and can be downloaded with download_weights.py:
$ python download_weights.py
Weights of all forecasting models (IB-MTS, LSTM, IB-LSTM, GRU, IB-GRU, NBeats) are available at Zenodo as a compressed .7z format (407.5MB). By default, they should be downloaded and unziped in the root directory of this project.
Weights for the VGG16 used in the loss are available from an external Google Drive source as a h5py format (528MB).
By default, they should be downloaded and unziped in the vgg_weights directory of this project.
And weights of the Classifier of solar activities on IRIS data are available at Zenodo as a compressed .7z format (46.3MB).
By default, they should be downloaded and unziped in the classifiers directory of this project.
You can run download_testres.py script:
$ python download_testres.py
All test results are available at Zenodo as a compressed .7z format (108MB). By default, they should be downloaded and unziped from the root directory of this project.
If you would like to directly test the models and skip the preprocessing and training part, follow From pretrained models and preprocessed data.
If you would like to preprocess the data and train the models, follow From Scratch.
This part assumes that you downloaded weights and preprocessed data as described in Download data and Download weights.
- Generate test results for each model
You can also download test results at Zenodo Download test results and skip the next line of code. To test your model, you can do it in the following way for iris data and IBMTS model:
$ python main.py --model_type=IBMTS --dataset=iris_level_2C --label_length=240 --labels=QS_AR_FL --preload_train --nochange_traindata --test --test_ds=TE_TEL --add_classifier --add_centercount --notrain --nopredict
mode_type can be IBMTS, LSTMS (simple LSTM), LSTM (named 'IB-LSTM' in the paper), GRUS (simple GRU), GRU (named 'IB-GRU' in the paper), NBeats.
For each 'IRIS', 'AL' or 'PB' evaluated in the paper, here are the corresponding parameters:
IRIS data:
dataset=iris_level_2C
dataset_address=iris_data
label_length=240
labels=QS_AR_FL
AL data:
dataset=al_2C
dataset_address=al_data
label_length=137
labels=AL
PB data:
dataset=pb_2C
dataset_address=pb_data
label_length=325
labels=PB
This may generate a lot of figures and data in the npz format: you can also download those results at Zenodo following Download test results.
- Compare models
To generate part of Table 2 and Figures 9, 10:
$python compare_mts_plots.py --fname=<path for results> --dataset=<data name> --label_length=<spatial length of data>
Default paths to model results must be modified in the py file.
To generate other part of Table 2 and Figures 11 and 13:
$python compare_mts_plots_long.py --fname=<path for results> --dataset=<data name> --label_length=<spatial length of data>
To generate part of Tables 3, 5 and Figures 14, 17:
$python compare_cv_plots.py --fname=<path for results> --dataset=<data name> --label_length=<spatial length of data>
$python get_psnrssim.py --fname=<path for results>
To generate other part of Tables 3, 5 and Figures 15, 18:
$python compare_cvnn_long.py --fname=<path for results> --dataset=<data name> --label_length=<spatial length of data>
$python get_psnrssim_long.py --fname=<path for results>
To generate Figure 12:
$python durations_counts.py
To generate Figure 16 and Table 4:
$python compare_centers.py --fname=<path for results> --dataset=<data name> --label_length=<spatial length of data>
$python get_ib.py --fname=<path for results> --dataset=<data name> --label_length=<spatial length of data>
$python get_ib_print.py --fname=<path for results>
$python compare_centers_long.py --fname=<path for results> --dataset=<data name> --label_length=<spatial length of data>
$python get_ib_long.py --fname=<path for results> --dataset=<data name> --label_length=<spatial length of data>
$python get_ib_long_print.py --fname=<path for results>
To generate Table 6:
$python compare_centers.py --fname=<path for results> --dataset=<data name> --label_length=<spatial length of data>
$python get_accthss_centers.py --fname=<path for results>
$python compare_centers_long.py --fname=<path for results> --dataset=<data name> --label_length=<spatial length of data>
$python get_accthss_centerslong.py --fname=<path for results>
To generate Figure 19:
$python compare_centers.py --fname=<path for results> --dataset=<data name> --label_length=<spatial length of data>
To generate Table 7:
$python get_accthss_class.py --fname=<path for results> --dataset=<data name> --label_length=<spatial length of data>
$python print_act.py --fname=<path for results>
$python get_accthss_class_long.py --fname=<path for results> --dataset=<data name> --label_length=<spatial length of data>
$python print_act_class_long.py --fname=<path for results>
To generate Figure 20:
$python compare_feat_plots.py --fname=<path for results> --dataset=<data name> --label_length=<spatial length of data>
$python compare_feat_long.py --fname=<path for results> --dataset=<data name> --label_length=<spatial length of data>
This part assumes that you downloaded raw or preprocessed data as described in Download data and that you downloaded VGG16 weights and eventually IRIS Classifiers weights as described in Download weights.
To train an IBMTS model from scratch on iris data:
$ python main.py --model_type=IBMTS --epoch=100 --dataset=iris_level_2C --dataset_address=iris_data --label_length=240 --labels=QS_AR_FL --train --preload_train --nopredict --notest
For IRIS data, you must set preload_train parameter to True because data was already preprocessed and the model should look for this data. When preload_train is True and no weights are found, it will just start training from scratch. If you start with AL or PB data from scratch (without preprocessing), set preload_train to False and define the train_ratio and test_ratio parameters, or set given_tvt to True if you specified in file names which ones are for train/valid/test.
To test your model, you can do it in the following way for iris data and IBMTS model:
$ python main.py --model_type=IBMTS --dataset=iris_level_2C --label_length=240 --labels=QS_AR_FL --preload_train --nochange_traindata --test --test_ds=TE_TEL --add_classifier --add_centercount --notrain --nopredict
This may generate a lot of figures and data in the npz format. To compare models between then and generate figures present in the paper, follow From pretrained models and preprocessed data.
Complete explanation can be found in paper.
