fixed bugs in data_graph, dpt and tsne

scanpy/data_structs/data_graph.py

@@ -174,12 +174,19 @@ def __init__(self,
                  recompute_graph=None,
                  flavor='haghverdi16'):
         self.sym = True  # we do not allow asymetric cases
+        self.flavor = flavor  # this is to experiment around
+        self.n_pcs = n_pcs
+        self.n_dcs = n_dcs
         # use the graph in adata
         if (not recompute_graph
+            # make sure X_diffmap is there
             and 'X_diffmap' in adata.smp
-            and adata.smp['X_diffmap'].shape[1] >= n_dcs-1):
-                self.n_pcs = n_pcs
-                self.n_dcs = n_dcs
+            # make sure enough DCs are there
+            and adata.smp['X_diffmap'].shape[1] >= n_dcs-1
+            # make sure that it's sparse
+            and issparse(adata.add['Ktilde']) == knn
+            # make sure n_neighbors matches
+            and k == adata.add['distance'][0].nonzero()[0].size + 1):
                 self.init_iroot_directly(adata)
                 self.X = adata.X  # this is a hack, PCA?
                 self.knn = issparse(adata.add['Ktilde'])
@@ -188,7 +195,7 @@ def __init__(self,
                 if self.knn:
                     self.k = adata.add['distance'][0].nonzero()[0].size + 1
                 else:
-                    self.k = adata.X.shape[0]
+                    self.k = None  # currently do not store this, is unknown
                 # for output of spectrum
                 self.X_diffmap = adata.smp['X_diffmap'][:, :n_dcs-1]
                 self.evals = np.r_[1, adata.add['diffmap_evals'][:n_dcs-1]]
@@ -205,7 +212,7 @@ def __init__(self,
         # recompute the graph
         else:
             self.k = k if k is not None else 30
-            logg.info('compute data graph with `n_neighbors={}`'
+            logg.info('    computing data graph with n_neighbors = {} '
                       .format(self.k))
             self.evals = None
             self.rbasis = None
@@ -214,13 +221,8 @@ def __init__(self,
             self.Dsq = None
             self.knn = knn
             self.n_jobs = sett.n_jobs if n_jobs is None else n_jobs
-            self.n_pcs = n_pcs
-            self.n_dcs = n_dcs
-            self.flavor = flavor  # this is to experiment around
-            self.iroot = None
             self.X = adata.X  # might be overwritten with X_pca below
             self.Dchosen = None
-            self.M = None
             self.init_iroot_and_X_from_PCA(adata, recompute_pca, n_pcs)
             if False:  # TODO
                 # in case we already computed distance relations
@@ -233,16 +235,15 @@ def __init__(self,
                         self.Dsq = adata.add['distance']
 
     def init_iroot_directly(self, adata):
+        self.iroot = None
         if 'iroot' in adata.add:
             if adata.add['iroot'] >= adata.n_smps:
                 logg.warn('Root cell index {} does not exist for {} samples. '
                           'Is ignored.'
                           .format(adata.add['iroot'], adata.n_smps))
-                self.iroot = None
             else:
                 self.iroot = adata.add['iroot']
-
-
+
     def init_iroot_and_X_from_PCA(self, adata, recompute_pca, n_pcs):
         # retrieve xroot
         xroot = None
@@ -281,7 +282,7 @@ def update_diffmap(self, n_comps=None):
             self.n_dcs = n_comps
             logg.info('    updating number of DCs to', self.n_dcs)
         if self.evals is None or self.evals.size < self.n_dcs:
-            logg.info('computing spectral decomposition ("diffmap") with',
+            logg.info('    computing spectral decomposition ("diffmap") with',
                       self.n_dcs, 'components', r=True)
             self.compute_transition_matrix()
             self.embed(n_evals=self.n_dcs)

scanpy/tools/dpt.py

@@ -107,6 +107,7 @@ def dpt(adata, n_branchings=0, n_neighbors=30, knn=True, n_pcs=50, n_dcs=10,
         logg.hint(msg)
     if n_branchings == 0:
         logg.m('set parameter `n_branchings` > 0 to detect branchings', v='hint')
+    logg.m('perform Diffusion Pseudotime analysis', r=True)
     dpt = DPT(adata, n_neighbors=n_neighbors, knn=knn, n_pcs=n_pcs, n_dcs=n_dcs,
               min_group_size=min_group_size, n_jobs=n_jobs,
               recompute_graph=recompute_graph, recompute_pca=recompute_pca,
@@ -118,8 +119,7 @@ def dpt(adata, n_branchings=0, n_neighbors=30, knn=True, n_pcs=50, n_dcs=10,
     adata.add['diffmap_evals'] = dpt.evals[1:]
     if knn: adata.add['distance'] = dpt.Dsq
     if knn: adata.add['Ktilde'] = dpt.Ktilde
-    logg.m('perform Diffusion Pseudotime analysis', r=True)
-    if n_branchings > 1: logg.info('... this uses a hierarchical implementation')
+    if n_branchings > 1: logg.info('    this uses a hierarchical implementation')
     # compute DPT distance matrix, which we refer to as 'Ddiff'
     if dpt.iroot is not None:
         dpt.set_pseudotime()  # pseudotimes are distances from root point

scanpy/tools/tsne.py

@@ -64,7 +64,7 @@ def tsne(adata, random_state=0, n_pcs=50, perplexity=30, learning_rate=None,
     Multicore-TSNE
     GitHub (2017)
     """
-    logg.info('compute tSNE', r=True)
+    logg.info('computing tSNE', r=True)
     adata = adata.copy() if copy else adata
     # preprocessing by PCA
     if 'X_pca' in adata.smp and adata.smp['X_pca'].shape[1] >= n_pcs:
@@ -74,7 +74,7 @@ def tsne(adata, random_state=0, n_pcs=50, perplexity=30, learning_rate=None,
     else:
         if n_pcs > 0 and adata.X.shape[1] > n_pcs:
             logg.info('    preprocess using PCA with', n_pcs, 'PCs')
-            logg.info('avoid this by setting n_pcs = 0', v='hint')
+            logg.hint('avoid this by setting n_pcs = 0')
             X = pca(adata.X, random_state=random_state, n_comps=n_pcs)
             adata.smp['X_pca'] = X
             logg.info('    using', n_pcs, 'principal components')
@@ -95,7 +95,7 @@ def tsne(adata, random_state=0, n_pcs=50, perplexity=30, learning_rate=None,
             from MulticoreTSNE import MulticoreTSNE as TSNE
             params_sklearn['learning_rate'] = 200 if learning_rate is None else learning_rate
             tsne = TSNE(n_jobs=n_jobs, **params_sklearn)
-            logg.info('    using package MulticoreTSNE')
+            logg.info('    using package MulticoreTSNE by D. Ulyanov')
             X_tsne = tsne.fit_transform(X.astype(np.float64))
         except ImportError:
             multicore_failed = True