Merge pull request #734 from mattsmart/master

Reproducibility for Aligned UMAP

umap/aligned_umap.py

@@ -352,11 +352,13 @@ def fit(self, X, y=None, **fit_params):
                 make_epochs_per_sample(mapper.graph_.tocoo().data, n_epochs)
             )
 
+        rng_state_transform = np.random.RandomState(self.transform_seed)
         regularisation_weights = build_neighborhood_similarities(
             indptr_list, indices_list, relations,
         )
         first_init = spectral_layout(
-            self.mappers_[0]._raw_data, self.mappers_[0].graph_, self.n_components, np.random,
+            self.mappers_[0]._raw_data, self.mappers_[0].graph_, self.n_components,
+            rng_state_transform,
         )
         expansion = 10.0 / np.abs(first_init).max()
         first_embedding = (first_init * expansion).astype(np.float32, order="C",)
@@ -365,7 +367,8 @@ def fit(self, X, y=None, **fit_params):
         embeddings.append(first_embedding)
         for i in range(1, self.n_models_):
             next_init = spectral_layout(
-                self.mappers_[i]._raw_data, self.mappers_[i].graph_, self.n_components, np.random,
+                self.mappers_[i]._raw_data, self.mappers_[i].graph_, self.n_components,
+                rng_state_transform,
             )
             expansion = 10.0 / np.abs(next_init).max()
             next_embedding = (next_init * expansion).astype(np.float32, order="C",)
@@ -380,9 +383,9 @@ def fit(self, X, y=None, **fit_params):
                 )
             )
 
-        random_state = check_random_state(self.random_state)
-        rng_state = random_state.randint(INT32_MIN, INT32_MAX, 3).astype(np.int64)
-
+        seed_triplet = rng_state_transform.randint(
+            INT32_MIN, INT32_MAX, 3
+        ).astype(np.int64)
         self.embeddings_ = optimize_layout_aligned_euclidean(
             embeddings,
             embeddings,
@@ -392,7 +395,7 @@ def fit(self, X, y=None, **fit_params):
             epochs_per_samples,
             regularisation_weights,
             relations,
-            rng_state,
+            seed_triplet,
             lambda_=self.alignment_regularisation,
         )
 
@@ -436,6 +439,8 @@ def update(self, X, y=None, **fit_params):
             set_op_mix_ratio=get_nth_item_or_val(self.set_op_mix_ratio, self.n_models_),
             unique=get_nth_item_or_val(self.unique, self.n_models_),
             n_components=self.n_components,
+            random_state=self.random_state,
+            transform_seed=self.transform_seed,
         ).fit(X)
 
         self.mappers_ += [new_mapper]
@@ -477,11 +482,12 @@ def update(self, X, y=None, **fit_params):
             self.embeddings_[-1], new_mapper.graph_, new_dict_relations
         )
 
-        random_state = check_random_state(self.random_state)
-        rng_state = random_state.randint(INT32_MIN, INT32_MAX, 3).astype(np.int64)
-
         self.embeddings_.append(new_embedding)
 
+        rng_state_transform = np.random.RandomState(self.transform_seed)
+        seed_triplet = rng_state_transform.randint(
+            INT32_MIN, INT32_MAX, 3
+        ).astype(np.int64)
         self.embeddings_ = optimize_layout_aligned_euclidean(
             self.embeddings_,
             self.embeddings_,
@@ -491,6 +497,6 @@ def update(self, X, y=None, **fit_params):
             epochs_per_samples,
             new_regularisation_weights,
             new_relations,
-            rng_state,
+            seed_triplet,
             lambda_=self.alignment_regularisation,
         )

umap/layouts.py

@@ -722,6 +722,7 @@ def _optimize_layout_aligned_euclidean_single_epoch(
             max_n_edges = e_p_s.shape[0]
 
     embedding_order = np.arange(n_embeddings).astype(np.int32)
+    np.random.seed(abs(rng_state[0]))
     np.random.shuffle(embedding_order)
 
     for i in range(max_n_edges):