Merge branch 'vijay_gpt3_pairwise_constraints' into vijay_refactor_se…

…mi_clustering

active_semi_clustering/semi_supervised/labeled_data/kmeans.py

@@ -17,7 +17,7 @@
 
 
 class KMeans:
-    def __init__(self, n_clusters=3, max_iter=100, num_reinit=1, normalize_vectors=False, split_normalization=False, init="random", verbose=False):
+    def __init__(self, n_clusters=3, max_iter=100, num_reinit=1, normalize_vectors=False, split_normalization=False, init="random", split_point=300, verbose=False):
         self.n_clusters = n_clusters
         self.max_iter = max_iter
         self.normalize_vectors = normalize_vectors
@@ -26,6 +26,7 @@ def __init__(self, n_clusters=3, max_iter=100, num_reinit=1, normalize_vectors=F
         self.init = init
         self.verbose = verbose
         self.num_reinit = num_reinit
+        self.split_point = split_point
 
     def fit(self, X, y=None, **kwargs):
         # Initialize cluster centers
@@ -51,8 +52,8 @@ def fit(self, X, y=None, **kwargs):
                 timer = time.perf_counter()
                 if self.normalize_vectors:
                     if self.split_normalization:
-                        kg_centers = normalize(cluster_centers[:, :300], axis=1, norm="l2")
-                        bert_centers = normalize(cluster_centers[:, 300:], axis=1, norm="l2")
+                        kg_centers = normalize(cluster_centers[:, :self.split_point], axis=1, norm="l2")
+                        bert_centers = normalize(cluster_centers[:, self.split_point:], axis=1, norm="l2")
                         cluster_centers = np.hstack([kg_centers, bert_centers])
                     else:
                         cluster_centers = normalize(cluster_centers, axis=1, norm="l2")

active_semi_clustering/semi_supervised/pairwise_constraints/kmeans_corrector.py

@@ -5,13 +5,15 @@
 from sklearn.preprocessing import normalize
 from tqdm import tqdm
 
+from active_semi_clustering.active.pairwise_constraints import ExampleOracle
 from active_semi_clustering.active.pairwise_constraints.example_oracle import MaximumQueriesExceeded
 
 class KMeansCorrection:
-    def __init__(self, oracle, cluster_predictions, cluster_centers):
+    def __init__(self, oracle, cluster_predictions, cluster_centers, labels):
         self.oracle = oracle
         self.cluster_predictions = cluster_predictions
         self.cluster_centers = cluster_centers
+        self.labels = labels
 
     @staticmethod
     def normalize_features(features):
@@ -47,7 +49,84 @@ def fit(self, X, num_corrections):
 
         corrected_labels = copy.deepcopy(self.cluster_predictions)
 
+        example_oracle = ExampleOracle(self.labels, max_queries_cnt=100000)
+        top_cluster_labels = []
+        for i, ent_idx in tqdm(enumerate(closest_top_two)):
+            correct_found = False
+            top_clust = ambiguous_points_top_five_clust_idxs[i][0]
+            top_cluster_rep_points = representative_points[top_clust]
+            top_cluster_bad = False
+            for top_cluster_rep in top_cluster_rep_points:
+                top_cluster_class = example_oracle.query(ent_idx, top_cluster_rep)
+                if top_cluster_class is False:
+                    top_cluster_bad = True
+                    break
+            if not top_cluster_bad:
+                top_cluster_labels.append(0)
+                continue
+            for i, next_best_clust in enumerate(ambiguous_points_top_five_clust_idxs[i][1:]):
+                next_best_cluster_rep_points = representative_points[next_best_clust]
+                make_correction = False
+                for next_best_cluster_rep in next_best_cluster_rep_points:
+                    next_cluster_class = example_oracle.query(ent_idx, next_best_cluster_rep) 
+                    if top_cluster_bad and next_cluster_class is True:
+                        make_correction = True
+                        break
+                if make_correction:
+                    top_cluster_labels.append(i+1)
+                    correct_found = True
+                    break
+            if not correct_found:
+                top_cluster_labels.append(-1)
+
+
+        top_cluster_predictions = []
+        for i, ent_idx in tqdm(enumerate(closest_top_two)):
+            correct_found = False
+            top_clust = ambiguous_points_top_five_clust_idxs[i][0]
+            top_cluster_rep_points = representative_points[top_clust]
+            top_cluster_bad = False
+            for top_cluster_rep in top_cluster_rep_points:
+                top_cluster_class = self.oracle.query(ent_idx, top_cluster_rep)
+                if top_cluster_class is False:
+                    top_cluster_bad = True
+                    break
+            if not top_cluster_bad:
+                top_cluster_predictions.append(0)
+                continue
+            for i, next_best_clust in enumerate(ambiguous_points_top_five_clust_idxs[i][1:]):
+                next_best_cluster_rep_points = representative_points[next_best_clust]
+                make_correction = False
+                for next_best_cluster_rep in next_best_cluster_rep_points:
+                    next_cluster_class = self.oracle.query(ent_idx, next_best_cluster_rep) 
+                    if top_cluster_bad and next_cluster_class is True:
+                        make_correction = True
+                        break
+                if make_correction:
+                    top_cluster_predictions.append(i+1)
+                    correct_found = True
+                    break
+            if not correct_found:
+                top_cluster_predictions.append(-1)
+
+        num_gpt_corrections_false_positive =  len([x for i, x in enumerate(top_cluster_predictions) if top_cluster_predictions[i] > 0 and top_cluster_labels[i] == 0 ]); num_gpt_corrections_false_negative =  len([x for i, x in enumerate(top_cluster_predictions) if top_cluster_predictions[i] == 0 and top_cluster_labels[i] > 0 ]); num_gpt_corrections_true_positive =  len([x for i, x in enumerate(top_cluster_predictions) if top_cluster_predictions[i] > 0 and top_cluster_labels[i] == top_cluster_predictions[i] ]); num_gpt_corrections_true_negative =  len([x for i, x in enumerate(top_cluster_predictions) if top_cluster_predictions[i] == 0 and top_cluster_labels[i] == top_cluster_predictions[i] ])
+
+
+        num_gpt_corrections_false_positive =  len([x for i, x in enumerate(top_cluster_predictions) if top_cluster_predictions[i] > 0 and top_cluster_labels[i] == 0 ])
+        num_gpt_corrections_false_negative =  len([x for i, x in enumerate(top_cluster_predictions) if top_cluster_predictions[i] == 0 and top_cluster_labels[i] > 0 ])
+        num_gpt_corrections_true_positive =  len([x for i, x in enumerate(top_cluster_predictions) if top_cluster_predictions[i] > 0 and top_cluster_labels[i] == top_cluster_predictions[i] ])
+        num_gpt_corrections_true_negative =  len([x for i, x in enumerate(top_cluster_predictions) if top_cluster_predictions[i] == 0 and top_cluster_labels[i] == top_cluster_predictions[i] ])
+
+        print(f"num_gpt_corrections_false_positive: {num_gpt_corrections_false_positive}")
+        print(f"num_gpt_corrections_false_negative: {num_gpt_corrections_false_negative}")
+        print(f"num_gpt_corrections_true_positive: {num_gpt_corrections_true_positive}")
+        print(f"num_gpt_corrections_true_negative: {num_gpt_corrections_true_negative}")
+
+
+        kmeans_corrections = []
+
         num_corrections_made = 0
+        num_correct_corrections_made = 0
         num_queries = 0
         try:
             for i, ent_idx in tqdm(enumerate(closest_top_two)):
@@ -69,17 +148,35 @@ def fit(self, X, num_corrections):
                         if top_cluster_bad and next_cluster_class is True:
                             make_correction = True
                             break
+
+                    gt_make_correction = False
+                    for next_best_cluster_rep in next_best_cluster_rep_points:
+                        next_cluster_class = example_oracle.query(ent_idx, next_best_cluster_rep) 
+                        num_queries += 1
+                        if top_cluster_bad and next_cluster_class is True:
+                            gt_make_correction = True
+                            break
+
                     if make_correction == True:
                         num_corrections_made += 1
+                        if gt_make_correction:
+                            num_correct_corrections_made += 1
+                        kmeans_corrections.append({"ent_idx": int(ent_idx), 
+                                                   "previous_label": int(corrected_labels[ent_idx]),
+                                                   "corrected_label": int(next_best_clust)})
                         corrected_labels[ent_idx] = next_best_clust
+
                         break
         except MaximumQueriesExceeded:
             pass
 
         self.labels_ = corrected_labels
 
+        breakpoint()
+
         print(f"Num oracle queries: {num_queries}")
         print(f"Num Corrections: {num_corrections_made}")
+        print(f"Num Correct Corrections: {num_correct_corrections_made}")
 
         '''
         from active_semi_clustering.active.pairwise_constraints import ExampleOracle

active_semi_clustering/semi_supervised/pairwise_constraints/pckmeans.py

@@ -37,7 +37,7 @@ def fit(self, X, y=None, ml=[], cl=[]):
         # Initialize centroids
         start = time.perf_counter()
         cluster_centers = self._init_cluster_centers(X)
-        #cluster_centers = self._initialize_cluster_centers(X, neighborhoods)
+        # cluster_centers = self._initialize_cluster_centers(X, neighborhoods)
         elapsed = time.perf_counter() - start
         print(f"Initializing neighborhoods took {round(elapsed, 4)} seconds")