@@ -423,6 +407,7 @@
$("#ExecuteBut").click(function(){
var mode = document.getElementById('param-EX-view').value
mode = parseInt(mode)
+
if (mode == 1) {
$("#myModal").modal('show');
$('.modal-backdrop').removeClass("modal-backdrop");
@@ -434,7 +419,7 @@
$('#myModal').modal('hide');
}
- $("#cost").html('(Unknown Iteration and Cost Values)');
+ $("#cost").html('(Unknown Overall Cost)');
$("#datasetDetails").html('(Unknown Number of Dimensions and Instances)');
$("#CategoryName").html('No Classification');
$("#knnBarChartDetails").html('(Number of Selected Points: 0/0)');
diff --git a/js/tsne_vis.js b/js/tsne_vis.js
index 46dd900..8cb7709 100755
--- a/js/tsne_vis.js
+++ b/js/tsne_vis.js
@@ -19,7 +19,7 @@ var dim = document.getElementById('overviewRect').offsetWidth-2; var dimensions
// Category = the name of the category if it exists. The user has to add an asterisk ("*") mark in order to let the program identify this feature as a label/category name.
// ColorsCategorical = the categorical colors (maximum value = 10).
-var Category; var ColorsCategorical; var valCategExists = 0;
+var Category; var ColorsCategorical; var valCategExists = 0; var insideCall = false;
// This is for the removal of the distances cache.
var returnVal = false;
@@ -34,7 +34,7 @@ var sliderTrigger = false; var sliderInsideTrigger = false; var parameters; var
var inside = 0; var kValuesLegend = []; var findNearestTable = []; var howManyPoints;
var maxKNN = 0
-var mode = 1; var colors = ['#a6cee3','#fb9a99','#b2df8a','#33a02c','#1f78b4','#e31a1c','#fdbf6f','#ff7f00','#cab2d6','#6a3d9a']; var projections = []; var metricsSorting = []; var dataReceivedFromServer = []; var dataReceivedFromServerOptimized = []; var metrics = []; var FocusedIDs = [];
+var mode = 1; var colors = ['#a6cee3','#fb9a99','#b2df8a','#33a02c','#1f78b4','#e31a1c','#fdbf6f','#ff7f00','#cab2d6','#6a3d9a']; var projections = []; var betas = []; var cost_per_point = []; var cost_overall; var metricsSorting = []; var dataReceivedFromServer = []; var dataReceivedFromServerOptimized = []; var metrics = []; var FocusedIDs = [];
var Category; var target_names = []
@@ -140,6 +140,9 @@ function ReSort(flagInitialize) {
projections = dataReceivedFromServer['projections']
parameters = dataReceivedFromServer['parameters']
metricsSorting = dataReceivedFromServer['metrics']
+ betas = dataReceivedFromServer['betas']
+ cost_per_point = dataReceivedFromServer['cpp']
+ cost_overall = dataReceivedFromServer['cpi']
if (FocusedIDs.length != 0) {
if (globalFlagCheck) {
@@ -147,11 +150,12 @@ function ReSort(flagInitialize) {
metricsCopy = metrics
}
globalFlagCheck = false
- document.getElementById("textToChange").innerHTML = "[Sorting Projections According to Metric for the Current Points Selection:";
metricsSorting = dataReceivedFromServerOptimized['metrics']
metrics = dataReceivedFromServerOptimized['metricsEntire']
if (FocusedIDs.length == points.length) {
- document.getElementById("textToChange").innerHTML = "[Sorting Metric:";
+ document.getElementById("textToChange").innerHTML = "[Sorting Projections According to Metric:";
+ } else {
+ document.getElementById("textToChange").innerHTML = "[Sorting Projections According to Metric for Current Selection:";
}
} else {
if (!globalFlagCheck) {
@@ -669,6 +673,7 @@ if(k >= 6) {
Plotly.relayout(graphDiv, update)
sliderInsideTrigger = true
+ insideCall = true;
activeProjectionNumberProv = order[SelProjIDSProv[0]]
getData()
@@ -676,6 +681,7 @@ if(k >= 6) {
if (flagInitialize) {
closeModalFun()
+ insideCall = true;
getData()
}
@@ -804,19 +810,68 @@ function parseData(url) {
// results.data variable is all the columns except strings, undefined values, or "Version" plus beta and cost values."
// results.meta.fields variable is all the features (columns) plus beta and cost strings.
if (mode == 2) {
- init(results.data, results_all, results.meta.fields); // Call the init() function that starts everything!
+ if (insideCall) {
+ init(results.data, results_all, results.meta.fields); // Call the init() function that starts everything!
+ } else {
+ $.post("http://127.0.0.1:5000/resetAll", JSON.stringify(''), function(){
+ });
+ var parametersLocal = []
+ parametersLocal.push(document.getElementById("param-perplexity-value").value)
+ parametersLocal.push(document.getElementById("param-learningrate-value").value)
+ parametersLocal.push(document.getElementById("param-maxiter-value").value)
+ parametersLocal.push(results_all)
+ $.post("http://127.0.0.1:5000/receiverSingle", JSON.stringify(parametersLocal), function(){
+ $.get("http://127.0.0.1:5000/senderSingle", function( data ) {
+ dataReceivedFromServer = data
+ projections = dataReceivedFromServer['projections']
+ betas = dataReceivedFromServer['betas']
+ cost_per_point = dataReceivedFromServer['cpp']
+ cost_overall = dataReceivedFromServer['cpi']
+ activeProjectionNumber = 0
+ activeProjectionNumberProv = 0
+ init(results.data, results_all, results.meta.fields); // Call the init() function that starts everything!
+ });
+
+ });
+ }
} else {
- // ajax the JSON to the server
- $.post("http://127.0.0.1:5000/resetAll", JSON.stringify(''), function(){
- });
- $.post("http://127.0.0.1:5000/receiver", JSON.stringify(results_all), function(){
- $.get("http://127.0.0.1:5000/sender", function( data ) {
- dataReceivedFromServer = data
- ReSortOver()
-
- });
-
- });
+ if (flagAnalysis) {
+
+ $('#myModal').modal('hide');
+ $.post("http://127.0.0.1:5000/resetAll", JSON.stringify(''), function(){
+ });
+
+ var parametersLocal = []
+ parametersLocal.push(document.getElementById("param-perplexity-value").value)
+ parametersLocal.push(document.getElementById("param-learningrate-value").value)
+ parametersLocal.push(document.getElementById("param-maxiter-value").value)
+ parametersLocal.push(results_all)
+ $.post("http://127.0.0.1:5000/receiverSingle", JSON.stringify(parametersLocal), function(){
+ $.get("http://127.0.0.1:5000/senderSingle", function( data ) {
+ dataReceivedFromServer = data
+ projections = dataReceivedFromServer['projections']
+ betas = dataReceivedFromServer['betas']
+ cost_per_point = dataReceivedFromServer['cpp']
+ cost_overall = dataReceivedFromServer['cpi']
+ activeProjectionNumber = 0
+ init(results.data, results_all, results.meta.fields); // Call the init() function that starts everything!
+ });
+
+ });
+ } else {
+ // ajax the JSON to the server
+ $.post("http://127.0.0.1:5000/resetAll", JSON.stringify(''), function(){
+ });
+ $.post("http://127.0.0.1:5000/receiver", JSON.stringify(results_all), function(){
+ $.get("http://127.0.0.1:5000/sender", function( data ) {
+ dataReceivedFromServer = data
+ ReSortOver()
+
+ });
+
+ });
+ }
+
}
}
@@ -834,6 +889,9 @@ function ReSortOver() {
parameters = dataReceivedFromServer['parameters']
metricsSorting = dataReceivedFromServer['metrics']
metrics = dataReceivedFromServer['metricsEntire']
+ betas = dataReceivedFromServer['betas']
+ cost_per_point = dataReceivedFromServer['cpp']
+ cost_overall = dataReceivedFromServer['cpi']
var traces = []
var target_names = []
@@ -2629,10 +2687,10 @@ function init(data, results_all, fields) {
} else{
tsne = new tsnejs.tSNE(opt); // Set new t-SNE with specific perplexity.
dists = [];
- dists = computeDistances(data, document.getElementById("param-distance").value, document.getElementById("param-transform").value); // Compute the distances in the high-dimensional space.
+ dists = computeDistances(data, 'euclideanDist', 'noTrans'); // Compute the distances in the high-dimensional space.
InitialFormDists.push(dists);
- tsne.initDataDist(dists); // Init t-SNE with dists.
- for(var i = 0; i < final_dataset.length; i++) {final_dataset[i].beta = tsne.beta[i]; beta_all[i] = tsne.beta[i];} // Calculate beta and bring it back from the t-SNE algorithm.
+ //tsne.initDataDist(dists); // Init t-SNE with dists.
+ //for(var i = 0; i < final_dataset.length; i++) {final_dataset[i].beta = tsne.beta[i]; beta_all[i] = tsne.beta[i];} // Calculate beta and bring it back from the t-SNE algorithm.
}
var object;
all_labels = [];
@@ -2748,28 +2806,6 @@ function euclideanDist(data) {
}
-// Calculate jaccard dist
-function jaccardDist(data) {
-
- dist = initDist(data);
- for(var i = 0; i < data.length; i++) {
- for(var j = i + 1; j < data.length; j++) {
- for(var d in data[0]) {
- if(d != Category) {
- x = data[i][d];
- y = data[j][d];
- if(x == y) {
- dist[i][j] += 1;
- }
- }
- }
- dist[j][i] = dist[i][j];
- }
- }
- return dist;
-
-}
-
// Normalize distances to prevent numerical issues.
function normDist(data, dist) {
@@ -2793,51 +2829,6 @@ function noTrans(data) {
return data;
}
-// Log tranformation
-function logTrans(data) {
-
- for(var i = 0; i < data.length; i++) {
- for(var d in data[0]) {
- if(d != Category) {
- X = data[i][d];
- data[i][d] = Math.log10(X + 1);
- }
- }
- }
- return data;
-
-}
-
-// asinh tranformation
-function asinhTrans(data) {
-
- for(var i = 0; i < data.length; i++) {
- for(var d in data[0]) {
- if(d != Category) {
- X = data[i][d];
- data[i][d] = Math.log(X + Math.sqrt(X * X + 1));
- }
- }
- }
- return data;
-
-}
-// Binarize tranformation
-function binTrans(data) {
-
- for(var i = 0; i < data.length; i++) {
- for(var d in data[0]) {
- if(d != Category) {
- X = data[i][d];
- if(X > 0) data[i][d] = 1;
- if(X < 0) data[i][d] = 0;
- }
- }
- }
- return data;
-
-}
-
// Compute the distances by applying the chosen distance functions and transformation functions.
function computeDistances(data, distFunc, transFunc) {
dist = eval(distFunc)(eval(transFunc)(data));
@@ -2884,7 +2875,7 @@ function OverallCostLineChart(){
margin: {
l: 40,
r: 15,
- b: 26,
+ b: 30,
t: 5
},
};
@@ -2899,7 +2890,15 @@ function updateEmbedding(AnalysisResults) {
points = [];
points2d = [];
if (AnalysisResults == ""){ // Check if the embedding does not need to load because we had a previous analysis uploaded.
- var Y = tsne.getSolution(); // We receive the solution from the t-SNE
+ if (sliderTrigger) {
+ if (sliderInsideTrigger) {
+ var Y = projections[activeProjectionNumberProv]
+ } else {
+ var Y = projections[activeProjectionNumber]
+ }
+ } else {
+ var Y = projections[activeProjectionNumber]
+ }
var xExt = d3.extent(Y, d => d[0]);
var yExt = d3.extent(Y, d => d[1]);
var maxExt = [Math.min(xExt[0], yExt[0]), Math.max(xExt[1], yExt[1])];
@@ -2911,10 +2910,11 @@ function updateEmbedding(AnalysisResults) {
var y = d3.scaleLinear() // Scale the y points into the canvas width/height
.domain(maxExt)
.range([10, +dimensionsY-10]);
- for(var i = 0; i < final_dataset.length; i++) {
+
+ for(var i = 0; i < betas[activeProjectionNumber].length; i++) {
x_position[i] = x(Y[i][0]); // x points position
y_position[i] = y(Y[i][1]); // y points position
- points[i] = {id: i, x: x_position[i], y: y_position[i], beta: final_dataset[i].beta, cost: final_dataset[i].cost, selected: true, schemaInv: false, DimON: null, pcp: false}; // Create the points and points2D (2 dimensions)
+ points[i] = {id: i, x: x_position[i], y: y_position[i], beta: betas[activeProjectionNumber][i], cost: cost_per_point[activeProjectionNumber][i], selected: true, schemaInv: false, DimON: null, pcp: false}; // Create the points and points2D (2 dimensions)
points2d[i] = {x: x_position[i], y: y_position[i]}; // and add everything that we know about the points (e.g., selected = true, pcp = false in the beginning and so on)
points[i] = extend(points[i], ArrayContainsDataFeaturesCleared[i]);
points[i] = extend(points[i], dataFeatures[i]);
@@ -2934,12 +2934,10 @@ function updateEmbedding(AnalysisResults) {
ArrayWithCostsList = AnalysisResults.slice(2*dataFeatures.length+length+10, 2*dataFeatures.length+length+11);
Iterations = IterationsList[0];
ArrayWithCosts = ArrayWithCostsList[0];
- $("#cost").html("(Number. of Iter.: " + ParametersSet[3] + ", Ov. Cost: " + overallCost + ")");
+ $("#cost").html("(Overall Cost: " + overallCost + ")");
$('#param-perplexity-value').text(ParametersSet[1]);
$('#param-learningrate-value').text(ParametersSet[2]);
$('#param-maxiter-value').text(ParametersSet[3]);
- document.getElementById("param-distance").value = ParametersSet[4];
- document.getElementById("param-transform").value = ParametersSet[5];
} else{
var length = (AnalysisResults.length - 9) / 2;
points = AnalysisResults.slice(0, length); // Load the points from the previous analysis
@@ -2950,12 +2948,10 @@ function updateEmbedding(AnalysisResults) {
ArrayWithCostsList = AnalysisResults.slice(2*length+8, 2*length+9);
Iterations = IterationsList[0];
ArrayWithCosts = ArrayWithCostsList[0];
- $("#cost").html("(Number of Iter.: " + ParametersSet[3] + ", Ov. Cost: " + overallCost + ")");
+ $("#cost").html("(Overall Cost: " + overallCost + ")");
$('#param-perplexity-value').text(ParametersSet[1]);
$('#param-learningrate-value').text(ParametersSet[2]);
$('#param-maxiter-value').text(ParametersSet[3]);
- document.getElementById("param-distance").value = ParametersSet[4];
- document.getElementById("param-transform").value = ParametersSet[5];
}
$("#data").html(ParametersSet[0]); // Print on the screen the classification label.
$("#param-dataset").html('-');
@@ -3014,7 +3010,7 @@ function ShepardHeatMap () {
dist_list2d = []; // Distances lists empty
dist_list = [];
// Calculate the 2D distances.
- dists2d = computeDistances(points2d, document.getElementById("param-distance").value, document.getElementById("param-transform").value);
+ dists2d = computeDistances(points2d, 'euclideanDist', 'noTrans');
InitialFormDists2D.push(dists2d);
for (var j=0; j
points[i].cost){
- minSize1 = points[i].cost;
+ minSize1 = points[i].cost;
}
}
-
+
var rscale1 = d3.scaleLinear()
.domain([minSize1, maxSize1])
.range([5,parseInt(12-(1-document.getElementById("param-costlim").value)*7)]);
@@ -5233,7 +5264,12 @@ if (points.length) { // If points exist (at least 1 point)
.attr("transform", "translate(10,20)");
var SizeRange1 = [];
- SizeRange1.push((minSize1).toFixed(4));
+
+ if (minSize1 < 0) {
+ SizeRange1.push(0.0000)
+ } else {
+ SizeRange1.push((minSize1).toFixed(4));
+ }
SizeRange1.push(((maxSize1-minSize1)/2).toFixed(4));
SizeRange1.push((maxSize1).toFixed(4));
@@ -5267,7 +5303,11 @@ if (points.length) { // If points exist (at least 1 point)
var min = points[temp].cost;
for (var i=temp+1; i points[i].cost){
- min = points[i].cost;
+ if (points[i].cost < 0) {
+ min = 0;
+ } else {
+ min = points[i].cost;
+ }
}
}
@@ -5294,7 +5334,7 @@ if (points.length) { // If points exist (at least 1 point)
labels_cost = d3.range(min, max+calcStep, calcStep);
for (var i=0; i<9; i++){
labels_cost[i] = labels_cost[i].toFixed(5);
- abbr_labels_cost[i] = abbreviateNumber(labels_cost[i]);
+ abbr_labels_cost[i] = Math.abs(abbreviateNumber(labels_cost[i]));
}
var svg = d3.select("#legend1"); // Add the legend for the beta/cost
@@ -5575,8 +5615,8 @@ if (points.length) { // If points exist (at least 1 point)
var abbr_labels_cost = [];
labels_cost = d3.range(min, max+calcStep, calcStep);
for (var i=0; i<7; i++){
- labels_cost[i] = labels_cost[i].toFixed(5);
- abbr_labels_cost[i] = abbreviateNumber(labels_cost[i]);
+ labels_cost[i] = labels_cost[i].toFixed(4);
+ abbr_labels_cost[i] = Math.abs(abbreviateNumber(labels_cost[i]));
}
var svg = d3.select("#legend1"); // Add the legend for the beta/cost
@@ -5588,7 +5628,7 @@ if (points.length) { // If points exist (at least 1 point)
.labelFormat(d3.format(",.5f"))
.cells(7)
.labels([abbr_labels_cost[0],abbr_labels_cost[1],abbr_labels_cost[2],abbr_labels_cost[3],abbr_labels_cost[4],abbr_labels_cost[5],abbr_labels_cost[6]])
- .title("Remaining Cost")
+ .title("Rem. Cost")
.scale(colorScale);
svg.select(".legendLinear")
@@ -6027,8 +6067,8 @@ function SaveAnalysis(){ // Save the analysis into a .txt file
let perplexity = document.getElementById("param-perplexity-value").value;
let learningRate = document.getElementById("param-learningrate-value").value;
let IterValue = document.getElementById("param-maxiter-value").value;
- let parDist = document.getElementById("param-distance").value;
- let parTrans = document.getElementById("param-transform").value;
+ let parDist = 'euclideanDist';
+ let parTrans = 'noTrans';
let Parameters = [];
if (dataset == "empty"){
Parameters.push(new_file.name);
diff --git a/sptree.cpp b/sptree.cpp
old mode 100644
new mode 100755
diff --git a/sptree.h b/sptree.h
old mode 100644
new mode 100755
diff --git a/tsne.cpp b/tsne.cpp
old mode 100644
new mode 100755
index 77812ba..5420b33
--- a/tsne.cpp
+++ b/tsne.cpp
@@ -43,9 +43,23 @@
using namespace std;
+static double sign(double x) { return (x == .0 ? .0 : (x < .0 ? -1.0 : 1.0)); }
+
+static void zeroMean(double* X, int N, int D);
+static void computeGaussianPerplexity(double* X, int N, int D, double* P, double perplexity);
+static void computeGaussianPerplexity(double* X, int N, int D, unsigned int** _row_P, unsigned int** _col_P, double** _val_P, double perplexity, int K);
+static double randn();
+static void computeExactGradient(double* P, double* Y, int N, int D, double* dC);
+static void computeGradient(unsigned int* inp_row_P, unsigned int* inp_col_P, double* inp_val_P, double* Y, int N, int D, double* dC, double theta);
+static double evaluateError(double* P, double* Y, int N, int D);
+static double evaluateError(unsigned int* row_P, unsigned int* col_P, double* val_P, double* Y, int N, int D, double theta, bool save_cost_per_point = false);
+static void computeSquaredEuclideanDistance(double* X, int N, int D, double* DD);
+static void symmetrizeMatrix(unsigned int** row_P, unsigned int** col_P, double** val_P, int N);
+static void save_1D_array(double* array, int n, const char* filename);
+
// Perform t-SNE
void TSNE::run(double* X, int N, int D, double* Y, int no_dims, double perplexity, double theta, int rand_seed,
- bool skip_random_init, int max_iter, int stop_lying_iter, int mom_switch_iter) {
+ bool skip_random_init, int max_iter, int stop_lying_iter, int mom_switch_iter, int cost_iter_step) {
// Set random seed
if (skip_random_init != true) {
@@ -133,20 +147,22 @@ void TSNE::run(double* X, int N, int D, double* Y, int no_dims, double perplexit
else { for(int i = 0; i < row_P[N]; i++) val_P[i] *= 12.0; }
// Initialize solution (randomly)
- if (skip_random_init != true) {
- for(int i = 0; i < N * no_dims; i++) Y[i] = randn() * .0001;
- }
+ if (skip_random_init != true) {
+ for(int i = 0; i < N * no_dims; i++) Y[i] = randn() * .0001;
+ }
// Perform main training loop
if(exact) printf("Input similarities computed in %4.2f seconds!\nLearning embedding...\n", (float) (end - start) / CLOCKS_PER_SEC);
else printf("Input similarities computed in %4.2f seconds (sparsity = %f)!\nLearning embedding...\n", (float) (end - start) / CLOCKS_PER_SEC, (double) row_P[N] / ((double) N * (double) N));
start = clock();
+ double* cost_per_iter = new double[max_iter];
+
for(int iter = 0; iter < max_iter; iter++) {
// Compute (approximate) gradient
if(exact) computeExactGradient(P, Y, N, no_dims, dY);
- else computeGradient(P, row_P, col_P, val_P, Y, N, no_dims, dY, theta);
+ else computeGradient(row_P, col_P, val_P, Y, N, no_dims, dY, theta);
// Update gains
for(int i = 0; i < N * no_dims; i++) gains[i] = (sign(dY[i]) != sign(uY[i])) ? (gains[i] + .2) : (gains[i] * .8);
@@ -167,22 +183,29 @@ void TSNE::run(double* X, int N, int D, double* Y, int no_dims, double perplexit
if(iter == mom_switch_iter) momentum = final_momentum;
// Print out progress
- if (iter > 0 && (iter % 50 == 0 || iter == max_iter - 1)) {
+ if (iter % cost_iter_step == 0 || iter == max_iter - 1) {
end = clock();
double C = .0;
if(exact) C = evaluateError(P, Y, N, no_dims);
else C = evaluateError(row_P, col_P, val_P, Y, N, no_dims, theta); // doing approximate computation here!
if(iter == 0)
- printf("Iteration %d: error is %f\n", iter + 1, C);
+ printf("Iteration %d: error is %f\n", iter, C);
else {
total_time += (float) (end - start) / CLOCKS_PER_SEC;
- printf("Iteration %d: error is %f (50 iterations in %4.2f seconds)\n", iter, C, (float) (end - start) / CLOCKS_PER_SEC);
+ printf("Iteration %d: error is %f (%d iterations in %4.2f seconds)\n", iter, C, cost_iter_step, (float) (end - start) / CLOCKS_PER_SEC);
}
+ cost_per_iter[iter] = C;
start = clock();
}
}
end = clock(); total_time += (float) (end - start) / CLOCKS_PER_SEC;
+ // Save the cost per iteration to file
+ save_1D_array(cost_per_iter, max_iter, "cost_per_iter.txt");
+
+ // This time it's for saving the final cost per point
+ evaluateError(row_P, col_P, val_P, Y, N, no_dims, theta, true);
+
// Clean up memory
free(dY);
free(uY);
@@ -193,12 +216,13 @@ void TSNE::run(double* X, int N, int D, double* Y, int no_dims, double perplexit
free(col_P); col_P = NULL;
free(val_P); val_P = NULL;
}
+ delete[] cost_per_iter;
printf("Fitting performed in %4.2f seconds.\n", total_time);
}
// Compute gradient of the t-SNE cost function (using Barnes-Hut algorithm)
-void TSNE::computeGradient(double* P, unsigned int* inp_row_P, unsigned int* inp_col_P, double* inp_val_P, double* Y, int N, int D, double* dC, double theta)
+static void computeGradient(unsigned int* inp_row_P, unsigned int* inp_col_P, double* inp_val_P, double* Y, int N, int D, double* dC, double theta)
{
// Construct space-partitioning tree on current map
@@ -222,7 +246,7 @@ void TSNE::computeGradient(double* P, unsigned int* inp_row_P, unsigned int* inp
}
// Compute gradient of the t-SNE cost function (exact)
-void TSNE::computeExactGradient(double* P, double* Y, int N, int D, double* dC) {
+static void computeExactGradient(double* P, double* Y, int N, int D, double* dC) {
// Make sure the current gradient contains zeros
for(int i = 0; i < N * D; i++) dC[i] = 0.0;
@@ -272,7 +296,7 @@ void TSNE::computeExactGradient(double* P, double* Y, int N, int D, double* dC)
// Evaluate t-SNE cost function (exactly)
-double TSNE::evaluateError(double* P, double* Y, int N, int D) {
+static double evaluateError(double* P, double* Y, int N, int D) {
// Compute the squared Euclidean distance matrix
double* DD = (double*) malloc(N * N * sizeof(double));
@@ -308,7 +332,7 @@ double TSNE::evaluateError(double* P, double* Y, int N, int D) {
}
// Evaluate t-SNE cost function (approximately)
-double TSNE::evaluateError(unsigned int* row_P, unsigned int* col_P, double* val_P, double* Y, int N, int D, double theta)
+static double evaluateError(unsigned int* row_P, unsigned int* col_P, double* val_P, double* Y, int N, int D, double theta, bool save_cost_per_point)
{
// Get estimate of normalization term
@@ -316,11 +340,13 @@ double TSNE::evaluateError(unsigned int* row_P, unsigned int* col_P, double* val
double* buff = (double*) calloc(D, sizeof(double));
double sum_Q = .0;
for(int n = 0; n < N; n++) tree->computeNonEdgeForces(n, theta, buff, &sum_Q);
+ double* cost_per_point = new double[N];
// Loop over all edges to compute t-SNE error
int ind1, ind2;
- double C = .0, Q;
+ double C = .0, Q, c;
for(int n = 0; n < N; n++) {
+ cost_per_point[n] = .0;
ind1 = n * D;
for(int i = row_P[n]; i < row_P[n + 1]; i++) {
Q = .0;
@@ -329,19 +355,25 @@ double TSNE::evaluateError(unsigned int* row_P, unsigned int* col_P, double* val
for(int d = 0; d < D; d++) buff[d] -= Y[ind2 + d];
for(int d = 0; d < D; d++) Q += buff[d] * buff[d];
Q = (1.0 / (1.0 + Q)) / sum_Q;
- C += val_P[i] * log((val_P[i] + FLT_MIN) / (Q + FLT_MIN));
+ c = val_P[i] * log((val_P[i] + FLT_MIN) / (Q + FLT_MIN));
+ cost_per_point[n] += c;
+ C += c;
}
}
+ if (save_cost_per_point)
+ save_1D_array(cost_per_point, N, "cost_per_point.txt");
+
// Clean up memory
free(buff);
delete tree;
+ delete[] cost_per_point;
return C;
}
// Compute input similarities with a fixed perplexity
-void TSNE::computeGaussianPerplexity(double* X, int N, int D, double* P, double perplexity) {
+static void computeGaussianPerplexity(double* X, int N, int D, double* P, double perplexity) {
// Compute the squared Euclidean distance matrix
double* DD = (double*) malloc(N * N * sizeof(double));
@@ -412,7 +444,7 @@ void TSNE::computeGaussianPerplexity(double* X, int N, int D, double* P, double
// Compute input similarities with a fixed perplexity using ball trees (this function allocates memory another function should free)
-void TSNE::computeGaussianPerplexity(double* X, int N, int D, unsigned int** _row_P, unsigned int** _col_P, double** _val_P, double perplexity, int K) {
+static void computeGaussianPerplexity(double* X, int N, int D, unsigned int** _row_P, unsigned int** _col_P, double** _val_P, double perplexity, int K) {
if(perplexity > K) printf("Perplexity should be lower than K!\n");
@@ -439,6 +471,7 @@ void TSNE::computeGaussianPerplexity(double* X, int N, int D, unsigned int** _ro
printf("Building tree...\n");
vector indices;
vector distances;
+ double* betas = new double[N];
for(int n = 0; n < N; n++) {
if(n % 10000 == 0) printf(" - point %d of %d\n", n, N);
@@ -495,6 +528,9 @@ void TSNE::computeGaussianPerplexity(double* X, int N, int D, unsigned int** _ro
iter++;
}
+ // Save the found beta to a vector of all betas
+ betas[n] = beta;
+
// Row-normalize current row of P and store in matrix
for(unsigned int m = 0; m < K; m++) cur_P[m] /= sum_P;
for(unsigned int m = 0; m < K; m++) {
@@ -503,15 +539,19 @@ void TSNE::computeGaussianPerplexity(double* X, int N, int D, unsigned int** _ro
}
}
+ // Save the betas to disk, so we can recover them later
+ save_1D_array(betas, N, "betas.txt");
+
// Clean up memory
obj_X.clear();
free(cur_P);
delete tree;
+ delete[] betas;
}
// Symmetrizes a sparse matrix
-void TSNE::symmetrizeMatrix(unsigned int** _row_P, unsigned int** _col_P, double** _val_P, int N) {
+static void symmetrizeMatrix(unsigned int** _row_P, unsigned int** _col_P, double** _val_P, int N) {
// Get sparse matrix
unsigned int* row_P = *_row_P;
@@ -599,7 +639,7 @@ void TSNE::symmetrizeMatrix(unsigned int** _row_P, unsigned int** _col_P, double
}
// Compute squared Euclidean distance matrix
-void TSNE::computeSquaredEuclideanDistance(double* X, int N, int D, double* DD) {
+static void computeSquaredEuclideanDistance(double* X, int N, int D, double* DD) {
const double* XnD = X;
for(int n = 0; n < N; ++n, XnD += D) {
const double* XmD = XnD + D;
@@ -618,7 +658,7 @@ void TSNE::computeSquaredEuclideanDistance(double* X, int N, int D, double* DD)
// Makes data zero-mean
-void TSNE::zeroMean(double* X, int N, int D) {
+static void zeroMean(double* X, int N, int D) {
// Compute data mean
double* mean = (double*) calloc(D, sizeof(double));
@@ -647,7 +687,7 @@ void TSNE::zeroMean(double* X, int N, int D) {
// Generates a Gaussian random number
-double TSNE::randn() {
+static double randn() {
double x, y, radius;
do {
x = 2 * (rand() / ((double) RAND_MAX + 1)) - 1;
@@ -702,3 +742,19 @@ void TSNE::save_data(double* data, int* landmarks, double* costs, int n, int d)
fclose(h);
printf("Wrote the %i x %i data matrix successfully!\n", n, d);
}
+
+// Function that saves a 1D array in the numpy.savetxt format
+void save_1D_array(double* array, int n, const char* filename) {
+
+ // Open file, write first 2 integers and then the data
+ FILE *h;
+ if((h = fopen(filename, "w")) == NULL) {
+ printf("Error: could not open [%s] for writing.\n", filename);
+ return;
+ }
+ //fprintf(h, "%d\n", n);
+ for (int i = 0; i < n; ++i)
+ fprintf(h, "%f\n", array[i]);
+ fclose(h);
+ printf("Wrote [%s] successfully, with %i rows!\n", filename, n);
+}
diff --git a/tsne.h b/tsne.h
old mode 100644
new mode 100755
index 2600855..7e3ae0a
--- a/tsne.h
+++ b/tsne.h
@@ -34,30 +34,18 @@
#ifndef TSNE_H
#define TSNE_H
-
-static inline double sign(double x) { return (x == .0 ? .0 : (x < .0 ? -1.0 : 1.0)); }
-
-
-class TSNE
-{
-public:
+#ifdef __cplusplus
+extern "C" {
+namespace TSNE {
+#endif
void run(double* X, int N, int D, double* Y, int no_dims, double perplexity, double theta, int rand_seed,
- bool skip_random_init, int max_iter=1000, int stop_lying_iter=250, int mom_switch_iter=250);
+ bool skip_random_init, int max_iter, int stop_lying_iter, int mom_switch_iter, int cost_iter_step);
bool load_data(double** data, int* n, int* d, int* no_dims, double* theta, double* perplexity, int* rand_seed, int* max_iter);
void save_data(double* data, int* landmarks, double* costs, int n, int d);
- void symmetrizeMatrix(unsigned int** row_P, unsigned int** col_P, double** val_P, int N); // should be static!
-
-
-private:
- void computeGradient(double* P, unsigned int* inp_row_P, unsigned int* inp_col_P, double* inp_val_P, double* Y, int N, int D, double* dC, double theta);
- void computeExactGradient(double* P, double* Y, int N, int D, double* dC);
- double evaluateError(double* P, double* Y, int N, int D);
- double evaluateError(unsigned int* row_P, unsigned int* col_P, double* val_P, double* Y, int N, int D, double theta);
- void zeroMean(double* X, int N, int D);
- void computeGaussianPerplexity(double* X, int N, int D, double* P, double perplexity);
- void computeGaussianPerplexity(double* X, int N, int D, unsigned int** _row_P, unsigned int** _col_P, double** _val_P, double perplexity, int K);
- void computeSquaredEuclideanDistance(double* X, int N, int D, double* DD);
- double randn();
+ //void save_1D_array(double* data, int N, char* filename);
+#ifdef __cplusplus
};
+}
+#endif
#endif
diff --git a/tsneGrid.py b/tsneGrid.py
index 64ea4c8..704466c 100644
--- a/tsneGrid.py
+++ b/tsneGrid.py
@@ -42,12 +42,30 @@ def Reset():
global projectionsAll
projectionsAll = []
+ global betas
+ betas = []
+
+ global cpp
+ cpp = []
+
+ global cpi
+ cpi = []
+
global SelectedListofParams
SelectedListofParams = []
global SelectedProjectionsReturn
SelectedProjectionsReturn = []
+ global SelectedProjectionsBeta
+ SelectedProjectionsBeta = []
+
+ global SelectedProjectionsCPP
+ SelectedProjectionsCPP = []
+
+ global SelectedProjectionsCPI
+ SelectedProjectionsCPI = []
+
global clusterIndex
clusterIndex = []
@@ -160,8 +178,9 @@ def preprocess(data):
return dataNP, length, gatherLabels
def multi_run_wrapper(args):
- embedding_array = bhtsne.run_bh_tsne(*args)
- return embedding_array
+ projectionsAllLoc, betasL, cppL, cpiL = bhtsne.run_bh_tsne(*args)
+
+ return projectionsAllLoc, betasL, cppL, cpiL
def procrustesFun(projections):
similarityList = []
@@ -198,7 +217,7 @@ memory = Memory(location, verbose=0)
def wrapGetResults(listofParamsPlusData):
pool = Pool()
- return pool.map(multi_run_wrapper, listofParamsPlusData)
+ return zip(*pool.map(multi_run_wrapper, listofParamsPlusData))
wrapGetResults = memory.cache(wrapGetResults)
@@ -213,12 +232,9 @@ def calculateGrid():
D_highSpace = distance.squareform(distance.pdist(dataProc))
DEFAULT_NO_DIMS = 2
- INITIAL_DIMENSIONS = 50
- DEFAULT_PERPLEXITY = 50
- DEFAULT_THETA = 0.5
- EMPTY_SEED = -1
- VERBOSE = True
- DEFAULT_USE_PCA = False
+ VERBOSE = False
+ DEFAULT_USE_PCA = True
+ randseed=1137
# all other data sets
perplexity = [5,10,15,20,25,30,35,40,45,50] # 10 perplexity
@@ -249,15 +265,25 @@ def calculateGrid():
for k in n_iter:
for j in learning_rate:
for i in perplexity:
- listofParamsPlusData.append((dataProc,DEFAULT_NO_DIMS,length,i,j,EMPTY_SEED,VERBOSE,DEFAULT_USE_PCA,k))
+ listofParamsPlusData.append((dataProc,DEFAULT_NO_DIMS,i,j,randseed,VERBOSE,length,DEFAULT_USE_PCA,k,True,True,True))
listofParamsAll.append((i,j,k))
- projectionsAll = wrapGetResults(listofParamsPlusData)
-
+ projectionsAll, betas, cpp, cpi = wrapGetResults(listofParamsPlusData)
+
global SelectedListofParams
SelectedListofParams = []
+
global SelectedProjectionsReturn
SelectedProjectionsReturn = []
+
+ global SelectedProjectionsBeta
+ SelectedProjectionsBeta = []
+
+ global SelectedProjectionsCPP
+ SelectedProjectionsCPP = []
+
+ global SelectedProjectionsCPI
+ SelectedProjectionsCPI = []
global clusterIndex
clusterIndex = procrustesFun(projectionsAll)
@@ -295,10 +321,18 @@ def calculateGrid():
global KeepKs
KeepKs = []
+
+
for index in clusterIndex:
SelectedProjectionsReturn.append(projectionsAll[index].tolist())
SelectedListofParams.append(listofParamsAll[index])
+ SelectedProjectionsBeta.append(betas[index].tolist())
+
+ SelectedProjectionsCPP.append(cpp[index].tolist())
+
+ SelectedProjectionsCPI.append(cpi[index].tolist())
+
D_lowSpace = distance.squareform(distance.pdist(projectionsAll[index]))
D_lowSpaceList.append(D_lowSpace)
@@ -372,10 +406,13 @@ def background_process():
global overalProjectionsNumber
global metricsMatrix
global metricsMatrixEntire
+ global SelectedProjectionsBeta
+ global SelectedProjectionsCPP
+ global SelectedProjectionsCPI
while (len(projectionsAll) != overalProjectionsNumber):
pass
- return jsonify({ 'projections': SelectedProjectionsReturn, 'parameters': SelectedListofParams, 'metrics': metricsMatrix, 'metricsEntire': metricsMatrixEntire })
+ return jsonify({ 'projections': SelectedProjectionsReturn, 'parameters': SelectedListofParams, 'metrics': metricsMatrix, 'metricsEntire': metricsMatrixEntire, 'betas': SelectedProjectionsBeta, 'cpp': SelectedProjectionsCPP, 'cpi': SelectedProjectionsCPI})
@app.route('/receiverOptimizer', methods = ['POST'])
def OptimizeSelection():
@@ -462,6 +499,69 @@ def SendOptimizedProjections():
return jsonify({'metrics': metricsMatrixSel, 'metricsEntire': metricsMatrixEntireSel })
+@app.route('/receiverSingle', methods = ['POST'])
+def singleParameters():
+ data = request.get_data().decode('utf8').replace("'", '"')
+ data = json.loads(data)
+
+ global dataProc
+ dataProc, length, labels = preprocess(data[3])
+
+ DEFAULT_NO_DIMS = 2
+ VERBOSE = False
+ DEFAULT_USE_PCA = True
+ randseed=1137
+
+ perplexity = int(data[0])
+ learning_rate = int(data[1])
+ n_iter = int(data[2])
+
+ global projectionsAll
+
+ listofParamsPlusData = []
+ listofParamsAll= []
+ listofParamsPlusData.append((dataProc,DEFAULT_NO_DIMS,perplexity,learning_rate,randseed,VERBOSE,length,DEFAULT_USE_PCA,n_iter,True,True,True))
+ listofParamsAll.append((perplexity,learning_rate,n_iter))
+
+ projectionsAll, betas, cpp, cpi = wrapGetResults(listofParamsPlusData)
+
+
+ global SelectedProjectionsReturn
+ SelectedProjectionsReturn = []
+
+ global SelectedProjectionsBeta
+ SelectedProjectionsBeta = []
+
+ global SelectedProjectionsCPP
+ SelectedProjectionsCPP = []
+
+ global SelectedProjectionsCPI
+ SelectedProjectionsCPI = []
+
+ SelectedProjectionsReturn.append(projectionsAll[0].tolist())
+
+ SelectedProjectionsBeta.append(betas[0].tolist())
+
+ SelectedProjectionsCPP.append(cpp[0].tolist())
+
+ SelectedProjectionsCPI.append(cpi[0].tolist())
+
+ return 'OK'
+
+@app.route('/senderSingle')
+def sendSingle():
+
+ global projectionsAll
+ global SelectedProjectionsReturn
+ global SelectedProjectionsBeta
+ global SelectedProjectionsCPP
+ global SelectedProjectionsCPI
+ while (len(projectionsAll) != 1):
+ pass
+ return jsonify({ 'projections': SelectedProjectionsReturn, 'betas': SelectedProjectionsBeta, 'cpp': SelectedProjectionsCPP, 'cpi': SelectedProjectionsCPI})
+
+
if __name__ == '__main__':
app.run("0.0.0.0", "5000")
+
diff --git a/tsne_main.cpp b/tsne_main.cpp
old mode 100644
new mode 100755
index 8efa8fc..a5ad763
--- a/tsne_main.cpp
+++ b/tsne_main.cpp
@@ -10,14 +10,12 @@
int main() {
// Define some variables
- int origN, N, D, no_dims, max_iter, *landmarks;
- double perc_landmarks;
+ int origN, N, D, no_dims, max_iter;
double perplexity, theta, *data;
int rand_seed = -1;
- TSNE* tsne = new TSNE();
// Read the parameters and the dataset
- if(tsne->load_data(&data, &origN, &D, &no_dims, &theta, &perplexity, &rand_seed, &max_iter)) {
+ if(TSNE::load_data(&data, &origN, &D, &no_dims, &theta, &perplexity, &rand_seed, &max_iter)) {
// Make dummy landmarks
N = origN;
@@ -29,10 +27,10 @@ int main() {
double* Y = (double*) malloc(N * no_dims * sizeof(double));
double* costs = (double*) calloc(N, sizeof(double));
if(Y == NULL || costs == NULL) { printf("Memory allocation failed!\n"); exit(1); }
- tsne->run(data, N, D, Y, no_dims, perplexity, theta, rand_seed, false, max_iter);
+ TSNE::run(data, N, D, Y, no_dims, perplexity, theta, rand_seed, false, max_iter, 250, 250, 1);
// Save the results
- tsne->save_data(Y, landmarks, costs, N, no_dims);
+ TSNE::save_data(Y, landmarks, costs, N, no_dims);
// Clean up the memory
free(data); data = NULL;
@@ -40,5 +38,4 @@ int main() {
free(costs); costs = NULL;
free(landmarks); landmarks = NULL;
}
- delete(tsne);
}
diff --git a/vptree.h b/vptree.h
old mode 100644
new mode 100755