Fixed a bug with the encoding in the vocoder. Trying a bigger vocoder network

wave-rnn/inference_demo.py

@@ -3,7 +3,7 @@ from vocoder.vocoder_dataset import VocoderDataset
 from vocoder import inference
 from vocoder import audio
 import numpy as np
-from vocoder.params import print_params, model_name
+from vocoder.params import print_params, model_name, use_mu_law
 
 print_params()
 
@@ -28,6 +28,8 @@ for i in sorted(np.random.choice(len(dataset), n_samples)):
     out_pred_fpath = fileio.join(gen_path, "%s_%d_pred.wav" % (model_name, i))
     
     wav_gt = audio.restore_signal(wav_gt)
+    if use_mu_law:
+        wav_gt = audio.expand_signal(wav_gt)
     wav_pred = inference.infer_waveform(mel, normalize=False)   # The dataloader already normalizes
 
     audio.save_wav(out_pred_fpath, wav_pred)

wave-rnn/train.py

@@ -34,8 +34,8 @@ model_dir = 'checkpoints'
 fileio.ensure_dir(model_dir)
 model_fpath = fileio.join(model_dir, model_name + '.pt')
 
-# data_path = "../data/Synthesizer"
-data_path = "E:/Datasets/Synthesizer"
+data_path = "../data/Synthesizer"
+# data_path = "E:/Datasets/Synthesizer"
 gen_path = 'model_outputs'
 fileio.ensure_dir(gen_path)
 
@@ -125,6 +125,6 @@ if __name__ == '__main__':
             print('<saved>')
             
     optimizer = optim.Adam(model.parameters())
-    train(model, optimizer, epochs=60, batch_size=64, classes=2 ** bits,
+    train(model, optimizer, epochs=100, batch_size=100, classes=2 ** bits,
           seq_len=seq_len, step=step, lr=1e-4)
     
\ No newline at end of file

wave-rnn/vocoder/params.py

@@ -17,38 +17,38 @@ use_mu_law = True
 # Minimum number of mel frames below which samples are discarded for training
 min_n_frames = 10
 
-## Model parameters
-model_name = 'mu_law'
-# Number of bits for the encoding. Higher means higher quality output but longer training time 
-# and training memory required.
-bits = 9
-pad = 2
-seq_len = hop_length * 5
-mel_win = seq_len // hop_length + 2 * pad
-rnn_dims = 512
-fc_dims = 512
-upsample_factors = (5, 5, 8)
-feat_dims = 80
-compute_dims = 128
-res_out_dims = 128
-res_blocks = 10
-
 # ## Model parameters
-# model_name = 'mu_law_big'
+# model_name = 'mu_law'
 # # Number of bits for the encoding. Higher means higher quality output but longer training time 
 # # and training memory required.
 # bits = 9
 # pad = 2
 # seq_len = hop_length * 5
 # mel_win = seq_len // hop_length + 2 * pad
-# rnn_dims = 768
-# fc_dims = 768
+# rnn_dims = 512
+# fc_dims = 512
 # upsample_factors = (5, 5, 8)
 # feat_dims = 80
-# compute_dims = 196
-# res_out_dims = 196
+# compute_dims = 128
+# res_out_dims = 128
 # res_blocks = 10
 
+## Model parameters
+model_name = 'mu_law_big'
+# Number of bits for the encoding. Higher means higher quality output but longer training time 
+# and training memory required.
+bits = 9
+pad = 2
+seq_len = hop_length * 5
+mel_win = seq_len // hop_length + 2 * pad
+rnn_dims = 768
+fc_dims = 768
+upsample_factors = (5, 5, 8)
+feat_dims = 80
+compute_dims = 196
+res_out_dims = 196
+res_blocks = 10
+
 
 def print_params():
     for param_name in sorted(globals()):

wave-rnn/vocoder/vocoder_dataset.py

@@ -22,6 +22,8 @@ class VocoderDataset(Dataset):
         
         # Load the wav and quantize it
         wav = np.load(wav_path)
+        if use_mu_law:
+            wav = audio.compand_signal(wav)
         quant = audio.quantize_signal(wav)
         
         # Load the mel spectrogram and adjust its range to [0, 1]