aebilgrau / gmcm Goto Github PK

The defaults of xlab and ylab in plot.theta-function are wrong.

Create MStep in Rcpp.

Create EStep in Rcpp.

The usage in the get.IDR docs should not display get.idr as it is no longer exported to the users.

The automatic download for the vignette is broken due to a redirect.

trying URL ‘http://i.space.com/images/i/000/010/634/original/shuttle-atlantis-lifts-off.jpg?1309297321’
Warning in download.file(url, destfile = fig7.file, method = "internal", :
  "internal" method cannot handle https redirection to: 'https://www.space.com/images/i/000/010/634/original/shuttle-atlantis-lifts-off.jpg?1309297321'
Quitting from lines 128-128 (GMCM-JStatSoft.Rnw) 

The file does not appear to be hosted at space.com anymore.

See file.

Warning when using a large mu/sd:

real.par <- c(0.6, 1.3, 2, 0.3)
sim <- SimulateGMCMData(n = 500, par = real.par, d = 6)
res5 <- fit.meta.GMCM(u = sim$u,
                          init.par = c(0.1, 11, 11, 0.9),
                          method = "L-BFGS-B", max.ite = 1000,
                          verbose = TRUE)
Warning messages:
1: In regularize.values(x, y, ties, missing(ties)) :
  collapsing to unique 'x' values
2: In regularize.values(x, y, ties, missing(ties)) :
  collapsing to unique 'x' values
3: In regularize.values(x, y, ties, missing(ties)) :
  collapsing to unique 'x' values
4: In regularize.values(x, y, ties, missing(ties)) :
  collapsing to unique 'x' values
5: In regularize.values(x, y, ties, missing(ties)) :
  collapsing to unique 'x' values
6: In regularize.values(x, y, ties, missing(ties)) :
  collapsing to unique 'x' values

• fit.special.GMCM for fit.full.GMCM
• fit.general.GMCM for fit.meta.GMCM

For at least the special model, the report from the shiny app are not always in agreement. This should be expected for fitting methods other than SANN.

Somewhat like #33, is.theta fails and warns with either rownames or colnames (or both) are present for the covariance matrices.

library("GMCM")
set.seed(1)
n <- 1000
d <- 5
m <- 5
data <- SimulateGMMData(n = n, d = d, m = m) # Simulate some data
start.theta <- rtheta(d = d, m = m)
res <- GMCM:::EMAlgorithm(data$z, theta = start.theta, max.ite = 1000,
trace.theta = TRUE, verbose = TRUE)
iteration 2 Delta loglik = 43233.53

error: inv_sympd(): matrix appears to be singular

Error: inv_sympd(): matrix appears to be singular

Without a license, this project isn't open source and no one can use the code.

Please use something like http://choosealicense.com/ to decide what license to use. I recommend MIT or GPL.

A novel effort by Add a License Please

and vice versa...

A option for users to provide the marginal distributions for SimulateGMCMData as a list.

Due to too many spuriously committed PDF Rplorts ("Rplots1.pdf" through "Rplots999.pdf"), dev.new() can no longer choose a new name.

See

https://travis-ci.org/AEBilgrau/GMCM/jobs/570560854

dev.new(): using pdf(file="Rplots999.pdf")
Error in dev.new(file = NULL) : no suitable unused file name for pdf()

By default, show only the used columns in the "Classified data" boxes. Have a checkbox that toggles if the results should be attached to the "input data" or only the used columns.

Will allow for including NEWS automatically in the pkgdown site

… in the general GMCM to avoid get.prob combined with apply(probs, 1, which.max), etc.

Have some within app test data to try out the app without bringing data yourself.

Re. #19, the following features could be considered for the GMCM:

Input file handling:

• Subset rows which are passed to model
• Better "pre-processing". Allow to apply basic functions on chosen cols.

Special case GMCM

• Displaying more clearly exactly which cols have been used and the dimension of the model

• A "fit log" -- captured output from fit.meta.GMCM (?)

• Ability to provide new values to predict

• 3d plots when applicable

Other

• Add various alters, models, popups, etc, to guide users safely through
• Add usage to README, update NEWS, deploy app somewhere

Add a tooltip to "Important: Check if larger values in data indicate stronger evidence" to explain further.

plot.theta sometimes fail when add.ellipses = TRUE and non-default which.dims = c(3,2). Seems to be due to the built in which argument of ellipse::ellipse.

The 'Scatter plot of raw data' panel gives a momentary error Error: figure margins too large when selected the second model column.

Consider adding some goodness of fit measures (request by e-mail).

currently only plotted.

Possibly also the reverse is true

Parallelize dgmcm.loglik.R for faster optimization.

Sometimes, optim throws an error from fit.meta.GMCM about infinite objective function values.

An example which reproduce this error is:

library(GMCM)
set.seed(3)
true.par <- c(0.95, 3, 2, 0.5)
data <- SimulateGMCMData(d = 3, n = 10000, par = true.par)
start.par <- c(0.32, 3.62, 3.74, 0.08)
fit <- fit.meta.GMCM(Uhat(data$u), init.par = start.par, method = "L-BFGS")

And this yield the error:

Error in optim(inv.tt(init.par, d = d, positive.rho = positive.rho), meta.gmcm.loglik,  : 
  L-BFGS-B needs finite values of 'fn'

In the shiny app, add process spinners when computing.

Add some site articles:

• PDF JStatSoft paper or Vignette (see https://pkgdown.r-lib.org/articles/pkgdown.html)
• General model example
• Special model example (move from README)
• Shiny

Benchmarking?

Reproducible example:

library("GMCM")

n <- 1000
set.seed(123)
data <- data.frame(X1 = rnorm(n),
                   X2 = rchisq(n, df = 10),
                   X3 = rpois(n, lambda = 100))

udata <- Uhat(data)

fitted.theta <- fit.full.GMCM(u = udata, m = 1, method = "NM", verbose = FALSE)

http://cran.r-project.org/package=clue

Make a custom badge in the README showing CRAN test results from here.

Set up a very basic test suite to expand.

Instead of manually relisting the "unlisted" parameter list as done in the functions tt and tt.inv use relist. Probably more safe and faster?. See ?relist.

Implement the theta format as proper S3 object class and define methods.

Instead of sliders (which are only accurate to some degree) have a numeric input as well.

fit.full.GMCM(..., trace.theta=TRUE, method="PEM") does not appear to work properly.

The lower-left hand notification are easily missed; improve their visibility.

Much like the special model, show the results of the classification.

Roxygenize all functions in GMCM.

Improve the interface with a theta() constructor. Looks like list(…) but evaluates lazily and checks specified names, and applies as.theta, etc.

Code like this is annoying:

start.theta <- 
  list(m = 3,
       d = 4,
       pie = c(1,1,1)/3,
       mu = list(comp1 = rep(0, 4), comp2 = c(1,0,0,0), comp3 = c(0,1,0,0)),
       sigma = replicate(3, diag(4), simplify = FALSE))

start.theta <- as.theta(start.theta)