Reproducibility in scientific computing - linux

A common workflow in scientific computing is to first write code (perhaps a simulation), run it and analyse its results, then make modifications based on what the previous round of results show. This cycle may go round tens, possibly even hundreds of times before the project is finished.
A key problem of this development cycle is one of reproducibility. As I have gone through this cycle, I will have produced results, graphs, and various other output. I want to be able to take any graph (from yesterday, last week, month, or longer) and reliably reconstruct the code and environment which were used to produce this. How can I solve this problem? The "obvious" solution appears to be one of organisation and recording everything, however this has the potential to create much additional work. I'm interested in the balance of achieving this without handicapping productivity.

http://ipython.org/notebook.html
for people who want to share reproducible research.
http://jupyter.org/
Not just python, there are many languages supported.
Recently, i was experimenting with julia language,
this is one of the advised tutorials.
It's using IJulia which is based on IPython, very nice intro.

Related

Software to Tune/Calibrate Properties for Heuristic Algorithms

Today I read that there is a software called WinCalibra (scroll a bit down) which can take a text file with properties as input.
This program can then optimize the input properties based on the output values of your algorithm. See this paper or the user documentation for more information (see link above; sadly doc is a zipped exe).
Do you know other software which can do the same which runs under Linux? (preferable Open Source)
EDIT: Since I need this for a java application: should I invest my research in java libraries like gaul or watchmaker? The problem is that I don't want to roll out my own solution nor I have time to do so. Do you have pointers to an out-of-the-box applications like Calibra? (internet searches weren't successfull; I only found libraries)
I decided to give away the bounty (otherwise no one would have a benefit) although I didn't found a satisfactory solution :-( (out-of-the-box application)
Some kind of (Metropolis algorithm-like) probability selected random walk is a possibility in this instance. Perhaps with simulated annealing to improve the final selection. Though the timing parameters you've supplied are not optimal for getting a really great result this way.
It works like this:
You start at some point. Use your existing data to pick one that look promising (like the highest value you've got). Set o to the output value at this point.
You propose a randomly selected step in the input space, assign the output value there to n.
Accept the step (that is update the working position) if 1) n>o or 2) the new value is lower, but a random number on [0,1) is less than f(n/o) for some monotonically increasing f() with range and domain on [0,1).
Repeat steps 2 and 3 as long as you can afford, collecting statistics at each step.
Finally compute the result. In your case an average of all points is probably sufficient.
Important frill: This approach has trouble if the space has many local maxima with deep dips between them unless the step size is big enough to get past the dips; but big steps makes the whole thing slow to converge. To fix this you do two things:
Do simulated annealing (start with a large step size and gradually reduce it, thus allowing the walker to move between local maxima early on, but trapping it in one region later to accumulate precision results.
Use several (many if you can afford it) independent walkers so that they can get trapped in different local maxima. The more you use, and the bigger the difference in output values, the more likely you are to get the best maxima.
This is not necessary if you know that you only have one, big, broad, nicely behaved local extreme.
Finally, the selection of f(). You can just use f(x) = x, but you'll get optimal convergence if you use f(x) = exp(-(1/x)).
Again, you don't have enough time for a great many steps (though if you have multiple computers, you can run separate instances to get the multiple walkers effect, which will help), so you might be better off with some kind of deterministic approach. But that is not a subject I know enough about to offer any advice.
There are a lot of genetic algorithm based software that can do exactly that. Wrote a PHD about it a decade or two ago.
A google for Genetic Algorithms Linux shows a load of starting points.
Intrigued by the question, I did a bit of poking around, trying to get a better understanding of the nature of CALIBRA, its standing in academic circles and the existence of similar software of projects, in the Open Source and Linux world.
Please be kind (and, please, edit directly, or suggest editing) for the likely instances where my assertions are incomplete, inexact and even flat-out incorrect. While working in related fields, I'm by no mean an Operational Research (OR) authority!
[Algorithm] Parameter tuning problem is a relatively well defined problem, typically framed as one of a solution search problem whereby, the combination of all possible parameter values constitute a solution space and the parameter tuning logic's aim is to "navigate" [portions of] this space in search of an optimal (or locally optimal) set of parameters.
The optimality of a given solution is measured in various ways and such metrics help direct the search. In the case of the Parameter Tuning problem, the validity of a given solution is measured, directly or through a function, from the output of the algorithm [i.e. the algorithm being tuned not the algorithm of the tuning logic!].
Framed as a search problem, the discipline of Algorithm Parameter Tuning doesn't differ significantly from other other Solution Search problems where the solution space is defined by something else than the parameters to a given algorithm. But because it works on algorithms which are in themselves solutions of sorts, this discipline is sometimes referred as Metaheuristics or Metasearch. (A metaheuristics approach can be applied to various algorihms)
Certainly there are many specific features of the parameter tuning problem as compared to the other optimization applications but with regard to the solution searching per-se, the approaches and problems are generally the same.
Indeed, while well defined, the search problem is generally still broadly unsolved, and is the object of active research in very many different directions, for many different domains. Various approaches offer mixed success depending on the specific conditions and requirements of the domain, and this vibrant and diverse mix of academic research and practical applications is a common trait to Metaheuristics and to Optimization at large.
So... back to CALIBRA...
From its own authors' admission, Calibra has several limitations
Limit of 5 parameters, maximum
Requirement of a range of values for [some of ?] the parameters
Works better when the parameters are relatively independent (but... wait, when that is the case, isn't the whole search problem much easier ;-) )
CALIBRA is based on a combination of approaches, which are repeated in a sequence. A mix of guided search and local optimization.
The paper where CALIBRA was presented is dated 2006. Since then, there's been relatively few references to this paper and to CALIBRA at large. Its two authors have since published several other papers in various disciplines related to Operational Research (OR).
This may be indicative that CALIBRA hasn't been perceived as a breakthrough.
State of the art in that area ("parameter tuning", "algorithm configuration") is the SPOT package in R. You can connect external fitness functions using a language of your choice. It is really powerful.
I am working on adapters for e.g. C++ and Java that simplify the experimental setup, which requires some getting used to in SPOT. The project goes under name InPUT, and a first version of the tuning part will be up soon.

How to keep track of performance testing

I'm currently doing performance and load testing of a complex many-tier system investigating the effect of different changes, but I'm having problems keeping track of everything:
There are many copies of different assemblies
Orignally released assemblies
Officially released hotfixes
Assemblies that I've built containing further additional fixes
Assemblies that I've build containing additional diagnostic logging or tracing
There are many database patches, some of the above assemblies depend on certain database patches being applied
Many different logging levels exist, in different tiers (Application logging, Application performance statistics, SQL server profiling)
There are many different scenarios, sometimes it is useful to test only 1 scenario, other times I need to test combinations of different scenarios.
Load may be split across multiple machines or only a single machine
The data present in the database can change, for example some tests might be done with generated data, and then later with data taken from a live system.
There is a massive amount of potential performance data to be collected after each test, for example:
Many different types of application specific logging
SQL Profiler traces
Event logs
DMVs
Perfmon counters
The database(s) are several Gb in size so where I would have used backups to revert to a previous state I tend to apply changes to whatever database is present after the last test, causing me to quickly loose track of things.
I collect as much information as I can about each test I do (the scenario tested, which patches are applied what data is in the database), but I still find myself having to repeat tests because of inconsistent results. For example I just did a test which I believed to be an exact duplicate of a test I ran a few months ago, however with updated data in the database. I know for a fact that the new data should cause a performance degregation, however the results show the opposite!
At the same time I find myself sepdning disproportionate amounts of time recording these all these details.
One thing I considered was using scripting to automate the collection of performance data etc..., but I wasnt sure this was such a good idea - not only is it time spent developing scripts instead of testing, but bugs in my scripts could cause me to loose track of things even quicker.
I'm after some advice / hints on how better to manage the test environment, in particular how to strike a balance between collecting everything and actually getting some testing done at the risk of missing something important?
Scripting the collection of the test parameters + environment is a very good idea to check out. If you're testing across several days, and the scripting takes a day, it's time well spent. If after a day you see it won't finish soon, reevaluate and possibly stop pursuing this direction.
But you owe it to yourself to try it.
I would tend to agree with #orip, scripting at least part of your workload is likely to save you time. You might consider taking a moment to ask what tasks are the most time consuming in terms of your labor and how amenable are they to automation? Scripts are especially good at collecting and summarizing data - much better then people, typically. If the performance data requires a lot of interpretation on your part, you may have problems.
An advantage to scripting some of these tasks is that you can then check them in along side the source / patches / branches and you may find you benefit from organizational structure of your systems complexity rather than struggling to chase it as you do now.
If you can get away with testing only against a few set configurations that will keep the admin simple. It may also make it easier to put one on each of several virtual machines which can be quickly redeployed to give clean baselines.
If you genuinely need the complexity you describe I'd recommend building a simple database to allow you to query the multivariate results you have. Having a column for each of the important factors will a allow you to query in for questions like "what testing config had the lowest variance in latency?" and "which test database allowed the raising of most bugs?". I use sqlite3 (probably through the Python wrapper or the Firefox plug-in) for this kind of lightweight collection, because it keeps maintenance overhead relatively low and allows you to avoid perturbing the system under test too far, even if you need to run on the same box.
Scripting the tests will make them quicker to execute and permit results to be gathered in an already-ordered way, but it sounds like your system may be too complex to make this easy to do.

How to choose a Feature Selection Algorithm? - advice

Is there a research paper/book that I can read which can tell me for the problem at hand what sort of feature selection algorithm would work best.
I am trying to simply identify twitter messages as pos/neg (to begin with). I started out with Frequency based feature selection (having started with NLTK book) but soon realised that for a similar problem various individuals have choosen different algorithms
Although I can try Frequency based, mutual information, information gain and various other algorithms the list seems endless.. and was wondering if there an efficient way then trial and error.
any advice
Have you tried the book I recommended upon your last question? It's freely available online and entirely about the task you are dealing with: Sentiment Analysis and Opinion Mining by Pang and Lee. Chapter 4 ("Extraction and Classification") is just what you need!
I did an NLP course last term, and it came pretty clear that sentiment analysis is something that nobody really knows how to do well (yet). Doing this with unsupervised learning is of course even harder.
There's quite a lot of research going on regarding this, some of it commercial and thus not open to the public. I can't point you to any research papers but the book we used for the course was this (google books preview). That said, the book covers a lot of material and might not be the quickest way to find a solution to this particular problem.
The only other thing I can point you towards is to try googling around, maybe in scholar.google.com for "sentiment analysis" or "opinion mining".
Have a look at the NLTK movie_reviews corpus. The reviews are already pos/neg categorized and might help you with training your classifier. Although the language you find in Twitter is probably very different from those.
As a last note, please post any successes (or failures for that matter) here. This issue will come up later for sure at some point.
Unfortunately, there is no silver bullet for anything when dealing with machine learning. It's usually referred to as the "No Free Lunch" theorem. Basically a number of algorithms work for a problem, and some do better on some problems and worse on others. Over all, they all perform about the same. The same feature set may cause one algorithm to perform better and another to perform worse for a given data set. For a different data set, the situation could be completely reversed.
Usually what I do is pick a few feature selection algorithms that have worked for others on similar tasks and then start with those. If the performance I get using my favorite classifiers is acceptable, scrounging for another half percentage point probably isn't worth my time. But if it's not acceptable, then it's time to re-evaluate my approach, or to look for more feature selection methods.

Pair Programming with an uneven number of team members? [closed]

Closed. This question is opinion-based. It is not currently accepting answers.
Want to improve this question? Update the question so it can be answered with facts and citations by editing this post.
Closed 5 years ago.
Improve this question
Recently, we've come across an issue at work where if one person is working on some code by themselves, it seems to come out with the other team members looking at it and going "Huh? That's ugly, unmanageable, I need to rewrite that"
In fact, recently, I myself have had to re-factor something that was written the week before so that I'd be able to add in my (related) feature.
I know that Pair programming is the way to go for this, but we have an uneven team (3 members). As our team is being pushed pretty hard at the moment, we really don't have time for Peer Reviews (though we can do Pair Programming, as we're allowed to estimate that into our task estimates)
I'm just curious as to how people would suggest we overcome these issues with poor code being generated.
When you work alone, and produce code which your colleagues find ugly and unmanageable and needs to be rewritten, then do you:
(a) agree with them when you look at it a second time,
(b) disagree?
If (a), then the problem is that on your own, you aren't fully clarifying your code when you write it. Since pair programming is the only thing making you write decent code, I suppose I'd recommend that the "odd one out" should work on tasks which do not involve writing long tracts of bad code: bug-hunting; maybe writing test code, since that tends to be a bit less fiendish. Meanwhile, work on improving your skills at writing better code - perhaps do reviews of your own code from a few months ago, and make notes as to what was wrong with it.
If (b), then the problem you have is incompatible ways of expressing your ideas. The code may not be bad by your standards, but it's mutually incomprehensible, which in a corporate setting means it's bad code. Pair programming means what you write is a compromise that 2 out of 3 of you understand, but that's not really a solution. You need to come to some mutual agreements about what you find most difficult about each other's code, and stop doing that. And you all urgently need to start thinking of "code quality" in terms of "my 2 colleagues will like this code", not "I like this code".
Either way, you all need to work on writing code for the purpose of being read, rather than for the purpose of getting the immediate job done as quickly as you possibly can. Personally I have done this by trying to express things in the way that I think other people might express and understand them, rather than just what makes sense to me at the time. Eventually it becomes habitual. When I write code, I write it for a public audience just like I'm writing this post for a public audience. OK, so on my personal projects it's an audience of people who think just like me, whereas at work it's an audience that thinks like my colleagues. But the principle is to write code as if someone's reading it. You're explaining yourself to them, not the compiler.
Not that my code is the best in the world, but I do think I benefited in that my first job was in a company with 30-odd programmers, so I got to see a wide range of ways of thinking about things. Also a few examples of "what not to do", where one programmer had done something that nobody else could easily understand, and therefore could definitively be said to be bad. With only 3 people, it's not clear whether a 2 v. 1 difference of opinion means that the 1 is a freak or a reasonable minority. When I did something and 4 or 5 people could glance at it and immediately say "eeew, don't do that", then I started to really believe it was just a dumb idea in the first place.
I'd also recommend that if you aren't allowed to budget for code review, lie and cheat. If you're heavily re-writing someone else's code, you're effectively taking the time to review it anyway, you just aren't providing the feedback which is the worthwhile part of code review. So sneak the review in under the radar - write a function or three, then ask a colleague to look at it and give you instant feedback on whether it makes sense to them. It helps to have a conversation as soon as you've done it, with the code on the monitor, but do try not to interrupt people when they have "flow", or to get into lengthy arguments. It's not pair programming, and it's not formal code review, but it might help you figure out what it is you're doing on your own that's so bad.
I'm surprised that you don't have time to do peer reviews but you have time to do paired programming. Is the latter not a much bigger sink of time?
We also have three developers only at our company and, surprise, surprise, we're being pushed hard at the moment. I'm pretty sure my boss would laugh at me if I suggested paired programming because that would be viewed as doubling the number of man hours for a task even though in practice that's not the result it should produce. Our peer reviews are never more than an hour and that is an extreme case. On average I would say they are probably about 10 minutes and, per developer, only happen once or twice in a day.
IMO you should give peer reviews a try. You often find that the offending people (i.e. the people writing the lower quality code) eventually realise that they need to make more of an effort and the quality improves over time.
If you have three developers and each of you think the others code is not good, you urgently need peer reviews.
So:
you are being pushed pretty hard
your code is of poor quality
Do you think the two could possibly be related? The answer is to fix the schedule.
Pair up all three at once.
Set up some coding standards.
Use a dunce cap for build breaking developers.
Perform daily stand up meetings to communicate progress.
Also try peer reviews twice a week, like Tuesday and Friday.
Pair Programming doesn't have to be all day every day to be effective. I have seen good results from even an hour or two working together each week. One way to go would be to pair A & B for a while, then A & C, then A & B... with a lot of individual time in between.
It also depends a lot on the personalities and chemistry of the team members. Two of the three might work exceptionally well together and you'd want to benefit from that.
You should still pair. Set up sessions say 1 day per week and rotate the pairs. This should keep your manager happy and increase the quality of the code, improve communication. If you keep metrics on how many faults happen in paired vs solitary coding you should start to see the benfit and display this to your manager,
eg This took x man hours but saved on average y in defect fixing. Additionally the clode is cleaner and will take less time to alter then next time we touch it.
From there you will have hard statistics and you can start to code more.
Basically your story seems to be the same as mine.
No time to do things.
Mistakes happen.
Rush to fix it (taking more time)
Go to 1
You need to stop the rot.
Code reviews
Enable Stylecop that will force you to write readable, standardised and manageable code
We use code reviews. Additionally there are some single task: changing a diagram, installing some stuff...

How do I calculate these statistics?

I'm writing an app to help facilitate some research, and part of this involves doing some statistical calculations. Right now, the researchers are using a program called SPSS. Part of the output that they care about looks like this:
They're really only concerned about the F and Sig. values. My problem is that I have no background in statistics, and I can't figure out what the tests are called, or how to calculate them.
I thought the F value might be the result of the F-test, but after following the steps given on Wikipedia, I got a result that was different from what SPSS gives.
This website might help you out a bit more. Also this one.
I'm working from a fairly rusty memory of a statistics course, but here goes nothing:
When you're doing analysis of variance (ANOVA), you actually calculate the F statistic as the ratio from the mean-square variances "between the groups" and the mean-square variances "within the groups". The second link above seems pretty good for this calculation.
This makes the F statistic measure exactly how powerful your model is, because the "between the groups" variance is explanatory power, and "within the groups" variance is random error. High F implies a highly significant model.
As in many statistical operations, you back-determine Sig. using the F statistic. Here's where your Wikipedia information comes in slightly handy. What you want to do is - using the degrees of freedom given to you by SPSS - find the proper P value at which an F table will give you the F statistic you calculated. The P value where this happens [F(table) = F(calculated)] is the significance.
Conceptually, a lower significance value shows a very strong ability to reject the null hypothesis (which for these purposes means to determine your model has explanatory power).
Sorry to any math folks if any of this is wrong. I'll be checking back to make edits!!!
Good luck to you. Stats is fun, just maybe not this part. =)
I assume from your question that your research colleagues want to automate the process by which certain statistical analyses are performed (i.e., they want to batch process data sets). You have two options:
1) SPSS is now scriptable through python (as of version 15) - go to spss.com and search for python. You can write python scripts to automate data analyses and extract key values from pivot tables, and then process the answers any way you like. This has the virtue of allowing an exact comparison between the results from your python script and the hand-calculated efforts in SPSS of your collaborators. Thus you won't have to really know any statistics to do this work (which is a key advantage)
2) You could do this in R, a free statistics environment, which could probably be scripted. This has the disadvantage that you will have to learn statistics to ensure that you are doing it correctly.
Statistics is hard :-). After a year of reading and re-reading books and papers and can only say with confidence that I understand the very basics of it.
You might wish to investigate ready-made libraries for whichever programming language you are using, because they are many gotcha's in math in general and statistics in particular (rounding errors being an obvious example).
As an example you could take a look at the R project, which is both an interactive environment and a library which you can use from your C++ code, distributed under the GPL (ie if you are using it only internally and publishing only the results, you don't need to open your code).
In short: don't do this by hand, link/use existing software. And sain_grocen's answer is incorrect. :(
These are all tests for significance of parameter estimates that are typically used in Multivariate response Multiple Regressions. These would not be simple things to do outside of a statistical programming environment. I would suggest either getting the output from a pre-existing statistical program, or using one that you can link to and use that code.
I'm afraid that the first answer (sain_grocen's) will lead you down the wrong path. His explanation is likely of a special case of what you are actually dealing with. The anova explained in his links is for a single variate response, in a balanced design. These aren't the F statistics you are seeing. The names in your output (Pillai's Trace, Hotelling's Trace,...) are some of the available multivariate versions. They have F distributions under certain assumptions. I can't explain a text books worth of material here, I would advise you to start by looking at
"Applied Multivariate Statistical Analysis" by Johnson and Wichern
Can you explain more why SPSS itself isn't a fine solution to the problem? Is it that it generates pivot tables as output that are hard to manipulate? Is it the cost of the program?
F-statistics can arise from any number of particular tests. The F is just a distribution (loosely: a description of the "frequencies" of groups of values), like a Normal (Gaussian), or Uniform. In general they arise from ratios of variances. Opinion: many statisticians (myself included), find F-based tests to be unstable (jargon: non-robust).
The particular output statistics (Pillai's trace, etc.) suggest that the original analysis is a MANOVA example, which as other posters describe is a complicated, and hard to get right procedure.
I'm guess also that, based on the MANOVA, and the use of SPSS, this is a psychology or sociology project... if not please enlighten. It might be that other, simpler models might actually be easier to understand and more repeatable. Consult your local university statistical consulting group, if you have one.
Good luck!
Here's an explanation of MANOVA ouptput, from a very good site on statistics and on SPSS:
Output with explanation:
http://faculty.chass.ncsu.edu/garson/PA765/manospss.htm
How and why to do MANOVA or multivariate GLM:
(same path as above, but terminating in '/manova.htm')
Writing software from scratch to calculate these outputs would be both lengthy and difficult;
there's lots of numerical problems and matrix inversions to do.
As Henry said, use Python scripts, or R. I'd suggest working with somebody who knows SPSS if scripting.
In addition, SPSS itself is capable of exporting the output tables to files using something called OMS.
A script within SPSS can do this.
Find out who in your research group knows SPSS and work with them.

Resources