Experiments on ecapa with different audio streams and aug. models

nauman-daw · nauman-daw · commit 37ce61368028 · 2021-04-03T15:03:51.000-04:00
diff --git a/recipes/AMI/Diarization/experiment_beam.py b/recipes/AMI/Diarization/experiment_beam.py
@@ -276,6 +276,7 @@ def diarize_dataset(full_csv, split_type, n_lambdas, pval, n_neighbors=10):
                 params["affinity"],
                 n_neighbors,
             )
+        # Maybe used for AHC
         if params["backend"] == "AHC":
             # call AHC
             threshold = pval  # TODO: have to update the calling function
@@ -625,6 +626,11 @@ def check_dirs():
             "Tuning for p-value for SC (Multiple iterations over AMI Dev set)"
         )
         best_pval = dev_p_tuner(full_csv, "dev")
+
+    elif params["backend"] == "AHC":
+        logger.info("Tuning for threshold-value for AHC")
+        best_threshold = dev_threshold_tuner(full_csv, "dev")
+        best_pval = best_threshold
     else:
         # This part (NN for unknown num of speakers) is WIP
         if params["oracle_n_spkrs"] is False:
diff --git a/speechbrain/processing/diarization.py b/speechbrain/processing/diarization.py
@@ -1147,13 +1147,39 @@ def do_AHC(diary_obj, out_rttm_file, rec_id, k_oracle=4, p_val=0.3):
 
     # p_val --> threshold_val (for AHC)
 
+    # Mean centering and normalizing the embeddings
+    # mu = diary_obj.get_mean_stat1()
+    # print ("mu = " , mu)
+    # print ("\n\n\n\n")
+    # print("stat0: ", diary_obj.stat0)
+    # diary_obj.center_stat1(mu)
+    diary_obj.norm_stat1()
+
+    import fastcluster
+    from scipy.cluster.hierarchy import fcluster
+
+    linkage_matrix = fastcluster.linkage(
+        diary_obj.stat1, method="ward", metric="cosine"
+    )
+
+    # processing
     if k_oracle is not None:
         print("ORACLE SPKRs...")
         num_of_spk = k_oracle
-        clustering = AgglomerativeClustering(
-            n_clusters=num_of_spk, linkage="ward"
-        ).fit(diary_obj.stat1)
-        labels = clustering.labels_
+
+        predicted_label = (
+            fcluster(linkage_matrix, num_of_spk, criterion="maxclust") - 1
+        )
+
+        # clustering = AgglomerativeClustering(
+        #    n_clusters=num_of_spk,
+        #    affinity="cosine",
+        #    linkage="ward",
+        # ).fit(diary_obj.stat1)
+        # labels = clustering.labels_
+
+        labels = predicted_label
+
         print("labels.shape (Oracle) = ", labels.shape)
     else:
         print("Using AHC threshold pval = ", p_val)
