From 1c1f0f29a768a0519933f876ecc8a9494e8984a4 Mon Sep 17 00:00:00 2001
From: quillotm <mathias.quillot@univ-avignon.fr>
Date: Tue, 24 Aug 2021 09:39:35 +0200
Subject: [PATCH] Now we train n_init time with the basic multidistance
 implementation of k-means.

---
 volia/clustering_modules/kmeans_multidistance.py | 53 +++++++++++++++++++++---
 1 file changed, 48 insertions(+), 5 deletions(-)

diff --git a/volia/clustering_modules/kmeans_multidistance.py b/volia/clustering_modules/kmeans_multidistance.py
index f1c7944..47fe159 100644
--- a/volia/clustering_modules/kmeans_multidistance.py
+++ b/volia/clustering_modules/kmeans_multidistance.py
@@ -2,12 +2,18 @@
 import pickle
 from abstract_clustering import AbstractClustering
 from KMeans_Multidistance.KMeans_Class import KMeans
+from random import seed
+from random import random
+import numpy as np
+from sklearn.metrics import pairwise_distances
 
 class kmeansMultidistance():
     def __init__(self, distance="cosine"):
-        self.kmeans_model = None
+        self.kmeans_model = None # Best model
         self.centroids = None
         self.distance = distance
+        self.seed = None # Seed of the best
+        self.seeds = None
 
     def predict(self, features):
         """
@@ -48,7 +54,44 @@ class kmeansMultidistance():
         @param k:
         @return:
         """
-        model = KMeans(k=5, maxiter=maxiter, distance=self.distance, record_heterogeneity=[], verbose=True, seed=123)
-        centroids, _ = model.fit(features)
-        self.centroids = centroids
-        self.kmeans_model = model
\ No newline at end of file
+        # Initialization
+        self.kmeans_model = None
+        self.centroids = None
+        self.seed = None
+
+        # Compute seeds before using seeds
+        seed()
+        self.seeds = [random() for i in range(ninit)]
+
+        # Learning k-means model
+        results = []
+        for i in range(ninit):
+            model = KMeans(k=5,
+                           maxiter=maxiter,
+                           distance=self.distance,
+                           record_heterogeneity=[],
+                           verbose=True,
+                           seed=self.seeds[i])
+            centroids, closest_cluster = model.fit(features)
+
+            # Compute distance
+            kwds = {}
+            if self.distance == "mahalanobis":
+                VI = np.linalg.pinv(np.cov(features.T)).T
+                kwds = {"VI": VI}
+            distances = pairwise_distances(features, centroids, metric=self.distance, **kwds)
+
+            # Then compute the loss
+            loss = np.sum(distances[np.arange(len(distances)), closest_cluster])
+
+            results.append({
+                "model": model,
+                "centroids": centroids,
+                "seed": self.seeds[i],
+                "loss": loss
+            })
+        losses = [result["loss"] for result in results]
+        best = results[losses.index(min(losses))]
+        self.kmeans_model = results[best]["model"]
+        self.centroids = results[best]["centroids"]
+        self.seed = results[best]["seed"]
-- 
1.8.2.3