use silouette method for scoring

1 files changed, 15 insertions, 17 deletions

diff --git a/main.c b/main.c
index 5629cd5..ed46af7 100644
--- a/main.c
+++ b/main.c
@@ -13,7 +13,7 @@
 #include "km.h"
 
 #define MAX_CLUSTERS 10
-#define NUM_TESTS 100
+#define NUM_TESTS 300
 #define MAX_BEST 10
 
 #define IM_WIDTH 128
@@ -34,9 +34,9 @@ typedef struct {
 
   struct {
     float distance,
-          variance,
+          silouette,
           cluster_size;
-    size_t num_empty_clusters;
+    size_t num_empty;
   } rows[MAX_CLUSTERS - 2];
 
   // best clusters
@@ -62,7 +62,7 @@ ctx_best_sort(
   // sort best sets by ascending score (worst to best)
   qsort(
     ctx->best,
-    (ctx->num_best % MAX_BEST),
+    MIN(ctx->num_best, MAX_BEST),
     sizeof(best_item_t),
     best_score_cmp
   );
@@ -162,10 +162,10 @@ find_on_data(
   ctx_t * const ctx = cb_data;
   const size_t ofs = data->num_clusters - 2;
 
-  ctx->rows[ofs].distance += data->mean_distance;
-  ctx->rows[ofs].variance += data->mean_variance;
+  ctx->rows[ofs].distance += data->distance_sum;
+  ctx->rows[ofs].silouette += data->silouette;
   ctx->rows[ofs].cluster_size += data->mean_cluster_size;
-  ctx->rows[ofs].num_empty_clusters += data->num_empty_clusters;
+  ctx->rows[ofs].num_empty += data->num_empty_clusters;
 }
 
 static bool
@@ -184,7 +184,6 @@ find_on_best(
 
   if (ctx->num_best >= MAX_BEST) {
     // finalize old best data set
-    // D("finalizing old best %zu", ctx->num_best);
     km_set_fini(dst);
   }
 
@@ -220,17 +219,15 @@ ctx_csv_print_row(
 ) {
   const size_t num_clusters = i + 2;
   const float mean_distance = ctx->rows[i].distance / NUM_TESTS,
-              mean_variance = ctx->rows[i].variance / NUM_TESTS,
               mean_cluster_size = ctx->rows[i].cluster_size / NUM_TESTS,
-              mean_empty = 1.0 * ctx->rows[i].num_empty_clusters / NUM_TESTS,
-              score = km_score(mean_distance, mean_empty);
+              mean_empty = 1.0 * ctx->rows[i].num_empty / NUM_TESTS,
+              score = ctx->rows[i].silouette / NUM_TESTS;
 
   // print result
-  fprintf(fh, "%zu,%0.3f,%0.3f,%0.3f,%0.3f,%0.3f\n",
+  fprintf(fh, "%zu,%0.3f,%0.3f,%0.3f,%0.3f\n",
     num_clusters,
     score,
     mean_distance,
-    mean_variance,
     mean_cluster_size,
     mean_empty
   );
@@ -242,7 +239,7 @@ ctx_csv_print(
   FILE * const fh
 ) {
   // print headers
-  fprintf(fh, "#,score,distance,variance,cluster_size,empty_clusters\n");
+  fprintf(fh, "#,score,distance,size,empty\n");
 
   // print rows
   for (size_t i = 0; i < MAX_CLUSTERS - 2; i++) {
@@ -260,12 +257,13 @@ save_on_best(
   const float score,
   void * const cb_data
 ) {
+  const ctx_t * const ctx = cb_data;
   UNUSED(score);
-  UNUSED(cb_data);
 
   // convert rank to channel brightness
   const uint8_t ch = 0x33 + (0xff - 0x33) * (1.0 * rank + 1) / (MAX_BEST);
-  const uint32_t color = (ch & 0xff) << 16;
+  const uint8_t shift = (rank == MIN(ctx->num_best, MAX_BEST) - 1) ? 8 : 16;
+  const uint32_t color = (ch & 0xff) << shift;
   // const uint32_t color = 0xff0000;
   // D("rank = %zu, score = %0.3f, size = %zu, color = %06x", rank, score, set->num_rows, color);
 
@@ -289,7 +287,7 @@ ctx_save_png(
   }
 
   // draw best cluster points
-  ctx_best_each(ctx, save_on_best, NULL);
+  ctx_best_each(ctx, save_on_best, (void*) ctx);
 
   // save png
   if (!stbi_write_png(png_path, IM_WIDTH, IM_HEIGHT, 3, im_data, IM_STRIDE)) {