debug xvector extraction on a sliding window

nnet/xvector.py

@@ -314,7 +314,7 @@ class TrainingMonitor():
         """
         # TODO
         self.logger.critical(f"***Validation metrics - Cross validation accuracy = {self.val_acc[-1]} %, EER = {self.val_eer[-1] * 100} %")
-        self.logger.critical(f"***Test metrics - Test EER = {self.test_eer[-1] * 100} %")
+        #self.logger.critical(f"***Test metrics - Test EER = {self.test_eer[-1] * 100} %")
 
     def display_final(self):
         """
@@ -1673,7 +1673,7 @@ def cross_validation(model, validation_loader, device, validation_shape, tar_ind
     return (100. * accuracy.cpu().numpy() / validation_shape[0],
             loss.cpu().numpy() / ((batch_idx + 1) * batch_size),
             equal_error_rate)
-    return 0, 0, 0 
+    #return 0, 0, 0 
 
 
 def extract_embeddings(idmap_name,
@@ -1715,6 +1715,7 @@ def extract_embeddings(idmap_name,
         checkpoint = torch.load(model_filename, map_location=device)
         speaker_number = checkpoint["speaker_number"]
         model_opts = checkpoint["model_archi"]
+        model_opts["embedding_size"] = 256
         model = Xtractor(speaker_number,
                          model_archi=model_opts["model_type"],
                          loss=model_opts["loss"]["type"],
@@ -1770,11 +1771,11 @@ def extract_embeddings(idmap_name,
                 for td in tmp_data:
                     _, vec = model(x=td.to(device), is_eval=True, norm_embedding=norm_embeddings)
                     embed.append(vec.detach().cpu())
-
                 modelset.extend(mod * data.shape[0])
                 segset.extend(seg * data.shape[0])
                 if sliding_window:
-                    starts.extend(numpy.arange(start, start + data.shape[0] * win_shift, win_shift))
+                    tmp_start = numpy.arange(0, data.shape[0] * win_shift, win_shift)
+                    starts.extend(tmp_start * sample_rate + start.detach().cpu().numpy())
                 else:
                     starts.append(start)