To see a demonstration of the capabilities of liquidSVM from an R viewpoint, please look at the demo.

Disclaimer: liquidSVM and the R-bindings are in general quite stable and well tested by several people. However, use in production is at your own risk.

If you run into problems please check first the documentation for more details, or report the bug to the maintainer.

# Installation

There are several options to install the package.

The most convenient way is to use the standard install to get it from CRAN:

install.packages("liquidSVM")


You can also use our repository:

install.packages("liquidSVM", repos="http://www.isa.uni-stuttgart.de/software/R")


Remark that in R a package can be installed either as source or binary

Source (default on Linux Systems): : Allows for optimization to the system and liquidSVM can benefit a lot from these optimizations. The drawback is that this needs a C++ compiler. This is usually okay on Linux-systems, but on Windows one has to install https://cran.r-project.org/bin/windows/Rtools/, and on MacOS X https://itunes.apple.com/de/app/xcode/id497799835?mt=12 (https://developer.apple.com/download/.)

Binary (default on Windows and MacOS X): : compiled versions are provided, so you do not need compilers. However, these are optimized for generic processors (e.g. they do not use AVX), and hence you might do much better on your machine if you compile it yourself.

You can change the default behaviour of install.packages(...) under Windows/MacOS by using the parameter type="source".

The binaries in our repository are only compiled using R version 3.*. If you use another version, they might not work and you have to try source installation (type="source").

Note: on MacOS X there can be an issue with binary package installation. If you get the error tar: Failed to set default locale then consult

https://cran.r-project.org/bin/macosx/RMacOSX-FAQ.html#Internationalization-of-the-R_002eapp

Download the source or binary package from http://www.isa.uni-stuttgart.de/software/. On the command line use:

R CMD INSTALL path-to-package/liquidSVM_1.0.1.tar.gz
# Windows
Rcmd INSTALL path-to-package/liquidSVM_1.0.1.zip
# MacOS X using Termninal
R CMD INSTALL path-to-package/liquidSVM_1.0.1.tgz


or in a running R session:

install.packages("path-to-package/liquidSVM_1.0.1.tar.gz",repos=NULL)
# Windows binary
install.packages("path-to-package/liquidSVM_1.0.1.zip",repos=NULL)
# MacOS X binary
install.packages("path-to-package/liquidSVM_1.0.1.tgz",repos=NULL)


You can use also the means of any R-IDE. E.g. in RStudio go to the menu

Tools > Install packages...


Then set install from to package archive file (.tar.gz or .tgz) and choose your package and install the package.

## Advanced Configuration Options with Source Install

liquidSVM can be configured for different uses of available hardware. We provide the following configurations:

native : compiles for the current system, e.g. uses AVX or even AVX2 if available. This uses g++/clang++ -march=native -O3.

generic : compiles for a wide range of currently deployed CPUs (uses SSE). This uses g++/clang++ -mtune=generic -O3. Our binary packages are compiled with this configuration.

default : compiles using the default values provided by R.

debug : compiles with debugging enabled.

empty : gives no default compile arguments.

Additional compiler flags can be provided as well. On the command line, here are some examples:

R CMD INSTALL --configure-args=native path-to-package/liquidSVM_1.0.1.tar.gz
R CMD INSTALL --configure-args=generic path-to-package/liquidSVM_1.0.1.tar.gz
R CMD INSTALL --configure-args="empty -march=core2 -O3" path-to-package/liquidSVM_1.0.1.tar.gz


or in a running R session:

install.packages("liquidSVM",configure.args="native")
install.packages("liquidSVM",configure.args="generic")
install.packages("liquidSVM",configure.args="empty -march=core2 -O3")


Under MacOS you have to add the paramter type="source" in order to trigger compilation.

Hint: to see whether liquidSVM got compiled with SSE and/or AVX use:

compilationInfo()
#> [1] "Compiled without vectorization"


### Windows configuration

On Windows unfortunately neither --configure-args nor configure.args have any effect. We enable compilation configuration by reading the environment variable LIQUIDSVM_CONFIGURE_ARGS and using it in the same way as the configure args on the other platforms (see above). So on the Windows command line use

set LIQUIDSVM_CONFIGURE_ARGS=native
R CMD INSTALL path-to-package/liquidSVM_1.0.1.tar.gz

set LIQUIDSVM_CONFIGURE_ARGS=empty -march=core2 -O3
R CMD INSTALL path-to-package/liquidSVM_1.0.1.tar.gz


Remark that no quotation has to be used. It is not tested whether paths with spaces will work in this setting.

If you wish to install from within R you can specify the environment variable as well:

Sys.setenv(LIQUIDSVM_CONFIGURE_ARGS="native")
install.packages("liquidSVM")

Sys.setenv(LIQUIDSVM_CONFIGURE_ARGS="empty -march=core2 -O3")
install.packages("liquidSVM")


If you have https://cran.r-project.org/bin/windows/Rtools/ installed then you should definitely try to use native, because on Windows we use generic as the default configuration even for source installs.

### Common Problems

• MacOS X: It seems that in some cases clang++ -march=native does not activate AVX even if it is available. Hence if you know it is available, use configure.args="native -mavx" or even configure.args="native -mavx2".
• Windows: On one machine set LIQUIDSVM_CONFIGURE_ARGS=native compiled but crashed on execution: the compiler thought that FusedMultiplyAdd was available but it was not. The solution was to set LIQUIDSVM_CONFIGURE_ARGS=native -mno-fma

For GCC it can help to use g++ -Q --help=target -march=native ... to figure out which options trigger what optimizations. For both GCC and clang you can also print the compilation headers by g++ -march=native ... -dM -E - < /dev/null | egrep "SSE|AVX".

## CUDA

liquidSVM also is able to calculate the kernel on a GPU if it is compiled with CUDA-support. Since there is a big overhead in moving the kernel matrix from the GPU memory, this is most useful for problems with many feature-dimensions (see demo)

To activate CUDA support you have to specify its location (usually /usr/local/cuda) as a parameter to the configure arguments:

R CMD INSTALL --configure-args="native /my/path/to/cuda" path-to-package/liquidSVM_1.0.1.tar.gz


or again in R

install.packages('liquidSVM',configure.args="native /my/path/to/cuda")


Note that due to lack of testing machines this is known to work only on some Linux machines. The above instructions will probably not work on Windows!

If you have compiled with CUDA-support, you can activate it for a computation by using svm(..., GPUs=1):

# Configuration parameters

The uses of svm(...), lsSVM(...), mcSVM(...), etc. can be configured using the following parameters.

## Overview of Configuration Parameters

display : This parameter determines the amount of output of you see at the screen: The larger its value is, the more you see. This can help as a progress indication.

scale : If set to a true value then for every feature in the training data a scaling is calculated so that its values lie in the interval $$[0,1]$$. The training then is performed using these scaled values and any testing data is scaled transparently as well.

Because SVMs are not scale-invariant any data should be scaled
for two main reasons: First that all features have the same weight,
and second to assure that the default gamma parameters that liquidSVM
provide remain meaningful.

If you do not have scaled the data previously this is an easy option.


threads : This parameter determines the number of cores used for computing the kernel matrices, the validation error, and the test error.

* threads=0 (default) means that all physical cores of your CPU run one thread.
* threads=-1 means that all but one physical cores of your CPU run one thread.


partition_choice : This parameter determines the way the input space is partitioned. This allows larger data sets for which the kernel matrix does not fit into memory.

* partition_choice=0 (default) disables partitioning.
* partition_choice=6 gives usually highest speed.
* partition_choice=5 gives usually the best test error.


grid_choice : This parameter determines the size of the hyper- parameter grid used during the training phase. Larger values correspond to larger grids. By default, a 10x10 grid is used. Exact descriptions are given in the next section.

adaptivity_control : This parameter determines, whether an adaptive grid search heuristic is employed. Larger values lead to more aggressive strategies. The default adaptivity_control = 0 disables the heuristic.

random_seed : This parameter determines the seed for the random generator. random_seed = -1 uses the internal timer create the seed. All other values lead to repeatable behavior of the svm.

folds : How many folds should be used.

## Specialized configuration parameters

Parameters for regression (least-squares, quantile, and expectile)

clipping : This parameter determines whether the decision functions should be clipped at the specified value. The value clipping = -1.0 leads to an adaptive clipping value, whereas clipping = 0 disables clipping.

Parameter for multiclass classification determine the multiclass strategy: mc-type=0 : AvA with hinge loss. mc-type=1 : OvA with least squares loss. mc-type=2 : OvA with hinge loss. mc-type=3 : AvA with least squares loss.

Parameters for Neyman-Pearson Learning

class : The class, the constraint is enforced on.

constraint : The constraint on the false alarm rate. The script actually considers a couple of values around the value of constraint to give the user an informed choice.

## Hyperparameter Grid

For Support Vector Machines two hyperparameters need to be determined:

• gamma the bandwith of the kernel
• lambda has to be chosen such that neither over- nor underfitting happen. lambda values are the classical regularization parameter in front of the norm term.

liquidSVM has a built-in a cross-validation scheme to calculate validation errors for many values of these hyperparameters and then to choose the best pair. Since there are two parameters this means we consider a two-dimensional grid.

For both parameters either specific values can be given or a geometrically spaced grid can be specified.

gamma_steps, min_gamma, max_gamma : specifies in the interval between min_gamma and max_gamma there should be gamma_steps many values

gammas : e.g. gammas=c(0.1,1,10,100) will do these four gamma values

lambda_steps, min_lambda, max_lambda : specifies in the interval between min_lambda and max_lambda there should be lambda_steps many values

lambdas : e.g. lambdas=c(0.1,1,10,100) will do these four lambda values

c_values : the classical term in front of the empirical error term, e.g. c_values=c(0.1,1,10,100) will do these four cost values (basically inverse of lambdas)

Note the min and max values are scaled according the the number of samples, the dimensionality of the data sets, the number of folds used, and the estimated diameter of the data set.

Using grid_choice allows for some general choices of these parameters

grid_choice 0 1 2
gamma_steps 10 15 20
lambda_steps 10 15 20
min_gamma 0.2 0.1 0.05
max_gamma 5.0 10.0 20.0
min_lambda 0.001 0.0001 0.00001
max_lambda 0.01 0.01 0.01

Using negative values of grid_choice we create a grid with listed gamma and lambda values:

grid_choice -1
gammas c(10.0, 5.0, 2.0, 1.0, 0.5, 0.25, 0.1, 0.05)
lambdas c(1.0, 0.1, 0.01, 0.001, 0.0001, 0.00001, 0.000001, 0.0000001)
grid_choice -2
gammas c(10.0, 5.0, 2.0, 1.0, 0.5, 0.25, 0.1, 0.05)
c_values c(0.01, 0.1, 1, 10, 100, 1000, 10000)

An adaptive grid search can be activated. The higher the values of MAX_LAMBDA_INCREASES and MAX_NUMBER_OF_WORSE_GAMMAS are set the more conservative the search strategy is. The values can be freely modified.

ADAPTIVITY_CONTROL 1 2
MAX_LAMBDA_INCREASES 4 3
MAX_NUMBER_OF_WORSE_GAMMAS 4 3

## Cells

A major issue with SVMs is that for larger sample sizes the kernel matrix does not fit into the memory any more. Classically this gives an upper limit for the class of problems that traditional SVMs can handle without significant runtime increase. Furthermore also the time complexity is at least $$O(n^2)$$.

liquidSVM implements two major concepts to circumvent these issues. One is random chunks which is known well in the literature. However we prefer the new alternative of splitting the space into spatial cells and use local SVMs on every cell.

If you specify useCells=TRUE then the sample space $$X$$ gets partitioned into a number of cells. The training is done first for cell 1 then for cell 2 and so on. Now, to predict the label for a value $$x\in X$$ liquidSVM first finds out to which cell this $$x$$ belongs and then uses the SVM of that cell to predict a label for it.

If you run into memory issues turn cells on: useCells=TRUE

This is quite performant, since the complexity in both time and memore are both $$O(\mbox{CELLSIZE} \times n)$$ and this holds both for training as well as testing! It also can be shown that the quality of the solution is comparable, at least for moderate dimensions.

The cells can be configured using the partition_choice:

1) This gives a partition into random chunks of size 2000

VORONOI=c(1, 2000)


2) This gives a partition into 10 random chunks

VORONOI=c(2, 10)


3) This gives a Voronoi partition into cells with radius not larger than 1.0. For its creation a subsample containing at most 50.000 samples is used.

  VORONOI=c(3, 1.0, 50000)


4) This gives a Voronoi partition into cells with at most 2000 samples (approximately). For its creation a subsample containing at most 50.000 samples is used. A shrinking heuristic is used to reduce the number of cells.

  VORONOI=c(4, 2000, 1, 50000)


5) This gives a overlapping regions with at most 2000 samples (approximately). For its creation a subsample containing at most 50.000 samples is used. A stopping heuristic is used to stop the creation of regions if 0.5 * 2000 samples have not been assigned to a region, yet.

VORONOI=c(5, 2000, 0.5, 50000, 1)


6) This splits the working sets into Voronoi like with PARTITION_TYPE=4. Unlike that case, the centers for the Voronoi partition are found by a recursive tree approach, which in many cases may be faster.

 VORONOI=c(6, 2000, 1, 50000, 2.0, 20, 4,)


The first parameter values correspond to NO_PARTITION, RANDOM_CHUNK_BY_SIZE, RANDOM_CHUNK_BY_NUMBER, VORONOI_BY_RADIUS, VORONOI_BY_SIZE, OVERLAP_BY_SIZE

## Weights

• qt, ex: Here the number of considered tau-quantiles/expectiles as well as the considered tau-values are defined. You can freely change these values but notice that the list of tau-values is space-separated!

• npl, roc: Here, you define, which weighted classification problems will be considered. The choice is usually a bit tricky. Good luck …

NPL:
WEIGHT_STEPS=10
MIN_WEIGHT=0.001
MAX_WEIGHT=0.5
GEO_WEIGHTS=1

ROC:
WEIGHT_STEPS=9
MAX_WEIGHT=0.9
MIN_WEIGHT=0.1
GEO_WEIGHTS=0


## Grouped Cross Validation

By specifying groupIds when initializing an SVM samples obtain group ids. This by default also sets FOLDS_KIND to GROUPED. If the latter is the case then samples with the same group id will be put into the same fold at cross validation. This is important if e.g. there are several patients with several measurements each.

## More Advanced Parameters

The following parameters should only employed by experienced users and are self-explanatory for these:

KERNEL : specifies the kernel to use, at the moment either GAUSS_RBF or POISSON

RETRAIN_METHOD : After training on grids and folds there are only solutions on folds. In order to construct a global solution one can either retrain on the whole training data (SELECT_ON_ENTIRE_TRAIN_SET) or the (partial) solutions from the training are kept and combined using voting (SELECT_ON_EACH_FOLD default)

store_solutions_internally : If this is true (default in all applicable cases) then the solutions of the train phase are stored and can be just reused in the select phase. If you slowly run out of memory during the train phase maybe disable this. However then in the select phase the best models have to be trained again.

For completeness here are some values that usually get set by the learning scenario

SVM_TYPE : KERNEL_RULE, SVM_LS_2D, SVM_HINGE_2D, SVM_QUANTILE, SVM_EXPECTILE_2D, SVM_TEMPLATE

LOSS_TYPE : CLASSIFICATION_LOSS, MULTI_CLASS_LOSS, LEAST_SQUARES_LOSS, WEIGHTED_LEAST_SQUARES_LOSS, PINBALL_LOSS, TEMPLATE_LOSS

VOTE_SCENARIO : VOTE_CLASSIFICATION, VOTE_REGRESSION, VOTE_NPL

KERNEL_MEMORY_MODEL : LINE_BY_LINE, BLOCK, CACHE, EMPTY

FOLDS_KIND : BLOCKS, ALTERNATING, RANDOM, STRATIFIED, GROUPED, RANDOM_SUBSET

WS_TYPE : FULL_SET, MULTI_CLASS_ALL_VS_ALL, MULTI_CLASS_ONE_VS_ALL, BOOT_STRAP

# Known Issues / Common Problems

• Ctrl-C / Interrupt is tricky. It works most of the time, but it can fail. If you get weird results or errors save your models and restart the R session.

• CUDA has not been tested neither on Windows nor on macOS.

• 32-bit has been seen to work but is not supported.

## Using external parallelization

liquidSVM does its own threading - hence do not parallelize on top of that, unless you know what you are doing. Hence just give the parameter threads=n or let the default use all of your physical cores.

If you really want to do it yourself you have to serialze the solutions. Furthermore you have to be carefule to assign disjoint cores else they will fight for the same core:

library(parallel)
## how big should the cluster be
workers <- 2
cl <- makeCluster(workers)
## how many threads should each worker use
## to make it interesting use disjoint parts of sml$train data <- sml$train[ seq(i,nrow(sml$train),workers) , ] ## the second argument to threads sets the offset of cores model <- lsSVM(Y~., data, threads=c(threads,threads*(i-1)) ) ## finally return the serialized solution serialize.liquidSVM(model) }) for(i in 1:workers){ ## get the solution in the master session model <- unserialize.liquidSVM(obj[[i]]) print(errors(test(model,sml$test)))