tSNE performance in FCS Express

Given that t-SNE is an highly demanding algorithm, the development team at De Novo Software made a great effort to improve the speed of running the algorithm with FCS Express.

 

To give you an idea of how t-SNE is performing within FCS Express, a performance test has been run comparing FCS Express and the independent R package in running both the exact tSNE algorithm and the Barnes-Hut approximation.

 

Here is a brief description of how the test was run:

 

File

Machine

FCS Express

R

FCS format version: 3.0

Total number of events: 500397 (different sample sizes were used in the performance test)

Total number of parameters: 40 (20 parameters have been selected in the performance test)

CPU: Intel Core i5-4200M CPU 2.5GHz

Memory: 8Gb

System type: Windows 8.1 64 bit

Version: 6.00.0045 (64 bit)

R version 3.3.1 (2016-06-21)

R Studio Version 0.99.903

 

 

 

 

The graph below (Figure 29.59) show the elapsed time (in minutes) for t-SNE calculation (using both the exact t-SNE and the Burnes-Hut Approximation (Amount of Approximation = 0.5)) in both FCS Express and R. A fixed number of parameters (i.e. n=20) and different numbers of events have been used. t-SNE was not estimated for unsampled events. Please refer to the Defining a tSNE transformation chapter for more information about all the above mentioned options.

 

Figure 29.59 - Performance comparison between FCS Express and R in running both the exact tSNE and the Barnes-Hut Approximation.

Figure 29.59 - Performance comparison between FCS Express and R in running both the exact tSNE and the Barnes-Hut Approximation.