Implementation of FIt-SNE is available from https://github.com/KlugerLab/FIt-SNE.
This function uses fftRtsne to run FIt-SNE.
Usage
run.fitsne(dat, use.cols, seed = 42, fitsne.x.name = "FItSNE_X",
fitsne.y.name = "FItSNE_Y", dims = 2, perplexity = 30, theta = 0.5,
max_iter = 750, fft_not_bh = TRUE, ann_not_vptree = TRUE,
stop_early_exag_iter = 250, exaggeration_factor = 12.0,
no_momentum_during_exag = FALSE,start_late_exag_iter = -1,
late_exag_coeff = 1.0, mom_switch_iter = 250, momentum = 0.5,
final_momentum = 0.8, learning_rate = 'auto', n_trees = 50,
search_k = -1, nterms = 3, intervals_per_integer = 1,
min_num_intervals = 50, K = -1, sigma = -30, initialization = 'pca',
max_step_norm = 5, load_affinities = NULL, fast_tsne_path = NULL,
nthreads = 0, perplexity_list = NULL, get_costs = FALSE, df = 1.0)Arguments
- dat
NO DEFAULT. Input data.table or data.frame.
- use.cols
NO DEFAULT. Vector of column names or numbers for clustering.
- seed
Default = 42. Seed value for reproducibility.
- fitsne.x.name
Default = "FItSNE_X". Character. Name of FItSNE x-axis.
- fitsne.y.name
Default = "FItSNE_Y". Character. Name of FItSNE y-axis.
- dims
Default = 2. Dimensionality of the embedding (reduced data).
- perplexity
Default = 30. Perplexity is used to determine the bandwidth of the Gaussian kernel in the input space
- theta
Default = 0.5. For exact t-SNE, set to 0. If non-zero, then will use either Barnes Hut or FIt-SNE based on nbody_algo. If Barnes Hut, then this determines the accuracy of BH approximation.
- max_iter
Default = 750. Number of iterations of t-SNE to run.
- fft_not_bh
Default = TRUE. If theta is nonzero, this determines whether to use FIt-SNE or Barnes Hut approximation.
- ann_not_vptree
Default = TRUE. Use vp-trees (as in bhtsne) or approximate nearest neighbors (default). Set to be TRUE for approximate nearest neighbors.
- stop_early_exag_iter
Default = 250. When to switch off early exaggeration.
- exaggeration_factor
Default = 12. Coefficient for early exaggeration (>1).
- no_momentum_during_exag
Default = FALSE. Set to 0 to use momentum and other optimization tricks. Can be set to 1 to do plain, vanilla gradient descent (useful for testing large exaggeration coefficients).
- start_late_exag_iter
Default = -1. When to start late exaggeration. Set to -1 by default to not use late exaggeration.
- late_exag_coeff
Default = 1. Late exaggeration coefficient. Set to 1 by default to not use late exaggeration.
- mom_switch_iter
Default = 250. Iteration number to switch from momentum to final_momentum.
- momentum
Default = 0.5.Initial value of momentum.
- final_momentum
Default = 0.8. Value of momentum to use later in the optimisation.
- learning_rate
Default = 'auto'. Set to desired learning rate or 'auto', which sets learning rate to N/exaggeration_factor where N is the sample size, or to 200 if N/exaggeration_factor < 200.
- n_trees
Default = 50. When using Annoy, the number of search trees to use.
- search_k
Default = -1. When using Annoy, the number of nodes to inspect during search. Default is -1 which translate to 3perplexityn_trees (or K*n_trees when using fixed sigma).
- nterms
Default = 3. If using FIt-SNE, this is the number of interpolation points per sub-interval.
- intervals_per_integer
Default = 1. See min_num_intervals.
- min_num_intervals
Default = 50. Let maxloc = ceil(max(max(X))) and minloc = floor(min(min(X))). i.e. the points are in a minloc^no_dims by maxloc^no_dims interval/square. The number of intervals in each dimension is either min_num_intervals or ceil((maxloc - minloc)/intervals_per_integer), whichever is larger. min_num_intervals must be an integer >0, and intervals_per_integer must be >0. Defaults are min_num_intervals=50 and intervals_per_integer = 1.
- K
Default = -1. Number of nearest neighbours to get when using fixed sigma.
- sigma
Default = -30. Fixed sigma value to use when perplexity==-1.
- initialization
Default = 'pca'. pca', 'random', or N x no_dims array to intialize the solution.
- max_step_norm
Default = 5. Maximum distance that a point is allowed to move on one iteration. Larger steps are clipped to this value. This prevents possible instabilities during gradient descent. Set to -1 to switch it off.
- load_affinities
Default = NULL. If 1, input similarities are loaded from a file and not computed. If 2, input similarities are saved into a file. If 0, affinities are neither saved nor loaded.
- fast_tsne_path
Default = NULL. Path to FItSNE executable.
- nthreads
Default = 0. Number of threads to use, set to use all available threads by default.
- perplexity_list
Default = NULL. If perplexity==0 then perplexity combination will be used with values taken from perplexity_list.
- get_costs
Default = FALSE. Logical indicating whether the KL-divergence costs computed every 50 iterations should be returned.
- df
Default = 1.0. Positive numeric that controls the degree of freedom of t-distribution. The actual degree of freedom is 2*df-1. The standard t-SNE choice of 1 degree of freedom corresponds to df=1. Large df approximates Gaussian kernel. df<1 corresponds to heavier tails, which can often resolve substructure in the embedding. See Kobak et al. (2019) for details.
Examples
if (FALSE) { # \dontrun{
dat <- Spectre::demo.clustered
dat.sub <- Spectre::do.subsample(dat, 30000)
use.cols <- names(dat)[12:19]
dat.reduced <- run.fitsne(dat = dat.sub, use.cols = use.cols)
} # }