Cosine optimization

Fixed inner product ordering

apis/python/test/test_distance_metrics.py

@@ -60,16 +60,17 @@ def test_inner_product(tmp_path):
 
     inner_products_sklearn = np.dot(query_vectors, dataset_vectors.T)
 
-    sorted_inner_products_sklearn = np.sort(inner_products_sklearn, axis=1)
+    sorted_inner_products_sklearn = np.sort(inner_products_sklearn, axis=1)[:, ::-1]
 
     distances, _ = index.query(query_vectors, k=5)
 
-    sorted_distances = np.sort(distances, axis=1)
+    # multiply distances by -1 to get inner products, since library returns negative inner products
+    distances = -1 * distances
 
     for i in range(len(sorted_inner_products_sklearn)):
         compare = sorted_inner_products_sklearn[i][:5]
         assert np.allclose(
-            compare, sorted_distances[i], 1e-4
+            compare, distances[i], 1e-4
         ), "Inner products do not match"
 
 

src/include/scoring.h

@@ -260,22 +260,22 @@ struct inner_product_distance {
 #ifdef __AVX2__
   template <feature_vector V, feature_vector U>
   constexpr inline float operator()(const V& a, const U& b) const {
-    return avx2_inner_product(a, b);
+    return -avx2_inner_product(a, b);
   }
 
   template <feature_vector V>
   constexpr inline float operator()(const V& a) const {
-    return avx2_inner_product(a);
+    return -avx2_inner_product(a);
   }
 #else
   template <feature_vector V, feature_vector U>
   constexpr inline float operator()(const V& a, const U& b) const {
-    return unroll4_inner_product(a, b);
+    return -unroll4_inner_product(a, b);
   }
 
   template <feature_vector V>
   constexpr inline float operator()(const V& a) const {
-    return unroll4_inner_product(a);
+    return -unroll4_inner_product(a);
   }
 #endif
 };
@@ -296,7 +296,7 @@ struct cosine_distance {
   template <feature_vector V, feature_vector U>
   constexpr inline float operator()(const V& a, const U& b) const {
     return 1 -
-           inner_product(a, b) / (sqrt(l2_distance(a)) * sqrt(l2_distance(b)));
+           (-inner_product(a, b)) / sqrt(l2_distance(a) * l2_distance(b)); // our inner product is negative so we have to flip it
   }
 };