complex conditioning

audio_data.py

@@ -185,7 +185,7 @@ def __init__(self, *args, **kwargs):
  self.conditioning_channels = conditioning_channels
 
  print("Conditioning captures time scales ranging from " + str(self.min_conditioning_breadth) + " to " \
- + str(self.min_conditioning_breadth * self.conditioning_channels) + " seconds")
+ + str(self.min_conditioning_breadth * self.conditioning_channels // 2) + " seconds")
 
  super().__init__(*args, **kwargs)
 
@@ -199,8 +199,10 @@ def conditioning(self, file_index, position_in_file, item_length):
  x = np.linspace(phase_start, phase_start + phase_length, num=conditioning_count)
  conditioning = np.zeros((self.conditioning_channels, conditioning_count))
 
- for c in range(self.conditioning_channels):
- conditioning[c, :] = np.cos(x * ((c + 1) / self.conditioning_channels))
+ frequency_count = self.conditioning_channels // 2
+ for c in range(frequency_count):
+ conditioning[2 * c, :] = np.cos(x * ((c + 1) / frequency_count))
+ conditioning[2 * c + 1, :] = np.sin(x * ((c + 1) / frequency_count))
 
  file_encoding = np.zeros((len(self.files), conditioning_count))
  file_encoding[file_index, :] = 1.

wavenet_model.py

@@ -383,7 +383,7 @@ def __init__(self, args_dict):
  self.file_encoding_layers.append(nn.Conv1d(in_channels=self.file_encoding_channels[i],
  out_channels=self.file_encoding_channels[i + 1],
  kernel_size=1,
- bias=False))
+ bias=self.use_bias))
 
  self.conditioning_layers = nn.ModuleList()
  self.file_conditioning_cross_layers = nn.ModuleList()
@@ -396,7 +396,7 @@ def __init__(self, args_dict):
  self.file_conditioning_cross_layers.append(nn.Conv1d(in_channels=self.file_encoding_channels[-1],
  out_channels=self.conditioning_channels[i],
  kernel_size=1,
- bias=False))
+ bias=self.use_bias))
 
  def activation_unit_init(self):
  super().activation_unit_init()