I noticed a difference in output when computing the TFSD using the seewave R package and the scikit-maad implementation in python.
Is there any justification and is it possible to get the same result from the same input?
Regards,
1
-
$\begingroup$ please provide additional information about the differences, routines used, what is the input, what did you expect and what are the outputs. $\endgroup$– WMXZCommented Feb 15 at 17:12
Add a comment
|
2 Answers
$\begingroup$
$\endgroup$
I think it's essentially the same answer to both questions.
In general (regarding development of software algorithms), a clearly defined specification and/or standard is required to get the same result from the same input. I wasn't familiar with TFSD but after googling see that it is an acronym for time-frequency second derivative. It looks like this is more of a concept than a clearly defined algorithms, so it seems reasonable that there could be more than one way to estimate this quantity.
Without a clearly defined and well-standardised specification that can be followed, software developers have some latitude when translating a concept into an algorithm.
There are a number of factors that can lead to differences in implementation of a concept -- even if all input parameters are the same. These factors might include: the specific details of the algorithm/software implementation, the programming language, low-level data types, and additional libraries that are used.
If you really need the scikid-maad result, but want to do everything else in R you could look into the R package reticulate (https://cran.r-project.org/web/packages/reticulate/vignettes/calling_python.html) to call Python code from within R. Or if it's the seewave results that you require, but you're working in Python, then the rpy2 package (https://github.com/rpy2/rpy2) seems to be the way to achieve that. NB: Though I'm aware of these packages, I've never actually used them myself, so can't comment on how easy they are to use or how well they work.
$\begingroup$
$\endgroup$
1
Thank you Brian for the answer. I will look into these packages.
@WMXZ input: any wav file, routine used: TFSD from seewave and TFSD from scikit-maad using the same parameters, I expected the same output.
So far, I think the main difference is in the spectrogram computation that feed the algorithm.
-
2$\begingroup$ bernabeu-a, OK, but this is better added as comment and not as a answer to original question. Also, in order to dive into the code, one would need all the parameters used. Otherwise, it results only to philosophical discussion $\endgroup$– WMXZCommented Feb 16 at 9:00