cleanup

TTS/bin/train_encoder.py

@@ -42,7 +42,8 @@ def setup_loader(ap, is_val=False, verbose=False):
         dataset = MyDataset(ap,
                             meta_data_eval if is_val else meta_data_train,
                             voice_len=1.6,
-                            num_utter_per_speaker=10,
+                            num_utter_per_speaker=c.num_utters_per_speaker,
+                            num_speakers_in_batch=c.num_speakers_in_batch,
                             skip_speakers=False,
                             storage_size=c.storage["storage_size"],
                             sample_from_storage_p=c.storage["sample_from_storage_p"],
@@ -98,11 +99,10 @@ def train(model, criterion, optimizer, scheduler, ap, global_step):
         epoch_time += step_time
 
         # Averaged Loss and Averaged Loader Time
-        dataset_number_prefetched = 2 * c.num_loader_workers  # this is hardcoded in pytorch
         avg_loss = 0.01 * loss.item() \
                    + 0.99 * avg_loss if avg_loss != 0 else loss.item()
-        avg_loader_time = 1/dataset_number_prefetched * loader_time\
+        avg_loader_time = 1/c.num_loader_workers * loader_time + \
-                          + (dataset_number_prefetched-1) / dataset_number_prefetched * avg_loader_time if avg_loader_time != 0 else loader_time
+                          (c.num_loader_workers-1) / c.num_loader_workers * avg_loader_time if avg_loader_time != 0 else loader_time
         current_lr = optimizer.param_groups[0]['lr']
 
         if global_step % c.steps_plot_stats == 0:

TTS/speaker_encoder/config.json

@@ -36,7 +36,8 @@
     "tb_model_param_stats": false, // true, plots param stats per layer on tensorboard. Might be memory consuming, but good for debugging. 
     "steps_plot_stats": 10, // number of steps to plot embeddings.
     "num_speakers_in_batch": 64, // Batch size for training. Lower values than 32 might cause hard to learn attention. It is overwritten by 'gradual_training'.
-    "num_loader_workers": 4,        // number of training data loader processes. Don't set it too big. 4-8 are good values.
+    "num_utters_per_speaker": 10,  //
+    "num_loader_workers": 8,        // number of training data loader processes. Don't set it too big. 4-8 are good values.
     "wd": 0.000001, // Weight decay weight.
     "checkpoint": true, // If true, it saves checkpoints per "save_step"
     "save_step": 1000, // Number of training steps expected to save traning stats and checkpoints.
@@ -50,8 +51,8 @@
         "use_lstm_with_projection": true
     },
     "storage": {
-        "sample_from_storage_p": 0.42,  // the probability with which we'll sample from the DataSet in-memory storage
+        "sample_from_storage_p": 0.66,  // the probability with which we'll sample from the DataSet in-memory storage
-        "storage_size": 5,   // the size of the in-memory storage with respect to a single batch
+        "storage_size": 15,   // the size of the in-memory storage with respect to a single batch
         "additive_noise": 1e-5   // add very small gaussian noise to the data in order to increase robustness
     },
     "datasets": 

TTS/speaker_encoder/dataset.py

@@ -33,8 +33,10 @@ class MyDataset(Dataset):
         self.additive_noise = float(additive_noise)
         if self.verbose:
             print("\n > DataLoader initialization")
+            print(f" | > Speakers per Batch: {num_speakers_in_batch}")
             print(f" | > Storage Size: {self.storage.maxsize} speakers, each with {num_utter_per_speaker} utters")
             print(f" | > Sample_from_storage_p : {self.sample_from_storage_p}")
+            print(f" | > Noise added : {self.additive_noise}")
             print(f" | > Number of instances : {len(self.items)}")
             print(f" | > Sequence length: {self.seq_len}")
             print(f" | > Num speakers: {len(self.speakers)}")

TTS/tts/utils/generic_utils.py

@@ -7,11 +7,9 @@ from TTS.utils.generic_utils import check_argument
 
 
 def split_dataset(items):
-    is_multi_speaker = False
     speakers = [item[-1] for item in items]
     is_multi_speaker = len(set(speakers)) > 1
-    eval_split_size = 500 if len(items) * 0.01 > 500 else int(
+    eval_split_size = min(500, int(len(items) * 0.01))
-        len(items) * 0.01)
     assert eval_split_size > 0, " [!] You do not have enough samples to train. You need at least 100 samples."
     np.random.seed(0)
     np.random.shuffle(items)
@@ -142,6 +140,11 @@ def check_config(c):
     check_argument('do_trim_silence', c['audio'], restricted=True, val_type=bool)
     check_argument('trim_db', c['audio'], restricted=True, val_type=int)
 
+    # storage parameters
+    check_argument('sample_from_storage_p', c['storage'], restricted=True, val_type=float, min_val=0.0, max_val=1.0)
+    check_argument('storage_size', c['storage'], restricted=True, val_type=int, min_val=1, max_val=100)
+    check_argument('additive_noise', c['storage'], restricted=True, val_type=float, min_val=0.0, max_val=1.0)
+
     # training parameters
     check_argument('batch_size', c, restricted=True, val_type=int, min_val=1)
     check_argument('eval_batch_size', c, restricted=True, val_type=int, min_val=1)