Modified predict to compute representations just once; added predict_score with same input as predict_rank, tests

Author: paoloRais

lightfm/_lightfm_fast.pyx.template

@@ -1139,51 +1139,129 @@ def fit_bpr(CSRMatrix item_features,
                user_alpha)
 
 
-def predict_lightfm(CSRMatrix item_features,
-                    CSRMatrix user_features,
-                    int[::1] user_ids,
-                    int[::1] item_ids,
-                    double[::1] predictions,
-                    FastLightFM lightfm,
-                    int num_threads):
+cdef precompute_unique(CSRMatrix item_features,
+                       CSRMatrix user_features,
+                       int[::1] unique_users,
+                       int[::1] unique_items,
+                       flt *user_reprs,
+                       flt *it_reprs,
+                       FastLightFM lightfm,
+                       int num_threads):
     """
-    Generate predictions.
+    Precomputes the representations for all the users in unique_users and
+    all the items in unique_items
     """
-
-    cdef int i, no_examples
-    cdef flt *user_repr
+    cdef int i, j
     cdef flt *it_repr
+    cdef flt *user_repr
+    cdef int no_features
+    cdef int no_users
 
-    no_examples = predictions.shape[0]
-
+    no_features = unique_items.shape[0]
+    no_users = unique_users.shape[0]
     {nogil_block}
-
         user_repr = <flt *>malloc(sizeof(flt) * (lightfm.no_components + 1))
         it_repr = <flt *>malloc(sizeof(flt) * (lightfm.no_components + 1))
-
-        for i in {range_block}(no_examples):
-
+        # users representations
+        for i in {range_block}(no_users):
             compute_representation(user_features,
                                    lightfm.user_features,
                                    lightfm.user_biases,
                                    lightfm,
-                                   user_ids[i],
+                                   unique_users[i],
                                    lightfm.user_scale,
                                    user_repr)
+            for j in {range_block}(lightfm.no_components + 1):
+                user_reprs[i * (lightfm.no_components + 1) + j] = user_repr[j]
+
+        # items representations
+        for i in {range_block}(no_features):
             compute_representation(item_features,
                                    lightfm.item_features,
                                    lightfm.item_biases,
                                    lightfm,
-                                   item_ids[i],
+                                   unique_items[i],
                                    lightfm.item_scale,
                                    it_repr)
+            for j in {range_block}(lightfm.no_components + 1):
+                it_reprs[i * (lightfm.no_components + 1) + j] = it_repr[j]
+
+
+def predict_lightfm(CSRMatrix item_features,
+                    CSRMatrix user_features,
+                    int[::1] user_ids,
+                    int[::1] item_ids,
+                    double[::1] predictions,
+                    FastLightFM lightfm,
+                    int num_threads,
+                    bint precompute):
+    """
+    Generate predictions.
+    """
+    cdef int i, j, no_examples
+    cdef flt *user_repr
+    cdef flt *it_repr
+    cdef flt *user_reprs
+    cdef flt *it_reprs
+    cdef int[::1] unique_users
+    cdef int[::1] unique_items
+    cdef long[::1] inverse_users
+    cdef long[::1] inverse_items
+    cdef int no_features
+    cdef int no_users
+
+    no_examples = predictions.shape[0]
+
+    if precompute:
+        unique_users, inverse_users = np.unique(user_ids, return_inverse=True)
+        unique_items, inverse_items = np.unique(item_ids, return_inverse=True)
+        no_features = unique_items.shape[0]
+        no_users = unique_users.shape[0]
+
+        user_reprs = <flt *>malloc(sizeof(flt) * no_users * (lightfm.no_components + 1))
+        it_reprs = <flt *>malloc(sizeof(flt) * no_features *(lightfm.no_components + 1))
+        precompute_unique(item_features,
+                          user_features,
+                          unique_users,
+                          unique_items,
+                          user_reprs,
+                          it_reprs,
+                          lightfm,
+                          num_threads)
+
+    {nogil_block}
+        user_repr = <flt *>malloc(sizeof(flt) * (lightfm.no_components + 1))
+        it_repr = <flt *>malloc(sizeof(flt) * (lightfm.no_components + 1))
+        for i in {range_block}(no_examples):
+            if precompute:
+                for j in {range_block}(lightfm.no_components + 1):
+                    user_repr[j] = user_reprs[inverse_users[i] * (lightfm.no_components + 1) + j]
+                    it_repr[j] = it_reprs[inverse_items[i] * (lightfm.no_components + 1) + j]
+            else:
+                compute_representation(user_features,
+                                       lightfm.user_features,
+                                       lightfm.user_biases,
+                                       lightfm,
+                                       user_ids[i],
+                                       lightfm.user_scale,
+                                       user_repr)
+                compute_representation(item_features,
+                                       lightfm.item_features,
+                                       lightfm.item_biases,
+                                       lightfm,
+                                       item_ids[i],
+                                       lightfm.item_scale,
+                                       it_repr)
 
             predictions[i] = compute_prediction_from_repr(user_repr,
-                                                          it_repr,
-                                                          lightfm.no_components)
+                                                              it_repr,
+                                                              lightfm.no_components)
 
         free(user_repr)
         free(it_repr)
+    if precompute:
+        free(user_reprs)
+        free(it_reprs)
 
 
 def predict_ranks(CSRMatrix item_features,
@@ -1341,3 +1419,4 @@ def __test_in_positives(int row, int col, CSRMatrix mat):
         return True
     else:
         return False
+