feat: update pretrained model url, raise error if there are no files …

…to preprocess, shuffle files consistently (#715)

* feat: update urls for pretrained models
* fix(preprocess_resample): raise if no files found
* fix(preprocess_flist_config): consistent random shuffling

BREAKING CHANGE: default preset changed

README.md

@@ -32,7 +32,7 @@
   <img src="https://img.shields.io/pypi/l/so-vits-svc-fork.svg?style=flat-square" alt="License">
 </p>
 
-A fork of [`so-vits-svc`](https://github.com/svc-develop-team/so-vits-svc) with **realtime support** and **greatly improved interface**. Based on branch `4.0` (v1) and the models are compatible.
+A fork of [`so-vits-svc`](https://github.com/svc-develop-team/so-vits-svc) with **realtime support** and **greatly improved interface**. Based on branch `4.0` (v1) (or `4.1`) and the models are compatible.
 
 ## Features not available in the original repo
 

src/so_vits_svc_fork/__main__.py

@@ -546,7 +546,7 @@ def pre_resample(
     "-t",
     "--config-type",
     type=click.Choice([x.stem for x in CONFIG_TEMPLATE_DIR.rglob("*.json")]),
-    default="so-vits-svc-4.0v1-legacy",
+    default="so-vits-svc-4.0v1",
     help="config type",
 )
 def pre_config(

src/so_vits_svc_fork/preprocessing/config_templates/quickvc.json

@@ -68,7 +68,11 @@
     "type_": "ms-istft",
     "gen_istft_n_fft": 16,
     "gen_istft_hop_size": 4,
-    "subbands": 4
+    "subbands": 4,
+    "pretrained": {
+      "D_0.pth": "https://huggingface.co/datasets/ms903/sovits4.0-768vec-layer12/resolve/main/sovits_768l12_pre_large_320k/clean_D_320000.pth",
+      "G_0.pth": "https://huggingface.co/datasets/ms903/sovits4.0-768vec-layer12/resolve/main/sovits_768l12_pre_large_320k/clean_G_320000.pth"
+    }
   },
   "spk": {}
 }

src/so_vits_svc_fork/preprocessing/config_templates/so-vits-svc-4.0v1-legacy.json

@@ -59,7 +59,11 @@
     "use_spectral_norm": false,
     "gin_channels": 256,
     "ssl_dim": 256,
-    "n_speakers": 200
+    "n_speakers": 200,
+    "pretrained": {
+      "D_0.pth": "https://huggingface.co/therealvul/so-vits-svc-4.0-init/resolve/main/D_0.pth",
+      "G_0.pth": "https://huggingface.co/therealvul/so-vits-svc-4.0-init/resolve/main/G_0.pth"
+    }
   },
   "spk": {}
 }

src/so_vits_svc_fork/preprocessing/config_templates/so-vits-svc-4.0v1.json

@@ -61,7 +61,11 @@
     "gin_channels": 256,
     "ssl_dim": 768,
     "n_speakers": 200,
-    "type_": "hifi-gan"
+    "type_": "hifi-gan",
+    "pretrained": {
+      "D_0.pth": "https://huggingface.co/datasets/ms903/sovits4.0-768vec-layer12/resolve/main/sovits_768l12_pre_large_320k/clean_D_320000.pth",
+      "G_0.pth": "https://huggingface.co/datasets/ms903/sovits4.0-768vec-layer12/resolve/main/sovits_768l12_pre_large_320k/clean_G_320000.pth"
+    }
   },
   "spk": {}
 }

src/so_vits_svc_fork/preprocessing/preprocess_flist_config.py

@@ -5,8 +5,8 @@
 from copy import deepcopy
 from logging import getLogger
 from pathlib import Path
-from random import shuffle
 
+import numpy as np
 from librosa import get_duration
 from tqdm import tqdm
 
@@ -32,6 +32,7 @@ def preprocess_config(
     test = []
     spk_dict = {}
     spk_id = 0
+    random = np.random.RandomState(1234)
     for speaker in os.listdir(input_dir):
         spk_dict[speaker] = spk_id
         spk_id += 1
@@ -41,7 +42,7 @@ def preprocess_config(
                 LOG.warning(f"skip {path} because it is too short.")
                 continue
             paths.append(path)
-        shuffle(paths)
+        random.shuffle(paths)
         if len(paths) <= 4:
             raise ValueError(
                 f"too few files in {input_dir / speaker} (expected at least 5)."

src/so_vits_svc_fork/preprocessing/preprocess_resample.py

@@ -103,9 +103,11 @@ def preprocess_resample(
     output_dir = Path(output_dir)
     """Preprocess audio files in input_dir and save them to output_dir."""
 
-    in_paths = []
     out_paths = []
-    for in_path in input_dir.rglob("*.*"):
+    in_paths = list(input_dir.rglob("*.*"))
+    if not in_paths:
+        raise ValueError(f"No audio files found in {input_dir}")
+    for in_path in in_paths:
         in_path_relative = in_path.relative_to(input_dir)
         if not in_path.is_absolute() and is_relative_to(
             in_path, Path("dataset_raw") / "44k"
@@ -125,7 +127,6 @@ def preprocess_resample(
         out_path = output_dir / speaker_name / file_name
         out_path = _get_unique_filename(out_path, out_paths)
         out_path.parent.mkdir(parents=True, exist_ok=True)
-        in_paths.append(in_path)
         out_paths.append(out_path)
 
     in_and_out_paths = list(zip(in_paths, out_paths))

src/so_vits_svc_fork/train.py

@@ -10,6 +10,7 @@
 import lightning.pytorch as pl
 import torch
 from lightning.pytorch.accelerators import MPSAccelerator, TPUAccelerator
+from lightning.pytorch.callbacks import DeviceStatsMonitor
 from lightning.pytorch.loggers import TensorBoardLogger
 from lightning.pytorch.strategies.ddp import DDPStrategy
 from lightning.pytorch.tuner import Tuner
@@ -73,7 +74,16 @@ def train(
     model_path = Path(model_path)
 
     hparams = utils.get_backup_hparams(config_path, model_path)
-    utils.ensure_pretrained_model(model_path, hparams.model.get("type_", "hifi-gan"))
+    utils.ensure_pretrained_model(
+        model_path,
+        hparams.model.get(
+            "pretrained",
+            {
+                "D_0.pth": "https://huggingface.co/therealvul/so-vits-svc-4.0-init/resolve/main/D_0.pth",
+                "G_0.pth": "https://huggingface.co/therealvul/so-vits-svc-4.0-init/resolve/main/G_0.pth",
+            },
+        ),
+    )
 
     datamodule = VCDataModule(hparams)
     strategy = (
@@ -100,7 +110,8 @@ def train(
         if hparams.train.get("bf16_run", False)
         else 32,
         strategy=strategy,
-        callbacks=[pl.callbacks.RichProgressBar()] if not is_notebook() else None,
+        callbacks=([pl.callbacks.RichProgressBar()] if not is_notebook() else [])
+        + [DeviceStatsMonitor()],
         benchmark=True,
         enable_checkpointing=False,
     )

src/so_vits_svc_fork/utils.py

@@ -113,9 +113,31 @@ def download_file(
 
 
 def ensure_pretrained_model(
-    folder_path: Path | str, type_: str, **tqdm_kwargs: Any
+    folder_path: Path | str, type_: str | dict[str, str], **tqdm_kwargs: Any
 ) -> tuple[Path, ...] | None:
     folder_path = Path(folder_path)
+
+    # new code
+    if not isinstance(type_, str):
+        try:
+            Parallel(n_jobs=len(type_))(
+                [
+                    delayed(download_file)(
+                        url,
+                        folder_path / filename,
+                        position=i,
+                        skip_if_exists=True,
+                        **tqdm_kwargs,
+                    )
+                    for i, (filename, url) in enumerate(type_.items())
+                ]
+            )
+            return tuple(folder_path / filename for filename in type_.values())
+        except Exception as e:
+            LOG.error(f"Failed to download {type_}")
+            LOG.exception(e)
+
+    # old code
     models_candidates = PRETRAINED_MODEL_URLS.get(type_, None)
     if models_candidates is None:
         LOG.warning(f"Unknown pretrained model type: {type_}")
@@ -133,8 +155,8 @@ def ensure_pretrained_model(
             )
             return tuple(paths)
         except Exception as e:
+            LOG.error(f"Failed to download {model_urls}")
             LOG.exception(e)
-    return
 
 
 class HubertModelWithFinalProj(HubertModel):