change dilation back to 2

LJ_speech.json

@@ -55,11 +55,11 @@
   "trainer": {
     "steps": 580000,
     "save_dir": "/media/ycy/86A4D88BA4D87F5D/DataSet/waveglow-result/",
-    "save_freq": 20000,
+    "save_freq": 50,
     "verbosity": 2
   },
   "visualization": {
-    "tensorboardX": false,
-    "log_dir": "saved/runs"
+    "tensorboardX": true,
+    "log_dir": "/media/ycy/86A4D88BA4D87F5D/DataSet/waveglow-result/runs"
   }
 }

data_loader/data_loaders.py

@@ -103,22 +103,21 @@ def get_nframes(info_str):
                 else:
                     assert f_obj.samplerate == self.sr
         self.file_lengths = np.array(self.file_lengths)
-        self.boundaries = np.cumsum(self.file_lengths)
+        self.boundaries = np.cumsum(self.file_lengths) / (self.file_lengths.sum() - 1)
 
         # normalization value based on each file
-        # will updated online
-        self.max_values = np.ones_like(self.boundaries) * 0.01
+        # will updated on the fily
+        self.max_values = np.zeros_like(self.boundaries)
 
     def __len__(self):
         return self.size
 
     def __getitem__(self, index):
-        index = random.randint(0, self.boundaries[-1] - 1)
-        index = np.digitize(index, self.boundaries)
+        index = np.digitize(random.uniform(0, 1), self.boundaries)
         f, length = self.files[index], self.file_lengths[index]
-        pos = random.randint(0, length - self.segment - 1)
+        pos = random.randint(0, length - 1)
         f.seek(pos)
-        x = f.read(self.segment, dtype='float32', always_2d=True).mean(1)
+        x = f.read(self.segment, dtype='float32', always_2d=True, fill_value=0.).mean(1)
         max_abs = np.abs(x).max()
         if max_abs > self.max_values[index]:
             self.max_values[index] = max_abs
@@ -149,7 +148,9 @@ def __init__(self, steps, data_dir, batch_size, num_workers, **kwargs):
 
 
 if __name__ == '__main__':
-    loader = RandomWaveFileLoader(100, '/media/ycy/86A4D88BA4D87F5D/DataSet/LJSpeech-1.1/wavs', 64, 0, segment=16000)
+    loader = RandomWaveFileLoader(100, '/media/ycy/86A4D88BA4D87F5D/DataSet/LJSpeech-1.1/wavs', 1, 0, segment=16000)
 
+    import matplotlib.pyplot as plt
     for x in loader:
-        print(x[0, :10])
+        plt.plot(x[0].numpy())
+        plt.show()

model/model.py

@@ -51,7 +51,7 @@ def __init__(self,
                  bias=False,
                  zero_init=True):
         super().__init__()
-        dilations = radix ** torch.arange(depth)
+        dilations = 2 ** torch.arange(depth)
         self.dilations = dilations.tolist()
         self.in_chs = in_channels
         self.res_chs = residual_channels

trainer/trainer.py

@@ -30,13 +30,17 @@ def train(self):
             data = data.to(self.device)
 
             self.optimizer.zero_grad()
-            *output, mels = self.model(data)
-            loss = self.loss(*output)
+            z, logdet, mels = self.model(data)
+            loss = self.loss(z, logdet)
             loss.backward()
             self.optimizer.step()
 
             self.writer.set_step(step)
             self.writer.add_scalar('loss', loss.item())
+            self.writer.add_scalar('log_determinant', logdet.mean().item())
+            self.writer.add_scalar('z_mean', z.mean().item())
+            self.writer.add_scalar('z_std', z.std().item())
+            self.writer.add_scalar('max_memory_allocated', torch.cuda.max_memory_allocated() / (1024 ** 2))
 
             if self.verbosity >= 2 and step % self.log_step == 0:
                 self.logger.info('Train Step: [{}/{} ({:.0f}%)] Loss: {:.6f}'.format(
@@ -45,12 +49,14 @@ def train(self):
                     100.0 * step / self.steps,
                     loss.item()))
                 #self.writer.add_image('input', make_grid(data.cpu(), nrow=8, normalize=True))
-                self.writer.add_image('input mel-spectrum', mels[0].cpu(), dataformats='HW')
-                with torch.no_grad():
-                    z = torch.randn_like(output[0][:1]) * output[0].std()
-                    x, _ = self.model.inverse(z, mels[:1])
-                    torch.clamp(x, -1, 1, out=x)
-                self.writer.add_audio('reconstruct audio', x.cpu(), sample_rate=self.model.sr)
+                mel_spec = mels[0].cpu()
+                mel_spec -= mel_spec.min()
+                mel_spec /= mel_spec.max()
+                self.writer.add_image('input_mel-spectrum', mel_spec.flip(0), dataformats='HW')
+
+                x = self.model.infer(mels[0], z[0].std().item())
+                torch.clamp(x, -1, 1, out=x)
+                self.writer.add_audio('reconstruct_audio', x.cpu()[None, :], sample_rate=self.model.sr)
 
             if self.lr_scheduler is not None:
                 self.lr_scheduler.step()