version 0.1-3

cran · Jul 15, 2016 · 0d67ff8 · 0d67ff8
1 parent 5864315
commit 0d67ff8
Show file tree

Hide file tree

Showing 7 changed files with 34 additions and 34 deletions.
diff --git a/DESCRIPTION b/DESCRIPTION
@@ -1,13 +1,16 @@
 Package: tsne
 Type: Package
-Title: T-distributed Stochastic Neighbor Embedding for R (t-SNE)
-Version: 0.1-2
-Date: 2010-02-19
+Title: T-Distributed Stochastic Neighbor Embedding for R (t-SNE)
+Version: 0.1-3
+Date: 2016-06-04
 Author: Justin Donaldson <jdonaldson@gmail.com>
 Maintainer: Justin Donaldson <jdonaldson@gmail.com>
 Description: A "pure R" implementation of the t-SNE algorithm.
 License: GPL
 LazyLoad: yes
-Packaged: 2012-05-01 16:27:35 UTC; justindonaldson
+NeedsCompilation: no
+URL: https://github.com/jdonaldson/rtsne/
+BugReports: https://github.com/jdonaldson/rtsne/issues
+Packaged: 2016-07-15 15:40:42 UTC; jdonaldson
 Repository: CRAN
-Date/Publication: 2012-05-02 06:34:28
+Date/Publication: 2016-07-15 20:02:16
diff --git a/MD5 b/MD5
@@ -1,7 +1,7 @@
-c0b2fb9370262ac6c5a189ea2aab6606 *DESCRIPTION
-8b54e5a89fbda3af5e077053d40bec76 *NAMESPACE
-c62fd78c2dbd326aee4a0f43037d2449 *NEWS
+89ec2b9a3dc063fc9141420faf288338 *DESCRIPTION
+d65aaf552fc71192a426c3b41f3b0c48 *NAMESPACE
+a63efcd4c418f6199317015c874204de *NEWS
 5b83328297146a191456b5e2299c5123 *R/tsne-internal.R
-13e8ed8f8f943eaa5731f88b11ba8148 *R/tsne.R
-1bb74773e41979acecc6c87375bef147 *man/tsne-package.Rd
-42113a4da50526fa264e890b27861e90 *man/tsne.Rd
+a444f7bd180097f3bb7f3b1b61d77d43 *R/tsne.R
+bb49103a06187a42f001e2671f2072f7 *man/tsne-package.Rd
+b085bd2cdbbc2f88aab4172f4312be83 *man/tsne.Rd
diff --git a/NAMESPACE b/NAMESPACE
@@ -1 +1,2 @@
 exportPattern("^[[:alpha:]]+")
+importFrom("stats", "dist", "rnorm")
diff --git a/NEWS b/NEWS
@@ -1,3 +1,4 @@
 * version 0.1 - Initial release.  main tsne function included.
 * version 0.1-1 - fixed misc. documentation omissions.
-* version 0.1-2 - using an initial configuration parameter automatically places the tsne in the 'late stage' mode, which is mainly focused on small scale adjustments.
+* version 0.1-2 - using an initial configuration parameter automatically places the tsne in the 'late stage' mode, which is mainly focused on small scale adjustments.
+* version 0.1-3 - fix a broken gain equation, and made distance matrix checks more reliable
diff --git a/R/tsne.R b/R/tsne.R
@@ -1,6 +1,6 @@
 tsne <-
 function(X, initial_config = NULL, k=2, initial_dims=30, perplexity=30, max_iter = 1000, min_cost=0, epoch_callback=NULL,whiten=TRUE, epoch=100 ){
-	if (class(X) == 'dist') { 
+	if ('dist' %in% class(X)) {
 		n = attr(X,'Size')
 		}
 	else 	{
@@ -22,32 +22,28 @@ function(X, initial_config = NULL, k=2, initial_dims=30, perplexity=30, max_iter
 
 	eps = 2^(-52) # typical machine precision
 
-	if (!is.null(initial_config) && is.matrix(initial_config)) { 		
+	if (!is.null(initial_config) && is.matrix(initial_config)) {
 		if (nrow(initial_config) != n | ncol(initial_config) != k){
 			stop('initial_config argument does not match necessary configuration for X')
 		}
 		ydata = initial_config
 		initial_P_gain = 1
-		
+
 	} else {
 		ydata = matrix(rnorm(k * n),n)
 	}
-	
+
 	P = .x2p(X,perplexity, 1e-5)$P
-	# P[is.nan(P)]<-eps
 	P = .5 * (P + t(P))
 
 	P[P < eps]<-eps
 	P = P/sum(P)
-
 
-
 	P = P * initial_P_gain
 	grads =  matrix(0,nrow(ydata),ncol(ydata))
 	incs =  matrix(0,nrow(ydata),ncol(ydata))
 	gains = matrix(1,nrow(ydata),ncol(ydata))
 
-
 	for (iter in 1:max_iter){
 		if (iter %% epoch == 0) { # epoch
 			cost =  sum(apply(P * log((P+eps)/(Q+eps)),1,sum))
@@ -57,9 +53,8 @@ function(X, initial_config = NULL, k=2, initial_dims=30, perplexity=30, max_iter
 
 		}
 
-
 		sum_ydata = apply(ydata^2, 1, sum)
-		num =  1/(1 + sum_ydata +    sweep(-2 * ydata %*% t(ydata),2, -t(sum_ydata))) 
+		num =  1/(1 + sum_ydata +    sweep(-2 * ydata %*% t(ydata),2, -t(sum_ydata)))
 		diag(num)=0
 		Q = num / sum(num)
 		if (any(is.nan(num))) message ('NaN in grad. descent')
@@ -68,21 +63,19 @@ function(X, initial_config = NULL, k=2, initial_dims=30, perplexity=30, max_iter
 		for (i in 1:n){
 			grads[i,] = apply(sweep(-ydata, 2, -ydata[i,]) * stiffnesses[,i],2,sum)
 		}
-
-		gains = (gains + .2) * abs(sign(grads) != sign(incs)) 
-				+ gains * .8 * abs(sign(grads) == sign(incs))		
+
+		gains = ((gains + .2) * abs(sign(grads) != sign(incs)) +
+				 gains * .8 * abs(sign(grads) == sign(incs)))
+
 		gains[gains < min_gain] = min_gain
 
 		incs = momentum * incs - epsilon * (gains * grads)
 		ydata = ydata + incs
 		ydata = sweep(ydata,2,apply(ydata,2,mean))
 		if (iter == mom_switch_iter) momentum = final_momentum
-		
+
 		if (iter == 100 && is.null(initial_config)) P = P/4
-
 
-
-
 	}
 	ydata
 }

diff --git a/man/tsne-package.Rd b/man/tsne-package.Rd
@@ -21,7 +21,7 @@ LazyLoad: \tab yes\cr
 }
 \author{
 Justin Donaldson
-http://www.scwn.net
+https://github.com/jdonaldson/rtsne
 Maintainer: Justin Donaldson (jdonaldson@gmail.com)
 }
 \references{

diff --git a/man/tsne.Rd b/man/tsne.Rd
@@ -9,7 +9,9 @@ The t-SNE method for dimensionality reduction
 Provides a simple function interface for specifying t-SNE dimensionality reduction on R matrices or "dist" objects.
 }
 \usage{
-tsne(X, initial_config = NULL, k = 2, initial_dims = 30, perplexity = 30, max_iter = 1000, min_cost = 0, epoch_callback = NULL, whiten = TRUE, epoch=100)
+tsne(X, initial_config = NULL, k = 2, initial_dims = 30, perplexity = 30,
+     max_iter = 1000, min_cost = 0, epoch_callback = NULL, whiten = TRUE,
+     epoch=100)
 }
 %- maybe also 'usage' for other objects documented here.
 \arguments{
@@ -23,7 +25,7 @@ tsne(X, initial_config = NULL, k = 2, initial_dims = 30, perplexity = 30, max_it
 	the dimension of the resulting embedding.
 }
   \item{initial_dims}{
-	The number of dimensions to use in reduction method. 
+	The number of dimensions to use in reduction method.
 }
   \item{perplexity}{
 	Perplexity parameter.  (optimal number of neighbors)
@@ -55,9 +57,9 @@ tsne(X, initial_config = NULL, k = 2, initial_dims = 30, perplexity = 30, max_it
 
 }
 \references{
- L.J.P. van der Maaten and G.E. Hinton. Visualizing High-Dimensional Data Using t-SNE. \emph{Journal of Machine Learning Research} 9 (Nov) : 2579-2605, 2008. 
+ L.J.P. van der Maaten and G.E. Hinton. Visualizing High-Dimensional Data Using t-SNE. \emph{Journal of Machine Learning Research} 9 (Nov) : 2579-2605, 2008.
 
- L.J.P. van der Maaten. Learning a Parametric Embedding by Preserving Local Structure. In \emph{Proceedings of the Twelfth International Conference on Artificial Intelligence and Statistics} (AISTATS), JMLR W&CP 5:384-391, 2009. 
+ L.J.P. van der Maaten. Learning a Parametric Embedding by Preserving Local Structure. In \emph{Proceedings of the Twelfth International Conference on Artificial Intelligence and Statistics} (AISTATS), JMLR W&CP 5:384-391, 2009.
 }
 \author{
 Justin Donaldson (jdonaldson@gmail.com)