from packaging import version from typing import Optional, Union import warnings from anndata import AnnData from .._utils import _doc_params, AnyRandom from ..tools._utils import _choose_representation, doc_use_rep, doc_n_pcs from .._settings import settings from .. import logging as logg @_doc_params(doc_n_pcs=doc_n_pcs, use_rep=doc_use_rep) def tsne( adata: AnnData, n_pcs: Optional[int] = None, use_rep: Optional[str] = None, perplexity: Union[float, int] = 30, early_exaggeration: Union[float, int] = 12, learning_rate: Union[float, int] = 1000, random_state: AnyRandom = 0, use_fast_tsne: bool = False, n_jobs: Optional[int] = None, copy: bool = False, *, metric: str = "euclidean", ) -> Optional[AnnData]: """\ t-SNE [Maaten08]_ [Amir13]_ [Pedregosa11]_. t-distributed stochastic neighborhood embedding (tSNE) [Maaten08]_ has been proposed for visualizating single-cell data by [Amir13]_. Here, by default, we use the implementation of *scikit-learn* [Pedregosa11]_. You can achieve a huge speedup and better convergence if you install `Multicore-tSNE `__ by [Ulyanov16]_, which will be automatically detected by Scanpy. Parameters ---------- adata Annotated data matrix. {doc_n_pcs} {use_rep} perplexity The perplexity is related to the number of nearest neighbors that is used in other manifold learning algorithms. Larger datasets usually require a larger perplexity. Consider selecting a value between 5 and 50. The choice is not extremely critical since t-SNE is quite insensitive to this parameter. metric Distance metric calculate neighbors on. early_exaggeration Controls how tight natural clusters in the original space are in the embedded space and how much space will be between them. For larger values, the space between natural clusters will be larger in the embedded space. Again, the choice of this parameter is not very critical. If the cost function increases during initial optimization, the early exaggeration factor or the learning rate might be too high. learning_rate Note that the R-package "Rtsne" uses a default of 200. The learning rate can be a critical parameter. It should be between 100 and 1000. If the cost function increases during initial optimization, the early exaggeration factor or the learning rate might be too high. If the cost function gets stuck in a bad local minimum increasing the learning rate helps sometimes. random_state Change this to use different intial states for the optimization. If `None`, the initial state is not reproducible. n_jobs Number of jobs for parallel computation. `None` means using :attr:`scanpy._settings.ScanpyConfig.n_jobs`. copy Return a copy instead of writing to `adata`. Returns ------- Depending on `copy`, returns or updates `adata` with the following fields. **X_tsne** : `np.ndarray` (`adata.obs`, dtype `float`) tSNE coordinates of data. """ import sklearn start = logg.info('computing tSNE') adata = adata.copy() if copy else adata X = _choose_representation(adata, use_rep=use_rep, n_pcs=n_pcs) # params for sklearn n_jobs = settings.n_jobs if n_jobs is None else n_jobs params_sklearn = dict( perplexity=perplexity, random_state=random_state, verbose=settings.verbosity > 3, early_exaggeration=early_exaggeration, learning_rate=learning_rate, n_jobs=n_jobs, metric=metric, ) # square_distances will default to true in the future, we'll get ahead of the # warning for now if metric != "euclidean": sklearn_version = version.parse(sklearn.__version__) if sklearn_version >= version.parse("0.24.0"): params_sklearn["square_distances"] = True else: warnings.warn( "Results for non-euclidean metrics changed in sklearn 0.24.0, while " f"you are using {sklearn.__version__}.", UserWarning, ) # Backwards compat handling: Remove in scanpy 1.9.0 if n_jobs != 1 and not use_fast_tsne: warnings.warn( UserWarning( "In previous versions of scanpy, calling tsne with n_jobs > 1 would use " "MulticoreTSNE. Now this uses the scikit-learn version of TSNE by default. " "If you'd like the old behaviour (which is deprecated), pass " "'use_fast_tsne=True'. Note, MulticoreTSNE is not actually faster anymore." ) ) if use_fast_tsne: warnings.warn( FutureWarning( "Argument `use_fast_tsne` is deprecated, and support for MulticoreTSNE " "will be dropped in a future version of scanpy." ) ) # deal with different tSNE implementations if use_fast_tsne: try: from MulticoreTSNE import MulticoreTSNE as TSNE tsne = TSNE(**params_sklearn) logg.info(" using the 'MulticoreTSNE' package by Ulyanov (2017)") # need to transform to float64 for MulticoreTSNE... X_tsne = tsne.fit_transform(X.astype('float64')) except ImportError: use_fast_tsne = False warnings.warn( UserWarning( "Could not import 'MulticoreTSNE'. Falling back to scikit-learn." ) ) if use_fast_tsne is False: # In case MultiCore failed to import from sklearn.manifold import TSNE # unfortunately, sklearn does not allow to set a minimum number # of iterations for barnes-hut tSNE tsne = TSNE(**params_sklearn) logg.info(' using sklearn.manifold.TSNE') X_tsne = tsne.fit_transform(X) # update AnnData instance adata.obsm['X_tsne'] = X_tsne # annotate samples with tSNE coordinates adata.uns["tsne"] = { "params": { k: v for k, v in { "perplexity": perplexity, "early_exaggeration": early_exaggeration, "learning_rate": learning_rate, "n_jobs": n_jobs, "metric": metric, "use_rep": use_rep, }.items() if v is not None } } logg.info( ' finished', time=start, deep="added\n 'X_tsne', tSNE coordinates (adata.obsm)", ) return adata if copy else None