Merge pull request #298 from Renumics/fix/234-scalar-columns-not-embe…

…ddable

Fix checking scalar dtypes for data alignment before embedding

renumics/spotlight/backend/tasks/reduction.py

@@ -6,11 +6,10 @@
 
 import numpy as np
 import pandas as pd
-from sklearn import preprocessing
 
 from renumics.spotlight.dataset.exceptions import ColumnNotExistsError
 from renumics.spotlight.data_store import DataStore
-from renumics.spotlight.dtypes import is_category_dtype, is_embedding_dtype
+from renumics.spotlight import dtypes
 
 SEED = 42
 
@@ -27,6 +26,7 @@ def align_data(
     """
     Align data from table's columns, remove `NaN`'s.
     """
+    from sklearn import preprocessing
 
     if not column_names or not indices:
         return np.empty(0, np.float64), []
@@ -35,7 +35,7 @@ def align_data(
     for column_name in column_names:
         dtype = data_store.dtypes[column_name]
         column_values = data_store.get_converted_values(column_name, indices)
-        if is_embedding_dtype(dtype):
+        if dtypes.is_embedding_dtype(dtype):
             embedding_length = max(
                 0 if x is None else len(cast(np.ndarray, x)) for x in column_values
             )
@@ -49,17 +49,19 @@ def align_data(
                         ]
                     )
                 )
-        elif is_category_dtype(dtype):
+        elif dtypes.is_category_dtype(dtype):
             na_mask = np.array(column_values) == -1
             one_hot_values = preprocessing.label_binarize(
                 column_values, classes=sorted(set(column_values).difference({-1}))  # type: ignore
             ).astype(float)
             one_hot_values[na_mask] = np.nan
             aligned_values.append(one_hot_values)
-        elif dtype in (int, bool, float):
+        elif dtypes.is_scalar_dtype(dtype):
             aligned_values.append(np.array(column_values, dtype=float))
         else:
-            raise ColumnNotEmbeddable
+            raise ColumnNotEmbeddable(
+                f"Column '{column_name}' of type {dtype} is not embeddable."
+            )
 
     data = np.hstack([col.reshape((len(indices), -1)) for col in aligned_values])
     mask = ~pd.isna(data).any(axis=1)
@@ -118,7 +120,7 @@ def compute_pca(
 
     try:
         data, indices = align_data(data_store, column_names, indices)
-    except (ColumnNotExistsError, ValueError):
+    except (ColumnNotExistsError, ColumnNotEmbeddable):
         return np.empty(0, np.float64), []
     if data.size == 0:
         return np.empty(0, np.float64), []
@@ -129,5 +131,6 @@ def compute_pca(
     elif normalization == "robust standardize":
         data = preprocessing.RobustScaler(copy=False).fit_transform(data)
     reducer = decomposition.PCA(n_components=2, copy=False, random_state=SEED)
-    embeddings = reducer.fit_transform(data)
+    # `fit_transform` returns Fortran-ordered array.
+    embeddings = np.ascontiguousarray(reducer.fit_transform(data))
     return embeddings, indices