Quanteda regression

[eddelbuettel]

Those watching this repo and hence seeing this probably also follow the rcpp-devel list.

The quanteda by @kbenoit et al package came up as it at times seems to fail tests for me / us / CRAN. Now that RcppArmadillo 0.760.1.0 is on CRAN, it is seen to let quanteda fail 14 times:

* checking tests ... [59s/36s] ERROR
  Running ‘testthat.R’ [58s/36s]
Running the tests in ‘tests/testthat.R’ failed.
Last 13 lines of output:
  testthat results ================================================================
  OK: 1390 SKIPPED: 19 FAILED: 17
  1. Error: test textplot_scale1d wordfish in the most basic way (@test-plots.R#96) 
  2. Error: textmodel-ca works as expected as ca::ca : use mt (@test-textmodel_ca.R#27) 
  3. Error: textmodel-ca(sparse) works as expected on another dataset (@test-textmodel_ca.R#68) 
  4. Error: textmodel-wordfish works as expected: dense vs sparse vs sparse+mt (@test-textmodel_wordfish.R#22) 
  5. Error: print/show/summary method works as expected (@test-textmodel_wordfish.R#32) 
  6. Error: coef works for wordfish fitted (@test-textmodel_wordfish.R#64) 
  7. Error: textmodel-wordfish works for quasipoisson - feature as expected: dense vs sparse vs sparse+mt (@test-textmodel_wordfish.R#81) 
  8. Error: textmodel-wordfish works for quasipoisson - overall as expected: dense vs sparse vs sparse+mt (@test-textmodel_wordfish.R#103) 
  9. Error: textmodel-wordfish (sparse) works as expected on another dataset (@test-textmodel_wordfish.R#119) 
  1. ...
  
  Error: testthat unit tests failed
  Execution halted
* checking for unstated dependencies in vignettes ... OK

Most of these are of the dfmSparse is not supported variety leading to our as<>() caster.

For testing, I downgraded to the previously release version 0.7.900.2.0 of RcppArmadillo -- and with it quateda 0.99 (which itself only got onto CRAN after we submitted) passes.

@binxiangni Could you possibly take a quick look if either we (or quanteda) can accomodate with a simple cast? I can take a look too but I am simply not that deep in the sparse code ...

[kbenoit]

If we can help please let me know, I am on holidays for a week from Sunday but @koheiw and/or @HaiyanLW could assist with this. dfmSparse inherits from Matrix::dgCMatrix and we could easily change our code drop back to that before sending to RcppArmadillo functions if needed.

[eddelbuettel]

Appreciate the super-quick reply, Ken! I suspected something exactly along those lines. It could be our code is too simplistic ie in R I learned to test via inherits() rather than == for objects that subclass as eg in

R> x <- xts(1:3, order.by=Sys.time()+0:2)
R> class(x)
[1] "xts" "zoo"
R> 

Can you point us too where dfmSparse is set up? Maybe we need to change how we check, maybe it is your side, ... but I am sure we get this squared.

This was essentially a "regression test race condition" as we both made changes at the same time. But there is a lot more sparse matrix code on our side which will hopefully help you all too.

[eddelbuettel]

It think it may be as simple as your

setClass("dfmSparse",
         contains = c("dfm", "dgCMatrix"))

not meshing with our side of

            // Get the type of sparse matrix
            std::string type = Rcpp::as<std::string>(mat.slot("class"));
            if (type == "dgCMatrix") {

[eddelbuettel]

It all works if we just make this change to RcppArmadillo:

modified   inst/include/RcppArmadilloAs.h
@@ -97,7 +97,7 @@ namespace traits {
             
             // Get the type of sparse matrix
             std::string type = Rcpp::as<std::string>(mat.slot("class"));
-            if (type == "dgCMatrix") {
+            if (type == "dgCMatrix" || type == "dfmSparse") {
                 IntegerVector i = mat.slot("i");
                 IntegerVector p = mat.slot("p");
                 Vector<RTYPE> x = mat.slot("x");

@dmbates Is there a more S4-ish trick which could be done from client-package side? We can't possibly enumerate all possible S4 classes on our end, but we also shouldn't have to force people to stick with dgcMatrix and the like. I am clearly missing a (possibly simple) piece here...

[eddelbuettel]

I think I have idea:

• In quanteda, we provide a proxy class quantedaMat (or qmat for short) which inherist from arma::spmat.
• It checks that the incoming class is dfmSparse, and then sets it to dgCMatrix.

Thoughts?

[eddelbuettel]

Sounds like we exactly that (from today's R-devel NEWS feed)

CHANGES IN R-devel BUG FIXES

Internal C-level ‘R_check_class_and_super()’ and hence ‘R_check_class_etc()’ now also consider non-direct super classes and hence return a match in more cases. This e.g., fixes behaviour of derived classes in package ‘Matrix’.

[eddelbuettel]

Realized (on train) that the shim class may create problems as it will likely affect behaviour once back in R too.

So the better bet may be to go with is() as suggsted by Kurt:

R> class(dfm1)
[1] "dfmSparse"
attr(,"package")
[1] "quanteda"
R> "dgCMatrix" %in% methods::is(dfm1)          ## this does the Right Thing (TM)
[1] TRUE
R> 

[binxiangni]

Sorry, I don't understand your latest solution. Do you mean a minor modification on the quanteda side? I found it hard to change the conditional statement in as<>(), since the slot type is just a character vector and we cannot detect its inheritance in Rcpp.

[eddelbuettel]

We can access the line "dgCMatrix" %in% methods::is(dfm1) because we can access R functions. It is kinda ugly, but the thing we need here.

So I guess I will implement that, and we then change all your class == "thisClass" string comparisons with an additional || isIt(class, "thisClass") where the to-be-witten isIt() will basically do this one line using is().

With the news from R-devel we probably have a chance to conditionally on R-devel (and then R 3.5.* next) do it in C code. And as the || is less expensive at the C++ level, I think we're fine.

In short: we just generalize your comparisons. At least, that is my current plan ... but then I am also currently at work and not looking at this.

[eddelbuettel]

The following works (which is a start ;-) but feels a little ... rushed.

Should the helper function be somewhere else? In Rcpp?

modified   inst/include/RcppArmadilloAs.h
@@ -25,6 +25,19 @@
 
 namespace Rcpp{
 
+    inline bool isIt(const std::string cls, Rcpp::S4 s) {
+        static Rcpp::Environment mthds("package:methods");
+        static Rcpp::Function is = mthds["is"];
+  
+        Rcpp::CharacterVector res = is(s);
+        bool itis = false;
+        for (int i=0; !itis && i<res.size(); i++) {
+            itis = strcmp(cls.c_str(), res[i]) == 0;
+        }
+        return itis;
+    }
+
+    
 namespace traits {
 
     template <typename T> 
@@ -97,7 +110,7 @@ namespace traits {
             
             // Get the type of sparse matrix
             std::string type = Rcpp::as<std::string>(mat.slot("class"));
-            if (type == "dgCMatrix") {
+            if (type == "dgCMatrix" || isIt("dgCMatrix", mat)) {
                 IntegerVector i = mat.slot("i");
                 IntegerVector p = mat.slot("p");
                 Vector<RTYPE> x = mat.slot("x");
@@ -110,7 +123,7 @@ namespace traits {
                 std::copy(p.begin(), p.end(), arma::access::rwp(res.col_ptrs));
                 std::copy(x.begin(), x.end(), arma::access::rwp(res.values));
             }
-            else if (type == "dtCMatrix") {
+            else if (type == "dtCMatrix" || isIt("dtCMatrix", mat)) {
                 IntegerVector i = mat.slot("i");
                 IntegerVector p = mat.slot("p");
                 Vector<RTYPE> x = mat.slot("x");
@@ -128,7 +141,7 @@ namespace traits {
                     res.diag().ones();
                 }
             }
-            else if (type == "dsCMatrix") {
+            else if (type == "dsCMatrix" || isIt("dsCMatrix", mat)) {
                 IntegerVector i = mat.slot("i");
                 IntegerVector p = mat.slot("p");
                 Vector<RTYPE> x = mat.slot("x");
@@ -148,7 +161,7 @@ namespace traits {
                     res = symmatl(res);
                 }
             }
-            else if (type == "dgTMatrix") {
+            else if (type == "dgTMatrix" || isIt("dgTMatrix", mat)) {
                 IntegerVector ti = mat.slot("i");
                 IntegerVector tj = mat.slot("j");
                 Vector<RTYPE> tx = mat.slot("x");
@@ -200,7 +213,7 @@ namespace traits {
                 std::copy(p.begin(), p.end(), arma::access::rwp(res.col_ptrs));
                 std::copy(x.begin(), x.end(), arma::access::rwp(res.values));
             }
-            else if (type == "dtTMatrix") {
+            else if (type == "dtTMatrix" || isIt("dtTMatrix", mat)) {
                 IntegerVector ti = mat.slot("i");
                 IntegerVector tj = mat.slot("j");
                 Vector<RTYPE> tx = mat.slot("x");
@@ -257,7 +270,7 @@ namespace traits {
                     res.diag().ones();
                 }
             }
-            else if (type == "dsTMatrix") {
+            else if (type == "dsTMatrix" || isIt("dsTMatrix", mat)) {
                 IntegerVector ti = mat.slot("i");
                 IntegerVector tj = mat.slot("j");
                 Vector<RTYPE> tx = mat.slot("x");
@@ -316,7 +329,7 @@ namespace traits {
                     res = symmatl(res);
                 }
             }
-            else if (type == "dgRMatrix") {
+            else if (type == "dgRMatrix" || isIt("dgRMatrix", mat)) {
                 IntegerVector rj = mat.slot("j");
                 IntegerVector rp = mat.slot("p");
                 Vector<RTYPE> rx = mat.slot("x");
@@ -366,7 +379,7 @@ namespace traits {
                 std::copy(p.begin(), p.end(), arma::access::rwp(res.col_ptrs));
                 std::copy(x.begin(), x.end(), arma::access::rwp(res.values));
             }
-            else if (type == "dtRMatrix") {
+            else if (type == "dtRMatrix" || isIt("dtRMatrix", mat)) {
                 IntegerVector rj = mat.slot("j");
                 IntegerVector rp = mat.slot("p");
                 Vector<RTYPE> rx = mat.slot("x");
@@ -421,7 +434,7 @@ namespace traits {
                     res.diag().ones();
                 }
             }
-            else if (type == "dsRMatrix") {
+            else if (type == "dsRMatrix" || isIt("dsRMatrix", mat)) {
                 IntegerVector rj = mat.slot("j");
                 IntegerVector rp = mat.slot("p");
                 Vector<RTYPE> rx = mat.slot("x");
@@ -478,7 +491,7 @@ namespace traits {
                     res = symmatl(res);
                 }
             }
-            else if (type == "indMatrix") {
+            else if (type == "indMatrix" || isIt("indMatrix", mat)) {
                 std::vector<int> i;
                 IntegerVector p(ncol + 1);
                 IntegerVector x(nrow, 1);
@@ -519,7 +532,7 @@ namespace traits {
                 std::copy(p.begin(), p.end(), arma::access::rwp(res.col_ptrs));
                 std::copy(x.begin(), x.end(), arma::access::rwp(res.values));
             }
-            else if (type == "pMatrix") {
+            else if (type == "pMatrix" || isIt("pMatrix", mat)) {
                 std::vector<int> i;
                 IntegerVector p(ncol + 1);
                 IntegerVector x(ncol, 1);
@@ -550,7 +563,7 @@ namespace traits {
                 std::copy(p.begin(), p.end(), arma::access::rwp(res.col_ptrs));
                 std::copy(x.begin(), x.end(), arma::access::rwp(res.values));
             }
-            else if (type == "ddiMatrix") {
+            else if (type == "ddiMatrix" || isIt("ddiMatrix", mat)) {
                 std::vector<int> i;
                 std::vector<int> p;
                 std::vector<double> x;

[kevinushey]

It looks like S4 objects already do have an is() method -- perhaps we can just call that? E.g.

#include <Rcpp.h>
using namespace Rcpp;

// [[Rcpp::export]]
bool cpp_inherits(S4 object, std::string clazz)
{
  return object.is(clazz);
}

/*** R
library(Matrix)
m <- Matrix(1)
cpp_inherits(m, "dsyMatrix")
cpp_inherits(m, "Matrix")
*/

[eddelbuettel]

Crikey. Did I just reimplement that?

[kevinushey]

More or less :-)

The Rcpp code does something similar, but extracts the class definition directly and looks to see if the compared class is a known parent class.

[eddelbuettel]

And exactly what we needed. Thanks for cluebat, Kevin :)

R> cpp_inherits(dfm1, "dgCMatrix")     # after library(quanteda); example(dfm)
[1] TRUE
R> 

[eddelbuettel]

Well I didn't commit it so maybe it doesn't count, right? ;-)

That'll make the right generalization for the test and is what we needed. All good.