from types import MappingProxyType
from typing import Union, Optional, Any, Mapping, Callable, NamedTuple, Generator, Tuple
import numpy as np
import scipy
from anndata import AnnData
from scipy.sparse import issparse, coo_matrix, csr_matrix
from sklearn.utils import check_random_state
from .. import logging as logg
from .. import _utils
from .._utils import _doc_params, AnyRandom, NeighborsView
from .._compat import Literal
from ..tools._utils import _choose_representation, doc_use_rep, doc_n_pcs
from .. import settings
N_DCS = 15 # default number of diffusion components
N_PCS = settings.N_PCS # Backwards compat, constants should be defined in only one place.
_Method = Literal['umap', 'gauss', 'rapids']
_MetricFn = Callable[[np.ndarray, np.ndarray], float]
# from sklearn.metrics.pairwise_distances.__doc__:
_MetricSparseCapable = Literal[
'cityblock', 'cosine', 'euclidean', 'l1', 'l2', 'manhattan'
]
_MetricScipySpatial = Literal[
'braycurtis', 'canberra', 'chebyshev', 'correlation', 'dice', 'hamming',
'jaccard', 'kulsinski', 'mahalanobis', 'minkowski', 'rogerstanimoto',
'russellrao', 'seuclidean', 'sokalmichener', 'sokalsneath', 'sqeuclidean',
'yule'
]
_Metric = Union[_MetricSparseCapable, _MetricScipySpatial]
@_doc_params(n_pcs=doc_n_pcs, use_rep=doc_use_rep)
def neighbors(
adata: AnnData,
n_neighbors: int = 15,
n_pcs: Optional[int] = None,
use_rep: Optional[str] = None,
knn: bool = True,
random_state: AnyRandom = 0,
method: Optional[_Method] = 'umap',
metric: Union[_Metric, _MetricFn] = 'euclidean',
metric_kwds: Mapping[str, Any] = MappingProxyType({}),
key_added: Optional[str] = None,
copy: bool = False,
) -> Optional[AnnData]:
"""\
Compute a neighborhood graph of observations [McInnes18]_.
The neighbor search efficiency of this heavily relies on UMAP [McInnes18]_,
which also provides a method for estimating connectivities of data points -
the connectivity of the manifold (`method=='umap'`). If `method=='gauss'`,
connectivities are computed according to [Coifman05]_, in the adaption of
[Haghverdi16]_.
Parameters
----------
adata
Annotated data matrix.
n_neighbors
The size of local neighborhood (in terms of number of neighboring data
points) used for manifold approximation. Larger values result in more
global views of the manifold, while smaller values result in more local
data being preserved. In general values should be in the range 2 to 100.
If `knn` is `True`, number of nearest neighbors to be searched. If `knn`
is `False`, a Gaussian kernel width is set to the distance of the
`n_neighbors` neighbor.
{n_pcs}
{use_rep}
knn
If `True`, use a hard threshold to restrict the number of neighbors to
`n_neighbors`, that is, consider a knn graph. Otherwise, use a Gaussian
Kernel to assign low weights to neighbors more distant than the
`n_neighbors` nearest neighbor.
random_state
A numpy random seed.
method
Use 'umap' [McInnes18]_ or 'gauss' (Gauss kernel following [Coifman05]_
with adaptive width [Haghverdi16]_) for computing connectivities.
Use 'rapids' for the RAPIDS implementation of UMAP (experimental, GPU
only).
metric
A known metric’s name or a callable that returns a distance.
metric_kwds
Options for the metric.
key_added
If not specified, the neighbors data is stored in .uns['neighbors'],
distances and connectivities are stored in .obsp['distances'] and
.obsp['connectivities'] respectively.
If specified, the neighbors data is added to .uns[key_added],
distances are stored in .obsp[key_added+'_distances'] and
connectivities in .obsp[key_added+'_connectivities'].
copy
Return a copy instead of writing to adata.
Returns
-------
Depending on `copy`, updates or returns `adata` with the following:
See `key_added` parameter description for the storage path of
connectivities and distances.
**connectivities** : sparse matrix of dtype `float32`.
Weighted adjacency matrix of the neighborhood graph of data
points. Weights should be interpreted as connectivities.
**distances** : sparse matrix of dtype `float32`.
Instead of decaying weights, this stores distances for each pair of
neighbors.
"""
start = logg.info('computing neighbors')
adata = adata.copy() if copy else adata
if adata.is_view: # we shouldn't need this here...
adata._init_as_actual(adata.copy())
neighbors = Neighbors(adata)
neighbors.compute_neighbors(
n_neighbors=n_neighbors, knn=knn, n_pcs=n_pcs, use_rep=use_rep,
method=method, metric=metric, metric_kwds=metric_kwds,
random_state=random_state,
)
if key_added is None:
key_added = 'neighbors'
conns_key = 'connectivities'
dists_key = 'distances'
else:
conns_key = key_added + '_connectivities'
dists_key = key_added + '_distances'
adata.uns[key_added] = {}
neighbors_dict = adata.uns[key_added]
neighbors_dict['connectivities_key'] = conns_key
neighbors_dict['distances_key'] = dists_key
neighbors_dict['params'] = {'n_neighbors': neighbors.n_neighbors, 'method': method}
neighbors_dict['params']['metric'] = metric
if metric_kwds:
neighbors_dict['params']['metric_kwds'] = metric_kwds
if use_rep is not None:
neighbors_dict['params']['use_rep'] = use_rep
if n_pcs is not None:
neighbors_dict['params']['n_pcs'] = n_pcs
adata.obsp[dists_key] = neighbors.distances
adata.obsp[conns_key] = neighbors.connectivities
if neighbors.rp_forest is not None:
neighbors_dict['rp_forest'] = neighbors.rp_forest
logg.info(
' finished',
time=start,
deep=(
f'added to `.uns[{key_added!r}]`\n'
f' `.obsp[{dists_key!r}]`, distances for each pair of neighbors\n'
f' `.obsp[{conns_key!r}]`, weighted adjacency matrix'
),
)
return adata if copy else None
class FlatTree(NamedTuple):
hyperplanes: None
offsets: None
children: None
indices: None
RPForestDict = Mapping[str, Mapping[str, np.ndarray]]
def _rp_forest_generate(rp_forest_dict: RPForestDict) -> Generator[FlatTree, None, None]:
props = FlatTree._fields
num_trees = len(rp_forest_dict[props[0]]['start'])-1
for i in range(num_trees):
tree = []
for prop in props:
start = rp_forest_dict[prop]['start'][i]
end = rp_forest_dict[prop]['start'][i+1]
tree.append(rp_forest_dict[prop]['data'][start:end])
yield FlatTree(*tree)
tree = []
for prop in props:
start = rp_forest_dict[prop]['start'][num_trees]
tree.append(rp_forest_dict[prop]['data'][start:])
yield FlatTree(*tree)
def compute_neighbors_umap(
X: Union[np.ndarray, csr_matrix],
n_neighbors: int,
random_state: AnyRandom = None,
metric: Union[_Metric, _MetricFn] = 'euclidean',
metric_kwds: Mapping[str, Any] = MappingProxyType({}),
angular: bool = False,
verbose: bool = False,
):
"""This is from umap.fuzzy_simplicial_set [McInnes18]_.
Given a set of data X, a neighborhood size, and a measure of distance
compute the fuzzy simplicial set (here represented as a fuzzy graph in
the form of a sparse matrix) associated to the data. This is done by
locally approximating geodesic distance at each point, creating a fuzzy
simplicial set for each such point, and then combining all the local
fuzzy simplicial sets into a global one via a fuzzy union.
Parameters
----------
X: array of shape (n_samples, n_features)
The data to be modelled as a fuzzy simplicial set.
n_neighbors
The number of neighbors to use to approximate geodesic distance.
Larger numbers induce more global estimates of the manifold that can
miss finer detail, while smaller values will focus on fine manifold
structure to the detriment of the larger picture.
random_state
A state capable being used as a numpy random state.
metric
The metric to use to compute distances in high dimensional space.
If a string is passed it must match a valid predefined metric. If
a general metric is required a function that takes two 1d arrays and
returns a float can be provided. For performance purposes it is
required that this be a numba jit'd function. Valid string metrics
include:
* euclidean
* manhattan
* chebyshev
* minkowski
* canberra
* braycurtis
* mahalanobis
* wminkowski
* seuclidean
* cosine
* correlation
* haversine
* hamming
* jaccard
* dice
* russelrao
* kulsinski
* rogerstanimoto
* sokalmichener
* sokalsneath
* yule
Metrics that take arguments (such as minkowski, mahalanobis etc.)
can have arguments passed via the metric_kwds dictionary. At this
time care must be taken and dictionary elements must be ordered
appropriately; this will hopefully be fixed in the future.
metric_kwds
Arguments to pass on to the metric, such as the ``p`` value for
Minkowski distance.
angular
Whether to use angular/cosine distance for the random projection
forest for seeding NN-descent to determine approximate nearest
neighbors.
verbose
Whether to report information on the current progress of the algorithm.
Returns
-------
**knn_indices**, **knn_dists** : np.arrays of shape (n_observations, n_neighbors)
"""
from umap.umap_ import nearest_neighbors
random_state = check_random_state(random_state)
knn_indices, knn_dists, forest = nearest_neighbors(
X, n_neighbors, random_state=random_state,
metric=metric, metric_kwds=metric_kwds,
angular=angular, verbose=verbose,
)
return knn_indices, knn_dists, forest
def compute_neighbors_rapids(
X: np.ndarray,
n_neighbors: int
):
"""Compute nearest neighbors using RAPIDS cuml.
Parameters
----------
X: array of shape (n_samples, n_features)
The data to compute nearest neighbors for.
n_neighbors
The number of neighbors to use.
Returns
-------
**knn_indices**, **knn_dists** : np.arrays of shape (n_observations, n_neighbors)
"""
from cuml.neighbors import NearestNeighbors
nn = NearestNeighbors(n_neighbors=n_neighbors)
X_contiguous = np.ascontiguousarray(X, dtype=np.float32)
nn.fit(X_contiguous)
knn_distsq, knn_indices = nn.kneighbors(X_contiguous)
return knn_indices, np.sqrt(knn_distsq) # cuml uses sqeuclidean metric so take sqrt
def _get_sparse_matrix_from_indices_distances_umap(knn_indices, knn_dists, n_obs, n_neighbors):
rows = np.zeros((n_obs * n_neighbors), dtype=np.int64)
cols = np.zeros((n_obs * n_neighbors), dtype=np.int64)
vals = np.zeros((n_obs * n_neighbors), dtype=np.float64)
for i in range(knn_indices.shape[0]):
for j in range(n_neighbors):
if knn_indices[i, j] == -1:
continue # We didn't get the full knn for i
if knn_indices[i, j] == i:
val = 0.0
else:
val = knn_dists[i, j]
rows[i * n_neighbors + j] = i
cols[i * n_neighbors + j] = knn_indices[i, j]
vals[i * n_neighbors + j] = val
result = coo_matrix((vals, (rows, cols)),
shape=(n_obs, n_obs))
result.eliminate_zeros()
return result.tocsr()
def _compute_connectivities_umap(
knn_indices, knn_dists,
n_obs, n_neighbors, set_op_mix_ratio=1.0,
local_connectivity=1.0,
):
"""\
This is from umap.fuzzy_simplicial_set [McInnes18]_.
Given a set of data X, a neighborhood size, and a measure of distance
compute the fuzzy simplicial set (here represented as a fuzzy graph in
the form of a sparse matrix) associated to the data. This is done by
locally approximating geodesic distance at each point, creating a fuzzy
simplicial set for each such point, and then combining all the local
fuzzy simplicial sets into a global one via a fuzzy union.
"""
from umap.umap_ import fuzzy_simplicial_set
X = coo_matrix(([], ([], [])), shape=(n_obs, 1))
connectivities = fuzzy_simplicial_set(
X,
n_neighbors,
None,
None,
knn_indices=knn_indices,
knn_dists=knn_dists,
set_op_mix_ratio=set_op_mix_ratio,
local_connectivity=local_connectivity,
)
if isinstance(connectivities, tuple):
# In umap-learn 0.4, this returns (result, sigmas, rhos)
connectivities = connectivities[0]
distances = _get_sparse_matrix_from_indices_distances_umap(
knn_indices, knn_dists, n_obs, n_neighbors
)
return distances, connectivities.tocsr()
def _get_sparse_matrix_from_indices_distances_numpy(indices, distances, n_obs, n_neighbors):
n_nonzero = n_obs * n_neighbors
indptr = np.arange(0, n_nonzero + 1, n_neighbors)
D = csr_matrix((
distances.copy().ravel(), # copy the data, otherwise strange behavior here
indices.copy().ravel(),
indptr,
), shape=(n_obs, n_obs))
D.eliminate_zeros()
return D
def _get_indices_distances_from_sparse_matrix(D, n_neighbors: int):
indices = np.zeros((D.shape[0], n_neighbors), dtype=int)
distances = np.zeros((D.shape[0], n_neighbors), dtype=D.dtype)
n_neighbors_m1 = n_neighbors - 1
for i in range(indices.shape[0]):
neighbors = D[i].nonzero() # 'true' and 'spurious' zeros
indices[i, 0] = i
distances[i, 0] = 0
# account for the fact that there might be more than n_neighbors
# due to an approximate search
# [the point itself was not detected as its own neighbor during the search]
if len(neighbors[1]) > n_neighbors_m1:
sorted_indices = np.argsort(D[i][neighbors].A1)[:n_neighbors_m1]
indices[i, 1:] = neighbors[1][sorted_indices]
distances[i, 1:] = D[i][
neighbors[0][sorted_indices], neighbors[1][sorted_indices]]
else:
indices[i, 1:] = neighbors[1]
distances[i, 1:] = D[i][neighbors]
return indices, distances
def _get_indices_distances_from_dense_matrix(D, n_neighbors: int):
sample_range = np.arange(D.shape[0])[:, None]
indices = np.argpartition(D, n_neighbors-1, axis=1)[:, :n_neighbors]
indices = indices[sample_range, np.argsort(D[sample_range, indices])]
distances = D[sample_range, indices]
return indices, distances
def _backwards_compat_get_full_X_diffmap(adata: AnnData) -> np.ndarray:
if 'X_diffmap0' in adata.obs:
return np.c_[adata.obs['X_diffmap0'].values[:, None],
adata.obsm['X_diffmap']]
else:
return adata.obsm['X_diffmap']
def _backwards_compat_get_full_eval(adata: AnnData):
if 'X_diffmap0' in adata.obs:
return np.r_[1, adata.uns['diffmap_evals']]
else:
return adata.uns['diffmap_evals']
def _make_forest_dict(forest):
d = {}
props = ('hyperplanes', 'offsets', 'children', 'indices')
for prop in props:
d[prop] = {}
sizes = np.fromiter((getattr(tree, prop).shape[0] for tree in forest), dtype=int)
d[prop]['start'] = np.zeros_like(sizes)
if prop == 'offsets':
dims = sizes.sum()
else:
dims = (sizes.sum(), getattr(forest[0], prop).shape[1])
dtype = getattr(forest[0], prop).dtype
dat = np.empty(dims, dtype=dtype)
start = 0
for i, size in enumerate(sizes):
d[prop]['start'][i] = start
end = start+size
dat[start:end] = getattr(forest[i], prop)
start = end
d[prop]['data'] = dat
return d
class OnFlySymMatrix:
"""Emulate a matrix where elements are calculated on the fly.
"""
def __init__(
self,
get_row: Callable[[Any], np.ndarray],
shape: Tuple[int, int],
DC_start: int = 0,
DC_end: int = -1,
rows: Optional[Mapping[Any, np.ndarray]] = None,
restrict_array: Optional[np.ndarray] = None,
):
self.get_row = get_row
self.shape = shape
self.DC_start = DC_start
self.DC_end = DC_end
self.rows = {} if rows is None else rows
self.restrict_array = restrict_array # restrict the array to a subset
def __getitem__(self, index):
if isinstance(index, (int, np.integer)):
if self.restrict_array is None:
glob_index = index
else:
# map the index back to the global index
glob_index = self.restrict_array[index]
if glob_index not in self.rows:
self.rows[glob_index] = self.get_row(glob_index)
row = self.rows[glob_index]
if self.restrict_array is None:
return row
else:
return row[self.restrict_array]
else:
if self.restrict_array is None:
glob_index_0, glob_index_1 = index
else:
glob_index_0 = self.restrict_array[index[0]]
glob_index_1 = self.restrict_array[index[1]]
if glob_index_0 not in self.rows:
self.rows[glob_index_0] = self.get_row(glob_index_0)
return self.rows[glob_index_0][glob_index_1]
def restrict(self, index_array):
"""Generate a view restricted to a subset of indices.
"""
new_shape = index_array.shape[0], index_array.shape[0]
return OnFlySymMatrix(
self.get_row, new_shape, DC_start=self.DC_start,
DC_end=self.DC_end,
rows=self.rows, restrict_array=index_array,
)
class Neighbors:
"""\
Data represented as graph of nearest neighbors.
Represent a data matrix as a graph of nearest neighbor relations (edges)
among data points (nodes).
Parameters
----------
adata
Annotated data object.
n_dcs
Number of diffusion components to use.
neighbors_key
Where to look in .uns and .obsp for neighbors data
"""
def __init__(self, adata: AnnData, n_dcs: Optional[int] = None, neighbors_key: Optional[str] = None):
self._adata = adata
self._init_iroot()
# use the graph in adata
info_str = ''
self.knn: Optional[bool] = None
self._distances: Union[np.ndarray, csr_matrix, None] = None
self._connectivities: Union[np.ndarray, csr_matrix, None] = None
self._transitions_sym: Union[np.ndarray, csr_matrix, None] = None
self._number_connected_components: Optional[int] = None
self._rp_forest: Optional[RPForestDict] = None
if neighbors_key is None:
neighbors_key = 'neighbors'
if neighbors_key in adata.uns:
neighbors = NeighborsView(adata, neighbors_key)
if 'distances' in neighbors:
self.knn = issparse(neighbors['distances'])
self._distances = neighbors['distances']
if 'connectivities' in neighbors:
self.knn = issparse(neighbors['connectivities'])
self._connectivities = neighbors['connectivities']
if 'rp_forest' in neighbors:
self._rp_forest = neighbors['rp_forest']
if 'params' in neighbors:
self.n_neighbors = neighbors['params']['n_neighbors']
else:
def count_nonzero(a: Union[np.ndarray, csr_matrix]) -> int:
return a.count_nonzero() if issparse(a) else np.count_nonzero(a)
# estimating n_neighbors
if self._connectivities is None:
self.n_neighbors = int(count_nonzero(self._distances) / self._distances.shape[0])
else:
self.n_neighbors = int(count_nonzero(self._connectivities) / self._connectivities.shape[0] / 2)
info_str += '`.distances` `.connectivities` '
self._number_connected_components = 1
if issparse(self._connectivities):
from scipy.sparse.csgraph import connected_components
self._connected_components = connected_components(self._connectivities)
self._number_connected_components = self._connected_components[0]
if 'X_diffmap' in adata.obsm_keys():
self._eigen_values = _backwards_compat_get_full_eval(adata)
self._eigen_basis = _backwards_compat_get_full_X_diffmap(adata)
if n_dcs is not None:
if n_dcs > len(self._eigen_values):
raise ValueError(
'Cannot instantiate using `n_dcs`={}. '
'Compute diffmap/spectrum with more components first.'
.format(n_dcs))
self._eigen_values = self._eigen_values[:n_dcs]
self._eigen_basis = self._eigen_basis[:, :n_dcs]
self.n_dcs = len(self._eigen_values)
info_str += '`.eigen_values` `.eigen_basis` `.distances_dpt`'
else:
self._eigen_values = None
self._eigen_basis = None
self.n_dcs = None
if info_str != '':
logg.debug(f' initialized {info_str}')
@property
def rp_forest(self) -> Optional[RPForestDict]:
return self._rp_forest
@property
def distances(self) -> Union[np.ndarray, csr_matrix, None]:
"""Distances between data points (sparse matrix).
"""
return self._distances
@property
def connectivities(self) -> Union[np.ndarray, csr_matrix, None]:
"""Connectivities between data points (sparse matrix).
"""
return self._connectivities
@property
def transitions(self) -> Union[np.ndarray, csr_matrix]:
"""Transition matrix (sparse matrix).
Is conjugate to the symmetrized transition matrix via::
self.transitions = self.Z * self.transitions_sym / self.Z
where ``self.Z`` is the diagonal matrix storing the normalization of the
underlying kernel matrix.
Notes
-----
This has not been tested, in contrast to `transitions_sym`.
"""
if issparse(self.Z):
Zinv = self.Z.power(-1)
else:
Zinv = np.diag(1./np.diag(self.Z))
return self.Z @ self.transitions_sym @ Zinv
@property
def transitions_sym(self) -> Union[np.ndarray, csr_matrix, None]:
"""Symmetrized transition matrix (sparse matrix).
Is conjugate to the transition matrix via::
self.transitions_sym = self.Z / self.transitions * self.Z
where ``self.Z`` is the diagonal matrix storing the normalization of the
underlying kernel matrix.
"""
return self._transitions_sym
@property
def eigen_values(self):
"""Eigen values of transition matrix (numpy array).
"""
return self._eigen_values
@property
def eigen_basis(self):
"""Eigen basis of transition matrix (numpy array).
"""
return self._eigen_basis
@property
def distances_dpt(self):
"""DPT distances (on-fly matrix).
This is yields [Haghverdi16]_, Eq. 15 from the supplement with the
extensions of [Wolf19]_, supplement on random-walk based distance
measures.
"""
return OnFlySymMatrix(self._get_dpt_row, shape=self._adata.shape)
def to_igraph(self):
"""Generate igraph from connectiviies.
"""
return _utils.get_igraph_from_adjacency(self.connectivities)
@_doc_params(n_pcs=doc_n_pcs, use_rep=doc_use_rep)
def compute_neighbors(
self,
n_neighbors: int = 30,
knn: bool = True,
n_pcs: Optional[int] = None,
use_rep: Optional[str] = None,
method: _Method = 'umap',
random_state: AnyRandom = 0,
write_knn_indices: bool = False,
metric: _Metric = 'euclidean',
metric_kwds: Mapping[str, Any] = MappingProxyType({}),
) -> None:
"""\
Compute distances and connectivities of neighbors.
Parameters
----------
n_neighbors
Use this number of nearest neighbors.
knn
Restrict result to `n_neighbors` nearest neighbors.
{n_pcs}
{use_rep}
Returns
-------
Writes sparse graph attributes `.distances` and `.connectivities`.
Also writes `.knn_indices` and `.knn_distances` if
`write_knn_indices==True`.
"""
from sklearn.metrics import pairwise_distances
start_neighbors = logg.debug('computing neighbors')
if n_neighbors > self._adata.shape[0]: # very small datasets
n_neighbors = 1 + int(0.5*self._adata.shape[0])
logg.warning(f'n_obs too small: adjusting to `n_neighbors = {n_neighbors}`')
if method == 'umap' and not knn:
raise ValueError('`method = \'umap\' only with `knn = True`.')
if method == 'rapids' and metric != 'euclidean':
raise ValueError("`method` 'rapids' only supports the 'euclidean' `metric`.")
if method not in {'umap', 'gauss', 'rapids'}:
raise ValueError("`method` needs to be 'umap', 'gauss', or 'rapids'.")
if self._adata.shape[0] >= 10000 and not knn:
logg.warning('Using high n_obs without `knn=True` takes a lot of memory...')
self.n_neighbors = n_neighbors
self.knn = knn
X = _choose_representation(self._adata, use_rep=use_rep, n_pcs=n_pcs)
# neighbor search
use_dense_distances = (metric == 'euclidean' and X.shape[0] < 8192) or knn == False
if use_dense_distances:
_distances = pairwise_distances(X, metric=metric, **metric_kwds)
knn_indices, knn_distances = _get_indices_distances_from_dense_matrix(
_distances, n_neighbors)
if knn:
self._distances = _get_sparse_matrix_from_indices_distances_numpy(
knn_indices, knn_distances, X.shape[0], n_neighbors)
else:
self._distances = _distances
elif method == 'rapids':
knn_indices, knn_distances = compute_neighbors_rapids(X, n_neighbors)
else:
# non-euclidean case and approx nearest neighbors
if X.shape[0] < 4096:
X = pairwise_distances(X, metric=metric, **metric_kwds)
metric = 'precomputed'
knn_indices, knn_distances, forest = compute_neighbors_umap(
X, n_neighbors, random_state, metric=metric, metric_kwds=metric_kwds)
# very cautious here
try:
if forest:
self._rp_forest = _make_forest_dict(forest)
except:
pass
# write indices as attributes
if write_knn_indices:
self.knn_indices = knn_indices
self.knn_distances = knn_distances
start_connect = logg.debug('computed neighbors', time=start_neighbors)
if not use_dense_distances or method in {'umap', 'rapids'}:
# we need self._distances also for method == 'gauss' if we didn't
# use dense distances
self._distances, self._connectivities = _compute_connectivities_umap(
knn_indices,
knn_distances,
self._adata.shape[0],
self.n_neighbors,
)
# overwrite the umap connectivities if method is 'gauss'
# self._distances is unaffected by this
if method == 'gauss':
self._compute_connectivities_diffmap()
logg.debug('computed connectivities', time=start_connect)
self._number_connected_components = 1
if issparse(self._connectivities):
from scipy.sparse.csgraph import connected_components
self._connected_components = connected_components(self._connectivities)
self._number_connected_components = self._connected_components[0]
def _compute_connectivities_diffmap(self, density_normalize=True):
# init distances
if self.knn:
Dsq = self._distances.power(2)
indices, distances_sq = _get_indices_distances_from_sparse_matrix(
Dsq, self.n_neighbors)
else:
Dsq = np.power(self._distances, 2)
indices, distances_sq = _get_indices_distances_from_dense_matrix(
Dsq, self.n_neighbors)
# exclude the first point, the 0th neighbor
indices = indices[:, 1:]
distances_sq = distances_sq[:, 1:]
# choose sigma, the heuristic here doesn't seem to make much of a difference,
# but is used to reproduce the figures of Haghverdi et al. (2016)
if self.knn:
# as the distances are not sorted
# we have decay within the n_neighbors first neighbors
sigmas_sq = np.median(distances_sq, axis=1)
else:
# the last item is already in its sorted position through argpartition
# we have decay beyond the n_neighbors neighbors
sigmas_sq = distances_sq[:, -1]/4
sigmas = np.sqrt(sigmas_sq)
# compute the symmetric weight matrix
if not issparse(self._distances):
Num = 2 * np.multiply.outer(sigmas, sigmas)
Den = np.add.outer(sigmas_sq, sigmas_sq)
W = np.sqrt(Num/Den) * np.exp(-Dsq/Den)
# make the weight matrix sparse
if not self.knn:
mask = W > 1e-14
W[mask == False] = 0
else:
# restrict number of neighbors to ~k
# build a symmetric mask
mask = np.zeros(Dsq.shape, dtype=bool)
for i, row in enumerate(indices):
mask[i, row] = True
for j in row:
if i not in set(indices[j]):
W[j, i] = W[i, j]
mask[j, i] = True
# set all entries that are not nearest neighbors to zero
W[mask == False] = 0
else:
W = Dsq.copy() # need to copy the distance matrix here; what follows is inplace
for i in range(len(Dsq.indptr[:-1])):
row = Dsq.indices[Dsq.indptr[i]:Dsq.indptr[i+1]]
num = 2 * sigmas[i] * sigmas[row]
den = sigmas_sq[i] + sigmas_sq[row]
W.data[Dsq.indptr[i]:Dsq.indptr[i+1]] = np.sqrt(num/den) * np.exp(
-Dsq.data[Dsq.indptr[i]: Dsq.indptr[i+1]] / den)
W = W.tolil()
for i, row in enumerate(indices):
for j in row:
if i not in set(indices[j]):
W[j, i] = W[i, j]
W = W.tocsr()
self._connectivities = W
def compute_transitions(self, density_normalize: bool = True):
"""\
Compute transition matrix.
Parameters
----------
density_normalize
The density rescaling of Coifman and Lafon (2006): Then only the
geometry of the data matters, not the sampled density.
Returns
-------
Makes attributes `.transitions_sym` and `.transitions` available.
"""
start = logg.info('computing transitions')
W = self._connectivities
# density normalization as of Coifman et al. (2005)
# ensures that kernel matrix is independent of sampling density
if density_normalize:
# q[i] is an estimate for the sampling density at point i
# it's also the degree of the underlying graph
q = np.asarray(W.sum(axis=0))
if not issparse(W):
Q = np.diag(1.0/q)
else:
Q = scipy.sparse.spdiags(1.0/q, 0, W.shape[0], W.shape[0])
K = Q @ W @ Q
else:
K = W
# z[i] is the square root of the row sum of K
z = np.sqrt(np.asarray(K.sum(axis=0)))
if not issparse(K):
self.Z = np.diag(1.0/z)
else:
self.Z = scipy.sparse.spdiags(1.0/z, 0, K.shape[0], K.shape[0])
self._transitions_sym = self.Z @ K @ self.Z
logg.info(' finished', time=start)
def compute_eigen(
self,
n_comps: int = 15,
sym: Optional[bool] = None,
sort: Literal['decrease', 'increase'] = 'decrease',
):
"""\
Compute eigen decomposition of transition matrix.
Parameters
----------
n_comps
Number of eigenvalues/vectors to be computed, set `n_comps = 0` if
you need all eigenvectors.
sym
Instead of computing the eigendecomposition of the assymetric
transition matrix, computed the eigendecomposition of the symmetric
Ktilde matrix.
Returns
-------
Writes the following attributes.
eigen_values : numpy.ndarray
Eigenvalues of transition matrix.
eigen_basis : numpy.ndarray
Matrix of eigenvectors (stored in columns). `.eigen_basis` is
projection of data matrix on right eigenvectors, that is, the
projection on the diffusion components. these are simply the
components of the right eigenvectors and can directly be used for
plotting.
"""
np.set_printoptions(precision=10)
if self._transitions_sym is None:
raise ValueError('Run `.compute_transitions` first.')
matrix = self._transitions_sym
# compute the spectrum
if n_comps == 0:
evals, evecs = scipy.linalg.eigh(matrix)
else:
n_comps = min(matrix.shape[0]-1, n_comps)
# ncv = max(2 * n_comps + 1, int(np.sqrt(matrix.shape[0])))
ncv = None
which = 'LM' if sort == 'decrease' else 'SM'
# it pays off to increase the stability with a bit more precision
matrix = matrix.astype(np.float64)
evals, evecs = scipy.sparse.linalg.eigsh(
matrix, k=n_comps, which=which, ncv=ncv
)
evals, evecs = evals.astype(np.float32), evecs.astype(np.float32)
if sort == 'decrease':
evals = evals[::-1]
evecs = evecs[:, ::-1]
logg.info(
' eigenvalues of transition matrix\n'
' {}'.format(str(evals).replace('\n', '\n '))
)
if self._number_connected_components > len(evals)/2:
logg.warning('Transition matrix has many disconnected components!')
self._eigen_values = evals
self._eigen_basis = evecs
def _init_iroot(self):
self.iroot = None
# set iroot directly
if 'iroot' in self._adata.uns:
if self._adata.uns['iroot'] >= self._adata.n_obs:
logg.warning(
f'Root cell index {self._adata.uns["iroot"]} does not '
f'exist for {self._adata.n_obs} samples. It’s ignored.'
)
else:
self.iroot = self._adata.uns['iroot']
return
# set iroot via xroot
xroot = None
if 'xroot' in self._adata.uns: xroot = self._adata.uns['xroot']
elif 'xroot' in self._adata.var: xroot = self._adata.var['xroot']
# see whether we can set self.iroot using the full data matrix
if xroot is not None and xroot.size == self._adata.shape[1]:
self._set_iroot_via_xroot(xroot)
def _get_dpt_row(self, i):
mask = None
if self._number_connected_components > 1:
label = self._connected_components[1][i]
mask = self._connected_components[1] == label
row = sum(
(
self.eigen_values[l] / (1-self.eigen_values[l])
* (self.eigen_basis[i, l] - self.eigen_basis[:, l])
)**2
# account for float32 precision
for l in range(0, self.eigen_values.size)
if self.eigen_values[l] < 0.9994
)
# thanks to Marius Lange for pointing Alex to this:
# we will likely remove the contributions from the stationary state below when making
# backwards compat breaking changes, they originate from an early implementation in 2015
# they never seem to have deteriorated results, but also other distance measures (see e.g.
# PAGA paper) don't have it, which makes sense
row += sum(
(self.eigen_basis[i, l] - self.eigen_basis[:, l])**2
for l in range(0, self.eigen_values.size)
if self.eigen_values[l] >= 0.9994
)
if mask is not None:
row[~mask] = np.inf
return np.sqrt(row)
def _set_pseudotime(self):
"""Return pseudotime with respect to root point.
"""
self.pseudotime = self.distances_dpt[self.iroot].copy()
self.pseudotime /= np.max(self.pseudotime[self.pseudotime < np.inf])
def _set_iroot_via_xroot(self, xroot):
"""Determine the index of the root cell.
Given an expression vector, find the observation index that is closest
to this vector.
Parameters
----------
xroot : np.ndarray
Vector that marks the root cell, the vector storing the initial
condition, only relevant for computing pseudotime.
"""
if self._adata.shape[1] != xroot.size:
raise ValueError(
'The root vector you provided does not have the '
'correct dimension.')
# this is the squared distance
dsqroot = 1e10
iroot = 0
for i in range(self._adata.shape[0]):
diff = self._adata.X[i, :] - xroot
dsq = diff @ diff
if dsq < dsqroot:
dsqroot = dsq
iroot = i
if np.sqrt(dsqroot) < 1e-10: break
logg.debug(f'setting root index to {iroot}')
if self.iroot is not None and iroot != self.iroot:
logg.warning(f'Changing index of iroot from {self.iroot} to {iroot}.')
self.iroot = iroot
