Refactor to support PyTorch 2.0 and Lightning 2.0

augmentation/spec_stretch.py

@@ -31,20 +31,20 @@ def process_item(self, item: dict, key_shift=0., speed=1., replace_spk_id=None)
                 aug_item['wav_fn'], keyshift=key_shift, speed=speed
             )
 
-        aug_item['mel'] = torch.from_numpy(mel)
+        aug_item['mel'] = mel
 
         if speed != 1. or hparams.get('use_speed_embed', False):
             aug_item['length'] = mel.shape[0]
             aug_item['speed'] = int(np.round(hparams['hop_size'] * speed)) / hparams['hop_size'] # real speed
             aug_item['seconds'] /= aug_item['speed']
             aug_item['ph_dur'] /= aug_item['speed']
             aug_item['mel2ph'] = get_mel2ph_torch(
-                self.lr, aug_item['ph_dur'], aug_item['length'], hparams, device=self.device
-            )
+                self.lr, torch.from_numpy(aug_item['ph_dur']), aug_item['length'], hparams, device=self.device
+            ).cpu().numpy()
             f0, _, _ = get_pitch_parselmouth(
                 wav, aug_item['length'], hparams, speed=speed, interp_uv=hparams['interp_uv']
             )
-            aug_item['f0'] = torch.from_numpy(f0)
+            aug_item['f0'] = f0.astype(np.float32)
 
         if key_shift != 0. or hparams.get('use_key_shift_embed', False):
             aug_item['key_shift'] = key_shift

basics/base_dataset.py

@@ -9,12 +9,9 @@
 class BaseDataset(Dataset):
     '''
         Base class for datasets.
-        1. *ordered_indices*:
-            if self.shuffle == True, shuffle the indices;
-            if self.sort_by_len == True, sort data by length;
-        2. *sizes*:
+        1. *sizes*:
             clipped length if "max_frames" is set;
-        3. *num_tokens*:
+        2. *num_tokens*:
             unclipped length.
 
         Subclasses should define:
@@ -23,11 +20,9 @@ class BaseDataset(Dataset):
         2. *__getitem__*:
             the index function.
     '''
-    def __init__(self, shuffle):
+    def __init__(self):
         super().__init__()
         self.hparams = hparams
-        self.shuffle = shuffle
-        self.sort_by_len = hparams['sort_by_len']
         self.sizes = None
 
     @property
@@ -50,19 +45,3 @@ def size(self, index):
         """Return an example's size as a float or tuple. This value is used when
         filtering a dataset with ``--max-positions``."""
         return self._sizes[index]
-
-    def ordered_indices(self):
-        """Return an ordered list of indices. Batches will be constructed based
-        on this order."""
-        if self.shuffle:
-            indices = np.random.permutation(len(self))
-            if self.sort_by_len:
-                indices = indices[np.argsort(np.array(self._sizes)[indices], kind='mergesort')]
-                # 先random, 然后稳定排序, 保证排序后同长度的数据顺序是依照random permutation的 (被其随机打乱).
-        else:
-            indices = np.arange(len(self))
-        return indices
-
-    @property
-    def num_workers(self):
-        return int(hparams.get('ds_workers', os.getenv('NUM_WORKERS', 0)))