implemented grad norm penalty+moved logs to dataset

lib/audio.py

@@ -27,6 +27,7 @@ AUTOTUNE_NOTES = np.array([
     2093.00, 2217.46, 2349.32, 2489.02, 2637.02, 2793.83,
     2959.96, 3135.96, 3322.44, 3520.00, 3729.31, 3951.07
 ])
+SR_MAP = {"32k": 32000,"40k": 40000, "48k": 48000}
 
 class AudioProcessor:
     def __init__(self, normalize=True, threshold_silence=True, dynamic_threshold=True, sample_size=16000, multiplier=2.0, fill_method="median", kernel_size=5, silence_threshold_db=-50, normalize_threshold_db=-1):

lib/train/utils.py

@@ -382,6 +382,10 @@ def get_hparams(init=True):
     hparams.if_cache_data_in_gpu = args.if_cache_data_in_gpu
     hparams.data.training_files = os.path.join(experiment_dir,"filelist.txt")
     hparams.save_best_model = True
+    hparams.train.gradient_lambda = 0.
+    hparams.train.num_workers = 4
+    hparams.train.timbre_alpha = 0.
+    hparams.model_path = None
     return hparams
 
 
@@ -476,6 +480,10 @@ class HParams:
     def __repr__(self):
         return self.__dict__.__repr__()
     
+    def get(self, key, default=None):
+        try: return self[key]
+        except: return default
+    
     def sync_log_interval(self, dataset_length):
         log_interval = getattr(self,"log_every_epoch",1)
-        self.train.log_interval = int(self.train.batch_size*dataset_length*log_interval)
+        self.train.log_interval = int(dataset_length*log_interval)