Merge pull request #26 from ktonal/develop

v0.2.5

mimikit/__init__.py

@@ -1,4 +1,4 @@
-__version__ = '0.2.4'
+__version__ = '0.2.5'
 
 from . import audios
 from . import connectors

mimikit/data/feature.py

@@ -89,8 +89,8 @@ def encode(self, inputs):
         -------
 
         """
-        if hasattr(super(), 'encoders'):
-            inputs = super().encoders[type(inputs)](inputs)
+        if hasattr(super(type(self), self), 'encoders'):
+            inputs = super(type(self), self).encoders[type(inputs)](inputs)
         return self.encoders[type(inputs)](inputs)
 
     def decode(self, inputs):
@@ -106,8 +106,8 @@ def decode(self, inputs):
 
         """
         inputs = self.decoders[type(inputs)](inputs)
-        if hasattr(super(), 'decoders'):
-            inputs = super().decoders[type(inputs)](inputs)
+        if hasattr(super(type(self), self), 'decoders'):
+            inputs = super(type(self), self).decoders[type(inputs)](inputs)
         return inputs
 
     def load(self, path):

mimikit/models/s2s_lstm.py

@@ -105,7 +105,7 @@ def demo():
     # how many examples are used pro training steps
     batch_size = 16
     # the learning rate
-    max_lr = 1e-3
+    max_lr = 3e-4
     # betas control how fast the network changes its 'learning course'.
     # generally, betas should be close but smaller than 1. and be balanced with the batch_size :
     # the smaller the batch, the higher the betas 'could be'.

mimikit/models/sample_rnn.py

@@ -27,7 +27,7 @@ class SampleRNNData(IData):
 
     @classmethod
     def schema(cls, sr=22050, emphasis=0., q_levels=256):
-        return {'qx': MuLawSignal(sr=sr, emphasis=emphasis, q_levels=q_levels)}
+        return {'qx': MuLawSignal(sr=sr, emphasis=emphasis, q_levels=q_levels, normalize=True)}
 
     def batch_signature(self, stage='fit'):
         batch_seq_len, frame_sizes = tuple(getattr(self, key) for key in ["batch_seq_len", "frame_sizes"])