⚗️ | Added some stupid ways for training + some makeup

data.py

@@ -12,12 +12,15 @@ import AudioUtils
 class AudioDataset(Dataset):
     audio_sample_rates = [11025]
 
-    def __init__(self, input_dir, clip_length=16384):
+    def __init__(self, input_dir, clip_length: int = 8000, normalize: bool = True):
+        self.clip_length = clip_length
+        self.normalize = normalize
+
         input_files = [
-            os.path.join(root, f)
-            for root, _, files in os.walk(input_dir)
-            for f in files
-            if f.endswith(".wav") or f.endswith(".mp3") or f.endswith(".flac")
+            os.path.join(input_dir, f)
+            for f in os.listdir(input_dir)
+            if os.path.isfile(os.path.join(input_dir, f))
+            and f.lower().endswith((".wav", ".mp3", ".flac"))
         ]
 
         data = []
@@ -25,14 +28,15 @@ class AudioDataset(Dataset):
             input_files, desc=f"Processing {len(input_files)} audio file(s)"
         ):
             decoder = decoders.AudioDecoder(audio_clip)
-
             decoded_samples = decoder.get_all_samples()
-            audio = decoded_samples.data
+
+            audio = decoded_samples.data.float()  # ensure float32
             original_sample_rate = decoded_samples.sample_rate
 
             audio = AudioUtils.stereo_tensor_to_mono(audio)
+            if normalize:
+                audio = AudioUtils.normalize(audio)
 
-            # Generate low-quality audio with random downsampling
             mangled_sample_rate = random.choice(self.audio_sample_rates)
             resample_transform_low = torchaudio.transforms.Resample(
                 original_sample_rate, mangled_sample_rate
@@ -41,25 +45,27 @@ class AudioDataset(Dataset):
                 mangled_sample_rate, original_sample_rate
             )
 
-            low_audio = resample_transform_low(audio)
-            low_audio = resample_transform_high(low_audio)
+            low_audio = resample_transform_high(resample_transform_low(audio))
 
             splitted_high_quality_audio = AudioUtils.split_audio(audio, clip_length)
+            splitted_low_quality_audio = AudioUtils.split_audio(low_audio, clip_length)
+
+            if not splitted_high_quality_audio or not splitted_low_quality_audio:
+                continue  # skip empty or invalid clips
+
             splitted_high_quality_audio[-1] = AudioUtils.pad_tensor(
                 splitted_high_quality_audio[-1], clip_length
             )
-
-            splitted_low_quality_audio = AudioUtils.split_audio(low_audio, clip_length)
             splitted_low_quality_audio[-1] = AudioUtils.pad_tensor(
                 splitted_low_quality_audio[-1], clip_length
             )
 
-            for high_quality_sample, low_quality_sample in zip(
+            for high_quality_data, low_quality_data in zip(
                 splitted_high_quality_audio, splitted_low_quality_audio
             ):
                 data.append(
                     (
-                        (high_quality_sample, low_quality_sample),
+                        (high_quality_data, low_quality_data),
                         (original_sample_rate, mangled_sample_rate),
                     )
                 )