tf.keras wrappers fixes (issue Rayhane-mamah#230)

hparams.py

@@ -153,9 +153,9 @@
 	cbhg_rnn_units = 128, #Number of GRU units used in bidirectional RNN of CBHG block. CBHG output is 2x rnn_units in shape
 
 	#Loss params
-	mask_encoder = False, #whether to mask encoder padding while computing attention. Set to True for better prosody but slower convergence.
+	mask_encoder = True, #whether to mask encoder padding while computing attention. Set to True for better prosody but slower convergence.
 	mask_decoder = False, #Whether to use loss mask for padded sequences (if False, <stop_token> loss function will not be weighted, else recommended pos_weight = 20)
-	cross_entropy_pos_weight = 1, #Use class weights to reduce the stop token classes imbalance (by adding more penalty on False Negatives (FN)) (1 = disabled)
+	cross_entropy_pos_weight = 20, #Use class weights to reduce the stop token classes imbalance (by adding more penalty on False Negatives (FN)) (1 = disabled)
 	predict_linear = True, #Whether to add a post-processing network to the Tacotron to predict linear spectrograms (True mode Not tested!!)
 	###########################################################################################################################################
 

wavenet_vocoder/models/modules.py

@@ -292,7 +292,10 @@ def call(self, inputs, incremental=False, convolution_queue=None):
 				output = tf.nn.bias_add(output, self.layer.bias)
 
 			#[batch_size, 1(time_step), channels(filters)]
-			return tf.reshape(output, [batch_size, 1, self.layer.filters]), convolution_queue
+			if convolution_queue is None:
+				return tf.reshape(output, [batch_size, 1, self.layer.filters])
+			else:
+				return [tf.reshape(output, [batch_size, 1, self.layer.filters]), convolution_queue]
 
 		#Normal run
 		#Causal convolutions are only padded on the left side
@@ -322,7 +325,7 @@ def incremental_step(self, inputs, convolution_queue=None):
 
 		inputs: [batch_size, time_length, channels] ('NWC')! Channels last!
 		'''
-		return self(inputs, True, convolution_queue)
+		return self(inputs, incremental=True, convolution_queue=convolution_queue)
 
 
 class Conv1D1x1(CausalConv1D):
@@ -370,11 +373,11 @@ def __init__(self, filters,
 
 	def call(self, inputs, incremental=False, convolution_queue=None):
 		#Call parent class call function
-		return super(Conv1D1x1, self).call(inputs, incremental, convolution_queue)
+		return super(Conv1D1x1, self).call(inputs, incremental=incremental, convolution_queue=convolution_queue)
 
 	def incremental_step(self, inputs, unused_queue=None):
 		#Call parent class incremental function
-		output, _ = self(inputs, True, unused_queue) #Drop unused queue
+		output = self(inputs, incremental=True, convolution_queue=unused_queue) #Drop unused queue
 		return output
 
 
@@ -449,11 +452,11 @@ def set_mode(self, is_training):
 
 
 	def call(self, x, c=None, g=None):
-		x, s, _ = self.step(x, c, g, False)
-		return (x, s)
+		x, s, _ = self.step(x, c=c, g=g, is_incremental=False)
+		return [x, s]
 
 	def incremental_step(self, x, c=None, g=None, queue=None):
-		return self.step(x, c, g, True, queue=queue)
+		return self.step(x, c=c, g=g, is_incremental=True, queue=queue)
 
 	def step(self, x, c, g, is_incremental, queue=None):
 		'''

wavenet_vocoder/models/wavenet.py

@@ -402,7 +402,9 @@ def initialize(self, y, c, g, input_lengths, x=None, synthesis_length=None):
 		self.variables = tf.trainable_variables()
 		n_vars = np.sum([np.prod(v.shape) for v in tf.trainable_variables()])
 		log('  Receptive Field:           ({} samples / {:.1f} ms)'.format(self.receptive_field, self.receptive_field / hparams.sample_rate * 1000.))
-		log('  WaveNet Parameters:        {:.3f} Million.'.format(np.sum([np.prod(v.get_shape().as_list()) for v in self.variables]) / 1_000_000))
+
+		#1_000_000 is causing syntax problems for some people?! Python please :)
+		log('  WaveNet Parameters:        {:.3f} Million.'.format(np.sum([np.prod(v.get_shape().as_list()) for v in self.variables]) / 1000000))
 
 		self.ema = tf.train.ExponentialMovingAverage(decay=hparams.wavenet_ema_decay)