Options - Options - 5.3 English - 68552

The available Python options are detailed in the aoclda.dimension_reduction.tsne() class constructor.

The following options can be set using da_options_set_?:

Table 4.22 t-SNE options#

Option name	Type	Default	Description	Constraints
n_components	integer	\(i=2\)	Number of embedding dimensions.	\(1 \le i \le 3\)
perplexity	real	\(r=30\)	Target perplexity for conditional probabilities.	\(1 \le r\)
learning rate	real	\(r=-1\)	Gradient descent learning rate. Any non-positive value selects auto: \(\max(n_{\mathrm{samples}} /\) early exaggeration \(/ 4,\; 50)\).	Any \(r\)
max_iter	integer	\(i=1000\)	Maximum number of gradient descent iterations.	\(1 \le i\)
n_iter_without_progress	integer	\(i=300\)	Stop if no progress is made for this many iterations.	\(0 \le i\)
min_grad_norm	real	\(r=1e-07\)	Stop if the gradient norm is below this threshold.	\(0 \le r\)
early exaggeration	real	\(r=12\)	Exaggeration factor for early iterations.	\(1 \le r\)
theta	real	\(r=0.5\)	Barnes-Hut approximation parameter (0 for exact).	\(0 \le r \le 1\)
init	string	\(s=\) pca	Initialization method for the embedding.	\(s=\) pca, random, or supplied.
seed	integer	\(i=0\)	Seed for random number generation; set to -1 for non-deterministic results.	\(-1 \le i\)
mixed precision	string	\(s=\) no	Whether to use mixed precision iterative refinement, in which lower precision arithmetic is used before switching to the working precision for the final iterations.	\(s=\) no, or yes.
low precision max_iter	integer	\(i=200\)	If mixed precision iterative refinement is enabled, maximum number of iterations for the low precision phase.	\(1 \le i\)
low precision min_grad_norm	real	\(r=0.0001\)	If mixed precision iterative refinement is enabled, gradient norm convergence threshold for the low precision phase.	\(0 \le r\)
check data	string	\(s=\) no	Check input data for NaNs prior to performing computation.	\(s=\) no, or yes.
storage order	string	\(s=\) column-major	Whether data is supplied and returned in row- or column-major order.	\(s=\) c, column-major, f, fortran, or row-major.

If init is set to supplied, an explicit initial embedding must be provided via da_tsne_set_init_embedding_? before calling da_tsne_compute_?.

If learning rate is set to any non-positive value (the default is -1), the learning rate is automatically set to \(\max(n_{\mathrm{samples}} /\) early exaggeration \(/ 4,\; 50)\).

For options with data-dependent bounds, the valid maxima are determined after data is set:

• n_components must satisfy \(n_{\mathrm{components}} \le \min(3, n_{\mathrm{features}})\).

• perplexity must satisfy \(\text{perplexity} \le n_{\mathrm{samples}} - 1\).

If a user-provided value exceeds these bounds, it is reduced to the nearest valid value and the API returns an incompatible-options warning status.

The option mixed precision switches on an experimental mode in which an initial embedding is found in lower precision before refining the result in the working precision. The option low precision max_iter sets the maximum number of iterations for the low-precision phase, and low precision min_grad_norm sets the convergence tolerance for the low-precision phase.