Wildly varying time to run STAMP with query

[ginoa]

Description
Running time of STAMP varies hugely depending on size of data. For some data sizes it even gets stuck without showing the progress bar.
For example computing the distance profile between a dataset and a query of length 34 and window size 13, I observe the following

• average of 1 it/s for dataset of length 620579
• average of 0.5 it/s for dataset of length 620578
• average of 0.3 it/s for dataset of length 620577
• average of 2.5 it/s for dataset of length 620576
• average of 0.04 it/s for dataset of length 620574
• stamp gets stuck without error for some sizes of dataset - or more probably it takes a lot of time to get to the point of showing the progress bar - so much that I do not have patience to wait and want to interrupt it. I must then hard kill Rstudio to interrupt evaluation.

In other words, changing the size of the dataset by just one unit has a big, nonlinear and unintuitive effect on execution time of stamp. For some size of data the function gets stuck without error.

Working examples
I could reproduce this behavior with random data.

ref_data <- rnorm(620576)
query_data <- rnorm(34)
tst <- tsmp::stamp(ref_data,query_data,window_size=13)
## execution speed: 0.4 it/s

ref_data <- rnorm(620577)
query_data <- rnorm(34)
tst <- tsmp::stamp(ref_data,query_data,window_size=13)
## execution speed: 2.5 it/s

Expected behavior
I would expect that running time get shorter as data decreases in size.

[franzbischoff]

Confirmed. With STOMP this behaviour is not present.
I'll have to profile the code and check where is the bottleneck.

Thanks.

[franzbischoff]

The bottleneck is stats::fft(), this may have something to do with the FFT algorithm. Maybe padding the data with zeroes (the right amount) can solve this. Need to look into it.

[franzbischoff]

Implementing mass_v3 solves this problem

https://www.cs.unm.edu/~mueen/MASS_V3.m

[franzbischoff]

Changes STAMP to use MASS_V3