sort each posting-list by the distance to centroid

Signed-off-by: Keming <kemingy94@gmail.com>

Author: kemingy

src/rabitq.rs

@@ -176,8 +176,8 @@ impl RaBitQ {
             if i % 5000 == 0 {
                 debug!("\t> preprocessing {}...", i);
             }
-            let min_label = kmeans_nearest_cluster(&centroids.as_ref(), &xp);
-            labels[min_label].push(i as u32);
+            let (min_label, min_dist) = kmeans_nearest_cluster(&centroids.as_ref(), &xp);
+            labels[min_label].push((i as u32, min_dist));
             let x_c_quantized = xp - centroids.col(min_label);
             x_c_distance[i] = x_c_quantized.norm_l2();
             factors[i].center_distance_square = x_c_distance[i].powi(2);
@@ -205,6 +205,13 @@ impl RaBitQ {
         }
 
         // sort by labels
+        let labels = labels
+            .into_iter()
+            .map(|mut v| {
+                v.sort_by(|a, b| a.1.partial_cmp(&b.1).unwrap());
+                v.into_iter().map(|(i, _)| i).collect::<Vec<_>>()
+            })
+            .collect::<Vec<_>>();
         debug!("sort by labels...");
         let mut offsets = vec![0; k + 1];
         for i in 0..k {

src/utils.rs

@@ -248,7 +248,7 @@ pub fn project(vec: &ColRef<f32>, orthogonal: &MatRef<f32>) -> Col<f32> {
 }
 
 /// Find the nearest cluster for the given vector.
-pub fn kmeans_nearest_cluster(centroids: &MatRef<f32>, vec: &ColRef<f32>) -> usize {
+pub fn kmeans_nearest_cluster(centroids: &MatRef<f32>, vec: &ColRef<f32>) -> (usize, f32) {
     let mut min_dist = f32::MAX;
     let mut min_label = 0;
     for (j, centroid) in centroids.col_iter().enumerate() {
@@ -258,7 +258,7 @@ pub fn kmeans_nearest_cluster(centroids: &MatRef<f32>, vec: &ColRef<f32>) -> usi
             min_label = j;
         }
     }
-    min_label
+    (min_label, min_dist)
 }
 
 /// Read the fvces/ivces file.