A GP-GPU/CPU Dynamic Time Warping (DTW) implementation for the analysis of Multivariate Time Series (MTS).
MTSS is a GPU/CPU software designed for the classification and the sub-sequence similarity search of MTS. Originally inspired by [1], MTSS aims to improve the time performance and accuracy for classifying and sub-searching any kind of MTS by using the well known similarity measure: Dynamic Time Warping (DTW).
The Classification task aims to predict group membership for MTS instances belonging to set of data. Supposing we have already split the initial data set into train set and test set, the membership group for each MTS of the test set, is assigned by comparing once at time each i-th MTS of test set over all the MTS belonging to the train set and, then selecting the membership group for which the comparison against all the MTS in the train set is minimum in term of DTW similarity measure. As we will explain later, we can provide three different types of DTW similarity measure for comparing two MTS.
The Similarity subseq-search aims to find the best match for a time series (known as the query) onto another time series (known as instance). That's achieved by punctually shifting the instance time series on the query, and comparing it by using a similarity distance such as DTW (see figure below).
Although each of the two task can be perfomed both on CPU ans GPU, the latter implementation improves the speed-up of the above citated tasks when compared to single-threaded CPU implementation. It's important to stress out that the software's speed-up performace gets better when the variables ruling the speed complexity of those two tasks drastically increase.
The software allow to uses three different types of DTW, namely:
- D-MDTW: Dependent-Multivariate Dynamic Time Warping
- I-MDTW: Independent-Multivariate Dynamic Time Warping
- R-MDTW: Rotation-Multivariate Dynamic Time Warping
For more information, please refer to [1-2].
The software is purely written in CUDA, using the C/C++ language as support. In order to use it you must have installed:
-
A working gcc compiler.
-
A CUDA version 5.0 or greater. For installing, please refer to the official documention.
-
gnuplot (optional - required for the benchmark)
-
Check library unit testing framework. For installing, please refer to the official documentation
Runn the following command to clone the master repository into a target directory:
git clone https://github.com/DavideNardone/MTSS-Multivariate-Time-Series-Sofwtare.git <target>
Once you are on the top folder, you can install the software by running the following commands:
autoreconf -i && automake --add-missing
./configure && make && make install
These two commands will ensure that you'd have all the compiled files you need for using the software.
The software can be used as a standard command line tool with a list of options that are explained as follows.
The program can run with the following flag options:
-
-t: It's used to decide which task to perform (e.g., CLASSIFICATION, SUBSEQ_SEARCH)
-
-i: It represents some input parameters, that is:
- The version you want to use: CPU or GPU;
- The number of dimension for the MTS;
- Depending on the first parameter, it represents the #thread for each block (GPU) and/or the read mode.
NOTE: For more information about the read mode, please refer to the section Data Format.
-
-f: It's used to specify the file path of the data (refer to the section Data Format).
-
-k (optional): In the CLASSIFICATION task is possible to perform k-fold cross validation by specifying the number of folders (e.g., -k 10 0) only when the train and test set are provided. An additional flag(0-1) is used for performing the shuffling among the folders generated.
NB: Setting the flag to1does not allow the reproducibility of the results on the same data set among the GPU and CPU versions. -
-o: Depending on the task and read mode, the following parameters represents:
- CLASSIFICATION (read-mode=0 oppure 1):
- The number of MTS samples;
- The length of each MTS sample (same size for each dimension);
NOTE: For this combination it's necessary the -k flag.
- CLASSIFICATION (read-mode=2):
- The number of MTS sample in the TRAINING SET;
- The number of MTS sample in the TESTING SET;
- The MTS length (same size for each dimension).
- SUBSEQ_SEARCH (read-mode=0 oppure 1):
- The MTS length (same size for each dimension);
- The MTS query's length to search for.
-
-m: It's used to specify the type of MDTW to use:
- 0: Dependent similarity measure;
- 1: Independent similarity measure;
- 2: Rotation similarity measure (It suites only for the CLASSIFICATION task)
- <similarity distance>: ED or DTW.
-
-v: Verbose flag for skipping some or all of the output information (e.g., -v 50, display information at each 50th iteration)
-
-d: It specify the GPU's ID to use (e.g. 0: GeForce GTX TITAN X).
-
--help: It quickly explain how the MTSS software works.
-
--version: It shows the info version about the sofware.
At the moment, MTSS works only with txt file format. Depending on the type of task to perform, the data file must be adequayely formatted.
CLASSIFICATION
For this task, MTSS provides three different types of reading mode:
-
read-mode=0: It's possible to feed MTSS with two files,
[DATA, LABEL]. TheDATAfile must be formatted as a TD* data matrix, where each row must represents the t-th features values of the MTS at the time instant d-th (in this case, the MTS are appended in the file), while theLABELfile just contains the integer class label. (A template file is placed in data/classification/rm_0 -
read-mode=1: It's possible to feed MTSS with N files, where each of them is formatted as a NT* data matrix, where each row at the first position must contain an integer class label followed by T-1 values of the MTS. (A template file is placed in data/classification/rm_1).
-
read-mode=2: It's possible to feed MTSS with a
TRAIN SETand aTEST SETfile. Both the file must be formatted as D*T data matrix, where each d-th row must contain the MTS values, where the latter MTS are appended in a file. (A template file is placed in data/classification/rm_2).
SUBSEQ_SEARCH
For this task, MTSS takes in input two files [T_SERIES, Q_SERIES]. The T_SERIES represents the time series on which the Q_SERIES has to be searched. Both the files must be formatted as DT* data matrix where each column must contain the dimensional values of the T_SERIES at each time instant t and viceversa (depending on the read-mode (0|1)). (A template file is placed in data/subseq_search/.
NOTE: The MTSS presented here assume that all the MTS have the same time length.
All the information about the template data set [3] can be retrieved at the following source: https://sites.google.com/site/dtwadaptive/home
In this section we provide some basic examples with the aim of showing the usage of the software. The data set used are thoese referred in the previous section. For each task mentioned, we list both the CPU and GPU solution.
CLASSIFICATION
Compiling
nvcc -arch=sm_30 src/MD_DTW.cu src/module.cu -o mdtwObj
Running
CPU:
./mdtwObj -t CLASSIFICATION -i CPU 3 0 -f data/classification/rm_0/DATA data/classification/rm_0/LABEL -k 10 0 -o 1000 152 -m 0 DTW -v 0
./mdtwObj -t CLASSIFICATION -i CPU 3 1 -f data/classification/rm_1/X_MAT data/classification/rm_1/Y_MAT data/classification/rm_1/Z_MAT -k 10 0 -o 1000 152 -m 0 DTW -v 0
./mdtwObj -t CLASSIFICATION -i CPU 3 2 -f data/classification/rm_2/TRAIN data/classification/rm_2/TEST -o 150 850 152 -k 10 0 -m 0 DTW -v 0
GPU:
./mdtwObj -t CLASSIFICATION -i GPU 3 512 0 -f data/classification/rm_0/DATA data/classification/rm_0/LABEL -k 10 0 -o 1000 152 -m 0 DTW -d 0 -v 0
./mdtwObj -t CLASSIFICATION -i GPU 3 512 1 -f data/classification/rm_1/X_MAT data/classification/rm_1/Y_MAT data/classification/rm_1/Z_MAT -k 10 0 -o 1000 152 -m 0 DTW -d 0 -v 0
./mdtwObj -t CLASSIFICATION -i GPU 3 512 2 -f data/classification/rm_2/TRAIN data/classification/rm_2/TEST -o 150 850 152 -k 10 0 -m 0 DTW -d 0 -v 0
SUBSEQ_SEARCH
Compiling
nvcc -arch=sm_30 src/MD_DTW.cu src/module.cu -o mdtwObj
Running
CPU:
./mdtwObj -t SUBSEQ_SEARCH -i CPU 3 0 -f data/subseq_search/T_series data/subseq_search/Q_series -o 3907 421 -m 0 DTW -v 0
GPU:
./mdtwObj -t SUBSEQ_SEARCH -i GPU 3 512 0 -f data/subseq_search/T_series data/subseq_search/Q_series -o 3907 421 -m 0 DTW -d 0 -v 0
NOTE: The implementation presented here assumes that each compared MTS has the same time length.
In order to validate the software, some unit tests are provided in the tests folder. These units can be exectuted by installing the check framework and by running the following executable file:
/tests/unit_test
In order to compare the time performance between the CPU and GPU implementations when the keys task parameters change, we provided a benchmark test for exploring a range of all combined parameters. The benchmark run may take awhile to get completed, so in case the user is interested in testing only some parameters combination, he/she can modify the file benchmark.cu code ad-hoc and then run the executable file:
/tests/benchmark
Davide Nardone, University of Naples Parthenope, Science and Techonlogies Departement, Msc Applied Computer Science https://www.linkedin.com/in/davide-nardone-127428102/
For any kind of problem, questions, ideas or suggestions, please don't esitate to contact me at:
[1] Sart, Doruk, et al. "Accelerating dynamic time warping subsequence search with GPUs and FPGAs." Data Mining (ICDM), 2010 IEEE 10th International Conference on. IEEE, 2010. APA
[2] Shokoohi-Yekta, Mohammad, Jun Wang, and Eamonn Keogh. "On the non-trivial generalization of dynamic time warping to the multi-dimensional case." Proceedings of the 2015 SIAM International Conference on Data Mining. Society for Industrial and Applied Mathematics, 2015.
[3] Shokoohi-Yekta, M., Hu, B., Jin, H. et al. Data Min Knowl Disc (2017) 31: 1. https://doi.org/10.1007/s10618-016-0455-0
The following software is related to a scientific manuscript submitted to an official journal, so i wouldn't recommend you to use it for any plagiarism.