This is an extension of NeuTraj(https://github.com/yaodi833/NeuTraj), which accelerates the similarity computation by neural metric learning. The code has been tested successfully with Python 3.7.
Pytorch, Numpy
Please create 3 empty folders:
*data: Path of the original data which is organized to a time series.
*features: This folder contains the features that generated after the preprocessing.py. It contains four files: DATASETNAME_ts_value, DATASETNAME_ts_grid, DATASETNAME_ts_index and DATASETNAME_ts_label and seed_distance.
*model: It is used for placing the NeuTS model of each training epoch.
Due to the file limit of Github, we put the dataset on other sites. Please first download the data and put it in data folder. Three time series datasets and the learned best models can be download at:
*data: It contains original time series dataset. https://pan.baidu.com/s/1TxJ2StuCe5gfTtn_ELmOSg Extraction code: hiwq
*features: This folder contains the processed features and precomputed seed distances for three time series datasets. https://pan.baidu.com/s/16tmRwEsz4WSLBTG6k-_Dyg Extraction code: x2fx
*best_models: It contains the learned best models. https://pan.baidu.com/s/1kQzCd7oAxAII5FscTWrMLA Extraction code: 3jtc
*UCRSuite: Modified UCR Suite for time series classification. https://pan.baidu.com/s/1gthOreFy5FznWQqITNZL3Q Extraction code: cs7w
Run preprocessing.py. It filters the original data and maps the coordinates to grids. After such process, intermediate files which contain DATASETNAME_ts_value, DATASETNAME_ts_grid, DATASETNAME_ts_index and DATASETNAME_ts_label are generated. Then, we calculate the pair-wise distance under the distance measure and get the seed_distance.
Run train.py. It trains NeuTS under the supervision of seed distance. The parameters of NeuTS can be modified in /tools/config.py