Backing up decent results

.gitignore

@@ -16,3 +16,4 @@ tacotron2
 data
 SV2TTS/datasets/temp
 SV2TTS/_old/
+SV2TTS/saved_models/*_bak_*.pt
\ No newline at end of file

SV2TTS/data_objects/speaker_verification_dataset.py

@@ -38,7 +38,7 @@ class SpeakerVerificationDataset(Dataset):
         params = OrderedDict([
             ("Total speakers", len(self.speakers)),
             ("Average utterances per speaker", self.mean_n_utterances),
-            ("Datasets", ','.join(self.datasets)),
+            ("Datasets", ', '.join(self.datasets)),
         ])
         return params
     

SV2TTS/model.py

@@ -97,20 +97,20 @@ class SpeakerEncoder(nn.Module):
         sim_matrix = sim_matrix * self.similarity_weight + self.similarity_bias
 
         # Loss
-        ground_truth = torch.from_numpy(
-            np.repeat(np.arange(speakers_per_batch), utterances_per_speaker)
-        ).long()
-        loss = self.loss_fn(sim_matrix, ground_truth)
+        ground_truth = np.repeat(np.arange(speakers_per_batch), utterances_per_speaker)
+        loss = self.loss_fn(sim_matrix, torch.from_numpy(ground_truth).long())
         
         # EER (not backpropagated)
+        sim_matrix = sim_matrix.detach().numpy()
         with torch.no_grad():
+            ## Imabalanced EER
             inv_argmax = lambda i: np.eye(1, speakers_per_batch, i, dtype=np.int)[0]
             labels = np.array([inv_argmax(i) for i in ground_truth])
-            preds = sim_matrix.detach().numpy()
+            preds = sim_matrix
 
             # Snippet from https://yangcha.github.io/EER-ROC/
             fpr, tpr, thresholds = roc_curve(labels.flatten(), preds.flatten())
             eer = brentq(lambda x: 1. - x - interp1d(fpr, tpr)(x), 0., 1.)
             # thresh = interp1d(fpr, thresholds)(eer)
-        
+            
         return loss, eer
\ No newline at end of file

SV2TTS/params_model.py

@@ -6,7 +6,7 @@ model_num_layers = 3
 
 
 ## Training parameters
-learning_rate_init = 2e-5   
+learning_rate_init = 1e-4 #2e-5   
 # exponential_decay_beta = 0.9998
-speakers_per_batch = 64
+speakers_per_batch = 32
 utterances_per_speaker = 10
\ No newline at end of file