How to use Memoise in R - r

I have been trying to use memoise to speed up computations for a function I wrote.
In a nutshell, I have a function a() that performs do.call() on another function b().
function b() runs code to read files over the past year and then performs some calculations.
So if you wanted to run a() for 1/12/2019, 2/12/2019, 3/12/2019, b() is going to read files for one year prior to 1/12, 2/12, 3/12. E.g. Files read for 1/12/19 would be between 1/12/18 to 1/12/19, files read for 2/12/2019 would be between 2/12/2018 and 2/12/2019. Essentially, there will be a lot of overlap in the files that b() reads. so it would essentially be ideal to memoise an fread() function that b() uses.
I have written a sample function below
b<-function(){
##ideally would want to memoise it this way
if(!is.memoised(fread))
##fread=memoise(fread)
fread(...)
}
a<-function(){
do.call(b())
}
So I have discovered after many painful hours that this does not work in theory as the memoisation function has a constraint relating to how many times it can be called - i.e. if it is called more than once then it does not remember the previous function parameters that were used by the previously memoised fread(), essentially eliminating the usefulness of the function)
I suspect it has something to do with the scope of the memoised function, everytime b() is called it is memoising fread again
I tried to implement <<- for memoising fread, but I think I did not implement it the right way as it threw an error saying 'could not modify fread'
Would memoising fread in the global environment help? Again, I was not able to implement it the right way
is there something drastically long in the way i am thinking about these functions and memoisation?
Any input would be appreciated.

Related

Is it unwise to modify the class of functions in other packages?

There's a bit of a preamble before I get to my question, so hang with me!
For an R package I'm working on I'd like to make it as easy as possible for users to partially apply functions inline. I had the idea of using the [] operators to call my partial application function, which I've name "partialApplication." What I aim to achieve is this:
dnorm[mean = 3](1:10)
# Which would be exactly equivalent to:
dnorm(1:10, mean = 3)
To achieve this I tried defining a new [] method for objects of class function, i.e.
`[.function` <- function(...) partialApplication(...)
However, R gives a warning that the [] method for function objects is "locked." (Is there any way to override this?)
My idea seemed to be thwarted, but I thought of one simple solution: I can invent a new S3 class "partialAppliable" and create a [] method for it, i.e.
`[.partialAppliable` = function(...) partialApplication(...)
Then, I can take any function I want and append 'partialAppliable' to its class, and now my method will work.
class(dnorm) = append(class(dnorm), 'partialAppliable')
dnorm[mean = 3](1:10)
# It works!
Now here's my question/problem: I'd like users to be able to use any function they want, so I thought, what if I loop through all the objects in the active environment (using ls) and append 'partialAppliable' to the class of all functions? For instance:
allobjs = unlist(lapply(search(), ls))
#This lists all objects defined in all active packages
for(i in allobjs) {
if(is.function(get(i))) {
curfunc = get(i)
class(curfunc) = append(class(curfunc), 'partialAppliable')
assign(i, curfunc)
}
}
Voilà! It works. (I know, I should probably assign the modified functions back into their original package environments, but you get the picture).
Now, I'm not a professional programmer, but I've picked up that doing this sort of thing (globally modifying all variables in all packages) is generally considered unwise/risky. However, I can't think of any specific problems that will arise. So here's my question: what problems might arise from doing this? Can anyone think of specific functions/packages that will be broken by doing this?
Thanks!
This is similar to what the Defaults package did. The package is archived because the author decided that modifying other package's code is a "very bad thing". I think most people would agree. Just because you can do it, does not mean it's a good idea.
And, no, you most certainly should not assign the modified functions back to their original package environments. CRAN does not like when packages modify the users search path unnecessarily, so I would be surprised if they allowed a package to modify other package's function arguments.
You could work around that by putting all the modified functions in an environment on the search path. But then you have to ensure that environment is always searched first, which means modifying the search path every time another package is loaded.
Changing arguments for functions in other packages also has the potential to make it very difficult for others to reproduce your results because they must have all your argument settings. Unless you always call functions with all their arguments specified, which defeats the purpose of what you're trying to do.

should .RData files be used to store functions?

I use .RData files to store objects (e.g. lists, vectors, etc) then call them into other scripts, but I'm wondering whether they should also be used to store functions (most likely user-defined functions)?
I know source() is generally recommended for this purpose (and creating packages even more so), but an advantage as I see it is that a single .RData file can contain multiple objects - a list, dataframe, and the function, for example. Saves needing to call objects using load(), then the function separately, using source().
Are there reasons to be cautious of this approach, that I'm not seeing?
Thank you
At my old job, we used to serialize out closures:
> f <- (function(x) function() x)(2)
> f()
[1] 2
> saveRDS(f, file='/tmp/f')
and then
> f <- readRDS('/tmp/f')
> f()
[1] 2
This can let you bundle data (eg coefficients) with a function. Be careful though, your libraries won't get autoloaded.

how to make a function available at start up in R

I have a user defined function in R
blah=function(a,b){
something with a and b
}
is it possile to put this somewhere so that I do not need to remember to load in the workspace every time I start up R? Similar to a built in function like
summary(); t.test(); max(); sd()
You can put the function into your .rprofile file.
However, be very careful with what you put there, since it essentially makes your code non-reproducible — it now depends on your .rprofile:
Let’s say you have an R code file performing some analysis, and the code uses the function blah. Executing the code on any other system will fail due to the non-existence of the blah function.
As a consequence, this file should only contain system-specific setup. Don’t define helper functions in there — or if you do, make them defined only in interactive sessions, so that you have a clear environment when R is running a non-interactive script:
if (interactive()) {
# Helper functions go here.
}
And if you find yourself using the same helper functions over and over again, bundle them into packages (or modules) and reuse those.

R: Dealing with functions that sometimes crash the R session?

Have an R function ( let's call it MyFunction ) that sometimes crashes the R session , most of the time it does not.
Have to apply this function to a large number of objects in a sequential manner.
for(i in 1:nrow(objects))
{
result[i] <- MyFunction(objects[i]);
}
I'm coming from a C# background - where functions rarely crash the "session" and programmers normally surround such function calls in try - catch blocks. However, in R I've seen some functions that just crash the session and using tryCatch is of no help since the function does not cause an exception but a full blast session crash ;-)
Just wondering what's the best way of "catching" the crash.
I'm considering writing a Python script that calls the R function from Python ( via one of the R-Python connectors ) and catching the R crash in Python. Would that work ?
Any advise ?
Cheers !
Use the mcparallel function from the parallel package to run the function in a forked process. That way, if it crashes R, only the subprocess crashes, and an error is returned to the main process. If you want to apply this function to a large number of objects and collect the results in a list, use mclapply
Hello such behaviour is very rare in my experience. You might not know that there is a debuger that can help you to go step by step in you function.
install.packages('debug') #install the debug package
library(debug)
mtrace(myFunctionToBeDebuged) #this function will start the debuger
mtrace(myFunctionToBeDebuged, FALSE) #to stop the function to be traced
NOTE: when you are in the debuger, should you want to quit it do qqq()

How to prevent functions polluting global namespace?

My R project is getting increasingly complex, and I'm starting to look for some construct that's equivalent to classes in Java/C#, or modules in python, so that my global namespace doesn't become littered with functions that are never used outside of one particular .r file.
So, I guess my question is: to what extent is it possible to limit the scope of functions to within a specific .r file, or similar?
I think I can just make the entire .r file into one giant function, and put functions inside that, but that messes with the echoing:
myfile.r:
myfile <- function() {
somefunction <- function(a,b,c){}
anotherfunction <- function(a,b,c){}
# do some stuff here...
123
456
# ...
}
myfile()
Output:
> source("myfile.r",echo=T)
> myfile <- function() {
+ somefunction <- function(a,b,c){}
+ anotherfunction <- function(a,b,c){}
+
+ # do some stuff here...
+ # . .... [TRUNCATED]
> myfile()
>
You can see that "123" is not printed, even though we used echo=T in the source command.
I'm wondering if there is some other construct which is more standard, since putting everything inside a single function doesn't sound like something that is really standard? But perhaps it is? Also, if it means that echo=T works then that is a definite bonus for me.
Firstly, as #Spacedman has said, you'll be best served by a package but there are other options.
S3 Methods
R's original "object orientation" is known as S3. The majority of R's code base uses this particular paradigm. It is what makes plot() work for all kinds of objects. plot() is a generic function and the R Core Team and package developers can and have written their own methods for plot(). Strictly these methods might have names like plot.foo() where foo is a class of object for which the function defines a plot() method. The beauty of S3 is that you don't (hardly) ever need to know or call plot.foo() you just use plot(bar) and R works out which plot() method to dispatch to based on the class of object bar.
In your comments on your Question you mention that you have a function populate() that has methods (in effect) for classes "crossvalidate" and "prod" which you keep in separate .r files. The S3 way to set this up is to do:
populate <- function(x, ...) { ## add whatever args you want/need
UseMethod("populate")
}
populate.crossvalidate <-
function(x, y, z, ...) { ## add args but must those of generic
## function code here
}
populate.prod <-
function(x, y, z, ...) { ## add args but must have those of generic
## function code here
}
The given some object bar with class "prod", calling
populate(bar)
will result in R calling populate() (the generic), it then looks for a function with name populate.prod because that is the class of bar. It finds our populate.prod() and so dispatches that function passing on to it the arguments we initially specified.
So you see that you only ever refer to the methods using the name of the generic, not the full function name. R works out for you what method needs to be called.
The two populate() methods can have very different arguments, with exception that strictly they should have the same arguments as the generic function. So in the example above, all methods should have arguments x and .... (There is an exception for methods that employ formula objects but we don't need to worry about that here.)
Package Namespaces
Since R 2.14.0, all R packages have had their own namespace, even if one were not provided by the package author, although namespaces have been around for a lot longer in R than that.
In your example, we wish to register the populate() generic and it's two methods with the S3 system. We also wish to export the generic function. Usually we don't want or need to export the individual methods. So, pop your functions into .R files in the R folder of the package sources and then in the top level of the package sources create a file named NAMESPACE and add the following statements:
export(populate) ## export generic
S3method(populate, crossvalidate) ## register methods
S3method(populate, prod)
Then once you have installed your package, you will note that you can call populate() but R will complain if you try to call populate.prod() etc directly by name from the prompt or in another function. This is because the functions that are the individual methods have not been exported from the namespace and thence are not visible outside it. Any function in your package that call populate() will be able to access the methods you have defined, but any functions or code outside your package can't see the methods at all. If you want, you can call non-exported functions using the ::: operator, i.e.
mypkg:::populate.crossvalidate(foo, bar)
will work, where mypkg is the name of your package.
To be honest, you don't even need a NAMESPACE file as R will auto generate one when you install the package, one that automatically exports all functions. That way your two methods will be visible as populate.xxx() (where xxx is the particular method) and will operate as S3 methods.
Read Section 1 Creating R Packages in the Writing R Extensions manual for details of what is involved, but yuo won't need to do half of this if you don't want too, especially if the package is for your own use. Just create the appropriate package folders (i.e. R and man), stick your .R files in R. Write a single .Rd file in man where you add
\name{Misc Functions}
\alias{populate}
\alias{populate.crossvalidate}
\alias{populate.prod}
at the top of the file. Add \alias{} for any other functions you have. Then you'll need to build and install the package.
Alternative using sys.source()
Although I don't (can't!) really recommend what I mention below as a long-term viable option here, there is an alternative that will allow you to isolate the functions from individual .r files as you initially requested. This is achieved through the use of environments not namespaces and doesn't involve creating a package.
The sys.source() function can be used to source R code/functions from a .R file and evaluate it in an environment. As you .R file is creating/defining functions, if you source it inside another environment then those will functions will be defined there, in that environment. They won't be visible on the standard search path by default and hence a populate() function defined in crossvalidate.R will not clash with a populate() defined in prod.R as long as you use two separate environments. When you need to use one set of functions you can assign the environment to the search path, upon which it will then be miraculously visible to everything, and when you are done you can detach it. The attach the other environment, use it, detach etc. Or you can arrange for R code to be evaluated in a specific environment using things like eval().
Like I said, this isn't a recommended solution but it will work, after a fashion, in the manner you describe. For example
## two source files that both define the same function
writeLines("populate <- function(x) 1:10", con = "crossvalidate.R")
writeLines("populate <- function(x) letters[1:10]", con = "prod.R")
## create two environments
crossvalidate <- new.env()
prod <- new.env()
## source the .R files into their respective environments
sys.source("crossvalidate.R", envir = crossvalidate)
sys.source("prod.R", envir = prod)
## show that there are no populates find-able on the search path
> ls()
[1] "crossvalidate" "prod"
> find("populate")
character(0)
Now, attach one of the environments and call populate():
> attach(crossvalidate)
> populate()
[1] 1 2 3 4 5 6 7 8 9 10
> detach(crossvalidate)
Now call the function in the other environment
> attach(prod)
> populate()
[1] "a" "b" "c" "d" "e" "f" "g" "h" "i" "j"
> detach(prod)
Clearly, each time you want to use a particular function, you need to attach() its environment and then call it, followed by a detach() call. Which is a pain.
I did say you can arrange for R code (expressions really) to be evaluated in a stated environment. You can use eval() of with() for this for example.
> with(crossvalidate, populate())
[1] 1 2 3 4 5 6 7 8 9 10
At least now you only need a single call to run the version of populate() of your choice. However, if calling the functions by their full name, e.g. populate.crossvalidate() is too much effort (as per your comments) then I dare say that even the with() idea will be too much hassle? And anyway, why would you use this when you can quite easily have your own R package.
Don't worry about the complexity of 'making a package'. Stop thinking of it like that. What you are going to do is this:
in the folder where you are working on your project, make a folder called 'R'
put your R code in there, one function per file
make a DESCRIPTION file in your project directory. Check out existing examples for the exact format, but you only need a few fields.
Get devtools. install.packages("devtools")
Use devtools. library(devtools)
Now, write your functions in your R files in your R folder. To load them into R, DONT source them. Do load_all(). Your functions will be loaded but NOT into the global environment.
Edit one of your R files, then do load_all() again. This will load any modified files in the R folder, thus updating your function.
That's it. Edit, load_all, rinse and repeat. You have created a package, but its pretty lightweight and you don't have to deal with the bondage and discipline of R's package building tools.
I've seen, used, and even written code that tries to implement a lightweight packagey mechanism for loading objects, and none are as good as what devtools does.
All Hail Hadley!
You might want to consider making a package. As an alternative, you could look at environments. Finally, RStudio's projects may be closer to what would suit you.

Resources