feat(preprocessing): allow nested dataset

.gitignore

@@ -141,7 +141,7 @@ cython_debug/
 
 # additional files
 tests/**/*.wav
-!tests/dataset_raw/34j/*.wav
+!tests/dataset_raw/44k/34j/**/*.wav
 tests/**/*.npy
 tests/**/*.pt
 tests/**/*.txt

README.md

@@ -87,7 +87,7 @@ svc --model-path <model-path> source.wav
 
 #### Local
 
-Place your dataset like `dataset_raw/{speaker_id}/{wav_file}.wav` and run:
+Place your dataset like `dataset_raw/{speaker_id}/**/{wav_file}.{any_format}` (subfolders are acceptable) and run:
 
 ```shell
 svc pre-resample

src/so_vits_svc_fork/__main__.py

@@ -84,7 +84,7 @@ def wrap_text(
 def cli():
     """so-vits-svc allows any folder structure for training data.
     However, the following folder structure is recommended.\n
-        When training: dataset_raw/{speaker_name}/{wav_name}.wav\n
+        When training: dataset_raw/{speaker_name}/**/{wav_name}.{any_format}\n
         When inference: configs/44k/config.json, logs/44k/G_XXXX.pth\n
     If the folder structure is followed, you DO NOT NEED TO SPECIFY model path, config path, etc.
     (The latest model will be automatically loaded.)\n
@@ -476,7 +476,8 @@ def clean():
     folders = ["dataset", "filelists", "logs"]
     if pyip.inputYesNo(f"Are you sure you want to delete files in {folders}?") == "yes":
         for folder in folders:
-            shutil.rmtree(folder)
+            if Path(folder).exists():
+                shutil.rmtree(folder)
         LOG.info("Cleaned up files")
     else:
         LOG.info("Aborted")

src/so_vits_svc_fork/preprocess_resample.py

@@ -2,6 +2,7 @@
 
 from logging import getLogger
 from pathlib import Path
+from typing import Iterable
 
 import audioread.exceptions
 import librosa
@@ -22,6 +23,18 @@
 # - save as 16-bit wav file
 
 
+def _get_unique_filename(path: Path, existing_paths: Iterable[Path]) -> Path:
+    """Return a unique path by appending a number to the original path."""
+    if path not in existing_paths:
+        return path
+    i = 1
+    while True:
+        new_path = path.parent / f"{path.stem}_{i}{path.suffix}"
+        if new_path not in existing_paths:
+            return new_path
+        i += 1
+
+
 def preprocess_resample(
     input_dir: Path | str, output_dir: Path | str, sampling_rate: int
 ) -> None:
@@ -54,11 +67,21 @@ def preprocess_one(input_path: Path, output_path: Path) -> None:
         audio /= max(audio.max(), -audio.min())
         soundfile.write(output_path, audio, samplerate=sampling_rate, subtype="PCM_16")
 
-    in_and_out_paths = []
+    in_paths = []
+    out_paths = []
     for in_path in input_dir.rglob("*.*"):
-        out_path = output_dir / in_path.relative_to(input_dir).with_suffix(".wav")
+        in_path_relative = in_path.relative_to(input_dir)
+        if len(in_path_relative.parts) < 2:
+            continue
+        speaker_name = in_path_relative.parts[0]
+        file_name = in_path_relative.with_suffix(".wav").name
+        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_and_out_paths.append((in_path, out_path))
+        in_paths.append(in_path)
+        out_paths.append(out_path)
+
+    in_and_out_paths = list(zip(in_paths, out_paths))
 
     with tqdm_joblib(desc="Preprocessing", total=len(in_and_out_paths)):
         Parallel(n_jobs=-1)(delayed(preprocess_one)(*args) for args in in_and_out_paths)