some minor fixes

inference/ds_variance.py

@@ -241,14 +241,15 @@ def preprocess_input(
         batch['midi'] = ph_midi
 
         if load_pitch:
+            # Interpolate unvoiced parts before resampling.
             f0 = resample_align_curve(
-                np.array(param['f0_seq'].split(), np.float32),
+                interp_f0(np.array(param['f0_seq'].split(), np.float32))[0],
                 original_timestep=float(param['f0_timestep']),
                 target_timestep=self.timestep,
                 align_length=T_s
             )
             batch['pitch'] = torch.from_numpy(
-                librosa.hz_to_midi(interp_f0(f0)[0]).astype(np.float32)
+                librosa.hz_to_midi(f0).astype(np.float32)
             ).to(self.device)[None]
 
         if self.model.predict_dur:

preprocessing/acoustic_binarizer.py

@@ -337,7 +337,7 @@ def arrange_data_augmentation(self, data_iterator):
                     aug_list.append(aug_task)
                 elif aug_type == 1:
                     aug_task = {
-                        'name': aug_item,
+                        'name': aug_item['name'],
                         'func': aug_item['func'],
                         'kwargs': deepcopy(aug_item['kwargs'])
                     }

preprocessing/variance_binarizer.py

@@ -314,14 +314,21 @@ def process_item(self, item_name, meta_data, binarization_args):
         if self.prefer_ds:
             f0_seq = self.load_attr_from_ds(ds_id, name, 'f0_seq', idx=ds_seg_idx)
             if f0_seq is not None:
+                f0_timestep = float(self.load_attr_from_ds(ds_id, name, 'f0_timestep', idx=ds_seg_idx))
+                # Interpolate unvoiced parts before resampling.
+                f0_points, uv_points = interp_f0(np.array(f0_seq.split(), np.float32))
                 f0 = resample_align_curve(
-                    np.array(f0_seq.split(), np.float32),
-                    original_timestep=float(self.load_attr_from_ds(ds_id, name, 'f0_timestep', idx=ds_seg_idx)),
+                    f0_points,
+                    original_timestep=f0_timestep,
                     target_timestep=self.timestep,
                     align_length=length
                 )
-                uv = f0 == 0
-                f0, _ = interp_f0(f0, uv)
+                uv = resample_align_curve(
+                    uv_points.astype(np.float32),
+                    original_timestep=f0_timestep,
+                    target_timestep=self.timestep,
+                    align_length=length
+                ) > 0.5
         if f0 is None:
             f0, uv = pitch_extractor.get_pitch(
                 waveform, samplerate=hparams['audio_sample_rate'], length=length,

utils/infer_utils.py

@@ -39,17 +39,11 @@ def trans_key(raw_data, key):
 
 
 def resample_align_curve(points: np.ndarray, original_timestep: float, target_timestep: float, align_length: int):
-    t_max = (len(points) - 1) * original_timestep
     curve_interp = np.interp(
-        np.arange(0, t_max, target_timestep),
+        np.arange(align_length) * target_timestep,
         original_timestep * np.arange(len(points)),
         points
     ).astype(points.dtype)
-    delta_l = align_length - len(curve_interp)
-    if delta_l < 0:
-        curve_interp = curve_interp[:align_length]
-    elif delta_l > 0:
-        curve_interp = np.concatenate((curve_interp, np.full(delta_l, fill_value=curve_interp[-1])), axis=0)
     return curve_interp