In our paper, we propose loss "DTW-like method for variable-length data (DTWL-VL)".

This repository has the code for "Generation of Variable-Length Time Series from Text using Dynamic Time Warping-Based Method". The codes for implementation stated in the paper, and the code for creating dataset.

Required libraries can be installed by runninng

pip install -r requirements.txt　.

The strict library version we used in the experiment is below:

torch: 1.6.0 cu101

numpy: 1.17.2

tensorboardX: 2.1

Python: 3.7.4

Cython: 0.29.13

future: 0.17.1

joblib: 0.13.2

Pillow: 9.0.1

protobuf: 3.11.3

scikit-learn: 0.21.3

scipy: 1.4.1

setuptools: 41.4.0

six: 1.12.0

llvmlite: 0.29.0 (Install before numba)

numba: 0.45.1

tslearn: 0.4.1

h5py: 2.10.0

The path "../glove2/glove.6B.50d.txt" in "withtime_batch_dataload_inside.py" under the 4 directories for experiment("MSEVariant", "DILATE_VL_active", "DILATE_VL_pad", "DTWL_VL") is the path of GloVe(Jeffrey Pennington, Richard Socher, and Christopher Manning. 2014. GloVe: Global Vectors for Word Representation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics, Doha, Qatar, 1532–1543. https://doi.org/10.3115/v1/D14-1162) embedding file.

You need to download it from　 https://nlp.stanford.edu/data/glove.6B.zip 　(Jeffrey Pennington and Richard Socher and Christopher D. Manning. 2014. GloVe: Global Vectors for Word Representation.) and put it at the place indicated with the relative path, or edit the path "../glove2/glove.6B.50d.txt" written in "withtime_batch_dataload_inside.py" for GloVe file for yourself.

Also, you need to create the data for this experiment by running the code python textmake_withtime.py in "DatasetCreationCode_40". You need to put the folder where the same folder the folder for the experiment exists, or edit the path "../DatasetCreationCode_40" in "Models.py" for every experiment data.

You can start training by entering the folder ("MSEVariant", "DILATE_VL_active", "DILATE_VL_pad", "DTWL_VL") and running python Models.py {gpu_id} (input the id of gpu you want to use in {gpu_id}) .

Our proposed loss is implemented in DTWL_VL/SDTWVL.py .

We refer the code at https://github.com/vincent-leguen/DILATE, which is code for the NeurIPS 2019 paper "Shape and Time Distortion Loss for Training Deep Time Series Forecasting Models" (Le Guen, Vincent and Thome, Nicolas), for "DILATE_VL_pad" and "DILATE_VL_active".

We refer to https://github.com/jadore801120/attention-is-all-you-need-pytorch, for the structure of the model.

The license file of https://github.com/jadore801120/attention-is-all-you-need-pytorch is in the folder of each settings("MSEVariant", "DILATE_VL_active", "DILATE_VL_pad", "DTWL_VL"), and license file of https://github.com/vincent-leguen/DILATE is in the folder named "loss", which includes the code related to the repository(In "DILATE_VL_active", "DILATE_VL_pad").

The code for the model architecture (Written in "Models.py", "transformer/Layers.py", "transformer/Modules.py", "transformer/SubLayers.py" in all the training code folder) is based on the implementation of https://github.com/jadore801120/attention-is-all-you-need-pytorch, and the code for DILATE loss ("loss/path_soft_dtw.py", "loss/dilate_loss.py", "loss/soft_dtw.py" in "DILATE_VL_active", "DILATE_VL_pad") is based on https://github.com/vincent-leguen/DILATE.