General suggestions for debugging in R

问题:

I get an error when using an R function that I wrote:

Warning messages:
1: glm.fit: algorithm did not converge 
2: glm.fit: algorithm did not converge 

What I have done:

1. Step through the function
2. Adding print to find out at what line the error occurs suggests two functions that should not use glm.fit. They are window() and save().

My general approaches include adding print and stop commands, and stepping through a function line by line until I can locate the exception.

However, it is not clear to me using those techniques where this error comes from in the code. I am not even certain which functions within the code depend on glm.fit. How do I go about diagnosing this problem?

回答1:

I\'d say that debugging is an art form, so there\'s no clear silver bullet. There are good strategies for debugging in any language, and they apply here too (e.g. read this nice article). For instance, the first thing is to reproduce the problem...if you can\'t do that, then you need to get more information (e.g. with logging). Once you can reproduce it, you need to reduce it down to the source.

Rather than a \"trick\", I would say that I have a favorite debugging routine:

1. When an error occurs, the first thing that I usually do is look at the stack trace by calling traceback(): that shows you where the error occurred, which is especially useful if you have several nested functions.
2. Next I will set options(error=recover); this immediately switches into browser mode where the error occurs, so you can browse the workspace from there.
3. If I still don\'t have enough information, I usually use the debug() function and step through the script line by line.

The best new trick in R 2.10 (when working with script files) is to use the findLineNum() and setBreakpoint() functions.

As a final comment: depending upon the error, it is also very helpful to set try() or tryCatch() statements around external function calls (especially when dealing with S4 classes). That will sometimes provide even more information, and it also gives you more control over how errors are handled at run time.

These related questions have a lot of suggestions:

• Debugging tools for the R language
• Debugging lapply/sapply calls
• Getting the state of variables after an error occurs in R
• R script line numbers at error?

回答2:

The best walkthrough I\'ve seen so far is:

http://www.biostat.jhsph.edu/%7Erpeng/docs/R-debug-tools.pdf

Anybody agree/disagree?

回答3:

As was pointed out to me in another question, Rprof() and summaryRprof() are nice tools to find slow parts of your program that might benefit from speeding up or moving to a C/C++ implementation. This probably applies more if you\'re doing simulation work or other compute- or data-intensive activities. The profr package can help visualizing the results.

I\'m on a bit of a learn-about-debugging kick, so another suggestion from another thread:

• Set options(warn=2) to treat warnings like errors

You can also use options to drop you right into the heat of the action when an error or warning occurs, using your favorite debugging function of choice. For instance:

• Set options(error=recover) to run recover() when an error occurs, as Shane noted (and as is documented in the R debugging guide. Or any other handy function you would find useful to have run.

And another two methods from one of @Shane\'s links:

• Wrap an inner function call with try() to return more information on it.
• For *apply functions, use .inform=TRUE (from the plyr package) as an option to the apply command

@JoshuaUlrich also pointed out a neat way of using the conditional abilities of the classic browser() command to turn on/off debugging:

• Put inside the function you might want to debug browser(expr=isTRUE(getOption(\"myDebug\")))
• And set the global option by options(myDebug=TRUE)
• You could even wrap the browser call: myBrowse <- browser(expr=isTRUE(getOption(\"myDebug\"))) and then call with myBrowse() since it uses globals.

Then there are the new functions available in R 2.10:

• findLineNum() takes a source file name and line number and returns the function and environment. This seems to be helpful when you source() a .R file and it returns an error at line #n, but you need to know what function is located at line #n.
• setBreakpoint() takes a source file name and line number and sets a breakpoint there

The codetools package, and particularly its checkUsage function can be particularly helpful in quickly picking up syntax and stylistic errors that a compiler would typically report (unused locals, undefined global functions and variables, partial argument matching, and so forth).

setBreakpoint() is a more user-friendly front-end to trace(). Details on the internals of how this works are available in a recent R Journal article.

If you are trying to debug someone else\'s package, once you have located the problem you can over-write their functions with fixInNamespace and assignInNamespace, but do not use this in production code.

None of this should preclude the tried-and-true standard R debugging tools, some of which are above and others of which are not. In particular, the post-mortem debugging tools are handy when you have a time-consuming bunch of code that you\'d rather not re-run.

Finally, for tricky problems which don\'t seem to throw an error message, you can use options(error=dump.frames) as detailed in this question: Error without an error being thrown

回答4:

At some point, glm.fit is being called. That means one of the functions you call or one of the functions called by those functions is using either glm, glm.fit.

Also, as I mention in my comment above, that is a warning not an error, which makes a big difference. You can\'t trigger any of R\'s debugging tools from a warning (with default options before someone tells me I am wrong ;-).

If we change the options to turn warnings into errors then we can start to use R\'s debugging tools. From ?options we have:

 ‘warn’: sets the handling of warning messages.  If ‘warn’ is
      negative all warnings are ignored.  If ‘warn’ is zero (the
      default) warnings are stored until the top-level function
      returns.  If fewer than 10 warnings were signalled they will
      be printed otherwise a message saying how many (max 50) were
      signalled.  An object called ‘last.warning’ is created and
      can be printed through the function ‘warnings’.  If ‘warn’ is
      one, warnings are printed as they occur.  If ‘warn’ is two or
      larger all warnings are turned into errors.

So if you run

options(warn = 2)

then run your code, R will throw an error. At which point, you could run

traceback()

to see the call stack. Here is an example.

> options(warn = 2)
> foo <- function(x) bar(x + 2)
> bar <- function(y) warning(\"don\'t want to use \'y\'!\")
> foo(1)
Error in bar(x + 2) : (converted from warning) don\'t want to use \'y\'!
> traceback()
7: doWithOneRestart(return(expr), restart)
6: withOneRestart(expr, restarts[[1L]])
5: withRestarts({
       .Internal(.signalCondition(simpleWarning(msg, call), msg, 
           call))
       .Internal(.dfltWarn(msg, call))
   }, muffleWarning = function() NULL)
4: .signalSimpleWarning(\"don\'t want to use \'y\'!\", quote(bar(x + 
       2)))
3: warning(\"don\'t want to use \'y\'!\")
2: bar(x + 2)
1: foo(1)

Here you can ignore the frames marked 4: and higher. We see that foo called bar and that bar generated the warning. That should show you which functions were calling glm.fit.

If you now want to debug this, we can turn to another option to tell R to enter the debugger when it encounters an error, and as we have made warnings errors we will get a debugger when the original warning is triggered. For that you should run:

options(error = recover)

Here is an example:

> options(error = recover)
> foo(1)
Error in bar(x + 2) : (converted from warning) don\'t want to use \'y\'!

Enter a frame number, or 0 to exit   

1: foo(1)
2: bar(x + 2)
3: warning(\"don\'t want to use \'y\'!\")
4: .signalSimpleWarning(\"don\'t want to use \'y\'!\", quote(bar(x + 2)))
5: withRestarts({
6: withOneRestart(expr, restarts[[1]])
7: doWithOneRestart(return(expr), restart)

Selection:

You can then step into any of those frames to see what was happening when the warning was thrown.

To reset the above options to their default, enter

options(error = NULL, warn = 0)

As for the specific warning you quote, it is highly likely that you need to allow more iterations in the code. Once you\'ve found out what is calling glm.fit, work out how to pass it the control argument using glm.control - see ?glm.control.

回答5:

So browser(), traceback() and debug() walk into a bar, but trace() waits outside and keeps the motor running.

By inserting browser somewhere in your function, the execution will halt and wait for your input. You can move forward using n (or Enter), run the entire chunk (iteration) with c, finish the current loop/function with f, or quit with Q; see ?browser.

With debug, you get the same effect as with browser, but this stops the execution of a function at its beginning. Same shortcuts apply. This function will be in a \"debug\" mode until you turn it off using undebug (that is, after debug(foo), running the function foo will enter \"debug\" mode every time until you run undebug(foo)).

A more transient alternative is debugonce, which will remove the \"debug\" mode from the function after the next time it\'s evaluated.

traceback will give you the flow of execution of functions all the way up to where something went wrong (an actual error).

You can insert code bits (i.e. custom functions) in functions using trace, for example browser. This is useful for functions from packages and you\'re too lazy to get the nicely folded source code.

回答6:

My general strategy looks like:

1. Run traceback() to see look for obvious issues
2. Set options(warn=2) to treat warnings like errors
3. Set options(error=recover) to step into the call stack on error

回答7:

After going through all the steps suggested here I just learned that setting .verbose = TRUE in foreach() also gives me tons of useful information. In particular foreach(.verbose=TRUE) shows exactly where an error occurs inside the foreach loop, while traceback() does not look inside the foreach loop.

回答8:

Mark Bravington\'s debugger which is available as the package debug on CRAN is very good and pretty straight forward.

library(debug);
mtrace(myfunction);
myfunction(a,b);
#... debugging, can query objects, step, skip, run, breakpoints etc..
qqq(); # quit the debugger only
mtrace.off(); # turn off debugging

The code pops up in a highlighted Tk window so you can see what\'s going on and, of course you can call another mtrace() while in a different function.

HTH

回答9:

I like Gavin\'s answer: I did not know about options(error = recover). I also like to use the \'debug\' package that gives a visual way to step through your code.

require(debug)
mtrace(foo)
foo(1)

At this point it opens up a separate debug window showing your function, with a yellow line showing where you are in the code. In the main window the code enters debug mode, and you can keep hitting enter to step through the code (and there are other commands as well), and examine variable values, etc. The yellow line in the debug window keeps moving to show where you are in the code. When done with debugging, you can turn off tracing with:

mtrace.off()

回答10:

Based on the answer I received here, you should definitely check out the options(error=recover) setting. When this is set, upon encountering an error, you\'ll see text on the console similar to the following (traceback output):

> source(<my filename>)
Error in plot.window(...) : need finite \'xlim\' values
In addition: Warning messages:
1: In xy.coords(x, y, xlabel, ylabel, log) : NAs introduced by coercion
2: In min(x) : no non-missing arguments to min; returning Inf
3: In max(x) : no non-missing arguments to max; returning -Inf

Enter a frame number, or 0 to exit   

1: source(<my filename>)
2: eval.with.vis(ei, envir)
3: eval.with.vis(expr, envir, enclos)
4: LinearParamSearch(data = dataset, y = data.frame(LGD = dataset$LGD10), data.names = data
5: LinearParamSearch.R#66: plot(x = x, y = y.data, xlab = names(y), ylab = data.names[i])
6: LinearParamSearch.R#66: plot.default(x = x, y = y.data, xlab = names(y), ylab = data.nam
7: LinearParamSearch.R#66: localWindow(xlim, ylim, log, asp, ...)
8: LinearParamSearch.R#66: plot.window(...)

Selection:

At which point you can choose which \"frame\" to enter. When you make a selection, you\'ll be placed into browser() mode:

Selection: 4
Called from: stop(gettextf(\"replacement has %d rows, data has %d\", N, n), 
    domain = NA)
Browse[1]> 

And you can examine the environment as it was at the time of the error. When you\'re done, type c to bring you back to the frame selection menu. When you\'re done, as it tells you, type 0 to exit.

回答11:

I gave this answer to a more recent question, but am adding it here for completeness.

Personally I tend not to use functions to debug. I often find that this causes as much trouble as it solves. Also, coming from a Matlab background I like being able to do this in an integrated development environment (IDE) rather than doing this in the code. Using an IDE keeps your code clean and simple.

For R, I use an IDE called \"RStudio\" (http://www.rstudio.com), which is available for windows, mac, and linux and is pretty easy to use.

Versions of Rstudio since about October 2013 (0.98ish?) have the capability to add breakpoints in scripts and functions: to do this, just click on the left margin of the file to add a breakpoint. You can set a breakpoint and then step through from that point on. You also have access to all of the data in that environment, so you can try out commands.

See http://www.rstudio.com/ide/docs/debugging/overview for details. If you already have Rstudio installed, you may need to upgrade - this is a relatively new (late 2013) feature.

You may also find other IDEs that have similar functionality.

Admittedly, if it\'s a built-in function you may have to resort to some of the suggestions made by other people in this discussion. But, if it\'s your own code that needs fixing, an IDE-based solution might be just what you need.

回答12:

To debug Reference Class methods without instance reference

ClassName$trace(methodName, browser)

回答13:

I am beginning to think that not printing error line number - a most basic requirement - BY DEFAILT- is some kind of a joke in R/Rstudio. The only reliable method I have found to find where an error occurred is to make the additional effort of calloing traceback() and see the top line.