check user and item matrix for nan entries also in bpr gpu version (#731)

Co-authored-by: Jan Morlock <jan.morlock@1und1.de>
Co-authored-by: Ben Frederickson <github@benfrederickson.com>

Author: 3 people

implicit/cpu/matrix_factorization_base.py

@@ -5,7 +5,7 @@
 import numpy as np
 from scipy.sparse import csr_matrix, lil_matrix
 
-from ..recommender_base import ModelFitError, RecommenderBase
+from ..recommender_base import RecommenderBase
 from .topk import topk
 
 
@@ -247,10 +247,7 @@ def item_norms(self):
         return self._item_norms
 
     def _check_fit_errors(self):
-        is_nan = np.any(np.isnan(self.user_factors), axis=None)
-        is_nan |= np.any(np.isnan(self.item_factors), axis=None)
-        if is_nan:
-            raise ModelFitError("NaN encountered in factors")
+        self._check_factors(self.user_factors, self.item_factors)
 
 
 def _filter_items_from_sparse_matrix(items, query_items):

implicit/gpu/bpr.py

@@ -160,6 +160,8 @@ def fit(self, user_items, show_progress=True, callback=None):
                 if callback:
                     callback(_epoch, time.time() - s, correct, skipped)
 
+        self._check_fit_errors()
+
     def to_cpu(self) -> implicit.cpu.bpr.BayesianPersonalizedRanking:
         """Converts this model to an equivalent version running on the cpu"""
         ret = implicit.cpu.bpr.BayesianPersonalizedRanking(

implicit/gpu/matrix_factorization_base.py

@@ -201,6 +201,9 @@ def similar_items(
 
     similar_items.__doc__ = RecommenderBase.similar_items.__doc__
 
+    def _check_fit_errors(self):
+        self._check_factors(self.user_factors.to_numpy(), self.item_factors.to_numpy())
+
     def recalculate_user(self, userid, user_items):
         raise NotImplementedError("recalculate_user is not supported with this model")
 

implicit/recommender_base.py

@@ -214,3 +214,10 @@ def rank_items(self, userid, user_items, selected_items, recalculate_user=False)
             items=selected_items,
             filter_already_liked_items=False,
         )
+
+    @staticmethod
+    def _check_factors(user_factors, item_factors):
+        is_nan = np.any(np.isnan(user_factors), axis=None)
+        is_nan |= np.any(np.isnan(item_factors), axis=None)
+        if is_nan:
+            raise ModelFitError("NaN encountered in factors")